We'll also have to fix the argument ordering in different arches,
probably having multiple syscall_fmt entries with each possible order
and then use perf_evsel__env_arch() (if dealing with a perf.data file)
or the current system info, for live sessions.
perf trace: Allow specifying number of syscall args for tracepointless syscalls
When we don't have syscalls:sys_{enter,exit}_NAME, we had to resort to
dumping all the 6 syscall arguments, fix it by providing that info for
such syscalls, like 'clone'.
perf trace: Ditch __syscall__arg_val() variant, not needed anymore
All callers now can use syscall__arg_val(arg, idx), be it to iterate
thru the syscall arguments while taking into account alignment, or to
get values for other arguments that affect how the current argument
should be formatted (think of fcntl's 'cmd' and 'arg' arguments).
perf trace: Allow allocating sc->arg_fmt even without the syscall tracepoint
At least "clone" doesn't have (enter, exit) entries tracefs/events/syscalls/,
but we can provide a syscall_fmt and use it instead, as will be done for
"clone" in the next cset.
Kan Liang [Fri, 30 Jun 2017 14:16:56 +0000 (10:16 -0400)]
perf intel-pt: Always set no branch for dummy event
An earlier kernel patch allowed enabling PT and LBR at the same time on
Goldmont.
commit ccbebba4c6bf ("perf/x86/intel/pt: Bypass PT vs. LBR exclusivity
if the core supports it")
However, users still cannot use Intel PT and LBRs simultaneously. $
sudo perf record -e cycles,intel_pt//u -b -- sleep 1 Error: PMU
Hardware doesn't support sampling/overflow-interrupts.
PT implicitly adds dummy event in perf tool. dummy event is software
event which doesn't support LBR.
Always setting no branch for dummy event in Intel PT.
Kan Liang [Fri, 30 Jun 2017 14:16:55 +0000 (10:16 -0400)]
perf intel-pt: Set no_aux_samples for the tracking event
The reason of introducing the tracking event (a dummy software event) is
to collect side-band information. Additional sampling is wasteful.
no_aux_samples should be set for tracking event.
Merge tag 'perf-core-for-mingo-4.13-20170718' of git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux into perf/core
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
User visible changes:
- Initial support for namespaces, using setns to access files in
namespaces, grabbing their build-ids, etc. We still need to work
more to deal with namespaces that vanish before we can get the
needed data to do analysis, but this should be as good as what is
in bcc now (Krister Johansen)
- Add header record types to pipe-mode, now this command:
Will show the same as in non-pipe mode, i.e. involving a perf.data
file (David Carrillo-Cisneros)
- Implement a visual marker for fused x86 instructions in the annotate
TUI browser, available now in 'perf report', more work needed to have
it available as well in 'perf top' (Jin Yao)
- Beautify the fcntl syscall, which is an interesting one in the sense
that infrastructure had to be put in place to change the formatters of
some arguments according to the value in a previous one, i.e. cmd
dictates how arg and the syscall return will be formatted.
(Arnaldo Carvalho de Melo
Infrastructure changes:
- 'perf test attr' fixes (Jiri Olsa)
Vendor events changes:
- Add POWER9 PMU events Sukadev (Bhattiprolu)
- Support additional POWER8+ PVR in PMU mapfile (Shriya)
perf/core: Fix scheduling regression of pinned groups
Vince Weaver reported:
> I was tracking down some regressions in my perf_event_test testsuite.
> Some of the tests broke in the 4.11-rc1 timeframe.
>
> I've bisected one of them, this report is about
> tests/overflow/simul_oneshot_group_overflow
> This test creates an event group containing two sampling events, set
> to overflow to a signal handler (which disables and then refreshes the
> event).
