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>
14 #include <linux/mempolicy.h>
15 #include <linux/cpuset.h>
18 #include <asm/pgtable.h>
20 #include <linux/hugetlb.h>
23 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
24 static unsigned long nr_huge_pages, free_huge_pages;
25 unsigned long max_huge_pages;
26 static struct list_head hugepage_freelists[MAX_NUMNODES];
27 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
28 static unsigned int free_huge_pages_node[MAX_NUMNODES];
30 static void clear_huge_page(struct page *page, unsigned long addr)
35 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
37 clear_user_highpage(page + i, addr);
41 static void copy_huge_page(struct page *dst, struct page *src,
47 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
49 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE);
54 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
56 static DEFINE_SPINLOCK(hugetlb_lock);
58 static void enqueue_huge_page(struct page *page)
60 int nid = page_to_nid(page);
61 list_add(&page->lru, &hugepage_freelists[nid]);
63 free_huge_pages_node[nid]++;
66 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
67 unsigned long address)
69 int nid = numa_node_id();
70 struct page *page = NULL;
71 struct zonelist *zonelist = huge_zonelist(vma, address);
74 for (z = zonelist->zones; *z; z++) {
75 nid = (*z)->zone_pgdat->node_id;
76 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
77 !list_empty(&hugepage_freelists[nid]))
82 page = list_entry(hugepage_freelists[nid].next,
86 free_huge_pages_node[nid]--;
91 static int alloc_fresh_huge_page(void)
95 page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
97 nid = (nid + 1) % num_online_nodes();
99 page[1].lru.next = (void *)free_huge_page; /* dtor */
100 spin_lock(&hugetlb_lock);
102 nr_huge_pages_node[page_to_nid(page)]++;
103 spin_unlock(&hugetlb_lock);
104 put_page(page); /* free it into the hugepage allocator */
110 void free_huge_page(struct page *page)
112 BUG_ON(page_count(page));
114 INIT_LIST_HEAD(&page->lru);
116 spin_lock(&hugetlb_lock);
117 enqueue_huge_page(page);
118 spin_unlock(&hugetlb_lock);
121 struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
125 spin_lock(&hugetlb_lock);
126 page = dequeue_huge_page(vma, addr);
128 spin_unlock(&hugetlb_lock);
131 spin_unlock(&hugetlb_lock);
132 set_page_refcounted(page);
136 static int __init hugetlb_init(void)
140 if (HPAGE_SHIFT == 0)
143 for (i = 0; i < MAX_NUMNODES; ++i)
144 INIT_LIST_HEAD(&hugepage_freelists[i]);
146 for (i = 0; i < max_huge_pages; ++i) {
147 if (!alloc_fresh_huge_page())
150 max_huge_pages = free_huge_pages = nr_huge_pages = i;
151 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
154 module_init(hugetlb_init);
156 static int __init hugetlb_setup(char *s)
158 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
162 __setup("hugepages=", hugetlb_setup);
165 static void update_and_free_page(struct page *page)
169 nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
170 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
171 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
172 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
173 1 << PG_private | 1<< PG_writeback);
175 page[1].lru.next = NULL;
176 set_page_refcounted(page);
177 __free_pages(page, HUGETLB_PAGE_ORDER);
180 #ifdef CONFIG_HIGHMEM
181 static void try_to_free_low(unsigned long count)
184 for (i = 0; i < MAX_NUMNODES; ++i) {
185 struct page *page, *next;
186 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
187 if (PageHighMem(page))
189 list_del(&page->lru);
190 update_and_free_page(page);
191 nid = page_zone(page)->zone_pgdat->node_id;
193 free_huge_pages_node[nid]--;
194 if (count >= nr_huge_pages)
200 static inline void try_to_free_low(unsigned long count)
205 static unsigned long set_max_huge_pages(unsigned long count)
207 while (count > nr_huge_pages) {
208 if (!alloc_fresh_huge_page())
209 return nr_huge_pages;
211 if (count >= nr_huge_pages)
212 return nr_huge_pages;
214 spin_lock(&hugetlb_lock);
215 try_to_free_low(count);
216 while (count < nr_huge_pages) {
217 struct page *page = dequeue_huge_page(NULL, 0);
220 update_and_free_page(page);
222 spin_unlock(&hugetlb_lock);
223 return nr_huge_pages;
226 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
227 struct file *file, void __user *buffer,
228 size_t *length, loff_t *ppos)
230 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
231 max_huge_pages = set_max_huge_pages(max_huge_pages);
234 #endif /* CONFIG_SYSCTL */
236 int hugetlb_report_meminfo(char *buf)
239 "HugePages_Total: %5lu\n"
240 "HugePages_Free: %5lu\n"
241 "Hugepagesize: %5lu kB\n",
247 int hugetlb_report_node_meminfo(int nid, char *buf)
250 "Node %d HugePages_Total: %5u\n"
251 "Node %d HugePages_Free: %5u\n",
252 nid, nr_huge_pages_node[nid],
253 nid, free_huge_pages_node[nid]);
256 int is_hugepage_mem_enough(size_t size)
258 return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
261 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
262 unsigned long hugetlb_total_pages(void)
264 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
268 * We cannot handle pagefaults against hugetlb pages at all. They cause
269 * handle_mm_fault() to try to instantiate regular-sized pages in the
270 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
273 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
274 unsigned long address, int *unused)
280 struct vm_operations_struct hugetlb_vm_ops = {
281 .nopage = hugetlb_nopage,
284 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
291 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
293 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
295 entry = pte_mkyoung(entry);
296 entry = pte_mkhuge(entry);
301 static void set_huge_ptep_writable(struct vm_area_struct *vma,
302 unsigned long address, pte_t *ptep)
306 entry = pte_mkwrite(pte_mkdirty(*ptep));
307 ptep_set_access_flags(vma, address, ptep, entry, 1);
308 update_mmu_cache(vma, address, entry);
309 lazy_mmu_prot_update(entry);
313 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
314 struct vm_area_struct *vma)
316 pte_t *src_pte, *dst_pte, entry;
317 struct page *ptepage;
321 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
323 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
324 src_pte = huge_pte_offset(src, addr);
327 dst_pte = huge_pte_alloc(dst, addr);
330 spin_lock(&dst->page_table_lock);
331 spin_lock(&src->page_table_lock);
332 if (!pte_none(*src_pte)) {
334 ptep_set_wrprotect(src, addr, src_pte);
336 ptepage = pte_page(entry);
338 add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
339 set_huge_pte_at(dst, addr, dst_pte, entry);
341 spin_unlock(&src->page_table_lock);
342 spin_unlock(&dst->page_table_lock);
350 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
353 struct mm_struct *mm = vma->vm_mm;
354 unsigned long address;
359 WARN_ON(!is_vm_hugetlb_page(vma));
360 BUG_ON(start & ~HPAGE_MASK);
361 BUG_ON(end & ~HPAGE_MASK);
363 spin_lock(&mm->page_table_lock);
365 /* Update high watermark before we lower rss */
366 update_hiwater_rss(mm);
368 for (address = start; address < end; address += HPAGE_SIZE) {
369 ptep = huge_pte_offset(mm, address);
373 pte = huge_ptep_get_and_clear(mm, address, ptep);
377 page = pte_page(pte);
379 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
382 spin_unlock(&mm->page_table_lock);
383 flush_tlb_range(vma, start, end);
386 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
387 unsigned long address, pte_t *ptep, pte_t pte)
389 struct page *old_page, *new_page;
392 old_page = pte_page(pte);
394 /* If no-one else is actually using this page, avoid the copy
395 * and just make the page writable */
396 avoidcopy = (page_count(old_page) == 1);
398 set_huge_ptep_writable(vma, address, ptep);
399 return VM_FAULT_MINOR;
402 page_cache_get(old_page);
403 new_page = alloc_huge_page(vma, address);
406 page_cache_release(old_page);
410 spin_unlock(&mm->page_table_lock);
411 copy_huge_page(new_page, old_page, address);
412 spin_lock(&mm->page_table_lock);
414 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
415 if (likely(pte_same(*ptep, pte))) {
417 set_huge_pte_at(mm, address, ptep,
418 make_huge_pte(vma, new_page, 1));
419 /* Make the old page be freed below */
422 page_cache_release(new_page);
423 page_cache_release(old_page);
424 return VM_FAULT_MINOR;
427 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
428 unsigned long address, pte_t *ptep, int write_access)
430 int ret = VM_FAULT_SIGBUS;
434 struct address_space *mapping;
437 mapping = vma->vm_file->f_mapping;
438 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
439 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
442 * Use page lock to guard against racing truncation
443 * before we get page_table_lock.
446 page = find_lock_page(mapping, idx);
448 if (hugetlb_get_quota(mapping))
450 page = alloc_huge_page(vma, address);
452 hugetlb_put_quota(mapping);
456 clear_huge_page(page, address);
458 if (vma->vm_flags & VM_SHARED) {
461 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
464 hugetlb_put_quota(mapping);
473 spin_lock(&mm->page_table_lock);
474 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
478 ret = VM_FAULT_MINOR;
479 if (!pte_none(*ptep))
482 add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
483 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
484 && (vma->vm_flags & VM_SHARED)));
485 set_huge_pte_at(mm, address, ptep, new_pte);
487 if (write_access && !(vma->vm_flags & VM_SHARED)) {
488 /* Optimization, do the COW without a second fault */
489 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
492 spin_unlock(&mm->page_table_lock);
498 spin_unlock(&mm->page_table_lock);
499 hugetlb_put_quota(mapping);
505 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
506 unsigned long address, int write_access)
512 ptep = huge_pte_alloc(mm, address);
518 return hugetlb_no_page(mm, vma, address, ptep, write_access);
520 ret = VM_FAULT_MINOR;
522 spin_lock(&mm->page_table_lock);
523 /* Check for a racing update before calling hugetlb_cow */
524 if (likely(pte_same(entry, *ptep)))
525 if (write_access && !pte_write(entry))
526 ret = hugetlb_cow(mm, vma, address, ptep, entry);
527 spin_unlock(&mm->page_table_lock);
532 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
533 struct page **pages, struct vm_area_struct **vmas,
534 unsigned long *position, int *length, int i)
536 unsigned long vpfn, vaddr = *position;
537 int remainder = *length;
539 vpfn = vaddr/PAGE_SIZE;
540 spin_lock(&mm->page_table_lock);
541 while (vaddr < vma->vm_end && remainder) {
546 * Some archs (sparc64, sh*) have multiple pte_ts to
547 * each hugepage. We have to make * sure we get the
548 * first, for the page indexing below to work.
550 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
552 if (!pte || pte_none(*pte)) {
555 spin_unlock(&mm->page_table_lock);
556 ret = hugetlb_fault(mm, vma, vaddr, 0);
557 spin_lock(&mm->page_table_lock);
558 if (ret == VM_FAULT_MINOR)
568 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
581 spin_unlock(&mm->page_table_lock);
588 void hugetlb_change_protection(struct vm_area_struct *vma,
589 unsigned long address, unsigned long end, pgprot_t newprot)
591 struct mm_struct *mm = vma->vm_mm;
592 unsigned long start = address;
596 BUG_ON(address >= end);
597 flush_cache_range(vma, address, end);
599 spin_lock(&mm->page_table_lock);
600 for (; address < end; address += HPAGE_SIZE) {
601 ptep = huge_pte_offset(mm, address);
604 if (!pte_none(*ptep)) {
605 pte = huge_ptep_get_and_clear(mm, address, ptep);
606 pte = pte_mkhuge(pte_modify(pte, newprot));
607 set_huge_pte_at(mm, address, ptep, pte);
608 lazy_mmu_prot_update(pte);
611 spin_unlock(&mm->page_table_lock);
613 flush_tlb_range(vma, start, end);