xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a3xx_gpu.c (revision e7f127b2)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  *
6  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
7  */
8 
9 #include "a3xx_gpu.h"
10 
11 #define A3XX_INT0_MASK \
12 	(A3XX_INT0_RBBM_AHB_ERROR |        \
13 	 A3XX_INT0_RBBM_ATB_BUS_OVERFLOW | \
14 	 A3XX_INT0_CP_T0_PACKET_IN_IB |    \
15 	 A3XX_INT0_CP_OPCODE_ERROR |       \
16 	 A3XX_INT0_CP_RESERVED_BIT_ERROR | \
17 	 A3XX_INT0_CP_HW_FAULT |           \
18 	 A3XX_INT0_CP_IB1_INT |            \
19 	 A3XX_INT0_CP_IB2_INT |            \
20 	 A3XX_INT0_CP_RB_INT |             \
21 	 A3XX_INT0_CP_REG_PROTECT_FAULT |  \
22 	 A3XX_INT0_CP_AHB_ERROR_HALT |     \
23 	 A3XX_INT0_CACHE_FLUSH_TS |        \
24 	 A3XX_INT0_UCHE_OOB_ACCESS)
25 
26 extern bool hang_debug;
27 
28 static void a3xx_dump(struct msm_gpu *gpu);
29 static bool a3xx_idle(struct msm_gpu *gpu);
30 
31 static void a3xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
32 {
33 	struct msm_ringbuffer *ring = submit->ring;
34 	unsigned int i;
35 
36 	for (i = 0; i < submit->nr_cmds; i++) {
37 		switch (submit->cmd[i].type) {
38 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
39 			/* ignore IB-targets */
40 			break;
41 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
42 			/* ignore if there has not been a ctx switch: */
43 			if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno)
44 				break;
45 			fallthrough;
46 		case MSM_SUBMIT_CMD_BUF:
47 			OUT_PKT3(ring, CP_INDIRECT_BUFFER_PFD, 2);
48 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
49 			OUT_RING(ring, submit->cmd[i].size);
50 			OUT_PKT2(ring);
51 			break;
52 		}
53 	}
54 
55 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
56 	OUT_RING(ring, submit->seqno);
57 
58 	/* Flush HLSQ lazy updates to make sure there is nothing
59 	 * pending for indirect loads after the timestamp has
60 	 * passed:
61 	 */
62 	OUT_PKT3(ring, CP_EVENT_WRITE, 1);
63 	OUT_RING(ring, HLSQ_FLUSH);
64 
65 	/* wait for idle before cache flush/interrupt */
66 	OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
67 	OUT_RING(ring, 0x00000000);
68 
69 	/* BIT(31) of CACHE_FLUSH_TS triggers CACHE_FLUSH_TS IRQ from GPU */
70 	OUT_PKT3(ring, CP_EVENT_WRITE, 3);
71 	OUT_RING(ring, CACHE_FLUSH_TS | BIT(31));
72 	OUT_RING(ring, rbmemptr(ring, fence));
73 	OUT_RING(ring, submit->seqno);
74 
75 #if 0
76 	/* Dummy set-constant to trigger context rollover */
77 	OUT_PKT3(ring, CP_SET_CONSTANT, 2);
78 	OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
79 	OUT_RING(ring, 0x00000000);
80 #endif
81 
82 	adreno_flush(gpu, ring, REG_AXXX_CP_RB_WPTR);
83 }
84 
85 static bool a3xx_me_init(struct msm_gpu *gpu)
86 {
87 	struct msm_ringbuffer *ring = gpu->rb[0];
88 
89 	OUT_PKT3(ring, CP_ME_INIT, 17);
90 	OUT_RING(ring, 0x000003f7);
91 	OUT_RING(ring, 0x00000000);
92 	OUT_RING(ring, 0x00000000);
93 	OUT_RING(ring, 0x00000000);
94 	OUT_RING(ring, 0x00000080);
95 	OUT_RING(ring, 0x00000100);
96 	OUT_RING(ring, 0x00000180);
97 	OUT_RING(ring, 0x00006600);
98 	OUT_RING(ring, 0x00000150);
99 	OUT_RING(ring, 0x0000014e);
100 	OUT_RING(ring, 0x00000154);
101 	OUT_RING(ring, 0x00000001);
102 	OUT_RING(ring, 0x00000000);
103 	OUT_RING(ring, 0x00000000);
104 	OUT_RING(ring, 0x00000000);
105 	OUT_RING(ring, 0x00000000);
106 	OUT_RING(ring, 0x00000000);
107 
108 	adreno_flush(gpu, ring, REG_AXXX_CP_RB_WPTR);
109 	return a3xx_idle(gpu);
110 }
111 
112 static int a3xx_hw_init(struct msm_gpu *gpu)
113 {
114 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
115 	struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
116 	uint32_t *ptr, len;
117 	int i, ret;
118 
119 	DBG("%s", gpu->name);
120 
121 	if (adreno_is_a305(adreno_gpu)) {
122 		/* Set up 16 deep read/write request queues: */
123 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
124 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
125 		gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
126 		gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
127 		gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
128 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
129 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
130 		/* Enable WR-REQ: */
131 		gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
132 		/* Set up round robin arbitration between both AXI ports: */
133 		gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
134 		/* Set up AOOO: */
135 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
136 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
137 	} else if (adreno_is_a306(adreno_gpu)) {
138 		gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
139 		gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x0000000a);
140 		gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x0000000a);
141 	} else if (adreno_is_a320(adreno_gpu)) {
142 		/* Set up 16 deep read/write request queues: */
143 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
144 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
145 		gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
146 		gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
147 		gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
148 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
149 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
150 		/* Enable WR-REQ: */
151 		gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
152 		/* Set up round robin arbitration between both AXI ports: */
153 		gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
154 		/* Set up AOOO: */
155 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
156 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
157 		/* Enable 1K sort: */
158 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x000000ff);
159 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
160 
161 	} else if (adreno_is_a330v2(adreno_gpu)) {
162 		/*
163 		 * Most of the VBIF registers on 8974v2 have the correct
164 		 * values at power on, so we won't modify those if we don't
165 		 * need to
166 		 */
167 		/* Enable 1k sort: */
168 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
169 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
170 		/* Enable WR-REQ: */
171 		gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
172 		gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
173 		/* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
174 		gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
175 
176 	} else if (adreno_is_a330(adreno_gpu)) {
177 		/* Set up 16 deep read/write request queues: */
178 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x18181818);
179 		gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x18181818);
180 		gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x18181818);
181 		gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x18181818);
182 		gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
183 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x18181818);
184 		gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x18181818);
185 		/* Enable WR-REQ: */
186 		gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
187 		/* Set up round robin arbitration between both AXI ports: */
188 		gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
189 		/* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
190 		gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0001);
191 		/* Set up AOOO: */
192 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003f);
193 		gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003f003f);
194 		/* Enable 1K sort: */
195 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
196 		gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
197 		/* Disable VBIF clock gating. This is to enable AXI running
198 		 * higher frequency than GPU:
199 		 */
200 		gpu_write(gpu, REG_A3XX_VBIF_CLKON, 0x00000001);
201 
202 	} else {
203 		BUG();
204 	}
205 
206 	/* Make all blocks contribute to the GPU BUSY perf counter: */
207 	gpu_write(gpu, REG_A3XX_RBBM_GPU_BUSY_MASKED, 0xffffffff);
208 
209 	/* Tune the hystersis counters for SP and CP idle detection: */
210 	gpu_write(gpu, REG_A3XX_RBBM_SP_HYST_CNT, 0x10);
211 	gpu_write(gpu, REG_A3XX_RBBM_WAIT_IDLE_CLOCKS_CTL, 0x10);
212 
213 	/* Enable the RBBM error reporting bits.  This lets us get
214 	 * useful information on failure:
215 	 */
216 	gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL0, 0x00000001);
217 
218 	/* Enable AHB error reporting: */
219 	gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL1, 0xa6ffffff);
220 
221 	/* Turn on the power counters: */
222 	gpu_write(gpu, REG_A3XX_RBBM_RBBM_CTL, 0x00030000);
223 
224 	/* Turn on hang detection - this spews a lot of useful information
225 	 * into the RBBM registers on a hang:
226 	 */
227 	gpu_write(gpu, REG_A3XX_RBBM_INTERFACE_HANG_INT_CTL, 0x00010fff);
228 
229 	/* Enable 64-byte cacheline size. HW Default is 32-byte (0x000000E0): */
230 	gpu_write(gpu, REG_A3XX_UCHE_CACHE_MODE_CONTROL_REG, 0x00000001);
231 
232 	/* Enable Clock gating: */
233 	if (adreno_is_a306(adreno_gpu))
234 		gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
235 	else if (adreno_is_a320(adreno_gpu))
236 		gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbfffffff);
237 	else if (adreno_is_a330v2(adreno_gpu))
238 		gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
239 	else if (adreno_is_a330(adreno_gpu))
240 		gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbffcffff);
241 
242 	if (adreno_is_a330v2(adreno_gpu))
243 		gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x05515455);
244 	else if (adreno_is_a330(adreno_gpu))
245 		gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x00000000);
246 
247 	/* Set the OCMEM base address for A330, etc */
248 	if (a3xx_gpu->ocmem.hdl) {
249 		gpu_write(gpu, REG_A3XX_RB_GMEM_BASE_ADDR,
250 			(unsigned int)(a3xx_gpu->ocmem.base >> 14));
251 	}
252 
253 	/* Turn on performance counters: */
254 	gpu_write(gpu, REG_A3XX_RBBM_PERFCTR_CTL, 0x01);
255 
256 	/* Enable the perfcntrs that we use.. */
257 	for (i = 0; i < gpu->num_perfcntrs; i++) {
258 		const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i];
259 		gpu_write(gpu, perfcntr->select_reg, perfcntr->select_val);
260 	}
261 
262 	gpu_write(gpu, REG_A3XX_RBBM_INT_0_MASK, A3XX_INT0_MASK);
263 
264 	ret = adreno_hw_init(gpu);
265 	if (ret)
266 		return ret;
267 
268 	/*
269 	 * Use the default ringbuffer size and block size but disable the RPTR
270 	 * shadow
271 	 */
272 	gpu_write(gpu, REG_AXXX_CP_RB_CNTL,
273 		MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE);
274 
275 	/* Set the ringbuffer address */
276 	gpu_write(gpu, REG_AXXX_CP_RB_BASE, lower_32_bits(gpu->rb[0]->iova));
277 
278 	/* setup access protection: */
279 	gpu_write(gpu, REG_A3XX_CP_PROTECT_CTRL, 0x00000007);
280 
281 	/* RBBM registers */
282 	gpu_write(gpu, REG_A3XX_CP_PROTECT(0), 0x63000040);
283 	gpu_write(gpu, REG_A3XX_CP_PROTECT(1), 0x62000080);
284 	gpu_write(gpu, REG_A3XX_CP_PROTECT(2), 0x600000cc);
285 	gpu_write(gpu, REG_A3XX_CP_PROTECT(3), 0x60000108);
286 	gpu_write(gpu, REG_A3XX_CP_PROTECT(4), 0x64000140);
287 	gpu_write(gpu, REG_A3XX_CP_PROTECT(5), 0x66000400);
288 
289 	/* CP registers */
290 	gpu_write(gpu, REG_A3XX_CP_PROTECT(6), 0x65000700);
291 	gpu_write(gpu, REG_A3XX_CP_PROTECT(7), 0x610007d8);
292 	gpu_write(gpu, REG_A3XX_CP_PROTECT(8), 0x620007e0);
293 	gpu_write(gpu, REG_A3XX_CP_PROTECT(9), 0x61001178);
294 	gpu_write(gpu, REG_A3XX_CP_PROTECT(10), 0x64001180);
295 
296 	/* RB registers */
297 	gpu_write(gpu, REG_A3XX_CP_PROTECT(11), 0x60003300);
298 
299 	/* VBIF registers */
300 	gpu_write(gpu, REG_A3XX_CP_PROTECT(12), 0x6b00c000);
301 
302 	/* NOTE: PM4/micro-engine firmware registers look to be the same
303 	 * for a2xx and a3xx.. we could possibly push that part down to
304 	 * adreno_gpu base class.  Or push both PM4 and PFP but
305 	 * parameterize the pfp ucode addr/data registers..
306 	 */
307 
308 	/* Load PM4: */
309 	ptr = (uint32_t *)(adreno_gpu->fw[ADRENO_FW_PM4]->data);
310 	len = adreno_gpu->fw[ADRENO_FW_PM4]->size / 4;
311 	DBG("loading PM4 ucode version: %x", ptr[1]);
312 
313 	gpu_write(gpu, REG_AXXX_CP_DEBUG,
314 			AXXX_CP_DEBUG_DYNAMIC_CLK_DISABLE |
315 			AXXX_CP_DEBUG_MIU_128BIT_WRITE_ENABLE);
316 	gpu_write(gpu, REG_AXXX_CP_ME_RAM_WADDR, 0);
317 	for (i = 1; i < len; i++)
318 		gpu_write(gpu, REG_AXXX_CP_ME_RAM_DATA, ptr[i]);
319 
320 	/* Load PFP: */
321 	ptr = (uint32_t *)(adreno_gpu->fw[ADRENO_FW_PFP]->data);
322 	len = adreno_gpu->fw[ADRENO_FW_PFP]->size / 4;
323 	DBG("loading PFP ucode version: %x", ptr[5]);
324 
325 	gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_ADDR, 0);
326 	for (i = 1; i < len; i++)
327 		gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_DATA, ptr[i]);
328 
329 	/* CP ROQ queue sizes (bytes) - RB:16, ST:16, IB1:32, IB2:64 */
330 	if (adreno_is_a305(adreno_gpu) || adreno_is_a306(adreno_gpu) ||
331 			adreno_is_a320(adreno_gpu)) {
332 		gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS,
333 				AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB1_START(2) |
334 				AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB2_START(6) |
335 				AXXX_CP_QUEUE_THRESHOLDS_CSQ_ST_START(14));
336 	} else if (adreno_is_a330(adreno_gpu)) {
337 		/* NOTE: this (value take from downstream android driver)
338 		 * includes some bits outside of the known bitfields.  But
339 		 * A330 has this "MERCIU queue" thing too, which might
340 		 * explain a new bitfield or reshuffling:
341 		 */
342 		gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS, 0x003e2008);
343 	}
344 
345 	/* clear ME_HALT to start micro engine */
346 	gpu_write(gpu, REG_AXXX_CP_ME_CNTL, 0);
347 
348 	return a3xx_me_init(gpu) ? 0 : -EINVAL;
349 }
350 
351 static void a3xx_recover(struct msm_gpu *gpu)
352 {
353 	int i;
354 
355 	adreno_dump_info(gpu);
356 
357 	for (i = 0; i < 8; i++) {
358 		printk("CP_SCRATCH_REG%d: %u\n", i,
359 			gpu_read(gpu, REG_AXXX_CP_SCRATCH_REG0 + i));
360 	}
361 
362 	/* dump registers before resetting gpu, if enabled: */
363 	if (hang_debug)
364 		a3xx_dump(gpu);
365 
366 	gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 1);
367 	gpu_read(gpu, REG_A3XX_RBBM_SW_RESET_CMD);
368 	gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 0);
369 	adreno_recover(gpu);
370 }
371 
372 static void a3xx_destroy(struct msm_gpu *gpu)
373 {
374 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
375 	struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
376 
