2 * PPC Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
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>
23 #include <asm/setup.h>
24 #include <asm/hugetlb.h>
26 #ifdef CONFIG_HUGETLB_PAGE
28 #define PAGE_SHIFT_64K 16
29 #define PAGE_SHIFT_16M 24
30 #define PAGE_SHIFT_16G 34
32 unsigned int HPAGE_SHIFT;
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
41 #ifdef CONFIG_PPC_FSL_BOOK3E
42 #define MAX_NUMBER_GPAGES 128
44 u64 gpage_list[MAX_NUMBER_GPAGES];
45 unsigned int nr_gpages;
47 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
49 #define MAX_NUMBER_GPAGES 1024
50 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51 static unsigned nr_gpages;
54 #define hugepd_none(hpd) ((hpd).pd == 0)
56 #ifdef CONFIG_PPC_BOOK3S_64
58 * At this point we do the placement change only for BOOK3S 64. This would
59 * possibly work on other subarchs.
63 * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have
64 * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD;
66 * Defined in such a way that we can optimize away code block at build time
67 * if CONFIG_HUGETLB_PAGE=n.
69 int pmd_huge(pmd_t pmd)
72 * leaf pte for huge page, bottom two bits != 00
74 return ((pmd_val(pmd) & 0x3) != 0x0);
77 int pud_huge(pud_t pud)
80 * leaf pte for huge page, bottom two bits != 00
82 return ((pud_val(pud) & 0x3) != 0x0);
85 int pgd_huge(pgd_t pgd)
88 * leaf pte for huge page, bottom two bits != 00
90 return ((pgd_val(pgd) & 0x3) != 0x0);
93 #if defined(CONFIG_PPC_64K_PAGES) && defined(CONFIG_DEBUG_VM)
95 * This enables us to catch the wrong page directory format
96 * Moved here so that we can use WARN() in the call.
98 int hugepd_ok(hugepd_t hpd)
103 * We should not find this format in page directory, warn otherwise.
105 is_hugepd = (((hpd.pd & 0x3) == 0x0) && ((hpd.pd & HUGEPD_SHIFT_MASK) != 0));
106 WARN(is_hugepd, "Found wrong page directory format\n");
112 int pmd_huge(pmd_t pmd)
117 int pud_huge(pud_t pud)
122 int pgd_huge(pgd_t pgd)
128 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
130 /* Only called for hugetlbfs pages, hence can ignore THP */
131 return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
134 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
135 unsigned long address, unsigned pdshift, unsigned pshift)
137 struct kmem_cache *cachep;
140 #ifdef CONFIG_PPC_FSL_BOOK3E
142 int num_hugepd = 1 << (pshift - pdshift);
143 cachep = hugepte_cache;
145 cachep = PGT_CACHE(pdshift - pshift);
148 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
150 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
151 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
156 spin_lock(&mm->page_table_lock);
157 #ifdef CONFIG_PPC_FSL_BOOK3E
159 * We have multiple higher-level entries that point to the same
160 * actual pte location. Fill in each as we go and backtrack on error.
161 * We need all of these so the DTLB pgtable walk code can find the
162 * right higher-level entry without knowing if it's a hugepage or not.
164 for (i = 0; i < num_hugepd; i++, hpdp++) {
165 if (unlikely(!hugepd_none(*hpdp)))
168 /* We use the old format for PPC_FSL_BOOK3E */
169 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
171 /* If we bailed from the for loop early, an error occurred, clean up */
172 if (i < num_hugepd) {
173 for (i = i - 1 ; i >= 0; i--, hpdp--)
175 kmem_cache_free(cachep, new);
178 if (!hugepd_none(*hpdp))
179 kmem_cache_free(cachep, new);
181 #ifdef CONFIG_PPC_BOOK3S_64
182 hpdp->pd = (unsigned long)new |
183 (shift_to_mmu_psize(pshift) << 2);
185 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
189 spin_unlock(&mm->page_table_lock);
194 * These macros define how to determine which level of the page table holds
197 #ifdef CONFIG_PPC_FSL_BOOK3E
198 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
199 #define HUGEPD_PUD_SHIFT PUD_SHIFT
201 #define HUGEPD_PGD_SHIFT PUD_SHIFT
202 #define HUGEPD_PUD_SHIFT PMD_SHIFT
205 #ifdef CONFIG_PPC_BOOK3S_64
207 * At this point we do the placement change only for BOOK3S 64. This would
208 * possibly work on other subarchs.
