Quelle  v3d_submit.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0+
/*
 * Copyright (C) 2014-2018 Broadcom
 * Copyright (C) 2023 Raspberry Pi
 */

#include <drm/drm_syncobj.h>

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

/* Takes the reservation lock on all the BOs being referenced, so that
 * we can attach fences and update the reservations after pushing the job
 * to the queue.
 *
 * We don't lock the RCL the tile alloc/state BOs, or overflow memory
 * (all of which are on render->unref_list). They're entirely private
 * to v3d, so we don't attach dma-buf fences to them.
 */
static int
v3d_lock_bo_reservations(struct v3d_job *job,
    struct ww_acquire_ctx *acquire_ctx)
{
 int i, ret;

 ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
 if (ret)
  return ret;

 for (i = 0; i < job->bo_count; i++) {
  ret = dma_resv_reserve_fences(job->bo[i]->resv, 1);
  if (ret)
   goto fail;

  ret = drm_sched_job_add_implicit_dependencies(&job->base,
             job->bo[i], true);
  if (ret)
   goto fail;
 }

 return 0;

fail:
 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
 return ret;
}

/**
 * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
 * referenced by the job.
 * @dev: DRM device
 * @file_priv: DRM file for this fd
 * @job: V3D job being set up
 * @bo_handles: GEM handles
 * @bo_count: Number of GEM handles passed in
 *
 * The command validator needs to reference BOs by their index within
 * the submitted job's BO list. This does the validation of the job's
 * BO list and reference counting for the lifetime of the job.
 *
 * Note that this function doesn't need to unreference the BOs on
 * failure, because that will happen at `v3d_job_free()`.
 */
static int
v3d_lookup_bos(struct drm_device *dev,
        struct drm_file *file_priv,
        struct v3d_job *job,
        u64 bo_handles,
        u32 bo_count)
{
 job->bo_count = bo_count;

 if (!job->bo_count) {
  /* See comment on bo_index for why we have to check
 * this.
 */
  DRM_DEBUG("Rendering requires BOs\n");
  return -EINVAL;
 }

 return drm_gem_objects_lookup(file_priv,
          (void __user *)(uintptr_t)bo_handles,
          job->bo_count, &job->bo);
}

static void
v3d_job_free(struct kref *ref)
{
 struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
 int i;

 if (job->bo) {
  for (i = 0; i < job->bo_count; i++)
   drm_gem_object_put(job->bo[i]);
  kvfree(job->bo);
 }

 dma_fence_put(job->irq_fence);
 dma_fence_put(job->done_fence);

 if (job->perfmon)
  v3d_perfmon_put(job->perfmon);

 kfree(job);
}

static void
v3d_render_job_free(struct kref *ref)
{
 struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
        base.refcount);
 struct v3d_bo *bo, *save;

 list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
  drm_gem_object_put(&bo->base.base);
 }

 v3d_job_free(ref);
}

void v3d_job_cleanup(struct v3d_job *job)
{
 if (!job)
  return;

 drm_sched_job_cleanup(&job->base);
 v3d_job_put(job);
}

void v3d_job_put(struct v3d_job *job)
{
 if (!job)
  return;

 kref_put(&job->refcount, job->free);
}

static int
v3d_job_allocate(void **container, size_t size)
{
 *container = kcalloc(1, size, GFP_KERNEL);
 if (!*container) {
  DRM_ERROR("Cannot allocate memory for V3D job.\n");
  return -ENOMEM;
 }

 return 0;
}

static void
v3d_job_deallocate(void **container)
{
 kfree(*container);
 *container = NULL;
}

static int
v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
      struct v3d_job *job, void (*free)(struct kref *ref),
      u32 in_sync, struct v3d_submit_ext *se, enum v3d_queue queue)
{
 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
 int ret, i;

 job->v3d = v3d;
 job->free = free;
 job->file = file_priv;

 ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
     1, v3d_priv, file_priv->client_id);
 if (ret)
  return ret;

 if (has_multisync) {
  if (se->in_sync_count && se->wait_stage == queue) {
   struct drm_v3d_sem __user *handle = u64_to_user_ptr(se->in_syncs);

   for (i = 0; i < se->in_sync_count; i++) {
    struct drm_v3d_sem in;

    if (copy_from_user(&in, handle++, sizeof(in))) {
     ret = -EFAULT;
     DRM_DEBUG("Failed to copy wait dep handle.\n");
     goto fail_deps;
    }
    ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in.handle, 0);

    // TODO: Investigate why this was filtered out for the IOCTL.
    if (ret && ret != -ENOENT)
     goto fail_deps;
   }
  }
 } else {
  ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in_sync, 0);

