1 /* 2 * PPC Huge TLB Page Support for Kernel. 3 * 4 * Copyright (C) 2003 David Gibson, IBM Corporation. 5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor 6 * 7 * Based on the IA-32 version: 8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com> 9 */ 10 11 #include <linux/mm.h> 12 #include <linux/io.h> 13 #include <linux/slab.h> 14 #include <linux/hugetlb.h> 15 #include <linux/export.h> 16 #include <linux/of_fdt.h> 17 #include <linux/memblock.h> 18 #include <linux/bootmem.h> 19 #include <linux/moduleparam.h> 20 #include <asm/pgtable.h> 21 #include <asm/pgalloc.h> 22 #include <asm/tlb.h> 23 #include <asm/setup.h> 24 #include <asm/hugetlb.h> 25 26 #ifdef CONFIG_HUGETLB_PAGE 27 28 #define PAGE_SHIFT_64K 16 29 #define PAGE_SHIFT_16M 24 30 #define PAGE_SHIFT_16G 34 31 32 unsigned int HPAGE_SHIFT; 33 34 /* 35 * Tracks gpages after the device tree is scanned and before the 36 * huge_boot_pages list is ready. On non-Freescale implementations, this is 37 * just used to track 16G pages and so is a single array. FSL-based 38 * implementations may have more than one gpage size, so we need multiple 39 * arrays 40 */ 41 #ifdef CONFIG_PPC_FSL_BOOK3E 42 #define MAX_NUMBER_GPAGES 128 43 struct psize_gpages { 44 u64 gpage_list[MAX_NUMBER_GPAGES]; 45 unsigned int nr_gpages; 46 }; 47 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT]; 48 #else 49 #define MAX_NUMBER_GPAGES 1024 50 static u64 gpage_freearray[MAX_NUMBER_GPAGES]; 51 static unsigned nr_gpages; 52 #endif 53 54 #define hugepd_none(hpd) ((hpd).pd == 0) 55 56 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) 57 { 58 /* Only called for hugetlbfs pages, hence can ignore THP */ 59 return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL); 60 } 61 62 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp, 63 unsigned long address, unsigned pdshift, unsigned pshift) 64 { 65 struct kmem_cache *cachep; 66 pte_t *new; 67 68 #ifdef CONFIG_PPC_FSL_BOOK3E 69 int i; 70 int num_hugepd = 1 << (pshift - pdshift); 71 cachep = hugepte_cache; 72 #else 73 cachep = PGT_CACHE(pdshift - pshift); 74 #endif 75 76 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT); 77 78 BUG_ON(pshift > HUGEPD_SHIFT_MASK); 79 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK); 80 81 if (! new) 82 return -ENOMEM; 83 84 spin_lock(&mm->page_table_lock); 85 #ifdef CONFIG_PPC_FSL_BOOK3E 86 /* 87 * We have multiple higher-level entries that point to the same 88 * actual pte location. Fill in each as we go and backtrack on error. 89 * We need all of these so the DTLB pgtable walk code can find the 90 * right higher-level entry without knowing if it's a hugepage or not. 91 */ 92 for (i = 0; i < num_hugepd; i++, hpdp++) { 93 if (unlikely(!hugepd_none(*hpdp))) 94 break; 95 else 96 /* We use the old format for PPC_FSL_BOOK3E */ 97 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; 98 } 99 /* If we bailed from the for loop early, an error occurred, clean up */ 100 if (i < num_hugepd) { 101 for (i = i - 1 ; i >= 0; i--, hpdp--) 102 hpdp->pd = 0; 103 kmem_cache_free(cachep, new); 104 } 105 #else 106 if (!hugepd_none(*hpdp)) 107 kmem_cache_free(cachep, new); 108 else { 109 #ifdef CONFIG_PPC_BOOK3S_64 110 hpdp->pd = (unsigned long)new | 111 (shift_to_mmu_psize(pshift) << 2); 112 #else 113 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift; 114 #endif 115 } 116 #endif 117 spin_unlock(&mm->page_table_lock); 118 return 0; 119 } 120 121 /* 122 * These macros define how to determine which level of the page table holds 123 * the hpdp. 124 */ 125 #ifdef CONFIG_PPC_FSL_BOOK3E 126 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT 127 #define HUGEPD_PUD_SHIFT PUD_SHIFT 128 #else 129 #define HUGEPD_PGD_SHIFT PUD_SHIFT 130 #define HUGEPD_PUD_SHIFT PMD_SHIFT 131 #endif 132 133 #ifdef CONFIG_PPC_BOOK3S_64 134 /* 135 * At this point we do the placement change only for BOOK3S 64. This would 136 * possibly work on other subarchs. 137 */ 138 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) 139 { 140 pgd_t *pg; 141 pud_t *pu; 142 pmd_t *pm; 143 hugepd_t *hpdp = NULL; 144 unsigned pshift = __ffs(sz); 145 unsigned pdshift = PGDIR_SHIFT; 146 147 addr &= ~(sz-1); 148 pg = pgd_offset(mm, addr); 149 150 if (pshift == PGDIR_SHIFT) 151 /* 16GB huge page */ 152 return (pte_t *) pg; 153 else if (pshift > PUD_SHIFT) 154 /* 155 * We need to use hugepd table 156 */ 157 hpdp = (hugepd_t *)pg; 158 else { 159 pdshift = PUD_SHIFT; 160 pu = pud_alloc(mm, pg, addr); 161 if (pshift == PUD_SHIFT) 162 return (pte_t *)pu; 163 else if (pshift > PMD_SHIFT) 164 hpdp = (hugepd_t *)pu; 165 else { 166 pdshift = PMD_SHIFT; 167 pm = pmd_alloc(mm, pu, addr); 168 if (pshift == PMD_SHIFT) 169 /* 16MB hugepage */ 170 return (pte_t *)pm; 171 else 172 hpdp = (hugepd_t *)pm; 173 } 174 } 175 if (!hpdp) 176 return NULL; 177 178 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); 179 180 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) 181 return NULL; 182 183 return hugepte_offset(*hpdp, addr, pdshift); 184 } 185 186 #else 187 188 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz) 189 { 190 pgd_t *pg; 191 pud_t *pu; 192 pmd_t *pm; 193 hugepd_t *hpdp = NULL; 194 unsigned pshift = __ffs(sz); 195 unsigned pdshift = PGDIR_SHIFT; 196 197 addr &= ~(sz-1); 198 199 pg = pgd_offset(mm, addr); 200 201 if (pshift >= HUGEPD_PGD_SHIFT) { 202 hpdp = (hugepd_t *)pg; 203 } else { 204 pdshift = PUD_SHIFT; 205 pu = pud_alloc(mm, pg, addr); 206 if (pshift >= HUGEPD_PUD_SHIFT) { 207 hpdp = (hugepd_t *)pu; 208 } else { 209 pdshift = PMD_SHIFT; 210 pm = pmd_alloc(mm, pu, addr); 211 hpdp = (hugepd_t *)pm; 212 } 213 } 214 215 if (!hpdp) 216 return NULL; 217 218 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp)); 219 220 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift)) 221 return NULL; 222 223 return hugepte_offset(*hpdp, addr, pdshift); 224 } 225 #endif 226 227 #ifdef CONFIG_PPC_FSL_BOOK3E 228 /* Build list of addresses of gigantic pages. This function is used in early 229 * boot before the buddy allocator is setup. 230 */ 231 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) 232 { 233 unsigned int idx = shift_to_mmu_psize(__ffs(page_size)); 234 int i; 235 236 if (addr == 0) 237 return; 238 239 gpage_freearray[idx].nr_gpages = number_of_pages; 240 241 for (i = 0; i < number_of_pages; i++) { 242 gpage_freearray[idx].gpage_list[i] = addr; 243 addr += page_size; 244 } 245 } 246 247 /* 248 * Moves the gigantic page addresses from the temporary list to the 249 * huge_boot_pages list. 250 */ 251 int alloc_bootmem_huge_page(struct hstate *hstate) 252 { 253 struct huge_bootmem_page *m; 254 int idx = shift_to_mmu_psize(huge_page_shift(hstate)); 255 int nr_gpages = gpage_freearray[idx].nr_gpages; 256 257 if (nr_gpages == 0) 258 return 0; 259 260 #ifdef CONFIG_HIGHMEM 261 /* 262 * If gpages can be in highmem we can't use the trick of storing the 263 * data structure in the page; allocate space for this 264 */ 265 m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0); 266 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages]; 267 #else 268 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]); 269 #endif 270 271 list_add(&m->list, &huge_boot_pages); 272 gpage_freearray[idx].nr_gpages = nr_gpages; 273 gpage_freearray[idx].gpage_list[nr_gpages] = 0; 274 m->hstate = hstate; 275 276 return 1; 277 } 278 /* 279 * Scan the command line hugepagesz= options for gigantic pages; store those in 280 * a list that we use to allocate the memory once all options are parsed. 281 */ 282 283 unsigned long gpage_npages[MMU_PAGE_COUNT]; 284 285 static int __init do_gpage_early_setup(char *param, char *val, 286 const char *unused, void *arg) 287 { 288 static phys_addr_t size; 289 unsigned long npages; 290 291 /* 292 * The hugepagesz and hugepages cmdline options are interleaved. We 293 * use the size variable to keep track of whether or not this was done 294 * properly and skip over instances where it is incorrect. Other 295 * command-line parsing code will issue warnings, so we don't need to. 296 * 297 */ 298 if ((strcmp(param, "default_hugepagesz") == 0) || 299 (strcmp(param, "hugepagesz") == 0)) { 300 size = memparse(val, NULL); 301 } else if (strcmp(param, "hugepages") == 0) { 302 if (size != 0) { 303 if (sscanf(val, "%lu", &npages) <= 0) 304 npages = 0; 305 if (npages > MAX_NUMBER_GPAGES) { 306 pr_warn("MMU: %lu pages requested for page " 307 "size %llu KB, limiting to " 308 __stringify(MAX_NUMBER_GPAGES) "\n", 309 npages, size / 1024); 310 npages = MAX_NUMBER_GPAGES; 311 } 312 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages; 313 size = 0; 314 } 315 } 316 return 0; 317 } 318 319 320 /* 321 * This function allocates physical space for pages that are larger than the 322 * buddy allocator can handle. We want to allocate these in highmem because 323 * the amount of lowmem is limited. This means that this function MUST be 324 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb 325 * allocate to grab highmem. 326 */ 327 void __init reserve_hugetlb_gpages(void) 328 { 329 static __initdata char cmdline[COMMAND_LINE_SIZE]; 330 phys_addr_t size, base; 331 int i; 332 333 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE); 334 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0, 335 NULL, &do_gpage_early_setup); 336 337 /* 338 * Walk gpage list in reverse, allocating larger page sizes first. 339 * Skip over unsupported sizes, or sizes that have 0 gpages allocated. 340 * When we reach the point in the list where pages are no longer 341 * considered gpages, we're done. 342 */ 343 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) { 344 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0) 345 continue; 346 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT)) 347 break; 348 349 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i)); 350 base = memblock_alloc_base(size * gpage_npages[i], size, 351 MEMBLOCK_ALLOC_ANYWHERE); 352 add_gpage(base, size, gpage_npages[i]); 353 } 354 } 355 356 #else /* !PPC_FSL_BOOK3E */ 357 358 /* Build list of addresses of gigantic pages. This function is used in early 359 * boot before the buddy allocator is setup. 360 */ 361 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages) 362 { 363 if (!addr) 364 return; 365 while (number_of_pages > 0) { 366 gpage_freearray[nr_gpages] = addr; 367 nr_gpages++; 368 number_of_pages--; 369 addr += page_size; 370 } 371 } 372 373 /* Moves the gigantic page addresses from the temporary list to the 374 * huge_boot_pages list. 375 */ 376 int alloc_bootmem_huge_page(struct hstate *hstate) 377 { 378 struct huge_bootmem_page *m; 379 if (nr_gpages == 0) 380 return 0; 381 m = phys_to_virt(gpage_freearray[--nr_gpages]); 382 gpage_freearray[nr_gpages] = 0; 383 list_add(&m->list, &huge_boot_pages); 384 m->hstate = hstate; 385 return 1; 386 } 387 #endif 388 389 #ifdef CONFIG_PPC_FSL_BOOK3E 390 #define HUGEPD_FREELIST_SIZE \ 391 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t)) 392 393 struct hugepd_freelist { 394 struct rcu_head rcu; 395 unsigned int index; 396 void *ptes[0]; 397 }; 398 399 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur); 400 401 static void hugepd_free_rcu_callback(struct rcu_head *head) 402 { 403 struct hugepd_freelist *batch = 404 container_of(head, struct hugepd_freelist, rcu); 405 unsigned int i; 406 407 for (i = 0; i < batch->index; i++) 408 kmem_cache_free(hugepte_cache, batch->ptes[i]); 409 410 free_page((unsigned long)batch); 411 } 412 413 static void hugepd_free(struct mmu_gather *tlb, void *hugepte) 414 { 415 struct hugepd_freelist **batchp; 416 417 batchp = this_cpu_ptr(&hugepd_freelist_cur); 418 419 if (atomic_read(&tlb->mm->mm_users) < 2 || 420 cpumask_equal(mm_cpumask(tlb->mm), 421 cpumask_of(smp_processor_id()))) { 422 kmem_cache_free(hugepte_cache, hugepte); 423 put_cpu_var(hugepd_freelist_cur); 424 return; 425 } 426 427 if (*batchp == NULL) { 428 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC); 429 (*batchp)->index = 0; 430 } 431 432 (*batchp)->ptes[(*batchp)->index++] = hugepte; 433 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) { 434 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback); 435 *batchp = NULL; 436 } 437 put_cpu_var(hugepd_freelist_cur); 438 } 439 #endif 440 441 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift, 442 unsigned long start, unsigned long end, 443 unsigned long floor, unsigned long ceiling) 444 { 445 pte_t *hugepte = hugepd_page(*hpdp); 446 int i; 447 448 unsigned long pdmask = ~((1UL << pdshift) - 1); 449 unsigned int num_hugepd = 1; 450 451 #ifdef CONFIG_PPC_FSL_BOOK3E 452 /* Note: On fsl the hpdp may be the first of several */ 453 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift)); 454 #else 455 unsigned int shift = hugepd_shift(*hpdp); 456 #endif 457 458 start &= pdmask; 459 if (start < floor) 460 return; 461 if (ceiling) { 462 ceiling &= pdmask; 463 if (! ceiling) 464 return; 465 } 466 if (end - 1 > ceiling - 1) 467 return; 468 469 for (i = 0; i < num_hugepd; i++, hpdp++) 470 hpdp->pd = 0; 471 472 #ifdef CONFIG_PPC_FSL_BOOK3E 473 hugepd_free(tlb, hugepte); 474 #else 475 pgtable_free_tlb(tlb, hugepte, pdshift - shift); 476 #endif 477 } 478 479 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 480 unsigned long addr, unsigned long end, 481 unsigned long floor, unsigned long ceiling) 482 { 483 pmd_t *pmd; 484 unsigned long next; 485 unsigned long start; 486 487 start = addr; 488 do { 489 pmd = pmd_offset(pud, addr); 490 next = pmd_addr_end(addr, end); 491 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) { 492 /* 493 * if it is not hugepd pointer, we should already find 494 * it cleared. 495 */ 496 WARN_ON(!pmd_none_or_clear_bad(pmd)); 497 continue; 498 } 499 #ifdef CONFIG_PPC_FSL_BOOK3E 500 /* 501 * Increment next by the size of the huge mapping since 502 * there may be more than one entry at this level for a 503 * single hugepage, but all of them point to 504 * the same kmem cache that holds the hugepte. 