1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2015-2018 Broadcom */
3
4 #include <linux/delay.h>
5 #include <linux/mutex.h>
6 #include <linux/spinlock_types.h>
7 #include <linux/workqueue.h>
8
9 #include <drm/drm_encoder.h>
10 #include <drm/drm_gem.h>
11 #include <drm/drm_gem_shmem_helper.h>
12 #include <drm/gpu_scheduler.h>
13
14 #include "v3d_performance_counters.h"
15
16 #include "uapi/drm/v3d_drm.h"
17
18 struct clk;
19 struct platform_device;
20 struct reset_control;
21
22 #define V3D_MMU_PAGE_SHIFT 12
23 #define V3D_PAGE_FACTOR (PAGE_SIZE >> V3D_MMU_PAGE_SHIFT)
24
25 #define V3D_MAX_QUEUES (V3D_CPU + 1)
26
v3d_queue_to_string(enum v3d_queue queue)27 static inline char *v3d_queue_to_string(enum v3d_queue queue)
28 {
29 switch (queue) {
30 case V3D_BIN: return "bin";
31 case V3D_RENDER: return "render";
32 case V3D_TFU: return "tfu";
33 case V3D_CSD: return "csd";
34 case V3D_CACHE_CLEAN: return "cache_clean";
35 case V3D_CPU: return "cpu";
36 }
37 return "UNKNOWN";
38 }
39
40 struct v3d_stats {
41 u64 start_ns;
42 u64 enabled_ns;
43 u64 jobs_completed;
44
45 /*
46 * This seqcount is used to protect the access to the GPU stats
47 * variables. It must be used as, while we are reading the stats,
48 * IRQs can happen and the stats can be updated.
49 */
50 seqcount_t lock;
51 };
52
53 struct v3d_queue_state {
54 struct drm_gpu_scheduler sched;
55
56 u64 fence_context;
57 u64 emit_seqno;
58
59 /* Stores the GPU stats for this queue in the global context. */
60 struct v3d_stats stats;
61 };
62
63 /* Performance monitor object. The perform lifetime is controlled by userspace
64 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
65 * request, and when this is the case, HW perf counters will be activated just
66 * before the submit_cl is submitted to the GPU and disabled when the job is
67 * done. This way, only events related to a specific job will be counted.
68 */
69 struct v3d_perfmon {
70 /* Tracks the number of users of the perfmon, when this counter reaches
71 * zero the perfmon is destroyed.
72 */
73 refcount_t refcnt;
74
75 /* Protects perfmon stop, as it can be invoked from multiple places. */
76 struct mutex lock;
77
78 /* Number of counters activated in this perfmon instance
79 * (should be less than DRM_V3D_MAX_PERF_COUNTERS).
80 */
81 u8 ncounters;
82
83 /* Events counted by the HW perf counters. */
84 u8 counters[DRM_V3D_MAX_PERF_COUNTERS];
85
86 /* Storage for counter values. Counters are incremented by the
87 * HW perf counter values every time the perfmon is attached
88 * to a GPU job. This way, perfmon users don't have to
89 * retrieve the results after each job if they want to track
90 * events covering several submissions. Note that counter
91 * values can't be reset, but you can fake a reset by
92 * destroying the perfmon and creating a new one.
93 */
94 u64 values[] __counted_by(ncounters);
95 };
96
97 struct v3d_dev {
98 struct drm_device drm;
99
100 /* Short representation (e.g. 33, 41) of the V3D tech version */
101 int ver;
102
103 /* Short representation (e.g. 5, 6) of the V3D tech revision */
104 int rev;
105
106 bool single_irq_line;
107
108 struct v3d_perfmon_info perfmon_info;
109
110 void __iomem *hub_regs;
111 void __iomem *core_regs[3];
112 void __iomem *bridge_regs;
113 void __iomem *gca_regs;
114 struct clk *clk;
115 struct reset_control *reset;
116
117 /* Virtual and DMA addresses of the single shared page table. */
118 volatile u32 *pt;
119 dma_addr_t pt_paddr;
120
121 /* Virtual and DMA addresses of the MMU's scratch page. When
122 * a read or write is invalid in the MMU, it will be
123 * redirected here.
124 */
125 void *mmu_scratch;
126 dma_addr_t mmu_scratch_paddr;
127 /* virtual address bits from V3D to the MMU. */
128 int va_width;
129
130 /* Number of V3D cores. */
131 u32 cores;
132
133 /* Allocator managing the address space. All units are in
134 * number of pages.
