1// SPDX-License-Identifier: GPL-2.0
  2#include <linux/pagewalk.h>
  3#include <linux/hugetlb.h>
  4#include <linux/bitops.h>
  5#include <linux/mmu_notifier.h>
  6#include <asm/cacheflush.h>
  7#include <asm/tlbflush.h>
  8
  9/**
 10 * struct wp_walk - Private struct for pagetable walk callbacks
 11 * @range: Range for mmu notifiers
 12 * @tlbflush_start: Address of first modified pte
 13 * @tlbflush_end: Address of last modified pte + 1
 14 * @total: Total number of modified ptes
 15 */
 16struct wp_walk {
 17	struct mmu_notifier_range range;
 18	unsigned long tlbflush_start;
 19	unsigned long tlbflush_end;
 20	unsigned long total;
 21};
 22
 23/**
 24 * wp_pte - Write-protect a pte
 25 * @pte: Pointer to the pte
 26 * @addr: The start of protecting virtual address
 27 * @end: The end of protecting virtual address
 28 * @walk: pagetable walk callback argument
 29 *
 30 * The function write-protects a pte and records the range in
 31 * virtual address space of touched ptes for efficient range TLB flushes.
 32 */
 33static int wp_pte(pte_t *pte, unsigned long addr, unsigned long end,
 34		  struct mm_walk *walk)
 35{
 36	struct wp_walk *wpwalk = walk->private;
 37	pte_t ptent = *pte;
 38
 39	if (pte_write(ptent)) {
 40		pte_t old_pte = ptep_modify_prot_start(walk->vma, addr, pte);
 41
 42		ptent = pte_wrprotect(old_pte);
 43		ptep_modify_prot_commit(walk->vma, addr, pte, old_pte, ptent);
 44		wpwalk->total++;
 45		wpwalk->tlbflush_start = min(wpwalk->tlbflush_start, addr);
 46		wpwalk->tlbflush_end = max(wpwalk->tlbflush_end,
 47					   addr + PAGE_SIZE);
 48	}
 49
 50	return 0;
 51}
 52
 53/**
 54 * struct clean_walk - Private struct for the clean_record_pte function.
 55 * @base: struct wp_walk we derive from
 56 * @bitmap_pgoff: Address_space Page offset of the first bit in @bitmap
 57 * @bitmap: Bitmap with one bit for each page offset in the address_space range
 58 * covered.
 59 * @start: Address_space page offset of first modified pte relative
 60 * to @bitmap_pgoff
 61 * @end: Address_space page offset of last modified pte relative
 62 * to @bitmap_pgoff
 63 */
 64struct clean_walk {
 65	struct wp_walk base;
 66	pgoff_t bitmap_pgoff;
 67	unsigned long *bitmap;
 68	pgoff_t start;
 69	pgoff_t end;
 70};
 71
 72#define to_clean_walk(_wpwalk) container_of(_wpwalk, struct clean_walk, base)
 73
 74/**
 75 * clean_record_pte - Clean a pte and record its address space offset in a
 76 * bitmap
 77 * @pte: Pointer to the pte
 78 * @addr: The start of virtual address to be clean
 79 * @end: The end of virtual address to be clean
 80 * @walk: pagetable walk callback argument
 81 *
 82 * The function cleans a pte and records the range in
 83 * virtual address space of touched ptes for efficient TLB flushes.
 84 * It also records dirty ptes in a bitmap representing page offsets
 85 * in the address_space, as well as the first and last of the bits
 86 * touched.
 87 */
 88static int clean_record_pte(pte_t *pte, unsigned long addr,
 89			    unsigned long end, struct mm_walk *walk)
 90{
 91	struct wp_walk *wpwalk = walk->private;
 92	struct clean_walk *cwalk = to_clean_walk(wpwalk);
 93	pte_t ptent = *pte;
 94
 95	if (pte_dirty(ptent)) {
 96		pgoff_t pgoff = ((addr - walk->vma->vm_start) >> PAGE_SHIFT) +
 97			walk->vma->vm_pgoff - cwalk->bitmap_pgoff;
 98		pte_t old_pte = ptep_modify_prot_start(walk->vma, addr, pte);
 99
100		ptent = pte_mkclean(old_pte);
101		ptep_modify_prot_commit(walk->vma, addr, pte, old_pte, ptent);
102
103		wpwalk->total++;
104		wpwalk->tlbflush_start = min(wpwalk->tlbflush_start, addr);
105		wpwalk->tlbflush_end = max(wpwalk->tlbflush_end,
106					   addr + PAGE_SIZE);
107
108		__set_bit(pgoff, cwalk->bitmap);
109		cwalk->start = min(cwalk->start, pgoff);
110		cwalk->end = max(cwalk->end, pgoff + 1);
111	}
112
113	return 0;
114}
115
116/*
117 * wp_clean_pmd_entry - The pagewalk pmd callback.
