Andutt, On Thu, Jan 21, 2010 at 10:16:45AM +0100, andutt wrote: > I have some problem with an XFS filesystem, im running Suse SLES 10 > SP2, with 2.6.16.60-0.42.7-smp kernel untop of a veritas 1Tb volume, a > couple of times per day i get following in my kernel log. > > Badness in xfs_write at fs/xfs/linux-2.6/xfs_lrw.c:838 (I am picking up this old thread for the benefit of the mail archive, whose results I found to be inadequate. Hopefully what follows is sufficient explanation of this warning for the next poor fool. It may even be correct.) I was able to reproduce this (once) by running the following overnight: while true; do dd if=/dev/zero of=file conv=notrunc bs=4k count=1000 oflag=direct; done while true; do dd if=file of=/dev/null bs=1k count=10000 iflag=direct; done while true; do dd if=file of=/dev/null bs=4k count=10000; done while true; do dd if=file of=/dev/null bs=4k count=10000; done while true ; do dd if=file of=/dev/null bs=1k count=1000; done while true; do dd if=file of=/dev/null bs=4k count=10000; done while true ; do dd if=file of=/dev/null bs=1k count=1000; done * 'ish while true ; do dd if=file of=/dev/null bs=1k count=1000; done * 'ish * ish... maybe this is what I ran. It's not in scrollback. I believe what we are seeing is a thin race between the direct write and the buffered readers adding cached pages to the file. It looks to me that the direct write and the readers all take the iolock shared. So there is no exclusion between them when the direct write checks for cached pages in the file to determine whether to take the iolock exclusive to flush them out, and when the actual flushing and WARN_ON occur. The buffered readers can read data from disk into the cache during this window and this causes the WARN_ON to fire. The standard advice given here is: "Don't do direct and buffered io concurrently on the same file." It seems to me that having pages cached during the direct write isn't necessarily a problem. These pages might be at offsets that do not overlap with the direct write, and this would be a perfectly valid usage of concurrent direct and buffered io. I believe the required semantics of an overlapping buffered/direct situation are undefined (I could be horribly wrong). And XFS leaves it up to the application to use adequate locking to prevent overlapping concurrent buffered and direct io. It looks like running flushinval with iolock shared is not a problem here, and there cannot be any dirty pages because the direct write has the iolock shared and buffered writes take it exclusive. I suggest you 1) be certain that your application is doing the Right Thing, and then 2) ignore the warning. It really is up to the application to do the right thing in this area. And the warning doesn't (necessarily) indicate that the application did something wrong, just that it could have. This warning has been removed from xfs in v3.1 (see c58cb165). -Ben _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
