| blob | 7f855206e7fb2bb1f9a30fcf7745bff7a2ae3adb |
1 /*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@lxorguk.ukuu.org.uk>
7 */
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
13 #include <linux/backing-dev.h>
14 #include <linux/mm.h>
15 #include <linux/vmacache.h>
16 #include <linux/shm.h>
17 #include <linux/mman.h>
18 #include <linux/pagemap.h>
19 #include <linux/swap.h>
20 #include <linux/syscalls.h>
21 #include <linux/capability.h>
22 #include <linux/init.h>
23 #include <linux/file.h>
24 #include <linux/fs.h>
25 #include <linux/personality.h>
26 #include <linux/security.h>
27 #include <linux/hugetlb.h>
28 #include <linux/profile.h>
29 #include <linux/export.h>
30 #include <linux/mount.h>
31 #include <linux/mempolicy.h>
32 #include <linux/rmap.h>
33 #include <linux/mmu_notifier.h>
34 #include <linux/mmdebug.h>
35 #include <linux/perf_event.h>
36 #include <linux/audit.h>
37 #include <linux/khugepaged.h>
38 #include <linux/uprobes.h>
39 #include <linux/rbtree_augmented.h>
40 #include <linux/sched/sysctl.h>
41 #include <linux/notifier.h>
42 #include <linux/memory.h>
43 #include <linux/printk.h>
45 #include <asm/uaccess.h>
46 #include <asm/cacheflush.h>
47 #include <asm/tlb.h>
48 #include <asm/mmu_context.h>
52 #ifndef arch_mmap_check
53 #define arch_mmap_check(addr, len, flags) (0)
54 #endif
56 #ifndef arch_rebalance_pgtables
57 #define arch_rebalance_pgtables(addr, len) (addr)
58 #endif
64 /* description of effects of mapping type and prot in current implementation.
65 * this is due to the limited x86 page protection hardware. The expected
66 * behavior is in parens:
67 *
68 * map_type prot
69 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
70 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
71 * w: (no) no w: (no) no w: (yes) yes w: (no) no
72 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
73 *
74 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
75 * w: (no) no w: (no) no w: (copy) copy w: (no) no
76 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
77 *
78 */
82 };
85 {
89 }
93 {
95 }
97 /* Update vma->vm_page_prot to reflect vma->vm_flags. */
99 {
106 vm_flags);
107 }
108 }
117 /*
118 * Make sure vm_committed_as in one cacheline and not cacheline shared with
119 * other variables. It can be updated by several CPUs frequently.
120 */
123 /*
124 * The global memory commitment made in the system can be a metric
125 * that can be used to drive ballooning decisions when Linux is hosted
126 * as a guest. On Hyper-V, the host implements a policy engine for dynamically
127 * balancing memory across competing virtual machines that are hosted.
128 * Several metrics drive this policy engine including the guest reported
129 * memory commitment.
130 */
132 {
134 }
137 /*
138 * Check that a process has enough memory to allocate a new virtual
139 * mapping. 0 means there is enough memory for the allocation to
140 * succeed and -ENOMEM implies there is not.
141 *
142 * We currently support three overcommit policies, which are set via the
143 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
144 *
145 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
146 * Additional code 2002 Jul 20 by Robert Love.
147 *
148 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
149 *
150 * Note this is a helper function intended to be used by LSMs which
151 * wish to use this logic.
152 */
154 {
163 /*
164 * Sometimes we want to use more memory than we have
165 */
173 /*
174 * shmem pages shouldn't be counted as free in this
175 * case, they can't be purged, only swapped out, and
176 * that won't affect the overall amount of available
177 * memory in the system.
178 */
183 /*
184 * Any slabs which are created with the
185 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
186 * which are reclaimable, under pressure. The dentry
187 * cache and most inode caches should fall into this
188 */
191 /*
192 * Leave reserved pages. The pages are not for anonymous pages.
193 */
196 else
199 /*
200 * Reserve some for root
201 */
209 }
212 /*
213 * Reserve some for root
214 */
218 /*
219 * Don't let a single process grow so big a user can't recover
220 */
224 }
228 error:
232 }
234 /*
235 * Requires inode->i_mapping->i_mmap_mutex
236 */
239 {
248 else
251 }
253 /*
254 * Unlink a file-based vm structure from its interval tree, to hide
255 * vma from rmap and vmtruncate before freeing its page tables.
256 */
258 {
266 }
267 }
269 /*
270 * Close a vm structure and free it, returning the next.
271 */
273 {
284 }
289 {
298 #ifdef CONFIG_COMPAT_BRK
299 /*
300 * CONFIG_COMPAT_BRK can still be overridden by setting
301 * randomize_va_space to 2, which will still cause mm->start_brk
302 * to be arbitrarily shifted
303 */
306 else
308 #else
310 #endif
314 /*
315 * Check against rlimit here. If this check is done later after the test
316 * of oldbrk with newbrk then it can escape the test and let the data
317 * segment grow beyond its set limit the in case where the limit is
318 * not page aligned -Ram Gupta
319 */
329 /* Always allow shrinking brk. */
334 }
336 /* Check against existing mmap mappings. */
340 /* Ok, looks good - let it rip. */
344 set_brk:
352 out:
356 }
359 {
369 }
375 }
377 }
379 #ifdef CONFIG_DEBUG_VM_RB
381 {
393 }
398 }
403 }
409 }
410 i++;
414 }
417 j++;
421 }
423 }
426 {
434 vma);
435 }
436 }
439 {
454 i++;
455 }
459 }
464 }
470 }
472 }
473 #else
474 #define validate_mm_rb(root, ignore) do { } while (0)
475 #define validate_mm(mm) do { } while (0)
476 #endif
481 /*
482 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
483 * vma->vm_prev->vm_end values changed, without modifying the vma's position
484 * in the rbtree.
485 */
487 {
488 /*
489 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
490 * function that does exacltly what we want.
491 */
493 }
497 {
498 /* All rb_subtree_gap values must be consistent prior to insertion */
502 }
505 {
506 /*
507 * All rb_subtree_gap values must be consistent prior to erase,
508 * with the possible exception of the vma being erased.
509 */
512 /*
513 * Note rb_erase_augmented is a fairly large inline function,
514 * so make sure we instantiate it only once with our desired
515 * augmented rbtree callbacks.
516 */
518 }
520 /*
521 * vma has some anon_vma assigned, and is already inserted on that
522 * anon_vma's interval trees.
523 *
524 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
525 * vma must be removed from the anon_vma's interval trees using
526 * anon_vma_interval_tree_pre_update_vma().
527 *
528 * After the update, the vma will be reinserted using
529 * anon_vma_interval_tree_post_update_vma().
530 *
531 * The entire update must be protected by exclusive mmap_sem and by
532 * the root anon_vma's mutex.
533 */
536 {
541 }
545 {
550 }
555 {
568 /* Fail if an existing vma overlaps the area */
575 }
576 }
584 }
588 {
592 /* Find first overlaping mapping */
600 /* Iterate over the rest of the overlaps */
609 }
612 }
616 {
617 /* Update tracking information for the gap following the new vma. */
620 else
623 /*
624 * vma->vm_prev wasn't known when we followed the rbtree to find the
625 * correct insertion point for that vma. As a result, we could not
626 * update the vma vm_rb parents rb_subtree_gap values on the way down.
627 * So, we first insert the vma with a zero rb_subtree_gap value
628 * (to be consistent with what we did on the way down), and then
629 * immediately update the gap to the correct value. Finally we
630 * rebalance the rbtree after all augmented values have been set.
631 */
636 }
639 {
654 else
657 }
658 }
660 static void
664 {
667 }
672 {
678 }
688 }
690 /*
691 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
692 * mm's list and rbtree. It has already been inserted into the interval tree.
693 */
695 {
704 }
709 {
717 /* Kill the cache */
719 }
721 /*
722 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
723 * is already present in an i_mmap tree without adjusting the tree.
724 * The following helper function should be used when such adjustments
725 * are necessary. The "insert" vma (if any) is to be inserted
726 * before we drop the necessary locks.
727 */
730 {
746 /*
747 * vma expands, overlapping all the next, and
748 * perhaps the one after too (mprotect case 6).
749 */
755 /*
756 * vma expands, overlapping part of the next:
757 * mprotect case 5 shifting the boundary up.
758 */
763 /*
764 * vma shrinks, and !insert tells it's not
765 * split_vma inserting another: so it must be
766 * mprotect case 4 shifting the boundary down.
767 */
771 }
773 /*
774 * Easily overlooked: when mprotect shifts the boundary,
775 * make sure the expanding vma has anon_vma set if the
776 * shrinking vma had, to cover any anon pages imported.
777 */
782 }
783 }
794 }
798 /*
799 * Put into interval tree now, so instantiated pages
800 * are visible to arm/parisc __flush_dcache_page
801 * throughout; but we cannot insert into address
802 * space until vma start or end is updated.
803 */
805 }
806 }
820 }
827 }
832 }
836 }
841 }
848 }
851 /*
852 * vma_merge has merged next into vma, and needs
853 * us to remove next before dropping the locks.
854 */
859 /*
860 * split_vma has split insert from vma, and needs
861 * us to insert it before dropping the locks
862 * (it may either follow vma or precede it).
863 */
873 }
874 }
881 }
890 }
896 }
902 /*
903 * In mprotect's case 6 (see comments on vma_merge),
904 * we must remove another next too. It would clutter
905 * up the code too much to do both in one go.
906 */
912 else
914 }
921 }
923 /*
924 * If the vma has a ->close operation then the driver probably needs to release
925 * per-vma resources, so we don't attempt to merge those.
926 */
929 {
930 /*
931 * VM_SOFTDIRTY should not prevent from VMA merging, if we
932 * match the flags but dirty bit -- the caller should mark
933 * merged VMA as dirty. If dirty bit won't be excluded from
934 * comparison, we increase pressue on the memory system forcing
935 * the kernel to generate new VMAs when old one could be
936 * extended instead.
937 */
945 }
950 {
951 /*
952 * The list_is_singular() test is to avoid merging VMA cloned from
953 * parents. This can improve scalability caused by anon_vma lock.
954 */
959 }
961 /*
962 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
963 * in front of (at a lower virtual address and file offset than) the vma.
964 *
965 * We cannot merge two vmas if they have differently assigned (non-NULL)
966 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
967 *
968 * We don't check here for the merged mmap wrapping around the end of pagecache
969 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
970 * wrap, nor mmaps which cover the final page at index -1UL.
971 */
972 static int
975 {
980 }
982 }
984 /*
985 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
986 * beyond (at a higher virtual address and file offset than) the vma.
987 *
988 * We cannot merge two vmas if they have differently assigned (non-NULL)
989 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
990 */
991 static int
994 {
997 pgoff_t vm_pglen;
1001 }
1003 }
1005 /*
1006 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1007 * whether that can be merged with its predecessor or its successor.
1008 * Or both (it neatly fills a hole).
1009 *
1010 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1011 * certain not to be mapped by the time vma_merge is called; but when
1012 * called for mprotect, it is certain to be already mapped (either at
1013 * an offset within prev, or at the start of next), and the flags of
1014 * this area are about to be changed to vm_flags - and the no-change
1015 * case has already been eliminated.
1016 *
1017 * The following mprotect cases have to be considered, where AAAA is
1018 * the area passed down from mprotect_fixup, never extending beyond one
1019 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1020 *
1021 * AAAA AAAA AAAA AAAA
1022 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
1023 * cannot merge might become might become might become
1024 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
1025 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
1026 * mremap move: PPPPNNNNNNNN 8
1027 * AAAA
1028 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
1029 * might become case 1 below case 2 below case 3 below
1030 *
1031 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
1032 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
1033 */
1039 {
1044 /*
1045 * We later require that vma->vm_flags == vm_flags,
1046 * so this tests vma->vm_flags & VM_SPECIAL, too.
1047 */
1053 else
1059 /*
1060 * Can it merge with the predecessor?
1061 */
1066 /*
1067 * OK, it can. Can we now merge in the successor as well?
1068 */
1075 /* cases 1, 6 */
1085 }
1087 /*
1088 * Can this new request be merged in front of next?
1089 */
1104 }
1107 }
1109 /*
1110 * Rough compatbility check to quickly see if it's even worth looking
1111 * at sharing an anon_vma.
1112 *
1113 * They need to have the same vm_file, and the flags can only differ
1114 * in things that mprotect may change.
1115 *
1116 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1117 * we can merge the two vma's. For example, we refuse to merge a vma if
1118 * there is a vm_ops->close() function, because that indicates that the
1119 * driver is doing some kind of reference counting. But that doesn't
1120 * really matter for the anon_vma sharing case.
1121 */
1123 {
1129 }
1131 /*
1132 * Do some basic sanity checking to see if we can re-use the anon_vma
1133 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1134 * the same as 'old', the other will be the new one that is trying
1135 * to share the anon_vma.
1136 *
1137 * NOTE! This runs with mm_sem held for reading, so it is possible that
1138 * the anon_vma of 'old' is concurrently in the process of being set up
1139 * by another page fault trying to merge _that_. But that's ok: if it
1140 * is being set up, that automatically means that it will be a singleton
1141 * acceptable for merging, so we can do all of this optimistically. But
1142 * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1143 *
1144 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1145 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1146 * is to return an anon_vma that is "complex" due to having gone through
1147 * a fork).
1148 *
1149 * We also make sure that the two vma's are compatible (adjacent,
1150 * and with the same memory policies). That's all stable, even with just
1151 * a read lock on the mm_sem.
1152 */
1153 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1154 {
1160 }
1162 }
1164 /*
1165 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1166 * neighbouring vmas for a suitable anon_vma, before it goes off
1167 * to allocate a new anon_vma. It checks because a repetitive
1168 * sequence of mprotects and faults may otherwise lead to distinct
1169 * anon_vmas being allocated, preventing vma merge in subsequent
1170 * mprotect.
1171 */
1173 {
1184 try_prev:
1192 none:
1193 /*
1194 * There's no absolute need to look only at touching neighbours:
1195 * we could search further afield for "compatible" anon_vmas.
1196 * But it would probably just be a waste of time searching,
1197 * or lead to too many vmas hanging off the same anon_vma.
1198 * We're trying to allow mprotect remerging later on,
1199 * not trying to minimize memory used for anon_vmas.
1200 */
1202 }
1204 #ifdef CONFIG_PROC_FS
1207 {
1208 const unsigned long stack_flags
1219 }
1222 /*
1223 * If a hint addr is less than mmap_min_addr change hint to be as
1224 * low as possible but still greater than mmap_min_addr
1225 */
1227 {
1233 }
1238 {
1241 /* mlock MCL_FUTURE? */
1249 }
1251 }
1253 /*
1254 * The caller must hold down_write(¤t->mm->mmap_sem).
1255 */
1261 {
1263 vm_flags_t vm_flags;
1267 /*
1268 * Does the application expect PROT_READ to imply PROT_EXEC?
1269 *
1270 * (the exception is when the underlying filesystem is noexec
1271 * mounted, in which case we dont add PROT_EXEC.)
1272 */
1283 /* Careful about overflows.. */
1288 /* offset overflow? */
1292 /* Too many mappings? */
1296 /* Obtain the address to map to. we verify (or select) it and ensure
1297 * that it represents a valid section of the address space.
1298 */
1303 /* Do simple checking here so the lower-level routines won't have
1304 * to. we assume access permissions have been handled by the open
1305 * of the memory object, so we don't do any here.
1306 */
1325 /*
1326 * Make sure we don't allow writing to an append-only
1327 * file..
1328 */
1332 /*
1333 * Make sure there are no mandatory locks on the file.
1334 */
1342 /* fall through */
1350 }
1360 }
1366 /*
1367 * Ignore pgoff.
1368 */
1373 /*
1374 * Set pgoff according to addr for anon_vma.
1375 */
1380 }
1381 }
1383 /*
1384 * Set 'VM_NORESERVE' if we should not account for the
1385 * memory use of this mapping.
1386 */
1388 /* We honor MAP_NORESERVE if allowed to overcommit */
1392 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1395 }
1403 }
1408 {
1431 /*
1432 * VM_NORESERVE is used because the reservations will be
1433 * taken when vm_ops->mmap() is called
1434 * A dummy user value is used because we are not locking
1435 * memory so no accounting is necessary
1436 */
1438 VM_NORESERVE,
1443 }
1448 out_fput:
1451 out:
1453 }
1455 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1463 };
1466 {
1476 }
1479 /*
1480 * Some shared mappigns will want the pages marked read-only
1481 * to track write events. If so, we'll downgrade vm_page_prot
1482 * to the private version (using protection_map[] without the
1483 * VM_SHARED bit).
1484 */
1486 {
1489 /* If it was private or non-writable, the write bit is already clear */
1493 /* The backer wishes to know when pages are first written to? */
1497 /* The open routine did something to the protections that pgprot_modify
1498 * won't preserve? */
1503 /* Do we need to track softdirty? */
1507 /* Specialty mapping? */
1511 /* Can the mapping track the dirty pages? */
1514 }
1516 /*
1517 * We account for memory if it's a private writeable mapping,
1518 * not hugepages and VM_NORESERVE wasn't set.
1519 */
1521 {
1522 /*
1523 * hugetlb has its own accounting separate from the core VM
1524 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1525 */
1530 }
1534 {
1541 /* Check against address space limit. */
1545 /*
1546 * MAP_FIXED may remove pages of mappings that intersects with
1547 * requested mapping. Account for the pages it would unmap.
1548 */
1556 }
1558 /* Clear old maps */
1560 munmap_back:
1565 }
1567 /*
1568 * Private writable mapping: check memory availability
1569 */
1575 }
1577 /*
1578 * Can we just expand an old mapping?
1579 */
1584 /*
1585 * Determine the object being mapped and call the appropriate
1586 * specific mapper. the address has already been validated, but
1587 * not unmapped, but the maps are removed from the list.
1588 */
1593 }
1608 }
1613 }
1615 /* ->mmap() can change vma->vm_file, but must guarantee that
1616 * vma_link() below can deny write-access if VM_DENYWRITE is set
1617 * and map writably if VM_SHARED is set. This usually means the
1618 * new file must not have been exposed to user-space, yet.
1619 */
1625 /* Can addr have changed??
1626 *
1627 * Answer: Yes, several device drivers can do it in their
1628 * f_op->mmap method. -DaveM
1629 * Bug: If addr is changed, prev, rb_link, rb_parent should
1630 * be updated for vma_link()
1631 */
1640 }
1643 /* Once vma denies write, undo our temporary denial count */
1649 }
1651 out:
1659 else
1661 }
1666 /*
1667 * New (or expanded) vma always get soft dirty status.
1668 * Otherwise user-space soft-dirty page tracker won't
1669 * be able to distinguish situation when vma area unmapped,
1670 * then new mapped in-place (which must be aimed as
1671 * a completely new data area).
1672 */
1679 unmap_and_free_vma:
1683 /* Undo any partial mapping done by a device driver. */
1688 allow_write_and_free_vma:
1691 free_vma:
1693 unacct_error:
1697 }
1700 {
1701 /*
1702 * We implement the search by looking for an rbtree node that
1703 * immediately follows a suitable gap. That is,
1704 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1705 * - gap_end = vma->vm_start >= info->low_limit + length;
1706 * - gap_end - gap_start >= length
1707 */
1713 /* Adjust search length to account for worst case alignment overhead */
1718 /* Adjust search limits by the desired length */
1727 /* Check if rbtree root looks promising */
1735 /* Visit left subtree if it looks promising */
1744 }
1745 }
1748 check_current:
1749 /* Check if current node has a suitable gap */
1755 /* Visit right subtree if it looks promising */
1763 }
1764 }
1766 /* Go back up the rbtree to find next candidate node */
1777 }
1778 }
1779 }
1781 check_highest:
1782 /* Check highest gap, which does not precede any rbtree node */
1788 found:
1789 /* We found a suitable gap. Clip it with the original low_limit. */
1793 /* Adjust gap address to the desired alignment */
1799 }
1802 {
1807 /* Adjust search length to account for worst case alignment overhead */
1812 /*
1813 * Adjust search limits by the desired length.
1814 * See implementation comment at top of unmapped_area().
1815 */
1825 /* Check highest gap, which does not precede any rbtree node */
1830 /* Check if rbtree root looks promising */
1838 /* Visit right subtree if it looks promising */
1847 }
1848 }
1850 check_current:
1851 /* Check if current node has a suitable gap */
1858 /* Visit left subtree if it looks promising */
1866 }
1867 }
1869 /* Go back up the rbtree to find next candidate node */
1880 }
1881 }
1882 }
1884 found:
1885 /* We found a suitable gap. Clip it with the original high_limit. */
1889 found_highest:
1890 /* Compute highest gap address at the desired alignment */
1897 }
1899 /* Get an address range which is currently unmapped.
1900 * For shmat() with addr=0.
1901 *
1902 * Ugly calling convention alert:
1903 * Return value with the low bits set means error value,
1904 * ie
1905 * if (ret & ~PAGE_MASK)
1906 * error = ret;
1907 *
1908 * This function "knows" that -ENOMEM has the bits set.
1909 */
1910 #ifndef HAVE_ARCH_UNMAPPED_AREA
1911 unsigned long
1914 {
1931 }
1939 }
1940 #endif
1942 /*
1943 * This mmap-allocator allocates new areas top-down from below the
1944 * stack's low limit (the base):
1945 */
1946 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1947 unsigned long
1951 {
1957 /* requested length too big for entire address space */
1964 /* requesting a specific address */
1971 }
1980 /*
1981 * A failed mmap() very likely causes application failure,
1982 * so fall back to the bottom-up function here. This scenario
1983 * can happen with large stack limits and large mmap()
1984 * allocations.
1985 */
1992 }
1995 }
1996 #endif
1998 unsigned long
2001 {
2009 /* Careful about overflows.. */
2028 }
2032 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
2034 {
2038 /* Check the cache first. */
2058 }
2063 }
2067 /*
2068 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
2069 */
2073 {
2085 }
2086 }
2088 }
2090 /*
2091 * Verify that the stack growth is acceptable and
2092 * update accounting. This is shared with both the
2093 * grow-up and grow-down cases.
2094 */
2095 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
2096 {
2101 /* address space limit tests */
2105 /* Stack limit test */
2109 /* mlock limit tests */
2118 }
2120 /* Check to ensure the stack will not grow into a hugetlb-only region */
2126 /*
2127 * Overcommit.. This must be the final test, as it will
2128 * update security statistics.
2129 */
2133 /* Ok, everything looks good - let it rip */
2138 }
2140 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2141 /*
2142 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2143 * vma is the last one with address > vma->vm_end. Have to extend vma.
2144 */
2146 {
2152 /*
2153 * We must make sure the anon_vma is allocated
2154 * so that the anon_vma locking is not a noop.
2155 */
2160 /*
2161 * vma->vm_start/vm_end cannot change under us because the caller
2162 * is required to hold the mmap_sem in read mode. We need the
2163 * anon_vma lock to serialize against concurrent expand_stacks.
2164 * Also guard against wrapping around to address 0.
2165 */
2171 }
2174 /* Somebody else might have raced and expanded it already */
2185 /*
2186 * vma_gap_update() doesn't support concurrent
2187 * updates, but we only hold a shared mmap_sem
2188 * lock here, so we need to protect against
2189 * concurrent vma expansions.
2190 * vma_lock_anon_vma() doesn't help here, as
2191 * we don't guarantee that all growable vmas
2192 * in a mm share the same root anon vma.
2193 * So, we reuse mm->page_table_lock to guard
2194 * against concurrent vma expansions.
2195 */
2202 else
2207 }
2208 }
2209 }
2214 }
2217 /*
2218 * vma is the first one with address < vma->vm_start. Have to extend vma.
2219 */
2222 {
2225 /*
2226 * We must make sure the anon_vma is allocated
2227 * so that the anon_vma locking is not a noop.
2228 */
2239 /*
2240 * vma->vm_start/vm_end cannot change under us because the caller
2241 * is required to hold the mmap_sem in read mode. We need the
2242 * anon_vma lock to serialize against concurrent expand_stacks.
2243 */
2245 /* Somebody else might have raced and expanded it already */
2256 /*
2257 * vma_gap_update() doesn't support concurrent
2258 * updates, but we only hold a shared mmap_sem
2259 * lock here, so we need to protect against
2260 * concurrent vma expansions.
2261 * vma_lock_anon_vma() doesn't help here, as
2262 * we don't guarantee that all growable vmas
2263 * in a mm share the same root anon vma.
2264 * So, we reuse mm->page_table_lock to guard
2265 * against concurrent vma expansions.
2266 */
2276 }
2277 }
2278 }
2283 }
2285 /*
2286 * Note how expand_stack() refuses to expand the stack all the way to
2287 * abut the next virtual mapping, *unless* that mapping itself is also
2288 * a stack mapping. We want to leave room for a guard page, after all
2289 * (the guard page itself is not added here, that is done by the
2290 * actual page faulting logic)
2291 *
2292 * This matches the behavior of the guard page logic (see mm/memory.c:
2293 * check_stack_guard_page()), which only allows the guard page to be
2294 * removed under these circumstances.
2295 */
2296 #ifdef CONFIG_STACK_GROWSUP
2298 {
2306 }
2308 }
2312 {
2324 }
2325 #else
2327 {
2335 }
2337 }
2341 {
2359 }
2360 #endif
2362 /*
2363 * Ok - we have the memory areas we should free on the vma list,
2364 * so release them, and do the vma updates.
2365 *
2366 * Called with the mm semaphore held.
2367 */
2369 {
2372 /* Update high watermark before we lower total_vm */
2384 }
2386 /*
2387 * Get rid of page table information in the indicated region.
2388 *
2389 * Called with the mm semaphore held.
2390 */
2394 {
2405 }
2407 /*
2408 * Create a list of vma's touched by the unmap, removing them from the mm's
2409 * vma list as we go..
2410 */
2411 static void
2414 {
2434 /* Kill the cache */
2436 }
2438 /*
2439 * __split_vma() bypasses sysctl_max_map_count checking. We use this on the
2440 * munmap path where it doesn't make sense to fail.
2441 */
2444 {
2456 /* most fields are the same, copy all, and then fixup */
2466 }
2484 else
2487 /* Success. */
2491 /* Clean everything up if vma_adjust failed. */
2497 out_free_mpol:
2499 out_free_vma:
2501 out_err:
2503 }
2505 /*
2506 * Split a vma into two pieces at address 'addr', a new vma is allocated
2507 * either for the first part or the tail.
2508 */
2511 {
2516 }
2518 /* Munmap is split into 2 main parts -- this part which finds
2519 * what needs doing, and the areas themselves, which do the
2520 * work. This now handles partial unmappings.
2521 * Jeremy Fitzhardinge <jeremy@goop.org>
2522 */
2524 {
2535 /* Find the first overlapping VMA */
2540 /* we have start < vma->vm_end */
2542 /* if it doesn't overlap, we have nothing.. */
2547 /*
2548 * If we need to split any vma, do it now to save pain later.
2549 *
2550 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2551 * unmapped vm_area_struct will remain in use: so lower split_vma
2552 * places tmp vma above, and higher split_vma places tmp vma below.
2553 */
2557 /*
2558 * Make sure that map_count on return from munmap() will
2559 * not exceed its limit; but let map_count go just above
2560 * its limit temporarily, to help free resources as expected.
