Stephane reported that we don't set properly PERIOD sample type for
events with period term defined.
Before:
$ perf record -e cpu/cpu-cycles,period=1000/u ls
$ perf evlist -v
cpu/cpu-cycles,period=1000/u: ... sample_type: IP|TID|TIME|PERIOD, ...
After:
$ perf record -e cpu/cpu-cycles,period=1000/u ls
$ perf evlist -v
cpu/cpu-cycles,period=1000/u: ... sample_type: IP|TID|TIME, ...
Setting PERIOD sample type based on period term setup.
Committer note:
When we use -c or a period=N term in the event definition, then we don't
need to ask the kernel, for this event, via perf_event_attr.sample_type
|= PERF_SAMPLE_PERIOD, to put the event period in each sample for this
event, as we know it already, it is in perf_event_attr.sample_period.
Reported-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Stephane Eranian <eranian@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20180201083812.11359-2-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When setting the "dwarf" unwinder for a specific event and not
specifying the max-stack, the attr.sample_max_stack ended up using an
uninitialized callchain_param.max_stack, fix it by using designated
initializers for that callchain_param variable, zeroing all non
explicitely initialized struct members.
Here is what happened:
# perf trace -vv --no-syscalls --max-stack 4 -e probe_libc:inet_pton/call-graph=dwarf/ ping -6 -c 1 ::1
callchain: type DWARF
callchain: stack dump size 8192
perf_event_attr:
type 2
size 112
config 0x730
{ sample_period, sample_freq } 1
sample_type IP|TID|TIME|ADDR|CALLCHAIN|CPU|PERIOD|RAW|REGS_USER|STACK_USER|DATA_SRC
exclude_callchain_user 1
{ wakeup_events, wakeup_watermark } 1
sample_regs_user 0xff0fff
sample_stack_user 8192
sample_max_stack 50656
sys_perf_event_open failed, error -75
Value too large for defined data type
# perf trace -vv --no-syscalls --max-stack 4 -e probe_libc:inet_pton/call-graph=dwarf/ ping -6 -c 1 ::1
callchain: type DWARF
callchain: stack dump size 8192
perf_event_attr:
type 2
size 112
config 0x730
sample_type IP|TID|TIME|ADDR|CALLCHAIN|CPU|PERIOD|RAW|REGS_USER|STACK_USER|DATA_SRC
exclude_callchain_user 1
sample_regs_user 0xff0fff
sample_stack_user 8192
sample_max_stack 30448
sys_perf_event_open failed, error -75
Value too large for defined data type
#
Now the attr.sample_max_stack is set to zero and the above works as
expected:
# perf trace --no-syscalls --max-stack 4 -e probe_libc:inet_pton/call-graph=dwarf/ ping -6 -c 1 ::1
PING ::1(::1) 56 data bytes
64 bytes from ::1: icmp_seq=1 ttl=64 time=0.072 ms
--- ::1 ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.072/0.072/0.072/0.000 ms
0.000 probe_libc:inet_pton:(7feb7a998350))
__inet_pton (inlined)
gaih_inet.constprop.7 (/usr/lib64/libc-2.26.so)
__GI_getaddrinfo (inlined)
[0xffffaa39b6108f3f] (/usr/bin/ping)
#
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Hendrick Brueckner <brueckner@linux.vnet.ibm.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Thomas Richter <tmricht@linux.vnet.ibm.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-is9tramondqa9jlxxsgcm9iz@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Adding support to display sample misc field in form
of letter for each bit:
# perf script -F +misc ...
sched-messaging 1414 K 28690.636582: 4590 cycles ...
sched-messaging 1407 U 28690.636600: 325620 cycles ...
sched-messaging 1414 K 28690.636608: 19473 cycles ...
misc field __________/
The misc bits are assigned to following letters:
PERF_RECORD_MISC_KERNEL K
PERF_RECORD_MISC_USER U
PERF_RECORD_MISC_HYPERVISOR H
PERF_RECORD_MISC_GUEST_KERNEL G
PERF_RECORD_MISC_GUEST_USER g
PERF_RECORD_MISC_MMAP_DATA* M
PERF_RECORD_MISC_COMM_EXEC E
PERF_RECORD_MISC_SWITCH_OUT S
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20180107160356.28203-9-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
While monitoring a multithread process with pid option, perf sometimes
may return sys_perf_event_open failure with 3(No such process) if any of
the process's threads die before we open the event. However, we want
perf continue monitoring the remaining threads and do not exit with
error.
Here, the patch enables perf_evsel::ignore_missing_thread for -p option
to ignore complete failure if any of threads die before we open the event.
