Re: [PATCH v2 7/8] fuse: add extent map I/O support

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:
- [High] fuse: missing inode_lock_shared() in DAX read path causes lockdep splats and races
- [High] fuse: fuse_rw_clamp() truncates writes at EOF to 0 bytes, causing infinite loops
- [High] fuse: O_APPEND writes bypass fuse_rw_clamp limits due to ordering
- [High] fuse: ABBA deadlock between mmap_lock and i_rwsem in DAX page fault
--

commit 900b63ca4cd5e837e1d2beda0ce7ca37d122ab2d
Author: Miklos Szeredi <mszeredi@redhat.com>
Subject: fuse: add extent map I/O support

Wire up read, write, splice and mmap operations for extent-mapped
files through the iomap/DAX infrastructure.

> diff --git a/fs/fuse/ext_map.c b/fs/fuse/ext_map.c
> index ab01d678d08e5..61cb22e821b60 100644
> --- a/fs/fuse/ext_map.c
> +++ b/fs/fuse/ext_map.c
[ ... ]
> +static vm_fault_t fuse_ext_map_huge_fault(struct vm_fault *vmf, unsigned int order)
> +{
> +	struct inode *inode = file_inode(vmf->vma->vm_file);
> +	bool write_fault = (vmf->flags & FAULT_FLAG_WRITE) && (vmf->vma->vm_flags & VM_SHARED);
> +	vm_fault_t ret;
> +	unsigned long pfn;
> +
> +	if (WARN_ON_ONCE(!IS_DAX(inode)))
> +		return VM_FAULT_SIGBUS;
> +
> +	if (write_fault) {
> +		sb_start_pagefault(inode->i_sb);
> +		file_update_time(vmf->vma->vm_file);
> +	}
> +
> +	filemap_invalidate_lock_shared(inode->i_mapping);
> +
> +	ret = dax_iomap_fault(vmf, order, &pfn, NULL, &fuse_ext_map_iomap_ops);
> +	if (ret & VM_FAULT_NEEDDSYNC)
> +		ret = dax_finish_sync_fault(vmf, order, pfn);
        ^^^^

[Severity: High]
Can this create an ABBA deadlock between mmap_lock and i_rwsem?

This page fault handler is entered holding the VFS mmap_lock, and then
takes filemap_invalidate_lock_shared(). If the fault requires a data
sync (O_DSYNC), it calls dax_finish_sync_fault() which synchronously
invokes vfs_fsync_range(). 

The filesystem's fsync handler, fuse_fsync(), unconditionally acquires
inode_lock(inode). This creates an mmap_lock -> inode_lock dependency,
which violates the standard kernel lock hierarchy and can deadlock if
another thread holds the inode_lock while taking a page fault.

> +
> +	filemap_invalidate_unlock_shared(inode->i_mapping);
> +
> +	if (write_fault)
> +		sb_end_pagefault(inode->i_sb);
> +
> +	return ret;
> +}
[ ... ]
> +ssize_t fuse_ext_map_write_iter(struct kiocb *iocb, struct iov_iter *from)
> +{
> +	ssize_t res;
> +
> +	fuse_rw_clamp(iocb, from);
> +
> +	res = generic_write_checks(iocb, from);
        ^^^^

[Severity: High]
Can O_APPEND writes bypass the file size limits checked in
fuse_rw_clamp() due to this ordering?

Because fuse_rw_clamp() runs before generic_write_checks(),
kiocb->ki_pos is not yet updated to i_size for O_APPEND writes. Clamping
will evaluate based on the old, smaller position, and
generic_write_checks() will then advance ki_pos to i_size, allowing an
out-of-bounds write via dax_iomap_rw().

[Severity: High]
Also, does fuse_rw_clamp() cause writes at or past EOF to return 0
instead of a standard POSIX error?

When a write starts at or past i_size, fuse_rw_clamp() truncates the
iterator to 0. Then, generic_write_checks() returns 0, which propagates
up to the VFS. This can cause standard utilities (like cp and dd) to
retry indefinitely in an infinite loop instead of receiving -ENOSPC or
-EFBIG.

> +	if (res <= 0)
> +		return res;
> +
> +	res = kiocb_modified(iocb);
> +	if (res)
> +		return res;
> +
> +	return dax_iomap_rw(iocb, from, &fuse_ext_map_iomap_ops);
> +}

> diff --git a/fs/fuse/passthrough.c b/fs/fuse/passthrough.c
> index e9ab1aea34e27..040817ad76e94 100644
> --- a/fs/fuse/passthrough.c
> +++ b/fs/fuse/passthrough.c
> @@ -48,10 +47,10 @@ ssize_t fuse_passthrough_read_iter(struct kiocb *iocb, struct iov_iter *iter)
>  	if (!count)
>  		return 0;
>  
> -	ret = backing_file_read_iter(backing_file, iter, iocb, iocb->ki_flags,
> -				     &ctx);
> +	if (!backing_file)
> +		return fuse_ext_map_read_iter(iocb, iter);
        ^^^^

[Severity: High]
Does this path miss acquiring the inode read lock before calling into
the DAX read operations?

When fuse_ext_map_read_iter() processes the read, it delegates to
dax_iomap_rw() which expects the VFS inode lock to be held for
synchronization against truncates. Skipping
inode_lock_shared(&inode->i_rwsem) here could lead to lockdep splats or
out-of-bounds reads during concurrent mapping changes.

>  
> -	return ret;
> +	return backing_file_read_iter(backing_file, iter, iocb, iocb->ki_flags, &ctx);
>  }

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261001150935.655979-1-mszeredi@redhat.com?part=7