When running perf on the ftrace:function tracepoint, there is a bug
which can be reproduced by:
perf record -e ftrace:function -a sleep 20 &
perf record -e ftrace:function ls
perf script
ls 10304 [005] 171.853235: ftrace:function:
perf_output_begin
ls 10304 [005] 171.853237: ftrace:function:
perf_output_begin
ls 10304 [005] 171.853239: ftrace:function:
task_tgid_nr_ns
ls 10304 [005] 171.853240: ftrace:function:
task_tgid_nr_ns
ls 10304 [005] 171.853242: ftrace:function:
__task_pid_nr_ns
ls 10304 [005] 171.853244: ftrace:function:
__task_pid_nr_ns
We can see that all the function traces are doubled.
The problem is caused by the inconsistency of the register
function perf_ftrace_event_register() with the probe function
perf_ftrace_function_call(). The former registers one probe
for every perf_event. And the latter handles all perf_events
on the current cpu. So when two perf_events on the current cpu,
the traces of them will be doubled.
So this patch adds an extra parameter "event" for perf_tp_event,
only send sample data to this event when it's not NULL.
Signed-off-by: Zhou Chengming <zhouchengming1@huawei.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@kernel.org
Cc: alexander.shishkin@linux.intel.com
Cc: huawei.libin@huawei.com
Link: http://lkml.kernel.org/r/1503668977-12526-1-git-send-email-zhouchengming1@huawei.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
While examining the kernel source code, I found a dangerous operation that
could turn into a double-fetch situation (a race condition bug) where the same
userspace memory region are fetched twice into kernel with sanity checks after
the first fetch while missing checks after the second fetch.
1. The first fetch happens in line 9573 get_user(size, &uattr->size).
2. Subsequently the 'size' variable undergoes a few sanity checks and
transformations (line 9577 to 9584).
3. The second fetch happens in line 9610 copy_from_user(attr, uattr, size)
4. Given that 'uattr' can be fully controlled in userspace, an attacker can
race condition to override 'uattr->size' to arbitrary value (say, 0xFFFFFFFF)
after the first fetch but before the second fetch. The changed value will be
copied to 'attr->size'.
5. There is no further checks on 'attr->size' until the end of this function,
and once the function returns, we lose the context to verify that 'attr->size'
conforms to the sanity checks performed in step 2 (line 9577 to 9584).
6. My manual analysis shows that 'attr->size' is not used elsewhere later,
so, there is no working exploit against it right now. However, this could
easily turns to an exploitable one if careless developers start to use
'attr->size' later.
To fix this, override 'attr->size' from the second fetch to the one from the
first fetch, regardless of what is actually copied in.
In this way, it is assured that 'attr->size' is consistent with the checks
performed after the first fetch.
Signed-off-by: Meng Xu <mengxu.gatech@gmail.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@kernel.org
Cc: alexander.shishkin@linux.intel.com
Cc: meng.xu@gatech.edu
Cc: sanidhya@gatech.edu
Cc: taesoo@gatech.edu
Link: http://lkml.kernel.org/r/1503522470-35531-1-git-send-email-meng.xu@gatech.edu
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Regardless of which events form a group, it does not make sense for the
events to target different tasks and/or CPUs, as this leaves the group
inconsistent and impossible to schedule. The core perf code assumes that
these are consistent across (successfully intialised) groups.
Core perf code only verifies this when moving SW events into a HW
context. Thus, we can violate this requirement for pure SW groups and
pure HW groups, unless the relevant PMU driver happens to perform this
verification itself. These mismatched groups subsequently wreak havoc
elsewhere.
For example, we handle watchpoints as SW events, and reserve watchpoint
HW on a per-CPU basis at pmu::event_init() time to ensure that any event
that is initialised is guaranteed to have a slot at pmu::add() time.
However, the core code only checks the group leader's cpu filter (via
event_filter_match()), and can thus install follower events onto CPUs
violating thier (mismatched) CPU filters, potentially installing them
into a CPU without sufficient reserved slots.
This can be triggered with the below test case, resulting in warnings
from arch backends.
#define _GNU_SOURCE
#include <linux/hw_breakpoint.h>
#include <linux/perf_event.h>
#include <sched.h>
#include <stdio.h>
#include <sys/prctl.h>
#include <sys/syscall.h>
#include <unistd.h>
static int perf_event_open(struct perf_event_attr *attr, pid_t pid, int cpu,
int group_fd, unsigned long flags)
{
return syscall(__NR_perf_event_open, attr, pid, cpu, group_fd, flags);
}
char watched_char;
struct perf_event_attr wp_attr = {
.type = PERF_TYPE_BREAKPOINT,
.bp_type = HW_BREAKPOINT_RW,
.bp_addr = (unsigned long)&watched_char,
.bp_len = 1,
.size = sizeof(wp_attr),
};
int main(int argc, char *argv[])
{
int leader, ret;
cpu_set_t cpus;
/*
* Force use of CPU0 to ensure our CPU0-bound events get scheduled.
*/
CPU_ZERO(&cpus);
CPU_SET(0, &cpus);
ret = sched_setaffinity(0, sizeof(cpus), &cpus);
if (ret) {
printf("Unable to set cpu affinity\n");
return 1;
}
/* open leader event, bound to this task, CPU0 only */
leader = perf_event_open(&wp_attr, 0, 0, -1, 0);
if (leader < 0) {
printf("Couldn't open leader: %d\n", leader);
return 1;
}
/*
* Open a follower event that is bound to the same task, but a
* different CPU. This means that the group should never be possible to
* schedule.
*/
ret = perf_event_open(&wp_attr, 0, 1, leader, 0);
if (ret < 0) {
printf("Couldn't open mismatched follower: %d\n", ret);
return 1;
} else {
printf("Opened leader/follower with mismastched CPUs\n");
}
/*
* Open as many independent events as we can, all bound to the same
* task, CPU0 only.
*/
do {
ret = perf_event_open(&wp_attr, 0, 0, -1, 0);
} while (ret >= 0);
/*
* Force enable/disble all events to trigger the erronoeous
* installation of the follower event.
*/
printf("Opened all events. Toggling..\n");
for (;;) {
prctl(PR_TASK_PERF_EVENTS_DISABLE, 0, 0, 0, 0);
prctl(PR_TASK_PERF_EVENTS_ENABLE, 0, 0, 0, 0);
}
return 0;
}
Fix this by validating this requirement regardless of whether we're
moving events.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Zhou Chengming <zhouchengming1@huawei.com>
Cc: stable@vger.kernel.org
Link: http://lkml.kernel.org/r/1498142498-15758-1-git-send-email-mark.rutland@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Vince reported that when we do IOC_ENABLE/IOC_DISABLE while the task
is SIGSTOP'ed state the timestamps go wobbly.
It turns out we indeed fail to correctly account time while in 'OFF'
state and doing IOC_ENABLE without getting scheduled in exposes the
problem.
Further thinking about this problem, it occurred to me that we can
suffer a similar fate when we migrate an uncore event between CPUs.
The perf_event_install() on the 'new' CPU will do add_event_to_ctx()
which will reset all the time stamp, resulting in a subsequent
update_event_times() to overwrite the total_time_* fields with smaller
values.
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf fixes from Ingo Molnar:
"Two hw-enablement patches, two race fixes, three fixes for regressions
of semantics, plus a number of tooling fixes"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf/x86/intel: Add proper condition to run sched_task callbacks
perf/core: Fix locking for children siblings group read
perf/core: Fix scheduling regression of pinned groups
perf/x86/intel: Fix debug_store reset field for freq events
perf/x86/intel: Add Goldmont Plus CPU PMU support
perf/x86/intel: Enable C-state residency events for Apollo Lake
perf symbols: Accept zero as the kernel base address
Revert "perf/core: Drop kernel samples even though :u is specified"
perf annotate: Fix broken arrow at row 0 connecting jmp instruction to its target
perf evsel: State in the default event name if attr.exclude_kernel is set
perf evsel: Fix attr.exclude_kernel setting for default cycles:p
We're missing ctx lock when iterating children siblings
within the perf_read path for group reading. Following
race and crash can happen:
User space doing read syscall on event group leader:
T1:
perf_read
lock event->ctx->mutex
perf_read_group
lock leader->child_mutex
__perf_read_group_add(child)
list_for_each_entry(sub, &leader->sibling_list, group_entry)
----> sub might be invalid at this point, because it could
get removed via perf_event_exit_task_context in T2
Child exiting and cleaning up its events:
T2:
perf_event_exit_task_context
lock ctx->mutex
list_for_each_entry_safe(child_event, next, &child_ctx->event_list,...