118 *
119 * Dirty-tracking should take place on the PTE level, so
120 * WARN() if encountering a dirty huge pmd.
121 * Furthermore, never split huge pmds, since that currently
122 * causes dirty info loss. The pagefault handler should do
123 * that if needed.
124 */
125static int wp_clean_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long end,
126			      struct mm_walk *walk)
127{
128	pmd_t pmdval = pmd_read_atomic(pmd);
129
130	if (!pmd_trans_unstable(&pmdval))
131		return 0;
132
133	if (pmd_none(pmdval)) {
134		walk->action = ACTION_AGAIN;
135		return 0;
136	}
137
138	/* Huge pmd, present or migrated */
139	walk->action = ACTION_CONTINUE;
140	if (pmd_trans_huge(pmdval) || pmd_devmap(pmdval))
141		WARN_ON(pmd_write(pmdval) || pmd_dirty(pmdval));
142
143	return 0;
144}
145
146/*
147 * wp_clean_pud_entry - The pagewalk pud callback.
148 *
149 * Dirty-tracking should take place on the PTE level, so
150 * WARN() if encountering a dirty huge puds.
151 * Furthermore, never split huge puds, since that currently
152 * causes dirty info loss. The pagefault handler should do
153 * that if needed.
154 */
155static int wp_clean_pud_entry(pud_t *pud, unsigned long addr, unsigned long end,
156			      struct mm_walk *walk)
157{
158	pud_t pudval = READ_ONCE(*pud);
159
160	if (!pud_trans_unstable(&pudval))
161		return 0;
162
163	if (pud_none(pudval)) {
164		walk->action = ACTION_AGAIN;
165		return 0;
166	}
167
168#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
169	/* Huge pud */
170	walk->action = ACTION_CONTINUE;
171	if (pud_trans_huge(pudval) || pud_devmap(pudval))
172		WARN_ON(pud_write(pudval) || pud_dirty(pudval));
173#endif
174
175	return 0;
176}
177
178/*
179 * wp_clean_pre_vma - The pagewalk pre_vma callback.
180 *
181 * The pre_vma callback performs the cache flush, stages the tlb flush
182 * and calls the necessary mmu notifiers.
183 */
184static int wp_clean_pre_vma(unsigned long start, unsigned long end,
185			    struct mm_walk *walk)
186{
187	struct wp_walk *wpwalk = walk->private;
188
189	wpwalk->tlbflush_start = end;
190	wpwalk->tlbflush_end = start;
191
192	mmu_notifier_range_init(&wpwalk->range, MMU_NOTIFY_PROTECTION_PAGE, 0,
193				walk->vma, walk->mm, start, end);
194	mmu_notifier_invalidate_range_start(&wpwalk->range);
195	flush_cache_range(walk->vma, start, end);
196
197	/*
198	 * We're not using tlb_gather_mmu() since typically
199	 * only a small subrange of PTEs are affected, whereas
200	 * tlb_gather_mmu() records the full range.
201	 */
202	inc_tlb_flush_pending(walk->mm);
203
204	return 0;
205}
206
207/*
208 * wp_clean_post_vma - The pagewalk post_vma callback.
209 *
210 * The post_vma callback performs the tlb flush and calls necessary mmu
211 * notifiers.
212 */
213static void wp_clean_post_vma(struct mm_walk *walk)
214{
215	struct wp_walk *wpwalk = walk->private;
216
217	if (mm_tlb_flush_nested(walk->mm))
218		flush_tlb_range(walk->vma, wpwalk->range.start,
219				wpwalk->range.end);
220	else if (wpwalk->tlbflush_end > wpwalk->tlbflush_start)
221		flush_tlb_range(walk->vma, wpwalk->tlbflush_start,
222				wpwalk->tlbflush_end);
223
224	mmu_notifier_invalidate_range_end(&wpwalk->range);
225	dec_tlb_flush_pending(walk->mm);
226}
227
228/*
229 * wp_clean_test_walk - The pagewalk test_walk callback.
230 *
231 * Won't perform dirty-tracking on COW, read-only or HUGETLB vmas.
