xref: /linux/drivers/gpu/drm/v3d/v3d_drv.h (revision 2697b79a469b68e3ad3640f55284359c1396278d)
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 "uapi/drm/v3d_drm.h"
15 
16 struct clk;
17 struct platform_device;
18 struct reset_control;
19 
20 #define GMP_GRANULARITY (128 * 1024)
21 
22 #define V3D_MMU_PAGE_SHIFT 12
23 
24 #define V3D_MAX_QUEUES (V3D_CPU + 1)
25 
26 static inline char *v3d_queue_to_string(enum v3d_queue queue)
27 {
28 	switch (queue) {
29 	case V3D_BIN: return "bin";
30 	case V3D_RENDER: return "render";
31 	case V3D_TFU: return "tfu";
32 	case V3D_CSD: return "csd";
33 	case V3D_CACHE_CLEAN: return "cache_clean";
34 	case V3D_CPU: return "cpu";
35 	}
36 	return "UNKNOWN";
37 }
38 
39 struct v3d_stats {
40 	u64 start_ns;
41 	u64 enabled_ns;
42 	u64 jobs_completed;
43 
44 	/*
45 	 * This seqcount is used to protect the access to the GPU stats
46 	 * variables. It must be used as, while we are reading the stats,
47 	 * IRQs can happen and the stats can be updated.
48 	 */
49 	seqcount_t lock;
50 };
51 
52 struct v3d_queue_state {
53 	struct drm_gpu_scheduler sched;
54 
55 	u64 fence_context;
56 	u64 emit_seqno;
57 
58 	/* Stores the GPU stats for this queue in the global context. */
59 	struct v3d_stats stats;
60 };
61 
62 /* Performance monitor object. The perform lifetime is controlled by userspace
63  * using perfmon related ioctls. A perfmon can be attached to a submit_cl
64  * request, and when this is the case, HW perf counters will be activated just
65  * before the submit_cl is submitted to the GPU and disabled when the job is
66  * done. This way, only events related to a specific job will be counted.
67  */
68 struct v3d_perfmon {
69 	/* Tracks the number of users of the perfmon, when this counter reaches
70 	 * zero the perfmon is destroyed.
71 	 */
72 	refcount_t refcnt;
73 
74 	/* Protects perfmon stop, as it can be invoked from multiple places. */
75 	struct mutex lock;
76 
77 	/* Number of counters activated in this perfmon instance
78 	 * (should be less than DRM_V3D_MAX_PERF_COUNTERS).
79 	 */
80 	u8 ncounters;
81 
82 	/* Events counted by the HW perf counters. */
83 	u8 counters[DRM_V3D_MAX_PERF_COUNTERS];
84 
85 	/* Storage for counter values. Counters are incremented by the
86 	 * HW perf counter values every time the perfmon is attached
87 	 * to a GPU job.  This way, perfmon users don't have to
88 	 * retrieve the results after each job if they want to track
89 	 * events covering several submissions.  Note that counter
90 	 * values can't be reset, but you can fake a reset by
91 	 * destroying the perfmon and creating a new one.
92 	 */
93 	u64 values[] __counted_by(ncounters);
94 };
95 
96 struct v3d_dev {
97 	struct drm_device drm;
98 
99 	/* Short representation (e.g. 33, 41) of the V3D tech version
100 	 * and revision.
101 	 */
102 	int ver;
103 	bool single_irq_line;
104 
105 	void __iomem *hub_regs;
106 	void __iomem *core_regs[3];
107 	void __iomem *bridge_regs;
108 	void __iomem *gca_regs;
109 	struct clk *clk;
110 	struct reset_control *reset;
111 
112 	/* Virtual and DMA addresses of the single shared page table. */
113 	volatile u32 *pt;
114 	dma_addr_t pt_paddr;
115 
116 	/* Virtual and DMA addresses of the MMU's scratch page.  When
117 	 * a read or write is invalid in the MMU, it will be
118 	 * redirected here.
119 	 */
120 	void *mmu_scratch;
121 	dma_addr_t mmu_scratch_paddr;
122 	/* virtual address bits from V3D to the MMU. */
123 	int va_width;
124 
125 	/* Number of V3D cores. */
126 	u32 cores;
127 
128 	/* Allocator managing the address space.  All units are in
129 	 * number of pages.
130 	 */
131 	struct drm_mm mm;
132 	spinlock_t mm_lock;
133 
134 	struct work_struct overflow_mem_work;
135 
136 	struct v3d_bin_job *bin_job;
137 	struct v3d_render_job *render_job;
138 	struct v3d_tfu_job *tfu_job;
139 	struct v3d_csd_job *csd_job;
140 	struct v3d_cpu_job *cpu_job;
141 
142 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
143 
144 	/* Spinlock used to synchronize the overflow memory
145 	 * management against bin job submission.
146 	 */
147 	spinlock_t job_lock;
148 
149 	/* Used to track the active perfmon if any. */
150 	struct v3d_perfmon *active_perfmon;
151 
152 	/* Protects bo_stats */
153 	struct mutex bo_lock;
154 
155 	/* Lock taken when resetting the GPU, to keep multiple
156 	 * processes from trying to park the scheduler threads and
157 	 * reset at once.
158 	 */
159 	struct mutex reset_lock;
160 
161 	/* Lock taken when creating and pushing the GPU scheduler
162 	 * jobs, to keep the sched-fence seqnos in order.
163 	 */
164 	struct mutex sched_lock;
165 
166 	/* Lock taken during a cache clean and when initiating an L2
167 	 * flush, to keep L2 flushes from interfering with the
168 	 * synchronous L2 cleans.
169 	 */
170 	struct mutex cache_clean_lock;
171 
172 	struct {
173 		u32 num_allocated;
174 		u32 pages_allocated;
175 	} bo_stats;
176 };
177 
178 static inline struct v3d_dev *
179 to_v3d_dev(struct drm_device *dev)
180 {
181 	return container_of(dev, struct v3d_dev, drm);
182 }
183 
184 static inline bool
185 v3d_has_csd(struct v3d_dev *v3d)
186 {
187 	return v3d->ver >= 41;
188 }
189 
190 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
191 
192 /* The per-fd struct, which tracks the MMU mappings. */
193 struct v3d_file_priv {
194 	struct v3d_dev *v3d;
195 
196 	struct {
197 		struct idr idr;
198 		struct mutex lock;
199 	} perfmon;
200 
201 	struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
202 
203 	/* Stores the GPU stats for a specific queue for this fd. */
204 	struct v3d_stats stats[V3D_MAX_QUEUES];
205 };
206 
207 struct v3d_bo {
208 	struct drm_gem_shmem_object base;
209 
210 	struct drm_mm_node node;
211 
212 	/* List entry for the BO's position in
213 	 * v3d_render_job->unref_list
214 	 */
215 	struct list_head unref_head;
216 
217 	void *vaddr;
218 };
219 
220 static inline struct v3d_bo *
221 to_v3d_bo(struct drm_gem_object *bo)
222 {
223 	return (struct v3d_bo *)bo;
224 }
225 
226 struct v3d_fence {
227 	struct dma_fence base;
228 	struct drm_device *dev;
229 	/* v3d seqno for signaled() test */
230 	u64 seqno;
231 	enum v3d_queue queue;
232 };
233 
234 static inline struct v3d_fence *
235 to_v3d_fence(struct dma_fence *fence)
236 {
237 	return (struct v3d_fence *)fence;
238 }
239 
240 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
241 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
242 
243 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
244 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
245 
246 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
247 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
248 
249 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
250 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
251 
252 struct v3d_job {
253 	struct drm_sched_job base;
254 
255 	struct kref refcount;
256 
257 	struct v3d_dev *v3d;
258 
259 	/* This is the array of BOs that were looked up at the start
260 	 * of submission.
261 	 */
262 	struct drm_gem_object **bo;
263 	u32 bo_count;
264 
265 	/* v3d fence to be signaled by IRQ handler when the job is complete. */
266 	struct dma_fence *irq_fence;
267 
268 	/* scheduler fence for when the job is considered complete and
269 	 * the BO reservations can be released.
270 	 */
271 	struct dma_fence *done_fence;
272 
273 	/* Pointer to a performance monitor object if the user requested it,
274 	 * NULL otherwise.
275 	 */
276 	struct v3d_perfmon *perfmon;
277 
278 	/* File descriptor of the process that submitted the job that could be used
279 	 * for collecting stats by process of GPU usage.
280 	 */
281 	struct drm_file *file;
282 
283 	/* Callback for the freeing of the job on refcount going to 0. */
284 	void (*free)(struct kref *ref);
285 };
286 
287 struct v3d_bin_job {
288 	struct v3d_job base;
289 
290 	/* GPU virtual addresses of the start/end of the CL job. */
291 	u32 start, end;
292 
293 	u32 timedout_ctca, timedout_ctra;
294 
295 	/* Corresponding render job, for attaching our overflow memory. */
296 	struct v3d_render_job *render;
297 
298 	/* Submitted tile memory allocation start/size, tile state. */
299 	u32 qma, qms, qts;
300 };
301 
302 struct v3d_render_job {
303 	struct v3d_job base;
304 
305 	/* GPU virtual addresses of the start/end of the CL job. */
306 	u32 start, end;
307 
308 	u32 timedout_ctca, timedout_ctra;
309 
310 	/* List of overflow BOs used in the job that need to be
311 	 * released once the job is complete.
312 	 */
313 	struct list_head unref_list;
314 };
315 
316 struct v3d_tfu_job {
317 	struct v3d_job base;
318 
319 	struct drm_v3d_submit_tfu args;
320 };
321 
322 struct v3d_csd_job {
323 	struct v3d_job base;
324 
325 	u32 timedout_batches;
326 
327 	struct drm_v3d_submit_csd args;
328 };
329 
330 enum v3d_cpu_job_type {
331 	V3D_CPU_JOB_TYPE_INDIRECT_CSD = 1,
332 	V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY,
333 	V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY,
334 	V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY,
335 	V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY,
336 	V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY,
337 };
338 
339 struct v3d_timestamp_query {
340 	/* Offset of this query in the timestamp BO for its value. */
341 	u32 offset;
342 
343 	/* Syncobj that indicates the timestamp availability */
344 	struct drm_syncobj *syncobj;
345 };
346 
347 /* Number of perfmons required to handle all supported performance counters */
348 #define V3D_MAX_PERFMONS DIV_ROUND_UP(V3D_PERFCNT_NUM, \
349 				      DRM_V3D_MAX_PERF_COUNTERS)
350 
351 struct v3d_performance_query {
352 	/* Performance monitor IDs for this query */
353 	u32 kperfmon_ids[V3D_MAX_PERFMONS];
354 
355 	/* Syncobj that indicates the query availability */
356 	struct drm_syncobj *syncobj;
357 };
358 
359 struct v3d_indirect_csd_info {
360 	/* Indirect CSD */
361 	struct v3d_csd_job *job;
362 
363 	/* Clean cache job associated to the Indirect CSD job */
364 	struct v3d_job *clean_job;
365 
366 	/* Offset within the BO where the workgroup counts are stored */
367 	u32 offset;
368 
369 	/* Workgroups size */
370 	u32 wg_size;
371 
372 	/* Indices of the uniforms with the workgroup dispatch counts
373 	 * in the uniform stream.
374 	 */
375 	u32 wg_uniform_offsets[3];
376 
377 	/* Indirect BO */
378 	struct drm_gem_object *indirect;
379 
380 	/* Context of the Indirect CSD job */
381 	struct ww_acquire_ctx acquire_ctx;
382 };
383 
384 struct v3d_timestamp_query_info {
385 	struct v3d_timestamp_query *queries;
386 
387 	u32 count;
388 };
389 
390 struct v3d_performance_query_info {
391 	struct v3d_performance_query *queries;
392 
393 	/* Number of performance queries */
394 	u32 count;
395 
396 	/* Number of performance monitors related to that query pool */
397 	u32 nperfmons;
398 
399 	/* Number of performance counters related to that query pool */
400 	u32 ncounters;
401 };
402 
403 struct v3d_copy_query_results_info {
404 	/* Define if should write to buffer using 64 or 32 bits */
405 	bool do_64bit;
406 
407 	/* Define if it can write to buffer even if the query is not available */
408 	bool do_partial;
409 
410 	/* Define if it should write availability bit to buffer */
411 	bool availability_bit;
412 
413 	/* Offset of the copy buffer in the BO */
414 	u32 offset;
415 
416 	/* Stride of the copy buffer in the BO */
417 	u32 stride;
418 };
419 
420 struct v3d_cpu_job {
421 	struct v3d_job base;
422 
423 	enum v3d_cpu_job_type job_type;
424 
425 	struct v3d_indirect_csd_info indirect_csd;
426 
427 	struct v3d_timestamp_query_info timestamp_query;
428 
429 	struct v3d_copy_query_results_info copy;
430 
431 	struct v3d_performance_query_info performance_query;
432 };
433 
434 typedef void (*v3d_cpu_job_fn)(struct v3d_cpu_job *);
435 
436 struct v3d_submit_outsync {
437 	struct drm_syncobj *syncobj;
438 };
439 
440 struct v3d_submit_ext {
441 	u32 flags;
442 	u32 wait_stage;
443 
444 	u32 in_sync_count;
445 	u64 in_syncs;
446 
447 	u32 out_sync_count;
448 	struct v3d_submit_outsync *out_syncs;
449 };
450 
451 /**
452  * __wait_for - magic wait macro
453  *
454  * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
455  * important that we check the condition again after having timed out, since the
456  * timeout could be due to preemption or similar and we've never had a chance to
457  * check the condition before the timeout.
