1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2020 NVIDIA Corporation */
4 #include <linux/dma-fence-array.h>
5 #include <linux/dma-mapping.h>
6 #include <linux/file.h>
7 #include <linux/host1x.h>
8 #include <linux/iommu.h>
9 #include <linux/kref.h>
10 #include <linux/list.h>
11 #include <linux/nospec.h>
12 #include <linux/pm_runtime.h>
13 #include <linux/scatterlist.h>
14 #include <linux/slab.h>
15 #include <linux/sync_file.h>
17 #include <drm/drm_drv.h>
18 #include <drm/drm_file.h>
19 #include <drm/drm_syncobj.h>
26 #define SUBMIT_ERR(context, fmt, ...) \
27 dev_err_ratelimited(context->client->base.dev, \
28 "%s: job submission failed: " fmt "\n", \
29 current->comm, ##__VA_ARGS__)
32 struct host1x_bo base;
38 dma_addr_t gather_data_dma;
39 size_t gather_data_words;
42 static struct host1x_bo *gather_bo_get(struct host1x_bo *host_bo)
44 struct gather_bo *bo = container_of(host_bo, struct gather_bo, base);
51 static void gather_bo_release(struct kref *ref)
53 struct gather_bo *bo = container_of(ref, struct gather_bo, ref);
55 dma_free_attrs(bo->dev, bo->gather_data_words * 4, bo->gather_data, bo->gather_data_dma,
60 static void gather_bo_put(struct host1x_bo *host_bo)
62 struct gather_bo *bo = container_of(host_bo, struct gather_bo, base);
64 kref_put(&bo->ref, gather_bo_release);
67 static struct host1x_bo_mapping *
68 gather_bo_pin(struct device *dev, struct host1x_bo *bo, enum dma_data_direction direction)
70 struct gather_bo *gather = container_of(bo, struct gather_bo, base);
71 struct host1x_bo_mapping *map;
74 map = kzalloc(sizeof(*map), GFP_KERNEL);
76 return ERR_PTR(-ENOMEM);
79 map->bo = host1x_bo_get(bo);
80 map->direction = direction;
83 map->sgt = kzalloc(sizeof(*map->sgt), GFP_KERNEL);
89 err = dma_get_sgtable(gather->dev, map->sgt, gather->gather_data, gather->gather_data_dma,
90 gather->gather_data_words * 4);
94 err = dma_map_sgtable(dev, map->sgt, direction, 0);
98 map->phys = sg_dma_address(map->sgt->sgl);
99 map->size = gather->gather_data_words * 4;
105 sg_free_table(map->sgt);
112 static void gather_bo_unpin(struct host1x_bo_mapping *map)
117 dma_unmap_sgtable(map->dev, map->sgt, map->direction, 0);
118 sg_free_table(map->sgt);
120 host1x_bo_put(map->bo);
125 static void *gather_bo_mmap(struct host1x_bo *host_bo)
127 struct gather_bo *bo = container_of(host_bo, struct gather_bo, base);
129 return bo->gather_data;
132 static void gather_bo_munmap(struct host1x_bo *host_bo, void *addr)
136 static const struct host1x_bo_ops gather_bo_ops = {
137 .get = gather_bo_get,
138 .put = gather_bo_put,
139 .pin = gather_bo_pin,
140 .unpin = gather_bo_unpin,
141 .mmap = gather_bo_mmap,
142 .munmap = gather_bo_munmap,
145 static struct tegra_drm_mapping *
146 tegra_drm_mapping_get(struct tegra_drm_context *context, u32 id)
148 struct tegra_drm_mapping *mapping;
150 xa_lock(&context->mappings);
152 mapping = xa_load(&context->mappings, id);
154 kref_get(&mapping->ref);
156 xa_unlock(&context->mappings);
161 static void *alloc_copy_user_array(void __user *from, size_t count, size_t size)
166 if (check_mul_overflow(count, size, ©_len))
167 return ERR_PTR(-EINVAL);
169 if (copy_len > 0x4000)
170 return ERR_PTR(-E2BIG);
172 data = vmemdup_user(from, copy_len);
174 return ERR_CAST(data);
179 static int submit_copy_gather_data(struct gather_bo **pbo, struct device *dev,
180 struct tegra_drm_context *context,
181 struct drm_tegra_channel_submit *args)
183 struct gather_bo *bo;
186 if (args->gather_data_words == 0) {
187 SUBMIT_ERR(context, "gather_data_words cannot be zero");
191 if (check_mul_overflow((size_t)args->gather_data_words, (size_t)4, ©_len)) {
