This is Sleepycat bug #14908. I thought I posted it here before but apparently not. Fixed in current 4.5.20, but still broken in all earlier releases.
-------- Original Message -------- Subject: region size bug Re: [BDB-Alpha] Berkeley DB 4.5.8 ALPHA Date: Mon, 10 Jul 2006 13:37:33 -0700 From: Howard Chu hyc@symas.com To: support@sleepycat.com CC: support@symas.com References: 45A742B5-7DD5-4512-A204-A10FE8FC5DFC@oracle.com
I just ran into this in 4.2.52 but the same calculation occurs in 4.4 and 4.5.8 alpha:
This computation gives the wrong results when the number of cache regions is greater than the number of gigabytes (which we encounter on Linux using shared memory regions, which are constrained to much smaller than a gigabyte each).
in mp/mp_region.c:
roff_t reg_size;
/* Figure out how big each cache region is. */ reg_size = (roff_t)(dbenv->mp_gbytes / dbenv->mp_ncache) * GIGABYTE; reg_size += ((roff_t)(dbenv->mp_gbytes % dbenv->mp_ncache) * GIGABYTE) / dbenv->mp_ncache; reg_size += dbenv->mp_bytes / dbenv->mp_ncache; *reg_sizep = reg_size;
The first reg_size calculation always goes to zero when mp_ncache > mp_gbytes. This should have been, instead: reg_size = GIGABYTE / dbenv->mp_ncache * dbenv->mp_gbytes;
Or in unidiff: --- mp_region.c.O 2006-06-19 08:37:58.000000000 -0700 +++ mp_region.c 2006-07-10 13:36:43.000000000 -0700 @@ -260,9 +260,7 @@ roff_t reg_size;
/* Figure out how big each cache region is. */ - reg_size = (roff_t)(dbenv->mp_gbytes / dbenv->mp_ncache) * GIGABYTE; - reg_size += ((roff_t)(dbenv->mp_gbytes % - dbenv->mp_ncache) * GIGABYTE) / dbenv->mp_ncache; + reg_size = GIGABYTE / dbenv->mp_ncache * dbenv->mp_gbytes; reg_size += dbenv->mp_bytes / dbenv->mp_ncache; *reg_sizep = reg_size;