linux-stable/fs/pnode.c

641 lines
16 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* linux/fs/pnode.c
*
* (C) Copyright IBM Corporation 2005.
* Author : Ram Pai (linuxram@us.ibm.com)
*/
#include <linux/mnt_namespace.h>
#include <linux/mount.h>
#include <linux/fs.h>
#include <linux/nsproxy.h>
#include <uapi/linux/mount.h>
#include "internal.h"
#include "pnode.h"
/* return the next shared peer mount of @p */
static inline struct mount *next_peer(struct mount *p)
{
return list_entry(p->mnt_share.next, struct mount, mnt_share);
}
static inline struct mount *first_slave(struct mount *p)
{
return list_entry(p->mnt_slave_list.next, struct mount, mnt_slave);
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static inline struct mount *last_slave(struct mount *p)
{
return list_entry(p->mnt_slave_list.prev, struct mount, mnt_slave);
}
static inline struct mount *next_slave(struct mount *p)
{
return list_entry(p->mnt_slave.next, struct mount, mnt_slave);
}
static struct mount *get_peer_under_root(struct mount *mnt,
struct mnt_namespace *ns,
const struct path *root)
{
struct mount *m = mnt;
do {
/* Check the namespace first for optimization */
if (m->mnt_ns == ns && is_path_reachable(m, m->mnt.mnt_root, root))
return m;
m = next_peer(m);
} while (m != mnt);
return NULL;
}
/*
* Get ID of closest dominating peer group having a representative
* under the given root.
*
* Caller must hold namespace_sem
*/
int get_dominating_id(struct mount *mnt, const struct path *root)
{
struct mount *m;
for (m = mnt->mnt_master; m != NULL; m = m->mnt_master) {
struct mount *d = get_peer_under_root(m, mnt->mnt_ns, root);
if (d)
return d->mnt_group_id;
}
return 0;
}
static int do_make_slave(struct mount *mnt)
{
struct mount *master, *slave_mnt;
if (list_empty(&mnt->mnt_share)) {
if (IS_MNT_SHARED(mnt)) {
mnt_release_group_id(mnt);
CLEAR_MNT_SHARED(mnt);
}
master = mnt->mnt_master;
if (!master) {
struct list_head *p = &mnt->mnt_slave_list;
while (!list_empty(p)) {
slave_mnt = list_first_entry(p,
struct mount, mnt_slave);
list_del_init(&slave_mnt->mnt_slave);
slave_mnt->mnt_master = NULL;
}
return 0;
}
} else {
struct mount *m;
/*
* slave 'mnt' to a peer mount that has the
* same root dentry. If none is available then
* slave it to anything that is available.
*/
for (m = master = next_peer(mnt); m != mnt; m = next_peer(m)) {
if (m->mnt.mnt_root == mnt->mnt.mnt_root) {
master = m;
break;
}
}
list_del_init(&mnt->mnt_share);
mnt->mnt_group_id = 0;
CLEAR_MNT_SHARED(mnt);
}
list_for_each_entry(slave_mnt, &mnt->mnt_slave_list, mnt_slave)
slave_mnt->mnt_master = master;
list_move(&mnt->mnt_slave, &master->mnt_slave_list);
list_splice(&mnt->mnt_slave_list, master->mnt_slave_list.prev);
INIT_LIST_HEAD(&mnt->mnt_slave_list);
mnt->mnt_master = master;
return 0;
}
/*
* vfsmount lock must be held for write
*/
void change_mnt_propagation(struct mount *mnt, int type)
{
if (type == MS_SHARED) {
set_mnt_shared(mnt);
return;
}
do_make_slave(mnt);
if (type != MS_SLAVE) {
list_del_init(&mnt->mnt_slave);
mnt->mnt_master = NULL;
if (type == MS_UNBINDABLE)
mnt->mnt.mnt_flags |= MNT_UNBINDABLE;
else
mnt->mnt.mnt_flags &= ~MNT_UNBINDABLE;
}
}
/*
* get the next mount in the propagation tree.
* @m: the mount seen last
* @origin: the original mount from where the tree walk initiated
*
* Note that peer groups form contiguous segments of slave lists.
* We rely on that in get_source() to be able to find out if
* vfsmount found while iterating with propagation_next() is
* a peer of one we'd found earlier.
*/
static struct mount *propagation_next(struct mount *m,
struct mount *origin)
{
/* are there any slaves of this mount? */
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
return first_slave(m);
while (1) {
struct mount *master = m->mnt_master;
if (master == origin->mnt_master) {
struct mount *next = next_peer(m);
return (next == origin) ? NULL : next;
} else if (m->mnt_slave.next != &master->mnt_slave_list)
return next_slave(m);
/* back at master */
m = master;
}
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static struct mount *skip_propagation_subtree(struct mount *m,
struct mount *origin)
{
/*
* Advance m such that propagation_next will not return
* the slaves of m.
*/
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
m = last_slave(m);
return m;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
static struct mount *next_group(struct mount *m, struct mount *origin)
{
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
while (1) {
while (1) {
struct mount *next;
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
return first_slave(m);
next = next_peer(m);
if (m->mnt_group_id == origin->mnt_group_id) {
if (next == origin)
return NULL;
} else if (m->mnt_slave.next != &next->mnt_slave)
break;
m = next;
}
/* m is the last peer */
while (1) {
struct mount *master = m->mnt_master;
if (m->mnt_slave.next != &master->mnt_slave_list)
return next_slave(m);
m = next_peer(master);
if (master->mnt_group_id == origin->mnt_group_id)
break;
if (master->mnt_slave.next == &m->mnt_slave)
break;
m = master;
}
if (m == origin)
return NULL;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/* all accesses are serialized by namespace_sem */
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
static struct mount *last_dest, *first_source, *last_source, *dest_master;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
static struct hlist_head *list;
fs: allow to mount beneath top mount Various distributions are adding or are in the process of adding support for system extensions and in the future configuration extensions through various tools. A more detailed explanation on system and configuration extensions can be found on the manpage which is listed below at [1]. System extension images may – dynamically at runtime — extend the /usr/ and /opt/ directory hierarchies with additional files. This is particularly useful on immutable system images where a /usr/ and/or /opt/ hierarchy residing on a read-only file system shall be extended temporarily at runtime without making any persistent modifications. When one or more system extension images are activated, their /usr/ and /opt/ hierarchies are combined via overlayfs with the same hierarchies of the host OS, and the host /usr/ and /opt/ overmounted with it ("merging"). When they are deactivated, the mount point is disassembled — again revealing the unmodified original host version of the hierarchy ("unmerging"). Merging thus makes the extension's resources suddenly appear below the /usr/ and /opt/ hierarchies as if they were included in the base OS image itself. Unmerging makes them disappear again, leaving in place only the files that were shipped with the base OS image itself. System configuration images are similar but operate on directories containing system or service configuration. On nearly all modern distributions mount propagation plays a crucial role and the rootfs of the OS is a shared mount in a peer group (usually with peer group id 1): TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID / / ext4 shared:1 29 1 On such systems all services and containers run in a separate mount namespace and are pivot_root()ed into their rootfs. A separate mount namespace is almost always used as it is the minimal isolation mechanism services have. But usually they are even much more isolated up to the point where they almost become indistinguishable from containers. Mount propagation again plays a crucial role here. The rootfs of all these services is a slave mount to the peer group of the host rootfs. This is done so the service will receive mount propagation events from the host when certain files or directories are updated. In addition, the rootfs of each service, container, and sandbox is also a shared mount in its separate peer group: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID / / ext4 shared:24 master:1 71 47 For people not too familiar with mount propagation, the master:1 means that this is a slave mount to peer group 1. Which as one can see is the host rootfs as indicated by shared:1 above. The shared:24 indicates that the service rootfs is a shared mount in a separate peer group with peer group id 24. A service may run other services. Such nested services will also have a rootfs mount that is a slave to the peer group of the outer service rootfs mount. For containers things are just slighly different. A container's rootfs isn't a slave to the service's or host rootfs' peer group. The rootfs mount of a container is simply a shared mount in its own peer group: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID /home/ubuntu/debian-tree / ext4 shared:99 61 60 So whereas services are isolated OS components a container is treated like a separate world and mount propagation into it is restricted to a single well known mount that is a slave to the peer group of the shared mount /run on the host: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID /propagate/debian-tree /run/host/incoming tmpfs master:5 71 68 Here, the master:5 indicates that this mount is a slave to the peer group with peer group id 5. This allows to propagate mounts into the container and served as a workaround for not being able to insert mounts into mount namespaces directly. But the new mount api does support inserting mounts directly. For the interested reader the blogpost in [2] might be worth reading where I explain the old and the new approach to inserting mounts into mount namespaces. Containers of course, can themselves be run as services. They often run full systems themselves which means they again run services and containers with the exact same propagation settings explained above. The whole system is designed so that it can be easily updated, including all services in various fine-grained ways without having to enter every single service's mount namespace which would be prohibitively expensive. The mount propagation layout has been carefully chosen so it is possible to propagate updates for system extensions and configurations from the host into all services. The simplest model to update the whole system is to mount on top of /usr, /opt, or /etc on the host. The new mount on /usr, /opt, or /etc will then