Bug 95321

We are seeing failures due to slow disk writes in our video capture application on multiple systems.  The drive occasionally stalls on disk writes for 6-7 sec. before resuming.  This delay can also be seen by running 'dd', not just via our application.  The occurrences of the stall varies; on some systems it reliably occurs once immediately after booting, but on other systems is more itinerant.

The filesystems use LUKS encryption.  We have only seen this problem on systems updated from Fedora 15 to Fedora 20 with the filesystems in question having encryption added in-place via luksipc.  The same systems when re-imaged with Fedora 20 and LUKS do not exhibit the write stalls.

strace results of dd show that the write() command blocks for 6-7 seconds after 8MB of writes (regardless of the block size and number of write() calls); see attachment.

latencytop run during the dd write stall shows the block occurring either in an open() call (if the file being written did not exist) or trunc() (if the file did exist).

Alt-SysRq-w run during the dd write stall shows calls including vfs_write and related write functions; see attachment.

We request assistance in interpreting these results and/or running further tests to find the root cause of and solution for the write delay.

blkid:
/dev/sda8: UUID="12768db1-3446-4735-bb90-7a2a2aba17e5" TYPE="crypto_LUKS" PARTUUID="a89f101c-08" 
/dev/mapper/sda8: UUID="5448129e-7196-4f19-a79d-61ae7155d71d" TYPE="ext3" 

df:
Filesystem              1K-blocks      Used Available Use% Mounted on
/dev/mapper/sda8        756684360 466905176 251335120  66% /video/cases