Quelle  v3d_sched.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0+
/* Copyright (C) 2018 Broadcom */

/**
 * DOC: Broadcom V3D scheduling
 *
 * The shared DRM GPU scheduler is used to coordinate submitting jobs
 * to the hardware. Each DRM fd (roughly a client process) gets its
 * own scheduler entity, which will process jobs in order. The GPU
 * scheduler will schedule the clients with a FIFO scheduling algorithm.
 *
 * For simplicity, and in order to keep latency low for interactive
 * jobs when bulk background jobs are queued up, we submit a new job
 * to the HW only when it has completed the last one, instead of
 * filling up the CT[01]Q FIFOs with jobs. Similarly, we use
 * `drm_sched_job_add_dependency()` to manage the dependency between bin
 * and render, instead of having the clients submit jobs using the HW's
 * semaphores to interlock between them.
 */

#include <linux/sched/clock.h>
#include <linux/kthread.h>

#include <drm/drm_syncobj.h>

#include "v3d_drv.h"
#include "v3d_regs.h"
#include "v3d_trace.h"

#define V3D_CSD_CFG012_WG_COUNT_SHIFT 16

static struct v3d_job *
to_v3d_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_job, base);
}

static struct v3d_bin_job *
to_bin_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_bin_job, base.base);
}

static struct v3d_render_job *
to_render_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_render_job, base.base);
}

static struct v3d_tfu_job *
to_tfu_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_tfu_job, base.base);
}

static struct v3d_csd_job *
to_csd_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_csd_job, base.base);
}

static struct v3d_cpu_job *
to_cpu_job(struct drm_sched_job *sched_job)
{
 return container_of(sched_job, struct v3d_cpu_job, base.base);
}

static void
v3d_sched_job_free(struct drm_sched_job *sched_job)
{
 struct v3d_job *job = to_v3d_job(sched_job);

 v3d_job_cleanup(job);
}

void
v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
         unsigned int count)
{
 if (query_info->queries) {
  unsigned int i;

  for (i = 0; i < count; i++)
   drm_syncobj_put(query_info->queries[i].syncobj);

  kvfree(query_info->queries);
 }
}

void
v3d_performance_query_info_free(struct v3d_performance_query_info *query_info,
    unsigned int count)
{
 if (query_info->queries) {
  unsigned int i;

  for (i = 0; i < count; i++) {
   drm_syncobj_put(query_info->queries[i].syncobj);
   kvfree(query_info->queries[i].kperfmon_ids);
  }

  kvfree(query_info->queries);
 }
}

static void
v3d_cpu_job_free(struct drm_sched_job *sched_job)
{
 struct v3d_cpu_job *job = to_cpu_job(sched_job);

 v3d_timestamp_query_info_free(&job->timestamp_query,
          job->timestamp_query.count);

 v3d_performance_query_info_free(&job->performance_query,
     job->performance_query.count);

 v3d_job_cleanup(&job->base);
}

static void
v3d_switch_perfmon(struct v3d_dev *v3d, struct v3d_job *job)
{
 struct v3d_perfmon *perfmon = v3d->global_perfmon;

 if (!perfmon)
  perfmon = job->perfmon;

 if (perfmon == v3d->active_perfmon)
  return;

 if (perfmon != v3d->active_perfmon)
  v3d_perfmon_stop(v3d, v3d->active_perfmon, true);

 if (perfmon && v3d->active_perfmon != perfmon)
  v3d_perfmon_start(v3d, perfmon);
}

static void
v3d_job_start_stats(struct v3d_job *job, enum v3d_queue queue)
{
 struct v3d_dev *v3d = job->v3d;
 struct v3d_file_priv *file = job->file->driver_priv;
 struct v3d_stats *global_stats = &v3d->queue[queue].stats;
 struct v3d_stats *local_stats = &file->stats[queue];
 u64 now = local_clock();
 unsigned long flags;

