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 <asm/pgtable.h>
23 #include <asm/pgalloc.h>
25 #include <asm/setup.h>
26 #include <asm/hugetlb.h>
27 #include <asm/pte-walk.h>
30 #ifdef CONFIG_HUGETLB_PAGE
32 #define PAGE_SHIFT_64K 16
33 #define PAGE_SHIFT_512K 19
34 #define PAGE_SHIFT_8M 23
35 #define PAGE_SHIFT_16M 24
36 #define PAGE_SHIFT_16G 34
38 unsigned int HPAGE_SHIFT;
39 EXPORT_SYMBOL(HPAGE_SHIFT);
41 #define hugepd_none(hpd) (hpd_val(hpd) == 0)
43 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
46 * Only called for hugetlbfs pages, hence can ignore THP and the
49 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
52 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
53 unsigned long address, unsigned pdshift, unsigned pshift)
55 struct kmem_cache *cachep;
60 if (pshift >= pdshift) {
61 cachep = hugepte_cache;
62 num_hugepd = 1 << (pshift - pdshift);
64 cachep = PGT_CACHE(pdshift - pshift);
68 new = kmem_cache_zalloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
70 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
71 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
77 * Make sure other cpus find the hugepd set only after a
78 * properly initialized page table is visible to them.
79 * For more details look for comment in __pte_alloc().
83 spin_lock(&mm->page_table_lock);
86 * We have multiple higher-level entries that point to the same
87 * actual pte location. Fill in each as we go and backtrack on error.
88 * We need all of these so the DTLB pgtable walk code can find the
89 * right higher-level entry without knowing if it's a hugepage or not.
91 for (i = 0; i < num_hugepd; i++, hpdp++) {
92 if (unlikely(!hugepd_none(*hpdp)))
95 #ifdef CONFIG_PPC_BOOK3S_64
96 *hpdp = __hugepd(__pa(new) |
97 (shift_to_mmu_psize(pshift) << 2));
98 #elif defined(CONFIG_PPC_8xx)
99 *hpdp = __hugepd(__pa(new) | _PMD_USER |
100 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
101 _PMD_PAGE_512K) | _PMD_PRESENT);
103 /* We use the old format for PPC_FSL_BOOK3E */
104 *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
108 /* If we bailed from the for loop early, an error occurred, clean up */
109 if (i < num_hugepd) {
110 for (i = i - 1 ; i >= 0; i--, hpdp--)
112 kmem_cache_free(cachep, new);
114 spin_unlock(&mm->page_table_lock);
119 * These macros define how to determine which level of the page table holds
122 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
123 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
124 #define HUGEPD_PUD_SHIFT PUD_SHIFT
128 * At this point we do the placement change only for BOOK3S 64. This would
129 * possibly work on other subarchs.
131 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
136 hugepd_t *hpdp = NULL;
137 unsigned pshift = __ffs(sz);
138 unsigned pdshift = PGDIR_SHIFT;
141 pg = pgd_offset(mm, addr);
143 #ifdef CONFIG_PPC_BOOK3S_64
144 if (pshift == PGDIR_SHIFT)
147 else if (pshift > PUD_SHIFT)
149 * We need to use hugepd table
151 hpdp = (hugepd_t *)pg;
154 pu = pud_alloc(mm, pg, addr);
155 if (pshift == PUD_SHIFT)
157 else if (pshift > PMD_SHIFT)
158 hpdp = (hugepd_t *)pu;
161 pm = pmd_alloc(mm, pu, addr);
162 if (pshift == PMD_SHIFT)
166 hpdp = (hugepd_t *)pm;
170 if (pshift >= HUGEPD_PGD_SHIFT) {
171 hpdp = (hugepd_t *)pg;
174 pu = pud_alloc(mm, pg, addr);
175 if (pshift >= HUGEPD_PUD_SHIFT) {
176 hpdp = (hugepd_t *)pu;
179 pm = pmd_alloc(mm, pu, addr);
180 hpdp = (hugepd_t *)pm;
187 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
189 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
192 return hugepte_offset(*hpdp, addr, pdshift);
195 #ifdef CONFIG_PPC_BOOK3S_64
197 * Tracks gpages after the device tree is scanned and before the
198 * huge_boot_pages list is ready on pseries.
200 #define MAX_NUMBER_GPAGES 1024
201 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
202 __initdata static unsigned nr_gpages;
205 * Build list of addresses of gigantic pages. This function is used in early
206 * boot before the buddy allocator is setup.
