[v2,03/13] ceph: handle idmapped mounts in create_request_message()

Commit Message

Aleksandr Mikhalitsyn May 24, 2023, 3:33 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]>
Signed-off-by: Alexander Mikhalitsyn <[e-mail]>
---
 fs/ceph/mds_client.c | 22 ++++++++++++++++++----
 1 file changed, 18 insertions(+), 4 deletions(-)

Comments

Xiubo Li May 29, 2023, 3:52 a.m. UTC | #1

On 5/24/23 23:33, Alexander Mikhalitsyn 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.
>
>  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]>
> Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
> --- a/fs/ceph/mds_client.c
> +++ b/fs/ceph/mds_client.c
> @@ -2583,6 +2583,8 @@ 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,
> @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
> +	 * initial idmapping meaning the unmapped fs{g,u}id is sent.
> +	 */
> +	caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
> +					VFSUIDT_INIT(req->r_cred->fsuid));
> +	caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
> +					VFSGIDT_INIT(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));

Hi Alexander,

You didn't answer Jeff and Greg's concerns in the first version 
https://www.spinics.net/lists/ceph-devel/msg53356.html.

I am also confused as Greg mentioned. If we just map the ids as 1000:0 
and created a file and then map the ids 1000:10, then the file couldn't 
be accessible, right ? Is this normal and as expected ?

IMO the idmapping should be client-side feature and we should make it 
consistent by using the unmapped fs{g,u}id always here.

Thanks

- Xiubo

>   	head->ino = cpu_to_le64(req->r_deleg_ino);
>   	head->args = req->r_args;
>

Aleksandr Mikhalitsyn May 31, 2023, 4:32 p.m. UTC | #2

On Mon, May 29, 2023 at 5:52 AM Xiubo Li <[e-mail]> wrote:
>
>
> On 5/24/23 23:33, Alexander Mikhalitsyn 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.
> >
> >  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]>
> > Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
> > --- a/fs/ceph/mds_client.c
> > +++ b/fs/ceph/mds_client.c
> > @@ -2583,6 +2583,8 @@ 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,
> > @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
> > +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
> > +      */
> > +     caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
> > +                                     VFSUIDT_INIT(req->r_cred->fsuid));
> > +     caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
> > +                                     VFSGIDT_INIT(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));
>
> Hi Alexander,

Dear Xiubo,

Thanks for paying attention to this series!

>
> You didn't answer Jeff and Greg's concerns in the first version
> https://www.spinics.net/lists/ceph-devel/msg53356.html.

I've tried to respin discussion in the -v1 thread:
https://lore.kernel.org/all/[e-mail]/

No one replied, so I decided to send rebased and slightly changed -v2,
where I've fixed this:
https://lore.kernel.org/all/[e-mail]/

>
> I am also confused as Greg mentioned. If we just map the ids as 1000:0
> and created a file and then map the ids 1000:10, then the file couldn't
> be accessible, right ? Is this normal and as expected ?

This can be a problem only if filtering based on the UID is turned on
on the server side (which is a relatively rare case).

idmapped mounts are not about mapping a caller UID/GID, idmapped
mounts are about mapping inode owner's UID/GID.
So, for example if you have UID 1000 (on disk) and have an idmapping
1000:0 then it will be shown as owned by 0.
If you create a file from a user with UID 0 then you will get UID 1000
on disk. To achieve that, we map a current user fs{g,u}id
when sending a creation request according to the idmapping mount to
make things consistent. But when a user opens a file,
we are sending UID/GID as they are without applying an idmapping. Of
course, generic_permission() kernel helper is aware of
mount idmapping and before open request will go to the server we will
check that current user is allowed to open this file (and during
this check UID/GID of a current user and UID/GID of the file owner
will be properly compared). I.e. this issue is only relevant for the
case
when we have additional permission checks on the network file system
server side.

>
> IMO the idmapping should be client-side feature and we should make it
> consistent by using the unmapped fs{g,u}id always here.

To make the current user fs{g,u}id always idmapped we need to make
really big changes in the VFS layer. And it's not obvious
that it justifies the cost. Because this particular feature with
Cephfs idmapped mounts is already used/tested with LXD/LXC workloads
and it works perfectly well. And as far as I know, LXD/LXC were the
first idmapped mount adopters. IMHO, it's better to
start from this approach and if someone will want to extend this
functionality for network filesystems and want to map fs{g,u}id which
are sent over the
wire we will take a look at that. Because anyway, integration with
Cephfs is important for the LXD project and we are looking closely at
this.

Kind regards,
Alex

>
> Thanks
>
> - Xiubo
>
> >       head->ino = cpu_to_le64(req->r_deleg_ino);
> >       head->args = req->r_args;
> >
>

Xiubo Li June 1, 2023, 2:29 a.m. UTC | #3

On 6/1/23 00:32, Aleksandr Mikhalitsyn wrote:
> On Mon, May 29, 2023 at 5:52 AM Xiubo Li <[e-mail]> wrote:
>>
>> On 5/24/23 23:33, Alexander Mikhalitsyn 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.
>>>
>>>   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]>
>>> Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
>>> --- a/fs/ceph/mds_client.c
>>> +++ b/fs/ceph/mds_client.c
>>> @@ -2583,6 +2583,8 @@ 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,
>>> @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
>>> +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
>>> +      */
>>> +     caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
>>> +                                     VFSUIDT_INIT(req->r_cred->fsuid));
>>> +     caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
>>> +                                     VFSGIDT_INIT(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));
>> Hi Alexander,
> Dear Xiubo,
>
> Thanks for paying attention to this series!
>
>> You didn't answer Jeff and Greg's concerns in the first version
>> https://www.spinics.net/lists/ceph-devel/msg53356.html.
> I've tried to respin discussion in the -v1 thread:
> https://lore.kernel.org/all/[e-mail]/
>
> No one replied, so I decided to send rebased and slightly changed -v2,
> where I've fixed this:
> https://lore.kernel.org/all/[e-mail]/
>
>> I am also confused as Greg mentioned. If we just map the ids as 1000:0
>> and created a file and then map the ids 1000:10, then the file couldn't
>> be accessible, right ? Is this normal and as expected ?
> This can be a problem only if filtering based on the UID is turned on
> on the server side (which is a relatively rare case).
>
> idmapped mounts are not about mapping a caller UID/GID, idmapped
> mounts are about mapping inode owner's UID/GID.
> So, for example if you have UID 1000 (on disk) and have an idmapping
> 1000:0 then it will be shown as owned by 0.

My understanding was that on the disk the files' owner UID should be 
1000 always, while in the client side it will show file's owner as the 
mapped UID 0 with an idmapping 1000:0.

This should be the same as what you mentioned above, right ?

> If you create a file from a user with UID 0 then you will get UID 1000
> on disk. To achieve that, we map a current user fs{g,u}id
> when sending a creation request according to the idmapping mount to
> make things consistent.

As you know the cephfs MDSs will use the creation requests' caller UID 
as the owner's UID when creating new inodes.

Which means that if the creation requests switches to use the mapped UID 
0 as the caller UID then the file's owner will be UID 0 instead of UID 
1000 in cephfs MDSs. Does this what this patch want to do ?


>   But when a user opens a file,
> we are sending UID/GID as they are without applying an idmapping.

If my understanding is correct above, then when opening the file with 
non-mapped UID 1000 it may fail because the files' owner is UID 0.

Correct me if my understanding is wrong.

>   Of
> course, generic_permission() kernel helper is aware of
> mount idmapping

Yeah, this was also what I thought it should be.

There is another client auth feature [1] for cephfs. The MDS will allow 
us to set a path restriction for specify UID, more detail please see [2]:

  allow rw path=/dir1 uid=1000 gids=1000

This may cause the creation requests to fail if you set the caller UID 
to the mapped UID.


[1] https://docs.ceph.com/en/latest/cephfs/client-auth/
[2] https://tracker.ceph.com/issues/59388


Thanks

- Xiubo

> and before open request will go to the server we will
> check that current user is allowed to open this file (and during
> this check UID/GID of a current user and UID/GID of the file owner
> will be properly compared). I.e. this issue is only relevant for the
> case
> when we have additional permission checks on the network file system
> server side.
>
>> IMO the idmapping should be client-side feature and we should make it
>> consistent by using the unmapped fs{g,u}id always here.
> To make the current user fs{g,u}id always idmapped we need to make
> really big changes in the VFS layer. And it's not obvious
> that it justifies the cost. Because this particular feature with
> Cephfs idmapped mounts is already used/tested with LXD/LXC workloads
> and it works perfectly well. And as far as I know, LXD/LXC were the
> first idmapped mount adopters. IMHO, it's better to
> start from this approach and if someone will want to extend this
> functionality for network filesystems and want to map fs{g,u}id which
> are sent over the
> wire we will take a look at that. Because anyway, integration with
> Cephfs is important for the LXD project and we are looking closely at
> this.
>
> Kind regards,
> Alex
>
>> Thanks
>>
>> - Xiubo
>>
>>>        head->ino = cpu_to_le64(req->r_deleg_ino);
>>>        head->args = req->r_args;
>>>

