1 // SPDX-License-Identifier: GPL-2.0
3 * SPARC64 Huge TLB page support.
5 * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
10 #include <linux/sched/mm.h>
11 #include <linux/hugetlb.h>
12 #include <linux/pagemap.h>
13 #include <linux/sysctl.h>
16 #include <asm/pgalloc.h>
17 #include <asm/pgtable.h>
19 #include <asm/tlbflush.h>
20 #include <asm/cacheflush.h>
21 #include <asm/mmu_context.h>
23 /* Slightly simplified from the non-hugepage variant because by
24 * definition we don't have to worry about any page coloring stuff
27 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
33 struct hstate *h = hstate_file(filp);
34 unsigned long task_size = TASK_SIZE;
35 struct vm_unmapped_area_info info;
37 if (test_thread_flag(TIF_32BIT))
38 task_size = STACK_TOP32;
42 info.low_limit = TASK_UNMAPPED_BASE;
43 info.high_limit = min(task_size, VA_EXCLUDE_START);
44 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
45 info.align_offset = 0;
46 addr = vm_unmapped_area(&info);
48 if ((addr & ~PAGE_MASK) && task_size > VA_EXCLUDE_END) {
49 VM_BUG_ON(addr != -ENOMEM);
50 info.low_limit = VA_EXCLUDE_END;
51 info.high_limit = task_size;
52 addr = vm_unmapped_area(&info);
59 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
60 const unsigned long len,
61 const unsigned long pgoff,
62 const unsigned long flags)
64 struct hstate *h = hstate_file(filp);
65 struct mm_struct *mm = current->mm;
66 unsigned long addr = addr0;
67 struct vm_unmapped_area_info info;
69 /* This should only ever run for 32-bit processes. */
70 BUG_ON(!test_thread_flag(TIF_32BIT));
72 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
74 info.low_limit = PAGE_SIZE;
75 info.high_limit = mm->mmap_base;
76 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
77 info.align_offset = 0;
78 addr = vm_unmapped_area(&info);
81 * A failed mmap() very likely causes application failure,
82 * so fall back to the bottom-up function here. This scenario
83 * can happen with large stack limits and large mmap()
86 if (addr & ~PAGE_MASK) {
87 VM_BUG_ON(addr != -ENOMEM);
89 info.low_limit = TASK_UNMAPPED_BASE;
90 info.high_limit = STACK_TOP32;
91 addr = vm_unmapped_area(&info);
98 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
99 unsigned long len, unsigned long pgoff, unsigned long flags)
101 struct hstate *h = hstate_file(file);
102 struct mm_struct *mm = current->mm;
103 struct vm_area_struct *vma;
104 unsigned long task_size = TASK_SIZE;
106 if (test_thread_flag(TIF_32BIT))
107 task_size = STACK_TOP32;
109 if (len & ~huge_page_mask(h))
114 if (flags & MAP_FIXED) {
115 if (prepare_hugepage_range(file, addr, len))
121 addr = ALIGN(addr, huge_page_size(h));
122 vma = find_vma(mm, addr);
123 if (task_size - len >= addr &&
124 (!vma || addr + len <= vm_start_gap(vma)))
127 if (mm->get_unmapped_area == arch_get_unmapped_area)
128 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
131 return hugetlb_get_unmapped_area_topdown(file, addr, len,
135 static pte_t sun4u_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
140 static pte_t sun4v_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
142 unsigned long hugepage_size = _PAGE_SZ4MB_4V;
144 pte_val(entry) = pte_val(entry) & ~_PAGE_SZALL_4V;
147 case HPAGE_16GB_SHIFT:
148 hugepage_size = _PAGE_SZ16GB_4V;
149 pte_val(entry) |= _PAGE_PUD_HUGE;
151 case HPAGE_2GB_SHIFT:
152 hugepage_size = _PAGE_SZ2GB_4V;
153 pte_val(entry) |= _PAGE_PMD_HUGE;
155 case HPAGE_256MB_SHIFT:
156 hugepage_size = _PAGE_SZ256MB_4V;
157 pte_val(entry) |= _PAGE_PMD_HUGE;
160 pte_val(entry) |= _PAGE_PMD_HUGE;
162 case HPAGE_64K_SHIFT:
163 hugepage_size = _PAGE_SZ64K_4V;
166 WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift);
169 pte_val(entry) = pte_val(entry) | hugepage_size;
173 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift)
175 if (tlb_type == hypervisor)
176 return sun4v_hugepage_shift_to_tte(entry, shift);
178 return sun4u_hugepage_shift_to_tte(entry, shift);
181 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
182 struct page *page, int writeable)
184 unsigned int shift = huge_page_shift(hstate_vma(vma));
186 return hugepage_shift_to_tte(entry, shift);
189 static unsigned int sun4v_huge_tte_to_shift(pte_t entry)
191 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V;
194 switch (tte_szbits) {
195 case _PAGE_SZ16GB_4V:
196 shift = HPAGE_16GB_SHIFT;
199 shift = HPAGE_2GB_SHIFT;
201 case _PAGE_SZ256MB_4V:
202 shift = HPAGE_256MB_SHIFT;
