When an rindex entry is found to be corrupt, compute_bitstructs() calls
gfs2_consist_rgrpd() which calls gfs2_rgrp_dump() like this:
gfs2_rgrp_dump(NULL, rgd->rd_gl, fs_id_buf);
gfs2_rgrp_dump then dereferences the gl without checking it and we get
BUG: KASAN: null-ptr-deref in gfs2_rgrp_dump+0x28/0x280
because there's no rgrp glock involved while reading the rindex on mount.
Fix this by changing gfs2_rgrp_dump to take an rgrp argument.
Reported-by: syzbot+43fa87986bdd31df9de6@syzkaller.appspotmail.com
Signed-off-by: Andrew Price <anprice@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Keeping reservations and quotas separate helps reviewing the code.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Before this patch, multiple callers called gfs2_rsqa_alloc to force
the existence of a reservations structure and a quota data structure
if needed. However, now the reservations are handled separately, so
the quota data is only the quota data. So we eliminate the one in
favor of just calling gfs2_qa_alloc directly.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Before this patch, the rgrp code had a serious problem related to
how it managed buffer_heads for resource groups. The problem caused
file system corruption, especially in cases of journal replay.
When an rgrp glock was demoted to transfer ownership to a
different cluster node, do_xmote() first calls rgrp_go_sync and then
rgrp_go_inval, as expected. When it calls rgrp_go_sync, that called
gfs2_rgrp_brelse() that dropped the buffer_head reference count.
In most cases, the reference count went to zero, which is right.
However, there were other places where the buffers are handled
differently.
After rgrp_go_sync, do_xmote called rgrp_go_inval which called
gfs2_rgrp_brelse a second time, then rgrp_go_inval's call to
truncate_inode_pages_range would get rid of the pages in memory,
but only if the reference count drops to 0.
Unfortunately, gfs2_rgrp_brelse was setting bi->bi_bh = NULL.
So when rgrp_go_sync called gfs2_rgrp_brelse, it lost the pointer
to the buffer_heads in cases where the reference count was still 1.
Therefore, when rgrp_go_inval called gfs2_rgrp_brelse a second time,
it failed the check for "if (bi->bi_bh)" and thus failed to call
brelse a second time. Because of that, the reference count on those
buffers sometimes failed to drop from 1 to 0. And that caused
function truncate_inode_pages_range to keep the pages in page cache
rather than freeing them.
The next time the rgrp glock was acquired, the metadata read of
the rgrp buffers re-used the pages in memory, which were now
wrong because they were likely modified by the other node who
acquired the glock in EX (which is why we demoted the glock).
This re-use of the page cache caused corruption because changes
made by the other nodes were never seen, so the bitmaps were
inaccurate.
For some reason, the problem became most apparent when journal
replay forced the replay of rgrps in memory, which caused newer
rgrp data to be overwritten by the older in-core pages.
A big part of the problem was that the rgrp buffer were released
in multiple places: The go_unlock function would release them when
the glock was released rather than when the glock is demoted,
which is clearly wrong because our intent was to cache them until
the glock is demoted from SH or EX.
This patch attempts to clean up the mess and make one consistent
and centralized mechanism for managing the rgrp buffer_heads by
implementing several changes:
1. It eliminates the call to gfs2_rgrp_brelse() from rgrp_go_sync.
We don't want to release the buffers or zero the pointers when
syncing for the reasons stated above. It only makes sense to
release them when the glock is actually invalidated (go_inval).
And when we do, then we set the bh pointers to NULL.
2. The go_unlock function (which was only used for rgrps) is
eliminated, as we've talked about doing many times before.
The go_unlock function was called too early in the glock dq
process, and should not happen until the glock is invalidated.
3. It also eliminates the call to rgrp_brelse in gfs2_clear_rgrpd.
That will now happen automatically when the rgrp glocks are
demoted, and shouldn't happen any sooner or later than that.
Instead, function gfs2_clear_rgrpd has been modified to demote
the rgrp glocks, and therefore, free those pages, before the
remaining glocks are culled by gfs2_gl_hash_clear. This
prevents the gl_object from hanging around when the glocks are
culled.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, if a glock error was encountered, the glock with
the problem was dumped. But sometimes you may have lots of file systems
mounted, and that doesn't tell you which file system it was for.
