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 /* Currently active job for this queue */
63 struct v3d_job *active_job;
64 spinlock_t queue_lock;
65 /* Protect dma fence for signalling job completion */
66 spinlock_t fence_lock;
67 };
68
69 /* Performance monitor object. The perform lifetime is controlled by userspace
70 * using perfmon related ioctls. A perfmon can be attached to a submit_cl
71 * request, and when this is the case, HW perf counters will be activated just
72 * before the submit_cl is submitted to the GPU and disabled when the job is
73 * done. This way, only events related to a specific job will be counted.
74 */
75 struct v3d_perfmon {
76 /* Tracks the number of users of the perfmon, when this counter reaches
77 * zero the perfmon is destroyed.
78 */
79 refcount_t refcnt;
80
81 /* Protects perfmon stop, as it can be invoked from multiple places. */
82 struct mutex lock;
83
84 /* Number of counters activated in this perfmon instance
85 * (should be less than DRM_V3D_MAX_PERF_COUNTERS).
86 */
87 u8 ncounters;
88
89 /* Events counted by the HW perf counters. */
90 u8 counters[DRM_V3D_MAX_PERF_COUNTERS];
91
92 /* Storage for counter values. Counters are incremented by the
93 * HW perf counter values every time the perfmon is attached
94 * to a GPU job. This way, perfmon users don't have to
95 * retrieve the results after each job if they want to track
96 * events covering several submissions. Note that counter
97 * values can't be reset, but you can fake a reset by
98 * destroying the perfmon and creating a new one.
99 */
100 u64 values[] __counted_by(ncounters);
101 };
102
103 enum v3d_gen {
104 V3D_GEN_33 = 33,
105 V3D_GEN_41 = 41,
106 V3D_GEN_42 = 42,
107 V3D_GEN_71 = 71,
108 };
109
110 enum v3d_irq {
111 V3D_CORE_IRQ,
112 V3D_HUB_IRQ,
113 V3D_MAX_IRQS,
114 };
115
116 struct v3d_dev {
117 struct drm_device drm;
118
119 /* Short representation (e.g. 33, 41) of the V3D tech version */
120 enum v3d_gen ver;
121
122 /* Short representation (e.g. 5, 6) of the V3D tech revision */
123 int rev;
124
125 bool single_irq_line;
126
127 int irq[V3D_MAX_IRQS];
128
129 struct v3d_perfmon_info perfmon_info;
130
131 void __iomem *hub_regs;
132 void __iomem *core_regs[3];
133 void __iomem *bridge_regs;
134 void __iomem *gca_regs;
135 void __iomem *sms_regs;
136 struct clk *clk;
137 struct reset_control *reset;
138
139 /* Virtual and DMA addresses of the single shared page table. */
140 volatile u32 *pt;
141 dma_addr_t pt_paddr;
142
143 /* Virtual and DMA addresses of the MMU's scratch page. When
144 * a read or write is invalid in the MMU, it will be
145 * redirected here.
146 */
147 void *mmu_scratch;
148 dma_addr_t mmu_scratch_paddr;
149 /* virtual address bits from V3D to the MMU. */
150 int va_width;
151
152 /* Number of V3D cores. */
153 u32 cores;
154
155 /* Allocator managing the address space. All units are in
156 * number of pages.
157 */
158 struct drm_mm mm;
159 spinlock_t mm_lock;
160
161 /*
162 * tmpfs instance used for shmem backed objects
163 */
164 struct vfsmount *gemfs;
165
166 struct work_struct overflow_mem_work;
167
168 struct v3d_queue_state queue[V3D_MAX_QUEUES];
169
170 /* Used to track the active perfmon if any. */
171 struct v3d_perfmon *active_perfmon;
172
173 /* Protects bo_stats */
174 struct mutex bo_lock;
175
176 /* Lock taken when resetting the GPU, to keep multiple
177 * processes from trying to park the scheduler threads and
178 * reset at once.
179 */
180 struct mutex reset_lock;
181
182 /* Lock taken when creating and pushing the GPU scheduler
183 * jobs, to keep the sched-fence seqnos in order.
184 */
185 struct mutex sched_lock;
186
187 /* Lock taken during a cache clean and when initiating an L2
188 * flush, to keep L2 flushes from interfering with the
189 * synchronous L2 cleans.
190 */
191 struct mutex cache_clean_lock;
192
193 struct {
194 u32 num_allocated;
195 u32 pages_allocated;
196 } bo_stats;
197
198 /* To support a performance analysis tool in user space, we require
199 * a single, globally configured performance monitor (perfmon) for
200 * all jobs.
201 */
202 struct v3d_perfmon *global_perfmon;
203
204 /* Global reset counter. The counter must be incremented when
205 * a GPU reset happens. It must be protected by @reset_lock.
206 */
207 unsigned int reset_counter;
208 };
209
210 static inline struct v3d_dev *
to_v3d_dev(struct drm_device * dev)211 to_v3d_dev(struct drm_device *dev)
212 {
213 return container_of(dev, struct v3d_dev, drm);
214 }
215
216 static inline bool
v3d_has_csd(struct v3d_dev * v3d)217 v3d_has_csd(struct v3d_dev *v3d)
218 {
219 return v3d->ver >= V3D_GEN_41;
220 }
221
222 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
223
224 /* The per-fd struct, which tracks the MMU mappings. */
225 struct v3d_file_priv {
226 struct v3d_dev *v3d;
227
228 struct {
229 struct idr idr;
230 struct mutex lock;
231 } perfmon;
232
233 struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
234
235 /* Stores the GPU stats for a specific queue for this fd. */
236 struct v3d_stats stats[V3D_MAX_QUEUES];
237
238 /* Per-fd reset counter, must be incremented when a job submitted
239 * by this fd causes a GPU reset. It must be protected by
240 * &struct v3d_dev->reset_lock.
241 */
242 unsigned int reset_counter;
243 };
244
245 struct v3d_bo {
246 struct drm_gem_shmem_object base;
247
248 struct drm_mm_node node;
249
250 /* List entry for the BO's position in
251 * v3d_render_job->unref_list
252 */
253 struct list_head unref_head;
254
255 void *vaddr;
256 };
257
258 static inline struct v3d_bo *
to_v3d_bo(struct drm_gem_object * bo)259 to_v3d_bo(struct drm_gem_object *bo)
260 {
261 return (struct v3d_bo *)bo;
262 }
263
264 struct v3d_fence {
265 struct dma_fence base;
266 struct drm_device *dev;
267 /* v3d seqno for signaled() test */
268 u64 seqno;
269 enum v3d_queue queue;
270 };
271
272 static inline struct v3d_fence *
to_v3d_fence(struct dma_fence * fence)273 to_v3d_fence(struct dma_fence *fence)
274 {
275 return (struct v3d_fence *)fence;
276 }
277
278 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
279 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
280
281 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
282 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
283
284 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
285 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
286
287 #define V3D_SMS_IDLE 0x0
288 #define V3D_SMS_ISOLATING_FOR_RESET 0xa
289 #define V3D_SMS_RESETTING 0xb
290 #define V3D_SMS_ISOLATING_FOR_POWER_OFF 0xc
291 #define V3D_SMS_POWER_OFF_STATE 0xd
292
293 #define V3D_SMS_READ(offset) readl(v3d->sms_regs + (offset))
294 #define V3D_SMS_WRITE(offset, val) writel(val, v3d->sms_regs + (offset))
295
296 #define V3D_CORE_READ(core, offset) readl(v3d->core_regs[core] + offset)
297 #define V3D_CORE_WRITE(core, offset, val) writel(val, v3d->core_regs[core] + offset)
298
299 struct v3d_job {
300 struct drm_sched_job base;
301
302 struct kref refcount;
303
304 struct v3d_dev *v3d;
305
306 /* This is the array of BOs that were looked up at the start
307 * of submission.
308 */
309 struct drm_gem_object **bo;
310 u32 bo_count;
311
312 /* v3d fence to be signaled by IRQ handler when the job is complete. */
313 struct dma_fence *irq_fence;
314
315 /* scheduler fence for when the job is considered complete and
316 * the BO reservations can be released.
317 */
318 struct dma_fence *done_fence;
319
320 /* Pointer to a performance monitor object if the user requested it,
321 * NULL otherwise.
322 */
323 struct v3d_perfmon *perfmon;
324
325 /* File descriptor of the process that submitted the job that could be used
326 * to collect per-process information about the GPU.
327 */
328 struct v3d_file_priv *file_priv;
329
330 /* Callback for the freeing of the job on refcount going to 0. */
331 void (*free)(struct kref *ref);
332 };
333
334 struct v3d_bin_job {
335 struct v3d_job base;
336
337 /* GPU virtual addresses of the start/end of the CL job. */
338 u32 start, end;
339
340 u32 timedout_ctca, timedout_ctra;
341
342 /* Corresponding render job, for attaching our overflow memory. */
343 struct v3d_render_job *render;
344
345 /* Submitted tile memory allocation start/size, tile state. */
346 u32 qma, qms, qts;
347 };
348
349 struct v3d_render_job {
350 struct v3d_job base;
351
352 /* GPU virtual addresses of the start/end of the CL job. */
353 u32 start, end;
354
355 u32 timedout_ctca, timedout_ctra;
356
357 /* List of overflow BOs used in the job that need to be
358 * released once the job is complete.
359 */
360 struct list_head unref_list;
361 };
362
363 struct v3d_tfu_job {
364 struct v3d_job base;
365
366 struct drm_v3d_submit_tfu args;
367 };
368
369 struct v3d_csd_job {
370 struct v3d_job base;
371
372 u32 timedout_batches;
373
374 struct drm_v3d_submit_csd args;
375 };
376
377 enum v3d_cpu_job_type {
378 V3D_CPU_JOB_TYPE_INDIRECT_CSD = 1,
379 V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY,
380 V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY,
381 V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY,
382 V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY,
383 V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY,
384 };
385
386 struct v3d_timestamp_query {
387 /* Offset of this query in the timestamp BO for its value. */
388 u32 offset;
389
390 /* Syncobj that indicates the timestamp availability */
391 struct drm_syncobj *syncobj;
392 };
393
394 struct v3d_performance_query {
395 /* Performance monitor IDs for this query */
396 u32 *kperfmon_ids;
397
398 /* Syncobj that indicates the query availability */
399 struct drm_syncobj *syncobj;
400 };
401
402 struct v3d_indirect_csd_info {
403 /* Indirect CSD */
404 struct v3d_csd_job *job;
405
406 /* Clean cache job associated to the Indirect CSD job */
407 struct v3d_job *clean_job;
408
409 /* Offset within the BO where the workgroup counts are stored */
410 u32 offset;
411
412 /* Workgroups size */
413 u32 wg_size;
414
415 /* Indices of the uniforms with the workgroup dispatch counts
416 * in the uniform stream.
417 */
418 u32 wg_uniform_offsets[3];
419
420 /* Indirect BO */
421 struct drm_gem_object *indirect;
422
423 /* Context of the Indirect CSD job */
424 struct ww_acquire_ctx acquire_ctx;
425 };
426
427 struct v3d_timestamp_query_info {
428 struct v3d_timestamp_query *queries;
429
430 u32 count;
431 };
432
433 struct v3d_performance_query_info {
434 struct v3d_performance_query *queries;
435
436 /* Number of performance queries */
437 u32 count;
438
439 /* Number of performance monitors related to that query pool */
440 u32 nperfmons;
441
442 /* Number of performance counters related to that query pool */
443 u32 ncounters;
444 };
445
446 struct v3d_copy_query_results_info {
447 /* Define if should write to buffer using 64 or 32 bits */
448 bool do_64bit;
449
450 /* Define if it can write to buffer even if the query is not available */
451 bool do_partial;
452
453 /* Define if it should write availability bit to buffer */
454 bool availability_bit;
455
456 /* Offset of the copy buffer in the BO */
457 u32 offset;
458
459 /* Stride of the copy buffer in the BO */
460 u32 stride;
461 };
462
463 struct v3d_cpu_job {
464 struct v3d_job base;
465
466 enum v3d_cpu_job_type job_type;
467
468 struct v3d_indirect_csd_info indirect_csd;
469
470 struct v3d_timestamp_query_info timestamp_query;
471
472 struct v3d_copy_query_results_info copy;
473
474 struct v3d_performance_query_info performance_query;
475 };
476
477 typedef void (*v3d_cpu_job_fn)(struct v3d_cpu_job *);
478
479 struct v3d_submit_outsync {
480 struct drm_syncobj *syncobj;
481 };
482
483 struct v3d_submit_ext {
484 u32 flags;
485 u32 wait_stage;
486
487 u32 in_sync_count;
488 u64 in_syncs;
489
490 u32 out_sync_count;
491 struct v3d_submit_outsync *out_syncs;
492 };
493
494 /**
495 * __wait_for - magic wait macro
496 *
497 * Macro to help avoid open coding check/wait/timeout patterns. Note that it's
498 * important that we check the condition again after having timed out, since the
499 * timeout could be due to preemption or similar and we've never had a chance to
500 * check the condition before the timeout.
501 */
502 #define __wait_for(OP, COND, US, Wmin, Wmax) ({ \
503 const ktime_t end__ = ktime_add_ns(ktime_get_raw(), 1000ll * (US)); \
504 long wait__ = (Wmin); /* recommended min for usleep is 10 us */ \
505 int ret__; \
506 might_sleep(); \
507 for (;;) { \
508 const bool expired__ = ktime_after(ktime_get_raw(), end__); \
509 OP; \
510 /* Guarantee COND check prior to timeout */ \
511 barrier(); \
512 if (COND) { \
513 ret__ = 0; \
514 break; \
515 } \
516 if (expired__) { \
517 ret__ = -ETIMEDOUT; \
518 break; \
519 } \
520 usleep_range(wait__, wait__ * 2); \
521 if (wait__ < (Wmax)) \
522 wait__ <<= 1; \
523 } \
524 ret__; \
525 })
526
527 #define _wait_for(COND, US, Wmin, Wmax) __wait_for(, (COND), (US), (Wmin), \
528 (Wmax))
529 #define wait_for(COND, MS) _wait_for((COND), (MS) * 1000, 10, 1000)
530
nsecs_to_jiffies_timeout(const u64 n)531 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
532 {
533 /* nsecs_to_jiffies64() does not guard against overflow */
534 if ((NSEC_PER_SEC % HZ) != 0 &&
535 div_u64(n, NSEC_PER_SEC) >= MAX_JIFFY_OFFSET / HZ)
536 return MAX_JIFFY_OFFSET;
537
538 return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
539 }
540
541 /* v3d_bo.c */
542 struct drm_gem_object *v3d_create_object(struct drm_device *dev, size_t size);
543 void v3d_free_object(struct drm_gem_object *gem_obj);
544 struct v3d_bo *v3d_bo_create(struct drm_device *dev, struct drm_file *file_priv,
545 size_t size);
546 void v3d_get_bo_vaddr(struct v3d_bo *bo);
547 void v3d_put_bo_vaddr(struct v3d_bo *bo);
548 int v3d_create_bo_ioctl(struct drm_device *dev, void *data,
549 struct drm_file *file_priv);
550 int v3d_mmap_bo_ioctl(struct drm_device *dev, void *data,
551 struct drm_file *file_priv);
552 int v3d_get_bo_offset_ioctl(struct drm_device *dev, void *data,
553 struct drm_file *file_priv);
554 int v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
555 struct drm_file *file_priv);
556 struct drm_gem_object *v3d_prime_import_sg_table(struct drm_device *dev,
557 struct dma_buf_attachment *attach,
558 struct sg_table *sgt);
559
560 /* v3d_debugfs.c */
561 void v3d_debugfs_init(struct drm_minor *minor);
562
563 /* v3d_drv.c */
564 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
565 u64 *active_runtime, u64 *jobs_completed);
566
567 /* v3d_fence.c */
568 extern const struct dma_fence_ops v3d_fence_ops;
569 struct dma_fence *v3d_fence_create(struct v3d_dev *v3d, enum v3d_queue q);
570
571 /* v3d_gem.c */
572 int v3d_gem_init(struct drm_device *dev);
573 void v3d_gem_destroy(struct drm_device *dev);
574 void v3d_reset_sms(struct v3d_dev *v3d);
575 void v3d_reset(struct v3d_dev *v3d);
576 void v3d_invalidate_caches(struct v3d_dev *v3d);
577 void v3d_clean_caches(struct v3d_dev *v3d);
578
579 /* v3d_gemfs.c */
580 extern bool super_pages;
581 void v3d_gemfs_init(struct v3d_dev *v3d);
582 void v3d_gemfs_fini(struct v3d_dev *v3d);
583
584 /* v3d_submit.c */
585 void v3d_job_cleanup(struct v3d_job *job);
586 void v3d_job_put(struct v3d_job *job);
587 int v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
588 struct drm_file *file_priv);
589 int v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
590 struct drm_file *file_priv);
591 int v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
592 struct drm_file *file_priv);
593 int v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
594 struct drm_file *file_priv);
595
596 /* v3d_irq.c */
597 int v3d_irq_init(struct v3d_dev *v3d);
598 void v3d_irq_enable(struct v3d_dev *v3d);
599 void v3d_irq_disable(struct v3d_dev *v3d);
600 void v3d_irq_reset(struct v3d_dev *v3d);
601
602 /* v3d_mmu.c */
603 int v3d_mmu_flush_all(struct v3d_dev *v3d);
604 int v3d_mmu_set_page_table(struct v3d_dev *v3d);
605 void v3d_mmu_insert_ptes(struct v3d_bo *bo);
606 void v3d_mmu_remove_ptes(struct v3d_bo *bo);
607
608 /* v3d_sched.c */
609 void v3d_timestamp_query_info_free(struct v3d_timestamp_query_info *query_info,
610 unsigned int count);
611 void v3d_performance_query_info_free(struct v3d_performance_query_info *query_info,
612 unsigned int count);
613 void v3d_job_update_stats(struct v3d_job *job, enum v3d_queue q);
614 int v3d_sched_init(struct v3d_dev *v3d);
615 void v3d_sched_fini(struct v3d_dev *v3d);
616
617 /* v3d_perfmon.c */
618 void v3d_perfmon_init(struct v3d_dev *v3d);
619 void v3d_perfmon_get(struct v3d_perfmon *perfmon);
620 void v3d_perfmon_put(struct v3d_perfmon *perfmon);
621 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon);
622 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
623 bool capture);
624 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id);
625 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv);
626 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv);
627 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
628 struct drm_file *file_priv);
629 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
630 struct drm_file *file_priv);
631 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
632 struct drm_file *file_priv);
633 int v3d_perfmon_get_counter_ioctl(struct drm_device *dev, void *data,
634 struct drm_file *file_priv);
635 int v3d_perfmon_set_global_ioctl(struct drm_device *dev, void *data,
636 struct drm_file *file_priv);
637
638 /* v3d_sysfs.c */
639 int v3d_sysfs_init(struct device *dev);
640 void v3d_sysfs_destroy(struct device *dev);
641