13 Mar, 2010
15 commits
-
In current page-fault code,
handle_mm_fault()
-> ...
-> mem_cgroup_charge()
-> map page or handle error.
-> check return code.If page fault's return code is VM_FAULT_OOM, page_fault_out_of_memory() is
called. But if it's caused by memcg, OOM should have been already
invoked.Then, I added a patch: a636b327f731143ccc544b966cfd8de6cb6d72c6. That
patch records last_oom_jiffies for memcg's sub-hierarchy and prevents
page_fault_out_of_memory from being invoked in near future.But Nishimura-san reported that check by jiffies is not enough when the
system is terribly heavy.This patch changes memcg's oom logic as.
* If memcg causes OOM-kill, continue to retry.
* remove jiffies check which is used now.
* add memcg-oom-lock which works like perzone oom lock.
* If current is killed(as a process), bypass charge.Something more sophisticated can be added but this pactch does
fundamental things.
TODO:
- add oom notifier
- add permemcg disable-oom-kill flag and freezer at oom.
- more chances for wake up oom waiter (when changing memory limit etc..)Reviewed-by: Daisuke Nishimura
Tested-by: Daisuke Nishimura
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Cc: David Rientjes
Signed-off-by: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Events should be removed after rmdir of cgroup directory, but before
destroying subsystem state objects. Let's take reference to cgroup
directory dentry to do that.Signed-off-by: Kirill A. Shutemov
Acked-by: KAMEZAWA Hiroyuki
Cc: Paul Menage
Acked-by: Li Zefan
Cc: Balbir Singh
Cc: Pavel Emelyanov
Cc: Dan Malek
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Memcg has 2 eventcountes which counts "the same" event. Just usages are
different from each other. This patch tries to reduce event counter.Now logic uses "only increment, no reset" counter and masks for each
checks. Softlimit chesk was done per 1000 evetns. So, the similar check
can be done by !(new_counter & 0x3ff). Threshold check was done per 100
events. So, the similar check can be done by (!new_counter & 0x7f)ALL event checks are done right after EVENT percpu counter is updated.
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Kirill A. Shutemov
Cc: Balbir Singh
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Presently, move_task does "batched" precharge. Because res_counter or
css's refcnt are not-scalable jobs for memcg, try_charge_().. tend to be
done in batched manner if allowed.Now, softlimit and threshold check their event counter in try_charge, but
the charge is not a per-page event. And event counter is not updated at
charge(). Moreover, precharge doesn't pass "page" to try_charge() and
softlimit tree will be never updated until uncharge() causes an event."So the best place to check the event counter is commit_charge(). This is
per-page event by its nature. This patch move checks to there.Signed-off-by: KAMEZAWA Hiroyuki
Cc: Kirill A. Shutemov
Cc: Balbir Singh
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
When per-cpu counter for memcg was implemneted, dynamic percpu allocator
was not very good. But now, we have good one and useful macros. This
patch replaces memcg's private percpu counter implementation with generic
dynamic percpu allocator.The benefits are
- We can remove private implementation.
- The counters will be NUMA-aware. (Current one is not...)
- This patch makes sizeof struct mem_cgroup smaller. Then,
struct mem_cgroup may be fit in page size on small config.
- About basic performance aspects, see below.[Before]
# size mm/memcontrol.o
text data bss dec hex filename
24373 2528 4132 31033 7939 mm/memcontrol.o[page-fault-throuput test on 8cpu/SMP in root cgroup]
# /root/bin/perf stat -a -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8Performance counter stats for './multi-fault-fork 8' (5 runs):
45878618 page-faults ( +- 0.110% )
602635826 cache-misses ( +- 0.105% )61.005373262 seconds time elapsed ( +- 0.004% )
Then cache-miss/page fault = 13.14
[After]
#size mm/memcontrol.o
text data bss dec hex filename
23913 2528 4132 30573 776d mm/memcontrol.o
# /root/bin/perf stat -a -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8Performance counter stats for './multi-fault-fork 8' (5 runs):
48179400 page-faults ( +- 0.271% )
588628407 cache-misses ( +- 0.136% )61.004615021 seconds time elapsed ( +- 0.004% )
Then cache-miss/page fault = 12.22
Text size is reduced.
This performance improvement is not big and will be invisible in real world
applications. But this result shows this patch has some good effect even
on (small) SMP.Here is a test program I used.
1. fork() processes on each cpus.
2. do page fault repeatedly on each process.
3. after 60secs, kill all childredn and exit.(3 is necessary for getting stable data, this is improvement from previous one.)
#define _GNU_SOURCE
#include
#include
#include
#include
#include
#include
#include
#include/*
* For avoiding contention in page table lock, FAULT area is
* sparse. If FAULT_LENGTH is too large for your cpus, decrease it.
*/
#define FAULT_LENGTH (2 * 1024 * 1024)
#define PAGE_SIZE 4096
#define MAXNUM (128)void alarm_handler(int sig)
{
}void *worker(int cpu, int ppid)
{
void *start, *end;
char *c;
cpu_set_t set;
int i;CPU_ZERO(&set);
CPU_SET(cpu, &set);
sched_setaffinity(0, sizeof(set), &set);start = mmap(NULL, FAULT_LENGTH, PROT_READ|PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
if (start == MAP_FAILED) {
perror("mmap");
exit(1);
}
end = start + FAULT_LENGTH;pause();
//fprintf(stderr, "run%d", cpu);
while (1) {
for (c = (char*)start; (void *)c < end; c += PAGE_SIZE)
*c = 0;
madvise(start, FAULT_LENGTH, MADV_DONTNEED);
}
return NULL;
}int main(int argc, char *argv[])
{
int num, i, ret, pid, status;
int pids[MAXNUM];if (argc < 2)
return 0;setpgid(0, 0);
signal(SIGALRM, alarm_handler);
num = atoi(argv[1]);
pid = getpid();for (i = 0; i < num; ++i) {
ret = fork();
if (!ret) {
worker(i, pid);
exit(0);
}
pids[i] = ret;
}
sleep(1);
kill(-pid, SIGALRM);
sleep(60);
for (i = 0; i < num; i++)
kill(pids[i], SIGKILL);
for (i = 0; i < num; i++)
waitpid(pids[i], &status, 0);
return 0;
}Signed-off-by: KAMEZAWA Hiroyuki
Cc: Daisuke Nishimura
Cc: Balbir Singh
Cc: Pavel Emelyanov
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
s/mem_cgroup_print_mem_info/mem_cgroup_print_oom_info/
Signed-off-by: Kirill A. Shutemov
Cc: Balbir Singh
Cc: Pavel Emelyanov
Cc: KAMEZAWA Hiroyuki
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
It allows to register multiple memory and memsw thresholds and gets
notifications when it crosses.To register a threshold application need:
- create an eventfd;
- open memory.usage_in_bytes or memory.memsw.usage_in_bytes;
- write string like " " to
cgroup.event_control.Application will be notified through eventfd when memory usage crosses
threshold in any direction.It's applicable for root and non-root cgroup.
It uses stats to track memory usage, simmilar to soft limits. It checks
if we need to send event to userspace on every 100 page in/out. I guess
it's good compromise between performance and accuracy of thresholds.[akpm@linux-foundation.org: coding-style fixes]
[nishimura@mxp.nes.nec.co.jp: fix documentation merge issue]
Signed-off-by: Kirill A. Shutemov
Cc: Li Zefan
Cc: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Cc: Pavel Emelyanov
Cc: Dan Malek
Cc: Vladislav Buzov
Cc: Daisuke Nishimura
Cc: Alexander Shishkin
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Instead of incrementing counter on each page in/out and comparing it with
constant, we set counter to constant, decrement counter on each page
in/out and compare it with zero. We want to make comparing as fast as
possible. On many RISC systems (probably not only RISC) comparing with
zero is more effective than comparing with a constant, since not every
constant can be immediate operand for compare instruction.Also, I've renamed MEM_CGROUP_STAT_EVENTS to MEM_CGROUP_STAT_SOFTLIMIT,
since really it's not a generic counter.Signed-off-by: Kirill A. Shutemov
Cc: Li Zefan
Cc: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Cc: Pavel Emelyanov
Cc: Dan Malek
Cc: Vladislav Buzov
Cc: Daisuke Nishimura
Cc: Alexander Shishkin
Cc: Davide Libenzi
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Helper to get memory or mem+swap usage of the cgroup.
Signed-off-by: Kirill A. Shutemov
Acked-by: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Pavel Emelyanov
Cc: Dan Malek
Cc: Vladislav Buzov
Cc: Daisuke Nishimura
Cc: Alexander Shishkin
Cc: Davide Libenzi
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Try to reduce overheads in moving swap charge by:
- Adds a new function(__mem_cgroup_put), which takes "count" as a arg and
decrement mem->refcnt by "count".
- Removed res_counter_uncharge, css_put, and mem_cgroup_put from the path
of moving swap account, and consolidate all of them into mem_cgroup_clear_mc.
We cannot do that about mc.to->refcnt.These changes reduces the overhead from 1.35sec to 0.9sec to move charges
of 1G anonymous memory(including 500MB swap) in my test environment.Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
This patch is another core part of this move-charge-at-task-migration
feature. It enables moving charges of anonymous swaps.To move the charge of swap, we need to exchange swap_cgroup's record.
In current implementation, swap_cgroup's record is protected by:
- page lock: if the entry is on swap cache.
- swap_lock: if the entry is not on swap cache.This works well in usual swap-in/out activity.
But this behavior make the feature of moving swap charge check many
conditions to exchange swap_cgroup's record safely.So I changed modification of swap_cgroup's recored(swap_cgroup_record())
to use xchg, and define a new function to cmpxchg swap_cgroup's record.This patch also enables moving charge of non pte_present but not uncharged
swap caches, which can be exist on swap-out path, by getting the target
pages via find_get_page() as do_mincore() does.[kosaki.motohiro@jp.fujitsu.com: fix ia64 build]
[akpm@linux-foundation.org: fix typos]
Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: KOSAKI Motohiro
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
This move-charge-at-task-migration feature has extra charges on
"to"(pre-charges) and "from"(left-over charges) during moving charge.
This means unnecessary oom can happen.This patch tries to avoid such oom.
Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Try to reduce overheads in moving charge by:
- Instead of calling res_counter_uncharge() against the old cgroup in
__mem_cgroup_move_account() everytime, call res_counter_uncharge() at the end
of task migration once.
- removed css_get(&to->css) from __mem_cgroup_move_account() because callers
should have already called css_get(). And removed css_put(&to->css) too,
which was called by callers of move_account on success of move_account.
- Instead of calling __mem_cgroup_try_charge(), i.e. res_counter_charge(),
repeatedly, call res_counter_charge(PAGE_SIZE * count) in can_attach() if
possible.
- Instead of calling css_get()/css_put() repeatedly, make use of coalesce
__css_get()/__css_put() if possible.These changes reduces the overhead from 1.7sec to 0.6sec to move charges
of 1G anonymous memory in my test environment.Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
This patch is the core part of this move-charge-at-task-migration feature.
It implements functions to move charges of anonymous pages mapped only by
the target task.Implementation:
- define struct move_charge_struct and a valuable of it(mc) to remember the
count of pre-charges and other information.
- At can_attach(), get anon_rss of the target mm, call __mem_cgroup_try_charge()
repeatedly and count up mc.precharge.
- At attach(), parse the page table, find a target page to be move, and call
mem_cgroup_move_account() about the page.
- Cancel all precharges if mc.precharge > 0 on failure or at the end of
task move.[akpm@linux-foundation.org: a little simplification]
Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
In current memcg, charges associated with a task aren't moved to the new
cgroup at task migration. Some users feel this behavior to be strange.
These patches are for this feature, that is, for charging to the new
cgroup and, of course, uncharging from the old cgroup at task migration.This patch adds "memory.move_charge_at_immigrate" file, which is a flag
file to determine whether charges should be moved to the new cgroup at
task migration or not and what type of charges should be moved. This
patch also adds read and write handlers of the file.This patch also adds no-op handlers for this feature. These handlers will
be implemented in later patches. And you cannot write any values other
than 0 to move_charge_at_immigrate yet.Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: Paul Menage
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
07 Mar, 2010
1 commit
-
mm/memcontrol.c:2548:32: warning: Using plain integer as NULL pointer
Signed-off-by: Thiago Farina
Acked-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
17 Jan, 2010
1 commit
-
Current mem_cgroup_force_empty() only ensures mem->res.usage == 0 on
success. But this doesn't guarantee memcg's LRU is really empty, because
there are some cases in which !PageCgrupUsed pages exist on memcg's LRU.For example:
- Pages can be uncharged by its owner process while they are on LRU.
- race between mem_cgroup_add_lru_list() and __mem_cgroup_uncharge_common().So there can be a case in which the usage is zero but some of the LRUs are not empty.
OTOH, mem_cgroup_del_lru_list(), which can be called asynchronously with
rmdir, accesses the mem_cgroup, so this access can cause a problem if it
races with rmdir because the mem_cgroup might have been freed by rmdir.Actually, I saw a bug which seems to be caused by this race.
[1530745.949906] BUG: unable to handle kernel NULL pointer dereference at 0000000000000230
[1530745.950651] IP: [] mem_cgroup_del_lru_list+0x30/0x80
[1530745.950651] PGD 3863de067 PUD 3862c7067 PMD 0
[1530745.950651] Oops: 0002 [#1] SMP
[1530745.950651] last sysfs file: /sys/devices/system/cpu/cpu7/cache/index1/shared_cpu_map
[1530745.950651] CPU 3
[1530745.950651] Modules linked in: configs ipt_REJECT xt_tcpudp iptable_filter ip_tables x_tables bridge stp nfsd nfs_acl auth_rpcgss exportfs autofs4 hidp rfcomm l2cap crc16 bluetooth lockd sunrpc ib_iser rdma_cm ib_cm iw_cm ib_sa ib_mad ib_core ib_addr iscsi_tcp bnx2i cnic uio ipv6 cxgb3i cxgb3 mdio libiscsi_tcp libiscsi scsi_transport_iscsi dm_mirror dm_multipath scsi_dh video output sbs sbshc battery ac lp kvm_intel kvm sg ide_cd_mod cdrom serio_raw tpm_tis tpm tpm_bios acpi_memhotplug button parport_pc parport rtc_cmos rtc_core rtc_lib e1000 i2c_i801 i2c_core pcspkr dm_region_hash dm_log dm_mod ata_piix libata shpchp megaraid_mbox sd_mod scsi_mod megaraid_mm ext3 jbd uhci_hcd ohci_hcd ehci_hcd [last unloaded: freq_table]
[1530745.950651] Pid: 19653, comm: shmem_test_02 Tainted: G M 2.6.32-mm1-00701-g2b04386 #3 Express5800/140Rd-4 [N8100-1065]
[1530745.950651] RIP: 0010:[] [] mem_cgroup_del_lru_list+0x30/0x80
[1530745.950651] RSP: 0018:ffff8803863ddcb8 EFLAGS: 00010002
[1530745.950651] RAX: 00000000000001e0 RBX: ffff8803abc02238 RCX: 00000000000001e0
[1530745.950651] RDX: 0000000000000000 RSI: ffff88038611a000 RDI: ffff8803abc02238
[1530745.950651] RBP: ffff8803863ddcc8 R08: 0000000000000002 R09: ffff8803a04c8643
[1530745.950651] R10: 0000000000000000 R11: ffffffff810c7333 R12: 0000000000000000
[1530745.950651] R13: ffff880000017f00 R14: 0000000000000092 R15: ffff8800179d0310
[1530745.950651] FS: 0000000000000000(0000) GS:ffff880017800000(0000) knlGS:0000000000000000
[1530745.950651] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[1530745.950651] CR2: 0000000000000230 CR3: 0000000379d87000 CR4: 00000000000006e0
[1530745.950651] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[1530745.950651] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
[1530745.950651] Process shmem_test_02 (pid: 19653, threadinfo ffff8803863dc000, task ffff88038612a8a0)
[1530745.950651] Stack:
[1530745.950651] ffffea00040c2fe8 0000000000000000 ffff8803863ddd98 ffffffff810c739a
[1530745.950651] 00000000863ddd18 000000000000000c 0000000000000000 0000000000000000
[1530745.950651] 0000000000000002 0000000000000000 ffff8803863ddd68 0000000000000046
[1530745.950651] Call Trace:
[1530745.950651] [] release_pages+0x142/0x1e7
[1530745.950651] [] ? pagevec_move_tail+0x6e/0x112
[1530745.950651] [] pagevec_move_tail+0xfd/0x112
[1530745.950651] [] lru_add_drain+0x76/0x94
[1530745.950651] [] exit_mmap+0x6e/0x145
[1530745.950651] [] mmput+0x5e/0xcf
[1530745.950651] [] exit_mm+0x11c/0x129
[1530745.950651] [] ? audit_free+0x196/0x1c9
[1530745.950651] [] do_exit+0x1f5/0x6b7
[1530745.950651] [] ? up_read+0x2b/0x2f
[1530745.950651] [] ? lockdep_sys_exit_thunk+0x35/0x67
[1530745.950651] [] do_group_exit+0x83/0xb0
[1530745.950651] [] sys_exit_group+0x17/0x1b
[1530745.950651] [] system_call_fastpath+0x16/0x1b
[1530745.950651] Code: 54 53 0f 1f 44 00 00 83 3d cc 29 7c 00 00 41 89 f4 75 63 eb 4e 48 83 7b 08 00 75 04 0f 0b eb fe 48 89 df e8 18 f3 ff ff 44 89 e2 ff 4c d0 50 48 8b 05 2b 2d 7c 00 48 39 43 08 74 39 48 8b 4b
[1530745.950651] RIP [] mem_cgroup_del_lru_list+0x30/0x80
[1530745.950651] RSP
[1530745.950651] CR2: 0000000000000230
[1530745.950651] ---[ end trace c3419c1bb8acc34f ]---
[1530745.950651] Fixing recursive fault but reboot is needed!The problem here is pages on LRU may contain pointer to stale memcg. To
make res->usage to be 0, all pages on memcg must be uncharged or moved to
another(parent) memcg. Moved page_cgroup have already removed from
original LRU, but uncharged page_cgroup contains pointer to memcg withou
PCG_USED bit. (This asynchronous LRU work is for improving performance.)
If PCG_USED bit is not set, page_cgroup will never be added to memcg's
LRU. So, about pages not on LRU, they never access stale pointer. Then,
what we have to take care of is page_cgroup _on_ LRU list. This patch
fixes this problem by making mem_cgroup_force_empty() visit all LRUs
before exiting its loop and guarantee there are no pages on its LRU.Signed-off-by: Daisuke Nishimura
Acked-by: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Cc:
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
17 Dec, 2009
1 commit
-
* 'hwpoison' of git://git.kernel.org/pub/scm/linux/kernel/git/ak/linux-mce-2.6: (34 commits)
HWPOISON: Remove stray phrase in a comment
HWPOISON: Try to allocate migration page on the same node
HWPOISON: Don't do early filtering if filter is disabled
HWPOISON: Add a madvise() injector for soft page offlining
HWPOISON: Add soft page offline support
HWPOISON: Undefine short-hand macros after use to avoid namespace conflict
HWPOISON: Use new shake_page in memory_failure
HWPOISON: Use correct name for MADV_HWPOISON in documentation
HWPOISON: mention HWPoison in Kconfig entry
HWPOISON: Use get_user_page_fast in hwpoison madvise
HWPOISON: add an interface to switch off/on all the page filters
HWPOISON: add memory cgroup filter
memcg: add accessor to mem_cgroup.css
memcg: rename and export try_get_mem_cgroup_from_page()
HWPOISON: add page flags filter
mm: export stable page flags
HWPOISON: limit hwpoison injector to known page types
HWPOISON: add fs/device filters
HWPOISON: return 0 to indicate success reliably
HWPOISON: make semantics of IGNORED/DELAYED clear
...
16 Dec, 2009
12 commits
-
Variable `progress' isn't used in mem_cgroup_resize_limit() any more.
Remove it.[akpm@linux-foundation.org: cleanup]
Signed-off-by: Bob Liu
Cc: Daisuke Nishimura
Reviewed-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
memcg_tasklist was introduced at commit 7f4d454d(memcg: avoid deadlock
caused by race between oom and cpuset_attach) instead of cgroup_mutex to
fix a deadlock problem. The cgroup_mutex, which was removed by the
commit, in mem_cgroup_out_of_memory() was originally introduced at commit
c7ba5c9e (Memory controller: OOM handling).IIUC, the intention of this cgroup_mutex was to prevent task move during
select_bad_process() so that situations like below can be avoided.Assume cgroup "foo" has exceeded its limit and is about to trigger oom.
1. Process A, which has been in cgroup "baa" and uses large memory, is just
moved to cgroup "foo". Process A can be the candidates for being killed.
2. Process B, which has been in cgroup "foo" and uses large memory, is just
moved from cgroup "foo". Process B can be excluded from the candidates for
being killed.But these race window exists anyway even if we hold a lock, because
__mem_cgroup_try_charge() decides wether it should trigger oom or not
outside of the lock. So the original cgroup_mutex in
mem_cgroup_out_of_memory and thus current memcg_tasklist has no use. And
IMHO, those races are not so critical for users.This patch removes it and make codes simpler.
Signed-off-by: Daisuke Nishimura
Cc: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
task_in_mem_cgroup(), which is called by select_bad_process() to check
whether a task can be a candidate for being oom-killed from memcg's limit,
checks "curr->use_hierarchy"("curr" is the mem_cgroup the task belongs
to).But this check return true(it's false positive) when:
/aa use_hierarchy == 0 /aa/00 use_hierarchy == 1
Acked-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
mem_cgroup_move_parent() calls try_charge first and cancel_charge on
failure. IMHO, charge/uncharge(especially charge) is high cost operation,
so we should avoid it as far as possible.This patch tries to delay try_charge in mem_cgroup_move_parent() by
re-ordering checks it does.And this patch renames mem_cgroup_move_account() to
__mem_cgroup_move_account(), changes the return value of
__mem_cgroup_move_account() from int to void, and adds a new
wrapper(mem_cgroup_move_account()), which checks whether a @pc is valid
for moving account and calls __mem_cgroup_move_account().This patch removes the last caller of trylock_page_cgroup(), so removes
its definition too.Signed-off-by: Daisuke Nishimura
Acked-by: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
There are some places calling both res_counter_uncharge() and css_put() to
cancel the charge and the refcnt we have got by mem_cgroup_tyr_charge().This patch introduces mem_cgroup_cancel_charge() and call it in those
places.Signed-off-by: KAMEZAWA Hiroyuki
Signed-off-by: Daisuke Nishimura
Reviewed-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
In global VM, FILE_MAPPED is used but memcg uses MAPPED_FILE. This makes
grep difficult. Replace memcg's MAPPED_FILE with FILE_MAPPEDAnd in global VM, mapped shared memory is accounted into FILE_MAPPED.
But memcg doesn't. fix it.
Note:
page_is_file_cache() just checks SwapBacked or not.
So, we need to check PageAnon.Cc: Balbir Singh
Reviewed-by: Daisuke Nishimura
Signed-off-by: KAMEZAWA Hiroyuki
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
This is a patch for coalescing access to res_counter at charging by percpu
caching. At charge, memcg charges 64pages and remember it in percpu
cache. Because it's cache, drain/flush if necessary.This version uses public percpu area.
2 benefits for using public percpu area.
1. Sum of stocked charge in the system is limited to # of cpus
not to the number of memcg. This shows better synchonization.
2. drain code for flush/cpuhotplug is very easy (and quick)The most important point of this patch is that we never touch res_counter
in fast path. The res_counter is system-wide shared counter which is modified
very frequently. We shouldn't touch it as far as we can for avoiding
false sharing.On x86-64 8cpu server, I tested overheads of memcg at page fault by
running a program which does map/fault/unmap in a loop. Running
a task per a cpu by taskset and see sum of the number of page faults
in 60secs.[without memcg config]
40156968 page-faults # 0.085 M/sec ( +- 0.046% )
27.67 cache-miss/faults[root cgroup]
36659599 page-faults # 0.077 M/sec ( +- 0.247% )
31.58 cache miss/faults[in a child cgroup]
18444157 page-faults # 0.039 M/sec ( +- 0.133% )
69.96 cache miss/faults[ + coalescing uncharge patch]
27133719 page-faults # 0.057 M/sec ( +- 0.155% )
47.16 cache miss/faults[ + coalescing uncharge patch + this patch ]
34224709 page-faults # 0.072 M/sec ( +- 0.173% )
34.69 cache miss/faultsChangelog (since Oct/2):
- updated comments
- replaced get_cpu_var() with __get_cpu_var() if possible.
- removed mutex for system-wide drain. adds a counter instead of it.
- removed CONFIG_HOTPLUG_CPUChangelog (old):
- rebased onto the latest mmotm
- moved charge size check before __GFP_WAIT check for avoiding unnecesary
- added asynchronous flush routine.
