1 /* 2 * Copyright (c) 2008 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Eric Anholt <eric@anholt.net> 25 * Keith Packard <keithp@keithp.com> 26 * Mika Kuoppala <mika.kuoppala@intel.com> 27 * 28 */ 29 30 #include <linux/ascii85.h> 31 #include <linux/highmem.h> 32 #include <linux/nmi.h> 33 #include <linux/pagevec.h> 34 #include <linux/scatterlist.h> 35 #include <linux/string_helpers.h> 36 #include <linux/utsname.h> 37 #include <linux/zlib.h> 38 39 #include <drm/drm_cache.h> 40 #include <drm/drm_print.h> 41 42 #include "display/intel_dmc.h" 43 #include "display/intel_overlay.h" 44 45 #include "gem/i915_gem_context.h" 46 #include "gem/i915_gem_lmem.h" 47 #include "gt/intel_engine_regs.h" 48 #include "gt/intel_gt.h" 49 #include "gt/intel_gt_mcr.h" 50 #include "gt/intel_gt_pm.h" 51 #include "gt/intel_gt_regs.h" 52 #include "gt/uc/intel_guc_capture.h" 53 54 #include "i915_driver.h" 55 #include "i915_drv.h" 56 #include "i915_gpu_error.h" 57 #include "i915_memcpy.h" 58 #include "i915_reg.h" 59 #include "i915_scatterlist.h" 60 #include "i915_utils.h" 61 62 #define ALLOW_FAIL (__GFP_KSWAPD_RECLAIM | __GFP_RETRY_MAYFAIL | __GFP_NOWARN) 63 #define ATOMIC_MAYFAIL (GFP_ATOMIC | __GFP_NOWARN) 64 65 static void __sg_set_buf(struct scatterlist *sg, 66 void *addr, unsigned int len, loff_t it) 67 { 68 sg->page_link = (unsigned long)virt_to_page(addr); 69 sg->offset = offset_in_page(addr); 70 sg->length = len; 71 sg->dma_address = it; 72 } 73 74 static bool __i915_error_grow(struct drm_i915_error_state_buf *e, size_t len) 75 { 76 if (!len) 77 return false; 78 79 if (e->bytes + len + 1 <= e->size) 80 return true; 81 82 if (e->bytes) { 83 __sg_set_buf(e->cur++, e->buf, e->bytes, e->iter); 84 e->iter += e->bytes; 85 e->buf = NULL; 86 e->bytes = 0; 87 } 88 89 if (e->cur == e->end) { 90 struct scatterlist *sgl; 91 92 sgl = (typeof(sgl))__get_free_page(ALLOW_FAIL); 93 if (!sgl) { 94 e->err = -ENOMEM; 95 return false; 96 } 97 98 if (e->cur) { 99 e->cur->offset = 0; 100 e->cur->length = 0; 101 e->cur->page_link = 102 (unsigned long)sgl | SG_CHAIN; 103 } else { 104 e->sgl = sgl; 105 } 106 107 e->cur = sgl; 108 e->end = sgl + SG_MAX_SINGLE_ALLOC - 1; 109 } 110 111 e->size = ALIGN(len + 1, SZ_64K); 112 e->buf = kmalloc(e->size, ALLOW_FAIL); 113 if (!e->buf) { 114 e->size = PAGE_ALIGN(len + 1); 115 e->buf = kmalloc(e->size, GFP_KERNEL); 116 } 117 if (!e->buf) { 118 e->err = -ENOMEM; 119 return false; 120 } 121 122 return true; 123 } 124 125 __printf(2, 0) 126 static void i915_error_vprintf(struct drm_i915_error_state_buf *e, 127 const char *fmt, va_list args) 128 { 129 va_list ap; 130 int len; 131 132 if (e->err) 133 return; 134 135 va_copy(ap, args); 136 len = vsnprintf(NULL, 0, fmt, ap); 137 va_end(ap); 138 if (len <= 0) { 139 e->err = len; 140 return; 141 } 142 143 if (!__i915_error_grow(e, len)) 144 return; 145 146 GEM_BUG_ON(e->bytes >= e->size); 147 len = vscnprintf(e->buf + e->bytes, e->size - e->bytes, fmt, args); 148 if (len < 0) { 149 e->err = len; 150 return; 151 } 152 e->bytes += len; 153 } 154 155 static void i915_error_puts(struct drm_i915_error_state_buf *e, const char *str) 156 { 157 unsigned len; 158 159 if (e->err || !str) 160 return; 161 162 len = strlen(str); 163 if (!__i915_error_grow(e, len)) 164 return; 165 166 GEM_BUG_ON(e->bytes + len > e->size); 167 memcpy(e->buf + e->bytes, str, len); 168 e->bytes += len; 169 } 170 171 #define err_printf(e, ...) i915_error_printf(e, __VA_ARGS__) 172 #define err_puts(e, s) i915_error_puts(e, s) 173 174 static void __i915_printfn_error(struct drm_printer *p, struct va_format *vaf) 175 { 176 i915_error_vprintf(p->arg, vaf->fmt, *vaf->va); 177 } 178 179 static inline struct drm_printer 180 i915_error_printer(struct drm_i915_error_state_buf *e) 181 { 182 struct drm_printer p = { 183 .printfn = __i915_printfn_error, 184 .arg = e, 185 }; 186 return p; 187 } 188 189 /* single threaded page allocator with a reserved stash for emergencies */ 190 static void pool_fini(struct pagevec *pv) 191 { 192 pagevec_release(pv); 193 } 194 195 static int pool_refill(struct pagevec *pv, gfp_t gfp) 196 { 197 while (pagevec_space(pv)) { 198 struct page *p; 199 200 p = alloc_page(gfp); 201 if (!p) 202 return -ENOMEM; 203 204 pagevec_add(pv, p); 205 } 206 207 return 0; 208 } 209 210 static int pool_init(struct pagevec *pv, gfp_t gfp) 211 { 212 int err; 213 214 pagevec_init(pv); 215 216 err = pool_refill(pv, gfp); 217 if (err) 218 pool_fini(pv); 219 220 return err; 221 } 222 223 static void *pool_alloc(struct pagevec *pv, gfp_t gfp) 224 { 225 struct page *p; 226 227 p = alloc_page(gfp); 228 if (!p && pagevec_count(pv)) 229 p = pv->pages[--pv->nr]; 230 231 return p ? page_address(p) : NULL; 232 } 233 234 static void pool_free(struct pagevec *pv, void *addr) 235 { 236 struct page *p = virt_to_page(addr); 237 238 if (pagevec_space(pv)) 239 pagevec_add(pv, p); 240 else 241 __free_page(p); 242 } 243 244 #ifdef CONFIG_DRM_I915_COMPRESS_ERROR 245 246 struct i915_vma_compress { 247 struct pagevec pool; 248 struct z_stream_s zstream; 249 void *tmp; 250 }; 251 252 static bool compress_init(struct i915_vma_compress *c) 253 { 254 struct z_stream_s *zstream = &c->zstream; 255 256 if (pool_init(&c->pool, ALLOW_FAIL)) 257 return false; 258 259 zstream->workspace = 260 kmalloc(zlib_deflate_workspacesize(MAX_WBITS, MAX_MEM_LEVEL), 261 ALLOW_FAIL); 262 if (!zstream->workspace) { 263 pool_fini(&c->pool); 264 return false; 265 } 266 267 c->tmp = NULL; 268 if (i915_has_memcpy_from_wc()) 269 c->tmp = pool_alloc(&c->pool, ALLOW_FAIL); 270 271 return true; 272 } 273 274 static bool compress_start(struct i915_vma_compress *c) 275 { 276 struct z_stream_s *zstream = &c->zstream; 277 void *workspace = zstream->workspace; 278 279 memset(zstream, 0, sizeof(*zstream)); 280 zstream->workspace = workspace; 281 282 return zlib_deflateInit(zstream, Z_DEFAULT_COMPRESSION) == Z_OK; 283 } 284 285 static void *compress_next_page(struct i915_vma_compress *c, 286 struct i915_vma_coredump *dst) 287 { 288 void *page_addr; 289 struct page *page; 290 291 page_addr = pool_alloc(&c->pool, ALLOW_FAIL); 292 if (!page_addr) 293 return ERR_PTR(-ENOMEM); 294 295 page = virt_to_page(page_addr); 296 list_add_tail(&page->lru, &dst->page_list); 297 return page_addr; 298 } 299 300 static int compress_page(struct i915_vma_compress *c, 301 void *src, 302 struct i915_vma_coredump *dst, 303 bool wc) 304 { 305 struct z_stream_s *zstream = &c->zstream; 306 307 zstream->next_in = src; 308 if (wc && c->tmp && i915_memcpy_from_wc(c->tmp, src, PAGE_SIZE)) 309 zstream->next_in = c->tmp; 310 zstream->avail_in = PAGE_SIZE; 311 312 do { 313 if (zstream->avail_out == 0) { 314 zstream->next_out = compress_next_page(c, dst); 315 if (IS_ERR(zstream->next_out)) 316 return PTR_ERR(zstream->next_out); 317 318 zstream->avail_out = PAGE_SIZE; 319 } 320 321 if (zlib_deflate(zstream, Z_NO_FLUSH) != Z_OK) 322 return -EIO; 323 324 cond_resched(); 325 } while (zstream->avail_in); 326 327 /* Fallback to uncompressed if we increase size? */ 328 if (0 && zstream->total_out > zstream->total_in) 329 return -E2BIG; 330 331 return 0; 332 } 333 334 static int compress_flush(struct i915_vma_compress *c, 335 struct i915_vma_coredump *dst) 336 { 337 struct z_stream_s *zstream = &c->zstream; 338 339 do { 340 switch (zlib_deflate(zstream, Z_FINISH)) { 341 case Z_OK: /* more space requested */ 342 zstream->next_out = compress_next_page(c, dst); 343 if (IS_ERR(zstream->next_out)) 344 return PTR_ERR(zstream->next_out); 345 346 zstream->avail_out = PAGE_SIZE; 347 break; 348 349 case Z_STREAM_END: 350 goto end; 351 352 default: /* any error */ 353 return -EIO; 354 } 355 } while (1); 356 357 end: 358 memset(zstream->next_out, 0, zstream->avail_out); 359 dst->unused = zstream->avail_out; 360 return 0; 361 } 362 363 static void compress_finish(struct i915_vma_compress *c) 364 { 365 zlib_deflateEnd(&c->zstream); 366 } 367 368 static void compress_fini(struct i915_vma_compress *c) 369 { 370 kfree(c->zstream.workspace); 371 if (c->tmp) 372 pool_free(&c->pool, c->tmp); 373 pool_fini(&c->pool); 374 } 375 376 static void err_compression_marker(struct drm_i915_error_state_buf *m) 377 { 378 err_puts(m, ":"); 379 } 380 381 #else 382 383 struct i915_vma_compress { 384 struct pagevec pool; 385 }; 386 387 static bool compress_init(struct i915_vma_compress *c) 388 { 389 return pool_init(&c->pool, ALLOW_FAIL) == 0; 390 } 391 392 static bool compress_start(struct i915_vma_compress *c) 393 { 394 return true; 395 } 396 397 static int compress_page(struct i915_vma_compress *c, 398 void *src, 399 struct i915_vma_coredump *dst, 400 bool wc) 401 { 402 void *ptr; 403 404 ptr = pool_alloc(&c->pool, ALLOW_FAIL); 405 if (!ptr) 406 return -ENOMEM; 407 408 if (!(wc && i915_memcpy_from_wc(ptr, src, PAGE_SIZE))) 409 memcpy(ptr, src, PAGE_SIZE); 410 list_add_tail(&virt_to_page(ptr)->lru, &dst->page_list); 411 cond_resched(); 412 413 return 0; 414 } 415 416 static int compress_flush(struct i915_vma_compress *c, 417 struct i915_vma_coredump *dst) 418 { 419 return 0; 420 } 421 422 static void compress_finish(struct i915_vma_compress *c) 423 { 424 } 425 426 static void compress_fini(struct i915_vma_compress *c) 427 { 428 pool_fini(&c->pool); 429 } 430 431 static void err_compression_marker(struct drm_i915_error_state_buf *m) 432 { 433 err_puts(m, "~"); 434 } 435 436 #endif 437 438 static void error_print_instdone(struct drm_i915_error_state_buf *m, 439 const struct intel_engine_coredump *ee) 440 { 441 int slice; 442 int subslice; 443 int iter; 444 445 err_printf(m, " INSTDONE: 0x%08x\n", 446 ee->instdone.instdone); 447 448 if (ee->engine->class != RENDER_CLASS || GRAPHICS_VER(m->i915) <= 3) 449 return; 450 451 err_printf(m, " SC_INSTDONE: 0x%08x\n", 452 ee->instdone.slice_common); 453 454 if (GRAPHICS_VER(m->i915) <= 6) 455 return; 456 457 for_each_ss_steering(iter, ee->engine->gt, slice, subslice) 458 err_printf(m, " SAMPLER_INSTDONE[%d][%d]: 0x%08x\n", 459 slice, subslice, 460 ee->instdone.sampler[slice][subslice]); 461 462 for_each_ss_steering(iter, ee->engine->gt, slice, subslice) 463 err_printf(m, " ROW_INSTDONE[%d][%d]: 0x%08x\n", 464 slice, subslice, 465 ee->instdone.row[slice][subslice]); 466 467 if (GRAPHICS_VER(m->i915) < 12) 468 return; 469 470 if (GRAPHICS_VER_FULL(m->i915) >= IP_VER(12, 55)) { 471 for_each_ss_steering(iter, ee->engine->gt, slice, subslice) 472 err_printf(m, " GEOM_SVGUNIT_INSTDONE[%d][%d]: 0x%08x\n", 473 slice, subslice, 474 ee->instdone.geom_svg[slice][subslice]); 475 } 476 477 err_printf(m, " SC_INSTDONE_EXTRA: 0x%08x\n", 478 ee->instdone.slice_common_extra[0]); 479 err_printf(m, " SC_INSTDONE_EXTRA2: 0x%08x\n", 480 ee->instdone.slice_common_extra[1]); 481 } 482 483 static void error_print_request(struct drm_i915_error_state_buf *m, 484 const char *prefix, 485 const struct i915_request_coredump *erq) 486 { 487 if (!erq->seqno) 488 return; 489 490 err_printf(m, "%s pid %d, seqno %8x:%08x%s%s, prio %d, head %08x, tail %08x\n", 491 prefix, erq->pid, erq->context, erq->seqno, 492 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, 493 &erq->flags) ? "!" : "", 494 test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, 495 &erq->flags) ? "+" : "", 496 erq->sched_attr.priority, 497 erq->head, erq->tail); 498 } 499 500 static void error_print_context(struct drm_i915_error_state_buf *m, 501 const char *header, 502 const struct i915_gem_context_coredump *ctx) 503 { 504 err_printf(m, "%s%s[%d] prio %d, guilty %d active %d, runtime total %lluns, avg %lluns\n", 505 header, ctx->comm, ctx->pid, ctx->sched_attr.priority, 506 ctx->guilty, ctx->active, 507 ctx->total_runtime, ctx->avg_runtime); 508 } 509 510 static struct i915_vma_coredump * 511 __find_vma(struct i915_vma_coredump *vma, const char *name) 512 { 513 while (vma) { 514 if (strcmp(vma->name, name) == 0) 515 return vma; 516 vma = vma->next; 517 } 518 519 return NULL; 520 } 521 522 struct i915_vma_coredump * 523 intel_gpu_error_find_batch(const struct intel_engine_coredump *ee) 524 { 525 return __find_vma(ee->vma, "batch"); 526 } 527 528 static void error_print_engine(struct drm_i915_error_state_buf *m, 529 const struct intel_engine_coredump *ee) 530 { 531 struct i915_vma_coredump *batch; 532 int n; 533 534 err_printf(m, "%s command stream:\n", ee->engine->name); 535 err_printf(m, " CCID: 0x%08x\n", ee->ccid); 536 err_printf(m, " START: 0x%08x\n", ee->start); 537 err_printf(m, " HEAD: 0x%08x [0x%08x]\n", ee->head, ee->rq_head); 538 err_printf(m, " TAIL: 0x%08x [0x%08x, 0x%08x]\n", 539 ee->tail, ee->rq_post, ee->rq_tail); 540 err_printf(m, " CTL: 0x%08x\n", ee->ctl); 541 err_printf(m, " MODE: 0x%08x\n", ee->mode); 542 err_printf(m, " HWS: 0x%08x\n", ee->hws); 543 err_printf(m, " ACTHD: 0x%08x %08x\n", 544 (u32)(ee->acthd>>32), (u32)ee->acthd); 545 err_printf(m, " IPEIR: 0x%08x\n", ee->ipeir); 546 err_printf(m, " IPEHR: 0x%08x\n", ee->ipehr); 547 err_printf(m, " ESR: 0x%08x\n", ee->esr); 548 549 error_print_instdone(m, ee); 550 551 batch = intel_gpu_error_find_batch(ee); 552 if (batch) { 553 u64 start = batch->gtt_offset; 554 u64 end = start + batch->gtt_size; 555 556 err_printf(m, " batch: [0x%08x_%08x, 0x%08x_%08x]\n", 557 upper_32_bits(start), lower_32_bits(start), 558 upper_32_bits(end), lower_32_bits(end)); 559 } 560 if (GRAPHICS_VER(m->i915) >= 4) { 561 err_printf(m, " BBADDR: 0x%08x_%08x\n", 562 (u32)(ee->bbaddr>>32), (u32)ee->bbaddr); 563 err_printf(m, " BB_STATE: 0x%08x\n", ee->bbstate); 564 err_printf(m, " INSTPS: 0x%08x\n", ee->instps); 565 } 566 err_printf(m, " INSTPM: 0x%08x\n", ee->instpm); 567 err_printf(m, " FADDR: 0x%08x %08x\n", upper_32_bits(ee->faddr), 568 lower_32_bits(ee->faddr)); 569 if (GRAPHICS_VER(m->i915) >= 6) { 570 err_printf(m, " RC PSMI: 0x%08x\n", ee->rc_psmi); 571 err_printf(m, " FAULT_REG: 0x%08x\n", ee->fault_reg); 572 } 573 if (GRAPHICS_VER(m->i915) >= 11) { 574 err_printf(m, " NOPID: 0x%08x\n", ee->nopid); 575 err_printf(m, " EXCC: 0x%08x\n", ee->excc); 576 err_printf(m, " CMD_CCTL: 0x%08x\n", ee->cmd_cctl); 577 err_printf(m, " CSCMDOP: 0x%08x\n", ee->cscmdop); 578 err_printf(m, " CTX_SR_CTL: 0x%08x\n", ee->ctx_sr_ctl); 579 err_printf(m, " DMA_FADDR_HI: 0x%08x\n", ee->dma_faddr_hi); 580 err_printf(m, " DMA_FADDR_LO: 0x%08x\n", ee->dma_faddr_lo); 581 } 582 if (HAS_PPGTT(m->i915)) { 583 err_printf(m, " GFX_MODE: 0x%08x\n", ee->vm_info.gfx_mode); 584 585 if (GRAPHICS_VER(m->i915) >= 8) { 586 int i; 587 for (i = 0; i < 4; i++) 588 err_printf(m, " PDP%d: 0x%016llx\n", 589 i, ee->vm_info.pdp[i]); 590 } else { 591 err_printf(m, " PP_DIR_BASE: 0x%08x\n", 592 ee->vm_info.pp_dir_base); 593 } 594 } 595 596 for (n = 0; n < ee->num_ports; n++) { 597 err_printf(m, " ELSP[%d]:", n); 598 error_print_request(m, " ", &ee->execlist[n]); 599 } 600 } 601 602 void i915_error_printf(struct drm_i915_error_state_buf *e, const char *f, ...) 603 { 604 va_list args; 605 606 va_start(args, f); 607 i915_error_vprintf(e, f, args); 608 va_end(args); 609 } 610 611 void intel_gpu_error_print_vma(struct drm_i915_error_state_buf *m, 612 const struct intel_engine_cs *engine, 613 const struct i915_vma_coredump *vma) 614 { 615 char out[ASCII85_BUFSZ]; 616 struct page *page; 617 618 if (!vma) 619 return; 620 621 err_printf(m, "%s --- %s = 0x%08x %08x\n", 622 engine ? engine->name : "global", vma->name, 623 upper_32_bits(vma->gtt_offset), 624 lower_32_bits(vma->gtt_offset)); 625 626 if (vma->gtt_page_sizes > I915_GTT_PAGE_SIZE_4K) 627 err_printf(m, "gtt_page_sizes = 0x%08x\n", vma->gtt_page_sizes); 628 629 err_compression_marker(m); 630 list_for_each_entry(page, &vma->page_list, lru) { 631 int i, len; 632 const u32 *addr = page_address(page); 633 634 len = PAGE_SIZE; 635 if (page == list_last_entry(&vma->page_list, typeof(*page), lru)) 636 len -= vma->unused; 637 len = ascii85_encode_len(len); 638 639 for (i = 0; i < len; i++) 640 err_puts(m, ascii85_encode(addr[i], out)); 641 } 642 err_puts(m, "\n"); 643 } 644 645 static void err_print_capabilities(struct drm_i915_error_state_buf *m, 646 struct i915_gpu_coredump *error) 647 { 648 struct drm_printer p = i915_error_printer(m); 649 650 intel_device_info_print(&error->device_info, &error->runtime_info, &p); 651 intel_driver_caps_print(&error->driver_caps, &p); 652 } 653 654 static void err_print_params(struct drm_i915_error_state_buf *m, 655 const struct i915_params *params) 656 { 657 struct drm_printer p = i915_error_printer(m); 658 659 i915_params_dump(params, &p); 660 } 661 662 static void err_print_pciid(struct drm_i915_error_state_buf *m, 663 struct drm_i915_private *i915) 664 { 665 struct pci_dev *pdev = to_pci_dev(i915->drm.dev); 666 667 err_printf(m, "PCI ID: 0x%04x\n", pdev->device); 668 err_printf(m, "PCI Revision: 0x%02x\n", pdev->revision); 669 err_printf(m, "PCI Subsystem: %04x:%04x\n", 670 pdev->subsystem_vendor, 671 pdev->subsystem_device); 672 } 673 674 static void err_print_guc_ctb(struct drm_i915_error_state_buf *m, 675 const char *name, 676 const struct intel_ctb_coredump *ctb) 677 { 678 if (!ctb->size) 679 return; 680 681 err_printf(m, "GuC %s CTB: raw: 0x%08X, 0x%08X/%08X, cached: 0x%08X/%08X, desc = 0x%08X, buf = 0x%08X x 0x%08X\n", 682 name, ctb->raw_status, ctb->raw_head, ctb->raw_tail, 683 ctb->head, ctb->tail, ctb->desc_offset, ctb->cmds_offset, ctb->size); 684 } 685 686 static void err_print_uc(struct drm_i915_error_state_buf *m, 687 const struct intel_uc_coredump *error_uc) 688 { 689 struct drm_printer p = i915_error_printer(m); 690 691 intel_uc_fw_dump(&error_uc->guc_fw, &p); 692 intel_uc_fw_dump(&error_uc->huc_fw, &p); 693 err_printf(m, "GuC timestamp: 0x%08x\n", error_uc->guc.timestamp); 694 intel_gpu_error_print_vma(m, NULL, error_uc->guc.vma_log); 695 err_printf(m, "GuC CTB fence: %d\n", error_uc->guc.last_fence); 696 err_print_guc_ctb(m, "Send", error_uc->guc.ctb + 0); 697 err_print_guc_ctb(m, "Recv", error_uc->guc.ctb + 1); 698 intel_gpu_error_print_vma(m, NULL, error_uc->guc.vma_ctb); 699 } 700 701 static void err_free_sgl(struct scatterlist *sgl) 702 { 703 while (sgl) { 704 struct scatterlist *sg; 705 706 for (sg = sgl; !sg_is_chain(sg); sg++) { 707 kfree(sg_virt(sg)); 708 if (sg_is_last(sg)) 709 break; 710 } 711 712 sg = sg_is_last(sg) ? NULL : sg_chain_ptr(sg); 713 free_page((unsigned long)sgl); 714 sgl = sg; 715 } 716 } 717 718 static void err_print_gt_info(struct drm_i915_error_state_buf *m, 719 struct intel_gt_coredump *gt) 720 { 721 struct drm_printer p = i915_error_printer(m); 722 723 intel_gt_info_print(>->info, &p); 724 intel_sseu_print_topology(gt->_gt->i915, >->info.sseu, &p); 725 } 726 727 static void err_print_gt_display(struct drm_i915_error_state_buf *m, 728 struct intel_gt_coredump *gt) 729 { 730 err_printf(m, "IER: 0x%08x\n", gt->ier); 731 err_printf(m, "DERRMR: 0x%08x\n", gt->derrmr); 732 } 733 734 static void err_print_gt_global_nonguc(struct drm_i915_error_state_buf *m, 735 struct intel_gt_coredump *gt) 736 { 737 int i; 738 739 err_printf(m, "GT awake: %s\n", str_yes_no(gt->awake)); 740 err_printf(m, "CS timestamp frequency: %u Hz, %d ns\n", 741 gt->clock_frequency, gt->clock_period_ns); 742 err_printf(m, "EIR: 0x%08x\n", gt->eir); 743 err_printf(m, "PGTBL_ER: 0x%08x\n", gt->pgtbl_er); 744 745 for (i = 0; i < gt->ngtier; i++) 746 err_printf(m, "GTIER[%d]: 0x%08x\n", i, gt->gtier[i]); 747 } 748 749 static void err_print_gt_global(struct drm_i915_error_state_buf *m, 750 struct intel_gt_coredump *gt) 751 { 752 err_printf(m, "FORCEWAKE: 0x%08x\n", gt->forcewake); 753 754 if (IS_GRAPHICS_VER(m->i915, 6, 11)) { 755 err_printf(m, "ERROR: 0x%08x\n", gt->error); 756 err_printf(m, "DONE_REG: 0x%08x\n", gt->done_reg); 757 } 758 759 if (GRAPHICS_VER(m->i915) >= 8) 760 err_printf(m, "FAULT_TLB_DATA: 0x%08x 0x%08x\n", 761 gt->fault_data1, gt->fault_data0); 762 763 if (GRAPHICS_VER(m->i915) == 7) 764 err_printf(m, "ERR_INT: 0x%08x\n", gt->err_int); 765 766 if (IS_GRAPHICS_VER(m->i915, 8, 11)) 767 err_printf(m, "GTT_CACHE_EN: 0x%08x\n", gt->gtt_cache); 768 769 if (GRAPHICS_VER(m->i915) == 12) 770 err_printf(m, "AUX_ERR_DBG: 0x%08x\n", gt->aux_err); 771 772 if (GRAPHICS_VER(m->i915) >= 12) { 773 int i; 774 775 for (i = 0; i < I915_MAX_SFC; i++) { 776 /* 777 * SFC_DONE resides in the VD forcewake domain, so it 778 * only exists if the corresponding VCS engine is 779 * present. 780 */ 781 if ((gt->_gt->info.sfc_mask & BIT(i)) == 0 || 782 !HAS_ENGINE(gt->_gt, _VCS(i * 2))) 783 continue; 784 785 err_printf(m, " SFC_DONE[%d]: 0x%08x\n", i, 786 gt->sfc_done[i]); 787 } 788 789 err_printf(m, " GAM_DONE: 0x%08x\n", gt->gam_done); 790 } 791 } 792 793 static void err_print_gt_fences(struct drm_i915_error_state_buf *m, 794 struct intel_gt_coredump *gt) 795 { 796 int i; 797 798 for (i = 0; i < gt->nfence; i++) 799 err_printf(m, " fence[%d] = %08llx\n", i, gt->fence[i]); 800 } 801 802 static void err_print_gt_engines(struct drm_i915_error_state_buf *m, 803 struct intel_gt_coredump *gt) 804 { 805 const struct intel_engine_coredump *ee; 806 807 for (ee = gt->engine; ee; ee = ee->next) { 808 const struct i915_vma_coredump *vma; 809 810 if (ee->guc_capture_node) 811 intel_guc_capture_print_engine_node(m, ee); 812 else 813 error_print_engine(m, ee); 814 815 err_printf(m, " hung: %u\n", ee->hung); 816 err_printf(m, " engine reset count: %u\n", ee->reset_count); 817 error_print_context(m, " Active context: ", &ee->context); 818 819 for (vma = ee->vma; vma; vma = vma->next) 820 intel_gpu_error_print_vma(m, ee->engine, vma); 821 } 822 823 } 824 825 static void __err_print_to_sgl(struct drm_i915_error_state_buf *m, 826 struct i915_gpu_coredump *error) 827 { 828 const struct intel_engine_coredump *ee; 829 struct timespec64 ts; 830 831 if (*error->error_msg) 832 err_printf(m, "%s\n", error->error_msg); 833 err_printf(m, "Kernel: %s %s\n", 834 init_utsname()->release, 835 init_utsname()->machine); 836 err_printf(m, "Driver: %s\n", DRIVER_DATE); 837 ts = ktime_to_timespec64(error->time); 838 err_printf(m, "Time: %lld s %ld us\n", 839 (s64)ts.tv_sec, ts.tv_nsec / NSEC_PER_USEC); 840 ts = ktime_to_timespec64(error->boottime); 841 err_printf(m, "Boottime: %lld s %ld us\n", 842 (s64)ts.tv_sec, ts.tv_nsec / NSEC_PER_USEC); 843 ts = ktime_to_timespec64(error->uptime); 844 err_printf(m, "Uptime: %lld s %ld us\n", 845 (s64)ts.tv_sec, ts.tv_nsec / NSEC_PER_USEC); 846 err_printf(m, "Capture: %lu jiffies; %d ms ago\n", 847 error->capture, jiffies_to_msecs(jiffies - error->capture)); 848 849 for (ee = error->gt ? error->gt->engine : NULL; ee; ee = ee->next) 850 err_printf(m, "Active process (on ring %s): %s [%d]\n", 851 ee->engine->name, 852 ee->context.comm, 853 ee->context.pid); 854 855 err_printf(m, "Reset count: %u\n", error->reset_count); 856 err_printf(m, "Suspend count: %u\n", error->suspend_count); 857 err_printf(m, "Platform: %s\n", intel_platform_name(error->device_info.platform)); 858 err_printf(m, "Subplatform: 0x%x\n", 859 intel_subplatform(&error->runtime_info, 860 error->device_info.platform)); 861 err_print_pciid(m, m->i915); 862 863 err_printf(m, "IOMMU enabled?: %d\n", error->iommu); 864 865 intel_dmc_print_error_state(m, m->i915); 866 867 err_printf(m, "RPM wakelock: %s\n", str_yes_no(error->wakelock)); 868 err_printf(m, "PM suspended: %s\n", str_yes_no(error->suspended)); 869 870 if (error->gt) { 871 bool print_guc_capture = false; 872 873 if (error->gt->uc && error->gt->uc->guc.is_guc_capture) 874 print_guc_capture = true; 875 876 err_print_gt_display(m, error->gt); 877 err_print_gt_global_nonguc(m, error->gt); 878 err_print_gt_fences(m, error->gt); 879 880 /* 881 * GuC dumped global, eng-class and eng-instance registers together 882 * as part of engine state dump so we print in err_print_gt_engines 883 */ 884 if (!print_guc_capture) 885 err_print_gt_global(m, error->gt); 886 887 err_print_gt_engines(m, error->gt); 888 889 if (error->gt->uc) 890 err_print_uc(m, error->gt->uc); 891 892 err_print_gt_info(m, error->gt); 893 } 894 895 if (error->overlay) 896 intel_overlay_print_error_state(m, error->overlay); 897 898 err_print_capabilities(m, error); 899 err_print_params(m, &error->params); 900 } 901 902 static int err_print_to_sgl(struct i915_gpu_coredump *error) 903 { 904 struct drm_i915_error_state_buf m; 905 906 if (IS_ERR(error)) 907 return PTR_ERR(error); 908 909 if (READ_ONCE(error->sgl)) 910 return 0; 911 912 memset(&m, 0, sizeof(m)); 913 m.i915 = error->i915; 914 915 __err_print_to_sgl(&m, error); 916 917 if (m.buf) { 918 __sg_set_buf(m.cur++, m.buf, m.bytes, m.iter); 919 m.bytes = 0; 920 m.buf = NULL; 921 } 922 if (m.cur) { 923 GEM_BUG_ON(m.end < m.cur); 924 sg_mark_end(m.cur - 1); 925 } 926 GEM_BUG_ON(m.sgl && !m.cur); 927 928 if (m.err) { 929 err_free_sgl(m.sgl); 930 return m.err; 931 } 932 933 if (cmpxchg(&error->sgl, NULL, m.sgl)) 934 err_free_sgl(m.sgl); 935 936 return 0; 937 } 938 939 ssize_t i915_gpu_coredump_copy_to_buffer(struct i915_gpu_coredump *error, 940 char *buf, loff_t off, size_t rem) 941 { 942 struct scatterlist *sg; 943 size_t count; 944 loff_t pos; 945 int err; 946 947 if (!error || !rem) 948 return 0; 949 950 err = err_print_to_sgl(error); 951 if (err) 952 return err; 953 954 sg = READ_ONCE(error->fit); 955 if (!sg || off < sg->dma_address) 956 sg = error->sgl; 957 if (!sg) 958 return 0; 959 960 pos = sg->dma_address; 961 count = 0; 962 do { 963 size_t len, start; 964 965 if (sg_is_chain(sg)) { 966 sg = sg_chain_ptr(sg); 967 GEM_BUG_ON(sg_is_chain(sg)); 968 } 969 970 len = sg->length; 971 if (pos + len <= off) { 972 pos += len; 973 continue; 974 } 975 976 start = sg->offset; 977 if (pos < off) { 978 GEM_BUG_ON(off - pos > len); 979 len -= off - pos; 980 start += off - pos; 981 pos = off; 982 } 983 984 len = min(len, rem); 985 GEM_BUG_ON(!len || len > sg->length); 986 987 memcpy(buf, page_address(sg_page(sg)) + start, len); 988 989 count += len; 990 pos += len; 991 992 buf += len; 993 rem -= len; 994 if (!rem) { 995 WRITE_ONCE(error->fit, sg); 996 break; 997 } 998 } while (!sg_is_last(sg++)); 999 1000 return count; 1001 } 1002 1003 static void i915_vma_coredump_free(struct i915_vma_coredump *vma) 1004 { 1005 while (vma) { 1006 struct i915_vma_coredump *next = vma->next; 1007 struct page *page, *n; 1008 1009 list_for_each_entry_safe(page, n, &vma->page_list, lru) { 1010 list_del_init(&page->lru); 1011 __free_page(page); 1012 } 1013 1014 kfree(vma); 1015 vma = next; 1016 } 1017 } 1018 1019 static void cleanup_params(struct i915_gpu_coredump *error) 1020 { 1021 i915_params_free(&error->params); 1022 } 1023 1024 static void cleanup_uc(struct intel_uc_coredump *uc) 1025 { 1026 kfree(uc->guc_fw.file_selected.path); 1027 kfree(uc->huc_fw.file_selected.path); 1028 kfree(uc->guc_fw.file_wanted.path); 1029 kfree(uc->huc_fw.file_wanted.path); 1030 i915_vma_coredump_free(uc->guc.vma_log); 1031 i915_vma_coredump_free(uc->guc.vma_ctb); 1032 1033 kfree(uc); 1034 } 1035 1036 static void cleanup_gt(struct intel_gt_coredump *gt) 1037 { 1038 while (gt->engine) { 1039 struct intel_engine_coredump *ee = gt->engine; 1040 1041 gt->engine = ee->next; 1042 1043 i915_vma_coredump_free(ee->vma); 1044 intel_guc_capture_free_node(ee); 1045 kfree(ee); 1046 } 1047 1048 if (gt->uc) 1049 cleanup_uc(gt->uc); 1050 1051 kfree(gt); 1052 } 1053 1054 void __i915_gpu_coredump_free(struct kref *error_ref) 1055 { 1056 struct i915_gpu_coredump *error = 1057 container_of(error_ref, typeof(*error), ref); 1058 1059 while (error->gt) { 1060 struct intel_gt_coredump *gt = error->gt; 1061 1062 error->gt = gt->next; 1063 cleanup_gt(gt); 1064 } 1065 1066 kfree(error->overlay); 1067 1068 cleanup_params(error); 1069 1070 err_free_sgl(error->sgl); 1071 kfree(error); 1072 } 1073 1074 static struct i915_vma_coredump * 1075 i915_vma_coredump_create(const struct intel_gt *gt, 1076 const struct i915_vma_resource *vma_res, 1077 struct i915_vma_compress *compress, 1078 const char *name) 1079 1080 { 1081 struct i915_ggtt *ggtt = gt->ggtt; 1082 const u64 slot = ggtt->error_capture.start; 1083 struct i915_vma_coredump *dst; 1084 struct sgt_iter iter; 1085 int ret; 1086 1087 might_sleep(); 1088 1089 if (!vma_res || !vma_res->bi.pages || !compress) 1090 return NULL; 1091 1092 dst = kmalloc(sizeof(*dst), ALLOW_FAIL); 1093 if (!dst) 1094 return NULL; 1095 1096 if (!compress_start(compress)) { 1097 kfree(dst); 1098 return NULL; 1099 } 1100 1101 INIT_LIST_HEAD(&dst->page_list); 1102 strcpy(dst->name, name); 1103 dst->next = NULL; 1104 1105 dst->gtt_offset = vma_res->start; 1106 dst->gtt_size = vma_res->node_size; 1107 dst->gtt_page_sizes = vma_res->page_sizes_gtt; 1108 dst->unused = 0; 1109 1110 ret = -EINVAL; 1111 if (drm_mm_node_allocated(&ggtt->error_capture)) { 1112 void __iomem *s; 1113 dma_addr_t dma; 1114 1115 for_each_sgt_daddr(dma, iter, vma_res->bi.pages) { 1116 mutex_lock(&ggtt->error_mutex); 1117 if (ggtt->vm.raw_insert_page) 1118 ggtt->vm.raw_insert_page(&ggtt->vm, dma, slot, 1119 I915_CACHE_NONE, 0); 1120 else 1121 ggtt->vm.insert_page(&ggtt->vm, dma, slot, 1122 I915_CACHE_NONE, 0); 1123 mb(); 1124 1125 s = io_mapping_map_wc(&ggtt->iomap, slot, PAGE_SIZE); 1126 ret = compress_page(compress, 1127 (void __force *)s, dst, 1128 true); 1129 io_mapping_unmap(s); 1130 1131 mb(); 1132 ggtt->vm.clear_range(&ggtt->vm, slot, PAGE_SIZE); 1133 mutex_unlock(&ggtt->error_mutex); 1134 if (ret) 1135 break; 1136 } 1137 } else if (vma_res->bi.lmem) { 1138 struct intel_memory_region *mem = vma_res->mr; 1139 dma_addr_t dma; 1140 1141 for_each_sgt_daddr(dma, iter, vma_res->bi.pages) { 1142 dma_addr_t offset = dma - mem->region.start; 1143 void __iomem *s; 1144 1145 if (offset + PAGE_SIZE > mem->io_size) { 1146 ret = -EINVAL; 1147 break; 1148 } 1149 1150 s = io_mapping_map_wc(&mem->iomap, offset, PAGE_SIZE); 1151 ret = compress_page(compress, 1152 (void __force *)s, dst, 1153 true); 1154 io_mapping_unmap(s); 1155 if (ret) 1156 break; 1157 } 1158 } else { 1159 struct page *page; 1160 1161 for_each_sgt_page(page, iter, vma_res->bi.pages) { 1162 void *s; 1163 1164 drm_clflush_pages(&page, 1); 1165 1166 s = kmap(page); 1167 ret = compress_page(compress, s, dst, false); 1168 kunmap(page); 1169 1170 drm_clflush_pages(&page, 1); 1171 1172 if (ret) 1173 break; 1174 } 1175 } 1176 1177 if (ret || compress_flush(compress, dst)) { 1178 struct page *page, *n; 1179 1180 list_for_each_entry_safe_reverse(page, n, &dst->page_list, lru) { 1181 list_del_init(&page->lru); 1182 pool_free(&compress->pool, page_address(page)); 1183 } 1184 1185 kfree(dst); 1186 dst = NULL; 1187 } 1188 compress_finish(compress); 1189 1190 return dst; 1191 } 1192 1193 static void gt_record_fences(struct intel_gt_coredump *gt) 1194 { 1195 struct i915_ggtt *ggtt = gt->_gt->ggtt; 1196 struct intel_uncore *uncore = gt->_gt->uncore; 1197 int i; 1198 1199 if (GRAPHICS_VER(uncore->i915) >= 6) { 1200 for (i = 0; i < ggtt->num_fences; i++) 1201 gt->fence[i] = 1202 intel_uncore_read64(uncore, 1203 FENCE_REG_GEN6_LO(i)); 1204 } else if (GRAPHICS_VER(uncore->i915) >= 4) { 1205 for (i = 0; i < ggtt->num_fences; i++) 1206 gt->fence[i] = 1207 intel_uncore_read64(uncore, 1208 FENCE_REG_965_LO(i)); 1209 } else { 1210 for (i = 0; i < ggtt->num_fences; i++) 1211 gt->fence[i] = 1212 intel_uncore_read(uncore, FENCE_REG(i)); 1213 } 1214 gt->nfence = i; 1215 } 1216 1217 static void engine_record_registers(struct intel_engine_coredump *ee) 1218 { 1219 const struct intel_engine_cs *engine = ee->engine; 1220 struct drm_i915_private *i915 = engine->i915; 1221 1222 if (GRAPHICS_VER(i915) >= 6) { 1223 ee->rc_psmi = ENGINE_READ(engine, RING_PSMI_CTL); 1224 1225 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) 1226 ee->fault_reg = intel_gt_mcr_read_any(engine->gt, 1227 XEHP_RING_FAULT_REG); 1228 else if (GRAPHICS_VER(i915) >= 12) 1229 ee->fault_reg = intel_uncore_read(engine->uncore, 1230 GEN12_RING_FAULT_REG); 1231 else if (GRAPHICS_VER(i915) >= 8) 1232 ee->fault_reg = intel_uncore_read(engine->uncore, 1233 GEN8_RING_FAULT_REG); 1234 else 1235 ee->fault_reg = GEN6_RING_FAULT_REG_READ(engine); 1236 } 1237 1238 if (GRAPHICS_VER(i915) >= 4) { 1239 ee->esr = ENGINE_READ(engine, RING_ESR); 1240 ee->faddr = ENGINE_READ(engine, RING_DMA_FADD); 1241 ee->ipeir = ENGINE_READ(engine, RING_IPEIR); 1242 ee->ipehr = ENGINE_READ(engine, RING_IPEHR); 1243 ee->instps = ENGINE_READ(engine, RING_INSTPS); 1244 ee->bbaddr = ENGINE_READ(engine, RING_BBADDR); 1245 ee->ccid = ENGINE_READ(engine, CCID); 1246 if (GRAPHICS_VER(i915) >= 8) { 1247 ee->faddr |= (u64)ENGINE_READ(engine, RING_DMA_FADD_UDW) << 32; 1248 ee->bbaddr |= (u64)ENGINE_READ(engine, RING_BBADDR_UDW) << 32; 1249 } 1250 ee->bbstate = ENGINE_READ(engine, RING_BBSTATE); 1251 } else { 1252 ee->faddr = ENGINE_READ(engine, DMA_FADD_I8XX); 1253 ee->ipeir = ENGINE_READ(engine, IPEIR); 1254 ee->ipehr = ENGINE_READ(engine, IPEHR); 1255 } 1256 1257 if (GRAPHICS_VER(i915) >= 11) { 1258 ee->cmd_cctl = ENGINE_READ(engine, RING_CMD_CCTL); 1259 ee->cscmdop = ENGINE_READ(engine, RING_CSCMDOP); 1260 ee->ctx_sr_ctl = ENGINE_READ(engine, RING_CTX_SR_CTL); 1261 ee->dma_faddr_hi = ENGINE_READ(engine, RING_DMA_FADD_UDW); 1262 ee->dma_faddr_lo = ENGINE_READ(engine, RING_DMA_FADD); 1263 ee->nopid = ENGINE_READ(engine, RING_NOPID); 1264 ee->excc = ENGINE_READ(engine, RING_EXCC); 1265 } 1266 1267 intel_engine_get_instdone(engine, &ee->instdone); 1268 1269 ee->instpm = ENGINE_READ(engine, RING_INSTPM); 1270 ee->acthd = intel_engine_get_active_head(engine); 1271 ee->start = ENGINE_READ(engine, RING_START); 1272 ee->head = ENGINE_READ(engine, RING_HEAD); 1273 ee->tail = ENGINE_READ(engine, RING_TAIL); 1274 ee->ctl = ENGINE_READ(engine, RING_CTL); 1275 if (GRAPHICS_VER(i915) > 2) 1276 ee->mode = ENGINE_READ(engine, RING_MI_MODE); 1277 1278 if (!HWS_NEEDS_PHYSICAL(i915)) { 1279 i915_reg_t mmio; 1280 1281 if (GRAPHICS_VER(i915) == 7) { 1282 switch (engine->id) { 1283 default: 1284 MISSING_CASE(engine->id); 1285 fallthrough; 1286 case RCS0: 1287 mmio = RENDER_HWS_PGA_GEN7; 1288 break; 1289 case BCS0: 1290 mmio = BLT_HWS_PGA_GEN7; 1291 break; 1292 case VCS0: 1293 mmio = BSD_HWS_PGA_GEN7; 1294 break; 1295 case VECS0: 1296 mmio = VEBOX_HWS_PGA_GEN7; 1297 break; 1298 } 1299 } else if (GRAPHICS_VER(engine->i915) == 6) { 1300 mmio = RING_HWS_PGA_GEN6(engine->mmio_base); 1301 } else { 1302 /* XXX: gen8 returns to sanity */ 1303 mmio = RING_HWS_PGA(engine->mmio_base); 1304 } 1305 1306 ee->hws = intel_uncore_read(engine->uncore, mmio); 1307 } 1308 1309 ee->reset_count = i915_reset_engine_count(&i915->gpu_error, engine); 1310 1311 if (HAS_PPGTT(i915)) { 1312 int i; 1313 1314 ee->vm_info.gfx_mode = ENGINE_READ(engine, RING_MODE_GEN7); 1315 1316 if (GRAPHICS_VER(i915) == 6) { 1317 ee->vm_info.pp_dir_base = 1318 ENGINE_READ(engine, RING_PP_DIR_BASE_READ); 1319 } else if (GRAPHICS_VER(i915) == 7) { 1320 ee->vm_info.pp_dir_base = 1321 ENGINE_READ(engine, RING_PP_DIR_BASE); 1322 } else if (GRAPHICS_VER(i915) >= 8) { 1323 u32 base = engine->mmio_base; 1324 1325 for (i = 0; i < 4; i++) { 1326 ee->vm_info.pdp[i] = 1327 intel_uncore_read(engine->uncore, 1328 GEN8_RING_PDP_UDW(base, i)); 1329 ee->vm_info.pdp[i] <<= 32; 1330 ee->vm_info.pdp[i] |= 1331 intel_uncore_read(engine->uncore, 1332 GEN8_RING_PDP_LDW(base, i)); 1333 } 1334 } 1335 } 1336 } 1337 1338 static void record_request(const struct i915_request *request, 1339 struct i915_request_coredump *erq) 1340 { 1341 erq->flags = request->fence.flags; 1342 erq->context = request->fence.context; 1343 erq->seqno = request->fence.seqno; 1344 erq->sched_attr = request->sched.attr; 1345 erq->head = request->head; 1346 erq->tail = request->tail; 1347 1348 erq->pid = 0; 1349 rcu_read_lock(); 1350 if (!intel_context_is_closed(request->context)) { 1351 const struct i915_gem_context *ctx; 1352 1353 ctx = rcu_dereference(request->context->gem_context); 1354 if (ctx) 1355 erq->pid = pid_nr(ctx->pid); 1356 } 1357 rcu_read_unlock(); 1358 } 1359 1360 static void engine_record_execlists(struct intel_engine_coredump *ee) 1361 { 1362 const struct intel_engine_execlists * const el = &ee->engine->execlists; 1363 struct i915_request * const *port = el->active; 1364 unsigned int n = 0; 1365 1366 while (*port) 1367 record_request(*port++, &ee->execlist[n++]); 1368 1369 ee->num_ports = n; 1370 } 1371 1372 static bool record_context(struct i915_gem_context_coredump *e, 1373 struct intel_context *ce) 1374 { 1375 struct i915_gem_context *ctx; 1376 struct task_struct *task; 1377 bool simulated; 1378 1379 rcu_read_lock(); 1380 ctx = rcu_dereference(ce->gem_context); 1381 if (ctx && !kref_get_unless_zero(&ctx->ref)) 1382 ctx = NULL; 1383 rcu_read_unlock(); 1384 if (!ctx) 1385 return true; 1386 1387 rcu_read_lock(); 1388 task = pid_task(ctx->pid, PIDTYPE_PID); 1389 if (task) { 1390 strcpy(e->comm, task->comm); 1391 e->pid = task->pid; 1392 } 1393 rcu_read_unlock(); 1394 1395 e->sched_attr = ctx->sched; 1396 e->guilty = atomic_read(&ctx->guilty_count); 1397 e->active = atomic_read(&ctx->active_count); 1398 1399 e->total_runtime = intel_context_get_total_runtime_ns(ce); 1400 e->avg_runtime = intel_context_get_avg_runtime_ns(ce); 1401 1402 simulated = i915_gem_context_no_error_capture(ctx); 1403 1404 i915_gem_context_put(ctx); 1405 return simulated; 1406 } 1407 1408 struct intel_engine_capture_vma { 1409 struct intel_engine_capture_vma *next; 1410 struct i915_vma_resource *vma_res; 1411 char name[16]; 1412 bool lockdep_cookie; 1413 }; 1414 1415 static struct intel_engine_capture_vma * 1416 capture_vma_snapshot(struct intel_engine_capture_vma *next, 1417 struct i915_vma_resource *vma_res, 1418 gfp_t gfp, const char *name) 1419 { 1420 struct intel_engine_capture_vma *c; 1421 1422 if (!vma_res) 1423 return next; 1424 1425 c = kmalloc(sizeof(*c), gfp); 1426 if (!c) 1427 return next; 1428 1429 if (!i915_vma_resource_hold(vma_res, &c->lockdep_cookie)) { 1430 kfree(c); 1431 return next; 1432 } 1433 1434 strcpy(c->name, name); 1435 c->vma_res = i915_vma_resource_get(vma_res); 1436 1437 c->next = next; 1438 return c; 1439 } 1440 1441 static struct intel_engine_capture_vma * 1442 capture_vma(struct intel_engine_capture_vma *next, 1443 struct i915_vma *vma, 1444 const char *name, 1445 gfp_t gfp) 1446 { 1447 if (!vma) 1448 return next; 1449 1450 /* 1451 * If the vma isn't pinned, then the vma should be snapshotted 1452 * to a struct i915_vma_snapshot at command submission time. 1453 * Not here. 1454 */ 1455 if (GEM_WARN_ON(!i915_vma_is_pinned(vma))) 1456 return next; 1457 1458 next = capture_vma_snapshot(next, vma->resource, gfp, name); 1459 1460 return next; 1461 } 1462 1463 static struct intel_engine_capture_vma * 1464 capture_user(struct intel_engine_capture_vma *capture, 1465 const struct i915_request *rq, 1466 gfp_t gfp) 1467 { 1468 struct i915_capture_list *c; 1469 1470 for (c = rq->capture_list; c; c = c->next) 1471 capture = capture_vma_snapshot(capture, c->vma_res, gfp, 1472 "user"); 1473 1474 return capture; 1475 } 1476 1477 static void add_vma(struct intel_engine_coredump *ee, 1478 struct i915_vma_coredump *vma) 1479 { 1480 if (vma) { 1481 vma->next = ee->vma; 1482 ee->vma = vma; 1483 } 1484 } 1485 1486 static struct i915_vma_coredump * 1487 create_vma_coredump(const struct intel_gt *gt, struct i915_vma *vma, 1488 const char *name, struct i915_vma_compress *compress) 1489 { 1490 struct i915_vma_coredump *ret = NULL; 1491 struct i915_vma_resource *vma_res; 1492 bool lockdep_cookie; 1493 1494 if (!vma) 1495 return NULL; 1496 1497 vma_res = vma->resource; 1498 1499 if (i915_vma_resource_hold(vma_res, &lockdep_cookie)) { 1500 ret = i915_vma_coredump_create(gt, vma_res, compress, name); 1501 i915_vma_resource_unhold(vma_res, lockdep_cookie); 1502 } 1503 1504 return ret; 1505 } 1506 1507 static void add_vma_coredump(struct intel_engine_coredump *ee, 1508 const struct intel_gt *gt, 1509 struct i915_vma *vma, 1510 const char *name, 1511 struct i915_vma_compress *compress) 1512 { 1513 add_vma(ee, create_vma_coredump(gt, vma, name, compress)); 1514 } 1515 1516 struct intel_engine_coredump * 1517 intel_engine_coredump_alloc(struct intel_engine_cs *engine, gfp_t gfp, u32 dump_flags) 1518 { 1519 struct intel_engine_coredump *ee; 1520 1521 ee = kzalloc(sizeof(*ee), gfp); 1522 if (!ee) 1523 return NULL; 1524 1525 ee->engine = engine; 1526 1527 if (!(dump_flags & CORE_DUMP_FLAG_IS_GUC_CAPTURE)) { 1528 engine_record_registers(ee); 1529 engine_record_execlists(ee); 1530 } 1531 1532 return ee; 1533 } 1534 1535 static struct intel_engine_capture_vma * 1536 engine_coredump_add_context(struct intel_engine_coredump *ee, 1537 struct intel_context *ce, 1538 gfp_t gfp) 1539 { 1540 struct intel_engine_capture_vma *vma = NULL; 1541 1542 ee->simulated |= record_context(&ee->context, ce); 1543 if (ee->simulated) 1544 return NULL; 1545 1546 /* 1547 * We need to copy these to an anonymous buffer 1548 * as the simplest method to avoid being overwritten 1549 * by userspace. 1550 */ 1551 vma = capture_vma(vma, ce->ring->vma, "ring", gfp); 1552 vma = capture_vma(vma, ce->state, "HW context", gfp); 1553 1554 return vma; 1555 } 1556 1557 struct intel_engine_capture_vma * 1558 intel_engine_coredump_add_request(struct intel_engine_coredump *ee, 1559 struct i915_request *rq, 1560 gfp_t gfp) 1561 { 1562 struct intel_engine_capture_vma *vma; 1563 1564 vma = engine_coredump_add_context(ee, rq->context, gfp); 1565 if (!vma) 1566 return NULL; 1567 1568 /* 1569 * We need to copy these to an anonymous buffer 1570 * as the simplest method to avoid being overwritten 1571 * by userspace. 1572 */ 1573 vma = capture_vma_snapshot(vma, rq->batch_res, gfp, "batch"); 1574 vma = capture_user(vma, rq, gfp); 1575 1576 ee->rq_head = rq->head; 1577 ee->rq_post = rq->postfix; 1578 ee->rq_tail = rq->tail; 1579 1580 return vma; 1581 } 1582 1583 void 1584 intel_engine_coredump_add_vma(struct intel_engine_coredump *ee, 1585 struct intel_engine_capture_vma *capture, 1586 struct i915_vma_compress *compress) 1587 { 1588 const struct intel_engine_cs *engine = ee->engine; 1589 1590 while (capture) { 1591 struct intel_engine_capture_vma *this = capture; 1592 struct i915_vma_resource *vma_res = this->vma_res; 1593 1594 add_vma(ee, 1595 i915_vma_coredump_create(engine->gt, vma_res, 1596 compress, this->name)); 1597 1598 i915_vma_resource_unhold(vma_res, this->lockdep_cookie); 1599 i915_vma_resource_put(vma_res); 1600 1601 capture = this->next; 1602 kfree(this); 1603 } 1604 1605 add_vma_coredump(ee, engine->gt, engine->status_page.vma, 1606 "HW Status", compress); 1607 1608 add_vma_coredump(ee, engine->gt, engine->wa_ctx.vma, 1609 "WA context", compress); 1610 } 1611 1612 static struct intel_engine_coredump * 1613 capture_engine(struct intel_engine_cs *engine, 1614 struct i915_vma_compress *compress, 1615 u32 dump_flags) 1616 { 1617 struct intel_engine_capture_vma *capture = NULL; 1618 struct intel_engine_coredump *ee; 1619 struct intel_context *ce = NULL; 1620 struct i915_request *rq = NULL; 1621 1622 ee = intel_engine_coredump_alloc(engine, ALLOW_FAIL, dump_flags); 1623 if (!ee) 1624 return NULL; 1625 1626 intel_engine_get_hung_entity(engine, &ce, &rq); 1627 if (rq && !i915_request_started(rq)) 1628 drm_info(&engine->gt->i915->drm, "Got hung context on %s with active request %lld:%lld [0x%04X] not yet started\n", 1629 engine->name, rq->fence.context, rq->fence.seqno, ce->guc_id.id); 1630 1631 if (rq) { 1632 capture = intel_engine_coredump_add_request(ee, rq, ATOMIC_MAYFAIL); 1633 i915_request_put(rq); 1634 } else if (ce) { 1635 capture = engine_coredump_add_context(ee, ce, ATOMIC_MAYFAIL); 1636 } 1637 1638 if (capture) { 1639 intel_engine_coredump_add_vma(ee, capture, compress); 1640 1641 if (dump_flags & CORE_DUMP_FLAG_IS_GUC_CAPTURE) 1642 intel_guc_capture_get_matching_node(engine->gt, ee, ce); 1643 } else { 1644 kfree(ee); 1645 ee = NULL; 1646 } 1647 1648 return ee; 1649 } 1650 1651 static void 1652 gt_record_engines(struct intel_gt_coredump *gt, 1653 intel_engine_mask_t engine_mask, 1654 struct i915_vma_compress *compress, 1655 u32 dump_flags) 1656 { 1657 struct intel_engine_cs *engine; 1658 enum intel_engine_id id; 1659 1660 for_each_engine(engine, gt->_gt, id) { 1661 struct intel_engine_coredump *ee; 1662 1663 /* Refill our page pool before entering atomic section */ 1664 pool_refill(&compress->pool, ALLOW_FAIL); 1665 1666 ee = capture_engine(engine, compress, dump_flags); 1667 if (!ee) 1668 continue; 1669 1670 ee->hung = engine->mask & engine_mask; 1671 1672 gt->simulated |= ee->simulated; 1673 if (ee->simulated) { 1674 if (dump_flags & CORE_DUMP_FLAG_IS_GUC_CAPTURE) 1675 intel_guc_capture_free_node(ee); 1676 kfree(ee); 1677 continue; 1678 } 1679 1680 ee->next = gt->engine; 1681 gt->engine = ee; 1682 } 1683 } 1684 1685 static void gt_record_guc_ctb(struct intel_ctb_coredump *saved, 1686 const struct intel_guc_ct_buffer *ctb, 1687 const void *blob_ptr, struct intel_guc *guc) 1688 { 1689 if (!ctb || !ctb->desc) 1690 return; 1691 1692 saved->raw_status = ctb->desc->status; 1693 saved->raw_head = ctb->desc->head; 1694 saved->raw_tail = ctb->desc->tail; 1695 saved->head = ctb->head; 1696 saved->tail = ctb->tail; 1697 saved->size = ctb->size; 1698 saved->desc_offset = ((void *)ctb->desc) - blob_ptr; 1699 saved->cmds_offset = ((void *)ctb->cmds) - blob_ptr; 1700 } 1701 1702 static struct intel_uc_coredump * 1703 gt_record_uc(struct intel_gt_coredump *gt, 1704 struct i915_vma_compress *compress) 1705 { 1706 const struct intel_uc *uc = >->_gt->uc; 1707 struct intel_uc_coredump *error_uc; 1708 1709 error_uc = kzalloc(sizeof(*error_uc), ALLOW_FAIL); 1710 if (!error_uc) 1711 return NULL; 1712 1713 memcpy(&error_uc->guc_fw, &uc->guc.fw, sizeof(uc->guc.fw)); 1714 memcpy(&error_uc->huc_fw, &uc->huc.fw, sizeof(uc->huc.fw)); 1715 1716 error_uc->guc_fw.file_selected.path = kstrdup(uc->guc.fw.file_selected.path, ALLOW_FAIL); 1717 error_uc->huc_fw.file_selected.path = kstrdup(uc->huc.fw.file_selected.path, ALLOW_FAIL); 1718 error_uc->guc_fw.file_wanted.path = kstrdup(uc->guc.fw.file_wanted.path, ALLOW_FAIL); 1719 error_uc->huc_fw.file_wanted.path = kstrdup(uc->huc.fw.file_wanted.path, ALLOW_FAIL); 1720 1721 /* 1722 * Save the GuC log and include a timestamp reference for converting the 1723 * log times to system times (in conjunction with the error->boottime and 1724 * gt->clock_frequency fields saved elsewhere). 1725 */ 1726 error_uc->guc.timestamp = intel_uncore_read(gt->_gt->uncore, GUCPMTIMESTAMP); 1727 error_uc->guc.vma_log = create_vma_coredump(gt->_gt, uc->guc.log.vma, 1728 "GuC log buffer", compress); 1729 error_uc->guc.vma_ctb = create_vma_coredump(gt->_gt, uc->guc.ct.vma, 1730 "GuC CT buffer", compress); 1731 error_uc->guc.last_fence = uc->guc.ct.requests.last_fence; 1732 gt_record_guc_ctb(error_uc->guc.ctb + 0, &uc->guc.ct.ctbs.send, 1733 uc->guc.ct.ctbs.send.desc, (struct intel_guc *)&uc->guc); 1734 gt_record_guc_ctb(error_uc->guc.ctb + 1, &uc->guc.ct.ctbs.recv, 1735 uc->guc.ct.ctbs.send.desc, (struct intel_guc *)&uc->guc); 1736 1737 return error_uc; 1738 } 1739 1740 /* Capture display registers. */ 1741 static void gt_record_display_regs(struct intel_gt_coredump *gt) 1742 { 1743 struct intel_uncore *uncore = gt->_gt->uncore; 1744 struct drm_i915_private *i915 = uncore->i915; 1745 1746 if (GRAPHICS_VER(i915) >= 6) 1747 gt->derrmr = intel_uncore_read(uncore, DERRMR); 1748 1749 if (GRAPHICS_VER(i915) >= 8) 1750 gt->ier = intel_uncore_read(uncore, GEN8_DE_MISC_IER); 1751 else if (IS_VALLEYVIEW(i915)) 1752 gt->ier = intel_uncore_read(uncore, VLV_IER); 1753 else if (HAS_PCH_SPLIT(i915)) 1754 gt->ier = intel_uncore_read(uncore, DEIER); 1755 else if (GRAPHICS_VER(i915) == 2) 1756 gt->ier = intel_uncore_read16(uncore, GEN2_IER); 1757 else 1758 gt->ier = intel_uncore_read(uncore, GEN2_IER); 1759 } 1760 1761 /* Capture all other registers that GuC doesn't capture. */ 1762 static void gt_record_global_nonguc_regs(struct intel_gt_coredump *gt) 1763 { 1764 struct intel_uncore *uncore = gt->_gt->uncore; 1765 struct drm_i915_private *i915 = uncore->i915; 1766 int i; 1767 1768 if (IS_VALLEYVIEW(i915)) { 1769 gt->gtier[0] = intel_uncore_read(uncore, GTIER); 1770 gt->ngtier = 1; 1771 } else if (GRAPHICS_VER(i915) >= 11) { 1772 gt->gtier[0] = 1773 intel_uncore_read(uncore, 1774 GEN11_RENDER_COPY_INTR_ENABLE); 1775 gt->gtier[1] = 1776 intel_uncore_read(uncore, GEN11_VCS_VECS_INTR_ENABLE); 1777 gt->gtier[2] = 1778 intel_uncore_read(uncore, GEN11_GUC_SG_INTR_ENABLE); 1779 gt->gtier[3] = 1780 intel_uncore_read(uncore, 1781 GEN11_GPM_WGBOXPERF_INTR_ENABLE); 1782 gt->gtier[4] = 1783 intel_uncore_read(uncore, 1784 GEN11_CRYPTO_RSVD_INTR_ENABLE); 1785 gt->gtier[5] = 1786 intel_uncore_read(uncore, 1787 GEN11_GUNIT_CSME_INTR_ENABLE); 1788 gt->ngtier = 6; 1789 } else if (GRAPHICS_VER(i915) >= 8) { 1790 for (i = 0; i < 4; i++) 1791 gt->gtier[i] = 1792 intel_uncore_read(uncore, GEN8_GT_IER(i)); 1793 gt->ngtier = 4; 1794 } else if (HAS_PCH_SPLIT(i915)) { 1795 gt->gtier[0] = intel_uncore_read(uncore, GTIER); 1796 gt->ngtier = 1; 1797 } 1798 1799 gt->eir = intel_uncore_read(uncore, EIR); 1800 gt->pgtbl_er = intel_uncore_read(uncore, PGTBL_ER); 1801 } 1802 1803 /* 1804 * Capture all registers that relate to workload submission. 1805 * NOTE: In GuC submission, when GuC resets an engine, it can dump these for us 1806 */ 1807 static void gt_record_global_regs(struct intel_gt_coredump *gt) 1808 { 1809 struct intel_uncore *uncore = gt->_gt->uncore; 1810 struct drm_i915_private *i915 = uncore->i915; 1811 int i; 1812 1813 /* 1814 * General organization 1815 * 1. Registers specific to a single generation 1816 * 2. Registers which belong to multiple generations 1817 * 3. Feature specific registers. 1818 * 4. Everything else 1819 * Please try to follow the order. 1820 */ 1821 1822 /* 1: Registers specific to a single generation */ 1823 if (IS_VALLEYVIEW(i915)) 1824 gt->forcewake = intel_uncore_read_fw(uncore, FORCEWAKE_VLV); 1825 1826 if (GRAPHICS_VER(i915) == 7) 1827 gt->err_int = intel_uncore_read(uncore, GEN7_ERR_INT); 1828 1829 if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) { 1830 gt->fault_data0 = intel_gt_mcr_read_any((struct intel_gt *)gt->_gt, 1831 XEHP_FAULT_TLB_DATA0); 1832 gt->fault_data1 = intel_gt_mcr_read_any((struct intel_gt *)gt->_gt, 1833 XEHP_FAULT_TLB_DATA1); 1834 } else if (GRAPHICS_VER(i915) >= 12) { 1835 gt->fault_data0 = intel_uncore_read(uncore, 1836 GEN12_FAULT_TLB_DATA0); 1837 gt->fault_data1 = intel_uncore_read(uncore, 1838 GEN12_FAULT_TLB_DATA1); 1839 } else if (GRAPHICS_VER(i915) >= 8) { 1840 gt->fault_data0 = intel_uncore_read(uncore, 1841 GEN8_FAULT_TLB_DATA0); 1842 gt->fault_data1 = intel_uncore_read(uncore, 1843 GEN8_FAULT_TLB_DATA1); 1844 } 1845 1846 if (GRAPHICS_VER(i915) == 6) { 1847 gt->forcewake = intel_uncore_read_fw(uncore, FORCEWAKE); 1848 gt->gab_ctl = intel_uncore_read(uncore, GAB_CTL); 1849 gt->gfx_mode = intel_uncore_read(uncore, GFX_MODE); 1850 } 1851 1852 /* 2: Registers which belong to multiple generations */ 1853 if (GRAPHICS_VER(i915) >= 7) 1854 gt->forcewake = intel_uncore_read_fw(uncore, FORCEWAKE_MT); 1855 1856 if (GRAPHICS_VER(i915) >= 6) { 1857 if (GRAPHICS_VER(i915) < 12) { 1858 gt->error = intel_uncore_read(uncore, ERROR_GEN6); 1859 gt->done_reg = intel_uncore_read(uncore, DONE_REG); 1860 } 1861 } 1862 1863 /* 3: Feature specific registers */ 1864 if (IS_GRAPHICS_VER(i915, 6, 7)) { 1865 gt->gam_ecochk = intel_uncore_read(uncore, GAM_ECOCHK); 1866 gt->gac_eco = intel_uncore_read(uncore, GAC_ECO_BITS); 1867 } 1868 1869 if (IS_GRAPHICS_VER(i915, 8, 11)) 1870 gt->gtt_cache = intel_uncore_read(uncore, HSW_GTT_CACHE_EN); 1871 1872 if (GRAPHICS_VER(i915) == 12) 1873 gt->aux_err = intel_uncore_read(uncore, GEN12_AUX_ERR_DBG); 1874 1875 if (GRAPHICS_VER(i915) >= 12) { 1876 for (i = 0; i < I915_MAX_SFC; i++) { 1877 /* 1878 * SFC_DONE resides in the VD forcewake domain, so it 1879 * only exists if the corresponding VCS engine is 1880 * present. 1881 */ 1882 if ((gt->_gt->info.sfc_mask & BIT(i)) == 0 || 1883 !HAS_ENGINE(gt->_gt, _VCS(i * 2))) 1884 continue; 1885 1886 gt->sfc_done[i] = 1887 intel_uncore_read(uncore, GEN12_SFC_DONE(i)); 1888 } 1889 1890 gt->gam_done = intel_uncore_read(uncore, GEN12_GAM_DONE); 1891 } 1892 } 1893 1894 static void gt_record_info(struct intel_gt_coredump *gt) 1895 { 1896 memcpy(>->info, >->_gt->info, sizeof(struct intel_gt_info)); 1897 gt->clock_frequency = gt->_gt->clock_frequency; 1898 gt->clock_period_ns = gt->_gt->clock_period_ns; 1899 } 1900 1901 /* 1902 * Generate a semi-unique error code. The code is not meant to have meaning, The 1903 * code's only purpose is to try to prevent false duplicated bug reports by 1904 * grossly estimating a GPU error state. 1905 * 1906 * TODO Ideally, hashing the batchbuffer would be a very nice way to determine 1907 * the hang if we could strip the GTT offset information from it. 1908 * 1909 * It's only a small step better than a random number in its current form. 1910 */ 1911 static u32 generate_ecode(const struct intel_engine_coredump *ee) 1912 { 1913 /* 1914 * IPEHR would be an ideal way to detect errors, as it's the gross 1915 * measure of "the command that hung." However, has some very common 1916 * synchronization commands which almost always appear in the case 1917 * strictly a client bug. Use instdone to differentiate those some. 1918 */ 1919 return ee ? ee->ipehr ^ ee->instdone.instdone : 0; 1920 } 1921 1922 static const char *error_msg(struct i915_gpu_coredump *error) 1923 { 1924 struct intel_engine_coredump *first = NULL; 1925 unsigned int hung_classes = 0; 1926 struct intel_gt_coredump *gt; 1927 int len; 1928 1929 for (gt = error->gt; gt; gt = gt->next) { 1930 struct intel_engine_coredump *cs; 1931 1932 for (cs = gt->engine; cs; cs = cs->next) { 1933 if (cs->hung) { 1934 hung_classes |= BIT(cs->engine->uabi_class); 1935 if (!first) 1936 first = cs; 1937 } 1938 } 1939 } 1940 1941 len = scnprintf(error->error_msg, sizeof(error->error_msg), 1942 "GPU HANG: ecode %d:%x:%08x", 1943 GRAPHICS_VER(error->i915), hung_classes, 1944 generate_ecode(first)); 1945 if (first && first->context.pid) { 1946 /* Just show the first executing process, more is confusing */ 1947 len += scnprintf(error->error_msg + len, 1948 sizeof(error->error_msg) - len, 1949 ", in %s [%d]", 1950 first->context.comm, first->context.pid); 1951 } 1952 1953 return error->error_msg; 1954 } 1955 1956 static void capture_gen(struct i915_gpu_coredump *error) 1957 { 1958 struct drm_i915_private *i915 = error->i915; 1959 1960 error->wakelock = atomic_read(&i915->runtime_pm.wakeref_count); 1961 error->suspended = i915->runtime_pm.suspended; 1962 1963 error->iommu = i915_vtd_active(i915); 1964 error->reset_count = i915_reset_count(&i915->gpu_error); 1965 error->suspend_count = i915->suspend_count; 1966 1967 i915_params_copy(&error->params, &i915->params); 1968 memcpy(&error->device_info, 1969 INTEL_INFO(i915), 1970 sizeof(error->device_info)); 1971 memcpy(&error->runtime_info, 1972 RUNTIME_INFO(i915), 1973 sizeof(error->runtime_info)); 1974 error->driver_caps = i915->caps; 1975 } 1976 1977 struct i915_gpu_coredump * 1978 i915_gpu_coredump_alloc(struct drm_i915_private *i915, gfp_t gfp) 1979 { 1980 struct i915_gpu_coredump *error; 1981 1982 if (!i915->params.error_capture) 1983 return NULL; 1984 1985 error = kzalloc(sizeof(*error), gfp); 1986 if (!error) 1987 return NULL; 1988 1989 kref_init(&error->ref); 1990 error->i915 = i915; 1991 1992 error->time = ktime_get_real(); 1993 error->boottime = ktime_get_boottime(); 1994 error->uptime = ktime_sub(ktime_get(), to_gt(i915)->last_init_time); 1995 error->capture = jiffies; 1996 1997 capture_gen(error); 1998 1999 return error; 2000 } 2001 2002 #define DAY_AS_SECONDS(x) (24 * 60 * 60 * (x)) 2003 2004 struct intel_gt_coredump * 2005 intel_gt_coredump_alloc(struct intel_gt *gt, gfp_t gfp, u32 dump_flags) 2006 { 2007 struct intel_gt_coredump *gc; 2008 2009 gc = kzalloc(sizeof(*gc), gfp); 2010 if (!gc) 2011 return NULL; 2012 2013 gc->_gt = gt; 2014 gc->awake = intel_gt_pm_is_awake(gt); 2015 2016 gt_record_display_regs(gc); 2017 gt_record_global_nonguc_regs(gc); 2018 2019 /* 2020 * GuC dumps global, eng-class and eng-instance registers 2021 * (that can change as part of engine state during execution) 2022 * before an engine is reset due to a hung context. 2023 * GuC captures and reports all three groups of registers 2024 * together as a single set before the engine is reset. 2025 * Thus, if GuC triggered the context reset we retrieve 2026 * the register values as part of gt_record_engines. 2027 */ 2028 if (!(dump_flags & CORE_DUMP_FLAG_IS_GUC_CAPTURE)) 2029 gt_record_global_regs(gc); 2030 2031 gt_record_fences(gc); 2032 2033 return gc; 2034 } 2035 2036 struct i915_vma_compress * 2037 i915_vma_capture_prepare(struct intel_gt_coredump *gt) 2038 { 2039 struct i915_vma_compress *compress; 2040 2041 compress = kmalloc(sizeof(*compress), ALLOW_FAIL); 2042 if (!compress) 2043 return NULL; 2044 2045 if (!compress_init(compress)) { 2046 kfree(compress); 2047 return NULL; 2048 } 2049 2050 return compress; 2051 } 2052 2053 void i915_vma_capture_finish(struct intel_gt_coredump *gt, 2054 struct i915_vma_compress *compress) 2055 { 2056 if (!compress) 2057 return; 2058 2059 compress_fini(compress); 2060 kfree(compress); 2061 } 2062 2063 static struct i915_gpu_coredump * 2064 __i915_gpu_coredump(struct intel_gt *gt, intel_engine_mask_t engine_mask, u32 dump_flags) 2065 { 2066 struct drm_i915_private *i915 = gt->i915; 2067 struct i915_gpu_coredump *error; 2068 2069 /* Check if GPU capture has been disabled */ 2070 error = READ_ONCE(i915->gpu_error.first_error); 2071 if (IS_ERR(error)) 2072 return error; 2073 2074 error = i915_gpu_coredump_alloc(i915, ALLOW_FAIL); 2075 if (!error) 2076 return ERR_PTR(-ENOMEM); 2077 2078 error->gt = intel_gt_coredump_alloc(gt, ALLOW_FAIL, dump_flags); 2079 if (error->gt) { 2080 struct i915_vma_compress *compress; 2081 2082 compress = i915_vma_capture_prepare(error->gt); 2083 if (!compress) { 2084 kfree(error->gt); 2085 kfree(error); 2086 return ERR_PTR(-ENOMEM); 2087 } 2088 2089 if (INTEL_INFO(i915)->has_gt_uc) { 2090 error->gt->uc = gt_record_uc(error->gt, compress); 2091 if (error->gt->uc) { 2092 if (dump_flags & CORE_DUMP_FLAG_IS_GUC_CAPTURE) 2093 error->gt->uc->guc.is_guc_capture = true; 2094 else 2095 GEM_BUG_ON(error->gt->uc->guc.is_guc_capture); 2096 } 2097 } 2098 2099 gt_record_info(error->gt); 2100 gt_record_engines(error->gt, engine_mask, compress, dump_flags); 2101 2102 2103 i915_vma_capture_finish(error->gt, compress); 2104 2105 error->simulated |= error->gt->simulated; 2106 } 2107 2108 error->overlay = intel_overlay_capture_error_state(i915); 2109 2110 return error; 2111 } 2112 2113 struct i915_gpu_coredump * 2114 i915_gpu_coredump(struct intel_gt *gt, intel_engine_mask_t engine_mask, u32 dump_flags) 2115 { 2116 static DEFINE_MUTEX(capture_mutex); 2117 int ret = mutex_lock_interruptible(&capture_mutex); 2118 struct i915_gpu_coredump *dump; 2119 2120 if (ret) 2121 return ERR_PTR(ret); 2122 2123 dump = __i915_gpu_coredump(gt, engine_mask, dump_flags); 2124 mutex_unlock(&capture_mutex); 2125 2126 return dump; 2127 } 2128 2129 void i915_error_state_store(struct i915_gpu_coredump *error) 2130 { 2131 struct drm_i915_private *i915; 2132 static bool warned; 2133 2134 if (IS_ERR_OR_NULL(error)) 2135 return; 2136 2137 i915 = error->i915; 2138 drm_info(&i915->drm, "%s\n", error_msg(error)); 2139 2140 if (error->simulated || 2141 cmpxchg(&i915->gpu_error.first_error, NULL, error)) 2142 return; 2143 2144 i915_gpu_coredump_get(error); 2145 2146 if (!xchg(&warned, true) && 2147 ktime_get_real_seconds() - DRIVER_TIMESTAMP < DAY_AS_SECONDS(180)) { 2148 pr_info("GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.\n"); 2149 pr_info("Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/intel/issues/new.\n"); 2150 pr_info("Please see https://gitlab.freedesktop.org/drm/intel/-/wikis/How-to-file-i915-bugs for details.\n"); 2151 pr_info("drm/i915 developers can then reassign to the right component if it's not a kernel issue.\n"); 2152 pr_info("The GPU crash dump is required to analyze GPU hangs, so please always attach it.\n"); 2153 pr_info("GPU crash dump saved to /sys/class/drm/card%d/error\n", 2154 i915->drm.primary->index); 2155 } 2156 } 2157 2158 /** 2159 * i915_capture_error_state - capture an error record for later analysis 2160 * @gt: intel_gt which originated the hang 2161 * @engine_mask: hung engines 2162 * 2163 * 2164 * Should be called when an error is detected (either a hang or an error 2165 * interrupt) to capture error state from the time of the error. Fills 2166 * out a structure which becomes available in debugfs for user level tools 2167 * to pick up. 2168 */ 2169 void i915_capture_error_state(struct intel_gt *gt, 2170 intel_engine_mask_t engine_mask, u32 dump_flags) 2171 { 2172 struct i915_gpu_coredump *error; 2173 2174 error = i915_gpu_coredump(gt, engine_mask, dump_flags); 2175 if (IS_ERR(error)) { 2176 cmpxchg(>->i915->gpu_error.first_error, NULL, error); 2177 return; 2178 } 2179 2180 i915_error_state_store(error); 2181 i915_gpu_coredump_put(error); 2182 } 2183 2184 struct i915_gpu_coredump * 2185 i915_first_error_state(struct drm_i915_private *i915) 2186 { 2187 struct i915_gpu_coredump *error; 2188 2189 spin_lock_irq(&i915->gpu_error.lock); 2190 error = i915->gpu_error.first_error; 2191 if (!IS_ERR_OR_NULL(error)) 2192 i915_gpu_coredump_get(error); 2193 spin_unlock_irq(&i915->gpu_error.lock); 2194 2195 return error; 2196 } 2197 2198 void i915_reset_error_state(struct drm_i915_private *i915) 2199 { 2200 struct i915_gpu_coredump *error; 2201 2202 spin_lock_irq(&i915->gpu_error.lock); 2203 error = i915->gpu_error.first_error; 2204 if (error != ERR_PTR(-ENODEV)) /* if disabled, always disabled */ 2205 i915->gpu_error.first_error = NULL; 2206 spin_unlock_irq(&i915->gpu_error.lock); 2207 2208 if (!IS_ERR_OR_NULL(error)) 2209 i915_gpu_coredump_put(error); 2210 } 2211 2212 void i915_disable_error_state(struct drm_i915_private *i915, int err) 2213 { 2214 spin_lock_irq(&i915->gpu_error.lock); 2215 if (!i915->gpu_error.first_error) 2216 i915->gpu_error.first_error = ERR_PTR(err); 2217 spin_unlock_irq(&i915->gpu_error.lock); 2218 } 2219