perf creates a single 'struct thread' to represent the idle task. That
is because threads are identified by PID and TID, and the idle task
always has PID == TID == 0.
However, there are actually separate idle tasks for each CPU. That
creates a problem for thread stack processing which assumes that each
thread has a single stack, not one stack per CPU.
Fix that by passing through the CPU number, and in the case of the idle
"thread", pick the thread stack from an array based on the CPU number.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181221120620.9659-8-adrian.hunter@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In preparation for fixing thread stack processing for the idle task,
avoid direct reference to the thread's stack. The thread stack will
change to an array of thread stacks, at which point the meaning of the
direct reference will change.
Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/20181221120620.9659-4-adrian.hunter@intel.com
[ Rename thread_stack__ts() to thread__stack() since this operates on a 'thread' struct ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Joe suggested to have the coalesce default set just to 'iaddr', because
it's easier to read on the default 'perf c2c report' output.
By removing the "pid" field from the default -c/--coalesce option, the
'perf c2c' report will group all the relevant PIDs under the instruction
address ('iaddr') bucket. User can always run "-c pid,iaddr" for a more
fine grained output on particular PIDs.
Suggested-by: Joe Mario <jmario@redhat.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20181228101820.28010-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This is a fix for another instance of the skid problem Milian recently
found [1]
The LBRs don't freeze at the exact same time as the PMI is triggered.
The perf script brstackinsn code that dumps LBR assembler assumes that
the last branch in the LBR leads to the sample point. But with skid
it's possible that the CPU executes one or more branches before the
sample, but which do not appear in the LBR.
What happens then is either that the sample point is before the last LBR
branch. In this case the dumper sees a negative length and ignores it.
Or it the sample point is long after the last branch. Then the dumper
sees a very long block and dumps it upto its block limit (16k bytes),
which is noise in the output.
On typical sample session this can happen regularly.
This patch tries to detect and handle the situation. On the last block
that is dumped by the LBR dumper we always stop on the first branch. If
the block length is negative just scan forward to the first branch.
Otherwise scan until a branch is found.
The PT decoder already has a function that uses the instruction decoder
to detect branches, so we can just reuse it here.
Then when a terminating branch is found print an indication and stop
dumping. This might miss a few instructions, but at least shows no
runaway blocks.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Link: http://lkml.kernel.org/r/20181120050617.4119-1-andi@firstfloor.org
[ Resolved conflict with dd2e18e9ac ("perf tools: Support 'srccode' output") ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Ondřej reported that when compiled with python3, the python extension
regresses in evlist.get_pollfd function behaviour.
The evlist.get_pollfd function creates file objects from evlist's fds
and returns them in a list. The python3 version also sets them to 'close
the original descriptor' when the object dies (is closed), by passing
True via the 'closefd' arg in the PyFile_FromFd call.
The python's closefd doc says:
If closefd is False, the underlying file descriptor will be kept open
when the file is closed.
That's why the following line in python3 closes all evlist fds:
evlist.get_pollfd()
the returned list is immediately destroyed and that takes down the
original events fds.
Passing closefd as False to PyFile_FromFd to fix this.
Reported-by: Ondřej Lysoněk <olysonek@redhat.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Jaroslav Škarvada <jskarvad@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Fixes: 66dfdff03d ("perf tools: Add Python 3 support")
Link: http://lkml.kernel.org/r/20181226112121.5285-1-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We shouldn't hardcode the size of the tracepoint common_ fields, use the
offset of the 'id'/'__syscallnr' field in the sys_enter event instead.
This caused the augmented syscalls code to fail on a particular build of a
PREEMPT_RT_FULL kernel where these extra 'common_migrate_disable' and
'common_padding' fields were before the syscall id one:
# cat /sys/kernel/debug/tracing/events/raw_syscalls/sys_enter/format
name: sys_enter
ID: 22
format:
field:unsigned short common_type; offset:0; size:2; signed:0;
field:unsigned char common_flags; offset:2; size:1; signed:0;
field:unsigned char common_preempt_count; offset:3; size:1; signed:0;
field:int common_pid; offset:4; size:4; signed:1;
field:unsigned short common_migrate_disable; offset:8; size:2; signed:0;
field:unsigned short common_padding; offset:10; size:2; signed:0;
field:long id; offset:16; size:8; signed:1;
field:unsigned long args[6]; offset:24; size:48; signed:0;
print fmt: "NR %ld (%lx, %lx, %lx, %lx, %lx, %lx)", REC->id, REC->args[0], REC->args[1], REC->args[2], REC->args[3], REC->args[4], REC->args[5]
#
All those 'common_' prefixed fields are zeroed when they hit a BPF tracepoint
hook, we better just discard those, i.e. somehow pass an offset to the
BPF program from the start of the ctx and make adjustments in the 'perf trace'
handlers to adjust the offset of the syscall arg offsets obtained from tracefs.
Till then, fix it the quick way and add this to the augmented_raw_syscalls.c to
bet it to work in such kernels:
diff --git a/tools/perf/examples/bpf/augmented_raw_syscalls.c b/tools/perf/examples/bpf/augmented_raw_syscalls.c
index 53c233370fae..1f746f931e13 100644
--- a/tools/perf/examples/bpf/augmented_raw_syscalls.c
+++ b/tools/perf/examples/bpf/augmented_raw_syscalls.c
@@ -38,12 +38,14 @@ struct bpf_map SEC("maps") syscalls = {
struct syscall_enter_args {
unsigned long long common_tp_fields;
+ long rt_common_tp_fields;
long syscall_nr;
unsigned long args[6];
};
struct syscall_exit_args {
unsigned long long common_tp_fields;
+ long rt_common_tp_fields;
long syscall_nr;
long ret;
};
Just to check that this was the case. Fix it properly later, for now remove the
hardcoding of the offset in the 'perf trace' side and document the situation
with this patch.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Luis Cláudio Gonçalves <lclaudio@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-2pqavrktqkliu5b9nzouio21@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Current libbfd feature test unconditionally links against -liberty and -lz.
While it's required on some systems (e.g. opensuse), it's completely
unnecessary on the others, where only -lbdf is sufficient (debian).
This patch streamlines (and renames) the following feature checks:
feature-libbfd - only link against -lbfd (debian),
see commit 2cf9040714 ("perf tools: Fix bfd
dependency libraries detection")
feature-libbfd-liberty - link against -lbfd and -liberty
feature-libbfd-liberty-z - link against -lbfd, -liberty and -lz (opensuse),
see commit 280e7c48c3 ("perf tools: fix BFD
detection on opensuse")
(feature-liberty{,-z} were renamed to feature-libbfd-liberty{,z}
for clarity)
The main motivation is to fix this feature test for bpftool which is
currently broken on debian (libbfd feature shows OFF, but we still
unconditionally link against -lbfd and it works).
Tested on debian with only -lbfd installed (without -liberty); I'd
appreciate if somebody on the other systems can test this new detection
method.
Signed-off-by: Stanislav Fomichev <sdf@google.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/4dfc634cfcfb236883971b5107cf3c28ec8a31be.1542328222.git.sdf@google.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
While updating 'perf trace' on an machine with an old precompiled
augmented_raw_syscalls.o that didn't setup the syscall map the new 'perf
trace' codebase notices the augmented_raw_syscalls.o eBPF event, decides
to use it instead of the old raw_syscalls:sys_{enter,exit} method, but
then because we don't have the syscall map tries to set the tracepoint
filter on the sys_{enter,exit} evsels, that are NULL, segfaulting.
Make the code more robust by checking it those tracepoints have
their respective evsels in place before trying to set the tp filter.
With this we still get everything to work, just not setting up the
syscall filters, which is better than a segfault. Now to update the
precompiled augmented_raw_syscalls.o and continue development :-)
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Luis Cláudio Gonçalves <lclaudio@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-3ft5rjdl05wgz2pwpx2z8btu@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
- Implement BPF based syscall filtering in 'perf trace', using BPF maps and
the augmented_raw_syscalls.c BPF proggie (Arnaldo Carvalho de Melo)
- Allow specifying in .perfconfig a set of events use in 'perf trace' in
addition to any other specified from the command line. This initially
will be used to always use the augmented_raw_syscalls.o precompiled
BPF program for getting pointer contents. (Arnaldo Carvalho de Melo)
- Allow fine grained control about how the syscall output should be
formatted. This will be used to allow producing the same output produced
by the 'strace' tool, to then use in regression tests comparing the
output of 'perf trace' with the one produced from 'strace' (Arnaldo Carvalho de Melo)
- Beautify the renameat2 olddirfd, newdirfd and flags arguments (Arnaldo Carvalho de Melo)
- Beautify arch_prctl 'code' syscall arg (Arnaldo Carvalho de Melo)
- Beautify fadvise64 'advice' syscall arg (Arnaldo Carvalho de Melo)
- Relax checks on perf-PID.map ownership, resulting in symbols in
executable anonymous maps setup by JITs in things like node.js to
be resolved in a 'perf top' session run by root without the need
for --force to be used (Arnaldo Carvalho de Melo)
- Update asm-generic/unistd.h copy (Arnaldo Carvalho de Melo)
- Do not use the first and last symbols when setting up address filters in
auxtrace, this fails when we don't have a symbol table, filter the entire
area based on the dso size. (Adrian Hunter)
- Do not use kernel headers to build libsubcmd, we shouldn't use
anything from outside tools/, fixes the build with the Android NDK (Arnaldo Carvalho de Melo)
- Add several prototypes for systems lacking those, such as open_memstream(),
sigqueue(), fixing warnings building with Android's bionic libc that were
preventing the use of -Werror there (Arnaldo Carvalho de Melo)
- Use LDFLAGS in the libtraceevent build commands, allowing developers
to override its values (Jiri Olsa)
- Link libperf-jvmti.so with LDFLAGS variable, allowing distro
packages to propagate its settings when building this library (Jiri Olsa)
- cs-etm (ARM CoreSight) fixes: (Leo Yan)
- Correct packets swapping in cs_etm__flush()
- Avoid stale branch samples when flush packet
- Remove unused 'trace_on' in cs_etm_decoder
- Refactor enumeration cs_etm_sample_type
- Rename CS_ETM_TRACE_ON to CS_ETM_DISCONTINUITY
- Treat NO_SYNC element as trace discontinuity
- Treat EO_TRACE element as trace discontinuity
- Generate branch sample for exception packet
- Use shebangs in the 'perf test' shell scripts, making them identifiable as
shell scripts (Michael Petlan)
- Avoid segfaults caused by negated options in 'perf stat' (Michael Petlan)
- Fix processing of dereferenced args in bprintk events in libtracevent (Steven Rostedt)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This actually so far, AFAIK is available only in x86, so the code was
put in place with x86 prefixes, in arches where it is not available it
will just not be called, so no further mechanisms are needed at this
time.
Later, when other arches wire this up, we'll just look at the uname
(live sessions) or perf_env data in the perf.data header to auto-wire
the right beautifier.
With this the output is the same as produced by 'strace' when used with
the following ~/.perfconfig:
# cat ~/.perfconfig
[llvm]
dump-obj = true
[trace]
add_events = /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o
show_zeros = yes
show_duration = no
no_inherit = yes
show_timestamp = no
show_arg_names = no
args_alignment = -40
show_prefix = yes
#
And, on fedora 29, since the string tables are generated from the kernel
sources, we don't know about 0x3001, just like strace:
--- /tmp/strace 2018-12-17 11:22:08.707586721 -0300
+++ /tmp/trace 2018-12-18 11:11:32.037512729 -0300
@@ -1,49 +1,49 @@
-arch_prctl(0x3001 /* ARCH_??? */, 0x7ffc8a92dc80) = -1 EINVAL (Invalid argument)
+arch_prctl(0x3001 /* ARCH_??? */, 0x7ffe4eb93ae0) = -1 EINVAL (Invalid argument)
-arch_prctl(ARCH_SET_FS, 0x7faf6700f540) = 0
+arch_prctl(ARCH_SET_FS, 0x7fb507364540) = 0
And that seems to be related to the CET/Shadow Stack feature, that
userland in Fedora 29 (glibc 2.28) are querying the kernel about, that
0x3001 seems to be ARCH_CET_STATUS, I'll check the situation and test
with a fedora 29 kernel to see if the other codes are used.
A diff that ignores the different pointers for different runs needs to
be put in place in the upcoming regression tests comparing 'perf trace's
output to strace's.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Luis Cláudio Gonçalves <lclaudio@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-73a9prs8ktkrt97trtdmdjs8@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This will all come from userspace, but to test the changes to make 'perf
trace' output similar to strace's, do this one more now manually.
To update the precompiled augmented_raw_syscalls.o binary I just run:
# perf record -e ~acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.c sleep 1
LLVM: dumping /home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.022 MB perf.data ]
#
Because to have augmented_raw_syscalls to be always used and a fast
startup and remove the need to have the llvm toolchain installed, I'm
using:
# perf config | grep add_events
trace.add_events=/home/acme/git/perf/tools/perf/examples/bpf/augmented_raw_syscalls.o
#
So when doing changes to augmented_raw_syscals.c one needs to rebuild
the .o file.
This will be done automagically later, i.e. have a 'make' behaviour of
recompiling when the .c gets changed.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Luis Cláudio Gonçalves <lclaudio@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: https://lkml.kernel.org/n/tip-lw3i2atyq8549fpqwmszn3qp@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
