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/of_fdt.h>
16 #include <linux/memblock.h>
17 #include <linux/bootmem.h>
18 #include <linux/moduleparam.h>
19 #include <asm/pgtable.h>
20 #include <asm/pgalloc.h>
22 #include <asm/setup.h>
24 #define PAGE_SHIFT_64K 16
25 #define PAGE_SHIFT_16M 24
26 #define PAGE_SHIFT_16G 34
28 unsigned int HPAGE_SHIFT;
31 * Tracks gpages after the device tree is scanned and before the
32 * huge_boot_pages list is ready. On non-Freescale implementations, this is
33 * just used to track 16G pages and so is a single array. FSL-based
34 * implementations may have more than one gpage size, so we need multiple
37 #ifdef CONFIG_PPC_FSL_BOOK3E
38 #define MAX_NUMBER_GPAGES 128
40 u64 gpage_list[MAX_NUMBER_GPAGES];
41 unsigned int nr_gpages;
43 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
45 #define MAX_NUMBER_GPAGES 1024
46 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
47 static unsigned nr_gpages;
50 static inline int shift_to_mmu_psize(unsigned int shift)
54 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize)
55 if (mmu_psize_defs[psize].shift == shift)
60 static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
62 if (mmu_psize_defs[mmu_psize].shift)
63 return mmu_psize_defs[mmu_psize].shift;
67 #define hugepd_none(hpd) ((hpd).pd == 0)
69 pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
74 hugepd_t *hpdp = NULL;
75 unsigned pdshift = PGDIR_SHIFT;
80 pg = pgdir + pgd_index(ea);
82 hpdp = (hugepd_t *)pg;
83 } else if (!pgd_none(*pg)) {
85 pu = pud_offset(pg, ea);
87 hpdp = (hugepd_t *)pu;
88 else if (!pud_none(*pu)) {
90 pm = pmd_offset(pu, ea);
92 hpdp = (hugepd_t *)pm;
93 else if (!pmd_none(*pm)) {
94 return pte_offset_kernel(pm, ea);
103 *shift = hugepd_shift(*hpdp);
104 return hugepte_offset(hpdp, ea, pdshift);
107 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
109 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
112 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
113 unsigned long address, unsigned pdshift, unsigned pshift)
115 struct kmem_cache *cachep;
118 #ifdef CONFIG_PPC_FSL_BOOK3E
120 int num_hugepd = 1 << (pshift - pdshift);
121 cachep = hugepte_cache;
123 cachep = PGT_CACHE(pdshift - pshift);
126 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
128 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
129 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
134 spin_lock(&mm->page_table_lock);
135 #ifdef CONFIG_PPC_FSL_BOOK3E
137 * We have multiple higher-level entries that point to the same
138 * actual pte location. Fill in each as we go and backtrack on error.
139 * We need all of these so the DTLB pgtable walk code can find the
140 * right higher-level entry without knowing if it's a hugepage or not.
142 for (i = 0; i < num_hugepd; i++, hpdp++) {
143 if (unlikely(!hugepd_none(*hpdp)))
146 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
148 /* If we bailed from the for loop early, an error occurred, clean up */
149 if (i < num_hugepd) {
150 for (i = i - 1 ; i >= 0; i--, hpdp--)
152 kmem_cache_free(cachep, new);
155 if (!hugepd_none(*hpdp))
156 kmem_cache_free(cachep, new);
158 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
160 spin_unlock(&mm->page_table_lock);
165 * These macros define how to determine which level of the page table holds
168 #ifdef CONFIG_PPC_FSL_BOOK3E
169 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
170 #define HUGEPD_PUD_SHIFT PUD_SHIFT
172 #define HUGEPD_PGD_SHIFT PUD_SHIFT
173 #define HUGEPD_PUD_SHIFT PMD_SHIFT
176 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
181 hugepd_t *hpdp = NULL;
182 unsigned pshift = __ffs(sz);
183 unsigned pdshift = PGDIR_SHIFT;
187 pg = pgd_offset(mm, addr);
189 if (pshift >= HUGEPD_PGD_SHIFT) {
190 hpdp = (hugepd_t *)pg;
193 pu = pud_alloc(mm, pg, addr);
194 if (pshift >= HUGEPD_PUD_SHIFT) {
195 hpdp = (hugepd_t *)pu;
198 pm = pmd_alloc(mm, pu, addr);
199 hpdp = (hugepd_t *)pm;
206 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
208 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
211 return hugepte_offset(hpdp, addr, pdshift);
214 #ifdef CONFIG_PPC_FSL_BOOK3E
215 /* Build list of addresses of gigantic pages. This function is used in early
216 * boot before the buddy or bootmem allocator is setup.
218 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
220 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
226 gpage_freearray[idx].nr_gpages = number_of_pages;
228 for (i = 0; i < number_of_pages; i++) {
229 gpage_freearray[idx].gpage_list[i] = addr;
235 * Moves the gigantic page addresses from the temporary list to the
236 * huge_boot_pages list.
238 int alloc_bootmem_huge_page(struct hstate *hstate)
240 struct huge_bootmem_page *m;
241 int idx = shift_to_mmu_psize(hstate->order + PAGE_SHIFT);
242 int nr_gpages = gpage_freearray[idx].nr_gpages;
247 #ifdef CONFIG_HIGHMEM
249 * If gpages can be in highmem we can't use the trick of storing the
250 * data structure in the page; allocate space for this
252 m = alloc_bootmem(sizeof(struct huge_bootmem_page));
253 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
255 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
258 list_add(&m->list, &huge_boot_pages);
259 gpage_freearray[idx].nr_gpages = nr_gpages;
260 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
266 * Scan the command line hugepagesz= options for gigantic pages; store those in
267 * a list that we use to allocate the memory once all options are parsed.
270 unsigned long gpage_npages[MMU_PAGE_COUNT];
272 static int __init do_gpage_early_setup(char *param, char *val)
274 static phys_addr_t size;
275 unsigned long npages;
278 * The hugepagesz and hugepages cmdline options are interleaved. We
279 * use the size variable to keep track of whether or not this was done
280 * properly and skip over instances where it is incorrect. Other
281 * command-line parsing code will issue warnings, so we don't need to.
284 if ((strcmp(param, "default_hugepagesz") == 0) ||
285 (strcmp(param, "hugepagesz") == 0)) {
286 size = memparse(val, NULL);
287 } else if (strcmp(param, "hugepages") == 0) {
289 if (sscanf(val, "%lu", &npages) <= 0)
291 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
300 * This function allocates physical space for pages that are larger than the
301 * buddy allocator can handle. We want to allocate these in highmem because
302 * the amount of lowmem is limited. This means that this function MUST be
303 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
304 * allocate to grab highmem.
306 void __init reserve_hugetlb_gpages(void)
308 static __initdata char cmdline[COMMAND_LINE_SIZE];
309 phys_addr_t size, base;
312 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
313 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
314 &do_gpage_early_setup);
317 * Walk gpage list in reverse, allocating larger page sizes first.
318 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
319 * When we reach the point in the list where pages are no longer
320 * considered gpages, we're done.
322 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
323 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
325 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
328 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
329 base = memblock_alloc_base(size * gpage_npages[i], size,
330 MEMBLOCK_ALLOC_ANYWHERE);
331 add_gpage(base, size, gpage_npages[i]);
335 #else /* !PPC_FSL_BOOK3E */
337 /* Build list of addresses of gigantic pages. This function is used in early
338 * boot before the buddy or bootmem allocator is setup.
340 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
344 while (number_of_pages > 0) {
345 gpage_freearray[nr_gpages] = addr;
352 /* Moves the gigantic page addresses from the temporary list to the
353 * huge_boot_pages list.
355 int alloc_bootmem_huge_page(struct hstate *hstate)
357 struct huge_bootmem_page *m;
360 m = phys_to_virt(gpage_freearray[--nr_gpages]);
361 gpage_freearray[nr_gpages] = 0;
362 list_add(&m->list, &huge_boot_pages);
368 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
373 #ifdef CONFIG_PPC_FSL_BOOK3E
374 #define HUGEPD_FREELIST_SIZE \
375 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
377 struct hugepd_freelist {
383 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
385 static void hugepd_free_rcu_callback(struct rcu_head *head)
387 struct hugepd_freelist *batch =
388 container_of(head, struct hugepd_freelist, rcu);
391 for (i = 0; i < batch->index; i++)
392 kmem_cache_free(hugepte_cache, batch->ptes[i]);
394 free_page((unsigned long)batch);
397 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
399 struct hugepd_freelist **batchp;
401 batchp = &__get_cpu_var(hugepd_freelist_cur);
403 if (atomic_read(&tlb->mm->mm_users) < 2 ||
404 cpumask_equal(mm_cpumask(tlb->mm),
405 cpumask_of(smp_processor_id()))) {
406 kmem_cache_free(hugepte_cache, hugepte);
410 if (*batchp == NULL) {
411 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
412 (*batchp)->index = 0;
415 (*batchp)->ptes[(*batchp)->index++] = hugepte;
416 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
417 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
423 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
424 unsigned long start, unsigned long end,
425 unsigned long floor, unsigned long ceiling)
427 pte_t *hugepte = hugepd_page(*hpdp);
430 unsigned long pdmask = ~((1UL << pdshift) - 1);
431 unsigned int num_hugepd = 1;
433 #ifdef CONFIG_PPC_FSL_BOOK3E
434 /* Note: On fsl the hpdp may be the first of several */
435 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
437 unsigned int shift = hugepd_shift(*hpdp);
448 if (end - 1 > ceiling - 1)
451 for (i = 0; i < num_hugepd; i++, hpdp++)
456 #ifdef CONFIG_PPC_FSL_BOOK3E
457 hugepd_free(tlb, hugepte);
459 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
463 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
464 unsigned long addr, unsigned long end,
465 unsigned long floor, unsigned long ceiling)
473 pmd = pmd_offset(pud, addr);
474 next = pmd_addr_end(addr, end);
477 #ifdef CONFIG_PPC_FSL_BOOK3E
479 * Increment next by the size of the huge mapping since
480 * there may be more than one entry at this level for a
481 * single hugepage, but all of them point to
482 * the same kmem cache that holds the hugepte.
