2 * SPARC64 Huge TLB page support.
4 * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
9 #include <linux/sched/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/pagemap.h>
12 #include <linux/sysctl.h>
15 #include <asm/pgalloc.h>
16 #include <asm/pgtable.h>
18 #include <asm/tlbflush.h>
19 #include <asm/cacheflush.h>
20 #include <asm/mmu_context.h>
22 /* Slightly simplified from the non-hugepage variant because by
23 * definition we don't have to worry about any page coloring stuff
26 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
32 struct hstate *h = hstate_file(filp);
33 unsigned long task_size = TASK_SIZE;
34 struct vm_unmapped_area_info info;
36 if (test_thread_flag(TIF_32BIT))
37 task_size = STACK_TOP32;
41 info.low_limit = TASK_UNMAPPED_BASE;
42 info.high_limit = min(task_size, VA_EXCLUDE_START);
43 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
44 info.align_offset = 0;
45 addr = vm_unmapped_area(&info);
47 if ((addr & ~PAGE_MASK) && task_size > VA_EXCLUDE_END) {
48 VM_BUG_ON(addr != -ENOMEM);
49 info.low_limit = VA_EXCLUDE_END;
50 info.high_limit = task_size;
51 addr = vm_unmapped_area(&info);
58 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
59 const unsigned long len,
60 const unsigned long pgoff,
61 const unsigned long flags)
63 struct hstate *h = hstate_file(filp);
64 struct mm_struct *mm = current->mm;
65 unsigned long addr = addr0;
66 struct vm_unmapped_area_info info;
68 /* This should only ever run for 32-bit processes. */
69 BUG_ON(!test_thread_flag(TIF_32BIT));
71 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
73 info.low_limit = PAGE_SIZE;
74 info.high_limit = mm->mmap_base;
75 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
76 info.align_offset = 0;
77 addr = vm_unmapped_area(&info);
80 * A failed mmap() very likely causes application failure,
81 * so fall back to the bottom-up function here. This scenario
82 * can happen with large stack limits and large mmap()
85 if (addr & ~PAGE_MASK) {
86 VM_BUG_ON(addr != -ENOMEM);
88 info.low_limit = TASK_UNMAPPED_BASE;
89 info.high_limit = STACK_TOP32;
90 addr = vm_unmapped_area(&info);
97 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
98 unsigned long len, unsigned long pgoff, unsigned long flags)
100 struct hstate *h = hstate_file(file);
101 struct mm_struct *mm = current->mm;
102 struct vm_area_struct *vma;
103 unsigned long task_size = TASK_SIZE;
105 if (test_thread_flag(TIF_32BIT))
106 task_size = STACK_TOP32;
108 if (len & ~huge_page_mask(h))
113 if (flags & MAP_FIXED) {
114 if (prepare_hugepage_range(file, addr, len))
120 addr = ALIGN(addr, huge_page_size(h));
121 vma = find_vma(mm, addr);
122 if (task_size - len >= addr &&
123 (!vma || addr + len <= vma->vm_start))
126 if (mm->get_unmapped_area == arch_get_unmapped_area)
127 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
130 return hugetlb_get_unmapped_area_topdown(file, addr, len,
134 static pte_t sun4u_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
139 static pte_t sun4v_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
141 unsigned long hugepage_size = _PAGE_SZ4MB_4V;
143 pte_val(entry) = pte_val(entry) & ~_PAGE_SZALL_4V;
146 case HPAGE_2GB_SHIFT:
147 hugepage_size = _PAGE_SZ2GB_4V;
148 pte_val(entry) |= _PAGE_PMD_HUGE;
150 case HPAGE_256MB_SHIFT:
151 hugepage_size = _PAGE_SZ256MB_4V;
152 pte_val(entry) |= _PAGE_PMD_HUGE;
155 pte_val(entry) |= _PAGE_PMD_HUGE;
157 case HPAGE_64K_SHIFT:
158 hugepage_size = _PAGE_SZ64K_4V;
161 WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift);
164 pte_val(entry) = pte_val(entry) | hugepage_size;
168 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift)
170 if (tlb_type == hypervisor)
171 return sun4v_hugepage_shift_to_tte(entry, shift);
173 return sun4u_hugepage_shift_to_tte(entry, shift);
176 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
177 struct page *page, int writeable)
179 unsigned int shift = huge_page_shift(hstate_vma(vma));
181 return hugepage_shift_to_tte(entry, shift);
184 static unsigned int sun4v_huge_tte_to_shift(pte_t entry)
186 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V;
189 switch (tte_szbits) {
191 shift = HPAGE_2GB_SHIFT;
193 case _PAGE_SZ256MB_4V:
194 shift = HPAGE_256MB_SHIFT;
197 shift = REAL_HPAGE_SHIFT;
200 shift = HPAGE_64K_SHIFT;
209 static unsigned int sun4u_huge_tte_to_shift(pte_t entry)
211 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U;
214 switch (tte_szbits) {
215 case _PAGE_SZ256MB_4U:
216 shift = HPAGE_256MB_SHIFT;
219 shift = REAL_HPAGE_SHIFT;
222 shift = HPAGE_64K_SHIFT;
231 static unsigned int huge_tte_to_shift(pte_t entry)
235 if (tlb_type == hypervisor)
236 shift = sun4v_huge_tte_to_shift(entry);
238 shift = sun4u_huge_tte_to_shift(entry);
240 if (shift == PAGE_SHIFT)
241 WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n",
247 static unsigned long huge_tte_to_size(pte_t pte)
249 unsigned long size = 1UL << huge_tte_to_shift(pte);
251 if (size == REAL_HPAGE_SIZE)
256 pte_t *huge_pte_alloc(struct mm_struct *mm,
257 unsigned long addr, unsigned long sz)
264 pgd = pgd_offset(mm, addr);
265 pud = pud_alloc(mm, pgd, addr);
267 pmd = pmd_alloc(mm, pud, addr);
274 pte = pte_alloc_map(mm, pmd, addr);
280 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
287 pgd = pgd_offset(mm, addr);
288 if (!pgd_none(*pgd)) {
289 pud = pud_offset(pgd, addr);
290 if (!pud_none(*pud)) {
291 pmd = pmd_offset(pud, addr);
292 if (!pmd_none(*pmd)) {
293 if (is_hugetlb_pmd(*pmd))
296 pte = pte_offset_map(pmd, addr);
304 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
305 pte_t *ptep, pte_t entry)
307 unsigned int i, nptes, orig_shift, shift;
311 size = huge_tte_to_size(entry);
312 shift = size >= HPAGE_SIZE ? PMD_SHIFT : PAGE_SHIFT;
313 nptes = size >> shift;
315 if (!pte_present(*ptep) && pte_present(entry))
316 mm->context.hugetlb_pte_count += nptes;
320 orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig);
322 for (i = 0; i < nptes; i++)
323 ptep[i] = __pte(pte_val(entry) + (i << shift));
325 maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift);
326 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
327 if (size == HPAGE_SIZE)
328 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0,
332 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
335 unsigned int i, nptes, hugepage_shift;
340 size = huge_tte_to_size(entry);
341 if (size >= HPAGE_SIZE)
342 nptes = size >> PMD_SHIFT;
344 nptes = size >> PAGE_SHIFT;
346 hugepage_shift = pte_none(entry) ? PAGE_SHIFT :
347 huge_tte_to_shift(entry);
349 if (pte_present(entry))
350 mm->context.hugetlb_pte_count -= nptes;
353 for (i = 0; i < nptes; i++)
354 ptep[i] = __pte(0UL);
356 maybe_tlb_batch_add(mm, addr, ptep, entry, 0, hugepage_shift);
357 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
358 if (size == HPAGE_SIZE)
359 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0,
365 int pmd_huge(pmd_t pmd)
367 return !pmd_none(pmd) &&
368 (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID;
371 int pud_huge(pud_t pud)
376 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
379 pgtable_t token = pmd_pgtable(*pmd);
382 pte_free_tlb(tlb, token, addr);
383 atomic_long_dec(&tlb->mm->nr_ptes);
386 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
387 unsigned long addr, unsigned long end,
388 unsigned long floor, unsigned long ceiling)
395 pmd = pmd_offset(pud, addr);
397 next = pmd_addr_end(addr, end);
400 if (is_hugetlb_pmd(*pmd))
403 hugetlb_free_pte_range(tlb, pmd, addr);
404 } while (pmd++, addr = next, addr != end);
414 if (end - 1 > ceiling - 1)
417 pmd = pmd_offset(pud, start);
419 pmd_free_tlb(tlb, pmd, start);
420 mm_dec_nr_pmds(tlb->mm);
423 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
424 unsigned long addr, unsigned long end,
425 unsigned long floor, unsigned long ceiling)
432 pud = pud_offset(pgd, addr);
434 next = pud_addr_end(addr, end);
435 if (pud_none_or_clear_bad(pud))
437 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
439 } while (pud++, addr = next, addr != end);
445 ceiling &= PGDIR_MASK;
449 if (end - 1 > ceiling - 1)
452 pud = pud_offset(pgd, start);
454 pud_free_tlb(tlb, pud, start);
457 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
458 unsigned long addr, unsigned long end,
459 unsigned long floor, unsigned long ceiling)
475 if (end - 1 > ceiling - 1)
480 pgd = pgd_offset(tlb->mm, addr);
482 next = pgd_addr_end(addr, end);
483 if (pgd_none_or_clear_bad(pgd))
485 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
486 } while (pgd++, addr = next, addr != end);