mm, compaction: periodically drop lock and restore IRQs in scanners
authorVlastimil Babka <vbabka@suse.cz>
Thu, 9 Oct 2014 22:27:16 +0000 (15:27 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 10 Oct 2014 02:25:54 +0000 (22:25 -0400)
Compaction scanners regularly check for lock contention and need_resched()
through the compact_checklock_irqsave() function.  However, if there is no
contention, the lock can be held and IRQ disabled for potentially long
time.

This has been addressed by commit b2eef8c0d091 ("mm: compaction: minimise
the time IRQs are disabled while isolating pages for migration") for the
migration scanner.  However, the refactoring done by commit 2a1402aa044b
("mm: compaction: acquire the zone->lru_lock as late as possible") has
changed the conditions so that the lock is dropped only when there's
contention on the lock or need_resched() is true.  Also, need_resched() is
checked only when the lock is already held.  The comment "give a chance to
irqs before checking need_resched" is therefore misleading, as IRQs remain
disabled when the check is done.

This patch restores the behavior intended by commit b2eef8c0d091 and also
tries to better balance and make more deterministic the time spent by
checking for contention vs the time the scanners might run between the
checks.  It also avoids situations where checking has not been done often
enough before.  The result should be avoiding both too frequent and too
infrequent contention checking, and especially the potentially
long-running scans with IRQs disabled and no checking of need_resched() or
for fatal signal pending, which can happen when many consecutive pages or
pageblocks fail the preliminary tests and do not reach the later call site
to compact_checklock_irqsave(), as explained below.

Before the patch:

In the migration scanner, compact_checklock_irqsave() was called each
loop, if reached.  If not reached, some lower-frequency checking could
still be done if the lock was already held, but this would not result in
aborting contended async compaction until reaching
compact_checklock_irqsave() or end of pageblock.  In the free scanner, it
was similar but completely without the periodical checking, so lock can be
potentially held until reaching the end of pageblock.

After the patch, in both scanners:

The periodical check is done as the first thing in the loop on each
SWAP_CLUSTER_MAX aligned pfn, using the new compact_unlock_should_abort()
function, which always unlocks the lock (if locked) and aborts async
compaction if scheduling is needed.  It also aborts any type of compaction
when a fatal signal is pending.

The compact_checklock_irqsave() function is replaced with a slightly
different compact_trylock_irqsave().  The biggest difference is that the
function is not called at all if the lock is already held.  The periodical
need_resched() checking is left solely to compact_unlock_should_abort().
The lock contention avoidance for async compaction is achieved by the
periodical unlock by compact_unlock_should_abort() and by using trylock in
compact_trylock_irqsave() and aborting when trylock fails.  Sync
compaction does not use trylock.

Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Zhang Yanfei <zhangyanfei@cn.fujitsu.com>
Acked-by: Minchan Kim <minchan@kernel.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Michal Nazarewicz <mina86@mina86.com>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rik van Riel <riel@redhat.com>
Acked-by: David Rientjes <rientjes@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/compaction.c

index 26bb20ef853dbe7039e2cdca309f5a34426a3a5f..74770e40cfe59a32f1d4de0c512a24dfed2b36d5 100644 (file)
@@ -223,61 +223,72 @@ static void update_pageblock_skip(struct compact_control *cc,
 }
 #endif /* CONFIG_COMPACTION */
 
-static int should_release_lock(spinlock_t *lock)
+/*
+ * Compaction requires the taking of some coarse locks that are potentially
+ * very heavily contended. For async compaction, back out if the lock cannot
+ * be taken immediately. For sync compaction, spin on the lock if needed.
+ *
+ * Returns true if the lock is held
+ * Returns false if the lock is not held and compaction should abort
+ */
+static bool compact_trylock_irqsave(spinlock_t *lock, unsigned long *flags,
+                                               struct compact_control *cc)
 {
-       /*
-        * Sched contention has higher priority here as we may potentially
-        * have to abort whole compaction ASAP. Returning with lock contention
-        * means we will try another zone, and further decisions are
-        * influenced only when all zones are lock contended. That means
-        * potentially missing a lock contention is less critical.
-        */
-       if (need_resched())
-               return COMPACT_CONTENDED_SCHED;
-       else if (spin_is_contended(lock))
-               return COMPACT_CONTENDED_LOCK;
+       if (cc->mode == MIGRATE_ASYNC) {
+               if (!spin_trylock_irqsave(lock, *flags)) {
+                       cc->contended = COMPACT_CONTENDED_LOCK;
+                       return false;
+               }
+       } else {
+               spin_lock_irqsave(lock, *flags);
+       }
 
