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
vhost core.
Testing with the hacked virtio-scsi driver that use 1 descriptor for 1
byte in the response.
Reported-by: Al Viro <[email protected]>
Cc: Benjamin Coddington <[email protected]>
Cc: Nicholas Bellinger <[email protected]>
Fixes: a77ec83a5789 ("vhost/scsi: fix reuse of &vq->iov[out] in response")
Signed-off-by: Jason Wang <[email protected]>
---
drivers/vhost/scsi.c | 21 +++++++++++++++++----
1 file changed, 17 insertions(+), 4 deletions(-)
diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
index dca6346d75b3..7d6d70072603 100644
--- a/drivers/vhost/scsi.c
+++ b/drivers/vhost/scsi.c
@@ -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 page *),
+ 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);
--
2.25.1
On Wed, Jan 11, 2023 at 02:07:30PM +0800, Jason Wang wrote:
> 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
> vhost core.
>
> Testing with the hacked virtio-scsi driver that use 1 descriptor for 1
> byte in the response.
>
> Reported-by: Al Viro <[email protected]>
> Cc: Benjamin Coddington <[email protected]>
> Cc: Nicholas Bellinger <[email protected]>
> Fixes: a77ec83a5789 ("vhost/scsi: fix reuse of &vq->iov[out] in response")
> Signed-off-by: Jason Wang <[email protected]>
> ---
> drivers/vhost/scsi.c | 21 +++++++++++++++++----
> 1 file changed, 17 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> index dca6346d75b3..7d6d70072603 100644
> --- a/drivers/vhost/scsi.c
> +++ b/drivers/vhost/scsi.c
> @@ -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];
Where is the guarantee that vc.in < UIO_MAXIOV?
> 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 page *),
> + GFP_KERNEL);
Should sizeof(struct page *) be sizeof(struct iovec)?
> + 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);
> --
> 2.25.1
>
On Thu, Jan 12, 2023 at 3:18 AM Stefan Hajnoczi <[email protected]> wrote:
>
> On Wed, Jan 11, 2023 at 02:07:30PM +0800, Jason Wang wrote:
> > 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
> > vhost core.
> >
> > Testing with the hacked virtio-scsi driver that use 1 descriptor for 1
> > byte in the response.
> >
> > Reported-by: Al Viro <[email protected]>
> > Cc: Benjamin Coddington <[email protected]>
> > Cc: Nicholas Bellinger <[email protected]>
> > Fixes: a77ec83a5789 ("vhost/scsi: fix reuse of &vq->iov[out] in response")
> > Signed-off-by: Jason Wang <[email protected]>
> > ---
> > drivers/vhost/scsi.c | 21 +++++++++++++++++----
> > 1 file changed, 17 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> > index dca6346d75b3..7d6d70072603 100644
> > --- a/drivers/vhost/scsi.c
> > +++ b/drivers/vhost/scsi.c
> > @@ -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];
>
> Where is the guarantee that vc.in < UIO_MAXIOV?
>
We limit it here in the vhost_virtqueue structure:
struct iovec iov[UIO_MAXIOV];
And we pass sizeof(vq->iov) to vhost_get_vq_desc():
vc->head = vhost_get_vq_desc(vq, vq->iov,
ARRAY_SIZE(vq->iov), &vc->out, &vc->in,
NULL, NULL);
> > 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 page *),
> > + GFP_KERNEL);
>
> Should sizeof(struct page *) be sizeof(struct iovec)?
Yes, I will fix it.
Thanks
>
> > + 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);
> > --
> > 2.25.1
> >