linux.git/tools/perf/util/sort.c, branch v5.12

perf report: Fix -F for branch & mem modes

2021-03-06T19:54:32+00:00

perf report fails to add valid additional fields with -F when
used with branch or mem modes. Fix it.

Before patch:

  $ perf record -b
  $ perf report -b -F +srcline_from --stdio
  Error:
  Invalid --fields key: `srcline_from'

After patch:

  $ perf report -b -F +srcline_from --stdio
  # Samples: 8K of event 'cycles'
  # Event count (approx.): 8784
  ...

Committer notes:

There was an inversion: when looking at branch stack dimensions (keys)
it was checking if the sort mode was 'mem', not 'branch'.

Fixes: aa6b3c99236b ("perf report: Make -F more strict like -s")
Reported-by: Athira Jajeev 
Signed-off-by: Ravi Bangoria 
Reviewed-by: Athira Jajeev 
Tested-by: Arnaldo Carvalho de Melo 
Tested-by: Athira Jajeev 
Cc: Jiri Olsa 
Cc: Kan Liang 
Cc: Namhyung Kim 
Link: http://lore.kernel.org/lkml/20210304062958.85465-1-ravi.bangoria@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo

perf report: Support instruction latency

2021-02-08T19:25:00+00:00

The instruction latency information can be recorded on some platforms,
e.g., the Intel Sapphire Rapids server. With both memory latency
(weight) and the new instruction latency information, users can easily
locate the expensive load instructions, and also understand the time
spent in different stages. The users can optimize their applications in
different pipeline stages.

The 'weight' field is shared among different architectures. Reusing the
'weight' field may impacts other architectures. Add a new field to store
the instruction latency.

Like the 'weight' support, introduce a 'ins_lat' for the global
instruction latency, and a 'local_ins_lat' for the local instruction
latency version.

Add new sort functions, INSTR Latency and Local INSTR Latency,
accordingly.

Add local_ins_lat to the default_mem_sort_order[].

Signed-off-by: Kan Liang 
Cc: Andi Kleen 
Cc: Jin Yao 
Cc: Jiri Olsa 
Cc: Madhavan Srinivasan 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Stephane Eranian 
Link: http://lore.kernel.org/lkml/1612296553-21962-7-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf tools: Support data block and addr block

2021-02-08T19:25:00+00:00

Two new data source fields, to indicate the block reasons of a load
instruction, are introduced on the Intel Sapphire Rapids server. The
fields can be used by the memory profiling.

Add a new sort function, SORT_MEM_BLOCKED, for the two fields.

For the previous platforms or the block reason is unknown, print "N/A"
for the block reason.

Add blocked as a default mem sort key for perf report and perf mem
report.

Committer testing:

So in machines without this capability we get a "N/A" filling the new "Blocked"
column:

  $ perf mem record ls
  arch     certs	 CREDITS  Documentation  include  ipc     Kconfig  lib       MAINTAINERS  mm   samples  security  usr    block
  COPYING	 crypto	 drivers  fs             init     Kbuild  kernel   LICENSES  Makefile     net  README   scripts   sound  tools
  virt
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.008 MB perf.data (17 samples) ]
  $
  $ perf mem report --stdio
  # To display the perf.data header info, please use --header/--header-only options.
  #
  # Total Lost Samples: 0
  #
  # Samples: 6  of event 'cpu/mem-loads,ldlat=30/Pu'
  # Total weight : 1381
  # Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked
  #
  # Overhead  Samples  Local Weight  Memory access         Symbol                   Shared Object  Data Symbol             Data Object   Snoop  TLB access    Locked  Blocked
  # ........  .......  ............  ....................  .......................  .............  ......................  ............  .....  ............  ......  .......
  #
      32.87%        1  454           Local RAM or RAM hit  [.] _dl_relocate_object  ld-2.31.so     [.] 0x00007fe91cef3078  libc-2.31.so  Hit    L1 or L2 hit  No       N/A
      25.56%        1  353           LFB or LFB hit        [.] strcmp               ld-2.31.so     [.] 0x00005586973855ca  ls            None   L1 or L2 hit  No       N/A
      22.59%        1  312           LFB or LFB hit        [.] _dl_cache_libcmp     ld-2.31.so     [.] 0x00007fe91d0e3b18  ld.so.cache   None   L1 or L2 hit  No       N/A
       8.47%        1  117           LFB or LFB hit        [.] _dl_relocate_object  ld-2.31.so     [.] 0x00007fe91ceee570  libc-2.31.so  None   L1 or L2 hit  No       N/A
       6.88%        1  95            LFB or LFB hit        [.] _dl_relocate_object  ld-2.31.so     [.] 0x00007fe91ceed490  libc-2.31.so  None   L1 or L2 hit  No       N/A
       3.62%        1  50            LFB or LFB hit        [.] _dl_cache_libcmp     ld-2.31.so     [.] 0x00007fe91d0ebe60  ld.so.cache   None   L1 or L2 hit  No       N/A

