drm/i915: Use fence_write() from rpm resume
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 12 Oct 2016 11:48:26 +0000 (12:48 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 12 Oct 2016 15:27:14 +0000 (16:27 +0100)
During rpm resume we restore the fences, but we do not have the
protection of struct_mutex. This rules out updating the activity
tracking on the fences, and requires us to rely on the rpm as the
serialisation barrier instead.

[  350.298052] [drm:intel_runtime_resume [i915]] Resuming device
[  350.308606]
[  350.310520] ===============================
[  350.315560] [ INFO: suspicious RCU usage. ]
[  350.320554] 4.8.0-rc8-bsw-rapl+ #3133 Tainted: G     U  W
[  350.327208] -------------------------------
[  350.331977] ../drivers/gpu/drm/i915/i915_gem_request.h:371 suspicious rcu_dereference_protected() usage!
[  350.342619]
[  350.342619] other info that might help us debug this:
[  350.342619]
[  350.351593]
[  350.351593] rcu_scheduler_active = 1, debug_locks = 0
[  350.358952] 3 locks held by Xorg/320:
[  350.363077]  #0:  (&dev->mode_config.mutex){+.+.+.}, at: [<ffffffffa030589c>] drm_modeset_lock_all+0x3c/0xd0 [drm]
[  350.375162]  #1:  (crtc_ww_class_acquire){+.+.+.}, at: [<ffffffffa03058a6>] drm_modeset_lock_all+0x46/0xd0 [drm]
[  350.387022]  #2:  (crtc_ww_class_mutex){+.+.+.}, at: [<ffffffffa0305056>] drm_modeset_lock+0x36/0x110 [drm]
[  350.398236]
[  350.398236] stack backtrace:
[  350.403196] CPU: 1 PID: 320 Comm: Xorg Tainted: G     U  W       4.8.0-rc8-bsw-rapl+ #3133
[  350.412457] Hardware name: Intel Corporation CHERRYVIEW C0 PLATFORM/Braswell CRB, BIOS BRAS.X64.X088.R00.1510270350 10/27/2015
[  350.425212]  0000000000000000 ffff8801680a78c8 ffffffff81332187 ffff88016c5c5000
[  350.433611]  0000000000000001 ffff8801680a78f8 ffffffff810ca6da ffff88016cc8b0f0
[  350.442012]  ffff88016cc80000 ffff88016cc80000 ffff880177ad0000 ffff8801680a7948
[  350.450409] Call Trace:
[  350.453165]  [<ffffffff81332187>] dump_stack+0x67/0x90
[  350.458931]  [<ffffffff810ca6da>] lockdep_rcu_suspicious+0xea/0x120
[  350.466002]  [<ffffffffa039e8dd>] fence_update+0xbd/0x670 [i915]
[  350.472766]  [<ffffffffa039efe2>] i915_gem_restore_fences+0x52/0x70 [i915]
[  350.480496]  [<ffffffffa0368f42>] vlv_resume_prepare+0x72/0x570 [i915]
[  350.487839]  [<ffffffffa0369802>] intel_runtime_resume+0x102/0x210 [i915]
[  350.495442]  [<ffffffff8137f26f>] pci_pm_runtime_resume+0x7f/0xb0
[  350.502274]  [<ffffffff8137f1f0>] ? pci_restore_standard_config+0x40/0x40
[  350.509883]  [<ffffffff814401c5>] __rpm_callback+0x35/0x70
[  350.516037]  [<ffffffff8137f1f0>] ? pci_restore_standard_config+0x40/0x40
[  350.523646]  [<ffffffff81440224>] rpm_callback+0x24/0x80
[  350.529604]  [<ffffffff8137f1f0>] ? pci_restore_standard_config+0x40/0x40
[  350.537212]  [<ffffffff814417bd>] rpm_resume+0x4ad/0x740
[  350.543161]  [<ffffffff81441aa1>] __pm_runtime_resume+0x51/0x80
[  350.549824]  [<ffffffffa03889c8>] intel_runtime_pm_get+0x28/0x90 [i915]
[  350.557265]  [<ffffffffa0388a53>] intel_display_power_get+0x23/0x50 [i915]
[  350.565001]  [<ffffffffa03ef23d>] intel_atomic_commit_tail+0xdfd/0x10b0 [i915]
[  350.573106]  [<ffffffffa034b2e9>] ? drm_atomic_helper_swap_state+0x159/0x300 [drm_kms_helper]
[  350.582659]  [<ffffffff81615091>] ? _raw_spin_unlock+0x31/0x50
[  350.589205]  [<ffffffffa034b2e9>] ? drm_atomic_helper_swap_state+0x159/0x300 [drm_kms_helper]
[  350.598787]  [<ffffffffa03ef8a5>] intel_atomic_commit+0x3b5/0x500 [i915]
[  350.606319]  [<ffffffffa03061dc>] ? drm_atomic_set_crtc_for_connector+0xcc/0x100 [drm]
[  350.615209]  [<ffffffffa0306b49>] drm_atomic_commit+0x49/0x50 [drm]
[  350.622242]  [<ffffffffa034dee8>] drm_atomic_helper_set_config+0x88/0xc0 [drm_kms_helper]
[  350.631419]  [<ffffffffa02f94ac>] drm_mode_set_config_internal+0x6c/0x120 [drm]
[  350.639623]  [<ffffffffa02fa94c>] drm_mode_setcrtc+0x22c/0x4d0 [drm]
[  350.646760]  [<ffffffffa02f0f19>] drm_ioctl+0x209/0x460 [drm]
[  350.653217]  [<ffffffffa02fa720>] ? drm_mode_getcrtc+0x150/0x150 [drm]
[  350.660536]  [<ffffffff810c984a>] ? __lock_is_held+0x4a/0x70
[  350.666885]  [<ffffffff81202303>] do_vfs_ioctl+0x93/0x6b0
[  350.672939]  [<ffffffff8120f843>] ? __fget+0x113/0x200
[  350.678797]  [<ffffffff8120f735>] ? __fget+0x5/0x200
[  350.684361]  [<ffffffff81202964>] SyS_ioctl+0x44/0x80
[  350.690030]  [<ffffffff81001deb>] do_syscall_64+0x5b/0x120
[  350.696184]  [<ffffffff81615ada>] entry_SYSCALL64_slow_path+0x25/0x25

