LIMIT_NETDEBUG allows the admin to disable some warning messages (echo 0
>/proc/sys/net/core/warnings).
The "TCP: Treason uncloaked!" message can use this facility.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
If CONFIG_TASK_IO_ACCOUNTING is defined, we update io accounting counters for
each task.
This patch permits reporting of values using the well known getrusage()
syscall, filling ru_inblock and ru_oublock instead of null values.
As TASK_IO_ACCOUNTING currently counts bytes counts, we approximate blocks
count doing : nr_blocks = nr_bytes / 512
Example of use :
----------------------
After patch is applied, /usr/bin/time command can now give a good
approximation of IO that the process had to do.
$ /usr/bin/time grep tototo /usr/include/*
Command exited with non-zero status 1
0.00user 0.02system 0:02.11elapsed 1%CPU (0avgtext+0avgdata 0maxresident)k
24288inputs+0outputs (0major+259minor)pagefaults 0swaps
$ /usr/bin/time dd if=/dev/zero of=/tmp/testfile count=1000
1000+0 enregistrements lus
1000+0 enregistrements écrits
512000 octets (512 kB) copiés, 0,00326601 seconde, 157 MB/s
0.00user 0.00system 0:00.00elapsed 80%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+3000outputs (0major+299minor)pagefaults 0swaps
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Oleg Nesterov <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Analysis of current linux futex code :
--------------------------------------
A central hash table futex_queues[] holds all contexts (futex_q) of waiting
threads.
Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to
perform lookups or insert/deletion of a futex_q.
When a futex_wait() is done, calling thread has to :
1) - Obtain a read lock on mmap_sem to be able to validate the user pointer
(calling find_vma()). This validation tells us if the futex uses
an inode based store (mapped file), or mm based store (anonymous mem)
2) - compute a hash key
3) - Atomic increment of reference counter on an inode or a mm_struct
4) - lock part of futex_queues[] hash table
5) - perform the test on value of futex.
(rollback is value != expected_value, returns EWOULDBLOCK)
(various loops if test triggers mm faults)
6) queue the context into hash table, release the lock got in 4)
7) - release the read_lock on mmap_sem
<block>
8) Eventually unqueue the context (but rarely, as this part may be done
by the futex_wake())
Futexes were designed to improve scalability but current implementation has
various problems :
- Central hashtable :
This means scalability problems if many processes/threads want to use
futexes at the same time.
This means NUMA unbalance because this hashtable is located on one node.
- Using mmap_sem on every futex() syscall :
Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic
ops on mmap_sem, dirtying cache line :
- lot of cache line ping pongs on SMP configurations.
mmap_sem is also extensively used by mm code (page faults, mmap()/munmap())
Highly threaded processes might suffer from mmap_sem contention.
mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded
programs because of contention on the mmap_sem cache line.
- Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter:
It's also a cache line ping pong on SMP. It also increases mmap_sem hold time
because of cache misses.
Most of these scalability problems come from the fact that futexes are in
one global namespace. As we use a central hash table, we must make sure
they are all using the same reference (given by the mm subsystem). We
chose to force all futexes be 'shared'. This has a cost.
But fact is POSIX defined PRIVATE and SHARED, allowing clear separation,
and optimal performance if carefuly implemented. Time has come for linux
to have better threading performance.
The goal is to permit new futex commands to avoid :
- Taking the mmap_sem semaphore, conflicting with other subsystems.
- Modifying a ref_count on mm or an inode, still conflicting with mm or fs.
This is possible because, for one process using PTHREAD_PROCESS_PRIVATE
futexes, we only need to distinguish futexes by their virtual address, no
matter the underlying mm storage is.
If glibc wants to exploit this new infrastructure, it should use new
_PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be
prepared to fallback on old subcommands for old kernels. Using one global
variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK.
PTHREAD_PROCESS_SHARED futexes should still use the old subcommands.
Compatibility with old applications is preserved, they still hit the
scalability problems, but new applications can fly :)
Note : the same SHARED futex (mapped on a file) can be used by old binaries
*and* new binaries, because both binaries will use the old subcommands.
Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic,
as this is the default semantic. Almost all applications should benefit
of this changes (new kernel and updated libc)
Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine)
/* calling futex_wait(addr, value) with value != *addr */
433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes)
424 cycles per futex(FUTEX_WAIT) call (using one futex)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex)
For reference :
187 cycles per getppid() call
188 cycles per umask() call
181 cycles per ni_syscall() call
Signed-off-by: Eric Dumazet <[email protected]>
Pierre Peiffer <[email protected]>
Cc: "Ulrich Drepper" <[email protected]>
Cc: "Nick Piggin" <[email protected]>
Cc: "Ingo Molnar" <[email protected]>
Cc: Rusty Russell <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I noticed expensive divides done in try_to_wakeup() and
find_busiest_group() on a bi dual core Opteron machine (total of 4 cores),
moderatly loaded (15.000 context switch per second)
oprofile numbers :
CPU: AMD64 processors, speed 2600.05 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit
mask of 0x00 (No unit mask) count 50000
samples % symbol name
...
613914 1.0498 try_to_wake_up
834 0.0013 :ffffffff80227ae1: div %rcx
77513 0.1191 :ffffffff80227ae4: mov %rax,%r11
608893 1.0413 find_busiest_group
1841 0.0031 :ffffffff802260bf: div %rdi
140109 0.2394 :ffffffff802260c2: test %sil,%sil
Some of these divides can use the reciprocal divides we introduced some
time ago (currently used in slab AFAIK)
We can assume a load will fit in a 32bits number, because with a
SCHED_LOAD_SCALE=128 value, its still a theorical limit of 33554432
When/if we reach this limit one day, probably cpus will have a fast
hardware divide and we can zap the reciprocal divide trick.
Ingo suggested to rename cpu_power to __cpu_power to make clear it should
not be modified without changing its reciprocal value too.
I did not convert the divide in cpu_avg_load_per_task(), because tracking
nr_running changes may be not worth it ? We could use a static table of 32
reciprocal values but it would add a conditional branch and table lookup.
[[email protected]: !SMP build fix]
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Ingo Molnar <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
struct clocksource is a critical data structure.
Most of its fields are read only, some of them are heavily modified at each
timer interrupt.
It makes sense to separate those fields and make sure they all share one
cache line, or at least the minimum for machines with small cache lines.
[[email protected]: build fix]
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: John Stultz <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
1) Introduces a new method in 'struct dentry_operations'. This method
called d_dname() might be called from d_path() to build a pathname for
special filesystems. It is called without locks.
Future patches (if we succeed in having one common dentry for all
pipes/sockets) may need to change prototype of this method, but we now
use : char *d_dname(struct dentry *dentry, char *buffer, int buflen);
2) Adds a dynamic_dname() helper function that eases d_dname() implementations
3) Defines d_dname method for sockets : No more sprintf() at socket
creation. This is delayed up to the moment someone does an access to
/proc/pid/fd/...
4) Defines d_dname method for pipes : No more sprintf() at pipe
creation. This is delayed up to the moment someone does an access to
/proc/pid/fd/...
A benchmark consisting of 1.000.000 calls to pipe()/close()/close() gives a
*nice* speedup on my Pentium(M) 1.6 Ghz :
3.090 s instead of 3.450 s
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Christoph Hellwig <[email protected]>
Acked-by: Linus Torvalds <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Change the order of fields of struct pid_entry (file fs/proc/base.c) in order
to avoid a hole on 64bit archs. (8 bytes saved per object)
Also change all pid_entry arrays to be const qualified, to make clear they
must not be modified.
Before (on x86_64) :
# size fs/proc/base.o
text data bss dec hex filename
15549 2192 0 17741 454d fs/proc/base.o
After :
# size fs/proc/base.o
text data bss dec hex filename
17229 176 0 17405 43fd fs/proc/base.o
Thats 336 bytes saved on kernel size on x86_64
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: "Eric W. Biederman" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The first thing done by timespec_trunc() is :
if (gran <= jiffies_to_usecs(1) * 1000)
This should really be a test against a constant known at compile time.
Alas, it isnt. jiffies_to_usec() was unilined so C compiler emits a function
call and a multiply to compute : a CONSTANT.
mov $0x1,%edi
mov %rbx,0xffffffffffffffe8(%rbp)
mov %r12,0xfffffffffffffff0(%rbp)
mov %edx,%ebx
mov %rsi,0xffffffffffffffc8(%rbp)
mov %rsi,%r12
callq ffffffff80232010 <jiffies_to_usecs>
imul $0x3e8,%eax,%eax
cmp %ebx,%eax
This patch reorders kernel/time.c a bit so that jiffies_to_usecs() is defined
before timespec_trunc() so that compiler now generates :
cmp $0x3d0900,%edx (HZ=250 on my machine)
This gives a better code (timespec_trunc() becoming a leaf function), and
shorter kernel size as well.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: john stultz <[email protected]>
Cc: Roman Zippel <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Some NUMA machines have a big MAX_NUMNODES (possibly 1024), but fewer
possible nodes. This patch dynamically sizes the 'struct kmem_cache' to
allocate only needed space.
I moved nodelists[] field at the end of struct kmem_cache, and use the
following computation in kmem_cache_init()
cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
nr_node_ids * sizeof(struct kmem_list3 *);
On my two nodes x86_64 machine, kmem_cache.obj_size is now 192 instead of 704
(This is because on x86_64, MAX_NUMNODES is 64)
On bigger NUMA setups, this might reduce the gfporder of "cache_cache"
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Pekka Enberg <[email protected]>
Cc: Andy Whitcroft <[email protected]>
Cc: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We can avoid allocating empty shared caches and avoid unecessary check of
cache->limit. We save some memory. We avoid bringing into CPU cache
unecessary cache lines.
All accesses to l3->shared are already checking NULL pointers so this patch is
safe.
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Pekka Enberg <[email protected]>
Cc: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The existing comment in mm/slab.c is *perfect*, so I reproduce it :
/*
* CPU bound tasks (e.g. network routing) can exhibit cpu bound
* allocation behaviour: Most allocs on one cpu, most free operations
* on another cpu. For these cases, an efficient object passing between
* cpus is necessary. This is provided by a shared array. The array
* replaces Bonwick's magazine layer.
* On uniprocessor, it's functionally equivalent (but less efficient)
* to a larger limit. Thus disabled by default.
*/
As most shiped linux kernels are now compiled with CONFIG_SMP, there is no way
a preprocessor #if can detect if the machine is UP or SMP. Better to use
num_possible_cpus().
This means on UP we allocate a 'size=0 shared array', to be more efficient.
Another patch can later avoid the allocations of 'empty shared arrays', to
save some memory.
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Pekka Enberg <[email protected]>
Acked-by: Christoph Lameter <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
1) struct ip6_flowlabel : moves 'users' field to avoid two 32bits
holes for 64bit arches. Shrinks by 8 bytes sizeof(struct
ip6_flowlabel)
2) ipv6_addr_cmp() and ipv6_addr_copy() dont need (void *) casts :
Compiler might take into account natural alignement of in6_addr
structs to emit better code for memcpy()/memcmp() Casts to (void *)
force byte accesses.
3) ipv6_addr_prefix() optimization :
Better to clear whole struct, as compiler can emit better code for
memset(addr, 0, 16) (2 stores on x86_64), and avoid some conditional
branches.
# size vmlinux.after vmlinux.before
text data bss dec hex filename
5262262 647612 557432 6467306 62aeea vmlinux.after
5262550 647612 557432 6467594 62b00a vmlinux.before
thats 288 bytes saved.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
__HAVE_ARCH_ADDR_SET seems unused these days, just get rid of it.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We apparently hit the 1024 limit of vsyscall_0 zone when some debugging
options are set, or if __vsyscall_gtod_data is 64 bytes larger.
In order to save 128 bytes from the vsyscall_0 zone, we move __vgetcpu_mode
& __jiffies to vsyscall_2 zone where they really belong, since they are
used only from vgetcpu() (which is in this vsyscall_2 area).
After patch is applied, new layout is :
ffffffffff600000 T vgettimeofday
ffffffffff60004e t vsysc2
ffffffffff600140 t vread_hpet
ffffffffff600150 t vread_tsc
ffffffffff600180 D __vsyscall_gtod_data
ffffffffff600400 T vtime
ffffffffff600413 t vsysc1
ffffffffff600800 T vgetcpu
ffffffffff600870 D __vgetcpu_mode
ffffffffff600880 D __jiffies
ffffffffff600c00 T venosys_1
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Current vsyscall_gtod_data is large (3 or 4 cache lines dirtied at timer
interrupt). We can shrink it to exactly 64 bytes (1 cache line on AMD64)
Instead of copying a whole struct clocksource, we copy only needed fields.
I deleted an unused field : offset_base
This patch fixes one oddity in vgettimeofday(): It can returns a timeval with
tv_usec = 1000000. Maybe not a bug, but why not doing the right thing ?
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
There is a tiny probability that the return value from vtime(time_t *t) is
Signed-off-by: Andi Kleen <[email protected]>
different than the value stored in *t
Using a temporary variable solves the problem and gives a faster code.
17: 48 85 ff test %rdi,%rdi
1a: 48 8b 05 00 00 00 00 mov 0(%rip),%rax #
__vsyscall_gtod_data.wall_time_tv.tv_sec
21: 74 03 je 26
23: 48 89 07 mov %rax,(%rdi)
26: c9 leaveq
27: c3 retq
Signed-off-by: Eric Dumazet <[email protected]>
Some people want to have many UDP sockets, binded to a single port but
many different addresses. We currently hash all those sockets into a
single chain. Processing of incoming packets is very expensive,
because the whole chain must be examined to find the best match.
I chose in this patch to hash UDP sockets with a hash function that
take into account both their port number and address : This has a
drawback because we need two lookups : one with a given address, one
with a wildcard (null) address.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Rusty added a new 'stats' field to struct net_device.
loopback driver can use it instead of declaring another struct
net_device_stats This saves some memory.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In order to get more randomness for secure_tcpv6_sequence_number(),
secure_tcp_sequence_number(), secure_dccp_sequence_number() functions,
we can use the high resolution time services, providing nanosec
resolution.
I've also done two kmalloc()/kzalloc() conversions.
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: James Morris <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As IPPROTO_TCP is 6, it makes sense to make sure inet_protos[] array
is properly cache line aligned to avoid false sharing on SMP.
c0680540 b peer_total
c0680544 b inet_peer_unused_head
c0680560 B inet_protos
On i386 this example, we can see that inet_protos[IPPROTO_TCP] shares
a potentially hot (and modified) cache line.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
tcp_memory_pressure and tcp_socket currently share a cache line with tcp_memory_allocated, tcp_sockets_allocated.
(Very hot cache line)
It makes sense to declare these variables as __read_mostly, to avoid false sharing on SMP.
ffffffff8081d9c0 B tcp_orphan_count
ffffffff8081d9c4 B tcp_memory_allocated
ffffffff8081d9c8 B tcp_sockets_allocated
ffffffff8081d9cc B tcp_memory_pressure
ffffffff8081d9d0 b tcp_md5sig_users
ffffffff8081d9d8 b tcp_md5sig_pool
ffffffff8081d9e0 b warntime.31570
ffffffff8081d9e8 b tcp_socket
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
There is a very tiny probability that build_ehash_secret() is called
at the same time by different CPUS.
Also, using __read_mostly is a must for inet_ehash_secret
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Now that network timestamps use ktime_t infrastructure, we can add a new
SOL_SOCKET sockopt SO_TIMESTAMPNS.
This command is similar to SO_TIMESTAMP, but permits transmission of
a 'timespec struct' instead of a 'timeval struct' control message.
(nanosecond resolution instead of microsecond)
Control message is labelled SCM_TIMESTAMPNS instead of SCM_TIMESTAMP
A socket cannot mix SO_TIMESTAMP and SO_TIMESTAMPNS : the two modes are
mutually exclusive.
sock_recv_timestamp() became too big to be fully inlined so I added a
__sock_recv_timestamp() helper function.
Signed-off-by: Eric Dumazet <[email protected]>
CC: [email protected]
Signed-off-by: David S. Miller <[email protected]>
net_msg_warn should be placed in the read_mostly section, to avoid
performance problems on SMP
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Now network timestamps use ktime_t infrastructure, we can add a new
ioctl() SIOCGSTAMPNS command to get timestamps in 'struct timespec'.
User programs can thus access to nanosecond resolution.
Signed-off-by: Eric Dumazet <[email protected]>
CC: Stephen Hemminger <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
1) Some sysctl vars are declared __read_mostly
2) We can avoid updating stack[] when doing an AVL lookup only.
lookup() macro is extended to receive a second parameter, that may be NULL
in case of a pure lookup (no need to save the AVL path). This removes
unnecessary instructions, because compiler knows if this _stack parameter is
NULL or not.
text size of net/ipv4/inetpeer.o is 2063 bytes instead of 2107 on x86_64
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We currently use a special structure (struct skb_timeval) and plain
'struct timeval' to store packet timestamps in sk_buffs and struct
sock.
This has some drawbacks :
- Fixed resolution of micro second.
- Waste of space on 64bit platforms where sizeof(struct timeval)=16
I suggest using ktime_t that is a nice abstraction of high resolution
time services, currently capable of nanosecond resolution.
As sizeof(ktime_t) is 8 bytes, using ktime_t in 'struct sock' permits
a 8 byte shrink of this structure on 64bit architectures. Some other
structures also benefit from this size reduction (struct ipq in
ipv4/ip_fragment.c, struct frag_queue in ipv6/reassembly.c, ...)
Once this ktime infrastructure adopted, we can more easily provide
nanosecond resolution on top of it. (ioctl SIOCGSTAMPNS and/or
SO_TIMESTAMPNS/SCM_TIMESTAMPNS)
Note : this patch includes a bug correction in
compat_sock_get_timestamp() where a "err = 0;" was missing (so this
syscall returned -ENOENT instead of 0)
Signed-off-by: Eric Dumazet <[email protected]>
CC: Stephen Hemminger <[email protected]>
CC: John find <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
sk_backlog is a critical field of struct sock. (known famous words)
It is (ab)used in hot paths, in particular in release_sock(), tcp_recvmsg(),
tcp_v4_rcv(), sk_receive_skb().
It really makes sense to place it next to sk_lock, because sk_backlog is only
used after sk_lock locked (and thus memory cache line in L1 cache). This
should reduce cache misses and sk_lock acquisition time.
(In theory, we could only move the head pointer near sk_lock, and leaving tail
far away, because 'tail' is normally not so hot, but keep it simple :) )
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
I noticed in oprofile study a cache miss in tcp_rcv_established() to read
copied_seq.
ffffffff80400a80 <tcp_rcv_established>: /* tcp_rcv_established total: 4034293
2.0400 */
55493 0.0281 :ffffffff80400bc9: mov 0x4c8(%r12),%eax copied_seq
543103 0.2746 :ffffffff80400bd1: cmp 0x3e0(%r12),%eax rcv_nxt
if (tp->copied_seq == tp->rcv_nxt &&
len - tcp_header_len <= tp->ucopy.len) {
In this function, the cache line 0x4c0 -> 0x500 is used only for this
reading 'copied_seq' field.
rcv_wup and copied_seq should be next to rcv_nxt field, to lower number of
active cache lines in hot paths. (tcp_rcv_established(), tcp_poll(), ...)
As you suggested, I changed tcp_create_openreq_child() so that these fields
are changed together, to avoid adding a new store buffer stall.
Patch is 64bit friendly (no new hole because of alignment constraints)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
ARCH_HAVE_XTIME_LOCK is used by x86_64 arch . This arch needs to place a
read only copy of xtime_lock into vsyscall page. This read only copy is
named __xtime_lock, and xtime_lock is defined in
arch/x86_64/kernel/vmlinux.lds.S as an alias. So the declaration of
xtime_lock in kernel/timer.c was guarded by ARCH_HAVE_XTIME_LOCK define,
defined to true on x86_64.
We can get same result with _attribute__((weak)) in the declaration. linker
should do the job.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
oprofile hunting showed a stall in rw_verify_area(), because of triple
indirection and potential cache misses.
(file->f_path.dentry->d_inode->i_flock)
By moving initialization of 'struct inode' pointer before the pos/count
sanity tests, we allow the compiler and processor to perform two loads by
anticipation, reducing stall, without prefetch() hints. Even x86 arch has
enough registers to not use temporary variables and not increase text size.
I validated this patch running a bench and studied oprofile changes, and
absolute perf of the test program.
Results of my epoll_pipe_bench (source available on request) on a Pentium-M
1.6 GHz machine
Before :
# ./epoll_pipe_bench -l 30 -t 20
Avg: 436089 evts/sec read_count=8843037 write_count=8843040 21.218390 samples
per call
(best value out of 10 runs)
After :
# ./epoll_pipe_bench -l 30 -t 20
Avg: 470980 evts/sec read_count=9549871 write_count=9549894 21.216694 samples
per call
(best value out of 10 runs)
oprofile CPU_CLK_UNHALTED events gave a reduction from 5.3401 % to 2.5851 %
for the rw_verify_area() function.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I added IS_NOATIME(inode) macro definition in include/linux/fs.h, true if
the inode superblock is marked readonly or noatime.
This new macro is then used in touch_atime() instead of separatly testing
MS_RDONLY and MS_NOATIME
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I noticed cache misses in touch_atime() that can be avoided if we keep
mnt_count & mnt_expiry_mark in a different cache line than mnt_flags
(mostly read)
mnt_count & mnt_expiry_mark are modified each time a file is opened/closed
in a file system.
touch_atime() is called each time a file is read, and generally needs to
read mnt_flags.
Other fields of struct vfsmount are mostly read so I chose to move
mnt_count & mnt_expiry_mark at the end of struct vfsmount. And adding a
comment so that nobody tries to re-arrange fields to fill the holes :)
On 64bits platforms, the new offsetof(mnt_count) is 0xC0
On 32bits platforms, it is 0x60, so I didnot add a
____cacheline_aligned_in_smp because it would have a too big impact on the
size of this object (in particular if CONFIG_X86_L1_CACHE_SHIFT=7)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I forgot to test build this part of the networking code... Sorry guys.
This patch renames u.rt_next to u.dst.rt_next
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This last patch (but not least :) ) finally moves the next pointer at
the end of struct dst_entry. This permits to perform route cache
lookups with a minimal cost of one cache line per entry, instead of
two.
Both 32bits and 64bits platforms benefit from this new layout.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This patch removes the next pointer from 'struct dn_route.u' union,
and renames u.rt_next to u.dst.dn_next.
It also moves 'struct flowi' right after 'struct dst_entry' to prepare
speedup lookups.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This patch removes the next pointer from 'struct rt6_info.u' union,
and renames u.next to u.dst.rt6_next.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This patch removes the rt_next pointer from 'struct rtable.u' union,
and renames u.rt_next to u.dst_rt_next.
It also moves 'struct flowi' right after 'struct dst_entry' to prepare
the gain on lookups.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
This patch introduces an anonymous union to nicely express the fact that all
objects inherited from struct dst_entry should access to the generic 'next'
pointer but with appropriate type verification.
This patch is a prereq before following patches.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
In a prior patch, I introduced a sk_hash field (__sk_common.skc_hash) to let
tcp lookups use one cache line per unmatched entry instead of two.
We can also use sk_hash to speedup UDP part as well. We store in sk_hash the
hnum value, and use sk->sk_hash (same cache line than 'next' pointer),
instead of inet->num (different cache line)
Note : We still have a false sharing problem for SMP machines, because
sock_hold(sock) dirties the cache line containing the 'next' pointer. Not
counting the udp_hash_lock rwlock. (did someone mentioned RCU ? :) )
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
I believe dead code from sock_from_file() can be cleaned up.
All sockets are now built using sock_attach_fd(), that puts the 'sock' pointer
into file->private_data and &socket_file_ops into file->f_op
I could not find a place where file->private_data could be set to NULL,
keeping opened the file.
So to get 'sock' from a 'file' pointer, either :
- This is a socket file (f_op == &socket_file_ops), and we can directly get
'sock' from private_data.
- This is not a socket, we return -ENOTSOCK and dont even try to find a socket
via dentry/inode :)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
ehash table layout is currently this one :
First half of this table is used by sockets not in TIME_WAIT state
Second half of it is used by sockets in TIME_WAIT state.
This is non optimal because of for a given hash or socket, the two chain heads
are located in separate cache lines.
Moreover the locks of the second half are never used.
If instead of this halving, we use two list heads in inet_ehash_bucket instead
of only one, we probably can avoid one cache miss, and reduce ram usage,
particularly if sizeof(rwlock_t) is big (various CONFIG_DEBUG_SPINLOCK,
CONFIG_DEBUG_LOCK_ALLOC settings). So we still halves the table but we keep
together related chains to speedup lookups and socket state change.
In this patch I did not try to align struct inet_ehash_bucket, but a future
patch could try to make this structure have a convenient size (a power of two
or a multiple of L1_CACHE_SIZE).
I guess rwlock will just vanish as soon as RCU is plugged into ehash :) , so
maybe we dont need to scratch our heads to align the bucket...
Note : In case struct inet_ehash_bucket is not a power of two, we could
probably change alloc_large_system_hash() (in case it use __get_free_pages())
to free the unused space. It currently allocates a big zone, but the last
quarter of it could be freed. Again, this should be a temporary 'problem'.
Patch tested on ipv4 tcp only, but should be OK for IPV6 and DCCP.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
During an oprofile session of linux-2.6.20 on a dual opteron system, I noticed
an expensive divide was done in tg3_poll().
I am using gcc-4.1.1, so the following comment from drivers/net/tg3.c seems
over-optimistic :
/* Do not place this n-ring entries value into the tp struct itself,
* we really want to expose these constants to GCC so that modulo et
* al. operations are done with shifts and masks instead of with
* hw multiply/modulo instructions. Another solution would be to
* replace things like '% foo' with '& (foo - 1)'.
*/
#define TG3_RX_RCB_RING_SIZE(tp) \
((tp->tg3_flags2 & TG3_FLG2_5705_PLUS) ? 512 : 1024)
Assembly code before patch :
(oprofile results included)
6434 0.0088 :ffffffff803684b9: mov 0x6f0(%r15),%eax
587 8.0e-04 :ffffffff803684c0: and $0x40000,%eax
2170 0.0030 :ffffffff803684c5: cmp $0x1,%eax
:ffffffff803684c8: lea 0x1(%r13),%eax
:ffffffff803684cc: sbb %ecx,%ecx
2051 0.0028 :ffffffff803684ce: xor %edx,%edx
:ffffffff803684d0: and $0x200,%ecx
20 2.7e-05 :ffffffff803684d6: add $0x200,%ecx
1986 0.0027 :ffffffff803684dc: div %ecx
103427 0.1410 :ffffffff803684de: cmp %edx,0xffffffffffffff7c(%rbp)
Assembly code after the suggested patch :
ffffffff803684b9: mov 0x6f0(%r15),%eax
ffffffff803684c0: and $0x40000,%eax
ffffffff803684c5: cmp $0x1,%eax
ffffffff803684c8: sbb %eax,%eax
ffffffff803684ca: inc %r13d
ffffffff803684cd: and $0x200,%eax
ffffffff803684d2: add $0x1ff,%eax
ffffffff803684d7: and %eax,%r13d
ffffffff803684da: cmp %r13d,0xffffffffffffff7c(%rbp)
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Michael Chan <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
calc_load() is called by timer interrupt to update avenrun[]. It currently
calls nr_active() at each timer tick (HZ per second), while the update of
avenrun[] is done only once every 5 seconds. (LOAD_FREQ=5 Hz)
nr_active() is quite expensive on SMP machines, since it has to sum up
nr_running and nr_uninterruptible of all online CPUS, bringing foreign
dirty cache lines.