192 SUBMIT_ERR(context, "gather_data_words is too large");
196 bo = kzalloc(sizeof(*bo), GFP_KERNEL);
198 SUBMIT_ERR(context, "failed to allocate memory for bo info");
202 host1x_bo_init(&bo->base, &gather_bo_ops);
206 bo->gather_data = dma_alloc_attrs(dev, copy_len, &bo->gather_data_dma,
207 GFP_KERNEL | __GFP_NOWARN, 0);
208 if (!bo->gather_data) {
209 SUBMIT_ERR(context, "failed to allocate memory for gather data");
214 if (copy_from_user(bo->gather_data, u64_to_user_ptr(args->gather_data_ptr), copy_len)) {
215 SUBMIT_ERR(context, "failed to copy gather data from userspace");
216 dma_free_attrs(dev, copy_len, bo->gather_data, bo->gather_data_dma, 0);
221 bo->gather_data_words = args->gather_data_words;
228 static int submit_write_reloc(struct tegra_drm_context *context, struct gather_bo *bo,
229 struct drm_tegra_submit_buf *buf, struct tegra_drm_mapping *mapping)
231 /* TODO check that target_offset is within bounds */
232 dma_addr_t iova = mapping->iova + buf->reloc.target_offset;
235 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
236 if (buf->flags & DRM_TEGRA_SUBMIT_RELOC_SECTOR_LAYOUT)
240 written_ptr = iova >> buf->reloc.shift;
242 if (buf->reloc.gather_offset_words >= bo->gather_data_words) {
244 "relocation has too large gather offset (%u vs gather length %zu)",
245 buf->reloc.gather_offset_words, bo->gather_data_words);
249 buf->reloc.gather_offset_words = array_index_nospec(buf->reloc.gather_offset_words,
250 bo->gather_data_words);
252 bo->gather_data[buf->reloc.gather_offset_words] = written_ptr;
257 static int submit_process_bufs(struct tegra_drm_context *context, struct gather_bo *bo,
258 struct drm_tegra_channel_submit *args,
259 struct tegra_drm_submit_data *job_data)
261 struct tegra_drm_used_mapping *mappings;
262 struct drm_tegra_submit_buf *bufs;
266 bufs = alloc_copy_user_array(u64_to_user_ptr(args->bufs_ptr), args->num_bufs,
269 SUBMIT_ERR(context, "failed to copy bufs array from userspace");
270 return PTR_ERR(bufs);
273 mappings = kcalloc(args->num_bufs, sizeof(*mappings), GFP_KERNEL);
275 SUBMIT_ERR(context, "failed to allocate memory for mapping info");
280 for (i = 0; i < args->num_bufs; i++) {
281 struct drm_tegra_submit_buf *buf = &bufs[i];
282 struct tegra_drm_mapping *mapping;
284 if (buf->flags & ~DRM_TEGRA_SUBMIT_RELOC_SECTOR_LAYOUT) {
285 SUBMIT_ERR(context, "invalid flag specified for buffer");
290 mapping = tegra_drm_mapping_get(context, buf->mapping);
292 SUBMIT_ERR(context, "invalid mapping ID '%u' for buffer", buf->mapping);
297 err = submit_write_reloc(context, bo, buf, mapping);
299 tegra_drm_mapping_put(mapping);
303 mappings[i].mapping = mapping;
304 mappings[i].flags = buf->flags;
307 job_data->used_mappings = mappings;
308 job_data->num_used_mappings = i;
316 tegra_drm_mapping_put(mappings[i].mapping);
319 job_data->used_mappings = NULL;
327 static int submit_get_syncpt(struct tegra_drm_context *context, struct host1x_job *job,
328 struct xarray *syncpoints, struct drm_tegra_channel_submit *args)
330 struct host1x_syncpt *sp;
332 if (args->syncpt.flags) {
333 SUBMIT_ERR(context, "invalid flag specified for syncpt");
337 /* Syncpt ref will be dropped on job release */
338 sp = xa_load(syncpoints, args->syncpt.id);
340 SUBMIT_ERR(context, "syncpoint specified in syncpt was not allocated");
344 job->syncpt = host1x_syncpt_get(sp);
345 job->syncpt_incrs = args->syncpt.increments;
350 static int submit_job_add_gather(struct host1x_job *job, struct tegra_drm_context *context,
351 struct drm_tegra_submit_cmd_gather_uptr *cmd,
352 struct gather_bo *bo, u32 *offset,
353 struct tegra_drm_submit_data *job_data,
358 if (cmd->reserved[0] || cmd->reserved[1] || cmd->reserved[2]) {
359 SUBMIT_ERR(context, "non-zero reserved field in GATHER_UPTR command");
363 /* Check for maximum gather size */
364 if (cmd->words > 16383) {
365 SUBMIT_ERR(context, "too many words in GATHER_UPTR command");
369 if (check_add_overflow(*offset, cmd->words, &next_offset)) {
370 SUBMIT_ERR(context, "too many total words in job");
374 if (next_offset > bo->gather_data_words) {
375 SUBMIT_ERR(context, "GATHER_UPTR command overflows gather data");
379 if (tegra_drm_fw_validate(context->client, bo->gather_data, *offset,
380 cmd->words, job_data, class)) {
381 SUBMIT_ERR(context, "job was rejected by firewall");
385 host1x_job_add_gather(job, &bo->base, cmd->words, *offset * 4);
387 *offset = next_offset;
392 static struct host1x_job *
393 submit_create_job(struct tegra_drm_context *context, struct gather_bo *bo,
394 struct drm_tegra_channel_submit *args, struct tegra_drm_submit_data *job_data,
395 struct xarray *syncpoints)
397 struct drm_tegra_submit_cmd *cmds;
398 u32 i, gather_offset = 0, class;
399 struct host1x_job *job;
402 /* Set initial class for firewall. */
403 class = context->client->base.class;
405 cmds = alloc_copy_user_array(u64_to_user_ptr(args->cmds_ptr), args->num_cmds,
408 SUBMIT_ERR(context, "failed to copy cmds array from userspace");
409 return ERR_CAST(cmds);
412 job = host1x_job_alloc(context->channel, args->num_cmds, 0, true);
414 SUBMIT_ERR(context, "failed to allocate memory for job");
415 job = ERR_PTR(-ENOMEM);
419 err = submit_get_syncpt(context, job, syncpoints, args);
423 job->client = &context->client->base;
424 job->class = context->client->base.class;
425 job->serialize = true;
427 for (i = 0; i < args->num_cmds; i++) {
428 struct drm_tegra_submit_cmd *cmd = &cmds[i];
431 SUBMIT_ERR(context, "unknown flags given for cmd");
436 if (cmd->type == DRM_TEGRA_SUBMIT_CMD_GATHER_UPTR) {
437 err = submit_job_add_gather(job, context, &cmd->gather_uptr, bo,
438 &gather_offset, job_data, &class);
441 } else if (cmd->type == DRM_TEGRA_SUBMIT_CMD_WAIT_SYNCPT) {
442 if (cmd->wait_syncpt.reserved[0] || cmd->wait_syncpt.reserved[1]) {
443 SUBMIT_ERR(context, "non-zero reserved value");
448 host1x_job_add_wait(job, cmd->wait_syncpt.id, cmd->wait_syncpt.value,
450 } else if (cmd->type == DRM_TEGRA_SUBMIT_CMD_WAIT_SYNCPT_RELATIVE) {
451 if (cmd->wait_syncpt.reserved[0] || cmd->wait_syncpt.reserved[1]) {
452 SUBMIT_ERR(context, "non-zero reserved value");
457 if (cmd->wait_syncpt.id != args->syncpt.id) {
458 SUBMIT_ERR(context, "syncpoint ID in CMD_WAIT_SYNCPT_RELATIVE is not used by the job");
463 host1x_job_add_wait(job, cmd->wait_syncpt.id, cmd->wait_syncpt.value,
466 SUBMIT_ERR(context, "unknown cmd type");
472 if (gather_offset == 0) {
473 SUBMIT_ERR(context, "job must have at least one gather");
490 static void release_job(struct host1x_job *job)
492 struct tegra_drm_client *client = container_of(job->client, struct tegra_drm_client, base);
493 struct tegra_drm_submit_data *job_data = job->user_data;
496 if (job->memory_context)
497 host1x_memory_context_put(job->memory_context);
499 for (i = 0; i < job_data->num_used_mappings; i++)
500 tegra_drm_mapping_put(job_data->used_mappings[i].mapping);
502 kfree(job_data->used_mappings);
505 pm_runtime_mark_last_busy(client->base.dev);
506 pm_runtime_put_autosuspend(client->base.dev);
509 int tegra_drm_ioctl_channel_submit(struct drm_device *drm, void *data,
510 struct drm_file *file)
512 struct tegra_drm_file *fpriv = file->driver_priv;
513 struct drm_tegra_channel_submit *args = data;
514 struct tegra_drm_submit_data *job_data;
515 struct drm_syncobj *syncobj = NULL;
516 struct tegra_drm_context *context;
517 struct host1x_job *job;
518 struct gather_bo *bo;
522 mutex_lock(&fpriv->lock);
524 context = xa_load(&fpriv->contexts, args->context);
526 mutex_unlock(&fpriv->lock);
527 pr_err_ratelimited("%s: %s: invalid channel context '%#x'", __func__,
528 current->comm, args->context);
532 if (args->syncobj_in) {
533 struct dma_fence *fence;
535 err = drm_syncobj_find_fence(file, args->syncobj_in, 0, 0, &fence);
537 SUBMIT_ERR(context, "invalid syncobj_in '%#x'", args->syncobj_in);
541 err = dma_fence_wait_timeout(fence, true, msecs_to_jiffies(10000));
542 dma_fence_put(fence);
544 SUBMIT_ERR(context, "wait for syncobj_in timed out");
549 if (args->syncobj_out) {
550 syncobj = drm_syncobj_find(file, args->syncobj_out);
552 SUBMIT_ERR(context, "invalid syncobj_out '%#x'", args->syncobj_out);
558 /* Allocate gather BO and copy gather words in. */
559 err = submit_copy_gather_data(&bo, drm->dev, context, args);
563 job_data = kzalloc(sizeof(*job_data), GFP_KERNEL);
565 SUBMIT_ERR(context, "failed to allocate memory for job data");
570 /* Get data buffer mappings and do relocation patching. */
571 err = submit_process_bufs(context, bo, args, job_data);
575 /* Allocate host1x_job and add gathers and waits to it. */
576 job = submit_create_job(context, bo, args, job_data, &fpriv->syncpoints);
582 /* Map gather data for Host1x. */
583 err = host1x_job_pin(job, context->client->base.dev);
585 SUBMIT_ERR(context, "failed to pin job: %d", err);
589 if (context->client->ops->get_streamid_offset) {
590 err = context->client->ops->get_streamid_offset(
591 context->client, &job->engine_streamid_offset);
593 SUBMIT_ERR(context, "failed to get streamid offset: %d", err);
598 if (context->memory_context && context->client->ops->can_use_memory_ctx) {
601 err = context->client->ops->can_use_memory_ctx(context->client, &supported);
603 SUBMIT_ERR(context, "failed to detect if engine can use memory context: %d", err);
608 job->memory_context = context->memory_context;
609 host1x_memory_context_get(job->memory_context);
611 } else if (context->client->ops->get_streamid_offset) {
613 * Job submission will need to temporarily change stream ID,
614 * so need to tell it what to change it back to.
616 if (!tegra_dev_iommu_get_stream_id(context->client->base.dev,
617 &job->engine_fallback_streamid))
618 job->engine_fallback_streamid = TEGRA_STREAM_ID_BYPASS;
622 err = pm_runtime_resume_and_get(context->client->base.dev);
624 SUBMIT_ERR(context, "could not power up engine: %d", err);
625 goto put_memory_context;
628 job->user_data = job_data;
629 job->release = release_job;
630 job->timeout = 10000;
633 * job_data is now part of job reference counting, so don't release
638 /* Submit job to hardware. */
639 err = host1x_job_submit(job);
641 SUBMIT_ERR(context, "host1x job submission failed: %d", err);
645 /* Return postfences to userspace and add fences to DMA reservations. */
646 args->syncpt.value = job->syncpt_end;
649 struct dma_fence *fence = host1x_fence_create(job->syncpt, job->syncpt_end, true);
651 err = PTR_ERR(fence);
652 SUBMIT_ERR(context, "failed to create postfence: %d", err);
655 drm_syncobj_replace_fence(syncobj, fence);
661 if (job->memory_context)
662 host1x_memory_context_put(job->memory_context);
664 host1x_job_unpin(job);
668 if (job_data && job_data->used_mappings) {
669 for (i = 0; i < job_data->num_used_mappings; i++)
670 tegra_drm_mapping_put(job_data->used_mappings[i].mapping);
672 kfree(job_data->used_mappings);
677 gather_bo_put(&bo->base);
680 drm_syncobj_put(syncobj);
682 mutex_unlock(&fpriv->lock);