But it may still return sys_perf_event_open failure with 22(Invalid) if we
monitors several event groups.
sys_perf_event_open: pid 28960 cpu 40 group_fd 118202 flags 0x8
sys_perf_event_open: pid 28961 cpu 40 group_fd 118203 flags 0x8
WARNING: Ignored open failure for pid 28962
sys_perf_event_open: pid 28962 cpu 40 group_fd [118203] flags 0x8
sys_perf_event_open failed, error -22
That is because when we ignore a missing thread, we change the thread_idx
without dealing with its fds, FD(evsel, cpu, thread). Then get_group_fd()
may return a wrong group_fd for the next thread and sys_perf_event_open()
return with 22.
sys_perf_event_open(){
...
if (group_fd != -1)
perf_fget_light()//to get corresponding group_leader by group_fd
...
if (group_leader)
if (group_leader->ctx->task != ctx->task)//should on the same task
goto err_context
...
}
This patch also fixes this bug by introducing perf_evsel__remove_fd() and
update_fds to allow removing fds for the missing thread.
Changes since v1:
- Change group_fd__remove() into a more genetic way without changing code logic
- Remove redundant condition
Changes since v2:
- Use a proper function name and add some comment.
- Multiline comment style fixes.
Committer testing:
Before this patch the recently added 'perf stat --per-thread' for system
wide counting would race while enumerating all threads using /proc:
[root@jouet ~]# perf stat --per-thread
failed to parse CPUs map: No such file or directory
Usage: perf stat [<options>] [<command>]
-C, --cpu <cpu> list of cpus to monitor in system-wide
-a, --all-cpus system-wide collection from all CPUs
[root@jouet ~]# perf stat --per-thread
failed to parse CPUs map: No such file or directory
Usage: perf stat [<options>] [<command>]
-C, --cpu <cpu> list of cpus to monitor in system-wide
-a, --all-cpus system-wide collection from all CPUs
[root@jouet ~]#
When, say, the kernel was being built, so lots of shortlived threads,
after this patch this doesn't happen.
Signed-off-by: Mengting Zhang <zhangmengting@huawei.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Cheng Jian <cj.chengjian@huawei.com>
Cc: Li Bin <huawei.libin@huawei.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/1513148513-6974-1-git-send-email-zhangmengting@huawei.com
[ Remove one use 'evlist' alias variable ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When we detect a different endianity we swap event before processing.
It's tricky for samples because we have no idea what's inside. We treat
it as an array of u64s, swap them and later on we swap back parts which
are different.
We mangle this way also the tracepoint raw data, which ends up in report
showing wrong data:
1.95% comm=Q^B pid=29285 prio=16777216 target_cpu=000
1.67% comm=l^B pid=0 prio=16777216 target_cpu=000
Luckily the traceevent library handles the endianity by itself (thank
you Steven!), so we can pass the RAW data directly in the other
endianity.
2.51% comm=beah-rhts-task pid=1175 prio=120 target_cpu=002
2.23% comm=kworker/0:0 pid=11566 prio=120 target_cpu=000
The fix is basically to swap back the raw data if different endianity is
detected.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/20171129184346.3656-1-jolsa@kernel.org
[ Add util/memswap.c to python-ext-sources to link missing mem_bswap_64() ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Help identify to the user the event with the unsupported sampling error.
Also suggest a corrective action.
BEFORE:
$ sudo ./oldperf record -e armv8_pmuv3/mem_access/,ccn/cycles/,armv8_pmuv3/l2d_cache/ true
Error:
PMU Hardware doesn't support sampling/overflow-interrupts.
AFTER:
$ sudo ./newperf record -e armv8_pmuv3/mem_access/,ccn/cycles/,armv8_pmuv3/l2d_cache/ true
Error:
ccn/cycles/: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'
Signed-off-by: Kim Phillips <kim.phillips@arm.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20171114150452.e846f2e23684c7d7d8ee706f@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Use a typed enum for the perf_evsel_config_term type enum. This allows
gcc to do much stronger type checks, and also check for missing case
statements.
I removed the unused _MAX member from the number.
It found one missing case. I'm not sure it's a real problem, so I just
turned it into a BUG_ON for now.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20171020202755.21410-1-andi@firstfloor.org
[ Renamed the enum name to term_type as per jolsa's request ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The Intel PMU event aliases have a implicit period= specifier to set the
default period.
Unfortunately this breaks overriding these periods with -c or -F,
because the alias terms look like they are user specified to the
internal parser, and user specified event qualifiers override the
command line options.
Track that they are coming from aliases by adding a "weak" state to the
term. Any weak terms don't override command line options.
I only did it for -c/-F for now, I think that's the only case that's
broken currently.
Before:
$ perf record -c 1000 -vv -e uops_issued.any
...
{ sample_period, sample_freq } 2000003
After:
$ perf record -c 1000 -vv -e uops_issued.any
...
{ sample_period, sample_freq } 1000
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20171020202755.21410-2-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Yet another fix for probing the max attr.precise_ip setting: it is not
enough settting attr.exclude_kernel for !root users, as they _can_
profile the kernel if the kernel.perf_event_paranoid sysctl is set to
-1, so check that as well.
Testing it:
As non root:
$ sysctl kernel.perf_event_paranoid
kernel.perf_event_paranoid = 2
$ perf record sleep 1
$ perf evlist -v
cycles:uppp: ..., exclude_kernel: 1, ... precise_ip: 3, ...
Now as non-root, but with kernel.perf_event_paranoid set set to the
most permissive value, -1:
$ sysctl kernel.perf_event_paranoid
kernel.perf_event_paranoid = -1
$ perf record sleep 1
$ perf evlist -v
cycles:ppp: ..., exclude_kernel: 0, ... precise_ip: 3, ...
$
I.e. non-root, default kernel.perf_event_paranoid: :uppp modifier = not allowed to sample the kernel,
non-root, most permissible kernel.perf_event_paranoid: :ppp = allowed to sample the kernel.
In both cases, use the highest available precision: attr.precise_ip = 3.
Reported-and-Tested-by: Ingo Molnar <mingo@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Fixes: d37a369790 ("perf evsel: Fix attr.exclude_kernel setting for default cycles:p")
Link: http://lkml.kernel.org/n/tip-nj2qkf75xsd6pw6hhjzfqqdx@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix buffer overflow for:
% perf stat -e msr/tsc/,cstate_core/c7-residency/ true
that causes glibc free list corruption. For some reason it doesn't
trigger in valgrind, but it is visible in AS:
=================================================================
==32681==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x603000003f5c at pc 0x0000005671ef bp 0x7ffdaaac9ac0 sp 0x7ffdaaac9ab0
READ of size 4 at 0x603000003f5c thread T0
#0 0x5671ee in perf_evsel__close_fd util/evsel.c:1196
#1 0x56c57a in perf_evsel__close util/evsel.c:1717
#2 0x55ed5f in perf_evlist__close util/evlist.c:1631
#3 0x4647e1 in __run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:749
#4 0x4648e3 in run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:767
#5 0x46e1bc in cmd_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:2785
#6 0x52f83d in run_builtin /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:296
#7 0x52fd49 in handle_internal_command /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:348
#8 0x5300de in run_argv /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:392
#9 0x5308f3 in main /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:530
#10 0x7f0672d13400 in __libc_start_main (/lib64/libc.so.6+0x20400)
#11 0x428419 in _start (/home/ak/hle/obj-perf/perf+0x428419)
0x603000003f5c is located 0 bytes to the right of 28-byte region [0x603000003f40,0x603000003f5c)
allocated by thread T0 here:
#0 0x7f0675139020 in calloc (/lib64/libasan.so.3+0xc7020)
#1 0x648a2d in zalloc util/util.h:23
#2 0x648a88 in xyarray__new util/xyarray.c:9
#3 0x566419 in perf_evsel__alloc_fd util/evsel.c:1039
#4 0x56b427 in perf_evsel__open util/evsel.c:1529
#5 0x56c620 in perf_evsel__open_per_thread util/evsel.c:1730
#6 0x461dea in create_perf_stat_counter /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:263
#7 0x4637d7 in __run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:600
#8 0x4648e3 in run_perf_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:767
#9 0x46e1bc in cmd_stat /home/ak/hle/linux-hle-2.6/tools/perf/builtin-stat.c:2785
#10 0x52f83d in run_builtin /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:296
#11 0x52fd49 in handle_internal_command /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:348
#12 0x5300de in run_argv /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:392
#13 0x5308f3 in main /home/ak/hle/linux-hle-2.6/tools/perf/perf.c:530
#14 0x7f0672d13400 in __libc_start_main (/lib64/libc.so.6+0x20400)
The event is allocated with cpus == 1, but freed with cpus == real number
When the evsel close function walks the file descriptors it exceeds the
fd xyarray boundaries and reads random memory.
v2:
Now that xyarrays save their original dimensions we can use these to
iterate the two dimensional fd arrays. Fix some users (close, ioctl) in
evsel.c to use these fields directly. This allows simplifying the code
and dropping quite a few function arguments. Adjust all callers by
removing the unneeded arguments.
The actual perf event reading still uses the original values from the
evsel list.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170811232634.30465-2-andi@firstfloor.org
[ Fix up xy_max_[xy]() -> xyarray__max_[xy]() ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Make perf stat use group read if there are groups defined. The group
read will get the values for all member of groups within a single
syscall instead of calling read syscall for every event.
We can see considerable less amount of kernel cycles spent on single
group read, than reading each event separately, like for following perf
stat command:
# perf stat -e {cycles,instructions} -I 10 -a sleep 1
Monitored with "perf stat -r 5 -e '{cycles:u,cycles:k}'"
Before:
24,325,676 cycles:u
297,040,775 cycles:k
1.038554134 seconds time elapsed
After:
25,034,418 cycles:u
158,256,395 cycles:k
1.036864497 seconds time elapsed
The perf_evsel__open fallback changes contributed by Andi Kleen.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20170726120206.9099-4-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Macro fusion merges two instructions to a single micro-op. Intel core
platform performs this hardware optimization under limited
circumstances.
For example, CMP + JCC can be "fused" and executed /retired together.
While with sampling this can result in the sample sometimes being on the
JCC and sometimes on the CMP. So for the fused instruction pair, they
could be considered together.
On Nehalem, fused instruction pairs:
cmp/test + jcc.
On other new CPU:
cmp/test/add/sub/and/inc/dec + jcc.
This patch adds an x86-specific function which checks if 2 instructions
are in a "fused" pair. For non-x86 arch, the function is just NULL.
Changelog:
v4: Move the CPU model checking to symbol__disassemble and save the CPU
family/model in arch structure.
It avoids checking every time when jump arrow printed.
v3: Add checking for Nehalem (CMP, TEST). For other newer Intel CPUs
just check it by default (CMP, TEST, ADD, SUB, AND, INC, DEC).
v2: Remove the original weak function. Arnaldo points out that doing it
as a weak function that will be overridden by the host arch doesn't
work. So now it's implemented as an arch-specific function.
Committer fix:
Do not access evsel->evlist->env->cpuid, ->env can be null, introduce
perf_evsel__env_cpuid(), just like perf_evsel__env_arch(), also used in
this function call.
The original patch was segfaulting 'perf top' + annotation.
But this essentially disables this fused instructions augmentation in
'perf top', the right thing is to get the cpuid from the running kernel,
left for a later patch tho.
Signed-off-by: Yao Jin <yao.jin@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1499403995-19857-2-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When no event is specified perf will use the "cycles" hardware event
with the highest precision available in the processor, and excluding
kernel events for non-root users, so make that clear in the event name
by setting the "u" event modifier, i.e. "cycles:upp".
E.g.:
The default for root:
# perf record usleep 1
# perf evlist -v
cycles:ppp: ..., precise_ip: 3, exclude_kernel: 0, ...
#
And for !root:
$ perf record usleep 1
$ perf evlist -v
cycles:uppp: ... , precise_ip: 3, exclude_kernel: 1, ...
$
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-lf29zcdl422i9knrgde0uwy3@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/urgent fixes from Arnaldo Carvalho de Melo:
User visible changes:
- Fix max attr.precise_ip probing to make perf use the best cycles:p
available in the processor for non root users (Arnaldo Carvalho de Melo)
- Fix processing of MMAP events for 32-bit binaries on 64-bit systems
when unwind support is not fully integrated, fixing DSO and symbol
resolution (Jiri Olsa)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We should set attr.exclude_kernel when probing for attr.precise_ip
level, otherwise !CAP_SYS_ADMIN users will not default to skidless
samples in capable hardware.
The increase in the paranoid level in commit 0161028b7c ("perf/core:
Change the default paranoia level to 2") broke this, fix it by excluding
kernel samples when probing.
Before:
$ perf record usleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.018 MB perf.data (6 samples) ]
$ perf evlist -v
cycles:u: sample_freq: 4000, sample_type: IP|TID|TIME|PERIOD, exclude_kernel: 1
After:
$ perf record usleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.018 MB perf.data (8 samples) ]
$ perf evlist -v
cycles:ppp: sample_freq: 4000, sample_type: IP|TID|TIME|PERIOD, exclude_kernel: 1, precise_ip: 3
^^^^^^^^^^^^^
^^^^^^^^^^^^^
^^^^^^^^^^^^^
$
To further clarify: we always set .exclude_kernel when non !CAP_SYS_ADMIN
users profile, its just on the attr.precise_ip probing that we weren't doing
so, fix it.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Fixes: 7f8d1ade1b ("perf tools: By default use the most precise "cycles" hw counter available")
Link: http://lkml.kernel.org/n/tip-t2qttwhbnua62o5gt75cueml@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Since commit 18e7a45af9 ("perf/x86: Reject non sampling events with
precise_ip") returns -EINVAL for sys_perf_event_open() with an attribute
with (attr.precise_ip > 0 && attr.sample_period == 0), just like is done
in the routine used to probe the max precise level when no events were
passed to 'perf record' or 'perf top', i.e.:
perf_evsel__new_cycles()
perf_event_attr__set_max_precise_ip()
The x86 code, in x86_pmu_hw_config(), which is called all the way from
sys_perf_event_open() did, starting with the aforementioned commit:
/* There's no sense in having PEBS for non sampling events: */
if (!is_sampling_event(event))
return -EINVAL;
Which makes it fail for cycles:ppp, cycles:pp and cycles:p, always using
just the non precise cycles variant.
