Re: doPickSplit stack buffer overflow in XLogInsert?

On 03/31/2014 09:08 PM, Robert Haas wrote:
> On Wed, Mar 26, 2014 at 9:45 PM, Peter Geoghegan <pg(at)heroku(dot)com> wrote:
>> On Wed, Nov 27, 2013 at 9:10 AM, Noah Misch <noah(at)leadboat(dot)com> wrote:
>>> The threat is that rounding the read size up to the next MAXALIGN would cross
>>> into an unreadable memory page, resulting in a SIGSEGV. Every palloc chunk
>>> has MAXALIGN'd size under the hood, so the excess read of "toDelete" cannot
>>> cause a SIGSEGV. For a stack variable, it depends on the ABI. I'm not aware
>>> of an ABI where the four bytes past the end of this stack variable could be
>>> unreadable, which is not to claim I'm well-read on the topic. We should fix
>>> this in due course on code hygiene grounds, but I would not back-patch it.
>>
>> Attached patch silences the "Invalid read of size n" complaints of
>> Valgrind. I agree with your general thoughts around backpatching. Note
>> that the patch addresses a distinct complaint from Kevin's, as
>> Valgrind doesn't take issue with the invalid reads past the end of
>> spgxlogPickSplit variables on the stack.
>
> Is the needless zeroing this patch introduces apt to cause a
> performance problem?
>
> This function is actually pretty wacky. If we're stuffing bytes with
> undefined contents into the WAL record, maybe the answer isn't to
> force the contents of those bytes to be defined, but rather to elide
> them from the WAL record.

Agreed. I propose the attached, which removes the MAXALIGNs. It's not
suitable for backpatching, though, as it changes the format of the WAL
record.

- Heikki