| blob | 9d548512ff8a30498ce07f2e536f765e8bd54cca |
1 /*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/backing-dev.h>
12 #include <linux/mm.h>
13 #include <linux/shm.h>
14 #include <linux/mman.h>
15 #include <linux/pagemap.h>
16 #include <linux/swap.h>
17 #include <linux/syscalls.h>
18 #include <linux/capability.h>
19 #include <linux/init.h>
20 #include <linux/file.h>
21 #include <linux/fs.h>
22 #include <linux/personality.h>
23 #include <linux/security.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/export.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_event.h>
32 #include <linux/audit.h>
33 #include <linux/khugepaged.h>
34 #include <linux/uprobes.h>
35 #include <linux/rbtree_augmented.h>
36 #include <linux/sched/sysctl.h>
37 #include <linux/notifier.h>
38 #include <linux/memory.h>
40 #include <asm/uaccess.h>
41 #include <asm/cacheflush.h>
42 #include <asm/tlb.h>
43 #include <asm/mmu_context.h>
47 #ifndef arch_mmap_check
48 #define arch_mmap_check(addr, len, flags) (0)
49 #endif
51 #ifndef arch_rebalance_pgtables
52 #define arch_rebalance_pgtables(addr, len) (addr)
53 #endif
59 /* description of effects of mapping type and prot in current implementation.
60 * this is due to the limited x86 page protection hardware. The expected
61 * behavior is in parens:
62 *
63 * map_type prot
64 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
65 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
66 * w: (no) no w: (no) no w: (yes) yes w: (no) no
67 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
68 *
69 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
70 * w: (no) no w: (no) no w: (copy) copy w: (no) no
71 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
72 *
73 */
77 };
80 {
84 }
92 /*
93 * Make sure vm_committed_as in one cacheline and not cacheline shared with
94 * other variables. It can be updated by several CPUs frequently.
95 */
98 /*
99 * The global memory commitment made in the system can be a metric
100 * that can be used to drive ballooning decisions when Linux is hosted
101 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
102 * balancing memory across competing virtual machines that are hosted.
103 * Several metrics drive this policy engine including the guest reported
104 * memory commitment.
105 */
107 {
109 }
112 /*
113 * Check that a process has enough memory to allocate a new virtual
114 * mapping. 0 means there is enough memory for the allocation to
115 * succeed and -ENOMEM implies there is not.
116 *
117 * We currently support three overcommit policies, which are set via the
118 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
119 *
120 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
121 * Additional code 2002 Jul 20 by Robert Love.
122 *
123 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
124 *
125 * Note this is a helper function intended to be used by LSMs which
126 * wish to use this logic.
127 */
129 {
134 /*
135 * Sometimes we want to use more memory than we have
136 */
144 /*
145 * shmem pages shouldn't be counted as free in this
146 * case, they can't be purged, only swapped out, and
147 * that won't affect the overall amount of available
148 * memory in the system.
149 */
154 /*
155 * Any slabs which are created with the
156 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
157 * which are reclaimable, under pressure. The dentry
158 * cache and most inode caches should fall into this
159 */
162 /*
163 * Leave reserved pages. The pages are not for anonymous pages.
164 */
167 else
170 /*
171 * Reserve some for root
172 */
180 }
184 /*
185 * Reserve some for root
186 */
191 /*
192 * Don't let a single process grow so big a user can't recover
193 */
197 }
201 error:
205 }
207 /*
208 * Requires inode->i_mapping->i_mmap_mutex
209 */
212 {
221 else
224 }
226 /*
227 * Unlink a file-based vm structure from its interval tree, to hide
228 * vma from rmap and vmtruncate before freeing its page tables.
229 */
231 {
239 }
240 }
242 /*
243 * Close a vm structure and free it, returning the next.
244 */
246 {
257 }
262 {
271 #ifdef CONFIG_COMPAT_BRK
272 /*
273 * CONFIG_COMPAT_BRK can still be overridden by setting
274 * randomize_va_space to 2, which will still cause mm->start_brk
275 * to be arbitrarily shifted
276 */
279 else
281 #else
283 #endif
287 /*
288 * Check against rlimit here. If this check is done later after the test
289 * of oldbrk with newbrk then it can escape the test and let the data
290 * segment grow beyond its set limit the in case where the limit is
291 * not page aligned -Ram Gupta
292 */
303 /* Always allow shrinking brk. */
308 }
310 /* Check against existing mmap mappings. */
314 /* Ok, looks good - let it rip. */
318 set_brk:
326 out:
330 }
333 {
343 }
349 }
351 }
353 #ifdef CONFIG_DEBUG_VM_RB
355 {
366 }
370 }
375 }
381 }
382 i++;
386 }
389 j++;
393 }
395 }
398 {
406 }
407 }
410 {
423 i++;
424 }
428 }
433 }
438 }
440 }
441 #else
442 #define validate_mm_rb(root, ignore) do { } while (0)
443 #define validate_mm(mm) do { } while (0)
444 #endif
449 /*
450 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
451 * vma->vm_prev->vm_end values changed, without modifying the vma's position
452 * in the rbtree.
453 */
455 {
456 /*
457 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
458 * function that does exacltly what we want.
459 */
461 }
465 {
466 /* All rb_subtree_gap values must be consistent prior to insertion */
470 }
473 {
474 /*
475 * All rb_subtree_gap values must be consistent prior to erase,
476 * with the possible exception of the vma being erased.
477 */
480 /*
481 * Note rb_erase_augmented is a fairly large inline function,
482 * so make sure we instantiate it only once with our desired
483 * augmented rbtree callbacks.
484 */
486 }
488 /*
489 * vma has some anon_vma assigned, and is already inserted on that
490 * anon_vma's interval trees.
491 *
492 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
493 * vma must be removed from the anon_vma's interval trees using
494 * anon_vma_interval_tree_pre_update_vma().
495 *
496 * After the update, the vma will be reinserted using
497 * anon_vma_interval_tree_post_update_vma().
498 *
499 * The entire update must be protected by exclusive mmap_sem and by
500 * the root anon_vma's mutex.
501 */
504 {
509 }
513 {
518 }
523 {
536 /* Fail if an existing vma overlaps the area */
543 }
544 }
552 }
556 {
560 /* Find first overlaping mapping */
568 /* Iterate over the rest of the overlaps */
577 }
580 }
584 {
585 /* Update tracking information for the gap following the new vma. */
588 else
591 /*
592 * vma->vm_prev wasn't known when we followed the rbtree to find the
593 * correct insertion point for that vma. As a result, we could not
594 * update the vma vm_rb parents rb_subtree_gap values on the way down.
595 * So, we first insert the vma with a zero rb_subtree_gap value
596 * (to be consistent with what we did on the way down), and then
597 * immediately update the gap to the correct value. Finally we
598 * rebalance the rbtree after all augmented values have been set.
599 */
604 }
607 {
622 else
625 }
626 }
628 static void
632 {
635 }
640 {
657 }
659 /*
660 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
661 * mm's list and rbtree. It has already been inserted into the interval tree.
662 */
664 {
673 }
678 {
687 }
689 /*
690 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
691 * is already present in an i_mmap tree without adjusting the tree.
692 * The following helper function should be used when such adjustments
693 * are necessary. The "insert" vma (if any) is to be inserted
694 * before we drop the necessary locks.
695 */
698 {
714 /*
715 * vma expands, overlapping all the next, and
716 * perhaps the one after too (mprotect case 6).
717 */
723 /*
724 * vma expands, overlapping part of the next:
725 * mprotect case 5 shifting the boundary up.
726 */
731 /*
732 * vma shrinks, and !insert tells it's not
733 * split_vma inserting another: so it must be
734 * mprotect case 4 shifting the boundary down.
735 */
739 }
741 /*
742 * Easily overlooked: when mprotect shifts the boundary,
743 * make sure the expanding vma has anon_vma set if the
744 * shrinking vma had, to cover any anon pages imported.
745 */
750 }
751 }
762 }
766 /*
767 * Put into interval tree now, so instantiated pages
768 * are visible to arm/parisc __flush_dcache_page
769 * throughout; but we cannot insert into address
770 * space until vma start or end is updated.
771 */
773 }
774 }
788 }
795 }
800 }
804 }
809 }
816 }
819 /*
820 * vma_merge has merged next into vma, and needs
821 * us to remove next before dropping the locks.
822 */
827 /*
828 * split_vma has split insert from vma, and needs
829 * us to insert it before dropping the locks
830 * (it may either follow vma or precede it).
831 */
841 }
842 }
849 }
858 }
864 }
870 /*
871 * In mprotect's case 6 (see comments on vma_merge),
872 * we must remove another next too. It would clutter
873 * up the code too much to do both in one go.
874 */
880 else
882 }
889 }
891 /*
892 * If the vma has a ->close operation then the driver probably needs to release
893 * per-vma resources, so we don't attempt to merge those.
894 */
897 {
905 }
910 {
911 /*
912 * The list_is_singular() test is to avoid merging VMA cloned from
913 * parents. This can improve scalability caused by anon_vma lock.
914 */
919 }
921 /*
922 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
923 * in front of (at a lower virtual address and file offset than) the vma.
924 *
925 * We cannot merge two vmas if they have differently assigned (non-NULL)
926 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
927 *
928 * We don't check here for the merged mmap wrapping around the end of pagecache
929 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
930 * wrap, nor mmaps which cover the final page at index -1UL.
931 */
932 static int
935 {
940 }
942 }
944 /*
945 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
946 * beyond (at a higher virtual address and file offset than) the vma.
947 *
948 * We cannot merge two vmas if they have differently assigned (non-NULL)
949 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
950 */
951 static int
954 {
957 pgoff_t vm_pglen;
961 }
963 }
965 /*
966 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
967 * whether that can be merged with its predecessor or its successor.
968 * Or both (it neatly fills a hole).
969 *
970 * In most cases - when called for mmap, brk or mremap - [addr,end) is
971 * certain not to be mapped by the time vma_merge is called; but when
972 * called for mprotect, it is certain to be already mapped (either at
973 * an offset within prev, or at the start of next), and the flags of
974 * this area are about to be changed to vm_flags - and the no-change
975 * case has already been eliminated.
976 *
977 * The following mprotect cases have to be considered, where AAAA is
978 * the area passed down from mprotect_fixup, never extending beyond one
979 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
980 *
981 * AAAA AAAA AAAA AAAA
982 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
983 * cannot merge might become might become might become
984 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
985 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
986 * mremap move: PPPPNNNNNNNN 8
987 * AAAA
988 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
989 * might become case 1 below case 2 below case 3 below
990 *
991 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
992 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
993 */
999 {
1004 /*
1005 * We later require that vma->vm_flags == vm_flags,
1006 * so this tests vma->vm_flags & VM_SPECIAL, too.
1007 */
1013 else
1019 /*
1020 * Can it merge with the predecessor?
1021 */
1026 /*
1027 * OK, it can. Can we now merge in the successor as well?
1028 */
1035 /* cases 1, 6 */
1045 }
1047 /*
1048 * Can this new request be merged in front of next?
1049 */
1064 }
1067 }
1069 /*
1070 * Rough compatbility check to quickly see if it's even worth looking
1071 * at sharing an anon_vma.
1072 *
1073 * They need to have the same vm_file, and the flags can only differ
1074 * in things that mprotect may change.
1075 *
1076 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1077 * we can merge the two vma's. For example, we refuse to merge a vma if
1078 * there is a vm_ops->close() function, because that indicates that the
1079 * driver is doing some kind of reference counting. But that doesn't
1080 * really matter for the anon_vma sharing case.
1081 */
1083 {
1089 }
1091 /*
1092 * Do some basic sanity checking to see if we can re-use the anon_vma
1093 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1094 * the same as 'old', the other will be the new one that is trying
1095 * to share the anon_vma.
1096 *
1097 * NOTE! This runs with mm_sem held for reading, so it is possible that
1098 * the anon_vma of 'old' is concurrently in the process of being set up
1099 * by another page fault trying to merge _that_. But that's ok: if it
1100 * is being set up, that automatically means that it will be a singleton
1101 * acceptable for merging, so we can do all of this optimistically. But
1102 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1103 *
1104 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1105 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1106 * is to return an anon_vma that is "complex" due to having gone through
1107 * a fork).
1108 *
1109 * We also make sure that the two vma's are compatible (adjacent,
1110 * and with the same memory policies). That's all stable, even with just
1111 * a read lock on the mm_sem.
1112 */
1113 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1114 {
1120 }
1122 }
1124 /*
1125 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1126 * neighbouring vmas for a suitable anon_vma, before it goes off
1127 * to allocate a new anon_vma. It checks because a repetitive
1128 * sequence of mprotects and faults may otherwise lead to distinct
1129 * anon_vmas being allocated, preventing vma merge in subsequent
1130 * mprotect.
1131 */
1133 {
1144 try_prev:
1152 none:
1153 /*
1154 * There's no absolute need to look only at touching neighbours:
1155 * we could search further afield for "compatible" anon_vmas.
1156 * But it would probably just be a waste of time searching,
1157 * or lead to too many vmas hanging off the same anon_vma.
1158 * We're trying to allow mprotect remerging later on,
1159 * not trying to minimize memory used for anon_vmas.
