xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_v3d.c (revision b3f7787b)
1 /*
2  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published by
8  * the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License along with
16  * this program.  If not, see <http://www.gnu.org/licenses/>.
17  */
18 
19 #include "linux/clk.h"
20 #include "linux/component.h"
21 #include "linux/pm_runtime.h"
22 #include "vc4_drv.h"
23 #include "vc4_regs.h"
24 
25 #ifdef CONFIG_DEBUG_FS
26 #define REGDEF(reg) { reg, #reg }
27 static const struct {
28 	uint32_t reg;
29 	const char *name;
30 } vc4_reg_defs[] = {
31 	REGDEF(V3D_IDENT0),
32 	REGDEF(V3D_IDENT1),
33 	REGDEF(V3D_IDENT2),
34 	REGDEF(V3D_SCRATCH),
35 	REGDEF(V3D_L2CACTL),
36 	REGDEF(V3D_SLCACTL),
37 	REGDEF(V3D_INTCTL),
38 	REGDEF(V3D_INTENA),
39 	REGDEF(V3D_INTDIS),
40 	REGDEF(V3D_CT0CS),
41 	REGDEF(V3D_CT1CS),
42 	REGDEF(V3D_CT0EA),
43 	REGDEF(V3D_CT1EA),
44 	REGDEF(V3D_CT0CA),
45 	REGDEF(V3D_CT1CA),
46 	REGDEF(V3D_CT00RA0),
47 	REGDEF(V3D_CT01RA0),
48 	REGDEF(V3D_CT0LC),
49 	REGDEF(V3D_CT1LC),
50 	REGDEF(V3D_CT0PC),
51 	REGDEF(V3D_CT1PC),
52 	REGDEF(V3D_PCS),
53 	REGDEF(V3D_BFC),
54 	REGDEF(V3D_RFC),
55 	REGDEF(V3D_BPCA),
56 	REGDEF(V3D_BPCS),
57 	REGDEF(V3D_BPOA),
58 	REGDEF(V3D_BPOS),
59 	REGDEF(V3D_BXCF),
60 	REGDEF(V3D_SQRSV0),
61 	REGDEF(V3D_SQRSV1),
62 	REGDEF(V3D_SQCNTL),
63 	REGDEF(V3D_SRQPC),
64 	REGDEF(V3D_SRQUA),
65 	REGDEF(V3D_SRQUL),
66 	REGDEF(V3D_SRQCS),
67 	REGDEF(V3D_VPACNTL),
68 	REGDEF(V3D_VPMBASE),
69 	REGDEF(V3D_PCTRC),
70 	REGDEF(V3D_PCTRE),
71 	REGDEF(V3D_PCTR0),
72 	REGDEF(V3D_PCTRS0),
73 	REGDEF(V3D_PCTR1),
74 	REGDEF(V3D_PCTRS1),
75 	REGDEF(V3D_PCTR2),
76 	REGDEF(V3D_PCTRS2),
77 	REGDEF(V3D_PCTR3),
78 	REGDEF(V3D_PCTRS3),
79 	REGDEF(V3D_PCTR4),
80 	REGDEF(V3D_PCTRS4),
81 	REGDEF(V3D_PCTR5),
82 	REGDEF(V3D_PCTRS5),
83 	REGDEF(V3D_PCTR6),
84 	REGDEF(V3D_PCTRS6),
85 	REGDEF(V3D_PCTR7),
86 	REGDEF(V3D_PCTRS7),
87 	REGDEF(V3D_PCTR8),
88 	REGDEF(V3D_PCTRS8),
89 	REGDEF(V3D_PCTR9),
90 	REGDEF(V3D_PCTRS9),
91 	REGDEF(V3D_PCTR10),
92 	REGDEF(V3D_PCTRS10),
93 	REGDEF(V3D_PCTR11),
94 	REGDEF(V3D_PCTRS11),
95 	REGDEF(V3D_PCTR12),
96 	REGDEF(V3D_PCTRS12),
97 	REGDEF(V3D_PCTR13),
98 	REGDEF(V3D_PCTRS13),
99 	REGDEF(V3D_PCTR14),
100 	REGDEF(V3D_PCTRS14),
101 	REGDEF(V3D_PCTR15),
102 	REGDEF(V3D_PCTRS15),
103 	REGDEF(V3D_DBGE),
104 	REGDEF(V3D_FDBGO),
105 	REGDEF(V3D_FDBGB),
106 	REGDEF(V3D_FDBGR),
107 	REGDEF(V3D_FDBGS),
108 	REGDEF(V3D_ERRSTAT),
109 };
110 
111 int vc4_v3d_debugfs_regs(struct seq_file *m, void *unused)
112 {
113 	struct drm_info_node *node = (struct drm_info_node *)m->private;
114 	struct drm_device *dev = node->minor->dev;
115 	struct vc4_dev *vc4 = to_vc4_dev(dev);
116 	int i;
117 
118 	for (i = 0; i < ARRAY_SIZE(vc4_reg_defs); i++) {
119 		seq_printf(m, "%s (0x%04x): 0x%08x\n",
120 			   vc4_reg_defs[i].name, vc4_reg_defs[i].reg,
121 			   V3D_READ(vc4_reg_defs[i].reg));
122 	}
123 
124 	return 0;
125 }
126 
127 int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
128 {
129 	struct drm_info_node *node = (struct drm_info_node *)m->private;
130 	struct drm_device *dev = node->minor->dev;
131 	struct vc4_dev *vc4 = to_vc4_dev(dev);
132 	uint32_t ident1 = V3D_READ(V3D_IDENT1);
133 	uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
134 	uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
135 	uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
136 
137 	seq_printf(m, "Revision:   %d\n",
138 		   VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
139 	seq_printf(m, "Slices:     %d\n", nslc);
140 	seq_printf(m, "TMUs:       %d\n", nslc * tups);
141 	seq_printf(m, "QPUs:       %d\n", nslc * qups);
142 	seq_printf(m, "Semaphores: %d\n",
143 		   VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
144 
145 	return 0;
146 }
147 #endif /* CONFIG_DEBUG_FS */
148 
149 static void vc4_v3d_init_hw(struct drm_device *dev)
150 {
151 	struct vc4_dev *vc4 = to_vc4_dev(dev);
152 
153 	/* Take all the memory that would have been reserved for user
154 	 * QPU programs, since we don't have an interface for running
155 	 * them, anyway.
156 	 */
157 	V3D_WRITE(V3D_VPMBASE, 0);
158 }
159 
160 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
161 {
162 	struct drm_device *dev = vc4->dev;
163 	unsigned long irqflags;
164 	int slot;
165 	uint64_t seqno = 0;
166 	struct vc4_exec_info *exec;
167 
168 try_again:
169 	spin_lock_irqsave(&vc4->job_lock, irqflags);
170 	slot = ffs(~vc4->bin_alloc_used);
171 	if (slot != 0) {
172 		/* Switch from ffs() bit index to a 0-based index. */
173 		slot--;
174 		vc4->bin_alloc_used |= BIT(slot);
175 		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
176 		return slot;
177 	}
178 
179 	/* Couldn't find an open slot.  Wait for render to complete
180 	 * and try again.
181 	 */
182 	exec = vc4_last_render_job(vc4);
183 	if (exec)
184 		seqno = exec->seqno;
185 	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
186 
187 	if (seqno) {
188 		int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
189 
190 		if (ret == 0)
191 			goto try_again;
192 
193 		return ret;
194 	}
195 
196 	return -ENOMEM;
197 }
198 
199 /**
200  * vc4_allocate_bin_bo() - allocates the memory that will be used for
201  * tile binning.
202  *
203  * The binner has a limitation that the addresses in the tile state
204  * buffer that point into the tile alloc buffer or binner overflow
205  * memory only have 28 bits (256MB), and the top 4 on the bus for
206  * tile alloc references end up coming from the tile state buffer's
207  * address.
208  *
209  * To work around this, we allocate a single large buffer while V3D is
210  * in use, make sure that it has the top 4 bits constant across its
211  * entire extent, and then put the tile state, tile alloc, and binner
212  * overflow memory inside that buffer.
213  *
214  * This creates a limitation where we may not be able to execute a job
215  * if it doesn't fit within the buffer that we allocated up front.
216  * However, it turns out that 16MB is "enough for anybody", and
217  * real-world applications run into allocation failures from the
218  * overall CMA pool before they make scenes complicated enough to run
219  * out of bin space.