To make sure that this is the case, I tested it, before this patch,
with:
# perf probe -L x86_pmu_hw_config
<x86_pmu_hw_config@/home/acme/git/linux/arch/x86/events/core.c:0>
0 int x86_pmu_hw_config(struct perf_event *event)
1 {
2 if (event->attr.precise_ip) {
<SNIP>
17 if (event->attr.precise_ip > precise)
18 return -EOPNOTSUPP;
/* There's no sense in having PEBS for non sampling events: */
21 if (!is_sampling_event(event))
22 return -EINVAL;
}
<SNIP>
# perf probe x86_pmu_hw_config:22
Added new events:
probe:x86_pmu_hw_config (on x86_pmu_hw_config:22)
probe:x86_pmu_hw_config_1 (on x86_pmu_hw_config:22)
You can now use it in all perf tools, such as:
perf record -e probe:x86_pmu_hw_config_1 -aR sleep 1
# perf trace -e perf_event_open,probe:x86_pmu_hwconfig*/max-stack=16/ perf record usleep 1
0.000 ( 0.015 ms): perf/4150 perf_event_open(attr_uptr: 0x7ffebc8ba110, cpu: -1, group_fd: -1 ) ...
0.015 ( ): probe:x86_pmu_hw_config:(ffffffff9c0065e1))
x86_pmu_hw_config ([kernel.kallsyms])
hsw_hw_config ([kernel.kallsyms])
x86_pmu_event_init ([kernel.kallsyms])
perf_try_init_event ([kernel.kallsyms])
perf_event_alloc ([kernel.kallsyms])
SYSC_perf_event_open ([kernel.kallsyms])
sys_perf_event_open ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
return_from_SYSCALL_64 ([kernel.kallsyms])
syscall (/usr/lib64/libc-2.24.so)
perf_event_attr__set_max_precise_ip (/home/acme/bin/perf)
perf_evsel__new_cycles (/home/acme/bin/perf)
perf_evlist__add_default (/home/acme/bin/perf)
cmd_record (/home/acme/bin/perf)
run_builtin (/home/acme/bin/perf)
handle_internal_command (/home/acme/bin/perf)
0.000 ( 0.021 ms): perf/4150 ... [continued]: perf_event_open()) = -1 EINVAL Invalid argument
0.023 ( 0.002 ms): perf/4150 perf_event_open(attr_uptr: 0x7ffebc8ba110, cpu: -1, group_fd: -1 ) ...
0.025 ( ): probe:x86_pmu_hw_config:(ffffffff9c0065e1))
x86_pmu_hw_config ([kernel.kallsyms])
hsw_hw_config ([kernel.kallsyms])
x86_pmu_event_init ([kernel.kallsyms])
perf_try_init_event ([kernel.kallsyms])
perf_event_alloc ([kernel.kallsyms])
SYSC_perf_event_open ([kernel.kallsyms])
sys_perf_event_open ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
return_from_SYSCALL_64 ([kernel.kallsyms])
syscall (/usr/lib64/libc-2.24.so)
perf_event_attr__set_max_precise_ip (/home/acme/bin/perf)
perf_evsel__new_cycles (/home/acme/bin/perf)
perf_evlist__add_default (/home/acme/bin/perf)
cmd_record (/home/acme/bin/perf)
run_builtin (/home/acme/bin/perf)
handle_internal_command (/home/acme/bin/perf)
0.023 ( 0.004 ms): perf/4150 ... [continued]: perf_event_open()) = -1 EINVAL Invalid argument
0.028 ( 0.002 ms): perf/4150 perf_event_open(attr_uptr: 0x7ffebc8ba110, cpu: -1, group_fd: -1 ) ...
0.030 ( ): probe:x86_pmu_hw_config:(ffffffff9c0065e1))
x86_pmu_hw_config ([kernel.kallsyms])
hsw_hw_config ([kernel.kallsyms])
x86_pmu_event_init ([kernel.kallsyms])
perf_try_init_event ([kernel.kallsyms])
perf_event_alloc ([kernel.kallsyms])
SYSC_perf_event_open ([kernel.kallsyms])
sys_perf_event_open ([kernel.kallsyms])
do_syscall_64 ([kernel.kallsyms])
return_from_SYSCALL_64 ([kernel.kallsyms])
syscall (/usr/lib64/libc-2.24.so)
perf_event_attr__set_max_precise_ip (/home/acme/bin/perf)
perf_evsel__new_cycles (/home/acme/bin/perf)
perf_evlist__add_default (/home/acme/bin/perf)
cmd_record (/home/acme/bin/perf)
run_builtin (/home/acme/bin/perf)
handle_internal_command (/home/acme/bin/perf)
0.028 ( 0.004 ms): perf/4150 ... [continued]: perf_event_open()) = -1 EINVAL Invalid argument
41.018 ( 0.012 ms): perf/4150 perf_event_open(attr_uptr: 0x7ffebc8b5dd0, pid: -1, group_fd: -1, flags: FD_CLOEXEC) = 4
41.065 ( 0.011 ms): perf/4150 perf_event_open(attr_uptr: 0x3c7db78, pid: -1, group_fd: -1, flags: FD_CLOEXEC) = 4
41.080 ( 0.006 ms): perf/4150 perf_event_open(attr_uptr: 0x3c7db78, pid: -1, group_fd: -1, flags: FD_CLOEXEC) = 4
41.103 ( 0.010 ms): perf/4150 perf_event_open(attr_uptr: 0x3c4e748, pid: 4151 (perf), group_fd: -1, flags: FD_CLOEXEC) = 4
41.115 ( 0.006 ms): perf/4150 perf_event_open(attr_uptr: 0x3c4e748, pid: 4151 (perf), cpu: 1, group_fd: -1, flags: FD_CLOEXEC) = 5
41.122 ( 0.004 ms): perf/4150 perf_event_open(attr_uptr: 0x3c4e748, pid: 4151 (perf), cpu: 2, group_fd: -1, flags: FD_CLOEXEC) = 6
41.128 ( 0.008 ms): perf/4150 perf_event_open(attr_uptr: 0x3c4e748, pid: 4151 (perf), cpu: 3, group_fd: -1, flags: FD_CLOEXEC) = 8
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.017 MB perf.data (2 samples) ]
#
I.e. that return -EINVAL in x86_pmu_hw_config() is hit three times.
So fix it by just setting attr.sample_period
Now, after this patch:
# perf trace --max-stack=2 -e perf_event_open,probe:x86_pmu_hw_config* perf record usleep 1
[ perf record: Woken up 1 times to write data ]
0.000 ( 0.017 ms): perf/8469 perf_event_open(attr_uptr: 0x7ffe36c27d10, pid: -1, cpu: 3, group_fd: -1, flags: FD_CLOEXEC) = 4
syscall (/usr/lib64/libc-2.24.so)
perf_event_open_cloexec_flag (/home/acme/bin/perf)
0.050 ( 0.031 ms): perf/8469 perf_event_open(attr_uptr: 0x24ebb78, pid: -1, group_fd: -1, flags: FD_CLOEXEC) = 4
syscall (/usr/lib64/libc-2.24.so)
perf_evlist__config (/home/acme/bin/perf)
0.092 ( 0.040 ms): perf/8469 perf_event_open(attr_uptr: 0x24ebb78, pid: -1, group_fd: -1, flags: FD_CLOEXEC) = 4
syscall (/usr/lib64/libc-2.24.so)
perf_evlist__config (/home/acme/bin/perf)
0.143 ( 0.007 ms): perf/8469 perf_event_open(attr_uptr: 0x24bc748, cpu: -1, group_fd: -1 ) = 4
syscall (/usr/lib64/libc-2.24.so)
perf_event_attr__set_max_precise_ip (/home/acme/bin/perf)
0.161 ( 0.007 ms): perf/8469 perf_event_open(attr_uptr: 0x24bc748, pid: 8470 (perf), group_fd: -1, flags: FD_CLOEXEC) = 4
syscall (/usr/lib64/libc-2.24.so)
perf_evsel__open (/home/acme/bin/perf)
0.171 ( 0.005 ms): perf/8469 perf_event_open(attr_uptr: 0x24bc748, pid: 8470 (perf), cpu: 1, group_fd: -1, flags: FD_CLOEXEC) = 5
syscall (/usr/lib64/libc-2.24.so)
perf_evsel__open (/home/acme/bin/perf)
0.180 ( 0.007 ms): perf/8469 perf_event_open(attr_uptr: 0x24bc748, pid: 8470 (perf), cpu: 2, group_fd: -1, flags: FD_CLOEXEC) = 6
syscall (/usr/lib64/libc-2.24.so)
perf_evsel__open (/home/acme/bin/perf)
0.190 ( 0.005 ms): perf/8469 perf_event_open(attr_uptr: 0x24bc748, pid: 8470 (perf), cpu: 3, group_fd: -1, flags: FD_CLOEXEC) = 8
syscall (/usr/lib64/libc-2.24.so)
perf_evsel__open (/home/acme/bin/perf)
[ perf record: Captured and wrote 0.017 MB perf.data (7 samples) ]
#
The probe one called from perf_event_attr__set_max_precise_ip() works
the first time, with attr.precise_ip = 3, wit hthe next ones being the
per cpu ones for the cycles:ppp event.
And here is the text from a report and alternative proposed patch by
Thomas-Mich Richter:
---
On s390 the counter and sampling facility do not support a precise IP
skid level and sometimes returns EOPNOTSUPP when structure member
precise_ip in struct perf_event_attr is not set to zero.
On s390 commnd 'perf record -- true' fails with error EOPNOTSUPP. This
happens only when no events are specified on command line.
The functions called are
...
--> perf_evlist__add_default
--> perf_evsel__new_cycles
--> perf_event_attr__set_max_precise_ip
The last function determines the value of structure member precise_ip by
invoking the perf_event_open() system call and checking the return code.
The first successful open is the value for precise_ip.
However the value is determined without setting member sample_period and
indicates no sampling.
On s390 the counter facility and sampling facility are different. The
above procedure determines a precise_ip value of 3 using the counter
facility. Later it uses the sampling facility with a value of 3 and
fails with EOPNOTSUPP.
---
v2: Older compilers (e.g. gcc 4.4.7) don't support referencing members
of unnamed union members in the container struct initialization, so
move from:
struct perf_event_attr attr = {
...
.sample_period = 1,
};
to right after it as:
struct perf_event_attr attr = {
...
};
attr.sample_period = 1;
v3: We need to reset .sample_period to 0 to let the users of
perf_evsel__new_cycles() to properly setup attr.sample_period or
attr.sample_freq. Reported by Ingo Molnar.
Reported-and-Acked-by: Thomas-Mich Richter <tmricht@linux.vnet.ibm.com>
Acked-by: Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Fixes: 18e7a45af9 ("perf/x86: Reject non sampling events with precise_ip")
Link: http://lkml.kernel.org/n/tip-yv6nnkl7tzqocrm0hl3x7vf1@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The kernel has a special check for a specific irq_vectors trace event.
TRACE_EVENT_PERF_PERM(irq_work_exit,
is_sampling_event(p_event) ? -EPERM : 0);
The perf-record fails for this irq_vectors event when it is present,
like when using a wildcard:
root@skl:/tmp# perf record -a -e irq_vectors:* sleep 2
Error:
You may not have permission to collect system-wide stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid,
which controls use of the performance events system by
unprivileged users (without CAP_SYS_ADMIN).
The current value is 2:
-1: Allow use of (almost) all events by all users
>= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK
>= 1: Disallow CPU event access by users without CAP_SYS_ADMIN
>= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN
To make this setting permanent, edit /etc/sysctl.conf too, e.g.:
kernel.perf_event_paranoid = -1
This patch prints out the exact sub event that failed with EPERM for
wildcards to help in understanding what went wrong when this event is
present:
After the patch:
root@skl:/tmp# perf record -a -e irq_vectors:* sleep 2
Error:
No permission to enable irq_vectors:irq_work_exit event.
You may not have permission to collect system-wide stats.
......
Committer notes:
So we have a lot of irq_vectors events:
[root@jouet ~]# perf list irq_vectors:*
List of pre-defined events (to be used in -e):
irq_vectors:call_function_entry [Tracepoint event]
irq_vectors:call_function_exit [Tracepoint event]
irq_vectors:call_function_single_entry [Tracepoint event]
irq_vectors:call_function_single_exit [Tracepoint event]
irq_vectors:deferred_error_apic_entry [Tracepoint event]
irq_vectors:deferred_error_apic_exit [Tracepoint event]
irq_vectors:error_apic_entry [Tracepoint event]
irq_vectors:error_apic_exit [Tracepoint event]
irq_vectors:irq_work_entry [Tracepoint event]
irq_vectors:irq_work_exit [Tracepoint event]
irq_vectors:local_timer_entry [Tracepoint event]
irq_vectors:local_timer_exit [Tracepoint event]
irq_vectors:reschedule_entry [Tracepoint event]
irq_vectors:reschedule_exit [Tracepoint event]
irq_vectors:spurious_apic_entry [Tracepoint event]
irq_vectors:spurious_apic_exit [Tracepoint event]
irq_vectors:thermal_apic_entry [Tracepoint event]
irq_vectors:thermal_apic_exit [Tracepoint event]
irq_vectors:threshold_apic_entry [Tracepoint event]
irq_vectors:threshold_apic_exit [Tracepoint event]
irq_vectors:x86_platform_ipi_entry [Tracepoint event]
irq_vectors:x86_platform_ipi_exit [Tracepoint event]
#
And some may be sampled:
[root@jouet ~]# perf record -e irq_vectors:local* sleep 20s
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.020 MB perf.data (2 samples) ]
[root@jouet ~]# perf report -D | egrep 'stats:|events:'
Aggregated stats:
TOTAL events: 155
MMAP events: 144
COMM events: 2
EXIT events: 1
SAMPLE events: 2
MMAP2 events: 4
FINISHED_ROUND events: 1
TIME_CONV events: 1
irq_vectors:local_timer_entry stats:
TOTAL events: 1
SAMPLE events: 1
irq_vectors:local_timer_exit stats:
TOTAL events: 1
SAMPLE events: 1
[root@jouet ~]#
But, as shown in the tracepoint definition at the start of this message,
some, like "irq_vectors:irq_work_exit", may not be sampled, just counted,
i.e. if we try to sample, as when using 'perf record', we get an error:
[root@jouet ~]# perf record -e irq_vectors:irq_work_exit
Error:
You may not have permission to collect system-wide stats.
