drivers/gpu/drm/lima/lima_sched.c

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /* Copyright 2017-2019 Qiang Yu <yuq825@gmail.com> */
   3
   4 #include <linux/kthread.h>
   5 #include <linux/slab.h>
   6 #include <linux/xarray.h>
   7
   8 #include "lima_drv.h"
   9 #include "lima_sched.h"
  10 #include "lima_vm.h"
  11 #include "lima_mmu.h"
  12 #include "lima_l2_cache.h"
  13 #include "lima_gem.h"
  14
  15 struct lima_fence {
  16         struct dma_fence base;
  17         struct lima_sched_pipe *pipe;
  18 };
  19
  20 static struct kmem_cache *lima_fence_slab;
  21 static int lima_fence_slab_refcnt;
  22
  23 int lima_sched_slab_init(void)
  24 {
  25         if (!lima_fence_slab) {
  26                 lima_fence_slab = kmem_cache_create(
  27                         "lima_fence", sizeof(struct lima_fence), 0,
  28                         SLAB_HWCACHE_ALIGN, NULL);
  29                 if (!lima_fence_slab)
  30                         return -ENOMEM;
  31         }
  32
  33         lima_fence_slab_refcnt++;
  34         return 0;
  35 }
  36
  37 void lima_sched_slab_fini(void)
  38 {
  39         if (!--lima_fence_slab_refcnt) {
  40                 kmem_cache_destroy(lima_fence_slab);
  41                 lima_fence_slab = NULL;
  42         }
  43 }
  44
  45 static inline struct lima_fence *to_lima_fence(struct dma_fence *fence)
  46 {
  47         return container_of(fence, struct lima_fence, base);
  48 }
  49
  50 static const char *lima_fence_get_driver_name(struct dma_fence *fence)
  51 {
  52         return "lima";
  53 }
  54
  55 static const char *lima_fence_get_timeline_name(struct dma_fence *fence)
  56 {
  57         struct lima_fence *f = to_lima_fence(fence);
  58
  59         return f->pipe->base.name;
  60 }
  61
  62 static void lima_fence_release_rcu(struct rcu_head *rcu)
  63 {
  64         struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
  65         struct lima_fence *fence = to_lima_fence(f);
  66
  67         kmem_cache_free(lima_fence_slab, fence);
  68 }
  69
  70 static void lima_fence_release(struct dma_fence *fence)
  71 {
  72         struct lima_fence *f = to_lima_fence(fence);
  73
  74         call_rcu(&f->base.rcu, lima_fence_release_rcu);
  75 }
  76
  77 static const struct dma_fence_ops lima_fence_ops = {
  78         .get_driver_name = lima_fence_get_driver_name,
  79         .get_timeline_name = lima_fence_get_timeline_name,
  80         .release = lima_fence_release,
  81 };
  82
  83 static struct lima_fence *lima_fence_create(struct lima_sched_pipe *pipe)
  84 {
  85         struct lima_fence *fence;
  86
  87         fence = kmem_cache_zalloc(lima_fence_slab, GFP_KERNEL);
  88         if (!fence)
  89                 return NULL;
  90
  91         fence->pipe = pipe;
  92         dma_fence_init(&fence->base, &lima_fence_ops, &pipe->fence_lock,
  93                        pipe->fence_context, ++pipe->fence_seqno);
  94
  95         return fence;
  96 }
  97
  98 static inline struct lima_sched_task *to_lima_task(struct drm_sched_job *job)
  99 {
 100         return container_of(job, struct lima_sched_task, base);
 101 }
 102
 103 static inline struct lima_sched_pipe *to_lima_pipe(struct drm_gpu_scheduler *sched)
 104 {
 105         return container_of(sched, struct lima_sched_pipe, base);
 106 }
 107
 108 int lima_sched_task_init(struct lima_sched_task *task,
 109                          struct lima_sched_context *context,
 110                          struct lima_bo **bos, int num_bos,
 111                          struct lima_vm *vm)
 112 {
 113         int err, i;
 114
 115         task->bos = kmemdup(bos, sizeof(*bos) * num_bos, GFP_KERNEL);
 116         if (!task->bos)
 117                 return -ENOMEM;
 118
 119         for (i = 0; i < num_bos; i++)
 120                 drm_gem_object_get(&bos[i]->base.base);
 121
 122         err = drm_sched_job_init(&task->base, &context->base, vm);
 123         if (err) {
 124                 kfree(task->bos);
 125                 return err;
 126         }
 127
 128         task->num_bos = num_bos;
 129         task->vm = lima_vm_get(vm);
 130
 131         xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
 132
 133         return 0;
 134 }
 135
 136 void lima_sched_task_fini(struct lima_sched_task *task)
 137 {
 138         struct dma_fence *fence;
 139         unsigned long index;
 140         int i;
 141
 142         drm_sched_job_cleanup(&task->base);
 143
 144         xa_for_each(&task->deps, index, fence) {
 145                 dma_fence_put(fence);
 146         }
 147         xa_destroy(&task->deps);
 148
 149         if (task->bos) {
 150                 for (i = 0; i < task->num_bos; i++)
 151                         drm_gem_object_put_unlocked(&task->bos[i]->base.base);
 152                 kfree(task->bos);
 153         }
 154
 155         lima_vm_put(task->vm);
 156 }
 157
 158 int lima_sched_context_init(struct lima_sched_pipe *pipe,
 159                             struct lima_sched_context *context,
 160                             atomic_t *guilty)
 161 {
 162         struct drm_gpu_scheduler *sched = &pipe->base;
 163
 164         return drm_sched_entity_init(&context->base, DRM_SCHED_PRIORITY_NORMAL,
 165                                      &sched, 1, guilty);
 166 }
 167
 168 void lima_sched_context_fini(struct lima_sched_pipe *pipe,
 169                              struct lima_sched_context *context)
 170 {
 171         drm_sched_entity_fini(&context->base);
 172 }
 173
 174 struct dma_fence *lima_sched_context_queue_task(struct lima_sched_context *context,
 175                                                 struct lima_sched_task *task)
 176 {
 177         struct dma_fence *fence = dma_fence_get(&task->base.s_fence->finished);
 178
 179         drm_sched_entity_push_job(&task->base, &context->base);
 180         return fence;
 181 }
 182
 183 static struct dma_fence *lima_sched_dependency(struct drm_sched_job *job,
 184                                                struct drm_sched_entity *entity)
 185 {
 186         struct lima_sched_task *task = to_lima_task(job);
 187
 188         if (!xa_empty(&task->deps))
 189                 return xa_erase(&task->deps, task->last_dep++);
 190
 191         return NULL;
 192 }
 193
 194 static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job)
 195 {
 196         struct lima_sched_task *task = to_lima_task(job);
 197         struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
 198         struct lima_fence *fence;
 199         struct dma_fence *ret;
 200         struct lima_vm *vm = NULL, *last_vm = NULL;
 201         int i;
 202
 203         /* after GPU reset */
 204         if (job->s_fence->finished.error < 0)
 205                 return NULL;
 206
 207         fence = lima_fence_create(pipe);
 208         if (!fence)
 209                 return NULL;
 210         task->fence = &fence->base;
 211
 212         /* for caller usage of the fence, otherwise irq handler
 213          * may consume the fence before caller use it
 214          */
 215         ret = dma_fence_get(task->fence);
 216
 217         pipe->current_task = task;
 218
 219         /* this is needed for MMU to work correctly, otherwise GP/PP
 220          * will hang or page fault for unknown reason after running for
 221          * a while.
 222          *
 223          * Need to investigate:
 224          * 1. is it related to TLB
 225          * 2. how much performance will be affected by L2 cache flush
 226          * 3. can we reduce the calling of this function because all
 227          *    GP/PP use the same L2 cache on mali400
 228          *
 229          * TODO:
 230          * 1. move this to task fini to save some wait time?
