From: Zhaoyang Huang <[email protected]>
Let us look at the timeline of scenarios below with WMARK_LOW=25MB WMARK_MIN=5MB
(managed pages 1.9GB). We can find that CMA begin to be used until 'C' under the
method of 'fixed 2 times of free cma over free pages' which could have the
scenario 'A' and 'B' into a fault state, that is, free UNMOVABLE & RECLAIMABLE
pages is lower than corresponding watermark without reclaiming which should be
deemed as against current memory policy. This commit try to solve this by
checking zone_watermark_ok again with removing CMA pages which could lead to a
proper time point of CMA's utilization.
-- Free_pages
|
|
-- WMARK_LOW
|
-- Free_CMA
|
|
--
Free_CMA/Free_pages(MB) A(12/30) --> B(12/25) --> C(12/20)
fixed 1/2 ratio N N Y
this commit Y Y Y
Signed-off-by: Zhaoyang Huang <[email protected]>
---
v2: do proportion check when zone_watermark_ok, update commit message
v3: update coding style and simplify the logic when zone_watermark_ok
v4: code update according to Roman's suggest
v5: update commit message
---
---
mm/page_alloc.c | 44 ++++++++++++++++++++++++++++++++++++++++----
1 file changed, 40 insertions(+), 4 deletions(-)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0745aed..4719800 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3071,6 +3071,43 @@ static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
}
+#ifdef CONFIG_CMA
+/*
+ * GFP_MOVABLE allocation could drain UNMOVABLE & RECLAIMABLE page blocks via
+ * the help of CMA which makes GFP_KERNEL failed. Checking if zone_watermark_ok
+ * again without ALLOC_CMA to see if to use CMA first.
+ */
+static bool use_cma_first(struct zone *zone, unsigned int order, unsigned int alloc_flags)
+{
+ unsigned long watermark;
+ bool cma_first = false;
+
+ watermark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
+ /* check if GFP_MOVABLE pass previous zone_watermark_ok via the help of CMA */
+ if (zone_watermark_ok(zone, order, watermark, 0, alloc_flags & (~ALLOC_CMA))) {
+ /*
+ * Balance movable allocations between regular and CMA areas by
+ * allocating from CMA when over half of the zone's free memory
+ * is in the CMA area.
+ */
+ cma_first = (zone_page_state(zone, NR_FREE_CMA_PAGES) >
+ zone_page_state(zone, NR_FREE_PAGES) / 2);
+ } else {
+ /*
+ * watermark failed means UNMOVABLE & RECLAIMBLE is not enough
+ * now, we should use cma first to keep them stay around the
+ * corresponding watermark
+ */
+ cma_first = true;
+ }
+ return cma_first;
+}
+#else
+static bool use_cma_first(struct zone *zone, unsigned int order, unsigned int alloc_flags)
+{
+ return false;
+}
+#endif
/*
* Do the hard work of removing an element from the buddy allocator.
* Call me with the zone->lock already held.
@@ -3084,12 +3121,11 @@ static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
if (IS_ENABLED(CONFIG_CMA)) {
/*
* Balance movable allocations between regular and CMA areas by
- * allocating from CMA when over half of the zone's free memory
- * is in the CMA area.
+ * allocating from CMA base on judging zone_watermark_ok again
+ * to see if the latest check got pass via the help of CMA
*/
if (alloc_flags & ALLOC_CMA &&
- zone_page_state(zone, NR_FREE_CMA_PAGES) >
- zone_page_state(zone, NR_FREE_PAGES) / 2) {
+ use_cma_first(zone, order, alloc_flags)) {
page = __rmqueue_cma_fallback(zone, order);
if (page)
return page;
--
1.9.1
On Thu, 11 May 2023 13:22:30 +0800 "zhaoyang.huang" <[email protected]> wrote:
> From: Zhaoyang Huang <[email protected]>
>
> Let us look at the timeline of scenarios below with WMARK_LOW=25MB WMARK_MIN=5MB
> (managed pages 1.9GB). We can find that CMA begin to be used until 'C' under the
> method of 'fixed 2 times of free cma over free pages' which could have the
> scenario 'A' and 'B' into a fault state, that is, free UNMOVABLE & RECLAIMABLE
> pages is lower than corresponding watermark without reclaiming which should be
> deemed as against current memory policy. This commit try to solve this by
> checking zone_watermark_ok again with removing CMA pages which could lead to a
> proper time point of CMA's utilization.
>
> -- Free_pages
> |
> |
> -- WMARK_LOW
> |
> -- Free_CMA
> |
> |
> --
>
> Free_CMA/Free_pages(MB) A(12/30) --> B(12/25) --> C(12/20)
> fixed 1/2 ratio N N Y
> this commit Y Y Y
>
Roman previously asked
: Also I'm a bit concerned about potential performance implications.
: Would be great to provide some benchmarks or some data. Probably it's
: ok because of we have pcp caches on top, but I'm not 100% sure.
Are you able to perform such testing and tell us the result?
Thanks.
