[02/12] ceph: handle idmapped mounts in create_request_message()

Commit Message

Christian Brauner Jan. 4, 2022, 2:04 p.m. UTC

From: Christian Brauner <[e-mail]>

Inode operations that create a new filesystem object such as ->mknod,
->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
Instead the caller's fs{g,u}id is used for the {g,u}id of the new
filesystem object.

Cephfs mds creation request argument structures mirror this filesystem
behavior. They don't encode a {g,u}id explicitly. Instead the caller's
fs{g,u}id that is always sent as part of any mds request is used by the
servers to set the {g,u}id of the new filesystem object.

In order to ensure that the correct {g,u}id is used map the caller's
fs{g,u}id for creation requests. This doesn't require complex changes.
It suffices to pass in the relevant idmapping recorded in the request
message. If this request message was triggered from an inode operation
that creates filesystem objects it will have passed down the relevant
idmaping. If this is a request message that was triggered from an inode
operation that doens't need to take idmappings into account the initial
idmapping is passed down which is an identity mapping and thus is
guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.

The last few weeks before Christmas 2021 I have spent time not just
reading and poking the cephfs kernel code but also took a look at the
ceph mds server userspace to ensure I didn't miss some subtlety.

This made me aware of one complication to solve. All requests send the
caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
server in exactly two cases:

1. to set the ownership for creation requests
2. to determine whether this client is allowed access on this server

Case 1. we already covered and explained. Case 2. is only relevant for
servers where an explicit uid access restriction has been set. That is
to say the mds server restricts access to requests coming from a
specific uid. Servers without uid restrictions will grant access to
requests from any uid by setting MDS_AUTH_UID_ANY.

Case 2. introduces the complication because the caller's fs{g,u}id is
not just used to record ownership but also serves as the {g,u}id used
when checking access to the server.

Consider a user mounting a cephfs client and creating an idmapped mount
from it that maps files owned by uid 1000 to be owned uid 0:

mount -t cephfs -o [...] /unmapped
mount-idmapped --map-mount 1000:0:1 /idmapped

That is to say if the mounted cephfs filesystem contains a file "file1"
which is owned by uid 1000:

- looking at it via /unmapped/file1 will report it as owned by uid 1000
  (One can think of this as the on-disk value.)
- looking at it via /idmapped/file1 will report it as owned by uid 0

Now, consider creating new files via the idmapped mount at /idmapped.
When a caller with fs{g,u}id 1000 creates a file "file2" by going
through the idmapped mount mounted at /idmapped it will create a file
that is owned by uid 1000 on-disk, i.e.:

- looking at it via /unmapped/file2 will report it as owned by uid 1000
- looking at it via /idmapped/file2 will report it as owned by uid 0

Now consider an mds server that has a uid access restriction set and
only grants access to requests from uid 0.

If the client sends a creation request for a file e.g. /idmapped/file2
it will send the caller's fs{g,u}id idmapped according to the idmapped
mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
0 in the idmapped mount and will be sent over the wire allowing the
caller access to the mds server.

However, if the caller is not issuing a creation request the caller's
fs{g,u}id will be send without the mount's idmapping applied. So if the
caller that just successfully created a new file on the restricted mds
server sends a request as fs{g,u}id 1000 access will be refused. This
however is inconsistent.

From my perspective the root of the problem lies in the fact that
creation requests implicitly infer the ownership from the {g,u}id that
gets sent along with every mds request.

I have thought of multiple ways of addressing this problem but the one I
prefer is to give all mds requests that create a filesystem object a
proper, separate {g,u}id field entry in the argument struct. This is,
for example how ->setattr mds requests work.

This way the caller's fs{g,u}id can be used consistenly for server
access checks and is separated from the ownership for new filesystem
objects.

Servers could then be updated to refuse creation requests whenever the
{g,u}id used for access checking doesn't match the {g,u}id used for
creating the filesystem object just as is done for setattr requests on a
uid restricted server. But I am, of course, open to other suggestions.

Cc: Jeff Layton <[e-mail]>
Cc: Ilya Dryomov <[e-mail]>
Cc: [e-mail]
Signed-off-by: Christian Brauner <[e-mail]>
---
 fs/ceph/mds_client.c | 22 ++++++++++++++++++----
 1 file changed, 18 insertions(+), 4 deletions(-)

Comments

Jeff Layton Jan. 4, 2022, 5:40 p.m. UTC | #1

On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> From: Christian Brauner <[e-mail]>
> 
> Inode operations that create a new filesystem object such as ->mknod,
> ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> filesystem object.
> 
> Cephfs mds creation request argument structures mirror this filesystem
> behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> fs{g,u}id that is always sent as part of any mds request is used by the
> servers to set the {g,u}id of the new filesystem object.
> 
> In order to ensure that the correct {g,u}id is used map the caller's
> fs{g,u}id for creation requests. This doesn't require complex changes.
> It suffices to pass in the relevant idmapping recorded in the request
> message. If this request message was triggered from an inode operation
> that creates filesystem objects it will have passed down the relevant
> idmaping. If this is a request message that was triggered from an inode
> operation that doens't need to take idmappings into account the initial
> idmapping is passed down which is an identity mapping and thus is
> guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> 
> The last few weeks before Christmas 2021 I have spent time not just
> reading and poking the cephfs kernel code but also took a look at the
> ceph mds server userspace to ensure I didn't miss some subtlety.
> 
> This made me aware of one complication to solve. All requests send the
> caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> server in exactly two cases:
> 
> 1. to set the ownership for creation requests
> 2. to determine whether this client is allowed access on this server
> 
> Case 1. we already covered and explained. Case 2. is only relevant for
> servers where an explicit uid access restriction has been set. That is
> to say the mds server restricts access to requests coming from a
> specific uid. Servers without uid restrictions will grant access to
> requests from any uid by setting MDS_AUTH_UID_ANY.
> 
> Case 2. introduces the complication because the caller's fs{g,u}id is
> not just used to record ownership but also serves as the {g,u}id used
> when checking access to the server.
> 
> Consider a user mounting a cephfs client and creating an idmapped mount
> from it that maps files owned by uid 1000 to be owned uid 0:
> 
> mount -t cephfs -o [...] /unmapped
> mount-idmapped --map-mount 1000:0:1 /idmapped
> 
> That is to say if the mounted cephfs filesystem contains a file "file1"
> which is owned by uid 1000:
> 
> - looking at it via /unmapped/file1 will report it as owned by uid 1000
>   (One can think of this as the on-disk value.)
> - looking at it via /idmapped/file1 will report it as owned by uid 0
> 
> Now, consider creating new files via the idmapped mount at /idmapped.
> When a caller with fs{g,u}id 1000 creates a file "file2" by going
> through the idmapped mount mounted at /idmapped it will create a file
> that is owned by uid 1000 on-disk, i.e.:
> 
> - looking at it via /unmapped/file2 will report it as owned by uid 1000
> - looking at it via /idmapped/file2 will report it as owned by uid 0
> 
> Now consider an mds server that has a uid access restriction set and
> only grants access to requests from uid 0.
> 
> If the client sends a creation request for a file e.g. /idmapped/file2
> it will send the caller's fs{g,u}id idmapped according to the idmapped
> mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> 0 in the idmapped mount and will be sent over the wire allowing the
> caller access to the mds server.
> 
> However, if the caller is not issuing a creation request the caller's
> fs{g,u}id will be send without the mount's idmapping applied. So if the
> caller that just successfully created a new file on the restricted mds
> server sends a request as fs{g,u}id 1000 access will be refused. This
> however is inconsistent.
> 

IDGI, why would you send the fs{g,u}id without the mount's idmapping
applied in this case? ISTM that idmapping is wholly a client-side
feature, and that you should always map id's regardless of whether
you're creating or not.

> From my perspective the root of the problem lies in the fact that
> creation requests implicitly infer the ownership from the {g,u}id that
> gets sent along with every mds request.
> 
> I have thought of multiple ways of addressing this problem but the one I
> prefer is to give all mds requests that create a filesystem object a
> proper, separate {g,u}id field entry in the argument struct. This is,
> for example how ->setattr mds requests work.
> 
> This way the caller's fs{g,u}id can be used consistenly for server
> access checks and is separated from the ownership for new filesystem
> objects.
> 
> Servers could then be updated to refuse creation requests whenever the
> {g,u}id used for access checking doesn't match the {g,u}id used for
> creating the filesystem object just as is done for setattr requests on a
> uid restricted server. But I am, of course, open to other suggestions.
> 
> Cc: Jeff Layton <[e-mail]>
> Cc: Ilya Dryomov <[e-mail]>
> Cc: [e-mail]
> Signed-off-by: Christian Brauner <[e-mail]>
> ---
>  fs/ceph/mds_client.c | 22 ++++++++++++++++++----
>  1 file changed, 18 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> index ae2cc4ce1d48..1fb43a8fd64c 100644
> --- a/fs/ceph/mds_client.c
> +++ b/fs/ceph/mds_client.c
> @@ -2459,6 +2459,8 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
>  	void *p, *end;
>  	int ret;
>  	bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
> +	kuid_t caller_fsuid;
> +	kgid_t caller_fsgid;
>  
>  	ret = set_request_path_attr(req->r_inode, req->r_dentry,
>  			      req->r_parent, req->r_path1, req->r_ino1.ino,
> @@ -2524,10 +2526,22 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
>  
>  	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
>  	head->op = cpu_to_le32(req->r_op);
> -	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
> -						 req->r_cred->fsuid));
> -	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
> -						 req->r_cred->fsgid));
> +	/*
> +	 * Inode operations that create filesystem objects based on the
> +	 * caller's fs{g,u}id like ->mknod(), ->create(), ->mkdir() etc. don't
> +	 * have separate {g,u}id fields in their respective structs in the
> +	 * ceph_mds_request_args union. Instead the caller_{g,u}id field is
> +	 * used to set ownership of the newly created inode by the mds server.
> +	 * For these inode operations we need to send the mapped fs{g,u}id over
> +	 * the wire. For other cases we simple set req->mnt_userns to the
> +	 * initial idmapping meaning the unmapped fs{g,u}id is sent.
> +	 */
> +	caller_fsuid = mapped_kuid_user(req->mnt_userns, &init_user_ns,
> +					req->r_cred->fsuid);
> +	caller_fsgid = mapped_kgid_user(req->mnt_userns, &init_user_ns,
> +					req->r_cred->fsgid);
> +	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, caller_fsuid));
> +	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, caller_fsgid));
>  	head->ino = cpu_to_le64(req->r_deleg_ino);
>  	head->args = req->r_args;
>

