On Wed, May 28, 2014 at 12:43 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> - the shrinker list logic depends on the actual freeing of the dentry
> to be delayed until the RCU grace period (already true for RCU-lookup
> dentries)
Side note: I just unconditionally removed the immediate __d_free()
case, but it could as well have become
- if (!(dentry->d_flags & DCACHE_RCUACCESS))
+ if (!(dentry->d_flags & (DCACHE_RCUACCESS|DCACHE_SHRINK_LIST))
__d_free(&dentry->d_u.d_rcu);
else
call_rcu(&dentry->d_u.d_rcu, __d_free);
instead.
Anyway, one reason I looked at this is that most of the threads in
Mika's NMI watchdog traces were stuck on _raw_spin_lock() in the whole
d_walk() thing, and I had a *really* hard time convincing myself that
this was all safe without the RCU lock. I'm wondering if Mika perhaps
has CONFIG_PREEMPT_RCU set, which means that spinlocks (or the
rename_lock sequence lock) do not end up being RCU-safe points.
And once I started worrying about that, the whole "ok, who calls
dentry_kill() when shrinking races with dput()" question started just
making me worry about that whole DCACHE_MAY_FREE thing. Because one
of the things that Miklos/Al's series of patches did was to drop some
of the RCU locking, since it wasn't "necessary" any more as far as the
lru list itself was concerned..
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
