On Wed, Jul 22, 2026 at 2:40 AM Jia Jia <[email protected]> wrote:
>
> vhost_scsi_setup_vq_cmds() runs only from VHOST_SCSI_SET_ENDPOINT and
> allocates each command's protection scatterlist array (prot_sgl) only
> when VIRTIO_SCSI_F_T10_PI is already set at that moment. Command pools
> are not rebuilt later. vhost_scsi_set_features() may still flip that
> bit after the endpoint is live: it updates acked_features and returns
> success without reallocating prot_sgl.
>
> That is a feature-versus-resource lifetime mismatch. The broken order
> is:
>
> VHOST_SET_FEATURES(0)
> VHOST_SCSI_SET_ENDPOINT
> setup_vq_cmds() sees no T10-PI -> prot_sgl stays NULL
> VHOST_SET_FEATURES(VIRTIO_SCSI_F_T10_PI)
> only acked_features changes; command resources stay as above
VIRTIO_SCSI_F_T10_PI is defined in the VIRTIO spec. VIRTIO feature
bits do not change once negotiated. The device must be reset in order
to set the feature bits again.
VHOST_F_LOG_ALL is a vhost-specific feature bit that is not covered by
the VIRTIO spec and it can be changed at runtime. It is an exception.
Please generalize this fix to only allow VHOST_F_LOG_ALL to be changed
at runtime. Rather than protecting VIRTIO_SCSI_F_T10_PI specifically,
attempts to change feature bits other than VHOST_F_LOG_ALL once the
device is operational must be rejected.
Perhaps other vhost device implementations have the same issue?
Thanks,
Stefan
> submit a T10-PI WRITE with a 129-page protection payload
>
> The I/O path then follows the new feature bit while using the old
> command objects. vhost_scsi_mapal() (static, inlined into
> vhost_scsi_handle_vq() here) does:
>
> sg_alloc_table_chained(table, 129, first_chunk=NULL,
> nents_first_chunk=inline_sg_cnt)
>
> With first_chunk NULL, __sg_alloc_table() sets curr_max_ents from
> nents_first_chunk and calls sg_pool_alloc(129). sg_pool_index() then
> hits:
>
> BUG_ON(nents > SG_CHUNK_SIZE); /* 129 > 128 */
>
> The kernel reported the following call trace and register state:
>
> Call Trace:
> <TASK>
> ? __sg_alloc_table+0x1d8/0x250
> ? __pfx_vhost_run_work_list+0x10/0x10 [vhost]
> sg_alloc_table_chained+0x59/0xf0
> ? __pfx_sg_pool_alloc+0x10/0x10
> ? vhost_scsi_calc_sgls.constprop.0+0x43/0x60 [vhost_scsi]
> vhost_scsi_handle_vq+0xf02/0x1700 [vhost_scsi]
> ? __pfx_vhost_scsi_handle_vq+0x10/0x10 [vhost_scsi]
> vhost_scsi_handle_kick+0x37/0x50 [vhost_scsi]
> vhost_run_work_list+0x8e/0xd0 [vhost]
> vhost_task_fn+0xe1/0x210
> ret_from_fork+0x348/0x540
> </TASK>
>
> RIP: 0010:0x4
> CR2: 0000000000000004
> RSP: 0018:ffffc90000dbf940 EFLAGS: 00010202
> RAX: ffffffff82396810 RBX: ffff88811dc28b80 RCX: 0000000000000000
> RDX: 0000000000000000 RSI: 0000000000000820 RDI: 0000000000000081
>
> The host worker dies and the machine stops accepting network service
> until reset. The missing object is prot_sgl after a post-endpoint
> T10-PI enable introduced by commit bf2d650391be ("vhost-scsi: Allocate
> T10 PI structs only when enabled").
>
> I also checked the other feature bits handled around endpoint setup.
> Only T10-PI allocates this endpoint-time per-command resource and does
> not rebuild it on a later SET_FEATURES. LOG_ALL uses lazy tvc_log and
> already tears log storage down when cleared; HOTPLUG does not allocate
> prot_sgl. So the chosen fix is to refuse T10-PI enable/disable while
> an endpoint is active, rather than adding a second rebuild path in
> set_features().
>
> Return -EBUSY if vs->vs_tpg is set and the T10-PI bit would change.
> Userspace must CLEAR_ENDPOINT, SET_FEATURES, then SET_ENDPOINT again so
> setup_vq_cmds() can allocate protection SGLs when needed. That matches
> the existing "build command resources at endpoint" model.
>
> Fixes: bf2d650391be ("vhost-scsi: Allocate T10 PI structs only when enabled")
> Signed-off-by: Jia Jia <[email protected]>
> ---
> drivers/vhost/scsi.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/drivers/vhost/scsi.c b/drivers/vhost/scsi.c
> index 9a1253b9d8c5..000000000000 100644
> --- a/drivers/vhost/scsi.c
> +++ b/drivers/vhost/scsi.c
> @@ -2219,6 +2219,7 @@ static int vhost_scsi_set_features(struct vhost_scsi
> *vs, u64 features)
> {
> struct vhost_virtqueue *vq;
> bool is_log, was_log;
> + bool is_t10_pi, was_t10_pi;
> int i;
>
> if (features & ~VHOST_SCSI_FEATURES)
> @@ -2234,6 +2235,13 @@ static int vhost_scsi_set_features(struct vhost_scsi
> *vs, u64 features)
> if (!vs->dev.nvqs)
> goto out;
>
> + is_t10_pi = features & (1ULL << VIRTIO_SCSI_F_T10_PI);
> + was_t10_pi = vhost_has_feature(&vs->vqs[0].vq, VIRTIO_SCSI_F_T10_PI);
> + if (vs->vs_tpg && is_t10_pi != was_t10_pi) {
> + mutex_unlock(&vs->dev.mutex);
> + return -EBUSY;
> + }
> +
> is_log = features & (1 << VHOST_F_LOG_ALL);
> /*
> * All VQs should have same feature.
> --
> 2.43.0
>