 /*
 * We only need to disable local interrupts to appease lockdep who
 * otherwise would think v3d_job_start_stats vs v3d_stats_update has an
 * unsafe in-irq vs no-irq-off usage problem. This is a false positive
 * because all the locks are per queue and stats type, and all jobs are
 * completely one at a time serialised. More specifically:
 *
 * 1. Locks for GPU queues are updated from interrupt handlers under a
 *    spin lock and started here with preemption disabled.
 *
 * 2. Locks for CPU queues are updated from the worker with preemption
 *    disabled and equally started here with preemption disabled.
 *
 * Therefore both are consistent.
 *
 * 3. Because next job can only be queued after the previous one has
 *    been signaled, and locks are per queue, there is also no scope for
 *    the start part to race with the update part.
 */
 if (IS_ENABLED(CONFIG_LOCKDEP))
  local_irq_save(flags);
 else
  preempt_disable();

 write_seqcount_begin(&local_stats->lock);
 local_stats->start_ns = now;
 write_seqcount_end(&local_stats->lock);

 write_seqcount_begin(&global_stats->lock);
 global_stats->start_ns = now;
 write_seqcount_end(&global_stats->lock);

 if (IS_ENABLED(CONFIG_LOCKDEP))
  local_irq_restore(flags);
 else
  preempt_enable();
}

static void
v3d_stats_update(struct v3d_stats *stats, u64 now)
{
 write_seqcount_begin(&stats->lock);
 stats->enabled_ns += now - stats->start_ns;
 stats->jobs_completed++;
 stats->start_ns = 0;
 write_seqcount_end(&stats->lock);
}

void
v3d_job_update_stats(struct v3d_job *job, enum v3d_queue queue)
{
 struct v3d_dev *v3d = job->v3d;
 struct v3d_file_priv *file = job->file->driver_priv;
 struct v3d_stats *global_stats = &v3d->queue[queue].stats;
 u64 now = local_clock();
 unsigned long flags;

 /* See comment in v3d_job_start_stats() */
 if (IS_ENABLED(CONFIG_LOCKDEP))
  local_irq_save(flags);
 else
  preempt_disable();

 /* Don't update the local stats if the file context has already closed */
 if (file)
  v3d_stats_update(&file->stats[queue], now);
 else
  drm_dbg(&v3d->drm, "The file descriptor was closed before job completion\n");

 v3d_stats_update(global_stats, now);

 if (IS_ENABLED(CONFIG_LOCKDEP))
  local_irq_restore(flags);
 else
  preempt_enable();
}

static struct dma_fence *v3d_bin_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_bin_job *job = to_bin_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;
 struct drm_device *dev = &v3d->drm;
 struct dma_fence *fence;
 unsigned long irqflags;

 if (unlikely(job->base.base.s_fence->finished.error)) {
  spin_lock_irqsave(&v3d->job_lock, irqflags);
  v3d->bin_job = NULL;
  spin_unlock_irqrestore(&v3d->job_lock, irqflags);
  return NULL;
 }

 /* Lock required around bin_job update vs
 * v3d_overflow_mem_work().
 */
 spin_lock_irqsave(&v3d->job_lock, irqflags);
 v3d->bin_job = job;
 /* Clear out the overflow allocation, so we don't
 * reuse the overflow attached to a previous job.
 */
 V3D_CORE_WRITE(0, V3D_PTB_BPOS, 0);
 spin_unlock_irqrestore(&v3d->job_lock, irqflags);

 v3d_invalidate_caches(v3d);

 fence = v3d_fence_create(v3d, V3D_BIN);
 if (IS_ERR(fence))
  return NULL;

 if (job->base.irq_fence)
  dma_fence_put(job->base.irq_fence);
 job->base.irq_fence = dma_fence_get(fence);

 trace_v3d_submit_cl(dev, false, to_v3d_fence(fence)->seqno,
       job->start, job->end);

 v3d_job_start_stats(&job->base, V3D_BIN);
 v3d_switch_perfmon(v3d, &job->base);

