Message-ID: <49419670.2060107@cosmosbay.com>
Date:	Thu, 11 Dec 2008 23:38:40 +0100
From:	Eric Dumazet <dada1@...mosbay.com>
To:	Andrew Morton <akpm@...ux-foundation.org>
CC:	Ingo Molnar <mingo@...e.hu>, Christoph Hellwig <hch@...radead.org>,
	David Miller <davem@...emloft.net>,
	"Rafael J. Wysocki" <rjw@...k.pl>, linux-kernel@...r.kernel.org,
	"kernel-testers@...r.kernel.org >> Kernel Testers List" 
	<kernel-testers@...r.kernel.org>, Mike Galbraith <efault@....de>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	Linux Netdev List <netdev@...r.kernel.org>,
	Christoph Lameter <cl@...ux-foundation.org>,
	linux-fsdevel@...r.kernel.org, Al Viro <viro@...IV.linux.org.uk>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
Subject: [PATCH v3 0/7] fs: Scalability of sockets/pipes allocation/deallocation
 on SMP

Hi Andrew

Take v2 of this patch serie got no new feedback, maybe its time for mm
inclusion for a while ?

In this third version I added last two patches, one intialy from Christoph
Lameter, and one to avoid dirtying mnt->mnt_count on hardwired fs.

Many thanks to Christoph and Paul for this SLAB_DESTROY_PER_RCU work done
on "struct file".

Thank you

Short summary : Nice speedups for allocation/deallocation of sockets/pipes
(From 27.5 seconds to 1.62 s, on a 8 cpus machine)

Long version :

To allocate a socket or a pipe we :

0) Do the usual file table manipulation (pretty scalable these days,
but would be faster if 'struct file' were using SLAB_DESTROY_BY_RCU
and avoid call_rcu() cache killer). This point is addressed by 6th
patch.

1) allocate an inode with new_inode()
This function :
- locks inode_lock,
- dirties nr_inodes counter
- dirties inode_in_use list  (for sockets/pipes, this is useless)
- dirties superblock s_inodes.  - dirties last_ino counter
All these are in different cache lines unfortunatly.

2) allocate a dentry
d_alloc() takes dcache_lock,
insert dentry on its parent list (dirtying sock_mnt->mnt_sb->s_root)
dirties nr_dentry

3) d_instantiate() dentry  (dcache_lock taken again)

4) init_file() -> atomic_inc() on sock_mnt->refcount


At close() time, we must undo the things. Its even more expensive because
of the _atomic_dec_and_lock() that stress a lot, and because of two cache
lines that are touched when an element is deleted from a list
(previous and next items)

This is really bad, since sockets/pipes dont need to be visible in dcache
or an inode list per super block.

This patch series get rid of all but one contended cache lines for
sockets, pipes and anonymous fd  (signalfd, timerfd, ...)

socketallocbench is a very simple program (attached to this mail) that makes
a loop :

for (i = 0; i < 1000000; i++)
    close(socket(AF_INET, SOCK_STREAM, 0));

Cost if one cpu runs the program :

real    1.561s
user    0.092s
sys     1.469s

Cost if 8 processes are launched on a 8 CPU machine
(socketallocbench -n 8) :

real    27.496s   <<<< !!!! >>>>
user    0.657s
sys     3m39.092s

Oprofile results (for the 8 process run, 3 times):

CPU: Core 2, speed 3000.03 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit
mask of 0x00 (Unhalted core cycles) count 100000
samples  cum. samples  %        cum. %     symbol name
3347352  3347352       28.0232  28.0232    _atomic_dec_and_lock
3301428  6648780       27.6388  55.6620    d_instantiate
2971130  9619910       24.8736  80.5355    d_alloc
241318   9861228        2.0203  82.5558    init_file
146190   10007418       1.2239  83.7797    __slab_free
144149   10151567       1.2068  84.9864    inotify_d_instantiate
143971   10295538       1.2053  86.1917    inet_create
137168   10432706       1.1483  87.3401    new_inode
117549   10550255       0.9841  88.3242    add_partial
110795   10661050       0.9275  89.2517    generic_drop_inode
107137   10768187       0.8969  90.1486    kmem_cache_alloc
94029    10862216       0.7872  90.9358    tcp_close
82837    10945053       0.6935  91.6293    dput
67486    11012539       0.5650  92.1943    dentry_iput
57751    11070290       0.4835  92.6778    iput
54327    11124617       0.4548  93.1326    tcp_v4_init_sock
49921    11174538       0.4179  93.5505    sysenter_past_esp
47616    11222154       0.3986  93.9491    kmem_cache_free
30792    11252946       0.2578  94.2069    clear_inode
27540    11280486       0.2306  94.4375    copy_from_user
26509    11306995       0.2219  94.6594    init_timer
26363    11333358       0.2207  94.8801    discard_slab
25284    11358642       0.2117  95.0918    __fput
22482    11381124       0.1882  95.2800    __percpu_counter_add
20369    11401493       0.1705  95.4505    sock_alloc
18501    11419994       0.1549  95.6054    inet_csk_destroy_sock
17923    11437917       0.1500  95.7555    sys_close


This patch serie avoids all contented cache lines and makes this "bench"
pretty fast.


