mirror of
https://git.proxmox.com/git/mirror_zfs.git
synced 2024-12-25 18:59:33 +03:00
e16b3fcc61
5008 lock contention (rrw_exit) while running a read only load Reviewed by: Matthew Ahrens <matthew.ahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Alex Reece <alex.reece@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Richard Yao <ryao@gentoo.org> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Approved by: Garrett D'Amore <garrett@damore.org> 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 <tim@chase2k.com> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #3555
396 lines
11 KiB
C
396 lines
11 KiB
C
/*
|
|
* CDDL HEADER START
|
|
*
|
|
* The contents of this file are subject to the terms of the
|
|
* Common Development and Distribution License (the "License").
|
|
* You may not use this file except in compliance with the License.
|
|
*
|
|
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
|
|
* or http://www.opensolaris.org/os/licensing.
|
|
* See the License for the specific language governing permissions
|
|
* and limitations under the License.
|
|
*
|
|
* When distributing Covered Code, include this CDDL HEADER in each
|
|
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
|
|
* If applicable, add the following below this CDDL HEADER, with the
|
|
* fields enclosed by brackets "[]" replaced with your own identifying
|
|
* information: Portions Copyright [yyyy] [name of copyright owner]
|
|
*
|
|
* CDDL HEADER END
|
|
*/
|
|
/*
|
|
* Copyright 2009 Sun Microsystems, Inc. All rights reserved.
|
|
* Use is subject to license terms.
|
|
*/
|
|
/*
|
|
* Copyright (c) 2012 by Delphix. All rights reserved.
|
|
*/
|
|
|
|
#include <sys/refcount.h>
|
|
#include <sys/rrwlock.h>
|
|
|
|
/*
|
|
* This file contains the implementation of a re-entrant read
|
|
* reader/writer lock (aka "rrwlock").
|
|
*
|
|
* This is a normal reader/writer lock with the additional feature
|
|
* of allowing threads who have already obtained a read lock to
|
|
* re-enter another read lock (re-entrant read) - even if there are
|
|
* waiting writers.
|
|
*
|
|
* Callers who have not obtained a read lock give waiting writers priority.
|
|
*
|
|
* The rrwlock_t lock does not allow re-entrant writers, nor does it
|
|
* allow a re-entrant mix of reads and writes (that is, it does not
|
|
* allow a caller who has already obtained a read lock to be able to
|
|
* then grab a write lock without first dropping all read locks, and
|
|
* vice versa).
|
|
*
|
|
* The rrwlock_t uses tsd (thread specific data) to keep a list of
|
|
* nodes (rrw_node_t), where each node keeps track of which specific
|
|
* lock (rrw_node_t::rn_rrl) the thread has grabbed. Since re-entering
|
|
* should be rare, a thread that grabs multiple reads on the same rrwlock_t
|
|
* will store multiple rrw_node_ts of the same 'rrn_rrl'. Nodes on the
|
|
* tsd list can represent a different rrwlock_t. This allows a thread
|
|
* to enter multiple and unique rrwlock_ts for read locks at the same time.
|
|
*
|
|
* Since using tsd exposes some overhead, the rrwlock_t only needs to
|
|
* keep tsd data when writers are waiting. If no writers are waiting, then
|
|
* a reader just bumps the anonymous read count (rr_anon_rcount) - no tsd
|
|
* is needed. Once a writer attempts to grab the lock, readers then
|
|
* keep tsd data and bump the linked readers count (rr_linked_rcount).
|
|
*
|
|
* If there are waiting writers and there are anonymous readers, then a
|
|
* reader doesn't know if it is a re-entrant lock. But since it may be one,
|
|
* we allow the read to proceed (otherwise it could deadlock). Since once
|
|
* waiting writers are active, readers no longer bump the anonymous count,
|
|
* the anonymous readers will eventually flush themselves out. At this point,
|
|
* readers will be able to tell if they are a re-entrant lock (have a
|
|
* rrw_node_t entry for the lock) or not. If they are a re-entrant lock, then
|
|
* we must let the proceed. If they are not, then the reader blocks for the
|
|
* waiting writers. Hence, we do not starve writers.