On Sat, Oct 7, 2023 at 5:17 AM Andrew Morton <[email protected]> wrote:
>
> On Thu, 11 May 2023 13:22:30 +0800 "zhaoyang.huang" <[email protected]> wrote:
>
> > From: Zhaoyang Huang <[email protected]>
> >
> > Let us look at the timeline of scenarios below with WMARK_LOW=25MB WMARK_MIN=5MB
> > (managed pages 1.9GB). We can find that CMA begin to be used until 'C' under the
> > method of 'fixed 2 times of free cma over free pages' which could have the
> > scenario 'A' and 'B' into a fault state, that is, free UNMOVABLE & RECLAIMABLE
> > pages is lower than corresponding watermark without reclaiming which should be
> > deemed as against current memory policy. This commit try to solve this by
> > checking zone_watermark_ok again with removing CMA pages which could lead to a
> > proper time point of CMA's utilization.
> >
> > -- Free_pages
> > |
> > |
> > -- WMARK_LOW
> > |
> > -- Free_CMA
> > |
> > |
> > --
> >
> > Free_CMA/Free_pages(MB) A(12/30) --> B(12/25) --> C(12/20)
> > fixed 1/2 ratio N N Y
> > this commit Y Y Y
> >
>
> Roman previously asked
>
> : Also I'm a bit concerned about potential performance implications.
> : Would be great to provide some benchmarks or some data. Probably it's
> : ok because of we have pcp caches on top, but I'm not 100% sure.
>
> Are you able to perform such testing and tell us the result?
I have applied this patch in a v5.4 based ANDROID system and got no
regression problem. Actually, this commit is aimed to have
alloc_pages(GFP_USER) use CMA instead of stealing U&R(could lead to
GFP_KERNEL fail) only when zone's free pages and free cma are around
WATERMARK_LOW/MIN which would NOT affect most scenarios.
>
> Thanks.
>
On Sun, 8 Oct 2023 15:54:40 +0800 Zhaoyang Huang <[email protected]> wrote:
> > Roman previously asked
> >
> > : Also I'm a bit concerned about potential performance implications.
> > : Would be great to provide some benchmarks or some data. Probably it's
> > : ok because of we have pcp caches on top, but I'm not 100% sure.
> >
> > Are you able to perform such testing and tell us the result?
> I have applied this patch in a v5.4 based ANDROID system and got no
> regression problem. Actually, this commit is aimed to have
> alloc_pages(GFP_USER) use CMA instead of stealing U&R(could lead to
> GFP_KERNEL fail) only when zone's free pages and free cma are around
> WATERMARK_LOW/MIN which would NOT affect most scenarios.
OK, thanks.
Could the appropriate people please take a look at this? It has been
in mm-unstable since May.
Thanks.
On Mon, Oct 09, 2023 at 05:14:15PM -0700, Andrew Morton wrote:
> On Sun, 8 Oct 2023 15:54:40 +0800 Zhaoyang Huang <[email protected]> wrote:
>
> > > Roman previously asked
> > >
> > > : Also I'm a bit concerned about potential performance implications.
> > > : Would be great to provide some benchmarks or some data. Probably it's
> > > : ok because of we have pcp caches on top, but I'm not 100% sure.
> > >
> > > Are you able to perform such testing and tell us the result?
> > I have applied this patch in a v5.4 based ANDROID system and got no
> > regression problem. Actually, this commit is aimed to have
> > alloc_pages(GFP_USER) use CMA instead of stealing U&R(could lead to
> > GFP_KERNEL fail) only when zone's free pages and free cma are around
> > WATERMARK_LOW/MIN which would NOT affect most scenarios.
>
> OK, thanks.
>
> Could the appropriate people please take a look at this? It has been
> in mm-unstable since May.
I have 2 concerns:
1) it's still hard to understand the commit message and comments, I can
only reverse-engineer it based on the code changes;
2) performance concerns I expressed earlier are not addressed. Idk what's
a good benchmark for the page allocator, probably something i/o or
networking heavy.
On the positive side I believe that the patch is solving a real problem.
On Tue, Oct 10, 2023 at 9:31 AM Roman Gushchin <[email protected]> wrote:
>
> On Mon, Oct 09, 2023 at 05:14:15PM -0700, Andrew Morton wrote:
> > On Sun, 8 Oct 2023 15:54:40 +0800 Zhaoyang Huang <[email protected]> wrote:
> >
> > > > Roman previously asked
> > > >
> > > > : Also I'm a bit concerned about potential performance implications.
> > > > : Would be great to provide some benchmarks or some data. Probably it's
> > > > : ok because of we have pcp caches on top, but I'm not 100% sure.
> > > >
> > > > Are you able to perform such testing and tell us the result?
> > > I have applied this patch in a v5.4 based ANDROID system and got no
> > > regression problem. Actually, this commit is aimed to have
> > > alloc_pages(GFP_USER) use CMA instead of stealing U&R(could lead to
> > > GFP_KERNEL fail) only when zone's free pages and free cma are around
> > > WATERMARK_LOW/MIN which would NOT affect most scenarios.
> >
> > OK, thanks.
> >
> > Could the appropriate people please take a look at this? It has been
> > in mm-unstable since May.
>
> I have 2 concerns:
> 1) it's still hard to understand the commit message and comments, I can
> only reverse-engineer it based on the code changes;
> 2) performance concerns I expressed earlier are not addressed. Idk what's
> a good benchmark for the page allocator, probably something i/o or
> networking heavy.
>
ok, I will update the commit message
> On the positive side I believe that the patch is solving a real problem.