getblk: Avoid sleeping on wrong buf in lockless path

[cem]

Summary

If the buffer identity changed during lookup, sleeping could introduce a
lock order reversal. Since we do not know if the identity changed until we
get the lock, we must try-lock (LK_NOWAIT) only. EINTR and ERESTART error
handling becomes irrelevant, as we no longer sleep.

Reported by: kib
X-MFC-With: r363482

[kib]

sys/kern/vfs_bio.c

3868–3870

SLEEPFAIL does not make sense together with NOWAIT.

[cem]

sys/kern/vfs_bio.c

3888–3891

Should this logic be changed to avoid LK_SLEEPFAIL | LK_NOWAIT?

lockflags = LK_EXCLUSIVE | LK_INTERLOCK;

if ((flags & GB_LOCK_NOWAIT) != 0)
    lockflags |= LK_NOWAIT;
else
    lockflags |= LK_SLEEPFAIL;

[cem]

Drop LK_SLEEPFAIL for LK_NOWAIT lockless lookup path. No functional change.

[kib]

sys/kern/vfs_bio.c

3868–3870

Drop lockflags use there, pass LK_EXCLUSIVE | LK_NOWAIT directly to BUF_TIMELOCK().

3888–3891

I think it would clarify the code.

[cem]

sys/kern/vfs_bio.c

3868–3870

will do.

3888–3891

will do as a follow-up