Revision b91e1302ad9b80c174a4855533f7e3aa2873355e authored by Linus Torvalds on 27 December 2016, 19:40:38 UTC, committed by Linus Torvalds on 29 December 2016, 19:03:15 UTC
In commit 62906027091f ("mm: add PageWaiters indicating tasks are
waiting for a page bit") Nick Piggin made our page locking no longer
unconditionally touch the hashed page waitqueue, which not only helps
performance in general, but is particularly helpful on NUMA machines
where the hashed wait queues can bounce around a lot.
However, the "clear lock bit atomically and then test the waiters bit"
sequence turns out to be much more expensive than it needs to be,
because you get a nasty stall when trying to access the same word that
just got updated atomically.
On architectures where locking is done with LL/SC, this would be trivial
to fix with a new primitive that clears one bit and tests another
atomically, but that ends up not working on x86, where the only atomic
operations that return the result end up being cmpxchg and xadd. The
atomic bit operations return the old value of the same bit we changed,
not the value of an unrelated bit.
On x86, we could put the lock bit in the high bit of the byte, and use
"xadd" with that bit (where the overflow ends up not touching other
bits), and look at the other bits of the result. However, an even
simpler model is to just use a regular atomic "and" to clear the lock
bit, and then the sign bit in eflags will indicate the resulting state
of the unrelated bit #7.
So by moving the PageWaiters bit up to bit #7, we can atomically clear
the lock bit and test the waiters bit on x86 too. And architectures
with LL/SC (which is all the usual RISC suspects), the particular bit
doesn't matter, so they are fine with this approach too.
This avoids the extra access to the same atomic word, and thus avoids
the costly stall at page unlock time.
The only downside is that the interface ends up being a bit odd and
specialized: clear a bit in a byte, and test the sign bit. Nick doesn't
love the resulting name of the new primitive, but I'd rather make the
name be descriptive and very clear about the limitation imposed by
trying to work across all relevant architectures than make it be some
generic thing that doesn't make the odd semantics explicit.
So this introduces the new architecture primitive
clear_bit_unlock_is_negative_byte();
and adds the trivial implementation for x86. We have a generic
non-optimized fallback (that just does a "clear_bit()"+"test_bit(7)"
combination) which can be overridden by any architecture that can do
better. According to Nick, Power has the same hickup x86 has, for
example, but some other architectures may not even care.
All these optimizations mean that my page locking stress-test (which is
just executing a lot of small short-lived shell scripts: "make test" in
the git source tree) no longer makes our page locking look horribly bad.
Before all these optimizations, just the unlock_page() costs were just
over 3% of all CPU overhead on "make test". After this, it's down to
0.66%, so just a quarter of the cost it used to be.
(The difference on NUMA is bigger, but there this micro-optimization is
likely less noticeable, since the big issue on NUMA was not the accesses
to 'struct page', but the waitqueue accesses that were already removed
by Nick's earlier commit).
Acked-by: Nick Piggin <npiggin@gmail.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Bob Peterson <rpeterso@redhat.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Lutomirski <luto@kernel.org>
Cc: Andreas Gruenbacher <agruenba@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
1 parent 2d706e7
server.c
/* AFS server record management
*
* Copyright (C) 2002, 2007 Red Hat, Inc. All Rights Reserved.
* Written by David Howells (dhowells@redhat.com)
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*/
#include <linux/sched.h>
#include <linux/slab.h>
#include "internal.h"
static unsigned afs_server_timeout = 10; /* server timeout in seconds */
static void afs_reap_server(struct work_struct *);
/* tree of all the servers, indexed by IP address */
static struct rb_root afs_servers = RB_ROOT;
static DEFINE_RWLOCK(afs_servers_lock);
/* LRU list of all the servers not currently in use */
static LIST_HEAD(afs_server_graveyard);
static DEFINE_SPINLOCK(afs_server_graveyard_lock);
static DECLARE_DELAYED_WORK(afs_server_reaper, afs_reap_server);
/*
* install a server record in the master tree
*/
static int afs_install_server(struct afs_server *server)
{
struct afs_server *xserver;
struct rb_node **pp, *p;
int ret;
_enter("%p", server);
write_lock(&afs_servers_lock);
ret = -EEXIST;
pp = &afs_servers.rb_node;
p = NULL;
while (*pp) {
p = *pp;
_debug("- consider %p", p);
xserver = rb_entry(p, struct afs_server, master_rb);
if (server->addr.s_addr < xserver->addr.s_addr)
pp = &(*pp)->rb_left;
else if (server->addr.s_addr > xserver->addr.s_addr)
pp = &(*pp)->rb_right;
else
goto error;
}
rb_link_node(&server->master_rb, p, pp);
rb_insert_color(&server->master_rb, &afs_servers);
ret = 0;
error:
write_unlock(&afs_servers_lock);
return ret;
}
/*
* allocate a new server record
*/
static struct afs_server *afs_alloc_server(struct afs_cell *cell,
const struct in_addr *addr)
{
struct afs_server *server;
_enter("");
server = kzalloc(sizeof(struct afs_server), GFP_KERNEL);
if (server) {
atomic_set(&server->usage, 1);
server->cell = cell;
INIT_LIST_HEAD(&server->link);
INIT_LIST_HEAD(&server->grave);
init_rwsem(&server->sem);
spin_lock_init(&server->fs_lock);
server->fs_vnodes = RB_ROOT;
server->cb_promises = RB_ROOT;
spin_lock_init(&server->cb_lock);
init_waitqueue_head(&server->cb_break_waitq);
INIT_DELAYED_WORK(&server->cb_break_work,
afs_dispatch_give_up_callbacks);
memcpy(&server->addr, addr, sizeof(struct in_addr));
server->addr.s_addr = addr->s_addr;
_leave(" = %p{%d}", server, atomic_read(&server->usage));
} else {
_leave(" = NULL [nomem]");
}
return server;
}
/*
* get an FS-server record for a cell
*/
struct afs_server *afs_lookup_server(struct afs_cell *cell,
const struct in_addr *addr)
{
struct afs_server *server, *candidate;
_enter("%p,%pI4", cell, &addr->s_addr);
/* quick scan of the list to see if we already have the server */
read_lock(&cell->servers_lock);
list_for_each_entry(server, &cell->servers, link) {
if (server->addr.s_addr == addr->s_addr)
goto found_server_quickly;
}
read_unlock(&cell->servers_lock);
candidate = afs_alloc_server(cell, addr);
if (!candidate) {
_leave(" = -ENOMEM");
return ERR_PTR(-ENOMEM);
}
write_lock(&cell->servers_lock);
/* check the cell's server list again */
list_for_each_entry(server, &cell->servers, link) {
if (server->addr.s_addr == addr->s_addr)
goto found_server;
}
_debug("new");
server = candidate;
if (afs_install_server(server) < 0)
goto server_in_two_cells;
afs_get_cell(cell);
list_add_tail(&server->link, &cell->servers);
write_unlock(&cell->servers_lock);
_leave(" = %p{%d}", server, atomic_read(&server->usage));
return server;
/* found a matching server quickly */
found_server_quickly:
_debug("found quickly");
afs_get_server(server);
read_unlock(&cell->servers_lock);
no_longer_unused:
if (!list_empty(&server->grave)) {
spin_lock(&afs_server_graveyard_lock);
list_del_init(&server->grave);
spin_unlock(&afs_server_graveyard_lock);
}
_leave(" = %p{%d}", server, atomic_read(&server->usage));
return server;
/* found a matching server on the second pass */
found_server:
_debug("found");
afs_get_server(server);
