lkml.org 
[lkml]   [2013]   [Sep]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] dcache: Translating dentry into pathname without taking rename_lock
On 09/04/2013 05:31 PM, Linus Torvalds wrote:
> On Wed, Sep 4, 2013 at 12:05 PM, Waiman Long<Waiman.Long@hp.com> wrote:
>> + rcu_read_unlock();
>> + if (read_seqretry(&rename_lock, seq))
>> + goto restart;
> Btw, you have this pattern twice, and while it's not necessarily
> incorrect, it's a bit worrisome for performance.

The rcu_read_unlock sequence in the middle of prepend_path() is not
likely to executed. So it shouldn't affect performance exception for the
conditional check.

> The rcu_read_unlock() is very possibly going to trigger an immediate
> scheduling event, so checking the sequence lock after dropping the
> read-lock sounds like it would make it much easier to hit the race
> with some rename.

I can put read_seqbegin/read_seqretry within the
rcu_read_lock/rcu_read_unlock block. However, read_seqbegin() can spin
for a while if a rename is in progress. So it depends on which is more
important, a shorter RCU critical section at the expense of more retries
or vice versa.

> I'm also a tiny bit worried about livelocking on the sequence lock in
> the presence of lots of renames, so I'm wondering if the locking
> should try to approximate what we do for the RCU lookup path: start
> off optimistically using just the RCU lock and a sequence point, but
> if that fails, fall back to taking the real lock. Maybe using a
> counter ("try twice, then get the rename lock for writing")
>
> Hmm?

Yes, I can implement a counter that switch to taking the rename_lock if
the count reaches 0. It shouldn't be too hard. That should avoid the
possibility of a livelock.

Longman


\
 
 \ /
  Last update: 2013-09-05 04:41    [W:0.091 / U:0.196 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site