135 */
136 struct drm_mm mm;
137 spinlock_t mm_lock;
138
139 /*
140 * tmpfs instance used for shmem backed objects
141 */
142 struct vfsmount *gemfs;
143
144 struct work_struct overflow_mem_work;
145
146 struct v3d_bin_job *bin_job;
147 struct v3d_render_job *render_job;
148 struct v3d_tfu_job *tfu_job;
149 struct v3d_csd_job *csd_job;
150 struct v3d_cpu_job *cpu_job;
151
152 struct v3d_queue_state queue[V3D_MAX_QUEUES];
153
154 /* Spinlock used to synchronize the overflow memory
155 * management against bin job submission.
156 */
157 spinlock_t job_lock;
158
159 /* Used to track the active perfmon if any. */
160 struct v3d_perfmon *active_perfmon;
161
162 /* Protects bo_stats */
163 struct mutex bo_lock;
164
165 /* Lock taken when resetting the GPU, to keep multiple
166 * processes from trying to park the scheduler threads and
167 * reset at once.
168 */
169 struct mutex reset_lock;
170
171 /* Lock taken when creating and pushing the GPU scheduler
172 * jobs, to keep the sched-fence seqnos in order.
173 */
174 struct mutex sched_lock;
175
176 /* Lock taken during a cache clean and when initiating an L2
177 * flush, to keep L2 flushes from interfering with the
178 * synchronous L2 cleans.
179 */
180 struct mutex cache_clean_lock;
181
182 struct {
183 u32 num_allocated;
184 u32 pages_allocated;
185 } bo_stats;
186 };
187
188 static inline struct v3d_dev *
to_v3d_dev(struct drm_device * dev)189 to_v3d_dev(struct drm_device *dev)
190 {
191 return container_of(dev, struct v3d_dev, drm);
192 }
193
194 static inline bool
v3d_has_csd(struct v3d_dev * v3d)195 v3d_has_csd(struct v3d_dev *v3d)
196 {
197 return v3d->ver >= 41;
198 }
199
200 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
201
202 /* The per-fd struct, which tracks the MMU mappings. */
203 struct v3d_file_priv {
204 struct v3d_dev *v3d;
205
206 struct {
207 struct idr idr;
208 struct mutex lock;
209 } perfmon;
210
211 struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
212
213 /* Stores the GPU stats for a specific queue for this fd. */
214 struct v3d_stats stats[V3D_MAX_QUEUES];
215 };
216
217 struct v3d_bo {
218 struct drm_gem_shmem_object base;
219
220 struct drm_mm_node node;
221
222 /* List entry for the BO's position in
223 * v3d_render_job->unref_list
224 */
225 struct list_head unref_head;
226
227 void *vaddr;
228 };
229
230 static inline struct v3d_bo *
to_v3d_bo(struct drm_gem_object * bo)231 to_v3d_bo(struct drm_gem_object *bo)
232 {
233 return (struct v3d_bo *)bo;
234 }
235
236 struct v3d_fence {
237 struct dma_fence base;
238 struct drm_device *dev;
239 /* v3d seqno for signaled() test */
240 u64 seqno;
241 enum v3d_queue queue;
242 };
243
244 static inline struct v3d_fence *
to_v3d_fence(struct dma_fence * fence)245 to_v3d_fence(struct dma_fence *fence)
246 {
247 return (struct v3d_fence *)fence;
248 }
249
250 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
251 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
252
253 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
254 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
255
256 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
257 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
258
259 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
260 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
261
262 struct v3d_job {
263 struct drm_sched_job base;
264
265 struct kref refcount;
266
267 struct v3d_dev *v3d;
268
269 /* This is the array of BOs that were looked up at the start
270 * of submission.
271 */
272 struct drm_gem_object **bo;
273 u32 bo_count;
274
275 /* v3d fence to be signaled by IRQ handler when the job is complete. */
276 struct dma_fence *irq_fence;
277
278 /* scheduler fence for when the job is considered complete and
279 * the BO reservations can be released.
280 */
281 struct dma_fence *done_fence;
282
283 /* Pointer to a performance monitor object if the user requested it,
284 * NULL otherwise.
285 */
286 struct v3d_perfmon *perfmon;
287
288 /* File descriptor of the process that submitted the job that could be used
289 * for collecting stats by process of GPU usage.
