commit 4eb7b93e03 ("ocfs2: improve write IO performance when
fragmentation is high") introduced another regression.
The following ocfs2-test case can trigger this issue:
> discontig_runner.sh => activate_discontig_bg.sh => resv_unwritten:
> ${RESV_UNWRITTEN_BIN} -f ${WORK_PLACE}/large_testfile -s 0 -l \
> $((${FILE_MAJOR_SIZE_M}*1024*1024))
In my env, test disk size (by "fdisk -l <dev>"):
> 53687091200 bytes, 104857600 sectors.
Above command is:
> /usr/local/ocfs2-test/bin/resv_unwritten -f \
> /mnt/ocfs2/ocfs2-activate-discontig-bg-dir/large_testfile -s 0 -l \
> 53187969024
Error log:
> [*] Reserve 50724M space for a LARGE file, reserve 200M space for future test.
> ioctl error 28: "No space left on device"
> resv allocation failed Unknown error -1
> reserve unwritten region from 0 to 53187969024.
Call flow:
__ocfs2_change_file_space //by ioctl OCFS2_IOC_RESVSP64
ocfs2_allocate_unwritten_extents //start:0 len:53187969024
while()
+ ocfs2_get_clusters //cpos:0, alloc_size:1623168 (cluster number)
+ ocfs2_extend_allocation
+ ocfs2_lock_allocators
| + choose OCFS2_AC_USE_MAIN & ocfs2_cluster_group_search
|
+ ocfs2_add_inode_data
ocfs2_add_clusters_in_btree
__ocfs2_claim_clusters
ocfs2_claim_suballoc_bits
+ During the allocation of the final part of the large file
(after ~47GB), no chain had the required contiguous
bits_wanted. Consequently, the allocation failed.
How to fix:
When OCFS2 is encountering fragmented allocation, the file system should
stop attempting bits_wanted contiguous allocation and instead provide the
largest available contiguous free bits from the cluster groups.
Link: https://lkml.kernel.org/r/20250414060125.19938-2-heming.zhao@suse.com
Fixes: 4eb7b93e03 ("ocfs2: improve write IO performance when fragmentation is high")
Signed-off-by: Heming Zhao <heming.zhao@suse.com>
Reported-by: Gautham Ananthakrishna <gautham.ananthakrishna@oracle.com>
Reviewed-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Cc: Mark Fasheh <mark@fasheh.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Changwei Ge <gechangwei@live.cn>
Cc: Jun Piao <piaojun@huawei.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The group_search function ocfs2_cluster_group_search() should
bypass groups with insufficient space to avoid unnecessary
searches.
This patch is particularly useful when ocfs2 is handling huge
number small files, and volume fragmentation is very high.
In this case, ocfs2 is busy with looking up available la window
from //global_bitmap.
This patch introduces a new member in the Group Description (gd)
struct called 'bg_contig_free_bits', representing the max
contigous free bits in this gd. When ocfs2 allocates a new
la window from //global_bitmap, 'bg_contig_free_bits' helps
expedite the search process.
Let's image below path.
1. la state (->local_alloc_state) is set THROTTLED or DISABLED.
2. when user delete a large file and trigger
ocfs2_local_alloc_seen_free_bits set osb->local_alloc_state
unconditionally.
3. a write IOs thread run and trigger the worst performance path
```
ocfs2_reserve_clusters_with_limit
ocfs2_reserve_local_alloc_bits
ocfs2_local_alloc_slide_window //[1]
+ ocfs2_local_alloc_reserve_for_window //[2]
+ ocfs2_local_alloc_new_window //[3]
ocfs2_recalc_la_window
```
[1]:
will be called when la window bits used up.
[2]:
under la state is ENABLED, and this func only check global_bitmap
free bits, it will succeed in general.
[3]:
will use the default la window size to search clusters then fail.
ocfs2_recalc_la_window attempts other la window sizes.
the timing complexity is O(n^4), resulting in a significant time
cost for scanning global bitmap. This leads to a dramatic slowdown
in write I/Os (e.g., user space 'dd').
i.e.
an ocfs2 partition size: 1.45TB, cluster size: 4KB,
la window default size: 106MB.
The partition is fragmentation by creating & deleting huge mount of
small files.
before this patch, the timing of [3] should be
(the number got from real world):
- la window size change order (size: MB):
106, 53, 26.5, 13, 6.5, 3.25, 1.6, 0.8
only 0.8MB succeed, 0.8MB also triggers la window to disable.
ocfs2_local_alloc_new_window retries 8 times, first 7 times totally
runs in worst case.
- group chain number: 242
ocfs2_claim_suballoc_bits calls for-loop 242 times
- each chain has 49 block group
ocfs2_search_chain calls while-loop 49 times
- each bg has 32256 blocks
ocfs2_block_group_find_clear_bits calls while-loop for 32256 bits.
for ocfs2_find_next_zero_bit uses ffz() to find zero bit, let's use
(32256/64) (this is not worst value) for timing calucation.
the loop times: 7*242*49*(32256/64) = 41835024 (~42 million times)
In the worst case, user space writes 1MB data will trigger 42M scanning
times.
under this patch, the timing is '7*242*49 = 83006', reduced by three
orders of magnitude.
Link: https://lkml.kernel.org/r/20240328125203.20892-2-heming.zhao@suse.com
Signed-off-by: Heming Zhao <heming.zhao@suse.com>
Reviewed-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Cc: Changwei Ge <gechangwei@live.cn>
Cc: Gang He <ghe@suse.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Jun Piao <piaojun@huawei.com>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Cc: Mark Fasheh <mark@fasheh.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version this program is distributed in the
hope that it will be useful but without any warranty without even
the implied warranty of merchantability or fitness for a particular
purpose see the gnu general public license for more details you
should have received a copy of the gnu general public license along
with this program if not write to the free software foundation inc
59 temple place suite 330 boston ma 021110 1307 usa
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 84 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Richard Fontana <rfontana@redhat.com>
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/20190524100844.756442981@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
After updating alloc_dinode counts in ocfs2_alloc_dinode_update_counts(),
if ocfs2_alloc_dinode_update_bitmap() failed, there is a rare case that
some space may be lost.
So, roll back alloc_dinode counts when ocfs2_block_group_set_bits()
failed.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Younger Liu <younger.liucn@gmail.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ocfs2_alloc_dinode_update_counts() and ocfs2_block_group_set_bits() are
already provided in suballoc.c. So, the same functions in
move_extents.c are not needed any more.
Declare the functions in suballoc.h and remove redundant functions in
move_extents.c.
Signed-off-by: Younger Liu <liuyiyang@hisense.com>
Cc: Younger Liu <younger.liucn@gmail.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Joel Becker <jlbec@evilplan.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ocfs2_block_group_alloc_discontig() disables chain relink by setting
ac->ac_allow_chain_relink = 0 because it grabs clusters from multiple
cluster groups.
It doesn't keep the credits for all chain relink,but
ocfs2_claim_suballoc_bits overrides this in this call trace:
ocfs2_block_group_claim_bits()->ocfs2_claim_clusters()->
__ocfs2_claim_clusters()->ocfs2_claim_suballoc_bits()
ocfs2_claim_suballoc_bits set ac->ac_allow_chain_relink = 1; then call
ocfs2_search_chain() one time and disable it again, and then we run out
of credits.
Fix is to allow relink by default and disable it in
ocfs2_block_group_alloc_discontig.
Without this patch, End-users will run into a crash due to run out of
credits, backtrace like this:
RIP: 0010:[<ffffffffa0808b14>] [<ffffffffa0808b14>]
jbd2_journal_dirty_metadata+0x164/0x170 [jbd2]
RSP: 0018:ffff8801b919b5b8 EFLAGS: 00010246
RAX: 0000000000000000 RBX: ffff88022139ddc0 RCX: ffff880159f652d0
RDX: ffff880178aa3000 RSI: ffff880159f652d0 RDI: ffff880087f09bf8
RBP: ffff8801b919b5e8 R08: 0000000000000000 R09: 0000000000000000
R10: 0000000000001e00 R11: 00000000000150b0 R12: ffff880159f652d0
R13: ffff8801a0cae908 R14: ffff880087f09bf8 R15: ffff88018d177800
FS: 00007fc9b0b6b6e0(0000) GS:ffff88022fd40000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 000000000040819c CR3: 0000000184017000 CR4: 00000000000006e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process dd (pid: 9945, threadinfo ffff8801b919a000, task ffff880149a264c0)
Call Trace:
ocfs2_journal_dirty+0x2f/0x70 [ocfs2]
ocfs2_relink_block_group+0x111/0x480 [ocfs2]
ocfs2_search_chain+0x455/0x9a0 [ocfs2]
...
Signed-off-by: Xiaowei.Hu <xiaowei.hu@oracle.com>
Reviewed-by: Srinivas Eeda <srinivas.eeda@oracle.com>
Cc: Mark Fasheh <mfasheh@suse.com>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This allows code which needs to know the eventual block number of an inode
but can't allocate it yet due to transaction or lock ordering. For example,
ocfs2_create_inode_in_orphan() currently gives a junk blkno for preparation
of the orphan dir because it can't yet know where the actual inode is placed
- that code is actually in ocfs2_mknod_locked. This is a problem when the
orphan dirs are indexed as the junk inode number will create an index entry
which goes unused (and fails the later removal from the orphan dir). Now
with these interfaces, ocfs2_create_inode_in_orphan() can run the block
group search (and get back the inode block number) *before* any actual
allocation occurs.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Get the suballoc_loc from ocfs2_claim_new_inode() or
ocfs2_claim_metadata(). Store it on the appropriate field of the block
we just allocated.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
They all take an ocfs2_alloc_context, which has the allocation inode.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
We're going to be adding more info to a suballocator allocation. Rather
than growing every function in the chain, let's pass a result structure
around.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Tao Ma <tao.ma@oracle.com>
This patch improves Ocfs2 allocation policy by allowing an inode to
reserve a portion of the local alloc bitmap for itself. The reserved
portion (allocation window) is advisory in that other allocation
windows might steal it if the local alloc bitmap becomes
full. Otherwise, the reservations are honored and guaranteed to be
free. When the local alloc window is moved to a different portion of
the bitmap, existing reservations are discarded.
Reservation windows are represented internally by a red-black
tree. Within that tree, each node represents the reservation window of
one inode. An LRU of active reservations is also maintained. When new
data is written, we allocate it from the inodes window. When all bits
in a window are exhausted, we allocate a new one as close to the
previous one as possible. Should we not find free space, an existing
reservation is pulled off the LRU and cannibalized.
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
When the local alloc file changes windows, unused bits are freed back to the
global bitmap. By defnition, those bits can not be in use by any file. Also,
the local alloc will never have been able to allocate those bits if they
were part of a previous truncate. Therefore it makes sense that we should
clear unused local alloc bits in the undo buffer so that they can be used
immediatly.
[ Modified to call it ocfs2_release_clusters() -- Joel ]
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
This patch add extent block (metadata) stealing mechanism for
extent allocation. This mechanism is same as the inode stealing.
if no room in slot specific extent_alloc, we will try to
allocate extent block from the next slot.
Signed-off-by: Tiger Yang <tiger.yang@oracle.com>
Acked-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Joel Becker <joel.becker@oracle.com>
For nfs exporting, ocfs2_get_dentry() returns the dentry for fh.
ocfs2_get_dentry() may read from disk when the inode is not in memory,
without any cross cluster lock. this leads to the file system loading a
stale inode.
This patch fixes above problem.
Solution is that in case of inode is not in memory, we get the cluster
lock(PR) of alloc inode where the inode in question is allocated from (this
causes node on which deletion is done sync the alloc inode) before reading
out the inode itsself. then we check the bitmap in the group (the inode in
question allcated from) to see if the bit is clear. if it's clear then it's
stale. if the bit is set, we then check generation as the existing code
does.
We have to read out the inode in question from disk first to know its alloc
slot and allot bit. And if its not stale we read it out using ocfs2_iget().
The second read should then be from cache.
And also we have to add a per superblock nfs_sync_lock to cover the lock for
alloc inode and that for inode in question. this is because ocfs2_get_dentry()
and ocfs2_delete_inode() lock on them in reverse order. nfs_sync_lock is locked
in EX mode in ocfs2_get_dentry() and in PR mode in ocfs2_delete_inode(). so
that mutliple ocfs2_delete_inode() can run concurrently in normal case.
[mfasheh@suse.com: build warning fixes and comment cleanups]
Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>
Acked-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
In ocfs2, the inode block search looks for the "emptiest" inode
group to allocate from. So if an inode alloc file has many equally
(or almost equally) empty groups, new inodes will tend to get
spread out amongst them, which in turn can put them all over the
disk. This is undesirable because directory operations on conceptually
"nearby" inodes force a large number of seeks.
So we add ip_last_used_group in core directory inodes which records
the last used allocation group. Another field named ip_last_used_slot
is also added in case inode stealing happens. When claiming new inode,
we passed in directory's inode so that the allocation can use this
information.
For more details, please see
http://oss.oracle.com/osswiki/OCFS2/DesignDocs/InodeAllocationStrategy.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Add an optional validation hook to ocfs2_read_blocks(). Now the
validation function is only called when a block was actually read off of
disk. It is not called when the buffer was in cache.
We add a buffer state bit BH_NeedsValidate to flag these buffers. It
must always be one higher than the last JBD2 buffer state bit.
