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 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
58 /* Only called for hugetlbfs pages, hence can ignore THP */
59 return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
62 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
63 unsigned long address, unsigned pdshift, unsigned pshift)
65 struct kmem_cache *cachep;
68 #ifdef CONFIG_PPC_FSL_BOOK3E
70 int num_hugepd = 1 << (pshift - pdshift);
71 cachep = hugepte_cache;
73 cachep = PGT_CACHE(pdshift - pshift);
76 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
78 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
79 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
84 spin_lock(&mm->page_table_lock);
85 #ifdef CONFIG_PPC_FSL_BOOK3E
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.
92 for (i = 0; i < num_hugepd; i++, hpdp++) {
93 if (unlikely(!hugepd_none(*hpdp)))
96 /* We use the old format for PPC_FSL_BOOK3E */
97 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
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--)
103 kmem_cache_free(cachep, new);
106 if (!hugepd_none(*hpdp))
107 kmem_cache_free(cachep, new);
109 #ifdef CONFIG_PPC_BOOK3S_64
110 hpdp->pd = (unsigned long)new |
111 (shift_to_mmu_psize(pshift) << 2);
113 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
117 spin_unlock(&mm->page_table_lock);
122 * These macros define how to determine which level of the page table holds
125 #ifdef CONFIG_PPC_FSL_BOOK3E
126 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
127 #define HUGEPD_PUD_SHIFT PUD_SHIFT
129 #define HUGEPD_PGD_SHIFT PUD_SHIFT
130 #define HUGEPD_PUD_SHIFT PMD_SHIFT
133 #ifdef CONFIG_PPC_BOOK3S_64
135 * At this point we do the placement change only for BOOK3S 64. This would
136 * possibly work on other subarchs.
138 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
143 hugepd_t *hpdp = NULL;
144 unsigned pshift = __ffs(sz);
145 unsigned pdshift = PGDIR_SHIFT;
148 pg = pgd_offset(mm, addr);
150 if (pshift == PGDIR_SHIFT)
153 else if (pshift > PUD_SHIFT)
155 * We need to use hugepd table
157 hpdp = (hugepd_t *)pg;
160 pu = pud_alloc(mm, pg, addr);
161 if (pshift == PUD_SHIFT)
163 else if (pshift > PMD_SHIFT)
164 hpdp = (hugepd_t *)pu;
167 pm = pmd_alloc(mm, pu, addr);
168 if (pshift == PMD_SHIFT)
172 hpdp = (hugepd_t *)pm;
178 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
180 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
183 return hugepte_offset(*hpdp, addr, pdshift);
188 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
193 hugepd_t *hpdp = NULL;
194 unsigned pshift = __ffs(sz);
195 unsigned pdshift = PGDIR_SHIFT;
199 pg = pgd_offset(mm, addr);
201 if (pshift >= HUGEPD_PGD_SHIFT) {
202 hpdp = (hugepd_t *)pg;
205 pu = pud_alloc(mm, pg, addr);
206 if (pshift >= HUGEPD_PUD_SHIFT) {
207 hpdp = (hugepd_t *)pu;
210 pm = pmd_alloc(mm, pu, addr);
211 hpdp = (hugepd_t *)pm;
218 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
220 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
223 return hugepte_offset(*hpdp, addr, pdshift);
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.
231 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
233 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
239 gpage_freearray[idx].nr_gpages = number_of_pages;
241 for (i = 0; i < number_of_pages; i++) {
242 gpage_freearray[idx].gpage_list[i] = addr;
248 * Moves the gigantic page addresses from the temporary list to the
249 * huge_boot_pages list.
251 int alloc_bootmem_huge_page(struct hstate *hstate)
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;
260 #ifdef CONFIG_HIGHMEM
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
265 m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
266 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
268 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
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;
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.
283 unsigned long gpage_npages[MMU_PAGE_COUNT];
285 static int __init do_gpage_early_setup(char *param, char *val,
286 const char *unused, void *arg)
288 static phys_addr_t size;
289 unsigned long npages;
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.
298 if ((strcmp(param, "default_hugepagesz") == 0) ||
299 (strcmp(param, "hugepagesz") == 0)) {
300 size = memparse(val, NULL);
301 } else if (strcmp(param, "hugepages") == 0) {
303 if (sscanf(val, "%lu", &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;
312 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
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.
327 void __init reserve_hugetlb_gpages(void)
329 static __initdata char cmdline[COMMAND_LINE_SIZE];
330 phys_addr_t size, base;
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);
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.
343 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
344 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
346 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
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]);
356 #else /* !PPC_FSL_BOOK3E */
358 /* Build list of addresses of gigantic pages. This function is used in early
359 * boot before the buddy allocator is setup.
361 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
365 while (number_of_pages > 0) {
366 gpage_freearray[nr_gpages] = addr;
373 /* Moves the gigantic page addresses from the temporary list to the
374 * huge_boot_pages list.
376 int alloc_bootmem_huge_page(struct hstate *hstate)
378 struct huge_bootmem_page *m;
381 m = phys_to_virt(gpage_freearray[--nr_gpages]);
382 gpage_freearray[nr_gpages] = 0;
383 list_add(&m->list, &huge_boot_pages);
389 #ifdef CONFIG_PPC_FSL_BOOK3E
390 #define HUGEPD_FREELIST_SIZE \
391 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
393 struct hugepd_freelist {
399 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
401 static void hugepd_free_rcu_callback(struct rcu_head *head)
403 struct hugepd_freelist *batch =
404 container_of(head, struct hugepd_freelist, rcu);
407 for (i = 0; i < batch->index; i++)
408 kmem_cache_free(hugepte_cache, batch->ptes[i]);
410 free_page((unsigned long)batch);
413 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
415 struct hugepd_freelist **batchp;
417 batchp = this_cpu_ptr(&hugepd_freelist_cur);
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);
427 if (*batchp == NULL) {
428 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
429 (*batchp)->index = 0;
432 (*batchp)->ptes[(*batchp)->index++] = hugepte;
433 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
434 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
437 put_cpu_var(hugepd_freelist_cur);
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)
445 pte_t *hugepte = hugepd_page(*hpdp);
448 unsigned long pdmask = ~((1UL << pdshift) - 1);
449 unsigned int num_hugepd = 1;
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));
455 unsigned int shift = hugepd_shift(*hpdp);
466 if (end - 1 > ceiling - 1)
469 for (i = 0; i < num_hugepd; i++, hpdp++)
472 #ifdef CONFIG_PPC_FSL_BOOK3E
473 hugepd_free(tlb, hugepte);
475 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
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)
489 pmd = pmd_offset(pud, addr);
490 next = pmd_addr_end(addr, end);
491 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
493 * if it is not hugepd pointer, we should already find
496 WARN_ON(!pmd_none_or_clear_bad(pmd));
499 #ifdef CONFIG_PPC_FSL_BOOK3E
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.
506 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
508 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
509 addr, next, floor, ceiling);
510 } while (addr = next, addr != end);
520 if (end - 1 > ceiling - 1)
523 pmd = pmd_offset(pud, start);
525 pmd_free_tlb(tlb, pmd, start);
526 mm_dec_nr_pmds(tlb->mm);
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)
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))
544 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
547 #ifdef CONFIG_PPC_FSL_BOOK3E
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.
554 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
556 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
557 addr, next, floor, ceiling);
559 } while (addr = next, addr != end);
565 ceiling &= PGDIR_MASK;
569 if (end - 1 > ceiling - 1)
572 pud = pud_offset(pgd, start);
574 pud_free_tlb(tlb, pud, start);
578 * This function frees user-level page tables of a process.
580 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
581 unsigned long addr, unsigned long end,
582 unsigned long floor, unsigned long ceiling)
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.
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(),
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))
610 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
612 #ifdef CONFIG_PPC_FSL_BOOK3E
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.
619 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
621 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
622 addr, next, floor, ceiling);
624 } while (addr = next, addr != end);
628 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
629 * To prevent hugepage split, disable irq.
632 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
637 unsigned long mask, flags;
638 struct page *page = ERR_PTR(-EINVAL);
640 local_irq_save(flags);
641 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
644 pte = READ_ONCE(*ptep);
646 * Verify it is a huge page else bail.
647 * Transparent hugepages are handled by generic code. We can skip them
650 if (!shift || is_thp)
653 if (!pte_present(pte)) {
657 mask = (1UL << shift) - 1;
658 page = pte_page(pte);
660 page += (address & mask) / PAGE_SIZE;
663 local_irq_restore(flags);
668 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
669 pmd_t *pmd, int write)
676 follow_huge_pud(struct mm_struct *mm, unsigned long address,
677 pud_t *pud, int write)
683 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
686 unsigned long __boundary = (addr + sz) & ~(sz-1);
687 return (__boundary - 1 < end - 1) ? __boundary : end;
690 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
691 unsigned long end, int write, struct page **pages, int *nr)
694 unsigned long sz = 1UL << hugepd_shift(hugepd);
697 ptep = hugepte_offset(hugepd, addr, pdshift);
699 next = hugepte_addr_end(addr, end, sz);
700 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
702 } while (ptep++, addr = next, addr != end);
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,
712 struct hstate *hstate = hstate_file(file);
713 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
715 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
719 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
721 #ifdef CONFIG_PPC_MM_SLICES
722 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
724 return 1UL << mmu_psize_to_shift(psize);
726 if (!is_vm_hugetlb_page(vma))
729 return huge_page_size(hstate_vma(vma));
733 static inline bool is_power_of_4(unsigned long x)
735 if (is_power_of_2(x))
736 return (__ilog2(x) % 2) ? false : true;
740 static int __init add_huge_page_size(unsigned long long size)
742 int shift = __ffs(size);
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))
751 if (!is_power_of_2(size)
752 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
756 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
759 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
761 /* Return if huge page size has already been setup */
762 if (size_to_hstate(size))
765 hugetlb_add_hstate(shift - PAGE_SHIFT);
770 static int __init hugepage_setup_sz(char *str)
772 unsigned long long size;
774 size = memparse(str, &str);
776 if (add_huge_page_size(size) != 0)
777 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
781 __setup("hugepagesz=", hugepage_setup_sz);
783 #ifdef CONFIG_PPC_FSL_BOOK3E
784 struct kmem_cache *hugepte_cache;
785 static int __init hugetlbpage_init(void)
789 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
792 if (!mmu_psize_defs[psize].shift)
795 shift = mmu_psize_to_shift(psize);
797 /* Don't treat normal page sizes as huge... */
798 if (shift != PAGE_SHIFT)
799 if (add_huge_page_size(1ULL << shift) < 0)
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
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",
813 /* Default hpage size = 4M */
814 if (mmu_psize_defs[MMU_PAGE_4M].shift)
815 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
817 panic("%s: Unable to set default huge page size\n", __func__);
823 static int __init hugetlbpage_init(void)
827 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
830 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
834 if (!mmu_psize_defs[psize].shift)
837 shift = mmu_psize_to_shift(psize);
839 if (add_huge_page_size(1ULL << shift) < 0)
842 if (shift < PMD_SHIFT)
844 else if (shift < PUD_SHIFT)
847 pdshift = PGDIR_SHIFT;
849 * if we have pdshift and shift value same, we don't
850 * use pgt cache for hugepd.
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);
860 /* Set default large page size. Currently, we pick 16M or 1M
861 * depending on what is available
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;
871 arch_initcall(hugetlbpage_init);
873 void flush_dcache_icache_hugepage(struct page *page)
878 BUG_ON(!PageCompound(page));
880 for (i = 0; i < (1UL << compound_order(page)); i++) {
881 if (!PageHighMem(page)) {
882 __flush_dcache_icache(page_address(page+i));
884 start = kmap_atomic(page+i);
885 __flush_dcache_icache(start);
886 kunmap_atomic(start);
891 #endif /* CONFIG_HUGETLB_PAGE */
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
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
906 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
907 bool *is_thp, unsigned *shift)
913 hugepd_t *hpdp = NULL;
914 unsigned pdshift = PGDIR_SHIFT;
922 pgdp = pgdir + pgd_index(ea);
923 pgd = READ_ONCE(*pgdp);
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.
932 else if (pgd_huge(pgd)) {
933 ret_pte = (pte_t *) pgdp;
935 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
936 hpdp = (hugepd_t *)&pgd;
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
944 pudp = pud_offset(&pgd, ea);
945 pud = READ_ONCE(*pudp);
949 else if (pud_huge(pud)) {
950 ret_pte = (pte_t *) pudp;
952 } else if (is_hugepd(__hugepd(pud_val(pud))))
953 hpdp = (hugepd_t *)&pud;
956 pmdp = pmd_offset(&pud, ea);
957 pmd = READ_ONCE(*pmdp);
959 * A hugepage collapse is captured by pmd_none, because
960 * it mark the pmd none and do a hpte invalidate.
965 if (pmd_trans_huge(pmd)) {
968 ret_pte = (pte_t *) pmdp;
973 ret_pte = (pte_t *) pmdp;
975 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
976 hpdp = (hugepd_t *)&pmd;
978 return pte_offset_kernel(&pmd, ea);
984 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
985 pdshift = hugepd_shift(*hpdp);
991 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
993 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
994 unsigned long end, int write, struct page **pages, int *nr)
997 unsigned long pte_end;
998 struct page *head, *page;
1002 pte_end = (addr + sz) & ~(sz-1);
1006 pte = READ_ONCE(*ptep);
1007 mask = _PAGE_PRESENT | _PAGE_USER;
1011 if ((pte_val(pte) & mask) != mask)
1014 /* hugepages are never "special" */
1015 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1018 head = pte_page(pte);
1020 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1022 VM_BUG_ON(compound_head(page) != head);
1027 } while (addr += PAGE_SIZE, addr != end);
1029 if (!page_cache_add_speculative(head, refs)) {
1034 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1035 /* Could be optimized better */