On 07.11.24 09:58, Miklos Szeredi wrote:
On Thu, 24 Oct 2024 at 18:47, Hanna Czenczek <hreitz@xxxxxxxxxx> wrote:
To be able to issue INIT (and GETATTR), we need to at least partially
initialize the super_block structure, which is currently done via
fuse_fill_super_common().
What exactly is needed to be initialized?
It isn’t much, but I believe it’s most of fuse_fill_super_common()
(without restructuring the code so flags returned by INIT are put into a
separate structure and then re-joined into sb and fc later).
fuse_send_init() reads sb->s_bdi->ra_pages, process_init_reply() writes
it and sb->s_time_gran, ->s_flags, ->s_stack_depth, ->s_export_op, and
->s_iflags. In addition, process_init_reply() depends on several flags
and objects in fc being set up (among those are fc->dax and
fc->default_permissions), which is done by fuse_fill_super_common().
So I think what we need from fuse_fill_super_common() is:
- fuse_sb_defaults() (so these values can then be overwritten by
process_init_reply()),
- fuse_dax_conn_alloc(),
- fuse_bdi_init(),
- fc->default_permissions at least, but I’d just take the fc->[flag]
setting block as a whole then.
I assume we’ll also want the SB_MANDLOCK check then, and
rcu_assign_pointer(). Then we might as well also set the block sizes
and the subtype.
The problem is that I don’t know the order things in
fuse_fill_super_common() need to be in, and fuse_dev_alloc_install() is
called before fuse_bdi_init(), so I didn’t want to move that.
So what I understand is that calling fuse_dev_alloc_install() there
isn’t necessary? I’m happy to move that to part 2, as you suggest, but
I’m not sure we can really omit much from part 1 without changing how
process_init_reply() operates. We could in theory delay
process_init_reply() until after GETATTR (and thus after setting
s_root), but that seems kind of wrong, and would still require setting
up BDI and DAX for fuse_send_init().
@@ -1762,18 +1801,12 @@ int fuse_fill_super_common(struct super_block *sb, struct fuse_fs_context *ctx)
sb->s_d_op = &fuse_dentry_operations;
mutex_lock(&fuse_mutex);
- err = -EINVAL;
- if (ctx->fudptr && *ctx->fudptr)
- goto err_unlock;
-
err = fuse_ctl_add_conn(fc);
if (err)
goto err_unlock;
list_add_tail(&fc->entry, &fuse_conn_list);
sb->s_root = root_dentry;
- if (ctx->fudptr)
- *ctx->fudptr = fud;
This is wrong, because we need the fuse_mutex protection for checking
and setting the private_data on the fuse device file.
If this split is needed (which I'm not sure) then fud allocation
should probably be moved to part2 instead of moving the *ctx->fudptr
setup to part1.
@@ -1635,8 +1657,16 @@ static void virtio_kill_sb(struct super_block *sb)
struct fuse_mount *fm = get_fuse_mount_super(sb);
bool last;
- /* If mount failed, we can still be called without any fc */
- if (sb->s_root) {
+ /*
+ * Only destroy the connection after full initialization, i.e.
+ * once s_root is set (see commit d534d31d6a45d).
+ * One exception: For virtio-fs, we call INIT before s_root is
+ * set so we can determine the root node's mode. We must call
+ * DESTROY after INIT. So if an error occurs during that time
+ * window (specifically in fuse_make_root_inode()), we still
+ * need to call virtio_fs_conn_destroy() here.
+ */
+ if (sb->s_root || (fm->fc && fm->fc->initialized && !fm->submount)) {
How could fm->submount be set if sb->s_root isn't?
fuse_get_tree_submount(), specifically fuse_fill_super_submount() whose
error path leads to deactivate_locked_super(), can fail before
sb->s_root is set.
Still, the idea was rather to make it clear that this condition (INIT
sent but s_root not set) is unique to non-submounts, so as not to mess
with the submount code unintentionally.
Or sb->s_root set
and fc->initialized isn't?
That would be the non-virtio-fs non-submount case (fuse_fill_super()),
where s_root is set first and INIT sent after.
Hanna
Seems it would be sufficient to check fm->fc->initialized, no?
Thanks,
Miklos