290 */
291 struct drm_file *file;
292
293 /* Callback for the freeing of the job on refcount going to 0. */
294 void (*free)(struct kref *ref);
295 };
296
297 struct v3d_bin_job {
298 struct v3d_job base;
299
300 /* GPU virtual addresses of the start/end of the CL job. */
301 u32 start, end;
302
303 u32 timedout_ctca, timedout_ctra;
304
305 /* Corresponding render job, for attaching our overflow memory. */
306 struct v3d_render_job *render;
307
308 /* Submitted tile memory allocation start/size, tile state. */
309 u32 qma, qms, qts;
310 };
311
312 struct v3d_render_job {
313 struct v3d_job base;
314
315 /* GPU virtual addresses of the start/end of the CL job. */
316 u32 start, end;
317
318 u32 timedout_ctca, timedout_ctra;
319
320 /* List of overflow BOs used in the job that need to be
321 * released once the job is complete.
322 */
323 struct list_head unref_list;
324 };
325
326 struct v3d_tfu_job {
327 struct v3d_job base;
328
329 struct drm_v3d_submit_tfu args;
330 };
331
332 struct v3d_csd_job {
333 struct v3d_job base;
334
335 u32 timedout_batches;
336
337 struct drm_v3d_submit_csd args;
338 };
339
340 enum v3d_cpu_job_type {
341 V3D_CPU_JOB_TYPE_INDIRECT_CSD = 1,
342 V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY,
343 V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY,
344 V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY,
345 V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY,
346 V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY,
347 };
348
349 struct v3d_timestamp_query {
350 /* Offset of this query in the timestamp BO for its value. */
351 u32 offset;
352
353 /* Syncobj that indicates the timestamp availability */
354 struct drm_syncobj *syncobj;
355 };
356
357 struct v3d_performance_query {
358 /* Performance monitor IDs for this query */
359 u32 *kperfmon_ids;
360
361 /* Syncobj that indicates the query availability */
362 struct drm_syncobj *syncobj;
363 };
364
365 struct v3d_indirect_csd_info {
366 /* Indirect CSD */
367 struct v3d_csd_job *job;
368
369 /* Clean cache job associated to the Indirect CSD job */
370 struct v3d_job *clean_job;
371
372 /* Offset within the BO where the workgroup counts are stored */
373 u32 offset;
374
375 /* Workgroups size */
376 u32 wg_size;
377
378 /* Indices of the uniforms with the workgroup dispatch counts
379 * in the uniform stream.
380 */
381 u32 wg_uniform_offsets[3];
382
383 /* Indirect BO */
384 struct drm_gem_object *indirect;
385
386 /* Context of the Indirect CSD job */
387 struct ww_acquire_ctx acquire_ctx;
388 };
389
390 struct v3d_timestamp_query_info {
391 struct v3d_timestamp_query *queries;
392
393 u32 count;
394 };
395
396 struct v3d_performance_query_info {
397 struct v3d_performance_query *queries;
398
399 /* Number of performance queries */
400 u32 count;
401
402 /* Number of performance monitors related to that query pool */
403 u32 nperfmons;
404
405 /* Number of performance counters related to that query pool */
406 u32 ncounters;
407 };
408
409 struct v3d_copy_query_results_info {
410 /* Define if should write to buffer using 64 or 32 bits */
411 bool do_64bit;
412
413 /* Define if it can write to buffer even if the query is not available */
414 bool do_partial;
415
416 /* Define if it should write availability bit to buffer */
417 bool availability_bit;
418
419 /* Offset of the copy buffer in the BO */
420 u32 offset;
421
422 /* Stride of the copy buffer in the BO */
423 u32 stride;
424 };
425
426 struct v3d_cpu_job {
427 struct v3d_job base;
428
429 enum v3d_cpu_job_type job_type;
430
431 struct v3d_indirect_csd_info indirect_csd;
432
433 struct v3d_timestamp_query_info timestamp_query;
434
435 struct v3d_copy_query_results_info copy;
436
437 struct v3d_performance_query_info performance_query;
438 };
439
440 typedef void (*v3d_cpu_job_fn)(struct v3d_cpu_job *);
441
442 struct v3d_submit_outsync {
443 struct drm_syncobj *syncobj;
444 };
445
446 struct v3d_submit_ext {
447 u32 flags;
448 u32 wait_stage;
449
450 u32 in_sync_count;
451 u64 in_syncs;
452
453 u32 out_sync_count;
454 struct v3d_submit_outsync *out_syncs;
455 };
456
457 /**
458 * __wait_for - magic wait macro
459 *
460 * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
461 * important that we check the condition again after having timed out, since the
462 * timeout could be due to preemption or similar and we've never had a chance to
463 * check the condition before the timeout.
