Re: [PATCH] btrfs: do not leak -ERESTARTSYS from FITRIM to userspace

[Luca Stefani]

On Wed, 28 Oct 2026 at 09:57, Qu Wenruo <wqu@xxxxxxxx> wrote:
>
>
>
> 在 2026/10/28 19:21, Luca Stefani 写道:
> > On Wed, 28 Oct 2026 at 07:57, Zhiwei Liang via B4 Relay
> > <devnull+zhiwei.liang.zliang.me@xxxxxxxxxx> wrote:
> >>
> >> From: Zhiwei Liang <zhiwei.liang@xxxxxxxxx>
> >>
> >> When a suspend starts during FITRIM, fstrim can fail with:
> >>
> >>    fstrim: /: FITRIM ioctl failed: Unknown error 512
> >>
> >> On a Fedora 44 laptop with kernel 7.2.7, 3 of the last 21 weekly
> >> fstrim.service runs were in progress when an automatic suspend started.
> >> All three failed this way and left the filesystem untrimmed.
> >>
> >> Since commit 69313850dce3 ("btrfs: add cancellation points to trim
> >> loops"), the trim loops return -ERESTARTSYS when
> >> btrfs_trim_interrupted() sees freezing(current), and FITRIM returns it
> >> to userspace. Since commit bfb670b9183b ("btrfs: handle user interrupt
> >> properly in btrfs_trim_fs()"), btrfs_trim_fs() returns it right away.
> >>
> >> -ERESTARTSYS is only valid while a signal is pending, because only the
> >> signal handling on syscall exit turns it into a restart. But
> >> freeze_processes() first sets pm_freezing, which makes freezing() true
> >> for every user task. Then it walks the task list and sends each task a
> >> fake signal. The trim loops check btrfs_trim_interrupted() after every
> >> extent. If a check falls in this window, the loop returns -ERESTARTSYS
> >> with no signal pending. Nothing restarts the syscall. On that laptop,
> >> with a few thousand threads, freezing user space took 2 to 5 ms in each
> >> of the three suspends above, and all three runs hit this window.
> >>
> >> Make btrfs_trim_interrupted() treat a freeze as an interruption only
> >> once the fake signal has arrived, i.e. when
> >> freezing(current) && signal_pending(current). freeze_processes() sends
> >> that signal in the same call. The trim loop stops at the next check
> >> after it, so suspend is not delayed.
> > Looking at how ext4 handles this scenario it seems to me they don't set
> > the error code to -ERESTARTSYS and (from a quick glance) return the
> > bytes discarded to the caller.
> > And like before they continue the discard (which will skip/fail again) until
> > there's nothing to be done.
> >
> > I don't know what the intention of btrfs: handle user interrupt
> > properly in btrfs_trim_fs()
> > was, maybe we should just swallow the ERESTARTSYS case in those scenarios.
>
> I agree on that btrfs is the exception, we should follow ext4/xfs, not
> rely on some tricky signal handling behavior for suspension.
When I updated btrfs_trim_interrupted I used the same "checks" as
seen in ext4 so at least on that part it's identical.

What differs is our error handling: we set/handle -ERESTARTSYS, ext4 does not.
ext4_try_to_trim_range only ever succeeds by returning the count of
trimmed bytes
even if it checks for trim interruptions.

I think if we remove ret = -ERESTARTSYS; from the
btrfs_trim_interrupted branches
we'd align the behavior (plus reverting the previous change as that's
now useless)
>
> I think the tokens of the author should be spent on unifying the btrfs
> fstrim suspension behavior instead.
>
>
> >>
> >> The restarted syscall re-reads struct fstrim_range from user memory.
> >> When the device phase is interrupted, btrfs_trim_fs() has already stored
> >> the trimmed byte count in range->len, and btrfs_ioctl_fitrim() copies
> >> the struct back unconditionally. The restart is then asked to trim
> >> [start, start + trimmed) instead of the requested range, or fails with
> >> -EINVAL when nothing was counted yet. Skip the copy_to_user() for
> >> -ERESTARTSYS. In all other cases the partial count is still copied back.
> >>
> >> Tested in a VM with pm_test=freezer. Without the patch, 18 of 20 freezes
> >> during a trim leaked errno 512, and 10 of 10 restarts after a
> >> device-phase interruption failed with -EINVAL. With the patch, all 30
> >> runs restarted and trimmed the full amount.
> >>
> >> Fixes: 69313850dce3 ("btrfs: add cancellation points to trim loops")
> >> CC: stable@xxxxxxxxxxxxxxx # 5.15+
> >> Signed-off-by: Zhiwei Liang <zhiwei.liang@xxxxxxxxx>
> >> ---
> >>   fs/btrfs/free-space-cache.h | 3 ++-
> >>   fs/btrfs/ioctl.c            | 7 +++++++
> >>   2 files changed, 9 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/fs/btrfs/free-space-cache.h b/fs/btrfs/free-space-cache.h
> >> index e22443598b..cba2f5fd6b 100644
> >> --- a/fs/btrfs/free-space-cache.h
> >> +++ b/fs/btrfs/free-space-cache.h
> >> @@ -58,7 +58,8 @@ static inline bool btrfs_free_space_trimming_bitmap(
> >>
> >>   static inline bool btrfs_trim_interrupted(void)
> >>   {
> >> -       return fatal_signal_pending(current) || freezing(current);
> >> +       return fatal_signal_pending(current) ||
> >> +              (freezing(current) && signal_pending(current));
> >>   }
> >>
> >>   /*
> >> diff --git a/fs/btrfs/ioctl.c b/fs/btrfs/ioctl.c
> >> index ac070c745b..b02c59cbb0 100644
> >> --- a/fs/btrfs/ioctl.c
> >> +++ b/fs/btrfs/ioctl.c
> >> @@ -529,6 +529,13 @@ static noinline int btrfs_ioctl_fitrim(struct btrfs_fs_info *fs_info,
> >>          range.minlen = max(range.minlen, minlen);
> >>          ret = btrfs_trim_fs(fs_info, &range);
> >>
> >> +       /*
> >> +        * A restart re-reads the range from user space, so do not clobber it
> >> +        * with the partial len the rerun would otherwise take as its input.
> >> +        */
> >> +       if (ret == -ERESTARTSYS)
> >> +               return ret;
> >> +
> >>          if (copy_to_user(arg, &range, sizeof(range)))
> >>                  return -EFAULT;
> >>
> >>
> >> ---
> >> base-commit: e17e64baf63f7c85bec209ed73f14c1b2a7e39fb
> >> change-id: 20261028-btrfs-fitrim-freezer-d4cba16e7e13
> >>
> >> Best regards,
> >> --
> >> Zhiwei Liang <zhiwei.liang@xxxxxxxxx>
> >>
> >>
>