Merge tag 'perf-core-for-mingo-5.5-20191122' of git://git.kernel.org/pub/scm/linux...
authorIngo Molnar <mingo@kernel.org>
Sat, 23 Nov 2019 08:00:13 +0000 (09:00 +0100)
committerIngo Molnar <mingo@kernel.org>
Sat, 23 Nov 2019 08:00:13 +0000 (09:00 +0100)
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:

perf report:

  Jin Yao:

  - Allow entering the annotation view (symbol source/assembly +
    overhead/cycles/etc column) from the 'perf report --total-cycles'
    interface.

    E.g.:

      # perf record --all-cpus --branch-any --all-kernel
      ^C[ perf record: Woken up 5 times to write data ]
      #
      # perf evlist -v
      cycles: size: 120, { sample_period, sample_freq }: 4000,
      sample_type: IP|TID|TIME|CPU|PERIOD|BRANCH_STACK,
      read_format: ID, disabled: 1, inherit: 1, exclude_user: 1, mmap: 1, comm: 1, freq: 1, task: 1,
      precise_ip: 3, sample_id_all: 1, exclude_guest: 1, mmap2: 1, comm_exec: 1, ksymbol: 1,
      bpf_event: 1, branch_sample_type: ANY
      #
      # perf report --total-cycles
      #
      # Samples: 78762 of event 'cycles'
      Sampled  Sampled Avg      Avg
      Cycles%  Cycles  Cycles%  Cycles                           [Program Block Range]     Shared Object
        1.72%    95.8K   0.00%     254                        [msr.h:105 -> msr.h:166]  [kernel.vmlinux]
        1.56%   107.6K   0.00%     618                [compiler.h:199 -> common.c:301]  [kernel.vmlinux]
        0.83%    46.3K   0.00%     409              [entry_64.S:153 -> entry_64.S:175]  [kernel.vmlinux]
        0.83%    46.1K   0.00%      83                  [jump_label.h:41 -> tsc.c:230]  [kernel.vmlinux]
        0.64%    36.9K   0.01%    1.4K            [hda_intel.c:904 -> hda_intel.c:916]   [snd_hda_intel]
        0.57%    30.2K   0.00%     282                      [file.c:710 -> file.c:730]  [kernel.vmlinux]
        0.48%    25.8K   0.00%      82              [spinlock.c:158 -> spinlock.c:160]  [kernel.vmlinux]
        0.45%    23.7K   0.00%     369  [tick-broadcast.c:585 -> tick-broadcast.c:586]  [kernel.vmlinux]
        0.44%    24.4K   0.00%      73                       [msr.h:236 -> tsc.c:1088]  [kernel.vmlinux]
        0.43%    22.7K   0.00%     144                [cpuidle.c:229 -> cpuidle.c:232]  [kernel.vmlinux]

    Then press 'A' or Enter on one of those lines, just like with 'perf top', say
    the top one: [msr.h:105 -> msr.h:166], then this shows up:

      Samples: 78K of event 'cycles', 4000 Hz, Event count (approx.): 78762
      native_write_msr  /lib/modules/5.4.0-rc8/build/vmlinux [Percent: local period]
      Percent│ IPC Cycle (Average IPC: 0.02, IPC Coverage: 50.0%)
             │
             │             Disassembly of section .text:
             │
             │             ffffffff8106c480 <native_write_msr>:
             │             __wrmsr():
             │             return EAX_EDX_VAL(val, low, high);
             │             }
             │
             │             static inline void notrace __wrmsr(unsigned int msr, u32 low, u32 high)
             │             {
             │             asm volatile("1: wrmsr\n"
       49.16 │0.02           mov   %edi,%ecx
             │0.02           mov   %esi,%eax
             │0.02           wrmsr
             │             arch_static_branch():
             │             #include <linux/stringify.h>
             │             #include <linux/types.h>
             │
             │             static __always_inline bool arch_static_branch(struct static_key *key, bool branch)
             │             {
             │             asm_volatile_goto("1:"
        0.79 │0.02           nop
             │             native_write_msr():
             │             {
             │             __wrmsr(msr, low, high);
             │
             │             if (msr_tracepoint_active(__tracepoint_write_msr))
             │             do_trace_write_msr(msr, ((u64)high << 32 | low), 0);
             │             }
       50.05 │0.02  254    ← retq
             │             do_trace_write_msr(msr, ((u64)high << 32 | low), 0);
             │               shl   $0x20,%rdx
             │               mov   %esi,%esi
             │               or    %rdx,%rsi
             │               xor   %edx,%edx
             │             → jmpq  do_trace_write_msr

    We need to improve this to show the source code line numbers in the
    annotation view, so one can go from that program block to the annotation view
    and see those source code line numbers straight away.

auxtrace/Intel PT:

  Adrian Hunter:

  - Add support for AUX area sampling, requires new functionality that
    will land in 5.5, its already in tip.

    This includes kernel capability querying so that it fails gracefully
    with older kernels, duimping aux area samples in 'perf report -D' and
    'perf script'.

perf.data:

  Alexey Budankov:

  - Fix decompression of PERF_RECORD_COMPRESSED records.

core:

  Arnaldo Carvalho de Melo:

  - Use the 'dcacheline' cmp routine to find the right DSOs taking into
    account the 'maj', 'min', 'ino' and 'ino_generation', that got moved
    from 'struct map' to 'struct dso', where it belongs.

    This further reduces the size of 'struct map', there is still more
    work to do to maybe get it to max one cacheline.

libtraceevent:

  Hewenliang:

  - Fix memory leakage in copy_filter_type().

  Sudip Mukherjee:

  - Fix header installation.

perf parse:

  Ian Rogers :

  - Fix potential memory leak when handling tracepoint errors, found using
    LLVM's libFuzzer.

perf probe:

  Colin Ian King:

  - Fix spelling mistake "addrees" -> "address".

Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

Trivial merge