xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a5xx_gpu.c (revision bbecb07f)
1 /* Copyright (c) 2016-2017 The Linux Foundation. All rights reserved.
2  *
3  * This program is free software; you can redistribute it and/or modify
4  * it under the terms of the GNU General Public License version 2 and
5  * only version 2 as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful,
8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10  * GNU General Public License for more details.
11  *
12  */
13 
14 #include <linux/types.h>
15 #include <linux/cpumask.h>
16 #include <linux/qcom_scm.h>
17 #include <linux/dma-mapping.h>
18 #include <linux/of_address.h>
19 #include <linux/soc/qcom/mdt_loader.h>
20 #include "msm_gem.h"
21 #include "msm_mmu.h"
22 #include "a5xx_gpu.h"
23 
24 extern bool hang_debug;
25 static void a5xx_dump(struct msm_gpu *gpu);
26 
27 #define GPU_PAS_ID 13
28 
29 static int zap_shader_load_mdt(struct msm_gpu *gpu, const char *fwname)
30 {
31 	struct device *dev = &gpu->pdev->dev;
32 	const struct firmware *fw;
33 	struct device_node *np;
34 	struct resource r;
35 	phys_addr_t mem_phys;
36 	ssize_t mem_size;
37 	void *mem_region = NULL;
38 	int ret;
39 
40 	if (!IS_ENABLED(CONFIG_ARCH_QCOM))
41 		return -EINVAL;
42 
43 	np = of_get_child_by_name(dev->of_node, "zap-shader");
44 	if (!np)
45 		return -ENODEV;
46 
47 	np = of_parse_phandle(np, "memory-region", 0);
48 	if (!np)
49 		return -EINVAL;
50 
51 	ret = of_address_to_resource(np, 0, &r);
52 	if (ret)
53 		return ret;
54 
55 	mem_phys = r.start;
56 	mem_size = resource_size(&r);
57 
58 	/* Request the MDT file for the firmware */
59 	fw = adreno_request_fw(to_adreno_gpu(gpu), fwname);
60 	if (IS_ERR(fw)) {
61 		DRM_DEV_ERROR(dev, "Unable to load %s\n", fwname);
62 		return PTR_ERR(fw);
63 	}
64 
65 	/* Figure out how much memory we need */
66 	mem_size = qcom_mdt_get_size(fw);
67 	if (mem_size < 0) {
68 		ret = mem_size;
69 		goto out;
70 	}
71 
72 	/* Allocate memory for the firmware image */
73 	mem_region = memremap(mem_phys, mem_size,  MEMREMAP_WC);
74 	if (!mem_region) {
75 		ret = -ENOMEM;
76 		goto out;
77 	}
78 
79 	/*
80 	 * Load the rest of the MDT
81 	 *
82 	 * Note that we could be dealing with two different paths, since
83 	 * with upstream linux-firmware it would be in a qcom/ subdir..
84 	 * adreno_request_fw() handles this, but qcom_mdt_load() does
85 	 * not.  But since we've already gotten thru adreno_request_fw()
86 	 * we know which of the two cases it is:
87 	 */
88 	if (to_adreno_gpu(gpu)->fwloc == FW_LOCATION_LEGACY) {
89 		ret = qcom_mdt_load(dev, fw, fwname, GPU_PAS_ID,
90 				mem_region, mem_phys, mem_size);
91 	} else {
92 		char newname[strlen("qcom/") + strlen(fwname) + 1];
93 
94 		sprintf(newname, "qcom/%s", fwname);
95 
96 		ret = qcom_mdt_load(dev, fw, newname, GPU_PAS_ID,
97 				mem_region, mem_phys, mem_size);
98 	}
99 	if (ret)
100 		goto out;
101 
102 	/* Send the image to the secure world */
103 	ret = qcom_scm_pas_auth_and_reset(GPU_PAS_ID);
104 	if (ret)
105 		DRM_DEV_ERROR(dev, "Unable to authorize the image\n");
106 
107 out:
108 	if (mem_region)
109 		memunmap(mem_region);
110 
111 	release_firmware(fw);
112 
113 	return ret;
114 }
115 
116 static void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
117 {
118 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
119 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
120 	uint32_t wptr;
121 	unsigned long flags;
122 
123 	spin_lock_irqsave(&ring->lock, flags);
124 
125 	/* Copy the shadow to the actual register */
126 	ring->cur = ring->next;
127 
128 	/* Make sure to wrap wptr if we need to */
129 	wptr = get_wptr(ring);
130 
131 	spin_unlock_irqrestore(&ring->lock, flags);
132 
133 	/* Make sure everything is posted before making a decision */
134 	mb();
135 
136 	/* Update HW if this is the current ring and we are not in preempt */
137 	if (a5xx_gpu->cur_ring == ring && !a5xx_in_preempt(a5xx_gpu))
138 		gpu_write(gpu, REG_A5XX_CP_RB_WPTR, wptr);
139 }
140 
141 static void a5xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
142 	struct msm_file_private *ctx)
143 {
144 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
145 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
146 	struct msm_drm_private *priv = gpu->dev->dev_private;
147 	struct msm_ringbuffer *ring = submit->ring;
148 	unsigned int i, ibs = 0;
149 
150 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
151 	OUT_RING(ring, 0x02);
152 
153 	/* Turn off protected mode to write to special registers */
154 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
155 	OUT_RING(ring, 0);
156 
157 	/* Set the save preemption record for the ring/command */
158 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
159 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
160 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[submit->ring->id]));
161 
162 	/* Turn back on protected mode */
163 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
164 	OUT_RING(ring, 1);
165 
166 	/* Enable local preemption for finegrain preemption */
167 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
168 	OUT_RING(ring, 0x02);
169 
170 	/* Allow CP_CONTEXT_SWITCH_YIELD packets in the IB2 */
171 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
172 	OUT_RING(ring, 0x02);
173 
174 	/* Submit the commands */
175 	for (i = 0; i < submit->nr_cmds; i++) {
176 		switch (submit->cmd[i].type) {
177 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
178 			break;
179 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
180 			if (priv->lastctx == ctx)
181 				break;
182 		case MSM_SUBMIT_CMD_BUF:
183 			OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3);
184 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
185 			OUT_RING(ring, upper_32_bits(submit->cmd[i].iova));
186 			OUT_RING(ring, submit->cmd[i].size);
187 			ibs++;
188 			break;
189 		}
190 	}
191 
192 	/*
193 	 * Write the render mode to NULL (0) to indicate to the CP that the IBs
194 	 * are done rendering - otherwise a lucky preemption would start
195 	 * replaying from the last checkpoint
196 	 */
197 	OUT_PKT7(ring, CP_SET_RENDER_MODE, 5);
198 	OUT_RING(ring, 0);
199 	OUT_RING(ring, 0);
200 	OUT_RING(ring, 0);
201 	OUT_RING(ring, 0);
202 	OUT_RING(ring, 0);
203 
204 	/* Turn off IB level preemptions */
205 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
206 	OUT_RING(ring, 0x01);
207 
208 	/* Write the fence to the scratch register */
209 	OUT_PKT4(ring, REG_A5XX_CP_SCRATCH_REG(2), 1);
210 	OUT_RING(ring, submit->seqno);
211 
212 	/*
213 	 * Execute a CACHE_FLUSH_TS event. This will ensure that the
214 	 * timestamp is written to the memory and then triggers the interrupt
215 	 */
216 	OUT_PKT7(ring, CP_EVENT_WRITE, 4);
217 	OUT_RING(ring, CACHE_FLUSH_TS | (1 << 31));
218 	OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence)));
219 	OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence)));
220 	OUT_RING(ring, submit->seqno);
221 
222 	/* Yield the floor on command completion */
223 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
224 	/*
225 	 * If dword[2:1] are non zero, they specify an address for the CP to
226 	 * write the value of dword[3] to on preemption complete. Write 0 to
227 	 * skip the write
228 	 */
229 	OUT_RING(ring, 0x00);
230 	OUT_RING(ring, 0x00);
231 	/* Data value - not used if the address above is 0 */
232 	OUT_RING(ring, 0x01);
233 	/* Set bit 0 to trigger an interrupt on preempt complete */
234 	OUT_RING(ring, 0x01);
235 
236 	a5xx_flush(gpu, ring);
237 
238 	/* Check to see if we need to start preemption */
239 	a5xx_preempt_trigger(gpu);
240 }
241 
242 static const struct {
243 	u32 offset;
244 	u32 value;
245 } a5xx_hwcg[] = {
246 	{REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222},
247 	{REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222},
248 	{REG_A5XX_RBBM_CLOCK_CNTL_SP2, 0x02222222},
249 	{REG_A5XX_RBBM_CLOCK_CNTL_SP3, 0x02222222},
250 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220},
251 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220},
252 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP2, 0x02222220},
253 	{REG_A5XX_RBBM_CLOCK_CNTL2_SP3, 0x02222220},
254 	{REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF},
255 	{REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF},
256 	{REG_A5XX_RBBM_CLOCK_HYST_SP2, 0x0000F3CF},
257 	{REG_A5XX_RBBM_CLOCK_HYST_SP3, 0x0000F3CF},
258 	{REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080},
259 	{REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080},
260 	{REG_A5XX_RBBM_CLOCK_DELAY_SP2, 0x00000080},
261 	{REG_A5XX_RBBM_CLOCK_DELAY_SP3, 0x00000080},
262 	{REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222},
263 	{REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222},
264 	{REG_A5XX_RBBM_CLOCK_CNTL_TP2, 0x22222222},
265 	{REG_A5XX_RBBM_CLOCK_CNTL_TP3, 0x22222222},
266 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222},
267 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222},
268 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222},
269 	{REG_A5XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222},
270 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222},
271 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222},
272 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP2, 0x00002222},
273 	{REG_A5XX_RBBM_CLOCK_CNTL3_TP3, 0x00002222},
274 	{REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777},
275 	{REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777},
276 	{REG_A5XX_RBBM_CLOCK_HYST_TP2, 0x77777777},
277 	{REG_A5XX_RBBM_CLOCK_HYST_TP3, 0x77777777},
278 	{REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777},
279 	{REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777},
280 	{REG_A5XX_RBBM_CLOCK_HYST2_TP2, 0x77777777},
281 	{REG_A5XX_RBBM_CLOCK_HYST2_TP3, 0x77777777},
282 	{REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777},
283 	{REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777},
284 	{REG_A5XX_RBBM_CLOCK_HYST3_TP2, 0x00007777},
285 	{REG_A5XX_RBBM_CLOCK_HYST3_TP3, 0x00007777},
286 	{REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111},
287 	{REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111},
288 	{REG_A5XX_RBBM_CLOCK_DELAY_TP2, 0x11111111},
289 	{REG_A5XX_RBBM_CLOCK_DELAY_TP3, 0x11111111},
290 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111},
291 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111},
292 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111},
293 	{REG_A5XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111},
294 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111},
295 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111},
296 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP2, 0x00001111},
297 	{REG_A5XX_RBBM_CLOCK_DELAY3_TP3, 0x00001111},
298 	{REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222},
299 	{REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222},
300 	{REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222},
301 	{REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222},
302 	{REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444},
303 	{REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002},
304 	{REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222},
305 	{REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222},
306 	{REG_A5XX_RBBM_CLOCK_CNTL_RB2, 0x22222222},
307 	{REG_A5XX_RBBM_CLOCK_CNTL_RB3, 0x22222222},
308 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222},
309 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222},
310 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB2, 0x00222222},
311 	{REG_A5XX_RBBM_CLOCK_CNTL2_RB3, 0x00222222},
312 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220},
313 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220},
314 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU2, 0x00022220},
315 	{REG_A5XX_RBBM_CLOCK_CNTL_CCU3, 0x00022220},
316 	{REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222},
317 	{REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555},
318 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404},
319 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404},
320 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU2, 0x04040404},
321 	{REG_A5XX_RBBM_CLOCK_HYST_RB_CCU3, 0x04040404},
322 	{REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044},
323 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002},
324 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002},
325 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_2, 0x00000002},
326 	{REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_3, 0x00000002},
327 	{REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011},
328 	{REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222},
329 	{REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222},
330 	{REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222},
331 	{REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000},
332 	{REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004},
333 	{REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000},
334 	{REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000},
335 	{REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000},
336 	{REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200},
337 	{REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222}
338 };
339 
340 void a5xx_set_hwcg(struct msm_gpu *gpu, bool state)
341 {
342 	unsigned int i;
343 
344 	for (i = 0; i < ARRAY_SIZE(a5xx_hwcg); i++)
345 		gpu_write(gpu, a5xx_hwcg[i].offset,
346 			state ? a5xx_hwcg[i].value : 0);
347 
348 	gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, state ? 0xAAA8AA00 : 0);
349 	gpu_write(gpu, REG_A5XX_RBBM_ISDB_CNT, state ? 0x182 : 0x180);
350 }
351 
352 static int a5xx_me_init(struct msm_gpu *gpu)
353 {
354 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
355 	struct msm_ringbuffer *ring = gpu->rb[0];
356 
357 	OUT_PKT7(ring, CP_ME_INIT, 8);
358 
359 	OUT_RING(ring, 0x0000002F);
360 
361 	/* Enable multiple hardware contexts */
362 	OUT_RING(ring, 0x00000003);
363 
364 	/* Enable error detection */
365 	OUT_RING(ring, 0x20000000);
366 
367 	/* Don't enable header dump */
368 	OUT_RING(ring, 0x00000000);
369 	OUT_RING(ring, 0x00000000);
370 
371 	/* Specify workarounds for various microcode issues */
372 	if (adreno_is_a530(adreno_gpu)) {
373 		/* Workaround for token end syncs
374 		 * Force a WFI after every direct-render 3D mode draw and every
375 		 * 2D mode 3 draw
376 		 */
377 		OUT_RING(ring, 0x0000000B);
378 	} else {
379 		/* No workarounds enabled */
380 		OUT_RING(ring, 0x00000000);
381 	}
382 
383 	OUT_RING(ring, 0x00000000);
384 	OUT_RING(ring, 0x00000000);
385 
386 	gpu->funcs->flush(gpu, ring);
387 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
388 }
389 
390 static int a5xx_preempt_start(struct msm_gpu *gpu)
391 {
392 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
393 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
394 	struct msm_ringbuffer *ring = gpu->rb[0];
395 
396 	if (gpu->nr_rings == 1)
397 		return 0;
398 
399 	/* Turn off protected mode to write to special registers */
400 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
401 	OUT_RING(ring, 0);
402 
403 	/* Set the save preemption record for the ring/command */
404 	OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2);
405 	OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[ring->id]));
406 	OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[ring->id]));
407 
408 	/* Turn back on protected mode */
409 	OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1);
410 	OUT_RING(ring, 1);
411 
412 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1);
413 	OUT_RING(ring, 0x00);
414 
415 	OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1);
416 	OUT_RING(ring, 0x01);
417 
418 	OUT_PKT7(ring, CP_YIELD_ENABLE, 1);
419 	OUT_RING(ring, 0x01);
420 
421 	/* Yield the floor on command completion */
422 	OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4);
423 	OUT_RING(ring, 0x00);
424 	OUT_RING(ring, 0x00);
425 	OUT_RING(ring, 0x01);
426 	OUT_RING(ring, 0x01);
427 
428 	gpu->funcs->flush(gpu, ring);
429 
430 	return a5xx_idle(gpu, ring) ? 0 : -EINVAL;
431 }
432 
433 
434 static struct drm_gem_object *a5xx_ucode_load_bo(struct msm_gpu *gpu,
435 		const struct firmware *fw, u64 *iova)
436 {
437 	struct drm_gem_object *bo;
438 	void *ptr;
439 
440 	ptr = msm_gem_kernel_new_locked(gpu->dev, fw->size - 4,
441 		MSM_BO_UNCACHED | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
442 
443 	if (IS_ERR(ptr))
444 		return ERR_CAST(ptr);
445 
446 	memcpy(ptr, &fw->data[4], fw->size - 4);
447 
448 	msm_gem_put_vaddr(bo);
449 	return bo;
450 }
451 
452 static int a5xx_ucode_init(struct msm_gpu *gpu)
453 {
454 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
455 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
456 	int ret;
457 
458 	if (!a5xx_gpu->pm4_bo) {
459 		a5xx_gpu->pm4_bo = a5xx_ucode_load_bo(gpu, adreno_gpu->pm4,
460 			&a5xx_gpu->pm4_iova);
461 
462 		if (IS_ERR(a5xx_gpu->pm4_bo)) {
463 			ret = PTR_ERR(a5xx_gpu->pm4_bo);
464 			a5xx_gpu->pm4_bo = NULL;
465 			dev_err(gpu->dev->dev, "could not allocate PM4: %d\n",
466 				ret);
467 			return ret;
468 		}
469 	}
470 
471 	if (!a5xx_gpu->pfp_bo) {
472 		a5xx_gpu->pfp_bo = a5xx_ucode_load_bo(gpu, adreno_gpu->pfp,
473 			&a5xx_gpu->pfp_iova);
474 
475 		if (IS_ERR(a5xx_gpu->pfp_bo)) {
476 			ret = PTR_ERR(a5xx_gpu->pfp_bo);
477 			a5xx_gpu->pfp_bo = NULL;
478 			dev_err(gpu->dev->dev, "could not allocate PFP: %d\n",
479 				ret);
480 			return ret;
481 		}
482 	}
483 
484 	gpu_write64(gpu, REG_A5XX_CP_ME_INSTR_BASE_LO,
485 		REG_A5XX_CP_ME_INSTR_BASE_HI, a5xx_gpu->pm4_iova);
486 
487 	gpu_write64(gpu, REG_A5XX_CP_PFP_INSTR_BASE_LO,
488 		REG_A5XX_CP_PFP_INSTR_BASE_HI, a5xx_gpu->pfp_iova);
489 
490 	return 0;
491 }
492 
493 #define SCM_GPU_ZAP_SHADER_RESUME 0
494 
495 static int a5xx_zap_shader_resume(struct msm_gpu *gpu)
496 {
497 	int ret;
498 
499 	ret = qcom_scm_set_remote_state(SCM_GPU_ZAP_SHADER_RESUME, GPU_PAS_ID);
500 	if (ret)
501 		DRM_ERROR("%s: zap-shader resume failed: %d\n",
502 			gpu->name, ret);
503 
504 	return ret;
505 }
506 
507 static int a5xx_zap_shader_init(struct msm_gpu *gpu)
508 {
509 	static bool loaded;
510 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
511 	struct platform_device *pdev = gpu->pdev;
512 	int ret;
513 
514 	/*
515 	 * If the zap shader is already loaded into memory we just need to kick
516 	 * the remote processor to reinitialize it
517 	 */
518 	if (loaded)
519 		return a5xx_zap_shader_resume(gpu);
520 
521 	/* We need SCM to be able to load the firmware */
522 	if (!qcom_scm_is_available()) {
523 		DRM_DEV_ERROR(&pdev->dev, "SCM is not available\n");
524 		return -EPROBE_DEFER;
525 	}
526 
527 	/* Each GPU has a target specific zap shader firmware name to use */
528 	if (!adreno_gpu->info->zapfw) {
529 		DRM_DEV_ERROR(&pdev->dev,
530 			"Zap shader firmware file not specified for this target\n");
531 		return -ENODEV;
532 	}
533 
534 	ret = zap_shader_load_mdt(gpu, adreno_gpu->info->zapfw);
