Currently the rate of scanning for an address space is controlled
by the individual tasks. The next scan is simply determined by
2*p->numa_scan_period.
The 2*p->numa_scan_period is arbitrary and never changes. At this point
there is still no proper policy that decides if a task or process is
properly placed. It just scans and assumes the next NUMA fault will
place it properly. As it is assumed that pages will get properly placed
over time, increase the scan window each time a fault is incurred. This
is a big assumption as noted in the comments.
It should be noted that changing to p->numa_scan_period will increase
system CPU usage because now the scanning rate has effectively doubled.
If that is a problem then the min_rate should be made 200ms instead of
restoring the 2* logic.
Signed-off-by: Mel Gorman <mgorman@suse.de>
If there are a large number of NUMA hinting faults and all of them
are resulting in migrations it may indicate that memory is just
bouncing uselessly around. NUMA balancing cost is likely exceeding
any benefit from locality. Rate limit the PTE updates if the node
is migration rate-limited. As noted in the comments, this distorts
the NUMA faulting statistics.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Add a 1 second delay before starting to scan the working set of
a task and starting to balance it amongst nodes.
[ note that before the constant per task WSS sampling rate patch
the initial scan would happen much later still, in effect that
patch caused this regression. ]
The theory is that short-run tasks benefit very little from NUMA
placement: they come and go, and they better stick to the node
they were started on. As tasks mature and rebalance to other CPUs
and nodes, so does their NUMA placement have to change and so
does it start to matter more and more.
In practice this change fixes an observable kbuild regression:
# [ a perf stat --null --repeat 10 test of ten bzImage builds to /dev/shm ]
!NUMA:
45.291088843 seconds time elapsed ( +- 0.40% )
45.154231752 seconds time elapsed ( +- 0.36% )
+NUMA, no slow start:
46.172308123 seconds time elapsed ( +- 0.30% )
46.343168745 seconds time elapsed ( +- 0.25% )
+NUMA, 1 sec slow start:
45.224189155 seconds time elapsed ( +- 0.25% )
45.160866532 seconds time elapsed ( +- 0.17% )
and it also fixes an observable perf bench (hackbench) regression:
# perf stat --null --repeat 10 perf bench sched messaging
-NUMA:
-NUMA: 0.246225691 seconds time elapsed ( +- 1.31% )
+NUMA no slow start: 0.252620063 seconds time elapsed ( +- 1.13% )
+NUMA 1sec delay: 0.248076230 seconds time elapsed ( +- 1.35% )
The implementation is simple and straightforward, most of the patch
deals with adding the /proc/sys/kernel/numa_balancing_scan_delay_ms tunable
knob.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote the changelog, ran measurements, tuned the default. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
Previously, to probe the working set of a task, we'd use
a very simple and crude method: mark all of its address
space PROT_NONE.
That method has various (obvious) disadvantages:
- it samples the working set at dissimilar rates,
giving some tasks a sampling quality advantage
over others.
- creates performance problems for tasks with very
large working sets
- over-samples processes with large address spaces but
which only very rarely execute
Improve that method by keeping a rotating offset into the
address space that marks the current position of the scan,
and advance it by a constant rate (in a CPU cycles execution
proportional manner). If the offset reaches the last mapped
address of the mm then it then it starts over at the first
address.
The per-task nature of the working set sampling functionality in this tree
allows such constant rate, per task, execution-weight proportional sampling
of the working set, with an adaptive sampling interval/frequency that
goes from once per 100ms up to just once per 8 seconds. The current
sampling volume is 256 MB per interval.
As tasks mature and converge their working set, so does the
sampling rate slow down to just a trickle, 256 MB per 8
seconds of CPU time executed.
This, beyond being adaptive, also rate-limits rarely
executing systems and does not over-sample on overloaded
systems.
[ In AutoNUMA speak, this patch deals with the effective sampling
rate of the 'hinting page fault'. AutoNUMA's scanning is
currently rate-limited, but it is also fundamentally
single-threaded, executing in the knuma_scand kernel thread,
so the limit in AutoNUMA is global and does not scale up with
the number of CPUs, nor does it scan tasks in an execution
proportional manner.
