INFO: possible circular locking dependency detected [v2.6.37-rc1-152-g6c88078]

INFO: possible circular locking dependency detected [v2.6.37-rc1-152-g6c88078]

[Uwe Helm]

Hi,
I've been seeing these lately, Xorg version is 1.9.0 with latest git kernel.

=======================================================
[ INFO: possible circular locking dependency detected ]
2.6.37-rc1-uwe1+ #4
-------------------------------------------------------
Xorg/1401 is trying to acquire lock:
 (&mm->mmap_sem){++++++}, at: [<c01e4ddb>] might_fault+0x4b/0xa0

but task is already holding lock:
 (&dev->struct_mutex){+.+.+.}, at: [<f869c3ac>]
i915_mutex_lock_interruptible+0x3c/0x60 [i915]

which lock already depends on the new lock.


the existing dependency chain (in reverse order) is:

-> #1 (&dev->struct_mutex){+.+.+.}:
       [<c01738ec>] lock_acquire+0x7c/0x120
       [<c04d3560>] mutex_lock_nested+0x60/0x330
       [<f810c7aa>] drm_gem_mmap+0x3a/0x140 [drm]
       [<c01ec202>] mmap_region+0x2e2/0x430
       [<c01ec59f>] do_mmap_pgoff+0x24f/0x2f0
       [<c01ec6bb>] sys_mmap_pgoff+0x7b/0x110
       [<c0102f97>] sysenter_do_call+0x12/0x36

-> #0 (&mm->mmap_sem){++++++}:
       [<c0173549>] __lock_acquire+0x14a9/0x17d0
       [<c01738ec>] lock_acquire+0x7c/0x120
       [<c01e4e0c>] might_fault+0x7c/0xa0
       [<c032dfad>] _copy_from_user+0x3d/0x130
       [<f869cdec>] i915_gem_pwrite_ioctl+0x18c/0xb20 [i915]
       [<f810afd2>] drm_ioctl+0x1e2/0x440 [drm]
       [<c0216dcd>] do_vfs_ioctl+0x9d/0x6c0
       [<c0217457>] sys_ioctl+0x67/0x80
       [<c0102f97>] sysenter_do_call+0x12/0x36

other info that might help us debug this:

1 lock held by Xorg/1401:
 #0:  (&dev->struct_mutex){+.+.+.}, at: [<f869c3ac>]
i915_mutex_lock_interruptible+0x3c/0x60 [i915]

stack backtrace:
Pid: 1401, comm: Xorg Not tainted 2.6.37-rc1-uwe1+ #4
Call Trace:
 [<c04d12de>] ? printk+0x1d/0x1f
 [<c01714f2>] print_circular_bug+0xc2/0xd0
 [<c0173549>] __lock_acquire+0x14a9/0x17d0
 [<c04d82c9>] ? sub_preempt_count+0x9/0x50
 [<c01738ec>] lock_acquire+0x7c/0x120
 [<c01e4ddb>] ? might_fault+0x4b/0xa0
 [<c01e4e0c>] might_fault+0x7c/0xa0
 [<c01e4ddb>] ? might_fault+0x4b/0xa0
 [<c032dfad>] _copy_from_user+0x3d/0x130
 [<f869cdec>] i915_gem_pwrite_ioctl+0x18c/0xb20 [i915]
 [<c01e4ddb>] ? might_fault+0x4b/0xa0
 [<f810afd2>] drm_ioctl+0x1e2/0x440 [drm]
 [<f869cc60>] ? i915_gem_pwrite_ioctl+0x0/0xb20 [i915]
 [<c02f36d3>] ? smk_access+0xb3/0x1f0
 [<c02f3914>] ? smk_curacc+0x74/0xa0
 [<f810adf0>] ? drm_ioctl+0x0/0x440 [drm]
 [<c0216dcd>] do_vfs_ioctl+0x9d/0x6c0
 [<c0217457>] sys_ioctl+0x67/0x80
[<c0102f97>] sysenter_do_call+0x12/0x36

thanks,
Uwe

[PATCH] drm/i915: Avoid might_fault during pwrite whilst holding our mutex

[Chris Wilson]

... and so prevent a potential circular reference:

  [ INFO: possible circular locking dependency detected ]
  2.6.37-rc1-uwe1+ #4
  -------------------------------------------------------
  Xorg/1401 is trying to acquire lock:
   (&mm->mmap_sem){++++++}, at: [<c01e4ddb>] might_fault+0x4b/0xa0

  but task is already holding lock:
   (&dev->struct_mutex){+.+.+.}, at: [<f869c3ac>]
  i915_mutex_lock_interruptible+0x3c/0x60 [i915]

  which lock already depends on the new lock.