2561 */
2569 }
2571 /* Does it split the last one? */
2577 }
2580 /*
2581 * unlock any mlock()ed ranges before detaching vmas
2582 */
2589 }
2591 }
2592 }
2594 /*
2595 * Remove the vma's, and unmap the actual pages
2596 */
2600 /* Fix up all other VM information */
2604 }
2607 {
2615 }
2619 {
2622 }
2625 {
2626 #ifdef CONFIG_DEBUG_VM
2630 }
2631 #endif
2632 }
2634 /*
2635 * this is really a simplified "do_mmap". it only handles
2636 * anonymous maps. eventually we may be able to do some
2637 * brk-specific accounting here.
2638 */
2640 {
2662 /*
2663 * mm->mmap_sem is required to protect against another thread
2664 * changing the mappings in case we sleep.
2665 */
2668 /*
2669 * Clear old maps. this also does some error checking for us
2670 */
2671 munmap_back:
2676 }
2678 /* Check against address space limits *after* clearing old maps... */
2688 /* Can we just expand an old private anonymous mapping? */
2694 /*
2695 * create a vma struct for an anonymous mapping
2696 */
2701 }
2711 out:
2718 }
2721 {
2733 }
2736 /* Release all mmaps. */
2738 {
2743 /* mm's last user has gone, and its about to be pulled down */
2752 }
2753 }
2764 /* update_hiwater_rss(mm) here? but nobody should be looking */
2765 /* Use -1 here to ensure all VMAs in the mm are unmapped */
2771 /*
2772 * Walk the list again, actually closing and freeing it,
2773 * with preemption enabled, without holding any MM locks.
2774 */
2779 }
2784 }
2786 /* Insert vm structure into process list sorted by address
2787 * and into the inode's i_mmap tree. If vm_file is non-NULL
2788 * then i_mmap_mutex is taken here.
2789 */
2791 {
2795 /*
2796 * The vm_pgoff of a purely anonymous vma should be irrelevant
2797 * until its first write fault, when page's anon_vma and index
2798 * are set. But now set the vm_pgoff it will almost certainly
2799 * end up with (unless mremap moves it elsewhere before that
2800 * first wfault), so /proc/pid/maps tells a consistent story.
2801 *
2802 * By setting it to reflect the virtual start address of the
2803 * vma, merges and splits can happen in a seamless way, just
2804 * using the existing file pgoff checks and manipulations.
2805 * Similarly in do_mmap_pgoff and in do_brk.
2806 */
2810 }
2820 }
2822 /*
2823 * Copy the vma structure to a new location in the same mm,
2824 * prior to moving page table entries, to effect an mremap move.
2825 */
2829 {
2837 /*
2838 * If anonymous vma has not yet been faulted, update new pgoff
2839 * to match new location, to increase its chance of merging.
2840 */
2844 }
2851 /*
2852 * Source vma may have been merged into new_vma
2853 */
2856 /*
2857 * The only way we can get a vma_merge with
2858 * self during an mremap is if the vma hasn't
2859 * been faulted in yet and we were allowed to
2860 * reset the dst vma->vm_pgoff to the
2861 * destination address of the mremap to allow
2862 * the merge to happen. mremap must change the
2863 * vm_pgoff linearity between src and dst vmas
2864 * (in turn preventing a vma_merge) to be
2865 * safe. It is only safe to keep the vm_pgoff
2866 * linear if there are no pages mapped yet.
2867 */
2870 }
2890 }
2891 }
2894 out_free_mempol:
2896 out_free_vma:
2899 }
2901 /*
2902 * Return true if the calling process may expand its vm space by the passed
2903 * number of pages
2904 */
2906 {
2915 }
2920 /*
2921 * Having a close hook prevents vma merging regardless of flags.
2922 */
2924 {
2925 }
2928 {
2930 }
2936 };
2941 };
2945 {
2946 pgoff_t pgoff;
2949 /*
2950 * special mappings have no vm_file, and in that case, the mm
2951 * uses vm_pgoff internally. So we have to subtract it from here.
2952 * We are allowed to do this because we are the mm; do not copy
2953 * this code into drivers!
2954 */
2959 else
2961 pages;
2964 pgoff--;
2971 }
2974 }
2981 {
3010 out:
3013 }
3015 /*
3016 * Called with mm->mmap_sem held for writing.
3017 * Insert a new vma covering the given region, with the given flags.
3018 * Its pages are supplied by the given array of struct page *.
3019 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3020 * The region past the last page supplied will always produce SIGBUS.
3021 * The array pointer and the pages it points to are assumed to stay alive
3022 * for as long as this mapping might exist.
3023 */
3028 {
3031 }
3036 {
3042 }
3047 {
3049 /*
3050 * The LSB of head.next can't change from under us
3051 * because we hold the mm_all_locks_mutex.
3052 */
3054 /*
3055 * We can safely modify head.next after taking the
3056 * anon_vma->root->rwsem. If some other vma in this mm shares
3057 * the same anon_vma we won't take it again.
3058 *
3059 * No need of atomic instructions here, head.next
3060 * can't change from under us thanks to the
3061 * anon_vma->root->rwsem.
3062 */
3066 }
3067 }
3070 {
3072 /*
3073 * AS_MM_ALL_LOCKS can't change from under us because
3074 * we hold the mm_all_locks_mutex.
3075 *
3076 * Operations on ->flags have to be atomic because
3077 * even if AS_MM_ALL_LOCKS is stable thanks to the
3078 * mm_all_locks_mutex, there may be other cpus
3079 * changing other bitflags in parallel to us.
3080 */
3084 }
3085 }
3087 /*
3088 * This operation locks against the VM for all pte/vma/mm related
3089 * operations that could ever happen on a certain mm. This includes
3090 * vmtruncate, try_to_unmap, and all page faults.
3091 *
3092 * The caller must take the mmap_sem in write mode before calling
3093 * mm_take_all_locks(). The caller isn't allowed to release the
3094 * mmap_sem until mm_drop_all_locks() returns.
3095 *
3096 * mmap_sem in write mode is required in order to block all operations
3097 * that could modify pagetables and free pages without need of
3098 * altering the vma layout (for example populate_range() with
3099 * nonlinear vmas). It's also needed in write mode to avoid new
3100 * anon_vmas to be associated with existing vmas.
3101 *
3102 * A single task can't take more than one mm_take_all_locks() in a row
3103 * or it would deadlock.
3104 *
3105 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3106 * mapping->flags avoid to take the same lock twice, if more than one
3107 * vma in this mm is backed by the same anon_vma or address_space.
3108 *
3109 * We can take all the locks in random order because the VM code
3110 * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
3111 * takes more than one of them in a row. Secondly we're protected
3112 * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
3113 *
3114 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3115 * that may have to take thousand of locks.
3116 *
3117 * mm_take_all_locks() can fail if it's interrupted by signals.
3118 */
3120 {
3133 }
3141 }
3145 out_unlock:
3148 }
3151 {
3153 /*
3154 * The LSB of head.next can't change to 0 from under
3155 * us because we hold the mm_all_locks_mutex.
3156 *
3157 * We must however clear the bitflag before unlocking
3158 * the vma so the users using the anon_vma->rb_root will
3159 * never see our bitflag.
3160 *
3161 * No need of atomic instructions here, head.next
3162 * can't change from under us until we release the
3163 * anon_vma->root->rwsem.
3164 */
3169 }
3170 }
3173 {
3175 /*
3176 * AS_MM_ALL_LOCKS can't change to 0 from under us
3177 * because we hold the mm_all_locks_mutex.
3178 */
3183 }
3184 }
3186 /*
3187 * The mmap_sem cannot be released by the caller until
3188 * mm_drop_all_locks() returns.
3189 */
3191 {
3204 }
3207 }
3209 /*
3210 * initialise the VMA slab
3211 */
3213 {
3218 }
3220 /*
3221 * Initialise sysctl_user_reserve_kbytes.
3222 *
3223 * This is intended to prevent a user from starting a single memory hogging
3224 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3225 * mode.
3226 *
3227 * The default value is min(3% of free memory, 128MB)
3228 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3229 */
3231 {
3238 }
3241 /*
3242 * Initialise sysctl_admin_reserve_kbytes.
3243 *
3244 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3245 * to log in and kill a memory hogging process.
3246 *
3247 * Systems with more than 256MB will reserve 8MB, enough to recover
3248 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3249 * only reserve 3% of free pages by default.
3250 */
3252 {
3259 }
3262 /*
3263 * Reinititalise user and admin reserves if memory is added or removed.
3264 *
3265 * The default user reserve max is 128MB, and the default max for the
3266 * admin reserve is 8MB. These are usually, but not always, enough to
3267 * enable recovery from a memory hogging process using login/sshd, a shell,
3268 * and tools like top. It may make sense to increase or even disable the
3269 * reserve depending on the existence of swap or variations in the recovery
3270 * tools. So, the admin may have changed them.
3271 *
3272 * If memory is added and the reserves have been eliminated or increased above
3273 * the default max, then we'll trust the admin.
3274 *
3275 * If memory is removed and there isn't enough free memory, then we
3276 * need to reset the reserves.
3277 *
3278 * Otherwise keep the reserve set by the admin.
3279 */
3282 {
3287 /* Default max is 128MB. Leave alone if modified by operator. */
3292 /* Default max is 8MB. Leave alone if modified by operator. */
3304 sysctl_user_reserve_kbytes);
3305 }
3310 sysctl_admin_reserve_kbytes);
3311 }
3315 }
3317 }
3321 };
3324 {
3329 }