 /* Set the current and end address of the control list.
 * Writing the end register is what starts the job.
 */
 if (job->qma) {
  V3D_CORE_WRITE(0, V3D_CLE_CT0QMA, job->qma);
  V3D_CORE_WRITE(0, V3D_CLE_CT0QMS, job->qms);
 }
 if (job->qts) {
  V3D_CORE_WRITE(0, V3D_CLE_CT0QTS,
          V3D_CLE_CT0QTS_ENABLE |
          job->qts);
 }
 V3D_CORE_WRITE(0, V3D_CLE_CT0QBA, job->start);
 V3D_CORE_WRITE(0, V3D_CLE_CT0QEA, job->end);

 return fence;
}

static struct dma_fence *v3d_render_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_render_job *job = to_render_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;
 struct drm_device *dev = &v3d->drm;
 struct dma_fence *fence;

 if (unlikely(job->base.base.s_fence->finished.error)) {
  v3d->render_job = NULL;
  return NULL;
 }

 v3d->render_job = job;

 /* Can we avoid this flush?  We need to be careful of
 * scheduling, though -- imagine job0 rendering to texture and
 * job1 reading, and them being executed as bin0, bin1,
 * render0, render1, so that render1's flush at bin time
 * wasn't enough.
 */
 v3d_invalidate_caches(v3d);

 fence = v3d_fence_create(v3d, V3D_RENDER);
 if (IS_ERR(fence))
  return NULL;

 if (job->base.irq_fence)
  dma_fence_put(job->base.irq_fence);
 job->base.irq_fence = dma_fence_get(fence);

 trace_v3d_submit_cl(dev, true, to_v3d_fence(fence)->seqno,
       job->start, job->end);

 v3d_job_start_stats(&job->base, V3D_RENDER);
 v3d_switch_perfmon(v3d, &job->base);

 /* XXX: Set the QCFG */

 /* Set the current and end address of the control list.
 * Writing the end register is what starts the job.
 */
 V3D_CORE_WRITE(0, V3D_CLE_CT1QBA, job->start);
 V3D_CORE_WRITE(0, V3D_CLE_CT1QEA, job->end);

 return fence;
}

static struct dma_fence *
v3d_tfu_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_tfu_job *job = to_tfu_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;
 struct drm_device *dev = &v3d->drm;
 struct dma_fence *fence;

 if (unlikely(job->base.base.s_fence->finished.error)) {
  v3d->tfu_job = NULL;
  return NULL;
 }

 v3d->tfu_job = job;

 fence = v3d_fence_create(v3d, V3D_TFU);
 if (IS_ERR(fence))
  return NULL;

 if (job->base.irq_fence)
  dma_fence_put(job->base.irq_fence);
 job->base.irq_fence = dma_fence_get(fence);

 trace_v3d_submit_tfu(dev, to_v3d_fence(fence)->seqno);

 v3d_job_start_stats(&job->base, V3D_TFU);

 V3D_WRITE(V3D_TFU_IIA(v3d->ver), job->args.iia);
 V3D_WRITE(V3D_TFU_IIS(v3d->ver), job->args.iis);
 V3D_WRITE(V3D_TFU_ICA(v3d->ver), job->args.ica);
 V3D_WRITE(V3D_TFU_IUA(v3d->ver), job->args.iua);
 V3D_WRITE(V3D_TFU_IOA(v3d->ver), job->args.ioa);
 if (v3d->ver >= V3D_GEN_71)
  V3D_WRITE(V3D_V7_TFU_IOC, job->args.v71.ioc);
 V3D_WRITE(V3D_TFU_IOS(v3d->ver), job->args.ios);
 V3D_WRITE(V3D_TFU_COEF0(v3d->ver), job->args.coef[0]);
 if (v3d->ver >= V3D_GEN_71 || (job->args.coef[0] & V3D_TFU_COEF0_USECOEF)) {
  V3D_WRITE(V3D_TFU_COEF1(v3d->ver), job->args.coef[1]);
  V3D_WRITE(V3D_TFU_COEF2(v3d->ver), job->args.coef[2]);
  V3D_WRITE(V3D_TFU_COEF3(v3d->ver), job->args.coef[3]);
 }
 /* ICFG kicks off the job. */
 V3D_WRITE(V3D_TFU_ICFG(v3d->ver), job->args.icfg | V3D_TFU_ICFG_IOC);