propagate into every service. This works cleanly the first time. However, when the system is updated multiple times it becomes necessary to unmount the first update on /opt, /usr, /etc and then propagate the new update. But this means, there's an interval where the old base system is accessible. This has to be avoided to protect against downgrade attacks. The vfs already exposes a mechanism to userspace whereby mounts can be mounted beneath an existing mount. Such mounts are internally referred to as "tucked". The patch series exposes the ability to mount beneath a top mount through the new MOVE_MOUNT_BENEATH flag for the move_mount() system call. This allows userspace to seamlessly upgrade mounts. After this series the only thing that will have changed is that mounting beneath an existing mount can be done explicitly instead of just implicitly. Today, there are two scenarios where a mount can be mounted beneath an existing mount instead of on top of it: (1) When a service or container is started in a new mount namespace and pivot_root()s into its new rootfs. The way this is done is by mounting the new rootfs beneath the old rootfs: fd_newroot = open("/var/lib/machines/fedora", ...); fd_oldroot = open("/", ...); fchdir(fd_newroot); pivot_root(".", "."); After the pivot_root(".", ".") call the new rootfs is mounted beneath the old rootfs which can then be unmounted to reveal the underlying mount: fchdir(fd_oldroot); umount2(".", MNT_DETACH); Since pivot_root() moves the caller into a new rootfs no mounts must be propagated out of the new rootfs as a consequence of the pivot_root() call. Thus, the mounts cannot be shared. (2) When a mount is propagated to a mount that already has another mount mounted on the same dentry. The easiest example for this is to create a new mount namespace. The following commands will create a mount namespace where the rootfs mount / will be a slave to the peer group of the host rootfs / mount's peer group. IOW, it will receive propagation from the host: mount --make-shared / unshare --mount --propagation=slave Now a new mount on the /mnt dentry in that mount namespace is created. (As it can be confusing it should be spelled out that the tmpfs mount on the /mnt dentry that was just created doesn't propagate back to the host because the rootfs mount / of the mount namespace isn't a peer of the host rootfs.): mount -t tmpfs tmpfs /mnt TARGET SOURCE FSTYPE PROPAGATION └─/mnt tmpfs tmpfs Now another terminal in the host mount namespace can observe that the mount indeed hasn't propagated back to into the host mount namespace. A new mount can now be created on top of the /mnt dentry with the rootfs mount / as its parent: mount --bind /opt /mnt TARGET SOURCE FSTYPE PROPAGATION └─/mnt /dev/sda2[/opt] ext4 shared:1 The mount namespace that was created earlier can now observe that the bind mount created on the host has propagated into it: TARGET SOURCE FSTYPE PROPAGATION └─/mnt /dev/sda2[/opt] ext4 master:1 └─/mnt tmpfs tmpfs But instead of having been mounted on top of the tmpfs mount at the /mnt dentry the /opt mount has been mounted on top of the rootfs mount at the /mnt dentry. And the tmpfs mount has been remounted on top of the propagated /opt mount at the /opt dentry. So in other words, the propagated mount has been mounted beneath the preexisting mount in that mount namespace. Mount namespaces make this easy to illustrate but it's also easy to mount beneath an existing mount in the same mount namespace (The following example assumes a shared rootfs mount / with peer group id 1): mount --bind /opt /opt TARGET SOURCE FSTYPE MNT_ID PARENT_ID PROPAGATION └─/opt /dev/sda2[/opt] ext4 188 29 shared:1 If another mount is mounted on top of the /opt mount at the /opt dentry: mount --bind /tmp /opt The following clunky mount tree will result: TARGET SOURCE FSTYPE MNT_ID PARENT_ID PROPAGATION └─/opt /dev/sda2[/tmp] ext4 405 29 shared:1 └─/opt /dev/sda2[/opt] ext4 188 405 shared:1 └─/opt /dev/sda2[/tmp] ext4 404 188 shared:1 The /tmp mount is mounted beneath the /opt mount and another copy is mounted on top of the /opt mount. This happens because the rootfs / and the /opt mount are shared mounts in the same peer group. When the new /tmp mount is supposed to be mounted at the /opt dentry then the /tmp mount first propagates to the root mount at the /opt dentry. But there already is the /opt mount mounted at the /opt dentry. So the old /opt mount at the /opt dentry will be mounted on top of the new /tmp mount at the /tmp dentry, i.e. @opt->mnt_parent is @tmp and @opt->mnt_mountpoint is /tmp (Note that @opt->mnt_root is /opt which is what shows up as /opt under SOURCE). So again, a mount will be mounted beneath a preexisting mount. (Fwiw, a few iterations of mount --bind /opt /opt in a loop on a shared rootfs is a good example of what could be referred to as mount explosion.) The main point is that such mounts allows userspace to umount a top mount and reveal an underlying mount. So for example, umounting the tmpfs mount on /mnt that was created in example (1) using mount namespaces reveals the /opt mount which was mounted beneath it. In (2) where a mount was mounted beneath the top mount in the same mount namespace unmounting the top mount would unmount both the top mount and the mount beneath. In the process the original mount would be remounted on top of the rootfs mount / at the /opt dentry again. This again, is a result of mount propagation only this time it's umount propagation. However, this can be avoided by simply making the parent mount / of the @opt mount a private or slave mount. Then the top mount and the original mount can be unmounted to reveal the mount beneath. These two examples are fairly arcane and are merely added to make it clear how mount propagation has effects on current and future features. More common use-cases will just be things like: mount -t btrfs /dev/sdA /mnt mount -t xfs /dev/sdB --beneath /mnt umount /mnt after which we'll have updated from a btrfs filesystem to a xfs filesystem without ever revealing the underlying mountpoint. The crux is that the proposed mechanism already exists and that it is so powerful as to cover cases where mounts are supposed to be updated with new versions. Crucially, it offers an important flexibility. Namely that updates to a system may either be forced or can be delayed and the umount of the top mount be left to a service if it is a cooperative one. This adds a new flag to move_mount() that allows to explicitly move a beneath the top mount adhering to the following semantics: * Mounts cannot be mounted beneath the rootfs. This restriction encompasses the rootfs but also chroots via chroot() and pivot_root(). To mount a mount beneath the rootfs or a chroot, pivot_root() can be used as illustrated above. * The source mount must be a private mount to force the kernel to allocate a new, unused peer group id. This isn't a required restriction but a voluntary one. It avoids repeating a semantical quirk that already exists today. If bind mounts which already have a peer group id are inserted into mount trees that have the same peer group id this can cause a lot of mount propagation events to be generated (For example, consider running mount --bind /opt /opt in a loop where the parent mount is a shared mount.). * Avoid getting rid of the top mount in the kernel. Cooperative services need to be able to unmount the top mount themselves. This also avoids a good deal of additional complexity. The umount would have to be propagated which would be another rather expensive operation. So namespace_lock() and lock_mount_hash() would potentially have to be held for a long time for both a mount and umount propagation. That should be avoided. * The path to mount beneath must be mounted and attached. * The top mount and its parent must be in the caller's mount namespace and the caller must be able to mount in that mount namespace. * The caller must be able to unmount the top mount to prove that they could reveal the underlying mount. * The propagation tree is calculated based on the destination mount's parent mount and the destination mount's mountpoint on the parent mount. Of course, if the parent of the destination mount and the destination mount are shared mounts in the same peer group and the mountpoint of the new mount to be mounted is a subdir of their ->mnt_root then both will receive a mount of /opt. That's probably easier to understand with an example. Assuming a standard shared rootfs /: mount --bind /opt /opt mount --bind /tmp /opt will cause the same mount tree as: mount --bind /opt /opt mount --beneath /tmp /opt because both / and /opt are shared mounts/peers in the same peer group and the /opt dentry is a subdirectory of both the parent's and the child's ->mnt_root. If a mount tree like that is created it almost always is an accident or abuse of mount propagation. Realistically what most people probably mean in this scenarios is: mount --bind /opt /opt mount --make-private /opt mount --make-shared /opt This forces the allocation of a new separate peer group for the /opt mount. Aferwards a mount --bind or mount --beneath actually makes sense as the / and /opt mount belong to different peer groups. Before that it's likely just confusion about what the user wanted to achieve. * Refuse MOVE_MOUNT_BENEATH if: (1) the @mnt_from has been overmounted in between path resolution and acquiring @namespace_sem when locking @mnt_to. This avoids the proliferation of shadow mounts. (2) if @to_mnt is moved to a different mountpoint while acquiring @namespace_sem to lock @to_mnt. (3) if @to_mnt is unmounted while acquiring @namespace_sem to lock @to_mnt. (4) if the parent of the target mount propagates to the target mount at the same mountpoint. This would mean mounting @mnt_from on @mnt_to->mnt_parent and then propagating a copy @c of @mnt_from onto @mnt_to. This defeats the whole purpose of mounting @mnt_from beneath @mnt_to. (5) if the parent mount @mnt_to->mnt_parent propagates to @mnt_from at the same mountpoint. If @mnt_to->mnt_parent propagates to @mnt_from this would mean propagating a copy @c of @mnt_from on top of @mnt_from. Afterwards @mnt_from would be mounted on top of @mnt_to->mnt_parent and @mnt_to would be unmounted from @mnt->mnt_parent and remounted on @mnt_from. But since @c is already mounted on @mnt_from, @mnt_to would ultimately be remounted on top of @c. Afterwards, @mnt_from would be covered by a copy @c of @mnt_from and @c would be covered by @mnt_from itself. This defeats the whole purpose of mounting @mnt_from beneath @mnt_to. Cases (1) to (3) are required as they deal with races that would cause bugs or unexpected behavior for users. Cases (4) and (5) refuse semantical quirks that would not be a bug but would cause weird mount trees to be created. While they can already be created via other means (mount --bind /opt /opt x n) there's no reason to repeat past mistakes in new features. Link: https://man7.org/linux/man-pages/man8/systemd-sysext.8.html [1] Link: https://brauner.io/2023/02/28/mounting-into-mount-namespaces.html [2] Link: https://github.com/flatcar/sysext-bakery Link: https://fedoraproject.org/wiki/Changes/Unified_Kernel_Support_Phase_1 Link: https://fedoraproject.org/wiki/Changes/Unified_Kernel_Support_Phase_2 Link: https://github.com/systemd/systemd/pull/26013 Reviewed-by: Seth Forshee (DigitalOcean) <sforshee@kernel.org> Message-Id: <20230202-fs-move-mount-replace-v4-4-98f3d80d7eaa@kernel.org> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-05-03 11:18:42 +00:00
