On 01/28/2014 04:55 AM, Dave Chinner wrote: > Ok, so xfs_inode_item_format+0x4a is inside the very first call to > preapre the ilf structure. That tells us that the initial > xfs_log_vec/xfs_log_iovec array are resulting in an unaligned > buffer. > > Can you try the patch below, Michael? > > Cheers, > > Dave. To the best of my knowledge, it works fine. It was subjected to some of the xfstests xfs/* and generic/* series along with some of the "log" group. In addition, it was run through fs_mark and a homebrew benchmark idea of running a ( make clean; make ) loop for the kernel while two fsx processes were running. This was for default (4k, internal logdev) v5-superblock XFS filesystems. v4-superblock XFS went through similar testing, to a lesser degree. I did not know about a message like "log buf needs to be larger than stripe size" on tests that require v2 logs. [xfs/087 might be one of those tests.] This was also the case for the test-patch that padded the structs directly, and this might be an old message. At the very least, I'm keeping this patch. Your approach to the solution looks cool. Basically, you're stating that you can have either xfs_log_vec or xfs_log_iovec in there, and no matter what their size, they get aligned in the log, correct? Thanks! Michael _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs