xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision b34081f1)
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 "msm_gpu.h"
19 #include "msm_gem.h"
20 
21 
22 /*
23  * Power Management:
24  */
25 
26 #ifdef CONFIG_MSM_BUS_SCALING
27 #include <mach/board.h>
28 #include <mach/kgsl.h>
29 static void bs_init(struct msm_gpu *gpu, struct platform_device *pdev)
30 {
31 	struct drm_device *dev = gpu->dev;
32 	struct kgsl_device_platform_data *pdata = pdev->dev.platform_data;
33 
34 	if (!pdev) {
35 		dev_err(dev->dev, "could not find dtv pdata\n");
36 		return;
37 	}
38 
39 	if (pdata->bus_scale_table) {
40 		gpu->bsc = msm_bus_scale_register_client(pdata->bus_scale_table);
41 		DBG("bus scale client: %08x", gpu->bsc);
42 	}
43 }
44 
45 static void bs_fini(struct msm_gpu *gpu)
46 {
47 	if (gpu->bsc) {
48 		msm_bus_scale_unregister_client(gpu->bsc);
49 		gpu->bsc = 0;
50 	}
51 }
52 
53 static void bs_set(struct msm_gpu *gpu, int idx)
54 {
55 	if (gpu->bsc) {
56 		DBG("set bus scaling: %d", idx);
57 		msm_bus_scale_client_update_request(gpu->bsc, idx);
58 	}
59 }
60 #else
61 static void bs_init(struct msm_gpu *gpu, struct platform_device *pdev) {}
62 static void bs_fini(struct msm_gpu *gpu) {}
63 static void bs_set(struct msm_gpu *gpu, int idx) {}
64 #endif
65 
66 static int enable_pwrrail(struct msm_gpu *gpu)
67 {
68 	struct drm_device *dev = gpu->dev;
69 	int ret = 0;
70 
71 	if (gpu->gpu_reg) {
72 		ret = regulator_enable(gpu->gpu_reg);
73 		if (ret) {
74 			dev_err(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
75 			return ret;
76 		}
77 	}
78 
79 	if (gpu->gpu_cx) {
80 		ret = regulator_enable(gpu->gpu_cx);
81 		if (ret) {
82 			dev_err(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
83 			return ret;
84 		}
85 	}
86 
87 	return 0;
88 }
89 
90 static int disable_pwrrail(struct msm_gpu *gpu)
91 {
92 	if (gpu->gpu_cx)
93 		regulator_disable(gpu->gpu_cx);
94 	if (gpu->gpu_reg)
95 		regulator_disable(gpu->gpu_reg);
96 	return 0;
97 }
98 
99 static int enable_clk(struct msm_gpu *gpu)
100 {
101 	struct clk *rate_clk = NULL;
102 	int i;
103 
104 	/* NOTE: kgsl_pwrctrl_clk() ignores grp_clks[0].. */
105 	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--) {
106 		if (gpu->grp_clks[i]) {
107 			clk_prepare(gpu->grp_clks[i]);
108 			rate_clk = gpu->grp_clks[i];
109 		}
110 	}
111 
112 	if (rate_clk && gpu->fast_rate)
113 		clk_set_rate(rate_clk, gpu->fast_rate);
114 
115 	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--)
116 		if (gpu->grp_clks[i])
117 			clk_enable(gpu->grp_clks[i]);
118 
119 	return 0;
120 }
121 
122 static int disable_clk(struct msm_gpu *gpu)
123 {
124 	struct clk *rate_clk = NULL;
125 	int i;
126 
127 	/* NOTE: kgsl_pwrctrl_clk() ignores grp_clks[0].. */
128 	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--) {
129 		if (gpu->grp_clks[i]) {
130 			clk_disable(gpu->grp_clks[i]);
131 			rate_clk = gpu->grp_clks[i];
132 		}
133 	}
134 
135 	if (rate_clk && gpu->slow_rate)
136 		clk_set_rate(rate_clk, gpu->slow_rate);
137 
138 	for (i = ARRAY_SIZE(gpu->grp_clks) - 1; i > 0; i--)
139 		if (gpu->grp_clks[i])
140 			clk_unprepare(gpu->grp_clks[i]);
141 
142 	return 0;
143 }
144 
145 static int enable_axi(struct msm_gpu *gpu)
146 {
147 	if (gpu->ebi1_clk)
148 		clk_prepare_enable(gpu->ebi1_clk);
149 	if (gpu->bus_freq)
150 		bs_set(gpu, gpu->bus_freq);
151 	return 0;
152 }
153 
154 static int disable_axi(struct msm_gpu *gpu)
155 {
156 	if (gpu->ebi1_clk)
157 		clk_disable_unprepare(gpu->ebi1_clk);
158 	if (gpu->bus_freq)
159 		bs_set(gpu, 0);
160 	return 0;
161 }
162 
163 int msm_gpu_pm_resume(struct msm_gpu *gpu)
164 {
165 	int ret;
166 
167 	DBG("%s", gpu->name);
168 
169 	ret = enable_pwrrail(gpu);
170 	if (ret)
171 		return ret;
172 
173 	ret = enable_clk(gpu);
174 	if (ret)
175 		return ret;
176 
177 	ret = enable_axi(gpu);
178 	if (ret)
179 		return ret;
180 
181 	return 0;
182 }
183 
184 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
185 {
186 	int ret;
187 
188 	DBG("%s", gpu->name);
189 
190 	ret = disable_axi(gpu);
191 	if (ret)
192 		return ret;
193 
194 	ret = disable_clk(gpu);
195 	if (ret)
196 		return ret;
197 
198 	ret = disable_pwrrail(gpu);
199 	if (ret)
200 		return ret;
201 
202 	return 0;
203 }
204 
205 /*
206  * Hangcheck detection for locked gpu:
207  */
208 
209 static void recover_worker(struct work_struct *work)
210 {
211 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
212 	struct drm_device *dev = gpu->dev;
213 
214 	dev_err(dev->dev, "%s: hangcheck recover!\n", gpu->name);
215 
216 	mutex_lock(&dev->struct_mutex);
217 	gpu->funcs->recover(gpu);
218 	mutex_unlock(&dev->struct_mutex);
219 
220 	msm_gpu_retire(gpu);
221 }
222 
223 static void hangcheck_timer_reset(struct msm_gpu *gpu)
224 {
225 	DBG("%s", gpu->name);
226 	mod_timer(&gpu->hangcheck_timer,
227 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
228 }
229 
230 static void hangcheck_handler(unsigned long data)
231 {
232 	struct msm_gpu *gpu = (struct msm_gpu *)data;
233 	uint32_t fence = gpu->funcs->last_fence(gpu);
234 
235 	if (fence != gpu->hangcheck_fence) {
236 		/* some progress has been made.. ya! */
237 		gpu->hangcheck_fence = fence;
238 	} else if (fence < gpu->submitted_fence) {
239 		/* no progress and not done.. hung! */
240 		struct msm_drm_private *priv = gpu->dev->dev_private;
241 		gpu->hangcheck_fence = fence;
242 		queue_work(priv->wq, &gpu->recover_work);
243 	}
244 
245 	/* if still more pending work, reset the hangcheck timer: */
246 	if (gpu->submitted_fence > gpu->hangcheck_fence)
247 		hangcheck_timer_reset(gpu);
248 }
249 
250 /*
251  * Cmdstream submission/retirement:
252  */
253 
254 static void retire_worker(struct work_struct *work)
255 {
256 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
257 	struct drm_device *dev = gpu->dev;
258 	uint32_t fence = gpu->funcs->last_fence(gpu);
259 
260 	mutex_lock(&dev->struct_mutex);
261 
262 	while (!list_empty(&gpu->active_list)) {
263 		struct msm_gem_object *obj;
264 
265 		obj = list_first_entry(&gpu->active_list,
266 				struct msm_gem_object, mm_list);
267 
268 		if (obj->fence <= fence) {
269 			/* move to inactive: */
270 			msm_gem_move_to_inactive(&obj->base);
271 			msm_gem_put_iova(&obj->base, gpu->id);
272 			drm_gem_object_unreference(&obj->base);
273 		} else {
274 			break;
275 		}
276 	}
277 
278 	msm_update_fence(gpu->dev, fence);
279 
280 	mutex_unlock(&dev->struct_mutex);
281 }
282 
283 /* call from irq handler to schedule work to retire bo's */
284 void msm_gpu_retire(struct msm_gpu *gpu)
285 {
286 	struct msm_drm_private *priv = gpu->dev->dev_private;
287 	queue_work(priv->wq, &gpu->retire_work);
288 }
289 
290 /* add bo's to gpu's ring, and kick gpu: */
291 int msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
292 		struct msm_file_private *ctx)
293 {
294 	struct drm_device *dev = gpu->dev;
295 	struct msm_drm_private *priv = dev->dev_private;
296 	int i, ret;
297 
298 	mutex_lock(&dev->struct_mutex);
299 
300 	submit->fence = ++priv->next_fence;
301 
302 	gpu->submitted_fence = submit->fence;
303 
304 	ret = gpu->funcs->submit(gpu, submit, ctx);
305 	priv->lastctx = ctx;
306 
307 	for (i = 0; i < submit->nr_bos; i++) {
308 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
309 
310 		/* can't happen yet.. but when we add 2d support we'll have
311 		 * to deal w/ cross-ring synchronization:
312 		 */
313 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
314 
315 		if (!is_active(msm_obj)) {
316 			uint32_t iova;
317 
318 			/* ring takes a reference to the bo and iova: */
319 			drm_gem_object_reference(&msm_obj->base);
320 			msm_gem_get_iova_locked(&msm_obj->base,
321 					submit->gpu->id, &iova);
322 		}
323 
324 		msm_gem_move_to_active(&msm_obj->base, gpu, submit->fence);
325 	}
326 	hangcheck_timer_reset(gpu);
327 	mutex_unlock(&dev->struct_mutex);
328 
329 	return ret;
330 }
331 
332 /*
333  * Init/Cleanup:
334  */
335 
336 static irqreturn_t irq_handler(int irq, void *data)
337 {
338 	struct msm_gpu *gpu = data;
339 	return gpu->funcs->irq(gpu);
340 }
341 
342 static const char *clk_names[] = {
343 		"src_clk", "core_clk", "iface_clk", "mem_clk", "mem_iface_clk",
344 };
345 
346 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
347 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
348 		const char *name, const char *ioname, const char *irqname, int ringsz)
349 {
350 	int i, ret;
351 
352 	gpu->dev = drm;
353 	gpu->funcs = funcs;
354 	gpu->name = name;
355 
356 	INIT_LIST_HEAD(&gpu->active_list);
357 	INIT_WORK(&gpu->retire_work, retire_worker);
358 	INIT_WORK(&gpu->recover_work, recover_worker);
359 
360 	setup_timer(&gpu->hangcheck_timer, hangcheck_handler,
361 			(unsigned long)gpu);
362 
363 	BUG_ON(ARRAY_SIZE(clk_names) != ARRAY_SIZE(gpu->grp_clks));
364 
365 	/* Map registers: */
366 	gpu->mmio = msm_ioremap(pdev, ioname, name);
367 	if (IS_ERR(gpu->mmio)) {
368 		ret = PTR_ERR(gpu->mmio);
369 		goto fail;
370 	}
371 
372 	/* Get Interrupt: */
373 	gpu->irq = platform_get_irq_byname(pdev, irqname);
374 	if (gpu->irq < 0) {
375 		ret = gpu->irq;
376 		dev_err(drm->dev, "failed to get irq: %d\n", ret);
377 		goto fail;
378 	}
379 
380 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
381 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
382 	if (ret) {
383 		dev_err(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
384 		goto fail;
385 	}
386 
387 	/* Acquire clocks: */
388 	for (i = 0; i < ARRAY_SIZE(clk_names); i++) {
389 		gpu->grp_clks[i] = devm_clk_get(&pdev->dev, clk_names[i]);
390 		DBG("grp_clks[%s]: %p", clk_names[i], gpu->grp_clks[i]);
391 		if (IS_ERR(gpu->grp_clks[i]))
392 			gpu->grp_clks[i] = NULL;
393 	}
394 
395 	gpu->ebi1_clk = devm_clk_get(&pdev->dev, "bus_clk");
396 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
397 	if (IS_ERR(gpu->ebi1_clk))
398 		gpu->ebi1_clk = NULL;
399 
400 	/* Acquire regulators: */
401 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
402 	DBG("gpu_reg: %p", gpu->gpu_reg);
403 	if (IS_ERR(gpu->gpu_reg))
404 		gpu->gpu_reg = NULL;
405 
406 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
407 	DBG("gpu_cx: %p", gpu->gpu_cx);
408 	if (IS_ERR(gpu->gpu_cx))
409 		gpu->gpu_cx = NULL;
410 
411 	/* Setup IOMMU.. eventually we will (I think) do this once per context
412 	 * and have separate page tables per context.  For now, to keep things
413 	 * simple and to get something working, just use a single address space:
414 	 */
415 	gpu->iommu = iommu_domain_alloc(&platform_bus_type);
416 	if (!gpu->iommu) {
417 		dev_err(drm->dev, "failed to allocate IOMMU\n");
418 		ret = -ENOMEM;
419 		goto fail;
420 	}
421 	gpu->id = msm_register_iommu(drm, gpu->iommu);
422 
423 	/* Create ringbuffer: */
424 	gpu->rb = msm_ringbuffer_new(gpu, ringsz);
425 	if (IS_ERR(gpu->rb)) {
426 		ret = PTR_ERR(gpu->rb);
427 		gpu->rb = NULL;
428 		dev_err(drm->dev, "could not create ringbuffer: %d\n", ret);
429 		goto fail;
430 	}
431 
432 	ret = msm_gem_get_iova_locked(gpu->rb->bo, gpu->id, &gpu->rb_iova);
433 	if (ret) {
434 		gpu->rb_iova = 0;
435 		dev_err(drm->dev, "could not map ringbuffer: %d\n", ret);
436 		goto fail;
437 	}
438 
439 	bs_init(gpu, pdev);
440 
441 	return 0;
442 
443 fail:
444 	return ret;
445 }
446 
447 void msm_gpu_cleanup(struct msm_gpu *gpu)
448 {
449 	DBG("%s", gpu->name);
450 
451 	WARN_ON(!list_empty(&gpu->active_list));
452 
453 	bs_fini(gpu);
454 
455 	if (gpu->rb) {
456 		if (gpu->rb_iova)
457 			msm_gem_put_iova(gpu->rb->bo, gpu->id);
458 		msm_ringbuffer_destroy(gpu->rb);
459 	}
460 
461 	if (gpu->iommu)
462 		iommu_domain_free(gpu->iommu);
463 }
464