- fixed bugs pointed out by Nishimura-san.[akpm@linux-foundation.org: tweak comments]
[nishimura@mxp.nes.nec.co.jp: don't do INIT_WORK() repeatedly against the same work_struct]
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Signed-off-by: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
In massive parallel enviroment, res_counter can be a performance
bottleneck. One strong techinque to reduce lock contention is reducing
calls by coalescing some amount of calls into one.Considering charge/uncharge chatacteristic,
- charge is done one by one via demand-paging.
- uncharge is done by
- in chunk at munmap, truncate, exit, execve...
- one by one via vmscan/paging.It seems we have a chance to coalesce uncharges for improving scalability
at unmap/truncation.This patch is a for coalescing uncharge. For avoiding scattering memcg's
structure to functions under /mm, this patch adds memcg batch uncharge
information to the task. A reason for per-task batching is for making use
of caller's context information. We do batched uncharge (deleyed
uncharge) when truncation/unmap occurs but do direct uncharge when
uncharge is called by memory reclaim (vmscan.c).The degree of coalescing depends on callers
- at invalidate/trucate... pagevec size
- at unmap ....ZAP_BLOCK_SIZE
(memory itself will be freed in this degree.)
Then, we'll not coalescing too much.On x86-64 8cpu server, I tested overheads of memcg at page fault by
running a program which does map/fault/unmap in a loop. Running
a task per a cpu by taskset and see sum of the number of page faults
in 60secs.[without memcg config]
40156968 page-faults # 0.085 M/sec ( +- 0.046% )
27.67 cache-miss/faults
[root cgroup]
36659599 page-faults # 0.077 M/sec ( +- 0.247% )
31.58 miss/faults
[in a child cgroup]
18444157 page-faults # 0.039 M/sec ( +- 0.133% )
69.96 miss/faults
[child with this patch]
27133719 page-faults # 0.057 M/sec ( +- 0.155% )
47.16 miss/faultsWe can see some amounts of improvement.
(root cgroup doesn't affected by this patch)
Another patch for "charge" will follow this and above will be improved more.Changelog(since 2009/10/02):
- renamed filed of memcg_batch (as pages to bytes, memsw to memsw_bytes)
- some clean up and commentary/description updates.
- added initialize code to copy_process(). (possible bug fix)Changelog(old):
- fixed !CONFIG_MEM_CGROUP case.
- rebased onto the latest mmotm + softlimit fix patches.
- unified patch for callers
- added commetns.
- make ->do_batch as bool.
- removed css_get() at el. We don't need it.Signed-off-by: KAMEZAWA Hiroyuki
Cc: Balbir Singh
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
A memory cgroup has a memory.memsw.usage_in_bytes file. It shows the sum
of the usage of pages and swapents in the cgroup. Presently the root
cgroup's memsw.usage_in_bytes shows the wrong value - the number of
swapents are not added.So take MEM_CGROUP_STAT_SWAPOUT into account.
Signed-off-by: Kirill A. Shutemov
Reviewed-by: Daisuke Nishimura
Acked-by: KAMEZAWA Hiroyuki
Cc:
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
So that an outside user can free the reference count grabbed by
try_get_mem_cgroup_from_page().CC: KOSAKI Motohiro
CC: Hugh Dickins
CC: Daisuke Nishimura
CC: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Signed-off-by: Wu Fengguang
Signed-off-by: Andi Kleen -
So that the hwpoison injector can get mem_cgroup for arbitrary page
and thus know whether it is owned by some mem_cgroup task(s).[AK: Merged with latest git tree]
CC: KOSAKI Motohiro
CC: Hugh Dickins
CC: Daisuke Nishimura
CC: Balbir Singh
Acked-by: KAMEZAWA Hiroyuki
Signed-off-by: Wu Fengguang
Signed-off-by: Andi Kleen -
But ksm swapping does require one small change in mem cgroup handling.
When do_swap_page()'s call to ksm_might_need_to_copy() does indeed
substitute a duplicate page to accommodate a different anon_vma (or a the
!PageSwapCache check in mem_cgroup_try_charge_swapin().That was returning success without charging, on the assumption that
pte_same() would fail after, which is not the case here. Originally I
proposed that success, so that an unshrinkable mem cgroup at its limit
would not fail unnecessarily; but that's a minor point, and there are
plenty of other places where we may fail an overallocation which might
later prove unnecessary. So just go ahead and do what all the other
exceptions do: proceed to charge current mm.Signed-off-by: Hugh Dickins
Cc: Izik Eidus
Cc: Andrea Arcangeli
Cc: Chris Wright
Acked-by: KAMEZAWA Hiroyuki
Acked-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
04 Dec, 2009
1 commit
-
That is "success", "unknown", "through", "performance", "[re|un]mapping"
, "access", "default", "reasonable", "[con]currently", "temperature"
, "channel", "[un]used", "application", "example","hierarchy", "therefore"
, "[over|under]flow", "contiguous", "threshold", "enough" and others.Signed-off-by: André Goddard Rosa
Signed-off-by: Jiri Kosina
09 Nov, 2009
1 commit
-
This patch was generated by
git grep -E -i -l '[Aa]quire' | xargs -r perl -p -i -e 's/([Aa])quire/$1cquire/'
and the cumsumed was found by checking the diff for aquire.
