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>
22 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
23 static unsigned long nr_huge_pages, free_huge_pages;
24 unsigned long max_huge_pages;
25 static struct list_head hugepage_freelists[MAX_NUMNODES];
26 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
27 static unsigned int free_huge_pages_node[MAX_NUMNODES];
30 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
32 static DEFINE_SPINLOCK(hugetlb_lock);
34 static void enqueue_huge_page(struct page *page)
36 int nid = page_to_nid(page);
37 list_add(&page->lru, &hugepage_freelists[nid]);
39 free_huge_pages_node[nid]++;
42 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
43 unsigned long address)
45 int nid = numa_node_id();
46 struct page *page = NULL;
47 struct zonelist *zonelist = huge_zonelist(vma, address);
50 for (z = zonelist->zones; *z; z++) {
51 nid = (*z)->zone_pgdat->node_id;
52 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
53 !list_empty(&hugepage_freelists[nid]))
58 page = list_entry(hugepage_freelists[nid].next,
62 free_huge_pages_node[nid]--;
67 static int alloc_fresh_huge_page(void)
71 page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
73 nid = (nid + 1) % num_online_nodes();
75 page[1].lru.next = (void *)free_huge_page; /* dtor */
76 spin_lock(&hugetlb_lock);
78 nr_huge_pages_node[page_to_nid(page)]++;
79 spin_unlock(&hugetlb_lock);
80 put_page(page); /* free it into the hugepage allocator */
86 void free_huge_page(struct page *page)
88 BUG_ON(page_count(page));
90 INIT_LIST_HEAD(&page->lru);
92 spin_lock(&hugetlb_lock);
93 enqueue_huge_page(page);
94 spin_unlock(&hugetlb_lock);
97 struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
102 spin_lock(&hugetlb_lock);
103 page = dequeue_huge_page(vma, addr);
105 spin_unlock(&hugetlb_lock);
108 spin_unlock(&hugetlb_lock);
109 set_page_count(page, 1);
110 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
111 clear_user_highpage(&page[i], addr);
115 static int __init hugetlb_init(void)
119 if (HPAGE_SHIFT == 0)
122 for (i = 0; i < MAX_NUMNODES; ++i)
123 INIT_LIST_HEAD(&hugepage_freelists[i]);
125 for (i = 0; i < max_huge_pages; ++i) {
126 if (!alloc_fresh_huge_page())
129 max_huge_pages = free_huge_pages = nr_huge_pages = i;
130 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
133 module_init(hugetlb_init);
135 static int __init hugetlb_setup(char *s)
137 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
141 __setup("hugepages=", hugetlb_setup);
144 static void update_and_free_page(struct page *page)
148 nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
149 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
150 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
151 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
152 1 << PG_private | 1<< PG_writeback);
154 page[1].lru.next = NULL;
155 set_page_count(page, 1);
156 __free_pages(page, HUGETLB_PAGE_ORDER);
159 #ifdef CONFIG_HIGHMEM
160 static void try_to_free_low(unsigned long count)
163 for (i = 0; i < MAX_NUMNODES; ++i) {
164 struct page *page, *next;
165 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
166 if (PageHighMem(page))
168 list_del(&page->lru);
169 update_and_free_page(page);
170 nid = page_zone(page)->zone_pgdat->node_id;
172 free_huge_pages_node[nid]--;
173 if (count >= nr_huge_pages)
179 static inline void try_to_free_low(unsigned long count)
184 static unsigned long set_max_huge_pages(unsigned long count)
186 while (count > nr_huge_pages) {
187 if (!alloc_fresh_huge_page())
188 return nr_huge_pages;
190 if (count >= nr_huge_pages)
191 return nr_huge_pages;
193 spin_lock(&hugetlb_lock);
194 try_to_free_low(count);
195 while (count < nr_huge_pages) {
196 struct page *page = dequeue_huge_page(NULL, 0);
199 update_and_free_page(page);
201 spin_unlock(&hugetlb_lock);
202 return nr_huge_pages;
205 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
206 struct file *file, void __user *buffer,
207 size_t *length, loff_t *ppos)
209 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
210 max_huge_pages = set_max_huge_pages(max_huge_pages);
213 #endif /* CONFIG_SYSCTL */
215 int hugetlb_report_meminfo(char *buf)
218 "HugePages_Total: %5lu\n"
219 "HugePages_Free: %5lu\n"
220 "Hugepagesize: %5lu kB\n",
226 int hugetlb_report_node_meminfo(int nid, char *buf)
229 "Node %d HugePages_Total: %5u\n"
230 "Node %d HugePages_Free: %5u\n",
231 nid, nr_huge_pages_node[nid],
232 nid, free_huge_pages_node[nid]);
235 int is_hugepage_mem_enough(size_t size)
237 return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
240 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
241 unsigned long hugetlb_total_pages(void)
243 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
247 * We cannot handle pagefaults against hugetlb pages at all. They cause
248 * handle_mm_fault() to try to instantiate regular-sized pages in the
249 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
252 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
253 unsigned long address, int *unused)
259 struct vm_operations_struct hugetlb_vm_ops = {
260 .nopage = hugetlb_nopage,
263 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
270 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
272 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
274 entry = pte_mkyoung(entry);
275 entry = pte_mkhuge(entry);
280 static void set_huge_ptep_writable(struct vm_area_struct *vma,
281 unsigned long address, pte_t *ptep)
285 entry = pte_mkwrite(pte_mkdirty(*ptep));
286 ptep_set_access_flags(vma, address, ptep, entry, 1);
287 update_mmu_cache(vma, address, entry);
288 lazy_mmu_prot_update(entry);
292 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
293 struct vm_area_struct *vma)
295 pte_t *src_pte, *dst_pte, entry;
296 struct page *ptepage;
300 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
302 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
303 src_pte = huge_pte_offset(src, addr);
306 dst_pte = huge_pte_alloc(dst, addr);
309 spin_lock(&dst->page_table_lock);
310 spin_lock(&src->page_table_lock);
311 if (!pte_none(*src_pte)) {
313 ptep_set_wrprotect(src, addr, src_pte);
315 ptepage = pte_page(entry);
317 add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
318 set_huge_pte_at(dst, addr, dst_pte, entry);
320 spin_unlock(&src->page_table_lock);
321 spin_unlock(&dst->page_table_lock);
329 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
332 struct mm_struct *mm = vma->vm_mm;
333 unsigned long address;
338 WARN_ON(!is_vm_hugetlb_page(vma));
339 BUG_ON(start & ~HPAGE_MASK);
340 BUG_ON(end & ~HPAGE_MASK);
342 spin_lock(&mm->page_table_lock);
344 /* Update high watermark before we lower rss */
345 update_hiwater_rss(mm);
347 for (address = start; address < end; address += HPAGE_SIZE) {
348 ptep = huge_pte_offset(mm, address);
352 pte = huge_ptep_get_and_clear(mm, address, ptep);
356 page = pte_page(pte);
358 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
361 spin_unlock(&mm->page_table_lock);
362 flush_tlb_range(vma, start, end);
365 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
366 unsigned long address, pte_t *ptep, pte_t pte)
368 struct page *old_page, *new_page;
371 old_page = pte_page(pte);
373 /* If no-one else is actually using this page, avoid the copy
374 * and just make the page writable */
375 avoidcopy = (page_count(old_page) == 1);
377 set_huge_ptep_writable(vma, address, ptep);
378 return VM_FAULT_MINOR;
381 page_cache_get(old_page);
382 new_page = alloc_huge_page(vma, address);
385 page_cache_release(old_page);
389 spin_unlock(&mm->page_table_lock);
390 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++)
391 copy_user_highpage(new_page + i, old_page + i,
392 address + i*PAGE_SIZE);
393 spin_lock(&mm->page_table_lock);
395 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
396 if (likely(pte_same(*ptep, pte))) {
398 set_huge_pte_at(mm, address, ptep,
399 make_huge_pte(vma, new_page, 1));
400 /* Make the old page be freed below */
403 page_cache_release(new_page);
404 page_cache_release(old_page);
405 return VM_FAULT_MINOR;
408 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
409 unsigned long address, pte_t *ptep, int write_access)
411 int ret = VM_FAULT_SIGBUS;
415 struct address_space *mapping;
418 mapping = vma->vm_file->f_mapping;
419 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
420 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
423 * Use page lock to guard against racing truncation
424 * before we get page_table_lock.
427 page = find_lock_page(mapping, idx);
429 if (hugetlb_get_quota(mapping))
431 page = alloc_huge_page(vma, address);
433 hugetlb_put_quota(mapping);
438 if (vma->vm_flags & VM_SHARED) {
441 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
444 hugetlb_put_quota(mapping);
453 spin_lock(&mm->page_table_lock);
454 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
458 ret = VM_FAULT_MINOR;
459 if (!pte_none(*ptep))
462 add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
463 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
464 && (vma->vm_flags & VM_SHARED)));
465 set_huge_pte_at(mm, address, ptep, new_pte);
467 if (write_access && !(vma->vm_flags & VM_SHARED)) {
468 /* Optimization, do the COW without a second fault */
469 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
472 spin_unlock(&mm->page_table_lock);
478 spin_unlock(&mm->page_table_lock);
479 hugetlb_put_quota(mapping);
485 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
486 unsigned long address, int write_access)
492 ptep = huge_pte_alloc(mm, address);
498 return hugetlb_no_page(mm, vma, address, ptep, write_access);
500 ret = VM_FAULT_MINOR;
502 spin_lock(&mm->page_table_lock);
503 /* Check for a racing update before calling hugetlb_cow */
504 if (likely(pte_same(entry, *ptep)))
505 if (write_access && !pte_write(entry))
506 ret = hugetlb_cow(mm, vma, address, ptep, entry);
507 spin_unlock(&mm->page_table_lock);
512 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
513 struct page **pages, struct vm_area_struct **vmas,
514 unsigned long *position, int *length, int i)
516 unsigned long vpfn, vaddr = *position;
517 int remainder = *length;
519 vpfn = vaddr/PAGE_SIZE;
520 spin_lock(&mm->page_table_lock);
521 while (vaddr < vma->vm_end && remainder) {
526 * Some archs (sparc64, sh*) have multiple pte_ts to
527 * each hugepage. We have to make * sure we get the
528 * first, for the page indexing below to work.
530 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
532 if (!pte || pte_none(*pte)) {
535 spin_unlock(&mm->page_table_lock);
536 ret = hugetlb_fault(mm, vma, vaddr, 0);
537 spin_lock(&mm->page_table_lock);
538 if (ret == VM_FAULT_MINOR)
548 page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
561 spin_unlock(&mm->page_table_lock);