1 /* SPDX-License-Identifier: GPL-2.0-only */
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
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 * A PM QoS constraint to limit max freq while the GPU is idle.
116 struct dev_pm_qos_request idle_freq;
121 * A PM QoS constraint to boost min freq for a period of time
122 * until the boost expires.
124 struct dev_pm_qos_request boost_freq;
127 * busy_cycles: Last busy counter value, for calculating elapsed busy
128 * cycles since last sampling period.
132 /** time: Time of last sampling period. */
135 /** idle_time: Time of last transition to idle: */
138 struct devfreq_dev_status average_status;
143 * Used to delay clamping to idle freq on active->idle transition.
145 struct msm_hrtimer_work idle_work;
150 * Used to reset the boost_constraint after the boost period has
153 struct msm_hrtimer_work boost_work;
155 /** suspended: tracks if we're suspended */
161 struct drm_device *dev;
162 struct platform_device *pdev;
163 const struct msm_gpu_funcs *funcs;
165 struct adreno_smmu_priv adreno_smmu;
167 /* performance counters (hw & sw): */
168 spinlock_t perf_lock;
169 bool perfcntr_active;
174 uint32_t totaltime, activetime; /* sw counters */
175 uint32_t last_cntrs[5]; /* hw counters */
176 const struct msm_gpu_perfcntr *perfcntrs;
177 uint32_t num_perfcntrs;
179 struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
185 * The count of contexts that have enabled system profiling.
187 refcount_t sysprof_active;
192 * The ctx->seqno value of the last context to submit rendering,
193 * and the one with current pgtables installed (for generations
194 * that support per-context pgtables). Tracked by seqno rather
195 * than pointer value to avoid dangling pointers, and cases where
196 * a ctx can be freed and a new one created with the same address.
203 * General lock for serializing all the gpu things.
205 * TODO move to per-ring locking where feasible (ie. submit/retire
213 * The number of submitted but not yet retired submits, used to
214 * determine transitions between active and idle.
216 * Protected by active_lock
220 /** lock: protects active_submits and idle/active transitions */
221 struct mutex active_lock;
223 /* does gpu need hw_init? */
227 * global_faults: number of GPU hangs not attributed to a particular
235 struct msm_gem_address_space *aspace;
238 struct regulator *gpu_reg, *gpu_cx;
239 struct clk_bulk_data *grp_clks;
241 struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
244 /* Hang and Inactivity Detection:
246 #define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
248 #define DRM_MSM_HANGCHECK_DEFAULT_PERIOD 500 /* in ms */
249 #define DRM_MSM_HANGCHECK_PROGRESS_RETRIES 3
250 struct timer_list hangcheck_timer;
252 /* Fault info for most recent iova fault: */
253 struct msm_gpu_fault_info fault_info;
255 /* work for handling GPU ioval faults: */
256 struct kthread_work fault_work;
258 /* work for handling GPU recovery: */
259 struct kthread_work recover_work;
261 /** retire_event: notified when submits are retired: */
262 wait_queue_head_t retire_event;
264 /* work for handling active-list retiring: */
265 struct kthread_work retire_work;
267 /* worker for retire/recover: */
268 struct kthread_worker *worker;
270 struct drm_gem_object *memptrs_bo;
272 struct msm_gpu_devfreq devfreq;
274 uint32_t suspend_count;
276 struct msm_gpu_state *crashstate;
278 /* Enable clamping to idle freq when inactive: */
281 /* True if the hardware supports expanded apriv (a650 and newer) */
284 struct thermal_cooling_device *cooling;
286 /* To poll for cx gdsc collapse during gpu recovery */
287 struct reset_control *cx_collapse;
290 static inline struct msm_gpu *dev_to_gpu(struct device *dev)
292 struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(dev);
297 return container_of(adreno_smmu, struct msm_gpu, adreno_smmu);
300 /* It turns out that all targets use the same ringbuffer size */
301 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
302 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
304 #define MSM_GPU_RB_CNTL_DEFAULT \
305 (AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
306 AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
308 static inline bool msm_gpu_active(struct msm_gpu *gpu)
312 for (i = 0; i < gpu->nr_rings; i++) {
313 struct msm_ringbuffer *ring = gpu->rb[i];
315 if (fence_after(ring->fctx->last_fence, ring->memptrs->fence))
323 * The select_reg and select_val are just there for the benefit of the child
324 * class that actually enables the perf counter.. but msm_gpu base class
325 * will handle sampling/displaying the counters.
328 struct msm_gpu_perfcntr {
336 * The number of priority levels provided by drm gpu scheduler. The
337 * DRM_SCHED_PRIORITY_KERNEL priority level is treated specially in some
338 * cases, so we don't use it (no need for kernel generated jobs).
340 #define NR_SCHED_PRIORITIES (1 + DRM_SCHED_PRIORITY_HIGH - DRM_SCHED_PRIORITY_MIN)
343 * struct msm_file_private - per-drm_file context
345 * @queuelock: synchronizes access to submitqueues list
346 * @submitqueues: list of &msm_gpu_submitqueue created by userspace
347 * @queueid: counter incremented each time a submitqueue is created,
348 * used to assign &msm_gpu_submitqueue.id
349 * @aspace: the per-process GPU address-space
350 * @ref: reference count
351 * @seqno: unique per process seqno
353 struct msm_file_private {
355 struct list_head submitqueues;
357 struct msm_gem_address_space *aspace;
364 * The value of MSM_PARAM_SYSPROF set by userspace. This is
365 * intended to be used by system profiling tools like Mesa's
366 * pps-producer (perfetto), and restricted to CAP_SYS_ADMIN.
368 * Setting a value of 1 will preserve performance counters across
369 * context switches. Setting a value of 2 will in addition
370 * suppress suspend. (Performance counters lose state across
371 * power collapse, which is undesirable for profiling in some
374 * The value automatically reverts to zero when the drm device
380 * comm: Overridden task comm, see MSM_PARAM_COMM
382 * Accessed under msm_gpu::lock
387 * cmdline: Overridden task cmdline, see MSM_PARAM_CMDLINE
389 * Accessed under msm_gpu::lock
396 * The total (cumulative) elapsed time GPU was busy with rendering
397 * from this context in ns.
404 * The total (cumulative) GPU cycles elapsed attributed to this
412 * Table of per-priority-level sched entities used by submitqueues
413 * associated with this &drm_file. Because some userspace apps
414 * make assumptions about rendering from multiple gl contexts
415 * (of the same priority) within the process happening in FIFO
416 * order without requiring any fencing beyond MakeCurrent(), we
417 * create at most one &drm_sched_entity per-process per-priority-
420 struct drm_sched_entity *entities[NR_SCHED_PRIORITIES * MSM_GPU_MAX_RINGS];
424 * msm_gpu_convert_priority - Map userspace priority to ring # and sched priority
426 * @gpu: the gpu instance
427 * @prio: the userspace priority level
428 * @ring_nr: [out] the ringbuffer the userspace priority maps to
429 * @sched_prio: [out] the gpu scheduler priority level which the userspace
432 * With drm/scheduler providing it's own level of prioritization, our total
433 * number of available priority levels is (nr_rings * NR_SCHED_PRIORITIES).
434 * Each ring is associated with it's own scheduler instance. However, our
435 * UABI is that lower numerical values are higher priority. So mapping the
436 * single userspace priority level into ring_nr and sched_prio takes some
437 * care. The userspace provided priority (when a submitqueue is created)
438 * is mapped to ring nr and scheduler priority as such:
440 * ring_nr = userspace_prio / NR_SCHED_PRIORITIES
441 * sched_prio = NR_SCHED_PRIORITIES -
442 * (userspace_prio % NR_SCHED_PRIORITIES) - 1
444 * This allows generations without preemption (nr_rings==1) to have some
445 * amount of prioritization, and provides more priority levels for gens
446 * that do have preemption.
448 static inline int msm_gpu_convert_priority(struct msm_gpu *gpu, int prio,
449 unsigned *ring_nr, enum drm_sched_priority *sched_prio)
453 rn = div_u64_rem(prio, NR_SCHED_PRIORITIES, &sp);
455 /* invert sched priority to map to higher-numeric-is-higher-
456 * priority convention
458 sp = NR_SCHED_PRIORITIES - sp - 1;
460 if (rn >= gpu->nr_rings)
470 * struct msm_gpu_submitqueues - Userspace created context.
472 * A submitqueue is associated with a gl context or vk queue (or equiv)
475 * @id: userspace id for the submitqueue, unique within the drm_file
476 * @flags: userspace flags for the submitqueue, specified at creation
477 * (currently unusued)
478 * @ring_nr: the ringbuffer used by this submitqueue, which is determined
479 * by the submitqueue's priority
480 * @faults: the number of GPU hangs associated with this submitqueue
481 * @last_fence: the sequence number of the last allocated fence (for error
483 * @ctx: the per-drm_file context associated with the submitqueue (ie.
484 * which set of pgtables do submits jobs associated with the
486 * @node: node in the context's list of submitqueues
487 * @fence_idr: maps fence-id to dma_fence for userspace visible fence
488 * seqno, protected by submitqueue lock
489 * @idr_lock: for serializing access to fence_idr
490 * @lock: submitqueue lock for serializing submits on a queue
491 * @ref: reference count
492 * @entity: the submit job-queue
494 struct msm_gpu_submitqueue {
500 struct msm_file_private *ctx;
501 struct list_head node;
502 struct idr fence_idr;
503 struct mutex idr_lock;
506 struct drm_sched_entity *entity;
509 struct msm_gpu_state_bo {
517 struct msm_gpu_state {
519 struct timespec64 time;
530 } ring[MSM_GPU_MAX_RINGS];
540 struct msm_gpu_fault_info fault_info;
543 struct msm_gpu_state_bo *bos;
546 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
548 msm_writel(data, gpu->mmio + (reg << 2));
551 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
553 return msm_readl(gpu->mmio + (reg << 2));
556 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
558 msm_rmw(gpu->mmio + (reg << 2), mask, or);
561 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 reg)
566 * Why not a readq here? Two reasons: 1) many of the LO registers are
567 * not quad word aligned and 2) the GPU hardware designers have a bit
568 * of a history of putting registers where they fit, especially in
569 * spins. The longer a GPU family goes the higher the chance that
570 * we'll get burned. We could do a series of validity checks if we
571 * wanted to, but really is a readq() that much better? Nah.
575 * For some lo/hi registers (like perfcounters), the hi value is latched
576 * when the lo is read, so make sure to read the lo first to trigger
579 val = (u64) msm_readl(gpu->mmio + (reg << 2));
580 val |= ((u64) msm_readl(gpu->mmio + ((reg + 1) << 2)) << 32);
585 static inline void gpu_write64(struct msm_gpu *gpu, u32 reg, u64 val)
587 /* Why not a writeq here? Read the screed above */
588 msm_writel(lower_32_bits(val), gpu->mmio + (reg << 2));
589 msm_writel(upper_32_bits(val), gpu->mmio + ((reg + 1) << 2));
592 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
593 int msm_gpu_pm_resume(struct msm_gpu *gpu);
595 void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
596 struct drm_printer *p);
598 int msm_submitqueue_init(struct drm_device *drm, struct msm_file_private *ctx);
599 struct msm_gpu_submitqueue *msm_submitqueue_get(struct msm_file_private *ctx,
601 int msm_submitqueue_create(struct drm_device *drm,
602 struct msm_file_private *ctx,
603 u32 prio, u32 flags, u32 *id);
604 int msm_submitqueue_query(struct drm_device *drm, struct msm_file_private *ctx,
605 struct drm_msm_submitqueue_query *args);
606 int msm_submitqueue_remove(struct msm_file_private *ctx, u32 id);
607 void msm_submitqueue_close(struct msm_file_private *ctx);
609 void msm_submitqueue_destroy(struct kref *kref);
611 int msm_file_private_set_sysprof(struct msm_file_private *ctx,
612 struct msm_gpu *gpu, int sysprof);
613 void __msm_file_private_destroy(struct kref *kref);
615 static inline void msm_file_private_put(struct msm_file_private *ctx)
617 kref_put(&ctx->ref, __msm_file_private_destroy);
620 static inline struct msm_file_private *msm_file_private_get(
621 struct msm_file_private *ctx)
627 void msm_devfreq_init(struct msm_gpu *gpu);
628 void msm_devfreq_cleanup(struct msm_gpu *gpu);
629 void msm_devfreq_resume(struct msm_gpu *gpu);
630 void msm_devfreq_suspend(struct msm_gpu *gpu);
631 void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor);
632 void msm_devfreq_active(struct msm_gpu *gpu);
633 void msm_devfreq_idle(struct msm_gpu *gpu);
635 int msm_gpu_hw_init(struct msm_gpu *gpu);
637 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
638 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
639 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
640 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
642 void msm_gpu_retire(struct msm_gpu *gpu);
643 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit);
645 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
646 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
647 const char *name, struct msm_gpu_config *config);
649 struct msm_gem_address_space *
650 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task);
652 void msm_gpu_cleanup(struct msm_gpu *gpu);
654 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
655 void __init adreno_register(void);
656 void __exit adreno_unregister(void);
658 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
661 kref_put(&queue->ref, msm_submitqueue_destroy);
664 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
666 struct msm_gpu_state *state = NULL;
668 mutex_lock(&gpu->lock);
670 if (gpu->crashstate) {
671 kref_get(&gpu->crashstate->ref);
672 state = gpu->crashstate;
675 mutex_unlock(&gpu->lock);
680 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
682 mutex_lock(&gpu->lock);
684 if (gpu->crashstate) {
685 if (gpu->funcs->gpu_state_put(gpu->crashstate))
686 gpu->crashstate = NULL;
689 mutex_unlock(&gpu->lock);
693 * Simple macro to semi-cleanly add the MAP_PRIV flag for targets that can
694 * support expanded privileges
696 #define check_apriv(gpu, flags) \
697 (((gpu)->hw_apriv ? MSM_BO_MAP_PRIV : 0) | (flags))
700 #endif /* __MSM_GPU_H__ */