On Mon, 14 May 2007 10:37:18 +0800 Fengguang Wu <[email protected]> wrote:
> PG_lazyfree and PG_booked shares the same bit.
>
> Either it is a bug that shall fixed by the following patch, or
> the situation should be explicitly documented?
>
> Signed-off-by: Fengguang Wu <[email protected]>
> ---
> include/linux/page-flags.h | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> --- linux-2.6.21-mm2.orig/include/linux/page-flags.h
> +++ linux-2.6.21-mm2/include/linux/page-flags.h
> @@ -91,7 +91,7 @@
> #define PG_buddy 19 /* Page is free, on buddy lists */
> #define PG_booked 20 /* Has blocks reserved on-disk */
>
> -#define PG_lazyfree 20 /* MADV_FREE potential throwaway */
> +#define PG_lazyfree 21 /* MADV_FREE potential throwaway */
>
> /* PG_owner_priv_1 users should have descriptive aliases */
> #define PG_checked PG_owner_priv_1 /* Used by some filesystems */
That's an accident: PG_lazyfree got added but the out-of-tree ext4 patches
didn't get updated.
otoh, the intersection between pages which are PageBooked() and pages which
are PageLazyFree() should be zreo, so it'd be good to actually formalise
this reuse within the ext4 patches.
otoh2, PageLazyFree() could have reused PG_owner_priv_1.
Rik, Ted: any thoughts? We do need to scrimp on page flags: when we
finally run out, we're screwed.
On Sun, May 13, 2007 at 10:46:30PM -0700, Andrew Morton wrote:
> On Mon, 14 May 2007 10:37:18 +0800 Fengguang Wu <[email protected]> wrote:
>
> > PG_lazyfree and PG_booked shares the same bit.
> >
> > Either it is a bug that shall fixed by the following patch, or
> > the situation should be explicitly documented?
> >
> > Signed-off-by: Fengguang Wu <[email protected]>
> > ---
> > include/linux/page-flags.h | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > --- linux-2.6.21-mm2.orig/include/linux/page-flags.h
> > +++ linux-2.6.21-mm2/include/linux/page-flags.h
> > @@ -91,7 +91,7 @@
> > #define PG_buddy 19 /* Page is free, on buddy lists */
> > #define PG_booked 20 /* Has blocks reserved on-disk */
> >
> > -#define PG_lazyfree 20 /* MADV_FREE potential throwaway */
> > +#define PG_lazyfree 21 /* MADV_FREE potential throwaway */
> >
> > /* PG_owner_priv_1 users should have descriptive aliases */
> > #define PG_checked PG_owner_priv_1 /* Used by some filesystems */
>
> That's an accident: PG_lazyfree got added but the out-of-tree ext4 patches
> didn't get updated.
>
> otoh, the intersection between pages which are PageBooked() and pages which
> are PageLazyFree() should be zreo, so it'd be good to actually formalise
> this reuse within the ext4 patches.
>
> otoh2, PageLazyFree() could have reused PG_owner_priv_1.
otoh3: PG_lazyfree and PG_readahead can reuse the same bit, too.
PG_lazyfree applies to anonymous pages, while PG_readahead applies to
file backed pages.
On Sun, May 13, 2007 at 10:46:30PM -0700, Andrew Morton wrote:
> On Mon, 14 May 2007 10:37:18 +0800 Fengguang Wu <[email protected]> wrote:
>
> > PG_lazyfree and PG_booked shares the same bit.
> >
> > Either it is a bug that shall fixed by the following patch, or
> > the situation should be explicitly documented?
> >
> > Signed-off-by: Fengguang Wu <[email protected]>
> > ---
> > include/linux/page-flags.h | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > --- linux-2.6.21-mm2.orig/include/linux/page-flags.h
> > +++ linux-2.6.21-mm2/include/linux/page-flags.h
> > @@ -91,7 +91,7 @@
> > #define PG_buddy 19 /* Page is free, on buddy lists */
> > #define PG_booked 20 /* Has blocks reserved on-disk */
> >
> > -#define PG_lazyfree 20 /* MADV_FREE potential throwaway */
> > +#define PG_lazyfree 21 /* MADV_FREE potential throwaway */
> >
> > /* PG_owner_priv_1 users should have descriptive aliases */
> > #define PG_checked PG_owner_priv_1 /* Used by some filesystems */
>
> That's an accident: PG_lazyfree got added but the out-of-tree ext4 patches
> didn't get updated.
>
> otoh, the intersection between pages which are PageBooked() and pages which
> are PageLazyFree() should be zreo, so it'd be good to actually formalise
> this reuse within the ext4 patches.
>
> otoh2, PageLazyFree() could have reused PG_owner_priv_1.
otoh3: PG_lazyfree and PG_readahead can reuse the same bit, too.
PG_lazyfree applies to anonymous pages, while PG_readahead applies to
file backed pages.
On Sun, May 13, 2007 at 10:46:30PM -0700, Andrew Morton wrote:
> otoh, the intersection between pages which are PageBooked() and pages which
> are PageLazyFree() should be zreo, so it'd be good to actually formalise
> this reuse within the ext4 patches.
>
> otoh2, PageLazyFree() could have reused PG_owner_priv_1.
>
> Rik, Ted: any thoughts? We do need to scrimp on page flags: when we
> finally run out, we're screwed.
It makes sense to me. PG_lazyfree is currently only in -mm, right? I
don't see it in my git tree. It would probably would be a good idea
to make sure that we check to add some sanity checking code if it
isn't there already that PG_lazyfree isn't already set when try to set
PG_lazyfree (just in case there is a bug in the future which causes
the should-never-happen case of trying lazy free a PageBooked page).
- Ted
On Mon, 14 May 2007 14:06:19 -0400
Theodore Tso <[email protected]> wrote:
> On Sun, May 13, 2007 at 10:46:30PM -0700, Andrew Morton wrote:
> > otoh, the intersection between pages which are PageBooked() and pages which
> > are PageLazyFree() should be zreo, so it'd be good to actually formalise
> > this reuse within the ext4 patches.
> >
> > otoh2, PageLazyFree() could have reused PG_owner_priv_1.
> >
> > Rik, Ted: any thoughts? We do need to scrimp on page flags: when we
> > finally run out, we're screwed.
>
> It makes sense to me. PG_lazyfree is currently only in -mm, right?
Ah, yes, I got confused, sorry.
> I
> don't see it in my git tree. It would probably would be a good idea
> to make sure that we check to add some sanity checking code if it
> isn't there already that PG_lazyfree isn't already set when try to set
> PG_lazyfree (just in case there is a bug in the future which causes
> the should-never-happen case of trying lazy free a PageBooked page).
>
Actually, I think the current status of
lazy-freeing-of-memory-through-madv_free.patch is "might not be needed". I
_think_ we've determined that 0a27a14a62921b438bb6f33772690d345a089be6
sufficiently fixed the perfomance problems we had in there?
Andrew Morton wrote:
> On Mon, 14 May 2007 14:06:19 -0400
> Theodore Tso <[email protected]> wrote:
>
>
>>On Sun, May 13, 2007 at 10:46:30PM -0700, Andrew Morton wrote:
>>
>>>otoh, the intersection between pages which are PageBooked() and pages which
>>>are PageLazyFree() should be zreo, so it'd be good to actually formalise
>>>this reuse within the ext4 patches.
>>>
>>>otoh2, PageLazyFree() could have reused PG_owner_priv_1.
>>>
>>>Rik, Ted: any thoughts? We do need to scrimp on page flags: when we
>>>finally run out, we're screwed.
>>
>>It makes sense to me. PG_lazyfree is currently only in -mm, right?
>
>
> Ah, yes, I got confused, sorry.
>
>
>> I
>>don't see it in my git tree. It would probably would be a good idea
>>to make sure that we check to add some sanity checking code if it
>>isn't there already that PG_lazyfree isn't already set when try to set
>>PG_lazyfree (just in case there is a bug in the future which causes
>>the should-never-happen case of trying lazy free a PageBooked page).
>>
>
>
> Actually, I think the current status of
> lazy-freeing-of-memory-through-madv_free.patch is "might not be needed". I
> _think_ we've determined that 0a27a14a62921b438bb6f33772690d345a089be6
> sufficiently fixed the perfomance problems we had in there?
I think so far we've found that it fixes the MySQL scalability problem,
yes. I couldn't see any statistically significant difference with MySQL
in my tests with MADV_FREE (versus MADV_DONTNEED).
ebizzy is improved a bit at low concurrency but drops off slightly at
higher concurrency.
But basically, I don't think we've found a good reason to use a page
flag and introduce the potential performance regressions that the
MADV_FREE patch has.
--
SUSE Labs, Novell Inc.