1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2015-2018 Broadcom */
4 #include <linux/delay.h>
5 #include <linux/mutex.h>
6 #include <linux/spinlock_types.h>
7 #include <linux/workqueue.h>
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>
14 #include "uapi/drm/v3d_drm.h"
17 struct platform_device;
20 #define GMP_GRANULARITY (128 * 1024)
22 #define V3D_MMU_PAGE_SHIFT 12
24 #define V3D_MAX_QUEUES (V3D_CPU + 1)
26 static inline char *v3d_queue_to_string(enum v3d_queue 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";
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.
52 struct v3d_queue_state {
53 struct drm_gpu_scheduler sched;
58 /* Stores the GPU stats for this queue in the global context. */
59 struct v3d_stats stats;
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.
69 /* Tracks the number of users of the perfmon, when this counter reaches
70 * zero the perfmon is destroyed.
74 /* Protects perfmon stop, as it can be invoked from multiple places. */
77 /* Number of counters activated in this perfmon instance
78 * (should be less than DRM_V3D_MAX_PERF_COUNTERS).
82 /* Events counted by the HW perf counters. */
83 u8 counters[DRM_V3D_MAX_PERF_COUNTERS];
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.
93 u64 values[] __counted_by(ncounters);
97 struct drm_device drm;
99 /* Short representation (e.g. 33, 41) of the V3D tech version
103 bool single_irq_line;
105 void __iomem *hub_regs;
106 void __iomem *core_regs[3];
107 void __iomem *bridge_regs;
108 void __iomem *gca_regs;
110 struct reset_control *reset;
112 /* Virtual and DMA addresses of the single shared page table. */
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
121 dma_addr_t mmu_scratch_paddr;
122 /* virtual address bits from V3D to the MMU. */
125 /* Number of V3D cores. */
128 /* Allocator managing the address space. All units are in
134 struct work_struct overflow_mem_work;
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;
142 struct v3d_queue_state queue[V3D_MAX_QUEUES];
144 /* Spinlock used to synchronize the overflow memory
145 * management against bin job submission.
149 /* Used to track the active perfmon if any. */
150 struct v3d_perfmon *active_perfmon;
152 /* Protects bo_stats */
153 struct mutex bo_lock;
155 /* Lock taken when resetting the GPU, to keep multiple
156 * processes from trying to park the scheduler threads and
159 struct mutex reset_lock;
161 /* Lock taken when creating and pushing the GPU scheduler
162 * jobs, to keep the sched-fence seqnos in order.
164 struct mutex sched_lock;
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.
170 struct mutex cache_clean_lock;
178 static inline struct v3d_dev *
179 to_v3d_dev(struct drm_device *dev)
181 return container_of(dev, struct v3d_dev, drm);
185 v3d_has_csd(struct v3d_dev *v3d)
187 return v3d->ver >= 41;
190 #define v3d_to_pdev(v3d) to_platform_device((v3d)->drm.dev)
192 /* The per-fd struct, which tracks the MMU mappings. */
193 struct v3d_file_priv {
201 struct drm_sched_entity sched_entity[V3D_MAX_QUEUES];
203 /* Stores the GPU stats for a specific queue for this fd. */
204 struct v3d_stats stats[V3D_MAX_QUEUES];
208 struct drm_gem_shmem_object base;
210 struct drm_mm_node node;
212 /* List entry for the BO's position in
213 * v3d_render_job->unref_list
215 struct list_head unref_head;
220 static inline struct v3d_bo *
221 to_v3d_bo(struct drm_gem_object *bo)
223 return (struct v3d_bo *)bo;
227 struct dma_fence base;
228 struct drm_device *dev;
229 /* v3d seqno for signaled() test */
231 enum v3d_queue queue;
234 static inline struct v3d_fence *
235 to_v3d_fence(struct dma_fence *fence)
237 return (struct v3d_fence *)fence;
240 #define V3D_READ(offset) readl(v3d->hub_regs + offset)
241 #define V3D_WRITE(offset, val) writel(val, v3d->hub_regs + offset)
243 #define V3D_BRIDGE_READ(offset) readl(v3d->bridge_regs + offset)
244 #define V3D_BRIDGE_WRITE(offset, val) writel(val, v3d->bridge_regs + offset)
246 #define V3D_GCA_READ(offset) readl(v3d->gca_regs + offset)
247 #define V3D_GCA_WRITE(offset, val) writel(val, v3d->gca_regs + offset)
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)
253 struct drm_sched_job base;
255 struct kref refcount;
259 /* This is the array of BOs that were looked up at the start
262 struct drm_gem_object **bo;
265 /* v3d fence to be signaled by IRQ handler when the job is complete. */
266 struct dma_fence *irq_fence;
268 /* scheduler fence for when the job is considered complete and
269 * the BO reservations can be released.
271 struct dma_fence *done_fence;
273 /* Pointer to a performance monitor object if the user requested it,
276 struct v3d_perfmon *perfmon;
278 /* File descriptor of the process that submitted the job that could be used
279 * for collecting stats by process of GPU usage.
281 struct drm_file *file;
283 /* Callback for the freeing of the job on refcount going to 0. */
284 void (*free)(struct kref *ref);
290 /* GPU virtual addresses of the start/end of the CL job. */
293 u32 timedout_ctca, timedout_ctra;
295 /* Corresponding render job, for attaching our overflow memory. */
296 struct v3d_render_job *render;
298 /* Submitted tile memory allocation start/size, tile state. */
302 struct v3d_render_job {
305 /* GPU virtual addresses of the start/end of the CL job. */
308 u32 timedout_ctca, timedout_ctra;
310 /* List of overflow BOs used in the job that need to be
311 * released once the job is complete.
313 struct list_head unref_list;
319 struct drm_v3d_submit_tfu args;
325 u32 timedout_batches;
327 struct drm_v3d_submit_csd args;
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,
339 struct v3d_timestamp_query {
340 /* Offset of this query in the timestamp BO for its value. */
343 /* Syncobj that indicates the timestamp availability */
344 struct drm_syncobj *syncobj;
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)
351 struct v3d_performance_query {
352 /* Performance monitor IDs for this query */
353 u32 kperfmon_ids[V3D_MAX_PERFMONS];
355 /* Syncobj that indicates the query availability */
356 struct drm_syncobj *syncobj;
359 struct v3d_indirect_csd_info {
361 struct v3d_csd_job *job;
363 /* Clean cache job associated to the Indirect CSD job */
364 struct v3d_job *clean_job;
366 /* Offset within the BO where the workgroup counts are stored */
369 /* Workgroups size */
372 /* Indices of the uniforms with the workgroup dispatch counts
373 * in the uniform stream.
375 u32 wg_uniform_offsets[3];
378 struct drm_gem_object *indirect;
380 /* Context of the Indirect CSD job */
381 struct ww_acquire_ctx acquire_ctx;
384 struct v3d_timestamp_query_info {
385 struct v3d_timestamp_query *queries;
390 struct v3d_performance_query_info {
391 struct v3d_performance_query *queries;
393 /* Number of performance queries */
396 /* Number of performance monitors related to that query pool */
399 /* Number of performance counters related to that query pool */
403 struct v3d_copy_query_results_info {
404 /* Define if should write to buffer using 64 or 32 bits */
407 /* Define if it can write to buffer even if the query is not available */
410 /* Define if it should write availability bit to buffer */
411 bool availability_bit;
413 /* Offset of the copy buffer in the BO */
416 /* Stride of the copy buffer in the BO */
423 enum v3d_cpu_job_type job_type;
425 struct v3d_indirect_csd_info indirect_csd;
427 struct v3d_timestamp_query_info timestamp_query;
429 struct v3d_copy_query_results_info copy;
431 struct v3d_performance_query_info performance_query;
434 typedef void (*v3d_cpu_job_fn)(struct v3d_cpu_job *);
436 struct v3d_submit_outsync {
437 struct drm_syncobj *syncobj;
440 struct v3d_submit_ext {
448 struct v3d_submit_outsync *out_syncs;
452 * __wait_for - magic wait macro
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.
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 */ \
465 const bool expired__ = ktime_after(ktime_get_raw(), end__); \
467 /* Guarantee COND check prior to timeout */ \
474 ret__ = -ETIMEDOUT; \
477 usleep_range(wait__, wait__ * 2); \
478 if (wait__ < (Wmax)) \
484 #define _wait_for(COND, US, Wmin, Wmax) __wait_for(, (COND), (US), (Wmin), \
486 #define wait_for(COND, MS) _wait_for((COND), (MS) * 1000, 10, 1000)
488 static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
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;
495 return min_t(u64, MAX_JIFFY_OFFSET, nsecs_to_jiffies64(n) + 1);
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,
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);
518 void v3d_debugfs_init(struct drm_minor *minor);
521 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
522 u64 *active_runtime, u64 *jobs_completed);
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);
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);
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);
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);
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);
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);
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,
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);
580 int v3d_sysfs_init(struct device *dev);
581 void v3d_sysfs_destroy(struct device *dev);