]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
drm/i915: Rearrange i915_wait_request() accounting with callers
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 12:58:27 +0000 (13:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 19:53:43 +0000 (20:53 +0100)
Our low-level wait routine has evolved from our generic wait interface
that handled unlocked, RPS boosting, waits with time tracking. If we
push our GEM fence tracking to use reservation_objects (required for
handling multiple timelines), we lose the ability to pass the required
information down to i915_wait_request(). However, if we push the extra
functionality from i915_wait_request() to the individual callsites
(i915_gem_object_wait_rendering and i915_gem_wait_ioctl) that make use
of those extras, we can both simplify our low level wait and prepare for
extending the GEM interface for use of reservation_objects.

v2: Rewrite i915_wait_request() kerneldocs

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.william.auld@gmail.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-4-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gvt/scheduler.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_request.c
drivers/gpu/drm/i915/i915_gem_request.h
drivers/gpu/drm/i915/i915_gem_userptr.c
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_ringbuffer.c
drivers/gpu/drm/i915/intel_ringbuffer.h

index f7e320b9b17a3178c0d93040aac186a720427635..18acb45dd14dedc99f0c663561697ac94e4d3d6f 100644 (file)
@@ -400,6 +400,7 @@ static int workload_thread(void *priv)
        int ring_id = p->ring_id;
        struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
        struct intel_vgpu_workload *workload = NULL;
+       long lret;
        int ret;
        bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
        DEFINE_WAIT_FUNC(wait, woken_wake_function);
@@ -449,10 +450,12 @@ static int workload_thread(void *priv)
                gvt_dbg_sched("ring id %d wait workload %p\n",
                                workload->ring_id, workload);
 
-               workload->status = i915_wait_request(workload->req,
-                                                    0, NULL, NULL);
-               if (workload->status != 0)
+               lret = i915_wait_request(workload->req,
+                                        0, MAX_SCHEDULE_TIMEOUT);
+               if (lret < 0) {
+                       workload->status = lret;
                        gvt_err("fail to wait workload, skip\n");
+               }
 
 complete:
                gvt_dbg_sched("will complete workload %p\n, status: %d\n",
index e95352cc5ac236f43add6d4069413810302de64c..cf4b2427aff31288ee692dc2d8336cd1ff1d6590 100644 (file)
@@ -3319,9 +3319,10 @@ int __must_check i915_gem_wait_for_idle(struct drm_i915_private *dev_priv,
 int __must_check i915_gem_suspend(struct drm_device *dev);
 void i915_gem_resume(struct drm_device *dev);
 int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf);
-int __must_check
-i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
-                              bool readonly);
+int i915_gem_object_wait(struct drm_i915_gem_object *obj,
+                        unsigned int flags,
+                        long timeout,
+                        struct intel_rps_client *rps);
 int __must_check
 i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj,
                                  bool write);
index 1254143ab121ed3b5db977b2363965e3146a0ac1..537f502123ea2c3dd7d54999a27a73bdc0f360bf 100644 (file)
@@ -292,7 +292,12 @@ int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
         * must wait for all rendering to complete to the object (as unbinding
         * must anyway), and retire the requests.
         */
-       ret = i915_gem_object_wait_rendering(obj, false);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED |
+                                  I915_WAIT_ALL,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                return ret;
 
@@ -311,88 +316,172 @@ int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
        return ret;
 }
 
-/**
- * Ensures that all rendering to the object has completed and the object is
- * safe to unbind from the GTT or access from the CPU.
- * @obj: i915 gem object
- * @readonly: waiting for just read access or read-write access
- */
-int
-i915_gem_object_wait_rendering(struct drm_i915_gem_object *obj,
-                              bool readonly)
+static long
+i915_gem_object_wait_fence(struct dma_fence *fence,
+                          unsigned int flags,
+                          long timeout,
+                          struct intel_rps_client *rps)
 {
-       struct reservation_object *resv;
-       struct i915_gem_active *active;
-       unsigned long active_mask;
-       int idx;
+       struct drm_i915_gem_request *rq;
 
-       lockdep_assert_held(&obj->base.dev->struct_mutex);
+       BUILD_BUG_ON(I915_WAIT_INTERRUPTIBLE != 0x1);
 
