mm, compaction: move pageblock checks up from isolate_migratepages_range()
authorVlastimil Babka <vbabka@suse.cz>
Thu, 9 Oct 2014 22:27:09 +0000 (15:27 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 10 Oct 2014 02:25:54 +0000 (22:25 -0400)
isolate_migratepages_range() is the main function of the compaction
scanner, called either on a single pageblock by isolate_migratepages()
during regular compaction, or on an arbitrary range by CMA's
__alloc_contig_migrate_range().  It currently perfoms two pageblock-wide
compaction suitability checks, and because of the CMA callpath, it tracks
if it crossed a pageblock boundary in order to repeat those checks.

However, closer inspection shows that those checks are always true for CMA:
- isolation_suitable() is true because CMA sets cc->ignore_skip_hint to true
- migrate_async_suitable() check is skipped because CMA uses sync compaction

We can therefore move the compaction-specific checks to
isolate_migratepages() and simplify isolate_migratepages_range().
Furthermore, we can mimic the freepage scanner family of functions, which
has isolate_freepages_block() function called both by compaction from
isolate_freepages() and by CMA from isolate_freepages_range(), where each
use-case adds own specific glue code.  This allows further code
simplification.

Thus, we rename isolate_migratepages_range() to
isolate_migratepages_block() and limit its functionality to a single
pageblock (or its subset).  For CMA, a new different
isolate_migratepages_range() is created as a CMA-specific wrapper for the
_block() function.  The checks specific to compaction are moved to
isolate_migratepages().  As part of the unification of these two families
of functions, we remove the redundant zone parameter where applicable,
since zone pointer is already passed in cc->zone.

Furthermore, going back to compact_zone() and compact_finished() when
pageblock is found unsuitable (now by isolate_migratepages()) is wasteful
- the checks are meant to skip pageblocks quickly.  The patch therefore
also introduces a simple loop into isolate_migratepages() so that it does
not return immediately on failed pageblock checks, but keeps going until
isolate_migratepages_range() gets called once.  Similarily to
isolate_freepages(), the function periodically checks if it needs to
reschedule or abort async compaction.

[iamjoonsoo.kim@lge.com: fix isolated page counting bug in compaction]
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Minchan Kim <minchan@kernel.org>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
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>
Cc: 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
mm/internal.h
mm/page_alloc.c

index 7bf150d4e1c8bc1eebddac749ce77f6755b5d6a6..8058e3f98f08a8c98b57b6af0cfbcd64a6dedc0b 100644 (file)
@@ -132,7 +132,7 @@ void reset_isolation_suitable(pg_data_t *pgdat)
  */
 static void update_pageblock_skip(struct compact_control *cc,
                        struct page *page, unsigned long nr_isolated,
-                       bool set_unsuitable, bool migrate_scanner)
+                       bool migrate_scanner)
 {
        struct zone *zone = cc->zone;
        unsigned long pfn;
@@ -146,12 +146,7 @@ static void update_pageblock_skip(struct compact_control *cc,
        if (nr_isolated)
                return;
 
-       /*
-        * Only skip pageblocks when all forms of compaction will be known to
-        * fail in the near future.
-        */
-       if (set_unsuitable)
-               set_pageblock_skip(page);
+       set_pageblock_skip(page);
 
        pfn = page_to_pfn(page);
 
@@ -180,7 +175,7 @@ static inline bool isolation_suitable(struct compact_control *cc,
 
 static void update_pageblock_skip(struct compact_control *cc,
                        struct page *page, unsigned long nr_isolated,
-                       bool set_unsuitable, bool migrate_scanner)
+                       bool migrate_scanner)
 {
 }
 #endif /* CONFIG_COMPACTION */
@@ -348,8 +343,7 @@ isolate_fail:
 
        /* Update the pageblock-skip if the whole pageblock was scanned */
        if (blockpfn == end_pfn)
-               update_pageblock_skip(cc, valid_page, total_isolated, true,
-                                     false);
+               update_pageblock_skip(cc, valid_page, total_isolated, false);
 
        count_compact_events(COMPACTFREE_SCANNED, nr_scanned);
        if (total_isolated)
@@ -420,22 +414,19 @@ isolate_freepages_range(struct compact_control *cc,
 }
 
 /* Update the number of anon and file isolated pages in the zone */
-static void acct_isolated(struct zone *zone, bool locked, struct compact_control *cc)
+static void acct_isolated(struct zone *zone, struct compact_control *cc)
 {
        struct page *page;
        unsigned int count[2] = { 0, };
 
+       if (list_empty(&cc->migratepages))
+               return;
+
        list_for_each_entry(page, &cc->migratepages, lru)
                count[!!page_is_file_cache(page)]++;
 
