xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gpu.c (revision 249592bf)
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/bitfield.h>
12 #include <linux/devfreq.h>
13 #include <linux/nvmem-consumer.h>
14 #include <linux/soc/qcom/llcc-qcom.h>
15 
16 #define GPU_PAS_ID 13
17 
18 static inline bool _a6xx_check_idle(struct msm_gpu *gpu)
19 {
20 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
21 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
22 
23 	/* Check that the GMU is idle */
24 	if (!a6xx_gmu_isidle(&a6xx_gpu->gmu))
25 		return false;
26 
27 	/* Check tha the CX master is idle */
28 	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS) &
29 			~A6XX_RBBM_STATUS_CP_AHB_BUSY_CX_MASTER)
30 		return false;
31 
32 	return !(gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS) &
33 		A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT);
34 }
35 
36 static bool a6xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
37 {
38 	/* wait for CP to drain ringbuffer: */
39 	if (!adreno_idle(gpu, ring))
40 		return false;
41 
42 	if (spin_until(_a6xx_check_idle(gpu))) {
43 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
44 			gpu->name, __builtin_return_address(0),
45 			gpu_read(gpu, REG_A6XX_RBBM_STATUS),
46 			gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS),
47 			gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
48 			gpu_read(gpu, REG_A6XX_CP_RB_WPTR));
49 		return false;
50 	}
51 
52 	return true;
53 }
54 
55 static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
56 {
57 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
58 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
59 	uint32_t wptr;
60 	unsigned long flags;
61 
62 	/* Expanded APRIV doesn't need to issue the WHERE_AM_I opcode */
63 	if (a6xx_gpu->has_whereami && !adreno_gpu->base.hw_apriv) {
64 		struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
65 
66 		OUT_PKT7(ring, CP_WHERE_AM_I, 2);
67 		OUT_RING(ring, lower_32_bits(shadowptr(a6xx_gpu, ring)));
68 		OUT_RING(ring, upper_32_bits(shadowptr(a6xx_gpu, ring)));
69 	}
70 
71 	spin_lock_irqsave(&ring->preempt_lock, flags);
72 
73 	/* Copy the shadow to the actual register */
74 	ring->cur = ring->next;
75 
76 	/* Make sure to wrap wptr if we need to */
77 	wptr = get_wptr(ring);
78 
79 	spin_unlock_irqrestore(&ring->preempt_lock, flags);
80 
81 	/* Make sure everything is posted before making a decision */
82 	mb();
83 
84 	gpu_write(gpu, REG_A6XX_CP_RB_WPTR, wptr);
85 }
86 
87 static void get_stats_counter(struct msm_ringbuffer *ring, u32 counter,
88 		u64 iova)
89 {
90 	OUT_PKT7(ring, CP_REG_TO_MEM, 3);
91 	OUT_RING(ring, CP_REG_TO_MEM_0_REG(counter) |
92 		CP_REG_TO_MEM_0_CNT(2) |
93 		CP_REG_TO_MEM_0_64B);
94 	OUT_RING(ring, lower_32_bits(iova));
95 	OUT_RING(ring, upper_32_bits(iova));
96 }
97 
98 static void a6xx_set_pagetable(struct a6xx_gpu *a6xx_gpu,
99 		struct msm_ringbuffer *ring, struct msm_file_private *ctx)
100 {
101 	phys_addr_t ttbr;
102 	u32 asid;
103 	u64 memptr = rbmemptr(ring, ttbr0);
104 
105 	if (ctx == a6xx_gpu->cur_ctx)
106 		return;
107 
108 	if (msm_iommu_pagetable_params(ctx->aspace->mmu, &ttbr, &asid))
109 		return;
110 
111 	/* Execute the table update */
112 	OUT_PKT7(ring, CP_SMMU_TABLE_UPDATE, 4);
113 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_0_TTBR0_LO(lower_32_bits(ttbr)));
114 
115 	OUT_RING(ring,
116 		CP_SMMU_TABLE_UPDATE_1_TTBR0_HI(upper_32_bits(ttbr)) |
117 		CP_SMMU_TABLE_UPDATE_1_ASID(asid));
118 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_2_CONTEXTIDR(0));
119 	OUT_RING(ring, CP_SMMU_TABLE_UPDATE_3_CONTEXTBANK(0));
120 
121 	/*
122 	 * Write the new TTBR0 to the memstore. This is good for debugging.
123 	 */
124 	OUT_PKT7(ring, CP_MEM_WRITE, 4);
125 	OUT_RING(ring, CP_MEM_WRITE_0_ADDR_LO(lower_32_bits(memptr)));
126 	OUT_RING(ring, CP_MEM_WRITE_1_ADDR_HI(upper_32_bits(memptr)));
127 	OUT_RING(ring, lower_32_bits(ttbr));
128 	OUT_RING(ring, (asid << 16) | upper_32_bits(ttbr));
129 
130 	/*
131 	 * And finally, trigger a uche flush to be sure there isn't anything
132 	 * lingering in that part of the GPU
133 	 */
134 
135 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
136 	OUT_RING(ring, 0x31);
137 
138 	a6xx_gpu->cur_ctx = ctx;
139 }
140 
141 static void a6xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
142 {
143 	unsigned int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
144 	struct msm_drm_private *priv = gpu->dev->dev_private;
145 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
146 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
147 	struct msm_ringbuffer *ring = submit->ring;
148 	unsigned int i;
149 
150 	a6xx_set_pagetable(a6xx_gpu, ring, submit->queue->ctx);
151 
152 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
153 		rbmemptr_stats(ring, index, cpcycles_start));
154 
155 	/*
156 	 * For PM4 the GMU register offsets are calculated from the base of the
157 	 * GPU registers so we need to add 0x1a800 to the register value on A630
158 	 * to get the right value from PM4.
159 	 */
160 	get_stats_counter(ring, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L + 0x1a800,
161 		rbmemptr_stats(ring, index, alwayson_start));
162 
163 	/* Invalidate CCU depth and color */
164 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
165 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_DEPTH));
166 
167 	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
168 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(PC_CCU_INVALIDATE_COLOR));
169 
170 	/* Submit the commands */
171 	for (i = 0; i < submit->nr_cmds; i++) {
172 		switch (submit->cmd[i].type) {
173 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
174 			break;
175 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
176 			if (priv->lastctx == submit->queue->ctx)
177 				break;
178 			fallthrough;
179 		case MSM_SUBMIT_CMD_BUF:
180 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
181 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
182 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
183 			OUT_RING(ring, submit->cmd[i].size);
184 			break;
185 		}
186 	}
187 
188 	get_stats_counter(ring, REG_A6XX_RBBM_PERFCTR_CP_0_LO,
189 		rbmemptr_stats(ring, index, cpcycles_end));
190 	get_stats_counter(ring, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L + 0x1a800,
191 		rbmemptr_stats(ring, index, alwayson_end));
192 
193 	/* Write the fence to the scratch register */
194 	OUT_PKT4(ring, REG_A6XX_CP_SCRATCH_REG(2), 1);
195 	OUT_RING(ring, submit->seqno);
196 
197 	/*
198 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
199 	 * timestamp is written to the memory and then triggers the interrupt
200 	 */
201 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
202 	OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(CACHE_FLUSH_TS) |
203 		CP_EVENT_WRITE_0_IRQ);
204 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
205 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
206 	OUT_RING(ring, submit->seqno);
207 
208 	trace_msm_gpu_submit_flush(submit,
209 		gmu_read64(&a6xx_gpu->gmu, REG_A6XX_GMU_ALWAYS_ON_COUNTER_L,
210 			REG_A6XX_GMU_ALWAYS_ON_COUNTER_H));
211 
212 	a6xx_flush(gpu, ring);
213 }
214 
215 const struct adreno_reglist a630_hwcg[] = {
216 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x22222222},
217 	{REG_A6XX_RBBM_CLOCK_CNTL_SP1, 0x22222222},
218 	{REG_A6XX_RBBM_CLOCK_CNTL_SP2, 0x22222222},
219 	{REG_A6XX_RBBM_CLOCK_CNTL_SP3, 0x22222222},
220 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02022220},
221 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP1, 0x02022220},
222 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP2, 0x02022220},
223 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP3, 0x02022220},
224 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
225 	{REG_A6XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
226 	{REG_A6XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
227 	{REG_A6XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
228 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000f3cf},
229 	{REG_A6XX_RBBM_CLOCK_HYST_SP1, 0x0000f3cf},
230 	{REG_A6XX_RBBM_CLOCK_HYST_SP2, 0x0000f3cf},
231 	{REG_A6XX_RBBM_CLOCK_HYST_SP3, 0x0000f3cf},
232 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
233 	{REG_A6XX_RBBM_CLOCK_CNTL_TP1, 0x02222222},
234 	{REG_A6XX_RBBM_CLOCK_CNTL_TP2, 0x02222222},
235 	{REG_A6XX_RBBM_CLOCK_CNTL_TP3, 0x02222222},
236 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
237 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
238 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
239 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
240 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
241 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP1, 0x22222222},
242 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP2, 0x22222222},
243 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP3, 0x22222222},
244 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
245 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP1, 0x00022222},
246 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP2, 0x00022222},
247 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP3, 0x00022222},
248 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
249 	{REG_A6XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
250 	{REG_A6XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
251 	{REG_A6XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
252 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
253 	{REG_A6XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
254 	{REG_A6XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
255 	{REG_A6XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
256 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
257 	{REG_A6XX_RBBM_CLOCK_HYST3_TP1, 0x77777777},
258 	{REG_A6XX_RBBM_CLOCK_HYST3_TP2, 0x77777777},
259 	{REG_A6XX_RBBM_CLOCK_HYST3_TP3, 0x77777777},
260 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
261 	{REG_A6XX_RBBM_CLOCK_HYST4_TP1, 0x00077777},
262 	{REG_A6XX_RBBM_CLOCK_HYST4_TP2, 0x00077777},
263 	{REG_A6XX_RBBM_CLOCK_HYST4_TP3, 0x00077777},
264 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
265 	{REG_A6XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
266 	{REG_A6XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
267 	{REG_A6XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
268 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
269 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
270 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
271 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
272 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
273 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP1, 0x11111111},
274 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP2, 0x11111111},
275 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP3, 0x11111111},
276 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
277 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP1, 0x00011111},
278 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP2, 0x00011111},
279 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP3, 0x00011111},
280 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
281 	{REG_A6XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
282 	{REG_A6XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
283 	{REG_A6XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
284 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
285 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
286 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
287 	{REG_A6XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
288 	{REG_A6XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
289 	{REG_A6XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
290 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x00002222},
291 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB1, 0x00002222},
292 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB2, 0x00002222},
293 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB3, 0x00002222},
294 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
295 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU1, 0x00002220},
296 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU2, 0x00002220},
297 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU3, 0x00002220},
298 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040f00},
299 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU1, 0x00040f00},
300 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU2, 0x00040f00},
301 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU3, 0x00040f00},
302 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05022022},
303 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
304 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
305 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
306 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
307 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
308 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
309 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
310 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
311 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
312 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
313 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
314 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
315 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
316 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
317 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
318 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
319 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
320 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
321 	{},
322 };
323 
324 const struct adreno_reglist a640_hwcg[] = {
325 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
326 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
327 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
328 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
329 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
330 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
331 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
332 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
333 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
334 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
335 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
336 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
337 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
338 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
339 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
340 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
341 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
342 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x01002222},
343 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
344 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040F00},
345 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x05222022},
346 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
347 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
348 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
349 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
350 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
351 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
352 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
353 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
354 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
355 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
356 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
357 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
358 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
359 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
360 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
361 	{REG_A6XX_RBBM_CLOCK_HYST_HLSQ, 0x00000000},
362 	{REG_A6XX_RBBM_CLOCK_CNTL_TEX_FCHE, 0x00000222},
363 	{REG_A6XX_RBBM_CLOCK_DELAY_TEX_FCHE, 0x00000111},
364 	{REG_A6XX_RBBM_CLOCK_HYST_TEX_FCHE, 0x00000000},
365 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
366 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
367 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
368 	{REG_A6XX_RBBM_ISDB_CNT, 0x00000182},
369 	{REG_A6XX_RBBM_RAC_THRESHOLD_CNT, 0x00000000},
370 	{REG_A6XX_RBBM_SP_HYST_CNT, 0x00000000},
371 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
372 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
373 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
374 	{},
375 };
376 
377 const struct adreno_reglist a650_hwcg[] = {
378 	{REG_A6XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
379 	{REG_A6XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
380 	{REG_A6XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
381 	{REG_A6XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
382 	{REG_A6XX_RBBM_CLOCK_CNTL_TP0, 0x02222222},
383 	{REG_A6XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
384 	{REG_A6XX_RBBM_CLOCK_CNTL3_TP0, 0x22222222},
385 	{REG_A6XX_RBBM_CLOCK_CNTL4_TP0, 0x00022222},
386 	{REG_A6XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
387 	{REG_A6XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
388 	{REG_A6XX_RBBM_CLOCK_DELAY3_TP0, 0x11111111},
389 	{REG_A6XX_RBBM_CLOCK_DELAY4_TP0, 0x00011111},
390 	{REG_A6XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
391 	{REG_A6XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
392 	{REG_A6XX_RBBM_CLOCK_HYST3_TP0, 0x77777777},
393 	{REG_A6XX_RBBM_CLOCK_HYST4_TP0, 0x00077777},
394 	{REG_A6XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
395 	{REG_A6XX_RBBM_CLOCK_CNTL2_RB0, 0x01002222},
396 	{REG_A6XX_RBBM_CLOCK_CNTL_CCU0, 0x00002220},
397 	{REG_A6XX_RBBM_CLOCK_HYST_RB_CCU0, 0x00040F00},
398 	{REG_A6XX_RBBM_CLOCK_CNTL_RAC, 0x25222022},
399 	{REG_A6XX_RBBM_CLOCK_CNTL2_RAC, 0x00005555},
400 	{REG_A6XX_RBBM_CLOCK_DELAY_RAC, 0x00000011},
401 	{REG_A6XX_RBBM_CLOCK_HYST_RAC, 0x00445044},
402 	{REG_A6XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
403 	{REG_A6XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
404 	{REG_A6XX_RBBM_CLOCK_MODE_GPC, 0x00222222},
405 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ_2, 0x00000002},
406 	{REG_A6XX_RBBM_CLOCK_MODE_HLSQ, 0x00002222},
407 	{REG_A6XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
408 	{REG_A6XX_RBBM_CLOCK_DELAY_VFD, 0x00002222},
409 	{REG_A6XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
410 	{REG_A6XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
411 	{REG_A6XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
412 	{REG_A6XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
413 	{REG_A6XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
414 	{REG_A6XX_RBBM_CLOCK_HYST_HLSQ, 0x00000000},
415 	{REG_A6XX_RBBM_CLOCK_CNTL_TEX_FCHE, 0x00000222},
416 	{REG_A6XX_RBBM_CLOCK_DELAY_TEX_FCHE, 0x00000111},
417 	{REG_A6XX_RBBM_CLOCK_HYST_TEX_FCHE, 0x00000777},
418 	{REG_A6XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
419 	{REG_A6XX_RBBM_CLOCK_HYST_UCHE, 0x00000004},
420 	{REG_A6XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
421 	{REG_A6XX_RBBM_ISDB_CNT, 0x00000182},
422 	{REG_A6XX_RBBM_RAC_THRESHOLD_CNT, 0x00000000},
423 	{REG_A6XX_RBBM_SP_HYST_CNT, 0x00000000},
424 	{REG_A6XX_RBBM_CLOCK_CNTL_GMU_GX, 0x00000222},
425 	{REG_A6XX_RBBM_CLOCK_DELAY_GMU_GX, 0x00000111},
426 	{REG_A6XX_RBBM_CLOCK_HYST_GMU_GX, 0x00000555},
427 	{},
428 };
429 
430 static void a6xx_set_hwcg(struct msm_gpu *gpu, bool state)
431 {
432 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
433 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
434 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
435 	const struct adreno_reglist *reg;
436 	unsigned int i;
437 	u32 val, clock_cntl_on;
438 
439 	if (!adreno_gpu->info->hwcg)
440 		return;
441 
442 	if (adreno_is_a630(adreno_gpu))
443 		clock_cntl_on = 0x8aa8aa02;
444 	else
445 		clock_cntl_on = 0x8aa8aa82;
446 
447 	val = gpu_read(gpu, REG_A6XX_RBBM_CLOCK_CNTL);
448 
449 	/* Don't re-program the registers if they are already correct */
450 	if ((!state && !val) || (state && (val == clock_cntl_on)))
451 		return;
452 
453 	/* Disable SP clock before programming HWCG registers */
454 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 1, 0);
455 
456 	for (i = 0; (reg = &adreno_gpu->info->hwcg[i], reg->offset); i++)
457 		gpu_write(gpu, reg->offset, state ? reg->value : 0);
458 
459 	/* Enable SP clock */
460 	gmu_rmw(gmu, REG_A6XX_GPU_GMU_GX_SPTPRAC_CLOCK_CONTROL, 0, 1);
461 
462 	gpu_write(gpu, REG_A6XX_RBBM_CLOCK_CNTL, state ? clock_cntl_on : 0);
463 }
464 
465 static void a6xx_set_ubwc_config(struct msm_gpu *gpu)
466 {
467 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
468 	u32 lower_bit = 2;
469 	u32 amsbc = 0;
470 	u32 rgb565_predicator = 0;
471 	u32 uavflagprd_inv = 0;
472 
473 	/* a618 is using the hw default values */
474 	if (adreno_is_a618(adreno_gpu))
475 		return;
476 
477 	if (adreno_is_a640(adreno_gpu))
478 		amsbc = 1;
479 
480 	if (adreno_is_a650(adreno_gpu)) {
481 		/* TODO: get ddr type from bootloader and use 2 for LPDDR4 */
482 		lower_bit = 3;
483 		amsbc = 1;
484 		rgb565_predicator = 1;
485 		uavflagprd_inv = 2;
486 	}
487 
488 	gpu_write(gpu, REG_A6XX_RB_NC_MODE_CNTL,
489 		rgb565_predicator << 11 | amsbc << 4 | lower_bit << 1);
490 	gpu_write(gpu, REG_A6XX_TPL1_NC_MODE_CNTL, lower_bit << 1);
491 	gpu_write(gpu, REG_A6XX_SP_NC_MODE_CNTL,
492 		uavflagprd_inv >> 4 | lower_bit << 1);
493 	gpu_write(gpu, REG_A6XX_UCHE_MODE_CNTL, lower_bit << 21);
494 }
495 
496 static int a6xx_cp_init(struct msm_gpu *gpu)
497 {
498 	struct msm_ringbuffer *ring = gpu->rb[0];
499 
500 	OUT_PKT7(ring, CP_ME_INIT, 8);
501 
502 	OUT_RING(ring, 0x0000002f);
503 
504 	/* Enable multiple hardware contexts */
505 	OUT_RING(ring, 0x00000003);
506 
507 	/* Enable error detection */
508 	OUT_RING(ring, 0x20000000);
509 
510 	/* Don't enable header dump */
511 	OUT_RING(ring, 0x00000000);
512 	OUT_RING(ring, 0x00000000);
513 
514 	/* No workarounds enabled */
515 	OUT_RING(ring, 0x00000000);
516 
517 	/* Pad rest of the cmds with 0's */
518 	OUT_RING(ring, 0x00000000);
519 	OUT_RING(ring, 0x00000000);
520 
521 	a6xx_flush(gpu, ring);
522 	return a6xx_idle(gpu, ring) ? 0 : -EINVAL;
523 }
524 
525 /*
526  * Check that the microcode version is new enough to include several key
527  * security fixes. Return true if the ucode is safe.
528  */
529 static bool a6xx_ucode_check_version(struct a6xx_gpu *a6xx_gpu,
530 		struct drm_gem_object *obj)
531 {
532 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
533 	struct msm_gpu *gpu = &adreno_gpu->base;
534 	u32 *buf = msm_gem_get_vaddr(obj);
535 	bool ret = false;
536 
537 	if (IS_ERR(buf))
538 		return false;
539 
540 	/*
541 	 * Targets up to a640 (a618, a630 and a640) need to check for a
542 	 * microcode version that is patched to support the whereami opcode or
543 	 * one that is new enough to include it by default.
544 	 */
545 	if (adreno_is_a618(adreno_gpu) || adreno_is_a630(adreno_gpu) ||
546 		adreno_is_a640(adreno_gpu)) {
547 		/*
548 		 * If the lowest nibble is 0xa that is an indication that this
549 		 * microcode has been patched. The actual version is in dword
550 		 * [3] but we only care about the patchlevel which is the lowest
551 		 * nibble of dword [3]
552 		 *
553 		 * Otherwise check that the firmware is greater than or equal
554 		 * to 1.90 which was the first version that had this fix built
555 		 * in
556 		 */
557 		if ((((buf[0] & 0xf) == 0xa) && (buf[2] & 0xf) >= 1) ||
558 			(buf[0] & 0xfff) >= 0x190) {
559 			a6xx_gpu->has_whereami = true;
560 			ret = true;
561 			goto out;
562 		}
563 
564 		DRM_DEV_ERROR(&gpu->pdev->dev,
565 			"a630 SQE ucode is too old. Have version %x need at least %x\n",
566 			buf[0] & 0xfff, 0x190);
567 	}  else {
568 		/*
569 		 * a650 tier targets don't need whereami but still need to be
570 		 * equal to or newer than 0.95 for other security fixes
571 		 */
572 		if (adreno_is_a650(adreno_gpu)) {
573 			if ((buf[0] & 0xfff) >= 0x095) {
574 				ret = true;
575 				goto out;
576 			}
577 
578 			DRM_DEV_ERROR(&gpu->pdev->dev,
579 				"a650 SQE ucode is too old. Have version %x need at least %x\n",
580 				buf[0] & 0xfff, 0x095);
581 		}
582 
583 		/*
584 		 * When a660 is added those targets should return true here
585 		 * since those have all the critical security fixes built in
586 		 * from the start
587 		 */
588 	}
589 out:
590 	msm_gem_put_vaddr(obj);
591 	return ret;
592 }
593 
594 static int a6xx_ucode_init(struct msm_gpu *gpu)
595 {
596 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
597 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
598 
599 	if (!a6xx_gpu->sqe_bo) {
600 		a6xx_gpu->sqe_bo = adreno_fw_create_bo(gpu,
601 			adreno_gpu->fw[ADRENO_FW_SQE], &a6xx_gpu->sqe_iova);
602 
603 		if (IS_ERR(a6xx_gpu->sqe_bo)) {
604 			int ret = PTR_ERR(a6xx_gpu->sqe_bo);
605 
606 			a6xx_gpu->sqe_bo = NULL;
607 			DRM_DEV_ERROR(&gpu->pdev->dev,
608 				"Could not allocate SQE ucode: %d\n", ret);
609 
610 			return ret;
611 		}
612 
613 		msm_gem_object_set_name(a6xx_gpu->sqe_bo, "sqefw");
614 		if (!a6xx_ucode_check_version(a6xx_gpu, a6xx_gpu->sqe_bo)) {
615 			msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
616 			drm_gem_object_put(a6xx_gpu->sqe_bo);
617 
618 			a6xx_gpu->sqe_bo = NULL;
619 			return -EPERM;
620 		}
621 	}
622 
623 	gpu_write64(gpu, REG_A6XX_CP_SQE_INSTR_BASE_LO,
624 		REG_A6XX_CP_SQE_INSTR_BASE_HI, a6xx_gpu->sqe_iova);
625 
626 	return 0;
627 }
628 
629 static int a6xx_zap_shader_init(struct msm_gpu *gpu)
630 {
631 	static bool loaded;
632 	int ret;
633 
634 	if (loaded)
635 		return 0;
636 
637 	ret = adreno_zap_shader_load(gpu, GPU_PAS_ID);
638 
639 	loaded = !ret;
640 	return ret;
641 }
642 
643 #define A6XX_INT_MASK (A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR | \
644 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW | \
645 	  A6XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
646 	  A6XX_RBBM_INT_0_MASK_CP_IB2 | \
647 	  A6XX_RBBM_INT_0_MASK_CP_IB1 | \
648 	  A6XX_RBBM_INT_0_MASK_CP_RB | \
649 	  A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
650 	  A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW | \
651 	  A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT | \
652 	  A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
653 	  A6XX_RBBM_INT_0_MASK_UCHE_TRAP_INTR)
654 
655 static int a6xx_hw_init(struct msm_gpu *gpu)
656 {
657 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
658 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
659 	int ret;
660 
661 	/* Make sure the GMU keeps the GPU on while we set it up */
662 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
663 
664 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_CNTL, 0);
665 
666 	/*
667 	 * Disable the trusted memory range - we don't actually supported secure
668 	 * memory rendering at this point in time and we don't want to block off
669 	 * part of the virtual memory space.
670 	 */
671 	gpu_write64(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
672 		REG_A6XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
673 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
674 
675 	/* Turn on 64 bit addressing for all blocks */
676 	gpu_write(gpu, REG_A6XX_CP_ADDR_MODE_CNTL, 0x1);
677 	gpu_write(gpu, REG_A6XX_VSC_ADDR_MODE_CNTL, 0x1);
678 	gpu_write(gpu, REG_A6XX_GRAS_ADDR_MODE_CNTL, 0x1);
679 	gpu_write(gpu, REG_A6XX_RB_ADDR_MODE_CNTL, 0x1);
680 	gpu_write(gpu, REG_A6XX_PC_ADDR_MODE_CNTL, 0x1);
681 	gpu_write(gpu, REG_A6XX_HLSQ_ADDR_MODE_CNTL, 0x1);
682 	gpu_write(gpu, REG_A6XX_VFD_ADDR_MODE_CNTL, 0x1);
683 	gpu_write(gpu, REG_A6XX_VPC_ADDR_MODE_CNTL, 0x1);
684 	gpu_write(gpu, REG_A6XX_UCHE_ADDR_MODE_CNTL, 0x1);
685 	gpu_write(gpu, REG_A6XX_SP_ADDR_MODE_CNTL, 0x1);
686 	gpu_write(gpu, REG_A6XX_TPL1_ADDR_MODE_CNTL, 0x1);
687 	gpu_write(gpu, REG_A6XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1);
688 
689 	/* enable hardware clockgating */
690 	a6xx_set_hwcg(gpu, true);
691 
692 	/* VBIF/GBIF start*/
693 	if (adreno_is_a640(adreno_gpu) || adreno_is_a650(adreno_gpu)) {
694 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE0, 0x00071620);
695 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE1, 0x00071620);
696 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE2, 0x00071620);
697 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
698 		gpu_write(gpu, REG_A6XX_GBIF_QSB_SIDE3, 0x00071620);
699 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_CLIENT_QOS_CNTL, 0x3);
700 	} else {
701 		gpu_write(gpu, REG_A6XX_RBBM_VBIF_CLIENT_QOS_CNTL, 0x3);
702 	}
703 
704 	if (adreno_is_a630(adreno_gpu))
705 		gpu_write(gpu, REG_A6XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009);
706 
707 	/* Make all blocks contribute to the GPU BUSY perf counter */
708 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xffffffff);
709 
710 	/* Disable L2 bypass in the UCHE */
711 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_LO, 0xffffffc0);
712 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_RANGE_MAX_HI, 0x0001ffff);
713 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_LO, 0xfffff000);
714 	gpu_write(gpu, REG_A6XX_UCHE_TRAP_BASE_HI, 0x0001ffff);
715 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_LO, 0xfffff000);
716 	gpu_write(gpu, REG_A6XX_UCHE_WRITE_THRU_BASE_HI, 0x0001ffff);
717 
718 	if (!adreno_is_a650(adreno_gpu)) {
719 		/* Set the GMEM VA range [0x100000:0x100000 + gpu->gmem - 1] */
720 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MIN_LO,
721 			REG_A6XX_UCHE_GMEM_RANGE_MIN_HI, 0x00100000);
722 
723 		gpu_write64(gpu, REG_A6XX_UCHE_GMEM_RANGE_MAX_LO,
724 			REG_A6XX_UCHE_GMEM_RANGE_MAX_HI,
725 			0x00100000 + adreno_gpu->gmem - 1);
726 	}
727 
728 	gpu_write(gpu, REG_A6XX_UCHE_FILTER_CNTL, 0x804);
729 	gpu_write(gpu, REG_A6XX_UCHE_CACHE_WAYS, 0x4);
730 
731 	if (adreno_is_a640(adreno_gpu) || adreno_is_a650(adreno_gpu))
732 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x02000140);
733 	else
734 		gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_2, 0x010000c0);
735 	gpu_write(gpu, REG_A6XX_CP_ROQ_THRESHOLDS_1, 0x8040362c);
736 
737 	/* Setting the mem pool size */
738 	gpu_write(gpu, REG_A6XX_CP_MEM_POOL_SIZE, 128);
739 
740 	/* Setting the primFifo thresholds default values */
741 	if (adreno_is_a650(adreno_gpu))
742 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00300000);
743 	else if (adreno_is_a640(adreno_gpu))
744 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, 0x00200000);
745 	else
746 		gpu_write(gpu, REG_A6XX_PC_DBG_ECO_CNTL, (0x300 << 11));
747 
748 	/* Set the AHB default slave response to "ERROR" */
749 	gpu_write(gpu, REG_A6XX_CP_AHB_CNTL, 0x1);
750 
751 	/* Turn on performance counters */
752 	gpu_write(gpu, REG_A6XX_RBBM_PERFCTR_CNTL, 0x1);
753 
754 	/* Select CP0 to always count cycles */
755 	gpu_write(gpu, REG_A6XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT);
756 
757 	a6xx_set_ubwc_config(gpu);
758 
759 	/* Enable fault detection */
760 	gpu_write(gpu, REG_A6XX_RBBM_INTERFACE_HANG_INT_CNTL,
761 		(1 << 30) | 0x1fffff);
762 
763 	gpu_write(gpu, REG_A6XX_UCHE_CLIENT_PF, 1);
764 
765 	/* Set weights for bicubic filtering */
766 	if (adreno_is_a650(adreno_gpu)) {
767 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_0, 0);
768 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_1,
769 			0x3fe05ff4);
770 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_2,
771 			0x3fa0ebee);
772 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_3,
773 			0x3f5193ed);
774 		gpu_write(gpu, REG_A6XX_TPL1_BICUBIC_WEIGHTS_TABLE_4,
775 			0x3f0243f0);
776 	}
777 
778 	/* Protect registers from the CP */
779 	gpu_write(gpu, REG_A6XX_CP_PROTECT_CNTL, 0x00000003);
780 
781 	gpu_write(gpu, REG_A6XX_CP_PROTECT(0),
782 		A6XX_PROTECT_RDONLY(0x600, 0x51));
783 	gpu_write(gpu, REG_A6XX_CP_PROTECT(1), A6XX_PROTECT_RW(0xae50, 0x2));
784 	gpu_write(gpu, REG_A6XX_CP_PROTECT(2), A6XX_PROTECT_RW(0x9624, 0x13));
785 	gpu_write(gpu, REG_A6XX_CP_PROTECT(3), A6XX_PROTECT_RW(0x8630, 0x8));
786 	gpu_write(gpu, REG_A6XX_CP_PROTECT(4), A6XX_PROTECT_RW(0x9e70, 0x1));
787 	gpu_write(gpu, REG_A6XX_CP_PROTECT(5), A6XX_PROTECT_RW(0x9e78, 0x187));
788 	gpu_write(gpu, REG_A6XX_CP_PROTECT(6), A6XX_PROTECT_RW(0xf000, 0x810));
789 	gpu_write(gpu, REG_A6XX_CP_PROTECT(7),
790 		A6XX_PROTECT_RDONLY(0xfc00, 0x3));
791 	gpu_write(gpu, REG_A6XX_CP_PROTECT(8), A6XX_PROTECT_RW(0x50e, 0x0));
792 	gpu_write(gpu, REG_A6XX_CP_PROTECT(9), A6XX_PROTECT_RDONLY(0x50f, 0x0));
793 	gpu_write(gpu, REG_A6XX_CP_PROTECT(10), A6XX_PROTECT_RW(0x510, 0x0));
794 	gpu_write(gpu, REG_A6XX_CP_PROTECT(11),
795 		A6XX_PROTECT_RDONLY(0x0, 0x4f9));
796 	gpu_write(gpu, REG_A6XX_CP_PROTECT(12),
797 		A6XX_PROTECT_RDONLY(0x501, 0xa));
798 	gpu_write(gpu, REG_A6XX_CP_PROTECT(13),
799 		A6XX_PROTECT_RDONLY(0x511, 0x44));
800 	gpu_write(gpu, REG_A6XX_CP_PROTECT(14), A6XX_PROTECT_RW(0xe00, 0xe));
801 	gpu_write(gpu, REG_A6XX_CP_PROTECT(15), A6XX_PROTECT_RW(0x8e00, 0x0));
802 	gpu_write(gpu, REG_A6XX_CP_PROTECT(16), A6XX_PROTECT_RW(0x8e50, 0xf));
803 	gpu_write(gpu, REG_A6XX_CP_PROTECT(17), A6XX_PROTECT_RW(0xbe02, 0x0));
804 	gpu_write(gpu, REG_A6XX_CP_PROTECT(18),
805 		A6XX_PROTECT_RW(0xbe20, 0x11f3));
806 	gpu_write(gpu, REG_A6XX_CP_PROTECT(19), A6XX_PROTECT_RW(0x800, 0x82));
807 	gpu_write(gpu, REG_A6XX_CP_PROTECT(20), A6XX_PROTECT_RW(0x8a0, 0x8));
808 	gpu_write(gpu, REG_A6XX_CP_PROTECT(21), A6XX_PROTECT_RW(0x8ab, 0x19));
809 	gpu_write(gpu, REG_A6XX_CP_PROTECT(22), A6XX_PROTECT_RW(0x900, 0x4d));
810 	gpu_write(gpu, REG_A6XX_CP_PROTECT(23), A6XX_PROTECT_RW(0x98d, 0x76));
811 	gpu_write(gpu, REG_A6XX_CP_PROTECT(24),
812 			A6XX_PROTECT_RDONLY(0x980, 0x4));
813 	gpu_write(gpu, REG_A6XX_CP_PROTECT(25), A6XX_PROTECT_RW(0xa630, 0x0));
814 
815 	/* Enable expanded apriv for targets that support it */
816 	if (gpu->hw_apriv) {
817 		gpu_write(gpu, REG_A6XX_CP_APRIV_CNTL,
818 			(1 << 6) | (1 << 5) | (1 << 3) | (1 << 2) | (1 << 1));
819 	}
820 
821 	/* Enable interrupts */
822 	gpu_write(gpu, REG_A6XX_RBBM_INT_0_MASK, A6XX_INT_MASK);
823 
824 	ret = adreno_hw_init(gpu);
825 	if (ret)
826 		goto out;
827 
828 	ret = a6xx_ucode_init(gpu);
829 	if (ret)
830 		goto out;
831 
832 	/* Set the ringbuffer address */
833 	gpu_write64(gpu, REG_A6XX_CP_RB_BASE, REG_A6XX_CP_RB_BASE_HI,
834 		gpu->rb[0]->iova);
835 
836 	/* Targets that support extended APRIV can use the RPTR shadow from
837 	 * hardware but all the other ones need to disable the feature. Targets
838 	 * that support the WHERE_AM_I opcode can use that instead
839 	 */
840 	if (adreno_gpu->base.hw_apriv)
841 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL, MSM_GPU_RB_CNTL_DEFAULT);
842 	else
843 		gpu_write(gpu, REG_A6XX_CP_RB_CNTL,
844 			MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE);
845 
846 	/*
847 	 * Expanded APRIV and targets that support WHERE_AM_I both need a
848 	 * privileged buffer to store the RPTR shadow
849 	 */
850 
851 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami) {
852 		if (!a6xx_gpu->shadow_bo) {
853 			a6xx_gpu->shadow = msm_gem_kernel_new_locked(gpu->dev,
854 				sizeof(u32) * gpu->nr_rings,
855 				MSM_BO_UNCACHED | MSM_BO_MAP_PRIV,
856 				gpu->aspace, &a6xx_gpu->shadow_bo,
857 				&a6xx_gpu->shadow_iova);
858 
859 			if (IS_ERR(a6xx_gpu->shadow))
860 				return PTR_ERR(a6xx_gpu->shadow);
861 		}
862 
863 		gpu_write64(gpu, REG_A6XX_CP_RB_RPTR_ADDR_LO,
864 			REG_A6XX_CP_RB_RPTR_ADDR_HI,
865 			shadowptr(a6xx_gpu, gpu->rb[0]));
866 	}
867 
868 	/* Always come up on rb 0 */
869 	a6xx_gpu->cur_ring = gpu->rb[0];
870 
871 	a6xx_gpu->cur_ctx = NULL;
872 
873 	/* Enable the SQE_to start the CP engine */
874 	gpu_write(gpu, REG_A6XX_CP_SQE_CNTL, 1);
875 
876 	ret = a6xx_cp_init(gpu);
877 	if (ret)
878 		goto out;
879 
880 	/*
881 	 * Try to load a zap shader into the secure world. If successful
882 	 * we can use the CP to switch out of secure mode. If not then we
883 	 * have no resource but to try to switch ourselves out manually. If we
884 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
885 	 * be blocked and a permissions violation will soon follow.
886 	 */
887 	ret = a6xx_zap_shader_init(gpu);
888 	if (!ret) {
889 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
890 		OUT_RING(gpu->rb[0], 0x00000000);
891 
892 		a6xx_flush(gpu, gpu->rb[0]);
893 		if (!a6xx_idle(gpu, gpu->rb[0]))
894 			return -EINVAL;
895 	} else if (ret == -ENODEV) {
896 		/*
897 		 * This device does not use zap shader (but print a warning
898 		 * just in case someone got their dt wrong.. hopefully they
899 		 * have a debug UART to realize the error of their ways...
900 		 * if you mess this up you are about to crash horribly)
901 		 */
902 		dev_warn_once(gpu->dev->dev,
903 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
904 		gpu_write(gpu, REG_A6XX_RBBM_SECVID_TRUST_CNTL, 0x0);
905 		ret = 0;
906 	} else {
907 		return ret;
908 	}
909 
910 out:
911 	/*
912 	 * Tell the GMU that we are done touching the GPU and it can start power
913 	 * management
914 	 */
915 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
916 
917 	if (a6xx_gpu->gmu.legacy) {
918 		/* Take the GMU out of its special boot mode */
919 		a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_BOOT_SLUMBER);
920 	}
921 
922 	return ret;
923 }
924 
925 static void a6xx_dump(struct msm_gpu *gpu)
926 {
927 	DRM_DEV_INFO(&gpu->pdev->dev, "status:   %08x\n",
928 			gpu_read(gpu, REG_A6XX_RBBM_STATUS));
929 	adreno_dump(gpu);
930 }
931 
932 #define VBIF_RESET_ACK_TIMEOUT	100
933 #define VBIF_RESET_ACK_MASK	0x00f0
934 
935 static void a6xx_recover(struct msm_gpu *gpu)
936 {
937 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
938 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
939 	int i;
940 
941 	adreno_dump_info(gpu);
942 
943 	for (i = 0; i < 8; i++)
944 		DRM_DEV_INFO(&gpu->pdev->dev, "CP_SCRATCH_REG%d: %u\n", i,
945 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(i)));
946 
947 	if (hang_debug)
948 		a6xx_dump(gpu);
949 
950 	/*
951 	 * Turn off keep alive that might have been enabled by the hang
952 	 * interrupt
953 	 */
954 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
955 
956 	gpu->funcs->pm_suspend(gpu);
957 	gpu->funcs->pm_resume(gpu);
958 
959 	msm_gpu_hw_init(gpu);
960 }
961 
962 static int a6xx_fault_handler(void *arg, unsigned long iova, int flags)
963 {
964 	struct msm_gpu *gpu = arg;
965 
966 	pr_warn_ratelimited("*** gpu fault: iova=%08lx, flags=%d (%u,%u,%u,%u)\n",
967 			iova, flags,
968 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
969 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
970 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
971 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
972 
973 	return -EFAULT;
974 }
975 
976 static void a6xx_cp_hw_err_irq(struct msm_gpu *gpu)
977 {
978 	u32 status = gpu_read(gpu, REG_A6XX_CP_INTERRUPT_STATUS);
979 
980 	if (status & A6XX_CP_INT_CP_OPCODE_ERROR) {
981 		u32 val;
982 
983 		gpu_write(gpu, REG_A6XX_CP_SQE_STAT_ADDR, 1);
984 		val = gpu_read(gpu, REG_A6XX_CP_SQE_STAT_DATA);
985 		dev_err_ratelimited(&gpu->pdev->dev,
986 			"CP | opcode error | possible opcode=0x%8.8X\n",
987 			val);
988 	}
989 
990 	if (status & A6XX_CP_INT_CP_UCODE_ERROR)
991 		dev_err_ratelimited(&gpu->pdev->dev,
992 			"CP ucode error interrupt\n");
993 
994 	if (status & A6XX_CP_INT_CP_HW_FAULT_ERROR)
995 		dev_err_ratelimited(&gpu->pdev->dev, "CP | HW fault | status=0x%8.8X\n",
996 			gpu_read(gpu, REG_A6XX_CP_HW_FAULT));
997 
998 	if (status & A6XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
999 		u32 val = gpu_read(gpu, REG_A6XX_CP_PROTECT_STATUS);
1000 
1001 		dev_err_ratelimited(&gpu->pdev->dev,
1002 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
1003 			val & (1 << 20) ? "READ" : "WRITE",
1004 			(val & 0x3ffff), val);
1005 	}
1006 
1007 	if (status & A6XX_CP_INT_CP_AHB_ERROR)
1008 		dev_err_ratelimited(&gpu->pdev->dev, "CP AHB error interrupt\n");
1009 
1010 	if (status & A6XX_CP_INT_CP_VSD_PARITY_ERROR)
1011 		dev_err_ratelimited(&gpu->pdev->dev, "CP VSD decoder parity error\n");
1012 
1013 	if (status & A6XX_CP_INT_CP_ILLEGAL_INSTR_ERROR)
1014 		dev_err_ratelimited(&gpu->pdev->dev, "CP illegal instruction error\n");
1015 
1016 }
1017 
1018 static void a6xx_fault_detect_irq(struct msm_gpu *gpu)
1019 {
1020 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1021 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1022 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
1023 
1024 	/*
1025 	 * Force the GPU to stay on until after we finish
1026 	 * collecting information
1027 	 */
1028 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 1);
1029 
1030 	DRM_DEV_ERROR(&gpu->pdev->dev,
1031 		"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",
1032 		ring ? ring->id : -1, ring ? ring->seqno : 0,
1033 		gpu_read(gpu, REG_A6XX_RBBM_STATUS),
1034 		gpu_read(gpu, REG_A6XX_CP_RB_RPTR),
1035 		gpu_read(gpu, REG_A6XX_CP_RB_WPTR),
1036 		gpu_read64(gpu, REG_A6XX_CP_IB1_BASE, REG_A6XX_CP_IB1_BASE_HI),
1037 		gpu_read(gpu, REG_A6XX_CP_IB1_REM_SIZE),
1038 		gpu_read64(gpu, REG_A6XX_CP_IB2_BASE, REG_A6XX_CP_IB2_BASE_HI),
1039 		gpu_read(gpu, REG_A6XX_CP_IB2_REM_SIZE));
1040 
1041 	/* Turn off the hangcheck timer to keep it from bothering us */
1042 	del_timer(&gpu->hangcheck_timer);
1043 
1044 	kthread_queue_work(gpu->worker, &gpu->recover_work);
1045 }
1046 
1047 static irqreturn_t a6xx_irq(struct msm_gpu *gpu)
1048 {
1049 	u32 status = gpu_read(gpu, REG_A6XX_RBBM_INT_0_STATUS);
1050 
1051 	gpu_write(gpu, REG_A6XX_RBBM_INT_CLEAR_CMD, status);
1052 
1053 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_HANG_DETECT)
1054 		a6xx_fault_detect_irq(gpu);
1055 
1056 	if (status & A6XX_RBBM_INT_0_MASK_CP_AHB_ERROR)
1057 		dev_err_ratelimited(&gpu->pdev->dev, "CP | AHB bus error\n");
1058 
1059 	if (status & A6XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1060 		a6xx_cp_hw_err_irq(gpu);
1061 
1062 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNCFIFO_OVERFLOW)
1063 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB ASYNC overflow\n");
1064 
1065 	if (status & A6XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
1066 		dev_err_ratelimited(&gpu->pdev->dev, "RBBM | ATB bus overflow\n");
1067 
1068 	if (status & A6XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1069 		dev_err_ratelimited(&gpu->pdev->dev, "UCHE | Out of bounds access\n");
1070 
1071 	if (status & A6XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS)
1072 		msm_gpu_retire(gpu);
1073 
1074 	return IRQ_HANDLED;
1075 }
1076 
1077 static void a6xx_llc_rmw(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 mask, u32 or)
1078 {
1079 	return msm_rmw(a6xx_gpu->llc_mmio + (reg << 2), mask, or);
1080 }
1081 
1082 static void a6xx_llc_write(struct a6xx_gpu *a6xx_gpu, u32 reg, u32 value)
1083 {
1084 	return msm_writel(value, a6xx_gpu->llc_mmio + (reg << 2));
1085 }
1086 
1087 static void a6xx_llc_deactivate(struct a6xx_gpu *a6xx_gpu)
1088 {
1089 	llcc_slice_deactivate(a6xx_gpu->llc_slice);
1090 	llcc_slice_deactivate(a6xx_gpu->htw_llc_slice);
1091 }
1092 
1093 static void a6xx_llc_activate(struct a6xx_gpu *a6xx_gpu)
1094 {
1095 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1096 	struct msm_gpu *gpu = &adreno_gpu->base;
1097 	u32 cntl1_regval = 0;
1098 
1099 	if (IS_ERR(a6xx_gpu->llc_mmio))
1100 		return;
1101 
1102 	if (!llcc_slice_activate(a6xx_gpu->llc_slice)) {
1103 		u32 gpu_scid = llcc_get_slice_id(a6xx_gpu->llc_slice);
1104 
1105 		gpu_scid &= 0x1f;
1106 		cntl1_regval = (gpu_scid << 0) | (gpu_scid << 5) | (gpu_scid << 10) |
1107 			       (gpu_scid << 15) | (gpu_scid << 20);
1108 	}
1109 
1110 	/*
1111 	 * For targets with a MMU500, activate the slice but don't program the
1112 	 * register.  The XBL will take care of that.
1113 	 */
1114 	if (!llcc_slice_activate(a6xx_gpu->htw_llc_slice)) {
1115 		if (!a6xx_gpu->have_mmu500) {
1116 			u32 gpuhtw_scid = llcc_get_slice_id(a6xx_gpu->htw_llc_slice);
1117 
1118 			gpuhtw_scid &= 0x1f;
1119 			cntl1_regval |= FIELD_PREP(GENMASK(29, 25), gpuhtw_scid);
1120 		}
1121 	}
1122 
1123 	if (cntl1_regval) {
1124 		/*
1125 		 * Program the slice IDs for the various GPU blocks and GPU MMU
1126 		 * pagetables
1127 		 */
1128 		if (a6xx_gpu->have_mmu500)
1129 			gpu_rmw(gpu, REG_A6XX_GBIF_SCACHE_CNTL1, GENMASK(24, 0),
1130 				cntl1_regval);
1131 		else {
1132 			a6xx_llc_write(a6xx_gpu,
1133 				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_1, cntl1_regval);
1134 
1135 			/*
1136 			 * Program cacheability overrides to not allocate cache
1137 			 * lines on a write miss
1138 			 */
1139 			a6xx_llc_rmw(a6xx_gpu,
1140 				REG_A6XX_CX_MISC_SYSTEM_CACHE_CNTL_0, 0xF, 0x03);
1141 		}
1142 	}
1143 }
1144 
1145 static void a6xx_llc_slices_destroy(struct a6xx_gpu *a6xx_gpu)
1146 {
1147 	llcc_slice_putd(a6xx_gpu->llc_slice);
1148 	llcc_slice_putd(a6xx_gpu->htw_llc_slice);
1149 }
1150 
1151 static void a6xx_llc_slices_init(struct platform_device *pdev,
1152 		struct a6xx_gpu *a6xx_gpu)
1153 {
1154 	struct device_node *phandle;
1155 
1156 	a6xx_gpu->llc_mmio = msm_ioremap(pdev, "cx_mem", "gpu_cx");
1157 	if (IS_ERR(a6xx_gpu->llc_mmio))
1158 		return;
1159 
1160 	/*
1161 	 * There is a different programming path for targets with an mmu500
1162 	 * attached, so detect if that is the case
1163 	 */
1164 	phandle = of_parse_phandle(pdev->dev.of_node, "iommus", 0);
1165 	a6xx_gpu->have_mmu500 = (phandle &&
1166 		of_device_is_compatible(phandle, "arm,mmu-500"));
1167 	of_node_put(phandle);
1168 
1169 	a6xx_gpu->llc_slice = llcc_slice_getd(LLCC_GPU);
1170 	a6xx_gpu->htw_llc_slice = llcc_slice_getd(LLCC_GPUHTW);
1171 
1172 	if (IS_ERR_OR_NULL(a6xx_gpu->llc_slice) && IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice))
1173 		a6xx_gpu->llc_mmio = ERR_PTR(-EINVAL);
1174 }
1175 
1176 static int a6xx_pm_resume(struct msm_gpu *gpu)
1177 {
1178 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1179 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1180 	int ret;
1181 
1182 	gpu->needs_hw_init = true;
1183 
1184 	trace_msm_gpu_resume(0);
1185 
1186 	ret = a6xx_gmu_resume(a6xx_gpu);
1187 	if (ret)
1188 		return ret;
1189 
1190 	msm_gpu_resume_devfreq(gpu);
1191 
1192 	a6xx_llc_activate(a6xx_gpu);
1193 
1194 	return 0;
1195 }
1196 
1197 static int a6xx_pm_suspend(struct msm_gpu *gpu)
1198 {
1199 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1200 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1201 	int i, ret;
1202 
1203 	trace_msm_gpu_suspend(0);
1204 
1205 	a6xx_llc_deactivate(a6xx_gpu);
1206 
1207 	devfreq_suspend_device(gpu->devfreq.devfreq);
1208 
1209 	ret = a6xx_gmu_stop(a6xx_gpu);
1210 	if (ret)
1211 		return ret;
1212 
1213 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami)
1214 		for (i = 0; i < gpu->nr_rings; i++)
1215 			a6xx_gpu->shadow[i] = 0;
1216 
1217 	return 0;
1218 }
1219 
1220 static int a6xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1221 {
1222 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1223 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1224 	static DEFINE_MUTEX(perfcounter_oob);
1225 
1226 	mutex_lock(&perfcounter_oob);
1227 
1228 	/* Force the GPU power on so we can read this register */
1229 	a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1230 
1231 	*value = gpu_read64(gpu, REG_A6XX_CP_ALWAYS_ON_COUNTER_LO,
1232 		REG_A6XX_CP_ALWAYS_ON_COUNTER_HI);
1233 
1234 	a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_PERFCOUNTER_SET);
1235 	mutex_unlock(&perfcounter_oob);
1236 	return 0;
1237 }
1238 
1239 static struct msm_ringbuffer *a6xx_active_ring(struct msm_gpu *gpu)
1240 {
1241 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1242 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1243 
1244 	return a6xx_gpu->cur_ring;
1245 }
1246 
1247 static void a6xx_destroy(struct msm_gpu *gpu)
1248 {
1249 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1250 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1251 
1252 	if (a6xx_gpu->sqe_bo) {
1253 		msm_gem_unpin_iova(a6xx_gpu->sqe_bo, gpu->aspace);
1254 		drm_gem_object_put(a6xx_gpu->sqe_bo);
1255 	}
1256 
1257 	if (a6xx_gpu->shadow_bo) {
1258 		msm_gem_unpin_iova(a6xx_gpu->shadow_bo, gpu->aspace);
1259 		drm_gem_object_put(a6xx_gpu->shadow_bo);
1260 	}
1261 
1262 	a6xx_llc_slices_destroy(a6xx_gpu);
1263 
1264 	a6xx_gmu_remove(a6xx_gpu);
1265 
1266 	adreno_gpu_cleanup(adreno_gpu);
1267 
1268 	if (a6xx_gpu->opp_table)
1269 		dev_pm_opp_put_supported_hw(a6xx_gpu->opp_table);
1270 
1271 	kfree(a6xx_gpu);
1272 }
1273 
1274 static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
1275 {
1276 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1277 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1278 	u64 busy_cycles, busy_time;
1279 
1280 
1281 	/* Only read the gpu busy if the hardware is already active */
1282 	if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
1283 		return 0;
1284 
1285 	busy_cycles = gmu_read64(&a6xx_gpu->gmu,
1286 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
1287 			REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
1288 
1289 	busy_time = (busy_cycles - gpu->devfreq.busy_cycles) * 10;
1290 	do_div(busy_time, 192);
1291 
1292 	gpu->devfreq.busy_cycles = busy_cycles;
1293 
1294 	pm_runtime_put(a6xx_gpu->gmu.dev);
1295 
1296 	if (WARN_ON(busy_time > ~0LU))
1297 		return ~0LU;
1298 
1299 	return (unsigned long)busy_time;
1300 }
1301 
1302 static struct msm_gem_address_space *
1303 a6xx_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev)
1304 {
1305 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1306 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1307 	struct iommu_domain *iommu;
1308 	struct msm_mmu *mmu;
1309 	struct msm_gem_address_space *aspace;
1310 	u64 start, size;
1311 
1312 	iommu = iommu_domain_alloc(&platform_bus_type);
1313 	if (!iommu)
1314 		return NULL;
1315 
1316 	/*
1317 	 * This allows GPU to set the bus attributes required to use system
1318 	 * cache on behalf of the iommu page table walker.
1319 	 */
1320 	if (!IS_ERR_OR_NULL(a6xx_gpu->htw_llc_slice))
1321 		adreno_set_llc_attributes(iommu);
1322 
1323 	mmu = msm_iommu_new(&pdev->dev, iommu);
1324 	if (IS_ERR(mmu)) {
1325 		iommu_domain_free(iommu);
1326 		return ERR_CAST(mmu);
1327 	}
1328 
1329 	/*
1330 	 * Use the aperture start or SZ_16M, whichever is greater. This will
1331 	 * ensure that we align with the allocated pagetable range while still
1332 	 * allowing room in the lower 32 bits for GMEM and whatnot
1333 	 */
1334 	start = max_t(u64, SZ_16M, iommu->geometry.aperture_start);
1335 	size = iommu->geometry.aperture_end - start + 1;
1336 
1337 	aspace = msm_gem_address_space_create(mmu, "gpu",
1338 		start & GENMASK_ULL(48, 0), size);
1339 
1340 	if (IS_ERR(aspace) && !IS_ERR(mmu))
1341 		mmu->funcs->destroy(mmu);
1342 
1343 	return aspace;
1344 }
1345 
1346 static struct msm_gem_address_space *
1347 a6xx_create_private_address_space(struct msm_gpu *gpu)
1348 {
1349 	struct msm_mmu *mmu;
1350 
1351 	mmu = msm_iommu_pagetable_create(gpu->aspace->mmu);
1352 
1353 	if (IS_ERR(mmu))
1354 		return ERR_CAST(mmu);
1355 
1356 	return msm_gem_address_space_create(mmu,
1357 		"gpu", 0x100000000ULL, 0x1ffffffffULL);
1358 }
1359 
1360 static uint32_t a6xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
1361 {
1362 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1363 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
1364 
1365 	if (adreno_gpu->base.hw_apriv || a6xx_gpu->has_whereami)
1366 		return a6xx_gpu->shadow[ring->id];
1367 
1368 	return ring->memptrs->rptr = gpu_read(gpu, REG_A6XX_CP_RB_RPTR);
1369 }
1370 
1371 static u32 a618_get_speed_bin(u32 fuse)
1372 {
1373 	if (fuse == 0)
1374 		return 0;
1375 	else if (fuse == 169)
1376 		return 1;
1377 	else if (fuse == 174)
1378 		return 2;
1379 
1380 	return UINT_MAX;
1381 }
1382 
1383 static u32 fuse_to_supp_hw(struct device *dev, u32 revn, u32 fuse)
1384 {
1385 	u32 val = UINT_MAX;
1386 
1387 	if (revn == 618)
1388 		val = a618_get_speed_bin(fuse);
1389 
1390 	if (val == UINT_MAX) {
1391 		DRM_DEV_ERROR(dev,
1392 			"missing support for speed-bin: %u. Some OPPs may not be supported by hardware",
1393 			fuse);
1394 		return UINT_MAX;
1395 	}
1396 
1397 	return (1 << val);
1398 }
1399 
1400 static int a6xx_set_supported_hw(struct device *dev, struct a6xx_gpu *a6xx_gpu,
1401 		u32 revn)
1402 {
1403 	struct opp_table *opp_table;
1404 	u32 supp_hw = UINT_MAX;
1405 	u16 speedbin;
1406 	int ret;
1407 
1408 	ret = nvmem_cell_read_u16(dev, "speed_bin", &speedbin);
1409 	/*
1410 	 * -ENOENT means that the platform doesn't support speedbin which is
1411 	 * fine
1412 	 */
1413 	if (ret == -ENOENT) {
1414 		return 0;
1415 	} else if (ret) {
1416 		DRM_DEV_ERROR(dev,
1417 			      "failed to read speed-bin (%d). Some OPPs may not be supported by hardware",
1418 			      ret);
1419 		goto done;
1420 	}
1421 	speedbin = le16_to_cpu(speedbin);
1422 
1423 	supp_hw = fuse_to_supp_hw(dev, revn, speedbin);
1424 
1425 done:
1426 	opp_table = dev_pm_opp_set_supported_hw(dev, &supp_hw, 1);
1427 	if (IS_ERR(opp_table))
1428 		return PTR_ERR(opp_table);
1429 
1430 	a6xx_gpu->opp_table = opp_table;
1431 	return 0;
1432 }
1433 
1434 static const struct adreno_gpu_funcs funcs = {
1435 	.base = {
1436 		.get_param = adreno_get_param,
1437 		.hw_init = a6xx_hw_init,
1438 		.pm_suspend = a6xx_pm_suspend,
1439 		.pm_resume = a6xx_pm_resume,
1440 		.recover = a6xx_recover,
1441 		.submit = a6xx_submit,
1442 		.active_ring = a6xx_active_ring,
1443 		.irq = a6xx_irq,
1444 		.destroy = a6xx_destroy,
1445 #if defined(CONFIG_DRM_MSM_GPU_STATE)
1446 		.show = a6xx_show,
1447 #endif
1448 		.gpu_busy = a6xx_gpu_busy,
1449 		.gpu_get_freq = a6xx_gmu_get_freq,
1450 		.gpu_set_freq = a6xx_gmu_set_freq,
1451 #if defined(CONFIG_DRM_MSM_GPU_STATE)
1452 		.gpu_state_get = a6xx_gpu_state_get,
1453 		.gpu_state_put = a6xx_gpu_state_put,
1454 #endif
1455 		.create_address_space = a6xx_create_address_space,
1456 		.create_private_address_space = a6xx_create_private_address_space,
1457 		.get_rptr = a6xx_get_rptr,
1458 	},
1459 	.get_timestamp = a6xx_get_timestamp,
1460 };
1461 
1462 struct msm_gpu *a6xx_gpu_init(struct drm_device *dev)
1463 {
1464 	struct msm_drm_private *priv = dev->dev_private;
1465 	struct platform_device *pdev = priv->gpu_pdev;
1466 	struct adreno_platform_config *config = pdev->dev.platform_data;
1467 	const struct adreno_info *info;
1468 	struct device_node *node;
1469 	struct a6xx_gpu *a6xx_gpu;
1470 	struct adreno_gpu *adreno_gpu;
1471 	struct msm_gpu *gpu;
1472 	int ret;
1473 
1474 	a6xx_gpu = kzalloc(sizeof(*a6xx_gpu), GFP_KERNEL);
1475 	if (!a6xx_gpu)
1476 		return ERR_PTR(-ENOMEM);
1477 
1478 	adreno_gpu = &a6xx_gpu->base;
1479 	gpu = &adreno_gpu->base;
1480 
1481 	adreno_gpu->registers = NULL;
1482 
1483 	/*
1484 	 * We need to know the platform type before calling into adreno_gpu_init
1485 	 * so that the hw_apriv flag can be correctly set. Snoop into the info
1486 	 * and grab the revision number
1487 	 */
1488 	info = adreno_info(config->rev);
1489 
1490 	if (info && info->revn == 650)
1491 		adreno_gpu->base.hw_apriv = true;
1492 
1493 	a6xx_llc_slices_init(pdev, a6xx_gpu);
1494 
1495 	ret = a6xx_set_supported_hw(&pdev->dev, a6xx_gpu, info->revn);
1496 	if (ret) {
1497 		a6xx_destroy(&(a6xx_gpu->base.base));
1498 		return ERR_PTR(ret);
1499 	}
1500 
1501 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
1502 	if (ret) {
1503 		a6xx_destroy(&(a6xx_gpu->base.base));
1504 		return ERR_PTR(ret);
1505 	}
1506 
1507 	/* Check if there is a GMU phandle and set it up */
1508 	node = of_parse_phandle(pdev->dev.of_node, "qcom,gmu", 0);
1509 
1510 	/* FIXME: How do we gracefully handle this? */
1511 	BUG_ON(!node);
1512 
1513 	ret = a6xx_gmu_init(a6xx_gpu, node);
1514 	if (ret) {
1515 		a6xx_destroy(&(a6xx_gpu->base.base));
1516 		return ERR_PTR(ret);
1517 	}
1518 
1519 	if (gpu->aspace)
1520 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu,
1521 				a6xx_fault_handler);
1522 
1523 	return gpu;
1524 }
1525