Gregory Farnum Jan. 4, 2022, 7:33 p.m. UTC | #2

On Tue, Jan 4, 2022 at 9:41 AM Jeff Layton <[e-mail]> wrote:
>
> On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > From: Christian Brauner <[e-mail]>
> >
> > Inode operations that create a new filesystem object such as ->mknod,
> > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > filesystem object.
> >
> > Cephfs mds creation request argument structures mirror this filesystem
> > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > fs{g,u}id that is always sent as part of any mds request is used by the
> > servers to set the {g,u}id of the new filesystem object.
> >
> > In order to ensure that the correct {g,u}id is used map the caller's
> > fs{g,u}id for creation requests. This doesn't require complex changes.
> > It suffices to pass in the relevant idmapping recorded in the request
> > message. If this request message was triggered from an inode operation
> > that creates filesystem objects it will have passed down the relevant
> > idmaping. If this is a request message that was triggered from an inode
> > operation that doens't need to take idmappings into account the initial
> > idmapping is passed down which is an identity mapping and thus is
> > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> >
> > The last few weeks before Christmas 2021 I have spent time not just
> > reading and poking the cephfs kernel code but also took a look at the
> > ceph mds server userspace to ensure I didn't miss some subtlety.
> >
> > This made me aware of one complication to solve. All requests send the
> > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > server in exactly two cases:
> >
> > 1. to set the ownership for creation requests
> > 2. to determine whether this client is allowed access on this server
> >
> > Case 1. we already covered and explained. Case 2. is only relevant for
> > servers where an explicit uid access restriction has been set. That is
> > to say the mds server restricts access to requests coming from a
> > specific uid. Servers without uid restrictions will grant access to
> > requests from any uid by setting MDS_AUTH_UID_ANY.
> >
> > Case 2. introduces the complication because the caller's fs{g,u}id is
> > not just used to record ownership but also serves as the {g,u}id used
> > when checking access to the server.
> >
> > Consider a user mounting a cephfs client and creating an idmapped mount
> > from it that maps files owned by uid 1000 to be owned uid 0:
> >
> > mount -t cephfs -o [...] /unmapped
> > mount-idmapped --map-mount 1000:0:1 /idmapped
> >
> > That is to say if the mounted cephfs filesystem contains a file "file1"
> > which is owned by uid 1000:
> >
> > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> >   (One can think of this as the on-disk value.)
> > - looking at it via /idmapped/file1 will report it as owned by uid 0
> >
> > Now, consider creating new files via the idmapped mount at /idmapped.
> > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > through the idmapped mount mounted at /idmapped it will create a file
> > that is owned by uid 1000 on-disk, i.e.:
> >
> > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > - looking at it via /idmapped/file2 will report it as owned by uid 0
> >
> > Now consider an mds server that has a uid access restriction set and
> > only grants access to requests from uid 0.
> >
> > If the client sends a creation request for a file e.g. /idmapped/file2
> > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > 0 in the idmapped mount and will be sent over the wire allowing the
> > caller access to the mds server.
> >
> > However, if the caller is not issuing a creation request the caller's
> > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > caller that just successfully created a new file on the restricted mds
> > server sends a request as fs{g,u}id 1000 access will be refused. This
> > however is inconsistent.
> >
>
> IDGI, why would you send the fs{g,u}id without the mount's idmapping
> applied in this case? ISTM that idmapping is wholly a client-side
> feature, and that you should always map id's regardless of whether
> you're creating or not.

Yeah, I'm confused. We want the fs {g,u}id to be consistent throughout
the request pipeline and to reflect the actual Ceph user all the way
through the server-side pipeline. What if client.greg is only
authorized to work as uid 12345 and access /users/greg/; why would you
send in a bunch of requests as root just because I mounted that way
inside my own space?

This might be more obvious in the userspace Client, which is already
set up to be friendlier to mapped users for Ganesha etc:
mknod (https://github.com/ceph/ceph/blob/master/src/client/Client.cc#L7297)
and similar calls receive a "UserPerm" from the caller specifying who
the call should be performed as, and they pass that in to the generic
make_requst() function
(https://github.com/ceph/ceph/blob/master/src/client/Client.cc#L1778)
which uses it to set the uid and gid fields you found in the message.
-Greg

> > From my perspective the root of the problem lies in the fact that
> > creation requests implicitly infer the ownership from the {g,u}id that
> > gets sent along with every mds request.
> >
> > I have thought of multiple ways of addressing this problem but the one I
> > prefer is to give all mds requests that create a filesystem object a
> > proper, separate {g,u}id field entry in the argument struct. This is,
> > for example how ->setattr mds requests work.
> >
> > This way the caller's fs{g,u}id can be used consistenly for server
> > access checks and is separated from the ownership for new filesystem
> > objects.
> >
> > Servers could then be updated to refuse creation requests whenever the
> > {g,u}id used for access checking doesn't match the {g,u}id used for
> > creating the filesystem object just as is done for setattr requests on a
> > uid restricted server. But I am, of course, open to other suggestions.
> >
> > Cc: Jeff Layton <[e-mail]>
> > Cc: Ilya Dryomov <[e-mail]>
> > Cc: [e-mail]
> > Signed-off-by: Christian Brauner <[e-mail]>
> > ---
> >  fs/ceph/mds_client.c | 22 ++++++++++++++++++----
> >  1 file changed, 18 insertions(+), 4 deletions(-)
> >
> > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> > index ae2cc4ce1d48..1fb43a8fd64c 100644
> > --- a/fs/ceph/mds_client.c
> > +++ b/fs/ceph/mds_client.c
> > @@ -2459,6 +2459,8 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
> >       void *p, *end;
> >       int ret;
> >       bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
> > +     kuid_t caller_fsuid;
> > +     kgid_t caller_fsgid;
> >
> >       ret = set_request_path_attr(req->r_inode, req->r_dentry,
> >                             req->r_parent, req->r_path1, req->r_ino1.ino,
> > @@ -2524,10 +2526,22 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
> >
> >       head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
> >       head->op = cpu_to_le32(req->r_op);
> > -     head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
> > -                                              req->r_cred->fsuid));
> > -     head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
> > -                                              req->r_cred->fsgid));
> > +     /*
> > +      * Inode operations that create filesystem objects based on the
> > +      * caller's fs{g,u}id like ->mknod(), ->create(), ->mkdir() etc. don't
> > +      * have separate {g,u}id fields in their respective structs in the
> > +      * ceph_mds_request_args union. Instead the caller_{g,u}id field is
> > +      * used to set ownership of the newly created inode by the mds server.
> > +      * For these inode operations we need to send the mapped fs{g,u}id over
> > +      * the wire. For other cases we simple set req->mnt_userns to the
> > +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
> > +      */
> > +     caller_fsuid = mapped_kuid_user(req->mnt_userns, &init_user_ns,
> > +                                     req->r_cred->fsuid);
> > +     caller_fsgid = mapped_kgid_user(req->mnt_userns, &init_user_ns,
> > +                                     req->r_cred->fsgid);
> > +     head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, caller_fsuid));
> > +     head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, caller_fsgid));
> >       head->ino = cpu_to_le64(req->r_deleg_ino);
> >       head->args = req->r_args;
> >
>
> --
> Jeff Layton <[e-mail]>
>

