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);
v6.9.4
   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#include <linux/ksm.h>
  50
  51#include <linux/uaccess.h>
  52#include <asm/cacheflush.h>
  53#include <asm/tlb.h>
  54#include <asm/mmu_context.h>
  55
  56#define CREATE_TRACE_POINTS
  57#include <trace/events/mmap.h>
  58
  59#include "internal.h"
  60
  61#ifndef arch_mmap_check
  62#define arch_mmap_check(addr, len, flags)	(0)
  63#endif
  64
  65#ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
  66const int mmap_rnd_bits_min = CONFIG_ARCH_MMAP_RND_BITS_MIN;
  67int mmap_rnd_bits_max __ro_after_init = CONFIG_ARCH_MMAP_RND_BITS_MAX;
  68int mmap_rnd_bits __read_mostly = CONFIG_ARCH_MMAP_RND_BITS;
  69#endif
  70#ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
  71const int mmap_rnd_compat_bits_min = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MIN;
  72const int mmap_rnd_compat_bits_max = CONFIG_ARCH_MMAP_RND_COMPAT_BITS_MAX;
  73int mmap_rnd_compat_bits __read_mostly = CONFIG_ARCH_MMAP_RND_COMPAT_BITS;
  74#endif
  75
  76static bool ignore_rlimit_data;
  77core_param(ignore_rlimit_data, ignore_rlimit_data, bool, 0644);
  78
  79static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
  80		struct vm_area_struct *vma, struct vm_area_struct *prev,
  81		struct vm_area_struct *next, unsigned long start,
  82		unsigned long end, unsigned long tree_end, bool mm_wr_locked);
  83
  84static pgprot_t vm_pgprot_modify(pgprot_t oldprot, unsigned long vm_flags)
  85{
  86	return pgprot_modify(oldprot, vm_get_page_prot(vm_flags));
  87}
  88
  89/* Update vma->vm_page_prot to reflect vma->vm_flags. */
  90void vma_set_page_prot(struct vm_area_struct *vma)
  91{
  92	unsigned long vm_flags = vma->vm_flags;
  93	pgprot_t vm_page_prot;
  94
  95	vm_page_prot = vm_pgprot_modify(vma->vm_page_prot, vm_flags);
  96	if (vma_wants_writenotify(vma, vm_page_prot)) {
  97		vm_flags &= ~VM_SHARED;
  98		vm_page_prot = vm_pgprot_modify(vm_page_prot, vm_flags);
  99	}
 100	/* remove_protection_ptes reads vma->vm_page_prot without mmap_lock */
 101	WRITE_ONCE(vma->vm_page_prot, vm_page_prot);
 102}
 103
 104/*
 105 * Requires inode->i_mapping->i_mmap_rwsem
 106 */
 107static void __remove_shared_vm_struct(struct vm_area_struct *vma,
 108				      struct address_space *mapping)
 109{
 110	if (vma_is_shared_maywrite(vma))
 111		mapping_unmap_writable(mapping);
 112
 113	flush_dcache_mmap_lock(mapping);
 114	vma_interval_tree_remove(vma, &mapping->i_mmap);
 115	flush_dcache_mmap_unlock(mapping);
 116}
 117
 118/*
 119 * Unlink a file-based vm structure from its interval tree, to hide
 120 * vma from rmap and vmtruncate before freeing its page tables.
 121 */
 122void unlink_file_vma(struct vm_area_struct *vma)
 123{
 124	struct file *file = vma->vm_file;
 125
 126	if (file) {
 127		struct address_space *mapping = file->f_mapping;
 128		i_mmap_lock_write(mapping);
 129		__remove_shared_vm_struct(vma, mapping);
 130		i_mmap_unlock_write(mapping);
 131	}
 132}
 133
 134/*
 135 * Close a vm structure and free it.
 136 */
 137static void remove_vma(struct vm_area_struct *vma, bool unreachable)
 138{
 139	might_sleep();
 140	if (vma->vm_ops && vma->vm_ops->close)
 141		vma->vm_ops->close(vma);
 142	if (vma->vm_file)
 143		fput(vma->vm_file);
 144	mpol_put(vma_policy(vma));
 145	if (unreachable)
 146		__vm_area_free(vma);
 147	else
 148		vm_area_free(vma);
 149}
 150
 151static inline struct vm_area_struct *vma_prev_limit(struct vma_iterator *vmi,
 152						    unsigned long min)
 153{
 154	return mas_prev(&vmi->mas, min);
 155}
 156
 157/*
 158 * check_brk_limits() - Use platform specific check of range & verify mlock
 159 * limits.
 160 * @addr: The address to check
 161 * @len: The size of increase.
 162 *
 163 * Return: 0 on success.
 164 */
 165static int check_brk_limits(unsigned long addr, unsigned long len)
 166{
 167	unsigned long mapped_addr;
 168
 169	mapped_addr = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
 170	if (IS_ERR_VALUE(mapped_addr))
 171		return mapped_addr;
 172
 173	return mlock_future_ok(current->mm, current->mm->def_flags, len)
 174		? 0 : -EAGAIN;
 175}
 176static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *brkvma,
 
 
 
 177		unsigned long addr, unsigned long request, unsigned long flags);
 178SYSCALL_DEFINE1(brk, unsigned long, brk)
 179{
 180	unsigned long newbrk, oldbrk, origbrk;
 181	struct mm_struct *mm = current->mm;
 182	struct vm_area_struct *brkvma, *next = NULL;
 183	unsigned long min_brk;
 184	bool populate = false;
 
 185	LIST_HEAD(uf);
 186	struct vma_iterator vmi;
 187
 188	if (mmap_write_lock_killable(mm))
 189		return -EINTR;
 190
 191	origbrk = mm->brk;
 192
 193#ifdef CONFIG_COMPAT_BRK
 194	/*
 195	 * CONFIG_COMPAT_BRK can still be overridden by setting
 196	 * randomize_va_space to 2, which will still cause mm->start_brk
 197	 * to be arbitrarily shifted
 198	 */
 199	if (current->brk_randomized)
 200		min_brk = mm->start_brk;
 201	else
 202		min_brk = mm->end_data;
 203#else
 204	min_brk = mm->start_brk;
 205#endif
 206	if (brk < min_brk)
 207		goto out;
 208
 209	/*
 210	 * Check against rlimit here. If this check is done later after the test
 211	 * of oldbrk with newbrk then it can escape the test and let the data
 212	 * segment grow beyond its set limit the in case where the limit is
 213	 * not page aligned -Ram Gupta
 214	 */
 215	if (check_data_rlimit(rlimit(RLIMIT_DATA), brk, mm->start_brk,
 216			      mm->end_data, mm->start_data))
 217		goto out;
 218
 219	newbrk = PAGE_ALIGN(brk);
 220	oldbrk = PAGE_ALIGN(mm->brk);
 221	if (oldbrk == newbrk) {
 222		mm->brk = brk;
 223		goto success;
 224	}
 225
 226	/* Always allow shrinking brk. */
 
 
 
 227	if (brk <= mm->brk) {
 
 
 228		/* Search one past newbrk */
 229		vma_iter_init(&vmi, mm, newbrk);
 230		brkvma = vma_find(&vmi, oldbrk);
 231		if (!brkvma || brkvma->vm_start >= oldbrk)
 232			goto out; /* mapping intersects with an existing non-brk vma. */
 233		/*
 234		 * mm->brk must be protected by write mmap_lock.
 235		 * do_vma_munmap() will drop the lock on success,  so update it
 236		 * before calling do_vma_munmap().
 237		 */
 238		mm->brk = brk;
 239		if (do_vma_munmap(&vmi, brkvma, newbrk, oldbrk, &uf, true))
 240			goto out;
 
 
 
 
 241
 242		goto success_unlocked;
 
 243	}
 244
 245	if (check_brk_limits(oldbrk, newbrk - oldbrk))
 246		goto out;
 247
 248	/*
 249	 * Only check if the next VMA is within the stack_guard_gap of the
 250	 * expansion area
 251	 */
 252	vma_iter_init(&vmi, mm, oldbrk);
 253	next = vma_find(&vmi, newbrk + PAGE_SIZE + stack_guard_gap);
 254	if (next && newbrk + PAGE_SIZE > vm_start_gap(next))
 255		goto out;
 256
 257	brkvma = vma_prev_limit(&vmi, mm->start_brk);
 258	/* Ok, looks good - let it rip. */
 259	if (do_brk_flags(&vmi, brkvma, oldbrk, newbrk - oldbrk, 0) < 0)
 260		goto out;
 261
 262	mm->brk = brk;
 263	if (mm->def_flags & VM_LOCKED)
 264		populate = true;
 265
 266success:
 267	mmap_write_unlock(mm);
 268success_unlocked:
 
 
 
 269	userfaultfd_unmap_complete(mm, &uf);
 270	if (populate)
 271		mm_populate(oldbrk, newbrk - oldbrk);
 272	return brk;
 273
 274out:
 275	mm->brk = origbrk;
 276	mmap_write_unlock(mm);
 277	return origbrk;
 278}
 279
 280#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 281static void validate_mm(struct mm_struct *mm)
 282{
 283	int bug = 0;
 284	int i = 0;
 285	struct vm_area_struct *vma;
 286	VMA_ITERATOR(vmi, mm, 0);
 
 
 287
 288	mt_validate(&mm->mm_mt);
 289	for_each_vma(vmi, vma) {
 290#ifdef CONFIG_DEBUG_VM_RB
 291		struct anon_vma *anon_vma = vma->anon_vma;
 292		struct anon_vma_chain *avc;
 293#endif
 294		unsigned long vmi_start, vmi_end;
 295		bool warn = 0;
 296
 297		vmi_start = vma_iter_addr(&vmi);
 298		vmi_end = vma_iter_end(&vmi);
 299		if (VM_WARN_ON_ONCE_MM(vma->vm_end != vmi_end, mm))
 300			warn = 1;
 301
 302		if (VM_WARN_ON_ONCE_MM(vma->vm_start != vmi_start, mm))
 303			warn = 1;
 304
 305		if (warn) {
 306			pr_emerg("issue in %s\n", current->comm);
 307			dump_stack();
 308			dump_vma(vma);
 309			pr_emerg("tree range: %px start %lx end %lx\n", vma,
 310				 vmi_start, vmi_end - 1);
 311			vma_iter_dump_tree(&vmi);
 312		}
 313
 314#ifdef CONFIG_DEBUG_VM_RB
 315		if (anon_vma) {
 316			anon_vma_lock_read(anon_vma);
 317			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 318				anon_vma_interval_tree_verify(avc);
 319			anon_vma_unlock_read(anon_vma);
 320		}
 321#endif
 322		i++;
 323	}
 324	if (i != mm->map_count) {
 325		pr_emerg("map_count %d vma iterator %d\n", mm->map_count, i);
 326		bug = 1;
 327	}
 328	VM_BUG_ON_MM(bug, mm);
 329}
 330
 331#else /* !CONFIG_DEBUG_VM_MAPLE_TREE */
 
 332#define validate_mm(mm) do { } while (0)
 333#endif /* CONFIG_DEBUG_VM_MAPLE_TREE */
 334
 335/*
 336 * vma has some anon_vma assigned, and is already inserted on that
 337 * anon_vma's interval trees.
 338 *
 339 * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
 340 * vma must be removed from the anon_vma's interval trees using
 341 * anon_vma_interval_tree_pre_update_vma().
 342 *
 343 * After the update, the vma will be reinserted using
 344 * anon_vma_interval_tree_post_update_vma().
 345 *
 346 * The entire update must be protected by exclusive mmap_lock and by
 347 * the root anon_vma's mutex.
 348 */
 349static inline void
 350anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
 351{
 352	struct anon_vma_chain *avc;
 353
 354	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 355		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
 356}
 357
 358static inline void
 359anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
 360{
 361	struct anon_vma_chain *avc;
 362
 363	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
 364		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
 365}
 366
 367static unsigned long count_vma_pages_range(struct mm_struct *mm,
 368		unsigned long addr, unsigned long end)
 369{
 370	VMA_ITERATOR(vmi, mm, addr);
 371	struct vm_area_struct *vma;
 372	unsigned long nr_pages = 0;
 373
 374	for_each_vma_range(vmi, vma, end) {
 375		unsigned long vm_start = max(addr, vma->vm_start);
 376		unsigned long vm_end = min(end, vma->vm_end);
 377
 378		nr_pages += PHYS_PFN(vm_end - vm_start);
 379	}
 380
 381	return nr_pages;
 382}
 383
 384static void __vma_link_file(struct vm_area_struct *vma,
 385			    struct address_space *mapping)
 386{
 387	if (vma_is_shared_maywrite(vma))
 388		mapping_allow_writable(mapping);
 389
 390	flush_dcache_mmap_lock(mapping);
 391	vma_interval_tree_insert(vma, &mapping->i_mmap);
 392	flush_dcache_mmap_unlock(mapping);
 393}
 394
 395static void vma_link_file(struct vm_area_struct *vma)
 396{
 397	struct file *file = vma->vm_file;
 398	struct address_space *mapping;
 399
 400	if (file) {
 401		mapping = file->f_mapping;
 402		i_mmap_lock_write(mapping);
 403		__vma_link_file(vma, mapping);
 404		i_mmap_unlock_write(mapping);
 405	}
 406}
 407
 408static int vma_link(struct mm_struct *mm, struct vm_area_struct *vma)
 409{
 410	VMA_ITERATOR(vmi, mm, 0);
 411
 412	vma_iter_config(&vmi, vma->vm_start, vma->vm_end);
 413	if (vma_iter_prealloc(&vmi, vma))
 414		return -ENOMEM;
 415
 416	vma_start_write(vma);
 417	vma_iter_store(&vmi, vma);
 418	vma_link_file(vma);
 419	mm->map_count++;
 420	validate_mm(mm);
 421	return 0;
 422}
 423
 424/*
 425 * init_multi_vma_prep() - Initializer for struct vma_prepare
 426 * @vp: The vma_prepare struct
 427 * @vma: The vma that will be altered once locked
 428 * @next: The next vma if it is to be adjusted
 429 * @remove: The first vma to be removed
 430 * @remove2: The second vma to be removed
 
 
 431 */
 432static inline void init_multi_vma_prep(struct vma_prepare *vp,
 433		struct vm_area_struct *vma, struct vm_area_struct *next,
 434		struct vm_area_struct *remove, struct vm_area_struct *remove2)
 435{
 436	memset(vp, 0, sizeof(struct vma_prepare));
 437	vp->vma = vma;
 438	vp->anon_vma = vma->anon_vma;
 439	vp->remove = remove;
 440	vp->remove2 = remove2;
 441	vp->adj_next = next;
 442	if (!vp->anon_vma && next)
 443		vp->anon_vma = next->anon_vma;
 444
 445	vp->file = vma->vm_file;
 446	if (vp->file)
 447		vp->mapping = vma->vm_file->f_mapping;
 448
 449}
 450
 451/*
 452 * init_vma_prep() - Initializer wrapper for vma_prepare struct
 453 * @vp: The vma_prepare struct
 454 * @vma: The vma that will be altered once locked
 
 
 
 
 455 */
 456static inline void init_vma_prep(struct vma_prepare *vp,
 457				 struct vm_area_struct *vma)
 458{
 459	init_multi_vma_prep(vp, vma, NULL, NULL, NULL);
 
 
 
 460}
 461
 462
 463/*
 464 * vma_prepare() - Helper function for handling locking VMAs prior to altering
 465 * @vp: The initialized vma_prepare struct
 
 
 
 
 466 */
 467static inline void vma_prepare(struct vma_prepare *vp)
 
 468{
 469	if (vp->file) {
 470		uprobe_munmap(vp->vma, vp->vma->vm_start, vp->vma->vm_end);
 471
 472		if (vp->adj_next)
 473			uprobe_munmap(vp->adj_next, vp->adj_next->vm_start,
 474				      vp->adj_next->vm_end);
 475
 476		i_mmap_lock_write(vp->mapping);
 477		if (vp->insert && vp->insert->vm_file) {
 478			/*
 479			 * Put into interval tree now, so instantiated pages
 480			 * are visible to arm/parisc __flush_dcache_page
 481			 * throughout; but we cannot insert into address
 482			 * space until vma start or end is updated.
 483			 */
 484			__vma_link_file(vp->insert,
 485					vp->insert->vm_file->f_mapping);
 486		}
 487	}
 488
 489	if (vp->anon_vma) {
 490		anon_vma_lock_write(vp->anon_vma);
 491		anon_vma_interval_tree_pre_update_vma(vp->vma);
 492		if (vp->adj_next)
 493			anon_vma_interval_tree_pre_update_vma(vp->adj_next);
 494	}
 495
 496	if (vp->file) {
 497		flush_dcache_mmap_lock(vp->mapping);
 498		vma_interval_tree_remove(vp->vma, &vp->mapping->i_mmap);
 499		if (vp->adj_next)
 500			vma_interval_tree_remove(vp->adj_next,
 501						 &vp->mapping->i_mmap);
 502	}
 503
 504}
 505
 506/*
 507 * vma_complete- Helper function for handling the unlocking after altering VMAs,
 508 * or for inserting a VMA.
 509 *
 510 * @vp: The vma_prepare struct
 511 * @vmi: The vma iterator
 512 * @mm: The mm_struct
 513 */
 514static inline void vma_complete(struct vma_prepare *vp,
 515				struct vma_iterator *vmi, struct mm_struct *mm)
 516{
 517	if (vp->file) {
 518		if (vp->adj_next)
 519			vma_interval_tree_insert(vp->adj_next,
 520						 &vp->mapping->i_mmap);
 521		vma_interval_tree_insert(vp->vma, &vp->mapping->i_mmap);
 522		flush_dcache_mmap_unlock(vp->mapping);
 523	}
 524
 525	if (vp->remove && vp->file) {
 526		__remove_shared_vm_struct(vp->remove, vp->mapping);
 527		if (vp->remove2)
 528			__remove_shared_vm_struct(vp->remove2, vp->mapping);
 529	} else if (vp->insert) {
 530		/*
 531		 * split_vma has split insert from vma, and needs
 532		 * us to insert it before dropping the locks
 533		 * (it may either follow vma or precede it).
 534		 */
 535		vma_iter_store(vmi, vp->insert);
 536		mm->map_count++;
 537	}
 538
 539	if (vp->anon_vma) {
 540		anon_vma_interval_tree_post_update_vma(vp->vma);
 541		if (vp->adj_next)
 542			anon_vma_interval_tree_post_update_vma(vp->adj_next);
 543		anon_vma_unlock_write(vp->anon_vma);
 544	}
 545
 546	if (vp->file) {
 547		i_mmap_unlock_write(vp->mapping);
 548		uprobe_mmap(vp->vma);
 549
 550		if (vp->adj_next)
 551			uprobe_mmap(vp->adj_next);
 
 552	}
 553
 554	if (vp->remove) {
 555again:
 556		vma_mark_detached(vp->remove, true);
 557		if (vp->file) {
 558			uprobe_munmap(vp->remove, vp->remove->vm_start,
 559				      vp->remove->vm_end);
 560			fput(vp->file);
 561		}
 562		if (vp->remove->anon_vma)
 563			anon_vma_merge(vp->vma, vp->remove);
 564		mm->map_count--;
 565		mpol_put(vma_policy(vp->remove));
 566		if (!vp->remove2)
 567			WARN_ON_ONCE(vp->vma->vm_end < vp->remove->vm_end);
 568		vm_area_free(vp->remove);
 569
 570		/*
 571		 * In mprotect's case 6 (see comments on vma_merge),
 572		 * we are removing both mid and next vmas
 573		 */
 574		if (vp->remove2) {
 575			vp->remove = vp->remove2;
 576			vp->remove2 = NULL;
 577			goto again;
 578		}
 579	}
 580	if (vp->insert && vp->file)
 581		uprobe_mmap(vp->insert);
 582	validate_mm(mm);
 583}
 584
 585/*
 586 * dup_anon_vma() - Helper function to duplicate anon_vma
 587 * @dst: The destination VMA
 588 * @src: The source VMA
 589 * @dup: Pointer to the destination VMA when successful.
 590 *
 591 * Returns: 0 on success.
 592 */
 593static inline int dup_anon_vma(struct vm_area_struct *dst,
 594		struct vm_area_struct *src, struct vm_area_struct **dup)
 595{
 596	/*
 597	 * Easily overlooked: when mprotect shifts the boundary, make sure the
 598	 * expanding vma has anon_vma set if the shrinking vma had, to cover any
 599	 * anon pages imported.
 600	 */
 601	if (src->anon_vma && !dst->anon_vma) {
 602		int ret;
 603
 604		vma_assert_write_locked(dst);
 605		dst->anon_vma = src->anon_vma;
 606		ret = anon_vma_clone(dst, src);
 607		if (ret)
 608			return ret;
 609
 610		*dup = dst;
 611	}
 612
 613	return 0;
 614}
 615
 616/*
 617 * vma_expand - Expand an existing VMA
 618 *
 619 * @vmi: The vma iterator
 620 * @vma: The vma to expand
 621 * @start: The start of the vma
 622 * @end: The exclusive end of the vma
 623 * @pgoff: The page offset of vma
 624 * @next: The current of next vma.
 625 *
 626 * Expand @vma to @start and @end.  Can expand off the start and end.  Will
 627 * expand over @next if it's different from @vma and @end == @next->vm_end.
 628 * Checking if the @vma can expand and merge with @next needs to be handled by
 629 * the caller.
 630 *
 631 * Returns: 0 on success
 632 */
 633int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
 634	       unsigned long start, unsigned long end, pgoff_t pgoff,
 635	       struct vm_area_struct *next)
 636{
 637	struct vm_area_struct *anon_dup = NULL;
 
 
 
 
 638	bool remove_next = false;
 639	struct vma_prepare vp;
 640
 641	vma_start_write(vma);
 642	if (next && (vma != next) && (end == next->vm_end)) {
 643		int ret;
 
 
 644
 645		remove_next = true;
 646		vma_start_write(next);
 647		ret = dup_anon_vma(vma, next, &anon_dup);
 648		if (ret)
 649			return ret;
 
 650	}
 651
 652	init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
 653	/* Not merging but overwriting any part of next is not handled. */
 654	VM_WARN_ON(next && !vp.remove &&
 655		  next != vma && end > next->vm_start);
 656	/* Only handles expanding */
 657	VM_WARN_ON(vma->vm_start < start || vma->vm_end > end);
 658
 659	/* Note: vma iterator must be pointing to 'start' */
 660	vma_iter_config(vmi, start, end);
 661	if (vma_iter_prealloc(vmi, vma))
 662		goto nomem;
 663
 664	vma_prepare(&vp);
 665	vma_adjust_trans_huge(vma, start, end, 0);
 666	vma_set_range(vma, start, end, pgoff);
 667	vma_iter_store(vmi, vma);
 668
 669	vma_complete(&vp, vmi, vma->vm_mm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 670	return 0;
 671
 672nomem:
 673	if (anon_dup)
 674		unlink_anon_vmas(anon_dup);
 675	return -ENOMEM;
 676}
 677
 678/*
 679 * vma_shrink() - Reduce an existing VMAs memory area
 680 * @vmi: The vma iterator
 681 * @vma: The VMA to modify
 682 * @start: The new start
 683 * @end: The new end
 684 *
 685 * Returns: 0 on success, -ENOMEM otherwise
 686 */
 687int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
 688	       unsigned long start, unsigned long end, pgoff_t pgoff)
 689{
 690	struct vma_prepare vp;
 
 
 
 
 
 
 
 
 
 
 
 
 691
 692	WARN_ON((vma->vm_start != start) && (vma->vm_end != end));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 693
 694	if (vma->vm_start < start)
 695		vma_iter_config(vmi, vma->vm_start, start);
 696	else
 697		vma_iter_config(vmi, end, vma->vm_end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 698
 699	if (vma_iter_prealloc(vmi, NULL))
 700		return -ENOMEM;
 701
 702	vma_start_write(vma);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 703
 704	init_vma_prep(&vp, vma);
 705	vma_prepare(&vp);
 706	vma_adjust_trans_huge(vma, start, end, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 707
 708	vma_iter_clear(vmi);
 709	vma_set_range(vma, start, end, pgoff);
 710	vma_complete(&vp, vmi, vma->vm_mm);
 711	return 0;
 712}
 713
 714/*
 715 * If the vma has a ->close operation then the driver probably needs to release
 716 * per-vma resources, so we don't attempt to merge those if the caller indicates
 717 * the current vma may be removed as part of the merge.
 718 */
 719static inline bool is_mergeable_vma(struct vm_area_struct *vma,
 720		struct file *file, unsigned long vm_flags,
 721		struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 722		struct anon_vma_name *anon_name, bool may_remove_vma)
 723{
 724	/*
 725	 * VM_SOFTDIRTY should not prevent from VMA merging, if we
 726	 * match the flags but dirty bit -- the caller should mark
 727	 * merged VMA as dirty. If dirty bit won't be excluded from
 728	 * comparison, we increase pressure on the memory system forcing
 729	 * the kernel to generate new VMAs when old one could be
 730	 * extended instead.
 731	 */
 732	if ((vma->vm_flags ^ vm_flags) & ~VM_SOFTDIRTY)
 733		return false;
 734	if (vma->vm_file != file)
 735		return false;
 736	if (may_remove_vma && vma->vm_ops && vma->vm_ops->close)
 737		return false;
 738	if (!is_mergeable_vm_userfaultfd_ctx(vma, vm_userfaultfd_ctx))
 739		return false;
 740	if (!anon_vma_name_eq(anon_vma_name(vma), anon_name))
 741		return false;
 742	return true;
 743}
 744
 745static inline bool is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 746		 struct anon_vma *anon_vma2, struct vm_area_struct *vma)
 
 747{
 748	/*
 749	 * The list_is_singular() test is to avoid merging VMA cloned from
 750	 * parents. This can improve scalability caused by anon_vma lock.
 751	 */
 752	if ((!anon_vma1 || !anon_vma2) && (!vma ||
 753		list_is_singular(&vma->anon_vma_chain)))
 754		return true;
 755	return anon_vma1 == anon_vma2;
 756}
 757
 758/*
 759 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 760 * in front of (at a lower virtual address and file offset than) the vma.
 761 *
 762 * We cannot merge two vmas if they have differently assigned (non-NULL)
 763 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 764 *
 765 * We don't check here for the merged mmap wrapping around the end of pagecache
 766 * indices (16TB on ia32) because do_mmap() does not permit mmap's which
 767 * wrap, nor mmaps which cover the final page at index -1UL.
 768 *
 769 * We assume the vma may be removed as part of the merge.
 770 */
 771static bool
 772can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 773		struct anon_vma *anon_vma, struct file *file,
 774		pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 775		struct anon_vma_name *anon_name)
 
 776{
 777	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, true) &&
 778	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 779		if (vma->vm_pgoff == vm_pgoff)
 780			return true;
 781	}
 782	return false;
 783}
 784
 785/*
 786 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 787 * beyond (at a higher virtual address and file offset than) the vma.
 788 *
 789 * We cannot merge two vmas if they have differently assigned (non-NULL)
 790 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 791 *
 792 * We assume that vma is not removed as part of the merge.
 793 */
 794static bool
 795can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 796		struct anon_vma *anon_vma, struct file *file,
 797		pgoff_t vm_pgoff, struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 798		struct anon_vma_name *anon_name)
 
 799{
 800	if (is_mergeable_vma(vma, file, vm_flags, vm_userfaultfd_ctx, anon_name, false) &&
 801	    is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
 802		pgoff_t vm_pglen;
 803		vm_pglen = vma_pages(vma);
 804		if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 805			return true;
 806	}
 807	return false;
 808}
 809
 810/*
 811 * Given a mapping request (addr,end,vm_flags,file,pgoff,anon_name),
 812 * figure out whether that can be merged with its predecessor or its
 813 * successor.  Or both (it neatly fills a hole).
 814 *
 815 * In most cases - when called for mmap, brk or mremap - [addr,end) is
 816 * certain not to be mapped by the time vma_merge is called; but when
 817 * called for mprotect, it is certain to be already mapped (either at
 818 * an offset within prev, or at the start of next), and the flags of
 819 * this area are about to be changed to vm_flags - and the no-change
 820 * case has already been eliminated.
 821 *
 822 * The following mprotect cases have to be considered, where **** is
 823 * the area passed down from mprotect_fixup, never extending beyond one
 824 * vma, PPPP is the previous vma, CCCC is a concurrent vma that starts
 825 * at the same address as **** and is of the same or larger span, and
 826 * NNNN the next vma after ****:
 827 *
 828 *     ****             ****                   ****
 829 *    PPPPPPNNNNNN    PPPPPPNNNNNN       PPPPPPCCCCCC
 830 *    cannot merge    might become       might become
 831 *                    PPNNNNNNNNNN       PPPPPPPPPPCC
 832 *    mmap, brk or    case 4 below       case 5 below
 833 *    mremap move:
 834 *                        ****               ****
 835 *                    PPPP    NNNN       PPPPCCCCNNNN
 836 *                    might become       might become
 837 *                    PPPPPPPPPPPP 1 or  PPPPPPPPPPPP 6 or
 838 *                    PPPPPPPPNNNN 2 or  PPPPPPPPNNNN 7 or
 839 *                    PPPPNNNNNNNN 3     PPPPNNNNNNNN 8
 840 *
 841 * It is important for case 8 that the vma CCCC overlapping the
 842 * region **** is never going to extended over NNNN. Instead NNNN must
 843 * be extended in region **** and CCCC must be removed. This way in
 844 * all cases where vma_merge succeeds, the moment vma_merge drops the
 845 * rmap_locks, the properties of the merged vma will be already
 846 * correct for the whole merged range. Some of those properties like
 847 * vm_page_prot/vm_flags may be accessed by rmap_walks and they must
 848 * be correct for the whole merged range immediately after the
 849 * rmap_locks are released. Otherwise if NNNN would be removed and
 850 * CCCC would be extended over the NNNN range, remove_migration_ptes
 851 * or other rmap walkers (if working on addresses beyond the "end"
 852 * parameter) may establish ptes with the wrong permissions of CCCC
 853 * instead of the right permissions of NNNN.
 854 *
 855 * In the code below:
 856 * PPPP is represented by *prev
 857 * CCCC is represented by *curr or not represented at all (NULL)
 858 * NNNN is represented by *next or not represented at all (NULL)
 859 * **** is not represented - it will be merged and the vma containing the
 860 *      area is returned, or the function will return NULL
 861 */
 862static struct vm_area_struct
 863*vma_merge(struct vma_iterator *vmi, struct vm_area_struct *prev,
 864	   struct vm_area_struct *src, unsigned long addr, unsigned long end,
 865	   unsigned long vm_flags, pgoff_t pgoff, struct mempolicy *policy,
 866	   struct vm_userfaultfd_ctx vm_userfaultfd_ctx,
 867	   struct anon_vma_name *anon_name)
 868{
 869	struct mm_struct *mm = src->vm_mm;
 870	struct anon_vma *anon_vma = src->anon_vma;
 871	struct file *file = src->vm_file;
 872	struct vm_area_struct *curr, *next, *res;
 873	struct vm_area_struct *vma, *adjust, *remove, *remove2;
 874	struct vm_area_struct *anon_dup = NULL;
 875	struct vma_prepare vp;
 876	pgoff_t vma_pgoff;
 877	int err = 0;
 878	bool merge_prev = false;
 879	bool merge_next = false;
 880	bool vma_expanded = false;
 881	unsigned long vma_start = addr;
 882	unsigned long vma_end = end;
 883	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
 884	long adj_start = 0;
 885
 886	/*
 887	 * We later require that vma->vm_flags == vm_flags,
 888	 * so this tests vma->vm_flags & VM_SPECIAL, too.
 889	 */
 890	if (vm_flags & VM_SPECIAL)
 891		return NULL;
 892
 893	/* Does the input range span an existing VMA? (cases 5 - 8) */
 894	curr = find_vma_intersection(mm, prev ? prev->vm_end : 0, end);
 
 
 895
 896	if (!curr ||			/* cases 1 - 4 */
 897	    end == curr->vm_end)	/* cases 6 - 8, adjacent VMA */
 898		next = vma_lookup(mm, end);
 899	else
 900		next = NULL;		/* case 5 */
 901
 902	if (prev) {
 903		vma_start = prev->vm_start;
 904		vma_pgoff = prev->vm_pgoff;
 905
 906		/* Can we merge the predecessor? */
 907		if (addr == prev->vm_end && mpol_equal(vma_policy(prev), policy)
 908		    && can_vma_merge_after(prev, vm_flags, anon_vma, file,
 909					   pgoff, vm_userfaultfd_ctx, anon_name)) {
 910			merge_prev = true;
 911			vma_prev(vmi);
 912		}
 913	}
 914
 915	/* Can we merge the successor? */
 916	if (next && mpol_equal(policy, vma_policy(next)) &&
 917	    can_vma_merge_before(next, vm_flags, anon_vma, file, pgoff+pglen,
 918				 vm_userfaultfd_ctx, anon_name)) {
 
 
 919		merge_next = true;
 920	}
 921
 922	/* Verify some invariant that must be enforced by the caller. */
 923	VM_WARN_ON(prev && addr <= prev->vm_start);
 924	VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
 925	VM_WARN_ON(addr >= end);
 926
 927	if (!merge_prev && !merge_next)
 928		return NULL; /* Not mergeable. */
 929
 930	if (merge_prev)
 931		vma_start_write(prev);
 932
 933	res = vma = prev;
 934	remove = remove2 = adjust = NULL;
 935
 936	/* Can we merge both the predecessor and the successor? */
 937	if (merge_prev && merge_next &&
 938	    is_mergeable_anon_vma(prev->anon_vma, next->anon_vma, NULL)) {
 939		vma_start_write(next);
 940		remove = next;				/* case 1 */
 941		vma_end = next->vm_end;
 942		err = dup_anon_vma(prev, next, &anon_dup);
 943		if (curr) {				/* case 6 */
 944			vma_start_write(curr);
 945			remove = curr;
 946			remove2 = next;
 947			/*
 948			 * Note that the dup_anon_vma below cannot overwrite err
 949			 * since the first caller would do nothing unless next
 950			 * has an anon_vma.
 951			 */
 952			if (!next->anon_vma)
 953				err = dup_anon_vma(prev, curr, &anon_dup);
 954		}
 955	} else if (merge_prev) {			/* case 2 */
 956		if (curr) {
 957			vma_start_write(curr);
 958			if (end == curr->vm_end) {	/* case 7 */
 959				/*
 960				 * can_vma_merge_after() assumed we would not be
 961				 * removing prev vma, so it skipped the check
 962				 * for vm_ops->close, but we are removing curr
 963				 */
 964				if (curr->vm_ops && curr->vm_ops->close)
 965					err = -EINVAL;
 966				remove = curr;
 967			} else {			/* case 5 */
 968				adjust = curr;
 969				adj_start = (end - curr->vm_start);
 970			}
 971			if (!err)
 972				err = dup_anon_vma(prev, curr, &anon_dup);
 973		}
 974	} else { /* merge_next */
 975		vma_start_write(next);
 976		res = next;
 977		if (prev && addr < prev->vm_end) {	/* case 4 */
 978			vma_start_write(prev);
 979			vma_end = addr;
 980			adjust = next;
 981			adj_start = -(prev->vm_end - addr);
 982			err = dup_anon_vma(next, prev, &anon_dup);
 983		} else {
 984			/*
 985			 * Note that cases 3 and 8 are the ONLY ones where prev
 986			 * is permitted to be (but is not necessarily) NULL.
 987			 */
 988			vma = next;			/* case 3 */
 989			vma_start = addr;
 990			vma_end = next->vm_end;
 991			vma_pgoff = next->vm_pgoff - pglen;
 992			if (curr) {			/* case 8 */
 993				vma_pgoff = curr->vm_pgoff;
 994				vma_start_write(curr);
 995				remove = curr;
 996				err = dup_anon_vma(next, curr, &anon_dup);
 997			}
 998		}
 999	}
1000
1001	/* Error in anon_vma clone. */
 
 
1002	if (err)
1003		goto anon_vma_fail;
1004
1005	if (vma_start < vma->vm_start || vma_end > vma->vm_end)
1006		vma_expanded = true;
1007
1008	if (vma_expanded) {
1009		vma_iter_config(vmi, vma_start, vma_end);
1010	} else {
1011		vma_iter_config(vmi, adjust->vm_start + adj_start,
1012				adjust->vm_end);
1013	}
1014
1015	if (vma_iter_prealloc(vmi, vma))
1016		goto prealloc_fail;
1017
1018	init_multi_vma_prep(&vp, vma, adjust, remove, remove2);
1019	VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
1020		   vp.anon_vma != adjust->anon_vma);
1021
1022	vma_prepare(&vp);
1023	vma_adjust_trans_huge(vma, vma_start, vma_end, adj_start);
1024	vma_set_range(vma, vma_start, vma_end, vma_pgoff);
1025
1026	if (vma_expanded)
1027		vma_iter_store(vmi, vma);
1028
1029	if (adj_start) {
1030		adjust->vm_start += adj_start;
1031		adjust->vm_pgoff += adj_start >> PAGE_SHIFT;
1032		if (adj_start < 0) {
1033			WARN_ON(vma_expanded);
1034			vma_iter_store(vmi, next);
1035		}
1036	}
1037
1038	vma_complete(&vp, vmi, mm);
1039	khugepaged_enter_vma(res, vm_flags);
1040	return res;
1041
1042prealloc_fail:
1043	if (anon_dup)
1044		unlink_anon_vmas(anon_dup);
1045
1046anon_vma_fail:
1047	vma_iter_set(vmi, addr);
1048	vma_iter_load(vmi);
1049	return NULL;
1050}
1051
1052/*
1053 * Rough compatibility check to quickly see if it's even worth looking
1054 * at sharing an anon_vma.
1055 *
1056 * They need to have the same vm_file, and the flags can only differ
1057 * in things that mprotect may change.
1058 *
1059 * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1060 * we can merge the two vma's. For example, we refuse to merge a vma if
1061 * there is a vm_ops->close() function, because that indicates that the
1062 * driver is doing some kind of reference counting. But that doesn't
1063 * really matter for the anon_vma sharing case.
1064 */
1065static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1066{
1067	return a->vm_end == b->vm_start &&
1068		mpol_equal(vma_policy(a), vma_policy(b)) &&
1069		a->vm_file == b->vm_file &&
1070		!((a->vm_flags ^ b->vm_flags) & ~(VM_ACCESS_FLAGS | VM_SOFTDIRTY)) &&
1071		b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1072}
1073
1074/*
1075 * Do some basic sanity checking to see if we can re-use the anon_vma
1076 * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1077 * the same as 'old', the other will be the new one that is trying
1078 * to share the anon_vma.
1079 *
1080 * NOTE! This runs with mmap_lock held for reading, so it is possible that
1081 * the anon_vma of 'old' is concurrently in the process of being set up
1082 * by another page fault trying to merge _that_. But that's ok: if it
1083 * is being set up, that automatically means that it will be a singleton
1084 * acceptable for merging, so we can do all of this optimistically. But
1085 * we do that READ_ONCE() to make sure that we never re-load the pointer.
1086 *
1087 * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1088 * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1089 * is to return an anon_vma that is "complex" due to having gone through
1090 * a fork).
1091 *
1092 * We also make sure that the two vma's are compatible (adjacent,
1093 * and with the same memory policies). That's all stable, even with just
1094 * a read lock on the mmap_lock.
1095 */
1096static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1097{
1098	if (anon_vma_compatible(a, b)) {
1099		struct anon_vma *anon_vma = READ_ONCE(old->anon_vma);
1100
1101		if (anon_vma && list_is_singular(&old->anon_vma_chain))
1102			return anon_vma;
1103	}
1104	return NULL;
1105}
1106
1107/*
1108 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1109 * neighbouring vmas for a suitable anon_vma, before it goes off
1110 * to allocate a new anon_vma.  It checks because a repetitive
1111 * sequence of mprotects and faults may otherwise lead to distinct
1112 * anon_vmas being allocated, preventing vma merge in subsequent
1113 * mprotect.
1114 */
1115struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1116{
1117	MA_STATE(mas, &vma->vm_mm->mm_mt, vma->vm_end, vma->vm_end);
1118	struct anon_vma *anon_vma = NULL;
1119	struct vm_area_struct *prev, *next;
1120
1121	/* Try next first. */
1122	next = mas_walk(&mas);
1123	if (next) {
1124		anon_vma = reusable_anon_vma(next, vma, next);
1125		if (anon_vma)
1126			return anon_vma;
1127	}
1128
1129	prev = mas_prev(&mas, 0);
1130	VM_BUG_ON_VMA(prev != vma, vma);
1131	prev = mas_prev(&mas, 0);
1132	/* Try prev next. */
1133	if (prev)
1134		anon_vma = reusable_anon_vma(prev, prev, vma);
1135
1136	/*
1137	 * We might reach here with anon_vma == NULL if we can't find
1138	 * any reusable anon_vma.
1139	 * There's no absolute need to look only at touching neighbours:
1140	 * we could search further afield for "compatible" anon_vmas.
1141	 * But it would probably just be a waste of time searching,
1142	 * or lead to too many vmas hanging off the same anon_vma.
1143	 * We're trying to allow mprotect remerging later on,
1144	 * not trying to minimize memory used for anon_vmas.
1145	 */
1146	return anon_vma;
1147}
1148
1149/*
1150 * If a hint addr is less than mmap_min_addr change hint to be as
1151 * low as possible but still greater than mmap_min_addr
1152 */
1153static inline unsigned long round_hint_to_min(unsigned long hint)
1154{
1155	hint &= PAGE_MASK;
1156	if (((void *)hint != NULL) &&
1157	    (hint < mmap_min_addr))
1158		return PAGE_ALIGN(mmap_min_addr);
1159	return hint;
1160}
1161
1162bool mlock_future_ok(struct mm_struct *mm, unsigned long flags,
1163			unsigned long bytes)
1164{
1165	unsigned long locked_pages, limit_pages;
1166
1167	if (!(flags & VM_LOCKED) || capable(CAP_IPC_LOCK))
1168		return true;
1169
1170	locked_pages = bytes >> PAGE_SHIFT;
1171	locked_pages += mm->locked_vm;
1172
1173	limit_pages = rlimit(RLIMIT_MEMLOCK);
1174	limit_pages >>= PAGE_SHIFT;
1175
1176	return locked_pages <= limit_pages;
1177}
1178
1179static inline u64 file_mmap_size_max(struct file *file, struct inode *inode)
1180{
1181	if (S_ISREG(inode->i_mode))
1182		return MAX_LFS_FILESIZE;
1183
1184	if (S_ISBLK(inode->i_mode))
1185		return MAX_LFS_FILESIZE;
1186
1187	if (S_ISSOCK(inode->i_mode))
1188		return MAX_LFS_FILESIZE;
1189
1190	/* Special "we do even unsigned file positions" case */
1191	if (file->f_mode & FMODE_UNSIGNED_OFFSET)
1192		return 0;
1193
1194	/* Yes, random drivers might want more. But I'm tired of buggy drivers */
1195	return ULONG_MAX;
1196}
1197
1198static inline bool file_mmap_ok(struct file *file, struct inode *inode,
1199				unsigned long pgoff, unsigned long len)
1200{
1201	u64 maxsize = file_mmap_size_max(file, inode);
1202
1203	if (maxsize && len > maxsize)
1204		return false;
1205	maxsize -= len;
1206	if (pgoff > maxsize >> PAGE_SHIFT)
1207		return false;
1208	return true;
1209}
1210
1211/*
1212 * The caller must write-lock current->mm->mmap_lock.
1213 */
1214unsigned long do_mmap(struct file *file, unsigned long addr,
1215			unsigned long len, unsigned long prot,
1216			unsigned long flags, vm_flags_t vm_flags,
1217			unsigned long pgoff, unsigned long *populate,
1218			struct list_head *uf)
1219{
1220	struct mm_struct *mm = current->mm;
 
1221	int pkey = 0;
1222
 
1223	*populate = 0;
1224
1225	if (!len)
1226		return -EINVAL;
1227
1228	/*
1229	 * Does the application expect PROT_READ to imply PROT_EXEC?
1230	 *
1231	 * (the exception is when the underlying filesystem is noexec
1232	 *  mounted, in which case we don't add PROT_EXEC.)
1233	 */
1234	if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1235		if (!(file && path_noexec(&file->f_path)))
1236			prot |= PROT_EXEC;
1237
1238	/* force arch specific MAP_FIXED handling in get_unmapped_area */
1239	if (flags & MAP_FIXED_NOREPLACE)
1240		flags |= MAP_FIXED;
1241
1242	if (!(flags & MAP_FIXED))
1243		addr = round_hint_to_min(addr);
1244
1245	/* Careful about overflows.. */
1246	len = PAGE_ALIGN(len);
1247	if (!len)
1248		return -ENOMEM;
1249
1250	/* offset overflow? */
1251	if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1252		return -EOVERFLOW;
1253
1254	/* Too many mappings? */
1255	if (mm->map_count > sysctl_max_map_count)
1256		return -ENOMEM;
1257
1258	/* Obtain the address to map to. we verify (or select) it and ensure
1259	 * that it represents a valid section of the address space.
1260	 */
1261	addr = get_unmapped_area(file, addr, len, pgoff, flags);
1262	if (IS_ERR_VALUE(addr))
1263		return addr;
1264
1265	if (flags & MAP_FIXED_NOREPLACE) {
1266		if (find_vma_intersection(mm, addr, addr + len))
1267			return -EEXIST;
1268	}
1269
1270	if (prot == PROT_EXEC) {
1271		pkey = execute_only_pkey(mm);
1272		if (pkey < 0)
1273			pkey = 0;
1274	}
1275
1276	/* Do simple checking here so the lower-level routines won't have
1277	 * to. we assume access permissions have been handled by the open
1278	 * of the memory object, so we don't do any here.
1279	 */
1280	vm_flags |= calc_vm_prot_bits(prot, pkey) | calc_vm_flag_bits(flags) |
1281			mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1282
1283	if (flags & MAP_LOCKED)
1284		if (!can_do_mlock())
1285			return -EPERM;
1286
1287	if (!mlock_future_ok(mm, vm_flags, len))
1288		return -EAGAIN;
1289
1290	if (file) {
1291		struct inode *inode = file_inode(file);
1292		unsigned long flags_mask;
1293
1294		if (!file_mmap_ok(file, inode, pgoff, len))
1295			return -EOVERFLOW;
1296
1297		flags_mask = LEGACY_MAP_MASK | file->f_op->mmap_supported_flags;
1298
1299		switch (flags & MAP_TYPE) {
1300		case MAP_SHARED:
1301			/*
1302			 * Force use of MAP_SHARED_VALIDATE with non-legacy
1303			 * flags. E.g. MAP_SYNC is dangerous to use with
1304			 * MAP_SHARED as you don't know which consistency model
1305			 * you will get. We silently ignore unsupported flags
1306			 * with MAP_SHARED to preserve backward compatibility.
1307			 */
1308			flags &= LEGACY_MAP_MASK;
1309			fallthrough;
1310		case MAP_SHARED_VALIDATE:
1311			if (flags & ~flags_mask)
1312				return -EOPNOTSUPP;
1313			if (prot & PROT_WRITE) {
1314				if (!(file->f_mode & FMODE_WRITE))
1315					return -EACCES;
1316				if (IS_SWAPFILE(file->f_mapping->host))
1317					return -ETXTBSY;
1318			}
1319
1320			/*
1321			 * Make sure we don't allow writing to an append-only
1322			 * file..
1323			 */
1324			if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1325				return -EACCES;
1326
1327			vm_flags |= VM_SHARED | VM_MAYSHARE;
1328			if (!(file->f_mode & FMODE_WRITE))
1329				vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1330			fallthrough;
1331		case MAP_PRIVATE:
1332			if (!(file->f_mode & FMODE_READ))
1333				return -EACCES;
1334			if (path_noexec(&file->f_path)) {
1335				if (vm_flags & VM_EXEC)
1336					return -EPERM;
1337				vm_flags &= ~VM_MAYEXEC;
1338			}
1339
1340			if (!file->f_op->mmap)
1341				return -ENODEV;
1342			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1343				return -EINVAL;
1344			break;
1345
1346		default:
1347			return -EINVAL;
1348		}
1349	} else {
1350		switch (flags & MAP_TYPE) {
1351		case MAP_SHARED:
1352			if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1353				return -EINVAL;
1354			/*
1355			 * Ignore pgoff.
1356			 */
1357			pgoff = 0;
1358			vm_flags |= VM_SHARED | VM_MAYSHARE;
1359			break;
1360		case MAP_PRIVATE:
1361			/*
1362			 * Set pgoff according to addr for anon_vma.
1363			 */
1364			pgoff = addr >> PAGE_SHIFT;
1365			break;
1366		default:
1367			return -EINVAL;
1368		}
1369	}
1370
1371	/*
1372	 * Set 'VM_NORESERVE' if we should not account for the
1373	 * memory use of this mapping.
1374	 */
1375	if (flags & MAP_NORESERVE) {
1376		/* We honor MAP_NORESERVE if allowed to overcommit */
1377		if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1378			vm_flags |= VM_NORESERVE;
1379
1380		/* hugetlb applies strict overcommit unless MAP_NORESERVE */
1381		if (file && is_file_hugepages(file))
1382			vm_flags |= VM_NORESERVE;
1383	}
1384
1385	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
1386	if (!IS_ERR_VALUE(addr) &&
1387	    ((vm_flags & VM_LOCKED) ||
1388	     (flags & (MAP_POPULATE | MAP_NONBLOCK)) == MAP_POPULATE))
1389		*populate = len;
1390	return addr;
1391}
1392
1393unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1394			      unsigned long prot, unsigned long flags,
1395			      unsigned long fd, unsigned long pgoff)
1396{
1397	struct file *file = NULL;
1398	unsigned long retval;
1399
1400	if (!(flags & MAP_ANONYMOUS)) {
1401		audit_mmap_fd(fd, flags);
1402		file = fget(fd);
1403		if (!file)
1404			return -EBADF;
1405		if (is_file_hugepages(file)) {
1406			len = ALIGN(len, huge_page_size(hstate_file(file)));
1407		} else if (unlikely(flags & MAP_HUGETLB)) {
1408			retval = -EINVAL;
1409			goto out_fput;
1410		}
1411	} else if (flags & MAP_HUGETLB) {
1412		struct hstate *hs;
1413
1414		hs = hstate_sizelog((flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1415		if (!hs)
1416			return -EINVAL;
1417
1418		len = ALIGN(len, huge_page_size(hs));
1419		/*
1420		 * VM_NORESERVE is used because the reservations will be
1421		 * taken when vm_ops->mmap() is called
1422		 */
1423		file = hugetlb_file_setup(HUGETLB_ANON_FILE, len,
1424				VM_NORESERVE,
1425				HUGETLB_ANONHUGE_INODE,
1426				(flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1427		if (IS_ERR(file))
1428			return PTR_ERR(file);
1429	}
1430
1431	retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1432out_fput:
1433	if (file)
1434		fput(file);
1435	return retval;
1436}
1437
1438SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1439		unsigned long, prot, unsigned long, flags,
1440		unsigned long, fd, unsigned long, pgoff)
1441{
1442	return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1443}
1444
1445#ifdef __ARCH_WANT_SYS_OLD_MMAP
1446struct mmap_arg_struct {
1447	unsigned long addr;
1448	unsigned long len;
1449	unsigned long prot;
1450	unsigned long flags;
1451	unsigned long fd;
1452	unsigned long offset;
1453};
1454
1455SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1456{
1457	struct mmap_arg_struct a;
1458
1459	if (copy_from_user(&a, arg, sizeof(a)))
1460		return -EFAULT;
1461	if (offset_in_page(a.offset))
1462		return -EINVAL;
1463
1464	return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1465			       a.offset >> PAGE_SHIFT);
1466}
1467#endif /* __ARCH_WANT_SYS_OLD_MMAP */
1468
1469static bool vm_ops_needs_writenotify(const struct vm_operations_struct *vm_ops)
1470{
1471	return vm_ops && (vm_ops->page_mkwrite || vm_ops->pfn_mkwrite);
1472}
1473
1474static bool vma_is_shared_writable(struct vm_area_struct *vma)
1475{
1476	return (vma->vm_flags & (VM_WRITE | VM_SHARED)) ==
1477		(VM_WRITE | VM_SHARED);
1478}
1479
1480static bool vma_fs_can_writeback(struct vm_area_struct *vma)
1481{
1482	/* No managed pages to writeback. */
1483	if (vma->vm_flags & VM_PFNMAP)
1484		return false;
1485
1486	return vma->vm_file && vma->vm_file->f_mapping &&
1487		mapping_can_writeback(vma->vm_file->f_mapping);
1488}
1489
1490/*
1491 * Does this VMA require the underlying folios to have their dirty state
1492 * tracked?
1493 */
1494bool vma_needs_dirty_tracking(struct vm_area_struct *vma)
1495{
1496	/* Only shared, writable VMAs require dirty tracking. */
1497	if (!vma_is_shared_writable(vma))
1498		return false;
1499
1500	/* Does the filesystem need to be notified? */
1501	if (vm_ops_needs_writenotify(vma->vm_ops))
1502		return true;
1503
1504	/*
1505	 * Even if the filesystem doesn't indicate a need for writenotify, if it
1506	 * can writeback, dirty tracking is still required.
1507	 */
1508	return vma_fs_can_writeback(vma);
1509}
1510
1511/*
1512 * Some shared mappings will want the pages marked read-only
1513 * to track write events. If so, we'll downgrade vm_page_prot
1514 * to the private version (using protection_map[] without the
1515 * VM_SHARED bit).
1516 */
1517int vma_wants_writenotify(struct vm_area_struct *vma, pgprot_t vm_page_prot)
1518{
 
 
 
1519	/* If it was private or non-writable, the write bit is already clear */
1520	if (!vma_is_shared_writable(vma))
1521		return 0;
1522
1523	/* The backer wishes to know when pages are first written to? */
1524	if (vm_ops_needs_writenotify(vma->vm_ops))
1525		return 1;
1526
1527	/* The open routine did something to the protections that pgprot_modify
1528	 * won't preserve? */
1529	if (pgprot_val(vm_page_prot) !=
1530	    pgprot_val(vm_pgprot_modify(vm_page_prot, vma->vm_flags)))
1531		return 0;
1532
1533	/*
1534	 * Do we need to track softdirty? hugetlb does not support softdirty
1535	 * tracking yet.
1536	 */
1537	if (vma_soft_dirty_enabled(vma) && !is_vm_hugetlb_page(vma))
1538		return 1;
1539
1540	/* Do we need write faults for uffd-wp tracking? */
1541	if (userfaultfd_wp(vma))
1542		return 1;
1543
 
 
 
 
1544	/* Can the mapping track the dirty pages? */
1545	return vma_fs_can_writeback(vma);
 
1546}
1547
1548/*
1549 * We account for memory if it's a private writeable mapping,
1550 * not hugepages and VM_NORESERVE wasn't set.
1551 */
1552static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1553{
1554	/*
1555	 * hugetlb has its own accounting separate from the core VM
1556	 * VM_HUGETLB may not be set yet so we cannot check for that flag.
1557	 */
1558	if (file && is_file_hugepages(file))
1559		return 0;
1560
1561	return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1562}
1563
1564/**
1565 * unmapped_area() - Find an area between the low_limit and the high_limit with
1566 * the correct alignment and offset, all from @info. Note: current->mm is used
1567 * for the search.
1568 *
1569 * @info: The unmapped area information including the range [low_limit -
1570 * high_limit), the alignment offset and mask.
1571 *
1572 * Return: A memory address or -ENOMEM.
1573 */
1574static unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1575{
1576	unsigned long length, gap;
1577	unsigned long low_limit, high_limit;
1578	struct vm_area_struct *tmp;
1579
1580	MA_STATE(mas, &current->mm->mm_mt, 0, 0);
1581
1582	/* Adjust search length to account for worst case alignment overhead */
1583	length = info->length + info->align_mask;
1584	if (length < info->length)
1585		return -ENOMEM;
1586
1587	low_limit = info->low_limit;
1588	if (low_limit < mmap_min_addr)
1589		low_limit = mmap_min_addr;
1590	high_limit = info->high_limit;
1591retry:
1592	if (mas_empty_area(&mas, low_limit, high_limit - 1, length))
1593		return -ENOMEM;
1594
1595	gap = mas.index;
1596	gap += (info->align_offset - gap) & info->align_mask;
1597	tmp = mas_next(&mas, ULONG_MAX);
1598	if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1599		if (vm_start_gap(tmp) < gap + length - 1) {
1600			low_limit = tmp->vm_end;
1601			mas_reset(&mas);
1602			goto retry;
1603		}
1604	} else {
1605		tmp = mas_prev(&mas, 0);
1606		if (tmp && vm_end_gap(tmp) > gap) {
1607			low_limit = vm_end_gap(tmp);
1608			mas_reset(&mas);
1609			goto retry;
1610		}
1611	}
1612
1613	return gap;
1614}
1615
1616/**
1617 * unmapped_area_topdown() - Find an area between the low_limit and the
1618 * high_limit with the correct alignment and offset at the highest available
1619 * address, all from @info. Note: current->mm is used for the search.
1620 *
1621 * @info: The unmapped area information including the range [low_limit -
1622 * high_limit), the alignment offset and mask.
1623 *
1624 * Return: A memory address or -ENOMEM.
1625 */
1626static unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1627{
1628	unsigned long length, gap, gap_end;
1629	unsigned long low_limit, high_limit;
1630	struct vm_area_struct *tmp;
1631
1632	MA_STATE(mas, &current->mm->mm_mt, 0, 0);
1633	/* Adjust search length to account for worst case alignment overhead */
1634	length = info->length + info->align_mask;
1635	if (length < info->length)
1636		return -ENOMEM;
1637
1638	low_limit = info->low_limit;
1639	if (low_limit < mmap_min_addr)
1640		low_limit = mmap_min_addr;
1641	high_limit = info->high_limit;
1642retry:
1643	if (mas_empty_area_rev(&mas, low_limit, high_limit - 1, length))
1644		return -ENOMEM;
1645
1646	gap = mas.last + 1 - info->length;
1647	gap -= (gap - info->align_offset) & info->align_mask;
1648	gap_end = mas.last;
1649	tmp = mas_next(&mas, ULONG_MAX);
1650	if (tmp && (tmp->vm_flags & VM_STARTGAP_FLAGS)) { /* Avoid prev check if possible */
1651		if (vm_start_gap(tmp) <= gap_end) {
1652			high_limit = vm_start_gap(tmp);
1653			mas_reset(&mas);
1654			goto retry;
1655		}
1656	} else {
1657		tmp = mas_prev(&mas, 0);
1658		if (tmp && vm_end_gap(tmp) > gap) {
1659			high_limit = tmp->vm_start;
1660			mas_reset(&mas);
1661			goto retry;
1662		}
1663	}
1664
1665	return gap;
1666}
1667
1668/*
1669 * Search for an unmapped address range.
1670 *
1671 * We are looking for a range that:
1672 * - does not intersect with any VMA;
1673 * - is contained within the [low_limit, high_limit) interval;
1674 * - is at least the desired size.
1675 * - satisfies (begin_addr & align_mask) == (align_offset & align_mask)
1676 */
1677unsigned long vm_unmapped_area(struct vm_unmapped_area_info *info)
1678{
1679	unsigned long addr;
1680
1681	if (info->flags & VM_UNMAPPED_AREA_TOPDOWN)
1682		addr = unmapped_area_topdown(info);
1683	else
1684		addr = unmapped_area(info);
1685
1686	trace_vm_unmapped_area(addr, info);
1687	return addr;
1688}
1689
1690/* Get an address range which is currently unmapped.
1691 * For shmat() with addr=0.
1692 *
1693 * Ugly calling convention alert:
1694 * Return value with the low bits set means error value,
1695 * ie
1696 *	if (ret & ~PAGE_MASK)
1697 *		error = ret;
1698 *
1699 * This function "knows" that -ENOMEM has the bits set.
1700 */
1701unsigned long
1702generic_get_unmapped_area(struct file *filp, unsigned long addr,
1703			  unsigned long len, unsigned long pgoff,
1704			  unsigned long flags)
1705{
1706	struct mm_struct *mm = current->mm;
1707	struct vm_area_struct *vma, *prev;
1708	struct vm_unmapped_area_info info;
1709	const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1710
1711	if (len > mmap_end - mmap_min_addr)
1712		return -ENOMEM;
1713
1714	if (flags & MAP_FIXED)
1715		return addr;
1716
1717	if (addr) {
1718		addr = PAGE_ALIGN(addr);
1719		vma = find_vma_prev(mm, addr, &prev);
1720		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1721		    (!vma || addr + len <= vm_start_gap(vma)) &&
1722		    (!prev || addr >= vm_end_gap(prev)))
1723			return addr;
1724	}
1725
1726	info.flags = 0;
1727	info.length = len;
1728	info.low_limit = mm->mmap_base;
1729	info.high_limit = mmap_end;
1730	info.align_mask = 0;
1731	info.align_offset = 0;
1732	return vm_unmapped_area(&info);
1733}
1734
1735#ifndef HAVE_ARCH_UNMAPPED_AREA
1736unsigned long
1737arch_get_unmapped_area(struct file *filp, unsigned long addr,
1738		       unsigned long len, unsigned long pgoff,
1739		       unsigned long flags)
1740{
1741	return generic_get_unmapped_area(filp, addr, len, pgoff, flags);
1742}
1743#endif
1744
1745/*
1746 * This mmap-allocator allocates new areas top-down from below the
1747 * stack's low limit (the base):
1748 */
1749unsigned long
1750generic_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1751				  unsigned long len, unsigned long pgoff,
1752				  unsigned long flags)
1753{
1754	struct vm_area_struct *vma, *prev;
1755	struct mm_struct *mm = current->mm;
1756	struct vm_unmapped_area_info info;
1757	const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
1758
1759	/* requested length too big for entire address space */
1760	if (len > mmap_end - mmap_min_addr)
1761		return -ENOMEM;
1762
1763	if (flags & MAP_FIXED)
1764		return addr;
1765
1766	/* requesting a specific address */
1767	if (addr) {
1768		addr = PAGE_ALIGN(addr);
1769		vma = find_vma_prev(mm, addr, &prev);
1770		if (mmap_end - len >= addr && addr >= mmap_min_addr &&
1771				(!vma || addr + len <= vm_start_gap(vma)) &&
1772				(!prev || addr >= vm_end_gap(prev)))
1773			return addr;
1774	}
1775
1776	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1777	info.length = len;
1778	info.low_limit = PAGE_SIZE;
1779	info.high_limit = arch_get_mmap_base(addr, mm->mmap_base);
1780	info.align_mask = 0;
1781	info.align_offset = 0;
1782	addr = vm_unmapped_area(&info);
1783
1784	/*
1785	 * A failed mmap() very likely causes application failure,
1786	 * so fall back to the bottom-up function here. This scenario
1787	 * can happen with large stack limits and large mmap()
1788	 * allocations.
1789	 */
1790	if (offset_in_page(addr)) {
1791		VM_BUG_ON(addr != -ENOMEM);
1792		info.flags = 0;
1793		info.low_limit = TASK_UNMAPPED_BASE;
1794		info.high_limit = mmap_end;
1795		addr = vm_unmapped_area(&info);
1796	}
1797
1798	return addr;
1799}
1800
1801#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1802unsigned long
1803arch_get_unmapped_area_topdown(struct file *filp, unsigned long addr,
1804			       unsigned long len, unsigned long pgoff,
1805			       unsigned long flags)
1806{
1807	return generic_get_unmapped_area_topdown(filp, addr, len, pgoff, flags);
1808}
1809#endif
1810
1811unsigned long
1812get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1813		unsigned long pgoff, unsigned long flags)
1814{
1815	unsigned long (*get_area)(struct file *, unsigned long,
1816				  unsigned long, unsigned long, unsigned long);
1817
1818	unsigned long error = arch_mmap_check(addr, len, flags);
1819	if (error)
1820		return error;
1821
1822	/* Careful about overflows.. */
1823	if (len > TASK_SIZE)
1824		return -ENOMEM;
1825
1826	get_area = current->mm->get_unmapped_area;
1827	if (file) {
1828		if (file->f_op->get_unmapped_area)
1829			get_area = file->f_op->get_unmapped_area;
1830	} else if (flags & MAP_SHARED) {
1831		/*
1832		 * mmap_region() will call shmem_zero_setup() to create a file,
1833		 * so use shmem's get_unmapped_area in case it can be huge.
 
1834		 */
 
1835		get_area = shmem_get_unmapped_area;
1836	} else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
1837		/* Ensures that larger anonymous mappings are THP aligned. */
1838		get_area = thp_get_unmapped_area;
1839	}
1840
1841	/* Always treat pgoff as zero for anonymous memory. */
1842	if (!file)
1843		pgoff = 0;
1844
1845	addr = get_area(file, addr, len, pgoff, flags);
1846	if (IS_ERR_VALUE(addr))
1847		return addr;
1848
1849	if (addr > TASK_SIZE - len)
1850		return -ENOMEM;
1851	if (offset_in_page(addr))
1852		return -EINVAL;
1853
1854	error = security_mmap_addr(addr);
1855	return error ? error : addr;
1856}
1857
1858EXPORT_SYMBOL(get_unmapped_area);
1859
1860/**
1861 * find_vma_intersection() - Look up the first VMA which intersects the interval
1862 * @mm: The process address space.
1863 * @start_addr: The inclusive start user address.
1864 * @end_addr: The exclusive end user address.
1865 *
1866 * Returns: The first VMA within the provided range, %NULL otherwise.  Assumes
1867 * start_addr < end_addr.
1868 */
1869struct vm_area_struct *find_vma_intersection(struct mm_struct *mm,
1870					     unsigned long start_addr,
1871					     unsigned long end_addr)
1872{
1873	unsigned long index = start_addr;
1874
1875	mmap_assert_locked(mm);
1876	return mt_find(&mm->mm_mt, &index, end_addr - 1);
1877}
1878EXPORT_SYMBOL(find_vma_intersection);
1879
1880/**
1881 * find_vma() - Find the VMA for a given address, or the next VMA.
1882 * @mm: The mm_struct to check
1883 * @addr: The address
1884 *
1885 * Returns: The VMA associated with addr, or the next VMA.
1886 * May return %NULL in the case of no VMA at addr or above.
1887 */
1888struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1889{
1890	unsigned long index = addr;
1891
1892	mmap_assert_locked(mm);
1893	return mt_find(&mm->mm_mt, &index, ULONG_MAX);
1894}
1895EXPORT_SYMBOL(find_vma);
1896
1897/**
1898 * find_vma_prev() - Find the VMA for a given address, or the next vma and
1899 * set %pprev to the previous VMA, if any.
1900 * @mm: The mm_struct to check
1901 * @addr: The address
1902 * @pprev: The pointer to set to the previous VMA
1903 *
1904 * Note that RCU lock is missing here since the external mmap_lock() is used
1905 * instead.
1906 *
1907 * Returns: The VMA associated with @addr, or the next vma.
1908 * May return %NULL in the case of no vma at addr or above.
1909 */
1910struct vm_area_struct *
1911find_vma_prev(struct mm_struct *mm, unsigned long addr,
1912			struct vm_area_struct **pprev)
1913{
1914	struct vm_area_struct *vma;
1915	MA_STATE(mas, &mm->mm_mt, addr, addr);
1916
1917	vma = mas_walk(&mas);
1918	*pprev = mas_prev(&mas, 0);
1919	if (!vma)
1920		vma = mas_next(&mas, ULONG_MAX);
1921	return vma;
1922}
1923
1924/*
1925 * Verify that the stack growth is acceptable and
1926 * update accounting. This is shared with both the
1927 * grow-up and grow-down cases.
1928 */
1929static int acct_stack_growth(struct vm_area_struct *vma,
1930			     unsigned long size, unsigned long grow)
1931{
1932	struct mm_struct *mm = vma->vm_mm;
1933	unsigned long new_start;
1934
1935	/* address space limit tests */
1936	if (!may_expand_vm(mm, vma->vm_flags, grow))
1937		return -ENOMEM;
1938
1939	/* Stack limit test */
1940	if (size > rlimit(RLIMIT_STACK))
1941		return -ENOMEM;
1942
1943	/* mlock limit tests */
1944	if (!mlock_future_ok(mm, vma->vm_flags, grow << PAGE_SHIFT))
1945		return -ENOMEM;
1946
1947	/* Check to ensure the stack will not grow into a hugetlb-only region */
1948	new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1949			vma->vm_end - size;
1950	if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1951		return -EFAULT;
1952
1953	/*
1954	 * Overcommit..  This must be the final test, as it will
1955	 * update security statistics.
1956	 */
1957	if (security_vm_enough_memory_mm(mm, grow))
1958		return -ENOMEM;
1959
1960	return 0;
1961}
1962
1963#if defined(CONFIG_STACK_GROWSUP)
1964/*
1965 * PA-RISC uses this for its stack.
1966 * vma is the last one with address > vma->vm_end.  Have to extend vma.
1967 */
1968static int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1969{
1970	struct mm_struct *mm = vma->vm_mm;
1971	struct vm_area_struct *next;
1972	unsigned long gap_addr;
1973	int error = 0;
1974	MA_STATE(mas, &mm->mm_mt, vma->vm_start, address);
1975
1976	if (!(vma->vm_flags & VM_GROWSUP))
1977		return -EFAULT;
1978
1979	/* Guard against exceeding limits of the address space. */
1980	address &= PAGE_MASK;
1981	if (address >= (TASK_SIZE & PAGE_MASK))
1982		return -ENOMEM;
1983	address += PAGE_SIZE;
1984
1985	/* Enforce stack_guard_gap */
1986	gap_addr = address + stack_guard_gap;
1987
1988	/* Guard against overflow */
1989	if (gap_addr < address || gap_addr > TASK_SIZE)
1990		gap_addr = TASK_SIZE;
1991
1992	next = find_vma_intersection(mm, vma->vm_end, gap_addr);
1993	if (next && vma_is_accessible(next)) {
1994		if (!(next->vm_flags & VM_GROWSUP))
1995			return -ENOMEM;
1996		/* Check that both stack segments have the same anon_vma? */
1997	}
1998
1999	if (next)
2000		mas_prev_range(&mas, address);
2001
2002	__mas_set_range(&mas, vma->vm_start, address - 1);
2003	if (mas_preallocate(&mas, vma, GFP_KERNEL))
2004		return -ENOMEM;
2005
2006	/* We must make sure the anon_vma is allocated. */
2007	if (unlikely(anon_vma_prepare(vma))) {
2008		mas_destroy(&mas);
2009		return -ENOMEM;
2010	}
2011
2012	/* Lock the VMA before expanding to prevent concurrent page faults */
2013	vma_start_write(vma);
2014	/*
2015	 * vma->vm_start/vm_end cannot change under us because the caller
2016	 * is required to hold the mmap_lock in read mode.  We need the
2017	 * anon_vma lock to serialize against concurrent expand_stacks.
2018	 */
2019	anon_vma_lock_write(vma->anon_vma);
2020
2021	/* Somebody else might have raced and expanded it already */
2022	if (address > vma->vm_end) {
2023		unsigned long size, grow;
2024
2025		size = address - vma->vm_start;
2026		grow = (address - vma->vm_end) >> PAGE_SHIFT;
2027
2028		error = -ENOMEM;
2029		if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2030			error = acct_stack_growth(vma, size, grow);
2031			if (!error) {
2032				/*
2033				 * We only hold a shared mmap_lock lock here, so
2034				 * we need to protect against concurrent vma
2035				 * expansions.  anon_vma_lock_write() doesn't
2036				 * help here, as we don't guarantee that all
2037				 * growable vmas in a mm share the same root
2038				 * anon vma.  So, we reuse mm->page_table_lock
2039				 * to guard against concurrent vma expansions.
2040				 */
2041				spin_lock(&mm->page_table_lock);
2042				if (vma->vm_flags & VM_LOCKED)
2043					mm->locked_vm += grow;
2044				vm_stat_account(mm, vma->vm_flags, grow);
2045				anon_vma_interval_tree_pre_update_vma(vma);
2046				vma->vm_end = address;
2047				/* Overwrite old entry in mtree. */
2048				mas_store_prealloc(&mas, vma);
2049				anon_vma_interval_tree_post_update_vma(vma);
2050				spin_unlock(&mm->page_table_lock);
2051
2052				perf_event_mmap(vma);
2053			}
2054		}
2055	}
2056	anon_vma_unlock_write(vma->anon_vma);
 
2057	mas_destroy(&mas);
2058	validate_mm(mm);
2059	return error;
2060}
2061#endif /* CONFIG_STACK_GROWSUP */
2062
2063/*
2064 * vma is the first one with address < vma->vm_start.  Have to extend vma.
2065 * mmap_lock held for writing.
2066 */
2067int expand_downwards(struct vm_area_struct *vma, unsigned long address)
2068{
2069	struct mm_struct *mm = vma->vm_mm;
2070	MA_STATE(mas, &mm->mm_mt, vma->vm_start, vma->vm_start);
2071	struct vm_area_struct *prev;
2072	int error = 0;
2073
2074	if (!(vma->vm_flags & VM_GROWSDOWN))
2075		return -EFAULT;
2076
2077	address &= PAGE_MASK;
2078	if (address < mmap_min_addr || address < FIRST_USER_ADDRESS)
2079		return -EPERM;
2080
2081	/* Enforce stack_guard_gap */
2082	prev = mas_prev(&mas, 0);
2083	/* Check that both stack segments have the same anon_vma? */
2084	if (prev) {
2085		if (!(prev->vm_flags & VM_GROWSDOWN) &&
2086		    vma_is_accessible(prev) &&
2087		    (address - prev->vm_end < stack_guard_gap))
2088			return -ENOMEM;
2089	}
2090
2091	if (prev)
2092		mas_next_range(&mas, vma->vm_start);
2093
2094	__mas_set_range(&mas, address, vma->vm_end - 1);
2095	if (mas_preallocate(&mas, vma, GFP_KERNEL))
2096		return -ENOMEM;
2097
2098	/* We must make sure the anon_vma is allocated. */
2099	if (unlikely(anon_vma_prepare(vma))) {
2100		mas_destroy(&mas);
2101		return -ENOMEM;
2102	}
2103
2104	/* Lock the VMA before expanding to prevent concurrent page faults */
2105	vma_start_write(vma);
2106	/*
2107	 * vma->vm_start/vm_end cannot change under us because the caller
2108	 * is required to hold the mmap_lock in read mode.  We need the
2109	 * anon_vma lock to serialize against concurrent expand_stacks.
2110	 */
2111	anon_vma_lock_write(vma->anon_vma);
2112
2113	/* Somebody else might have raced and expanded it already */
2114	if (address < vma->vm_start) {
2115		unsigned long size, grow;
2116
2117		size = vma->vm_end - address;
2118		grow = (vma->vm_start - address) >> PAGE_SHIFT;
2119
2120		error = -ENOMEM;
2121		if (grow <= vma->vm_pgoff) {
2122			error = acct_stack_growth(vma, size, grow);
2123			if (!error) {
2124				/*
2125				 * We only hold a shared mmap_lock lock here, so
2126				 * we need to protect against concurrent vma
2127				 * expansions.  anon_vma_lock_write() doesn't
2128				 * help here, as we don't guarantee that all
2129				 * growable vmas in a mm share the same root
2130				 * anon vma.  So, we reuse mm->page_table_lock
2131				 * to guard against concurrent vma expansions.
2132				 */
2133				spin_lock(&mm->page_table_lock);
2134				if (vma->vm_flags & VM_LOCKED)
2135					mm->locked_vm += grow;
2136				vm_stat_account(mm, vma->vm_flags, grow);
2137				anon_vma_interval_tree_pre_update_vma(vma);
2138				vma->vm_start = address;
2139				vma->vm_pgoff -= grow;
2140				/* Overwrite old entry in mtree. */
2141				mas_store_prealloc(&mas, vma);
2142				anon_vma_interval_tree_post_update_vma(vma);
2143				spin_unlock(&mm->page_table_lock);
2144
2145				perf_event_mmap(vma);
2146			}
2147		}
2148	}
2149	anon_vma_unlock_write(vma->anon_vma);
 
2150	mas_destroy(&mas);
2151	validate_mm(mm);
2152	return error;
2153}
2154
2155/* enforced gap between the expanding stack and other mappings. */
2156unsigned long stack_guard_gap = 256UL<<PAGE_SHIFT;
2157
2158static int __init cmdline_parse_stack_guard_gap(char *p)
2159{
2160	unsigned long val;
2161	char *endptr;
2162
2163	val = simple_strtoul(p, &endptr, 10);
2164	if (!*endptr)
2165		stack_guard_gap = val << PAGE_SHIFT;
2166
2167	return 1;
2168}
2169__setup("stack_guard_gap=", cmdline_parse_stack_guard_gap);
2170
2171#ifdef CONFIG_STACK_GROWSUP
2172int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2173{
2174	return expand_upwards(vma, address);
2175}
2176
2177struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
 
2178{
2179	struct vm_area_struct *vma, *prev;
2180
2181	addr &= PAGE_MASK;
2182	vma = find_vma_prev(mm, addr, &prev);
2183	if (vma && (vma->vm_start <= addr))
2184		return vma;
2185	if (!prev)
2186		return NULL;
2187	if (expand_stack_locked(prev, addr))
2188		return NULL;
2189	if (prev->vm_flags & VM_LOCKED)
2190		populate_vma_page_range(prev, addr, prev->vm_end, NULL);
2191	return prev;
2192}
2193#else
2194int expand_stack_locked(struct vm_area_struct *vma, unsigned long address)
2195{
2196	return expand_downwards(vma, address);
2197}
2198
2199struct vm_area_struct *find_extend_vma_locked(struct mm_struct *mm, unsigned long addr)
 
2200{
2201	struct vm_area_struct *vma;
2202	unsigned long start;
2203
2204	addr &= PAGE_MASK;
2205	vma = find_vma(mm, addr);
2206	if (!vma)
2207		return NULL;
2208	if (vma->vm_start <= addr)
2209		return vma;
 
 
2210	start = vma->vm_start;
2211	if (expand_stack_locked(vma, addr))
2212		return NULL;
2213	if (vma->vm_flags & VM_LOCKED)
2214		populate_vma_page_range(vma, addr, start, NULL);
2215	return vma;
2216}
2217#endif
2218
2219#if defined(CONFIG_STACK_GROWSUP)
2220
2221#define vma_expand_up(vma,addr) expand_upwards(vma, addr)
2222#define vma_expand_down(vma, addr) (-EFAULT)
2223
2224#else
2225
2226#define vma_expand_up(vma,addr) (-EFAULT)
2227#define vma_expand_down(vma, addr) expand_downwards(vma, addr)
2228
2229#endif
2230
2231/*
2232 * expand_stack(): legacy interface for page faulting. Don't use unless
2233 * you have to.
2234 *
2235 * This is called with the mm locked for reading, drops the lock, takes
2236 * the lock for writing, tries to look up a vma again, expands it if
2237 * necessary, and downgrades the lock to reading again.
2238 *
2239 * If no vma is found or it can't be expanded, it returns NULL and has
2240 * dropped the lock.
2241 */
2242struct vm_area_struct *expand_stack(struct mm_struct *mm, unsigned long addr)
2243{
2244	struct vm_area_struct *vma, *prev;
2245
2246	mmap_read_unlock(mm);
2247	if (mmap_write_lock_killable(mm))
2248		return NULL;
2249
2250	vma = find_vma_prev(mm, addr, &prev);
2251	if (vma && vma->vm_start <= addr)
2252		goto success;
2253
2254	if (prev && !vma_expand_up(prev, addr)) {
2255		vma = prev;
2256		goto success;
2257	}
2258
2259	if (vma && !vma_expand_down(vma, addr))
2260		goto success;
2261
2262	mmap_write_unlock(mm);
2263	return NULL;
2264
2265success:
2266	mmap_write_downgrade(mm);
2267	return vma;
2268}
2269
2270/*
2271 * Ok - we have the memory areas we should free on a maple tree so release them,
2272 * and do the vma updates.
2273 *
2274 * Called with the mm semaphore held.
2275 */
2276static inline void remove_mt(struct mm_struct *mm, struct ma_state *mas)
2277{
2278	unsigned long nr_accounted = 0;
2279	struct vm_area_struct *vma;
2280
2281	/* Update high watermark before we lower total_vm */
2282	update_hiwater_vm(mm);
2283	mas_for_each(mas, vma, ULONG_MAX) {
2284		long nrpages = vma_pages(vma);
2285
2286		if (vma->vm_flags & VM_ACCOUNT)
2287			nr_accounted += nrpages;
2288		vm_stat_account(mm, vma->vm_flags, -nrpages);
2289		remove_vma(vma, false);
2290	}
2291	vm_unacct_memory(nr_accounted);
 
2292}
2293
2294/*
2295 * Get rid of page table information in the indicated region.
2296 *
2297 * Called with the mm semaphore held.
2298 */
2299static void unmap_region(struct mm_struct *mm, struct ma_state *mas,
2300		struct vm_area_struct *vma, struct vm_area_struct *prev,
2301		struct vm_area_struct *next, unsigned long start,
2302		unsigned long end, unsigned long tree_end, bool mm_wr_locked)
2303{
2304	struct mmu_gather tlb;
2305	unsigned long mt_start = mas->index;
2306
2307	lru_add_drain();
2308	tlb_gather_mmu(&tlb, mm);
2309	update_hiwater_rss(mm);
2310	unmap_vmas(&tlb, mas, vma, start, end, tree_end, mm_wr_locked);
2311	mas_set(mas, mt_start);
2312	free_pgtables(&tlb, mas, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2313				 next ? next->vm_start : USER_PGTABLES_CEILING,
2314				 mm_wr_locked);
2315	tlb_finish_mmu(&tlb);
2316}
2317
2318/*
2319 * __split_vma() bypasses sysctl_max_map_count checking.  We use this where it
2320 * has already been checked or doesn't make sense to fail.
2321 * VMA Iterator will point to the end VMA.
2322 */
2323static int __split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2324		       unsigned long addr, int new_below)
2325{
2326	struct vma_prepare vp;
2327	struct vm_area_struct *new;
2328	int err;
2329
2330	WARN_ON(vma->vm_start >= addr);
2331	WARN_ON(vma->vm_end <= addr);
2332
2333	if (vma->vm_ops && vma->vm_ops->may_split) {
2334		err = vma->vm_ops->may_split(vma, addr);
2335		if (err)
2336			return err;
2337	}
2338
2339	new = vm_area_dup(vma);
2340	if (!new)
2341		return -ENOMEM;
2342
2343	if (new_below) {
2344		new->vm_end = addr;
2345	} else {
2346		new->vm_start = addr;
2347		new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2348	}
2349
2350	err = -ENOMEM;
2351	vma_iter_config(vmi, new->vm_start, new->vm_end);
2352	if (vma_iter_prealloc(vmi, new))
2353		goto out_free_vma;
2354
2355	err = vma_dup_policy(vma, new);
2356	if (err)
2357		goto out_free_vmi;
2358
2359	err = anon_vma_clone(new, vma);
2360	if (err)
2361		goto out_free_mpol;
2362
2363	if (new->vm_file)
2364		get_file(new->vm_file);
2365
2366	if (new->vm_ops && new->vm_ops->open)
2367		new->vm_ops->open(new);
2368
2369	vma_start_write(vma);
2370	vma_start_write(new);
2371
2372	init_vma_prep(&vp, vma);
2373	vp.insert = new;
2374	vma_prepare(&vp);
2375	vma_adjust_trans_huge(vma, vma->vm_start, addr, 0);
2376
2377	if (new_below) {
2378		vma->vm_start = addr;
2379		vma->vm_pgoff += (addr - new->vm_start) >> PAGE_SHIFT;
2380	} else {
2381		vma->vm_end = addr;
2382	}
2383
2384	/* vma_complete stores the new vma */
2385	vma_complete(&vp, vmi, vma->vm_mm);
2386
2387	/* Success. */
2388	if (new_below)
2389		vma_next(vmi);
2390	return 0;
2391
2392out_free_mpol:
 
 
 
 
 
 
 
 
 
2393	mpol_put(vma_policy(new));
2394out_free_vmi:
2395	vma_iter_free(vmi);
2396out_free_vma:
2397	vm_area_free(new);
 
2398	return err;
2399}
2400
2401/*
2402 * Split a vma into two pieces at address 'addr', a new vma is allocated
2403 * either for the first part or the tail.
2404 */
2405static int split_vma(struct vma_iterator *vmi, struct vm_area_struct *vma,
2406		     unsigned long addr, int new_below)
2407{
2408	if (vma->vm_mm->map_count >= sysctl_max_map_count)
2409		return -ENOMEM;
2410
2411	return __split_vma(vmi, vma, addr, new_below);
2412}
2413
2414/*
2415 * We are about to modify one or multiple of a VMA's flags, policy, userfaultfd
2416 * context and anonymous VMA name within the range [start, end).
2417 *
2418 * As a result, we might be able to merge the newly modified VMA range with an
2419 * adjacent VMA with identical properties.
2420 *
2421 * If no merge is possible and the range does not span the entirety of the VMA,
2422 * we then need to split the VMA to accommodate the change.
2423 *
2424 * The function returns either the merged VMA, the original VMA if a split was
2425 * required instead, or an error if the split failed.
2426 */
2427struct vm_area_struct *vma_modify(struct vma_iterator *vmi,
2428				  struct vm_area_struct *prev,
2429				  struct vm_area_struct *vma,
2430				  unsigned long start, unsigned long end,
2431				  unsigned long vm_flags,
2432				  struct mempolicy *policy,
2433				  struct vm_userfaultfd_ctx uffd_ctx,
2434				  struct anon_vma_name *anon_name)
2435{
2436	pgoff_t pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
2437	struct vm_area_struct *merged;
 
2438
2439	merged = vma_merge(vmi, prev, vma, start, end, vm_flags,
2440			   pgoff, policy, uffd_ctx, anon_name);
2441	if (merged)
2442		return merged;
2443
2444	if (vma->vm_start < start) {
2445		int err = split_vma(vmi, vma, start, 1);
2446
2447		if (err)
2448			return ERR_PTR(err);
2449	}
2450
2451	if (vma->vm_end > end) {
2452		int err = split_vma(vmi, vma, end, 0);
2453
2454		if (err)
2455			return ERR_PTR(err);
2456	}
2457
2458	return vma;
2459}
2460
2461/*
2462 * Attempt to merge a newly mapped VMA with those adjacent to it. The caller
2463 * must ensure that [start, end) does not overlap any existing VMA.
2464 */
2465static struct vm_area_struct
2466*vma_merge_new_vma(struct vma_iterator *vmi, struct vm_area_struct *prev,
2467		   struct vm_area_struct *vma, unsigned long start,
2468		   unsigned long end, pgoff_t pgoff)
2469{
2470	return vma_merge(vmi, prev, vma, start, end, vma->vm_flags, pgoff,
2471			 vma_policy(vma), vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2472}
2473
2474/*
2475 * Expand vma by delta bytes, potentially merging with an immediately adjacent
2476 * VMA with identical properties.
2477 */
2478struct vm_area_struct *vma_merge_extend(struct vma_iterator *vmi,
2479					struct vm_area_struct *vma,
2480					unsigned long delta)
2481{
2482	pgoff_t pgoff = vma->vm_pgoff + vma_pages(vma);
2483
2484	/* vma is specified as prev, so case 1 or 2 will apply. */
2485	return vma_merge(vmi, vma, vma, vma->vm_end, vma->vm_end + delta,
2486			 vma->vm_flags, pgoff, vma_policy(vma),
2487			 vma->vm_userfaultfd_ctx, anon_vma_name(vma));
2488}
2489
2490/*
2491 * do_vmi_align_munmap() - munmap the aligned region from @start to @end.
2492 * @vmi: The vma iterator
2493 * @vma: The starting vm_area_struct
2494 * @mm: The mm_struct
2495 * @start: The aligned start address to munmap.
2496 * @end: The aligned end address to munmap.
2497 * @uf: The userfaultfd list_head
2498 * @unlock: Set to true to drop the mmap_lock.  unlocking only happens on
2499 * success.
2500 *
2501 * Return: 0 on success and drops the lock if so directed, error and leaves the
2502 * lock held otherwise.
2503 */
2504static int
2505do_vmi_align_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
2506		    struct mm_struct *mm, unsigned long start,
2507		    unsigned long end, struct list_head *uf, bool unlock)
2508{
2509	struct vm_area_struct *prev, *next = NULL;
2510	struct maple_tree mt_detach;
2511	int count = 0;
2512	int error = -ENOMEM;
2513	unsigned long locked_vm = 0;
2514	MA_STATE(mas_detach, &mt_detach, 0, 0);
2515	mt_init_flags(&mt_detach, vmi->mas.tree->ma_flags & MT_FLAGS_LOCK_MASK);
2516	mt_on_stack(mt_detach);
 
 
 
2517
 
2518	/*
2519	 * If we need to split any vma, do it now to save pain later.
2520	 *
2521	 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2522	 * unmapped vm_area_struct will remain in use: so lower split_vma
2523	 * places tmp vma above, and higher split_vma places tmp vma below.
2524	 */
2525
2526	/* Does it split the first one? */
2527	if (start > vma->vm_start) {
2528
2529		/*
2530		 * Make sure that map_count on return from munmap() will
2531		 * not exceed its limit; but let map_count go just above
2532		 * its limit temporarily, to help free resources as expected.
2533		 */
2534		if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2535			goto map_count_exceeded;
2536
2537		error = __split_vma(vmi, vma, start, 1);
 
 
 
 
2538		if (error)
2539			goto start_split_failed;
 
 
 
2540	}
2541
 
 
 
 
2542	/*
2543	 * Detach a range of VMAs from the mm. Using next as a temp variable as
2544	 * it is always overwritten.
2545	 */
2546	next = vma;
2547	do {
2548		/* Does it split the end? */
2549		if (next->vm_end > end) {
2550			error = __split_vma(vmi, next, end, 0);
 
 
2551			if (error)
2552				goto end_split_failed;
 
 
 
 
 
 
 
 
 
 
 
2553		}
2554		vma_start_write(next);
2555		mas_set(&mas_detach, count);
2556		error = mas_store_gfp(&mas_detach, next, GFP_KERNEL);
2557		if (error)
2558			goto munmap_gather_failed;
2559		vma_mark_detached(next, true);
2560		if (next->vm_flags & VM_LOCKED)
2561			locked_vm += vma_pages(next);
2562
2563		count++;
2564		if (unlikely(uf)) {
2565			/*
2566			 * If userfaultfd_unmap_prep returns an error the vmas
2567			 * will remain split, but userland will get a
2568			 * highly unexpected error anyway. This is no
2569			 * different than the case where the first of the two
2570			 * __split_vma fails, but we don't undo the first
2571			 * split, despite we could. This is unlikely enough
2572			 * failure that it's not worth optimizing it for.
2573			 */
2574			error = userfaultfd_unmap_prep(next, start, end, uf);
2575
2576			if (error)
2577				goto userfaultfd_error;
2578		}
2579#ifdef CONFIG_DEBUG_VM_MAPLE_TREE
2580		BUG_ON(next->vm_start < start);
2581		BUG_ON(next->vm_start > end);
2582#endif
2583	} for_each_vma_range(*vmi, next, end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2584
 
 
2585#if defined(CONFIG_DEBUG_VM_MAPLE_TREE)
2586	/* Make sure no VMAs are about to be lost. */
2587	{
2588		MA_STATE(test, &mt_detach, 0, 0);
2589		struct vm_area_struct *vma_mas, *vma_test;
2590		int test_count = 0;
2591
2592		vma_iter_set(vmi, start);
2593		rcu_read_lock();
2594		vma_test = mas_find(&test, count - 1);
2595		for_each_vma_range(*vmi, vma_mas, end) {
2596			BUG_ON(vma_mas != vma_test);
2597			test_count++;
2598			vma_test = mas_next(&test, count - 1);
2599		}
2600		rcu_read_unlock();
2601		BUG_ON(count != test_count);
 
2602	}
2603#endif
2604
2605	while (vma_iter_addr(vmi) > start)
2606		vma_iter_prev_range(vmi);
2607
2608	error = vma_iter_clear_gfp(vmi, start, end, GFP_KERNEL);
2609	if (error)
2610		goto clear_tree_failed;
2611
2612	/* Point of no return */
2613	mm->locked_vm -= locked_vm;
2614	mm->map_count -= count;
2615	if (unlock)
2616		mmap_write_downgrade(mm);
 
 
 
 
 
 
 
 
 
 
 
2617
2618	prev = vma_iter_prev_range(vmi);
2619	next = vma_next(vmi);
2620	if (next)
2621		vma_iter_prev_range(vmi);
2622
2623	/*
2624	 * We can free page tables without write-locking mmap_lock because VMAs
2625	 * were isolated before we downgraded mmap_lock.
2626	 */
2627	mas_set(&mas_detach, 1);
2628	unmap_region(mm, &mas_detach, vma, prev, next, start, end, count,
2629		     !unlock);
2630	/* Statistics and freeing VMAs */
2631	mas_set(&mas_detach, 0);
2632	remove_mt(mm, &mas_detach);
 
 
 
2633	validate_mm(mm);
2634	if (unlock)
2635		mmap_read_unlock(mm);
2636
2637	__mt_destroy(&mt_detach);
2638	return 0;
2639
2640clear_tree_failed:
2641userfaultfd_error:
2642munmap_gather_failed:
2643end_split_failed:
2644	mas_set(&mas_detach, 0);
2645	mas_for_each(&mas_detach, next, end)
2646		vma_mark_detached(next, false);
2647
2648	__mt_destroy(&mt_detach);
2649start_split_failed:
2650map_count_exceeded:
2651	validate_mm(mm);
2652	return error;
2653}
2654
2655/*
2656 * do_vmi_munmap() - munmap a given range.
2657 * @vmi: The vma iterator
2658 * @mm: The mm_struct
2659 * @start: The start address to munmap
2660 * @len: The length of the range to munmap
2661 * @uf: The userfaultfd list_head
2662 * @unlock: set to true if the user wants to drop the mmap_lock on success
 
2663 *
2664 * This function takes a @mas that is either pointing to the previous VMA or set
2665 * to MA_START and sets it up to remove the mapping(s).  The @len will be
2666 * aligned and any arch_unmap work will be preformed.
2667 *
2668 * Return: 0 on success and drops the lock if so directed, error and leaves the
2669 * lock held otherwise.
2670 */
2671int do_vmi_munmap(struct vma_iterator *vmi, struct mm_struct *mm,
2672		  unsigned long start, size_t len, struct list_head *uf,
2673		  bool unlock)
2674{
2675	unsigned long end;
2676	struct vm_area_struct *vma;
2677
2678	if ((offset_in_page(start)) || start > TASK_SIZE || len > TASK_SIZE-start)
2679		return -EINVAL;
2680
2681	end = start + PAGE_ALIGN(len);
2682	if (end == start)
2683		return -EINVAL;
2684
2685	 /* arch_unmap() might do unmaps itself.  */
2686	arch_unmap(mm, start, end);
2687
2688	/* Find the first overlapping VMA */
2689	vma = vma_find(vmi, end);
2690	if (!vma) {
2691		if (unlock)
2692			mmap_write_unlock(mm);
2693		return 0;
2694	}
2695
2696	return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
2697}
2698
2699/* do_munmap() - Wrapper function for non-maple tree aware do_munmap() calls.
2700 * @mm: The mm_struct
2701 * @start: The start address to munmap
2702 * @len: The length to be munmapped.
2703 * @uf: The userfaultfd list_head
2704 *
2705 * Return: 0 on success, error otherwise.
2706 */
2707int do_munmap(struct mm_struct *mm, unsigned long start, size_t len,
2708	      struct list_head *uf)
2709{
2710	VMA_ITERATOR(vmi, mm, start);
2711
2712	return do_vmi_munmap(&vmi, mm, start, len, uf, false);
2713}
2714
2715unsigned long mmap_region(struct file *file, unsigned long addr,
2716		unsigned long len, vm_flags_t vm_flags, unsigned long pgoff,
2717		struct list_head *uf)
2718{
2719	struct mm_struct *mm = current->mm;
2720	struct vm_area_struct *vma = NULL;
2721	struct vm_area_struct *next, *prev, *merge;
2722	pgoff_t pglen = len >> PAGE_SHIFT;
2723	unsigned long charged = 0;
2724	unsigned long end = addr + len;
2725	unsigned long merge_start = addr, merge_end = end;
2726	bool writable_file_mapping = false;
2727	pgoff_t vm_pgoff;
2728	int error;
2729	VMA_ITERATOR(vmi, mm, addr);
2730
2731	/* Check against address space limit. */
2732	if (!may_expand_vm(mm, vm_flags, len >> PAGE_SHIFT)) {
2733		unsigned long nr_pages;
2734
2735		/*
2736		 * MAP_FIXED may remove pages of mappings that intersects with
2737		 * requested mapping. Account for the pages it would unmap.
2738		 */
2739		nr_pages = count_vma_pages_range(mm, addr, end);
2740
2741		if (!may_expand_vm(mm, vm_flags,
2742					(len >> PAGE_SHIFT) - nr_pages))
2743			return -ENOMEM;
2744	}
2745
2746	/* Unmap any existing mapping in the area */
2747	if (do_vmi_munmap(&vmi, mm, addr, len, uf, false))
2748		return -ENOMEM;
2749
2750	/*
2751	 * Private writable mapping: check memory availability
2752	 */
2753	if (accountable_mapping(file, vm_flags)) {
2754		charged = len >> PAGE_SHIFT;
2755		if (security_vm_enough_memory_mm(mm, charged))
2756			return -ENOMEM;
2757		vm_flags |= VM_ACCOUNT;
2758	}
2759
2760	next = vma_next(&vmi);
2761	prev = vma_prev(&vmi);
2762	if (vm_flags & VM_SPECIAL) {
2763		if (prev)
2764			vma_iter_next_range(&vmi);
2765		goto cannot_expand;
2766	}
2767
2768	/* Attempt to expand an old mapping */
2769	/* Check next */
2770	if (next && next->vm_start == end && !vma_policy(next) &&
2771	    can_vma_merge_before(next, vm_flags, NULL, file, pgoff+pglen,
2772				 NULL_VM_UFFD_CTX, NULL)) {
2773		merge_end = next->vm_end;
2774		vma = next;
2775		vm_pgoff = next->vm_pgoff - pglen;
2776	}
2777
2778	/* Check prev */
2779	if (prev && prev->vm_end == addr && !vma_policy(prev) &&
2780	    (vma ? can_vma_merge_after(prev, vm_flags, vma->anon_vma, file,
2781				       pgoff, vma->vm_userfaultfd_ctx, NULL) :
2782		   can_vma_merge_after(prev, vm_flags, NULL, file, pgoff,
2783				       NULL_VM_UFFD_CTX, NULL))) {
2784		merge_start = prev->vm_start;
2785		vma = prev;
2786		vm_pgoff = prev->vm_pgoff;
2787	} else if (prev) {
2788		vma_iter_next_range(&vmi);
2789	}
2790
 
2791	/* Actually expand, if possible */
2792	if (vma &&
2793	    !vma_expand(&vmi, vma, merge_start, merge_end, vm_pgoff, next)) {
2794		khugepaged_enter_vma(vma, vm_flags);
2795		goto expanded;
2796	}
2797
2798	if (vma == prev)
2799		vma_iter_set(&vmi, addr);
2800cannot_expand:
2801
2802	/*
2803	 * Determine the object being mapped and call the appropriate
2804	 * specific mapper. the address has already been validated, but
2805	 * not unmapped, but the maps are removed from the list.
2806	 */
2807	vma = vm_area_alloc(mm);
2808	if (!vma) {
2809		error = -ENOMEM;
2810		goto unacct_error;
2811	}
2812
2813	vma_iter_config(&vmi, addr, end);
2814	vma_set_range(vma, addr, end, pgoff);
2815	vm_flags_init(vma, vm_flags);
2816	vma->vm_page_prot = vm_get_page_prot(vm_flags);
 
2817
2818	if (file) {
 
 
 
 
 
 
2819		vma->vm_file = get_file(file);
2820		error = call_mmap(file, vma);
2821		if (error)
2822			goto unmap_and_free_vma;
2823
2824		if (vma_is_shared_maywrite(vma)) {
2825			error = mapping_map_writable(file->f_mapping);
2826			if (error)
2827				goto close_and_free_vma;
2828
2829			writable_file_mapping = true;
2830		}
2831
2832		/*
2833		 * Expansion is handled above, merging is handled below.
2834		 * Drivers should not alter the address of the VMA.
2835		 */
2836		error = -EINVAL;
2837		if (WARN_ON((addr != vma->vm_start)))
2838			goto close_and_free_vma;
 
 
2839
2840		vma_iter_config(&vmi, addr, end);
2841		/*
2842		 * If vm_flags changed after call_mmap(), we should try merge
2843		 * vma again as we may succeed this time.
2844		 */
2845		if (unlikely(vm_flags != vma->vm_flags && prev)) {
2846			merge = vma_merge_new_vma(&vmi, prev, vma,
2847						  vma->vm_start, vma->vm_end,
2848						  vma->vm_pgoff);
2849			if (merge) {
2850				/*
2851				 * ->mmap() can change vma->vm_file and fput
2852				 * the original file. So fput the vma->vm_file
2853				 * here or we would add an extra fput for file
2854				 * and cause general protection fault
2855				 * ultimately.
2856				 */
2857				fput(vma->vm_file);
2858				vm_area_free(vma);
2859				vma = merge;
2860				/* Update vm_flags to pick up the change. */
2861				vm_flags = vma->vm_flags;
2862				goto unmap_writable;
2863			}
2864		}
2865
2866		vm_flags = vma->vm_flags;
2867	} else if (vm_flags & VM_SHARED) {
2868		error = shmem_zero_setup(vma);
2869		if (error)
2870			goto free_vma;
2871	} else {
2872		vma_set_anonymous(vma);
2873	}
2874
2875	if (map_deny_write_exec(vma, vma->vm_flags)) {
2876		error = -EACCES;
2877		goto close_and_free_vma;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2878	}
2879
2880	/* Allow architectures to sanity-check the vm_flags */
2881	error = -EINVAL;
2882	if (!arch_validate_flags(vma->vm_flags))
2883		goto close_and_free_vma;
2884
2885	error = -ENOMEM;
2886	if (vma_iter_prealloc(&vmi, vma))
2887		goto close_and_free_vma;
2888
2889	/* Lock the VMA since it is modified after insertion into VMA tree */
2890	vma_start_write(vma);
2891	vma_iter_store(&vmi, vma);
2892	mm->map_count++;
2893	vma_link_file(vma);
 
 
 
 
 
 
 
 
2894
2895	/*
2896	 * vma_merge() calls khugepaged_enter_vma() either, the below
2897	 * call covers the non-merge case.
2898	 */
2899	khugepaged_enter_vma(vma, vma->vm_flags);
2900
2901	/* Once vma denies write, undo our temporary denial count */
2902unmap_writable:
2903	if (writable_file_mapping)
2904		mapping_unmap_writable(file->f_mapping);
2905	file = vma->vm_file;
2906	ksm_add_vma(vma);
2907expanded:
2908	perf_event_mmap(vma);
2909
2910	vm_stat_account(mm, vm_flags, len >> PAGE_SHIFT);
2911	if (vm_flags & VM_LOCKED) {
2912		if ((vm_flags & VM_SPECIAL) || vma_is_dax(vma) ||
2913					is_vm_hugetlb_page(vma) ||
2914					vma == get_gate_vma(current->mm))
2915			vm_flags_clear(vma, VM_LOCKED_MASK);
2916		else
2917			mm->locked_vm += (len >> PAGE_SHIFT);
2918	}
2919
2920	if (file)
2921		uprobe_mmap(vma);
2922
2923	/*
2924	 * New (or expanded) vma always get soft dirty status.
2925	 * Otherwise user-space soft-dirty page tracker won't
2926	 * be able to distinguish situation when vma area unmapped,
2927	 * then new mapped in-place (which must be aimed as
2928	 * a completely new data area).
2929	 */
2930	vm_flags_set(vma, VM_SOFTDIRTY);
2931
2932	vma_set_page_prot(vma);
2933
2934	validate_mm(mm);
2935	return addr;
2936
2937close_and_free_vma:
2938	if (file && vma->vm_ops && vma->vm_ops->close)
2939		vma->vm_ops->close(vma);
2940
2941	if (file || vma->vm_file) {
2942unmap_and_free_vma:
2943		fput(vma->vm_file);
2944		vma->vm_file = NULL;
2945
2946		vma_iter_set(&vmi, vma->vm_end);
2947		/* Undo any partial mapping done by a device driver. */
2948		unmap_region(mm, &vmi.mas, vma, prev, next, vma->vm_start,
2949			     vma->vm_end, vma->vm_end, true);
2950	}
2951	if (writable_file_mapping)
2952		mapping_unmap_writable(file->f_mapping);
2953free_vma:
2954	vm_area_free(vma);
2955unacct_error:
2956	if (charged)
2957		vm_unacct_memory(charged);
2958	validate_mm(mm);
2959	return error;
2960}
2961
2962static int __vm_munmap(unsigned long start, size_t len, bool unlock)
2963{
2964	int ret;
2965	struct mm_struct *mm = current->mm;
2966	LIST_HEAD(uf);
2967	VMA_ITERATOR(vmi, mm, start);
2968
2969	if (mmap_write_lock_killable(mm))
2970		return -EINTR;
2971
2972	ret = do_vmi_munmap(&vmi, mm, start, len, &uf, unlock);
2973	if (ret || !unlock)
 
 
 
 
 
 
 
 
2974		mmap_write_unlock(mm);
2975
2976	userfaultfd_unmap_complete(mm, &uf);
2977	return ret;
2978}
2979
2980int vm_munmap(unsigned long start, size_t len)
2981{
2982	return __vm_munmap(start, len, false);
2983}
2984EXPORT_SYMBOL(vm_munmap);
2985
2986SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2987{
2988	addr = untagged_addr(addr);
2989	return __vm_munmap(addr, len, true);
2990}
2991
2992
2993/*
2994 * Emulation of deprecated remap_file_pages() syscall.
2995 */
2996SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
2997		unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
2998{
2999
3000	struct mm_struct *mm = current->mm;
3001	struct vm_area_struct *vma;
3002	unsigned long populate = 0;
3003	unsigned long ret = -EINVAL;
3004	struct file *file;
3005
3006	pr_warn_once("%s (%d) uses deprecated remap_file_pages() syscall. See Documentation/mm/remap_file_pages.rst.\n",
3007		     current->comm, current->pid);
3008
3009	if (prot)
3010		return ret;
3011	start = start & PAGE_MASK;
3012	size = size & PAGE_MASK;
3013
3014	if (start + size <= start)
3015		return ret;
3016
3017	/* Does pgoff wrap? */
3018	if (pgoff + (size >> PAGE_SHIFT) < pgoff)
3019		return ret;
3020
3021	if (mmap_write_lock_killable(mm))
3022		return -EINTR;
3023
3024	vma = vma_lookup(mm, start);
3025
3026	if (!vma || !(vma->vm_flags & VM_SHARED))
3027		goto out;
3028
3029	if (start + size > vma->vm_end) {
3030		VMA_ITERATOR(vmi, mm, vma->vm_end);
3031		struct vm_area_struct *next, *prev = vma;
3032
3033		for_each_vma_range(vmi, next, start + size) {
3034			/* hole between vmas ? */
3035			if (next->vm_start != prev->vm_end)
3036				goto out;
3037
3038			if (next->vm_file != vma->vm_file)
3039				goto out;
3040
3041			if (next->vm_flags != vma->vm_flags)
3042				goto out;
3043
3044			if (start + size <= next->vm_end)
3045				break;
3046
3047			prev = next;
3048		}
3049
3050		if (!next)
3051			goto out;
3052	}
3053
3054	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
3055	prot |= vma->vm_flags & VM_WRITE ? PROT_WRITE : 0;
3056	prot |= vma->vm_flags & VM_EXEC ? PROT_EXEC : 0;
3057
3058	flags &= MAP_NONBLOCK;
3059	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
3060	if (vma->vm_flags & VM_LOCKED)
3061		flags |= MAP_LOCKED;
3062
3063	file = get_file(vma->vm_file);
3064	ret = do_mmap(vma->vm_file, start, size,
3065			prot, flags, 0, pgoff, &populate, NULL);
3066	fput(file);
3067out:
3068	mmap_write_unlock(mm);
3069	if (populate)
3070		mm_populate(ret, populate);
3071	if (!IS_ERR_VALUE(ret))
3072		ret = 0;
3073	return ret;
3074}
3075
3076/*
3077 * do_vma_munmap() - Unmap a full or partial vma.
3078 * @vmi: The vma iterator pointing at the vma
3079 * @vma: The first vma to be munmapped
3080 * @start: the start of the address to unmap
3081 * @end: The end of the address to unmap
3082 * @uf: The userfaultfd list_head
3083 * @unlock: Drop the lock on success
3084 *
3085 * unmaps a VMA mapping when the vma iterator is already in position.
3086 * Does not handle alignment.
3087 *
3088 * Return: 0 on success drops the lock of so directed, error on failure and will
3089 * still hold the lock.
3090 */
3091int do_vma_munmap(struct vma_iterator *vmi, struct vm_area_struct *vma,
3092		unsigned long start, unsigned long end, struct list_head *uf,
3093		bool unlock)
3094{
3095	struct mm_struct *mm = vma->vm_mm;
 
3096
3097	arch_unmap(mm, start, end);
3098	return do_vmi_align_munmap(vmi, vma, mm, start, end, uf, unlock);
 
 
3099}
3100
3101/*
3102 * do_brk_flags() - Increase the brk vma if the flags match.
3103 * @vmi: The vma iterator
3104 * @addr: The start address
3105 * @len: The length of the increase
3106 * @vma: The vma,
3107 * @flags: The VMA Flags
3108 *
3109 * Extend the brk VMA from addr to addr + len.  If the VMA is NULL or the flags
3110 * do not match then create a new anonymous VMA.  Eventually we may be able to
3111 * do some brk-specific accounting here.
3112 */
3113static int do_brk_flags(struct vma_iterator *vmi, struct vm_area_struct *vma,
3114		unsigned long addr, unsigned long len, unsigned long flags)
3115{
3116	struct mm_struct *mm = current->mm;
3117	struct vma_prepare vp;
3118
 
3119	/*
3120	 * Check against address space limits by the changed size
3121	 * Note: This happens *after* clearing old mappings in some code paths.
3122	 */
3123	flags |= VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
3124	if (!may_expand_vm(mm, flags, len >> PAGE_SHIFT))
3125		return -ENOMEM;
3126
3127	if (mm->map_count > sysctl_max_map_count)
3128		return -ENOMEM;
3129
3130	if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
3131		return -ENOMEM;
3132
3133	/*
3134	 * Expand the existing vma if possible; Note that singular lists do not
3135	 * occur after forking, so the expand will only happen on new VMAs.
3136	 */
3137	if (vma && vma->vm_end == addr && !vma_policy(vma) &&
3138	    can_vma_merge_after(vma, flags, NULL, NULL,
3139				addr >> PAGE_SHIFT, NULL_VM_UFFD_CTX, NULL)) {
3140		vma_iter_config(vmi, vma->vm_start, addr + len);
3141		if (vma_iter_prealloc(vmi, vma))
3142			goto unacct_fail;
3143
3144		vma_start_write(vma);
3145
3146		init_vma_prep(&vp, vma);
3147		vma_prepare(&vp);
3148		vma_adjust_trans_huge(vma, vma->vm_start, addr + len, 0);
 
 
 
 
3149		vma->vm_end = addr + len;
3150		vm_flags_set(vma, VM_SOFTDIRTY);
3151		vma_iter_store(vmi, vma);
3152
3153		vma_complete(&vp, vmi, mm);
 
 
 
3154		khugepaged_enter_vma(vma, flags);
3155		goto out;
3156	}
3157
3158	if (vma)
3159		vma_iter_next_range(vmi);
3160	/* create a vma struct for an anonymous mapping */
3161	vma = vm_area_alloc(mm);
3162	if (!vma)
3163		goto unacct_fail;
3164
3165	vma_set_anonymous(vma);
3166	vma_set_range(vma, addr, addr + len, addr >> PAGE_SHIFT);
3167	vm_flags_init(vma, flags);
 
 
3168	vma->vm_page_prot = vm_get_page_prot(flags);
3169	vma_start_write(vma);
3170	if (vma_iter_store_gfp(vmi, vma, GFP_KERNEL))
3171		goto mas_store_fail;
3172
3173	mm->map_count++;
3174	validate_mm(mm);
3175	ksm_add_vma(vma);
3176out:
3177	perf_event_mmap(vma);
3178	mm->total_vm += len >> PAGE_SHIFT;
3179	mm->data_vm += len >> PAGE_SHIFT;
3180	if (flags & VM_LOCKED)
3181		mm->locked_vm += (len >> PAGE_SHIFT);
3182	vm_flags_set(vma, VM_SOFTDIRTY);
 
3183	return 0;
3184
3185mas_store_fail:
3186	vm_area_free(vma);
3187unacct_fail:
3188	vm_unacct_memory(len >> PAGE_SHIFT);
3189	return -ENOMEM;
3190}
3191
3192int vm_brk_flags(unsigned long addr, unsigned long request, unsigned long flags)
3193{
3194	struct mm_struct *mm = current->mm;
3195	struct vm_area_struct *vma = NULL;
3196	unsigned long len;
3197	int ret;
3198	bool populate;
3199	LIST_HEAD(uf);
3200	VMA_ITERATOR(vmi, mm, addr);
3201
3202	len = PAGE_ALIGN(request);
3203	if (len < request)
3204		return -ENOMEM;
3205	if (!len)
3206		return 0;
3207
 
 
 
3208	/* Until we need other flags, refuse anything except VM_EXEC. */
3209	if ((flags & (~VM_EXEC)) != 0)
3210		return -EINVAL;
3211
3212	if (mmap_write_lock_killable(mm))
3213		return -EINTR;
3214
3215	ret = check_brk_limits(addr, len);
3216	if (ret)
3217		goto limits_failed;
3218
3219	ret = do_vmi_munmap(&vmi, mm, addr, len, &uf, 0);
3220	if (ret)
3221		goto munmap_failed;
3222
3223	vma = vma_prev(&vmi);
3224	ret = do_brk_flags(&vmi, vma, addr, len, flags);
3225	populate = ((mm->def_flags & VM_LOCKED) != 0);
3226	mmap_write_unlock(mm);
3227	userfaultfd_unmap_complete(mm, &uf);
3228	if (populate && !ret)
3229		mm_populate(addr, len);
3230	return ret;
3231
3232munmap_failed:
3233limits_failed:
3234	mmap_write_unlock(mm);
3235	return ret;
3236}
3237EXPORT_SYMBOL(vm_brk_flags);
3238
 
 
 
 
 
 
3239/* Release all mmaps. */
3240void exit_mmap(struct mm_struct *mm)
3241{
3242	struct mmu_gather tlb;
3243	struct vm_area_struct *vma;
3244	unsigned long nr_accounted = 0;
3245	MA_STATE(mas, &mm->mm_mt, 0, 0);
3246	int count = 0;
3247
3248	/* mm's last user has gone, and its about to be pulled down */
3249	mmu_notifier_release(mm);
3250
3251	mmap_read_lock(mm);
3252	arch_exit_mmap(mm);
3253
3254	vma = mas_find(&mas, ULONG_MAX);
3255	if (!vma || unlikely(xa_is_zero(vma))) {
3256		/* Can happen if dup_mmap() received an OOM */
3257		mmap_read_unlock(mm);
3258		mmap_write_lock(mm);
3259		goto destroy;
3260	}
3261
3262	lru_add_drain();
3263	flush_cache_mm(mm);
3264	tlb_gather_mmu_fullmm(&tlb, mm);
3265	/* update_hiwater_rss(mm) here? but nobody should be looking */
3266	/* Use ULONG_MAX here to ensure all VMAs in the mm are unmapped */
3267	unmap_vmas(&tlb, &mas, vma, 0, ULONG_MAX, ULONG_MAX, false);
3268	mmap_read_unlock(mm);
3269
3270	/*
3271	 * Set MMF_OOM_SKIP to hide this task from the oom killer/reaper
3272	 * because the memory has been already freed.
3273	 */
3274	set_bit(MMF_OOM_SKIP, &mm->flags);
3275	mmap_write_lock(mm);
3276	mt_clear_in_rcu(&mm->mm_mt);
3277	mas_set(&mas, vma->vm_end);
3278	free_pgtables(&tlb, &mas, vma, FIRST_USER_ADDRESS,
3279		      USER_PGTABLES_CEILING, true);
3280	tlb_finish_mmu(&tlb);
3281
3282	/*
3283	 * Walk the list again, actually closing and freeing it, with preemption
3284	 * enabled, without holding any MM locks besides the unreachable
3285	 * mmap_write_lock.
3286	 */
3287	mas_set(&mas, vma->vm_end);
3288	do {
3289		if (vma->vm_flags & VM_ACCOUNT)
3290			nr_accounted += vma_pages(vma);
3291		remove_vma(vma, true);
3292		count++;
3293		cond_resched();
3294		vma = mas_find(&mas, ULONG_MAX);
3295	} while (vma && likely(!xa_is_zero(vma)));
3296
3297	BUG_ON(count != mm->map_count);
3298
3299	trace_exit_mmap(mm);
3300destroy:
3301	__mt_destroy(&mm->mm_mt);
3302	mmap_write_unlock(mm);
3303	vm_unacct_memory(nr_accounted);
3304}
3305
3306/* Insert vm structure into process list sorted by address
3307 * and into the inode's i_mmap tree.  If vm_file is non-NULL
3308 * then i_mmap_rwsem is taken here.
3309 */
3310int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
3311{
3312	unsigned long charged = vma_pages(vma);
3313
3314
3315	if (find_vma_intersection(mm, vma->vm_start, vma->vm_end))
3316		return -ENOMEM;
3317
3318	if ((vma->vm_flags & VM_ACCOUNT) &&
3319	     security_vm_enough_memory_mm(mm, charged))
3320		return -ENOMEM;
3321
3322	/*
3323	 * The vm_pgoff of a purely anonymous vma should be irrelevant
3324	 * until its first write fault, when page's anon_vma and index
3325	 * are set.  But now set the vm_pgoff it will almost certainly
3326	 * end up with (unless mremap moves it elsewhere before that
3327	 * first wfault), so /proc/pid/maps tells a consistent story.
3328	 *
3329	 * By setting it to reflect the virtual start address of the
3330	 * vma, merges and splits can happen in a seamless way, just
3331	 * using the existing file pgoff checks and manipulations.
3332	 * Similarly in do_mmap and in do_brk_flags.
3333	 */
3334	if (vma_is_anonymous(vma)) {
3335		BUG_ON(vma->anon_vma);
3336		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
3337	}
3338
3339	if (vma_link(mm, vma)) {
3340		if (vma->vm_flags & VM_ACCOUNT)
3341			vm_unacct_memory(charged);
3342		return -ENOMEM;
3343	}
3344
3345	return 0;
3346}
3347
3348/*
3349 * Copy the vma structure to a new location in the same mm,
3350 * prior to moving page table entries, to effect an mremap move.
3351 */
3352struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
3353	unsigned long addr, unsigned long len, pgoff_t pgoff,
3354	bool *need_rmap_locks)
3355{
3356	struct vm_area_struct *vma = *vmap;
3357	unsigned long vma_start = vma->vm_start;
3358	struct mm_struct *mm = vma->vm_mm;
3359	struct vm_area_struct *new_vma, *prev;
3360	bool faulted_in_anon_vma = true;
3361	VMA_ITERATOR(vmi, mm, addr);
3362
 
3363	/*
3364	 * If anonymous vma has not yet been faulted, update new pgoff
3365	 * to match new location, to increase its chance of merging.
3366	 */
3367	if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma)) {
3368		pgoff = addr >> PAGE_SHIFT;
3369		faulted_in_anon_vma = false;
3370	}
3371
3372	new_vma = find_vma_prev(mm, addr, &prev);
3373	if (new_vma && new_vma->vm_start < addr + len)
3374		return NULL;	/* should never get here */
3375
3376	new_vma = vma_merge_new_vma(&vmi, prev, vma, addr, addr + len, pgoff);
 
 
3377	if (new_vma) {
3378		/*
3379		 * Source vma may have been merged into new_vma
3380		 */
3381		if (unlikely(vma_start >= new_vma->vm_start &&
3382			     vma_start < new_vma->vm_end)) {
3383			/*
3384			 * The only way we can get a vma_merge with
3385			 * self during an mremap is if the vma hasn't
3386			 * been faulted in yet and we were allowed to
3387			 * reset the dst vma->vm_pgoff to the
3388			 * destination address of the mremap to allow
3389			 * the merge to happen. mremap must change the
3390			 * vm_pgoff linearity between src and dst vmas
3391			 * (in turn preventing a vma_merge) to be
3392			 * safe. It is only safe to keep the vm_pgoff
3393			 * linear if there are no pages mapped yet.
3394			 */
3395			VM_BUG_ON_VMA(faulted_in_anon_vma, new_vma);
3396			*vmap = vma = new_vma;
3397		}
3398		*need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
3399	} else {
3400		new_vma = vm_area_dup(vma);
3401		if (!new_vma)
3402			goto out;
3403		vma_set_range(new_vma, addr, addr + len, pgoff);
 
 
3404		if (vma_dup_policy(vma, new_vma))
3405			goto out_free_vma;
3406		if (anon_vma_clone(new_vma, vma))
3407			goto out_free_mempol;
3408		if (new_vma->vm_file)
3409			get_file(new_vma->vm_file);
3410		if (new_vma->vm_ops && new_vma->vm_ops->open)
3411			new_vma->vm_ops->open(new_vma);
3412		if (vma_link(mm, new_vma))
3413			goto out_vma_link;
3414		*need_rmap_locks = false;
3415	}
 
3416	return new_vma;
3417
3418out_vma_link:
3419	if (new_vma->vm_ops && new_vma->vm_ops->close)
3420		new_vma->vm_ops->close(new_vma);
3421
3422	if (new_vma->vm_file)
3423		fput(new_vma->vm_file);
3424
3425	unlink_anon_vmas(new_vma);
3426out_free_mempol:
3427	mpol_put(vma_policy(new_vma));
3428out_free_vma:
3429	vm_area_free(new_vma);
3430out:
 
3431	return NULL;
3432}
3433
3434/*
3435 * Return true if the calling process may expand its vm space by the passed
3436 * number of pages
3437 */
3438bool may_expand_vm(struct mm_struct *mm, vm_flags_t flags, unsigned long npages)
3439{
3440	if (mm->total_vm + npages > rlimit(RLIMIT_AS) >> PAGE_SHIFT)
3441		return false;
3442
3443	if (is_data_mapping(flags) &&
3444	    mm->data_vm + npages > rlimit(RLIMIT_DATA) >> PAGE_SHIFT) {
3445		/* Workaround for Valgrind */
3446		if (rlimit(RLIMIT_DATA) == 0 &&
3447		    mm->data_vm + npages <= rlimit_max(RLIMIT_DATA) >> PAGE_SHIFT)
3448			return true;
3449
3450		pr_warn_once("%s (%d): VmData %lu exceed data ulimit %lu. Update limits%s.\n",
3451			     current->comm, current->pid,
3452			     (mm->data_vm + npages) << PAGE_SHIFT,
3453			     rlimit(RLIMIT_DATA),
3454			     ignore_rlimit_data ? "" : " or use boot option ignore_rlimit_data");
3455
3456		if (!ignore_rlimit_data)
3457			return false;
3458	}
3459
3460	return true;
3461}
3462
3463void vm_stat_account(struct mm_struct *mm, vm_flags_t flags, long npages)
3464{
3465	WRITE_ONCE(mm->total_vm, READ_ONCE(mm->total_vm)+npages);
3466
3467	if (is_exec_mapping(flags))
3468		mm->exec_vm += npages;
3469	else if (is_stack_mapping(flags))
3470		mm->stack_vm += npages;
3471	else if (is_data_mapping(flags))
3472		mm->data_vm += npages;
3473}
3474
3475static vm_fault_t special_mapping_fault(struct vm_fault *vmf);
3476
3477/*
3478 * Having a close hook prevents vma merging regardless of flags.
3479 */
3480static void special_mapping_close(struct vm_area_struct *vma)
3481{
3482}
3483
3484static const char *special_mapping_name(struct vm_area_struct *vma)
3485{
3486	return ((struct vm_special_mapping *)vma->vm_private_data)->name;
3487}
3488
3489static int special_mapping_mremap(struct vm_area_struct *new_vma)
3490{
3491	struct vm_special_mapping *sm = new_vma->vm_private_data;
3492
3493	if (WARN_ON_ONCE(current->mm != new_vma->vm_mm))
3494		return -EFAULT;
3495
3496	if (sm->mremap)
3497		return sm->mremap(sm, new_vma);
3498
3499	return 0;
3500}
3501
3502static int special_mapping_split(struct vm_area_struct *vma, unsigned long addr)
3503{
3504	/*
3505	 * Forbid splitting special mappings - kernel has expectations over
3506	 * the number of pages in mapping. Together with VM_DONTEXPAND
3507	 * the size of vma should stay the same over the special mapping's
3508	 * lifetime.
3509	 */
3510	return -EINVAL;
3511}
3512
3513static const struct vm_operations_struct special_mapping_vmops = {
3514	.close = special_mapping_close,
3515	.fault = special_mapping_fault,
3516	.mremap = special_mapping_mremap,
3517	.name = special_mapping_name,
3518	/* vDSO code relies that VVAR can't be accessed remotely */
3519	.access = NULL,
3520	.may_split = special_mapping_split,
3521};
3522
3523static const struct vm_operations_struct legacy_special_mapping_vmops = {
3524	.close = special_mapping_close,
3525	.fault = special_mapping_fault,
3526};
3527
3528static vm_fault_t special_mapping_fault(struct vm_fault *vmf)
3529{
3530	struct vm_area_struct *vma = vmf->vma;
3531	pgoff_t pgoff;
3532	struct page **pages;
3533
3534	if (vma->vm_ops == &legacy_special_mapping_vmops) {
3535		pages = vma->vm_private_data;
3536	} else {
3537		struct vm_special_mapping *sm = vma->vm_private_data;
3538
3539		if (sm->fault)
3540			return sm->fault(sm, vmf->vma, vmf);
3541
3542		pages = sm->pages;
3543	}
3544
3545	for (pgoff = vmf->pgoff; pgoff && *pages; ++pages)
3546		pgoff--;
3547
3548	if (*pages) {
3549		struct page *page = *pages;
3550		get_page(page);
3551		vmf->page = page;
3552		return 0;
3553	}
3554
3555	return VM_FAULT_SIGBUS;
3556}
3557
3558static struct vm_area_struct *__install_special_mapping(
3559	struct mm_struct *mm,
3560	unsigned long addr, unsigned long len,
3561	unsigned long vm_flags, void *priv,
3562	const struct vm_operations_struct *ops)
3563{
3564	int ret;
3565	struct vm_area_struct *vma;
3566
 
3567	vma = vm_area_alloc(mm);
3568	if (unlikely(vma == NULL))
3569		return ERR_PTR(-ENOMEM);
3570
3571	vma_set_range(vma, addr, addr + len, 0);
3572	vm_flags_init(vma, (vm_flags | mm->def_flags |
3573		      VM_DONTEXPAND | VM_SOFTDIRTY) & ~VM_LOCKED_MASK);
 
 
3574	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
3575
3576	vma->vm_ops = ops;
3577	vma->vm_private_data = priv;
3578
3579	ret = insert_vm_struct(mm, vma);
3580	if (ret)
3581		goto out;
3582
3583	vm_stat_account(mm, vma->vm_flags, len >> PAGE_SHIFT);
3584
3585	perf_event_mmap(vma);
3586
 
3587	return vma;
3588
3589out:
3590	vm_area_free(vma);
 
3591	return ERR_PTR(ret);
3592}
3593
3594bool vma_is_special_mapping(const struct vm_area_struct *vma,
3595	const struct vm_special_mapping *sm)
3596{
3597	return vma->vm_private_data == sm &&
3598		(vma->vm_ops == &special_mapping_vmops ||
3599		 vma->vm_ops == &legacy_special_mapping_vmops);
3600}
3601
3602/*
3603 * Called with mm->mmap_lock held for writing.
3604 * Insert a new vma covering the given region, with the given flags.
3605 * Its pages are supplied by the given array of struct page *.
3606 * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
3607 * The region past the last page supplied will always produce SIGBUS.
3608 * The array pointer and the pages it points to are assumed to stay alive
3609 * for as long as this mapping might exist.
3610 */
3611struct vm_area_struct *_install_special_mapping(
3612	struct mm_struct *mm,
3613	unsigned long addr, unsigned long len,
3614	unsigned long vm_flags, const struct vm_special_mapping *spec)
3615{
3616	return __install_special_mapping(mm, addr, len, vm_flags, (void *)spec,
3617					&special_mapping_vmops);
3618}
3619
3620int install_special_mapping(struct mm_struct *mm,
3621			    unsigned long addr, unsigned long len,
3622			    unsigned long vm_flags, struct page **pages)
3623{
3624	struct vm_area_struct *vma = __install_special_mapping(
3625		mm, addr, len, vm_flags, (void *)pages,
3626		&legacy_special_mapping_vmops);
3627
3628	return PTR_ERR_OR_ZERO(vma);
3629}
3630
3631static DEFINE_MUTEX(mm_all_locks_mutex);
3632
3633static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
3634{
3635	if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3636		/*
3637		 * The LSB of head.next can't change from under us
3638		 * because we hold the mm_all_locks_mutex.
3639		 */
3640		down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_lock);
3641		/*
3642		 * We can safely modify head.next after taking the
3643		 * anon_vma->root->rwsem. If some other vma in this mm shares
3644		 * the same anon_vma we won't take it again.
3645		 *
3646		 * No need of atomic instructions here, head.next
3647		 * can't change from under us thanks to the
3648		 * anon_vma->root->rwsem.
3649		 */
3650		if (__test_and_set_bit(0, (unsigned long *)
3651				       &anon_vma->root->rb_root.rb_root.rb_node))
3652			BUG();
3653	}
3654}
3655
3656static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
3657{
3658	if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3659		/*
3660		 * AS_MM_ALL_LOCKS can't change from under us because
3661		 * we hold the mm_all_locks_mutex.
3662		 *
3663		 * Operations on ->flags have to be atomic because
3664		 * even if AS_MM_ALL_LOCKS is stable thanks to the
3665		 * mm_all_locks_mutex, there may be other cpus
3666		 * changing other bitflags in parallel to us.
3667		 */
3668		if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
3669			BUG();
3670		down_write_nest_lock(&mapping->i_mmap_rwsem, &mm->mmap_lock);
3671	}
3672}
3673
3674/*
3675 * This operation locks against the VM for all pte/vma/mm related
3676 * operations that could ever happen on a certain mm. This includes
3677 * vmtruncate, try_to_unmap, and all page faults.
3678 *
3679 * The caller must take the mmap_lock in write mode before calling
3680 * mm_take_all_locks(). The caller isn't allowed to release the
3681 * mmap_lock until mm_drop_all_locks() returns.
3682 *
3683 * mmap_lock in write mode is required in order to block all operations
3684 * that could modify pagetables and free pages without need of
3685 * altering the vma layout. It's also needed in write mode to avoid new
3686 * anon_vmas to be associated with existing vmas.
3687 *
3688 * A single task can't take more than one mm_take_all_locks() in a row
3689 * or it would deadlock.
3690 *
3691 * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
3692 * mapping->flags avoid to take the same lock twice, if more than one
3693 * vma in this mm is backed by the same anon_vma or address_space.
3694 *
3695 * We take locks in following order, accordingly to comment at beginning
3696 * of mm/rmap.c:
3697 *   - all hugetlbfs_i_mmap_rwsem_key locks (aka mapping->i_mmap_rwsem for
3698 *     hugetlb mapping);
3699 *   - all vmas marked locked
3700 *   - all i_mmap_rwsem locks;
3701 *   - all anon_vma->rwseml
3702 *
3703 * We can take all locks within these types randomly because the VM code
3704 * doesn't nest them and we protected from parallel mm_take_all_locks() by
3705 * mm_all_locks_mutex.
3706 *
3707 * mm_take_all_locks() and mm_drop_all_locks are expensive operations
3708 * that may have to take thousand of locks.
3709 *
3710 * mm_take_all_locks() can fail if it's interrupted by signals.
3711 */
3712int mm_take_all_locks(struct mm_struct *mm)
3713{
3714	struct vm_area_struct *vma;
3715	struct anon_vma_chain *avc;
3716	MA_STATE(mas, &mm->mm_mt, 0, 0);
3717
3718	mmap_assert_write_locked(mm);
3719
3720	mutex_lock(&mm_all_locks_mutex);
3721
3722	/*
3723	 * vma_start_write() does not have a complement in mm_drop_all_locks()
3724	 * because vma_start_write() is always asymmetrical; it marks a VMA as
3725	 * being written to until mmap_write_unlock() or mmap_write_downgrade()
3726	 * is reached.
3727	 */
3728	mas_for_each(&mas, vma, ULONG_MAX) {
3729		if (signal_pending(current))
3730			goto out_unlock;
3731		vma_start_write(vma);
3732	}
3733
3734	mas_set(&mas, 0);
3735	mas_for_each(&mas, vma, ULONG_MAX) {
3736		if (signal_pending(current))
3737			goto out_unlock;
3738		if (vma->vm_file && vma->vm_file->f_mapping &&
3739				is_vm_hugetlb_page(vma))
3740			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3741	}
3742
3743	mas_set(&mas, 0);
3744	mas_for_each(&mas, vma, ULONG_MAX) {
3745		if (signal_pending(current))
3746			goto out_unlock;
3747		if (vma->vm_file && vma->vm_file->f_mapping &&
3748				!is_vm_hugetlb_page(vma))
3749			vm_lock_mapping(mm, vma->vm_file->f_mapping);
3750	}
3751
3752	mas_set(&mas, 0);
3753	mas_for_each(&mas, vma, ULONG_MAX) {
3754		if (signal_pending(current))
3755			goto out_unlock;
3756		if (vma->anon_vma)
3757			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3758				vm_lock_anon_vma(mm, avc->anon_vma);
3759	}
3760
3761	return 0;
3762
3763out_unlock:
3764	mm_drop_all_locks(mm);
3765	return -EINTR;
3766}
3767
3768static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3769{
3770	if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_root.rb_node)) {
3771		/*
3772		 * The LSB of head.next can't change to 0 from under
3773		 * us because we hold the mm_all_locks_mutex.
3774		 *
3775		 * We must however clear the bitflag before unlocking
3776		 * the vma so the users using the anon_vma->rb_root will
3777		 * never see our bitflag.
3778		 *
3779		 * No need of atomic instructions here, head.next
3780		 * can't change from under us until we release the
3781		 * anon_vma->root->rwsem.
3782		 */
3783		if (!__test_and_clear_bit(0, (unsigned long *)
3784					  &anon_vma->root->rb_root.rb_root.rb_node))
3785			BUG();
3786		anon_vma_unlock_write(anon_vma);
3787	}
3788}
3789
3790static void vm_unlock_mapping(struct address_space *mapping)
3791{
3792	if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3793		/*
3794		 * AS_MM_ALL_LOCKS can't change to 0 from under us
3795		 * because we hold the mm_all_locks_mutex.
3796		 */
3797		i_mmap_unlock_write(mapping);
3798		if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3799					&mapping->flags))
3800			BUG();
3801	}
3802}
3803
3804/*
3805 * The mmap_lock cannot be released by the caller until
3806 * mm_drop_all_locks() returns.
3807 */
3808void mm_drop_all_locks(struct mm_struct *mm)
3809{
3810	struct vm_area_struct *vma;
3811	struct anon_vma_chain *avc;
3812	MA_STATE(mas, &mm->mm_mt, 0, 0);
3813
3814	mmap_assert_write_locked(mm);
3815	BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3816
3817	mas_for_each(&mas, vma, ULONG_MAX) {
3818		if (vma->anon_vma)
3819			list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3820				vm_unlock_anon_vma(avc->anon_vma);
3821		if (vma->vm_file && vma->vm_file->f_mapping)
3822			vm_unlock_mapping(vma->vm_file->f_mapping);
3823	}
3824
3825	mutex_unlock(&mm_all_locks_mutex);
3826}
3827
3828/*
3829 * initialise the percpu counter for VM
3830 */
3831void __init mmap_init(void)
3832{
3833	int ret;
3834
3835	ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
3836	VM_BUG_ON(ret);
3837}
3838
3839/*
3840 * Initialise sysctl_user_reserve_kbytes.
3841 *
3842 * This is intended to prevent a user from starting a single memory hogging
3843 * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
3844 * mode.
3845 *
3846 * The default value is min(3% of free memory, 128MB)
3847 * 128MB is enough to recover with sshd/login, bash, and top/kill.
3848 */
3849static int init_user_reserve(void)
3850{
3851	unsigned long free_kbytes;
3852
3853	free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3854
3855	sysctl_user_reserve_kbytes = min(free_kbytes / 32, SZ_128K);
3856	return 0;
3857}
3858subsys_initcall(init_user_reserve);
3859
3860/*
3861 * Initialise sysctl_admin_reserve_kbytes.
3862 *
3863 * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
3864 * to log in and kill a memory hogging process.
3865 *
3866 * Systems with more than 256MB will reserve 8MB, enough to recover
3867 * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
3868 * only reserve 3% of free pages by default.
3869 */
3870static int init_admin_reserve(void)
3871{
3872	unsigned long free_kbytes;
3873
3874	free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3875
3876	sysctl_admin_reserve_kbytes = min(free_kbytes / 32, SZ_8K);
3877	return 0;
3878}
3879subsys_initcall(init_admin_reserve);
3880
3881/*
3882 * Reinititalise user and admin reserves if memory is added or removed.
3883 *
3884 * The default user reserve max is 128MB, and the default max for the
3885 * admin reserve is 8MB. These are usually, but not always, enough to
3886 * enable recovery from a memory hogging process using login/sshd, a shell,
3887 * and tools like top. It may make sense to increase or even disable the
3888 * reserve depending on the existence of swap or variations in the recovery
3889 * tools. So, the admin may have changed them.
3890 *
3891 * If memory is added and the reserves have been eliminated or increased above
3892 * the default max, then we'll trust the admin.
3893 *
3894 * If memory is removed and there isn't enough free memory, then we
3895 * need to reset the reserves.
3896 *
3897 * Otherwise keep the reserve set by the admin.
3898 */
3899static int reserve_mem_notifier(struct notifier_block *nb,
3900			     unsigned long action, void *data)
3901{
3902	unsigned long tmp, free_kbytes;
3903
3904	switch (action) {
3905	case MEM_ONLINE:
3906		/* Default max is 128MB. Leave alone if modified by operator. */
3907		tmp = sysctl_user_reserve_kbytes;
3908		if (tmp > 0 && tmp < SZ_128K)
3909			init_user_reserve();
3910
3911		/* Default max is 8MB.  Leave alone if modified by operator. */
3912		tmp = sysctl_admin_reserve_kbytes;
3913		if (tmp > 0 && tmp < SZ_8K)
3914			init_admin_reserve();
3915
3916		break;
3917	case MEM_OFFLINE:
3918		free_kbytes = K(global_zone_page_state(NR_FREE_PAGES));
3919
3920		if (sysctl_user_reserve_kbytes > free_kbytes) {
3921			init_user_reserve();
3922			pr_info("vm.user_reserve_kbytes reset to %lu\n",
3923				sysctl_user_reserve_kbytes);
3924		}
3925
3926		if (sysctl_admin_reserve_kbytes > free_kbytes) {
3927			init_admin_reserve();
3928			pr_info("vm.admin_reserve_kbytes reset to %lu\n",
3929				sysctl_admin_reserve_kbytes);
3930		}
3931		break;
3932	default:
3933		break;
3934	}
3935	return NOTIFY_OK;
3936}
3937
3938static int __meminit init_reserve_notifier(void)
3939{
3940	if (hotplug_memory_notifier(reserve_mem_notifier, DEFAULT_CALLBACK_PRI))
3941		pr_err("Failed registering memory add/remove notifier for admin reserve\n");
3942
3943	return 0;
3944}
3945subsys_initcall(init_reserve_notifier);