perf_event_exit_event(child)
lock ctx->lock
perf_group_detach(child)
unlock ctx->lock
----> child is removed from sibling_list without any sync
with T1 path above
...
free_event(child)
Before the child is removed from the leader's child_list,
(and thus is omitted from perf_read_group processing), we
need to ensure that perf_read_group touches child's
siblings under its ctx->lock.
Peter further notes:
| One additional note; this bug got exposed by commit:
|
| ba5213ae6b ("perf/core: Correct event creation with PERF_FORMAT_GROUP")
|
| which made it possible to actually trigger this code-path.
Tested-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: ba5213ae6b ("perf/core: Correct event creation with PERF_FORMAT_GROUP")
Link: http://lkml.kernel.org/r/20170720141455.2106-1-jolsa@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Vince Weaver reported:
> I was tracking down some regressions in my perf_event_test testsuite.
> Some of the tests broke in the 4.11-rc1 timeframe.
>
> I've bisected one of them, this report is about
> tests/overflow/simul_oneshot_group_overflow
> This test creates an event group containing two sampling events, set
> to overflow to a signal handler (which disables and then refreshes the
> event).
>
> On a good kernel you get the following:
> Event perf::instructions with period 1000000
> Event perf::instructions with period 2000000
> fd 3 overflows: 946 (perf::instructions/1000000)
> fd 4 overflows: 473 (perf::instructions/2000000)
> Ending counts:
> Count 0: 946379875
> Count 1: 946365218
>
> With the broken kernels you get:
> Event perf::instructions with period 1000000
> Event perf::instructions with period 2000000
> fd 3 overflows: 938 (perf::instructions/1000000)
> fd 4 overflows: 318 (perf::instructions/2000000)
> Ending counts:
> Count 0: 946373080
> Count 1: 653373058
The root cause of the bug is that the following commit:
487f05e18a ("perf/core: Optimize event rescheduling on active contexts")
erronously assumed that event's 'pinned' setting determines whether the
event belongs to a pinned group or not, but in fact, it's the group
leader's pinned state that matters.
This was discovered by Vince in the test case described above, where two instruction
counters are grouped, the group leader is pinned, but the other event is not;
in the regressed case the counters were off by 33% (the difference between events'
periods), but should be the same within the error margin.
Fix the problem by looking at the group leader's pinning.
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Tested-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Fixes: 487f05e18a ("perf/core: Optimize event rescheduling on active contexts")
Link: http://lkml.kernel.org/r/87lgnmvw7h.fsf@ashishki-desk.ger.corp.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull networking updates from David Miller:
"Reasonably busy this cycle, but perhaps not as busy as in the 4.12
merge window:
1) Several optimizations for UDP processing under high load from
Paolo Abeni.
2) Support pacing internally in TCP when using the sch_fq packet
scheduler for this is not practical. From Eric Dumazet.
3) Support mutliple filter chains per qdisc, from Jiri Pirko.
4) Move to 1ms TCP timestamp clock, from Eric Dumazet.
5) Add batch dequeueing to vhost_net, from Jason Wang.
6) Flesh out more completely SCTP checksum offload support, from
Davide Caratti.
7) More plumbing of extended netlink ACKs, from David Ahern, Pablo
Neira Ayuso, and Matthias Schiffer.
8) Add devlink support to nfp driver, from Simon Horman.
9) Add RTM_F_FIB_MATCH flag to RTM_GETROUTE queries, from Roopa
Prabhu.
10) Add stack depth tracking to BPF verifier and use this information
in the various eBPF JITs. From Alexei Starovoitov.