  # Samples: 11  of event 'cpu/mem-stores/Pu'
  # Total weight : 11
  # Sort order   : local_weight,mem,sym,dso,symbol_daddr,dso_daddr,snoop,tlb,locked,blocked
  #
  # Overhead  Samples  Local Weight  Memory access  Symbol                   Shared Object  Data Symbol             Data Object  Snoop  TLB access  Locked  Blocked
  # ........  .......  ............  .............  .......................  .............  ......................  ...........  .....  ..........  ......  .......
  #
       9.09%        1  0             L1 hit         [.] __strcoll_l          libc-2.31.so   [.] 0x00007fffe5648fc8  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] _dl_lookup_symbol_x  ld-2.31.so     [.] 0x00007fffe56490b8  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] _dl_name_match_p     ld-2.31.so     [.] 0x00007fffe56487d8  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] _dl_start            ld-2.31.so     [.] start_time+0x0      ld-2.31.so   N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] _dl_sysdep_start     ld-2.31.so     [.] 0x00007fffe56494b8  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] do_lookup_x          ld-2.31.so     [.] 0x00007fffe5648ff8  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] do_lookup_x          ld-2.31.so     [.] 0x00007fffe5649064  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 hit         [.] do_lookup_x          ld-2.31.so     [.] 0x00007fffe5649130  [stack]      N/A    N/A         N/A      N/A
       9.09%        1  0             L1 miss        [.] _dl_start            ld-2.31.so     [.] _rtld_global+0xaf8  ld-2.31.so   N/A    N/A         N/A      N/A
       9.09%        1  0             L1 miss        [.] _dl_start            ld-2.31.so     [.] _rtld_global+0xc28  ld-2.31.so   N/A    N/A         N/A      N/A
       9.09%        1  0             L1 miss        [.] _dl_start            ld-2.31.so     [.] 0x00007fffe56495b8  [stack]      N/A    N/A         N/A      N/A

  # (Tip: Show user configuration overrides: perf config --user --list)
  $

Signed-off-by: Kan Liang 
Tested-by: Arnaldo Carvalho de Melo 
Cc: Andi Kleen 
Cc: Jin Yao 
Cc: Jiri Olsa 
Cc: Madhavan Srinivasan 
Cc: Namhyung Kim 
Cc: Peter Zijlstra 
Cc: Stephane Eranian 
Link: http://lore.kernel.org/lkml/1612296553-21962-4-git-send-email-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf report: Add support for PERF_SAMPLE_CODE_PAGE_SIZE

2021-01-20T17:34:20+00:00

Add a new sort dimension "code_page_size" for common sort.
With this option applied, perf can sort and report by sample's code page
size.

For example:

  # perf report --stdio --sort=comm,symbol,code_page_size
  # To display the perf.data header info, please use
  # --header/--header-only options.
  #
  #
  # Total Lost Samples: 0
  #
  # Samples: 3K of event 'mem-loads:uP'
  # Event count (approx.): 1470769
  #
  # Overhead  Command  Symbol                        Code Page Size IPC [IPC Coverage]
  # ........  .......  ............................  .............. ....................
  #
      69.56%  dtlb     [.] GetTickCount              4K              -   -
      17.93%  dtlb     [.] Calibrate                 4K              -   -
      11.40%  dtlb     [.] __gettimeofday            4K              -   -
  #

Signed-off-by: Stephane Eranian 
Acked-by: Jiri Olsa 
Acked-by: Namhyung Kim 
Cc: Andi Kleen 
Cc: Ingo Molnar 
Cc: Mark Rutland 
Cc: Michael Ellerman 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20210105195752.43489-6-kan.liang@linux.intel.com
Signed-off-by: Kan Liang 
Signed-off-by: Arnaldo Carvalho de Melo

perf sort: Add sort option for data page size

2020-12-19T20:52:24+00:00

Add a new sort option "data_page_size" for --mem-mode sort.  With this
option applied, perf can sort and report by sample's data page size.

Here is an example:

perf report --stdio --mem-mode
--sort=comm,symbol,phys_daddr,data_page_size

 # To display the perf.data header info, please use
 # --header/--header-only options.
 #
 #
 # Total Lost Samples: 0
 #
 # Samples: 9K of event 'mem-loads:uP'
 # Total weight : 9028
 # Sort order   : comm,symbol,phys_daddr,data_page_size
 #
 # Overhead  Command  Symbol                        Data Physical
 # Address
 # Data Page Size
 # ........  .......  ............................
 # ......................  ......................
 #
    11.19%  dtlb     [.] touch_buffer              [.] 0x00000003fec82ea8  4K
     8.61%  dtlb     [.] GetTickCount              [.] 0x00000003c4f2c8a8  4K
     4.52%  dtlb     [.] GetTickCount              [.] 0x00000003fec82f58  4K
     4.33%  dtlb     [.] __gettimeofday            [.] 0x00000003fec82f48  4K
     4.32%  dtlb     [.] GetTickCount              [.] 0x00000003fec82f78  4K
     4.28%  dtlb     [.] GetTickCount              [.] 0x00000003fec82f50  4K
     4.23%  dtlb     [.] GetTickCount              [.] 0x00000003fec82f70  4K
     4.11%  dtlb     [.] GetTickCount              [.] 0x00000003fec82f68  4K
     4.00%  dtlb     [.] Calibrate                 [.] 0x00000003fec82f98  4K
     3.91%  dtlb     [.] Calibrate                 [.] 0x00000003fec82f90  4K
     3.43%  dtlb     [.] touch_buffer              [.] 0x00000003fec82e98  4K
     3.42%  dtlb     [.] touch_buffer              [.] 0x00000003fec82e90  4K
     0.09%  dtlb     [.] DoDependentLoads          [.] 0x000000036ea084c0  2M
     0.08%  dtlb     [.] DoDependentLoads          [.] 0x000000032b010b80  2M

Signed-off-by: Kan Liang 
Acked-by: Jiri Olsa 
Acked-by: Namhyung Kim 
Cc: Andi Kleen 
Cc: Mark Rutland 
Cc: Michael Ellerman 
Cc: Stephane Eranian 
Cc: Will Deacon 
Link: http://lore.kernel.org/lkml/20201216185805.9981-3-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo

perf evlist: Use the right prefix for 'struct evlist' evsel list methods

2020-11-30T17:52:44+00:00

perf_evlist__ is for 'struct perf_evlist' methods, in tools/lib/perf/,
go on completing this split.

Cc: Adrian Hunter 
Cc: Ian Rogers 
Cc: Jiri Olsa 
Cc: Namhyung Kim 
Signed-off-by: Arnaldo Carvalho de Melo

perf util: Fix memory leak of prefix_if_not_in

2020-05-28T13:03:27+00:00

Need to free "str" before return when asprintf() failed to avoid memory
leak.

Signed-off-by: Xie XiuQi 
Cc: Alexander Shishkin 
Cc: Hongbo Yao 
Cc: Jiri Olsa 
Cc: Li Bin 
Cc: Mark Rutland 
Cc: Namhyung Kim 
Link: http://lore.kernel.org/lkml/20200521133218.30150-4-liwei391@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo

perf symbol: Fix kernel symbol address display

2020-05-05T19:35:32+00:00

Running commands

   ./perf record -e rb0000 -- find .
   ./perf report -v

reveals symbol names and its addresses. There is a mismatch between
kernel symbol and address. Here is an example for kernel symbol
check_chain_key:

 3.55%  find /lib/modules/.../build/vmlinux  0xf11ec  v [k] check_chain_key

This address is off by 0xff000 as can be seen with:

[root@t35lp46 ~]# fgrep check_chain_key /proc/kallsyms
00000000001f00d0 t check_chain_key
[root@t35lp46 ~]# objdump -t ~/linux/vmlinux| fgrep check_chain_key
00000000001f00d0 l     F .text	00000000000001e8 check_chain_key
[root@t35lp46 ~]#

This function is located in main memory 0x1f00d0 - 0x1f02b4. It has
several entries in the perf data file with the correct address:

[root@t35lp46 perf]# ./perf report -D -i perf.data.find-bad | \
				fgrep SAMPLE| fgrep 0x1f01ec
PERF_RECORD_SAMPLE(IP, 0x1): 22228/22228: 0x1f01ec period: 1300000 addr: 0
PERF_RECORD_SAMPLE(IP, 0x1): 22228/22228: 0x1f01ec period: 1300000 addr: 0

The root cause happens when reading symbol tables during perf report.
A long gdb call chain leads to

   machine__deliver_events
     perf_evlist__deliver_event
       perf_evlist__deliver_sample
         build_id__mark_dso_hits
	   thread__find_map(1)      Read correct address from sample entry
	     map__load
	       dso__load            Some more functions to end up in
	         ....
		 dso__load_sym.

Function dso__load_syms  checks for kernel relocation and symbol
adjustment for the kernel and results in kernel map adjustment of
	 kernel .text segment address (0x100000 on s390)
	 kernel .text segment offset in file (0x1000 on s390).
This results in all kernel symbol addresses to be changed by subtracting
0xff000 (on s390). For the symbol check_chain_key we end up with

    0x1f00d0 - 0x100000 + 0x1000 = 0xf11d0

and this address is saved in the perf symbol table. This calculation is
also applied by the mapping functions map__mapip() and map__unmapip()
to map IP addresses to dso mappings.

During perf report processing functions

   process_sample_event    (builtin-report.c)
     machine__resolve
       thread__find_map
     hist_entry_iter_add

are called. Function thread__find_map(1)
takes the correct sample address and applies the mapping function
map__mapip() from the kernel dso and saves the modified address
in struct addr_location for further reference. From now on this address
is used.

Funktion process_sample_event() then calls hist_entry_iter_add() to save
the address in member ip of struct hist_entry.

When samples are displayed using

    perf_evlist__tty_browse_hists
      hists__fprintf
        hist_entry__fprintf
	  hist_entry__snprintf
	    __hist_entry__snprintf
	      _hist_entry__sym_snprintf()

This simply displays the address of the symbol and ignores the dso <-> map
mappings done in function thread__find_map. This leads to the address
mismatch.

Output before:

ot@t35lp46 perf]# ./perf report -v | fgrep check_chain_key
     3.55%  find     /lib/modules/5.6.0d-perf+/build/vmlinux
     						0xf11ec v [k] check_chain_key
[root@t35lp46 perf]#

Output after:

[root@t35lp46 perf]# ./perf report -v | fgrep check_chain_key
     3.55%  find     /lib/modules/5.6.0d-perf+/build/vmlinux
     						0x1f01ec v [k] check_chain_key
[root@t35lp46 perf]#

Signed-off-by: Thomas Richter 
Acked-by: Sumanth Korikkar 
Cc: Heiko Carstens 
Cc: Vasily Gorbik 
Link: http://lore.kernel.org/lkml/20200415070744.59919-1-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo

perf evsel: Rename perf_evsel{prev,next}() to evsel{prev,next}()

2020-05-05T19:35:31+00:00

As those are 'struct evsel' methods, not part of tools/lib/perf/, aka
libperf, to whom the perf_ prefix belongs.

Cc: Adrian Hunter 
Cc: Jiri Olsa 
Cc: Namhyung Kim 
Signed-off-by: Arnaldo Carvalho de Melo

perf hist: Add fast path for duplicate entries check

2020-04-18T12:05:01+00:00

Perf checks the duplicate entries in a callchain before adding an entry.
However the check is very slow especially with deeper call stack.
Almost ~50% elapsed time of perf report is spent on the check when the
call stack is always depth of 32.

The hist_entry__cmp() is used to compare the new entry with the old
entries. It will go through all the available sorts in the sort_list,
and call the specific cmp of each sort, which is very slow.

Actually, for most cases, there are no duplicate entries in callchain.
The symbols are usually different. It's much faster to do a quick check
for symbols first. Only do the full cmp when the symbols are exactly the
same.

The quick check is only to check symbols, not dso. Export
_sort__sym_cmp.

  $ perf record --call-graph lbr ./tchain_edit_64

  Without the patch
  $time perf report --stdio
  real    0m21.142s
  user    0m21.110s
  sys     0m0.033s

  With the patch
  $time perf report --stdio
  real    0m10.977s
  user    0m10.948s
  sys     0m0.027s

Signed-off-by: Kan Liang 
Acked-by: Jiri Olsa 
Cc: Adrian Hunter 
Cc: Alexey Budankov 
Cc: Andi Kleen 
Cc: Mathieu Poirier 
Cc: Michael Ellerman 
Cc: Namhyung Kim 
Cc: Pavel Gerasimov 
Cc: Peter Zijlstra 
Cc: Ravi Bangoria 
Cc: Stephane Eranian 
Cc: Vitaly Slobodskoy 
Link: http://lore.kernel.org/lkml/20200319202517.23423-18-kan.liang@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo