vhost-scsi: unbreak any layout for response
authorJason Wang <jasowang@redhat.com>
Thu, 19 Jan 2023 07:36:47 +0000 (15:36 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 9 Feb 2023 10:28:04 +0000 (11:28 +0100)
[ Upstream commit 6dd88fd59da84631b5fe5c8176931c38cfa3b265 ]

Al Viro said:

"""
Since "vhost/scsi: fix reuse of &vq->iov[out] in response"
we have this:
                cmd->tvc_resp_iov = vq->iov[vc.out];
                cmd->tvc_in_iovs = vc.in;
combined with
                iov_iter_init(&iov_iter, ITER_DEST, &cmd->tvc_resp_iov,
                              cmd->tvc_in_iovs, sizeof(v_rsp));
in vhost_scsi_complete_cmd_work().  We used to have ->tvc_resp_iov
_pointing_ to vq->iov[vc.out]; back then iov_iter_init() asked to
set an iovec-backed iov_iter over the tail of vq->iov[], with
length being the amount of iovecs in the tail.

Now we have a copy of one element of that array.  Fortunately, the members
following it in the containing structure are two non-NULL kernel pointers,
so copy_to_iter() will not copy anything beyond the first iovec - kernel
pointer is not (on the majority of architectures) going to be accepted by
access_ok() in copyout() and it won't be skipped since the "length" (in
reality - another non-NULL kernel pointer) won't be zero.

So it's not going to give a guest-to-qemu escalation, but it's definitely
a bug.  Frankly, my preference would be to verify that the very first iovec
is long enough to hold rsp_size.  Due to the above, any users that try to
give us vq->iov[vc.out].iov_len < sizeof(struct virtio_scsi_cmd_resp)
would currently get a failure in vhost_scsi_complete_cmd_work()
anyway.
"""

However, the spec doesn't say anything about the legacy descriptor
layout for the respone. So this patch tries to not assume the response
to reside in a single separate descriptor which is what commit
79c14141a487 ("vhost/scsi: Convert completion path to use") tries to
achieve towards to ANY_LAYOUT.

This is done by allocating and using dedicate resp iov in the
command. To be safety, start with UIO_MAXIOV to be consistent with the
limitation that we advertise to the vhost_get_vq_desc().

Testing with the hacked virtio-scsi driver that use 1 descriptor for 1
byte in the response.

Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Benjamin Coddington <bcodding@redhat.com>
Cc: Nicholas Bellinger <nab@linux-iscsi.org>
Fixes: a77ec83a5789 ("vhost/scsi: fix reuse of &vq->iov[out] in response")
Signed-off-by: Jason Wang <jasowang@redhat.com>
Message-Id: <20230119073647.76467-1-jasowang@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/vhost/scsi.c

index dca6346..d5ecb88 100644 (file)
@@ -80,7 +80,7 @@ struct vhost_scsi_cmd {
        struct scatterlist *tvc_prot_sgl;
        struct page **tvc_upages;
        /* Pointer to response header iovec */
-       struct iovec tvc_resp_iov;
+       struct iovec *tvc_resp_iov;
        /* Pointer to vhost_scsi for our device */
        struct vhost_scsi *tvc_vhost;
        /* Pointer to vhost_virtqueue for the cmd */
@@ -563,7 +563,7 @@ static void vhost_scsi_complete_cmd_work(struct vhost_work *work)
                memcpy(v_rsp.sense, cmd->tvc_sense_buf,
                       se_cmd->scsi_sense_length);
 
-               iov_iter_init(&iov_iter, ITER_DEST, &cmd->tvc_resp_iov,
+               iov_iter_init(&iov_iter, ITER_DEST, cmd->tvc_resp_iov,
                              cmd->tvc_in_iovs, sizeof(v_rsp));
                ret = copy_to_iter(&v_rsp, sizeof(v_rsp), &iov_iter);
                if (likely(ret == sizeof(v_rsp))) {
@@ -594,6 +594,7 @@ vhost_scsi_get_cmd(struct vhost_virtqueue *vq, struct vhost_scsi_tpg *tpg,
        struct vhost_scsi_cmd *cmd;
        struct vhost_scsi_nexus *tv_nexus;
        struct scatterlist *sg, *prot_sg;
+       struct iovec *tvc_resp_iov;
        struct page **pages;
        int tag;
 
@@ -613,6 +614,7 @@ vhost_scsi_get_cmd(struct vhost_virtqueue *vq, struct vhost_scsi_tpg *tpg,
        sg = cmd->tvc_sgl;
        prot_sg = cmd->tvc_prot_sgl;
        pages = cmd->tvc_upages;
+       tvc_resp_iov = cmd->tvc_resp_iov;
        memset(cmd, 0, sizeof(*cmd));
        cmd->tvc_sgl = sg;
        cmd->tvc_prot_sgl = prot_sg;
@@ -625,6 +627,7 @@ vhost_scsi_get_cmd(struct vhost_virtqueue *vq, struct vhost_scsi_tpg *tpg,
        cmd->tvc_data_direction = data_direction;
        cmd->tvc_nexus = tv_nexus;
        cmd->inflight = vhost_scsi_get_inflight(vq);
+       cmd->tvc_resp_iov = tvc_resp_iov;
 
        memcpy(cmd->tvc_cdb, cdb, VHOST_SCSI_MAX_CDB_SIZE);
 
@@ -935,7 +938,7 @@ vhost_scsi_handle_vq(struct vhost_scsi *vs, struct vhost_virtqueue *vq)
        struct iov_iter in_iter, prot_iter, data_iter;
        u64 tag;
        u32 exp_data_len, data_direction;
-       int ret, prot_bytes, c = 0;
+       int ret, prot_bytes, i, c = 0;
        u16 lun;
        u8 task_attr;
        bool t10_pi = vhost_has_feature(vq, VIRTIO_SCSI_F_T10_PI);
@@ -1092,7 +1095,8 @@ vhost_scsi_handle_vq(struct vhost_scsi *vs, struct vhost_virtqueue *vq)
                }
                cmd->tvc_vhost = vs;
                cmd->tvc_vq = vq;
-               cmd->tvc_resp_iov = vq->iov[vc.out];
+               for (i = 0; i < vc.in ; i++)
+                       cmd->tvc_resp_iov[i] = vq->iov[vc.out + i];
                cmd->tvc_in_iovs = vc.in;
 
                pr_debug("vhost_scsi got command opcode: %#02x, lun: %d\n",
@@ -1461,6 +1465,7 @@ static void vhost_scsi_destroy_vq_cmds(struct vhost_virtqueue *vq)
                kfree(tv_cmd->tvc_sgl);
                kfree(tv_cmd->tvc_prot_sgl);
                kfree(tv_cmd->tvc_upages);
+               kfree(tv_cmd->tvc_resp_iov);
        }
 
        sbitmap_free(&svq->scsi_tags);
@@ -1508,6 +1513,14 @@ static int vhost_scsi_setup_vq_cmds(struct vhost_virtqueue *vq, int max_cmds)
                        goto out;
                }
 
+               tv_cmd->tvc_resp_iov = kcalloc(UIO_MAXIOV,
+                                              sizeof(struct iovec),
+                                              GFP_KERNEL);
+               if (!tv_cmd->tvc_resp_iov) {
+                       pr_err("Unable to allocate tv_cmd->tvc_resp_iov\n");
+                       goto out;
+               }
+
                tv_cmd->tvc_prot_sgl = kcalloc(VHOST_SCSI_PREALLOC_PROT_SGLS,
                                               sizeof(struct scatterlist),
                                               GFP_KERNEL);