This patch adds a new boolean parameter fsid to the dump_glock family
of functions. For non-error cases, such as dumping the glocks debugfs
file, the fsid is not dumped in order to keep lock dumps and glocktop
as clean as possible. For all error cases, such as GLOCK_BUG_ON, the
file system id is now printed. This will make it easier to debug.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Based on 1 normalized pattern(s):
this copyrighted material is made available to anyone wishing to use
modify copy or redistribute it subject to the terms and conditions
of the gnu general public license version 2
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 44 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190531081038.653000175@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This patch is based on an idea from Steve Whitehouse. The idea is
to dump the number of pages for inodes in the glock dumps.
The additional locking required me to drop const from quite a few
places.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Function rgblk_free can only deal with one resource group at a time, so
pass that resource group is as a parameter. Several of the callers
already have the resource group at hand, so we only need additional
lookup code in a few places.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Steven Whitehouse <swhiteho@redhat.com>
The state parameter of gfs2_rlist_alloc is set to LM_ST_EXCLUSIVE in all
calls, so remove it and hardcode that state in gfs2_rlist_alloc instead.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Steven Whitehouse <swhiteho@redhat.com>
This definition is only used to define RGRP_RSRV_MINBLKS, with no
benefit over defining RGRP_RSRV_MINBLKS directly.
In addition, instead of forcing RGRP_RSRV_MINBLKS to be of type u32,
cast it to that type where that type is required.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Steven Whitehouse <swhiteho@redhat.com>
Implement truncate/delete as a non-recursive algorithm. The older
algorithm was implemented with recursion to strip off each layer
at a time (going by height, starting with the maximum height.
This version tries to do the same thing but without recursion,
and without needing to allocate new structures or lists in memory.
For example, say you want to truncate a very large file to 1 byte,
and its end-of-file metapath is: 0.505.463.428. The starting
metapath would be 0.0.0.0. Since it's a truncate to non-zero, it
needs to preserve that byte, and all metadata pointing to it.
So it would start at 0.0.0.0, look up all its metadata buffers,
then free all data blocks pointed to at the highest level.
After that buffer is "swept", it moves on to 0.0.0.1, then
0.0.0.2, etc., reading in buffers and sweeping them clean.
When it gets to the end of the 0.0.0 metadata buffer (for 4K
blocks the last valid one is 0.0.0.508), it backs up to the
previous height and starts working on 0.0.1.0, then 0.0.1.1,
and so forth. After it reaches the end and sweeps 0.0.1.508,
it continues with 0.0.2.0, and so on. When that height is
exhausted, and it reaches 0.0.508.508 it backs up another level,
to 0.1.0.0, then 0.1.0.1, through 0.1.0.508. So it has to keep
marching backwards and forwards through the metadata until it's
all swept clean. Once it has all the data blocks freed, it
lowers the strip height, and begins the process all over again,
but with one less height. This time it sweeps 0.0.0 through
0.505.463. When that's clean, it lowers the strip height again
and works to free 0.505. Eventually it strips the lowest height, 0.
For a delete or truncate to 0, all metadata for all heights of
0.0.0.0 would be freed. For a truncate to 1 byte, 0.0.0.0 would
be preserved.
This isn't much different from normal integer incrementing,
where an integer gets incremented from 0000 (0.0.0.0) to 3021
(3.0.2.1). So 0000 gets increments to 0001, 0002, up to 0009,
then on to 0010, 0011 up to 0099, then 0100 and so forth. It's
just that each "digit" goes from 0 to 508 (for a total of 509
pointers) rather than from 0 to 9.
Note that the dinode will only have 483 pointers due to the
dinode structure itself.
Also note: this is just an example. These numbers (509 and 483)
are based on a standard 4K block size. Smaller block sizes will
yield smaller numbers of indirect pointers accordingly.
The truncation process is accomplished with the help of two
major functions and a few helper functions.
Functions do_strip and recursive_scan are obsolete, so removed.
New function sweep_bh_for_rgrps cleans a buffer_head pointed to
by the given metapath and height. By cleaning, I mean it frees
all blocks starting at the offset passed in metapath. It starts
at the first block in the buffer pointed to by the metapath and
identifies its resource group (rgrp). From there it frees all
subsequent block pointers that lie within that rgrp. If it's
already inside a transaction, it stays within it as long as it
can. In other words, it doesn't close a transaction until it knows
it's freed what it can from the resource group. In this way,
multiple buffers may be cleaned in a single transaction, as long
as those blocks in the buffer all lie within the same rgrp.
