2003-03-17 23:46:54

by Nigel Cunningham

[permalink] [raw]
Subject: [PATCH] Don't refill pcp lists during SWSUSP.

Hi.

Here's another patch (the last for a little while, I promise!). It stops
the pcp lists from being refilled while SWSUSP is running. Despite the
comment in the page, drain_local_pages does only need to get called once
right now, but I have patches coming that will (DV) change that. This
patch is thus groundwork for them.

I see Linus has just done 2.5.65 - I'll rediff if need be.

Regards,

Nigel
--
Nigel Cunningham
495 St Georges Road South, Hastings 4201, New Zealand

Be diligent to present yourself approved to God as a workman who does
not need to be ashamed, handling accurately the word of truth.
-- 2 Timothy 2:14, NASB.

diff -ruN linux-2.5.64-02-free-page-map/include/linux/suspend.h linux-2.5.64-03-do-not-refill-hot-pages/include/linux/suspend.h
--- linux-2.5.64-02-free-page-map/include/linux/suspend.h 2003-03-10 12:36:16.000000000 +1300
+++ linux-2.5.64-03-do-not-refill-hot-pages/include/linux/suspend.h 2003-03-18 11:21:53.000000000 +1200
@@ -63,6 +63,8 @@
extern unsigned int nr_copy_pages __nosavedata;
extern suspend_pagedir_t *pagedir_nosave __nosavedata;

+extern unsigned int suspend_task;
+
/* Communication between kernel/suspend.c and arch/i386/suspend.c */

extern void do_magic_resume_1(void);
diff -ruN linux-2.5.64-02-free-page-map/kernel/suspend.c linux-2.5.64-03-do-not-refill-hot-pages/kernel/suspend.c
--- linux-2.5.64-02-free-page-map/kernel/suspend.c 2003-03-18 10:54:31.000000000 +1200
+++ linux-2.5.64-03-do-not-refill-hot-pages/kernel/suspend.c 2003-03-18 11:21:53.000000000 +1200
@@ -68,6 +68,7 @@
extern int sys_sync(void);

unsigned char software_suspend_enabled = 0;
+unsigned int suspend_task = 0;

#define SUSPEND_CONSOLE (MAX_NR_CONSOLES-1)
/* With SUSPEND_CONSOLE defined, it suspend looks *really* cool, but
@@ -232,6 +233,8 @@
}
} while(todo);

+ suspend_task = current->pid;
+
printk( "|\n" );
BUG_ON(in_atomic());
return 0;
@@ -253,6 +256,7 @@
} while_each_thread(g, p);

read_unlock(&tasklist_lock);
+ suspend_task = 0;
printk( " done\n" );
MDELAY(500);
}
diff -ruN linux-2.5.64-02-free-page-map/mm/page_alloc.c linux-2.5.64-03-do-not-refill-hot-pages/mm/page_alloc.c
--- linux-2.5.64-02-free-page-map/mm/page_alloc.c 2003-03-18 10:53:22.000000000 +1200
+++ linux-2.5.64-03-do-not-refill-hot-pages/mm/page_alloc.c 2003-03-18 11:21:53.000000000 +1200
@@ -486,6 +486,10 @@
* Really, prep_compound_page() should be called from __rmqueue_bulk(). But
* we cheat by calling it from here, in the order > 0 path. Saves a branch
* or two.
+ *
+ * While suspending, we don't use the pcp structure. It mucks up our
+ * accounting for all the pages and necessitates calling drain_local_pages
+ * multiple times.
*/

static struct page *buffered_rmqueue(struct zone *zone, int order, int cold)
@@ -493,7 +497,7 @@
unsigned long flags;
struct page *page = NULL;

- if (order == 0) {
+ if ((order == 0) && (!suspend_task)) {
struct per_cpu_pages *pcp;

pcp = &zone->pageset[get_cpu()].pcp[cold];
@@ -700,7 +704,7 @@
void __free_pages(struct page *page, unsigned int order)
{
if (!PageReserved(page) && put_page_testzero(page)) {
- if (order == 0)
+ if ((order == 0) && (!suspend_task))
free_hot_page(page);
else
__free_pages_ok(page, order);




2003-03-18 00:27:16

by Andrew Morton

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

Nigel Cunningham <[email protected]> wrote:
>
> +extern unsigned int suspend_task;


Please do:

#ifdef CONFIG_SOFTWARE_SUSPEND
unsigned int suspend_task;
#else
#define suspend_task 0
#endif

so the compiler can remove the few fast-path instructions which you have
added.

>
> + suspend_task = current->pid;
> +

Zero is a valid PID (the idle task...). It might be clearer to make
suspend_task a task_struct*.

2003-03-18 05:18:27

by Nigel Cunningham

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

On Tue, 2003-03-18 at 12:05, Andrew Morton wrote:
> Nigel Cunningham <[email protected]> wrote:
> >
> > +extern unsigned int suspend_task;
>
>
> Please do:
>
> #ifdef CONFIG_SOFTWARE_SUSPEND
> unsigned int suspend_task;
> #else
> #define suspend_task 0
> #endif
>
> so the compiler can remove the few fast-path instructions which you have
> added.

Okee doke. Will do.

I said I was going to use dynamically allocated bitmaps instead of page
flags. Do you mind if I do use pageflags after all (at least for the
moment)? I've used one in the generate_free_page_map patch, and need one
more to mark pages which will be saved in another pageset.

> >
> > + suspend_task = current->pid;
> > +
>
> Zero is a valid PID (the idle task...). It might be clearer to make
> suspend_task a task_struct*.

Mmm, but I will use suspend_task elsewhere and wanted to make the test
simple (zero = not suspending at the moment; since init will never be
the suspend task, it's safe usage).

Regards,

Nigel

--
Nigel Cunningham
495 St Georges Road South, Hastings 4201, New Zealand

Be diligent to present yourself approved to God as a workman who does
not need to be ashamed, handling accurately the word of truth.
-- 2 Timothy 2:14, NASB.

2003-03-18 06:26:27

by Nigel Cunningham

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

On Tue, 2003-03-18 at 18:20, Andrew Morton wrote:
> Nigel Cunningham <[email protected]> wrote:
> > I said I was going to use dynamically allocated bitmaps instead of page
> > flags. Do you mind if I do use pageflags after all (at least for the
> > moment)? I've used one in the generate_free_page_map patch, and need one
> > more to mark pages which will be saved in another pageset.
> >
>
> I think it'd be best to avoid using more flags if poss. We are getting
> short, and designing for this now is probably less trauma than going
> through and redoing your stuff later.

Okay. I'll add the dynamic bitmap code and redo the previous patch.

Regards,

Nigel

--
Nigel Cunningham
495 St Georges Road South, Hastings 4201, New Zealand

Be diligent to present yourself approved to God as a workman who does
not need to be ashamed, handling accurately the word of truth.
-- 2 Timothy 2:14, NASB.

2003-03-18 06:32:25

by Nigel Cunningham

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

Here's the updated version.

Regards,

Nigel

diff -ruN linux-2.5.65-02/include/linux/suspend.h linux-2.5.65-03/include/linux/suspend.h
--- linux-2.5.65-02/include/linux/suspend.h 2003-03-10 12:36:16.000000000 +1300
+++ linux-2.5.65-03/include/linux/suspend.h 2003-03-18 17:56:37.000000000 +1200
@@ -63,6 +63,8 @@
extern unsigned int nr_copy_pages __nosavedata;
extern suspend_pagedir_t *pagedir_nosave __nosavedata;

+extern unsigned int suspend_task;
+
/* Communication between kernel/suspend.c and arch/i386/suspend.c */

extern void do_magic_resume_1(void);
@@ -79,6 +81,7 @@
static inline void software_suspend(void)
{
}
+#define suspend_task 0
#define software_resume() do { } while(0)
#define register_suspend_notifier(a) do { } while(0)
#define unregister_suspend_notifier(a) do { } while(0)
diff -ruN linux-2.5.65-02/kernel/suspend.c linux-2.5.65-03/kernel/suspend.c
--- linux-2.5.65-02/kernel/suspend.c 2003-03-18 17:53:44.000000000 +1200
+++ linux-2.5.65-03/kernel/suspend.c 2003-03-18 17:55:55.000000000 +1200
@@ -68,6 +68,7 @@
extern int sys_sync(void);

unsigned char software_suspend_enabled = 0;
+unsigned int suspend_task = 0;

#define SUSPEND_CONSOLE (MAX_NR_CONSOLES-1)
/* With SUSPEND_CONSOLE defined, it suspend looks *really* cool, but
@@ -232,6 +233,8 @@
}
} while(todo);