11) Support XDP on qed device VFs, from Yuval Mintz.
12) Introduce BPF PROG ID for better introspection of installed BPF
programs. From Martin KaFai Lau.
13) Add bpf_set_hash helper for TC bpf programs, from Daniel Borkmann.
14) For loads, allow narrower accesses in bpf verifier checking, from
Yonghong Song.
15) Support MIPS in the BPF selftests and samples infrastructure, the
MIPS eBPF JIT will be merged in via the MIPS GIT tree. From David
Daney.
16) Support kernel based TLS, from Dave Watson and others.
17) Remove completely DST garbage collection, from Wei Wang.
18) Allow installing TCP MD5 rules using prefixes, from Ivan
Delalande.
19) Add XDP support to Intel i40e driver, from Björn Töpel
20) Add support for TC flower offload in nfp driver, from Simon
Horman, Pieter Jansen van Vuuren, Benjamin LaHaise, Jakub
Kicinski, and Bert van Leeuwen.
21) IPSEC offloading support in mlx5, from Ilan Tayari.
22) Add HW PTP support to macb driver, from Rafal Ozieblo.
23) Networking refcount_t conversions, From Elena Reshetova.
24) Add sock_ops support to BPF, from Lawrence Brako. This is useful
for tuning the TCP sockopt settings of a group of applications,
currently via CGROUPs"
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1899 commits)
net: phy: dp83867: add workaround for incorrect RX_CTRL pin strap
dt-bindings: phy: dp83867: provide a workaround for incorrect RX_CTRL pin strap
cxgb4: Support for get_ts_info ethtool method
cxgb4: Add PTP Hardware Clock (PHC) support
cxgb4: time stamping interface for PTP
nfp: default to chained metadata prepend format
nfp: remove legacy MAC address lookup
nfp: improve order of interfaces in breakout mode
net: macb: remove extraneous return when MACB_EXT_DESC is defined
bpf: add missing break in for the TCP_BPF_SNDCWND_CLAMP case
bpf: fix return in load_bpf_file
mpls: fix rtm policy in mpls_getroute
net, ax25: convert ax25_cb.refcount from atomic_t to refcount_t
net, ax25: convert ax25_route.refcount from atomic_t to refcount_t
net, ax25: convert ax25_uid_assoc.refcount from atomic_t to refcount_t
net, sctp: convert sctp_ep_common.refcnt from atomic_t to refcount_t
net, sctp: convert sctp_transport.refcnt from atomic_t to refcount_t
net, sctp: convert sctp_chunk.refcnt from atomic_t to refcount_t
net, sctp: convert sctp_datamsg.refcnt from atomic_t to refcount_t
net, sctp: convert sctp_auth_bytes.refcnt from atomic_t to refcount_t
...
Pull SMP hotplug updates from Thomas Gleixner:
"This update is primarily a cleanup of the CPU hotplug locking code.
The hotplug locking mechanism is an open coded RWSEM, which allows
recursive locking. The main problem with that is the recursive nature
as it evades the full lockdep coverage and hides potential deadlocks.
The rework replaces the open coded RWSEM with a percpu RWSEM and
establishes full lockdep coverage that way.
The bulk of the changes fix up recursive locking issues and address
the now fully reported potential deadlocks all over the place. Some of
these deadlocks have been observed in the RT tree, but on mainline the
probability was low enough to hide them away."
* 'smp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (37 commits)
cpu/hotplug: Constify attribute_group structures
powerpc: Only obtain cpu_hotplug_lock if called by rtasd
ARM/hw_breakpoint: Fix possible recursive locking for arch_hw_breakpoint_init
cpu/hotplug: Remove unused check_for_tasks() function
perf/core: Don't release cred_guard_mutex if not taken
cpuhotplug: Link lock stacks for hotplug callbacks
acpi/processor: Prevent cpu hotplug deadlock
sched: Provide is_percpu_thread() helper
cpu/hotplug: Convert hotplug locking to percpu rwsem
s390: Prevent hotplug rwsem recursion
arm: Prevent hotplug rwsem recursion
arm64: Prevent cpu hotplug rwsem recursion
kprobes: Cure hotplug lock ordering issues
jump_label: Reorder hotplug lock and jump_label_lock
perf/tracing/cpuhotplug: Fix locking order
ACPI/processor: Use cpu_hotplug_disable() instead of get_online_cpus()
PCI: Replace the racy recursion prevention
PCI: Use cpu_hotplug_disable() instead of get_online_cpus()
perf/x86/intel: Drop get_online_cpus() in intel_snb_check_microcode()
x86/perf: Drop EXPORT of perf_check_microcode
...
Allow BPF_PROG_TYPE_PERF_EVENT program types to attach to all
perf_event types, including HW_CACHE, RAW, and dynamic pmu events.
Only tracepoint/kprobe events are treated differently which require
BPF_PROG_TYPE_TRACEPOINT/BPF_PROG_TYPE_KPROBE program types accordingly.
Also add support for reading all event counters using
bpf_perf_event_read() helper.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
perf, tracing, kprobes and jump_labels have a gazillion of ways to create
dependency lock chains. Some of those involve nested invocations of
get_online_cpus().
The conversion of the hotplug locking to a percpu rwsem requires to avoid
such nested calls. sys_perf_event_open() protects most of the syscall logic
against cpu hotplug. This causes nested calls and lock inversions versus
ftrace and kprobes in various interesting ways.
It's impossible to move the hotplug locking to the outer end of all call
chains in the involved facilities, so the hotplug protection in
sys_perf_event_open() needs to be solved differently.
Introduce 'pmus_mutex' which protects a perf private online cpumask. This
mutex is taken when the mask is updated in the cpu hotplug callbacks and
can be taken in sys_perf_event_open() to protect the swhash setup/teardown
code and when the final judgement about a valid event has to be made.
[ tglx: Produced changelog and fixed the swhash interaction ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@kernel.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Sebastian Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Link: http://lkml.kernel.org/r/20170524081548.930941109@linutronix.de
While going through the event inheritance code Oleg got confused.
Add some comments to better explain the silent dissapearance of
orphaned events.
So what happens is that at perf_event_release_kernel() time; when an
event looses its connection to userspace (and ceases to exist from the
user's perspective) we can still have an arbitrary amount of inherited
copies of the event. We want to synchronously find and remove all
these child events.
Since that requires a bit of lock juggling, there is the possibility
that concurrent clone()s will create new child events. Therefore we
first mark the parent event as DEAD, which marks all the extant child
events as orphaned.
We then avoid copying orphaned events; in order to avoid getting more
of them.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Dmitry Vyukov <dvyukov@google.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: fweisbec@gmail.com
Link: http://lkml.kernel.org/r/20170316125823.289567442@infradead.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We are going to split <linux/sched/mm.h> out of <linux/sched.h>, which
will have to be picked up from other headers and a couple of .c files.
Create a trivial placeholder <linux/sched/mm.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
The APIs that are going to be moved first are:
mm_alloc()
__mmdrop()
mmdrop()
mmdrop_async_fn()
mmdrop_async()
mmget_not_zero()
mmput()
mmput_async()
get_task_mm()
mm_access()
mm_release()
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We are going to split <linux/sched/clock.h> out of <linux/sched.h>, which
will have to be picked up from other headers and .c files.
Create a trivial placeholder <linux/sched/clock.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull cgroup updates from Tejun Heo:
"Several noteworthy changes.
- Parav's rdma controller is finally merged. It is very straight
forward and can limit the abosolute numbers of common rdma
constructs used by different cgroups.
- kernel/cgroup.c got too chubby and disorganized. Created
kernel/cgroup/ subdirectory and moved all cgroup related files
under kernel/ there and reorganized the core code. This hurts for
backporting patches but was long overdue.
- cgroup v2 process listing reimplemented so that it no longer
depends on allocating a buffer large enough to cache the entire
result to sort and uniq the output. v2 has always mangled the sort
order to ensure that users don't depend on the sorted output, so
this shouldn't surprise anybody. This makes the pid listing
functions use the same iterators that are used internally, which
have to have the same iterating capabilities anyway.
- perf cgroup filtering now works automatically on cgroup v2. This
patch was posted a long time ago but somehow fell through the
cracks.
- misc fixes asnd documentation updates"
* 'for-4.11' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup: (27 commits)
kernfs: fix locking around kernfs_ops->release() callback
cgroup: drop the matching uid requirement on migration for cgroup v2
cgroup, perf_event: make perf_event controller work on cgroup2 hierarchy
cgroup: misc cleanups
cgroup: call subsys->*attach() only for subsystems which are actually affected by migration
cgroup: track migration context in cgroup_mgctx
cgroup: cosmetic update to cgroup_taskset_add()
rdmacg: Fixed uninitialized current resource usage
cgroup: Add missing cgroup-v2 PID controller documentation.
rdmacg: Added documentation for rdmacg
IB/core: added support to use rdma cgroup controller
rdmacg: Added rdma cgroup controller
cgroup: fix a comment typo
cgroup: fix RCU related sparse warnings
cgroup: move namespace code to kernel/cgroup/namespace.c
cgroup: rename functions for consistency
cgroup: move v1 mount functions to kernel/cgroup/cgroup-v1.c
cgroup: separate out cgroup1_kf_syscall_ops
cgroup: refactor mount path and clearly distinguish v1 and v2 paths
cgroup: move cgroup v1 specific code to kernel/cgroup/cgroup-v1.c
...
Since commit:
321027c1fe ("perf/core: Fix concurrent sys_perf_event_open() vs. 'move_group' race")
... the code looks like (assuming move_group==1):
gctx = __perf_event_ctx_lock_double(group_leader, ctx);
perf_remove_from_context(group_leader, 0);
list_for_each_entry(sibling, &group_leader->sibling_list, group_entry) {
perf_remove_from_context(sibling, 0);
put_ctx(gctx);
}
/* ... */
/* misleading comment about how this is the last reference */
put_ctx(gctx);
perf_event_ctx_unlock(group_leader, gctx);
What that 'last' put_ctx() does is drop @group_leader's reference on
gctx after having dropped all its potential sibling references.
But the thing is that __perf_event_ctx_lock_double() returns with a
reference _and_ a held lock, and perf_event_ctx_unlock() unlocks that
lock and drops that reference. Therefore that put_ctx() cannot be the
'last' of anything, nor is there an unbalance in puts.
To reduce confusion, remove the comment and place the put_ctx() next
to the remove_from_context() call.
Reported-by: Ben Hutchings <ben@decadent.org.uk>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
perf_event is a utility controller whose primary role is identifying
cgroup membership to filter perf events; however, because it also
tracks some per-css state, it can't be replaced by pure cgroup
membership test. Mark the controller as implicitly enabled on the
default hierarchy so that perf events can always be filtered based on
cgroup v2 path as long as the controller is not mounted on a legacy
hierarchy.
"perf record" is updated accordingly so that it searches for both v1
and v2 hierarchies. A v1 hierarchy is used if perf_event is mounted
on it; otherwise, it uses the v2 hierarchy.
v2: Doc updated to reflect more flexible rebinding behavior.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
perf has additional overhead when monitoring the task which
frequently generates child tasks.
perf_init_event() is one of the hotspots for the additional overhead:
Currently, to get the PMU, it tries to search the type in pmu_idr at
first. But it is not always successful, especially for the widely used
PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE events. So it has to go to the
slow path which go through the whole PMUs list.
It will be a big performance issue, if the PMUs list is long (e.g. server
with many uncore boxes) and the task frequently generates child tasks.
The child event inherits its parent event. So the child event should
try its parent PMU first.
Here is some data from the overhead test on Broadwell server:
perf record -e $TEST_EVENTS -- ./loop.sh 50000
loop.sh
start=$(date +%s%N)
i=0
while [ "$i" -le "$1" ]
do
date > /dev/null
i=`expr $i + 1`
done
end=$(date +%s%N)
elapsed=`expr $end - $start`
Event# Original elapsed time Elapsed time with patch delta
1 196,573,192,397 189,162,029,998 -3.77%
2 257,567,753,013 241,620,788,683 -6.19%
4 398,730,726,971 370,518,938,714 -7.08%
8 824,983,761,120 740,702,489,329 -10.22%
16 1,883,411,923,498 1,672,027,508,355 -11.22%
... which shows a nice performance improvement.