 return fence;
}

static struct dma_fence *
v3d_csd_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_csd_job *job = to_csd_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;
 struct drm_device *dev = &v3d->drm;
 struct dma_fence *fence;
 int i, csd_cfg0_reg;

 if (unlikely(job->base.base.s_fence->finished.error)) {
  v3d->csd_job = NULL;
  return NULL;
 }

 v3d->csd_job = job;

 v3d_invalidate_caches(v3d);

 fence = v3d_fence_create(v3d, V3D_CSD);
 if (IS_ERR(fence))
  return NULL;

 if (job->base.irq_fence)
  dma_fence_put(job->base.irq_fence);
 job->base.irq_fence = dma_fence_get(fence);

 trace_v3d_submit_csd(dev, to_v3d_fence(fence)->seqno);

 v3d_job_start_stats(&job->base, V3D_CSD);
 v3d_switch_perfmon(v3d, &job->base);

 csd_cfg0_reg = V3D_CSD_QUEUED_CFG0(v3d->ver);
 for (i = 1; i <= 6; i++)
  V3D_CORE_WRITE(0, csd_cfg0_reg + 4 * i, job->args.cfg[i]);

 /* Although V3D 7.1 has an eighth configuration register, we are not
 * using it. Therefore, make sure it remains unused.
 *
 * XXX: Set the CFG7 register
 */
 if (v3d->ver >= V3D_GEN_71)
  V3D_CORE_WRITE(0, V3D_V7_CSD_QUEUED_CFG7, 0);

 /* CFG0 write kicks off the job. */
 V3D_CORE_WRITE(0, csd_cfg0_reg, job->args.cfg[0]);

 return fence;
}

static void
v3d_rewrite_csd_job_wg_counts_from_indirect(struct v3d_cpu_job *job)
{
 struct v3d_indirect_csd_info *indirect_csd = &job->indirect_csd;
 struct v3d_bo *bo = to_v3d_bo(job->base.bo[0]);
 struct v3d_bo *indirect = to_v3d_bo(indirect_csd->indirect);
 struct drm_v3d_submit_csd *args = &indirect_csd->job->args;
 struct v3d_dev *v3d = job->base.v3d;
 u32 num_batches, *wg_counts;

 v3d_get_bo_vaddr(bo);
 v3d_get_bo_vaddr(indirect);

 wg_counts = (uint32_t *)(bo->vaddr + indirect_csd->offset);

 if (wg_counts[0] == 0 || wg_counts[1] == 0 || wg_counts[2] == 0)
  return;

 args->cfg[0] = wg_counts[0] << V3D_CSD_CFG012_WG_COUNT_SHIFT;
 args->cfg[1] = wg_counts[1] << V3D_CSD_CFG012_WG_COUNT_SHIFT;
 args->cfg[2] = wg_counts[2] << V3D_CSD_CFG012_WG_COUNT_SHIFT;

 num_batches = DIV_ROUND_UP(indirect_csd->wg_size, 16) *
        (wg_counts[0] * wg_counts[1] * wg_counts[2]);

 /* V3D 7.1.6 and later don't subtract 1 from the number of batches */
 if (v3d->ver < 71 || (v3d->ver == 71 && v3d->rev < 6))
  args->cfg[4] = num_batches - 1;
 else
  args->cfg[4] = num_batches;

