2022-09-09 09:46:05

by Miaohe Lin

Subject: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
kswapd doesn't accidentally go to sleep. But when reserve_flags is
set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
go to sleep accidentally.

Fixes: 0a79cdad5eb2 ("mm: use alloc_flags to record if kswapd can wake")
Signed-off-by: Miaohe Lin <[email protected]>
---
mm/page_alloc.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index ba5c9402a5cb..4b97a03fa2dd 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5147,7 +5147,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,

reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
if (reserve_flags)
- alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
+ alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
+ (alloc_flags & ALLOC_KSWAPD);

/*
* Reset the nodemask and zonelist iterators if memory policies can be
--
2.23.0

2022-09-09 12:09:46

by David Hildenbrand

Subject: Re: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

On 09.09.22 11:24, Miaohe Lin wrote:
> If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
> kswapd doesn't accidentally go to sleep. But when reserve_flags is
> set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
> Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
> go to sleep accidentally.
>
> Fixes: 0a79cdad5eb2 ("mm: use alloc_flags to record if kswapd can wake")
> Signed-off-by: Miaohe Lin <[email protected]>
> ---
> mm/page_alloc.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index ba5c9402a5cb..4b97a03fa2dd 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5147,7 +5147,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
>
> reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
> if (reserve_flags)
> - alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
> + alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
> + (alloc_flags & ALLOC_KSWAPD);
>
> /*
> * Reset the nodemask and zonelist iterators if memory policies can be

Acked-by: David Hildenbrand <[email protected]>

--
Thanks,

David / dhildenb

2022-09-13 07:16:41

by Anshuman Khandual

Subject: Re: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

On 9/9/22 14:54, Miaohe Lin wrote:
> If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
> kswapd doesn't accidentally go to sleep. But when reserve_flags is
> set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
> Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
> go to sleep accidentally.

Currently wake_all_kswapds() gets skipped subsequently if ALLOC_KSWAPD
is lost, but this only happens when the 'retry:' loops is taken ?

>
> Fixes: 0a79cdad5eb2 ("mm: use alloc_flags to record if kswapd can wake")
> Signed-off-by: Miaohe Lin <[email protected]>
> ---
> mm/page_alloc.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index ba5c9402a5cb..4b97a03fa2dd 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5147,7 +5147,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
>
> reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
> if (reserve_flags)
> - alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
> + alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
> + (alloc_flags & ALLOC_KSWAPD);
>
> /*
> * Reset the nodemask and zonelist iterators if memory policies can be

2022-09-13 08:16:00

by Matthew Wilcox

Subject: Re: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

On Tue, Sep 13, 2022 at 12:32:50PM +0530, Anshuman Khandual wrote:
>
>
> On 9/9/22 14:54, Miaohe Lin wrote:
> > If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
> > kswapd doesn't accidentally go to sleep. But when reserve_flags is
> > set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
> > Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
> > go to sleep accidentally.
>
> Currently wake_all_kswapds() gets skipped subsequently if ALLOC_KSWAPD
> is lost, but this only happens when the 'retry:' loops is taken ?

Right, but see the comment:

/* Ensure kswapd doesn't accidentally go to sleep as long as we loop */

and that is not currently true. I think that was an inadvertent change.
Particularly since the changelog for 0a79cdad5eb2 says "No functional
change".

> >
> > Fixes: 0a79cdad5eb2 ("mm: use alloc_flags to record if kswapd can wake")
> > Signed-off-by: Miaohe Lin <[email protected]>
> > ---
> > mm/page_alloc.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index ba5c9402a5cb..4b97a03fa2dd 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -5147,7 +5147,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
> >
> > reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
> > if (reserve_flags)
> > - alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
> > + alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
> > + (alloc_flags & ALLOC_KSWAPD);
> >
> > /*
> > * Reset the nodemask and zonelist iterators if memory policies can be
>

2022-09-13 12:45:43

by Anshuman Khandual

Subject: Re: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

On 9/13/22 13:28, Matthew Wilcox wrote:
> On Tue, Sep 13, 2022 at 12:32:50PM +0530, Anshuman Khandual wrote:
>>
>> On 9/9/22 14:54, Miaohe Lin wrote:
>>> If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
>>> kswapd doesn't accidentally go to sleep. But when reserve_flags is
>>> set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
>>> Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
>>> go to sleep accidentally.
>> Currently wake_all_kswapds() gets skipped subsequently if ALLOC_KSWAPD
>> is lost, but this only happens when the 'retry:' loops is taken ?
> Right, but see the comment:
>
> /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
>
> and that is not currently true. I think that was an inadvertent change.
> Particularly since the changelog for 0a79cdad5eb2 says "No functional
> change".

Got it, thanks for the explanation.

2022-09-15 04:45:30

by Oscar Salvador

Subject: Re: [PATCH 01/16] mm/page_alloc: ensure kswapd doesn't accidentally go to sleep

On Fri, Sep 09, 2022 at 05:24:36PM +0800, Miaohe Lin wrote:
> If ALLOC_KSWAPD is set, wake_all_kswapds() will be called to ensure
> kswapd doesn't accidentally go to sleep. But when reserve_flags is
> set, alloc_flags will be overwritten and ALLOC_KSWAPD is thus lost.
> Preserve the ALLOC_KSWAPD flag in alloc_flags to ensure kswapd won't
> go to sleep accidentally.
>
> Fixes: 0a79cdad5eb2 ("mm: use alloc_flags to record if kswapd can wake")
> Signed-off-by: Miaohe Lin <[email protected]>

Reviewed-by: Oscar Salvador <[email protected]>

> ---
> mm/page_alloc.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index ba5c9402a5cb..4b97a03fa2dd 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5147,7 +5147,8 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
>
> reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
> if (reserve_flags)
> - alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
> + alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
> + (alloc_flags & ALLOC_KSWAPD);
>
> /*
> * Reset the nodemask and zonelist iterators if memory policies can be
> --
> 2.23.0
>

--
Oscar Salvador
SUSE Labs