xfs: Don't allocate new buffers on every call to _xfs_buf_find

Stats show that for an 8-way unlink @ ~80,000 unlinks/s we are doing
~1 million cache hit lookups to ~3000 buffer creates. That's almost
3 orders of magnitude more cahce hits than misses, so optimising for
cache hits is quite important. In the cache hit case, we do not need
to allocate a new buffer in case of a cache miss, so we are
effectively hitting the allocator for no good reason for vast the
majority of calls to _xfs_buf_find. 8-way create workloads are
showing similar cache hit/miss ratios.

The result is profiles that look like this:

     samples  pcnt function                        DSO
     _______ _____ _______________________________ _________________

     1036.00 10.0% _xfs_buf_find                   [kernel.kallsyms]
      582.00  5.6% kmem_cache_alloc                [kernel.kallsyms]
      519.00  5.0% __memcpy                        [kernel.kallsyms]
      468.00  4.5% __ticket_spin_lock              [kernel.kallsyms]
      388.00  3.7% kmem_cache_free                 [kernel.kallsyms]
      331.00  3.2% xfs_log_commit_cil              [kernel.kallsyms]


Further, there is a fair bit of work involved in initialising a new
buffer once a cache miss has occurred and we currently do that under
the rbtree spinlock. That increases spinlock hold time on what are
heavily used trees.

To fix this, remove the initialisation of the buffer from
_xfs_buf_find() and only allocate the new buffer once we've had a
cache miss. Initialise the buffer immediately after allocating it in
xfs_buf_get, too, so that is it ready for insert if we get another
cache miss after allocation. This minimises lock hold time and
avoids unnecessary allocator churn. The resulting profiles look
like:

     samples  pcnt function                    DSO
     _______ _____ ___________________________ _________________

     8111.00  9.1% _xfs_buf_find               [kernel.kallsyms]
     4380.00  4.9% __memcpy                    [kernel.kallsyms]
     4341.00  4.8% __ticket_spin_lock          [kernel.kallsyms]
     3401.00  3.8% kmem_cache_alloc            [kernel.kallsyms]
     2856.00  3.2% xfs_log_commit_cil          [kernel.kallsyms]
     2625.00  2.9% __kmalloc                   [kernel.kallsyms]
     2380.00  2.7% kfree                       [kernel.kallsyms]
     2016.00  2.3% kmem_cache_free             [kernel.kallsyms]

Showing a significant reduction in time spent doing allocation and
freeing from slabs (kmem_cache_alloc and kmem_cache_free).

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>

1 files changed, 28 insertions, 20 deletions

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index e3af85095ddb..6785b7bd952d 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -477,8 +477,6 @@ _xfs_buf_find(
 
         /* No match found */
         if (new_bp) {
-                _xfs_buf_initialize(new_bp, btp, range_base,
-                                range_length, flags);
                 rb_link_node(&new_bp->b_rbnode, parent, rbp);
                 rb_insert_color(&new_bp->b_rbnode, &pag->pag_buf_tree);
                 /* the buffer keeps the perag reference until it is freed */
@@ -521,35 +519,53 @@ found:
 }
 
 /*
- *      Assembles a buffer covering the specified range.
- *      Storage in memory for all portions of the buffer will be allocated,
- *      although backing storage may not be.
+ * Assembles a buffer covering the specified range. The code is optimised for
+ * cache hits, as metadata intensive workloads will see 3 orders of magnitude
+ * more hits than misses.
  */
-xfs_buf_t *
+struct xfs_buf *
 xfs_buf_get(
         xfs_buftarg_t           *target,/* target for buffer            */
         xfs_off_t               ioff,   /* starting offset of range     */
         size_t                  isize,  /* length of range              */
         xfs_buf_flags_t         flags)
 {
-        xfs_buf_t               *bp, *new_bp;
+        struct xfs_buf          *bp;
+        struct xfs_buf          *new_bp;
         int                     error = 0;
 
+        bp = _xfs_buf_find(target, ioff, isize, flags, NULL);
+        if (likely(bp))
+                goto found;
+
         new_bp = xfs_buf_allocate(flags);
         if (unlikely(!new_bp))
                 return NULL;
 
+        _xfs_buf_initialize(new_bp, target,
+                            ioff << BBSHIFT, isize << BBSHIFT, flags);
+
         bp = _xfs_buf_find(target, ioff, isize, flags, new_bp);
+        if (!bp) {
+                xfs_buf_deallocate(new_bp);
+                return NULL;
+        }
+
         if (bp == new_bp) {
                 error = xfs_buf_allocate_memory(bp, flags);
                 if (error)
                         goto no_buffer;
-        } else {
+        } else
                 xfs_buf_deallocate(new_bp);
-                if (unlikely(bp == NULL))
-                        return NULL;
-        }
 
+        /*
+         * Now we have a workable buffer, fill in the block number so
+         * that we can do IO on it.
+         */
+        bp->b_bn = ioff;
+        bp->b_count_desired = bp->b_buffer_length;
+
+found:
         if (!(bp->b_flags & XBF_MAPPED)) {
                 error = _xfs_buf_map_pages(bp, flags);
                 if (unlikely(error)) {
@@ -560,18 +576,10 @@ xfs_buf_get(
         }
 
         XFS_STATS_INC(xb_get);
-
-        /*
-         * Always fill in the block number now, the mapped cases can do
-         * their own overlay of this later.
-         */
-        bp->b_bn = ioff;
-        bp->b_count_desired = bp->b_buffer_length;
-
         trace_xfs_buf_get(bp, flags, _RET_IP_);
         return bp;
 
- no_buffer:
+no_buffer:
         if (flags & (XBF_LOCK | XBF_TRYLOCK))
                 xfs_buf_unlock(bp);
         xfs_buf_rele(bp);