Fix fakephp deadlock

From: Ian Abbott <[email protected]>

I last sent this 11 days ago and didn't get any replies, so my timeout handler has kicked in. :)

The patch is same as last time except I've bumped the revision from rc7 to rc8 and have remembered to use diff's -p option this time!

---
From: Ian Abbott <[email protected]>

If the fakephp driver is used to emulate removal of a PCI device by writing text string "0" to the "power" sysfs attribute file, this causes its parent directory and its contents (including the "power" file) to be deleted before the write operation returns. Unfortunately, it ends up in a deadlock waiting for itself to complete.

The deadlock is as follows: sysfs_write_file calls flush_write_buffer which calls sysfs_get_active_two before calling power_write_file in pci_hotplug_core.c via the sysfs store operation. The power_write_file function calls disable_slot in fakephp.c via the slot operation. The disable_slot function calls remove_slot which calls pci_hp_deregister (back in pci_hotplug_core.c) which calls fs_remove_slot which calls sysfs_remove_file to remove the "power" file. The sysfs_remove_file function calls sysfs_hash_and_remove which calls sysfs_addrm_finish which calls sysfs_deactivate. The sysfs_deactivate function sees that something has an active reference on the sysfs_dirent (from the previous call to sysfs_get_active_two back up the call stack somewhere) so waits for the active reference to go away, which is of course impossible.

The problem has been present since 2.6.21.

This patch breaks the deadlock by queuing a work queue item on a single- threaded workqueue to do the removal from sysfs. There is also some protection against performing store operations on the "power" file of a slot that has already been disabled.

---

@@ -39,6 +39,7 @@

#include <linux/init.h>
#include <linux/string.h>
#include <linux/slab.h>
+#include <linux/workqueue.h>
#include "../pci.h"

#if !defined(MODULE)

@@ -63,10 +64,13 @@ struct dummy_slot {
	struct list_head node;
	struct hotplug_slot *slot;
	struct pci_dev *dev;

+ struct work_struct remove_work;
+ unsigned long removed;
};

static int debug;
static LIST_HEAD(slot_list);
+static struct workqueue_struct *dummyphp_wq;

static int enable_slot (struct hotplug_slot *slot); static int disable_slot (struct hotplug_slot *slot); @@ -109,7 +113,7 @@ static int add_slot(struct pci_dev *dev)

	slot->name = &dev->dev.bus_id[0];
	dbg("slot->name = %s\n", slot->name);

- dslot = kmalloc(sizeof(struct dummy_slot), GFP_KERNEL);

+	dslot = kzalloc(sizeof(struct dummy_slot), GFP_KERNEL);
	if (!dslot)
		goto error_info;

@@ -164,6 +168,13 @@ static void remove_slot(struct dummy_slo

		err("Problem unregistering a slot %s\n", dslot->slot->name);
}

+static void remove_slot_worker(struct work_struct *work) +{
+ struct dummy_slot *dslot =
+ container_of(work, struct dummy_slot, remove_work);
+ remove_slot(dslot);
+}
+
/**
* pci_rescan_slot - Rescan slot
* @temp: Device template. Should be set: bus and devfn.
@@ -271,6 +282,20 @@ static inline void pci_rescan(void) { static int enable_slot(struct hotplug_slot *hotplug_slot)

{
	/* mis-use enable_slot for rescanning of the pci bus */

+ struct dummy_slot *dslot;
+
+ if (hotplug_slot) {
+ /* make sure slot is not scheduled for removal
+ * otherwise flush_workqueue() could deadlock */
+ dslot = hotplug_slot->private;
+ if (test_bit(0, &dslot->removed)) {
+ /* FIXME: -ENODEV is also returned when function
+ * succeeds. */
+ return -ENODEV;
+ }
+ }
+ /* first, complete any pending removals */

+	flush_workqueue(dummyphp_wq);
	pci_rescan();
	return -ENODEV;
}
@@ -306,6 +331,10 @@ static int disable_slot(struct hotplug_s
		err("Can't remove PCI devices with other PCI devices behind it yet.\n");
		return -ENODEV;
	}

+ if (test_and_set_bit(0, &dslot->removed)) {
+ dbg("Slot already scheduled for removal\n");
+ return -ENODEV;
+ }

	/* search for subfunctions and disable them first */
	if (!(dslot->dev->devfn & 7)) {
		for (func = 1; func < 8; func++) {
@@ -328,8 +357,9 @@ static int disable_slot(struct hotplug_s
	/* remove the device from the pci core */
	pci_remove_bus_device(dslot->dev);

- /* blow away this sysfs entry and other parts. */
- remove_slot(dslot);
+ /* queue work item to blow away this sysfs entry and other parts. */
+ INIT_WORK(&dslot->remove_work, remove_slot_worker);
+ queue_work(dummyphp_wq, &dslot->remove_work);

	return 0;
}
@@ -340,6 +370,7 @@ static void cleanup_slots (void)
	struct list_head *next;
	struct dummy_slot *dslot;

+	destroy_workqueue(dummyphp_wq);
	list_for_each_safe (tmp, next, &slot_list) {
		dslot = list_entry (tmp, struct dummy_slot, node);
		remove_slot(dslot);
@@ -351,6 +382,10 @@ static int __init dummyphp_init(void)
{
	info(DRIVER_DESC "\n");

+ dummyphp_wq = create_singlethread_workqueue(MY_NAME);
+ if (!dummyphp_wq)
+ return -ENOMEM;
+

	return pci_scan_buses();
}

---
Note that the return values from the enable_slot function needs some attention. The patch causes this to return -ENODEV if the slot has already been disabled, but that is also the normal return value of the function. The return values should be different. Is there any reason why the normal return value can't be 0?

-=( Ian Abbott @ MEV Ltd.    E-mail: <[email protected]>        )=-
-=( Tel: +44 (0)161 477 1898   FAX: +44 (0)161 718 3587         )=-

Because I'm using it as a boolean?

Okay, it looks like there might be a race condition between enable_slot() and disable_slot() if some other task calls disable_slot() while enable_slot() is between the test_bit() and flush_workqueue() calls. I can fix that by avoiding the call to flush_workqueue() in enable_slot() and allocating and queueing a work queue item to defer the call to pci_rescan(). And enable_slot() won't then need to check if the slot was marked as removed - it can just go ahead and allocate and queue a work item.

-=( Ian Abbott @ MEV Ltd.    E-mail: <[email protected]>        )=-
-=( Tel: +44 (0)161 477 1898   FAX: +44 (0)161 718 3587         )=-

Heh, an unsigned long as a boolean? Come on... :)

That sounds reasonable.

thanks,

greg k-h