perf c2c report: Fix crash for empty browser
Jiri Olsa
reviewerPATCH 1/1 perf c2c report: Fix crash for empty browser
Do not try to display entry details if there's
not any. Currently this ends up in crash:
$ perf c2c report perf: Segmentation fault
--- tools/perf/builtin-c2c.c | 3 +++ 1 file changed, 3 insertions(+)
tools/perf/builtin-c2c.c +3 -0
@@ -2349,6 +2349,9 @@ static int perf_c2c__browse_cacheline(struct hist_entry *he)" s Toggle full length of symbol and source line columns \n" " q Return back to cacheline list \n" ; +if (!he) +return 0 ; +/* Display compact version first. */ c2c.symbol_full = false;
Signature · 1 line
2.17.1
Arnaldo Carvalho de Melo
maintainerEm Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu:
Quoted 4 lines
Do not try to display entry details if there's
not any. Currently this ends up in crash:$ perf c2c report perf: Segmentation fault
How to replicate this?
I tried:
$ perf record sleep 1 $ perf c2c report
But it didn't segfault
Quoted 23 lines
Reported-by [email protected]Signed-off-by Jiri Olsa <[email protected]>--- tools/perf/builtin-c2c.c | 3 +++ 1 file changed, 3 insertions(+)tools/perf/builtin-c2c.c +3
-0@@ -2349,6 +2349,9 @@ static int perf_c2c__browse_cacheline(struct hist_entry *he)" s Toggle full length of symbol and source line columns \n" " q Return back to cacheline list \n" ; +if (!he) +return 0 ; +/* Display compact version first. */ c2c.symbol_full = false;Signature · 1 line
2.17.1
Quoted 6 lines · Arnaldo Carvalho de Melo
Em Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu:
Quoted 4 lines
Do not try to display entry details if there's
not any. Currently this ends up in crash:$ perf c2c report perf: Segmentation fault
Quoted 8 lines
How to replicate this?
I tried:
$ perf record sleep 1 $ perf c2c reportBut it didn't segfault
Similarly I have tried :
$ perf record sleep 1 $ perf c2c report Then Press `d` to show the cache-line contents. This replies the segfault on my machine (4.17.8-1). The patch mentioned above should solve it, even tough I am not sure as I haven't been able to recompile the kernel.
The segfault by itself seems to be due to the report logic, as it did not expect to report on an empty browser.
What has stepped me back is that application which I have been testing with rely on multiple threads instantiated through pthread, which should be counted in user-level threads right? But they still seem to return an empty browser.
When instead c2c is runned system-wide, with an application running on multiple threads like firefox or julia, cache hits are measured and also they are traced back in the source code.
Anyway, thanks for your support!
Quoted 14 lines · On Thu, Jul 26, 2018 at 11:31:34PM +0000, [email protected] wrote:
Quoted 13 lines · Arnaldo Carvalho de Melo
Em Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu:
Quoted 4 lines
Do not try to display entry details if there's
not any. Currently this ends up in crash:$ perf c2c report perf: Segmentation faultHow to replicate this?
I tried:
$ perf record sleep 1 $ perf c2c report
Quoted 6 lines
Quoted 1 line
But it didn't segfault
Similarly I have tried :
$ perf record sleep 1 $ perf c2c report Then Press `d` to show the cache-line contents.
yep, sry I forgot to mention you need to press the 'd' to show details
Quoted 3 lines
This replies the segfault on my machine (4.17.8-1).
The patch mentioned above should solve it, even tough I am not sure as I haven't been able to recompile the kernel.
no need to recompile kernel
Quoted 7 lines
The segfault by itself seems to be due to the report logic, as it did not expect to report on an empty browser.
What has stepped me back is that application which I have been testing with rely on multiple threads instantiated through pthread, which should be counted in user-level threads right? But they still seem to return an empty browser.
right, c2c scans read/write accesses and tries to find false sharing cases maybe there was nothing to be found
Quoted 3 lines
When instead c2c is runned system-wide, with an application running on multiple threads like firefox or julia, cache hits are measured and also they are traced back in the source code.
I got a cache line (attached) for 'perf bench sched messaging' NOT being traced system wide and just for user (you'll get plenty of detected cachelines in kernel space):
jirka
---
[root@krava perf]# ./perf c2c record --all-user -- ./perf bench sched messaging -l 100000 [root@krava perf]# ./perf c2c report --stdio
=================================================
Shared Data Cache Line Table
=================================================
#
# ----------- Cacheline ---------- Total Tot ----- LLC Load Hitm ----- ---- Store Reference ---- --- Load Dram ---- LLC Total ----- Core Load Hit ----- -- LLC Load Hit --
# Index Address Node PA cnt records Hitm Total Lcl Rmt Total L1Hit L1Miss Lcl Rmt Ld Miss Loads FB L1 L2 Llc Rmt
# ..... .................. .... ...... ....... ....... ....... ....... ....... ....... ....... ....... ........ ........ ....... ....... ....... ....... ....... ........ ........
