On Wed, Feb 26, 2020 at 09:59:41AM +1100, Dave Chinner wrote: > On Tue, Feb 25, 2020 at 01:12:28PM -0800, Ira Weiny wrote: > > On Tue, Feb 25, 2020 at 09:32:45AM +1100, Dave Chinner wrote: > > > On Mon, Feb 24, 2020 at 11:57:36AM -0800, Ira Weiny wrote: > > > > On Mon, Feb 24, 2020 at 11:34:55AM +1100, Dave Chinner wrote: > > > > > On Thu, Feb 20, 2020 at 04:41:30PM -0800, ira.weiny@xxxxxxxxx wrote: > > > > > > From: Ira Weiny <ira.weiny@xxxxxxxxx> > > > > > > > > > > > > > > [snip] > > > > > > > > > > > > > > > > diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c > > > > > > index 35df324875db..5b014c428f0f 100644 > > > > > > --- a/fs/xfs/xfs_inode.c > > > > > > +++ b/fs/xfs/xfs_inode.c > > > > > > @@ -142,12 +142,12 @@ xfs_ilock_attr_map_shared( > > > > > > * > > > > > > * Basic locking order: > > > > > > * > > > > > > - * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock > > > > > > + * s_dax_sem -> i_rwsem -> i_mmap_lock -> page_lock -> i_ilock > > > > > > * > > > > > > * mmap_sem locking order: > > > > > > * > > > > > > * i_rwsem -> page lock -> mmap_sem > > > > > > - * mmap_sem -> i_mmap_lock -> page_lock > > > > > > + * s_dax_sem -> mmap_sem -> i_mmap_lock -> page_lock > > > > > > * > > > > > > * The difference in mmap_sem locking order mean that we cannot hold the > > > > > > * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can > > > > > > @@ -182,6 +182,9 @@ xfs_ilock( > > > > > > (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)); > > > > > > ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0); > > > > > > > > > > > > + if (lock_flags & XFS_DAX_EXCL) > > > > > > + inode_aops_down_write(VFS_I(ip)); > > > > > > > > > > I largely don't see the point of adding this to xfs_ilock/iunlock. > > > > > > > > > > It's only got one caller, so I don't see much point in adding it to > > > > > an interface that has over a hundred other call sites that don't > > > > > need or use this lock. just open code it where it is needed in the > > > > > ioctl code. > > > > > > > > I know it seems overkill but if we don't do this we need to code a flag to be > > > > returned from xfs_ioctl_setattr_dax_invalidate(). This flag is then used in > > > > xfs_ioctl_setattr_get_trans() to create the transaction log item which can then > > > > be properly used to unlock the lock in xfs_inode_item_release() > > > > > > > > I don't know of a cleaner way to communicate to xfs_inode_item_release() to > > > > unlock i_aops_sem after the transaction is complete. > > > > > > We manually unlock inodes after transactions in many cases - > > > anywhere we do a rolling transaction, the inode locks do not get > > > released by the transaction. Hence for a one-off case like this it > > > doesn't really make sense to push all this infrastructure into the > > > transaction subsystem. Especially as we can manually lock before and > > > unlock after the transaction context without any real complexity. > > > > So does xfs_trans_commit() operate synchronously? > > What do you mean by "synchronously", and what are you expecting to > occur (a)synchronously with respect to filesystem objects and/or > on-disk state? > > Keep in mid that the xfs transaction subsystem is a complex > asynchronous IO engine full of feedback loops and resource > management, This is precisely why I added the lock to the transaction state. So that I could guarantee that the lock will be released in the proper order when the complicated transaction subsystem was done with it. I did not see any reason to allow operations to proceed before that time. And so this seemed safe... > so asking if something is "synchronous" without any > other context is a difficult question to answer :) Or apparently it is difficult to even ask... ;-) (...not trying to be sarcastic.) Seriously, I'm not an expert in this area so I did what I thought was most safe. Which for me was the number 1 goal. > > > I want to understand this better because I have fought with a lot of ABBA > > issues with these locks. So... can I hold the lock until after > > xfs_trans_commit() and safely unlock it there... because the XFS_MMAPLOCK_EXCL, > > XFS_IOLOCK_EXCL, and XFS_ILOCK_EXCL will be released at that point? Thus > > preserving the following lock order. > > See how operations like xfs_create, xfs_unlink, etc work. The don't > specify flags to xfs_ijoin(), and so the transaction commits don't > automatically unlock the inode. xfs_ijoin()? Do you mean xfs_trans_ijoin()? > This is necessary so that rolling > transactions are executed atomically w.r.t. inode access - no-one > can lock and access the inode while a multi-commit rolling > transaction on the inode is on-going. > > In this case it's just a single commit and we don't need to keep > it locked after the change is made, so we can unlock the inode > on commit. So for the XFS internal locks the code is fine and > doesn't need to change. We just need to wrap the VFS aops lock (if > we keep it) around the outside of all the XFS locking until the > transaction commits and unlocks the XFS locks... Ok, I went ahead and coded it up and it is testing now. Everything looks good. I have to say that at this point I have to agree that I can't see how a deadlock could occur so... Thanks for the review, Ira > > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx
