Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * mm/mmap.c
   4 *
   5 * Written by obz.
   6 *
   7 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   8 */
   9
  10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12#include <linux/kernel.h>
  13#include <linux/slab.h>
  14#include <linux/backing-dev.h>
  15#include <linux/mm.h>
  16#include <linux/mm_inline.h>
  17#include <linux/shm.h>
  18#include <linux/mman.h>
  19#include <linux/pagemap.h>
  20#include <linux/swap.h>
  21#include <linux/syscalls.h>
  22#include <linux/capability.h>
  23#include <linux/init.h>
  24#include <linux/file.h>
  25#include <linux/fs.h>
  26#include <linux/personality.h>
  27#include <linux/security.h>
  28#include <linux/hugetlb.h>
  29#include <linux/shmem_fs.h>
  30#include <linux/profile.h>
  31#include <linux/export.h>
  32#include <linux/mount.h>
  33#include <linux/mempolicy.h>
  34#include <linux/rmap.h>
  35#include <linux/mmu_notifier.h>
  36#include <linux/mmdebug.h>
  37#include <linux/perf_event.h>
  38#include <linux/audit.h>
  39#include <linux/khugepaged.h>
  40#include <linux/uprobes.h>
 
  41#include <linux/notifier.h>
  42#include <linux/memory.h>
  43#include <linux/printk.h>
  44#include <linux/userfaultfd_k.h>
  45#include <linux/moduleparam.h>
  46#include <linux/pkeys.h>
  47#include <linux/oom.h>
  48#include <linux/sched/mm.h>
  49
  50#include <linux/uaccess.h>
  51#include <asm/cacheflush.h>
  52#include <asm/tlb.h>
  53#include <asm/mmu_context.h>
  54
  55#define CREATE_TRACE_POINTS
  56#include <trace/events/mmap.h>
  57
  58#include "internal.h"
  59
  60#ifndef arch_mmap_check
  61#define arch_mmap_check(addr, len, flags)	(0)
  62#endif
  63
  64#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  65const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  66const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  67int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  68#endif
  69#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  70const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  71const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  72int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  73#endif
  74
  75static bool ignore_rlimit_data;
  76core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  77
  78static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
  79		struct vm_area_struct *vma, struct vm_area_struct *prev,
  80		struct vm_area_struct *next, unsigned long start,
  81		unsigned long end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  82
  83static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
  84{
  85	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
  86}
  87
  88/* Update vma->vm_page_prot to reflect vma->vm_flags. */
  89void vma_set_page_prot(struct vm_area_struct *vma)
  90{
  91	unsigned long vm_flags = vma->vm_flags;
  92	pgprot_t vm_page_prot;
  93
  94	vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
  95	if (vma_wants_writenotify(vma, vm_page_prot)) {
  96		vm_flags &= ~VM_SHARED;
  97		vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
  98	}
  99	/* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
 100	WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
 101}
 102
 103/*
 104 * Requires inode->i_mapping->i_mmap_rwsem
 105 */
 106static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 107		struct file *file, struct address_space *mapping)
 108{
 
 
 109	if (vma->vm_flags & VM_SHARED)
 110		mapping_unmap_writable(mapping);
 111
 112	flush_dcache_mmap_lock(mapping);
 113	vma_interval_tree_remove(vma, &mapping->i_mmap);
 114	flush_dcache_mmap_unlock(mapping);
 115}
 116
 117/*
 118 * Unlink a file-based vm structure from its interval tree, to hide
 119 * vma from rmap and vmtruncate before freeing its page tables.
 120 */
 121void unlink_file_vma(struct vm_area_struct *vma)
 122{
 123	struct file *file = vma->vm_file;
 124
 125	if (file) {
 126		struct address_space *mapping = file->f_mapping;
 127		i_mmap_lock_write(mapping);
 128		__remove_shared_vm_struct(vma, file, mapping);
 129		i_mmap_unlock_write(mapping);
 130	}
 131}
 132
 133/*
 134 * Close a vm structure and free it.
 135 */
 136static void remove_vma(struct vm_area_struct *vma)
 137{
 
 
 138	might_sleep();
 139	if (vma->vm_ops && vma->vm_ops->close)
 140		vma->vm_ops->close(vma);
 141	if (vma->vm_file)
 142		fput(vma->vm_file);
 143	mpol_put(vma_policy(vma));
 144	vm_area_free(vma);
 
 145}
 146
 147/*
 148 * check_brk_limits() - Use platform specific check of range & verify mlock
 149 * limits.
 150 * @addr: The address to check
 151 * @len: The size of increase.
 152 *
 153 * Return: 0 on success.
 154 */
 155static int check_brk_limits(unsigned long addr, unsigned long len)
 156{
 157	unsigned long mapped_addr;
 158
 159	mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
 160	if (IS_ERR_VALUE(mapped_addr))
 161		return mapped_addr;
 162
 163	return mlock_future_check(current->mm, current->mm->def_flags, len);
 164}
 165static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
 166			 unsigned long newbrk, unsigned long oldbrk,
 167			 struct list_head *uf);
 168static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *brkvma,
 169		unsigned long addr, unsigned long request, unsigned long flags);
 170SYSCALL_DEFINE1(brk, unsigned long, brk)
 171{
 172	unsigned long newbrk, oldbrk, origbrk;
 
 173	struct mm_struct *mm = current->mm;
 174	struct vm_area_struct *brkvma, *next = NULL;
 175	unsigned long min_brk;
 176	bool populate;
 177	bool downgraded = false;
 178	LIST_HEAD(uf);
 179	MA_STATE(mas, &mm->mm_mt, 0, 0);
 180
 181	if (mmap_write_lock_killable(mm))
 182		return -EINTR;
 183
 184	origbrk = mm->brk;
 185
 186#ifdef CONFIG_COMPAT_BRK
 187	/*
 188	 * CONFIG_COMPAT_BRK can still be overridden by setting
 189	 * randomize_va_space to 2, which will still cause mm->start_brk
 190	 * to be arbitrarily shifted
 191	 */
 192	if (current->brk_randomized)
 193		min_brk = mm->start_brk;
 194	else
 195		min_brk = mm->end_data;
 196#else
 197	min_brk = mm->start_brk;
 198#endif
 199	if (brk < min_brk)
 200		goto out;
 201
 202	/*
 203	 * Check against rlimit here. If this check is done later after the test
 204	 * of oldbrk with newbrk then it can escape the test and let the data
 205	 * segment grow beyond its set limit the in case where the limit is
 206	 * not page aligned -Ram Gupta
 207	 */
 208	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
 209			      mm->end_data, mm->start_data))
 210		goto out;
 211
 212	newbrk = PAGE_ALIGN(brk);
 213	oldbrk = PAGE_ALIGN(mm->brk);
 214	if (oldbrk == newbrk) {
 215		mm->brk = brk;
 216		goto success;
 217	}
 218
 219	/*
 220	 * Always allow shrinking brk.
 221	 * do_brk_munmap() may downgrade mmap_lock to read.
 222	 */
 223	if (brk <= mm->brk) {
 224		int ret;
 225
 226		/* Search one past newbrk */
 227		mas_set(&mas, newbrk);
 228		brkvma = mas_find(&mas, oldbrk);
 229		if (!brkvma || brkvma->vm_start >= oldbrk)
 230			goto out; /* mapping intersects with an existing non-brk vma. */
 231		/*
 232		 * mm->brk must be protected by write mmap_lock.
 233		 * do_brk_munmap() may downgrade the lock,  so update it
 234		 * before calling do_brk_munmap().
 235		 */
 236		mm->brk = brk;
 237		ret = do_brk_munmap(&mas, brkvma, newbrk, oldbrk, &uf);
 238		if (ret == 1)  {
 239			downgraded = true;
 240			goto success;
 241		} else if (!ret)
 242			goto success;
 243
 244		mm->brk = origbrk;
 245		goto out;
 246	}
 247
 248	if (check_brk_limits(oldbrk, newbrk - oldbrk))
 249		goto out;
 250
 251	/*
 252	 * Only check if the next VMA is within the stack_guard_gap of the
 253	 * expansion area
 254	 */
 255	mas_set(&mas, oldbrk);
 256	next = mas_find(&mas, newbrk - 1 + PAGE_SIZE + stack_guard_gap);
 257	if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
 258		goto out;
 259
 260	brkvma = mas_prev(&mas, mm->start_brk);
 261	/* Ok, looks good - let it rip. */
 262	if (do_brk_flags(&mas, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
 263		goto out;
 264
 
 265	mm->brk = brk;
 266
 267success:
 268	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 269	if (downgraded)
 270		mmap_read_unlock(mm);
 271	else
 272		mmap_write_unlock(mm);
 273	userfaultfd_unmap_complete(mm, &uf);
 274	if (populate)
 275		mm_populate(oldbrk, newbrk - oldbrk);
 276	return brk;
 277
 278out:
 279	mmap_write_unlock(mm);
 280	return origbrk;
 
 281}
 282
 283#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
 284extern void mt_validate(struct maple_tree *mt);
 285extern void mt_dump(const struct maple_tree *mt);
 286
 287/* Validate the maple tree */
 288static void validate_mm_mt(struct mm_struct *mm)
 289{
 290	struct maple_tree *mt = &mm->mm_mt;
 291	struct vm_area_struct *vma_mt;
 292
 293	MA_STATE(mas, mt, 0, 0);
 294
 295	mt_validate(&mm->mm_mt);
 296	mas_for_each(&mas, vma_mt, ULONG_MAX) {
 297		if ((vma_mt->vm_start != mas.index) ||
 298		    (vma_mt->vm_end - 1 != mas.last)) {
 299			pr_emerg("issue in %s\n", current->comm);
 300			dump_stack();
 301			dump_vma(vma_mt);
 302			pr_emerg("mt piv: %p %lu - %lu\n", vma_mt,
 303				 mas.index, mas.last);
 304			pr_emerg("mt vma: %p %lu - %lu\n", vma_mt,
 305				 vma_mt->vm_start, vma_mt->vm_end);
 306
 307			mt_dump(mas.tree);
 308			if (vma_mt->vm_end != mas.last + 1) {
 309				pr_err("vma: %p vma_mt %lu-%lu\tmt %lu-%lu\n",
 310						mm, vma_mt->vm_start, vma_mt->vm_end,
 311						mas.index, mas.last);
 312				mt_dump(mas.tree);
 313			}
 314			VM_BUG_ON_MM(vma_mt->vm_end != mas.last + 1, mm);
 315			if (vma_mt->vm_start != mas.index) {
 316				pr_err("vma: %p vma_mt %p %lu - %lu doesn't match\n",
 317						mm, vma_mt, vma_mt->vm_start, vma_mt->vm_end);
 318				mt_dump(mas.tree);
 319			}
 320			VM_BUG_ON_MM(vma_mt->vm_start != mas.index, mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 321		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 322	}
 323}
 324
 325static void validate_mm(struct mm_struct *mm)
 326{
 327	int bug = 0;
 328	int i = 0;
 329	struct vm_area_struct *vma;
 330	MA_STATE(mas, &mm->mm_mt, 0, 0);
 331
 332	validate_mm_mt(mm);
 333
 334	mas_for_each(&mas, vma, ULONG_MAX) {
 335#ifdef CONFIG_DEBUG_VM_RB
 336		struct anon_vma *anon_vma = vma->anon_vma;
 337		struct anon_vma_chain *avc;
 338
 339		if (anon_vma) {
 340			anon_vma_lock_read(anon_vma);
 341			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 342				anon_vma_interval_tree_verify(avc);
 343			anon_vma_unlock_read(anon_vma);
 344		}
 345#endif
 
 
 346		i++;
 347	}
 348	if (i != mm->map_count) {
 349		pr_emerg("map_count %d mas_for_each %d\n", mm->map_count, i);
 
 
 
 
 
 
 
 
 
 
 
 350		bug = 1;
 351	}
 352	VM_BUG_ON_MM(bug, mm);
 353}
 354
 355#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
 356#define validate_mm_mt(root) do { } while (0)
 357#define validate_mm(mm) do { } while (0)
 358#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 359
 360/*
 361 * vma has some anon_vma assigned, and is already inserted on that
 362 * anon_vma's interval trees.
 363 *
 364 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 365 * vma must be removed from the anon_vma's interval trees using
 366 * anon_vma_interval_tree_pre_update_vma().
 367 *
 368 * After the update, the vma will be reinserted using
 369 * anon_vma_interval_tree_post_update_vma().
 370 *
 371 * The entire update must be protected by exclusive mmap_lock and by
 372 * the root anon_vma's mutex.
 373 */
 374static inline void
 375anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 376{
 377	struct anon_vma_chain *avc;
 378
 379	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 380		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 381}
 382
 383static inline void
 384anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 385{
 386	struct anon_vma_chain *avc;
 387
 388	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 389		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 390}
 391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 392static unsigned long count_vma_pages_range(struct mm_struct *mm,
 393		unsigned long addr, unsigned long end)
 394{
 395	VMA_ITERATOR(vmi, mm, addr);
 396	struct vm_area_struct *vma;
 397	unsigned long nr_pages = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 398
 399	for_each_vma_range(vmi, vma, end) {
 400		unsigned long vm_start = max(addr, vma->vm_start);
 401		unsigned long vm_end = min(end, vma->vm_end);
 402
 403		nr_pages += PHYS_PFN(vm_end - vm_start);
 
 404	}
 405
 406	return nr_pages;
 407}
 408
 409static void __vma_link_file(struct vm_area_struct *vma,
 410			    struct address_space *mapping)
 411{
 412	if (vma->vm_flags & VM_SHARED)
 413		mapping_allow_writable(mapping);
 
 
 
 414
 415	flush_dcache_mmap_lock(mapping);
 416	vma_interval_tree_insert(vma, &mapping->i_mmap);
 417	flush_dcache_mmap_unlock(mapping);
 
 
 
 
 
 
 
 
 
 
 418}
 419
 420/*
 421 * vma_mas_store() - Store a VMA in the maple tree.
 422 * @vma: The vm_area_struct
 423 * @mas: The maple state
 424 *
 425 * Efficient way to store a VMA in the maple tree when the @mas has already
 426 * walked to the correct location.
 427 *
 428 * Note: the end address is inclusive in the maple tree.
 429 */
 430void vma_mas_store(struct vm_area_struct *vma, struct ma_state *mas)
 431{
 432	trace_vma_store(mas->tree, vma);
 433	mas_set_range(mas, vma->vm_start, vma->vm_end - 1);
 434	mas_store_prealloc(mas, vma);
 435}
 436
 437/*
 438 * vma_mas_remove() - Remove a VMA from the maple tree.
 439 * @vma: The vm_area_struct
 440 * @mas: The maple state
 441 *
 442 * Efficient way to remove a VMA from the maple tree when the @mas has already
 443 * been established and points to the correct location.
 444 * Note: the end address is inclusive in the maple tree.
 445 */
 446void vma_mas_remove(struct vm_area_struct *vma, struct ma_state *mas)
 447{
 448	trace_vma_mas_szero(mas->tree, vma->vm_start, vma->vm_end - 1);
 449	mas->index = vma->vm_start;
 450	mas->last = vma->vm_end - 1;
 451	mas_store_prealloc(mas, NULL);
 452}
 453
 454/*
 455 * vma_mas_szero() - Set a given range to zero.  Used when modifying a
 456 * vm_area_struct start or end.
 457 *
 458 * @mas: The maple tree ma_state
 459 * @start: The start address to zero
 460 * @end: The end address to zero.
 461 */
 462static inline void vma_mas_szero(struct ma_state *mas, unsigned long start,
 463				unsigned long end)
 464{
 465	trace_vma_mas_szero(mas->tree, start, end - 1);
 466	mas_set_range(mas, start, end - 1);
 467	mas_store_prealloc(mas, NULL);
 468}
 469
 470static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
 
 
 471{
 472	MA_STATE(mas, &mm->mm_mt, 0, 0);
 473	struct address_space *mapping = NULL;
 474
 475	if (mas_preallocate(&mas, vma, GFP_KERNEL))
 476		return -ENOMEM;
 477
 478	if (vma->vm_file) {
 479		mapping = vma->vm_file->f_mapping;
 480		i_mmap_lock_write(mapping);
 481	}
 482
 483	vma_mas_store(vma, &mas);
 
 484
 485	if (mapping) {
 486		__vma_link_file(vma, mapping);
 487		i_mmap_unlock_write(mapping);
 488	}
 489
 490	mm->map_count++;
 491	validate_mm(mm);
 492	return 0;
 493}
 494
 495/*
 496 * vma_expand - Expand an existing VMA
 497 *
 498 * @mas: The maple state
 499 * @vma: The vma to expand
 500 * @start: The start of the vma
 501 * @end: The exclusive end of the vma
 502 * @pgoff: The page offset of vma
 503 * @next: The current of next vma.
 504 *
 505 * Expand @vma to @start and @end.  Can expand off the start and end.  Will
 506 * expand over @next if it's different from @vma and @end == @next->vm_end.
 507 * Checking if the @vma can expand and merge with @next needs to be handled by
 508 * the caller.
 509 *
 510 * Returns: 0 on success
 511 */
 512inline int vma_expand(struct ma_state *mas, struct vm_area_struct *vma,
 513		      unsigned long start, unsigned long end, pgoff_t pgoff,
 514		      struct vm_area_struct *next)
 515{
 516	struct mm_struct *mm = vma->vm_mm;
 517	struct address_space *mapping = NULL;
 518	struct rb_root_cached *root = NULL;
 519	struct anon_vma *anon_vma = vma->anon_vma;
 520	struct file *file = vma->vm_file;
 521	bool remove_next = false;
 522
 523	if (next && (vma != next) && (end == next->vm_end)) {
 524		remove_next = true;
 525		if (next->anon_vma && !vma->anon_vma) {
 526			int error;
 527
 528			anon_vma = next->anon_vma;
 529			vma->anon_vma = anon_vma;
 530			error = anon_vma_clone(vma, next);
 531			if (error)
 532				return error;
 533		}
 534	}
 535
 536	/* Not merging but overwriting any part of next is not handled. */
 537	VM_BUG_ON(next && !remove_next && next != vma && end > next->vm_start);
 538	/* Only handles expanding */
 539	VM_BUG_ON(vma->vm_start < start || vma->vm_end > end);
 540
 541	if (mas_preallocate(mas, vma, GFP_KERNEL))
 542		goto nomem;
 543
 544	vma_adjust_trans_huge(vma, start, end, 0);
 545
 546	if (file) {
 547		mapping = file->f_mapping;
 548		root = &mapping->i_mmap;
 549		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 550		i_mmap_lock_write(mapping);
 551	}
 552
 553	if (anon_vma) {
 554		anon_vma_lock_write(anon_vma);
 555		anon_vma_interval_tree_pre_update_vma(vma);
 556	}
 557
 558	if (file) {
 559		flush_dcache_mmap_lock(mapping);
 560		vma_interval_tree_remove(vma, root);
 561	}
 562
 563	vma->vm_start = start;
 564	vma->vm_end = end;
 565	vma->vm_pgoff = pgoff;
 566	/* Note: mas must be pointing to the expanding VMA */
 567	vma_mas_store(vma, mas);
 568
 569	if (file) {
 570		vma_interval_tree_insert(vma, root);
 571		flush_dcache_mmap_unlock(mapping);
 572	}
 573
 574	/* Expanding over the next vma */
 575	if (remove_next && file) {
 576		__remove_shared_vm_struct(next, file, mapping);
 577	}
 578
 579	if (anon_vma) {
 580		anon_vma_interval_tree_post_update_vma(vma);
 581		anon_vma_unlock_write(anon_vma);
 582	}
 
 
 583
 584	if (file) {
 585		i_mmap_unlock_write(mapping);
 586		uprobe_mmap(vma);
 587	}
 
 
 
 588
 589	if (remove_next) {
 590		if (file) {
 591			uprobe_munmap(next, next->vm_start, next->vm_end);
 592			fput(file);
 593		}
 594		if (next->anon_vma)
 595			anon_vma_merge(vma, next);
 596		mm->map_count--;
 597		mpol_put(vma_policy(next));
 598		vm_area_free(next);
 599	}
 
 
 600
 601	validate_mm(mm);
 602	return 0;
 
 603
 604nomem:
 605	return -ENOMEM;
 
 
 
 606}
 607
 608/*
 609 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 610 * is already present in an i_mmap tree without adjusting the tree.
 611 * The following helper function should be used when such adjustments
 612 * are necessary.  The "insert" vma (if any) is to be inserted
 613 * before we drop the necessary locks.
 614 */
 615int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
 616	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
 617	struct vm_area_struct *expand)
 618{
 619	struct mm_struct *mm = vma->vm_mm;
 620	struct vm_area_struct *next_next = NULL;	/* uninit var warning */
 621	struct vm_area_struct *next = find_vma(mm, vma->vm_end);
 622	struct vm_area_struct *orig_vma = vma;
 623	struct address_space *mapping = NULL;
 624	struct rb_root_cached *root = NULL;
 625	struct anon_vma *anon_vma = NULL;
 626	struct file *file = vma->vm_file;
 627	bool vma_changed = false;
 628	long adjust_next = 0;
 629	int remove_next = 0;
 630	MA_STATE(mas, &mm->mm_mt, 0, 0);
 631	struct vm_area_struct *exporter = NULL, *importer = NULL;
 632
 633	if (next && !insert) {
 
 
 634		if (end >= next->vm_end) {
 635			/*
 636			 * vma expands, overlapping all the next, and
 637			 * perhaps the one after too (mprotect case 6).
 638			 * The only other cases that gets here are
 639			 * case 1, case 7 and case 8.
 640			 */
 641			if (next == expand) {
 642				/*
 643				 * The only case where we don't expand "vma"
 644				 * and we expand "next" instead is case 8.
 645				 */
 646				VM_WARN_ON(end != next->vm_end);
 647				/*
 648				 * remove_next == 3 means we're
 649				 * removing "vma" and that to do so we
 650				 * swapped "vma" and "next".
 651				 */
 652				remove_next = 3;
 653				VM_WARN_ON(file != next->vm_file);
 654				swap(vma, next);
 655			} else {
 656				VM_WARN_ON(expand != vma);
 657				/*
 658				 * case 1, 6, 7, remove_next == 2 is case 6,
 659				 * remove_next == 1 is case 1 or 7.
 660				 */
 661				remove_next = 1 + (end > next->vm_end);
 662				if (remove_next == 2)
 663					next_next = find_vma(mm, next->vm_end);
 664
 665				VM_WARN_ON(remove_next == 2 &&
 666					   end != next_next->vm_end);
 
 
 
 
 667			}
 668
 669			exporter = next;
 670			importer = vma;
 671
 672			/*
 673			 * If next doesn't have anon_vma, import from vma after
 674			 * next, if the vma overlaps with it.
 675			 */
 676			if (remove_next == 2 && !next->anon_vma)
 677				exporter = next_next;
 678
 679		} else if (end > next->vm_start) {
 680			/*
 681			 * vma expands, overlapping part of the next:
 682			 * mprotect case 5 shifting the boundary up.
 683			 */
 684			adjust_next = (end - next->vm_start);
 685			exporter = next;
 686			importer = vma;
 687			VM_WARN_ON(expand != importer);
 688		} else if (end < vma->vm_end) {
 689			/*
 690			 * vma shrinks, and !insert tells it's not
 691			 * split_vma inserting another: so it must be
 692			 * mprotect case 4 shifting the boundary down.
 693			 */
 694			adjust_next = -(vma->vm_end - end);
 695			exporter = vma;
 696			importer = next;
 697			VM_WARN_ON(expand != importer);
 698		}
 699
 700		/*
 701		 * Easily overlooked: when mprotect shifts the boundary,
 702		 * make sure the expanding vma has anon_vma set if the
 703		 * shrinking vma had, to cover any anon pages imported.
 704		 */
 705		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 706			int error;
 707
 708			importer->anon_vma = exporter->anon_vma;
 709			error = anon_vma_clone(importer, exporter);
 710			if (error)
 711				return error;
 712		}
 713	}
 714
 715	if (mas_preallocate(&mas, vma, GFP_KERNEL))
 716		return -ENOMEM;
 717
 718	vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
 
 719	if (file) {
 720		mapping = file->f_mapping;
 721		root = &mapping->i_mmap;
 722		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 723
 724		if (adjust_next)
 725			uprobe_munmap(next, next->vm_start, next->vm_end);
 726
 727		i_mmap_lock_write(mapping);
 728		if (insert && insert->vm_file) {
 729			/*
 730			 * Put into interval tree now, so instantiated pages
 731			 * are visible to arm/parisc __flush_dcache_page
 732			 * throughout; but we cannot insert into address
 733			 * space until vma start or end is updated.
 734			 */
 735			__vma_link_file(insert, insert->vm_file->f_mapping);
 736		}
 737	}
 738
 739	anon_vma = vma->anon_vma;
 740	if (!anon_vma && adjust_next)
 741		anon_vma = next->anon_vma;
 742	if (anon_vma) {
 743		VM_WARN_ON(adjust_next && next->anon_vma &&
 744			   anon_vma != next->anon_vma);
 745		anon_vma_lock_write(anon_vma);
 746		anon_vma_interval_tree_pre_update_vma(vma);
 747		if (adjust_next)
 748			anon_vma_interval_tree_pre_update_vma(next);
 749	}
 750
 751	if (file) {
 752		flush_dcache_mmap_lock(mapping);
 753		vma_interval_tree_remove(vma, root);
 754		if (adjust_next)
 755			vma_interval_tree_remove(next, root);
 756	}
 757
 758	if (start != vma->vm_start) {
 759		if ((vma->vm_start < start) &&
 760		    (!insert || (insert->vm_end != start))) {
 761			vma_mas_szero(&mas, vma->vm_start, start);
 762			VM_WARN_ON(insert && insert->vm_start > vma->vm_start);
 763		} else {
 764			vma_changed = true;
 765		}
 766		vma->vm_start = start;
 
 767	}
 768	if (end != vma->vm_end) {
 769		if (vma->vm_end > end) {
 770			if (!insert || (insert->vm_start != end)) {
 771				vma_mas_szero(&mas, end, vma->vm_end);
 772				mas_reset(&mas);
 773				VM_WARN_ON(insert &&
 774					   insert->vm_end < vma->vm_end);
 775			}
 776		} else {
 777			vma_changed = true;
 778		}
 779		vma->vm_end = end;
 
 780	}
 781
 782	if (vma_changed)
 783		vma_mas_store(vma, &mas);
 784
 785	vma->vm_pgoff = pgoff;
 786	if (adjust_next) {
 787		next->vm_start += adjust_next;
 788		next->vm_pgoff += adjust_next >> PAGE_SHIFT;
 789		vma_mas_store(next, &mas);
 790	}
 791
 792	if (file) {
 793		if (adjust_next)
 794			vma_interval_tree_insert(next, root);
 795		vma_interval_tree_insert(vma, root);
 796		flush_dcache_mmap_unlock(mapping);
 797	}
 798
 799	if (remove_next && file) {
 800		__remove_shared_vm_struct(next, file, mapping);
 801		if (remove_next == 2)
 802			__remove_shared_vm_struct(next_next, file, mapping);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 803	} else if (insert) {
 804		/*
 805		 * split_vma has split insert from vma, and needs
 806		 * us to insert it before dropping the locks
 807		 * (it may either follow vma or precede it).
 808		 */
 809		mas_reset(&mas);
 810		vma_mas_store(insert, &mas);
 811		mm->map_count++;
 
 
 
 
 
 
 
 812	}
 813
 814	if (anon_vma) {
 815		anon_vma_interval_tree_post_update_vma(vma);
 816		if (adjust_next)
 817			anon_vma_interval_tree_post_update_vma(next);
 818		anon_vma_unlock_write(anon_vma);
 819	}
 820
 821	if (file) {
 822		i_mmap_unlock_write(mapping);
 
 
 823		uprobe_mmap(vma);
 824
 825		if (adjust_next)
 826			uprobe_mmap(next);
 827	}
 828
 829	if (remove_next) {
 830again:
 831		if (file) {
 832			uprobe_munmap(next, next->vm_start, next->vm_end);
 833			fput(file);
 834		}
 835		if (next->anon_vma)
 836			anon_vma_merge(vma, next);
 837		mm->map_count--;
 838		mpol_put(vma_policy(next));
 839		if (remove_next != 2)
 840			BUG_ON(vma->vm_end < next->vm_end);
 841		vm_area_free(next);
 842
 843		/*
 844		 * In mprotect's case 6 (see comments on vma_merge),
 845		 * we must remove next_next too.
 
 846		 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 847		if (remove_next == 2) {
 848			remove_next = 1;
 849			next = next_next;
 850			goto again;
 851		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 852	}
 853	if (insert && file)
 854		uprobe_mmap(insert);
 855
 856	mas_destroy(&mas);
 857	validate_mm(mm);
 858
 859	return 0;
 860}
 861
 862/*
 863 * If the vma has a ->close operation then the driver probably needs to release
 864 * per-vma resources, so we don't attempt to merge those.
 865 */
 866static inline int is_mergeable_vma(struct vm_area_struct *vma,
 867				struct file *file, unsigned long vm_flags,
 868				struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 869				struct anon_vma_name *anon_name)
 870{
 871	/*
 872	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
 873	 * match the flags but dirty bit -- the caller should mark
 874	 * merged VMA as dirty. If dirty bit won't be excluded from
 875	 * comparison, we increase pressure on the memory system forcing
 876	 * the kernel to generate new VMAs when old one could be
 877	 * extended instead.
 878	 */
 879	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
 880		return 0;
 881	if (vma->vm_file != file)
 882		return 0;
 883	if (vma->vm_ops && vma->vm_ops->close)
 884		return 0;
 885	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
 886		return 0;
 887	if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
 888		return 0;
 889	return 1;
 890}
 891
 892static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 893					struct anon_vma *anon_vma2,
 894					struct vm_area_struct *vma)
 895{
 896	/*
 897	 * The list_is_singular() test is to avoid merging VMA cloned from
 898	 * parents. This can improve scalability caused by anon_vma lock.
 899	 */
 900	if ((!anon_vma1 || !anon_vma2) && (!vma ||
 901		list_is_singular(&vma->anon_vma_chain)))
 902		return 1;
 903	return anon_vma1 == anon_vma2;
 904}
 905
 906/*
 907 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 908 * in front of (at a lower virtual address and file offset than) the vma.
 909 *
 910 * We cannot merge two vmas if they have differently assigned (non-NULL)
 911 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 912 *
 913 * We don't check here for the merged mmap wrapping around the end of pagecache
 914 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
 915 * wrap, nor mmaps which cover the final page at index -1UL.
 916 */
 917static int
 918can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 919		     struct anon_vma *anon_vma, struct file *file,
 920		     pgoff_t vm_pgoff,
 921		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 922		     struct anon_vma_name *anon_name)
 923{
 924	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
 925	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 926		if (vma->vm_pgoff == vm_pgoff)
 927			return 1;
 928	}
 929	return 0;
 930}
 931
 932/*
 933 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 934 * beyond (at a higher virtual address and file offset than) the vma.
 935 *
 936 * We cannot merge two vmas if they have differently assigned (non-NULL)
 937 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 938 */
 939static int
 940can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 941		    struct anon_vma *anon_vma, struct file *file,
 942		    pgoff_t vm_pgoff,
 943		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 944		    struct anon_vma_name *anon_name)
 945{
 946	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name) &&
 947	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 948		pgoff_t vm_pglen;
 949		vm_pglen = vma_pages(vma);
 950		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 951			return 1;
 952	}
 953	return 0;
 954}
 955
 956/*
 957 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
 958 * figure out whether that can be merged with its predecessor or its
 959 * successor.  Or both (it neatly fills a hole).
 960 *
 961 * In most cases - when called for mmap, brk or mremap - [addr,end) is
 962 * certain not to be mapped by the time vma_merge is called; but when
 963 * called for mprotect, it is certain to be already mapped (either at
 964 * an offset within prev, or at the start of next), and the flags of
 965 * this area are about to be changed to vm_flags - and the no-change
 966 * case has already been eliminated.
 967 *
 968 * The following mprotect cases have to be considered, where AAAA is
 969 * the area passed down from mprotect_fixup, never extending beyond one
 970 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
 971 *
 972 *     AAAA             AAAA                   AAAA
 973 *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPNNNNNN
 974 *    cannot merge    might become       might become
 975 *                    PPNNNNNNNNNN       PPPPPPPPPPNN
 976 *    mmap, brk or    case 4 below       case 5 below
 977 *    mremap move:
 978 *                        AAAA               AAAA
 979 *                    PPPP    NNNN       PPPPNNNNXXXX
 980 *                    might become       might become
 981 *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
 982 *                    PPPPPPPPNNNN 2 or  PPPPPPPPXXXX 7 or
 983 *                    PPPPNNNNNNNN 3     PPPPXXXXXXXX 8
 984 *
 985 * It is important for case 8 that the vma NNNN overlapping the
 986 * region AAAA is never going to extended over XXXX. Instead XXXX must
 987 * be extended in region AAAA and NNNN must be removed. This way in
 988 * all cases where vma_merge succeeds, the moment vma_adjust drops the
 989 * rmap_locks, the properties of the merged vma will be already
 990 * correct for the whole merged range. Some of those properties like
 991 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
 992 * be correct for the whole merged range immediately after the
 993 * rmap_locks are released. Otherwise if XXXX would be removed and
 994 * NNNN would be extended over the XXXX range, remove_migration_ptes
 995 * or other rmap walkers (if working on addresses beyond the "end"
 996 * parameter) may establish ptes with the wrong permissions of NNNN
 997 * instead of the right permissions of XXXX.
 998 */
 999struct vm_area_struct *vma_merge(struct mm_struct *mm,
1000			struct vm_area_struct *prev, unsigned long addr,
1001			unsigned long end, unsigned long vm_flags,
1002			struct anon_vma *anon_vma, struct file *file,
1003			pgoff_t pgoff, struct mempolicy *policy,
1004			struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
1005			struct anon_vma_name *anon_name)
1006{
1007	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1008	struct vm_area_struct *mid, *next, *res;
1009	int err = -1;
1010	bool merge_prev = false;
1011	bool merge_next = false;
1012
1013	/*
1014	 * We later require that vma->vm_flags == vm_flags,
1015	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1016	 */
1017	if (vm_flags & VM_SPECIAL)
1018		return NULL;
1019
1020	next = find_vma(mm, prev ? prev->vm_end : 0);
1021	mid = next;
1022	if (next && next->vm_end == end)		/* cases 6, 7, 8 */
1023		next = find_vma(mm, next->vm_end);
 
 
 
1024
1025	/* verify some invariant that must be enforced by the caller */
1026	VM_WARN_ON(prev && addr <= prev->vm_start);
1027	VM_WARN_ON(mid && end > mid->vm_end);
1028	VM_WARN_ON(addr >= end);
1029
1030	/* Can we merge the predecessor? */
 
 
1031	if (prev && prev->vm_end == addr &&
1032			mpol_equal(vma_policy(prev), policy) &&
1033			can_vma_merge_after(prev, vm_flags,
1034					    anon_vma, file, pgoff,
1035					    vm_userfaultfd_ctx, anon_name)) {
1036		merge_prev = true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1037	}
1038	/* Can we merge the successor? */
 
 
 
1039	if (next && end == next->vm_start &&
1040			mpol_equal(policy, vma_policy(next)) &&
1041			can_vma_merge_before(next, vm_flags,
1042					     anon_vma, file, pgoff+pglen,
1043					     vm_userfaultfd_ctx, anon_name)) {
1044		merge_next = true;
1045	}
1046	/* Can we merge both the predecessor and the successor? */
1047	if (merge_prev && merge_next &&
1048			is_mergeable_anon_vma(prev->anon_vma,
1049				next->anon_vma, NULL)) {	 /* cases 1, 6 */
1050		err = __vma_adjust(prev, prev->vm_start,
1051					next->vm_end, prev->vm_pgoff, NULL,
1052					prev);
1053		res = prev;
1054	} else if (merge_prev) {			/* cases 2, 5, 7 */
1055		err = __vma_adjust(prev, prev->vm_start,
1056					end, prev->vm_pgoff, NULL, prev);
1057		res = prev;
1058	} else if (merge_next) {
1059		if (prev && addr < prev->vm_end)	/* case 4 */
1060			err = __vma_adjust(prev, prev->vm_start,
1061					addr, prev->vm_pgoff, NULL, next);
1062		else					/* cases 3, 8 */
1063			err = __vma_adjust(mid, addr, next->vm_end,
1064					next->vm_pgoff - pglen, NULL, next);
1065		res = next;
 
 
 
 
 
 
 
 
 
 
1066	}
1067
1068	/*
1069	 * Cannot merge with predecessor or successor or error in __vma_adjust?
1070	 */
1071	if (err)
1072		return NULL;
1073	khugepaged_enter_vma(res, vm_flags);
1074	return res;
1075}
1076
1077/*
1078 * Rough compatibility check to quickly see if it's even worth looking
1079 * at sharing an anon_vma.
1080 *
1081 * They need to have the same vm_file, and the flags can only differ
1082 * in things that mprotect may change.
1083 *
1084 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1085 * we can merge the two vma's. For example, we refuse to merge a vma if
1086 * there is a vm_ops->close() function, because that indicates that the
1087 * driver is doing some kind of reference counting. But that doesn't
1088 * really matter for the anon_vma sharing case.
1089 */
1090static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1091{
1092	return a->vm_end == b->vm_start &&
1093		mpol_equal(vma_policy(a), vma_policy(b)) &&
1094		a->vm_file == b->vm_file &&
1095		!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1096		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1097}
1098
1099/*
1100 * Do some basic sanity checking to see if we can re-use the anon_vma
1101 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1102 * the same as 'old', the other will be the new one that is trying
1103 * to share the anon_vma.
1104 *
1105 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1106 * the anon_vma of 'old' is concurrently in the process of being set up
1107 * by another page fault trying to merge _that_. But that's ok: if it
1108 * is being set up, that automatically means that it will be a singleton
1109 * acceptable for merging, so we can do all of this optimistically. But
1110 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1111 *
1112 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1113 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1114 * is to return an anon_vma that is "complex" due to having gone through
1115 * a fork).
1116 *
1117 * We also make sure that the two vma's are compatible (adjacent,
1118 * and with the same memory policies). That's all stable, even with just
1119 * a read lock on the mmap_lock.
1120 */
1121static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1122{
1123	if (anon_vma_compatible(a, b)) {
1124		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1125
1126		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1127			return anon_vma;
1128	}
1129	return NULL;
1130}
1131
1132/*
1133 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1134 * neighbouring vmas for a suitable anon_vma, before it goes off
1135 * to allocate a new anon_vma.  It checks because a repetitive
1136 * sequence of mprotects and faults may otherwise lead to distinct
1137 * anon_vmas being allocated, preventing vma merge in subsequent
1138 * mprotect.
1139 */
1140struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1141{
1142	MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1143	struct anon_vma *anon_vma = NULL;
1144	struct vm_area_struct *prev, *next;
1145
1146	/* Try next first. */
1147	next = mas_walk(&mas);
1148	if (next) {
1149		anon_vma = reusable_anon_vma(next, vma, next);
1150		if (anon_vma)
1151			return anon_vma;
1152	}
1153
1154	prev = mas_prev(&mas, 0);
1155	VM_BUG_ON_VMA(prev != vma, vma);
1156	prev = mas_prev(&mas, 0);
1157	/* Try prev next. */
1158	if (prev)
1159		anon_vma = reusable_anon_vma(prev, prev, vma);
1160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1161	/*
1162	 * We might reach here with anon_vma == NULL if we can't find
1163	 * any reusable anon_vma.
1164	 * There's no absolute need to look only at touching neighbours:
1165	 * we could search further afield for "compatible" anon_vmas.
1166	 * But it would probably just be a waste of time searching,
1167	 * or lead to too many vmas hanging off the same anon_vma.
1168	 * We're trying to allow mprotect remerging later on,
1169	 * not trying to minimize memory used for anon_vmas.
1170	 */
1171	return anon_vma;
1172}
1173
1174/*
1175 * If a hint addr is less than mmap_min_addr change hint to be as
1176 * low as possible but still greater than mmap_min_addr
1177 */
1178static inline unsigned long round_hint_to_min(unsigned long hint)
1179{
1180	hint &= PAGE_MASK;
1181	if (((void *)hint != NULL) &&
1182	    (hint < mmap_min_addr))
1183		return PAGE_ALIGN(mmap_min_addr);
1184	return hint;
1185}
1186
1187int mlock_future_check(struct mm_struct *mm, unsigned long flags,
1188		       unsigned long len)
 
1189{
1190	unsigned long locked, lock_limit;
1191
1192	/*  mlock MCL_FUTURE? */
1193	if (flags & VM_LOCKED) {
1194		locked = len >> PAGE_SHIFT;
1195		locked += mm->locked_vm;
1196		lock_limit = rlimit(RLIMIT_MEMLOCK);
1197		lock_limit >>= PAGE_SHIFT;
1198		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1199			return -EAGAIN;
1200	}
1201	return 0;
1202}
1203
1204static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1205{
1206	if (S_ISREG(inode->i_mode))
1207		return MAX_LFS_FILESIZE;
1208
1209	if (S_ISBLK(inode->i_mode))
1210		return MAX_LFS_FILESIZE;
1211
1212	if (S_ISSOCK(inode->i_mode))
1213		return MAX_LFS_FILESIZE;
1214
1215	/* Special "we do even unsigned file positions" case */
1216	if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1217		return 0;
1218
1219	/* Yes, random drivers might want more. But I'm tired of buggy drivers */
1220	return ULONG_MAX;
1221}
1222
1223static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1224				unsigned long pgoff, unsigned long len)
1225{
1226	u64 maxsize = file_mmap_size_max(file, inode);
1227
1228	if (maxsize && len > maxsize)
1229		return false;
1230	maxsize -= len;
1231	if (pgoff > maxsize >> PAGE_SHIFT)
1232		return false;
1233	return true;
1234}
1235
1236/*
1237 * The caller must write-lock current->mm->mmap_lock.
1238 */
1239unsigned long do_mmap(struct file *file, unsigned long addr,
1240			unsigned long len, unsigned long prot,
1241			unsigned long flags, unsigned long pgoff,
1242			unsigned long *populate, struct list_head *uf)
 
1243{
1244	struct mm_struct *mm = current->mm;
1245	vm_flags_t vm_flags;
1246	int pkey = 0;
1247
1248	validate_mm(mm);
1249	*populate = 0;
1250
1251	if (!len)
1252		return -EINVAL;
1253
1254	/*
1255	 * Does the application expect PROT_READ to imply PROT_EXEC?
1256	 *
1257	 * (the exception is when the underlying filesystem is noexec
1258	 *  mounted, in which case we dont add PROT_EXEC.)
1259	 */
1260	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1261		if (!(file && path_noexec(&file->f_path)))
1262			prot |= PROT_EXEC;
1263
1264	/* force arch specific MAP_FIXED handling in get_unmapped_area */
1265	if (flags & MAP_FIXED_NOREPLACE)
1266		flags |= MAP_FIXED;
1267
1268	if (!(flags & MAP_FIXED))
1269		addr = round_hint_to_min(addr);
1270
1271	/* Careful about overflows.. */
1272	len = PAGE_ALIGN(len);
1273	if (!len)
1274		return -ENOMEM;
1275
1276	/* offset overflow? */
1277	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1278		return -EOVERFLOW;
1279
1280	/* Too many mappings? */
1281	if (mm->map_count > sysctl_max_map_count)
1282		return -ENOMEM;
1283
1284	/* Obtain the address to map to. we verify (or select) it and ensure
1285	 * that it represents a valid section of the address space.
1286	 */
1287	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1288	if (IS_ERR_VALUE(addr))
1289		return addr;
1290
1291	if (flags & MAP_FIXED_NOREPLACE) {
1292		if (find_vma_intersection(mm, addr, addr + len))
 
 
1293			return -EEXIST;
1294	}
1295
1296	if (prot == PROT_EXEC) {
1297		pkey = execute_only_pkey(mm);
1298		if (pkey < 0)
1299			pkey = 0;
1300	}
1301
1302	/* Do simple checking here so the lower-level routines won't have
1303	 * to. we assume access permissions have been handled by the open
1304	 * of the memory object, so we don't do any here.
1305	 */
1306	vm_flags = calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1307			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1308
1309	if (flags & MAP_LOCKED)
1310		if (!can_do_mlock())
1311			return -EPERM;
1312
1313	if (mlock_future_check(mm, vm_flags, len))
1314		return -EAGAIN;
1315
1316	if (file) {
1317		struct inode *inode = file_inode(file);
1318		unsigned long flags_mask;
1319
1320		if (!file_mmap_ok(file, inode, pgoff, len))
1321			return -EOVERFLOW;
1322
1323		flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1324
1325		switch (flags & MAP_TYPE) {
1326		case MAP_SHARED:
1327			/*
1328			 * Force use of MAP_SHARED_VALIDATE with non-legacy
1329			 * flags. E.g. MAP_SYNC is dangerous to use with
1330			 * MAP_SHARED as you don't know which consistency model
1331			 * you will get. We silently ignore unsupported flags
1332			 * with MAP_SHARED to preserve backward compatibility.
1333			 */
1334			flags &= LEGACY_MAP_MASK;
1335			fallthrough;
1336		case MAP_SHARED_VALIDATE:
1337			if (flags & ~flags_mask)
1338				return -EOPNOTSUPP;
1339			if (prot & PROT_WRITE) {
1340				if (!(file->f_mode & FMODE_WRITE))
1341					return -EACCES;
1342				if (IS_SWAPFILE(file->f_mapping->host))
1343					return -ETXTBSY;
1344			}
1345
1346			/*
1347			 * Make sure we don't allow writing to an append-only
1348			 * file..
1349			 */
1350			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1351				return -EACCES;
1352
 
 
 
 
 
 
1353			vm_flags |= VM_SHARED | VM_MAYSHARE;
1354			if (!(file->f_mode & FMODE_WRITE))
1355				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1356			fallthrough;
 
1357		case MAP_PRIVATE:
1358			if (!(file->f_mode & FMODE_READ))
1359				return -EACCES;
1360			if (path_noexec(&file->f_path)) {
1361				if (vm_flags & VM_EXEC)
1362					return -EPERM;
1363				vm_flags &= ~VM_MAYEXEC;
1364			}
1365
1366			if (!file->f_op->mmap)
1367				return -ENODEV;
1368			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1369				return -EINVAL;
1370			break;
1371
1372		default:
1373			return -EINVAL;
1374		}
1375	} else {
1376		switch (flags & MAP_TYPE) {
1377		case MAP_SHARED:
1378			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1379				return -EINVAL;
1380			/*
1381			 * Ignore pgoff.
1382			 */
1383			pgoff = 0;
1384			vm_flags |= VM_SHARED | VM_MAYSHARE;
1385			break;
1386		case MAP_PRIVATE:
1387			/*
1388			 * Set pgoff according to addr for anon_vma.
1389			 */
1390			pgoff = addr >> PAGE_SHIFT;
1391			break;
1392		default:
1393			return -EINVAL;
1394		}
1395	}
1396
1397	/*
1398	 * Set 'VM_NORESERVE' if we should not account for the
1399	 * memory use of this mapping.
1400	 */
1401	if (flags & MAP_NORESERVE) {
1402		/* We honor MAP_NORESERVE if allowed to overcommit */
1403		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1404			vm_flags |= VM_NORESERVE;
1405
1406		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1407		if (file && is_file_hugepages(file))
1408			vm_flags |= VM_NORESERVE;
1409	}
1410
1411	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1412	if (!IS_ERR_VALUE(addr) &&
1413	    ((vm_flags & VM_LOCKED) ||
1414	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1415		*populate = len;
1416	return addr;
1417}
1418
1419unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1420			      unsigned long prot, unsigned long flags,
1421			      unsigned long fd, unsigned long pgoff)
1422{
1423	struct file *file = NULL;
1424	unsigned long retval;
1425
1426	if (!(flags & MAP_ANONYMOUS)) {
1427		audit_mmap_fd(fd, flags);
1428		file = fget(fd);
1429		if (!file)
1430			return -EBADF;
1431		if (is_file_hugepages(file)) {
1432			len = ALIGN(len, huge_page_size(hstate_file(file)));
1433		} else if (unlikely(flags & MAP_HUGETLB)) {
1434			retval = -EINVAL;
1435			goto out_fput;
1436		}
1437	} else if (flags & MAP_HUGETLB) {
 
1438		struct hstate *hs;
1439
1440		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1441		if (!hs)
1442			return -EINVAL;
1443
1444		len = ALIGN(len, huge_page_size(hs));
1445		/*
1446		 * VM_NORESERVE is used because the reservations will be
1447		 * taken when vm_ops->mmap() is called
 
 
1448		 */
1449		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1450				VM_NORESERVE,
1451				HUGETLB_ANONHUGE_INODE,
1452				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1453		if (IS_ERR(file))
1454			return PTR_ERR(file);
1455	}
1456
 
 
1457	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1458out_fput:
1459	if (file)
1460		fput(file);
1461	return retval;
1462}
1463
1464SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1465		unsigned long, prot, unsigned long, flags,
1466		unsigned long, fd, unsigned long, pgoff)
1467{
1468	return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1469}
1470
1471#ifdef __ARCH_WANT_SYS_OLD_MMAP
1472struct mmap_arg_struct {
1473	unsigned long addr;
1474	unsigned long len;
1475	unsigned long prot;
1476	unsigned long flags;
1477	unsigned long fd;
1478	unsigned long offset;
1479};
1480
1481SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1482{
1483	struct mmap_arg_struct a;
1484
1485	if (copy_from_user(&a, arg, sizeof(a)))
1486		return -EFAULT;
1487	if (offset_in_page(a.offset))
1488		return -EINVAL;
1489
1490	return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1491			       a.offset >> PAGE_SHIFT);
1492}
1493#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1494
1495/*
1496 * Some shared mappings will want the pages marked read-only
1497 * to track write events. If so, we'll downgrade vm_page_prot
1498 * to the private version (using protection_map[] without the
1499 * VM_SHARED bit).
1500 */
1501int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1502{
1503	vm_flags_t vm_flags = vma->vm_flags;
1504	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1505
1506	/* If it was private or non-writable, the write bit is already clear */
1507	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1508		return 0;
1509
1510	/* The backer wishes to know when pages are first written to? */
1511	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1512		return 1;
1513
1514	/* The open routine did something to the protections that pgprot_modify
1515	 * won't preserve? */
1516	if (pgprot_val(vm_page_prot) !=
1517	    pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1518		return 0;
1519
1520	/*
1521	 * Do we need to track softdirty? hugetlb does not support softdirty
1522	 * tracking yet.
1523	 */
1524	if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1525		return 1;
1526
1527	/* Do we need write faults for uffd-wp tracking? */
1528	if (userfaultfd_wp(vma))
1529		return 1;
1530
1531	/* Specialty mapping? */
1532	if (vm_flags & VM_PFNMAP)
1533		return 0;
1534
1535	/* Can the mapping track the dirty pages? */
1536	return vma->vm_file && vma->vm_file->f_mapping &&
1537		mapping_can_writeback(vma->vm_file->f_mapping);
1538}
1539
1540/*
1541 * We account for memory if it's a private writeable mapping,
1542 * not hugepages and VM_NORESERVE wasn't set.
1543 */
1544static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1545{
1546	/*
1547	 * hugetlb has its own accounting separate from the core VM
1548	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1549	 */
1550	if (file && is_file_hugepages(file))
1551		return 0;
1552
1553	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1554}
1555
1556/**
1557 * unmapped_area() - Find an area between the low_limit and the high_limit with
1558 * the correct alignment and offset, all from @info. Note: current->mm is used
1559 * for the search.
1560 *
1561 * @info: The unmapped area information including the range (low_limit -
1562 * hight_limit), the alignment offset and mask.
1563 *
1564 * Return: A memory address or -ENOMEM.
1565 */
1566static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1567{
1568	unsigned long length, gap;
 
 
 
 
1569
1570	MA_STATE(mas, &current->mm->mm_mt, 0, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1571
1572	/* Adjust search length to account for worst case alignment overhead */
1573	length = info->length + info->align_mask;
1574	if (length < info->length)
1575		return -ENOMEM;
1576
1577	if (mas_empty_area(&mas, info->low_limit, info->high_limit - 1,
1578				  length))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1579		return -ENOMEM;
1580
1581	gap = mas.index;
1582	gap += (info->align_offset - gap) & info->align_mask;
1583	return gap;
 
 
 
 
 
 
 
 
1584}
1585
1586/**
1587 * unmapped_area_topdown() - Find an area between the low_limit and the
1588 * high_limit with * the correct alignment and offset at the highest available
1589 * address, all from @info. Note: current->mm is used for the search.
1590 *
1591 * @info: The unmapped area information including the range (low_limit -
1592 * hight_limit), the alignment offset and mask.
1593 *
1594 * Return: A memory address or -ENOMEM.
1595 */
1596static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1597{
1598	unsigned long length, gap;
 
 
1599
1600	MA_STATE(mas, &current->mm->mm_mt, 0, 0);
1601	/* Adjust search length to account for worst case alignment overhead */
1602	length = info->length + info->align_mask;
1603	if (length < info->length)
1604		return -ENOMEM;
1605
1606	if (mas_empty_area_rev(&mas, info->low_limit, info->high_limit - 1,
1607				length))
 
 
 
 
1608		return -ENOMEM;
 
1609
1610	gap = mas.last + 1 - info->length;
1611	gap -= (gap - info->align_offset) & info->align_mask;
1612	return gap;
1613}
1614
1615/*
1616 * Search for an unmapped address range.
1617 *
1618 * We are looking for a range that:
1619 * - does not intersect with any VMA;
1620 * - is contained within the [low_limit, high_limit) interval;
1621 * - is at least the desired size.
1622 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1623 */
1624unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1625{
1626	unsigned long addr;
1627
1628	if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1629		addr = unmapped_area_topdown(info);
1630	else
1631		addr = unmapped_area(info);
 
 
1632
1633	trace_vm_unmapped_area(addr, info);
1634	return addr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1635}
1636
1637/* Get an address range which is currently unmapped.
1638 * For shmat() with addr=0.
1639 *
1640 * Ugly calling convention alert:
1641 * Return value with the low bits set means error value,
1642 * ie
1643 *	if (ret & ~PAGE_MASK)
1644 *		error = ret;
1645 *
1646 * This function "knows" that -ENOMEM has the bits set.
1647 */
 
1648unsigned long
1649generic_get_unmapped_area(struct file *filp, unsigned long addr,
1650			  unsigned long len, unsigned long pgoff,
1651			  unsigned long flags)
1652{
1653	struct mm_struct *mm = current->mm;
1654	struct vm_area_struct *vma, *prev;
1655	struct vm_unmapped_area_info info;
1656	const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1657
1658	if (len > mmap_end - mmap_min_addr)
1659		return -ENOMEM;
1660
1661	if (flags & MAP_FIXED)
1662		return addr;
1663
1664	if (addr) {
1665		addr = PAGE_ALIGN(addr);
1666		vma = find_vma_prev(mm, addr, &prev);
1667		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1668		    (!vma || addr + len <= vm_start_gap(vma)) &&
1669		    (!prev || addr >= vm_end_gap(prev)))
1670			return addr;
1671	}
1672
1673	info.flags = 0;
1674	info.length = len;
1675	info.low_limit = mm->mmap_base;
1676	info.high_limit = mmap_end;
1677	info.align_mask = 0;
1678	info.align_offset = 0;
1679	return vm_unmapped_area(&info);
1680}
1681
1682#ifndef HAVE_ARCH_UNMAPPED_AREA
1683unsigned long
1684arch_get_unmapped_area(struct file *filp, unsigned long addr,
1685		       unsigned long len, unsigned long pgoff,
1686		       unsigned long flags)
1687{
1688	return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1689}
1690#endif
1691
1692/*
1693 * This mmap-allocator allocates new areas top-down from below the
1694 * stack's low limit (the base):
1695 */
 
1696unsigned long
1697generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1698				  unsigned long len, unsigned long pgoff,
1699				  unsigned long flags)
1700{
1701	struct vm_area_struct *vma, *prev;
1702	struct mm_struct *mm = current->mm;
 
1703	struct vm_unmapped_area_info info;
1704	const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1705
1706	/* requested length too big for entire address space */
1707	if (len > mmap_end - mmap_min_addr)
1708		return -ENOMEM;
1709
1710	if (flags & MAP_FIXED)
1711		return addr;
1712
1713	/* requesting a specific address */
1714	if (addr) {
1715		addr = PAGE_ALIGN(addr);
1716		vma = find_vma_prev(mm, addr, &prev);
1717		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1718				(!vma || addr + len <= vm_start_gap(vma)) &&
1719				(!prev || addr >= vm_end_gap(prev)))
1720			return addr;
1721	}
1722
1723	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1724	info.length = len;
1725	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
1726	info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1727	info.align_mask = 0;
1728	info.align_offset = 0;
1729	addr = vm_unmapped_area(&info);
1730
1731	/*
1732	 * A failed mmap() very likely causes application failure,
1733	 * so fall back to the bottom-up function here. This scenario
1734	 * can happen with large stack limits and large mmap()
1735	 * allocations.
1736	 */
1737	if (offset_in_page(addr)) {
1738		VM_BUG_ON(addr != -ENOMEM);
1739		info.flags = 0;
1740		info.low_limit = TASK_UNMAPPED_BASE;
1741		info.high_limit = mmap_end;
1742		addr = vm_unmapped_area(&info);
1743	}
1744
1745	return addr;
1746}
1747
1748#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1749unsigned long
1750arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751			       unsigned long len, unsigned long pgoff,
1752			       unsigned long flags)
1753{
1754	return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1755}
1756#endif
1757
1758unsigned long
1759get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1760		unsigned long pgoff, unsigned long flags)
1761{
1762	unsigned long (*get_area)(struct file *, unsigned long,
1763				  unsigned long, unsigned long, unsigned long);
1764
1765	unsigned long error = arch_mmap_check(addr, len, flags);
1766	if (error)
1767		return error;
1768
1769	/* Careful about overflows.. */
1770	if (len > TASK_SIZE)
1771		return -ENOMEM;
1772
1773	get_area = current->mm->get_unmapped_area;
1774	if (file) {
1775		if (file->f_op->get_unmapped_area)
1776			get_area = file->f_op->get_unmapped_area;
1777	} else if (flags & MAP_SHARED) {
1778		/*
1779		 * mmap_region() will call shmem_zero_setup() to create a file,
1780		 * so use shmem's get_unmapped_area in case it can be huge.
1781		 * do_mmap() will clear pgoff, so match alignment.
1782		 */
1783		pgoff = 0;
1784		get_area = shmem_get_unmapped_area;
1785	}
1786
1787	addr = get_area(file, addr, len, pgoff, flags);
1788	if (IS_ERR_VALUE(addr))
1789		return addr;
1790
1791	if (addr > TASK_SIZE - len)
1792		return -ENOMEM;
1793	if (offset_in_page(addr))
1794		return -EINVAL;
1795
1796	error = security_mmap_addr(addr);
1797	return error ? error : addr;
1798}
1799
1800EXPORT_SYMBOL(get_unmapped_area);
1801
1802/**
1803 * find_vma_intersection() - Look up the first VMA which intersects the interval
1804 * @mm: The process address space.
1805 * @start_addr: The inclusive start user address.
1806 * @end_addr: The exclusive end user address.
1807 *
1808 * Returns: The first VMA within the provided range, %NULL otherwise.  Assumes
1809 * start_addr < end_addr.
1810 */
1811struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1812					     unsigned long start_addr,
1813					     unsigned long end_addr)
1814{
1815	unsigned long index = start_addr;
 
1816
1817	mmap_assert_locked(mm);
1818	return mt_find(&mm->mm_mt, &index, end_addr - 1);
1819}
1820EXPORT_SYMBOL(find_vma_intersection);
1821
1822/**
1823 * find_vma() - Find the VMA for a given address, or the next VMA.
1824 * @mm: The mm_struct to check
1825 * @addr: The address
1826 *
1827 * Returns: The VMA associated with addr, or the next VMA.
1828 * May return %NULL in the case of no VMA at addr or above.
1829 */
1830struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1831{
1832	unsigned long index = addr;
1833
1834	mmap_assert_locked(mm);
1835	return mt_find(&mm->mm_mt, &index, ULONG_MAX);
 
 
 
 
 
 
 
 
 
 
1836}
 
1837EXPORT_SYMBOL(find_vma);
1838
1839/**
1840 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1841 * set %pprev to the previous VMA, if any.
1842 * @mm: The mm_struct to check
1843 * @addr: The address
1844 * @pprev: The pointer to set to the previous VMA
1845 *
1846 * Note that RCU lock is missing here since the external mmap_lock() is used
1847 * instead.
1848 *
1849 * Returns: The VMA associated with @addr, or the next vma.
1850 * May return %NULL in the case of no vma at addr or above.
1851 */
1852struct vm_area_struct *
1853find_vma_prev(struct mm_struct *mm, unsigned long addr,
1854			struct vm_area_struct **pprev)
1855{
1856	struct vm_area_struct *vma;
1857	MA_STATE(mas, &mm->mm_mt, addr, addr);
1858
1859	vma = mas_walk(&mas);
1860	*pprev = mas_prev(&mas, 0);
1861	if (!vma)
1862		vma = mas_next(&mas, ULONG_MAX);
 
 
 
 
 
 
 
1863	return vma;
1864}
1865
1866/*
1867 * Verify that the stack growth is acceptable and
1868 * update accounting. This is shared with both the
1869 * grow-up and grow-down cases.
1870 */
1871static int acct_stack_growth(struct vm_area_struct *vma,
1872			     unsigned long size, unsigned long grow)
1873{
1874	struct mm_struct *mm = vma->vm_mm;
1875	unsigned long new_start;
1876
1877	/* address space limit tests */
1878	if (!may_expand_vm(mm, vma->vm_flags, grow))
1879		return -ENOMEM;
1880
1881	/* Stack limit test */
1882	if (size > rlimit(RLIMIT_STACK))
1883		return -ENOMEM;
1884
1885	/* mlock limit tests */
1886	if (mlock_future_check(mm, vma->vm_flags, grow << PAGE_SHIFT))
1887		return -ENOMEM;
 
 
 
 
 
 
 
1888
1889	/* Check to ensure the stack will not grow into a hugetlb-only region */
1890	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1891			vma->vm_end - size;
1892	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1893		return -EFAULT;
1894
1895	/*
1896	 * Overcommit..  This must be the final test, as it will
1897	 * update security statistics.
1898	 */
1899	if (security_vm_enough_memory_mm(mm, grow))
1900		return -ENOMEM;
1901
1902	return 0;
1903}
1904
1905#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1906/*
1907 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1908 * vma is the last one with address > vma->vm_end.  Have to extend vma.
1909 */
1910int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1911{
1912	struct mm_struct *mm = vma->vm_mm;
1913	struct vm_area_struct *next;
1914	unsigned long gap_addr;
1915	int error = 0;
1916	MA_STATE(mas, &mm->mm_mt, 0, 0);
1917
1918	if (!(vma->vm_flags & VM_GROWSUP))
1919		return -EFAULT;
1920
1921	/* Guard against exceeding limits of the address space. */
1922	address &= PAGE_MASK;
1923	if (address >= (TASK_SIZE & PAGE_MASK))
1924		return -ENOMEM;
1925	address += PAGE_SIZE;
1926
1927	/* Enforce stack_guard_gap */
1928	gap_addr = address + stack_guard_gap;
1929
1930	/* Guard against overflow */
1931	if (gap_addr < address || gap_addr > TASK_SIZE)
1932		gap_addr = TASK_SIZE;
1933
1934	next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1935	if (next && vma_is_accessible(next)) {
 
1936		if (!(next->vm_flags & VM_GROWSUP))
1937			return -ENOMEM;
1938		/* Check that both stack segments have the same anon_vma? */
1939	}
1940
1941	if (mas_preallocate(&mas, vma, GFP_KERNEL))
1942		return -ENOMEM;
1943
1944	/* We must make sure the anon_vma is allocated. */
1945	if (unlikely(anon_vma_prepare(vma))) {
1946		mas_destroy(&mas);
1947		return -ENOMEM;
1948	}
1949
1950	/*
1951	 * vma->vm_start/vm_end cannot change under us because the caller
1952	 * is required to hold the mmap_lock in read mode.  We need the
1953	 * anon_vma lock to serialize against concurrent expand_stacks.
1954	 */
1955	anon_vma_lock_write(vma->anon_vma);
1956
1957	/* Somebody else might have raced and expanded it already */
1958	if (address > vma->vm_end) {
1959		unsigned long size, grow;
1960
1961		size = address - vma->vm_start;
1962		grow = (address - vma->vm_end) >> PAGE_SHIFT;
1963
1964		error = -ENOMEM;
1965		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1966			error = acct_stack_growth(vma, size, grow);
1967			if (!error) {
1968				/*
1969				 * We only hold a shared mmap_lock lock here, so
1970				 * we need to protect against concurrent vma
1971				 * expansions.  anon_vma_lock_write() doesn't
1972				 * help here, as we don't guarantee that all
1973				 * growable vmas in a mm share the same root
1974				 * anon vma.  So, we reuse mm->page_table_lock
1975				 * to guard against concurrent vma expansions.
 
 
1976				 */
1977				spin_lock(&mm->page_table_lock);
1978				if (vma->vm_flags & VM_LOCKED)
1979					mm->locked_vm += grow;
1980				vm_stat_account(mm, vma->vm_flags, grow);
1981				anon_vma_interval_tree_pre_update_vma(vma);
1982				vma->vm_end = address;
1983				/* Overwrite old entry in mtree. */
1984				vma_mas_store(vma, &mas);
1985				anon_vma_interval_tree_post_update_vma(vma);
 
 
 
 
1986				spin_unlock(&mm->page_table_lock);
1987
1988				perf_event_mmap(vma);
1989			}
1990		}
1991	}
1992	anon_vma_unlock_write(vma->anon_vma);
1993	khugepaged_enter_vma(vma, vma->vm_flags);
1994	mas_destroy(&mas);
1995	return error;
1996}
1997#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1998
1999/*
2000 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2001 */
2002int expand_downwards(struct vm_area_struct *vma, unsigned long address)
 
2003{
2004	struct mm_struct *mm = vma->vm_mm;
2005	MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2006	struct vm_area_struct *prev;
2007	int error = 0;
2008
2009	address &= PAGE_MASK;
2010	if (address < mmap_min_addr)
2011		return -EPERM;
 
2012
2013	/* Enforce stack_guard_gap */
2014	prev = mas_prev(&mas, 0);
2015	/* Check that both stack segments have the same anon_vma? */
2016	if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2017			vma_is_accessible(prev)) {
2018		if (address - prev->vm_end < stack_guard_gap)
2019			return -ENOMEM;
2020	}
2021
2022	if (mas_preallocate(&mas, vma, GFP_KERNEL))
2023		return -ENOMEM;
2024
2025	/* We must make sure the anon_vma is allocated. */
2026	if (unlikely(anon_vma_prepare(vma))) {
2027		mas_destroy(&mas);
2028		return -ENOMEM;
2029	}
2030
2031	/*
2032	 * vma->vm_start/vm_end cannot change under us because the caller
2033	 * is required to hold the mmap_lock in read mode.  We need the
2034	 * anon_vma lock to serialize against concurrent expand_stacks.
2035	 */
2036	anon_vma_lock_write(vma->anon_vma);
2037
2038	/* Somebody else might have raced and expanded it already */
2039	if (address < vma->vm_start) {
2040		unsigned long size, grow;
2041
2042		size = vma->vm_end - address;
2043		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2044
2045		error = -ENOMEM;
2046		if (grow <= vma->vm_pgoff) {
2047			error = acct_stack_growth(vma, size, grow);
2048			if (!error) {
2049				/*
2050				 * We only hold a shared mmap_lock lock here, so
2051				 * we need to protect against concurrent vma
2052				 * expansions.  anon_vma_lock_write() doesn't
2053				 * help here, as we don't guarantee that all
2054				 * growable vmas in a mm share the same root
2055				 * anon vma.  So, we reuse mm->page_table_lock
2056				 * to guard against concurrent vma expansions.
 
 
2057				 */
2058				spin_lock(&mm->page_table_lock);
2059				if (vma->vm_flags & VM_LOCKED)
2060					mm->locked_vm += grow;
2061				vm_stat_account(mm, vma->vm_flags, grow);
2062				anon_vma_interval_tree_pre_update_vma(vma);
2063				vma->vm_start = address;
2064				vma->vm_pgoff -= grow;
2065				/* Overwrite old entry in mtree. */
2066				vma_mas_store(vma, &mas);
2067				anon_vma_interval_tree_post_update_vma(vma);
 
2068				spin_unlock(&mm->page_table_lock);
2069
2070				perf_event_mmap(vma);
2071			}
2072		}
2073	}
2074	anon_vma_unlock_write(vma->anon_vma);
2075	khugepaged_enter_vma(vma, vma->vm_flags);
2076	mas_destroy(&mas);
2077	return error;
2078}
2079
2080/* enforced gap between the expanding stack and other mappings. */
2081unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2082
2083static int __init cmdline_parse_stack_guard_gap(char *p)
2084{
2085	unsigned long val;
2086	char *endptr;
2087
2088	val = simple_strtoul(p, &endptr, 10);
2089	if (!*endptr)
2090		stack_guard_gap = val << PAGE_SHIFT;
2091
2092	return 1;
2093}
2094__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2095
2096#ifdef CONFIG_STACK_GROWSUP
2097int expand_stack(struct vm_area_struct *vma, unsigned long address)
2098{
2099	return expand_upwards(vma, address);
2100}
2101
2102struct vm_area_struct *
2103find_extend_vma(struct mm_struct *mm, unsigned long addr)
2104{
2105	struct vm_area_struct *vma, *prev;
2106
2107	addr &= PAGE_MASK;
2108	vma = find_vma_prev(mm, addr, &prev);
2109	if (vma && (vma->vm_start <= addr))
2110		return vma;
2111	if (!prev || expand_stack(prev, addr))
2112		return NULL;
2113	if (prev->vm_flags & VM_LOCKED)
2114		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2115	return prev;
2116}
2117#else
2118int expand_stack(struct vm_area_struct *vma, unsigned long address)
2119{
2120	return expand_downwards(vma, address);
2121}
2122
2123struct vm_area_struct *
2124find_extend_vma(struct mm_struct *mm, unsigned long addr)
2125{
2126	struct vm_area_struct *vma;
2127	unsigned long start;
2128
2129	addr &= PAGE_MASK;
2130	vma = find_vma(mm, addr);
2131	if (!vma)
2132		return NULL;
2133	if (vma->vm_start <= addr)
2134		return vma;
2135	if (!(vma->vm_flags & VM_GROWSDOWN))
2136		return NULL;
2137	start = vma->vm_start;
2138	if (expand_stack(vma, addr))
2139		return NULL;
2140	if (vma->vm_flags & VM_LOCKED)
2141		populate_vma_page_range(vma, addr, start, NULL);
2142	return vma;
2143}
2144#endif
2145
2146EXPORT_SYMBOL_GPL(find_extend_vma);
2147
2148/*
2149 * Ok - we have the memory areas we should free on a maple tree so release them,
2150 * and do the vma updates.
2151 *
2152 * Called with the mm semaphore held.
2153 */
2154static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2155{
2156	unsigned long nr_accounted = 0;
2157	struct vm_area_struct *vma;
2158
2159	/* Update high watermark before we lower total_vm */
2160	update_hiwater_vm(mm);
2161	mas_for_each(mas, vma, ULONG_MAX) {
2162		long nrpages = vma_pages(vma);
2163
2164		if (vma->vm_flags & VM_ACCOUNT)
2165			nr_accounted += nrpages;
2166		vm_stat_account(mm, vma->vm_flags, -nrpages);
2167		remove_vma(vma);
2168	}
2169	vm_unacct_memory(nr_accounted);
2170	validate_mm(mm);
2171}
2172
2173/*
2174 * Get rid of page table information in the indicated region.
2175 *
2176 * Called with the mm semaphore held.
2177 */
2178static void unmap_region(struct mm_struct *mm, struct maple_tree *mt,
2179		struct vm_area_struct *vma, struct vm_area_struct *prev,
2180		struct vm_area_struct *next,
2181		unsigned long start, unsigned long end)
2182{
 
2183	struct mmu_gather tlb;
2184
2185	lru_add_drain();
2186	tlb_gather_mmu(&tlb, mm);
2187	update_hiwater_rss(mm);
2188	unmap_vmas(&tlb, mt, vma, start, end);
2189	free_pgtables(&tlb, mt, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2190				 next ? next->vm_start : USER_PGTABLES_CEILING);
2191	tlb_finish_mmu(&tlb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2192}
2193
2194/*
2195 * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2196 * has already been checked or doesn't make sense to fail.
2197 */
2198int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2199		unsigned long addr, int new_below)
2200{
2201	struct vm_area_struct *new;
2202	int err;
2203	validate_mm_mt(mm);
2204
2205	if (vma->vm_ops && vma->vm_ops->may_split) {
2206		err = vma->vm_ops->may_split(vma, addr);
2207		if (err)
2208			return err;
2209	}
2210
2211	new = vm_area_dup(vma);
2212	if (!new)
2213		return -ENOMEM;
2214
 
 
 
 
 
2215	if (new_below)
2216		new->vm_end = addr;
2217	else {
2218		new->vm_start = addr;
2219		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2220	}
2221
2222	err = vma_dup_policy(vma, new);
2223	if (err)
2224		goto out_free_vma;
2225
2226	err = anon_vma_clone(new, vma);
2227	if (err)
2228		goto out_free_mpol;
2229
2230	if (new->vm_file)
2231		get_file(new->vm_file);
2232
2233	if (new->vm_ops && new->vm_ops->open)
2234		new->vm_ops->open(new);
2235
2236	if (new_below)
2237		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2238			((addr - new->vm_start) >> PAGE_SHIFT), new);
2239	else
2240		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2241
2242	/* Success. */
2243	if (!err)
2244		return 0;
2245
2246	/* Avoid vm accounting in close() operation */
2247	new->vm_start = new->vm_end;
2248	new->vm_pgoff = 0;
2249	/* Clean everything up if vma_adjust failed. */
2250	if (new->vm_ops && new->vm_ops->close)
2251		new->vm_ops->close(new);
2252	if (new->vm_file)
2253		fput(new->vm_file);
2254	unlink_anon_vmas(new);
2255 out_free_mpol:
2256	mpol_put(vma_policy(new));
2257 out_free_vma:
2258	vm_area_free(new);
2259	validate_mm_mt(mm);
2260	return err;
2261}
2262
2263/*
2264 * Split a vma into two pieces at address 'addr', a new vma is allocated
2265 * either for the first part or the tail.
2266 */
2267int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2268	      unsigned long addr, int new_below)
2269{
2270	if (mm->map_count >= sysctl_max_map_count)
2271		return -ENOMEM;
2272
2273	return __split_vma(mm, vma, addr, new_below);
2274}
2275
2276static inline int munmap_sidetree(struct vm_area_struct *vma,
2277				   struct ma_state *mas_detach)
 
 
 
 
 
2278{
2279	mas_set_range(mas_detach, vma->vm_start, vma->vm_end - 1);
2280	if (mas_store_gfp(mas_detach, vma, GFP_KERNEL))
2281		return -ENOMEM;
2282
2283	if (vma->vm_flags & VM_LOCKED)
2284		vma->vm_mm->locked_vm -= vma_pages(vma);
2285
2286	return 0;
2287}
 
2288
2289/*
2290 * do_mas_align_munmap() - munmap the aligned region from @start to @end.
2291 * @mas: The maple_state, ideally set up to alter the correct tree location.
2292 * @vma: The starting vm_area_struct
2293 * @mm: The mm_struct
2294 * @start: The aligned start address to munmap.
2295 * @end: The aligned end address to munmap.
2296 * @uf: The userfaultfd list_head
2297 * @downgrade: Set to true to attempt a write downgrade of the mmap_lock
2298 *
2299 * If @downgrade is true, check return code for potential release of the lock.
2300 */
2301static int
2302do_mas_align_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2303		    struct mm_struct *mm, unsigned long start,
2304		    unsigned long end, struct list_head *uf, bool downgrade)
2305{
2306	struct vm_area_struct *prev, *next = NULL;
2307	struct maple_tree mt_detach;
2308	int count = 0;
2309	int error = -ENOMEM;
2310	MA_STATE(mas_detach, &mt_detach, 0, 0);
2311	mt_init_flags(&mt_detach, MT_FLAGS_LOCK_EXTERN);
2312	mt_set_external_lock(&mt_detach, &mm->mmap_lock);
2313
2314	if (mas_preallocate(mas, vma, GFP_KERNEL))
2315		return -ENOMEM;
 
 
2316
2317	mas->last = end - 1;
2318	/*
2319	 * If we need to split any vma, do it now to save pain later.
2320	 *
2321	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2322	 * unmapped vm_area_struct will remain in use: so lower split_vma
2323	 * places tmp vma above, and higher split_vma places tmp vma below.
2324	 */
2325
2326	/* Does it split the first one? */
2327	if (start > vma->vm_start) {
 
2328
2329		/*
2330		 * Make sure that map_count on return from munmap() will
2331		 * not exceed its limit; but let map_count go just above
2332		 * its limit temporarily, to help free resources as expected.
2333		 */
2334		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2335			goto map_count_exceeded;
2336
2337		/*
2338		 * mas_pause() is not needed since mas->index needs to be set
2339		 * differently than vma->vm_end anyways.
2340		 */
2341		error = __split_vma(mm, vma, start, 0);
2342		if (error)
2343			goto start_split_failed;
2344
2345		mas_set(mas, start);
2346		vma = mas_walk(mas);
2347	}
2348
2349	prev = mas_prev(mas, 0);
2350	if (unlikely((!prev)))
2351		mas_set(mas, start);
2352
2353	/*
2354	 * Detach a range of VMAs from the mm. Using next as a temp variable as
2355	 * it is always overwritten.
2356	 */
2357	mas_for_each(mas, next, end - 1) {
2358		/* Does it split the end? */
2359		if (next->vm_end > end) {
2360			struct vm_area_struct *split;
2361
2362			error = __split_vma(mm, next, end, 1);
2363			if (error)
2364				goto end_split_failed;
2365
2366			mas_set(mas, end);
2367			split = mas_prev(mas, 0);
2368			error = munmap_sidetree(split, &mas_detach);
2369			if (error)
2370				goto munmap_sidetree_failed;
2371
2372			count++;
2373			if (vma == next)
2374				vma = split;
2375			break;
2376		}
2377		error = munmap_sidetree(next, &mas_detach);
2378		if (error)
2379			goto munmap_sidetree_failed;
2380
2381		count++;
2382#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2383		BUG_ON(next->vm_start < start);
2384		BUG_ON(next->vm_start > end);
2385#endif
2386	}
2387
2388	if (!next)
2389		next = mas_next(mas, ULONG_MAX);
2390
2391	if (unlikely(uf)) {
2392		/*
2393		 * If userfaultfd_unmap_prep returns an error the vmas
2394		 * will remain split, but userland will get a
2395		 * highly unexpected error anyway. This is no
2396		 * different than the case where the first of the two
2397		 * __split_vma fails, but we don't undo the first
2398		 * split, despite we could. This is unlikely enough
2399		 * failure that it's not worth optimizing it for.
2400		 */
2401		error = userfaultfd_unmap_prep(mm, start, end, uf);
2402
2403		if (error)
2404			goto userfaultfd_error;
2405	}
2406
2407	/* Point of no return */
2408	mas_set_range(mas, start, end - 1);
2409#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2410	/* Make sure no VMAs are about to be lost. */
2411	{
2412		MA_STATE(test, &mt_detach, start, end - 1);
2413		struct vm_area_struct *vma_mas, *vma_test;
2414		int test_count = 0;
2415
2416		rcu_read_lock();
2417		vma_test = mas_find(&test, end - 1);
2418		mas_for_each(mas, vma_mas, end - 1) {
2419			BUG_ON(vma_mas != vma_test);
2420			test_count++;
2421			vma_test = mas_next(&test, end - 1);
2422		}
2423		rcu_read_unlock();
2424		BUG_ON(count != test_count);
2425		mas_set_range(mas, start, end - 1);
2426	}
2427#endif
2428	mas_store_prealloc(mas, NULL);
2429	mm->map_count -= count;
2430	/*
2431	 * Do not downgrade mmap_lock if we are next to VM_GROWSDOWN or
2432	 * VM_GROWSUP VMA. Such VMAs can change their size under
2433	 * down_read(mmap_lock) and collide with the VMA we are about to unmap.
2434	 */
2435	if (downgrade) {
2436		if (next && (next->vm_flags & VM_GROWSDOWN))
2437			downgrade = false;
2438		else if (prev && (prev->vm_flags & VM_GROWSUP))
2439			downgrade = false;
2440		else
2441			mmap_write_downgrade(mm);
2442	}
2443
2444	unmap_region(mm, &mt_detach, vma, prev, next, start, end);
2445	/* Statistics and freeing VMAs */
2446	mas_set(&mas_detach, start);
2447	remove_mt(mm, &mas_detach);
2448	__mt_destroy(&mt_detach);
2449
2450
2451	validate_mm(mm);
2452	return downgrade ? 1 : 0;
2453
2454userfaultfd_error:
2455munmap_sidetree_failed:
2456end_split_failed:
2457	__mt_destroy(&mt_detach);
2458start_split_failed:
2459map_count_exceeded:
2460	mas_destroy(mas);
2461	return error;
2462}
2463
2464/*
2465 * do_mas_munmap() - munmap a given range.
2466 * @mas: The maple state
2467 * @mm: The mm_struct
2468 * @start: The start address to munmap
2469 * @len: The length of the range to munmap
2470 * @uf: The userfaultfd list_head
2471 * @downgrade: set to true if the user wants to attempt to write_downgrade the
2472 * mmap_lock
2473 *
2474 * This function takes a @mas that is either pointing to the previous VMA or set
2475 * to MA_START and sets it up to remove the mapping(s).  The @len will be
2476 * aligned and any arch_unmap work will be preformed.
2477 *
2478 * Returns: -EINVAL on failure, 1 on success and unlock, 0 otherwise.
2479 */
2480int do_mas_munmap(struct ma_state *mas, struct mm_struct *mm,
2481		  unsigned long start, size_t len, struct list_head *uf,
2482		  bool downgrade)
2483{
2484	unsigned long end;
2485	struct vm_area_struct *vma;
2486
2487	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2488		return -EINVAL;
2489
2490	end = start + PAGE_ALIGN(len);
2491	if (end == start)
2492		return -EINVAL;
2493
2494	 /* arch_unmap() might do unmaps itself.  */
2495	arch_unmap(mm, start, end);
2496
2497	/* Find the first overlapping VMA */
2498	vma = mas_find(mas, end - 1);
2499	if (!vma)
2500		return 0;
2501
2502	return do_mas_align_munmap(mas, vma, mm, start, end, uf, downgrade);
2503}
2504
2505/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2506 * @mm: The mm_struct
2507 * @start: The start address to munmap
2508 * @len: The length to be munmapped.
2509 * @uf: The userfaultfd list_head
2510 */
2511int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2512	      struct list_head *uf)
2513{
2514	MA_STATE(mas, &mm->mm_mt, start, start);
2515
2516	return do_mas_munmap(&mas, mm, start, len, uf, false);
2517}
2518
2519unsigned long mmap_region(struct file *file, unsigned long addr,
2520		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2521		struct list_head *uf)
2522{
2523	struct mm_struct *mm = current->mm;
2524	struct vm_area_struct *vma = NULL;
2525	struct vm_area_struct *next, *prev, *merge;
2526	pgoff_t pglen = len >> PAGE_SHIFT;
2527	unsigned long charged = 0;
2528	unsigned long end = addr + len;
2529	unsigned long merge_start = addr, merge_end = end;
2530	pgoff_t vm_pgoff;
2531	int error;
2532	MA_STATE(mas, &mm->mm_mt, addr, end - 1);
2533
2534	/* Check against address space limit. */
2535	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2536		unsigned long nr_pages;
2537
2538		/*
2539		 * MAP_FIXED may remove pages of mappings that intersects with
2540		 * requested mapping. Account for the pages it would unmap.
2541		 */
2542		nr_pages = count_vma_pages_range(mm, addr, end);
2543
2544		if (!may_expand_vm(mm, vm_flags,
2545					(len >> PAGE_SHIFT) - nr_pages))
2546			return -ENOMEM;
2547	}
2548
2549	/* Unmap any existing mapping in the area */
2550	if (do_mas_munmap(&mas, mm, addr, len, uf, false))
2551		return -ENOMEM;
2552
2553	/*
2554	 * Private writable mapping: check memory availability
2555	 */
2556	if (accountable_mapping(file, vm_flags)) {
2557		charged = len >> PAGE_SHIFT;
2558		if (security_vm_enough_memory_mm(mm, charged))
2559			return -ENOMEM;
2560		vm_flags |= VM_ACCOUNT;
2561	}
2562
2563	next = mas_next(&mas, ULONG_MAX);
2564	prev = mas_prev(&mas, 0);
2565	if (vm_flags & VM_SPECIAL)
2566		goto cannot_expand;
2567
2568	/* Attempt to expand an old mapping */
2569	/* Check next */
2570	if (next && next->vm_start == end && !vma_policy(next) &&
2571	    can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2572				 NULL_VM_UFFD_CTX, NULL)) {
2573		merge_end = next->vm_end;
2574		vma = next;
2575		vm_pgoff = next->vm_pgoff - pglen;
2576	}
2577
2578	/* Check prev */
2579	if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2580	    (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2581				       pgoff, vma->vm_userfaultfd_ctx, NULL) :
2582		   can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2583				       NULL_VM_UFFD_CTX, NULL))) {
2584		merge_start = prev->vm_start;
2585		vma = prev;
2586		vm_pgoff = prev->vm_pgoff;
2587	}
2588
2589
2590	/* Actually expand, if possible */
2591	if (vma &&
2592	    !vma_expand(&mas, vma, merge_start, merge_end, vm_pgoff, next)) {
2593		khugepaged_enter_vma(vma, vm_flags);
2594		goto expanded;
2595	}
2596
2597	mas.index = addr;
2598	mas.last = end - 1;
2599cannot_expand:
2600	/*
2601	 * Determine the object being mapped and call the appropriate
2602	 * specific mapper. the address has already been validated, but
2603	 * not unmapped, but the maps are removed from the list.
2604	 */
2605	vma = vm_area_alloc(mm);
2606	if (!vma) {
2607		error = -ENOMEM;
2608		goto unacct_error;
2609	}
2610
2611	vma->vm_start = addr;
2612	vma->vm_end = end;
2613	vma->vm_flags = vm_flags;
2614	vma->vm_page_prot = vm_get_page_prot(vm_flags);
2615	vma->vm_pgoff = pgoff;
2616
2617	if (file) {
2618		if (vm_flags & VM_SHARED) {
2619			error = mapping_map_writable(file->f_mapping);
2620			if (error)
2621				goto free_vma;
2622		}
2623
2624		vma->vm_file = get_file(file);
2625		error = call_mmap(file, vma);
2626		if (error)
2627			goto unmap_and_free_vma;
2628
2629		/*
2630		 * Expansion is handled above, merging is handled below.
2631		 * Drivers should not alter the address of the VMA.
2632		 */
2633		if (WARN_ON((addr != vma->vm_start))) {
2634			error = -EINVAL;
2635			goto close_and_free_vma;
2636		}
2637		mas_reset(&mas);
2638
2639		/*
2640		 * If vm_flags changed after call_mmap(), we should try merge
2641		 * vma again as we may succeed this time.
2642		 */
2643		if (unlikely(vm_flags != vma->vm_flags && prev)) {
2644			merge = vma_merge(mm, prev, vma->vm_start, vma->vm_end, vma->vm_flags,
2645				NULL, vma->vm_file, vma->vm_pgoff, NULL, NULL_VM_UFFD_CTX, NULL);
2646			if (merge) {
2647				/*
2648				 * ->mmap() can change vma->vm_file and fput
2649				 * the original file. So fput the vma->vm_file
2650				 * here or we would add an extra fput for file
2651				 * and cause general protection fault
2652				 * ultimately.
2653				 */
2654				fput(vma->vm_file);
2655				vm_area_free(vma);
2656				vma = merge;
2657				/* Update vm_flags to pick up the change. */
2658				vm_flags = vma->vm_flags;
2659				goto unmap_writable;
2660			}
 
2661		}
2662
2663		vm_flags = vma->vm_flags;
2664	} else if (vm_flags & VM_SHARED) {
2665		error = shmem_zero_setup(vma);
2666		if (error)
2667			goto free_vma;
2668	} else {
2669		vma_set_anonymous(vma);
2670	}
2671
2672	/* Allow architectures to sanity-check the vm_flags */
2673	if (!arch_validate_flags(vma->vm_flags)) {
2674		error = -EINVAL;
2675		if (file)
2676			goto close_and_free_vma;
2677		else if (vma->vm_file)
2678			goto unmap_and_free_vma;
2679		else
2680			goto free_vma;
2681	}
2682
2683	if (mas_preallocate(&mas, vma, GFP_KERNEL)) {
2684		error = -ENOMEM;
2685		if (file)
2686			goto close_and_free_vma;
2687		else if (vma->vm_file)
2688			goto unmap_and_free_vma;
2689		else
2690			goto free_vma;
2691	}
2692
2693	if (vma->vm_file)
2694		i_mmap_lock_write(vma->vm_file->f_mapping);
2695
2696	vma_mas_store(vma, &mas);
2697	mm->map_count++;
2698	if (vma->vm_file) {
2699		if (vma->vm_flags & VM_SHARED)
2700			mapping_allow_writable(vma->vm_file->f_mapping);
2701
2702		flush_dcache_mmap_lock(vma->vm_file->f_mapping);
2703		vma_interval_tree_insert(vma, &vma->vm_file->f_mapping->i_mmap);
2704		flush_dcache_mmap_unlock(vma->vm_file->f_mapping);
2705		i_mmap_unlock_write(vma->vm_file->f_mapping);
2706	}
2707
2708	/*
2709	 * vma_merge() calls khugepaged_enter_vma() either, the below
2710	 * call covers the non-merge case.
2711	 */
2712	khugepaged_enter_vma(vma, vma->vm_flags);
 
2713
2714	/* Once vma denies write, undo our temporary denial count */
2715unmap_writable:
2716	if (file && vm_flags & VM_SHARED)
2717		mapping_unmap_writable(file->f_mapping);
2718	file = vma->vm_file;
2719expanded:
2720	perf_event_mmap(vma);
2721
2722	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2723	if (vm_flags & VM_LOCKED) {
2724		if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2725					is_vm_hugetlb_page(vma) ||
2726					vma == get_gate_vma(current->mm))
2727			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
2728		else
2729			mm->locked_vm += (len >> PAGE_SHIFT);
2730	}
2731
2732	if (file)
2733		uprobe_mmap(vma);
2734
2735	/*
2736	 * New (or expanded) vma always get soft dirty status.
2737	 * Otherwise user-space soft-dirty page tracker won't
2738	 * be able to distinguish situation when vma area unmapped,
2739	 * then new mapped in-place (which must be aimed as
2740	 * a completely new data area).
2741	 */
2742	vma->vm_flags |= VM_SOFTDIRTY;
2743
2744	vma_set_page_prot(vma);
2745
2746	validate_mm(mm);
2747	return addr;
2748
2749close_and_free_vma:
2750	if (vma->vm_ops && vma->vm_ops->close)
2751		vma->vm_ops->close(vma);
2752unmap_and_free_vma:
2753	fput(vma->vm_file);
2754	vma->vm_file = NULL;
2755
2756	/* Undo any partial mapping done by a device driver. */
2757	unmap_region(mm, mas.tree, vma, prev, next, vma->vm_start, vma->vm_end);
2758	if (file && (vm_flags & VM_SHARED))
2759		mapping_unmap_writable(file->f_mapping);
2760free_vma:
2761	vm_area_free(vma);
2762unacct_error:
2763	if (charged)
2764		vm_unacct_memory(charged);
2765	validate_mm(mm);
2766	return error;
2767}
2768
2769static int __vm_munmap(unsigned long start, size_t len, bool downgrade)
2770{
2771	int ret;
2772	struct mm_struct *mm = current->mm;
2773	LIST_HEAD(uf);
2774	MA_STATE(mas, &mm->mm_mt, start, start);
2775
2776	if (mmap_write_lock_killable(mm))
2777		return -EINTR;
2778
2779	ret = do_mas_munmap(&mas, mm, start, len, &uf, downgrade);
2780	/*
2781	 * Returning 1 indicates mmap_lock is downgraded.
2782	 * But 1 is not legal return value of vm_munmap() and munmap(), reset
2783	 * it to 0 before return.
2784	 */
2785	if (ret == 1) {
2786		mmap_read_unlock(mm);
2787		ret = 0;
2788	} else
2789		mmap_write_unlock(mm);
2790
2791	userfaultfd_unmap_complete(mm, &uf);
2792	return ret;
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797	return __vm_munmap(start, len, false);
2798}
2799EXPORT_SYMBOL(vm_munmap);
2800
2801SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2802{
2803	addr = untagged_addr(addr);
2804	return __vm_munmap(addr, len, true);
2805}
2806
2807
2808/*
2809 * Emulation of deprecated remap_file_pages() syscall.
2810 */
2811SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2812		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2813{
2814
2815	struct mm_struct *mm = current->mm;
2816	struct vm_area_struct *vma;
2817	unsigned long populate = 0;
2818	unsigned long ret = -EINVAL;
2819	struct file *file;
2820
2821	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
2822		     current->comm, current->pid);
2823
2824	if (prot)
2825		return ret;
2826	start = start & PAGE_MASK;
2827	size = size & PAGE_MASK;
2828
2829	if (start + size <= start)
2830		return ret;
2831
2832	/* Does pgoff wrap? */
2833	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2834		return ret;
2835
2836	if (mmap_write_lock_killable(mm))
2837		return -EINTR;
2838
2839	vma = vma_lookup(mm, start);
2840
2841	if (!vma || !(vma->vm_flags & VM_SHARED))
2842		goto out;
2843
 
 
 
2844	if (start + size > vma->vm_end) {
2845		VMA_ITERATOR(vmi, mm, vma->vm_end);
2846		struct vm_area_struct *next, *prev = vma;
2847
2848		for_each_vma_range(vmi, next, start + size) {
2849			/* hole between vmas ? */
2850			if (next->vm_start != prev->vm_end)
2851				goto out;
2852
2853			if (next->vm_file != vma->vm_file)
2854				goto out;
2855
2856			if (next->vm_flags != vma->vm_flags)
2857				goto out;
2858
2859			if (start + size <= next->vm_end)
2860				break;
2861
2862			prev = next;
2863		}
2864
2865		if (!next)
2866			goto out;
2867	}
2868
2869	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2870	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2871	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2872
2873	flags &= MAP_NONBLOCK;
2874	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2875	if (vma->vm_flags & VM_LOCKED)
 
2876		flags |= MAP_LOCKED;
2877
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2878	file = get_file(vma->vm_file);
2879	ret = do_mmap(vma->vm_file, start, size,
2880			prot, flags, pgoff, &populate, NULL);
2881	fput(file);
2882out:
2883	mmap_write_unlock(mm);
2884	if (populate)
2885		mm_populate(ret, populate);
2886	if (!IS_ERR_VALUE(ret))
2887		ret = 0;
2888	return ret;
2889}
2890
2891/*
2892 * brk_munmap() - Unmap a parital vma.
2893 * @mas: The maple tree state.
2894 * @vma: The vma to be modified
2895 * @newbrk: the start of the address to unmap
2896 * @oldbrk: The end of the address to unmap
2897 * @uf: The userfaultfd list_head
2898 *
2899 * Returns: 1 on success.
2900 * unmaps a partial VMA mapping.  Does not handle alignment, downgrades lock if
2901 * possible.
2902 */
2903static int do_brk_munmap(struct ma_state *mas, struct vm_area_struct *vma,
2904			 unsigned long newbrk, unsigned long oldbrk,
2905			 struct list_head *uf)
2906{
2907	struct mm_struct *mm = vma->vm_mm;
2908	int ret;
2909
2910	arch_unmap(mm, newbrk, oldbrk);
2911	ret = do_mas_align_munmap(mas, vma, mm, newbrk, oldbrk, uf, true);
2912	validate_mm_mt(mm);
2913	return ret;
2914}
2915
2916/*
2917 * do_brk_flags() - Increase the brk vma if the flags match.
2918 * @mas: The maple tree state.
2919 * @addr: The start address
2920 * @len: The length of the increase
2921 * @vma: The vma,
2922 * @flags: The VMA Flags
2923 *
2924 * Extend the brk VMA from addr to addr + len.  If the VMA is NULL or the flags
2925 * do not match then create a new anonymous VMA.  Eventually we may be able to
2926 * do some brk-specific accounting here.
2927 */
2928static int do_brk_flags(struct ma_state *mas, struct vm_area_struct *vma,
2929		unsigned long addr, unsigned long len, unsigned long flags)
2930{
2931	struct mm_struct *mm = current->mm;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2932
2933	validate_mm_mt(mm);
2934	/*
2935	 * Check against address space limits by the changed size
2936	 * Note: This happens *after* clearing old mappings in some code paths.
2937	 */
2938	flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
 
 
 
 
 
 
 
 
 
 
 
2939	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2940		return -ENOMEM;
2941
2942	if (mm->map_count > sysctl_max_map_count)
2943		return -ENOMEM;
2944
2945	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2946		return -ENOMEM;
2947
 
 
 
 
 
 
2948	/*
2949	 * Expand the existing vma if possible; Note that singular lists do not
2950	 * occur after forking, so the expand will only happen on new VMAs.
2951	 */
2952	if (vma && vma->vm_end == addr && !vma_policy(vma) &&
2953	    can_vma_merge_after(vma, flags, NULL, NULL,
2954				addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
2955		mas_set_range(mas, vma->vm_start, addr + len - 1);
2956		if (mas_preallocate(mas, vma, GFP_KERNEL))
2957			goto unacct_fail;
2958
2959		vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
2960		if (vma->anon_vma) {
2961			anon_vma_lock_write(vma->anon_vma);
2962			anon_vma_interval_tree_pre_update_vma(vma);
2963		}
2964		vma->vm_end = addr + len;
2965		vma->vm_flags |= VM_SOFTDIRTY;
2966		mas_store_prealloc(mas, vma);
2967
2968		if (vma->anon_vma) {
2969			anon_vma_interval_tree_post_update_vma(vma);
2970			anon_vma_unlock_write(vma->anon_vma);
2971		}
2972		khugepaged_enter_vma(vma, flags);
2973		goto out;
2974	}
2975
2976	/* create a vma struct for an anonymous mapping */
2977	vma = vm_area_alloc(mm);
2978	if (!vma)
2979		goto unacct_fail;
2980
2981	vma_set_anonymous(vma);
2982	vma->vm_start = addr;
2983	vma->vm_end = addr + len;
2984	vma->vm_pgoff = addr >> PAGE_SHIFT;
2985	vma->vm_flags = flags;
2986	vma->vm_page_prot = vm_get_page_prot(flags);
2987	mas_set_range(mas, vma->vm_start, addr + len - 1);
2988	if (mas_store_gfp(mas, vma, GFP_KERNEL))
2989		goto mas_store_fail;
2990
2991	mm->map_count++;
2992out:
2993	perf_event_mmap(vma);
2994	mm->total_vm += len >> PAGE_SHIFT;
2995	mm->data_vm += len >> PAGE_SHIFT;
2996	if (flags & VM_LOCKED)
2997		mm->locked_vm += (len >> PAGE_SHIFT);
2998	vma->vm_flags |= VM_SOFTDIRTY;
2999	validate_mm(mm);
3000	return 0;
 
3001
3002mas_store_fail:
3003	vm_area_free(vma);
3004unacct_fail:
3005	vm_unacct_memory(len >> PAGE_SHIFT);
3006	return -ENOMEM;
3007}
3008
3009int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3010{
3011	struct mm_struct *mm = current->mm;
3012	struct vm_area_struct *vma = NULL;
3013	unsigned long len;
3014	int ret;
3015	bool populate;
3016	LIST_HEAD(uf);
3017	MA_STATE(mas, &mm->mm_mt, addr, addr);
3018
3019	len = PAGE_ALIGN(request);
3020	if (len < request)
3021		return -ENOMEM;
3022	if (!len)
3023		return 0;
3024
3025	if (mmap_write_lock_killable(mm))
3026		return -EINTR;
3027
3028	/* Until we need other flags, refuse anything except VM_EXEC. */
3029	if ((flags & (~VM_EXEC)) != 0)
3030		return -EINVAL;
3031
3032	ret = check_brk_limits(addr, len);
3033	if (ret)
3034		goto limits_failed;
3035
3036	ret = do_mas_munmap(&mas, mm, addr, len, &uf, 0);
3037	if (ret)
3038		goto munmap_failed;
3039
3040	vma = mas_prev(&mas, 0);
3041	ret = do_brk_flags(&mas, vma, addr, len, flags);
3042	populate = ((mm->def_flags & VM_LOCKED) != 0);
3043	mmap_write_unlock(mm);
3044	userfaultfd_unmap_complete(mm, &uf);
3045	if (populate && !ret)
3046		mm_populate(addr, len);
3047	return ret;
3048
3049munmap_failed:
3050limits_failed:
3051	mmap_write_unlock(mm);
3052	return ret;
3053}
3054EXPORT_SYMBOL(vm_brk_flags);
3055
3056int vm_brk(unsigned long addr, unsigned long len)
3057{
3058	return vm_brk_flags(addr, len, 0);
3059}
3060EXPORT_SYMBOL(vm_brk);
3061
3062/* Release all mmaps. */
3063void exit_mmap(struct mm_struct *mm)
3064{
3065	struct mmu_gather tlb;
3066	struct vm_area_struct *vma;
3067	unsigned long nr_accounted = 0;
3068	MA_STATE(mas, &mm->mm_mt, 0, 0);
3069	int count = 0;
3070
3071	/* mm's last user has gone, and its about to be pulled down */
3072	mmu_notifier_release(mm);
3073
3074	mmap_read_lock(mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3075	arch_exit_mmap(mm);
3076
3077	vma = mas_find(&mas, ULONG_MAX);
3078	if (!vma) {
3079		/* Can happen if dup_mmap() received an OOM */
3080		mmap_read_unlock(mm);
3081		return;
3082	}
3083
3084	lru_add_drain();
3085	flush_cache_mm(mm);
3086	tlb_gather_mmu_fullmm(&tlb, mm);
3087	/* update_hiwater_rss(mm) here? but nobody should be looking */
3088	/* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3089	unmap_vmas(&tlb, &mm->mm_mt, vma, 0, ULONG_MAX);
3090	mmap_read_unlock(mm);
3091
3092	/*
3093	 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3094	 * because the memory has been already freed.
3095	 */
3096	set_bit(MMF_OOM_SKIP, &mm->flags);
3097	mmap_write_lock(mm);
3098	free_pgtables(&tlb, &mm->mm_mt, vma, FIRST_USER_ADDRESS,
3099		      USER_PGTABLES_CEILING);
3100	tlb_finish_mmu(&tlb);
3101
3102	/*
3103	 * Walk the list again, actually closing and freeing it, with preemption
3104	 * enabled, without holding any MM locks besides the unreachable
3105	 * mmap_write_lock.
3106	 */
3107	do {
3108		if (vma->vm_flags & VM_ACCOUNT)
3109			nr_accounted += vma_pages(vma);
3110		remove_vma(vma);
3111		count++;
3112		cond_resched();
3113	} while ((vma = mas_find(&mas, ULONG_MAX)) != NULL);
3114
3115	BUG_ON(count != mm->map_count);
3116
3117	trace_exit_mmap(mm);
3118	__mt_destroy(&mm->mm_mt);
3119	mmap_write_unlock(mm);
3120	vm_unacct_memory(nr_accounted);
3121}
3122
3123/* Insert vm structure into process list sorted by address
3124 * and into the inode's i_mmap tree.  If vm_file is non-NULL
3125 * then i_mmap_rwsem is taken here.
3126 */
3127int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3128{
3129	unsigned long charged = vma_pages(vma);
3130
3131
3132	if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
 
3133		return -ENOMEM;
3134
3135	if ((vma->vm_flags & VM_ACCOUNT) &&
3136	     security_vm_enough_memory_mm(mm, charged))
3137		return -ENOMEM;
3138
3139	/*
3140	 * The vm_pgoff of a purely anonymous vma should be irrelevant
3141	 * until its first write fault, when page's anon_vma and index
3142	 * are set.  But now set the vm_pgoff it will almost certainly
3143	 * end up with (unless mremap moves it elsewhere before that
3144	 * first wfault), so /proc/pid/maps tells a consistent story.
3145	 *
3146	 * By setting it to reflect the virtual start address of the
3147	 * vma, merges and splits can happen in a seamless way, just
3148	 * using the existing file pgoff checks and manipulations.
3149	 * Similarly in do_mmap and in do_brk_flags.
3150	 */
3151	if (vma_is_anonymous(vma)) {
3152		BUG_ON(vma->anon_vma);
3153		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3154	}
3155
3156	if (vma_link(mm, vma)) {
3157		vm_unacct_memory(charged);
3158		return -ENOMEM;
3159	}
3160
3161	return 0;
3162}
3163
3164/*
3165 * Copy the vma structure to a new location in the same mm,
3166 * prior to moving page table entries, to effect an mremap move.
3167 */
3168struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3169	unsigned long addr, unsigned long len, pgoff_t pgoff,
3170	bool *need_rmap_locks)
3171{
3172	struct vm_area_struct *vma = *vmap;
3173	unsigned long vma_start = vma->vm_start;
3174	struct mm_struct *mm = vma->vm_mm;
3175	struct vm_area_struct *new_vma, *prev;
 
3176	bool faulted_in_anon_vma = true;
3177
3178	validate_mm_mt(mm);
3179	/*
3180	 * If anonymous vma has not yet been faulted, update new pgoff
3181	 * to match new location, to increase its chance of merging.
3182	 */
3183	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3184		pgoff = addr >> PAGE_SHIFT;
3185		faulted_in_anon_vma = false;
3186	}
3187
3188	new_vma = find_vma_prev(mm, addr, &prev);
3189	if (new_vma && new_vma->vm_start < addr + len)
3190		return NULL;	/* should never get here */
3191
3192	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3193			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3194			    vma->vm_userfaultfd_ctx, anon_vma_name(vma));
3195	if (new_vma) {
3196		/*
3197		 * Source vma may have been merged into new_vma
3198		 */
3199		if (unlikely(vma_start >= new_vma->vm_start &&
3200			     vma_start < new_vma->vm_end)) {
3201			/*
3202			 * The only way we can get a vma_merge with
3203			 * self during an mremap is if the vma hasn't
3204			 * been faulted in yet and we were allowed to
3205			 * reset the dst vma->vm_pgoff to the
3206			 * destination address of the mremap to allow
3207			 * the merge to happen. mremap must change the
3208			 * vm_pgoff linearity between src and dst vmas
3209			 * (in turn preventing a vma_merge) to be
3210			 * safe. It is only safe to keep the vm_pgoff
3211			 * linear if there are no pages mapped yet.
3212			 */
3213			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3214			*vmap = vma = new_vma;
3215		}
3216		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3217	} else {
3218		new_vma = vm_area_dup(vma);
3219		if (!new_vma)
3220			goto out;
 
3221		new_vma->vm_start = addr;
3222		new_vma->vm_end = addr + len;
3223		new_vma->vm_pgoff = pgoff;
3224		if (vma_dup_policy(vma, new_vma))
3225			goto out_free_vma;
 
3226		if (anon_vma_clone(new_vma, vma))
3227			goto out_free_mempol;
3228		if (new_vma->vm_file)
3229			get_file(new_vma->vm_file);
3230		if (new_vma->vm_ops && new_vma->vm_ops->open)
3231			new_vma->vm_ops->open(new_vma);
3232		if (vma_link(mm, new_vma))
3233			goto out_vma_link;
3234		*need_rmap_locks = false;
3235	}
3236	validate_mm_mt(mm);
3237	return new_vma;
3238
3239out_vma_link:
3240	if (new_vma->vm_ops && new_vma->vm_ops->close)
3241		new_vma->vm_ops->close(new_vma);
3242
3243	if (new_vma->vm_file)
3244		fput(new_vma->vm_file);
3245
3246	unlink_anon_vmas(new_vma);
3247out_free_mempol:
3248	mpol_put(vma_policy(new_vma));
3249out_free_vma:
3250	vm_area_free(new_vma);
3251out:
3252	validate_mm_mt(mm);
3253	return NULL;
3254}
3255
3256/*
3257 * Return true if the calling process may expand its vm space by the passed
3258 * number of pages
3259 */
3260bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3261{
3262	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3263		return false;
3264
3265	if (is_data_mapping(flags) &&
3266	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3267		/* Workaround for Valgrind */
3268		if (rlimit(RLIMIT_DATA) == 0 &&
3269		    mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3270			return true;
3271
3272		pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3273			     current->comm, current->pid,
3274			     (mm->data_vm + npages) << PAGE_SHIFT,
3275			     rlimit(RLIMIT_DATA),
3276			     ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3277
3278		if (!ignore_rlimit_data)
3279			return false;
3280	}
3281
3282	return true;
3283}
3284
3285void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3286{
3287	WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3288
3289	if (is_exec_mapping(flags))
3290		mm->exec_vm += npages;
3291	else if (is_stack_mapping(flags))
3292		mm->stack_vm += npages;
3293	else if (is_data_mapping(flags))
3294		mm->data_vm += npages;
3295}
3296
3297static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3298
3299/*
3300 * Having a close hook prevents vma merging regardless of flags.
3301 */
3302static void special_mapping_close(struct vm_area_struct *vma)
3303{
3304}
3305
3306static const char *special_mapping_name(struct vm_area_struct *vma)
3307{
3308	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3309}
3310
3311static int special_mapping_mremap(struct vm_area_struct *new_vma)
3312{
3313	struct vm_special_mapping *sm = new_vma->vm_private_data;
3314
3315	if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3316		return -EFAULT;
3317
3318	if (sm->mremap)
3319		return sm->mremap(sm, new_vma);
3320
3321	return 0;
3322}
3323
3324static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3325{
3326	/*
3327	 * Forbid splitting special mappings - kernel has expectations over
3328	 * the number of pages in mapping. Together with VM_DONTEXPAND
3329	 * the size of vma should stay the same over the special mapping's
3330	 * lifetime.
3331	 */
3332	return -EINVAL;
3333}
3334
3335static const struct vm_operations_struct special_mapping_vmops = {
3336	.close = special_mapping_close,
3337	.fault = special_mapping_fault,
3338	.mremap = special_mapping_mremap,
3339	.name = special_mapping_name,
3340	/* vDSO code relies that VVAR can't be accessed remotely */
3341	.access = NULL,
3342	.may_split = special_mapping_split,
3343};
3344
3345static const struct vm_operations_struct legacy_special_mapping_vmops = {
3346	.close = special_mapping_close,
3347	.fault = special_mapping_fault,
3348};
3349
3350static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3351{
3352	struct vm_area_struct *vma = vmf->vma;
3353	pgoff_t pgoff;
3354	struct page **pages;
3355
3356	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3357		pages = vma->vm_private_data;
3358	} else {
3359		struct vm_special_mapping *sm = vma->vm_private_data;
3360
3361		if (sm->fault)
3362			return sm->fault(sm, vmf->vma, vmf);
3363
3364		pages = sm->pages;
3365	}
3366
3367	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3368		pgoff--;
3369
3370	if (*pages) {
3371		struct page *page = *pages;
3372		get_page(page);
3373		vmf->page = page;
3374		return 0;
3375	}
3376
3377	return VM_FAULT_SIGBUS;
3378}
3379
3380static struct vm_area_struct *__install_special_mapping(
3381	struct mm_struct *mm,
3382	unsigned long addr, unsigned long len,
3383	unsigned long vm_flags, void *priv,
3384	const struct vm_operations_struct *ops)
3385{
3386	int ret;
3387	struct vm_area_struct *vma;
3388
3389	validate_mm_mt(mm);
3390	vma = vm_area_alloc(mm);
3391	if (unlikely(vma == NULL))
3392		return ERR_PTR(-ENOMEM);
3393
 
 
3394	vma->vm_start = addr;
3395	vma->vm_end = addr + len;
3396
3397	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
3398	vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
3399	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3400
3401	vma->vm_ops = ops;
3402	vma->vm_private_data = priv;
3403
3404	ret = insert_vm_struct(mm, vma);
3405	if (ret)
3406		goto out;
3407
3408	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3409
3410	perf_event_mmap(vma);
3411
3412	validate_mm_mt(mm);
3413	return vma;
3414
3415out:
3416	vm_area_free(vma);
3417	validate_mm_mt(mm);
3418	return ERR_PTR(ret);
3419}
3420
3421bool vma_is_special_mapping(const struct vm_area_struct *vma,
3422	const struct vm_special_mapping *sm)
3423{
3424	return vma->vm_private_data == sm &&
3425		(vma->vm_ops == &special_mapping_vmops ||
3426		 vma->vm_ops == &legacy_special_mapping_vmops);
3427}
3428
3429/*
3430 * Called with mm->mmap_lock held for writing.
3431 * Insert a new vma covering the given region, with the given flags.
3432 * Its pages are supplied by the given array of struct page *.
3433 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3434 * The region past the last page supplied will always produce SIGBUS.
3435 * The array pointer and the pages it points to are assumed to stay alive
3436 * for as long as this mapping might exist.
3437 */
3438struct vm_area_struct *_install_special_mapping(
3439	struct mm_struct *mm,
3440	unsigned long addr, unsigned long len,
3441	unsigned long vm_flags, const struct vm_special_mapping *spec)
3442{
3443	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3444					&special_mapping_vmops);
3445}
3446
3447int install_special_mapping(struct mm_struct *mm,
3448			    unsigned long addr, unsigned long len,
3449			    unsigned long vm_flags, struct page **pages)
3450{
3451	struct vm_area_struct *vma = __install_special_mapping(
3452		mm, addr, len, vm_flags, (void *)pages,
3453		&legacy_special_mapping_vmops);
3454
3455	return PTR_ERR_OR_ZERO(vma);
3456}
3457
3458static DEFINE_MUTEX(mm_all_locks_mutex);
3459
3460static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3461{
3462	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3463		/*
3464		 * The LSB of head.next can't change from under us
3465		 * because we hold the mm_all_locks_mutex.
3466		 */
3467		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3468		/*
3469		 * We can safely modify head.next after taking the
3470		 * anon_vma->root->rwsem. If some other vma in this mm shares
3471		 * the same anon_vma we won't take it again.
3472		 *
3473		 * No need of atomic instructions here, head.next
3474		 * can't change from under us thanks to the
3475		 * anon_vma->root->rwsem.
3476		 */
3477		if (__test_and_set_bit(0, (unsigned long *)
3478				       &anon_vma->root->rb_root.rb_root.rb_node))
3479			BUG();
3480	}
3481}
3482
3483static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3484{
3485	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3486		/*
3487		 * AS_MM_ALL_LOCKS can't change from under us because
3488		 * we hold the mm_all_locks_mutex.
3489		 *
3490		 * Operations on ->flags have to be atomic because
3491		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3492		 * mm_all_locks_mutex, there may be other cpus
3493		 * changing other bitflags in parallel to us.
3494		 */
3495		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3496			BUG();
3497		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3498	}
3499}
3500
3501/*
3502 * This operation locks against the VM for all pte/vma/mm related
3503 * operations that could ever happen on a certain mm. This includes
3504 * vmtruncate, try_to_unmap, and all page faults.
3505 *
3506 * The caller must take the mmap_lock in write mode before calling
3507 * mm_take_all_locks(). The caller isn't allowed to release the
3508 * mmap_lock until mm_drop_all_locks() returns.
3509 *
3510 * mmap_lock in write mode is required in order to block all operations
3511 * that could modify pagetables and free pages without need of
3512 * altering the vma layout. It's also needed in write mode to avoid new
3513 * anon_vmas to be associated with existing vmas.
3514 *
3515 * A single task can't take more than one mm_take_all_locks() in a row
3516 * or it would deadlock.
3517 *
3518 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3519 * mapping->flags avoid to take the same lock twice, if more than one
3520 * vma in this mm is backed by the same anon_vma or address_space.
3521 *
3522 * We take locks in following order, accordingly to comment at beginning
3523 * of mm/rmap.c:
3524 *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3525 *     hugetlb mapping);
3526 *   - all i_mmap_rwsem locks;
3527 *   - all anon_vma->rwseml
3528 *
3529 * We can take all locks within these types randomly because the VM code
3530 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3531 * mm_all_locks_mutex.
3532 *
3533 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3534 * that may have to take thousand of locks.
3535 *
3536 * mm_take_all_locks() can fail if it's interrupted by signals.
3537 */
3538int mm_take_all_locks(struct mm_struct *mm)
3539{
3540	struct vm_area_struct *vma;
3541	struct anon_vma_chain *avc;
3542	MA_STATE(mas, &mm->mm_mt, 0, 0);
3543
3544	mmap_assert_write_locked(mm);
3545
3546	mutex_lock(&mm_all_locks_mutex);
3547
3548	mas_for_each(&mas, vma, ULONG_MAX) {
3549		if (signal_pending(current))
3550			goto out_unlock;
3551		if (vma->vm_file && vma->vm_file->f_mapping &&
3552				is_vm_hugetlb_page(vma))
3553			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3554	}
3555
3556	mas_set(&mas, 0);
3557	mas_for_each(&mas, vma, ULONG_MAX) {
3558		if (signal_pending(current))
3559			goto out_unlock;
3560		if (vma->vm_file && vma->vm_file->f_mapping &&
3561				!is_vm_hugetlb_page(vma))
3562			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3563	}
3564
3565	mas_set(&mas, 0);
3566	mas_for_each(&mas, vma, ULONG_MAX) {
3567		if (signal_pending(current))
3568			goto out_unlock;
3569		if (vma->anon_vma)
3570			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3571				vm_lock_anon_vma(mm, avc->anon_vma);
3572	}
3573
3574	return 0;
3575
3576out_unlock:
3577	mm_drop_all_locks(mm);
3578	return -EINTR;
3579}
3580
3581static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3582{
3583	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3584		/*
3585		 * The LSB of head.next can't change to 0 from under
3586		 * us because we hold the mm_all_locks_mutex.
3587		 *
3588		 * We must however clear the bitflag before unlocking
3589		 * the vma so the users using the anon_vma->rb_root will
3590		 * never see our bitflag.
3591		 *
3592		 * No need of atomic instructions here, head.next
3593		 * can't change from under us until we release the
3594		 * anon_vma->root->rwsem.
3595		 */
3596		if (!__test_and_clear_bit(0, (unsigned long *)
3597					  &anon_vma->root->rb_root.rb_root.rb_node))
3598			BUG();
3599		anon_vma_unlock_write(anon_vma);
3600	}
3601}
3602
3603static void vm_unlock_mapping(struct address_space *mapping)
3604{
3605	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3606		/*
3607		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3608		 * because we hold the mm_all_locks_mutex.
3609		 */
3610		i_mmap_unlock_write(mapping);
3611		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3612					&mapping->flags))
3613			BUG();
3614	}
3615}
3616
3617/*
3618 * The mmap_lock cannot be released by the caller until
3619 * mm_drop_all_locks() returns.
3620 */
3621void mm_drop_all_locks(struct mm_struct *mm)
3622{
3623	struct vm_area_struct *vma;
3624	struct anon_vma_chain *avc;
3625	MA_STATE(mas, &mm->mm_mt, 0, 0);
3626
3627	mmap_assert_write_locked(mm);
3628	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3629
3630	mas_for_each(&mas, vma, ULONG_MAX) {
3631		if (vma->anon_vma)
3632			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3633				vm_unlock_anon_vma(avc->anon_vma);
3634		if (vma->vm_file && vma->vm_file->f_mapping)
3635			vm_unlock_mapping(vma->vm_file->f_mapping);
3636	}
3637
3638	mutex_unlock(&mm_all_locks_mutex);
3639}
3640
3641/*
3642 * initialise the percpu counter for VM
3643 */
3644void __init mmap_init(void)
3645{
3646	int ret;
3647
3648	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3649	VM_BUG_ON(ret);
3650}
3651
3652/*
3653 * Initialise sysctl_user_reserve_kbytes.
3654 *
3655 * This is intended to prevent a user from starting a single memory hogging
3656 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3657 * mode.
3658 *
3659 * The default value is min(3% of free memory, 128MB)
3660 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3661 */
3662static int init_user_reserve(void)
3663{
3664	unsigned long free_kbytes;
3665
3666	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3667
3668	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3669	return 0;
3670}
3671subsys_initcall(init_user_reserve);
3672
3673/*
3674 * Initialise sysctl_admin_reserve_kbytes.
3675 *
3676 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3677 * to log in and kill a memory hogging process.
3678 *
3679 * Systems with more than 256MB will reserve 8MB, enough to recover
3680 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3681 * only reserve 3% of free pages by default.
3682 */
3683static int init_admin_reserve(void)
3684{
3685	unsigned long free_kbytes;
3686
3687	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3688
3689	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3690	return 0;
3691}
3692subsys_initcall(init_admin_reserve);
3693
3694/*
3695 * Reinititalise user and admin reserves if memory is added or removed.
3696 *
3697 * The default user reserve max is 128MB, and the default max for the
3698 * admin reserve is 8MB. These are usually, but not always, enough to
3699 * enable recovery from a memory hogging process using login/sshd, a shell,
3700 * and tools like top. It may make sense to increase or even disable the
3701 * reserve depending on the existence of swap or variations in the recovery
3702 * tools. So, the admin may have changed them.
3703 *
3704 * If memory is added and the reserves have been eliminated or increased above
3705 * the default max, then we'll trust the admin.
3706 *
3707 * If memory is removed and there isn't enough free memory, then we
3708 * need to reset the reserves.
3709 *
3710 * Otherwise keep the reserve set by the admin.
3711 */
3712static int reserve_mem_notifier(struct notifier_block *nb,
3713			     unsigned long action, void *data)
3714{
3715	unsigned long tmp, free_kbytes;
3716
3717	switch (action) {
3718	case MEM_ONLINE:
3719		/* Default max is 128MB. Leave alone if modified by operator. */
3720		tmp = sysctl_user_reserve_kbytes;
3721		if (0 < tmp && tmp < (1UL << 17))
3722			init_user_reserve();
3723
3724		/* Default max is 8MB.  Leave alone if modified by operator. */
3725		tmp = sysctl_admin_reserve_kbytes;
3726		if (0 < tmp && tmp < (1UL << 13))
3727			init_admin_reserve();
3728
3729		break;
3730	case MEM_OFFLINE:
3731		free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3732
3733		if (sysctl_user_reserve_kbytes > free_kbytes) {
3734			init_user_reserve();
3735			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3736				sysctl_user_reserve_kbytes);
3737		}
3738
3739		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3740			init_admin_reserve();
3741			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3742				sysctl_admin_reserve_kbytes);
3743		}
3744		break;
3745	default:
3746		break;
3747	}
3748	return NOTIFY_OK;
3749}
3750
 
 
 
 
3751static int __meminit init_reserve_notifier(void)
3752{
3753	if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3754		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3755
3756	return 0;
3757}
3758subsys_initcall(init_reserve_notifier);
v4.17
 
   1/*
   2 * mm/mmap.c
   3 *
   4 * Written by obz.
   5 *
   6 * Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   7 */
   8
   9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  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/shmem_fs.h>
  29#include <linux/profile.h>
  30#include <linux/export.h>
  31#include <linux/mount.h>
  32#include <linux/mempolicy.h>
  33#include <linux/rmap.h>
  34#include <linux/mmu_notifier.h>
  35#include <linux/mmdebug.h>
  36#include <linux/perf_event.h>
  37#include <linux/audit.h>
  38#include <linux/khugepaged.h>
  39#include <linux/uprobes.h>
  40#include <linux/rbtree_augmented.h>
  41#include <linux/notifier.h>
  42#include <linux/memory.h>
  43#include <linux/printk.h>
  44#include <linux/userfaultfd_k.h>
  45#include <linux/moduleparam.h>
  46#include <linux/pkeys.h>
  47#include <linux/oom.h>
 
  48
  49#include <linux/uaccess.h>
  50#include <asm/cacheflush.h>
  51#include <asm/tlb.h>
  52#include <asm/mmu_context.h>
  53
 
 
 
  54#include "internal.h"
  55
  56#ifndef arch_mmap_check
  57#define arch_mmap_check(addr, len, flags)	(0)
  58#endif
  59
  60#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  61const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  62const int mmap_rnd_bits_max = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  63int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  64#endif
  65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  66const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  67const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  68int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  69#endif
  70
  71static bool ignore_rlimit_data;
  72core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  73
  74static void unmap_region(struct mm_struct *mm,
  75		struct vm_area_struct *vma, struct vm_area_struct *prev,
  76		unsigned long start, unsigned long end);
  77
  78/* description of effects of mapping type and prot in current implementation.
  79 * this is due to the limited x86 page protection hardware.  The expected
  80 * behavior is in parens:
  81 *
  82 * map_type	prot
  83 *		PROT_NONE	PROT_READ	PROT_WRITE	PROT_EXEC
  84 * MAP_SHARED	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  85 *		w: (no) no	w: (no) no	w: (yes) yes	w: (no) no
  86 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  87 *
  88 * MAP_PRIVATE	r: (no) no	r: (yes) yes	r: (no) yes	r: (no) yes
  89 *		w: (no) no	w: (no) no	w: (copy) copy	w: (no) no
  90 *		x: (no) no	x: (no) yes	x: (no) yes	x: (yes) yes
  91 *
  92 * On arm64, PROT_EXEC has the following behaviour for both MAP_SHARED and
  93 * MAP_PRIVATE:
  94 *								r: (no) no
  95 *								w: (no) no
  96 *								x: (yes) yes
  97 */
  98pgprot_t protection_map[16] __ro_after_init = {
  99	__P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
 100	__S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
 101};
 102
 103#ifndef CONFIG_ARCH_HAS_FILTER_PGPROT
 104static inline pgprot_t arch_filter_pgprot(pgprot_t prot)
 105{
 106	return prot;
 107}
 108#endif
 109
 110pgprot_t vm_get_page_prot(unsigned long vm_flags)
 111{
 112	pgprot_t ret = __pgprot(pgprot_val(protection_map[vm_flags &
 113				(VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
 114			pgprot_val(arch_vm_get_page_prot(vm_flags)));
 115
 116	return arch_filter_pgprot(ret);
 117}
 118EXPORT_SYMBOL(vm_get_page_prot);
 119
 120static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
 121{
 122	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
 123}
 124
 125/* Update vma->vm_page_prot to reflect vma->vm_flags. */
 126void vma_set_page_prot(struct vm_area_struct *vma)
 127{
 128	unsigned long vm_flags = vma->vm_flags;
 129	pgprot_t vm_page_prot;
 130
 131	vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
 132	if (vma_wants_writenotify(vma, vm_page_prot)) {
 133		vm_flags &= ~VM_SHARED;
 134		vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
 135	}
 136	/* remove_protection_ptes reads vma->vm_page_prot without mmap_sem */
 137	WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
 138}
 139
 140/*
 141 * Requires inode->i_mapping->i_mmap_rwsem
 142 */
 143static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 144		struct file *file, struct address_space *mapping)
 145{
 146	if (vma->vm_flags & VM_DENYWRITE)
 147		atomic_inc(&file_inode(file)->i_writecount);
 148	if (vma->vm_flags & VM_SHARED)
 149		mapping_unmap_writable(mapping);
 150
 151	flush_dcache_mmap_lock(mapping);
 152	vma_interval_tree_remove(vma, &mapping->i_mmap);
 153	flush_dcache_mmap_unlock(mapping);
 154}
 155
 156/*
 157 * Unlink a file-based vm structure from its interval tree, to hide
 158 * vma from rmap and vmtruncate before freeing its page tables.
 159 */
 160void unlink_file_vma(struct vm_area_struct *vma)
 161{
 162	struct file *file = vma->vm_file;
 163
 164	if (file) {
 165		struct address_space *mapping = file->f_mapping;
 166		i_mmap_lock_write(mapping);
 167		__remove_shared_vm_struct(vma, file, mapping);
 168		i_mmap_unlock_write(mapping);
 169	}
 170}
 171
 172/*
 173 * Close a vm structure and free it, returning the next.
 174 */
 175static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
 176{
 177	struct vm_area_struct *next = vma->vm_next;
 178
 179	might_sleep();
 180	if (vma->vm_ops && vma->vm_ops->close)
 181		vma->vm_ops->close(vma);
 182	if (vma->vm_file)
 183		fput(vma->vm_file);
 184	mpol_put(vma_policy(vma));
 185	kmem_cache_free(vm_area_cachep, vma);
 186	return next;
 187}
 188
 189static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 190
 
 
 
 
 
 
 
 191SYSCALL_DEFINE1(brk, unsigned long, brk)
 192{
 193	unsigned long retval;
 194	unsigned long newbrk, oldbrk;
 195	struct mm_struct *mm = current->mm;
 196	struct vm_area_struct *next;
 197	unsigned long min_brk;
 198	bool populate;
 
 199	LIST_HEAD(uf);
 
 200
 201	if (down_write_killable(&mm->mmap_sem))
 202		return -EINTR;
 203
 
 
 204#ifdef CONFIG_COMPAT_BRK
 205	/*
 206	 * CONFIG_COMPAT_BRK can still be overridden by setting
 207	 * randomize_va_space to 2, which will still cause mm->start_brk
 208	 * to be arbitrarily shifted
 209	 */
 210	if (current->brk_randomized)
 211		min_brk = mm->start_brk;
 212	else
 213		min_brk = mm->end_data;
 214#else
 215	min_brk = mm->start_brk;
 216#endif
 217	if (brk < min_brk)
 218		goto out;
 219
 220	/*
 221	 * Check against rlimit here. If this check is done later after the test
 222	 * of oldbrk with newbrk then it can escape the test and let the data
 223	 * segment grow beyond its set limit the in case where the limit is
 224	 * not page aligned -Ram Gupta
 225	 */
 226	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
 227			      mm->end_data, mm->start_data))
 228		goto out;
 229
 230	newbrk = PAGE_ALIGN(brk);
 231	oldbrk = PAGE_ALIGN(mm->brk);
 232	if (oldbrk == newbrk)
 233		goto set_brk;
 
 
 234
 235	/* Always allow shrinking brk. */
 
 
 
 236	if (brk <= mm->brk) {
 237		if (!do_munmap(mm, newbrk, oldbrk-newbrk, &uf))
 238			goto set_brk;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 239		goto out;
 240	}
 241
 242	/* Check against existing mmap mappings. */
 243	next = find_vma(mm, oldbrk);
 
 
 
 
 
 
 
 244	if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
 245		goto out;
 246
 
 247	/* Ok, looks good - let it rip. */
 248	if (do_brk(oldbrk, newbrk-oldbrk, &uf) < 0)
 249		goto out;
 250
 251set_brk:
 252	mm->brk = brk;
 
 
 253	populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
 254	up_write(&mm->mmap_sem);
 
 
 
 255	userfaultfd_unmap_complete(mm, &uf);
 256	if (populate)
 257		mm_populate(oldbrk, newbrk - oldbrk);
 258	return brk;
 259
 260out:
 261	retval = mm->brk;
 262	up_write(&mm->mmap_sem);
 263	return retval;
 264}
 265
 266static long vma_compute_subtree_gap(struct vm_area_struct *vma)
 267{
 268	unsigned long max, prev_end, subtree_gap;
 269
 270	/*
 271	 * Note: in the rare case of a VM_GROWSDOWN above a VM_GROWSUP, we
 272	 * allow two stack_guard_gaps between them here, and when choosing
 273	 * an unmapped area; whereas when expanding we only require one.
 274	 * That's a little inconsistent, but keeps the code here simpler.
 275	 */
 276	max = vm_start_gap(vma);
 277	if (vma->vm_prev) {
 278		prev_end = vm_end_gap(vma->vm_prev);
 279		if (max > prev_end)
 280			max -= prev_end;
 281		else
 282			max = 0;
 283	}
 284	if (vma->vm_rb.rb_left) {
 285		subtree_gap = rb_entry(vma->vm_rb.rb_left,
 286				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 287		if (subtree_gap > max)
 288			max = subtree_gap;
 289	}
 290	if (vma->vm_rb.rb_right) {
 291		subtree_gap = rb_entry(vma->vm_rb.rb_right,
 292				struct vm_area_struct, vm_rb)->rb_subtree_gap;
 293		if (subtree_gap > max)
 294			max = subtree_gap;
 295	}
 296	return max;
 297}
 298
 299#ifdef CONFIG_DEBUG_VM_RB
 300static int browse_rb(struct mm_struct *mm)
 301{
 302	struct rb_root *root = &mm->mm_rb;
 303	int i = 0, j, bug = 0;
 304	struct rb_node *nd, *pn = NULL;
 305	unsigned long prev = 0, pend = 0;
 306
 307	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 308		struct vm_area_struct *vma;
 309		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 310		if (vma->vm_start < prev) {
 311			pr_emerg("vm_start %lx < prev %lx\n",
 312				  vma->vm_start, prev);
 313			bug = 1;
 314		}
 315		if (vma->vm_start < pend) {
 316			pr_emerg("vm_start %lx < pend %lx\n",
 317				  vma->vm_start, pend);
 318			bug = 1;
 319		}
 320		if (vma->vm_start > vma->vm_end) {
 321			pr_emerg("vm_start %lx > vm_end %lx\n",
 322				  vma->vm_start, vma->vm_end);
 323			bug = 1;
 324		}
 325		spin_lock(&mm->page_table_lock);
 326		if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
 327			pr_emerg("free gap %lx, correct %lx\n",
 328			       vma->rb_subtree_gap,
 329			       vma_compute_subtree_gap(vma));
 330			bug = 1;
 331		}
 332		spin_unlock(&mm->page_table_lock);
 333		i++;
 334		pn = nd;
 335		prev = vma->vm_start;
 336		pend = vma->vm_end;
 337	}
 338	j = 0;
 339	for (nd = pn; nd; nd = rb_prev(nd))
 340		j++;
 341	if (i != j) {
 342		pr_emerg("backwards %d, forwards %d\n", j, i);
 343		bug = 1;
 344	}
 345	return bug ? -1 : i;
 346}
 347
 348static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
 349{
 350	struct rb_node *nd;
 351
 352	for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 353		struct vm_area_struct *vma;
 354		vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 355		VM_BUG_ON_VMA(vma != ignore &&
 356			vma->rb_subtree_gap != vma_compute_subtree_gap(vma),
 357			vma);
 358	}
 359}
 360
 361static void validate_mm(struct mm_struct *mm)
 362{
 363	int bug = 0;
 364	int i = 0;
 365	unsigned long highest_address = 0;
 366	struct vm_area_struct *vma = mm->mmap;
 367
 368	while (vma) {
 
 
 
 369		struct anon_vma *anon_vma = vma->anon_vma;
 370		struct anon_vma_chain *avc;
 371
 372		if (anon_vma) {
 373			anon_vma_lock_read(anon_vma);
 374			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 375				anon_vma_interval_tree_verify(avc);
 376			anon_vma_unlock_read(anon_vma);
 377		}
 378
 379		highest_address = vm_end_gap(vma);
 380		vma = vma->vm_next;
 381		i++;
 382	}
 383	if (i != mm->map_count) {
 384		pr_emerg("map_count %d vm_next %d\n", mm->map_count, i);
 385		bug = 1;
 386	}
 387	if (highest_address != mm->highest_vm_end) {
 388		pr_emerg("mm->highest_vm_end %lx, found %lx\n",
 389			  mm->highest_vm_end, highest_address);
 390		bug = 1;
 391	}
 392	i = browse_rb(mm);
 393	if (i != mm->map_count) {
 394		if (i != -1)
 395			pr_emerg("map_count %d rb %d\n", mm->map_count, i);
 396		bug = 1;
 397	}
 398	VM_BUG_ON_MM(bug, mm);
 399}
 400#else
 401#define validate_mm_rb(root, ignore) do { } while (0)
 
 402#define validate_mm(mm) do { } while (0)
 403#endif
 404
 405RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
 406		     unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
 407
 408/*
 409 * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
 410 * vma->vm_prev->vm_end values changed, without modifying the vma's position
 411 * in the rbtree.
 412 */
 413static void vma_gap_update(struct vm_area_struct *vma)
 414{
 415	/*
 416	 * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
 417	 * function that does exacltly what we want.
 418	 */
 419	vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
 420}
 421
 422static inline void vma_rb_insert(struct vm_area_struct *vma,
 423				 struct rb_root *root)
 424{
 425	/* All rb_subtree_gap values must be consistent prior to insertion */
 426	validate_mm_rb(root, NULL);
 427
 428	rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 429}
 430
 431static void __vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
 432{
 433	/*
 434	 * Note rb_erase_augmented is a fairly large inline function,
 435	 * so make sure we instantiate it only once with our desired
 436	 * augmented rbtree callbacks.
 437	 */
 438	rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
 439}
 440
 441static __always_inline void vma_rb_erase_ignore(struct vm_area_struct *vma,
 442						struct rb_root *root,
 443						struct vm_area_struct *ignore)
 444{
 445	/*
 446	 * All rb_subtree_gap values must be consistent prior to erase,
 447	 * with the possible exception of the "next" vma being erased if
 448	 * next->vm_start was reduced.
 449	 */
 450	validate_mm_rb(root, ignore);
 451
 452	__vma_rb_erase(vma, root);
 453}
 454
 455static __always_inline void vma_rb_erase(struct vm_area_struct *vma,
 456					 struct rb_root *root)
 457{
 458	/*
 459	 * All rb_subtree_gap values must be consistent prior to erase,
 460	 * with the possible exception of the vma being erased.
 461	 */
 462	validate_mm_rb(root, vma);
 463
 464	__vma_rb_erase(vma, root);
 465}
 466
 467/*
 468 * vma has some anon_vma assigned, and is already inserted on that
 469 * anon_vma's interval trees.
 470 *
 471 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 472 * vma must be removed from the anon_vma's interval trees using
 473 * anon_vma_interval_tree_pre_update_vma().
 474 *
 475 * After the update, the vma will be reinserted using
 476 * anon_vma_interval_tree_post_update_vma().
 477 *
 478 * The entire update must be protected by exclusive mmap_sem and by
 479 * the root anon_vma's mutex.
 480 */
 481static inline void
 482anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 483{
 484	struct anon_vma_chain *avc;
 485
 486	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 487		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 488}
 489
 490static inline void
 491anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 492{
 493	struct anon_vma_chain *avc;
 494
 495	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 496		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 497}
 498
 499static int find_vma_links(struct mm_struct *mm, unsigned long addr,
 500		unsigned long end, struct vm_area_struct **pprev,
 501		struct rb_node ***rb_link, struct rb_node **rb_parent)
 502{
 503	struct rb_node **__rb_link, *__rb_parent, *rb_prev;
 504
 505	__rb_link = &mm->mm_rb.rb_node;
 506	rb_prev = __rb_parent = NULL;
 507
 508	while (*__rb_link) {
 509		struct vm_area_struct *vma_tmp;
 510
 511		__rb_parent = *__rb_link;
 512		vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 513
 514		if (vma_tmp->vm_end > addr) {
 515			/* Fail if an existing vma overlaps the area */
 516			if (vma_tmp->vm_start < end)
 517				return -ENOMEM;
 518			__rb_link = &__rb_parent->rb_left;
 519		} else {
 520			rb_prev = __rb_parent;
 521			__rb_link = &__rb_parent->rb_right;
 522		}
 523	}
 524
 525	*pprev = NULL;
 526	if (rb_prev)
 527		*pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 528	*rb_link = __rb_link;
 529	*rb_parent = __rb_parent;
 530	return 0;
 531}
 532
 533static unsigned long count_vma_pages_range(struct mm_struct *mm,
 534		unsigned long addr, unsigned long end)
 535{
 
 
 536	unsigned long nr_pages = 0;
 537	struct vm_area_struct *vma;
 538
 539	/* Find first overlaping mapping */
 540	vma = find_vma_intersection(mm, addr, end);
 541	if (!vma)
 542		return 0;
 543
 544	nr_pages = (min(end, vma->vm_end) -
 545		max(addr, vma->vm_start)) >> PAGE_SHIFT;
 546
 547	/* Iterate over the rest of the overlaps */
 548	for (vma = vma->vm_next; vma; vma = vma->vm_next) {
 549		unsigned long overlap_len;
 550
 551		if (vma->vm_start > end)
 552			break;
 
 553
 554		overlap_len = min(end, vma->vm_end) - vma->vm_start;
 555		nr_pages += overlap_len >> PAGE_SHIFT;
 556	}
 557
 558	return nr_pages;
 559}
 560
 561void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 562		struct rb_node **rb_link, struct rb_node *rb_parent)
 563{
 564	/* Update tracking information for the gap following the new vma. */
 565	if (vma->vm_next)
 566		vma_gap_update(vma->vm_next);
 567	else
 568		mm->highest_vm_end = vm_end_gap(vma);
 569
 570	/*
 571	 * vma->vm_prev wasn't known when we followed the rbtree to find the
 572	 * correct insertion point for that vma. As a result, we could not
 573	 * update the vma vm_rb parents rb_subtree_gap values on the way down.
 574	 * So, we first insert the vma with a zero rb_subtree_gap value
 575	 * (to be consistent with what we did on the way down), and then
 576	 * immediately update the gap to the correct value. Finally we
 577	 * rebalance the rbtree after all augmented values have been set.
 578	 */
 579	rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 580	vma->rb_subtree_gap = 0;
 581	vma_gap_update(vma);
 582	vma_rb_insert(vma, &mm->mm_rb);
 583}
 584
 585static void __vma_link_file(struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 586{
 587	struct file *file;
 
 
 
 588
 589	file = vma->vm_file;
 590	if (file) {
 591		struct address_space *mapping = file->f_mapping;
 592
 593		if (vma->vm_flags & VM_DENYWRITE)
 594			atomic_dec(&file_inode(file)->i_writecount);
 595		if (vma->vm_flags & VM_SHARED)
 596			atomic_inc(&mapping->i_mmap_writable);
 597
 598		flush_dcache_mmap_lock(mapping);
 599		vma_interval_tree_insert(vma, &mapping->i_mmap);
 600		flush_dcache_mmap_unlock(mapping);
 601	}
 
 
 602}
 603
 604static void
 605__vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 606	struct vm_area_struct *prev, struct rb_node **rb_link,
 607	struct rb_node *rb_parent)
 608{
 609	__vma_link_list(mm, vma, prev, rb_parent);
 610	__vma_link_rb(mm, vma, rb_link, rb_parent);
 
 
 
 
 
 
 
 611}
 612
 613static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 614			struct vm_area_struct *prev, struct rb_node **rb_link,
 615			struct rb_node *rb_parent)
 616{
 
 617	struct address_space *mapping = NULL;
 618
 
 
 
 619	if (vma->vm_file) {
 620		mapping = vma->vm_file->f_mapping;
 621		i_mmap_lock_write(mapping);
 622	}
 623
 624	__vma_link(mm, vma, prev, rb_link, rb_parent);
 625	__vma_link_file(vma);
 626
 627	if (mapping)
 
 628		i_mmap_unlock_write(mapping);
 
 629
 630	mm->map_count++;
 631	validate_mm(mm);
 
 632}
 633
 634/*
 635 * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
 636 * mm's list and rbtree.  It has already been inserted into the interval tree.
 637 */
 638static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 639{
 640	struct vm_area_struct *prev;
 641	struct rb_node **rb_link, *rb_parent;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 642
 643	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
 644			   &prev, &rb_link, &rb_parent))
 645		BUG();
 646	__vma_link(mm, vma, prev, rb_link, rb_parent);
 647	mm->map_count++;
 648}
 649
 650static __always_inline void __vma_unlink_common(struct mm_struct *mm,
 651						struct vm_area_struct *vma,
 652						struct vm_area_struct *prev,
 653						bool has_prev,
 654						struct vm_area_struct *ignore)
 655{
 656	struct vm_area_struct *next;
 657
 658	vma_rb_erase_ignore(vma, &mm->mm_rb, ignore);
 659	next = vma->vm_next;
 660	if (has_prev)
 661		prev->vm_next = next;
 662	else {
 663		prev = vma->vm_prev;
 664		if (prev)
 665			prev->vm_next = next;
 666		else
 667			mm->mmap = next;
 668	}
 669	if (next)
 670		next->vm_prev = prev;
 671
 672	/* Kill the cache */
 673	vmacache_invalidate(mm);
 674}
 675
 676static inline void __vma_unlink_prev(struct mm_struct *mm,
 677				     struct vm_area_struct *vma,
 678				     struct vm_area_struct *prev)
 679{
 680	__vma_unlink_common(mm, vma, prev, true, vma);
 681}
 682
 683/*
 684 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 685 * is already present in an i_mmap tree without adjusting the tree.
 686 * The following helper function should be used when such adjustments
 687 * are necessary.  The "insert" vma (if any) is to be inserted
 688 * before we drop the necessary locks.
 689 */
 690int __vma_adjust(struct vm_area_struct *vma, unsigned long start,
 691	unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert,
 692	struct vm_area_struct *expand)
 693{
 694	struct mm_struct *mm = vma->vm_mm;
 695	struct vm_area_struct *next = vma->vm_next, *orig_vma = vma;
 
 
 696	struct address_space *mapping = NULL;
 697	struct rb_root_cached *root = NULL;
 698	struct anon_vma *anon_vma = NULL;
 699	struct file *file = vma->vm_file;
 700	bool start_changed = false, end_changed = false;
 701	long adjust_next = 0;
 702	int remove_next = 0;
 
 
 703
 704	if (next && !insert) {
 705		struct vm_area_struct *exporter = NULL, *importer = NULL;
 706
 707		if (end >= next->vm_end) {
 708			/*
 709			 * vma expands, overlapping all the next, and
 710			 * perhaps the one after too (mprotect case 6).
 711			 * The only other cases that gets here are
 712			 * case 1, case 7 and case 8.
 713			 */
 714			if (next == expand) {
 715				/*
 716				 * The only case where we don't expand "vma"
 717				 * and we expand "next" instead is case 8.
 718				 */
 719				VM_WARN_ON(end != next->vm_end);
 720				/*
 721				 * remove_next == 3 means we're
 722				 * removing "vma" and that to do so we
 723				 * swapped "vma" and "next".
 724				 */
 725				remove_next = 3;
 726				VM_WARN_ON(file != next->vm_file);
 727				swap(vma, next);
 728			} else {
 729				VM_WARN_ON(expand != vma);
 730				/*
 731				 * case 1, 6, 7, remove_next == 2 is case 6,
 732				 * remove_next == 1 is case 1 or 7.
 733				 */
 734				remove_next = 1 + (end > next->vm_end);
 
 
 
 735				VM_WARN_ON(remove_next == 2 &&
 736					   end != next->vm_next->vm_end);
 737				VM_WARN_ON(remove_next == 1 &&
 738					   end != next->vm_end);
 739				/* trim end to next, for case 6 first pass */
 740				end = next->vm_end;
 741			}
 742
 743			exporter = next;
 744			importer = vma;
 745
 746			/*
 747			 * If next doesn't have anon_vma, import from vma after
 748			 * next, if the vma overlaps with it.
 749			 */
 750			if (remove_next == 2 && !next->anon_vma)
 751				exporter = next->vm_next;
 752
 753		} else if (end > next->vm_start) {
 754			/*
 755			 * vma expands, overlapping part of the next:
 756			 * mprotect case 5 shifting the boundary up.
 757			 */
 758			adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
 759			exporter = next;
 760			importer = vma;
 761			VM_WARN_ON(expand != importer);
 762		} else if (end < vma->vm_end) {
 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			 */
 768			adjust_next = -((vma->vm_end - end) >> PAGE_SHIFT);
 769			exporter = vma;
 770			importer = next;
 771			VM_WARN_ON(expand != importer);
 772		}
 773
 774		/*
 775		 * Easily overlooked: when mprotect shifts the boundary,
 776		 * make sure the expanding vma has anon_vma set if the
 777		 * shrinking vma had, to cover any anon pages imported.
 778		 */
 779		if (exporter && exporter->anon_vma && !importer->anon_vma) {
 780			int error;
 781
 782			importer->anon_vma = exporter->anon_vma;
 783			error = anon_vma_clone(importer, exporter);
 784			if (error)
 785				return error;
 786		}
 787	}
 788again:
 
 
 
 789	vma_adjust_trans_huge(orig_vma, start, end, adjust_next);
 790
 791	if (file) {
 792		mapping = file->f_mapping;
 793		root = &mapping->i_mmap;
 794		uprobe_munmap(vma, vma->vm_start, vma->vm_end);
 795
 796		if (adjust_next)
 797			uprobe_munmap(next, next->vm_start, next->vm_end);
 798
 799		i_mmap_lock_write(mapping);
 800		if (insert) {
 801			/*
 802			 * Put into interval tree now, so instantiated pages
 803			 * are visible to arm/parisc __flush_dcache_page
 804			 * throughout; but we cannot insert into address
 805			 * space until vma start or end is updated.
 806			 */
 807			__vma_link_file(insert);
 808		}
 809	}
 810
 811	anon_vma = vma->anon_vma;
 812	if (!anon_vma && adjust_next)
 813		anon_vma = next->anon_vma;
 814	if (anon_vma) {
 815		VM_WARN_ON(adjust_next && next->anon_vma &&
 816			   anon_vma != next->anon_vma);
 817		anon_vma_lock_write(anon_vma);
 818		anon_vma_interval_tree_pre_update_vma(vma);
 819		if (adjust_next)
 820			anon_vma_interval_tree_pre_update_vma(next);
 821	}
 822
 823	if (root) {
 824		flush_dcache_mmap_lock(mapping);
 825		vma_interval_tree_remove(vma, root);
 826		if (adjust_next)
 827			vma_interval_tree_remove(next, root);
 828	}
 829
 830	if (start != vma->vm_start) {
 
 
 
 
 
 
 
 831		vma->vm_start = start;
 832		start_changed = true;
 833	}
 834	if (end != vma->vm_end) {
 
 
 
 
 
 
 
 
 
 
 835		vma->vm_end = end;
 836		end_changed = true;
 837	}
 
 
 
 
 838	vma->vm_pgoff = pgoff;
 839	if (adjust_next) {
 840		next->vm_start += adjust_next << PAGE_SHIFT;
 841		next->vm_pgoff += adjust_next;
 
 842	}
 843
 844	if (root) {
 845		if (adjust_next)
 846			vma_interval_tree_insert(next, root);
 847		vma_interval_tree_insert(vma, root);
 848		flush_dcache_mmap_unlock(mapping);
 849	}
 850
 851	if (remove_next) {
 852		/*
 853		 * vma_merge has merged next into vma, and needs
 854		 * us to remove next before dropping the locks.
 855		 */
 856		if (remove_next != 3)
 857			__vma_unlink_prev(mm, next, vma);
 858		else
 859			/*
 860			 * vma is not before next if they've been
 861			 * swapped.
 862			 *
 863			 * pre-swap() next->vm_start was reduced so
 864			 * tell validate_mm_rb to ignore pre-swap()
 865			 * "next" (which is stored in post-swap()
 866			 * "vma").
 867			 */
 868			__vma_unlink_common(mm, next, NULL, false, vma);
 869		if (file)
 870			__remove_shared_vm_struct(next, file, mapping);
 871	} else if (insert) {
 872		/*
 873		 * split_vma has split insert from vma, and needs
 874		 * us to insert it before dropping the locks
 875		 * (it may either follow vma or precede it).
 876		 */
 877		__insert_vm_struct(mm, insert);
 878	} else {
 879		if (start_changed)
 880			vma_gap_update(vma);
 881		if (end_changed) {
 882			if (!next)
 883				mm->highest_vm_end = vm_end_gap(vma);
 884			else if (!adjust_next)
 885				vma_gap_update(next);
 886		}
 887	}
 888
 889	if (anon_vma) {
 890		anon_vma_interval_tree_post_update_vma(vma);
 891		if (adjust_next)
 892			anon_vma_interval_tree_post_update_vma(next);
 893		anon_vma_unlock_write(anon_vma);
 894	}
 895	if (mapping)
 
 896		i_mmap_unlock_write(mapping);
 897
 898	if (root) {
 899		uprobe_mmap(vma);
 900
 901		if (adjust_next)
 902			uprobe_mmap(next);
 903	}
 904
 905	if (remove_next) {
 
 906		if (file) {
 907			uprobe_munmap(next, next->vm_start, next->vm_end);
 908			fput(file);
 909		}
 910		if (next->anon_vma)
 911			anon_vma_merge(vma, next);
 912		mm->map_count--;
 913		mpol_put(vma_policy(next));
 914		kmem_cache_free(vm_area_cachep, next);
 
 
 
 915		/*
 916		 * In mprotect's case 6 (see comments on vma_merge),
 917		 * we must remove another next too. It would clutter
 918		 * up the code too much to do both in one go.
 919		 */
 920		if (remove_next != 3) {
 921			/*
 922			 * If "next" was removed and vma->vm_end was
 923			 * expanded (up) over it, in turn
 924			 * "next->vm_prev->vm_end" changed and the
 925			 * "vma->vm_next" gap must be updated.
 926			 */
 927			next = vma->vm_next;
 928		} else {
 929			/*
 930			 * For the scope of the comment "next" and
 931			 * "vma" considered pre-swap(): if "vma" was
 932			 * removed, next->vm_start was expanded (down)
 933			 * over it and the "next" gap must be updated.
 934			 * Because of the swap() the post-swap() "vma"
 935			 * actually points to pre-swap() "next"
 936			 * (post-swap() "next" as opposed is now a
 937			 * dangling pointer).
 938			 */
 939			next = vma;
 940		}
 941		if (remove_next == 2) {
 942			remove_next = 1;
 943			end = next->vm_end;
 944			goto again;
 945		}
 946		else if (next)
 947			vma_gap_update(next);
 948		else {
 949			/*
 950			 * If remove_next == 2 we obviously can't
 951			 * reach this path.
 952			 *
 953			 * If remove_next == 3 we can't reach this
 954			 * path because pre-swap() next is always not
 955			 * NULL. pre-swap() "next" is not being
 956			 * removed and its next->vm_end is not altered
 957			 * (and furthermore "end" already matches
 958			 * next->vm_end in remove_next == 3).
 959			 *
 960			 * We reach this only in the remove_next == 1
 961			 * case if the "next" vma that was removed was
 962			 * the highest vma of the mm. However in such
 963			 * case next->vm_end == "end" and the extended
 964			 * "vma" has vma->vm_end == next->vm_end so
 965			 * mm->highest_vm_end doesn't need any update
 966			 * in remove_next == 1 case.
 967			 */
 968			VM_WARN_ON(mm->highest_vm_end != vm_end_gap(vma));
 969		}
 970	}
 971	if (insert && file)
 972		uprobe_mmap(insert);
 973
 
 974	validate_mm(mm);
 975
 976	return 0;
 977}
 978
 979/*
 980 * If the vma has a ->close operation then the driver probably needs to release
 981 * per-vma resources, so we don't attempt to merge those.
 982 */
 983static inline int is_mergeable_vma(struct vm_area_struct *vma,
 984				struct file *file, unsigned long vm_flags,
 985				struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 
 986{
 987	/*
 988	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
 989	 * match the flags but dirty bit -- the caller should mark
 990	 * merged VMA as dirty. If dirty bit won't be excluded from
 991	 * comparison, we increase pressue on the memory system forcing
 992	 * the kernel to generate new VMAs when old one could be
 993	 * extended instead.
 994	 */
 995	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
 996		return 0;
 997	if (vma->vm_file != file)
 998		return 0;
 999	if (vma->vm_ops && vma->vm_ops->close)
1000		return 0;
1001	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
1002		return 0;
 
 
1003	return 1;
1004}
1005
1006static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
1007					struct anon_vma *anon_vma2,
1008					struct vm_area_struct *vma)
1009{
1010	/*
1011	 * The list_is_singular() test is to avoid merging VMA cloned from
1012	 * parents. This can improve scalability caused by anon_vma lock.
1013	 */
1014	if ((!anon_vma1 || !anon_vma2) && (!vma ||
1015		list_is_singular(&vma->anon_vma_chain)))
1016		return 1;
1017	return anon_vma1 == anon_vma2;
1018}
1019
1020/*
1021 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1022 * in front of (at a lower virtual address and file offset than) the vma.
1023 *
1024 * We cannot merge two vmas if they have differently assigned (non-NULL)
1025 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1026 *
1027 * We don't check here for the merged mmap wrapping around the end of pagecache
1028 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
1029 * wrap, nor mmaps which cover the final page at index -1UL.
1030 */
1031static int
1032can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
1033		     struct anon_vma *anon_vma, struct file *file,
1034		     pgoff_t vm_pgoff,
1035		     struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 
1036{
1037	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1038	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1039		if (vma->vm_pgoff == vm_pgoff)
1040			return 1;
1041	}
1042	return 0;
1043}
1044
1045/*
1046 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
1047 * beyond (at a higher virtual address and file offset than) the vma.
1048 *
1049 * We cannot merge two vmas if they have differently assigned (non-NULL)
1050 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
1051 */
1052static int
1053can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
1054		    struct anon_vma *anon_vma, struct file *file,
1055		    pgoff_t vm_pgoff,
1056		    struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 
1057{
1058	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx) &&
1059	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
1060		pgoff_t vm_pglen;
1061		vm_pglen = vma_pages(vma);
1062		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
1063			return 1;
1064	}
1065	return 0;
1066}
1067
1068/*
1069 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
1070 * whether that can be merged with its predecessor or its successor.
1071 * Or both (it neatly fills a hole).
1072 *
1073 * In most cases - when called for mmap, brk or mremap - [addr,end) is
1074 * certain not to be mapped by the time vma_merge is called; but when
1075 * called for mprotect, it is certain to be already mapped (either at
1076 * an offset within prev, or at the start of next), and the flags of
1077 * this area are about to be changed to vm_flags - and the no-change
1078 * case has already been eliminated.
1079 *
1080 * The following mprotect cases have to be considered, where AAAA is
1081 * the area passed down from mprotect_fixup, never extending beyond one
1082 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
1083 *
1084 *     AAAA             AAAA                AAAA          AAAA
1085 *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
1086 *    cannot merge    might become    might become    might become
1087 *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
1088 *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
1089 *    mremap move:                                    PPPPXXXXXXXX 8
1090 *        AAAA
1091 *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
1092 *    might become    case 1 below    case 2 below    case 3 below
 
 
 
1093 *
1094 * It is important for case 8 that the the vma NNNN overlapping the
1095 * region AAAA is never going to extended over XXXX. Instead XXXX must
1096 * be extended in region AAAA and NNNN must be removed. This way in
1097 * all cases where vma_merge succeeds, the moment vma_adjust drops the
1098 * rmap_locks, the properties of the merged vma will be already
1099 * correct for the whole merged range. Some of those properties like
1100 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
1101 * be correct for the whole merged range immediately after the
1102 * rmap_locks are released. Otherwise if XXXX would be removed and
1103 * NNNN would be extended over the XXXX range, remove_migration_ptes
1104 * or other rmap walkers (if working on addresses beyond the "end"
1105 * parameter) may establish ptes with the wrong permissions of NNNN
1106 * instead of the right permissions of XXXX.
1107 */
1108struct vm_area_struct *vma_merge(struct mm_struct *mm,
1109			struct vm_area_struct *prev, unsigned long addr,
1110			unsigned long end, unsigned long vm_flags,
1111			struct anon_vma *anon_vma, struct file *file,
1112			pgoff_t pgoff, struct mempolicy *policy,
1113			struct vm_userfaultfd_ctx vm_userfaultfd_ctx)
 
1114{
1115	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
1116	struct vm_area_struct *area, *next;
1117	int err;
 
 
1118
1119	/*
1120	 * We later require that vma->vm_flags == vm_flags,
1121	 * so this tests vma->vm_flags & VM_SPECIAL, too.
1122	 */
1123	if (vm_flags & VM_SPECIAL)
1124		return NULL;
1125
1126	if (prev)
1127		next = prev->vm_next;
1128	else
1129		next = mm->mmap;
1130	area = next;
1131	if (area && area->vm_end == end)		/* cases 6, 7, 8 */
1132		next = next->vm_next;
1133
1134	/* verify some invariant that must be enforced by the caller */
1135	VM_WARN_ON(prev && addr <= prev->vm_start);
1136	VM_WARN_ON(area && end > area->vm_end);
1137	VM_WARN_ON(addr >= end);
1138
1139	/*
1140	 * Can it merge with the predecessor?
1141	 */
1142	if (prev && prev->vm_end == addr &&
1143			mpol_equal(vma_policy(prev), policy) &&
1144			can_vma_merge_after(prev, vm_flags,
1145					    anon_vma, file, pgoff,
1146					    vm_userfaultfd_ctx)) {
1147		/*
1148		 * OK, it can.  Can we now merge in the successor as well?
1149		 */
1150		if (next && end == next->vm_start &&
1151				mpol_equal(policy, vma_policy(next)) &&
1152				can_vma_merge_before(next, vm_flags,
1153						     anon_vma, file,
1154						     pgoff+pglen,
1155						     vm_userfaultfd_ctx) &&
1156				is_mergeable_anon_vma(prev->anon_vma,
1157						      next->anon_vma, NULL)) {
1158							/* cases 1, 6 */
1159			err = __vma_adjust(prev, prev->vm_start,
1160					 next->vm_end, prev->vm_pgoff, NULL,
1161					 prev);
1162		} else					/* cases 2, 5, 7 */
1163			err = __vma_adjust(prev, prev->vm_start,
1164					 end, prev->vm_pgoff, NULL, prev);
1165		if (err)
1166			return NULL;
1167		khugepaged_enter_vma_merge(prev, vm_flags);
1168		return prev;
1169	}
1170
1171	/*
1172	 * Can this new request be merged in front of next?
1173	 */
1174	if (next && end == next->vm_start &&
1175			mpol_equal(policy, vma_policy(next)) &&
1176			can_vma_merge_before(next, vm_flags,
1177					     anon_vma, file, pgoff+pglen,
1178					     vm_userfaultfd_ctx)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1179		if (prev && addr < prev->vm_end)	/* case 4 */
1180			err = __vma_adjust(prev, prev->vm_start,
1181					 addr, prev->vm_pgoff, NULL, next);
1182		else {					/* cases 3, 8 */
1183			err = __vma_adjust(area, addr, next->vm_end,
1184					 next->vm_pgoff - pglen, NULL, next);
1185			/*
1186			 * In case 3 area is already equal to next and
1187			 * this is a noop, but in case 8 "area" has
1188			 * been removed and next was expanded over it.
1189			 */
1190			area = next;
1191		}
1192		if (err)
1193			return NULL;
1194		khugepaged_enter_vma_merge(area, vm_flags);
1195		return area;
1196	}
1197
1198	return NULL;
 
 
 
 
 
 
1199}
1200
1201/*
1202 * Rough compatbility check to quickly see if it's even worth looking
1203 * at sharing an anon_vma.
1204 *
1205 * They need to have the same vm_file, and the flags can only differ
1206 * in things that mprotect may change.
1207 *
1208 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1209 * we can merge the two vma's. For example, we refuse to merge a vma if
1210 * there is a vm_ops->close() function, because that indicates that the
1211 * driver is doing some kind of reference counting. But that doesn't
1212 * really matter for the anon_vma sharing case.
1213 */
1214static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1215{
1216	return a->vm_end == b->vm_start &&
1217		mpol_equal(vma_policy(a), vma_policy(b)) &&
1218		a->vm_file == b->vm_file &&
1219		!((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC|VM_SOFTDIRTY)) &&
1220		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1221}
1222
1223/*
1224 * Do some basic sanity checking to see if we can re-use the anon_vma
1225 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1226 * the same as 'old', the other will be the new one that is trying
1227 * to share the anon_vma.
1228 *
1229 * NOTE! This runs with mm_sem held for reading, so it is possible that
1230 * the anon_vma of 'old' is concurrently in the process of being set up
1231 * by another page fault trying to merge _that_. But that's ok: if it
1232 * is being set up, that automatically means that it will be a singleton
1233 * acceptable for merging, so we can do all of this optimistically. But
1234 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1235 *
1236 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1237 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1238 * is to return an anon_vma that is "complex" due to having gone through
1239 * a fork).
1240 *
1241 * We also make sure that the two vma's are compatible (adjacent,
1242 * and with the same memory policies). That's all stable, even with just
1243 * a read lock on the mm_sem.
1244 */
1245static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1246{
1247	if (anon_vma_compatible(a, b)) {
1248		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1249
1250		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1251			return anon_vma;
1252	}
1253	return NULL;
1254}
1255
1256/*
1257 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1258 * neighbouring vmas for a suitable anon_vma, before it goes off
1259 * to allocate a new anon_vma.  It checks because a repetitive
1260 * sequence of mprotects and faults may otherwise lead to distinct
1261 * anon_vmas being allocated, preventing vma merge in subsequent
1262 * mprotect.
1263 */
1264struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1265{
1266	struct anon_vma *anon_vma;
1267	struct vm_area_struct *near;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1268
1269	near = vma->vm_next;
1270	if (!near)
1271		goto try_prev;
1272
1273	anon_vma = reusable_anon_vma(near, vma, near);
1274	if (anon_vma)
1275		return anon_vma;
1276try_prev:
1277	near = vma->vm_prev;
1278	if (!near)
1279		goto none;
1280
1281	anon_vma = reusable_anon_vma(near, near, vma);
1282	if (anon_vma)
1283		return anon_vma;
1284none:
1285	/*
 
 
1286	 * There's no absolute need to look only at touching neighbours:
1287	 * we could search further afield for "compatible" anon_vmas.
1288	 * But it would probably just be a waste of time searching,
1289	 * or lead to too many vmas hanging off the same anon_vma.
1290	 * We're trying to allow mprotect remerging later on,
1291	 * not trying to minimize memory used for anon_vmas.
1292	 */
1293	return NULL;
1294}
1295
1296/*
1297 * If a hint addr is less than mmap_min_addr change hint to be as
1298 * low as possible but still greater than mmap_min_addr
1299 */
1300static inline unsigned long round_hint_to_min(unsigned long hint)
1301{
1302	hint &= PAGE_MASK;
1303	if (((void *)hint != NULL) &&
1304	    (hint < mmap_min_addr))
1305		return PAGE_ALIGN(mmap_min_addr);
1306	return hint;
1307}
1308
1309static inline int mlock_future_check(struct mm_struct *mm,
1310				     unsigned long flags,
1311				     unsigned long len)
1312{
1313	unsigned long locked, lock_limit;
1314
1315	/*  mlock MCL_FUTURE? */
1316	if (flags & VM_LOCKED) {
1317		locked = len >> PAGE_SHIFT;
1318		locked += mm->locked_vm;
1319		lock_limit = rlimit(RLIMIT_MEMLOCK);
1320		lock_limit >>= PAGE_SHIFT;
1321		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1322			return -EAGAIN;
1323	}
1324	return 0;
1325}
1326
1327static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1328{
1329	if (S_ISREG(inode->i_mode))
1330		return MAX_LFS_FILESIZE;
1331
1332	if (S_ISBLK(inode->i_mode))
1333		return MAX_LFS_FILESIZE;
1334
 
 
 
1335	/* Special "we do even unsigned file positions" case */
1336	if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1337		return 0;
1338
1339	/* Yes, random drivers might want more. But I'm tired of buggy drivers */
1340	return ULONG_MAX;
1341}
1342
1343static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1344				unsigned long pgoff, unsigned long len)
1345{
1346	u64 maxsize = file_mmap_size_max(file, inode);
1347
1348	if (maxsize && len > maxsize)
1349		return false;
1350	maxsize -= len;
1351	if (pgoff > maxsize >> PAGE_SHIFT)
1352		return false;
1353	return true;
1354}
1355
1356/*
1357 * The caller must hold down_write(&current->mm->mmap_sem).
1358 */
1359unsigned long do_mmap(struct file *file, unsigned long addr,
1360			unsigned long len, unsigned long prot,
1361			unsigned long flags, vm_flags_t vm_flags,
1362			unsigned long pgoff, unsigned long *populate,
1363			struct list_head *uf)
1364{
1365	struct mm_struct *mm = current->mm;
 
1366	int pkey = 0;
1367
 
1368	*populate = 0;
1369
1370	if (!len)
1371		return -EINVAL;
1372
1373	/*
1374	 * Does the application expect PROT_READ to imply PROT_EXEC?
1375	 *
1376	 * (the exception is when the underlying filesystem is noexec
1377	 *  mounted, in which case we dont add PROT_EXEC.)
1378	 */
1379	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1380		if (!(file && path_noexec(&file->f_path)))
1381			prot |= PROT_EXEC;
1382
1383	/* force arch specific MAP_FIXED handling in get_unmapped_area */
1384	if (flags & MAP_FIXED_NOREPLACE)
1385		flags |= MAP_FIXED;
1386
1387	if (!(flags & MAP_FIXED))
1388		addr = round_hint_to_min(addr);
1389
1390	/* Careful about overflows.. */
1391	len = PAGE_ALIGN(len);
1392	if (!len)
1393		return -ENOMEM;
1394
1395	/* offset overflow? */
1396	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1397		return -EOVERFLOW;
1398
1399	/* Too many mappings? */
1400	if (mm->map_count > sysctl_max_map_count)
1401		return -ENOMEM;
1402
1403	/* Obtain the address to map to. we verify (or select) it and ensure
1404	 * that it represents a valid section of the address space.
1405	 */
1406	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1407	if (offset_in_page(addr))
1408		return addr;
1409
1410	if (flags & MAP_FIXED_NOREPLACE) {
1411		struct vm_area_struct *vma = find_vma(mm, addr);
1412
1413		if (vma && vma->vm_start <= addr)
1414			return -EEXIST;
1415	}
1416
1417	if (prot == PROT_EXEC) {
1418		pkey = execute_only_pkey(mm);
1419		if (pkey < 0)
1420			pkey = 0;
1421	}
1422
1423	/* Do simple checking here so the lower-level routines won't have
1424	 * to. we assume access permissions have been handled by the open
1425	 * of the memory object, so we don't do any here.
1426	 */
1427	vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1428			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1429
1430	if (flags & MAP_LOCKED)
1431		if (!can_do_mlock())
1432			return -EPERM;
1433
1434	if (mlock_future_check(mm, vm_flags, len))
1435		return -EAGAIN;
1436
1437	if (file) {
1438		struct inode *inode = file_inode(file);
1439		unsigned long flags_mask;
1440
1441		if (!file_mmap_ok(file, inode, pgoff, len))
1442			return -EOVERFLOW;
1443
1444		flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1445
1446		switch (flags & MAP_TYPE) {
1447		case MAP_SHARED:
1448			/*
1449			 * Force use of MAP_SHARED_VALIDATE with non-legacy
1450			 * flags. E.g. MAP_SYNC is dangerous to use with
1451			 * MAP_SHARED as you don't know which consistency model
1452			 * you will get. We silently ignore unsupported flags
1453			 * with MAP_SHARED to preserve backward compatibility.
1454			 */
1455			flags &= LEGACY_MAP_MASK;
1456			/* fall through */
1457		case MAP_SHARED_VALIDATE:
1458			if (flags & ~flags_mask)
1459				return -EOPNOTSUPP;
1460			if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1461				return -EACCES;
 
 
 
 
1462
1463			/*
1464			 * Make sure we don't allow writing to an append-only
1465			 * file..
1466			 */
1467			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1468				return -EACCES;
1469
1470			/*
1471			 * Make sure there are no mandatory locks on the file.
1472			 */
1473			if (locks_verify_locked(file))
1474				return -EAGAIN;
1475
1476			vm_flags |= VM_SHARED | VM_MAYSHARE;
1477			if (!(file->f_mode & FMODE_WRITE))
1478				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1479
1480			/* fall through */
1481		case MAP_PRIVATE:
1482			if (!(file->f_mode & FMODE_READ))
1483				return -EACCES;
1484			if (path_noexec(&file->f_path)) {
1485				if (vm_flags & VM_EXEC)
1486					return -EPERM;
1487				vm_flags &= ~VM_MAYEXEC;
1488			}
1489
1490			if (!file->f_op->mmap)
1491				return -ENODEV;
1492			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1493				return -EINVAL;
1494			break;
1495
1496		default:
1497			return -EINVAL;
1498		}
1499	} else {
1500		switch (flags & MAP_TYPE) {
1501		case MAP_SHARED:
1502			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1503				return -EINVAL;
1504			/*
1505			 * Ignore pgoff.
1506			 */
1507			pgoff = 0;
1508			vm_flags |= VM_SHARED | VM_MAYSHARE;
1509			break;
1510		case MAP_PRIVATE:
1511			/*
1512			 * Set pgoff according to addr for anon_vma.
1513			 */
1514			pgoff = addr >> PAGE_SHIFT;
1515			break;
1516		default:
1517			return -EINVAL;
1518		}
1519	}
1520
1521	/*
1522	 * Set 'VM_NORESERVE' if we should not account for the
1523	 * memory use of this mapping.
1524	 */
1525	if (flags & MAP_NORESERVE) {
1526		/* We honor MAP_NORESERVE if allowed to overcommit */
1527		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1528			vm_flags |= VM_NORESERVE;
1529
1530		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1531		if (file && is_file_hugepages(file))
1532			vm_flags |= VM_NORESERVE;
1533	}
1534
1535	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1536	if (!IS_ERR_VALUE(addr) &&
1537	    ((vm_flags & VM_LOCKED) ||
1538	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1539		*populate = len;
1540	return addr;
1541}
1542
1543unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1544			      unsigned long prot, unsigned long flags,
1545			      unsigned long fd, unsigned long pgoff)
1546{
1547	struct file *file = NULL;
1548	unsigned long retval;
1549
1550	if (!(flags & MAP_ANONYMOUS)) {
1551		audit_mmap_fd(fd, flags);
1552		file = fget(fd);
1553		if (!file)
1554			return -EBADF;
1555		if (is_file_hugepages(file))
1556			len = ALIGN(len, huge_page_size(hstate_file(file)));
1557		retval = -EINVAL;
1558		if (unlikely(flags & MAP_HUGETLB && !is_file_hugepages(file)))
1559			goto out_fput;
 
1560	} else if (flags & MAP_HUGETLB) {
1561		struct user_struct *user = NULL;
1562		struct hstate *hs;
1563
1564		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1565		if (!hs)
1566			return -EINVAL;
1567
1568		len = ALIGN(len, huge_page_size(hs));
1569		/*
1570		 * VM_NORESERVE is used because the reservations will be
1571		 * taken when vm_ops->mmap() is called
1572		 * A dummy user value is used because we are not locking
1573		 * memory so no accounting is necessary
1574		 */
1575		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1576				VM_NORESERVE,
1577				&user, HUGETLB_ANONHUGE_INODE,
1578				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1579		if (IS_ERR(file))
1580			return PTR_ERR(file);
1581	}
1582
1583	flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1584
1585	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1586out_fput:
1587	if (file)
1588		fput(file);
1589	return retval;
1590}
1591
1592SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1593		unsigned long, prot, unsigned long, flags,
1594		unsigned long, fd, unsigned long, pgoff)
1595{
1596	return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1597}
1598
1599#ifdef __ARCH_WANT_SYS_OLD_MMAP
1600struct mmap_arg_struct {
1601	unsigned long addr;
1602	unsigned long len;
1603	unsigned long prot;
1604	unsigned long flags;
1605	unsigned long fd;
1606	unsigned long offset;
1607};
1608
1609SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1610{
1611	struct mmap_arg_struct a;
1612
1613	if (copy_from_user(&a, arg, sizeof(a)))
1614		return -EFAULT;
1615	if (offset_in_page(a.offset))
1616		return -EINVAL;
1617
1618	return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1619			       a.offset >> PAGE_SHIFT);
1620}
1621#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1622
1623/*
1624 * Some shared mappigns will want the pages marked read-only
1625 * to track write events. If so, we'll downgrade vm_page_prot
1626 * to the private version (using protection_map[] without the
1627 * VM_SHARED bit).
1628 */
1629int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1630{
1631	vm_flags_t vm_flags = vma->vm_flags;
1632	const struct vm_operations_struct *vm_ops = vma->vm_ops;
1633
1634	/* If it was private or non-writable, the write bit is already clear */
1635	if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1636		return 0;
1637
1638	/* The backer wishes to know when pages are first written to? */
1639	if (vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite))
1640		return 1;
1641
1642	/* The open routine did something to the protections that pgprot_modify
1643	 * won't preserve? */
1644	if (pgprot_val(vm_page_prot) !=
1645	    pgprot_val(vm_pgprot_modify(vm_page_prot, vm_flags)))
1646		return 0;
1647
1648	/* Do we need to track softdirty? */
1649	if (IS_ENABLED(CONFIG_MEM_SOFT_DIRTY) && !(vm_flags & VM_SOFTDIRTY))
 
 
 
 
 
 
 
1650		return 1;
1651
1652	/* Specialty mapping? */
1653	if (vm_flags & VM_PFNMAP)
1654		return 0;
1655
1656	/* Can the mapping track the dirty pages? */
1657	return vma->vm_file && vma->vm_file->f_mapping &&
1658		mapping_cap_account_dirty(vma->vm_file->f_mapping);
1659}
1660
1661/*
1662 * We account for memory if it's a private writeable mapping,
1663 * not hugepages and VM_NORESERVE wasn't set.
1664 */
1665static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1666{
1667	/*
1668	 * hugetlb has its own accounting separate from the core VM
1669	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1670	 */
1671	if (file && is_file_hugepages(file))
1672		return 0;
1673
1674	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1675}
1676
1677unsigned long mmap_region(struct file *file, unsigned long addr,
1678		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
1679		struct list_head *uf)
 
 
 
 
 
 
 
 
1680{
1681	struct mm_struct *mm = current->mm;
1682	struct vm_area_struct *vma, *prev;
1683	int error;
1684	struct rb_node **rb_link, *rb_parent;
1685	unsigned long charged = 0;
1686
1687	/* Check against address space limit. */
1688	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
1689		unsigned long nr_pages;
1690
1691		/*
1692		 * MAP_FIXED may remove pages of mappings that intersects with
1693		 * requested mapping. Account for the pages it would unmap.
1694		 */
1695		nr_pages = count_vma_pages_range(mm, addr, addr + len);
1696
1697		if (!may_expand_vm(mm, vm_flags,
1698					(len >> PAGE_SHIFT) - nr_pages))
1699			return -ENOMEM;
1700	}
1701
1702	/* Clear old maps */
1703	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
1704			      &rb_parent)) {
1705		if (do_munmap(mm, addr, len, uf))
1706			return -ENOMEM;
1707	}
1708
1709	/*
1710	 * Private writable mapping: check memory availability
1711	 */
1712	if (accountable_mapping(file, vm_flags)) {
1713		charged = len >> PAGE_SHIFT;
1714		if (security_vm_enough_memory_mm(mm, charged))
1715			return -ENOMEM;
1716		vm_flags |= VM_ACCOUNT;
1717	}
1718
1719	/*
1720	 * Can we just expand an old mapping?
1721	 */
1722	vma = vma_merge(mm, prev, addr, addr + len, vm_flags,
1723			NULL, file, pgoff, NULL, NULL_VM_UFFD_CTX);
1724	if (vma)
1725		goto out;
1726
1727	/*
1728	 * Determine the object being mapped and call the appropriate
1729	 * specific mapper. the address has already been validated, but
1730	 * not unmapped, but the maps are removed from the list.
1731	 */
1732	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1733	if (!vma) {
1734		error = -ENOMEM;
1735		goto unacct_error;
1736	}
1737
1738	vma->vm_mm = mm;
1739	vma->vm_start = addr;
1740	vma->vm_end = addr + len;
1741	vma->vm_flags = vm_flags;
1742	vma->vm_page_prot = vm_get_page_prot(vm_flags);
1743	vma->vm_pgoff = pgoff;
1744	INIT_LIST_HEAD(&vma->anon_vma_chain);
1745
1746	if (file) {
1747		if (vm_flags & VM_DENYWRITE) {
1748			error = deny_write_access(file);
1749			if (error)
1750				goto free_vma;
1751		}
1752		if (vm_flags & VM_SHARED) {
1753			error = mapping_map_writable(file->f_mapping);
1754			if (error)
1755				goto allow_write_and_free_vma;
1756		}
1757
1758		/* ->mmap() can change vma->vm_file, but must guarantee that
1759		 * vma_link() below can deny write-access if VM_DENYWRITE is set
1760		 * and map writably if VM_SHARED is set. This usually means the
1761		 * new file must not have been exposed to user-space, yet.
1762		 */
1763		vma->vm_file = get_file(file);
1764		error = call_mmap(file, vma);
1765		if (error)
1766			goto unmap_and_free_vma;
1767
1768		/* Can addr have changed??
1769		 *
1770		 * Answer: Yes, several device drivers can do it in their
1771		 *         f_op->mmap method. -DaveM
1772		 * Bug: If addr is changed, prev, rb_link, rb_parent should
1773		 *      be updated for vma_link()
1774		 */
1775		WARN_ON_ONCE(addr != vma->vm_start);
1776
1777		addr = vma->vm_start;
1778		vm_flags = vma->vm_flags;
1779	} else if (vm_flags & VM_SHARED) {
1780		error = shmem_zero_setup(vma);
1781		if (error)
1782			goto free_vma;
1783	}
1784
1785	vma_link(mm, vma, prev, rb_link, rb_parent);
1786	/* Once vma denies write, undo our temporary denial count */
1787	if (file) {
1788		if (vm_flags & VM_SHARED)
1789			mapping_unmap_writable(file->f_mapping);
1790		if (vm_flags & VM_DENYWRITE)
1791			allow_write_access(file);
1792	}
1793	file = vma->vm_file;
1794out:
1795	perf_event_mmap(vma);
1796
1797	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
1798	if (vm_flags & VM_LOCKED) {
1799		if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1800					vma == get_gate_vma(current->mm)))
1801			mm->locked_vm += (len >> PAGE_SHIFT);
1802		else
1803			vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
1804	}
1805
1806	if (file)
1807		uprobe_mmap(vma);
1808
1809	/*
1810	 * New (or expanded) vma always get soft dirty status.
1811	 * Otherwise user-space soft-dirty page tracker won't
1812	 * be able to distinguish situation when vma area unmapped,
1813	 * then new mapped in-place (which must be aimed as
1814	 * a completely new data area).
1815	 */
1816	vma->vm_flags |= VM_SOFTDIRTY;
1817
1818	vma_set_page_prot(vma);
1819
1820	return addr;
1821
1822unmap_and_free_vma:
1823	vma->vm_file = NULL;
1824	fput(file);
1825
1826	/* Undo any partial mapping done by a device driver. */
1827	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1828	charged = 0;
1829	if (vm_flags & VM_SHARED)
1830		mapping_unmap_writable(file->f_mapping);
1831allow_write_and_free_vma:
1832	if (vm_flags & VM_DENYWRITE)
1833		allow_write_access(file);
1834free_vma:
1835	kmem_cache_free(vm_area_cachep, vma);
1836unacct_error:
1837	if (charged)
1838		vm_unacct_memory(charged);
1839	return error;
1840}
1841
1842unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1843{
1844	/*
1845	 * We implement the search by looking for an rbtree node that
1846	 * immediately follows a suitable gap. That is,
1847	 * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1848	 * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1849	 * - gap_end - gap_start >= length
1850	 */
1851
1852	struct mm_struct *mm = current->mm;
1853	struct vm_area_struct *vma;
1854	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1855
1856	/* Adjust search length to account for worst case alignment overhead */
1857	length = info->length + info->align_mask;
1858	if (length < info->length)
1859		return -ENOMEM;
1860
1861	/* Adjust search limits by the desired length */
1862	if (info->high_limit < length)
1863		return -ENOMEM;
1864	high_limit = info->high_limit - length;
1865
1866	if (info->low_limit > high_limit)
1867		return -ENOMEM;
1868	low_limit = info->low_limit + length;
1869
1870	/* Check if rbtree root looks promising */
1871	if (RB_EMPTY_ROOT(&mm->mm_rb))
1872		goto check_highest;
1873	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1874	if (vma->rb_subtree_gap < length)
1875		goto check_highest;
1876
1877	while (true) {
1878		/* Visit left subtree if it looks promising */
1879		gap_end = vm_start_gap(vma);
1880		if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1881			struct vm_area_struct *left =
1882				rb_entry(vma->vm_rb.rb_left,
1883					 struct vm_area_struct, vm_rb);
1884			if (left->rb_subtree_gap >= length) {
1885				vma = left;
1886				continue;
1887			}
1888		}
1889
1890		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1891check_current:
1892		/* Check if current node has a suitable gap */
1893		if (gap_start > high_limit)
1894			return -ENOMEM;
1895		if (gap_end >= low_limit &&
1896		    gap_end > gap_start && gap_end - gap_start >= length)
1897			goto found;
1898
1899		/* Visit right subtree if it looks promising */
1900		if (vma->vm_rb.rb_right) {
1901			struct vm_area_struct *right =
1902				rb_entry(vma->vm_rb.rb_right,
1903					 struct vm_area_struct, vm_rb);
1904			if (right->rb_subtree_gap >= length) {
1905				vma = right;
1906				continue;
1907			}
1908		}
1909
1910		/* Go back up the rbtree to find next candidate node */
1911		while (true) {
1912			struct rb_node *prev = &vma->vm_rb;
1913			if (!rb_parent(prev))
1914				goto check_highest;
1915			vma = rb_entry(rb_parent(prev),
1916				       struct vm_area_struct, vm_rb);
1917			if (prev == vma->vm_rb.rb_left) {
1918				gap_start = vm_end_gap(vma->vm_prev);
1919				gap_end = vm_start_gap(vma);
1920				goto check_current;
1921			}
1922		}
1923	}
1924
1925check_highest:
1926	/* Check highest gap, which does not precede any rbtree node */
1927	gap_start = mm->highest_vm_end;
1928	gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1929	if (gap_start > high_limit)
1930		return -ENOMEM;
1931
1932found:
1933	/* We found a suitable gap. Clip it with the original low_limit. */
1934	if (gap_start < info->low_limit)
1935		gap_start = info->low_limit;
1936
1937	/* Adjust gap address to the desired alignment */
1938	gap_start += (info->align_offset - gap_start) & info->align_mask;
1939
1940	VM_BUG_ON(gap_start + info->length > info->high_limit);
1941	VM_BUG_ON(gap_start + info->length > gap_end);
1942	return gap_start;
1943}
1944
1945unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
 
 
 
 
 
 
 
 
 
 
1946{
1947	struct mm_struct *mm = current->mm;
1948	struct vm_area_struct *vma;
1949	unsigned long length, low_limit, high_limit, gap_start, gap_end;
1950
 
1951	/* Adjust search length to account for worst case alignment overhead */
1952	length = info->length + info->align_mask;
1953	if (length < info->length)
1954		return -ENOMEM;
1955
1956	/*
1957	 * Adjust search limits by the desired length.
1958	 * See implementation comment at top of unmapped_area().
1959	 */
1960	gap_end = info->high_limit;
1961	if (gap_end < length)
1962		return -ENOMEM;
1963	high_limit = gap_end - length;
1964
1965	if (info->low_limit > high_limit)
1966		return -ENOMEM;
1967	low_limit = info->low_limit + length;
 
1968
1969	/* Check highest gap, which does not precede any rbtree node */
1970	gap_start = mm->highest_vm_end;
1971	if (gap_start <= high_limit)
1972		goto found_highest;
 
 
 
 
 
 
 
 
1973
1974	/* Check if rbtree root looks promising */
1975	if (RB_EMPTY_ROOT(&mm->mm_rb))
1976		return -ENOMEM;
1977	vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1978	if (vma->rb_subtree_gap < length)
1979		return -ENOMEM;
1980
1981	while (true) {
1982		/* Visit right subtree if it looks promising */
1983		gap_start = vma->vm_prev ? vm_end_gap(vma->vm_prev) : 0;
1984		if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1985			struct vm_area_struct *right =
1986				rb_entry(vma->vm_rb.rb_right,
1987					 struct vm_area_struct, vm_rb);
1988			if (right->rb_subtree_gap >= length) {
1989				vma = right;
1990				continue;
1991			}
1992		}
1993
1994check_current:
1995		/* Check if current node has a suitable gap */
1996		gap_end = vm_start_gap(vma);
1997		if (gap_end < low_limit)
1998			return -ENOMEM;
1999		if (gap_start <= high_limit &&
2000		    gap_end > gap_start && gap_end - gap_start >= length)
2001			goto found;
2002
2003		/* Visit left subtree if it looks promising */
2004		if (vma->vm_rb.rb_left) {
2005			struct vm_area_struct *left =
2006				rb_entry(vma->vm_rb.rb_left,
2007					 struct vm_area_struct, vm_rb);
2008			if (left->rb_subtree_gap >= length) {
2009				vma = left;
2010				continue;
2011			}
2012		}
2013
2014		/* Go back up the rbtree to find next candidate node */
2015		while (true) {
2016			struct rb_node *prev = &vma->vm_rb;
2017			if (!rb_parent(prev))
2018				return -ENOMEM;
2019			vma = rb_entry(rb_parent(prev),
2020				       struct vm_area_struct, vm_rb);
2021			if (prev == vma->vm_rb.rb_right) {
2022				gap_start = vma->vm_prev ?
2023					vm_end_gap(vma->vm_prev) : 0;
2024				goto check_current;
2025			}
2026		}
2027	}
2028
2029found:
2030	/* We found a suitable gap. Clip it with the original high_limit. */
2031	if (gap_end > info->high_limit)
2032		gap_end = info->high_limit;
2033
2034found_highest:
2035	/* Compute highest gap address at the desired alignment */
2036	gap_end -= info->length;
2037	gap_end -= (gap_end - info->align_offset) & info->align_mask;
2038
2039	VM_BUG_ON(gap_end < info->low_limit);
2040	VM_BUG_ON(gap_end < gap_start);
2041	return gap_end;
2042}
2043
2044/* Get an address range which is currently unmapped.
2045 * For shmat() with addr=0.
2046 *
2047 * Ugly calling convention alert:
2048 * Return value with the low bits set means error value,
2049 * ie
2050 *	if (ret & ~PAGE_MASK)
2051 *		error = ret;
2052 *
2053 * This function "knows" that -ENOMEM has the bits set.
2054 */
2055#ifndef HAVE_ARCH_UNMAPPED_AREA
2056unsigned long
2057arch_get_unmapped_area(struct file *filp, unsigned long addr,
2058		unsigned long len, unsigned long pgoff, unsigned long flags)
 
2059{
2060	struct mm_struct *mm = current->mm;
2061	struct vm_area_struct *vma, *prev;
2062	struct vm_unmapped_area_info info;
 
2063
2064	if (len > TASK_SIZE - mmap_min_addr)
2065		return -ENOMEM;
2066
2067	if (flags & MAP_FIXED)
2068		return addr;
2069
2070	if (addr) {
2071		addr = PAGE_ALIGN(addr);
2072		vma = find_vma_prev(mm, addr, &prev);
2073		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2074		    (!vma || addr + len <= vm_start_gap(vma)) &&
2075		    (!prev || addr >= vm_end_gap(prev)))
2076			return addr;
2077	}
2078
2079	info.flags = 0;
2080	info.length = len;
2081	info.low_limit = mm->mmap_base;
2082	info.high_limit = TASK_SIZE;
2083	info.align_mask = 0;
 
2084	return vm_unmapped_area(&info);
2085}
 
 
 
 
 
 
 
 
 
2086#endif
2087
2088/*
2089 * This mmap-allocator allocates new areas top-down from below the
2090 * stack's low limit (the base):
2091 */
2092#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
2093unsigned long
2094arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
2095			  const unsigned long len, const unsigned long pgoff,
2096			  const unsigned long flags)
2097{
2098	struct vm_area_struct *vma, *prev;
2099	struct mm_struct *mm = current->mm;
2100	unsigned long addr = addr0;
2101	struct vm_unmapped_area_info info;
 
2102
2103	/* requested length too big for entire address space */
2104	if (len > TASK_SIZE - mmap_min_addr)
2105		return -ENOMEM;
2106
2107	if (flags & MAP_FIXED)
2108		return addr;
2109
2110	/* requesting a specific address */
2111	if (addr) {
2112		addr = PAGE_ALIGN(addr);
2113		vma = find_vma_prev(mm, addr, &prev);
2114		if (TASK_SIZE - len >= addr && addr >= mmap_min_addr &&
2115				(!vma || addr + len <= vm_start_gap(vma)) &&
2116				(!prev || addr >= vm_end_gap(prev)))
2117			return addr;
2118	}
2119
2120	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
2121	info.length = len;
2122	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
2123	info.high_limit = mm->mmap_base;
2124	info.align_mask = 0;
 
2125	addr = vm_unmapped_area(&info);
2126
2127	/*
2128	 * A failed mmap() very likely causes application failure,
2129	 * so fall back to the bottom-up function here. This scenario
2130	 * can happen with large stack limits and large mmap()
2131	 * allocations.
2132	 */
2133	if (offset_in_page(addr)) {
2134		VM_BUG_ON(addr != -ENOMEM);
2135		info.flags = 0;
2136		info.low_limit = TASK_UNMAPPED_BASE;
2137		info.high_limit = TASK_SIZE;
2138		addr = vm_unmapped_area(&info);
2139	}
2140
2141	return addr;
2142}
 
 
 
 
 
 
 
 
 
2143#endif
2144
2145unsigned long
2146get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
2147		unsigned long pgoff, unsigned long flags)
2148{
2149	unsigned long (*get_area)(struct file *, unsigned long,
2150				  unsigned long, unsigned long, unsigned long);
2151
2152	unsigned long error = arch_mmap_check(addr, len, flags);
2153	if (error)
2154		return error;
2155
2156	/* Careful about overflows.. */
2157	if (len > TASK_SIZE)
2158		return -ENOMEM;
2159
2160	get_area = current->mm->get_unmapped_area;
2161	if (file) {
2162		if (file->f_op->get_unmapped_area)
2163			get_area = file->f_op->get_unmapped_area;
2164	} else if (flags & MAP_SHARED) {
2165		/*
2166		 * mmap_region() will call shmem_zero_setup() to create a file,
2167		 * so use shmem's get_unmapped_area in case it can be huge.
2168		 * do_mmap_pgoff() will clear pgoff, so match alignment.
2169		 */
2170		pgoff = 0;
2171		get_area = shmem_get_unmapped_area;
2172	}
2173
2174	addr = get_area(file, addr, len, pgoff, flags);
2175	if (IS_ERR_VALUE(addr))
2176		return addr;
2177
2178	if (addr > TASK_SIZE - len)
2179		return -ENOMEM;
2180	if (offset_in_page(addr))
2181		return -EINVAL;
2182
2183	error = security_mmap_addr(addr);
2184	return error ? error : addr;
2185}
2186
2187EXPORT_SYMBOL(get_unmapped_area);
2188
2189/* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
2190struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 
 
 
 
 
 
 
 
 
 
2191{
2192	struct rb_node *rb_node;
2193	struct vm_area_struct *vma;
2194
2195	/* Check the cache first. */
2196	vma = vmacache_find(mm, addr);
2197	if (likely(vma))
2198		return vma;
2199
2200	rb_node = mm->mm_rb.rb_node;
2201
2202	while (rb_node) {
2203		struct vm_area_struct *tmp;
2204
2205		tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
 
 
 
 
 
2206
2207		if (tmp->vm_end > addr) {
2208			vma = tmp;
2209			if (tmp->vm_start <= addr)
2210				break;
2211			rb_node = rb_node->rb_left;
2212		} else
2213			rb_node = rb_node->rb_right;
2214	}
2215
2216	if (vma)
2217		vmacache_update(addr, vma);
2218	return vma;
2219}
2220
2221EXPORT_SYMBOL(find_vma);
2222
2223/*
2224 * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
 
 
 
 
 
 
 
 
 
 
2225 */
2226struct vm_area_struct *
2227find_vma_prev(struct mm_struct *mm, unsigned long addr,
2228			struct vm_area_struct **pprev)
2229{
2230	struct vm_area_struct *vma;
 
2231
2232	vma = find_vma(mm, addr);
2233	if (vma) {
2234		*pprev = vma->vm_prev;
2235	} else {
2236		struct rb_node *rb_node = mm->mm_rb.rb_node;
2237		*pprev = NULL;
2238		while (rb_node) {
2239			*pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
2240			rb_node = rb_node->rb_right;
2241		}
2242	}
2243	return vma;
2244}
2245
2246/*
2247 * Verify that the stack growth is acceptable and
2248 * update accounting. This is shared with both the
2249 * grow-up and grow-down cases.
2250 */
2251static int acct_stack_growth(struct vm_area_struct *vma,
2252			     unsigned long size, unsigned long grow)
2253{
2254	struct mm_struct *mm = vma->vm_mm;
2255	unsigned long new_start;
2256
2257	/* address space limit tests */
2258	if (!may_expand_vm(mm, vma->vm_flags, grow))
2259		return -ENOMEM;
2260
2261	/* Stack limit test */
2262	if (size > rlimit(RLIMIT_STACK))
2263		return -ENOMEM;
2264
2265	/* mlock limit tests */
2266	if (vma->vm_flags & VM_LOCKED) {
2267		unsigned long locked;
2268		unsigned long limit;
2269		locked = mm->locked_vm + grow;
2270		limit = rlimit(RLIMIT_MEMLOCK);
2271		limit >>= PAGE_SHIFT;
2272		if (locked > limit && !capable(CAP_IPC_LOCK))
2273			return -ENOMEM;
2274	}
2275
2276	/* Check to ensure the stack will not grow into a hugetlb-only region */
2277	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2278			vma->vm_end - size;
2279	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2280		return -EFAULT;
2281
2282	/*
2283	 * Overcommit..  This must be the final test, as it will
2284	 * update security statistics.
2285	 */
2286	if (security_vm_enough_memory_mm(mm, grow))
2287		return -ENOMEM;
2288
2289	return 0;
2290}
2291
2292#if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2293/*
2294 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2295 * vma is the last one with address > vma->vm_end.  Have to extend vma.
2296 */
2297int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2298{
2299	struct mm_struct *mm = vma->vm_mm;
2300	struct vm_area_struct *next;
2301	unsigned long gap_addr;
2302	int error = 0;
 
2303
2304	if (!(vma->vm_flags & VM_GROWSUP))
2305		return -EFAULT;
2306
2307	/* Guard against exceeding limits of the address space. */
2308	address &= PAGE_MASK;
2309	if (address >= (TASK_SIZE & PAGE_MASK))
2310		return -ENOMEM;
2311	address += PAGE_SIZE;
2312
2313	/* Enforce stack_guard_gap */
2314	gap_addr = address + stack_guard_gap;
2315
2316	/* Guard against overflow */
2317	if (gap_addr < address || gap_addr > TASK_SIZE)
2318		gap_addr = TASK_SIZE;
2319
2320	next = vma->vm_next;
2321	if (next && next->vm_start < gap_addr &&
2322			(next->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2323		if (!(next->vm_flags & VM_GROWSUP))
2324			return -ENOMEM;
2325		/* Check that both stack segments have the same anon_vma? */
2326	}
2327
 
 
 
2328	/* We must make sure the anon_vma is allocated. */
2329	if (unlikely(anon_vma_prepare(vma)))
 
2330		return -ENOMEM;
 
2331
2332	/*
2333	 * vma->vm_start/vm_end cannot change under us because the caller
2334	 * is required to hold the mmap_sem in read mode.  We need the
2335	 * anon_vma lock to serialize against concurrent expand_stacks.
2336	 */
2337	anon_vma_lock_write(vma->anon_vma);
2338
2339	/* Somebody else might have raced and expanded it already */
2340	if (address > vma->vm_end) {
2341		unsigned long size, grow;
2342
2343		size = address - vma->vm_start;
2344		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2345
2346		error = -ENOMEM;
2347		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2348			error = acct_stack_growth(vma, size, grow);
2349			if (!error) {
2350				/*
2351				 * vma_gap_update() doesn't support concurrent
2352				 * updates, but we only hold a shared mmap_sem
2353				 * lock here, so we need to protect against
2354				 * concurrent vma expansions.
2355				 * anon_vma_lock_write() doesn't help here, as
2356				 * we don't guarantee that all growable vmas
2357				 * in a mm share the same root anon vma.
2358				 * So, we reuse mm->page_table_lock to guard
2359				 * against concurrent vma expansions.
2360				 */
2361				spin_lock(&mm->page_table_lock);
2362				if (vma->vm_flags & VM_LOCKED)
2363					mm->locked_vm += grow;
2364				vm_stat_account(mm, vma->vm_flags, grow);
2365				anon_vma_interval_tree_pre_update_vma(vma);
2366				vma->vm_end = address;
 
 
2367				anon_vma_interval_tree_post_update_vma(vma);
2368				if (vma->vm_next)
2369					vma_gap_update(vma->vm_next);
2370				else
2371					mm->highest_vm_end = vm_end_gap(vma);
2372				spin_unlock(&mm->page_table_lock);
2373
2374				perf_event_mmap(vma);
2375			}
2376		}
2377	}
2378	anon_vma_unlock_write(vma->anon_vma);
2379	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2380	validate_mm(mm);
2381	return error;
2382}
2383#endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2384
2385/*
2386 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2387 */
2388int expand_downwards(struct vm_area_struct *vma,
2389				   unsigned long address)
2390{
2391	struct mm_struct *mm = vma->vm_mm;
 
2392	struct vm_area_struct *prev;
2393	int error;
2394
2395	address &= PAGE_MASK;
2396	error = security_mmap_addr(address);
2397	if (error)
2398		return error;
2399
2400	/* Enforce stack_guard_gap */
2401	prev = vma->vm_prev;
2402	/* Check that both stack segments have the same anon_vma? */
2403	if (prev && !(prev->vm_flags & VM_GROWSDOWN) &&
2404			(prev->vm_flags & (VM_WRITE|VM_READ|VM_EXEC))) {
2405		if (address - prev->vm_end < stack_guard_gap)
2406			return -ENOMEM;
2407	}
2408
 
 
 
2409	/* We must make sure the anon_vma is allocated. */
2410	if (unlikely(anon_vma_prepare(vma)))
 
2411		return -ENOMEM;
 
2412
2413	/*
2414	 * vma->vm_start/vm_end cannot change under us because the caller
2415	 * is required to hold the mmap_sem in read mode.  We need the
2416	 * anon_vma lock to serialize against concurrent expand_stacks.
2417	 */
2418	anon_vma_lock_write(vma->anon_vma);
2419
2420	/* Somebody else might have raced and expanded it already */
2421	if (address < vma->vm_start) {
2422		unsigned long size, grow;
2423
2424		size = vma->vm_end - address;
2425		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2426
2427		error = -ENOMEM;
2428		if (grow <= vma->vm_pgoff) {
2429			error = acct_stack_growth(vma, size, grow);
2430			if (!error) {
2431				/*
2432				 * vma_gap_update() doesn't support concurrent
2433				 * updates, but we only hold a shared mmap_sem
2434				 * lock here, so we need to protect against
2435				 * concurrent vma expansions.
2436				 * anon_vma_lock_write() doesn't help here, as
2437				 * we don't guarantee that all growable vmas
2438				 * in a mm share the same root anon vma.
2439				 * So, we reuse mm->page_table_lock to guard
2440				 * against concurrent vma expansions.
2441				 */
2442				spin_lock(&mm->page_table_lock);
2443				if (vma->vm_flags & VM_LOCKED)
2444					mm->locked_vm += grow;
2445				vm_stat_account(mm, vma->vm_flags, grow);
2446				anon_vma_interval_tree_pre_update_vma(vma);
2447				vma->vm_start = address;
2448				vma->vm_pgoff -= grow;
 
 
2449				anon_vma_interval_tree_post_update_vma(vma);
2450				vma_gap_update(vma);
2451				spin_unlock(&mm->page_table_lock);
2452
2453				perf_event_mmap(vma);
2454			}
2455		}
2456	}
2457	anon_vma_unlock_write(vma->anon_vma);
2458	khugepaged_enter_vma_merge(vma, vma->vm_flags);
2459	validate_mm(mm);
2460	return error;
2461}
2462
2463/* enforced gap between the expanding stack and other mappings. */
2464unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2465
2466static int __init cmdline_parse_stack_guard_gap(char *p)
2467{
2468	unsigned long val;
2469	char *endptr;
2470
2471	val = simple_strtoul(p, &endptr, 10);
2472	if (!*endptr)
2473		stack_guard_gap = val << PAGE_SHIFT;
2474
2475	return 0;
2476}
2477__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2478
2479#ifdef CONFIG_STACK_GROWSUP
2480int expand_stack(struct vm_area_struct *vma, unsigned long address)
2481{
2482	return expand_upwards(vma, address);
2483}
2484
2485struct vm_area_struct *
2486find_extend_vma(struct mm_struct *mm, unsigned long addr)
2487{
2488	struct vm_area_struct *vma, *prev;
2489
2490	addr &= PAGE_MASK;
2491	vma = find_vma_prev(mm, addr, &prev);
2492	if (vma && (vma->vm_start <= addr))
2493		return vma;
2494	if (!prev || expand_stack(prev, addr))
2495		return NULL;
2496	if (prev->vm_flags & VM_LOCKED)
2497		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2498	return prev;
2499}
2500#else
2501int expand_stack(struct vm_area_struct *vma, unsigned long address)
2502{
2503	return expand_downwards(vma, address);
2504}
2505
2506struct vm_area_struct *
2507find_extend_vma(struct mm_struct *mm, unsigned long addr)
2508{
2509	struct vm_area_struct *vma;
2510	unsigned long start;
2511
2512	addr &= PAGE_MASK;
2513	vma = find_vma(mm, addr);
2514	if (!vma)
2515		return NULL;
2516	if (vma->vm_start <= addr)
2517		return vma;
2518	if (!(vma->vm_flags & VM_GROWSDOWN))
2519		return NULL;
2520	start = vma->vm_start;
2521	if (expand_stack(vma, addr))
2522		return NULL;
2523	if (vma->vm_flags & VM_LOCKED)
2524		populate_vma_page_range(vma, addr, start, NULL);
2525	return vma;
2526}
2527#endif
2528
2529EXPORT_SYMBOL_GPL(find_extend_vma);
2530
2531/*
2532 * Ok - we have the memory areas we should free on the vma list,
2533 * so release them, and do the vma updates.
2534 *
2535 * Called with the mm semaphore held.
2536 */
2537static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2538{
2539	unsigned long nr_accounted = 0;
 
2540
2541	/* Update high watermark before we lower total_vm */
2542	update_hiwater_vm(mm);
2543	do {
2544		long nrpages = vma_pages(vma);
2545
2546		if (vma->vm_flags & VM_ACCOUNT)
2547			nr_accounted += nrpages;
2548		vm_stat_account(mm, vma->vm_flags, -nrpages);
2549		vma = remove_vma(vma);
2550	} while (vma);
2551	vm_unacct_memory(nr_accounted);
2552	validate_mm(mm);
2553}
2554
2555/*
2556 * Get rid of page table information in the indicated region.
2557 *
2558 * Called with the mm semaphore held.
2559 */
2560static void unmap_region(struct mm_struct *mm,
2561		struct vm_area_struct *vma, struct vm_area_struct *prev,
 
2562		unsigned long start, unsigned long end)
2563{
2564	struct vm_area_struct *next = prev ? prev->vm_next : mm->mmap;
2565	struct mmu_gather tlb;
2566
2567	lru_add_drain();
2568	tlb_gather_mmu(&tlb, mm, start, end);
2569	update_hiwater_rss(mm);
2570	unmap_vmas(&tlb, vma, start, end);
2571	free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2572				 next ? next->vm_start : USER_PGTABLES_CEILING);
2573	tlb_finish_mmu(&tlb, start, end);
2574}
2575
2576/*
2577 * Create a list of vma's touched by the unmap, removing them from the mm's
2578 * vma list as we go..
2579 */
2580static void
2581detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2582	struct vm_area_struct *prev, unsigned long end)
2583{
2584	struct vm_area_struct **insertion_point;
2585	struct vm_area_struct *tail_vma = NULL;
2586
2587	insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2588	vma->vm_prev = NULL;
2589	do {
2590		vma_rb_erase(vma, &mm->mm_rb);
2591		mm->map_count--;
2592		tail_vma = vma;
2593		vma = vma->vm_next;
2594	} while (vma && vma->vm_start < end);
2595	*insertion_point = vma;
2596	if (vma) {
2597		vma->vm_prev = prev;
2598		vma_gap_update(vma);
2599	} else
2600		mm->highest_vm_end = prev ? vm_end_gap(prev) : 0;
2601	tail_vma->vm_next = NULL;
2602
2603	/* Kill the cache */
2604	vmacache_invalidate(mm);
2605}
2606
2607/*
2608 * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2609 * has already been checked or doesn't make sense to fail.
2610 */
2611int __split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2612		unsigned long addr, int new_below)
2613{
2614	struct vm_area_struct *new;
2615	int err;
 
2616
2617	if (vma->vm_ops && vma->vm_ops->split) {
2618		err = vma->vm_ops->split(vma, addr);
2619		if (err)
2620			return err;
2621	}
2622
2623	new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2624	if (!new)
2625		return -ENOMEM;
2626
2627	/* most fields are the same, copy all, and then fixup */
2628	*new = *vma;
2629
2630	INIT_LIST_HEAD(&new->anon_vma_chain);
2631
2632	if (new_below)
2633		new->vm_end = addr;
2634	else {
2635		new->vm_start = addr;
2636		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2637	}
2638
2639	err = vma_dup_policy(vma, new);
2640	if (err)
2641		goto out_free_vma;
2642
2643	err = anon_vma_clone(new, vma);
2644	if (err)
2645		goto out_free_mpol;
2646
2647	if (new->vm_file)
2648		get_file(new->vm_file);
2649
2650	if (new->vm_ops && new->vm_ops->open)
2651		new->vm_ops->open(new);
2652
2653	if (new_below)
2654		err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2655			((addr - new->vm_start) >> PAGE_SHIFT), new);
2656	else
2657		err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2658
2659	/* Success. */
2660	if (!err)
2661		return 0;
2662
 
 
 
2663	/* Clean everything up if vma_adjust failed. */
2664	if (new->vm_ops && new->vm_ops->close)
2665		new->vm_ops->close(new);
2666	if (new->vm_file)
2667		fput(new->vm_file);
2668	unlink_anon_vmas(new);
2669 out_free_mpol:
2670	mpol_put(vma_policy(new));
2671 out_free_vma:
2672	kmem_cache_free(vm_area_cachep, new);
 
2673	return err;
2674}
2675
2676/*
2677 * Split a vma into two pieces at address 'addr', a new vma is allocated
2678 * either for the first part or the tail.
2679 */
2680int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2681	      unsigned long addr, int new_below)
2682{
2683	if (mm->map_count >= sysctl_max_map_count)
2684		return -ENOMEM;
2685
2686	return __split_vma(mm, vma, addr, new_below);
2687}
2688
2689/* Munmap is split into 2 main parts -- this part which finds
2690 * what needs doing, and the areas themselves, which do the
2691 * work.  This now handles partial unmappings.
2692 * Jeremy Fitzhardinge <jeremy@goop.org>
2693 */
2694int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2695	      struct list_head *uf)
2696{
2697	unsigned long end;
2698	struct vm_area_struct *vma, *prev, *last;
 
2699
2700	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2701		return -EINVAL;
2702
2703	len = PAGE_ALIGN(len);
2704	if (len == 0)
2705		return -EINVAL;
2706
2707	/* Find the first overlapping VMA */
2708	vma = find_vma(mm, start);
2709	if (!vma)
2710		return 0;
2711	prev = vma->vm_prev;
2712	/* we have  start < vma->vm_end  */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2713
2714	/* if it doesn't overlap, we have nothing.. */
2715	end = start + len;
2716	if (vma->vm_start >= end)
2717		return 0;
2718
 
2719	/*
2720	 * If we need to split any vma, do it now to save pain later.
2721	 *
2722	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2723	 * unmapped vm_area_struct will remain in use: so lower split_vma
2724	 * places tmp vma above, and higher split_vma places tmp vma below.
2725	 */
 
 
2726	if (start > vma->vm_start) {
2727		int error;
2728
2729		/*
2730		 * Make sure that map_count on return from munmap() will
2731		 * not exceed its limit; but let map_count go just above
2732		 * its limit temporarily, to help free resources as expected.
2733		 */
2734		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2735			return -ENOMEM;
2736
 
 
 
 
2737		error = __split_vma(mm, vma, start, 0);
2738		if (error)
2739			return error;
2740		prev = vma;
 
 
2741	}
2742
2743	/* Does it split the last one? */
2744	last = find_vma(mm, end);
2745	if (last && end > last->vm_start) {
2746		int error = __split_vma(mm, last, end, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2747		if (error)
2748			return error;
 
 
 
 
 
 
2749	}
2750	vma = prev ? prev->vm_next : mm->mmap;
 
 
2751
2752	if (unlikely(uf)) {
2753		/*
2754		 * If userfaultfd_unmap_prep returns an error the vmas
2755		 * will remain splitted, but userland will get a
2756		 * highly unexpected error anyway. This is no
2757		 * different than the case where the first of the two
2758		 * __split_vma fails, but we don't undo the first
2759		 * split, despite we could. This is unlikely enough
2760		 * failure that it's not worth optimizing it for.
2761		 */
2762		int error = userfaultfd_unmap_prep(vma, start, end, uf);
 
2763		if (error)
2764			return error;
2765	}
2766
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2767	/*
2768	 * unlock any mlock()ed ranges before detaching vmas
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2769	 */
2770	if (mm->locked_vm) {
2771		struct vm_area_struct *tmp = vma;
2772		while (tmp && tmp->vm_start < end) {
2773			if (tmp->vm_flags & VM_LOCKED) {
2774				mm->locked_vm -= vma_pages(tmp);
2775				munlock_vma_pages_all(tmp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2776			}
2777			tmp = tmp->vm_next;
2778		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2779	}
2780
2781	/*
2782	 * Remove the vma's, and unmap the actual pages
 
2783	 */
2784	detach_vmas_to_be_unmapped(mm, vma, prev, end);
2785	unmap_region(mm, vma, prev, start, end);
2786
2787	arch_unmap(mm, vma, start, end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2788
2789	/* Fix up all other VM information */
2790	remove_vma_list(mm, vma);
2791
2792	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2793}
2794
2795int vm_munmap(unsigned long start, size_t len)
2796{
2797	int ret;
2798	struct mm_struct *mm = current->mm;
2799	LIST_HEAD(uf);
 
2800
2801	if (down_write_killable(&mm->mmap_sem))
2802		return -EINTR;
2803
2804	ret = do_munmap(mm, start, len, &uf);
2805	up_write(&mm->mmap_sem);
 
 
 
 
 
 
 
 
 
 
2806	userfaultfd_unmap_complete(mm, &uf);
2807	return ret;
2808}
 
 
 
 
 
2809EXPORT_SYMBOL(vm_munmap);
2810
2811SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2812{
2813	profile_munmap(addr);
2814	return vm_munmap(addr, len);
2815}
2816
2817
2818/*
2819 * Emulation of deprecated remap_file_pages() syscall.
2820 */
2821SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2822		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2823{
2824
2825	struct mm_struct *mm = current->mm;
2826	struct vm_area_struct *vma;
2827	unsigned long populate = 0;
2828	unsigned long ret = -EINVAL;
2829	struct file *file;
2830
2831	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/vm/remap_file_pages.txt.\n",
2832		     current->comm, current->pid);
2833
2834	if (prot)
2835		return ret;
2836	start = start & PAGE_MASK;
2837	size = size & PAGE_MASK;
2838
2839	if (start + size <= start)
2840		return ret;
2841
2842	/* Does pgoff wrap? */
2843	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
2844		return ret;
2845
2846	if (down_write_killable(&mm->mmap_sem))
2847		return -EINTR;
2848
2849	vma = find_vma(mm, start);
2850
2851	if (!vma || !(vma->vm_flags & VM_SHARED))
2852		goto out;
2853
2854	if (start < vma->vm_start)
2855		goto out;
2856
2857	if (start + size > vma->vm_end) {
2858		struct vm_area_struct *next;
 
2859
2860		for (next = vma->vm_next; next; next = next->vm_next) {
2861			/* hole between vmas ? */
2862			if (next->vm_start != next->vm_prev->vm_end)
2863				goto out;
2864
2865			if (next->vm_file != vma->vm_file)
2866				goto out;
2867
2868			if (next->vm_flags != vma->vm_flags)
2869				goto out;
2870
2871			if (start + size <= next->vm_end)
2872				break;
 
 
2873		}
2874
2875		if (!next)
2876			goto out;
2877	}
2878
2879	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
2880	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
2881	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
2882
2883	flags &= MAP_NONBLOCK;
2884	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
2885	if (vma->vm_flags & VM_LOCKED) {
2886		struct vm_area_struct *tmp;
2887		flags |= MAP_LOCKED;
2888
2889		/* drop PG_Mlocked flag for over-mapped range */
2890		for (tmp = vma; tmp->vm_start >= start + size;
2891				tmp = tmp->vm_next) {
2892			/*
2893			 * Split pmd and munlock page on the border
2894			 * of the range.
2895			 */
2896			vma_adjust_trans_huge(tmp, start, start + size, 0);
2897
2898			munlock_vma_pages_range(tmp,
2899					max(tmp->vm_start, start),
2900					min(tmp->vm_end, start + size));
2901		}
2902	}
2903
2904	file = get_file(vma->vm_file);
2905	ret = do_mmap_pgoff(vma->vm_file, start, size,
2906			prot, flags, pgoff, &populate, NULL);
2907	fput(file);
2908out:
2909	up_write(&mm->mmap_sem);
2910	if (populate)
2911		mm_populate(ret, populate);
2912	if (!IS_ERR_VALUE(ret))
2913		ret = 0;
2914	return ret;
2915}
2916
2917static inline void verify_mm_writelocked(struct mm_struct *mm)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2918{
2919#ifdef CONFIG_DEBUG_VM
2920	if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2921		WARN_ON(1);
2922		up_read(&mm->mmap_sem);
2923	}
2924#endif
 
2925}
2926
2927/*
2928 *  this is really a simplified "do_mmap".  it only handles
2929 *  anonymous maps.  eventually we may be able to do some
2930 *  brk-specific accounting here.
 
 
 
 
 
 
 
2931 */
2932static int do_brk_flags(unsigned long addr, unsigned long request, unsigned long flags, struct list_head *uf)
 
2933{
2934	struct mm_struct *mm = current->mm;
2935	struct vm_area_struct *vma, *prev;
2936	unsigned long len;
2937	struct rb_node **rb_link, *rb_parent;
2938	pgoff_t pgoff = addr >> PAGE_SHIFT;
2939	int error;
2940
2941	len = PAGE_ALIGN(request);
2942	if (len < request)
2943		return -ENOMEM;
2944	if (!len)
2945		return 0;
2946
2947	/* Until we need other flags, refuse anything except VM_EXEC. */
2948	if ((flags & (~VM_EXEC)) != 0)
2949		return -EINVAL;
2950	flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2951
2952	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2953	if (offset_in_page(error))
2954		return error;
2955
2956	error = mlock_future_check(mm, mm->def_flags, len);
2957	if (error)
2958		return error;
2959
 
2960	/*
2961	 * mm->mmap_sem is required to protect against another thread
2962	 * changing the mappings in case we sleep.
2963	 */
2964	verify_mm_writelocked(mm);
2965
2966	/*
2967	 * Clear old maps.  this also does some error checking for us
2968	 */
2969	while (find_vma_links(mm, addr, addr + len, &prev, &rb_link,
2970			      &rb_parent)) {
2971		if (do_munmap(mm, addr, len, uf))
2972			return -ENOMEM;
2973	}
2974
2975	/* Check against address space limits *after* clearing old maps... */
2976	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
2977		return -ENOMEM;
2978
2979	if (mm->map_count > sysctl_max_map_count)
2980		return -ENOMEM;
2981
2982	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2983		return -ENOMEM;
2984
2985	/* Can we just expand an old private anonymous mapping? */
2986	vma = vma_merge(mm, prev, addr, addr + len, flags,
2987			NULL, NULL, pgoff, NULL, NULL_VM_UFFD_CTX);
2988	if (vma)
2989		goto out;
2990
2991	/*
2992	 * create a vma struct for an anonymous mapping
 
2993	 */
2994	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2995	if (!vma) {
2996		vm_unacct_memory(len >> PAGE_SHIFT);
2997		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2998	}
2999
3000	INIT_LIST_HEAD(&vma->anon_vma_chain);
3001	vma->vm_mm = mm;
 
 
 
 
3002	vma->vm_start = addr;
3003	vma->vm_end = addr + len;
3004	vma->vm_pgoff = pgoff;
3005	vma->vm_flags = flags;
3006	vma->vm_page_prot = vm_get_page_prot(flags);
3007	vma_link(mm, vma, prev, rb_link, rb_parent);
 
 
 
 
3008out:
3009	perf_event_mmap(vma);
3010	mm->total_vm += len >> PAGE_SHIFT;
3011	mm->data_vm += len >> PAGE_SHIFT;
3012	if (flags & VM_LOCKED)
3013		mm->locked_vm += (len >> PAGE_SHIFT);
3014	vma->vm_flags |= VM_SOFTDIRTY;
 
3015	return 0;
3016}
3017
3018static int do_brk(unsigned long addr, unsigned long len, struct list_head *uf)
3019{
3020	return do_brk_flags(addr, len, 0, uf);
 
 
3021}
3022
3023int vm_brk_flags(unsigned long addr, unsigned long len, unsigned long flags)
3024{
3025	struct mm_struct *mm = current->mm;
 
 
3026	int ret;
3027	bool populate;
3028	LIST_HEAD(uf);
 
 
 
 
 
 
 
3029
3030	if (down_write_killable(&mm->mmap_sem))
3031		return -EINTR;
3032
3033	ret = do_brk_flags(addr, len, flags, &uf);
 
 
 
 
 
 
 
 
 
 
 
 
 
3034	populate = ((mm->def_flags & VM_LOCKED) != 0);
3035	up_write(&mm->mmap_sem);
3036	userfaultfd_unmap_complete(mm, &uf);
3037	if (populate && !ret)
3038		mm_populate(addr, len);
3039	return ret;
 
 
 
 
 
3040}
3041EXPORT_SYMBOL(vm_brk_flags);
3042
3043int vm_brk(unsigned long addr, unsigned long len)
3044{
3045	return vm_brk_flags(addr, len, 0);
3046}
3047EXPORT_SYMBOL(vm_brk);
3048
3049/* Release all mmaps. */
3050void exit_mmap(struct mm_struct *mm)
3051{
3052	struct mmu_gather tlb;
3053	struct vm_area_struct *vma;
3054	unsigned long nr_accounted = 0;
 
 
3055
3056	/* mm's last user has gone, and its about to be pulled down */
3057	mmu_notifier_release(mm);
3058
3059	if (unlikely(mm_is_oom_victim(mm))) {
3060		/*
3061		 * Manually reap the mm to free as much memory as possible.
3062		 * Then, as the oom reaper does, set MMF_OOM_SKIP to disregard
3063		 * this mm from further consideration.  Taking mm->mmap_sem for
3064		 * write after setting MMF_OOM_SKIP will guarantee that the oom
3065		 * reaper will not run on this mm again after mmap_sem is
3066		 * dropped.
3067		 *
3068		 * Nothing can be holding mm->mmap_sem here and the above call
3069		 * to mmu_notifier_release(mm) ensures mmu notifier callbacks in
3070		 * __oom_reap_task_mm() will not block.
3071		 *
3072		 * This needs to be done before calling munlock_vma_pages_all(),
3073		 * which clears VM_LOCKED, otherwise the oom reaper cannot
3074		 * reliably test it.
3075		 */
3076		mutex_lock(&oom_lock);
3077		__oom_reap_task_mm(mm);
3078		mutex_unlock(&oom_lock);
3079
3080		set_bit(MMF_OOM_SKIP, &mm->flags);
3081		down_write(&mm->mmap_sem);
3082		up_write(&mm->mmap_sem);
3083	}
3084
3085	if (mm->locked_vm) {
3086		vma = mm->mmap;
3087		while (vma) {
3088			if (vma->vm_flags & VM_LOCKED)
3089				munlock_vma_pages_all(vma);
3090			vma = vma->vm_next;
3091		}
3092	}
3093
3094	arch_exit_mmap(mm);
3095
3096	vma = mm->mmap;
3097	if (!vma)	/* Can happen if dup_mmap() received an OOM */
 
 
3098		return;
 
3099
3100	lru_add_drain();
3101	flush_cache_mm(mm);
3102	tlb_gather_mmu(&tlb, mm, 0, -1);
3103	/* update_hiwater_rss(mm) here? but nobody should be looking */
3104	/* Use -1 here to ensure all VMAs in the mm are unmapped */
3105	unmap_vmas(&tlb, vma, 0, -1);
3106	free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
3107	tlb_finish_mmu(&tlb, 0, -1);
 
 
 
 
 
 
 
 
 
3108
3109	/*
3110	 * Walk the list again, actually closing and freeing it,
3111	 * with preemption enabled, without holding any MM locks.
 
3112	 */
3113	while (vma) {
3114		if (vma->vm_flags & VM_ACCOUNT)
3115			nr_accounted += vma_pages(vma);
3116		vma = remove_vma(vma);
3117	}
 
 
 
 
 
 
 
 
3118	vm_unacct_memory(nr_accounted);
3119}
3120
3121/* Insert vm structure into process list sorted by address
3122 * and into the inode's i_mmap tree.  If vm_file is non-NULL
3123 * then i_mmap_rwsem is taken here.
3124 */
3125int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3126{
3127	struct vm_area_struct *prev;
3128	struct rb_node **rb_link, *rb_parent;
3129
3130	if (find_vma_links(mm, vma->vm_start, vma->vm_end,
3131			   &prev, &rb_link, &rb_parent))
3132		return -ENOMEM;
 
3133	if ((vma->vm_flags & VM_ACCOUNT) &&
3134	     security_vm_enough_memory_mm(mm, vma_pages(vma)))
3135		return -ENOMEM;
3136
3137	/*
3138	 * The vm_pgoff of a purely anonymous vma should be irrelevant
3139	 * until its first write fault, when page's anon_vma and index
3140	 * are set.  But now set the vm_pgoff it will almost certainly
3141	 * end up with (unless mremap moves it elsewhere before that
3142	 * first wfault), so /proc/pid/maps tells a consistent story.
3143	 *
3144	 * By setting it to reflect the virtual start address of the
3145	 * vma, merges and splits can happen in a seamless way, just
3146	 * using the existing file pgoff checks and manipulations.
3147	 * Similarly in do_mmap_pgoff and in do_brk.
3148	 */
3149	if (vma_is_anonymous(vma)) {
3150		BUG_ON(vma->anon_vma);
3151		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3152	}
3153
3154	vma_link(mm, vma, prev, rb_link, rb_parent);
 
 
 
 
3155	return 0;
3156}
3157
3158/*
3159 * Copy the vma structure to a new location in the same mm,
3160 * prior to moving page table entries, to effect an mremap move.
3161 */
3162struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3163	unsigned long addr, unsigned long len, pgoff_t pgoff,
3164	bool *need_rmap_locks)
3165{
3166	struct vm_area_struct *vma = *vmap;
3167	unsigned long vma_start = vma->vm_start;
3168	struct mm_struct *mm = vma->vm_mm;
3169	struct vm_area_struct *new_vma, *prev;
3170	struct rb_node **rb_link, *rb_parent;
3171	bool faulted_in_anon_vma = true;
3172
 
3173	/*
3174	 * If anonymous vma has not yet been faulted, update new pgoff
3175	 * to match new location, to increase its chance of merging.
3176	 */
3177	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3178		pgoff = addr >> PAGE_SHIFT;
3179		faulted_in_anon_vma = false;
3180	}
3181
3182	if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
 
3183		return NULL;	/* should never get here */
 
3184	new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
3185			    vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
3186			    vma->vm_userfaultfd_ctx);
3187	if (new_vma) {
3188		/*
3189		 * Source vma may have been merged into new_vma
3190		 */
3191		if (unlikely(vma_start >= new_vma->vm_start &&
3192			     vma_start < new_vma->vm_end)) {
3193			/*
3194			 * The only way we can get a vma_merge with
3195			 * self during an mremap is if the vma hasn't
3196			 * been faulted in yet and we were allowed to
3197			 * reset the dst vma->vm_pgoff to the
3198			 * destination address of the mremap to allow
3199			 * the merge to happen. mremap must change the
3200			 * vm_pgoff linearity between src and dst vmas
3201			 * (in turn preventing a vma_merge) to be
3202			 * safe. It is only safe to keep the vm_pgoff
3203			 * linear if there are no pages mapped yet.
3204			 */
3205			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3206			*vmap = vma = new_vma;
3207		}
3208		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3209	} else {
3210		new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
3211		if (!new_vma)
3212			goto out;
3213		*new_vma = *vma;
3214		new_vma->vm_start = addr;
3215		new_vma->vm_end = addr + len;
3216		new_vma->vm_pgoff = pgoff;
3217		if (vma_dup_policy(vma, new_vma))
3218			goto out_free_vma;
3219		INIT_LIST_HEAD(&new_vma->anon_vma_chain);
3220		if (anon_vma_clone(new_vma, vma))
3221			goto out_free_mempol;
3222		if (new_vma->vm_file)
3223			get_file(new_vma->vm_file);
3224		if (new_vma->vm_ops && new_vma->vm_ops->open)
3225			new_vma->vm_ops->open(new_vma);
3226		vma_link(mm, new_vma, prev, rb_link, rb_parent);
 
3227		*need_rmap_locks = false;
3228	}
 
3229	return new_vma;
3230
 
 
 
 
 
 
 
 
3231out_free_mempol:
3232	mpol_put(vma_policy(new_vma));
3233out_free_vma:
3234	kmem_cache_free(vm_area_cachep, new_vma);
3235out:
 
3236	return NULL;
3237}
3238
3239/*
3240 * Return true if the calling process may expand its vm space by the passed
3241 * number of pages
3242 */
3243bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3244{
3245	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3246		return false;
3247
3248	if (is_data_mapping(flags) &&
3249	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3250		/* Workaround for Valgrind */
3251		if (rlimit(RLIMIT_DATA) == 0 &&
3252		    mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3253			return true;
3254
3255		pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3256			     current->comm, current->pid,
3257			     (mm->data_vm + npages) << PAGE_SHIFT,
3258			     rlimit(RLIMIT_DATA),
3259			     ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3260
3261		if (!ignore_rlimit_data)
3262			return false;
3263	}
3264
3265	return true;
3266}
3267
3268void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3269{
3270	mm->total_vm += npages;
3271
3272	if (is_exec_mapping(flags))
3273		mm->exec_vm += npages;
3274	else if (is_stack_mapping(flags))
3275		mm->stack_vm += npages;
3276	else if (is_data_mapping(flags))
3277		mm->data_vm += npages;
3278}
3279
3280static int special_mapping_fault(struct vm_fault *vmf);
3281
3282/*
3283 * Having a close hook prevents vma merging regardless of flags.
3284 */
3285static void special_mapping_close(struct vm_area_struct *vma)
3286{
3287}
3288
3289static const char *special_mapping_name(struct vm_area_struct *vma)
3290{
3291	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3292}
3293
3294static int special_mapping_mremap(struct vm_area_struct *new_vma)
3295{
3296	struct vm_special_mapping *sm = new_vma->vm_private_data;
3297
3298	if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3299		return -EFAULT;
3300
3301	if (sm->mremap)
3302		return sm->mremap(sm, new_vma);
3303
3304	return 0;
3305}
3306
 
 
 
 
 
 
 
 
 
 
 
3307static const struct vm_operations_struct special_mapping_vmops = {
3308	.close = special_mapping_close,
3309	.fault = special_mapping_fault,
3310	.mremap = special_mapping_mremap,
3311	.name = special_mapping_name,
 
 
 
3312};
3313
3314static const struct vm_operations_struct legacy_special_mapping_vmops = {
3315	.close = special_mapping_close,
3316	.fault = special_mapping_fault,
3317};
3318
3319static int special_mapping_fault(struct vm_fault *vmf)
3320{
3321	struct vm_area_struct *vma = vmf->vma;
3322	pgoff_t pgoff;
3323	struct page **pages;
3324
3325	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3326		pages = vma->vm_private_data;
3327	} else {
3328		struct vm_special_mapping *sm = vma->vm_private_data;
3329
3330		if (sm->fault)
3331			return sm->fault(sm, vmf->vma, vmf);
3332
3333		pages = sm->pages;
3334	}
3335
3336	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3337		pgoff--;
3338
3339	if (*pages) {
3340		struct page *page = *pages;
3341		get_page(page);
3342		vmf->page = page;
3343		return 0;
3344	}
3345
3346	return VM_FAULT_SIGBUS;
3347}
3348
3349static struct vm_area_struct *__install_special_mapping(
3350	struct mm_struct *mm,
3351	unsigned long addr, unsigned long len,
3352	unsigned long vm_flags, void *priv,
3353	const struct vm_operations_struct *ops)
3354{
3355	int ret;
3356	struct vm_area_struct *vma;
3357
3358	vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
 
3359	if (unlikely(vma == NULL))
3360		return ERR_PTR(-ENOMEM);
3361
3362	INIT_LIST_HEAD(&vma->anon_vma_chain);
3363	vma->vm_mm = mm;
3364	vma->vm_start = addr;
3365	vma->vm_end = addr + len;
3366
3367	vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND | VM_SOFTDIRTY;
 
3368	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3369
3370	vma->vm_ops = ops;
3371	vma->vm_private_data = priv;
3372
3373	ret = insert_vm_struct(mm, vma);
3374	if (ret)
3375		goto out;
3376
3377	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3378
3379	perf_event_mmap(vma);
3380
 
3381	return vma;
3382
3383out:
3384	kmem_cache_free(vm_area_cachep, vma);
 
3385	return ERR_PTR(ret);
3386}
3387
3388bool vma_is_special_mapping(const struct vm_area_struct *vma,
3389	const struct vm_special_mapping *sm)
3390{
3391	return vma->vm_private_data == sm &&
3392		(vma->vm_ops == &special_mapping_vmops ||
3393		 vma->vm_ops == &legacy_special_mapping_vmops);
3394}
3395
3396/*
3397 * Called with mm->mmap_sem held for writing.
3398 * Insert a new vma covering the given region, with the given flags.
3399 * Its pages are supplied by the given array of struct page *.
3400 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3401 * The region past the last page supplied will always produce SIGBUS.
3402 * The array pointer and the pages it points to are assumed to stay alive
3403 * for as long as this mapping might exist.
3404 */
3405struct vm_area_struct *_install_special_mapping(
3406	struct mm_struct *mm,
3407	unsigned long addr, unsigned long len,
3408	unsigned long vm_flags, const struct vm_special_mapping *spec)
3409{
3410	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3411					&special_mapping_vmops);
3412}
3413
3414int install_special_mapping(struct mm_struct *mm,
3415			    unsigned long addr, unsigned long len,
3416			    unsigned long vm_flags, struct page **pages)
3417{
3418	struct vm_area_struct *vma = __install_special_mapping(
3419		mm, addr, len, vm_flags, (void *)pages,
3420		&legacy_special_mapping_vmops);
3421
3422	return PTR_ERR_OR_ZERO(vma);
3423}
3424
3425static DEFINE_MUTEX(mm_all_locks_mutex);
3426
3427static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3428{
3429	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3430		/*
3431		 * The LSB of head.next can't change from under us
3432		 * because we hold the mm_all_locks_mutex.
3433		 */
3434		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
3435		/*
3436		 * We can safely modify head.next after taking the
3437		 * anon_vma->root->rwsem. If some other vma in this mm shares
3438		 * the same anon_vma we won't take it again.
3439		 *
3440		 * No need of atomic instructions here, head.next
3441		 * can't change from under us thanks to the
3442		 * anon_vma->root->rwsem.
3443		 */
3444		if (__test_and_set_bit(0, (unsigned long *)
3445				       &anon_vma->root->rb_root.rb_root.rb_node))
3446			BUG();
3447	}
3448}
3449
3450static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3451{
3452	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3453		/*
3454		 * AS_MM_ALL_LOCKS can't change from under us because
3455		 * we hold the mm_all_locks_mutex.
3456		 *
3457		 * Operations on ->flags have to be atomic because
3458		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3459		 * mm_all_locks_mutex, there may be other cpus
3460		 * changing other bitflags in parallel to us.
3461		 */
3462		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3463			BUG();
3464		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_sem);
3465	}
3466}
3467
3468/*
3469 * This operation locks against the VM for all pte/vma/mm related
3470 * operations that could ever happen on a certain mm. This includes
3471 * vmtruncate, try_to_unmap, and all page faults.
3472 *
3473 * The caller must take the mmap_sem in write mode before calling
3474 * mm_take_all_locks(). The caller isn't allowed to release the
3475 * mmap_sem until mm_drop_all_locks() returns.
3476 *
3477 * mmap_sem in write mode is required in order to block all operations
3478 * that could modify pagetables and free pages without need of
3479 * altering the vma layout. It's also needed in write mode to avoid new
3480 * anon_vmas to be associated with existing vmas.
3481 *
3482 * A single task can't take more than one mm_take_all_locks() in a row
3483 * or it would deadlock.
3484 *
3485 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3486 * mapping->flags avoid to take the same lock twice, if more than one
3487 * vma in this mm is backed by the same anon_vma or address_space.
3488 *
3489 * We take locks in following order, accordingly to comment at beginning
3490 * of mm/rmap.c:
3491 *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3492 *     hugetlb mapping);
3493 *   - all i_mmap_rwsem locks;
3494 *   - all anon_vma->rwseml
3495 *
3496 * We can take all locks within these types randomly because the VM code
3497 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3498 * mm_all_locks_mutex.
3499 *
3500 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3501 * that may have to take thousand of locks.
3502 *
3503 * mm_take_all_locks() can fail if it's interrupted by signals.
3504 */
3505int mm_take_all_locks(struct mm_struct *mm)
3506{
3507	struct vm_area_struct *vma;
3508	struct anon_vma_chain *avc;
 
3509
3510	BUG_ON(down_read_trylock(&mm->mmap_sem));
3511
3512	mutex_lock(&mm_all_locks_mutex);
3513
3514	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3515		if (signal_pending(current))
3516			goto out_unlock;
3517		if (vma->vm_file && vma->vm_file->f_mapping &&
3518				is_vm_hugetlb_page(vma))
3519			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3520	}
3521
3522	for (vma = mm->mmap; vma; vma = vma->vm_next) {
 
3523		if (signal_pending(current))
3524			goto out_unlock;
3525		if (vma->vm_file && vma->vm_file->f_mapping &&
3526				!is_vm_hugetlb_page(vma))
3527			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3528	}
3529
3530	for (vma = mm->mmap; vma; vma = vma->vm_next) {
 
3531		if (signal_pending(current))
3532			goto out_unlock;
3533		if (vma->anon_vma)
3534			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3535				vm_lock_anon_vma(mm, avc->anon_vma);
3536	}
3537
3538	return 0;
3539
3540out_unlock:
3541	mm_drop_all_locks(mm);
3542	return -EINTR;
3543}
3544
3545static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3546{
3547	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3548		/*
3549		 * The LSB of head.next can't change to 0 from under
3550		 * us because we hold the mm_all_locks_mutex.
3551		 *
3552		 * We must however clear the bitflag before unlocking
3553		 * the vma so the users using the anon_vma->rb_root will
3554		 * never see our bitflag.
3555		 *
3556		 * No need of atomic instructions here, head.next
3557		 * can't change from under us until we release the
3558		 * anon_vma->root->rwsem.
3559		 */
3560		if (!__test_and_clear_bit(0, (unsigned long *)
3561					  &anon_vma->root->rb_root.rb_root.rb_node))
3562			BUG();
3563		anon_vma_unlock_write(anon_vma);
3564	}
3565}
3566
3567static void vm_unlock_mapping(struct address_space *mapping)
3568{
3569	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3570		/*
3571		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3572		 * because we hold the mm_all_locks_mutex.
3573		 */
3574		i_mmap_unlock_write(mapping);
3575		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3576					&mapping->flags))
3577			BUG();
3578	}
3579}
3580
3581/*
3582 * The mmap_sem cannot be released by the caller until
3583 * mm_drop_all_locks() returns.
3584 */
3585void mm_drop_all_locks(struct mm_struct *mm)
3586{
3587	struct vm_area_struct *vma;
3588	struct anon_vma_chain *avc;
 
3589
3590	BUG_ON(down_read_trylock(&mm->mmap_sem));
3591	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3592
3593	for (vma = mm->mmap; vma; vma = vma->vm_next) {
3594		if (vma->anon_vma)
3595			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3596				vm_unlock_anon_vma(avc->anon_vma);
3597		if (vma->vm_file && vma->vm_file->f_mapping)
3598			vm_unlock_mapping(vma->vm_file->f_mapping);
3599	}
3600
3601	mutex_unlock(&mm_all_locks_mutex);
3602}
3603
3604/*
3605 * initialise the percpu counter for VM
3606 */
3607void __init mmap_init(void)
3608{
3609	int ret;
3610
3611	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3612	VM_BUG_ON(ret);
3613}
3614
3615/*
3616 * Initialise sysctl_user_reserve_kbytes.
3617 *
3618 * This is intended to prevent a user from starting a single memory hogging
3619 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3620 * mode.
3621 *
3622 * The default value is min(3% of free memory, 128MB)
3623 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3624 */
3625static int init_user_reserve(void)
3626{
3627	unsigned long free_kbytes;
3628
3629	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3630
3631	sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
3632	return 0;
3633}
3634subsys_initcall(init_user_reserve);
3635
3636/*
3637 * Initialise sysctl_admin_reserve_kbytes.
3638 *
3639 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3640 * to log in and kill a memory hogging process.
3641 *
3642 * Systems with more than 256MB will reserve 8MB, enough to recover
3643 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3644 * only reserve 3% of free pages by default.
3645 */
3646static int init_admin_reserve(void)
3647{
3648	unsigned long free_kbytes;
3649
3650	free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3651
3652	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
3653	return 0;
3654}
3655subsys_initcall(init_admin_reserve);
3656
3657/*
3658 * Reinititalise user and admin reserves if memory is added or removed.
3659 *
3660 * The default user reserve max is 128MB, and the default max for the
3661 * admin reserve is 8MB. These are usually, but not always, enough to
3662 * enable recovery from a memory hogging process using login/sshd, a shell,
3663 * and tools like top. It may make sense to increase or even disable the
3664 * reserve depending on the existence of swap or variations in the recovery
3665 * tools. So, the admin may have changed them.
3666 *
3667 * If memory is added and the reserves have been eliminated or increased above
3668 * the default max, then we'll trust the admin.
3669 *
3670 * If memory is removed and there isn't enough free memory, then we
3671 * need to reset the reserves.
3672 *
3673 * Otherwise keep the reserve set by the admin.
3674 */
3675static int reserve_mem_notifier(struct notifier_block *nb,
3676			     unsigned long action, void *data)
3677{
3678	unsigned long tmp, free_kbytes;
3679
3680	switch (action) {
3681	case MEM_ONLINE:
3682		/* Default max is 128MB. Leave alone if modified by operator. */
3683		tmp = sysctl_user_reserve_kbytes;
3684		if (0 < tmp && tmp < (1UL << 17))
3685			init_user_reserve();
3686
3687		/* Default max is 8MB.  Leave alone if modified by operator. */
3688		tmp = sysctl_admin_reserve_kbytes;
3689		if (0 < tmp && tmp < (1UL << 13))
3690			init_admin_reserve();
3691
3692		break;
3693	case MEM_OFFLINE:
3694		free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
3695
3696		if (sysctl_user_reserve_kbytes > free_kbytes) {
3697			init_user_reserve();
3698			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3699				sysctl_user_reserve_kbytes);
3700		}
3701
3702		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3703			init_admin_reserve();
3704			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3705				sysctl_admin_reserve_kbytes);
3706		}
3707		break;
3708	default:
3709		break;
3710	}
3711	return NOTIFY_OK;
3712}
3713
3714static struct notifier_block reserve_mem_nb = {
3715	.notifier_call = reserve_mem_notifier,
3716};
3717
3718static int __meminit init_reserve_notifier(void)
3719{
3720	if (register_hotmemory_notifier(&reserve_mem_nb))
3721		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3722
3723	return 0;
3724}
3725subsys_initcall(init_reserve_notifier);