Bo Ørsted Andresen <boa@neogrid.dk> writes:
>> On 2016-06-07 19:41, Andres Freund wrote:
>> Any chance the running version of postgres is out of date with the installed
>> binaries / debug symbols?
> You mean that I upgraded without restarting postgres before the segfault?
I think the reason for the lack of useful backtrace info is that we've
smashed the stack. Note that the original report shows i == 3324 which is
much larger than the available length of the local items[] array (408).
So presumably, the passed-in "len" was bogus (much too large).
If you're prepared to build a custom version of Postgres, you could
try adding this to _bt_restore_page():
/* Need to copy tuple header due to alignment considerations */ memcpy(&itupdata, from, sizeof(IndexTupleData));
itemsz = IndexTupleDSize(itupdata); itemsz = MAXALIGN(itemsz);
+ if (i >= lengthof(items))
+ elog(PANIC, "too many items on btree page");
+ items[i] = (Item) from; itemsizes[i] = itemsz; i++;
from += itemsz;
and then you should get a core dump before the stack is clobbered.
I wonder whether we shouldn't add such a check to the regular sources...
regards, tom lane