Allowing one to hook into the syscalls:sys_enter_NAME tracepoints,
an example is provided that hooks into the 'openat' syscall.
Using it with the probe:vfs_getname probe into getname_flags to get the
filename args as it is copied from userspace:
# perf probe -l
probe:vfs_getname (on getname_flags:73@acme/git/linux/fs/namei.c with pathname)
# perf trace -e probe:*getname,tools/perf/examples/bpf/sys_enter_openat.c cat /etc/passwd > /dev/null
0.000 probe:vfs_getname:(ffffffffbd2a8983) pathname="/etc/ld.so.preload"
0.022 syscalls:sys_enter_openat:dfd: CWD, filename: 0xafbe8da8, flags: CLOEXEC
0.027 probe:vfs_getname:(ffffffffbd2a8983) pathname="/etc/ld.so.cache"
0.054 syscalls:sys_enter_openat:dfd: CWD, filename: 0xafdf0ce0, flags: CLOEXEC
0.057 probe:vfs_getname:(ffffffffbd2a8983) pathname="/lib64/libc.so.6"
0.316 probe:vfs_getname:(ffffffffbd2a8983) pathname="/usr/lib/locale/locale-archive"
0.375 syscalls:sys_enter_openat:dfd: CWD, filename: 0xe2b2b0b4
0.379 probe:vfs_getname:(ffffffffbd2a8983) pathname="/etc/passwd"
#
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: https://lkml.kernel.org/n/tip-2po9jcqv1qgj0koxlg8kkg30@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support for s390 auxiliary trace support.
Use 'perf record -e rbd000 -- ls' to create the perf.data file.
Use 'perf report' to display the auxiliary trace data.
Output before:
[root@s35lp76 perf]# ./perf report --stdio
0x128 [0x10]: failed to process type: 70
Error:
failed to process sample
[root@s35lp76 perf]#
Output after:
[root@s35lp76 perf]# ./perf report --stdio
18.21% 18.21% ls [kernel.kallsyms] [k] ftrace_likely_update
9.52% 9.52% ls [kernel.kallsyms] [k] lock_acquire
9.38% 9.38% ls [kernel.kallsyms] [k] lock_release
3.45% 3.45% ls [kernel.kallsyms] [k] lock_acquired
2.88% 2.88% ls [kernel.kallsyms] [k] link_path_walk
2.63% 2.63% ls [kernel.kallsyms] [k] __d_lookup
2.38% 2.38% ls [kernel.kallsyms] [k] __d_lookup_rcu
2.04% 2.04% ls [kernel.kallsyms] [k] ___might_sleep
1.83% 1.83% ls [kernel.kallsyms] [k] debug_lockdep_rcu_enabled
1.44% 1.44% ls [kernel.kallsyms] [k] dput
....
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Link: http://lkml.kernel.org/r/20180802074622.13641-4-tmricht@linux.ibm.com
[ Use PRI[xd]64 to fix the build on debian:experimental-x-mips (gcc 8.1.0) and others ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add support for s390 auxiliary trace support.
Use 'perf record -e rbd000' to create the perf.data file. The event
also has the symbolic name SF_CYCLES_BASIC_DIAG, using 'perf record -e
SF_CYCLES_BASIC_DIAG' is equivalent.
Use 'perf report -D' to display the auxiliary trace data.
Output before:
0 0 0x25a66 [0x30]: PERF_RECORD_AUXTRACE size: 0x40000
offset: 0 ref: 0 idx: 4 tid: -1 cpu: 4
Nothing else
Output after:
0 0 0x25a66 [0x30]: PERF_RECORD_AUXTRACE size: 0x40000
offset: 0 ref: 0 idx: 4 tid: -1 cpu: 4
.
. ... s390 AUX data: size 262144 bytes
[00000000] Basic Def:0001 Inst:0000 TW AS:3 ASN:0xffff IA:0x0000000000c2f1bc
CL:1 HPP:0x8000000000000000 GPP:000000000000000000
[0x000020] Diag Def:8005
[0x0000bf] Basic Def:0001 Inst:0000 TW AS:3 ASN:0xffff IA:0x0000000000c2f1bc
CL:1 HPP:0x8000000000000000 GPP:000000000000000000
[0x0000df] Diag Def:8005
[0x00017e] Basic Def:0001 Inst:0000 TW AS:3 ASN:0xffff IA:0x0000000000c2f1bc
CL:1 HPP:0x8000000000000000 GPP:000000000000000000
....
[0x000fc0] Trailer F T bsdes:32 dsdes:159 Overflow:0 Time:0xd4ab59a8450fa108
C:1 TOD:0xd4ab4ec98ceb3832 1:0x8000000000000000 2:0xd4ab4ec98ceb3832
This output is shown for every sampled data block. The
output contains the
- basic-sampling data entry
- diagnostic-sampling data entry
- trailer entry
The basic sampling entry and diagnostic sampling entry sizes can be
extracted using the trailer entries in the SDB. On older hardware these
values (bsdes and dsdes in the trailer entry) are reserved and zero.
Older hardware use hard coded values based on the s390 machine type.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Link: http://lkml.kernel.org/r/20180802074622.13641-3-tmricht@linux.ibm.com
Link: http://lkml.kernel.org/r/eda2632e-7919-5ffd-5f68-821e77d216fa@linux.ibm.com
[ Merged a fix for a 'tipe puned' problem reported by Michael Ellerman see last Link tag. ]
[ Removed __packed from two structs, they're already naturally packed and having that. ]
[ attribute breaks the build in gcc 8.1.1 mips, 4.4.7 x86_64, 7.1.1 ARCompact ISA, etc) ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Add initial support for s390 auxiliary traces using the CPU-Measurement
Sampling Facility.
Support and ignore PERF_REPORT_AUXTRACE_INFO records in the perf data
file. Later patches will show the contents of the auxiliary traces.
Setup the auxtrace queues and data structures for s390. A raw dump of
the perf.data file now does not show an error when an auxtrace event is
encountered.
Output before:
[root@s35lp76 perf]# ./perf report -D -i perf.data.auxtrace
0x128 [0x10]: failed to process type: 70
Error:
failed to process sample
0x128 [0x10]: event: 70
.
. ... raw event: size 16 bytes
. 0000: 00 00 00 46 00 00 00 10 00 00 00 00 00 00 00 00 ...F............
0x128 [0x10]: PERF_RECORD_AUXTRACE_INFO type: 0
[root@s35lp76 perf]#
Output after:
# ./perf report -D -i perf.data.auxtrace |fgrep PERF_RECORD_AUXTRACE
0 0 0x128 [0x10]: PERF_RECORD_AUXTRACE_INFO type: 5
0 0 0x25a66 [0x30]: PERF_RECORD_AUXTRACE size: 0x40000
offset: 0 ref: 0 idx: 4 tid: -1 cpu: 4
....
Additional notes about the underlying hardware and software
implementation, provided by Hendrik Brueckner (see Link: below).
=============================================================================
The CPU-Measurement Facility (CPU-MF) provides a set of functions to obtain
performance information on the mainframe. Basically, it was introduced
with System z10 years ago for the z/Architecture, that means, 64-bit.
For Linux, there are two facilities of interest, counter facility and sampling
facility. The counter facility provides hardware counters for instructions,
cycles, crypto-activities, and many more.
The sampling facility is a hardware sampler that when started will write
samples at a particular interval into a sampling buffer. At some point,
for example, if a sample block is full, it generates an interrupt to collect
samples (while the sampler continues to run).
Few years ago, I started to provide the a perf PMU to use the counter
and sampling facilities. Recently, the device driver was updated to also
"export" the sampling buffer into the AUX area. Thomas now completed the
related perf work to interpret and process these AUX data.
If people are more interested in the sampling facility, they can have a
look into:
- The Load-Program-Parameter and the CPU-Measurement Facilities, SA23-2260-05
http://www-01.ibm.com/support/docview.wss?uid=isg26fcd1cc32246f4c8852574ce0044734a
and to learn how-to use it for Linux on Z, have look at chapter 54,
"Using the CPU-measurement facilities" in the:
- Device Drivers, Features, and Commands, SC33-8411-34
http://public.dhe.ibm.com/software/dw/linux390/docu/l416dd34.pdf
=============================================================================
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Link: http://lkml.kernel.org/r/20180803100758.GA28475@linux.ibm.com
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Link: http://lkml.kernel.org/r/20180802074622.13641-2-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To avoid having to ask libtraceevent to find a field by name when
handling each tracepoint event, we setup a struct syscall_tp with
a tp_field struct having an extractor function + the offset for the
"id", "args" and "ret" raw_syscalls:sys_{enter,exit} tracepoints.
Now that we want to do the same with syscalls:sys_{entry,exit}_NAME
individual syscall tracepoints, where we have "id" as "__syscall_nr" and
"args" as the actual series of per syscall parameters, we need more
flexibility from the routines that set up these pre-looked up syscall
tracepoint arg fields.
The next cset will use it.
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: Steven Rostedt <rostedt@goodmis.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-v59q5e0jrlzkpl9a1c7t81ni@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We were using the beautifiers only when processing the
raw_syscalls:sys_enter events, but we can as well use them for the
syscalls:sys_enter_NAME events, as the layout is the same.
Some more tweaking is needed as we're processing them straight away,
i.e. there is no buffering in the sys_enter_NAME event to wait for
things like vfs_getname to provide pointer contents and then flushing
at sys_exit_NAME, so we need to state in the syscall_arg that this
is unbuffered, just print the pointer values, beautifying just
non-pointer syscall args.
This just shows an alternative way of processing tracepoints, that we
will end up using when creating "tracepoint" payloads that already copy
pointer contents (or chunks of it, i.e. not the whole filename, but just
the end of it, not all the bf for a read/write, but just the start,
etc), directly in the kernel using eBPF.
E.g.:
# perf trace -e syscalls:*enter*sleep,*sleep sleep 1
0.303 ( ): syscalls:sys_enter_nanosleep:rqtp: 0x7ffc93d5ecc0
0.305 (1000.229 ms): sleep/8746 nanosleep(rqtp: 0x7ffc93d5ecc0) = 0
# perf trace -e syscalls:*_*sleep,*sleep sleep 1
0.288 ( ): syscalls:sys_enter_nanosleep:rqtp: 0x7ffecde87e40
0.289 ( ): sleep/8748 nanosleep(rqtp: 0x7ffecde87e40) ...
1000.479 ( ): syscalls:sys_exit_nanosleep:0x0
0.289 (1000.208 ms): sleep/8748 ... [continued]: nanosleep()) = 0
#
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: Steven Rostedt <rostedt@goodmis.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-jehyd2zwhw00z3p7v7mg9632@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the vfs_getname() wannabe tracepoint is in place:
# perf probe -l
probe:vfs_getname (on getname_flags:73@acme/git/linux/fs/namei.c with pathname)
#
'perf trace' will use it to get the pathname when it is copied from
userspace to the kernel, right after syscalls:sys_enter_open, copied
in the 'probe:vfs_getname', stash it somewhere and then, at
syscalls:sys_exit_open time, if the 'open' return is not -1, i.e. a
successfull open syscall, associate that pathname to this return, i.e.
the fd.
We were not doing this for the 'openat' syscall, which would cause 'perf
trace' to fallback to using /proc to get the fd, change it so that we
use what we got from probe:vfs_getname, reducing the 'openat'
beautification process cost, ditching the syscalls performed to read
procfs state and avoiding some possible races in the process.
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: https://lkml.kernel.org/n/tip-xnp44ao3bkb6ejeczxfnjwsh@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The next example scripts need the definition for the BPF functions, i.e.
things like BPF_FUNC_probe_read, and in time will require lots of other
definitions found in uapi/linux/bpf.h, so include it from the bpf.h file
included from the eBPF scripts build with clang via '-e bpf_script.c'
like in this example:
$ tail -8 tools/perf/examples/bpf/5sec.c
#include <bpf.h>
int probe(hrtimer_nanosleep, rqtp->tv_sec)(void *ctx, int err, long sec)
{
return sec == 5;
}
license(GPL);
$
That 'bpf.h' include in the 5sec.c eBPF example will come from a set of
header files crafted for building eBPF objects, that in a end-user
system will come from:
/usr/lib/perf/include/bpf/bpf.h
And will include <uapi/linux/bpf.h> either from the place where the
kernel was built, or from a kernel-devel rpm package like:
-working-directory /lib/modules/4.17.9-100.fc27.x86_64/build
That is set up by tools/perf/util/llvm-utils.c, and can be overriden
by setting the 'kbuild-dir' variable in the "llvm" ~/.perfconfig file,
like:
# cat ~/.perfconfig
[llvm]
kbuild-dir = /home/foo/git/build/linux
This usually doesn't need any change, just documenting here my findings
while working with this code.
In the future we may want to instead just use what is in
/usr/include/linux/bpf.h, that comes from the UAPI provided from the
kernel sources, for now, to avoid getting the kernel's non-UAPI
"linux/bpf.h" file, that will cause clang to fail and is not what we
want anyway (no BPF function definitions, etc), do it explicitely by
asking for "uapi/linux/bpf.h".
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: https://lkml.kernel.org/n/tip-zd8zeyhr2sappevojdem9xxt@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
After update of kernel, the perf tool doesn't run anymore on my 32MB RAM
powerpc board, but still runs on a 128MB RAM board:
~# strace perf
execve("/usr/sbin/perf", ["perf"], [/* 12 vars */]) = -1 ENOMEM (Cannot allocate memory)
--- SIGSEGV {si_signo=SIGSEGV, si_code=SI_KERNEL, si_addr=0} ---
+++ killed by SIGSEGV +++
Segmentation fault
objdump -x shows that .bss section has a huge size of 24Mbytes:
27 .bss 016baca8 101cebb8 101cebb8 001cd988 2**3
With especially the following objects having quite big size:
10205f80 l O .bss 00140000 runtime_cycles_stats
10345f80 l O .bss 00140000 runtime_stalled_cycles_front_stats
10485f80 l O .bss 00140000 runtime_stalled_cycles_back_stats
105c5f80 l O .bss 00140000 runtime_branches_stats
10705f80 l O .bss 00140000 runtime_cacherefs_stats
10845f80 l O .bss 00140000 runtime_l1_dcache_stats
10985f80 l O .bss 00140000 runtime_l1_icache_stats
10ac5f80 l O .bss 00140000 runtime_ll_cache_stats
10c05f80 l O .bss 00140000 runtime_itlb_cache_stats
10d45f80 l O .bss 00140000 runtime_dtlb_cache_stats
10e85f80 l O .bss 00140000 runtime_cycles_in_tx_stats
10fc5f80 l O .bss 00140000 runtime_transaction_stats
11105f80 l O .bss 00140000 runtime_elision_stats
11245f80 l O .bss 00140000 runtime_topdown_total_slots
11385f80 l O .bss 00140000 runtime_topdown_slots_retired
114c5f80 l O .bss 00140000 runtime_topdown_slots_issued
11605f80 l O .bss 00140000 runtime_topdown_fetch_bubbles
11745f80 l O .bss 00140000 runtime_topdown_recovery_bubbles
This is due to commit 4d255766d2 ("perf: Bump max number of cpus
to 1024"), because many tables are sized with MAX_NR_CPUS
This patch gives the opportunity to redefine MAX_NR_CPUS via
$ make EXTRA_CFLAGS=-DMAX_NR_CPUS=1
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: linuxppc-dev@lists.ozlabs.org
Link: http://lkml.kernel.org/r/20170922112043.8349468C57@po15668-vm-win7.idsi0.si.c-s.fr
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Before:
libbpf: license of tools/perf/examples/bpf/etcsnoop.c is GPL
libbpf: section(6) version, size 4, link 0, flags 3, type=1
libbpf: kernel version of tools/perf/examples/bpf/etcsnoop.c is 41200
libbpf: section(7) .symtab, size 120, link 1, flags 0, type=2
bpf: config program 'syscalls:sys_enter_openat'
libbpf: load bpf program failed: Operation not permitted
libbpf: failed to load program 'syscalls:sys_enter_openat'
libbpf: failed to load object 'tools/perf/examples/bpf/etcsnoop.c'
bpf: load objects failed
After: (just the last line changes)
bpf: load objects failed: err=-4009: (Incorrect kernel version)
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: https://lkml.kernel.org/n/tip-wi44iid0yjfht3lcvplc75fm@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
PMU event descriptions use 7 spaces + '[' or 8 spaces as indentation.
Metric groups used a tab + '['. This patch unifies it to the way PMU
event descriptions are indented.
BEFORE:
$ perf list
[...]
Metric Groups:
DSB:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
[...]
AFTER:
$ perf list
[...]
Metric Groups:
DSB:
DSB_Coverage
[Fraction of Uops delivered by the DSB (aka Decoded Icache; or Uop Cache)]
[...]
Signed-off-by: Michael Petlan <mpetlan@redhat.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Kim Phillips <kim.phillips@arm.com>
LPU-Reference: 771439042.22924766.1532986504631.JavaMail.zimbra@redhat.com
Link: https://lkml.kernel.org/n/tip-mlo850517m6u1rbjndvd1bwr@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf tool build and install is controlled via a Makefile. The
'install' rule creates directories and copies files. Among them are
header files installed in /usr/lib/include/perf/bpf/.
However all listed examples are installing its header files in
/usr/lib/<tool-name>/...[/include]/header.h
and not in
/usr/lib/include/<tool-name>/.../header.h.
Background information:
Building the Fedora 28 glibc RPM on s390x and s390 fails on s390 (gcc
-m31) as gcc is not able to find header-files like stdbool.h.
In the glibc.spec file, you can see that glibc is configured with
"--with-headers". In this case, first -nostdinc is added to the CFLAGS
and then further include paths are added via -isystem. One of those
paths should contain header files like stdbool.h.
In order to get this path, gcc is invoked with:
- on Fedora 28 (with 4.18 kernel):
$ gcc -print-file-name=include
/usr/lib/gcc/s390x-redhat-linux/8/include
$ gcc -m31 -print-file-name=include
/usr/lib/gcc/s390x-redhat-linux/8/../../../../lib/include
=> If perf is installed, this is: /usr/lib/include
On my machine this directory is only containing the directory "perf".
If perf is not installed gcc returns: /usr/lib/gcc/s390x-redhat-linux/8/include
- on Ubuntu 18.04 (with 4.15 kernel):
$ gcc -print-file-name=include
/usr/lib/gcc/s390x-linux-gnu/7/include
$ gcc -m31 -print-file-name=include
/usr/lib/gcc/s390x-linux-gnu/7/include
=> gcc returns the correct path even if perf is installed.
In each case, the introduction of the subdirectory /usr/lib/include
leads to the regression that one can not build the glibc RPM for s390
anymore as gcc can not find headers like stdbool.h.
To remedy this install bpf.h to /usr/lib/perf/include/bpf/bpf.h
Output before using the command 'perf test -Fv 40':
echo '...[bpf-program-source]...' | /usr/bin/clang ... \
-I/root/lib/include/perf/bpf ...
^^^^^^^^^^^^
...
[root@p23lp27 perf]# perf test -F 40
40: BPF filter :
40.1: Basic BPF filtering : Ok
40.2: BPF pinning : Ok
40.3: BPF prologue generation : Ok
40.4: BPF relocation checker : Ok
[root@p23lp27 perf]#
Output after using command 'perf test -Fv 40':
echo '...[bpf-program-source]...' | /usr/bin/clang ... \
-I/root/lib/perf/include/bpf ...
^^^^^^^^^^^^
...
[root@p23lp27 perf]# perf test -F 40
40: BPF filter :
40.1: Basic BPF filtering : Ok
40.2: BPF pinning : Ok
40.3: BPF prologue generation : Ok
40.4: BPF relocation checker : Ok
[root@p23lp27 perf]#
Committer testing:
While the above 'perf test -F 40' (or 'perf test bpf') will allow us
to see that the correct path is now added via -I, to actually test this
we better try to use a bpf script that includes files in the changed
directory.
We have the files that now reside in /root/lib/perf/examples/bpf/ to do
just that:
# tail -8 /root/lib/perf/examples/bpf/5sec.c
#include <bpf.h>
int probe(hrtimer_nanosleep, rqtp->tv_sec)(void *ctx, int err, long sec)
{
return sec == 5;
}
license(GPL);
# perf trace -e *sleep -e /root/lib/perf/examples/bpf/5sec.c sleep 4
0.333 (4000.086 ms): sleep/9248 nanosleep(rqtp: 0x7ffc155f3300) = 0
# perf trace -e *sleep -e /root/lib/perf/examples/bpf/5sec.c sleep 5
0.287 ( ): sleep/9659 nanosleep(rqtp: 0x7ffeafe38200) ...
0.290 ( ): perf_bpf_probe:hrtimer_nanosleep:(ffffffff9911efe0) tv_sec=5
0.287 (5000.059 ms): sleep/9659 ... [continued]: nanosleep()) = 0
# perf trace -e *sleep -e /root/lib/perf/examples/bpf/5sec.c sleep 6
0.247 (5999.951 ms): sleep/10068 nanosleep(rqtp: 0x7fff2086d900) = 0
# perf trace -e *sleep -e /root/lib/perf/examples/bpf/5sec.c sleep 5.987
0.293 ( ): sleep/10489 nanosleep(rqtp: 0x7ffdd4fc10e0) ...
0.296 ( ): perf_bpf_probe:hrtimer_nanosleep:(ffffffff9911efe0) tv_sec=5
0.293 (5986.912 ms): sleep/10489 ... [continued]: nanosleep()) = 0
#
Suggested-by: Stefan Liebler <stli@linux.ibm.com>
Suggested-by: Arnaldo Carvalho de Melo <acme@kernel.org>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Fixes: 1b16fffa38 ("perf llvm-utils: Add bpf include path to clang command line")
Link: http://lkml.kernel.org/r/20180731073254.91090-1-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
'perf c2c' scans read/write accesses and tries to find false sharing
cases, so when the events it wants were not asked for or ended up not
taking place, we get no histograms.
So do not try to display entry details if there's not any. Currently
this ends up in crash:
$ perf c2c report # then press 'd'
perf: Segmentation fault
$
Committer testing:
Before:
Record a perf.data file without events of interest to 'perf c2c report',
then call it and press 'd':
# perf record sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.001 MB perf.data (6 samples) ]
# perf c2c report
perf: Segmentation fault
-------- backtrace --------
perf[0x5b1d2a]
/lib64/libc.so.6(+0x346df)[0x7fcb566e36df]
perf[0x46fcae]
perf[0x4a9f1e]
perf[0x4aa220]
perf(main+0x301)[0x42c561]
/lib64/libc.so.6(__libc_start_main+0xe9)[0x7fcb566cff29]
perf(_start+0x29)[0x42c999]
#
After the patch the segfault doesn't take place, a follow up patch to
tell the user why nothing changes when 'd' is pressed would be good.
Reported-by: rodia@autistici.org
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: David Ahern <dsahern@gmail.com>
Cc: Don Zickus <dzickus@redhat.com>
Cc: Joe Mario <jmario@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Fixes: f1c5fd4d0b ("perf c2c report: Add TUI cacheline browser")
Link: http://lkml.kernel.org/r/20180724062008.26126-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Recently, the subtest numbering was changed to start from 1. While it
is fine for displaying results, this should not be the case when the
subtests are actually invoked.
Typically, the subtests are stored in zero-indexed arrays and invoked
based on the index passed to the main test function. Since the index
now starts from 1, the second subtest in the array (index 1) gets
invoked instead of the first (index 0). This applies to all of the
following subtests but for the last one, the subtest always fails
because it does not meet the boundary condition of the subtest index
being lesser than the number of subtests.
This can be observed on powerpc64 and x86_64 systems running Fedora 28
as shown below.
Before:
# perf test "builtin clang support"
55: builtin clang support :
55.1: builtin clang compile C source to IR : Ok
55.2: builtin clang compile C source to ELF object : FAILED!
# perf test "LLVM search and compile"
38: LLVM search and compile :
38.1: Basic BPF llvm compile : Ok
38.2: kbuild searching : Ok
38.3: Compile source for BPF prologue generation : Ok
38.4: Compile source for BPF relocation : FAILED!
# perf test "BPF filter"
40: BPF filter :
40.1: Basic BPF filtering : Ok
40.2: BPF pinning : Ok
40.3: BPF prologue generation : Ok
40.4: BPF relocation checker : FAILED!
After:
# perf test "builtin clang support"
55: builtin clang support :
55.1: builtin clang compile C source to IR : Ok
55.2: builtin clang compile C source to ELF object : Ok
# perf test "LLVM search and compile"
38: LLVM search and compile :
38.1: Basic BPF llvm compile : Ok
38.2: kbuild searching : Ok
38.3: Compile source for BPF prologue generation : Ok
38.4: Compile source for BPF relocation : Ok
# perf test "BPF filter"
40: BPF filter :
40.1: Basic BPF filtering : Ok
40.2: BPF pinning : Ok
40.3: BPF prologue generation : Ok
40.4: BPF relocation checker : Ok
Signed-off-by: Sandipan Das <sandipan@linux.ibm.com>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Hendrik Brueckner <brueckner@linux.ibm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Cc: Thomas Richter <tmricht@linux.ibm.com>
Fixes: 9ef0112442 ("perf test: Fix subtest number when showing results")
Link: http://lkml.kernel.org/r/20180726171733.33208-1-sandipan@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For instance:
$ trace -e socket* ssh sandy
0.000 ( 0.031 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
0.052 ( 0.015 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
1.568 ( 0.020 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
1.603 ( 0.012 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
1.699 ( 0.014 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
1.724 ( 0.012 ms): ssh/19919 socket(family: LOCAL, type: STREAM|CLOEXEC|NONBLOCK ) = 3
1.804 ( 0.020 ms): ssh/19919 socket(family: INET, type: STREAM, protocol: TCP ) = 3
17.549 ( 0.098 ms): ssh/19919 socket(family: LOCAL, type: STREAM ) = 4
acme@sandy's password:
Just like with other syscall args, the common bits are supressed so that
the output is more compact, i.e. we use "TCP" instead of "IPPROTO_TCP",
but we can make this show the original constant names if we like it by
using some command line knob or ~/.perfconfig "[trace]" section
variable.
