blob: f127940ec24fc8c52e6492de4e5729ce330c487b [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();
45 clear_user_highpage(page + i, addr);
46 }
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;
Mel Gorman396faf02007-07-17 04:03:13 -070074 struct zonelist *zonelist = huge_zonelist(vma, address,
75 htlb_alloc_mask);
Christoph Lameter96df9332006-01-06 00:10:45 -080076 struct zone **z;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Christoph Lameter96df9332006-01-06 00:10:45 -080078 for (z = zonelist->zones; *z; z++) {
Christoph Lameter89fa3022006-09-25 23:31:55 -070079 nid = zone_to_nid(*z);
Mel Gorman396faf02007-07-17 04:03:13 -070080 if (cpuset_zone_allowed_softwall(*z, htlb_alloc_mask) &&
Andrew Morton3abf7af2007-07-19 01:49:08 -070081 !list_empty(&hugepage_freelists[nid])) {
82 page = list_entry(hugepage_freelists[nid].next,
83 struct page, lru);
84 list_del(&page->lru);
85 free_huge_pages--;
86 free_huge_pages_node[nid]--;
87 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 }
89 return page;
90}
91
David Gibson27a85ef2006-03-22 00:08:56 -080092static void free_huge_page(struct page *page)
93{
94 BUG_ON(page_count(page));
95
96 INIT_LIST_HEAD(&page->lru);
97
98 spin_lock(&hugetlb_lock);
99 enqueue_huge_page(page);
100 spin_unlock(&hugetlb_lock);
101}
102
Nick Piggina4822892006-03-22 00:08:08 -0800103static int alloc_fresh_huge_page(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104{
Joe Jinf96efd52007-07-15 23:38:12 -0700105 static int prev_nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 struct page *page;
Joe Jinf96efd52007-07-15 23:38:12 -0700107 int nid;
108
Hugh Dickins7ed5cb22007-07-19 01:49:11 -0700109 /*
110 * Copy static prev_nid to local nid, work on that, then copy it
111 * back to prev_nid afterwards: otherwise there's a window in which
112 * a racer might pass invalid nid MAX_NUMNODES to alloc_pages_node.
113 * But we don't need to use a spin_lock here: it really doesn't
114 * matter if occasionally a racer chooses the same nid as we do.
115 */
Joe Jinf96efd52007-07-15 23:38:12 -0700116 nid = next_node(prev_nid, node_online_map);
Paul Jacksonfdb7cc52006-03-22 00:09:10 -0800117 if (nid == MAX_NUMNODES)
118 nid = first_node(node_online_map);
Joe Jinf96efd52007-07-15 23:38:12 -0700119 prev_nid = nid;
Joe Jinf96efd52007-07-15 23:38:12 -0700120
Mel Gorman396faf02007-07-17 04:03:13 -0700121 page = alloc_pages_node(nid, htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
Joe Jinf96efd52007-07-15 23:38:12 -0700122 HUGETLB_PAGE_ORDER);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 if (page) {
Andy Whitcroft33f2ef82006-12-06 20:33:32 -0800124 set_compound_page_dtor(page, free_huge_page);
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800125 spin_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126 nr_huge_pages++;
127 nr_huge_pages_node[page_to_nid(page)]++;
Eric Paris0bd0f9f2005-11-21 21:32:28 -0800128 spin_unlock(&hugetlb_lock);
Nick Piggina4822892006-03-22 00:08:08 -0800129 put_page(page); /* free it into the hugepage allocator */
130 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 }
Nick Piggina4822892006-03-22 00:08:08 -0800132 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133}
134
David Gibson27a85ef2006-03-22 00:08:56 -0800135static struct page *alloc_huge_page(struct vm_area_struct *vma,
136 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137{
138 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
140 spin_lock(&hugetlb_lock);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700141 if (vma->vm_flags & VM_MAYSHARE)
142 resv_huge_pages--;
143 else if (free_huge_pages <= resv_huge_pages)
144 goto fail;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800145
146 page = dequeue_huge_page(vma, addr);
147 if (!page)
148 goto fail;
149
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150 spin_unlock(&hugetlb_lock);
Nick Piggin7835e982006-03-22 00:08:40 -0800151 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 return page;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800153
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700154fail:
Ken Chenace4bd22007-05-09 02:33:09 -0700155 if (vma->vm_flags & VM_MAYSHARE)
156 resv_huge_pages++;
David Gibsonb45b5bd2006-03-22 00:08:55 -0800157 spin_unlock(&hugetlb_lock);
158 return NULL;
159}
160
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161static int __init hugetlb_init(void)
162{
163 unsigned long i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164
Benjamin Herrenschmidt3c726f82005-11-07 11:06:55 +1100165 if (HPAGE_SHIFT == 0)
166 return 0;
167
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 for (i = 0; i < MAX_NUMNODES; ++i)
169 INIT_LIST_HEAD(&hugepage_freelists[i]);
170
171 for (i = 0; i < max_huge_pages; ++i) {
Nick Piggina4822892006-03-22 00:08:08 -0800172 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 }
175 max_huge_pages = free_huge_pages = nr_huge_pages = i;
176 printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
177 return 0;
178}
179module_init(hugetlb_init);
180
181static int __init hugetlb_setup(char *s)
182{
183 if (sscanf(s, "%lu", &max_huge_pages) <= 0)
184 max_huge_pages = 0;
185 return 1;
186}
187__setup("hugepages=", hugetlb_setup);
188
Ken Chen8a630112007-05-09 02:33:34 -0700189static unsigned int cpuset_mems_nr(unsigned int *array)
190{
191 int node;
192 unsigned int nr = 0;
193
194 for_each_node_mask(node, cpuset_current_mems_allowed)
195 nr += array[node];
196
197 return nr;
198}
199
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200#ifdef CONFIG_SYSCTL
201static void update_and_free_page(struct page *page)
202{
203 int i;
204 nr_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700205 nr_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
207 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
208 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
209 1 << PG_private | 1<< PG_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 }
Akinobu Mitaf8af0bb2007-07-19 01:49:12 -0700211 set_compound_page_dtor(page, NULL);
Nick Piggin7835e982006-03-22 00:08:40 -0800212 set_page_refcounted(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 __free_pages(page, HUGETLB_PAGE_ORDER);
214}
215
216#ifdef CONFIG_HIGHMEM
217static void try_to_free_low(unsigned long count)
218{
Christoph Lameter4415cc82006-09-25 23:31:55 -0700219 int i;
220
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 for (i = 0; i < MAX_NUMNODES; ++i) {
222 struct page *page, *next;
223 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
224 if (PageHighMem(page))
225 continue;
226 list_del(&page->lru);
227 update_and_free_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 free_huge_pages--;
Christoph Lameter4415cc82006-09-25 23:31:55 -0700229 free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 if (count >= nr_huge_pages)
231 return;
232 }
233 }
234}
235#else
236static inline void try_to_free_low(unsigned long count)
237{
238}
239#endif
240
241static unsigned long set_max_huge_pages(unsigned long count)
242{
243 while (count > nr_huge_pages) {
Nick Piggina4822892006-03-22 00:08:08 -0800244 if (!alloc_fresh_huge_page())
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 return nr_huge_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 }
247 if (count >= nr_huge_pages)
248 return nr_huge_pages;
249
250 spin_lock(&hugetlb_lock);
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700251 count = max(count, resv_huge_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 try_to_free_low(count);
253 while (count < nr_huge_pages) {
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800254 struct page *page = dequeue_huge_page(NULL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 if (!page)
256 break;
257 update_and_free_page(page);
258 }
259 spin_unlock(&hugetlb_lock);
260 return nr_huge_pages;
261}
262
263int hugetlb_sysctl_handler(struct ctl_table *table, int write,
264 struct file *file, void __user *buffer,
265 size_t *length, loff_t *ppos)
266{
267 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
268 max_huge_pages = set_max_huge_pages(max_huge_pages);
269 return 0;
270}
Mel Gorman396faf02007-07-17 04:03:13 -0700271
272int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
273 struct file *file, void __user *buffer,
274 size_t *length, loff_t *ppos)
275{
276 proc_dointvec(table, write, file, buffer, length, ppos);
277 if (hugepages_treat_as_movable)
278 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
279 else
280 htlb_alloc_mask = GFP_HIGHUSER;
281 return 0;
282}
283
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284#endif /* CONFIG_SYSCTL */
285
286int hugetlb_report_meminfo(char *buf)
287{
288 return sprintf(buf,
289 "HugePages_Total: %5lu\n"
290 "HugePages_Free: %5lu\n"
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700291 "HugePages_Rsvd: %5lu\n"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 "Hugepagesize: %5lu kB\n",
293 nr_huge_pages,
294 free_huge_pages,
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700295 resv_huge_pages,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296 HPAGE_SIZE/1024);
297}
298
299int hugetlb_report_node_meminfo(int nid, char *buf)
300{
301 return sprintf(buf,
302 "Node %d HugePages_Total: %5u\n"
303 "Node %d HugePages_Free: %5u\n",
304 nid, nr_huge_pages_node[nid],
305 nid, free_huge_pages_node[nid]);
306}
307
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
309unsigned long hugetlb_total_pages(void)
310{
311 return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
312}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
314/*
315 * We cannot handle pagefaults against hugetlb pages at all. They cause
316 * handle_mm_fault() to try to instantiate regular-sized pages in the
317 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
318 * this far.
319 */
Nick Piggind0217ac2007-07-19 01:47:03 -0700320static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321{
322 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -0700323 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324}
325
326struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -0700327 .fault = hugetlb_vm_op_fault,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328};
329
David Gibson1e8f8892006-01-06 00:10:44 -0800330static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
331 int writable)
David Gibson63551ae2005-06-21 17:14:44 -0700332{
333 pte_t entry;
334
David Gibson1e8f8892006-01-06 00:10:44 -0800335 if (writable) {
David Gibson63551ae2005-06-21 17:14:44 -0700336 entry =
337 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
338 } else {
339 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
340 }
341 entry = pte_mkyoung(entry);
342 entry = pte_mkhuge(entry);
343
344 return entry;
345}
346
David Gibson1e8f8892006-01-06 00:10:44 -0800347static void set_huge_ptep_writable(struct vm_area_struct *vma,
348 unsigned long address, pte_t *ptep)
349{
350 pte_t entry;
351
352 entry = pte_mkwrite(pte_mkdirty(*ptep));
Benjamin Herrenschmidt8dab5242007-06-16 10:16:12 -0700353 if (ptep_set_access_flags(vma, address, ptep, entry, 1)) {
354 update_mmu_cache(vma, address, entry);
355 lazy_mmu_prot_update(entry);
356 }
David Gibson1e8f8892006-01-06 00:10:44 -0800357}
358
359
David Gibson63551ae2005-06-21 17:14:44 -0700360int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
361 struct vm_area_struct *vma)
362{
363 pte_t *src_pte, *dst_pte, entry;
364 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -0700365 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -0800366 int cow;
367
368 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -0700369
Hugh Dickins1c598272005-10-19 21:23:43 -0700370 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
Hugh Dickinsc74df322005-10-29 18:16:23 -0700371 src_pte = huge_pte_offset(src, addr);
372 if (!src_pte)
373 continue;
David Gibson63551ae2005-06-21 17:14:44 -0700374 dst_pte = huge_pte_alloc(dst, addr);
375 if (!dst_pte)
376 goto nomem;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700377 spin_lock(&dst->page_table_lock);
Hugh Dickins1c598272005-10-19 21:23:43 -0700378 spin_lock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700379 if (!pte_none(*src_pte)) {
David Gibson1e8f8892006-01-06 00:10:44 -0800380 if (cow)
381 ptep_set_wrprotect(src, addr, src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -0700382 entry = *src_pte;
383 ptepage = pte_page(entry);
384 get_page(ptepage);
Hugh Dickins1c598272005-10-19 21:23:43 -0700385 set_huge_pte_at(dst, addr, dst_pte, entry);
386 }
387 spin_unlock(&src->page_table_lock);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700388 spin_unlock(&dst->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700389 }
390 return 0;
391
392nomem:
393 return -ENOMEM;
394}
395
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700396void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
397 unsigned long end)
David Gibson63551ae2005-06-21 17:14:44 -0700398{
399 struct mm_struct *mm = vma->vm_mm;
400 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -0700401 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -0700402 pte_t pte;
403 struct page *page;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700404 struct page *tmp;
Chen, Kenneth Wc0a499c2006-12-06 20:31:39 -0800405 /*
406 * A page gathering list, protected by per file i_mmap_lock. The
407 * lock is used to avoid list corruption from multiple unmapping
408 * of the same page since we are using page->lru.
