Skip to contentSkip to conversation

xen/scsiback: fix command-tag handling on pre-completion error paths

PATCH Applied Michael Bommarito 4 participants 8 messages Jun 11 – Jun 16

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover xen/scsiback: fix command-tag handling on pre-completion error paths

scsiback_get_pend_req() hands a pvSCSI frontend request a session tag and a zeroed se_cmd. Two error paths that run before the command completes through the target core mishandle that command and leak (or, in one case, underflow) the tag.

Impact: a pvSCSI guest can exhaust a LUN's per-session command tag pool, stopping the LUN, via crafted ring requests; for the first case the refcount underflow also panics the host under panic_on_warn.

Patch 1 fixes scsiback_do_cmd_fn(): on a failed grant map and on an unknown request type the never-initialised command (cmd_kref == 0) is freed with transport_generic_free_cmd(), which underflows the zero refcount and leaks the tag.

Patch 2 fixes scsiback_device_action(): when target_submit_tmr() fails the
err: path frees nothing. transport_generic_free_cmd() cannot be used there either, since the command is initialised by then and se_tmr_req has already been freed on one error sub-path.

Both paths go through one helper that returns just the tag.

Patch 1's underflow was reproduced on a Xen dom0 (guest to host, with a panic_on_warn host panic); with the series applied the same request is handled with no underflow.

Michael Bommarito (2):
xen/scsiback: free unsubmitted command instead of double-putting it xen/scsiback: free the command tag on the TMR submit-failure path

 drivers/xen/xen-scsiback.c | 30 +++++++++++++++++++++++-------
 1 file changed, 23 insertions(+), 7 deletions(-)

base-commit: 5200f5f493f79f14bbdc349e402a40dfb32f23c8

Signature · 1 line
2.53.0
#1 lore raw

PATCH 1/2 xen/scsiback: free unsubmitted command instead of double-putting it

scsiback_get_pend_req() obtains a command tag and returns a vscsibk_pend whose embedded se_cmd has only been memset to 0, so its cmd_kref is 0; the se_cmd is initialised (kref_init() via target_init_cmd()) only later, in scsiback_cmd_exec(), on the successful VSCSIIF_ACT_SCSI_CDB path. The two error paths in scsiback_do_cmd_fn() taken before the command is submitted -- a failed scsiback_gnttab_data_map() and an unknown ring_req.act -- call transport_generic_free_cmd(&pending_req->se_cmd, 0), which kref_put()s a refcount of 0. That underflows it ("refcount_t:
underflow; use-after-free") and, as the release function is not run, leaks the command tag.

Impact: a pvSCSI guest can leak every command tag of a LUN's session, stopping the LUN, by submitting requests with a bad grant reference or an unknown request type; under panic_on_warn the refcount underflow panics the host.

Add a helper that just returns the tag with target_free_tag() and sends the error response. It frees the tag while the v2p reference still pins the session, and snapshots the response fields beforehand because freeing the tag can let another ring reuse the pending_req slot.

Fixes 2dbcdf33dbf6 ("xen-scsiback: Convert to percpu_ida tag allocation")
Assisted-by Claude:claude-opus-4-8

---
Reproduced on a Xen dom0 (Linux 6.1.y) exporting a pvSCSI LUN to a guest. A frontend that sends a single ring request with an unknown action type drives scsiback_do_cmd_fn() into transport_generic_free_cmd() on the never-initialised command and logs

refcount_t: underflow; use-after-free
WARNING: ... refcount_warn_saturate
transport_generic_free_cmd+0x... [target_core_mod] scsiback_do_cmd_fn+0x... [xen_scsiback] scsiback_irq_fn+0x... [xen_scsiback]

from the vscsiif IRQ thread, and panics the dom0 under panic_on_warn. The failed grant-map path reaches the same free. With this patch the same request is answered with DID_ERROR and the tag is returned, with no underflow. These error paths are unchanged since 2dbcdf33dbf6, so mainline is affected identically.