So the idea of rate-limiting the scanning was first implemented
in the AutoNUMA tree via a global rate limit. This patch goes
beyond that by implementing an execution rate proportional
working set sampling rate that is not implemented via a single
global scanning daemon. ]
[ Dan Carpenter pointed out a possible NULL pointer dereference in the
first version of this patch. ]
Based-on-idea-by: Andrea Arcangeli <aarcange@redhat.com>
Bug-Found-By: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
[ Wrote changelog and fixed bug. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Mel Gorman <mgorman@suse.de>
Reviewed-by: Rik van Riel <riel@redhat.com>
NOTE: This patch is based on "sched, numa, mm: Add fault driven
placement and migration policy" but as it throws away all the policy
to just leave a basic foundation I had to drop the signed-offs-by.
This patch creates a bare-bones method for setting PTEs pte_numa in the
context of the scheduler that when faulted later will be faulted onto the
node the CPU is running on. In itself this does nothing useful but any
placement policy will fundamentally depend on receiving hints on placement
from fault context and doing something intelligent about it.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Pull futex fix from Thomas Gleixner:
"Single fix for a long standing futex race when taking over a futex
whose owner died. You can end up with two owners, which violates
quite some rules."
* 'core-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
futex: Handle futex_pi OWNER_DIED take over correctly
Siddhesh analyzed a failure in the take over of pi futexes in case the
owner died and provided a workaround.
See: http://sourceware.org/bugzilla/show_bug.cgi?id=14076
The detailed problem analysis shows:
Futex F is initialized with PTHREAD_PRIO_INHERIT and
PTHREAD_MUTEX_ROBUST_NP attributes.
T1 lock_futex_pi(F);
T2 lock_futex_pi(F);
--> T2 blocks on the futex and creates pi_state which is associated
to T1.
T1 exits
--> exit_robust_list() runs
--> Futex F userspace value TID field is set to 0 and
FUTEX_OWNER_DIED bit is set.
T3 lock_futex_pi(F);
--> Succeeds due to the check for F's userspace TID field == 0
--> Claims ownership of the futex and sets its own TID into the
userspace TID field of futex F
--> returns to user space
T1 --> exit_pi_state_list()
--> Transfers pi_state to waiter T2 and wakes T2 via
rt_mutex_unlock(&pi_state->mutex)
T2 --> acquires pi_state->mutex and gains real ownership of the
pi_state
--> Claims ownership of the futex and sets its own TID into the
userspace TID field of futex F
--> returns to user space
T3 --> observes inconsistent state
This problem is independent of UP/SMP, preemptible/non preemptible
kernels, or process shared vs. private. The only difference is that
certain configurations are more likely to expose it.
So as Siddhesh correctly analyzed the following check in
futex_lock_pi_atomic() is the culprit:
if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
We check the userspace value for a TID value of 0 and take over the
futex unconditionally if that's true.
AFAICT this check is there as it is correct for a different corner
case of futexes: the WAITERS bit became stale.
Now the proposed change
- if (unlikely(ownerdied || !(curval & FUTEX_TID_MASK))) {
+ if (unlikely(ownerdied ||
+ !(curval & (FUTEX_TID_MASK | FUTEX_WAITERS)))) {
solves the problem, but it's not obvious why and it wreckages the
"stale WAITERS bit" case.
What happens is, that due to the WAITERS bit being set (T2 is blocked
on that futex) it enforces T3 to go through lookup_pi_state(), which
in the above case returns an existing pi_state and therefor forces T3
to legitimately fight with T2 over the ownership of the pi_state (via
pi_state->mutex). Probelm solved!
Though that does not work for the "WAITERS bit is stale" problem
because if lookup_pi_state() does not find existing pi_state it
returns -ERSCH (due to TID == 0) which causes futex_lock_pi() to
return -ESRCH to user space because the OWNER_DIED bit is not set.
