commit 81e9d6f864 upstream.
Commit e4a0d3e720 ("aio: Make it possible to remap aio ring") introduced
a null-deref if mremap is called on an old aio mapping after fork as
mm->ioctx_table will be set to NULL.
[jmoyer@redhat.com: fix 80 column issue]
Link: https://lkml.kernel.org/r/x49sffq4nvg.fsf@segfault.boston.devel.redhat.com
Fixes: e4a0d3e720 ("aio: Make it possible to remap aio ring")
Signed-off-by: Seth Jenkins <sethjenkins@google.com>
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Benjamin LaHaise <bcrl@kvack.org>
Cc: Jann Horn <jannh@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 50252e4b5e upstream.
signalfd_poll() and binder_poll() are special in that they use a
waitqueue whose lifetime is the current task, rather than the struct
file as is normally the case. This is okay for blocking polls, since a
blocking poll occurs within one task; however, non-blocking polls
require another solution. This solution is for the queue to be cleared
before it is freed, by sending a POLLFREE notification to all waiters.
Unfortunately, only eventpoll handles POLLFREE. A second type of
non-blocking poll, aio poll, was added in kernel v4.18, and it doesn't
handle POLLFREE. This allows a use-after-free to occur if a signalfd or
binder fd is polled with aio poll, and the waitqueue gets freed.
Fix this by making aio poll handle POLLFREE.
A patch by Ramji Jiyani <ramjiyani@google.com>
(https://lore.kernel.org/r/20211027011834.2497484-1-ramjiyani@google.com)
tried to do this by making aio_poll_wake() always complete the request
inline if POLLFREE is seen. However, that solution had two bugs.
First, it introduced a deadlock, as it unconditionally locked the aio
context while holding the waitqueue lock, which inverts the normal
locking order. Second, it didn't consider that POLLFREE notifications
are missed while the request has been temporarily de-queued.
The second problem was solved by my previous patch. This patch then
properly fixes the use-after-free by handling POLLFREE in a
deadlock-free way. It does this by taking advantage of the fact that
freeing of the waitqueue is RCU-delayed, similar to what eventpoll does.
Fixes: 2c14fa838c ("aio: implement IOCB_CMD_POLL")
Cc: <stable@vger.kernel.org> # v4.18+
Link: https://lore.kernel.org/r/20211209010455.42744-6-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 363bee27e2 upstream.
Currently, aio_poll_wake() will always remove the poll request from the
waitqueue. Then, if aio_poll_complete_work() sees that none of the
polled events are ready and the request isn't cancelled, it re-adds the
request to the waitqueue. (This can easily happen when polling a file
that doesn't pass an event mask when waking up its waitqueue.)
This is fundamentally broken for two reasons:
1. If a wakeup occurs between vfs_poll() and the request being
re-added to the waitqueue, it will be missed because the request
wasn't on the waitqueue at the time. Therefore, IOCB_CMD_POLL
might never complete even if the polled file is ready.
2. When the request isn't on the waitqueue, there is no way to be
notified that the waitqueue is being freed (which happens when its
lifetime is shorter than the struct file's). This is supposed to
happen via the waitqueue entries being woken up with POLLFREE.
Therefore, leave the requests on the waitqueue until they are actually
completed (or cancelled). To keep track of when aio_poll_complete_work
needs to be scheduled, use new fields in struct poll_iocb. Remove the
'done' field which is now redundant.
Note that this is consistent with how sys_poll() and eventpoll work;
their wakeup functions do *not* remove the waitqueue entries.
Fixes: 2c14fa838c ("aio: implement IOCB_CMD_POLL")
Cc: <stable@vger.kernel.org> # v4.18+
Link: https://lore.kernel.org/r/20211209010455.42744-5-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 530f32fc37 upstream.
Avi Kivity reports that on fuse filesystems running in a user namespace
asyncronous fsync fails with EOVERFLOW.
The reason is that f_ops->fsync() is called with the creds of the kthread
performing aio work instead of the creds of the process originally
submitting IOCB_CMD_FSYNC.
Fuse sends the creds of the caller in the request header and it needs to
translate the uid and gid into the server's user namespace. Since the
kthread is running in init_user_ns, the translation will fail and the
operation returns an error.
It can be argued that fsync doesn't actually need any creds, but just
zeroing out those fields in the header (as with requests that currently
don't take creds) is a backward compatibility risk.
Instead of working around this issue in fuse, solve the core of the problem
by calling the filesystem with the proper creds.
Reported-by: Avi Kivity <avi@scylladb.com>
Tested-by: Giuseppe Scrivano <gscrivan@redhat.com>
Fixes: c9582eb0ff ("fuse: Fail all requests with invalid uids or gids")
Cc: stable@vger.kernel.org # 4.18+
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 01d7a35687 upstream.
If we have nested or circular eventfd wakeups, then we can deadlock if
we run them inline from our poll waitqueue wakeup handler. It's also
possible to have very long chains of notifications, to the extent where
we could risk blowing the stack.
Check the eventfd recursion count before calling eventfd_signal(). If
it's non-zero, then punt the signaling to async context. This is always
safe, as it takes us out-of-line in terms of stack and locking context.
Cc: stable@vger.kernel.org # 4.19+
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit af5c72b1fc upstream.
aio_poll() has to cope with several unpleasant problems:
* requests that might stay around indefinitely need to
be made visible for io_cancel(2); that must not be done to
a request already completed, though.
* in cases when ->poll() has placed us on a waitqueue,
wakeup might have happened (and request completed) before ->poll()
returns.
* worse, in some early wakeup cases request might end
up re-added into the queue later - we can't treat "woken up and
currently not in the queue" as "it's not going to stick around
indefinitely"
* ... moreover, ->poll() might have decided not to
put it on any queues to start with, and that needs to be distinguished
from the previous case
* ->poll() might have tried to put us on more than one queue.
Only the first will succeed for aio poll, so we might end up missing
wakeups. OTOH, we might very well notice that only after the
wakeup hits and request gets completed (all before ->poll() gets
around to the second poll_wait()). In that case it's too late to
decide that we have an error.
req->woken was an attempt to deal with that. Unfortunately, it was
broken. What we need to keep track of is not that wakeup has happened -
the thing might come back after that. It's that async reference is
already gone and won't come back, so we can't (and needn't) put the
request on the list of cancellables.
The easiest case is "request hadn't been put on any waitqueues"; we
can tell by seeing NULL apt.head, and in that case there won't be
anything async. We should either complete the request ourselves
(if vfs_poll() reports anything of interest) or return an error.
In all other cases we get exclusion with wakeups by grabbing the
queue lock.
If request is currently on queue and we have something interesting
from vfs_poll(), we can steal it and complete the request ourselves.
If it's on queue and vfs_poll() has not reported anything interesting,
we either put it on the cancellable list, or, if we know that it
hadn't been put on all queues ->poll() wanted it on, we steal it and
return an error.
If it's _not_ on queue, it's either been already dealt with (in which
case we do nothing), or there's aio_poll_complete_work() about to be
executed. In that case we either put it on the cancellable list,
or, if we know it hadn't been put on all queues ->poll() wanted it on,
simulate what cancel would've done.
It's a lot more convoluted than I'd like it to be. Single-consumer APIs
suck, and unfortunately aio is not an exception...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 2bb874c0d8 upstream.
Instead of having aio_complete() set ->ki_res.{res,res2}, do that
explicitly in its callers, drop the reference (as aio_complete()
used to do) and delay the rest until the final iocb_put().
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a9339b7855 upstream.
We want to separate forming the resulting io_event from putting it
into the ring buffer.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit b53119f13a upstream.
aio_poll() is not the only case that needs file pinned; worse, while
aio_read()/aio_write() can live without pinning iocb itself, the
proof is rather brittle and can easily break on later changes.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 84c4e1f89f upstream.
Al Viro root-caused a race where the IOCB_CMD_POLL handling of
fget/fput() could cause us to access the file pointer after it had
already been freed:
"In more details - normally IOCB_CMD_POLL handling looks so:
1) io_submit(2) allocates aio_kiocb instance and passes it to
aio_poll()
2) aio_poll() resolves the descriptor to struct file by req->file =
fget(iocb->aio_fildes)
3) aio_poll() sets ->woken to false and raises ->ki_refcnt of that
aio_kiocb to 2 (bumps by 1, that is).
4) aio_poll() calls vfs_poll(). After sanity checks (basically,
"poll_wait() had been called and only once") it locks the queue.
That's what the extra reference to iocb had been for - we know we
can safely access it.
5) With queue locked, we check if ->woken has already been set to
true (by aio_poll_wake()) and, if it had been, we unlock the
queue, drop a reference to aio_kiocb and bugger off - at that
point it's a responsibility to aio_poll_wake() and the stuff
called/scheduled by it. That code will drop the reference to file
in req->file, along with the other reference to our aio_kiocb.
