Re: [PATCH] perf ordered_events: fix crash in free_dup_event()

From: Jiri Olsa
Date: Thu Aug 09 2018 - 03:39:51 EST

On Wed, Aug 08, 2018 at 02:47:42PM -0700, Stephane Eranian wrote:
> Hi,
> Ok, I found the problem. It still exists upstream , just very tricky to trigger.
> Took me lots of time with gdb + watchpoints to track this down, where
> in fact it was just in front of me.
> From the crashdump:
> Program received signal SIGSEGV, Segmentation fault.
> free_dup_event (oe=0x26a39a0, event=0xffffffff00000000)
> I was puzzled by the 0xffffffff00000000. I tracked down where this
> value was coming from using watchpoints.
> In my case the memory was used before by elfutils to back the struct
> Elf. The -1 in the upper bits came from:
> file_read_elf () at third_party/elfutils/libelf/elf_begin.c:451
> elf->[cnt].shndx_index = -1;
> And yet the next access to that memory location was in the crash. That
> meant the memory was released by
> elfutils and reused by ordered_events, yet without any initialization.
> But looking at alloc_event(), it was
> not obvious to figure out how a new_event->event could be uninitialized.
> Well, it turns out there is this little hack where the code
> commandeers the first element in the oe->buffer to
> use as a list_head for the oe->to_free freelist. The problem is that
> this entry also gets freed, but its
> event->event field is NEVER initialized. So depending on how the
> memory was previously used, you
> could get a on NULL value and crash in free_dup_event(). This is what
> happened to me. I am glad
> I pursued this further because the bug is still in the upstream
> version. The patch is a one-liner fixing
> the initialization of the event->event = NULL. For the other elements
> in the list, the initialization is
> already done at the end of alloc_event().
> I will send the patch separately.

nice ;-) thanks