535 
536 	loaded = !ret;
537 
538 	return ret;
539 }
540 
541 #define A5XX_INT_MASK (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
542 	  A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
543 	  A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
544 	  A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
545 	  A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
546 	  A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW | \
547 	  A5XX_RBBM_INT_0_MASK_CP_HW_ERROR | \
548 	  A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT | \
549 	  A5XX_RBBM_INT_0_MASK_CP_SW | \
550 	  A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \
551 	  A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \
552 	  A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
553 
554 static int a5xx_hw_init(struct msm_gpu *gpu)
555 {
556 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
557 	int ret;
558 
559 	gpu_write(gpu, REG_A5XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x00000003);
560 
561 	/* Make all blocks contribute to the GPU BUSY perf counter */
562 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xFFFFFFFF);
563 
564 	/* Enable RBBM error reporting bits */
565 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL0, 0x00000001);
566 
567 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_FAULT_DETECT_MASK) {
568 		/*
569 		 * Mask out the activity signals from RB1-3 to avoid false
570 		 * positives
571 		 */
572 
573 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL11,
574 			0xF0000000);
575 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL12,
576 			0xFFFFFFFF);
577 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL13,
578 			0xFFFFFFFF);
579 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL14,
580 			0xFFFFFFFF);
581 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL15,
582 			0xFFFFFFFF);
583 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL16,
584 			0xFFFFFFFF);
585 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL17,
586 			0xFFFFFFFF);
587 		gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL18,
588 			0xFFFFFFFF);
589 	}
590 
591 	/* Enable fault detection */
592 	gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_INT_CNTL,
593 		(1 << 30) | 0xFFFF);
594 
595 	/* Turn on performance counters */
596 	gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_CNTL, 0x01);
597 
598 	/* Increase VFD cache access so LRZ and other data gets evicted less */
599 	gpu_write(gpu, REG_A5XX_UCHE_CACHE_WAYS, 0x02);
600 
601 	/* Disable L2 bypass in the UCHE */
602 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_LO, 0xFFFF0000);
603 	gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_HI, 0x0001FFFF);
604 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_LO, 0xFFFF0000);
605 	gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_HI, 0x0001FFFF);
606 
607 	/* Set the GMEM VA range (0 to gpu->gmem) */
608 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_LO, 0x00100000);
609 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_HI, 0x00000000);
610 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_LO,
611 		0x00100000 + adreno_gpu->gmem - 1);
612 	gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_HI, 0x00000000);
613 
614 	gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x40);
615 	gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x40);
616 	gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x80000060);
617 	gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x40201B16);
618 
619 	gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, (0x400 << 11 | 0x300 << 22));
620 
621 	if (adreno_gpu->info->quirks & ADRENO_QUIRK_TWO_PASS_USE_WFI)
622 		gpu_rmw(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0, (1 << 8));
623 
624 	gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0xc0200100);
625 
626 	/* Enable USE_RETENTION_FLOPS */
627 	gpu_write(gpu, REG_A5XX_CP_CHICKEN_DBG, 0x02000000);
628 
629 	/* Enable ME/PFP split notification */
630 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL1, 0xA6FFFFFF);
631 
632 	/* Enable HWCG */
633 	a5xx_set_hwcg(gpu, true);
634 
635 	gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL2, 0x0000003F);
636 
637 	/* Set the highest bank bit */
638 	gpu_write(gpu, REG_A5XX_TPL1_MODE_CNTL, 2 << 7);
639 	gpu_write(gpu, REG_A5XX_RB_MODE_CNTL, 2 << 1);
640 
641 	/* Protect registers from the CP */
642 	gpu_write(gpu, REG_A5XX_CP_PROTECT_CNTL, 0x00000007);
643 
644 	/* RBBM */
645 	gpu_write(gpu, REG_A5XX_CP_PROTECT(0), ADRENO_PROTECT_RW(0x04, 4));
646 	gpu_write(gpu, REG_A5XX_CP_PROTECT(1), ADRENO_PROTECT_RW(0x08, 8));
647 	gpu_write(gpu, REG_A5XX_CP_PROTECT(2), ADRENO_PROTECT_RW(0x10, 16));
648 	gpu_write(gpu, REG_A5XX_CP_PROTECT(3), ADRENO_PROTECT_RW(0x20, 32));
649 	gpu_write(gpu, REG_A5XX_CP_PROTECT(4), ADRENO_PROTECT_RW(0x40, 64));
650 	gpu_write(gpu, REG_A5XX_CP_PROTECT(5), ADRENO_PROTECT_RW(0x80, 64));
651 
652 	/* Content protect */
653 	gpu_write(gpu, REG_A5XX_CP_PROTECT(6),
654 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
655 			16));
656 	gpu_write(gpu, REG_A5XX_CP_PROTECT(7),
657 		ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TRUST_CNTL, 2));
658 
659 	/* CP */
660 	gpu_write(gpu, REG_A5XX_CP_PROTECT(8), ADRENO_PROTECT_RW(0x800, 64));
661 	gpu_write(gpu, REG_A5XX_CP_PROTECT(9), ADRENO_PROTECT_RW(0x840, 8));
662 	gpu_write(gpu, REG_A5XX_CP_PROTECT(10), ADRENO_PROTECT_RW(0x880, 32));
663 	gpu_write(gpu, REG_A5XX_CP_PROTECT(11), ADRENO_PROTECT_RW(0xAA0, 1));
664 
665 	/* RB */
666 	gpu_write(gpu, REG_A5XX_CP_PROTECT(12), ADRENO_PROTECT_RW(0xCC0, 1));
667 	gpu_write(gpu, REG_A5XX_CP_PROTECT(13), ADRENO_PROTECT_RW(0xCF0, 2));
668 
669 	/* VPC */
670 	gpu_write(gpu, REG_A5XX_CP_PROTECT(14), ADRENO_PROTECT_RW(0xE68, 8));
671 	gpu_write(gpu, REG_A5XX_CP_PROTECT(15), ADRENO_PROTECT_RW(0xE70, 4));
672 
673 	/* UCHE */
674 	gpu_write(gpu, REG_A5XX_CP_PROTECT(16), ADRENO_PROTECT_RW(0xE80, 16));
675 
676 	if (adreno_is_a530(adreno_gpu))
677 		gpu_write(gpu, REG_A5XX_CP_PROTECT(17),
678 			ADRENO_PROTECT_RW(0x10000, 0x8000));
679 
680 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_CNTL, 0);
681 	/*
682 	 * Disable the trusted memory range - we don't actually supported secure
683 	 * memory rendering at this point in time and we don't want to block off
684 	 * part of the virtual memory space.
685 	 */
686 	gpu_write64(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO,
687 		REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_HI, 0x00000000);
688 	gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000);
689 
690 	ret = adreno_hw_init(gpu);
691 	if (ret)
692 		return ret;
693 
694 	a5xx_preempt_hw_init(gpu);
695 
696 	a5xx_gpmu_ucode_init(gpu);
697 
698 	ret = a5xx_ucode_init(gpu);
699 	if (ret)
700 		return ret;
701 
702 	/* Disable the interrupts through the initial bringup stage */
703 	gpu_write(gpu, REG_A5XX_RBBM_INT_0_MASK, A5XX_INT_MASK);
704 
705 	/* Clear ME_HALT to start the micro engine */
706 	gpu_write(gpu, REG_A5XX_CP_PFP_ME_CNTL, 0);
707 	ret = a5xx_me_init(gpu);
708 	if (ret)
709 		return ret;
710 
711 	ret = a5xx_power_init(gpu);
712 	if (ret)
713 		return ret;
714 
715 	/*
716 	 * Send a pipeline event stat to get misbehaving counters to start
717 	 * ticking correctly
718 	 */
719 	if (adreno_is_a530(adreno_gpu)) {
720 		OUT_PKT7(gpu->rb[0], CP_EVENT_WRITE, 1);
721 		OUT_RING(gpu->rb[0], 0x0F);
722 
723 		gpu->funcs->flush(gpu, gpu->rb[0]);
724 		if (!a5xx_idle(gpu, gpu->rb[0]))
725 			return -EINVAL;
726 	}
727 
728 	/*
729 	 * Try to load a zap shader into the secure world. If successful
730 	 * we can use the CP to switch out of secure mode. If not then we
731 	 * have no resource but to try to switch ourselves out manually. If we
732 	 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will
733 	 * be blocked and a permissions violation will soon follow.
734 	 */
735 	ret = a5xx_zap_shader_init(gpu);
736 	if (!ret) {
737 		OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1);
738 		OUT_RING(gpu->rb[0], 0x00000000);
739 
740 		gpu->funcs->flush(gpu, gpu->rb[0]);
741 		if (!a5xx_idle(gpu, gpu->rb[0]))
742 			return -EINVAL;
743 	} else {
744 		/* Print a warning so if we die, we know why */
745 		dev_warn_once(gpu->dev->dev,
746 			"Zap shader not enabled - using SECVID_TRUST_CNTL instead\n");
747 		gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0);
748 	}
749 
750 	/* Last step - yield the ringbuffer */
751 	a5xx_preempt_start(gpu);
752 
753 	return 0;
754 }
755 
756 static void a5xx_recover(struct msm_gpu *gpu)
757 {
758 	int i;
759 
760 	adreno_dump_info(gpu);
761 
762 	for (i = 0; i < 8; i++) {
763 		printk("CP_SCRATCH_REG%d: %u\n", i,
764 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(i)));
765 	}
766 
767 	if (hang_debug)
768 		a5xx_dump(gpu);
769 
770 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 1);
771 	gpu_read(gpu, REG_A5XX_RBBM_SW_RESET_CMD);
772 	gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 0);
773 	adreno_recover(gpu);
774 }
775 
776 static void a5xx_destroy(struct msm_gpu *gpu)
777 {
778 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
779 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
780 
781 	DBG("%s", gpu->name);
782 
783 	a5xx_preempt_fini(gpu);
784 
785 	if (a5xx_gpu->pm4_bo) {
786 		if (a5xx_gpu->pm4_iova)
787 			msm_gem_put_iova(a5xx_gpu->pm4_bo, gpu->aspace);
788 		drm_gem_object_unreference_unlocked(a5xx_gpu->pm4_bo);
789 	}
790 
791 	if (a5xx_gpu->pfp_bo) {
792 		if (a5xx_gpu->pfp_iova)
793 			msm_gem_put_iova(a5xx_gpu->pfp_bo, gpu->aspace);
794 		drm_gem_object_unreference_unlocked(a5xx_gpu->pfp_bo);
795 	}
796 
797 	if (a5xx_gpu->gpmu_bo) {
798 		if (a5xx_gpu->gpmu_iova)
799 			msm_gem_put_iova(a5xx_gpu->gpmu_bo, gpu->aspace);
800 		drm_gem_object_unreference_unlocked(a5xx_gpu->gpmu_bo);
801 	}
802 
803 	adreno_gpu_cleanup(adreno_gpu);
804 	kfree(a5xx_gpu);
805 }
806 
807 static inline bool _a5xx_check_idle(struct msm_gpu *gpu)
808 {
809 	if (gpu_read(gpu, REG_A5XX_RBBM_STATUS) & ~A5XX_RBBM_STATUS_HI_BUSY)
810 		return false;
811 
812 	/*
813 	 * Nearly every abnormality ends up pausing the GPU and triggering a
814 	 * fault so we can safely just watch for this one interrupt to fire
815 	 */
816 	return !(gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS) &
817 		A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT);
818 }
819 
820 bool a5xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
821 {
822 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
823 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
824 
825 	if (ring != a5xx_gpu->cur_ring) {
826 		WARN(1, "Tried to idle a non-current ringbuffer\n");
827 		return false;
828 	}
829 
830 	/* wait for CP to drain ringbuffer: */
831 	if (!adreno_idle(gpu, ring))
832 		return false;
833 
834 	if (spin_until(_a5xx_check_idle(gpu))) {
835 		DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n",
836 			gpu->name, __builtin_return_address(0),
837 			gpu_read(gpu, REG_A5XX_RBBM_STATUS),
838 			gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS),
839 			gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
840 			gpu_read(gpu, REG_A5XX_CP_RB_WPTR));
841 		return false;
842 	}
843 
844 	return true;
845 }
846 
847 static int a5xx_fault_handler(void *arg, unsigned long iova, int flags)
848 {
849 	struct msm_gpu *gpu = arg;
850 	pr_warn_ratelimited("*** gpu fault: iova=%08lx, flags=%d (%u,%u,%u,%u)\n",
851 			iova, flags,
852 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(4)),
853 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(5)),
854 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(6)),
855 			gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(7)));
856 
857 	return -EFAULT;
858 }
859 
860 static void a5xx_cp_err_irq(struct msm_gpu *gpu)
861 {
862 	u32 status = gpu_read(gpu, REG_A5XX_CP_INTERRUPT_STATUS);
863 
864 	if (status & A5XX_CP_INT_CP_OPCODE_ERROR) {
865 		u32 val;
866 
867 		gpu_write(gpu, REG_A5XX_CP_PFP_STAT_ADDR, 0);
868 
869 		/*
870 		 * REG_A5XX_CP_PFP_STAT_DATA is indexed, and we want index 1 so
871 		 * read it twice
872 		 */
873 
874 		gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
875 		val = gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA);
876 
877 		dev_err_ratelimited(gpu->dev->dev, "CP | opcode error | possible opcode=0x%8.8X\n",
878 			val);
879 	}
880 
881 	if (status & A5XX_CP_INT_CP_HW_FAULT_ERROR)
882 		dev_err_ratelimited(gpu->dev->dev, "CP | HW fault | status=0x%8.8X\n",
883 			gpu_read(gpu, REG_A5XX_CP_HW_FAULT));
884 
885 	if (status & A5XX_CP_INT_CP_DMA_ERROR)
886 		dev_err_ratelimited(gpu->dev->dev, "CP | DMA error\n");
887 
888 	if (status & A5XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) {
889 		u32 val = gpu_read(gpu, REG_A5XX_CP_PROTECT_STATUS);
890 
891 		dev_err_ratelimited(gpu->dev->dev,
892 			"CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n",
893 			val & (1 << 24) ? "WRITE" : "READ",
894 			(val & 0xFFFFF) >> 2, val);
895 	}
896 
897 	if (status & A5XX_CP_INT_CP_AHB_ERROR) {
898 		u32 status = gpu_read(gpu, REG_A5XX_CP_AHB_FAULT);
899 		const char *access[16] = { "reserved", "reserved",
900 			"timestamp lo", "timestamp hi", "pfp read", "pfp write",
901 			"", "", "me read", "me write", "", "", "crashdump read",
902 			"crashdump write" };
903 
904 		dev_err_ratelimited(gpu->dev->dev,
905 			"CP | AHB error | addr=%X access=%s error=%d | status=0x%8.8X\n",
906 			status & 0xFFFFF, access[(status >> 24) & 0xF],
907 			(status & (1 << 31)), status);
908 	}
909 }
910 
911 static void a5xx_rbbm_err_irq(struct msm_gpu *gpu, u32 status)
912 {
913 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR) {
914 		u32 val = gpu_read(gpu, REG_A5XX_RBBM_AHB_ERROR_STATUS);
915 
916 		dev_err_ratelimited(gpu->dev->dev,
917 			"RBBM | AHB bus error | %s | addr=0x%X | ports=0x%X:0x%X\n",
918 			val & (1 << 28) ? "WRITE" : "READ",
919 			(val & 0xFFFFF) >> 2, (val >> 20) & 0x3,
920 			(val >> 24) & 0xF);
921 
922 		/* Clear the error */
923 		gpu_write(gpu, REG_A5XX_RBBM_AHB_CMD, (1 << 4));
924 
925 		/* Clear the interrupt */
926 		gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
927 			A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
928 	}
929 
930 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT)
931 		dev_err_ratelimited(gpu->dev->dev, "RBBM | AHB transfer timeout\n");
932 
933 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT)
934 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ME master split | status=0x%X\n",
935 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ME_SPLIT_STATUS));
936 
937 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT)
938 		dev_err_ratelimited(gpu->dev->dev, "RBBM | PFP master split | status=0x%X\n",
939 			gpu_read(gpu, REG_A5XX_RBBM_AHB_PFP_SPLIT_STATUS));
940 
941 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT)
942 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ETS master split | status=0x%X\n",
943 			gpu_read(gpu, REG_A5XX_RBBM_AHB_ETS_SPLIT_STATUS));
944 
945 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
946 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB ASYNC overflow\n");
947 
948 	if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW)
949 		dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB bus overflow\n");
950 }
951 
952 static void a5xx_uche_err_irq(struct msm_gpu *gpu)
953 {
954 	uint64_t addr = (uint64_t) gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_HI);
955 
956 	addr |= gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_LO);
957 
958 	dev_err_ratelimited(gpu->dev->dev, "UCHE | Out of bounds access | addr=0x%llX\n",
959 		addr);
960 }
961 
962 static void a5xx_gpmu_err_irq(struct msm_gpu *gpu)
963 {
964 	dev_err_ratelimited(gpu->dev->dev, "GPMU | voltage droop\n");
965 }
966 
967 static void a5xx_fault_detect_irq(struct msm_gpu *gpu)
968 {
969 	struct drm_device *dev = gpu->dev;
970 	struct msm_drm_private *priv = dev->dev_private;
971 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
972 
973 	dev_err(dev->dev, "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",
974 		ring ? ring->id : -1, ring ? ring->seqno : 0,
975 		gpu_read(gpu, REG_A5XX_RBBM_STATUS),
976 		gpu_read(gpu, REG_A5XX_CP_RB_RPTR),
977 		gpu_read(gpu, REG_A5XX_CP_RB_WPTR),
978 		gpu_read64(gpu, REG_A5XX_CP_IB1_BASE, REG_A5XX_CP_IB1_BASE_HI),
979 		gpu_read(gpu, REG_A5XX_CP_IB1_BUFSZ),
980 		gpu_read64(gpu, REG_A5XX_CP_IB2_BASE, REG_A5XX_CP_IB2_BASE_HI),
981 		gpu_read(gpu, REG_A5XX_CP_IB2_BUFSZ));
982 
983 	/* Turn off the hangcheck timer to keep it from bothering us */
984 	del_timer(&gpu->hangcheck_timer);
985 
986 	queue_work(priv->wq, &gpu->recover_work);
987 }
988 
989 #define RBBM_ERROR_MASK \
990 	(A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \
991 	A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \
992 	A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \
993 	A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \
994 	A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \
995 	A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW)
996 
997 static irqreturn_t a5xx_irq(struct msm_gpu *gpu)
998 {
999 	u32 status = gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS);
1000 
1001 	/*
1002 	 * Clear all the interrupts except RBBM_AHB_ERROR - if we clear it
1003 	 * before the source is cleared the interrupt will storm.