#
0 0x7fff5b729cc0 0 1 44 100.00% 1 1 0 21 21 0 2 0 2 23 0 0 9 11 0=================================================
Shared Cache Line Distribution Pareto
=================================================
#
# ----- HITM ----- -- Store Refs -- --------- Data address --------- ---------- cycles ---------- Total cpu Shared
# Num Rmt Lcl L1 Hit L1 Miss Offset Node PA cnt Pid Code address rmt hitm lcl hitm load records cnt Symbol Object Source:Line Node
# ..... ....... ....... ....... ....... .................. .... ...... ....... .................. ........ ........ ........ ....... ........ ............... .................. ..................... ....
#
-------------------------------------------------------------
0 0 1 21 0 0x7fff5b729cc0
-------------------------------------------------------------
0.00% 100.00% 0.00% 0.00% 0x38 0 1 17356 0x7febaf7e1a46 0 142 101 23 4 [.] __libc_read libpthread-2.27.so read.c:28 0
0.00% 0.00% 100.00% 0.00% 0x38 0 1 17356 0x494e4e 0 0 0 21 4 [.] receiver perf sched-messaging.c:129 0Adding Joe to the CC list.
Em Fri, Jul 27, 2018 at 09:06:23AM +0200, Jiri Olsa escreveu:
Quoted 40 lines · On Thu, Jul 26, 2018 at 11:31:34PM +0000, [email protected] wrote:
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 12 lines
Em Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu:
Quoted 4 lines
Do not try to display entry details if there's
not any. Currently this ends up in crash:$ perf c2c report perf: Segmentation faultHow to replicate this?
I tried:
$ perf record sleep 1 $ perf c2c reportQuoted 6 lines
Quoted 1 line
But it didn't segfault
Similarly I have tried :
$ perf record sleep 1 $ perf c2c report Then Press `d` to show the cache-line contents.yep, sry I forgot to mention you need to press the 'd' to show details
Quoted 3 lines
This replies the segfault on my machine (4.17.8-1).
The patch mentioned above should solve it, even tough I am not sure as I haven't been able to recompile the kernel.no need to recompile kernel
Quoted 6 lines
The segfault by itself seems to be due to the report logic, as it did not expect to report on an empty browser.
What has stepped me back is that application which I have been testing with rely on multiple threads instantiated through pthread, which should be counted in user-level threads right? But they still seem to return an empty browser.right, c2c scans read/write accesses and tries to find false sharing cases maybe there was nothing to be found
Quoted 7 lines
Quoted 3 lines
When instead c2c is runned system-wide, with an application running on multiple threads like firefox or julia, cache hits are measured and also they are traced back in the source code.
I got a cache line (attached) for 'perf bench sched messaging' NOT being traced system wide and just for user (you'll get plenty of detected cachelines in kernel space):
With that info in mind, we get:
[root@seventh ~]# perf record sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.001 MB perf.data (6 samples) ] [root@seventh ~]#
[root@seventh ~]#
[root@seventh ~]# perf c2c report # and press 'd'
perf: Segmentation fault
-------- backtrace --------
perf[0x5b1d2a] /lib64/libc.so.6(+0x346df)[0x7fcb566e36df] perf[0x46fcae]
perf[0x4a9f1e]
perf[0x4aa220]
perf(main+0x301)[0x42c561] /lib64/libc.so.6(__libc_start_main+0xe9)[0x7fcb566cff29] perf(_start+0x29)[0x42c999]
[root@seventh ~]#
With your patches the segfault is gone, but I'd do a follow up patch to show some message telling the user why 'd' showed nothing and instructing him about what is missing, i.e. is this done on a perf.data file that has no events of interest? Suggest using 'perf c2c record' or 'perf record -e events,of,interest,to,perf,c2c', was this done on some workload where no false sharing was detected? Say so, etc.
I applied your patch with a more detailed commit log to state how this can reproduced, etc, as usual:
https://git.kernel.org/acme/c/983eb6aa7098
- Arnaldo
Quoted 71 lines · Arnaldo Carvalho de Melo
Adding Joe to the CC list.
