Re: [PATCH] cifs: Fix the smbd_request and smbd_reponse slabs to allow usercopy

[Stefan Metzmacher <metze@samba.org>]

Am 25.06.25 um 15:37 schrieb David Howells:
> The handling of received data in the smbdirect client code involves using
> copy_to_iter() to copy data from the smbd_reponse struct's packet trailer
> to a folioq buffer provided by netfslib that encapsulates a chunk of
> pagecache.
> 
> If, however, CONFIG_HARDENED_USERCOPY=y, this will result in the checks
> then performed in copy_to_iter() oopsing with something like the following:
> 
>   CIFS: Attempting to mount //172.31.9.1/test
>   CIFS: VFS: RDMA transport established
>   usercopy: Kernel memory exposure attempt detected from SLUB object 'smbd_response_0000000091e24ea1' (offset 81, size 63)!
>   ------------[ cut here ]------------
>   kernel BUG at mm/usercopy.c:102!
>   ...
>   RIP: 0010:usercopy_abort+0x6c/0x80
>   ...
>   Call Trace:
>    <TASK>
>    __check_heap_object+0xe3/0x120
>    __check_object_size+0x4dc/0x6d0
>    smbd_recv+0x77f/0xfe0 [cifs]
>    cifs_readv_from_socket+0x276/0x8f0 [cifs]
>    cifs_read_from_socket+0xcd/0x120 [cifs]
>    cifs_demultiplex_thread+0x7e9/0x2d50 [cifs]
>    kthread+0x396/0x830
>    ret_from_fork+0x2b8/0x3b0
>    ret_from_fork_asm+0x1a/0x30
> 
> The problem is that the smbd_response slab's packet field isn't marked as
> being permitted for usercopy.
> 
> Fix this by passing parameters to kmem_slab_create() to indicate that
> copy_to_iter() is permitted from the packet region of the smbd_response
> slab objects.
> 
> Further, do the same thing for smbd_request slab objects and their packet
> field.
> 
> Fixes: ee4cdf7ba857 ("netfs: Speed up buffered reading")
> Reported-by: Stefan Metzmacher <metze@samba.org>
> Link: https://lore.kernel.org/r/acb7f612-df26-4e2a-a35d-7cd040f513e1@samba.org/
> Signed-off-by: David Howells <dhowells@redhat.com>
> cc: Steve French <stfrench@microsoft.com>
> cc: Paulo Alcantara <pc@manguebit.com>
> cc: linux-cifs@vger.kernel.org
> cc: netfs@lists.linux.dev
> cc: linux-fsdevel@vger.kernel.org
> ---
>   fs/smb/client/smbdirect.c |   21 +++++++++++++++------
>   1 file changed, 15 insertions(+), 6 deletions(-)
> 
> diff --git a/fs/smb/client/smbdirect.c b/fs/smb/client/smbdirect.c
> index ef6bf8d6808d..5915273636ad 100644
> --- a/fs/smb/client/smbdirect.c
> +++ b/fs/smb/client/smbdirect.c
> @@ -1476,12 +1476,17 @@ static int allocate_caches_and_workqueue(struct smbd_connection *info)
>   	int rc;
>   
>   	scnprintf(name, MAX_NAME_LEN, "smbd_request_%p", info);
> +	struct kmem_cache_args request_args = {
> +		.align		= __alignof__(struct smbd_request),
> +		.useroffset	= offsetof(struct smbd_request, packet),
> +		.usersize	= sizeof(struct smbdirect_data_transfer),

This looks wrong, the smbdirect_data_transfer header itself
should be written by userspace.

So I guess we don't need this at all.

> +	};
>   	info->request_cache =
>   		kmem_cache_create(
>   			name,
>   			sizeof(struct smbd_request) +
>   				sizeof(struct smbdirect_data_transfer),
> -			0, SLAB_HWCACHE_ALIGN, NULL);
> +			&request_args, SLAB_HWCACHE_ALIGN);
>   	if (!info->request_cache)
>   		return -ENOMEM;
>   
> @@ -1492,12 +1497,16 @@ static int allocate_caches_and_workqueue(struct smbd_connection *info)
>   		goto out1;
>   
>   	scnprintf(name, MAX_NAME_LEN, "smbd_response_%p", info);
> +
> +	struct kmem_cache_args response_args = {
> +		.align		= __alignof__(struct smbd_response),
> +		.useroffset	= offsetof(struct smbd_response, packet),
> +		.usersize	= sp->max_recv_size,

This should be have + sizeof(struct smbdirect_data_transfer) for useroffset
and - sizeof(struct smbdirect_data_transfer) for usersize

As the smbdirect_data_transfer header should not accessed by userspace.

My attempt looks like this:

-               kmem_cache_create(
+               kmem_cache_create_usercopy(
                         name,
                         sizeof(struct smbd_response) +
                                 sp->max_recv_size,
-                       0, SLAB_HWCACHE_ALIGN, NULL);
+                       __alignof__(struct smbd_response),
+                       SLAB_HWCACHE_ALIGN,
+                       /*
+                        * only the payload should be exposed
+                        */
+                       offsetof(struct smbd_response, packet) +
+                                sizeof(struct smbdirect_data_transfer),
+                       sp->max_recv_size -
+                               sizeof(struct smbdirect_data_transfer),
+                       NULL);

But I noticed that kmem_cache_create_usercopy is a legacy wrapper.