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1484745662-15928-2-git-send-email-kan.liang@intel.com
[ Tidied up the changelog and the code comment. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch follows from a conversation in CQM/CMT's last series about
speeding up the context switch for cgroup events:
https://patchwork.kernel.org/patch/9478617/
This is a low-hanging fruit optimization. It replaces the iteration over
the "pmus" list in cgroup switch by an iteration over a new list that
contains only cpuctxs with at least one cgroup event.
This is necessary because the number of PMUs have increased over the years
e.g modern x86 server systems have well above 50 PMUs.
The iteration over the full PMU list is unneccessary and can be costly in
heavy cache contention scenarios.
Below are some instrumentation measurements with 10, 50 and 90 percentiles
of the total cost of context switch before and after this optimization for
a simple array read/write microbenchark.
Contention
Level Nr events Before (us) After (us) Median
L2 L3 types (10%, 50%, 90%) (10%, 50%, 90% Speedup
--------------------------------------------------------------------------
Low Low 1 (1.72, 2.42, 5.85) (1.35, 1.64, 5.46) 29%
High Low 1 (2.08, 4.56, 19.8) (1720, 2.20, 13.7) 51%
High High 1 (2.86, 10.4, 12.7) (2.54, 4.32, 12.1) 58%
Low Low 2 (1.98, 3.20, 6.89) (1.68, 2.41, 8.89) 24%
High Low 2 (2.48, 5.28, 22.4) (2150, 3.69, 14.6) 30%
High High 2 (3.32, 8.09, 13.9) (2.80, 5.15, 13.7) 36%
where:
1 event type = cycles
2 event types = cycles,intel_cqm/llc_occupancy/
Contention L2 Low: workset < L2 cache size.
High: " >> L2 " " .
Contention L3 Low: workset of task on all sockets < L3 cache size.
High: " " " " " " >> L3 " " .
Median Speedup is (50%ile Before - 50%ile After) / 50%ile Before
Unsurprisingly, the benefits of this optimization decrease with the number
of cpuctxs with a cgroup events, yet, is never detrimental.
Tested-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vikas Shivappa <vikas.shivappa@linux.intel.com>
Cc: Vince Weaver <vince@deater.net>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/20170118192454.58008-2-davidcc@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
It's possible to set up PEBS events to get only errors and not
any data, like on SNB-X (model 45) and IVB-EP (model 62)
via 2 perf commands running simultaneously:
taskset -c 1 ./perf record -c 4 -e branches:pp -j any -C 10
This leads to a soft lock up, because the error path of the
intel_pmu_drain_pebs_nhm() does not account event->hw.interrupt
for error PEBS interrupts, so in case you're getting ONLY
errors you don't have a way to stop the event when it's over
the max_samples_per_tick limit:
NMI watchdog: BUG: soft lockup - CPU#22 stuck for 22s! [perf_fuzzer:5816]
...
RIP: 0010:[<ffffffff81159232>] [<ffffffff81159232>] smp_call_function_single+0xe2/0x140
...
Call Trace:
? trace_hardirqs_on_caller+0xf5/0x1b0
? perf_cgroup_attach+0x70/0x70
perf_install_in_context+0x199/0x1b0
? ctx_resched+0x90/0x90
SYSC_perf_event_open+0x641/0xf90
SyS_perf_event_open+0x9/0x10
do_syscall_64+0x6c/0x1f0
entry_SYSCALL64_slow_path+0x25/0x25
Add perf_event_account_interrupt() which does the interrupt
and frequency checks and call it from intel_pmu_drain_pebs_nhm()'s
error path.
We keep the pending_kill and pending_wakeup logic only in the
__perf_event_overflow() path, because they make sense only if
there's any data to deliver.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vince@deater.net>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1482931866-6018-2-git-send-email-jolsa@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