If it's not in a transaction, it starts one. If the buffer_head
has references to blocks within multiple rgrps, it frees all the
blocks inside the first rgrp it finds, then closes the
transaction. Then it repeats the cycle: identifies the next
unfreed block, uses it to find its rgrp, then starts a new
transaction for that set. It repeats this process repeatedly
until the buffer_head contains no more references to any blocks
past the given metapath.
Function trunc_dealloc has been reworked into a finite state
automaton. It has basically 3 active states:
DEALLOC_MP_FULL, DEALLOC_MP_LOWER, and DEALLOC_FILL_MP:
The DEALLOC_MP_FULL state implies the metapath has a full set
of buffers out to the "shrink height", and therefore, it can
call function sweep_bh_for_rgrps to free the blocks within the
highest height of the metapath. If it's just swept the lowest
level (or an error has occurred) the state machine is ended.
Otherwise it proceeds to the DEALLOC_MP_LOWER state.
The DEALLOC_MP_LOWER state implies we are finished with a given
buffer_head, which may now be released, and therefore we are
then missing some buffer information from the metapath. So we
need to find more buffers to read in. In most cases, this is
just a matter of releasing the buffer_head and moving to the
next pointer from the previous height, so it may be read in and
swept as well. If it can't find another non-null pointer to
process, it checks whether it's reached the end of a height
and needs to lower the strip height, or whether it still needs
move forward through the previous height's metadata. In this
state, all zero-pointers are skipped. From this state, it can
only loop around (once more backing up another height) or,
once a valid metapath is found (one that has non-zero
pointers), proceed to state DEALLOC_FILL_MP.
The DEALLOC_FILL_MP state implies that we have a metapath
but not all its buffers are read in. So we must proceed to read
in buffer_heads until the metapath has a valid buffer for every
height. If the previous state backed us up 3 heights, we may
need to read in a buffer, increment the height, then repeat the
process until buffers have been read in for all required heights.
If it's successful reading a buffer, and it's at the highest
height we need, it proceeds back to the DEALLOC_MP_FULL state.
If it's unable to fill in a buffer, (encounters a hole, etc.)
it tries to find another non-zero block pointer. If they're all
zero, it lowers the height and returns to the DEALLOC_MP_LOWER
state. If it finds a good non-null pointer, it loops around and
reads it in, while keeping the metapath in lock-step with the
pointers it examines.
The state machine runs until the truncation request is
satisfied. Then any transactions are ended, the quota and
statfs data are updated, and the function is complete.
Helper function metaptr1 was introduced to be an easy way to
determine the start of a buffer_head's indirect pointers.
Helper function lookup_mp_height was introduced to find a
metapath index and read in the buffer that corresponds to it.
In this way, function lookup_metapath becomes a simple loop to
call it for every height.
Helper function fillup_metapath is similar to lookup_metapath
except it can do partial lookups. If the state machine
backed up multiple levels (like 2999 wrapping to 3000) it
needs to find out the next starting point and start issuing
metadata reads at that point.
Helper function hptrs is a shortcut to determine how many
pointers should be expected in a buffer. Height 0 is the dinode
which has fewer pointers than the others.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Before this patch, multi-block reservation structures were allocated
from a special slab. This patch folds the structure into the gfs2_inode
structure. The disadvantage is that the gfs2_inode needs more memory,
even when a file is opened read-only. The advantages are: (a) we don't
need the special slab and the extra time it takes to allocate and
deallocate from it. (b) we no longer need to worry that the structure
exists for things like quota management. (c) This also allows us to
remove the calls to get_write_access and put_write_access since we
know the structure will exist.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
This patch basically reverts the majority of patch 5407e24.
That patch eliminated the gfs2_qadata structure in favor of just
using the reservations structure. The problem with doing that is that
it increases the size of the reservations structure. That is not an
issue until it comes time to fold the reservations structure into the
inode in memory so we know it's always there. By separating out the
quota structure again, we aren't punishing the non-quota users by
making all the inodes bigger, requiring more slab space. This patch
creates a new slab area to allocate the quota stuff so it's managed
a little more sanely.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
This patch allows the block allocation code to retain the buffers
for the resource groups so they don't need to be re-read from buffer
cache with every request. This is a performance improvement that's
especially noticeable when resource groups are very large. For
example, with 2GB resource groups and 4K blocks, there can be 33
blocks for every resource group. This patch allows those 33 buffers
to be kept around and not read in and thrown away with every
operation. The buffers are released when the resource group is
either synced or invalidated.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Reviewed-by: Steven Whitehouse <swhiteho@redhat.com>
Reviewed-by: Benjamin Marzinski <bmarzins@redhat.com>
struct gfs2_alloc_parms is passed to gfs2_quota_check() and
gfs2_inplace_reserve() with ap->target containing the number of
blocks being requested for allocation in the current operation.
We add a new field to struct gfs2_alloc_parms called 'allowed'.
gfs2_quota_check() and gfs2_inplace_reserve() return the max
blocks allowed by quota and the max blocks allowed by the chosen
rgrp respectively in 'allowed'.
A new field 'min_target', when non-zero, tells gfs2_quota_check()
and gfs2_inplace_reserve() to not return -EDQUOT/-ENOSPC when
there are atleast 'min_target' blocks allowable/available. The
assumption is that the caller is ok with just 'min_target' blocks
and will likely proceed with allocating them.
Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Acked-by: Steven Whitehouse <swhiteho@redhat.com>
If we run out of blocks for a given multi-block allocation, we obviously
did not reserve enough. We should reserve more blocks for the next
reservation to reduce fragmentation. This patch increases the size hint
for reservations when they run out.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch checks if i_goal is either zero or if doesn't exist
within any rgrp (i.e gfs2_blk2rgrpd() returns NULL). If so, it
assigns the ip->i_no_addr block as the i_goal.
There are two scenarios where a bad i_goal can result in a
-EBADSLT error.
1. Attempting to allocate to an existing inode:
Control reaches gfs2_inplace_reserve() and ip->i_goal is bad.
We need to fix i_goal here.
2. A new inode is created in a directory whose i_goal is hosed:
In this case, the parent dir's i_goal is copied onto the new
inode. Since the new inode is not yet created, the ip->i_no_addr
field is invalid and so, the fix in gfs2_inplace_reserve() as per
1) won't work in this scenario. We need to catch and fix it sooner
in the parent dir itself (gfs2_create_inode()), before it is
copied to the new inode.
Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Al Viro has tactfully pointed out that we are using the incorrect
error code in some cases. This patch fixes that, and also removes
the (unused) return value for glock dumping.
> * gfs2_iget() - ENOBUFS instead of ENOMEM. ENOBUFS is
> "No buffer space available (POSIX.1 (XSI STREAMS option))" and since
> we don't support STREAMS it's probably fair game, but... what the hell?
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
This patch adds a structure to contain allocation parameters with
the intention of future expansion of this structure. The idea is
that we should be able to add more information about the allocation
in the future in order to allow the allocator to make a better job
of placing the requests on-disk.
There is no functional difference from applying this patch.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The reservation for an inode should be cleared when it is truncated so
that we can start again at a different offset for future allocations.
We could try and do better than that, by resetting the search based on
where the truncation started from, but this is only a first step.
In addition, there are three callers of gfs2_rs_delete() but only one
of those should really be testing the value of i_writecount. While
we get away with that in the other cases currently, I think it would
be better if we made that test specific to the one case which
requires it.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The functions that delete block reservations from the rgrp block
reservations rbtree no longer use the ip parameter. This patch
eliminates the parameter.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Just like ext3, this works on the root directory and any directory
with the +T flag set. Also, just like ext3, any subdirectory created
in one of the just mentioned cases will be allocated to a random
resource group (GFS2 equivalent of a block group).
If you are creating a set of directories, each of which will contain a
job running on a different node, then by setting +T on the parent
directory before creating the subdirectories, each will land up in a
different resource group, and thus resource group contention between
nodes will be kept to a minimum.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
With the recently added block reservation code, an additional function
was added to search for free blocks. This had a restriction of only being
able to search for aligned extents of free blocks. As a result the
allocation patterns when reserving blocks were suboptimal when the
existing allocation of blocks for an inode was not aligned to the same
boundary.
This patch resolves that problem by adding the ability for gfs2_rbm_find
to search for extents of a particular minimum size. We can then use
gfs2_rbm_find for both looking for reservations, and also looking for
free blocks on an individual basis when we actually come to do the
allocation later on. As a result we only need a single set of code
to deal with both situations.
The function gfs2_rbm_from_block() is moved up rgrp.c so that it
occurs before all of its callers.
Many thanks are due to Bob for helping track down the final issue in
this patch. That fix to the rb_tree traversal and to not share
block reservations from a dirctory to its children is included here.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
This patch introduces a new structure, gfs2_rbm, which is a
tuple of a resource group, a bitmap within the resource group
and an offset within that bitmap. This is designed to make
manipulating these sets of variables easier. There is also a
new helper function which converts this representation back
to a disk block address.
In addition, the rbtree nodes which are used for the reservations
were not being correctly initialised, which is now fixed. Also,
the tracing was not passing through the inode where it should
have been. That is mostly fixed aside from one corner case. This
needs to be revisited since there can also be a NULL rgrp in
some cases which results in the device being incorrect in the
trace.
This is intended to be the first step towards cleaning up some
of the allocation code, and some further bug fixes.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The rs_requested field is left over from the original allocation
code, however this should have been a parameter passed to the
various functions from gfs2_inplace_reserve() and not a member of the
reservation structure as the value is not required after the
initial allocation.
This also helps simplify the code since we no longer need to set
the rs_requested to zero. Also the gfs2_inplace_release()
function can also be simplified since the reservation structure
will always be defined when it is called, and the only remaining
task is to unlock the rgrp if required. It can also now be
called unconditionally too, resulting in a further simplification.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch reduces GFS2 file fragmentation by pre-reserving blocks. The
resulting improved on disk layout greatly speeds up operations in cases
which would have resulted in interlaced allocation of blocks previously.
A typical example of this is 10 parallel dd processes, each writing to a
file in a common dirctory.
The implementation uses an rbtree of reservations attached to each
resource group (and each inode).
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch moves the ancillary quota data structures into the
block reservations structure. This saves GFS2 some time and
effort in allocating and deallocating the qadata structure.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch lengthens the lifespan of the reservations structure for
inodes. Before, they were allocated and deallocated for every write
operation. With this patch, they are allocated when the first write
occurs, and deallocated when the last process closes the file.
It's more efficient to do it this way because it saves GFS2 a lot of
unnecessary allocates and frees. It also gives us more flexibility
for the future: (1) we can now fold the qadata structure back into
the structure and save those alloc/frees, (2) we can use this for
multi-block reservations.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The FITRIM ioctl provides an alternative way to send discard requests to
the underlying device. Using the discard mount option results in every
freed block generating a discard request to the block device. This can
be slow, since many block devices can only process discard requests of
larger sizes, and also such operations can be time consuming.
Rather than using the discard mount option, FITRIM allows a sweep of the
filesystem on an occasional basis, and also to optionally avoid sending
down discard requests for smaller regions.
In GFS2 FITRIM will work at resource group granularity. There is a flag
for each resource group which keeps track of which resource groups have
been trimmed. This flag is reset whenever a deallocation occurs in the
resource group, and set whenever a successful FITRIM of that resource
group has taken place. This helps to reduce repeated discard requests
for the same block ranges, again improving performance.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch separates the code pertaining to allocations into two
parts: quota-related information and block reservations.
This patch also moves all the block reservation structure allocations to
function gfs2_inplace_reserve to simplify the code, and moves
the frees to function gfs2_inplace_release.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is a revision of the one I previously posted.
I tried to integrate all the suggestions Steve gave.
The purpose of the patch is to change function gfs2_alloc_block
(allocate either a dinode block or an extent of data blocks)
to a more generic gfs2_alloc_blocks function that can
allocate both a dinode _and_ an extent of data blocks in the
same call. This will ultimately help us create a multi-block
reservation scheme to reduce file fragmentation.
This patch moves more toward a generic multi-block allocator that
takes a pointer to the number of data blocks to allocate, plus whether
or not to allocate a dinode. In theory, it could be called to allocate
(1) a single dinode block, (2) a group of one or more data blocks, or
(3) a dinode plus several data blocks.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
GFS2 functions gfs2_alloc_block and gfs2_alloc_di do basically
the same things, with a few exceptions. This patch combines
the two functions into a slightly more generic gfs2_alloc_block.
Having one centralized block allocation function will reduce
code redundancy and make it easier to implement multi-block
reservations to reduce file fragmentation in the future.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
The two variables being initialised in gfs2_inplace_reserve
to track the file & line number of the caller are never
used, so we might as well remove them.
If something does go wrong, then a stack trace is probably
more useful anyway.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Each block which is deallocated, requires a call to gfs2_rlist_add()
and each of those calls was calling gfs2_blk2rgrpd() in order to
figure out which rgrp the block belonged in. This can be speeded up
by making use of the rgrp cached in the inode. We also reset this
cached rgrp in case the block has changed rgrp. This should provide
a big reduction in gfs2_blk2rgrpd() calls during deallocation.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Since we have ruled out supporting online filesystem shrink,
it is possible to make the resource group list append only
during the life of a super block. This gives several benefits:
Firstly, we only need to read new rindex elements as they are added
rather than needing to reread the whole rindex file each time one
element is added.
Secondly, the rindex glock can be held for much shorter periods of
time, and is completely removed from the fast path for allocations.
The lock is taken in shared mode only when updating the resource
groups when the first allocation occurs, and after a grow has
taken place.
Thirdly, this results in a reduction in code size, and everything
gets a lot simpler to understand in this area.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Here is an update of Bob's original rbtree patch which, in addition, also
resolves the rather strange ref counting that was being done relating to
the bitmap blocks.
Originally we had a dual system for journaling resource groups. The metadata
blocks were journaled and also the rgrp itself was added to a list. The reason
for adding the rgrp to the list in the journal was so that the "repolish
clones" code could be run to update the free space, and potentially send any
discard requests when the log was flushed. This was done by comparing the
"cloned" bitmap with what had been written back on disk during the transaction
commit.
Due to this, there was a requirement to hang on to the rgrps' bitmap buffers
until the journal had been flushed. For that reason, there was a rather
complicated set up in the ->go_lock ->go_unlock functions for rgrps involving
both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference
count on the buffers.
However, the journal maintains a reference count on the buffers anyway, since
they are being journaled as metadata buffers. So by moving the code which deals
with the post-journal accounting for bitmap blocks to the metadata journaling
code, we can entirely dispense with the rather strange buffer ref counting
scheme and also the requirement to journal the rgrps.
The net result of all this is that the ->sd_rindex_spin is left to do exactly
one job, and that is to look after the rbtree or rgrps.
This patch is designed to be a stepping stone towards using RCU for the rbtree
of resource groups, however the reduction in the number of uses of the
->sd_rindex_spin is likely to have benefits for multi-threaded workloads,
anyway.
The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also
be removed in future in favour of calling the functions directly where required
in the code. That will allow locking of resource groups without needing to
actually read them in - something that could be useful in speeding up statfs.
In the mean time though it is valid to dereference ->bi_bh only when the rgrp
is locked. This is basically the same rule as before, modulo the references not
being valid until the following journal flush.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Cc: Benjamin Marzinski <bmarzins@redhat.com>
__gfs2_free_data and __gfs2_free_meta are almost identical, and
can be trivially combined.
[This is as per Eric's original patch minus gfs2_free_data() which had
no callers left and plus the conversion of the bmap.c calls to these
functions. All in all, a nice clean up]
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch is a performance improvement to GFS2's dealloc code.
Rather than update the quota file and statfs file for every
single block that's stripped off in unlink function do_strip,
this patch keeps track and updates them once for every layer
that's stripped. This is done entirely inside the existing
transaction, so there should be no risk of corruption.
The other functions that deallocate blocks will be unaffected
because they are using wrapper functions that do the same
thing that they do today.
