NTFS: - Modify ->readpage and ->writepage (fs/ntfs/aops.c) so they detect

        and handle the case where an attribute is converted from resident
        to non-resident by a concurrent file write.
      - Reorder some operations when converting an attribute from resident
        to non-resident (fs/ntfs/attrib.c) so it is safe wrt concurrent
        ->readpage and ->writepage.

Signed-off-by: Anton Altaparmakov <aia21@cantab.net>

3 files changed, 46 insertions, 16 deletions

diff --git a/fs/ntfs/ChangeLog b/fs/ntfs/ChangeLog
index e0b4adf5adc..b29e0618f35 100644
--- a/fs/ntfs/ChangeLog
+++ b/fs/ntfs/ChangeLog
@@ -100,6 +100,9 @@ ToDo/Notes:
 	- Add fs/ntfs/attrib.[hc]::ntfs_attr_make_non_resident().
 	- Fix sign of various error return values to be negative in
 	  fs/ntfs/lcnalloc.c.
+	- Modify ->readpage and ->writepage (fs/ntfs/aops.c) so they detect and
+	  handle the case where an attribute is converted from resident to
+	  non-resident by a concurrent file write.
 
 2.1.22 - Many bug and race fixes and error handling improvements.
 
diff --git a/fs/ntfs/aops.c b/fs/ntfs/aops.c
index 2a7cba258cc..6241c4cfbe2 100644
--- a/fs/ntfs/aops.c
+++ b/fs/ntfs/aops.c
@@ -355,6 +355,7 @@ static int ntfs_readpage(struct file *file, struct page *page)
 	u32 attr_len;
 	int err = 0;
 
+retry_readpage:
 	BUG_ON(!PageLocked(page));
 	/*
 	 * This can potentially happen because we clear PageUptodate() during
@@ -408,6 +409,14 @@ static int ntfs_readpage(struct file *file, struct page *page)
 		err = PTR_ERR(mrec);
 		goto err_out;
 	}
+	/*
+	 * If a parallel write made the attribute non-resident, drop the mft
+	 * record and retry the readpage.
+	 */
+	if (unlikely(NInoNonResident(ni))) {
+		unmap_mft_record(base_ni);
+		goto retry_readpage;
+	}
 	ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
 	if (unlikely(!ctx)) {
 		err = -ENOMEM;
@@ -1248,6 +1257,7 @@ static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
 	u32 attr_len;
 	int err;
 
+retry_writepage:
 	BUG_ON(!PageLocked(page));
 	i_size = i_size_read(vi);
 	/* Is the page fully outside i_size? (truncate in progress) */
@@ -1338,6 +1348,14 @@ static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
 		ctx = NULL;
 		goto err_out;
 	}
+	/*
+	 * If a parallel write made the attribute non-resident, drop the mft
+	 * record and retry the writepage.
+	 */
+	if (unlikely(NInoNonResident(ni))) {
+		unmap_mft_record(base_ni);
+		goto retry_writepage;
+	}
 	ctx = ntfs_attr_get_search_ctx(base_ni, m);
 	if (unlikely(!ctx)) {
 		err = -ENOMEM;
diff --git a/fs/ntfs/attrib.c b/fs/ntfs/attrib.c
index 3b9de404021..41859343a0c 100644
--- a/fs/ntfs/attrib.c
+++ b/fs/ntfs/attrib.c
@@ -1376,19 +1376,6 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 	err = ntfs_attr_record_resize(m, a, arec_size);
 	if (unlikely(err))
 		goto err_out;
-	/* Setup the in-memory attribute structure to be non-resident. */
-	NInoSetNonResident(ni);
-	ni->runlist.rl = rl;
-	write_lock_irqsave(&ni->size_lock, flags);
-	ni->allocated_size = new_size;
-	write_unlock_irqrestore(&ni->size_lock, flags);
-	/*
-	 * FIXME: For now just clear all of these as we do not support them
-	 * when writing.
-	 */
-	NInoClearCompressed(ni);
-	NInoClearSparse(ni);
-	NInoClearEncrypted(ni);
 	/*
 	 * Convert the resident part of the attribute record to describe a
 	 * non-resident attribute.
@@ -1399,7 +1386,10 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 		memmove((u8*)a + name_ofs, (u8*)a + le16_to_cpu(a->name_offset),
 				a->name_length * sizeof(ntfschar));
 	a->name_offset = cpu_to_le16(name_ofs);
-	/* Update the flags to match the in-memory ones. */
+	/*
+	 * FIXME: For now just clear all of these as we do not support them
+	 * when writing.
+	 */
 	a->flags &= cpu_to_le16(0xffff & ~le16_to_cpu(ATTR_IS_SPARSE |
 			ATTR_IS_ENCRYPTED | ATTR_COMPRESSION_MASK));
 	/* Setup the fields specific to non-resident attributes. */
@@ -1422,6 +1412,25 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 				err);
 		goto undo_err_out;
 	}
+	/* Setup the in-memory attribute structure to be non-resident. */
+	/*
+	 * FIXME: For now just clear all of these as we do not support them
+	 * when writing.
+	 */
+	NInoClearSparse(ni);
+	NInoClearEncrypted(ni);
+	NInoClearCompressed(ni);
+	ni->runlist.rl = rl;
+	write_lock_irqsave(&ni->size_lock, flags);
+	ni->allocated_size = new_size;
+	write_unlock_irqrestore(&ni->size_lock, flags);
+	/*
+	 * This needs to be last since the address space operations ->readpage
+	 * and ->writepage can run concurrently with us as they are not
+	 * serialized on i_sem.  Note, we are not allowed to fail once we flip
+	 * this switch, which is another reason to do this last.
+	 */
+	NInoSetNonResident(ni);
 	/* Mark the mft record dirty, so it gets written back. */
 	flush_dcache_mft_record_page(ctx->ntfs_ino);
 	mark_mft_record_dirty(ctx->ntfs_ino);
@@ -1431,6 +1440,7 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 	if (page) {
 		set_page_dirty(page);
 		unlock_page(page);
+		mark_page_accessed(page);
 		page_cache_release(page);
 	}
 	ntfs_debug("Done.");
@@ -1492,11 +1502,10 @@ undo_err_out:
 		memcpy((u8*)a + mp_ofs, kaddr, attr_size);
 		kunmap_atomic(kaddr, KM_USER0);
 	}
-	/* Finally setup the ntfs inode appropriately. */
+	/* Setup the allocated size in the ntfs inode in case it changed. */
 	write_lock_irqsave(&ni->size_lock, flags);
 	ni->allocated_size = arec_size - mp_ofs;
 	write_unlock_irqrestore(&ni->size_lock, flags);
-	NInoClearNonResident(ni);
 	/* Mark the mft record dirty, so it gets written back. */
 	flush_dcache_mft_record_page(ctx->ntfs_ino);
 	mark_mft_record_dirty(ctx->ntfs_ino);