|
|
*/
|
|
|
|
/* global key for TSD */
|
|
uint_t rrw_tsd_key;
|
|
|
|
typedef struct rrw_node {
|
|
struct rrw_node *rn_next;
|
|
rrwlock_t *rn_rrl;
|
|
void *rn_tag;
|
|
} rrw_node_t;
|
|
|
|
static rrw_node_t *
|
|
rrn_find(rrwlock_t *rrl)
|
|
{
|
|
rrw_node_t *rn;
|
|
|
|
if (refcount_count(&rrl->rr_linked_rcount) == 0)
|
|
return (NULL);
|
|
|
|
for (rn = tsd_get(rrw_tsd_key); rn != NULL; rn = rn->rn_next) {
|
|
if (rn->rn_rrl == rrl)
|
|
return (rn);
|
|
}
|
|
return (NULL);
|
|
}
|
|
|
|
/*
|
|
* Add a node to the head of the singly linked list.
|
|
*/
|
|
static void
|
|
rrn_add(rrwlock_t *rrl, void *tag)
|
|
{
|
|
rrw_node_t *rn;
|
|
|
|
rn = kmem_alloc(sizeof (*rn), KM_SLEEP);
|
|
rn->rn_rrl = rrl;
|
|
rn->rn_next = tsd_get(rrw_tsd_key);
|
|
rn->rn_tag = tag;
|
|
VERIFY(tsd_set(rrw_tsd_key, rn) == 0);
|
|
}
|
|
|
|
/*
|
|
* If a node is found for 'rrl', then remove the node from this
|
|
* thread's list and return TRUE; otherwise return FALSE.
|
|
*/
|
|
static boolean_t
|
|
rrn_find_and_remove(rrwlock_t *rrl, void *tag)
|
|
{
|
|
rrw_node_t *rn;
|
|
rrw_node_t *prev = NULL;
|
|
|
|
if (refcount_count(&rrl->rr_linked_rcount) == 0)
|
|
return (B_FALSE);
|
|
|
|
for (rn = tsd_get(rrw_tsd_key); rn != NULL; rn = rn->rn_next) {
|
|
if (rn->rn_rrl == rrl && rn->rn_tag == tag) {
|
|
if (prev)
|
|
prev->rn_next = rn->rn_next;
|
|
else
|
|
VERIFY(tsd_set(rrw_tsd_key, rn->rn_next) == 0);
|
|
kmem_free(rn, sizeof (*rn));
|
|
return (B_TRUE);
|
|
}
|
|
prev = rn;
|
|
}
|
|
return (B_FALSE);
|
|
}
|
|
|
|
void
|
|
rrw_init(rrwlock_t *rrl, boolean_t track_all)
|
|
{
|
|
mutex_init(&rrl->rr_lock, NULL, MUTEX_DEFAULT, NULL);
|
|
cv_init(&rrl->rr_cv, NULL, CV_DEFAULT, NULL);
|
|
rrl->rr_writer = NULL;
|
|
refcount_create(&rrl->rr_anon_rcount);
|
|
refcount_create(&rrl->rr_linked_rcount);
|
|
rrl->rr_writer_wanted = B_FALSE;
|
|
rrl->rr_track_all = track_all;
|
|
}
|
|
|
|
void
|
|
rrw_destroy(rrwlock_t *rrl)
|
|
{
|
|
mutex_destroy(&rrl->rr_lock);
|
|
cv_destroy(&rrl->rr_cv);
|
|
ASSERT(rrl->rr_writer == NULL);
|
|
refcount_destroy(&rrl->rr_anon_rcount);
|
|
refcount_destroy(&rrl->rr_linked_rcount);
|
|
}
|
|
|
|
static void
|
|
rrw_enter_read_impl(rrwlock_t *rrl, boolean_t prio, void *tag)
|
|
{
|
|
mutex_enter(&rrl->rr_lock);
|
|
#if !defined(DEBUG) && defined(_KERNEL)
|
|
if (rrl->rr_writer == NULL && !rrl->rr_writer_wanted &&
|
|
!rrl->rr_track_all) {
|
|
rrl->rr_anon_rcount.rc_count++;
|
|
mutex_exit(&rrl->rr_lock);
|
|
return;
|
|
}
|
|
DTRACE_PROBE(zfs__rrwfastpath__rdmiss);
|
|
#endif
|
|
ASSERT(rrl->rr_writer != curthread);
|
|
ASSERT(refcount_count(&rrl->rr_anon_rcount) >= 0);
|
|
|
|
while (rrl->rr_writer != NULL || (rrl->rr_writer_wanted &&
|
|
refcount_is_zero(&rrl->rr_anon_rcount) && !prio &&
|
|
rrn_find(rrl) == NULL))
|
|
cv_wait(&rrl->rr_cv, &rrl->rr_lock);
|
|
|
|
if (rrl->rr_writer_wanted || rrl->rr_track_all) {
|
|
/* may or may not be a re-entrant enter */
|
|
rrn_add(rrl, tag);
|
|
(void) refcount_add(&rrl->rr_linked_rcount, tag);
|
|
} else {
|
|
(void) refcount_add(&rrl->rr_anon_rcount, tag);
|
|
}
|
|
ASSERT(rrl->rr_writer == NULL);
|
|
mutex_exit(&rrl->rr_lock);
|
|
}
|
|
|
|
void
|
|
rrw_enter_read(rrwlock_t *rrl, void *tag)
|
|
{
|
|
rrw_enter_read_impl(rrl, B_FALSE, tag);
|
|
}
|
|
|
|
/*
|
|
* take a read lock even if there are pending write lock requests. if we want
|
|
* to take a lock reentrantly, but from different threads (that have a
|
|
* relationship to each other), the normal detection mechanism to overrule
|
|
* the pending writer does not work, so we have to give an explicit hint here.
