2006-05-18 03:31:03

by Shaohua Li

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

On Sun, 2006-04-30 at 14:04 +0200, Pavel Machek wrote:
> On So 29-04-06 23:57:21, Andrew Morton wrote:
> > Shaohua Li <[email protected]> wrote:
> > >
> > > On Sun, 2006-04-30 at 06:45 +0000, Pavel Machek wrote:
> > > > Hi!
> > > >
> > > > > flush_tlb_all uses on_each_cpu, which will disable/enable interrupt.
> > > > > In suspend/resume time, this will make interrupt wrongly enabled.
> > > >
> > > > > diff -puN arch/i386/mm/init.c~flush_tlb_all_check arch/i386/mm/init.c
> > > > > --- linux-2.6.17-rc3/arch/i386/mm/init.c~flush_tlb_all_check 2006-04-29 08:47:05.000000000 +0800
> > > > > +++ linux-2.6.17-rc3-root/arch/i386/mm/init.c 2006-04-29 08:48:15.000000000 +0800
> > > > > @@ -420,7 +420,10 @@ void zap_low_mappings (void)
> > > > > #else
> > > > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > > > #endif
> > > > > - flush_tlb_all();
> > > > > + if (cpus_weight(cpu_online_map) == 1)
> > > > > + local_flush_tlb();
> > > > > + else
> > > > > + flush_tlb_all();
> > > > > }
> > > > >
> > > >
> > > > Either it is okay to enable interrupts here -> unneccessary and ugly
> > > > test, or it is not, and then we are broken in SMP case.
> > > It's not broken in SMP case, APs are offlined here in suspend/resume.
> > >
> >
> > In which case, how's about this?
>
> Certainly better, I'd say.
>
> > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > #else
> > set_pgd(swapper_pg_dir+i, __pgd(0));
> > #endif
> > - if (cpus_weight(cpu_online_map) == 1)
> > + /*
> > + * We can be called at suspend/resume time, with local interrupts
> > + * disabled. But flush_tlb_all() requires that local interrupts be
> > + * enabled.
> > + *
> > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > + */
> > + if (num_online_cpus() == 1)
> > local_flush_tlb();
> > else
> > flush_tlb_all();
>
> But this still scares. It means calling convention is "may enable
> interrupts with >1 cpu, may not with == 1 cpu".
Below patch should make things clean. How do you think?

Signed-off-by: Shaohua Li <[email protected]>
---

linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/sleep.c | 10 +++-------
linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/wakeup.S | 2 +-
linux-2.6.17-rc4-mm1-root/arch/i386/mm/init.c | 13 +------------
3 files changed, 5 insertions(+), 20 deletions(-)

diff -puN arch/i386/kernel/acpi/wakeup.S~i386_direct_mapping arch/i386/kernel/acpi/wakeup.S
--- linux-2.6.17-rc4-mm1/arch/i386/kernel/acpi/wakeup.S~i386_direct_mapping 2006-05-16 14:52:37.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/wakeup.S 2006-05-16 15:10:57.000000000 +0800
@@ -56,7 +56,7 @@ wakeup_code:
1:

# set up page table
- movl $swapper_pg_dir-__PAGE_OFFSET, %eax
+ movl $sleep_pg_dir-__PAGE_OFFSET, %eax
movl %eax, %cr3

testl $1, real_efer_save_restore - wakeup_code
diff -puN arch/i386/kernel/acpi/sleep.c~i386_direct_mapping arch/i386/kernel/acpi/sleep.c
--- linux-2.6.17-rc4-mm1/arch/i386/kernel/acpi/sleep.c~i386_direct_mapping 2006-05-16 14:52:51.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/sleep.c 2006-05-17 09:30:26.000000000 +0800
@@ -11,16 +11,14 @@
#include <linux/cpumask.h>

#include <asm/smp.h>
-#include <asm/tlbflush.h>

/* address in low memory of the wakeup routine. */
unsigned long acpi_wakeup_address = 0;
unsigned long acpi_video_flags;
extern char wakeup_start, wakeup_end;

-extern void zap_low_mappings(void);
-
extern unsigned long FASTCALL(acpi_copy_wakeup_routine(unsigned long));
+char sleep_pg_dir[PAGE_SIZE];

static void init_low_mapping(pgd_t * pgd, int pgd_limit)
{
@@ -31,8 +29,6 @@ static void init_low_mapping(pgd_t * pgd
set_pgd(pgd, *(pgd + USER_PTRS_PER_PGD));
pgd_ofs++, pgd++;
}
- WARN_ON(num_online_cpus() != 1);
- local_flush_tlb();
}

