1// SPDX-License-Identifier: GPL-2.0 2 3/* 4 * Copyright 2016-2022 HabanaLabs, Ltd. 5 * All Rights Reserved. 6 */ 7 8#include <linux/slab.h> 9 10#include "../habanalabs.h" 11 12#include <trace/events/habanalabs.h> 13 14/** 15 * hl_mmu_get_funcs() - get MMU functions structure 16 * @hdev: habanalabs device structure. 17 * @pgt_residency: page table residency. 18 * @is_dram_addr: true if we need HMMU functions 19 * 20 * @return appropriate MMU functions structure 21 */ 22static struct hl_mmu_funcs *hl_mmu_get_funcs(struct hl_device *hdev, int pgt_residency, 23 bool is_dram_addr) 24{ 25 return &hdev->mmu_func[pgt_residency]; 26} 27 28bool hl_is_dram_va(struct hl_device *hdev, u64 virt_addr) 29{ 30 struct asic_fixed_properties *prop = &hdev->asic_prop; 31 32 return hl_mem_area_inside_range(virt_addr, prop->dmmu.page_size, 33 prop->dmmu.start_addr, 34 prop->dmmu.end_addr); 35} 36 37/** 38 * hl_mmu_init() - initialize the MMU module. 39 * @hdev: habanalabs device structure. 40 * 41 * Return: 0 for success, non-zero for failure. 42 */ 43int hl_mmu_init(struct hl_device *hdev) 44{ 45 int rc = -EOPNOTSUPP; 46 47 if (hdev->mmu_disable) 48 return 0; 49 50 mutex_init(&hdev->mmu_lock); 51 52 if (hdev->mmu_func[MMU_DR_PGT].init != NULL) { 53 rc = hdev->mmu_func[MMU_DR_PGT].init(hdev); 54 if (rc) 55 return rc; 56 } 57 58 if (hdev->mmu_func[MMU_HR_PGT].init != NULL) { 59 rc = hdev->mmu_func[MMU_HR_PGT].init(hdev); 60 if (rc) 61 goto fini_dr_mmu; 62 } 63 64 return 0; 65 66fini_dr_mmu: 67 if (hdev->mmu_func[MMU_DR_PGT].fini != NULL) 68 hdev->mmu_func[MMU_DR_PGT].fini(hdev); 69 70 return rc; 71} 72 73/** 74 * hl_mmu_fini() - release the MMU module. 75 * @hdev: habanalabs device structure. 76 * 77 * This function does the following: 78 * - Disable MMU in H/W. 79 * - Free the pgt_infos pool. 80 * 81 * All contexts should be freed before calling this function. 82 */ 83void hl_mmu_fini(struct hl_device *hdev) 84{ 85 if (hdev->mmu_disable) 86 return; 87 88 if (hdev->mmu_func[MMU_DR_PGT].fini != NULL) 89 hdev->mmu_func[MMU_DR_PGT].fini(hdev); 90 91 if (hdev->mmu_func[MMU_HR_PGT].fini != NULL) 92 hdev->mmu_func[MMU_HR_PGT].fini(hdev); 93 94 mutex_destroy(&hdev->mmu_lock); 95} 96 97/** 98 * hl_mmu_ctx_init() - initialize a context for using the MMU module. 99 * @ctx: pointer to the context structure to initialize. 100 * 101 * Initialize a mutex to protect the concurrent mapping flow, a hash to hold all 102 * page tables hops related to this context. 103 * Return: 0 on success, non-zero otherwise. 104 */ 105int hl_mmu_ctx_init(struct hl_ctx *ctx) 106{ 107 struct hl_device *hdev = ctx->hdev; 108 int rc = -EOPNOTSUPP; 109 110 if (hdev->mmu_disable) 111 return 0; 112 113 if (hdev->mmu_func[MMU_DR_PGT].ctx_init != NULL) { 114 rc = hdev->mmu_func[MMU_DR_PGT].ctx_init(ctx); 115 if (rc) 116 return rc; 117 } 118 119 if (hdev->mmu_func[MMU_HR_PGT].ctx_init != NULL) { 120 rc = hdev->mmu_func[MMU_HR_PGT].ctx_init(ctx); 121 if (rc) 122 goto fini_dr_ctx; 123 } 124 125 return 0; 126 127fini_dr_ctx: 128 if (hdev->mmu_func[MMU_DR_PGT].fini != NULL) 129 hdev->mmu_func[MMU_DR_PGT].fini(hdev); 130 131 return rc; 132} 133 134/* 135 * hl_mmu_ctx_fini - disable a ctx from using the mmu module 136 * 137 * @ctx: pointer to the context structure 138 * 139 * This function does the following: 140 * - Free any pgts which were not freed yet 141 * - Free the mutex 142 * - Free DRAM default page mapping hops 143 */ 144void hl_mmu_ctx_fini(struct hl_ctx *ctx) 145{ 146 struct hl_device *hdev = ctx->hdev; 147 148 if (hdev->mmu_disable) 149 return; 150 151 if (hdev->mmu_func[MMU_DR_PGT].ctx_fini != NULL) 152 hdev->mmu_func[MMU_DR_PGT].ctx_fini(ctx); 153 154 if (hdev->mmu_func[MMU_HR_PGT].ctx_fini != NULL) 155 hdev->mmu_func[MMU_HR_PGT].ctx_fini(ctx); 156} 157 158/* 159 * hl_mmu_get_real_page_size - get real page size to use in map/unmap operation 160 * 161 * @hdev: pointer to device data. 162 * @mmu_prop: MMU properties. 163 * @page_size: page size 164 * @real_page_size: set here the actual page size to use for the operation 165 * @is_dram_addr: true if DRAM address, otherwise false. 166 * 167 * @return 0 on success, otherwise non 0 error code 168 * 169 * note that this is general implementation that can fit most MMU arch. but as this is used as an 170 * MMU function: 171 * 1. it shall not be called directly- only from mmu_func structure instance 172 * 2. each MMU may modify the implementation internally 173 */ 174int hl_mmu_get_real_page_size(struct hl_device *hdev, struct hl_mmu_properties *mmu_prop, 175 u32 page_size, u32 *real_page_size, bool is_dram_addr) 176{ 177 /* 178 * The H/W handles mapping of specific page sizes. Hence if the page 179 * size is bigger, we break it to sub-pages and map them separately. 180 */ 181 if ((page_size % mmu_prop->page_size) == 0) { 182 *real_page_size = mmu_prop->page_size; 183 return 0; 184 } 185 186 dev_err(hdev->dev, "page size of %u is not %uKB aligned, can't map\n", 187 page_size, mmu_prop->page_size >> 10); 188 189 return -EFAULT; 190} 191 192static struct hl_mmu_properties *hl_mmu_get_prop(struct hl_device *hdev, u32 page_size, 193 bool is_dram_addr) 194{ 195 struct asic_fixed_properties *prop = &hdev->asic_prop; 196 197 if (is_dram_addr) 198 return &prop->dmmu; 199 else if ((page_size % prop->pmmu_huge.page_size) == 0) 200 return &prop->pmmu_huge; 201 202 return &prop->pmmu; 203} 204 205/* 206 * hl_mmu_unmap_page - unmaps a virtual addr 207 * 208 * @ctx: pointer to the context structure 209 * @virt_addr: virt addr to map from 210 * @page_size: size of the page to unmap 211 * @flush_pte: whether to do a PCI flush 212 * 213 * This function does the following: 214 * - Check that the virt addr is mapped 215 * - Unmap the virt addr and frees pgts if possible 216 * - Returns 0 on success, -EINVAL if the given addr is not mapped 217 * 218 * Because this function changes the page tables in the device and because it 219 * changes the MMU hash, it must be protected by a lock. 220 * However, because it maps only a single page, the lock should be implemented 221 * in a higher level in order to protect the entire mapping of the memory area 222 * 223 * For optimization reasons PCI flush may be requested once after unmapping of 224 * large area. 225 */ 226int hl_mmu_unmap_page(struct hl_ctx *ctx, u64 virt_addr, u32 page_size, bool flush_pte) 227{ 228 struct hl_device *hdev = ctx->hdev; 229 struct hl_mmu_properties *mmu_prop; 230 struct hl_mmu_funcs *mmu_funcs; 231 int i, pgt_residency, rc = 0; 232 u32 real_page_size, npages; 233 u64 real_virt_addr; 234 bool is_dram_addr; 235 236 if (hdev->mmu_disable) 237 return 0; 238 239 is_dram_addr = hl_is_dram_va(hdev, virt_addr); 240 mmu_prop = hl_mmu_get_prop(hdev, page_size, is_dram_addr); 241 242 pgt_residency = mmu_prop->host_resident ? MMU_HR_PGT : MMU_DR_PGT; 243 mmu_funcs = hl_mmu_get_funcs(hdev, pgt_residency, is_dram_addr); 244 245 rc = hdev->asic_funcs->mmu_get_real_page_size(hdev, mmu_prop, page_size, &real_page_size, 246 is_dram_addr); 247 if (rc) 248 return rc; 249 250 npages = page_size / real_page_size; 251 real_virt_addr = virt_addr; 252 253 for (i = 0 ; i < npages ; i++) { 254 rc = mmu_funcs->unmap(ctx, real_virt_addr, is_dram_addr); 255 if (rc) 256 break; 257 258 real_virt_addr += real_page_size; 259 } 260 261 if (flush_pte) 262 mmu_funcs->flush(ctx); 263 264 if (trace_habanalabs_mmu_unmap_enabled() && !rc) 265 trace_habanalabs_mmu_unmap(hdev->dev, virt_addr, 0, page_size, flush_pte); 266 267 return rc; 268} 269 270/* 271 * hl_mmu_map_page - maps a virtual addr to physical addr 272 * 273 * @ctx: pointer to the context structure 274 * @virt_addr: virt addr to map from 275 * @phys_addr: phys addr to map to 276 * @page_size: physical page size 277 * @flush_pte: whether to do a PCI flush 278 * 279 * This function does the following: 280 * - Check that the virt addr is not mapped 281 * - Allocate pgts as necessary in order to map the virt addr to the phys 282 * - Returns 0 on success, -EINVAL if addr is already mapped, or -ENOMEM. 283 * 284 * Because this function changes the page tables in the device and because it 285 * changes the MMU hash, it must be protected by a lock. 286 * However, because it maps only a single page, the lock should be implemented 287 * in a higher level in order to protect the entire mapping of the memory area 288 * 289 * For optimization reasons PCI flush may be requested once after mapping of 290 * large area. 291 */ 292int hl_mmu_map_page(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr, u32 page_size, 293 bool flush_pte) 294{ 295 int i, rc, pgt_residency, mapped_cnt = 0; 296 struct hl_device *hdev = ctx->hdev; 297 struct hl_mmu_properties *mmu_prop; 298 u64 real_virt_addr, real_phys_addr; 299 struct hl_mmu_funcs *mmu_funcs; 300 u32 real_page_size, npages; 301 bool is_dram_addr; 302 303 304 if (hdev->mmu_disable) 305 return 0; 306 307 is_dram_addr = hl_is_dram_va(hdev, virt_addr); 308 mmu_prop = hl_mmu_get_prop(hdev, page_size, is_dram_addr); 309 310 pgt_residency = mmu_prop->host_resident ? MMU_HR_PGT : MMU_DR_PGT; 311 mmu_funcs = hl_mmu_get_funcs(hdev, pgt_residency, is_dram_addr); 312 313 rc = hdev->asic_funcs->mmu_get_real_page_size(hdev, mmu_prop, page_size, &real_page_size, 314 is_dram_addr); 315 if (rc) 316 return rc; 317 318 /* 319 * Verify that the phys and virt addresses are aligned with the 320 * MMU page size (in dram this means checking the address and MMU 321 * after scrambling) 322 */ 323 if ((is_dram_addr && 324 ((hdev->asic_funcs->scramble_addr(hdev, phys_addr) & 325 (mmu_prop->page_size - 1)) || 326 (hdev->asic_funcs->scramble_addr(hdev, virt_addr) & 327 (mmu_prop->page_size - 1)))) || 328 (!is_dram_addr && ((phys_addr & (real_page_size - 1)) || 329 (virt_addr & (real_page_size - 1))))) 330 dev_crit(hdev->dev, 331 "Mapping address 0x%llx with virtual address 0x%llx and page size of 0x%x is erroneous! Addresses must be divisible by page size", 332 phys_addr, virt_addr, real_page_size); 333 334 npages = page_size / real_page_size; 335 real_virt_addr = virt_addr; 336 real_phys_addr = phys_addr; 337 338 for (i = 0 ; i < npages ; i++) { 339 rc = mmu_funcs->map(ctx, real_virt_addr, real_phys_addr, real_page_size, 340 is_dram_addr); 341 if (rc) 342 goto err; 343 344 real_virt_addr += real_page_size; 345 real_phys_addr += real_page_size; 346 mapped_cnt++; 347 } 348 349 if (flush_pte) 350 mmu_funcs->flush(ctx); 351 352 trace_habanalabs_mmu_map(hdev->dev, virt_addr, phys_addr, page_size, flush_pte); 353 354 return 0; 355 356err: 357 real_virt_addr = virt_addr; 358 for (i = 0 ; i < mapped_cnt ; i++) { 359 if (mmu_funcs->unmap(ctx, real_virt_addr, is_dram_addr)) 360 dev_warn_ratelimited(hdev->dev, 361 "failed to unmap va: 0x%llx\n", real_virt_addr); 362 363 real_virt_addr += real_page_size; 364 } 365 366 mmu_funcs->flush(ctx); 367 368 return rc; 369} 370 371/* 372 * hl_mmu_map_contiguous - implements a wrapper for hl_mmu_map_page 373 * for mapping contiguous physical memory 374 * 375 * @ctx: pointer to the context structure 376 * @virt_addr: virt addr to map from 377 * @phys_addr: phys addr to map to 378 * @size: size to map 379 * 380 */ 381int hl_mmu_map_contiguous(struct hl_ctx *ctx, u64 virt_addr, 382 u64 phys_addr, u32 size) 383{ 384 struct hl_device *hdev = ctx->hdev; 385 struct asic_fixed_properties *prop = &hdev->asic_prop; 386 u64 curr_va, curr_pa; 387 u32 page_size; 388 bool flush_pte; 389 int rc = 0, off; 390 391 if (hl_mem_area_inside_range(virt_addr, size, 392 prop->dmmu.start_addr, prop->dmmu.end_addr)) 393 page_size = prop->dmmu.page_size; 394 else if (hl_mem_area_inside_range(virt_addr, size, 395 prop->pmmu.start_addr, prop->pmmu.end_addr)) 396 page_size = prop->pmmu.page_size; 397 else if (hl_mem_area_inside_range(virt_addr, size, 398 prop->pmmu_huge.start_addr, prop->pmmu_huge.end_addr)) 399 page_size = prop->pmmu_huge.page_size; 400 else 401 return -EINVAL; 402 403 for (off = 0 ; off < size ; off += page_size) { 404 curr_va = virt_addr + off; 405 curr_pa = phys_addr + off; 406 flush_pte = (off + page_size) >= size; 407 rc = hl_mmu_map_page(ctx, curr_va, curr_pa, page_size, 408 flush_pte); 409 if (rc) { 410 dev_err(hdev->dev, 411 "Map failed for va 0x%llx to pa 0x%llx\n", 412 curr_va, curr_pa); 413 /* last mapping failed so don't try to unmap it - reduce off by page_size */ 414 off -= page_size; 415 goto unmap; 416 } 417 } 418 419 return rc; 420 421unmap: 422 for (; off >= 0 ; off -= page_size) { 423 curr_va = virt_addr + off; 424 flush_pte = (off - (s32) page_size) < 0; 425 if (hl_mmu_unmap_page(ctx, curr_va, page_size, flush_pte)) 426 dev_warn_ratelimited(hdev->dev, 427 "failed to unmap va 0x%llx\n", curr_va); 428 } 429 430 return rc; 431} 432 433/* 434 * hl_mmu_unmap_contiguous - implements a wrapper for hl_mmu_unmap_page 435 * for unmapping contiguous physical memory 436 * 437 * @ctx: pointer to the context structure 438 * @virt_addr: virt addr to unmap 439 * @size: size to unmap 440 * 441 */ 442int hl_mmu_unmap_contiguous(struct hl_ctx *ctx, u64 virt_addr, u32 size) 443{ 444 struct hl_device *hdev = ctx->hdev; 445 struct asic_fixed_properties *prop = &hdev->asic_prop; 446 u64 curr_va; 447 u32 page_size; 448 bool flush_pte; 449 int rc = 0, off; 450 451 if (hl_mem_area_inside_range(virt_addr, size, 452 prop->dmmu.start_addr, prop->dmmu.end_addr)) 453 page_size = prop->dmmu.page_size; 454 else if (hl_mem_area_inside_range(virt_addr, size, 455 prop->pmmu.start_addr, prop->pmmu.end_addr)) 456 page_size = prop->pmmu.page_size; 457 else if (hl_mem_area_inside_range(virt_addr, size, 458 prop->pmmu_huge.start_addr, prop->pmmu_huge.end_addr)) 459 page_size = prop->pmmu_huge.page_size; 460 else 461 return -EINVAL; 462 463 for (off = 0 ; off < size ; off += page_size) { 464 curr_va = virt_addr + off; 465 flush_pte = (off + page_size) >= size; 466 rc = hl_mmu_unmap_page(ctx, curr_va, page_size, flush_pte); 467 if (rc) 468 dev_warn_ratelimited(hdev->dev, 469 "Unmap failed for va 0x%llx\n", curr_va); 470 } 471 472 return rc; 473} 474 475static void hl_mmu_pa_page_with_offset(struct hl_ctx *ctx, u64 virt_addr, 476 struct hl_mmu_hop_info *hops, 477 u64 *phys_addr) 478{ 479 struct asic_fixed_properties *prop = &ctx->hdev->asic_prop; 480 u64 offset_mask, addr_mask, hop_shift, tmp_phys_addr; 481 struct hl_mmu_properties *mmu_prop; 482 483 /* last hop holds the phys address and flags */ 484 if (hops->unscrambled_paddr) 485 tmp_phys_addr = hops->unscrambled_paddr; 486 else 487 tmp_phys_addr = hops->hop_info[hops->used_hops - 1].hop_pte_val; 488 489 if (hops->range_type == HL_VA_RANGE_TYPE_HOST_HUGE) 490 mmu_prop = &prop->pmmu_huge; 491 else if (hops->range_type == HL_VA_RANGE_TYPE_HOST) 492 mmu_prop = &prop->pmmu; 493 else /* HL_VA_RANGE_TYPE_DRAM */ 494 mmu_prop = &prop->dmmu; 495 496 if ((hops->range_type == HL_VA_RANGE_TYPE_DRAM) && 497 !is_power_of_2(prop->dram_page_size)) { 498 u64 dram_page_size, dram_base, abs_phys_addr, abs_virt_addr, 499 page_id, page_start; 500 u32 page_off; 501 502 /* 503 * Bit arithmetic cannot be used for non power of two page 504 * sizes. In addition, since bit arithmetic is not used, 505 * we cannot ignore dram base. All that shall be considered. 506 */ 507 508 dram_page_size = prop->dram_page_size; 509 dram_base = prop->dram_base_address; 510 abs_phys_addr = tmp_phys_addr - dram_base; 511 abs_virt_addr = virt_addr - dram_base; 512 page_id = DIV_ROUND_DOWN_ULL(abs_phys_addr, dram_page_size); 513 page_start = page_id * dram_page_size; 514 div_u64_rem(abs_virt_addr, dram_page_size, &page_off); 515 516 *phys_addr = page_start + page_off + dram_base; 517 } else { 518 /* 519 * find the correct hop shift field in hl_mmu_properties 520 * structure in order to determine the right masks 521 * for the page offset. 522 */ 523 hop_shift = mmu_prop->hop_shifts[hops->used_hops - 1]; 524 offset_mask = (1ull << hop_shift) - 1; 525 addr_mask = ~(offset_mask); 526 *phys_addr = (tmp_phys_addr & addr_mask) | 527 (virt_addr & offset_mask); 528 } 529} 530 531int hl_mmu_va_to_pa(struct hl_ctx *ctx, u64 virt_addr, u64 *phys_addr) 532{ 533 struct hl_mmu_hop_info hops; 534 int rc; 535 536 memset(&hops, 0, sizeof(hops)); 537 538 rc = hl_mmu_get_tlb_info(ctx, virt_addr, &hops); 539 if (rc) 540 return rc; 541 542 hl_mmu_pa_page_with_offset(ctx, virt_addr, &hops, phys_addr); 543 544 return 0; 545} 546 547int hl_mmu_get_tlb_info(struct hl_ctx *ctx, u64 virt_addr, 548 struct hl_mmu_hop_info *hops) 549{ 550 struct hl_device *hdev = ctx->hdev; 551 struct asic_fixed_properties *prop; 552 struct hl_mmu_properties *mmu_prop; 553 struct hl_mmu_funcs *mmu_funcs; 554 int pgt_residency, rc; 555 bool is_dram_addr; 556 557 if (hdev->mmu_disable) 558 return -EOPNOTSUPP; 559 560 prop = &hdev->asic_prop; 561 hops->scrambled_vaddr = virt_addr; /* assume no scrambling */ 562 563 is_dram_addr = hl_mem_area_inside_range(virt_addr, prop->dmmu.page_size, 564 prop->dmmu.start_addr, 565 prop->dmmu.end_addr); 566 567 /* host-residency is the same in PMMU and PMMU huge, no need to distinguish here */ 568 mmu_prop = is_dram_addr ? &prop->dmmu : &prop->pmmu; 569 pgt_residency = mmu_prop->host_resident ? MMU_HR_PGT : MMU_DR_PGT; 570 mmu_funcs = hl_mmu_get_funcs(hdev, pgt_residency, is_dram_addr); 571 572 mutex_lock(&hdev->mmu_lock); 573 rc = mmu_funcs->get_tlb_info(ctx, virt_addr, hops); 574 mutex_unlock(&hdev->mmu_lock); 575 576 if (rc) 577 return rc; 578 579 /* add page offset to physical address */ 580 if (hops->unscrambled_paddr) 581 hl_mmu_pa_page_with_offset(ctx, virt_addr, hops, &hops->unscrambled_paddr); 582 583 return 0; 584} 585 586int hl_mmu_if_set_funcs(struct hl_device *hdev) 587{ 588 if (hdev->mmu_disable) 589 return 0; 590 591 switch (hdev->asic_type) { 592 case ASIC_GOYA: 593 case ASIC_GAUDI: 594 case ASIC_GAUDI_SEC: 595 hl_mmu_v1_set_funcs(hdev, &hdev->mmu_func[MMU_DR_PGT]); 596 break; 597 case ASIC_GAUDI2: 598 case ASIC_GAUDI2B: 599 case ASIC_GAUDI2C: 600 /* MMUs in Gaudi2 are always host resident */ 601 hl_mmu_v2_hr_set_funcs(hdev, &hdev->mmu_func[MMU_HR_PGT]); 602 break; 603 default: 604 dev_err(hdev->dev, "Unrecognized ASIC type %d\n", 605 hdev->asic_type); 606 return -EOPNOTSUPP; 607 } 608 609 return 0; 610} 611 612/** 613 * hl_mmu_scramble_addr() - The generic mmu address scrambling routine. 614 * @hdev: pointer to device data. 615 * @addr: The address to scramble. 616 * 617 * Return: The scrambled address. 618 */ 619u64 hl_mmu_scramble_addr(struct hl_device *hdev, u64 addr) 620{ 621 return addr; 622} 623 624/** 625 * hl_mmu_descramble_addr() - The generic mmu address descrambling 626 * routine. 627 * @hdev: pointer to device data. 628 * @addr: The address to descramble. 629 * 630 * Return: The un-scrambled address. 631 */ 632u64 hl_mmu_descramble_addr(struct hl_device *hdev, u64 addr) 633{ 634 return addr; 635} 636 637int hl_mmu_invalidate_cache(struct hl_device *hdev, bool is_hard, u32 flags) 638{ 639 int rc; 640 641 rc = hdev->asic_funcs->mmu_invalidate_cache(hdev, is_hard, flags); 642 if (rc) 643 dev_err_ratelimited(hdev->dev, 644 "%s cache invalidation failed, rc=%d\n", 645 flags == VM_TYPE_USERPTR ? "PMMU" : "HMMU", rc); 646 647 return rc; 648} 649 650int hl_mmu_invalidate_cache_range(struct hl_device *hdev, bool is_hard, 651 u32 flags, u32 asid, u64 va, u64 size) 652{ 653 int rc; 654 655 rc = hdev->asic_funcs->mmu_invalidate_cache_range(hdev, is_hard, flags, 656 asid, va, size); 657 if (rc) 658 dev_err_ratelimited(hdev->dev, 659 "%s cache range invalidation failed: va=%#llx, size=%llu, rc=%d", 660 flags == VM_TYPE_USERPTR ? "PMMU" : "HMMU", va, size, rc); 661 662 return rc; 663} 664 665static void hl_mmu_prefetch_work_function(struct work_struct *work) 666{ 667 struct hl_prefetch_work *pfw = container_of(work, struct hl_prefetch_work, prefetch_work); 668 struct hl_ctx *ctx = pfw->ctx; 669 struct hl_device *hdev = ctx->hdev; 670 671 if (!hl_device_operational(hdev, NULL)) 672 goto put_ctx; 673 674 mutex_lock(&hdev->mmu_lock); 675 676 hdev->asic_funcs->mmu_prefetch_cache_range(ctx, pfw->flags, pfw->asid, pfw->va, pfw->size); 677 678 mutex_unlock(&hdev->mmu_lock); 679 680put_ctx: 681 /* 682 * context was taken in the common mmu prefetch function- see comment there about 683 * context handling. 684 */ 685 hl_ctx_put(ctx); 686 kfree(pfw); 687} 688 689int hl_mmu_prefetch_cache_range(struct hl_ctx *ctx, u32 flags, u32 asid, u64 va, u64 size) 690{ 691 struct hl_prefetch_work *handle_prefetch_work; 692 693 handle_prefetch_work = kmalloc(sizeof(*handle_prefetch_work), GFP_KERNEL); 694 if (!handle_prefetch_work) 695 return -ENOMEM; 696 697 INIT_WORK(&handle_prefetch_work->prefetch_work, hl_mmu_prefetch_work_function); 698 handle_prefetch_work->ctx = ctx; 699 handle_prefetch_work->va = va; 700 handle_prefetch_work->size = size; 701 handle_prefetch_work->flags = flags; 702 handle_prefetch_work->asid = asid; 703 704 /* 705 * as actual prefetch is done in a WQ we must get the context (and put it 706 * at the end of the work function) 707 */ 708 hl_ctx_get(ctx); 709 queue_work(ctx->hdev->prefetch_wq, &handle_prefetch_work->prefetch_work); 710 711 return 0; 712} 713 714u64 hl_mmu_get_next_hop_addr(struct hl_ctx *ctx, u64 curr_pte) 715{ 716 return (curr_pte & PAGE_PRESENT_MASK) ? (curr_pte & HOP_PHYS_ADDR_MASK) : ULLONG_MAX; 717} 718 719/** 720 * hl_mmu_get_hop_pte_phys_addr() - extract PTE address from HOP 721 * @ctx: pointer to the context structure to initialize. 722 * @mmu_prop: MMU properties. 723 * @hop_idx: HOP index. 724 * @hop_addr: HOP address. 725 * @virt_addr: virtual address for the translation. 726 * 727 * @return the matching PTE value on success, otherwise U64_MAX. 728 */ 729u64 hl_mmu_get_hop_pte_phys_addr(struct hl_ctx *ctx, struct hl_mmu_properties *mmu_prop, 730 u8 hop_idx, u64 hop_addr, u64 virt_addr) 731{ 732 u64 mask, shift; 733 734 if (hop_idx >= mmu_prop->num_hops) { 735 dev_err_ratelimited(ctx->hdev->dev, "Invalid hop index %d\n", hop_idx); 736 return U64_MAX; 737 } 738 739 shift = mmu_prop->hop_shifts[hop_idx]; 740 mask = mmu_prop->hop_masks[hop_idx]; 741 742 return hop_addr + ctx->hdev->asic_prop.mmu_pte_size * ((virt_addr & mask) >> shift); 743} 744 745static void mmu_dma_mem_free_from_chunk(struct gen_pool *pool, 746 struct gen_pool_chunk *chunk, 747 void *data) 748{ 749 struct hl_device *hdev = data; 750 751 hl_asic_dma_free_coherent(hdev, (chunk->end_addr - chunk->start_addr) + 1, 752 (void *)chunk->start_addr, chunk->phys_addr); 753} 754 755void hl_mmu_hr_flush(struct hl_ctx *ctx) 756{ 757 /* a flush operation requires memory barrier */ 758 mb(); 759} 760 761/** 762 * hl_mmu_hr_pool_destroy() - destroy genpool 763 * @hdev: habanalabs device structure. 