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 = __pa(new) | (shift_to_mmu_psize(pshift) << 2);
112 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
116 spin_unlock(&mm->page_table_lock);
121 * These macros define how to determine which level of the page table holds
124 #ifdef CONFIG_PPC_FSL_BOOK3E
125 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
126 #define HUGEPD_PUD_SHIFT PUD_SHIFT
128 #define HUGEPD_PGD_SHIFT PUD_SHIFT
129 #define HUGEPD_PUD_SHIFT PMD_SHIFT
132 #ifdef CONFIG_PPC_BOOK3S_64
134 * At this point we do the placement change only for BOOK3S 64. This would
135 * possibly work on other subarchs.
137 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
142 hugepd_t *hpdp = NULL;
143 unsigned pshift = __ffs(sz);
144 unsigned pdshift = PGDIR_SHIFT;
147 pg = pgd_offset(mm, addr);
149 if (pshift == PGDIR_SHIFT)
152 else if (pshift > PUD_SHIFT)
154 * We need to use hugepd table
156 hpdp = (hugepd_t *)pg;
159 pu = pud_alloc(mm, pg, addr);
160 if (pshift == PUD_SHIFT)
162 else if (pshift > PMD_SHIFT)
163 hpdp = (hugepd_t *)pu;
166 pm = pmd_alloc(mm, pu, addr);
167 if (pshift == PMD_SHIFT)
171 hpdp = (hugepd_t *)pm;
177 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
179 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
182 return hugepte_offset(*hpdp, addr, pdshift);
187 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
192 hugepd_t *hpdp = NULL;
193 unsigned pshift = __ffs(sz);
194 unsigned pdshift = PGDIR_SHIFT;
198 pg = pgd_offset(mm, addr);
200 if (pshift >= HUGEPD_PGD_SHIFT) {
201 hpdp = (hugepd_t *)pg;
204 pu = pud_alloc(mm, pg, addr);
205 if (pshift >= HUGEPD_PUD_SHIFT) {
206 hpdp = (hugepd_t *)pu;
209 pm = pmd_alloc(mm, pu, addr);
210 hpdp = (hugepd_t *)pm;
217 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
219 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
222 return hugepte_offset(*hpdp, addr, pdshift);
226 #ifdef CONFIG_PPC_FSL_BOOK3E
227 /* Build list of addresses of gigantic pages. This function is used in early
228 * boot before the buddy allocator is setup.
230 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
232 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
238 gpage_freearray[idx].nr_gpages = number_of_pages;
240 for (i = 0; i < number_of_pages; i++) {
241 gpage_freearray[idx].gpage_list[i] = addr;
247 * Moves the gigantic page addresses from the temporary list to the
248 * huge_boot_pages list.
250 int alloc_bootmem_huge_page(struct hstate *hstate)
252 struct huge_bootmem_page *m;
253 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
254 int nr_gpages = gpage_freearray[idx].nr_gpages;
259 #ifdef CONFIG_HIGHMEM
261 * If gpages can be in highmem we can't use the trick of storing the
262 * data structure in the page; allocate space for this
264 m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
265 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
267 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
270 list_add(&m->list, &huge_boot_pages);
271 gpage_freearray[idx].nr_gpages = nr_gpages;
272 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
278 * Scan the command line hugepagesz= options for gigantic pages; store those in
279 * a list that we use to allocate the memory once all options are parsed.
282 unsigned long gpage_npages[MMU_PAGE_COUNT];
284 static int __init do_gpage_early_setup(char *param, char *val,
285 const char *unused, void *arg)
287 static phys_addr_t size;
288 unsigned long npages;
291 * The hugepagesz and hugepages cmdline options are interleaved. We
292 * use the size variable to keep track of whether or not this was done
293 * properly and skip over instances where it is incorrect. Other
294 * command-line parsing code will issue warnings, so we don't need to.
297 if ((strcmp(param, "default_hugepagesz") == 0) ||
298 (strcmp(param, "hugepagesz") == 0)) {
299 size = memparse(val, NULL);
300 } else if (strcmp(param, "hugepages") == 0) {
302 if (sscanf(val, "%lu", &npages) <= 0)
304 if (npages > MAX_NUMBER_GPAGES) {
305 pr_warn("MMU: %lu pages requested for page "
306 "size %llu KB, limiting to "
307 __stringify(MAX_NUMBER_GPAGES) "\n",
308 npages, size / 1024);
309 npages = MAX_NUMBER_GPAGES;
311 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
320 * This function allocates physical space for pages that are larger than the
321 * buddy allocator can handle. We want to allocate these in highmem because
322 * the amount of lowmem is limited. This means that this function MUST be
323 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
324 * allocate to grab highmem.