Signed-off-by: Uwe Kleine-König
Signed-off-by: Jiri Kosina
02 Oct, 2009
3 commits
-
In charge/uncharge/reclaim path, usage_in_excess is calculated repeatedly
and it takes res_counter's spin_lock every time.This patch removes unnecessary calls for res_count_soft_limit_excess.
Reviewed-by: Daisuke Nishimura
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Paul Menage
Cc: Li Zefan
Cc: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
This patch clean up/fixes for memcg's uncharge soft limit path.
Problems:
Now, res_counter_charge()/uncharge() handles softlimit information at
charge/uncharge and softlimit-check is done when event counter per memcg
goes over limit. Now, event counter per memcg is updated only when
memory usage is over soft limit. Here, considering hierarchical memcg
management, ancesotors should be taken care of.Now, ancerstors(hierarchy) are handled in charge() but not in uncharge().
This is not good.Prolems:
1. memcg's event counter incremented only when softlimit hits. That's bad.
It makes event counter hard to be reused for other purpose.2. At uncharge, only the lowest level rescounter is handled. This is bug.
Because ancesotor's event counter is not incremented, children should
take care of them.3. res_counter_uncharge()'s 3rd argument is NULL in most case.
ops under res_counter->lock should be small. No "if" sentense is better.Fixes:
* Removed soft_limit_xx poitner and checks in charge and uncharge.
Do-check-only-when-necessary scheme works enough well without them.* make event-counter of memcg incremented at every charge/uncharge.
(per-cpu area will be accessed soon anyway)* All ancestors are checked at soft-limit-check. This is necessary because
ancesotor's event counter may never be modified. Then, they should be
checked at the same time.Reviewed-by: Daisuke Nishimura
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Paul Menage
Cc: Li Zefan
Cc: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
__mem_cgroup_largest_soft_limit_node() returns a mem_cgroup_per_zone "mz"
with incremnted mz->mem->css's refcnt. Then, the caller of this function
has to call css_put(mz->mem->css).But, mz can be !NULL even if "not found" i.e. without css_get(). By
this, css->refcnt will go down to minus.This may cause various things...one of results will be
initite-loop in css_tryget() as this.INFO: RCU detected CPU 0 stall (t=10000 jiffies)
sending NMI to all CPUs:
NMI backtrace for cpu 0
CPU 0:<> [] trace_hardirqs_off+0xd/0x10
[] flat_send_IPI_mask+0x90/0xb0
[] flat_send_IPI_all+0x69/0x70
[] arch_trigger_all_cpu_backtrace+0x62/0xa0
[] __rcu_pending+0x7e/0x370
[] rcu_check_callbacks+0x47/0x130
[] update_process_times+0x46/0x70
[] tick_sched_timer+0x60/0x160
[] ? tick_sched_timer+0x0/0x160
[] __run_hrtimer+0xba/0x150
[] hrtimer_interrupt+0xd5/0x1b0
[] ? trace_hardirqs_off_thunk+0x3a/0x3c
[] smp_apic_timer_interrupt+0x6d/0x9b
[] apic_timer_interrupt+0x13/0x20
[] ? mem_cgroup_walk_tree+0x156/0x180
[] ? mem_cgroup_walk_tree+0x73/0x180
[] ? mem_cgroup_walk_tree+0x32/0x180
[] ? mem_cgroup_get_local_stat+0x0/0x110
[] ? mem_control_stat_show+0x14b/0x330
[] ? cgroup_seqfile_show+0x3d/0x60Above shows CPU0 caught in css_tryget()'s inifinite loop because
of bad refcnt.This is a fix to set mz=NULL at the top of retry path.
Signed-off-by: KAMEZAWA Hiroyuki
Acked-by: Paul Menage
Cc: Li Zefan
Cc: Balbir Singh
Cc: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
24 Sep, 2009
5 commits
-
We now count MEM_CGROUP_STAT_SWAPOUT, so we can show swap usage. It would
be useful for users to show swap usage in memory.stat file, because they
don't need calculate memsw.usage - res.usage to know swap usage.Signed-off-by: Daisuke Nishimura
Reviewed-by: KAMEZAWA Hiroyuki
Reviewed-by: Balbir Singh
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Reduce the resource counter overhead (mostly spinlock) associated with the
root cgroup. This is a part of the several patches to reduce mem cgroup
overhead. I had posted other approaches earlier (including using percpu
counters). Those patches will be a natural addition and will be added
iteratively on top of these.The patch stops resource counter accounting for the root cgroup. The data
for display is derived from the statisitcs we maintain via
mem_cgroup_charge_statistics (which is more scalable). What happens today
is that, we do double accounting, once using res_counter_charge() and once
using memory_cgroup_charge_statistics(). For the root, since we don't
implement limits any more, we don't need to track every charge via
res_counter_charge() and check for limit being exceeded and reclaim.The main mem->res usage_in_bytes can be derived by summing the cache and
rss usage data from memory statistics (MEM_CGROUP_STAT_RSS and
MEM_CGROUP_STAT_CACHE). However, for memsw->res usage_in_bytes, we need
additional data about swapped out memory. This patch adds a
