2 * SPARC64 Huge TLB page support.
4 * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
7 #include <linux/init.h>
8 #include <linux/module.h>
11 #include <linux/hugetlb.h>
12 #include <linux/pagemap.h>
13 #include <linux/smp_lock.h>
14 #include <linux/slab.h>
15 #include <linux/sysctl.h>
18 #include <asm/pgalloc.h>
20 #include <asm/tlbflush.h>
21 #include <asm/cacheflush.h>
22 #include <asm/mmu_context.h>
24 /* Slightly simplified from the non-hugepage variant because by
25 * definition we don't have to worry about any page coloring stuff
27 #define VA_EXCLUDE_START (0x0000080000000000UL - (1UL << 32UL))
28 #define VA_EXCLUDE_END (0xfffff80000000000UL + (1UL << 32UL))
30 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
36 struct mm_struct *mm = current->mm;
37 struct vm_area_struct * vma;
38 unsigned long task_size = TASK_SIZE;
39 unsigned long start_addr;
41 if (test_thread_flag(TIF_32BIT))
42 task_size = STACK_TOP32;
43 if (unlikely(len >= VA_EXCLUDE_START))
46 if (len > mm->cached_hole_size) {
47 start_addr = addr = mm->free_area_cache;
49 start_addr = addr = TASK_UNMAPPED_BASE;
50 mm->cached_hole_size = 0;
56 addr = ALIGN(addr, HPAGE_SIZE);
58 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
59 /* At this point: (!vma || addr < vma->vm_end). */
60 if (addr < VA_EXCLUDE_START &&
61 (addr + len) >= VA_EXCLUDE_START) {
62 addr = VA_EXCLUDE_END;
63 vma = find_vma(mm, VA_EXCLUDE_END);
65 if (unlikely(task_size < addr)) {
66 if (start_addr != TASK_UNMAPPED_BASE) {
67 start_addr = addr = TASK_UNMAPPED_BASE;
68 mm->cached_hole_size = 0;
73 if (likely(!vma || addr + len <= vma->vm_start)) {
75 * Remember the place where we stopped the search:
77 mm->free_area_cache = addr + len;
80 if (addr + mm->cached_hole_size < vma->vm_start)
81 mm->cached_hole_size = vma->vm_start - addr;
83 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
88 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
89 const unsigned long len,
90 const unsigned long pgoff,
91 const unsigned long flags)
93 struct vm_area_struct *vma;
94 struct mm_struct *mm = current->mm;
95 unsigned long addr = addr0;
97 /* This should only ever run for 32-bit processes. */
98 BUG_ON(!test_thread_flag(TIF_32BIT));
100 /* check if free_area_cache is useful for us */
101 if (len <= mm->cached_hole_size) {
102 mm->cached_hole_size = 0;
103 mm->free_area_cache = mm->mmap_base;
106 /* either no address requested or can't fit in requested address hole */
107 addr = mm->free_area_cache & HPAGE_MASK;
109 /* make sure it can fit in the remaining address space */
110 if (likely(addr > len)) {
111 vma = find_vma(mm, addr-len);
112 if (!vma || addr <= vma->vm_start) {
113 /* remember the address as a hint for next time */
114 return (mm->free_area_cache = addr-len);
118 if (unlikely(mm->mmap_base < len))
121 addr = (mm->mmap_base-len) & HPAGE_MASK;
125 * Lookup failure means no vma is above this address,
126 * else if new region fits below vma->vm_start,
127 * return with success:
129 vma = find_vma(mm, addr);
130 if (likely(!vma || addr+len <= vma->vm_start)) {
131 /* remember the address as a hint for next time */
132 return (mm->free_area_cache = addr);
135 /* remember the largest hole we saw so far */
136 if (addr + mm->cached_hole_size < vma->vm_start)
137 mm->cached_hole_size = vma->vm_start - addr;
139 /* try just below the current vma->vm_start */
140 addr = (vma->vm_start-len) & HPAGE_MASK;
141 } while (likely(len < vma->vm_start));
145 * A failed mmap() very likely causes application failure,
146 * so fall back to the bottom-up function here. This scenario
147 * can happen with large stack limits and large mmap()
150 mm->cached_hole_size = ~0UL;
151 mm->free_area_cache = TASK_UNMAPPED_BASE;
152 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
154 * Restore the topdown base:
156 mm->free_area_cache = mm->mmap_base;
157 mm->cached_hole_size = ~0UL;
163 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
164 unsigned long len, unsigned long pgoff, unsigned long flags)
166 struct mm_struct *mm = current->mm;
167 struct vm_area_struct *vma;
168 unsigned long task_size = TASK_SIZE;
170 if (test_thread_flag(TIF_32BIT))
171 task_size = STACK_TOP32;
173 if (len & ~HPAGE_MASK)
179 addr = ALIGN(addr, HPAGE_SIZE);
180 vma = find_vma(mm, addr);
181 if (task_size - len >= addr &&
182 (!vma || addr + len <= vma->vm_start))
185 if (mm->get_unmapped_area == arch_get_unmapped_area)
186 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
189 return hugetlb_get_unmapped_area_topdown(file, addr, len,
193 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
200 /* We must align the address, because our caller will run
201 * set_huge_pte_at() on whatever we return, which writes out
202 * all of the sub-ptes for the hugepage range. So we have
203 * to give it the first such sub-pte.
207 pgd = pgd_offset(mm, addr);
208 pud = pud_alloc(mm, pgd, addr);
210 pmd = pmd_alloc(mm, pud, addr);
212 pte = pte_alloc_map(mm, pmd, addr);
217 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
226 pgd = pgd_offset(mm, addr);
227 if (!pgd_none(*pgd)) {
228 pud = pud_offset(pgd, addr);
229 if (!pud_none(*pud)) {
230 pmd = pmd_offset(pud, addr);
232 pte = pte_offset_map(pmd, addr);
238 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
243 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
244 pte_t *ptep, pte_t entry)
248 if (!pte_present(*ptep) && pte_present(entry))
249 mm->context.huge_pte_count++;
252 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
253 set_pte_at(mm, addr, ptep, entry);
256 pte_val(entry) += PAGE_SIZE;
260 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
267 if (pte_present(entry))
268 mm->context.huge_pte_count--;
272 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
273 pte_clear(mm, addr, ptep);
281 struct page *follow_huge_addr(struct mm_struct *mm,
282 unsigned long address, int write)
284 return ERR_PTR(-EINVAL);
287 int pmd_huge(pmd_t pmd)
292 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
293 pmd_t *pmd, int write)
298 static void context_reload(void *__data)
300 struct mm_struct *mm = __data;
302 if (mm == current->mm)
303 load_secondary_context(mm);
306 void hugetlb_prefault_arch_hook(struct mm_struct *mm)
308 struct tsb_config *tp = &mm->context.tsb_block[MM_TSB_HUGE];
310 if (likely(tp->tsb != NULL))
313 tsb_grow(mm, MM_TSB_HUGE, 0);
314 tsb_context_switch(mm);
317 /* On UltraSPARC-III+ and later, configure the second half of
318 * the Data-TLB for huge pages.
320 if (tlb_type == cheetah_plus) {
323 spin_lock(&ctx_alloc_lock);
324 ctx = mm->context.sparc64_ctx_val;
325 ctx &= ~CTX_PGSZ_MASK;
326 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
327 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
329 if (ctx != mm->context.sparc64_ctx_val) {
330 /* When changing the page size fields, we
331 * must perform a context flush so that no
332 * stale entries match. This flush must
333 * occur with the original context register
338 /* Reload the context register of all processors
339 * also executing in this address space.
341 mm->context.sparc64_ctx_val = ctx;
342 on_each_cpu(context_reload, mm, 0, 0);
344 spin_unlock(&ctx_alloc_lock);