208 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
212 while (number_of_pages > 0) {
213 gpage_freearray[nr_gpages] = addr;
220 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
222 struct huge_bootmem_page *m;
225 m = phys_to_virt(gpage_freearray[--nr_gpages]);
226 gpage_freearray[nr_gpages] = 0;
227 list_add(&m->list, &huge_boot_pages);
234 int __init alloc_bootmem_huge_page(struct hstate *h)
237 #ifdef CONFIG_PPC_BOOK3S_64
238 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
239 return pseries_alloc_bootmem_huge_page(h);
241 return __alloc_bootmem_huge_page(h);
244 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
245 #define HUGEPD_FREELIST_SIZE \
246 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
248 struct hugepd_freelist {
254 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
256 static void hugepd_free_rcu_callback(struct rcu_head *head)
258 struct hugepd_freelist *batch =
259 container_of(head, struct hugepd_freelist, rcu);
262 for (i = 0; i < batch->index; i++)
263 kmem_cache_free(hugepte_cache, batch->ptes[i]);
265 free_page((unsigned long)batch);
268 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
270 struct hugepd_freelist **batchp;
272 batchp = &get_cpu_var(hugepd_freelist_cur);
274 if (atomic_read(&tlb->mm->mm_users) < 2 ||
275 mm_is_thread_local(tlb->mm)) {
276 kmem_cache_free(hugepte_cache, hugepte);
277 put_cpu_var(hugepd_freelist_cur);
281 if (*batchp == NULL) {
282 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
283 (*batchp)->index = 0;
286 (*batchp)->ptes[(*batchp)->index++] = hugepte;
287 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
288 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
291 put_cpu_var(hugepd_freelist_cur);
294 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
297 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
298 unsigned long start, unsigned long end,
299 unsigned long floor, unsigned long ceiling)
301 pte_t *hugepte = hugepd_page(*hpdp);
304 unsigned long pdmask = ~((1UL << pdshift) - 1);
305 unsigned int num_hugepd = 1;
306 unsigned int shift = hugepd_shift(*hpdp);
308 /* Note: On fsl the hpdp may be the first of several */
310 num_hugepd = 1 << (shift - pdshift);
320 if (end - 1 > ceiling - 1)
323 for (i = 0; i < num_hugepd; i++, hpdp++)
326 if (shift >= pdshift)
327 hugepd_free(tlb, hugepte);
329 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
332 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
333 unsigned long addr, unsigned long end,
334 unsigned long floor, unsigned long ceiling)
344 pmd = pmd_offset(pud, addr);
345 next = pmd_addr_end(addr, end);
346 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
348 * if it is not hugepd pointer, we should already find
351 WARN_ON(!pmd_none_or_clear_bad(pmd));
355 * Increment next by the size of the huge mapping since
356 * there may be more than one entry at this level for a
357 * single hugepage, but all of them point to
358 * the same kmem cache that holds the hugepte.
360 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
364 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
365 addr, next, floor, ceiling);
366 } while (addr = next, addr != end);
376 if (end - 1 > ceiling - 1)
379 pmd = pmd_offset(pud, start);
381 pmd_free_tlb(tlb, pmd, start);
382 mm_dec_nr_pmds(tlb->mm);
385 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
386 unsigned long addr, unsigned long end,
387 unsigned long floor, unsigned long ceiling)
395 pud = pud_offset(pgd, addr);
396 next = pud_addr_end(addr, end);
397 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
398 if (pud_none_or_clear_bad(pud))
400 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
405 * Increment next by the size of the huge mapping since
406 * there may be more than one entry at this level for a
407 * single hugepage, but all of them point to
408 * the same kmem cache that holds the hugepte.
410 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
414 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
415 addr, next, floor, ceiling);
417 } while (addr = next, addr != end);
423 ceiling &= PGDIR_MASK;
427 if (end - 1 > ceiling - 1)
430 pud = pud_offset(pgd, start);
432 pud_free_tlb(tlb, pud, start);
433 mm_dec_nr_puds(tlb->mm);
437 * This function frees user-level page tables of a process.
439 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
440 unsigned long addr, unsigned long end,
441 unsigned long floor, unsigned long ceiling)
447 * Because there are a number of different possible pagetable
448 * layouts for hugepage ranges, we limit knowledge of how
449 * things should be laid out to the allocation path
450 * (huge_pte_alloc(), above). Everything else works out the
451 * structure as it goes from information in the hugepd
452 * pointers. That means that we can't here use the
453 * optimization used in the normal page free_pgd_range(), of
454 * checking whether we're actually covering a large enough
455 * range to have to do anything at the top level of the walk
456 * instead of at the bottom.
458 * To make sense of this, you should probably go read the big
459 * block comment at the top of the normal free_pgd_range(),
464 next = pgd_addr_end(addr, end);
465 pgd = pgd_offset(tlb->mm, addr);
466 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
467 if (pgd_none_or_clear_bad(pgd))
469 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
473 * Increment next by the size of the huge mapping since
474 * there may be more than one entry at the pgd level
475 * for a single hugepage, but all of them point to the
476 * same kmem cache that holds the hugepte.
478 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
482 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
483 addr, next, floor, ceiling);
485 } while (addr = next, addr != end);
488 struct page *follow_huge_pd(struct vm_area_struct *vma,
489 unsigned long address, hugepd_t hpd,
490 int flags, int pdshift)
494 struct page *page = NULL;
496 int shift = hugepd_shift(hpd);
497 struct mm_struct *mm = vma->vm_mm;
500 ptl = &mm->page_table_lock;
503 ptep = hugepte_offset(hpd, address, pdshift);
504 if (pte_present(*ptep)) {
505 mask = (1UL << shift) - 1;
506 page = pte_page(*ptep);
507 page += ((address & mask) >> PAGE_SHIFT);
508 if (flags & FOLL_GET)
511 if (is_hugetlb_entry_migration(*ptep)) {
513 __migration_entry_wait(mm, ptep, ptl);
521 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
524 unsigned long __boundary = (addr + sz) & ~(sz-1);
525 return (__boundary - 1 < end - 1) ? __boundary : end;
528 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
529 unsigned long end, int write, struct page **pages, int *nr)
532 unsigned long sz = 1UL << hugepd_shift(hugepd);
535 ptep = hugepte_offset(hugepd, addr, pdshift);
537 next = hugepte_addr_end(addr, end, sz);
538 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
540 } while (ptep++, addr = next, addr != end);
545 #ifdef CONFIG_PPC_MM_SLICES
546 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
547 unsigned long len, unsigned long pgoff,
550 struct hstate *hstate = hstate_file(file);
551 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
553 #ifdef CONFIG_PPC_RADIX_MMU
555 return radix__hugetlb_get_unmapped_area(file, addr, len,
558 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
562 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
564 #ifdef CONFIG_PPC_MM_SLICES
565 /* With radix we don't use slice, so derive it from vma*/
566 if (!radix_enabled()) {
567 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
569 return 1UL << mmu_psize_to_shift(psize);
572 if (!is_vm_hugetlb_page(vma))
575 return huge_page_size(hstate_vma(vma));
578 static inline bool is_power_of_4(unsigned long x)
580 if (is_power_of_2(x))
581 return (__ilog2(x) % 2) ? false : true;
585 static int __init add_huge_page_size(unsigned long long size)
587 int shift = __ffs(size);
590 /* Check that it is a page size supported by the hardware and
591 * that it fits within pagetable and slice limits. */
592 if (size <= PAGE_SIZE)
594 #if defined(CONFIG_PPC_FSL_BOOK3E)
595 if (!is_power_of_4(size))
597 #elif !defined(CONFIG_PPC_8xx)
598 if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
602 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
605 #ifdef CONFIG_PPC_BOOK3S_64
607 * We need to make sure that for different page sizes reported by
608 * firmware we only add hugetlb support for page sizes that can be
609 * supported by linux page table layout.
614 if (radix_enabled()) {
615 if (mmu_psize != MMU_PAGE_2M) {
616 if (cpu_has_feature(CPU_FTR_POWER9_DD1) ||
617 (mmu_psize != MMU_PAGE_1G))
621 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
626 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
628 /* Return if huge page size has already been setup */
629 if (size_to_hstate(size))
632 hugetlb_add_hstate(shift - PAGE_SHIFT);
637 static int __init hugepage_setup_sz(char *str)
639 unsigned long long size;
641 size = memparse(str, &str);
643 if (add_huge_page_size(size) != 0) {
645 pr_err("Invalid huge page size specified(%llu)\n", size);
650 __setup("hugepagesz=", hugepage_setup_sz);
652 struct kmem_cache *hugepte_cache;
653 static int __init hugetlbpage_init(void)
657 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
658 if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
661 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
665 if (!mmu_psize_defs[psize].shift)
668 shift = mmu_psize_to_shift(psize);
670 #ifdef CONFIG_PPC_BOOK3S_64
671 if (shift > PGDIR_SHIFT)
673 else if (shift > PUD_SHIFT)
674 pdshift = PGDIR_SHIFT;
675 else if (shift > PMD_SHIFT)
680 if (shift < HUGEPD_PUD_SHIFT)
682 else if (shift < HUGEPD_PGD_SHIFT)
685 pdshift = PGDIR_SHIFT;
688 if (add_huge_page_size(1ULL << shift) < 0)
691 * if we have pdshift and shift value same, we don't
692 * use pgt cache for hugepd.
695 pgtable_cache_add(pdshift - shift, NULL);
696 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
697 else if (!hugepte_cache) {
699 * Create a kmem cache for hugeptes. The bottom bits in
700 * the pte have size information encoded in them, so
701 * align them to allow this
703 hugepte_cache = kmem_cache_create("hugepte-cache",
705 HUGEPD_SHIFT_MASK + 1,
707 if (hugepte_cache == NULL)
708 panic("%s: Unable to create kmem cache "
709 "for hugeptes\n", __func__);
715 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
716 /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
717 if (mmu_psize_defs[MMU_PAGE_4M].shift)
718 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
719 else if (mmu_psize_defs[MMU_PAGE_512K].shift)
720 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
722 /* Set default large page size. Currently, we pick 16M or 1M
723 * depending on what is available
725 if (mmu_psize_defs[MMU_PAGE_16M].shift)
726 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
727 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
728 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
729 else if (mmu_psize_defs[MMU_PAGE_2M].shift)
730 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
735 arch_initcall(hugetlbpage_init);
737 void flush_dcache_icache_hugepage(struct page *page)
742 BUG_ON(!PageCompound(page));
744 for (i = 0; i < (1UL << compound_order(page)); i++) {
745 if (!PageHighMem(page)) {
746 __flush_dcache_icache(page_address(page+i));
748 start = kmap_atomic(page+i);
749 __flush_dcache_icache(start);
750 kunmap_atomic(start);
755 #endif /* CONFIG_HUGETLB_PAGE */
758 * We have 4 cases for pgds and pmds:
759 * (1) invalid (all zeroes)
760 * (2) pointer to next table, as normal; bottom 6 bits == 0
761 * (3) leaf pte for huge page _PAGE_PTE set
762 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
764 * So long as we atomically load page table pointers we are safe against teardown,
765 * we can follow the address down to the the page and take a ref on it.
766 * This function need to be called with interrupts disabled. We use this variant
767 * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
769 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
770 bool *is_thp, unsigned *hpage_shift)
776 hugepd_t *hpdp = NULL;
777 unsigned pdshift = PGDIR_SHIFT;
785 pgdp = pgdir + pgd_index(ea);
786 pgd = READ_ONCE(*pgdp);
788 * Always operate on the local stack value. This make sure the
789 * value don't get updated by a parallel THP split/collapse,
790 * page fault or a page unmap. The return pte_t * is still not
791 * stable. So should be checked there for above conditions.
795 else if (pgd_huge(pgd)) {
796 ret_pte = (pte_t *) pgdp;
798 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
799 hpdp = (hugepd_t *)&pgd;
802 * Even if we end up with an unmap, the pgtable will not
803 * be freed, because we do an rcu free and here we are
807 pudp = pud_offset(&pgd, ea);
808 pud = READ_ONCE(*pudp);
812 else if (pud_huge(pud)) {
813 ret_pte = (pte_t *) pudp;
815 } else if (is_hugepd(__hugepd(pud_val(pud))))
816 hpdp = (hugepd_t *)&pud;
819 pmdp = pmd_offset(&pud, ea);
820 pmd = READ_ONCE(*pmdp);
822 * A hugepage collapse is captured by pmd_none, because
823 * it mark the pmd none and do a hpte invalidate.
828 if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
831 ret_pte = (pte_t *) pmdp;
836 ret_pte = (pte_t *) pmdp;
838 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
839 hpdp = (hugepd_t *)&pmd;
841 return pte_offset_kernel(&pmd, ea);
847 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
848 pdshift = hugepd_shift(*hpdp);
851 *hpage_shift = pdshift;
854 EXPORT_SYMBOL_GPL(__find_linux_pte);
856 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
857 unsigned long end, int write, struct page **pages, int *nr)
859 unsigned long pte_end;
860 struct page *head, *page;
864 pte_end = (addr + sz) & ~(sz-1);
868 pte = READ_ONCE(*ptep);
870 if (!pte_access_permitted(pte, write))
873 /* hugepages are never "special" */
874 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
877 head = pte_page(pte);
879 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
881 VM_BUG_ON(compound_head(page) != head);
886 } while (addr += PAGE_SIZE, addr != end);
888 if (!page_cache_add_speculative(head, refs)) {
893 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
894 /* Could be optimized better */