blob: ba029d6407402c35aac9cdef714e6a60d6662096 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
4 */
5#include <linux/gfp.h>
6#include <linux/list.h>
7#include <linux/init.h>
8#include <linux/module.h>
9#include <linux/mm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
12#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070013#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080014#include <linux/mempolicy.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080015#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080016#include <linux/mutex.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080017
David Gibson63551ae2005-06-21 17:14:44 -070018#include <asm/page.h>
19#include <asm/pgtable.h>
20
21#include <linux/hugetlb.h>
Nick Piggin7835e982006-03-22 00:08:40 -080022#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
24const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -070025static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -070026unsigned long max_huge_pages;
27static struct list_head hugepage_freelists[MAX_NUMNODES];
28static unsigned int nr_huge_pages_node[MAX_NUMNODES];
29static unsigned int free_huge_pages_node[MAX_NUMNODES];
Mel Gorman396faf02007-07-17 04:03:13 -070030static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31unsigned long hugepages_treat_as_movable;
32
David Gibson3935baa2006-03-22 00:08:53 -080033/*
34 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
35 */
36static DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080037
David Gibson79ac6ba2006-03-22 00:08:51 -080038static void clear_huge_page(struct page *page, unsigned long addr)
39{
40 int i;
41
42 might_sleep();
43 for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); i++) {
44 cond_resched();
Ralf Baechle281e0e32007-10-01 01:20:10 -070045 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
David Gibson79ac6ba2006-03-22 00:08:51 -080046 }
47}
48
49static void copy_huge_page(struct page *dst, struct page *src,
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000050 unsigned long addr, struct vm_area_struct *vma)
David Gibson79ac6ba2006-03-22 00:08:51 -080051{
52 int i;
53
54 might_sleep();
55 for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++) {
56 cond_resched();
Atsushi Nemoto9de455b2006-12-12 17:14:55 +000057 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
David Gibson79ac6ba2006-03-22 00:08:51 -080058 }
59}
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061static void enqueue_huge_page(struct page *page)
62{
63 int nid = page_to_nid(page);
64 list_add(&page->lru, &hugepage_freelists[nid]);
65 free_huge_pages++;
66 free_huge_pages_node[nid]++;
67}
68
Christoph Lameter5da7ca82006-01-06 00:10:46 -080069static struct page *dequeue_huge_page(struct vm_area_struct *vma,
70 unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -070071{
Nishanth Aravamudan31a5c6e2007-07-15 23:38:02 -070072 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 struct page *page = NULL;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070074 struct mempolicy *mpol;
Mel Gorman396faf02007-07-17 04:03:13 -070075 struct zonelist *zonelist = huge_zonelist(vma, address,
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070076 htlb_alloc_mask, &mpol);
Christoph Lameter96df9332006-01-06 00:10:45 -080077 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Christoph Lameter96df9332006-01-06 00:10:45 -080079 for (z = zonelist->zones; *z; z++) {
Christoph Lameter89fa3022006-09-25 23:31:55 -070080 nid = zone_to_nid(*z);
Mel Gorman396faf02007-07-17 04:03:13 -070081 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
Andrew Morton3abf7af2007-07-19 01:49:08 -070082 !list_empty(&hugepage_freelists[nid])) {
83 page = list_entry(hugepage_freelists[nid].next,
84 struct page, lru);
85 list_del(&page->lru);
86 free_huge_pages--;
87 free_huge_pages_node[nid]--;
Ken Chen5ab3ee72007-07-23 18:44:00 -070088 break;
Andrew Morton3abf7af2007-07-19 01:49:08 -070089 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070090 }
Lee Schermerhorn480eccf2007-09-18 22:46:47 -070091 mpol_free(mpol); /* unref if mpol !NULL */
Linus Torvalds1da177e2005-04-16 15:20:36 -070092 return page;
93}
94
Adam Litke6af2acb2007-10-16 01:26:16 -070095static void update_and_free_page(struct page *page)
96{
97 int i;
98 nr_huge_pages--;
99 nr_huge_pages_node[page_to_nid(page)]--;
100 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
101 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
102 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
103 1 << PG_private | 1<< PG_writeback);
104 }
105 set_compound_page_dtor(page, NULL);
106 set_page_refcounted(page);
107 __free_pages(page, HUGETLB_PAGE_ORDER);
108}
109
David Gibson27a85ef2006-03-22 00:08:56 -0800110static void free_huge_page(struct page *page)
111{
112 BUG_ON(page_count(page));
113
114 INIT_LIST_HEAD(&page->lru);
115
116 spin_lock(&hugetlb_lock);
117 enqueue_huge_page(page);
118 spin_unlock(&hugetlb_lock);
119}
120
Nick Piggina4822892006-03-22 00:08:08 -0800121static int alloc_fresh_huge_page(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122{
Joe Jinf96efd52007-07-15 23:38:12 -0700123 static int prev_nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700125 int nid;
126
Hugh Dickins7ed5cb22007-07-19 01:49:11 -0700127 /*
128 * Copy static prev_nid to local nid, work on that, then copy it
129 * back to prev_nid afterwards: otherwise there's a window in which
130 * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
131 * But we don't need to use a spin_lock here: it really doesn't
132 * matter if occasionally a racer chooses the same nid as we do.
