1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (c) 2010-2015, NVIDIA Corporation.
8 #include <linux/dma-mapping.h>
10 #include <linux/host1x.h>
11 #include <linux/kref.h>
12 #include <linux/module.h>
13 #include <linux/scatterlist.h>
14 #include <linux/slab.h>
15 #include <linux/vmalloc.h>
16 #include <trace/events/host1x.h>
23 #define HOST1X_WAIT_SYNCPT_OFFSET 0x8
25 struct host1x_job *host1x_job_alloc(struct host1x_channel *ch,
26 u32 num_cmdbufs, u32 num_relocs)
28 struct host1x_job *job = NULL;
29 unsigned int num_unpins = num_cmdbufs + num_relocs;
33 /* Check that we're not going to overflow */
34 total = sizeof(struct host1x_job) +
35 (u64)num_relocs * sizeof(struct host1x_reloc) +
36 (u64)num_unpins * sizeof(struct host1x_job_unpin_data) +
37 (u64)num_cmdbufs * sizeof(struct host1x_job_gather) +
38 (u64)num_unpins * sizeof(dma_addr_t) +
39 (u64)num_unpins * sizeof(u32 *);
40 if (total > ULONG_MAX)
43 mem = job = kzalloc(total, GFP_KERNEL);
50 /* Redistribute memory to the structs */
51 mem += sizeof(struct host1x_job);
52 job->relocs = num_relocs ? mem : NULL;
53 mem += num_relocs * sizeof(struct host1x_reloc);
54 job->unpins = num_unpins ? mem : NULL;
55 mem += num_unpins * sizeof(struct host1x_job_unpin_data);
56 job->gathers = num_cmdbufs ? mem : NULL;
57 mem += num_cmdbufs * sizeof(struct host1x_job_gather);
58 job->addr_phys = num_unpins ? mem : NULL;
60 job->reloc_addr_phys = job->addr_phys;
61 job->gather_addr_phys = &job->addr_phys[num_relocs];
65 EXPORT_SYMBOL(host1x_job_alloc);
67 struct host1x_job *host1x_job_get(struct host1x_job *job)
72 EXPORT_SYMBOL(host1x_job_get);
74 static void job_free(struct kref *ref)
76 struct host1x_job *job = container_of(ref, struct host1x_job, ref);
81 void host1x_job_put(struct host1x_job *job)
83 kref_put(&job->ref, job_free);
85 EXPORT_SYMBOL(host1x_job_put);
87 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo,
88 unsigned int words, unsigned int offset)
90 struct host1x_job_gather *gather = &job->gathers[job->num_gathers];
92 gather->words = words;
94 gather->offset = offset;
98 EXPORT_SYMBOL(host1x_job_add_gather);
100 static unsigned int pin_job(struct host1x *host, struct host1x_job *job)
107 for (i = 0; i < job->num_relocs; i++) {
108 struct host1x_reloc *reloc = &job->relocs[i];
109 struct sg_table *sgt;
110 dma_addr_t phys_addr;
112 reloc->target.bo = host1x_bo_get(reloc->target.bo);
113 if (!reloc->target.bo) {
118 phys_addr = host1x_bo_pin(reloc->target.bo, &sgt);
120 job->addr_phys[job->num_unpins] = phys_addr;
121 job->unpins[job->num_unpins].bo = reloc->target.bo;
122 job->unpins[job->num_unpins].sgt = sgt;
126 for (i = 0; i < job->num_gathers; i++) {
127 struct host1x_job_gather *g = &job->gathers[i];
128 size_t gather_size = 0;
129 struct scatterlist *sg;
130 struct sg_table *sgt;
131 dma_addr_t phys_addr;
136 g->bo = host1x_bo_get(g->bo);
142 phys_addr = host1x_bo_pin(g->bo, &sgt);
144 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && host->domain) {
145 for_each_sg(sgt->sgl, sg, sgt->nents, j)
146 gather_size += sg->length;
147 gather_size = iova_align(&host->iova, gather_size);
149 shift = iova_shift(&host->iova);
150 alloc = alloc_iova(&host->iova, gather_size >> shift,
151 host->iova_end >> shift, true);
157 err = iommu_map_sg(host->domain,
158 iova_dma_addr(&host->iova, alloc),
159 sgt->sgl, sgt->nents, IOMMU_READ);
161 __free_iova(&host->iova, alloc);
166 job->addr_phys[job->num_unpins] =
167 iova_dma_addr(&host->iova, alloc);
168 job->unpins[job->num_unpins].size = gather_size;
170 job->addr_phys[job->num_unpins] = phys_addr;
173 job->gather_addr_phys[i] = job->addr_phys[job->num_unpins];
175 job->unpins[job->num_unpins].bo = g->bo;
176 job->unpins[job->num_unpins].sgt = sgt;
183 host1x_job_unpin(job);
187 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g)
190 void *cmdbuf_page_addr = NULL;
191 struct host1x_bo *cmdbuf = g->bo;
194 /* pin & patch the relocs for one gather */
195 for (i = 0; i < job->num_relocs; i++) {
196 struct host1x_reloc *reloc = &job->relocs[i];
197 u32 reloc_addr = (job->reloc_addr_phys[i] +
198 reloc->target.offset) >> reloc->shift;
201 /* skip all other gathers */
202 if (cmdbuf != reloc->cmdbuf.bo)
205 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
206 target = (u32 *)job->gather_copy_mapped +
207 reloc->cmdbuf.offset / sizeof(u32) +
208 g->offset / sizeof(u32);
212 if (last_page != reloc->cmdbuf.offset >> PAGE_SHIFT) {
213 if (cmdbuf_page_addr)
214 host1x_bo_kunmap(cmdbuf, last_page,
217 cmdbuf_page_addr = host1x_bo_kmap(cmdbuf,
218 reloc->cmdbuf.offset >> PAGE_SHIFT);
219 last_page = reloc->cmdbuf.offset >> PAGE_SHIFT;
221 if (unlikely(!cmdbuf_page_addr)) {
222 pr_err("Could not map cmdbuf for relocation\n");
227 target = cmdbuf_page_addr + (reloc->cmdbuf.offset & ~PAGE_MASK);
229 *target = reloc_addr;
232 if (cmdbuf_page_addr)
233 host1x_bo_kunmap(cmdbuf, last_page, cmdbuf_page_addr);
238 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf,
241 offset *= sizeof(u32);
243 if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset)
246 /* relocation shift value validation isn't implemented yet */
253 struct host1x_firewall {
254 struct host1x_job *job;
257 unsigned int num_relocs;
258 struct host1x_reloc *reloc;
260 struct host1x_bo *cmdbuf;
270 static int check_register(struct host1x_firewall *fw, unsigned long offset)
272 if (!fw->job->is_addr_reg)
275 if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) {
279 if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset))
289 static int check_class(struct host1x_firewall *fw, u32 class)
291 if (!fw->job->is_valid_class) {
292 if (fw->class != class)
295 if (!fw->job->is_valid_class(fw->class))
302 static int check_mask(struct host1x_firewall *fw)
313 ret = check_register(fw, reg);
327 static int check_incr(struct host1x_firewall *fw)
329 u32 count = fw->count;
337 ret = check_register(fw, reg);
350 static int check_nonincr(struct host1x_firewall *fw)
352 u32 count = fw->count;
359 ret = check_register(fw, fw->reg);
371 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g)
373 u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped +
374 (g->offset / sizeof(u32));
375 u32 job_class = fw->class;
378 fw->words = g->words;
382 while (fw->words && !err) {
383 u32 word = cmdbuf_base[fw->offset];
384 u32 opcode = (word & 0xf0000000) >> 28;
394 fw->class = word >> 6 & 0x3ff;
395 fw->mask = word & 0x3f;
396 fw->reg = word >> 16 & 0xfff;
397 err = check_class(fw, job_class);
399 err = check_mask(fw);
404 fw->reg = word >> 16 & 0xfff;
405 fw->count = word & 0xffff;
406 err = check_incr(fw);
412 fw->reg = word >> 16 & 0xfff;
413 fw->count = word & 0xffff;
414 err = check_nonincr(fw);
420 fw->mask = word & 0xffff;
421 fw->reg = word >> 16 & 0xfff;
422 err = check_mask(fw);
439 static inline int copy_gathers(struct host1x_job *job, struct device *dev)
441 struct host1x_firewall fw;
448 fw.reloc = job->relocs;
449 fw.num_relocs = job->num_relocs;
450 fw.class = job->class;
452 for (i = 0; i < job->num_gathers; i++) {
453 struct host1x_job_gather *g = &job->gathers[i];
455 size += g->words * sizeof(u32);
459 * Try a non-blocking allocation from a higher priority pools first,
460 * as awaiting for the allocation here is a major performance hit.
462 job->gather_copy_mapped = dma_alloc_wc(dev, size, &job->gather_copy,
465 /* the higher priority allocation failed, try the generic-blocking */
466 if (!job->gather_copy_mapped)
467 job->gather_copy_mapped = dma_alloc_wc(dev, size,
470 if (!job->gather_copy_mapped)
473 job->gather_copy_size = size;
475 for (i = 0; i < job->num_gathers; i++) {
476 struct host1x_job_gather *g = &job->gathers[i];
479 /* Copy the gather */
480 gather = host1x_bo_mmap(g->bo);
481 memcpy(job->gather_copy_mapped + offset, gather + g->offset,
482 g->words * sizeof(u32));
483 host1x_bo_munmap(g->bo, gather);
485 /* Store the location in the buffer */
486 g->base = job->gather_copy;
489 /* Validate the job */
490 if (validate(&fw, g))
493 offset += g->words * sizeof(u32);
496 /* No relocs should remain at this point */
503 int host1x_job_pin(struct host1x_job *job, struct device *dev)
507 struct host1x *host = dev_get_drvdata(dev->parent);
510 err = pin_job(host, job);
514 if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL)) {
515 err = copy_gathers(job, dev);
521 for (i = 0; i < job->num_gathers; i++) {
522 struct host1x_job_gather *g = &job->gathers[i];
524 /* process each gather mem only once */
528 /* copy_gathers() sets gathers base if firewall is enabled */
529 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
530 g->base = job->gather_addr_phys[i];
532 for (j = i + 1; j < job->num_gathers; j++) {
533 if (job->gathers[j].bo == g->bo) {
534 job->gathers[j].handled = true;
535 job->gathers[j].base = g->base;
539 err = do_relocs(job, g);
546 host1x_job_unpin(job);
551 EXPORT_SYMBOL(host1x_job_pin);
553 void host1x_job_unpin(struct host1x_job *job)
555 struct host1x *host = dev_get_drvdata(job->channel->dev->parent);
558 for (i = 0; i < job->num_unpins; i++) {
559 struct host1x_job_unpin_data *unpin = &job->unpins[i];
561 if (!IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) &&
562 unpin->size && host->domain) {
563 iommu_unmap(host->domain, job->addr_phys[i],
565 free_iova(&host->iova,
566 iova_pfn(&host->iova, job->addr_phys[i]));
569 host1x_bo_unpin(unpin->bo, unpin->sgt);
570 host1x_bo_put(unpin->bo);
575 if (job->gather_copy_size)
576 dma_free_wc(job->channel->dev, job->gather_copy_size,
577 job->gather_copy_mapped, job->gather_copy);
579 EXPORT_SYMBOL(host1x_job_unpin);
582 * Debug routine used to dump job entries
584 void host1x_job_dump(struct device *dev, struct host1x_job *job)
586 dev_dbg(dev, " SYNCPT_ID %d\n", job->syncpt_id);
587 dev_dbg(dev, " SYNCPT_VAL %d\n", job->syncpt_end);
588 dev_dbg(dev, " FIRST_GET 0x%x\n", job->first_get);
589 dev_dbg(dev, " TIMEOUT %d\n", job->timeout);
590 dev_dbg(dev, " NUM_SLOTS %d\n", job->num_slots);
591 dev_dbg(dev, " NUM_HANDLES %d\n", job->num_unpins);