2 * Copyright (C) 2013 Red Hat
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
21 #include <linux/clk.h>
22 #include <linux/regulator/consumer.h>
25 #include "msm_fence.h"
26 #include "msm_ringbuffer.h"
28 struct msm_gem_submit;
29 struct msm_gpu_perfcntr;
32 struct msm_gpu_config {
37 unsigned int nr_rings;
40 /* So far, with hardware that I've seen to date, we can have:
41 * + zero, one, or two z180 2d cores
42 * + a3xx or a2xx 3d core, which share a common CP (the firmware
43 * for the CP seems to implement some different PM4 packet types
44 * but the basics of cmdstream submission are the same)
46 * Which means that the eventual complete "class" hierarchy, once
47 * support for all past and present hw is in place, becomes:
54 struct msm_gpu_funcs {
55 int (*get_param)(struct msm_gpu *gpu, uint32_t param, uint64_t *value);
56 int (*hw_init)(struct msm_gpu *gpu);
57 int (*pm_suspend)(struct msm_gpu *gpu);
58 int (*pm_resume)(struct msm_gpu *gpu);
59 void (*submit)(struct msm_gpu *gpu, struct msm_gem_submit *submit,
60 struct msm_file_private *ctx);
61 void (*flush)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
62 irqreturn_t (*irq)(struct msm_gpu *irq);
63 struct msm_ringbuffer *(*active_ring)(struct msm_gpu *gpu);
64 void (*recover)(struct msm_gpu *gpu);
65 void (*destroy)(struct msm_gpu *gpu);
66 #ifdef CONFIG_DEBUG_FS
67 /* show GPU status in debugfs: */
68 void (*show)(struct msm_gpu *gpu, struct msm_gpu_state *state,
69 struct drm_printer *p);
70 /* for generation specific debugfs: */
71 int (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
73 unsigned long (*gpu_busy)(struct msm_gpu *gpu);
74 struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
75 int (*gpu_state_put)(struct msm_gpu_state *state);
76 unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
77 void (*gpu_set_freq)(struct msm_gpu *gpu, unsigned long freq);
82 struct drm_device *dev;
83 struct platform_device *pdev;
84 const struct msm_gpu_funcs *funcs;
86 /* performance counters (hw & sw): */
93 uint32_t totaltime, activetime; /* sw counters */
94 uint32_t last_cntrs[5]; /* hw counters */
95 const struct msm_gpu_perfcntr *perfcntrs;
96 uint32_t num_perfcntrs;
98 struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
101 /* list of GEM active objects: */
102 struct list_head active_list;
104 /* does gpu need hw_init? */
107 /* worker for handling active-list retiring: */
108 struct work_struct retire_work;
113 struct msm_gem_address_space *aspace;
116 struct regulator *gpu_reg, *gpu_cx;
117 struct clk_bulk_data *grp_clks;
119 struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
122 /* Hang and Inactivity Detection:
124 #define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
126 #define DRM_MSM_HANGCHECK_PERIOD 500 /* in ms */
127 #define DRM_MSM_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_MSM_HANGCHECK_PERIOD)
128 struct timer_list hangcheck_timer;
129 struct work_struct recover_work;
131 struct drm_gem_object *memptrs_bo;
134 struct devfreq *devfreq;
139 struct msm_gpu_state *crashstate;
142 /* It turns out that all targets use the same ringbuffer size */
143 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
144 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
146 #define MSM_GPU_RB_CNTL_DEFAULT \
147 (AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
148 AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
150 static inline bool msm_gpu_active(struct msm_gpu *gpu)
154 for (i = 0; i < gpu->nr_rings; i++) {
155 struct msm_ringbuffer *ring = gpu->rb[i];
157 if (ring->seqno > ring->memptrs->fence)
165 * The select_reg and select_val are just there for the benefit of the child
166 * class that actually enables the perf counter.. but msm_gpu base class
167 * will handle sampling/displaying the counters.
170 struct msm_gpu_perfcntr {
177 struct msm_gpu_submitqueue {
182 struct list_head node;
186 struct msm_gpu_state_bo {
192 struct msm_gpu_state {
194 struct timespec64 time;
204 } ring[MSM_GPU_MAX_RINGS];
215 struct msm_gpu_state_bo *bos;
218 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
220 msm_writel(data, gpu->mmio + (reg << 2));
223 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
225 return msm_readl(gpu->mmio + (reg << 2));
228 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
230 uint32_t val = gpu_read(gpu, reg);
233 gpu_write(gpu, reg, val | or);
236 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
241 * Why not a readq here? Two reasons: 1) many of the LO registers are
242 * not quad word aligned and 2) the GPU hardware designers have a bit
243 * of a history of putting registers where they fit, especially in
244 * spins. The longer a GPU family goes the higher the chance that
245 * we'll get burned. We could do a series of validity checks if we
246 * wanted to, but really is a readq() that much better? Nah.
250 * For some lo/hi registers (like perfcounters), the hi value is latched
251 * when the lo is read, so make sure to read the lo first to trigger
254 val = (u64) msm_readl(gpu->mmio + (lo << 2));
255 val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
260 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
262 /* Why not a writeq here? Read the screed above */
263 msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
264 msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
267 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
268 int msm_gpu_pm_resume(struct msm_gpu *gpu);
269 void msm_gpu_resume_devfreq(struct msm_gpu *gpu);
271 int msm_gpu_hw_init(struct msm_gpu *gpu);
273 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
274 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
275 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
276 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
278 void msm_gpu_retire(struct msm_gpu *gpu);
279 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
280 struct msm_file_private *ctx);
282 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
283 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
284 const char *name, struct msm_gpu_config *config);
286 void msm_gpu_cleanup(struct msm_gpu *gpu);
288 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
289 void __init adreno_register(void);
290 void __exit adreno_unregister(void);
292 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
295 kref_put(&queue->ref, msm_submitqueue_destroy);
298 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
300 struct msm_gpu_state *state = NULL;
302 mutex_lock(&gpu->dev->struct_mutex);
304 if (gpu->crashstate) {
305 kref_get(&gpu->crashstate->ref);
306 state = gpu->crashstate;
309 mutex_unlock(&gpu->dev->struct_mutex);
314 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
316 mutex_lock(&gpu->dev->struct_mutex);
318 if (gpu->crashstate) {
319 if (gpu->funcs->gpu_state_put(gpu->crashstate))
320 gpu->crashstate = NULL;
323 mutex_unlock(&gpu->dev->struct_mutex);
326 #endif /* __MSM_GPU_H__ */