505 */ 506 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd)); 507 #endif 508 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT, 509 addr, next, floor, ceiling); 510 } while (addr = next, addr != end); 511 512 start &= PUD_MASK; 513 if (start < floor) 514 return; 515 if (ceiling) { 516 ceiling &= PUD_MASK; 517 if (!ceiling) 518 return; 519 } 520 if (end - 1 > ceiling - 1) 521 return; 522 523 pmd = pmd_offset(pud, start); 524 pud_clear(pud); 525 pmd_free_tlb(tlb, pmd, start); 526 mm_dec_nr_pmds(tlb->mm); 527 } 528 529 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 530 unsigned long addr, unsigned long end, 531 unsigned long floor, unsigned long ceiling) 532 { 533 pud_t *pud; 534 unsigned long next; 535 unsigned long start; 536 537 start = addr; 538 do { 539 pud = pud_offset(pgd, addr); 540 next = pud_addr_end(addr, end); 541 if (!is_hugepd(__hugepd(pud_val(*pud)))) { 542 if (pud_none_or_clear_bad(pud)) 543 continue; 544 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, 545 ceiling); 546 } else { 547 #ifdef CONFIG_PPC_FSL_BOOK3E 548 /* 549 * Increment next by the size of the huge mapping since 550 * there may be more than one entry at this level for a 551 * single hugepage, but all of them point to 552 * the same kmem cache that holds the hugepte. 553 */ 554 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud)); 555 #endif 556 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT, 557 addr, next, floor, ceiling); 558 } 559 } while (addr = next, addr != end); 560 561 start &= PGDIR_MASK; 562 if (start < floor) 563 return; 564 if (ceiling) { 565 ceiling &= PGDIR_MASK; 566 if (!ceiling) 567 return; 568 } 569 if (end - 1 > ceiling - 1) 570 return; 571 572 pud = pud_offset(pgd, start); 573 pgd_clear(pgd); 574 pud_free_tlb(tlb, pud, start); 575 } 576 577 /* 578 * This function frees user-level page tables of a process. 579 */ 580 void hugetlb_free_pgd_range(struct mmu_gather *tlb, 581 unsigned long addr, unsigned long end, 582 unsigned long floor, unsigned long ceiling) 583 { 584 pgd_t *pgd; 585 unsigned long next; 586 587 /* 588 * Because there are a number of different possible pagetable 589 * layouts for hugepage ranges, we limit knowledge of how 590 * things should be laid out to the allocation path 591 * (huge_pte_alloc(), above). Everything else works out the 592 * structure as it goes from information in the hugepd 593 * pointers. That means that we can't here use the 594 * optimization used in the normal page free_pgd_range(), of 595 * checking whether we're actually covering a large enough 596 * range to have to do anything at the top level of the walk 597 * instead of at the bottom. 598 * 599 * To make sense of this, you should probably go read the big 600 * block comment at the top of the normal free_pgd_range(), 601 * too. 602 */ 603 604 do { 605 next = pgd_addr_end(addr, end); 606 pgd = pgd_offset(tlb->mm, addr); 607 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) { 608 if (pgd_none_or_clear_bad(pgd)) 609 continue; 610 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); 611 } else { 612 #ifdef CONFIG_PPC_FSL_BOOK3E 613 /* 614 * Increment next by the size of the huge mapping since 615 * there may be more than one entry at the pgd level 616 * for a single hugepage, but all of them point to the 617 * same kmem cache that holds the hugepte. 618 */ 619 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd)); 620 #endif 621 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT, 622 addr, next, floor, ceiling); 623 } 624 } while (addr = next, addr != end); 625 } 626 627 /* 628 * We are holding mmap_sem, so a parallel huge page collapse cannot run. 629 * To prevent hugepage split, disable irq. 630 */ 631 struct page * 632 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) 633 { 634 bool is_thp; 635 pte_t *ptep, pte; 636 unsigned shift; 637 unsigned long mask, flags; 638 struct page *page = ERR_PTR(-EINVAL); 639 640 local_irq_save(flags); 641 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift); 642 if (!ptep) 643 goto no_page; 644 pte = READ_ONCE(*ptep); 645 /* 646 * Verify it is a huge page else bail. 647 * Transparent hugepages are handled by generic code. We can skip them 648 * here. 649 */ 650 if (!shift || is_thp) 651 goto no_page; 652 653 if (!pte_present(pte)) { 654 page = NULL; 655 goto no_page; 656 } 657 mask = (1UL << shift) - 1; 658 page = pte_page(pte); 659 if (page) 660 page += (address & mask) / PAGE_SIZE; 661 662 no_page: 663 local_irq_restore(flags); 664 return page; 665 } 666 667 struct page * 668 follow_huge_pmd(struct mm_struct *mm, unsigned long address, 669 pmd_t *pmd, int write) 670 { 671 BUG(); 672 return NULL; 673 } 674 675 struct page * 676 follow_huge_pud(struct mm_struct *mm, unsigned long address, 677 pud_t *pud, int write) 678 { 679 BUG(); 680 return NULL; 681 } 682 683 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end, 684 unsigned long sz) 685 { 686 unsigned long __boundary = (addr + sz) & ~(sz-1); 687 return (__boundary - 1 < end - 1) ? __boundary : end; 688 } 689 690 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift, 691 unsigned long end, int write, struct page **pages, int *nr) 692 { 693 pte_t *ptep; 694 unsigned long sz = 1UL << hugepd_shift(hugepd); 695 unsigned long next; 696 697 ptep = hugepte_offset(hugepd, addr, pdshift); 698 do { 699 next = hugepte_addr_end(addr, end, sz); 700 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr)) 701 return 0; 702 } while (ptep++, addr = next, addr != end); 703 704 return 1; 705 } 706 707 #ifdef CONFIG_PPC_MM_SLICES 708 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 709 unsigned long len, unsigned long pgoff, 710 unsigned long flags) 711 { 712 struct hstate *hstate = hstate_file(file); 713 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate)); 714 715 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1); 716 } 717 #endif 718 719 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) 720 { 721 #ifdef CONFIG_PPC_MM_SLICES 722 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start); 723 724 return 1UL << mmu_psize_to_shift(psize); 725 #else 726 if (!is_vm_hugetlb_page(vma)) 727 return PAGE_SIZE; 728 729 return huge_page_size(hstate_vma(vma)); 730 #endif 731 } 732 733 static inline bool is_power_of_4(unsigned long x) 734 { 735 if (is_power_of_2(x)) 736 return (__ilog2(x) % 2) ? false : true; 737 return false; 738 } 739 740 static int __init add_huge_page_size(unsigned long long size) 741 { 742 int shift = __ffs(size); 743 int mmu_psize; 744 745 /* Check that it is a page size supported by the hardware and 746 * that it fits within pagetable and slice limits. */ 747 #ifdef CONFIG_PPC_FSL_BOOK3E 748 if ((size < PAGE_SIZE) || !is_power_of_4(size)) 749 return -EINVAL; 750 #else 751 if (!is_power_of_2(size) 752 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT)) 753 return -EINVAL; 754 #endif 755 756 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0) 757 return -EINVAL; 758 759 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift); 760 761 /* Return if huge page size has already been setup */ 762 if (size_to_hstate(size)) 763 return 0; 764 765 hugetlb_add_hstate(shift - PAGE_SHIFT); 766 767 return 0; 768 } 769 770 static int __init hugepage_setup_sz(char *str) 771 { 772 unsigned long long size; 773 774 size = memparse(str, &str); 775 776 if (add_huge_page_size(size) != 0) 777 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size); 778 779 return 1; 780 } 781 __setup("hugepagesz=", hugepage_setup_sz); 782 783 #ifdef CONFIG_PPC_FSL_BOOK3E 784 struct kmem_cache *hugepte_cache; 785 static int __init hugetlbpage_init(void) 786 { 787 int psize; 788 789 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { 790 unsigned shift; 791 792 if (!mmu_psize_defs[psize].shift) 793 continue; 794 795 shift = mmu_psize_to_shift(psize); 796 797 /* Don't treat normal page sizes as huge... */ 798 if (shift != PAGE_SHIFT) 799 if (add_huge_page_size(1ULL << shift) < 0) 800 continue; 801 } 802 803 /* 804 * Create a kmem cache for hugeptes. The bottom bits in the pte have 805 * size information encoded in them, so align them to allow this 806 */ 807 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t), 808 HUGEPD_SHIFT_MASK + 1, 0, NULL); 809 if (hugepte_cache == NULL) 810 panic("%s: Unable to create kmem cache for hugeptes\n", 811 __func__); 812 813 /* Default hpage size = 4M */ 814 if (mmu_psize_defs[MMU_PAGE_4M].shift) 815 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift; 816 else 817 panic("%s: Unable to set default huge page size\n", __func__); 818 819 820 return 0; 821 } 822 #else 823 static int __init hugetlbpage_init(void) 824 { 825 int psize; 826 827 if (!mmu_has_feature(MMU_FTR_16M_PAGE)) 828 return -ENODEV; 829 830 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) { 831 unsigned shift; 832 unsigned pdshift; 833 834 if (!mmu_psize_defs[psize].shift) 835 continue; 836 837 shift = mmu_psize_to_shift(psize); 838 839 if (add_huge_page_size(1ULL << shift) < 0) 840 continue; 841 842 if (shift < PMD_SHIFT) 843 pdshift = PMD_SHIFT; 844 else if (shift < PUD_SHIFT) 845 pdshift = PUD_SHIFT; 846 else 847 pdshift = PGDIR_SHIFT; 848 /* 849 * if we have pdshift and shift value same, we don't 850 * use pgt cache for hugepd. 851 */ 852 if (pdshift != shift) { 853 pgtable_cache_add(pdshift - shift, NULL); 854 if (!PGT_CACHE(pdshift - shift)) 855 panic("hugetlbpage_init(): could not create " 856 "pgtable cache for %d bit pagesize\n", shift); 857 } 858 } 859 860 /* Set default large page size. Currently, we pick 16M or 1M 861 * depending on what is available 862 */ 863 if (mmu_psize_defs[MMU_PAGE_16M].shift) 864 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift; 865 else if (mmu_psize_defs[MMU_PAGE_1M].shift) 866 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift; 867 868 return 0; 869 } 870 #endif 871 arch_initcall(hugetlbpage_init); 872 873 void flush_dcache_icache_hugepage(struct page *page) 874 { 875 int i; 876 void *start; 877 878 BUG_ON(!PageCompound(page)); 879 880 for (i = 0; i < (1UL << compound_order(page)); i++) { 881 if (!PageHighMem(page)) { 882 __flush_dcache_icache(page_address(page+i)); 883 } else { 884 start = kmap_atomic(page+i); 885 __flush_dcache_icache(start); 886 kunmap_atomic(start); 887 } 888 } 889 } 890 891 #endif /* CONFIG_HUGETLB_PAGE */ 892 893 /* 894 * We have 4 cases for pgds and pmds: 895 * (1) invalid (all zeroes) 896 * (2) pointer to next table, as normal; bottom 6 bits == 0 897 * (3) leaf pte for huge page _PAGE_PTE set 898 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table 899 * 900 * So long as we atomically load page table pointers we are safe against teardown, 901 * we can follow the address down to the the page and take a ref on it. 902 * This function need to be called with interrupts disabled. We use this variant 903 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1 904 */ 905 906 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, 907 bool *is_thp, unsigned *shift) 908 { 909 pgd_t pgd, *pgdp; 910 pud_t pud, *pudp; 911 pmd_t pmd, *pmdp; 912 pte_t *ret_pte; 913 hugepd_t *hpdp = NULL; 914 unsigned pdshift = PGDIR_SHIFT; 915 916 if (shift) 917 *shift = 0; 918 919 if (is_thp) 920 *is_thp = false; 921 922 pgdp = pgdir + pgd_index(ea); 923 pgd = READ_ONCE(*pgdp); 924 /* 925 * Always operate on the local stack value. This make sure the 926 * value don't get updated by a parallel THP split/collapse, 927 * page fault or a page unmap. The return pte_t * is still not 928 * stable. So should be checked there for above conditions. 929 */ 930 if (pgd_none(pgd)) 931 return NULL; 932 else if (pgd_huge(pgd)) { 933 ret_pte = (pte_t *) pgdp; 934 goto out; 935 } else if (is_hugepd(__hugepd(pgd_val(pgd)))) 936 hpdp = (hugepd_t *)&pgd; 937 else { 938 /* 939 * Even if we end up with an unmap, the pgtable will not 940 * be freed, because we do an rcu free and here we are 941 * irq disabled 942 */ 943 pdshift = PUD_SHIFT; 944 pudp = pud_offset(&pgd, ea); 945 pud = READ_ONCE(*pudp); 946 947 if (pud_none(pud)) 948 return NULL; 949 else if (pud_huge(pud)) { 950 ret_pte = (pte_t *) pudp; 951 goto out; 952 } else if (is_hugepd(__hugepd(pud_val(pud)))) 953 hpdp = (hugepd_t *)&pud; 954 else { 955 pdshift = PMD_SHIFT; 956 pmdp = pmd_offset(&pud, ea); 957 pmd = READ_ONCE(*pmdp); 958 /* 959 * A hugepage collapse is captured by pmd_none, because 960 * it mark the pmd none and do a hpte invalidate. 961 * 962 * We don't worry about pmd_trans_splitting here, The 963 * caller if it needs to handle the splitting case 964 * should check for that. 965 */ 966 if (pmd_none(pmd)) 967 return NULL; 968 969 if (pmd_trans_huge(pmd)) { 970 if (is_thp) 971 *is_thp = true; 972 ret_pte = (pte_t *) pmdp; 973 goto out; 974 } 975 976 if (pmd_huge(pmd)) { 977 ret_pte = (pte_t *) pmdp; 978 goto out; 979 } else if (is_hugepd(__hugepd(pmd_val(pmd)))) 980 hpdp = (hugepd_t *)&pmd; 981 else 982 return pte_offset_kernel(&pmd, ea); 983 } 984 } 985 if (!hpdp) 986 return NULL; 987 988 ret_pte = hugepte_offset(*hpdp, ea, pdshift); 989 pdshift = hugepd_shift(*hpdp); 990 out: 991 if (shift) 992 *shift = pdshift; 993 return ret_pte; 994 } 995 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte); 996 997 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr, 998 unsigned long end, int write, struct page **pages, int *nr) 999 { 1000 unsigned long mask; 1001 unsigned long pte_end; 1002 struct page *head, *page, *tail; 1003 pte_t pte; 1004 int refs; 1005 1006 pte_end = (addr + sz) & ~(sz-1); 1007 if (pte_end < end) 1008 end = pte_end; 1009 1010 pte = READ_ONCE(*ptep); 1011 mask = _PAGE_PRESENT | _PAGE_USER; 1012 if (write) 1013 mask |= _PAGE_RW; 1014 1015 if ((pte_val(pte) & mask) != mask) 1016 return 0; 1017 1018 /* hugepages are never "special" */ 1019 VM_BUG_ON(!pfn_valid(pte_pfn(pte))); 1020 1021 refs = 0; 1022 head = pte_page(pte); 1023 1024 page = head + ((addr & (sz-1)) >> PAGE_SHIFT); 1025 tail = page; 1026 do { 1027 VM_BUG_ON(compound_head(page) != head); 1028 pages[*nr] = page; 1029 (*nr)++; 1030 page++; 1031 refs++; 1032 } while (addr += PAGE_SIZE, addr != end); 1033 1034 if (!page_cache_add_speculative(head, refs)) { 1035 *nr -= refs; 1036 return 0; 1037 } 1038 1039 if (unlikely(pte_val(pte) != pte_val(*ptep))) { 1040 /* Could be optimized better */ 1041 *nr -= refs; 1042 while (refs--) 1043 put_page(head); 1044 return 0; 1045 } 1046 1047 /* 1048 * Any tail page need their mapcount reference taken before we 1049 * return. 1050 */ 1051 while (refs--) { 1052 if (PageTail(tail)) 1053 get_huge_page_tail(tail); 1054 tail++; 1055 } 1056 1057 return 1; 1058 } 1059