1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 as published by
7  * the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * You should have received a copy of the GNU General Public License along with
15  * this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #include "adreno_gpu.h"
19 #include "msm_gem.h"
20 
21 struct adreno_info {
22 	struct adreno_rev rev;
23 	uint32_t revn;
24 	const char *name;
25 	const char *pm4fw, *pfpfw;
26 	uint32_t gmem;
27 };
28 
29 #define ANY_ID 0xff
30 
31 static const struct adreno_info gpulist[] = {
32 	{
33 		.rev   = ADRENO_REV(3, 0, 5, ANY_ID),
34 		.revn  = 305,
35 		.name  = "A305",
36 		.pm4fw = "a300_pm4.fw",
37 		.pfpfw = "a300_pfp.fw",
38 		.gmem  = SZ_256K,
39 	}, {
40 		.rev   = ADRENO_REV(3, 2, ANY_ID, ANY_ID),
41 		.revn  = 320,
42 		.name  = "A320",
43 		.pm4fw = "a300_pm4.fw",
44 		.pfpfw = "a300_pfp.fw",
45 		.gmem  = SZ_512K,
46 	}, {
47 		.rev   = ADRENO_REV(3, 3, 0, 0),
48 		.revn  = 330,
49 		.name  = "A330",
50 		.pm4fw = "a330_pm4.fw",
51 		.pfpfw = "a330_pfp.fw",
52 		.gmem  = SZ_1M,
53 	},
54 };
55 
56 #define RB_SIZE    SZ_32K
57 #define RB_BLKSIZE 16
58 
59 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
60 {
61 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
62 
63 	switch (param) {
64 	case MSM_PARAM_GPU_ID:
65 		*value = adreno_gpu->info->revn;
66 		return 0;
67 	case MSM_PARAM_GMEM_SIZE:
68 		*value = adreno_gpu->info->gmem;
69 		return 0;
70 	default:
71 		DBG("%s: invalid param: %u", gpu->name, param);
72 		return -EINVAL;
73 	}
74 }
75 
76 #define rbmemptr(adreno_gpu, member)  \
77 	((adreno_gpu)->memptrs_iova + offsetof(struct adreno_rbmemptrs, member))
78 
79 int adreno_hw_init(struct msm_gpu *gpu)
80 {
81 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
82 
83 	DBG("%s", gpu->name);
84 
85 	/* Setup REG_CP_RB_CNTL: */
86 	gpu_write(gpu, REG_AXXX_CP_RB_CNTL,
87 			/* size is log2(quad-words): */
88 			AXXX_CP_RB_CNTL_BUFSZ(ilog2(gpu->rb->size / 8)) |
89 			AXXX_CP_RB_CNTL_BLKSZ(RB_BLKSIZE));
90 
91 	/* Setup ringbuffer address: */
92 	gpu_write(gpu, REG_AXXX_CP_RB_BASE, gpu->rb_iova);
93 	gpu_write(gpu, REG_AXXX_CP_RB_RPTR_ADDR, rbmemptr(adreno_gpu, rptr));
94 
95 	/* Setup scratch/timestamp: */
96 	gpu_write(gpu, REG_AXXX_SCRATCH_ADDR, rbmemptr(adreno_gpu, fence));
97 
98 	gpu_write(gpu, REG_AXXX_SCRATCH_UMSK, 0x1);
99 
100 	return 0;
101 }
102 
103 static uint32_t get_wptr(struct msm_ringbuffer *ring)
104 {
105 	return ring->cur - ring->start;
106 }
107 
108 uint32_t adreno_last_fence(struct msm_gpu *gpu)
109 {
110 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
111 	return adreno_gpu->memptrs->fence;
112 }
113 
114 void adreno_recover(struct msm_gpu *gpu)
115 {
116 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
117 	struct drm_device *dev = gpu->dev;
118 	int ret;
119 
120 	gpu->funcs->pm_suspend(gpu);
121 
122 	/* reset ringbuffer: */
123 	gpu->rb->cur = gpu->rb->start;
124 
125 	/* reset completed fence seqno, just discard anything pending: */
126 	adreno_gpu->memptrs->fence = gpu->submitted_fence;
127 	adreno_gpu->memptrs->rptr  = 0;
128 	adreno_gpu->memptrs->wptr  = 0;
129 
130 	gpu->funcs->pm_resume(gpu);
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 }
137 
138 int adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
139 		struct msm_file_private *ctx)
140 {
141 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
142 	struct msm_drm_private *priv = gpu->dev->dev_private;
143 	struct msm_ringbuffer *ring = gpu->rb;
