The goal of this patch is to include more information about the host
environment into the perf.data so it is more self-descriptive. Overtime,
profiles are captured on various machines and it becomes hard to track
what was recorded, on what machine and when.
This patch provides a way to solve this by extending the perf.data file
with basic information about the host machine. To add those extensions,
we leverage the feature bits capabilities of the perf.data format. The
change is backward compatible with existing perf.data files.
We define the following useful new extensions:
- HEADER_HOSTNAME: the hostname
- HEADER_OSRELEASE: the kernel release number
- HEADER_ARCH: the hw architecture
- HEADER_CPUDESC: generic CPU description
- HEADER_NRCPUS: number of online/avail cpus
- HEADER_CMDLINE: perf command line
- HEADER_VERSION: perf version
- HEADER_TOPOLOGY: cpu topology
- HEADER_EVENT_DESC: full event description (attrs)
- HEADER_CPUID: easy-to-parse low level CPU identication
The small granularity for the entries is to make it easier to extend
without breaking backward compatiblity. Many entries are provided as
ASCII strings.
Perf report/script have been modified to print the basic information as
easy-to-parse ASCII strings. Extended information about CPU and NUMA
topology may be requested with the -I option.
Thanks to David Ahern for reviewing and testing the many versions of
this patch.
$ perf report --stdio
# ========
# captured on : Mon Sep 26 15:22:14 2011
# hostname : quad
# os release : 3.1.0-rc4-tip
# perf version : 3.1.0-rc4
# arch : x86_64
# nrcpus online : 4
# nrcpus avail : 4
# cpudesc : Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz
# cpuid : GenuineIntel,6,15,11
# total memory : 8105360 kB
# cmdline : /home/eranian/perfmon/official/tip/build/tools/perf/perf record date
# event : name = cycles, type = 0, config = 0x0, config1 = 0x0, config2 = 0x0, excl_usr = 0, excl_kern = 0, id = { 29, 30, 31,
# HEADER_CPU_TOPOLOGY info available, use -I to display
# HEADER_NUMA_TOPOLOGY info available, use -I to display
# ========
#
...
$ perf report --stdio -I
# ========
# captured on : Mon Sep 26 15:22:14 2011
# hostname : quad
# os release : 3.1.0-rc4-tip
# perf version : 3.1.0-rc4
# arch : x86_64
# nrcpus online : 4
# nrcpus avail : 4
# cpudesc : Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz
# cpuid : GenuineIntel,6,15,11
# total memory : 8105360 kB
# cmdline : /home/eranian/perfmon/official/tip/build/tools/perf/perf record date
# event : name = cycles, type = 0, config = 0x0, config1 = 0x0, config2 = 0x0, excl_usr = 0, excl_kern = 0, id = { 29, 30, 31,
# sibling cores : 0-3
# sibling threads : 0
# sibling threads : 1
# sibling threads : 2
# sibling threads : 3
# node0 meminfo : total = 8320608 kB, free = 7571024 kB
# node0 cpu list : 0-3
# ========
#
...
Reviewed-by: David Ahern <dsahern@gmail.com>
Tested-by: David Ahern <dsahern@gmail.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Richter <robert.richter@amd.com>
Cc: Andi Kleen <ak@linux.intel.com>
Link: http://lkml.kernel.org/r/20110930134040.GA5575@quad
Signed-off-by: Stephane Eranian <eranian@google.com>
[ committer notes: Use --show-info in the tools as was in the docs, rename
perf_header_fprintf_info to perf_file_section__fprintf_info, fixup
conflict with f69b64f7 "perf: Support setting the disassembler style" ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, analyzing PPC data files on x86 the cpu field is always 0 and
the tid and pid are backwards. For example, analyzing a PPC file on PPC
the pid/tid fields show:
rsyslogd 1210/1212
and analyzing the same PPC file using an x86 perf binary shows:
rsyslogd 1212/1210
The problem is that the swap_op method for samples is
perf_event__all64_swap which assumes all elements in the sample_data
struct are u64s. cpu, tid and pid are u32s and need to be handled
individually. Given that the swap is done before the sample is parsed,
the simplest solution is to undo the 64-bit swap of those elements when
the sample is parsed and do the proper swap.