static inline bool peers(const struct mount *m1, const struct mount *m2)
{
return m1->mnt_group_id == m2->mnt_group_id && m1->mnt_group_id;
}
static int propagate_one(struct mount *m, struct mountpoint *dest_mp)
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
{
struct mount *child;
int type;
/* skip ones added by this propagate_mnt() */
if (IS_MNT_NEW(m))
return 0;
/* skip if mountpoint isn't covered by it */
if (!is_subdir(dest_mp->m_dentry, m->mnt.mnt_root))
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
return 0;
if (peers(m, last_dest)) {
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
type = CL_MAKE_SHARED;
} else {
struct mount *n, *p;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
bool done;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
for (n = m; ; n = p) {
p = n->mnt_master;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
if (p == dest_master || IS_MNT_MARKED(p))
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
break;
}
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
do {
struct mount *parent = last_source->mnt_parent;
pnode: terminate at peers of source The propagate_mnt() function handles mount propagation when creating mounts and propagates the source mount tree @source_mnt to all applicable nodes of the destination propagation mount tree headed by @dest_mnt. Unfortunately it contains a bug where it fails to terminate at peers of @source_mnt when looking up copies of the source mount that become masters for copies of the source mount tree mounted on top of slaves in the destination propagation tree causing a NULL dereference. Once the mechanics of the bug are understood it's easy to trigger. Because of unprivileged user namespaces it is available to unprivileged users. While fixing this bug we've gotten confused multiple times due to unclear terminology or missing concepts. So let's start this with some clarifications: * The terms "master" or "peer" denote a shared mount. A shared mount belongs to a peer group. * A peer group is a set of shared mounts that propagate to each other. They are identified by a peer group id. The peer group id is available in @shared_mnt->mnt_group_id. Shared mounts within the same peer group have the same peer group id. The peers in a peer group can be reached via @shared_mnt->mnt_share. * The terms "slave mount" or "dependent mount" denote a mount that receives propagation from a peer in a peer group. IOW, shared mounts may have slave mounts and slave mounts have shared mounts as their master. Slave mounts of a given peer in a peer group are listed on that peers slave list available at @shared_mnt->mnt_slave_list. * The term "master mount" denotes a mount in a peer group. IOW, it denotes a shared mount or a peer mount in a peer group. The term "master mount" - or "master" for short - is mostly used when talking in the context of slave mounts that receive propagation from a master mount. A master mount of a slave identifies the closest peer group a slave mount receives propagation from. The master mount of a slave can be identified via @slave_mount->mnt_master. Different slaves may point to different masters in the same peer group. * Multiple peers in a peer group can have non-empty ->mnt_slave_lists. Non-empty ->mnt_slave_lists of peers don't intersect. Consequently, to ensure all slave mounts of a peer group are visited the ->mnt_slave_lists of all peers in a peer group have to be walked. * Slave mounts point to a peer in the closest peer group they receive propagation from via @slave_mnt->mnt_master (see above). Together with these peers they form a propagation group (see below). The closest peer group can thus be identified through the peer group id @slave_mnt->mnt_master->mnt_group_id of the peer/master that a slave mount receives propagation from. * A shared-slave mount is a slave mount to a peer group pg1 while also a peer in another peer group pg2. IOW, a peer group may receive propagation from another peer group. If a peer group pg1 is a slave to another peer group pg2 then all peers in peer group pg1 point to the same peer in peer group pg2 via ->mnt_master. IOW, all peers in peer group pg1 appear on the same ->mnt_slave_list. IOW, they cannot be slaves to different peer groups. * A pure slave mount is a slave mount that is a slave to a peer group but is not a peer in another peer group. * A propagation group denotes the set of mounts consisting of a single peer group pg1 and all slave mounts and shared-slave mounts that point to a peer in that peer group via ->mnt_master. IOW, all slave mounts such that @slave_mnt->mnt_master->mnt_group_id is equal to @shared_mnt->mnt_group_id. The concept of a propagation group makes it easier to talk about a single propagation level in a propagation tree. For example, in propagate_mnt() the immediate peers of @dest_mnt and all slaves of @dest_mnt's peer group form a propagation group propg1. So a shared-slave mount that is a slave in propg1 and that is a peer in another peer group pg2 forms another propagation group propg2 together with all slaves that point to that shared-slave mount in their ->mnt_master. * A propagation tree refers to all mounts that receive propagation starting from a specific shared mount. For example, for propagate_mnt() @dest_mnt is the start of a propagation tree. The propagation tree ecompasses all mounts that receive propagation from @dest_mnt's peer group down to the leafs. With that out of the way let's get to the actual algorithm. We know that @dest_mnt is guaranteed to be a pure shared mount or a shared-slave mount. This is guaranteed by a check in attach_recursive_mnt(). So propagate_mnt() will first propagate the source mount tree to all peers in @dest_mnt's peer group: for (n = next_peer(dest_mnt); n != dest_mnt; n = next_peer(n)) { ret = propagate_one(n); if (ret) goto out; } Notice, that the peer propagation loop of propagate_mnt() doesn't propagate @dest_mnt itself. @dest_mnt is mounted directly in attach_recursive_mnt() after we propagated to the destination propagation tree. The mount that will be mounted on top of @dest_mnt is @source_mnt. This copy was created earlier even before we entered attach_recursive_mnt() and doesn't concern us a lot here. It's just important to notice that when propagate_mnt() is called @source_mnt will not yet have been mounted on top of @dest_mnt. Thus, @source_mnt->mnt_parent will either still point to @source_mnt or - in the case @source_mnt is moved and thus already attached - still to its former parent. For each peer @m in @dest_mnt's peer group propagate_one() will create a new copy of the source mount tree and mount that copy @child on @m such that @child->mnt_parent points to @m after propagate_one() returns. propagate_one() will stash the last destination propagation node @m in @last_dest and the last copy it created for the source mount tree in @last_source. Hence, if we call into propagate_one() again for the next destination propagation node @m, @last_dest will point to the previous destination propagation node and @last_source will point to the previous copy of the source mount tree and mounted on @last_dest. Each new copy of the source mount tree is created from the previous copy of the source mount tree. This will become important later. The peer loop in propagate_mnt() is straightforward. We iterate through the peers copying and updating @last_source and @last_dest as we go through them and mount each copy of the source mount tree @child on a peer @m in @dest_mnt's peer group. After propagate_mnt() handled the peers in @dest_mnt's peer group propagate_mnt() will propagate the source mount tree down the propagation tree that @dest_mnt's peer group propagates to: for (m = next_group(dest_mnt, dest_mnt); m; m = next_group(m, dest_mnt)) { /* everything in that slave group */ n = m; do { ret = propagate_one(n); if (ret) goto out; n = next_peer(n); } while (n != m); } The next_group() helper will recursively walk the destination propagation tree, descending into each propagation group of the propagation tree. The important part is that it takes care to propagate the source mount tree to all peers in the peer group of a propagation group before it propagates to the slaves to those peers in the propagation group. IOW, it creates and mounts copies of the source mount tree that become masters before it creates and mounts copies of the source mount tree that become slaves to these masters. It is important to remember that propagating the source mount tree to each mount @m in the destination propagation tree simply means that we create and mount new copies @child of the source mount tree on @m such that @child->mnt_parent points to @m. Since we know that each node @m in the destination propagation tree headed by @dest_mnt's peer group will be overmounted