/**
@@ -45,7 +41,8 @@ int acpi_save_state_mem(void)
{
if (!acpi_wakeup_address)
return 1;
- init_low_mapping(swapper_pg_dir, USER_PTRS_PER_PGD);
+ memcpy(sleep_pg_dir, swapper_pg_dir, PAGE_SIZE);
+ init_low_mapping((pgd_t *)sleep_pg_dir, USER_PTRS_PER_PGD);
memcpy((void *)acpi_wakeup_address, &wakeup_start,
&wakeup_end - &wakeup_start);
acpi_copy_wakeup_routine(acpi_wakeup_address);
@@ -58,7 +55,6 @@ int acpi_save_state_mem(void)
*/
void acpi_restore_state_mem(void)
{
- zap_low_mappings();
}

/**
diff -puN arch/i386/mm/init.c~i386_direct_mapping arch/i386/mm/init.c
--- linux-2.6.17-rc4-mm1/arch/i386/mm/init.c~i386_direct_mapping 2006-05-17 09:03:39.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/mm/init.c 2006-05-17 09:03:49.000000000 +0800
@@ -422,18 +422,7 @@ void zap_low_mappings (void)
#else
set_pgd(swapper_pg_dir+i, __pgd(0));
#endif
- /*
- * We can be called at suspend/resume time, with local interrupts
- * disabled. But flush_tlb_all() requires that local interrupts be
- * enabled.
- *
- * Happily, the APs are not yet started, so we can use local_flush_tlb()
- * in that case
- */
- if (num_online_cpus() == 1)
- local_flush_tlb();
- else
- flush_tlb_all();
+ flush_tlb_all();
}

static int disable_nx __initdata = 0;
_


2006-05-18 08:32:38

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

Hi!

> > > In which case, how's about this?
> >
> > Certainly better, I'd say.
> >
> > > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > > #else
> > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > #endif
> > > - if (cpus_weight(cpu_online_map) == 1)
> > > + /*
> > > + * We can be called at suspend/resume time, with local interrupts
> > > + * disabled. But flush_tlb_all() requires that local interrupts be
> > > + * enabled.
> > > + *
> > > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > > + */
> > > + if (num_online_cpus() == 1)
> > > local_flush_tlb();
> > > else
> > > flush_tlb_all();
> >
> > But this still scares. It means calling convention is "may enable
> > interrupts with >1 cpu, may not with == 1 cpu".
> Below patch should make things clean. How do you think?

Nice...

Could we perhaps reuse swsusp_pg_dir (just make it used for swsusp &
suspend-to-ram) to save a bit more code? It is in arch/i386/mm/init.c
.

Pavel

--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

2006-05-18 08:39:48

by Shaohua Li

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

On Thu, 2006-05-18 at 10:31 +0200, Pavel Machek wrote:
> Hi!
>
> > > > In which case, how's about this?
> > >
> > > Certainly better, I'd say.
> > >
> > > > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > > > #else
> > > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > > #endif
> > > > - if (cpus_weight(cpu_online_map) == 1)
> > > > + /*
> > > > + * We can be called at suspend/resume time, with local interrupts
> > > > + * disabled. But flush_tlb_all() requires that local interrupts be
> > > > + * enabled.
> > > > + *
> > > > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > > > + */
> > > > + if (num_online_cpus() == 1)
> > > > local_flush_tlb();
> > > > else
> > > > flush_tlb_all();
> > >
> > > But this still scares. It means calling convention is "may enable
> > > interrupts with >1 cpu, may not with == 1 cpu".
> > Below patch should make things clean. How do you think?
>
> Nice...
>
> Could we perhaps reuse swsusp_pg_dir (just make it used for swsusp &
> suspend-to-ram) to save a bit more code? It is in arch/i386/mm/init.c
Sure. But it's under CONFIG_SOFTWARE_SUSPEND. That part needs cleanup I
think and it's a little strange to me (why should we simply copy
swapper_pg_dir to swsusp_pg_dir, instead do it in zap_low_mappings?).

Thanks,
Shaohua

2006-05-18 08:47:27

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

On Čt 18-05-06 16:38:21, Shaohua Li wrote:
> On Thu, 2006-05-18 at 10:31 +0200, Pavel Machek wrote:
> > Hi!
> >
> > > > > In which case, how's about this?
> > > >
> > > > Certainly better, I'd say.
> > > >
> > > > > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > > > > #else
> > > > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > > > #endif
> > > > > - if (cpus_weight(cpu_online_map) == 1)
> > > > > + /*
> > > > > + * We can be called at suspend/resume time, with local interrupts
> > > > > + * disabled. But flush_tlb_all() requires that local interrupts be
> > > > > + * enabled.
> > > > > + *
> > > > > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > > > > + */
> > > > > + if (num_online_cpus() == 1)
> > > > > local_flush_tlb();
> > > > > else
> > > > > flush_tlb_all();
> > > >
> > > > But this still scares. It means calling convention is "may enable
> > > > interrupts with >1 cpu, may not with == 1 cpu".
> > > Below patch should make things clean. How do you think?
> >
> > Nice...
> >
> > Could we perhaps reuse swsusp_pg_dir (just make it used for swsusp &
> > suspend-to-ram) to save a bit more code? It is in arch/i386/mm/init.c
> Sure. But it's under CONFIG_SOFTWARE_SUSPEND.

Yes, #ifdef definitely does need to change there.

> That part needs cleanup I
> think and it's a little strange to me (why should we simply copy
> swapper_pg_dir to swsusp_pg_dir, instead do it in zap_low_mappings?).

?? Sorry, I do not understand. "instead of doing"?
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

2006-05-19 01:17:04

by Shaohua Li

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

On Thu, 2006-05-18 at 10:31 +0200, Pavel Machek wrote:
> Hi!
>
> > > > In which case, how's about this?
> > >
> > > Certainly better, I'd say.
> > >
> > > > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > > > #else
> > > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > > #endif
> > > > - if (cpus_weight(cpu_online_map) == 1)
> > > > + /*
> > > > + * We can be called at suspend/resume time, with local interrupts
> > > > + * disabled. But flush_tlb_all() requires that local interrupts be
> > > > + * enabled.
> > > > + *
> > > > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > > > + */
> > > > + if (num_online_cpus() == 1)
> > > > local_flush_tlb();
> > > > else
> > > > flush_tlb_all();
> > >
> > > But this still scares. It means calling convention is "may enable
> > > interrupts with >1 cpu, may not with == 1 cpu".
> > Below patch should make things clean. How do you think?
>
> Nice...
>
> Could we perhaps reuse swsusp_pg_dir (just make it used for swsusp &
> suspend-to-ram) to save a bit more code? It is in arch/i386/mm/init.c
Ok, I guess this is what you want.

Signed-off-by: Shaohua Li <[email protected]>
---

linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/sleep.c | 18 ---------------
linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/wakeup.S | 2 -
linux-2.6.17-rc4-mm1-root/arch/i386/mm/init.c | 15 +-----------
3 files changed, 3 insertions(+), 32 deletions(-)

diff -puN arch/i386/kernel/acpi/wakeup.S~i386_direct_mapping arch/i386/kernel/acpi/wakeup.S
--- linux-2.6.17-rc4-mm1/arch/i386/kernel/acpi/wakeup.S~i386_direct_mapping 2006-05-16 14:52:37.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/wakeup.S 2006-05-18 07:12:40.000000000 +0800
@@ -56,7 +56,7 @@ wakeup_code:
1:

# set up page table
- movl $swapper_pg_dir-__PAGE_OFFSET, %eax
+ movl $swsusp_pg_dir-__PAGE_OFFSET, %eax
movl %eax, %cr3

testl $1, real_efer_save_restore - wakeup_code
diff -puN arch/i386/kernel/acpi/sleep.c~i386_direct_mapping arch/i386/kernel/acpi/sleep.c
--- linux-2.6.17-rc4-mm1/arch/i386/kernel/acpi/sleep.c~i386_direct_mapping 2006-05-16 14:52:51.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/kernel/acpi/sleep.c 2006-05-18 07:11:38.000000000 +0800
@@ -11,30 +11,14 @@
#include <linux/cpumask.h>

#include <asm/smp.h>
-#include <asm/tlbflush.h>

/* address in low memory of the wakeup routine. */
unsigned long acpi_wakeup_address = 0;
unsigned long acpi_video_flags;
extern char wakeup_start, wakeup_end;

-extern void zap_low_mappings(void);
-
extern unsigned long FASTCALL(acpi_copy_wakeup_routine(unsigned long));

-static void init_low_mapping(pgd_t * pgd, int pgd_limit)
-{
- int pgd_ofs = 0;
-
- while ((pgd_ofs < pgd_limit)
- && (pgd_ofs + USER_PTRS_PER_PGD < PTRS_PER_PGD)) {
- set_pgd(pgd, *(pgd + USER_PTRS_PER_PGD));
- pgd_ofs++, pgd++;
- }
- WARN_ON(num_online_cpus() != 1);
- local_flush_tlb();
-}
-
/**
* acpi_save_state_mem - save kernel state
*
@@ -45,7 +29,6 @@ int acpi_save_state_mem(void)
{
if (!acpi_wakeup_address)
return 1;
- init_low_mapping(swapper_pg_dir, USER_PTRS_PER_PGD);
memcpy((void *)acpi_wakeup_address, &wakeup_start,
&wakeup_end - &wakeup_start);
acpi_copy_wakeup_routine(acpi_wakeup_address);
@@ -58,7 +41,6 @@ int acpi_save_state_mem(void)
*/
void acpi_restore_state_mem(void)
{
- zap_low_mappings();
}

/**
diff -puN arch/i386/mm/init.c~i386_direct_mapping arch/i386/mm/init.c
--- linux-2.6.17-rc4-mm1/arch/i386/mm/init.c~i386_direct_mapping 2006-05-17 09:03:39.000000000 +0800
+++ linux-2.6.17-rc4-mm1-root/arch/i386/mm/init.c 2006-05-18 07:11:10.000000000 +0800
@@ -386,7 +386,7 @@ static void __init pagetable_init (void)
#endif
}

-#ifdef CONFIG_SOFTWARE_SUSPEND
+#if defined(CONFIG_SOFTWARE_SUSPEND) || defined(CONFIG_ACPI_SLEEP)
/*
* Swap suspend & friends need this for resume because things like the intel-agp
* driver might have split up a kernel 4MB mapping.
@@ -422,18 +422,7 @@ void zap_low_mappings (void)
#else
set_pgd(swapper_pg_dir+i, __pgd(0));
#endif
- /*
- * We can be called at suspend/resume time, with local interrupts
- * disabled. But flush_tlb_all() requires that local interrupts be
- * enabled.
- *
- * Happily, the APs are not yet started, so we can use local_flush_tlb()
- * in that case
- */
- if (num_online_cpus() == 1)
- local_flush_tlb();
- else
- flush_tlb_all();
+ flush_tlb_all();
}

static int disable_nx __initdata = 0;
_

2006-05-20 15:50:50

by Pavel Machek

[permalink] [raw]
Subject: Re: [PATCH] don't use flush_tlb_all in suspend time

Hi!

> > > > > @@ -420,7 +421,14 @@ void zap_low_mappings (void)
> > > > > #else
> > > > > set_pgd(swapper_pg_dir+i, __pgd(0));
> > > > > #endif
> > > > > - if (cpus_weight(cpu_online_map) == 1)
> > > > > + /*
> > > > > + * We can be called at suspend/resume time, with local interrupts
> > > > > + * disabled. But flush_tlb_all() requires that local interrupts be
> > > > > + * enabled.
> > > > > + *
> > > > > + * Happily, the APs are not yet started, so we can use local_flush_tlb() * in that case
> > > > > + */
> > > > > + if (num_online_cpus() == 1)
> > > > > local_flush_tlb();
> > > > > else
> > > > > flush_tlb_all();
> > > >
> > > > But this still scares. It means calling convention is "may enable
> > > > interrupts with >1 cpu, may not with == 1 cpu".
> > > Below patch should make things clean. How do you think?
> >
> > Nice...
> >
> > Could we perhaps reuse swsusp_pg_dir (just make it used for swsusp &
> > suspend-to-ram) to save a bit more code? It is in arch/i386/mm/init.c
> Ok, I guess this is what you want.

Yes, thanks a lot for patience.

> Signed-off-by: Shaohua Li <[email protected]>

Acked-by: Pavel Machek <[email protected]>

--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html