[dpdk-dev] [dpdk-stable] [PATCH 2/2] malloc: fix realloc padded element size

Xueming(Steven) Li xuemingl at mellanox.com
Thu Nov 21 15:11:59 CET 2019



> -----Original Message-----
> From: Burakov, Anatoly <anatoly.burakov at intel.com>
> Sent: Thursday, November 21, 2019 9:17 PM
> To: Xueming(Steven) Li <xuemingl at mellanox.com>; David Marchand
> <david.marchand at redhat.com>
> Cc: Asaf Penso <asafp at mellanox.com>; dev <dev at dpdk.org>; dpdk stable
> <stable at dpdk.org>
> Subject: Re: [dpdk-dev] [dpdk-stable] [PATCH 2/2] malloc: fix realloc padded
> element size
> 
> On 21-Nov-19 1:15 PM, Burakov, Anatoly wrote:
> > On 21-Nov-19 12:55 PM, Xueming(Steven) Li wrote:
> >> Hi Anatoly,
> >>
> >>> -----Original Message-----
> >>> From: Burakov, Anatoly <anatoly.burakov at intel.com>
> >>> Sent: Thursday, November 21, 2019 8:30 PM
> >>> To: David Marchand <david.marchand at redhat.com>; Xueming(Steven)
> Li
> >>> <xuemingl at mellanox.com>
> >>> Cc: Asaf Penso <asafp at mellanox.com>; dev <dev at dpdk.org>; dpdk
> stable
> >>> <stable at dpdk.org>
> >>> Subject: Re: [dpdk-stable] [PATCH 2/2] malloc: fix realloc padded
> >>> element size
> >>>
> >>> On 20-Nov-19 1:25 PM, David Marchand wrote:
> >>>> On Wed, Nov 20, 2019 at 3:12 AM Xueming(Steven) Li
> >>>> <xuemingl at mellanox.com> wrote:
> >>>>>
> >>>>>> -----Original Message-----
> >>>>>> From: David Marchand <david.marchand at redhat.com>
> >>>>>> Sent: Wednesday, November 20, 2019 4:47 AM
> >>>>>> To: Xueming(Steven) Li <xuemingl at mellanox.com>; Anatoly
> Burakov
> >>>>>> <anatoly.burakov at intel.com>
> >>>>>> Cc: Asaf Penso <asafp at mellanox.com>; dev <dev at dpdk.org>; dpdk
> >>>>>> stable <stable at dpdk.org>
> >>>>>> Subject: Re: [dpdk-stable] [PATCH 2/2] malloc: fix realloc padded
> >>>>>> element size
> >>>>>>
> >>>>>> On Tue, Nov 12, 2019 at 3:50 PM Xueming Li
> >>>>>> <xuemingl at mellanox.com>
> >>>>>> wrote:
> >>>>>>>
> >>>>>>> When resize a memory with next element, the original element
> >>>>>>> size grows.
> >>>>>>> If the orginal element has padding, the real inner element size
> >>>>>>> didn't grow as well and this causes trailer verification failure
> >>>>>>> when malloc debug enabled.
> >>>>>>
> >>>>>> I did not see this when running the malloc_autotest with debug
> >>>>>> enabled.
> >>>>>> What is missing for me to catch it?
> >>>>>>
> >>>>> Yes, it happens rarely, depends on memory fragment. I only caught
> >>>>> this in
> >>> middle of a long test.
> >>>>>
> >>>>>>
> >>>>>> Just a bit chilly to apply this series.
> >>>>>> The first patch seems an optimisation.
> >>>>>> The second one seems more interesting if we fix the debug mode,
> >>>>>> but I suppose we can live without them in 19.11.
> >>>>> Few people enable memory debug option, they are there for years.
> >>>>
> >>>> Had a discussion offlist with Anatoly.
> >>>> Those two issues are hard to catch but the fixes are relevant and
> >>>> Anatoly is confident.
> >>>> I will take this in rc3.
> >>>>
> >>>> Series applied, thanks.
> >>>>
> >>>
> >>> To test them, i had to modify malloc to always create padded
> >>> elements :)
> >>
> >> I fix another issue in element join, as I made some local enhancement
> >> patch on memory, it's required, not sure whether it help on public code.
> >>
> >> diff --git a/lib/librte_eal/common/malloc_elem.c
> >> b/lib/librte_eal/common/malloc_elem.c
> >> index afacb1813c..c3fa0d1039 100644
> >> --- a/lib/librte_eal/common/malloc_elem.c
> >> +++ b/lib/librte_eal/common/malloc_elem.c
> >> @@ -487,6 +487,10 @@ join_elem(struct malloc_elem *elem1, struct
> >> malloc_elem *elem2)
> >>          else
> >>                  elem1->heap->last = elem1;
> >>          elem1->next = next;
> >> +       if (elem1->pad) {
> >> +               struct malloc_elem *inner = RTE_PTR_ADD(elem1,
> >> elem1->pad);
> >> +               inner->size += elem2->size;
> >> +       }
> >>   }
> >>
> >>
> >
> > This looks like a good change as well - while we /mostly/ join
> > elements when they're free, there is one case where we can join a free
> > element with one already occupied - and /that/ element could be
> > padded, which we currently don't update. So, this patch would help the
> public code as well.

Correct, it happens in rte_realloc.

> 
> (that said, the += should be replaced with explicit calculation of correct size,
> as this joining may be run multiple times, and we don't want incorrect size
> accumulating...)

I'll update and send a new patch.

> 
> >
> >>
> >>>
> >>> --
> >>> Thanks,
> >>> Anatoly
> >
> >
> 
> 
> --
> Thanks,
> Anatoly


More information about the dev mailing list