xref: /linux/drivers/gpu/drm/v3d/v3d_drv.h (revision 92d6295a29dba56148406a8452c69ab49787741b)
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 
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 enum v3d_gen {
98 	V3D_GEN_33 = 33,
99 	V3D_GEN_41 = 41,
100 	V3D_GEN_42 = 42,
101 	V3D_GEN_71 = 71,
102 };
103 
104 enum v3d_irq {
105 	V3D_CORE_IRQ,
106 	V3D_HUB_IRQ,
107 	V3D_MAX_IRQS,
108 };
109 
110 struct v3d_dev {
111 	struct drm_device drm;
112 
113 	/* Short representation (e.g. 33, 41) of the V3D tech version */
114 	enum v3d_gen ver;
115 
116 	/* Short representation (e.g. 5, 6) of the V3D tech revision */
117 	int rev;
118 
119 	bool single_irq_line;
120 
121 	int irq[V3D_MAX_IRQS];
122 
123 	struct v3d_perfmon_info perfmon_info;
124 
125 	void __iomem *hub_regs;
126 	void __iomem *core_regs[3];
127 	void __iomem *bridge_regs;
128 	void __iomem *gca_regs;
129 	void __iomem *sms_regs;
130 	struct clk *clk;
131 	struct reset_control *reset;
132 
133 	/* Virtual and DMA addresses of the single shared page table. */
134 	volatile u32 *pt;
135 	dma_addr_t pt_paddr;
136 
137 	/* Virtual and DMA addresses of the MMU's scratch page.  When
138 	 * a read or write is invalid in the MMU, it will be
139 	 * redirected here.
140 	 */
141 	void *mmu_scratch;
142 	dma_addr_t mmu_scratch_paddr;
143 	/* virtual address bits from V3D to the MMU. */
144 	int va_width;
145 
146 	/* Number of V3D cores. */
147 	u32 cores;
148 
149 	/* Allocator managing the address space.  All units are in
150 	 * number of pages.
151 	 */
152 	struct drm_mm mm;
153 	spinlock_t mm_lock;
154 
155 	/*
156 	 * tmpfs instance used for shmem backed objects
157 	 */
158 	struct vfsmount *gemfs;
159 
160 	struct work_struct overflow_mem_work;
161 
162 	struct v3d_bin_job *bin_job;
163 	struct v3d_render_job *render_job;
164 	struct v3d_tfu_job *tfu_job;
165 	struct v3d_csd_job *csd_job;
166 
167 	struct v3d_queue_state queue[V3D_MAX_QUEUES];
168 
169 	/* Spinlock used to synchronize the overflow memory
170 	 * management against bin job submission.
171 	 */
172 	spinlock_t job_lock;
173 
174 	/* Used to track the active perfmon if any. */
175 	struct v3d_perfmon *active_perfmon;
176 
177 	/* Protects bo_stats */
178 	struct mutex bo_lock;
179 
180 	/* Lock taken when resetting the GPU, to keep multiple
181 	 * processes from trying to park the scheduler threads and
182 	 * reset at once.
183 	 */
184 	struct mutex reset_lock;
185 
186 	/* Lock taken when creating and pushing the GPU scheduler
187 	 * jobs, to keep the sched-fence seqnos in order.
188 	 */
189 	struct mutex sched_lock;
190 
191 	/* Lock taken during a cache clean and when initiating an L2
192 	 * flush, to keep L2 flushes from interfering with the
193 	 * synchronous L2 cleans.
194 	 */
195 	struct mutex cache_clean_lock;
196 
197 	struct {
198 		u32 num_allocated;
199 		u32 pages_allocated;
200 	} bo_stats;
201 
202 	/* To support a performance analysis tool in user space, we require
203 	 * a single, globally configured performance monitor (perfmon) for
204 	 * all jobs.
205 	 */
206 	struct v3d_perfmon *global_perfmon;
207 
208 	/* Global reset counter. The counter must be incremented when
209 	 * a GPU reset happens. It must be protected by @reset_lock.
210 	 */
211 	unsigned int reset_counter;
212 };
213 
214 static inline struct v3d_dev *
215 to_v3d_dev(struct drm_device *dev)
216 {
217 	return container_of(dev, struct v3d_dev, drm);
218 }
219 
220 static inline bool
221 v3d_has_csd(struct v3d_dev *v3d)
222 {
223 	return v3d->ver >= V3D_GEN_41;
224 }
225 
226 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
227 
228 /* The per-fd struct, which tracks the MMU mappings. */
229 struct v3d_file_priv {
230 	struct v3d_dev *v3d;
231 
232 	struct {
233 		struct idr idr;
234 		struct mutex lock;
235 	} perfmon;
236 
237 	struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
238 
239 	/* Stores the GPU stats for a specific queue for this fd. */
240 	struct v3d_stats stats[V3D_MAX_QUEUES];
241 
242 	/* Per-fd reset counter, must be incremented when a job submitted
243 	 * by this fd causes a GPU reset. It must be protected by
244 	 * &struct v3d_dev->reset_lock.
245 	 */
246 	unsigned int reset_counter;
247 };
248 
249 struct v3d_bo {
250 	struct drm_gem_shmem_object base;
251 
252 	struct drm_mm_node node;
253 
254 	/* List entry for the BO's position in
255 	 * v3d_render_job->unref_list
256 	 */
257 	struct list_head unref_head;
258 
259 	void *vaddr;
260 };
261 
262 static inline struct v3d_bo *
263 to_v3d_bo(struct drm_gem_object *bo)
264 {
265 	return (struct v3d_bo *)bo;
266 }
267 
268 struct v3d_fence {
269 	struct dma_fence base;
270 	struct drm_device *dev;
271 	/* v3d seqno for signaled() test */
272 	u64 seqno;
273 	enum v3d_queue queue;
274 };
275 
276 static inline struct v3d_fence *
277 to_v3d_fence(struct dma_fence *fence)
278 {
279 	return (struct v3d_fence *)fence;
280 }
281 
282 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
283 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
284 
285 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
286 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
287 
288 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
289 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
290 
291 #define V3D_SMS_IDLE				0x0
292 #define V3D_SMS_ISOLATING_FOR_RESET		0xa
293 #define V3D_SMS_RESETTING			0xb
294 #define V3D_SMS_ISOLATING_FOR_POWER_OFF	0xc
295 #define V3D_SMS_POWER_OFF_STATE		0xd
296 
297 #define V3D_SMS_READ(offset) readl(v3d->sms_regs + (offset))
298 #define V3D_SMS_WRITE(offset, val) writel(val, v3d->sms_regs + (offset))
299 
300 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
301 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
302 
303 struct v3d_job {
304 	struct drm_sched_job base;
305 
306 	struct kref refcount;
307 
308 	struct v3d_dev *v3d;
309 
310 	/* This is the array of BOs that were looked up at the start
311 	 * of submission.
312 	 */
313 	struct drm_gem_object **bo;
314 	u32 bo_count;
315 
316 	/* v3d fence to be signaled by IRQ handler when the job is complete. */
317 	struct dma_fence *irq_fence;
318 
319 	/* scheduler fence for when the job is considered complete and
320 	 * the BO reservations can be released.
321 	 */
322 	struct dma_fence *done_fence;
323 
324 	/* Pointer to a performance monitor object if the user requested it,
325 	 * NULL otherwise.
326 	 */
327 	struct v3d_perfmon *perfmon;
328 
329 	/* File descriptor of the process that submitted the job that could be used
330 	 * for collecting stats by process of GPU usage.
331 	 */
332 	struct drm_file *file;
333 
334 	/* Callback for the freeing of the job on refcount going to 0. */
335 	void (*free)(struct kref *ref);
336 };
337 
338 struct v3d_bin_job {
339 	struct v3d_job base;
340 
341 	/* GPU virtual addresses of the start/end of the CL job. */
342 	u32 start, end;
343 
344 	u32 timedout_ctca, timedout_ctra;
345 
346 	/* Corresponding render job, for attaching our overflow memory. */
347 	struct v3d_render_job *render;
348 
349 	/* Submitted tile memory allocation start/size, tile state. */
350 	u32 qma, qms, qts;
351 };
352 
353 struct v3d_render_job {
354 	struct v3d_job base;
355 
356 	/* GPU virtual addresses of the start/end of the CL job. */
357 	u32 start, end;
358 
359 	u32 timedout_ctca, timedout_ctra;
360 
361 	/* List of overflow BOs used in the job that need to be
362 	 * released once the job is complete.
363 	 */
364 	struct list_head unref_list;
365 };
366 
367 struct v3d_tfu_job {
368 	struct v3d_job base;
369 
370 	struct drm_v3d_submit_tfu args;
371 };
372 
373 struct v3d_csd_job {
374 	struct v3d_job base;
375 
376 	u32 timedout_batches;
377 
378 	struct drm_v3d_submit_csd args;
379 };
380 
381 enum v3d_cpu_job_type {
382 	V3D_CPU_JOB_TYPE_INDIRECT_CSD = 1,
383 	V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY,
384 	V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY,
385 	V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY,
386 	V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY,
387 	V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY,
388 };
389 
390 struct v3d_timestamp_query {
391 	/* Offset of this query in the timestamp BO for its value. */
392 	u32 offset;
393 
394 	/* Syncobj that indicates the timestamp availability */
395 	struct drm_syncobj *syncobj;
396 };
397 
398 struct v3d_performance_query {
399 	/* Performance monitor IDs for this query */
400 	u32 *kperfmon_ids;
401 
402 	/* Syncobj that indicates the query availability */
403 	struct drm_syncobj *syncobj;
404 };
405 
406 struct v3d_indirect_csd_info {
407 	/* Indirect CSD */
408 	struct v3d_csd_job *job;
409 
410 	/* Clean cache job associated to the Indirect CSD job */
411 	struct v3d_job *clean_job;
412 
413 	/* Offset within the BO where the workgroup counts are stored */
414 	u32 offset;
415 
416 	/* Workgroups size */
417 	u32 wg_size;
418 
419 	/* Indices of the uniforms with the workgroup dispatch counts
420 	 * in the uniform stream.
421 	 */
422 	u32 wg_uniform_offsets[3];
423 
424 	/* Indirect BO */
425 	struct drm_gem_object *indirect;
426 
427 	/* Context of the Indirect CSD job */
428 	struct ww_acquire_ctx acquire_ctx;
429 };
430 
431 struct v3d_timestamp_query_info {
432 	struct v3d_timestamp_query *queries;
433 
434 	u32 count;
435 };
436 
437 struct v3d_performance_query_info {
438 	struct v3d_performance_query *queries;
439 
440 	/* Number of performance queries */
441 	u32 count;
442 
443 	/* Number of performance monitors related to that query pool */
444 	u32 nperfmons;
445 
446 	/* Number of performance counters related to that query pool */
447 	u32 ncounters;
448 };
449 
450 struct v3d_copy_query_results_info {
451 	/* Define if should write to buffer using 64 or 32 bits */
452 	bool do_64bit;
453 
454 	/* Define if it can write to buffer even if the query is not available */
455 	bool do_partial;
456 
457 	/* Define if it should write availability bit to buffer */
458 	bool availability_bit;
459 
460 	/* Offset of the copy buffer in the BO */
461 	u32 offset;
462 
463 	/* Stride of the copy buffer in the BO */
464 	u32 stride;
465 };
466 
467 struct v3d_cpu_job {
468 	struct v3d_job base;
469 
470 	enum v3d_cpu_job_type job_type;
471 
472 	struct v3d_indirect_csd_info indirect_csd;
473 
474 	struct v3d_timestamp_query_info timestamp_query;
475 
476 	struct v3d_copy_query_results_info copy;
477 
478 	struct v3d_performance_query_info performance_query;
479 };
480 
481 typedef void (*v3d_cpu_job_fn)(struct v3d_cpu_job *);
482 
483 struct v3d_submit_outsync {
484 	struct drm_syncobj *syncobj;
485 };
486 
487 struct v3d_submit_ext {
488 	u32 flags;
489 	u32 wait_stage;
490 
491 	u32 in_sync_count;
492 	u64 in_syncs;
493 
494 	u32 out_sync_count;
495 	struct v3d_submit_outsync *out_syncs;
496 };
497 
498 /**
499  * __wait_for - magic wait macro
500  *
501  * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
502  * important that we check the condition again after having timed out, since the
503  * timeout could be due to preemption or similar and we've never had a chance to
504  * check the condition before the timeout.
505  */
506 #define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
507 	const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
508 	long wait__ = (Wmin); /* recommended min for usleep is 10 us */	\
509 	int ret__;							\
510 	might_sleep();							\
511 	for (;;) {							\
512 		const bool expired__ = ktime_after(ktime_get_raw(), end__); \
513 		OP;							\
514 		/* Guarantee COND check prior to timeout */		\
515 		barrier();						\
516 		if (COND) {						\
517 			ret__ = 0;					\
518 			break;						\
519 		}							\
520 		if (expired__) {					\
521 			ret__ = -ETIMEDOUT;				\
522 			break;						\
523 		}							\
524 		usleep_range(wait__, wait__ * 2);			\
525 		if (wait__ < (Wmax))					\
526 			wait__ <<= 1;					\
527 	}								\
528 	ret__;								\
529 })
530 
531 #define _wait_for(COND, US, Wmin, Wmax)	__wait_for(, (COND), (US), (Wmin), \
532 						   (Wmax))
533 #define wait_for(COND, MS)		_wait_for((COND), (MS) * 1000, 10, 1000)
534 
535 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
536 {
537 	/* nsecs_to_jiffies64() does not guard against overflow */
538 	if ((NSEC_PER_SEC % HZ) != 0 &&
539 	    div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
540 		return MAX_JIFFY_OFFSET;
541 
542 	return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
543 }
544 
545 /* v3d_bo.c */
546 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size);
547 void v3d_free_object(struct drm_gem_object *gem_obj);
548 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
549 			     size_t size);
550 void v3d_get_bo_vaddr(struct v3d_bo *bo);
551 void v3d_put_bo_vaddr(struct v3d_bo *bo);
552 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
553 			struct drm_file *file_priv);
554 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
555 		      struct drm_file *file_priv);
556 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
557 			    struct drm_file *file_priv);
558 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
559 		      struct drm_file *file_priv);
560 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
561 						 struct dma_buf_attachment *attach,
562 						 struct sg_table *sgt);
563 
564 /* v3d_debugfs.c */
565 void v3d_debugfs_init(struct drm_minor *minor);
566 
567 /* v3d_drv.c */
568 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
569 		   u64 *active_runtime, u64 *jobs_completed);
570 
571 /* v3d_fence.c */
572 extern const struct dma_fence_ops v3d_fence_ops;
573 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue queue);
574 
575 /* v3d_gem.c */
576 int v3d_gem_init(struct drm_device *dev);
577 void v3d_gem_destroy(struct drm_device *dev);
578 void v3d_reset_sms(struct v3d_dev *v3d);
579 void v3d_reset(struct v3d_dev *v3d);
580 void v3d_invalidate_caches(struct v3d_dev *v3d);
581 void v3d_clean_caches(struct v3d_dev *v3d);
582 
583 /* v3d_gemfs.c */
584 extern bool super_pages;
585 void v3d_gemfs_init(struct v3d_dev *v3d);
586 void v3d_gemfs_fini(struct v3d_dev *v3d);
587 
588 /* v3d_submit.c */
589 void v3d_job_cleanup(struct v3d_job *job);
590 void v3d_job_put(struct v3d_job *job);
591 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
592 			struct drm_file *file_priv);
593 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
594 			 struct drm_file *file_priv);
595 int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
596 			 struct drm_file *file_priv);
597 int v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
598 			 struct drm_file *file_priv);
599 
600 /* v3d_irq.c */
601 int v3d_irq_init(struct v3d_dev *v3d);
602 void v3d_irq_enable(struct v3d_dev *v3d);
603 void v3d_irq_disable(struct v3d_dev *v3d);
604 void v3d_irq_reset(struct v3d_dev *v3d);
605 
606 /* v3d_mmu.c */
607 int v3d_mmu_flush_all(struct v3d_dev *v3d);
608 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
609 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
610 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
611 
612 /* v3d_sched.c */
613 void v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
614 				   unsigned int count);
615 void v3d_performance_query_info_free(struct v3d_performance_query_info *query_info,
616 				     unsigned int count);
617 void v3d_job_update_stats(struct v3d_job *job, enum v3d_queue queue);
618 int v3d_sched_init(struct v3d_dev *v3d);
619 void v3d_sched_fini(struct v3d_dev *v3d);
620 
621 /* v3d_perfmon.c */
622 void v3d_perfmon_init(struct v3d_dev *v3d);
623 void v3d_perfmon_get(struct v3d_perfmon *perfmon);
624 void v3d_perfmon_put(struct v3d_perfmon *perfmon);
625 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon);
626 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
627 		      bool capture);
628 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id);
629 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv);
630 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv);
631 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
632 			     struct drm_file *file_priv);
633 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
634 			      struct drm_file *file_priv);
635 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
636 				 struct drm_file *file_priv);
637 int v3d_perfmon_get_counter_ioctl(struct drm_device *dev, void *data,
638 				  struct drm_file *file_priv);
639 int v3d_perfmon_set_global_ioctl(struct drm_device *dev, void *data,
640 				 struct drm_file *file_priv);
641 
642 /* v3d_sysfs.c */
643 int v3d_sysfs_init(struct device *dev);
644 void v3d_sysfs_destroy(struct device *dev);
645