-       return COMPACT_CONTENDED_NONE;
+       return true;
 }
 
 /*
  * Compaction requires the taking of some coarse locks that are potentially
- * very heavily contended. Check if the process needs to be scheduled or
- * if the lock is contended. For async compaction, back out in the event
- * if contention is severe. For sync compaction, schedule.
+ * very heavily contended. The lock should be periodically unlocked to avoid
+ * having disabled IRQs for a long time, even when there is nobody waiting on
+ * the lock. It might also be that allowing the IRQs will result in
+ * need_resched() becoming true. If scheduling is needed, async compaction
+ * aborts. Sync compaction schedules.
+ * Either compaction type will also abort if a fatal signal is pending.
+ * In either case if the lock was locked, it is dropped and not regained.
  *
- * Returns true if the lock is held.
- * Returns false if the lock is released and compaction should abort
+ * Returns true if compaction should abort due to fatal signal pending, or
+ *             async compaction due to need_resched()
+ * Returns false when compaction can continue (sync compaction might have
+ *             scheduled)
  */
-static bool compact_checklock_irqsave(spinlock_t *lock, unsigned long *flags,
-                                     bool locked, struct compact_control *cc)
+static bool compact_unlock_should_abort(spinlock_t *lock,
+               unsigned long flags, bool *locked, struct compact_control *cc)
 {
-       int contended = should_release_lock(lock);
+       if (*locked) {
+               spin_unlock_irqrestore(lock, flags);
+               *locked = false;
+       }
 
-       if (contended) {
-               if (locked) {
-                       spin_unlock_irqrestore(lock, *flags);
-                       locked = false;
-               }
+       if (fatal_signal_pending(current)) {
+               cc->contended = COMPACT_CONTENDED_SCHED;
+               return true;
+       }
 
-               /* async aborts if taking too long or contended */
+       if (need_resched()) {
                if (cc->mode == MIGRATE_ASYNC) {
-                       cc->contended = contended;
-                       return false;
+                       cc->contended = COMPACT_CONTENDED_SCHED;
+                       return true;
                }
-
                cond_resched();
        }
 
-       if (!locked)
-               spin_lock_irqsave(lock, *flags);
-       return true;
+       return false;
 }
 
 /*
  * Aside from avoiding lock contention, compaction also periodically checks
  * need_resched() and either schedules in sync compaction or aborts async
- * compaction. This is similar to what compact_checklock_irqsave() does, but
+ * compaction. This is similar to what compact_unlock_should_abort() does, but
  * is used where no lock is concerned.
  *
  * Returns false when no scheduling was needed, or sync compaction scheduled.
@@ -336,6 +347,16 @@ static unsigned long isolate_freepages_block(struct compact_control *cc,
                int isolated, i;
                struct page *page = cursor;
 
+               /*
+                * Periodically drop the lock (if held) regardless of its
+                * contention, to give chance to IRQs. Abort if fatal signal
+                * pending or async compaction detects need_resched()
+                */
+               if (!(blockpfn % SWAP_CLUSTER_MAX)
+                   && compact_unlock_should_abort(&cc->zone->lock, flags,
+                                                               &locked, cc))
+                       break;
+
                nr_scanned++;
                if (!pfn_valid_within(blockpfn))
                        goto isolate_fail;
@@ -353,8 +374,9 @@ static unsigned long isolate_freepages_block(struct compact_control *cc,
                 * spin on the lock and we acquire the lock as late as
                 * possible.
                 */
-               locked = compact_checklock_irqsave(&cc->zone->lock, &flags,
-                                                               locked, cc);
+               if (!locked)
+                       locked = compact_trylock_irqsave(&cc->zone->lock,
+                                                               &flags, cc);
                if (!locked)
                        break;
 
@@ -552,13 +574,15 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
 
        /* Time to isolate some pages for migration */
        for (; low_pfn < end_pfn; low_pfn++) {
-               /* give a chance to irqs before checking need_resched() */
-               if (locked && !(low_pfn % SWAP_CLUSTER_MAX)) {
-                       if (should_release_lock(&zone->lru_lock)) {
-                               spin_unlock_irqrestore(&zone->lru_lock, flags);
-                               locked = false;
-                       }
-               }
+               /*
+                * Periodically drop the lock (if held) regardless of its
+                * contention, to give chance to IRQs. Abort async compaction
+                * if contended.
+                */
+               if (!(low_pfn % SWAP_CLUSTER_MAX)
+                   && compact_unlock_should_abort(&zone->lru_lock, flags,
+                                                               &locked, cc))
+                       break;
 
                if (!pfn_valid_within(low_pfn))
                        continue;
@@ -620,10 +644,11 @@ isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
                    page_count(page) > page_mapcount(page))
                        continue;
 
-               /* Check if it is ok to still hold the lock */
-               locked = compact_checklock_irqsave(&zone->lru_lock, &flags,
-                                                               locked, cc);
-               if (!locked || fatal_signal_pending(current))
+               /* If the lock is not held, try to take it */
+               if (!locked)
+                       locked = compact_trylock_irqsave(&zone->lru_lock,
+                                                               &flags, cc);
+               if (!locked)
                        break;
 
                /* Recheck PageLRU and PageTransHuge under lock */