>Number:         144446
>Category:       bin
>Synopsis:       [patch] db(3) fails with large block sizes
>Confidential:   no
>Severity:       non-critical
>Priority:       low
>Responsible:    freebsd-bugs
>State:          open
>Quarter:        
>Keywords:       
>Date-Required:
>Class:          sw-bug
>Submitter-Id:   current-users
>Arrival-Date:   Wed Mar 03 11:30:01 UTC 2010
>Closed-Date:
>Last-Modified:
>Originator:     Peter Jeremy
>Release:        FreeBSD 8.0-STABLE amd64
>Organization:
Alcatel-Lucent Australia
>Environment:
System: FreeBSD server.vk2pj.dyndns.org 8.0-STABLE FreeBSD 8.0-STABLE #1: Wed 
Jan 27 06:55:10 EST 2010 
[email protected]:/var/obj/usr/src/sys/server amd64

>Description:
        Whilst trying to port db(3) to a Solaris system, I have identified
        two issues with the existing hash(3) code.

        Firstly, when creating a new hash database, the bucket size
        defaults to the st_blksize of the file (hash/hash.c::init_hash()).
        There is no sanity checking to ensure that st_blksize is within
        valid limits (hash/hash.h defined MAX_BSIZE as 65536).

        In FreeBSD, st_blksize is currently hardwired to PAGE_SIZE in
        kern/vfs_vnops.c::vn_stat() so this is purely a theoretical
        issue on FreeBSD.  Solaris exposes the blocksize from the
        underlying filesystem - and in the case of ZFS, this is 128KB,
        which exceeds MAX_BSIZE.  In my case, the symptoms were that
        when sequentially reading the database (via DB->seq()), the
        returned keys were padded with 64KB of NULs.

        Secondly, when the bucket size is set to 64KB (MAX_BSIZE),
        non-trivial databases crash reporting:
            "HASH: Out of overflow pages.  Increase page size"
        It's not clear what triggers this.

        Thirdly, whilst writing this PR, I've noticed that hash(3)
        states that the default hash table bucket size is 256 bytes.
        The actual default (as per DEF_BUCKET_SIZE in hash/hash.h)
        is 4096 bytes.

>How-To-Repeat:
        First problem can't be reproduced on FreeBSD but occurs on
        at least Solaris and OpenSolaris with ZFS.

        The second problem can be reproduced by extending the db
        test tool (lib/libc/db/test/run.test, test20) to include
        a bucket size of 65536.  Based on the progression of
        fill factor's in test20, the logical fill factors are
        2735 3647 5471, however the test fails with each of these
        values as well as 8, 341 and 10001.  All other bucket
        sizes appear to work successfully - which suggests there
        is a problem with this particular bucket size.

>Fix:
        First patch adds check for excessive st_blksize.  Note that this
        assumes that st_blksize is a power of 2.  It might be reasonable
        to verify that st_blksize isn't too small but I'm not sure what
        the lower limit is.

Index: hash/hash.c
===================================================================
RCS file: /usr/ncvs/src/lib/libc/db/hash/hash.c,v
retrieving revision 1.21.2.2
diff -u -r1.21.2.2 hash.c
--- hash/hash.c 28 Aug 2009 19:48:06 -0000      1.21.2.2
+++ hash/hash.c 3 Mar 2010 09:39:45 -0000
@@ -293,6 +293,8 @@
                if (stat(file, &statbuf))
                        return (NULL);
                hashp->BSIZE = statbuf.st_blksize;
+               if (hashp->BSIZE > MAX_BSIZE)
+                       hashp->BSIZE = MAX_BSIZE;
                hashp->BSHIFT = __log2(hashp->BSIZE);
        }
 

        This is simply a work-around for the second problem.  I'm not
        certain what the actual problem is bu this avoids it.

Index: hash/hash.h
===================================================================
RCS file: /usr/ncvs/src/lib/libc/db/hash/hash.h,v
retrieving revision 1.9.2.1
diff -u -r1.9.2.1 hash.h
--- hash/hash.h 3 Aug 2009 08:13:06 -0000       1.9.2.1
+++ hash/hash.h 3 Mar 2010 09:39:45 -0000
@@ -118,7 +118,7 @@
 /*
  * Constants
  */
-#define        MAX_BSIZE               65536           /* 2^16 */
+#define        MAX_BSIZE               32768           /* 2^15 but should be 
65536 */
 #define MIN_BUFFERS            6
 #define MINHDRSIZE             512
 #define DEF_BUFSIZE            65536           /* 64 K */


        The following updates the man page to contain the current
        default bucket size as per DEF_BUCKET_SIZE in hash/hash.h

Index: man/hash.3
===================================================================
RCS file: /usr/ncvs/src/lib/libc/db/man/hash.3,v
retrieving revision 1.9.10.1
diff -u -r1.9.10.1 hash.3
--- man/hash.3  3 Aug 2009 08:13:06 -0000       1.9.10.1
+++ man/hash.3  3 Mar 2010 09:50:52 -0000
@@ -78,7 +78,7 @@
 element
 defines the
 .Nm
-table bucket size, and is, by default, 256 bytes.
+table bucket size, and is, by default, 4096 bytes.
 It may be preferable to increase the page size for disk-resident tables
 and tables with large data items.
 .It Va ffactor
>Release-Note:
>Audit-Trail:
>Unformatted:
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-bugs
To unsubscribe, send any mail to "[email protected]"

Reply via email to