210 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
215 hugepd_t *hpdp = NULL;
216 unsigned pshift = __ffs(sz);
217 unsigned pdshift = PGDIR_SHIFT;
220 pg = pgd_offset(mm, addr);
222 if (pshift == PGDIR_SHIFT)
225 else if (pshift > PUD_SHIFT)
227 * We need to use hugepd table
229 hpdp = (hugepd_t *)pg;
232 pu = pud_alloc(mm, pg, addr);
233 if (pshift == PUD_SHIFT)
235 else if (pshift > PMD_SHIFT)
236 hpdp = (hugepd_t *)pu;
239 pm = pmd_alloc(mm, pu, addr);
240 if (pshift == PMD_SHIFT)
244 hpdp = (hugepd_t *)pm;
250 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
252 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
255 return hugepte_offset(*hpdp, addr, pdshift);
260 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
265 hugepd_t *hpdp = NULL;
266 unsigned pshift = __ffs(sz);
267 unsigned pdshift = PGDIR_SHIFT;
271 pg = pgd_offset(mm, addr);
273 if (pshift >= HUGEPD_PGD_SHIFT) {
274 hpdp = (hugepd_t *)pg;
277 pu = pud_alloc(mm, pg, addr);
278 if (pshift >= HUGEPD_PUD_SHIFT) {
279 hpdp = (hugepd_t *)pu;
282 pm = pmd_alloc(mm, pu, addr);
283 hpdp = (hugepd_t *)pm;
290 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
292 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
295 return hugepte_offset(*hpdp, addr, pdshift);
299 #ifdef CONFIG_PPC_FSL_BOOK3E
300 /* Build list of addresses of gigantic pages. This function is used in early
301 * boot before the buddy allocator is setup.
303 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
305 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
311 gpage_freearray[idx].nr_gpages = number_of_pages;
313 for (i = 0; i < number_of_pages; i++) {
314 gpage_freearray[idx].gpage_list[i] = addr;
320 * Moves the gigantic page addresses from the temporary list to the
321 * huge_boot_pages list.
323 int alloc_bootmem_huge_page(struct hstate *hstate)
325 struct huge_bootmem_page *m;
326 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
327 int nr_gpages = gpage_freearray[idx].nr_gpages;
332 #ifdef CONFIG_HIGHMEM
334 * If gpages can be in highmem we can't use the trick of storing the
335 * data structure in the page; allocate space for this
337 m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
338 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
340 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
343 list_add(&m->list, &huge_boot_pages);
344 gpage_freearray[idx].nr_gpages = nr_gpages;
345 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
351 * Scan the command line hugepagesz= options for gigantic pages; store those in
352 * a list that we use to allocate the memory once all options are parsed.
355 unsigned long gpage_npages[MMU_PAGE_COUNT];
357 static int __init do_gpage_early_setup(char *param, char *val,
358 const char *unused, void *arg)
360 static phys_addr_t size;
361 unsigned long npages;
364 * The hugepagesz and hugepages cmdline options are interleaved. We
365 * use the size variable to keep track of whether or not this was done
366 * properly and skip over instances where it is incorrect. Other
367 * command-line parsing code will issue warnings, so we don't need to.
370 if ((strcmp(param, "default_hugepagesz") == 0) ||
371 (strcmp(param, "hugepagesz") == 0)) {
372 size = memparse(val, NULL);
373 } else if (strcmp(param, "hugepages") == 0) {
375 if (sscanf(val, "%lu", &npages) <= 0)
377 if (npages > MAX_NUMBER_GPAGES) {
378 pr_warn("MMU: %lu pages requested for page "
379 "size %llu KB, limiting to "
380 __stringify(MAX_NUMBER_GPAGES) "\n",
381 npages, size / 1024);
382 npages = MAX_NUMBER_GPAGES;
384 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
393 * This function allocates physical space for pages that are larger than the
394 * buddy allocator can handle. We want to allocate these in highmem because
395 * the amount of lowmem is limited. This means that this function MUST be
396 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
397 * allocate to grab highmem.
399 void __init reserve_hugetlb_gpages(void)
401 static __initdata char cmdline[COMMAND_LINE_SIZE];
402 phys_addr_t size, base;
405 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
406 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
407 NULL, &do_gpage_early_setup);
410 * Walk gpage list in reverse, allocating larger page sizes first.
411 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
412 * When we reach the point in the list where pages are no longer
413 * considered gpages, we're done.
415 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
416 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
418 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
421 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
422 base = memblock_alloc_base(size * gpage_npages[i], size,
423 MEMBLOCK_ALLOC_ANYWHERE);
424 add_gpage(base, size, gpage_npages[i]);
428 #else /* !PPC_FSL_BOOK3E */
430 /* Build list of addresses of gigantic pages. This function is used in early
431 * boot before the buddy allocator is setup.
433 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
437 while (number_of_pages > 0) {
438 gpage_freearray[nr_gpages] = addr;
445 /* Moves the gigantic page addresses from the temporary list to the
446 * huge_boot_pages list.
448 int alloc_bootmem_huge_page(struct hstate *hstate)
450 struct huge_bootmem_page *m;
453 m = phys_to_virt(gpage_freearray[--nr_gpages]);
454 gpage_freearray[nr_gpages] = 0;
455 list_add(&m->list, &huge_boot_pages);
461 #ifdef CONFIG_PPC_FSL_BOOK3E
462 #define HUGEPD_FREELIST_SIZE \
463 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
465 struct hugepd_freelist {
471 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
473 static void hugepd_free_rcu_callback(struct rcu_head *head)
475 struct hugepd_freelist *batch =
476 container_of(head, struct hugepd_freelist, rcu);
479 for (i = 0; i < batch->index; i++)
480 kmem_cache_free(hugepte_cache, batch->ptes[i]);
482 free_page((unsigned long)batch);
485 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
487 struct hugepd_freelist **batchp;
489 batchp = this_cpu_ptr(&hugepd_freelist_cur);
491 if (atomic_read(&tlb->mm->mm_users) < 2 ||
492 cpumask_equal(mm_cpumask(tlb->mm),
493 cpumask_of(smp_processor_id()))) {
494 kmem_cache_free(hugepte_cache, hugepte);
495 put_cpu_var(hugepd_freelist_cur);
499 if (*batchp == NULL) {
500 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
501 (*batchp)->index = 0;
504 (*batchp)->ptes[(*batchp)->index++] = hugepte;
505 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
506 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
509 put_cpu_var(hugepd_freelist_cur);
513 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
514 unsigned long start, unsigned long end,
515 unsigned long floor, unsigned long ceiling)
517 pte_t *hugepte = hugepd_page(*hpdp);
520 unsigned long pdmask = ~((1UL << pdshift) - 1);
521 unsigned int num_hugepd = 1;
523 #ifdef CONFIG_PPC_FSL_BOOK3E
524 /* Note: On fsl the hpdp may be the first of several */
525 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
527 unsigned int shift = hugepd_shift(*hpdp);
538 if (end - 1 > ceiling - 1)
541 for (i = 0; i < num_hugepd; i++, hpdp++)
544 #ifdef CONFIG_PPC_FSL_BOOK3E
545 hugepd_free(tlb, hugepte);
547 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
551 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
552 unsigned long addr, unsigned long end,
553 unsigned long floor, unsigned long ceiling)
561 pmd = pmd_offset(pud, addr);
562 next = pmd_addr_end(addr, end);
563 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
565 * if it is not hugepd pointer, we should already find
568 WARN_ON(!pmd_none_or_clear_bad(pmd));
571 #ifdef CONFIG_PPC_FSL_BOOK3E
573 * Increment next by the size of the huge mapping since
574 * there may be more than one entry at this level for a
575 * single hugepage, but all of them point to
576 * the same kmem cache that holds the hugepte.
578 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
580 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
581 addr, next, floor, ceiling);
582 } while (addr = next, addr != end);
592 if (end - 1 > ceiling - 1)
595 pmd = pmd_offset(pud, start);
597 pmd_free_tlb(tlb, pmd, start);
598 mm_dec_nr_pmds(tlb->mm);
601 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
602 unsigned long addr, unsigned long end,
603 unsigned long floor, unsigned long ceiling)
611 pud = pud_offset(pgd, addr);
612 next = pud_addr_end(addr, end);
613 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
614 if (pud_none_or_clear_bad(pud))
616 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
619 #ifdef CONFIG_PPC_FSL_BOOK3E
621 * Increment next by the size of the huge mapping since
622 * there may be more than one entry at this level for a
623 * single hugepage, but all of them point to
624 * the same kmem cache that holds the hugepte.
626 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
628 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
629 addr, next, floor, ceiling);
631 } while (addr = next, addr != end);
637 ceiling &= PGDIR_MASK;
641 if (end - 1 > ceiling - 1)
644 pud = pud_offset(pgd, start);
646 pud_free_tlb(tlb, pud, start);
650 * This function frees user-level page tables of a process.
652 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
653 unsigned long addr, unsigned long end,
654 unsigned long floor, unsigned long ceiling)
660 * Because there are a number of different possible pagetable
661 * layouts for hugepage ranges, we limit knowledge of how
662 * things should be laid out to the allocation path
663 * (huge_pte_alloc(), above). Everything else works out the
664 * structure as it goes from information in the hugepd
665 * pointers. That means that we can't here use the
666 * optimization used in the normal page free_pgd_range(), of
667 * checking whether we're actually covering a large enough
668 * range to have to do anything at the top level of the walk
669 * instead of at the bottom.
