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 unsigned int HPAGE_SHIFT;
40 EXPORT_SYMBOL(HPAGE_SHIFT);
42 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
44 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
47 * Only called for hugetlbfs pages, hence can ignore THP and the
50 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
53 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
54 unsigned long address, unsigned pdshift, unsigned pshift)
56 struct kmem_cache *cachep;
61 if (pshift >= pdshift) {
62 cachep = hugepte_cache;
63 num_hugepd = 1 << (pshift - pdshift);
65 cachep = PGT_CACHE(pdshift - pshift);
69 new = kmem_cache_zalloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
71 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
72 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
78 * Make sure other cpus find the hugepd set only after a
79 * properly initialized page table is visible to them.
80 * For more details look for comment in __pte_alloc().
84 spin_lock(&mm->page_table_lock);
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 #ifdef CONFIG_PPC_BOOK3S_64
97 *hpdp = __hugepd(__pa(new) |
98 (shift_to_mmu_psize(pshift) << 2));
99 #elif defined(CONFIG_PPC_8xx)
100 *hpdp = __hugepd(__pa(new) |
101 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
102 _PMD_PAGE_512K) | _PMD_PRESENT);
104 /* We use the old format for PPC_FSL_BOOK3E */
105 *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
109 /* If we bailed from the for loop early, an error occurred, clean up */
110 if (i < num_hugepd) {
111 for (i = i - 1 ; i >= 0; i--, hpdp--)
113 kmem_cache_free(cachep, new);
115 kmemleak_ignore(new);
117 spin_unlock(&mm->page_table_lock);
122 * These macros define how to determine which level of the page table holds
125 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
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
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 #ifdef CONFIG_PPC_BOOK3S_64
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;
176 if (pshift >= HUGEPD_PGD_SHIFT) {
177 hpdp = (hugepd_t *)pg;
180 pu = pud_alloc(mm, pg, addr);
181 if (pshift >= HUGEPD_PUD_SHIFT) {
182 hpdp = (hugepd_t *)pu;
185 pm = pmd_alloc(mm, pu, addr);
186 hpdp = (hugepd_t *)pm;
193 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
195 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
198 return hugepte_offset(*hpdp, addr, pdshift);
201 #ifdef CONFIG_PPC_BOOK3S_64
203 * Tracks gpages after the device tree is scanned and before the
204 * huge_boot_pages list is ready on pseries.
206 #define MAX_NUMBER_GPAGES 1024
207 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
208 __initdata static unsigned nr_gpages;
211 * Build list of addresses of gigantic pages. This function is used in early
212 * boot before the buddy allocator is setup.
214 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
218 while (number_of_pages > 0) {
219 gpage_freearray[nr_gpages] = addr;
226 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
228 struct huge_bootmem_page *m;
231 m = phys_to_virt(gpage_freearray[--nr_gpages]);
232 gpage_freearray[nr_gpages] = 0;
233 list_add(&m->list, &huge_boot_pages);
240 int __init alloc_bootmem_huge_page(struct hstate *h)
243 #ifdef CONFIG_PPC_BOOK3S_64
244 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
245 return pseries_alloc_bootmem_huge_page(h);
247 return __alloc_bootmem_huge_page(h);
250 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
251 #define HUGEPD_FREELIST_SIZE \
252 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
254 struct hugepd_freelist {
260 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
262 static void hugepd_free_rcu_callback(struct rcu_head *head)
264 struct hugepd_freelist *batch =
265 container_of(head, struct hugepd_freelist, rcu);
268 for (i = 0; i < batch->index; i++)
269 kmem_cache_free(hugepte_cache, batch->ptes[i]);
271 free_page((unsigned long)batch);
274 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
276 struct hugepd_freelist **batchp;
278 batchp = &get_cpu_var(hugepd_freelist_cur);
280 if (atomic_read(&tlb->mm->mm_users) < 2 ||
281 mm_is_thread_local(tlb->mm)) {
282 kmem_cache_free(hugepte_cache, hugepte);
283 put_cpu_var(hugepd_freelist_cur);
287 if (*batchp == NULL) {
288 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
289 (*batchp)->index = 0;
292 (*batchp)->ptes[(*batchp)->index++] = hugepte;
293 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
294 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
297 put_cpu_var(hugepd_freelist_cur);
300 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
303 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
304 unsigned long start, unsigned long end,
305 unsigned long floor, unsigned long ceiling)
307 pte_t *hugepte = hugepd_page(*hpdp);
310 unsigned long pdmask = ~((1UL << pdshift) - 1);
311 unsigned int num_hugepd = 1;
312 unsigned int shift = hugepd_shift(*hpdp);
314 /* Note: On fsl the hpdp may be the first of several */
316 num_hugepd = 1 << (shift - pdshift);
326 if (end - 1 > ceiling - 1)
329 for (i = 0; i < num_hugepd; i++, hpdp++)
332 if (shift >= pdshift)
333 hugepd_free(tlb, hugepte);
335 pgtable_free_tlb(tlb, hugepte, 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);
442 * This function frees user-level page tables of a process.
444 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
445 unsigned long addr, unsigned long end,
446 unsigned long floor, unsigned long ceiling)
452 * Because there are a number of different possible pagetable
453 * layouts for hugepage ranges, we limit knowledge of how
454 * things should be laid out to the allocation path
455 * (huge_pte_alloc(), above). Everything else works out the
456 * structure as it goes from information in the hugepd
457 * pointers. That means that we can't here use the
458 * optimization used in the normal page free_pgd_range(), of
459 * checking whether we're actually covering a large enough
460 * range to have to do anything at the top level of the walk
461 * instead of at the bottom.
463 * To make sense of this, you should probably go read the big
464 * block comment at the top of the normal free_pgd_range(),
469 next = pgd_addr_end(addr, end);
470 pgd = pgd_offset(tlb->mm, addr);
471 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
472 if (pgd_none_or_clear_bad(pgd))
474 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
478 * Increment next by the size of the huge mapping since
479 * there may be more than one entry at the pgd level
480 * for a single hugepage, but all of them point to the
481 * same kmem cache that holds the hugepte.
483 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
487 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
488 addr, next, floor, ceiling);
490 } while (addr = next, addr != end);
493 struct page *follow_huge_pd(struct vm_area_struct *vma,
494 unsigned long address, hugepd_t hpd,
495 int flags, int pdshift)
499 struct page *page = NULL;
501 int shift = hugepd_shift(hpd);
502 struct mm_struct *mm = vma->vm_mm;
505 ptl = &mm->page_table_lock;
508 ptep = hugepte_offset(hpd, address, pdshift);
509 if (pte_present(*ptep)) {
510 mask = (1UL << shift) - 1;
511 page = pte_page(*ptep);
512 page += ((address & mask) >> PAGE_SHIFT);
513 if (flags & FOLL_GET)
516 if (is_hugetlb_entry_migration(*ptep)) {
518 __migration_entry_wait(mm, ptep, ptl);
526 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
529 unsigned long __boundary = (addr + sz) & ~(sz-1);
530 return (__boundary - 1 < end - 1) ? __boundary : end;
533 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
534 unsigned long end, int write, struct page **pages, int *nr)
537 unsigned long sz = 1UL << hugepd_shift(hugepd);
540 ptep = hugepte_offset(hugepd, addr, pdshift);
542 next = hugepte_addr_end(addr, end, sz);
543 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
545 } while (ptep++, addr = next, addr != end);
550 #ifdef CONFIG_PPC_MM_SLICES
551 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
552 unsigned long len, unsigned long pgoff,
555 struct hstate *hstate = hstate_file(file);
556 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
558 #ifdef CONFIG_PPC_RADIX_MMU
560 return radix__hugetlb_get_unmapped_area(file, addr, len,
563 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
567 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
569 #ifdef CONFIG_PPC_MM_SLICES
570 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
571 /* With radix we don't use slice, so derive it from vma*/
572 if (!radix_enabled())
573 return 1UL << mmu_psize_to_shift(psize);
575 if (!is_vm_hugetlb_page(vma))
578 return huge_page_size(hstate_vma(vma));
581 static inline bool is_power_of_4(unsigned long x)
583 if (is_power_of_2(x))
584 return (__ilog2(x) % 2) ? false : true;
588 static int __init add_huge_page_size(unsigned long long size)
590 int shift = __ffs(size);
593 /* Check that it is a page size supported by the hardware and
594 * that it fits within pagetable and slice limits. */
595 if (size <= PAGE_SIZE)
597 #if defined(CONFIG_PPC_FSL_BOOK3E)
598 if (!is_power_of_4(size))
600 #elif !defined(CONFIG_PPC_8xx)
601 if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
605 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
608 #ifdef CONFIG_PPC_BOOK3S_64
610 * We need to make sure that for different page sizes reported by
611 * firmware we only add hugetlb support for page sizes that can be
612 * supported by linux page table layout.
617 if (radix_enabled()) {
618 if (mmu_psize != MMU_PAGE_2M) {
619 if (cpu_has_feature(CPU_FTR_POWER9_DD1) ||
620 (mmu_psize != MMU_PAGE_1G))
624 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
629 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
631 /* Return if huge page size has already been setup */
632 if (size_to_hstate(size))
635 hugetlb_add_hstate(shift - PAGE_SHIFT);
640 static int __init hugepage_setup_sz(char *str)
642 unsigned long long size;
644 size = memparse(str, &str);
646 if (add_huge_page_size(size) != 0) {
648 pr_err("Invalid huge page size specified(%llu)\n", size);
653 __setup("hugepagesz=", hugepage_setup_sz);
655 struct kmem_cache *hugepte_cache;
656 static int __init hugetlbpage_init(void)
660 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
661 if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
664 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
668 if (!mmu_psize_defs[psize].shift)
671 shift = mmu_psize_to_shift(psize);
673 if (add_huge_page_size(1ULL << shift) < 0)
676 if (shift < HUGEPD_PUD_SHIFT)
678 else if (shift < HUGEPD_PGD_SHIFT)
681 pdshift = PGDIR_SHIFT;
683 * if we have pdshift and shift value same, we don't
684 * use pgt cache for hugepd.
