2 * Copyright (c) 2014 The Linux Foundation. All rights reserved.
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
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.
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
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/>.
19 #include <linux/clk.h>
20 #include <linux/component.h>
21 #include <linux/pm_runtime.h>
25 static const struct debugfs_reg32 v3d_regs[] = {
26 VC4_REG32(V3D_IDENT0),
27 VC4_REG32(V3D_IDENT1),
28 VC4_REG32(V3D_IDENT2),
29 VC4_REG32(V3D_SCRATCH),
30 VC4_REG32(V3D_L2CACTL),
31 VC4_REG32(V3D_SLCACTL),
32 VC4_REG32(V3D_INTCTL),
33 VC4_REG32(V3D_INTENA),
34 VC4_REG32(V3D_INTDIS),
41 VC4_REG32(V3D_CT00RA0),
42 VC4_REG32(V3D_CT01RA0),
55 VC4_REG32(V3D_SQRSV0),
56 VC4_REG32(V3D_SQRSV1),
57 VC4_REG32(V3D_SQCNTL),
62 VC4_REG32(V3D_VPACNTL),
63 VC4_REG32(V3D_VPMBASE),
66 VC4_REG32(V3D_PCTR(0)),
67 VC4_REG32(V3D_PCTRS(0)),
68 VC4_REG32(V3D_PCTR(1)),
69 VC4_REG32(V3D_PCTRS(1)),
70 VC4_REG32(V3D_PCTR(2)),
71 VC4_REG32(V3D_PCTRS(2)),
72 VC4_REG32(V3D_PCTR(3)),
73 VC4_REG32(V3D_PCTRS(3)),
74 VC4_REG32(V3D_PCTR(4)),
75 VC4_REG32(V3D_PCTRS(4)),
76 VC4_REG32(V3D_PCTR(5)),
77 VC4_REG32(V3D_PCTRS(5)),
78 VC4_REG32(V3D_PCTR(6)),
79 VC4_REG32(V3D_PCTRS(6)),
80 VC4_REG32(V3D_PCTR(7)),
81 VC4_REG32(V3D_PCTRS(7)),
82 VC4_REG32(V3D_PCTR(8)),
83 VC4_REG32(V3D_PCTRS(8)),
84 VC4_REG32(V3D_PCTR(9)),
85 VC4_REG32(V3D_PCTRS(9)),
86 VC4_REG32(V3D_PCTR(10)),
87 VC4_REG32(V3D_PCTRS(10)),
88 VC4_REG32(V3D_PCTR(11)),
89 VC4_REG32(V3D_PCTRS(11)),
90 VC4_REG32(V3D_PCTR(12)),
91 VC4_REG32(V3D_PCTRS(12)),
92 VC4_REG32(V3D_PCTR(13)),
93 VC4_REG32(V3D_PCTRS(13)),
94 VC4_REG32(V3D_PCTR(14)),
95 VC4_REG32(V3D_PCTRS(14)),
96 VC4_REG32(V3D_PCTR(15)),
97 VC4_REG32(V3D_PCTRS(15)),
100 VC4_REG32(V3D_FDBGB),
101 VC4_REG32(V3D_FDBGR),
102 VC4_REG32(V3D_FDBGS),
103 VC4_REG32(V3D_ERRSTAT),
106 static int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
108 struct drm_info_node *node = (struct drm_info_node *)m->private;
109 struct drm_device *dev = node->minor->dev;
110 struct vc4_dev *vc4 = to_vc4_dev(dev);
111 int ret = vc4_v3d_pm_get(vc4);
114 uint32_t ident1 = V3D_READ(V3D_IDENT1);
115 uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
116 uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
117 uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
119 seq_printf(m, "Revision: %d\n",
120 VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
121 seq_printf(m, "Slices: %d\n", nslc);
122 seq_printf(m, "TMUs: %d\n", nslc * tups);
123 seq_printf(m, "QPUs: %d\n", nslc * qups);
124 seq_printf(m, "Semaphores: %d\n",
125 VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
133 * Wraps pm_runtime_get_sync() in a refcount, so that we can reliably
134 * get the pm_runtime refcount to 0 in vc4_reset().
137 vc4_v3d_pm_get(struct vc4_dev *vc4)
139 mutex_lock(&vc4->power_lock);
140 if (vc4->power_refcount++ == 0) {
141 int ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
144 vc4->power_refcount--;
145 mutex_unlock(&vc4->power_lock);
149 mutex_unlock(&vc4->power_lock);
155 vc4_v3d_pm_put(struct vc4_dev *vc4)
157 mutex_lock(&vc4->power_lock);
158 if (--vc4->power_refcount == 0) {
159 pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
160 pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
162 mutex_unlock(&vc4->power_lock);
165 static void vc4_v3d_init_hw(struct drm_device *dev)
167 struct vc4_dev *vc4 = to_vc4_dev(dev);
169 /* Take all the memory that would have been reserved for user
170 * QPU programs, since we don't have an interface for running
173 V3D_WRITE(V3D_VPMBASE, 0);
176 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
178 struct drm_device *dev = vc4->dev;
179 unsigned long irqflags;
182 struct vc4_exec_info *exec;
185 spin_lock_irqsave(&vc4->job_lock, irqflags);
186 slot = ffs(~vc4->bin_alloc_used);
188 /* Switch from ffs() bit index to a 0-based index. */
190 vc4->bin_alloc_used |= BIT(slot);
191 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
195 /* Couldn't find an open slot. Wait for render to complete
198 exec = vc4_last_render_job(vc4);
201 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
204 int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
216 * vc4_allocate_bin_bo() - allocates the memory that will be used for
219 * The binner has a limitation that the addresses in the tile state
220 * buffer that point into the tile alloc buffer or binner overflow
221 * memory only have 28 bits (256MB), and the top 4 on the bus for
222 * tile alloc references end up coming from the tile state buffer's
225 * To work around this, we allocate a single large buffer while V3D is
226 * in use, make sure that it has the top 4 bits constant across its
227 * entire extent, and then put the tile state, tile alloc, and binner
228 * overflow memory inside that buffer.
230 * This creates a limitation where we may not be able to execute a job
231 * if it doesn't fit within the buffer that we allocated up front.
232 * However, it turns out that 16MB is "enough for anybody", and
233 * real-world applications run into allocation failures from the
234 * overall CMA pool before they make scenes complicated enough to run
237 static int vc4_allocate_bin_bo(struct drm_device *drm)
239 struct vc4_dev *vc4 = to_vc4_dev(drm);
240 struct vc4_v3d *v3d = vc4->v3d;
241 uint32_t size = 16 * 1024 * 1024;
243 struct list_head list;
245 /* We may need to try allocating more than once to get a BO
246 * that doesn't cross 256MB. Track the ones we've allocated
247 * that failed so far, so that we can free them when we've got
248 * one that succeeded (if we freed them right away, our next
249 * allocation would probably be the same chunk of memory).
251 INIT_LIST_HEAD(&list);
254 struct vc4_bo *bo = vc4_bo_create(drm, size, true,
260 dev_err(&v3d->pdev->dev,
261 "Failed to allocate memory for tile binning: "
262 "%d. You may need to enable CMA or give it "
268 /* Check if this BO won't trigger the addressing bug. */
269 if ((bo->base.paddr & 0xf0000000) ==
270 ((bo->base.paddr + bo->base.base.size - 1) & 0xf0000000)) {
273 /* Set up for allocating 512KB chunks of
274 * binner memory. The biggest allocation we
275 * need to do is for the initial tile alloc +
276 * tile state buffer. We can render to a
277 * maximum of ((2048*2048) / (32*32) = 4096
278 * tiles in a frame (until we do floating
279 * point rendering, at which point it would be
280 * 8192). Tile state is 48b/tile (rounded to
281 * a page), and tile alloc is 32b/tile
282 * (rounded to a page), plus a page of extra,
283 * for a total of 320kb for our worst-case.
284 * We choose 512kb so that it divides evenly
285 * into our 16MB, and the rest of the 512kb
286 * will be used as storage for the overflow
287 * from the initial 32b CL per bin.
289 vc4->bin_alloc_size = 512 * 1024;
290 vc4->bin_alloc_used = 0;
291 vc4->bin_alloc_overflow = 0;
292 WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
293 bo->base.base.size / vc4->bin_alloc_size);
298 /* Put it on the list to free later, and try again. */
299 list_add(&bo->unref_head, &list);
302 /* Free all the BOs we allocated but didn't choose. */
303 while (!list_empty(&list)) {
304 struct vc4_bo *bo = list_last_entry(&list,
305 struct vc4_bo, unref_head);
307 list_del(&bo->unref_head);
308 drm_gem_object_put_unlocked(&bo->base.base);
315 static int vc4_v3d_runtime_suspend(struct device *dev)
317 struct vc4_v3d *v3d = dev_get_drvdata(dev);
318 struct vc4_dev *vc4 = v3d->vc4;
320 vc4_irq_uninstall(vc4->dev);
322 drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
325 clk_disable_unprepare(v3d->clk);
330 static int vc4_v3d_runtime_resume(struct device *dev)
332 struct vc4_v3d *v3d = dev_get_drvdata(dev);
333 struct vc4_dev *vc4 = v3d->vc4;
336 ret = vc4_allocate_bin_bo(vc4->dev);
340 ret = clk_prepare_enable(v3d->clk);
344 vc4_v3d_init_hw(vc4->dev);
346 /* We disabled the IRQ as part of vc4_irq_uninstall in suspend. */
347 enable_irq(vc4->dev->irq);
348 vc4_irq_postinstall(vc4->dev);
354 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
356 struct platform_device *pdev = to_platform_device(dev);
357 struct drm_device *drm = dev_get_drvdata(master);
358 struct vc4_dev *vc4 = to_vc4_dev(drm);
359 struct vc4_v3d *v3d = NULL;
362 v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
366 dev_set_drvdata(dev, v3d);
370 v3d->regs = vc4_ioremap_regs(pdev, 0);
371 if (IS_ERR(v3d->regs))
372 return PTR_ERR(v3d->regs);
373 v3d->regset.base = v3d->regs;
374 v3d->regset.regs = v3d_regs;
375 v3d->regset.nregs = ARRAY_SIZE(v3d_regs);
380 v3d->clk = devm_clk_get(dev, NULL);
381 if (IS_ERR(v3d->clk)) {
382 int ret = PTR_ERR(v3d->clk);
384 if (ret == -ENOENT) {
385 /* bcm2835 didn't have a clock reference in the DT. */
389 if (ret != -EPROBE_DEFER)
390 dev_err(dev, "Failed to get V3D clock: %d\n",
396 if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
397 DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
398 V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
402 ret = clk_prepare_enable(v3d->clk);
406 ret = vc4_allocate_bin_bo(drm);
408 clk_disable_unprepare(v3d->clk);
412 /* Reset the binner overflow address/size at setup, to be sure
413 * we don't reuse an old one.
415 V3D_WRITE(V3D_BPOA, 0);
416 V3D_WRITE(V3D_BPOS, 0);
418 vc4_v3d_init_hw(drm);
420 ret = drm_irq_install(drm, platform_get_irq(pdev, 0));
422 DRM_ERROR("Failed to install IRQ handler\n");
426 pm_runtime_set_active(dev);
427 pm_runtime_use_autosuspend(dev);
428 pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
429 pm_runtime_enable(dev);
431 vc4_debugfs_add_file(drm, "v3d_ident", vc4_v3d_debugfs_ident, NULL);
432 vc4_debugfs_add_regset32(drm, "v3d_regs", &v3d->regset);
437 static void vc4_v3d_unbind(struct device *dev, struct device *master,
440 struct drm_device *drm = dev_get_drvdata(master);
441 struct vc4_dev *vc4 = to_vc4_dev(drm);
443 pm_runtime_disable(dev);
445 drm_irq_uninstall(drm);
447 /* Disable the binner's overflow memory address, so the next
448 * driver probe (if any) doesn't try to reuse our old
451 V3D_WRITE(V3D_BPOA, 0);
452 V3D_WRITE(V3D_BPOS, 0);
457 static const struct dev_pm_ops vc4_v3d_pm_ops = {
458 SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
461 static const struct component_ops vc4_v3d_ops = {
462 .bind = vc4_v3d_bind,
463 .unbind = vc4_v3d_unbind,
466 static int vc4_v3d_dev_probe(struct platform_device *pdev)
468 return component_add(&pdev->dev, &vc4_v3d_ops);
471 static int vc4_v3d_dev_remove(struct platform_device *pdev)
473 component_del(&pdev->dev, &vc4_v3d_ops);
477 static const struct of_device_id vc4_v3d_dt_match[] = {
478 { .compatible = "brcm,bcm2835-v3d" },
479 { .compatible = "brcm,cygnus-v3d" },
480 { .compatible = "brcm,vc4-v3d" },
484 struct platform_driver vc4_v3d_driver = {
485 .probe = vc4_v3d_dev_probe,
486 .remove = vc4_v3d_dev_remove,
489 .of_match_table = vc4_v3d_dt_match,
490 .pm = &vc4_v3d_pm_ops,