Add support for the new LBRv5 format used on Intel Skylake CPUs.
The flags for mispredict, abort, in_tx etc. moved to range of separate
LBR_INFO_* MSRs. Teach the LBR code to read those. The original
LBR registers stay the same, except they have full sign
extension now.
LBR_INFO also reports a cycle count to the last branch.
Report the cycle information using the new "cycles" branch_info
output field.
In addition we have to context switch and clear the new INFO
MSRs to avoid any information leaks.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1431285767-27027-6-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
PEBSv3 is the same as the existing PEBSv2 used on Haswell,
but it adds a new TSC field. Add support to the generic
PEBS handler to handle the new format, and overwrite
the perf time stamp using the new native_sched_clock_from_tsc().
Right now the time stamp is just slightly more accurate,
as it is nearer the actual event trigger point. With
the PEBS threshold > 1 patchkit it will be much more accurate,
avoid the problems with MMAP mismatches earlier.
The accurate time stamping is only implemented for
the default trace clock for now.
v2: Use _skl prefix. Check for default clock_id.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: eranian@google.com
Link: http://lkml.kernel.org/r/1431285767-27027-3-git-send-email-andi@firstfloor.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Intel PT chapter in the new Intel Architecture SDM adds several packets
corresponding enable bits and registers that control packet generation.
Also, additional bits in the Intel PT CPUID leaf were added to enumerate
presence and parameters of these new packets and features.
The packets and enables are:
* CYC: cycle accurate mode, provides the number of cycles elapsed since
previous CYC packet; its presence and available threshold values are
enumerated via CPUID;
* MTC: mini time counter packets, used for tracking TSC time between
full TSC packets; its presence and available resolution options are
enumerated via CPUID;
* PSB packet period is now configurable, available period values are
enumerated via CPUID.
This patch adds corresponding bit and register definitions, pmu driver
capabilities based on CPUID enumeration, new attribute format bits for
the new featurens and extends event configuration validation function
to take these into account.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@infradead.org
Cc: adrian.hunter@intel.com
Cc: hpa@zytor.com
Link: http://lkml.kernel.org/r/1438262131-12725-1-git-send-email-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently, the PT driver zeroes out the status register every time before
starting the event. However, all the writable bits are already taken care
of in pt_handle_status() function, except the new PacketByteCnt field,
which in new versions of PT contains the number of packet bytes written
since the last sync (PSB) packet. Zeroing it out before enabling PT forces
a sync packet to be written. This means that, with the existing code, a
sync packet (PSB and PSBEND, 18 bytes in total) will be generated every
time a PT event is scheduled in.
To avoid these unnecessary syncs and save a WRMSR in the fast path, this
patch changes the default behavior to not clear PacketByteCnt field, so
that the sync packets will be generated with the period specified as
"psb_period" attribute config field. This has little impact on the trace
data as the other packets that are normally sent within PSB+ (between PSB
and PSBEND) have their own generation scenarios which do not depend on the
sync packets.
One exception where we do need to force PSB like this when tracing starts,
so that the decoder has a clear sync point in the trace. For this purpose
we aready have hw::itrace_started flag, which we are currently using to
output PERF_RECORD_ITRACE_START. This patch moves setting itrace_started
from perf core to the pmu::start, where it should still be 0 on the very
first run.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: acme@infradead.org
Cc: adrian.hunter@intel.com
Cc: hpa@zytor.com
Link: http://lkml.kernel.org/r/1438264104-16189-1-git-send-email-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The previous change documents that cleanup_return_instances()
can't always detect the dead frames, the stack can grow. But
there is one special case which imho worth fixing:
arch_uretprobe_is_alive() can return true when the stack didn't
actually grow, but the next "call" insn uses the already
invalidated frame.
Test-case:
#include <stdio.h>
#include <setjmp.h>
jmp_buf jmp;
int nr = 1024;
void func_2(void)
{
if (--nr == 0)
return;
longjmp(jmp, 1);
}
void func_1(void)
{
setjmp(jmp);
func_2();
}
int main(void)
{
func_1();
return 0;
}
If you ret-probe func_1() and func_2() prepare_uretprobe() hits
the MAX_URETPROBE_DEPTH limit and "return" from func_2() is not
reported.
