Function fuse_direntplus_link() might call fuse_iget() to initialize a new
fuse_inode and change its attributes. If fi->attr_version is always
initialized with 0, even if the attributes returned by the FUSE_READDIR
request is staled, as the new fi->attr_version is 0, fuse_change_attributes
will still set the staled attributes to inode. This wrong behaviour may
cause file size inconsistency even when there is no changes from
server-side.

To reproduce the issue, consider the following 2 programs (A and B) are
running concurrently,

        A                                               B
----------------------------------      --------------------------------
{ /fusemnt/dir/f is a file path in a fuse mount, the size of f is 0. }

readdir(/fusemnt/dir) start
//Daemon set size 0 to f direntry
                                        fallocate(f, 1024)
                                        stat(f) // B see size 1024
                                        echo 2 > /proc/sys/vm/drop_caches
readdir(/fusemnt/dir) reply to kernel
Kernel set 0 to the I_NEW inode

                                        stat(f) // B see size 0

In the above case, only program B is modifying the file size, however, B
observes file size changing between the 2 'readonly' stat() calls. To fix
this issue, we should make sure readdirplus still follows the rule of
attr_version staleness checking even if the fi->attr_version is lost due to
inode eviction. So this patch increases fc->attr_version on inode eviction,
and compares request attr_version and the fc->attr_version when a
FUSE_READDIRPLUS request is finished.

Signed-off-by: Jiachen Zhang <zhangjiachen.jaycee@bytedance.com>
---
 fs/fuse/inode.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
index 660be31aaabc..3e0b1fb1db17 100644
--- a/fs/fuse/inode.c
+++ b/fs/fuse/inode.c
@@ -115,6 +115,7 @@ static void fuse_free_inode(struct inode *inode)
 
 static void fuse_evict_inode(struct inode *inode)
 {
+	struct fuse_conn *fc = get_fuse_conn(inode);
 	struct fuse_inode *fi = get_fuse_inode(inode);
 
 	/* Will write inode on close/munmap and in all other dirtiers */
@@ -137,6 +138,8 @@ static void fuse_evict_inode(struct inode *inode)
 		WARN_ON(!list_empty(&fi->write_files));
 		WARN_ON(!list_empty(&fi->queued_writes));
 	}
+
+	atomic64_inc(&fc->attr_version);
 }
 
 static int fuse_reconfigure(struct fs_context *fsc)
@@ -409,6 +412,10 @@ struct inode *fuse_iget(struct super_block *sb, u64 nodeid,
 	fi->nlookup++;
 	spin_unlock(&fi->lock);
 	fuse_change_attributes(inode, attr, attr_valid, attr_version);
+	spin_lock(&fi->lock);
+	if (attr_version < atomic64_read(&fc->attr_version))
+		fuse_invalidate_attr(inode);
+	spin_unlock(&fi->lock);
 
 	return inode;
 }
-- 
2.20.1

On Tue, 11 Jul 2023 at 06:36, Jiachen Zhang
<zhangjiachen.jaycee@bytedance.com> wrote:
>
> Function fuse_direntplus_link() might call fuse_iget() to initialize a new
> fuse_inode and change its attributes. If fi->attr_version is always
> initialized with 0, even if the attributes returned by the FUSE_READDIR
> request is staled, as the new fi->attr_version is 0, fuse_change_attributes
> will still set the staled attributes to inode. This wrong behaviour may
> cause file size inconsistency even when there is no changes from
> server-side.
>
> To reproduce the issue, consider the following 2 programs (A and B) are
> running concurrently,
>
>         A                                               B
> ----------------------------------      --------------------------------
> { /fusemnt/dir/f is a file path in a fuse mount, the size of f is 0. }
>
> readdir(/fusemnt/dir) start
> //Daemon set size 0 to f direntry
>                                         fallocate(f, 1024)
>                                         stat(f) // B see size 1024
>                                         echo 2 > /proc/sys/vm/drop_caches
> readdir(/fusemnt/dir) reply to kernel
> Kernel set 0 to the I_NEW inode
>
>                                         stat(f) // B see size 0
>
> In the above case, only program B is modifying the file size, however, B
> observes file size changing between the 2 'readonly' stat() calls. To fix
> this issue, we should make sure readdirplus still follows the rule of
> attr_version staleness checking even if the fi->attr_version is lost due to
> inode eviction. So this patch increases fc->attr_version on inode eviction,
> and compares request attr_version and the fc->attr_version when a
> FUSE_READDIRPLUS request is finished.

Thanks for the report.  It's really interesting that you are the first
to notice this misbehavior, even though it's been there since the
beginning.

The fix looks correct but suboptimal: fc->attr_version will get
incremented due to non-evict events as well, which could lead to false
positives.   I'd add another counter (e.g. fc->evict_ctr) and manage
that separately for lookup type operations (LOOKUP/READDIRPLUS).

Thanks,
Miklos