2 * PPC64 (POWER4) Huge TLB Page Support for Kernel.
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
6 * Based on the IA-32 version:
7 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
12 #include <linux/hugetlb.h>
13 #include <asm/pgtable.h>
14 #include <asm/pgalloc.h>
17 #define PAGE_SHIFT_64K 16
18 #define PAGE_SHIFT_16M 24
19 #define PAGE_SHIFT_16G 34
21 #define MAX_NUMBER_GPAGES 1024
23 /* Tracks the 16G pages after the device tree is scanned and before the
24 * huge_boot_pages list is ready. */
25 static unsigned long gpage_freearray[MAX_NUMBER_GPAGES];
26 static unsigned nr_gpages;
28 /* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad()
29 * will choke on pointers to hugepte tables, which is handy for
30 * catching screwups early. */
32 static inline int shift_to_mmu_psize(unsigned int shift)
36 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize)
37 if (mmu_psize_defs[psize].shift == shift)
42 static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
44 if (mmu_psize_defs[mmu_psize].shift)
45 return mmu_psize_defs[mmu_psize].shift;
49 #define hugepd_none(hpd) ((hpd).pd == 0)
51 static inline pte_t *hugepd_page(hugepd_t hpd)
53 BUG_ON(!hugepd_ok(hpd));
54 return (pte_t *)((hpd.pd & ~HUGEPD_SHIFT_MASK) | 0xc000000000000000);
57 static inline unsigned int hugepd_shift(hugepd_t hpd)
59 return hpd.pd & HUGEPD_SHIFT_MASK;
62 static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr, unsigned pdshift)
64 unsigned long idx = (addr & ((1UL << pdshift) - 1)) >> hugepd_shift(*hpdp);
65 pte_t *dir = hugepd_page(*hpdp);
70 pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
75 hugepd_t *hpdp = NULL;
76 unsigned pdshift = PGDIR_SHIFT;
81 pg = pgdir + pgd_index(ea);
83 hpdp = (hugepd_t *)pg;
84 } else if (!pgd_none(*pg)) {
86 pu = pud_offset(pg, ea);
88 hpdp = (hugepd_t *)pu;
89 else if (!pud_none(*pu)) {
91 pm = pmd_offset(pu, ea);
93 hpdp = (hugepd_t *)pm;
94 else if (!pmd_none(*pm)) {
95 return pte_offset_map(pm, ea);
104 *shift = hugepd_shift(*hpdp);
105 return hugepte_offset(hpdp, ea, pdshift);
108 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
113 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
114 unsigned long address, unsigned pdshift, unsigned pshift)
116 pte_t *new = kmem_cache_zalloc(PGT_CACHE(pdshift - pshift),
117 GFP_KERNEL|__GFP_REPEAT);
119 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
120 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
125 spin_lock(&mm->page_table_lock);
126 if (!hugepd_none(*hpdp))
127 kmem_cache_free(PGT_CACHE(pdshift - pshift), new);
129 hpdp->pd = ((unsigned long)new & ~0x8000000000000000) | pshift;
130 spin_unlock(&mm->page_table_lock);
134 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
139 hugepd_t *hpdp = NULL;
140 unsigned pshift = __ffs(sz);
141 unsigned pdshift = PGDIR_SHIFT;
145 pg = pgd_offset(mm, addr);
146 if (pshift >= PUD_SHIFT) {
147 hpdp = (hugepd_t *)pg;
150 pu = pud_alloc(mm, pg, addr);
151 if (pshift >= PMD_SHIFT) {
152 hpdp = (hugepd_t *)pu;
155 pm = pmd_alloc(mm, pu, addr);
156 hpdp = (hugepd_t *)pm;
163 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
165 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
168 return hugepte_offset(hpdp, addr, pdshift);
171 /* Build list of addresses of gigantic pages. This function is used in early
172 * boot before the buddy or bootmem allocator is setup.
174 void add_gpage(unsigned long addr, unsigned long page_size,
175 unsigned long number_of_pages)
179 while (number_of_pages > 0) {
180 gpage_freearray[nr_gpages] = addr;
187 /* Moves the gigantic page addresses from the temporary list to the
188 * huge_boot_pages list.
190 int alloc_bootmem_huge_page(struct hstate *hstate)
192 struct huge_bootmem_page *m;
195 m = phys_to_virt(gpage_freearray[--nr_gpages]);
196 gpage_freearray[nr_gpages] = 0;
197 list_add(&m->list, &huge_boot_pages);
202 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
207 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
208 unsigned long start, unsigned long end,
209 unsigned long floor, unsigned long ceiling)
211 pte_t *hugepte = hugepd_page(*hpdp);
212 unsigned shift = hugepd_shift(*hpdp);
213 unsigned long pdmask = ~((1UL << pdshift) - 1);
223 if (end - 1 > ceiling - 1)
228 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
231 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
232 unsigned long addr, unsigned long end,
233 unsigned long floor, unsigned long ceiling)
240 pmd = pmd_offset(pud, addr);
242 next = pmd_addr_end(addr, end);
245 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
246 addr, next, floor, ceiling);
247 } while (pmd++, addr = next, addr != end);
257 if (end - 1 > ceiling - 1)
260 pmd = pmd_offset(pud, start);
262 pmd_free_tlb(tlb, pmd, start);
265 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
266 unsigned long addr, unsigned long end,
267 unsigned long floor, unsigned long ceiling)
274 pud = pud_offset(pgd, addr);
276 next = pud_addr_end(addr, end);
277 if (!is_hugepd(pud)) {
278 if (pud_none_or_clear_bad(pud))
280 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
283 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
284 addr, next, floor, ceiling);
286 } while (pud++, addr = next, addr != end);
292 ceiling &= PGDIR_MASK;
296 if (end - 1 > ceiling - 1)
299 pud = pud_offset(pgd, start);
301 pud_free_tlb(tlb, pud, start);
305 * This function frees user-level page tables of a process.
