Created attachment 100976 [details]
Patch to add valgrind support for perf events ioctls.

Here's a patched I whipped up that seems to work fine. This is my first patch to valgrind, please let me know if I should do anything else.

Keno, thank you for the patch.   It looks OK, apart from this fragment

+   case VKI_PERF_EVENT_IOC_SET_FILTER: {
+      char *filter = (char *)ARG3;
+      PRE_MEM_READ("ioctl(VKI_PERF_EVENT_IOC_SET_FILTER)",
+        (Addr)ARG3, VG_(strnlen)(filter, VKI_PAGE_SIZE)+1);
+      break;
+   }

If I interpret this correctly, ARG3 is a C style zero-terminated string that the syscall
reads.  Is that correct?

If so, there's already a macro for that: PRE_MEM_RASCIIZ.  I think you want
something like  PRE_MEM_RASCIIZ(ioctl(VKI_PERF_EVENT_IOC_SET_FILTER).filter", ARG3).

Can you redo the patch and re-test it?  Thanks.

> If I interpret this correctly, ARG3 is a C style zero-terminated string that the syscall
> reads.  Is that correct?

Essentially yes, though with the caveat that if there's no NUL after PAGE_SIZE-1 bytes, it'll still accept that as far as I know (as a PAGE_SIZE-1 sized string), since it uses strndup_user(str, PAGE_SIZE) on the kernel side of things.

(In reply to Keno Fischer from comment #3)
> > If I interpret this correctly, ARG3 is a C style zero-terminated string that the syscall
> > reads.  Is that correct?
> 
> Essentially yes, though with the caveat that if there's no NUL after
> PAGE_SIZE-1 bytes, it'll still accept that as far as I know (as a
> PAGE_SIZE-1 sized string), since it uses strndup_user(str, PAGE_SIZE) on the
> kernel side of things.

That seems a funny corner case that I cannot imagine anybody relies on (the caller would have to explicitly not pass a zero terminated string larger than PAGE_SIZE, which is different on different arches). If someone does I think a warning from valgrind seems somewhat justified. So just using PRE_MEM_RASCIIZ here should be fine.

Created attachment 101305 [details]
Updated patch.

Committed, r16077.  Thanks for the patch.