We keep track of what we've so far been referring to as reservations in
rd_rstree: the nodes in that tree indicate where in a resource group we'd
like to allocate the next couple of blocks for a particular inode. Local
processes take those as hints, but they may still "steal" blocks from those
extents, so when actually allocating a block, we must double check in the
bitmap whether that block is actually still free. Likewise, other cluster
nodes may "steal" such blocks as well.
One of the following patches introduces resource group glock sharing, i.e.,
sharing of an exclusively locked resource group glock among local processes to
speed up allocations. To make that work, we'll need to keep track of how many
blocks we've actually reserved for each inode, so we end up with two different
kinds of reservations.
Distinguish these two kinds by referring to blocks which are reserved but may
still be "stolen" as "requested". This rename also makes it more obvious that
rs_requested and rd_requested are strongly related.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
In gfs2_release, check if the inode has an active reservation to avoid
unnecessary lock taking.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
If gfs2_inplace_reserve has chosen a resource group but it couldn't make a
reservation there, there are too many other reservations in that resource
group. In that case, don't even try to respect existing reservations in
gfs2_alloc_blocks.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Only pass the current reservation down to gfs2_rbm_find rather than the entire
inode; we don't need any of the other information.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Pass a non-NULL minext to gfs2_rbm_find even for single-block allocations. In
gfs2_rbm_find, also set rgd->rd_extfail_pt when a single-block allocation
fails in a resource group: there is no reason for treating that case
differently. In gfs2_reservation_check_and_update, only check how many free
blocks we have if more than one block is requested; we already know there's at
least one free block.
In addition, when allocating N blocks fails in gfs2_rbm_find, we need to set
rd_extfail_pt to N - 1 rather than N: rd_extfail_pt defines the biggest
allocation that might still succeed.
Finally, reset rd_extfail_pt when updating the resource group statistics in
update_rgrp_lvb, as we already do in gfs2_rgrp_bh_get.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
When reading a resource group from disk or when receiving the resource group
statistics from a Lock Value Block (LVB), set/clear the GBF_FULL flags of all
bitmaps in that resource group according to whether or not the resource group
is full.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Removing a reservation doesn't make any actual space available, so don't clear
the GBF_FULL flags in that case. Otherwise, we'll only spend more time
scanning the bitmaps unnecessarily.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
This reverts commit e79e0e1428.
It turns out that we're only setting the GBF_FULL flag of a bitmap if we've
been scanning from the beginning of the bitmap until the end and we haven't
found a single free block, and we're not skipping reservations in that process,
either. This means that in gfs2_rbm_find, we can always skip bitmaps with the
GBF_FULL flag set.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Variable ndata is only used inside "if (!dinode)", so it can be replaced
entirely with *nblocks.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
GFS2 uses struct gfs2_rbm to represent a filesystem block number as a
bit position within a resource group. This representation is used in
the bitmap manipulation code to prevent excessive conversions between
block numbers and bit positions, but also in struct gfs2_blkreserv which
is part of struct gfs2_inode, to mark the start of a reservation. In
the inode, the bit position representation makes less sense: first, the
start position is used as a block number about as often as a bit
position; second, the bit position representation makes the code
unnecessarily complicated and difficult to read.
Therefore, change struct gfs2_blkreserv to represent the start of a
reservation as a block number instead of a bit position. (This requires
keeping track of the resource group in gfs2_blkreserv separately.) With
that change, various things can be slightly simplified, and struct
gfs2_rbm can be moved to rgrp.c.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Change gfs2_rbm_incr to advance an rbm by a given number of blocks. Use that
in gfs2_reservation_check_and_update to save a gfs2_rbm_to_block ->
gfs2_rbm_from_block round trip.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Function signal_our_withdraw needs to work on file systems that have been
partially frozen. To do this, it called flush_workqueue(gfs2_freeze_wq).
This this wrong because it waits for *ALL* file systems to be unfrozen, not
just the one we're withdrawing from. It should only wait for the targetted
file system to be unfrozen. Otherwise it would wait until ALL file systems
are thawed before signaling the withdraw.
This patch changes signal_our_withdraw so it calls flush_work() for the target
file system's freeze work (only) to be completed.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, function gfs2_statfs_sync called sb_start_write and
sb_end_write. This is completely unnecessary because, aside from grabbing
glocks, gfs2_statfs_sync does all its updates to statfs with a transaction:
gfs2_trans_begin and _end. And transactions always do sb_start_intwrite in
gfs2_trans_begin and sb_end_intwrite in gfs2_trans_end.
This patch simply removes the call to sb_start_write.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Since commit a0e3cc65fa ("gfs2: Turn gl_delete into a delayed work"), we're
cancelling any pending delete work of an iopen glock before attaching a new
inode to that glock in gfs2_create_inode. This means that delete_work_func can
no longer be queued or running when attaching the iopen glock to the new inode,
and we can revert commit a4923865ea ("GFS2: Prevent delete work from
occurring on glocks used for create"), which tried to achieve the same but in a
racy way.
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
In gfs2_create_inode and gfs2_inode_lookup, make sure to cancel any pending
delete work before taking the inode glock. Otherwise, gfs2_cancel_delete_work
may block waiting for delete_work_func to complete, and delete_work_func may
block trying to acquire the inode glock in gfs2_inode_lookup.
Reported-by: Alexander Aring <aahringo@redhat.com>
Fixes: a0e3cc65fa ("gfs2: Turn gl_delete into a delayed work")
Cc: stable@vger.kernel.org # v5.8+
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Commit 20f829999c ("gfs2: Rework read and page fault locking") lifted
the glock lock taking from the low-level ->readpage and ->readahead
address space operations to the higher-level ->read_iter file and
->fault vm operations. The glocks are still taken in LM_ST_SHARED mode
only. On filesystems mounted without the noatime option, ->read_iter
sometimes needs to update the atime as well, though. Right now, this
leads to a failed locking mode assertion in gfs2_dirty_inode.
Fix that by introducing a new update_time inode operation. There, if
the glock is held non-exclusively, upgrade it to an exclusive lock.
Reported-by: Alexander Aring <aahringo@redhat.com>
Fixes: 20f829999c ("gfs2: Rework read and page fault locking")
Cc: stable@vger.kernel.org # v5.8+
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
GFS2's freeze/thaw mechanism uses a special freeze glock to control its
operation. It does this with a sync glock operation (glops.c) called
freeze_go_sync. When the freeze glock is demoted (glock's do_xmote) the
glops function causes the file system to be frozen. This is intended. However,
GFS2's mount and unmount processes also hold the freeze glock to prevent other
processes, perhaps on different cluster nodes, from mounting the frozen file
system in read-write mode.
Before this patch, there was no check in freeze_go_sync for whether a freeze
in intended or whether the glock demote was caused by a normal unmount.
So it was trying to freeze the file system it's trying to unmount, which
ends up in a deadlock.
This patch adds an additional check to freeze_go_sync so that demotes of the
freeze glock are ignored if they come from the unmount process.
Fixes: 20b3291290 ("gfs2: Fix regression in freeze_go_sync")
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
If gfs2 tries to mount a (corrupt) file system that has no resource
groups it still tries to set preferences on the first one, which causes
a kernel null pointer dereference. This patch adds a check to function
gfs2_ri_update so this condition is detected and reported back as an
error.