307 * Must be called with pagetable lock held.
309 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
310 unsigned long addr, unsigned long end,
311 unsigned long floor, unsigned long ceiling)
317 * Because there are a number of different possible pagetable
318 * layouts for hugepage ranges, we limit knowledge of how
319 * things should be laid out to the allocation path
320 * (huge_pte_alloc(), above). Everything else works out the
321 * structure as it goes from information in the hugepd
322 * pointers. That means that we can't here use the
323 * optimization used in the normal page free_pgd_range(), of
324 * checking whether we're actually covering a large enough
325 * range to have to do anything at the top level of the walk
326 * instead of at the bottom.
328 * To make sense of this, you should probably go read the big
329 * block comment at the top of the normal free_pgd_range(),
333 pgd = pgd_offset(tlb->mm, addr);
335 next = pgd_addr_end(addr, end);
336 if (!is_hugepd(pgd)) {
337 if (pgd_none_or_clear_bad(pgd))
339 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
341 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
342 addr, next, floor, ceiling);
344 } while (pgd++, addr = next, addr != end);
348 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
355 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
357 /* Verify it is a huge page else bail. */
359 return ERR_PTR(-EINVAL);
361 mask = (1UL << shift) - 1;
362 page = pte_page(*ptep);
364 page += (address & mask) / PAGE_SIZE;
369 int pmd_huge(pmd_t pmd)
374 int pud_huge(pud_t pud)
380 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
381 pmd_t *pmd, int write)
387 static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
388 unsigned long end, int write, struct page **pages, int *nr)
391 unsigned long pte_end;
392 struct page *head, *page;
396 pte_end = (addr + sz) & ~(sz-1);
401 mask = _PAGE_PRESENT | _PAGE_USER;
405 if ((pte_val(pte) & mask) != mask)
408 /* hugepages are never "special" */
409 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
412 head = pte_page(pte);
414 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
416 VM_BUG_ON(compound_head(page) != head);
421 } while (addr += PAGE_SIZE, addr != end);
423 if (!page_cache_add_speculative(head, refs)) {
428 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
429 /* Could be optimized better */
439 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
442 unsigned long __boundary = (addr + sz) & ~(sz-1);
443 return (__boundary - 1 < end - 1) ? __boundary : end;
446 int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
447 unsigned long addr, unsigned long end,
448 int write, struct page **pages, int *nr)
451 unsigned long sz = 1UL << hugepd_shift(*hugepd);
454 ptep = hugepte_offset(hugepd, addr, pdshift);
456 next = hugepte_addr_end(addr, end, sz);
457 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
459 } while (ptep++, addr = next, addr != end);
464 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
465 unsigned long len, unsigned long pgoff,
468 struct hstate *hstate = hstate_file(file);
469 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
471 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
474 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
476 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
478 return 1UL << mmu_psize_to_shift(psize);
481 static int __init add_huge_page_size(unsigned long long size)
483 int shift = __ffs(size);
486 /* Check that it is a page size supported by the hardware and
487 * that it fits within pagetable and slice limits. */
488 if (!is_power_of_2(size)
489 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
492 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
495 #ifdef CONFIG_SPU_FS_64K_LS
496 /* Disable support for 64K huge pages when 64K SPU local store
497 * support is enabled as the current implementation conflicts.
499 if (shift == PAGE_SHIFT_64K)
501 #endif /* CONFIG_SPU_FS_64K_LS */
503 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
505 /* Return if huge page size has already been setup */
506 if (size_to_hstate(size))
509 hugetlb_add_hstate(shift - PAGE_SHIFT);
514 static int __init hugepage_setup_sz(char *str)
516 unsigned long long size;
518 size = memparse(str, &str);
520 if (add_huge_page_size(size) != 0)
521 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
525 __setup("hugepagesz=", hugepage_setup_sz);
527 static int __init hugetlbpage_init(void)
531 if (!cpu_has_feature(CPU_FTR_16M_PAGE))
534 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
538 if (!mmu_psize_defs[psize].shift)
541 shift = mmu_psize_to_shift(psize);
543 if (add_huge_page_size(1ULL << shift) < 0)
546 if (shift < PMD_SHIFT)
548 else if (shift < PUD_SHIFT)
551 pdshift = PGDIR_SHIFT;
553 pgtable_cache_add(pdshift - shift, NULL);
554 if (!PGT_CACHE(pdshift - shift))
555 panic("hugetlbpage_init(): could not create "
556 "pgtable cache for %d bit pagesize\n", shift);
559 /* Set default large page size. Currently, we pick 16M or 1M
560 * depending on what is available
562 if (mmu_psize_defs[MMU_PAGE_16M].shift)
563 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
564 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
565 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
570 module_init(hugetlbpage_init);
572 void flush_dcache_icache_hugepage(struct page *page)
576 BUG_ON(!PageCompound(page));
578 for (i = 0; i < (1UL << compound_order(page)); i++)
579 __flush_dcache_icache(page_address(page+i));