687 pgtable_cache_add(pdshift - shift, NULL);
688 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
689 else if (!hugepte_cache) {
691 * Create a kmem cache for hugeptes. The bottom bits in
692 * the pte have size information encoded in them, so
693 * align them to allow this
695 hugepte_cache = kmem_cache_create("hugepte-cache",
697 HUGEPD_SHIFT_MASK + 1,
699 if (hugepte_cache == NULL)
700 panic("%s: Unable to create kmem cache "
701 "for hugeptes\n", __func__);
707 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
708 /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
709 if (mmu_psize_defs[MMU_PAGE_4M].shift)
710 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
711 else if (mmu_psize_defs[MMU_PAGE_512K].shift)
712 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
714 /* Set default large page size. Currently, we pick 16M or 1M
715 * depending on what is available
717 if (mmu_psize_defs[MMU_PAGE_16M].shift)
718 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
719 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
720 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
721 else if (mmu_psize_defs[MMU_PAGE_2M].shift)
722 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
727 arch_initcall(hugetlbpage_init);
729 void flush_dcache_icache_hugepage(struct page *page)
734 BUG_ON(!PageCompound(page));
736 for (i = 0; i < (1UL << compound_order(page)); i++) {
737 if (!PageHighMem(page)) {
738 __flush_dcache_icache(page_address(page+i));
740 start = kmap_atomic(page+i);
741 __flush_dcache_icache(start);
742 kunmap_atomic(start);
747 #endif /* CONFIG_HUGETLB_PAGE */
750 * We have 4 cases for pgds and pmds:
751 * (1) invalid (all zeroes)
752 * (2) pointer to next table, as normal; bottom 6 bits == 0
753 * (3) leaf pte for huge page _PAGE_PTE set
754 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
756 * So long as we atomically load page table pointers we are safe against teardown,
757 * we can follow the address down to the the page and take a ref on it.
758 * This function need to be called with interrupts disabled. We use this variant
759 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
761 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
762 bool *is_thp, unsigned *hpage_shift)
768 hugepd_t *hpdp = NULL;
769 unsigned pdshift = PGDIR_SHIFT;
777 pgdp = pgdir + pgd_index(ea);
778 pgd = READ_ONCE(*pgdp);
780 * Always operate on the local stack value. This make sure the
781 * value don't get updated by a parallel THP split/collapse,
782 * page fault or a page unmap. The return pte_t * is still not
783 * stable. So should be checked there for above conditions.
787 else if (pgd_huge(pgd)) {
788 ret_pte = (pte_t *) pgdp;
790 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
791 hpdp = (hugepd_t *)&pgd;
794 * Even if we end up with an unmap, the pgtable will not
795 * be freed, because we do an rcu free and here we are
799 pudp = pud_offset(&pgd, ea);
800 pud = READ_ONCE(*pudp);
804 else if (pud_huge(pud)) {
805 ret_pte = (pte_t *) pudp;
807 } else if (is_hugepd(__hugepd(pud_val(pud))))
808 hpdp = (hugepd_t *)&pud;
811 pmdp = pmd_offset(&pud, ea);
812 pmd = READ_ONCE(*pmdp);
814 * A hugepage collapse is captured by pmd_none, because
815 * it mark the pmd none and do a hpte invalidate.
820 if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
823 ret_pte = (pte_t *) pmdp;
828 ret_pte = (pte_t *) pmdp;
830 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
831 hpdp = (hugepd_t *)&pmd;
833 return pte_offset_kernel(&pmd, ea);
839 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
840 pdshift = hugepd_shift(*hpdp);
843 *hpage_shift = pdshift;
846 EXPORT_SYMBOL_GPL(__find_linux_pte);
848 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
849 unsigned long end, int write, struct page **pages, int *nr)
851 unsigned long pte_end;
852 struct page *head, *page;
856 pte_end = (addr + sz) & ~(sz-1);
860 pte = READ_ONCE(*ptep);
862 if (!pte_present(pte) || !pte_read(pte))
864 if (write && !pte_write(pte))
867 /* hugepages are never "special" */
868 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
871 head = pte_page(pte);
873 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
875 VM_BUG_ON(compound_head(page) != head);
880 } while (addr += PAGE_SIZE, addr != end);
882 if (!page_cache_add_speculative(head, refs)) {
887 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
888 /* Could be optimized better */