1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License version 2 as published by
9  * the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful, but WITHOUT
12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
14  * more details.
15  *
16  * You should have received a copy of the GNU General Public License along with
17  * this program.  If not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include "adreno_gpu.h"
21 #include "msm_gem.h"
22 #include "msm_mmu.h"
23 
24 #define RB_SIZE    SZ_32K
25 #define RB_BLKSIZE 32
26 
27 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
28 {
29 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
30 
31 	switch (param) {
32 	case MSM_PARAM_GPU_ID:
33 		*value = adreno_gpu->info->revn;
34 		return 0;
35 	case MSM_PARAM_GMEM_SIZE:
36 		*value = adreno_gpu->gmem;
37 		return 0;
38 	case MSM_PARAM_CHIP_ID:
39 		*value = adreno_gpu->rev.patchid |
40 				(adreno_gpu->rev.minor << 8) |
41 				(adreno_gpu->rev.major << 16) |
42 				(adreno_gpu->rev.core << 24);
43 		return 0;
44 	case MSM_PARAM_MAX_FREQ:
45 		*value = adreno_gpu->base.fast_rate;
46 		return 0;
47 	case MSM_PARAM_TIMESTAMP:
48 		if (adreno_gpu->funcs->get_timestamp)
49 			return adreno_gpu->funcs->get_timestamp(gpu, value);
50 		return -EINVAL;
51 	default:
52 		DBG("%s: invalid param: %u", gpu->name, param);
53 		return -EINVAL;
54 	}
55 }
56 
57 int adreno_hw_init(struct msm_gpu *gpu)
58 {
59 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
60 	int ret;
61 
62 	DBG("%s", gpu->name);
63 
64 	ret = msm_gem_get_iova(gpu->rb->bo, gpu->id, &gpu->rb_iova);
65 	if (ret) {
66 		gpu->rb_iova = 0;
67 		dev_err(gpu->dev->dev, "could not map ringbuffer: %d\n", ret);
68 		return ret;
69 	}
70 
71 	/* Setup REG_CP_RB_CNTL: */
72 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL,
73 			/* size is log2(quad-words): */
74 			AXXX_CP_RB_CNTL_BUFSZ(ilog2(gpu->rb->size / 8)) |
75 			AXXX_CP_RB_CNTL_BLKSZ(ilog2(RB_BLKSIZE / 8)) |
76 			(adreno_is_a430(adreno_gpu) ? AXXX_CP_RB_CNTL_NO_UPDATE : 0));
77 
78 	/* Setup ringbuffer address: */
79 	adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_BASE,
80 		REG_ADRENO_CP_RB_BASE_HI, gpu->rb_iova);
81 
82 	if (!adreno_is_a430(adreno_gpu)) {
83 		adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR,
84 			REG_ADRENO_CP_RB_RPTR_ADDR_HI,
85 			rbmemptr(adreno_gpu, rptr));
86 	}
87 
88 	return 0;
89 }
90 
91 static uint32_t get_wptr(struct msm_ringbuffer *ring)
92 {
93 	return ring->cur - ring->start;
94 }
95 
96 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
97 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu)
98 {
99 	if (adreno_is_a430(adreno_gpu))
100 		return adreno_gpu->memptrs->rptr = adreno_gpu_read(
101 			adreno_gpu, REG_ADRENO_CP_RB_RPTR);
102 	else
103 		return adreno_gpu->memptrs->rptr;
104 }
105 
106 uint32_t adreno_last_fence(struct msm_gpu *gpu)
107 {
108 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
109 	return adreno_gpu->memptrs->fence;
110 }
111 
112 void adreno_recover(struct msm_gpu *gpu)
113 {
114 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
115 	struct drm_device *dev = gpu->dev;
116 	int ret;
117 
118 	gpu->funcs->pm_suspend(gpu);
119 
120 	/* reset ringbuffer: */
121 	gpu->rb->cur = gpu->rb->start;
122 
123 	/* reset completed fence seqno: */
124 	adreno_gpu->memptrs->fence = gpu->fctx->completed_fence;
125 	adreno_gpu->memptrs->rptr  = 0;
126 	adreno_gpu->memptrs->wptr  = 0;
127 
128 	gpu->funcs->pm_resume(gpu);
129 
130 	disable_irq(gpu->irq);
131 	ret = gpu->funcs->hw_init(gpu);
132 	if (ret) {
133 		dev_err(dev->dev, "gpu hw init failed: %d\n", ret);
134 		/* hmm, oh well? */
135 	}
136 	enable_irq(gpu->irq);
137 }
138 
139 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
140 		struct msm_file_private *ctx)
141 {
142 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
143 	struct msm_drm_private *priv = gpu->dev->dev_private;
144 	struct msm_ringbuffer *ring = gpu->rb;
145 	unsigned i;
146 
147 	for (i = 0; i < submit->nr_cmds; i++) {
148 		switch (submit->cmd[i].type) {
149 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
150 			/* ignore IB-targets */
151 			break;
152 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
153 			/* ignore if there has not been a ctx switch: */
154 			if (priv->lastctx == ctx)
155 				break;
156 		case MSM_SUBMIT_CMD_BUF:
157 			OUT_PKT3(ring, adreno_is_a430(adreno_gpu) ?
158 				CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
159 			OUT_RING(ring, submit->cmd[i].iova);
160 			OUT_RING(ring, submit->cmd[i].size);
161 			OUT_PKT2(ring);
162 			break;
163 		}
164 	}
165 
166 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
167 	OUT_RING(ring, submit->fence->seqno);
168 
169 	if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) {
170 		/* Flush HLSQ lazy updates to make sure there is nothing
171 		 * pending for indirect loads after the timestamp has
172 		 * passed:
173 		 */
174 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
175 		OUT_RING(ring, HLSQ_FLUSH);
176 
177 		OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
178 		OUT_RING(ring, 0x00000000);
179 	}
180 
181 	OUT_PKT3(ring, CP_EVENT_WRITE, 3);
182 	OUT_RING(ring, CACHE_FLUSH_TS);
183 	OUT_RING(ring, rbmemptr(adreno_gpu, fence));
184 	OUT_RING(ring, submit->fence->seqno);
185 
186 	/* we could maybe be clever and only CP_COND_EXEC the interrupt: */
187 	OUT_PKT3(ring, CP_INTERRUPT, 1);
188 	OUT_RING(ring, 0x80000000);
189 
190 	/* Workaround for missing irq issue on 8x16/a306.  Unsure if the
191 	 * root cause is a platform issue or some a306 quirk, but this
192 	 * keeps things humming along:
193 	 */
194 	if (adreno_is_a306(adreno_gpu)) {
195 		OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
196 		OUT_RING(ring, 0x00000000);
197 		OUT_PKT3(ring, CP_INTERRUPT, 1);
198 		OUT_RING(ring, 0x80000000);
199 	}
200 
201 #if 0
202 	if (adreno_is_a3xx(adreno_gpu)) {
203 		/* Dummy set-constant to trigger context rollover */
204 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
205 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
206 		OUT_RING(ring, 0x00000000);
207 	}
208 #endif
209 
210 	gpu->funcs->flush(gpu);
211 }
212 
213 void adreno_flush(struct msm_gpu *gpu)
214 {
215 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
216 	uint32_t wptr;
217 
218 	/*
219 	 * Mask wptr value that we calculate to fit in the HW range. This is
220 	 * to account for the possibility that the last command fit exactly into
221 	 * the ringbuffer and rb->next hasn't wrapped to zero yet
222 	 */
223 	wptr = get_wptr(gpu->rb) & ((gpu->rb->size / 4) - 1);
224 
225 	/* ensure writes to ringbuffer have hit system memory: */
226 	mb();
227 
228 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr);
229 }
230 
231 bool adreno_idle(struct msm_gpu *gpu)
232 {
233 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
234 	uint32_t wptr = get_wptr(gpu->rb);
235 
236 	/* wait for CP to drain ringbuffer: */
237 	if (!spin_until(get_rptr(adreno_gpu) == wptr))
238 		return true;
239 
240 	/* TODO maybe we need to reset GPU here to recover from hang? */
241 	DRM_ERROR("%s: timeout waiting to drain ringbuffer!\n", gpu->name);
242 	return false;
243 }
244 
245 #ifdef CONFIG_DEBUG_FS
246 void adreno_show(struct msm_gpu *gpu, struct seq_file *m)
247 {
248 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
249 	int i;
250 
251 	seq_printf(m, "revision: %d (%d.%d.%d.%d)\n",
252 			adreno_gpu->info->revn, adreno_gpu->rev.core,
253 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
254 			adreno_gpu->rev.patchid);
255 
256 	seq_printf(m, "fence:    %d/%d\n", adreno_gpu->memptrs->fence,
257 			gpu->fctx->last_fence);
258 	seq_printf(m, "rptr:     %d\n", get_rptr(adreno_gpu));
259 	seq_printf(m, "wptr:     %d\n", adreno_gpu->memptrs->wptr);
260 	seq_printf(m, "rb wptr:  %d\n", get_wptr(gpu->rb));
261 
262 	gpu->funcs->pm_resume(gpu);
263 
264 	/* dump these out in a form that can be parsed by demsm: */
265 	seq_printf(m, "IO:region %s 00000000 00020000\n", gpu->name);
266 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
267 		uint32_t start = adreno_gpu->registers[i];
268 		uint32_t end   = adreno_gpu->registers[i+1];
269 		uint32_t addr;
270 
271 		for (addr = start; addr <= end; addr++) {
272 			uint32_t val = gpu_read(gpu, addr);
273 			seq_printf(m, "IO:R %08x %08x\n", addr<<2, val);
274 		}
275 	}
276 
277 	gpu->funcs->pm_suspend(gpu);
278 }
279 #endif
280 
281 /* Dump common gpu status and scratch registers on any hang, to make
282  * the hangcheck logs more useful.  The scratch registers seem always
283  * safe to read when GPU has hung (unlike some other regs, depending
284  * on how the GPU hung), and they are useful to match up to cmdstream
285  * dumps when debugging hangs:
286  */
287 void adreno_dump_info(struct msm_gpu *gpu)
288 {
289 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
290 
291 	printk("revision: %d (%d.%d.%d.%d)\n",
292 			adreno_gpu->info->revn, adreno_gpu->rev.core,
293 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
294 			adreno_gpu->rev.patchid);
295 
296 	printk("fence:    %d/%d\n", adreno_gpu->memptrs->fence,
297 			gpu->fctx->last_fence);
298 	printk("rptr:     %d\n", get_rptr(adreno_gpu));
299 	printk("wptr:     %d\n", adreno_gpu->memptrs->wptr);
300 	printk("rb wptr:  %d\n", get_wptr(gpu->rb));
301 }
302 
303 /* would be nice to not have to duplicate the _show() stuff with printk(): */
304 void adreno_dump(struct msm_gpu *gpu)
305 {
306 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
307 	int i;
308 
309 	/* dump these out in a form that can be parsed by demsm: */
310 	printk("IO:region %s 00000000 00020000\n", gpu->name);
311 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
312 		uint32_t start = adreno_gpu->registers[i];
313 		uint32_t end   = adreno_gpu->registers[i+1];
314 		uint32_t addr;
315 
316 		for (addr = start; addr <= end; addr++) {
317 			uint32_t val = gpu_read(gpu, addr);
318 			printk("IO:R %08x %08x\n", addr<<2, val);
319 		}
320 	}
321 }
322 
323 static uint32_t ring_freewords(struct msm_gpu *gpu)
324 {
325 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
326 	uint32_t size = gpu->rb->size / 4;
327 	uint32_t wptr = get_wptr(gpu->rb);
328 	uint32_t rptr = get_rptr(adreno_gpu);
329 	return (rptr + (size - 1) - wptr) % size;
330 }
331 
332 void adreno_wait_ring(struct msm_gpu *gpu, uint32_t ndwords)
333 {
334 	if (spin_until(ring_freewords(gpu) >= ndwords))
335 		DRM_ERROR("%s: timeout waiting for ringbuffer space\n", gpu->name);
336 }
337 
