[v2] usb: cdnsp: Fix deadlock issue in cdnsp_thread_irq_handler

Commit Message

Pawel Laszczak May 26, 2021, 6:05 a.m. UTC

From: Pawel Laszczak <[e-mail]>

Patch fixes the following critical issue caused by deadlock which has been
detected during testing NCM class:

smp: csd: Detected non-responsive CSD lock (#1) on CPU#0
smp:     csd: CSD lock (#1) unresponsive.
....
RIP: 0010:native_queued_spin_lock_slowpath+0x61/0x1d0
RSP: 0018:ffffbc494011cde0 EFLAGS: 00000002
RAX: 0000000000000101 RBX: ffff9ee8116b4a68 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9ee8116b4658
RBP: ffffbc494011cde0 R08: 0000000000000001 R09: 0000000000000000
R10: ffff9ee8116b4670 R11: 0000000000000000 R12: ffff9ee8116b4658
R13: ffff9ee8116b4670 R14: 0000000000000246 R15: ffff9ee8116b4658
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f7bcc41a830 CR3: 000000007a612003 CR4: 00000000001706e0
Call Trace:
 <IRQ>
 do_raw_spin_lock+0xc0/0xd0
 _raw_spin_lock_irqsave+0x95/0xa0
 cdnsp_gadget_ep_queue.cold+0x88/0x107 [cdnsp_udc_pci]
 usb_ep_queue+0x35/0x110
 eth_start_xmit+0x220/0x3d0 [u_ether]
 ncm_tx_timeout+0x34/0x40 [usb_f_ncm]
 ? ncm_free_inst+0x50/0x50 [usb_f_ncm]
 __hrtimer_run_queues+0xac/0x440
 hrtimer_run_softirq+0x8c/0xb0
 __do_softirq+0xcf/0x428
 asm_call_irq_on_stack+0x12/0x20
 </IRQ>
 do_softirq_own_stack+0x61/0x70
 irq_exit_rcu+0xc1/0xd0
 sysvec_apic_timer_interrupt+0x52/0xb0
 asm_sysvec_apic_timer_interrupt+0x12/0x20
RIP: 0010:do_raw_spin_trylock+0x18/0x40
RSP: 0018:ffffbc494138bda8 EFLAGS: 00000246
RAX: 0000000000000000 RBX: ffff9ee8116b4658 RCX: 0000000000000000
RDX: 0000000000000001 RSI: 0000000000000000 RDI: ffff9ee8116b4658
RBP: ffffbc494138bda8 R08: 0000000000000001 R09: 0000000000000000
R10: ffff9ee8116b4670 R11: 0000000000000000 R12: ffff9ee8116b4658
R13: ffff9ee8116b4670 R14: ffff9ee7b5c73d80 R15: ffff9ee8116b4000
 _raw_spin_lock+0x3d/0x70
 ? cdnsp_thread_irq_handler.cold+0x32/0x112c [cdnsp_udc_pci]
 cdnsp_thread_irq_handler.cold+0x32/0x112c [cdnsp_udc_pci]
 ? cdnsp_remove_request+0x1f0/0x1f0 [cdnsp_udc_pci]
 ? cdnsp_thread_irq_handler+0x5/0xa0 [cdnsp_udc_pci]
 ? irq_thread+0xa0/0x1c0
 irq_thread_fn+0x28/0x60
 irq_thread+0x105/0x1c0
 ? __kthread_parkme+0x42/0x90
 ? irq_forced_thread_fn+0x90/0x90
 ? wake_threads_waitq+0x30/0x30
 ? irq_thread_check_affinity+0xe0/0xe0
 kthread+0x12a/0x160
 ? kthread_park+0x90/0x90
 ret_from_fork+0x22/0x30

The root cause of issue is spin_lock/spin_unlock instruction instead
spin_lock_irqsave/spin_lock_irqrestore in cdnsp_thread_irq_handler
function.

Cc: [e-mail]
Fixes: 3d82904559f4 ("usb: cdnsp: cdns3 Add main part of Cadence USBSSP DRD Driver")
Signed-off-by: Pawel Laszczak <[e-mail]>

---
Changelog:
v2:
- updated commit message

 drivers/usb/cdns3/cdnsp-ring.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

Patch

diff --git a/drivers/usb/cdns3/cdnsp-ring.c b/drivers/usb/cdns3/cdnsp-ring.c
index 5f0513c96c04..68972746e363 100644
--- a/drivers/usb/cdns3/cdnsp-ring.c
+++ b/drivers/usb/cdns3/cdnsp-ring.c
@@ -1517,13 +1517,14 @@  irqreturn_t cdnsp_thread_irq_handler(int irq, void *data)
 {
 	struct cdnsp_device *pdev = (struct cdnsp_device *)data;
 	union cdnsp_trb *event_ring_deq;
+	unsigned long flags;
 	int counter = 0;
 
-	spin_lock(&pdev->lock);
+	spin_lock_irqsave(&pdev->lock, flags);
 
 	if (pdev->cdnsp_state & (CDNSP_STATE_HALTED | CDNSP_STATE_DYING)) {
 		cdnsp_died(pdev);
-		spin_unlock(&pdev->lock);
+		spin_unlock_irqrestore(&pdev->lock, flags);
 		return IRQ_HANDLED;
 	}
 
@@ -1539,7 +1540,7 @@  irqreturn_t cdnsp_thread_irq_handler(int irq, void *data)
 
 	cdnsp_update_erst_dequeue(pdev, event_ring_deq, 1);
 
-	spin_unlock(&pdev->lock);
+	spin_unlock_irqrestore(&pdev->lock, flags);
 
 	return IRQ_HANDLED;
 }