[Date Prev][Date Next]
[Chronological]
[Thread]
[Top]
Re: LMDB: issue with mdb_cursor_del
Hi, Timur.
There two different cases: with MDB_WRITEMAP, and without it.
1) With MDB_WRITEMAP the mdb_mid2l_append() will be used.
Internally mdb_mid2l_append() don't check for duplicated, but just
append a given page number to the list.
Therefore when database opens with MDB_WRITEMAP this bug leads to
duplicates inside the dirty-list and then (seems) to database
corruption.
2) Without MDB_WRITEMAP the mdb_mid2l_insert() will be used.
Internally mdb_mid2l_insert() made insertion into a sorted list, so it
checks for duplicates and returns -1 for such case.
Therefore when database opens without MDB_WRITEMAP this bug leads only
to assertion failure, or nothing if assertions checking was disabled.
Regards,
Leonid.
2017-10-19 19:11 GMT+03:00 <timur.kristof@gmail.com>:
> On Thu, 2017-10-19 at 08:12 +0200, Hallvard Breien Furuseth wrote:
>> On 18/10/17 22:44, timur.kristof@gmail.com wrote:
>> > On Wed, 2017-10-18 at 20:08 +0300, Леонид Юрьев wrote:
>> > > Sure this will not create problems when working without
>> > > MDB_WRITEMAP.
>> > > But in MDB_WRITEMAP mode duplicates in the dirty-list seems may
>> > > leads
>> > > some side affects.
>> > > Therefore in MDBX I had planned to completely rewrite
>> > > freelist_save().
>> >
>> > I think simply accepting -1 would fix this problem for everyone but
>> > users of MDB_WRITEMAP in which case it doesn't affect the outcome
>> > because a different function is called: mdb_mid2l_append which
>> > cannot
>> > return -1.
>>
>> No, Leo is right. Duplicates in the freelist give a broken DB.
>>
>
> I didn't suggest duplicates. I suggested accepting -1 from insert()
> which merely indicated that the page is already on the dirty list. I
> does not make a duplicate.
>
> Or am I getting it wrong?
>
>> A cleaner fix is for freelist_save() to quit messing with loose
>> pages and just reserve room for them, like it does for me_pghead.
>>
>> I've reported ITS#8756 <http://www.openldap.org/its/?findid=8756>.
>>
>> --
>> Hallvard