Skip to contentSkip to conversation

phylib: Fix deadlock on resume

PATCH Applied Anton Vorontsov 2 participants 3 messages Dec 22, 2009 – Dec 28, 2009

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/3 phylib: Fix deadlock on resume

Sometimes kernel hangs on resume with the following trace:

ucc_geth e0102000.ucc: resume

 INFO: task bash:1764 blocked for more than 120 seconds.
 "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
 bash          D 0fecf43c     0  1764   1763 0x00000000
 Call Trace:
 [cf9a7c10] [c0012868] ret_from_except+0x0/0x14 (unreliable)
 --- Exception: cf9a7ce0 at __switch_to+0x4c/0x6c
     LR = 0xcf9a7cc0
 [cf9a7cd0] [c0008c14] __switch_to+0x4c/0x6c (unreliable)
 [cf9a7ce0] [c028bcfc] schedule+0x158/0x260
 [cf9a7d10] [c028c720] __mutex_lock_slowpath+0x80/0xd8
 [cf9a7d40] [c01cf388] phy_stop+0x20/0x70
 [cf9a7d50] [c01d514c] ugeth_resume+0x6c/0x13c
 [...]

Here is why.

On suspend:

- PM core starts suspending devices, ucc_geth_suspend gets called;

- ucc_geth calls phy_stop() on suspend. Note that phy_stop() is mostly asynchronous so it doesn't block ucc_geth's suspend routine, it just sets PHY_HALTED state and disables PHY's interrupts;

- Suddenly the state machine gets scheduled, it grabs the phydev->lock mutex and tries to process the PHY_HALTED state, so it calls phydev->adjust_link(phydev->attached_dev). In ucc_geth case adjust_link() calls msleep(), which reschedules the code flow back to PM core, which now finishes suspend and so we end up sleeping with phydev->lock mutex held.

On resume:

- PM core starts resuming devices (notice that nobody rescheduled the state machine yet, so the mutex is still held), the core calls ucc_geth's resume routine;

- ucc_geth_resume restarts the PHY with phy_stop()/phy_start() sequence, and the phy_*() calls are trying to grab the phydev->lock mutex. Here comes the deadlock.

This patch fixes the issue by stopping the state machine on suspend and starting it again on resume.

---
 drivers/net/phy/mdio_bus.c |   24 ++++++++++++++++++++++--
 1 files changed, 22 insertions(+), 2 deletions(-)
drivers/net/phy/mdio_bus.c +22 -2
index bd4e8d7..49252d3 100644
@@ -303,8 +303,18 @@ static int mdio_bus_suspend(struct device * dev, pm_message_t state)
 	struct phy_driver *phydrv = to_phy_driver(dev->driver);
 	struct phy_device *phydev = to_phy_device(dev);
 
+	/*
+	 * We must stop the state machine manually, otherwise it stops out of
+	 * control, possibly with the phydev->lock held. Upon resume, netdev
+	 * may call phy routines that try to grab the same lock, and that may
+	 * lead to a deadlock.
+	 */
+	if (phydev->attached_dev)
+		phy_stop_machine(phydev);
+
 	if (!mdio_bus_phy_may_suspend(phydev))
 		return 0;
+
 	return phydrv->suspend(phydev);
 }
 
@@ -312,10 +322,20 @@ static int mdio_bus_resume(struct device * dev)
 {
 	struct phy_driver *phydrv = to_phy_driver(dev->driver);
 	struct phy_device *phydev = to_phy_device(dev);
+	int ret;
 
 	if (!mdio_bus_phy_may_suspend(phydev))
-		return 0;
-	return phydrv->resume(phydev);
+		goto no_resume;
+
+	ret = phydrv->resume(phydev);
+	if (ret < 0)
+		return ret;
+
+no_resume:
+	if (phydev->attached_dev)
+		phy_start_machine(phydev, NULL);
+
+	return 0;
 }
 
 struct bus_type mdio_bus_type = {
Signature · 1 line
1.6.3.3
maintainer Applied
#1 lore raw

From: Anton Vorontsov <[email protected]>
Date: Wed, 23 Dec 2009 01:14:47 +0300

Quoted 1 line

Sometimes kernel hangs on resume with the following trace:

...

Quoted 4 lines

This patch fixes the issue by stopping the state machine on suspend and starting it again on resume.

Applied.