Commit | Line | Data |
---|---|---|
3f49584b SP |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * DAMON Primitives for Virtual Address Spaces | |
4 | * | |
6ad59a38 | 5 | * Author: SeongJae Park <sj@kernel.org> |
3f49584b SP |
6 | */ |
7 | ||
8 | #define pr_fmt(fmt) "damon-va: " fmt | |
9 | ||
46c3a0ac | 10 | #include <linux/highmem.h> |
3f49584b | 11 | #include <linux/hugetlb.h> |
1ae41dff | 12 | #include <linux/mman.h> |
3f49584b | 13 | #include <linux/mmu_notifier.h> |
3f49584b SP |
14 | #include <linux/page_idle.h> |
15 | #include <linux/pagewalk.h> | |
8581fd40 | 16 | #include <linux/sched/mm.h> |
46c3a0ac | 17 | |
f7d911c3 | 18 | #include "ops-common.h" |
3f49584b | 19 | |
17ccae8b SP |
20 | #ifdef CONFIG_DAMON_VADDR_KUNIT_TEST |
21 | #undef DAMON_MIN_REGION | |
22 | #define DAMON_MIN_REGION 1 | |
23 | #endif | |
24 | ||
3f49584b | 25 | /* |
1971bd63 | 26 | * 't->pid' should be the pointer to the relevant 'struct pid' having reference |
3f49584b SP |
27 | * count. Caller must put the returned task, unless it is NULL. |
28 | */ | |
88f86dcf SP |
29 | static inline struct task_struct *damon_get_task_struct(struct damon_target *t) |
30 | { | |
1971bd63 | 31 | return get_pid_task(t->pid, PIDTYPE_PID); |
88f86dcf | 32 | } |
3f49584b SP |
33 | |
34 | /* | |
35 | * Get the mm_struct of the given target | |
36 | * | |
37 | * Caller _must_ put the mm_struct after use, unless it is NULL. | |
38 | * | |
39 | * Returns the mm_struct of the target on success, NULL on failure | |
40 | */ | |
41 | static struct mm_struct *damon_get_mm(struct damon_target *t) | |
42 | { | |
43 | struct task_struct *task; | |
44 | struct mm_struct *mm; | |
45 | ||
46 | task = damon_get_task_struct(t); | |
47 | if (!task) | |
48 | return NULL; | |
49 | ||
50 | mm = get_task_mm(task); | |
51 | put_task_struct(task); | |
52 | return mm; | |
53 | } | |
54 | ||
55 | /* | |
56 | * Functions for the initial monitoring target regions construction | |
57 | */ | |
58 | ||
59 | /* | |
60 | * Size-evenly split a region into 'nr_pieces' small regions | |
61 | * | |
62 | * Returns 0 on success, or negative error code otherwise. | |
63 | */ | |
64 | static int damon_va_evenly_split_region(struct damon_target *t, | |
65 | struct damon_region *r, unsigned int nr_pieces) | |
66 | { | |
67 | unsigned long sz_orig, sz_piece, orig_end; | |
68 | struct damon_region *n = NULL, *next; | |
69 | unsigned long start; | |
70 | ||
71 | if (!r || !nr_pieces) | |
72 | return -EINVAL; | |
73 | ||
74 | orig_end = r->ar.end; | |
ab63f63f | 75 | sz_orig = damon_sz_region(r); |
3f49584b SP |
76 | sz_piece = ALIGN_DOWN(sz_orig / nr_pieces, DAMON_MIN_REGION); |
77 | ||
78 | if (!sz_piece) | |
79 | return -EINVAL; | |
80 | ||
81 | r->ar.end = r->ar.start + sz_piece; | |
82 | next = damon_next_region(r); | |
83 | for (start = r->ar.end; start + sz_piece <= orig_end; | |
84 | start += sz_piece) { | |
85 | n = damon_new_region(start, start + sz_piece); | |
86 | if (!n) | |
87 | return -ENOMEM; | |
88 | damon_insert_region(n, r, next, t); | |
89 | r = n; | |
90 | } | |
91 | /* complement last region for possible rounding error */ | |
92 | if (n) | |
93 | n->ar.end = orig_end; | |
94 | ||
95 | return 0; | |
96 | } | |
97 | ||
98 | static unsigned long sz_range(struct damon_addr_range *r) | |
99 | { | |
100 | return r->end - r->start; | |
101 | } | |
102 | ||
3f49584b SP |
103 | /* |
104 | * Find three regions separated by two biggest unmapped regions | |
105 | * | |
106 | * vma the head vma of the target address space | |
107 | * regions an array of three address ranges that results will be saved | |
108 | * | |
109 | * This function receives an address space and finds three regions in it which | |
110 | * separated by the two biggest unmapped regions in the space. Please refer to | |
111 | * below comments of '__damon_va_init_regions()' function to know why this is | |
112 | * necessary. | |
113 | * | |
114 | * Returns 0 if success, or negative error code otherwise. | |
115 | */ | |
d0cf3dd4 | 116 | static int __damon_va_three_regions(struct mm_struct *mm, |
3f49584b SP |
117 | struct damon_addr_range regions[3]) |
118 | { | |
d0cf3dd4 LH |
119 | struct damon_addr_range first_gap = {0}, second_gap = {0}; |
120 | VMA_ITERATOR(vmi, mm, 0); | |
121 | struct vm_area_struct *vma, *prev = NULL; | |
122 | unsigned long start; | |
3f49584b | 123 | |
d0cf3dd4 LH |
124 | /* |
125 | * Find the two biggest gaps so that first_gap > second_gap > others. | |
126 | * If this is too slow, it can be optimised to examine the maple | |
127 | * tree gaps. | |
128 | */ | |
129 | for_each_vma(vmi, vma) { | |
130 | unsigned long gap; | |
131 | ||
132 | if (!prev) { | |
133 | start = vma->vm_start; | |
3f49584b SP |
134 | goto next; |
135 | } | |
d0cf3dd4 LH |
136 | gap = vma->vm_start - prev->vm_end; |
137 | ||
138 | if (gap > sz_range(&first_gap)) { | |
139 | second_gap = first_gap; | |
140 | first_gap.start = prev->vm_end; | |
141 | first_gap.end = vma->vm_start; | |
142 | } else if (gap > sz_range(&second_gap)) { | |
143 | second_gap.start = prev->vm_end; | |
144 | second_gap.end = vma->vm_start; | |
3f49584b SP |
145 | } |
146 | next: | |
d0cf3dd4 | 147 | prev = vma; |
3f49584b SP |
148 | } |
149 | ||
150 | if (!sz_range(&second_gap) || !sz_range(&first_gap)) | |
151 | return -EINVAL; | |
152 | ||
153 | /* Sort the two biggest gaps by address */ | |
154 | if (first_gap.start > second_gap.start) | |
8bd0b9da | 155 | swap(first_gap, second_gap); |
3f49584b SP |
156 | |
157 | /* Store the result */ | |
158 | regions[0].start = ALIGN(start, DAMON_MIN_REGION); | |
159 | regions[0].end = ALIGN(first_gap.start, DAMON_MIN_REGION); | |
160 | regions[1].start = ALIGN(first_gap.end, DAMON_MIN_REGION); | |
161 | regions[1].end = ALIGN(second_gap.start, DAMON_MIN_REGION); | |
162 | regions[2].start = ALIGN(second_gap.end, DAMON_MIN_REGION); | |
d0cf3dd4 | 163 | regions[2].end = ALIGN(prev->vm_end, DAMON_MIN_REGION); |
3f49584b SP |
164 | |
165 | return 0; | |
166 | } | |
167 | ||
168 | /* | |
169 | * Get the three regions in the given target (task) | |
170 | * | |
171 | * Returns 0 on success, negative error code otherwise. | |
172 | */ | |
173 | static int damon_va_three_regions(struct damon_target *t, | |
174 | struct damon_addr_range regions[3]) | |
175 | { | |
176 | struct mm_struct *mm; | |
177 | int rc; | |
178 | ||
179 | mm = damon_get_mm(t); | |
180 | if (!mm) | |
181 | return -EINVAL; | |
182 | ||
183 | mmap_read_lock(mm); | |
d0cf3dd4 | 184 | rc = __damon_va_three_regions(mm, regions); |
3f49584b SP |
185 | mmap_read_unlock(mm); |
186 | ||
187 | mmput(mm); | |
188 | return rc; | |
189 | } | |
190 | ||
191 | /* | |
192 | * Initialize the monitoring target regions for the given target (task) | |
193 | * | |
194 | * t the given target | |
195 | * | |
196 | * Because only a number of small portions of the entire address space | |
197 | * is actually mapped to the memory and accessed, monitoring the unmapped | |
198 | * regions is wasteful. That said, because we can deal with small noises, | |
199 | * tracking every mapping is not strictly required but could even incur a high | |
200 | * overhead if the mapping frequently changes or the number of mappings is | |
201 | * high. The adaptive regions adjustment mechanism will further help to deal | |
202 | * with the noise by simply identifying the unmapped areas as a region that | |
203 | * has no access. Moreover, applying the real mappings that would have many | |
204 | * unmapped areas inside will make the adaptive mechanism quite complex. That | |
205 | * said, too huge unmapped areas inside the monitoring target should be removed | |
206 | * to not take the time for the adaptive mechanism. | |
207 | * | |
208 | * For the reason, we convert the complex mappings to three distinct regions | |
209 | * that cover every mapped area of the address space. Also the two gaps | |
210 | * between the three regions are the two biggest unmapped areas in the given | |
211 | * address space. In detail, this function first identifies the start and the | |
212 | * end of the mappings and the two biggest unmapped areas of the address space. | |
213 | * Then, it constructs the three regions as below: | |
214 | * | |
215 | * [mappings[0]->start, big_two_unmapped_areas[0]->start) | |
216 | * [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start) | |
217 | * [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end) | |
218 | * | |
219 | * As usual memory map of processes is as below, the gap between the heap and | |
220 | * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed | |
221 | * region and the stack will be two biggest unmapped regions. Because these | |
222 | * gaps are exceptionally huge areas in usual address space, excluding these | |
223 | * two biggest unmapped regions will be sufficient to make a trade-off. | |
224 | * | |
225 | * <heap> | |
226 | * <BIG UNMAPPED REGION 1> | |
227 | * <uppermost mmap()-ed region> | |
228 | * (other mmap()-ed regions and small unmapped regions) | |
229 | * <lowermost mmap()-ed region> | |
230 | * <BIG UNMAPPED REGION 2> | |
231 | * <stack> | |
232 | */ | |
233 | static void __damon_va_init_regions(struct damon_ctx *ctx, | |
234 | struct damon_target *t) | |
235 | { | |
962fe7a6 | 236 | struct damon_target *ti; |
3f49584b SP |
237 | struct damon_region *r; |
238 | struct damon_addr_range regions[3]; | |
239 | unsigned long sz = 0, nr_pieces; | |
962fe7a6 | 240 | int i, tidx = 0; |
3f49584b SP |
241 | |
242 | if (damon_va_three_regions(t, regions)) { | |
962fe7a6 SP |
243 | damon_for_each_target(ti, ctx) { |
244 | if (ti == t) | |
245 | break; | |
246 | tidx++; | |
247 | } | |
248 | pr_debug("Failed to get three regions of %dth target\n", tidx); | |
3f49584b SP |
249 | return; |
250 | } | |
251 | ||
252 | for (i = 0; i < 3; i++) | |
253 | sz += regions[i].end - regions[i].start; | |
cbeaa77b SP |
254 | if (ctx->attrs.min_nr_regions) |
255 | sz /= ctx->attrs.min_nr_regions; | |
3f49584b SP |
256 | if (sz < DAMON_MIN_REGION) |
257 | sz = DAMON_MIN_REGION; | |
258 | ||
259 | /* Set the initial three regions of the target */ | |
260 | for (i = 0; i < 3; i++) { | |
