aboutsummaryrefslogtreecommitdiffstats
path: root/fs/inode.c
diff options
context:
space:
mode:
authorEric Dumazet <eric.dumazet@gmail.com>2010-10-23 11:18:01 -0400
committerAl Viro <viro@zeniv.linux.org.uk>2010-10-25 21:26:11 -0400
commitf991bd2e14210fb93d722cb23e54991de20e8a3d (patch)
tree60643fb9d828c388d963ceb26504dce9890eabbb /fs/inode.c
parent7de9c6ee3ecffd99e1628e81a5ea5468f7581a1f (diff)
fs: introduce a per-cpu last_ino allocator
new_inode() dirties a contended cache line to get increasing inode numbers. This limits performance on workloads that cause significant parallel inode allocation. Solve this problem by using a per_cpu variable fed by the shared last_ino in batches of 1024 allocations. This reduces contention on the shared last_ino, and give same spreading ino numbers than before (i.e. same wraparound after 2^32 allocations). Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Diffstat (limited to 'fs/inode.c')
-rw-r--r--fs/inode.c45
1 files changed, 38 insertions, 7 deletions
diff --git a/fs/inode.c b/fs/inode.c
index 05ea293d5f3..46a3e120b19 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -717,6 +717,43 @@ repeat:
717 return NULL; 717 return NULL;
718} 718}
719 719
720/*
721 * Each cpu owns a range of LAST_INO_BATCH numbers.
722 * 'shared_last_ino' is dirtied only once out of LAST_INO_BATCH allocations,
723 * to renew the exhausted range.
724 *
725 * This does not significantly increase overflow rate because every CPU can
726 * consume at most LAST_INO_BATCH-1 unused inode numbers. So there is
727 * NR_CPUS*(LAST_INO_BATCH-1) wastage. At 4096 and 1024, this is ~0.1% of the
728 * 2^32 range, and is a worst-case. Even a 50% wastage would only increase
729 * overflow rate by 2x, which does not seem too significant.
730 *
731 * On a 32bit, non LFS stat() call, glibc will generate an EOVERFLOW
732 * error if st_ino won't fit in target struct field. Use 32bit counter
733 * here to attempt to avoid that.
734 */
735#define LAST_INO_BATCH 1024
736static DEFINE_PER_CPU(unsigned int, last_ino);
737
738static unsigned int get_next_ino(void)
739{
740 unsigned int *p = &get_cpu_var(last_ino);
741 unsigned int res = *p;
742
743#ifdef CONFIG_SMP
744 if (unlikely((res & (LAST_INO_BATCH-1)) == 0)) {
745 static atomic_t shared_last_ino;
746 int next = atomic_add_return(LAST_INO_BATCH, &shared_last_ino);
747
748 res = next - LAST_INO_BATCH;
749 }
750#endif
751
752 *p = ++res;
753 put_cpu_var(last_ino);
754 return res;
755}
756
720/** 757/**
721 * new_inode - obtain an inode 758 * new_inode - obtain an inode
722 * @sb: superblock 759 * @sb: superblock
@@ -731,12 +768,6 @@ repeat:
731 */ 768 */
732struct inode *new_inode(struct super_block *sb) 769struct inode *new_inode(struct super_block *sb)
733{ 770{
734 /*
735 * On a 32bit, non LFS stat() call, glibc will generate an EOVERFLOW
736 * error if st_ino won't fit in target struct field. Use 32bit counter
737 * here to attempt to avoid that.
738 */
739 static unsigned int last_ino;
740 struct inode *inode; 771 struct inode *inode;
741 772
742 spin_lock_prefetch(&inode_lock); 773 spin_lock_prefetch(&inode_lock);
@@ -745,7 +776,7 @@ struct inode *new_inode(struct super_block *sb)
745 if (inode) { 776 if (inode) {
746 spin_lock(&inode_lock); 777 spin_lock(&inode_lock);
747 __inode_sb_list_add(inode); 778 __inode_sb_list_add(inode);
748 inode->i_ino = ++last_ino; 779 inode->i_ino = get_next_ino();
749 inode->i_state = 0; 780 inode->i_state = 0;
750 spin_unlock(&inode_lock); 781 spin_unlock(&inode_lock);
751 } 782 }