 WARN_ON(args->cfg[4] == ~0);

 for (int i = 0; i < 3; i++) {
  /* 0xffffffff indicates that the uniform rewrite is not needed */
  if (indirect_csd->wg_uniform_offsets[i] != 0xffffffff) {
   u32 uniform_idx = indirect_csd->wg_uniform_offsets[i];
   ((uint32_t *)indirect->vaddr)[uniform_idx] = wg_counts[i];
  }
 }

 v3d_put_bo_vaddr(indirect);
 v3d_put_bo_vaddr(bo);
}

static void
v3d_timestamp_query(struct v3d_cpu_job *job)
{
 struct v3d_timestamp_query_info *timestamp_query = &job->timestamp_query;
 struct v3d_bo *bo = to_v3d_bo(job->base.bo[0]);
 u8 *value_addr;

 v3d_get_bo_vaddr(bo);

 for (int i = 0; i < timestamp_query->count; i++) {
  value_addr = ((u8 *)bo->vaddr) + timestamp_query->queries[i].offset;
  *((u64 *)value_addr) = i == 0 ? ktime_get_ns() : 0ull;

  drm_syncobj_replace_fence(timestamp_query->queries[i].syncobj,
       job->base.done_fence);
 }

 v3d_put_bo_vaddr(bo);
}

static void
v3d_reset_timestamp_queries(struct v3d_cpu_job *job)
{
 struct v3d_timestamp_query_info *timestamp_query = &job->timestamp_query;
 struct v3d_timestamp_query *queries = timestamp_query->queries;
 struct v3d_bo *bo = to_v3d_bo(job->base.bo[0]);
 u8 *value_addr;

 v3d_get_bo_vaddr(bo);

 for (int i = 0; i < timestamp_query->count; i++) {
  value_addr = ((u8 *)bo->vaddr) + queries[i].offset;
  *((u64 *)value_addr) = 0;

  drm_syncobj_replace_fence(queries[i].syncobj, NULL);
 }

 v3d_put_bo_vaddr(bo);
}

static void write_to_buffer_32(u32 *dst, unsigned int idx, u32 value)
{
 dst[idx] = value;
}

static void write_to_buffer_64(u64 *dst, unsigned int idx, u64 value)
{
 dst[idx] = value;
}

static void
write_to_buffer(void *dst, unsigned int idx, bool do_64bit, u64 value)
{
 if (do_64bit)
  write_to_buffer_64(dst, idx, value);
 else
  write_to_buffer_32(dst, idx, value);
}

static void
v3d_copy_query_results(struct v3d_cpu_job *job)
{
 struct v3d_timestamp_query_info *timestamp_query = &job->timestamp_query;
 struct v3d_timestamp_query *queries = timestamp_query->queries;
 struct v3d_bo *bo = to_v3d_bo(job->base.bo[0]);
 struct v3d_bo *timestamp = to_v3d_bo(job->base.bo[1]);
 struct v3d_copy_query_results_info *copy = &job->copy;
 struct dma_fence *fence;
 u8 *query_addr;
 bool available, write_result;
 u8 *data;
 int i;

 v3d_get_bo_vaddr(bo);
 v3d_get_bo_vaddr(timestamp);

 data = ((u8 *)bo->vaddr) + copy->offset;

 for (i = 0; i < timestamp_query->count; i++) {
  fence = drm_syncobj_fence_get(queries[i].syncobj);
  available = fence ? dma_fence_is_signaled(fence) : false;

  write_result = available || copy->do_partial;
  if (write_result) {
   query_addr = ((u8 *)timestamp->vaddr) + queries[i].offset;
   write_to_buffer(data, 0, copy->do_64bit, *((u64 *)query_addr));
  }

  if (copy->availability_bit)
   write_to_buffer(data, 1, copy->do_64bit, available ? 1u : 0u);

  data += copy->stride;

  dma_fence_put(fence);
 }

 v3d_put_bo_vaddr(timestamp);
 v3d_put_bo_vaddr(bo);
}

static void
v3d_reset_performance_queries(struct v3d_cpu_job *job)
{
 struct v3d_performance_query_info *performance_query = &job->performance_query;
 struct v3d_file_priv *v3d_priv = job->base.file->driver_priv;
 struct v3d_dev *v3d = job->base.v3d;
 struct v3d_perfmon *perfmon;

 for (int i = 0; i < performance_query->count; i++) {
  for (int j = 0; j < performance_query->nperfmons; j++) {
   perfmon = v3d_perfmon_find(v3d_priv,
         performance_query->queries[i].kperfmon_ids[j]);
   if (!perfmon) {
    DRM_DEBUG("Failed to find perfmon.");
    continue;
   }

   v3d_perfmon_stop(v3d, perfmon, false);

   memset(perfmon->values, 0, perfmon->ncounters * sizeof(u64));

   v3d_perfmon_put(perfmon);
  }

  drm_syncobj_replace_fence(performance_query->queries[i].syncobj, NULL);
 }
}

static void
v3d_write_performance_query_result(struct v3d_cpu_job *job, void *data,
       unsigned int query)
{
 struct v3d_performance_query_info *performance_query =
      &job->performance_query;
 struct v3d_file_priv *v3d_priv = job->base.file->driver_priv;
 struct v3d_performance_query *perf_query =
   &performance_query->queries[query];
 struct v3d_dev *v3d = job->base.v3d;
 unsigned int i, j, offset;

 for (i = 0, offset = 0;
      i < performance_query->nperfmons;
      i++, offset += DRM_V3D_MAX_PERF_COUNTERS) {
  struct v3d_perfmon *perfmon;

  perfmon = v3d_perfmon_find(v3d_priv,
        perf_query->kperfmon_ids[i]);
  if (!perfmon) {
   DRM_DEBUG("Failed to find perfmon.");
   continue;
  }

  v3d_perfmon_stop(v3d, perfmon, true);

  if (job->copy.do_64bit) {
   for (j = 0; j < perfmon->ncounters; j++)
    write_to_buffer_64(data, offset + j,
         perfmon->values[j]);
  } else {
   for (j = 0; j < perfmon->ncounters; j++)
    write_to_buffer_32(data, offset + j,
         perfmon->values[j]);
  }

  v3d_perfmon_put(perfmon);
 }
}

static void
v3d_copy_performance_query(struct v3d_cpu_job *job)
{
 struct v3d_performance_query_info *performance_query = &job->performance_query;
 struct v3d_copy_query_results_info *copy = &job->copy;
 struct v3d_bo *bo = to_v3d_bo(job->base.bo[0]);
 struct dma_fence *fence;
 bool available, write_result;
 u8 *data;

 v3d_get_bo_vaddr(bo);

 data = ((u8 *)bo->vaddr) + copy->offset;

 for (int i = 0; i < performance_query->count; i++) {
  fence = drm_syncobj_fence_get(performance_query->queries[i].syncobj);
  available = fence ? dma_fence_is_signaled(fence) : false;

  write_result = available || copy->do_partial;
  if (write_result)
   v3d_write_performance_query_result(job, data, i);

  if (copy->availability_bit)
   write_to_buffer(data, performance_query->ncounters,
     copy->do_64bit, available ? 1u : 0u);

  data += copy->stride;

  dma_fence_put(fence);
 }

 v3d_put_bo_vaddr(bo);
}

static const v3d_cpu_job_fn cpu_job_function[] = {
 [V3D_CPU_JOB_TYPE_INDIRECT_CSD] = v3d_rewrite_csd_job_wg_counts_from_indirect,
 [V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY] = v3d_timestamp_query,
 [V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY] = v3d_reset_timestamp_queries,
 [V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY] = v3d_copy_query_results,
 [V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY] = v3d_reset_performance_queries,
 [V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY] = v3d_copy_performance_query,
};

static struct dma_fence *
v3d_cpu_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_cpu_job *job = to_cpu_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;

 if (job->job_type >= ARRAY_SIZE(cpu_job_function)) {
  DRM_DEBUG_DRIVER("Unknown CPU job: %d\n", job->job_type);
  return NULL;
 }