1004 	 */
1005 	gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD,
1006 		status & ~A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR);
1007 
1008 	/* Pass status to a5xx_rbbm_err_irq because we've already cleared it */
1009 	if (status & RBBM_ERROR_MASK)
1010 		a5xx_rbbm_err_irq(gpu, status);
1011 
1012 	if (status & A5XX_RBBM_INT_0_MASK_CP_HW_ERROR)
1013 		a5xx_cp_err_irq(gpu);
1014 
1015 	if (status & A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT)
1016 		a5xx_fault_detect_irq(gpu);
1017 
1018 	if (status & A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS)
1019 		a5xx_uche_err_irq(gpu);
1020 
1021 	if (status & A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP)
1022 		a5xx_gpmu_err_irq(gpu);
1023 
1024 	if (status & A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS) {
1025 		a5xx_preempt_trigger(gpu);
1026 		msm_gpu_retire(gpu);
1027 	}
1028 
1029 	if (status & A5XX_RBBM_INT_0_MASK_CP_SW)
1030 		a5xx_preempt_irq(gpu);
1031 
1032 	return IRQ_HANDLED;
1033 }
1034 
1035 static const u32 a5xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
1036 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_A5XX_CP_RB_BASE),
1037 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE_HI, REG_A5XX_CP_RB_BASE_HI),
1038 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR, REG_A5XX_CP_RB_RPTR_ADDR),
1039 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR_HI,
1040 		REG_A5XX_CP_RB_RPTR_ADDR_HI),
1041 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_A5XX_CP_RB_RPTR),
1042 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_A5XX_CP_RB_WPTR),
1043 	REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_A5XX_CP_RB_CNTL),
1044 };
1045 
1046 static const u32 a5xx_registers[] = {
1047 	0x0000, 0x0002, 0x0004, 0x0020, 0x0022, 0x0026, 0x0029, 0x002B,
1048 	0x002E, 0x0035, 0x0038, 0x0042, 0x0044, 0x0044, 0x0047, 0x0095,
1049 	0x0097, 0x00BB, 0x03A0, 0x0464, 0x0469, 0x046F, 0x04D2, 0x04D3,
1050 	0x04E0, 0x0533, 0x0540, 0x0555, 0x0800, 0x081A, 0x081F, 0x0841,
1051 	0x0860, 0x0860, 0x0880, 0x08A0, 0x0B00, 0x0B12, 0x0B15, 0x0B28,
1052 	0x0B78, 0x0B7F, 0x0BB0, 0x0BBD, 0x0BC0, 0x0BC6, 0x0BD0, 0x0C53,
1053 	0x0C60, 0x0C61, 0x0C80, 0x0C82, 0x0C84, 0x0C85, 0x0C90, 0x0C98,
1054 	0x0CA0, 0x0CA0, 0x0CB0, 0x0CB2, 0x2180, 0x2185, 0x2580, 0x2585,
1055 	0x0CC1, 0x0CC1, 0x0CC4, 0x0CC7, 0x0CCC, 0x0CCC, 0x0CD0, 0x0CD8,
1056 	0x0CE0, 0x0CE5, 0x0CE8, 0x0CE8, 0x0CEC, 0x0CF1, 0x0CFB, 0x0D0E,
1057 	0x2100, 0x211E, 0x2140, 0x2145, 0x2500, 0x251E, 0x2540, 0x2545,
1058 	0x0D10, 0x0D17, 0x0D20, 0x0D23, 0x0D30, 0x0D30, 0x20C0, 0x20C0,
1059 	0x24C0, 0x24C0, 0x0E40, 0x0E43, 0x0E4A, 0x0E4A, 0x0E50, 0x0E57,
1060 	0x0E60, 0x0E7C, 0x0E80, 0x0E8E, 0x0E90, 0x0E96, 0x0EA0, 0x0EA8,
1061 	0x0EB0, 0x0EB2, 0xE140, 0xE147, 0xE150, 0xE187, 0xE1A0, 0xE1A9,
1062 	0xE1B0, 0xE1B6, 0xE1C0, 0xE1C7, 0xE1D0, 0xE1D1, 0xE200, 0xE201,
1063 	0xE210, 0xE21C, 0xE240, 0xE268, 0xE000, 0xE006, 0xE010, 0xE09A,
1064 	0xE0A0, 0xE0A4, 0xE0AA, 0xE0EB, 0xE100, 0xE105, 0xE380, 0xE38F,
1065 	0xE3B0, 0xE3B0, 0xE400, 0xE405, 0xE408, 0xE4E9, 0xE4F0, 0xE4F0,
1066 	0xE280, 0xE280, 0xE282, 0xE2A3, 0xE2A5, 0xE2C2, 0xE940, 0xE947,
1067 	0xE950, 0xE987, 0xE9A0, 0xE9A9, 0xE9B0, 0xE9B6, 0xE9C0, 0xE9C7,
1068 	0xE9D0, 0xE9D1, 0xEA00, 0xEA01, 0xEA10, 0xEA1C, 0xEA40, 0xEA68,
1069 	0xE800, 0xE806, 0xE810, 0xE89A, 0xE8A0, 0xE8A4, 0xE8AA, 0xE8EB,
1070 	0xE900, 0xE905, 0xEB80, 0xEB8F, 0xEBB0, 0xEBB0, 0xEC00, 0xEC05,
1071 	0xEC08, 0xECE9, 0xECF0, 0xECF0, 0xEA80, 0xEA80, 0xEA82, 0xEAA3,
1072 	0xEAA5, 0xEAC2, 0xA800, 0xA8FF, 0xAC60, 0xAC60, 0xB000, 0xB97F,
1073 	0xB9A0, 0xB9BF, ~0
1074 };
1075 
1076 static void a5xx_dump(struct msm_gpu *gpu)
1077 {
1078 	dev_info(gpu->dev->dev, "status:   %08x\n",
1079 		gpu_read(gpu, REG_A5XX_RBBM_STATUS));
1080 	adreno_dump(gpu);
1081 }
1082 
1083 static int a5xx_pm_resume(struct msm_gpu *gpu)
1084 {
1085 	int ret;
1086 
1087 	/* Turn on the core power */
1088 	ret = msm_gpu_pm_resume(gpu);
1089 	if (ret)
1090 		return ret;
1091 
1092 	/* Turn the RBCCU domain first to limit the chances of voltage droop */