Em Fri, Jul 27, 2018 at 09:06:23AM +0200, Jiri Olsa escreveu:
Quoted 40 lines · On Thu, Jul 26, 2018 at 11:31:34PM +0000, [email protected] wrote:
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 12 lines
Em Tue, Jul 24, 2018 at 08:20:08AM +0200, Jiri Olsa escreveu:
Quoted 4 lines
Do not try to display entry details if there's
not any. Currently this ends up in crash:$ perf c2c report perf: Segmentation faultHow to replicate this?
I tried:
$ perf record sleep 1 $ perf c2c reportQuoted 6 lines
Quoted 1 line
But it didn't segfault
Similarly I have tried :
$ perf record sleep 1 $ perf c2c report Then Press `d` to show the cache-line contents.yep, sry I forgot to mention you need to press the 'd' to show details
Quoted 3 lines
This replies the segfault on my machine (4.17.8-1).
The patch mentioned above should solve it, even tough I am not sure as I haven't been able to recompile the kernel.no need to recompile kernel
Quoted 6 lines
The segfault by itself seems to be due to the report logic, as it did not expect to report on an empty browser.
What has stepped me back is that application which I have been testing with rely on multiple threads instantiated through pthread, which should be counted in user-level threads right? But they still seem to return an empty browser.right, c2c scans read/write accesses and tries to find false sharing cases maybe there was nothing to be found
Quoted 7 lines
Quoted 3 lines
When instead c2c is runned system-wide, with an application running on multiple threads like firefox or julia, cache hits are measured and also they are traced back in the source code.
I got a cache line (attached) for 'perf bench sched messaging' NOT being traced system wide and just for user (you'll get plenty of detected cachelines in kernel space):
With that info in mind, we get:
[root@seventh ~]# perf record sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.001 MB perf.data (6 samples) ] [root@seventh ~]#
[root@seventh ~]#
[root@seventh ~]# perf c2c report # and press 'd'
perf: Segmentation fault
-------- backtrace --------
perf[0x5b1d2a] /lib64/libc.so.6(+0x346df)[0x7fcb566e36df] perf[0x46fcae]
perf[0x4a9f1e]
perf[0x4aa220]
perf(main+0x301)[0x42c561] /lib64/libc.so.6(__libc_start_main+0xe9)[0x7fcb566cff29] perf(_start+0x29)[0x42c999]
[root@seventh ~]#
Quoted 6 lines
With your patches the segfault is gone, but I'd do a follow up patch to show some message telling the user why 'd' showed nothing and instructing him about what is missing, i.e. is this done on a perf.data file that has no events of interest? Suggest using 'perf c2c record' or 'perf record -e events,of,interest,to,perf,c2c', was this done on some workload where no false sharing was detected? Say so, etc.
ok, will try to come up with something
Quoted 5 lines
I applied your patch with a more detailed commit log to state how this can reproduced, etc, as usual:
I checked, looks good, thanks
jirka
tip-bot for Jiri Olsa
merge botperf c2c report: Fix crash for empty browser
Commit-ID: 73978332572ccf5e364c31e9a70ba953f8202b46 Gitweb: https://git.kernel.org/tip/73978332572ccf5e364c31e9a70ba953f8202b46 Author: Jiri Olsa <[email protected]> AuthorDate: Tue, 24 Jul 2018 08:20:08 +0200 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Tue, 31 Jul 2018 10:53:20 -0300
perf c2c report: Fix crash for empty browser
'perf c2c' scans read/write accesses and tries to find false sharing cases, so when the events it wants were not asked for or ended up not taking place, we get no histograms.
So do not try to display entry details if there's not any. Currently this ends up in crash:
$ perf c2c report # then press 'd' perf: Segmentation fault $
Committer testing:
Before:
Record a perf.data file without events of interest to 'perf c2c report', then call it and press 'd':
# perf record sleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.001 MB perf.data (6 samples) ] # perf c2c report perf: Segmentation fault -------- backtrace -------- perf[0x5b1d2a] /lib64/libc.so.6(+0x346df)[0x7fcb566e36df] perf[0x46fcae] perf[0x4a9f1e] perf[0x4aa220] perf(main+0x301)[0x42c561] /lib64/libc.so.6(__libc_start_main+0xe9)[0x7fcb566cff29] perf(_start+0x29)[0x42c999] #
After the patch the segfault doesn't take place, a follow up patch to tell the user why nothing changes when 'd' is pressed would be good.
--- tools/perf/builtin-c2c.c | 3 +++ 1 file changed, 3 insertions(+)
tools/perf/builtin-c2c.c +3 -0
@@ -2349,6 +2349,9 @@ static int perf_c2c__browse_cacheline(struct hist_entry *he)" s Toggle full length of symbol and source line columns \n" " q Return back to cacheline list \n" ; +if (!he) +return 0 ; +/* Display compact version first. */ c2c.symbol_full = false;