Reported-by: syzbot+e3f23ce40269a4c9053a@syzkaller.appspotmail.com
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
This patch introduce a new globs attribute to define the subclass of the
glock lockref spinlock. This avoid the following lockdep warning, which
occurs when we lock an inode lock while an iopen lock is held:
============================================
WARNING: possible recursive locking detected
5.10.0-rc3+ #4990 Not tainted
--------------------------------------------
kworker/0:1/12 is trying to acquire lock:
ffff9067d45672d8 (&gl->gl_lockref.lock){+.+.}-{3:3}, at: lockref_get+0x9/0x20
but task is already holding lock:
ffff9067da308588 (&gl->gl_lockref.lock){+.+.}-{3:3}, at: delete_work_func+0x164/0x260
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0
----
lock(&gl->gl_lockref.lock);
lock(&gl->gl_lockref.lock);
*** DEADLOCK ***
May be due to missing lock nesting notation
3 locks held by kworker/0:1/12:
#0: ffff9067c1bfdd38 ((wq_completion)delete_workqueue){+.+.}-{0:0}, at: process_one_work+0x1b7/0x540
#1: ffffac594006be70 ((work_completion)(&(&gl->gl_delete)->work)){+.+.}-{0:0}, at: process_one_work+0x1b7/0x540
#2: ffff9067da308588 (&gl->gl_lockref.lock){+.+.}-{3:3}, at: delete_work_func+0x164/0x260
stack backtrace:
CPU: 0 PID: 12 Comm: kworker/0:1 Not tainted 5.10.0-rc3+ #4990
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.13.0-2.fc32 04/01/2014
Workqueue: delete_workqueue delete_work_func
Call Trace:
dump_stack+0x8b/0xb0
__lock_acquire.cold+0x19e/0x2e3
lock_acquire+0x150/0x410
? lockref_get+0x9/0x20
_raw_spin_lock+0x27/0x40
? lockref_get+0x9/0x20
lockref_get+0x9/0x20
delete_work_func+0x188/0x260
process_one_work+0x237/0x540
worker_thread+0x4d/0x3b0
? process_one_work+0x540/0x540
kthread+0x127/0x140
? __kthread_bind_mask+0x60/0x60
ret_from_fork+0x22/0x30
Suggested-by: Andreas Gruenbacher <agruenba@redhat.com>
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Commit 0e539ca1bb ("gfs2: Fix NULL pointer dereference in gfs2_rgrp_dump")
introduced additional locking in gfs2_rgrp_go_dump, which is also used for
dumping resource group glocks via debugfs. However, on that code path, the
glock spin lock is already taken in dump_glock, and taking it again in
gfs2_glock2rgrp leads to deadlock. This can be reproduced with:
$ mkfs.gfs2 -O -p lock_nolock /dev/FOO
$ mount /dev/FOO /mnt/foo
$ touch /mnt/foo/bar
$ cat /sys/kernel/debug/gfs2/FOO/glocks
Fix that by not taking the glock spin lock inside the go_dump callback.
Fixes: 0e539ca1bb ("gfs2: Fix NULL pointer dereference in gfs2_rgrp_dump")
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Patch 541656d3a5 ("gfs2: freeze should work on read-only mounts") changed
the check for glock state in function freeze_go_sync() from "gl->gl_state
== LM_ST_SHARED" to "gl->gl_req == LM_ST_EXCLUSIVE". That's wrong and it
regressed gfs2's freeze/thaw mechanism because it caused only the freezing
node (which requests the glock in EX) to queue freeze work.
All nodes go through this go_sync code path during the freeze to drop their
SHared hold on the freeze glock, allowing the freezing node to acquire it
in EXclusive mode. But all the nodes must freeze access to the file system
locally, so they ALL must queue freeze work. The freeze_work calls
freeze_func, which makes a request to reacquire the freeze glock in SH,
effectively blocking until the thaw from the EX holder. Once thawed, the
freezing node drops its EX hold on the freeze glock, then the (blocked)
freeze_func reacquires the freeze glock in SH again (on all nodes, including
the freezer) so all nodes go back to a thawed state.
This patch changes the check back to gl_state == LM_ST_SHARED like it was
prior to 541656d3a5.
Fixes: 541656d3a5 ("gfs2: freeze should work on read-only mounts")
Cc: stable@vger.kernel.org # v5.8+
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Patch b2a846dbef ("gfs2: Ignore journal log writes for jdata holes")
tried (unsuccessfully) to fix a case in which writes were done to jdata
blocks, the blocks are sent to the ail list, then a punch_hole or truncate
operation caused the blocks to be freed. In other words, the ail items
are for jdata holes. Before b2a846dbef, the jdata hole caused function
gfs2_block_map to return -EIO, which was eventually interpreted as an
IO error to the journal, and then withdraw.
This patch changes function gfs2_get_block_noalloc, which is only used
for jdata writes, so it returns -ENODATA rather than -EIO, and when
-ENODATA is returned to gfs2_ail1_start_one, the error is ignored.
We can safely ignore it because gfs2_ail1_start_one is only called
when the jdata pages have already been written and truncated, so the
ail1 content no longer applies.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
This reverts commit b2a846dbef.
That commit changed the behavior of function gfs2_block_map to return
-ENODATA in cases where a hole (IOMAP_HOLE) is encountered and create is
false. While that fixed the intended problem for jdata, it also broke
other callers of gfs2_block_map such as some jdata block reads. Before
the patch, an encountered hole would be skipped and the buffer seen as
unmapped by the caller. The patch changed the behavior to return
-ENODATA, which is interpreted as an error by the caller.