326 void __init reserve_hugetlb_gpages(void)
328 static __initdata char cmdline[COMMAND_LINE_SIZE];
329 phys_addr_t size, base;
332 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
333 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
334 NULL, &do_gpage_early_setup);
337 * Walk gpage list in reverse, allocating larger page sizes first.
338 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
339 * When we reach the point in the list where pages are no longer
340 * considered gpages, we're done.
342 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
343 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
345 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
348 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
349 base = memblock_alloc_base(size * gpage_npages[i], size,
350 MEMBLOCK_ALLOC_ANYWHERE);
351 add_gpage(base, size, gpage_npages[i]);
355 #else /* !PPC_FSL_BOOK3E */
357 /* Build list of addresses of gigantic pages. This function is used in early
358 * boot before the buddy allocator is setup.
360 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
364 while (number_of_pages > 0) {
365 gpage_freearray[nr_gpages] = addr;
372 /* Moves the gigantic page addresses from the temporary list to the
373 * huge_boot_pages list.
375 int alloc_bootmem_huge_page(struct hstate *hstate)
377 struct huge_bootmem_page *m;
380 m = phys_to_virt(gpage_freearray[--nr_gpages]);
381 gpage_freearray[nr_gpages] = 0;
382 list_add(&m->list, &huge_boot_pages);
388 #ifdef CONFIG_PPC_FSL_BOOK3E
389 #define HUGEPD_FREELIST_SIZE \
390 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
392 struct hugepd_freelist {
398 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
400 static void hugepd_free_rcu_callback(struct rcu_head *head)
402 struct hugepd_freelist *batch =
403 container_of(head, struct hugepd_freelist, rcu);
406 for (i = 0; i < batch->index; i++)
407 kmem_cache_free(hugepte_cache, batch->ptes[i]);
409 free_page((unsigned long)batch);
412 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
414 struct hugepd_freelist **batchp;
416 batchp = &get_cpu_var(hugepd_freelist_cur);
418 if (atomic_read(&tlb->mm->mm_users) < 2 ||
419 cpumask_equal(mm_cpumask(tlb->mm),
420 cpumask_of(smp_processor_id()))) {
421 kmem_cache_free(hugepte_cache, hugepte);
422 put_cpu_var(hugepd_freelist_cur);
426 if (*batchp == NULL) {
427 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
428 (*batchp)->index = 0;
431 (*batchp)->ptes[(*batchp)->index++] = hugepte;
432 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
433 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
436 put_cpu_var(hugepd_freelist_cur);
440 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
441 unsigned long start, unsigned long end,
442 unsigned long floor, unsigned long ceiling)
444 pte_t *hugepte = hugepd_page(*hpdp);
447 unsigned long pdmask = ~((1UL << pdshift) - 1);
448 unsigned int num_hugepd = 1;
450 #ifdef CONFIG_PPC_FSL_BOOK3E
451 /* Note: On fsl the hpdp may be the first of several */
452 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
454 unsigned int shift = hugepd_shift(*hpdp);
465 if (end - 1 > ceiling - 1)
468 for (i = 0; i < num_hugepd; i++, hpdp++)
471 #ifdef CONFIG_PPC_FSL_BOOK3E
472 hugepd_free(tlb, hugepte);
474 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
478 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
479 unsigned long addr, unsigned long end,
480 unsigned long floor, unsigned long ceiling)
488 pmd = pmd_offset(pud, addr);
489 next = pmd_addr_end(addr, end);
490 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
492 * if it is not hugepd pointer, we should already find
495 WARN_ON(!pmd_none_or_clear_bad(pmd));
498 #ifdef CONFIG_PPC_FSL_BOOK3E
500 * Increment next by the size of the huge mapping since
501 * there may be more than one entry at this level for a
502 * single hugepage, but all of them point to
503 * the same kmem cache that holds the hugepte.
505 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
507 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
508 addr, next, floor, ceiling);
509 } while (addr = next, addr != end);
519 if (end - 1 > ceiling - 1)
522 pmd = pmd_offset(pud, start);
524 pmd_free_tlb(tlb, pmd, start);
525 mm_dec_nr_pmds(tlb->mm);
528 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
529 unsigned long addr, unsigned long end,
530 unsigned long floor, unsigned long ceiling)
538 pud = pud_offset(pgd, addr);
539 next = pud_addr_end(addr, end);
540 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
541 if (pud_none_or_clear_bad(pud))
543 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
546 #ifdef CONFIG_PPC_FSL_BOOK3E
548 * Increment next by the size of the huge mapping since
549 * there may be more than one entry at this level for a
550 * single hugepage, but all of them point to
551 * the same kmem cache that holds the hugepte.