-       if (!readonly) {
-               active = obj->last_read;
-               active_mask = i915_gem_object_get_active(obj);
-       } else {
-               active_mask = 1;
-               active = &obj->last_write;
+       if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
+               return timeout;
+
+       if (!dma_fence_is_i915(fence))
+               return dma_fence_wait_timeout(fence,
+                                             flags & I915_WAIT_INTERRUPTIBLE,
+                                             timeout);
+
+       rq = to_request(fence);
+       if (i915_gem_request_completed(rq))
+               goto out;
+
+       /* This client is about to stall waiting for the GPU. In many cases
+        * this is undesirable and limits the throughput of the system, as
+        * many clients cannot continue processing user input/output whilst
+        * blocked. RPS autotuning may take tens of milliseconds to respond
+        * to the GPU load and thus incurs additional latency for the client.
+        * We can circumvent that by promoting the GPU frequency to maximum
+        * before we wait. This makes the GPU throttle up much more quickly
+        * (good for benchmarks and user experience, e.g. window animations),
+        * but at a cost of spending more power processing the workload
+        * (bad for battery). Not all clients even want their results
+        * immediately and for them we should just let the GPU select its own
+        * frequency to maximise efficiency. To prevent a single client from
+        * forcing the clocks too high for the whole system, we only allow
+        * each client to waitboost once in a busy period.
+        */
+       if (rps) {
+               if (INTEL_GEN(rq->i915) >= 6)
+                       gen6_rps_boost(rq->i915, rps, rq->emitted_jiffies);
+               else
+                       rps = NULL;
        }
 
-       for_each_active(active_mask, idx) {
+       timeout = i915_wait_request(rq, flags, timeout);
+
+out:
+       if (flags & I915_WAIT_LOCKED && i915_gem_request_completed(rq))
+               i915_gem_request_retire_upto(rq);
+
+       if (rps && rq->fence.seqno == rq->engine->last_submitted_seqno) {
+               /* The GPU is now idle and this client has stalled.
+                * Since no other client has submitted a request in the
+                * meantime, assume that this client is the only one
+                * supplying work to the GPU but is unable to keep that
+                * work supplied because it is waiting. Since the GPU is
+                * then never kept fully busy, RPS autoclocking will
+                * keep the clocks relatively low, causing further delays.
+                * Compensate by giving the synchronous client credit for
+                * a waitboost next time.
+                */
+               spin_lock(&rq->i915->rps.client_lock);
+               list_del_init(&rps->link);
+               spin_unlock(&rq->i915->rps.client_lock);
+       }
+
+       return timeout;
+}
+
+static long
+i915_gem_object_wait_reservation(struct reservation_object *resv,
+                                unsigned int flags,
+                                long timeout,
+                                struct intel_rps_client *rps)
+{
+       struct dma_fence *excl;
+
+       if (flags & I915_WAIT_ALL) {
+               struct dma_fence **shared;
+               unsigned int count, i;
                int ret;
 
-               ret = i915_gem_active_wait(&active[idx],
-                                          &obj->base.dev->struct_mutex);
+               ret = reservation_object_get_fences_rcu(resv,
+                                                       &excl, &count, &shared);
                if (ret)
                        return ret;
-       }
 
-       resv = i915_gem_object_get_dmabuf_resv(obj);
-       if (resv) {
-               long err;
+               for (i = 0; i < count; i++) {
+                       timeout = i915_gem_object_wait_fence(shared[i],
+                                                            flags, timeout,
+                                                            rps);
+                       if (timeout <= 0)
+                               break;
+
+                       dma_fence_put(shared[i]);
+               }
 
-               err = reservation_object_wait_timeout_rcu(resv, !readonly, true,
-                                                         MAX_SCHEDULE_TIMEOUT);
-               if (err < 0)
-                       return err;
+               for (; i < count; i++)
+                       dma_fence_put(shared[i]);
+               kfree(shared);
+       } else {
+               excl = reservation_object_get_excl_rcu(resv);
        }
 
-       return 0;
+       if (excl && timeout > 0)
+               timeout = i915_gem_object_wait_fence(excl, flags, timeout, rps);
+
+       dma_fence_put(excl);
+
+       return timeout;
 }
 
-/* A nonblocking variant of the above wait. Must be called prior to
- * acquiring the mutex for the object, as the object state may change
- * during this call. A reference must be held by the caller for the object.
+/**
+ * Waits for rendering to the object to be completed
+ * @obj: i915 gem object
+ * @flags: how to wait (under a lock, for all rendering or just for writes etc)
+ * @timeout: how long to wait
+ * @rps: client (user process) to charge for any waitboosting
  */
-static __must_check int
-__unsafe_wait_rendering(struct drm_i915_gem_object *obj,
-                       struct intel_rps_client *rps,
-                       bool readonly)
+int
+i915_gem_object_wait(struct drm_i915_gem_object *obj,
+                    unsigned int flags,
+                    long timeout,
+                    struct intel_rps_client *rps)
 {
+       struct reservation_object *resv;
        struct i915_gem_active *active;
        unsigned long active_mask;
        int idx;
 
