Mounted xfs slows down block device

Mounted xfs slows down block device

[Jan Engelhardt]

Hi,


I am seeing a really odd interaction between dm-crypt and xfs (not 
happening with other filesystems): when a crypto block device is 
mounted, reading from it slows by a significant amount.

# umount /home
# dd_rescue /dev/mapper/home /dev/null
dd_rescue: (info): ipos:     15360.0k, opos:     15360.0k, xferd:     15360.0k
                   errs:      0, errxfer:         0.0k, succxfer:     15360.0k
             +curr.rate:    19623kB/s, avg.rate:    19623kB/s, avg.load: 15.3%

# mount /home
# dd_rescue /dev/mapper/home /dev/null
dd_rescue: (info): ipos:      5120.0k, opos:      5120.0k, xferd:      5120.0k
                   errs:      0, errxfer:         0.0k, succxfer:      5120.0k
             +curr.rate:     4578kB/s, avg.rate:     4578kB/s, avg.load: 17.0%

/proc/mounts contains:
/dev/mapper/home /home xfs rw,relatime,nobarrier,usrquota,prjquota,grpquota 0 0

xfs_info:
meta-data=/dev/mapper/home       isize=256    agcount=16, agsize=3259186 blks
         =                       sectsz=512   attr=1
data     =                       bsize=4096   blocks=52146976, imaxpct=25
         =                       sunit=0      swidth=0 blks
naming   =version 2              bsize=4096   ascii-ci=0
log      =internal               bsize=4096   blocks=25462, version=2
         =                       sectsz=512   sunit=0 blks, lazy-count=0
realtime =none                   extsz=65536  blocks=0, rtextents=0


Values vary; on my home desktop it goes from 19->4MB/s, and even on
server it goes from 80->31MB/sec. It is independent of any mount options
or the age of the filesystem.

This happens for me on at least 2.6.31-rc6.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs

Re: Mounted xfs slows down block device

[Christoph Hellwig]

Strange, as XFS doesn't actually use the block device mapping at all,
so all the caching doesn't interact with each other.  Maybe some
throtteling code in the VM doesn't like these parallel accesses.

But I need to add that reading the block device on a mounted filesystem
is not a good idea anyay - you will not get any sort of concistency
guarantee.

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs