1 // SPDX-License-Identifier: GPL-2.0+
2 /* Copyright (C) 2014-2018 Broadcom */
3
4 /**
5 * DOC: Broadcom V3D Graphics Driver
6 *
7 * This driver supports the Broadcom V3D 3.3 and 4.1 OpenGL ES GPUs.
8 * For V3D 2.x support, see the VC4 driver.
9 *
10 * The V3D GPU includes a tiled render (composed of a bin and render
11 * pipelines), the TFU (texture formatting unit), and the CSD (compute
12 * shader dispatch).
13 */
14
15 #include <linux/clk.h>
16 #include <linux/device.h>
17 #include <linux/dma-mapping.h>
18 #include <linux/io.h>
19 #include <linux/module.h>
20 #include <linux/of.h>
21 #include <linux/of_platform.h>
22 #include <linux/platform_device.h>
23 #include <linux/sched/clock.h>
24 #include <linux/reset.h>
25
26 #include <drm/drm_drv.h>
27 #include <drm/drm_managed.h>
28 #include <drm/drm_print.h>
29 #include <uapi/drm/v3d_drm.h>
30
31 #include "v3d_drv.h"
32 #include "v3d_regs.h"
33
34 #define DRIVER_NAME "v3d"
35 #define DRIVER_DESC "Broadcom V3D graphics"
36 #define DRIVER_MAJOR 1
37 #define DRIVER_MINOR 0
38 #define DRIVER_PATCHLEVEL 0
39
40 /* Only expose the `super_pages` modparam if THP is enabled. */
41 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
42 bool super_pages = true;
43 module_param_named(super_pages, super_pages, bool, 0400);
44 MODULE_PARM_DESC(super_pages, "Enable/Disable Super Pages support.");
45 #endif
46
v3d_get_param_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)47 static int v3d_get_param_ioctl(struct drm_device *dev, void *data,
48 struct drm_file *file_priv)
49 {
50 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
51 struct v3d_dev *v3d = to_v3d_dev(dev);
52 struct drm_v3d_get_param *args = data;
53 static const u32 reg_map[] = {
54 [DRM_V3D_PARAM_V3D_UIFCFG] = V3D_HUB_UIFCFG,
55 [DRM_V3D_PARAM_V3D_HUB_IDENT1] = V3D_HUB_IDENT1,
56 [DRM_V3D_PARAM_V3D_HUB_IDENT2] = V3D_HUB_IDENT2,
57 [DRM_V3D_PARAM_V3D_HUB_IDENT3] = V3D_HUB_IDENT3,
58 [DRM_V3D_PARAM_V3D_CORE0_IDENT0] = V3D_CTL_IDENT0,
59 [DRM_V3D_PARAM_V3D_CORE0_IDENT1] = V3D_CTL_IDENT1,
60 [DRM_V3D_PARAM_V3D_CORE0_IDENT2] = V3D_CTL_IDENT2,
61 };
62
63 if (args->pad != 0)
64 return -EINVAL;
65
66 /* Note that DRM_V3D_PARAM_V3D_CORE0_IDENT0 is 0, so we need
67 * to explicitly allow it in the "the register in our
68 * parameter map" check.
69 */
70 if (args->param < ARRAY_SIZE(reg_map) &&
71 (reg_map[args->param] ||
72 args->param == DRM_V3D_PARAM_V3D_CORE0_IDENT0)) {
73 u32 offset = reg_map[args->param];
74
75 if (args->value != 0)
76 return -EINVAL;
77
78 if (args->param >= DRM_V3D_PARAM_V3D_CORE0_IDENT0 &&
79 args->param <= DRM_V3D_PARAM_V3D_CORE0_IDENT2) {
80 args->value = V3D_CORE_READ(0, offset);
81 } else {
82 args->value = V3D_READ(offset);
83 }
84 return 0;
85 }
86
87 switch (args->param) {
88 case DRM_V3D_PARAM_SUPPORTS_TFU:
89 args->value = 1;
90 return 0;
91 case DRM_V3D_PARAM_SUPPORTS_CSD:
92 args->value = v3d_has_csd(v3d);
93 return 0;
94 case DRM_V3D_PARAM_SUPPORTS_CACHE_FLUSH:
95 args->value = 1;
96 return 0;
97 case DRM_V3D_PARAM_SUPPORTS_PERFMON:
98 args->value = (v3d->ver >= V3D_GEN_41);
99 return 0;
100 case DRM_V3D_PARAM_SUPPORTS_MULTISYNC_EXT:
101 args->value = 1;
102 return 0;
103 case DRM_V3D_PARAM_SUPPORTS_CPU_QUEUE:
104 args->value = 1;
105 return 0;
106 case DRM_V3D_PARAM_MAX_PERF_COUNTERS:
107 args->value = v3d->perfmon_info.max_counters;
108 return 0;
109 case DRM_V3D_PARAM_SUPPORTS_SUPER_PAGES:
110 args->value = !!drm_gem_get_huge_mnt(dev);
111 return 0;
112 case DRM_V3D_PARAM_GLOBAL_RESET_COUNTER:
113 mutex_lock(&v3d->reset_lock);
114 args->value = v3d->reset_counter;
115 mutex_unlock(&v3d->reset_lock);
116 return 0;
117 case DRM_V3D_PARAM_CONTEXT_RESET_COUNTER:
118 mutex_lock(&v3d->reset_lock);
119 args->value = v3d_priv->reset_counter;
120 mutex_unlock(&v3d->reset_lock);
121 return 0;
122 default:
123 drm_dbg(dev, "Unknown parameter %d\n", args->param);
124 return -EINVAL;
125 }
126 }
127
128 static int
v3d_open(struct drm_device * dev,struct drm_file * file)129 v3d_open(struct drm_device *dev, struct drm_file *file)
130 {
131 struct v3d_dev *v3d = to_v3d_dev(dev);
132 struct v3d_file_priv *v3d_priv;
133 struct drm_gpu_scheduler *sched;
134 int i;
135
136 v3d_priv = kzalloc_obj(*v3d_priv);
137 if (!v3d_priv)
138 return -ENOMEM;
139
140 v3d_priv->v3d = v3d;
141
142 for (i = 0; i < V3D_MAX_QUEUES; i++) {
143 sched = &v3d->queue[i].sched;
144 drm_sched_entity_init(&v3d_priv->sched_entity[i],
145 DRM_SCHED_PRIORITY_NORMAL, &sched,
146 1, NULL);
147
148 memset(&v3d_priv->stats[i], 0, sizeof(v3d_priv->stats[i]));
149 seqcount_init(&v3d_priv->stats[i].lock);
150 }
151
152 v3d_perfmon_open_file(v3d_priv);
153 file->driver_priv = v3d_priv;
154
155 return 0;
156 }
157
158 static void
v3d_postclose(struct drm_device * dev,struct drm_file * file)159 v3d_postclose(struct drm_device *dev, struct drm_file *file)
160 {
161 struct v3d_dev *v3d = to_v3d_dev(dev);
162 struct v3d_file_priv *v3d_priv = file->driver_priv;
163 unsigned long irqflags;
164 enum v3d_queue q;
165
166 for (q = 0; q < V3D_MAX_QUEUES; q++) {
167 struct v3d_queue_state *queue = &v3d->queue[q];
168 struct v3d_job *job = queue->active_job;
169
170 drm_sched_entity_destroy(&v3d_priv->sched_entity[q]);
171
172 if (job && job->base.entity == &v3d_priv->sched_entity[q]) {
173 spin_lock_irqsave(&queue->queue_lock, irqflags);
174 job->file_priv = NULL;
175 spin_unlock_irqrestore(&queue->queue_lock, irqflags);
176 }
177 }
178
179 v3d_perfmon_close_file(v3d_priv);
180 kfree(v3d_priv);
181 }
182
v3d_get_stats(const struct v3d_stats * stats,u64 timestamp,u64 * active_runtime,u64 * jobs_completed)183 void v3d_get_stats(const struct v3d_stats *stats, u64 timestamp,
184 u64 *active_runtime, u64 *jobs_completed)
185 {
186 unsigned int seq;
187
188 do {
189 seq = read_seqcount_begin(&stats->lock);
190 *active_runtime = stats->enabled_ns;
191 if (stats->start_ns)
192 *active_runtime += timestamp - stats->start_ns;
193 *jobs_completed = stats->jobs_completed;
194 } while (read_seqcount_retry(&stats->lock, seq));
195 }
196
v3d_show_fdinfo(struct drm_printer * p,struct drm_file * file)197 static void v3d_show_fdinfo(struct drm_printer *p, struct drm_file *file)
198 {
199 struct v3d_file_priv *file_priv = file->driver_priv;
200 u64 timestamp = local_clock();
201 enum v3d_queue queue;
202
203 for (queue = 0; queue < V3D_MAX_QUEUES; queue++) {
204 struct v3d_stats *stats = &file_priv->stats[queue];
205 u64 active_runtime, jobs_completed;
206
207 v3d_get_stats(stats, timestamp, &active_runtime, &jobs_completed);
208
209 /* Note that, in case of a GPU reset, the time spent during an
210 * attempt of executing the job is not computed in the runtime.
211 */
212 drm_printf(p, "drm-engine-%s: \t%llu ns\n",
213 v3d_queue_to_string(queue), active_runtime);
214
215 /* Note that we only count jobs that completed. Therefore, jobs
216 * that were resubmitted due to a GPU reset are not computed.
217 */
218 drm_printf(p, "v3d-jobs-%s: \t%llu jobs\n",
219 v3d_queue_to_string(queue), jobs_completed);
220 }
221
222 drm_show_memory_stats(p, file);
223 }
224
225 static const struct file_operations v3d_drm_fops = {
226 .owner = THIS_MODULE,
227 DRM_GEM_FOPS,
228 .show_fdinfo = drm_show_fdinfo,
229 };
230
231 /* DRM_AUTH is required on SUBMIT_CL for now, while we don't have GMP
232 * protection between clients. Note that render nodes would be
233 * able to submit CLs that could access BOs from clients authenticated
234 * with the master node. The TFU doesn't use the GMP, so it would
235 * need to stay DRM_AUTH until we do buffer size/offset validation.
236 */
237 static const struct drm_ioctl_desc v3d_drm_ioctls[] = {
238 DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CL, v3d_submit_cl_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
239 DRM_IOCTL_DEF_DRV(V3D_WAIT_BO, v3d_wait_bo_ioctl, DRM_RENDER_ALLOW),
240 DRM_IOCTL_DEF_DRV(V3D_CREATE_BO, v3d_create_bo_ioctl, DRM_RENDER_ALLOW),
241 DRM_IOCTL_DEF_DRV(V3D_MMAP_BO, v3d_mmap_bo_ioctl, DRM_RENDER_ALLOW),
242 DRM_IOCTL_DEF_DRV(V3D_GET_PARAM, v3d_get_param_ioctl, DRM_RENDER_ALLOW),
243 DRM_IOCTL_DEF_DRV(V3D_GET_BO_OFFSET, v3d_get_bo_offset_ioctl, DRM_RENDER_ALLOW),
244 DRM_IOCTL_DEF_DRV(V3D_SUBMIT_TFU, v3d_submit_tfu_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
245 DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CSD, v3d_submit_csd_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
246 DRM_IOCTL_DEF_DRV(V3D_PERFMON_CREATE, v3d_perfmon_create_ioctl, DRM_RENDER_ALLOW),
247 DRM_IOCTL_DEF_DRV(V3D_PERFMON_DESTROY, v3d_perfmon_destroy_ioctl, DRM_RENDER_ALLOW),
248 DRM_IOCTL_DEF_DRV(V3D_PERFMON_GET_VALUES, v3d_perfmon_get_values_ioctl, DRM_RENDER_ALLOW),
249 DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CPU, v3d_submit_cpu_ioctl, DRM_RENDER_ALLOW | DRM_AUTH),
250 DRM_IOCTL_DEF_DRV(V3D_PERFMON_GET_COUNTER, v3d_perfmon_get_counter_ioctl, DRM_RENDER_ALLOW),
251 DRM_IOCTL_DEF_DRV(V3D_PERFMON_SET_GLOBAL, v3d_perfmon_set_global_ioctl, DRM_RENDER_ALLOW),
252 };
253
254 static const struct drm_driver v3d_drm_driver = {
255 .driver_features = (DRIVER_GEM |
256 DRIVER_RENDER |
257 DRIVER_SYNCOBJ),
258
259 .open = v3d_open,
260 .postclose = v3d_postclose,
261
262 #if defined(CONFIG_DEBUG_FS)
263 .debugfs_init = v3d_debugfs_init,
264 #endif
265
266 .gem_create_object = v3d_create_object,
267 .gem_prime_import_sg_table = v3d_prime_import_sg_table,
268
269 .ioctls = v3d_drm_ioctls,
270 .num_ioctls = ARRAY_SIZE(v3d_drm_ioctls),
271 .fops = &v3d_drm_fops,
272 .show_fdinfo = v3d_show_fdinfo,
273
274 .name = DRIVER_NAME,
275 .desc = DRIVER_DESC,
276 .major = DRIVER_MAJOR,
277 .minor = DRIVER_MINOR,
278 .patchlevel = DRIVER_PATCHLEVEL,
279 };
280
281 static const struct of_device_id v3d_of_match[] = {
282 { .compatible = "brcm,2711-v3d", .data = (void *)V3D_GEN_42 },
283 { .compatible = "brcm,2712-v3d", .data = (void *)V3D_GEN_71 },
284 { .compatible = "brcm,7268-v3d", .data = (void *)V3D_GEN_33 },
285 { .compatible = "brcm,7278-v3d", .data = (void *)V3D_GEN_41 },
286 {},
287 };
288 MODULE_DEVICE_TABLE(of, v3d_of_match);
289
290 static void
v3d_idle_sms(struct v3d_dev * v3d)291 v3d_idle_sms(struct v3d_dev *v3d)
292 {
293 if (v3d->ver < V3D_GEN_71)
294 return;
295
296 V3D_SMS_WRITE(V3D_SMS_TEE_CS, V3D_SMS_CLEAR_POWER_OFF);
297
298 if (wait_for((V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_TEE_CS),
299 V3D_SMS_STATE) == V3D_SMS_IDLE), 100)) {
300 drm_err(&v3d->drm, "Failed to power up SMS\n");
301 }
302
303 v3d_reset_sms(v3d);
304 }
305
306 static void
v3d_power_off_sms(struct v3d_dev * v3d)307 v3d_power_off_sms(struct v3d_dev *v3d)
308 {
309 if (v3d->ver < V3D_GEN_71)
310 return;
311
312 V3D_SMS_WRITE(V3D_SMS_TEE_CS, V3D_SMS_POWER_OFF);
313
314 if (wait_for((V3D_GET_FIELD(V3D_SMS_READ(V3D_SMS_TEE_CS),
315 V3D_SMS_STATE) == V3D_SMS_POWER_OFF_STATE), 100)) {
316 drm_err(&v3d->drm, "Failed to power off SMS\n");
317 }
318 }
319
320 static int
map_regs(struct v3d_dev * v3d,void __iomem ** regs,const char * name)321 map_regs(struct v3d_dev *v3d, void __iomem **regs, const char *name)
322 {
323 *regs = devm_platform_ioremap_resource_byname(v3d_to_pdev(v3d), name);
324 return PTR_ERR_OR_ZERO(*regs);
325 }
326
v3d_platform_drm_probe(struct platform_device * pdev)327 static int v3d_platform_drm_probe(struct platform_device *pdev)
328 {
329 struct device *dev = &pdev->dev;
330 struct drm_device *drm;
331 struct v3d_dev *v3d;
332 enum v3d_gen gen;
333 int ret;
334 u32 mmu_debug;
335 u32 ident1, ident3;
336 u64 mask;
337
338 v3d = devm_drm_dev_alloc(dev, &v3d_drm_driver, struct v3d_dev, drm);
339 if (IS_ERR(v3d))
340 return PTR_ERR(v3d);
341
342 drm = &v3d->drm;
343
344 platform_set_drvdata(pdev, drm);
345
346 gen = (uintptr_t)of_device_get_match_data(dev);
347 v3d->ver = gen;
348
349 ret = map_regs(v3d, &v3d->hub_regs, "hub");
350 if (ret)
351 return ret;
352
353 ret = map_regs(v3d, &v3d->core_regs[0], "core0");
354 if (ret)
355 return ret;
356
357 if (v3d->ver >= V3D_GEN_71) {
358 ret = map_regs(v3d, &v3d->sms_regs, "sms");
359 if (ret)
360 return ret;
361 }
362
363 v3d->clk = devm_clk_get_optional(dev, NULL);
364 if (IS_ERR(v3d->clk))
365 return dev_err_probe(dev, PTR_ERR(v3d->clk), "Failed to get V3D clock\n");
366
367 ret = clk_prepare_enable(v3d->clk);
368 if (ret) {
369 dev_err(&pdev->dev, "Couldn't enable the V3D clock\n");
370 return ret;
371 }
372
373 v3d_idle_sms(v3d);
374
375 mmu_debug = V3D_READ(V3D_MMU_DEBUG_INFO);
376 mask = DMA_BIT_MASK(30 + V3D_GET_FIELD(mmu_debug, V3D_MMU_PA_WIDTH));
377 ret = dma_set_mask_and_coherent(dev, mask);
378 if (ret)
379 goto clk_disable;
380
381 dma_set_max_seg_size(&pdev->dev, UINT_MAX);
382
383 v3d->va_width = 30 + V3D_GET_FIELD(mmu_debug, V3D_MMU_VA_WIDTH);
384
385 ident1 = V3D_READ(V3D_HUB_IDENT1);
386 v3d->ver = (V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_TVER) * 10 +
387 V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_REV));
388 /* Make sure that the V3D tech version retrieved from the HW is equal
389 * to the one advertised by the device tree.
390 */
391 WARN_ON(v3d->ver != gen);
392
393 v3d->cores = V3D_GET_FIELD(ident1, V3D_HUB_IDENT1_NCORES);
394 WARN_ON(v3d->cores > 1); /* multicore not yet implemented */
395
396 ident3 = V3D_READ(V3D_HUB_IDENT3);
397 v3d->rev = V3D_GET_FIELD(ident3, V3D_HUB_IDENT3_IPREV);
398
399 v3d_perfmon_init(v3d);
400
401 v3d->reset = devm_reset_control_get_exclusive(dev, NULL);
402 if (IS_ERR(v3d->reset)) {
403 ret = PTR_ERR(v3d->reset);
404
405 if (ret == -EPROBE_DEFER)
406 goto clk_disable;
407
408 v3d->reset = NULL;
409 ret = map_regs(v3d, &v3d->bridge_regs, "bridge");
410 if (ret) {
411 dev_err(dev,
412 "Failed to get reset control or bridge regs\n");
413 goto clk_disable;
414 }
415 }
416
417 if (v3d->ver < V3D_GEN_41) {
418 ret = map_regs(v3d, &v3d->gca_regs, "gca");
419 if (ret)
420 goto clk_disable;
421 }
422
423 v3d->mmu_scratch = dma_alloc_wc(dev, 4096, &v3d->mmu_scratch_paddr,
424 GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
425 if (!v3d->mmu_scratch) {
426 dev_err(dev, "Failed to allocate MMU scratch page\n");
427 ret = -ENOMEM;
428 goto clk_disable;
429 }
430
431 ret = v3d_gem_init(drm);
432 if (ret)
433 goto dma_free;
434
435 ret = v3d_irq_init(v3d);
436 if (ret)
437 goto gem_destroy;
438
439 ret = drm_dev_register(drm, 0);
440 if (ret)
441 goto irq_disable;
442
443 ret = v3d_sysfs_init(dev);
444 if (ret)
445 goto drm_unregister;
446
447 return 0;
448
449 drm_unregister:
450 drm_dev_unregister(drm);
451 irq_disable:
452 v3d_irq_disable(v3d);
453 gem_destroy:
454 v3d_gem_destroy(drm);
455 dma_free:
456 dma_free_wc(dev, 4096, v3d->mmu_scratch, v3d->mmu_scratch_paddr);
457 clk_disable:
458 clk_disable_unprepare(v3d->clk);
459 return ret;
460 }
461
v3d_platform_drm_remove(struct platform_device * pdev)462 static void v3d_platform_drm_remove(struct platform_device *pdev)
463 {
464 struct drm_device *drm = platform_get_drvdata(pdev);
465 struct v3d_dev *v3d = to_v3d_dev(drm);
466 struct device *dev = &pdev->dev;
467
468 v3d_sysfs_destroy(dev);
469
470 drm_dev_unregister(drm);
471
472 v3d_gem_destroy(drm);
473
474 dma_free_wc(v3d->drm.dev, 4096, v3d->mmu_scratch,
475 v3d->mmu_scratch_paddr);
476
477 v3d_power_off_sms(v3d);
478
479 clk_disable_unprepare(v3d->clk);
480 }
481
482 static struct platform_driver v3d_platform_driver = {
483 .probe = v3d_platform_drm_probe,
484 .remove = v3d_platform_drm_remove,
485 .driver = {
486 .name = "v3d",
487 .of_match_table = v3d_of_match,
488 },
489 };
490
491 module_platform_driver(v3d_platform_driver);
492
493 MODULE_ALIAS("platform:v3d-drm");
494 MODULE_DESCRIPTION("Broadcom V3D DRM Driver");
495 MODULE_AUTHOR("Eric Anholt <eric@anholt.net>");
496 MODULE_LICENSE("GPL v2");
497