671 * To make sense of this, you should probably go read the big
672 * block comment at the top of the normal free_pgd_range(),
677 next = pgd_addr_end(addr, end);
678 pgd = pgd_offset(tlb->mm, addr);
679 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
680 if (pgd_none_or_clear_bad(pgd))
682 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
684 #ifdef CONFIG_PPC_FSL_BOOK3E
686 * Increment next by the size of the huge mapping since
687 * there may be more than one entry at the pgd level
688 * for a single hugepage, but all of them point to the
689 * same kmem cache that holds the hugepte.
691 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
693 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
694 addr, next, floor, ceiling);
696 } while (addr = next, addr != end);
700 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
701 * To prevent hugepage split, disable irq.
704 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
709 unsigned long mask, flags;
710 struct page *page = ERR_PTR(-EINVAL);
712 local_irq_save(flags);
713 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
716 pte = READ_ONCE(*ptep);
718 * Verify it is a huge page else bail.
719 * Transparent hugepages are handled by generic code. We can skip them
722 if (!shift || is_thp)
725 if (!pte_present(pte)) {
729 mask = (1UL << shift) - 1;
730 page = pte_page(pte);
732 page += (address & mask) / PAGE_SIZE;
735 local_irq_restore(flags);
740 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
741 pmd_t *pmd, int write)
748 follow_huge_pud(struct mm_struct *mm, unsigned long address,
749 pud_t *pud, int write)
755 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
758 unsigned long __boundary = (addr + sz) & ~(sz-1);
759 return (__boundary - 1 < end - 1) ? __boundary : end;
762 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
763 unsigned long end, int write, struct page **pages, int *nr)
766 unsigned long sz = 1UL << hugepd_shift(hugepd);
769 ptep = hugepte_offset(hugepd, addr, pdshift);
771 next = hugepte_addr_end(addr, end, sz);
772 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
774 } while (ptep++, addr = next, addr != end);
779 #ifdef CONFIG_PPC_MM_SLICES
780 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
781 unsigned long len, unsigned long pgoff,
784 struct hstate *hstate = hstate_file(file);
785 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
787 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
791 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
793 #ifdef CONFIG_PPC_MM_SLICES
794 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
796 return 1UL << mmu_psize_to_shift(psize);
798 if (!is_vm_hugetlb_page(vma))
801 return huge_page_size(hstate_vma(vma));
805 static inline bool is_power_of_4(unsigned long x)
807 if (is_power_of_2(x))
808 return (__ilog2(x) % 2) ? false : true;
812 static int __init add_huge_page_size(unsigned long long size)
814 int shift = __ffs(size);
817 /* Check that it is a page size supported by the hardware and
818 * that it fits within pagetable and slice limits. */
819 #ifdef CONFIG_PPC_FSL_BOOK3E
820 if ((size < PAGE_SIZE) || !is_power_of_4(size))
823 if (!is_power_of_2(size)
824 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
828 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
831 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
833 /* Return if huge page size has already been setup */
834 if (size_to_hstate(size))
837 hugetlb_add_hstate(shift - PAGE_SHIFT);
842 static int __init hugepage_setup_sz(char *str)
844 unsigned long long size;
846 size = memparse(str, &str);
848 if (add_huge_page_size(size) != 0)
849 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
853 __setup("hugepagesz=", hugepage_setup_sz);
855 #ifdef CONFIG_PPC_FSL_BOOK3E
856 struct kmem_cache *hugepte_cache;
857 static int __init hugetlbpage_init(void)
861 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
864 if (!mmu_psize_defs[psize].shift)
867 shift = mmu_psize_to_shift(psize);
869 /* Don't treat normal page sizes as huge... */
870 if (shift != PAGE_SHIFT)
871 if (add_huge_page_size(1ULL << shift) < 0)
876 * Create a kmem cache for hugeptes. The bottom bits in the pte have
877 * size information encoded in them, so align them to allow this
879 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
880 HUGEPD_SHIFT_MASK + 1, 0, NULL);
881 if (hugepte_cache == NULL)
882 panic("%s: Unable to create kmem cache for hugeptes\n",
885 /* Default hpage size = 4M */
886 if (mmu_psize_defs[MMU_PAGE_4M].shift)
887 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
889 panic("%s: Unable to set default huge page size\n", __func__);
895 static int __init hugetlbpage_init(void)
899 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
902 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
906 if (!mmu_psize_defs[psize].shift)
909 shift = mmu_psize_to_shift(psize);
911 if (add_huge_page_size(1ULL << shift) < 0)
914 if (shift < PMD_SHIFT)
916 else if (shift < PUD_SHIFT)
919 pdshift = PGDIR_SHIFT;
921 * if we have pdshift and shift value same, we don't
922 * use pgt cache for hugepd.
924 if (pdshift != shift) {
925 pgtable_cache_add(pdshift - shift, NULL);
926 if (!PGT_CACHE(pdshift - shift))
927 panic("hugetlbpage_init(): could not create "
928 "pgtable cache for %d bit pagesize\n", shift);
932 /* Set default large page size. Currently, we pick 16M or 1M
933 * depending on what is available
935 if (mmu_psize_defs[MMU_PAGE_16M].shift)
936 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
937 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
938 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
943 arch_initcall(hugetlbpage_init);
945 void flush_dcache_icache_hugepage(struct page *page)
950 BUG_ON(!PageCompound(page));
952 for (i = 0; i < (1UL << compound_order(page)); i++) {
953 if (!PageHighMem(page)) {
954 __flush_dcache_icache(page_address(page+i));
956 start = kmap_atomic(page+i);
957 __flush_dcache_icache(start);
958 kunmap_atomic(start);
963 #endif /* CONFIG_HUGETLB_PAGE */
966 * We have 4 cases for pgds and pmds:
967 * (1) invalid (all zeroes)
968 * (2) pointer to next table, as normal; bottom 6 bits == 0
969 * (3) leaf pte for huge page, bottom two bits != 00
970 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
972 * So long as we atomically load page table pointers we are safe against teardown,
973 * we can follow the address down to the the page and take a ref on it.
974 * This function need to be called with interrupts disabled. We use this variant
975 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
978 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
979 bool *is_thp, unsigned *shift)
985 hugepd_t *hpdp = NULL;
986 unsigned pdshift = PGDIR_SHIFT;
994 pgdp = pgdir + pgd_index(ea);
995 pgd = READ_ONCE(*pgdp);
997 * Always operate on the local stack value. This make sure the
998 * value don't get updated by a parallel THP split/collapse,
999 * page fault or a page unmap. The return pte_t * is still not
1000 * stable. So should be checked there for above conditions.
1004 else if (pgd_huge(pgd)) {
1005 ret_pte = (pte_t *) pgdp;
1007 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
1008 hpdp = (hugepd_t *)&pgd;
1011 * Even if we end up with an unmap, the pgtable will not
1012 * be freed, because we do an rcu free and here we are
1015 pdshift = PUD_SHIFT;
1016 pudp = pud_offset(&pgd, ea);
1017 pud = READ_ONCE(*pudp);
1021 else if (pud_huge(pud)) {
1022 ret_pte = (pte_t *) pudp;
1024 } else if (is_hugepd(__hugepd(pud_val(pud))))
1025 hpdp = (hugepd_t *)&pud;
1027 pdshift = PMD_SHIFT;
1028 pmdp = pmd_offset(&pud, ea);
1029 pmd = READ_ONCE(*pmdp);
1031 * A hugepage collapse is captured by pmd_none, because
1032 * it mark the pmd none and do a hpte invalidate.
1034 * We don't worry about pmd_trans_splitting here, The
1035 * caller if it needs to handle the splitting case
1036 * should check for that.
1041 if (pmd_trans_huge(pmd)) {
1044 ret_pte = (pte_t *) pmdp;
1048 if (pmd_huge(pmd)) {
1049 ret_pte = (pte_t *) pmdp;
1051 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
1052 hpdp = (hugepd_t *)&pmd;
1054 return pte_offset_kernel(&pmd, ea);
1060 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
1061 pdshift = hugepd_shift(*hpdp);
1067 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
1069 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1070 unsigned long end, int write, struct page **pages, int *nr)
1073 unsigned long pte_end;
1074 struct page *head, *page, *tail;
1078 pte_end = (addr + sz) & ~(sz-1);
1082 pte = READ_ONCE(*ptep);
1083 mask = _PAGE_PRESENT | _PAGE_USER;
1087 if ((pte_val(pte) & mask) != mask)
1090 /* hugepages are never "special" */
1091 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1094 head = pte_page(pte);
1096 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1099 VM_BUG_ON(compound_head(page) != head);
1104 } while (addr += PAGE_SIZE, addr != end);
1106 if (!page_cache_add_speculative(head, refs)) {
1111 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1112 /* Could be optimized better */
1120 * Any tail page need their mapcount reference taken before we
1125 get_huge_page_tail(tail);