blob: 563f320454902df04e782f73aa8d0473656034c9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/mincore.c
3 *
Linus Torvalds2f77d102006-12-16 09:44:32 -08004 * Copyright (C) 1994-2006 Linus Torvalds
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 */
6
7/*
8 * The mincore() system call.
9 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/pagemap.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090011#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070012#include <linux/mm.h>
13#include <linux/mman.h>
14#include <linux/syscalls.h>
Nick Piggin42da9cb2007-02-12 00:51:39 -080015#include <linux/swap.h>
16#include <linux/swapops.h>
Naoya Horiguchi4f16fc12009-12-14 17:59:58 -080017#include <linux/hugetlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070018
19#include <asm/uaccess.h>
20#include <asm/pgtable.h>
21
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080022static int mincore_hugetlb(pte_t *pte, unsigned long hmask, unsigned long addr,
23 unsigned long end, struct mm_walk *walk)
Johannes Weinerf4884012010-05-24 14:32:10 -070024{
25#ifdef CONFIG_HUGETLB_PAGE
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080026 unsigned char present;
27 unsigned char *vec = walk->private;
Johannes Weinerf4884012010-05-24 14:32:10 -070028
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080029 /*
30 * Hugepages under user process are always in RAM and never
31 * swapped out, but theoretically it needs to be checked.
32 */
33 present = pte && !huge_pte_none(huge_ptep_get(pte));
34 for (; addr != end; vec++, addr += PAGE_SIZE)
35 *vec = present;
36 walk->private = vec;
Johannes Weinerf4884012010-05-24 14:32:10 -070037#else
38 BUG();
39#endif
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080040 return 0;
Johannes Weinerf4884012010-05-24 14:32:10 -070041}
42
Linus Torvalds1da177e2005-04-16 15:20:36 -070043/*
44 * Later we can get more picky about what "in core" means precisely.
45 * For now, simply check to see if the page is in the page cache,
46 * and is up to date; i.e. that no page-in operation would be required
47 * at this time if an application were to map and access this page.
48 */
Nick Piggin42da9cb2007-02-12 00:51:39 -080049static unsigned char mincore_page(struct address_space *mapping, pgoff_t pgoff)
Linus Torvalds1da177e2005-04-16 15:20:36 -070050{
51 unsigned char present = 0;
Nick Piggin42da9cb2007-02-12 00:51:39 -080052 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -070053
Nick Piggin42da9cb2007-02-12 00:51:39 -080054 /*
55 * When tmpfs swaps out a page from a file, any process mapping that
56 * file will not get a swp_entry_t in its pte, but rather it is like
57 * any other file mapping (ie. marked !present and faulted in with
Nick Piggin3c18ddd2008-04-28 02:12:10 -070058 * tmpfs's .fault). So swapped out tmpfs mappings are tested here.
Nick Piggin42da9cb2007-02-12 00:51:39 -080059 */
Hugh Dickins31475dd2011-08-03 16:21:27 -070060#ifdef CONFIG_SWAP
Johannes Weiner0cd61442014-04-03 14:47:46 -070061 if (shmem_mapping(mapping)) {
62 page = find_get_entry(mapping, pgoff);
63 /*
64 * shmem/tmpfs may return swap: account for swapcache
65 * page too.
66 */
67 if (radix_tree_exceptional_entry(page)) {
68 swp_entry_t swp = radix_to_swp_entry(page);
69 page = find_get_page(swap_address_space(swp), swp.val);
70 }
71 } else
72 page = find_get_page(mapping, pgoff);
73#else
74 page = find_get_page(mapping, pgoff);
Hugh Dickins31475dd2011-08-03 16:21:27 -070075#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 if (page) {
77 present = PageUptodate(page);
78 page_cache_release(page);
79 }
80
81 return present;
82}
83
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080084static int __mincore_unmapped_range(unsigned long addr, unsigned long end,
85 struct vm_area_struct *vma, unsigned char *vec)
Linus Torvalds1da177e2005-04-16 15:20:36 -070086{
Johannes Weiner25ef0e52010-05-24 14:32:11 -070087 unsigned long nr = (end - addr) >> PAGE_SHIFT;
Nick Piggin42da9cb2007-02-12 00:51:39 -080088 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
Johannes Weinerf4884012010-05-24 14:32:10 -070090 if (vma->vm_file) {
91 pgoff_t pgoff;
Linus Torvalds2f77d102006-12-16 09:44:32 -080092
Johannes Weinerf4884012010-05-24 14:32:10 -070093 pgoff = linear_page_index(vma, addr);
94 for (i = 0; i < nr; i++, pgoff++)
95 vec[i] = mincore_page(vma->vm_file->f_mapping, pgoff);
96 } else {
97 for (i = 0; i < nr; i++)
98 vec[i] = 0;
Naoya Horiguchi4f16fc12009-12-14 17:59:58 -080099 }
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800100 return nr;
Johannes Weinerf4884012010-05-24 14:32:10 -0700101}
Naoya Horiguchi4f16fc12009-12-14 17:59:58 -0800102
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800103static int mincore_unmapped_range(unsigned long addr, unsigned long end,
104 struct mm_walk *walk)
Johannes Weinerf4884012010-05-24 14:32:10 -0700105{
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800106 walk->private += __mincore_unmapped_range(addr, end,
107 walk->vma, walk->private);
108 return 0;
109}
Linus Torvalds2f77d102006-12-16 09:44:32 -0800110
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800111static int mincore_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
112 struct mm_walk *walk)
113{
114 spinlock_t *ptl;
115 struct vm_area_struct *vma = walk->vma;
116 pte_t *ptep;
117 unsigned char *vec = walk->private;
118 int nr = (end - addr) >> PAGE_SHIFT;
119
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -0800120 ptl = pmd_trans_huge_lock(pmd, vma);
121 if (ptl) {
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800122 memset(vec, 1, nr);
123 spin_unlock(ptl);
124 goto out;
125 }
126
127 if (pmd_trans_unstable(pmd)) {
128 __mincore_unmapped_range(addr, end, vma, vec);
129 goto out;
130 }
131
132 ptep = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
133 for (; addr != end; ptep++, addr += PAGE_SIZE) {
Nick Piggin42da9cb2007-02-12 00:51:39 -0800134 pte_t pte = *ptep;
135
Johannes Weinerf4884012010-05-24 14:32:10 -0700136 if (pte_none(pte))
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800137 __mincore_unmapped_range(addr, addr + PAGE_SIZE,
138 vma, vec);
Johannes Weinerf4884012010-05-24 14:32:10 -0700139 else if (pte_present(pte))
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700140 *vec = 1;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800141 else { /* pte is a swap entry */
Nick Piggin42da9cb2007-02-12 00:51:39 -0800142 swp_entry_t entry = pte_to_swp_entry(pte);
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700143
Weijie Yangc313dc5d2014-12-12 16:55:07 -0800144 if (non_swap_entry(entry)) {
145 /*
146 * migration or hwpoison entries are always
147 * uptodate
148 */
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700149 *vec = 1;
Nick Piggin42da9cb2007-02-12 00:51:39 -0800150 } else {
Nick Piggin30fcffe2007-02-14 12:35:02 +0100151#ifdef CONFIG_SWAP
Shaohua Li33806f02013-02-22 16:34:37 -0800152 *vec = mincore_page(swap_address_space(entry),
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800153 entry.val);
Nick Piggin30fcffe2007-02-14 12:35:02 +0100154#else
155 WARN_ON(1);
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700156 *vec = 1;
Nick Piggin30fcffe2007-02-14 12:35:02 +0100157#endif
Nick Piggin42da9cb2007-02-12 00:51:39 -0800158 }
159 }
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700160 vec++;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800161 }
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700162 pte_unmap_unlock(ptep - 1, ptl);
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800163out:
164 walk->private += nr;
165 cond_resched();
166 return 0;
Johannes Weinere48293f2010-05-24 14:32:11 -0700167}
168
Johannes Weinerf4884012010-05-24 14:32:10 -0700169/*
170 * Do a chunk of "sys_mincore()". We've already checked
171 * all the arguments, we hold the mmap semaphore: we should
172 * just return the amount of info we're asked for.
173 */
174static long do_mincore(unsigned long addr, unsigned long pages, unsigned char *vec)
175{
Johannes Weinerf4884012010-05-24 14:32:10 -0700176 struct vm_area_struct *vma;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700177 unsigned long end;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800178 int err;
179 struct mm_walk mincore_walk = {
180 .pmd_entry = mincore_pte_range,
181 .pte_hole = mincore_unmapped_range,
182 .hugetlb_entry = mincore_hugetlb,
183 .private = vec,
184 };
Johannes Weinerf4884012010-05-24 14:32:10 -0700185
186 vma = find_vma(current->mm, addr);
187 if (!vma || addr < vma->vm_start)
188 return -ENOMEM;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800189 mincore_walk.mm = vma->vm_mm;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700190 end = min(vma->vm_end, addr + (pages << PAGE_SHIFT));
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800191 err = walk_page_range(addr, end, &mincore_walk);
192 if (err < 0)
193 return err;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700194 return (end - addr) >> PAGE_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195}
196
197/*
198 * The mincore(2) system call.
199 *
200 * mincore() returns the memory residency status of the pages in the
201 * current process's address space specified by [addr, addr + len).
202 * The status is returned in a vector of bytes. The least significant
203 * bit of each byte is 1 if the referenced page is in memory, otherwise
204 * it is zero.
205 *
206 * Because the status of a page can change after mincore() checks it
207 * but before it returns to the application, the returned vector may
208 * contain stale information. Only locked pages are guaranteed to
209 * remain in memory.
210 *
211 * return values:
212 * zero - success
213 * -EFAULT - vec points to an illegal address
214 * -EINVAL - addr is not a multiple of PAGE_CACHE_SIZE
215 * -ENOMEM - Addresses in the range [addr, addr + len] are
216 * invalid for the address space of this process, or
217 * specify one or more pages which are not currently
218 * mapped
219 * -EAGAIN - A kernel resource was temporarily unavailable.
220 */
Heiko Carstens3480b252009-01-14 14:14:16 +0100221SYSCALL_DEFINE3(mincore, unsigned long, start, size_t, len,
222 unsigned char __user *, vec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223{
Linus Torvalds2f77d102006-12-16 09:44:32 -0800224 long retval;
225 unsigned long pages;
226 unsigned char *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227
Linus Torvalds2f77d102006-12-16 09:44:32 -0800228 /* Check the start address: needs to be page-aligned.. */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 if (start & ~PAGE_CACHE_MASK)
Linus Torvalds2f77d102006-12-16 09:44:32 -0800230 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
Linus Torvalds2f77d102006-12-16 09:44:32 -0800232 /* ..and we need to be passed a valid user-space range */
233 if (!access_ok(VERIFY_READ, (void __user *) start, len))
234 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235
Linus Torvalds2f77d102006-12-16 09:44:32 -0800236 /* This also avoids any overflows on PAGE_CACHE_ALIGN */
237 pages = len >> PAGE_SHIFT;
Alexander Kuleshove7bbdd02015-11-05 18:46:38 -0800238 pages += (offset_in_page(len)) != 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
Linus Torvalds2f77d102006-12-16 09:44:32 -0800240 if (!access_ok(VERIFY_WRITE, vec, pages))
241 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Linus Torvalds2f77d102006-12-16 09:44:32 -0800243 tmp = (void *) __get_free_page(GFP_USER);
244 if (!tmp)
Linus Torvalds4fb23e42006-12-16 16:01:50 -0800245 return -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
Linus Torvalds2f77d102006-12-16 09:44:32 -0800247 retval = 0;
248 while (pages) {
249 /*
250 * Do at most PAGE_SIZE entries per iteration, due to
251 * the temporary buffer size.
252 */
253 down_read(&current->mm->mmap_sem);
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700254 retval = do_mincore(start, min(pages, PAGE_SIZE), tmp);
Linus Torvalds2f77d102006-12-16 09:44:32 -0800255 up_read(&current->mm->mmap_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256
Linus Torvalds2f77d102006-12-16 09:44:32 -0800257 if (retval <= 0)
258 break;
259 if (copy_to_user(vec, tmp, retval)) {
260 retval = -EFAULT;
261 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 }
Linus Torvalds2f77d102006-12-16 09:44:32 -0800263 pages -= retval;
264 vec += retval;
265 start += retval << PAGE_SHIFT;
266 retval = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 }
Linus Torvalds2f77d102006-12-16 09:44:32 -0800268 free_page((unsigned long) tmp);
269 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270}