blob: ddb872da3f5b065e11937ed5b236643731d24a3c [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
Linus Torvalds7c0f6ba2016-12-24 11:46:01 -080019#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070020#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);
Huang Yingf6ab1f72016-10-07 17:00:21 -070069 page = find_get_page(swap_address_space(swp),
70 swp_offset(swp));
Johannes Weiner0cd61442014-04-03 14:47:46 -070071 }
72 } else
73 page = find_get_page(mapping, pgoff);
74#else
75 page = find_get_page(mapping, pgoff);
Hugh Dickins31475dd2011-08-03 16:21:27 -070076#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -070077 if (page) {
78 present = PageUptodate(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +030079 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -070080 }
81
82 return present;
83}
84
Naoya Horiguchi1e25a272015-02-11 15:28:11 -080085static int __mincore_unmapped_range(unsigned long addr, unsigned long end,
86 struct vm_area_struct *vma, unsigned char *vec)
Linus Torvalds1da177e2005-04-16 15:20:36 -070087{
Johannes Weiner25ef0e52010-05-24 14:32:11 -070088 unsigned long nr = (end - addr) >> PAGE_SHIFT;
Nick Piggin42da9cb2007-02-12 00:51:39 -080089 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
Johannes Weinerf4884012010-05-24 14:32:10 -070091 if (vma->vm_file) {
92 pgoff_t pgoff;
Linus Torvalds2f77d102006-12-16 09:44:32 -080093
Johannes Weinerf4884012010-05-24 14:32:10 -070094 pgoff = linear_page_index(vma, addr);
95 for (i = 0; i < nr; i++, pgoff++)
96 vec[i] = mincore_page(vma->vm_file->f_mapping, pgoff);
97 } else {
98 for (i = 0; i < nr; i++)
99 vec[i] = 0;
Naoya Horiguchi4f16fc12009-12-14 17:59:58 -0800100 }
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800101 return nr;
Johannes Weinerf4884012010-05-24 14:32:10 -0700102}
Naoya Horiguchi4f16fc12009-12-14 17:59:58 -0800103
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800104static int mincore_unmapped_range(unsigned long addr, unsigned long end,
105 struct mm_walk *walk)
Johannes Weinerf4884012010-05-24 14:32:10 -0700106{
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800107 walk->private += __mincore_unmapped_range(addr, end,
108 walk->vma, walk->private);
109 return 0;
110}
Linus Torvalds2f77d102006-12-16 09:44:32 -0800111
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800112static int mincore_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
113 struct mm_walk *walk)
114{
115 spinlock_t *ptl;
116 struct vm_area_struct *vma = walk->vma;
117 pte_t *ptep;
118 unsigned char *vec = walk->private;
119 int nr = (end - addr) >> PAGE_SHIFT;
120
Kirill A. Shutemovb6ec57f2016-01-21 16:40:25 -0800121 ptl = pmd_trans_huge_lock(pmd, vma);
122 if (ptl) {
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800123 memset(vec, 1, nr);
124 spin_unlock(ptl);
125 goto out;
126 }
127
128 if (pmd_trans_unstable(pmd)) {
129 __mincore_unmapped_range(addr, end, vma, vec);
130 goto out;
131 }
132
133 ptep = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
134 for (; addr != end; ptep++, addr += PAGE_SIZE) {
Nick Piggin42da9cb2007-02-12 00:51:39 -0800135 pte_t pte = *ptep;
136
Johannes Weinerf4884012010-05-24 14:32:10 -0700137 if (pte_none(pte))
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800138 __mincore_unmapped_range(addr, addr + PAGE_SIZE,
139 vma, vec);
Johannes Weinerf4884012010-05-24 14:32:10 -0700140 else if (pte_present(pte))
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700141 *vec = 1;
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800142 else { /* pte is a swap entry */
Nick Piggin42da9cb2007-02-12 00:51:39 -0800143 swp_entry_t entry = pte_to_swp_entry(pte);
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700144
Weijie Yangc313dc5d2014-12-12 16:55:07 -0800145 if (non_swap_entry(entry)) {
146 /*
147 * migration or hwpoison entries are always
148 * uptodate
149 */
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700150 *vec = 1;
Nick Piggin42da9cb2007-02-12 00:51:39 -0800151 } else {
Nick Piggin30fcffe2007-02-14 12:35:02 +0100152#ifdef CONFIG_SWAP
Shaohua Li33806f02013-02-22 16:34:37 -0800153 *vec = mincore_page(swap_address_space(entry),
Huang Yingf6ab1f72016-10-07 17:00:21 -0700154 swp_offset(entry));
Nick Piggin30fcffe2007-02-14 12:35:02 +0100155#else
156 WARN_ON(1);
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700157 *vec = 1;
Nick Piggin30fcffe2007-02-14 12:35:02 +0100158#endif
Nick Piggin42da9cb2007-02-12 00:51:39 -0800159 }
160 }
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700161 vec++;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800162 }
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700163 pte_unmap_unlock(ptep - 1, ptl);
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800164out:
165 walk->private += nr;
166 cond_resched();
167 return 0;
Johannes Weinere48293f2010-05-24 14:32:11 -0700168}
169
Johannes Weinerf4884012010-05-24 14:32:10 -0700170/*
171 * Do a chunk of "sys_mincore()". We've already checked
172 * all the arguments, we hold the mmap semaphore: we should
173 * just return the amount of info we're asked for.
