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/interconnect.h>
13 #include <linux/pm_opp.h>
14 #include <linux/regulator/consumer.h>
17 #include "msm_fence.h"
18 #include "msm_ringbuffer.h"
21 struct msm_gem_submit;
22 struct msm_gpu_perfcntr;
25 struct msm_gpu_config {
27 unsigned int nr_rings;
30 /* So far, with hardware that I've seen to date, we can have:
31 * + zero, one, or two z180 2d cores
32 * + a3xx or a2xx 3d core, which share a common CP (the firmware
33 * for the CP seems to implement some different PM4 packet types
34 * but the basics of cmdstream submission are the same)
36 * Which means that the eventual complete "class" hierarchy, once
37 * support for all past and present hw is in place, becomes:
44 struct msm_gpu_funcs {
45 int (*get_param)(struct msm_gpu *gpu, uint32_t param, uint64_t *value);
46 int (*hw_init)(struct msm_gpu *gpu);
47 int (*pm_suspend)(struct msm_gpu *gpu);
48 int (*pm_resume)(struct msm_gpu *gpu);
49 void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit);
50 void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
51 irqreturn_t (*irq)(struct msm_gpu *irq);
52 struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
53 void (*recover)(struct msm_gpu *gpu);
54 void (*destroy)(struct msm_gpu *gpu);
55 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
56 /* show GPU status in debugfs: */
57 void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
58 struct drm_printer *p);
59 /* for generation specific debugfs: */
60 void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
62 unsigned long (*gpu_busy)(struct msm_gpu *gpu);
63 struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
64 int (*gpu_state_put)(struct msm_gpu_state *state);
65 unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
66 void (*gpu_set_freq)(struct msm_gpu *gpu, struct dev_pm_opp *opp);
67 struct msm_gem_address_space *(*create_address_space)
68 (struct msm_gpu *gpu, struct platform_device *pdev);
69 struct msm_gem_address_space *(*create_private_address_space)
70 (struct msm_gpu *gpu);
71 uint32_t (*get_rptr)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
74 /* Additional state for iommu faults: */
75 struct msm_gpu_fault_info {
84 * struct msm_gpu_devfreq - devfreq related state
86 struct msm_gpu_devfreq {
87 /** devfreq: devfreq instance */
88 struct devfreq *devfreq;
93 * Used by implementation of gpu->gpu_busy() to track the last
94 * busy counter value, for calculating elapsed busy cycles since
95 * last sampling period.
99 /** time: Time of last sampling period. */
102 /** idle_time: Time of last transition to idle: */
108 * Shadow frequency used while the GPU is idle. From the PoV of
109 * the devfreq governor, we are continuing to sample busyness and
110 * adjust frequency while the GPU is idle, but we use this shadow
111 * value as the GPU is actually clamped to minimum frequency while
114 unsigned long idle_freq;
119 struct drm_device *dev;
120 struct platform_device *pdev;
121 const struct msm_gpu_funcs *funcs;
123 struct adreno_smmu_priv adreno_smmu;
125 /* performance counters (hw & sw): */
126 spinlock_t perf_lock;
127 bool perfcntr_active;
132 uint32_t totaltime, activetime; /* sw counters */
133 uint32_t last_cntrs[5]; /* hw counters */
134 const struct msm_gpu_perfcntr *perfcntrs;
135 uint32_t num_perfcntrs;
137 struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
141 * List of GEM active objects on this gpu. Protected by
142 * msm_drm_private::mm_lock
144 struct list_head active_list;
149 * The number of submitted but not yet retired submits, used to
150 * determine transitions between active and idle.
156 /** lock: protects active_submits and idle/active transitions */
157 struct mutex active_lock;
159 /* does gpu need hw_init? */
162 /* number of GPU hangs (for all contexts) */
168 struct msm_gem_address_space *aspace;
171 struct regulator *gpu_reg, *gpu_cx;
172 struct clk_bulk_data *grp_clks;
174 struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
177 /* Hang and Inactivity Detection:
179 #define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
181 #define DRM_MSM_HANGCHECK_DEFAULT_PERIOD 500 /* in ms */
182 struct timer_list hangcheck_timer;
184 /* Fault info for most recent iova fault: */
185 struct msm_gpu_fault_info fault_info;
187 /* work for handling GPU ioval faults: */
188 struct kthread_work fault_work;
190 /* work for handling GPU recovery: */
191 struct kthread_work recover_work;
193 /* work for handling active-list retiring: */
194 struct kthread_work retire_work;
196 /* worker for retire/recover: */
197 struct kthread_worker *worker;
199 struct drm_gem_object *memptrs_bo;
201 struct msm_gpu_devfreq devfreq;
203 uint32_t suspend_count;
205 struct msm_gpu_state *crashstate;
206 /* True if the hardware supports expanded apriv (a650 and newer) */
209 struct thermal_cooling_device *cooling;
212 static inline struct msm_gpu *dev_to_gpu(struct device *dev)
214 struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(dev);
215 return container_of(adreno_smmu, struct msm_gpu, adreno_smmu);
218 /* It turns out that all targets use the same ringbuffer size */
219 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
220 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
222 #define MSM_GPU_RB_CNTL_DEFAULT \
223 (AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
224 AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
226 static inline bool msm_gpu_active(struct msm_gpu *gpu)
230 for (i = 0; i < gpu->nr_rings; i++) {
231 struct msm_ringbuffer *ring = gpu->rb[i];
233 if (ring->seqno > ring->memptrs->fence)
241 * The select_reg and select_val are just there for the benefit of the child
242 * class that actually enables the perf counter.. but msm_gpu base class
243 * will handle sampling/displaying the counters.
246 struct msm_gpu_perfcntr {
254 * The number of priority levels provided by drm gpu scheduler. The
255 * DRM_SCHED_PRIORITY_KERNEL priority level is treated specially in some
256 * cases, so we don't use it (no need for kernel generated jobs).
258 #define NR_SCHED_PRIORITIES (1 + DRM_SCHED_PRIORITY_HIGH - DRM_SCHED_PRIORITY_MIN)
261 * msm_gpu_convert_priority - Map userspace priority to ring # and sched priority
263 * @gpu: the gpu instance
264 * @prio: the userspace priority level
265 * @ring_nr: [out] the ringbuffer the userspace priority maps to
266 * @sched_prio: [out] the gpu scheduler priority level which the userspace
269 * With drm/scheduler providing it's own level of prioritization, our total
270 * number of available priority levels is (nr_rings * NR_SCHED_PRIORITIES).
271 * Each ring is associated with it's own scheduler instance. However, our
272 * UABI is that lower numerical values are higher priority. So mapping the
273 * single userspace priority level into ring_nr and sched_prio takes some
274 * care. The userspace provided priority (when a submitqueue is created)
275 * is mapped to ring nr and scheduler priority as such:
277 * ring_nr = userspace_prio / NR_SCHED_PRIORITIES
278 * sched_prio = NR_SCHED_PRIORITIES -
279 * (userspace_prio % NR_SCHED_PRIORITIES) - 1
281 * This allows generations without preemption (nr_rings==1) to have some
282 * amount of prioritization, and provides more priority levels for gens
283 * that do have preemption.
285 static inline int msm_gpu_convert_priority(struct msm_gpu *gpu, int prio,
286 unsigned *ring_nr, enum drm_sched_priority *sched_prio)
290 rn = div_u64_rem(prio, NR_SCHED_PRIORITIES, &sp);
292 /* invert sched priority to map to higher-numeric-is-higher-
293 * priority convention
295 sp = NR_SCHED_PRIORITIES - sp - 1;
297 if (rn >= gpu->nr_rings)
307 * A submitqueue is associated with a gl context or vk queue (or equiv)
310 * @id: userspace id for the submitqueue, unique within the drm_file
311 * @flags: userspace flags for the submitqueue, specified at creation
312 * (currently unusued)
313 * @ring_nr: the ringbuffer used by this submitqueue, which is determined
314 * by the submitqueue's priority
315 * @faults: the number of GPU hangs associated with this submitqueue
316 * @ctx: the per-drm_file context associated with the submitqueue (ie.
317 * which set of pgtables do submits jobs associated with the
319 * @node: node in the context's list of submitqueues
320 * @fence_idr: maps fence-id to dma_fence for userspace visible fence
321 * seqno, protected by submitqueue lock
322 * @lock: submitqueue lock
323 * @ref: reference count
324 * @entity: the submit job-queue
326 struct msm_gpu_submitqueue {
331 struct msm_file_private *ctx;
332 struct list_head node;
333 struct idr fence_idr;
336 struct drm_sched_entity entity;
339 struct msm_gpu_state_bo {
346 struct msm_gpu_state {
348 struct timespec64 time;
359 } ring[MSM_GPU_MAX_RINGS];
369 struct msm_gpu_fault_info fault_info;
372 struct msm_gpu_state_bo *bos;
375 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
377 msm_writel(data, gpu->mmio + (reg << 2));
380 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
382 return msm_readl(gpu->mmio + (reg << 2));
385 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
387 msm_rmw(gpu->mmio + (reg << 2), mask, or);
390 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
395 * Why not a readq here? Two reasons: 1) many of the LO registers are
396 * not quad word aligned and 2) the GPU hardware designers have a bit
397 * of a history of putting registers where they fit, especially in
398 * spins. The longer a GPU family goes the higher the chance that
399 * we'll get burned. We could do a series of validity checks if we
400 * wanted to, but really is a readq() that much better? Nah.
404 * For some lo/hi registers (like perfcounters), the hi value is latched
405 * when the lo is read, so make sure to read the lo first to trigger
408 val = (u64) msm_readl(gpu->mmio + (lo << 2));
409 val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
414 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
416 /* Why not a writeq here? Read the screed above */
417 msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
418 msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
421 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
422 int msm_gpu_pm_resume(struct msm_gpu *gpu);
424 void msm_devfreq_init(struct msm_gpu *gpu);
425 void msm_devfreq_cleanup(struct msm_gpu *gpu);
426 void msm_devfreq_resume(struct msm_gpu *gpu);
427 void msm_devfreq_suspend(struct msm_gpu *gpu);
428 void msm_devfreq_active(struct msm_gpu *gpu);
429 void msm_devfreq_idle(struct msm_gpu *gpu);
431 int msm_gpu_hw_init(struct msm_gpu *gpu);
433 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
434 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
435 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
436 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
438 void msm_gpu_retire(struct msm_gpu *gpu);
439 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit);
441 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
442 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
443 const char *name, struct msm_gpu_config *config);
445 struct msm_gem_address_space *
446 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task);
448 void msm_gpu_cleanup(struct msm_gpu *gpu);
450 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
451 void __init adreno_register(void);
452 void __exit adreno_unregister(void);
454 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
457 kref_put(&queue->ref, msm_submitqueue_destroy);
460 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
462 struct msm_gpu_state *state = NULL;
464 mutex_lock(&gpu->dev->struct_mutex);
466 if (gpu->crashstate) {
467 kref_get(&gpu->crashstate->ref);
468 state = gpu->crashstate;
471 mutex_unlock(&gpu->dev->struct_mutex);
476 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
478 mutex_lock(&gpu->dev->struct_mutex);
480 if (gpu->crashstate) {
481 if (gpu->funcs->gpu_state_put(gpu->crashstate))
482 gpu->crashstate = NULL;
485 mutex_unlock(&gpu->dev->struct_mutex);
489 * Simple macro to semi-cleanly add the MAP_PRIV flag for targets that can
490 * support expanded privileges
492 #define check_apriv(gpu, flags) \
493 (((gpu)->hw_apriv ? MSM_BO_MAP_PRIV : 0) | (flags))
496 #endif /* __MSM_GPU_H__ */