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 if (HPAGE_SHIFT == 0)
109 for (i = 0; i < MAX_NUMNODES; ++i)
110 INIT_LIST_HEAD(&hugepage_freelists[i]);
112 for (i = 0; i < max_huge_pages; ++i) {
113 page = alloc_fresh_huge_page();
116 spin_lock(&hugetlb_lock);
117 enqueue_huge_page(page);
118 spin_unlock(&hugetlb_lock);
120 max_huge_pages = free_huge_pages = nr_huge_pages = i;
121 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
124 module_init(hugetlb_init);
126 static int __init hugetlb_setup(char *s)
128 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
132 __setup("hugepages=", hugetlb_setup);
135 static void update_and_free_page(struct page *page)
139 nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
140 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
141 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
142 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
143 1 << PG_private | 1<< PG_writeback);
144 set_page_count(&page[i], 0);
146 set_page_count(page, 1);
147 __free_pages(page, HUGETLB_PAGE_ORDER);
150 #ifdef CONFIG_HIGHMEM
151 static void try_to_free_low(unsigned long count)
154 for (i = 0; i < MAX_NUMNODES; ++i) {
155 struct page *page, *next;
156 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
157 if (PageHighMem(page))
159 list_del(&page->lru);
160 update_and_free_page(page);
161 nid = page_zone(page)->zone_pgdat->node_id;
163 free_huge_pages_node[nid]--;
164 if (count >= nr_huge_pages)
170 static inline void try_to_free_low(unsigned long count)
175 static unsigned long set_max_huge_pages(unsigned long count)
177 while (count > nr_huge_pages) {
178 struct page *page = alloc_fresh_huge_page();
180 return nr_huge_pages;
181 spin_lock(&hugetlb_lock);
182 enqueue_huge_page(page);
183 spin_unlock(&hugetlb_lock);
185 if (count >= nr_huge_pages)
186 return nr_huge_pages;
188 spin_lock(&hugetlb_lock);
189 try_to_free_low(count);
190 while (count < nr_huge_pages) {
191 struct page *page = dequeue_huge_page();
194 update_and_free_page(page);
196 spin_unlock(&hugetlb_lock);
197 return nr_huge_pages;
200 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
201 struct file *file, void __user *buffer,
202 size_t *length, loff_t *ppos)
204 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
205 max_huge_pages = set_max_huge_pages(max_huge_pages);
208 #endif /* CONFIG_SYSCTL */
210 int hugetlb_report_meminfo(char *buf)
213 "HugePages_Total: %5lu\n"
214 "HugePages_Free: %5lu\n"
215 "Hugepagesize: %5lu kB\n",
221 int hugetlb_report_node_meminfo(int nid, char *buf)
224 "Node %d HugePages_Total: %5u\n"
225 "Node %d HugePages_Free: %5u\n",
226 nid, nr_huge_pages_node[nid],
227 nid, free_huge_pages_node[nid]);
230 int is_hugepage_mem_enough(size_t size)
232 return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
235 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
236 unsigned long hugetlb_total_pages(void)
238 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
240 EXPORT_SYMBOL(hugetlb_total_pages);
243 * We cannot handle pagefaults against hugetlb pages at all. They cause
244 * handle_mm_fault() to try to instantiate regular-sized pages in the
245 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
248 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
249 unsigned long address, int *unused)
255 struct vm_operations_struct hugetlb_vm_ops = {
256 .nopage = hugetlb_nopage,
259 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page)
263 if (vma->vm_flags & VM_WRITE) {
265 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
267 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
269 entry = pte_mkyoung(entry);
270 entry = pte_mkhuge(entry);
275 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
276 struct vm_area_struct *vma)
278 pte_t *src_pte, *dst_pte, entry;
279 struct page *ptepage;
282 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
283 src_pte = huge_pte_offset(src, addr);
286 dst_pte = huge_pte_alloc(dst, addr);
289 spin_lock(&dst->page_table_lock);
290 spin_lock(&src->page_table_lock);
291 if (!pte_none(*src_pte)) {
293 ptepage = pte_page(entry);
295 add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
296 set_huge_pte_at(dst, addr, dst_pte, entry);
298 spin_unlock(&src->page_table_lock);
299 spin_unlock(&dst->page_table_lock);
307 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
310 struct mm_struct *mm = vma->vm_mm;
311 unsigned long address;
316 WARN_ON(!is_vm_hugetlb_page(vma));
317 BUG_ON(start & ~HPAGE_MASK);
318 BUG_ON(end & ~HPAGE_MASK);
320 spin_lock(&mm->page_table_lock);
322 /* Update high watermark before we lower rss */
323 update_hiwater_rss(mm);
325 for (address = start; address < end; address += HPAGE_SIZE) {
326 ptep = huge_pte_offset(mm, address);
330 pte = huge_ptep_get_and_clear(mm, address, ptep);
334 page = pte_page(pte);
336 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
339 spin_unlock(&mm->page_table_lock);
340 flush_tlb_range(vma, start, end);
343 static struct page *find_lock_huge_page(struct address_space *mapping,
348 struct inode *inode = mapping->host;
352 page = find_lock_page(mapping, idx);
356 /* Check to make sure the mapping hasn't been truncated */
357 size = i_size_read(inode) >> HPAGE_SHIFT;
361 if (hugetlb_get_quota(mapping))
363 page = alloc_huge_page();
365 hugetlb_put_quota(mapping);
369 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
372 hugetlb_put_quota(mapping);
381 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
382 unsigned long address, int write_access)
384 int ret = VM_FAULT_SIGBUS;
389 struct address_space *mapping;
391 pte = huge_pte_alloc(mm, address);
395 mapping = vma->vm_file->f_mapping;
396 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
397 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
400 * Use page lock to guard against racing truncation
401 * before we get page_table_lock.
403 page = find_lock_huge_page(mapping, idx);
407 spin_lock(&mm->page_table_lock);
408 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
412 ret = VM_FAULT_MINOR;
416 add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
417 set_huge_pte_at(mm, address, pte, make_huge_pte(vma, page));
418 spin_unlock(&mm->page_table_lock);
424 spin_unlock(&mm->page_table_lock);
425 hugetlb_put_quota(mapping);
431 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
432 struct page **pages, struct vm_area_struct **vmas,
433 unsigned long *position, int *length, int i)
435 unsigned long vpfn, vaddr = *position;
436 int remainder = *length;
438 vpfn = vaddr/PAGE_SIZE;
439 spin_lock(&mm->page_table_lock);
440 while (vaddr < vma->vm_end && remainder) {
445 * Some archs (sparc64, sh*) have multiple pte_ts to
446 * each hugepage. We have to make * sure we get the
447 * first, for the page indexing below to work.
449 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
451 if (!pte || pte_none(*pte)) {
454 spin_unlock(&mm->page_table_lock);
455 ret = hugetlb_fault(mm, vma, vaddr, 0);
456 spin_lock(&mm->page_table_lock);
457 if (ret == VM_FAULT_MINOR)
467 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
480 spin_unlock(&mm->page_table_lock);