2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
15 #include <asm/pgtable.h>
17 #include <linux/hugetlb.h>
19 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
20 static unsigned long nr_huge_pages, free_huge_pages;
21 unsigned long max_huge_pages;
22 static struct list_head hugepage_freelists[MAX_NUMNODES];
23 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
24 static unsigned int free_huge_pages_node[MAX_NUMNODES];
25 static DEFINE_SPINLOCK(hugetlb_lock);
27 static void enqueue_huge_page(struct page *page)
29 int nid = page_to_nid(page);
30 list_add(&page->lru, &hugepage_freelists[nid]);
32 free_huge_pages_node[nid]++;
35 static struct page *dequeue_huge_page(void)
37 int nid = numa_node_id();
38 struct page *page = NULL;
40 if (list_empty(&hugepage_freelists[nid])) {
41 for (nid = 0; nid < MAX_NUMNODES; ++nid)
42 if (!list_empty(&hugepage_freelists[nid]))
45 if (nid >= 0 && nid < MAX_NUMNODES &&
46 !list_empty(&hugepage_freelists[nid])) {
47 page = list_entry(hugepage_freelists[nid].next,
51 free_huge_pages_node[nid]--;
56 static struct page *alloc_fresh_huge_page(void)
60 page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
62 nid = (nid + 1) % num_online_nodes();
65 nr_huge_pages_node[page_to_nid(page)]++;
70 void free_huge_page(struct page *page)
72 BUG_ON(page_count(page));
74 INIT_LIST_HEAD(&page->lru);
75 page[1].mapping = NULL;
77 spin_lock(&hugetlb_lock);
78 enqueue_huge_page(page);
79 spin_unlock(&hugetlb_lock);
82 struct page *alloc_huge_page(void)
87 spin_lock(&hugetlb_lock);
88 page = dequeue_huge_page();
90 spin_unlock(&hugetlb_lock);
93 spin_unlock(&hugetlb_lock);
94 set_page_count(page, 1);
95 page[1].mapping = (void *)free_huge_page;
96 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
97 clear_highpage(&page[i]);
101 static int __init hugetlb_init(void)
106 for (i = 0; i < MAX_NUMNODES; ++i)
107 INIT_LIST_HEAD(&hugepage_freelists[i]);
109 for (i = 0; i < max_huge_pages; ++i) {
110 page = alloc_fresh_huge_page();
113 spin_lock(&hugetlb_lock);
114 enqueue_huge_page(page);
115 spin_unlock(&hugetlb_lock);
117 max_huge_pages = free_huge_pages = nr_huge_pages = i;
118 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
121 module_init(hugetlb_init);
123 static int __init hugetlb_setup(char *s)
125 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
129 __setup("hugepages=", hugetlb_setup);
132 static void update_and_free_page(struct page *page)
136 nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
137 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
138 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
139 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
140 1 << PG_private | 1<< PG_writeback);
141 set_page_count(&page[i], 0);
143 set_page_count(page, 1);
144 __free_pages(page, HUGETLB_PAGE_ORDER);
147 #ifdef CONFIG_HIGHMEM
148 static void try_to_free_low(unsigned long count)
151 for (i = 0; i < MAX_NUMNODES; ++i) {
152 struct page *page, *next;
153 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
154 if (PageHighMem(page))
156 list_del(&page->lru);
157 update_and_free_page(page);
158 nid = page_zone(page)->zone_pgdat->node_id;
160 free_huge_pages_node[nid]--;
161 if (count >= nr_huge_pages)
167 static inline void try_to_free_low(unsigned long count)
172 static unsigned long set_max_huge_pages(unsigned long count)
174 while (count > nr_huge_pages) {
175 struct page *page = alloc_fresh_huge_page();
177 return nr_huge_pages;
178 spin_lock(&hugetlb_lock);
179 enqueue_huge_page(page);
180 spin_unlock(&hugetlb_lock);
182 if (count >= nr_huge_pages)
183 return nr_huge_pages;
185 spin_lock(&hugetlb_lock);
186 try_to_free_low(count);
187 while (count < nr_huge_pages) {
188 struct page *page = dequeue_huge_page();
191 update_and_free_page(page);
193 spin_unlock(&hugetlb_lock);
194 return nr_huge_pages;
197 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
198 struct file *file, void __user *buffer,
199 size_t *length, loff_t *ppos)
201 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
202 max_huge_pages = set_max_huge_pages(max_huge_pages);
205 #endif /* CONFIG_SYSCTL */
207 int hugetlb_report_meminfo(char *buf)
210 "HugePages_Total: %5lu\n"
211 "HugePages_Free: %5lu\n"
212 "Hugepagesize: %5lu kB\n",
218 int hugetlb_report_node_meminfo(int nid, char *buf)
221 "Node %d HugePages_Total: %5u\n"
222 "Node %d HugePages_Free: %5u\n",
223 nid, nr_huge_pages_node[nid],
224 nid, free_huge_pages_node[nid]);
227 int is_hugepage_mem_enough(size_t size)
229 return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
232 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
233 unsigned long hugetlb_total_pages(void)
235 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
237 EXPORT_SYMBOL(hugetlb_total_pages);
240 * We cannot handle pagefaults against hugetlb pages at all. They cause
241 * handle_mm_fault() to try to instantiate regular-sized pages in the
242 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
245 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
246 unsigned long address, int *unused)
252 struct vm_operations_struct hugetlb_vm_ops = {
253 .nopage = hugetlb_nopage,
256 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page)
260 if (vma->vm_flags & VM_WRITE) {
262 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
264 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
266 entry = pte_mkyoung(entry);
267 entry = pte_mkhuge(entry);
272 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
273 struct vm_area_struct *vma)
275 pte_t *src_pte, *dst_pte, entry;
276 struct page *ptepage;
277 unsigned long addr = vma->vm_start;
278 unsigned long end = vma->vm_end;
281 dst_pte = huge_pte_alloc(dst, addr);
284 src_pte = huge_pte_offset(src, addr);
285 BUG_ON(!src_pte || pte_none(*src_pte)); /* prefaulted */
287 ptepage = pte_page(entry);
289 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
290 set_huge_pte_at(dst, addr, dst_pte, entry);
299 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
302 struct mm_struct *mm = vma->vm_mm;
303 unsigned long address;
308 WARN_ON(!is_vm_hugetlb_page(vma));
309 BUG_ON(start & ~HPAGE_MASK);
310 BUG_ON(end & ~HPAGE_MASK);
312 for (address = start; address < end; address += HPAGE_SIZE) {
313 ptep = huge_pte_offset(mm, address);
315 /* This can happen on truncate, or if an
316 * mmap() is aborted due to an error before
320 pte = huge_ptep_get_and_clear(mm, address, ptep);
324 page = pte_page(pte);
327 add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
328 flush_tlb_range(vma, start, end);
331 void zap_hugepage_range(struct vm_area_struct *vma,
332 unsigned long start, unsigned long length)
334 struct mm_struct *mm = vma->vm_mm;
336 spin_lock(&mm->page_table_lock);
337 unmap_hugepage_range(vma, start, start + length);
338 spin_unlock(&mm->page_table_lock);
341 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
343 struct mm_struct *mm = current->mm;
347 WARN_ON(!is_vm_hugetlb_page(vma));
348 BUG_ON(vma->vm_start & ~HPAGE_MASK);
349 BUG_ON(vma->vm_end & ~HPAGE_MASK);
351 hugetlb_prefault_arch_hook(mm);
353 spin_lock(&mm->page_table_lock);
354 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
356 pte_t *pte = huge_pte_alloc(mm, addr);
364 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
365 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
366 page = find_get_page(mapping, idx);
368 /* charge the fs quota first */
369 if (hugetlb_get_quota(mapping)) {
373 page = alloc_huge_page();
375 hugetlb_put_quota(mapping);
379 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
383 hugetlb_put_quota(mapping);
384 free_huge_page(page);
388 add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE);
389 set_huge_pte_at(mm, addr, pte, make_huge_pte(vma, page));
392 spin_unlock(&mm->page_table_lock);
396 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
397 struct page **pages, struct vm_area_struct **vmas,
398 unsigned long *position, int *length, int i)
400 unsigned long vpfn, vaddr = *position;
401 int remainder = *length;
403 BUG_ON(!is_vm_hugetlb_page(vma));
405 vpfn = vaddr/PAGE_SIZE;
406 while (vaddr < vma->vm_end && remainder) {
412 /* Some archs (sparc64, sh*) have multiple
413 * pte_ts to each hugepage. We have to make
414 * sure we get the first, for the page
415 * indexing below to work. */
416 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
418 /* hugetlb should be locked, and hence, prefaulted */
419 WARN_ON(!pte || pte_none(*pte));
421 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
423 WARN_ON(!PageCompound(page));