 v3d_job_start_stats(&job->base, V3D_CPU);
 trace_v3d_cpu_job_begin(&v3d->drm, job->job_type);

 cpu_job_function[job->job_type](job);

 trace_v3d_cpu_job_end(&v3d->drm, job->job_type);
 v3d_job_update_stats(&job->base, V3D_CPU);

 return NULL;
}

static struct dma_fence *
v3d_cache_clean_job_run(struct drm_sched_job *sched_job)
{
 struct v3d_job *job = to_v3d_job(sched_job);
 struct v3d_dev *v3d = job->v3d;

 v3d_job_start_stats(job, V3D_CACHE_CLEAN);

 v3d_clean_caches(v3d);

 v3d_job_update_stats(job, V3D_CACHE_CLEAN);

 return NULL;
}

static enum drm_gpu_sched_stat
v3d_gpu_reset_for_timeout(struct v3d_dev *v3d, struct drm_sched_job *sched_job)
{
 enum v3d_queue q;

 mutex_lock(&v3d->reset_lock);

 /* block scheduler */
 for (q = 0; q < V3D_MAX_QUEUES; q++)
  drm_sched_stop(&v3d->queue[q].sched, sched_job);

 if (sched_job)
  drm_sched_increase_karma(sched_job);

 /* get the GPU back into the init state */
 v3d_reset(v3d);

 for (q = 0; q < V3D_MAX_QUEUES; q++)
  drm_sched_resubmit_jobs(&v3d->queue[q].sched);

 /* Unblock schedulers and restart their jobs. */
 for (q = 0; q < V3D_MAX_QUEUES; q++) {
  drm_sched_start(&v3d->queue[q].sched, 0);
 }

 mutex_unlock(&v3d->reset_lock);

 return DRM_GPU_SCHED_STAT_RESET;
}

static enum drm_gpu_sched_stat
v3d_cl_job_timedout(struct drm_sched_job *sched_job, enum v3d_queue q,
      u32 *timedout_ctca, u32 *timedout_ctra)
{
 struct v3d_job *job = to_v3d_job(sched_job);
 struct v3d_dev *v3d = job->v3d;
 u32 ctca = V3D_CORE_READ(0, V3D_CLE_CTNCA(q));
 u32 ctra = V3D_CORE_READ(0, V3D_CLE_CTNRA(q));

 /* If the current address or return address have changed, then the GPU
 * has probably made progress and we should delay the reset. This
 * could fail if the GPU got in an infinite loop in the CL, but that
 * is pretty unlikely outside of an i-g-t testcase.
 */
 if (*timedout_ctca != ctca || *timedout_ctra != ctra) {
  *timedout_ctca = ctca;
  *timedout_ctra = ctra;

  return DRM_GPU_SCHED_STAT_NO_HANG;
 }

 return v3d_gpu_reset_for_timeout(v3d, sched_job);
}

static enum drm_gpu_sched_stat
v3d_bin_job_timedout(struct drm_sched_job *sched_job)
{
 struct v3d_bin_job *job = to_bin_job(sched_job);

 return v3d_cl_job_timedout(sched_job, V3D_BIN,
       &job->timedout_ctca, &job->timedout_ctra);
}

static enum drm_gpu_sched_stat
v3d_render_job_timedout(struct drm_sched_job *sched_job)
{
 struct v3d_render_job *job = to_render_job(sched_job);

 return v3d_cl_job_timedout(sched_job, V3D_RENDER,
       &job->timedout_ctca, &job->timedout_ctra);
}

static enum drm_gpu_sched_stat
v3d_generic_job_timedout(struct drm_sched_job *sched_job)
{
 struct v3d_job *job = to_v3d_job(sched_job);

 return v3d_gpu_reset_for_timeout(job->v3d, sched_job);
}

static enum drm_gpu_sched_stat
v3d_csd_job_timedout(struct drm_sched_job *sched_job)
{
 struct v3d_csd_job *job = to_csd_job(sched_job);
 struct v3d_dev *v3d = job->base.v3d;
 u32 batches = V3D_CORE_READ(0, V3D_CSD_CURRENT_CFG4(v3d->ver));

