2022-05-11 11:42:34

by Wonhyuk Yang

[permalink] [raw]
Subject: [PATCH] mm/page_alloc: Fix tracepoint mm_page_alloc_zone_locked()

Currently, trace point mm_page_alloc_zone_locked() doesn't show
correct information.

First, when alloc_flag has ALLOC_HARDER/ALLOC_CMA, page can
be allocated from MIGRATE_HIGHATOMIC/MIGRATE_CMA. Nevertheless,
tracepoint use requested migration type not MIGRATE_HIGHATOMIC and
MIGRATE_CMA.

Second, after Commit 44042b4498728 ("mm/page_alloc: allow high-order
pages to be stored on the per-cpu lists") percpu-list can store
high order pages. But trace point determine whether it is a refiil
of percpu-list by comparing requested order and 0.

To handle these problems, use cached migration type by
get_pcppage_migratetype() instead of requested migration type.
Then, make mm_page_alloc_zone_locked() be called only two contexts
(rmqueue_bulk, rmqueue). With a new argument called percpu_refill,
it can show whether it is a refill of percpu-list correctly.

Cc: Baik Song An <[email protected]>
Cc: Hong Yeon Kim <[email protected]>
Cc: Taeung Song <[email protected]>
Cc: [email protected]
Signed-off-by: Wonhyuk Yang <[email protected]>
---
include/trace/events/kmem.h | 14 +++++++++-----
mm/page_alloc.c | 18 +++++++++---------
2 files changed, 18 insertions(+), 14 deletions(-)

diff --git a/include/trace/events/kmem.h b/include/trace/events/kmem.h
index ddc8c944f417..f89fb3afcd46 100644
--- a/include/trace/events/kmem.h
+++ b/include/trace/events/kmem.h
@@ -229,20 +229,23 @@ TRACE_EVENT(mm_page_alloc,

DECLARE_EVENT_CLASS(mm_page,

- TP_PROTO(struct page *page, unsigned int order, int migratetype),
+ TP_PROTO(struct page *page, unsigned int order, int migratetype,
+ int percpu_refill),

- TP_ARGS(page, order, migratetype),
+ TP_ARGS(page, order, migratetype, percpu_refill),

TP_STRUCT__entry(
__field( unsigned long, pfn )
__field( unsigned int, order )
__field( int, migratetype )
+ __field( int, percpu_refill )
),

TP_fast_assign(
__entry->pfn = page ? page_to_pfn(page) : -1UL;
__entry->order = order;
__entry->migratetype = migratetype;
+ __entry->percpu_refill = percpu_refill;
),

TP_printk("page=%p pfn=0x%lx order=%u migratetype=%d percpu_refill=%d",
@@ -250,14 +253,15 @@ DECLARE_EVENT_CLASS(mm_page,
__entry->pfn != -1UL ? __entry->pfn : 0,
__entry->order,
__entry->migratetype,
- __entry->order == 0)
+ __entry->percpu_refill)
);

DEFINE_EVENT(mm_page, mm_page_alloc_zone_locked,

- TP_PROTO(struct page *page, unsigned int order, int migratetype),
+ TP_PROTO(struct page *page, unsigned int order, int migratetype,
+ int percpu_refill),

- TP_ARGS(page, order, migratetype)
+ TP_ARGS(page, order, migratetype, percpu_refill)
);

TRACE_EVENT(mm_page_pcpu_drain,
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0e42038382c1..0351808322ba 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2999,7 +2999,7 @@ __rmqueue(struct zone *zone, unsigned int order, int migratetype,
zone_page_state(zone, NR_FREE_PAGES) / 2) {
page = __rmqueue_cma_fallback(zone, order);
if (page)
- goto out;
+ return page;
}
}
retry:
@@ -3012,9 +3012,6 @@ __rmqueue(struct zone *zone, unsigned int order, int migratetype,
alloc_flags))
goto retry;
}
-out:
- if (page)
- trace_mm_page_alloc_zone_locked(page, order, migratetype);
return page;
}

@@ -3028,6 +3025,7 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
int migratetype, unsigned int alloc_flags)
{
int i, allocated = 0;
+ int mt;

/*
* local_lock_irq held so equivalent to spin_lock_irqsave for
@@ -3055,7 +3053,9 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
*/
list_add_tail(&page->lru, list);
allocated++;
- if (is_migrate_cma(get_pcppage_migratetype(page)))
+ mt = get_pcppage_migratetype(page);
+ trace_mm_page_alloc_zone_locked(page, order, mt, true);
+ if (is_migrate_cma(mt))
__mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
-(1 << order));
}
@@ -3704,6 +3704,7 @@ struct page *rmqueue(struct zone *preferred_zone,
{
unsigned long flags;
struct page *page;
+ int mt;

if (likely(pcp_allowed_order(order))) {
/*
@@ -3735,16 +3736,15 @@ struct page *rmqueue(struct zone *preferred_zone,
*/
if (order > 0 && alloc_flags & ALLOC_HARDER) {
page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
- if (page)
- trace_mm_page_alloc_zone_locked(page, order, migratetype);
}
if (!page) {
page = __rmqueue(zone, order, migratetype, alloc_flags);
if (!page)
goto failed;
}
- __mod_zone_freepage_state(zone, -(1 << order),
- get_pcppage_migratetype(page));
+ mt = get_pcppage_migratetype(page);
+ trace_mm_page_alloc_zone_locked(page, order, mt, false);
+ __mod_zone_freepage_state(zone, -(1 << order), mt);
spin_unlock_irqrestore(&zone->lock, flags);
} while (check_new_pages(page, order));

--
2.30.2



2022-05-11 20:01:22

by Mel Gorman

[permalink] [raw]
Subject: Re: [PATCH] mm/page_alloc: Fix tracepoint mm_page_alloc_zone_locked()

On Wed, May 11, 2022 at 05:12:07PM +0900, Wonhyuk Yang wrote:
> Currently, trace point mm_page_alloc_zone_locked() doesn't show
> correct information.
>
> First, when alloc_flag has ALLOC_HARDER/ALLOC_CMA, page can
> be allocated from MIGRATE_HIGHATOMIC/MIGRATE_CMA. Nevertheless,
> tracepoint use requested migration type not MIGRATE_HIGHATOMIC and
> MIGRATE_CMA.
>
> Second, after Commit 44042b4498728 ("mm/page_alloc: allow high-order
> pages to be stored on the per-cpu lists") percpu-list can store
> high order pages. But trace point determine whether it is a refiil
> of percpu-list by comparing requested order and 0.
>
> To handle these problems, use cached migration type by
> get_pcppage_migratetype() instead of requested migration type.
> Then, make mm_page_alloc_zone_locked() be called only two contexts
> (rmqueue_bulk, rmqueue). With a new argument called percpu_refill,
> it can show whether it is a refill of percpu-list correctly.
>

You're definitely right that the current tracepoint is broken.

I got momentarily confused because HIGHATOMIC and CMA are not stored on
PCP lists even though they are a pageblock migrate type. Superficially
calling get_pcppage_migratetype on a page that cannot be a PCP page
seems silly but in the context of this patch, it happens to work because
it was isolated with __rmqueue_smallest which sets the PCP type even if
the page is not going to a PCP list.

The original intent of that tracepoint was to trace when pages were
removed from the buddy list. That would suggest this untested patch on
top of yours as a simplication;

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0351808322ba..66a70b898130 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2476,6 +2476,8 @@ struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
del_page_from_free_list(page, zone, current_order);
expand(zone, page, order, current_order, migratetype);
set_pcppage_migratetype(page, migratetype);
+ trace_mm_page_alloc_zone_locked(page, order, migratetype,
+ pcp_allowed_order(order) && migratetype < MIGRATE_PCPTYPES);
return page;
}

@@ -3025,7 +3027,6 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
int migratetype, unsigned int alloc_flags)
{
int i, allocated = 0;
- int mt;

/*
* local_lock_irq held so equivalent to spin_lock_irqsave for
@@ -3053,9 +3054,7 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
*/
list_add_tail(&page->lru, list);
allocated++;
- mt = get_pcppage_migratetype(page);
- trace_mm_page_alloc_zone_locked(page, order, mt, true);
- if (is_migrate_cma(mt))
+ if (is_migrate_cma(get_pcppage_migratetype(page)))
__mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
-(1 << order));
}
@@ -3704,7 +3703,6 @@ struct page *rmqueue(struct zone *preferred_zone,
{
unsigned long flags;
struct page *page;
- int mt;

if (likely(pcp_allowed_order(order))) {
/*
@@ -3734,17 +3732,15 @@ struct page *rmqueue(struct zone *preferred_zone,
* reserved for high-order atomic allocation, so order-0
* request should skip it.
*/
- if (order > 0 && alloc_flags & ALLOC_HARDER) {
+ if (order > 0 && alloc_flags & ALLOC_HARDER)
page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
- }
if (!page) {
page = __rmqueue(zone, order, migratetype, alloc_flags);
if (!page)
goto failed;
}
- mt = get_pcppage_migratetype(page);
- trace_mm_page_alloc_zone_locked(page, order, mt, false);
- __mod_zone_freepage_state(zone, -(1 << order), mt);
+ __mod_zone_freepage_state(zone, -(1 << order),
+ get_pcppage_migratetype(page));
spin_unlock_irqrestore(&zone->lock, flags);
} while (check_new_pages(page, order));


2022-05-12 05:03:21

by Wonhyuk Yang

[permalink] [raw]
Subject: Re: [PATCH] mm/page_alloc: Fix tracepoint mm_page_alloc_zone_locked()

On Wed, May 11, 2022 at 11:23 PM Mel Gorman <[email protected]> wrote:
>
> On Wed, May 11, 2022 at 05:12:07PM +0900, Wonhyuk Yang wrote:
> > Currently, trace point mm_page_alloc_zone_locked() doesn't show
> > correct information.
> >
> > First, when alloc_flag has ALLOC_HARDER/ALLOC_CMA, page can
> > be allocated from MIGRATE_HIGHATOMIC/MIGRATE_CMA. Nevertheless,
> > tracepoint use requested migration type not MIGRATE_HIGHATOMIC and
> > MIGRATE_CMA.
> >
> > Second, after Commit 44042b4498728 ("mm/page_alloc: allow high-order
> > pages to be stored on the per-cpu lists") percpu-list can store
> > high order pages. But trace point determine whether it is a refiil
> > of percpu-list by comparing requested order and 0.
> >
> > To handle these problems, use cached migration type by
> > get_pcppage_migratetype() instead of requested migration type.
> > Then, make mm_page_alloc_zone_locked() be called only two contexts
> > (rmqueue_bulk, rmqueue). With a new argument called percpu_refill,
> > it can show whether it is a refill of percpu-list correctly.
> >
>
> You're definitely right that the current tracepoint is broken.
>
> I got momentarily confused because HIGHATOMIC and CMA are not stored on
> PCP lists even though they are a pageblock migrate type. Superficially
> calling get_pcppage_migratetype on a page that cannot be a PCP page
> seems silly but in the context of this patch, it happens to work because
> it was isolated with __rmqueue_smallest which sets the PCP type even if
> the page is not going to a PCP list.

Yes, I agree that calling get_pcppage_migratetype look quite confusing.

> The original intent of that tracepoint was to trace when pages were
> removed from the buddy list. That would suggest this untested patch on
> top of yours as a simplication;
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 0351808322ba..66a70b898130 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2476,6 +2476,8 @@ struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
> del_page_from_free_list(page, zone, current_order);
> expand(zone, page, order, current_order, migratetype);
> set_pcppage_migratetype(page, migratetype);
> + trace_mm_page_alloc_zone_locked(page, order, migratetype,
> + pcp_allowed_order(order) && migratetype < MIGRATE_PCPTYPES);
> return page;
> }

Interestingly, my first approach was quite similar your suggestion. But I
noticed that there can be a request whose migration type is MOVABLE
and alloc_flags doen't have ALLOC_CMA. In that case, page are marked
as percpu-refill even though it was allocated from buddy-list directly.
Is there no problem if we just ignore this case?


> @@ -3025,7 +3027,6 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
> int migratetype, unsigned int alloc_flags)
> {
> int i, allocated = 0;
> - int mt;
>
> /*
> * local_lock_irq held so equivalent to spin_lock_irqsave for
> @@ -3053,9 +3054,7 @@ static int rmqueue_bulk(struct zone *zone, unsigned int order,
> */
> list_add_tail(&page->lru, list);
> allocated++;
> - mt = get_pcppage_migratetype(page);
> - trace_mm_page_alloc_zone_locked(page, order, mt, true);
> - if (is_migrate_cma(mt))
> + if (is_migrate_cma(get_pcppage_migratetype(page)))
> __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
> -(1 << order));
> }
> @@ -3704,7 +3703,6 @@ struct page *rmqueue(struct zone *preferred_zone,
> {
> unsigned long flags;
> struct page *page;
> - int mt;
>
> if (likely(pcp_allowed_order(order))) {
> /*
> @@ -3734,17 +3732,15 @@ struct page *rmqueue(struct zone *preferred_zone,
> * reserved for high-order atomic allocation, so order-0
> * request should skip it.
> */
> - if (order > 0 && alloc_flags & ALLOC_HARDER) {
> + if (order > 0 && alloc_flags & ALLOC_HARDER)
> page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
> - }
> if (!page) {
> page = __rmqueue(zone, order, migratetype, alloc_flags);
> if (!page)
> goto failed;
> }
> - mt = get_pcppage_migratetype(page);
> - trace_mm_page_alloc_zone_locked(page, order, mt, false);
> - __mod_zone_freepage_state(zone, -(1 << order), mt);
> + __mod_zone_freepage_state(zone, -(1 << order),
> + get_pcppage_migratetype(page));
> spin_unlock_irqrestore(&zone->lock, flags);
> } while (check_new_pages(page, order));
>

2022-05-12 21:34:23

by Wonhyuk Yang

[permalink] [raw]
Subject: Re: [PATCH] mm/page_alloc: Fix tracepoint mm_page_alloc_zone_locked()

On Thu, May 12, 2022 at 12:47 AM Mel Gorman <[email protected]> wrote:
>
> On Thu, May 12, 2022 at 12:02:30AM +0900, Wonhyuk Yang wrote:
> > > The original intent of that tracepoint was to trace when pages were
> > > removed from the buddy list. That would suggest this untested patch on
> > > top of yours as a simplication;
> > >
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 0351808322ba..66a70b898130 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -2476,6 +2476,8 @@ struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
> > > del_page_from_free_list(page, zone, current_order);
> > > expand(zone, page, order, current_order, migratetype);
> > > set_pcppage_migratetype(page, migratetype);
> > > + trace_mm_page_alloc_zone_locked(page, order, migratetype,
> > > + pcp_allowed_order(order) && migratetype < MIGRATE_PCPTYPES);
> > > return page;
> > > }
> >
> > Interestingly, my first approach was quite similar your suggestion. But I
> > noticed that there can be a request whose migration type is MOVABLE
> > and alloc_flags doen't have ALLOC_CMA. In that case, page are marked
> > as percpu-refill even though it was allocated from buddy-list directly.
> > Is there no problem if we just ignore this case?
> >
>
> I assume you are referring to the case where CMA allocations are being
> balanced between regular and CMA areas. I think it's relatively harmless
> if percpu_refill field is not 100% accurate for that case. There are
> also cases like the percpu list is too small to hold a THP and it's not a
> percpu_refill either. If 100% accuracy is an issue, I would prefer renaming
> it to percpu_eligible or just deleting it instead of adding complexity
> for a tracepoint. The main value of that tracepoint is determining what
> percentage of allocations are potentially contending on zone lock at a
> particular time.
>

Okay, I'll send a new one with your suggestions. Thanks

2022-05-14 03:35:21

by Mel Gorman

[permalink] [raw]
Subject: Re: [PATCH] mm/page_alloc: Fix tracepoint mm_page_alloc_zone_locked()

On Thu, May 12, 2022 at 12:02:30AM +0900, Wonhyuk Yang wrote:
> > The original intent of that tracepoint was to trace when pages were
> > removed from the buddy list. That would suggest this untested patch on
> > top of yours as a simplication;
> >
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 0351808322ba..66a70b898130 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -2476,6 +2476,8 @@ struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
> > del_page_from_free_list(page, zone, current_order);
> > expand(zone, page, order, current_order, migratetype);
> > set_pcppage_migratetype(page, migratetype);
> > + trace_mm_page_alloc_zone_locked(page, order, migratetype,
> > + pcp_allowed_order(order) && migratetype < MIGRATE_PCPTYPES);
> > return page;
> > }
>
> Interestingly, my first approach was quite similar your suggestion. But I
> noticed that there can be a request whose migration type is MOVABLE
> and alloc_flags doen't have ALLOC_CMA. In that case, page are marked
> as percpu-refill even though it was allocated from buddy-list directly.
> Is there no problem if we just ignore this case?
>

I assume you are referring to the case where CMA allocations are being
balanced between regular and CMA areas. I think it's relatively harmless
if percpu_refill field is not 100% accurate for that case. There are
also cases like the percpu list is too small to hold a THP and it's not a
percpu_refill either. If 100% accuracy is an issue, I would prefer renaming
it to percpu_eligible or just deleting it instead of adding complexity
for a tracepoint. The main value of that tracepoint is determining what
percentage of allocations are potentially contending on zone lock at a
particular time.

--
Mel Gorman
SUSE Labs