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:
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 int pmd_huge(pmd_t pmd)
69 * leaf pte for huge page, bottom two bits != 00
71 return ((pmd_val(pmd) & 0x3) != 0x0);
74 int pud_huge(pud_t pud)
77 * leaf pte for huge page, bottom two bits != 00
79 return ((pud_val(pud) & 0x3) != 0x0);
82 int pgd_huge(pgd_t pgd)
85 * leaf pte for huge page, bottom two bits != 00
87 return ((pgd_val(pgd) & 0x3) != 0x0);
90 int pmd_huge_support(void)
95 int pmd_huge(pmd_t pmd)
100 int pud_huge(pud_t pud)
105 int pgd_huge(pgd_t pgd)
110 int pmd_huge_support(void)
116 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
118 /* Only called for hugetlbfs pages, hence can ignore THP */
119 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
122 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
123 unsigned long address, unsigned pdshift, unsigned pshift)
125 struct kmem_cache *cachep;
128 #ifdef CONFIG_PPC_FSL_BOOK3E
130 int num_hugepd = 1 << (pshift - pdshift);
131 cachep = hugepte_cache;
133 cachep = PGT_CACHE(pdshift - pshift);
136 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
138 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
139 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
144 spin_lock(&mm->page_table_lock);
145 #ifdef CONFIG_PPC_FSL_BOOK3E
147 * We have multiple higher-level entries that point to the same
148 * actual pte location. Fill in each as we go and backtrack on error.
149 * We need all of these so the DTLB pgtable walk code can find the
150 * right higher-level entry without knowing if it's a hugepage or not.
152 for (i = 0; i < num_hugepd; i++, hpdp++) {
153 if (unlikely(!hugepd_none(*hpdp)))
156 /* We use the old format for PPC_FSL_BOOK3E */
157 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
159 /* If we bailed from the for loop early, an error occurred, clean up */
160 if (i < num_hugepd) {
161 for (i = i - 1 ; i >= 0; i--, hpdp--)
163 kmem_cache_free(cachep, new);
166 if (!hugepd_none(*hpdp))
167 kmem_cache_free(cachep, new);
169 #ifdef CONFIG_PPC_BOOK3S_64
170 hpdp->pd = (unsigned long)new |
171 (shift_to_mmu_psize(pshift) << 2);
173 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
177 spin_unlock(&mm->page_table_lock);
182 * These macros define how to determine which level of the page table holds
185 #ifdef CONFIG_PPC_FSL_BOOK3E
186 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
187 #define HUGEPD_PUD_SHIFT PUD_SHIFT
189 #define HUGEPD_PGD_SHIFT PUD_SHIFT
190 #define HUGEPD_PUD_SHIFT PMD_SHIFT
193 #ifdef CONFIG_PPC_BOOK3S_64
195 * At this point we do the placement change only for BOOK3S 64. This would
196 * possibly work on other subarchs.
198 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
203 hugepd_t *hpdp = NULL;
204 unsigned pshift = __ffs(sz);
205 unsigned pdshift = PGDIR_SHIFT;
208 pg = pgd_offset(mm, addr);
210 if (pshift == PGDIR_SHIFT)
213 else if (pshift > PUD_SHIFT)
215 * We need to use hugepd table
217 hpdp = (hugepd_t *)pg;
220 pu = pud_alloc(mm, pg, addr);
221 if (pshift == PUD_SHIFT)
223 else if (pshift > PMD_SHIFT)
224 hpdp = (hugepd_t *)pu;
227 pm = pmd_alloc(mm, pu, addr);
228 if (pshift == PMD_SHIFT)
232 hpdp = (hugepd_t *)pm;
238 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
240 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
243 return hugepte_offset(hpdp, addr, pdshift);
248 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
253 hugepd_t *hpdp = NULL;
254 unsigned pshift = __ffs(sz);
255 unsigned pdshift = PGDIR_SHIFT;
259 pg = pgd_offset(mm, addr);
261 if (pshift >= HUGEPD_PGD_SHIFT) {
262 hpdp = (hugepd_t *)pg;
265 pu = pud_alloc(mm, pg, addr);
266 if (pshift >= HUGEPD_PUD_SHIFT) {
267 hpdp = (hugepd_t *)pu;
270 pm = pmd_alloc(mm, pu, addr);
271 hpdp = (hugepd_t *)pm;
278 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
280 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
283 return hugepte_offset(hpdp, addr, pdshift);
287 #ifdef CONFIG_PPC_FSL_BOOK3E
288 /* Build list of addresses of gigantic pages. This function is used in early
289 * boot before the buddy or bootmem allocator is setup.
291 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
293 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
299 gpage_freearray[idx].nr_gpages = number_of_pages;
301 for (i = 0; i < number_of_pages; i++) {
302 gpage_freearray[idx].gpage_list[i] = addr;
308 * Moves the gigantic page addresses from the temporary list to the
309 * huge_boot_pages list.
311 int alloc_bootmem_huge_page(struct hstate *hstate)
313 struct huge_bootmem_page *m;
314 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
315 int nr_gpages = gpage_freearray[idx].nr_gpages;
320 #ifdef CONFIG_HIGHMEM
322 * If gpages can be in highmem we can't use the trick of storing the
323 * data structure in the page; allocate space for this
325 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
326 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
328 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
331 list_add(&m->list, &huge_boot_pages);
332 gpage_freearray[idx].nr_gpages = nr_gpages;
333 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
339 * Scan the command line hugepagesz= options for gigantic pages; store those in
340 * a list that we use to allocate the memory once all options are parsed.
343 unsigned long gpage_npages[MMU_PAGE_COUNT];
345 static int __init do_gpage_early_setup(char *param, char *val,
348 static phys_addr_t size;
349 unsigned long npages;
352 * The hugepagesz and hugepages cmdline options are interleaved. We
353 * use the size variable to keep track of whether or not this was done
354 * properly and skip over instances where it is incorrect. Other
355 * command-line parsing code will issue warnings, so we don't need to.
358 if ((strcmp(param, "default_hugepagesz") == 0) ||
359 (strcmp(param, "hugepagesz") == 0)) {
360 size = memparse(val, NULL);
361 } else if (strcmp(param, "hugepages") == 0) {
363 if (sscanf(val, "%lu", &npages) <= 0)
365 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
374 * This function allocates physical space for pages that are larger than the
375 * buddy allocator can handle. We want to allocate these in highmem because
376 * the amount of lowmem is limited. This means that this function MUST be
377 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
378 * allocate to grab highmem.
380 void __init reserve_hugetlb_gpages(void)
382 static __initdata char cmdline[COMMAND_LINE_SIZE];
383 phys_addr_t size, base;
386 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
387 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
388 &do_gpage_early_setup);
391 * Walk gpage list in reverse, allocating larger page sizes first.
392 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
393 * When we reach the point in the list where pages are no longer
394 * considered gpages, we're done.
396 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
397 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
399 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
402 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
403 base = memblock_alloc_base(size * gpage_npages[i], size,
404 MEMBLOCK_ALLOC_ANYWHERE);
405 add_gpage(base, size, gpage_npages[i]);
409 #else /* !PPC_FSL_BOOK3E */
411 /* Build list of addresses of gigantic pages. This function is used in early
412 * boot before the buddy or bootmem allocator is setup.
414 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
418 while (number_of_pages > 0) {
419 gpage_freearray[nr_gpages] = addr;
426 /* Moves the gigantic page addresses from the temporary list to the
427 * huge_boot_pages list.
429 int alloc_bootmem_huge_page(struct hstate *hstate)
431 struct huge_bootmem_page *m;
434 m = phys_to_virt(gpage_freearray[--nr_gpages]);
435 gpage_freearray[nr_gpages] = 0;
436 list_add(&m->list, &huge_boot_pages);
442 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
447 #ifdef CONFIG_PPC_FSL_BOOK3E
448 #define HUGEPD_FREELIST_SIZE \
449 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
451 struct hugepd_freelist {
457 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
459 static void hugepd_free_rcu_callback(struct rcu_head *head)
461 struct hugepd_freelist *batch =
462 container_of(head, struct hugepd_freelist, rcu);
465 for (i = 0; i < batch->index; i++)
466 kmem_cache_free(hugepte_cache, batch->ptes[i]);
468 free_page((unsigned long)batch);
471 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
473 struct hugepd_freelist **batchp;
475 batchp = &__get_cpu_var(hugepd_freelist_cur);
477 if (atomic_read(&tlb->mm->mm_users) < 2 ||
478 cpumask_equal(mm_cpumask(tlb->mm),
479 cpumask_of(smp_processor_id()))) {
480 kmem_cache_free(hugepte_cache, hugepte);
484 if (*batchp == NULL) {
485 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
486 (*batchp)->index = 0;
489 (*batchp)->ptes[(*batchp)->index++] = hugepte;
490 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
491 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
497 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
498 unsigned long start, unsigned long end,
499 unsigned long floor, unsigned long ceiling)
501 pte_t *hugepte = hugepd_page(*hpdp);
504 unsigned long pdmask = ~((1UL << pdshift) - 1);
505 unsigned int num_hugepd = 1;
507 #ifdef CONFIG_PPC_FSL_BOOK3E
508 /* Note: On fsl the hpdp may be the first of several */
509 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
511 unsigned int shift = hugepd_shift(*hpdp);
522 if (end - 1 > ceiling - 1)
525 for (i = 0; i < num_hugepd; i++, hpdp++)
530 #ifdef CONFIG_PPC_FSL_BOOK3E
531 hugepd_free(tlb, hugepte);
533 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
537 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
538 unsigned long addr, unsigned long end,
539 unsigned long floor, unsigned long ceiling)
547 pmd = pmd_offset(pud, addr);
548 next = pmd_addr_end(addr, end);
549 if (!is_hugepd(pmd)) {
551 * if it is not hugepd pointer, we should already find
554 WARN_ON(!pmd_none_or_clear_bad(pmd));
557 #ifdef CONFIG_PPC_FSL_BOOK3E
559 * Increment next by the size of the huge mapping since
560 * there may be more than one entry at this level for a
561 * single hugepage, but all of them point to
562 * the same kmem cache that holds the hugepte.
564 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
566 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
567 addr, next, floor, ceiling);
568 } while (addr = next, addr != end);
578 if (end - 1 > ceiling - 1)
581 pmd = pmd_offset(pud, start);
583 pmd_free_tlb(tlb, pmd, start);
586 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
587 unsigned long addr, unsigned long end,
588 unsigned long floor, unsigned long ceiling)
596 pud = pud_offset(pgd, addr);
597 next = pud_addr_end(addr, end);
598 if (!is_hugepd(pud)) {
599 if (pud_none_or_clear_bad(pud))
601 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
604 #ifdef CONFIG_PPC_FSL_BOOK3E
606 * Increment next by the size of the huge mapping since
607 * there may be more than one entry at this level for a
608 * single hugepage, but all of them point to
609 * the same kmem cache that holds the hugepte.
611 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
613 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
614 addr, next, floor, ceiling);
616 } while (addr = next, addr != end);
622 ceiling &= PGDIR_MASK;
626 if (end - 1 > ceiling - 1)
629 pud = pud_offset(pgd, start);
631 pud_free_tlb(tlb, pud, start);
635 * This function frees user-level page tables of a process.
637 * Must be called with pagetable lock held.
639 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
640 unsigned long addr, unsigned long end,
641 unsigned long floor, unsigned long ceiling)
647 * Because there are a number of different possible pagetable
648 * layouts for hugepage ranges, we limit knowledge of how
649 * things should be laid out to the allocation path
650 * (huge_pte_alloc(), above). Everything else works out the
651 * structure as it goes from information in the hugepd
652 * pointers. That means that we can't here use the
653 * optimization used in the normal page free_pgd_range(), of
654 * checking whether we're actually covering a large enough
655 * range to have to do anything at the top level of the walk
656 * instead of at the bottom.
658 * To make sense of this, you should probably go read the big
659 * block comment at the top of the normal free_pgd_range(),
664 next = pgd_addr_end(addr, end);
665 pgd = pgd_offset(tlb->mm, addr);
666 if (!is_hugepd(pgd)) {
667 if (pgd_none_or_clear_bad(pgd))
669 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
671 #ifdef CONFIG_PPC_FSL_BOOK3E
673 * Increment next by the size of the huge mapping since
674 * there may be more than one entry at the pgd level
675 * for a single hugepage, but all of them point to the
676 * same kmem cache that holds the hugepte.
678 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
680 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
681 addr, next, floor, ceiling);
683 } while (addr = next, addr != end);
687 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
694 * Transparent hugepages are handled by generic code. We can skip them
697 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
699 /* Verify it is a huge page else bail. */
700 if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
701 return ERR_PTR(-EINVAL);
703 mask = (1UL << shift) - 1;
704 page = pte_page(*ptep);
706 page += (address & mask) / PAGE_SIZE;
712 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
713 pmd_t *pmd, int write)
719 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
722 unsigned long __boundary = (addr + sz) & ~(sz-1);
723 return (__boundary - 1 < end - 1) ? __boundary : end;
726 int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
727 unsigned long addr, unsigned long end,
728 int write, struct page **pages, int *nr)
731 unsigned long sz = 1UL << hugepd_shift(*hugepd);
734 ptep = hugepte_offset(hugepd, addr, pdshift);
736 next = hugepte_addr_end(addr, end, sz);
737 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
739 } while (ptep++, addr = next, addr != end);
744 #ifdef CONFIG_PPC_MM_SLICES
745 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
746 unsigned long len, unsigned long pgoff,
749 struct hstate *hstate = hstate_file(file);
750 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
752 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
756 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
758 #ifdef CONFIG_PPC_MM_SLICES
759 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
761 return 1UL << mmu_psize_to_shift(psize);
763 if (!is_vm_hugetlb_page(vma))
766 return huge_page_size(hstate_vma(vma));
770 static inline bool is_power_of_4(unsigned long x)
772 if (is_power_of_2(x))
773 return (__ilog2(x) % 2) ? false : true;
777 static int __init add_huge_page_size(unsigned long long size)
779 int shift = __ffs(size);
782 /* Check that it is a page size supported by the hardware and
783 * that it fits within pagetable and slice limits. */
784 #ifdef CONFIG_PPC_FSL_BOOK3E
785 if ((size < PAGE_SIZE) || !is_power_of_4(size))
788 if (!is_power_of_2(size)
789 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
793 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
796 #ifdef CONFIG_SPU_FS_64K_LS
797 /* Disable support for 64K huge pages when 64K SPU local store
798 * support is enabled as the current implementation conflicts.
800 if (shift == PAGE_SHIFT_64K)
802 #endif /* CONFIG_SPU_FS_64K_LS */
804 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
806 /* Return if huge page size has already been setup */
807 if (size_to_hstate(size))
810 hugetlb_add_hstate(shift - PAGE_SHIFT);
815 static int __init hugepage_setup_sz(char *str)
817 unsigned long long size;
819 size = memparse(str, &str);
821 if (add_huge_page_size(size) != 0)
822 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
826 __setup("hugepagesz=", hugepage_setup_sz);
828 #ifdef CONFIG_PPC_FSL_BOOK3E
829 struct kmem_cache *hugepte_cache;
830 static int __init hugetlbpage_init(void)
834 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
837 if (!mmu_psize_defs[psize].shift)
840 shift = mmu_psize_to_shift(psize);
842 /* Don't treat normal page sizes as huge... */
843 if (shift != PAGE_SHIFT)
844 if (add_huge_page_size(1ULL << shift) < 0)
849 * Create a kmem cache for hugeptes. The bottom bits in the pte have
850 * size information encoded in them, so align them to allow this
852 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
853 HUGEPD_SHIFT_MASK + 1, 0, NULL);
854 if (hugepte_cache == NULL)
855 panic("%s: Unable to create kmem cache for hugeptes\n",
858 /* Default hpage size = 4M */
859 if (mmu_psize_defs[MMU_PAGE_4M].shift)
860 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
862 panic("%s: Unable to set default huge page size\n", __func__);
868 static int __init hugetlbpage_init(void)
872 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
875 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
879 if (!mmu_psize_defs[psize].shift)
882 shift = mmu_psize_to_shift(psize);
884 if (add_huge_page_size(1ULL << shift) < 0)
887 if (shift < PMD_SHIFT)
889 else if (shift < PUD_SHIFT)
892 pdshift = PGDIR_SHIFT;
894 * if we have pdshift and shift value same, we don't
895 * use pgt cache for hugepd.
897 if (pdshift != shift) {
898 pgtable_cache_add(pdshift - shift, NULL);
899 if (!PGT_CACHE(pdshift - shift))
900 panic("hugetlbpage_init(): could not create "
901 "pgtable cache for %d bit pagesize\n", shift);
905 /* Set default large page size. Currently, we pick 16M or 1M
906 * depending on what is available
908 if (mmu_psize_defs[MMU_PAGE_16M].shift)
909 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
910 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
911 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
916 module_init(hugetlbpage_init);
918 void flush_dcache_icache_hugepage(struct page *page)
923 BUG_ON(!PageCompound(page));
925 for (i = 0; i < (1UL << compound_order(page)); i++) {
926 if (!PageHighMem(page)) {
927 __flush_dcache_icache(page_address(page+i));
929 start = kmap_atomic(page+i);
930 __flush_dcache_icache(start);
931 kunmap_atomic(start);
936 #endif /* CONFIG_HUGETLB_PAGE */
939 * We have 4 cases for pgds and pmds:
940 * (1) invalid (all zeroes)
941 * (2) pointer to next table, as normal; bottom 6 bits == 0
942 * (3) leaf pte for huge page, bottom two bits != 00
943 * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
945 * So long as we atomically load page table pointers we are safe against teardown,
946 * we can follow the address down to the the page and take a ref on it.
949 pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
955 hugepd_t *hpdp = NULL;
956 unsigned pdshift = PGDIR_SHIFT;
961 pgdp = pgdir + pgd_index(ea);
962 pgd = ACCESS_ONCE(*pgdp);
964 * Always operate on the local stack value. This make sure the
965 * value don't get updated by a parallel THP split/collapse,
966 * page fault or a page unmap. The return pte_t * is still not
967 * stable. So should be checked there for above conditions.
971 else if (pgd_huge(pgd)) {
972 ret_pte = (pte_t *) pgdp;
974 } else if (is_hugepd(&pgd))
975 hpdp = (hugepd_t *)&pgd;
978 * Even if we end up with an unmap, the pgtable will not
979 * be freed, because we do an rcu free and here we are
983 pudp = pud_offset(&pgd, ea);
984 pud = ACCESS_ONCE(*pudp);
988 else if (pud_huge(pud)) {
989 ret_pte = (pte_t *) pudp;
991 } else if (is_hugepd(&pud))
992 hpdp = (hugepd_t *)&pud;
995 pmdp = pmd_offset(&pud, ea);
996 pmd = ACCESS_ONCE(*pmdp);
998 * A hugepage collapse is captured by pmd_none, because
999 * it mark the pmd none and do a hpte invalidate.
1001 * A hugepage split is captured by pmd_trans_splitting
1002 * because we mark the pmd trans splitting and do a
1006 if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1009 if (pmd_huge(pmd) || pmd_large(pmd)) {
1010 ret_pte = (pte_t *) pmdp;
1012 } else if (is_hugepd(&pmd))
1013 hpdp = (hugepd_t *)&pmd;
1015 return pte_offset_kernel(&pmd, ea);
1021 ret_pte = hugepte_offset(hpdp, ea, pdshift);
1022 pdshift = hugepd_shift(*hpdp);
1028 EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
1030 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1031 unsigned long end, int write, struct page **pages, int *nr)
1034 unsigned long pte_end;
1035 struct page *head, *page, *tail;
1039 pte_end = (addr + sz) & ~(sz-1);
1043 pte = ACCESS_ONCE(*ptep);
1044 mask = _PAGE_PRESENT | _PAGE_USER;
1048 if ((pte_val(pte) & mask) != mask)
1051 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1053 * check for splitting here
1055 if (pmd_trans_splitting(pte_pmd(pte)))
1059 /* hugepages are never "special" */
1060 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1063 head = pte_page(pte);
1065 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1068 VM_BUG_ON(compound_head(page) != head);
1073 } while (addr += PAGE_SIZE, addr != end);
1075 if (!page_cache_add_speculative(head, refs)) {
1080 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1081 /* Could be optimized better */
1089 * Any tail page need their mapcount reference taken before we
1094 get_huge_page_tail(tail);