174 */
175static long do_mincore(unsigned long addr, unsigned long pages, unsigned char *vec)
176{
Johannes Weinerf4884012010-05-24 14:32:10 -0700177 struct vm_area_struct *vma;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700178 unsigned long end;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800179 int err;
180 struct mm_walk mincore_walk = {
181 .pmd_entry = mincore_pte_range,
182 .pte_hole = mincore_unmapped_range,
183 .hugetlb_entry = mincore_hugetlb,
184 .private = vec,
185 };
Johannes Weinerf4884012010-05-24 14:32:10 -0700186
187 vma = find_vma(current->mm, addr);
188 if (!vma || addr < vma->vm_start)
189 return -ENOMEM;
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800190 mincore_walk.mm = vma->vm_mm;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700191 end = min(vma->vm_end, addr + (pages << PAGE_SHIFT));
Naoya Horiguchi1e25a272015-02-11 15:28:11 -0800192 err = walk_page_range(addr, end, &mincore_walk);
193 if (err < 0)
194 return err;
Johannes Weiner25ef0e52010-05-24 14:32:11 -0700195 return (end - addr) >> PAGE_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196}
197
198/*
199 * The mincore(2) system call.
200 *
201 * mincore() returns the memory residency status of the pages in the
202 * current process's address space specified by [addr, addr + len).
203 * The status is returned in a vector of bytes. The least significant
204 * bit of each byte is 1 if the referenced page is in memory, otherwise
205 * it is zero.
206 *
207 * Because the status of a page can change after mincore() checks it
208 * but before it returns to the application, the returned vector may
209 * contain stale information. Only locked pages are guaranteed to
210 * remain in memory.
211 *
212 * return values:
213 * zero - success
214 * -EFAULT - vec points to an illegal address
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +0300215 * -EINVAL - addr is not a multiple of PAGE_SIZE
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 * -ENOMEM - Addresses in the range [addr, addr + len] are
217 * invalid for the address space of this process, or
218 * specify one or more pages which are not currently
219 * mapped
220 * -EAGAIN - A kernel resource was temporarily unavailable.
221 */
Heiko Carstens3480b252009-01-14 14:14:16 +0100222SYSCALL_DEFINE3(mincore, unsigned long, start, size_t, len,
223 unsigned char __user *, vec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224{
Linus Torvalds2f77d102006-12-16 09:44:32 -0800225 long retval;
226 unsigned long pages;
227 unsigned char *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
Linus Torvalds2f77d102006-12-16 09:44:32 -0800229 /* Check the start address: needs to be page-aligned.. */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +0300230 if (start & ~PAGE_MASK)
Linus Torvalds2f77d102006-12-16 09:44:32 -0800231 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232
Linus Torvalds2f77d102006-12-16 09:44:32 -0800233 /* ..and we need to be passed a valid user-space range */
234 if (!access_ok(VERIFY_READ, (void __user *) start, len))
235 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236
Kirill A. Shutemovea1754a2016-04-01 15:29:48 +0300237 /* This also avoids any overflows on PAGE_ALIGN */
Linus Torvalds2f77d102006-12-16 09:44:32 -0800238 pages = len >> PAGE_SHIFT;
Alexander Kuleshove7bbdd02015-11-05 18:46:38 -0800239 pages += (offset_in_page(len)) != 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240
Linus Torvalds2f77d102006-12-16 09:44:32 -0800241 if (!access_ok(VERIFY_WRITE, vec, pages))
242 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243
Linus Torvalds2f77d102006-12-16 09:44:32 -0800244 tmp = (void *) __get_free_page(GFP_USER);
245 if (!tmp)
Linus Torvalds4fb23e42006-12-16 16:01:50 -0800246 return -EAGAIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247
Linus Torvalds2f77d102006-12-16 09:44:32 -0800248 retval = 0;
249 while (pages) {
250 /*
251 * Do at most PAGE_SIZE entries per iteration, due to
252 * the temporary buffer size.
253 */
254 down_read(&current->mm->mmap_sem);
Johannes Weiner6a60f1b2010-05-24 14:32:09 -0700255 retval = do_mincore(start, min(pages, PAGE_SIZE), tmp);
Linus Torvalds2f77d102006-12-16 09:44:32 -0800256 up_read(&current->mm->mmap_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
Linus Torvalds2f77d102006-12-16 09:44:32 -0800258 if (retval <= 0)
259 break;
260 if (copy_to_user(vec, tmp, retval)) {
261 retval = -EFAULT;
262 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 }
Linus Torvalds2f77d102006-12-16 09:44:32 -0800264 pages -= retval;
265 vec += retval;
266 start += retval << PAGE_SHIFT;
267 retval = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 }
Linus Torvalds2f77d102006-12-16 09:44:32 -0800269 free_page((unsigned long) tmp);
270 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}