The RAW data field is generic and perf cannot have programmatic knowledge
of how to treat that data. Instead a warning is given to the user.
Thanks to Anton Blanchard for providing a data file for a mult-CPU
PPC system so I could verify the fix for the CPU fields.
v3 -> v4:
- fixed use of WARN_ONCE
v2 -> v3:
- used WARN_ONCE for message regarding raw data
- removed struct wrapper around union
- fixed whitespace issues
v1 -> v2:
- added a union for undoing the byte-swap on u64 and redoing swap on
u32's to address compiler errors (see git commit 65014ab3)
Cc: Anton Blanchard <anton@samba.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1315321946-16993-1-git-send-email-dsahern@gmail.com
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The perf_event_attr struct has two __u32's at the top and
they need to be swapped individually.
With this change I was able to analyze a perf.data collected in a
32-bit PPC VM on an x86 system. I tested both 32-bit and 64-bit
binaries for the Intel analysis side; both read the PPC perf.data
file correctly.
-v2:
- changed the existing perf_event__attr_swap() to swap only elements
of perf_event_attr and exported it for use in swapping the
attributes in the file header
- updated swap_ops used for processing events
Signed-off-by: David Ahern <dsahern@gmail.com>
Acked-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: acme@ghostprotocols.net
Cc: peterz@infradead.org
Cc: paulus@samba.org
Cc: <stable@kernel.org>
Link: http://lkml.kernel.org/r/1310754849-12474-1-git-send-email-dsahern@gmail.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add an option to perf report/annotate/script to specify which
CPUs to operate on. This enables us to take a single system wide
profile and analyse each CPU (or group of CPUs) in isolation.
This was useful when profiling a multiprocess workload where the
bottleneck was on one CPU but this was hidden in the overall
profile. Per process and per thread breakdowns didn't help
because multiple processes were running on each CPU and no
single process consumed an entire CPU.
The patch converts the list of CPUs returned by cpu_map__new
into a bitmap for fast lookup. I wanted to use -C to be
consistent with perf top/record/stat, but unfortunately perf
report already uses -C <comms>.
v2: Incorporate suggestions from David Ahern:
- Added -c to perf script
- Check that SAMPLE_CPU is set when -c is used
- Update documentation
v3: Create perf_session__cpu_bitmap()
Signed-off-by: Anton Blanchard <anton@samba.org>
Acked-by: David Ahern <dsahern@gmail.com>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Link: http://lkml.kernel.org/r/20110704215750.11647eb9@kryten
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The 'sym' option displays both the function name and the DSO it comes
from. Split the display of the dso into a separate option. This allows
display of the ip address and symbol without the dso, thus shortening
line lengths - and decluttering the output a bit.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1306528124-25861-3-git-send-email-dsahern@gmail.com
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently the "sym" output field is used to dump instruction pointers
and callchain stack. Sample addresses can also be converted to symbols,
so the meaning of "sym" needs to be fixed. This patch adds an "ip"
option and if it is selected the user can also opt to dump symbols for
them. If the user opts to dump IP without syms only the address is
shown.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1306528124-25861-2-git-send-email-dsahern@gmail.com
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Check that the total size of the sample fields having a fixed
size do not exceed the one of the whole event. This robustifies
the sample parsing.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Stephane Eranian <eranian@google.com>
Check for required sample attributes using evsel rather than sample_type
in the session header. If the attribute for a default field is not
present for the event type (e.g., new command operating on file from
older kernel) the field is removed from the output list.
