Horton
Menu

Thread

  1. hash_search and out of memory

    Hitoshi Harada <umi.tanuki@gmail.com> — 2012-10-18T05:42:45Z

    If OOM happens during expand_table() in hash_search_with_hash_value()
for RelationCacheInsert, the hash table entry is allocated and stored
in the hash table, but idhentry->reldesc remains NULL.  Since OOM
causes AbortTransaction(), in AtEOXact_RelationCache() this NULL
pointer is referenced and we hit SIGSEGV.

The fix would be either catch OOM error with PG_TRY() and undo the
hash entry, or do the expansion first and put the entry later.  The
latter is a bit ugly because we have to re-calculate hash bucket after
we decided to expand, so the former looks better.  Do you think of
other solutions?

Thanks,
-- 
Hitoshi Harada

  2. Re: hash_search and out of memory

    Tom Lane <tgl@sss.pgh.pa.us> — 2012-10-18T13:56:29Z

    Hitoshi Harada <umi.tanuki@gmail.com> writes:
> If OOM happens during expand_table() in hash_search_with_hash_value()
> for RelationCacheInsert,

What OOM?  expand_table is supposed to return without doing anything
if it can't expand the table.  If that's not happening, that's a bug
in the hash code.

			regards, tom lane

  3. Re: hash_search and out of memory

    Tom Lane <tgl@sss.pgh.pa.us> — 2012-10-18T15:35:09Z

    I wrote:
> Hitoshi Harada <umi.tanuki@gmail.com> writes:
>> If OOM happens during expand_table() in hash_search_with_hash_value()
>> for RelationCacheInsert,

> What OOM?  expand_table is supposed to return without doing anything
> if it can't expand the table.  If that's not happening, that's a bug
> in the hash code.

Oh, wait, I take that back --- the palloc-based allocator does throw
errors.  I think that when that was designed, we were thinking that
palloc-based hash tables would be thrown away anyway after an error,
but of course that's not true for long-lived tables such as the relcache
hash table.

I'm not terribly comfortable with trying to use a PG_TRY block to catch
an OOM error - there are too many ways that could break, and this code
path is by definition not very testable.  I think moving up the
expand_table action is probably the best bet.  Will you submit a patch?

			regards, tom lane

  4. Re: hash_search and out of memory

    Hitoshi Harada <umi.tanuki@gmail.com> — 2012-10-19T03:31:21Z

    On Thu, Oct 18, 2012 at 8:35 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I wrote:
>> Hitoshi Harada <umi.tanuki@gmail.com> writes:
>>> If OOM happens during expand_table() in hash_search_with_hash_value()
>>> for RelationCacheInsert,
>
> the palloc-based allocator does throw
> errors.  I think that when that was designed, we were thinking that
> palloc-based hash tables would be thrown away anyway after an error,
> but of course that's not true for long-lived tables such as the relcache
> hash table.
>
> I'm not terribly comfortable with trying to use a PG_TRY block to catch
> an OOM error - there are too many ways that could break, and this code
> path is by definition not very testable.  I think moving up the
> expand_table action is probably the best bet.  Will you submit a patch?

Here it is. I factored out the bucket finding code to re-calculate it
after expansion.

Thanks,
-- 
Hitoshi Harada

  5. Re: hash_search and out of memory

    Tom Lane <tgl@sss.pgh.pa.us> — 2012-10-19T18:40:18Z

    Hitoshi Harada <umi.tanuki@gmail.com> writes:
> On Thu, Oct 18, 2012 at 8:35 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I'm not terribly comfortable with trying to use a PG_TRY block to catch
>> an OOM error - there are too many ways that could break, and this code
>> path is by definition not very testable.  I think moving up the
>> expand_table action is probably the best bet.  Will you submit a patch?

> Here it is. I factored out the bucket finding code to re-calculate it
> after expansion.

I didn't like that too much.  I think a better solution is just to do
the table expansion at the very start of the function, along the lines
of the attached patch.  This requires an extra test of the "action"
parameter, but I think that's probably cheaper than an extra function
call.  It's definitely cheaper than recalculating the hash etc when
a split does occur.

			regards, tom lane

  6. Re: hash_search and out of memory

    Hitoshi Harada <umi.tanuki@gmail.com> — 2012-10-19T19:31:41Z

    On Fri, Oct 19, 2012 at 11:40 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Hitoshi Harada <umi.tanuki@gmail.com> writes:
>> On Thu, Oct 18, 2012 at 8:35 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>> I'm not terribly comfortable with trying to use a PG_TRY block to catch
>>> an OOM error - there are too many ways that could break, and this code
>>> path is by definition not very testable.  I think moving up the
>>> expand_table action is probably the best bet.  Will you submit a patch?
>
>> Here it is. I factored out the bucket finding code to re-calculate it
>> after expansion.
>
> I didn't like that too much.  I think a better solution is just to do
> the table expansion at the very start of the function, along the lines
> of the attached patch.  This requires an extra test of the "action"
> parameter, but I think that's probably cheaper than an extra function
> call.  It's definitely cheaper than recalculating the hash etc when
> a split does occur.
>

OK.  Looks better.  But nentries should be bogus a little now?

+ 		/*
+ 		 * Can't split if running in partitioned mode, nor if frozen, nor if
+ 		 * table is the subject of any active hash_seq_search scans.  Strange
+ 		 * order of these tests is to try to check cheaper conditions first.
+ 		 */
+ 		if (!IS_PARTITIONED(hctl) && !hashp->frozen &&
+ 			hctl->nentries / (long) (hctl->max_bucket + 1) >= hctl->ffactor &&
+ 			!has_seq_scans(hashp))
+ 			(void) expand_table(hashp);

hctl->nentries + 1 sounds appropriate?

Thanks,
-- 
Hitoshi Harada

  7. Re: hash_search and out of memory

    Tom Lane <tgl@sss.pgh.pa.us> — 2012-10-19T20:02:52Z

    Hitoshi Harada <umi.tanuki@gmail.com> writes:
> OK.  Looks better.  But nentries should be bogus a little now?

No, I think it's fine as is.  Essentially this logic says "attempt to
split when the new insertion would make us go over the target fill
factor", whereas the old logic split when the just-completed insertion
reached the target.  There's not a lot of reason to care about the
precise boundary anyway, but to the extent that you believe that the
target is exact, I think this behavior is actually a bit more precise.

			regards, tom lane