Aleksandr Mikhalitsyn June 1, 2023, 6:29 p.m. UTC | #4

On Thu, Jun 1, 2023 at 4:29 AM Xiubo Li <[e-mail]> wrote:
>
>
> On 6/1/23 00:32, Aleksandr Mikhalitsyn wrote:
> > On Mon, May 29, 2023 at 5:52 AM Xiubo Li <[e-mail]> wrote:
> >>
> >> On 5/24/23 23:33, Alexander Mikhalitsyn 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.
> >>>
> >>>   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]>
> >>> Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
> >>> --- a/fs/ceph/mds_client.c
> >>> +++ b/fs/ceph/mds_client.c
> >>> @@ -2583,6 +2583,8 @@ 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,
> >>> @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
> >>> +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
> >>> +      */
> >>> +     caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
> >>> +                                     VFSUIDT_INIT(req->r_cred->fsuid));
> >>> +     caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
> >>> +                                     VFSGIDT_INIT(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));
> >> Hi Alexander,
> > Dear Xiubo,
> >
> > Thanks for paying attention to this series!
> >
> >> You didn't answer Jeff and Greg's concerns in the first version
> >> https://www.spinics.net/lists/ceph-devel/msg53356.html.
> > I've tried to respin discussion in the -v1 thread:
> > https://lore.kernel.org/all/[e-mail]/
> >
> > No one replied, so I decided to send rebased and slightly changed -v2,
> > where I've fixed this:
> > https://lore.kernel.org/all/[e-mail]/
> >
> >> I am also confused as Greg mentioned. If we just map the ids as 1000:0
> >> and created a file and then map the ids 1000:10, then the file couldn't
> >> be accessible, right ? Is this normal and as expected ?
> > This can be a problem only if filtering based on the UID is turned on
> > on the server side (which is a relatively rare case).
> >
> > idmapped mounts are not about mapping a caller UID/GID, idmapped
> > mounts are about mapping inode owner's UID/GID.
> > So, for example if you have UID 1000 (on disk) and have an idmapping
> > 1000:0 then it will be shown as owned by 0.
>
> My understanding was that on the disk the files' owner UID should be
> 1000 always, while in the client side it will show file's owner as the
> mapped UID 0 with an idmapping 1000:0.

Hi, Xiubo!

>
> This should be the same as what you mentioned above, right ?

Right.

Let me show a real output from a real command line experiment :-)

1. Mount cephfs

mount.ceph [e-mail]=/ /mnt/ceph -o
mon_addr=127.0.0.1:6789,secret=very_secret_key

2. Make 1000:1000 a root dentry owner (it will be convenient because
we want to use mapping 1000:0:1 for simplicity)

chown 1000:1000 /mnt/ceph

3. create an idmapped mount based on a regular /mnt/ceph mount using a
mount-idmapped tool that was written by Christian.
[ taken from https://raw.githubusercontent.com/brauner/mount-idmapped/master/mount-idmapped.c
]

./mount-idmapped --map-mount b:1000:0:1 /mnt/ceph /mnt/ceph_idmapped

"b" stands for "both", so we are creating a mapping of length 1 for
both UID and GID.
1000 is a UID/GID "on-disk", 0 is a mapped UID/GID.

4. Just to be precise, let's look at which UID/GID we have now.

root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
total 4
drwxrwxrwx 2 1000 1000    0 Jun  1 17:51 .
drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..

root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
total 4
drwxrwxrwx 2 0 0    0 Jun  1 17:51 .
drwxr-xr-x 4 0 0 4096 Jun  1 16:55 ..

5. Now let's create a bunch of files with different owners and through
different mounts (idmapped/non-idmapped).

5.1. Create a file from 0:0 through the idmapped mount (it should
appear as 1000:1000 on disk)
root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid0

5.2. Create a file from 1000:1000 through the idmapped mount (should
fail because 1000:1000 is not a valid UID/GID as it can't be mapped
back to the "on-disk" UID/GID set).
root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000
touch: cannot touch
'/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000': Value
too large for defined data type

... and we've got EOVERFLOW. That's correct!

5.3. Create a file from 0:0 but through the regular mount. (it should
appear as overflowuid(=65534) in idmapped mount, because 0:0 on-disk
is not mapped to the UID/GID set).

root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
/mnt/ceph/created_directly_with_uid0

5.4. Create a file from 1000:1000 but through the regular mount. (it
should appear as 0:0 in idmapped mount, because 1000 (on-disk) mapped
to 0).

root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
/mnt/ceph/created_directly_with_uid1000

6. Now let's look on the result:

root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
total 4
drwxrwxrwx 2 1000 1000    3 Jun  1 17:54 .
drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..
-rw-r--r-- 1    0    0    0 Jun  1 17:54 created_directly_with_uid0
-rw-rw-r-- 1 1000 1000    0 Jun  1 17:54 created_directly_with_uid1000
-rw-r--r-- 1 1000 1000    0 Jun  1 17:53 created_through_idmapped_mnt_with_uid0

root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
total 4
drwxrwxrwx 2     0     0    3 Jun  1 17:54 .
drwxr-xr-x 4     0     0 4096 Jun  1 16:55 ..
-rw-r--r-- 1 65534 65534    0 Jun  1 17:54 created_directly_with_uid0
-rw-rw-r-- 1     0     0    0 Jun  1 17:54 created_directly_with_uid1000
-rw-r--r-- 1     0     0    0 Jun  1 17:53
created_through_idmapped_mnt_with_uid0

>
> > If you create a file from a user with UID 0 then you will get UID 1000
> > on disk. To achieve that, we map a current user fs{g,u}id
> > when sending a creation request according to the idmapping mount to
> > make things consistent.
>
> As you know the cephfs MDSs will use the creation requests' caller UID
> as the owner's UID when creating new inodes.

Yes, that's why we have to map a caller UID to end up with the correct
value of a file owner.

>
> Which means that if the creation requests switches to use the mapped UID
> 0 as the caller UID then the file's owner will be UID 0 instead of UID
> 1000 in cephfs MDSs. Does this what this patch want to do ?

In my example we have a caller with UID equal 0, then the mapped UID
will be 1000. So, the file will be created with UID = 1000.

>
>
> >   But when a user opens a file,
> > we are sending UID/GID as they are without applying an idmapping.
>
> If my understanding is correct above, then when opening the file with
> non-mapped UID 1000 it may fail because the files' owner is UID 0.
>
> Correct me if my understanding is wrong.
>
> >   Of
> > course, generic_permission() kernel helper is aware of
> > mount idmapping
>
> Yeah, this was also what I thought it should be.
>
> There is another client auth feature [1] for cephfs. The MDS will allow
> us to set a path restriction for specify UID, more detail please see [2]:
>
>   allow rw path=/dir1 uid=1000 gids=1000
>
> This may cause the creation requests to fail if you set the caller UID
> to the mapped UID.

Yes, that can be a problem of course. But it will only affect users
who want to use this feature and it doesn't open any security holes.
It's just a limitation of this approach. Unfortunately it's barely
fixable without massive VFS changes and until we have no real use
cases
for this combination of idmapped mounts + MDS UID/GID-based path
restriction we are not sure that it makes sense to implement this
right now.

>
>
> [1] https://docs.ceph.com/en/latest/cephfs/client-auth/
> [2] https://tracker.ceph.com/issues/59388

Thanks, I'll take a look closer at that!

Thanks for closely looking into this patchset, Xiubo!

Kind regards,
Alex

>
>
> Thanks
>
> - Xiubo
>
> > and before open request will go to the server we will
> > check that current user is allowed to open this file (and during
> > this check UID/GID of a current user and UID/GID of the file owner
> > will be properly compared). I.e. this issue is only relevant for the
> > case
> > when we have additional permission checks on the network file system
> > server side.
> >
> >> IMO the idmapping should be client-side feature and we should make it
> >> consistent by using the unmapped fs{g,u}id always here.
> > To make the current user fs{g,u}id always idmapped we need to make
> > really big changes in the VFS layer. And it's not obvious
> > that it justifies the cost. Because this particular feature with
> > Cephfs idmapped mounts is already used/tested with LXD/LXC workloads
> > and it works perfectly well. And as far as I know, LXD/LXC were the
> > first idmapped mount adopters. IMHO, it's better to
> > start from this approach and if someone will want to extend this
> > functionality for network filesystems and want to map fs{g,u}id which
> > are sent over the
> > wire we will take a look at that. Because anyway, integration with
> > Cephfs is important for the LXD project and we are looking closely at
> > this.
> >
> > Kind regards,
> > Alex
> >
> >> Thanks
> >>
> >> - Xiubo
> >>
> >>>        head->ino = cpu_to_le64(req->r_deleg_ino);
> >>>        head->args = req->r_args;
> >>>
>

