Linux Audio

Check our new training course

Loading...
v3.15
 
  1/*
  2 *	mm/mremap.c
  3 *
  4 *	(C) Copyright 1996 Linus Torvalds
  5 *
  6 *	Address space accounting code	<alan@lxorguk.ukuu.org.uk>
  7 *	(C) Copyright 2002 Red Hat Inc, All Rights Reserved
  8 */
  9
 10#include <linux/mm.h>
 11#include <linux/hugetlb.h>
 12#include <linux/shm.h>
 13#include <linux/ksm.h>
 14#include <linux/mman.h>
 15#include <linux/swap.h>
 16#include <linux/capability.h>
 17#include <linux/fs.h>
 18#include <linux/swapops.h>
 19#include <linux/highmem.h>
 20#include <linux/security.h>
 21#include <linux/syscalls.h>
 22#include <linux/mmu_notifier.h>
 23#include <linux/sched/sysctl.h>
 
 24
 25#include <asm/uaccess.h>
 26#include <asm/cacheflush.h>
 27#include <asm/tlbflush.h>
 
 28
 29#include "internal.h"
 30
 31static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
 32{
 33	pgd_t *pgd;
 
 34	pud_t *pud;
 35	pmd_t *pmd;
 36
 37	pgd = pgd_offset(mm, addr);
 38	if (pgd_none_or_clear_bad(pgd))
 39		return NULL;
 40
 41	pud = pud_offset(pgd, addr);
 
 
 
 
 42	if (pud_none_or_clear_bad(pud))
 43		return NULL;
 44
 
 
 
 
 
 
 
 
 
 
 
 
 45	pmd = pmd_offset(pud, addr);
 46	if (pmd_none(*pmd))
 47		return NULL;
 48
 49	return pmd;
 50}
 51
 52static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
 53			    unsigned long addr)
 54{
 55	pgd_t *pgd;
 
 
 
 
 
 
 
 
 
 
 
 
 
 56	pud_t *pud;
 57	pmd_t *pmd;
 58
 59	pgd = pgd_offset(mm, addr);
 60	pud = pud_alloc(mm, pgd, addr);
 61	if (!pud)
 62		return NULL;
 63
 64	pmd = pmd_alloc(mm, pud, addr);
 65	if (!pmd)
 66		return NULL;
 67
 68	VM_BUG_ON(pmd_trans_huge(*pmd));
 69
 70	return pmd;
 71}
 72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 73static pte_t move_soft_dirty_pte(pte_t pte)
 74{
 75	/*
 76	 * Set soft dirty bit so we can notice
 77	 * in userspace the ptes were moved.
 78	 */
 79#ifdef CONFIG_MEM_SOFT_DIRTY
 80	if (pte_present(pte))
 81		pte = pte_mksoft_dirty(pte);
 82	else if (is_swap_pte(pte))
 83		pte = pte_swp_mksoft_dirty(pte);
 84	else if (pte_file(pte))
 85		pte = pte_file_mksoft_dirty(pte);
 86#endif
 87	return pte;
 88}
 89
 90static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 91		unsigned long old_addr, unsigned long old_end,
 92		struct vm_area_struct *new_vma, pmd_t *new_pmd,
 93		unsigned long new_addr, bool need_rmap_locks)
 94{
 95	struct address_space *mapping = NULL;
 96	struct anon_vma *anon_vma = NULL;
 97	struct mm_struct *mm = vma->vm_mm;
 98	pte_t *old_pte, *new_pte, pte;
 99	spinlock_t *old_ptl, *new_ptl;
 
 
100
101	/*
102	 * When need_rmap_locks is true, we take the i_mmap_mutex and anon_vma
103	 * locks to ensure that rmap will always observe either the old or the
104	 * new ptes. This is the easiest way to avoid races with
105	 * truncate_pagecache(), page migration, etc...
106	 *
107	 * When need_rmap_locks is false, we use other ways to avoid
108	 * such races:
109	 *
110	 * - During exec() shift_arg_pages(), we use a specially tagged vma
111	 *   which rmap call sites look for using is_vma_temporary_stack().
112	 *
113	 * - During mremap(), new_vma is often known to be placed after vma
114	 *   in rmap traversal order. This ensures rmap will always observe
115	 *   either the old pte, or the new pte, or both (the page table locks
116	 *   serialize access to individual ptes, but only rmap traversal
117	 *   order guarantees that we won't miss both the old and new ptes).
118	 */
119	if (need_rmap_locks) {
120		if (vma->vm_file) {
121			mapping = vma->vm_file->f_mapping;
122			mutex_lock(&mapping->i_mmap_mutex);
123		}
124		if (vma->anon_vma) {
125			anon_vma = vma->anon_vma;
126			anon_vma_lock_write(anon_vma);
127		}
128	}
129
130	/*
131	 * We don't have to worry about the ordering of src and dst
132	 * pte locks because exclusive mmap_sem prevents deadlock.
133	 */
134	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
135	new_pte = pte_offset_map(new_pmd, new_addr);
136	new_ptl = pte_lockptr(mm, new_pmd);
137	if (new_ptl != old_ptl)
138		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 
139	arch_enter_lazy_mmu_mode();
140
141	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
142				   new_pte++, new_addr += PAGE_SIZE) {
143		if (pte_none(*old_pte))
144			continue;
 
145		pte = ptep_get_and_clear(mm, old_addr, old_pte);
 
 
 
 
 
 
 
 
 
 
 
 
 
146		pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
147		pte = move_soft_dirty_pte(pte);
148		set_pte_at(mm, new_addr, new_pte, pte);
149	}
150
151	arch_leave_lazy_mmu_mode();
 
 
152	if (new_ptl != old_ptl)
153		spin_unlock(new_ptl);
154	pte_unmap(new_pte - 1);
155	pte_unmap_unlock(old_pte - 1, old_ptl);
156	if (anon_vma)
157		anon_vma_unlock_write(anon_vma);
158	if (mapping)
159		mutex_unlock(&mapping->i_mmap_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
160}
161
162#define LATENCY_LIMIT	(64 * PAGE_SIZE)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
163
164unsigned long move_page_tables(struct vm_area_struct *vma,
165		unsigned long old_addr, struct vm_area_struct *new_vma,
166		unsigned long new_addr, unsigned long len,
167		bool need_rmap_locks)
168{
169	unsigned long extent, next, old_end;
 
170	pmd_t *old_pmd, *new_pmd;
171	bool need_flush = false;
172	unsigned long mmun_start;	/* For mmu_notifiers */
173	unsigned long mmun_end;		/* For mmu_notifiers */
174
175	old_end = old_addr + len;
176	flush_cache_range(vma, old_addr, old_end);
177
178	mmun_start = old_addr;
179	mmun_end   = old_end;
180	mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
181
182	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
183		cond_resched();
184		next = (old_addr + PMD_SIZE) & PMD_MASK;
185		/* even if next overflowed, extent below will be ok */
186		extent = next - old_addr;
187		if (extent > old_end - old_addr)
188			extent = old_end - old_addr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
189		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
190		if (!old_pmd)
191			continue;
192		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
193		if (!new_pmd)
194			break;
195		if (pmd_trans_huge(*old_pmd)) {
196			int err = 0;
197			if (extent == HPAGE_PMD_SIZE) {
198				VM_BUG_ON(vma->vm_file || !vma->anon_vma);
199				/* See comment in move_ptes() */
200				if (need_rmap_locks)
201					anon_vma_lock_write(vma->anon_vma);
202				err = move_huge_pmd(vma, new_vma, old_addr,
203						    new_addr, old_end,
204						    old_pmd, new_pmd);
205				if (need_rmap_locks)
206					anon_vma_unlock_write(vma->anon_vma);
207			}
208			if (err > 0) {
209				need_flush = true;
 
 
210				continue;
211			} else if (!err) {
212				split_huge_page_pmd(vma, old_addr, old_pmd);
213			}
214			VM_BUG_ON(pmd_trans_huge(*old_pmd));
215		}
216		if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
217						      new_pmd, new_addr))
218			break;
219		next = (new_addr + PMD_SIZE) & PMD_MASK;
220		if (extent > next - new_addr)
221			extent = next - new_addr;
222		if (extent > LATENCY_LIMIT)
223			extent = LATENCY_LIMIT;
224		move_ptes(vma, old_pmd, old_addr, old_addr + extent,
225			  new_vma, new_pmd, new_addr, need_rmap_locks);
226		need_flush = true;
227	}
228	if (likely(need_flush))
229		flush_tlb_range(vma, old_end-len, old_addr);
230
231	mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
232
233	return len + old_addr - old_end;	/* how much done */
234}
235
236static unsigned long move_vma(struct vm_area_struct *vma,
237		unsigned long old_addr, unsigned long old_len,
238		unsigned long new_len, unsigned long new_addr, bool *locked)
 
 
239{
240	struct mm_struct *mm = vma->vm_mm;
241	struct vm_area_struct *new_vma;
242	unsigned long vm_flags = vma->vm_flags;
243	unsigned long new_pgoff;
244	unsigned long moved_len;
245	unsigned long excess = 0;
246	unsigned long hiwater_vm;
247	int split = 0;
248	int err;
249	bool need_rmap_locks;
250
251	/*
252	 * We'd prefer to avoid failure later on in do_munmap:
253	 * which may split one vma into three before unmapping.
254	 */
255	if (mm->map_count >= sysctl_max_map_count - 3)
256		return -ENOMEM;
257
 
 
 
 
 
 
 
 
 
258	/*
259	 * Advise KSM to break any KSM pages in the area to be moved:
260	 * it would be confusing if they were to turn up at the new
261	 * location, where they happen to coincide with different KSM
262	 * pages recently unmapped.  But leave vma->vm_flags as it was,
263	 * so KSM can come around to merge on vma and new_vma afterwards.
264	 */
265	err = ksm_madvise(vma, old_addr, old_addr + old_len,
266						MADV_UNMERGEABLE, &vm_flags);
267	if (err)
268		return err;
269
 
 
 
 
 
270	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
271	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
272			   &need_rmap_locks);
273	if (!new_vma)
 
 
274		return -ENOMEM;
 
275
276	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
277				     need_rmap_locks);
278	if (moved_len < old_len) {
 
 
 
 
 
 
279		/*
280		 * On error, move entries back from new area to old,
281		 * which will succeed since page tables still there,
282		 * and then proceed to unmap new area instead of old.
283		 */
284		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
285				 true);
286		vma = new_vma;
287		old_len = new_len;
288		old_addr = new_addr;
289		new_addr = -ENOMEM;
 
 
290	}
291
292	/* Conceal VM_ACCOUNT so old reservation is not undone */
293	if (vm_flags & VM_ACCOUNT) {
294		vma->vm_flags &= ~VM_ACCOUNT;
295		excess = vma->vm_end - vma->vm_start - old_len;
296		if (old_addr > vma->vm_start &&
297		    old_addr + old_len < vma->vm_end)
298			split = 1;
299	}
300
301	/*
302	 * If we failed to move page tables we still do total_vm increment
303	 * since do_munmap() will decrement it by old_len == new_len.
304	 *
305	 * Since total_vm is about to be raised artificially high for a
306	 * moment, we need to restore high watermark afterwards: if stats
307	 * are taken meanwhile, total_vm and hiwater_vm appear too high.
308	 * If this were a serious issue, we'd add a flag to do_munmap().
309	 */
310	hiwater_vm = mm->hiwater_vm;
311	vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
312
313	if (do_munmap(mm, old_addr, old_len) < 0) {
314		/* OOM: unable to split vma, just get accounts right */
315		vm_unacct_memory(excess >> PAGE_SHIFT);
 
316		excess = 0;
317	}
 
 
 
 
 
 
318	mm->hiwater_vm = hiwater_vm;
319
320	/* Restore VM_ACCOUNT if one or two pieces of vma left */
321	if (excess) {
322		vma->vm_flags |= VM_ACCOUNT;
323		if (split)
324			vma->vm_next->vm_flags |= VM_ACCOUNT;
325	}
326
327	if (vm_flags & VM_LOCKED) {
328		mm->locked_vm += new_len >> PAGE_SHIFT;
329		*locked = true;
330	}
331
332	return new_addr;
333}
334
335static struct vm_area_struct *vma_to_resize(unsigned long addr,
336	unsigned long old_len, unsigned long new_len, unsigned long *p)
 
337{
338	struct mm_struct *mm = current->mm;
339	struct vm_area_struct *vma = find_vma(mm, addr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
340
341	if (!vma || vma->vm_start > addr)
342		goto Efault;
 
343
344	if (is_vm_hugetlb_page(vma))
345		goto Einval;
346
347	/* We can't remap across vm area boundaries */
348	if (old_len > vma->vm_end - addr)
349		goto Efault;
 
 
 
350
351	/* Need to be careful about a growing mapping */
352	if (new_len > old_len) {
353		unsigned long pgoff;
 
 
354
355		if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
356			goto Efault;
357		pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
358		pgoff += vma->vm_pgoff;
359		if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
360			goto Einval;
361	}
362
363	if (vma->vm_flags & VM_LOCKED) {
364		unsigned long locked, lock_limit;
365		locked = mm->locked_vm << PAGE_SHIFT;
366		lock_limit = rlimit(RLIMIT_MEMLOCK);
367		locked += new_len - old_len;
368		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
369			goto Eagain;
370	}
371
372	if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
373		goto Enomem;
 
374
375	if (vma->vm_flags & VM_ACCOUNT) {
376		unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
377		if (security_vm_enough_memory_mm(mm, charged))
378			goto Efault;
379		*p = charged;
380	}
381
382	return vma;
383
384Efault:	/* very odd choice for most of the cases, but... */
385	return ERR_PTR(-EFAULT);
386Einval:
387	return ERR_PTR(-EINVAL);
388Enomem:
389	return ERR_PTR(-ENOMEM);
390Eagain:
391	return ERR_PTR(-EAGAIN);
392}
393
394static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
395		unsigned long new_addr, unsigned long new_len, bool *locked)
 
 
 
396{
397	struct mm_struct *mm = current->mm;
398	struct vm_area_struct *vma;
399	unsigned long ret = -EINVAL;
400	unsigned long charged = 0;
401	unsigned long map_flags;
402
403	if (new_addr & ~PAGE_MASK)
404		goto out;
405
406	if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
407		goto out;
408
409	/* Check if the location we're moving into overlaps the
410	 * old location at all, and fail if it does.
411	 */
412	if ((new_addr <= addr) && (new_addr+new_len) > addr)
413		goto out;
414
415	if ((addr <= new_addr) && (addr+old_len) > new_addr)
416		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
417
418	ret = do_munmap(mm, new_addr, new_len);
419	if (ret)
420		goto out;
 
 
421
422	if (old_len >= new_len) {
423		ret = do_munmap(mm, addr+new_len, old_len - new_len);
424		if (ret && old_len != new_len)
425			goto out;
426		old_len = new_len;
427	}
428
429	vma = vma_to_resize(addr, old_len, new_len, &charged);
430	if (IS_ERR(vma)) {
431		ret = PTR_ERR(vma);
432		goto out;
433	}
434
435	map_flags = MAP_FIXED;
 
 
 
 
 
 
 
 
 
436	if (vma->vm_flags & VM_MAYSHARE)
437		map_flags |= MAP_SHARED;
438
439	ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
440				((addr - vma->vm_start) >> PAGE_SHIFT),
441				map_flags);
442	if (ret & ~PAGE_MASK)
443		goto out1;
444
445	ret = move_vma(vma, addr, old_len, new_len, new_addr, locked);
446	if (!(ret & ~PAGE_MASK))
 
 
 
 
 
 
447		goto out;
 
448out1:
449	vm_unacct_memory(charged);
450
451out:
452	return ret;
453}
454
455static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
456{
457	unsigned long end = vma->vm_end + delta;
458	if (end < vma->vm_end) /* overflow */
459		return 0;
460	if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
461		return 0;
462	if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
463			      0, MAP_FIXED) & ~PAGE_MASK)
464		return 0;
465	return 1;
466}
467
468/*
469 * Expand (or shrink) an existing mapping, potentially moving it at the
470 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
471 *
472 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
473 * This option implies MREMAP_MAYMOVE.
474 */
475SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
476		unsigned long, new_len, unsigned long, flags,
477		unsigned long, new_addr)
478{
479	struct mm_struct *mm = current->mm;
480	struct vm_area_struct *vma;
481	unsigned long ret = -EINVAL;
482	unsigned long charged = 0;
483	bool locked = false;
 
 
 
 
484
485	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 
 
 
 
 
 
 
 
 
 
 
 
486		return ret;
487
488	if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
489		return ret;
490
491	if (addr & ~PAGE_MASK)
 
 
 
 
 
 
 
 
 
492		return ret;
493
494	old_len = PAGE_ALIGN(old_len);
495	new_len = PAGE_ALIGN(new_len);
496
497	/*
498	 * We allow a zero old-len as a special case
499	 * for DOS-emu "duplicate shm area" thing. But
500	 * a zero new-len is nonsensical.
501	 */
502	if (!new_len)
503		return ret;
504
505	down_write(&current->mm->mmap_sem);
 
506
507	if (flags & MREMAP_FIXED) {
508		ret = mremap_to(addr, old_len, new_addr, new_len,
509				&locked);
 
510		goto out;
511	}
512
513	/*
514	 * Always allow a shrinking remap: that just unmaps
515	 * the unnecessary pages..
516	 * do_munmap does all the needed commit accounting
 
517	 */
518	if (old_len >= new_len) {
519		ret = do_munmap(mm, addr+new_len, old_len - new_len);
520		if (ret && old_len != new_len)
 
 
 
 
521			goto out;
 
 
 
522		ret = addr;
523		goto out;
524	}
525
526	/*
527	 * Ok, we need to grow..
528	 */
529	vma = vma_to_resize(addr, old_len, new_len, &charged);
530	if (IS_ERR(vma)) {
531		ret = PTR_ERR(vma);
532		goto out;
533	}
534
535	/* old_len exactly to the end of the area..
536	 */
537	if (old_len == vma->vm_end - addr) {
538		/* can we just expand the current mapping? */
539		if (vma_expandable(vma, new_len - old_len)) {
540			int pages = (new_len - old_len) >> PAGE_SHIFT;
541
542			if (vma_adjust(vma, vma->vm_start, addr + new_len,
543				       vma->vm_pgoff, NULL)) {
544				ret = -ENOMEM;
545				goto out;
546			}
547
548			vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
549			if (vma->vm_flags & VM_LOCKED) {
550				mm->locked_vm += pages;
551				locked = true;
552				new_addr = addr;
553			}
554			ret = addr;
555			goto out;
556		}
557	}
558
559	/*
560	 * We weren't able to just expand or shrink the area,
561	 * we need to create a new one and move it..
562	 */
563	ret = -ENOMEM;
564	if (flags & MREMAP_MAYMOVE) {
565		unsigned long map_flags = 0;
566		if (vma->vm_flags & VM_MAYSHARE)
567			map_flags |= MAP_SHARED;
568
569		new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
570					vma->vm_pgoff +
571					((addr - vma->vm_start) >> PAGE_SHIFT),
572					map_flags);
573		if (new_addr & ~PAGE_MASK) {
574			ret = new_addr;
575			goto out;
576		}
577
578		ret = move_vma(vma, addr, old_len, new_len, new_addr, &locked);
 
579	}
580out:
581	if (ret & ~PAGE_MASK)
582		vm_unacct_memory(charged);
583	up_write(&current->mm->mmap_sem);
 
 
 
 
 
584	if (locked && new_len > old_len)
585		mm_populate(new_addr + old_len, new_len - old_len);
 
 
 
586	return ret;
587}
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *	mm/mremap.c
   4 *
   5 *	(C) Copyright 1996 Linus Torvalds
   6 *
   7 *	Address space accounting code	<alan@lxorguk.ukuu.org.uk>
   8 *	(C) Copyright 2002 Red Hat Inc, All Rights Reserved
   9 */
  10
  11#include <linux/mm.h>
  12#include <linux/hugetlb.h>
  13#include <linux/shm.h>
  14#include <linux/ksm.h>
  15#include <linux/mman.h>
  16#include <linux/swap.h>
  17#include <linux/capability.h>
  18#include <linux/fs.h>
  19#include <linux/swapops.h>
  20#include <linux/highmem.h>
  21#include <linux/security.h>
  22#include <linux/syscalls.h>
  23#include <linux/mmu_notifier.h>
  24#include <linux/uaccess.h>
  25#include <linux/userfaultfd_k.h>
  26
 
  27#include <asm/cacheflush.h>
  28#include <asm/tlb.h>
  29#include <asm/pgalloc.h>
  30
  31#include "internal.h"
  32
  33static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
  34{
  35	pgd_t *pgd;
  36	p4d_t *p4d;
  37	pud_t *pud;
 
  38
  39	pgd = pgd_offset(mm, addr);
  40	if (pgd_none_or_clear_bad(pgd))
  41		return NULL;
  42
  43	p4d = p4d_offset(pgd, addr);
  44	if (p4d_none_or_clear_bad(p4d))
  45		return NULL;
  46
  47	pud = pud_offset(p4d, addr);
  48	if (pud_none_or_clear_bad(pud))
  49		return NULL;
  50
  51	return pud;
  52}
  53
  54static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
  55{
  56	pud_t *pud;
  57	pmd_t *pmd;
  58
  59	pud = get_old_pud(mm, addr);
  60	if (!pud)
  61		return NULL;
  62
  63	pmd = pmd_offset(pud, addr);
  64	if (pmd_none(*pmd))
  65		return NULL;
  66
  67	return pmd;
  68}
  69
  70static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma,
  71			    unsigned long addr)
  72{
  73	pgd_t *pgd;
  74	p4d_t *p4d;
  75
  76	pgd = pgd_offset(mm, addr);
  77	p4d = p4d_alloc(mm, pgd, addr);
  78	if (!p4d)
  79		return NULL;
  80
  81	return pud_alloc(mm, p4d, addr);
  82}
  83
  84static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
  85			    unsigned long addr)
  86{
  87	pud_t *pud;
  88	pmd_t *pmd;
  89
  90	pud = alloc_new_pud(mm, vma, addr);
 
  91	if (!pud)
  92		return NULL;
  93
  94	pmd = pmd_alloc(mm, pud, addr);
  95	if (!pmd)
  96		return NULL;
  97
  98	VM_BUG_ON(pmd_trans_huge(*pmd));
  99
 100	return pmd;
 101}
 102
 103static void take_rmap_locks(struct vm_area_struct *vma)
 104{
 105	if (vma->vm_file)
 106		i_mmap_lock_write(vma->vm_file->f_mapping);
 107	if (vma->anon_vma)
 108		anon_vma_lock_write(vma->anon_vma);
 109}
 110
 111static void drop_rmap_locks(struct vm_area_struct *vma)
 112{
 113	if (vma->anon_vma)
 114		anon_vma_unlock_write(vma->anon_vma);
 115	if (vma->vm_file)
 116		i_mmap_unlock_write(vma->vm_file->f_mapping);
 117}
 118
 119static pte_t move_soft_dirty_pte(pte_t pte)
 120{
 121	/*
 122	 * Set soft dirty bit so we can notice
 123	 * in userspace the ptes were moved.
 124	 */
 125#ifdef CONFIG_MEM_SOFT_DIRTY
 126	if (pte_present(pte))
 127		pte = pte_mksoft_dirty(pte);
 128	else if (is_swap_pte(pte))
 129		pte = pte_swp_mksoft_dirty(pte);
 
 
 130#endif
 131	return pte;
 132}
 133
 134static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 135		unsigned long old_addr, unsigned long old_end,
 136		struct vm_area_struct *new_vma, pmd_t *new_pmd,
 137		unsigned long new_addr, bool need_rmap_locks)
 138{
 
 
 139	struct mm_struct *mm = vma->vm_mm;
 140	pte_t *old_pte, *new_pte, pte;
 141	spinlock_t *old_ptl, *new_ptl;
 142	bool force_flush = false;
 143	unsigned long len = old_end - old_addr;
 144
 145	/*
 146	 * When need_rmap_locks is true, we take the i_mmap_rwsem and anon_vma
 147	 * locks to ensure that rmap will always observe either the old or the
 148	 * new ptes. This is the easiest way to avoid races with
 149	 * truncate_pagecache(), page migration, etc...
 150	 *
 151	 * When need_rmap_locks is false, we use other ways to avoid
 152	 * such races:
 153	 *
 154	 * - During exec() shift_arg_pages(), we use a specially tagged vma
 155	 *   which rmap call sites look for using vma_is_temporary_stack().
 156	 *
 157	 * - During mremap(), new_vma is often known to be placed after vma
 158	 *   in rmap traversal order. This ensures rmap will always observe
 159	 *   either the old pte, or the new pte, or both (the page table locks
 160	 *   serialize access to individual ptes, but only rmap traversal
 161	 *   order guarantees that we won't miss both the old and new ptes).
 162	 */
 163	if (need_rmap_locks)
 164		take_rmap_locks(vma);
 
 
 
 
 
 
 
 
 165
 166	/*
 167	 * We don't have to worry about the ordering of src and dst
 168	 * pte locks because exclusive mmap_lock prevents deadlock.
 169	 */
 170	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
 171	new_pte = pte_offset_map(new_pmd, new_addr);
 172	new_ptl = pte_lockptr(mm, new_pmd);
 173	if (new_ptl != old_ptl)
 174		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 175	flush_tlb_batched_pending(vma->vm_mm);
 176	arch_enter_lazy_mmu_mode();
 177
 178	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 179				   new_pte++, new_addr += PAGE_SIZE) {
 180		if (pte_none(*old_pte))
 181			continue;
 182
 183		pte = ptep_get_and_clear(mm, old_addr, old_pte);
 184		/*
 185		 * If we are remapping a valid PTE, make sure
 186		 * to flush TLB before we drop the PTL for the
 187		 * PTE.
 188		 *
 189		 * NOTE! Both old and new PTL matter: the old one
 190		 * for racing with page_mkclean(), the new one to
 191		 * make sure the physical page stays valid until
 192		 * the TLB entry for the old mapping has been
 193		 * flushed.
 194		 */
 195		if (pte_present(pte))
 196			force_flush = true;
 197		pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
 198		pte = move_soft_dirty_pte(pte);
 199		set_pte_at(mm, new_addr, new_pte, pte);
 200	}
 201
 202	arch_leave_lazy_mmu_mode();
 203	if (force_flush)
 204		flush_tlb_range(vma, old_end - len, old_end);
 205	if (new_ptl != old_ptl)
 206		spin_unlock(new_ptl);
 207	pte_unmap(new_pte - 1);
 208	pte_unmap_unlock(old_pte - 1, old_ptl);
 209	if (need_rmap_locks)
 210		drop_rmap_locks(vma);
 211}
 212
 213#ifndef arch_supports_page_table_move
 214#define arch_supports_page_table_move arch_supports_page_table_move
 215static inline bool arch_supports_page_table_move(void)
 216{
 217	return IS_ENABLED(CONFIG_HAVE_MOVE_PMD) ||
 218		IS_ENABLED(CONFIG_HAVE_MOVE_PUD);
 219}
 220#endif
 221
 222#ifdef CONFIG_HAVE_MOVE_PMD
 223static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 224		  unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
 225{
 226	spinlock_t *old_ptl, *new_ptl;
 227	struct mm_struct *mm = vma->vm_mm;
 228	pmd_t pmd;
 229
 230	if (!arch_supports_page_table_move())
 231		return false;
 232	/*
 233	 * The destination pmd shouldn't be established, free_pgtables()
 234	 * should have released it.
 235	 *
 236	 * However, there's a case during execve() where we use mremap
 237	 * to move the initial stack, and in that case the target area
 238	 * may overlap the source area (always moving down).
 239	 *
 240	 * If everything is PMD-aligned, that works fine, as moving
 241	 * each pmd down will clear the source pmd. But if we first
 242	 * have a few 4kB-only pages that get moved down, and then
 243	 * hit the "now the rest is PMD-aligned, let's do everything
 244	 * one pmd at a time", we will still have the old (now empty
 245	 * of any 4kB pages, but still there) PMD in the page table
 246	 * tree.
 247	 *
 248	 * Warn on it once - because we really should try to figure
 249	 * out how to do this better - but then say "I won't move
 250	 * this pmd".
 251	 *
 252	 * One alternative might be to just unmap the target pmd at
 253	 * this point, and verify that it really is empty. We'll see.
 254	 */
 255	if (WARN_ON_ONCE(!pmd_none(*new_pmd)))
 256		return false;
 257
 258	/*
 259	 * We don't have to worry about the ordering of src and dst
 260	 * ptlocks because exclusive mmap_lock prevents deadlock.
 261	 */
 262	old_ptl = pmd_lock(vma->vm_mm, old_pmd);
 263	new_ptl = pmd_lockptr(mm, new_pmd);
 264	if (new_ptl != old_ptl)
 265		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 266
 267	/* Clear the pmd */
 268	pmd = *old_pmd;
 269	pmd_clear(old_pmd);
 270
 271	VM_BUG_ON(!pmd_none(*new_pmd));
 272
 273	pmd_populate(mm, new_pmd, pmd_pgtable(pmd));
 274	flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
 275	if (new_ptl != old_ptl)
 276		spin_unlock(new_ptl);
 277	spin_unlock(old_ptl);
 278
 279	return true;
 280}
 281#else
 282static inline bool move_normal_pmd(struct vm_area_struct *vma,
 283		unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
 284		pmd_t *new_pmd)
 285{
 286	return false;
 287}
 288#endif
 289
 290#if CONFIG_PGTABLE_LEVELS > 2 && defined(CONFIG_HAVE_MOVE_PUD)
 291static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
 292		  unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 293{
 294	spinlock_t *old_ptl, *new_ptl;
 295	struct mm_struct *mm = vma->vm_mm;
 296	pud_t pud;
 297
 298	if (!arch_supports_page_table_move())
 299		return false;
 300	/*
 301	 * The destination pud shouldn't be established, free_pgtables()
 302	 * should have released it.
 303	 */
 304	if (WARN_ON_ONCE(!pud_none(*new_pud)))
 305		return false;
 306
 307	/*
 308	 * We don't have to worry about the ordering of src and dst
 309	 * ptlocks because exclusive mmap_lock prevents deadlock.
 310	 */
 311	old_ptl = pud_lock(vma->vm_mm, old_pud);
 312	new_ptl = pud_lockptr(mm, new_pud);
 313	if (new_ptl != old_ptl)
 314		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 315
 316	/* Clear the pud */
 317	pud = *old_pud;
 318	pud_clear(old_pud);
 319
 320	VM_BUG_ON(!pud_none(*new_pud));
 321
 322	pud_populate(mm, new_pud, pud_pgtable(pud));
 323	flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
 324	if (new_ptl != old_ptl)
 325		spin_unlock(new_ptl);
 326	spin_unlock(old_ptl);
 327
 328	return true;
 329}
 330#else
 331static inline bool move_normal_pud(struct vm_area_struct *vma,
 332		unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
 333		pud_t *new_pud)
 334{
 335	return false;
 336}
 337#endif
 338
 339#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
 340static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 341			  unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 342{
 343	spinlock_t *old_ptl, *new_ptl;
 344	struct mm_struct *mm = vma->vm_mm;
 345	pud_t pud;
 346
 347	/*
 348	 * The destination pud shouldn't be established, free_pgtables()
 349	 * should have released it.
 350	 */
 351	if (WARN_ON_ONCE(!pud_none(*new_pud)))
 352		return false;
 353
 354	/*
 355	 * We don't have to worry about the ordering of src and dst
 356	 * ptlocks because exclusive mmap_lock prevents deadlock.
 357	 */
 358	old_ptl = pud_lock(vma->vm_mm, old_pud);
 359	new_ptl = pud_lockptr(mm, new_pud);
 360	if (new_ptl != old_ptl)
 361		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
 362
 363	/* Clear the pud */
 364	pud = *old_pud;
 365	pud_clear(old_pud);
 366
 367	VM_BUG_ON(!pud_none(*new_pud));
 368
 369	/* Set the new pud */
 370	/* mark soft_ditry when we add pud level soft dirty support */
 371	set_pud_at(mm, new_addr, new_pud, pud);
 372	flush_pud_tlb_range(vma, old_addr, old_addr + HPAGE_PUD_SIZE);
 373	if (new_ptl != old_ptl)
 374		spin_unlock(new_ptl);
 375	spin_unlock(old_ptl);
 376
 377	return true;
 378}
 379#else
 380static bool move_huge_pud(struct vm_area_struct *vma, unsigned long old_addr,
 381			  unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
 382{
 383	WARN_ON_ONCE(1);
 384	return false;
 385
 386}
 387#endif
 388
 389enum pgt_entry {
 390	NORMAL_PMD,
 391	HPAGE_PMD,
 392	NORMAL_PUD,
 393	HPAGE_PUD,
 394};
 395
 396/*
 397 * Returns an extent of the corresponding size for the pgt_entry specified if
 398 * valid. Else returns a smaller extent bounded by the end of the source and
 399 * destination pgt_entry.
 400 */
 401static __always_inline unsigned long get_extent(enum pgt_entry entry,
 402			unsigned long old_addr, unsigned long old_end,
 403			unsigned long new_addr)
 404{
 405	unsigned long next, extent, mask, size;
 406
 407	switch (entry) {
 408	case HPAGE_PMD:
 409	case NORMAL_PMD:
 410		mask = PMD_MASK;
 411		size = PMD_SIZE;
 412		break;
 413	case HPAGE_PUD:
 414	case NORMAL_PUD:
 415		mask = PUD_MASK;
 416		size = PUD_SIZE;
 417		break;
 418	default:
 419		BUILD_BUG();
 420		break;
 421	}
 422
 423	next = (old_addr + size) & mask;
 424	/* even if next overflowed, extent below will be ok */
 425	extent = next - old_addr;
 426	if (extent > old_end - old_addr)
 427		extent = old_end - old_addr;
 428	next = (new_addr + size) & mask;
 429	if (extent > next - new_addr)
 430		extent = next - new_addr;
 431	return extent;
 432}
 433
 434/*
 435 * Attempts to speedup the move by moving entry at the level corresponding to
 436 * pgt_entry. Returns true if the move was successful, else false.
 437 */
 438static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
 439			unsigned long old_addr, unsigned long new_addr,
 440			void *old_entry, void *new_entry, bool need_rmap_locks)
 441{
 442	bool moved = false;
 443
 444	/* See comment in move_ptes() */
 445	if (need_rmap_locks)
 446		take_rmap_locks(vma);
 447
 448	switch (entry) {
 449	case NORMAL_PMD:
 450		moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
 451					new_entry);
 452		break;
 453	case NORMAL_PUD:
 454		moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
 455					new_entry);
 456		break;
 457	case HPAGE_PMD:
 458		moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 459			move_huge_pmd(vma, old_addr, new_addr, old_entry,
 460				      new_entry);
 461		break;
 462	case HPAGE_PUD:
 463		moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
 464			move_huge_pud(vma, old_addr, new_addr, old_entry,
 465				      new_entry);
 466		break;
 467
 468	default:
 469		WARN_ON_ONCE(1);
 470		break;
 471	}
 472
 473	if (need_rmap_locks)
 474		drop_rmap_locks(vma);
 475
 476	return moved;
 477}
 478
 479unsigned long move_page_tables(struct vm_area_struct *vma,
 480		unsigned long old_addr, struct vm_area_struct *new_vma,
 481		unsigned long new_addr, unsigned long len,
 482		bool need_rmap_locks)
 483{
 484	unsigned long extent, old_end;
 485	struct mmu_notifier_range range;
 486	pmd_t *old_pmd, *new_pmd;
 487	pud_t *old_pud, *new_pud;
 
 
 488
 489	old_end = old_addr + len;
 490	flush_cache_range(vma, old_addr, old_end);
 491
 492	mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
 493				old_addr, old_end);
 494	mmu_notifier_invalidate_range_start(&range);
 495
 496	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 497		cond_resched();
 498		/*
 499		 * If extent is PUD-sized try to speed up the move by moving at the
 500		 * PUD level if possible.
 501		 */
 502		extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
 503
 504		old_pud = get_old_pud(vma->vm_mm, old_addr);
 505		if (!old_pud)
 506			continue;
 507		new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
 508		if (!new_pud)
 509			break;
 510		if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
 511			if (extent == HPAGE_PUD_SIZE) {
 512				move_pgt_entry(HPAGE_PUD, vma, old_addr, new_addr,
 513					       old_pud, new_pud, need_rmap_locks);
 514				/* We ignore and continue on error? */
 515				continue;
 516			}
 517		} else if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
 518
 519			if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
 520					   old_pud, new_pud, true))
 521				continue;
 522		}
 523
 524		extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 525		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 526		if (!old_pmd)
 527			continue;
 528		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 529		if (!new_pmd)
 530			break;
 531		if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
 532		    pmd_devmap(*old_pmd)) {
 533			if (extent == HPAGE_PMD_SIZE &&
 534			    move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
 535					   old_pmd, new_pmd, need_rmap_locks))
 536				continue;
 537			split_huge_pmd(vma, old_pmd, old_addr);
 538			if (pmd_trans_unstable(old_pmd))
 539				continue;
 540		} else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
 541			   extent == PMD_SIZE) {
 542			/*
 543			 * If the extent is PMD-sized, try to speed the move by
 544			 * moving at the PMD level if possible.
 545			 */
 546			if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
 547					   old_pmd, new_pmd, true))
 548				continue;
 
 
 
 
 549		}
 550
 551		if (pte_alloc(new_vma->vm_mm, new_pmd))
 552			break;
 553		move_ptes(vma, old_pmd, old_addr, old_addr + extent, new_vma,
 554			  new_pmd, new_addr, need_rmap_locks);
 
 
 
 
 
 
 555	}
 
 
 556
 557	mmu_notifier_invalidate_range_end(&range);
 558
 559	return len + old_addr - old_end;	/* how much done */
 560}
 561
 562static unsigned long move_vma(struct vm_area_struct *vma,
 563		unsigned long old_addr, unsigned long old_len,
 564		unsigned long new_len, unsigned long new_addr,
 565		bool *locked, unsigned long flags,
 566		struct vm_userfaultfd_ctx *uf, struct list_head *uf_unmap)
 567{
 568	struct mm_struct *mm = vma->vm_mm;
 569	struct vm_area_struct *new_vma;
 570	unsigned long vm_flags = vma->vm_flags;
 571	unsigned long new_pgoff;
 572	unsigned long moved_len;
 573	unsigned long excess = 0;
 574	unsigned long hiwater_vm;
 575	int split = 0;
 576	int err = 0;
 577	bool need_rmap_locks;
 578
 579	/*
 580	 * We'd prefer to avoid failure later on in do_munmap:
 581	 * which may split one vma into three before unmapping.
 582	 */
 583	if (mm->map_count >= sysctl_max_map_count - 3)
 584		return -ENOMEM;
 585
 586	if (vma->vm_ops && vma->vm_ops->may_split) {
 587		if (vma->vm_start != old_addr)
 588			err = vma->vm_ops->may_split(vma, old_addr);
 589		if (!err && vma->vm_end != old_addr + old_len)
 590			err = vma->vm_ops->may_split(vma, old_addr + old_len);
 591		if (err)
 592			return err;
 593	}
 594
 595	/*
 596	 * Advise KSM to break any KSM pages in the area to be moved:
 597	 * it would be confusing if they were to turn up at the new
 598	 * location, where they happen to coincide with different KSM
 599	 * pages recently unmapped.  But leave vma->vm_flags as it was,
 600	 * so KSM can come around to merge on vma and new_vma afterwards.
 601	 */
 602	err = ksm_madvise(vma, old_addr, old_addr + old_len,
 603						MADV_UNMERGEABLE, &vm_flags);
 604	if (err)
 605		return err;
 606
 607	if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT)) {
 608		if (security_vm_enough_memory_mm(mm, new_len >> PAGE_SHIFT))
 609			return -ENOMEM;
 610	}
 611
 612	new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 613	new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff,
 614			   &need_rmap_locks);
 615	if (!new_vma) {
 616		if (unlikely(flags & MREMAP_DONTUNMAP && vm_flags & VM_ACCOUNT))
 617			vm_unacct_memory(new_len >> PAGE_SHIFT);
 618		return -ENOMEM;
 619	}
 620
 621	moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len,
 622				     need_rmap_locks);
 623	if (moved_len < old_len) {
 624		err = -ENOMEM;
 625	} else if (vma->vm_ops && vma->vm_ops->mremap) {
 626		err = vma->vm_ops->mremap(new_vma);
 627	}
 628
 629	if (unlikely(err)) {
 630		/*
 631		 * On error, move entries back from new area to old,
 632		 * which will succeed since page tables still there,
 633		 * and then proceed to unmap new area instead of old.
 634		 */
 635		move_page_tables(new_vma, new_addr, vma, old_addr, moved_len,
 636				 true);
 637		vma = new_vma;
 638		old_len = new_len;
 639		old_addr = new_addr;
 640		new_addr = err;
 641	} else {
 642		mremap_userfaultfd_prep(new_vma, uf);
 643	}
 644
 645	/* Conceal VM_ACCOUNT so old reservation is not undone */
 646	if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP)) {
 647		vma->vm_flags &= ~VM_ACCOUNT;
 648		excess = vma->vm_end - vma->vm_start - old_len;
 649		if (old_addr > vma->vm_start &&
 650		    old_addr + old_len < vma->vm_end)
 651			split = 1;
 652	}
 653
 654	/*
 655	 * If we failed to move page tables we still do total_vm increment
 656	 * since do_munmap() will decrement it by old_len == new_len.
 657	 *
 658	 * Since total_vm is about to be raised artificially high for a
 659	 * moment, we need to restore high watermark afterwards: if stats
 660	 * are taken meanwhile, total_vm and hiwater_vm appear too high.
 661	 * If this were a serious issue, we'd add a flag to do_munmap().
 662	 */
 663	hiwater_vm = mm->hiwater_vm;
 664	vm_stat_account(mm, vma->vm_flags, new_len >> PAGE_SHIFT);
 665
 666	/* Tell pfnmap has moved from this vma */
 667	if (unlikely(vma->vm_flags & VM_PFNMAP))
 668		untrack_pfn_moved(vma);
 669
 670	if (unlikely(!err && (flags & MREMAP_DONTUNMAP))) {
 671		/* We always clear VM_LOCKED[ONFAULT] on the old vma */
 672		vma->vm_flags &= VM_LOCKED_CLEAR_MASK;
 673
 674		/*
 675		 * anon_vma links of the old vma is no longer needed after its page
 676		 * table has been moved.
 677		 */
 678		if (new_vma != vma && vma->vm_start == old_addr &&
 679			vma->vm_end == (old_addr + old_len))
 680			unlink_anon_vmas(vma);
 681
 682		/* Because we won't unmap we don't need to touch locked_vm */
 683		return new_addr;
 684	}
 685
 686	if (do_munmap(mm, old_addr, old_len, uf_unmap) < 0) {
 687		/* OOM: unable to split vma, just get accounts right */
 688		if (vm_flags & VM_ACCOUNT && !(flags & MREMAP_DONTUNMAP))
 689			vm_acct_memory(new_len >> PAGE_SHIFT);
 690		excess = 0;
 691	}
 692
 693	if (vm_flags & VM_LOCKED) {
 694		mm->locked_vm += new_len >> PAGE_SHIFT;
 695		*locked = true;
 696	}
 697
 698	mm->hiwater_vm = hiwater_vm;
 699
 700	/* Restore VM_ACCOUNT if one or two pieces of vma left */
 701	if (excess) {
 702		vma->vm_flags |= VM_ACCOUNT;
 703		if (split)
 704			vma->vm_next->vm_flags |= VM_ACCOUNT;
 705	}
 706
 
 
 
 
 
 707	return new_addr;
 708}
 709
 710static struct vm_area_struct *vma_to_resize(unsigned long addr,
 711	unsigned long old_len, unsigned long new_len, unsigned long flags,
 712	unsigned long *p)
 713{
 714	struct mm_struct *mm = current->mm;
 715	struct vm_area_struct *vma;
 716	unsigned long pgoff;
 717
 718	vma = vma_lookup(mm, addr);
 719	if (!vma)
 720		return ERR_PTR(-EFAULT);
 721
 722	/*
 723	 * !old_len is a special case where an attempt is made to 'duplicate'
 724	 * a mapping.  This makes no sense for private mappings as it will
 725	 * instead create a fresh/new mapping unrelated to the original.  This
 726	 * is contrary to the basic idea of mremap which creates new mappings
 727	 * based on the original.  There are no known use cases for this
 728	 * behavior.  As a result, fail such attempts.
 729	 */
 730	if (!old_len && !(vma->vm_flags & (VM_SHARED | VM_MAYSHARE))) {
 731		pr_warn_once("%s (%d): attempted to duplicate a private mapping with mremap.  This is not supported.\n", current->comm, current->pid);
 732		return ERR_PTR(-EINVAL);
 733	}
 734
 735	if ((flags & MREMAP_DONTUNMAP) &&
 736			(vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)))
 737		return ERR_PTR(-EINVAL);
 738
 739	if (is_vm_hugetlb_page(vma))
 740		return ERR_PTR(-EINVAL);
 741
 742	/* We can't remap across vm area boundaries */
 743	if (old_len > vma->vm_end - addr)
 744		return ERR_PTR(-EFAULT);
 745
 746	if (new_len == old_len)
 747		return vma;
 748
 749	/* Need to be careful about a growing mapping */
 750	pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
 751	pgoff += vma->vm_pgoff;
 752	if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
 753		return ERR_PTR(-EINVAL);
 754
 755	if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
 756		return ERR_PTR(-EFAULT);
 
 
 
 
 
 757
 758	if (vma->vm_flags & VM_LOCKED) {
 759		unsigned long locked, lock_limit;
 760		locked = mm->locked_vm << PAGE_SHIFT;
 761		lock_limit = rlimit(RLIMIT_MEMLOCK);
 762		locked += new_len - old_len;
 763		if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 764			return ERR_PTR(-EAGAIN);
 765	}
 766
 767	if (!may_expand_vm(mm, vma->vm_flags,
 768				(new_len - old_len) >> PAGE_SHIFT))
 769		return ERR_PTR(-ENOMEM);
 770
 771	if (vma->vm_flags & VM_ACCOUNT) {
 772		unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
 773		if (security_vm_enough_memory_mm(mm, charged))
 774			return ERR_PTR(-ENOMEM);
 775		*p = charged;
 776	}
 777
 778	return vma;
 
 
 
 
 
 
 
 
 
 779}
 780
 781static unsigned long mremap_to(unsigned long addr, unsigned long old_len,
 782		unsigned long new_addr, unsigned long new_len, bool *locked,
 783		unsigned long flags, struct vm_userfaultfd_ctx *uf,
 784		struct list_head *uf_unmap_early,
 785		struct list_head *uf_unmap)
 786{
 787	struct mm_struct *mm = current->mm;
 788	struct vm_area_struct *vma;
 789	unsigned long ret = -EINVAL;
 790	unsigned long charged = 0;
 791	unsigned long map_flags = 0;
 792
 793	if (offset_in_page(new_addr))
 794		goto out;
 795
 796	if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 797		goto out;
 798
 799	/* Ensure the old/new locations do not overlap */
 800	if (addr + old_len > new_addr && new_addr + new_len > addr)
 
 
 801		goto out;
 802
 803	/*
 804	 * move_vma() need us to stay 4 maps below the threshold, otherwise
 805	 * it will bail out at the very beginning.
 806	 * That is a problem if we have already unmaped the regions here
 807	 * (new_addr, and old_addr), because userspace will not know the
 808	 * state of the vma's after it gets -ENOMEM.
 809	 * So, to avoid such scenario we can pre-compute if the whole
 810	 * operation has high chances to success map-wise.
 811	 * Worst-scenario case is when both vma's (new_addr and old_addr) get
 812	 * split in 3 before unmapping it.
 813	 * That means 2 more maps (1 for each) to the ones we already hold.
 814	 * Check whether current map count plus 2 still leads us to 4 maps below
 815	 * the threshold, otherwise return -ENOMEM here to be more safe.
 816	 */
 817	if ((mm->map_count + 2) >= sysctl_max_map_count - 3)
 818		return -ENOMEM;
 819
 820	if (flags & MREMAP_FIXED) {
 821		ret = do_munmap(mm, new_addr, new_len, uf_unmap_early);
 822		if (ret)
 823			goto out;
 824	}
 825
 826	if (old_len >= new_len) {
 827		ret = do_munmap(mm, addr+new_len, old_len - new_len, uf_unmap);
 828		if (ret && old_len != new_len)
 829			goto out;
 830		old_len = new_len;
 831	}
 832
 833	vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 834	if (IS_ERR(vma)) {
 835		ret = PTR_ERR(vma);
 836		goto out;
 837	}
 838
 839	/* MREMAP_DONTUNMAP expands by old_len since old_len == new_len */
 840	if (flags & MREMAP_DONTUNMAP &&
 841		!may_expand_vm(mm, vma->vm_flags, old_len >> PAGE_SHIFT)) {
 842		ret = -ENOMEM;
 843		goto out;
 844	}
 845
 846	if (flags & MREMAP_FIXED)
 847		map_flags |= MAP_FIXED;
 848
 849	if (vma->vm_flags & VM_MAYSHARE)
 850		map_flags |= MAP_SHARED;
 851
 852	ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
 853				((addr - vma->vm_start) >> PAGE_SHIFT),
 854				map_flags);
 855	if (IS_ERR_VALUE(ret))
 856		goto out1;
 857
 858	/* We got a new mapping */
 859	if (!(flags & MREMAP_FIXED))
 860		new_addr = ret;
 861
 862	ret = move_vma(vma, addr, old_len, new_len, new_addr, locked, flags, uf,
 863		       uf_unmap);
 864
 865	if (!(offset_in_page(ret)))
 866		goto out;
 867
 868out1:
 869	vm_unacct_memory(charged);
 870
 871out:
 872	return ret;
 873}
 874
 875static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
 876{
 877	unsigned long end = vma->vm_end + delta;
 878	if (end < vma->vm_end) /* overflow */
 879		return 0;
 880	if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
 881		return 0;
 882	if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
 883			      0, MAP_FIXED) & ~PAGE_MASK)
 884		return 0;
 885	return 1;
 886}
 887
 888/*
 889 * Expand (or shrink) an existing mapping, potentially moving it at the
 890 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 891 *
 892 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 893 * This option implies MREMAP_MAYMOVE.
 894 */
 895SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
 896		unsigned long, new_len, unsigned long, flags,
 897		unsigned long, new_addr)
 898{
 899	struct mm_struct *mm = current->mm;
 900	struct vm_area_struct *vma;
 901	unsigned long ret = -EINVAL;
 902	unsigned long charged = 0;
 903	bool locked = false;
 904	bool downgraded = false;
 905	struct vm_userfaultfd_ctx uf = NULL_VM_UFFD_CTX;
 906	LIST_HEAD(uf_unmap_early);
 907	LIST_HEAD(uf_unmap);
 908
 909	/*
 910	 * There is a deliberate asymmetry here: we strip the pointer tag
 911	 * from the old address but leave the new address alone. This is
 912	 * for consistency with mmap(), where we prevent the creation of
 913	 * aliasing mappings in userspace by leaving the tag bits of the
 914	 * mapping address intact. A non-zero tag will cause the subsequent
 915	 * range checks to reject the address as invalid.
 916	 *
 917	 * See Documentation/arm64/tagged-address-abi.rst for more information.
 918	 */
 919	addr = untagged_addr(addr);
 920
 921	if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE | MREMAP_DONTUNMAP))
 922		return ret;
 923
 924	if (flags & MREMAP_FIXED && !(flags & MREMAP_MAYMOVE))
 925		return ret;
 926
 927	/*
 928	 * MREMAP_DONTUNMAP is always a move and it does not allow resizing
 929	 * in the process.
 930	 */
 931	if (flags & MREMAP_DONTUNMAP &&
 932			(!(flags & MREMAP_MAYMOVE) || old_len != new_len))
 933		return ret;
 934
 935
 936	if (offset_in_page(addr))
 937		return ret;
 938
 939	old_len = PAGE_ALIGN(old_len);
 940	new_len = PAGE_ALIGN(new_len);
 941
 942	/*
 943	 * We allow a zero old-len as a special case
 944	 * for DOS-emu "duplicate shm area" thing. But
 945	 * a zero new-len is nonsensical.
 946	 */
 947	if (!new_len)
 948		return ret;
 949
 950	if (mmap_write_lock_killable(current->mm))
 951		return -EINTR;
 952
 953	if (flags & (MREMAP_FIXED | MREMAP_DONTUNMAP)) {
 954		ret = mremap_to(addr, old_len, new_addr, new_len,
 955				&locked, flags, &uf, &uf_unmap_early,
 956				&uf_unmap);
 957		goto out;
 958	}
 959
 960	/*
 961	 * Always allow a shrinking remap: that just unmaps
 962	 * the unnecessary pages..
 963	 * __do_munmap does all the needed commit accounting, and
 964	 * downgrades mmap_lock to read if so directed.
 965	 */
 966	if (old_len >= new_len) {
 967		int retval;
 968
 969		retval = __do_munmap(mm, addr+new_len, old_len - new_len,
 970				  &uf_unmap, true);
 971		if (retval < 0 && old_len != new_len) {
 972			ret = retval;
 973			goto out;
 974		/* Returning 1 indicates mmap_lock is downgraded to read. */
 975		} else if (retval == 1)
 976			downgraded = true;
 977		ret = addr;
 978		goto out;
 979	}
 980
 981	/*
 982	 * Ok, we need to grow..
 983	 */
 984	vma = vma_to_resize(addr, old_len, new_len, flags, &charged);
 985	if (IS_ERR(vma)) {
 986		ret = PTR_ERR(vma);
 987		goto out;
 988	}
 989
 990	/* old_len exactly to the end of the area..
 991	 */
 992	if (old_len == vma->vm_end - addr) {
 993		/* can we just expand the current mapping? */
 994		if (vma_expandable(vma, new_len - old_len)) {
 995			int pages = (new_len - old_len) >> PAGE_SHIFT;
 996
 997			if (vma_adjust(vma, vma->vm_start, addr + new_len,
 998				       vma->vm_pgoff, NULL)) {
 999				ret = -ENOMEM;
1000				goto out;
1001			}
1002
1003			vm_stat_account(mm, vma->vm_flags, pages);
1004			if (vma->vm_flags & VM_LOCKED) {
1005				mm->locked_vm += pages;
1006				locked = true;
1007				new_addr = addr;
1008			}
1009			ret = addr;
1010			goto out;
1011		}
1012	}
1013
1014	/*
1015	 * We weren't able to just expand or shrink the area,
1016	 * we need to create a new one and move it..
1017	 */
1018	ret = -ENOMEM;
1019	if (flags & MREMAP_MAYMOVE) {
1020		unsigned long map_flags = 0;
1021		if (vma->vm_flags & VM_MAYSHARE)
1022			map_flags |= MAP_SHARED;
1023
1024		new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
1025					vma->vm_pgoff +
1026					((addr - vma->vm_start) >> PAGE_SHIFT),
1027					map_flags);
1028		if (IS_ERR_VALUE(new_addr)) {
1029			ret = new_addr;
1030			goto out;
1031		}
1032
1033		ret = move_vma(vma, addr, old_len, new_len, new_addr,
1034			       &locked, flags, &uf, &uf_unmap);
1035	}
1036out:
1037	if (offset_in_page(ret)) {
1038		vm_unacct_memory(charged);
1039		locked = false;
1040	}
1041	if (downgraded)
1042		mmap_read_unlock(current->mm);
1043	else
1044		mmap_write_unlock(current->mm);
1045	if (locked && new_len > old_len)
1046		mm_populate(new_addr + old_len, new_len - old_len);
1047	userfaultfd_unmap_complete(mm, &uf_unmap_early);
1048	mremap_userfaultfd_complete(&uf, addr, ret, old_len);
1049	userfaultfd_unmap_complete(mm, &uf_unmap);
1050	return ret;
1051}