220  */
221 int
222 vc4_allocate_bin_bo(struct drm_device *drm)
223 {
224 	struct vc4_dev *vc4 = to_vc4_dev(drm);
225 	struct vc4_v3d *v3d = vc4->v3d;
226 	uint32_t size = 16 * 1024 * 1024;
227 	int ret = 0;
228 	struct list_head list;
229 
230 	/* We may need to try allocating more than once to get a BO
231 	 * that doesn't cross 256MB.  Track the ones we've allocated
232 	 * that failed so far, so that we can free them when we've got
233 	 * one that succeeded (if we freed them right away, our next
234 	 * allocation would probably be the same chunk of memory).
235 	 */
236 	INIT_LIST_HEAD(&list);
237 
238 	while (true) {
239 		struct vc4_bo *bo = vc4_bo_create(drm, size, true);
240 
241 		if (IS_ERR(bo)) {
242 			ret = PTR_ERR(bo);
243 
244 			dev_err(&v3d->pdev->dev,
245 				"Failed to allocate memory for tile binning: "
246 				"%d. You may need to enable CMA or give it "
247 				"more memory.",
248 				ret);
249 			break;
250 		}
251 
252 		/* Check if this BO won't trigger the addressing bug. */
253 		if ((bo->base.paddr & 0xf0000000) ==
254 		    ((bo->base.paddr + bo->base.base.size - 1) & 0xf0000000)) {
255 			vc4->bin_bo = bo;
256 
257 			/* Set up for allocating 512KB chunks of
258 			 * binner memory.  The biggest allocation we
259 			 * need to do is for the initial tile alloc +
260 			 * tile state buffer.  We can render to a
261 			 * maximum of ((2048*2048) / (32*32) = 4096
262 			 * tiles in a frame (until we do floating
263 			 * point rendering, at which point it would be
264 			 * 8192).  Tile state is 48b/tile (rounded to
265 			 * a page), and tile alloc is 32b/tile
266 			 * (rounded to a page), plus a page of extra,
267 			 * for a total of 320kb for our worst-case.
268 			 * We choose 512kb so that it divides evenly
269 			 * into our 16MB, and the rest of the 512kb
270 			 * will be used as storage for the overflow
271 			 * from the initial 32b CL per bin.
272 			 */
273 			vc4->bin_alloc_size = 512 * 1024;
274 			vc4->bin_alloc_used = 0;
275 			vc4->bin_alloc_overflow = 0;
276 			WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
277 				     bo->base.base.size / vc4->bin_alloc_size);
278 
279 			break;
280 		}
281 
282 		/* Put it on the list to free later, and try again. */
283 		list_add(&bo->unref_head, &list);
284 	}
285 
286 	/* Free all the BOs we allocated but didn't choose. */
287 	while (!list_empty(&list)) {
288 		struct vc4_bo *bo = list_last_entry(&list,
289 						    struct vc4_bo, unref_head);
290 
291 		list_del(&bo->unref_head);
292 		drm_gem_object_put_unlocked(&bo->base.base);
293 	}
294 
295 	return ret;
296 }
297 
298 #ifdef CONFIG_PM
299 static int vc4_v3d_runtime_suspend(struct device *dev)
300 {
301 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
302 	struct vc4_dev *vc4 = v3d->vc4;
303 
304 	vc4_irq_uninstall(vc4->dev);
305 
306 	drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
307 	vc4->bin_bo = NULL;
308 
309 	clk_disable_unprepare(v3d->clk);
310 
311 	return 0;
312 }
313 
314 static int vc4_v3d_runtime_resume(struct device *dev)
315 {
316 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
317 	struct vc4_dev *vc4 = v3d->vc4;
318 	int ret;
319 
320 	ret = vc4_allocate_bin_bo(vc4->dev);
321 	if (ret)
322 		return ret;
323 
324 	ret = clk_prepare_enable(v3d->clk);
325 	if (ret != 0)
326 		return ret;
327 
328 	vc4_v3d_init_hw(vc4->dev);
329 	vc4_irq_postinstall(vc4->dev);
330 
331 	return 0;
332 }
333 #endif
334 
335 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
336 {
337 	struct platform_device *pdev = to_platform_device(dev);
338 	struct drm_device *drm = dev_get_drvdata(master);
339 	struct vc4_dev *vc4 = to_vc4_dev(drm);
340 	struct vc4_v3d *v3d = NULL;
341 	int ret;
342 
343 	v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
344 	if (!v3d)
345 		return -ENOMEM;
346 
347 	dev_set_drvdata(dev, v3d);
348 
349 	v3d->pdev = pdev;
350 
351 	v3d->regs = vc4_ioremap_regs(pdev, 0);
352 	if (IS_ERR(v3d->regs))
353 		return PTR_ERR(v3d->regs);
354 
355 	vc4->v3d = v3d;
356 	v3d->vc4 = vc4;
357 
358 	v3d->clk = devm_clk_get(dev, NULL);
359 	if (IS_ERR(v3d->clk)) {
360 		int ret = PTR_ERR(v3d->clk);
361 
362 		if (ret == -ENOENT) {
363 			/* bcm2835 didn't have a clock reference in the DT. */
364 			ret = 0;
365 			v3d->clk = NULL;
366 		} else {
367 			if (ret != -EPROBE_DEFER)
368 				dev_err(dev, "Failed to get V3D clock: %d\n",
369 					ret);
370 			return ret;
371 		}
372 	}
373 
374 	if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
375 		DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
376 			  V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
377 		return -EINVAL;
378 	}
379 
380 	ret = clk_prepare_enable(v3d->clk);
381 	if (ret != 0)
382 		return ret;
383 
384 	ret = vc4_allocate_bin_bo(drm);
385 	if (ret) {
386 		clk_disable_unprepare(v3d->clk);
387 		return ret;
388 	}
389 
390 	/* Reset the binner overflow address/size at setup, to be sure
391 	 * we don't reuse an old one.
392 	 */
393 	V3D_WRITE(V3D_BPOA, 0);
394 	V3D_WRITE(V3D_BPOS, 0);
395 
396 	vc4_v3d_init_hw(drm);
397 
398 	ret = drm_irq_install(drm, platform_get_irq(pdev, 0));
399 	if (ret) {
400 		DRM_ERROR("Failed to install IRQ handler\n");
401 		return ret;
402 	}
403 
404 	pm_runtime_use_autosuspend(dev);
405 	pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
406 	pm_runtime_enable(dev);
407 
408 	return 0;
409 }
410 
411 static void vc4_v3d_unbind(struct device *dev, struct device *master,
412 			   void *data)
413 {
414 	struct drm_device *drm = dev_get_drvdata(master);
415 	struct vc4_dev *vc4 = to_vc4_dev(drm);
416 
417 	pm_runtime_disable(dev);
418 
419 	drm_irq_uninstall(drm);
420 
421 	/* Disable the binner's overflow memory address, so the next
422 	 * driver probe (if any) doesn't try to reuse our old
423 	 * allocation.
424 	 */
425 	V3D_WRITE(V3D_BPOA, 0);
426 	V3D_WRITE(V3D_BPOS, 0);
427 
428 	vc4->v3d = NULL;
429 }
430 
431 static const struct dev_pm_ops vc4_v3d_pm_ops = {
432 	SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
433 };
434 
435 static const struct component_ops vc4_v3d_ops = {
436 	.bind   = vc4_v3d_bind,
437 	.unbind = vc4_v3d_unbind,
438 };
439 
440 static int vc4_v3d_dev_probe(struct platform_device *pdev)
441 {
442 	return component_add(&pdev->dev, &vc4_v3d_ops);
443 }
444 
445 static int vc4_v3d_dev_remove(struct platform_device *pdev)
446 {
447 	component_del(&pdev->dev, &vc4_v3d_ops);
448 	return 0;
449 }
450 
451 static const struct of_device_id vc4_v3d_dt_match[] = {
452 	{ .compatible = "brcm,bcm2835-v3d" },
453 	{ .compatible = "brcm,cygnus-v3d" },
454 	{ .compatible = "brcm,vc4-v3d" },
455 	{}
456 };
457 
458 struct platform_driver vc4_v3d_driver = {
459 	.probe = vc4_v3d_dev_probe,
460 	.remove = vc4_v3d_dev_remove,
461 	.driver = {
462 		.name = "vc4_v3d",
463 		.of_match_table = vc4_v3d_dt_match,
464 		.pm = &vc4_v3d_pm_ops,
465 	},
466 };
467