  // TODO: Investigate why this was filtered out for the IOCTL.
  if (ret && ret != -ENOENT)
   goto fail_deps;
 }

 kref_init(&job->refcount);

 return 0;

fail_deps:
 drm_sched_job_cleanup(&job->base);
 return ret;
}

static void
v3d_push_job(struct v3d_job *job)
{
 drm_sched_job_arm(&job->base);

 job->done_fence = dma_fence_get(&job->base.s_fence->finished);

 /* put by scheduler job completion */
 kref_get(&job->refcount);

 drm_sched_entity_push_job(&job->base);
}

static void
v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
      struct v3d_job *job,
      struct ww_acquire_ctx *acquire_ctx,
      u32 out_sync,
      struct v3d_submit_ext *se,
      struct dma_fence *done_fence)
{
 struct drm_syncobj *sync_out;
 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
 int i;

 for (i = 0; i < job->bo_count; i++) {
  /* XXX: Use shared fences for read-only objects. */
  dma_resv_add_fence(job->bo[i]->resv, job->done_fence,
       DMA_RESV_USAGE_WRITE);
 }

 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);

 /* Update the return sync object for the job */
 /* If it only supports a single signal semaphore*/
 if (!has_multisync) {
  sync_out = drm_syncobj_find(file_priv, out_sync);
  if (sync_out) {
   drm_syncobj_replace_fence(sync_out, done_fence);
   drm_syncobj_put(sync_out);
  }
  return;
 }

 /* If multiple semaphores extension is supported */
 if (se->out_sync_count) {
  for (i = 0; i < se->out_sync_count; i++) {
   drm_syncobj_replace_fence(se->out_syncs[i].syncobj,
        done_fence);
   drm_syncobj_put(se->out_syncs[i].syncobj);
  }
  kvfree(se->out_syncs);
 }
}

static int
v3d_setup_csd_jobs_and_bos(struct drm_file *file_priv,
      struct v3d_dev *v3d,
      struct drm_v3d_submit_csd *args,
      struct v3d_csd_job **job,
      struct v3d_job **clean_job,
      struct v3d_submit_ext *se,
      struct ww_acquire_ctx *acquire_ctx)
{
 int ret;

 ret = v3d_job_allocate((void *)job, sizeof(**job));
 if (ret)
  return ret;

 ret = v3d_job_init(v3d, file_priv, &(*job)->base,
      v3d_job_free, args->in_sync, se, V3D_CSD);
 if (ret) {
  v3d_job_deallocate((void *)job);
  return ret;
 }

 ret = v3d_job_allocate((void *)clean_job, sizeof(**clean_job));
 if (ret)
  return ret;

 ret = v3d_job_init(v3d, file_priv, *clean_job,
      v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
 if (ret) {
  v3d_job_deallocate((void *)clean_job);
  return ret;
 }

 (*job)->args = *args;

 ret = v3d_lookup_bos(&v3d->drm, file_priv, *clean_job,
        args->bo_handles, args->bo_handle_count);
 if (ret)
  return ret;

 return v3d_lock_bo_reservations(*clean_job, acquire_ctx);
}

static void
v3d_put_multisync_post_deps(struct v3d_submit_ext *se)
{
 unsigned int i;

 if (!(se && se->out_sync_count))
  return;

 for (i = 0; i < se->out_sync_count; i++)
  drm_syncobj_put(se->out_syncs[i].syncobj);
 kvfree(se->out_syncs);
}

static int
v3d_get_multisync_post_deps(struct drm_file *file_priv,
       struct v3d_submit_ext *se,
       u32 count, u64 handles)
{
 struct drm_v3d_sem __user *post_deps;
 int i, ret;

 if (!count)
  return 0;

 se->out_syncs = (struct v3d_submit_outsync *)
   kvmalloc_array(count,
           sizeof(struct v3d_submit_outsync),
           GFP_KERNEL);
 if (!se->out_syncs)
  return -ENOMEM;

 post_deps = u64_to_user_ptr(handles);

 for (i = 0; i < count; i++) {
  struct drm_v3d_sem out;

  if (copy_from_user(&out, post_deps++, sizeof(out))) {
   ret = -EFAULT;
   DRM_DEBUG("Failed to copy post dep handles\n");
   goto fail;
  }

  se->out_syncs[i].syncobj = drm_syncobj_find(file_priv,
           out.handle);
  if (!se->out_syncs[i].syncobj) {
   ret = -EINVAL;
   goto fail;
  }
 }
 se->out_sync_count = count;

 return 0;

fail:
 for (i--; i >= 0; i--)
  drm_syncobj_put(se->out_syncs[i].syncobj);
 kvfree(se->out_syncs);

