blob: c7407074c09b8cd0acd5e51d71b8aef8a96da730 [file] [log] [blame]
Jerome Glisse771fe6b2009-06-05 14:42:42 +02001/*
2 * Copyright 2008 Jerome Glisse.
3 * All Rights Reserved.
4 *
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:
11 *
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
14 * Software.
15 *
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.
23 *
24 * Authors:
25 * Jerome Glisse <glisse@freedesktop.org>
26 */
David Howells760285e2012-10-02 18:01:07 +010027#include <drm/drmP.h>
28#include <drm/radeon_drm.h>
Jerome Glisse771fe6b2009-06-05 14:42:42 +020029#include "radeon_reg.h"
30#include "radeon.h"
31
Lauri Kasanen1109ca02012-08-31 13:43:50 -040032static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
Jerome Glisse771fe6b2009-06-05 14:42:42 +020033{
34 struct drm_device *ddev = p->rdev->ddev;
35 struct radeon_cs_chunk *chunk;
36 unsigned i, j;
37 bool duplicate;
38
39 if (p->chunk_relocs_idx == -1) {
40 return 0;
41 }
42 chunk = &p->chunks[p->chunk_relocs_idx];
Alex Deuchercf4ccd02011-11-18 10:19:47 -050043 p->dma_reloc_idx = 0;
Jerome Glisse771fe6b2009-06-05 14:42:42 +020044 /* FIXME: we assume that each relocs use 4 dwords */
45 p->nrelocs = chunk->length_dw / 4;
46 p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
47 if (p->relocs_ptr == NULL) {
48 return -ENOMEM;
49 }
50 p->relocs = kcalloc(p->nrelocs, sizeof(struct radeon_cs_reloc), GFP_KERNEL);
51 if (p->relocs == NULL) {
52 return -ENOMEM;
53 }
54 for (i = 0; i < p->nrelocs; i++) {
55 struct drm_radeon_cs_reloc *r;
56
57 duplicate = false;
58 r = (struct drm_radeon_cs_reloc *)&chunk->kdata[i*4];
Christian König16557f12011-10-24 14:59:17 +020059 for (j = 0; j < i; j++) {
Jerome Glisse771fe6b2009-06-05 14:42:42 +020060 if (r->handle == p->relocs[j].handle) {
61 p->relocs_ptr[i] = &p->relocs[j];
62 duplicate = true;
63 break;
64 }
65 }
Christian König4474f3a2013-04-08 12:41:28 +020066 if (duplicate) {
Christian König16557f12011-10-24 14:59:17 +020067 p->relocs[i].handle = 0;
Christian König4474f3a2013-04-08 12:41:28 +020068 continue;
69 }
70
71 p->relocs[i].gobj = drm_gem_object_lookup(ddev, p->filp,
72 r->handle);
73 if (p->relocs[i].gobj == NULL) {
74 DRM_ERROR("gem object lookup failed 0x%x\n",
75 r->handle);
76 return -ENOENT;
77 }
78 p->relocs_ptr[i] = &p->relocs[i];
79 p->relocs[i].robj = gem_to_radeon_bo(p->relocs[i].gobj);
80 p->relocs[i].lobj.bo = p->relocs[i].robj;
81 p->relocs[i].lobj.written = !!r->write_domain;
82
Christian Königf2ba57b2013-04-08 12:41:29 +020083 /* the first reloc of an UVD job is the
84 msg and that must be in VRAM */
85 if (p->ring == R600_RING_TYPE_UVD_INDEX && i == 0) {
86 /* TODO: is this still needed for NI+ ? */
87 p->relocs[i].lobj.domain =
88 RADEON_GEM_DOMAIN_VRAM;
89
90 p->relocs[i].lobj.alt_domain =
91 RADEON_GEM_DOMAIN_VRAM;
92
93 } else {
94 uint32_t domain = r->write_domain ?
95 r->write_domain : r->read_domains;
96
97 p->relocs[i].lobj.domain = domain;
98 if (domain == RADEON_GEM_DOMAIN_VRAM)
99 domain |= RADEON_GEM_DOMAIN_GTT;
100 p->relocs[i].lobj.alt_domain = domain;
101 }
Christian König4474f3a2013-04-08 12:41:28 +0200102
103 p->relocs[i].lobj.tv.bo = &p->relocs[i].robj->tbo;
104 p->relocs[i].handle = r->handle;
105
106 radeon_bo_list_add_object(&p->relocs[i].lobj,
107 &p->validated);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200108 }
Christian Königf2ba57b2013-04-08 12:41:29 +0200109 return radeon_bo_list_validate(&p->validated, p->ring);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200110}
111
Jerome Glisse721604a2012-01-05 22:11:05 -0500112static int radeon_cs_get_ring(struct radeon_cs_parser *p, u32 ring, s32 priority)
113{
114 p->priority = priority;
115
116 switch (ring) {
117 default:
118 DRM_ERROR("unknown ring id: %d\n", ring);
119 return -EINVAL;
120 case RADEON_CS_RING_GFX:
121 p->ring = RADEON_RING_TYPE_GFX_INDEX;
122 break;
123 case RADEON_CS_RING_COMPUTE:
Alex Deucher8d5ef7b2012-03-20 17:18:24 -0400124 if (p->rdev->family >= CHIP_TAHITI) {
125 if (p->priority > 0)
126 p->ring = CAYMAN_RING_TYPE_CP1_INDEX;
127 else
128 p->ring = CAYMAN_RING_TYPE_CP2_INDEX;
129 } else
130 p->ring = RADEON_RING_TYPE_GFX_INDEX;
Jerome Glisse721604a2012-01-05 22:11:05 -0500131 break;
Alex Deucher278a3342012-12-13 12:27:28 -0500132 case RADEON_CS_RING_DMA:
133 if (p->rdev->family >= CHIP_CAYMAN) {
134 if (p->priority > 0)
135 p->ring = R600_RING_TYPE_DMA_INDEX;
136 else
137 p->ring = CAYMAN_RING_TYPE_DMA1_INDEX;
138 } else if (p->rdev->family >= CHIP_R600) {
139 p->ring = R600_RING_TYPE_DMA_INDEX;
140 } else {
141 return -EINVAL;
142 }
143 break;
Christian Königf2ba57b2013-04-08 12:41:29 +0200144 case RADEON_CS_RING_UVD:
145 p->ring = R600_RING_TYPE_UVD_INDEX;
146 break;
Jerome Glisse721604a2012-01-05 22:11:05 -0500147 }
148 return 0;
149}
150
Christian König220907d2012-05-10 16:46:43 +0200151static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
Christian König93504fc2012-01-05 22:11:06 -0500152{
Christian König220907d2012-05-10 16:46:43 +0200153 int i;
Christian König93504fc2012-01-05 22:11:06 -0500154
Christian Königcdac5502012-02-23 15:18:42 +0100155 for (i = 0; i < p->nrelocs; i++) {
Christian Königf82cbdd2012-08-09 16:35:36 +0200156 if (!p->relocs[i].robj)
Christian Königcdac5502012-02-23 15:18:42 +0100157 continue;
158
Alex Deucher43f12142013-02-01 17:32:42 +0100159 radeon_ib_sync_to(&p->ib, p->relocs[i].robj->tbo.sync_obj);
Christian Königcdac5502012-02-23 15:18:42 +0100160 }
Christian König93504fc2012-01-05 22:11:06 -0500161}
162
Alex Deucher9b001472012-05-30 10:09:30 -0400163/* XXX: note that this is called from the legacy UMS CS ioctl as well */
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200164int radeon_cs_parser_init(struct radeon_cs_parser *p, void *data)
165{
166 struct drm_radeon_cs *cs = data;
167 uint64_t *chunk_array_ptr;
Jerome Glisse721604a2012-01-05 22:11:05 -0500168 unsigned size, i;
169 u32 ring = RADEON_CS_RING_GFX;
170 s32 priority = 0;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200171
172 if (!cs->num_chunks) {
173 return 0;
174 }
175 /* get chunks */
176 INIT_LIST_HEAD(&p->validated);
177 p->idx = 0;
Jerome Glissef2e39222012-05-09 15:35:02 +0200178 p->ib.sa_bo = NULL;
179 p->ib.semaphore = NULL;
180 p->const_ib.sa_bo = NULL;
181 p->const_ib.semaphore = NULL;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200182 p->chunk_ib_idx = -1;
183 p->chunk_relocs_idx = -1;
Jerome Glisse721604a2012-01-05 22:11:05 -0500184 p->chunk_flags_idx = -1;
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400185 p->chunk_const_ib_idx = -1;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200186 p->chunks_array = kcalloc(cs->num_chunks, sizeof(uint64_t), GFP_KERNEL);
187 if (p->chunks_array == NULL) {
188 return -ENOMEM;
189 }
190 chunk_array_ptr = (uint64_t *)(unsigned long)(cs->chunks);
191 if (DRM_COPY_FROM_USER(p->chunks_array, chunk_array_ptr,
192 sizeof(uint64_t)*cs->num_chunks)) {
193 return -EFAULT;
194 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500195 p->cs_flags = 0;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200196 p->nchunks = cs->num_chunks;
197 p->chunks = kcalloc(p->nchunks, sizeof(struct radeon_cs_chunk), GFP_KERNEL);
198 if (p->chunks == NULL) {
199 return -ENOMEM;
200 }
201 for (i = 0; i < p->nchunks; i++) {
202 struct drm_radeon_cs_chunk __user **chunk_ptr = NULL;
203 struct drm_radeon_cs_chunk user_chunk;
204 uint32_t __user *cdata;
205
206 chunk_ptr = (void __user*)(unsigned long)p->chunks_array[i];
207 if (DRM_COPY_FROM_USER(&user_chunk, chunk_ptr,
208 sizeof(struct drm_radeon_cs_chunk))) {
209 return -EFAULT;
210 }
Dave Airlie5176fdc2009-06-30 11:47:14 +1000211 p->chunks[i].length_dw = user_chunk.length_dw;
212 p->chunks[i].kdata = NULL;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200213 p->chunks[i].chunk_id = user_chunk.chunk_id;
Ilija Hadzic580f8392013-01-02 18:27:37 -0500214 p->chunks[i].user_ptr = (void __user *)(unsigned long)user_chunk.chunk_data;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200215 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) {
216 p->chunk_relocs_idx = i;
217 }
218 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_IB) {
219 p->chunk_ib_idx = i;
Dave Airlie5176fdc2009-06-30 11:47:14 +1000220 /* zero length IB isn't useful */
221 if (p->chunks[i].length_dw == 0)
222 return -EINVAL;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200223 }
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400224 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_CONST_IB) {
225 p->chunk_const_ib_idx = i;
226 /* zero length CONST IB isn't useful */
227 if (p->chunks[i].length_dw == 0)
228 return -EINVAL;
229 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500230 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
231 p->chunk_flags_idx = i;
232 /* zero length flags aren't useful */
233 if (p->chunks[i].length_dw == 0)
234 return -EINVAL;
Marek Olšáke70f2242011-10-25 01:38:45 +0200235 }
Dave Airlie5176fdc2009-06-30 11:47:14 +1000236
Dave Airlie513bcb42009-09-23 16:56:27 +1000237 cdata = (uint32_t *)(unsigned long)user_chunk.chunk_data;
Jerome Glisse721604a2012-01-05 22:11:05 -0500238 if ((p->chunks[i].chunk_id == RADEON_CHUNK_ID_RELOCS) ||
239 (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS)) {
Dave Airlie513bcb42009-09-23 16:56:27 +1000240 size = p->chunks[i].length_dw * sizeof(uint32_t);
241 p->chunks[i].kdata = kmalloc(size, GFP_KERNEL);
242 if (p->chunks[i].kdata == NULL) {
243 return -ENOMEM;
244 }
245 if (DRM_COPY_FROM_USER(p->chunks[i].kdata,
246 p->chunks[i].user_ptr, size)) {
247 return -EFAULT;
248 }
Marek Olšáke70f2242011-10-25 01:38:45 +0200249 if (p->chunks[i].chunk_id == RADEON_CHUNK_ID_FLAGS) {
Jerome Glisse721604a2012-01-05 22:11:05 -0500250 p->cs_flags = p->chunks[i].kdata[0];
251 if (p->chunks[i].length_dw > 1)
252 ring = p->chunks[i].kdata[1];
253 if (p->chunks[i].length_dw > 2)
254 priority = (s32)p->chunks[i].kdata[2];
Marek Olšáke70f2242011-10-25 01:38:45 +0200255 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200256 }
257 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500258
Alex Deucher9b001472012-05-30 10:09:30 -0400259 /* these are KMS only */
260 if (p->rdev) {
261 if ((p->cs_flags & RADEON_CS_USE_VM) &&
262 !p->rdev->vm_manager.enabled) {
263 DRM_ERROR("VM not active on asic!\n");
264 return -EINVAL;
265 }
266
Alex Deucher9b001472012-05-30 10:09:30 -0400267 if (radeon_cs_get_ring(p, ring, priority))
268 return -EINVAL;
Christian König57449042013-04-08 12:41:27 +0200269
270 /* we only support VM on some SI+ rings */
271 if ((p->rdev->asic->ring[p->ring].cs_parse == NULL) &&
272 ((p->cs_flags & RADEON_CS_USE_VM) == 0)) {
273 DRM_ERROR("Ring %d requires VM!\n", p->ring);
274 return -EINVAL;
275 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200276 }
Marek Olšáke70f2242011-10-25 01:38:45 +0200277
Jerome Glisse721604a2012-01-05 22:11:05 -0500278 /* deal with non-vm */
279 if ((p->chunk_ib_idx != -1) &&
280 ((p->cs_flags & RADEON_CS_USE_VM) == 0) &&
281 (p->chunks[p->chunk_ib_idx].chunk_id == RADEON_CHUNK_ID_IB)) {
282 if (p->chunks[p->chunk_ib_idx].length_dw > (16 * 1024)) {
283 DRM_ERROR("cs IB too big: %d\n",
284 p->chunks[p->chunk_ib_idx].length_dw);
285 return -EINVAL;
286 }
Ilija Hadzicff4bd082013-01-07 18:21:57 -0500287 if (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) {
Dave Airlie6a7068b2012-04-03 16:23:41 +0100288 p->chunks[p->chunk_ib_idx].kpage[0] = kmalloc(PAGE_SIZE, GFP_KERNEL);
289 p->chunks[p->chunk_ib_idx].kpage[1] = kmalloc(PAGE_SIZE, GFP_KERNEL);
290 if (p->chunks[p->chunk_ib_idx].kpage[0] == NULL ||
291 p->chunks[p->chunk_ib_idx].kpage[1] == NULL) {
Ilija Hadzic25d89992013-01-07 18:21:59 -0500292 kfree(p->chunks[p->chunk_ib_idx].kpage[0]);
293 kfree(p->chunks[p->chunk_ib_idx].kpage[1]);
Ilija Hadzic1da80cf2013-01-23 13:59:05 -0500294 p->chunks[p->chunk_ib_idx].kpage[0] = NULL;
295 p->chunks[p->chunk_ib_idx].kpage[1] = NULL;
Dave Airlie6a7068b2012-04-03 16:23:41 +0100296 return -ENOMEM;
297 }
298 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500299 p->chunks[p->chunk_ib_idx].kpage_idx[0] = -1;
300 p->chunks[p->chunk_ib_idx].kpage_idx[1] = -1;
301 p->chunks[p->chunk_ib_idx].last_copied_page = -1;
302 p->chunks[p->chunk_ib_idx].last_page_index =
303 ((p->chunks[p->chunk_ib_idx].length_dw * 4) - 1) / PAGE_SIZE;
304 }
305
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200306 return 0;
307}
308
309/**
310 * cs_parser_fini() - clean parser states
311 * @parser: parser structure holding parsing context.
312 * @error: error number
313 *
314 * If error is set than unvalidate buffer, otherwise just free memory
315 * used by parsing context.
316 **/
317static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error)
318{
319 unsigned i;
320
Jerome Glissee43b5ec2012-08-06 12:32:21 -0400321 if (!error) {
Thomas Hellstrom147666f2010-11-17 12:38:32 +0000322 ttm_eu_fence_buffer_objects(&parser->validated,
Jerome Glissef2e39222012-05-09 15:35:02 +0200323 parser->ib.fence);
Jerome Glissee43b5ec2012-08-06 12:32:21 -0400324 } else {
Thomas Hellstrom147666f2010-11-17 12:38:32 +0000325 ttm_eu_backoff_reservation(&parser->validated);
Jerome Glissee43b5ec2012-08-06 12:32:21 -0400326 }
Thomas Hellstrom147666f2010-11-17 12:38:32 +0000327
Pauli Nieminenfcbc4512010-03-19 07:44:33 +0000328 if (parser->relocs != NULL) {
329 for (i = 0; i < parser->nrelocs; i++) {
330 if (parser->relocs[i].gobj)
331 drm_gem_object_unreference_unlocked(parser->relocs[i].gobj);
332 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200333 }
Michel Dänzer48e113e2009-09-15 17:09:32 +0200334 kfree(parser->track);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200335 kfree(parser->relocs);
336 kfree(parser->relocs_ptr);
337 for (i = 0; i < parser->nchunks; i++) {
338 kfree(parser->chunks[i].kdata);
Dave Airlie6a7068b2012-04-03 16:23:41 +0100339 if ((parser->rdev->flags & RADEON_IS_AGP)) {
340 kfree(parser->chunks[i].kpage[0]);
341 kfree(parser->chunks[i].kpage[1]);
342 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200343 }
344 kfree(parser->chunks);
345 kfree(parser->chunks_array);
346 radeon_ib_free(parser->rdev, &parser->ib);
Jerome Glissef2e39222012-05-09 15:35:02 +0200347 radeon_ib_free(parser->rdev, &parser->const_ib);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200348}
349
Jerome Glisse721604a2012-01-05 22:11:05 -0500350static int radeon_cs_ib_chunk(struct radeon_device *rdev,
351 struct radeon_cs_parser *parser)
352{
353 struct radeon_cs_chunk *ib_chunk;
354 int r;
355
356 if (parser->chunk_ib_idx == -1)
357 return 0;
358
359 if (parser->cs_flags & RADEON_CS_USE_VM)
360 return 0;
361
362 ib_chunk = &parser->chunks[parser->chunk_ib_idx];
363 /* Copy the packet into the IB, the parser will read from the
364 * input memory (cached) and write to the IB (which can be
365 * uncached).
366 */
367 r = radeon_ib_get(rdev, parser->ring, &parser->ib,
Christian König4bf3dd92012-08-06 18:57:44 +0200368 NULL, ib_chunk->length_dw * 4);
Jerome Glisse721604a2012-01-05 22:11:05 -0500369 if (r) {
370 DRM_ERROR("Failed to get ib !\n");
371 return r;
372 }
Jerome Glissef2e39222012-05-09 15:35:02 +0200373 parser->ib.length_dw = ib_chunk->length_dw;
Christian Königeb0c19c2012-02-23 15:18:44 +0100374 r = radeon_cs_parse(rdev, parser->ring, parser);
Jerome Glisse721604a2012-01-05 22:11:05 -0500375 if (r || parser->parser_error) {
376 DRM_ERROR("Invalid command stream !\n");
377 return r;
378 }
379 r = radeon_cs_finish_pages(parser);
380 if (r) {
381 DRM_ERROR("Invalid command stream !\n");
382 return r;
383 }
Christian König220907d2012-05-10 16:46:43 +0200384 radeon_cs_sync_rings(parser);
Christian König4ef72562012-07-13 13:06:00 +0200385 r = radeon_ib_schedule(rdev, &parser->ib, NULL);
Jerome Glisse721604a2012-01-05 22:11:05 -0500386 if (r) {
387 DRM_ERROR("Failed to schedule IB !\n");
388 }
Christian König93bf8882012-07-03 14:05:41 +0200389 return r;
Jerome Glisse721604a2012-01-05 22:11:05 -0500390}
391
392static int radeon_bo_vm_update_pte(struct radeon_cs_parser *parser,
393 struct radeon_vm *vm)
394{
Jerome Glisse3e8970f2012-08-13 12:07:33 -0400395 struct radeon_device *rdev = parser->rdev;
Jerome Glisse721604a2012-01-05 22:11:05 -0500396 struct radeon_bo_list *lobj;
397 struct radeon_bo *bo;
398 int r;
399
Jerome Glisse3e8970f2012-08-13 12:07:33 -0400400 r = radeon_vm_bo_update_pte(rdev, vm, rdev->ring_tmp_bo.bo, &rdev->ring_tmp_bo.bo->tbo.mem);
401 if (r) {
402 return r;
403 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500404 list_for_each_entry(lobj, &parser->validated, tv.head) {
405 bo = lobj->bo;
406 r = radeon_vm_bo_update_pte(parser->rdev, vm, bo, &bo->tbo.mem);
407 if (r) {
408 return r;
409 }
410 }
411 return 0;
412}
413
414static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
415 struct radeon_cs_parser *parser)
416{
417 struct radeon_cs_chunk *ib_chunk;
418 struct radeon_fpriv *fpriv = parser->filp->driver_priv;
419 struct radeon_vm *vm = &fpriv->vm;
420 int r;
421
422 if (parser->chunk_ib_idx == -1)
423 return 0;
Jerome Glisse721604a2012-01-05 22:11:05 -0500424 if ((parser->cs_flags & RADEON_CS_USE_VM) == 0)
425 return 0;
426
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400427 if ((rdev->family >= CHIP_TAHITI) &&
428 (parser->chunk_const_ib_idx != -1)) {
429 ib_chunk = &parser->chunks[parser->chunk_const_ib_idx];
430 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
431 DRM_ERROR("cs IB CONST too big: %d\n", ib_chunk->length_dw);
432 return -EINVAL;
433 }
434 r = radeon_ib_get(rdev, parser->ring, &parser->const_ib,
Christian König4bf3dd92012-08-06 18:57:44 +0200435 vm, ib_chunk->length_dw * 4);
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400436 if (r) {
437 DRM_ERROR("Failed to get const ib !\n");
438 return r;
439 }
Jerome Glissef2e39222012-05-09 15:35:02 +0200440 parser->const_ib.is_const_ib = true;
441 parser->const_ib.length_dw = ib_chunk->length_dw;
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400442 /* Copy the packet into the IB */
Jerome Glissef2e39222012-05-09 15:35:02 +0200443 if (DRM_COPY_FROM_USER(parser->const_ib.ptr, ib_chunk->user_ptr,
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400444 ib_chunk->length_dw * 4)) {
445 return -EFAULT;
446 }
Jerome Glissef2e39222012-05-09 15:35:02 +0200447 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->const_ib);
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400448 if (r) {
449 return r;
450 }
451 }
452
Jerome Glisse721604a2012-01-05 22:11:05 -0500453 ib_chunk = &parser->chunks[parser->chunk_ib_idx];
454 if (ib_chunk->length_dw > RADEON_IB_VM_MAX_SIZE) {
455 DRM_ERROR("cs IB too big: %d\n", ib_chunk->length_dw);
456 return -EINVAL;
457 }
458 r = radeon_ib_get(rdev, parser->ring, &parser->ib,
Christian König4bf3dd92012-08-06 18:57:44 +0200459 vm, ib_chunk->length_dw * 4);
Jerome Glisse721604a2012-01-05 22:11:05 -0500460 if (r) {
461 DRM_ERROR("Failed to get ib !\n");
462 return r;
463 }
Jerome Glissef2e39222012-05-09 15:35:02 +0200464 parser->ib.length_dw = ib_chunk->length_dw;
Jerome Glisse721604a2012-01-05 22:11:05 -0500465 /* Copy the packet into the IB */
Jerome Glissef2e39222012-05-09 15:35:02 +0200466 if (DRM_COPY_FROM_USER(parser->ib.ptr, ib_chunk->user_ptr,
Jerome Glisse721604a2012-01-05 22:11:05 -0500467 ib_chunk->length_dw * 4)) {
468 return -EFAULT;
469 }
Jerome Glissef2e39222012-05-09 15:35:02 +0200470 r = radeon_ring_ib_parse(rdev, parser->ring, &parser->ib);
Jerome Glisse721604a2012-01-05 22:11:05 -0500471 if (r) {
472 return r;
473 }
474
Christian König36ff39c2012-05-09 10:07:08 +0200475 mutex_lock(&rdev->vm_manager.lock);
Jerome Glisse721604a2012-01-05 22:11:05 -0500476 mutex_lock(&vm->mutex);
Christian Königddf03f52012-08-09 20:02:28 +0200477 r = radeon_vm_alloc_pt(rdev, vm);
Jerome Glisse721604a2012-01-05 22:11:05 -0500478 if (r) {
479 goto out;
480 }
481 r = radeon_bo_vm_update_pte(parser, vm);
482 if (r) {
483 goto out;
484 }
Christian König220907d2012-05-10 16:46:43 +0200485 radeon_cs_sync_rings(parser);
Alex Deucher43f12142013-02-01 17:32:42 +0100486 radeon_ib_sync_to(&parser->ib, vm->fence);
487 radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
488 rdev, vm, parser->ring));
Christian König4ef72562012-07-13 13:06:00 +0200489
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400490 if ((rdev->family >= CHIP_TAHITI) &&
491 (parser->chunk_const_ib_idx != -1)) {
Christian König4ef72562012-07-13 13:06:00 +0200492 r = radeon_ib_schedule(rdev, &parser->ib, &parser->const_ib);
493 } else {
494 r = radeon_ib_schedule(rdev, &parser->ib, NULL);
Alex Deucherdfcf5f32012-03-20 17:18:14 -0400495 }
496
Jerome Glisse721604a2012-01-05 22:11:05 -0500497 if (!r) {
Christian Königee60e292012-08-09 16:21:08 +0200498 radeon_vm_fence(rdev, vm, parser->ib.fence);
Jerome Glisse721604a2012-01-05 22:11:05 -0500499 }
Christian Königee60e292012-08-09 16:21:08 +0200500
501out:
Christian König13e55c32012-10-09 13:31:19 +0200502 radeon_vm_add_to_lru(rdev, vm);
Christian König36ff39c2012-05-09 10:07:08 +0200503 mutex_unlock(&vm->mutex);
504 mutex_unlock(&rdev->vm_manager.lock);
Jerome Glisse721604a2012-01-05 22:11:05 -0500505 return r;