Expected event types must exist. For example, if a user specifies
-f trace:time,trace -f sw:time,cpu,sym
the perf.data file must contain both tracepoints and software events
(ie., it is an error if either does not exist in the file).
Attribute checking is done once at the beginning of perf-script rather
than for each sample.
v1 -> v2:
- addressed comments from acme
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1302148460-570-1-git-send-email-daahern@cisco.com
Signed-off-by: David Ahern <daahern@cisco.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Resolving the sample->id to an evsel since the most advanced tools,
report and annotate, and the others will too when they evolve to
properly support multi-event perf.data files.
Good also because it does an extra validation, checking that the ID is
valid when present. When that is not the case, the overhead is just a
branch + function call (perf_evlist__id2evsel).
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that we can reuse things like the id to attr lookup routine
(perf_evlist__id2evsel) that uses a hash table instead of the linear
lookup done in the older perf_header_attr routines, etc.
Also to make evsels/evlist more pervasive an API, simplyfing using the
emerging perf lib.
cc: Arun Sharma <arun@sharma-home.net>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
By creating an perf_evlist out of the attributes in the perf.data file
header, so that we can use evlists and evsels when reading recorded
sessions in addition to when we record sessions.
More work is needed to allow tools to allow the user to select which
events are wanted when browsing sessions, be it just one or a subset of
them, aggregated or showed at the same time but with different
indications on the UI to allow seeing workloads thru different views at
the same time.
But the overall goal/trend is to more uniformly use evsels and evlists.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
And move the event_t methods to the perf_event__ too.
No code changes, just namespace consistency.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To avoid linking more stuff in the python binding I'm working on, future
csets will make the sample type be taken from the evsel itself, but for
that we need to first have one file per cpu and per sample_type, not a
single perf.data file.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The callchains are fed with an array of a fixed size.
As a result we iterate over each callchains three times:
- 1st to resolve symbols
- 2nd to filter out context boundaries
- 3rd for the insertion into the tree
This also involves some pairs of memory allocation/deallocation
everytime we insert a callchain, for the filtered out array of
addresses and for the array of symbols that comes along.
Instead, feed the callchains through a linked list with persistent
allocations. It brings several pros like:
- Merge the 1st and 2nd iterations in one. That was possible before
but in a way that would involve allocating an array slightly taller
than necessary because we don't know in advance the number of context
boundaries to filter out.
- Much lesser allocations/deallocations. The linked list keeps
persistent empty entries for the next usages and is extendable at
will.
- Makes it easier for multiple sources of callchains to feed a
stacktrace together. This is deemed to pave the way for cfi based
callchains wherein traditional frame pointer based kernel
stacktraces will precede cfi based user ones, producing an overall
callchain which size is hardly predictable. This requirement
makes the static array obsolete and makes a linked list based
iterator a much more flexible fit.
Basic testing on a big perf file containing callchains (~ 176 MB)
has shown a throughput gain of about 11% with perf report.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1294977121-5700-2-git-send-email-fweisbec@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Not really something to be exported from session.c. Rename it to
'readn' as others did in the past.
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If we are running the new perf on an old kernel without support for
sample_id_all, we should fall back to the old unordered processing of
events. If we didn't than we would *always* process events without
timestamps out of order, whether or not we hit a reordering race. In
other words, instead of there being a chance of not attributing samples
correctly, we would guarantee that samples would not be attributed.
While processing all events without timestamps before events with
timestamps may seem like an intuitive solution, it falls down as
PERF_RECORD_EXIT events would also be processed before any samples.
Even with a workaround for that case, samples before/after an exec would
not be attributed correctly.
This patch allows commands to indicate whether they need to fall back to
unordered processing, so that commands that do not care about timestamps
on every event will not be affected. If we do fallback, this will print
out a warning if report -D was invoked.
This patch adds the test in perf_session__new so that we only need to
test once per session. Commands that do not use an event_ops (such as
record and top) can simply pass NULL in it's place.