Xiubo Li June 2, 2023, 12:41 a.m. UTC | #5

On 6/2/23 02:29, Aleksandr Mikhalitsyn wrote:
> On Thu, Jun 1, 2023 at 4:29 AM Xiubo Li <[e-mail]> wrote:
>>
>> On 6/1/23 00:32, Aleksandr Mikhalitsyn wrote:
>>> On Mon, May 29, 2023 at 5:52 AM Xiubo Li <[e-mail]> wrote:
>>>> On 5/24/23 23:33, Alexander Mikhalitsyn 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.
>>>>>
>>>>>    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]>
>>>>> Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
>>>>> --- a/fs/ceph/mds_client.c
>>>>> +++ b/fs/ceph/mds_client.c
>>>>> @@ -2583,6 +2583,8 @@ 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,
>>>>> @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
>>>>> +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
>>>>> +      */
>>>>> +     caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
>>>>> +                                     VFSUIDT_INIT(req->r_cred->fsuid));
>>>>> +     caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
>>>>> +                                     VFSGIDT_INIT(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));
>>>> Hi Alexander,
>>> Dear Xiubo,
>>>
>>> Thanks for paying attention to this series!
>>>
>>>> You didn't answer Jeff and Greg's concerns in the first version
>>>> https://www.spinics.net/lists/ceph-devel/msg53356.html.
>>> I've tried to respin discussion in the -v1 thread:
>>> https://lore.kernel.org/all/[e-mail]/
>>>
>>> No one replied, so I decided to send rebased and slightly changed -v2,
>>> where I've fixed this:
>>> https://lore.kernel.org/all/[e-mail]/
>>>
>>>> I am also confused as Greg mentioned. If we just map the ids as 1000:0
>>>> and created a file and then map the ids 1000:10, then the file couldn't
>>>> be accessible, right ? Is this normal and as expected ?
>>> This can be a problem only if filtering based on the UID is turned on
>>> on the server side (which is a relatively rare case).
>>>
>>> idmapped mounts are not about mapping a caller UID/GID, idmapped
>>> mounts are about mapping inode owner's UID/GID.
>>> So, for example if you have UID 1000 (on disk) and have an idmapping
>>> 1000:0 then it will be shown as owned by 0.
>> My understanding was that on the disk the files' owner UID should be
>> 1000 always, while in the client side it will show file's owner as the
>> mapped UID 0 with an idmapping 1000:0.
> Hi, Xiubo!
>
>> This should be the same as what you mentioned above, right ?
> Right.
>
> Let me show a real output from a real command line experiment :-)
>
> 1. Mount cephfs
>
> mount.ceph [e-mail]=/ /mnt/ceph -o
> mon_addr=127.0.0.1:6789,secret=very_secret_key
>
> 2. Make 1000:1000 a root dentry owner (it will be convenient because
> we want to use mapping 1000:0:1 for simplicity)
>
> chown 1000:1000 /mnt/ceph
>
> 3. create an idmapped mount based on a regular /mnt/ceph mount using a
> mount-idmapped tool that was written by Christian.
> [ taken from https://raw.githubusercontent.com/brauner/mount-idmapped/master/mount-idmapped.c
> ]
>
> ./mount-idmapped --map-mount b:1000:0:1 /mnt/ceph /mnt/ceph_idmapped
>
> "b" stands for "both", so we are creating a mapping of length 1 for
> both UID and GID.
> 1000 is a UID/GID "on-disk", 0 is a mapped UID/GID.
>
> 4. Just to be precise, let's look at which UID/GID we have now.
>
> root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
> total 4
> drwxrwxrwx 2 1000 1000    0 Jun  1 17:51 .
> drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..
>
> root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
> total 4
> drwxrwxrwx 2 0 0    0 Jun  1 17:51 .
> drwxr-xr-x 4 0 0 4096 Jun  1 16:55 ..
>
> 5. Now let's create a bunch of files with different owners and through
> different mounts (idmapped/non-idmapped).
>
> 5.1. Create a file from 0:0 through the idmapped mount (it should
> appear as 1000:1000 on disk)
> root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
> /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid0
>
> 5.2. Create a file from 1000:1000 through the idmapped mount (should
> fail because 1000:1000 is not a valid UID/GID as it can't be mapped
> back to the "on-disk" UID/GID set).
> root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
> /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000
> touch: cannot touch
> '/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000': Value
> too large for defined data type
>
> ... and we've got EOVERFLOW. That's correct!
>
> 5.3. Create a file from 0:0 but through the regular mount. (it should
> appear as overflowuid(=65534) in idmapped mount, because 0:0 on-disk
> is not mapped to the UID/GID set).
>
> root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
> /mnt/ceph/created_directly_with_uid0
>
> 5.4. Create a file from 1000:1000 but through the regular mount. (it
> should appear as 0:0 in idmapped mount, because 1000 (on-disk) mapped
> to 0).
>
> root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
> /mnt/ceph/created_directly_with_uid1000
>
> 6. Now let's look on the result:
>
> root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
> total 4
> drwxrwxrwx 2 1000 1000    3 Jun  1 17:54 .
> drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..
> -rw-r--r-- 1    0    0    0 Jun  1 17:54 created_directly_with_uid0
> -rw-rw-r-- 1 1000 1000    0 Jun  1 17:54 created_directly_with_uid1000
> -rw-r--r-- 1 1000 1000    0 Jun  1 17:53 created_through_idmapped_mnt_with_uid0
>
> root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
> total 4
> drwxrwxrwx 2     0     0    3 Jun  1 17:54 .
> drwxr-xr-x 4     0     0 4096 Jun  1 16:55 ..
> -rw-r--r-- 1 65534 65534    0 Jun  1 17:54 created_directly_with_uid0
> -rw-rw-r-- 1     0     0    0 Jun  1 17:54 created_directly_with_uid1000
> -rw-r--r-- 1     0     0    0 Jun  1 17:53
> created_through_idmapped_mnt_with_uid0
>
>>> If you create a file from a user with UID 0 then you will get UID 1000
>>> on disk. To achieve that, we map a current user fs{g,u}id
>>> when sending a creation request according to the idmapping mount to
>>> make things consistent.
>> As you know the cephfs MDSs will use the creation requests' caller UID
>> as the owner's UID when creating new inodes.
> Yes, that's why we have to map a caller UID to end up with the correct
> value of a file owner.
>
Hmm, I think my understanding was incorrect. This patch here is trying 
to get the correct value of UID 1000 from a mapped mount, which the UID 0.


>> Which means that if the creation requests switches to use the mapped UID
>> 0 as the caller UID then the file's owner will be UID 0 instead of UID
>> 1000 in cephfs MDSs. Does this what this patch want to do ?
> In my example we have a caller with UID equal 0, then the mapped UID
> will be 1000. So, the file will be created with UID = 1000.

Okay, thanks for your above example it helped me to understand the idmap 
logic. Before I tried to read the xfstests test cases and VFS code about 
the idmap but didn't totally done yet.

I will test and review the patches again today or next week.

Thanks

- Xiubo

>>
>>>    But when a user opens a file,
>>> we are sending UID/GID as they are without applying an idmapping.
>> If my understanding is correct above, then when opening the file with
>> non-mapped UID 1000 it may fail because the files' owner is UID 0.
>>
>> Correct me if my understanding is wrong.
>>
>>>    Of
>>> course, generic_permission() kernel helper is aware of
>>> mount idmapping
>> Yeah, this was also what I thought it should be.
>>
>> There is another client auth feature [1] for cephfs. The MDS will allow
>> us to set a path restriction for specify UID, more detail please see [2]:
>>
>>    allow rw path=/dir1 uid=1000 gids=1000
>>
>> This may cause the creation requests to fail if you set the caller UID
>> to the mapped UID.
> Yes, that can be a problem of course. But it will only affect users
> who want to use this feature and it doesn't open any security holes.
> It's just a limitation of this approach. Unfortunately it's barely
> fixable without massive VFS changes and until we have no real use
> cases
> for this combination of idmapped mounts + MDS UID/GID-based path
> restriction we are not sure that it makes sense to implement this
> right now.
>
>>
>> [1] https://docs.ceph.com/en/latest/cephfs/client-auth/
>> [2] https://tracker.ceph.com/issues/59388
> Thanks, I'll take a look closer at that!
>
> Thanks for closely looking into this patchset, Xiubo!
>
> Kind regards,
> Alex
>
>>
>> Thanks
>>
>> - Xiubo
>>
>>> and before open request will go to the server we will
>>> check that current user is allowed to open this file (and during
>>> this check UID/GID of a current user and UID/GID of the file owner
>>> will be properly compared). I.e. this issue is only relevant for the
>>> case
>>> when we have additional permission checks on the network file system
>>> server side.
>>>
>>>> IMO the idmapping should be client-side feature and we should make it
>>>> consistent by using the unmapped fs{g,u}id always here.
>>> To make the current user fs{g,u}id always idmapped we need to make
>>> really big changes in the VFS layer. And it's not obvious
>>> that it justifies the cost. Because this particular feature with
>>> Cephfs idmapped mounts is already used/tested with LXD/LXC workloads
>>> and it works perfectly well. And as far as I know, LXD/LXC were the
>>> first idmapped mount adopters. IMHO, it's better to
>>> start from this approach and if someone will want to extend this
>>> functionality for network filesystems and want to map fs{g,u}id which
>>> are sent over the
>>> wire we will take a look at that. Because anyway, integration with
>>> Cephfs is important for the LXD project and we are looking closely at
>>> this.
>>>
>>> Kind regards,
>>> Alex
>>>
>>>> Thanks
>>>>
>>>> - Xiubo
>>>>
>>>>>         head->ino = cpu_to_le64(req->r_deleg_ino);
>>>>>         head->args = req->r_args;
>>>>>