261 | r = damon_new_region(regions[i].start, regions[i].end); | |
262 | if (!r) { | |
263 | pr_err("%d'th init region creation failed\n", i); | |
264 | return; | |
265 | } | |
266 | damon_add_region(r, t); | |
267 | ||
268 | nr_pieces = (regions[i].end - regions[i].start) / sz; | |
269 | damon_va_evenly_split_region(t, r, nr_pieces); | |
270 | } | |
271 | } | |
272 | ||
273 | /* Initialize '->regions_list' of every target (task) */ | |
cdeed009 | 274 | static void damon_va_init(struct damon_ctx *ctx) |
3f49584b SP |
275 | { |
276 | struct damon_target *t; | |
277 | ||
278 | damon_for_each_target(t, ctx) { | |
279 | /* the user may set the target regions as they want */ | |
280 | if (!damon_nr_regions(t)) | |
281 | __damon_va_init_regions(ctx, t); | |
282 | } | |
283 | } | |
284 | ||
3f49584b SP |
285 | /* |
286 | * Update regions for current memory mappings | |
287 | */ | |
cdeed009 | 288 | static void damon_va_update(struct damon_ctx *ctx) |
3f49584b SP |
289 | { |
290 | struct damon_addr_range three_regions[3]; | |
291 | struct damon_target *t; | |
292 | ||
293 | damon_for_each_target(t, ctx) { | |
294 | if (damon_va_three_regions(t, three_regions)) | |
295 | continue; | |
dae0087a | 296 | damon_set_regions(t, three_regions, 3); |
3f49584b SP |
297 | } |
298 | } | |
299 | ||
3f49584b SP |
300 | static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr, |
301 | unsigned long next, struct mm_walk *walk) | |
302 | { | |
303 | pte_t *pte; | |
e7ee3f97 | 304 | pmd_t pmde; |
3f49584b SP |
305 | spinlock_t *ptl; |
306 | ||
e7ee3f97 | 307 | if (pmd_trans_huge(pmdp_get(pmd))) { |
3f49584b | 308 | ptl = pmd_lock(walk->mm, pmd); |
e7ee3f97 LY |
309 | pmde = pmdp_get(pmd); |
310 | ||
311 | if (!pmd_present(pmde)) { | |
c8b9aff4 BW |
312 | spin_unlock(ptl); |
313 | return 0; | |
314 | } | |
315 | ||
e7ee3f97 | 316 | if (pmd_trans_huge(pmde)) { |
c11d34fa | 317 | damon_pmdp_mkold(pmd, walk->vma, addr); |
3f49584b SP |
318 | spin_unlock(ptl); |
319 | return 0; | |
320 | } | |
321 | spin_unlock(ptl); | |
322 | } | |
323 | ||
3f49584b | 324 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
7780d040 HD |
325 | if (!pte) { |
326 | walk->action = ACTION_AGAIN; | |
327 | return 0; | |
328 | } | |
c33c7948 | 329 | if (!pte_present(ptep_get(pte))) |
3f49584b | 330 | goto out; |
c11d34fa | 331 | damon_ptep_mkold(pte, walk->vma, addr); |
3f49584b SP |
332 | out: |
333 | pte_unmap_unlock(pte, ptl); | |
334 | return 0; | |
335 | } | |
336 | ||
49f4203a BW |
337 | #ifdef CONFIG_HUGETLB_PAGE |
338 | static void damon_hugetlb_mkold(pte_t *pte, struct mm_struct *mm, | |
339 | struct vm_area_struct *vma, unsigned long addr) | |
340 | { | |
341 | bool referenced = false; | |
342 | pte_t entry = huge_ptep_get(pte); | |
6b7cea90 | 343 | struct folio *folio = pfn_folio(pte_pfn(entry)); |
935d4f0c | 344 | unsigned long psize = huge_page_size(hstate_vma(vma)); |
49f4203a | 345 | |
6b7cea90 | 346 | folio_get(folio); |
49f4203a BW |
347 | |
348 | if (pte_young(entry)) { | |
349 | referenced = true; | |
350 | entry = pte_mkold(entry); | |
935d4f0c | 351 | set_huge_pte_at(mm, addr, pte, entry, psize); |
49f4203a BW |
352 | } |
353 | ||
354 | #ifdef CONFIG_MMU_NOTIFIER | |
355 | if (mmu_notifier_clear_young(mm, addr, | |
356 | addr + huge_page_size(hstate_vma(vma)))) | |
357 | referenced = true; | |
358 | #endif /* CONFIG_MMU_NOTIFIER */ | |