-       /* If locked we can use the interrupt unsafe versions */
-       if (locked) {
-               __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
-               __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
-       } else {
-               mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
-               mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
-       }
+       mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
+       mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
 }
 
 /* Similar to reclaim, but different enough that they don't share logic */
@@ -454,40 +445,34 @@ static bool too_many_isolated(struct zone *zone)
 }
 
 /**
- * isolate_migratepages_range() - isolate all migrate-able pages in range.
- * @zone:      Zone pages are in.
+ * isolate_migratepages_block() - isolate all migrate-able pages within
+ *                               a single pageblock
  * @cc:                Compaction control structure.
- * @low_pfn:   The first PFN of the range.
- * @end_pfn:   The one-past-the-last PFN of the range.
- * @unevictable: true if it allows to isolate unevictable pages
+ * @low_pfn:   The first PFN to isolate
+ * @end_pfn:   The one-past-the-last PFN to isolate, within same pageblock
+ * @isolate_mode: Isolation mode to be used.
  *
  * Isolate all pages that can be migrated from the range specified by
- * [low_pfn, end_pfn).  Returns zero if there is a fatal signal
- * pending), otherwise PFN of the first page that was not scanned
- * (which may be both less, equal to or more then end_pfn).
+ * [low_pfn, end_pfn). The range is expected to be within same pageblock.
+ * Returns zero if there is a fatal signal pending, otherwise PFN of the
+ * first page that was not scanned (which may be both less, equal to or more
+ * than end_pfn).
  *
- * Assumes that cc->migratepages is empty and cc->nr_migratepages is
- * zero.
- *
- * Apart from cc->migratepages and cc->nr_migratetypes this function
- * does not modify any cc's fields, in particular it does not modify
- * (or read for that matter) cc->migrate_pfn.
+ * The pages are isolated on cc->migratepages list (not required to be empty),
+ * and cc->nr_migratepages is updated accordingly. The cc->migrate_pfn field
+ * is neither read nor updated.
  */
-unsigned long
-isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
-               unsigned long low_pfn, unsigned long end_pfn, bool unevictable)
+static unsigned long
+isolate_migratepages_block(struct compact_control *cc, unsigned long low_pfn,
+                       unsigned long end_pfn, isolate_mode_t isolate_mode)
 {
-       unsigned long last_pageblock_nr = 0, pageblock_nr;
+       struct zone *zone = cc->zone;
        unsigned long nr_scanned = 0, nr_isolated = 0;
        struct list_head *migratelist = &cc->migratepages;
        struct lruvec *lruvec;
        unsigned long flags;
        bool locked = false;
        struct page *page = NULL, *valid_page = NULL;
-       bool set_unsuitable = true;
-       const isolate_mode_t mode = (cc->mode == MIGRATE_ASYNC ?
-                                       ISOLATE_ASYNC_MIGRATE : 0) |
-                                   (unevictable ? ISOLATE_UNEVICTABLE : 0);
 
        /*
         * Ensure that there are not too many pages isolated from the LRU
@@ -518,19 +503,6 @@ isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
                        }
                }
 
-               /*
-                * migrate_pfn does not necessarily start aligned to a
-                * pageblock. Ensure that pfn_valid is called when moving
-                * into a new MAX_ORDER_NR_PAGES range in case of large
-                * memory holes within the zone
-                */
-               if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
-                       if (!pfn_valid(low_pfn)) {
-                               low_pfn += MAX_ORDER_NR_PAGES - 1;
-                               continue;
-                       }
-               }
-
                if (!pfn_valid_within(low_pfn))
                        continue;
                nr_scanned++;
@@ -548,28 +520,6 @@ isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
                if (!valid_page)
                        valid_page = page;
 
-               /* If isolation recently failed, do not retry */
-               pageblock_nr = low_pfn >> pageblock_order;
-               if (last_pageblock_nr != pageblock_nr) {
-                       int mt;
-
-                       last_pageblock_nr = pageblock_nr;
-                       if (!isolation_suitable(cc, page))
-                               goto next_pageblock;
-
-                       /*
-                        * For async migration, also only scan in MOVABLE
-                        * blocks. Async migration is optimistic to see if
-                        * the minimum amount of work satisfies the allocation
-                        */
-                       mt = get_pageblock_migratetype(page);
-                       if (cc->mode == MIGRATE_ASYNC &&
-                           !migrate_async_suitable(mt)) {
-                               set_unsuitable = false;
-                               goto next_pageblock;
-                       }
-               }
-
                /*
                 * Skip if free. page_order cannot be used without zone->lock
                 * as nothing prevents parallel allocations or buddy merging.
@@ -604,8 +554,11 @@ isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
                 */
                if (PageTransHuge(page)) {
                        if (!locked)
-                               goto next_pageblock;
-                       low_pfn += (1 << compound_order(page)) - 1;
+                               low_pfn = ALIGN(low_pfn + 1,
+                                               pageblock_nr_pages) - 1;
+                       else
+                               low_pfn += (1 << compound_order(page)) - 1;
+
                        continue;
                }
 