144 	unsigned i, ibs = 0;
145 
146 	for (i = 0; i < submit->nr_cmds; i++) {
147 		switch (submit->cmd[i].type) {
148 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
149 			/* ignore IB-targets */
150 			break;
151 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
152 			/* ignore if there has not been a ctx switch: */
153 			if (priv->lastctx == ctx)
154 				break;
155 		case MSM_SUBMIT_CMD_BUF:
156 			OUT_PKT3(ring, CP_INDIRECT_BUFFER_PFD, 2);
157 			OUT_RING(ring, submit->cmd[i].iova);
158 			OUT_RING(ring, submit->cmd[i].size);
159 			ibs++;
160 			break;
161 		}
162 	}
163 
164 	/* on a320, at least, we seem to need to pad things out to an
165 	 * even number of qwords to avoid issue w/ CP hanging on wrap-
166 	 * around:
167 	 */
168 	if (ibs % 2)
169 		OUT_PKT2(ring);
170 
171 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
172 	OUT_RING(ring, submit->fence);
173 
174 	if (adreno_is_a3xx(adreno_gpu)) {
175 		/* Flush HLSQ lazy updates to make sure there is nothing
176 		 * pending for indirect loads after the timestamp has
177 		 * passed:
178 		 */
179 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
180 		OUT_RING(ring, HLSQ_FLUSH);
181 
182 		OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
183 		OUT_RING(ring, 0x00000000);
184 	}
185 
186 	OUT_PKT3(ring, CP_EVENT_WRITE, 3);
187 	OUT_RING(ring, CACHE_FLUSH_TS);
188 	OUT_RING(ring, rbmemptr(adreno_gpu, fence));
189 	OUT_RING(ring, submit->fence);
190 
191 	/* we could maybe be clever and only CP_COND_EXEC the interrupt: */
192 	OUT_PKT3(ring, CP_INTERRUPT, 1);
193 	OUT_RING(ring, 0x80000000);
194 
195 #if 0
196 	if (adreno_is_a3xx(adreno_gpu)) {
197 		/* Dummy set-constant to trigger context rollover */
198 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
199 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
200 		OUT_RING(ring, 0x00000000);
201 	}
202 #endif
203 
204 	gpu->funcs->flush(gpu);
205 
206 	return 0;
207 }
208 
209 void adreno_flush(struct msm_gpu *gpu)
210 {
211 	uint32_t wptr = get_wptr(gpu->rb);
212 
213 	/* ensure writes to ringbuffer have hit system memory: */
214 	mb();
215 
216 	gpu_write(gpu, REG_AXXX_CP_RB_WPTR, wptr);
217 }
218 
219 void adreno_idle(struct msm_gpu *gpu)
220 {
221 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
222 	uint32_t rptr, wptr = get_wptr(gpu->rb);
223 	unsigned long t;
224 
225 	t = jiffies + ADRENO_IDLE_TIMEOUT;
226 
227 	/* then wait for CP to drain ringbuffer: */
228 	do {
229 		rptr = adreno_gpu->memptrs->rptr;
230 		if (rptr == wptr)
231 			return;
232 	} while(time_before(jiffies, t));
233 
234 	DRM_ERROR("%s: timeout waiting to drain ringbuffer!\n", gpu->name);
235 
236 	/* TODO maybe we need to reset GPU here to recover from hang? */
237 }
238 
239 #ifdef CONFIG_DEBUG_FS
240 void adreno_show(struct msm_gpu *gpu, struct seq_file *m)
241 {
242 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
243 
244 	seq_printf(m, "revision: %d (%d.%d.%d.%d)\n",
245 			adreno_gpu->info->revn, adreno_gpu->rev.core,
246 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
247 			adreno_gpu->rev.patchid);
248 
249 	seq_printf(m, "fence:    %d/%d\n", adreno_gpu->memptrs->fence,
250 			gpu->submitted_fence);
251 	seq_printf(m, "rptr:     %d\n", adreno_gpu->memptrs->rptr);
252 	seq_printf(m, "wptr:     %d\n", adreno_gpu->memptrs->wptr);
253 	seq_printf(m, "rb wptr:  %d\n", get_wptr(gpu->rb));
254 }
255 #endif
256 
257 void adreno_wait_ring(struct msm_gpu *gpu, uint32_t ndwords)
258 {
259 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
260 	uint32_t freedwords;
261 	unsigned long t = jiffies + ADRENO_IDLE_TIMEOUT;
262 	do {
263 		uint32_t size = gpu->rb->size / 4;
264 		uint32_t wptr = get_wptr(gpu->rb);
265 		uint32_t rptr = adreno_gpu->memptrs->rptr;
266 		freedwords = (rptr + (size - 1) - wptr) % size;
267 
268 		if (time_after(jiffies, t)) {
269 			DRM_ERROR("%s: timeout waiting for ringbuffer space\n", gpu->name);
270 			break;
271 		}
272 	} while(freedwords < ndwords);
273 }
274 
275 static const char *iommu_ports[] = {
276 		"gfx3d_user", "gfx3d_priv",
277 		"gfx3d1_user", "gfx3d1_priv",
278 };
279 
280 static inline bool _rev_match(uint8_t entry, uint8_t id)
281 {
282 	return (entry == ANY_ID) || (entry == id);
283 }
284 
285 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
286 		struct adreno_gpu *gpu, const struct adreno_gpu_funcs *funcs,
287 		struct adreno_rev rev)
288 {
289 	int i, ret;
290 
291 	/* identify gpu: */
292 	for (i = 0; i < ARRAY_SIZE(gpulist); i++) {
293 		const struct adreno_info *info = &gpulist[i];
294 		if (_rev_match(info->rev.core, rev.core) &&
295 				_rev_match(info->rev.major, rev.major) &&
296 				_rev_match(info->rev.minor, rev.minor) &&
297 				_rev_match(info->rev.patchid, rev.patchid)) {
298 			gpu->info = info;
299 			gpu->revn = info->revn;
300 			break;
301 		}
302 	}
303 
304 	if (i == ARRAY_SIZE(gpulist)) {
305 		dev_err(drm->dev, "Unknown GPU revision: %u.%u.%u.%u\n",
306 				rev.core, rev.major, rev.minor, rev.patchid);
307 		return -ENXIO;
308 	}
309 
310 	DBG("Found GPU: %s (%u.%u.%u.%u)", gpu->info->name,
311 			rev.core, rev.major, rev.minor, rev.patchid);
312 
313 	gpu->funcs = funcs;
314 	gpu->rev = rev;
315 
316 	ret = request_firmware(&gpu->pm4, gpu->info->pm4fw, drm->dev);
317 	if (ret) {
318 		dev_err(drm->dev, "failed to load %s PM4 firmware: %d\n",
319 				gpu->info->pm4fw, ret);
320 		return ret;
321 	}
322 
323 	ret = request_firmware(&gpu->pfp, gpu->info->pfpfw, drm->dev);
324 	if (ret) {
325 		dev_err(drm->dev, "failed to load %s PFP firmware: %d\n",
326 				gpu->info->pfpfw, ret);
327 		return ret;
328 	}
329 
330 	ret = msm_gpu_init(drm, pdev, &gpu->base, &funcs->base,
331 			gpu->info->name, "kgsl_3d0_reg_memory", "kgsl_3d0_irq",
332 			RB_SIZE);
333 	if (ret)
334 		return ret;
335 
336 	ret = msm_iommu_attach(drm, gpu->base.iommu,
337 			iommu_ports, ARRAY_SIZE(iommu_ports));
338 	if (ret)
339 		return ret;
340 
341 	gpu->memptrs_bo = msm_gem_new(drm, sizeof(*gpu->memptrs),
342 			MSM_BO_UNCACHED);
343 	if (IS_ERR(gpu->memptrs_bo)) {
344 		ret = PTR_ERR(gpu->memptrs_bo);
345 		gpu->memptrs_bo = NULL;
346 		dev_err(drm->dev, "could not allocate memptrs: %d\n", ret);
347 		return ret;
348 	}
349 
350 	gpu->memptrs = msm_gem_vaddr_locked(gpu->memptrs_bo);
351 	if (!gpu->memptrs) {
352 		dev_err(drm->dev, "could not vmap memptrs\n");
353 		return -ENOMEM;
354 	}
355 
356 	ret = msm_gem_get_iova_locked(gpu->memptrs_bo, gpu->base.id,
357 			&gpu->memptrs_iova);
358 	if (ret) {
359 		dev_err(drm->dev, "could not map memptrs: %d\n", ret);
360 		return ret;
361 	}
362 
363 	return 0;
364 }
365 
366 void adreno_gpu_cleanup(struct adreno_gpu *gpu)
367 {
368 	if (gpu->memptrs_bo) {
369 		if (gpu->memptrs_iova)
370 			msm_gem_put_iova(gpu->memptrs_bo, gpu->base.id);
371 		drm_gem_object_unreference(gpu->memptrs_bo);
372 	}
373 	if (gpu->pm4)
374 		release_firmware(gpu->pm4);
375 	if (gpu->pfp)
376 		release_firmware(gpu->pfp);
377 	msm_gpu_cleanup(&gpu->base);
378 }
379