drm/i915: Fix race on unreferencing the wrong mmio-flip-request
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 27 Apr 2015 12:41:16 +0000 (13:41 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 20 May 2015 09:25:46 +0000 (11:25 +0200)
As we perform the mmio-flip without any locking and then try to acquire
the struct_mutex prior to dereferencing the request, it is possible for
userspace to queue a new pageflip before the worker can finish clearing
the old state - and then it will clear the new flip request. The result
is that the new flip could be completed before the GPU has finished
rendering.

The bugs stems from removing the seqno checking in
commit 536f5b5e86b225dab94c7ff8061ae482b6077387
Author: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Date:   Thu Nov 6 11:03:40 2014 +0200

    drm/i915: Make mmio flip wait for seqno in the work function

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_drv.h

index 1b539057b2646bb99dce5791db575e5c1e020b73..6a66d6b7d33b4a8aef237ba5e4faa0fd3445c439 100644 (file)
@@ -2161,10 +2161,12 @@ i915_gem_request_get_ring(struct drm_i915_gem_request *req)
        return req ? req->ring : NULL;
 }
 
-static inline void
+static inline struct drm_i915_gem_request *
 i915_gem_request_reference(struct drm_i915_gem_request *req)
 {
-       kref_get(&req->ref);
+       if (req)
+               kref_get(&req->ref);
+       return req;
 }
 
 static inline void
index 3c8801cecd3ff7fd824501acbbf5fe2e13c8241d..f2ec1d6f9c6abea8b266d752bdf10e8a87aecddc 100644 (file)
@@ -10793,22 +10793,18 @@ static void intel_do_mmio_flip(struct intel_crtc *intel_crtc)
 
 static void intel_mmio_flip_work_func(struct work_struct *work)
 {
-       struct intel_crtc *crtc =
-               container_of(work, struct intel_crtc, mmio_flip.work);
-       struct intel_mmio_flip *mmio_flip;
+       struct intel_mmio_flip *mmio_flip =
+               container_of(work, struct intel_mmio_flip, work);
 
-       mmio_flip = &crtc->mmio_flip;
-       if (mmio_flip->req)
-               WARN_ON(__i915_wait_request(mmio_flip->req,
-                                           crtc->reset_counter,
-                                           false, NULL, NULL) != 0);
+       if (mmio_flip->rq)
+               WARN_ON(__i915_wait_request(mmio_flip->rq,
+                                           mmio_flip->crtc->reset_counter,
+                                           false, NULL, NULL));
 
-       intel_do_mmio_flip(crtc);
-       if (mmio_flip->req) {
-               mutex_lock(&crtc->base.dev->struct_mutex);
-               i915_gem_request_assign(&mmio_flip->req, NULL);
-               mutex_unlock(&crtc->base.dev->struct_mutex);
-       }
+       intel_do_mmio_flip(mmio_flip->crtc);
+
+       i915_gem_request_unreference__unlocked(mmio_flip->rq);
+       kfree(mmio_flip);
 }
 
 static int intel_queue_mmio_flip(struct drm_device *dev,
@@ -10818,12 +10814,17 @@ static int intel_queue_mmio_flip(struct drm_device *dev,
                                 struct intel_engine_cs *ring,
                                 uint32_t flags)
 {
-       struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+       struct intel_mmio_flip *mmio_flip;
+
+       mmio_flip = kmalloc(sizeof(*mmio_flip), GFP_KERNEL);
+       if (mmio_flip == NULL)
+               return -ENOMEM;
 
-       i915_gem_request_assign(&intel_crtc->mmio_flip.req,
-                               obj->last_write_req);
+       mmio_flip->rq = i915_gem_request_reference(obj->last_write_req);
+       mmio_flip->crtc = to_intel_crtc(crtc);
 
-       schedule_work(&intel_crtc->mmio_flip.work);
+       INIT_WORK(&mmio_flip->work, intel_mmio_flip_work_func);
+       schedule_work(&mmio_flip->work);
 
        return 0;
 }
@@ -13564,8 +13565,6 @@ static void intel_crtc_init(struct drm_device *dev, int pipe)
        dev_priv->plane_to_crtc_mapping[intel_crtc->plane] = &intel_crtc->base;
        dev_priv->pipe_to_crtc_mapping[intel_crtc->pipe] = &intel_crtc->base;
 
-       INIT_WORK(&intel_crtc->mmio_flip.work, intel_mmio_flip_work_func);
-
        drm_crtc_helper_add(&intel_crtc->base, &intel_helper_funcs);
 
        WARN_ON(drm_crtc_index(&intel_crtc->base) != intel_crtc->pipe);
index 29262228f2641ebd0f81a077fdc91833c0158cd4..47bc729043c5522640cb6d28438e7d9601dd43d8 100644 (file)
@@ -459,8 +459,9 @@ struct intel_pipe_wm {
 };
 
 struct intel_mmio_flip {
-       struct drm_i915_gem_request *req;
        struct work_struct work;
+       struct drm_i915_gem_request *rq;
+       struct intel_crtc *crtc;
 };
 
 struct skl_pipe_wm {
@@ -544,7 +545,6 @@ struct intel_crtc {
        } wm;
 
        int scanline_offset;
-       struct intel_mmio_flip mmio_flip;
 
        struct intel_crtc_atomic_commit atomic;