1 /* SPDX-License-Identifier: GPL-2.0-only */
3 * Copyright (C) 2013 Red Hat
10 #include <linux/adreno-smmu-priv.h>
11 #include <linux/clk.h>
12 #include <linux/devfreq.h>
13 #include <linux/interconnect.h>
14 #include <linux/pm_opp.h>
15 #include <linux/regulator/consumer.h>
16 #include <linux/reset.h>
19 #include "msm_fence.h"
20 #include "msm_ringbuffer.h"
23 struct msm_gem_submit;
24 struct msm_gpu_perfcntr;
26 struct msm_file_private;
28 struct msm_gpu_config {
30 unsigned int nr_rings;
33 /* So far, with hardware that I've seen to date, we can have:
34 * + zero, one, or two z180 2d cores
35 * + a3xx or a2xx 3d core, which share a common CP (the firmware
36 * for the CP seems to implement some different PM4 packet types
37 * but the basics of cmdstream submission are the same)
39 * Which means that the eventual complete "class" hierarchy, once
40 * support for all past and present hw is in place, becomes:
47 struct msm_gpu_funcs {
48 int (*get_param)(struct msm_gpu *gpu, struct msm_file_private *ctx,
49 uint32_t param, uint64_t *value, uint32_t *len);
50 int (*set_param)(struct msm_gpu *gpu, struct msm_file_private *ctx,
51 uint32_t param, uint64_t value, uint32_t len);
52 int (*hw_init)(struct msm_gpu *gpu);
53 int (*pm_suspend)(struct msm_gpu *gpu);
54 int (*pm_resume)(struct msm_gpu *gpu);
55 void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit);
56 void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
57 irqreturn_t (*irq)(struct msm_gpu *irq);
58 struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
59 void (*recover)(struct msm_gpu *gpu);
60 void (*destroy)(struct msm_gpu *gpu);
61 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
62 /* show GPU status in debugfs: */
63 void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
64 struct drm_printer *p);
65 /* for generation specific debugfs: */
66 void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
68 /* note: gpu_busy() can assume that we have been pm_resumed */
69 u64 (*gpu_busy)(struct msm_gpu *gpu, unsigned long *out_sample_rate);
70 struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
71 int (*gpu_state_put)(struct msm_gpu_state *state);
72 unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
73 /* note: gpu_set_freq() can assume that we have been pm_resumed */
74 void (*gpu_set_freq)(struct msm_gpu *gpu, struct dev_pm_opp *opp,
76 struct msm_gem_address_space *(*create_address_space)
77 (struct msm_gpu *gpu, struct platform_device *pdev);
78 struct msm_gem_address_space *(*create_private_address_space)
79 (struct msm_gpu *gpu);
80 uint32_t (*get_rptr)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
83 * progress: Has the GPU made progress?
85 * Return true if GPU position in cmdstream has advanced (or changed)
86 * since the last call. To avoid false negatives, this should account
87 * for cmdstream that is buffered in this FIFO upstream of the CP fw.
89 bool (*progress)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
92 /* Additional state for iommu faults: */
93 struct msm_gpu_fault_info {
102 * struct msm_gpu_devfreq - devfreq related state
104 struct msm_gpu_devfreq {
105 /** devfreq: devfreq instance */
106 struct devfreq *devfreq;
108 /** lock: lock for "suspended", "busy_cycles", and "time" */
114 * Shadow frequency used while the GPU is idle. From the PoV of
115 * the devfreq governor, we are continuing to sample busyness and
116 * adjust frequency while the GPU is idle, but we use this shadow
117 * value as the GPU is actually clamped to minimum frequency while
120 unsigned long idle_freq;
125 * A PM QoS constraint to boost min freq for a period of time
126 * until the boost expires.
128 struct dev_pm_qos_request boost_freq;
131 * busy_cycles: Last busy counter value, for calculating elapsed busy
132 * cycles since last sampling period.
136 /** time: Time of last sampling period. */
139 /** idle_time: Time of last transition to idle: */
145 * Used to delay clamping to idle freq on active->idle transition.
147 struct msm_hrtimer_work idle_work;
152 * Used to reset the boost_constraint after the boost period has
155 struct msm_hrtimer_work boost_work;
157 /** suspended: tracks if we're suspended */
163 struct drm_device *dev;
164 struct platform_device *pdev;
165 const struct msm_gpu_funcs *funcs;
167 struct adreno_smmu_priv adreno_smmu;
169 /* performance counters (hw & sw): */
170 spinlock_t perf_lock;
171 bool perfcntr_active;
176 uint32_t totaltime, activetime; /* sw counters */
177 uint32_t last_cntrs[5]; /* hw counters */
178 const struct msm_gpu_perfcntr *perfcntrs;
179 uint32_t num_perfcntrs;
181 struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
187 * The count of contexts that have enabled system profiling.
189 refcount_t sysprof_active;
194 * The ctx->seqno value of the last context to submit rendering,
195 * and the one with current pgtables installed (for generations
196 * that support per-context pgtables). Tracked by seqno rather
197 * than pointer value to avoid dangling pointers, and cases where
198 * a ctx can be freed and a new one created with the same address.
205 * General lock for serializing all the gpu things.
207 * TODO move to per-ring locking where feasible (ie. submit/retire
215 * The number of submitted but not yet retired submits, used to
216 * determine transitions between active and idle.
218 * Protected by active_lock
222 /** lock: protects active_submits and idle/active transitions */
223 struct mutex active_lock;
225 /* does gpu need hw_init? */
229 * global_faults: number of GPU hangs not attributed to a particular
237 struct msm_gem_address_space *aspace;
240 struct regulator *gpu_reg, *gpu_cx;
241 struct clk_bulk_data *grp_clks;
243 struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
246 /* Hang and Inactivity Detection:
248 #define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
250 #define DRM_MSM_HANGCHECK_DEFAULT_PERIOD 500 /* in ms */
251 #define DRM_MSM_HANGCHECK_PROGRESS_RETRIES 3
252 struct timer_list hangcheck_timer;
254 /* Fault info for most recent iova fault: */
255 struct msm_gpu_fault_info fault_info;
257 /* work for handling GPU ioval faults: */
258 struct kthread_work fault_work;
260 /* work for handling GPU recovery: */
261 struct kthread_work recover_work;
263 /** retire_event: notified when submits are retired: */
264 wait_queue_head_t retire_event;
266 /* work for handling active-list retiring: */
267 struct kthread_work retire_work;
269 /* worker for retire/recover: */
270 struct kthread_worker *worker;
272 struct drm_gem_object *memptrs_bo;
274 struct msm_gpu_devfreq devfreq;
276 uint32_t suspend_count;
278 struct msm_gpu_state *crashstate;
280 /* True if the hardware supports expanded apriv (a650 and newer) */
283 struct thermal_cooling_device *cooling;
285 /* To poll for cx gdsc collapse during gpu recovery */
286 struct reset_control *cx_collapse;
289 static inline struct msm_gpu *dev_to_gpu(struct device *dev)
291 struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(dev);
296 return container_of(adreno_smmu, struct msm_gpu, adreno_smmu);
299 /* It turns out that all targets use the same ringbuffer size */
300 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
301 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
303 #define MSM_GPU_RB_CNTL_DEFAULT \
304 (AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
305 AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
307 static inline bool msm_gpu_active(struct msm_gpu *gpu)
311 for (i = 0; i < gpu->nr_rings; i++) {
312 struct msm_ringbuffer *ring = gpu->rb[i];
314 if (fence_after(ring->fctx->last_fence, ring->memptrs->fence))
322 * The select_reg and select_val are just there for the benefit of the child
323 * class that actually enables the perf counter.. but msm_gpu base class
324 * will handle sampling/displaying the counters.
327 struct msm_gpu_perfcntr {
335 * The number of priority levels provided by drm gpu scheduler. The
336 * DRM_SCHED_PRIORITY_KERNEL priority level is treated specially in some
337 * cases, so we don't use it (no need for kernel generated jobs).
339 #define NR_SCHED_PRIORITIES (1 + DRM_SCHED_PRIORITY_HIGH - DRM_SCHED_PRIORITY_MIN)
342 * struct msm_file_private - per-drm_file context
344 * @queuelock: synchronizes access to submitqueues list
345 * @submitqueues: list of &msm_gpu_submitqueue created by userspace
346 * @queueid: counter incremented each time a submitqueue is created,
347 * used to assign &msm_gpu_submitqueue.id
348 * @aspace: the per-process GPU address-space
349 * @ref: reference count
350 * @seqno: unique per process seqno
352 struct msm_file_private {
354 struct list_head submitqueues;
356 struct msm_gem_address_space *aspace;
363 * The value of MSM_PARAM_SYSPROF set by userspace. This is
364 * intended to be used by system profiling tools like Mesa's
365 * pps-producer (perfetto), and restricted to CAP_SYS_ADMIN.
367 * Setting a value of 1 will preserve performance counters across
368 * context switches. Setting a value of 2 will in addition
369 * suppress suspend. (Performance counters lose state across
370 * power collapse, which is undesirable for profiling in some
373 * The value automatically reverts to zero when the drm device
379 * comm: Overridden task comm, see MSM_PARAM_COMM
381 * Accessed under msm_gpu::lock
386 * cmdline: Overridden task cmdline, see MSM_PARAM_CMDLINE
388 * Accessed under msm_gpu::lock
395 * The total (cumulative) elapsed time GPU was busy with rendering
396 * from this context in ns.
403 * The total (cumulative) GPU cycles elapsed attributed to this
411 * Table of per-priority-level sched entities used by submitqueues
412 * associated with this &drm_file. Because some userspace apps
413 * make assumptions about rendering from multiple gl contexts
414 * (of the same priority) within the process happening in FIFO
415 * order without requiring any fencing beyond MakeCurrent(), we
416 * create at most one &drm_sched_entity per-process per-priority-
419 struct drm_sched_entity *entities[NR_SCHED_PRIORITIES * MSM_GPU_MAX_RINGS];
423 * msm_gpu_convert_priority - Map userspace priority to ring # and sched priority
425 * @gpu: the gpu instance
426 * @prio: the userspace priority level
427 * @ring_nr: [out] the ringbuffer the userspace priority maps to
428 * @sched_prio: [out] the gpu scheduler priority level which the userspace
431 * With drm/scheduler providing it's own level of prioritization, our total
432 * number of available priority levels is (nr_rings * NR_SCHED_PRIORITIES).
433 * Each ring is associated with it's own scheduler instance. However, our
434 * UABI is that lower numerical values are higher priority. So mapping the
435 * single userspace priority level into ring_nr and sched_prio takes some
436 * care. The userspace provided priority (when a submitqueue is created)
437 * is mapped to ring nr and scheduler priority as such:
439 * ring_nr = userspace_prio / NR_SCHED_PRIORITIES
440 * sched_prio = NR_SCHED_PRIORITIES -
441 * (userspace_prio % NR_SCHED_PRIORITIES) - 1
443 * This allows generations without preemption (nr_rings==1) to have some
444 * amount of prioritization, and provides more priority levels for gens
445 * that do have preemption.
447 static inline int msm_gpu_convert_priority(struct msm_gpu *gpu, int prio,
448 unsigned *ring_nr, enum drm_sched_priority *sched_prio)
452 rn = div_u64_rem(prio, NR_SCHED_PRIORITIES, &sp);
454 /* invert sched priority to map to higher-numeric-is-higher-
455 * priority convention
457 sp = NR_SCHED_PRIORITIES - sp - 1;
459 if (rn >= gpu->nr_rings)
469 * struct msm_gpu_submitqueues - Userspace created context.
471 * A submitqueue is associated with a gl context or vk queue (or equiv)
474 * @id: userspace id for the submitqueue, unique within the drm_file
475 * @flags: userspace flags for the submitqueue, specified at creation
476 * (currently unusued)
477 * @ring_nr: the ringbuffer used by this submitqueue, which is determined
478 * by the submitqueue's priority
479 * @faults: the number of GPU hangs associated with this submitqueue
480 * @last_fence: the sequence number of the last allocated fence (for error
482 * @ctx: the per-drm_file context associated with the submitqueue (ie.
483 * which set of pgtables do submits jobs associated with the
485 * @node: node in the context's list of submitqueues
486 * @fence_idr: maps fence-id to dma_fence for userspace visible fence
487 * seqno, protected by submitqueue lock
488 * @idr_lock: for serializing access to fence_idr
489 * @lock: submitqueue lock for serializing submits on a queue
490 * @ref: reference count
491 * @entity: the submit job-queue
493 struct msm_gpu_submitqueue {
499 struct msm_file_private *ctx;
500 struct list_head node;
501 struct idr fence_idr;
502 struct mutex idr_lock;
505 struct drm_sched_entity *entity;
508 struct msm_gpu_state_bo {
516 struct msm_gpu_state {
518 struct timespec64 time;
529 } ring[MSM_GPU_MAX_RINGS];
539 struct msm_gpu_fault_info fault_info;
542 struct msm_gpu_state_bo *bos;
545 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
547 msm_writel(data, gpu->mmio + (reg << 2));
550 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
552 return msm_readl(gpu->mmio + (reg << 2));
555 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
557 msm_rmw(gpu->mmio + (reg << 2), mask, or);
560 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 reg)
565 * Why not a readq here? Two reasons: 1) many of the LO registers are
566 * not quad word aligned and 2) the GPU hardware designers have a bit
567 * of a history of putting registers where they fit, especially in
568 * spins. The longer a GPU family goes the higher the chance that
569 * we'll get burned. We could do a series of validity checks if we
570 * wanted to, but really is a readq() that much better? Nah.
574 * For some lo/hi registers (like perfcounters), the hi value is latched
575 * when the lo is read, so make sure to read the lo first to trigger
578 val = (u64) msm_readl(gpu->mmio + (reg << 2));
579 val |= ((u64) msm_readl(gpu->mmio + ((reg + 1) << 2)) << 32);
584 static inline void gpu_write64(struct msm_gpu *gpu, u32 reg, u64 val)
586 /* Why not a writeq here? Read the screed above */
587 msm_writel(lower_32_bits(val), gpu->mmio + (reg << 2));
588 msm_writel(upper_32_bits(val), gpu->mmio + ((reg + 1) << 2));
591 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
592 int msm_gpu_pm_resume(struct msm_gpu *gpu);
594 void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
595 struct drm_printer *p);
597 int msm_submitqueue_init(struct drm_device *drm, struct msm_file_private *ctx);
598 struct msm_gpu_submitqueue *msm_submitqueue_get(struct msm_file_private *ctx,
600 int msm_submitqueue_create(struct drm_device *drm,
601 struct msm_file_private *ctx,
602 u32 prio, u32 flags, u32 *id);
603 int msm_submitqueue_query(struct drm_device *drm, struct msm_file_private *ctx,
604 struct drm_msm_submitqueue_query *args);
605 int msm_submitqueue_remove(struct msm_file_private *ctx, u32 id);
606 void msm_submitqueue_close(struct msm_file_private *ctx);
608 void msm_submitqueue_destroy(struct kref *kref);
610 int msm_file_private_set_sysprof(struct msm_file_private *ctx,
611 struct msm_gpu *gpu, int sysprof);
612 void __msm_file_private_destroy(struct kref *kref);
614 static inline void msm_file_private_put(struct msm_file_private *ctx)
616 kref_put(&ctx->ref, __msm_file_private_destroy);
619 static inline struct msm_file_private *msm_file_private_get(
620 struct msm_file_private *ctx)
626 void msm_devfreq_init(struct msm_gpu *gpu);
627 void msm_devfreq_cleanup(struct msm_gpu *gpu);
628 void msm_devfreq_resume(struct msm_gpu *gpu);
629 void msm_devfreq_suspend(struct msm_gpu *gpu);
630 void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor);
631 void msm_devfreq_active(struct msm_gpu *gpu);
632 void msm_devfreq_idle(struct msm_gpu *gpu);
634 int msm_gpu_hw_init(struct msm_gpu *gpu);
636 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
637 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
638 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
639 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
641 void msm_gpu_retire(struct msm_gpu *gpu);
642 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit);
644 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
645 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
646 const char *name, struct msm_gpu_config *config);
648 struct msm_gem_address_space *
649 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task);
651 void msm_gpu_cleanup(struct msm_gpu *gpu);
653 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
654 void __init adreno_register(void);
655 void __exit adreno_unregister(void);
657 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
660 kref_put(&queue->ref, msm_submitqueue_destroy);
663 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
665 struct msm_gpu_state *state = NULL;
667 mutex_lock(&gpu->lock);
669 if (gpu->crashstate) {
670 kref_get(&gpu->crashstate->ref);
671 state = gpu->crashstate;
674 mutex_unlock(&gpu->lock);
679 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
681 mutex_lock(&gpu->lock);
683 if (gpu->crashstate) {
684 if (gpu->funcs->gpu_state_put(gpu->crashstate))
685 gpu->crashstate = NULL;
688 mutex_unlock(&gpu->lock);
692 * Simple macro to semi-cleanly add the MAP_PRIV flag for targets that can
693 * support expanded privileges
695 #define check_apriv(gpu, flags) \
696 (((gpu)->hw_apriv ? MSM_BO_MAP_PRIV : 0) | (flags))
699 #endif /* __MSM_GPU_H__ */