-       active_mask = __I915_BO_ACTIVE(obj);
-       if (!active_mask)
-               return 0;
+       might_sleep();
+#if IS_ENABLED(CONFIG_LOCKDEP)
+       GEM_BUG_ON(debug_locks &&
+                  !!lockdep_is_held(&obj->base.dev->struct_mutex) !=
+                  !!(flags & I915_WAIT_LOCKED));
+#endif
+       GEM_BUG_ON(timeout < 0);
 
-       if (!readonly) {
+       if (flags & I915_WAIT_ALL) {
                active = obj->last_read;
+               active_mask = i915_gem_object_get_active(obj);
        } else {
                active_mask = 1;
                active = &obj->last_write;
        }
 
        for_each_active(active_mask, idx) {
-               int ret;
-
-               ret = i915_gem_active_wait_unlocked(&active[idx],
-                                                   I915_WAIT_INTERRUPTIBLE,
-                                                   NULL, rps);
-               if (ret)
-                       return ret;
+               struct drm_i915_gem_request *request;
+
+               request = i915_gem_active_get_unlocked(&active[idx]);
+               if (request) {
+                       timeout = i915_gem_object_wait_fence(&request->fence,
+                                                            flags, timeout,
+                                                            rps);
+                       i915_gem_request_put(request);
+               }
+               if (timeout < 0)
+                       return timeout;
        }
 
-       return 0;
+       resv = i915_gem_object_get_dmabuf_resv(obj);
+       if (resv)
+               timeout = i915_gem_object_wait_reservation(resv,
+                                                          flags, timeout,
+                                                          rps);
+       return timeout < 0 ? timeout : 0;
 }
 
 static struct intel_rps_client *to_rps_client(struct drm_file *file)
@@ -449,12 +538,18 @@ i915_gem_phys_pwrite(struct drm_i915_gem_object *obj,
        struct drm_device *dev = obj->base.dev;
        void *vaddr = obj->phys_handle->vaddr + args->offset;
        char __user *user_data = u64_to_user_ptr(args->data_ptr);
-       int ret = 0;
+       int ret;
 
        /* We manually control the domain here and pretend that it
         * remains coherent i.e. in the GTT domain, like shmem_pwrite.
         */
-       ret = i915_gem_object_wait_rendering(obj, false);
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED |
+                                  I915_WAIT_ALL,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  to_rps_client(file_priv));
        if (ret)
                return ret;
 
@@ -614,12 +709,17 @@ int i915_gem_obj_prepare_shmem_read(struct drm_i915_gem_object *obj,
 {
        int ret;
 
-       *needs_clflush = 0;
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
 
+       *needs_clflush = 0;
        if (!i915_gem_object_has_struct_page(obj))
                return -ENODEV;
 
-       ret = i915_gem_object_wait_rendering(obj, true);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                return ret;
 
@@ -661,11 +761,18 @@ int i915_gem_obj_prepare_shmem_write(struct drm_i915_gem_object *obj,
 {
        int ret;
 
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
+
        *needs_clflush = 0;
        if (!i915_gem_object_has_struct_page(obj))
                return -ENODEV;
 
-       ret = i915_gem_object_wait_rendering(obj, false);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED |
+                                  I915_WAIT_ALL,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                return ret;
 
@@ -1051,7 +1158,10 @@ i915_gem_pread_ioctl(struct drm_device *dev, void *data,
 
        trace_i915_gem_object_pread(obj, args->offset, args->size);
 
-       ret = __unsafe_wait_rendering(obj, to_rps_client(file), true);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  to_rps_client(file));
        if (ret)
                goto err;
 
@@ -1449,7 +1559,11 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data,
 
        trace_i915_gem_object_pwrite(obj, args->offset, args->size);
 
-       ret = __unsafe_wait_rendering(obj, to_rps_client(file), false);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_ALL,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  to_rps_client(file));
        if (ret)
                goto err;
 
@@ -1536,7 +1650,11 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
         * We will repeat the flush holding the lock in the normal manner
         * to catch cases where we are gazumped.
         */
-       ret = __unsafe_wait_rendering(obj, to_rps_client(file), !write_domain);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  (write_domain ? I915_WAIT_ALL : 0),
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  to_rps_client(file));
        if (ret)
                goto err;
 
@@ -1772,7 +1890,10 @@ int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
         * repeat the flush holding the lock in the normal manner to catch cases
         * where we are gazumped.
         */
-       ret = __unsafe_wait_rendering(obj, NULL, !write);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE,
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                goto err;
 
@@ -2817,6 +2938,17 @@ void i915_gem_close_object(struct drm_gem_object *gem, struct drm_file *file)
        mutex_unlock(&obj->base.dev->struct_mutex);
 }
 
+static unsigned long to_wait_timeout(s64 timeout_ns)
+{
+       if (timeout_ns < 0)
+               return MAX_SCHEDULE_TIMEOUT;
+
+       if (timeout_ns == 0)
+               return 0;
+
+       return nsecs_to_jiffies_timeout(timeout_ns);
+}
+
 /**
  * i915_gem_wait_ioctl - implements DRM_IOCTL_I915_GEM_WAIT
  * @dev: drm device pointer
@@ -2845,10 +2977,9 @@ int
 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 {
        struct drm_i915_gem_wait *args = data;
-       struct intel_rps_client *rps = to_rps_client(file);
        struct drm_i915_gem_object *obj;
-       unsigned long active;
-       int idx, ret = 0;
+       ktime_t start;
+       long ret;
 
        if (args->flags != 0)
                return -EINVAL;
@@ -2857,14 +2988,17 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
        if (!obj)
                return -ENOENT;
 
-       active = __I915_BO_ACTIVE(obj);
-       for_each_active(active, idx) {
-               s64 *timeout = args->timeout_ns >= 0 ? &args->timeout_ns : NULL;
-               ret = i915_gem_active_wait_unlocked(&obj->last_read[idx],
-                                                   I915_WAIT_INTERRUPTIBLE,
-                                                   timeout, rps);
-               if (ret)
-                       break;
+       start = ktime_get();
+
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE | I915_WAIT_ALL,
+                                  to_wait_timeout(args->timeout_ns),
+                                  to_rps_client(file));
+
+       if (args->timeout_ns > 0) {
+               args->timeout_ns -= ktime_to_ns(ktime_sub(ktime_get(), start));
+               if (args->timeout_ns < 0)
+                       args->timeout_ns = 0;
        }
 
        i915_gem_object_put_unlocked(obj);
@@ -3283,7 +3417,13 @@ i915_gem_object_set_to_gtt_domain(struct drm_i915_gem_object *obj, bool write)
        uint32_t old_write_domain, old_read_domains;
        int ret;
 
-       ret = i915_gem_object_wait_rendering(obj, !write);
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED |
+                                  (write ? I915_WAIT_ALL : 0),
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                return ret;
 
@@ -3400,7 +3540,12 @@ restart:
                 * If we wait upon the object, we know that all the bound
                 * VMA are no longer active.
                 */
-               ret = i915_gem_object_wait_rendering(obj, false);
+               ret = i915_gem_object_wait(obj,
+                                          I915_WAIT_INTERRUPTIBLE |
+                                          I915_WAIT_LOCKED |
+                                          I915_WAIT_ALL,
+                                          MAX_SCHEDULE_TIMEOUT,
+                                          NULL);
                if (ret)
                        return ret;
 
@@ -3647,7 +3792,13 @@ i915_gem_object_set_to_cpu_domain(struct drm_i915_gem_object *obj, bool write)
        uint32_t old_write_domain, old_read_domains;
        int ret;
 
-       ret = i915_gem_object_wait_rendering(obj, !write);
+       lockdep_assert_held(&obj->base.dev->struct_mutex);
+       ret = i915_gem_object_wait(obj,
+                                  I915_WAIT_INTERRUPTIBLE |
+                                  I915_WAIT_LOCKED |
+                                  (write ? I915_WAIT_ALL : 0),
+                                  MAX_SCHEDULE_TIMEOUT,
+                                  NULL);
        if (ret)
                return ret;
 
@@ -3703,7 +3854,7 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
        struct drm_i915_file_private *file_priv = file->driver_priv;
        unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
        struct drm_i915_gem_request *request, *target = NULL;
-       int ret;
+       long ret;
 
        /* ABI: return -EIO if already wedged */
        if (i915_terminally_wedged(&dev_priv->gpu_error))
@@ -3730,10 +3881,12 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
        if (target == NULL)
                return 0;
 
-       ret = i915_wait_request(target, I915_WAIT_INTERRUPTIBLE, NULL, NULL);
+       ret = i915_wait_request(target,
+                               I915_WAIT_INTERRUPTIBLE,
+                               MAX_SCHEDULE_TIMEOUT);
        i915_gem_request_put(target);
 
-       return ret;
+       return ret < 0 ? ret : 0;
 }
 
 static bool
index 5e38bc04a4f0c917c7ab27974b3f3f8161c737da..fbe0923fe0bcbf1ca9e9f337ab1a535ec8a30a3c 100644 (file)
@@ -59,31 +59,9 @@ static bool i915_fence_enable_signaling(struct dma_fence *fence)
 
 static signed long i915_fence_wait(struct dma_fence *fence,
                                   bool interruptible,
-                                  signed long timeout_jiffies)
+                                  signed long timeout)
 {
-       s64 timeout_ns, *timeout;
-       int ret;
-
-       if (timeout_jiffies != MAX_SCHEDULE_TIMEOUT) {
-               timeout_ns = jiffies_to_nsecs(timeout_jiffies);
-               timeout = &timeout_ns;
-       } else {
-               timeout = NULL;
-       }
-
-       ret = i915_wait_request(to_request(fence),
-                               interruptible, timeout,
-                               NO_WAITBOOST);
-       if (ret == -ETIME)
-               return 0;
-
-       if (ret < 0)
-               return ret;
-
-       if (timeout_jiffies != MAX_SCHEDULE_TIMEOUT)
-               timeout_jiffies = nsecs_to_jiffies(timeout_ns);
-
-       return timeout_jiffies;
+       return i915_wait_request(to_request(fence), interruptible, timeout);
 }
 
 static void i915_fence_value_str(struct dma_fence *fence, char *str, int size)
@@ -166,7 +144,7 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
        struct i915_gem_active *active, *next;
 
        trace_i915_gem_request_retire(request);
-       list_del(&request->link);
+       list_del_init(&request->link);
 
        /* We know the GPU must have read the request to have
         * sent us the seqno + interrupt, so use the position
@@ -224,7 +202,8 @@ void i915_gem_request_retire_upto(struct drm_i915_gem_request *req)
        struct drm_i915_gem_request *tmp;
 
        lockdep_assert_held(&req->i915->drm.struct_mutex);
-       GEM_BUG_ON(list_empty(&req->link));
+       if (list_empty(&req->link))
+               return;
 
        do {
                tmp = list_first_entry(&engine->request_list,
@@ -780,75 +759,48 @@ bool __i915_spin_request(const struct drm_i915_gem_request *req,
 
 /**
  * i915_wait_request - wait until execution of request has finished
- * @req: duh!
+ * @req: the request to wait upon
  * @flags: how to wait
- * @timeout: in - how long to wait (NULL forever); out - how much time remaining
- * @rps: client to charge for RPS boosting
+ * @timeout: how long to wait in jiffies
+ *
+ * i915_wait_request() waits for the request to be completed, for a
+ * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
+ * unbounded wait).
  *
- * Note: It is of utmost importance that the passed in seqno and reset_counter
- * values have been read by the caller in an smp safe manner. Where read-side
- * locks are involved, it is sufficient to read the reset_counter before
- * unlocking the lock that protects the seqno. For lockless tricks, the
- * reset_counter _must_ be read before, and an appropriate smp_rmb must be
- * inserted.
+ * If the caller holds the struct_mutex, the caller must pass I915_WAIT_LOCKED
+ * in via the flags, and vice versa if the struct_mutex is not held, the caller
+ * must not specify that the wait is locked.
  *
- * Returns 0 if the request was found within the alloted time. Else returns the
- * errno with remaining time filled in timeout argument.
+ * Returns the remaining time (in jiffies) if the request completed, which may
+ * be zero or -ETIME if the request is unfinished after the timeout expires.
+ * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
+ * pending before the request completes.
  */
-int i915_wait_request(struct drm_i915_gem_request *req,
-                     unsigned int flags,
-                     s64 *timeout,
-                     struct intel_rps_client *rps)
+long i915_wait_request(struct drm_i915_gem_request *req,
+                      unsigned int flags,
+                      long timeout)
 {
        const int state = flags & I915_WAIT_INTERRUPTIBLE ?
                TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
        DEFINE_WAIT(reset);
        struct intel_wait wait;
-       unsigned long timeout_remain;
-       int ret = 0;
 
        might_sleep();
 #if IS_ENABLED(CONFIG_LOCKDEP)
-       GEM_BUG_ON(!!lockdep_is_held(&req->i915->drm.struct_mutex) !=
+       GEM_BUG_ON(debug_locks &&
+                  !!lockdep_is_held(&req->i915->drm.struct_mutex) !=
                   !!(flags & I915_WAIT_LOCKED));
 #endif
+       GEM_BUG_ON(timeout < 0);
 
        if (i915_gem_request_completed(req))
-               return 0;
+               return timeout;
 
-       timeout_remain = MAX_SCHEDULE_TIMEOUT;
-       if (timeout) {
-               if (WARN_ON(*timeout < 0))
-                       return -EINVAL;
-
-               if (*timeout == 0)
-                       return -ETIME;
-
-               /* Record current time in case interrupted, or wedged */
-               timeout_remain = nsecs_to_jiffies_timeout(*timeout);
-               *timeout += ktime_get_raw_ns();
-       }
+       if (!timeout)
+               return -ETIME;
 
        trace_i915_gem_request_wait_begin(req);
 
-       /* This client is about to stall waiting for the GPU. In many cases
-        * this is undesirable and limits the throughput of the system, as
-        * many clients cannot continue processing user input/output whilst
-        * blocked. RPS autotuning may take tens of milliseconds to respond
-        * to the GPU load and thus incurs additional latency for the client.
-        * We can circumvent that by promoting the GPU frequency to maximum
-        * before we wait. This makes the GPU throttle up much more quickly
-        * (good for benchmarks and user experience, e.g. window animations),
-        * but at a cost of spending more power processing the workload
-        * (bad for battery). Not all clients even want their results
-        * immediately and for them we should just let the GPU select its own
-        * frequency to maximise efficiency. To prevent a single client from
-        * forcing the clocks too high for the whole system, we only allow
-        * each client to waitboost once in a busy period.
-        */
-       if (IS_RPS_CLIENT(rps) && INTEL_GEN(req->i915) >= 6)
-               gen6_rps_boost(req->i915, rps, req->emitted_jiffies);
-
        /* Optimistic short spin before touching IRQs */
        if (i915_spin_request(req, state, 5))
                goto complete;
@@ -867,16 +819,17 @@ int i915_wait_request(struct drm_i915_gem_request *req,
 
        for (;;) {
                if (signal_pending_state(state, current)) {
-                       ret = -ERESTARTSYS;
+                       timeout = -ERESTARTSYS;
                        break;
                }
 
-               timeout_remain = io_schedule_timeout(timeout_remain);
-               if (timeout_remain == 0) {
-                       ret = -ETIME;
+               if (!timeout) {
+                       timeout = -ETIME;
                        break;
                }
 
+               timeout = io_schedule_timeout(timeout);
+
                if (intel_wait_complete(&wait))
                        break;
 
@@ -923,40 +876,7 @@ wakeup:
 complete:
        trace_i915_gem_request_wait_end(req);
 
-       if (timeout) {
-               *timeout -= ktime_get_raw_ns();
-               if (*timeout < 0)
-                       *timeout = 0;
-
-               /*
-                * Apparently ktime isn't accurate enough and occasionally has a
-                * bit of mismatch in the jiffies<->nsecs<->ktime loop. So patch
-                * things up to make the test happy. We allow up to 1 jiffy.
-                *
-                * This is a regrssion from the timespec->ktime conversion.
-                */
-               if (ret == -ETIME && *timeout < jiffies_to_usecs(1)*1000)
-                       *timeout = 0;
-       }
-
-       if (IS_RPS_USER(rps) &&
-           req->fence.seqno == req->engine->last_submitted_seqno) {
-               /* The GPU is now idle and this client has stalled.
-                * Since no other client has submitted a request in the
-                * meantime, assume that this client is the only one
-                * supplying work to the GPU but is unable to keep that
-                * work supplied because it is waiting. Since the GPU is
-                * then never kept fully busy, RPS autoclocking will
-                * keep the clocks relatively low, causing further delays.
-                * Compensate by giving the synchronous client credit for
-                * a waitboost next time.
-                */
-               spin_lock(&req->i915->rps.client_lock);
-               list_del_init(&rps->link);
-               spin_unlock(&req->i915->rps.client_lock);
-       }
-
-       return ret;
+       return timeout;
 }
 
 static bool engine_retire_requests(struct intel_engine_cs *engine)
index 4e6d038cc9de78956456d8a135431129714c43d1..ae0913adfec62be89be36fab581231ef56bf0cfb 100644 (file)
@@ -228,13 +228,13 @@ struct intel_rps_client;
 #define IS_RPS_CLIENT(p) (!IS_ERR(p))
 #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
 
-int i915_wait_request(struct drm_i915_gem_request *req,
-                     unsigned int flags,
-                     s64 *timeout,
-                     struct intel_rps_client *rps)
+long i915_wait_request(struct drm_i915_gem_request *req,
+                      unsigned int flags,
+                      long timeout)
        __attribute__((nonnull(1)));
 #define I915_WAIT_INTERRUPTIBLE        BIT(0)
 #define I915_WAIT_LOCKED       BIT(1) /* struct_mutex held, handle GPU reset */
+#define I915_WAIT_ALL          BIT(2) /* used by i915_gem_object_wait() */
 
 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine);
 
@@ -583,14 +583,16 @@ static inline int __must_check
 i915_gem_active_wait(const struct i915_gem_active *active, struct mutex *mutex)
 {
        struct drm_i915_gem_request *request;
+       long ret;
 
        request = i915_gem_active_peek(active, mutex);
        if (!request)
                return 0;
 
-       return i915_wait_request(request,
-                                I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
-                                NULL, NULL);
+       ret = i915_wait_request(request,
+                               I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
+                               MAX_SCHEDULE_TIMEOUT);
+       return ret < 0 ? ret : 0;
 }
 
 /**
@@ -617,20 +619,18 @@ i915_gem_active_wait(const struct i915_gem_active *active, struct mutex *mutex)
  */
 static inline int
 i915_gem_active_wait_unlocked(const struct i915_gem_active *active,
-                             unsigned int flags,
-                             s64 *timeout,
-                             struct intel_rps_client *rps)
+                             unsigned int flags)
 {
        struct drm_i915_gem_request *request;
-       int ret = 0;
+       long ret = 0;
 
        request = i915_gem_active_get_unlocked(active);
        if (request) {
-               ret = i915_wait_request(request, flags, timeout, rps);
+               ret = i915_wait_request(request, flags, MAX_SCHEDULE_TIMEOUT);
                i915_gem_request_put(request);
        }
 
-       return ret;
+       return ret < 0 ? ret : 0;
 }
 
 /**
@@ -647,7 +647,7 @@ i915_gem_active_retire(struct i915_gem_active *active,
                       struct mutex *mutex)
 {
        struct drm_i915_gem_request *request;
-       int ret;
+       long ret;
 
        request = i915_gem_active_raw(active, mutex);
        if (!request)
@@ -655,8 +655,8 @@ i915_gem_active_retire(struct i915_gem_active *active,
 
        ret = i915_wait_request(request,
                                I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
-                               NULL, NULL);
-       if (ret)
+                               MAX_SCHEDULE_TIMEOUT);
+       if (ret < 0)
                return ret;
 
        list_del_init(&active->link);
index c6f780f5abc9cf3776edcfa67a6fa08a6b53239a..c49dd95413bd42373863495aec6a06704e432f48 100644 (file)
@@ -61,23 +61,13 @@ struct i915_mmu_object {
        bool attached;
 };
 
-static void wait_rendering(struct drm_i915_gem_object *obj)
-{
-       unsigned long active = __I915_BO_ACTIVE(obj);
-       int idx;
-
-       for_each_active(active, idx)
-               i915_gem_active_wait_unlocked(&obj->last_read[idx],
-                                             0, NULL, NULL);
-}
-
 static void cancel_userptr(struct work_struct *work)
 {
        struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
        struct drm_i915_gem_object *obj = mo->obj;
        struct drm_device *dev = obj->base.dev;
 
-       wait_rendering(obj);
+       i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
 
        mutex_lock(&dev->struct_mutex);
        /* Cancel any active worker and force us to re-evaluate gup */
index cb7dd11277fdab0969a69918289f53de4c132cf1..e4800b81c59e1b95d9bb368bbbfbfbdcea1fae24 100644 (file)
@@ -12072,7 +12072,7 @@ static void intel_mmio_flip_work_func(struct work_struct *w)
 
        if (work->flip_queued_req)
                WARN_ON(i915_wait_request(work->flip_queued_req,
-                                         0, NULL, NO_WAITBOOST));
+                                         0, MAX_SCHEDULE_TIMEOUT) < 0);
 
        /* For framebuffer backed by dmabuf, wait for fence */
        resv = i915_gem_object_get_dmabuf_resv(obj);
@@ -14187,19 +14187,21 @@ static int intel_atomic_prepare_commit(struct drm_device *dev,
                for_each_plane_in_state(state, plane, plane_state, i) {
                        struct intel_plane_state *intel_plane_state =
                                to_intel_plane_state(plane_state);
+                       long timeout;
 
                        if (!intel_plane_state->wait_req)
                                continue;
 
-                       ret = i915_wait_request(intel_plane_state->wait_req,
-                                               I915_WAIT_INTERRUPTIBLE,
-                                               NULL, NULL);
-                       if (ret) {
+                       timeout = i915_wait_request(intel_plane_state->wait_req,
+                                                   I915_WAIT_INTERRUPTIBLE,
+                                                   MAX_SCHEDULE_TIMEOUT);
+                       if (timeout < 0) {
                                /* Any hang should be swallowed by the wait */
-                               WARN_ON(ret == -EIO);
+                               WARN_ON(timeout == -EIO);
                                mutex_lock(&dev->struct_mutex);
                                drm_atomic_helper_cleanup_planes(dev, state);
                                mutex_unlock(&dev->struct_mutex);
+                               ret = timeout;
                                break;
                        }
                }
@@ -14403,7 +14405,7 @@ static void intel_atomic_commit_tail(struct drm_atomic_state *state)
        bool hw_check = intel_state->modeset;
        unsigned long put_domains[I915_MAX_PIPES] = {};
        unsigned crtc_vblank_mask = 0;
-       int i, ret;
+       int i;
 
        for_each_plane_in_state(state, plane, plane_state, i) {
                struct intel_plane_state *intel_plane_state =
@@ -14412,11 +14414,10 @@ static void intel_atomic_commit_tail(struct drm_atomic_state *state)
                if (!intel_plane_state->wait_req)
                        continue;
 
-               ret = i915_wait_request(intel_plane_state->wait_req,
-                                       0, NULL, NULL);
                /* EIO should be eaten, and we can't get interrupted in the
                 * worker, and blocking commits have waited already. */
-               WARN_ON(ret);
+               WARN_ON(i915_wait_request(intel_plane_state->wait_req,
+                                         0, MAX_SCHEDULE_TIMEOUT) < 0);
        }
 
        drm_atomic_helper_wait_for_dependencies(state);
@@ -14780,7 +14781,11 @@ intel_prepare_plane_fb(struct drm_plane *plane,
                 * can safely continue.
                 */
                if (needs_modeset(crtc_state))
-                       ret = i915_gem_object_wait_rendering(old_obj, true);
+                       ret = i915_gem_object_wait(old_obj,
+                                                  I915_WAIT_INTERRUPTIBLE |
+                                                  I915_WAIT_LOCKED,
+                                                  MAX_SCHEDULE_TIMEOUT,
+                                                  NULL);
                if (ret) {
                        /* GPU hangs should have been swallowed by the wait */
                        WARN_ON(ret == -EIO);
index 67a70c5e64534bb96f8bc4ad131e28cceee870dc..8ef735faa6034a6c6491e35d8a1fb8097d7d134e 100644 (file)
@@ -2155,7 +2155,9 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
 {
        struct intel_ring *ring = req->ring;
        struct drm_i915_gem_request *target;
-       int ret;
+       long timeout;
+
+       lockdep_assert_held(&req->i915->drm.struct_mutex);
 
        intel_ring_update_space(ring);
        if (ring->space >= bytes)
@@ -2185,11 +2187,11 @@ static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
        if (WARN_ON(&target->ring_link == &ring->request_list))
                return -ENOSPC;
 
-       ret = i915_wait_request(target,
-                               I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
-                               NULL, NO_WAITBOOST);
-       if (ret)
-               return ret;
+       timeout = i915_wait_request(target,
+                                   I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
+                                   MAX_SCHEDULE_TIMEOUT);
+       if (timeout < 0)
+               return timeout;
 
        i915_gem_request_retire_upto(target);
 
index 32b2e6332ccff4a46374263844b2e0669ad88c24..884a5ae2225d3f5c3f6cd8df50923fefed089f75 100644 (file)
@@ -524,8 +524,7 @@ static inline int intel_engine_idle(struct intel_engine_cs *engine,
                                    unsigned int flags)
 {
        /* Wait upon the last request to be completed */
-       return i915_gem_active_wait_unlocked(&engine->last_request,
-                                            flags, NULL, NULL);
+       return i915_gem_active_wait_unlocked(&engine->last_request, flags);
 }
 
 int intel_init_render_ring_buffer(struct intel_engine_cs *engine);