New cost if run on one cpu :

real    1.245s (instead of 1.561s)
user    0.074s
sys     1.161s


If run on 8 CPUS :

real    1.624s
user    0.580s
sys     12.296s


On oprofile, we finally can see network stuff coming at the front of
expensive stuff. (with the exception of kmem_cache_[z]alloc(), because
it has to clear 192 bytes of file structures, this takes half of the time)

CPU: Core 2, speed 3000.09 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a unit mask of 0x00 (Unhalted core cycles) count 100
000
samples  cum. samples  %        cum. %     symbol name
176586   176586        10.9376  10.9376    kmem_cache_alloc
169838   346424        10.5196  21.4572    tcp_close
105331   451755         6.5241  27.9813    tcp_v4_init_sock
105146   556901         6.5126  34.4939    tcp_v4_destroy_sock
83307    640208         5.1600  39.6539    sysenter_past_esp
80241    720449         4.9701  44.6239    inet_csk_destroy_sock
74263    794712         4.5998  49.2237    kmem_cache_free
56806    851518         3.5185  52.7422    __percpu_counter_add
48619    900137         3.0114  55.7536    copy_from_user
44803    944940         2.7751  58.5287    init_timer
28539    973479         1.7677  60.2964    d_alloc
27795    1001274        1.7216  62.0180    alloc_fd
26747    1028021        1.6567  63.6747    __fput
24312    1052333        1.5059  65.1805    sys_close
24205    1076538        1.4992  66.6798    inet_create
22409    1098947        1.3880  68.0677    alloc_inode
21359    1120306        1.3230  69.3907    release_sock
19865    1140171        1.2304  70.6211    fd_install
19472    1159643        1.2061  71.8272    lock_sock_nested
18956    1178599        1.1741  73.0013    sock_init_data
17301    1195900        1.0716  74.0729    drop_file_write_access
17113    1213013        1.0600  75.1329    inotify_d_instantiate
16384    1229397        1.0148  76.1477    dput
15173    1244570        0.9398  77.0875    local_bh_enable_ip
15017    1259587        0.9301  78.0176    local_bh_enable
13354    1272941        0.8271  78.8448    __sock_create
13139    1286080        0.8138  79.6586    inet_release
13062    1299142        0.8090  80.4676    sysenter_do_call
11935    1311077        0.7392  81.2069    iput_single


This patch serie contains 7 patches, against linux-2.6 tree,
plus one patch in mm (fs: filp_cachep can be static in fs/file_table.c)

[PATCH 1/7] fs: Use a percpu_counter to track nr_dentry

Adding a percpu_counter nr_dentry avoids cache line ping pongs
between cpus to maintain this metric, and dcache_lock is
no more needed to protect dentry_stat.nr_dentry

We centralize nr_dentry updates at the right place :
- increments in d_alloc()
- decrements in d_free()

d_alloc() can avoid taking dcache_lock if parent is NULL

("socketallocbench -n 8" bench result : 27.5s to 25s)

[PATCH 2/7] fs: Use a percpu_counter to track nr_inodes

Avoids cache line ping pongs between cpus and prepare next patch,
because updates of nr_inodes dont need inode_lock anymore.

("socketallocbench -n 8" bench result : no difference at this point)

[PATCH 3/7] fs: Introduce a per_cpu last_ino allocator

new_inode() dirties a contended cache line to get increasing
inode numbers.

Solve this problem by providing to each cpu a per_cpu variable,
feeded by the shared last_ino, but once every 1024 allocations.

This reduce contention on the shared last_ino, and give same
spreading ino numbers than before.
(same wraparound after 232 allocations)

("socketallocbench -n 8" result : no difference)


[PATCH 4/7] fs: Introduce SINGLE dentries for pipes, socket, anon fd

Sockets, pipes and anonymous fds have interesting properties.

Like other files, they use a dentry and an inode.

But dentries for these kind of files are not hashed into dcache,
since there is no way someone can lookup such a file in the vfs tree.
(/proc/{pid}/fd/{number} uses a different mechanism)

Still, allocating and freeing such dentries are expensive processes,
because we currently take dcache_lock inside d_alloc(), d_instantiate(),
and dput(). This lock is very contended on SMP machines.

This patch defines a new DCACHE_SINGLE flag, to mark a dentry as
a single one (for sockets, pipes, anonymous fd), and a new
d_alloc_single(const struct qstr *name, struct inode *inode)
method, called by the three subsystems.

Internally, dput() can take a fast path to dput_single() for
SINGLE dentries. No more atomic_dec_and_lock()
for such dentries.


Differences betwen an SINGLE dentry and a normal one are :

1) SINGLE dentry has the DCACHE_SINGLE flag
2) SINGLE dentry's parent is itself (DCACHE_DISCONNECTED)
This to avoid taking a reference on sb 'root' dentry, shared
by too many dentries.
3) They are not hashed into global hash table (DCACHE_UNHASHED)
4) Their d_alias list is empty

(socket8 bench result : from 25s to 19.9s)

[PATCH 5/7] fs: new_inode_single() and iput_single()

Goal of this patch is to not touch inode_lock for socket/pipes/anonfd
inodes allocation/freeing.

SINGLE dentries are attached to inodes that dont need to be linked
in a list of inodes, being "inode_in_use" or "sb->s_inodes"
As inode_lock was taken only to protect these lists, we avoid taking it as well.

Using iput_single() from dput_single() avoids taking inode_lock
at freeing time.

This patch has a very noticeable effect, because we avoid dirtying of three contended cache lines in new_inode(), and five cache lines
in iput()

("socketallocbench -n 8" result : from 19.9s to 3.01s)


[PATH 6/7] fs: struct file move from call_rcu() to SLAB_DESTROY_BY_RCU

From: Christoph Lameter <cl@...ux-foundation.org>

Currently we schedule RCU frees for each file we free separately. That has
several drawbacks against the earlier file handling (in 2.6.5 f.e.), which
did not require RCU callbacks:

1. Excessive number of RCU callbacks can be generated causing long RCU
  queues that in turn cause long latencies. We hit SLUB page allocation
  more often than necessary.

2. The cache hot object is not preserved between free and realloc. A close
  followed by another open is very fast with the RCUless approach because
  the last freed object is returned by the slab allocator that is
  still cache hot. RCU free means that the object is not immediately
  available again. The new object is cache cold and therefore open/close
  performance tests show a significant degradation with the RCU
  implementation.

One solution to this problem is to move the RCU freeing into the Slab
allocator by specifying SLAB_DESTROY_BY_RCU as an option at slab creation
time. The slab allocator will do RCU frees only when it is necessary
to dispose of slabs of objects (rare). So with that approach we can cut
out the RCU overhead significantly.

However, the slab allocator may return the object for another use even
before the RCU period has expired under SLAB_DESTROY_BY_RCU. This means
there is the (unlikely) possibility that the object is going to be
switched under us in sections protected by rcu_read_lock() and
rcu_read_unlock(). So we need to verify that we have acquired the correct
object after establishing a stable object reference (incrementing the
refcounter does that).


Signed-off-by: Christoph Lameter <cl@...ux-foundation.org>
Signed-off-by: Eric Dumazet <dada1@...mosbay.com>
Signed-off-by: Paul E. McKenney <paulmck@...ux.vnet.ibm.com>

("socketallocbench -n 8" result : from 3.01s to 2.20s)

[PATCH 7/7] fs: MS_NOREFCOUNT

Some fs are hardwired into kernel, and mntput()/mntget() hit a contended
cache line. We define a new superblock flag, MS_NOREFCOUNT, that is set
on socket, pipes and anonymous fd superblocks. mntput()/mntget() become
null ops on these fs.

("socketallocbench -n 8" result : from 2.20s to 1.64s)

cat socketallocbench.c
/*
 * socketallocbench benchmark
 *
 * Usage : socket [-n procs]  [-l loops]
 */
#include <sys/socket.h>
#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/wait.h>

void dowork(int loops)
{
        int i;

        for (i = 0; i < loops; i++)
                close(socket(AF_INET, SOCK_STREAM, 0));
}

int main(int argc, char *argv[])
{
        int i;
        int n = 1;
        int loops = 1000000;
        pid_t *pidtable;

        while ((i = getopt(argc, argv, "n:l:")) != EOF) {
                if (i == 'n')
                        n = atoi(optarg);
                if (i == 'l')
                        loops = atoi(optarg);
        }
        pidtable = malloc(n * sizeof(pid_t));
        for (i = 1; i < n; i++) {
                pidtable[i] = fork();
                if (pidtable[i] == 0) {
                        dowork(loops);
                        _exit(0);
                }
                if (pidtable[i] == -1) {
                        perror("fork");
                        n = i;
                        break;
                }
        }
        dowork(loops);
        for (i = 1; i < n; i++) {
                int status;

                wait(&status);
                }
        return 0;
}
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists