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/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>
24 #define HOST1X_WAIT_SYNCPT_OFFSET 0x8
26 struct host1x_job *host1x_job_alloc(struct host1x_channel *ch,
27 u32 num_cmdbufs, u32 num_relocs,
30 struct host1x_job *job = NULL;
31 unsigned int num_unpins = num_relocs;
36 enable_firewall = IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL) && !skip_firewall;
39 num_unpins += num_cmdbufs;
41 /* Check that we're not going to overflow */
42 total = sizeof(struct host1x_job) +
43 (u64)num_relocs * sizeof(struct host1x_reloc) +
44 (u64)num_unpins * sizeof(struct host1x_job_unpin_data) +
45 (u64)num_cmdbufs * sizeof(struct host1x_job_cmd) +
46 (u64)num_unpins * sizeof(dma_addr_t) +
47 (u64)num_unpins * sizeof(u32 *);
48 if (total > ULONG_MAX)
51 mem = job = kzalloc(total, GFP_KERNEL);
55 job->enable_firewall = enable_firewall;
60 /* Redistribute memory to the structs */
61 mem += sizeof(struct host1x_job);
62 job->relocs = num_relocs ? mem : NULL;
63 mem += num_relocs * sizeof(struct host1x_reloc);
64 job->unpins = num_unpins ? mem : NULL;
65 mem += num_unpins * sizeof(struct host1x_job_unpin_data);
66 job->cmds = num_cmdbufs ? mem : NULL;
67 mem += num_cmdbufs * sizeof(struct host1x_job_cmd);
68 job->addr_phys = num_unpins ? mem : NULL;
70 job->reloc_addr_phys = job->addr_phys;
71 job->gather_addr_phys = &job->addr_phys[num_relocs];
75 EXPORT_SYMBOL(host1x_job_alloc);
77 struct host1x_job *host1x_job_get(struct host1x_job *job)
82 EXPORT_SYMBOL(host1x_job_get);
84 static void job_free(struct kref *ref)
86 struct host1x_job *job = container_of(ref, struct host1x_job, ref);
93 * remove_callback is atomic w.r.t. fence signaling, so
94 * after the call returns, we know that the callback is not
95 * in execution, and the fence can be safely freed.
97 dma_fence_remove_callback(job->fence, &job->fence_cb);
98 dma_fence_put(job->fence);
102 host1x_syncpt_put(job->syncpt);
107 void host1x_job_put(struct host1x_job *job)
109 kref_put(&job->ref, job_free);
111 EXPORT_SYMBOL(host1x_job_put);
113 void host1x_job_add_gather(struct host1x_job *job, struct host1x_bo *bo,
114 unsigned int words, unsigned int offset)
116 struct host1x_job_gather *gather = &job->cmds[job->num_cmds].gather;
118 gather->words = words;
120 gather->offset = offset;
124 EXPORT_SYMBOL(host1x_job_add_gather);
126 void host1x_job_add_wait(struct host1x_job *job, u32 id, u32 thresh,
127 bool relative, u32 next_class)
129 struct host1x_job_cmd *cmd = &job->cmds[job->num_cmds];
133 cmd->wait.threshold = thresh;
134 cmd->wait.next_class = next_class;
135 cmd->wait.relative = relative;
139 EXPORT_SYMBOL(host1x_job_add_wait);
141 static unsigned int pin_job(struct host1x *host, struct host1x_job *job)
143 unsigned long mask = HOST1X_RELOC_READ | HOST1X_RELOC_WRITE;
144 struct host1x_client *client = job->client;
145 struct device *dev = client->dev;
146 struct host1x_job_gather *g;
152 for (i = 0; i < job->num_relocs; i++) {
153 struct host1x_reloc *reloc = &job->relocs[i];
154 enum dma_data_direction direction;
155 struct host1x_bo_mapping *map;
156 struct host1x_bo *bo;
158 reloc->target.bo = host1x_bo_get(reloc->target.bo);
159 if (!reloc->target.bo) {
164 bo = reloc->target.bo;
166 switch (reloc->flags & mask) {
167 case HOST1X_RELOC_READ:
168 direction = DMA_TO_DEVICE;
171 case HOST1X_RELOC_WRITE:
172 direction = DMA_FROM_DEVICE;
175 case HOST1X_RELOC_READ | HOST1X_RELOC_WRITE:
176 direction = DMA_BIDIRECTIONAL;
184 map = host1x_bo_pin(dev, bo, direction, NULL);
191 * host1x clients are generally not able to do scatter-gather themselves, so fail
192 * if the buffer is discontiguous and we fail to map its SG table to a single
193 * contiguous chunk of I/O virtual memory.
195 if (map->chunks > 1) {
200 job->addr_phys[job->num_unpins] = map->phys;
201 job->unpins[job->num_unpins].map = map;
206 * We will copy gathers BO content later, so there is no need to
209 if (job->enable_firewall)
212 for (i = 0; i < job->num_cmds; i++) {
213 struct host1x_bo_mapping *map;
214 size_t gather_size = 0;
215 struct scatterlist *sg;
220 if (job->cmds[i].is_wait)
223 g = &job->cmds[i].gather;
225 g->bo = host1x_bo_get(g->bo);
231 map = host1x_bo_pin(host->dev, g->bo, DMA_TO_DEVICE, NULL);
238 for_each_sgtable_sg(map->sgt, sg, j)
239 gather_size += sg->length;
241 gather_size = iova_align(&host->iova, gather_size);
243 shift = iova_shift(&host->iova);
244 alloc = alloc_iova(&host->iova, gather_size >> shift,
245 host->iova_end >> shift, true);
251 err = iommu_map_sgtable(host->domain, iova_dma_addr(&host->iova, alloc),
252 map->sgt, IOMMU_READ);
254 __free_iova(&host->iova, alloc);
259 map->phys = iova_dma_addr(&host->iova, alloc);
260 map->size = gather_size;
263 job->addr_phys[job->num_unpins] = map->phys;
264 job->unpins[job->num_unpins].map = map;
267 job->gather_addr_phys[i] = map->phys;
273 host1x_bo_put(g->bo);
275 host1x_job_unpin(job);
279 static int do_relocs(struct host1x_job *job, struct host1x_job_gather *g)
281 void *cmdbuf_addr = NULL;
282 struct host1x_bo *cmdbuf = g->bo;
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;
292 /* skip all other gathers */
293 if (cmdbuf != reloc->cmdbuf.bo)
296 if (job->enable_firewall) {
297 target = (u32 *)job->gather_copy_mapped +
298 reloc->cmdbuf.offset / sizeof(u32) +
299 g->offset / sizeof(u32);
304 cmdbuf_addr = host1x_bo_mmap(cmdbuf);
306 if (unlikely(!cmdbuf_addr)) {
307 pr_err("Could not map cmdbuf for relocation\n");
312 target = cmdbuf_addr + reloc->cmdbuf.offset;
314 *target = reloc_addr;
318 host1x_bo_munmap(cmdbuf, cmdbuf_addr);
323 static bool check_reloc(struct host1x_reloc *reloc, struct host1x_bo *cmdbuf,
326 offset *= sizeof(u32);
328 if (reloc->cmdbuf.bo != cmdbuf || reloc->cmdbuf.offset != offset)
331 /* relocation shift value validation isn't implemented yet */
338 struct host1x_firewall {
339 struct host1x_job *job;
342 unsigned int num_relocs;
343 struct host1x_reloc *reloc;
345 struct host1x_bo *cmdbuf;
355 static int check_register(struct host1x_firewall *fw, unsigned long offset)
357 if (!fw->job->is_addr_reg)
360 if (fw->job->is_addr_reg(fw->dev, fw->class, offset)) {
364 if (!check_reloc(fw->reloc, fw->cmdbuf, fw->offset))
374 static int check_class(struct host1x_firewall *fw, u32 class)
376 if (!fw->job->is_valid_class) {
377 if (fw->class != class)
380 if (!fw->job->is_valid_class(fw->class))
387 static int check_mask(struct host1x_firewall *fw)
398 ret = check_register(fw, reg);
412 static int check_incr(struct host1x_firewall *fw)
414 u32 count = fw->count;
422 ret = check_register(fw, reg);
435 static int check_nonincr(struct host1x_firewall *fw)
437 u32 count = fw->count;
444 ret = check_register(fw, fw->reg);
456 static int validate(struct host1x_firewall *fw, struct host1x_job_gather *g)
458 u32 *cmdbuf_base = (u32 *)fw->job->gather_copy_mapped +
459 (g->offset / sizeof(u32));
460 u32 job_class = fw->class;
463 fw->words = g->words;
467 while (fw->words && !err) {
468 u32 word = cmdbuf_base[fw->offset];
469 u32 opcode = (word & 0xf0000000) >> 28;
479 fw->class = word >> 6 & 0x3ff;
480 fw->mask = word & 0x3f;
481 fw->reg = word >> 16 & 0xfff;
482 err = check_class(fw, job_class);
484 err = check_mask(fw);
489 fw->reg = word >> 16 & 0xfff;
490 fw->count = word & 0xffff;
491 err = check_incr(fw);
497 fw->reg = word >> 16 & 0xfff;
498 fw->count = word & 0xffff;
499 err = check_nonincr(fw);
505 fw->mask = word & 0xffff;
506 fw->reg = word >> 16 & 0xfff;
507 err = check_mask(fw);
524 static inline int copy_gathers(struct device *host, struct host1x_job *job,
527 struct host1x_firewall fw;
534 fw.reloc = job->relocs;
535 fw.num_relocs = job->num_relocs;
536 fw.class = job->class;
538 for (i = 0; i < job->num_cmds; i++) {
539 struct host1x_job_gather *g;
541 if (job->cmds[i].is_wait)
544 g = &job->cmds[i].gather;
546 size += g->words * sizeof(u32);
550 * Try a non-blocking allocation from a higher priority pools first,
551 * as awaiting for the allocation here is a major performance hit.
553 job->gather_copy_mapped = dma_alloc_wc(host, size, &job->gather_copy,
556 /* the higher priority allocation failed, try the generic-blocking */
557 if (!job->gather_copy_mapped)
558 job->gather_copy_mapped = dma_alloc_wc(host, size,
561 if (!job->gather_copy_mapped)
564 job->gather_copy_size = size;
566 for (i = 0; i < job->num_cmds; i++) {
567 struct host1x_job_gather *g;
570 if (job->cmds[i].is_wait)
572 g = &job->cmds[i].gather;
574 /* Copy the gather */
575 gather = host1x_bo_mmap(g->bo);
576 memcpy(job->gather_copy_mapped + offset, gather + g->offset,
577 g->words * sizeof(u32));
578 host1x_bo_munmap(g->bo, gather);
580 /* Store the location in the buffer */
581 g->base = job->gather_copy;
584 /* Validate the job */
585 if (validate(&fw, g))
588 offset += g->words * sizeof(u32);
591 /* No relocs should remain at this point */
598 int host1x_job_pin(struct host1x_job *job, struct device *dev)
602 struct host1x *host = dev_get_drvdata(dev->parent);
605 err = pin_job(host, job);
609 if (job->enable_firewall) {
610 err = copy_gathers(host->dev, job, dev);
616 for (i = 0; i < job->num_cmds; i++) {
617 struct host1x_job_gather *g;
619 if (job->cmds[i].is_wait)
621 g = &job->cmds[i].gather;
623 /* process each gather mem only once */
627 /* copy_gathers() sets gathers base if firewall is enabled */
628 if (!job->enable_firewall)
629 g->base = job->gather_addr_phys[i];
631 for (j = i + 1; j < job->num_cmds; j++) {
632 if (!job->cmds[j].is_wait &&
633 job->cmds[j].gather.bo == g->bo) {
634 job->cmds[j].gather.handled = true;
635 job->cmds[j].gather.base = g->base;
639 err = do_relocs(job, g);
646 host1x_job_unpin(job);
651 EXPORT_SYMBOL(host1x_job_pin);
653 void host1x_job_unpin(struct host1x_job *job)
655 struct host1x *host = dev_get_drvdata(job->channel->dev->parent);
658 for (i = 0; i < job->num_unpins; i++) {
659 struct host1x_bo_mapping *map = job->unpins[i].map;
660 struct host1x_bo *bo = map->bo;
662 if (!job->enable_firewall && map->size && host->domain) {
663 iommu_unmap(host->domain, job->addr_phys[i], map->size);
664 free_iova(&host->iova, iova_pfn(&host->iova, job->addr_phys[i]));
667 host1x_bo_unpin(map);
673 if (job->gather_copy_size)
674 dma_free_wc(host->dev, job->gather_copy_size,
675 job->gather_copy_mapped, job->gather_copy);
677 EXPORT_SYMBOL(host1x_job_unpin);
680 * Debug routine used to dump job entries
682 void host1x_job_dump(struct device *dev, struct host1x_job *job)
684 dev_dbg(dev, " SYNCPT_ID %d\n", job->syncpt->id);
685 dev_dbg(dev, " SYNCPT_VAL %d\n", job->syncpt_end);
686 dev_dbg(dev, " FIRST_GET 0x%x\n", job->first_get);
687 dev_dbg(dev, " TIMEOUT %d\n", job->timeout);
688 dev_dbg(dev, " NUM_SLOTS %d\n", job->num_slots);
689 dev_dbg(dev, " NUM_HANDLES %d\n", job->num_unpins);