464 */
465 #define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
466 const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
467 long wait__ = (Wmin); /* recommended min for usleep is 10 us */ \
468 int ret__; \
469 might_sleep(); \
470 for (;;) { \
471 const bool expired__ = ktime_after(ktime_get_raw(), end__); \
472 OP; \
473 /* Guarantee COND check prior to timeout */ \
474 barrier(); \
475 if (COND) { \
476 ret__ = 0; \
477 break; \
478 } \
479 if (expired__) { \
480 ret__ = -ETIMEDOUT; \
481 break; \
482 } \
483 usleep_range(wait__, wait__ * 2); \
484 if (wait__ < (Wmax)) \
485 wait__ <<= 1; \
486 } \
487 ret__; \
488 })
489
490 #define _wait_for(COND, US, Wmin, Wmax) __wait_for(, (COND), (US), (Wmin), \
491 (Wmax))
492 #define wait_for(COND, MS) _wait_for((COND), (MS) * 1000, 10, 1000)
493
nsecs_to_jiffies_timeout(const u64 n)494 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
495 {
496 /* nsecs_to_jiffies64() does not guard against overflow */
497 if ((NSEC_PER_SEC % HZ) != 0 &&
498 div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
499 return MAX_JIFFY_OFFSET;
500
501 return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
502 }
503
504 /* v3d_bo.c */
505 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size);
506 void v3d_free_object(struct drm_gem_object *gem_obj);
507 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
508 size_t size);
509 void v3d_get_bo_vaddr(struct v3d_bo *bo);
510 void v3d_put_bo_vaddr(struct v3d_bo *bo);
511 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
512 struct drm_file *file_priv);
513 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
514 struct drm_file *file_priv);
515 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
516 struct drm_file *file_priv);
517 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
518 struct drm_file *file_priv);
519 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
520 struct dma_buf_attachment *attach,
521 struct sg_table *sgt);
522
523 /* v3d_debugfs.c */
524 void v3d_debugfs_init(struct drm_minor *minor);
525
526 /* v3d_drv.c */
527 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
528 u64 *active_runtime, u64 *jobs_completed);
529
530 /* v3d_fence.c */
531 extern const struct dma_fence_ops v3d_fence_ops;
532 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
533
534 /* v3d_gem.c */
535 int v3d_gem_init(struct drm_device *dev);
536 void v3d_gem_destroy(struct drm_device *dev);
537 void v3d_reset(struct v3d_dev *v3d);
538 void v3d_invalidate_caches(struct v3d_dev *v3d);
539 void v3d_clean_caches(struct v3d_dev *v3d);
540
541 /* v3d_gemfs.c */
542 extern bool super_pages;
543 void v3d_gemfs_init(struct v3d_dev *v3d);
544 void v3d_gemfs_fini(struct v3d_dev *v3d);
545
546 /* v3d_submit.c */
547 void v3d_job_cleanup(struct v3d_job *job);
548 void v3d_job_put(struct v3d_job *job);
549 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
550 struct drm_file *file_priv);
551 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
552 struct drm_file *file_priv);
553 int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
554 struct drm_file *file_priv);
555 int v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
556 struct drm_file *file_priv);
557
558 /* v3d_irq.c */
559 int v3d_irq_init(struct v3d_dev *v3d);
560 void v3d_irq_enable(struct v3d_dev *v3d);
561 void v3d_irq_disable(struct v3d_dev *v3d);
562 void v3d_irq_reset(struct v3d_dev *v3d);
563
564 /* v3d_mmu.c */
565 int v3d_mmu_flush_all(struct v3d_dev *v3d);
566 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
567 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
568 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
569
570 /* v3d_sched.c */
571 void v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
572 unsigned int count);
573 void v3d_performance_query_info_free(struct v3d_performance_query_info *query_info,
574 unsigned int count);
575 void v3d_job_update_stats(struct v3d_job *job, enum v3d_queue queue);
576 int v3d_sched_init(struct v3d_dev *v3d);
577 void v3d_sched_fini(struct v3d_dev *v3d);
578
579 /* v3d_perfmon.c */
580 void v3d_perfmon_init(struct v3d_dev *v3d);
581 void v3d_perfmon_get(struct v3d_perfmon *perfmon);
582 void v3d_perfmon_put(struct v3d_perfmon *perfmon);
583 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon);
584 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
585 bool capture);
586 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id);
587 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv);
588 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv);
589 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
590 struct drm_file *file_priv);
591 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
592 struct drm_file *file_priv);
593 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
594 struct drm_file *file_priv);
595 int v3d_perfmon_get_counter_ioctl(struct drm_device *dev, void *data,
596 struct drm_file *file_priv);
597
598 /* v3d_sysfs.c */
599 int v3d_sysfs_init(struct device *dev);
600 void v3d_sysfs_destroy(struct device *dev);
601