 return ret;
}

/* Get data for multiple binary semaphores synchronization. Parse syncobj
 * to be signaled when job completes (out_sync).
 */
static int
v3d_get_multisync_submit_deps(struct drm_file *file_priv,
         struct drm_v3d_extension __user *ext,
         struct v3d_submit_ext *se)
{
 struct drm_v3d_multi_sync multisync;
 int ret;

 if (se->in_sync_count || se->out_sync_count) {
  DRM_DEBUG("Two multisync extensions were added to the same job.");
  return -EINVAL;
 }

 if (copy_from_user(&multisync, ext, sizeof(multisync)))
  return -EFAULT;

 if (multisync.pad)
  return -EINVAL;

 ret = v3d_get_multisync_post_deps(file_priv, se, multisync.out_sync_count,
       multisync.out_syncs);
 if (ret)
  return ret;

 se->in_sync_count = multisync.in_sync_count;
 se->in_syncs = multisync.in_syncs;
 se->flags |= DRM_V3D_EXT_ID_MULTI_SYNC;
 se->wait_stage = multisync.wait_stage;

 return 0;
}

/* Get data for the indirect CSD job submission. */
static int
v3d_get_cpu_indirect_csd_params(struct drm_file *file_priv,
    struct drm_v3d_extension __user *ext,
    struct v3d_cpu_job *job)
{
 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 struct v3d_dev *v3d = v3d_priv->v3d;
 struct drm_v3d_indirect_csd indirect_csd;
 struct v3d_indirect_csd_info *info = &job->indirect_csd;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(&indirect_csd, ext, sizeof(indirect_csd)))
  return -EFAULT;

 if (!v3d_has_csd(v3d)) {
  DRM_DEBUG("Attempting CSD submit on non-CSD hardware.\n");
  return -EINVAL;
 }

 job->job_type = V3D_CPU_JOB_TYPE_INDIRECT_CSD;
 info->offset = indirect_csd.offset;
 info->wg_size = indirect_csd.wg_size;
 memcpy(&info->wg_uniform_offsets, &indirect_csd.wg_uniform_offsets,
        sizeof(indirect_csd.wg_uniform_offsets));

 info->indirect = drm_gem_object_lookup(file_priv, indirect_csd.indirect);

 return v3d_setup_csd_jobs_and_bos(file_priv, v3d, &indirect_csd.submit,
       &info->job, &info->clean_job,
       NULL, &info->acquire_ctx);
}

/* Get data for the query timestamp job submission. */
static int
v3d_get_cpu_timestamp_query_params(struct drm_file *file_priv,
       struct drm_v3d_extension __user *ext,
       struct v3d_cpu_job *job)
{
 u32 __user *offsets, *syncs;
 struct drm_v3d_timestamp_query timestamp;
 struct v3d_timestamp_query_info *query_info = &job->timestamp_query;
 unsigned int i;
 int err;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(×tamp, ext, sizeof(timestamp)))
  return -EFAULT;

 if (timestamp.pad)
  return -EINVAL;

 job->job_type = V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY;

 query_info->queries = kvmalloc_array(timestamp.count,
          sizeof(struct v3d_timestamp_query),
          GFP_KERNEL);
 if (!query_info->queries)
  return -ENOMEM;

 offsets = u64_to_user_ptr(timestamp.offsets);
 syncs = u64_to_user_ptr(timestamp.syncs);

 for (i = 0; i < timestamp.count; i++) {
  u32 offset, sync;

  if (get_user(offset, offsets++)) {
   err = -EFAULT;
   goto error;
  }

  query_info->queries[i].offset = offset;

  if (get_user(sync, syncs++)) {
   err = -EFAULT;
   goto error;
  }

  query_info->queries[i].syncobj = drm_syncobj_find(file_priv,
          sync);
  if (!query_info->queries[i].syncobj) {
   err = -ENOENT;
   goto error;
  }
 }
 query_info->count = timestamp.count;

 return 0;

error:
 v3d_timestamp_query_info_free(&job->timestamp_query, i);
 return err;
}

static int
v3d_get_cpu_reset_timestamp_params(struct drm_file *file_priv,
       struct drm_v3d_extension __user *ext,
       struct v3d_cpu_job *job)
{
 u32 __user *syncs;
 struct drm_v3d_reset_timestamp_query reset;
 struct v3d_timestamp_query_info *query_info = &job->timestamp_query;
 unsigned int i;
 int err;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(&reset, ext, sizeof(reset)))
  return -EFAULT;

 job->job_type = V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY;

 query_info->queries = kvmalloc_array(reset.count,
          sizeof(struct v3d_timestamp_query),
          GFP_KERNEL);
 if (!query_info->queries)
  return -ENOMEM;

 syncs = u64_to_user_ptr(reset.syncs);

 for (i = 0; i < reset.count; i++) {
  u32 sync;

  query_info->queries[i].offset = reset.offset + 8 * i;

  if (get_user(sync, syncs++)) {
   err = -EFAULT;
   goto error;
  }

  query_info->queries[i].syncobj = drm_syncobj_find(file_priv,
          sync);
  if (!query_info->queries[i].syncobj) {
   err = -ENOENT;
   goto error;
  }
 }
 query_info->count = reset.count;

 return 0;

error:
 v3d_timestamp_query_info_free(&job->timestamp_query, i);
 return err;
}

/* Get data for the copy timestamp query results job submission. */
static int
v3d_get_cpu_copy_query_results_params(struct drm_file *file_priv,
          struct drm_v3d_extension __user *ext,
          struct v3d_cpu_job *job)
{
 u32 __user *offsets, *syncs;
 struct drm_v3d_copy_timestamp_query copy;
 struct v3d_timestamp_query_info *query_info = &job->timestamp_query;
 unsigned int i;
 int err;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(©, ext, sizeof(copy)))
  return -EFAULT;

 if (copy.pad)
  return -EINVAL;

 job->job_type = V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY;

 query_info->queries = kvmalloc_array(copy.count,
          sizeof(struct v3d_timestamp_query),
          GFP_KERNEL);
 if (!query_info->queries)
  return -ENOMEM;

 offsets = u64_to_user_ptr(copy.offsets);
 syncs = u64_to_user_ptr(copy.syncs);

 for (i = 0; i < copy.count; i++) {
  u32 offset, sync;

  if (get_user(offset, offsets++)) {
   err = -EFAULT;
   goto error;
  }

  query_info->queries[i].offset = offset;

  if (get_user(sync, syncs++)) {
   err = -EFAULT;
   goto error;
  }

  query_info->queries[i].syncobj = drm_syncobj_find(file_priv,
          sync);
  if (!query_info->queries[i].syncobj) {
   err = -ENOENT;
   goto error;
  }
 }
 query_info->count = copy.count;

 job->copy.do_64bit = copy.do_64bit;
 job->copy.do_partial = copy.do_partial;
 job->copy.availability_bit = copy.availability_bit;
 job->copy.offset = copy.offset;
 job->copy.stride = copy.stride;

 return 0;

error:
 v3d_timestamp_query_info_free(&job->timestamp_query, i);
 return err;
}

static int
v3d_copy_query_info(struct v3d_performance_query_info *query_info,
      unsigned int count,
      unsigned int nperfmons,
      u32 __user *syncs,
      u64 __user *kperfmon_ids,
      struct drm_file *file_priv)
{
 unsigned int i, j;
 int err;

 for (i = 0; i < count; i++) {
  struct v3d_performance_query *query = &query_info->queries[i];
  u32 __user *ids_pointer;
  u32 sync, id;
  u64 ids;

  if (get_user(sync, syncs++)) {
   err = -EFAULT;
   goto error;
  }

  if (get_user(ids, kperfmon_ids++)) {
   err = -EFAULT;
   goto error;
  }

  query->kperfmon_ids =
   kvmalloc_array(nperfmons,
           sizeof(struct v3d_performance_query *),
           GFP_KERNEL);
  if (!query->kperfmon_ids) {
   err = -ENOMEM;
   goto error;
  }

  ids_pointer = u64_to_user_ptr(ids);

  for (j = 0; j < nperfmons; j++) {
   if (get_user(id, ids_pointer++)) {
    kvfree(query->kperfmon_ids);
    err = -EFAULT;
    goto error;
   }

   query->kperfmon_ids[j] = id;
  }

  query->syncobj = drm_syncobj_find(file_priv, sync);
  if (!query->syncobj) {
   kvfree(query->kperfmon_ids);
   err = -ENOENT;
   goto error;
  }
 }

 return 0;

error:
 v3d_performance_query_info_free(query_info, i);
 return err;
}

static int
v3d_get_cpu_reset_performance_params(struct drm_file *file_priv,
         struct drm_v3d_extension __user *ext,
         struct v3d_cpu_job *job)
{
 struct v3d_performance_query_info *query_info = &job->performance_query;
 struct drm_v3d_reset_performance_query reset;
 int err;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(&reset, ext, sizeof(reset)))
  return -EFAULT;

 job->job_type = V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY;

 query_info->queries =
  kvmalloc_array(reset.count,
          sizeof(struct v3d_performance_query),
          GFP_KERNEL);
 if (!query_info->queries)
  return -ENOMEM;

 err = v3d_copy_query_info(query_info,
      reset.count,
      reset.nperfmons,
      u64_to_user_ptr(reset.syncs),
      u64_to_user_ptr(reset.kperfmon_ids),
      file_priv);
 if (err)
  return err;

