"perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

"perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

[Alexey Brodkin]

Hi Adrian,

Just noticed that starting from Linux v4.2-rc1 "perf record"
doesn't work on ARC. That's what I see:
 ------------>8------------
 # perf record ls -la
 Cannot use AUX area tracing mmaps
 failed to mmap with 38 (Function not implemented)
 ------------>8------------

I believe that happens because by default auxtrace is enabled
(NO_AUXTRACE=0) and so auxtrace_mmap__mmap() from
"tools/perf/util/auxtrace.c" gets compiled in and following
check fails:
 ------------>8------------
 #if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
 	pr_err("Cannot use AUX area tracing mmaps\n");
 	return -1;
 #endif
 ------------>8------------

Unfortunately we don't have __sync_val_compare_and_swap()
in our current toolchain. And ARC as of today is 32-bit architecture.

Now if I pass NO_AUXTRACE=1 in perf building command then everything
works as expected.

So I'm wondering what would be the best way to get perf properly
working for ARC (at least) and probably other architecture/toolchain
combos?

I see at least 2 options:

 [1] Add feature check for BITS_PER_LONG. And then if
     "BITS_PER_LONG != 32 & HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT"
     automatically set "NO_AUXTRACE=1" on building perf.

 [2] By default enable auxtrace only on selected architectures.
     For example Intel 64-bit and maybe others who really capable of
     running auxtrace (and may make use of it).

Any suggestions are much appreciated.

Just in case that's my back-trace to auxtrace_mmap__mmap():
 ------------>8------------
#0  auxtrace_mmap__mmap (mm=0x11a1a44, mp=0x5fa468c0, userpg=0x2021e000, fd=4)
    at util/auxtrace.c:57
#1  0x0005d3b0 in __perf_evlist__mmap (evlist=0x11a0f30, evlist=0x11a0f30, 
    fd=4, mp=0x5fa468b8, idx=0) at util/evlist.c:838
#2  perf_evlist__mmap_per_evsel (evlist=evlist@entry=0x11a0f30, 
    idx=idx@entry=0, mp=mp@entry=0x5fa468b8, cpu=cpu@entry=0, 
    thread=thread@entry=0, output=output@entry=0x5fa468b4) at util/evlist.c:861
#3  0x0005de1e in perf_evlist__mmap_per_cpu (mp=0x5fa468b8, evlist=0x11a0f30)
    at util/evlist.c:910
#4  perf_evlist__mmap_ex (evlist=0x11a0f30, evlist@entry=0x0, 
    pages=<optimized out>, overwrite=overwrite@entry=false, auxtrace_pages=0, 
    auxtrace_overwrite=<optimized out>) at util/evlist.c:1095
#5  0x0001e906 in record__open (rec=0xf11a0 <record>) at builtin-record.c:305
#6  __cmd_record (rec=0xf11a0 <record>, argv=<optimized out>, 
    argc=<optimized out>) at builtin-record.c:519
#7  cmd_record (argc=<optimized out>, argv=<optimized out>, 
    prefix=<optimized out>) at builtin-record.c:1168
#8  0x00053758 in run_builtin (p=p@entry=0xf2938 <commands+72>, 
    argc=argc@entry=2, argv=argv@entry=0x11a1730) at perf.c:370
#9  0x00053f10 in handle_internal_command (argv=0x11a1730, argc=2)
    at perf.c:429
#10 run_argv (argv=0x5fa47c70, argcp=0x5fa47c74) at perf.c:473
#11 main (argc=2, argv=0x5fa47d98) at perf.c:588
 ------------>8------------

-Alexey

Re: "perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

[Adrian Hunter]

On 14/07/15 15:12, Alexey Brodkin wrote:
> Hi Adrian,
> 
> Just noticed that starting from Linux v4.2-rc1 "perf record"
> doesn't work on ARC. That's what I see:
>  ------------>8------------
>  # perf record ls -la
>  Cannot use AUX area tracing mmaps
>  failed to mmap with 38 (Function not implemented)
>  ------------>8------------
> 
> I believe that happens because by default auxtrace is enabled
> (NO_AUXTRACE=0) and so auxtrace_mmap__mmap() from
> "tools/perf/util/auxtrace.c" gets compiled in and following
> check fails:
>  ------------>8------------
>  #if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
>  	pr_err("Cannot use AUX area tracing mmaps\n");
>  	return -1;
>  #endif
>  ------------>8------------
> 
> Unfortunately we don't have __sync_val_compare_and_swap()
> in our current toolchain. And ARC as of today is 32-bit architecture.
> 
> Now if I pass NO_AUXTRACE=1 in perf building command then everything
> works as expected.
> 
> So I'm wondering what would be the best way to get perf properly
> working for ARC (at least) and probably other architecture/toolchain
> combos?
> 
> I see at least 2 options:
> 
>  [1] Add feature check for BITS_PER_LONG. And then if
>      "BITS_PER_LONG != 32 & HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT"
>      automatically set "NO_AUXTRACE=1" on building perf.
> 
>  [2] By default enable auxtrace only on selected architectures.
>      For example Intel 64-bit and maybe others who really capable of
>      running auxtrace (and may make use of it).
> 
> Any suggestions are much appreciated.
> 
> Just in case that's my back-trace to auxtrace_mmap__mmap():
>  ------------>8------------
> #0  auxtrace_mmap__mmap (mm=0x11a1a44, mp=0x5fa468c0, userpg=0x2021e000, fd=4)
>     at util/auxtrace.c:57
> #1  0x0005d3b0 in __perf_evlist__mmap (evlist=0x11a0f30, evlist=0x11a0f30, 
>     fd=4, mp=0x5fa468b8, idx=0) at util/evlist.c:838
> #2  perf_evlist__mmap_per_evsel (evlist=evlist@entry=0x11a0f30, 
>     idx=idx@entry=0, mp=mp@entry=0x5fa468b8, cpu=cpu@entry=0, 
>     thread=thread@entry=0, output=output@entry=0x5fa468b4) at util/evlist.c:861
> #3  0x0005de1e in perf_evlist__mmap_per_cpu (mp=0x5fa468b8, evlist=0x11a0f30)
>     at util/evlist.c:910
> #4  perf_evlist__mmap_ex (evlist=0x11a0f30, evlist@entry=0x0, 
>     pages=<optimized out>, overwrite=overwrite@entry=false, auxtrace_pages=0, 
>     auxtrace_overwrite=<optimized out>) at util/evlist.c:1095
> #5  0x0001e906 in record__open (rec=0xf11a0 <record>) at builtin-record.c:305
> #6  __cmd_record (rec=0xf11a0 <record>, argv=<optimized out>, 
>     argc=<optimized out>) at builtin-record.c:519
> #7  cmd_record (argc=<optimized out>, argv=<optimized out>, 
>     prefix=<optimized out>) at builtin-record.c:1168
> #8  0x00053758 in run_builtin (p=p@entry=0xf2938 <commands+72>, 
>     argc=argc@entry=2, argv=argv@entry=0x11a1730) at perf.c:370
> #9  0x00053f10 in handle_internal_command (argv=0x11a1730, argc=2)
>     at perf.c:429
> #10 run_argv (argv=0x5fa47c70, argcp=0x5fa47c74) at perf.c:473
> #11 main (argc=2, argv=0x5fa47d98) at perf.c:588
>  ------------>8------------
> 

How about this:

From: Adrian Hunter <adrian.hunter@intel.com>
Date: Tue, 14 Jul 2015 15:32:41 +0300
Subject: [PATCH] perf tools: Fix misplaced check for
 HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT

Move the checking for HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT
for AUX area mmaps until after checking if such mmaps are
used anyway.

Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
---
 tools/perf/util/auxtrace.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/tools/perf/util/auxtrace.c b/tools/perf/util/auxtrace.c
index 7e7405c9b936..83d9dd96fe08 100644
--- a/tools/perf/util/auxtrace.c
+++ b/tools/perf/util/auxtrace.c
@@ -53,11 +53,6 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
 {
 	struct perf_event_mmap_page *pc = userpg;
 
-#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
-	pr_err("Cannot use AUX area tracing mmaps\n");
-	return -1;
-#endif
-
 	WARN_ONCE(mm->base, "Uninitialized auxtrace_mmap\n");
 
 	mm->userpg = userpg;
@@ -73,6 +68,11 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
 		return 0;
 	}
 
+#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
+	pr_err("Cannot use AUX area tracing mmaps\n");
+	return -1;
+#endif
+
 	pc->aux_offset = mp->offset;
 	pc->aux_size = mp->len;
 
-- 
1.9.1

Re: "perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

[Alexey Brodkin]

Hi Adrian,

On Tue, 2015-07-14 at 15:36 +0300, Adrian Hunter wrote:
> On 14/07/15 15:12, Alexey Brodkin wrote:
> How about this:
> 
> From: Adrian Hunter <adrian.hunter@intel.com>
> Date: Tue, 14 Jul 2015 15:32:41 +0300
> Subject: [PATCH] perf tools: Fix misplaced check for
>  HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT
> 
> Move the checking for HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT
> for AUX area mmaps until after checking if such mmaps are
> used anyway.
> 
> Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
> ---
>  tools/perf/util/auxtrace.c | 10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/tools/perf/util/auxtrace.c b/tools/perf/util/auxtrace.c
> index 7e7405c9b936..83d9dd96fe08 100644
> --- a/tools/perf/util/auxtrace.c
> +++ b/tools/perf/util/auxtrace.c
> @@ -53,11 +53,6 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
>  {
>  	struct perf_event_mmap_page *pc = userpg;
>  
> -#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
> -	pr_err("Cannot use AUX area tracing mmaps\n");
> -	return -1;
> -#endif
> -
>  	WARN_ONCE(mm->base, "Uninitialized auxtrace_mmap\n");
>  
>  	mm->userpg = userpg;
> @@ -73,6 +68,11 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
>  		return 0;
>  	}
>  
> +#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
> +	pr_err("Cannot use AUX area tracing mmaps\n");
> +	return -1;
> +#endif
> +
>  	pc->aux_offset = mp->offset;
>  	pc->aux_size = mp->len;

That was really fast!

And indeed fixes reported problem with "record".
 --------------------------->8-------------------------
# perf record ls -la
total 32
drwx------    2 root     root           140 Jan  1 00:00 .
drwxrwxr-x   18 root     root           420 Jul 14  2015 ..
-rw-------    1 root     root            19 Jan  1 00:00 .ash_history
-rw-rw-r--    1 root     root             0 Jul 14  2015 .bash_history
-rw-rw-r--    1 root     root           175 Jul 14  2015 .bash_logout
-rw-rw-r--    1 root     root            78 Jul 14  2015 .bash_profile
-rw-------    1 root     root           312 Jan  1 00:00 perf.data
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.059 MB perf.data (1528 samples) ]

# perf report
# To display the perf.data header info, please use --header/--header-only options.
#
#
# Total Lost Samples: 0
#
# Samples: 1K of event 'cycles'
# Event count (approx.): 2902149
#
# Overhead  Command  Shared Object            Symbol                                
# ........  .......  .......................  ......................................
#
     7.37%  ls       libuClibc-0.9.34-git.so  [.] _ppfs_parsespec                   
     4.40%  ls       [kernel.kallsyms]        [k] memset                            
     3.96%  ls       [kernel.kallsyms]        [k] aligndestination                  
     3.74%  ls       ld-uClibc-0.9.34-git.so  [.] _dl_find_hash                     
     2.34%  ls       [kernel.kallsyms]        [k] perf_event_exec                   
     2.28%  ls       [kernel.kallsyms]        [k] flush_signal_handlers             
     2.28%  ls       [kernel.kallsyms]        [k] _raw_spin_unlock_irqrestore       
     1.77%  ls       libuClibc-0.9.34-git.so  [.] __stdio_fwrite                    
     1.43%  ls       [kernel.kallsyms]        [k] copy_page_to_iter                 
     1.36%  ls       libuClibc-0.9.34-git.so  [.] _ppfs_init                        
     1.22%  ls       [kernel.kallsyms]        [k] filemap_map_pages 
 --------------------------->8-------------------------

Thanks a lot for your prompt response.
Feel free to add my Tested-by then.