Christian Brauner Jan. 5, 2022, 2:10 p.m. UTC | #3

On Tue, Jan 04, 2022 at 12:40:51PM -0500, Jeff Layton wrote:
> On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > From: Christian Brauner <[e-mail]>
> > 
> > Inode operations that create a new filesystem object such as ->mknod,
> > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > filesystem object.
> > 
> > Cephfs mds creation request argument structures mirror this filesystem
> > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > fs{g,u}id that is always sent as part of any mds request is used by the
> > servers to set the {g,u}id of the new filesystem object.
> > 
> > In order to ensure that the correct {g,u}id is used map the caller's
> > fs{g,u}id for creation requests. This doesn't require complex changes.
> > It suffices to pass in the relevant idmapping recorded in the request
> > message. If this request message was triggered from an inode operation
> > that creates filesystem objects it will have passed down the relevant
> > idmaping. If this is a request message that was triggered from an inode
> > operation that doens't need to take idmappings into account the initial
> > idmapping is passed down which is an identity mapping and thus is
> > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> > 
> > The last few weeks before Christmas 2021 I have spent time not just
> > reading and poking the cephfs kernel code but also took a look at the
> > ceph mds server userspace to ensure I didn't miss some subtlety.
> > 
> > This made me aware of one complication to solve. All requests send the
> > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > server in exactly two cases:
> > 
> > 1. to set the ownership for creation requests
> > 2. to determine whether this client is allowed access on this server
> > 
> > Case 1. we already covered and explained. Case 2. is only relevant for
> > servers where an explicit uid access restriction has been set. That is
> > to say the mds server restricts access to requests coming from a
> > specific uid. Servers without uid restrictions will grant access to
> > requests from any uid by setting MDS_AUTH_UID_ANY.
> > 
> > Case 2. introduces the complication because the caller's fs{g,u}id is
> > not just used to record ownership but also serves as the {g,u}id used
> > when checking access to the server.
> > 
> > Consider a user mounting a cephfs client and creating an idmapped mount
> > from it that maps files owned by uid 1000 to be owned uid 0:
> > 
> > mount -t cephfs -o [...] /unmapped
> > mount-idmapped --map-mount 1000:0:1 /idmapped
> > 
> > That is to say if the mounted cephfs filesystem contains a file "file1"
> > which is owned by uid 1000:
> > 
> > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> >   (One can think of this as the on-disk value.)
> > - looking at it via /idmapped/file1 will report it as owned by uid 0
> > 
> > Now, consider creating new files via the idmapped mount at /idmapped.
> > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > through the idmapped mount mounted at /idmapped it will create a file
> > that is owned by uid 1000 on-disk, i.e.:
> > 
> > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > - looking at it via /idmapped/file2 will report it as owned by uid 0
> > 
> > Now consider an mds server that has a uid access restriction set and
> > only grants access to requests from uid 0.
> > 
> > If the client sends a creation request for a file e.g. /idmapped/file2
> > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > 0 in the idmapped mount and will be sent over the wire allowing the
> > caller access to the mds server.
> > 
> > However, if the caller is not issuing a creation request the caller's
> > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > caller that just successfully created a new file on the restricted mds
> > server sends a request as fs{g,u}id 1000 access will be refused. This
> > however is inconsistent.
> > 
> 
> IDGI, why would you send the fs{g,u}id without the mount's idmapping
> applied in this case? ISTM that idmapping is wholly a client-side
> feature, and that you should always map id's regardless of whether
> you're creating or not.

Since the idmapping is a property of the mount and not a property of the
caller the caller's fs{g,u}id aren't mapped. What is mapped are the
inode's i{g,u}id when accessed from a particular mount.

The fs{g,u}id are only ever mapped when a new filesystem object is
created. So if I have an idmapped mount that makes it so that files
owned by 1000 on-disk appear to be owned by uid 0 then a user with uid 0
creating a new file will create files with uid 1000 on-disk when going
through that mount. For cephfs that'd be the uid we would be sending
with creation requests as I've currently written it.

So then when the user looks at the file it created it will see it as
being owned by uid 0 from that idmapped mount (whereas on-disk it's
1000). But the user's fs{g,u}id isn't per se changed when going through
that mount. So in my opinion I was thinking that the server with access
permissions set would want to always check permissions on the users
"raw" fs{g,u}id. That would mean I'd have to change the patch obviously.
My suggestion was to send the {g,u}id the file will be created with
separately. The alternative would be to not just pass the idmapping into
the creation iop's but into all iops so that we can always map it for
cephfs. But this would mean a lot of vfs changes for one filesystem. So
if we could first explore alternatives approaches I'd be grateful.

(I'll be traveling for the latter half of this week starting today at
CET afternoon so apologies but I'll probably take some time to respond.)

Christian Brauner Jan. 5, 2022, 2:11 p.m. UTC | #4

On Tue, Jan 04, 2022 at 11:33:33AM -0800, Gregory Farnum wrote:
> On Tue, Jan 4, 2022 at 9:41 AM Jeff Layton <[e-mail]> wrote:
> >
> > On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > > From: Christian Brauner <[e-mail]>
> > >
> > > Inode operations that create a new filesystem object such as ->mknod,
> > > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > > filesystem object.
> > >
> > > Cephfs mds creation request argument structures mirror this filesystem
> > > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > > fs{g,u}id that is always sent as part of any mds request is used by the
> > > servers to set the {g,u}id of the new filesystem object.
> > >
> > > In order to ensure that the correct {g,u}id is used map the caller's
> > > fs{g,u}id for creation requests. This doesn't require complex changes.
> > > It suffices to pass in the relevant idmapping recorded in the request
> > > message. If this request message was triggered from an inode operation
> > > that creates filesystem objects it will have passed down the relevant
> > > idmaping. If this is a request message that was triggered from an inode
> > > operation that doens't need to take idmappings into account the initial
> > > idmapping is passed down which is an identity mapping and thus is
> > > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> > >
> > > The last few weeks before Christmas 2021 I have spent time not just
> > > reading and poking the cephfs kernel code but also took a look at the
> > > ceph mds server userspace to ensure I didn't miss some subtlety.
> > >
> > > This made me aware of one complication to solve. All requests send the
> > > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > > server in exactly two cases:
> > >
> > > 1. to set the ownership for creation requests
> > > 2. to determine whether this client is allowed access on this server
> > >
> > > Case 1. we already covered and explained. Case 2. is only relevant for
> > > servers where an explicit uid access restriction has been set. That is
> > > to say the mds server restricts access to requests coming from a
> > > specific uid. Servers without uid restrictions will grant access to
> > > requests from any uid by setting MDS_AUTH_UID_ANY.
> > >
> > > Case 2. introduces the complication because the caller's fs{g,u}id is
> > > not just used to record ownership but also serves as the {g,u}id used
> > > when checking access to the server.
> > >
> > > Consider a user mounting a cephfs client and creating an idmapped mount
> > > from it that maps files owned by uid 1000 to be owned uid 0:
> > >
> > > mount -t cephfs -o [...] /unmapped
> > > mount-idmapped --map-mount 1000:0:1 /idmapped
> > >
> > > That is to say if the mounted cephfs filesystem contains a file "file1"
> > > which is owned by uid 1000:
> > >
> > > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> > >   (One can think of this as the on-disk value.)
> > > - looking at it via /idmapped/file1 will report it as owned by uid 0
> > >
> > > Now, consider creating new files via the idmapped mount at /idmapped.
> > > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > > through the idmapped mount mounted at /idmapped it will create a file
> > > that is owned by uid 1000 on-disk, i.e.:
> > >
> > > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > > - looking at it via /idmapped/file2 will report it as owned by uid 0
> > >
> > > Now consider an mds server that has a uid access restriction set and
> > > only grants access to requests from uid 0.
> > >
> > > If the client sends a creation request for a file e.g. /idmapped/file2
> > > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > > 0 in the idmapped mount and will be sent over the wire allowing the
> > > caller access to the mds server.
> > >
> > > However, if the caller is not issuing a creation request the caller's
> > > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > > caller that just successfully created a new file on the restricted mds
> > > server sends a request as fs{g,u}id 1000 access will be refused. This
> > > however is inconsistent.
> > >
> >
> > IDGI, why would you send the fs{g,u}id without the mount's idmapping
> > applied in this case? ISTM that idmapping is wholly a client-side
> > feature, and that you should always map id's regardless of whether
> > you're creating or not.
> 
> Yeah, I'm confused. We want the fs {g,u}id to be consistent throughout
> the request pipeline and to reflect the actual Ceph user all the way
> through the server-side pipeline. What if client.greg is only
> authorized to work as uid 12345 and access /users/greg/; why would you
> send in a bunch of requests as root just because I mounted that way
> inside my own space?
> 
> This might be more obvious in the userspace Client, which is already
> set up to be friendlier to mapped users for Ganesha etc:
> mknod (https://github.com/ceph/ceph/blob/master/src/client/Client.cc#L7297)
> and similar calls receive a "UserPerm" from the caller specifying who
> the call should be performed as, and they pass that in to the generic
> make_requst() function
> (https://github.com/ceph/ceph/blob/master/src/client/Client.cc#L1778)
> which uses it to set the uid and gid fields you found in the message.

Thank you for those links. I think I read through this code before and
I'll give it another read.

Jeff Layton Jan. 5, 2022, 3:03 p.m. UTC | #5

On Wed, 2022-01-05 at 15:10 +0100, Christian Brauner wrote:
> On Tue, Jan 04, 2022 at 12:40:51PM -0500, Jeff Layton wrote:
> > On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > > From: Christian Brauner <[e-mail]>
> > > 
> > > Inode operations that create a new filesystem object such as ->mknod,
> > > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > > filesystem object.
> > > 
> > > Cephfs mds creation request argument structures mirror this filesystem
> > > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > > fs{g,u}id that is always sent as part of any mds request is used by the
> > > servers to set the {g,u}id of the new filesystem object.
> > > 
> > > In order to ensure that the correct {g,u}id is used map the caller's
> > > fs{g,u}id for creation requests. This doesn't require complex changes.
> > > It suffices to pass in the relevant idmapping recorded in the request
> > > message. If this request message was triggered from an inode operation
> > > that creates filesystem objects it will have passed down the relevant
> > > idmaping. If this is a request message that was triggered from an inode
> > > operation that doens't need to take idmappings into account the initial
> > > idmapping is passed down which is an identity mapping and thus is
> > > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> > > 
> > > The last few weeks before Christmas 2021 I have spent time not just
> > > reading and poking the cephfs kernel code but also took a look at the
> > > ceph mds server userspace to ensure I didn't miss some subtlety.
> > > 
> > > This made me aware of one complication to solve. All requests send the
> > > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > > server in exactly two cases:
> > > 
> > > 1. to set the ownership for creation requests
> > > 2. to determine whether this client is allowed access on this server
> > > 
> > > Case 1. we already covered and explained. Case 2. is only relevant for
> > > servers where an explicit uid access restriction has been set. That is
> > > to say the mds server restricts access to requests coming from a
> > > specific uid. Servers without uid restrictions will grant access to
> > > requests from any uid by setting MDS_AUTH_UID_ANY.
> > > 
> > > Case 2. introduces the complication because the caller's fs{g,u}id is
> > > not just used to record ownership but also serves as the {g,u}id used
> > > when checking access to the server.
> > > 
> > > Consider a user mounting a cephfs client and creating an idmapped mount
> > > from it that maps files owned by uid 1000 to be owned uid 0:
> > > 
> > > mount -t cephfs -o [...] /unmapped
> > > mount-idmapped --map-mount 1000:0:1 /idmapped
> > > 
> > > That is to say if the mounted cephfs filesystem contains a file "file1"
> > > which is owned by uid 1000:
> > > 
> > > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> > >   (One can think of this as the on-disk value.)
> > > - looking at it via /idmapped/file1 will report it as owned by uid 0
> > > 
> > > Now, consider creating new files via the idmapped mount at /idmapped.
> > > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > > through the idmapped mount mounted at /idmapped it will create a file
> > > that is owned by uid 1000 on-disk, i.e.:
> > > 
> > > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > > - looking at it via /idmapped/file2 will report it as owned by uid 0
> > > 
> > > Now consider an mds server that has a uid access restriction set and
> > > only grants access to requests from uid 0.
> > > 
> > > If the client sends a creation request for a file e.g. /idmapped/file2
> > > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > > 0 in the idmapped mount and will be sent over the wire allowing the
> > > caller access to the mds server.
> > > 
> > > However, if the caller is not issuing a creation request the caller's
> > > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > > caller that just successfully created a new file on the restricted mds
> > > server sends a request as fs{g,u}id 1000 access will be refused. This
> > > however is inconsistent.
> > > 
> > 
> > IDGI, why would you send the fs{g,u}id without the mount's idmapping
> > applied in this case? ISTM that idmapping is wholly a client-side
> > feature, and that you should always map id's regardless of whether
> > you're creating or not.
> 
> Since the idmapping is a property of the mount and not a property of the
> caller the caller's fs{g,u}id aren't mapped. What is mapped are the
> inode's i{g,u}id when accessed from a particular mount.
> 
> The fs{g,u}id are only ever mapped when a new filesystem object is
> created. So if I have an idmapped mount that makes it so that files
> owned by 1000 on-disk appear to be owned by uid 0 then a user with uid 0
> creating a new file will create files with uid 1000 on-disk when going
> through that mount. For cephfs that'd be the uid we would be sending
> with creation requests as I've currently written it.
> 
> So then when the user looks at the file it created it will see it as
> being owned by uid 0 from that idmapped mount (whereas on-disk it's
> 1000). But the user's fs{g,u}id isn't per se changed when going through
> that mount. So in my opinion I was thinking that the server with access
> permissions set would want to always check permissions on the users
> "raw" fs{g,u}id. That would mean I'd have to change the patch obviously.
> My suggestion was to send the {g,u}id the file will be created with
> separately. The alternative would be to not just pass the idmapping into
> the creation iop's but into all iops so that we can always map it for
> cephfs. But this would mean a lot of vfs changes for one filesystem. So
> if we could first explore alternatives approaches I'd be grateful.
> 

You'll probably need to do this for NFS anyway, if you have plans in
that direction. Extending the protocol there will be much more
difficult. I think that approach sounds much cleaner overall.

> (I'll be traveling for the latter half of this week starting today at
> CET afternoon so apologies but I'll probably take some time to respond.)

Ok. I guess you can get away with this on a local fs because the backend
storage doesn't really care about uid/gids at all. The only permission
checking is done in the kernel and you (presumably) can just map the
inode's uid/gid prior to checking permissions.

I'm a little confused as to what you mean by "raw" id here. In your
earlier example with a mapping of 1000:0:1, which one is the raw id for
the actual user?

Christian Brauner Jan. 5, 2022, 3:35 p.m. UTC | #6

On Wed, Jan 05, 2022 at 10:03:06AM -0500, Jeff Layton wrote:
> On Wed, 2022-01-05 at 15:10 +0100, Christian Brauner wrote:
> > On Tue, Jan 04, 2022 at 12:40:51PM -0500, Jeff Layton wrote:
> > > On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > > > From: Christian Brauner <[e-mail]>
> > > > 
> > > > Inode operations that create a new filesystem object such as ->mknod,
> > > > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > > > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > > > filesystem object.
> > > > 
> > > > Cephfs mds creation request argument structures mirror this filesystem
> > > > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > > > fs{g,u}id that is always sent as part of any mds request is used by the
> > > > servers to set the {g,u}id of the new filesystem object.
> > > > 
> > > > In order to ensure that the correct {g,u}id is used map the caller's
> > > > fs{g,u}id for creation requests. This doesn't require complex changes.
> > > > It suffices to pass in the relevant idmapping recorded in the request
> > > > message. If this request message was triggered from an inode operation
> > > > that creates filesystem objects it will have passed down the relevant
> > > > idmaping. If this is a request message that was triggered from an inode
> > > > operation that doens't need to take idmappings into account the initial
> > > > idmapping is passed down which is an identity mapping and thus is
> > > > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> > > > 
> > > > The last few weeks before Christmas 2021 I have spent time not just
> > > > reading and poking the cephfs kernel code but also took a look at the
> > > > ceph mds server userspace to ensure I didn't miss some subtlety.
> > > > 
> > > > This made me aware of one complication to solve. All requests send the
> > > > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > > > server in exactly two cases:
> > > > 
> > > > 1. to set the ownership for creation requests
> > > > 2. to determine whether this client is allowed access on this server
> > > > 
> > > > Case 1. we already covered and explained. Case 2. is only relevant for
> > > > servers where an explicit uid access restriction has been set. That is
> > > > to say the mds server restricts access to requests coming from a
> > > > specific uid. Servers without uid restrictions will grant access to
> > > > requests from any uid by setting MDS_AUTH_UID_ANY.
> > > > 
> > > > Case 2. introduces the complication because the caller's fs{g,u}id is
> > > > not just used to record ownership but also serves as the {g,u}id used
> > > > when checking access to the server.
> > > > 
> > > > Consider a user mounting a cephfs client and creating an idmapped mount
> > > > from it that maps files owned by uid 1000 to be owned uid 0:
> > > > 
> > > > mount -t cephfs -o [...] /unmapped
> > > > mount-idmapped --map-mount 1000:0:1 /idmapped
> > > > 
> > > > That is to say if the mounted cephfs filesystem contains a file "file1"
> > > > which is owned by uid 1000:
> > > > 
> > > > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> > > >   (One can think of this as the on-disk value.)
> > > > - looking at it via /idmapped/file1 will report it as owned by uid 0
> > > > 
> > > > Now, consider creating new files via the idmapped mount at /idmapped.
> > > > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > > > through the idmapped mount mounted at /idmapped it will create a file
> > > > that is owned by uid 1000 on-disk, i.e.:
> > > > 
> > > > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > > > - looking at it via /idmapped/file2 will report it as owned by uid 0
> > > > 
> > > > Now consider an mds server that has a uid access restriction set and
> > > > only grants access to requests from uid 0.
> > > > 
> > > > If the client sends a creation request for a file e.g. /idmapped/file2
> > > > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > > > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > > > 0 in the idmapped mount and will be sent over the wire allowing the
> > > > caller access to the mds server.
> > > > 
> > > > However, if the caller is not issuing a creation request the caller's
> > > > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > > > caller that just successfully created a new file on the restricted mds
> > > > server sends a request as fs{g,u}id 1000 access will be refused. This
> > > > however is inconsistent.
> > > > 
> > > 
> > > IDGI, why would you send the fs{g,u}id without the mount's idmapping
> > > applied in this case? ISTM that idmapping is wholly a client-side
> > > feature, and that you should always map id's regardless of whether
> > > you're creating or not.
> > 
> > Since the idmapping is a property of the mount and not a property of the
> > caller the caller's fs{g,u}id aren't mapped. What is mapped are the
> > inode's i{g,u}id when accessed from a particular mount.
> > 
> > The fs{g,u}id are only ever mapped when a new filesystem object is
> > created. So if I have an idmapped mount that makes it so that files
> > owned by 1000 on-disk appear to be owned by uid 0 then a user with uid 0
> > creating a new file will create files with uid 1000 on-disk when going
> > through that mount. For cephfs that'd be the uid we would be sending
> > with creation requests as I've currently written it.
> > 
> > So then when the user looks at the file it created it will see it as
> > being owned by uid 0 from that idmapped mount (whereas on-disk it's
> > 1000). But the user's fs{g,u}id isn't per se changed when going through
> > that mount. So in my opinion I was thinking that the server with access
> > permissions set would want to always check permissions on the users
> > "raw" fs{g,u}id. That would mean I'd have to change the patch obviously.
> > My suggestion was to send the {g,u}id the file will be created with
> > separately. The alternative would be to not just pass the idmapping into
> > the creation iop's but into all iops so that we can always map it for
> > cephfs. But this would mean a lot of vfs changes for one filesystem. So
> > if we could first explore alternatives approaches I'd be grateful.
> > 
> 
> You'll probably need to do this for NFS anyway, if you have plans in
> that direction. Extending the protocol there will be much more
> difficult. I think that approach sounds much cleaner overall.