 query_info->count = reset.count;
 query_info->nperfmons = reset.nperfmons;

 return 0;
}

static int
v3d_get_cpu_copy_performance_query_params(struct drm_file *file_priv,
       struct drm_v3d_extension __user *ext,
       struct v3d_cpu_job *job)
{
 struct v3d_performance_query_info *query_info = &job->performance_query;
 struct drm_v3d_copy_performance_query copy;
 int err;

 if (!job) {
  DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
  return -EINVAL;
 }

 if (job->job_type) {
  DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
  return -EINVAL;
 }

 if (copy_from_user(©, ext, sizeof(copy)))
  return -EFAULT;

 if (copy.pad)
  return -EINVAL;

 job->job_type = V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY;

 query_info->queries =
  kvmalloc_array(copy.count,
          sizeof(struct v3d_performance_query),
          GFP_KERNEL);
 if (!query_info->queries)
  return -ENOMEM;

 err = v3d_copy_query_info(query_info,
      copy.count,
      copy.nperfmons,
      u64_to_user_ptr(copy.syncs),
      u64_to_user_ptr(copy.kperfmon_ids),
      file_priv);
 if (err)
  return err;

 query_info->count = copy.count;
 query_info->nperfmons = copy.nperfmons;
 query_info->ncounters = copy.ncounters;

 job->copy.do_64bit = copy.do_64bit;
 job->copy.do_partial = copy.do_partial;
 job->copy.availability_bit = copy.availability_bit;
 job->copy.offset = copy.offset;
 job->copy.stride = copy.stride;

 return 0;
}

/* Whenever userspace sets ioctl extensions, v3d_get_extensions parses data
 * according to the extension id (name).
 */
static int
v3d_get_extensions(struct drm_file *file_priv,
     u64 ext_handles,
     struct v3d_submit_ext *se,
     struct v3d_cpu_job *job)
{
 struct drm_v3d_extension __user *user_ext;
 int ret;

 user_ext = u64_to_user_ptr(ext_handles);
 while (user_ext) {
  struct drm_v3d_extension ext;

  if (copy_from_user(&ext, user_ext, sizeof(ext))) {
   DRM_DEBUG("Failed to copy submit extension\n");
   return -EFAULT;
  }

  switch (ext.id) {
  case DRM_V3D_EXT_ID_MULTI_SYNC:
   ret = v3d_get_multisync_submit_deps(file_priv, user_ext, se);
   break;
  case DRM_V3D_EXT_ID_CPU_INDIRECT_CSD:
   ret = v3d_get_cpu_indirect_csd_params(file_priv, user_ext, job);
   break;
  case DRM_V3D_EXT_ID_CPU_TIMESTAMP_QUERY:
   ret = v3d_get_cpu_timestamp_query_params(file_priv, user_ext, job);
   break;
  case DRM_V3D_EXT_ID_CPU_RESET_TIMESTAMP_QUERY:
   ret = v3d_get_cpu_reset_timestamp_params(file_priv, user_ext, job);
   break;
  case DRM_V3D_EXT_ID_CPU_COPY_TIMESTAMP_QUERY:
   ret = v3d_get_cpu_copy_query_results_params(file_priv, user_ext, job);
   break;
  case DRM_V3D_EXT_ID_CPU_RESET_PERFORMANCE_QUERY:
   ret = v3d_get_cpu_reset_performance_params(file_priv, user_ext, job);
   break;
  case DRM_V3D_EXT_ID_CPU_COPY_PERFORMANCE_QUERY:
   ret = v3d_get_cpu_copy_performance_query_params(file_priv, user_ext, job);
   break;
  default:
   DRM_DEBUG_DRIVER("Unknown extension id: %d\n", ext.id);
   return -EINVAL;
  }

  if (ret)
   return ret;

  user_ext = u64_to_user_ptr(ext.next);
 }

 return 0;
}

/**
 * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
 * @dev: DRM device
 * @data: ioctl argument
 * @file_priv: DRM file for this fd
 *
 * This is the main entrypoint for userspace to submit a 3D frame to
 * the GPU.  Userspace provides the binner command list (if
 * applicable), and the kernel sets up the render command list to draw
 * to the framebuffer described in the ioctl, using the command lists
 * that the 3D engine's binner will produce.
 */
int
v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
      struct drm_file *file_priv)
{
 struct v3d_dev *v3d = to_v3d_dev(dev);
 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 struct drm_v3d_submit_cl *args = data;
 struct v3d_submit_ext se = {0};
 struct v3d_bin_job *bin = NULL;
 struct v3d_render_job *render = NULL;
 struct v3d_job *clean_job = NULL;
 struct v3d_job *last_job;
 struct ww_acquire_ctx acquire_ctx;
 int ret = 0;

 trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);

 if (args->pad)
  return -EINVAL;

 if (args->flags &&
     args->flags & ~(DRM_V3D_SUBMIT_CL_FLUSH_CACHE |
       DRM_V3D_SUBMIT_EXTENSION)) {
  DRM_INFO("invalid flags: %d\n", args->flags);
  return -EINVAL;
 }

 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
  ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
  if (ret) {
   DRM_DEBUG("Failed to get extensions.\n");
   return ret;
  }
 }

 ret = v3d_job_allocate((void *)&render, sizeof(*render));
 if (ret)
  return ret;

 ret = v3d_job_init(v3d, file_priv, &render->base,
      v3d_render_job_free, args->in_sync_rcl, &se, V3D_RENDER);
 if (ret) {
  v3d_job_deallocate((void *)&render);
  goto fail;
 }

 render->start = args->rcl_start;
 render->end = args->rcl_end;
 INIT_LIST_HEAD(&render->unref_list);

 if (args->bcl_start != args->bcl_end) {
  ret = v3d_job_allocate((void *)&bin, sizeof(*bin));
  if (ret)
   goto fail;

  ret = v3d_job_init(v3d, file_priv, &bin->base,
       v3d_job_free, args->in_sync_bcl, &se, V3D_BIN);
  if (ret) {
   v3d_job_deallocate((void *)&bin);
   goto fail;
  }

  bin->start = args->bcl_start;
  bin->end = args->bcl_end;
  bin->qma = args->qma;
  bin->qms = args->qms;
  bin->qts = args->qts;
  bin->render = render;
 }

 if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) {
  ret = v3d_job_allocate((void *)&clean_job, sizeof(*clean_job));
  if (ret)
   goto fail;

  ret = v3d_job_init(v3d, file_priv, clean_job,
       v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
  if (ret) {
   v3d_job_deallocate((void *)&clean_job);
   goto fail;
  }

  last_job = clean_job;
 } else {
  last_job = &render->base;
 }

 ret = v3d_lookup_bos(dev, file_priv, last_job,
        args->bo_handles, args->bo_handle_count);
 if (ret)
  goto fail;

 ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
 if (ret)
  goto fail;

 if (args->perfmon_id) {
  if (v3d->global_perfmon) {
   ret = -EAGAIN;
   goto fail_perfmon;
  }

  render->base.perfmon = v3d_perfmon_find(v3d_priv,
       args->perfmon_id);

  if (!render->base.perfmon) {
   ret = -ENOENT;
   goto fail_perfmon;
  }
 }

 mutex_lock(&v3d->sched_lock);
 if (bin) {
  bin->base.perfmon = render->base.perfmon;
  v3d_perfmon_get(bin->base.perfmon);
  v3d_push_job(&bin->base);

  ret = drm_sched_job_add_dependency(&render->base.base,
         dma_fence_get(bin->base.done_fence));
  if (ret)
   goto fail_unreserve;
 }

 v3d_push_job(&render->base);

 if (clean_job) {
  struct dma_fence *render_fence =
   dma_fence_get(render->base.done_fence);
  ret = drm_sched_job_add_dependency(&clean_job->base,
         render_fence);
  if (ret)
   goto fail_unreserve;
  clean_job->perfmon = render->base.perfmon;
  v3d_perfmon_get(clean_job->perfmon);
  v3d_push_job(clean_job);
 }

 mutex_unlock(&v3d->sched_lock);

 v3d_attach_fences_and_unlock_reservation(file_priv,
       last_job,
       &acquire_ctx,
       args->out_sync,
       &se,
       last_job->done_fence);

 v3d_job_put(&bin->base);
 v3d_job_put(&render->base);
 v3d_job_put(clean_job);

 return 0;

fail_unreserve:
 mutex_unlock(&v3d->sched_lock);
fail_perfmon:
 drm_gem_unlock_reservations(last_job->bo,
        last_job->bo_count, &acquire_ctx);
fail:
 v3d_job_cleanup((void *)bin);
 v3d_job_cleanup((void *)render);
 v3d_job_cleanup(clean_job);
 v3d_put_multisync_post_deps(&se);

 return ret;
}

/**
 * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D.
 * @dev: DRM device
 * @data: ioctl argument
 * @file_priv: DRM file for this fd
 *
 * Userspace provides the register setup for the TFU, which we don't
 * need to validate since the TFU is behind the MMU.
 */
int
v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
       struct drm_file *file_priv)
{
 struct v3d_dev *v3d = to_v3d_dev(dev);
 struct drm_v3d_submit_tfu *args = data;
 struct v3d_submit_ext se = {0};
 struct v3d_tfu_job *job = NULL;
 struct ww_acquire_ctx acquire_ctx;
 int ret = 0;

 trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);