1160 */
1162 }
1164 #ifdef CONFIG_PROC_FS
1167 {
1168 const unsigned long stack_flags
1179 }
1182 /*
1183 * If a hint addr is less than mmap_min_addr change hint to be as
1184 * low as possible but still greater than mmap_min_addr
1185 */
1187 {
1193 }
1195 /*
1196 * The caller must hold down_write(¤t->mm->mmap_sem).
1197 */
1203 {
1205 vm_flags_t vm_flags;
1209 /*
1210 * Does the application expect PROT_READ to imply PROT_EXEC?
1211 *
1212 * (the exception is when the underlying filesystem is noexec
1213 * mounted, in which case we dont add PROT_EXEC.)
1214 */
1225 /* Careful about overflows.. */
1230 /* offset overflow? */
1234 /* Too many mappings? */
1238 /* Obtain the address to map to. we verify (or select) it and ensure
1239 * that it represents a valid section of the address space.
1240 */
1245 /* Do simple checking here so the lower-level routines won't have
1246 * to. we assume access permissions have been handled by the open
1247 * of the memory object, so we don't do any here.
1248 */
1256 /* mlock MCL_FUTURE? */
1265 }
1275 /*
1276 * Make sure we don't allow writing to an append-only
1277 * file..
1278 */
1282 /*
1283 * Make sure there are no mandatory locks on the file.
1284 */
1292 /* fall through */
1300 }
1310 }
1316 /*
1317 * Ignore pgoff.
1318 */
1323 /*
1324 * Set pgoff according to addr for anon_vma.
1325 */
1330 }
1331 }
1333 /*
1334 * Set 'VM_NORESERVE' if we should not account for the
1335 * memory use of this mapping.
1336 */
1338 /* We honor MAP_NORESERVE if allowed to overcommit */
1342 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1345 }
1353 }
1358 {
1381 /*
1382 * VM_NORESERVE is used because the reservations will be
1383 * taken when vm_ops->mmap() is called
1384 * A dummy user value is used because we are not locking
1385 * memory so no accounting is necessary
1386 */
1388 VM_NORESERVE,
1393 }
1398 out_fput:
1401 out:
1403 }
1405 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1413 };
1416 {
1426 }
1429 /*
1430 * Some shared mappigns will want the pages marked read-only
1431 * to track write events. If so, we'll downgrade vm_page_prot
1432 * to the private version (using protection_map[] without the
1433 * VM_SHARED bit).
1434 */
1436 {
1439 /* If it was private or non-writable, the write bit is already clear */
1443 /* The backer wishes to know when pages are first written to? */
1447 /* The open routine did something to the protections already? */
1452 /* Specialty mapping? */
1456 /* Can the mapping track the dirty pages? */
1459 }
1461 /*
1462 * We account for memory if it's a private writeable mapping,
1463 * not hugepages and VM_NORESERVE wasn't set.
1464 */
1466 {
1467 /*
1468 * hugetlb has its own accounting separate from the core VM
1469 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1470 */
1475 }
1479 {
1486 /* Check against address space limit. */
1490 /*
1491 * MAP_FIXED may remove pages of mappings that intersects with
1492 * requested mapping. Account for the pages it would unmap.
1493 */
1501 }
1503 /* Clear old maps */
1505 munmap_back:
1510 }
1512 /*
1513 * Private writable mapping: check memory availability
1514 */
1520 }
1522 /*
1523 * Can we just expand an old mapping?
1524 */
1529 /*
1530 * Determine the object being mapped and call the appropriate
1531 * specific mapper. the address has already been validated, but
1532 * not unmapped, but the maps are removed from the list.
1533 */
1538 }
1553 }
1559 /* Can addr have changed??
1560 *
1561 * Answer: Yes, several device drivers can do it in their
1562 * f_op->mmap method. -DaveM
1563 * Bug: If addr is changed, prev, rb_link, rb_parent should
1564 * be updated for vma_link()
1565 */
1574 }
1579 /* Can vma->vm_page_prot have changed??
1580 *
1581 * Answer: Yes, drivers may have changed it in their
1582 * f_op->mmap method.
1583 *
1584 * Ensures that vmas marked as uncached stay that way.
1585 */
1589 }
1592 /* Once vma denies write, undo our temporary denial count */
1596 out:
1604 else
1606 }
1611 /*
1612 * New (or expanded) vma always get soft dirty status.
1613 * Otherwise user-space soft-dirty page tracker won't
1614 * be able to distinguish situation when vma area unmapped,
1615 * then new mapped in-place (which must be aimed as
1616 * a completely new data area).
1617 */
1622 unmap_and_free_vma:
1628 /* Undo any partial mapping done by a device driver. */
1631 free_vma:
1633 unacct_error:
1637 }
1640 {
1641 /*
1642 * We implement the search by looking for an rbtree node that
1643 * immediately follows a suitable gap. That is,
1644 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1645 * - gap_end = vma->vm_start >= info->low_limit + length;
1646 * - gap_end - gap_start >= length
1647 */
1653 /* Adjust search length to account for worst case alignment overhead */
1658 /* Adjust search limits by the desired length */
1667 /* Check if rbtree root looks promising */
1675 /* Visit left subtree if it looks promising */
1684 }
1685 }
1688 check_current:
1689 /* Check if current node has a suitable gap */
1695 /* Visit right subtree if it looks promising */
1703 }
1704 }
1706 /* Go back up the rbtree to find next candidate node */
1717 }
1718 }
1719 }
1721 check_highest:
1722 /* Check highest gap, which does not precede any rbtree node */
1728 found:
1729 /* We found a suitable gap. Clip it with the original low_limit. */
1733 /* Adjust gap address to the desired alignment */
1739 }
1742 {
1747 /* Adjust search length to account for worst case alignment overhead */
1752 /*
1753 * Adjust search limits by the desired length.
1754 * See implementation comment at top of unmapped_area().
1755 */
1765 /* Check highest gap, which does not precede any rbtree node */
1770 /* Check if rbtree root looks promising */
1778 /* Visit right subtree if it looks promising */
1787 }
1788 }
1790 check_current:
1791 /* Check if current node has a suitable gap */
1798 /* Visit left subtree if it looks promising */
1806 }
1807 }
1809 /* Go back up the rbtree to find next candidate node */
1820 }
1821 }
1822 }
1824 found:
1825 /* We found a suitable gap. Clip it with the original high_limit. */
1829 found_highest:
1830 /* Compute highest gap address at the desired alignment */
1837 }
1839 /* Get an address range which is currently unmapped.
1840 * For shmat() with addr=0.
1841 *
1842 * Ugly calling convention alert:
1843 * Return value with the low bits set means error value,
1844 * ie
1845 * if (ret & ~PAGE_MASK)
1846 * error = ret;
1847 *
1848 * This function "knows" that -ENOMEM has the bits set.
1849 */
1850 #ifndef HAVE_ARCH_UNMAPPED_AREA
1851 unsigned long
1854 {
1871 }
1879 }
1880 #endif
1882 /*
1883 * This mmap-allocator allocates new areas top-down from below the
1884 * stack's low limit (the base):
1885 */
1886 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1887 unsigned long
1891 {
1897 /* requested length too big for entire address space */
1904 /* requesting a specific address */
1911 }
1920 /*
1921 * A failed mmap() very likely causes application failure,
1922 * so fall back to the bottom-up function here. This scenario
1923 * can happen with large stack limits and large mmap()
1924 * allocations.
1925 */
1932 }
1935 }
1936 #endif
1938 unsigned long
1941 {
1949 /* Careful about overflows.. */
1968 }
1972 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1974 {
1977 /* Check the cache first. */
1978 /* (Cache hit rate is typically around 35%.) */
1999 }
2002 }
2004 }
2008 /*
2009 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2010 */
2014 {
2026 }
2027 }
2029 }
2031 /*
2032 * Verify that the stack growth is acceptable and
2033 * update accounting. This is shared with both the
2034 * grow-up and grow-down cases.
2035 */
2036 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2037 {
2042 /* address space limit tests */
2046 /* Stack limit test */
2050 /* mlock limit tests */
2059 }
2061 /* Check to ensure the stack will not grow into a hugetlb-only region */
2067 /*
2068 * Overcommit.. This must be the final test, as it will
2069 * update security statistics.
2070 */
2074 /* Ok, everything looks good - let it rip */
2079 }
2081 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2082 /*
2083 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2084 * vma is the last one with address > vma->vm_end. Have to extend vma.
2085 */
2087 {
2093 /*
2094 * We must make sure the anon_vma is allocated
2095 * so that the anon_vma locking is not a noop.
2096 */
2101 /*
2102 * vma->vm_start/vm_end cannot change under us because the caller
2103 * is required to hold the mmap_sem in read mode. We need the
2104 * anon_vma lock to serialize against concurrent expand_stacks.
2105 * Also guard against wrapping around to address 0.
2106 */
2112 }
2115 /* Somebody else might have raced and expanded it already */
2126 /*
2127 * vma_gap_update() doesn't support concurrent
2128 * updates, but we only hold a shared mmap_sem
2129 * lock here, so we need to protect against
2130 * concurrent vma expansions.
2131 * vma_lock_anon_vma() doesn't help here, as
2132 * we don't guarantee that all growable vmas
2133 * in a mm share the same root anon vma.
2134 * So, we reuse mm->page_table_lock to guard
2135 * against concurrent vma expansions.
2136 */
2143 else
2148 }
2149 }
2150 }
2155 }
2158 /*
2159 * vma is the first one with address < vma->vm_start. Have to extend vma.
2160 */
2163 {
2166 /*
2167 * We must make sure the anon_vma is allocated
2168 * so that the anon_vma locking is not a noop.
2169 */
2180 /*
2181 * vma->vm_start/vm_end cannot change under us because the caller
2182 * is required to hold the mmap_sem in read mode. We need the
2183 * anon_vma lock to serialize against concurrent expand_stacks.
2184 */
2186 /* Somebody else might have raced and expanded it already */
2197 /*
2198 * vma_gap_update() doesn't support concurrent
2199 * updates, but we only hold a shared mmap_sem
2200 * lock here, so we need to protect against
2201 * concurrent vma expansions.
2202 * vma_lock_anon_vma() doesn't help here, as
2203 * we don't guarantee that all growable vmas
2204 * in a mm share the same root anon vma.
2205 * So, we reuse mm->page_table_lock to guard
2206 * against concurrent vma expansions.
2207 */
2217 }
2218 }
2219 }
2224 }
2226 /*
2227 * Note how expand_stack() refuses to expand the stack all the way to
2228 * abut the next virtual mapping, *unless* that mapping itself is also
2229 * a stack mapping. We want to leave room for a guard page, after all
2230 * (the guard page itself is not added here, that is done by the
2231 * actual page faulting logic)
2232 *
2233 * This matches the behavior of the guard page logic (see mm/memory.c:
2234 * check_stack_guard_page()), which only allows the guard page to be
2235 * removed under these circumstances.
2236 */
2237 #ifdef CONFIG_STACK_GROWSUP
2239 {
2247 }
2249 }
2253 {
2265 }
2266 #else
2268 {
2276 }
2278 }
2282 {
2300 }
2301 #endif
2303 /*
2304 * Ok - we have the memory areas we should free on the vma list,
2305 * so release them, and do the vma updates.
2306 *
2307 * Called with the mm semaphore held.
2308 */
2310 {
2313 /* Update high watermark before we lower total_vm */
2325 }
2327 /*
2328 * Get rid of page table information in the indicated region.
2329 *
2330 * Called with the mm semaphore held.
2331 */
2335 {
2346 }
2348 /*
2349 * Create a list of vma's touched by the unmap, removing them from the mm's
2350 * vma list as we go..
2351 */
2352 static void
2355 {
2375 }
2377 /*
2378 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2379 * munmap path where it doesn't make sense to fail.
2380 */
2383 {
2395 /* most fields are the same, copy all, and then fixup */
2405 }
2423 else
2426 /* Success. */
2430 /* Clean everything up if vma_adjust failed. */
2436 out_free_mpol:
2438 out_free_vma:
2440 out_err:
2442 }
2444 /*
2445 * Split a vma into two pieces at address 'addr', a new vma is allocated
2446 * either for the first part or the tail.
2447 */
2450 {
2455 }
2457 /* Munmap is split into 2 main parts -- this part which finds
2458 * what needs doing, and the areas themselves, which do the
2459 * work. This now handles partial unmappings.
2460 * Jeremy Fitzhardinge <jeremy@goop.org>
2461 */
2463 {
2473 /* Find the first overlapping VMA */
2478 /* we have start < vma->vm_end */
2480 /* if it doesn't overlap, we have nothing.. */
2485 /*
2486 * If we need to split any vma, do it now to save pain later.
2487 *
2488 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2489 * unmapped vm_area_struct will remain in use: so lower split_vma
2490 * places tmp vma above, and higher split_vma places tmp vma below.
2491 */
2495 /*
2496 * Make sure that map_count on return from munmap() will
2497 * not exceed its limit; but let map_count go just above
2498 * its limit temporarily, to help free resources as expected.
2499 */
2507 }
2509 /* Does it split the last one? */
2515 }
2518 /*
2519 * unlock any mlock()ed ranges before detaching vmas
2520 */
2527 }
2529 }
2530 }
2532 /*
2533 * Remove the vma's, and unmap the actual pages
2534 */
2538 /* Fix up all other VM information */
2542 }
2545 {
2553 }
2557 {
2560 }
2563 {
2564 #ifdef CONFIG_DEBUG_VM
2568 }
2569 #endif
2570 }
2572 /*
2573 * this is really a simplified "do_mmap". it only handles
2574 * anonymous maps. eventually we may be able to do some
2575 * brk-specific accounting here.
2576 */
2578 {
2596 /*
2597 * mlock MCL_FUTURE?
2598 */
2607 }
2609 /*
2610 * mm->mmap_sem is required to protect against another thread
2611 * changing the mappings in case we sleep.
2612 */
2615 /*
2616 * Clear old maps. this also does some error checking for us
2617 */
2618 munmap_back:
2623 }
2625 /* Check against address space limits *after* clearing old maps... */
2635 /* Can we just expand an old private anonymous mapping? */
2641 /*
2642 * create a vma struct for an anonymous mapping
2643 */
2648 }
2658 out:
2665 }
2668 {
2680 }
2683 /* Release all mmaps. */
2685 {
2690 /* mm's last user has gone, and its about to be pulled down */
2699 }
2700 }
2711 /* update_hiwater_rss(mm) here? but nobody should be looking */
2712 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2718 /*
2719 * Walk the list again, actually closing and freeing it,
2720 * with preemption enabled, without holding any MM locks.
2721 */
2726 }
2730 }
2732 /* Insert vm structure into process list sorted by address
2733 * and into the inode's i_mmap tree. If vm_file is non-NULL
2734 * then i_mmap_mutex is taken here.
2735 */
2737 {
2741 /*
2742 * The vm_pgoff of a purely anonymous vma should be irrelevant
2743 * until its first write fault, when page's anon_vma and index
2744 * are set. But now set the vm_pgoff it will almost certainly
2745 * end up with (unless mremap moves it elsewhere before that
2746 * first wfault), so /proc/pid/maps tells a consistent story.
2747 *
2748 * By setting it to reflect the virtual start address of the
2749 * vma, merges and splits can happen in a seamless way, just
2750 * using the existing file pgoff checks and manipulations.
2751 * Similarly in do_mmap_pgoff and in do_brk.
2752 */
2756 }
2766 }
2768 /*
2769 * Copy the vma structure to a new location in the same mm,
2770 * prior to moving page table entries, to effect an mremap move.
2771 */
2775 {
2783 /*
2784 * If anonymous vma has not yet been faulted, update new pgoff
2785 * to match new location, to increase its chance of merging.
2786 */
2790 }
2797 /*
2798 * Source vma may have been merged into new_vma
2799 */
2802 /*
2803 * The only way we can get a vma_merge with
2804 * self during an mremap is if the vma hasn't
2805 * been faulted in yet and we were allowed to
2806 * reset the dst vma->vm_pgoff to the
2807 * destination address of the mremap to allow
2808 * the merge to happen. mremap must change the
2809 * vm_pgoff linearity between src and dst vmas
2810 * (in turn preventing a vma_merge) to be
2811 * safe. It is only safe to keep the vm_pgoff
2812 * linear if there are no pages mapped yet.
2813 */
2816 }
2836 }
2837 }
2840 out_free_mempol:
2842 out_free_vma:
2845 }
2847 /*
2848 * Return true if the calling process may expand its vm space by the passed
2849 * number of pages
2850 */
2852 {
2861 }
2866 {
2867 pgoff_t pgoff;
2870 /*
2871 * special mappings have no vm_file, and in that case, the mm
2872 * uses vm_pgoff internally. So we have to subtract it from here.
2873 * We are allowed to do this because we are the mm; do not copy
2874 * this code into drivers!
2875 */
2879 pgoff--;
2886 }
2889 }
2891 /*
2892 * Having a close hook prevents vma merging regardless of flags.
2893 */
2895 {
2896 }
2901 };
2903 /*
2904 * Called with mm->mmap_sem held for writing.
2905 * Insert a new vma covering the given region, with the given flags.
2906 * Its pages are supplied by the given array of struct page *.
2907 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2908 * The region past the last page supplied will always produce SIGBUS.
2909 * The array pointer and the pages it points to are assumed to stay alive
2910 * for as long as this mapping might exist.
2911 */
2915 {
2944 out:
2947 }
2952 {
2954 /*
2955 * The LSB of head.next can't change from under us
2956 * because we hold the mm_all_locks_mutex.
2957 */
2959 /*
2960 * We can safely modify head.next after taking the
2961 * anon_vma->root->rwsem. If some other vma in this mm shares
2962 * the same anon_vma we won't take it again.
2963 *
2964 * No need of atomic instructions here, head.next
2965 * can't change from under us thanks to the
2966 * anon_vma->root->rwsem.