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
LKML-Reference: <1291951882-sup-6069@au1.ibm.com>
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
At perf_session__process_event, so that we reduce the number of lines in eache
tool sample processing routine that now receives a sample_data pointer already
parsed.
This will also be useful in the next patch, where we'll allow sample the
identity fields in MMAP, FORK, EXIT, etc, when it will be possible to see (cpu,
timestamp) just after before every event.
Also validate callchains in perf_session__process_event, i.e. as early as
possible, and keep a counter of the number of events discarded due to invalid
callchains, warning the user about it if it happens.
There is an assumption that was kept that all events have the same sample_type,
that will be dealt with in the future, when this preexisting limitation will be
removed.
Tested-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Ian Munsie <imunsie@au1.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Stephane Eranian <eranian@google.com>
LKML-Reference: <1291318772-30880-4-git-send-email-acme@infradead.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The ordered sample code allocates singular reference objects struct
sample_queue which have 48byte size on 64bit and 20 bytes on 32bit. That's
silly. Allocate ~64k sized chunks and hand them out.
Performance gain: ~ 15%
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <20101130163820.398713983@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When the sample queue is flushed we free the sample reference objects. Though
we need to malloc new objects when we process further. Stop the malloc/free
orgy and cache the already allocated object for resuage. Only allocate when
the cache is empty.
Performance gain: ~ 10%
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <20101130163820.338488630@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The homebrewn sort algorithm fails to sort in time order. One of the problem
spots is that it fails to deal with equal timestamps correctly.
My first gut reaction was to replace the fancy list with an rbtree, but the
performance is 3 times worse.
Rewrite it so it works.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <20101130163819.908482530@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Move them to a session->dead_threads list just like we do with maps that
are replaced, because we may have hist_entries pointing to them.
This fixes a bug when inserting maps for a new thread that reused the
TID, mixing maps for two different threads, causing an endless loop.
The code for insering maps should be made more robust but for .35 this
is the minimalistic patch.
Reported-by: Ingo Molnar <mingo@elte.hu>
Cc: David S. Miller <davem@davemloft.net>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The changes made to support host and guest machines in a session, that
started when the 'perf kvm' tool was introduced ended up introducing a
bug where the host_machine was not having its DSOs traversed for
build-id processing.
Fix it by moving some methods to the right classes and considering the
host_machine when processing build-ids.
Reported-by: Tom Zanussi <tzanussi@gmail.com>
Reported-by: Stephane Eranian <eranian@google.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This is one more thing that started global but are more useful per hist
or per session.
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Those are really not specific to the newt code, can be used by other UI
frontends.
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
In cbbc79a we introduced support for multiple events by introducing a
new "event_stat_id" struct and then made several perf_session methods
receive a point to it instead of a pointer to perf_session, and kept the
event_stats and hists rb_tree in perf_session.
While working on the new newt based browser, I realised that it would be
better to introduce a new class, "hists" (short for "histograms"),
renaming the "event_stat_id" struct and the perf_session methods that
were really "hists" methods, as they manipulate only struct hists
members, not touching anything in the other perf_session members.
Other optimizations, such as calculating the maximum lenght of a symbol
name present in an hists instance will be possible as we add them,
avoiding a re-traversal just for finding that information.
The rationale for the name "hists" to replace "event_stat_id" is that we
may have multiple sets of hists for the same event_stat id, as, for
instance, the 'perf diff' tool has, so event stat id is not what
characterizes what this struct and the functions that manipulate it do.
Cc: Eric B Munson <ebmunson@us.ibm.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We have just one host on a given session, and that is the most common
setup right now, so embed a ->host_machine struct machine instance
directly in the perf_session class, check if we're looking for it before
going to the rb_tree.
This also fixes a problem found when we try to process old perf.data
files where we didn't have MMAP events for the kernel and modules and
thus don't create the kernel maps, do it in event__preprocess_sample if
it wasn't already.
Reported-by: Ingo Molnar <mingo@elte.hu>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Tom Zanussi <tzanussi@gmail.com>
Cc: Zhang, Yanmin <yanmin_zhang@linux.intel.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The current events reordering algorithm is based on a heuristic that
gets broken once we deal with a very fast flow of events.
Indeed the time period based flushing is not suitable anymore
in the following case, assuming we have a flush period of two
seconds.
CPU 0 | CPU 1
|
cnt1 timestamps | cnt1 timestamps
|
0 | 0
1 | 1
2 | 2
3 | 3
[...] | [...]
4 seconds later
If we spend too much time to read the buffers (case of a lot of
events to record in each buffers or when we have a lot of CPU buffers
to read), in the next pass the CPU 0 buffer could contain a slice
of several seconds of events. We'll read them all and notice we've
reached the period to flush. In the above example we flush the first
half of the CPU 0 buffer, then we read the CPU 1 buffer where we
have events that were on the flush slice and then the reordering
fails.
It's simple to reproduce with:
perf lock record perf bench sched messaging
To solve this, we use a new solution that doesn't rely on an
heuristical time slice period anymore but on a deterministic basis
based on how perf record does its job.
perf record saves the buffers through passes. A pass is a tour
on every buffers from every CPUs. This is made in order: for
each CPU we read the buffers of every counters. So the more
buffers we visit, the later will be the timstamps of their events.
When perf record finishes a pass it records a
PERF_RECORD_FINISHED_ROUND pseudo event.
We record the max timestamp t found in the pass n. Assuming these
timestamps are monotonic across cpus, we know that if a buffer
still has events with timestamps below t, they will be all available
and then read in the pass n + 1.
Hence when we start to read the pass n + 2, we can safely flush every
events with timestamps below t.
============ PASS n =================
CPU 0 | CPU 1
|
cnt1 timestamps | cnt2 timestamps
1 | 2
2 | 3
- | 4 <--- max recorded
============ PASS n + 1 ==============
CPU 0 | CPU 1
|
cnt1 timestamps | cnt2 timestamps
3 | 5
4 | 6
5 | 7 <---- max recorded
Flush every events below timestamp 4
============ PASS n + 2 ==============
CPU 0 | CPU 1
|
cnt1 timestamps | cnt2 timestamps
6 | 8
7 | 9
- | 10
Flush every events below timestamp 7
etc...
It also works on perf.data versions that don't have
PERF_RECORD_FINISHED_ROUND pseudo events. The difference is that
the events will be only flushed in the end of the perf.data
processing. It will then consume more memory and scale less with
large perf.data files.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Tom Zanussi <tzanussi@gmail.com>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Currently, perf 'live mode' writes build-ids at the end of the
session, which isn't actually useful for processing live mode events.
What would be better would be to have the build-ids sent before any of
the samples that reference them, which can be done by processing the
event stream and retrieving the build-ids on the first hit. Doing
that in perf-record itself, however, is off-limits.
This patch introduces perf-inject, which does the same job while
leaving perf-record untouched. Normal mode perf still records the
build-ids at the end of the session as it should, but for live mode,
perf-inject can be injected in between the record and report steps
e.g.:
perf record -o - ./hackbench 10 | perf inject -v -b | perf report -v -i -
perf-inject reads a perf-record event stream and repipes it to stdout.
At any point the processing code can inject other events into the
event stream - in this case build-ids (-b option) are read and
injected as needed into the event stream.
Build-ids are just the first user of perf-inject - potentially
anything that needs userspace processing to augment the trace stream
with additional information could make use of this facility.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
LKML-Reference: <1272696080-16435-3-git-send-email-tzanussi@gmail.com>
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Now those methods don't operate on a global list of dsos, but on lists
of machines, so make this clear by renaming the functions.
Cc: Avi Kivity <avi@redhat.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Zhang, Yanmin <yanmin_zhang@linux.intel.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
struct kernel_info and kerninfo__ are too vague, what they really
describe are machines, virtual ones or hosts.
There are more changes to introduce helpers to shorten function calls
and to make more clear what is really being done, but I left that for
subsequent patches.
Cc: Avi Kivity <avi@redhat.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Zhang, Yanmin <yanmin_zhang@linux.intel.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The sample events recorded by perf record are not time ordered
because we have one buffer per cpu for each event (even demultiplexed
per task/per cpu for task bound events). But when we read trace events
we want them to be ordered by time because many state machines are
involved.
There are currently two ways perf tools deal with that:
- use -M to multiplex every buffers (perf sched, perf kmem)
But this creates a lot of contention in SMP machines on
record time.
- use a post-processing time reordering (perf timechart, perf lock)
The reordering used by timechart is simple but doesn't scale well
with huge flow of events, in terms of performance and memory use
(unusable with perf lock for example).
Perf lock has its own samples reordering that flushes its memory
use in a regular basis and that uses a sorting based on the
previous event queued (a new event to be queued is close to the
previous one most of the time).
This patch proposes to export perf lock's samples reordering facility
to the session layer that reads the events. So if a tool wants to
get ordered sample events, it needs to set its
struct perf_event_ops::ordered_samples to true and that's it.
This prepares tracing based perf tools to get rid of the need to
use buffers multiplexing (-M) or to implement their own
reordering.
Also lower the flush period to 2 as it's sufficient already.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
Bypasses the build_id perf header code and replaces it with a
synthesized event and processing function that accomplishes the
same thing, used when reading/writing perf data to/from a pipe.
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: fweisbec@gmail.com
Cc: rostedt@goodmis.org
Cc: k-keiichi@bx.jp.nec.com
Cc: acme@ghostprotocols.net
LKML-Reference: <1270184365-8281-9-git-send-email-tzanussi@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Bypasses the tracing_data perf header code and replaces it with
a synthesized event and processing function that accomplishes
the same thing, used when reading/writing perf data to/from a
pipe.
The tracing data is pretty large, and this patch doesn't attempt
to break it down into component events. The tracing_data event
itself doesn't actually contain the tracing data, rather it
arranges for the event processing code to skip over it after
it's read, using the skip return value added to the event
processing loop in a previous patch.
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: fweisbec@gmail.com
Cc: rostedt@goodmis.org
Cc: k-keiichi@bx.jp.nec.com
Cc: acme@ghostprotocols.net
LKML-Reference: <1270184365-8281-8-git-send-email-tzanussi@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Bypasses the event type perf header code and replaces it with a
synthesized event and processing function that accomplishes the
same thing, used when reading/writing perf data to/from a pipe.
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: fweisbec@gmail.com
Cc: rostedt@goodmis.org
Cc: k-keiichi@bx.jp.nec.com
Cc: acme@ghostprotocols.net
LKML-Reference: <1270184365-8281-7-git-send-email-tzanussi@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Bypasses the attr perf header code and replaces it with a
synthesized event and processing function that accomplishes the
same thing, used when reading/writing perf data to/from a pipe.
Making the attrs into events allows them to be streamed over a
pipe along with the rest of the header data (in later patches).
It also paves the way to allowing events to be added and removed
from perf sessions dynamically.
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: fweisbec@gmail.com
Cc: rostedt@goodmis.org
Cc: k-keiichi@bx.jp.nec.com
Cc: acme@ghostprotocols.net
LKML-Reference: <1270184365-8281-6-git-send-email-tzanussi@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch makes several changes to allow the perf event stream
to be sent and received over a pipe:
- adds pipe-specific versions of the header read/write code
- adds pipe-specific version of the event processing code
- adds a range of event types to be used for header or other
pseudo events, above the range used by the kernel
- checks the return value of event handlers, which they can use
to skip over large events during event processing rather than actually
reading them into event objects.
- unifies the multiple do_read() functions and updates its
users.
Note that none of these changes affect the existing perf data
file format or processing - this code only comes into play if
perf output is sent to stdout (or is read from stdin).
Signed-off-by: Tom Zanussi <tzanussi@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: fweisbec@gmail.com
Cc: rostedt@goodmis.org
Cc: k-keiichi@bx.jp.nec.com
Cc: acme@ghostprotocols.net
LKML-Reference: <1270184365-8281-2-git-send-email-tzanussi@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
So that it can use it in the 'perf annotate' command line, otherwise
it'll use the default and not the specified -i filename passed to 'perf
report'.
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Due to the assumption in perf_session__new that the kernel maps would be
created using the fake PERF_RECORD_MMAP event in a perf.data file 'perf
kmem --stat caller', that doesn't have such event, ends up not being
able to resolve the kernel addresses.
Fix it by calling perf_session__create_kernel_maps() in __cmd_kmem().
LKML-Reference: <new-submission>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For when we are processing the events and inserting the entries in the
browser.
Experimentation here: naming "ui_something" we may be treading into
creating a TUI/GUI set of routines that can then be implemented in terms
of multiple backends.
Also the time it takes for adding things to the "browser" takes, visually
(I guess I should do some profiling here ;-) ), more time than for
processing the events...
That means we probably need to create a custom hist_entry browser, so
that we reuse the structures we have in place instead of duplicating
them in newt.
But progress was made and at least we can see something while long files
are being loaded, that must be one of UI 101 bullet points :-)
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We need this to know where a symbol in a callchain came from,
for various reasons, among them precise annotation from a
TUI/GUI tool.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1269459619-982-5-git-send-email-acme@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Newt has widespread availability and provides a rather simple
API as can be seen by the size of this patch.
The work needed to support it will benefit other frontends too.
In this initial patch it just checks if the output is a tty, if
not it falls back to the previous behaviour, also if
newt-devel/libnewt-dev is not installed the previous behaviour
is maintaned.
Pressing enter on a symbol will annotate it, ESC in the
annotation window will return to the report symbol list.
More work will be done to remove the special casing in
color_fprintf, stop using fmemopen/FILE in the printing of
hist_entries, etc.
Also the annotation doesn't need to be done via spawning "perf
annotate" and then browsing its output, we can do better by
calling directly the builtin-annotate.c functions, that would
then be moved to tools/perf/util/annotate.c and shared with perf
top, etc
But lets go by baby steps, this patch already improves perf
usability by allowing to quickly do annotations on symbols from
the report screen and provides a first experimentation with
libnewt/TUI integration of tools.
Tested on RHEL5 and Fedora12 X86_64 and on Debian PARISC64 to
browse a perf.data file collected on a Fedora12 x86_64 box.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1268349164-5822-5-git-send-email-acme@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch adds the structures necessary to count each event
type independently in perf report.
Signed-off-by: Eric B Munson <ebmunson@us.ibm.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1267804269-22660-4-git-send-email-acme@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We want to stream events as fast as possible to perf.data, and
also in the future we want to have splice working, when no
interception will be possible.
Using build_id__mark_dso_hit_ops to create the list of DSOs that
back MMAPs we also optimize disk usage in the build-id cache by
only caching DSOs that had hits.
Suggested-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1265223128-11786-6-git-send-email-acme@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
I noticed while writing the first test in 'perf regtest' that to
just test the symbol handling routines one needs to create a
perf session, that is a layer centered on a perf.data file,
events, etc, so I untied these layers.
This reduces the complexity for the users as the number of
parameters to most of the symbols and session APIs now was
reduced while not adding more state to all the map instances by
only having data that is needed to split the kernel (kallsyms
and ELF symtab sections) maps and do vmlinux relocation on the
main kernel map.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1265223128-11786-1-git-send-email-acme@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
