xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gpu.c (revision 002dff36)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2019 The Linux Foundation. All rights reserved. */
3 
4 
5 #include "msm_gem.h"
6 #include "msm_mmu.h"
7 #include "msm_gpu_trace.h"
8 #include "a6xx_gpu.h"
9 #include "a6xx_gmu.xml.h"
10 
11 #include <linux/devfreq.h>
12 
13 #define GPU_PAS_ID 13
14 
15 static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
16 {
17 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
18 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
19 
20 	/* Check that the GMU is idle */
21 	if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
22 		return false;
23 
24 	/* Check tha the CX master is idle */
25 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
26 			~A6XX_RBBM_STATUS_CP_AHB_BUSY_CX_MASTER)
27 		return false;
28 
29 	return !(gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS) &
30 		A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT);
31 }
32 
33 bool a6xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
34 {
35 	/* wait for CP to drain ringbuffer: */
36 	if (!adreno_idle(gpu, ring))
37 		return false;
38 
39 	if (spin_until(_a6xx_check_idle(gpu))) {
40 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
41 			gpu->name, __builtin_return_address(0),
42 			gpu_read(gpu, REG_A6XX_RBBM_STATUS),
43 			gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS),
44 			gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
45 			gpu_read(gpu, REG_A6XX_CP_RB_WPTR));
46 		return false;
47 	}
48 
49 	return true;
50 }
51 
52 static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
53 {
54 	uint32_t wptr;
55 	unsigned long flags;
56 
57 	spin_lock_irqsave(&ring->lock, flags);
58 
59 	/* Copy the shadow to the actual register */
60 	ring->cur = ring->next;
61 
62 	/* Make sure to wrap wptr if we need to */
63 	wptr = get_wptr(ring);
64 
65 	spin_unlock_irqrestore(&ring->lock, flags);
66 
67 	/* Make sure everything is posted before making a decision */
68 	mb();
69 
70 	gpu_write(gpu, REG_A6XX_CP_RB_WPTR, wptr);
71 }
72 
73 static void get_stats_counter(struct msm_ringbuffer *ring, u32 counter,
74 		u64 iova)
75 {
76 	OUT_PKT7(ring, CP_REG_TO_MEM, 3);
77 	OUT_RING(ring, counter | (1 << 30) | (2 << 18));
78 	OUT_RING(ring, lower_32_bits(iova));
79 	OUT_RING(ring, upper_32_bits(iova));
80 }
81 
82 static void a6xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
83 	struct msm_file_private *ctx)
84 {
85 	unsigned int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
86 	struct msm_drm_private *priv = gpu->dev->dev_private;
87 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
88 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
89 	struct msm_ringbuffer *ring = submit->ring;
90 	unsigned int i;
91 
92 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
93 		rbmemptr_stats(ring, index, cpcycles_start));
94 
95 	/*
96 	 * For PM4 the GMU register offsets are calculated from the base of the
97 	 * GPU registers so we need to add 0x1a800 to the register value on A630
98 	 * to get the right value from PM4.
99 	 */
100 	get_stats_counter(ring, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L + 0x1a800,
101 		rbmemptr_stats(ring, index, alwayson_start));
102 
103 	/* Invalidate CCU depth and color */
104 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
105 	OUT_RING(ring, PC_CCU_INVALIDATE_DEPTH);
106 
107 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
108 	OUT_RING(ring, PC_CCU_INVALIDATE_COLOR);
109 
110 	/* Submit the commands */
111 	for (i = 0; i < submit->nr_cmds; i++) {
112 		switch (submit->cmd[i].type) {
113 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
114 			break;
115 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
116 			if (priv->lastctx == ctx)
117 				break;
118 			/* fall-thru */
119 		case MSM_SUBMIT_CMD_BUF:
120 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
121 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
122 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
123 			OUT_RING(ring, submit->cmd[i].size);
124 			break;
125 		}
126 	}
127 
128 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
129 		rbmemptr_stats(ring, index, cpcycles_end));
130 	get_stats_counter(ring, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L + 0x1a800,
131 		rbmemptr_stats(ring, index, alwayson_end));
132 
133 	/* Write the fence to the scratch register */
134 	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
135 	OUT_RING(ring, submit->seqno);
136 
137 	/*
138 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
139 	 * timestamp is written to the memory and then triggers the interrupt
140 	 */
141 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
142 	OUT_RING(ring, CACHE_FLUSH_TS | (1 << 31));
143 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
144 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
145 	OUT_RING(ring, submit->seqno);
146 
147 	trace_msm_gpu_submit_flush(submit,
148 		gmu_read64(&a6xx_gpu->gmu, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L,
149 			REG_A6XX_GMU_ALWAYS_ON_COUNTER_H));
150 
151 	a6xx_flush(gpu, ring);
152 }
153 
154 static const struct {
155 	u32 offset;
156 	u32 value;
157 } a6xx_hwcg[] = {
158 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x22222222},
159 	{REG_A6XX_RBBM_CLOCK_CNTL_SP1, 0x22222222},
160 	{REG_A6XX_RBBM_CLOCK_CNTL_SP2, 0x22222222},
161 	{REG_A6XX_RBBM_CLOCK_CNTL_SP3, 0x22222222},
162 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02022220},
163 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP1, 0x02022220},
164 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP2, 0x02022220},
165 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP3, 0x02022220},
166 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
167 	{REG_A6XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
168 	{REG_A6XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
169 	{REG_A6XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
170 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000f3cf},
171 	{REG_A6XX_RBBM_CLOCK_HYST_SP1, 0x0000f3cf},
172 	{REG_A6XX_RBBM_CLOCK_HYST_SP2, 0x0000f3cf},
173 	{REG_A6XX_RBBM_CLOCK_HYST_SP3, 0x0000f3cf},
174 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
175 	{REG_A6XX_RBBM_CLOCK_CNTL_TP1, 0x02222222},
176 	{REG_A6XX_RBBM_CLOCK_CNTL_TP2, 0x02222222},
177 	{REG_A6XX_RBBM_CLOCK_CNTL_TP3, 0x02222222},
178 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
179 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
180 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
181 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
182 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
183 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP1, 0x22222222},
184 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP2, 0x22222222},
185 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP3, 0x22222222},
186 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
187 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP1, 0x00022222},
188 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP2, 0x00022222},
189 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP3, 0x00022222},
190 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
191 	{REG_A6XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
192 	{REG_A6XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
193 	{REG_A6XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
194 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
195 	{REG_A6XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
196 	{REG_A6XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
197 	{REG_A6XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
198 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
199 	{REG_A6XX_RBBM_CLOCK_HYST3_TP1, 0x77777777},
200 	{REG_A6XX_RBBM_CLOCK_HYST3_TP2, 0x77777777},
201 	{REG_A6XX_RBBM_CLOCK_HYST3_TP3, 0x77777777},
202 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
203 	{REG_A6XX_RBBM_CLOCK_HYST4_TP1, 0x00077777},
204 	{REG_A6XX_RBBM_CLOCK_HYST4_TP2, 0x00077777},
205 	{REG_A6XX_RBBM_CLOCK_HYST4_TP3, 0x00077777},
206 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
207 	{REG_A6XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
208 	{REG_A6XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
209 	{REG_A6XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
210 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
211 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
212 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
213 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
214 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
215 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP1, 0x11111111},
216 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP2, 0x11111111},
217 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP3, 0x11111111},
218 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
219 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP1, 0x00011111},
220 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP2, 0x00011111},
221 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP3, 0x00011111},
222 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
223 	{REG_A6XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
224 	{REG_A6XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
225 	{REG_A6XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
226 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
227 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
228 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
229 	{REG_A6XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
230 	{REG_A6XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
231 	{REG_A6XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
232 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x00002222},
233 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB1, 0x00002222},
234 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB2, 0x00002222},
235 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB3, 0x00002222},
236 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
237 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU1, 0x00002220},
238 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU2, 0x00002220},
239 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU3, 0x00002220},
240 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040f00},
241 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU1, 0x00040f00},
242 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU2, 0x00040f00},
243 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU3, 0x00040f00},
244 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05022022},
245 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
246 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
247 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
248 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
249 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
250 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
251 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
252 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
253 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
254 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
255 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
256 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
257 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
258 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
259 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
260 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
261 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
262 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555}
263 };
264 
265 static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
266 {
267 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
268 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
269 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
270 	unsigned int i;
271 	u32 val;
272 
273 	val = gpu_read(gpu, REG_A6XX_RBBM_CLOCK_CNTL);
274 
275 	/* Don't re-program the registers if they are already correct */
276 	if ((!state && !val) || (state && (val == 0x8aa8aa02)))
277 		return;
278 
279 	/* Disable SP clock before programming HWCG registers */
280 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
281 
282 	for (i = 0; i < ARRAY_SIZE(a6xx_hwcg); i++)
283 		gpu_write(gpu, a6xx_hwcg[i].offset,
284 			state ? a6xx_hwcg[i].value : 0);
285 
286 	/* Enable SP clock */
287 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
288 
289 	gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? 0x8aa8aa02 : 0);
290 }
291 
292 static int a6xx_cp_init(struct msm_gpu *gpu)
293 {
294 	struct msm_ringbuffer *ring = gpu->rb[0];
295 
296 	OUT_PKT7(ring, CP_ME_INIT, 8);
297 
298 	OUT_RING(ring, 0x0000002f);
299 
300 	/* Enable multiple hardware contexts */
301 	OUT_RING(ring, 0x00000003);
302 
303 	/* Enable error detection */
304 	OUT_RING(ring, 0x20000000);
305 
306 	/* Don't enable header dump */
307 	OUT_RING(ring, 0x00000000);
308 	OUT_RING(ring, 0x00000000);
309 
310 	/* No workarounds enabled */
311 	OUT_RING(ring, 0x00000000);
312 
313 	/* Pad rest of the cmds with 0's */
314 	OUT_RING(ring, 0x00000000);
315 	OUT_RING(ring, 0x00000000);
316 
317 	a6xx_flush(gpu, ring);
318 	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
319 }
320 
321 static int a6xx_ucode_init(struct msm_gpu *gpu)
322 {
323 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
324 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
325 
326 	if (!a6xx_gpu->sqe_bo) {
327 		a6xx_gpu->sqe_bo = adreno_fw_create_bo(gpu,
328 			adreno_gpu->fw[ADRENO_FW_SQE], &a6xx_gpu->sqe_iova);
329 
330 		if (IS_ERR(a6xx_gpu->sqe_bo)) {
331 			int ret = PTR_ERR(a6xx_gpu->sqe_bo);
332 
333 			a6xx_gpu->sqe_bo = NULL;
334 			DRM_DEV_ERROR(&gpu->pdev->dev,
335 				"Could not allocate SQE ucode: %d\n", ret);
336 
337 			return ret;
338 		}
339 
340 		msm_gem_object_set_name(a6xx_gpu->sqe_bo, "sqefw");
341 	}
342 
343 	gpu_write64(gpu, REG_A6XX_CP_SQE_INSTR_BASE_LO,
344 		REG_A6XX_CP_SQE_INSTR_BASE_HI, a6xx_gpu->sqe_iova);
345 
346 	return 0;
347 }
348 
349 static int a6xx_zap_shader_init(struct msm_gpu *gpu)
350 {
351 	static bool loaded;
352 	int ret;
353 
354 	if (loaded)
355 		return 0;
356 
357 	ret = adreno_zap_shader_load(gpu, GPU_PAS_ID);
358 
359 	loaded = !ret;
360 	return ret;
361 }
362 
363 #define A6XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
364 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
365 	  A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
366 	  A6XX_RBBM_INT_0_MASK_CP_IB2 | \
367 	  A6XX_RBBM_INT_0_MASK_CP_IB1 | \
368 	  A6XX_RBBM_INT_0_MASK_CP_RB | \
369 	  A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
370 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
371 	  A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
372 	  A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
373 	  A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR)
374 
375 static int a6xx_hw_init(struct msm_gpu *gpu)
376 {
377 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
378 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
379 	int ret;
380 
381 	/* Make sure the GMU keeps the GPU on while we set it up */
382 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
383 
384 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
385 
386 	/*
387 	 * Disable the trusted memory range - we don't actually supported secure
388 	 * memory rendering at this point in time and we don't want to block off
389 	 * part of the virtual memory space.
390 	 */
391 	gpu_write64(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
392 		REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
393 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
394 
395 	/* Turn on 64 bit addressing for all blocks */
396 	gpu_write(gpu, REG_A6XX_CP_ADDR_MODE_CNTL, 0x1);
397 	gpu_write(gpu, REG_A6XX_VSC_ADDR_MODE_CNTL, 0x1);
398 	gpu_write(gpu, REG_A6XX_GRAS_ADDR_MODE_CNTL, 0x1);
399 	gpu_write(gpu, REG_A6XX_RB_ADDR_MODE_CNTL, 0x1);
400 	gpu_write(gpu, REG_A6XX_PC_ADDR_MODE_CNTL, 0x1);
401 	gpu_write(gpu, REG_A6XX_HLSQ_ADDR_MODE_CNTL, 0x1);
402 	gpu_write(gpu, REG_A6XX_VFD_ADDR_MODE_CNTL, 0x1);
403 	gpu_write(gpu, REG_A6XX_VPC_ADDR_MODE_CNTL, 0x1);
404 	gpu_write(gpu, REG_A6XX_UCHE_ADDR_MODE_CNTL, 0x1);
405 	gpu_write(gpu, REG_A6XX_SP_ADDR_MODE_CNTL, 0x1);
406 	gpu_write(gpu, REG_A6XX_TPL1_ADDR_MODE_CNTL, 0x1);
407 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1);
408 
409 	/*
410 	 * enable hardware clockgating
411 	 * For now enable clock gating only for a630
412 	 */
413 	if (adreno_is_a630(adreno_gpu))
414 		a6xx_set_hwcg(gpu, true);
415 
416 	/* VBIF/GBIF start*/
417 	if (adreno_is_a640(adreno_gpu) || adreno_is_a650(adreno_gpu)) {
418 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE0, 0x00071620);
419 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE1, 0x00071620);
420 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE2, 0x00071620);
421 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
422 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
423 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL, 0x3);
424 	} else {
425 		gpu_write(gpu, REG_A6XX_RBBM_VBIF_CLIENT_QOS_CNTL, 0x3);
426 	}
427 
428 	if (adreno_is_a630(adreno_gpu))
429 		gpu_write(gpu, REG_A6XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
430 
431 	/* Make all blocks contribute to the GPU BUSY perf counter */
432 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xffffffff);
433 
434 	/* Disable L2 bypass in the UCHE */
435 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_LO, 0xffffffc0);
436 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_HI, 0x0001ffff);
437 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_LO, 0xfffff000);
438 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_HI, 0x0001ffff);
439 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_LO, 0xfffff000);
440 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_HI, 0x0001ffff);
441 
442 	if (!adreno_is_a650(adreno_gpu)) {
443 		/* Set the GMEM VA range [0x100000:0x100000 + gpu->gmem - 1] */
444 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MIN_LO,
445 			REG_A6XX_UCHE_GMEM_RANGE_MIN_HI, 0x00100000);
446 
447 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MAX_LO,
448 			REG_A6XX_UCHE_GMEM_RANGE_MAX_HI,
449 			0x00100000 + adreno_gpu->gmem - 1);
450 	}
451 
452 	gpu_write(gpu, REG_A6XX_UCHE_FILTER_CNTL, 0x804);
453 	gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, 0x4);
454 
455 	if (adreno_is_a640(adreno_gpu) || adreno_is_a650(adreno_gpu))
456 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x02000140);
457 	else
458 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x010000c0);
459 	gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
460 
461 	/* Setting the mem pool size */
462 	gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 128);
463 
464 	/* Setting the primFifo thresholds default values */
465 	if (adreno_is_a650(adreno_gpu))
466 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00300000);
467 	else if (adreno_is_a640(adreno_gpu))
468 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00200000);
469 	else
470 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, (0x300 << 11));
471 
472 	/* Set the AHB default slave response to "ERROR" */
473 	gpu_write(gpu, REG_A6XX_CP_AHB_CNTL, 0x1);
474 
475 	/* Turn on performance counters */
476 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_CNTL, 0x1);
477 
478 	/* Select CP0 to always count cycles */
479 	gpu_write(gpu, REG_A6XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT);
480 
481 	if (adreno_is_a630(adreno_gpu)) {
482 		gpu_write(gpu, REG_A6XX_RB_NC_MODE_CNTL, 2 << 1);
483 		gpu_write(gpu, REG_A6XX_TPL1_NC_MODE_CNTL, 2 << 1);
484 		gpu_write(gpu, REG_A6XX_SP_NC_MODE_CNTL, 2 << 1);
485 		gpu_write(gpu, REG_A6XX_UCHE_MODE_CNTL, 2 << 21);
486 	}
487 
488 	/* Enable fault detection */
489 	gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL,
490 		(1 << 30) | 0x1fffff);
491 
492 	gpu_write(gpu, REG_A6XX_UCHE_CLIENT_PF, 1);
493 
494 	/* Set weights for bicubic filtering */
495 	if (adreno_is_a650(adreno_gpu)) {
496 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_0, 0);
497 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_1,
498 			0x3fe05ff4);
499 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_2,
500 			0x3fa0ebee);
501 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_3,
502 			0x3f5193ed);
503 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_4,
504 			0x3f0243f0);
505 	}
506 
507 	/* Protect registers from the CP */
508 	gpu_write(gpu, REG_A6XX_CP_PROTECT_CNTL, 0x00000003);
509 
510 	gpu_write(gpu, REG_A6XX_CP_PROTECT(0),
511 		A6XX_PROTECT_RDONLY(0x600, 0x51));
512 	gpu_write(gpu, REG_A6XX_CP_PROTECT(1), A6XX_PROTECT_RW(0xae50, 0x2));
513 	gpu_write(gpu, REG_A6XX_CP_PROTECT(2), A6XX_PROTECT_RW(0x9624, 0x13));
514 	gpu_write(gpu, REG_A6XX_CP_PROTECT(3), A6XX_PROTECT_RW(0x8630, 0x8));
515 	gpu_write(gpu, REG_A6XX_CP_PROTECT(4), A6XX_PROTECT_RW(0x9e70, 0x1));
516 	gpu_write(gpu, REG_A6XX_CP_PROTECT(5), A6XX_PROTECT_RW(0x9e78, 0x187));
517 	gpu_write(gpu, REG_A6XX_CP_PROTECT(6), A6XX_PROTECT_RW(0xf000, 0x810));
518 	gpu_write(gpu, REG_A6XX_CP_PROTECT(7),
519 		A6XX_PROTECT_RDONLY(0xfc00, 0x3));
520 	gpu_write(gpu, REG_A6XX_CP_PROTECT(8), A6XX_PROTECT_RW(0x50e, 0x0));
521 	gpu_write(gpu, REG_A6XX_CP_PROTECT(9), A6XX_PROTECT_RDONLY(0x50f, 0x0));
522 	gpu_write(gpu, REG_A6XX_CP_PROTECT(10), A6XX_PROTECT_RW(0x510, 0x0));
523 	gpu_write(gpu, REG_A6XX_CP_PROTECT(11),
524 		A6XX_PROTECT_RDONLY(0x0, 0x4f9));
525 	gpu_write(gpu, REG_A6XX_CP_PROTECT(12),
526 		A6XX_PROTECT_RDONLY(0x501, 0xa));
527 	gpu_write(gpu, REG_A6XX_CP_PROTECT(13),
528 		A6XX_PROTECT_RDONLY(0x511, 0x44));
529 	gpu_write(gpu, REG_A6XX_CP_PROTECT(14), A6XX_PROTECT_RW(0xe00, 0xe));
530 	gpu_write(gpu, REG_A6XX_CP_PROTECT(15), A6XX_PROTECT_RW(0x8e00, 0x0));
531 	gpu_write(gpu, REG_A6XX_CP_PROTECT(16), A6XX_PROTECT_RW(0x8e50, 0xf));
532 	gpu_write(gpu, REG_A6XX_CP_PROTECT(17), A6XX_PROTECT_RW(0xbe02, 0x0));
533 	gpu_write(gpu, REG_A6XX_CP_PROTECT(18),
534 		A6XX_PROTECT_RW(0xbe20, 0x11f3));
535 	gpu_write(gpu, REG_A6XX_CP_PROTECT(19), A6XX_PROTECT_RW(0x800, 0x82));
536 	gpu_write(gpu, REG_A6XX_CP_PROTECT(20), A6XX_PROTECT_RW(0x8a0, 0x8));
537 	gpu_write(gpu, REG_A6XX_CP_PROTECT(21), A6XX_PROTECT_RW(0x8ab, 0x19));
538 	gpu_write(gpu, REG_A6XX_CP_PROTECT(22), A6XX_PROTECT_RW(0x900, 0x4d));
539 	gpu_write(gpu, REG_A6XX_CP_PROTECT(23), A6XX_PROTECT_RW(0x98d, 0x76));
540 	gpu_write(gpu, REG_A6XX_CP_PROTECT(24),
541 			A6XX_PROTECT_RDONLY(0x980, 0x4));
542 	gpu_write(gpu, REG_A6XX_CP_PROTECT(25), A6XX_PROTECT_RW(0xa630, 0x0));
543 
544 	if (adreno_is_a650(adreno_gpu)) {
545 		gpu_write(gpu, REG_A6XX_CP_APRIV_CNTL,
546 			(1 << 6) | (1 << 5) | (1 << 3) | (1 << 2) | (1 << 1));
547 	}
548 
549 	/* Enable interrupts */
550 	gpu_write(gpu, REG_A6XX_RBBM_INT_0_MASK, A6XX_INT_MASK);
551 
552 	ret = adreno_hw_init(gpu);
553 	if (ret)
554 		goto out;
555 
556 	ret = a6xx_ucode_init(gpu);
557 	if (ret)
558 		goto out;
559 
560 	/* Always come up on rb 0 */
561 	a6xx_gpu->cur_ring = gpu->rb[0];
562 
563 	/* Enable the SQE_to start the CP engine */
564 	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 1);
565 
566 	ret = a6xx_cp_init(gpu);
567 	if (ret)
568 		goto out;
569 
570 	/*
571 	 * Try to load a zap shader into the secure world. If successful
572 	 * we can use the CP to switch out of secure mode. If not then we
573 	 * have no resource but to try to switch ourselves out manually. If we
574 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
575 	 * be blocked and a permissions violation will soon follow.
576 	 */
577 	ret = a6xx_zap_shader_init(gpu);
578 	if (!ret) {
579 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
580 		OUT_RING(gpu->rb[0], 0x00000000);
581 
582 		a6xx_flush(gpu, gpu->rb[0]);
583 		if (!a6xx_idle(gpu, gpu->rb[0]))
584 			return -EINVAL;
585 	} else if (ret == -ENODEV) {
586 		/*
587 		 * This device does not use zap shader (but print a warning
588 		 * just in case someone got their dt wrong.. hopefully they
589 		 * have a debug UART to realize the error of their ways...
590 		 * if you mess this up you are about to crash horribly)
591 		 */
592 		dev_warn_once(gpu->dev->dev,
593 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
594 		gpu_write(gpu, REG_A6XX_RBBM_SECVID_TRUST_CNTL, 0x0);
595 		ret = 0;
596 	} else {
597 		return ret;
598 	}
599 
600 out:
601 	/*
602 	 * Tell the GMU that we are done touching the GPU and it can start power
603 	 * management
604 	 */
605 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
606 
607 	if (a6xx_gpu->gmu.legacy) {
608 		/* Take the GMU out of its special boot mode */
609 		a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_BOOT_SLUMBER);
610 	}
611 
612 	return ret;
613 }
614 
615 static void a6xx_dump(struct msm_gpu *gpu)
616 {
617 	DRM_DEV_INFO(&gpu->pdev->dev, "status:   %08x\n",
618 			gpu_read(gpu, REG_A6XX_RBBM_STATUS));
619 	adreno_dump(gpu);
620 }
621 
622 #define VBIF_RESET_ACK_TIMEOUT	100
623 #define VBIF_RESET_ACK_MASK	0x00f0
624 
625 static void a6xx_recover(struct msm_gpu *gpu)
626 {
627 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
628 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
629 	int i;
630 
631 	adreno_dump_info(gpu);
632 
633 	for (i = 0; i < 8; i++)
634 		DRM_DEV_INFO(&gpu->pdev->dev, "CP_SCRATCH_REG%d: %u\n", i,
635 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(i)));
636 
637 	if (hang_debug)
638 		a6xx_dump(gpu);
639 
640 	/*
641 	 * Turn off keep alive that might have been enabled by the hang
642 	 * interrupt
643 	 */
644 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
645 
646 	gpu->funcs->pm_suspend(gpu);
647 	gpu->funcs->pm_resume(gpu);
648 
649 	msm_gpu_hw_init(gpu);
650 }
651 
652 static int a6xx_fault_handler(void *arg, unsigned long iova, int flags)
653 {
654 	struct msm_gpu *gpu = arg;
655 
656 	pr_warn_ratelimited("*** gpu fault: iova=%08lx, flags=%d (%u,%u,%u,%u)\n",
657 			iova, flags,
658 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
659 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
660 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
661 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
662 
663 	return -EFAULT;
664 }
665 
666 static void a6xx_cp_hw_err_irq(struct msm_gpu *gpu)
667 {
668 	u32 status = gpu_read(gpu, REG_A6XX_CP_INTERRUPT_STATUS);
669 
670 	if (status & A6XX_CP_INT_CP_OPCODE_ERROR) {
671 		u32 val;
672 
673 		gpu_write(gpu, REG_A6XX_CP_SQE_STAT_ADDR, 1);
674 		val = gpu_read(gpu, REG_A6XX_CP_SQE_STAT_DATA);
675 		dev_err_ratelimited(&gpu->pdev->dev,
676 			"CP | opcode error | possible opcode=0x%8.8X\n",
677 			val);
678 	}
679 
680 	if (status & A6XX_CP_INT_CP_UCODE_ERROR)
681 		dev_err_ratelimited(&gpu->pdev->dev,
682 			"CP ucode error interrupt\n");
683 
684 	if (status & A6XX_CP_INT_CP_HW_FAULT_ERROR)
685 		dev_err_ratelimited(&gpu->pdev->dev, "CP | HW fault | status=0x%8.8X\n",
686 			gpu_read(gpu, REG_A6XX_CP_HW_FAULT));
687 
688 	if (status & A6XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
689 		u32 val = gpu_read(gpu, REG_A6XX_CP_PROTECT_STATUS);
690 
691 		dev_err_ratelimited(&gpu->pdev->dev,
692 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
693 			val & (1 << 20) ? "READ" : "WRITE",
694 			(val & 0x3ffff), val);
695 	}
696 
697 	if (status & A6XX_CP_INT_CP_AHB_ERROR)
698 		dev_err_ratelimited(&gpu->pdev->dev, "CP AHB error interrupt\n");
699 
700 	if (status & A6XX_CP_INT_CP_VSD_PARITY_ERROR)
701 		dev_err_ratelimited(&gpu->pdev->dev, "CP VSD decoder parity error\n");
702 
703 	if (status & A6XX_CP_INT_CP_ILLEGAL_INSTR_ERROR)
704 		dev_err_ratelimited(&gpu->pdev->dev, "CP illegal instruction error\n");
705 
706 }
707 
708 static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
709 {
710 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
711 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
712 	struct drm_device *dev = gpu->dev;
713 	struct msm_drm_private *priv = dev->dev_private;
714 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
715 
716 	/*
717 	 * Force the GPU to stay on until after we finish
718 	 * collecting information
719 	 */
720 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
721 
722 	DRM_DEV_ERROR(&gpu->pdev->dev,
723 		"gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
724 		ring ? ring->id : -1, ring ? ring->seqno : 0,
725 		gpu_read(gpu, REG_A6XX_RBBM_STATUS),
726 		gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
727 		gpu_read(gpu, REG_A6XX_CP_RB_WPTR),
728 		gpu_read64(gpu, REG_A6XX_CP_IB1_BASE, REG_A6XX_CP_IB1_BASE_HI),
729 		gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
730 		gpu_read64(gpu, REG_A6XX_CP_IB2_BASE, REG_A6XX_CP_IB2_BASE_HI),
731 		gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE));
732 
733 	/* Turn off the hangcheck timer to keep it from bothering us */
734 	del_timer(&gpu->hangcheck_timer);
735 
736 	queue_work(priv->wq, &gpu->recover_work);
737 }
738 
739 static irqreturn_t a6xx_irq(struct msm_gpu *gpu)
740 {
741 	u32 status = gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS);
742 
743 	gpu_write(gpu, REG_A6XX_RBBM_INT_CLEAR_CMD, status);
744 
745 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT)
746 		a6xx_fault_detect_irq(gpu);
747 
748 	if (status & A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR)
749 		dev_err_ratelimited(&gpu->pdev->dev, "CP | AHB bus error\n");
750 
751 	if (status & A6XX_RBBM_INT_0_MASK_CP_HW_ERROR)
752 		a6xx_cp_hw_err_irq(gpu);
753 
754 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW)
755 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB ASYNC overflow\n");
756 
757 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
758 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB bus overflow\n");
759 
760 	if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
761 		dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of bounds access\n");
762 
763 	if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS)
764 		msm_gpu_retire(gpu);
765 
766 	return IRQ_HANDLED;
767 }
768 
769 static const u32 a6xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
770 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_A6XX_CP_RB_BASE),
771 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE_HI, REG_A6XX_CP_RB_BASE_HI),
772 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR,
773 		REG_A6XX_CP_RB_RPTR_ADDR_LO),
774 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR_HI,
775 		REG_A6XX_CP_RB_RPTR_ADDR_HI),
776 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_A6XX_CP_RB_RPTR),
777 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_A6XX_CP_RB_WPTR),
778 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_A6XX_CP_RB_CNTL),
779 };
780 
781 static int a6xx_pm_resume(struct msm_gpu *gpu)
782 {
783 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
784 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
785 	int ret;
786 
787 	gpu->needs_hw_init = true;
788 
789 	ret = a6xx_gmu_resume(a6xx_gpu);
790 	if (ret)
791 		return ret;
792 
793 	msm_gpu_resume_devfreq(gpu);
794 
795 	return 0;
796 }
797 
798 static int a6xx_pm_suspend(struct msm_gpu *gpu)
799 {
800 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
801 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
802 
803 	devfreq_suspend_device(gpu->devfreq.devfreq);
804 
805 	return a6xx_gmu_stop(a6xx_gpu);
806 }
807 
808 static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
809 {
810 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
811 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
812 
813 	/* Force the GPU power on so we can read this register */
814 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
815 
816 	*value = gpu_read64(gpu, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
817 		REG_A6XX_RBBM_PERFCTR_CP_0_HI);
818 
819 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
820 	return 0;
821 }
822 
823 static struct msm_ringbuffer *a6xx_active_ring(struct msm_gpu *gpu)
824 {
825 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
826 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
827 
828 	return a6xx_gpu->cur_ring;
829 }
830 
831 static void a6xx_destroy(struct msm_gpu *gpu)
832 {
833 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
834 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
835 
836 	if (a6xx_gpu->sqe_bo) {
837 		msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
838 		drm_gem_object_put_unlocked(a6xx_gpu->sqe_bo);
839 	}
840 
841 	a6xx_gmu_remove(a6xx_gpu);
842 
843 	adreno_gpu_cleanup(adreno_gpu);
844 	kfree(a6xx_gpu);
845 }
846 
847 static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
848 {
849 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
850 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
851 	u64 busy_cycles, busy_time;
852 
853 
854 	/* Only read the gpu busy if the hardware is already active */
855 	if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
856 		return 0;
857 
858 	busy_cycles = gmu_read64(&a6xx_gpu->gmu,
859 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
860 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
861 
862 	busy_time = (busy_cycles - gpu->devfreq.busy_cycles) * 10;
863 	do_div(busy_time, 192);
864 
865 	gpu->devfreq.busy_cycles = busy_cycles;
866 
867 	pm_runtime_put(a6xx_gpu->gmu.dev);
868 
869 	if (WARN_ON(busy_time > ~0LU))
870 		return ~0LU;
871 
872 	return (unsigned long)busy_time;
873 }
874 
875 static const struct adreno_gpu_funcs funcs = {
876 	.base = {
877 		.get_param = adreno_get_param,
878 		.hw_init = a6xx_hw_init,
879 		.pm_suspend = a6xx_pm_suspend,
880 		.pm_resume = a6xx_pm_resume,
881 		.recover = a6xx_recover,
882 		.submit = a6xx_submit,
883 		.flush = a6xx_flush,
884 		.active_ring = a6xx_active_ring,
885 		.irq = a6xx_irq,
886 		.destroy = a6xx_destroy,
887 #if defined(CONFIG_DRM_MSM_GPU_STATE)
888 		.show = a6xx_show,
889 #endif
890 		.gpu_busy = a6xx_gpu_busy,
891 		.gpu_get_freq = a6xx_gmu_get_freq,
892 		.gpu_set_freq = a6xx_gmu_set_freq,
893 #if defined(CONFIG_DRM_MSM_GPU_STATE)
894 		.gpu_state_get = a6xx_gpu_state_get,
895 		.gpu_state_put = a6xx_gpu_state_put,
896 		.create_address_space = adreno_iommu_create_address_space,
897 #endif
898 	},
899 	.get_timestamp = a6xx_get_timestamp,
900 };
901 
902 struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
903 {
904 	struct msm_drm_private *priv = dev->dev_private;
905 	struct platform_device *pdev = priv->gpu_pdev;
906 	struct device_node *node;
907 	struct a6xx_gpu *a6xx_gpu;
908 	struct adreno_gpu *adreno_gpu;
909 	struct msm_gpu *gpu;
910 	int ret;
911 
912 	a6xx_gpu = kzalloc(sizeof(*a6xx_gpu), GFP_KERNEL);
913 	if (!a6xx_gpu)
914 		return ERR_PTR(-ENOMEM);
915 
916 	adreno_gpu = &a6xx_gpu->base;
917 	gpu = &adreno_gpu->base;
918 
919 	adreno_gpu->registers = NULL;
920 	adreno_gpu->reg_offsets = a6xx_register_offsets;
921 
922 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
923 	if (ret) {
924 		a6xx_destroy(&(a6xx_gpu->base.base));
925 		return ERR_PTR(ret);
926 	}
927 
928 	/* Check if there is a GMU phandle and set it up */
929 	node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
930 
931 	/* FIXME: How do we gracefully handle this? */
932 	BUG_ON(!node);
933 
934 	ret = a6xx_gmu_init(a6xx_gpu, node);
935 	if (ret) {
936 		a6xx_destroy(&(a6xx_gpu->base.base));
937 		return ERR_PTR(ret);
938 	}
939 
940 	if (gpu->aspace)
941 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu,
942 				a6xx_fault_handler);
943 
944 	return gpu;
945 }
946