6) otherwise, we see whether we need to wait. If we do, we unlock the
queue, drop one reference to aio_kiocb and go away - eventual
wakeup (or cancel) will deal with the reference to file and with
the other reference to aio_kiocb
7) otherwise we remove ourselves from waitqueue (still under the
queue lock), so that wakeup won't get us. No async activity will
be happening, so we can safely drop req->file and iocb ourselves.
If wakeup happens while we are in vfs_poll(), we are fine - aio_kiocb
won't get freed under us, so we can do all the checks and locking
safely. And we don't touch ->file if we detect that case.
However, vfs_poll() most certainly *does* touch the file it had been
given. So wakeup coming while we are still in ->poll() might end up
doing fput() on that file. That case is not too rare, and usually we
are saved by the still present reference from descriptor table - that
fput() is not the final one.
But if another thread closes that descriptor right after our fget()
and wakeup does happen before ->poll() returns, we are in trouble -
final fput() done while we are in the middle of a method:
Al also wrote a patch to take an extra reference to the file descriptor
to fix this, but I instead suggested we just streamline the whole file
pointer handling by submit_io() so that the generic aio submission code
simply keeps the file pointer around until the aio has completed.
Fixes: bfe4037e72 ("aio: implement IOCB_CMD_POLL")
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Reported-by: syzbot+503d4cc169fcec1cb18c@syzkaller.appspotmail.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 88a6f18b95 upstream.
In preparation of handing in iocbs in a different fashion as well. Also
make it clear that the iocb being passed in isn't modified, by marking
it const throughout.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 2bc4ca9bb6 upstream.
It's 192 bytes, fairly substantial. Most items don't need to be cleared,
especially not upfront. Clear the ones we do need to clear, and leave
the other ones for setup when the iocb is prepared and submitted.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 432c79978c upstream.
This is in preparation for certain types of IO not needing a ring
reserveration.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit bc9bff6162 upstream.
We know this is a read/write request, but in preparation for
having different kinds of those, ensure that we call the assigned
handler instead of assuming it's aio_complete_rq().
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 154989e45f upstream.
No one is going to poll for aio (yet), so we must clear the HIPRI
flag, as we would otherwise send it down the poll queues, where no
one will be polling for completions.
Signed-off-by: Christoph Hellwig <hch@lst.de>
IOCB_HIPRI, not RWF_HIPRI.
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit a538e3ff9d upstream.
Matthew pointed out that the ioctx_table is susceptible to spectre v1,
because the index can be controlled by an attacker. The below patch
should mitigate the attack for all of the aio system calls.
Cc: stable@vger.kernel.org
Reported-by: Matthew Wilcox <willy@infradead.org>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Pull vfs aio updates from Al Viro:
"Christoph's aio poll, saner this time around.
This time it's pretty much local to fs/aio.c. Hopefully race-free..."
* 'work.aio' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
aio: allow direct aio poll comletions for keyed wakeups
aio: implement IOCB_CMD_POLL
aio: add a iocb refcount
timerfd: add support for keyed wakeups
Pull vfs open-related updates from Al Viro:
- "do we need fput() or put_filp()" rules are gone - it's always fput()
now. We keep track of that state where it belongs - in ->f_mode.
- int *opened mess killed - in finish_open(), in ->atomic_open()
instances and in fs/namei.c code around do_last()/lookup_open()/atomic_open().
- alloc_file() wrappers with saner calling conventions are introduced
(alloc_file_clone() and alloc_file_pseudo()); callers converted, with
much simplification.
- while we are at it, saner calling conventions for path_init() and
link_path_walk(), simplifying things inside fs/namei.c (both on
open-related paths and elsewhere).
* 'work.open3' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (40 commits)
few more cleanups of link_path_walk() callers
allow link_path_walk() to take ERR_PTR()
make path_init() unconditionally paired with terminate_walk()
document alloc_file() changes
make alloc_file() static
do_shmat(): grab shp->shm_file earlier, switch to alloc_file_clone()
new helper: alloc_file_clone()
create_pipe_files(): switch the first allocation to alloc_file_pseudo()
anon_inode_getfile(): switch to alloc_file_pseudo()
hugetlb_file_setup(): switch to alloc_file_pseudo()
ocxlflash_getfile(): switch to alloc_file_pseudo()
cxl_getfile(): switch to alloc_file_pseudo()
... and switch shmem_file_setup() to alloc_file_pseudo()
__shmem_file_setup(): reorder allocations
new wrapper: alloc_file_pseudo()
kill FILE_{CREATED,OPENED}
switch atomic_open() and lookup_open() to returning 0 in all success cases
document ->atomic_open() changes
->atomic_open(): return 0 in all success cases
get rid of 'opened' in path_openat() and the helpers downstream
...
