Messages in this thread | | | Date | Sat, 24 Nov 2012 16:09:37 -0500 (EST) | From | Mikulas Patocka <> | Subject | Re: Recent kernel "mount" slow |
| |
On Sat, 24 Nov 2012, Jeff Chua wrote:
> On Fri, Nov 23, 2012 at 9:24 PM, Jens Axboe <axboe@kernel.dk> wrote: > > On 2012-11-22 20:21, Linus Torvalds wrote: > >> Doesn't sound like a fsdevel issue since it seems to be independent of > >> filesystems. More like some generic block layer thing. Adding Jens > >> (and quoting the whole thing) > >> > >> Jens, any ideas? Most of your stuff came in after -rc2, which would > >> fit with the fact that most of the slowdown seems to be after -rc2 > >> according to Jeff. > > > > No ideas. Looking at what went in from my side, only the rq plug sorting > > is a core change, and that should not cause any change in behaviour for > > a single device. That's commit 975927b9. > > > >> Jeff, more bisecting would be good, though. > > > > Probably required, yes... > > > This one slows mount from 0.012s to 0.168s. > > commit 62ac665ff9fc07497ca524bd20d6a96893d11071 > Author: Mikulas Patocka <mpatocka@redhat.com> > Date: Wed Sep 26 07:46:43 2012 +0200 > > blockdev: turn a rw semaphore into a percpu rw semaphore > > > There were couple of more changes to percpu-rw-semaphores after > 3.7.0-rc2 and those slows mount further from 0.168s to 0.500s. I don't > really know, but I'm suspecting these. Still bisecting.
The problem there is that you either use normal semaphores and slow down I/O or you use percpu-semaphores, you don't slow down I/O, but you slow down mount.
So it's better to slow down mount.
(if you don't use any semaphore at all, as it was in 3.6 kernel and before, there is a race condition that can crash the kernel if someone does mount and direct I/O read on the same device at the same time)
You can improve mount time if you change all occurences of synchronize_sched() in include/linux/percpu-rwsem.h to synchronize_sched_expedited().
But some people say that synchronize_sched_expedited() is bad for real time latency. (can there be something like: if (realtime) synchronize_sched(); else synchronize_sched_expedited(); ?)
Mikulas
| |