I ran an ALTER TABLE ... ADD FULLTEXT (...) statement on a ~900M MyISAM table in MariaDB 5.5.28a.
Issuing a "KILL <id>" via the MySQL client resulted in a "*** glibc detected *** mysqld free() invalid pointer" trace in the syslog and mysqld got signal 6.
Additionally, mysqld hung trying to gather a stack trace of its own and needed a sig9, which made it a bit nasty. Nothing except the generic sig6 message in the error log.
I havn't been able to reproduce it yet. I'm filing this bug in case someone more informed can suggest how to repeat it other than using brute force and luck... 
*** glibc detected *** /home/db/sandboxes/5.5.28/bin/mysqld: free(): invalid pointer: 0x0ae82e20 ***
|
======= Backtrace: =========
|
/lib/tls/i686/cmov/libc.so.6(+0x6b591)[0xb7504591]
|
/lib/tls/i686/cmov/libc.so.6(+0x6cde8)[0xb7505de8]
|
/lib/tls/i686/cmov/libc.so.6(cfree+0x6d)[0xb7508ecd]
|
/home/db/sandboxes/5.5.28/bin/mysqld(my_free+0x1d)[0x875872d]
|
/home/db/sandboxes/5.5.28/bin/mysqld[0x86843fb]
|
/home/db/sandboxes/5.5.28/bin/mysqld[0x8675e51]
|
/home/db/sandboxes/5.5.28/bin/mysqld[0x8675ff2]
|
/home/db/sandboxes/5.5.28/bin/mysqld[0x867400e]
|
/home/db/sandboxes/5.5.28/bin/mysqld[0x823563c]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z17mysql_alter_tableP3THDPcS1_P24st_ha_create_informationP10TABLE_LISTP10Alter_infojP8st_orderbb+0x237b)[0x823e3fb]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_ZN21Alter_table_statement7executeEP3THD+0x56f)[0x847a1ef]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z21mysql_execute_commandP3THD+0x727)[0x81bd8e7]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z11mysql_parseP3THDPcjP12Parser_state+0x1bb)[0x81c58fb]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z16dispatch_command19enum_server_commandP3THDPcj+0x1beb)[0x81c74eb]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z10do_commandP3THD+0xd8)[0x81c7a18]
|
/home/db/sandboxes/5.5.28/bin/mysqld(_Z24do_handle_one_connectionP3THD+0x154)[0x8287f44]
|
/home/db/sandboxes/5.5.28/bin/mysqld(handle_one_connection+0x3c)[0x828808c]
|
/lib/tls/i686/cmov/libpthread.so.0(+0x596e)[0xb775496e]
|
/lib/tls/i686/cmov/libc.so.6(clone+0x5e)[0xb7566a4e]
|
Using objdump on the mysqld binary and the return addresses in the trace, I think the problem is mi_repair_by_sort() blindly freeing sort_info pointers after some lower layer detected the kill, specifically here in the 5.5.28a source:
mi_check.c:2557 my_free(sort_info.key_block);
|
Objdump excerpt:
86843cf: 0f 84 a8 00 00 00 je 868447d <mi_repair_by_sort+0x57d>
|
86843d5: 8b 85 68 f9 ff ff mov -0x698(%ebp),%eax
|
86843db: 31 c9 xor %ecx,%ecx
|
86843dd: 8d 50 e8 lea -0x18(%eax),%edx
|
86843e0: 85 c0 test %eax,%eax
|
86843e2: 0f 45 ca cmovne %edx,%ecx
|
86843e5: 89 0c 24 mov %ecx,(%esp)
|
86843e8: e8 23 43 0d 00 call 8758710 <my_free>
|
86843ed: 8b 85 7c ff ff ff mov -0x84(%ebp),%eax
|
86843f3: 89 04 24 mov %eax,(%esp)
|
86843f6: e8 15 43 0d 00 call 8758710 <my_free> <<< HERE
|
86843fb: 8b 45 84 mov -0x7c(%ebp),%eax
|
86843fe: 89 04 24 mov %eax,(%esp)
|
8684401: e8 0a 43 0d 00 call 8758710 <my_free>
|
8684406: 8b 85 78 ff ff ff mov -0x88(%ebp),%eax
|
868440c: 89 04 24 mov %eax,(%esp)
|
868440f: e8 fc 42 0d 00 call 8758710 <my_free>
|
8684414: 8b 8d 54 f6 ff ff mov -0x9ac(%ebp),%ecx
|
868441a: 89 0c 24 mov %ecx,(%esp)
|
868441d: e8 3e 05 0c 00 call 8744960 <end_io_cache>
|
8684422: 8b 9d 60 f6 ff ff mov -0x9a0(%ebp),%ebx
|
Or rather, some lower layer is leaving sort_info in a bad state after a kill.
Obviously this is MyISAM and the FULLTEXT focus may be on Aria now, but fwiw the code in ma_check.c and maria_repair_by_sort() looks very similar.