Ok. Is it ok if I take a little while to work on this?
I have some other work I need to be looking at first and then I have
Februrary "free".

> 
> > (I'll be traveling for the latter half of this week starting today at
> > CET afternoon so apologies but I'll probably take some time to respond.)
> 
> Ok. I guess you can get away with this on a local fs because the backend
> storage doesn't really care about uid/gids at all. The only permission
> checking is done in the kernel and you (presumably) can just map the
> inode's uid/gid prior to checking permissions.

Yes, we always map the inode as that's semantically cleaner and easier
to reason about in my opinion.

> 
> I'm a little confused as to what you mean by "raw" id here. In your
> earlier example with a mapping of 1000:0:1, which one is the raw id for
> the actual user?

Oh, sorry. In this context I really just meant the values gotten from
current_fs{g,u}id() as they are sent now.

Alexander Mikhalitsyn May 19, 2023, 11:44 a.m. UTC | #7

Dear colleagues,

On Wed, 5 Jan 2022 15:10:23 +0100
Christian Brauner <[e-mail]> wrote:

> On Tue, Jan 04, 2022 at 12:40:51PM -0500, Jeff Layton wrote:
> > On Tue, 2022-01-04 at 15:04 +0100, Christian Brauner wrote:
> > > From: Christian Brauner <[e-mail]>
> > > 
> > > Inode operations that create a new filesystem object such as ->mknod,
> > > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly.
> > > Instead the caller's fs{g,u}id is used for the {g,u}id of the new
> > > filesystem object.
> > > 
> > > Cephfs mds creation request argument structures mirror this filesystem
> > > behavior. They don't encode a {g,u}id explicitly. Instead the caller's
> > > fs{g,u}id that is always sent as part of any mds request is used by the
> > > servers to set the {g,u}id of the new filesystem object.
> > > 
> > > In order to ensure that the correct {g,u}id is used map the caller's
> > > fs{g,u}id for creation requests. This doesn't require complex changes.
> > > It suffices to pass in the relevant idmapping recorded in the request
> > > message. If this request message was triggered from an inode operation
> > > that creates filesystem objects it will have passed down the relevant
> > > idmaping. If this is a request message that was triggered from an inode
> > > operation that doens't need to take idmappings into account the initial
> > > idmapping is passed down which is an identity mapping and thus is
> > > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent.
> > > 
> > > The last few weeks before Christmas 2021 I have spent time not just
> > > reading and poking the cephfs kernel code but also took a look at the
> > > ceph mds server userspace to ensure I didn't miss some subtlety.
> > > 
> > > This made me aware of one complication to solve. All requests send the
> > > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the
> > > server in exactly two cases:
> > > 
> > > 1. to set the ownership for creation requests
> > > 2. to determine whether this client is allowed access on this server
> > > 
> > > Case 1. we already covered and explained. Case 2. is only relevant for
> > > servers where an explicit uid access restriction has been set. That is
> > > to say the mds server restricts access to requests coming from a
> > > specific uid. Servers without uid restrictions will grant access to
> > > requests from any uid by setting MDS_AUTH_UID_ANY.
> > > 
> > > Case 2. introduces the complication because the caller's fs{g,u}id is
> > > not just used to record ownership but also serves as the {g,u}id used
> > > when checking access to the server.
> > > 
> > > Consider a user mounting a cephfs client and creating an idmapped mount
> > > from it that maps files owned by uid 1000 to be owned uid 0:
> > > 
> > > mount -t cephfs -o [...] /unmapped
> > > mount-idmapped --map-mount 1000:0:1 /idmapped
> > > 
> > > That is to say if the mounted cephfs filesystem contains a file "file1"
> > > which is owned by uid 1000:
> > > 
> > > - looking at it via /unmapped/file1 will report it as owned by uid 1000
> > >   (One can think of this as the on-disk value.)
> > > - looking at it via /idmapped/file1 will report it as owned by uid 0
> > > 
> > > Now, consider creating new files via the idmapped mount at /idmapped.
> > > When a caller with fs{g,u}id 1000 creates a file "file2" by going
> > > through the idmapped mount mounted at /idmapped it will create a file
> > > that is owned by uid 1000 on-disk, i.e.:
> > > 
> > > - looking at it via /unmapped/file2 will report it as owned by uid 1000
> > > - looking at it via /idmapped/file2 will report it as owned by uid 0
> > > 
> > > Now consider an mds server that has a uid access restriction set and
> > > only grants access to requests from uid 0.
> > > 
> > > If the client sends a creation request for a file e.g. /idmapped/file2
> > > it will send the caller's fs{g,u}id idmapped according to the idmapped
> > > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id
> > > 0 in the idmapped mount and will be sent over the wire allowing the
> > > caller access to the mds server.
> > > 
> > > However, if the caller is not issuing a creation request the caller's
> > > fs{g,u}id will be send without the mount's idmapping applied. So if the
> > > caller that just successfully created a new file on the restricted mds
> > > server sends a request as fs{g,u}id 1000 access will be refused. This
> > > however is inconsistent.
> > > 
> > 
> > IDGI, why would you send the fs{g,u}id without the mount's idmapping
> > applied in this case? ISTM that idmapping is wholly a client-side
> > feature, and that you should always map id's regardless of whether
> > you're creating or not.
> 
> Since the idmapping is a property of the mount and not a property of the
> caller the caller's fs{g,u}id aren't mapped. What is mapped are the
> inode's i{g,u}id when accessed from a particular mount.
> 
> The fs{g,u}id are only ever mapped when a new filesystem object is
> created. So if I have an idmapped mount that makes it so that files
> owned by 1000 on-disk appear to be owned by uid 0 then a user with uid 0
> creating a new file will create files with uid 1000 on-disk when going
> through that mount. For cephfs that'd be the uid we would be sending
> with creation requests as I've currently written it.
> 
> So then when the user looks at the file it created it will see it as
> being owned by uid 0 from that idmapped mount (whereas on-disk it's
> 1000). But the user's fs{g,u}id isn't per se changed when going through
> that mount. So in my opinion I was thinking that the server with access
> permissions set would want to always check permissions on the users
> "raw" fs{g,u}id. That would mean I'd have to change the patch obviously.
> My suggestion was to send the {g,u}id the file will be created with
> separately. The alternative would be to not just pass the idmapping into
> the creation iop's but into all iops so that we can always map it for
> cephfs. But this would mean a lot of vfs changes for one filesystem. So
> if we could first explore alternatives approaches I'd be grateful.

I can't understand which kind of operations we are talking about here.
Right now almost all inode_operations are taking struct mnt_idmap as a parameter
(at the moment of this series was posted it was struct user_namespace, but that's not important).

The only iops those are not taking idmap is lookup, readlink, fiemap, update_time, atomic_open
and a few more. So, we want to pass struct mnt_idmap to them to always map current_fs{g,u}id
according to a mount idmapping? As Christian pointed above:

> Since the idmapping is a property of the mount and not a property of the
> caller the caller's fs{g,u}id aren't mapped. What is mapped are the
> inode's i{g,u}id when accessed from a particular mount.

If we want to go this way then we don't need to pass mnt_idmap to any additional inode ops
and the current approach works fine. Please, correct me if I'm wrong.

> 
> (I'll be traveling for the latter half of this week starting today at
> CET afternoon so apologies but I'll probably take some time to respond.)
>

P.S.

I'm trying to make a respin for this series, I've made a formal rebase on top of
the current Linux kernel tree and fixed it according to the Jeff's review comment:
https://lore.kernel.org/all/[e-mail]/

This thing is really important for LXD/LXC project so I'll be happy to help with pushing
this forward.

Current tree:
https://github.com/mihalicyn/linux/commits/fs.idmapped.ceph.v2

Kind regards,
Alex

Patch

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index ae2cc4ce1d48..1fb43a8fd64c 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -2459,6 +2459,8 @@  static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
 	void *p, *end;
 	int ret;
 	bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME);
+	kuid_t caller_fsuid;
+	kgid_t caller_fsgid;
 
 	ret = set_request_path_attr(req->r_inode, req->r_dentry,
 			      req->r_parent, req->r_path1, req->r_ino1.ino,
@@ -2524,10 +2526,22 @@  static struct ceph_msg *create_request_message(struct ceph_mds_session *session,
 
 	head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch);
 	head->op = cpu_to_le32(req->r_op);
-	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns,
-						 req->r_cred->fsuid));
-	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns,
-						 req->r_cred->fsgid));
+	/*
+	 * Inode operations that create filesystem objects based on the
+	 * caller's fs{g,u}id like ->mknod(), ->create(), ->mkdir() etc. don't
+	 * have separate {g,u}id fields in their respective structs in the
+	 * ceph_mds_request_args union. Instead the caller_{g,u}id field is
+	 * used to set ownership of the newly created inode by the mds server.
+	 * For these inode operations we need to send the mapped fs{g,u}id over
+	 * the wire. For other cases we simple set req->mnt_userns to the
+	 * initial idmapping meaning the unmapped fs{g,u}id is sent.
+	 */
+	caller_fsuid = mapped_kuid_user(req->mnt_userns, &init_user_ns,
+					req->r_cred->fsuid);
+	caller_fsgid = mapped_kgid_user(req->mnt_userns, &init_user_ns,
+					req->r_cred->fsgid);
+	head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, caller_fsuid));
+	head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, caller_fsgid));
 	head->ino = cpu_to_le64(req->r_deleg_ino);
 	head->args = req->r_args;