 231          * 2. when GP/PP use different l2 cache, need PP wait GP l2
 232          *    cache flush?
 233          */
 234         for (i = 0; i < pipe->num_l2_cache; i++)
 235                 lima_l2_cache_flush(pipe->l2_cache[i]);
 236
 237         if (task->vm != pipe->current_vm) {
 238                 vm = lima_vm_get(task->vm);
 239                 last_vm = pipe->current_vm;
 240                 pipe->current_vm = task->vm;
 241         }
 242
 243         if (pipe->bcast_mmu)
 244                 lima_mmu_switch_vm(pipe->bcast_mmu, vm);
 245         else {
 246                 for (i = 0; i < pipe->num_mmu; i++)
 247                         lima_mmu_switch_vm(pipe->mmu[i], vm);
 248         }
 249
 250         if (last_vm)
 251                 lima_vm_put(last_vm);
 252
 253         pipe->error = false;
 254         pipe->task_run(pipe, task);
 255
 256         return task->fence;
 257 }
 258
 259 static void lima_sched_timedout_job(struct drm_sched_job *job)
 260 {
 261         struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
 262         struct lima_sched_task *task = to_lima_task(job);
 263
 264         if (!pipe->error)
 265                 DRM_ERROR("lima job timeout\n");
 266
 267         drm_sched_stop(&pipe->base, &task->base);
 268
 269         drm_sched_increase_karma(&task->base);
 270
 271         pipe->task_error(pipe);
 272
 273         if (pipe->bcast_mmu)
 274                 lima_mmu_page_fault_resume(pipe->bcast_mmu);
 275         else {
 276                 int i;
 277
 278                 for (i = 0; i < pipe->num_mmu; i++)
 279                         lima_mmu_page_fault_resume(pipe->mmu[i]);
 280         }
 281
 282         if (pipe->current_vm)
 283                 lima_vm_put(pipe->current_vm);
 284
 285         pipe->current_vm = NULL;
 286         pipe->current_task = NULL;
 287
 288         drm_sched_resubmit_jobs(&pipe->base);
 289         drm_sched_start(&pipe->base, true);
 290 }
 291
 292 static void lima_sched_free_job(struct drm_sched_job *job)
 293 {
 294         struct lima_sched_task *task = to_lima_task(job);
 295         struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
 296         struct lima_vm *vm = task->vm;
 297         struct lima_bo **bos = task->bos;
 298         int i;
 299
 300         dma_fence_put(task->fence);
 301
 302         for (i = 0; i < task->num_bos; i++)
 303                 lima_vm_bo_del(vm, bos[i]);
 304
 305         lima_sched_task_fini(task);
 306         kmem_cache_free(pipe->task_slab, task);
 307 }
 308
 309 static const struct drm_sched_backend_ops lima_sched_ops = {
 310         .dependency = lima_sched_dependency,
 311         .run_job = lima_sched_run_job,
 312         .timedout_job = lima_sched_timedout_job,
 313         .free_job = lima_sched_free_job,
 314 };
 315
 316 int lima_sched_pipe_init(struct lima_sched_pipe *pipe, const char *name)
 317 {
 318         unsigned int timeout = lima_sched_timeout_ms > 0 ?
 319                                lima_sched_timeout_ms : 500;
 320
 321         pipe->fence_context = dma_fence_context_alloc(1);
 322         spin_lock_init(&pipe->fence_lock);
 323
 324         return drm_sched_init(&pipe->base, &lima_sched_ops, 1, 0,
 325                               msecs_to_jiffies(timeout), name);
 326 }
 327
 328 void lima_sched_pipe_fini(struct lima_sched_pipe *pipe)
 329 {
 330         drm_sched_fini(&pipe->base);
 331 }
 332
 333 void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe)
 334 {
 335         if (pipe->error)
 336                 drm_sched_fault(&pipe->base);
 337         else {
 338                 struct lima_sched_task *task = pipe->current_task;
 339
 340                 pipe->task_fini(pipe);
 341                 dma_fence_signal(task->fence);
 342         }
 343 }