On Wed 26-05-21 07:40:41, Dave Chinner wrote:

On Tue, May 25, 2021 at 03:50:44PM +0200, Jan Kara wrote:

quoted

Use invalidate_lock instead of XFS internal i_mmap_lock. The intended
purpose of invalidate_lock is exactly the same. Note that the locking in
__xfs_filemap_fault() slightly changes as filemap_fault() already takes
invalidate_lock.

Reviewed-by: Christoph Hellwig <hch@lst.de>
CC: <redacted>
CC: "Darrick J. Wong" <redacted>
Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/xfs/xfs_file.c  | 12 ++++++-----
 fs/xfs/xfs_inode.c | 52 ++++++++++++++++++++++++++--------------------
 fs/xfs/xfs_inode.h |  1 -
 fs/xfs/xfs_super.c |  2 --
 4 files changed, 36 insertions(+), 31 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 396ef36dcd0a..dc9cb5c20549 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c

@@ -1282,7 +1282,7 @@ xfs_file_llseek(
  *
  * mmap_lock (MM)
  *   sb_start_pagefault(vfs, freeze)
- *     i_mmaplock (XFS - truncate serialisation)
+ *     invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
  *       page_lock (MM)
  *         i_lock (XFS - extent map serialisation)
  */

@@ -1303,24 +1303,26 @@ __xfs_filemap_fault(
 		file_update_time(vmf->vma->vm_file);
 	}
 
-	xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
 	if (IS_DAX(inode)) {
 		pfn_t pfn;
 
+		xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
 		ret = dax_iomap_fault(vmf, pe_size, &pfn, NULL,
 				(write_fault && !vmf->cow_page) ?
 				 &xfs_direct_write_iomap_ops :
 				 &xfs_read_iomap_ops);
 		if (ret & VM_FAULT_NEEDDSYNC)
 			ret = dax_finish_sync_fault(vmf, pe_size, pfn);
+		xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
 	} else {
-		if (write_fault)
+		if (write_fault) {
+			xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
 			ret = iomap_page_mkwrite(vmf,
 					&xfs_buffered_write_iomap_ops);
-		else
+			xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
+		} else
 			ret = filemap_fault(vmf);
 	}
-	xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);

This seems kinda messy. filemap_fault() basically takes the
invalidate lock around the entire operation, it runs, so maybe it
would be cleaner to implement it as:

filemap_fault_locked(vmf)
{
	/* does the filemap fault work */
}

filemap_fault(vmf)
{
	filemap_invalidate_down_read(...)
	ret = filemap_fault_locked(vmf)
	filemap_invalidate_up_read(...)
	return ret;
}

And that means XFS could just call filemap_fault_locked() and not 
have to do all this messy locking just to avoid holding the lock
that filemap_fault has now internalised.

Sure, I can do that.

quoted

@@ -355,8 +358,11 @@ xfs_isilocked(
 
 	if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
 		if (!(lock_flags & XFS_MMAPLOCK_SHARED))
-			return !!ip->i_mmaplock.mr_writer;
-		return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
+			return !debug_locks ||
+				lockdep_is_held_type(
+					&VFS_I(ip)->i_mapping->invalidate_lock,
+					0);
+		return rwsem_is_locked(&VFS_I(ip)->i_mapping->invalidate_lock);
 	}

<sigh>

And so here we are again, losing more of our read vs write debug
checks on debug kernels when lockdep is not enabled....

Can we please add rwsem_is_locked_read() and rwsem_is_locked_write()
wrappers that just look at the rwsem counter value to determine how
the lock is held? Then the mrlock_t can go away entirely....

Apparently someone already did that for XFS as Darrick pointed out. So we
just have to sort out how to merge it.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR