The following two patches are aimed at reducing the amount of time IRQs are
disabled. It was reported by some ALSA people that transparent hugepages was
causing slowdowns on MIDI playback but I strongly suspect that compaction
running for smaller orders was also a factor. The theory was that IRQs
were being disabled for too long and sure enough, compaction was found to
be disabling IRQs for a long time. The patches reduce the length of time
IRQs are disabled when scanning for free pages and for pages to migrate.
It's late in the cycle but the IRQs disabled times are sufficiently bad
that it would be desirable to have these merged for 2.6.38 if possible.
mm/compaction.c | 35 ++++++++++++++++++++++++++++++-----
1 files changed, 30 insertions(+), 5 deletions(-)
--
1.7.2.3
From: Andrea Arcangeli <[email protected]>
compaction_alloc() isolates pages for migration in isolate_migratepages. While
it's scanning, IRQs are disabled on the mistaken assumption the scanning
should be short. Tests show this to be true for the most part but
contention times on the LRU lock can be increased. Before this patch,
the IRQ disabled times for a simple test looked like
Total sampled time IRQs off (not real total time): 5493
Event shrink_inactive_list..shrink_zone 1596 us count 1
Event shrink_inactive_list..shrink_zone 1530 us count 1
Event shrink_inactive_list..shrink_zone 956 us count 1
Event shrink_inactive_list..shrink_zone 541 us count 1
Event shrink_inactive_list..shrink_zone 531 us count 1
Event split_huge_page..add_to_swap 232 us count 1
Event save_args..call_softirq 36 us count 1
Event save_args..call_softirq 35 us count 2
Event __wake_up..__wake_up 1 us count 1
This patch reduces the worst-case IRQs-disabled latencies by releasing the
lock every SWAP_CLUSTER_MAX pages that are scanned and releasing the CPU if
necessary. The cost of this is that the processing performing compaction will
be slower but IRQs being disabled for too long a time has worse consequences
as the following report shows;
Total sampled time IRQs off (not real total time): 4367
Event shrink_inactive_list..shrink_zone 881 us count 1
Event shrink_inactive_list..shrink_zone 875 us count 1
Event shrink_inactive_list..shrink_zone 868 us count 1
Event shrink_inactive_list..shrink_zone 555 us count 1
Event split_huge_page..add_to_swap 495 us count 1
Event compact_zone..compact_zone_order 269 us count 1
Event split_huge_page..add_to_swap 266 us count 1
Event shrink_inactive_list..shrink_zone 85 us count 1
Event save_args..call_softirq 36 us count 2
Event __wake_up..__wake_up 1 us count 1
Signed-off-by: Andrea Arcangeli <[email protected]>
Signed-off-by: Mel Gorman <[email protected]>
---
mm/compaction.c | 18 ++++++++++++++++++
1 files changed, 18 insertions(+), 0 deletions(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index f47de94..5babbbb 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -278,9 +278,27 @@ static unsigned long isolate_migratepages(struct zone *zone,
}
/* Time to isolate some pages for migration */
+ cond_resched();
spin_lock_irq(&zone->lru_lock);
for (; low_pfn < end_pfn; low_pfn++) {
struct page *page;
+ bool unlocked = false;
+
+ /* give a chance to irqs before checking need_resched() */
+ if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
+ spin_unlock_irq(&zone->lru_lock);
+ unlocked = true;
+ }
+ if (need_resched() || spin_is_contended(&zone->lru_lock)) {
+ if (!unlocked)
+ spin_unlock_irq(&zone->lru_lock);
+ cond_resched();
+ spin_lock_irq(&zone->lru_lock);
+ if (fatal_signal_pending(current))
+ break;
+ } else if (unlocked)
+ spin_lock_irq(&zone->lru_lock);
+
if (!pfn_valid_within(low_pfn))
continue;
nr_scanned++;
--
1.7.2.3
compaction_alloc() isolates free pages to be used as migration targets.
While its scanning, IRQs are disabled on the mistaken assumption the scanning
should be short. Analysis showed that IRQs were in fact being disabled for
substantial time. A simple test was run using large anonymous mappings with
transparent hugepage support enabled to trigger frequent compactions. A
monitor sampled what the worst IRQ-off latencies were and a post-processing
tool found the following;
Total sampled time IRQs off (not real total time): 22355
Event compaction_alloc..compaction_alloc 8409 us count 1
Event compaction_alloc..compaction_alloc 7341 us count 1
Event compaction_alloc..compaction_alloc 2463 us count 1
Event compaction_alloc..compaction_alloc 2054 us count 1
Event shrink_inactive_list..shrink_zone 1864 us count 1
Event shrink_inactive_list..shrink_zone 88 us count 1
Event save_args..call_softirq 36 us count 1
Event save_args..call_softirq 35 us count 2
Event __make_request..__blk_run_queue 24 us count 1
Event __alloc_pages_nodemask..__alloc_pages_nodemask 6 us count 1
i.e. compaction is disabled IRQs for a prolonged period of time - 8ms in
one instance. The full report generated by the tool can be found at
http://www.csn.ul.ie/~mel/postings/minfree-20110225/irqsoff-vanilla-micro.report .
This patch reduces the time IRQs are disabled by simply disabling IRQs
at the last possible minute. An updated IRQs-off summary report then
looks like;
Total sampled time IRQs off (not real total time): 5493
Event shrink_inactive_list..shrink_zone 1596 us count 1
Event shrink_inactive_list..shrink_zone 1530 us count 1
Event shrink_inactive_list..shrink_zone 956 us count 1
Event shrink_inactive_list..shrink_zone 541 us count 1
Event shrink_inactive_list..shrink_zone 531 us count 1
Event split_huge_page..add_to_swap 232 us count 1
Event save_args..call_softirq 36 us count 1
Event save_args..call_softirq 35 us count 2
Event __wake_up..__wake_up 1 us count 1
A full report is again available at
http://www.csn.ul.ie/~mel/postings/minfree-20110225/irqsoff-minimiseirq-free-v1r4-micro.report .
. As should be obvious, IRQ disabled latencies due to compaction are
almost elimimnated for this particular test.
Signed-off-by: Mel Gorman <[email protected]>
---
mm/compaction.c | 17 ++++++++++++-----
1 files changed, 12 insertions(+), 5 deletions(-)
diff --git a/mm/compaction.c b/mm/compaction.c
index 8be430b8..f47de94 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -155,7 +155,6 @@ static void isolate_freepages(struct zone *zone,
* pages on cc->migratepages. We stop searching if the migrate
* and free page scanners meet or enough free pages are isolated.
*/
- spin_lock_irqsave(&zone->lock, flags);
for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
pfn -= pageblock_nr_pages) {
unsigned long isolated;
@@ -178,9 +177,18 @@ static void isolate_freepages(struct zone *zone,
if (!suitable_migration_target(page))
continue;
- /* Found a block suitable for isolating free pages from */
- isolated = isolate_freepages_block(zone, pfn, freelist);
- nr_freepages += isolated;
+ /*
+ * Found a block suitable for isolating free pages from. Now
+ * we disabled interrupts, double check things are ok and
+ * isolate the pages. This is to minimise the time IRQs
+ * are disabled
+ */
+ spin_lock_irqsave(&zone->lock, flags);
+ if (suitable_migration_target(page)) {
+ isolated = isolate_freepages_block(zone, pfn, freelist);
+ nr_freepages += isolated;
+ }
+ spin_unlock_irqrestore(&zone->lock, flags);
/*
* Record the highest PFN we isolated pages from. When next
@@ -190,7 +198,6 @@ static void isolate_freepages(struct zone *zone,
if (isolated)
high_pfn = max(high_pfn, pfn);
}
- spin_unlock_irqrestore(&zone->lock, flags);
/* split_free_page does not map the pages */
list_for_each_entry(page, freelist, lru) {
--
1.7.2.3