Consider tweaking /proc/sys/kernel/perf_event_paranoid,
<SNIP>
The error message is misleading, this patch will help in pointing out
what is the event causing such an error, but the error message needs
improvement, i.e. we need to figure out a way to check if a tracepoint
is counting only, like this one, when all we can do is to count it with
'perf stat', at most printing the delta using interval printing, as in:
[root@jouet ~]# perf stat -I 5000 -e irq_vectors:irq_work_*
# time counts unit events
5.000168871 0 irq_vectors:irq_work_entry
5.000168871 0 irq_vectors:irq_work_exit
10.000676730 0 irq_vectors:irq_work_entry
10.000676730 0 irq_vectors:irq_work_exit
15.001122415 0 irq_vectors:irq_work_entry
15.001122415 0 irq_vectors:irq_work_exit
20.001298051 0 irq_vectors:irq_work_entry
20.001298051 0 irq_vectors:irq_work_exit
25.001485020 1 irq_vectors:irq_work_entry
25.001485020 1 irq_vectors:irq_work_exit
30.001658706 0 irq_vectors:irq_work_entry
30.001658706 0 irq_vectors:irq_work_exit
^C 32.045711878 0 irq_vectors:irq_work_entry
32.045711878 0 irq_vectors:irq_work_exit
[root@jouet ~]#
But at least, when we use a wildcard, this patch helps a bit.
Signed-off-by: Yao Jin <yao.jin@linux.intel.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1491566932-503-1-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support for a new JSON event attribute to name MetricExpr for better
output in perf stat.
If the event has no MetricName it uses the normal event name instead to
describe the metric.
Before
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
time unc_p_freq_max_os_cycles
1.000149775 15.7
2.000344807 19.3
3.000502544 16.7
4.000640656 6.6
5.000779955 9.9
After
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
time freq_max_os_cycles %
1.000149775 15.7
2.000344807 19.3
3.000502544 16.7
4.000640656 6.6
5.000779955 9.9
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-13-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add generic infrastructure to perf stat to output ratios for
"MetricExpr" entries in the event lists. Many events are more useful as
ratios than in raw form, typically some count in relation to total
ticks.
Transfer the MetricExpr information from the alias to the evsel.
We mark the events that need to be collected for MetricExpr, and also
link the events using them with a pointer. The code is careful to always
prefer the right event in the same group to minimize multiplexing
errors. At the moment only a single relation is supported.
Then add a rblist to the stat shadow code that remembers stats based on
the cpu and context.
Then finally update and retrieve and print these values similarly to the
existing hardcoded perf metrics. We use the simple expression parser
added earlier to evaluate the expression.
Normally we just output the result without further commentary, but for
--metric-only this would lead to empty columns. So for this case use the
original event as description.
There is no attempt to automatically add the MetricExpr event, if it is
missing, however we suggest it to the user, because the user tool
doesn't have enough information to reliably construct a group that is
guaranteed to schedule. So we leave that to the user.
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}'
1.000147889 800,085,181 unc_p_clockticks
1.000147889 93,126,241 unc_p_freq_max_os_cycles # 11.6
2.000448381 800,218,217 unc_p_clockticks
2.000448381 142,516,095 unc_p_freq_max_os_cycles # 17.8
3.000639852 800,243,057 unc_p_clockticks
3.000639852 162,292,689 unc_p_freq_max_os_cycles # 20.3
% perf stat -a -I 1000 -e '{unc_p_clockticks,unc_p_freq_max_os_cycles}' --metric-only
# time freq_max_os_cycles %
1.000127077 0.9
2.000301436 0.7
3.000456379 0.0
v2: Change from DivideBy to MetricExpr
v3: Use expr__ prefix. Support more than one other event.
v4: Update description
v5: Only print warning message once for multiple PMUs.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20170320201711.14142-11-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
