Pull perf updates from Ingo Molnar:
"This update is pretty big and almost exclusively includes tooling
changes, because v4.9's LTS status forced to completion most of the
pending kernel side hardware enablement work and because we tried to
freeze core perf work a bit to give a time window for the fuzzing
efforts.
The diff is large mostly due to the JSON hardware event tables added
for Intel and Power8 CPUs. This was a popular feature request from
people working close to hardware and from the HPC community.
Tree size is big because this added the CPU event tables for over a
decade of Intel CPUs. Future changes for a CPU vendor alrady support
should be much smaller, as events for new models are added. The new
events are listed in 'perf list', for the CPU model the tool is
running on. If you find an interesting event it can be used as-is:
$ perf stat -a -e l2_lines_out.pf_clean sleep 1
Performance counter stats for 'system wide':
7,860,403 l2_lines_out.pf_clean
1.000624918 seconds time elapsed
The event lists can be searched the usual 'perf list' fashion for
(case insensitive) substrings as well:
$ perf list l2_lines_out
List of pre-defined events (to be used in -e):
cache:
l2_lines_out.demand_clean
[Clean L2 cache lines evicted by demand]
l2_lines_out.demand_dirty
[Dirty L2 cache lines evicted by demand]
l2_lines_out.dirty_all
[Dirty L2 cache lines filling the L2]
l2_lines_out.pf_clean
[Clean L2 cache lines evicted by L2 prefetch]
l2_lines_out.pf_dirty
[Dirty L2 cache lines evicted by L2 prefetch]
etc.
There's a few high level categories as well that can be listed:
'cache', 'floating point', 'frontend', 'memory', 'pipeline', 'virtual
memory'.
Existing generic events and workflows should work as-is.
The only kernel side change is a late breaking fix for an older
regression, related to Intel BTS, LBR and PT feature interaction.
On the tooling side there are three new tools / major features:
- The new 'perf c2c' tool provides means for Shared Data C2C/HITM
analysis.
This allows you to track down cacheline contention. The tool is
based on x86's load latency and precise store facility events
provided by Intel CPUs.
It was tested by Joe Mario and has proven to be useful, finding
some cacheline contentions. Joe also wrote a blog about c2c tool
with examples:
https://joemario.github.io/blog/2016/09/01/c2c-blog/
excerpt of the content on this site:
At a high level, “perf c2c” will show you:
* The cachelines where false sharing was detected.
* The readers and writers to those cachelines, and the offsets where those accesses occurred.
* The pid, tid, instruction addr, function name, binary object name for those readers and writers.
* The source file and line number for each reader and writer.
* The average load latency for the loads to those cachelines.
* Which numa nodes the samples a cacheline came from and which CPUs were involved.
Using perf c2c is similar to using the Linux perf tool today.
First collect data with “perf c2c record”, then generate a
report output with “perf c2c report”
There one finds extensive details on using the tool, with tips on
reducing the volume of samples while still capturing enough to do
its job. (Dick Fowles, Joe Mario, Don Zickus, Jiri Olsa)
- The new 'perf sched timehist' tool provides tailored analysis of
scheduling events.
Example usage:
perf sched record -- sleep 1
perf sched timehist
By default it shows the individual schedule events, including the
wait time (time between sched-out and next sched-in events for the
task), the task scheduling delay (time between wakeup and actually
running) and run time for the task:
time cpu task name wait time sch delay run time
[tid/pid] (msec) (msec) (msec)
-------- ------ ---------------- --------- --------- --------
1.874569 [0011] gcc[31949] 0.014 0.000 1.148
1.874591 [0010] gcc[31951] 0.000 0.000 0.024
1.874603 [0010] migration/10[59] 3.350 0.004 0.011
1.874604 [0011] <idle> 1.148 0.000 0.035
1.874723 [0005] <idle> 0.016 0.000 1.383
1.874746 [0005] gcc[31949] 0.153 0.078 0.022
...
Times are in msec.usec. (David Ahern, Namhyung Kim)
- Add CPU vendor hardware event tables:
Add JSON files with vendor event naming for Intel and Power8
processors, allowing users of tools like oprofile to keep using the
event names they are used to, as well as people reading vendor
documentation, where such naming is used. (Andi Kleen, Sukadev
Bhattiprolu)
You should see all the new events with 'perf list' and you should
be able to search them, for example 'perf list miss' will list all
the myriads of miss events.
Other tooling features added were:
- Cross-arch annotation support:
o Improve ARM support in the annotation code, affecting 'perf
annotate', 'perf report' and live annotation in 'perf top' (Kim
Phillips)
o Initial support for PowerPC in the annotation code (Ravi
Bangoria)
o Support AArch64 in the 'annotate' code, native/local and
cross-arch/remote (Kim Phillips)
- Allow considering just events in a given time interval, via the
'--time start.s.ms,end.s.ms' command line, added to 'perf kmem',
'perf report', 'perf sched timehist' and 'perf script' (David
Ahern)
- Add option to stop printing a callchain at one of a given group of
symbol names (David Ahern)
- Track memory freed in 'perf kmem stat' (David Ahern)
- Allow querying and setting .perfconfig variables (Taeung Song)
- Show branch information in callchains (predicted, TSX aborts, loop
iteractions, etc) (Jin Yao)
- Dynamicly change verbosity level by pressing 'V' in the 'perf
top/report' hists TUI browser (Alexis Berlemont)
- Implement 'perf trace --delay' in the same fashion as in 'perf
record --delay', to skip sampling workload initialization events
(Alexis Berlemont)
- Make vendor named events case insensitive in 'perf list', i.e.
'perf list LONGEST_LAT' works just the same as 'perf list
longest_lat' (Andi Kleen)
- Add unwinding support for jitdump (Stefano Sanfilippo)
Tooling infrastructure changes:
- Support linking perf with clang and LLVM libraries, initially
statically, but this limitation will be lifted and shared
libraries, when available, will be preferred to the static build,
that should, as with other features, be enabled explicitly (Wang
Nan)
- Add initial support (and perf test entry) for tooling hooks,
starting with 'record_start' and 'record_end', that will have as
its initial user the eBPF infrastructure, where perf_ prefixed
functions will be JITed and run when such hooks are called (Wang
Nan)
- Implement assorted libbpf improvements (Wang Nan)"
... and lots of other changes, features, cleanups and refactorings I
did not list, see the shortlog and the git log for details"
* 'perf-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (220 commits)
perf/x86: Fix exclusion of BTS and LBR for Goldmont
perf tools: Explicitly document that --children is enabled by default
perf sched timehist: Cleanup idle_max_cpu handling
perf sched timehist: Handle zero sample->tid properly
perf callchain: Introduce callchain_cursor__copy()
perf sched: Cleanup option processing
perf sched timehist: Improve error message when analyzing wrong file
perf tools: Move perf build related variables under non fixdep leg
perf tools: Force fixdep compilation at the start of the build
perf tools: Move PERF-VERSION-FILE target into rules area
perf build: Check LLVM version in feature check
perf annotate: Show raw form for jump instruction with indirect target
perf tools: Add non config targets
perf tools: Cleanup build directory before each test
perf tools: Move python/perf.so target into rules area
perf tools: Move install-gtk target into rules area
tools build: Move tabs to spaces where suitable
tools build: Make the .cmd file more readable
perf clang: Compile BPF script using builtin clang support
perf clang: Support compile IR to BPF object and add testcase
...