232 */
233static int wp_clean_test_walk(unsigned long start, unsigned long end,
234			      struct mm_walk *walk)
235{
236	unsigned long vm_flags = READ_ONCE(walk->vma->vm_flags);
237
238	/* Skip non-applicable VMAs */
239	if ((vm_flags & (VM_SHARED | VM_MAYWRITE | VM_HUGETLB)) !=
240	    (VM_SHARED | VM_MAYWRITE))
241		return 1;
242
243	return 0;
244}
245
246static const struct mm_walk_ops clean_walk_ops = {
247	.pte_entry = clean_record_pte,
248	.pmd_entry = wp_clean_pmd_entry,
249	.pud_entry = wp_clean_pud_entry,
250	.test_walk = wp_clean_test_walk,
251	.pre_vma = wp_clean_pre_vma,
252	.post_vma = wp_clean_post_vma
253};
254
255static const struct mm_walk_ops wp_walk_ops = {
256	.pte_entry = wp_pte,
257	.pmd_entry = wp_clean_pmd_entry,
258	.pud_entry = wp_clean_pud_entry,
259	.test_walk = wp_clean_test_walk,
260	.pre_vma = wp_clean_pre_vma,
261	.post_vma = wp_clean_post_vma
262};
263
264/**
265 * wp_shared_mapping_range - Write-protect all ptes in an address space range
266 * @mapping: The address_space we want to write protect
267 * @first_index: The first page offset in the range
268 * @nr: Number of incremental page offsets to cover
269 *
270 * Note: This function currently skips transhuge page-table entries, since
271 * it's intended for dirty-tracking on the PTE level. It will warn on
272 * encountering transhuge write-enabled entries, though, and can easily be
273 * extended to handle them as well.
274 *
275 * Return: The number of ptes actually write-protected. Note that
276 * already write-protected ptes are not counted.
277 */
278unsigned long wp_shared_mapping_range(struct address_space *mapping,
279				      pgoff_t first_index, pgoff_t nr)
280{
281	struct wp_walk wpwalk = { .total = 0 };
282
283	i_mmap_lock_read(mapping);
284	WARN_ON(walk_page_mapping(mapping, first_index, nr, &wp_walk_ops,
285				  &wpwalk));
286	i_mmap_unlock_read(mapping);
287
288	return wpwalk.total;
289}
290EXPORT_SYMBOL_GPL(wp_shared_mapping_range);
291
292/**
293 * clean_record_shared_mapping_range - Clean and record all ptes in an
294 * address space range
295 * @mapping: The address_space we want to clean
296 * @first_index: The first page offset in the range
297 * @nr: Number of incremental page offsets to cover
298 * @bitmap_pgoff: The page offset of the first bit in @bitmap
299 * @bitmap: Pointer to a bitmap of at least @nr bits. The bitmap needs to
300 * cover the whole range @first_index..@first_index + @nr.
301 * @start: Pointer to number of the first set bit in @bitmap.
302 * is modified as new bits are set by the function.
303 * @end: Pointer to the number of the last set bit in @bitmap.
304 * none set. The value is modified as new bits are set by the function.
305 *
306 * Note: When this function returns there is no guarantee that a CPU has
307 * not already dirtied new ptes. However it will not clean any ptes not
308 * reported in the bitmap. The guarantees are as follows:
309 * a) All ptes dirty when the function starts executing will end up recorded
310 *    in the bitmap.
311 * b) All ptes dirtied after that will either remain dirty, be recorded in the
312 *    bitmap or both.
313 *
314 * If a caller needs to make sure all dirty ptes are picked up and none
315 * additional are added, it first needs to write-protect the address-space
316 * range and make sure new writers are blocked in page_mkwrite() or
317 * pfn_mkwrite(). And then after a TLB flush following the write-protection
318 * pick up all dirty bits.
319 *
320 * This function currently skips transhuge page-table entries, since
321 * it's intended for dirty-tracking on the PTE level. It will warn on
322 * encountering transhuge dirty entries, though, and can easily be extended
323 * to handle them as well.
324 *
325 * Return: The number of dirty ptes actually cleaned.
326 */
327unsigned long clean_record_shared_mapping_range(struct address_space *mapping,
328						pgoff_t first_index, pgoff_t nr,
329						pgoff_t bitmap_pgoff,
330						unsigned long *bitmap,
331						pgoff_t *start,
332						pgoff_t *end)
333{
334	bool none_set = (*start >= *end);
335	struct clean_walk cwalk = {
336		.base = { .total = 0 },
337		.bitmap_pgoff = bitmap_pgoff,
338		.bitmap = bitmap,
339		.start = none_set ? nr : *start,
340		.end = none_set ? 0 : *end,
341	};
342
343	i_mmap_lock_read(mapping);
344	WARN_ON(walk_page_mapping(mapping, first_index, nr, &clean_walk_ops,
345				  &cwalk.base));
346	i_mmap_unlock_read(mapping);
347
348	*start = cwalk.start;
349	*end = cwalk.end;
350
351	return cwalk.base.total;
352}
353EXPORT_SYMBOL_GPL(clean_record_shared_mapping_range);