Re: Possibly too stringent Assert() in b-tree code

On Mon, Sep 19, 2016 at 1:32 PM, Antonin Houska <ah(at)cybertec(dot)at> wrote:
> I've recently seen this when using 9.6:
>
> #0 0x00007f147892f0c7 in raise () from /lib64/libc.so.6
> #1 0x00007f1478930478 in abort () from /lib64/libc.so.6
> #2 0x00000000009683a1 in ExceptionalCondition (conditionName=0x9f2ef8 "!(((PageHeader) (page))->pd_special >= (__builtin_offsetof (PageHeaderData, pd_linp)))", errorType=0x9f2e8a "FailedAssertion",
> fileName=0x9f2e60 "../../../../src/include/storage/bufpage.h", lineNumber=314) at assert.c:54
> #3 0x00000000004ee3d2 in PageValidateSpecialPointer (page=0x7f14700f3480 "") at ../../../../src/include/storage/bufpage.h:314
> #4 0x00000000004ef913 in _bt_getbuf (rel=0x7f14795a1a50, blkno=11, access=2) at nbtpage.c:629
> #5 0x00000000004eae8c in _bt_split (rel=0x7f14795a1a50, buf=136, cbuf=0, firstright=367, newitemoff=408, newitemsz=16, newitem=0x2608330, newitemonleft=0 '\000') at nbtinsert.c:986
> #6 0x00000000004ea563 in _bt_insertonpg (rel=0x7f14795a1a50, buf=136, cbuf=0, stack=0x26090c8, itup=0x2608330, newitemoff=408, split_only_page=0 '\000') at nbtinsert.c:781
> #7 0x00000000004e954c in _bt_doinsert (rel=0x7f14795a1a50, itup=0x2608330, checkUnique=UNIQUE_CHECK_YES, heapRel=0x25aa0a0) at nbtinsert.c:204
> #8 0x00000000004f3b73 in btinsert (rel=0x7f14795a1a50, values=0x7ffe46c6f7f0, isnull=0x7ffe46c6f7d0 "", ht_ctid=0x260927c, heapRel=0x25aa0a0, checkUnique=UNIQUE_CHECK_YES) at nbtree.c:279
> #9 0x00000000004e7964 in index_insert (indexRelation=0x7f14795a1a50, values=0x7ffe46c6f7f0, isnull=0x7ffe46c6f7d0 "", heap_t_ctid=0x260927c, heapRelation=0x25aa0a0, checkUnique=UNIQUE_CHECK_YES)
> at indexam.c:204
>
> Of course, the database could have been corrupted after having encountered
> many crashes during my experiments. Neverthelesss, even without in-depth
> knowledge of the b-tree code I suspect that this stack trace might reflect a
> legal situation. In partcular, if _bt_page_recyclable() returned on this
> condition:
>
> if (PageIsNew(page))
> return true;
>

I think you have a valid point. It seems we don't need to write WAL
for reuse page (aka don't call _bt_log_reuse_page()), if the page is
new, as the only purpose of that log is to handle conflict based on
transaction id stored in special area which will be anyway zero.

--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com