409 */
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700410 LIST_HEAD(page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700411
412 WARN_ON(!is_vm_hugetlb_page(vma));
413 BUG_ON(start & ~HPAGE_MASK);
414 BUG_ON(end & ~HPAGE_MASK);
415
Hugh Dickins508034a2005-10-29 18:16:30 -0700416 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700417 for (address = start; address < end; address += HPAGE_SIZE) {
David Gibsonc7546f82005-08-05 11:59:35 -0700418 ptep = huge_pte_offset(mm, address);
Adam Litke4c887262005-10-29 18:16:46 -0700419 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -0700420 continue;
421
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800422 if (huge_pmd_unshare(mm, &address, ptep))
423 continue;
424
David Gibsonc7546f82005-08-05 11:59:35 -0700425 pte = huge_ptep_get_and_clear(mm, address, ptep);
David Gibson63551ae2005-06-21 17:14:44 -0700426 if (pte_none(pte))
427 continue;
David Gibsonc7546f82005-08-05 11:59:35 -0700428
David Gibson63551ae2005-06-21 17:14:44 -0700429 page = pte_page(pte);
Ken Chen6649a382007-02-08 14:20:27 -0800430 if (pte_dirty(pte))
431 set_page_dirty(page);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700432 list_add(&page->lru, &page_list);
David Gibson63551ae2005-06-21 17:14:44 -0700433 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 spin_unlock(&mm->page_table_lock);
Hugh Dickins508034a2005-10-29 18:16:30 -0700435 flush_tlb_range(vma, start, end);
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -0700436 list_for_each_entry_safe(page, tmp, &page_list, lru) {
437 list_del(&page->lru);
438 put_page(page);
439 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440}
David Gibson63551ae2005-06-21 17:14:44 -0700441
Chen, Kenneth W502717f2006-10-11 01:20:46 -0700442void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
443 unsigned long end)
444{
445 /*
446 * It is undesirable to test vma->vm_file as it should be non-null
447 * for valid hugetlb area. However, vm_file will be NULL in the error
448 * cleanup path of do_mmap_pgoff. When hugetlbfs ->mmap method fails,
449 * do_mmap_pgoff() nullifies vma->vm_file before calling this function
450 * to clean up. Since no pte has actually been setup, it is safe to
451 * do nothing in this case.
452 */
453 if (vma->vm_file) {
454 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
455 __unmap_hugepage_range(vma, start, end);
456 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
457 }
458}
459
David Gibson1e8f8892006-01-06 00:10:44 -0800460static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
461 unsigned long address, pte_t *ptep, pte_t pte)
462{
463 struct page *old_page, *new_page;
David Gibson79ac6ba2006-03-22 00:08:51 -0800464 int avoidcopy;
David Gibson1e8f8892006-01-06 00:10:44 -0800465
466 old_page = pte_page(pte);
467
468 /* If no-one else is actually using this page, avoid the copy
469 * and just make the page writable */
470 avoidcopy = (page_count(old_page) == 1);
471 if (avoidcopy) {
472 set_huge_ptep_writable(vma, address, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -0700473 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800474 }
475
476 page_cache_get(old_page);
Christoph Lameter5da7ca82006-01-06 00:10:46 -0800477 new_page = alloc_huge_page(vma, address);
David Gibson1e8f8892006-01-06 00:10:44 -0800478
479 if (!new_page) {
480 page_cache_release(old_page);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800481 return VM_FAULT_OOM;
David Gibson1e8f8892006-01-06 00:10:44 -0800482 }
483
484 spin_unlock(&mm->page_table_lock);
Atsushi Nemoto9de455b2006-12-12 17:14:55 +0000485 copy_huge_page(new_page, old_page, address, vma);
David Gibson1e8f8892006-01-06 00:10:44 -0800486 spin_lock(&mm->page_table_lock);
487
488 ptep = huge_pte_offset(mm, address & HPAGE_MASK);
489 if (likely(pte_same(*ptep, pte))) {
490 /* Break COW */
491 set_huge_pte_at(mm, address, ptep,
492 make_huge_pte(vma, new_page, 1));
493 /* Make the old page be freed below */
494 new_page = old_page;
495 }
496 page_cache_release(new_page);
497 page_cache_release(old_page);
Nick Piggin83c54072007-07-19 01:47:05 -0700498 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800499}
500
Robert P. J. Daya1ed3dd2007-07-17 04:03:33 -0700501static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
David Gibson1e8f8892006-01-06 00:10:44 -0800502 unsigned long address, pte_t *ptep, int write_access)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100503{
504 int ret = VM_FAULT_SIGBUS;
Adam Litke4c887262005-10-29 18:16:46 -0700505 unsigned long idx;
506 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -0700507 struct page *page;
508 struct address_space *mapping;
David Gibson1e8f8892006-01-06 00:10:44 -0800509 pte_t new_pte;
Adam Litke4c887262005-10-29 18:16:46 -0700510
Adam Litke4c887262005-10-29 18:16:46 -0700511 mapping = vma->vm_file->f_mapping;
512 idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
513 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
514
515 /*
516 * Use page lock to guard against racing truncation
517 * before we get page_table_lock.
518 */
Christoph Lameter6bda6662006-01-06 00:10:49 -0800519retry:
520 page = find_lock_page(mapping, idx);
521 if (!page) {
Hugh Dickinsebed4bf2006-10-28 10:38:43 -0700522 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
523 if (idx >= size)
524 goto out;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800525 if (hugetlb_get_quota(mapping))
526 goto out;
527 page = alloc_huge_page(vma, address);
528 if (!page) {
529 hugetlb_put_quota(mapping);
Christoph Lameter0df420d2006-02-07 12:58:30 -0800530 ret = VM_FAULT_OOM;
Christoph Lameter6bda6662006-01-06 00:10:49 -0800531 goto out;
532 }
David Gibson79ac6ba2006-03-22 00:08:51 -0800533 clear_huge_page(page, address);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100534
Christoph Lameter6bda6662006-01-06 00:10:49 -0800535 if (vma->vm_flags & VM_SHARED) {
536 int err;
537
538 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
539 if (err) {
540 put_page(page);
541 hugetlb_put_quota(mapping);
542 if (err == -EEXIST)
543 goto retry;
544 goto out;
545 }
546 } else
547 lock_page(page);
548 }
David Gibson1e8f8892006-01-06 00:10:44 -0800549
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100550 spin_lock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700551 size = i_size_read(mapping->host) >> HPAGE_SHIFT;
552 if (idx >= size)
553 goto backout;
554
Nick Piggin83c54072007-07-19 01:47:05 -0700555 ret = 0;
Adam Litke86e52162006-01-06 00:10:43 -0800556 if (!pte_none(*ptep))
Adam Litke4c887262005-10-29 18:16:46 -0700557 goto backout;
558
David Gibson1e8f8892006-01-06 00:10:44 -0800559 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
560 && (vma->vm_flags & VM_SHARED)));
561 set_huge_pte_at(mm, address, ptep, new_pte);
562
563 if (write_access && !(vma->vm_flags & VM_SHARED)) {
564 /* Optimization, do the COW without a second fault */
565 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
566 }
567
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100568 spin_unlock(&mm->page_table_lock);
Adam Litke4c887262005-10-29 18:16:46 -0700569 unlock_page(page);
570out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100571 return ret;
Adam Litke4c887262005-10-29 18:16:46 -0700572
573backout:
574 spin_unlock(&mm->page_table_lock);
575 hugetlb_put_quota(mapping);
576 unlock_page(page);
577 put_page(page);
578 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +0100579}
580
Adam Litke86e52162006-01-06 00:10:43 -0800581int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
582 unsigned long address, int write_access)
583{
584 pte_t *ptep;
585 pte_t entry;
David Gibson1e8f8892006-01-06 00:10:44 -0800586 int ret;
David Gibson3935baa2006-03-22 00:08:53 -0800587 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800588
589 ptep = huge_pte_alloc(mm, address);
590 if (!ptep)
591 return VM_FAULT_OOM;
592
David Gibson3935baa2006-03-22 00:08:53 -0800593 /*
594 * Serialize hugepage allocation and instantiation, so that we don't
595 * get spurious allocation failures if two CPUs race to instantiate
596 * the same page in the page cache.
597 */
598 mutex_lock(&hugetlb_instantiation_mutex);
Adam Litke86e52162006-01-06 00:10:43 -0800599 entry = *ptep;
David Gibson3935baa2006-03-22 00:08:53 -0800600 if (pte_none(entry)) {
601 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
602 mutex_unlock(&hugetlb_instantiation_mutex);
603 return ret;
604 }
Adam Litke86e52162006-01-06 00:10:43 -0800605
Nick Piggin83c54072007-07-19 01:47:05 -0700606 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -0800607
608 spin_lock(&mm->page_table_lock);
609 /* Check for a racing update before calling hugetlb_cow */
610 if (likely(pte_same(entry, *ptep)))
611 if (write_access && !pte_write(entry))
612 ret = hugetlb_cow(mm, vma, address, ptep, entry);
613 spin_unlock(&mm->page_table_lock);
David Gibson3935baa2006-03-22 00:08:53 -0800614 mutex_unlock(&hugetlb_instantiation_mutex);
David Gibson1e8f8892006-01-06 00:10:44 -0800615
616 return ret;
Adam Litke86e52162006-01-06 00:10:43 -0800617}
618
David Gibson63551ae2005-06-21 17:14:44 -0700619int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
620 struct page **pages, struct vm_area_struct **vmas,
621 unsigned long *position, int *length, int i)
622{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800623 unsigned long pfn_offset;
624 unsigned long vaddr = *position;
David Gibson63551ae2005-06-21 17:14:44 -0700625 int remainder = *length;
626
Hugh Dickins1c598272005-10-19 21:23:43 -0700627 spin_lock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700628 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -0700629 pte_t *pte;
630 struct page *page;
631
632 /*
633 * Some archs (sparc64, sh*) have multiple pte_ts to
634 * each hugepage. We have to make * sure we get the
635 * first, for the page indexing below to work.
