2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
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 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
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 int (*gpu_busy)(struct msm_gpu *gpu, uint64_t *value);
74 struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
75 int (*gpu_state_put)(struct msm_gpu_state *state);
80 struct drm_device *dev;
81 struct platform_device *pdev;
82 const struct msm_gpu_funcs *funcs;
84 /* performance counters (hw & sw): */
91 uint32_t totaltime, activetime; /* sw counters */
92 uint32_t last_cntrs[5]; /* hw counters */
93 const struct msm_gpu_perfcntr *perfcntrs;
94 uint32_t num_perfcntrs;
96 struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
99 /* list of GEM active objects: */
100 struct list_head active_list;
102 /* does gpu need hw_init? */
105 /* worker for handling active-list retiring: */
106 struct work_struct retire_work;
111 struct msm_gem_address_space *aspace;
114 struct regulator *gpu_reg, *gpu_cx;
115 struct clk_bulk_data *grp_clks;
117 struct clk *ebi1_clk, *core_clk, *rbbmtimer_clk;
120 /* Hang and Inactivity Detection:
122 #define DRM_MSM_INACTIVE_PERIOD 66 /* in ms (roughly four frames) */
124 #define DRM_MSM_HANGCHECK_PERIOD 500 /* in ms */
125 #define DRM_MSM_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_MSM_HANGCHECK_PERIOD)
126 struct timer_list hangcheck_timer;
127 struct work_struct recover_work;
129 struct drm_gem_object *memptrs_bo;
132 struct devfreq *devfreq;
137 struct msm_gpu_state *crashstate;
140 /* It turns out that all targets use the same ringbuffer size */
141 #define MSM_GPU_RINGBUFFER_SZ SZ_32K
142 #define MSM_GPU_RINGBUFFER_BLKSIZE 32
144 #define MSM_GPU_RB_CNTL_DEFAULT \
145 (AXXX_CP_RB_CNTL_BUFSZ(ilog2(MSM_GPU_RINGBUFFER_SZ / 8)) | \
146 AXXX_CP_RB_CNTL_BLKSZ(ilog2(MSM_GPU_RINGBUFFER_BLKSIZE / 8)))
148 static inline bool msm_gpu_active(struct msm_gpu *gpu)
152 for (i = 0; i < gpu->nr_rings; i++) {
153 struct msm_ringbuffer *ring = gpu->rb[i];
155 if (ring->seqno > ring->memptrs->fence)
163 * The select_reg and select_val are just there for the benefit of the child
164 * class that actually enables the perf counter.. but msm_gpu base class
165 * will handle sampling/displaying the counters.
168 struct msm_gpu_perfcntr {
175 struct msm_gpu_submitqueue {
180 struct list_head node;
184 struct msm_gpu_state_bo {
190 struct msm_gpu_state {
192 struct timespec64 time;
202 } ring[MSM_GPU_MAX_RINGS];
213 struct msm_gpu_state_bo *bos;
216 static inline void gpu_write(struct msm_gpu *gpu, u32 reg, u32 data)
218 msm_writel(data, gpu->mmio + (reg << 2));
221 static inline u32 gpu_read(struct msm_gpu *gpu, u32 reg)
223 return msm_readl(gpu->mmio + (reg << 2));
226 static inline void gpu_rmw(struct msm_gpu *gpu, u32 reg, u32 mask, u32 or)
228 uint32_t val = gpu_read(gpu, reg);
231 gpu_write(gpu, reg, val | or);
234 static inline u64 gpu_read64(struct msm_gpu *gpu, u32 lo, u32 hi)
239 * Why not a readq here? Two reasons: 1) many of the LO registers are
240 * not quad word aligned and 2) the GPU hardware designers have a bit
241 * of a history of putting registers where they fit, especially in
242 * spins. The longer a GPU family goes the higher the chance that
243 * we'll get burned. We could do a series of validity checks if we
244 * wanted to, but really is a readq() that much better? Nah.
248 * For some lo/hi registers (like perfcounters), the hi value is latched
249 * when the lo is read, so make sure to read the lo first to trigger
252 val = (u64) msm_readl(gpu->mmio + (lo << 2));
253 val |= ((u64) msm_readl(gpu->mmio + (hi << 2)) << 32);
258 static inline void gpu_write64(struct msm_gpu *gpu, u32 lo, u32 hi, u64 val)
260 /* Why not a writeq here? Read the screed above */
261 msm_writel(lower_32_bits(val), gpu->mmio + (lo << 2));
262 msm_writel(upper_32_bits(val), gpu->mmio + (hi << 2));
265 int msm_gpu_pm_suspend(struct msm_gpu *gpu);
266 int msm_gpu_pm_resume(struct msm_gpu *gpu);
268 int msm_gpu_hw_init(struct msm_gpu *gpu);
270 void msm_gpu_perfcntr_start(struct msm_gpu *gpu);
271 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu);
272 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
273 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs);
275 void msm_gpu_retire(struct msm_gpu *gpu);
276 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
277 struct msm_file_private *ctx);
279 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
280 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
281 const char *name, struct msm_gpu_config *config);
283 void msm_gpu_cleanup(struct msm_gpu *gpu);
285 struct msm_gpu *adreno_load_gpu(struct drm_device *dev);
286 void __init adreno_register(void);
287 void __exit adreno_unregister(void);
289 static inline void msm_submitqueue_put(struct msm_gpu_submitqueue *queue)
292 kref_put(&queue->ref, msm_submitqueue_destroy);
295 static inline struct msm_gpu_state *msm_gpu_crashstate_get(struct msm_gpu *gpu)
297 struct msm_gpu_state *state = NULL;
299 mutex_lock(&gpu->dev->struct_mutex);
301 if (gpu->crashstate) {
302 kref_get(&gpu->crashstate->ref);
303 state = gpu->crashstate;
306 mutex_unlock(&gpu->dev->struct_mutex);
311 static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
313 mutex_lock(&gpu->dev->struct_mutex);
315 if (gpu->crashstate) {
316 if (gpu->funcs->gpu_state_put(gpu->crashstate))
317 gpu->crashstate = NULL;
320 mutex_unlock(&gpu->dev->struct_mutex);
323 #endif /* __MSM_GPU_H__ */