git.libre-soc.org Git - mesa.git/blob - src/mesa/drivers/dri/intel/intel

1 /**************************************************************************

2 *

5 *

6 * Permission is hereby granted, free of charge, to any person obtaining a

7 * copy of this software and associated documentation files (the

8 * "Software"), to deal in the Software without restriction, including

9 * without limitation the rights to use, copy, modify, merge, publish,

10 * distribute, sub license, and/or sell copies of the Software, and to

11 * permit persons to whom the Software is furnished to do so, subject to

12 * the following conditions:

13 *

14 * The above copyright notice and this permission notice (including the

15 * next paragraph) shall be included in all copies or substantial portions

16 * of the Software.

17 *

18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS

19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF

20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.

21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR

22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,

23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE

24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

25 *

26 **************************************************************************/

28 /* Provide additional functionality on top of bufmgr buffers:

29 * - 2d semantics and blit operations

30 * - refcounting of buffers for multiple images in a buffer.

31 * - refcounting of buffer mappings.

32 * - some logic for moving the buffers to the best memory pools for

33 * given operations.

34 *

35 * Most of this is to make it easier to implement the fixed-layout

36 * mipmap tree required by intel hardware in the face of GL's

37 * programming interface where each image can be specifed in random

38 * order and it isn't clear what layout the tree should have until the

39 * last moment.

40 */

42 #include <sys/ioctl.h>

43 #include <errno.h>

45 #include "intel_context.h"

46 #include "intel_regions.h"

47 #include "intel_blit.h"

48 #include "intel_buffer_objects.h"

49 #include "intel_bufmgr.h"

50 #include "intel_batchbuffer.h"

51 #include "intel_chipset.h"

53 #define FILE_DEBUG_FLAG DEBUG_REGION

55 /* This should be set to the maximum backtrace size desired.

56 * Set it to 0 to disable backtrace debugging.

57 */

58 #define DEBUG_BACKTRACE_SIZE 0

60 #if DEBUG_BACKTRACE_SIZE == 0

61 /* Use the standard debug output */

62 #define _DBG(...) DBG(__VA_ARGS__)

63 #else

64 /* Use backtracing debug output */

65 #define _DBG(...) {debug_backtrace(); DBG(__VA_ARGS__);}

67 /* Backtracing debug support */

68 #include <execinfo.h>

70 static void

71 debug_backtrace(void)

72 {

73 void *trace[DEBUG_BACKTRACE_SIZE];

74 char **strings = NULL;

75 int traceSize;

76 register int i;

    traceSize = backtrace(trace, DEBUG_BACKTRACE_SIZE);

    strings = backtrace_symbols(trace, traceSize);

80 if (strings == NULL) {

81 DBG("no backtrace:");

82 return;

83 }

85 /* Spit out all the strings with a colon separator. Ignore

86 * the first, since we don't really care about the call

87 * to debug_backtrace() itself. Skip until the final "/" in

88 * the trace to avoid really long lines.

89 */

    for (i = 1; i < traceSize; i++) {

       char *p = strings[i], *slash = strings[i];

92 while (*p) {

          if (*p++ == '/') {

94 slash = p;

95 }

96 }

       DBG("%s:", slash);

99 }

100

101 /* Free up the memory, and we're done */

102 free(strings);

103 }

105 #endif

109 /* XXX: Thread safety?

110 */

111 GLubyte *

 intel_region_map(struct intel_context *intel, struct intel_region *region)

113 {

114 intelFlush(&intel->ctx);

115

    _DBG("%s %p\n", __FUNCTION__, region);

117 if (!region->map_refcount++) {

118 if (region->pbo)

119 intel_region_cow(intel, region);

120

       if (region->tiling != I915_TILING_NONE &&

122 intel->intelScreen->kernel_exec_fencing)

123 drm_intel_gem_bo_map_gtt(region->buffer);

124 else

          dri_bo_map(region->buffer, GL_TRUE);

       region->map = region->buffer->virtual;

127 }

128

129 return region->map;

130 }

131

132 void

 intel_region_unmap(struct intel_context *intel, struct intel_region *region)

134 {

    _DBG("%s %p\n", __FUNCTION__, region);

136 if (!--region->map_refcount) {

       if (region->tiling != I915_TILING_NONE &&

138 intel->intelScreen->kernel_exec_fencing)

139 drm_intel_gem_bo_unmap_gtt(region->buffer);

140 else

141 dri_bo_unmap(region->buffer);

142 region->map = NULL;

143 }

144 }

145

146 static struct intel_region *

 intel_region_alloc_internal(struct intel_context *intel,

148 GLuint cpp,

149 GLuint width, GLuint height, GLuint pitch,

150 dri_bo *buffer)

151 {

152 struct intel_region *region;

153

154 if (buffer == NULL) {

       _DBG("%s <-- NULL\n", __FUNCTION__);

156 return NULL;

157 }

158

    region = calloc(sizeof(*region), 1);

160 region->cpp = cpp;

161 region->width = width;

162 region->height = height;

163 region->pitch = pitch;

164 region->refcount = 1;

165 region->buffer = buffer;

166

167 /* Default to no tiling */

168 region->tiling = I915_TILING_NONE;

169 region->bit_6_swizzle = I915_BIT_6_SWIZZLE_NONE;

170

    _DBG("%s <-- %p\n", __FUNCTION__, region);

172 return region;

173 }

174

175 struct intel_region *

 intel_region_alloc(struct intel_context *intel,

177 uint32_t tiling,

178 GLuint cpp, GLuint width, GLuint height, GLuint pitch,

179 GLboolean expect_accelerated_upload)

180 {

181 dri_bo *buffer;

182 struct intel_region *region;

183

184 if (expect_accelerated_upload) {

       buffer = drm_intel_bo_alloc_for_render(intel->bufmgr, "region",

                                              pitch * cpp * height, 64);

187 } else {

       buffer = drm_intel_bo_alloc(intel->bufmgr, "region",

                                   pitch * cpp * height, 64);

190 }

191

    region = intel_region_alloc_internal(intel, cpp, width, height,

193 pitch, buffer);

194

195 if (tiling != I915_TILING_NONE) {

       assert(((pitch * cpp) & 127) == 0);

       drm_intel_bo_set_tiling(buffer, &tiling, pitch * cpp);

       drm_intel_bo_get_tiling(buffer, &region->tiling, &region->bit_6_swizzle);

199 }

200

201 return region;

202 }

203

204 struct intel_region *

 intel_region_alloc_for_handle(struct intel_context *intel,

206 GLuint cpp,

207 GLuint width, GLuint height, GLuint pitch,

208 GLuint handle, const char *name)

209 {

210 struct intel_region *region;

211 dri_bo *buffer;

212 int ret;

213

    buffer = intel_bo_gem_create_from_name(intel->bufmgr, name, handle);

215

    region = intel_region_alloc_internal(intel, cpp,

217 width, height, pitch, buffer);

218 if (region == NULL)

219 return region;

220

    ret = dri_bo_get_tiling(region->buffer, &region->tiling,

222 &region->bit_6_swizzle);

    if (ret != 0) {

       fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",

               handle, name, strerror(-ret));

226 intel_region_release(&region);

227 return NULL;

228 }

229

230 return region;

231 }

232

233 void

 intel_region_reference(struct intel_region **dst, struct intel_region *src)

235 {

236 if (src)

       _DBG("%s %p %d\n", __FUNCTION__, src, src->refcount);

238

239 assert(*dst == NULL);

240 if (src) {

241 src->refcount++;

242 *dst = src;

243 }

244 }

245

246 void

 intel_region_release(struct intel_region **region_handle)

248 {

249 struct intel_region *region = *region_handle;

250

251 if (region == NULL) {

       _DBG("%s NULL\n", __FUNCTION__);

253 return;

254 }

255

    _DBG("%s %p %d\n", __FUNCTION__, region, region->refcount - 1);

257

    ASSERT(region->refcount > 0);

259 region->refcount--;

260

    if (region->refcount == 0) {

       assert(region->map_refcount == 0);

263

264 if (region->pbo)

265 region->pbo->region = NULL;

266 region->pbo = NULL;

267 dri_bo_unreference(region->buffer);

268

       if (region->classic_map != NULL) {

270 drmUnmap(region->classic_map,

                         region->pitch * region->cpp * region->height);

272 }

273

274 free(region);

275 }

276 *region_handle = NULL;

277 }

278

279 /*

280 * XXX Move this into core Mesa?

281 */

282 void

283 _mesa_copy_rect(GLubyte * dst,

284 GLuint cpp,

285 GLuint dst_pitch,

286 GLuint dst_x,

287 GLuint dst_y,

288 GLuint width,

289 GLuint height,

290 const GLubyte * src,

291 GLuint src_pitch, GLuint src_x, GLuint src_y)

292 {

293 GLuint i;

294

295 dst_pitch *= cpp;

296 src_pitch *= cpp;

297 dst += dst_x * cpp;

298 src += src_x * cpp;

299 dst += dst_y * dst_pitch;

300 src += src_y * dst_pitch;

301 width *= cpp;

302

    if (width == dst_pitch && width == src_pitch)

       memcpy(dst, src, height * width);

305 else {

       for (i = 0; i < height; i++) {

          memcpy(dst, src, width);

308 dst += dst_pitch;

309 src += src_pitch;

310 }

311 }

312 }

313

314

315 /* Upload data to a rectangular sub-region. Lots of choices how to do this:

316 *

317 * - memcpy by span to current destination

318 * - upload data as new buffer and blit

319 *

320 * Currently always memcpy.

321 */

322 void

 intel_region_data(struct intel_context *intel,

324 struct intel_region *dst,

325 GLuint dst_offset,

326 GLuint dstx, GLuint dsty,

327 const void *src, GLuint src_pitch,

328 GLuint srcx, GLuint srcy, GLuint width, GLuint height)

329 {

    _DBG("%s\n", __FUNCTION__);

331

332 if (intel == NULL)

333 return;

334

335 if (dst->pbo) {

       if (dstx == 0 &&

           dsty == 0 && width == dst->pitch && height == dst->height)

338 intel_region_release_pbo(intel, dst);

339 else

340 intel_region_cow(intel, dst);

341 }

342

343 LOCK_HARDWARE(intel);

    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,

345 dst->cpp,

346 dst->pitch,

                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);

348

349 intel_region_unmap(intel, dst);

350 UNLOCK_HARDWARE(intel);

351 }

352

353 /* Copy rectangular sub-regions. Need better logic about when to

354 * push buffers into AGP - will currently do so whenever possible.

355 */

356 GLboolean

 intel_region_copy(struct intel_context *intel,

358 struct intel_region *dst,

359 GLuint dst_offset,

360 GLuint dstx, GLuint dsty,

361 struct intel_region *src,

362 GLuint src_offset,

363 GLuint srcx, GLuint srcy, GLuint width, GLuint height,

364 GLenum logicop)

365 {

    _DBG("%s\n", __FUNCTION__);

367

368 if (intel == NULL)

369 return GL_FALSE;

370

371 if (dst->pbo) {

       if (dstx == 0 &&

           dsty == 0 && width == dst->pitch && height == dst->height)

374 intel_region_release_pbo(intel, dst);

375 else

376 intel_region_cow(intel, dst);

377 }

378

    assert(src->cpp == dst->cpp);

380

381 return intelEmitCopyBlit(intel,

382 dst->cpp,

                             src->pitch, src->buffer, src_offset, src->tiling,

                             dst->pitch, dst->buffer, dst_offset, dst->tiling,

                             srcx, srcy, dstx, dsty, width, height,

386 logicop);

387 }

388

389 /* Attach to a pbo, discarding our data. Effectively zero-copy upload

390 * the pbo's data.

391 */

392 void

 intel_region_attach_pbo(struct intel_context *intel,

394 struct intel_region *region,

395 struct intel_buffer_object *pbo)

396 {

397 dri_bo *buffer;

398

    if (region->pbo == pbo)

400 return;

401

    _DBG("%s %p %p\n", __FUNCTION__, region, pbo);

403

404 /* If there is already a pbo attached, break the cow tie now.

405 * Don't call intel_region_release_pbo() as that would

406 * unnecessarily allocate a new buffer we would have to immediately

407 * discard.

408 */

409 if (region->pbo) {

410 region->pbo->region = NULL;

411 region->pbo = NULL;

412 }

413

414 if (region->buffer) {

415 dri_bo_unreference(region->buffer);

416 region->buffer = NULL;

417 }

418

419 /* make sure pbo has a buffer of its own */

    buffer = intel_bufferobj_buffer(intel, pbo, INTEL_WRITE_FULL);

421

422 region->pbo = pbo;

423 region->pbo->region = region;

424 dri_bo_reference(buffer);

425 region->buffer = buffer;

426 }

427

428

429 /* Break the COW tie to the pbo and allocate a new buffer.

430 * The pbo gets to keep the data.

431 */

432 void

 intel_region_release_pbo(struct intel_context *intel,

434 struct intel_region *region)

435 {

    _DBG("%s %p\n", __FUNCTION__, region);

    assert(region->buffer == region->pbo->buffer);

438 region->pbo->region = NULL;

439 region->pbo = NULL;

440 dri_bo_unreference(region->buffer);

441 region->buffer = NULL;

442

    region->buffer = dri_bo_alloc(intel->bufmgr, "region",

                                  region->pitch * region->cpp * region->height,

445 64);

446 }

447

448 /* Break the COW tie to the pbo. Both the pbo and the region end up

449 * with a copy of the data.

450 */

451 void

 intel_region_cow(struct intel_context *intel, struct intel_region *region)

453 {

    struct intel_buffer_object *pbo = region->pbo;

455

456 intel_region_release_pbo(intel, region);

457

    assert(region->cpp * region->pitch * region->height == pbo->Base.Size);

459

    _DBG("%s %p (%d bytes)\n", __FUNCTION__, region, pbo->Base.Size);

461

462 /* Now blit from the texture buffer to the new buffer:

463 */

464

465 LOCK_HARDWARE(intel);

    assert(intelEmitCopyBlit(intel,

467 region->cpp,

                             region->pitch, pbo->buffer, 0, region->tiling,

                             region->pitch, region->buffer, 0, region->tiling,

                             0, 0, 0, 0,

471 region->pitch, region->height,

472 GL_COPY));

473 UNLOCK_HARDWARE(intel);

474 }

475

476 dri_bo *

 intel_region_buffer(struct intel_context *intel,

478 struct intel_region *region, GLuint flag)

479 {

480 if (region->pbo) {

481 if (flag == INTEL_WRITE_PART)

482 intel_region_cow(intel, region);

483 else if (flag == INTEL_WRITE_FULL)

484 intel_region_release_pbo(intel, region);

485 }

486

487 return region->buffer;

488 }

489

490 static struct intel_region *

 intel_recreate_static(struct intel_context *intel,

492 const char *name,

493 struct intel_region *region,

494 intelRegion *region_desc)

495 {

496 intelScreenPrivate *intelScreen = intel->intelScreen;

497 int ret;

498

499 if (region == NULL) {

       region = calloc(sizeof(*region), 1);

501 region->refcount = 1;

       _DBG("%s creating new region %p\n", __FUNCTION__, region);

503 }

504 else {

       _DBG("%s %p\n", __FUNCTION__, region);

506 }

507

    if (intel->ctx.Visual.rgbBits == 24)

509 region->cpp = 4;

510 else

       region->cpp = intel->ctx.Visual.rgbBits / 8;

512 region->pitch = intelScreen->pitch;

513 region->width = intelScreen->width;

514 region->height = intelScreen->height;

515

    if (region->buffer != NULL) {

517 dri_bo_unreference(region->buffer);

518 region->buffer = NULL;

519 }

520

521 if (intel->ttm) {

       assert(region_desc->bo_handle != -1);

       region->buffer = intel_bo_gem_create_from_name(intel->bufmgr,

524 name,

525 region_desc->bo_handle);

526

       ret = dri_bo_get_tiling(region->buffer, &region->tiling,

528 &region->bit_6_swizzle);

       if (ret != 0) {

          fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",

                  region_desc->bo_handle, name, strerror(-ret));

532 intel_region_release(&region);

533 return NULL;

534 }

535 } else {

       if (region->classic_map != NULL) {

537 drmUnmap(region->classic_map,

                   region->pitch * region->cpp * region->height);

539 region->classic_map = NULL;

540 }

       ret = drmMap(intel->driFd, region_desc->handle,

                    region->pitch * region->cpp * region->height,

543 &region->classic_map);

       if (ret != 0) {

          fprintf(stderr, "Failed to drmMap %s buffer\n", name);

546 free(region);

547 return NULL;

548 }

549

       region->buffer = intel_bo_fake_alloc_static(intel->bufmgr,

551 name,

552 region_desc->offset,

553 region->pitch * region->cpp *

554 region->height,

555 region->classic_map);

556

557 /* The sarea just gives us a boolean for whether it's tiled or not,

558 * instead of which tiling mode it is. Guess.

559 */

560 if (region_desc->tiled) {

          if (IS_965(intel->intelScreen->deviceID) &&

562 region_desc == &intelScreen->depth)

563 region->tiling = I915_TILING_Y;

564 else

565 region->tiling = I915_TILING_X;

566 } else {

567 region->tiling = I915_TILING_NONE;

568 }

569

570 region->bit_6_swizzle = I915_BIT_6_SWIZZLE_NONE;

571 }

572

    assert(region->buffer != NULL);

574

575 return region;

576 }

577

578 /**

579 * Create intel_region structs to describe the static front, back, and depth

580 * buffers created by the xserver.

581 *

582 * Although FBO's mean we now no longer use these as render targets in

583 * all circumstances, they won't go away until the back and depth

584 * buffers become private, and the front buffer will remain even then.

585 *

586 * Note that these don't allocate video memory, just describe

587 * allocations alread made by the X server.

588 */

589 void

 intel_recreate_static_regions(struct intel_context *intel)

591 {

592 intelScreenPrivate *intelScreen = intel->intelScreen;

593

594 intel->front_region =

       intel_recreate_static(intel, "front",

596 intel->front_region,

597 &intelScreen->front);

598

599 intel->back_region =

       intel_recreate_static(intel, "back",

601 intel->back_region,

602 &intelScreen->back);

603

604 /* Still assumes front.cpp == depth.cpp. We can kill this when we move to

605 * private buffers.

606 */

607 intel->depth_region =

       intel_recreate_static(intel, "depth",

609 intel->depth_region,

610 &intelScreen->depth);

611 }