636 */
637 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
638
639 if (!pte || pte_none(*pte)) {
640 int ret;
641
642 spin_unlock(&mm->page_table_lock);
643 ret = hugetlb_fault(mm, vma, vaddr, 0);
644 spin_lock(&mm->page_table_lock);
Nick Piggin83c54072007-07-19 01:47:05 -0700645 if (!(ret & VM_FAULT_MAJOR))
Adam Litke4c887262005-10-29 18:16:46 -0700646 continue;
647
648 remainder = 0;
649 if (!i)
650 i = -EFAULT;
651 break;
652 }
David Gibson63551ae2005-06-21 17:14:44 -0700653
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800654 pfn_offset = (vaddr & ~HPAGE_MASK) >> PAGE_SHIFT;
655 page = pte_page(*pte);
656same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800657 if (pages) {
658 get_page(page);
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800659 pages[i] = page + pfn_offset;
Chen, Kenneth Wd6692182006-03-31 02:29:57 -0800660 }
David Gibson63551ae2005-06-21 17:14:44 -0700661
662 if (vmas)
663 vmas[i] = vma;
664
665 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800666 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -0700667 --remainder;
668 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -0800669 if (vaddr < vma->vm_end && remainder &&
670 pfn_offset < HPAGE_SIZE/PAGE_SIZE) {
671 /*
672 * We use pfn_offset to avoid touching the pageframes
673 * of this compound page.
674 */
675 goto same_page;
676 }
David Gibson63551ae2005-06-21 17:14:44 -0700677 }
Hugh Dickins1c598272005-10-19 21:23:43 -0700678 spin_unlock(&mm->page_table_lock);
David Gibson63551ae2005-06-21 17:14:44 -0700679 *length = remainder;
680 *position = vaddr;
681
682 return i;
683}
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800684
685void hugetlb_change_protection(struct vm_area_struct *vma,
686 unsigned long address, unsigned long end, pgprot_t newprot)
687{
688 struct mm_struct *mm = vma->vm_mm;
689 unsigned long start = address;
690 pte_t *ptep;
691 pte_t pte;
692
693 BUG_ON(address >= end);
694 flush_cache_range(vma, address, end);
695
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800696 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800697 spin_lock(&mm->page_table_lock);
698 for (; address < end; address += HPAGE_SIZE) {
699 ptep = huge_pte_offset(mm, address);
700 if (!ptep)
701 continue;
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800702 if (huge_pmd_unshare(mm, &address, ptep))
703 continue;
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800704 if (!pte_none(*ptep)) {
705 pte = huge_ptep_get_and_clear(mm, address, ptep);
706 pte = pte_mkhuge(pte_modify(pte, newprot));
707 set_huge_pte_at(mm, address, ptep, pte);
708 lazy_mmu_prot_update(pte);
709 }
710 }
711 spin_unlock(&mm->page_table_lock);
Chen, Kenneth W39dde652006-12-06 20:32:03 -0800712 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
Zhang, Yanmin8f860592006-03-22 00:08:50 -0800713
714 flush_tlb_range(vma, start, end);
715}
716
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700717struct file_region {
718 struct list_head link;
719 long from;
720 long to;
721};
722
723static long region_add(struct list_head *head, long f, long t)
724{
725 struct file_region *rg, *nrg, *trg;
726
727 /* Locate the region we are either in or before. */
728 list_for_each_entry(rg, head, link)
729 if (f <= rg->to)
730 break;
731
732 /* Round our left edge to the current segment if it encloses us. */
733 if (f > rg->from)
734 f = rg->from;
735
736 /* Check for and consume any regions we now overlap with. */
737 nrg = rg;
738 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
739 if (&rg->link == head)
740 break;
741 if (rg->from > t)
742 break;
743
744 /* If this area reaches higher then extend our area to
745 * include it completely. If this is not the first area
746 * which we intend to reuse, free it. */
747 if (rg->to > t)
748 t = rg->to;
749 if (rg != nrg) {
750 list_del(&rg->link);
751 kfree(rg);
752 }
753 }
754 nrg->from = f;
755 nrg->to = t;
756 return 0;
757}
758
759static long region_chg(struct list_head *head, long f, long t)
760{
761 struct file_region *rg, *nrg;
762 long chg = 0;
763
764 /* Locate the region we are before or in. */
765 list_for_each_entry(rg, head, link)
766 if (f <= rg->to)
767 break;
768
769 /* If we are below the current region then a new region is required.