Note we also have to remember the lesson from commit 4fc788f5ee3d
("drm/i915: Flush delayed fence releases after reset") where we have to
flush any changes to the fence on restore.

v2: Replace call to release user mmaps with an assertion that they have
already been zapped.

Fixes: 49ef5294cda2 ("drm/i915: Move fence tracking from object to vma")
Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161012114827.17031-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_fence.c

index 8df1fa7234e8e031e9c61da4608c5c4b7610476b..2c7ba0ee127c6a230eff6bf7deff705a1259364b 100644 (file)
@@ -290,6 +290,8 @@ i915_vma_put_fence(struct i915_vma *vma)
 {
        struct drm_i915_fence_reg *fence = vma->fence;
 
+       assert_rpm_wakelock_held(to_i915(vma->vm->dev));
+
        if (!fence)
                return 0;
 
@@ -341,6 +343,8 @@ i915_vma_get_fence(struct i915_vma *vma)
        struct drm_i915_fence_reg *fence;
        struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
 
+       assert_rpm_wakelock_held(to_i915(vma->vm->dev));
+
        /* Just update our place in the LRU if our fence is getting reused. */
        if (vma->fence) {
                fence = vma->fence;
@@ -371,6 +375,12 @@ void i915_gem_restore_fences(struct drm_device *dev)
        struct drm_i915_private *dev_priv = to_i915(dev);
        int i;
 
+       /* Note that this may be called outside of struct_mutex, by
+        * runtime suspend/resume. The barrier we require is enforced by
+        * rpm itself - all access to fences/GTT are only within an rpm
+        * wakeref, and to acquire that wakeref you must pass through here.
+        */
+
        for (i = 0; i < dev_priv->num_fence_regs; i++) {
                struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
                struct i915_vma *vma = reg->vma;
@@ -379,10 +389,17 @@ void i915_gem_restore_fences(struct drm_device *dev)
                 * Commit delayed tiling changes if we have an object still
                 * attached to the fence, otherwise just clear the fence.
                 */
-               if (vma && !i915_gem_object_is_tiled(vma->obj))
+               if (vma && !i915_gem_object_is_tiled(vma->obj)) {
+                       GEM_BUG_ON(!reg->dirty);
+                       GEM_BUG_ON(vma->obj->fault_mappable);
+
+                       list_move(&reg->link, &dev_priv->mm.fence_list);
+                       vma->fence = NULL;
                        vma = NULL;
+               }
 
-               fence_update(reg, vma);
+               fence_write(reg, vma);
+               reg->vma = vma;
        }
 }