Skip to contentSkip to conversation

fuse: invalidate inode pagecache when atomic_o_trunc flag is enabled

PATCH In review Chad Austin 3 participants 3 messages Dec 5, 2017 – Feb 9, 2018

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fuse: invalidate inode pagecache when atomic_o_trunc flag is enabled

When the atomic_o_trunc flag is set, fuse_do_setattr skips most of its logic, including truncating the page cache. Run the same invalidation logic in fuse_finish_open when truncating. This fixes a bug where open(O_TRUNC) followed by a sequence of writes does not flush the page cache, resulting in an mmap longer than the new file's length (from another process) not properly zero-filling the last page.

Does fuse_finish_open() need additional locking? I am not sure of the locking requirements of truncate_pagecache or invalidate_inode_pages2.

The following C++ program reproduces the bug. The FUSE daemon must have enabled the ATOMIC_O_TRUNC flag
https://gist.github.com/chadaustin/8e4ba1e2cd2d023ff6f9eff921eb8bfc

---
 fs/fuse/dir.c    |  7 ++-----
 fs/fuse/file.c   | 12 ++++++++++++
 fs/fuse/fuse_i.h |  2 ++
 3 files changed, 16 insertions(+), 5 deletions(-)
fs/fuse/dir.c +2 -5
index 2496738..c387171 100644
@@ -1698,11 +1698,8 @@ int fuse_do_setattr(struct dentry *dentry, struct iattr *attr,
 	 * Only call invalidate_inode_pages2() after removing
 	 * FUSE_NOWRITE, otherwise fuse_launder_page() would deadlock.
 	 */
-	if ((is_truncate || !is_wb) &&
-	    S_ISREG(inode->i_mode) && oldsize != outarg.attr.size) {
-		truncate_pagecache(inode, outarg.attr.size);
-		invalidate_inode_pages2(inode->i_mapping);
-	}
+	if ((is_truncate || !is_wb) && oldsize != outarg.attr.size)
+		fuse_truncate_inode_cache(inode, outarg.attr.size);
 
 	clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
 	return 0;
fs/fuse/file.c +12 -0
index cb7dff5..cb1968e 100644
@@ -188,6 +188,10 @@ void fuse_finish_open(struct inode *inode, struct file *file)
 		i_size_write(inode, 0);
 		spin_unlock(&fc->lock);
 		fuse_invalidate_attr(inode);
+
+		if (!fc->writeback_cache)
+			fuse_truncate_inode_cache(inode, 0);
+
 		if (fc->writeback_cache)
 			file_update_time(file);
 	}
@@ -195,6 +199,14 @@ void fuse_finish_open(struct inode *inode, struct file *file)
 		fuse_link_write_file(file);
 }
 
+void fuse_truncate_inode_cache(struct inode *inode, loff_t newsize)
+{
+	if (S_ISREG(inode->i_mode)) {
+		truncate_pagecache(inode, newsize);
+		invalidate_inode_pages2(inode->i_mapping);
+	}
+}
+
 int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 {
 	struct fuse_conn *fc = get_fuse_conn(inode);
fs/fuse/fuse_i.h +2 -0
index d5773ca..47baea4 100644
@@ -913,6 +913,8 @@ void fuse_release_nowrite(struct inode *inode);
 
 u64 fuse_get_attr_version(struct fuse_conn *fc);
 
+void fuse_truncate_inode_cache(struct inode *inode, loff_t newsize);
+
 /**
  * File-system tells the kernel to invalidate cache for the given node id.
  */
Signature · 1 line
2.9.5
maintainer
#1 lore raw
Quoted 7 lines · Chad Austin

When the atomic_o_trunc flag is set, fuse_do_setattr skips most of its logic, including truncating the page cache. Run the same invalidation logic in fuse_finish_open when truncating. This fixes a bug where open(O_TRUNC) followed by a sequence of writes does not flush the page cache, resulting in an mmap longer than the new file's length (from another process) not properly zero-filling the last page.

Quoted 6 lines

Does fuse_finish_open() need additional locking? I am not sure of the locking requirements of truncate_pagecache or invalidate_inode_pages2.

The following C++ program reproduces the bug. The FUSE daemon must have enabled the ATOMIC_O_TRUNC flag
https://gist.github.com/chadaustin/8e4ba1e2cd2d023ff6f9eff921eb8bfc

Thanks for the report and the patch.

Here's a simpler patch (we don't actually need the invalidate_inode_pages2(), because all the pages have already been truncated).

I tested it with your reproducer and it fixes the issue for me.

Unless you see a problem with it, I'll queue this up for 4.16.

Thanks,
Miklos

---
 fs/fuse/dir.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
fs/fuse/dir.c +5 -1
@@ -1629,8 +1629,12 @@ int fuse_do_setattr(struct dentry *dentr
 		return err;
 
 	if (attr->ia_valid & ATTR_OPEN) {
-		if (fc->atomic_o_trunc)
+		if (fc->atomic_o_trunc) {
+			WARN_ON(!(attr->ia_valid & ATTR_SIZE));
+			WARN_ON(attr->ia_size != 0);
+			truncate_pagecache(inode, 0);
 			return 0;
+		}
 		file = NULL;
 	}
 
#2 lore raw
Quoted 9 lines · Miklos Szeredi

Thanks for the report and the patch.

Here's a simpler patch (we don't actually need the invalidate_inode_pages2(), because all the pages have already been truncated).

I tested it with your reproducer and it fixes the issue for me.

Unless you see a problem with it, I'll queue this up for 4.16.

Thanks! I verified that your simpler patch also fixes our more complex internal reproduction. Looking forward to it landing!