764 * @hr_priv: MMU HR private data. 765 * @hop_table_size: HOP table size. 766 * 767 * This function does the following: 768 * - free entries allocated for shadow HOP0 769 * - free pool chunks 770 * - free pool 771 */ 772static void hl_mmu_hr_pool_destroy(struct hl_device *hdev, struct hl_mmu_hr_priv *hr_priv, 773 u32 hop_table_size) 774{ 775 struct asic_fixed_properties *prop = &hdev->asic_prop; 776 struct gen_pool **pool = &hr_priv->mmu_pgt_pool; 777 struct pgt_info *hop0_pgt; 778 int asid; 779 780 if (ZERO_OR_NULL_PTR(*pool)) 781 return; 782 783 /* Free the Fixed allocation of HOPs0 */ 784 if (hr_priv->mmu_asid_hop0) { 785 for (asid = 0 ; asid < prop->max_asid ; asid++) { 786 hop0_pgt = &hr_priv->mmu_asid_hop0[asid]; 787 if (ZERO_OR_NULL_PTR(hop0_pgt->virt_addr)) 788 continue; 789 790 gen_pool_free(*pool, (uintptr_t) hop0_pgt->virt_addr, hop_table_size); 791 } 792 } 793 794 gen_pool_for_each_chunk(*pool, mmu_dma_mem_free_from_chunk, hdev); 795 gen_pool_destroy(*pool); 796 797 /* Make sure that if we arrive here again without init was called we 798 * won't cause kernel panic. This can happen for example if we fail 799 * during hard reset code at certain points 800 */ 801 *pool = NULL; 802} 803 804/** 805 * hl_mmu_hr_init() - initialize the MMU module. 806 * @hdev: habanalabs device structure. 807 * @hr_priv: MMU HR private data. 808 * @hop_table_size: HOP table size. 809 * @pgt_size: memory size allocated for the page table 810 * 811 * @return 0 on success otherwise non-zero error code 812 * 813 * This function does the following: 814 * - Create a pool of pages for pgt_infos. 815 * - Create a shadow table for pgt 816 */ 817int hl_mmu_hr_init(struct hl_device *hdev, struct hl_mmu_hr_priv *hr_priv, u32 hop_table_size, 818 u64 pgt_size) 819{ 820 struct asic_fixed_properties *prop = &hdev->asic_prop; 821 size_t pool_chunk_size = SZ_4M; 822 struct pgt_info *hop0_pgt; 823 dma_addr_t dma_addr; 824 u64 virt_addr; 825 int i, rc; 826 827 /* 828 * we set alloc size as PAGE_SIZE (sine dma_alloc_coherent allocation order/size is 829 * PAGE_SHIFT/PAGE_SIZE) in order to be able to control the allocations alignment. 830 * This way we can call "DMA alloc align" according to dma_alloc granularity and supply 831 * allocations with higher-order alignment restrictions 832 */ 833 hr_priv->mmu_pgt_pool = gen_pool_create(PAGE_SHIFT, -1); 834 if (ZERO_OR_NULL_PTR(hr_priv->mmu_pgt_pool)) { 835 dev_err(hdev->dev, "Failed to create hr page pool\n"); 836 return -ENOMEM; 837 } 838 839 hr_priv->mmu_asid_hop0 = kvcalloc(prop->max_asid, sizeof(struct pgt_info), GFP_KERNEL); 840 if (ZERO_OR_NULL_PTR(hr_priv->mmu_asid_hop0)) { 841 dev_err(hdev->dev, "Failed to allocate hr-mmu hop0 table\n"); 842 rc = -ENOMEM; 843 goto destroy_mmu_pgt_pool; 844 } 845 846 for (i = 0 ; i < pgt_size ; i += pool_chunk_size) { 847 virt_addr = (uintptr_t) hl_asic_dma_alloc_coherent(hdev, pool_chunk_size, 848 &dma_addr, 849 GFP_KERNEL | __GFP_ZERO); 850 if (ZERO_OR_NULL_PTR(virt_addr)) { 851 dev_err(hdev->dev, 852 "Failed to allocate memory for host-resident page pool\n"); 853 rc = -ENOMEM; 854 goto destroy_mmu_pgt_pool; 855 } 856 857 rc = gen_pool_add_virt(hr_priv->mmu_pgt_pool, virt_addr, (phys_addr_t) dma_addr, 858 pool_chunk_size, -1); 859 if (rc) { 860 dev_err(hdev->dev, "Failed to fill host-resident page pool\n"); 861 goto destroy_mmu_pgt_pool; 862 } 863 } 864 865 for (i = 0 ; i < prop->max_asid ; i++) { 866 hop0_pgt = &hr_priv->mmu_asid_hop0[i]; 867 hop0_pgt->virt_addr = (uintptr_t) 868 gen_pool_dma_zalloc_align(hr_priv->mmu_pgt_pool, 869 hop_table_size, 870 (dma_addr_t *) &hop0_pgt->phys_addr, 871 hop_table_size); 872 if (!hop0_pgt->virt_addr) { 873 dev_err(hdev->dev, "Failed to allocate HOP from pgt pool\n"); 874 rc = -ENOMEM; 875 goto destroy_mmu_pgt_pool; 876 } 877 } 878 879 /* MMU H/W init will be done in device hw_init() */ 880 881 return 0; 882 883destroy_mmu_pgt_pool: 884 hl_mmu_hr_pool_destroy(hdev, hr_priv, hop_table_size); 885 if (!ZERO_OR_NULL_PTR(hr_priv->mmu_asid_hop0)) 886 kvfree(hr_priv->mmu_asid_hop0); 887 888 return rc; 889} 890 891/** 892 * hl_mmu_hr_fini() - release the MMU module. 893 * @hdev: habanalabs device structure. 894 * @hr_priv: MMU host resident private info. 895 * @hop_table_size: HOP table size 896 * 897 * This function does the following: 898 * - Disable MMU in H/W. 899 * - Free the pgt_infos pool. 900 * 901 * All contexts should be freed before calling this function. 902 */ 903void hl_mmu_hr_fini(struct hl_device *hdev, struct hl_mmu_hr_priv *hr_priv, u32 hop_table_size) 904{ 905 /* MMU H/W fini was already done in device hw_fini() */ 906 907 hl_mmu_hr_pool_destroy(hdev, hr_priv, hop_table_size); 908 909 if (!ZERO_OR_NULL_PTR(hr_priv->mmu_asid_hop0)) { 910 kvfree(hr_priv->mmu_asid_hop0); 911 912 /* Make sure that if we arrive here again without init was 913 * called we won't cause kernel panic. This can happen for 914 * example if we fail during hard reset code at certain points 915 */ 916 hr_priv->mmu_asid_hop0 = NULL; 917 } 918} 919 920/** 921 * hl_mmu_hr_free_hop_remove_pgt() - free HOP and remove PGT from hash 922 * @pgt_info: page table info structure. 923 * @hr_priv: MMU HR private data. 924 * @hop_table_size: HOP table size. 925 */ 926void hl_mmu_hr_free_hop_remove_pgt(struct pgt_info *pgt_info, struct hl_mmu_hr_priv *hr_priv, 927 u32 hop_table_size) 928{ 929 gen_pool_free(hr_priv->mmu_pgt_pool, pgt_info->virt_addr, hop_table_size); 930 hash_del(&pgt_info->node); 931 kfree(pgt_info); 932} 933 934/** 935 * hl_mmu_hr_pte_phys_to_virt() - translate PTE phys addr to virt addr 936 * @ctx: pointer to the context structure 937 * @pgt: pgt_info for the HOP hosting the PTE 938 * @phys_pte_addr: phys address of the PTE 939 * @hop_table_size: HOP table size 940 * 941 * @return PTE virtual address 942 * 943 * The function use the pgt_info to get HOP base virt addr and obtain the PTE's virt addr 944 * by adding the PTE offset. 945 */ 946u64 hl_mmu_hr_pte_phys_to_virt(struct hl_ctx *ctx, struct pgt_info *pgt, 947 u64 phys_pte_addr, u32 hop_table_size) 948{ 949 u64 page_mask = (hop_table_size - 1); 950 u64 pte_offset = phys_pte_addr & page_mask; 951 952 return pgt->virt_addr + pte_offset; 953} 954 955/** 956 * hl_mmu_hr_write_pte() - write HR PTE 957 * @ctx: pointer to the context structure 958 * @pgt_info: HOP's page table info structure 959 * @phys_pte_addr: phys PTE address 960 * @val: raw PTE data 961 * @hop_table_size: HOP table size 962 */ 963void hl_mmu_hr_write_pte(struct hl_ctx *ctx, struct pgt_info *pgt_info, u64 phys_pte_addr, 964 u64 val, u32 hop_table_size) 965{ 966 /* 967 * The value to write is the phys address of the next hop + 968 * flags at the 12 LSBs. 969 */ 970 u64 virt_addr = hl_mmu_hr_pte_phys_to_virt(ctx, pgt_info, phys_pte_addr, hop_table_size); 971 972 *((u64 *) (uintptr_t) virt_addr) = val; 973} 974 975/** 976 * hl_mmu_hr_clear_pte() - clear HR PTE 977 * @ctx: pointer to the context structure 978 * @pgt_info: HOP's page table info structure 979 * @phys_pte_addr: phys PTE address 980 * @hop_table_size: HOP table size 981 */ 982void hl_mmu_hr_clear_pte(struct hl_ctx *ctx, struct pgt_info *pgt_info, u64 phys_pte_addr, 983 u32 hop_table_size) 984{ 985 /* no need to transform the value to physical address */ 986 hl_mmu_hr_write_pte(ctx, pgt_info, phys_pte_addr, 0, hop_table_size); 987} 988 989/** 990 * hl_mmu_hr_put_pte() - put HR PTE and remove it if necessary (no more PTEs) 991 * @ctx: pointer to the context structure 992 * @pgt_info: HOP's page table info structure 993 * @hr_priv: HR MMU private info 994 * @hop_table_size: HOP table size 995 * 996 * @return number of PTEs still in the HOP 997 */ 998int hl_mmu_hr_put_pte(struct hl_ctx *ctx, struct pgt_info *pgt_info, 999 struct hl_mmu_hr_priv *hr_priv, 1000 u32 hop_table_size) 1001{ 1002 int num_of_ptes_left; 1003 1004 pgt_info->num_of_ptes--; 1005 1006 /* 1007 * Need to save the number of ptes left because free_hop might free 1008 * the pgt_info 1009 */ 1010 num_of_ptes_left = pgt_info->num_of_ptes; 1011 if (!num_of_ptes_left) 1012 hl_mmu_hr_free_hop_remove_pgt(pgt_info, hr_priv, hop_table_size); 1013 1014 return num_of_ptes_left; 1015} 1016 1017/** 1018 * hl_mmu_hr_get_pte() - increase PGT PTE count 1019 * @ctx: pointer to the context structure 1020 * @hr_func: host resident functions 1021 * @phys_hop_addr: HOP phys address 1022 */ 1023void hl_mmu_hr_get_pte(struct hl_ctx *ctx, struct hl_hr_mmu_funcs *hr_func, u64 phys_hop_addr) 1024{ 1025 hr_func->get_pgt_info(ctx, phys_hop_addr)->num_of_ptes++; 1026} 1027 1028/** 1029 * hl_mmu_hr_get_next_hop_pgt_info() - get pgt_info structure for the next HOP 1030 * @ctx: pointer to the context structure. 1031 * @hr_func: host resident functions. 1032 * @curr_pte: current PTE value. 1033 * 1034 * @return pgt_info structure on success, otherwise NULL. 1035 */ 1036struct pgt_info *hl_mmu_hr_get_next_hop_pgt_info(struct hl_ctx *ctx, 1037 struct hl_hr_mmu_funcs *hr_func, 1038 u64 curr_pte) 1039{ 1040 u64 next_hop_phys_addr = hl_mmu_get_next_hop_addr(ctx, curr_pte); 1041 1042 if (next_hop_phys_addr == ULLONG_MAX) 1043 return NULL; 1044 1045 return hr_func->get_pgt_info(ctx, next_hop_phys_addr); 1046} 1047 1048/** 1049 * hl_mmu_hr_alloc_hop() - allocate HOP 1050 * @ctx: pointer to the context structure. 1051 * @hr_priv: host resident private info structure. 1052 * @hr_func: host resident functions. 1053 * @mmu_prop: MMU properties. 1054 * 1055 * @return pgt_info structure associated with the allocated HOP on success, otherwise NULL. 1056 */ 1057struct pgt_info *hl_mmu_hr_alloc_hop(struct hl_ctx *ctx, struct hl_mmu_hr_priv *hr_priv, 1058 struct hl_hr_mmu_funcs *hr_func, 1059 struct hl_mmu_properties *mmu_prop) 1060{ 1061 struct hl_device *hdev = ctx->hdev; 1062 struct pgt_info *pgt_info; 1063 dma_addr_t phys_addr; 1064 void *virt_addr; 1065 int i, retry = 1; 1066 1067 pgt_info = kmalloc(sizeof(*pgt_info), GFP_KERNEL); 1068 if (!pgt_info) 1069 return NULL; 1070 1071 for (i = 0; i <= retry; i++) { 1072 virt_addr = gen_pool_dma_zalloc_align(hr_priv->mmu_pgt_pool, 1073 mmu_prop->hop_table_size, 1074 &phys_addr, 1075 mmu_prop->hop_table_size); 1076 if (virt_addr) 1077 break; 1078 1079 /* No memory in pool - get some and try again */ 1080 virt_addr = hl_asic_dma_alloc_coherent(hdev, SZ_2M, &phys_addr, 1081 GFP_KERNEL | __GFP_ZERO); 1082 if (ZERO_OR_NULL_PTR(virt_addr)) 1083 break; 1084 1085 if (gen_pool_add_virt(hr_priv->mmu_pgt_pool, (unsigned long)virt_addr, 1086 phys_addr, SZ_2M, -1)) { 1087 hl_asic_dma_free_coherent(hdev, SZ_2M, virt_addr, phys_addr); 1088 virt_addr = NULL; 1089 break; 1090 } 1091 } 1092 1093 if (ZERO_OR_NULL_PTR(virt_addr)) { 1094 dev_err(hdev->dev, "failed to allocate page\n"); 1095 goto pool_alloc_err; 1096 } 1097 1098 pgt_info->phys_addr = phys_addr; 1099 pgt_info->shadow_addr = (unsigned long) NULL; 1100 pgt_info->virt_addr = (unsigned long)virt_addr; 1101 pgt_info->ctx = ctx; 1102 pgt_info->num_of_ptes = 0; 1103 hr_func->add_pgt_info(ctx, pgt_info, phys_addr); 1104 1105 return pgt_info; 1106 1107pool_alloc_err: 1108 kfree(pgt_info); 1109 1110 return NULL; 1111} 1112 1113/** 1114 * hl_mmu_hr_get_alloc_next_hop() - get the next HOP, allocate it if it does not exist 1115 * @ctx: pointer to the context structure. 1116 * @hr_priv: host resident private info structure. 1117 * @hr_func: host resident functions. 1118 * @mmu_prop: MMU properties. 1119 * @curr_pte: current PTE value. 1120 * @is_new_hop: set to true if HOP is new (caller responsibility to set it to false). 1121 * 1122 * @return pgt_info structure associated with the allocated HOP on success, otherwise NULL. 1123 */ 1124struct pgt_info *hl_mmu_hr_get_alloc_next_hop(struct hl_ctx *ctx, 1125 struct hl_mmu_hr_priv *hr_priv, 1126 struct hl_hr_mmu_funcs *hr_func, 1127 struct hl_mmu_properties *mmu_prop, 1128 u64 curr_pte, bool *is_new_hop) 1129{ 1130 u64 hop_addr = hl_mmu_get_next_hop_addr(ctx, curr_pte); 1131 1132 if (hop_addr != ULLONG_MAX) 1133 return hr_func->get_pgt_info(ctx, hop_addr); 1134 1135 *is_new_hop = true; 1136 return hl_mmu_hr_alloc_hop(ctx, hr_priv, hr_func, mmu_prop); 1137} 1138 1139/** 1140 * hl_mmu_hr_get_tlb_info() - get the TLB info (info for a specific mapping) 1141 * @ctx: pointer to the context structure. 1142 * @virt_addr: the virt address for which to get info. 1143 * @hops: HOPs info structure. 1144 * @hr_func: host resident functions. 1145 * 1146 * @return 0 on success, otherwise non 0 error code.. 1147 */ 1148int hl_mmu_hr_get_tlb_info(struct hl_ctx *ctx, u64 virt_addr, struct hl_mmu_hop_info *hops, 1149 struct hl_hr_mmu_funcs *hr_func) 1150{ 1151 /* using 6 HOPs as this is the maximum number of HOPs */ 1152 struct pgt_info *hops_pgt_info[MMU_ARCH_6_HOPS] = { NULL }; 1153 struct hl_device *hdev = ctx->hdev; 1154 struct hl_mmu_properties *mmu_prop; 1155 int rc, i, used_hops; 1156 bool is_huge; 1157 1158 rc = hr_func->get_tlb_mapping_params(hdev, &mmu_prop, hops, virt_addr, &is_huge); 1159 if (rc) 1160 return rc; 1161 1162 used_hops = mmu_prop->num_hops; 1163 1164 /* huge pages use one less hop */ 1165 if (is_huge) 1166 used_hops--; 1167 1168 hops->scrambled_vaddr = hdev->asic_funcs->scramble_addr(hdev, virt_addr); 1169 1170 for (i = 0 ; i < used_hops ; i++) { 1171 if (i == 0) 1172 hops_pgt_info[i] = hr_func->get_hop0_pgt_info(ctx); 1173 else 1174 hops_pgt_info[i] = hl_mmu_hr_get_next_hop_pgt_info(ctx, hr_func, 1175 hops->hop_info[i - 1].hop_pte_val); 1176 1177 if (!hops_pgt_info[i]) 1178 return -EFAULT; 1179 1180 hops->hop_info[i].hop_addr = hops_pgt_info[i]->phys_addr; 1181 hops->hop_info[i].hop_pte_addr = 1182 hl_mmu_get_hop_pte_phys_addr(ctx, mmu_prop, i, 1183 hops->hop_info[i].hop_addr, 1184 hops->scrambled_vaddr); 1185 hops->hop_info[i].hop_pte_val = *(u64 *) (uintptr_t) 1186 hl_mmu_hr_pte_phys_to_virt(ctx, hops_pgt_info[i], 1187 hops->hop_info[i].hop_pte_addr, 1188 mmu_prop->hop_table_size); 1189 1190 if (!(hops->hop_info[i].hop_pte_val & PAGE_PRESENT_MASK)) 1191 return -EFAULT; 1192 1193 if (hops->hop_info[i].hop_pte_val & mmu_prop->last_mask) 1194 break; 1195 } 1196 1197 /* if passed over all hops then no last hop was found */ 1198 if (i == mmu_prop->num_hops) 1199 return -EFAULT; 1200 1201 if (hops->scrambled_vaddr != virt_addr) 1202 hops->unscrambled_paddr = hdev->asic_funcs->descramble_addr 1203 (hdev, hops->hop_info[i].hop_pte_val); 1204 else 1205 hops->unscrambled_paddr = hops->hop_info[i].hop_pte_val; 1206 1207 hops->used_hops = i + 1; 1208 1209 return 0; 1210} 1211 1212