|
|
*/
|
|
void
|
|
rrw_enter_read_prio(rrwlock_t *rrl, void *tag)
|
|
{
|
|
rrw_enter_read_impl(rrl, B_TRUE, tag);
|
|
}
|
|
|
|
|
|
void
|
|
rrw_enter_write(rrwlock_t *rrl)
|
|
{
|
|
mutex_enter(&rrl->rr_lock);
|
|
ASSERT(rrl->rr_writer != curthread);
|
|
|
|
while (refcount_count(&rrl->rr_anon_rcount) > 0 ||
|
|
refcount_count(&rrl->rr_linked_rcount) > 0 ||
|
|
rrl->rr_writer != NULL) {
|
|
rrl->rr_writer_wanted = B_TRUE;
|
|
cv_wait(&rrl->rr_cv, &rrl->rr_lock);
|
|
}
|
|
rrl->rr_writer_wanted = B_FALSE;
|
|
rrl->rr_writer = curthread;
|
|
mutex_exit(&rrl->rr_lock);
|
|
}
|
|
|
|
void
|
|
rrw_enter(rrwlock_t *rrl, krw_t rw, void *tag)
|
|
{
|
|
if (rw == RW_READER)
|
|
rrw_enter_read(rrl, tag);
|
|
else
|
|
rrw_enter_write(rrl);
|
|
}
|
|
|
|
void
|
|
rrw_exit(rrwlock_t *rrl, void *tag)
|
|
{
|
|
mutex_enter(&rrl->rr_lock);
|
|
#if !defined(DEBUG) && defined(_KERNEL)
|
|
if (!rrl->rr_writer && rrl->rr_linked_rcount.rc_count == 0) {
|
|
rrl->rr_anon_rcount.rc_count--;
|
|
if (rrl->rr_anon_rcount.rc_count == 0)
|
|
cv_broadcast(&rrl->rr_cv);
|
|
mutex_exit(&rrl->rr_lock);
|
|
return;
|
|
}
|
|
DTRACE_PROBE(zfs__rrwfastpath__exitmiss);
|
|
#endif
|
|
ASSERT(!refcount_is_zero(&rrl->rr_anon_rcount) ||
|
|
!refcount_is_zero(&rrl->rr_linked_rcount) ||
|
|
rrl->rr_writer != NULL);
|
|
|
|
if (rrl->rr_writer == NULL) {
|
|
int64_t count;
|
|
if (rrn_find_and_remove(rrl, tag)) {
|
|
count = refcount_remove(&rrl->rr_linked_rcount, tag);
|
|
} else {
|
|
ASSERT(!rrl->rr_track_all);
|
|
count = refcount_remove(&rrl->rr_anon_rcount, tag);
|
|
}
|
|
if (count == 0)
|
|
cv_broadcast(&rrl->rr_cv);
|
|
} else {
|
|
ASSERT(rrl->rr_writer == curthread);
|
|
ASSERT(refcount_is_zero(&rrl->rr_anon_rcount) &&
|
|
refcount_is_zero(&rrl->rr_linked_rcount));
|
|
rrl->rr_writer = NULL;
|
|
cv_broadcast(&rrl->rr_cv);
|
|
}
|
|
mutex_exit(&rrl->rr_lock);
|
|
}
|
|
|
|
/*
|
|
* If the lock was created with track_all, rrw_held(RW_READER) will return
|
|
* B_TRUE iff the current thread has the lock for reader. Otherwise it may
|
|
* return B_TRUE if any thread has the lock for reader.