506}
507
Christian König6c6f4782012-05-02 15:11:19 +0200508static int radeon_cs_handle_lockup(struct radeon_device *rdev, int r)
509{
510 if (r == -EDEADLK) {
511 r = radeon_gpu_reset(rdev);
512 if (!r)
513 r = -EAGAIN;
514 }
515 return r;
516}
517
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200518int radeon_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
519{
520 struct radeon_device *rdev = dev->dev_private;
521 struct radeon_cs_parser parser;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200522 int r;
523
Jerome Glissedee53e72012-07-02 12:45:19 -0400524 down_read(&rdev->exclusive_lock);
Jerome Glisse6b7746e2012-02-20 17:57:20 -0500525 if (!rdev->accel_working) {
Jerome Glissedee53e72012-07-02 12:45:19 -0400526 up_read(&rdev->exclusive_lock);
Jerome Glisse6b7746e2012-02-20 17:57:20 -0500527 return -EBUSY;
528 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200529 /* initialize parser */
530 memset(&parser, 0, sizeof(struct radeon_cs_parser));
531 parser.filp = filp;
532 parser.rdev = rdev;
Jerome Glissec8c15ff2010-01-18 13:01:36 +0100533 parser.dev = rdev->dev;
Dave Airlie428c6e32011-06-08 19:58:29 +1000534 parser.family = rdev->family;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200535 r = radeon_cs_parser_init(&parser, data);
536 if (r) {
537 DRM_ERROR("Failed to initialize parser !\n");
538 radeon_cs_parser_fini(&parser, r);
Jerome Glissedee53e72012-07-02 12:45:19 -0400539 up_read(&rdev->exclusive_lock);
Christian König6c6f4782012-05-02 15:11:19 +0200540 r = radeon_cs_handle_lockup(rdev, r);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200541 return r;
542 }
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200543 r = radeon_cs_parser_relocs(&parser);
544 if (r) {
Dave Airlie97f23b32010-03-19 10:33:44 +1000545 if (r != -ERESTARTSYS)
546 DRM_ERROR("Failed to parse relocation %d!\n", r);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200547 radeon_cs_parser_fini(&parser, r);
Jerome Glissedee53e72012-07-02 12:45:19 -0400548 up_read(&rdev->exclusive_lock);
Christian König6c6f4782012-05-02 15:11:19 +0200549 r = radeon_cs_handle_lockup(rdev, r);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200550 return r;
551 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500552 r = radeon_cs_ib_chunk(rdev, &parser);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200553 if (r) {
Jerome Glisse721604a2012-01-05 22:11:05 -0500554 goto out;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200555 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500556 r = radeon_cs_ib_vm_chunk(rdev, &parser);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200557 if (r) {
Jerome Glisse721604a2012-01-05 22:11:05 -0500558 goto out;
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200559 }
Jerome Glisse721604a2012-01-05 22:11:05 -0500560out:
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200561 radeon_cs_parser_fini(&parser, r);
Jerome Glissedee53e72012-07-02 12:45:19 -0400562 up_read(&rdev->exclusive_lock);
Christian König6c6f4782012-05-02 15:11:19 +0200563 r = radeon_cs_handle_lockup(rdev, r);
Jerome Glisse771fe6b2009-06-05 14:42:42 +0200564 return r;
565}
Dave Airlie513bcb42009-09-23 16:56:27 +1000566
567int radeon_cs_finish_pages(struct radeon_cs_parser *p)
568{
569 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
570 int i;
571 int size = PAGE_SIZE;
572
573 for (i = ibc->last_copied_page + 1; i <= ibc->last_page_index; i++) {
574 if (i == ibc->last_page_index) {
575 size = (ibc->length_dw * 4) % PAGE_SIZE;
576 if (size == 0)
577 size = PAGE_SIZE;
578 }
579
Jerome Glissef2e39222012-05-09 15:35:02 +0200580 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
Dave Airlie513bcb42009-09-23 16:56:27 +1000581 ibc->user_ptr + (i * PAGE_SIZE),
582 size))
583 return -EFAULT;
584 }
585 return 0;
586}
587
Dave Airliec4c7f312012-05-26 17:34:24 +0100588static int radeon_cs_update_pages(struct radeon_cs_parser *p, int pg_idx)
Dave Airlie513bcb42009-09-23 16:56:27 +1000589{
590 int new_page;
Dave Airlie513bcb42009-09-23 16:56:27 +1000591 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
592 int i;
593 int size = PAGE_SIZE;
Ilija Hadzicff4bd082013-01-07 18:21:57 -0500594 bool copy1 = (p->rdev && (p->rdev->flags & RADEON_IS_AGP)) ?
595 false : true;
Dave Airlie513bcb42009-09-23 16:56:27 +1000596
Dave Airliec5e617e2009-09-26 09:03:39 +1000597 for (i = ibc->last_copied_page + 1; i < pg_idx; i++) {
Jerome Glissef2e39222012-05-09 15:35:02 +0200598 if (DRM_COPY_FROM_USER(p->ib.ptr + (i * (PAGE_SIZE/4)),
Dave Airlie513bcb42009-09-23 16:56:27 +1000599 ibc->user_ptr + (i * PAGE_SIZE),
600 PAGE_SIZE)) {
601 p->parser_error = -EFAULT;
602 return 0;
603 }
604 }
605
Dave Airlie513bcb42009-09-23 16:56:27 +1000606 if (pg_idx == ibc->last_page_index) {
607 size = (ibc->length_dw * 4) % PAGE_SIZE;
Dave Airlie6a7068b2012-04-03 16:23:41 +0100608 if (size == 0)
609 size = PAGE_SIZE;
Dave Airlie513bcb42009-09-23 16:56:27 +1000610 }
611
Dave Airlie6a7068b2012-04-03 16:23:41 +0100612 new_page = ibc->kpage_idx[0] < ibc->kpage_idx[1] ? 0 : 1;
613 if (copy1)
Jerome Glissef2e39222012-05-09 15:35:02 +0200614 ibc->kpage[new_page] = p->ib.ptr + (pg_idx * (PAGE_SIZE / 4));
Dave Airlie6a7068b2012-04-03 16:23:41 +0100615
Dave Airlie513bcb42009-09-23 16:56:27 +1000616 if (DRM_COPY_FROM_USER(ibc->kpage[new_page],
617 ibc->user_ptr + (pg_idx * PAGE_SIZE),
618 size)) {
619 p->parser_error = -EFAULT;
620 return 0;
621 }
622
Dave Airlie6a7068b2012-04-03 16:23:41 +0100623 /* copy to IB for non single case */
624 if (!copy1)
Jerome Glissef2e39222012-05-09 15:35:02 +0200625 memcpy((void *)(p->ib.ptr+(pg_idx*(PAGE_SIZE/4))), ibc->kpage[new_page], size);
Dave Airlie513bcb42009-09-23 16:56:27 +1000626
627 ibc->last_copied_page = pg_idx;
628 ibc->kpage_idx[new_page] = pg_idx;
629
630 return new_page;
631}
Dave Airliec4c7f312012-05-26 17:34:24 +0100632
633u32 radeon_get_ib_value(struct radeon_cs_parser *p, int idx)
634{
635 struct radeon_cs_chunk *ibc = &p->chunks[p->chunk_ib_idx];
636 u32 pg_idx, pg_offset;
637 u32 idx_value = 0;
638 int new_page;
639
640 pg_idx = (idx * 4) / PAGE_SIZE;
641 pg_offset = (idx * 4) % PAGE_SIZE;
642
643 if (ibc->kpage_idx[0] == pg_idx)
644 return ibc->kpage[0][pg_offset/4];
645 if (ibc->kpage_idx[1] == pg_idx)
646 return ibc->kpage[1][pg_offset/4];
647
648 new_page = radeon_cs_update_pages(p, pg_idx);
649 if (new_page < 0) {
650 p->parser_error = new_page;
651 return 0;
652 }
653
654 idx_value = ibc->kpage[new_page][pg_offset/4];
655 return idx_value;
656}
Ilija Hadzic4db01312013-01-02 18:27:40 -0500657
658/**
659 * radeon_cs_packet_parse() - parse cp packet and point ib index to next packet
660 * @parser: parser structure holding parsing context.
661 * @pkt: where to store packet information
662 *
663 * Assume that chunk_ib_index is properly set. Will return -EINVAL
664 * if packet is bigger than remaining ib size. or if packets is unknown.
665 **/
666int radeon_cs_packet_parse(struct radeon_cs_parser *p,
667 struct radeon_cs_packet *pkt,
668 unsigned idx)
669{
670 struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
671 struct radeon_device *rdev = p->rdev;
672 uint32_t header;
673
674 if (idx >= ib_chunk->length_dw) {
675 DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
676 idx, ib_chunk->length_dw);
677 return -EINVAL;
678 }
679 header = radeon_get_ib_value(p, idx);
680 pkt->idx = idx;
681 pkt->type = RADEON_CP_PACKET_GET_TYPE(header);
682 pkt->count = RADEON_CP_PACKET_GET_COUNT(header);
683 pkt->one_reg_wr = 0;
684 switch (pkt->type) {
685 case RADEON_PACKET_TYPE0:
686 if (rdev->family < CHIP_R600) {
687 pkt->reg = R100_CP_PACKET0_GET_REG(header);
688 pkt->one_reg_wr =
689 RADEON_CP_PACKET0_GET_ONE_REG_WR(header);
690 } else
691 pkt->reg = R600_CP_PACKET0_GET_REG(header);
692 break;
693 case RADEON_PACKET_TYPE3:
694 pkt->opcode = RADEON_CP_PACKET3_GET_OPCODE(header);
695 break;
696 case RADEON_PACKET_TYPE2:
697 pkt->count = -1;
698 break;
699 default:
700 DRM_ERROR("Unknown packet type %d at %d !\n", pkt->type, idx);
701 return -EINVAL;
702 }
703 if ((pkt->count + 1 + pkt->idx) >= ib_chunk->length_dw) {
704 DRM_ERROR("Packet (%d:%d:%d) end after CS buffer (%d) !\n",
705 pkt->idx, pkt->type, pkt->count, ib_chunk->length_dw);
706 return -EINVAL;
707 }
708 return 0;
709}
Ilija Hadzic9ffb7a62013-01-02 18:27:42 -0500710
711/**
712 * radeon_cs_packet_next_is_pkt3_nop() - test if the next packet is P3 NOP
713 * @p: structure holding the parser context.
714 *
715 * Check if the next packet is NOP relocation packet3.
716 **/
717bool radeon_cs_packet_next_is_pkt3_nop(struct radeon_cs_parser *p)
718{
719 struct radeon_cs_packet p3reloc;
720 int r;
721
722 r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
723 if (r)
724 return false;
725 if (p3reloc.type != RADEON_PACKET_TYPE3)
726 return false;
727 if (p3reloc.opcode != RADEON_PACKET3_NOP)
728 return false;
729 return true;
730}
Ilija Hadzicc3ad63a2013-01-02 18:27:45 -0500731
732/**
733 * radeon_cs_dump_packet() - dump raw packet context
734 * @p: structure holding the parser context.
735 * @pkt: structure holding the packet.
736 *
737 * Used mostly for debugging and error reporting.
738 **/
739void radeon_cs_dump_packet(struct radeon_cs_parser *p,
740 struct radeon_cs_packet *pkt)
741{
742 volatile uint32_t *ib;
743 unsigned i;
744 unsigned idx;
745
746 ib = p->ib.ptr;
747 idx = pkt->idx;
748 for (i = 0; i <= (pkt->count + 1); i++, idx++)
749 DRM_INFO("ib[%d]=0x%08X\n", idx, ib[idx]);
750}
751
Ilija Hadzice9716992013-01-02 18:27:46 -0500752/**
753 * radeon_cs_packet_next_reloc() - parse next (should be reloc) packet
754 * @parser: parser structure holding parsing context.
755 * @data: pointer to relocation data
756 * @offset_start: starting offset
757 * @offset_mask: offset mask (to align start offset on)
758 * @reloc: reloc informations
759 *
760 * Check if next packet is relocation packet3, do bo validation and compute
761 * GPU offset using the provided start.
762 **/
763int radeon_cs_packet_next_reloc(struct radeon_cs_parser *p,
764 struct radeon_cs_reloc **cs_reloc,
765 int nomm)
766{
767 struct radeon_cs_chunk *relocs_chunk;
768 struct radeon_cs_packet p3reloc;
769 unsigned idx;
770 int r;
771
772 if (p->chunk_relocs_idx == -1) {
773 DRM_ERROR("No relocation chunk !\n");
774 return -EINVAL;
775 }
776 *cs_reloc = NULL;
777 relocs_chunk = &p->chunks[p->chunk_relocs_idx];
778 r = radeon_cs_packet_parse(p, &p3reloc, p->idx);
779 if (r)
780 return r;
781 p->idx += p3reloc.count + 2;
782 if (p3reloc.type != RADEON_PACKET_TYPE3 ||
783 p3reloc.opcode != RADEON_PACKET3_NOP) {
784 DRM_ERROR("No packet3 for relocation for packet at %d.\n",
785 p3reloc.idx);
786 radeon_cs_dump_packet(p, &p3reloc);
787 return -EINVAL;
788 }
789 idx = radeon_get_ib_value(p, p3reloc.idx + 1);
790 if (idx >= relocs_chunk->length_dw) {
791 DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
792 idx, relocs_chunk->length_dw);
793 radeon_cs_dump_packet(p, &p3reloc);
794 return -EINVAL;
795 }
796 /* FIXME: we assume reloc size is 4 dwords */
797 if (nomm) {
798 *cs_reloc = p->relocs;
799 (*cs_reloc)->lobj.gpu_offset =
800 (u64)relocs_chunk->kdata[idx + 3] << 32;
801 (*cs_reloc)->lobj.gpu_offset |= relocs_chunk->kdata[idx + 0];
802 } else
803 *cs_reloc = p->relocs_ptr[(idx / 4)];
804 return 0;
805}