xref: /openbmc/linux/drivers/gpu/host1x/job.c (revision 43ffe817)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Tegra host1x Job
4  *
5  * Copyright (c) 2010-2015, NVIDIA Corporation.
6  */
7 
8 #include <linux/dma-mapping.h>
9 #include <linux/err.h>
10 #include <linux/host1x.h>
11 #include <linux/iommu.h>
12 #include <linux/kref.h>
13 #include <linux/module.h>
14 #include <linux/scatterlist.h>
15 #include <linux/slab.h>
16 #include <linux/vmalloc.h>
17 #include <trace/events/host1x.h>
18 
19 #include "channel.h"
20 #include "dev.h"
21 #include "job.h"
22 #include "syncpt.h"
23 
24 #define HOST1X_WAIT_SYNCPT_OFFSET 0x8
25 
26 struct host1x_job *host1x_job_alloc(struct host1x_channel *ch,
27 				    u32 num_cmdbufs, u32 num_relocs)
28 {
29 	struct host1x_job *job = NULL;
30 	unsigned int num_unpins = num_relocs;
31 	u64 total;
32 	void *mem;
33 
34 	if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
35 		num_unpins += num_cmdbufs;
36 
37 	/* Check that we're not going to overflow */
38 	total = sizeof(struct host1x_job) +
39 		(u64)num_relocs * sizeof(struct host1x_reloc) +
40 		(u64)num_unpins * sizeof(struct host1x_job_unpin_data) +
41 		(u64)num_cmdbufs * sizeof(struct host1x_job_gather) +
42 		(u64)num_unpins * sizeof(dma_addr_t) +
43 		(u64)num_unpins * sizeof(u32 *);
44 	if (total > ULONG_MAX)
45 		return NULL;
46 
47 	mem = job = kzalloc(total, GFP_KERNEL);
48 	if (!job)
49 		return NULL;
50 
51 	kref_init(&job->ref);
52 	job->channel = ch;
53 
54 	/* Redistribute memory to the structs  */
55 	mem += sizeof(struct host1x_job);
56 	job->relocs = num_relocs ? mem : NULL;
57 	mem += num_relocs * sizeof(struct host1x_reloc);
58 	job->unpins = num_unpins ? mem : NULL;
59 	mem += num_unpins * sizeof(struct host1x_job_unpin_data);
60 	job->gathers = num_cmdbufs ? mem : NULL;
61 	mem += num_cmdbufs * sizeof(struct host1x_job_gather);
62 	job->addr_phys = num_unpins ? mem : NULL;
63 
64 	job->reloc_addr_phys = job->addr_phys;
65 	job->gather_addr_phys = &job->addr_phys[num_relocs];
66 
67 	return job;
68 }
69 EXPORT_SYMBOL(host1x_job_alloc);
70 
71 struct host1x_job *host1x_job_get(struct host1x_job *job)
72 {
73 	kref_get(&job->ref);
74 	return job;
75 }
76 EXPORT_SYMBOL(host1x_job_get);
77 
78 static void job_free(struct kref *ref)
79 {
80 	struct host1x_job *job = container_of(ref, struct host1x_job, ref);
81 
82 	kfree(job);
83 }
84 
85 void host1x_job_put(struct host1x_job *job)
86 {
87 	kref_put(&job->ref, job_free);
88 }
89 EXPORT_SYMBOL(host1x_job_put);
90 
91 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo,
92 			   unsigned int words, unsigned int offset)
93 {
94 	struct host1x_job_gather *gather = &job->gathers[job->num_gathers];
95 
96 	gather->words = words;
97 	gather->bo = bo;
98 	gather->offset = offset;
99 
100 	job->num_gathers++;
101 }
102 EXPORT_SYMBOL(host1x_job_add_gather);
103 
104 static unsigned int pin_job(struct host1x *host, struct host1x_job *job)
105 {
106 	struct host1x_client *client = job->client;
107 	struct device *dev = client->dev;
108 	struct host1x_job_gather *g;
109 	struct iommu_domain *domain;
110 	unsigned int i;
111 	int err;
112 
113 	domain = iommu_get_domain_for_dev(dev);
114 	job->num_unpins = 0;
115 
116 	for (i = 0; i < job->num_relocs; i++) {
117 		struct host1x_reloc *reloc = &job->relocs[i];
118 		dma_addr_t phys_addr, *phys;
119 		struct sg_table *sgt;
120 
121 		reloc->target.bo = host1x_bo_get(reloc->target.bo);
122 		if (!reloc->target.bo) {
123 			err = -EINVAL;
124 			goto unpin;
125 		}
126 
127 		/*
128 		 * If the client device is not attached to an IOMMU, the
129 		 * physical address of the buffer object can be used.
130 		 *
131 		 * Similarly, when an IOMMU domain is shared between all
132 		 * host1x clients, the IOVA is already available, so no
133 		 * need to map the buffer object again.
134 		 *
135 		 * XXX Note that this isn't always safe to do because it
136 		 * relies on an assumption that no cache maintenance is
137 		 * needed on the buffer objects.
138 		 */
139 		if (!domain || client->group)
140 			phys = &phys_addr;
141 		else
142 			phys = NULL;
143 
144 		sgt = host1x_bo_pin(dev, reloc->target.bo, phys);
145 		if (IS_ERR(sgt)) {
146 			err = PTR_ERR(sgt);
147 			goto unpin;
148 		}
149 
150 		if (sgt) {
151 			unsigned long mask = HOST1X_RELOC_READ |
152 					     HOST1X_RELOC_WRITE;
153 			enum dma_data_direction dir;
154 
155 			switch (reloc->flags & mask) {
156 			case HOST1X_RELOC_READ:
157 				dir = DMA_TO_DEVICE;
158 				break;
159 
160 			case HOST1X_RELOC_WRITE:
161 				dir = DMA_FROM_DEVICE;
162 				break;
163 
164 			case HOST1X_RELOC_READ | HOST1X_RELOC_WRITE:
165 				dir = DMA_BIDIRECTIONAL;
166 				break;
167 
168 			default:
169 				err = -EINVAL;
170 				goto unpin;
171 			}
172 
173 			err = dma_map_sgtable(dev, sgt, dir, 0);
174 			if (err)
175 				goto unpin;
176 
177 			job->unpins[job->num_unpins].dev = dev;
178 			job->unpins[job->num_unpins].dir = dir;
179 			phys_addr = sg_dma_address(sgt->sgl);
180 		}
181 
182 		job->addr_phys[job->num_unpins] = phys_addr;
183 		job->unpins[job->num_unpins].bo = reloc->target.bo;
184 		job->unpins[job->num_unpins].sgt = sgt;
185 		job->num_unpins++;
186 	}
187 
188 	/*
189 	 * We will copy gathers BO content later, so there is no need to
190 	 * hold and pin them.
191 	 */
192 	if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
193 		return 0;
194 
195 	for (i = 0; i < job->num_gathers; i++) {
196 		size_t gather_size = 0;
197 		struct scatterlist *sg;
198 		struct sg_table *sgt;
199 		dma_addr_t phys_addr;
200 		unsigned long shift;
201 		struct iova *alloc;
202 		dma_addr_t *phys;
203 		unsigned int j;
204 
205 		g = &job->gathers[i];
206 		g->bo = host1x_bo_get(g->bo);
207 		if (!g->bo) {
208 			err = -EINVAL;
209 			goto unpin;
210 		}
211 
212 		/**
213 		 * If the host1x is not attached to an IOMMU, there is no need
214 		 * to map the buffer object for the host1x, since the physical
215 		 * address can simply be used.
216 		 */
217 		if (!iommu_get_domain_for_dev(host->dev))
218 			phys = &phys_addr;
219 		else
220 			phys = NULL;
221 
222 		sgt = host1x_bo_pin(host->dev, g->bo, phys);
223 		if (IS_ERR(sgt)) {
224 			err = PTR_ERR(sgt);
225 			goto put;
226 		}
227 
228 		if (host->domain) {
229 			for_each_sgtable_sg(sgt, sg, j)
230 				gather_size += sg->length;
231 			gather_size = iova_align(&host->iova, gather_size);
232 
233 			shift = iova_shift(&host->iova);
234 			alloc = alloc_iova(&host->iova, gather_size >> shift,
235 					   host->iova_end >> shift, true);
236 			if (!alloc) {
237 				err = -ENOMEM;
238 				goto put;
239 			}
240 
241 			err = iommu_map_sgtable(host->domain,
242 					iova_dma_addr(&host->iova, alloc),
243 					sgt, IOMMU_READ);
244 			if (err == 0) {
245 				__free_iova(&host->iova, alloc);
246 				err = -EINVAL;
247 				goto put;
248 			}
249 
250 			job->unpins[job->num_unpins].size = gather_size;
251 			phys_addr = iova_dma_addr(&host->iova, alloc);
252 		} else if (sgt) {
253 			err = dma_map_sgtable(host->dev, sgt, DMA_TO_DEVICE, 0);
254 			if (err)
255 				goto put;
256 
257 			job->unpins[job->num_unpins].dir = DMA_TO_DEVICE;
258 			job->unpins[job->num_unpins].dev = host->dev;
259 			phys_addr = sg_dma_address(sgt->sgl);
260 		}
261 
262 		job->addr_phys[job->num_unpins] = phys_addr;
263 		job->gather_addr_phys[i] = phys_addr;
264 
265 		job->unpins[job->num_unpins].bo = g->bo;
266 		job->unpins[job->num_unpins].sgt = sgt;
267 		job->num_unpins++;
268 	}
269 
270 	return 0;
271 
272 put:
273 	host1x_bo_put(g->bo);
274 unpin:
275 	host1x_job_unpin(job);
276 	return err;
277 }
278 
279 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g)
280 {
281 	void *cmdbuf_addr = NULL;
282 	struct host1x_bo *cmdbuf = g->bo;
283 	unsigned int i;
284 
285 	/* pin & patch the relocs for one gather */
286 	for (i = 0; i < job->num_relocs; i++) {
287 		struct host1x_reloc *reloc = &job->relocs[i];
288 		u32 reloc_addr = (job->reloc_addr_phys[i] +
289 				  reloc->target.offset) >> reloc->shift;
290 		u32 *target;
291 
292 		/* skip all other gathers */
293 		if (cmdbuf != reloc->cmdbuf.bo)
294 			continue;
295 
296 		if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
297 			target = (u32 *)job->gather_copy_mapped +
298 					reloc->cmdbuf.offset / sizeof(u32) +
299 						g->offset / sizeof(u32);
300 			goto patch_reloc;
301 		}
302 
303 		if (!cmdbuf_addr) {
304 			cmdbuf_addr = host1x_bo_mmap(cmdbuf);
305 
306 			if (unlikely(!cmdbuf_addr)) {
307 				pr_err("Could not map cmdbuf for relocation\n");
308 				return -ENOMEM;
309 			}
310 		}
311 
312 		target = cmdbuf_addr + reloc->cmdbuf.offset;
313 patch_reloc:
314 		*target = reloc_addr;
315 	}
316 
317 	if (cmdbuf_addr)
318 		host1x_bo_munmap(cmdbuf, cmdbuf_addr);
319 
320 	return 0;
321 }
322 
323 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf,
324 			unsigned int offset)
325 {
326 	offset *= sizeof(u32);
327 
328 	if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset)
329 		return false;
330 
331 	/* relocation shift value validation isn't implemented yet */
332 	if (reloc->shift)
333 		return false;
334 
335 	return true;
336 }
337 
338 struct host1x_firewall {
339 	struct host1x_job *job;
340 	struct device *dev;
341 
342 	unsigned int num_relocs;
343 	struct host1x_reloc *reloc;
344 
345 	struct host1x_bo *cmdbuf;
346 	unsigned int offset;
347 
348 	u32 words;
349 	u32 class;
350 	u32 reg;
351 	u32 mask;
352 	u32 count;
353 };
354 
355 static int check_register(struct host1x_firewall *fw, unsigned long offset)
356 {
357 	if (!fw->job->is_addr_reg)
358 		return 0;
359 
360 	if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) {
361 		if (!fw->num_relocs)
362 			return -EINVAL;
363 
364 		if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset))
365 			return -EINVAL;
366 
367 		fw->num_relocs--;
368 		fw->reloc++;
369 	}
370 
371 	return 0;
372 }
373 
374 static int check_class(struct host1x_firewall *fw, u32 class)
375 {
376 	if (!fw->job->is_valid_class) {
377 		if (fw->class != class)
378 			return -EINVAL;
379 	} else {
380 		if (!fw->job->is_valid_class(fw->class))
381 			return -EINVAL;
382 	}
383 
384 	return 0;
385 }
386 
387 static int check_mask(struct host1x_firewall *fw)
388 {
389 	u32 mask = fw->mask;
390 	u32 reg = fw->reg;
391 	int ret;
392 
393 	while (mask) {
394 		if (fw->words == 0)
395 			return -EINVAL;
396 
397 		if (mask & 1) {
398 			ret = check_register(fw, reg);
399 			if (ret < 0)
400 				return ret;
401 
402 			fw->words--;
403 			fw->offset++;
404 		}
405 		mask >>= 1;
406 		reg++;
407 	}
408 
409 	return 0;
410 }
411 
412 static int check_incr(struct host1x_firewall *fw)
413 {
414 	u32 count = fw->count;
415 	u32 reg = fw->reg;
416 	int ret;
417 
418 	while (count) {
419 		if (fw->words == 0)
420 			return -EINVAL;
421 
422 		ret = check_register(fw, reg);
423 		if (ret < 0)
424 			return ret;
425 
426 		reg++;
427 		fw->words--;
428 		fw->offset++;
429 		count--;
430 	}
431 
432 	return 0;
433 }
434 
435 static int check_nonincr(struct host1x_firewall *fw)
436 {
437 	u32 count = fw->count;
438 	int ret;
439 
440 	while (count) {
441 		if (fw->words == 0)
442 			return -EINVAL;
443 
444 		ret = check_register(fw, fw->reg);
445 		if (ret < 0)
446 			return ret;
447 
448 		fw->words--;
449 		fw->offset++;
450 		count--;
451 	}
452 
453 	return 0;
454 }
455 
456 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g)
457 {
458 	u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped +
459 		(g->offset / sizeof(u32));
460 	u32 job_class = fw->class;
461 	int err = 0;
462 
463 	fw->words = g->words;
464 	fw->cmdbuf = g->bo;
465 	fw->offset = 0;
466 
467 	while (fw->words && !err) {
468 		u32 word = cmdbuf_base[fw->offset];
469 		u32 opcode = (word & 0xf0000000) >> 28;
470 
471 		fw->mask = 0;
472 		fw->reg = 0;
473 		fw->count = 0;
474 		fw->words--;
475 		fw->offset++;
476 
477 		switch (opcode) {
478 		case 0:
479 			fw->class = word >> 6 & 0x3ff;
480 			fw->mask = word & 0x3f;
481 			fw->reg = word >> 16 & 0xfff;
482 			err = check_class(fw, job_class);
483 			if (!err)
484 				err = check_mask(fw);
485 			if (err)
486 				goto out;
487 			break;
488 		case 1:
489 			fw->reg = word >> 16 & 0xfff;
490 			fw->count = word & 0xffff;
491 			err = check_incr(fw);
492 			if (err)
493 				goto out;
494 			break;
495 
496 		case 2:
497 			fw->reg = word >> 16 & 0xfff;
498 			fw->count = word & 0xffff;
499 			err = check_nonincr(fw);
500 			if (err)
501 				goto out;
502 			break;
503 
504 		case 3:
505 			fw->mask = word & 0xffff;
506 			fw->reg = word >> 16 & 0xfff;
507 			err = check_mask(fw);
508 			if (err)
509 				goto out;
510 			break;
511 		case 4:
512 		case 14:
513 			break;
514 		default:
515 			err = -EINVAL;
516 			break;
517 		}
518 	}
519 
520 out:
521 	return err;
522 }
523 
524 static inline int copy_gathers(struct device *host, struct host1x_job *job,
525 			       struct device *dev)
526 {
527 	struct host1x_firewall fw;
528 	size_t size = 0;
529 	size_t offset = 0;
530 	unsigned int i;
531 
532 	fw.job = job;
533 	fw.dev = dev;
534 	fw.reloc = job->relocs;
535 	fw.num_relocs = job->num_relocs;
536 	fw.class = job->class;
537 
538 	for (i = 0; i < job->num_gathers; i++) {
539 		struct host1x_job_gather *g = &job->gathers[i];
540 
541 		size += g->words * sizeof(u32);
542 	}
543 
544 	/*
545 	 * Try a non-blocking allocation from a higher priority pools first,
546 	 * as awaiting for the allocation here is a major performance hit.
547 	 */
548 	job->gather_copy_mapped = dma_alloc_wc(host, size, &job->gather_copy,
549 					       GFP_NOWAIT);
550 
551 	/* the higher priority allocation failed, try the generic-blocking */
552 	if (!job->gather_copy_mapped)
553 		job->gather_copy_mapped = dma_alloc_wc(host, size,
554 						       &job->gather_copy,
555 						       GFP_KERNEL);
556 	if (!job->gather_copy_mapped)
557 		return -ENOMEM;
558 
559 	job->gather_copy_size = size;
560 
561 	for (i = 0; i < job->num_gathers; i++) {
562 		struct host1x_job_gather *g = &job->gathers[i];
563 		void *gather;
564 
565 		/* Copy the gather */
566 		gather = host1x_bo_mmap(g->bo);
567 		memcpy(job->gather_copy_mapped + offset, gather + g->offset,
568 		       g->words * sizeof(u32));
569 		host1x_bo_munmap(g->bo, gather);
570 
571 		/* Store the location in the buffer */
572 		g->base = job->gather_copy;
573 		g->offset = offset;
574 
575 		/* Validate the job */
576 		if (validate(&fw, g))
577 			return -EINVAL;
578 
579 		offset += g->words * sizeof(u32);
580 	}
581 
582 	/* No relocs should remain at this point */
583 	if (fw.num_relocs)
584 		return -EINVAL;
585 
586 	return 0;
587 }
588 
589 int host1x_job_pin(struct host1x_job *job, struct device *dev)
590 {
591 	int err;
592 	unsigned int i, j;
593 	struct host1x *host = dev_get_drvdata(dev->parent);
594 
595 	/* pin memory */
596 	err = pin_job(host, job);
597 	if (err)
598 		goto out;
599 
600 	if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
601 		err = copy_gathers(host->dev, job, dev);
602 		if (err)
603 			goto out;
604 	}
605 
606 	/* patch gathers */
607 	for (i = 0; i < job->num_gathers; i++) {
608 		struct host1x_job_gather *g = &job->gathers[i];
609 
610 		/* process each gather mem only once */
611 		if (g->handled)
612 			continue;
613 
614 		/* copy_gathers() sets gathers base if firewall is enabled */
615 		if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
616 			g->base = job->gather_addr_phys[i];
617 
618 		for (j = i + 1; j < job->num_gathers; j++) {
619 			if (job->gathers[j].bo == g->bo) {
620 				job->gathers[j].handled = true;
621 				job->gathers[j].base = g->base;
622 			}
623 		}
624 
625 		err = do_relocs(job, g);
626 		if (err)
627 			break;
628 	}
629 
630 out:
631 	if (err)
632 		host1x_job_unpin(job);
633 	wmb();
634 
635 	return err;
636 }
637 EXPORT_SYMBOL(host1x_job_pin);
638 
639 void host1x_job_unpin(struct host1x_job *job)
640 {
641 	struct host1x *host = dev_get_drvdata(job->channel->dev->parent);
642 	unsigned int i;
643 
644 	for (i = 0; i < job->num_unpins; i++) {
645 		struct host1x_job_unpin_data *unpin = &job->unpins[i];
646 		struct device *dev = unpin->dev ?: host->dev;
647 		struct sg_table *sgt = unpin->sgt;
648 
649 		if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) &&
650 		    unpin->size && host->domain) {
651 			iommu_unmap(host->domain, job->addr_phys[i],
652 				    unpin->size);
653 			free_iova(&host->iova,
654 				iova_pfn(&host->iova, job->addr_phys[i]));
655 		}
656 
657 		if (unpin->dev && sgt)
658 			dma_unmap_sgtable(unpin->dev, sgt, unpin->dir, 0);
659 
660 		host1x_bo_unpin(dev, unpin->bo, sgt);
661 		host1x_bo_put(unpin->bo);
662 	}
663 
664 	job->num_unpins = 0;
665 
666 	if (job->gather_copy_size)
667 		dma_free_wc(host->dev, job->gather_copy_size,
668 			    job->gather_copy_mapped, job->gather_copy);
669 }
670 EXPORT_SYMBOL(host1x_job_unpin);
671 
672 /*
673  * Debug routine used to dump job entries
674  */
675 void host1x_job_dump(struct device *dev, struct host1x_job *job)
676 {
677 	dev_dbg(dev, "    SYNCPT_ID   %d\n", job->syncpt_id);
678 	dev_dbg(dev, "    SYNCPT_VAL  %d\n", job->syncpt_end);
679 	dev_dbg(dev, "    FIRST_GET   0x%x\n", job->first_get);
680 	dev_dbg(dev, "    TIMEOUT     %d\n", job->timeout);
681 	dev_dbg(dev, "    NUM_SLOTS   %d\n", job->num_slots);
682 	dev_dbg(dev, "    NUM_HANDLES %d\n", job->num_unpins);
683 }
684