scsi: lpfc: Fix potential memory leak
Subsystems (2)
PATCH 1/1 scsi: lpfc: Fix potential memory leak
In lpfc_bsg_hba_get_event() and lpfc_bsg_hba_set_event(), we should keep the refcount balance when there is some error or the *evt* will be replaced.
--- drivers/scsi/lpfc/lpfc_bsg.c | 8 ++++++++ 1 file changed, 8 insertions(+)
drivers/scsi/lpfc/lpfc_bsg.c +8 -0
@@ -1200,6 +1200,11 @@ lpfc_bsg_hba_set_event(struct bsg_job *job)spin_unlock_irqrestore (&phba->ct_ev_lock, flags);if (&evt->node == &phba->ct_ev_waiters) { + +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); +/* no event waiting struct yet - first call */ dd_data =kmalloc (sizeof (struct bsg_job_data ),GFP_KERNEL );if (dd_data ==NULL ) {
@@ -1292,6 +1297,9 @@ lpfc_bsg_hba_get_event(struct bsg_job *job)* an error indicating that there isn't anymore */ if (evt_dat ==NULL ) { +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); bsg_reply->reply_payload_rcv_len =0 ; rc = -ENOENT ; goto job_error;
Signature · 1 line
2.25.1
Hi Liang,
@@ -1200,6 +1200,11 @@ lpfc_bsg_hba_set_event(struct bsg_job *job)
spin_unlock_irqrestore(&phba->ct_ev_lock, flags);
if (&evt->node == &phba->ct_ev_waiters) {
+
+ spin_lock_irqsave(&phba->ct_ev_lock, flags);
+ lpfc_bsg_event_unref(evt);
+ spin_unlock_irqrestore(&phba->ct_ev_lock, flags);
+
/* no event waiting struct yet - first call */
dd_data = kmalloc(sizeof(struct bsg_job_data), GFP_KERNEL);
if (dd_data == NULL) {if (&evt->node == &phba->ct_ev_waiters) was true, then that means the ct_ev_waiters list was empty so the *evt ptr is not pointing to a real lpfc_bsg_event structure. I think we would even crash trying to decrement evt’s kref because evt is an uninitialized lpfc_bsg_event structure and not pointing to anything real. We’re about to kzalloc a new evt with lpfc_bsg_event_new in this same if statement clause anyways. I’m not clear on the reasoning behind the suggestion to call lpfc_bsg_event_unref(evt) there.
@@ -1292,6 +1297,9 @@ lpfc_bsg_hba_get_event(struct bsg_job *job)
* an error indicating that there isn't anymore
*/
if (evt_dat == NULL) {
+ spin_lock_irqsave(&phba->ct_ev_lock, flags);
+ lpfc_bsg_event_unref(evt);
+ spin_unlock_irqrestore(&phba->ct_ev_lock, flags);
bsg_reply->reply_payload_rcv_len = 0;
rc = -ENOENT;
goto job_error;Similar argument here. If (evt_dat == NULL) was true, then that means there was nothing of interest in the ct_ev_waiters list or that the ct_ev_waiters list was empty. In either case, we shouldn’t call lpfc_bsg_event_unref(evt) because we would be decrementing the wrong reg_id evt’s kref or even crash if the ct_ev_waiters list was empty trying to kref_put an uninitialized lpfc_bsg_event structure.
Regards,
Justin
Quoted 40 lines · Liang He
In
lpfc_bsg_hba_get_event()andlpfc_bsg_hba_set_event(), we should keep the refcount balance when there is some error or the *evt* will be replaced.Fixes f1c3b0fcbb81 ("[SCSI] lpfc 8.3.4: Add bsg (SGIOv4) support for ELS/CT support")Signed-off-by Liang He <[email protected]>--- drivers/scsi/lpfc/lpfc_bsg.c | 8 ++++++++ 1 file changed, 8 insertions(+)drivers/scsi/lpfc/lpfc_bsg.c +8
-1@@ -1200,6 +1200,11 @@ lpfc_bsg_hba_set_event(struct bsg_job *job)spin_unlock_irqrestore (&phba->ct_ev_lock, flags);if (&evt->node == &phba->ct_ev_waiters) { + +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); +/* no event waiting struct yet - first call */ dd_data =kmalloc (sizeof (struct bsg_job_data ),GFP_KERNEL );if (dd_data ==NULL ) {@@ -1292,6 +1297,9 @@ lpfc_bsg_hba_get_event(struct bsg_job *job)* an error indicating that there isn't anymore */ if (evt_dat ==NULL ) { +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); bsg_reply->reply_payload_rcv_len =0 ; rc = -ENOENT ; goto job_error; --2.25.1
Quoted 43 lines · Justin Tee
Hi Liang,
drivers/scsi/lpfc/lpfc_bsg.c +5
-0@@ -1200,6 +1200,11 @@ lpfc_bsg_hba_set_event(struct bsg_job *job)spin_unlock_irqrestore (&phba->ct_ev_lock, flags);if (&evt->node == &phba->ct_ev_waiters) { + +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); +/* no event waiting struct yet - first call */ dd_data =kmalloc (sizeof (struct bsg_job_data ),GFP_KERNEL );if (dd_data ==NULL ) {if (
&evt->node==&phba->ct_ev_waiters) was true, then that means thect_ev_waiterslist was empty so the *evt ptr is not pointing to a reallpfc_bsg_eventstructure. I think we would even crash trying to decrement evt’s kref because evt is an uninitializedlpfc_bsg_eventstructure and not pointing to anything real. We’re about to kzalloc a new evt withlpfc_bsg_event_newin this same if statement clause anyways. I’m not clear on the reasoning behind the suggestion to calllpfc_bsg_event_unref(evt)there.drivers/scsi/lpfc/lpfc_bsg.c +3
-0@@ -1292,6 +1297,9 @@ lpfc_bsg_hba_get_event(struct bsg_job *job)* an error indicating that there isn't anymore */ if (evt_dat ==NULL ) { +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); bsg_reply->reply_payload_rcv_len =0 ; rc = -ENOENT ; goto job_error;Similar argument here. If (
evt_dat== NULL) was true, then that means there was nothing of interest in thect_ev_waiterslist or that thect_ev_waiterslist was empty. In either case, we shouldn’t calllpfc_bsg_event_unref(evt)because we would be decrementing the wrongreg_idevt’s kref or even crash if thect_ev_waiterslist was empty trying tokref_putan uninitializedlpfc_bsg_eventstructure.
Quoted 2 lines
Regards,
Justin
Hi, Justin,
Thanks for your kindly explaination.
My patch is indeed wrong, as I have not realized the semantic of "&evt->node == &phba->ct_ev_waiters".
Thanks and sorry for my trouble.
Liang
Quoted 42 lines · Liang He
Quoted 40 lines
In
lpfc_bsg_hba_get_event()andlpfc_bsg_hba_set_event(), we should keep the refcount balance when there is some error or the *evt* will be replaced.Fixes f1c3b0fcbb81 ("[SCSI] lpfc 8.3.4: Add bsg (SGIOv4) support for ELS/CT support")Signed-off-by Liang He <[email protected]>--- drivers/scsi/lpfc/lpfc_bsg.c | 8 ++++++++ 1 file changed, 8 insertions(+)drivers/scsi/lpfc/lpfc_bsg.c +8
-1@@ -1200,6 +1200,11 @@ lpfc_bsg_hba_set_event(struct bsg_job *job)spin_unlock_irqrestore (&phba->ct_ev_lock, flags);if (&evt->node == &phba->ct_ev_waiters) { + +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); +/* no event waiting struct yet - first call */ dd_data =kmalloc (sizeof (struct bsg_job_data ),GFP_KERNEL );if (dd_data ==NULL ) {@@ -1292,6 +1297,9 @@ lpfc_bsg_hba_get_event(struct bsg_job *job)* an error indicating that there isn't anymore */ if (evt_dat ==NULL ) { +spin_lock_irqsave (&phba->ct_ev_lock, flags); +lpfc_bsg_event_unref (evt); +spin_unlock_irqrestore (&phba->ct_ev_lock, flags); bsg_reply->reply_payload_rcv_len =0 ; rc = -ENOENT ; goto job_error; --2.25.1