If we get a keyed wakeup for a aio poll waitqueue and wake can acquire the
ctx_lock without spinning we can just complete the iocb straight from the
wakeup callback to avoid a context switch.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Avi Kivity <avi@scylladb.com>
Simple one-shot poll through the io_submit() interface. To poll for
a file descriptor the application should submit an iocb of type
IOCB_CMD_POLL. It will poll the fd for the events specified in the
the first 32 bits of the aio_buf field of the iocb.
Unlike poll or epoll without EPOLLONESHOT this interface always works
in one shot mode, that is once the iocb is completed, it will have to be
resubmitted.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Avi Kivity <avi@scylladb.com>
This is needed to prevent races caused by the way the ->poll API works.
To avoid introducing overhead for other users of the iocbs we initialize
it to zero and only do refcount operations if it is non-zero in the
completion path.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Avi Kivity <avi@scylladb.com>
Pull vfs fixes from Al Viro:
"Fix several places that screw up cleanups after failures halfway
through opening a file (one open-coding filp_clone_open() and getting
it wrong, two misusing alloc_file()). That part is -stable fodder from
the 'work.open' branch.
And Christoph's regression fix for uapi breakage in aio series;
include/uapi/linux/aio_abi.h shouldn't be pulling in the kernel
definition of sigset_t, the reason for doing so in the first place had
been bogus - there's no need to expose struct __aio_sigset in
aio_abi.h at all"
* 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
aio: don't expose __aio_sigset in uapi
ocxlflash_getfile(): fix double-iput() on alloc_file() failures
cxl_getfile(): fix double-iput() on alloc_file() failures
drm_mode_create_lease_ioctl(): fix open-coded filp_clone_open()
glibc uses a different defintion of sigset_t than the kernel does,
and the current version would pull in both. To fix this just do not
expose the type at all - this somewhat mirrors pselect() where we
do not even have a type for the magic sigmask argument, but just
use pointer arithmetics.
Fixes: 7a074e96 ("aio: implement io_pgetevents")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Adrian Reber <adrian@lisas.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
takes inode, vfsmount, name, O_... flags and file_operations and
either returns a new struct file (in which case inode reference we
held is consumed) or returns ERR_PTR(), in which case no refcounts
are altered.
converted aio_private_file() and sock_alloc_file() to it
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
... so that it could set both ->f_flags and ->f_mode, without callers
having to set ->f_flags manually.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The poll() changes were not well thought out, and completely
unexplained. They also caused a huge performance regression, because
"->poll()" was no longer a trivial file operation that just called down
to the underlying file operations, but instead did at least two indirect
calls.
Indirect calls are sadly slow now with the Spectre mitigation, but the
performance problem could at least be largely mitigated by changing the
"->get_poll_head()" operation to just have a per-file-descriptor pointer
to the poll head instead. That gets rid of one of the new indirections.
But that doesn't fix the new complexity that is completely unwarranted
for the regular case. The (undocumented) reason for the poll() changes
was some alleged AIO poll race fixing, but we don't make the common case
slower and more complex for some uncommon special case, so this all
really needs way more explanations and most likely a fundamental
redesign.
[ This revert is a revert of about 30 different commits, not reverted
individually because that would just be unnecessarily messy - Linus ]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The ->poll_mask() operation has a mask of events that the caller
is interested in, but not all implementations might take it into
account. Mask the return value to only the requested events,
similar to what the poll and epoll code does.
Reported-by: Avi Kivity <avi@scylladb.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This is the per-I/O equivalent of the ioprio_set system call.
When IOCB_FLAG_IOPRIO is set on the iocb aio_flags field, then we set the
newly added kiocb ki_ioprio field to the value in the iocb aio_reqprio field.
This patch depends on block: add ioprio_check_cap function.
Signed-off-by: Adam Manzanares <adam.manzanares@wdc.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
In order to avoid kiocb bloat for per command iopriority support, rw_hint
is converted from enum to a u16. Added a guard around ki_hint assignment.
Signed-off-by: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
as it is, the logics in native io_submit(2) is "if asked for
more than LONG_MAX/sizeof(pointer) iocbs to submit, don't
bother with more than LONG_MAX/sizeof(pointer)" (i.e.
512M requests on 32bit and 1E requests on 64bit) while
compat io_submit(2) goes with "stop after the first
PAGE_SIZE/sizeof(pointer) iocbs", i.e. 1K or so. Which is
* inconsistent
* *way* too much in native case
* possibly too little in compat one
and
* wrong anyway, since the natural point where we
ought to stop bothering is ctx->nr_events
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
get rid of insane "copy array of 32bit pointers into an array of
native ones" glue.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The logics for 'avail' is
* not past the tail of cyclic buffer
* no more than asked
* not past the end of buffer
* not past the end of a page
Unobfuscate the last part.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
... so just make them return 0 when caller does not need to destroy iocb
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
If we can acquire ctx_lock without spinning we can just remove our
iocb from the active_reqs list, and thus complete the iocbs from the
wakeup context.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Simple one-shot poll through the io_submit() interface. To poll for
a file descriptor the application should submit an iocb of type
IOCB_CMD_POLL. It will poll the fd for the events specified in the
the first 32 bits of the aio_buf field of the iocb.
Unlike poll or epoll without EPOLLONESHOT this interface always works
in one shot mode, that is once the iocb is completed, it will have to be
resubmitted.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
With the current aio code there is no need for the magic KIOCB_CANCELLED
value, as a cancelation just kicks the driver to queue the completion
ASAP, with all actual completion handling done in another thread. Given
that both the completion path and cancelation take the context lock there
is no need for magic cmpxchg loops either. If we remove iocbs from the
active list after calling ->ki_cancel (but with ctx_lock still held), we
can also rely on the invariant thay anything found on the list has a
->ki_cancel callback and can be cancelled, further simplifing the code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
No need to pass the key field to lookup_iocb to compare it with KIOCB_KEY,
as we can do that right after retrieving it from userspace. Also move the
KIOCB_KEY definition to aio.c as it is an internal value not used by any
other place in the kernel.
Signed-off-by: Christoph Hellwig <hch@lst.de>
If io_destroy() gets to cancelling everything that can be cancelled and
gets to kiocb_cancel() calling the function driver has left in ->ki_cancel,
it becomes vulnerable to a race with IO completion. At that point req
is already taken off the list and aio_complete() does *NOT* spin until
we (in free_ioctx_users()) releases ->ctx_lock. As the result, it proceeds
to kiocb_free(), freing req just it gets passed to ->ki_cancel().
Fix is simple - remove from the list after the call of kiocb_cancel(). All
instances of ->ki_cancel() already have to cope with the being called with
iocb still on list - that's what happens in io_cancel(2).
Cc: stable@kernel.org
Fixes: 0460fef2a9 "aio: use cancellation list lazily"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
kill_ioctx() used to have an explicit RCU delay between removing the
reference from ->ioctx_table and percpu_ref_kill() dropping the refcount.
At some point that delay had been removed, on the theory that
percpu_ref_kill() itself contained an RCU delay. Unfortunately, that was
the wrong kind of RCU delay and it didn't care about rcu_read_lock() used
by lookup_ioctx(). As the result, we could get ctx freed right under
lookup_ioctx(). Tejun has fixed that in a6d7cff472 ("fs/aio: Add explicit
RCU grace period when freeing kioctx"); however, that fix is not enough.
Suppose io_destroy() from one thread races with e.g. io_setup() from another;
CPU1 removes the reference from current->mm->ioctx_table[...] just as CPU2
has picked it (under rcu_read_lock()). Then CPU1 proceeds to drop the
refcount, getting it to 0 and triggering a call of free_ioctx_users(),
which proceeds to drop the secondary refcount and once that reaches zero
calls free_ioctx_reqs(). That does
INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
queue_rcu_work(system_wq, &ctx->free_rwork);
and schedules freeing the whole thing after RCU delay.
In the meanwhile CPU2 has gotten around to percpu_ref_get(), bumping the
refcount from 0 to 1 and returned the reference to io_setup().
Tejun's fix (that queue_rcu_work() in there) guarantees that ctx won't get
freed until after percpu_ref_get(). Sure, we'd increment the counter before
ctx can be freed. Now we are out of rcu_read_lock() and there's nothing to
stop freeing of the whole thing. Unfortunately, CPU2 assumes that since it
has grabbed the reference, ctx is *NOT* going away until it gets around to
dropping that reference.
The fix is obvious - use percpu_ref_tryget_live() and treat failure as miss.
It's not costlier than what we currently do in normal case, it's safe to
call since freeing *is* delayed and it closes the race window - either
lookup_ioctx() comes before percpu_ref_kill() (in which case ctx->users
won't reach 0 until the caller of lookup_ioctx() drops it) or lookup_ioctx()
fails, ctx->users is unaffected and caller of lookup_ioctx() doesn't see
the object in question at all.
Cc: stable@kernel.org
Fixes: a6d7cff472 "fs/aio: Add explicit RCU grace period when freeing kioctx"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
