From e16b3fcc610fab2dcf3381486b2640dc2a2213cb Mon Sep 17 00:00:00 2001 From: Alexander Motin Date: Fri, 18 Jul 2014 08:53:38 -0800 Subject: Illumos 5008 - lock contention (rrw_exit) while running a read only load 5008 lock contention (rrw_exit) while running a read only load Reviewed by: Matthew Ahrens Reviewed by: George Wilson Reviewed by: Alex Reece Reviewed by: Christopher Siden Reviewed by: Richard Yao Reviewed by: Saso Kiselkov Approved by: Garrett D'Amore Porting notes: This patch ported perfectly cleanly to ZoL. During testing 100% cached small-block reads, extreme contention was noticed on rrl->rr_lock from rrw_exit() due to the frequent entering and leaving ZPL. Illumos picked up this patch from FreeBSD and it also helps under Linux. On a 1-minute 4K cached read test with 10 fio processes pinned to a single socket on a 4-socket (10 thread per socket) NUMA system, contentions on rrl->rr_lock were reduced from 508799 to 43085. Ported-by: Tim Chase Signed-off-by: Brian Behlendorf Closes #3555 --- include/sys/rrwlock.h | 25 +++++++++++++++++++++++++ include/sys/zfs_vfsops.h | 2 +- include/sys/zfs_znode.h | 4 ++-- 3 files changed, 28 insertions(+), 3 deletions(-) (limited to 'include/sys') diff --git a/include/sys/rrwlock.h b/include/sys/rrwlock.h index d2bdff495..7a328fd68 100644 --- a/include/sys/rrwlock.h +++ b/include/sys/rrwlock.h @@ -83,6 +83,31 @@ void rrw_tsd_destroy(void *arg); #define RRW_LOCK_HELD(x) \ (rrw_held(x, RW_WRITER) || rrw_held(x, RW_READER)) +/* + * A reader-mostly lock implementation, tuning above reader-writer locks + * for hightly parallel read acquisitions, pessimizing write acquisitions. + * + * This should be a prime number. See comment in rrwlock.c near + * RRM_TD_LOCK() for details. + */ +#define RRM_NUM_LOCKS 17 +typedef struct rrmlock { + rrwlock_t locks[RRM_NUM_LOCKS]; +} rrmlock_t; + +void rrm_init(rrmlock_t *rrl, boolean_t track_all); +void rrm_destroy(rrmlock_t *rrl); +void rrm_enter(rrmlock_t *rrl, krw_t rw, void *tag); +void rrm_enter_read(rrmlock_t *rrl, void *tag); +void rrm_enter_write(rrmlock_t *rrl); +void rrm_exit(rrmlock_t *rrl, void *tag); +boolean_t rrm_held(rrmlock_t *rrl, krw_t rw); + +#define RRM_READ_HELD(x) rrm_held(x, RW_READER) +#define RRM_WRITE_HELD(x) rrm_held(x, RW_WRITER) +#define RRM_LOCK_HELD(x) \ + (rrm_held(x, RW_WRITER) || rrm_held(x, RW_READER)) + #ifdef __cplusplus } #endif diff --git a/include/sys/zfs_vfsops.h b/include/sys/zfs_vfsops.h index c4db2a911..4073d82f8 100644 --- a/include/sys/zfs_vfsops.h +++ b/include/sys/zfs_vfsops.h @@ -67,7 +67,7 @@ typedef struct zfs_sb { boolean_t z_atime; /* enable atimes mount option */ boolean_t z_relatime; /* enable relatime mount option */ boolean_t z_unmounted; /* unmounted */ - rrwlock_t z_teardown_lock; + rrmlock_t z_teardown_lock; krwlock_t z_teardown_inactive_lock; list_t z_all_znodes; /* all znodes in the fs */ uint64_t z_nr_znodes; /* number of znodes in the fs */ diff --git a/include/sys/zfs_znode.h b/include/sys/zfs_znode.h index 79ca4f7e9..7e73cf991 100644 --- a/include/sys/zfs_znode.h +++ b/include/sys/zfs_znode.h @@ -250,7 +250,7 @@ typedef struct znode { /* Called on entry to each ZFS vnode and vfs operation */ #define ZFS_ENTER(zsb) \ { \ - rrw_enter_read(&(zsb)->z_teardown_lock, FTAG); \ + rrm_enter_read(&(zsb)->z_teardown_lock, FTAG); \ if ((zsb)->z_unmounted) { \ ZFS_EXIT(zsb); \ return (EIO); \ @@ -260,7 +260,7 @@ typedef struct znode { /* Must be called before exiting the vop */ #define ZFS_EXIT(zsb) \ { \ - rrw_exit(&(zsb)->z_teardown_lock, FTAG); \ + rrm_exit(&(zsb)->z_teardown_lock, FTAG); \ } /* Verifies the znode is valid */ -- cgit v1.2.3