On 10/22/19 12:02 PM, Stefan Priebe - Profihost AG wrote:
>
> Am 22.10.19 um 09:48 schrieb Vlastimil Babka:
>> On 10/22/19 9:41 AM, Stefan Priebe - Profihost AG wrote:
>>>> Hi, could you try the patch below? I suspect you're hitting a corner
>>>> case where compaction_suitable() returns COMPACT_SKIPPED for the
>>>> ZONE_DMA, triggering reclaim even if other zones have plenty of free
>>>> memory. And should_continue_reclaim() then returns true until twice the
>>>> requested page size is reclaimed (compact_gap()). That means 4MB
>>>> reclaimed for each THP allocation attempt, which roughly matches the
>>>> trace data you preovided previously.
>>>>
>>>> The amplification to 4MB should be removed in patches merged for 5.4, so
>>>> it would be only 32 pages reclaimed per THP allocation. The patch below
>>>> tries to remove this corner case completely, and it should be more
>>>> visible on your 5.2.x, so please apply it there.
>>>>
>>> is there any reason to not apply that one on top of 4.19?
>>>
>>> Greets,
>>> Stefan
>>>
>>
>> It should work, cherrypicks fine without conflict here.
>
> OK but does not work ;-)
>
>
> mm/compaction.c: In function '__compaction_suitable':
> mm/compaction.c:1451:19: error: implicit declaration of function
> 'zone_managed_pages'; did you mean 'node_spanned_pages'?
> [-Werror=implicit-function-declaration]
> alloc_flags, zone_managed_pages(zone)))
> ^~~~~~~~~~~~~~~~~~
> node_spanned_pages
Ah, this?
----8<----
>From f1335e1c0d4b74205fc0cc40b5960223d6f1dec7 Mon Sep 17 00:00:00 2001
From: Vlastimil Babka <vbabka@xxxxxxx>
Date: Thu, 12 Sep 2019 13:40:46 +0200
Subject: [PATCH] WIP
---
include/linux/compaction.h | 7 ++++++-
include/trace/events/mmflags.h | 1 +
mm/compaction.c | 16 +++++++++++++--
mm/vmscan.c | 36 ++++++++++++++++++++++++----------
4 files changed, 47 insertions(+), 13 deletions(-)
diff --git a/include/linux/compaction.h b/include/linux/compaction.h
index 68250a57aace..2f3b331c5239 100644
--- a/include/linux/compaction.h
+++ b/include/linux/compaction.h
@@ -17,8 +17,13 @@ enum compact_priority {
};
/* Return values for compact_zone() and try_to_compact_pages() */
-/* When adding new states, please adjust include/trace/events/compaction.h */
+/* When adding new states, please adjust include/trace/events/mmflags.h */
enum compact_result {
+ /*
+ * The zone is too small to provide the requested allocation even if
+ * fully freed (i.e. ZONE_DMA for THP allocation due to lowmem reserves)
+ */
+ COMPACT_IMPOSSIBLE,
/* For more detailed tracepoint output - internal to compaction */
COMPACT_NOT_SUITABLE_ZONE,
/*
diff --git a/include/trace/events/mmflags.h b/include/trace/events/mmflags.h
index a81cffb76d89..d7aa9cece234 100644
--- a/include/trace/events/mmflags.h
+++ b/include/trace/events/mmflags.h
@@ -169,6 +169,7 @@ IF_HAVE_VM_SOFTDIRTY(VM_SOFTDIRTY, "softdirty" ) \
#ifdef CONFIG_COMPACTION
#define COMPACTION_STATUS \
+ EM( COMPACT_IMPOSSIBLE, "impossible") \
EM( COMPACT_SKIPPED, "skipped") \
EM( COMPACT_DEFERRED, "deferred") \
EM( COMPACT_CONTINUE, "continue") \
diff --git a/mm/compaction.c b/mm/compaction.c
index 5079ddbec8f9..7d2299c7faa2 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -1416,6 +1416,7 @@ static enum compact_result compact_finished(struct zone *zone,
/*
* compaction_suitable: Is this suitable to run compaction on this zone now?
* Returns
+ * COMPACT_IMPOSSIBLE If the allocation would fail even with all pages free
* COMPACT_SKIPPED - If there are too few free pages for compaction
* COMPACT_SUCCESS - If the allocation would succeed without compaction
* COMPACT_CONTINUE - If compaction should run now
@@ -1439,6 +1440,16 @@ static enum compact_result __compaction_suitable(struct zone *zone, int order,
alloc_flags))
return COMPACT_SUCCESS;
+ /*
+ * If the allocation would not succeed even with a fully free zone
+ * due to e.g. lowmem reserves, indicate that compaction can't possibly
+ * help and it would be pointless to reclaim.
+ */
+ watermark += 1UL << order;
+ if (!__zone_watermark_ok(zone, 0, watermark, classzone_idx,
+ alloc_flags, zone->managed_pages))
+ return COMPACT_IMPOSSIBLE;
+
/*
* Watermarks for order-0 must be met for compaction to be able to
* isolate free pages for migration targets. This means that the
@@ -1526,7 +1537,7 @@ bool compaction_zonelist_suitable(struct alloc_context *ac, int order,
available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
compact_result = __compaction_suitable(zone, order, alloc_flags,
ac_classzone_idx(ac), available);
- if (compact_result != COMPACT_SKIPPED)
+ if (compact_result > COMPACT_SKIPPED)
return true;
}
@@ -1555,7 +1566,8 @@ static enum compact_result compact_zone(struct zone *zone, struct compact_contro
ret = compaction_suitable(zone, cc->order, cc->alloc_flags,
cc->classzone_idx);
/* Compaction is likely to fail */
- if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED)
+ if (ret == COMPACT_SUCCESS || ret == COMPACT_SKIPPED
+ || ret == COMPACT_IMPOSSIBLE)
return ret;
/* huh, compaction_suitable is returning something unexpected */
diff --git a/mm/vmscan.c b/mm/vmscan.c
index b37610c0eac6..7ad331a64fc5 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2849,11 +2849,12 @@ static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
}
/*
- * Returns true if compaction should go ahead for a costly-order request, or
- * the allocation would already succeed without compaction. Return false if we
- * should reclaim first.
+ * Returns 1 if compaction should go ahead for a costly-order request, or the
+ * allocation would already succeed without compaction. Return 0 if we should
+ * reclaim first. Return -1 when compaction can't help at all due to zone being
+ * too small, which means there's no point in reclaim nor compaction.
*/
-static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
+static inline int compaction_ready(struct zone *zone, struct scan_control *sc)
{
unsigned long watermark;
enum compact_result suitable;
@@ -2861,10 +2862,16 @@ static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
if (suitable == COMPACT_SUCCESS)
/* Allocation should succeed already. Don't reclaim. */
- return true;
+ return 1;
if (suitable == COMPACT_SKIPPED)
/* Compaction cannot yet proceed. Do reclaim. */
- return false;
+ return 0;
+ if (suitable == COMPACT_IMPOSSIBLE)
+ /*
+ * Compaction can't possibly help. So don't reclaim, but keep
+ * checking other zones.
+ */
+ return -1;
/*
* Compaction is already possible, but it takes time to run and there
@@ -2910,6 +2917,7 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
for_each_zone_zonelist_nodemask(zone, z, zonelist,
sc->reclaim_idx, sc->nodemask) {
+ int compact_ready;
/*
* Take care memory controller reclaiming has small influence
* to global LRU.
@@ -2929,10 +2937,18 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
* page allocations.
*/
if (IS_ENABLED(CONFIG_COMPACTION) &&
- sc->order > PAGE_ALLOC_COSTLY_ORDER &&
- compaction_ready(zone, sc)) {
- sc->compaction_ready = true;
- continue;
+ sc->order > PAGE_ALLOC_COSTLY_ORDER) {
+ compact_ready = compaction_ready(zone, sc);
+ if (compact_ready == 1) {
+ sc->compaction_ready = true;
+ continue;
+ } else if (compact_ready == -1) {
+ /*
+ * In this zone, neither reclaim nor
+ * compaction can help.
+ */
+ continue;
+ }
}
/*
--
2.23.0