This patch is an optimization of calc_load() so that nr_active() is called
only if we need it.
The use of unlikely() is welcome since the condition is true only once every
5*HZ time.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Ingo Molnar <[email protected]>
Acked-by: "Siddha, Suresh B" <[email protected]>
Cc: Nick Piggin <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
When some objects are allocated by one CPU but freed by another CPU we can
consume lot of cycles doing divides in obj_to_index().
(Typical load on a dual processor machine where network interrupts are
handled by one particular CPU (allocating skbufs), and the other CPU is
running the application (consuming and freeing skbufs))
Here on one production server (dual-core AMD Opteron 285), I noticed this
divide took 1.20 % of CPU_CLK_UNHALTED events in kernel. But Opteron are
quite modern cpus and the divide is much more expensive on oldest
architectures :
On a 200 MHz sparcv9 machine, the division takes 64 cycles instead of 1
cycle for a multiply.
Doing some math, we can use a reciprocal multiplication instead of a divide.
If we want to compute V = (A / B) (A and B being u32 quantities)
we can instead use :
V = ((u64)A * RECIPROCAL(B)) >> 32 ;
where RECIPROCAL(B) is precalculated to ((1LL << 32) + (B - 1)) / B
Note :
I wrote pure C code for clarity. gcc output for i386 is not optimal but
acceptable :
mull 0x14(%ebx)
mov %edx,%eax // part of the >> 32
xor %edx,%edx // useless
mov %eax,(%esp) // could be avoided
mov %edx,0x4(%esp) // useless
mov (%esp),%ebx
[[email protected]: small cleanups]
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Christoph Lameter <[email protected]>
Cc: David Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Fields of struct pipe_buf_operations have not a precise layout (ie not
optimized to fit cache lines nor reduce cache line ping pongs)
The bufs[] array is *large* and is placed near the beginning of the
structure, so all following fields have a large offset. This is
unfortunate because many archs have smaller instructions when using small
offsets relative to a base register. On x86 for example, 7 bits offsets
have smaller instruction lengths.
Moving bufs[] at the end of pipe_buf_operations permits all fields to have
small offsets, and reduce text size, and icache pressure.
# size vmlinux.pre vmlinux
text data bss dec hex filename
3268989 664356 492196 4425541 438745 vmlinux.pre
3268765 664356 492196 4425317 438665 vmlinux
So this patch reduces text size by 224 bytes on my x86_64 machine. Similar
results on ia32.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
- pipe/splice should use const pipe_buf_operations and file_operations
- struct pipe_inode_info has an unused field "start" : get rid of it.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Jens Axboe <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
hh_lock was converted from rwlock to seqlock by Stephen.
To have a 100% benefit of this change, I suggest to place read mostly fields
of hh_cache in a separate cache line, because hh_refcnt may be changed quite
frequently on some busy machines.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
On 32bits SMP platforms, 64bits i_size is protected by a seqcount
(i_size_seqcount).
When i_size is read or written, i_size_seqcount is read/written as well, so
it make sense to group these two fields together in the same cache line.
This patch moves i_size_seqcount next to i_size, and also moves i_version
to let offsetof(struct inode, i_size) being 0x40 instead of 0x3c (for
32bits platforms).
For 64 bits platforms, i_size_seqcount doesnt exist, and the move of a
'long i_version' should not introduce a new hole because of padding.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We currently insert socket dentries into the global dentry hashtable. This
is suboptimal because there is currently no way these entries can be used
for a lookup(). (/proc/xxx/fd/xxx uses a different mechanism). Inserting
them in dentry hashtable slows dcache lookups.
To let __dpath() still work correctly (ie not adding a " (deleted)") after
dentry name, we do :
- Right after d_alloc(), pretend they are hashed by clearing the
DCACHE_UNHASHED bit.
- Call d_instantiate() instead of d_add() : dentry is not inserted in
hash table.
__dpath() & friends work as intended during dentry lifetime.
- At dismantle time, once dput() must clear the dentry, setting again
DCACHE_UNHASHED bit inside the custom d_delete() function provided by
socket code, so that dput() can just kill_it.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Al Viro <[email protected]>
Acked-by: "David S. Miller" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Some dentries don't need to be globally visible in dentry hashtable.
(pipes & sockets)
Such dentries dont need to wait for a RCU grace period at delete time.
Being able to free them permits a better CPU cache use (hot cache)
This patch combined with (dont insert pipe dentries into dentry_hashtable)
reduced time of { pipe(p); close(p[0]); close(p[1]);} on my UP machine (1.6
GHz Pentium-M) from 3.23 us to 2.86 us (But this patch does not depend on
other patches, only bench results)
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Al Viro <[email protected]>
Cc: Maneesh Soni <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Dipankar Sarma <[email protected]>
Acked-by: David Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We currently insert pipe dentries into the global dentry hashtable. This
is suboptimal because there is currently no way these entries can be used
for a lookup(). (/proc/xxx/fd/xxx uses a different mechanism). Inserting
them in dentry hashtable slows dcache lookups.
To let __dpath() still work correctly (ie not adding a " (deleted)") after
dentry name, we do :
- Right after d_alloc(), pretend they are hashed by clearing the
DCACHE_UNHASHED bit.
- Call d_instantiate() instead of d_add() : dentry is not inserted in
hash table.
__dpath() & friends work as intended during dentry lifetime.
- At dismantle time, once dput() must clear the dentry, setting again
DCACHE_UNHASHED bit inside the custom d_delete() function provided by
pipe code, so that dput() can just kill_it.
This patch, combined with (avoid RCU for never hashed dentries) reduced
time of { pipe(p); close(p[0]); close(p[1]);} on my UP machine (1.6GHz
Pentium-M) from 3.23 us to 2.86 us (But this patch does not depend on other
patches, only bench results)
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: David Miller <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
On some workloads, (for example when lot of close() syscalls are done), RCU
qlen can be quite large, and RCU heads are no longer in cpu cache when
rcu_do_batch() is called.
This patch adds a prefetch() in rcu_do_batch() to give CPU a hint to bring
back cache lines containing 'struct rcu_head's.
Most list manipulations macros include prefetch(), but not open coded ones
(at least with current C compilers :) )
I got a nice speedup on a trivial benchmark (3.48 us per iteration instead
of 3.95 us on a 1.6 GHz Pentium-M)
while (1) { pipe(p); close(fd[0]); close(fd[1]);}
Signed-off-by: Eric Dumazet <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
We currently allocate a fixed size (TCP_SYNQ_HSIZE=512) slots hash table for
each LISTEN socket, regardless of various parameters (listen backlog for
example)
On x86_64, this means order-1 allocations (might fail), even for 'small'
sockets, expecting few connections. On the contrary, a huge server wanting a
backlog of 50000 is slowed down a bit because of this fixed limit.
