Skip to contentSkip to conversation

scsi: fcoe: Fix potential deadlock on &fip->ctlr_lock

PATCH RESEND Superseded Chengfeng Ye 4 participants 7 messages Aug 16, 2023 – Aug 17, 2023

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH RESEND 1/1 scsi: fcoe: Fix potential deadlock on &fip->ctlr_lock

There is a long call chain that &fip->ctlr_lock is acquired by isr fnic_isr_msix_wq_copy() under hard irq context. Thus other process context code acquiring the lock should disable irq, otherwise deadlock could happen if the irq preempt the execution while the lock is held in process context on the same CPU.

[ISR]
fnic_isr_msix_wq_copy()
-> fnic_wq_copy_cmpl_handler()
-> fnic_fcpio_cmpl_handler()
-> fnic_fcpio_flogi_reg_cmpl_handler()
-> fnic_flush_tx()
-> fnic_send_frame()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

[Process Context]
1. fcoe_ctlr_timer_work()
-> fcoe_ctlr_flogi_send()
-> spin_lock_bh(&fip->ctlr_lock)

2. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_announce()
-> spin_lock_bh(&fip->ctlr_lock)

3. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_flogi_retry()
-> spin_lock_bh(&fip->ctlr_lock)

4. -> fcoe_xmit()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

spin_lock_bh() is not enough since fnic_isr_msix_wq_copy() is a hardirq.

These flaws were found by an experimental static analysis tool I am developing for irq-related deadlock.

The patch fix the potential deadlocks by spin_lock_irqsave() to disable hard irq.

---
 drivers/scsi/fcoe/fcoe_ctlr.c | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
drivers/scsi/fcoe/fcoe_ctlr.c +12 -8
index 5c8d1ba3f8f3..19eee108db02 100644
@@ -319,16 +319,17 @@ static void fcoe_ctlr_announce(struct fcoe_ctlr *fip)
 {
 	struct fcoe_fcf *sel;
 	struct fcoe_fcf *fcf;
+	unsigned long flags;
 
 	mutex_lock(&fip->ctlr_mutex);
-	spin_lock_bh(&fip->ctlr_lock);
+	spin_lock_irqsave(&fip->ctlr_lock, flags);
 
 	kfree_skb(fip->flogi_req);
 	fip->flogi_req = NULL;
 	list_for_each_entry(fcf, &fip->fcfs, list)
 		fcf->flogi_sent = 0;
 
-	spin_unlock_bh(&fip->ctlr_lock);
+	spin_unlock_irqrestore(&fip->ctlr_lock, flags);
 	sel = fip->sel_fcf;
 
 	if (sel && ether_addr_equal(sel->fcf_mac, fip->dest_addr))
@@ -699,6 +700,7 @@ int fcoe_ctlr_els_send(struct fcoe_ctlr *fip, struct fc_lport *lport,
 {
 	struct fc_frame *fp;
 	struct fc_frame_header *fh;
+	unsigned long flags;
 	u16 old_xid;
 	u8 op;
 	u8 mac[ETH_ALEN];
@@ -732,11 +734,11 @@ int fcoe_ctlr_els_send(struct fcoe_ctlr *fip, struct fc_lport *lport,
 		op = FIP_DT_FLOGI;
 		if (fip->mode == FIP_MODE_VN2VN)
 			break;
-		spin_lock_bh(&fip->ctlr_lock);
+		spin_lock_irqsave(&fip->ctlr_lock, flags);
 		kfree_skb(fip->flogi_req);
 		fip->flogi_req = skb;
 		fip->flogi_req_send = 1;
-		spin_unlock_bh(&fip->ctlr_lock);
+		spin_unlock_irqrestore(&fip->ctlr_lock, flags);
 		schedule_work(&fip->timer_work);
 		return -EINPROGRESS;
 	case ELS_FDISC:
@@ -1705,10 +1707,11 @@ static int fcoe_ctlr_flogi_send_locked(struct fcoe_ctlr *fip)
 static int fcoe_ctlr_flogi_retry(struct fcoe_ctlr *fip)
 {
 	struct fcoe_fcf *fcf;
+	unsigned long flags;
 	int error;
 
 	mutex_lock(&fip->ctlr_mutex);
-	spin_lock_bh(&fip->ctlr_lock);
+	spin_lock_irqsave(&fip->ctlr_lock, flags);
 	LIBFCOE_FIP_DBG(fip, "re-sending FLOGI - reselect\n");
 	fcf = fcoe_ctlr_select(fip);
 	if (!fcf || fcf->flogi_sent) {
@@ -1719,7 +1722,7 @@ static int fcoe_ctlr_flogi_retry(struct fcoe_ctlr *fip)
 		fcoe_ctlr_solicit(fip, NULL);
 		error = fcoe_ctlr_flogi_send_locked(fip);
 	}
-	spin_unlock_bh(&fip->ctlr_lock);
+	spin_unlock_irqrestore(&fip->ctlr_lock, flags);
 	mutex_unlock(&fip->ctlr_mutex);
 	return error;
 }
@@ -1736,8 +1739,9 @@ static int fcoe_ctlr_flogi_retry(struct fcoe_ctlr *fip)
 static void fcoe_ctlr_flogi_send(struct fcoe_ctlr *fip)
 {
 	struct fcoe_fcf *fcf;
+	unsigned long flags;
 
-	spin_lock_bh(&fip->ctlr_lock);
+	spin_lock_irqsave(&fip->ctlr_lock, flags);
 	fcf = fip->sel_fcf;
 	if (!fcf || !fip->flogi_req_send)
 		goto unlock;
@@ -1764,7 +1768,7 @@ static void fcoe_ctlr_flogi_send(struct fcoe_ctlr *fip)
 	} else /* XXX */
 		LIBFCOE_FIP_DBG(fip, "No FCF selected - defer send\n");
 unlock:
-	spin_unlock_bh(&fip->ctlr_lock);
+	spin_unlock_irqrestore(&fip->ctlr_lock, flags);
 }
 
 /**
Signature · 1 line
2.17.1
#1 lore raw

Dear maintainers,

May I ask if anyone could help review the patch?

Thanks,
Chengfeng

maintainer Reviewed-by on v1 1/1
#2 lore raw
In reply to Chengfeng Ye
Quoted 43 lines · Chengfeng Ye

There is a long call chain that &fip->ctlr_lock is acquired by isr fnic_isr_msix_wq_copy() under hard irq context. Thus other process context code acquiring the lock should disable irq, otherwise deadlock could happen if the irq preempt the execution while the lock is held in process context on the same CPU.

[ISR]
fnic_isr_msix_wq_copy()
-> fnic_wq_copy_cmpl_handler()
-> fnic_fcpio_cmpl_handler()
-> fnic_fcpio_flogi_reg_cmpl_handler()
-> fnic_flush_tx()
-> fnic_send_frame()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

[Process Context]
1. fcoe_ctlr_timer_work()
-> fcoe_ctlr_flogi_send()
-> spin_lock_bh(&fip->ctlr_lock)

2. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_announce()
-> spin_lock_bh(&fip->ctlr_lock)

3. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_flogi_retry()
-> spin_lock_bh(&fip->ctlr_lock)

4. -> fcoe_xmit()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

spin_lock_bh() is not enough since fnic_isr_msix_wq_copy() is a hardirq.

These flaws were found by an experimental static analysis tool I am developing for irq-related deadlock.

Quoted 8 lines

The patch fix the potential deadlocks by spin_lock_irqsave() to disable hard irq.

---
  drivers/scsi/fcoe/fcoe_ctlr.c | 20 ++++++++++++--------
  1 file changed, 12 insertions(+), 8 deletions(-)

Cheers,

Hannes

Signature · 5 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew
Myers, Andrew McDonald, Martje Boudien Moerman
Reviewed-by on v1 1/1
#4 lore raw
In reply to Chengfeng Ye
Quoted 2 lines · Chengfeng Ye

These flaws were found by an experimental static analysis tool I am developing for irq-related deadlock.

So you did not use lockdep or is the code path so unlikely that nobody stumbled upon it before?

Quoted 2 lines

The patch fix the potential deadlocks by spin_lock_irqsave() to disable hard irq.

Shouldn't this have

Fixes 794d98e77f590 ("[SCSI] libfcoe: retry rejected FLOGI to another FCF if possible")

?

Quoted 1 line

Sebastian

#5 lore raw
Quoted 2 lines

So you did not use lockdep or is the code path so unlikely that nobody stumbled upon it before?

I think so, the lock acquisition inside the interrupt is guarded by an "unlikely"
if-branch condition inside fnic_send_frame(), so it happens that no one has encountered the problem so far by using lockdep.

Quoted 1 line
Fixes 794d98e77f590 ("[SCSI] libfcoe: retry rejected FLOGI to another FCF if possible")

Just sent v2 patch to add the fixes tag

Thanks for the review,
Chengfeng

Reviewed-by on v1 1/1
#6 lore raw
In reply to Chengfeng Ye
Quoted 37 lines · Chengfeng Ye

There is a long call chain that &fip->ctlr_lock is acquired by isr fnic_isr_msix_wq_copy() under hard irq context. Thus other process context code acquiring the lock should disable irq, otherwise deadlock could happen if the irq preempt the execution while the lock is held in process context on the same CPU.

[ISR]
fnic_isr_msix_wq_copy()
-> fnic_wq_copy_cmpl_handler()
-> fnic_fcpio_cmpl_handler()
-> fnic_fcpio_flogi_reg_cmpl_handler()
-> fnic_flush_tx()
-> fnic_send_frame()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

[Process Context]
1. fcoe_ctlr_timer_work()
-> fcoe_ctlr_flogi_send()
-> spin_lock_bh(&fip->ctlr_lock)

2. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_announce()
-> spin_lock_bh(&fip->ctlr_lock)

3. fcoe_ctlr_recv_work()
-> fcoe_ctlr_recv_handler()
-> fcoe_ctlr_recv_els()
-> fcoe_ctlr_flogi_retry()
-> spin_lock_bh(&fip->ctlr_lock)

4. -> fcoe_xmit()
-> fcoe_ctlr_els_send()
-> spin_lock_bh(&fip->ctlr_lock)

Quoted 8 lines

spin_lock_bh() is not enough since fnic_isr_msix_wq_copy() is a hardirq.

These flaws were found by an experimental static analysis tool I am developing for irq-related deadlock.

The patch fix the potential deadlocks by spin_lock_irqsave() to disable hard irq.