ceph already splices the correct dentry (in splice_dentry()) from the
result of mkdir but does nothing more with it.

Now that ->mkdir can return a dentry, return the correct dentry.

Signed-off-by: NeilBrown <neilb@suse.de>
---
 fs/ceph/dir.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index 39e0f240de06..c1a1c168bb27 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 	struct ceph_client *cl = mdsc->fsc->client;
 	struct ceph_mds_request *req;
 	struct ceph_acl_sec_ctx as_ctx = {};
+	struct dentry *ret = NULL;
 	int err;
 	int op;
 
@@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 	    !req->r_reply_info.head->is_dentry)
 		err = ceph_handle_notrace_create(dir, dentry);
 out_req:
+	if (!err && req->r_dentry != dentry)
+		/* Some other dentry was spliced in */
+		ret = dget(req->r_dentry);
 	ceph_mdsc_put_request(req);
 out:
 	if (!err)
+		/* Should this use 'ret' ?? */
 		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 	else
 		d_drop(dentry);
 	ceph_release_acl_sec_ctx(&as_ctx);
-	return ERR_PTR(err);
+	if (err)
+		return ERR_PTR(err);
+	return ret;
 }
 
 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
-- 
2.47.1

On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> ceph already splices the correct dentry (in splice_dentry()) from the
> result of mkdir but does nothing more with it.
> 
> Now that ->mkdir can return a dentry, return the correct dentry.
> 
> Signed-off-by: NeilBrown <neilb@suse.de>
> ---
>  fs/ceph/dir.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> index 39e0f240de06..c1a1c168bb27 100644
> --- a/fs/ceph/dir.c
> +++ b/fs/ceph/dir.c
> @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	struct ceph_client *cl = mdsc->fsc->client;
>  	struct ceph_mds_request *req;
>  	struct ceph_acl_sec_ctx as_ctx = {};
> +	struct dentry *ret = NULL;
>  	int err;
>  	int op;
>  
> @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	    !req->r_reply_info.head->is_dentry)
>  		err = ceph_handle_notrace_create(dir, dentry);
>  out_req:
> +	if (!err && req->r_dentry != dentry)
> +		/* Some other dentry was spliced in */
> +		ret = dget(req->r_dentry);
>  	ceph_mdsc_put_request(req);
>  out:
>  	if (!err)
> +		/* Should this use 'ret' ?? */

Probably? Is there a guarantee that "dentry" will even have an inode
attached if it got replaced by an disconnected one in the dcache?

>  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
>  	else
>  		d_drop(dentry);
>  	ceph_release_acl_sec_ctx(&as_ctx);
> -	return ERR_PTR(err);
> +	if (err)
> +		return ERR_PTR(err);
> +	return ret;
>  }
>  
>  static int ceph_link(struct dentry *old_dentry, struct inode *dir,

-- 
Jeff Layton <jlayton@kernel.org>

On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> ceph already splices the correct dentry (in splice_dentry()) from the
> result of mkdir but does nothing more with it.
> 
> Now that ->mkdir can return a dentry, return the correct dentry.
> 
> Signed-off-by: NeilBrown <neilb@suse.de>
> ---
>  fs/ceph/dir.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> index 39e0f240de06..c1a1c168bb27 100644
> --- a/fs/ceph/dir.c
> +++ b/fs/ceph/dir.c
> @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	struct ceph_client *cl = mdsc->fsc->client;
>  	struct ceph_mds_request *req;
>  	struct ceph_acl_sec_ctx as_ctx = {};
> +	struct dentry *ret = NULL;

I believe that it makes sense to initialize pointer by error here and always
return ret as output. If something goes wrong in the logic, then we already have
error.

>  	int err;
>  	int op;
>  
> @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	    !req->r_reply_info.head->is_dentry)
>  		err = ceph_handle_notrace_create(dir, dentry);
>  out_req:
> +	if (!err && req->r_dentry != dentry)
> +		/* Some other dentry was spliced in */
> +		ret = dget(req->r_dentry);
>  	ceph_mdsc_put_request(req);
>  out:
>  	if (!err)
> +		/* Should this use 'ret' ?? */

Could we make a decision should or shouldn't? :)
It looks not good to leave this comment instead of proper implementation. Do we
have some obstacles to make this decision?

