User mode stalls - can it be...?

User mode stalls - can it be...?

[Roger Larsson]

Hi,


I have an idea why we get the stalls!

If we end up in a situation where our write attempts are rejected
due to lack of resources - we will continue happily with next and
next...

Count hits zero and we return

But kswapd finds out that there is more to do and calls another
shrink_mmap...
(and need_resched is not set due to lower priority)

Gives, busy wait in shrink_mmap until pages are written...

I tried to renice 15 kswapd with improved results (I am using my patched
kernels)
Another approach would be to always sleep...

/RogerL

(Sent privately to Quintela and Riel - but it might be of value for
 someone else too...)

--
Home page:
  http://www.norran.net/nra02596/
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: User mode stalls - can it be...?

[Roger Larsson]

Roger Larsson wrote:

What about this modification (buffer.c):

 static void sync_page_buffers(struct buffer_head *bh, int wait)
{
	struct buffer_head * tmp = bh;

	do {
		struct buffer_head *p = tmp;
		tmp = tmp->b_this_page;
		if (buffer_locked(p)) {
			if (wait)
				__wait_on_buffer(p);
		} else if (buffer_dirty(p))
			ll_rw_block(WRITE, 1, &p);
	} while (tmp != bh);
}

-to-

static void sync_page_buffers(struct buffer_head *bh, int wait)
{
	struct buffer_head * tmp = bh;

	do {
		struct buffer_head *p = tmp;
		tmp = tmp->b_this_page;
		if (!buffer_locked(p) &&
                    buffer_dirty(p))
			ll_rw_block(WRITE, 1, &p);
		if (wait)
			__wait_on_buffer(p);
	} while (tmp != bh);
}


Not tested... (tomorrow - good night)

/RogerL

> 
> Hi,
> 
> I have an idea why we get the stalls!
> 
> If we end up in a situation where our write attempts are rejected
> due to lack of resources - we will continue happily with next and
> next...
> 
> Count hits zero and we return
> 
> But kswapd finds out that there is more to do and calls another
> shrink_mmap...
> (and need_resched is not set due to lower priority)
> 
> Gives, busy wait in shrink_mmap until pages are written...
> 
> I tried to renice 15 kswapd with improved results (I am using my patched
> kernels)
> Another approach would be to always sleep...
> 
> /RogerL
> 
> (Sent privately to Quintela and Riel - but it might be of value for
>  someone else too...)
> 
> --
> Home page:
>   http://www.norran.net/nra02596/
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org.  For more info on Linux MM,
> see: http://www.linux.eu.org/Linux-MM/

--
Home page:
  http://www.norran.net/nra02596/
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: User mode stalls - can it be...?

[Stephen C. Tweedie]

Hi,

On Wed, Jul 05, 2000 at 03:29:22AM +0200, Roger Larsson wrote:
> 
> static void sync_page_buffers(struct buffer_head *bh, int wait)
> {
> 	struct buffer_head * tmp = bh;
> 
> 	do {
> 		struct buffer_head *p = tmp;
> 		tmp = tmp->b_this_page;
> 		if (!buffer_locked(p) &&
>                     buffer_dirty(p))
> 			ll_rw_block(WRITE, 1, &p);
> 		if (wait)
> 			__wait_on_buffer(p);
> 	} while (tmp != bh);

Yikes no.  That will make the syncer wait for every single write
synchronously.  It will cause a page full of 1k buffers to be written
out as four separate 1k writes instead of being streamed.  This will
_kill_ performance.

Cheers,
 Stephen
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/

Re: User mode stalls - can it be...?

[Roger Larsson]

"Stephen C. Tweedie" wrote:
> 
> Hi,
> 
> On Wed, Jul 05, 2000 at 03:29:22AM +0200, Roger Larsson wrote:
> >
> > static void sync_page_buffers(struct buffer_head *bh, int wait)
> > {
> >       struct buffer_head * tmp = bh;
> >
> >       do {
> >               struct buffer_head *p = tmp;
> >               tmp = tmp->b_this_page;
> >               if (!buffer_locked(p) &&
> >                     buffer_dirty(p))
> >                       ll_rw_block(WRITE, 1, &p);
> >               if (wait)
> >                       __wait_on_buffer(p);
> >       } while (tmp != bh);
> 
> Yikes no.  That will make the syncer wait for every single write
> synchronously.  It will cause a page full of 1k buffers to be written
> out as four separate 1k writes instead of being streamed.  This will
> _kill_ performance.
> 
> Cheers,
>  Stephen


Well, 'wait' is not set at each call...

But I agree that it was not a wery bright idea...
(I am more awake now :-)

/RogerL

--
Home page:
  http://www.norran.net/nra02596/
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux.eu.org/Linux-MM/