The -ENODATA return code should be restricted to the specific case where
jdata holes are encountered during ail1 writes. That will be done in a
later patch.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
In the fail path of gfs2_check_blk_type, forgetting to call
gfs2_glock_dq_uninit will result in rgd_gh reference leak.
Signed-off-by: Zhang Qilong <zhangqilong3@huawei.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Commit fc0e38dae6 ("GFS2: Fix glock deallocation race") fixed a
sd_glock_disposal accounting bug by adding a missing atomic_dec
statement, but it failed to wake up sd_glock_wait when that decrement
causes sd_glock_disposal to reach zero. As a consequence,
gfs2_gl_hash_clear can now run into a 10-minute timeout instead of
being woken up. Add the missing wakeup.
Fixes: fc0e38dae6 ("GFS2: Fix glock deallocation race")
Cc: stable@vger.kernel.org # v2.6.39+
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Right now, we can end up calling cancel_delayed_work_sync from within
delete_work_func via gfs2_lookup_by_inum -> gfs2_inode_lookup ->
gfs2_cancel_delete_work. When that happens, it will result in a
deadlock. Instead, gfs2_inode_lookup should skip the call to
gfs2_cancel_delete_work when called from delete_work_func (blktype ==
GFS2_BLKST_UNLINKED).
Reported-by: Alexander Ahring Oder Aring <aahringo@redhat.com>
Fixes: a0e3cc65fa ("gfs2: Turn gl_delete into a delayed work")
Cc: stable@vger.kernel.org # v5.8+
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, gfs2_fitrim was not properly checking for a "live" file
system. If the file system had something to trim and the file system
was read-only (or spectator) it would start the trim, but when it starts
the transaction, gfs2_trans_begin returns -EROFS (read-only file system)
and it errors out. However, if the file system was already trimmed so
there's no work to do, it never called gfs2_trans_begin. That code is
bypassed so it never returns the error. Instead, it returns a good
return code with 0 work. All this makes for inconsistent behavior:
The same fstrim command can return -EROFS in one case and 0 in another.
This tripped up xfstests generic/537 which reports the error as:
+fstrim with unrecovered metadata just ate your filesystem
This patch adds a check for a "live" (iow, active journal, iow, RW)
file system, and if not, returns the error properly.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before commit 97fd734ba1, the local statfs_changeX inode was never
initialized for spectator mounts. However, it still checks for
spectator mounts when unmounting everything. There's no good reason to
lookup the statfs_changeX files because spectators cannot perform recovery.
It still, however, needs the master statfs file for statfs calls.
This patch adds the check for spectator mounts to init_statfs.
Fixes: 97fd734ba1 ("gfs2: lookup local statfs inodes prior to journal recovery")
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, function gfs2_meta_sync called filemap_fdatawrite to write
the address space for the metadata being synced. That's great for inodes, but
resource groups all point to the same superblock-address space, sdp->sd_aspace.
Each rgrp has its own range of blocks on which it should operate. That meant
every time an rgrp's metadata was synced, it would write all of them instead
of just the range.
This patch eliminates function gfs2_meta_sync and tailors specific metasync
functions for inodes and rgrps.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Hi,
Before this patch, function init_journal's "undo" directive jumped to label
fail_jinode_gh. But now that it does statfs initialization, it needs to
jump to fail_statfs instead. Failure to do so means that mount failures
after init_journal is successful will neglect to let go of the proper
statfs information, stranding the statfs_changeX inodes. This makes it
impossible to free its glocks, and results in:
gfs2: fsid=sda.s: G: s:EX n:2/805f f:Dqob t:EX d:UN/603701000 a:0 v:0 r:4 m:200 p:1
gfs2: fsid=sda.s: H: s:EX f:H e:0 p:1397947 [(ended)] init_journal+0x548/0x890 [gfs2]
gfs2: fsid=sda.s: I: n:6/32863 t:8 f:0x00 d:0x00000201 s:24 p:0
gfs2: fsid=sda.s: G: s:SH n:5/805f f:Dqob t:SH d:UN/603712000 a:0 v:0 r:3 m:200 p:0
gfs2: fsid=sda.s: H: s:SH f:EH e:0 p:1397947 [(ended)] gfs2_inode_lookup+0x1fb/0x410 [gfs2]
VFS: Busy inodes after unmount of sda. Self-destruct in 5 seconds. Have a nice day...
The next time the file system is mounted, it then reuses the same glocks,
which ends in a kernel NULL pointer dereference when trying to dump the
reused glock.
This patch makes the "undo" function of init_journal jump to fail_statfs
so the statfs files are properly deconstructed upon failure.
Fixes: 97fd734ba1 ("gfs2: lookup local statfs inodes prior to journal recovery")
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Gfs2 creates an address space for its rgrps called sd_aspace, but it never
called truncate_inode_pages_final on it. This confused vfs greatly which
tried to reference the address space after gfs2 had freed the superblock
that contained it.
This patch adds a call to truncate_inode_pages_final for sd_aspace, thus
avoiding the use-after-free.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Function gfs2_clear_rgrpd calls kfree(rgd->rd_bits) before calling
return_all_reservations, but return_all_reservations still dereferences
rgd->rd_bits in __rs_deltree. Fix that by moving the call to kfree below the
call to return_all_reservations.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Apply the outstanding statfs changes in the journal head to the
master statfs file. Zero out the local statfs file for good measure.
Previously, statfs updates would be read in from the local statfs inode and
synced to the master statfs inode during recovery.
We now use the statfs updates in the journal head to update the master statfs
inode instead of reading in from the local statfs inode. To preserve backward
compatibility with kernels that can't do this, we still need to keep the
local statfs inode up to date by writing changes to it. At some point in the
future, we can do away with the local statfs inodes altogether and keep the
statfs changes solely in the journal.
Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
We need to lookup the master statfs inode and the local statfs
inodes earlier in the mount process (in init_journal) so journal
recovery can use them when it attempts to recover the statfs info.
We lookup all the local statfs inodes and store them in a linked
list to allow a node to recover statfs info for other nodes in the
cluster.
Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
And read these in __get_log_header() from the log header.
Also make gfs2_statfs_change_out() non-static so it can be used
outside of super.c
Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
The gfs2_glock structure has a gl_vm member, introduced in commit 7005c3e4ae
("GFS2: Use range based functions for rgrp sync/invalidation"), which stores
the location of resource groups within their address space. This structure is
in a union with iopen glock specific fields. It was introduced because at
unmount time, the resource group objects were destroyed before flushing out any
pending resource group glock work, and flushing out such work could require
flushing / truncating the address space.
Since commit b3422cacdd ("gfs2: Rework how rgrp buffer_heads are managed"),
any pending resource group glock work is flushed out before destroying the
resource group objects. So the resource group objects will now always exist in
rgrp_go_sync and rgrp_go_inval, and we now simply compute the gl_vm values
where needed instead of caching them. This also eliminates the union.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Only initialize gl_delete for iopen glocks, but more importantly, only access
it for iopen glocks in flush_delete_work: flush_delete_work is called for
different types of glocks including rgrp glocks, and those use gl_vm which is
in a union with gl_delete. Without this fix, we'll end up clobbering gl_vm,
which results in general memory corruption.
Fixes: a0e3cc65fa ("gfs2: Turn gl_delete into a delayed work")
Cc: stable@vger.kernel.org # v5.8+
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
The comments before function glock_hash_walk had the wrong name and
an extra parameter. This simply fixes the comments.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, glock.c maintained a flag, GLF_QUEUED, which indicated
when a glock had a holder queued. It was only checked for inode glocks,
although set and cleared by all glocks, and it was only used to determine
whether the glock should be held for the minimum hold time before releasing.
The problem is that the flag is not accurate at all. If a process holds
the glock, the flag is set. When they dequeue the glock, it only cleared
the flag in cases when the state actually changed. So if the state doesn't
change, the flag may still be set, even when nothing is queued.
This happens to iopen glocks often: the get held in SH, then the file is
closed, but the glock remains in SH mode.
We don't need a special flag to indicate this: we can simply tell whether
the glock has any items queued to the holders queue. It's a waste of cpu
time to maintain it.
