[PATCH v2 2/7] libnvdimm/bus: Prevent duplicate device_unregister() calls

[Dan Williams]

A multithreaded namespace creation/destruction stress test currently
fails with signatures like the following:

    sysfs group 'power' not found for kobject 'dax1.1'
    RIP: 0010:sysfs_remove_group+0x76/0x80
    Call Trace:
     device_del+0x73/0x370
     device_unregister+0x16/0x50
     nd_async_device_unregister+0x1e/0x30 [libnvdimm]
     async_run_entry_fn+0x39/0x160
     process_one_work+0x23c/0x5e0
     worker_thread+0x3c/0x390

    BUG: kernel NULL pointer dereference, address: 0000000000000020
    RIP: 0010:klist_put+0x1b/0x6c
    Call Trace:
     klist_del+0xe/0x10
     device_del+0x8a/0x2c9
     ? __switch_to_asm+0x34/0x70
     ? __switch_to_asm+0x40/0x70
     device_unregister+0x44/0x4f
     nd_async_device_unregister+0x22/0x2d [libnvdimm]
     async_run_entry_fn+0x47/0x15a
     process_one_work+0x1a2/0x2eb
     worker_thread+0x1b8/0x26e

Use the kill_device() helper to atomically resolve the race of multiple
threads issuing kill, device_unregister(), requests.

Reported-by: Jane Chu <jane.chu@xxxxxxxxxx>
Reported-by: Erwin Tsaur <erwin.tsaur@xxxxxxxxxx>
Fixes: 4d88a97aa9e8 ("libnvdimm, nvdimm: dimm driver and base libnvdimm device-driver...")
Cc: <stable@xxxxxxxxxxxxxxx>
Link: https://github.com/pmem/ndctl/issues/96
Tested-by: Tested-by: Jane Chu <jane.chu@xxxxxxxxxx>
Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
---
 drivers/nvdimm/bus.c |   25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/drivers/nvdimm/bus.c b/drivers/nvdimm/bus.c
index 2dca3034fee0..42713b210f51 100644
--- a/drivers/nvdimm/bus.c
+++ b/drivers/nvdimm/bus.c
@@ -547,13 +547,38 @@ EXPORT_SYMBOL(nd_device_register);
 
 void nd_device_unregister(struct device *dev, enum nd_async_mode mode)
 {
+	bool killed;
+
 	switch (mode) {
 	case ND_ASYNC:
+		/*
+		 * In the async case this is being triggered with the
+		 * device lock held and the unregistration work needs to
+		 * be moved out of line iff this is thread has won the
+		 * race to schedule the deletion.
+		 */
+		if (!kill_device(dev))
+			return;
+
 		get_device(dev);
 		async_schedule_domain(nd_async_device_unregister, dev,
 				&nd_async_domain);
 		break;
 	case ND_SYNC:
+		/*
+		 * In the sync case the device is being unregistered due
+		 * to a state change of the parent. Claim the kill state
+		 * to synchronize against other unregistration requests,
+		 * or otherwise let the async path handle it if the
+		 * unregistration was already queued.
+		 */
+		device_lock(dev);
+		killed = kill_device(dev);
+		device_unlock(dev);
+
+		if (!killed)
+			return;
+
 		nd_synchronize();
 		device_unregister(dev);
 		break;