2 * Copyright 2008 Jerome Glisse.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
25 * Jerome Glisse <glisse@freedesktop.org>
28 #include "radeon_drm.h"
29 #include "radeon_reg.h"
32 void r100_cs_dump_packet(struct radeon_cs_parser *p,
33 struct radeon_cs_packet *pkt);
35 int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
37 struct drm_device *ddev = p->rdev->ddev;
38 struct radeon_cs_chunk *chunk;
42 if (p->chunk_relocs_idx == -1) {
45 chunk = &p->chunks[p->chunk_relocs_idx];
46 /* FIXME: we assume that each relocs use 4 dwords */
47 p->nrelocs = chunk->length_dw / 4;
48 p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
49 if (p->relocs_ptr == NULL) {
52 p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
53 if (p->relocs == NULL) {
56 for (i = 0; i < p->nrelocs; i++) {
57 struct drm_radeon_cs_reloc *r;
60 r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
61 for (j = 0; j < i; j++) {
62 if (r->handle == p->relocs[j].handle) {
63 p->relocs_ptr[i] = &p->relocs[j];
69 p->relocs[i].gobj = drm_gem_object_lookup(ddev,
72 if (p->relocs[i].gobj == NULL) {
73 DRM_ERROR("gem object lookup failed 0x%x\n",
77 p->relocs_ptr[i] = &p->relocs[i];
78 p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
79 p->relocs[i].lobj.bo = p->relocs[i].robj;
80 p->relocs[i].lobj.wdomain = r->write_domain;
81 p->relocs[i].lobj.rdomain = r->read_domains;
82 p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
83 p->relocs[i].handle = r->handle;
84 p->relocs[i].flags = r->flags;
85 radeon_bo_list_add_object(&p->relocs[i].lobj,
89 p->relocs[i].handle = 0;
91 return radeon_bo_list_validate(&p->validated);
94 static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
96 p->priority = priority;
100 DRM_ERROR("unknown ring id: %d\n", ring);
102 case RADEON_CS_RING_GFX:
103 p->ring = RADEON_RING_TYPE_GFX_INDEX;
105 case RADEON_CS_RING_COMPUTE:
106 if (p->rdev->family >= CHIP_TAHITI) {
108 p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
110 p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
112 p->ring = RADEON_RING_TYPE_GFX_INDEX;
118 static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
122 for (i = 0; i < p->nrelocs; i++) {
123 struct radeon_fence *a, *b;
125 if (!p->relocs[i].robj || !p->relocs[i].robj->tbo.sync_obj)
128 a = p->relocs[i].robj->tbo.sync_obj;
129 b = p->ib.sync_to[a->ring];
130 p->ib.sync_to[a->ring] = radeon_fence_later(a, b);
134 /* XXX: note that this is called from the legacy UMS CS ioctl as well */
135 int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
137 struct drm_radeon_cs *cs = data;
138 uint64_t *chunk_array_ptr;
140 u32 ring = RADEON_CS_RING_GFX;
143 if (!cs->num_chunks) {
147 INIT_LIST_HEAD(&p->validated);
150 p->ib.semaphore = NULL;
151 p->const_ib.sa_bo = NULL;
152 p->const_ib.semaphore = NULL;
153 p->chunk_ib_idx = -1;
154 p->chunk_relocs_idx = -1;
155 p->chunk_flags_idx = -1;
156 p->chunk_const_ib_idx = -1;
157 p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
158 if (p->chunks_array == NULL) {
161 chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
162 if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
163 sizeof(uint64_t)*cs->num_chunks)) {
167 p->nchunks = cs->num_chunks;
168 p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
169 if (p->chunks == NULL) {
172 for (i = 0; i < p->nchunks; i++) {
173 struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
174 struct drm_radeon_cs_chunk user_chunk;
175 uint32_t __user *cdata;
177 chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
178 if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
179 sizeof(struct drm_radeon_cs_chunk))) {
182 p->chunks[i].length_dw = user_chunk.length_dw;
183 p->chunks[i].kdata = NULL;
184 p->chunks[i].chunk_id = user_chunk.chunk_id;
186 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
187 p->chunk_relocs_idx = i;
189 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
191 /* zero length IB isn't useful */
192 if (p->chunks[i].length_dw == 0)
195 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
196 p->chunk_const_ib_idx = i;
197 /* zero length CONST IB isn't useful */
198 if (p->chunks[i].length_dw == 0)
201 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
202 p->chunk_flags_idx = i;
203 /* zero length flags aren't useful */
204 if (p->chunks[i].length_dw == 0)
208 p->chunks[i].length_dw = user_chunk.length_dw;
209 p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
211 cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
212 if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
213 (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
214 size = p->chunks[i].length_dw * sizeof(uint32_t);
215 p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
216 if (p->chunks[i].kdata == NULL) {
219 if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
220 p->chunks[i].user_ptr, size)) {
223 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
224 p->cs_flags = p->chunks[i].kdata[0];
225 if (p->chunks[i].length_dw > 1)
226 ring = p->chunks[i].kdata[1];
227 if (p->chunks[i].length_dw > 2)
228 priority = (s32)p->chunks[i].kdata[2];
233 /* these are KMS only */
235 if ((p->cs_flags & RADEON_CS_USE_VM) &&
236 !p->rdev->vm_manager.enabled) {
237 DRM_ERROR("VM not active on asic!\n");
241 /* we only support VM on SI+ */
242 if ((p->rdev->family >= CHIP_TAHITI) &&
243 ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
244 DRM_ERROR("VM required on SI+!\n");
248 if (radeon_cs_get_ring(p, ring, priority))
252 /* deal with non-vm */
253 if ((p->chunk_ib_idx != -1) &&
254 ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
255 (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
256 if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
257 DRM_ERROR("cs IB too big: %d\n",
258 p->chunks[p->chunk_ib_idx].length_dw);
261 if ((p->rdev->flags & RADEON_IS_AGP)) {
262 p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
263 p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
264 if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
265 p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
266 kfree(p->chunks[i].kpage[0]);
267 kfree(p->chunks[i].kpage[1]);
271 p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
272 p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
273 p->chunks[p->chunk_ib_idx].last_copied_page = -1;
274 p->chunks[p->chunk_ib_idx].last_page_index =
275 ((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
282 * cs_parser_fini() - clean parser states
283 * @parser: parser structure holding parsing context.
284 * @error: error number
286 * If error is set than unvalidate buffer, otherwise just free memory
287 * used by parsing context.
289 static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
294 ttm_eu_fence_buffer_objects(&parser->validated,
297 ttm_eu_backoff_reservation(&parser->validated);
299 if (parser->relocs != NULL) {
300 for (i = 0; i < parser->nrelocs; i++) {
301 if (parser->relocs[i].gobj)
302 drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
305 kfree(parser->track);
306 kfree(parser->relocs);
307 kfree(parser->relocs_ptr);
308 for (i = 0; i < parser->nchunks; i++) {
309 kfree(parser->chunks[i].kdata);
310 if ((parser->rdev->flags & RADEON_IS_AGP)) {
311 kfree(parser->chunks[i].kpage[0]);
312 kfree(parser->chunks[i].kpage[1]);
315 kfree(parser->chunks);
316 kfree(parser->chunks_array);
317 radeon_ib_free(parser->rdev, &parser->ib);
318 radeon_ib_free(parser->rdev, &parser->const_ib);
321 static int radeon_cs_ib_chunk(struct radeon_device *rdev,
322 struct radeon_cs_parser *parser)
324 struct radeon_cs_chunk *ib_chunk;
327 if (parser->chunk_ib_idx == -1)
330 if (parser->cs_flags & RADEON_CS_USE_VM)
333 ib_chunk = &parser->chunks[parser->chunk_ib_idx];
334 /* Copy the packet into the IB, the parser will read from the
335 * input memory (cached) and write to the IB (which can be
338 r = radeon_ib_get(rdev, parser->ring, &parser->ib,
339 ib_chunk->length_dw * 4);
341 DRM_ERROR("Failed to get ib !\n");
344 parser->ib.length_dw = ib_chunk->length_dw;
345 r = radeon_cs_parse(rdev, parser->ring, parser);
346 if (r || parser->parser_error) {
347 DRM_ERROR("Invalid command stream !\n");
350 r = radeon_cs_finish_pages(parser);
352 DRM_ERROR("Invalid command stream !\n");
355 radeon_cs_sync_rings(parser);
356 parser->ib.vm_id = 0;
357 r = radeon_ib_schedule(rdev, &parser->ib, NULL);
359 DRM_ERROR("Failed to schedule IB !\n");
364 static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
365 struct radeon_vm *vm)
367 struct radeon_bo_list *lobj;
368 struct radeon_bo *bo;
371 list_for_each_entry(lobj, &parser->validated, tv.head) {
373 r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
381 static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
382 struct radeon_cs_parser *parser)
384 struct radeon_cs_chunk *ib_chunk;
385 struct radeon_fpriv *fpriv = parser->filp->driver_priv;
386 struct radeon_vm *vm = &fpriv->vm;
389 if (parser->chunk_ib_idx == -1)
392 if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
395 if ((rdev->family >= CHIP_TAHITI) &&
396 (parser->chunk_const_ib_idx != -1)) {
397 ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
398 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
399 DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
402 r = radeon_ib_get(rdev, parser->ring, &parser->const_ib,
403 ib_chunk->length_dw * 4);
405 DRM_ERROR("Failed to get const ib !\n");
408 parser->const_ib.is_const_ib = true;
409 parser->const_ib.length_dw = ib_chunk->length_dw;
410 /* Copy the packet into the IB */
411 if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
412 ib_chunk->length_dw * 4)) {
415 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
421 ib_chunk = &parser->chunks[parser->chunk_ib_idx];
422 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
423 DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
426 r = radeon_ib_get(rdev, parser->ring, &parser->ib,
427 ib_chunk->length_dw * 4);
429 DRM_ERROR("Failed to get ib !\n");
432 parser->ib.length_dw = ib_chunk->length_dw;
433 /* Copy the packet into the IB */
434 if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
435 ib_chunk->length_dw * 4)) {
438 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
443 mutex_lock(&rdev->vm_manager.lock);
444 mutex_lock(&vm->mutex);
445 r = radeon_vm_bind(rdev, vm);
449 r = radeon_bo_vm_update_pte(parser, vm);
453 radeon_cs_sync_rings(parser);
455 parser->ib.vm_id = vm->id;
456 /* ib pool is bind at 0 in virtual address space,
457 * so gpu_addr is the offset inside the pool bo
459 parser->ib.gpu_addr = parser->ib.sa_bo->soffset;
461 if ((rdev->family >= CHIP_TAHITI) &&
462 (parser->chunk_const_ib_idx != -1)) {
463 parser->const_ib.vm_id = vm->id;
464 /* ib pool is bind at 0 in virtual address space,
465 * so gpu_addr is the offset inside the pool bo
467 parser->const_ib.gpu_addr = parser->const_ib.sa_bo->soffset;
468 r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
470 r = radeon_ib_schedule(rdev, &parser->ib, NULL);
476 radeon_fence_unref(&vm->fence);
478 vm->fence = radeon_fence_ref(parser->ib.fence);
480 mutex_unlock(&vm->mutex);
481 mutex_unlock(&rdev->vm_manager.lock);
485 static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
488 r = radeon_gpu_reset(rdev);
495 int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
497 struct radeon_device *rdev = dev->dev_private;
498 struct radeon_cs_parser parser;
501 down_read(&rdev->exclusive_lock);
502 if (!rdev->accel_working) {
503 up_read(&rdev->exclusive_lock);
506 /* initialize parser */
507 memset(&parser, 0, sizeof(struct radeon_cs_parser));
510 parser.dev = rdev->dev;
511 parser.family = rdev->family;
512 r = radeon_cs_parser_init(&parser, data);
514 DRM_ERROR("Failed to initialize parser !\n");
515 radeon_cs_parser_fini(&parser, r);
516 up_read(&rdev->exclusive_lock);
517 r = radeon_cs_handle_lockup(rdev, r);
520 r = radeon_cs_parser_relocs(&parser);
522 if (r != -ERESTARTSYS)
523 DRM_ERROR("Failed to parse relocation %d!\n", r);
524 radeon_cs_parser_fini(&parser, r);
525 up_read(&rdev->exclusive_lock);
526 r = radeon_cs_handle_lockup(rdev, r);
529 r = radeon_cs_ib_chunk(rdev, &parser);
533 r = radeon_cs_ib_vm_chunk(rdev, &parser);
538 radeon_cs_parser_fini(&parser, r);
539 up_read(&rdev->exclusive_lock);
540 r = radeon_cs_handle_lockup(rdev, r);
544 int radeon_cs_finish_pages(struct radeon_cs_parser *p)
546 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
548 int size = PAGE_SIZE;
550 for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
551 if (i == ibc->last_page_index) {
552 size = (ibc->length_dw * 4) % PAGE_SIZE;
557 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
558 ibc->user_ptr + (i * PAGE_SIZE),
565 static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
568 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
570 int size = PAGE_SIZE;
571 bool copy1 = (p->rdev->flags & RADEON_IS_AGP) ? false : true;
573 for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
574 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
575 ibc->user_ptr + (i * PAGE_SIZE),
577 p->parser_error = -EFAULT;
582 if (pg_idx == ibc->last_page_index) {
583 size = (ibc->length_dw * 4) % PAGE_SIZE;
588 new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
590 ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
592 if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
593 ibc->user_ptr + (pg_idx * PAGE_SIZE),
595 p->parser_error = -EFAULT;
599 /* copy to IB for non single case */
601 memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
603 ibc->last_copied_page = pg_idx;
604 ibc->kpage_idx[new_page] = pg_idx;
609 u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
611 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
612 u32 pg_idx, pg_offset;
616 pg_idx = (idx * 4) / PAGE_SIZE;
617 pg_offset = (idx * 4) % PAGE_SIZE;
619 if (ibc->kpage_idx[0] == pg_idx)
620 return ibc->kpage[0][pg_offset/4];
621 if (ibc->kpage_idx[1] == pg_idx)
622 return ibc->kpage[1][pg_offset/4];
624 new_page = radeon_cs_update_pages(p, pg_idx);
626 p->parser_error = new_page;
630 idx_value = ibc->kpage[new_page][pg_offset/4];