 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
  DRM_DEBUG("invalid flags: %d\n", args->flags);
  return -EINVAL;
 }

 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
  ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
  if (ret) {
   DRM_DEBUG("Failed to get extensions.\n");
   return ret;
  }
 }

 ret = v3d_job_allocate((void *)&job, sizeof(*job));
 if (ret)
  return ret;

 ret = v3d_job_init(v3d, file_priv, &job->base,
      v3d_job_free, args->in_sync, &se, V3D_TFU);
 if (ret) {
  v3d_job_deallocate((void *)&job);
  goto fail;
 }

 job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
          sizeof(*job->base.bo), GFP_KERNEL);
 if (!job->base.bo) {
  ret = -ENOMEM;
  goto fail;
 }

 job->args = *args;

 for (job->base.bo_count = 0;
      job->base.bo_count < ARRAY_SIZE(args->bo_handles);
      job->base.bo_count++) {
  struct drm_gem_object *bo;

  if (!args->bo_handles[job->base.bo_count])
   break;

  bo = drm_gem_object_lookup(file_priv, args->bo_handles[job->base.bo_count]);
  if (!bo) {
   DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
      job->base.bo_count,
      args->bo_handles[job->base.bo_count]);
   ret = -ENOENT;
   goto fail;
  }
  job->base.bo[job->base.bo_count] = bo;
 }

 ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
 if (ret)
  goto fail;

 mutex_lock(&v3d->sched_lock);
 v3d_push_job(&job->base);
 mutex_unlock(&v3d->sched_lock);

 v3d_attach_fences_and_unlock_reservation(file_priv,
       &job->base, &acquire_ctx,
       args->out_sync,
       &se,
       job->base.done_fence);

 v3d_job_put(&job->base);

 return 0;

fail:
 v3d_job_cleanup((void *)job);
 v3d_put_multisync_post_deps(&se);

 return ret;
}

/**
 * v3d_submit_csd_ioctl() - Submits a CSD (compute shader) job to the V3D.
 * @dev: DRM device
 * @data: ioctl argument
 * @file_priv: DRM file for this fd
 *
 * Userspace provides the register setup for the CSD, which we don't
 * need to validate since the CSD is behind the MMU.
 */
int
v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
       struct drm_file *file_priv)
{
 struct v3d_dev *v3d = to_v3d_dev(dev);
 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 struct drm_v3d_submit_csd *args = data;
 struct v3d_submit_ext se = {0};
 struct v3d_csd_job *job = NULL;
 struct v3d_job *clean_job = NULL;
 struct ww_acquire_ctx acquire_ctx;
 int ret;

 trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);

 if (args->pad)
  return -EINVAL;

 if (!v3d_has_csd(v3d)) {
  DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
  return -EINVAL;
 }

 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
  DRM_INFO("invalid flags: %d\n", args->flags);
  return -EINVAL;
 }

 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
  ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
  if (ret) {
   DRM_DEBUG("Failed to get extensions.\n");
   return ret;
  }
 }

 ret = v3d_setup_csd_jobs_and_bos(file_priv, v3d, args,
      &job, &clean_job, &se,
      &acquire_ctx);
 if (ret)
  goto fail;

 if (args->perfmon_id) {
  if (v3d->global_perfmon) {
   ret = -EAGAIN;
   goto fail_perfmon;
  }

  job->base.perfmon = v3d_perfmon_find(v3d_priv,
           args->perfmon_id);
  if (!job->base.perfmon) {
   ret = -ENOENT;
   goto fail_perfmon;
  }
 }

 mutex_lock(&v3d->sched_lock);
 v3d_push_job(&job->base);

 ret = drm_sched_job_add_dependency(&clean_job->base,
        dma_fence_get(job->base.done_fence));
 if (ret)
  goto fail_unreserve;

 v3d_push_job(clean_job);
 mutex_unlock(&v3d->sched_lock);

 v3d_attach_fences_and_unlock_reservation(file_priv,
       clean_job,
       &acquire_ctx,
       args->out_sync,
       &se,
       clean_job->done_fence);

 v3d_job_put(&job->base);
 v3d_job_put(clean_job);

 return 0;

fail_unreserve:
 mutex_unlock(&v3d->sched_lock);
fail_perfmon:
 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
        &acquire_ctx);
fail:
 v3d_job_cleanup((void *)job);
 v3d_job_cleanup(clean_job);
 v3d_put_multisync_post_deps(&se);

 return ret;
}

static const unsigned int cpu_job_bo_handle_count[] = {
 [V3D_CPU_JOB_TYPE_INDIRECT_CSD] = 1,
 [V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY] = 1,
 [V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY] = 1,
 [V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY] = 2,
 [V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY] = 0,
 [V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY] = 1,
};