Now there is a different solution to that problem. Do not look at the
user space value at all and enforce a lookup of possibly available
pi_state. If pi_state can be found, then the new incoming locker T3
blocks on that pi_state and legitimately races with T2 to acquire the
rt_mutex and the pi_state and therefor the proper ownership of the
user space futex.
lookup_pi_state() has the correct order of checks. It first tries to
find a pi_state associated with the user space futex and only if that
fails it checks for futex TID value = 0. If no pi_state is available
nothing can create new state at that point because this happens with
the hash bucket lock held.
So the above scenario changes to:
T1 lock_futex_pi(F);
T2 lock_futex_pi(F);
--> T2 blocks on the futex and creates pi_state which is associated
to T1.
T1 exits
--> exit_robust_list() runs
--> Futex F userspace value TID field is set to 0 and
FUTEX_OWNER_DIED bit is set.
T3 lock_futex_pi(F);
--> Finds pi_state and blocks on pi_state->rt_mutex
T1 --> exit_pi_state_list()
--> Transfers pi_state to waiter T2 and wakes it via
rt_mutex_unlock(&pi_state->mutex)
T2 --> acquires pi_state->mutex and gains ownership of the pi_state
--> Claims ownership of the futex and sets its own TID into the
userspace TID field of futex F
--> returns to user space
This covers all gazillion points on which T3 might come in between
T1's exit_robust_list() clearing the TID field and T2 fixing it up. It
also solves the "WAITERS bit stale" problem by forcing the take over.
Another benefit of changing the code this way is that it makes it less
dependent on untrusted user space values and therefor minimizes the
possible wreckage which might be inflicted.
As usual after staring for too long at the futex code my brain hurts
so much that I really want to ditch that whole optimization of
avoiding the syscall for the non contended case for PI futexes and rip
out the maze of corner case handling code. Unfortunately we can't as
user space relies on that existing behaviour, but at least thinking
about it helps me to preserve my mental sanity. Maybe we should
nevertheless :)
Reported-and-tested-by: Siddhesh Poyarekar <siddhesh.poyarekar@gmail.com>
Link: http://lkml.kernel.org/r/alpine.LFD.2.02.1210232138540.2756@ionos
Acked-by: Darren Hart <dvhart@linux.intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Masaki found and patched a kallsyms issue: the last symbol in a
module's symtab wasn't transferred. This is because we manually copy
the zero'th entry (which is always empty) then copy the rest in a loop
starting at 1, though from src[0]. His fix was minimal, I prefer to
rewrite the loops in more standard form.
There are two loops: one to get the size, and one to copy. Make these
identical: always count entry 0 and any defined symbol in an allocated
non-init section.
This bug exists since the following commit was introduced.
module: reduce symbol table for loaded modules (v2)
commit: 4a4962263f
LKML: http://lkml.org/lkml/2012/10/24/27
Reported-by: Masaki Kimura <masaki.kimura.kz@hitachi.com>
Cc: stable@kernel.org
Merge misc fixes from Andrew Morton:
"18 total. 15 fixes and some updates to a device_cgroup patchset which
bring it up to date with the version which I should have merged in the
first place."
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (18 patches)
fs/compat_ioctl.c: VIDEO_SET_SPU_PALETTE missing error check
gen_init_cpio: avoid stack overflow when expanding
drivers/rtc/rtc-imxdi.c: add missing spin lock initialization
mm, numa: avoid setting zone_reclaim_mode unless a node is sufficiently distant
pidns: limit the nesting depth of pid namespaces
drivers/dma/dw_dmac: make driver's endianness configurable
mm/mmu_notifier: allocate mmu_notifier in advance
tools/testing/selftests/epoll/test_epoll.c: fix build
UAPI: fix tools/vm/page-types.c
mm/page_alloc.c:alloc_contig_range(): return early for err path
rbtree: include linux/compiler.h for definition of __always_inline
genalloc: stop crashing the system when destroying a pool
backlight: ili9320: add missing SPI dependency
device_cgroup: add proper checking when changing default behavior
device_cgroup: stop using simple_strtoul()
device_cgroup: rename deny_all to behavior
cgroup: fix invalid rcu dereference
mm: fix XFS oops due to dirty pages without buffers on s390
If one includes documentation for an external tool, it should be
correct. This is not:
1. Overriding the input to rngd should typically be neither
necessary nor desired. This is especially so since newer
versions of rngd support a number of different *types* of sources.
2. The default kernel-exported device is called /dev/hwrng not
/dev/hwrandom nor /dev/hw_random (both of which were used in the
past; however, kernel and udev seem to have converged on
/dev/hwrng.)
Overall it is better if the documentation for rngd is kept with rngd
rather than in a kernel Makefile.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Jeff Garzik <jgarzik@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
'struct pid' is a "variable sized struct" - a header with an array of
upids at the end.
The size of the array depends on a level (depth) of pid namespaces. Now a
level of pidns is not limited, so 'struct pid' can be more than one page.
Looks reasonable, that it should be less than a page. MAX_PIS_NS_LEVEL is
not calculated from PAGE_SIZE, because in this case it depends on
architectures, config options and it will be reduced, if someone adds a
new fields in struct pid or struct upid.
I suggest to set MAX_PIS_NS_LEVEL = 32, because it saves ability to expand
"struct pid" and it's more than enough for all known for me use-cases.
When someone finds a reasonable use case, we can add a config option or a
sysctl parameter.
In addition it will reduce the effect of another problem, when we have
many nested namespaces and the oldest one starts dying.
zap_pid_ns_processe will be called for each namespace and find_vpid will
be called for each process in a namespace. find_vpid will be called
minimum max_level^2 / 2 times. The reason of that is that when we found a
bit in pidmap, we can't determine this pidns is top for this process or it
isn't.
vpid is a heavy operation, so a fork bomb, which create many nested
namespace, can make a system inaccessible for a long time. For example my
system becomes inaccessible for a few minutes with 4000 processes.
[akpm@linux-foundation.org: return -EINVAL in response to excessive nesting, not -ENOMEM]
Signed-off-by: Andrew Vagin <avagin@openvz.org>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull cgroup fixes from Tejun Heo:
"This pull request contains three fixes.
Two are reverts of task_lock() removal in cgroup fork path. The
optimizations incorrectly assumed that threadgroup_lock can protect
process forks (as opposed to thread creations) too. Further cleanup
of cgroup fork path is scheduled.
The third fixes cgroup emptiness notification loss."
* 'for-3.7-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup:
Revert "cgroup: Remove task_lock() from cgroup_post_fork()"
Revert "cgroup: Drop task_lock(parent) on cgroup_fork()"
cgroup: notify_on_release may not be triggered in some cases
Pull workqueue fix from Tejun Heo:
"This pull request contains one patch from Dan Magenheimer to fix
cancel_delayed_work() regression introduced by its reimplementation
using try_to_grab_pending(). The reimplementation made it incorrectly
return %true when the work item is idle.
There aren't too many consumers of the return value but it broke at
least ramster."
* 'for-3.7-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: cancel_delayed_work() should return %false if work item is idle
57b30ae77b ("workqueue: reimplement cancel_delayed_work() using
try_to_grab_pending()") made cancel_delayed_work() always return %true
unless someone else is also trying to cancel the work item, which is
broken - if the target work item is idle, the return value should be
%false.
try_to_grab_pending() indicates that the target work item was idle by
zero return value. Use it for return. Note that this brings
cancel_delayed_work() in line with __cancel_work_timer() in return
value handling.
Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <444a6439-b1a4-4740-9e7e-bc37267cfe73@default>
Pull perf fixes from Ingo Molnar:
"Most of these are uprobes race fixes from Oleg, and their preparatory
cleanups. (It's larger than what I'd normally send for an -rc kernel,
but they looked significant enough to not delay them.)
There's also an oprofile fix and an uncore PMU fix."
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (22 commits)
perf/x86: Disable uncore on virtualized CPUs
oprofile, x86: Fix wrapping bug in op_x86_get_ctrl()
ring-buffer: Check for uninitialized cpu buffer before resizing
uprobes: Fix the racy uprobe->flags manipulation
uprobes: Fix prepare_uprobe() race with itself
uprobes: Introduce prepare_uprobe()
uprobes: Fix handle_swbp() vs unregister() + register() race
uprobes: Do not delete uprobe if uprobe_unregister() fails
uprobes: Don't return success if alloc_uprobe() fails
uprobes/x86: Only rep+nop can be emulated correctly
uprobes: Simplify is_swbp_at_addr(), remove stale comments
uprobes: Kill set_orig_insn()->is_swbp_at_addr()
uprobes: Introduce copy_opcode(), kill read_opcode()
uprobes: Kill set_swbp()->is_swbp_at_addr()
uprobes: Restrict valid_vma(false) to skip VM_SHARED vmas
uprobes: Change valid_vma() to demand VM_MAYEXEC rather than VM_EXEC
uprobes: Change write_opcode() to use FOLL_FORCE
uprobes: Move clear_thread_flag(TIF_UPROBE) to uprobe_notify_resume()
uprobes: Kill UTASK_BP_HIT state
uprobes: Fix UPROBE_SKIP_SSTEP checks in handle_swbp()
...
Fix the warning:
kernel/module_signing.c:195:2: warning: format '%lu' expects type 'long unsigned int', but argument 3 has type 'size_t'
by using the proper 'z' modifier for printing a size_t.
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The min/max call needed to have explicit types on some architectures
(e.g. mn10300). Use clamp_t instead to avoid the warning:
kernel/sys.c: In function 'override_release':
kernel/sys.c:1287:10: warning: comparison of distinct pointer types lacks a cast [enabled by default]
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Emit the magic string that indicates a module has a signature after the
signature data instead of before it. This allows module_sig_check() to
be made simpler and faster by the elimination of the search for the
magic string. Instead we just need to do a single memcmp().
This works because at the end of the signature data there is the
fixed-length signature information block. This block then falls
immediately prior to the magic number.
From the contents of the information block, it is trivial to calculate
the size of the signature data and thus the size of the actual module
data.
Signed-off-by: David Howells <dhowells@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit 7e3aa30ac8.
The commit incorrectly assumed that fork path always performed
threadgroup_change_begin/end() and depended on that for
synchronization against task exit and cgroup migration paths instead
of explicitly grabbing task_lock().
threadgroup_change is not locked when forking a new process (as
opposed to a new thread in the same process) and even if it were it
wouldn't be effective as different processes use different threadgroup
locks.
Revert the incorrect optimization.
Signed-off-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <20121008020000.GB2575@localhost>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: stable@vger.kernel.org
This reverts commit 7e381b0eb1.
The commit incorrectly assumed that fork path always performed
threadgroup_change_begin/end() and depended on that for
synchronization against task exit and cgroup migration paths instead
of explicitly grabbing task_lock().
threadgroup_change is not locked when forking a new process (as
opposed to a new thread in the same process) and even if it were it
wouldn't be effective as different processes use different threadgroup
locks.
Revert the incorrect optimization.
Signed-off-by: Tejun Heo <tj@kernel.org>
LKML-Reference: <20121008020000.GB2575@localhost>
Acked-by: Li Zefan <lizefan@huawei.com>
Bitterly-Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: stable@vger.kernel.org
free_pid_ns() operates in a recursive fashion:
free_pid_ns(parent)
put_pid_ns(parent)
kref_put(&ns->kref, free_pid_ns);
free_pid_ns
thus if there was a huge nesting of namespaces the userspace may trigger
avalanche calling of free_pid_ns leading to kernel stack exhausting and a
panic eventually.
This patch turns the recursion into an iterative loop.
Based on a patch by Andrew Vagin.
[akpm@linux-foundation.org: export put_pid_ns() to modules]
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Andrew Vagin <avagin@openvz.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