484 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
486 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
487 addr, next, floor, ceiling);
488 } while (addr = next, addr != end);
498 if (end - 1 > ceiling - 1)
501 pmd = pmd_offset(pud, start);
503 pmd_free_tlb(tlb, pmd, start);
506 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
507 unsigned long addr, unsigned long end,
508 unsigned long floor, unsigned long ceiling)
516 pud = pud_offset(pgd, addr);
517 next = pud_addr_end(addr, end);
518 if (!is_hugepd(pud)) {
519 if (pud_none_or_clear_bad(pud))
521 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
524 #ifdef CONFIG_PPC_FSL_BOOK3E
526 * Increment next by the size of the huge mapping since
527 * there may be more than one entry at this level for a
528 * single hugepage, but all of them point to
529 * the same kmem cache that holds the hugepte.
531 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
533 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
534 addr, next, floor, ceiling);
536 } while (addr = next, addr != end);
542 ceiling &= PGDIR_MASK;
546 if (end - 1 > ceiling - 1)
549 pud = pud_offset(pgd, start);
551 pud_free_tlb(tlb, pud, start);
555 * This function frees user-level page tables of a process.
557 * Must be called with pagetable lock held.
559 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
560 unsigned long addr, unsigned long end,
561 unsigned long floor, unsigned long ceiling)
567 * Because there are a number of different possible pagetable
568 * layouts for hugepage ranges, we limit knowledge of how
569 * things should be laid out to the allocation path
570 * (huge_pte_alloc(), above). Everything else works out the
571 * structure as it goes from information in the hugepd
572 * pointers. That means that we can't here use the
573 * optimization used in the normal page free_pgd_range(), of
574 * checking whether we're actually covering a large enough
575 * range to have to do anything at the top level of the walk
576 * instead of at the bottom.
578 * To make sense of this, you should probably go read the big
579 * block comment at the top of the normal free_pgd_range(),
584 next = pgd_addr_end(addr, end);
585 pgd = pgd_offset(tlb->mm, addr);
586 if (!is_hugepd(pgd)) {
587 if (pgd_none_or_clear_bad(pgd))
589 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
591 #ifdef CONFIG_PPC_FSL_BOOK3E
593 * Increment next by the size of the huge mapping since
594 * there may be more than one entry at the pgd level
595 * for a single hugepage, but all of them point to the
596 * same kmem cache that holds the hugepte.
598 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
600 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
601 addr, next, floor, ceiling);
603 } while (addr = next, addr != end);
607 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
614 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
616 /* Verify it is a huge page else bail. */
618 return ERR_PTR(-EINVAL);
620 mask = (1UL << shift) - 1;
621 page = pte_page(*ptep);
623 page += (address & mask) / PAGE_SIZE;
628 int pmd_huge(pmd_t pmd)
633 int pud_huge(pud_t pud)
639 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
640 pmd_t *pmd, int write)
646 static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
647 unsigned long end, int write, struct page **pages, int *nr)
650 unsigned long pte_end;
651 struct page *head, *page, *tail;
655 pte_end = (addr + sz) & ~(sz-1);
660 mask = _PAGE_PRESENT | _PAGE_USER;
664 if ((pte_val(pte) & mask) != mask)
667 /* hugepages are never "special" */
668 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
671 head = pte_page(pte);
673 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
676 VM_BUG_ON(compound_head(page) != head);
681 } while (addr += PAGE_SIZE, addr != end);
683 if (!page_cache_add_speculative(head, refs)) {
688 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
689 /* Could be optimized better */
697 * Any tail page need their mapcount reference taken before we
702 get_huge_page_tail(tail);
709 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
712 unsigned long __boundary = (addr + sz) & ~(sz-1);
713 return (__boundary - 1 < end - 1) ? __boundary : end;
716 int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
717 unsigned long addr, unsigned long end,
718 int write, struct page **pages, int *nr)
721 unsigned long sz = 1UL << hugepd_shift(*hugepd);
724 ptep = hugepte_offset(hugepd, addr, pdshift);
726 next = hugepte_addr_end(addr, end, sz);
727 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
729 } while (ptep++, addr = next, addr != end);
734 #ifdef CONFIG_PPC_MM_SLICES
735 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
736 unsigned long len, unsigned long pgoff,
739 struct hstate *hstate = hstate_file(file);
740 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
742 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
746 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
748 #ifdef CONFIG_PPC_MM_SLICES
749 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
751 return 1UL << mmu_psize_to_shift(psize);
753 if (!is_vm_hugetlb_page(vma))
756 return huge_page_size(hstate_vma(vma));
760 static inline bool is_power_of_4(unsigned long x)
762 if (is_power_of_2(x))
763 return (__ilog2(x) % 2) ? false : true;
767 static int __init add_huge_page_size(unsigned long long size)
769 int shift = __ffs(size);
772 /* Check that it is a page size supported by the hardware and
773 * that it fits within pagetable and slice limits. */
774 #ifdef CONFIG_PPC_FSL_BOOK3E
775 if ((size < PAGE_SIZE) || !is_power_of_4(size))
778 if (!is_power_of_2(size)
779 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
783 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
786 #ifdef CONFIG_SPU_FS_64K_LS
787 /* Disable support for 64K huge pages when 64K SPU local store
788 * support is enabled as the current implementation conflicts.
790 if (shift == PAGE_SHIFT_64K)
792 #endif /* CONFIG_SPU_FS_64K_LS */
794 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
796 /* Return if huge page size has already been setup */
797 if (size_to_hstate(size))
800 hugetlb_add_hstate(shift - PAGE_SHIFT);
805 static int __init hugepage_setup_sz(char *str)
807 unsigned long long size;
809 size = memparse(str, &str);
811 if (add_huge_page_size(size) != 0)
812 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
816 __setup("hugepagesz=", hugepage_setup_sz);
818 #ifdef CONFIG_PPC_FSL_BOOK3E
819 struct kmem_cache *hugepte_cache;
820 static int __init hugetlbpage_init(void)
824 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
827 if (!mmu_psize_defs[psize].shift)
830 shift = mmu_psize_to_shift(psize);
832 /* Don't treat normal page sizes as huge... */
833 if (shift != PAGE_SHIFT)
834 if (add_huge_page_size(1ULL << shift) < 0)
839 * Create a kmem cache for hugeptes. The bottom bits in the pte have
840 * size information encoded in them, so align them to allow this
842 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
843 HUGEPD_SHIFT_MASK + 1, 0, NULL);
844 if (hugepte_cache == NULL)
845 panic("%s: Unable to create kmem cache for hugeptes\n",
848 /* Default hpage size = 4M */
849 if (mmu_psize_defs[MMU_PAGE_4M].shift)
850 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
852 panic("%s: Unable to set default huge page size\n", __func__);
858 static int __init hugetlbpage_init(void)
862 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
865 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
869 if (!mmu_psize_defs[psize].shift)
872 shift = mmu_psize_to_shift(psize);
874 if (add_huge_page_size(1ULL << shift) < 0)
877 if (shift < PMD_SHIFT)
879 else if (shift < PUD_SHIFT)
882 pdshift = PGDIR_SHIFT;
884 pgtable_cache_add(pdshift - shift, NULL);
885 if (!PGT_CACHE(pdshift - shift))
886 panic("hugetlbpage_init(): could not create "
887 "pgtable cache for %d bit pagesize\n", shift);
890 /* Set default large page size. Currently, we pick 16M or 1M
891 * depending on what is available
893 if (mmu_psize_defs[MMU_PAGE_16M].shift)
894 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
895 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
896 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
901 module_init(hugetlbpage_init);
903 void flush_dcache_icache_hugepage(struct page *page)
908 BUG_ON(!PageCompound(page));
910 for (i = 0; i < (1UL << compound_order(page)); i++) {
911 if (!PageHighMem(page)) {
912 __flush_dcache_icache(page_address(page+i));
914 start = kmap_atomic(page+i);
915 __flush_dcache_icache(start);
916 kunmap_atomic(start);