>  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
>  	else
>  		d_drop(dentry);
>  	ceph_release_acl_sec_ctx(&as_ctx);
> -	return ERR_PTR(err);
> +	if (err)
> +		return ERR_PTR(err);
> +	return ret;

What's about this?

return err ? ERR_PTR(err) : ret;

Thanks,
Slava.

>  }
>  
>  static int ceph_link(struct dentry *old_dentry, struct inode *dir,

On Fri, 21 Feb 2025, Viacheslav Dubeyko wrote:
> On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> > ceph already splices the correct dentry (in splice_dentry()) from the
> > result of mkdir but does nothing more with it.
> > 
> > Now that ->mkdir can return a dentry, return the correct dentry.
> > 
> > Signed-off-by: NeilBrown <neilb@suse.de>
> > ---
> >  fs/ceph/dir.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > index 39e0f240de06..c1a1c168bb27 100644
> > --- a/fs/ceph/dir.c
> > +++ b/fs/ceph/dir.c
> > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	struct ceph_client *cl = mdsc->fsc->client;
> >  	struct ceph_mds_request *req;
> >  	struct ceph_acl_sec_ctx as_ctx = {};
> > +	struct dentry *ret = NULL;
> 
> I believe that it makes sense to initialize pointer by error here and always
> return ret as output. If something goes wrong in the logic, then we already have
> error.

I'm not certain that I understand, but I have made a change which seems
to be consistent with the above and included it below.  Please let me
know if it is what you intended.

> 
> >  	int err;
> >  	int op;
> >  
> > @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	    !req->r_reply_info.head->is_dentry)
> >  		err = ceph_handle_notrace_create(dir, dentry);
> >  out_req:
> > +	if (!err && req->r_dentry != dentry)
> > +		/* Some other dentry was spliced in */
> > +		ret = dget(req->r_dentry);
> >  	ceph_mdsc_put_request(req);
> >  out:
> >  	if (!err)
> > +		/* Should this use 'ret' ?? */
> 
> Could we make a decision should or shouldn't? :)
> It looks not good to leave this comment instead of proper implementation. Do we
> have some obstacles to make this decision?

I suspect we should use ret, but I didn't want to make a change which
wasn't directly required by my needed.  So I highlighted this which
looks to me like a possible bug, hoping that someone more familiar with
the code would give an opinion.  Do you agree that 'ret' (i.e.
->r_dentry) should be used when ret is not NULL?

> 
> >  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> >  	else
> >  		d_drop(dentry);
> >  	ceph_release_acl_sec_ctx(&as_ctx);
> > -	return ERR_PTR(err);
> > +	if (err)
> > +		return ERR_PTR(err);
> > +	return ret;
> 
> What's about this?
> 
> return err ? ERR_PTR(err) : ret;

We could do that, but you said above that you thought we should always
return 'ret' - which does make some sense.

What do you think of the following alternate patch?

Thanks,
NeilBrown

diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
index 39e0f240de06..d2e5c557df83 100644
--- a/fs/ceph/dir.c
+++ b/fs/ceph/dir.c
@@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 	struct ceph_client *cl = mdsc->fsc->client;
 	struct ceph_mds_request *req;
 	struct ceph_acl_sec_ctx as_ctx = {};
+	struct dentry *ret;
 	int err;
 	int op;
 
@@ -1116,32 +1117,32 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 		      ceph_vinop(dir), dentry, dentry, mode);
 		op = CEPH_MDS_OP_MKDIR;
 	} else {
-		err = -EROFS;
+		ret = ERR_PTR(-EROFS);
 		goto out;
 	}
 
 	if (op == CEPH_MDS_OP_MKDIR &&
 	    ceph_quota_is_max_files_exceeded(dir)) {
-		err = -EDQUOT;
+		ret = ERR_PTR(-EDQUOT);
 		goto out;
 	}
 	if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
 	    !fscrypt_has_encryption_key(dir)) {
-		err = -ENOKEY;
+		ret = ERR_PTR(-ENOKEY);
 		goto out;
 	}
 
