swap: make cluster allocation per-cpu
authorShaohua Li <shli@kernel.org>
Wed, 11 Sep 2013 21:20:32 +0000 (14:20 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 11 Sep 2013 22:57:17 +0000 (15:57 -0700)
swap cluster allocation is to get better request merge to improve
performance.  But the cluster is shared globally, if multiple tasks are
doing swap, this will cause interleave disk access.  While multiple tasks
swap is quite common, for example, each numa node has a kswapd thread
doing swap and multiple threads/processes doing direct page reclaim.

ioscheduler can't help too much here, because tasks don't send swapout IO
down to block layer in the meantime.  Block layer does merge some IOs, but
a lot not, depending on how many tasks are doing swapout concurrently.  In
practice, I've seen a lot of small size IO in swapout workloads.

We makes the cluster allocation per-cpu here.  The interleave disk access
issue goes away.  All tasks swapout to their own cluster, so swapout will
become sequential, which can be easily merged to big size IO.  If one CPU
can't get its per-cpu cluster (for example, there is no free cluster
anymore in the swap), it will fallback to scan swap_map.  The CPU can
still continue swap.  We don't need recycle free swap entries of other
CPUs.

In my test (swap to a 2-disk raid0 partition), this improves around 10%
swapout throughput, and request size is increased significantly.

How does this impact swap readahead is uncertain though.  On one side,
page reclaim always isolates and swaps several adjancent pages, this will
make page reclaim write the pages sequentially and benefit readahead.  On
the other side, several CPU write pages interleave means the pages don't
live _sequentially_ but relatively _near_.  In the per-cpu allocation
case, if adjancent pages are written by different cpus, they will live
relatively _far_.  So how this impacts swap readahead depends on how many
pages page reclaim isolates and swaps one time.  If the number is big,
this patch will benefit swap readahead.  Of course, this is about
sequential access pattern.  The patch has no impact for random access
pattern, because the new cluster allocation algorithm is just for SSD.

Alternative solution is organizing swap layout to be per-mm instead of
this per-cpu approach.  In the per-mm layout, we allocate a disk range for
each mm, so pages of one mm live in swap disk adjacently.  per-mm layout
has potential issues of lock contention if multiple reclaimers are swap
pages from one mm.  For a sequential workload, per-mm layout is better to
implement swap readahead, because pages from the mm are adjacent in disk.
But per-cpu layout isn't very bad in this workload, as page reclaim always
isolates and swaps several pages one time, such pages will still live in
disk sequentially and readahead can utilize this.  For a random workload,
per-mm layout isn't beneficial of request merge, because it's quite
possible pages from different mm are swapout in the meantime and IO can't
be merged in per-mm layout.  while with per-cpu layout we can merge
requests from any mm.  Considering random workload is more popular in
workloads with swap (and per-cpu approach isn't too bad for sequential
workload too), I'm choosing per-cpu layout.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Shaohua Li <shli@fusionio.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Kyungmin Park <kmpark@infradead.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Rafael Aquini <aquini@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/swap.h
mm/swapfile.c

index 8a3c4a1caa14f501bac7d812f69ca4cc359ed0f7..24db9142e93b0ec438c7b04983c5205e78a0c7d0 100644 (file)
@@ -198,6 +198,16 @@ struct swap_cluster_info {
 #define CLUSTER_FLAG_FREE 1 /* This cluster is free */
 #define CLUSTER_FLAG_NEXT_NULL 2 /* This cluster has no next cluster */
 
+/*
+ * We assign a cluster to each CPU, so each CPU can allocate swap entry from
+ * its own cluster and swapout sequentially. The purpose is to optimize swapout
+ * throughput.
+ */
+struct percpu_cluster {
+       struct swap_cluster_info index; /* Current cluster index */
+       unsigned int next; /* Likely next allocation offset */
+};
+
 /*
  * The in-memory structure used to track swap areas.
  */
@@ -217,6 +227,7 @@ struct swap_info_struct {
        unsigned int inuse_pages;       /* number of those currently in use */
        unsigned int cluster_next;      /* likely index for next allocation */
        unsigned int cluster_nr;        /* countdown to next cluster search */
+       struct percpu_cluster __percpu *percpu_cluster; /* per cpu's swap location */
        struct swap_extent *curr_swap_extent;
        struct swap_extent first_swap_extent;
        struct block_device *bdev;      /* swap device or bdev of swap file */
index 98e52e373bd80be9a35f4197556994d40c2acc7d..3963fc24fcc1b6f8c4d365de4d99bda8993311b7 100644 (file)
@@ -392,13 +392,78 @@ static void dec_cluster_info_page(struct swap_info_struct *p,
  * It's possible scan_swap_map() uses a free cluster in the middle of free
  * cluster list. Avoiding such abuse to avoid list corruption.
  */
-static inline bool scan_swap_map_recheck_cluster(struct swap_info_struct *si,
+static bool
+scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
        unsigned long offset)
 {
+       struct percpu_cluster *percpu_cluster;
+       bool conflict;
+
        offset /= SWAPFILE_CLUSTER;
-       return !cluster_is_null(&si->free_cluster_head) &&
+       conflict = !cluster_is_null(&si->free_cluster_head) &&
                offset != cluster_next(&si->free_cluster_head) &&
                cluster_is_free(&si->cluster_info[offset]);
+
+       if (!conflict)
+               return false;
+
+       percpu_cluster = this_cpu_ptr(si->percpu_cluster);
+       cluster_set_null(&percpu_cluster->index);
+       return true;
+}
+
+/*
+ * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
+ * might involve allocating a new cluster for current CPU too.
+ */
+static void scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
+       unsigned long *offset, unsigned long *scan_base)
+{
+       struct percpu_cluster *cluster;
+       bool found_free;
+       unsigned long tmp;
+
+new_cluster:
+       cluster = this_cpu_ptr(si->percpu_cluster);
+       if (cluster_is_null(&cluster->index)) {
+               if (!cluster_is_null(&si->free_cluster_head)) {
+                       cluster->index = si->free_cluster_head;
+                       cluster->next = cluster_next(&cluster->index) *
+                                       SWAPFILE_CLUSTER;
+               } else if (!cluster_is_null(&si->discard_cluster_head)) {
+                       /*
+                        * we don't have free cluster but have some clusters in
+                        * discarding, do discard now and reclaim them
+                        */
+                       swap_do_scheduled_discard(si);
+                       *scan_base = *offset = si->cluster_next;
+                       goto new_cluster;
+               } else
+                       return;
+       }
+
+       found_free = false;
+
+       /*
+        * Other CPUs can use our cluster if they can't find a free cluster,
+        * check if there is still free entry in the cluster
+        */
+       tmp = cluster->next;
+       while (tmp < si->max && tmp < (cluster_next(&cluster->index) + 1) *
+              SWAPFILE_CLUSTER) {
+               if (!si->swap_map[tmp]) {
+                       found_free = true;
+                       break;
+               }
+               tmp++;
+       }
+       if (!found_free) {
+               cluster_set_null(&cluster->index);
+               goto new_cluster;
+       }
+       cluster->next = tmp + 1;
+       *offset = tmp;
+       *scan_base = tmp;
 }
 
 static unsigned long scan_swap_map(struct swap_info_struct *si,
@@ -423,41 +488,17 @@ static unsigned long scan_swap_map(struct swap_info_struct *si,
        si->flags += SWP_SCANNING;
        scan_base = offset = si->cluster_next;
 
+       /* SSD algorithm */
+       if (si->cluster_info) {
+               scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
+               goto checks;
+       }
+
        if (unlikely(!si->cluster_nr--)) {
                if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
                        si->cluster_nr = SWAPFILE_CLUSTER - 1;
                        goto checks;
                }
-check_cluster:
-               if (!cluster_is_null(&si->free_cluster_head)) {
-                       offset = cluster_next(&si->free_cluster_head) *
-                                               SWAPFILE_CLUSTER;
-                       last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
-                       si->cluster_next = offset;
-                       si->cluster_nr = SWAPFILE_CLUSTER - 1;
-                       goto checks;
-               } else if (si->cluster_info) {
-                       /*
-                        * we don't have free cluster but have some clusters in
-                        * discarding, do discard now and reclaim them
-                        */
-                       if (!cluster_is_null(&si->discard_cluster_head)) {
-                               si->cluster_nr = 0;
-                               swap_do_scheduled_discard(si);
-                               scan_base = offset = si->cluster_next;
-                               if (!si->cluster_nr)
-                                       goto check_cluster;
-                               si->cluster_nr--;
-                               goto checks;
-                       }
-
-                       /*
-                        * Checking free cluster is fast enough, we can do the
-                        * check every time
-                        */
-                       si->cluster_nr = 0;
-                       goto checks;
-               }
 
                spin_unlock(&si->lock);
 
@@ -516,8 +557,10 @@ check_cluster:
        }
 
 checks:
-       if (scan_swap_map_recheck_cluster(si, offset))
-               goto check_cluster;
+       if (si->cluster_info) {
+               while (scan_swap_map_ssd_cluster_conflict(si, offset))
+                       scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
+       }
        if (!(si->flags & SWP_WRITEOK))
                goto no_page;
        if (!si->highest_bit)
@@ -1884,6 +1927,8 @@ SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
        spin_unlock(&swap_lock);
        frontswap_invalidate_area(type);
        mutex_unlock(&swapon_mutex);
+       free_percpu(p->percpu_cluster);
+       p->percpu_cluster = NULL;
        vfree(swap_map);
        vfree(cluster_info);
        vfree(frontswap_map);
@@ -2403,6 +2448,16 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
                        error = -ENOMEM;
                        goto bad_swap;
                }
+               p->percpu_cluster = alloc_percpu(struct percpu_cluster);
+               if (!p->percpu_cluster) {
+                       error = -ENOMEM;
+                       goto bad_swap;
+               }
+               for_each_possible_cpu(i) {
+                       struct percpu_cluster *cluster;
+                       cluster = per_cpu_ptr(p->percpu_cluster, i);
+                       cluster_set_null(&cluster->index);
+               }
        }
 
        error = swap_cgroup_swapon(p->type, maxpages);
@@ -2475,6 +2530,8 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
        error = 0;
        goto out;
 bad_swap:
+       free_percpu(p->percpu_cluster);
+       p->percpu_cluster = NULL;
        if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
                set_blocksize(p->bdev, p->old_block_size);
                blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);