I tested this code on my roth cluster by creating 200
files in a directory, each of which is 100MB, then on
four nodes, I simultaneously deleted the files, thus competing
for GFS2 resources (but different files). The commands
I used were:
[root@roth-01]# time for i in `seq 1 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done
[root@roth-02]# time for i in `seq 2 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done
[root@roth-03]# time for i in `seq 3 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done
[root@roth-05]# time for i in `seq 4 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done
The performance increase was significant:
roth-01 roth-02 roth-03 roth-05
--------- --------- --------- ---------
old: real 0m34.027 0m25.021s 0m23.906s 0m35.646s
new: real 0m22.379s 0m24.362s 0m24.133s 0m18.562s
Total time spent deleting:
old: 118.6s
new: 89.4
For this particular case, this showed a 25% performance increase for
GFS2 unlinks.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
When you truncate the rindex file, you need to avoid calling gfs2_rindex_hold,
since you already hold it. However, if you haven't already read in the
resource groups, you need to do that.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch fixes a GFS2 problem whereby the first rename after a
mount can result in a file system consistency error being flagged
improperly and cause the file system to withdraw. The problem is
that the rename code tries to run the rgrp list with function
gfs2_blk2rgrpd before the rgrp list is guaranteed to be read in
from disk. The patch makes the rename function hold the rindex
glock (as the gfs2_unlink code does today) which reads in the rgrp
list if need be. There were a total of three places in the rename
code that improperly referenced the rgrp list without the rindex
glock and this patch fixes all three.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
There is a potential race in the inode deallocation code if two
nodes try to deallocate the same inode at the same time. Most of
the issue is solved by the iopen locking. There is still a small
window which is not covered by the iopen lock. This patches fixes
that and also makes the deallocation code more robust in the face of
any errors in the rgrp bitmaps, or erroneous iopen callbacks from
other nodes.
This does introduce one extra disk read, but that is generally not
an issue since its the same block that must be written to later
in the deallocation process. The total disk accesses therefore stay
the same,
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
A little while back, block allocation was given some improved
error handling which meant that -EIO was returned in the case
of there being a problem in the resource group data. In addition
a message is printed explaning what went wrong and how to fix it.
This extends that error handling so that it also covers inode
allocation too.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch improves the error handling in the case where we
discover that the summary information in the resource group
doesn't match the bitmap information while in the process of
allocating blocks. Originally this resulted in a kernel bug,
but this patch changes that so that we return -EIO and print
some messages explaining what went wrong, and how to fix it.
We also remember locally not to try and allocate from the
same rgrp again, so that a subsequent allocation in a
different rgrp should succeed.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Rather than having to allocate a single block at a time, this patch
allows the block allocator to allocate an extent. Since there is
no difference (so far as the block allocator is concerned) between
data blocks and indirect blocks, it is posible to allocate a single
extent and for the caller to unrevoke just the blocks required
for indirect blocks.
Currently the only bit of GFS2 to make use of this feature is the
build height function. The intention is that gfs2_block_map will
be changed to make use of this feature in future patches.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Thanks to the preceeding patches, the only difference between
these two functions is their name. We can thus merge them
and call the new function gfs2_alloc_block to reflect the
fact that it can allocate either kind of block.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch removed the unnecessary parameter from function
gfs2_rlist_alloc. The parameter was always passed in as 0.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
It is possible to reduce the size of GFS2 inodes by taking the i_alloc
structure out of the gfs2_inode. This patch allocates the i_alloc
structure whenever its needed, and frees it afterward. This decreases
the amount of low memory we use at the expense of requiring a memory
allocation for each page or partial page that we write. A quick test
with postmark shows that the overhead is not measurable and I also note
that OCFS2 use the same approach.
In the future I'd like to solve the problem by shrinking down the size
of the members of the i_alloc structure, but for now, this reduces the
immediate problem of using too much low-memory on x86 and doesn't add
too much overhead.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This addendum patch 2 corrects three things:
1. It fixes a stupid mistake in the previous addendum that broke gfs2.
Ref: https://www.redhat.com/archives/cluster-devel/2007-May/msg00162.html
2. It fixes a problem that Dave Teigland pointed out regarding the
external declarations in ops_address.h being in the wrong place.
3. It recasts a couple more %llu printks to (unsigned long long)
as requested by Steve Whitehouse.
I would have loved to put this all in one revised patch, but there was
a rush to get some patches for RHEL5. Therefore, the previous patches
were applied to the git tree "as is" and therefore, I'm posting another
addendum. Sorry.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This fixes a bug where, in certain cases an uninitialised variable
could cause a dereference of a NULL pointer in gfs2_commit_write().
Also a typo in a comment is fixed at the same time.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Andrew Price