The dinode, dirblock, extent_block, and xattr_block validators are
lifted to this scheme directly. The group_descriptor validator needs to
be split into two pieces. The first part only needs the gd buffer and
is passed to ocfs2_read_block(). The second part requires the dinode as
well, and is called every time. It's only 3 compares, so it's tiny.
This also allows us to clean up the non-fatal gd check used by resize.c.
It now has no magic argument.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
We have a clean call for validating group descriptors, but every place
that wants the always does a read_block()+validate() call pair. Create
a toplevel ocfs2_read_group_descriptor() that does the right
thing. This allows us to leverage the single call point later for
fancier handling. We also add validation of gd->bg_generation against
the superblock and gd->bg_blkno against the block we thought we read.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Currently the validation of group descriptors is directly duplicated so
that one version can error the filesystem and the other (resize) can
just report the problem. Consolidate to one function that takes a
boolean. Wrap that function with the old call for the old users.
This is in preparation for lifting the read+validate step into a
single function.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
ocfs2 inode numbers are block numbers. For any filesystem with less
than 2^32 blocks, this is not a problem. However, when ocfs2 starts
using JDB2, it will be able to support filesystems with more than 2^32
blocks. This would result in inode numbers higher than 2^32.
The problem is that stat(2) can't handle those numbers on 32bit
machines. The simple solution is to have ocfs2 allocate all inodes
below that boundary.
The suballoc code is changed to honor an optional block limit. Only the
inode suballocator sets that limit - all other allocations stay unlimited.
The biggest trick is to grow the inode suballocator beneath that limit.
There's no point in allocating block groups that are above the limit,
then rejecting their elements later on. We want to prevent the inode
allocator from ever having block groups above the limit. This involves
a little gyration with the local alloc code. If the local alloc window
is above the limit, it signals the caller to try the global bitmap but
does not disable the local alloc file (which can be used for other
allocations).
[ Minor cleanup - removed an ML_NOTICE comment. --Mark ]
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
We now have three different kinds of extent trees in ocfs2: inode data
(dinode), extended attributes (xattr_tree), and extended attribute
values (xattr_value). There is a nice abstraction for them,
ocfs2_extent_tree, but it is hidden in alloc.c. All the calling
functions have to pick amongst a varied API and pass in type bits and
often extraneous pointers.
A better way is to make ocfs2_extent_tree a first-class object.
Everyone converts their object to an ocfs2_extent_tree() via the
ocfs2_get_*_extent_tree() calls, then uses the ocfs2_extent_tree for all
tree calls to alloc.c.
This simplifies a lot of callers, making for readability. It also
provides an easy way to add additional extent tree types, as they only
need to be defined in alloc.c with a ocfs2_get_<new>_extent_tree()
function.
Signed-off-by: Joel Becker <joel.becker@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
This patch implements storing extended attributes both in inode or a single
external block. We only store EA's in-inode when blocksize > 512 or that
inode block has free space for it. When an EA's value is larger than 80
bytes, we will store the value via b-tree outside inode or block.
Signed-off-by: Tiger Yang <tiger.yang@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Add some thin wrappers around ocfs2_insert_extent() for each of the 3
different btree types, ocfs2_inode_insert_extent(),
ocfs2_xattr_value_insert_extent() and ocfs2_xattr_tree_insert_extent(). The
last is for the xattr index btree, which will be used in a followup patch.
All the old callers in file.c etc will call ocfs2_dinode_insert_extent(),
while the other two handle the xattr issue. And the init of extent tree are
handled by these functions.
When storing xattr value which is too large, we will allocate some clusters
for it and here ocfs2_extent_list and ocfs2_extent_rec will also be used. In
order to re-use the b-tree operation code, a new parameter named "private"
is added into ocfs2_extent_tree and it is used to indicate the root of
ocfs2_exent_list. The reason is that we can't deduce the root from the
buffer_head now. It may be in an inode, an ocfs2_xattr_block or even worse,
in any place in an ocfs2_xattr_bucket.
Signed-off-by: Tao Ma <tao.ma@oracle.com>
Signed-off-by: Mark Fasheh <mfasheh@suse.com>
Heming ZhaoandAndrew Morton
Jiangshan YiandAndrew Morton
Masahiro YamadaandLinus Torvalds
Randy DunlapandLinus Torvalds
Thomas GleixnerandGreg Kroah-Hartman
Younger LiuandLinus Torvalds
Younger LiuandLinus Torvalds
Xiaowei.HuandLinus Torvalds
Mark FashehandTao Ma
Joel BeckerandTao Ma
Tiger YangandJoel Becker
wengang wangandMark Fasheh