Re: [xfs-masters] Re: [BUG] Lockdep warning with XFS on 2.6.22-rc6
Patch looks good, Dave. (though, I stuffed up reviewing that bit of code previously:-) Oh, previous typo: s/inodes at the some time/inodes at the same time/ --Tim David Chinner wrote: On Tue, Jun 26, 2007 at 11:35:20AM +0200, Jarek Poplawski wrote: On 26-06-2007 04:16, David Chinner wrote: It does both - parent-first/child-second and ascending inode # order, which is where the problem is. standing alone, these seem fine, but they don't appear to work when the child has a lower inode number than the parent. ... >From xfs_inode.h: /* * Flags for lockdep annotations. * * XFS_I[O]LOCK_PARENT - for operations that require locking two inodes * (ie directory operations that require locking a directory inode and * an entry inode). The first inode gets locked with this flag so it * gets a lockdep subclass of 1 and the second lock will have a lockdep * subclass of 0. * * XFS_I[O]LOCK_INUMORDER - for locking several inodes at the some time * with xfs_lock_inodes(). This flag is used as the starting subclass * and each subsequent lock acquired will increment the subclass by one. * So the first lock acquired will have a lockdep subclass of 2, the * second lock will have a lockdep subclass of 3, and so on. */ I don't know xfs code, and probably miss something, but it seems there could be some inconsistency: lockdep warning shows mr_lock/1 taken both before and after mr_lock (i.e. /0). According to the above comment there should be always 1 before 0... That just fired some rusty neurons. #define XFS_IOLOCK_SHIFT16 #define XFS_IOLOCK_PARENT (1 << XFS_IOLOCK_SHIFT) #define XFS_IOLOCK_INUMORDER(2 << XFS_IOLOCK_SHIFT) #define XFS_ILOCK_SHIFT 24 #define XFS_ILOCK_PARENT(1 << XFS_ILOCK_SHIFT) #define XFS_ILOCK_INUMORDER (2 << XFS_ILOCK_SHIFT) So, in a lock_mode parameter, the upper 8 bits are for the ILOCK lockdep subclass, and the 16..23 bits are for the IOLOCK lockdep subclass. Where do we add them? static inline int xfs_lock_inumorder(int lock_mode, int subclass) { if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) lock_mode |= (subclass + XFS_IOLOCK_INUMORDER) << XFS_IOLOCK_SHIFT; if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) lock_mode |= (subclass + XFS_ILOCK_INUMORDER) << XFS_ILOCK_SHIFT; return lock_mode; } OH, look at those nice overflow bugs in that in that code. We shift the XFS_IOLOCK_INUMORDER and XFS_ILOCK_INUMORDER bits out the far side of the lock_mode variable result in lock subclasses of 0-3 instead of 2-5 Bugger, eh? Patch below should fix this (untested). Jarek - thanks for pointing what I should have seen earlier. Cheers, Dave. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [xfs-masters] Re: [BUG] Lockdep warning with XFS on 2.6.22-rc6
Patch looks good, Dave. (though, I stuffed up reviewing that bit of code previously:-) Oh, previous typo: s/inodes at the some time/inodes at the same time/ --Tim David Chinner wrote: On Tue, Jun 26, 2007 at 11:35:20AM +0200, Jarek Poplawski wrote: On 26-06-2007 04:16, David Chinner wrote: It does both - parent-first/child-second and ascending inode # order, which is where the problem is. standing alone, these seem fine, but they don't appear to work when the child has a lower inode number than the parent. ... From xfs_inode.h: /* * Flags for lockdep annotations. * * XFS_I[O]LOCK_PARENT - for operations that require locking two inodes * (ie directory operations that require locking a directory inode and * an entry inode). The first inode gets locked with this flag so it * gets a lockdep subclass of 1 and the second lock will have a lockdep * subclass of 0. * * XFS_I[O]LOCK_INUMORDER - for locking several inodes at the some time * with xfs_lock_inodes(). This flag is used as the starting subclass * and each subsequent lock acquired will increment the subclass by one. * So the first lock acquired will have a lockdep subclass of 2, the * second lock will have a lockdep subclass of 3, and so on. */ I don't know xfs code, and probably miss something, but it seems there could be some inconsistency: lockdep warning shows mr_lock/1 taken both before and after mr_lock (i.e. /0). According to the above comment there should be always 1 before 0... That just fired some rusty neurons. #define XFS_IOLOCK_SHIFT16 #define XFS_IOLOCK_PARENT (1 XFS_IOLOCK_SHIFT) #define XFS_IOLOCK_INUMORDER(2 XFS_IOLOCK_SHIFT) #define XFS_ILOCK_SHIFT 24 #define XFS_ILOCK_PARENT(1 XFS_ILOCK_SHIFT) #define XFS_ILOCK_INUMORDER (2 XFS_ILOCK_SHIFT) So, in a lock_mode parameter, the upper 8 bits are for the ILOCK lockdep subclass, and the 16..23 bits are for the IOLOCK lockdep subclass. Where do we add them? static inline int xfs_lock_inumorder(int lock_mode, int subclass) { if (lock_mode (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) lock_mode |= (subclass + XFS_IOLOCK_INUMORDER) XFS_IOLOCK_SHIFT; if (lock_mode (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) lock_mode |= (subclass + XFS_ILOCK_INUMORDER) XFS_ILOCK_SHIFT; return lock_mode; } OH, look at those nice overflow bugs in that in that code. We shift the XFS_IOLOCK_INUMORDER and XFS_ILOCK_INUMORDER bits out the far side of the lock_mode variable result in lock subclasses of 0-3 instead of 2-5 Bugger, eh? Patch below should fix this (untested). Jarek - thanks for pointing what I should have seen earlier. Cheers, Dave. - To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/