Re: [PATCH] btrfs: Don't block system suspend during fstrim
From: Luca Stefani
Date: Mon Sep 02 2024 - 05:31:26 EST
On 02/09/24 11:17, Qu Wenruo wrote:
在 2024/9/2 18:42, Qu Wenruo 写道:
在 2024/9/2 18:30, Luca Stefani 写道:
On Mon, 2 Sept 2024 at 10:49, Qu Wenruo <quwenruo.btrfs@xxxxxxx
<mailto:quwenruo.btrfs@xxxxxxx>> wrote:
在 2024/9/2 18:02, Luca Stefani 写道:
> Any update on this? It's not critical but I'd like to know if
it's in
> some part proper.
> Thanks, Luca.
Sorry I didn't see your patch in the list, thus sent a different
fix for
it later:
https://lore.kernel.org/linux-btrfs/20240830185113.GW25962@xxxxxxxxxxxxx/T/#t <https://lore.kernel.org/linux-btrfs/20240830185113.GW25962@xxxxxxxxxxxxx/T/#t>
>> Sometimes the system isn't able to suspend because
>> the task responsible for trimming the device isn't
>> able to finish in time.
>>
>> Since discard isn't a critical call it can be interrupted
>> at any time, we can simply report the amount of discarded
>> bytes in such cases and stop the trim.
>>
>> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=219180
<https://bugzilla.kernel.org/show_bug.cgi?id=219180>
>> Signed-off-by: Luca Stefani <luca.stefani.ge1@xxxxxxxxx
<mailto:luca.stefani.ge1@xxxxxxxxx>>
>> ---
>> I have no idea if that's correct, just something I implemented
>> looking at the same solution made in ext4 by 5229a658f645.
>>
>> The patch in itself seems to solve the issue.
>>
>> repro is as follows:
>> sudo /sbin/fstrim --listed-in /etc/fstab:/proc/self/mountinfo
>> --verbose --quiet-unsupported &
>> sudo ./sleepgraph.py -m mem -rtcwake 5
>>
>> [836563.289069] PM: suspend exit
>> [836563.909298] PM: suspend entry (s2idle)
>> [836563.935447] Filesystems sync: 0.026 seconds
>> [836563.951391] Freezing user space processes
>> [836583.958957] Freezing user space processes failed after
20.007
>> seconds (1 tasks refusing to freeze, wq_busy=0):
>> [836583.959582] task:fstrim state:D stack:0 pid:241865
>> tgid:241865 ppid:241864 flags:0x00004006
>> [836583.959592] Call Trace:
>> [836583.959595] <TASK>
>> [836583.959600] __schedule+0x400/0x1720
>> [836583.959612] ? mod_delayed_work_on+0xa4/0xb0
>> [836583.959622] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959628] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959631] ? blk_mq_flush_plug_list.part.0+0x1e3/0x610
>> [836583.959640] schedule+0x27/0xf0
>> [836583.959644] schedule_timeout+0x12f/0x160
>> [836583.959652] io_schedule_timeout+0x51/0x70
>> [836583.959657] wait_for_completion_io+0x8a/0x160
>> [836583.959663] submit_bio_wait+0x60/0x90
>> [836583.959671] blkdev_issue_discard+0x91/0x100
>> [836583.959680] btrfs_issue_discard+0xc4/0x140
>> [836583.959689] btrfs_discard_extent+0x241/0x2a0
>> [836583.959695] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959702] do_trimming+0xd2/0x240
>> [836583.959712] trim_bitmaps+0x350/0x4c0
>> [836583.959723] btrfs_trim_block_group+0xb8/0x110
>> [836583.959729] btrfs_trim_fs+0x118/0x440
>> [836583.959734] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959738] ? security_capable+0x41/0x70
>> [836583.959746] btrfs_ioctl_fitrim+0x113/0x180
>> [836583.959752] btrfs_ioctl+0xdaf/0x2670
>> [836583.959759] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959763] ? ioctl_has_perm.constprop.0.isra.0+0xd8/0x130
>> [836583.959774] __x64_sys_ioctl+0x94/0xd0
>> [836583.959782] do_syscall_64+0x82/0x160
>> [836583.959790] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959793] ? syscall_exit_to_user_mode+0x72/0x220
>> [836583.959799] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959802] ? do_syscall_64+0x8e/0x160
>> [836583.959807] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959811] ? do_sys_openat2+0x9c/0xe0
>> [836583.959821] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959825] ? syscall_exit_to_user_mode+0x72/0x220
>> [836583.959828] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959832] ? do_syscall_64+0x8e/0x160
>> [836583.959835] ? syscall_exit_to_user_mode+0x72/0x220
>> [836583.959838] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959842] ? do_syscall_64+0x8e/0x160
>> [836583.959845] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959849] ? do_syscall_64+0x8e/0x160
>> [836583.959851] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959855] ? do_syscall_64+0x8e/0x160
>> [836583.959858] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959861] ? do_syscall_64+0x8e/0x160
>> [836583.959864] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959868] ? srso_alias_return_thunk+0x5/0xfbef5
>> [836583.959873] entry_SYSCALL_64_after_hwframe+0x76/0x7e
>> [836583.959878] RIP: 0033:0x7f3e4261af2d
>> [836583.959944] RSP: 002b:00007ffec002f400 EFLAGS: 00000246
ORIG_RAX:
>> 0000000000000010
>> [836583.959950] RAX: ffffffffffffffda RBX: 00007ffec002f570 RCX:
>> 00007f3e4261af2d
>> [836583.959952] RDX: 00007ffec002f470 RSI: 00000000c0185879 RDI:
>> 0000000000000003
>> [836583.959955] RBP: 00007ffec002f450 R08: 0000562d74da7010 R09:
>> 00007ffec002e7f2
>> [836583.959957] R10: 0000000000000000 R11: 0000000000000246 R12:
>> 0000562d74daafc0
>> [836583.959960] R13: 0000000000000003 R14: 0000562d74daa970 R15:
>> 0000562d74daad40
>> [836583.959967] </TASK>
>> ---
>> fs/btrfs/extent-tree.c | 20 ++++++++++++++++----
>> 1 file changed, 16 insertions(+), 4 deletions(-)
>>
>> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
>> index feec49e6f9c8..7e4c1d4f2f7c 100644
>> --- a/fs/btrfs/extent-tree.c
>> +++ b/fs/btrfs/extent-tree.c
>> @@ -16,6 +16,7 @@
>> #include <linux/percpu_counter.h>
>> #include <linux/lockdep.h>
>> #include <linux/crc32c.h>
>> +#include <linux/freezer.h>
>> #include "ctree.h"
>> #include "extent-tree.h"
>> #include "transaction.h"
>> @@ -6361,6 +6362,11 @@ void
btrfs_error_unpin_extent_range(struct
>> btrfs_fs_info *fs_info, u64 start, u6
>> unpin_extent_range(fs_info, start, end, false);
>> }
>> +static bool btrfs_trim_interrupted(void)
>> +{
>> + return fatal_signal_pending(current) || freezing(current);
>> +}
>> +
>> /*
>> * It used to be that old block groups would be left around
forever.
>> * Iterating over them would be enough to trim unused space.
Since we
>> @@ -6459,8 +6465,8 @@ static int btrfs_trim_free_extents(struct
>> btrfs_device *device, u64 *trimmed)
>> start += len;
>> *trimmed += bytes;
>> - if (fatal_signal_pending(current)) {
>> - ret = -ERESTARTSYS;
>> + if (btrfs_trim_interrupted()) {
>> + ret = 0;
>> break;
Here we should still return the same error number other than 0,
to let
the caller know the operation is interrupted, other than finished
normally.
Here I was following how ext4 did it, my explanation for that was that
the kernel
may have still discarded some data before the thread was interrupted
thus it made
sense to report success.
In that case, progress is reported through fstrim_range structure, not
through the return value.
Even if we returned some error code, the fstrim_range::len is still
updated to indicate the progress.
So we need to keep the error code.
>> }
>> @@ -6508,6 +6514,9 @@ int btrfs_trim_fs(struct btrfs_fs_info
*fs_info,
>> struct fstrim_range *range)
>> cache = btrfs_lookup_first_block_group(fs_info,
range->start);
>> for (; cache; cache = btrfs_next_block_group(cache)) {
>> + if (btrfs_trim_interrupted())
>> + break;
>> +
The same here.
>> if (cache->start >= range_end) {
>> btrfs_put_block_group(cache);
>> break;
>> @@ -6547,17 +6556,20 @@ int btrfs_trim_fs(struct btrfs_fs_info
>> *fs_info, struct fstrim_range *range)
>> mutex_lock(&fs_devices->device_list_mutex);
>> list_for_each_entry(device, &fs_devices->devices,
dev_list) {
>> + if (btrfs_trim_interrupted())
>> + break;
>> +
The same here.
Furthermore, I think we may not need the extra checks.
The fstrim is based on block groups, and a block group is
normally 1GiB,
at most 10GiB (for RAID0/5/6/10 only), thus exiting at each block
group
boundary should be enough to meet the hibernation/suspension
timeout.
That's probably true, but 10seconds here wasn't enough and forcing the
early return in the other cases
was also required.
I tried the current patch you linked earlier in my testing and that was
the conclusion that led to me adding more checks.
My bad, I forgot that we have free extents trimming, which is not
limited to block group boundary, and that may be the root cause.
So in that case your extra checks are indeed needed.
Just need to change the return value.
>> if (test_bit(BTRFS_DEV_STATE_MISSING,
&device->dev_state))
>> continue;
>> ret = btrfs_trim_free_extents(device, &group_trimmed);
>> +
>> + trimmed += group_trimmed;
>> if (ret) {
>> dev_failed++;
>> dev_ret = ret;
>> break;
>> }
>> -
>> - trimmed += group_trimmed;
Any special reason moving the code here?
Same as not returning errno before in case of interrupt, I checked the
code paths and it's still
possible to trim some data (group_trimmed != 0) even in case of failure.
Oh, then it's fine.
Except the return code, everything looks fine to me now.
Forgot to mention that, even for error case, we should copy the
fstrim_range structure to the ioctl parameter to indicate any progress
we made.
This seems to be already the case.
range->len = trimmed; is always executed regardless of previous failures
and there doesn't seem to be any early return.
Will try adding back the errno and try the repro.
Thanks.
Thanks,
Qu
Just please update the commit message to explicitly mention that, we
have a free extent discarding phase, which can trim a lot of unallocated
space, and there is no limits on the trim size (unlike the block group
part).
Thanks,
Qu
Thanks,
Qu
>> }
>> mutex_unlock(&fs_devices->device_list_mutex);
>