 
 	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 	if (IS_ERR(req)) {
-		err = PTR_ERR(req);
+		ret = ERR_CAST(req);
 		goto out;
 	}
 
 	mode |= S_IFDIR;
 	req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
 	if (IS_ERR(req->r_new_inode)) {
-		err = PTR_ERR(req->r_new_inode);
+		ret = ERR_CAST(req->r_new_inode);
 		req->r_new_inode = NULL;
 		goto out_req;
 	}
@@ -1165,15 +1166,23 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 	    !req->r_reply_info.head->is_target &&
 	    !req->r_reply_info.head->is_dentry)
 		err = ceph_handle_notrace_create(dir, dentry);
+	ret = ERR_PTR(err);
 out_req:
+	if (!IS_ERR(ret) && req->r_dentry != dentry)
+		/* Some other dentry was spliced in */
+		ret = dget(req->r_dentry);
 	ceph_mdsc_put_request(req);
 out:
-	if (!err)
-		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
-	else
+	if (!IS_ERR(ret)) {
+		if (ret)
+			ceph_init_inode_acls(d_inode(ret), &as_ctx);
+		else
+			ceph_init_inode_acls(d_inode(dentry), &as_ctx);
+	} else {
 		d_drop(dentry);
+	}
 	ceph_release_acl_sec_ctx(&as_ctx);
-	return ERR_PTR(err);
+	return ret;
 }
 
 static int ceph_link(struct dentry *old_dentry, struct inode *dir,

On Mon, 2025-02-24 at 13:15 +1100, NeilBrown wrote:
> On Fri, 21 Feb 2025, Viacheslav Dubeyko wrote:
> > On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> > > ceph already splices the correct dentry (in splice_dentry()) from the
> > > result of mkdir but does nothing more with it.
> > > 
> > > Now that ->mkdir can return a dentry, return the correct dentry.
> > > 
> > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > ---
> > >  fs/ceph/dir.c | 9 ++++++++-
> > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > > index 39e0f240de06..c1a1c168bb27 100644
> > > --- a/fs/ceph/dir.c
> > > +++ b/fs/ceph/dir.c
> > > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > >  	struct ceph_client *cl = mdsc->fsc->client;
> > >  	struct ceph_mds_request *req;
> > >  	struct ceph_acl_sec_ctx as_ctx = {};
> > > +	struct dentry *ret = NULL;
> > 
> > I believe that it makes sense to initialize pointer by error here and always
> > return ret as output. If something goes wrong in the logic, then we already have
> > error.
> 
> I'm not certain that I understand, but I have made a change which seems
> to be consistent with the above and included it below.  Please let me
> know if it is what you intended.
> 
> > 
> > >  	int err;
> > >  	int op;
> > >  
> > > @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > >  	    !req->r_reply_info.head->is_dentry)
> > >  		err = ceph_handle_notrace_create(dir, dentry);
> > >  out_req:
> > > +	if (!err && req->r_dentry != dentry)
> > > +		/* Some other dentry was spliced in */
> > > +		ret = dget(req->r_dentry);
> > >  	ceph_mdsc_put_request(req);
> > >  out:
> > >  	if (!err)
> > > +		/* Should this use 'ret' ?? */
> > 
> > Could we make a decision should or shouldn't? :)
> > It looks not good to leave this comment instead of proper implementation. Do we
> > have some obstacles to make this decision?
> 
> I suspect we should use ret, but I didn't want to make a change which
> wasn't directly required by my needed.  So I highlighted this which
> looks to me like a possible bug, hoping that someone more familiar with
> the code would give an opinion.  Do you agree that 'ret' (i.e.
> ->r_dentry) should be used when ret is not NULL?
> 

I think if we are going to return ret as a dentry, then it makes sense to call
the ceph_init_inode_acls() for d_inode(ret). I don't see the point to call
ceph_init_inode_acls() for d_inode(dentry) then.

> > 
> > >  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > >  	else
> > >  		d_drop(dentry);
> > >  	ceph_release_acl_sec_ctx(&as_ctx);
> > > -	return ERR_PTR(err);
> > > +	if (err)
> > > +		return ERR_PTR(err);
> > > +	return ret;
> > 
> > What's about this?
> > 
> > return err ? ERR_PTR(err) : ret;
> 
> We could do that, but you said above that you thought we should always
> return 'ret' - which does make some sense.
> 
> What do you think of the following alternate patch?
> 

Patch looks good to me. Thanks.

Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>

> Thanks,
> NeilBrown
> 
> diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> index 39e0f240de06..d2e5c557df83 100644
> --- a/fs/ceph/dir.c
> +++ b/fs/ceph/dir.c
> @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	struct ceph_client *cl = mdsc->fsc->client;
>  	struct ceph_mds_request *req;
>  	struct ceph_acl_sec_ctx as_ctx = {};
> +	struct dentry *ret;
>  	int err;
>  	int op;
>  
> @@ -1116,32 +1117,32 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  		      ceph_vinop(dir), dentry, dentry, mode);
>  		op = CEPH_MDS_OP_MKDIR;
>  	} else {
> -		err = -EROFS;
> +		ret = ERR_PTR(-EROFS);
>  		goto out;
>  	}
>  
>  	if (op == CEPH_MDS_OP_MKDIR &&
>  	    ceph_quota_is_max_files_exceeded(dir)) {
> -		err = -EDQUOT;
> +		ret = ERR_PTR(-EDQUOT);
>  		goto out;
>  	}
>  	if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
>  	    !fscrypt_has_encryption_key(dir)) {
> -		err = -ENOKEY;
> +		ret = ERR_PTR(-ENOKEY);
>  		goto out;
>  	}
>  
>  
>  	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
>  	if (IS_ERR(req)) {
> -		err = PTR_ERR(req);
> +		ret = ERR_CAST(req);
>  		goto out;
>  	}
>  
>  	mode |= S_IFDIR;
>  	req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
>  	if (IS_ERR(req->r_new_inode)) {
> -		err = PTR_ERR(req->r_new_inode);
> +		ret = ERR_CAST(req->r_new_inode);
>  		req->r_new_inode = NULL;
>  		goto out_req;
>  	}
> @@ -1165,15 +1166,23 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
>  	    !req->r_reply_info.head->is_target &&
>  	    !req->r_reply_info.head->is_dentry)
>  		err = ceph_handle_notrace_create(dir, dentry);
> +	ret = ERR_PTR(err);
>  out_req:
> +	if (!IS_ERR(ret) && req->r_dentry != dentry)
> +		/* Some other dentry was spliced in */
> +		ret = dget(req->r_dentry);
>  	ceph_mdsc_put_request(req);
>  out:
> -	if (!err)
> -		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> -	else
> +	if (!IS_ERR(ret)) {
> +		if (ret)
> +			ceph_init_inode_acls(d_inode(ret), &as_ctx);
> +		else
> +			ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> +	} else {
>  		d_drop(dentry);
> +	}
>  	ceph_release_acl_sec_ctx(&as_ctx);
> -	return ERR_PTR(err);
> +	return ret;
>  }
>  
>  static int ceph_link(struct dentry *old_dentry, struct inode *dir,
> 

Thanks,
Slava.

On Tue, 25 Feb 2025, Viacheslav Dubeyko wrote:
> On Mon, 2025-02-24 at 13:15 +1100, NeilBrown wrote:
> > On Fri, 21 Feb 2025, Viacheslav Dubeyko wrote:
> > > On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> > > > ceph already splices the correct dentry (in splice_dentry()) from the
> > > > result of mkdir but does nothing more with it.
> > > > 
> > > > Now that ->mkdir can return a dentry, return the correct dentry.
> > > > 
> > > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > > ---
> > > >  fs/ceph/dir.c | 9 ++++++++-
> > > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > > > index 39e0f240de06..c1a1c168bb27 100644
> > > > --- a/fs/ceph/dir.c
> > > > +++ b/fs/ceph/dir.c
> > > > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > > >  	struct ceph_client *cl = mdsc->fsc->client;
> > > >  	struct ceph_mds_request *req;
> > > >  	struct ceph_acl_sec_ctx as_ctx = {};
> > > > +	struct dentry *ret = NULL;
> > > 
> > > I believe that it makes sense to initialize pointer by error here and always
> > > return ret as output. If something goes wrong in the logic, then we already have
> > > error.
> > 
> > I'm not certain that I understand, but I have made a change which seems
> > to be consistent with the above and included it below.  Please let me
> > know if it is what you intended.
> > 
> > > 
> > > >  	int err;
> > > >  	int op;
> > > >  
> > > > @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > > >  	    !req->r_reply_info.head->is_dentry)
> > > >  		err = ceph_handle_notrace_create(dir, dentry);
> > > >  out_req:
> > > > +	if (!err && req->r_dentry != dentry)
> > > > +		/* Some other dentry was spliced in */
> > > > +		ret = dget(req->r_dentry);
> > > >  	ceph_mdsc_put_request(req);
> > > >  out:
> > > >  	if (!err)
> > > > +		/* Should this use 'ret' ?? */
> > > 
> > > Could we make a decision should or shouldn't? :)
> > > It looks not good to leave this comment instead of proper implementation. Do we
> > > have some obstacles to make this decision?
> > 
> > I suspect we should use ret, but I didn't want to make a change which
> > wasn't directly required by my needed.  So I highlighted this which
> > looks to me like a possible bug, hoping that someone more familiar with
> > the code would give an opinion.  Do you agree that 'ret' (i.e.
> > ->r_dentry) should be used when ret is not NULL?
> > 
> 
> I think if we are going to return ret as a dentry, then it makes sense to call
> the ceph_init_inode_acls() for d_inode(ret). I don't see the point to call
> ceph_init_inode_acls() for d_inode(dentry) then.

If the mkdir used the original dentry, then ->mkdir returns NULL so ret
is NULL.  If the mkdir used a different dentry it returns that, so ret
is not NULL.

I'll try to re-organise the code so that "dentry" is the correct dentry
on success, and "ret" is the returned dentry, which might be NULL.

Thanks,
NeilBrown


> 
> > > 
> > > >  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > > >  	else
> > > >  		d_drop(dentry);
> > > >  	ceph_release_acl_sec_ctx(&as_ctx);
> > > > -	return ERR_PTR(err);
> > > > +	if (err)
> > > > +		return ERR_PTR(err);
> > > > +	return ret;
> > > 
> > > What's about this?
> > > 
> > > return err ? ERR_PTR(err) : ret;
> > 
> > We could do that, but you said above that you thought we should always
> > return 'ret' - which does make some sense.
> > 
> > What do you think of the following alternate patch?
> > 
> 
> Patch looks good to me. Thanks.
> 
> Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
> 
> > Thanks,
> > NeilBrown
> > 
> > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > index 39e0f240de06..d2e5c557df83 100644
> > --- a/fs/ceph/dir.c
> > +++ b/fs/ceph/dir.c
> > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	struct ceph_client *cl = mdsc->fsc->client;
> >  	struct ceph_mds_request *req;
> >  	struct ceph_acl_sec_ctx as_ctx = {};
> > +	struct dentry *ret;
> >  	int err;
> >  	int op;
> >  
> > @@ -1116,32 +1117,32 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  		      ceph_vinop(dir), dentry, dentry, mode);
> >  		op = CEPH_MDS_OP_MKDIR;
> >  	} else {
> > -		err = -EROFS;
> > +		ret = ERR_PTR(-EROFS);
> >  		goto out;
> >  	}
> >  
> >  	if (op == CEPH_MDS_OP_MKDIR &&
> >  	    ceph_quota_is_max_files_exceeded(dir)) {
> > -		err = -EDQUOT;
> > +		ret = ERR_PTR(-EDQUOT);
> >  		goto out;
> >  	}
> >  	if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
> >  	    !fscrypt_has_encryption_key(dir)) {
> > -		err = -ENOKEY;
> > +		ret = ERR_PTR(-ENOKEY);
> >  		goto out;
> >  	}
> >  
> >  
> >  	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
> >  	if (IS_ERR(req)) {
> > -		err = PTR_ERR(req);
> > +		ret = ERR_CAST(req);
> >  		goto out;
> >  	}
> >  
> >  	mode |= S_IFDIR;
> >  	req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
> >  	if (IS_ERR(req->r_new_inode)) {
> > -		err = PTR_ERR(req->r_new_inode);
> > +		ret = ERR_CAST(req->r_new_inode);
> >  		req->r_new_inode = NULL;
> >  		goto out_req;
> >  	}
> > @@ -1165,15 +1166,23 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	    !req->r_reply_info.head->is_target &&
> >  	    !req->r_reply_info.head->is_dentry)
> >  		err = ceph_handle_notrace_create(dir, dentry);
> > +	ret = ERR_PTR(err);
> >  out_req:
> > +	if (!IS_ERR(ret) && req->r_dentry != dentry)
> > +		/* Some other dentry was spliced in */
> > +		ret = dget(req->r_dentry);
> >  	ceph_mdsc_put_request(req);
> >  out:
> > -	if (!err)
> > -		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > -	else
> > +	if (!IS_ERR(ret)) {
> > +		if (ret)
> > +			ceph_init_inode_acls(d_inode(ret), &as_ctx);
> > +		else
> > +			ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > +	} else {
> >  		d_drop(dentry);
> > +	}
> >  	ceph_release_acl_sec_ctx(&as_ctx);
> > -	return ERR_PTR(err);
> > +	return ret;
> >  }
> >  
> >  static int ceph_link(struct dentry *old_dentry, struct inode *dir,
> > 
> 
> Thanks,
> Slava.
> 
> 

On Mon, 2025-02-24 at 22:09 +0000, Viacheslav Dubeyko wrote:
> On Mon, 2025-02-24 at 13:15 +1100, NeilBrown wrote:
> > On Fri, 21 Feb 2025, Viacheslav Dubeyko wrote:
> > > On Fri, 2025-02-21 at 10:36 +1100, NeilBrown wrote:
> > > > ceph already splices the correct dentry (in splice_dentry()) from the
> > > > result of mkdir but does nothing more with it.
> > > > 
> > > > Now that ->mkdir can return a dentry, return the correct dentry.
> > > > 
> > > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > > ---
> > > >  fs/ceph/dir.c | 9 ++++++++-
> > > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > > > index 39e0f240de06..c1a1c168bb27 100644
> > > > --- a/fs/ceph/dir.c
> > > > +++ b/fs/ceph/dir.c
> > > > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > > >  	struct ceph_client *cl = mdsc->fsc->client;
> > > >  	struct ceph_mds_request *req;
> > > >  	struct ceph_acl_sec_ctx as_ctx = {};
> > > > +	struct dentry *ret = NULL;
> > > 
> > > I believe that it makes sense to initialize pointer by error here and always
> > > return ret as output. If something goes wrong in the logic, then we already have
> > > error.
> > 
> > I'm not certain that I understand, but I have made a change which seems
> > to be consistent with the above and included it below.  Please let me
> > know if it is what you intended.
> > 
> > > 
> > > >  	int err;
> > > >  	int op;
> > > >  
> > > > @@ -1166,14 +1167,20 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> > > >  	    !req->r_reply_info.head->is_dentry)
> > > >  		err = ceph_handle_notrace_create(dir, dentry);
> > > >  out_req:
> > > > +	if (!err && req->r_dentry != dentry)
> > > > +		/* Some other dentry was spliced in */
> > > > +		ret = dget(req->r_dentry);
> > > >  	ceph_mdsc_put_request(req);
> > > >  out:
> > > >  	if (!err)
> > > > +		/* Should this use 'ret' ?? */
> > > 
> > > Could we make a decision should or shouldn't? :)
> > > It looks not good to leave this comment instead of proper implementation. Do we
> > > have some obstacles to make this decision?
> > 
> > I suspect we should use ret, but I didn't want to make a change which
> > wasn't directly required by my needed.  So I highlighted this which
> > looks to me like a possible bug, hoping that someone more familiar with
> > the code would give an opinion.  Do you agree that 'ret' (i.e.
> > ->r_dentry) should be used when ret is not NULL?
> > 
> 
> I think if we are going to return ret as a dentry, then it makes sense to call
> the ceph_init_inode_acls() for d_inode(ret). I don't see the point to call
> ceph_init_inode_acls() for d_inode(dentry) then.
> 

My assumption when looking at this was that they should point to the
same inode. That said, working with d_inode(ret) after that point is
less confusing to the casual reader.

> > > 
> > > >  		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > > >  	else
> > > >  		d_drop(dentry);
> > > >  	ceph_release_acl_sec_ctx(&as_ctx);
> > > > -	return ERR_PTR(err);
> > > > +	if (err)
> > > > +		return ERR_PTR(err);
> > > > +	return ret;
> > > 
> > > What's about this?
> > > 
> > > return err ? ERR_PTR(err) : ret;
> > 
> > We could do that, but you said above that you thought we should always
> > return 'ret' - which does make some sense.
> > 
> > What do you think of the following alternate patch?
> > 
> 
> Patch looks good to me. Thanks.
> 
> Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
> 
> > Thanks,
> > NeilBrown
> > 
> > diff --git a/fs/ceph/dir.c b/fs/ceph/dir.c
> > index 39e0f240de06..d2e5c557df83 100644
> > --- a/fs/ceph/dir.c
> > +++ b/fs/ceph/dir.c
> > @@ -1099,6 +1099,7 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	struct ceph_client *cl = mdsc->fsc->client;
> >  	struct ceph_mds_request *req;
> >  	struct ceph_acl_sec_ctx as_ctx = {};
> > +	struct dentry *ret;
> >  	int err;
> >  	int op;
> >  
> > @@ -1116,32 +1117,32 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  		      ceph_vinop(dir), dentry, dentry, mode);
> >  		op = CEPH_MDS_OP_MKDIR;
> >  	} else {
> > -		err = -EROFS;
> > +		ret = ERR_PTR(-EROFS);
> >  		goto out;
> >  	}
> >  
> >  	if (op == CEPH_MDS_OP_MKDIR &&
> >  	    ceph_quota_is_max_files_exceeded(dir)) {
> > -		err = -EDQUOT;
> > +		ret = ERR_PTR(-EDQUOT);
> >  		goto out;
> >  	}
> >  	if ((op == CEPH_MDS_OP_MKSNAP) && IS_ENCRYPTED(dir) &&
> >  	    !fscrypt_has_encryption_key(dir)) {
> > -		err = -ENOKEY;
> > +		ret = ERR_PTR(-ENOKEY);
> >  		goto out;
> >  	}
> >  
> >  
> >  	req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
> >  	if (IS_ERR(req)) {
> > -		err = PTR_ERR(req);
> > +		ret = ERR_CAST(req);
> >  		goto out;
> >  	}
> >  
> >  	mode |= S_IFDIR;
> >  	req->r_new_inode = ceph_new_inode(dir, dentry, &mode, &as_ctx);
> >  	if (IS_ERR(req->r_new_inode)) {
> > -		err = PTR_ERR(req->r_new_inode);
> > +		ret = ERR_CAST(req->r_new_inode);
> >  		req->r_new_inode = NULL;
> >  		goto out_req;
> >  	}
> > @@ -1165,15 +1166,23 @@ static struct dentry *ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
> >  	    !req->r_reply_info.head->is_target &&
> >  	    !req->r_reply_info.head->is_dentry)
> >  		err = ceph_handle_notrace_create(dir, dentry);
> > +	ret = ERR_PTR(err);
> >  out_req:
> > +	if (!IS_ERR(ret) && req->r_dentry != dentry)
> > +		/* Some other dentry was spliced in */
> > +		ret = dget(req->r_dentry);
> >  	ceph_mdsc_put_request(req);
> >  out:
> > -	if (!err)
> > -		ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > -	else
> > +	if (!IS_ERR(ret)) {
> > +		if (ret)
> > +			ceph_init_inode_acls(d_inode(ret), &as_ctx);
> > +		else
> > +			ceph_init_inode_acls(d_inode(dentry), &as_ctx);
> > +	} else {
> >  		d_drop(dentry);
> > +	}
> >  	ceph_release_acl_sec_ctx(&as_ctx);
> > -	return ERR_PTR(err);
> > +	return ret;
> >  }
> >  
> >  static int ceph_link(struct dentry *old_dentry, struct inode *dir,
> > 
> 
> Thanks,
> Slava.
> 

-- 
Jeff Layton <jlayton@kernel.org>