MEM_CGROUP_STAT_SWAPOUT and uses that along with MEM_CGROUP_STAT_RSS and
MEM_CGROUP_STAT_CACHE to derive the memsw data. This data is computed
recursively when hierarchy is enabled.The tests results I see on a 24 way show that
1. The lock contention disappears from /proc/lock_stats
2. The results of the test are comparable to running with
cgroup_disable=memory.Here is a sample of my program runs
Without Patch
Performance counter stats for '/home/balbir/parallel_pagefault':
7192804.124144 task-clock-msecs # 23.937 CPUs
424691 context-switches # 0.000 M/sec
267 CPU-migrations # 0.000 M/sec
28498113 page-faults # 0.004 M/sec
5826093739340 cycles # 809.989 M/sec
408883496292 instructions # 0.070 IPC
7057079452 cache-references # 0.981 M/sec
3036086243 cache-misses # 0.422 M/sec300.485365680 seconds time elapsed
With cgroup_disable=memory
Performance counter stats for '/home/balbir/parallel_pagefault':
7182183.546587 task-clock-msecs # 23.915 CPUs
425458 context-switches # 0.000 M/sec
203 CPU-migrations # 0.000 M/sec
92545093 page-faults # 0.013 M/sec
6034363609986 cycles # 840.185 M/sec
437204346785 instructions # 0.072 IPC
6636073192 cache-references # 0.924 M/sec
2358117732 cache-misses # 0.328 M/sec300.320905827 seconds time elapsed
With this patch applied
Performance counter stats for '/home/balbir/parallel_pagefault':
7191619.223977 task-clock-msecs # 23.955 CPUs
422579 context-switches # 0.000 M/sec
88 CPU-migrations # 0.000 M/sec
91946060 page-faults # 0.013 M/sec
5957054385619 cycles # 828.333 M/sec
1058117350365 instructions # 0.178 IPC
9161776218 cache-references # 1.274 M/sec
1920494280 cache-misses # 0.267 M/sec300.218764862 seconds time elapsed
Data from Prarit (kernel compile with make -j64 on a 64
CPU/32G machine)For a single run
Without patch
real 27m8.988s
user 87m24.916s
sys 382m6.037sWith patch
real 4m18.607s
user 84m58.943s
sys 50m52.682sWith config turned off
real 4m54.972s
user 90m13.456s
sys 50m19.711sNOTE: The data looks counterintuitive due to the increased performance
with the patch, even over the config being turned off. We probably need
more runs, but so far all testing has shown that the patches definitely
help.[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Balbir Singh
Cc: Prarit Bhargava
Cc: Andi Kleen
Reviewed-by: KAMEZAWA Hiroyuki
Reviewed-by: Daisuke Nishimura
Cc: KOSAKI Motohiro
Cc: Paul Menage
Cc: Li Zefan
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Implement reclaim from groups over their soft limit
Permit reclaim from memory cgroups on contention (via the direct reclaim
path).memory cgroup soft limit reclaim finds the group that exceeds its soft
limit by the largest number of pages and reclaims pages from it and then
reinserts the cgroup into its correct place in the rbtree.Add additional checks to mem_cgroup_hierarchical_reclaim() to detect long
loops in case all swap is turned off. The code has been refactored and
the loop check (loop < 2) has been enhanced for soft limits. For soft
limits, we try to do more targetted reclaim. Instead of bailing out after
two loops, the routine now reclaims memory proportional to the size by
which the soft limit is exceeded. The proportion has been empirically
determined.[akpm@linux-foundation.org: build fix]
[kamezawa.hiroyu@jp.fujitsu.com: fix softlimit css refcnt handling]
[nishimura@mxp.nes.nec.co.jp: refcount of the "victim" should be decremented before exiting the loop]
Signed-off-by: Balbir Singh
Cc: KAMEZAWA Hiroyuki
Cc: Li Zefan
Acked-by: KOSAKI Motohiro
Signed-off-by: KAMEZAWA Hiroyuki
Signed-off-by: Daisuke Nishimura
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Refactor mem_cgroup_hierarchical_reclaim()
Refactor the arguments passed to mem_cgroup_hierarchical_reclaim() into
flags, so that new parameters don't have to be passed as we make the
reclaim routine more flexibleSigned-off-by: Balbir Singh
Cc: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: KOSAKI Motohiro
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds -
Organize cgroups over soft limit in a RB-Tree
Introduce an RB-Tree for storing memory cgroups that are over their soft
limit. The overall goal is to1. Add a memory cgroup to the RB-Tree when the soft limit is exceeded.
We are careful about updates, updates take place only after a particular
time interval has passed
2. We remove the node from the RB-Tree when the usage goes below the soft
limitThe next set of patches will exploit the RB-Tree to get the group that is
over its soft limit by the largest amount and reclaim from it, when we
face memory contention.[hugh.dickins@tiscali.co.uk: CONFIG_CGROUP_MEM_RES_CTLR=y CONFIG_PREEMPT=y fails to boot]
Signed-off-by: Balbir Singh
Signed-off-by: KAMEZAWA Hiroyuki
Cc: Li Zefan
Cc: KOSAKI Motohiro
Signed-off-by: Hugh Dickins
Cc: Jiri Slaby
Signed-off-by: Andrew Morton
Signed-off-by: Linus Torvalds