1093 	gpu_write(gpu, REG_A5XX_GPMU_RBCCU_POWER_CNTL, 0x778000);
1094 
1095 	/* Wait 3 usecs before polling */
1096 	udelay(3);
1097 
1098 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS,
1099 		(1 << 20), (1 << 20));
1100 	if (ret) {
1101 		DRM_ERROR("%s: timeout waiting for RBCCU GDSC enable: %X\n",
1102 			gpu->name,
1103 			gpu_read(gpu, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS));
1104 		return ret;
1105 	}
1106 
1107 	/* Turn on the SP domain */
1108 	gpu_write(gpu, REG_A5XX_GPMU_SP_POWER_CNTL, 0x778000);
1109 	ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_SP_PWR_CLK_STATUS,
1110 		(1 << 20), (1 << 20));
1111 	if (ret)
1112 		DRM_ERROR("%s: timeout waiting for SP GDSC enable\n",
1113 			gpu->name);
1114 
1115 	return ret;
1116 }
1117 
1118 static int a5xx_pm_suspend(struct msm_gpu *gpu)
1119 {
1120 	/* Clear the VBIF pipe before shutting down */
1121 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0xF);
1122 	spin_until((gpu_read(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL1) & 0xF) == 0xF);
1123 
1124 	gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0);
1125 
1126 	/*
1127 	 * Reset the VBIF before power collapse to avoid issue with FIFO
1128 	 * entries
1129 	 */
1130 	gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x003C0000);
1131 	gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x00000000);
1132 
1133 	return msm_gpu_pm_suspend(gpu);
1134 }
1135 
1136 static int a5xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value)
1137 {
1138 	*value = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_CP_0_LO,
1139 		REG_A5XX_RBBM_PERFCTR_CP_0_HI);
1140 
1141 	return 0;
1142 }
1143 
1144 #ifdef CONFIG_DEBUG_FS
1145 static void a5xx_show(struct msm_gpu *gpu, struct seq_file *m)
1146 {
1147 	seq_printf(m, "status:   %08x\n",
1148 			gpu_read(gpu, REG_A5XX_RBBM_STATUS));
1149 
1150 	/*
1151 	 * Temporarily disable hardware clock gating before going into
1152 	 * adreno_show to avoid issues while reading the registers
1153 	 */
1154 	a5xx_set_hwcg(gpu, false);
1155 	adreno_show(gpu, m);
1156 	a5xx_set_hwcg(gpu, true);
1157 }
1158 #endif
1159 
1160 static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu)
1161 {
1162 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
1163 	struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu);
1164 
1165 	return a5xx_gpu->cur_ring;
1166 }
1167 
1168 static const struct adreno_gpu_funcs funcs = {
1169 	.base = {
1170 		.get_param = adreno_get_param,
1171 		.hw_init = a5xx_hw_init,
1172 		.pm_suspend = a5xx_pm_suspend,
1173 		.pm_resume = a5xx_pm_resume,
1174 		.recover = a5xx_recover,
1175 		.submit = a5xx_submit,
1176 		.flush = a5xx_flush,
1177 		.active_ring = a5xx_active_ring,
1178 		.irq = a5xx_irq,
1179 		.destroy = a5xx_destroy,
1180 #ifdef CONFIG_DEBUG_FS
1181 		.show = a5xx_show,
1182 #endif
1183 	},
1184 	.get_timestamp = a5xx_get_timestamp,
1185 };
1186 
1187 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev)
1188 {
1189 	struct msm_drm_private *priv = dev->dev_private;
1190 	struct platform_device *pdev = priv->gpu_pdev;
1191 	struct a5xx_gpu *a5xx_gpu = NULL;
1192 	struct adreno_gpu *adreno_gpu;
1193 	struct msm_gpu *gpu;
1194 	int ret;
1195 
1196 	if (!pdev) {
1197 		dev_err(dev->dev, "No A5XX device is defined\n");
1198 		return ERR_PTR(-ENXIO);
1199 	}
1200 
1201 	a5xx_gpu = kzalloc(sizeof(*a5xx_gpu), GFP_KERNEL);
1202 	if (!a5xx_gpu)
1203 		return ERR_PTR(-ENOMEM);
1204 
1205 	adreno_gpu = &a5xx_gpu->base;
1206 	gpu = &adreno_gpu->base;
1207 
1208 	adreno_gpu->registers = a5xx_registers;
1209 	adreno_gpu->reg_offsets = a5xx_register_offsets;
1210 
1211 	a5xx_gpu->lm_leakage = 0x4E001A;
1212 
1213 	ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, 4);
1214 	if (ret) {
1215 		a5xx_destroy(&(a5xx_gpu->base.base));
1216 		return ERR_PTR(ret);
1217 	}
1218 
1219 	if (gpu->aspace)
1220 		msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu, a5xx_fault_handler);
1221 
1222 	/* Set up the preemption specific bits and pieces for each ringbuffer */
1223 	a5xx_preempt_init(gpu);
1224 
1225 	return gpu;
1226 }
1227