This patch eliminates the flag in favor of simply checking list_empty
on the glock holders.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
When flushing out its ail1 list, gfs2_write_jdata_page calls function
__block_write_full_page passing in function gfs2_get_block_noalloc.
But there was a problem when a process wrote to a jdata file, then
truncated it or punched a hole, leaving references to the blocks within
the new hole in its ail list, which are to be written to the journal log.
In writing them to the journal, after calling gfs2_block_map, function
gfs2_get_block_noalloc determined that the (hole-punched) block was not
mapped, so it returned -EIO to generic_writepages, which passed it back
to gfs2_ail1_start_one. This, in turn, performed a withdraw, assuming
there was a real IO error writing to the journal.
This might be a valid error when writing metadata to the journal, but for
journaled data writes, it does not warrant a withdraw.
This patch adds a check to function gfs2_block_map that makes an exception
for journaled data writes that correspond to jdata holes: If the iomap
get function returns a block type of IOMAP_HOLE, it instead returns
-ENODATA which does not cause the withdraw. Other errors are returned as
before.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Function gfs2_block_map had a lot of redundancy between its create and
no_create paths. This patch simplifies the code to eliminate the redundancy.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
With jdata writes, we frequently got into situations where gfs2 deadlocked
because of this calling sequence:
gfs2_ail1_start
gfs2_ail1_flush - for every tr on the sd_ail1_list:
gfs2_ail1_start_one - for every bd on the tr's tr_ail1_list:
generic_writepages
write_cache_pages passing __writepage()
calls clear_page_dirty_for_io which calls set_page_dirty:
which calls jdata_set_page_dirty which sets PageChecked.
__writepage() calls
mapping->a_ops->writepage AKA gfs2_jdata_writepage
However, gfs2_jdata_writepage checks if PageChecked is set, and if so, it
ignores the write and redirties the page. The problem is that write_cache_pages
calls clear_page_dirty_for_io, which often calls set_page_dirty(). See comments
in page-writeback.c starting with "Yes, Virginia". If it's jdata,
set_page_dirty will call jdata_set_page_dirty which will set PageChecked.
That causes a conflict because it makes it look like the page has been
redirtied by another writer, in which case we need to skip writing it and
redirty the page. That ends up in a deadlock because it isn't a "real" writer
and nothing will ever clear PageChecked.
If we do have a real writer, it will have started a transaction. So this
patch checks if a transaction is in use, and if not, it skips setting
PageChecked. That way, the page will be dirtied, cleaned, and written
appropriately.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Patch 380f7c65a7 changed gfs2_releasepage
so that it held the sd_ail_lock spin_lock for most of its processing.
It did this for some mysterious undocumented bug somewhere in the
evict code path. But in the nine years since, evict has been reworked
and fixed many times, and so have the transactions and ail list.
I can't see a reason to hold the sd_ail_lock unless it's protecting
the actual ail lists hung off the transactions. Therefore, this patch
removes the locking to increase speed and efficiency, and to further help
us rework the log flush code to be more concurrent with transactions.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
This patch is one baby step toward simplifying the journal management.
It simply changes function gfs2_ail1_empty_one from a void to an int and
makes it return a count of active items. This allows the caller to check
the return code rather than list_empty on the tr_ail1_list. This way
we can, in a later patch, combine transaction ail1 and ail2 lists.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
Before this patch, when blocks were freed, it called gfs2_meta_wipe to
take the metadata out of the pending journal blocks. It did this mostly
by calling another function called gfs2_remove_from_journal. This is
shortsighted because it does not do anything with jdata blocks which
may also be in the journal.
This patch expands the function so that it wipes out jdata blocks from
the journal as well, and it wipes it from the ail1 list if it hasn't
been written back yet. Since it now processes jdata blocks as well,
the function has been renamed from gfs2_meta_wipe to gfs2_journal_wipe.
New function gfs2_ail1_wipe wants a static view of the ail list, so it
locks the sd_ail_lock when removing items. To accomplish this, function
gfs2_remove_from_journal no longer locks the sd_ail_lock, and it's now
the caller's responsibility to do so.
I was going to make sd_ail_lock locking conditional, but the practice is
generally frowned upon. For details, see: https://lwn.net/Articles/109066/
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
This patch adds some code to enhance the log_blocks trace point. It
reports the number of free log blocks. This makes the trace point much
more useful, especially for debugging performance problems when we can
tell when the journal gets full and needs to wait for flushes, etc.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