133 */
Joe Jinf96efd52007-07-15 23:38:12 -0700134 nid = next_node(prev_nid, node_online_map);
Paul Jacksonfdb7cc52006-03-22 00:09:10 -0800135 if (nid == MAX_NUMNODES)
136 nid = first_node(node_online_map);
Joe Jinf96efd52007-07-15 23:38:12 -0700137 prev_nid = nid;
Joe Jinf96efd52007-07-15 23:38:12 -0700138
Mel Gorman396faf02007-07-17 04:03:13 -0700139 page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
Joe Jinf96efd52007-07-15 23:38:12 -0700140 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800142 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800143 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 nr_huge_pages++;
145 nr_huge_pages_node[page_to_nid(page)]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800146 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800147 put_page(page); /* free it into the hugepage allocator */
148 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 }
Nick Piggina4822892006-03-22 00:08:08 -0800150 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151}
152
David Gibson27a85ef2006-03-22 00:08:56 -0800153static struct page *alloc_huge_page(struct vm_area_struct *vma,
154 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155{
156 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
158 spin_lock(&hugetlb_lock);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700159 if (vma->vm_flags & VM_MAYSHARE)
160 resv_huge_pages--;
161 else if (free_huge_pages <= resv_huge_pages)
162 goto fail;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800163
164 page = dequeue_huge_page(vma, addr);
165 if (!page)
166 goto fail;
167
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 spin_unlock(&hugetlb_lock);
Nick Piggin7835e982006-03-22 00:08:40 -0800169 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800171
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700172fail:
Ken Chenace4bd22007-05-09 02:33:09 -0700173 if (vma->vm_flags & VM_MAYSHARE)
174 resv_huge_pages++;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800175 spin_unlock(&hugetlb_lock);
176 return NULL;
177}
178
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179static int __init hugetlb_init(void)
180{
181 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100183 if (HPAGE_SHIFT == 0)
184 return 0;
185
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 for (i = 0; i < MAX_NUMNODES; ++i)
187 INIT_LIST_HEAD(&hugepage_freelists[i]);
188
189 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800190 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 }
193 max_huge_pages = free_huge_pages = nr_huge_pages = i;
194 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
195 return 0;
196}
197module_init(hugetlb_init);
198
199static int __init hugetlb_setup(char *s)
200{
201 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
202 max_huge_pages = 0;
203 return 1;
204}
205__setup("hugepages=", hugetlb_setup);
206
Ken Chen8a630112007-05-09 02:33:34 -0700207static unsigned int cpuset_mems_nr(unsigned int *array)
208{
209 int node;
210 unsigned int nr = 0;
211
212 for_each_node_mask(node, cpuset_current_mems_allowed)
213 nr += array[node];
214
215 return nr;
216}
217
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218#ifdef CONFIG_SYSCTL
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219#ifdef CONFIG_HIGHMEM
220static void try_to_free_low(unsigned long count)
221{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700222 int i;
223
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 for (i = 0; i < MAX_NUMNODES; ++i) {
225 struct page *page, *next;
226 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
227 if (PageHighMem(page))
228 continue;
229 list_del(&page->lru);
230 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700232 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 if (count >= nr_huge_pages)
234 return;
235 }
236 }
237}
238#else
239static inline void try_to_free_low(unsigned long count)
240{
241}
242#endif
243
244static unsigned long set_max_huge_pages(unsigned long count)
245{
246 while (count > nr_huge_pages) {
Nick Piggina4822892006-03-22 00:08:08 -0800247 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 return nr_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 }
250 if (count >= nr_huge_pages)
251 return nr_huge_pages;
252
253 spin_lock(&hugetlb_lock);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700254 count = max(count, resv_huge_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 try_to_free_low(count);
256 while (count < nr_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800257 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 if (!page)
259 break;
260 update_and_free_page(page);
261 }
262 spin_unlock(&hugetlb_lock);
263 return nr_huge_pages;
264}
265
266int hugetlb_sysctl_handler(struct ctl_table *table, int write,
267 struct file *file, void __user *buffer,
268 size_t *length, loff_t *ppos)
269{
270 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
271 max_huge_pages = set_max_huge_pages(max_huge_pages);
272 return 0;
273}
Mel Gorman396faf02007-07-17 04:03:13 -0700274
275int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
276 struct file *file, void __user *buffer,
277 size_t *length, loff_t *ppos)
278{
279 proc_dointvec(table, write, file, buffer, length, ppos);
280 if (hugepages_treat_as_movable)
281 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
282 else
283 htlb_alloc_mask = GFP_HIGHUSER;
284 return 0;
285}
286
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287#endif /* CONFIG_SYSCTL */
288
289int hugetlb_report_meminfo(char *buf)
290{
291 return sprintf(buf,
292 "HugePages_Total: %5lu\n"
293 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700294 "HugePages_Rsvd: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 "Hugepagesize: %5lu kB\n",
296 nr_huge_pages,
297 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700298 resv_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 HPAGE_SIZE/1024);
300}
301
302int hugetlb_report_node_meminfo(int nid, char *buf)
303{
304 return sprintf(buf,
305 "Node %d HugePages_Total: %5u\n"
306 "Node %d HugePages_Free: %5u\n",
307 nid, nr_huge_pages_node[nid],
308 nid, free_huge_pages_node[nid]);
309}
310
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
312unsigned long hugetlb_total_pages(void)
313{
314 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
315}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
317/*
318 * We cannot handle pagefaults against hugetlb pages at all. They cause
319 * handle_mm_fault() to try to instantiate regular-sized pages in the
320 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
321 * this far.
322 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700323static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324{
325 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700326 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327}
328
329struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700330 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331};
332
David Gibson1e8f8892006-01-06 00:10:44 -0800333static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
334 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700335{
336 pte_t entry;
337
David Gibson1e8f8892006-01-06 00:10:44 -0800338 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700339 entry =
340 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
341 } else {
342 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
343 }
344 entry = pte_mkyoung(entry);
345 entry = pte_mkhuge(entry);
346
347 return entry;
348}
349
David Gibson1e8f8892006-01-06 00:10:44 -0800350static void set_huge_ptep_writable(struct vm_area_struct *vma,
351 unsigned long address, pte_t *ptep)
352{
353 pte_t entry;
354
355 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700356 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
357 update_mmu_cache(vma, address, entry);
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700358 }
David Gibson1e8f8892006-01-06 00:10:44 -0800359}
360
361
David Gibson63551ae2005-06-21 17:14:44 -0700362int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
363 struct vm_area_struct *vma)
364{
365 pte_t *src_pte, *dst_pte, entry;
366 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700367 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800368 int cow;
369
370 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700371
Hugh Dickins1c598272005-10-19 21:23:43 -0700372 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700373 src_pte = huge_pte_offset(src, addr);
374 if (!src_pte)
375 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700376 dst_pte = huge_pte_alloc(dst, addr);
377 if (!dst_pte)
378 goto nomem;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700379 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700380 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700381 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800382 if (cow)
383 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700384 entry = *src_pte;
385 ptepage = pte_page(entry);
386 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700387 set_huge_pte_at(dst, addr, dst_pte, entry);
388 }
389 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700390 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700391 }
392 return 0;
393
394nomem:
395 return -ENOMEM;
396}
397
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700398void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
399 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700400{
401 struct mm_struct *mm = vma->vm_mm;
402 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700403 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700404 pte_t pte;
405 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700406 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800407 /*
408 * A page gathering list, protected by per file i_mmap_lock. The
409 * lock is used to avoid list corruption from multiple unmapping
410 * of the same page since we are using page->lru.
411 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700412 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700413
414 WARN_ON(!is_vm_hugetlb_page(vma));
415 BUG_ON(start & ~HPAGE_MASK);
416 BUG_ON(end & ~HPAGE_MASK);
417
Hugh Dickins508034a2005-10-29 18:16:30 -0700418 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700419 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700420 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700421 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700422 continue;
423
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800424 if (huge_pmd_unshare(mm, &address, ptep))
425 continue;
426
David Gibsonc7546f82005-08-05 11:59:35 -0700427 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700428 if (pte_none(pte))
429 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700430
David Gibson63551ae2005-06-21 17:14:44 -0700431 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800432 if (pte_dirty(pte))
433 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700434 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700435 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700437 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700438 list_for_each_entry_safe(page, tmp, &page_list, lru) {
439 list_del(&page->lru);
440 put_page(page);
441 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442}
David Gibson63551ae2005-06-21 17:14:44 -0700443
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700444void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
445 unsigned long end)
446{
447 /*
448 * It is undesirable to test vma->vm_file as it should be non-null
449 * for valid hugetlb area. However, vm_file will be NULL in the error
450 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
451 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
452 * to clean up. Since no pte has actually been setup, it is safe to
453 * do nothing in this case.
454 */
455 if (vma->vm_file) {
456 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
457 __unmap_hugepage_range(vma, start, end);
458 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
459 }
460}
461
David Gibson1e8f8892006-01-06 00:10:44 -0800462static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
463 unsigned long address, pte_t *ptep, pte_t pte)
464{
465 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800466 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800467
468 old_page = pte_page(pte);
469
470 /* If no-one else is actually using this page, avoid the copy
471 * and just make the page writable */
472 avoidcopy = (page_count(old_page) == 1);
473 if (avoidcopy) {
474 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700475 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800476 }
477
478 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800479 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800480
481 if (!new_page) {
482 page_cache_release(old_page);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800483 return VM_FAULT_OOM;
David Gibson1e8f8892006-01-06 00:10:44 -0800484 }
485
486 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000487 copy_huge_page(new_page, old_page, address, vma);
David Gibson1e8f8892006-01-06 00:10:44 -0800488 spin_lock(&mm->page_table_lock);
489
490 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
491 if (likely(pte_same(*ptep, pte))) {
492 /* Break COW */
493 set_huge_pte_at(mm, address, ptep,
494 make_huge_pte(vma, new_page, 1));
495 /* Make the old page be freed below */
496 new_page = old_page;
497 }
498 page_cache_release(new_page);
499 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700500 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800501}
502
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700503static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800504 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100505{
506 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700507 unsigned long idx;
508 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700509 struct page *page;
510 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800511 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700512
Adam Litke4c887262005-10-29 18:16:46 -0700513 mapping = vma->vm_file->f_mapping;
514 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
515 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
516
517 /*
518 * Use page lock to guard against racing truncation
519 * before we get page_table_lock.
520 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800521retry:
522 page = find_lock_page(mapping, idx);
523 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700524 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
525 if (idx >= size)
526 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800527 if (hugetlb_get_quota(mapping))
528 goto out;
529 page = alloc_huge_page(vma, address);
530 if (!page) {
531 hugetlb_put_quota(mapping);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800532 ret = VM_FAULT_OOM;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800533 goto out;
534 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800535 clear_huge_page(page, address);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100536
Christoph Lameter6bda6662006-01-06 00:10:49 -0800537 if (vma->vm_flags & VM_SHARED) {
538 int err;
539
540 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
541 if (err) {
542 put_page(page);
543 hugetlb_put_quota(mapping);
544 if (err == -EEXIST)
545 goto retry;
546 goto out;
547 }
548 } else
549 lock_page(page);
550 }
David Gibson1e8f8892006-01-06 00:10:44 -0800551
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100552 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700553 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
554 if (idx >= size)
555 goto backout;
556
Nick Piggin83c54072007-07-19 01:47:05 -0700557 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800558 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700559 goto backout;
560
David Gibson1e8f8892006-01-06 00:10:44 -0800561 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
562 && (vma->vm_flags & VM_SHARED)));
563 set_huge_pte_at(mm, address, ptep, new_pte);
564
565 if (write_access && !(vma->vm_flags & VM_SHARED)) {
566 /* Optimization, do the COW without a second fault */
567 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
568 }
569
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100570 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700571 unlock_page(page);
572out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100573 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700574
575backout:
576 spin_unlock(&mm->page_table_lock);
577 hugetlb_put_quota(mapping);
578 unlock_page(page);
579 put_page(page);
580 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100581}
582
Adam Litke86e52162006-01-06 00:10:43 -0800583int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
584 unsigned long address, int write_access)
585{
586 pte_t *ptep;
587 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800588 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800589 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800590
591 ptep = huge_pte_alloc(mm, address);
592 if (!ptep)
593 return VM_FAULT_OOM;
594
David Gibson3935baa2006-03-22 00:08:53 -0800595 /*
596 * Serialize hugepage allocation and instantiation, so that we don't
597 * get spurious allocation failures if two CPUs race to instantiate
598 * the same page in the page cache.
599 */
600 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800601 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800602 if (pte_none(entry)) {
603 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
604 mutex_unlock(&hugetlb_instantiation_mutex);
605 return ret;
606 }
Adam Litke86e52162006-01-06 00:10:43 -0800607
Nick Piggin83c54072007-07-19 01:47:05 -0700608 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800609
610 spin_lock(&mm->page_table_lock);
611 /* Check for a racing update before calling hugetlb_cow */
612 if (likely(pte_same(entry, *ptep)))
613 if (write_access && !pte_write(entry))
614 ret = hugetlb_cow(mm, vma, address, ptep, entry);
615 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800616 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800617
618 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800619}
620
David Gibson63551ae2005-06-21 17:14:44 -0700621int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
622 struct page **pages, struct vm_area_struct **vmas,
623 unsigned long *position, int *length, int i)
624{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800625 unsigned long pfn_offset;
626 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700627 int remainder = *length;
628
Hugh Dickins1c598272005-10-19 21:23:43 -0700629 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700630 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700631 pte_t *pte;
632 struct page *page;
633
634 /*
635 * Some archs (sparc64, sh*) have multiple pte_ts to
636 * each hugepage. We have to make * sure we get the
637 * first, for the page indexing below to work.
638 */
639 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
640
641 if (!pte || pte_none(*pte)) {
642 int ret;
643
644 spin_unlock(&mm->page_table_lock);
645 ret = hugetlb_fault(mm, vma, vaddr, 0);
646 spin_lock(&mm->page_table_lock);
Adam Litkea89182c2007-08-22 14:01:51 -0700647 if (!(ret & VM_FAULT_ERROR))
Adam Litke4c887262005-10-29 18:16:46 -0700648 continue;
649
650 remainder = 0;
651 if (!i)
652 i = -EFAULT;
653 break;
654 }
David Gibson63551ae2005-06-21 17:14:44 -0700655
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800656 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
657 page = pte_page(*pte);
658same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800659 if (pages) {
660 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800661 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800662 }
David Gibson63551ae2005-06-21 17:14:44 -0700663
664 if (vmas)
665 vmas[i] = vma;
666
667 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800668 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -0700669 --remainder;
670 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800671 if (vaddr < vma->vm_end && remainder &&
672 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
673 /*
674 * We use pfn_offset to avoid touching the pageframes
675 * of this compound page.
676 */
677 goto same_page;
678 }
David Gibson63551ae2005-06-21 17:14:44 -0700679 }
Hugh Dickins1c598272005-10-19 21:23:43 -0700680 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700681 *length = remainder;
682 *position = vaddr;
683
684 return i;
685}
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800686
687void hugetlb_change_protection(struct vm_area_struct *vma,
688 unsigned long address, unsigned long end, pgprot_t newprot)
689{
690 struct mm_struct *mm = vma->vm_mm;
691 unsigned long start = address;
692 pte_t *ptep;
693 pte_t pte;
694
695 BUG_ON(address >= end);
696 flush_cache_range(vma, address, end);
697
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800698 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800699 spin_lock(&mm->page_table_lock);
700 for (; address < end; address += HPAGE_SIZE) {
701 ptep = huge_pte_offset(mm, address);
702 if (!ptep)
703 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800704 if (huge_pmd_unshare(mm, &address, ptep))
705 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800706 if (!pte_none(*ptep)) {
707 pte = huge_ptep_get_and_clear(mm, address, ptep);
708 pte = pte_mkhuge(pte_modify(pte, newprot));
709 set_huge_pte_at(mm, address, ptep, pte);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800710 }
711 }
712 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800713 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800714
715 flush_tlb_range(vma, start, end);
716}
717
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700718struct file_region {
719 struct list_head link;
720 long from;
721 long to;
722};
723
724static long region_add(struct list_head *head, long f, long t)
725{
726 struct file_region *rg, *nrg, *trg;
727
728 /* Locate the region we are either in or before. */
729 list_for_each_entry(rg, head, link)
730 if (f <= rg->to)
731 break;
732
733 /* Round our left edge to the current segment if it encloses us. */
734 if (f > rg->from)
735 f = rg->from;
736
737 /* Check for and consume any regions we now overlap with. */
738 nrg = rg;
739 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
740 if (&rg->link == head)
741 break;
742 if (rg->from > t)
743 break;
744
745 /* If this area reaches higher then extend our area to
746 * include it completely. If this is not the first area
747 * which we intend to reuse, free it. */
748 if (rg->to > t)
749 t = rg->to;
750 if (rg != nrg) {
751 list_del(&rg->link);
752 kfree(rg);
753 }
754 }
755 nrg->from = f;
756 nrg->to = t;
757 return 0;
758}
759
760static long region_chg(struct list_head *head, long f, long t)
761{
762 struct file_region *rg, *nrg;
763 long chg = 0;
764
765 /* Locate the region we are before or in. */
766 list_for_each_entry(rg, head, link)
767 if (f <= rg->to)
768 break;
769
770 /* If we are below the current region then a new region is required.
771 * Subtle, allocate a new region at the position but make it zero
772 * size such that we can guarentee to record the reservation. */
773 if (&rg->link == head || t < rg->from) {
774 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
775 if (nrg == 0)
776 return -ENOMEM;
777 nrg->from = f;
778 nrg->to = f;
779 INIT_LIST_HEAD(&nrg->link);
780 list_add(&nrg->link, rg->link.prev);
781
782 return t - f;
783 }
784
785 /* Round our left edge to the current segment if it encloses us. */
786 if (f > rg->from)
787 f = rg->from;
788 chg = t - f;
789
790 /* Check for and consume any regions we now overlap with. */
791 list_for_each_entry(rg, rg->link.prev, link) {
792 if (&rg->link == head)
793 break;
794 if (rg->from > t)
795 return chg;
796
797 /* We overlap with this area, if it extends futher than
798 * us then we must extend ourselves. Account for its
799 * existing reservation. */
800 if (rg->to > t) {
801 chg += rg->to - t;
802 t = rg->to;
803 }
804 chg -= rg->to - rg->from;
805 }
806 return chg;
807}
808
809static long region_truncate(struct list_head *head, long end)
810{
811 struct file_region *rg, *trg;
812 long chg = 0;
813
814 /* Locate the region we are either in or before. */
815 list_for_each_entry(rg, head, link)
816 if (end <= rg->to)
817 break;
818 if (&rg->link == head)
819 return 0;
820
821 /* If we are in the middle of a region then adjust it. */
822 if (end > rg->from) {
823 chg = rg->to - end;
824 rg->to = end;
825 rg = list_entry(rg->link.next, typeof(*rg), link);
826 }
827
828 /* Drop any remaining regions. */
829 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
830 if (&rg->link == head)
831 break;
832 chg += rg->to - rg->from;
833 list_del(&rg->link);
834 kfree(rg);
835 }
836 return chg;
837}
838
839static int hugetlb_acct_memory(long delta)
840{
841 int ret = -ENOMEM;
842
843 spin_lock(&hugetlb_lock);
844 if ((delta + resv_huge_pages) <= free_huge_pages) {
845 resv_huge_pages += delta;
846 ret = 0;
847 }
848 spin_unlock(&hugetlb_lock);
849 return ret;
850}
851
852int hugetlb_reserve_pages(struct inode *inode, long from, long to)
853{
854 long ret, chg;
855
856 chg = region_chg(&inode->i_mapping->private_list, from, to);
857 if (chg < 0)
858 return chg;
Ken Chen8a630112007-05-09 02:33:34 -0700859 /*
860 * When cpuset is configured, it breaks the strict hugetlb page
861 * reservation as the accounting is done on a global variable. Such
862 * reservation is completely rubbish in the presence of cpuset because
863 * the reservation is not checked against page availability for the
864 * current cpuset. Application can still potentially OOM'ed by kernel
865 * with lack of free htlb page in cpuset that the task is in.
866 * Attempt to enforce strict accounting with cpuset is almost
867 * impossible (or too ugly) because cpuset is too fluid that
868 * task or memory node can be dynamically moved between cpusets.
869 *
870 * The change of semantics for shared hugetlb mapping with cpuset is
871 * undesirable. However, in order to preserve some of the semantics,
872 * we fall back to check against current free page availability as
873 * a best attempt and hopefully to minimize the impact of changing
874 * semantics that cpuset has.
875 */
876 if (chg > cpuset_mems_nr(free_huge_pages_node))
877 return -ENOMEM;
878
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700879 ret = hugetlb_acct_memory(chg);
880 if (ret < 0)
881 return ret;
882 region_add(&inode->i_mapping->private_list, from, to);
883 return 0;
884}
885
886void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
887{
888 long chg = region_truncate(&inode->i_mapping->private_list, offset);
889 hugetlb_acct_memory(freed - chg);
890}