+ suspend_task = current->pid;
+
printk( "|\n" );
BUG_ON(in_atomic());
return 0;
@@ -253,6 +256,7 @@
} while_each_thread(g, p);

read_unlock(&tasklist_lock);
+ suspend_task = 0;
printk( " done\n" );
MDELAY(500);
}
diff -ruN linux-2.5.65-02/mm/page_alloc.c linux-2.5.65-03/mm/page_alloc.c
--- linux-2.5.65-02/mm/page_alloc.c 2003-03-18 17:53:44.000000000 +1200
+++ linux-2.5.65-03/mm/page_alloc.c 2003-03-18 17:55:55.000000000 +1200
@@ -486,6 +486,10 @@
* Really, prep_compound_page() should be called from __rmqueue_bulk(). But
* we cheat by calling it from here, in the order > 0 path. Saves a branch
* or two.
+ *
+ * While suspending, we don't use the pcp structure. It mucks up our
+ * accounting for all the pages and necessitates calling drain_local_pages
+ * multiple times.
*/

static struct page *buffered_rmqueue(struct zone *zone, int order, int cold)
@@ -493,7 +497,7 @@
unsigned long flags;
struct page *page = NULL;

- if (order == 0) {
+ if ((order == 0) && (!suspend_task)) {
struct per_cpu_pages *pcp;

pcp = &zone->pageset[get_cpu()].pcp[cold];
@@ -700,7 +704,7 @@
void __free_pages(struct page *page, unsigned int order)
{
if (!PageReserved(page) && put_page_testzero(page)) {
- if (order == 0)
+ if ((order == 0) && (!suspend_task))
free_hot_page(page);
else
__free_pages_ok(page, order);



--
Nigel Cunningham
495 St Georges Road South, Hastings 4201, New Zealand

Be diligent to present yourself approved to God as a workman who does
not need to be ashamed, handling accurately the word of truth.
-- 2 Timothy 2:14, NASB.

2003-03-18 08:07:15

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

Hi!

> Here's another patch (the last for a little while, I promise!). It stops
> the pcp lists from being refilled while SWSUSP is running. Despite the
> comment in the page, drain_local_pages does only need to get called once
> right now, but I have patches coming that will (DV) change that. This
> patch is thus groundwork for them.

This adds external (and pretty ugly) dependency of swsusp on the
outside. And as it still needs to drain_local_pages(), nothing is
gained. I believe it is better to just call drain_local_pages few
times. Magic hooks "if suspending, don't do this" seem like wrong
approach to me.

Pavel
--
Horseback riding is like software...
...vgf orggre jura vgf serr.

2003-03-18 09:57:46

by Nigel Cunningham

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

Hi Pavel.

You need to remember that this is infrastructure for later. I tried it
other ways and would have needed a number of calls to drain local pages.
Perhaps I'm taking the wrong approach, trying to feed a patch at a time
when you can't see how it all fits together. If you like, I'll send a
'whole kit and caboodle' patch as soon as I get the last bugs ironed
out. I have it suspending and resuming at the moment, but have one last
bug to iron out that's stopping me getting back from
do_suspend_lowlevel. It's just a matter of time, and then that will be
fixed. I could send you a monster patch then :> (I'll be posting it
somewhere anyway - I'll want others to test it, of course).

Regards,

Nigel

On Tue, 2003-03-18 at 20:18, Pavel Machek wrote:
> Hi!
>
> > Here's another patch (the last for a little while, I promise!). It stops
> > the pcp lists from being refilled while SWSUSP is running. Despite the
> > comment in the page, drain_local_pages does only need to get called once
> > right now, but I have patches coming that will (DV) change that. This
> > patch is thus groundwork for them.
>
> This adds external (and pretty ugly) dependency of swsusp on the
> outside. And as it still needs to drain_local_pages(), nothing is
> gained. I believe it is better to just call drain_local_pages few
> times. Magic hooks "if suspending, don't do this" seem like wrong
> approach to me.
>
> Pavel
--
Nigel Cunningham
495 St Georges Road South, Hastings 4201, New Zealand

Be diligent to present yourself approved to God as a workman who does
not need to be ashamed, handling accurately the word of truth.
-- 2 Timothy 2:14, NASB.

2003-03-18 16:47:35

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH] Don't refill pcp lists during SWSUSP.

Hi!

> You need to remember that this is infrastructure for later. I tried it
> other ways and would have needed a number of calls to drain local
> pages.

If you need 10x drain_local_pages(), that's okay -- because it does
not slow down non-suspend paths. If you need more of them... well then
I understand this is the way to go (but you should be able to hide
drain_local_pages in some local funcion or something...).
Pavel
--
Horseback riding is like software...
...vgf orggre jura vgf serr.