458  */
459 #define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
460 	const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
461 	long wait__ = (Wmin); /* recommended min for usleep is 10 us */	\
462 	int ret__;							\
463 	might_sleep();							\
464 	for (;;) {							\
465 		const bool expired__ = ktime_after(ktime_get_raw(), end__); \
466 		OP;							\
467 		/* Guarantee COND check prior to timeout */		\
468 		barrier();						\
469 		if (COND) {						\
470 			ret__ = 0;					\
471 			break;						\
472 		}							\
473 		if (expired__) {					\
474 			ret__ = -ETIMEDOUT;				\
475 			break;						\
476 		}							\
477 		usleep_range(wait__, wait__ * 2);			\
478 		if (wait__ < (Wmax))					\
479 			wait__ <<= 1;					\
480 	}								\
481 	ret__;								\
482 })
483 
484 #define _wait_for(COND, US, Wmin, Wmax)	__wait_for(, (COND), (US), (Wmin), \
485 						   (Wmax))
486 #define wait_for(COND, MS)		_wait_for((COND), (MS) * 1000, 10, 1000)
487 
488 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
489 {
490 	/* nsecs_to_jiffies64() does not guard against overflow */
491 	if ((NSEC_PER_SEC % HZ) != 0 &&
492 	    div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
493 		return MAX_JIFFY_OFFSET;
494 
495 	return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
496 }
497 
498 /* v3d_bo.c */
499 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size);
500 void v3d_free_object(struct drm_gem_object *gem_obj);
501 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
502 			     size_t size);
503 void v3d_get_bo_vaddr(struct v3d_bo *bo);
504 void v3d_put_bo_vaddr(struct v3d_bo *bo);
505 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
506 			struct drm_file *file_priv);
507 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
508 		      struct drm_file *file_priv);
509 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
510 			    struct drm_file *file_priv);
511 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
512 		      struct drm_file *file_priv);
513 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
514 						 struct dma_buf_attachment *attach,
515 						 struct sg_table *sgt);
516 
517 /* v3d_debugfs.c */
518 void v3d_debugfs_init(struct drm_minor *minor);
519 
520 /* v3d_drv.c */
521 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
522 		   u64 *active_runtime, u64 *jobs_completed);
523 
524 /* v3d_fence.c */
525 extern const struct dma_fence_ops v3d_fence_ops;
526 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
527 
528 /* v3d_gem.c */
529 int v3d_gem_init(struct drm_device *dev);
530 void v3d_gem_destroy(struct drm_device *dev);
531 void v3d_reset(struct v3d_dev *v3d);
532 void v3d_invalidate_caches(struct v3d_dev *v3d);
533 void v3d_clean_caches(struct v3d_dev *v3d);
534 
535 /* v3d_submit.c */
536 void v3d_job_cleanup(struct v3d_job *job);
537 void v3d_job_put(struct v3d_job *job);
538 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
539 			struct drm_file *file_priv);
540 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
541 			 struct drm_file *file_priv);
542 int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
543 			 struct drm_file *file_priv);
544 int v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
545 			 struct drm_file *file_priv);
546 
547 /* v3d_irq.c */
548 int v3d_irq_init(struct v3d_dev *v3d);
549 void v3d_irq_enable(struct v3d_dev *v3d);
550 void v3d_irq_disable(struct v3d_dev *v3d);
551 void v3d_irq_reset(struct v3d_dev *v3d);
552 
553 /* v3d_mmu.c */
554 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
555 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
556 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
557 
558 /* v3d_sched.c */
559 void v3d_job_update_stats(struct v3d_job *job, enum v3d_queue queue);
560 int v3d_sched_init(struct v3d_dev *v3d);
561 void v3d_sched_fini(struct v3d_dev *v3d);
562 
563 /* v3d_perfmon.c */
564 void v3d_perfmon_get(struct v3d_perfmon *perfmon);
565 void v3d_perfmon_put(struct v3d_perfmon *perfmon);
566 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon);
567 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
568 		      bool capture);
569 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id);
570 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv);
571 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv);
572 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
573 			     struct drm_file *file_priv);
574 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
575 			      struct drm_file *file_priv);
576 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
577 				 struct drm_file *file_priv);
578 
579 /* v3d_sysfs.c */
580 int v3d_sysfs_init(struct device *dev);
581 void v3d_sysfs_destroy(struct device *dev);
582