On Mon, Dec 19, 2011 at 02:06:37AM +0000, Al Viro wrote: > On Mon, Dec 19, 2011 at 02:03:40AM +0000, Al Viro wrote: > > > OK, I'm definitely missing something. The very first thing > > xfs_file_aio_write_checks() does is > > xfs_rw_ilock(ip, XFS_ILOCK_EXCL); > > which really makes me wonder how the hell does that manage to avoid an > > instant deadlock in case of call via xfs_file_buffered_aio_write() > > where we have: > > struct address_space *mapping = file->f_mapping; > > struct inode *inode = mapping->host; > > struct xfs_inode *ip = XFS_I(inode); > > *iolock = XFS_IOLOCK_EXCL; > > xfs_rw_ilock(ip, *iolock); > > ret = xfs_file_aio_write_checks(file, &pos, &count, new_size, iolock); > > which leads to > > struct inode *inode = file->f_mapping->host; > > struct xfs_inode *ip = XFS_I(inode); > > (IOW, inode and ip are the same as in the caller) followed by > > xfs_rw_ilock(ip, XFS_ILOCK_EXCL); > > and with both xfs_rw_ilock() calls turning into > > mutex_lock(&VFS_I(ip)->i_mutex); > > xfs_ilock(ip, XFS_ILOCK_EXCL); > > we ought to deadlock on that i_mutex. What am I missing and how do we manage > > to survive that? > > Arrrgh... OK, I see... What I missed is that XFS_IOLOCK_EXCL is not > XFS_ILOCK_EXCL. Nice naming, that... Been that way for 15 years. :/ However, the naming makes sense to me - the IO lock is for serialising IO operations on the inode, while the I lock is for serialising metadata operations on the inode. I guess I'm used to it, though, so I'll conceed that it might look strange/confusing to someone who only occassionally looks at the internal XFS locking code.... Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>