>
> On a good kernel you get the following:
> Event perf::instructions with period 1000000
> Event perf::instructions with period 2000000
> fd 3 overflows: 946 (perf::instructions/1000000)
> fd 4 overflows: 473 (perf::instructions/2000000)
> Ending counts:
> Count 0: 946379875
> Count 1: 946365218
>
> With the broken kernels you get:
> Event perf::instructions with period 1000000
> Event perf::instructions with period 2000000
> fd 3 overflows: 938 (perf::instructions/1000000)
> fd 4 overflows: 318 (perf::instructions/2000000)
> Ending counts:
> Count 0: 946373080
> Count 1: 653373058
The root cause of the bug is that the following commit:
487f05e18a ("perf/core: Optimize event rescheduling on active contexts")
erronously assumed that event's 'pinned' setting determines whether the
event belongs to a pinned group or not, but in fact, it's the group
leader's pinned state that matters.
This was discovered by Vince in the test case described above, where two instruction
counters are grouped, the group leader is pinned, but the other event is not;
in the regressed case the counters were off by 33% (the difference between events'
periods), but should be the same within the error margin.
Fix the problem by looking at the group leader's pinning.
COND_FWD: conditional forward
COND_BWD: conditional backward
COND: conditional branch
UNCOND: unconditional branch
IND: indirect
CALL: function call
IND_CALL: indirect function call
RET: function return
SYSCALL: syscall
SYSRET: syscall return
COND_CALL: conditional function call
COND_RET: conditional function return
CROSS_4K and CROSS_2M:
They are the metrics checking for branches cross 4K or 2MB pages.
It's an approximate computing. We don't know if the area is 4K or
2MB, so always compute both.
To make the output simple, if a branch crosses 2M area, CROSS_4K
will not be incremented.
Change log
v7: Since the common branch type definitions are changed, some
tags/strings are updated accordingly.
v6: Remove branch_type_stat_display() since it's moved to branch.c.
v5: Remove the unnecessary sort__mode checking in
hist_iter__branch_callback().
v4: Comparing to previous version, the major changes are:
Add the computing of JCC forward/JCC backward and cross page checking
by using the from and to addresses.
Jin Yao [Tue, 18 Jul 2017 12:13:13 +0000 (20:13 +0800)]
perf util: Create branch.c/.h for common branch functions
Create new util/branch.c and util/branch.h to contain the common branch
functions. Such as:
branch_type_count(): Count the numbers of branch types
branch_type_name() : Return the name of branch type
branch_type_stat_display(): Display branch type statistics info
branch_type_str(): Construct the branch type string.
The branch type is saved in branch_flags.
Change log:
v8: Change PERF_BR_NONE to PERF_BR_UNKNOWN.
v7: Since the common branch type name is changed (e.g. JCC->COND),
this patch is performed the modification accordingly.
v6: Move that multiline conditional code inside {} brackets.
Move branch_type_stat_display() from builtin-report.c to
branch.c.
Move branch_type_str() from callchain.c to branch.c.
Jin Yao [Tue, 18 Jul 2017 12:13:10 +0000 (20:13 +0800)]
perf/x86/intel: Record branch type
Perf already has support for disassembling the branch instruction
and using the branch type for filtering. The patch just records
the branch type in perf_branch_entry.
Before recording, the patch converts the x86 branch type to
common branch type.
Change log:
v10: Set the branch_map array to be static. The previous version
has it on stack then makes the compiler to create it every
time when the function gets called.
v9: Use __ffs() to find first bit in type in common_branch_type().
It lets the code be clear.
v8: Change PERF_BR_NONE to PERF_BR_UNKNOWN.
v7: Just convert following x86 branch types to common branch types.
Jin Yao [Tue, 18 Jul 2017 12:13:09 +0000 (20:13 +0800)]
perf/core: Define the common branch type classification
It is often useful to know the branch types while analyzing branch data.
For example, a call is very different from a conditional branch.
Currently we have to look it up in binary while the binary may later not
be available and even the binary is available but user has to take some
time. It is very useful for user to check it directly in perf report.
Perf already has support for disassembling the branch instruction to get
the x86 branch type.
To keep consistent on kernel and userspace and make the classification
more common, the patch adds the common branch type classification
in perf_event.h.
The patch only defines a minimum but most common set of branch types.
PERF_BR_UNKNOWN : unknown
PERF_BR_COND ：conditional
PERF_BR_UNCOND : unconditional
PERF_BR_IND : indirect
PERF_BR_CALL : function call
PERF_BR_IND_CALL : indirect function call
PERF_BR_RET : function return
PERF_BR_SYSCALL : syscall
PERF_BR_SYSRET : syscall return
PERF_BR_COND_CALL : conditional function call
PERF_BR_COND_RET : conditional function return
The patch also adds a new field type (4 bits) in perf_branch_entry
to record the branch type.
Since the disassembling of branch instruction needs some overhead,
a new PERF_SAMPLE_BRANCH_TYPE_SAVE is introduced to indicate if it
needs to disassemble the branch instruction and record the branch
type.
Change log:
v10: Not changed.
v9: Not changed.
v8: Change PERF_BR_NONE to PERF_BR_UNKNOWN.
No other change.
v7: Just keep the most common branch types.
Others are removed.
v6: Not changed.
v5: Not changed. The v5 patch series just change the userspace.
v4: Comparing to previous version, the major changes are:
1. Remove the PERF_BR_JCC_FWD/PERF_BR_JCC_BWD, they will be
computed later in userspace.
2. Remove the "cross" field in perf_branch_entry. The cross page
computing will be done later in userspace.
There are three FEAT_OP* macros:
- FEAT_OPA: for features without process record.
- FEAT_OPP: for features with process record.
- FEAT_OPF: like FEAT_OPP but to show only if show_full_info flags
is set.
To add pipe-mode headers we need yet another variation of the macros
(one to specify whether a feature generates an auxiliar record).
Instead, we redefine macros so that:
- show_full_info is specified as an argument (to remove the
FEAT_OPF variation) and,
- it always sets "process" handler (to remove the FEAT_OPA variation).
Individual process handlers can be NULLed individually.
This allows to define two variations only:
- FEAT_OPR: synthesizes auxiliar event record.
- FEAT_OPN: doesn't synthesize an auxiliar event record.
Extend struct feat_fd to use a temporal buffer in pipe-mode, instead of
perf.data's file descriptor.
The header features build_id and aux_trace already have logic to print
in file-mode that heavily rely on lseek the file. For now, leave such
features inactive in pipe-mode and print a warning if their functions
are called in pipe-mode.
perf header: Don't pass struct perf_file_section to process_##_feat
struct perf_file_section is used in process_##_feat as container for
size and offset in the file descriptor. These attributes are meaninful
in pipe-mode but struct perf_file_section is not.
Add offset and size variables to struct feat_fd to store
perf_file_section's values in file-mode. Later on, the same variables
can be reused for pipe-mode.
Introduce struct feat_fd. This patch uses it as a wrapper around fd in
write_* functions for feature headers. Next patches will extend its
functionality to other feature header functions.
perf util: Add const modifier to buf in "writen" function
Make buf in helper function "writen" constant to simplify the life of
its callers.
This requires to hack a cast of buf prior to passing it to "ion" which
is simpler than the alternative of reworking the "ion" function to
provide a read and a write paths, the latter with constant buf argument.
Currently perf supports a mode to query inline stack. It works well for
finding user space inline functions but it doesn't work for kernel ones,
due to some unnecessary check.
This patch removes these unnecessary checks. Now kernel inline functions
can be reported.
Removing syscall_fmt::err_msg and instead always formatting negative
returns as errno values.
With this we can remove a lot of entries that have no special handling
besides the ones we can do by looking at the tracefs format files, i.e.
the types for the fields (e.g. pid_t), well known names (e.g. fd).
perf trace: Allow syscall arg formatters to request non suppression of zeros
The 'perf trace' tool is suppressing args set to zero, with the
exception of string tables (strarrays), which are kinda like enums, i.e.
we have maps to go from numbers to strings.
But the 'cmd' fcntl arg requires more specialized treatment, as its
value will regulate if the next fcntl syscall arg, 'arg', should be
ignored (not used) and also how to format the syscall return (fd, file
flags, etc), so add a 'show_zero" bool to struct syscall_arg_fmt, to
regulate this more explicitely.
Will be used in a following patch with fcntl, here is just the
mechanism.
perf trace: Group per syscall arg formatter info into one struct
Instead of having syscall_fmt.{arg_scnprintf,arg_parm}, introduce
struct syscall_arg_fmt and have these two, paving the way for more
state to change the formatting algorithms.
For instance, in the 'fcntl' 'cmd' case it is better not to suppress
it when being zero, showing instead its name "DUPFD".
We had that in an ad-hoc way just for strarrays, but with more involved
cases like fcntl, that can't be done with just a strarray, we'll need
a ".show_zero" arg in the 'cmd' syscall_arg_fmt.
Jiri Olsa [Mon, 3 Jul 2017 14:50:24 +0000 (16:50 +0200)]
perf tests attr: Add proper return values
The record command now properly returns the status of the tracee if
there's any. We need to properly set the expected return value of the
tracee in the attr tests.
Jiri Olsa [Mon, 3 Jul 2017 14:50:21 +0000 (16:50 +0200)]
perf tests attr: Add 1s for exclude_kernel and task base bits
There's an event open fallback which set exclude_kernel=1 in case use
does not have enough privileges. Adding both 0|1 for this attribute,
because we don't know what value it is.
Jiri Olsa [Mon, 3 Jul 2017 14:50:18 +0000 (16:50 +0200)]
perf tests attr: Add test_attr__ready function
We create many test events before the real ones just to test specific
features. But there's no way for attr tests to separate those test
events from those it needs to check.
Adding 'ready' call from the events open interface to trigger/start
events collection for attr test.
It can return NULL, in which case we should bail out and remove the
directory created with mkdtemp(), which is stored in the "__tempdir"
variable, not in "tempdir".
Cc: Masami Hiramatsu <[email protected]> Fixes: 8e5dc848356e ("perf test: Add a test case for SDT event") Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>
perf record: Do not ask for precise_ip with --no-samples
When the user doesn't specify an event with -e/--event, 'perf record'
will use as a default the "cycles" event with the highest level of
precision in perf_event_attr.precise_ip, but --no-samples, if present,
is incompatible with precise_ip != 0, so use the newly introduced
__perf_event__add_default(precise = false) to fix that:
Before:
# perf record -n usleep 1
Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
Error:
The sys_perf_event_open() syscall returned with 22 (Invalid argument) for event (cycles:ppp).
/bin/dmesg may provide additional information.
No CONFIG_PERF_EVENTS=y kernel support configured?
#
After:
# perf record -n usleep 1
Please consider tweaking /proc/sys/kernel/perf_event_max_sample_rate.
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.018 MB perf.data ]
[root@jouet /]# perf evlist -v
cycles: size: 112, sample_type: IP|TID|TIME|PERIOD, disabled: 1, inherit: 1, mmap: 1, comm: 1, freq: 1, enable_on_exec: 1, task: 1, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1
[root@jouet /]#
perf evlist: Allow asking for max precise_ip in add_default()
There are cases where we want to leave attr.precise_ip as zero, such
as when using 'perf record --no-samples', where this would make the
kernel return -EINVAL.
perf evsel: Allow asking for max precise_ip in new_cycles()
There are cases where we want to leave attr.precise_ip as zero, such
as when using 'perf record --no-samples', where this would make the
kernel return -EINVAL.
If a stripped binary is placed in the cache, the user is in a situation
where there's a cached elf file present, but it doesn't have any symtab
to use for name resolution. Grab the debuginfo for binaries that don't
end in .ko. This yields a better chance of resolving symbols from older
traces.
perf buildid-cache: Support binary objects from other namespaces
Teach buildid-cache how to add, remove, and update binary objects from
other mount namespaces. Allow probe events tracing binaries in
different namespaces to add their objects to the probe and build-id
caches too. As a handy side effect, this also lets us access SDT probes
in binaries from alternate mount namespaces.
perf probe: Allow placing uprobes in alternate namespaces.
Teaches perf how to place a uprobe on a file that's in a different mount
namespace. The user must add the probe using the --target-ns argument
to perf probe. Once it has been placed, it may be recorded against
without further namespace-specific commands.
perf maps: Lookup maps in both intitial mountns and inner mountns.
If a process is in a mountns and has symbols in /tmp/perf-<pid>.map,
look first in the namespace using the tgid for the pidns that the
process might be in. If the map isn't found there, try looking in the
mountns where perf is running, and use the tgid that's appropriate for
perf's pid namespace. If all else fails, use the original pid.
This allows us to locate a symbol map file in the mount namespace, if it
was generated there. However, we also try the tool's /tmp in case it's
there instead.
perf symbols: Find symbols in different mount namespace
Teach perf how to resolve symbols from binaries that are in a different
mount namespace from the tool. This allows perf to generate meaningful
stack traces even if the binary resides in a different mount namespace
from the tool.
In older distros we were not including our copies of unistd_{32,64}.h,
as we were relying on the system's asm/unistd.h, and a log time ago
the files to be included were asm-{x86_64,i386}/unistd.h.
Fix it by also carrying a copy of asm/unistd.h, that will be the same
as in modern distros and will allow us to provide missing __NR_setns,
for instance, in older distros.
perf trace beauty: Export the pid beautifier for use in more places
Now that the beautifiers are being split into multiple source and object
files, we will need more of them exported, do it for the 'pid' one, will
be used to augment the return of some syscalls that may return a 'pid',
such as fcntl(fd, F_GETOWN).
perf trace beauty: Export the fd beautifier for use in more places
Now that the beautifiers are being split into multiple source and object
files, we will need more of them exported, do it for the 'fd' one, will
be used to augment the return of some syscalls that may return an 'fd',
such as fcntl(fd, F_DUPFD).
perf trace beauty: Give syscall return beautifier more context
We need the current thread and the trace internal state so that we can
use the fd beautifier to augment syscall returns, so use struct
syscall_arg with some fields that make sense on returns (val, thread,
trace).
perf beauty open: Detach the syscall_arg agnostic bits from the flags formatter
We may want to use this in other contexts, like when formatting the
return of fcntl(fd, F_GETFL).
Make it have the following signature, so that we can set the formatter
for the return argument while processing the arguments of a syscall, as
fcntl, for instance, may return fds, flags, etc, so need different
return value formatters:
size_t formatter(unsigned long value, char *bf, size_t size);
This gets so detached from 'perf trace' internals that we may well get
all these and move to a tools/lib/syscall_beauty/ library at some point
and use it in other tools/ living utilities.
perf trace: Beautify new write hint fcntl commands
Those introduced by the commit c75b1d9421f8 ("fs: add fcntl() interface for
setting/getting write life time hints"), tested using the proggie in that
commit comment:
Sometimes it should be printed as an hex number, like with F_SETLK,
F_SETLKW and F_GETLK, that treat 'arg' as a struct flock pointer, in
other cases it is just an integer.
tools include uapi asm-generic: Grab a copy of fcntl.h
We'll need defines for beautifying fcntl arguments that are not
available in older distros, these:
trace/beauty/fcntl.c: In function 'syscall_arg__scnprintf_fcntl_arg':
trace/beauty/fcntl.c:93: error: 'F_OFD_SETLK' undeclared (first use in this function)
trace/beauty/fcntl.c:93: error: (Each undeclared identifier is reported only once
trace/beauty/fcntl.c:93: error: for each function it appears in.)
trace/beauty/fcntl.c:93: error: 'F_OFD_SETLKW' undeclared (first use in this function)
trace/beauty/fcntl.c:93: error: 'F_OFD_GETLK' undeclared (first use in this function)
trace/beauty/fcntl.c:94: error: 'F_GETOWN_EX' undeclared (first use in this function)
trace/beauty/fcntl.c:94: error: 'F_SETOWN_EX' undeclared (first use in this function)
perf trace beauty: Introduce syscall arg beautifier for long integers
Will be used in the fcntl arg beautifier, that nowadays formats as '%ld'
because there is no explicit arg beautifier attached, but as we will
have to first decide what beautifier to use (i.e. it may be a pointer,
etc) then we need to have this exported as a separate beautifier to be
called from there.
perf trace beauty: Allow accessing syscall args values in a syscall arg formatter
For instance, fcntl's upcoming 'arg' formatter needs to look at the
'cmd' value to decide how to format its value, sometimes it is a file
flags, sometimes an fd, a pointer to a structure, etc.
tools: Update include/uapi/linux/fcntl.h copy from the kernel
To get the changes in the commit c75b1d9421f8 ("fs: add fcntl()
interface for setting/getting write life time hints").
Silencing this perf build warning:
Warning: include/uapi/linux/fcntl.h differs from kernel
We already beautify the fcntl cmd argument, so an upcoming cset will
update the 'cmd' strarray to cover these new commands.
The hints are in the 3rd arg, a pointer, so not yet supported in 'perf
trace', for that we need to copy it somehow, probably using eBPF, a new
attempt at doing that is planned.
perf trace: Remove F_ from some of the fcntl command strings
The initial ones already had that "F_" prefix stripped to make things
shorter, some hadn't, do it now.
We do this to make the 'perf trace' output more compact. At some point
perhaps the best thing to do is to have the tool do this stripping
automatically, letting the user also decide if this is to be done or
not. For now, be consistent.
Jin Yao [Fri, 7 Jul 2017 05:06:35 +0000 (13:06 +0800)]
perf annotate: Implement visual marker for macro fusion
For marking fused instructions clearly this patch adds a line before the
first instruction of pair and joins it with the arrow of the jump to its
target.
For example, when "je" is selected in annotate view, the line before
cmpl is displayed and joins the arrow of "je".
Jin Yao [Fri, 7 Jul 2017 05:06:34 +0000 (13:06 +0800)]
perf annotate: Check for fused instructions
Macro fusion merges two instructions to a single micro-op. Intel core
platform performs this hardware optimization under limited
circumstances.
For example, CMP + JCC can be "fused" and executed /retired together.
While with sampling this can result in the sample sometimes being on the
JCC and sometimes on the CMP. So for the fused instruction pair, they
could be considered together.
On Nehalem, fused instruction pairs:
cmp/test + jcc.
On other new CPU:
cmp/test/add/sub/and/inc/dec + jcc.
This patch adds an x86-specific function which checks if 2 instructions
are in a "fused" pair. For non-x86 arch, the function is just NULL.
Changelog:
v4: Move the CPU model checking to symbol__disassemble and save the CPU
family/model in arch structure.
It avoids checking every time when jump arrow printed.
v3: Add checking for Nehalem (CMP, TEST). For other newer Intel CPUs
just check it by default (CMP, TEST, ADD, SUB, AND, INC, DEC).
v2: Remove the original weak function. Arnaldo points out that doing it
as a weak function that will be overridden by the host arch doesn't
work. So now it's implemented as an arch-specific function.
Committer fix:
Do not access evsel->evlist->env->cpuid, ->env can be null, introduce
perf_evsel__env_cpuid(), just like perf_evsel__env_arch(), also used in
this function call.
The original patch was segfaulting 'perf top' + annotation.
But this essentially disables this fused instructions augmentation in
'perf top', the right thing is to get the cpuid from the running kernel,
left for a later patch tho.
Jiri Olsa [Fri, 14 Jul 2017 16:35:51 +0000 (18:35 +0200)]
perf/x86/intel: Fix debug_store reset field for freq events
There's a bug in PEBs event enabling code, that prevents PEBS
freq events to work properly after non freq PEBS event was run.
freq events - perf_event_attr::freq set
-F <freq> option of perf record
PEBS events - perf_event_attr::precise_ip > 0
default for perf record
Like in following example with CPU 0 busy, we expect ~10000 samples
for following perf tool run:
# perf record -F 10000 -C 0 sleep 1
[ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 0.640 MB perf.data (10031 samples) ]
Everything's fine, but once we run non freq PEBS event like:
# perf record -c 10000 -C 0 sleep 1
[ perf record: Woken up 4 times to write data ]
[ perf record: Captured and wrote 1.053 MB perf.data (20061 samples) ]
the freq events start to fail like this:
# perf record -F 10000 -C 0 sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.185 MB perf.data (40 samples) ]
The issue is in non freq PEBs event initialization of debug_store reset
field, which value is used to auto-reload the counter value after PEBS
event drain. This value is not being used for PEBS freq events, but once
we run non freq event it stays in debug_store data and screws the
sample_freq counting for PEBS freq events.
projects / linux.git / log