]> git.ipfire.org Git - people/arne_f/kernel.git/commitdiff
vfio: Fix WARNING "do not call blocking ops when !TASK_RUNNING"
authorFarhan Ali <alifm@linux.ibm.com>
Wed, 3 Apr 2019 18:22:27 +0000 (14:22 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 15 Jun 2019 09:54:07 +0000 (11:54 +0200)
[ Upstream commit 41be3e2618174fdf3361e49e64f2bf530f40c6b0 ]

vfio_dev_present() which is the condition to
wait_event_interruptible_timeout(), will call vfio_group_get_device
and try to acquire the mutex group->device_lock.

wait_event_interruptible_timeout() will set the state of the current
task to TASK_INTERRUPTIBLE, before doing the condition check. This
means that we will try to acquire the mutex while already in a
sleeping state. The scheduler warns us by giving the following
warning:

[ 4050.264464] ------------[ cut here ]------------
[ 4050.264508] do not call blocking ops when !TASK_RUNNING; state=1 set at [<00000000b33c00e2>] prepare_to_wait_event+0x14a/0x188
[ 4050.264529] WARNING: CPU: 12 PID: 35924 at kernel/sched/core.c:6112 __might_sleep+0x76/0x90
....

 4050.264756] Call Trace:
[ 4050.264765] ([<000000000017bbaa>] __might_sleep+0x72/0x90)
[ 4050.264774]  [<0000000000b97edc>] __mutex_lock+0x44/0x8c0
[ 4050.264782]  [<0000000000b9878a>] mutex_lock_nested+0x32/0x40
[ 4050.264793]  [<000003ff800d7abe>] vfio_group_get_device+0x36/0xa8 [vfio]
[ 4050.264803]  [<000003ff800d87c0>] vfio_del_group_dev+0x238/0x378 [vfio]
[ 4050.264813]  [<000003ff8015f67c>] mdev_remove+0x3c/0x68 [mdev]
[ 4050.264825]  [<00000000008e01b0>] device_release_driver_internal+0x168/0x268
[ 4050.264834]  [<00000000008de692>] bus_remove_device+0x162/0x190
[ 4050.264843]  [<00000000008daf42>] device_del+0x1e2/0x368
[ 4050.264851]  [<00000000008db12c>] device_unregister+0x64/0x88
[ 4050.264862]  [<000003ff8015ed84>] mdev_device_remove+0xec/0x130 [mdev]
[ 4050.264872]  [<000003ff8015f074>] remove_store+0x6c/0xa8 [mdev]
[ 4050.264881]  [<000000000046f494>] kernfs_fop_write+0x14c/0x1f8
[ 4050.264890]  [<00000000003c1530>] __vfs_write+0x38/0x1a8
[ 4050.264899]  [<00000000003c187c>] vfs_write+0xb4/0x198
[ 4050.264908]  [<00000000003c1af2>] ksys_write+0x5a/0xb0
[ 4050.264916]  [<0000000000b9e270>] system_call+0xdc/0x2d8
[ 4050.264925] 4 locks held by sh/35924:
[ 4050.264933]  #0: 000000001ef90325 (sb_writers#4){.+.+}, at: vfs_write+0x9e/0x198
[ 4050.264948]  #1: 000000005c1ab0b3 (&of->mutex){+.+.}, at: kernfs_fop_write+0x1cc/0x1f8
[ 4050.264963]  #2: 0000000034831ab8 (kn->count#297){++++}, at: kernfs_remove_self+0x12e/0x150
[ 4050.264979]  #3: 00000000e152484f (&dev->mutex){....}, at: device_release_driver_internal+0x5c/0x268
[ 4050.264993] Last Breaking-Event-Address:
[ 4050.265002]  [<000000000017bbaa>] __might_sleep+0x72/0x90
[ 4050.265010] irq event stamp: 7039
[ 4050.265020] hardirqs last  enabled at (7047): [<00000000001cee7a>] console_unlock+0x6d2/0x740
[ 4050.265029] hardirqs last disabled at (7054): [<00000000001ce87e>] console_unlock+0xd6/0x740
[ 4050.265040] softirqs last  enabled at (6416): [<0000000000b8fe26>] __udelay+0xb6/0x100
[ 4050.265049] softirqs last disabled at (6415): [<0000000000b8fe06>] __udelay+0x96/0x100
[ 4050.265057] ---[ end trace d04a07d39d99a9f9 ]---

Let's fix this as described in the article
https://lwn.net/Articles/628628/.

Signed-off-by: Farhan Ali <alifm@linux.ibm.com>
[remove now redundant vfio_dev_present()]
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/vfio/vfio.c

index 64833879f75d3cf2d0739f03e4ca4b587c639cca..7a386fb30bf1fc448e4515946bf4aaf2fbb25797 100644 (file)
@@ -34,6 +34,7 @@
 #include <linux/uaccess.h>
 #include <linux/vfio.h>
 #include <linux/wait.h>
+#include <linux/sched/signal.h>
 
 #define DRIVER_VERSION "0.3"
 #define DRIVER_AUTHOR  "Alex Williamson <alex.williamson@redhat.com>"
@@ -904,30 +905,17 @@ void *vfio_device_data(struct vfio_device *device)
 }
 EXPORT_SYMBOL_GPL(vfio_device_data);
 
-/* Given a referenced group, check if it contains the device */
-static bool vfio_dev_present(struct vfio_group *group, struct device *dev)
-{
-       struct vfio_device *device;
-
-       device = vfio_group_get_device(group, dev);
-       if (!device)
-               return false;
-
-       vfio_device_put(device);
-       return true;
-}
-
 /*
  * Decrement the device reference count and wait for the device to be
  * removed.  Open file descriptors for the device... */
 void *vfio_del_group_dev(struct device *dev)
 {
+       DEFINE_WAIT_FUNC(wait, woken_wake_function);
        struct vfio_device *device = dev_get_drvdata(dev);
        struct vfio_group *group = device->group;
        void *device_data = device->device_data;
        struct vfio_unbound_dev *unbound;
        unsigned int i = 0;
-       long ret;
        bool interrupted = false;
 
        /*
@@ -964,6 +952,8 @@ void *vfio_del_group_dev(struct device *dev)
         * interval with counter to allow the driver to take escalating
         * measures to release the device if it has the ability to do so.
         */
+       add_wait_queue(&vfio.release_q, &wait);
+
        do {
                device = vfio_group_get_device(group, dev);
                if (!device)
@@ -975,12 +965,10 @@ void *vfio_del_group_dev(struct device *dev)
                vfio_device_put(device);
 
                if (interrupted) {
-                       ret = wait_event_timeout(vfio.release_q,
-                                       !vfio_dev_present(group, dev), HZ * 10);
+                       wait_woken(&wait, TASK_UNINTERRUPTIBLE, HZ * 10);
                } else {
-                       ret = wait_event_interruptible_timeout(vfio.release_q,
-                                       !vfio_dev_present(group, dev), HZ * 10);
-                       if (ret == -ERESTARTSYS) {
+                       wait_woken(&wait, TASK_INTERRUPTIBLE, HZ * 10);
+                       if (signal_pending(current)) {
                                interrupted = true;
                                dev_warn(dev,
                                         "Device is currently in use, task"
@@ -989,8 +977,10 @@ void *vfio_del_group_dev(struct device *dev)
                                         current->comm, task_pid_nr(current));
                        }
                }
-       } while (ret <= 0);
 
+       } while (1);
+
+       remove_wait_queue(&vfio.release_q, &wait);
        /*
         * In order to support multiple devices per group, devices can be
         * plucked from the group while other devices in the group are still