/**
 * v3d_submit_cpu_ioctl() - Submits a CPU job to the V3D.
 * @dev: DRM device
 * @data: ioctl argument
 * @file_priv: DRM file for this fd
 *
 * Userspace specifies the CPU job type and data required to perform its
 * operations through the drm_v3d_extension struct.
 */
int
v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
       struct drm_file *file_priv)
{
 struct v3d_dev *v3d = to_v3d_dev(dev);
 struct drm_v3d_submit_cpu *args = data;
 struct v3d_submit_ext se = {0};
 struct v3d_submit_ext *out_se = NULL;
 struct v3d_cpu_job *cpu_job = NULL;
 struct v3d_csd_job *csd_job = NULL;
 struct v3d_job *clean_job = NULL;
 struct ww_acquire_ctx acquire_ctx;
 int ret;

 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
  DRM_INFO("Invalid flags: %d\n", args->flags);
  return -EINVAL;
 }

 ret = v3d_job_allocate((void *)&cpu_job, sizeof(*cpu_job));
 if (ret)
  return ret;

 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
  ret = v3d_get_extensions(file_priv, args->extensions, &se, cpu_job);
  if (ret) {
   DRM_DEBUG("Failed to get extensions.\n");
   goto fail;
  }
 }

 /* Every CPU job must have a CPU job user extension */
 if (!cpu_job->job_type) {
  DRM_DEBUG("CPU job must have a CPU job user extension.\n");
  ret = -EINVAL;
  goto fail;
 }

 if (args->bo_handle_count != cpu_job_bo_handle_count[cpu_job->job_type]) {
  DRM_DEBUG("This CPU job was not submitted with the proper number of BOs.\n");
  ret = -EINVAL;
  goto fail;
 }

 trace_v3d_submit_cpu_ioctl(&v3d->drm, cpu_job->job_type);

 ret = v3d_job_init(v3d, file_priv, &cpu_job->base,
      v3d_job_free, 0, &se, V3D_CPU);
 if (ret) {
  v3d_job_deallocate((void *)&cpu_job);
  goto fail;
 }

 clean_job = cpu_job->indirect_csd.clean_job;
 csd_job = cpu_job->indirect_csd.job;

 if (args->bo_handle_count) {
  ret = v3d_lookup_bos(dev, file_priv, &cpu_job->base,
         args->bo_handles, args->bo_handle_count);
  if (ret)
   goto fail;

  ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx);
  if (ret)
   goto fail;
 }

 mutex_lock(&v3d->sched_lock);
 v3d_push_job(&cpu_job->base);

 switch (cpu_job->job_type) {
 case V3D_CPU_JOB_TYPE_INDIRECT_CSD:
  ret = drm_sched_job_add_dependency(&csd_job->base.base,
         dma_fence_get(cpu_job->base.done_fence));
  if (ret)
   goto fail_unreserve;

  v3d_push_job(&csd_job->base);

  ret = drm_sched_job_add_dependency(&clean_job->base,
         dma_fence_get(csd_job->base.done_fence));
  if (ret)
   goto fail_unreserve;

  v3d_push_job(clean_job);

  break;
 default:
  break;
 }
 mutex_unlock(&v3d->sched_lock);

 out_se = (cpu_job->job_type == V3D_CPU_JOB_TYPE_INDIRECT_CSD) ? NULL : &se;

 v3d_attach_fences_and_unlock_reservation(file_priv,
       &cpu_job->base,
       &acquire_ctx, 0,
       out_se, cpu_job->base.done_fence);

 switch (cpu_job->job_type) {
 case V3D_CPU_JOB_TYPE_INDIRECT_CSD:
  v3d_attach_fences_and_unlock_reservation(file_priv,
        clean_job,
        &cpu_job->indirect_csd.acquire_ctx,
        0, &se, clean_job->done_fence);
  break;
 default:
  break;
 }

 v3d_job_put(&cpu_job->base);
 v3d_job_put(&csd_job->base);
 v3d_job_put(clean_job);

 return 0;

fail_unreserve:
 mutex_unlock(&v3d->sched_lock);

 drm_gem_unlock_reservations(cpu_job->base.bo, cpu_job->base.bo_count,
        &acquire_ctx);

 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
        &cpu_job->indirect_csd.acquire_ctx);

fail:
 v3d_job_cleanup((void *)cpu_job);
 v3d_job_cleanup((void *)csd_job);
 v3d_job_cleanup(clean_job);
 v3d_put_multisync_post_deps(&se);
 kvfree(cpu_job->timestamp_query.queries);
 kvfree(cpu_job->performance_query.queries);

 return ret;
}