When the locking around the pwrite ioctl was simplified, I did not spot
that the phys path never took any locks and so we introduced this
potential circular reference.

Reported-by: Uwe Helm <uwe.helm@googlemail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem.c |   25 ++++++++++++++++---------
 1 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 984eb6e..eba9b16 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -4878,17 +4878,24 @@ i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
 		     struct drm_file *file_priv)
 {
 	struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
-	void *obj_addr;
-	int ret;
-	char __user *user_data;
+	void *vaddr = obj_priv->phys_obj->handle->vaddr + args->offset;
+	char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
 
-	user_data = (char __user *) (uintptr_t) args->data_ptr;
-	obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
+	DRM_DEBUG_DRIVER("vaddr %p, %lld\n", vaddr, args->size);
 
-	DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
-	ret = copy_from_user(obj_addr, user_data, args->size);
-	if (ret)
-		return -EFAULT;
+	if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
+		unsigned long unwritten;
+
+		/* The physical object once assigned is fixed for the lifetime
+		 * of the obj, so we can safely drop the lock and continue
+		 * to access vaddr.
+		 */
+		mutex_unlock(&dev->struct_mutex);
+		unwritten = copy_from_user(vaddr, user_data, args->size);
+		mutex_lock(&dev->struct_mutex);
+		if (unwritten)
+			return -EFAULT;
+	}
 
 	drm_agp_chipset_flush(dev);
 	return 0;
-- 
1.7.2.3

Re: [PATCH] drm/i915: Avoid might_fault during pwrite whilst holding our mutex

[Uwe Helm]

On Sun, Nov 7, 2010 at 5:34 PM, Chris Wilson <chris@chris-wilson.co.uk> wrote:
> ... and so prevent a potential circular reference:
>
>  [ INFO: possible circular locking dependency detected ]
>  2.6.37-rc1-uwe1+ #4
>  -------------------------------------------------------
>  Xorg/1401 is trying to acquire lock:
>   (&mm->mmap_sem){++++++}, at: [<c01e4ddb>] might_fault+0x4b/0xa0
>
>  but task is already holding lock:
>   (&dev->struct_mutex){+.+.+.}, at: [<f869c3ac>]
>  i915_mutex_lock_interruptible+0x3c/0x60 [i915]
>
>  which lock already depends on the new lock.
>
> When the locking around the pwrite ioctl was simplified, I did not spot
> that the phys path never took any locks and so we introduced this
> potential circular reference.
>
> Reported-by: Uwe Helm <uwe.helm@googlemail.com>
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_gem.c |   25 ++++++++++++++++---------
>  1 files changed, 16 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 984eb6e..eba9b16 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -4878,17 +4878,24 @@ i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj,
>                     struct drm_file *file_priv)
>  {
>        struct drm_i915_gem_object *obj_priv = to_intel_bo(obj);
> -       void *obj_addr;
> -       int ret;
> -       char __user *user_data;
> +       void *vaddr = obj_priv->phys_obj->handle->vaddr + args->offset;
> +       char __user *user_data = (char __user *) (uintptr_t) args->data_ptr;
>
> -       user_data = (char __user *) (uintptr_t) args->data_ptr;
> -       obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset;
> +       DRM_DEBUG_DRIVER("vaddr %p, %lld\n", vaddr, args->size);
>
> -       DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size);
> -       ret = copy_from_user(obj_addr, user_data, args->size);
> -       if (ret)
> -               return -EFAULT;
> +       if (__copy_from_user_inatomic_nocache(vaddr, user_data, args->size)) {
> +               unsigned long unwritten;
> +
> +               /* The physical object once assigned is fixed for the lifetime
> +                * of the obj, so we can safely drop the lock and continue
> +                * to access vaddr.
> +                */
> +               mutex_unlock(&dev->struct_mutex);
> +               unwritten = copy_from_user(vaddr, user_data, args->size);
> +               mutex_lock(&dev->struct_mutex);
> +               if (unwritten)
> +                       return -EFAULT;
> +       }
>
>        drm_agp_chipset_flush(dev);
>        return 0;
> --
> 1.7.2.3
>
>

works, thank you.