 /* If we've made progress, skip reset, add the job to the pending
 * list, and let the timer get rearmed.
 */
 if (job->timedout_batches != batches) {
  job->timedout_batches = batches;

  return DRM_GPU_SCHED_STAT_NO_HANG;
 }

 return v3d_gpu_reset_for_timeout(v3d, sched_job);
}

static const struct drm_sched_backend_ops v3d_bin_sched_ops = {
 .run_job = v3d_bin_job_run,
 .timedout_job = v3d_bin_job_timedout,
 .free_job = v3d_sched_job_free,
};

static const struct drm_sched_backend_ops v3d_render_sched_ops = {
 .run_job = v3d_render_job_run,
 .timedout_job = v3d_render_job_timedout,
 .free_job = v3d_sched_job_free,
};

static const struct drm_sched_backend_ops v3d_tfu_sched_ops = {
 .run_job = v3d_tfu_job_run,
 .timedout_job = v3d_generic_job_timedout,
 .free_job = v3d_sched_job_free,
};

static const struct drm_sched_backend_ops v3d_csd_sched_ops = {
 .run_job = v3d_csd_job_run,
 .timedout_job = v3d_csd_job_timedout,
 .free_job = v3d_sched_job_free
};

static const struct drm_sched_backend_ops v3d_cache_clean_sched_ops = {
 .run_job = v3d_cache_clean_job_run,
 .timedout_job = v3d_generic_job_timedout,
 .free_job = v3d_sched_job_free
};

static const struct drm_sched_backend_ops v3d_cpu_sched_ops = {
 .run_job = v3d_cpu_job_run,
 .timedout_job = v3d_generic_job_timedout,
 .free_job = v3d_cpu_job_free
};

static int
v3d_queue_sched_init(struct v3d_dev *v3d, const struct drm_sched_backend_ops *ops,
       enum v3d_queue queue, const char *name)
{
 struct drm_sched_init_args args = {
  .num_rqs = DRM_SCHED_PRIORITY_COUNT,
  .credit_limit = 1,
  .timeout = msecs_to_jiffies(500),
  .dev = v3d->drm.dev,
 };

 args.ops = ops;
 args.name = name;

 return drm_sched_init(&v3d->queue[queue].sched, &args);
}

int
v3d_sched_init(struct v3d_dev *v3d)
{
 int ret;

 ret = v3d_queue_sched_init(v3d, &v3d_bin_sched_ops, V3D_BIN, "v3d_bin");
 if (ret)
  return ret;

 ret = v3d_queue_sched_init(v3d, &v3d_render_sched_ops, V3D_RENDER,
       "v3d_render");
 if (ret)
  goto fail;

 ret = v3d_queue_sched_init(v3d, &v3d_tfu_sched_ops, V3D_TFU, "v3d_tfu");
 if (ret)
  goto fail;

 if (v3d_has_csd(v3d)) {
  ret = v3d_queue_sched_init(v3d, &v3d_csd_sched_ops, V3D_CSD,
        "v3d_csd");
  if (ret)
   goto fail;

  ret = v3d_queue_sched_init(v3d, &v3d_cache_clean_sched_ops,
        V3D_CACHE_CLEAN, "v3d_cache_clean");
  if (ret)
   goto fail;
 }

 ret = v3d_queue_sched_init(v3d, &v3d_cpu_sched_ops, V3D_CPU, "v3d_cpu");
 if (ret)
  goto fail;

 return 0;

fail:
 v3d_sched_fini(v3d);
 return ret;
}

void
v3d_sched_fini(struct v3d_dev *v3d)
{
 enum v3d_queue q;

 for (q = 0; q < V3D_MAX_QUEUES; q++) {
  if (v3d->queue[q].sched.ready)
   drm_sched_fini(&v3d->queue[q].sched);
 }
}