Skip to contentSkip to conversation

scsi: mpi3mr: fix use-after-free on tgt_dev->starget during target device refresh/update

PATCH Applied Laurence Oberman 5 participants 6 messages Aug 31 – Sep 18

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 scsi: mpi3mr: fix use-after-free on tgt_dev->starget during target device refresh/update

mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh() read tgt_dev->starget and immediately pass it to starget_for_each_device() without holding mrioc->tgtdev_lock. Every writer of this field -- mpi3mr_target_alloc(), mpi3mr_target_destroy(), mpi3mr_slave_destroy() and mpi3mr_sdev_init() -- correctly serializes access under tgtdev_lock, but these two read sites do not, which leaves a check-then-use window against the SCSI core's target teardown path (scsi_remove_target(), invoked e.g. via a concurrent host reset, sysfs "delete", or SCSI EH device offlining running independently of the fwevt workqueue).

Sequence observed on production hardware, triggered on the mpi3mr0_fwevt_wrkr workqueue during a SAS topology change shortly after a controller reset:

  BUG: kernel NULL pointer dereference, address: 0000000000000058
  RIP: scsi_is_host_device+0x7/0x20
  Call Trace:
   starget_for_each_device+0x34/0x100
   mpi3mr_refresh_tgtdevs+0x152/0x1d0 [mpi3mr]
   mpi3mr_fwevt_bh+0x514/0x6c0 [mpi3mr]
   mpi3mr_fwevt_worker+0x1a/0x50 [mpi3mr]
   process_one_work+0x194/0x380
   worker_thread+0x2fe/0x410

mpi3mr_refresh_tgtdevs() reads tgt_dev->starget as non-NULL, but by the time starget_for_each_device() dereferences it, a concurrent mpi3mr_target_destroy() has already cleared tgt_dev->starget under tgtdev_lock and the SCSI/device core has freed the underlying scsi_target (and its embedded struct device). The stale pointer is then walked by dev_to_shost() -> scsi_is_host_device(), producing the NULL/garbage dereference above.

Fix this by taking mrioc->tgtdev_lock around every read of tgt_dev->starget, matching the existing writer-side discipline. Since starget_for_each_device() and mpi3mr_update_sdev() can end up doing non-atomic work (e.g. queue_limits_commit_update()), the lock cannot be held across the whole call, so instead pin the target's device with get_device() while holding the lock, drop the lock, then run starget_for_each_device() against the pinned reference and put_device() afterwards. This closes the TOCTOU window instead of merely narrowing it.

The same unlocked read-and-dereference pattern also exists earlier in mpi3mr_refresh_tgtdevs()'s first removal-scan loop (tgt_dev->starget->hostdata); fix it the same way by holding tgtdev_lock across that check, which is cheap since it only touches plain struct fields.

Assisted-by Claude:Sonnet5 [Claude Code]
---
 drivers/scsi/mpi3mr/mpi3mr_os.c | 45 +++++++++++++++++++++++++--------
 1 file changed, 35 insertions(+), 10 deletions(-)
drivers/scsi/mpi3mr/mpi3mr_os.c +35 -10
index 402d1f35d214..922b28c3fe12 100644
@@ -1094,10 +1094,13 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 {
 	struct mpi3mr_tgt_dev *tgtdev, *tgtdev_next;
 	struct mpi3mr_stgt_priv_data *tgt_priv;
+	struct scsi_target *starget;
+	unsigned long flags;
 
 	dprint_reset(mrioc, "refresh target devices: check for removals\n");
 	list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc->tgtdev_list,
 	    list) {
+		spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
 		if (((tgtdev->dev_handle == MPI3MR_INVALID_DEV_HANDLE) ||
 		     tgtdev->is_hidden) &&
 		     tgtdev->host_exposed && tgtdev->starget &&
@@ -1106,6 +1109,7 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 			tgt_priv->dev_removed = 1;
 			atomic_set(&tgt_priv->block_io, 0);
 		}
+		spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
 	}
 
 	list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc->tgtdev_list,
@@ -1127,15 +1131,25 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 	tgtdev = NULL;
 	list_for_each_entry(tgtdev, &mrioc->tgtdev_list, list) {
 		if ((tgtdev->dev_handle != MPI3MR_INVALID_DEV_HANDLE) &&
-		    !tgtdev->is_hidden) {
-			if (!tgtdev->host_exposed)
+				!tgtdev->is_hidden) {
+			if (!tgtdev->host_exposed) {
 				mpi3mr_report_tgtdev_to_host(mrioc,
-							     tgtdev->perst_id);
-			else if (tgtdev->starget)
-				starget_for_each_device(tgtdev->starget,
-							(void *)tgtdev, mpi3mr_update_sdev);
-	}
+				     tgtdev->perst_id);
+				continue;
+			}
+			spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
+			starget = tgtdev->starget;
+			if (starget)
+				get_device(&starget->dev);
+			spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
+			if (starget) {
+				starget_for_each_device(starget, (void *)tgtdev,
+							mpi3mr_update_sdev);
+				put_device(&starget->dev);
+			}
+		}
 	}
+	dprint_reset(mrioc, "refresh target devices: done\n");
 }
 
 /**
@@ -1515,6 +1529,8 @@ static void mpi3mr_devinfochg_evt_bh(struct mpi3mr_ioc *mrioc,
 	struct mpi3_device_page0 *dev_pg0)
 {
 	struct mpi3mr_tgt_dev *tgtdev = NULL;
+	struct scsi_target *starget;
+	unsigned long flags;
 	u16 dev_handle = 0, perst_id = 0;
 
 	perst_id = le16_to_cpu(dev_pg0->persistent_id);
@@ -1535,9 +1551,18 @@ static void mpi3mr_devinfochg_evt_bh(struct mpi3mr_ioc *mrioc,
 		mpi3mr_report_tgtdev_to_host(mrioc, perst_id);
 	if (tgtdev->is_hidden && tgtdev->host_exposed)
 		mpi3mr_remove_tgtdev_from_host(mrioc, tgtdev);
-	if (!tgtdev->is_hidden && tgtdev->host_exposed && tgtdev->starget)
-		starget_for_each_device(tgtdev->starget, (void *)tgtdev,
-		    mpi3mr_update_sdev);
+	if (!tgtdev->is_hidden && tgtdev->host_exposed) {
+		spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
+		starget = tgtdev->starget;
+		if (starget)
+			get_device(&starget->dev);
+		spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
+		if (starget) {
+			starget_for_each_device(starget, (void *)tgtdev,
+						mpi3mr_update_sdev);
+			put_device(&starget->dev);
+		}
+	}
 out:
 	if (tgtdev)
 		mpi3mr_tgtdev_put(tgtdev);
Signature · 1 line
2.55.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

Pre-existing issues:
- [High] The patch fixes the unprotected check-then-use window for tgtdev->starget in mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh(), but identical vulnerable patterns remain in other functions like mpi3mr_remove_tgtdev_from_host(), mpi3mr_update_tgtdev(), and various event/BSG handlers.
- [High] Unprotected traversal of mrioc->tgtdev_list without holding tgtdev_lock can lead to list corruption and crashes due to concurrent insertions.
- [High] An AB-BA lock inversion exists between mrioc->tgtdev_lock and the SCSI host lock (shost->host_lock).
--

commit 22ac835bdc4964e1608be9181c0e9ec7c639df2b
Author: Laurence Oberman <[email protected]>

scsi: mpi3mr: fix use-after-free on tgt_dev->starget during target device refresh/update

This commit addresses a check-then-use window against the SCSI core's target teardown path in mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh(). It protects reads of tgt_dev->starget by acquiring tgtdev_lock and pinning the target's device with get_device(), closing a TOCTOU race condition that could lead to a NULL pointer dereference.

Quoted 5 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +0 -0
index f80a21ec161b6..0f73804487188 100644

drivers/scsi/mpi3mr/mpi3mr_os.c static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc) · line 1102

Quoted 9 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +2 -0
@@ -1094,10 +1094,13 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 {
 	struct mpi3mr_tgt_dev *tgtdev, *tgtdev_next;
 	struct mpi3mr_stgt_priv_data *tgt_priv;
+	struct scsi_target *starget;
+	unsigned long flags;
 
 	dprint_reset(mrioc, "refresh target devices: check for removals\n");
 	list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc->tgtdev_list,
 	    list) {

[Severity: High]
This is a pre-existing issue, but does iterating over mrioc->tgtdev_list without holding tgtdev_lock risk list corruption?

If the top-half event handler (mpi3mr_os_handle_events()) receives a device addition event concurrently, it can call mpi3mr_create_tgtdev(), which mutates the list using list_add_tail() inside mpi3mr_tgtdev_add_to_list() while holding the lock. The lockless reader here will race with the writer.

The same unprotected traversal appears again on the third loop in mpi3mr_refresh_tgtdevs() and in mpi3mr_invalidate_devhandles().

Quoted 4 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
+		spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
 		if (((tgtdev->dev_handle == MPI3MR_INVALID_DEV_HANDLE) ||
 		     tgtdev->is_hidden) &&
 		     tgtdev->host_exposed && tgtdev->starget &&

[ ... ]

Quoted 8 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +2 -2
@@ -1127,15 +1131,25 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 	tgtdev = NULL;
 	list_for_each_entry(tgtdev, &mrioc->tgtdev_list, list) {
 		if ((tgtdev->dev_handle != MPI3MR_INVALID_DEV_HANDLE) &&
-		    !tgtdev->is_hidden) {
-			if (!tgtdev->host_exposed)
+				!tgtdev->is_hidden) {
+			if (!tgtdev->host_exposed) {

drivers/scsi/mpi3mr/mpi3mr_os.c static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc) · line 1140

Quoted 10 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +4 -5
 				mpi3mr_report_tgtdev_to_host(mrioc,
-							     tgtdev->perst_id);
-			else if (tgtdev->starget)
-				starget_for_each_device(tgtdev->starget,
-							(void *)tgtdev, mpi3mr_update_sdev);
-	}
+				     tgtdev->perst_id);
+				continue;
+			}
+			spin_lock_irqsave(&mrioc->tgtdev_lock, flags);

[Severity: High]
This is a pre-existing issue, but does acquiring tgtdev_lock here create a lock inversion against shost->host_lock?

In mpi3mr_set_qd_for_all_vd_in_tg(), the driver acquires mrioc->tgtdev_lock and then calls starget_for_each_device(), which acquires shost->host_lock via __scsi_iterate_devices().

Concurrently, a target teardown (e.g., via sysfs or EH) acquires shost->host_lock before calling the driver's ->target_destroy callback (mpi3mr_target_destroy()). That callback then attempts to acquire mrioc->tgtdev_lock, creating a potential AB-BA deadlock.

drivers/scsi/mpi3mr/mpi3mr_os.c static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc) · line 1144

Quoted 4 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +4 -0
+			starget = tgtdev->starget;
+			if (starget)
+				get_device(&starget->dev);
+			spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);

[Severity: High]
This isn't a bug introduced by this patch, but while this change fixes the unprotected check-then-use window for tgtdev->starget in this function, are there identical vulnerable patterns remaining elsewhere?

For example, mpi3mr_remove_tgtdev_from_host() reads tgtdev->starget and tgtdev->starget->hostdata locklessly:

drivers/scsi/mpi3mr/mpi3mr_os.c:mpi3mr_remove_tgtdev_from_host() {
    ...
    if (tgtdev->starget && tgtdev->starget->hostdata) {
        tgt_priv = tgtdev->starget->hostdata;
    ...
}

Similar unprotected reads occur in mpi3mr_update_tgtdev(), mpi3mr_invalidate_devhandles(), mpi3mr_pcietopochg_evt_th(), and mpi3mr_bsg_process_mpt_cmds(). Can concurrent target teardowns trigger use-after-free conditions in these locations as well?

Quoted 9 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +7 -0
+			if (starget) {
+				starget_for_each_device(starget, (void *)tgtdev,
+							mpi3mr_update_sdev);
+				put_device(&starget->dev);
+			}
+		}
 	}
+	dprint_reset(mrioc, "refresh target devices: done\n");
 }
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Laurence Oberman
Quoted 170 lines · Laurence Oberman

mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh() read tgt_dev->starget and immediately pass it to starget_for_each_device() without holding mrioc->tgtdev_lock. Every writer of this field -- mpi3mr_target_alloc(), mpi3mr_target_destroy(), mpi3mr_slave_destroy()
and mpi3mr_sdev_init() -- correctly serializes access under tgtdev_lock, but these two read sites do not, which leaves a check-then-use window against the SCSI core's target teardown path (scsi_remove_target(), invoked e.g. via a concurrent host reset, sysfs "delete", or SCSI EH device offlining running independently of the fwevt workqueue).

Sequence observed on production hardware, triggered on the mpi3mr0_fwevt_wrkr workqueue during a SAS topology change shortly after a controller reset:

  BUG: kernel NULL pointer dereference, address: 0000000000000058
  RIP: scsi_is_host_device+0x7/0x20
  Call Trace:
   starget_for_each_device+0x34/0x100
   mpi3mr_refresh_tgtdevs+0x152/0x1d0 [mpi3mr]
   mpi3mr_fwevt_bh+0x514/0x6c0 [mpi3mr]
   mpi3mr_fwevt_worker+0x1a/0x50 [mpi3mr]
   process_one_work+0x194/0x380
   worker_thread+0x2fe/0x410

mpi3mr_refresh_tgtdevs() reads tgt_dev->starget as non-NULL, but by the time starget_for_each_device() dereferences it, a concurrent mpi3mr_target_destroy() has already cleared tgt_dev->starget under tgtdev_lock and the SCSI/device core has freed the underlying scsi_target (and its embedded struct device). The stale pointer is then walked by dev_to_shost() -> scsi_is_host_device(), producing the NULL/garbage dereference above.

Fix this by taking mrioc->tgtdev_lock around every read of tgt_dev->starget, matching the existing writer-side discipline. Since starget_for_each_device() and mpi3mr_update_sdev() can end up doing non-atomic work (e.g. queue_limits_commit_update()), the lock cannot be held across the whole call, so instead pin the target's device with get_device() while holding the lock, drop the lock, then run starget_for_each_device() against the pinned reference and put_device() afterwards. This closes the TOCTOU window instead of merely narrowing it.

The same unlocked read-and-dereference pattern also exists earlier in mpi3mr_refresh_tgtdevs()'s first removal-scan loop (tgt_dev->starget->hostdata); fix it the same way by holding tgtdev_lock across that check, which is cheap since it only touches plain struct fields.

Assisted-by Claude:Sonnet5 [Claude Code]

---
 drivers/scsi/mpi3mr/mpi3mr_os.c | 45 +++++++++++++++++++++++++------ --

 1 file changed, 35 insertions(+), 10 deletions(-)
drivers/scsi/mpi3mr/mpi3mr_os.c +2 -0
index 402d1f35d214..922b28c3fe12 100644
@@ -1094,10 +1094,13 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
  {
  	struct mpi3mr_tgt_dev *tgtdev, *tgtdev_next;
  	struct mpi3mr_stgt_priv_data *tgt_priv;
+	struct scsi_target *starget;
+	unsigned long flags;


  dprint_reset(mrioc, "refresh target devices: check for removals\n");
  list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc-

Quoted 1 line

tgtdev_list,

      list) {

drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
+		spin_lock_irqsave(&mrioc->tgtdev_lock, flags);

  if (((tgtdev->dev_handle == MPI3MR_INVALID_DEV_HANDLE) ||
       tgtdev->is_hidden) && tgtdev->host_exposed && tgtdev->starget &&

drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
@@ -1106,6 +1109,7 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
  			tgt_priv->dev_removed = 1;
  			atomic_set(&tgt_priv->block_io, 0);
  		}
+		spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
  	}


  list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc-

Quoted 1 line

tgtdev_list,

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -0
@@ -1127,15 +1131,25 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
  	tgtdev = NULL;
  	list_for_each_entry(tgtdev, &mrioc->tgtdev_list, list) {

  if ((tgtdev->dev_handle !=
MPI3MR_INVALID_DEV_HANDLE) &&

drivers/scsi/mpi3mr/mpi3mr_os.c +2 -2
-		    !tgtdev->is_hidden) {
-			if (!tgtdev->host_exposed)
+				!tgtdev->is_hidden) {
+			if (!tgtdev->host_exposed) {

  mpi3mr_report_tgtdev_to_host(mrioc,

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -1
-							     tgtdev-
Quoted 1 line

perst_id);

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -2
-			else if (tgtdev->starget)
-				starget_for_each_device(tgtdev-
Quoted 1 line

starget,

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -1
-							(void

*)tgtdev, mpi3mr_update_sdev);

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -1
-	}
+				     tgtdev->perst_id);
+				continue;
+			}
+			spin_lock_irqsave(&mrioc->tgtdev_lock,

flags);

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -0
+			starget = tgtdev->starget;
+			if (starget)
+				get_device(&starget->dev);
+			spin_unlock_irqrestore(&mrioc->tgtdev_lock,

flags);

drivers/scsi/mpi3mr/mpi3mr_os.c +2 -0
+			if (starget) {
+				starget_for_each_device(starget,

(void *)tgtdev,

drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
+							mpi3mr_updat

e_sdev);

drivers/scsi/mpi3mr/mpi3mr_os.c +3 -0
+				put_device(&starget->dev);
+			}
+		}

  }

drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
+	dprint_reset(mrioc, "refresh target devices: done\n");

 }

 /**

drivers/scsi/mpi3mr/mpi3mr_os.c +2 -0
@@ -1515,6 +1529,8 @@ static void mpi3mr_devinfochg_evt_bh(struct
 mpi3mr_ioc *mrioc,
  	struct mpi3_device_page0 *dev_pg0)
  {
  	struct mpi3mr_tgt_dev *tgtdev = NULL;
+	struct scsi_target *starget;
+	unsigned long flags;
  	u16 dev_handle = 0, perst_id = 0;


  perst_id = le16_to_cpu(dev_pg0->persistent_id);

drivers/scsi/mpi3mr/mpi3mr_os.c +8 -3
@@ -1535,9 +1551,18 @@ static void mpi3mr_devinfochg_evt_bh(struct
 mpi3mr_ioc *mrioc,
  		mpi3mr_report_tgtdev_to_host(mrioc, perst_id);
  	if (tgtdev->is_hidden && tgtdev->host_exposed)
  		mpi3mr_remove_tgtdev_from_host(mrioc, tgtdev);
-	if (!tgtdev->is_hidden && tgtdev->host_exposed && tgtdev-
 >starget)
-		starget_for_each_device(tgtdev->starget, (void
 *)tgtdev,
-		    mpi3mr_update_sdev);
+	if (!tgtdev->is_hidden && tgtdev->host_exposed) {
+		spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
+		starget = tgtdev->starget;
+		if (starget)
+			get_device(&starget->dev);
+		spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
+		if (starget) {
+			starget_for_each_device(starget, (void

*)tgtdev,

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -0
+						mpi3mr_update_sdev);
+			put_device(&starget->dev);
+		}
+	}

 out:
  if (tgtdev)
  mpi3mr_tgtdev_put(tgtdev);

Quoted 1 line

Re:

sashiko found Pre-existing issues:
- [High] The patch fixes the unprotected check-then-use window for tgtdev->starget in mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh(), but identical vulnerable patterns remain in other functions like mpi3mr_remove_tgtdev_from_host(), mpi3mr_update_tgtdev(), and various event/BSG handlers.
- [High] Unprotected traversal of mrioc->tgtdev_list without holding tgtdev_lock can lead to list corruption and crashes due to concurrent insertions.
- [High] An AB-BA lock inversion exists between mrioc->tgtdev_lock and the SCSI host lock (shost->host_lock).

The sashiko reviews were all for pre-existing issues and won't be fixed in the specific patch submission here.

I will make an attempt to deal with those in a seperate submission as already agreed with Chandrekanth

Thanks
Laurence

#3 lore raw
Quoted 186 lines · Laurence Oberman
Quoted 170 lines · Laurence Oberman

mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh() read tgt_dev->starget and immediately pass it to starget_for_each_device() without holding mrioc->tgtdev_lock. Every writer of this field -- mpi3mr_target_alloc(), mpi3mr_target_destroy(), mpi3mr_slave_destroy()
and mpi3mr_sdev_init() -- correctly serializes access under tgtdev_lock, but these two read sites do not, which leaves a check-then-use window against the SCSI core's target teardown path (scsi_remove_target(), invoked e.g. via a concurrent host reset, sysfs "delete", or SCSI EH device offlining running independently of the fwevt workqueue).

Sequence observed on production hardware, triggered on the mpi3mr0_fwevt_wrkr workqueue during a SAS topology change shortly after a controller reset:

  BUG: kernel NULL pointer dereference, address: 0000000000000058
  RIP: scsi_is_host_device+0x7/0x20
  Call Trace:
   starget_for_each_device+0x34/0x100
   mpi3mr_refresh_tgtdevs+0x152/0x1d0 [mpi3mr]
   mpi3mr_fwevt_bh+0x514/0x6c0 [mpi3mr]
   mpi3mr_fwevt_worker+0x1a/0x50 [mpi3mr]
   process_one_work+0x194/0x380
   worker_thread+0x2fe/0x410

mpi3mr_refresh_tgtdevs() reads tgt_dev->starget as non-NULL, but by the time starget_for_each_device() dereferences it, a concurrent mpi3mr_target_destroy() has already cleared tgt_dev->starget under tgtdev_lock and the SCSI/device core has freed the underlying scsi_target (and its embedded struct device). The stale pointer is then walked by dev_to_shost() -> scsi_is_host_device(), producing the NULL/garbage dereference above.

Fix this by taking mrioc->tgtdev_lock around every read of tgt_dev->starget, matching the existing writer-side discipline. Since starget_for_each_device() and mpi3mr_update_sdev() can end up doing non-atomic work (e.g. queue_limits_commit_update()), the lock cannot be held across the whole call, so instead pin the target's device with get_device() while holding the lock, drop the lock, then run starget_for_each_device() against the pinned reference and put_device() afterwards. This closes the TOCTOU window instead of merely narrowing it.

The same unlocked read-and-dereference pattern also exists earlier in mpi3mr_refresh_tgtdevs()'s first removal-scan loop (tgt_dev->starget->hostdata); fix it the same way by holding tgtdev_lock across that check, which is cheap since it only touches plain struct fields.

Assisted-by Claude:Sonnet5 [Claude Code]
---
 drivers/scsi/mpi3mr/mpi3mr_os.c | 45 +++++++++++++++++++++++++------
--
 1 file changed, 35 insertions(+), 10 deletions(-)
drivers/scsi/mpi3mr/mpi3mr_os.c +3 -0
index 402d1f35d214..922b28c3fe12 100644
@@ -1094,10 +1094,13 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
 {
      struct mpi3mr_tgt_dev *tgtdev, *tgtdev_next;
      struct mpi3mr_stgt_priv_data *tgt_priv;
+     struct scsi_target *starget;
+     unsigned long flags;
 
      dprint_reset(mrioc, "refresh target devices: check for
 removals\n");
      list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc-
 >tgtdev_list,
          list) {
+             spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
              if (((tgtdev->dev_handle ==
MPI3MR_INVALID_DEV_HANDLE) ||
                   tgtdev->is_hidden) &&
                   tgtdev->host_exposed && tgtdev->starget &&
drivers/scsi/mpi3mr/mpi3mr_os.c +1 -0
@@ -1106,6 +1109,7 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
                      tgt_priv->dev_removed = 1;
                      atomic_set(&tgt_priv->block_io, 0);
              }
+             spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
      }
 

list_for_each_entry_safe(tgtdev, tgtdev_next, &mrioc-

Quoted 1 line

tgtdev_list,

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -0
@@ -1127,15 +1131,25 @@ static void mpi3mr_refresh_tgtdevs(struct
 mpi3mr_ioc *mrioc)
      tgtdev = NULL;
      list_for_each_entry(tgtdev, &mrioc->tgtdev_list, list) {
              if ((tgtdev->dev_handle !=

MPI3MR_INVALID_DEV_HANDLE) &&

drivers/scsi/mpi3mr/mpi3mr_os.c +2 -3
-                 !tgtdev->is_hidden) {
-                     if (!tgtdev->host_exposed)
+                             !tgtdev->is_hidden) {
+                     if (!tgtdev->host_exposed) {
                              mpi3mr_report_tgtdev_to_host(mrioc,
-                                                          tgtdev-
Quoted 1 line

perst_id);

drivers/scsi/mpi3mr/mpi3mr_os.c +0 -2
-                     else if (tgtdev->starget)
-                             starget_for_each_device(tgtdev-
Quoted 1 line

starget,

- (void
*)tgtdev, mpi3mr_update_sdev);

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -1
-     }
+                                  tgtdev->perst_id);
+                             continue;
+                     }
+                     spin_lock_irqsave(&mrioc->tgtdev_lock,

flags);

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -0
+                     starget = tgtdev->starget;
+                     if (starget)
+                             get_device(&starget->dev);
+                     spin_unlock_irqrestore(&mrioc->tgtdev_lock,

flags);

drivers/scsi/mpi3mr/mpi3mr_os.c +2 -0
+                     if (starget) {
+                             starget_for_each_device(starget,

(void *)tgtdev,
+ mpi3mr_updat
e_sdev);

drivers/scsi/mpi3mr/mpi3mr_os.c +6 -0
+                             put_device(&starget->dev);
+                     }
+             }
      }
+     dprint_reset(mrioc, "refresh target devices: done\n");
 }
 
 /**
@@ -1515,6 +1529,8 @@ static void mpi3mr_devinfochg_evt_bh(struct
 mpi3mr_ioc *mrioc,
      struct mpi3_device_page0 *dev_pg0)
 {
      struct mpi3mr_tgt_dev *tgtdev = NULL;
+     struct scsi_target *starget;
+     unsigned long flags;
      u16 dev_handle = 0, perst_id = 0;
 
      perst_id = le16_to_cpu(dev_pg0->persistent_id);
drivers/scsi/mpi3mr/mpi3mr_os.c +8 -3
@@ -1535,9 +1551,18 @@ static void mpi3mr_devinfochg_evt_bh(struct
 mpi3mr_ioc *mrioc,
              mpi3mr_report_tgtdev_to_host(mrioc, perst_id);
      if (tgtdev->is_hidden && tgtdev->host_exposed)
              mpi3mr_remove_tgtdev_from_host(mrioc, tgtdev);
-     if (!tgtdev->is_hidden && tgtdev->host_exposed && tgtdev-
 >starget)
-             starget_for_each_device(tgtdev->starget, (void
 *)tgtdev,
-                 mpi3mr_update_sdev);
+     if (!tgtdev->is_hidden && tgtdev->host_exposed) {
+             spin_lock_irqsave(&mrioc->tgtdev_lock, flags);
+             starget = tgtdev->starget;
+             if (starget)
+                     get_device(&starget->dev);
+             spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);
+             if (starget) {
+                     starget_for_each_device(starget, (void

*)tgtdev,

drivers/scsi/mpi3mr/mpi3mr_os.c +4 -0
+                                             mpi3mr_update_sdev);
+                     put_device(&starget->dev);
+             }
+     }
 out:
      if (tgtdev)
              mpi3mr_tgtdev_put(tgtdev);
Quoted 1 line

Re:

sashiko found Pre-existing issues:
- [High] The patch fixes the unprotected check-then-use window for tgtdev->starget in mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh(), but identical vulnerable patterns remain in other functions like mpi3mr_remove_tgtdev_from_host(), mpi3mr_update_tgtdev(), and various event/BSG handlers.
- [High] Unprotected traversal of mrioc->tgtdev_list without holding tgtdev_lock can lead to list corruption and crashes due to concurrent insertions.
- [High] An AB-BA lock inversion exists between mrioc->tgtdev_lock and the SCSI host lock (shost->host_lock).

Quoted 9 lines

The sashiko reviews were all for pre-existing issues and won't be fixed in the specific patch submission here.

I will make an attempt to deal with those in a seperate submission as already agreed with Chandrekanth

Thanks
Laurence

Looks good to me. Thanks.

#4 lore raw

[BUG] scsi: mpi3mr: deadlock when throttling changes queue depth

In reply to Laurence Oberman

Hi there,

Quoted 39 lines · On Aug 31, 2026, at 5:18 AM, [email protected] wrote:
Quoted 9 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +0 -0
@@ -1127,15 +1131,25 @@ static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc)
 tgtdev = NULL;
 list_for_each_entry(tgtdev, &mrioc->tgtdev_list, list) {

if ((tgtdev->dev_handle != MPI3MR_INVALID_DEV_HANDLE) &&
- !tgtdev->is_hidden) {
- if (!tgtdev->host_exposed)
+ !tgtdev->is_hidden) {
+ if (!tgtdev->host_exposed) { mpi3mr_report_tgtdev_to_host(mrioc,

drivers/scsi/mpi3mr/mpi3mr_os.c static void mpi3mr_refresh_tgtdevs(struct mpi3mr_ioc *mrioc) · line 1140

Quoted 9 lines
drivers/scsi/mpi3mr/mpi3mr_os.c +4 -5
-      tgtdev->perst_id);
- else if (tgtdev->starget)
- starget_for_each_device(tgtdev->starget,
- (void *)tgtdev, mpi3mr_update_sdev);
- }
+      tgtdev->perst_id);
+ continue;
+ }
+ spin_lock_irqsave(&mrioc->tgtdev_lock, flags);

[Severity: High]
This is a pre-existing issue, but does acquiring tgtdev_lock here create a lock inversion against shost->host_lock?

In mpi3mr_set_qd_for_all_vd_in_tg(), the driver acquires mrioc->tgtdev_lock and then calls starget_for_each_device(), which acquires shost->host_lock via __scsi_iterate_devices().

Concurrently, a target teardown (e.g., via sysfs or EH) acquires shost->host_lock before calling the driver's ->target_destroy callback (mpi3mr_target_destroy()). That callback then attempts to acquire mrioc->tgtdev_lock, creating a potential AB-BA deadlock.

Quoted 4 lines

+ starget = tgtdev->starget;
+ if (starget)
+ get_device(&starget->dev);
+ spin_unlock_irqrestore(&mrioc->tgtdev_lock, flags);

--
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1

Quoted 5 lines · Laurence Oberman

sashiko found Pre-existing issues:
[...]
- [High] An AB-BA lock inversion exists between mrioc->tgtdev_lock and the SCSI host lock (shost->host_lock).

Quoted 8 lines

The sashiko reviews were all for pre-existing issues and won't be fixed in the specific patch submission here.

I will make an attempt to deal with those in a seperate submission as already agreed with Chandrekanth

Thanks
Laurence

Some of our production servers have been deadlocking due to the pre-existing lock inversion that Sashiko pointed out here, which was introduced in cf1ce8b71524 ("scsi: mpi3mr: Reduce VD queue depth on detecting throttling"). Has anyone followed up on this?

Our quick fix was to change mpi3mr_set_qd_for_all_vd_in_tg() to take shost->host_lock first, and use __starget_for_each_device() instead of starget_for_each_device(). We also changed mpi3mr_update_sdev_qd() to filter out devices in the SDEV_CANCEL or SDEV_DEL state, since that was previously done by scsi_device_get() in starget_for_each_device(), but we're not sure if that's strictly necessary.

But our LLMs pointed out another issue in this code path: mpi3mr_update_sdev_qd() runs with interrupts disabled by these spinlocks, but it calls scsi_change_queue_depth(), which seemingly isn't meant to be called in that context. In particular, if blk-iocost is attached, ioc_rqos_queue_depth_changed() will call spin_unlock_irq() and re-enable interrupts prematurely. The LLMs proposed a more involved patch to avoid holding any locks when calling scsi_change_queue_depth(), but we weren't sure if it was correct. But we can send it along if it would be useful.

(Thanks to our summer intern Albert Du (CC'ed) for all his great work on this issue!)

maintainer Applied
#5 lore raw
In reply to Laurence Oberman
Quoted 11 lines · Laurence Oberman

mpi3mr_refresh_tgtdevs() and mpi3mr_devinfochg_evt_bh() read tgt_dev->starget and immediately pass it to starget_for_each_device() without holding mrioc->tgtdev_lock. Every writer of this field -- mpi3mr_target_alloc(), mpi3mr_target_destroy(), mpi3mr_slave_destroy() and mpi3mr_sdev_init() -- correctly serializes access under tgtdev_lock, but these two read sites do not, which leaves a check-then-use window against the SCSI core's target teardown path (scsi_remove_target(), invoked e.g. via a concurrent host reset, sysfs "delete", or SCSI EH device offlining running independently of the fwevt workqueue).

Quoted 1 line

[...]

Applied to 7.3/scsi-fixes, thanks!

[1/1] scsi: mpi3mr: fix use-after-free on tgt_dev->starget during target device refresh/update

      https://git.kernel.org/mkp/scsi/c/139f57343b3d
Signature · 1 line
Martin K. Petersen