drm/i915: Move the global sync optimisation to the timeline

[karo-tx-linux.git] / drivers / gpu / drm / i915 / i915_gem_request.c

/*
 * Copyright © 2008-2015 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 */

#include <linux/prefetch.h>
#include <linux/dma-fence-array.h>

#include "i915_drv.h"

static const char *i915_fence_get_driver_name(struct dma_fence *fence)
{
        return "i915";
}

static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
{
        /* Timelines are bound by eviction to a VM. However, since
         * we only have a global seqno at the moment, we only have
         * a single timeline. Note that each timeline will have
         * multiple execution contexts (fence contexts) as we allow
         * engines within a single timeline to execute in parallel.
         */
        return to_request(fence)->timeline->common->name;
}

static bool i915_fence_signaled(struct dma_fence *fence)
{
        return i915_gem_request_completed(to_request(fence));
}

static bool i915_fence_enable_signaling(struct dma_fence *fence)
{
        if (i915_fence_signaled(fence))
                return false;

        intel_engine_enable_signaling(to_request(fence));
        return true;
}

static signed long i915_fence_wait(struct dma_fence *fence,
                                   bool interruptible,
                                   signed long timeout)
{
        return i915_wait_request(to_request(fence), interruptible, timeout);
}

static void i915_fence_value_str(struct dma_fence *fence, char *str, int size)
{
        snprintf(str, size, "%u", fence->seqno);
}

static void i915_fence_timeline_value_str(struct dma_fence *fence, char *str,
                                          int size)
{
        snprintf(str, size, "%u",
                 intel_engine_get_seqno(to_request(fence)->engine));
}

static void i915_fence_release(struct dma_fence *fence)
{
        struct drm_i915_gem_request *req = to_request(fence);

        kmem_cache_free(req->i915->requests, req);
}

const struct dma_fence_ops i915_fence_ops = {
        .get_driver_name = i915_fence_get_driver_name,
        .get_timeline_name = i915_fence_get_timeline_name,
        .enable_signaling = i915_fence_enable_signaling,
        .signaled = i915_fence_signaled,
        .wait = i915_fence_wait,
        .release = i915_fence_release,
        .fence_value_str = i915_fence_value_str,
        .timeline_value_str = i915_fence_timeline_value_str,
};

int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
                                   struct drm_file *file)
{
        struct drm_i915_private *dev_private;
        struct drm_i915_file_private *file_priv;

        WARN_ON(!req || !file || req->file_priv);

        if (!req || !file)
                return -EINVAL;

        if (req->file_priv)
                return -EINVAL;

        dev_private = req->i915;
        file_priv = file->driver_priv;

        spin_lock(&file_priv->mm.lock);
        req->file_priv = file_priv;
        list_add_tail(&req->client_list, &file_priv->mm.request_list);
        spin_unlock(&file_priv->mm.lock);

        return 0;
}

static inline void
i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
{
        struct drm_i915_file_private *file_priv = request->file_priv;

        if (!file_priv)
                return;

        spin_lock(&file_priv->mm.lock);
        list_del(&request->client_list);
        request->file_priv = NULL;
        spin_unlock(&file_priv->mm.lock);
}

void i915_gem_retire_noop(struct i915_gem_active *active,
                          struct drm_i915_gem_request *request)
{
        /* Space left intentionally blank */
}

static void i915_gem_request_retire(struct drm_i915_gem_request *request)
{
        struct i915_gem_active *active, *next;

        lockdep_assert_held(&request->i915->drm.struct_mutex);
        GEM_BUG_ON(!i915_gem_request_completed(request));

        trace_i915_gem_request_retire(request);
        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
         * of tail of the request to update the last known position
         * of the GPU head.
         *
         * Note this requires that we are always called in request
         * completion order.
         */
        list_del(&request->ring_link);
        request->ring->last_retired_head = request->postfix;

        /* Walk through the active list, calling retire on each. This allows
         * objects to track their GPU activity and mark themselves as idle
         * when their *last* active request is completed (updating state
         * tracking lists for eviction, active references for GEM, etc).
         *
         * As the ->retire() may free the node, we decouple it first and
         * pass along the auxiliary information (to avoid dereferencing
         * the node after the callback).
         */
        list_for_each_entry_safe(active, next, &request->active_list, link) {
                /* In microbenchmarks or focusing upon time inside the kernel,
                 * we may spend an inordinate amount of time simply handling
                 * the retirement of requests and processing their callbacks.
                 * Of which, this loop itself is particularly hot due to the
                 * cache misses when jumping around the list of i915_gem_active.
                 * So we try to keep this loop as streamlined as possible and
                 * also prefetch the next i915_gem_active to try and hide
                 * the likely cache miss.
                 */
                prefetchw(next);

                INIT_LIST_HEAD(&active->link);
                RCU_INIT_POINTER(active->request, NULL);

                active->retire(active, request);
        }

        i915_gem_request_remove_from_client(request);

        if (request->previous_context) {
                if (i915.enable_execlists)
                        intel_lr_context_unpin(request->previous_context,
                                               request->engine);
        }

        i915_gem_context_put(request->ctx);

        dma_fence_signal(&request->fence);
        i915_gem_request_put(request);
}

void i915_gem_request_retire_upto(struct drm_i915_gem_request *req)
{
        struct intel_engine_cs *engine = req->engine;
        struct drm_i915_gem_request *tmp;

        lockdep_assert_held(&req->i915->drm.struct_mutex);
        if (list_empty(&req->link))
                return;

        do {
                tmp = list_first_entry(&engine->timeline->requests,
                                       typeof(*tmp), link);

                i915_gem_request_retire(tmp);
        } while (tmp != req);
}

static int i915_gem_check_wedge(struct drm_i915_private *dev_priv)
{
        struct i915_gpu_error *error = &dev_priv->gpu_error;

        if (i915_terminally_wedged(error))
                return -EIO;

        if (i915_reset_in_progress(error)) {
                /* Non-interruptible callers can't handle -EAGAIN, hence return
                 * -EIO unconditionally for these.
                 */
                if (!dev_priv->mm.interruptible)
                        return -EIO;

                return -EAGAIN;
        }

        return 0;
}

static int i915_gem_init_global_seqno(struct drm_i915_private *i915, u32 seqno)
{
        struct i915_gem_timeline *timeline = &i915->gt.global_timeline;
        struct intel_engine_cs *engine;
        enum intel_engine_id id;
        int ret;

        /* Carefully retire all requests without writing to the rings */
        ret = i915_gem_wait_for_idle(i915,
                                     I915_WAIT_INTERRUPTIBLE |
                                     I915_WAIT_LOCKED);
        if (ret)
                return ret;

        i915_gem_retire_requests(i915);

        /* If the seqno wraps around, we need to clear the breadcrumb rbtree */
        if (!i915_seqno_passed(seqno, timeline->next_seqno)) {
                while (intel_kick_waiters(i915) || intel_kick_signalers(i915))
                        yield();
                yield();
        }

        /* Finally reset hw state */
        for_each_engine(engine, i915, id)
                intel_engine_init_global_seqno(engine, seqno);

        list_for_each_entry(timeline, &i915->gt.timelines, link) {
                for_each_engine(engine, i915, id) {
                        struct intel_timeline *tl = &timeline->engine[id];

                        memset(tl->sync_seqno, 0, sizeof(tl->sync_seqno));
                }
        }

        return 0;
}

int i915_gem_set_global_seqno(struct drm_device *dev, u32 seqno)
{
        struct drm_i915_private *dev_priv = to_i915(dev);
        int ret;

        lockdep_assert_held(&dev_priv->drm.struct_mutex);

        if (seqno == 0)
                return -EINVAL;

        /* HWS page needs to be set less than what we
         * will inject to ring
         */
        ret = i915_gem_init_global_seqno(dev_priv, seqno - 1);
        if (ret)
                return ret;

        dev_priv->gt.global_timeline.next_seqno = seqno;
        return 0;
}

static int i915_gem_get_global_seqno(struct drm_i915_private *dev_priv,
                                     u32 *seqno)
{
        struct i915_gem_timeline *tl = &dev_priv->gt.global_timeline;

        /* reserve 0 for non-seqno */
        if (unlikely(tl->next_seqno == 0)) {
                int ret;

                ret = i915_gem_init_global_seqno(dev_priv, 0);
                if (ret)
                        return ret;

                tl->next_seqno = 1;
        }

        *seqno = tl->next_seqno++;
        return 0;
}

static int __i915_sw_fence_call
submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
{
        struct drm_i915_gem_request *request =
                container_of(fence, typeof(*request), submit);
        struct intel_engine_cs *engine = request->engine;

        if (state != FENCE_COMPLETE)
                return NOTIFY_DONE;

        /* Will be called from irq-context when using foreign DMA fences */

        engine->timeline->last_submitted_seqno = request->fence.seqno;

        engine->emit_breadcrumb(request,
                                request->ring->vaddr + request->postfix);
        engine->submit_request(request);

        return NOTIFY_DONE;
}

/**
 * i915_gem_request_alloc - allocate a request structure
 *
 * @engine: engine that we wish to issue the request on.
 * @ctx: context that the request will be associated with.
 *       This can be NULL if the request is not directly related to
 *       any specific user context, in which case this function will
 *       choose an appropriate context to use.
 *
 * Returns a pointer to the allocated request if successful,
 * or an error code if not.
 */
struct drm_i915_gem_request *
i915_gem_request_alloc(struct intel_engine_cs *engine,
                       struct i915_gem_context *ctx)
{
        struct drm_i915_private *dev_priv = engine->i915;
        struct drm_i915_gem_request *req;
        u32 seqno;
        int ret;

        /* ABI: Before userspace accesses the GPU (e.g. execbuffer), report
         * EIO if the GPU is already wedged, or EAGAIN to drop the struct_mutex
         * and restart.
         */
        ret = i915_gem_check_wedge(dev_priv);
        if (ret)
                return ERR_PTR(ret);

        /* Move the oldest request to the slab-cache (if not in use!) */
        req = list_first_entry_or_null(&engine->timeline->requests,
                                       typeof(*req), link);
        if (req && i915_gem_request_completed(req))
                i915_gem_request_retire(req);

        /* Beware: Dragons be flying overhead.
         *
         * We use RCU to look up requests in flight. The lookups may
         * race with the request being allocated from the slab freelist.
         * That is the request we are writing to here, may be in the process
         * of being read by __i915_gem_active_get_rcu(). As such,
         * we have to be very careful when overwriting the contents. During
         * the RCU lookup, we change chase the request->engine pointer,
         * read the request->global_seqno and increment the reference count.
         *
         * The reference count is incremented atomically. If it is zero,
         * the lookup knows the request is unallocated and complete. Otherwise,
         * it is either still in use, or has been reallocated and reset
         * with dma_fence_init(). This increment is safe for release as we
         * check that the request we have a reference to and matches the active
         * request.
         *
         * Before we increment the refcount, we chase the request->engine
         * pointer. We must not call kmem_cache_zalloc() or else we set
         * that pointer to NULL and cause a crash during the lookup. If
         * we see the request is completed (based on the value of the
         * old engine and seqno), the lookup is complete and reports NULL.
         * If we decide the request is not completed (new engine or seqno),
         * then we grab a reference and double check that it is still the
         * active request - which it won't be and restart the lookup.
         *
         * Do not use kmem_cache_zalloc() here!
         */
        req = kmem_cache_alloc(dev_priv->requests, GFP_KERNEL);
        if (!req)
                return ERR_PTR(-ENOMEM);

        ret = i915_gem_get_global_seqno(dev_priv, &seqno);
        if (ret)
                goto err;

        req->timeline = engine->timeline;

        spin_lock_init(&req->lock);
        dma_fence_init(&req->fence,
                       &i915_fence_ops,
                       &req->lock,
                       req->timeline->fence_context,
                       seqno);

        i915_sw_fence_init(&req->submit, submit_notify);

        INIT_LIST_HEAD(&req->active_list);
        req->i915 = dev_priv;
        req->engine = engine;
        req->global_seqno = seqno;
        req->ctx = i915_gem_context_get(ctx);

        /* No zalloc, must clear what we need by hand */
        req->previous_context = NULL;
        req->file_priv = NULL;
        req->batch = NULL;

        /*
         * Reserve space in the ring buffer for all the commands required to
         * eventually emit this request. This is to guarantee that the
         * i915_add_request() call can't fail. Note that the reserve may need
         * to be redone if the request is not actually submitted straight
         * away, e.g. because a GPU scheduler has deferred it.
         */
        req->reserved_space = MIN_SPACE_FOR_ADD_REQUEST;
        GEM_BUG_ON(req->reserved_space < engine->emit_breadcrumb_sz);

        if (i915.enable_execlists)
                ret = intel_logical_ring_alloc_request_extras(req);
        else
                ret = intel_ring_alloc_request_extras(req);
        if (ret)
                goto err_ctx;

        /* Record the position of the start of the request so that
         * should we detect the updated seqno part-way through the
         * GPU processing the request, we never over-estimate the
         * position of the head.
         */
        req->head = req->ring->tail;

        return req;

err_ctx:
        i915_gem_context_put(ctx);
err:
        kmem_cache_free(dev_priv->requests, req);
        return ERR_PTR(ret);
}

static int
i915_gem_request_await_request(struct drm_i915_gem_request *to,
                               struct drm_i915_gem_request *from)
{
        int ret;

        GEM_BUG_ON(to == from);

        if (to->timeline == from->timeline)
                return 0;

        if (to->engine == from->engine) {
                ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
                                                       &from->submit,
                                                       GFP_KERNEL);
                return ret < 0 ? ret : 0;
        }

        if (!from->global_seqno) {
                ret = i915_sw_fence_await_dma_fence(&to->submit,
                                                    &from->fence, 0,
                                                    GFP_KERNEL);
                return ret < 0 ? ret : 0;
        }

        if (from->global_seqno <= to->timeline->sync_seqno[from->engine->id])
                return 0;

        trace_i915_gem_ring_sync_to(to, from);
        if (!i915.semaphores) {
                if (!i915_spin_request(from, TASK_INTERRUPTIBLE, 2)) {
                        ret = i915_sw_fence_await_dma_fence(&to->submit,
                                                            &from->fence, 0,
                                                            GFP_KERNEL);
                        if (ret < 0)
                                return ret;
                }
        } else {
                ret = to->engine->semaphore.sync_to(to, from);
                if (ret)
                        return ret;
        }

        to->timeline->sync_seqno[from->engine->id] = from->global_seqno;
        return 0;
}

int
i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
                                 struct dma_fence *fence)
{
        struct dma_fence_array *array;
        int ret;
        int i;

        if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
                return 0;

        if (dma_fence_is_i915(fence))
                return i915_gem_request_await_request(req, to_request(fence));

        if (!dma_fence_is_array(fence)) {
                ret = i915_sw_fence_await_dma_fence(&req->submit,
                                                    fence, I915_FENCE_TIMEOUT,
                                                    GFP_KERNEL);
                return ret < 0 ? ret : 0;
        }

        /* Note that if the fence-array was created in signal-on-any mode,
         * we should *not* decompose it into its individual fences. However,
         * we don't currently store which mode the fence-array is operating
         * in. Fortunately, the only user of signal-on-any is private to
         * amdgpu and we should not see any incoming fence-array from
         * sync-file being in signal-on-any mode.
         */

        array = to_dma_fence_array(fence);
        for (i = 0; i < array->num_fences; i++) {
                struct dma_fence *child = array->fences[i];

                if (dma_fence_is_i915(child))
                        ret = i915_gem_request_await_request(req,
                                                             to_request(child));
                else
                        ret = i915_sw_fence_await_dma_fence(&req->submit,
                                                            child, I915_FENCE_TIMEOUT,
                                                            GFP_KERNEL);
                if (ret < 0)
                        return ret;
        }

        return 0;
}

/**
 * i915_gem_request_await_object - set this request to (async) wait upon a bo
 *
 * @to: request we are wishing to use
 * @obj: object which may be in use on another ring.
 *
 * This code is meant to abstract object synchronization with the GPU.
 * Conceptually we serialise writes between engines inside the GPU.
 * We only allow one engine to write into a buffer at any time, but
 * multiple readers. To ensure each has a coherent view of memory, we must:
 *
 * - If there is an outstanding write request to the object, the new
 *   request must wait for it to complete (either CPU or in hw, requests
 *   on the same ring will be naturally ordered).
 *
 * - If we are a write request (pending_write_domain is set), the new
 *   request must wait for outstanding read requests to complete.
 *
 * Returns 0 if successful, else propagates up the lower layer error.
 */
int
i915_gem_request_await_object(struct drm_i915_gem_request *to,
                              struct drm_i915_gem_object *obj,
                              bool write)
{
        struct dma_fence *excl;
        int ret = 0;

        if (write) {
                struct dma_fence **shared;
                unsigned int count, i;

                ret = reservation_object_get_fences_rcu(obj->resv,
                                                        &excl, &count, &shared);
                if (ret)
                        return ret;

                for (i = 0; i < count; i++) {
                        ret = i915_gem_request_await_dma_fence(to, shared[i]);
                        if (ret)
                                break;

                        dma_fence_put(shared[i]);
                }

                for (; i < count; i++)
                        dma_fence_put(shared[i]);
                kfree(shared);
        } else {
                excl = reservation_object_get_excl_rcu(obj->resv);
        }

        if (excl) {
                if (ret == 0)
                        ret = i915_gem_request_await_dma_fence(to, excl);

                dma_fence_put(excl);
        }

        return ret;
}

static void i915_gem_mark_busy(const struct intel_engine_cs *engine)
{
        struct drm_i915_private *dev_priv = engine->i915;

        dev_priv->gt.active_engines |= intel_engine_flag(engine);
        if (dev_priv->gt.awake)
                return;

        intel_runtime_pm_get_noresume(dev_priv);
        dev_priv->gt.awake = true;

        intel_enable_gt_powersave(dev_priv);
        i915_update_gfx_val(dev_priv);
        if (INTEL_GEN(dev_priv) >= 6)
                gen6_rps_busy(dev_priv);

        queue_delayed_work(dev_priv->wq,
                           &dev_priv->gt.retire_work,
                           round_jiffies_up_relative(HZ));
}

/*
 * NB: This function is not allowed to fail. Doing so would mean the the
 * request is not being tracked for completion but the work itself is
 * going to happen on the hardware. This would be a Bad Thing(tm).
 */
void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
{
        struct intel_engine_cs *engine = request->engine;
        struct intel_ring *ring = request->ring;
        struct intel_timeline *timeline = request->timeline;
        struct drm_i915_gem_request *prev;
        int err;

        lockdep_assert_held(&request->i915->drm.struct_mutex);
        trace_i915_gem_request_add(request);

        /*
         * To ensure that this call will not fail, space for its emissions
         * should already have been reserved in the ring buffer. Let the ring
         * know that it is time to use that space up.
         */
        request->reserved_space = 0;

        /*
         * Emit any outstanding flushes - execbuf can fail to emit the flush
         * after having emitted the batchbuffer command. Hence we need to fix
         * things up similar to emitting the lazy request. The difference here
         * is that the flush _must_ happen before the next request, no matter
         * what.
         */
        if (flush_caches) {
                err = engine->emit_flush(request, EMIT_FLUSH);

                /* Not allowed to fail! */
                WARN(err, "engine->emit_flush() failed: %d!\n", err);
        }

        /* Record the position of the start of the breadcrumb so that
         * should we detect the updated seqno part-way through the
         * GPU processing the request, we never over-estimate the
         * position of the ring's HEAD.
         */
        err = intel_ring_begin(request, engine->emit_breadcrumb_sz);
        GEM_BUG_ON(err);
        request->postfix = ring->tail;
        ring->tail += engine->emit_breadcrumb_sz * sizeof(u32);

        /* Seal the request and mark it as pending execution. Note that
         * we may inspect this state, without holding any locks, during
         * hangcheck. Hence we apply the barrier to ensure that we do not
         * see a more recent value in the hws than we are tracking.
         */

        prev = i915_gem_active_raw(&timeline->last_request,
                                   &request->i915->drm.struct_mutex);
        if (prev)
                i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
                                             &request->submitq);

        request->emitted_jiffies = jiffies;
        request->previous_seqno = timeline->last_pending_seqno;
        timeline->last_pending_seqno = request->fence.seqno;
        i915_gem_active_set(&timeline->last_request, request);
        list_add_tail(&request->link, &timeline->requests);
        list_add_tail(&request->ring_link, &ring->request_list);

        i915_gem_mark_busy(engine);

        local_bh_disable();
        i915_sw_fence_commit(&request->submit);
        local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
}

static void reset_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)
{
        unsigned long flags;

        spin_lock_irqsave(&q->lock, flags);
        if (list_empty(&wait->task_list))
                __add_wait_queue(q, wait);
        spin_unlock_irqrestore(&q->lock, flags);
}

static unsigned long local_clock_us(unsigned int *cpu)
{
        unsigned long t;

        /* Cheaply and approximately convert from nanoseconds to microseconds.
         * The result and subsequent calculations are also defined in the same
         * approximate microseconds units. The principal source of timing
         * error here is from the simple truncation.
         *
         * Note that local_clock() is only defined wrt to the current CPU;
         * the comparisons are no longer valid if we switch CPUs. Instead of
         * blocking preemption for the entire busywait, we can detect the CPU
         * switch and use that as indicator of system load and a reason to
         * stop busywaiting, see busywait_stop().
         */
        *cpu = get_cpu();
        t = local_clock() >> 10;
        put_cpu();

        return t;
}

static bool busywait_stop(unsigned long timeout, unsigned int cpu)
{
        unsigned int this_cpu;

        if (time_after(local_clock_us(&this_cpu), timeout))
                return true;

        return this_cpu != cpu;
}

bool __i915_spin_request(const struct drm_i915_gem_request *req,
                         int state, unsigned long timeout_us)
{
        unsigned int cpu;

        /* When waiting for high frequency requests, e.g. during synchronous
         * rendering split between the CPU and GPU, the finite amount of time
         * required to set up the irq and wait upon it limits the response
         * rate. By busywaiting on the request completion for a short while we
         * can service the high frequency waits as quick as possible. However,
         * if it is a slow request, we want to sleep as quickly as possible.
         * The tradeoff between waiting and sleeping is roughly the time it
         * takes to sleep on a request, on the order of a microsecond.
         */

        timeout_us += local_clock_us(&cpu);
        do {
                if (__i915_gem_request_completed(req))
                        return true;

                if (signal_pending_state(state, current))
                        break;

                if (busywait_stop(timeout_us, cpu))
                        break;

                cpu_relax_lowlatency();
        } while (!need_resched());

        return false;
}

static long
__i915_request_wait_for_submit(struct drm_i915_gem_request *request,
                               unsigned int flags,
                               long timeout)
{
        const int state = flags & I915_WAIT_INTERRUPTIBLE ?
                TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
        wait_queue_head_t *q = &request->i915->gpu_error.wait_queue;
        DEFINE_WAIT(reset);
        DEFINE_WAIT(wait);

        if (flags & I915_WAIT_LOCKED)
                add_wait_queue(q, &reset);

        do {
                prepare_to_wait(&request->submit.wait, &wait, state);

                if (i915_sw_fence_done(&request->submit))
                        break;

                if (flags & I915_WAIT_LOCKED &&
                    i915_reset_in_progress(&request->i915->gpu_error)) {
                        __set_current_state(TASK_RUNNING);
                        i915_reset(request->i915);
                        reset_wait_queue(q, &reset);
                        continue;
                }

                if (signal_pending_state(state, current)) {
                        timeout = -ERESTARTSYS;
                        break;
                }

                timeout = io_schedule_timeout(timeout);
        } while (timeout);
        finish_wait(&request->submit.wait, &wait);

        if (flags & I915_WAIT_LOCKED)
                remove_wait_queue(q, &reset);

        return timeout;
}

/**
 * i915_wait_request - wait until execution of request has finished
 * @req: the request to wait upon
 * @flags: how to wait
 * @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).
 *
 * 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 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.
 */
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;

        might_sleep();
#if IS_ENABLED(CONFIG_LOCKDEP)
        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 timeout;

        if (!timeout)
                return -ETIME;

        trace_i915_gem_request_wait_begin(req);

        if (!i915_sw_fence_done(&req->submit)) {
                timeout = __i915_request_wait_for_submit(req, flags, timeout);
                if (timeout < 0)
                        goto complete;

                GEM_BUG_ON(!i915_sw_fence_done(&req->submit));
        }
        GEM_BUG_ON(!req->global_seqno);

        /* Optimistic short spin before touching IRQs */
        if (i915_spin_request(req, state, 5))
                goto complete;

        set_current_state(state);
        if (flags & I915_WAIT_LOCKED)
                add_wait_queue(&req->i915->gpu_error.wait_queue, &reset);

        intel_wait_init(&wait, req->global_seqno);
        if (intel_engine_add_wait(req->engine, &wait))
                /* In order to check that we haven't missed the interrupt
                 * as we enabled it, we need to kick ourselves to do a
                 * coherent check on the seqno before we sleep.
                 */
                goto wakeup;

        for (;;) {
                if (signal_pending_state(state, current)) {
                        timeout = -ERESTARTSYS;
                        break;
                }

                if (!timeout) {
                        timeout = -ETIME;
                        break;
                }

                timeout = io_schedule_timeout(timeout);

                if (intel_wait_complete(&wait))
                        break;

                set_current_state(state);

wakeup:
                /* Carefully check if the request is complete, giving time
                 * for the seqno to be visible following the interrupt.
                 * We also have to check in case we are kicked by the GPU
                 * reset in order to drop the struct_mutex.
                 */
                if (__i915_request_irq_complete(req))
                        break;

                /* If the GPU is hung, and we hold the lock, reset the GPU
                 * and then check for completion. On a full reset, the engine's
                 * HW seqno will be advanced passed us and we are complete.
                 * If we do a partial reset, we have to wait for the GPU to
                 * resume and update the breadcrumb.
                 *
                 * If we don't hold the mutex, we can just wait for the worker
                 * to come along and update the breadcrumb (either directly
                 * itself, or indirectly by recovering the GPU).
                 */
                if (flags & I915_WAIT_LOCKED &&
                    i915_reset_in_progress(&req->i915->gpu_error)) {
                        __set_current_state(TASK_RUNNING);
                        i915_reset(req->i915);
                        reset_wait_queue(&req->i915->gpu_error.wait_queue,
                                         &reset);
                        continue;
                }

                /* Only spin if we know the GPU is processing this request */
                if (i915_spin_request(req, state, 2))
                        break;
        }

        intel_engine_remove_wait(req->engine, &wait);
        if (flags & I915_WAIT_LOCKED)
                remove_wait_queue(&req->i915->gpu_error.wait_queue, &reset);
        __set_current_state(TASK_RUNNING);

complete:
        trace_i915_gem_request_wait_end(req);

        return timeout;
}

static bool engine_retire_requests(struct intel_engine_cs *engine)
{
        struct drm_i915_gem_request *request, *next;

        list_for_each_entry_safe(request, next,
                                 &engine->timeline->requests, link) {
                if (!i915_gem_request_completed(request))
                        return false;

                i915_gem_request_retire(request);
        }

        return true;
}

void i915_gem_retire_requests(struct drm_i915_private *dev_priv)
{
        struct intel_engine_cs *engine;
        unsigned int tmp;

        lockdep_assert_held(&dev_priv->drm.struct_mutex);

        if (dev_priv->gt.active_engines == 0)
                return;

        GEM_BUG_ON(!dev_priv->gt.awake);

        for_each_engine_masked(engine, dev_priv, dev_priv->gt.active_engines, tmp)
                if (engine_retire_requests(engine))
                        dev_priv->gt.active_engines &= ~intel_engine_flag(engine);

        if (dev_priv->gt.active_engines == 0)
                queue_delayed_work(dev_priv->wq,
                                   &dev_priv->gt.idle_work,
                                   msecs_to_jiffies(100));
}