perf-<pid>.map I generated for jit codes lose effect when hugepage used

perf-<pid>.map I generated for jit codes lose effect when hugepage used

[Zhu, Wen-Jie (Joshua)]

Hello,

I use perf and the perf-<pid>.map I generated for Java jit codes to help profiling java application.
And recently I found jit symbols cannot be listed any more in perf report.
By "sysctl vm.nr_hugepages=0" to disable hugepage temporarily, it works.

Without hugepage, by judging anonymous memory's vm_area_struct, perf_mmap_event's filename will be set to "//anon" indicating this vma is belong to anonymous memory.
Once hugepage is used, vma's filename point to hugetlbfs. In this way, this vma will not be regarded as anonymous memory by is_anon_memory() in perf user space utility.

Therefore in order to avoid this problem, I temporarily change the codes of is_anon_memory() into following:
            return (strcmp(filename, "//anon") && strcmp(filename, "/anon_hugepage (deleted)")) == 0;

I already verify this problem in  kernel 3.5.0-21-generic and RHEL kernel 2.6.32-279.el6.

Do you have any better solution replacement?

Thank you!
Joshua

Re: perf-<pid>.map I generated for jit codes lose effect when hugepage used

[Andi Kleen]

"Zhu, Wen-Jie (Joshua)" <zhu.wen-jie@hp.com> writes:
>
> Therefore in order to avoid this problem, I temporarily change the codes of is_anon_memory() into following:
>             return (strcmp(filename, "//anon") && strcmp(filename, "/anon_hugepage (deleted)")) == 0;
>
> I already verify this problem in  kernel 3.5.0-21-generic and RHEL kernel 2.6.32-279.el6.
>
> Do you have any better solution replacement?


Your fix looks good to me, probably you just need to send it as a more
formal patch.

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only