338 static const char *iommu_ports[] = {
339 		"gfx3d_user", "gfx3d_priv",
340 		"gfx3d1_user", "gfx3d1_priv",
341 };
342 
343 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
344 		struct adreno_gpu *adreno_gpu, const struct adreno_gpu_funcs *funcs)
345 {
346 	struct adreno_platform_config *config = pdev->dev.platform_data;
347 	struct msm_gpu *gpu = &adreno_gpu->base;
348 	struct msm_mmu *mmu;
349 	int ret;
350 
351 	adreno_gpu->funcs = funcs;
352 	adreno_gpu->info = adreno_info(config->rev);
353 	adreno_gpu->gmem = adreno_gpu->info->gmem;
354 	adreno_gpu->revn = adreno_gpu->info->revn;
355 	adreno_gpu->rev = config->rev;
356 	adreno_gpu->quirks = config->quirks;
357 
358 	gpu->fast_rate = config->fast_rate;
359 	gpu->slow_rate = config->slow_rate;
360 	gpu->bus_freq  = config->bus_freq;
361 #ifdef DOWNSTREAM_CONFIG_MSM_BUS_SCALING
362 	gpu->bus_scale_table = config->bus_scale_table;
363 #endif
364 
365 	DBG("fast_rate=%u, slow_rate=%u, bus_freq=%u",
366 			gpu->fast_rate, gpu->slow_rate, gpu->bus_freq);
367 
368 	ret = msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
369 			adreno_gpu->info->name, "kgsl_3d0_reg_memory", "kgsl_3d0_irq",
370 			RB_SIZE);
371 	if (ret)
372 		return ret;
373 
374 	ret = request_firmware(&adreno_gpu->pm4, adreno_gpu->info->pm4fw, drm->dev);
375 	if (ret) {
376 		dev_err(drm->dev, "failed to load %s PM4 firmware: %d\n",
377 				adreno_gpu->info->pm4fw, ret);
378 		return ret;
379 	}
380 
381 	ret = request_firmware(&adreno_gpu->pfp, adreno_gpu->info->pfpfw, drm->dev);
382 	if (ret) {
383 		dev_err(drm->dev, "failed to load %s PFP firmware: %d\n",
384 				adreno_gpu->info->pfpfw, ret);
385 		return ret;
386 	}
387 
388 	mmu = gpu->aspace->mmu;
389 	if (mmu) {
390 		ret = mmu->funcs->attach(mmu, iommu_ports,
391 				ARRAY_SIZE(iommu_ports));
392 		if (ret)
393 			return ret;
394 	}
395 
396 	mutex_lock(&drm->struct_mutex);
397 	adreno_gpu->memptrs_bo = msm_gem_new(drm, sizeof(*adreno_gpu->memptrs),
398 			MSM_BO_UNCACHED);
399 	mutex_unlock(&drm->struct_mutex);
400 	if (IS_ERR(adreno_gpu->memptrs_bo)) {
401 		ret = PTR_ERR(adreno_gpu->memptrs_bo);
402 		adreno_gpu->memptrs_bo = NULL;
403 		dev_err(drm->dev, "could not allocate memptrs: %d\n", ret);
404 		return ret;
405 	}
406 
407 	adreno_gpu->memptrs = msm_gem_get_vaddr(adreno_gpu->memptrs_bo);
408 	if (IS_ERR(adreno_gpu->memptrs)) {
409 		dev_err(drm->dev, "could not vmap memptrs\n");
410 		return -ENOMEM;
411 	}
412 
413 	ret = msm_gem_get_iova(adreno_gpu->memptrs_bo, gpu->id,
414 			&adreno_gpu->memptrs_iova);
415 	if (ret) {
416 		dev_err(drm->dev, "could not map memptrs: %d\n", ret);
417 		return ret;
418 	}
419 
420 	return 0;
421 }
422 
423 void adreno_gpu_cleanup(struct adreno_gpu *gpu)
424 {
425 	if (gpu->memptrs_bo) {
426 		if (gpu->memptrs)
427 			msm_gem_put_vaddr(gpu->memptrs_bo);
428 
429 		if (gpu->memptrs_iova)
430 			msm_gem_put_iova(gpu->memptrs_bo, gpu->base.id);
431 
432 		drm_gem_object_unreference_unlocked(gpu->memptrs_bo);
433 	}
434 	release_firmware(gpu->pm4);
435 	release_firmware(gpu->pfp);
436 	msm_gpu_cleanup(&gpu->base);
437 }
438