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 <linux/swap.h>
21 #include <linux/swapops.h>
22 #include <linux/kmemleak.h>
23 #include <asm/pgtable.h>
24 #include <asm/pgalloc.h>
26 #include <asm/setup.h>
27 #include <asm/hugetlb.h>
28 #include <asm/pte-walk.h>
31 #ifdef CONFIG_HUGETLB_PAGE
33 #define PAGE_SHIFT_64K 16
34 #define PAGE_SHIFT_512K 19
35 #define PAGE_SHIFT_8M 23
36 #define PAGE_SHIFT_16M 24
37 #define PAGE_SHIFT_16G 34
39 bool hugetlb_disabled = false;
41 unsigned int HPAGE_SHIFT;
42 EXPORT_SYMBOL(HPAGE_SHIFT);
44 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
46 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
49 * Only called for hugetlbfs pages, hence can ignore THP and the
52 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
55 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
56 unsigned long address, unsigned int pdshift,
57 unsigned int pshift, spinlock_t *ptl)
59 struct kmem_cache *cachep;
64 if (pshift >= pdshift) {
65 cachep = hugepte_cache;
66 num_hugepd = 1 << (pshift - pdshift);
68 cachep = PGT_CACHE(pdshift - pshift);
72 new = kmem_cache_zalloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
74 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
75 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
81 * Make sure other cpus find the hugepd set only after a
82 * properly initialized page table is visible to them.
83 * For more details look for comment in __pte_alloc().
89 * We have multiple higher-level entries that point to the same
90 * actual pte location. Fill in each as we go and backtrack on error.
91 * We need all of these so the DTLB pgtable walk code can find the
92 * right higher-level entry without knowing if it's a hugepage or not.
94 for (i = 0; i < num_hugepd; i++, hpdp++) {
95 if (unlikely(!hugepd_none(*hpdp)))
98 #ifdef CONFIG_PPC_BOOK3S_64
99 *hpdp = __hugepd(__pa(new) | HUGEPD_VAL_BITS |
100 (shift_to_mmu_psize(pshift) << 2));
101 #elif defined(CONFIG_PPC_8xx)
102 *hpdp = __hugepd(__pa(new) | _PMD_USER |
103 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
104 _PMD_PAGE_512K) | _PMD_PRESENT);
106 /* We use the old format for PPC_FSL_BOOK3E */
107 *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
111 /* If we bailed from the for loop early, an error occurred, clean up */
112 if (i < num_hugepd) {
113 for (i = i - 1 ; i >= 0; i--, hpdp--)
115 kmem_cache_free(cachep, new);
117 kmemleak_ignore(new);
124 * At this point we do the placement change only for BOOK3S 64. This would
125 * possibly work on other subarchs.
127 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
132 hugepd_t *hpdp = NULL;
133 unsigned pshift = __ffs(sz);
134 unsigned pdshift = PGDIR_SHIFT;
138 pg = pgd_offset(mm, addr);
140 #ifdef CONFIG_PPC_BOOK3S_64
141 if (pshift == PGDIR_SHIFT)
144 else if (pshift > PUD_SHIFT) {
146 * We need to use hugepd table
148 ptl = &mm->page_table_lock;
149 hpdp = (hugepd_t *)pg;
152 pu = pud_alloc(mm, pg, addr);
153 if (pshift == PUD_SHIFT)
155 else if (pshift > PMD_SHIFT) {
156 ptl = pud_lockptr(mm, pu);
157 hpdp = (hugepd_t *)pu;
160 pm = pmd_alloc(mm, pu, addr);
161 if (pshift == PMD_SHIFT)
165 ptl = pmd_lockptr(mm, pm);
166 hpdp = (hugepd_t *)pm;
171 if (pshift >= PGDIR_SHIFT) {
172 ptl = &mm->page_table_lock;
173 hpdp = (hugepd_t *)pg;
176 pu = pud_alloc(mm, pg, addr);
177 if (pshift >= PUD_SHIFT) {
178 ptl = pud_lockptr(mm, pu);
179 hpdp = (hugepd_t *)pu;
182 pm = pmd_alloc(mm, pu, addr);
183 ptl = pmd_lockptr(mm, pm);
184 hpdp = (hugepd_t *)pm;
191 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
193 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
194 pdshift, pshift, ptl))
197 return hugepte_offset(*hpdp, addr, pdshift);
200 #ifdef CONFIG_PPC_BOOK3S_64
202 * Tracks gpages after the device tree is scanned and before the
203 * huge_boot_pages list is ready on pseries.
205 #define MAX_NUMBER_GPAGES 1024
206 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
207 __initdata static unsigned nr_gpages;
210 * Build list of addresses of gigantic pages. This function is used in early
211 * boot before the buddy allocator is setup.
213 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
217 while (number_of_pages > 0) {
218 gpage_freearray[nr_gpages] = addr;
225 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
227 struct huge_bootmem_page *m;
230 m = phys_to_virt(gpage_freearray[--nr_gpages]);
231 gpage_freearray[nr_gpages] = 0;
232 list_add(&m->list, &huge_boot_pages);
239 int __init alloc_bootmem_huge_page(struct hstate *h)
242 #ifdef CONFIG_PPC_BOOK3S_64
243 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
244 return pseries_alloc_bootmem_huge_page(h);
246 return __alloc_bootmem_huge_page(h);
249 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
250 #define HUGEPD_FREELIST_SIZE \
251 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
253 struct hugepd_freelist {
259 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
261 static void hugepd_free_rcu_callback(struct rcu_head *head)
263 struct hugepd_freelist *batch =
264 container_of(head, struct hugepd_freelist, rcu);
267 for (i = 0; i < batch->index; i++)
268 kmem_cache_free(hugepte_cache, batch->ptes[i]);
270 free_page((unsigned long)batch);
273 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
275 struct hugepd_freelist **batchp;
277 batchp = &get_cpu_var(hugepd_freelist_cur);
279 if (atomic_read(&tlb->mm->mm_users) < 2 ||
280 mm_is_thread_local(tlb->mm)) {
281 kmem_cache_free(hugepte_cache, hugepte);
282 put_cpu_var(hugepd_freelist_cur);
286 if (*batchp == NULL) {
287 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
288 (*batchp)->index = 0;
291 (*batchp)->ptes[(*batchp)->index++] = hugepte;
292 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
293 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
296 put_cpu_var(hugepd_freelist_cur);
299 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
302 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
303 unsigned long start, unsigned long end,
304 unsigned long floor, unsigned long ceiling)
306 pte_t *hugepte = hugepd_page(*hpdp);
309 unsigned long pdmask = ~((1UL << pdshift) - 1);
310 unsigned int num_hugepd = 1;
311 unsigned int shift = hugepd_shift(*hpdp);
313 /* Note: On fsl the hpdp may be the first of several */
315 num_hugepd = 1 << (shift - pdshift);
325 if (end - 1 > ceiling - 1)
328 for (i = 0; i < num_hugepd; i++, hpdp++)
331 if (shift >= pdshift)
332 hugepd_free(tlb, hugepte);
334 pgtable_free_tlb(tlb, hugepte,
335 get_hugepd_cache_index(pdshift - shift));
338 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
339 unsigned long addr, unsigned long end,
340 unsigned long floor, unsigned long ceiling)
350 pmd = pmd_offset(pud, addr);
351 next = pmd_addr_end(addr, end);
352 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
354 * if it is not hugepd pointer, we should already find
357 WARN_ON(!pmd_none_or_clear_bad(pmd));
361 * Increment next by the size of the huge mapping since
362 * there may be more than one entry at this level for a
363 * single hugepage, but all of them point to
364 * the same kmem cache that holds the hugepte.
366 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
370 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
371 addr, next, floor, ceiling);
372 } while (addr = next, addr != end);
382 if (end - 1 > ceiling - 1)
385 pmd = pmd_offset(pud, start);
387 pmd_free_tlb(tlb, pmd, start);
388 mm_dec_nr_pmds(tlb->mm);
391 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
392 unsigned long addr, unsigned long end,
393 unsigned long floor, unsigned long ceiling)
401 pud = pud_offset(pgd, addr);
402 next = pud_addr_end(addr, end);
403 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
404 if (pud_none_or_clear_bad(pud))
406 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
411 * Increment next by the size of the huge mapping since
412 * there may be more than one entry at this level for a
413 * single hugepage, but all of them point to
414 * the same kmem cache that holds the hugepte.
416 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
420 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
421 addr, next, floor, ceiling);
423 } while (addr = next, addr != end);
429 ceiling &= PGDIR_MASK;
433 if (end - 1 > ceiling - 1)
436 pud = pud_offset(pgd, start);
438 pud_free_tlb(tlb, pud, start);
439 mm_dec_nr_puds(tlb->mm);
443 * This function frees user-level page tables of a process.
445 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
446 unsigned long addr, unsigned long end,
447 unsigned long floor, unsigned long ceiling)
453 * Because there are a number of different possible pagetable
454 * layouts for hugepage ranges, we limit knowledge of how
455 * things should be laid out to the allocation path
456 * (huge_pte_alloc(), above). Everything else works out the
457 * structure as it goes from information in the hugepd
458 * pointers. That means that we can't here use the
459 * optimization used in the normal page free_pgd_range(), of
460 * checking whether we're actually covering a large enough
461 * range to have to do anything at the top level of the walk
462 * instead of at the bottom.
464 * To make sense of this, you should probably go read the big
465 * block comment at the top of the normal free_pgd_range(),
470 next = pgd_addr_end(addr, end);
471 pgd = pgd_offset(tlb->mm, addr);
472 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
473 if (pgd_none_or_clear_bad(pgd))
475 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
479 * Increment next by the size of the huge mapping since
480 * there may be more than one entry at the pgd level
481 * for a single hugepage, but all of them point to the
482 * same kmem cache that holds the hugepte.
484 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
488 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
489 addr, next, floor, ceiling);
491 } while (addr = next, addr != end);
494 struct page *follow_huge_pd(struct vm_area_struct *vma,
495 unsigned long address, hugepd_t hpd,
496 int flags, int pdshift)
500 struct page *page = NULL;
502 int shift = hugepd_shift(hpd);
503 struct mm_struct *mm = vma->vm_mm;
507 * hugepage directory entries are protected by mm->page_table_lock
508 * Use this instead of huge_pte_lockptr
510 ptl = &mm->page_table_lock;
513 ptep = hugepte_offset(hpd, address, pdshift);
514 if (pte_present(*ptep)) {
515 mask = (1UL << shift) - 1;
516 page = pte_page(*ptep);
517 page += ((address & mask) >> PAGE_SHIFT);
518 if (flags & FOLL_GET)
521 if (is_hugetlb_entry_migration(*ptep)) {
523 __migration_entry_wait(mm, ptep, ptl);
531 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
534 unsigned long __boundary = (addr + sz) & ~(sz-1);
535 return (__boundary - 1 < end - 1) ? __boundary : end;
538 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
539 unsigned long end, int write, struct page **pages, int *nr)
542 unsigned long sz = 1UL << hugepd_shift(hugepd);
545 ptep = hugepte_offset(hugepd, addr, pdshift);
547 next = hugepte_addr_end(addr, end, sz);
548 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
550 } while (ptep++, addr = next, addr != end);
555 #ifdef CONFIG_PPC_MM_SLICES
556 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
557 unsigned long len, unsigned long pgoff,
560 struct hstate *hstate = hstate_file(file);
561 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
563 #ifdef CONFIG_PPC_RADIX_MMU
565 return radix__hugetlb_get_unmapped_area(file, addr, len,
568 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
572 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
574 #ifdef CONFIG_PPC_MM_SLICES
575 /* With radix we don't use slice, so derive it from vma*/
576 if (!radix_enabled()) {
577 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
579 return 1UL << mmu_psize_to_shift(psize);
582 return vma_kernel_pagesize(vma);
585 static inline bool is_power_of_4(unsigned long x)
587 if (is_power_of_2(x))
588 return (__ilog2(x) % 2) ? false : true;
592 static int __init add_huge_page_size(unsigned long long size)
594 int shift = __ffs(size);
597 /* Check that it is a page size supported by the hardware and
598 * that it fits within pagetable and slice limits. */
599 if (size <= PAGE_SIZE)
601 #if defined(CONFIG_PPC_FSL_BOOK3E)
602 if (!is_power_of_4(size))
604 #elif !defined(CONFIG_PPC_8xx)
605 if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
609 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
612 #ifdef CONFIG_PPC_BOOK3S_64
614 * We need to make sure that for different page sizes reported by
615 * firmware we only add hugetlb support for page sizes that can be
616 * supported by linux page table layout.
621 if (radix_enabled()) {
622 if (mmu_psize != MMU_PAGE_2M && mmu_psize != MMU_PAGE_1G)
625 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
630 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
632 /* Return if huge page size has already been setup */
633 if (size_to_hstate(size))
636 hugetlb_add_hstate(shift - PAGE_SHIFT);
641 static int __init hugepage_setup_sz(char *str)
643 unsigned long long size;
645 size = memparse(str, &str);
647 if (add_huge_page_size(size) != 0) {
649 pr_err("Invalid huge page size specified(%llu)\n", size);
654 __setup("hugepagesz=", hugepage_setup_sz);
656 struct kmem_cache *hugepte_cache;
657 static int __init hugetlbpage_init(void)
661 if (hugetlb_disabled) {
662 pr_info("HugeTLB support is disabled!\n");
666 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
667 if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
670 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
674 if (!mmu_psize_defs[psize].shift)
677 shift = mmu_psize_to_shift(psize);
679 #ifdef CONFIG_PPC_BOOK3S_64
680 if (shift > PGDIR_SHIFT)
682 else if (shift > PUD_SHIFT)
683 pdshift = PGDIR_SHIFT;
684 else if (shift > PMD_SHIFT)
689 if (shift < PUD_SHIFT)
691 else if (shift < PGDIR_SHIFT)
694 pdshift = PGDIR_SHIFT;
697 if (add_huge_page_size(1ULL << shift) < 0)
700 * if we have pdshift and shift value same, we don't
701 * use pgt cache for hugepd.
704 pgtable_cache_add(pdshift - shift, NULL);
705 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
706 else if (!hugepte_cache) {
708 * Create a kmem cache for hugeptes. The bottom bits in
709 * the pte have size information encoded in them, so
710 * align them to allow this
712 hugepte_cache = kmem_cache_create("hugepte-cache",
714 HUGEPD_SHIFT_MASK + 1,
716 if (hugepte_cache == NULL)
717 panic("%s: Unable to create kmem cache "
718 "for hugeptes\n", __func__);
724 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
725 /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
726 if (mmu_psize_defs[MMU_PAGE_4M].shift)
727 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
728 else if (mmu_psize_defs[MMU_PAGE_512K].shift)
729 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
731 /* Set default large page size. Currently, we pick 16M or 1M
732 * depending on what is available
734 if (mmu_psize_defs[MMU_PAGE_16M].shift)
735 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
736 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
737 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
738 else if (mmu_psize_defs[MMU_PAGE_2M].shift)
739 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
744 arch_initcall(hugetlbpage_init);
746 void flush_dcache_icache_hugepage(struct page *page)
751 BUG_ON(!PageCompound(page));
753 for (i = 0; i < (1UL << compound_order(page)); i++) {
754 if (!PageHighMem(page)) {
755 __flush_dcache_icache(page_address(page+i));
757 start = kmap_atomic(page+i);
758 __flush_dcache_icache(start);
759 kunmap_atomic(start);
764 #endif /* CONFIG_HUGETLB_PAGE */
767 * We have 4 cases for pgds and pmds:
768 * (1) invalid (all zeroes)
769 * (2) pointer to next table, as normal; bottom 6 bits == 0
770 * (3) leaf pte for huge page _PAGE_PTE set
771 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
773 * So long as we atomically load page table pointers we are safe against teardown,
774 * we can follow the address down to the the page and take a ref on it.
775 * This function need to be called with interrupts disabled. We use this variant
776 * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
778 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
779 bool *is_thp, unsigned *hpage_shift)
785 hugepd_t *hpdp = NULL;
786 unsigned pdshift = PGDIR_SHIFT;
794 pgdp = pgdir + pgd_index(ea);
795 pgd = READ_ONCE(*pgdp);
797 * Always operate on the local stack value. This make sure the
798 * value don't get updated by a parallel THP split/collapse,
799 * page fault or a page unmap. The return pte_t * is still not
800 * stable. So should be checked there for above conditions.
804 else if (pgd_huge(pgd)) {
805 ret_pte = (pte_t *) pgdp;
807 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
808 hpdp = (hugepd_t *)&pgd;
811 * Even if we end up with an unmap, the pgtable will not
812 * be freed, because we do an rcu free and here we are
816 pudp = pud_offset(&pgd, ea);
817 pud = READ_ONCE(*pudp);
821 else if (pud_huge(pud)) {
822 ret_pte = (pte_t *) pudp;
824 } else if (is_hugepd(__hugepd(pud_val(pud))))
825 hpdp = (hugepd_t *)&pud;
828 pmdp = pmd_offset(&pud, ea);
829 pmd = READ_ONCE(*pmdp);
831 * A hugepage collapse is captured by pmd_none, because
832 * it mark the pmd none and do a hpte invalidate.
837 if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
840 ret_pte = (pte_t *) pmdp;
844 * pmd_large check below will handle the swap pmd pte
845 * we need to do both the check because they are config
848 if (pmd_huge(pmd) || pmd_large(pmd)) {
849 ret_pte = (pte_t *) pmdp;
851 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
852 hpdp = (hugepd_t *)&pmd;
854 return pte_offset_kernel(&pmd, ea);
860 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
861 pdshift = hugepd_shift(*hpdp);
864 *hpage_shift = pdshift;
867 EXPORT_SYMBOL_GPL(__find_linux_pte);
869 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
870 unsigned long end, int write, struct page **pages, int *nr)
872 unsigned long pte_end;
873 struct page *head, *page;
877 pte_end = (addr + sz) & ~(sz-1);
881 pte = READ_ONCE(*ptep);
883 if (!pte_access_permitted(pte, write))
886 /* hugepages are never "special" */
887 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
890 head = pte_page(pte);
892 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
894 VM_BUG_ON(compound_head(page) != head);
899 } while (addr += PAGE_SIZE, addr != end);
901 if (!page_cache_add_speculative(head, refs)) {
906 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
907 /* Could be optimized better */