This patch makes the sizing of listen hash table a dynamic parameter,
depending of :
- net.core.somaxconn tunable (default is 128)
- net.ipv4.tcp_max_syn_backlog tunable (default : 256, 1024 or 128)
- backlog value given by user application (2nd parameter of listen())
For large allocations (bigger than PAGE_SIZE), we use vmalloc() instead of
kmalloc().
We still limit memory allocation with the two existing tunables (somaxconn &
tcp_max_syn_backlog). So for standard setups, this patch actually reduce RAM
usage.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
- reorder 'struct vm_struct' to speedup lookups on CPUS with small cache
lines. The fields 'next,addr,size' should be now in the same cache line,
to speedup lookups.
- One minor cleanup in __get_vm_area_node()
- Bugfixes in vmalloc_user() and vmalloc_32_user() NULL returns from
__vmalloc() and __find_vm_area() were not tested.
[[email protected]: remove redundant BUG_ONs]
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Nick Piggin <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
As suggested by David, just kill off some unused fields in dnports to
reduce sizef(struct flowi). If they come back, they should be moved to
nl_u.dn_u in order not to enlarge again struct flowi
[ Modified to really delete this stuff instead of using #if 0. -DaveM ]
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As BHs are off in loopback_xmit(), preemption cannot occurs, so we can
use __get_cpu_var() instead of per_cpu() (and avoid a
preempt_enable()/preempt_disable() pair)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Lot of routers/embedded devices still use CPUS with 16/32 bytes cache
lines. (486, Pentium, ... PIII) It makes sense to group together
fields used at lookup time so they fit in one cache line. This reduce
cache footprint and speedup lookups.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
I believe this NET_INC_STATS() call can be replaced by
NET_INC_STATS_BH(), a little bit cheaper.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
We dont need a full struct net_device_stats (currently 23 long : 184 bytes on
x86_64) per possible CPU, but only two counters : bytes and packets
We save few CPU cycles too in loopback_xmit() not updating 4 fields, but 2.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
It seems latest kernel has a wrong/missing __read_mostly implementation
for x86_64
__read_mostly macro should be declared outside of #if CONFIG_X86_VSMP block
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I discovered on oprofile hunting on a SMP platform that dentry lookups were
slowed down because d_hash_mask, d_hash_shift and dentry_hashtable were in
a cache line that contained inodes_stat. So each time inodes_stats is
changed by a cpu, other cpus have to refill their cache line.
This patch moves some variables to the __read_mostly section, in order to
avoid false sharing. RCU dentry lookups can go full speed.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
pfn_to_page() and others need to access both memnode_shift and the very
first bytes of memnodemap[]. If we force memnode_shift to be just before the
memnodemap array, we can reduce the memory footprint to one cache line
instead of two for most setups. This patch introduce a 'memnode' structure
where shift and map[] are carefully placed.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch avoids arithmetic on 'signed' types that are slower than
'unsigned'. This saves space and cpu cycles.
size of kernel/sys.o before the patch (gcc-3.4.5)
text data bss dec hex filename
10924 252 4 11180 2bac kernel/sys.o
size of kernel/sys.o after the patch
text data bss dec hex filename
10903 252 4 11159 2b97 kernel/sys.o
I noticed that gcc-4.1.0 (from Fedora Core 5) even uses idiv instruction for
(a+b)/2 if a and b are signed.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Instead of using for_each_cpu(i), we can use for_each_online_cpu(i).
When a CPU goes offline (ie removed from online map), it might have a non
null bh_accounting.nr, so this patch adds a transfer of this counter to an
online CPU counter.
We already have a hotcpu_notifier, (function buffer_cpu_notify()), where we
can do this bh_accounting.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Now CONFIG_DEBUG_INITDATA is in, initial percpu data
[__per_cpu_start,__per_cpu_end] can be declared as a redzone, and invalid
accesses after boot can be detected, at least for i386.
We can let non possible cpus percpu data point to this 'redzone' instead of
NULL .
NULL was not a good choice because part of [0..32768] memory may be
readable and invalid accesses may happen unnoticed.
If CONFIG_DEBUG_INITDATA is not defined, each non possible cpu points to
the initial percpu data (__per_cpu_offset[cpu] == 0), thus invalid accesses
wont be detected/crash.
This patch also moves __per_cpu_offset[] to read_mostly area to avoid false
sharing.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
1) Reduce the size of (struct fdtable) to exactly 64 bytes on 32bits
platforms, lowering kmalloc() allocated space by 50%.
2) Reduce the size of (files_struct), using a special 32 bits (or
64bits) embedded_fd_set, instead of a 1024 bits fd_set for the
close_on_exec_init and open_fds_init fields. This save some ram (248
bytes per task) as most tasks dont open more than 32 files. D-Cache
footprint for such tasks is also reduced to the minimum.
3) Reduce size of allocated fdset. Currently two full pages are
allocated, that is 32768 bits on x86 for example, and way too much. The
minimum is now L1_CACHE_BYTES.
UP and SMP should benefit from this patch, because most tasks will touch
only one cache line when open()/close() stdin/stdout/stderr (0/1/2),
(next_fd, close_on_exec_init, open_fds_init, fd_array[0 .. 2] being in the
same cache line)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Avoid an atomic operation in kref_put() when the last reference is
dropped. On most platforms, atomic_read() is a plan read of the counter
and involves no atomic at all.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Greg Kroah-Hartman <[email protected]>
percpu_data blindly allocates bootmem memory to store NR_CPUS instances of
cpudata, instead of allocating memory only for possible cpus.
As a preparation for changing that, we need to convert various 0 -> NR_CPUS
loops to use for_each_cpu().
(The above only applies to users of asm-generic/percpu.h. powerpc has gone it
alone and is presently only allocating memory for present CPUs, so it's
currently corrupting memory).
Signed-off-by: Eric Dumazet <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: James Bottomley <[email protected]>
Acked-by: Ingo Molnar <[email protected]>
Cc: Jens Axboe <[email protected]>
Cc: Anton Blanchard <[email protected]>
Acked-by: William Irwin <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Using __get_cpu_var(obj) is slightly faster than per_cpu_ptr(obj,
raw_smp_processor_id()).
1) Smaller code and memory use
For static and small objects, DEFINE_PER_CPU(type, object) is preferred over a
alloc_percpu() : Better and smaller code to access them, and no extra memory
(storing the pointer, and the percpu array of pointers)
x86_64 code before patch
mov 1237577(%rip),%rax # ffffffff803e5990 <rt_cache_stat>
not %rax # part of per_cpu machinery
mov %gs:0x3c,%edx # get cpu number
movslq %edx,%rdx # extend 32 bits cpu number to 64 bits
mov (%rax,%rdx,8),%rax # get the pointer for this cpu
incl 0x38(%rax)
x86_64 code after patch
mov $per_cpu__rt_cache_stat,%rdx
mov %gs:0x48,%rax # get percpu data offset
incl 0x38(%rax,%rdx,1)
2) False sharing avoidance for SMP :
For a small NR_CPUS, the array of per cpu pointers allocated in alloc_percpu()
can be <= 32 bytes. This let slab code gives a part of a cache line. If the
other part of this 64 bytes (or 128 bytes) cache line is used by a mostly
written object, we can have false sharing and expensive per_cpu_ptr() operations.
Size of rt_cache_stat is 64 bytes, so this patch is not a danger of a too big
increase of bss (in UP mode) or static per_cpu data for SMP
(PERCPU_ENOUGH_ROOM is currently 32768 bytes)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
I found that blkp field was not used in kernel tree.
As most of the times NR_CPUS is a power of two and kmalloc() memory blocks
too, this extra field basically doubles the memory space allocated in
__alloc_percpu() to store the 'struct percpu_data'
(for example, if NR_CPUS=8 on i386, kmalloc(4*8+4) returns a 64 bytes block
instead of a 32 bytes block after this patch)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Some long time ago, dentry struct was carefully tuned so that on 32 bits
UP, sizeof(struct dentry) was exactly 128, ie a power of 2, and a multiple
of memory cache lines.
Then RCU was added and dentry struct enlarged by two pointers, with nice
results for SMP, but not so good on UP, because breaking the above tuning
(128 + 8 = 136 bytes)
This patch reverts this unwanted side effect, by using an union (d_u),
where d_rcu and d_child are placed so that these two fields can share their
memory needs.
At the time d_free() is called (and d_rcu is really used), d_child is known
to be empty and not touched by the dentry freeing.
Lockless lookups only access d_name, d_parent, d_lock, d_op, d_flags (so
the previous content of d_child is not needed if said dentry was unhashed
but still accessed by a CPU because of RCU constraints)
As dentry cache easily contains millions of entries, a size reduction is
worth the extra complexity of the ugly C union.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Dipankar Sarma <[email protected]>
Cc: Maneesh Soni <[email protected]>
Cc: Miklos Szeredi <[email protected]>
Cc: "Paul E. McKenney" <[email protected]>
Cc: Ian Kent <[email protected]>
Cc: Paul Jackson <[email protected]>
Cc: Al Viro <[email protected]>
Cc: Christoph Hellwig <[email protected]>
Cc: Trond Myklebust <[email protected]>
Cc: Neil Brown <[email protected]>
Cc: James Morris <[email protected]>
Cc: Stephen Smalley <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
I noticed that some of 'struct proto_ops' used in the kernel may share
a cache line used by locks or other heavily modified data. (default
linker alignement is 32 bytes, and L1_CACHE_LINE is 64 or 128 at
least)
This patch makes sure a 'struct proto_ops' can be declared as const,
so that all cpus can share all parts of it without false sharing.
This is not mandatory : a driver can still use a read/write structure
if it needs to (and eventually a __read_mostly)
I made a global stubstitute to change all existing occurences to make
them const.
This should reduce the possibility of false sharing on SMP, and
speedup some socket system calls.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Part of a performance problem with ip_tables is that memory allocation
is not NUMA aware, but 'only' SMP aware (ie each CPU normally touch
separate cache lines)
Even with small iptables rules, the cost of this misplacement can be
high on common workloads. Instead of using one vmalloc() area
(located in the node of the iptables process), we now allocate an area
for each possible CPU, using vmalloc_node() so that memory should be
allocated in the CPU's node if possible.
Port to arp_tables and ip6_tables by Harald Welte.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
As reported by Keith Mannthey, there are problems in populate_memnodemap()
The bug was that the compute_hash_shift() was returning 31, with incorrect
initialization of memnodemap[]
To correct the bug, we must use (1UL << shift) instead of (1 << shift) to
avoid an integer overflow, and we must check that shift < 64 to avoid an
infinite loop.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Compute the highest possible value for memnode_shift, in order to reduce
footprint of memnodemap[] to the minimum, thus making all users
(phys_to_nid(), kfree()), more cache friendly.
Before the patch :
Node 0 MemBase 0000000000000000 Limit 00000001ffffffff
Node 1 MemBase 0000000200000000 Limit 00000003ffffffff
Using 23 for the hash shift. Max adder is 3ffffffff
After the patch :
Node 0 MemBase 0000000000000000 Limit 00000001ffffffff
Node 1 MemBase 0000000200000000 Limit 00000003ffffffff
Using 33 for the hash shift.
In this case, only 2 bytes of memnodemap[] are used, instead of 2048
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andi Kleen <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
The file_lock spinlock sits close to mostly read fields of 'struct
files_struct'
In SMP (and NUMA) environments, each time a thread wants to open or close
a file, it has to acquire the spinlock, thus invalidating the cache line
containing this spinlock on other CPUS. So other threads doing
read()/write()/... calls that use RCU to access the file table are going
to ask further memory (possibly NUMA) transactions to read again this
memory line.
Move the spinlock to another cache line, so that concurrent threads can
share the cache line containing 'count' and 'fdt' fields.
It's worth up to 9% on a microbenchmark using a 4-thread 2-package x86
machine. See
http://marc.theaimsgroup.com/?l=linux-kernel&m=112680448713342&w=2
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
Now that RCU applied on 'struct file' seems stable, we can place f_rcuhead
in a memory location that is not anymore used at call_rcu(&f->f_rcuhead,
file_free_rcu) time, to reduce the size of this critical kernel object.
The trick I used is to move f_rcuhead and f_list in an union called f_u
The callers are changed so that f_rcuhead becomes f_u.fu_rcuhead and f_list
becomes f_u.f_list
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This makes call_rcu() keep track of how many events there are on the RCU
list, and cause a reschedule event when the list gets too long.
This helps keep RCU event lists down.
Signed-off-by: Linus Torvalds <[email protected]>
No need to align struct inet_ehash_bucket on a 8 bytes boundary.
On 32 bits Uniprocessor, that's a waste of 4 bytes per struct (50 %)
On other platforms, the attribute is useless, natual alignement is already 8.
platform | Size before | Size after patch
-------------+-------------+------------------
32 bits, UP | 8 | 4
32 bits, SMP | 8 | 8
64 bits, UP | 8 | 8
64 bits, SMP | 16 | 16
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Arnaldo and I agreed it could be applied now, because I have other
pending patches depending on this one (Thank you Arnaldo)
(The other important patch moves skc_refcnt in a separate cache line,
so that the SMP/NUMA performance doesnt suffer from cache line ping pongs)
1) First some performance data :
--------------------------------
tcp_v4_rcv() wastes a *lot* of time in __inet_lookup_established()
The most time critical code is :
sk_for_each(sk, node, &head->chain) {
if (INET_MATCH(sk, acookie, saddr, daddr, ports, dif))
goto hit; /* You sunk my battleship! */
}
The sk_for_each() does use prefetch() hints but only the begining of
"struct sock" is prefetched.
As INET_MATCH first comparison uses inet_sk(__sk)->daddr, wich is far
away from the begining of "struct sock", it has to bring into CPU
cache cold cache line. Each iteration has to use at least 2 cache
lines.
This can be problematic if some chains are very long.
2) The goal
-----------
The idea I had is to change things so that INET_MATCH() may return
FALSE in 99% of cases only using the data already in the CPU cache,
using one cache line per iteration.
3) Description of the patch
---------------------------
Adds a new 'unsigned int skc_hash' field in 'struct sock_common',
filling a 32 bits hole on 64 bits platform.
struct sock_common {
unsigned short skc_family;
volatile unsigned char skc_state;
unsigned char skc_reuse;
int skc_bound_dev_if;
struct hlist_node skc_node;
struct hlist_node skc_bind_node;
atomic_t skc_refcnt;
+ unsigned int skc_hash;
struct proto *skc_prot;
};
Store in this 32 bits field the full hash, not masked by (ehash_size -
1) Using this full hash as the first comparison done in INET_MATCH
permits us immediatly skip the element without touching a second cache
line in case of a miss.
Suppress the sk_hashent/tw_hashent fields since skc_hash (aliased to
sk_hash and tw_hash) already contains the slot number if we mask with
(ehash_size - 1)
File include/net/inet_hashtables.h
64 bits platforms :
#define INET_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
(((__sk)->sk_hash == (__hash))
((*((__u64 *)&(inet_sk(__sk)->daddr)))== (__cookie)) && \
((*((__u32 *)&(inet_sk(__sk)->dport))) == (__ports)) && \
(!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
32bits platforms:
#define TCP_IPV4_MATCH(__sk, __hash, __cookie, __saddr, __daddr, __ports, __dif)\
(((__sk)->sk_hash == (__hash)) && \
(inet_sk(__sk)->daddr == (__saddr)) && \
(inet_sk(__sk)->rcv_saddr == (__daddr)) && \
(!((__sk)->sk_bound_dev_if) || ((__sk)->sk_bound_dev_if == (__dif))))
- Adds a prefetch(head->chain.first) in
__inet_lookup_established()/__tcp_v4_check_established() and
__inet6_lookup_established()/__tcp_v6_check_established() and
__dccp_v4_check_established() to bring into cache the first element of the
list, before the {read|write}_lock(&head->lock);
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Arnaldo Carvalho de Melo <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Place them on separate cache lines in SMP to lower memory bouncing
between multiple CPU accessing the device.
- One part is mostly used on receive path (including
eth_type_trans()) (poll_list, poll, quota, weight, last_rx,
dev_addr, broadcast)
- One part is mostly used on queue transmit path (qdisc)
(queue_lock, qdisc, qdisc_sleeping, qdisc_list, tx_queue_len)
- One part is mostly used on xmit path (device)
(xmit_lock, xmit_lock_owner, priv, hard_start_xmit, trans_start)
'features' is placed outside of these hot points, in a location that
may be shared by all cpus (because mostly read)
name_hlist is moved close to name[IFNAMSIZ] to speedup __dev_get_by_name()
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
dentry cache uses sophisticated RCU technology (and prefetching if
available) but touches 2 cache lines per dentry during hlist lookup.
This patch moves d_hash in the same cache line than d_parent and d_name
fields so that :
1) One cache line is needed instead of two.
2) the hlist_for_each_rcu() prefetching has a chance to bring all the
needed data in advance, not only the part that includes d_hash.next.
I also changed one old comment that was wrong for 64bits.
A further optimisation would be to separate dentry in two parts, one that
is mostly read, and one writen (d_count/d_lock) to avoid false sharing on
SMP/NUMA but this would need different field placement depending on 32bits
or 64bits platform.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
struct file cleanup: f_maxcount has an unique value (INT_MAX). Just use
the hard-wired value.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This is a multi-part message in MIME format. If the cpu lacks 3DNOW
feature, we can use a normal prefetcht0 instruction instead of NOP5.
"prefetchw (%rxx)" and "prefetcht0 (%rxx)" have the same length, ranging
from 3 to 5 bytes depending on the register. So this patch even helps
AMD64, shortening the length of the code.
Signed-off-by: Eric Dumazet <[email protected]>
Acked-by: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
One of my x86_64 (linux 2.6.13) server log is filled with :
schedule_timeout: wrong timeout value ffffffffffffff06 from ffffffff802e63ca
schedule_timeout: wrong timeout value ffffffffffffff06 from ffffffff802e63ca
schedule_timeout: wrong timeout value ffffffffffffff06 from ffffffff802e63ca
schedule_timeout: wrong timeout value ffffffffffffff06 from ffffffff802e63ca
schedule_timeout: wrong timeout value ffffffffffffff06 from ffffffff802e63ca
This is because some application does a
struct linger li;
li.l_onoff = 1;
li.l_linger = -1;
setsockopt(sock, SOL_SOCKET, SO_LINGER, &li, sizeof(li));
And unfortunatly l_linger is defined as a 'signed int' in
include/linux/socket.h:
struct linger {
int l_onoff; /* Linger active */
int l_linger; /* How long to linger for */
};
I dont know if it's safe to change l_linger to 'unsigned int' in the
include file (It might be defined as int in ABI specs)
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Avoid touching file->f_dentry on sockets, since file->private_data
directly gives us the socket pointer.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Mostobjects returned by __cache_alloc() will be written by the caller,
(but not all callers want to write all the object, but just at the
begining) prefetchw() tells the modern CPU to think about the future
writes, ie start some memory transactions in advance.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
This patch puts mostly read only data in the right section
(read_mostly), to help sharing of these data between CPUS without
memory ping pongs.
On one of my production machine, tcp_statistics was sitting in a
heavily modified cache line, so *every* SNMP update had to force a
reload.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
A kernel BUG() is triggered by a call to set_mempolicy() with a negative
first argument. This is because the mode is declared as an int, and the
validity check doesnt check < 0 values. Alternatively, mode could be
declared as unsigned int or unsigned long.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Andi Kleen <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
- rt_check_expire() fixes (an overflow occured if size of the hash
was >= 65536)
reminder of the bugfix:
The rt_check_expire() has a serious problem on machines with large
route caches, and a standard HZ value of 1000.
With default values, ie ip_rt_gc_interval = 60*HZ = 60000 ;
the loop count :
for (t = ip_rt_gc_interval << rt_hash_log; t >= 0;
overflows (t is a 31 bit value) as soon rt_hash_log is >= 16 (65536
slots in route cache hash table).
In this case, rt_check_expire() does nothing at all
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
- rt hash table allocated using alloc_large_system_hash() function.
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
- Locking abstraction
- Spinlocks moved out of rt hash table : Less memory (50%) used by rt
hash table. it's a win even on UP.
- Sizing of spinlocks table depends on NR_CPUS
Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: David S. Miller <[email protected]>
Eric Dumazet