2 * SPARC64 Huge TLB page support.
4 * Copyright (C) 2002, 2003 David S. Miller (davem@redhat.com)
7 #include <linux/config.h>
8 #include <linux/init.h>
9 #include <linux/module.h>
12 #include <linux/hugetlb.h>
13 #include <linux/pagemap.h>
14 #include <linux/smp_lock.h>
15 #include <linux/slab.h>
16 #include <linux/sysctl.h>
19 #include <asm/pgalloc.h>
21 #include <asm/tlbflush.h>
22 #include <asm/cacheflush.h>
23 #include <asm/mmu_context.h>
25 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
32 pgd = pgd_offset(mm, addr);
34 pud = pud_offset(pgd, addr);
36 pmd = pmd_alloc(mm, pud, addr);
38 pte = pte_alloc_map(mm, pmd, addr);
44 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
51 pgd = pgd_offset(mm, addr);
53 pud = pud_offset(pgd, addr);
55 pmd = pmd_offset(pud, addr);
57 pte = pte_offset_map(pmd, addr);
63 #define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0)
65 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
67 struct page *page, pte_t * page_table, int write_access)
72 add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
75 entry = pte_mkwrite(pte_mkdirty(mk_pte(page,
78 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
79 entry = pte_mkyoung(entry);
82 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
83 set_pte_at(mm, addr, page_table, entry);
87 pte_val(entry) += PAGE_SIZE;
92 * This function checks for proper alignment of input addr and len parameters.
94 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
96 if (len & ~HPAGE_MASK)
98 if (addr & ~HPAGE_MASK)
103 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
104 struct vm_area_struct *vma)
106 pte_t *src_pte, *dst_pte, entry;
107 struct page *ptepage;
108 unsigned long addr = vma->vm_start;
109 unsigned long end = vma->vm_end;
113 dst_pte = huge_pte_alloc(dst, addr);
116 src_pte = huge_pte_offset(src, addr);
117 BUG_ON(!src_pte || pte_none(*src_pte));
119 ptepage = pte_page(entry);
121 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
122 set_pte_at(dst, addr, dst_pte, entry);
123 pte_val(entry) += PAGE_SIZE;
127 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
135 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
136 struct page **pages, struct vm_area_struct **vmas,
137 unsigned long *position, int *length, int i)
139 unsigned long vaddr = *position;
140 int remainder = *length;
142 WARN_ON(!is_vm_hugetlb_page(vma));
144 while (vaddr < vma->vm_end && remainder) {
149 pte = huge_pte_offset(mm, vaddr);
151 /* hugetlb should be locked, and hence, prefaulted */
152 BUG_ON(!pte || pte_none(*pte));
154 page = pte_page(*pte);
156 WARN_ON(!PageCompound(page));
176 struct page *follow_huge_addr(struct mm_struct *mm,
177 unsigned long address, int write)
179 return ERR_PTR(-EINVAL);
182 int pmd_huge(pmd_t pmd)
187 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
188 pmd_t *pmd, int write)
193 void unmap_hugepage_range(struct vm_area_struct *vma,
194 unsigned long start, unsigned long end)
196 struct mm_struct *mm = vma->vm_mm;
197 unsigned long address;
202 BUG_ON(start & (HPAGE_SIZE - 1));
203 BUG_ON(end & (HPAGE_SIZE - 1));
205 for (address = start; address < end; address += HPAGE_SIZE) {
206 pte = huge_pte_offset(mm, address);
210 page = pte_page(*pte);
212 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
213 pte_clear(mm, address+(i*PAGE_SIZE), pte);
217 add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
218 flush_tlb_range(vma, start, end);
221 static void context_reload(void *__data)
223 struct mm_struct *mm = __data;
225 if (mm == current->mm)
226 load_secondary_context(mm);
229 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
231 struct mm_struct *mm = current->mm;
235 /* On UltraSPARC-III+ and later, configure the second half of
236 * the Data-TLB for huge pages.
238 if (tlb_type == cheetah_plus) {
241 spin_lock(&ctx_alloc_lock);
242 ctx = mm->context.sparc64_ctx_val;
243 ctx &= ~CTX_PGSZ_MASK;
244 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
245 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
247 if (ctx != mm->context.sparc64_ctx_val) {
248 /* When changing the page size fields, we
249 * must perform a context flush so that no
250 * stale entries match. This flush must
251 * occur with the original context register
256 /* Reload the context register of all processors
257 * also executing in this address space.
259 mm->context.sparc64_ctx_val = ctx;
260 on_each_cpu(context_reload, mm, 0, 0);
262 spin_unlock(&ctx_alloc_lock);
265 BUG_ON(vma->vm_start & ~HPAGE_MASK);
266 BUG_ON(vma->vm_end & ~HPAGE_MASK);
268 spin_lock(&mm->page_table_lock);
269 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
271 pte_t *pte = huge_pte_alloc(mm, addr);
281 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
282 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
283 page = find_get_page(mapping, idx);
285 /* charge the fs quota first */
286 if (hugetlb_get_quota(mapping)) {
290 page = alloc_huge_page();
292 hugetlb_put_quota(mapping);
296 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
300 hugetlb_put_quota(mapping);
301 free_huge_page(page);
305 set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
308 spin_unlock(&mm->page_table_lock);