 drivers/xen/xen-scsiback.c | 28 ++++++++++++++++++++++------
 1 file changed, 22 insertions(+), 6 deletions(-)
drivers/xen/xen-scsiback.c +22 -6
index e33f95c91b096..f324732eba7f8 100644
@@ -611,6 +611,25 @@ static void scsiback_disconnect(struct vscsibk_info *info)
 	xenbus_unmap_ring_vfree(info->dev, info->ring.sring);
 }
 
+/*
+ * Send the error response for a request that did not reach the target core
+ * and return its tag.  Free the tag before the response drops the v2p
+ * reference that keeps the session alive, and snapshot what the response
+ * needs since returning the tag can let the slot be reused.
+ */
+static void scsiback_resp_and_free(struct vscsibk_pend *pending_req,
+				   int32_t result)
+{
+	struct vscsibk_info *info = pending_req->info;
+	struct v2p_entry *v2p = pending_req->v2p;
+	struct se_session *se_sess = v2p->tpg->tpg_nexus->tvn_se_sess;
+	u16 rqid = pending_req->rqid;
+
+	target_free_tag(se_sess, &pending_req->se_cmd);
+	scsiback_send_response(info, NULL, result, 0, rqid);
+	kref_put(&v2p->kref, scsiback_free_translation_entry);
+}
+
 static void scsiback_device_action(struct vscsibk_pend *pending_req,
 	enum tcm_tmreq_table act, int tag)
 {
@@ -792,9 +811,8 @@ static int scsiback_do_cmd_fn(struct vscsibk_info *info,
 		case VSCSIIF_ACT_SCSI_CDB:
 			if (scsiback_gnttab_data_map(&ring_req, pending_req)) {
 				scsiback_fast_flush_area(pending_req);
-				scsiback_do_resp_with_sense(NULL,
-						DID_ERROR << 16, 0, pending_req);
-				transport_generic_free_cmd(&pending_req->se_cmd, 0);
+				scsiback_resp_and_free(pending_req,
+						       DID_ERROR << 16);
 			} else {
 				scsiback_cmd_exec(pending_req);
 			}
@@ -808,9 +826,7 @@ static int scsiback_do_cmd_fn(struct vscsibk_info *info,
 			break;
 		default:
 			pr_err_ratelimited("invalid request\n");
-			scsiback_do_resp_with_sense(NULL, DID_ERROR << 16, 0,
-						    pending_req);
-			transport_generic_free_cmd(&pending_req->se_cmd, 0);
+			scsiback_resp_and_free(pending_req, DID_ERROR << 16);
 			break;
 		}
 
Signature · 1 line
2.53.0
maintainer Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] xen/scsiback: free unsubmitted command instead of double-putting it

Quoted 23 lines · Michael Bommarito

scsiback_get_pend_req() obtains a command tag and returns a vscsibk_pend whose embedded se_cmd has only been memset to 0, so its cmd_kref is 0; the se_cmd is initialised (kref_init() via target_init_cmd()) only later, in scsiback_cmd_exec(), on the successful VSCSIIF_ACT_SCSI_CDB path. The two error paths in scsiback_do_cmd_fn() taken before the command is submitted -- a failed scsiback_gnttab_data_map() and an unknown ring_req.act -- call transport_generic_free_cmd(&pending_req->se_cmd, 0), which kref_put()s a refcount of 0. That underflows it ("refcount_t:
underflow; use-after-free") and, as the release function is not run, leaks the command tag.

Impact: a pvSCSI guest can leak every command tag of a LUN's session, stopping the LUN, by submitting requests with a bad grant reference or an unknown request type; under panic_on_warn the refcount underflow panics the host.

Add a helper that just returns the tag with target_free_tag() and sends the error response. It frees the tag while the v2p reference still pins the session, and snapshots the response fields beforehand because freeing the tag can let another ring reuse the pending_req slot.

Quoted 4 lines
Fixes 2dbcdf33dbf6 ("xen-scsiback: Convert to percpu_ida tag allocation")
Assisted-by Claude:claude-opus-4-8

Juergen

#3 lore raw

PATCH 2/2 xen/scsiback: free the command tag on the TMR submit-failure path

In reply to Michael Bommarito

scsiback_device_action() obtains a command tag in scsiback_get_pend_req() and submits a task-management request with target_submit_tmr(). When target_submit_tmr() fails it returns < 0 and scsiback jumps to the err: label, which sends a response but frees nothing, leaking the tag.

Impact: a pvSCSI guest can leak the command tags of a LUN's session, stopping the LUN, by issuing VSCSIIF_ACT_SCSI_ABORT or RESET requests whenever target_submit_tmr() fails.

transport_generic_free_cmd() cannot be used here. By the time target_submit_tmr() returns an error it has already run __target_init_cmd() (so se_cmd->cmd_kref is one, not zero), and on its target_get_sess_cmd() error path it has freed se_cmd->se_tmr_req via core_tmr_release_req() while leaving SCF_SCSI_TMR_CDB set and the pointer dangling. Letting the command release run target_free_cmd_mem() would then double-free se_tmr_req.

Use the same helper, which returns just the tag, on this path too.

Fixes 2dbcdf33dbf6 ("xen-scsiback: Convert to percpu_ida tag allocation")
Assisted-by Claude:claude-opus-4-8
---
 drivers/xen/xen-scsiback.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/xen/xen-scsiback.c +1 -1
index f324732eba7f8..c7036e0e41bda 100644
@@ -658,7 +658,7 @@ static void scsiback_device_action(struct vscsibk_pend *pending_req,
 	return;
 
 err:
-	scsiback_do_resp_with_sense(NULL, err, 0, pending_req);
+	scsiback_resp_and_free(pending_req, err);
 }
 
 /*
Signature · 1 line
2.53.0
#4 lore raw

Re: [PATCH 2/2] xen/scsiback: free the command tag on the TMR submit-failure path

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [Critical] The return value of target_init_cmd() is ignored in scsiback_cmd_exec(), leading to a percpu_ref underflow, kref underflow, and use-after-free if the session is shutting down.
- [High] Callers of target_submit_tmr() in other fabric drivers unconditionally dereference the se_tmr_req pointer on the error path, leading to a UAF or NULL pointer dereference. --

commit 0671b280117e9d31e5e4fbe8dc07f0a60a155bfb
Author: Michael Bommarito <[email protected]>

xen/scsiback: free the command tag on the TMR submit-failure path

This commit addresses a tag leak on the task-management request submit failure path in scsiback_device_action(). When target_submit_tmr() returns an error, the code now properly frees the tag by calling scsiback_resp_and_free().

drivers/xen/xen-scsiback.c static void scsiback_device_action(struct vscsibk_pend *pending_req, · line 661

Quoted 10 lines
drivers/xen/xen-scsiback.c +1 -1
index f324732eba7f8..c7036e0e41bda 100644
@@ -658,7 +658,7 @@ static void scsiback_device_action(struct vscsibk_pend *pending_req,
 	return;
 
 err:
-	scsiback_do_resp_with_sense(NULL, err, 0, pending_req);
+	scsiback_resp_and_free(pending_req, err);

[Severity: High]
This isn't a bug introduced by this patch, but since we are fixing error paths for target_submit_tmr() failures here, does this same issue affect other fabric drivers?

When target_submit_tmr() returns an error, the se_tmr_req pointer is either left dangling or remains unallocated. However, looking at srpt_handle_tsk_mgmt() in drivers/infiniband/ulp/srpt/ib_srpt.c, the error path unconditionally writes to it:

    rc = target_submit_tmr(&send_ioctx->cmd, ...);
    if (rc != 0) {
        send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
        cmd->se_tfo->queue_tm_rsp(cmd);
    }

The same pattern exists in drivers/scsi/ibmvscsi_tgt/ibmvscsi_tgt.c in ibmvscsis_parse_task():

    rc = target_submit_tmr(&cmd->se_cmd, ...);
    if (rc) {
        cmd->se_cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
        ...
    }

Could this lead to a use-after-free or null pointer dereference in those drivers under similar error conditions?