Aleksandr Mikhalitsyn June 2, 2023, 10:01 a.m. UTC | #6

On Fri, Jun 2, 2023 at 2:41 AM Xiubo Li <[e-mail]> wrote:
>
>
> On 6/2/23 02:29, Aleksandr Mikhalitsyn wrote:
> > On Thu, Jun 1, 2023 at 4:29 AM Xiubo Li <[e-mail]> wrote:
> >>
> >> On 6/1/23 00:32, Aleksandr Mikhalitsyn wrote:
> >>> On Mon, May 29, 2023 at 5:52 AM Xiubo Li <[e-mail]> wrote:
> >>>> On 5/24/23 23:33, Alexander Mikhalitsyn 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.
> >>>>>
> >>>>>    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]>
> >>>>> Signed-off-by: Alexander Mikhalitsyn <[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 810c3db2e369..e4265843b838 100644
> >>>>> --- a/fs/ceph/mds_client.c
> >>>>> +++ b/fs/ceph/mds_client.c
> >>>>> @@ -2583,6 +2583,8 @@ 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,
> >>>>> @@ -2651,10 +2653,22 @@ 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->r_mnt_idmap to the
> >>>>> +      * initial idmapping meaning the unmapped fs{g,u}id is sent.
> >>>>> +      */
> >>>>> +     caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
> >>>>> +                                     VFSUIDT_INIT(req->r_cred->fsuid));
> >>>>> +     caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
> >>>>> +                                     VFSGIDT_INIT(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));
> >>>> Hi Alexander,
> >>> Dear Xiubo,
> >>>
> >>> Thanks for paying attention to this series!
> >>>
> >>>> You didn't answer Jeff and Greg's concerns in the first version
> >>>> https://www.spinics.net/lists/ceph-devel/msg53356.html.
> >>> I've tried to respin discussion in the -v1 thread:
> >>> https://lore.kernel.org/all/[e-mail]/
> >>>
> >>> No one replied, so I decided to send rebased and slightly changed -v2,
> >>> where I've fixed this:
> >>> https://lore.kernel.org/all/[e-mail]/
> >>>
> >>>> I am also confused as Greg mentioned. If we just map the ids as 1000:0
> >>>> and created a file and then map the ids 1000:10, then the file couldn't
> >>>> be accessible, right ? Is this normal and as expected ?
> >>> This can be a problem only if filtering based on the UID is turned on
> >>> on the server side (which is a relatively rare case).
> >>>
> >>> idmapped mounts are not about mapping a caller UID/GID, idmapped
> >>> mounts are about mapping inode owner's UID/GID.
> >>> So, for example if you have UID 1000 (on disk) and have an idmapping
> >>> 1000:0 then it will be shown as owned by 0.
> >> My understanding was that on the disk the files' owner UID should be
> >> 1000 always, while in the client side it will show file's owner as the
> >> mapped UID 0 with an idmapping 1000:0.
> > Hi, Xiubo!
> >
> >> This should be the same as what you mentioned above, right ?
> > Right.
> >
> > Let me show a real output from a real command line experiment :-)
> >
> > 1. Mount cephfs
> >
> > mount.ceph [e-mail]=/ /mnt/ceph -o
> > mon_addr=127.0.0.1:6789,secret=very_secret_key
> >
> > 2. Make 1000:1000 a root dentry owner (it will be convenient because
> > we want to use mapping 1000:0:1 for simplicity)
> >
> > chown 1000:1000 /mnt/ceph
> >
> > 3. create an idmapped mount based on a regular /mnt/ceph mount using a
> > mount-idmapped tool that was written by Christian.
> > [ taken from https://raw.githubusercontent.com/brauner/mount-idmapped/master/mount-idmapped.c
> > ]
> >
> > ./mount-idmapped --map-mount b:1000:0:1 /mnt/ceph /mnt/ceph_idmapped
> >
> > "b" stands for "both", so we are creating a mapping of length 1 for
> > both UID and GID.
> > 1000 is a UID/GID "on-disk", 0 is a mapped UID/GID.
> >
> > 4. Just to be precise, let's look at which UID/GID we have now.
> >
> > root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
> > total 4
> > drwxrwxrwx 2 1000 1000    0 Jun  1 17:51 .
> > drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..
> >
> > root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
> > total 4
> > drwxrwxrwx 2 0 0    0 Jun  1 17:51 .
> > drwxr-xr-x 4 0 0 4096 Jun  1 16:55 ..
> >
> > 5. Now let's create a bunch of files with different owners and through
> > different mounts (idmapped/non-idmapped).
> >
> > 5.1. Create a file from 0:0 through the idmapped mount (it should
> > appear as 1000:1000 on disk)
> > root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
> > /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid0
> >
> > 5.2. Create a file from 1000:1000 through the idmapped mount (should
> > fail because 1000:1000 is not a valid UID/GID as it can't be mapped
> > back to the "on-disk" UID/GID set).
> > root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
> > /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000
> > touch: cannot touch
> > '/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000': Value
> > too large for defined data type
> >
> > ... and we've got EOVERFLOW. That's correct!
> >
> > 5.3. Create a file from 0:0 but through the regular mount. (it should
> > appear as overflowuid(=65534) in idmapped mount, because 0:0 on-disk
> > is not mapped to the UID/GID set).
> >
> > root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch
> > /mnt/ceph/created_directly_with_uid0
> >
> > 5.4. Create a file from 1000:1000 but through the regular mount. (it
> > should appear as 0:0 in idmapped mount, because 1000 (on-disk) mapped
> > to 0).
> >
> > root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch
> > /mnt/ceph/created_directly_with_uid1000
> >
> > 6. Now let's look on the result:
> >
> > root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph
> > total 4
> > drwxrwxrwx 2 1000 1000    3 Jun  1 17:54 .
> > drwxr-xr-x 4    0    0 4096 Jun  1 16:55 ..
> > -rw-r--r-- 1    0    0    0 Jun  1 17:54 created_directly_with_uid0
> > -rw-rw-r-- 1 1000 1000    0 Jun  1 17:54 created_directly_with_uid1000
> > -rw-r--r-- 1 1000 1000    0 Jun  1 17:53 created_through_idmapped_mnt_with_uid0
> >
> > root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped
> > total 4
> > drwxrwxrwx 2     0     0    3 Jun  1 17:54 .
> > drwxr-xr-x 4     0     0 4096 Jun  1 16:55 ..
> > -rw-r--r-- 1 65534 65534    0 Jun  1 17:54 created_directly_with_uid0
> > -rw-rw-r-- 1     0     0    0 Jun  1 17:54 created_directly_with_uid1000
> > -rw-r--r-- 1     0     0    0 Jun  1 17:53
> > created_through_idmapped_mnt_with_uid0
> >
> >>> If you create a file from a user with UID 0 then you will get UID 1000
> >>> on disk. To achieve that, we map a current user fs{g,u}id
> >>> when sending a creation request according to the idmapping mount to
> >>> make things consistent.
> >> As you know the cephfs MDSs will use the creation requests' caller UID
> >> as the owner's UID when creating new inodes.
> > Yes, that's why we have to map a caller UID to end up with the correct
> > value of a file owner.
> >
> Hmm, I think my understanding was incorrect. This patch here is trying
> to get the correct value of UID 1000 from a mapped mount, which the UID 0.
>
>
> >> Which means that if the creation requests switches to use the mapped UID
> >> 0 as the caller UID then the file's owner will be UID 0 instead of UID
> >> 1000 in cephfs MDSs. Does this what this patch want to do ?
> > In my example we have a caller with UID equal 0, then the mapped UID
> > will be 1000. So, the file will be created with UID = 1000.
>
> Okay, thanks for your above example it helped me to understand the idmap
> logic. Before I tried to read the xfstests test cases and VFS code about
> the idmap but didn't totally done yet.

