On 11/02/2020 23:30, Jens Axboe wrote:
> On 2/11/20 1:20 PM, Pavel Begunkov wrote:
>> Hi,
>>
>> I've been looking for hashed io-wq enqueuing, but not fully understand the
>> issue. As I remember, it was something about getting around blocking for
>> buffered I/O. Is that so? Is there any write-up of a thread for this issue?
> 
> Not sure if there's a writeup, but the issue is that buffered writes all
> need to grab the per-inode mutex. Hence if you don't serialize these writes,
> you end up having potentially quite a few threads banging on the same mutex.
> This causes a high level of lock contention (and scheduling). By serializing
> by inode hash we avoid that, and yield the same performance.
> 
> Dave Chinner is working on lifting this restriction, at which point we'll
> have to gate the hashing based on whether or not the fs is smart or dumb
> when it comes to buffered writes and locking.

Got it, thanks!

>> My case is 2-fd request, and I'm not sure how it should look. For splice() it's
>> probably Ok to hash the non-pipe end (if any), but I wonder how it should work,
>> if, for example, the restriction will be removed.
> 
> Probably just do the same, but for the output fd only (and following the
> same restrictions in terms of file type).
> 

-- 
Pavel Begunkov