205 shift = REAL_HPAGE_SHIFT;
208 shift = HPAGE_64K_SHIFT;
217 static unsigned int sun4u_huge_tte_to_shift(pte_t entry)
219 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U;
222 switch (tte_szbits) {
223 case _PAGE_SZ256MB_4U:
224 shift = HPAGE_256MB_SHIFT;
227 shift = REAL_HPAGE_SHIFT;
230 shift = HPAGE_64K_SHIFT;
239 static unsigned int huge_tte_to_shift(pte_t entry)
243 if (tlb_type == hypervisor)
244 shift = sun4v_huge_tte_to_shift(entry);
246 shift = sun4u_huge_tte_to_shift(entry);
248 if (shift == PAGE_SHIFT)
249 WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n",
255 static unsigned long huge_tte_to_size(pte_t pte)
257 unsigned long size = 1UL << huge_tte_to_shift(pte);
259 if (size == REAL_HPAGE_SIZE)
264 pte_t *huge_pte_alloc(struct mm_struct *mm,
265 unsigned long addr, unsigned long sz)
271 pgd = pgd_offset(mm, addr);
272 pud = pud_alloc(mm, pgd, addr);
277 pmd = pmd_alloc(mm, pud, addr);
282 return pte_alloc_map(mm, pmd, addr);
285 pte_t *huge_pte_offset(struct mm_struct *mm,
286 unsigned long addr, unsigned long sz)
292 pgd = pgd_offset(mm, addr);
295 pud = pud_offset(pgd, addr);
298 if (is_hugetlb_pud(*pud))
300 pmd = pmd_offset(pud, addr);
303 if (is_hugetlb_pmd(*pmd))
305 return pte_offset_map(pmd, addr);
308 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
309 pte_t *ptep, pte_t entry)
311 unsigned int nptes, orig_shift, shift;
312 unsigned long i, size;
315 size = huge_tte_to_size(entry);
318 if (size >= PUD_SIZE)
320 else if (size >= PMD_SIZE)
325 nptes = size >> shift;
327 if (!pte_present(*ptep) && pte_present(entry))
328 mm->context.hugetlb_pte_count += nptes;
332 orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig);
334 for (i = 0; i < nptes; i++)
335 ptep[i] = __pte(pte_val(entry) + (i << shift));
337 maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift);
338 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
339 if (size == HPAGE_SIZE)
340 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0,
344 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
347 unsigned int i, nptes, orig_shift, shift;
352 size = huge_tte_to_size(entry);
355 if (size >= PUD_SIZE)
357 else if (size >= PMD_SIZE)
362 nptes = size >> shift;
363 orig_shift = pte_none(entry) ? PAGE_SHIFT : huge_tte_to_shift(entry);
365 if (pte_present(entry))
366 mm->context.hugetlb_pte_count -= nptes;
369 for (i = 0; i < nptes; i++)
370 ptep[i] = __pte(0UL);
372 maybe_tlb_batch_add(mm, addr, ptep, entry, 0, orig_shift);
373 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
374 if (size == HPAGE_SIZE)
375 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0,
381 int pmd_huge(pmd_t pmd)
383 return !pmd_none(pmd) &&
384 (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID;
387 int pud_huge(pud_t pud)
389 return !pud_none(pud) &&
390 (pud_val(pud) & (_PAGE_VALID|_PAGE_PUD_HUGE)) != _PAGE_VALID;
393 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
396 pgtable_t token = pmd_pgtable(*pmd);
399 pte_free_tlb(tlb, token, addr);
400 atomic_long_dec(&tlb->mm->nr_ptes);
403 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
404 unsigned long addr, unsigned long end,
405 unsigned long floor, unsigned long ceiling)
412 pmd = pmd_offset(pud, addr);
414 next = pmd_addr_end(addr, end);
417 if (is_hugetlb_pmd(*pmd))
420 hugetlb_free_pte_range(tlb, pmd, addr);
421 } while (pmd++, addr = next, addr != end);
431 if (end - 1 > ceiling - 1)
434 pmd = pmd_offset(pud, start);
436 pmd_free_tlb(tlb, pmd, start);
437 mm_dec_nr_pmds(tlb->mm);
440 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
441 unsigned long addr, unsigned long end,
442 unsigned long floor, unsigned long ceiling)
449 pud = pud_offset(pgd, addr);
451 next = pud_addr_end(addr, end);
452 if (pud_none_or_clear_bad(pud))
454 if (is_hugetlb_pud(*pud))
457 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
459 } while (pud++, addr = next, addr != end);
465 ceiling &= PGDIR_MASK;
469 if (end - 1 > ceiling - 1)
472 pud = pud_offset(pgd, start);
474 pud_free_tlb(tlb, pud, start);
477 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
478 unsigned long addr, unsigned long end,
479 unsigned long floor, unsigned long ceiling)
495 if (end - 1 > ceiling - 1)
500 pgd = pgd_offset(tlb->mm, addr);
502 next = pgd_addr_end(addr, end);
503 if (pgd_none_or_clear_bad(pgd))
505 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
506 } while (pgd++, addr = next, addr != end);