Also needed is to make perf's event parser accept things like:
$ perf trace -e socket*/protocol=TCP/
By using both the tracefs event 'format' files and these tables built
from the kernel sources.
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: https://lkml.kernel.org/n/tip-l39jz1vnyda0b6jsufuc8bz7@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We may have string tables where not all slots have values, in those
cases its better to print the numeric value, for instance:
In the table below we would show "protocol: (null)" for
socket_ipproto[3]
Where it would be better to show "protocol: 3".
$ tools/perf/trace/beauty/socket_ipproto.sh
static const char *socket_ipproto[] = {
[0] = "IP",
[103] = "PIM",
[108] = "COMP",
[12] = "PUP",
[132] = "SCTP",
[136] = "UDPLITE",
[137] = "MPLS",
[17] = "UDP",
[1] = "ICMP",
[22] = "IDP",
[255] = "RAW",
[29] = "TP",
[2] = "IGMP",
[33] = "DCCP",
[41] = "IPV6",
[46] = "RSVP",
[47] = "GRE",
[4] = "IPIP",
[50] = "ESP",
[51] = "AH",
[6] = "TCP",
[8] = "EGP",
[92] = "MTP",
[94] = "BEETPH",
[98] = "ENCAP",
};
$
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: https://lkml.kernel.org/n/tip-7djfak94eb3b9ltr79cpn3ti@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It'll use tools/include copy of linux/in.h to generate a table to be
used by tools, initially by the 'socket' and 'socketpair' beautifiers in
'perf trace', but that could also be used to translate from a string
constant to the integer value to be used in a eBPF or tracefs tracepoint
filter.
When used without any args it produces:
$ tools/perf/trace/beauty/socket_ipproto.sh
static const char *socket_ipproto[] = {
[0] = "IP",
[103] = "PIM",
[108] = "COMP",
[12] = "PUP",
[132] = "SCTP",
[136] = "UDPLITE",
[137] = "MPLS",
[17] = "UDP",
[1] = "ICMP",
[22] = "IDP",
[255] = "RAW",
[29] = "TP",
[2] = "IGMP",
[33] = "DCCP",
[41] = "IPV6",
[46] = "RSVP",
[47] = "GRE",
[4] = "IPIP",
[50] = "ESP",
[51] = "AH",
[6] = "TCP",
[8] = "EGP",
[92] = "MTP",
[94] = "BEETPH",
[98] = "ENCAP",
};
$
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: https://lkml.kernel.org/n/tip-v9rafqh3qn6b9kp9vfvj9f8s@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The 'umask' event parameter is unsupported on some architectures like
powerpc64.
This can be observed on a powerpc64le system running Fedora 27 as shown
below.
# perf test "Parse event definition strings" -v
6: Parse event definition strings :
--- start ---
test child forked, pid 45915
...
running test 3 'cpu/name='COMPLEX_CYCLES_NAME:orig=cycles,desc=chip-clock-ticks',period=0x1,event=0x2,umask=0x3/ukp'Invalid event/parameter 'umask'
Invalid event/parameter 'umask'
failed to parse event 'cpu/name='COMPLEX_CYCLES_NAME:orig=cycles,desc=chip-clock-ticks',period=0x1,event=0x2,umask=0x3/ukp', err 1, str 'unknown term'
event syntax error: '..,event=0x2,umask=0x3/ukp'
\___ unknown term
valid terms: event,mark,pmc,cache_sel,pmcxsel,unit,thresh_stop,thresh_start,combine,thresh_sel,thresh_cmp,sample_mode,config,config1,config2,name,period,freq,branch_type,time,call-graph,stack-size,no-inherit,inherit,max-stack,no-overwrite,overwrite,driver-config
mem_access -> cpu/event=0x10401e0/
running test 0 'config=10,config1,config2=3,umask=1'
test child finished with 1
---- end ----
Parse event definition strings: FAILED!
Committer testing:
After applying the patch these test passes and in verbose mode we get:
# perf test -v "event definition"
6: Parse event definition strings:
--- start ---
test child forked, pid 11061
running test 0 'syscalls:sys_enter_openat'Using CPUID GenuineIntel-6-9E
<SNIP>
running test 53 'cycles/name='COMPLEX_CYCLES_NAME:orig=cycles,desc=chip-clock-ticks'/Duk'
running test 0 'cpu/config=10,config1,config2=3,period=1000/u'
running test 1 'cpu/config=1,name=krava/u,cpu/config=2/u'
running test 2 'cpu/config=1,call-graph=fp,time,period=100000/,cpu/config=2,call-graph=no,time=0,period=2000/'
running test 3 'cpu/name='COMPLEX_CYCLES_NAME:orig=cycles,desc=chip-clock-ticks',period=0x1,event=0x2/ukp'
<SNIP>
test child finished with 0
---- end ----
Parse event definition strings: Ok
#
Suggested-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Signed-off-by: Sandipan Das <sandipan@linux.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Fixes: 06dc5bf21f ("perf tests: Check that complex event name is parsed correctly")
Link: http://lkml.kernel.org/r/20180726105502.31670-1-sandipan@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
'perf record' will error out if both --delay and LBR are applied.
For example:
# perf record -D 1000 -a -e cycles -j any -- sleep 2
Error:
dummy:HG: PMU Hardware doesn't support sampling/overflow-interrupts.
Try 'perf stat'
#
A dummy event is added implicitly for initial delay, which has the same
configurations as real sampling events. The dummy event is a software
event. If LBR is configured, perf must error out.
The dummy event will only be used to track PERF_RECORD_MMAP while perf
waits for the initial delay to enable the real events. The BRANCH_STACK
bit can be safely cleared for the dummy event.
After applying the patch:
# perf record -D 1000 -a -e cycles -j any -- sleep 2
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.054 MB perf.data (828 samples) ]
#
Reported-by: Sunil K Pandey <sunil.k.pandey@intel.com>
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1531145722-16404-1-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The UAPI file byteorder/little_endian.h uses the __always_inline define
without including the header where it is defined, linux/stddef.h, this
ends up working in all the other distros because that file gets included
seemingly by luck from one of the files included from little_endian.h.
But not on Alpine:edge, that fails for all files where perf_event.h is
included but linux/stddef.h isn't include before that.
Adding the missing linux/stddef.h file where it breaks on Alpine:edge to
fix that, in all other distros, that is just a very small header anyway.
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: https://lkml.kernel.org/n/tip-9r1pifftxvuxms8l7ir73p5l@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To cope with the changes in:
12c89130a5 ("x86/asm/memcpy_mcsafe: Add write-protection-fault handling")
60622d6822 ("x86/asm/memcpy_mcsafe: Return bytes remaining")
bd131544aa ("x86/asm/memcpy_mcsafe: Add labels for __memcpy_mcsafe() write fault handling")
da7bc9c57e ("x86/asm/memcpy_mcsafe: Remove loop unrolling")
This needed introducing a file with a copy of the mcsafe_handle_tail()
function, that is used in the new memcpy_64.S file, as well as a dummy
mcsafe_test.h header.
Testing it:
$ nm ~/bin/perf | grep mcsafe
0000000000484130 T mcsafe_handle_tail
0000000000484300 T __memcpy_mcsafe
$
$ perf bench mem memcpy
# Running 'mem/memcpy' benchmark:
# function 'default' (Default memcpy() provided by glibc)
# Copying 1MB bytes ...
44.389205 GB/sec
# function 'x86-64-unrolled' (unrolled memcpy() in arch/x86/lib/memcpy_64.S)
# Copying 1MB bytes ...
22.710756 GB/sec
# function 'x86-64-movsq' (movsq-based memcpy() in arch/x86/lib/memcpy_64.S)
# Copying 1MB bytes ...
42.459239 GB/sec
# function 'x86-64-movsb' (movsb-based memcpy() in arch/x86/lib/memcpy_64.S)
# Copying 1MB bytes ...
42.459239 GB/sec
$
This silences this perf tools build warning:
Warning: Kernel ABI header at 'tools/arch/x86/lib/memcpy_64.S' differs from latest version at 'arch/x86/lib/memcpy_64.S'
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Mika Penttilä <mika.penttila@nextfour.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Tony Luck <tony.luck@intel.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-igdpciheradk3gb3qqal52d0@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Perf test 40 for example has several subtests numbered 1-4 when
displaying the start of the subtest. When the subtest results
are displayed the subtests are numbered 0-3.
Use this command to generate trace output:
[root@s35lp76 perf]# ./perf test -Fv 40 2>/tmp/bpf1
Fix this by adjusting the subtest number when show the
subtest result.
Output before:
[root@s35lp76 perf]# egrep '(^40\.[0-4]| subtest [0-4]:)' /tmp/bpf1
40.1: Basic BPF filtering :
BPF filter subtest 0: Ok
40.2: BPF pinning :
BPF filter subtest 1: Ok
40.3: BPF prologue generation :
BPF filter subtest 2: Ok
40.4: BPF relocation checker :
BPF filter subtest 3: Ok
[root@s35lp76 perf]#
Output after:
root@s35lp76 ~]# egrep '(^40\.[0-4]| subtest [0-4]:)' /tmp/bpf1
40.1: Basic BPF filtering :
BPF filter subtest 1: Ok
40.2: BPF pinning :
BPF filter subtest 2: Ok
40.3: BPF prologue generation :
BPF filter subtest 3: Ok
40.4: BPF relocation checker :
BPF filter subtest 4: Ok
[root@s35lp76 ~]#
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Link: http://lkml.kernel.org/r/20180724134858.100644-1-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
There's no reason to have separate function to display clock events.
It's only purpose was to convert the nanosecond value into microseconds.
We do that now in generic code, if the unit and scale values are
properly set, which this patch do for clock events.
The output differs in the unit field being displayed in its columns
rather than having it added as a suffix of the event name. Plus the
value is rounded into 2 decimal numbers as for any other event.
Before:
# perf stat -e cpu-clock,task-clock -C 0 sleep 3
Performance counter stats for 'CPU(s) 0':
3001.123137 cpu-clock (msec) # 1.000 CPUs utilized
3001.133250 task-clock (msec) # 1.000 CPUs utilized
3.001159813 seconds time elapsed
Now:
# perf stat -e cpu-clock,task-clock -C 0 sleep 3
Performance counter stats for 'CPU(s) 0':
3,001.05 msec cpu-clock # 1.000 CPUs utilized
3,001.05 msec task-clock # 1.000 CPUs utilized
3.001077794 seconds time elapsed
There's a small difference in csv output, as we now output the unit
field, which was empty before. It's in the proper spot, so there's no
compatibility issue.
Before:
# perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3
3001.065177,,cpu-clock,3001064187,100.00,1.000,CPUs utilized
3001.077085,,task-clock,3001077085,100.00,1.000,CPUs utilized
# perf stat -e cpu-clock,task-clock -C 0 -x, sleep 3
3000.80,msec,cpu-clock,3000799026,100.00,1.000,CPUs utilized
3000.80,msec,task-clock,3000799550,100.00,1.000,CPUs utilized
Add perf_evsel__is_clock to replace nsec_counter.
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/20180720110036.32251-2-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We occasionaly hit following assert failure in 'perf top', when processing the
/proc info in multiple threads.
perf: ...include/linux/refcount.h:109: refcount_inc:
Assertion `!(!refcount_inc_not_zero(r))' failed.
The gdb backtrace looks like this:
[Switching to Thread 0x7ffff11ba700 (LWP 13749)]
0x00007ffff50839fb in raise () from /lib64/libc.so.6
(gdb)
#0 0x00007ffff50839fb in raise () from /lib64/libc.so.6
#1 0x00007ffff5085800 in abort () from /lib64/libc.so.6
#2 0x00007ffff507c0da in __assert_fail_base () from /lib64/libc.so.6
#3 0x00007ffff507c152 in __assert_fail () from /lib64/libc.so.6
#4 0x0000000000535373 in refcount_inc (r=0x7fffdc009be0)
at ...include/linux/refcount.h:109
#5 0x00000000005354f1 in comm_str__get (cs=0x7fffdc009bc0)
at util/comm.c:24
#6 0x00000000005356bd in __comm_str__findnew (str=0x7fffd000b260 ":2",
root=0xbed5c0 <comm_str_root>) at util/comm.c:72
#7 0x000000000053579e in comm_str__findnew (str=0x7fffd000b260 ":2",
root=0xbed5c0 <comm_str_root>) at util/comm.c:95
#8 0x000000000053582e in comm__new (str=0x7fffd000b260 ":2",
timestamp=0, exec=false) at util/comm.c:111
#9 0x00000000005363bc in thread__new (pid=2, tid=2) at util/thread.c:57
#10 0x0000000000523da0 in ____machine__findnew_thread (machine=0xbfde38,
threads=0xbfdf28, pid=2, tid=2, create=true) at util/machine.c:457
#11 0x0000000000523eb4 in __machine__findnew_thread (machine=0xbfde38,
...
The failing assertion is this one:
REFCOUNT_WARN(!refcount_inc_not_zero(r), ...