770 * Subtle, allocate a new region at the position but make it zero
771 * size such that we can guarentee to record the reservation. */
772 if (&rg->link == head || t < rg->from) {
773 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
774 if (nrg == 0)
775 return -ENOMEM;
776 nrg->from = f;
777 nrg->to = f;
778 INIT_LIST_HEAD(&nrg->link);
779 list_add(&nrg->link, rg->link.prev);
780
781 return t - f;
782 }
783
784 /* Round our left edge to the current segment if it encloses us. */
785 if (f > rg->from)
786 f = rg->from;
787 chg = t - f;
788
789 /* Check for and consume any regions we now overlap with. */
790 list_for_each_entry(rg, rg->link.prev, link) {
791 if (&rg->link == head)
792 break;
793 if (rg->from > t)
794 return chg;
795
796 /* We overlap with this area, if it extends futher than
797 * us then we must extend ourselves. Account for its
798 * existing reservation. */
799 if (rg->to > t) {
800 chg += rg->to - t;
801 t = rg->to;
802 }
803 chg -= rg->to - rg->from;
804 }
805 return chg;
806}
807
808static long region_truncate(struct list_head *head, long end)
809{
810 struct file_region *rg, *trg;
811 long chg = 0;
812
813 /* Locate the region we are either in or before. */
814 list_for_each_entry(rg, head, link)
815 if (end <= rg->to)
816 break;
817 if (&rg->link == head)
818 return 0;
819
820 /* If we are in the middle of a region then adjust it. */
821 if (end > rg->from) {
822 chg = rg->to - end;
823 rg->to = end;
824 rg = list_entry(rg->link.next, typeof(*rg), link);
825 }
826
827 /* Drop any remaining regions. */
828 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
829 if (&rg->link == head)
830 break;
831 chg += rg->to - rg->from;
832 list_del(&rg->link);
833 kfree(rg);
834 }
835 return chg;
836}
837
838static int hugetlb_acct_memory(long delta)
839{
840 int ret = -ENOMEM;
841
842 spin_lock(&hugetlb_lock);
843 if ((delta + resv_huge_pages) <= free_huge_pages) {
844 resv_huge_pages += delta;
845 ret = 0;
846 }
847 spin_unlock(&hugetlb_lock);
848 return ret;
849}
850
851int hugetlb_reserve_pages(struct inode *inode, long from, long to)
852{
853 long ret, chg;
854
855 chg = region_chg(&inode->i_mapping->private_list, from, to);
856 if (chg < 0)
857 return chg;
Ken Chen8a630112007-05-09 02:33:34 -0700858 /*
859 * When cpuset is configured, it breaks the strict hugetlb page
860 * reservation as the accounting is done on a global variable. Such
861 * reservation is completely rubbish in the presence of cpuset because
862 * the reservation is not checked against page availability for the
863 * current cpuset. Application can still potentially OOM'ed by kernel
864 * with lack of free htlb page in cpuset that the task is in.
865 * Attempt to enforce strict accounting with cpuset is almost
866 * impossible (or too ugly) because cpuset is too fluid that
867 * task or memory node can be dynamically moved between cpusets.
868 *
869 * The change of semantics for shared hugetlb mapping with cpuset is
870 * undesirable. However, in order to preserve some of the semantics,
871 * we fall back to check against current free page availability as
872 * a best attempt and hopefully to minimize the impact of changing
873 * semantics that cpuset has.
874 */
875 if (chg > cpuset_mems_nr(free_huge_pages_node))
876 return -ENOMEM;
877
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -0700878 ret = hugetlb_acct_memory(chg);
879 if (ret < 0)
880 return ret;
881 region_add(&inode->i_mapping->private_list, from, to);
882 return 0;
883}
884
885void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
886{
887 long chg = region_truncate(&inode->i_mapping->private_list, offset);
888 hugetlb_acct_memory(freed - chg);
889}