2 * IA-64 Huge TLB Page Support for Kernel.
4 * Copyright (C) 2002-2004 Rohit Seth <rohit.seth@intel.com>
5 * Copyright (C) 2003-2004 Ken Chen <kenneth.w.chen@intel.com>
7 * Sep, 2003: add numa support
8 * Feb, 2004: dynamic hugetlb page size via boot parameter
11 #include <linux/config.h>
12 #include <linux/init.h>
15 #include <linux/hugetlb.h>
16 #include <linux/pagemap.h>
17 #include <linux/smp_lock.h>
18 #include <linux/slab.h>
19 #include <linux/sysctl.h>
21 #include <asm/pgalloc.h>
23 #include <asm/tlbflush.h>
25 unsigned int hpage_shift=HPAGE_SHIFT_DEFAULT;
28 huge_pte_alloc (struct mm_struct *mm, unsigned long addr)
30 unsigned long taddr = htlbpage_to_page(addr);
36 pgd = pgd_offset(mm, taddr);
37 pud = pud_alloc(mm, pgd, taddr);
39 pmd = pmd_alloc(mm, pud, taddr);
41 pte = pte_alloc_map(mm, pmd, taddr);
47 huge_pte_offset (struct mm_struct *mm, unsigned long addr)
49 unsigned long taddr = htlbpage_to_page(addr);
55 pgd = pgd_offset(mm, taddr);
56 if (pgd_present(*pgd)) {
57 pud = pud_offset(pgd, taddr);
58 if (pud_present(*pud)) {
59 pmd = pmd_offset(pud, taddr);
60 if (pmd_present(*pmd))
61 pte = pte_offset_map(pmd, taddr);
68 #define mk_pte_huge(entry) { pte_val(entry) |= _PAGE_P; }
71 set_huge_pte (struct mm_struct *mm, struct vm_area_struct *vma,
72 struct page *page, pte_t * page_table, int write_access)
76 add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
79 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
81 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
82 entry = pte_mkyoung(entry);
84 set_pte(page_table, entry);
88 * This function checks for proper alignment of input addr and len parameters.
90 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
92 if (len & ~HPAGE_MASK)
94 if (addr & ~HPAGE_MASK)
96 if (REGION_NUMBER(addr) != REGION_HPAGE)
102 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
103 struct vm_area_struct *vma)
105 pte_t *src_pte, *dst_pte, entry;
106 struct page *ptepage;
107 unsigned long addr = vma->vm_start;
108 unsigned long end = vma->vm_end;
111 dst_pte = huge_pte_alloc(dst, addr);
114 src_pte = huge_pte_offset(src, addr);
116 ptepage = pte_page(entry);
118 set_pte(dst_pte, entry);
119 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
128 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
129 struct page **pages, struct vm_area_struct **vmas,
130 unsigned long *st, int *length, int i)
133 unsigned long start = *st;
134 unsigned long pstart;
139 pstart = start & HPAGE_MASK;
140 ptep = huge_pte_offset(mm, start);
144 page = pte_page(pte);
146 page += ((start & ~HPAGE_MASK) >> PAGE_SHIFT);
155 if (((start & HPAGE_MASK) == pstart) && len &&
156 (start < vma->vm_end))
158 } while (len && start < vma->vm_end);
164 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long addr, int write)
169 if (REGION_NUMBER(addr) != REGION_HPAGE)
170 return ERR_PTR(-EINVAL);
172 ptep = huge_pte_offset(mm, addr);
173 if (!ptep || pte_none(*ptep))
175 page = pte_page(*ptep);
176 page += ((addr & ~HPAGE_MASK) >> PAGE_SHIFT);
179 int pmd_huge(pmd_t pmd)
184 follow_huge_pmd(struct mm_struct *mm, unsigned long address, pmd_t *pmd, int write)
190 * Same as generic free_pgtables(), except constant PGDIR_* and pgd_offset
191 * are hugetlb region specific.
193 void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
194 unsigned long start, unsigned long end)
196 unsigned long first = start & HUGETLB_PGDIR_MASK;
197 unsigned long last = end + HUGETLB_PGDIR_SIZE - 1;
198 struct mm_struct *mm = tlb->mm;
204 if (prev->vm_end > start) {
205 if (last > prev->vm_start)
206 last = prev->vm_start;
211 struct vm_area_struct *next = prev->vm_next;
214 if (next->vm_start < start) {
218 if (last > next->vm_start)
219 last = next->vm_start;
221 if (prev->vm_end > first)
222 first = prev->vm_end;
226 if (last < first) /* for arches with discontiguous pgd indices */
228 clear_page_range(tlb, first, last);
231 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
233 struct mm_struct *mm = vma->vm_mm;
234 unsigned long address;
238 BUG_ON(start & (HPAGE_SIZE - 1));
239 BUG_ON(end & (HPAGE_SIZE - 1));
241 for (address = start; address < end; address += HPAGE_SIZE) {
242 pte = huge_pte_offset(mm, address);
245 page = pte_page(*pte);
247 pte_clear(mm, address, pte);
249 add_mm_counter(mm, rss, - ((end - start) >> PAGE_SHIFT));
250 flush_tlb_range(vma, start, end);
253 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
255 struct mm_struct *mm = current->mm;
259 BUG_ON(vma->vm_start & ~HPAGE_MASK);
260 BUG_ON(vma->vm_end & ~HPAGE_MASK);
262 spin_lock(&mm->page_table_lock);
263 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
265 pte_t *pte = huge_pte_alloc(mm, addr);
275 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
276 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
277 page = find_get_page(mapping, idx);
279 /* charge the fs quota first */
280 if (hugetlb_get_quota(mapping)) {
284 page = alloc_huge_page();
286 hugetlb_put_quota(mapping);
290 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
294 hugetlb_put_quota(mapping);
295 page_cache_release(page);
299 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
302 spin_unlock(&mm->page_table_lock);
306 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
307 unsigned long pgoff, unsigned long flags)
309 struct vm_area_struct *vmm;
311 if (len > RGN_MAP_LIMIT)
313 if (len & ~HPAGE_MASK)
315 /* This code assumes that REGION_HPAGE != 0. */
316 if ((REGION_NUMBER(addr) != REGION_HPAGE) || (addr & (HPAGE_SIZE - 1)))
317 addr = HPAGE_REGION_BASE;
319 addr = ALIGN(addr, HPAGE_SIZE);
320 for (vmm = find_vma(current->mm, addr); ; vmm = vmm->vm_next) {
321 /* At this point: (!vmm || addr < vmm->vm_end). */
322 if (REGION_OFFSET(addr) + len > RGN_MAP_LIMIT)
324 if (!vmm || (addr + len) <= vmm->vm_start)
326 addr = ALIGN(vmm->vm_end, HPAGE_SIZE);
330 static int __init hugetlb_setup_sz(char *str)
333 unsigned long long size;
335 if (ia64_pal_vm_page_size(&tr_pages, NULL) != 0)
337 * shouldn't happen, but just in case.
339 tr_pages = 0x15557000UL;
341 size = memparse(str, &str);
342 if (*str || (size & (size-1)) || !(tr_pages & size) ||
344 size >= (1UL << PAGE_SHIFT << MAX_ORDER)) {
345 printk(KERN_WARNING "Invalid huge page size specified\n");
349 hpage_shift = __ffs(size);
351 * boot cpu already executed ia64_mmu_init, and has HPAGE_SHIFT_DEFAULT
352 * override here with new page shift.
354 ia64_set_rr(HPAGE_REGION_BASE, hpage_shift << 2);
357 __setup("hugepagesz=", hugetlb_setup_sz);