@@ -635,7 +588,7 @@ isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
                lruvec = mem_cgroup_page_lruvec(page, zone);
 
                /* Try isolate the page */
-               if (__isolate_lru_page(page, mode) != 0)
+               if (__isolate_lru_page(page, isolate_mode) != 0)
                        continue;
 
                VM_BUG_ON_PAGE(PageTransCompound(page), page);
@@ -654,15 +607,8 @@ isolate_success:
                        ++low_pfn;
                        break;
                }
-
-               continue;
-
-next_pageblock:
-               low_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages) - 1;
        }
 
-       acct_isolated(zone, locked, cc);
-
        if (locked)
                spin_unlock_irqrestore(&zone->lru_lock, flags);
 
@@ -671,8 +617,7 @@ next_pageblock:
         * if the whole pageblock was scanned without isolating any page.
         */
        if (low_pfn == end_pfn)
-               update_pageblock_skip(cc, valid_page, nr_isolated,
-                                     set_unsuitable, true);
+               update_pageblock_skip(cc, valid_page, nr_isolated, true);
 
        trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
 
@@ -683,15 +628,63 @@ next_pageblock:
        return low_pfn;
 }
 
+/**
+ * isolate_migratepages_range() - isolate migrate-able pages in a PFN range
+ * @cc:        Compaction control structure.
+ * @start_pfn: The first PFN to start isolating.
+ * @end_pfn:   The one-past-last PFN.
+ *
+ * Returns zero if isolation fails fatally due to e.g. pending signal.
+ * Otherwise, function returns one-past-the-last PFN of isolated page
+ * (which may be greater than end_pfn if end fell in a middle of a THP page).
+ */
+unsigned long
+isolate_migratepages_range(struct compact_control *cc, unsigned long start_pfn,
+                                                       unsigned long end_pfn)
+{
+       unsigned long pfn, block_end_pfn;
+
+       /* Scan block by block. First and last block may be incomplete */
+       pfn = start_pfn;
+       block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
+
+       for (; pfn < end_pfn; pfn = block_end_pfn,
+                               block_end_pfn += pageblock_nr_pages) {
+
+               block_end_pfn = min(block_end_pfn, end_pfn);
+
+               /* Skip whole pageblock in case of a memory hole */
+               if (!pfn_valid(pfn))
+                       continue;
+
+               pfn = isolate_migratepages_block(cc, pfn, block_end_pfn,
+                                                       ISOLATE_UNEVICTABLE);
+
+               /*
+                * In case of fatal failure, release everything that might
+                * have been isolated in the previous iteration, and signal
+                * the failure back to caller.
+                */
+               if (!pfn) {
+                       putback_movable_pages(&cc->migratepages);
+                       cc->nr_migratepages = 0;
+                       break;
+               }
+       }
+       acct_isolated(cc->zone, cc);
+
+       return pfn;
+}
+
 #endif /* CONFIG_COMPACTION || CONFIG_CMA */
 #ifdef CONFIG_COMPACTION
 /*
  * Based on information in the current compact_control, find blocks
  * suitable for isolating free pages from and then isolate them.
  */
-static void isolate_freepages(struct zone *zone,
-                               struct compact_control *cc)
+static void isolate_freepages(struct compact_control *cc)
 {
+       struct zone *zone = cc->zone;
        struct page *page;
        unsigned long block_start_pfn;  /* start of current pageblock */
        unsigned long block_end_pfn;    /* end of current pageblock */
@@ -809,7 +802,7 @@ static struct page *compaction_alloc(struct page *migratepage,
         */
        if (list_empty(&cc->freepages)) {
                if (!cc->contended)
-                       isolate_freepages(cc->zone, cc);
+                       isolate_freepages(cc);
 
                if (list_empty(&cc->freepages))
                        return NULL;
@@ -843,34 +836,82 @@ typedef enum {
 } isolate_migrate_t;
 
 /*
- * Isolate all pages that can be migrated from the block pointed to by
- * the migrate scanner within compact_control.
+ * Isolate all pages that can be migrated from the first suitable block,
+ * starting at the block pointed to by the migrate scanner pfn within
+ * compact_control.
  */
 static isolate_migrate_t isolate_migratepages(struct zone *zone,
                                        struct compact_control *cc)
 {
        unsigned long low_pfn, end_pfn;
+       struct page *page;
+       const isolate_mode_t isolate_mode =
+               (cc->mode == MIGRATE_ASYNC ? ISOLATE_ASYNC_MIGRATE : 0);
 
-       /* Do not scan outside zone boundaries */
-       low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
+       /*
+        * Start at where we last stopped, or beginning of the zone as
+        * initialized by compact_zone()
+        */
+       low_pfn = cc->migrate_pfn;
 
        /* Only scan within a pageblock boundary */
        end_pfn = ALIGN(low_pfn + 1, pageblock_nr_pages);
 
-       /* Do not cross the free scanner or scan within a memory hole */
-       if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
-               cc->migrate_pfn = end_pfn;
-               return ISOLATE_NONE;
-       }
+       /*
+        * Iterate over whole pageblocks until we find the first suitable.
+        * Do not cross the free scanner.
+        */
+       for (; end_pfn <= cc->free_pfn;
+                       low_pfn = end_pfn, end_pfn += pageblock_nr_pages) {
 
-       /* Perform the isolation */
-       low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn, false);
-       if (!low_pfn || cc->contended)
-               return ISOLATE_ABORT;
+               /*
+                * This can potentially iterate a massively long zone with
+                * many pageblocks unsuitable, so periodically check if we
+                * need to schedule, or even abort async compaction.
+                */
+               if (!(low_pfn % (SWAP_CLUSTER_MAX * pageblock_nr_pages))
+                                               && compact_should_abort(cc))
+                       break;
+
+               /* Skip whole pageblock in case of a memory hole */
+               if (!pfn_valid(low_pfn))
+                       continue;
 
+               page = pfn_to_page(low_pfn);
+
+               /* If isolation recently failed, do not retry */
+               if (!isolation_suitable(cc, page))
+                       continue;
+
+               /*
+                * For async compaction, also only scan in MOVABLE blocks.
+                * Async compaction is optimistic to see if the minimum amount
+                * of work satisfies the allocation.
+                */
+               if (cc->mode == MIGRATE_ASYNC &&
+                   !migrate_async_suitable(get_pageblock_migratetype(page)))
+                       continue;
+
+               /* Perform the isolation */
+               low_pfn = isolate_migratepages_block(cc, low_pfn, end_pfn,
+                                                               isolate_mode);
+
+               if (!low_pfn || cc->contended)
+                       return ISOLATE_ABORT;
+
+               /*
+                * Either we isolated something and proceed with migration. Or
+                * we failed and compact_zone should decide if we should
+                * continue or not.
+                */
+               break;
+       }
+
+       acct_isolated(zone, cc);
+       /* Record where migration scanner will be restarted */
        cc->migrate_pfn = low_pfn;
 
-       return ISOLATE_SUCCESS;
+       return cc->nr_migratepages ? ISOLATE_SUCCESS : ISOLATE_NONE;
 }
 
 static int compact_finished(struct zone *zone,
@@ -1043,9 +1084,6 @@ static int compact_zone(struct zone *zone, struct compact_control *cc)
                        ;
                }
 
-               if (!cc->nr_migratepages)
-                       continue;
-
                err = migrate_pages(&cc->migratepages, compaction_alloc,
                                compaction_free, (unsigned long)cc, cc->mode,
                                MR_COMPACTION);
index a1b651b11c5fcba7a0322bc19c93286896d08cbf..5a0738fa649c2d8869b82fd788cad262b2e8d4c3 100644 (file)
@@ -154,8 +154,8 @@ unsigned long
 isolate_freepages_range(struct compact_control *cc,
                        unsigned long start_pfn, unsigned long end_pfn);
 unsigned long
-isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
-       unsigned long low_pfn, unsigned long end_pfn, bool unevictable);
+isolate_migratepages_range(struct compact_control *cc,
+                          unsigned long low_pfn, unsigned long end_pfn);
 
 #endif
 
index 822babd808feedfc2fe89aff75eeccf8e436607a..dfbf54b51649ebefe70019778ae3303ff492b2b9 100644 (file)
@@ -6288,8 +6288,7 @@ static int __alloc_contig_migrate_range(struct compact_control *cc,
 
                if (list_empty(&cc->migratepages)) {
                        cc->nr_migratepages = 0;
-                       pfn = isolate_migratepages_range(cc->zone, cc,
-                                                        pfn, end, true);
+                       pfn = isolate_migratepages_range(cc, pfn, end);
                        if (!pfn) {
                                ret = -EINTR;
                                break;
This page took 0.032221 seconds and 5 git commands to generate.