[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: Segmentation fault when overwriting a k/v with MDB_RESERVE ?

Simon Majou wrote:

In your previous mail you said they were incompatible, but you provided a
logical reason why (you can't sort the data).
Now I found a case where the reason doesn't apply (you can append without
doing the comparison).


No, you can't.

http://symas.com/mdb/doc/group__mdb.html#ga4fa8573d9236d54687c61827ebf8cac0

"Loading unsorted keys with this flag will cause a MDB_KEYEXIST error."
The key is always checked, because not checking it would allow the DB to become corrupted. Again, with RESERVE, you're not providing anything that can be checked. 


So I thought maybe there was something else I missed.
But ok I understand this is not supported.


For exactly the same reason as I already told you.


On 21 August 2015 at 21:21, Howard Chu <hyc@symas.com <mailto:hyc@symas.com>>
wrote:

    Simon Majou wrote:

        Correction, with the flags MDB_RESERVE|MDB_APPENDDUP


    What part of "DUPSORT and RESERVE are incompatible" do you not understand?


        Here is the stack from gdb :

        Program received signal SIGSEGV, Segmentation fault.
        mdb_node_del (mc=mc@entry=0x7fffffffdb10, ksize=ksize@entry=0) at
        mdb.c:7183
        7183                mp->mp_ptrs[j] = mp->mp_ptrs[i];
        (gdb) bt
        #0  mdb_node_del (mc=mc@entry=0x7fffffffdb10, ksize=ksize@entry=0)
              at mdb.c:7183
        #1  0x00007ffff7bd06e2 in mdb_cursor_put (mc=0x7fffffffdb10,
              key=0x7fffffffd8d0, data=0x7fffffffd8c0, flags=<optimized out>)
              at mdb.c:6698
        #2  0x00007ffff7bd0f8e in mdb_cursor_put (mc=0x7fffffffd980,
              key=0x7fffffffdd70, data=0x7fffffffdda0, flags=<optimized out>)
              at mdb.c:6757
        #3  0x00007ffff7bd29b2 in mdb_put (txn=<optimized out>, dbi=<optimized
        out>,
              key=0x7fffffffdd70, data=0x7fffffffdda0, flags=327680) at mdb.c:8611
        #4  0x0000000000400dea in main (argc=1, argv=0x7fffffffded8)
              at test_lmdb_appenddup.c:59



        On 21 August 2015 at 19:25, Simon Majou <simon@majou.org
        <mailto:simon@majou.org>
        <mailto:simon@majou.org <mailto:simon@majou.org>>> wrote:

             And in the case where I want to append data in a DUPSORT database, is
             there a way to write directly in the persisted page ?
             I tried with the flags MDB_RESERVE|MDB_APPEND but it fails.

             On 18 August 2015 at 15:41, Simon Majou <simon@majou.org
        <mailto:simon@majou.org>
             <mailto:simon@majou.org <mailto:simon@majou.org>>> wrote:

                 Oh ok it makes sense now. Thank you for the heads up

                 On 18 August 2015 at 15:36, Howard Chu <hyc@symas.com
        <mailto:hyc@symas.com>
                 <mailto:hyc@symas.com <mailto:hyc@symas.com>>> wrote:

                     Simon Majou wrote:

                         This is version 0.9.16

                         I used sample-bdb.txt for instance, with the DUP_SORT
        option,
                         and the mdb_put
                         with the MDB_RESERVED flags, and execute it 2 times.
        The first
                         run (creation)
                         is ok, the second (update) gets the seg fault.


                     DUPSORT and RESERVED are incompatible. RESERVED means
        you're not
                     providing a data value, only a length. DUPSORT sorts
        *values* -
                     since you didn't provide a valid value, it probably chased a
                     pointer off into space.


                         I ran it in gdb and got the stack :
                         #0  mdb_cursor_set (mc=mc@entry=0x7fffffffdb40,
                         key=key@entry=0x7fffffffdda0,
                               data=data@entry=0x7fffffffd800,
        op=op@entry=MDB_SET,
                               exactp=exactp@entry=0x7fffffffd7f0) at mdb.c:5867
                         #1  0x00007ffff7bd024c in mdb_cursor_put
        (mc=0x7fffffffdb40,
                               key=0x7fffffffdda0, data=0x7fffffffd8f0,
        flags=32768) at
                         mdb.c:6391
                         #2  0x00007ffff7bd095f in mdb_cursor_put
        (mc=0x7fffffffd9b0,
                               key=0x7fffffffdd90, data=0x7fffffffdda0,
                         flags=<optimized out>)
                               at mdb.c:6780
                         #3  0x00007ffff7bd29b2 in mdb_put (txn=<optimized out>,
                         dbi=<optimized out>,
                               key=0x7fffffffdd90, data=0x7fffffffdda0,
        flags=65536) at
                         mdb.c:8611
                         #4  0x0000000000400ccf in main (argc=1,
        argv=0x7fffffffded8)
                         at sample-mdb.c:43

                         It seems the mp_lower value of the page is set to 0
        instead of
                         a value inside
                         the page.

                         On 18 August 2015 at 13:58, Howard Chu <hyc@symas.com
        <mailto:hyc@symas.com>
                         <mailto:hyc@symas.com <mailto:hyc@symas.com>>
        <mailto:hyc@symas.com <mailto:hyc@symas.com>

                         <mailto:hyc@symas.com <mailto:hyc@symas.com>>>>
                         wrote:

                              Simon Majou wrote:

                                  Hello,

                                  I wonder, why do I get a Segmentation fault
        when I
                         use mdb_put with
                                  MDB_RESERVE on a key that already exists ?


                              Have no idea. Provide some more details and
        maybe we can
                         guess.

                              LMDB version? stack trace from SEGV? What's the
        size of
                         the existing
                              record, what's the size of the new record? Show
        sample
                         code that
                              demonstrates the problem?

                              --
                                 -- Howard Chu
                                 CTO, Symas Corp. http://www.symas.com
                                 Director, Highland Sun
        http://highlandsun.com/hyc/
                                 Chief Architect, OpenLDAP
        http://www.openldap.org/project/







--
  -- Howard Chu
  CTO, Symas Corp.           http://www.symas.com
  Director, Highland Sun     http://highlandsun.com/hyc/
  Chief Architect, OpenLDAP  http://www.openldap.org/project/