Wireshark-dev: Re: [Wireshark-dev] [Bug 7775] Wireshark leaks memory when selecting packets
On Mon, Oct 08, 2012 at 04:29:15PM -0400, Evan Huus wrote:
> On Mon, Oct 8, 2012 at 3:31 PM, Jakub Zawadzki
> <darkjames-ws@xxxxxxxxxxxx> wrote:
> > Hi Evan,
> >
> > On Mon, Oct 08, 2012 at 03:06:29PM -0400, Evan Huus wrote:
> >> On Mon, Oct 8, 2012 at 2:58 PM, <bugzilla-daemon@xxxxxxxxxxxxx> wrote:
> >> > https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=7775
> >> >
> >> > --- Comment #10 from Anatoly <aries.nah@xxxxxxxxx> 2012-10-08 14:58:26 EDT ---
> >> > I've tried r45395, nothin is changed except wireshark has crashed:)
> >> > But I got one more probably interesting fact: r45380 (w/o your fixes) on Vista
> >> > 32 bit - memory consumption increases slightly and sometimes it decreases, but
> >> > the same revision of XP 32bit - it always increases by ~2MB.
> >>
> >> This appears to be an XP-specific bug if anyone still has an XP
> >> install to test with (I don't).
> >
> > It mighe be OOM, for every packet we mmap 10MiB chunk of memory.
> > Which is *not* unmaped in emem_free_all() nor ep_free_pool().
> > (AFAIR for performance reasons)
> >
> > Generally it'd be best to save somewhere mem->free_list before
> > 1333 g_free(mem);
> >
> > If it's hard to do than emem_destroy_chunk() is your friend.
> >
> > [PS: I haven't test it yet, but looking at diff it makes sense,
> > if not please move on ;)]
>
> I think you're right - I did all of my testing of the emem changes
> under the valgrind-wireshark.sh script, which disables chunks (and
> thus the mmap) so I never noticed it before. I have a patch (attached)
> which I believe ought to correctly share the free_list between the
> various pools, but for some reason I can't seem to track down it's
> causing random canary errors.
>
> Any ideas?
It works with your patch + attached patch.
diff --git a/epan/emem.c b/epan/emem.c
index 12a3070..b506bcf 100644
--- a/epan/emem.c
+++ b/epan/emem.c
@@ -1325,9 +1325,11 @@ ep_create_pool(void)
/* The free_list, specifically when using chunks, does
* not get freed (or unmapped, as the case may be) by
* emem_free_all(), so we share it between the pools. */
- if (ep_packet_mem->debug_use_chunks) {
+ /* We can only share unused chunks, so check if ->free_list->canary_last is NULL */
+ if (ep_packet_mem->debug_use_chunks && ep_packet_mem->free_list && ep_packet_mem->free_list->canary_last == NULL) {
mem->free_list = ep_packet_mem->free_list;
ep_packet_mem->free_list = NULL;
+ g_print("stolen\n");
}
ep_pool_stack = g_slist_prepend(ep_pool_stack, mem);
@@ -1362,8 +1364,14 @@ ep_free_pool(emem_pool_t *mem)
* via ep_create_pool() and so copying NULL over it here would
* just leak memory. */
if (ep_packet_mem->debug_use_chunks && free_list != NULL) {
- g_assert(ep_packet_mem->free_list == NULL);
- ep_packet_mem->free_list = free_list;
+ emem_chunk_t *p = ep_packet_mem->free_list;
+
+ if (p) {
+ while (p->next)
+ p = p->next;
+ p->next = free_list;
+ } else
+ ep_packet_mem->free_list = free_list;
}
}