write_unlock(&cell->servers_lock);
kfree(candidate);
goto no_longer_unused;
/* found a server that seems to be in two cells */
server_in_two_cells:
write_unlock(&cell->servers_lock);
kfree(candidate);
printk(KERN_NOTICE "kAFS: Server %pI4 appears to be in two cells\n",
addr);
_leave(" = -EEXIST");
return ERR_PTR(-EEXIST);
}
/*
* look up a server by its IP address
*/
struct afs_server *afs_find_server(const struct sockaddr_rxrpc *srx)
{
struct afs_server *server = NULL;
struct rb_node *p;
struct in_addr addr = srx->transport.sin.sin_addr;
_enter("{%d,%pI4}", srx->transport.family, &addr.s_addr);
if (srx->transport.family != AF_INET) {
WARN(true, "AFS does not yes support non-IPv4 addresses\n");
return NULL;
}
read_lock(&afs_servers_lock);
p = afs_servers.rb_node;
while (p) {
server = rb_entry(p, struct afs_server, master_rb);
_debug("- consider %p", p);
if (addr.s_addr < server->addr.s_addr) {
p = p->rb_left;
} else if (addr.s_addr > server->addr.s_addr) {
p = p->rb_right;
} else {
afs_get_server(server);
goto found;
}
}
server = NULL;
found:
read_unlock(&afs_servers_lock);
ASSERTIFCMP(server, server->addr.s_addr, ==, addr.s_addr);
_leave(" = %p", server);
return server;
}
/*
* destroy a server record
* - removes from the cell list
*/
void afs_put_server(struct afs_server *server)
{
if (!server)
return;
_enter("%p{%d}", server, atomic_read(&server->usage));
_debug("PUT SERVER %d", atomic_read(&server->usage));
ASSERTCMP(atomic_read(&server->usage), >, 0);
if (likely(!atomic_dec_and_test(&server->usage))) {
_leave("");
return;
}
afs_flush_callback_breaks(server);
spin_lock(&afs_server_graveyard_lock);
if (atomic_read(&server->usage) == 0) {
list_move_tail(&server->grave, &afs_server_graveyard);
server->time_of_death = get_seconds();
queue_delayed_work(afs_wq, &afs_server_reaper,
afs_server_timeout * HZ);
}
spin_unlock(&afs_server_graveyard_lock);
_leave(" [dead]");
}
/*
* destroy a dead server
*/
static void afs_destroy_server(struct afs_server *server)
{
_enter("%p", server);
ASSERTIF(server->cb_break_head != server->cb_break_tail,
delayed_work_pending(&server->cb_break_work));
ASSERTCMP(server->fs_vnodes.rb_node, ==, NULL);
ASSERTCMP(server->cb_promises.rb_node, ==, NULL);
ASSERTCMP(server->cb_break_head, ==, server->cb_break_tail);
ASSERTCMP(atomic_read(&server->cb_break_n), ==, 0);
afs_put_cell(server->cell);
kfree(server);
}
/*
* reap dead server records
*/
static void afs_reap_server(struct work_struct *work)
{
LIST_HEAD(corpses);
struct afs_server *server;
unsigned long delay, expiry;
time_t now;
now = get_seconds();
spin_lock(&afs_server_graveyard_lock);
while (!list_empty(&afs_server_graveyard)) {
server = list_entry(afs_server_graveyard.next,
struct afs_server, grave);
/* the queue is ordered most dead first */
expiry = server->time_of_death + afs_server_timeout;
if (expiry > now) {
delay = (expiry - now) * HZ;
mod_delayed_work(afs_wq, &afs_server_reaper, delay);
break;
}
write_lock(&server->cell->servers_lock);
write_lock(&afs_servers_lock);
if (atomic_read(&server->usage) > 0) {
list_del_init(&server->grave);
} else {
list_move_tail(&server->grave, &corpses);
list_del_init(&server->link);
rb_erase(&server->master_rb, &afs_servers);
}
write_unlock(&afs_servers_lock);
write_unlock(&server->cell->servers_lock);
}
spin_unlock(&afs_server_graveyard_lock);
/* now reap the corpses we've extracted */
while (!list_empty(&corpses)) {
server = list_entry(corpses.next, struct afs_server, grave);
list_del(&server->grave);
afs_destroy_server(server);
}
}
/*
* discard all the server records for rmmod
*/
void __exit afs_purge_servers(void)
{
afs_server_timeout = 0;
mod_delayed_work(afs_wq, &afs_server_reaper, 0);
}
Computing file changes ...