with a copy of the source mount tree and since we know that the propagation properties of each copy of the source mount tree we create and mount at @m will mostly mirror the propagation properties of @m. We can use that information to create and mount the copies of the source mount tree that become masters before their slaves. The easy case is always when @m and @last_dest are peers in a peer group of a given propagation group. In that case we know that we can simply copy @last_source without having to figure out what the master for the new copy @child of the source mount tree needs to be as we've done that in a previous call to propagate_one(). The hard case is when we're dealing with a slave mount or a shared-slave mount @m in a destination propagation group that we need to create and mount a copy of the source mount tree on. For each propagation group in the destination propagation tree we propagate the source mount tree to we want to make sure that the copies @child of the source mount tree we create and mount on slaves @m pick an ealier copy of the source mount tree that we mounted on a master @m of the destination propagation group as their master. This is a mouthful but as far as we can tell that's the core of it all. But, if we keep track of the masters in the destination propagation tree @m we can use the information to find the correct master for each copy of the source mount tree we create and mount at the slaves in the destination propagation tree @m. Let's walk through the base case as that's still fairly easy to grasp. If we're dealing with the first slave in the propagation group that @dest_mnt is in then we don't yet have marked any masters in the destination propagation tree. We know the master for the first slave to @dest_mnt's peer group is simple @dest_mnt. So we expect this algorithm to yield a copy of the source mount tree that was mounted on a peer in @dest_mnt's peer group as the master for the copy of the source mount tree we want to mount at the first slave @m: for (n = m; ; n = p) { p = n->mnt_master; if (p == dest_master || IS_MNT_MARKED(p)) break; } For the first slave we walk the destination propagation tree all the way up to a peer in @dest_mnt's peer group. IOW, the propagation hierarchy can be walked by walking up the @mnt->mnt_master hierarchy of the destination propagation tree @m. We will ultimately find a peer in @dest_mnt's peer group and thus ultimately @dest_mnt->mnt_master. Btw, here the assumption we listed at the beginning becomes important. Namely, that peers in a peer group pg1 that are slaves in another peer group pg2 appear on the same ->mnt_slave_list. IOW, all slaves who are peers in peer group pg1 point to the same peer in peer group pg2 via their ->mnt_master. Otherwise the termination condition in the code above would be wrong and next_group() would be broken too. So the first iteration sets: n = m; p = n->mnt_master; such that @p now points to a peer or @dest_mnt itself. We walk up one more level since we don't have any marked mounts. So we end up with: n = dest_mnt; p = dest_mnt->mnt_master; If @dest_mnt's peer group is not slave to another peer group then @p is now NULL. If @dest_mnt's peer group is a slave to another peer group then @p now points to @dest_mnt->mnt_master points which is a master outside the propagation tree we're dealing with. Now we need to figure out the master for the copy of the source mount tree we're about to create and mount on the first slave of @dest_mnt's peer group: do { struct mount *parent = last_source->mnt_parent; if (last_source == first_source) break; done = parent->mnt_master == p; if (done && peers(n, parent)) break; last_source = last_source->mnt_master; } while (!done); We know that @last_source->mnt_parent points to @last_dest and @last_dest is the last peer in @dest_mnt's peer group we propagated to in the peer loop in propagate_mnt(). Consequently, @last_source is the last copy we created and mount on that last peer in @dest_mnt's peer group. So @last_source is the master we want to pick. We know that @last_source->mnt_parent->mnt_master points to @last_dest->mnt_master. We also know that @last_dest->mnt_master is either NULL or points to a master outside of the destination propagation tree and so does @p. Hence: done = parent->mnt_master == p; is trivially true in the base condition. We also know that for the first slave mount of @dest_mnt's peer group that @last_dest either points @dest_mnt itself because it was initialized to: last_dest = dest_mnt; at the beginning of propagate_mnt() or it will point to a peer of @dest_mnt in its peer group. In both cases it is guaranteed that on the first iteration @n and @parent are peers (Please note the check for peers here as that's important.): if (done && peers(n, parent)) break; So, as we expected, we select @last_source, which referes to the last copy of the source mount tree we mounted on the last peer in @dest_mnt's peer group, as the master of the first slave in @dest_mnt's peer group. The rest is taken care of by clone_mnt(last_source, ...). We'll skip over that part otherwise this becomes a blogpost. At the end of propagate_mnt() we now mark @m->mnt_master as the first master in the destination propagation tree that is distinct from @dest_mnt->mnt_master. IOW, we mark @dest_mnt itself as a master. By marking @dest_mnt or one of it's peers we are able to easily find it again when we later lookup masters for other copies of the source mount tree we mount copies of the source mount tree on slaves @m to @dest_mnt's peer group. This, in turn allows us to find the master we selected for the copies of the source mount tree we mounted on master in the destination propagation tree again. The important part is to realize that the code makes use of the fact that the last copy of the source mount tree stashed in @last_source was mounted on top of the previous destination propagation node @last_dest. What this means is that @last_source allows us to walk the destination propagation hierarchy the same way each destination propagation node @m does. If we take @last_source, which is the copy of @source_mnt we have mounted on @last_dest in the previous iteration of propagate_one(), then we know @last_source->mnt_parent points to @last_dest but we also know that as we walk through the destination propagation tree that @last_source->mnt_master will point to an earlier copy of the source mount tree we mounted one an earlier destination propagation node @m. IOW, @last_source->mnt_parent will be our hook into the destination propagation tree and each consecutive @last_source->mnt_master will lead us to an earlier propagation node @m via @last_source->mnt_master->mnt_parent. Hence, by walking up @last_source->mnt_master, each of which is mounted on a node that is a master @m in the destination propagation tree we can also walk up the destination propagation hierarchy. So, for each new destination propagation node @m we use the previous copy of @last_source and the fact it's mounted on the previous propagation node @last_dest via @last_source->mnt_master->mnt_parent to determine what the master of the new copy of @last_source needs to be. The goal is to find the _closest_ master that the new copy of the source mount tree we are about to create and mount on a slave @m in the destination propagation tree needs to pick. IOW, we want to find a suitable master in the propagation group. As the propagation structure of the source mount propagation tree we create mirrors the propagation structure of the destination propagation tree we can find @m's closest master - i.e., a marked master - which is a peer in the closest peer group that @m receives propagation from. We store that closest master of @m in @p as before and record the slave to that master in @n We then search for this master @p via @last_source by walking up the master hierarchy starting from the last copy of the source mount tree stored in @last_source that we created and mounted on the previous destination propagation node @m. We will try to find the master by walking @last_source->mnt_master and by comparing @last_source->mnt_master->mnt_parent->mnt_master to @p. If we find @p then we can figure out what earlier copy of the source mount tree needs to be the master for the new copy of the source mount tree we're about to create and mount at the current destination propagation node @m. If @last_source->mnt_master->mnt_parent and @n are peers then we know that the closest master they receive propagation from is @last_source->mnt_master->mnt_parent->mnt_master. If not then the closest immediate peer group that they receive propagation from must be one level higher up. This builds on the earlier clarification at the beginning that all peers in a peer group which are slaves of other peer groups all point to the same ->mnt_master, i.e., appear on the same ->mnt_slave_list, of the closest peer group that they receive propagation from. However, terminating the walk has corner cases. If the closest marked master for a given destination node @m cannot be found by walking up the master hierarchy via @last_source->mnt_master then we need to terminate the walk when we encounter @source_mnt again. This isn't an arbitrary termination. It simply means that the new copy of the source mount tree we're about to create has a copy of the source mount tree we created and mounted on a peer in @dest_mnt's peer group as its master. IOW, @source_mnt is the peer in the closest peer group that the new copy of the source mount tree receives propagation from. We absolutely have to stop @source_mnt because @last_source->mnt_master either points outside the propagation hierarchy we're dealing with or it is NULL because @source_mnt isn't a shared-slave. So continuing the walk past @source_mnt would cause a NULL dereference via @last_source->mnt_master->mnt_parent. And so we have to stop the walk when we encounter @source_mnt again. One scenario where this can happen is when we first handled a series of slaves of @dest_mnt's peer group and then encounter peers in a new peer group that is a slave to @dest_mnt's peer group. We handle them and then we encounter another slave mount to @dest_mnt that is a pure slave to @dest_mnt's peer group. That pure slave will have a peer in @dest_mnt's peer group as its master. Consequently, the new copy of the source mount tree will need to have @source_mnt as it's master. So we walk the propagation hierarchy all the way up to @source_mnt based on @last_source->mnt_master. So terminate on @source_mnt, easy peasy. Except, that the check misses something that the rest of the algorithm already handles. If @dest_mnt has peers in it's peer group the peer loop in propagate_mnt(): for (n = next_peer(dest_mnt); n != dest_mnt; n = next_peer(n)) { ret = propagate_one(n); if (ret) goto out; } will consecutively update @last_source with each previous copy of the source mount tree we created and mounted at the previous peer in @dest_mnt's peer group. So after that loop terminates @last_source will point to whatever copy of the source mount tree was created and mounted on the last peer in @dest_mnt's peer group. Furthermore, if there is even a single additional peer in @dest_mnt's peer group then @last_source will __not__ point to @source_mnt anymore. Because, as we mentioned above, @dest_mnt isn't even handled in this loop but directly in attach_recursive_mnt(). So it can't even accidently come last in that peer loop. So the first time we handle a slave mount @m of @dest_mnt's peer group the copy of the source mount tree we create will make the __last copy of the source mount tree we created and mounted on the last peer in @dest_mnt's peer group the master of the new copy of the source mount tree we create and mount on the first slave of @dest_mnt's peer group__. But this means that the termination condition that checks for @source_mnt is wrong. The @source_mnt cannot be found anymore by propagate_one(). Instead it will find the last copy of the source mount tree we created and mounted for the last peer of @dest_mnt's peer group again. And that is a peer of @source_mnt not @source_mnt itself. IOW, we fail to terminate the loop correctly and ultimately dereference @last_source->mnt_master->mnt_parent. When @source_mnt's peer group isn't slave to another peer group then @last_source->mnt_master is NULL causing the splat below. For example, assume @dest_mnt is a pure shared mount and has three peers in its peer group: =================================================================================== mount-id mount-parent-id peer-group-id =================================================================================== (@dest_mnt) mnt_master[216] 309 297 shared:216 \ (@source_mnt) mnt_master[218]: 609 609 shared:218 (1) mnt_master[216]: 607 605 shared:216 \ (P1) mnt_master[218]: 624 607 shared:218 (2) mnt_master[216]: 576 574 shared:216 \ (P2) mnt_master[218]: 625 576 shared:218 (3) mnt_master[216]: 545 543 shared:216 \ (P3) mnt_master[218]: 626 545 shared:218 After this sequence has been processed @last_source will point to (P3), the copy generated for the third peer in @dest_mnt's peer group we handled. So the copy of the source mount tree (P4) we create and mount on the first slave of @dest_mnt's peer group: =================================================================================== mount-id mount-parent-id peer-group-id =================================================================================== mnt_master[216] 309 297 shared:216 / / (S0) mnt_slave 483 481 master:216 \ \ (P3) mnt_master[218] 626 545 shared:218 \ / \/ (P4) mnt_slave 627 483 master:218 will pick the last copy of the source mount tree (P3) as master, not (S0). When walking the propagation hierarchy via @last_source's master hierarchy we encounter (P3) but not (S0), i.e., @source_mnt. We can fix this in multiple ways: (1) By setting @last_source to @source_mnt after we processed the peers in @dest_mnt's peer group right after the peer loop in propagate_mnt(). (2) By changing the termination condition that relies on finding exactly @source_mnt to finding a peer of @source_mnt. (3) By only moving @last_source when we actually venture into a new peer group or some clever variant thereof. The first two options are minimally invasive and what we want as a fix. The third option is more intrusive but something we'd like to explore in the near future. This passes all LTP tests and specifically the mount propagation testsuite part of it. It also holds up against all known reproducers of this issues. Final words. First, this is a clever but __worringly__ underdocumented algorithm. There isn't a single detailed comment to be found in next_group(), propagate_one() or anywhere else in that file for that matter. This has been a giant pain to understand and work through and a bug like this is insanely difficult to fix without a detailed understanding of what's happening. Let's not talk about the amount of time that was sunk into fixing this. Second, all the cool kids with access to unshare --mount --user --map-root --propagation=unchanged are going to have a lot of fun. IOW, triggerable by unprivileged users while namespace_lock() lock is held. [ 115.848393] BUG: kernel NULL pointer dereference, address: 0000000000000010 [ 115.848967] #PF: supervisor read access in kernel mode [ 115.849386] #PF: error_code(0x0000) - not-present page [ 115.849803] PGD 0 P4D 0 [ 115.850012] Oops: 0000 [#1] PREEMPT SMP PTI [ 115.850354] CPU: 0 PID: 15591 Comm: mount Not tainted 6.1.0-rc7 #3 [ 115.850851] Hardware name: innotek GmbH VirtualBox/VirtualBox, BIOS VirtualBox 12/01/2006 [ 115.851510] RIP: 0010:propagate_one.part.0+0x7f/0x1a0 [ 115.851924] Code: 75 eb 4c 8b 05 c2 25 37 02 4c 89 ca 48 8b 4a 10 49 39 d0 74 1e 48 3b 81 e0 00 00 00 74 26 48 8b 92 e0 00 00 00 be 01 00 00 00 <48> 8b 4a 10 49 39 d0 75 e2 40 84 f6 74 38 4c 89 05 84 25 37 02 4d [ 115.853441] RSP: 0018:ffffb8d5443d7d50 EFLAGS: 00010282 [ 115.853865] RAX: ffff8e4d87c41c80 RBX: ffff8e4d88ded780 RCX: ffff8e4da4333a00 [ 115.854458] RDX: 0000000000000000 RSI: 0000000000000001 RDI: ffff8e4d88ded780 [ 115.855044] RBP: ffff8e4d88ded780 R08: ffff8e4da4338000 R09: ffff8e4da43388c0 [ 115.855693] R10: 0000000000000002 R11: ffffb8d540158000 R12: ffffb8d5443d7da8 [ 115.856304] R13: ffff8e4d88ded780 R14: 0000000000000000 R15: 0000000000000000 [ 115.856859] FS: 00007f92c90c9800(0000) GS:ffff8e4dfdc00000(0000) knlGS:0000000000000000 [ 115.857531] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 115.858006] CR2: 0000000000000010 CR3: 0000000022f4c002 CR4: 00000000000706f0 [ 115.858598] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 115.859393] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 115.860099] Call Trace: [ 115.860358] <TASK> [ 115.860535] propagate_mnt+0x14d/0x190 [ 115.860848] attach_recursive_mnt+0x274/0x3e0 [ 115.861212] path_mount+0x8c8/0xa60 [ 115.861503] __x64_sys_mount+0xf6/0x140 [ 115.861819] do_syscall_64+0x5b/0x80 [ 115.862117] ? do_faccessat+0x123/0x250 [ 115.862435] ? syscall_exit_to_user_mode+0x17/0x40 [ 115.862826] ? do_syscall_64+0x67/0x80 [ 115.863133] ? syscall_exit_to_user_mode+0x17/0x40 [ 115.863527] ? do_syscall_64+0x67/0x80 [ 115.863835] ? do_syscall_64+0x67/0x80 [ 115.864144] ? do_syscall_64+0x67/0x80 [ 115.864452] ? exc_page_fault+0x70/0x170 [ 115.864775] entry_SYSCALL_64_after_hwframe+0x63/0xcd [ 115.865187] RIP: 0033:0x7f92c92b0ebe [ 115.865480] Code: 48 8b 0d 75 4f 0c 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 a5 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 42 4f 0c 00 f7 d8 64 89 01 48 [ 115.866984] RSP: 002b:00007fff000aa728 EFLAGS: 00000246 ORIG_RAX: 00000000000000a5 [ 115.867607] RAX: ffffffffffffffda RBX: 000055a77888d6b0 RCX: 00007f92c92b0ebe [ 115.868240] RDX: 000055a77888d8e0 RSI: 000055a77888e6e0 RDI: 000055a77888e620 [ 115.868823] RBP: 0000000000000000 R08: 0000000000000000 R09: 0000000000000001 [ 115.869403] R10: 0000000000001000 R11: 0000000000000246 R12: 000055a77888e620 [ 115.869994] R13: 000055a77888d8e0 R14: 00000000ffffffff R15: 00007f92c93e4076 [ 115.870581] </TASK> [ 115.870763] Modules linked in: nft_fib_inet nft_fib_ipv4 nft_fib_ipv6 nft_fib nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject nft_ct nft_chain_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 ip_set rfkill nf_tables nfnetlink qrtr snd_intel8x0 sunrpc snd_ac97_codec ac97_bus snd_pcm snd_timer intel_rapl_msr intel_rapl_common snd vboxguest intel_powerclamp video rapl joydev soundcore i2c_piix4 wmi fuse zram xfs vmwgfx crct10dif_pclmul crc32_pclmul crc32c_intel polyval_clmulni polyval_generic drm_ttm_helper ttm e1000 ghash_clmulni_intel serio_raw ata_generic pata_acpi scsi_dh_rdac scsi_dh_emc scsi_dh_alua dm_multipath [ 115.875288] CR2: 0000000000000010 [ 115.875641] ---[ end trace 0000000000000000 ]--- [ 115.876135] RIP: 0010:propagate_one.part.0+0x7f/0x1a0 [ 115.876551] Code: 75 eb 4c 8b 05 c2 25 37 02 4c 89 ca 48 8b 4a 10 49 39 d0 74 1e 48 3b 81 e0 00 00 00 74 26 48 8b 92 e0 00 00 00 be 01 00 00 00 <48> 8b 4a 10 49 39 d0 75 e2 40 84 f6 74 38 4c 89 05 84 25 37 02 4d [ 115.878086] RSP: 0018:ffffb8d5443d7d50 EFLAGS: 00010282 [ 115.878511] RAX: ffff8e4d87c41c80 RBX: ffff8e4d88ded780 RCX: ffff8e4da4333a00 [ 115.879128] RDX: 0000000000000000 RSI: 0000000000000001 RDI: ffff8e4d88ded780 [ 115.879715] RBP: ffff8e4d88ded780 R08: ffff8e4da4338000 R09: ffff8e4da43388c0 [ 115.880359] R10: 0000000000000002 R11: ffffb8d540158000 R12: ffffb8d5443d7da8 [ 115.880962] R13: ffff8e4d88ded780 R14: 0000000000000000 R15: 0000000000000000 [ 115.881548] FS: 00007f92c90c9800(0000) GS:ffff8e4dfdc00000(0000) knlGS:0000000000000000 [ 115.882234] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 115.882713] CR2: 0000000000000010 CR3: 0000000022f4c002 CR4: 00000000000706f0 [ 115.883314] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 115.883966] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Fixes: f2ebb3a921c1 ("smarter propagate_mnt()") Fixes: 5ec0811d3037 ("propogate_mnt: Handle the first propogated copy being a slave") Cc: <stable@vger.kernel.org> Reported-by: Ditang Chen <ditang.c@gmail.com> Signed-off-by: Seth Forshee (Digital Ocean) <sforshee@kernel.org> Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org> --- If there are no big objections I'll get this to Linus rather sooner than later.
2022-12-17 21:28:40 +00:00
if (peers(last_source, first_source))
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
break;
done = parent->mnt_master == p;
if (done && peers(n, parent))
break;
last_source = last_source->mnt_master;
} while (!done);
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
type = CL_SLAVE;
/* beginning of peer group among the slaves? */
if (IS_MNT_SHARED(m))
type |= CL_MAKE_SHARED;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
child = copy_tree(last_source, last_source->mnt.mnt_root, type);
if (IS_ERR(child))
return PTR_ERR(child);
read_seqlock_excl(&mount_lock);
mnt_set_mountpoint(m, dest_mp, child);
if (m->mnt_master != dest_master)
SET_MNT_MARK(m->mnt_master);
read_sequnlock_excl(&mount_lock);
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
last_dest = m;
last_source = child;
hlist_add_head(&child->mnt_hash, list);
mnt: Add a per mount namespace limit on the number of mounts CAI Qian <caiqian@redhat.com> pointed out that the semantics of shared subtrees make it possible to create an exponentially increasing number of mounts in a mount namespace. mkdir /tmp/1 /tmp/2 mount --make-rshared / for i in $(seq 1 20) ; do mount --bind /tmp/1 /tmp/2 ; done Will create create 2^20 or 1048576 mounts, which is a practical problem as some people have managed to hit this by accident. As such CVE-2016-6213 was assigned. Ian Kent <raven@themaw.net> described the situation for autofs users as follows: > The number of mounts for direct mount maps is usually not very large because of > the way they are implemented, large direct mount maps can have performance > problems. There can be anywhere from a few (likely case a few hundred) to less > than 10000, plus mounts that have been triggered and not yet expired. > > Indirect mounts have one autofs mount at the root plus the number of mounts that > have been triggered and not yet expired. > > The number of autofs indirect map entries can range from a few to the common > case of several thousand and in rare cases up to between 30000 and 50000. I've > not heard of people with maps larger than 50000 entries. > > The larger the number of map entries the greater the possibility for a large > number of active mounts so it's not hard to expect cases of a 1000 or somewhat > more active mounts. So I am setting the default number of mounts allowed per mount namespace at 100,000. This is more than enough for any use case I know of, but small enough to quickly stop an exponential increase in mounts. Which should be perfect to catch misconfigurations and malfunctioning programs. For anyone who needs a higher limit this can be changed by writing to the new /proc/sys/fs/mount-max sysctl. Tested-by: CAI Qian <caiqian@redhat.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-09-28 05:27:17 +00:00
return count_mounts(m->mnt_ns, child);
}
/*
* mount 'source_mnt' under the destination 'dest_mnt' at
* dentry 'dest_dentry'. And propagate that mount to
* all the peer and slave mounts of 'dest_mnt'.
* Link all the new mounts into a propagation tree headed at
* source_mnt. Also link all the new mounts using ->mnt_list
* headed at source_mnt's ->mnt_list
*
* @dest_mnt: destination mount.
* @dest_dentry: destination dentry.
* @source_mnt: source mount.
* @tree_list : list of heads of trees to be attached.
*/
int propagate_mnt(struct mount *dest_mnt, struct mountpoint *dest_mp,
struct mount *source_mnt, struct hlist_head *tree_list)
{
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
struct mount *m, *n;
int ret = 0;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/*
* we don't want to bother passing tons of arguments to
* propagate_one(); everything is serialized by namespace_sem,
* so globals will do just fine.
*/
last_dest = dest_mnt;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
first_source = source_mnt;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
last_source = source_mnt;
list = tree_list;
dest_master = dest_mnt->mnt_master;
/* all peers of dest_mnt, except dest_mnt itself */
for (n = next_peer(dest_mnt); n != dest_mnt; n = next_peer(n)) {
ret = propagate_one(n, dest_mp);
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
if (ret)
goto out;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/* all slave groups */
for (m = next_group(dest_mnt, dest_mnt); m;
m = next_group(m, dest_mnt)) {
/* everything in that slave group */
n = m;
do {
ret = propagate_one(n, dest_mp);
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
if (ret)
goto out;
n = next_peer(n);
} while (n != m);
}
out:
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
read_seqlock_excl(&mount_lock);
hlist_for_each_entry(n, tree_list, mnt_hash) {
m = n->mnt_parent;
if (m->mnt_master != dest_mnt->mnt_master)
CLEAR_MNT_MARK(m->mnt_master);
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
read_sequnlock_excl(&mount_lock);
return ret;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
static struct mount *find_topper(struct mount *mnt)
{
/* If there is exactly one mount covering mnt completely return it. */
struct mount *child;
if (!list_is_singular(&mnt->mnt_mounts))
return NULL;
child = list_first_entry(&mnt->mnt_mounts, struct mount, mnt_child);
if (child->mnt_mountpoint != mnt->mnt.mnt_root)
return NULL;
return child;
}
/*
* return true if the refcount is greater than count
*/
static inline int do_refcount_check(struct mount *mnt, int count)
{
return mnt_get_count(mnt) > count;
}
fs: allow to mount beneath top mount Various distributions are adding or are in the process of adding support for system extensions and in the future configuration extensions through various tools. A more detailed explanation on system and configuration extensions can be found on the manpage which is listed below at [1]. System extension images may – dynamically at runtime — extend the /usr/ and /opt/ directory hierarchies with additional files. This is particularly useful on immutable system images where a /usr/ and/or /opt/ hierarchy residing on a read-only file system shall be extended temporarily at runtime without making any persistent modifications. When one or more system extension images are activated, their /usr/ and /opt/ hierarchies are combined via overlayfs with the same hierarchies of the host OS, and the host /usr/ and /opt/ overmounted with it ("merging"). When they are deactivated, the mount point is disassembled — again revealing the unmodified original host version of the hierarchy ("unmerging"). Merging thus makes the extension's resources suddenly appear below the /usr/ and /opt/ hierarchies as if they were included in the base OS image itself. Unmerging makes them disappear again, leaving in place only the files that were shipped with the base OS image itself. System configuration images are similar but operate on directories containing system or service configuration. On nearly all modern distributions mount propagation plays a crucial role and the rootfs of the OS is a shared mount in a peer group (usually with peer group id 1): TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID / / ext4 shared:1 29 1 On such systems all services and containers run in a separate mount namespace and are pivot_root()ed into their rootfs. A separate mount namespace is almost always used as it is the minimal isolation mechanism services have. But usually they are even much more isolated up to the point where they almost become indistinguishable from containers. Mount propagation again plays a crucial role here. The rootfs of all these services is a slave mount to the peer group of the host rootfs. This is done so the service will receive mount propagation events from the host when certain files or directories are updated. In addition, the rootfs of each service, container, and sandbox is also a shared mount in its separate peer group: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID / / ext4 shared:24 master:1 71 47 For people not too familiar with mount propagation, the master:1 means that this is a slave mount to peer group 1. Which as one can see is the host rootfs as indicated by shared:1 above. The shared:24 indicates that the service rootfs is a shared mount in a separate peer group with peer group id 24. A service may run other services. Such nested services will also have a rootfs mount that is a slave to the peer group of the outer service rootfs mount. For containers things are just slighly different. A container's rootfs isn't a slave to the service's or host rootfs' peer group. The rootfs mount of a container is simply a shared mount in its own peer group: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID /home/ubuntu/debian-tree / ext4 shared:99 61 60 So whereas services are isolated OS components a container is treated like a separate world and mount propagation into it is restricted to a single well known mount that is a slave to the peer group of the shared mount /run on the host: TARGET SOURCE FSTYPE PROPAGATION MNT_ID PARENT_ID /propagate/debian-tree /run/host/incoming tmpfs master:5 71 68 Here, the master:5 indicates that this mount is a slave to the peer group with peer group id 5. This allows to propagate mounts into the container and served as a workaround for not being able to insert mounts into mount namespaces directly. But the new mount api does support inserting mounts directly. For the interested reader the blogpost in [2] might be worth reading where I explain the old and the new approach to inserting mounts into mount namespaces. Containers of course, can themselves be run as services. They often run full systems themselves which means they again run services and containers with the exact same propagation settings explained above. The whole system is designed so that it can be easily updated, including all services in various fine-grained ways without having to enter every single service's mount namespace which would be prohibitively expensive. The mount propagation layout has been carefully chosen so it is possible to propagate updates for system extensions and configurations from the host into all services. The simplest model to update the whole system is to mount on top of /usr, /opt, or /etc on the host. The new mount on /usr, /opt, or /etc will then propagate into every service. This works cleanly the first time. However, when the system is updated multiple times it becomes necessary to unmount the first update on /opt, /usr, /etc and then propagate the new update. But this means, there's an interval where the old base system is accessible. This has to be avoided to protect against downgrade attacks. The vfs already exposes a mechanism to userspace whereby mounts can be mounted beneath an existing mount. Such mounts are internally referred to as "tucked". The patch series exposes the ability to mount beneath a top mount through the new MOVE_MOUNT_BENEATH flag for the move_mount() system call. This allows userspace to seamlessly upgrade mounts. After this series the only thing that will have changed is that mounting beneath an existing mount can be done explicitly instead of just implicitly. Today, there are two scenarios where a mount can be mounted beneath an existing mount instead of on top of it: (1) When a service or container is started in a new mount namespace and pivot_root()s into its new rootfs. The way this is done is by mounting the new rootfs beneath the old rootfs: fd_newroot = open("/var/lib/machines/fedora", ...); fd_oldroot = open("/", ...); fchdir(fd_newroot); pivot_root(".", "."); After the pivot_root(".", ".") call the new rootfs is mounted beneath the old rootfs which can then be unmounted to reveal the underlying mount: fchdir(fd_oldroot); umount2(".", MNT_DETACH); Since pivot_root() moves the caller into a new rootfs no mounts must be propagated out of the new rootfs as a consequence of the pivot_root() call. Thus, the mounts cannot be shared. (2) When a mount is propagated to a mount that already has another mount mounted on the same dentry. The easiest example for this is to create a new mount namespace. The following commands will create a mount namespace where the rootfs mount / will be a slave to the peer group of the host rootfs / mount's peer group. IOW, it will receive propagation from the host: mount --make-shared / unshare --mount --propagation=slave Now a new mount on the /mnt dentry in that mount namespace is created. (As it can be confusing it should be spelled out that the tmpfs mount on the /mnt dentry that was just created doesn't propagate back to the host because the rootfs mount / of the mount namespace isn't a peer of the host rootfs.): mount -t tmpfs tmpfs /mnt TARGET SOURCE FSTYPE PROPAGATION └─/mnt tmpfs tmpfs Now another terminal in the host mount namespace can observe that the mount indeed hasn't propagated back to into the host mount namespace. A new mount can now be created on top of the /mnt dentry with the rootfs mount / as its parent: mount --bind /opt /mnt TARGET SOURCE FSTYPE PROPAGATION └─/mnt /dev/sda2[/opt] ext4 shared:1 The mount namespace that was created earlier can now observe that the bind mount created on the host has propagated into it: TARGET SOURCE FSTYPE PROPAGATION └─/mnt /dev/sda2[/opt] ext4 master:1 └─/mnt tmpfs tmpfs But instead of having been mounted on top of the tmpfs mount at the /mnt dentry the /opt mount has been mounted on top of the rootfs mount at the /mnt dentry. And the tmpfs mount has been remounted on top of the propagated /opt mount at the /opt dentry. So in other words, the propagated mount has been mounted beneath the preexisting mount in that mount namespace. Mount namespaces make this easy to illustrate but it's also easy to mount beneath an existing mount in the same mount namespace (The following example assumes a shared rootfs mount / with peer group id 1): mount --bind /opt /opt TARGET SOURCE FSTYPE MNT_ID PARENT_ID PROPAGATION └─/opt /dev/sda2[/opt] ext4 188 29 shared:1 If another mount is mounted on top of the /opt mount at the /opt dentry: mount --bind /tmp /opt The following clunky mount tree will result: TARGET SOURCE FSTYPE MNT_ID PARENT_ID PROPAGATION └─/opt /dev/sda2[/tmp] ext4 405 29 shared:1 └─/opt /dev/sda2[/opt] ext4 188 405 shared:1 └─/opt /dev/sda2[/tmp] ext4 404 188 shared:1 The /tmp mount is mounted beneath the /opt mount and another copy is mounted on top of the /opt mount. This happens because the rootfs / and the /opt mount are shared mounts in the same peer group. When the new /tmp mount is supposed to be mounted at the /opt dentry then the /tmp mount first propagates to the root mount at the /opt dentry. But there already is the /opt mount mounted at the /opt dentry. So the old /opt mount at the /opt dentry will be mounted on top of the new /tmp mount at the /tmp dentry, i.e. @opt->mnt_parent is @tmp and @opt->mnt_mountpoint is /tmp (Note that @opt->mnt_root is /opt which is what shows up as /opt under SOURCE). So again, a mount will be mounted beneath a preexisting mount. (Fwiw, a few iterations of mount --bind /opt /opt in a loop on a shared rootfs is a good example of what could be referred to as mount explosion.) The main point is that such mounts allows userspace to umount a top mount and reveal an underlying mount. So for example, umounting the tmpfs mount on /mnt that was created in example (1) using mount namespaces reveals the /opt mount which was mounted beneath it. In (2) where a mount was mounted beneath the top mount in the same mount namespace unmounting the top mount would unmount both the top mount and the mount beneath. In the process the original mount would be remounted on top of the rootfs mount / at the /opt dentry again. This again, is a result of mount propagation only this time it's umount propagation. However, this can be avoided by simply making the parent mount / of the @opt mount a private or slave mount. Then the top mount and the original mount can be unmounted to reveal the mount beneath. These two examples are fairly arcane and are merely added to make it clear how mount propagation has effects on current and future features. More common use-cases will just be things like: mount -t btrfs /dev/sdA /mnt mount -t xfs /dev/sdB --beneath /mnt umount /mnt after which we'll have updated from a btrfs filesystem to a xfs filesystem without ever revealing the underlying mountpoint. The crux is that the proposed mechanism already exists and that it is so powerful as to cover cases where mounts are supposed to be updated with new versions. Crucially, it offers an important flexibility. Namely that updates to a system may either be forced or can be delayed and the umount of the top mount be left to a service if it is a cooperative one. This adds a new flag to move_mount() that allows to explicitly move a beneath the top mount adhering to the following semantics: * Mounts cannot be mounted beneath the rootfs. This restriction encompasses the rootfs but also chroots via chroot() and pivot_root(). To mount a mount beneath the rootfs or a chroot, pivot_root() can be used as illustrated above. * The source mount must be a private mount to force the kernel to allocate a new, unused peer group id. This isn't a required restriction but a voluntary one. It avoids repeating a semantical quirk that already exists today. If bind mounts which already have a peer group id are inserted into mount trees that have the same peer group id this can cause a lot of mount propagation events to be generated (For example, consider running mount --bind /opt /opt in a loop where the parent mount is a shared mount.). * Avoid getting rid of the top mount in the kernel. Cooperative services need to be able to unmount the top mount themselves. This also avoids a good deal of additional complexity. The umount would have to be propagated which would be another rather expensive operation. So namespace_lock() and lock_mount_hash() would potentially have to be held for a long time for both a mount and umount propagation. That should be avoided. * The path to mount beneath must be mounted and attached. * The top mount and its parent must be in the caller's mount namespace and the caller must be able to mount in that mount namespace. * The caller must be able to unmount the top mount to prove that they could reveal the underlying mount. * The propagation tree is calculated based on the destination mount's parent mount and the destination mount's mountpoint on the parent mount. Of course, if the parent of the destination mount and the destination mount are shared mounts in the same peer group and the mountpoint of the new mount to be mounted is a subdir of their ->mnt_root then both will receive a mount of /opt. That's probably easier to understand with an example. Assuming a standard shared rootfs /: mount --bind /opt /opt mount --bind /tmp /opt will cause the same mount tree as: mount --bind /opt /opt mount --beneath /tmp /opt because both / and /opt are shared mounts/peers in the same peer group and the /opt dentry is a subdirectory of both the parent's and the child's ->mnt_root. If a mount tree like that is created it almost always is an accident or abuse of mount propagation. Realistically what most people probably mean in this scenarios is: mount --bind /opt /opt mount --make-private /opt mount --make-shared /opt This forces the allocation of a new separate peer group for the /opt mount. Aferwards a mount --bind or mount --beneath actually makes sense as the / and /opt mount belong to different peer groups. Before that it's likely just confusion about what the user wanted to achieve. * Refuse MOVE_MOUNT_BENEATH if: (1) the @mnt_from has been overmounted in between path resolution and acquiring @namespace_sem when locking @mnt_to. This avoids the proliferation of shadow mounts. (2) if @to_mnt is moved to a different mountpoint while acquiring @namespace_sem to lock @to_mnt. (3) if @to_mnt is unmounted while acquiring @namespace_sem to lock @to_mnt. (4) if the parent of the target mount propagates to the target mount at the same mountpoint. This would mean mounting @mnt_from on @mnt_to->mnt_parent and then propagating a copy @c of @mnt_from onto @mnt_to. This defeats the whole purpose of mounting @mnt_from beneath @mnt_to. (5) if the parent mount @mnt_to->mnt_parent propagates to @mnt_from at the same mountpoint. If @mnt_to->mnt_parent propagates to @mnt_from this would mean propagating a copy @c of @mnt_from on top of @mnt_from. Afterwards @mnt_from would be mounted on top of @mnt_to->mnt_parent and @mnt_to would be unmounted from @mnt->mnt_parent and remounted on @mnt_from. But since @c is already mounted on @mnt_from, @mnt_to would ultimately be remounted on top of @c. Afterwards, @mnt_from would be covered by a copy @c of @mnt_from and @c would be covered by @mnt_from itself. This defeats the whole purpose of mounting @mnt_from beneath @mnt_to. Cases (1) to (3) are required as they deal with races that would cause bugs or unexpected behavior for users. Cases (4) and (5) refuse semantical quirks that would not be a bug but would cause weird mount trees to be created. While they can already be created via other means (mount --bind /opt /opt x n) there's no reason to repeat past mistakes in new features. Link: https://man7.org/linux/man-pages/man8/systemd-sysext.8.html [1] Link: https://brauner.io/2023/02/28/mounting-into-mount-namespaces.html [2] Link: https://github.com/flatcar/sysext-bakery Link: https://fedoraproject.org/wiki/Changes/Unified_Kernel_Support_Phase_1 Link: https://fedoraproject.org/wiki/Changes/Unified_Kernel_Support_Phase_2 Link: https://github.com/systemd/systemd/pull/26013 Reviewed-by: Seth Forshee (DigitalOcean) <sforshee@kernel.org> Message-Id: <20230202-fs-move-mount-replace-v4-4-98f3d80d7eaa@kernel.org> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-05-03 11:18:42 +00:00
/**
* propagation_would_overmount - check whether propagation from @from
* would overmount @to
* @from: shared mount
* @to: mount to check
* @mp: future mountpoint of @to on @from
*
* If @from propagates mounts to @to, @from and @to must either be peers
* or one of the masters in the hierarchy of masters of @to must be a
* peer of @from.
*
* If the root of the @to mount is equal to the future mountpoint @mp of
* the @to mount on @from then @to will be overmounted by whatever is
* propagated to it.
*
* Context: This function expects namespace_lock() to be held and that
* @mp is stable.
* Return: If @from overmounts @to, true is returned, false if not.
*/
bool propagation_would_overmount(const struct mount *from,
const struct mount *to,
const struct mountpoint *mp)
{
if (!IS_MNT_SHARED(from))
return false;
if (IS_MNT_NEW(to))
return false;
if (to->mnt.mnt_root != mp->m_dentry)
return false;
for (const struct mount *m = to; m; m = m->mnt_master) {
if (peers(from, m))
return true;
}
return false;
}
/*
* check if the mount 'mnt' can be unmounted successfully.
* @mnt: the mount to be checked for unmount
* NOTE: unmounting 'mnt' would naturally propagate to all
* other mounts its parent propagates to.
* Check if any of these mounts that **do not have submounts**
* have more references than 'refcnt'. If so return busy.
*
fs: scale mntget/mntput The problem that this patch aims to fix is vfsmount refcounting scalability. We need to take a reference on the vfsmount for every successful path lookup, which often go to the same mount point. The fundamental difficulty is that a "simple" reference count can never be made scalable, because any time a reference is dropped, we must check whether that was the last reference. To do that requires communication with all other CPUs that may have taken a reference count. We can make refcounts more scalable in a couple of ways, involving keeping distributed counters, and checking for the global-zero condition less frequently. - check the global sum once every interval (this will delay zero detection for some interval, so it's probably a showstopper for vfsmounts). - keep a local count and only taking the global sum when local reaches 0 (this is difficult for vfsmounts, because we can't hold preempt off for the life of a reference, so a counter would need to be per-thread or tied strongly to a particular CPU which requires more locking). - keep a local difference of increments and decrements, which allows us to sum the total difference and hence find the refcount when summing all CPUs. Then, keep a single integer "long" refcount for slow and long lasting references, and only take the global sum of local counters when the long refcount is 0. This last scheme is what I implemented here. Attached mounts and process root and working directory references are "long" references, and everything else is a short reference. This allows scalable vfsmount references during path walking over mounted subtrees and unattached (lazy umounted) mounts with processes still running in them. This results in one fewer atomic op in the fastpath: mntget is now just a per-CPU inc, rather than an atomic inc; and mntput just requires a spinlock and non-atomic decrement in the common case. However code is otherwise bigger and heavier, so single threaded performance is basically a wash. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 06:50:11 +00:00
* vfsmount lock must be held for write
*/
int propagate_mount_busy(struct mount *mnt, int refcnt)
{
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
struct mount *m, *child, *topper;
struct mount *parent = mnt->mnt_parent;
if (mnt == parent)
return do_refcount_check(mnt, refcnt);
/*
* quickly check if the current mount can be unmounted.
* If not, we don't have to go checking for all other
* mounts
*/
if (!list_empty(&mnt->mnt_mounts) || do_refcount_check(mnt, refcnt))
return 1;
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
int count = 1;
child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
if (!child)
continue;
/* Is there exactly one mount on the child that covers
* it completely whose reference should be ignored?
*/
topper = find_topper(child);
if (topper)
count += 1;
else if (!list_empty(&child->mnt_mounts))
continue;
if (do_refcount_check(child, count))
return 1;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
return 0;
}
/*
* Clear MNT_LOCKED when it can be shown to be safe.
*
* mount_lock lock must be held for write
*/
void propagate_mount_unlock(struct mount *mnt)
{
struct mount *parent = mnt->mnt_parent;
struct mount *m, *child;
BUG_ON(parent == mnt);
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
if (child)
child->mnt.mnt_flags &= ~MNT_LOCKED;
}
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static void umount_one(struct mount *mnt, struct list_head *to_umount)
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
CLEAR_MNT_MARK(mnt);
mnt->mnt.mnt_flags |= MNT_UMOUNT;
list_del_init(&mnt->mnt_child);
list_del_init(&mnt->mnt_umounting);
move_from_ns(mnt, to_umount);
}
/*
* NOTE: unmounting 'mnt' naturally propagates to all other mounts its
* parent propagates to.
*/
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static bool __propagate_umount(struct mount *mnt,
struct list_head *to_umount,
struct list_head *to_restore)
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
bool progress = false;
struct mount *child;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/*
* The state of the parent won't change if this mount is
* already unmounted or marked as without children.
*/
if (mnt->mnt.mnt_flags & (MNT_UMOUNT | MNT_MARKED))
goto out;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Verify topper is the only grandchild that has not been
* speculatively unmounted.
*/
list_for_each_entry(child, &mnt->mnt_mounts, mnt_child) {
if (child->mnt_mountpoint == mnt->mnt.mnt_root)
continue;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
if (!list_empty(&child->mnt_umounting) && IS_MNT_MARKED(child))
continue;
/* Found a mounted child */
goto children;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Mark mounts that can be unmounted if not locked */
SET_MNT_MARK(mnt);
progress = true;
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* If a mount is without children and not locked umount it. */
if (!IS_MNT_LOCKED(mnt)) {
umount_one(mnt, to_umount);
} else {
children:
list_move_tail(&mnt->mnt_umounting, to_restore);
}
out:
return progress;
}
static void umount_list(struct list_head *to_umount,
struct list_head *to_restore)
{
struct mount *mnt, *child, *tmp;
list_for_each_entry(mnt, to_umount, mnt_list) {
list_for_each_entry_safe(child, tmp, &mnt->mnt_mounts, mnt_child) {
/* topper? */
if (child->mnt_mountpoint == mnt->mnt.mnt_root)
list_move_tail(&child->mnt_umounting, to_restore);
else
umount_one(child, to_umount);
}
}
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static void restore_mounts(struct list_head *to_restore)
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Restore mounts to a clean working state */
while (!list_empty(to_restore)) {
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
struct mount *mnt, *parent;
struct mountpoint *mp;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
mnt = list_first_entry(to_restore, struct mount, mnt_umounting);
CLEAR_MNT_MARK(mnt);
list_del_init(&mnt->mnt_umounting);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Should this mount be reparented? */
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mp = mnt->mnt_mp;
parent = mnt->mnt_parent;
while (parent->mnt.mnt_flags & MNT_UMOUNT) {
mp = parent->mnt_mp;
parent = parent->mnt_parent;
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
if (parent != mnt->mnt_parent)
mnt_change_mountpoint(parent, mp, mnt);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
}
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static void cleanup_umount_visitations(struct list_head *visited)
{
while (!list_empty(visited)) {
struct mount *mnt =
list_first_entry(visited, struct mount, mnt_umounting);
list_del_init(&mnt->mnt_umounting);
}
}
/*
* collect all mounts that receive propagation from the mount in @list,
* and return these additional mounts in the same list.
* @list: the list of mounts to be unmounted.
*
* vfsmount lock must be held for write
*/
int propagate_umount(struct list_head *list)
{
struct mount *mnt;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
LIST_HEAD(to_restore);
LIST_HEAD(to_umount);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
LIST_HEAD(visited);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
/* Find candidates for unmounting */
list_for_each_entry_reverse(mnt, list, mnt_list) {
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
struct mount *parent = mnt->mnt_parent;
struct mount *m;
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
/*
* If this mount has already been visited it is known that it's
* entire peer group and all of their slaves in the propagation
* tree for the mountpoint has already been visited and there is
* no need to visit them again.
*/
if (!list_empty(&mnt->mnt_umounting))
continue;
list_add_tail(&mnt->mnt_umounting, &visited);
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
struct mount *child = __lookup_mnt(&m->mnt,
mnt->mnt_mountpoint);
if (!child)
continue;
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
if (!list_empty(&child->mnt_umounting)) {
/*
* If the child has already been visited it is
* know that it's entire peer group and all of
* their slaves in the propgation tree for the
* mountpoint has already been visited and there
* is no need to visit this subtree again.
*/
m = skip_propagation_subtree(m, parent);
continue;
} else if (child->mnt.mnt_flags & MNT_UMOUNT) {
/*
* We have come accross an partially unmounted
* mount in list that has not been visited yet.
* Remember it has been visited and continue
* about our merry way.
*/
list_add_tail(&child->mnt_umounting, &visited);
continue;
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Check the child and parents while progress is made */
while (__propagate_umount(child,
&to_umount, &to_restore)) {
/* Is the parent a umount candidate? */
child = child->mnt_parent;
if (list_empty(&child->mnt_umounting))
break;
}
}
}
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
umount_list(&to_umount, &to_restore);
restore_mounts(&to_restore);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
cleanup_umount_visitations(&visited);
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
list_splice_tail(&to_umount, list);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
return 0;
}