553 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
555 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
556 addr, next, floor, ceiling);
558 } while (addr = next, addr != end);
564 ceiling &= PGDIR_MASK;
568 if (end - 1 > ceiling - 1)
571 pud = pud_offset(pgd, start);
573 pud_free_tlb(tlb, pud, start);
577 * This function frees user-level page tables of a process.
579 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
580 unsigned long addr, unsigned long end,
581 unsigned long floor, unsigned long ceiling)
587 * Because there are a number of different possible pagetable
588 * layouts for hugepage ranges, we limit knowledge of how
589 * things should be laid out to the allocation path
590 * (huge_pte_alloc(), above). Everything else works out the
591 * structure as it goes from information in the hugepd
592 * pointers. That means that we can't here use the
593 * optimization used in the normal page free_pgd_range(), of
594 * checking whether we're actually covering a large enough
595 * range to have to do anything at the top level of the walk
596 * instead of at the bottom.
598 * To make sense of this, you should probably go read the big
599 * block comment at the top of the normal free_pgd_range(),
604 next = pgd_addr_end(addr, end);
605 pgd = pgd_offset(tlb->mm, addr);
606 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
607 if (pgd_none_or_clear_bad(pgd))
609 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
611 #ifdef CONFIG_PPC_FSL_BOOK3E
613 * Increment next by the size of the huge mapping since
614 * there may be more than one entry at the pgd level
615 * for a single hugepage, but all of them point to the
616 * same kmem cache that holds the hugepte.
618 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
620 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
621 addr, next, floor, ceiling);
623 } while (addr = next, addr != end);
627 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
628 * To prevent hugepage split, disable irq.
631 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
636 unsigned long mask, flags;
637 struct page *page = ERR_PTR(-EINVAL);
639 local_irq_save(flags);
640 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
643 pte = READ_ONCE(*ptep);
645 * Verify it is a huge page else bail.
646 * Transparent hugepages are handled by generic code. We can skip them
649 if (!shift || is_thp)
652 if (!pte_present(pte)) {
656 mask = (1UL << shift) - 1;
657 page = pte_page(pte);
659 page += (address & mask) / PAGE_SIZE;
662 local_irq_restore(flags);
667 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
668 pmd_t *pmd, int write)
675 follow_huge_pud(struct mm_struct *mm, unsigned long address,
676 pud_t *pud, int write)
682 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
685 unsigned long __boundary = (addr + sz) & ~(sz-1);
686 return (__boundary - 1 < end - 1) ? __boundary : end;
689 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
690 unsigned long end, int write, struct page **pages, int *nr)
693 unsigned long sz = 1UL << hugepd_shift(hugepd);
696 ptep = hugepte_offset(hugepd, addr, pdshift);
698 next = hugepte_addr_end(addr, end, sz);
699 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
701 } while (ptep++, addr = next, addr != end);
706 #ifdef CONFIG_PPC_MM_SLICES
707 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
708 unsigned long len, unsigned long pgoff,
711 struct hstate *hstate = hstate_file(file);
712 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
714 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
718 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
720 #ifdef CONFIG_PPC_MM_SLICES
721 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
723 return 1UL << mmu_psize_to_shift(psize);
725 if (!is_vm_hugetlb_page(vma))
728 return huge_page_size(hstate_vma(vma));
732 static inline bool is_power_of_4(unsigned long x)
734 if (is_power_of_2(x))
735 return (__ilog2(x) % 2) ? false : true;
739 static int __init add_huge_page_size(unsigned long long size)
741 int shift = __ffs(size);
744 /* Check that it is a page size supported by the hardware and
745 * that it fits within pagetable and slice limits. */
746 #ifdef CONFIG_PPC_FSL_BOOK3E
747 if ((size < PAGE_SIZE) || !is_power_of_4(size))
750 if (!is_power_of_2(size)
751 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
755 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
758 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
760 /* Return if huge page size has already been setup */
761 if (size_to_hstate(size))
764 hugetlb_add_hstate(shift - PAGE_SHIFT);
769 static int __init hugepage_setup_sz(char *str)
771 unsigned long long size;
773 size = memparse(str, &str);
775 if (add_huge_page_size(size) != 0)
776 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
780 __setup("hugepagesz=", hugepage_setup_sz);
782 #ifdef CONFIG_PPC_FSL_BOOK3E
783 struct kmem_cache *hugepte_cache;
784 static int __init hugetlbpage_init(void)
788 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
791 if (!mmu_psize_defs[psize].shift)
794 shift = mmu_psize_to_shift(psize);
796 /* Don't treat normal page sizes as huge... */
797 if (shift != PAGE_SHIFT)
798 if (add_huge_page_size(1ULL << shift) < 0)
803 * Create a kmem cache for hugeptes. The bottom bits in the pte have
804 * size information encoded in them, so align them to allow this
806 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
807 HUGEPD_SHIFT_MASK + 1, 0, NULL);
808 if (hugepte_cache == NULL)
809 panic("%s: Unable to create kmem cache for hugeptes\n",
812 /* Default hpage size = 4M */
813 if (mmu_psize_defs[MMU_PAGE_4M].shift)
814 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
816 panic("%s: Unable to set default huge page size\n", __func__);
822 static int __init hugetlbpage_init(void)
826 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
829 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
833 if (!mmu_psize_defs[psize].shift)
836 shift = mmu_psize_to_shift(psize);
838 if (add_huge_page_size(1ULL << shift) < 0)
841 if (shift < PMD_SHIFT)
843 else if (shift < PUD_SHIFT)
846 pdshift = PGDIR_SHIFT;
848 * if we have pdshift and shift value same, we don't
849 * use pgt cache for hugepd.
851 if (pdshift != shift) {
852 pgtable_cache_add(pdshift - shift, NULL);
853 if (!PGT_CACHE(pdshift - shift))
854 panic("hugetlbpage_init(): could not create "
855 "pgtable cache for %d bit pagesize\n", shift);
859 /* Set default large page size. Currently, we pick 16M or 1M
860 * depending on what is available
862 if (mmu_psize_defs[MMU_PAGE_16M].shift)
863 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
864 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
865 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
870 arch_initcall(hugetlbpage_init);
872 void flush_dcache_icache_hugepage(struct page *page)
877 BUG_ON(!PageCompound(page));
879 for (i = 0; i < (1UL << compound_order(page)); i++) {
880 if (!PageHighMem(page)) {
881 __flush_dcache_icache(page_address(page+i));
883 start = kmap_atomic(page+i);
884 __flush_dcache_icache(start);
885 kunmap_atomic(start);
890 #endif /* CONFIG_HUGETLB_PAGE */
893 * We have 4 cases for pgds and pmds:
894 * (1) invalid (all zeroes)
895 * (2) pointer to next table, as normal; bottom 6 bits == 0
896 * (3) leaf pte for huge page _PAGE_PTE set
897 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
899 * So long as we atomically load page table pointers we are safe against teardown,
900 * we can follow the address down to the the page and take a ref on it.
901 * This function need to be called with interrupts disabled. We use this variant
902 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
905 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
906 bool *is_thp, unsigned *shift)
912 hugepd_t *hpdp = NULL;
913 unsigned pdshift = PGDIR_SHIFT;
921 pgdp = pgdir + pgd_index(ea);
922 pgd = READ_ONCE(*pgdp);
924 * Always operate on the local stack value. This make sure the
925 * value don't get updated by a parallel THP split/collapse,
926 * page fault or a page unmap. The return pte_t * is still not
927 * stable. So should be checked there for above conditions.
931 else if (pgd_huge(pgd)) {
932 ret_pte = (pte_t *) pgdp;
934 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
935 hpdp = (hugepd_t *)&pgd;
938 * Even if we end up with an unmap, the pgtable will not
939 * be freed, because we do an rcu free and here we are
943 pudp = pud_offset(&pgd, ea);
944 pud = READ_ONCE(*pudp);
948 else if (pud_huge(pud)) {
949 ret_pte = (pte_t *) pudp;
951 } else if (is_hugepd(__hugepd(pud_val(pud))))
952 hpdp = (hugepd_t *)&pud;
955 pmdp = pmd_offset(&pud, ea);
956 pmd = READ_ONCE(*pmdp);
958 * A hugepage collapse is captured by pmd_none, because
959 * it mark the pmd none and do a hpte invalidate.
964 if (pmd_trans_huge(pmd)) {
967 ret_pte = (pte_t *) pmdp;
972 ret_pte = (pte_t *) pmdp;
974 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
975 hpdp = (hugepd_t *)&pmd;
977 return pte_offset_kernel(&pmd, ea);
983 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
984 pdshift = hugepd_shift(*hpdp);
990 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
992 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
993 unsigned long end, int write, struct page **pages, int *nr)
996 unsigned long pte_end;
997 struct page *head, *page;
1001 pte_end = (addr + sz) & ~(sz-1);
1005 pte = READ_ONCE(*ptep);
1006 mask = _PAGE_PRESENT | _PAGE_USER;
1010 if ((pte_val(pte) & mask) != mask)
1013 /* hugepages are never "special" */
1014 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1017 head = pte_page(pte);
1019 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1021 VM_BUG_ON(compound_head(page) != head);
1026 } while (addr += PAGE_SIZE, addr != end);
1028 if (!page_cache_add_speculative(head, refs)) {
1033 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1034 /* Could be optimized better */