Yeah, it's not trivial (especially when it's combined with a network
filesystem ;-) ).
But Christian had written a good documentation about mount idmappings:
https://github.com/torvalds/linux/blob/master/Documentation/filesystems/idmappings.rst

>
> I will test and review the patches again today or next week.

Huge thanks, Xiubo!

>
> Thanks
>
> - Xiubo
>
> >>
> >>>    But when a user opens a file,
> >>> we are sending UID/GID as they are without applying an idmapping.
> >> If my understanding is correct above, then when opening the file with
> >> non-mapped UID 1000 it may fail because the files' owner is UID 0.
> >>
> >> Correct me if my understanding is wrong.
> >>
> >>>    Of
> >>> course, generic_permission() kernel helper is aware of
> >>> mount idmapping
> >> Yeah, this was also what I thought it should be.
> >>
> >> There is another client auth feature [1] for cephfs. The MDS will allow
> >> us to set a path restriction for specify UID, more detail please see [2]:
> >>
> >>    allow rw path=/dir1 uid=1000 gids=1000
> >>
> >> This may cause the creation requests to fail if you set the caller UID
> >> to the mapped UID.
> > Yes, that can be a problem of course. But it will only affect users
> > who want to use this feature and it doesn't open any security holes.
> > It's just a limitation of this approach. Unfortunately it's barely
> > fixable without massive VFS changes and until we have no real use
> > cases
> > for this combination of idmapped mounts + MDS UID/GID-based path
> > restriction we are not sure that it makes sense to implement this
> > right now.
> >
> >>
> >> [1] https://docs.ceph.com/en/latest/cephfs/client-auth/
> >> [2] https://tracker.ceph.com/issues/59388
> > Thanks, I'll take a look closer at that!
> >
> > Thanks for closely looking into this patchset, Xiubo!
> >
> > Kind regards,
> > Alex
> >
> >>
> >> Thanks
> >>
> >> - Xiubo
> >>
> >>> and before open request will go to the server we will
> >>> check that current user is allowed to open this file (and during
> >>> this check UID/GID of a current user and UID/GID of the file owner
> >>> will be properly compared). I.e. this issue is only relevant for the
> >>> case
> >>> when we have additional permission checks on the network file system
> >>> server side.
> >>>
> >>>> IMO the idmapping should be client-side feature and we should make it
> >>>> consistent by using the unmapped fs{g,u}id always here.
> >>> To make the current user fs{g,u}id always idmapped we need to make
> >>> really big changes in the VFS layer. And it's not obvious
> >>> that it justifies the cost. Because this particular feature with
> >>> Cephfs idmapped mounts is already used/tested with LXD/LXC workloads
> >>> and it works perfectly well. And as far as I know, LXD/LXC were the
> >>> first idmapped mount adopters. IMHO, it's better to
> >>> start from this approach and if someone will want to extend this
> >>> functionality for network filesystems and want to map fs{g,u}id which
> >>> are sent over the
> >>> wire we will take a look at that. Because anyway, integration with
> >>> Cephfs is important for the LXD project and we are looking closely at
> >>> this.
> >>>
> >>> Kind regards,
> >>> Alex
> >>>
> >>>> Thanks
> >>>>
> >>>> - Xiubo
> >>>>
> >>>>>         head->ino = cpu_to_le64(req->r_deleg_ino);
> >>>>>         head->args = req->r_args;
> >>>>>
>

Patch

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 810c3db2e369..e4265843b838 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -2583,6 +2583,8 @@  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,
@@ -2651,10 +2653,22 @@  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->r_mnt_idmap to the
+	 * initial idmapping meaning the unmapped fs{g,u}id is sent.
+	 */
+	caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns,
+					VFSUIDT_INIT(req->r_cred->fsuid));
+	caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns,
+					VFSGIDT_INIT(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;