lkml.org 
[lkml]   [2019]   [Aug]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subjectperf tool issue following 'perf stat: Fix --no-scale' patch integration
Date
Dear Andi and all perf tool / arm debug  experts

This is about the following patch :
perf stat: Fix --no-scale
SHA-1 : 75998bb263bf48c1c85d78cd2d2f3a97d3747cab

Since it is applied in the kernel, I noticed that perf tool fails on my ARMv7 platform (STM32MP1 with Cortex-A7 and NEON) with the following error :
root@stm32mp1:~# perf stat --no-scale sleep 1
[10827.350202] Alignment trap: perf (631) PC=0x001139e8 Instr=0xf4640adf Address=0x0021a804 1
[10827.357704] Alignment trap: not handling instruction f4640adf at [<001139e8>]
[10827.364867] 8<--- cut here ---
[10827.367875] Unhandled fault: alignment exception (0x001) at 0x0021a804
[10827.374427] pgd = 8abc1568
[10827.377090] [0021a804] *pgd=ff2e8835
Bus error

The same error happens with or without the --no-scale option.
This is to give the context. I do not blame your patch, Andi :)

I analyzed the root cause of this issue, summarized below, but then I need your lights to imagine the best correction.

One of the changes in the patch concerns tools/perf/util/stat.c :
                                case AGGR_GLOBAL:
                               aggr->val += count->val;
-                              if (config->scale) {
-                                              aggr->ena += count->ena;
-                                              aggr->run += count->run;
-                              }
+                             aggr->ena += count->ena;
+                             aggr->run += count->run;

The consequence of this new writing is that GCC generates a NEON vectored instruction to load count->val and count->ena values in 64 bits registers, since they are sequential in memory and systematically initialized now:
                f4640adf              vld1.64  {d16-d17}, [r4 :64]

The problem comes from the ':64' specifying that the parameter has to be 8 bytes aligned.
The 'count' pointer points inside the 'contents[]' array from the 'struct xyarray'.
If I force this field to be 64 bits aligned, then perf works again:
struct xyarray {
                size_t row_size;
                size_t entry_size;
                size_t entries;
                size_t max_x;
                size_t max_y;
-              char contents[] ;
+             char contents[] __attribute__((aligned(64)));
};

But the xyarray structure is generic so I think this patch cannot be the final one.
Some GCC versions have a -mgeneral-regs-only option to forbid the generation of NEON instructions while compiling one file, but this does not seem to be mainlined (?).

Well, I am hesitating and don't know what kind of correction I should apply.
I also don't know very well perf tool source code, so this sets some borders to my imagination  :)

Can you help me please ?

Best regards

Gérald

\
 
 \ /
  Last update: 2019-08-21 16:59    [W:0.032 / U:5.208 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site