377 	DBG("%s", gpu->name);
378 
379 	adreno_gpu_cleanup(adreno_gpu);
380 
381 	adreno_gpu_ocmem_cleanup(&a3xx_gpu->ocmem);
382 
383 	kfree(a3xx_gpu);
384 }
385 
386 static bool a3xx_idle(struct msm_gpu *gpu)
387 {
388 	/* wait for ringbuffer to drain: */
389 	if (!adreno_idle(gpu, gpu->rb[0]))
390 		return false;
391 
392 	/* then wait for GPU to finish: */
393 	if (spin_until(!(gpu_read(gpu, REG_A3XX_RBBM_STATUS) &
394 			A3XX_RBBM_STATUS_GPU_BUSY))) {
395 		DRM_ERROR("%s: timeout waiting for GPU to idle!\n", gpu->name);
396 
397 		/* TODO maybe we need to reset GPU here to recover from hang? */
398 		return false;
399 	}
400 
401 	return true;
402 }
403 
404 static irqreturn_t a3xx_irq(struct msm_gpu *gpu)
405 {
406 	uint32_t status;
407 
408 	status = gpu_read(gpu, REG_A3XX_RBBM_INT_0_STATUS);
409 	DBG("%s: %08x", gpu->name, status);
410 
411 	// TODO
412 
413 	gpu_write(gpu, REG_A3XX_RBBM_INT_CLEAR_CMD, status);
414 
415 	msm_gpu_retire(gpu);
416 
417 	return IRQ_HANDLED;
418 }
419 
420 static const unsigned int a3xx_registers[] = {
421 	0x0000, 0x0002, 0x0010, 0x0012, 0x0018, 0x0018, 0x0020, 0x0027,
422 	0x0029, 0x002b, 0x002e, 0x0033, 0x0040, 0x0042, 0x0050, 0x005c,
423 	0x0060, 0x006c, 0x0080, 0x0082, 0x0084, 0x0088, 0x0090, 0x00e5,
424 	0x00ea, 0x00ed, 0x0100, 0x0100, 0x0110, 0x0123, 0x01c0, 0x01c1,
425 	0x01c3, 0x01c5, 0x01c7, 0x01c7, 0x01d5, 0x01d9, 0x01dc, 0x01dd,
426 	0x01ea, 0x01ea, 0x01ee, 0x01f1, 0x01f5, 0x01f5, 0x01fc, 0x01ff,
427 	0x0440, 0x0440, 0x0443, 0x0443, 0x0445, 0x0445, 0x044d, 0x044f,
428 	0x0452, 0x0452, 0x0454, 0x046f, 0x047c, 0x047c, 0x047f, 0x047f,
429 	0x0578, 0x057f, 0x0600, 0x0602, 0x0605, 0x0607, 0x060a, 0x060e,
430 	0x0612, 0x0614, 0x0c01, 0x0c02, 0x0c06, 0x0c1d, 0x0c3d, 0x0c3f,
431 	0x0c48, 0x0c4b, 0x0c80, 0x0c80, 0x0c88, 0x0c8b, 0x0ca0, 0x0cb7,
432 	0x0cc0, 0x0cc1, 0x0cc6, 0x0cc7, 0x0ce4, 0x0ce5, 0x0e00, 0x0e05,
433 	0x0e0c, 0x0e0c, 0x0e22, 0x0e23, 0x0e41, 0x0e45, 0x0e64, 0x0e65,
434 	0x0e80, 0x0e82, 0x0e84, 0x0e89, 0x0ea0, 0x0ea1, 0x0ea4, 0x0ea7,
435 	0x0ec4, 0x0ecb, 0x0ee0, 0x0ee0, 0x0f00, 0x0f01, 0x0f03, 0x0f09,
436 	0x2040, 0x2040, 0x2044, 0x2044, 0x2048, 0x204d, 0x2068, 0x2069,
437 	0x206c, 0x206d, 0x2070, 0x2070, 0x2072, 0x2072, 0x2074, 0x2075,
438 	0x2079, 0x207a, 0x20c0, 0x20d3, 0x20e4, 0x20ef, 0x2100, 0x2109,
439 	0x210c, 0x210c, 0x210e, 0x210e, 0x2110, 0x2111, 0x2114, 0x2115,
440 	0x21e4, 0x21e4, 0x21ea, 0x21ea, 0x21ec, 0x21ed, 0x21f0, 0x21f0,
441 	0x2200, 0x2212, 0x2214, 0x2217, 0x221a, 0x221a, 0x2240, 0x227e,
442 	0x2280, 0x228b, 0x22c0, 0x22c0, 0x22c4, 0x22ce, 0x22d0, 0x22d8,
443 	0x22df, 0x22e6, 0x22e8, 0x22e9, 0x22ec, 0x22ec, 0x22f0, 0x22f7,
444 	0x22ff, 0x22ff, 0x2340, 0x2343, 0x2440, 0x2440, 0x2444, 0x2444,
445 	0x2448, 0x244d, 0x2468, 0x2469, 0x246c, 0x246d, 0x2470, 0x2470,
446 	0x2472, 0x2472, 0x2474, 0x2475, 0x2479, 0x247a, 0x24c0, 0x24d3,
447 	0x24e4, 0x24ef, 0x2500, 0x2509, 0x250c, 0x250c, 0x250e, 0x250e,
448 	0x2510, 0x2511, 0x2514, 0x2515, 0x25e4, 0x25e4, 0x25ea, 0x25ea,
449 	0x25ec, 0x25ed, 0x25f0, 0x25f0, 0x2600, 0x2612, 0x2614, 0x2617,
450 	0x261a, 0x261a, 0x2640, 0x267e, 0x2680, 0x268b, 0x26c0, 0x26c0,
451 	0x26c4, 0x26ce, 0x26d0, 0x26d8, 0x26df, 0x26e6, 0x26e8, 0x26e9,
452 	0x26ec, 0x26ec, 0x26f0, 0x26f7, 0x26ff, 0x26ff, 0x2740, 0x2743,
453 	0x300c, 0x300e, 0x301c, 0x301d, 0x302a, 0x302a, 0x302c, 0x302d,
454 	0x3030, 0x3031, 0x3034, 0x3036, 0x303c, 0x303c, 0x305e, 0x305f,
455 	~0   /* sentinel */
456 };
457 
458 /* would be nice to not have to duplicate the _show() stuff with printk(): */
459 static void a3xx_dump(struct msm_gpu *gpu)
460 {
461 	printk("status:   %08x\n",
462 			gpu_read(gpu, REG_A3XX_RBBM_STATUS));
463 	adreno_dump(gpu);
464 }
465 
466 static struct msm_gpu_state *a3xx_gpu_state_get(struct msm_gpu *gpu)
467 {
468 	struct msm_gpu_state *state = kzalloc(sizeof(*state), GFP_KERNEL);
469 
470 	if (!state)
471 		return ERR_PTR(-ENOMEM);
472 
473 	adreno_gpu_state_get(gpu, state);
474 
475 	state->rbbm_status = gpu_read(gpu, REG_A3XX_RBBM_STATUS);
476 
477 	return state;
478 }
479 
480 static u32 a3xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
481 {
482 	ring->memptrs->rptr = gpu_read(gpu, REG_AXXX_CP_RB_RPTR);
483 	return ring->memptrs->rptr;
484 }
485 
486 static const struct adreno_gpu_funcs funcs = {
487 	.base = {
488 		.get_param = adreno_get_param,
489 		.hw_init = a3xx_hw_init,
490 		.pm_suspend = msm_gpu_pm_suspend,
491 		.pm_resume = msm_gpu_pm_resume,
492 		.recover = a3xx_recover,
493 		.submit = a3xx_submit,
494 		.active_ring = adreno_active_ring,
495 		.irq = a3xx_irq,
496 		.destroy = a3xx_destroy,
497 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
498 		.show = adreno_show,
499 #endif
500 		.gpu_state_get = a3xx_gpu_state_get,
501 		.gpu_state_put = adreno_gpu_state_put,
502 		.create_address_space = adreno_iommu_create_address_space,
503 		.get_rptr = a3xx_get_rptr,
504 	},
505 };
506 
507 static const struct msm_gpu_perfcntr perfcntrs[] = {
508 	{ REG_A3XX_SP_PERFCOUNTER6_SELECT, REG_A3XX_RBBM_PERFCTR_SP_6_LO,
509 			SP_ALU_ACTIVE_CYCLES, "ALUACTIVE" },
510 	{ REG_A3XX_SP_PERFCOUNTER7_SELECT, REG_A3XX_RBBM_PERFCTR_SP_7_LO,
511 			SP_FS_FULL_ALU_INSTRUCTIONS, "ALUFULL" },
512 };
513 
514 struct msm_gpu *a3xx_gpu_init(struct drm_device *dev)
515 {
516 	struct a3xx_gpu *a3xx_gpu = NULL;
517 	struct adreno_gpu *adreno_gpu;
518 	struct msm_gpu *gpu;
519 	struct msm_drm_private *priv = dev->dev_private;
520 	struct platform_device *pdev = priv->gpu_pdev;
521 	struct icc_path *ocmem_icc_path;
522 	struct icc_path *icc_path;
523 	int ret;
524 
525 	if (!pdev) {
526 		DRM_DEV_ERROR(dev->dev, "no a3xx device\n");
527 		ret = -ENXIO;
528 		goto fail;
529 	}
530 
531 	a3xx_gpu = kzalloc(sizeof(*a3xx_gpu), GFP_KERNEL);
532 	if (!a3xx_gpu) {
533 		ret = -ENOMEM;
534 		goto fail;
535 	}
536 
537 	adreno_gpu = &a3xx_gpu->base;
538 	gpu = &adreno_gpu->base;
539 
540 	gpu->perfcntrs = perfcntrs;
541 	gpu->num_perfcntrs = ARRAY_SIZE(perfcntrs);
542 
543 	adreno_gpu->registers = a3xx_registers;
544 
545 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 1);
546 	if (ret)
547 		goto fail;
548 
549 	/* if needed, allocate gmem: */
550 	if (adreno_is_a330(adreno_gpu)) {
551 		ret = adreno_gpu_ocmem_init(&adreno_gpu->base.pdev->dev,
552 					    adreno_gpu, &a3xx_gpu->ocmem);
553 		if (ret)
554 			goto fail;
555 	}
556 
557 	if (!gpu->aspace) {
558 		/* TODO we think it is possible to configure the GPU to
559 		 * restrict access to VRAM carveout.  But the required
560 		 * registers are unknown.  For now just bail out and
561 		 * limp along with just modesetting.  If it turns out
562 		 * to not be possible to restrict access, then we must
563 		 * implement a cmdstream validator.
564 		 */
565 		DRM_DEV_ERROR(dev->dev, "No memory protection without IOMMU\n");
566 		if (!allow_vram_carveout) {
567 			ret = -ENXIO;
568 			goto fail;
569 		}
570 	}
571 
572 	icc_path = devm_of_icc_get(&pdev->dev, "gfx-mem");
573 	if (IS_ERR(icc_path)) {
574 		ret = PTR_ERR(icc_path);
575 		goto fail;
576 	}
577 
578 	ocmem_icc_path = devm_of_icc_get(&pdev->dev, "ocmem");
579 	if (IS_ERR(ocmem_icc_path)) {
580 		ret = PTR_ERR(ocmem_icc_path);
581 		/* allow -ENODATA, ocmem icc is optional */
582 		if (ret != -ENODATA)
583 			goto fail;
584 		ocmem_icc_path = NULL;
585 	}
586 
587 
588 	/*
589 	 * Set the ICC path to maximum speed for now by multiplying the fastest
590 	 * frequency by the bus width (8). We'll want to scale this later on to
591 	 * improve battery life.
592 	 */
593 	icc_set_bw(icc_path, 0, Bps_to_icc(gpu->fast_rate) * 8);
594 	icc_set_bw(ocmem_icc_path, 0, Bps_to_icc(gpu->fast_rate) * 8);
595 
596 	return gpu;
597 
598 fail:
599 	if (a3xx_gpu)
600 		a3xx_destroy(&a3xx_gpu->base.base);
601 
602 	return ERR_PTR(ret);
603 }
604