Performance regression introduced by commit b667b8673443 ("pipe: Advance tail pointer inside of wait spinlock in pipe_read()")

Performance regression introduced by commit b667b8673443 ("pipe: Advance tail pointer inside of wait spinlock in pipe_read()")

[Waiman Long]

David,

I had found that parallel kernel build became much slower when a
5.5-based kernel is used. On a 2-socket 96-thread x86-64 system, the
"make -j88" time increased from less than 3 minutes with the 5.4 kernel
to more than double with the 5.5 kernel.

So I used bisection to try to find the culprit:

b667b867344301e24f21d4a4c844675ff61d89e1 is the first bad commit
commit b667b867344301e24f21d4a4c844675ff61d89e1
Author: David Howells <dhowells@redhat.com>
Date:   Tue Sep 24 16:09:04 2019 +0100

    pipe: Advance tail pointer inside of wait spinlock in pipe_read()
   
    Advance the pipe ring tail pointer inside of wait spinlock in
pipe_read()
    so that the pipe can be written into with kernel notifications from
    contexts where pipe->mutex cannot be taken.
   
    Signed-off-by: David Howells <dhowells@redhat.com>

diff --git a/fs/pipe.c b/fs/pipe.c
index 69afeab8a73a..ea134f69a292 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -325,9 +325,14 @@ pipe_read(struct kiocb *iocb, struct iov_iter *to)
 
                        if (!buf->len) {
                                pipe_buf_release(pipe, buf);
+                               spin_lock_irq(&pipe->wait.lock);
                                tail++;
                                pipe->tail = tail;
-                               do_wakeup = 1;
+                               do_wakeup = 0;
+                               wake_up_interruptible_sync_poll_locked(
+                                       &pipe->wait, EPOLLOUT |
EPOLLWRNORM);
+                               spin_unlock_irq(&pipe->wait.lock);
+                               kill_fasync(&pipe->fasync_writers,
SIGIO, POLL_O
                        }

I guess the make command may make heavy use of pipe. The adding of
spinlock code in your patch may probably over-serialize the pipe
operation. Could you achieve the same functionality without adding a lock?

Cheers,
Longman

Re: Performance regression introduced by commit b667b8673443 ("pipe: Advance tail pointer inside of wait spinlock in pipe_read()")

[Waiman Long]

On 1/17/20 12:29 PM, Waiman Long wrote:
> On 1/17/20 12:05 PM, Linus Torvalds wrote:
>> [ on mobile, sorry for html crud ]
>>
>> On Fri, Jan 17, 2020, 08:53 Waiman Long <longman@redhat.com
>> <mailto:longman@redhat.com>> wrote:
>>
>>
>>     I had found that parallel kernel build became much slower when a
>>     5.5-based kernel is used. On a 2-socket 96-thread x86-64 system, the
>>     "make -j88" time increased from less than 3 minutes with the 5.4
>>     kernel
>>     to more than double with the 5.5 kernel.
>>
>>
>> I suspect you may have hit the same bug in the GNU make jobserver
>> that I did.
>>
>> It's timing-sensitive, and under the right circumstances the make
>> jobserver loses job tickets to other jobservers that have a child
>> that died, but they are blocked waiting for a new ticket, so they
>> aren't releasing (or re-using) the one that the child death would
>> free up.
>>
>> End result: a big lack of parallelism, and a much slower build.
>>
>> GNU make v4.2.1 is buggy. The fix was done over two years ago, but
>> there hasn't been a new release since then, so a lot of distributions
>> have the buggy version..
>>
>> The fix is commit b552b05 ("[SV 51159] Use a non-blocking read with
>> pselect to avoid hangs.") In the make the git tree.
>>
>>
>>      Linus
>
> Thanks for the information.
>
> Yes, I did use make v4.2.1 which is the version that is shipped in
> RHEL8. I will build new make and try it.
>
> Thanks,
> Longman
>
I built a make with the lastest make git tree and the problem was gone
with the new make. So it was a bug in make not the kernel. Sorry for the
noise.

Cheers,
Longman

Re: Performance regression introduced by commit b667b8673443 ("pipe: Advance tail pointer inside of wait spinlock in pipe_read()")

[Linus Torvalds]

On Fri, Jan 17, 2020 at 10:11 AM Waiman Long <longman@redhat.com> wrote:
>
> I built a make with the lastest make git tree and the problem was gone
> with the new make. So it was a bug in make not the kernel. Sorry for the
> noise.

I think I spent about three days trying to figure it out. At least it
felt that way. I looked at the pipe code a _lot_, also blaming the
kernel for obvious reasons.

             Linus

Re: Performance regression introduced by commit b667b8673443 ("pipe: Advance tail pointer inside of wait spinlock in pipe_read()")

[Akemi Yagi]

On Fri, Jan 17, 2020 at 10:11 AM Waiman Long <longman@redhat.com> wrote:
>
> On 1/17/20 12:29 PM, Waiman Long wrote:
> > On 1/17/20 12:05 PM, Linus Torvalds wrote:

> >> GNU make v4.2.1 is buggy. The fix was done over two years ago, but
> >> there hasn't been a new release since then, so a lot of distributions
> >> have the buggy version..
> >>
> >> The fix is commit b552b05 ("[SV 51159] Use a non-blocking read with
> >> pselect to avoid hangs.") In the make the git tree.
> >>      Linus
> >
> > Yes, I did use make v4.2.1 which is the version that is shipped in
> > RHEL8. I will build new make and try it.

> > Longman
> >
> I built a make with the lastest make git tree and the problem was gone
> with the new make. So it was a bug in make not the kernel. Sorry for the
> noise.
>
> Longman

If you are using RHEL8, building your own make is the only solution at
this time. There is a bugzilla entry filed for this make bug but the
progress is slow:

https://bugzilla.redhat.com/show_bug.cgi?id=1774790

The same bug in Fedora make was dealt with fairly quickly, thanks to
the great "pressure" from Linus. ;-)

Akemi