-Alexey

Re: "perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

[Adrian Hunter]

On 14/07/15 15:52, Alexey Brodkin wrote:
> Hi Adrian,
> 
> On Tue, 2015-07-14 at 15:36 +0300, Adrian Hunter wrote:
>> On 14/07/15 15:12, Alexey Brodkin wrote:
>> How about this:
>>
>> From: Adrian Hunter <adrian.hunter@intel.com>
>> Date: Tue, 14 Jul 2015 15:32:41 +0300
>> Subject: [PATCH] perf tools: Fix misplaced check for
>>  HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT
>>
>> Move the checking for HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT
>> for AUX area mmaps until after checking if such mmaps are
>> used anyway.
>>
>> Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
>> ---
>>  tools/perf/util/auxtrace.c | 10 +++++-----
>>  1 file changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/tools/perf/util/auxtrace.c b/tools/perf/util/auxtrace.c
>> index 7e7405c9b936..83d9dd96fe08 100644
>> --- a/tools/perf/util/auxtrace.c
>> +++ b/tools/perf/util/auxtrace.c
>> @@ -53,11 +53,6 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
>>  {
>>  	struct perf_event_mmap_page *pc = userpg;
>>  
>> -#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
>> -	pr_err("Cannot use AUX area tracing mmaps\n");
>> -	return -1;
>> -#endif
>> -
>>  	WARN_ONCE(mm->base, "Uninitialized auxtrace_mmap\n");
>>  
>>  	mm->userpg = userpg;
>> @@ -73,6 +68,11 @@ int auxtrace_mmap__mmap(struct auxtrace_mmap *mm,
>>  		return 0;
>>  	}
>>  
>> +#if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)
>> +	pr_err("Cannot use AUX area tracing mmaps\n");
>> +	return -1;
>> +#endif
>> +
>>  	pc->aux_offset = mp->offset;
>>  	pc->aux_size = mp->len;
> 
> That was really fast!
> 
> And indeed fixes reported problem with "record".
>  --------------------------->8-------------------------
> # perf record ls -la
> total 32
> drwx------    2 root     root           140 Jan  1 00:00 .
> drwxrwxr-x   18 root     root           420 Jul 14  2015 ..
> -rw-------    1 root     root            19 Jan  1 00:00 .ash_history
> -rw-rw-r--    1 root     root             0 Jul 14  2015 .bash_history
> -rw-rw-r--    1 root     root           175 Jul 14  2015 .bash_logout
> -rw-rw-r--    1 root     root            78 Jul 14  2015 .bash_profile
> -rw-------    1 root     root           312 Jan  1 00:00 perf.data
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.059 MB perf.data (1528 samples) ]
> 
> # perf report
> # To display the perf.data header info, please use --header/--header-only options.
> #
> #
> # Total Lost Samples: 0
> #
> # Samples: 1K of event 'cycles'
> # Event count (approx.): 2902149
> #
> # Overhead  Command  Shared Object            Symbol                                
> # ........  .......  .......................  ......................................
> #
>      7.37%  ls       libuClibc-0.9.34-git.so  [.] _ppfs_parsespec                   
>      4.40%  ls       [kernel.kallsyms]        [k] memset                            
>      3.96%  ls       [kernel.kallsyms]        [k] aligndestination                  
>      3.74%  ls       ld-uClibc-0.9.34-git.so  [.] _dl_find_hash                     
>      2.34%  ls       [kernel.kallsyms]        [k] perf_event_exec                   
>      2.28%  ls       [kernel.kallsyms]        [k] flush_signal_handlers             
>      2.28%  ls       [kernel.kallsyms]        [k] _raw_spin_unlock_irqrestore       
>      1.77%  ls       libuClibc-0.9.34-git.so  [.] __stdio_fwrite                    
>      1.43%  ls       [kernel.kallsyms]        [k] copy_page_to_iter                 
>      1.36%  ls       libuClibc-0.9.34-git.so  [.] _ppfs_init                        
>      1.22%  ls       [kernel.kallsyms]        [k] filemap_map_pages 
>  --------------------------->8-------------------------
> 
> Thanks a lot for your prompt response.
> Feel free to add my Tested-by then.

+Arnaldo

Here is one for the urgent queue.

Re: "perf record" if BITS_PER_LONG != 64 && !defined(HAVE_SYNC_COMPARE_AND_SWAP_SUPPORT)

[Arnaldo Carvalho de Melo]

Em Tue, Jul 14, 2015 at 04:02:45PM +0300, Adrian Hunter escreveu:
> On 14/07/15 15:52, Alexey Brodkin wrote:
> > Thanks a lot for your prompt response.
> > Feel free to add my Tested-by then.
 
> +Arnaldo
 
> Here is one for the urgent queue.

Applied, thanks!

- Arnaldo