[Severity: Critical]
This is a pre-existing issue, but while looking at command handling in this file, I noticed a potential problem in scsiback_cmd_exec().

The code calls target_init_cmd() but ignores its return value:

    target_init_cmd(se_cmd, sess, pending_req->sense_buffer,
                    pending_req->v2p->lun, pending_req->data_len, 0,
                    pending_req->sc_data_direction, TARGET_SCF_ACK_KREF);

    if (target_submit_prep(se_cmd, pending_req->cmnd, pending_req->sgl, ...))
        return;

    target_submit(se_cmd);

If target_init_cmd() fails, which happens when target_get_sess_cmd() fails during session teardown, the session's cmd_cnt->refcnt is never incremented, and cmd_kref is reverted to 1 instead of 2.

Because scsiback_cmd_exec() blindly proceeds to map scatter-gather lists and submit the command, could this cause a percpu_ref underflow when scsiback_cmd_done() and transport_generic_free_cmd() both call target_put_sess_cmd() upon completion?

This would drop cmd_kref to 0 prematurely, triggering target_release_cmd_kref() and unconditionally invoking percpu_ref_put(), potentially leading to a kref underflow and use-after-free of the session.

Quoted 3 lines
drivers/xen/xen-scsiback.c +0 -0
 }
 
 /*
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
#5 lore raw

Re: [PATCH 2/2] xen/scsiback: free the command tag on the TMR submit-failure path

Quoted 1 line · On Thu, Jun 11, 2026 at 8:47 AM <[email protected]> wrote:

- [Critical] The return value of target_init_cmd() is ignored in scsiback_cmd_exec(), leading to a percpu_ref underflow, kref underflow, and use-after-free if the session is shutting down.

My understanding is that scsiback never calls target_stop_session, so per the comment at drivers/target/target_core_transport.c:1712, the return check is not required here.

The second point is real but the only two similarly affected sites I found, srpt and ibmvscsi_tgt, seemed low priority and not worth the churn. Happy to send in if someone wants.

Thanks,
Mike

maintainer Reviewed-by on v1 2/2
#6 lore raw

Re: [PATCH 2/2] xen/scsiback: free the command tag on the TMR submit-failure path

In reply to Michael Bommarito
Quoted 18 lines · Michael Bommarito

scsiback_device_action() obtains a command tag in scsiback_get_pend_req() and submits a task-management request with target_submit_tmr(). When target_submit_tmr() fails it returns < 0 and scsiback jumps to the err: label, which sends a response but frees nothing, leaking the tag.

Impact: a pvSCSI guest can leak the command tags of a LUN's session, stopping the LUN, by issuing VSCSIIF_ACT_SCSI_ABORT or RESET requests whenever target_submit_tmr() fails.

transport_generic_free_cmd() cannot be used here. By the time target_submit_tmr() returns an error it has already run __target_init_cmd() (so se_cmd->cmd_kref is one, not zero), and on its target_get_sess_cmd() error path it has freed se_cmd->se_tmr_req via core_tmr_release_req() while leaving SCF_SCSI_TMR_CDB set and the pointer dangling. Letting the command release run target_free_cmd_mem() would then double-free se_tmr_req.

Quoted 6 lines

Use the same helper, which returns just the tag, on this path too.

Fixes 2dbcdf33dbf6 ("xen-scsiback: Convert to percpu_ida tag allocation")
Assisted-by Claude:claude-opus-4-8

Juergen

#7 lore raw
In reply to Michael Bommarito

Michael,

Quoted 4 lines

scsiback_get_pend_req() hands a pvSCSI frontend request a session tag and a zeroed se_cmd. Two error paths that run before the command completes through the target core mishandle that command and leak (or, in one case, underflow) the tag.

Applied to 7.2/scsi-staging, thanks!

Signature · 1 line
Martin K. Petersen