|
|
*/
|
|
boolean_t
|
|
rrw_held(rrwlock_t *rrl, krw_t rw)
|
|
{
|
|
boolean_t held;
|
|
|
|
mutex_enter(&rrl->rr_lock);
|
|
if (rw == RW_WRITER) {
|
|
held = (rrl->rr_writer == curthread);
|
|
} else {
|
|
held = (!refcount_is_zero(&rrl->rr_anon_rcount) ||
|
|
rrn_find(rrl) != NULL);
|
|
}
|
|
mutex_exit(&rrl->rr_lock);
|
|
|
|
return (held);
|
|
}
|
|
|
|
void
|
|
rrw_tsd_destroy(void *arg)
|
|
{
|
|
rrw_node_t *rn = arg;
|
|
if (rn != NULL) {
|
|
panic("thread %p terminating with rrw lock %p held",
|
|
(void *)curthread, (void *)rn->rn_rrl);
|
|
}
|
|
}
|
|
|
|
/*
|
|
* A reader-mostly lock implementation, tuning above reader-writer locks
|
|
* for hightly parallel read acquisitions, while pessimizing writes.
|
|
*
|
|
* The idea is to split single busy lock into array of locks, so that
|
|
* each reader can lock only one of them for read, depending on result
|
|
* of simple hash function. That proportionally reduces lock congestion.
|
|
* Writer same time has to sequentially aquire write on all the locks.
|
|
* That makes write aquisition proportionally slower, but in places where
|
|
* it is used (filesystem unmount) performance is not critical.
|
|
*
|
|
* All the functions below are direct wrappers around functions above.
|
|
*/
|
|
void
|
|
rrm_init(rrmlock_t *rrl, boolean_t track_all)
|
|
{
|
|
int i;
|
|
|
|
for (i = 0; i < RRM_NUM_LOCKS; i++)
|
|
rrw_init(&rrl->locks[i], track_all);
|
|
}
|
|
|
|
void
|
|
rrm_destroy(rrmlock_t *rrl)
|
|
{
|
|
int i;
|
|
|
|
for (i = 0; i < RRM_NUM_LOCKS; i++)
|
|
rrw_destroy(&rrl->locks[i]);
|
|
}
|
|
|
|
void
|
|
rrm_enter(rrmlock_t *rrl, krw_t rw, void *tag)
|
|
{
|
|
if (rw == RW_READER)
|
|
rrm_enter_read(rrl, tag);
|
|
else
|
|
rrm_enter_write(rrl);
|
|
}
|
|
|
|
/*
|
|
* This maps the current thread to a specific lock. Note that the lock
|
|
* must be released by the same thread that acquired it. We do this
|
|
* mapping by taking the thread pointer mod a prime number. We examine
|
|
* only the low 32 bits of the thread pointer, because 32-bit division
|
|
* is faster than 64-bit division, and the high 32 bits have little
|
|
* entropy anyway.
|
|
*/
|
|
#define RRM_TD_LOCK() (((uint32_t)(uintptr_t)(curthread)) % RRM_NUM_LOCKS)
|
|
|
|
void
|
|
rrm_enter_read(rrmlock_t *rrl, void *tag)
|
|
{
|
|
rrw_enter_read(&rrl->locks[RRM_TD_LOCK()], tag);
|
|
}
|
|
|
|
void
|
|
rrm_enter_write(rrmlock_t *rrl)
|
|
{
|
|
int i;
|
|
|
|
for (i = 0; i < RRM_NUM_LOCKS; i++)
|
|
rrw_enter_write(&rrl->locks[i]);
|
|
}
|
|
|
|
void
|
|
rrm_exit(rrmlock_t *rrl, void *tag)
|
|
{
|
|
int i;
|
|
|
|
if (rrl->locks[0].rr_writer == curthread) {
|
|
for (i = 0; i < RRM_NUM_LOCKS; i++)
|
|
rrw_exit(&rrl->locks[i], tag);
|
|
} else {
|
|
rrw_exit(&rrl->locks[RRM_TD_LOCK()], tag);
|
|
}
|
|
}
|
|
|
|
boolean_t
|
|
rrm_held(rrmlock_t *rrl, krw_t rw)
|
|
{
|
|
if (rw == RW_WRITER) {
|
|
return (rrw_held(&rrl->locks[0], rw));
|
|
} else {
|
|
return (rrw_held(&rrl->locks[RRM_TD_LOCK()], rw));
|
|
}
|
|
}
|