The problem is that we keep global comm_str_root list, which
is accessed by multiple threads during the 'perf top' startup
and following 2 paths can race:
thread 1:
...
thread__new
comm__new
comm_str__findnew
down_write(&comm_str_lock);
__comm_str__findnew
comm_str__get
thread 2:
...
comm__override or comm__free
comm_str__put
refcount_dec_and_test
down_write(&comm_str_lock);
rb_erase(&cs->rb_node, &comm_str_root);
Because thread 2 first decrements the refcnt and only after then it removes the
struct comm_str from the list, the thread 1 can find this object on the list
with refcnt equls to 0 and hit the assert.
This patch fixes the thread 1 __comm_str__findnew path, by ignoring objects
that already dropped the refcnt to 0. For the rest of the objects we take the
refcnt before comparing its name and release it afterwards with comm_str__put,
which can also release the object completely.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Lukasz Odzioba <lukasz.odzioba@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: kernel-team@lge.com
Link: http://lkml.kernel.org/r/20180720101740.GA27176@krava
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
There's an issue with using threads::last_match in multithread mode
which is enabled during the perf top synthesize. It might crash with
following assertion:
perf: ...include/linux/refcount.h:109: refcount_inc:
Assertion `!(!refcount_inc_not_zero(r))' failed.
The gdb backtrace looks like this:
0x00007ffff50839fb in raise () from /lib64/libc.so.6
(gdb)
#0 0x00007ffff50839fb in raise () from /lib64/libc.so.6
#1 0x00007ffff5085800 in abort () from /lib64/libc.so.6
#2 0x00007ffff507c0da in __assert_fail_base () from /lib64/libc.so.6
#3 0x00007ffff507c152 in __assert_fail () from /lib64/libc.so.6
#4 0x0000000000535ff9 in refcount_inc (r=0x7fffe8009a70)
at ...include/linux/refcount.h:109
#5 0x0000000000536771 in thread__get (thread=0x7fffe8009a40)
at util/thread.c:115
#6 0x0000000000523cd0 in ____machine__findnew_thread (machine=0xbfde38,
threads=0xbfdf28, pid=2, tid=2, create=true) at util/machine.c:432
#7 0x0000000000523eb4 in __machine__findnew_thread (machine=0xbfde38,
pid=2, tid=2) at util/machine.c:489
#8 0x0000000000523f24 in machine__findnew_thread (machine=0xbfde38,
pid=2, tid=2) at util/machine.c:499
#9 0x0000000000526fbe in machine__process_fork_event (machine=0xbfde38,
...
The failing assertion is this one:
REFCOUNT_WARN(!refcount_inc_not_zero(r), ...
the problem is that we don't serialize access to threads::last_match.
We serialize the access to the threads tree, but we don't care how's
threads::last_match being accessed. Both locked/unlocked paths use
that data and can set it. In multithreaded mode we can end up with
invalid object in thread__get call, like in following paths race:
thread 1
...
machine__findnew_thread
down_write(&threads->lock);
__machine__findnew_thread
____machine__findnew_thread
th = threads->last_match;
if (th->tid == tid) {
thread__get
thread 2
...
machine__find_thread
down_read(&threads->lock);
__machine__findnew_thread
____machine__findnew_thread
th = threads->last_match;
if (th->tid == tid) {
thread__get
thread 3
...
machine__process_fork_event
machine__remove_thread
__machine__remove_thread
threads->last_match = NULL
thread__put
thread__put
Thread 1 and 2 might got stale last_match, before thread 3 clears
it. Thread 1 and 2 then race with thread 3's thread__put and they
might trigger the refcnt == 0 assertion above.
The patch is disabling the last_match cache for multiple thread
mode. It was originally meant for single thread scenarios, where
it's common to have multiple sequential searches of the same
thread.
In multithread mode this does not make sense, because top's threads
processes different /proc entries and so the 'struct threads' object
is queried for various threads. Moreover we'd need to add more locks
to make it work.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Lukasz Odzioba <lukasz.odzioba@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/r/20180719143345.12963-4-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Stephan reported, that pipe mode does not carry the group information
and thus the piped report won't display the grouped output for following
command:
# perf record -e '{cycles,instructions,branches}' -a sleep 4 | perf report
It has no idea about the group setup, so it will display events
separately:
# Overhead Command Shared Object ...
# ........ ............... .......................
#
6.71% swapper [kernel.kallsyms]
2.28% offlineimap libpython2.7.so.1.0
0.78% perf [kernel.kallsyms]
...
Fix GROUP_DESC feature record to be synthesized in pipe mode, so the
report output is grouped if there are groups defined in record:
# Overhead Command Shared ...
# ........................ ............... .......
#
7.57% 0.16% 0.30% swapper [kernel
1.87% 3.15% 2.46% offlineimap libpyth
1.33% 0.00% 0.00% perf [kernel
...
Reported-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Stephane Eranian <eranian@google.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: David Carrillo-Cisneros <davidcc@google.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20180712135202.14774-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When perf/data is recorded with the dwarf call-graph option, the
callchain shown by 'perf script' still shows the binary offsets of the
userspace symbols instead of their virtual addresses. Since the symbol
offset calculation is based on using virtual address as the ip, we see
incorrect offsets as well.
The use of virtual addresses affects the ability to find out the
line number in the corresponding source file to which an address
maps to as described in commit 6754075915 ("perf unwind: Use
addr_location::addr instead of ip for entries").
This has also been addressed by temporarily converting the virtual
address to the correponding binary offset so that it can be mapped
to the source line number correctly.
This is a follow-up for commit 1961018469 ("perf script: Show
virtual addresses instead of offsets").
This can be verified on a powerpc64le system running Fedora 27 as
shown below:
# perf probe -x /usr/lib64/libc-2.26.so -a inet_pton
# perf record -e probe_libc:inet_pton --call-graph=dwarf ping -6 -c 1 ::1
Before:
# perf report --stdio --no-children -s sym,srcline -g address
# Samples: 1 of event 'probe_libc:inet_pton'
# Event count (approx.): 1
#
# Overhead Symbol Source:Line
# ........ .................... ...........
#
100.00% [.] __GI___inet_pton inet_pton.c
|
---gaih_inet getaddrinfo.c:537 (inlined)
__GI_getaddrinfo getaddrinfo.c:2304 (inlined)
main ping.c:519
generic_start_main libc-start.c:308 (inlined)
__libc_start_main libc-start.c:102
...
# perf script -F comm,ip,sym,symoff,srcline,dso
ping
15af28 __GI___inet_pton+0xffff000099160008 (/usr/lib64/libc-2.26.so)
libc-2.26.so[ffff80004ca0af28]
10fa53 gaih_inet+0xffff000099160f43
libc-2.26.so[ffff80004c9bfa53] (inlined)
1105b3 __GI_getaddrinfo+0xffff000099160163
libc-2.26.so[ffff80004c9c05b3] (inlined)
2d6f main+0xfffffffd9f1003df (/usr/bin/ping)
ping[fffffffecf882d6f]
2369f generic_start_main+0xffff00009916013f
libc-2.26.so[ffff80004c8d369f] (inlined)
23897 __libc_start_main+0xffff0000991600b7 (/usr/lib64/libc-2.26.so)
libc-2.26.so[ffff80004c8d3897]
After:
# perf report --stdio --no-children -s sym,srcline -g address
# Samples: 1 of event 'probe_libc:inet_pton'
# Event count (approx.): 1
#
# Overhead Symbol Source:Line
# ........ .................... ...........
#
100.00% [.] __GI___inet_pton inet_pton.c
|
---gaih_inet.constprop.7 getaddrinfo.c:537
getaddrinfo getaddrinfo.c:2304
main ping.c:519
generic_start_main.isra.0 libc-start.c:308
__libc_start_main libc-start.c:102
...
# perf script -F comm,ip,sym,symoff,srcline,dso
ping
7fffb38aaf28 __GI___inet_pton+0x8 (/usr/lib64/libc-2.26.so)
inet_pton.c:68
7fffb385fa53 gaih_inet.constprop.7+0xf43 (/usr/lib64/libc-2.26.so)
getaddrinfo.c:537
7fffb38605b3 getaddrinfo+0x163 (/usr/lib64/libc-2.26.so)
getaddrinfo.c:2304
130782d6f main+0x3df (/usr/bin/ping)
ping.c:519
7fffb377369f generic_start_main.isra.0+0x13f (/usr/lib64/libc-2.26.so)
libc-start.c:308
7fffb3773897 __libc_start_main+0xb7 (/usr/lib64/libc-2.26.so)
libc-start.c:102
Signed-off-by: Sandipan Das <sandipan@linux.ibm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Cc: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Fixes: 6754075915 ("perf unwind: Use addr_location::addr instead of ip for entries")
Link: http://lkml.kernel.org/r/20180703120555.32971-1-sandipan@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This should speed up accessing new system calls introduced with the
kernel rather than waiting for libaudit updates to include them.
Using the existing other arch scripts resulted in this error:
tools/perf/arch/arm64/entry/syscalls//mksyscalltbl: 25: printf: __NR3264_ftruncate: expected numeric value
because, unlike other arches, asm-generic's unistd.h does things like:
#define __NR_ftruncate __NR3264_ftruncate
Turning the scripts printf's %d into a %s resulted in this in the
generated syscalls.c file:
static const char *syscalltbl_arm64[] = {
[__NR3264_ftruncate] = "ftruncate",
So we use the host C compiler to fold the macros, and print them out
from within a temporary C program, in order to get the correct output:
static const char *syscalltbl_arm64[] = {
[46] = "ftruncate",
Committer notes:
Testing this with a container with an old toolchain breaks because it
ends up using the system's /usr/include/asm-generic/unistd.h, included
from tools/arch/arm64/include/uapi/asm/unistd.h when what is desired is
for it to include tools/include/uapi/asm-generic/unistd.h.
Since all that tools/arch/arm64/include/uapi/asm/unistd.h is to set a
define and then include asm-generic/unistd.h, do that directly and use
tools/include/uapi/asm-generic/unistd.h as the file to get the syscall
definitions to expand.
Testing it:
tools/perf/arch/arm64/entry/syscalls/mksyscalltbl /gcc-linaro-5.4.1-2017.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu-gcc gcc tools/include/uapi/asm-generic/unistd.h
Now works and generates in the syscall string table.
Before it ended up as:
$ tools/perf/arch/arm64/entry/syscalls/mksyscalltbl /gcc-linaro-5.4.1-2017.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu-gcc gcc tools/arch/arm64/include/uapi/asm/unistd.h
static const char *syscalltbl_arm64[] = {
<stdin>: In function 'main':
<stdin>:257:38: error: '__NR_getrandom' undeclared (first use in this function)
<stdin>:257:38: note: each undeclared identifier is reported only once for each function it appears in
<stdin>:258:41: error: '__NR_memfd_create' undeclared (first use in this function)
<stdin>:259:32: error: '__NR_bpf' undeclared (first use in this function)
<stdin>:260:37: error: '__NR_execveat' undeclared (first use in this function)
tools/perf/arch/arm64/entry/syscalls/mksyscalltbl: 47: tools/perf/arch/arm64/entry/syscalls/mksyscalltbl: /tmp/create-table-60liya: Permission denied
};
$
Signed-off-by: Kim Phillips <kim.phillips@arm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Cc: Thomas Richter <tmricht@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/20180706163443.22626f5e9e10e5bab5e5c662@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If the event 'probe_libc:inet_pton' already exists, this test fails and
deletes the existing event before exiting. This will then pass for any
subsequent executions.
Instead of skipping to deleting the existing event because of failing to
add a new event, a duplicate event is now created and the script
continues with the usual checks. Only the new duplicate event that is
created at the beginning of the test is deleted as a part of the
cleanups in the end. All existing events remain as it is.
This can be observed on a powerpc64 system running Fedora 27 as shown
below.
# perf probe -x /usr/lib64/power8/libc-2.26.so -a inet_pton
Added new event:
probe_libc:inet_pton (on inet_pton in /usr/lib64/power8/libc-2.26.so)
Before:
# perf test -v "probe libc's inet_pton & backtrace it with ping"
62: probe libc's inet_pton & backtrace it with ping :
--- start ---
test child forked, pid 21302
test child finished with -1
---- end ----
probe libc's inet_pton & backtrace it with ping: FAILED!
# perf probe --list
After:
# perf test -v "probe libc's inet_pton & backtrace it with ping"
62: probe libc's inet_pton & backtrace it with ping :
--- start ---
test child forked, pid 21490
ping 21513 [035] 39357.565561: probe_libc:inet_pton_1: (7fffa4c623b0)
7fffa4c623b0 __GI___inet_pton+0x0 (/usr/lib64/power8/libc-2.26.so)
7fffa4c190dc gaih_inet.constprop.7+0xf4c (/usr/lib64/power8/libc-2.26.so)
7fffa4c19c4c getaddrinfo+0x15c (/usr/lib64/power8/libc-2.26.so)
111d93c20 main+0x3e0 (/usr/bin/ping)
test child finished with 0
---- end ----
probe libc's inet_pton & backtrace it with ping: Ok
# perf probe --list
probe_libc:inet_pton (on __inet_pton@resolv/inet_pton.c in /usr/lib64/power8/libc-2.26.so)
Signed-off-by: Sandipan Das <sandipan@linux.ibm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kim Phillips <kim.phillips@arm.com>
Cc: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Cc: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/e11fecff96e6cf4c65cdbd9012463513d7b8356c.1530724939.git.sandipan@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