2967 */
2971 }
2972 }
2975 {
2977 /*
2978 * AS_MM_ALL_LOCKS can't change from under us because
2979 * we hold the mm_all_locks_mutex.
2980 *
2981 * Operations on ->flags have to be atomic because
2982 * even if AS_MM_ALL_LOCKS is stable thanks to the
2983 * mm_all_locks_mutex, there may be other cpus
2984 * changing other bitflags in parallel to us.
2985 */
2989 }
2990 }
2992 /*
2993 * This operation locks against the VM for all pte/vma/mm related
2994 * operations that could ever happen on a certain mm. This includes
2995 * vmtruncate, try_to_unmap, and all page faults.
2996 *
2997 * The caller must take the mmap_sem in write mode before calling
2998 * mm_take_all_locks(). The caller isn't allowed to release the
2999 * mmap_sem until mm_drop_all_locks() returns.
3000 *
3001 * mmap_sem in write mode is required in order to block all operations
3002 * that could modify pagetables and free pages without need of
3003 * altering the vma layout (for example populate_range() with
3004 * nonlinear vmas). It's also needed in write mode to avoid new
3005 * anon_vmas to be associated with existing vmas.
3006 *
3007 * A single task can't take more than one mm_take_all_locks() in a row
3008 * or it would deadlock.
3009 *
3010 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3011 * mapping->flags avoid to take the same lock twice, if more than one
3012 * vma in this mm is backed by the same anon_vma or address_space.
3013 *
3014 * We can take all the locks in random order because the VM code
3015 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3016 * takes more than one of them in a row. Secondly we're protected
3017 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3018 *
3019 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3020 * that may have to take thousand of locks.
3021 *
3022 * mm_take_all_locks() can fail if it's interrupted by signals.
3023 */
3025 {
3038 }
3046 }
3050 out_unlock:
3053 }
3056 {
3058 /*
3059 * The LSB of head.next can't change to 0 from under
3060 * us because we hold the mm_all_locks_mutex.
3061 *
3062 * We must however clear the bitflag before unlocking
3063 * the vma so the users using the anon_vma->rb_root will
3064 * never see our bitflag.
3065 *
3066 * No need of atomic instructions here, head.next
3067 * can't change from under us until we release the
3068 * anon_vma->root->rwsem.
3069 */
3074 }
3075 }
3078 {
3080 /*
3081 * AS_MM_ALL_LOCKS can't change to 0 from under us
3082 * because we hold the mm_all_locks_mutex.
3083 */
3088 }
3089 }
3091 /*
3092 * The mmap_sem cannot be released by the caller until
3093 * mm_drop_all_locks() returns.
3094 */
3096 {
3109 }
3112 }
3114 /*
3115 * initialise the VMA slab
3116 */
3118 {
3123 }
3125 /*
3126 * Initialise sysctl_user_reserve_kbytes.
3127 *
3128 * This is intended to prevent a user from starting a single memory hogging
3129 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3130 * mode.
3131 *
3132 * The default value is min(3% of free memory, 128MB)
3133 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3134 */
3136 {
3143 }
3146 /*
3147 * Initialise sysctl_admin_reserve_kbytes.
3148 *
3149 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3150 * to log in and kill a memory hogging process.
3151 *
3152 * Systems with more than 256MB will reserve 8MB, enough to recover
3153 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3154 * only reserve 3% of free pages by default.
3155 */
3157 {
3164 }
3167 /*
3168 * Reinititalise user and admin reserves if memory is added or removed.
3169 *
3170 * The default user reserve max is 128MB, and the default max for the
3171 * admin reserve is 8MB. These are usually, but not always, enough to
3172 * enable recovery from a memory hogging process using login/sshd, a shell,
3173 * and tools like top. It may make sense to increase or even disable the
3174 * reserve depending on the existence of swap or variations in the recovery
3175 * tools. So, the admin may have changed them.
3176 *
3177 * If memory is added and the reserves have been eliminated or increased above
3178 * the default max, then we'll trust the admin.
3179 *
3180 * If memory is removed and there isn't enough free memory, then we
3181 * need to reset the reserves.
3182 *
3183 * Otherwise keep the reserve set by the admin.
3184 */
3187 {
3192 /* Default max is 128MB. Leave alone if modified by operator. */
3197 /* Default max is 8MB. Leave alone if modified by operator. */
3209 sysctl_user_reserve_kbytes);
3210 }
3215 sysctl_admin_reserve_kbytes);
3216 }
3220 }
3222 }
3226 };
3229 {
3234 }