When we know that the new call is not chained, we can do the
more strict check. In this case "sp" points to the new ret-addr,
so every frame which uses the same "sp" must be dead. The only
complication is that arch_uretprobe_is_alive() needs to know was
it chained or not, so we add the new RP_CHECK_CHAIN_CALL enum
and change prepare_uretprobe() to pass RP_CHECK_CALL only if
!chained.
Note: arch_uretprobe_is_alive() could also re-read *sp and check
if this word is still trampoline_vaddr. This could obviously
improve the logic, but I would like to avoid another
copy_from_user() especially in the case when we can't avoid the
false "alive == T" positives.
Tested-by: Pratyush Anand <panand@redhat.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Anton Arapov <arapov@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20150721134028.GA4786@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Test-case:
#include <stdio.h>
#include <setjmp.h>
jmp_buf jmp;
void func_2(void)
{
longjmp(jmp, 1);
}
void func_1(void)
{
if (setjmp(jmp))
return;
func_2();
printf("ERR!! I am running on the caller's stack\n");
}
int main(void)
{
func_1();
return 0;
}
fails if you probe func_1() and func_2() because
handle_trampoline() assumes that the probed function should must
return and hit the bp installed be prepare_uretprobe(). But in
this case func_2() does not return, so when func_1() returns the
kernel uses the no longer valid return_instance of func_2().
Change handle_trampoline() to unwind ->return_instances until we
know that the next chain is alive or NULL, this ensures that the
current chain is the last we need to report and free.
Alternatively, every return_instance could use unique
trampoline_vaddr, in this case we could use it as a key. And
this could solve the problem with sigaltstack() automatically.
But this approach needs more changes, and it puts the "hard"
limit on MAX_URETPROBE_DEPTH. Plus it can not solve another
problem partially fixed by the next patch.
Note: this change has no effect on !x86, the arch-agnostic
version of arch_uretprobe_is_alive() just returns "true".
TODO: as documented by the previous change, arch_uretprobe_is_alive()
can be fooled by sigaltstack/etc.
Tested-by: Pratyush Anand <panand@redhat.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Anton Arapov <arapov@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20150721134021.GA4773@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Add the x86 specific version of arch_uretprobe_is_alive()
helper. It returns true if the stack frame mangled by
prepare_uretprobe() is still on stack. So if it returns false,
we know that the probed function has already returned.
We add the new return_instance->stack member and change the
generic code to initialize it in prepare_uretprobe, but it
should be equally useful for other architectures.
TODO: this assumes that the probed application can't use
multiple stacks (say sigaltstack). We will try to improve
this logic later.
Tested-by: Pratyush Anand <panand@redhat.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Anton Arapov <arapov@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20150721134018.GA4766@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
No functional changes, preparation.
Add the new helper, find_next_ret_chain(), which finds the first
!chained entry and returns its ->next. Yes, it is suboptimal. We
probably want to turn ->chained into ->start_of_this_chain
pointer and avoid another loop. But this needs the boring
changes in dup_utask(), so lets do this later.
Change the main loop in handle_trampoline() to unwind the stack
until ri is equal to the pointer returned by this new helper.
Tested-by: Pratyush Anand <panand@redhat.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Acked-by: Anton Arapov <arapov@gmail.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/20150721134013.GA4755@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
User visible changes:
- Force period term to overload global settings, i.e. previously this
command line:
$ perf record -e 'cpu/instructions,period=20000/',cycles -c 1000 sleep 1
would result in both events having a period equal to 1000, with the fix we
get something saner:
$ perf evlist -v | grep period
cpu/instructions,period=20000/: ... { sample_period, sample_freq }: 20000, ...
cycles: ... { sample_period, sample_freq }: 1000 ...
$
(Jiri Olsa)
Infrastructure changes:
- Use the dummy software event with freq=0 in the twatch.py python
binding example, to avoid disabling nohz. (Arnaldo Carvalho de Melo)
- Add some missing constants to the python binding. (Arnaldo Carvalho de Melo)
- Fix mismatched declarations for elf_getphdrnum, that happens
only in the corner case where this function is not found on
the system. (Arnaldo Carvalho de Melo)
- Add build test for having ending double slash. (Jiri Olsa)
- Introduce callgraph_set for callgraph option. (Kan Liang)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pawel Moll reported build issue for having extra slash (/) at the end of
the prefix variable.
$ make prefix=/usr/local/
CC tests/attr.o
tests/attr.c: In function ‘test__attr’:
tests/attr.c:168:50: error: expected ‘)’ before ‘;’ token
snprintf(path_perf, PATH_MAX, "%s/perf", BINDIR);
^
tests/attr.c:176:1: error: expected ‘;’ before ‘}’ token
}
^
tests/attr.c:176:1: error: control reaches end of non-void function [-Werror=return-type]
}
^
cc1: all warnings being treated as errors
Adding automated test case for this.
Reported-by: Pawel Moll <pawel.moll@arm.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/20150727182417.GD20509@krava.brq.redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently the command line option settings beats the per event period
settings:
With no global settings, we get per-event configuration:
$ perf record -e 'cpu/instructions,period=20000/' sleep 1
$ perf evlist -v
... { sample_period, sample_freq }: 20000 ...
With 'c' option period setup, we get 'c' option value:
$ perf record -e 'cpu/instructions,period=20000/' -c 1000 sleep 1
$ perf evlist -v
... { sample_period, sample_freq }: 1000 ...
This patch makes the per-event settings overload the global 'c' option
setup:
$ perf record -e 'cpu/instructions,period=20000/' -c 1000 sleep 1
$ perf evlist -v
... { sample_period, sample_freq }: 20000 ...
I think the making the per-event settings to overload any other config
makes more sense than current state. However it breaks the current
'period' term handling, which might cause some noise.. so let's see ;-).
Also fixing parse event tests with the new behaviour.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1438162936-59698-3-git-send-email-kan.liang@intel.com
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The semantic associated in tools/perf/ with foo__delete(instance) is to
release all resources referenced by 'instance' members and then release
the memory for 'instance' itself.
The perf_session_env__delete() function isn't doing this, it just does
the first part, but the space used by 'instance' itself isn't freed, as
it is embedded in a larger structure, that will be freed at other stage.
For these cases we se foo__exit(), i.e. the usage is:
void foo__delete(foo)
{
if (foo) {
foo__exit(foo);
free(foo);
}
}
But when we have something like:
struct bar {
struct foo foo;
. . .
}
Then we can't really call foo__delete(&bar.foo), we must have this
instead:
void bar__exit(bar)
{
foo__exit(&bar.foo);
/* free other bar-> resources */
}
void bar__delete(bar)
{
if (bar) {
bar__exit(bar);
free(bar);
}
}
So just rename perf_session_env__delete() to perf_session_env__exit().
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/n/tip-djbgpcfo5udqptx3q0flwtmk@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When HAVE_ELF_GETPHDRNUM_SUPPORT is false we trip on this problem:
CC /tmp/build/perf/util/symbol-elf.o
util/symbol-elf.c:41:12: error: static declaration of ‘elf_getphdrnum’ follows non-static declaration
static int elf_getphdrnum(Elf *elf, size_t *dst)
^
In file included from util/symbol.h:19:0,
from util/symbol-elf.c:8:
/usr/include/libelf.h:206:12: note: previous declaration of ‘elf_getphdrnum’ was here
extern int elf_getphdrnum (Elf *__elf, size_t *__dst);
^
MKDIR /tmp/build/perf/bench/
/home/git/linux/tools/build/Makefile.build:68: recipe for target '/tmp/build/perf/util/symbol-elf.o' failed
make[3]: *** [/tmp/build/perf/util/symbol-elf.o] Error 1
Fix it.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/n/tip-qcmekyfedmov4sxr0wahcikr@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The python binding still doesn't provide symbol resolving facilities,
but the recent addition of the trace_event__register_resolver() function
made it add as a dependency the machine__resolve_kernel_addr() method,
that in turn drags all the symbol resolving code.
The problem:
[root@zoo ~]# perf test -v python
17: Try 'import perf' in python, checking link problems :
--- start ---
test child forked, pid 6853
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ImportError: /tmp/build/perf/python/perf.so: undefined symbol: machine__resolve_kernel_addr
test child finished with -1
---- end ----
Try 'import perf' in python, checking link problems: FAILED!
[root@zoo ~]#
Fix it by requiring this function to receive the resolver as a
parameter, just like pevent_register_function_resolver(), i.e. do
not explicitely refer to an object file not included in
tools/perf/util/python-ext-sources.
[root@zoo ~]# perf test python
17: Try 'import perf' in python, checking link problems : Ok
[root@zoo ~]#
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Borislav Petkov <bp@suse.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Stephane Eranian <eranian@google.com>
Fixes: c3168b0db9 ("perf symbols: Provide libtraceevent callback to resolve kernel symbols")
Link: http://lkml.kernel.org/n/tip-vxlhh95v2em9zdbgj3jm7xi5@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When building with a prefix ending with a slash, for example:
$ make prefix=/usr/local/
one of the perf tests fail to compile due to BUILD_STR macro mishandling
bindir_SQ string containing with two slashes:
-DBINDIR="BUILD_STR(/usr/local//bin)"
with the following error:
CC tests/attr.o
tests/attr.c: In function ‘test__attr’:
tests/attr.c:168:50: error: expected ‘)’ before ‘;’ token
snprintf(path_perf, PATH_MAX, "%s/perf", BINDIR);
^
tests/attr.c:176:1: error: expected ‘;’ before ‘}’ token
}
^
tests/attr.c:176:1: error: control reaches end of non-void function [-Werror=return-type]
}
^
cc1: all warnings being treated as errors
This patch works around the problem by "cleaning" the bindir string
using make's abspath function.
Signed-off-by: Pawel Moll <pawel.moll@arm.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1438092613-21014-1-git-send-email-pawel.moll@arm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The transaction length metrics in perf stat -T broke recently.
It would not match the metric correctly and always print K/sec.
This was caused by a incorrect update of the cycles_in_tx statistics.
Update the correct variable.
Also the check for zero division was reversed, which resulted in K/sec
being printed for no transactions. Fix this also up.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1438039491-22091-1-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
New features:
- Introduce PERF_RECORD_SWITCH(_CPU_WIDE) and use it in 'record' to
ask for context switches, allowing non priviledged tasks to know
when they are switched in and out, which wasn't possible with
the other context switch tracepoint and software events, see the
patch description for a comprehensive justification (Adrian Hunter)
- Stop collecting /proc/kallsyms in perf.data files, saving about
4.5MB on a typical x86-64 system, use the the symbol resolution
routines used in all the other tools (report, top, etc) now that
we can ask libtraceevent to use perf's symbol resolution code.
(Arnaldo Carvalho de Melo)
User visible fixes:
- Expose perf's symbol resolver to libtraceevent, so that its plugins can
resolve tracepoint fields to kernel functions, like the 'function' field
in the "timer:hrtimer_start tracepoint" (Arnaldo Carvalho de Melo)
Infrastructure changes:
- Map propagation of thread and cpu maps improvements, prep work for
'perf stat' new features (Jiri Olsa)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
A recent fix to the shadow timestamp inadvertly broke the running time
accounting.
We must not update the running timestamp if we fail to schedule the
event, the event will not have ran. This can (and did) result in
negative total runtime because the stopped timestamp was before the
running timestamp (we 'started' but never stopped the event -- because
it never really started we didn't have to stop it either).
Reported-and-Tested-by: Vince Weaver <vincent.weaver@maine.edu>
Fixes: 72f669c008 ("perf: Update shadow timestamp before add event")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: stable@vger.kernel.org # 4.1
Cc: Shaohua Li <shli@fb.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