359 | ||
360 | if (referenced) | |
6b7cea90 | 361 | folio_set_young(folio); |
49f4203a | 362 | |
6b7cea90 KW |
363 | folio_set_idle(folio); |
364 | folio_put(folio); | |
49f4203a BW |
365 | } |
366 | ||
367 | static int damon_mkold_hugetlb_entry(pte_t *pte, unsigned long hmask, | |
368 | unsigned long addr, unsigned long end, | |
369 | struct mm_walk *walk) | |
370 | { | |
371 | struct hstate *h = hstate_vma(walk->vma); | |
372 | spinlock_t *ptl; | |
373 | pte_t entry; | |
374 | ||
375 | ptl = huge_pte_lock(h, walk->mm, pte); | |
376 | entry = huge_ptep_get(pte); | |
377 | if (!pte_present(entry)) | |
378 | goto out; | |
379 | ||
380 | damon_hugetlb_mkold(pte, walk->mm, walk->vma, addr); | |
381 | ||
382 | out: | |
383 | spin_unlock(ptl); | |
384 | return 0; | |
385 | } | |
386 | #else | |
387 | #define damon_mkold_hugetlb_entry NULL | |
388 | #endif /* CONFIG_HUGETLB_PAGE */ | |
389 | ||
199b50f4 | 390 | static const struct mm_walk_ops damon_mkold_ops = { |
3f49584b | 391 | .pmd_entry = damon_mkold_pmd_entry, |
49f4203a | 392 | .hugetlb_entry = damon_mkold_hugetlb_entry, |
49b06385 | 393 | .walk_lock = PGWALK_RDLOCK, |
3f49584b SP |
394 | }; |
395 | ||
396 | static void damon_va_mkold(struct mm_struct *mm, unsigned long addr) | |
397 | { | |
398 | mmap_read_lock(mm); | |
399 | walk_page_range(mm, addr, addr + 1, &damon_mkold_ops, NULL); | |
400 | mmap_read_unlock(mm); | |
401 | } | |
402 | ||
403 | /* | |
404 | * Functions for the access checking of the regions | |
405 | */ | |
406 | ||
8ef4d5ca KX |
407 | static void __damon_va_prepare_access_check(struct mm_struct *mm, |
408 | struct damon_region *r) | |
3f49584b SP |
409 | { |
410 | r->sampling_addr = damon_rand(r->ar.start, r->ar.end); | |
411 | ||
412 | damon_va_mkold(mm, r->sampling_addr); | |
413 | } | |
414 | ||
cdeed009 | 415 | static void damon_va_prepare_access_checks(struct damon_ctx *ctx) |
3f49584b SP |
416 | { |
417 | struct damon_target *t; | |
418 | struct mm_struct *mm; | |
419 | struct damon_region *r; | |
420 | ||
421 | damon_for_each_target(t, ctx) { | |
422 | mm = damon_get_mm(t); | |
423 | if (!mm) | |
424 | continue; | |
425 | damon_for_each_region(r, t) | |
8ef4d5ca | 426 | __damon_va_prepare_access_check(mm, r); |
3f49584b SP |
427 | mmput(mm); |
428 | } | |
429 | } | |
430 | ||
431 | struct damon_young_walk_private { | |
fc8c7d23 SP |
432 | /* size of the folio for the access checked virtual memory address */ |
433 | unsigned long *folio_sz; | |
3f49584b SP |
434 | bool young; |
435 | }; | |
436 | ||
437 | static int damon_young_pmd_entry(pmd_t *pmd, unsigned long addr, | |
438 | unsigned long next, struct mm_walk *walk) | |
439 | { | |
440 | pte_t *pte; | |
c33c7948 | 441 | pte_t ptent; |
3f49584b | 442 | spinlock_t *ptl; |
dc1b7866 | 443 | struct folio *folio; |
3f49584b SP |
444 | struct damon_young_walk_private *priv = walk->private; |
445 | ||
446 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
e7ee3f97 LY |
447 | if (pmd_trans_huge(pmdp_get(pmd))) { |
448 | pmd_t pmde; | |
449 | ||
3f49584b | 450 | ptl = pmd_lock(walk->mm, pmd); |
e7ee3f97 LY |
451 | pmde = pmdp_get(pmd); |
452 | ||
453 | if (!pmd_present(pmde)) { | |
c8b9aff4 BW |
454 | spin_unlock(ptl); |
455 | return 0; | |
456 | } | |
457 | ||
e7ee3f97 | 458 | if (!pmd_trans_huge(pmde)) { |
3f49584b SP |
459 | spin_unlock(ptl); |
460 | goto regular_page; | |
461 | } | |
e7ee3f97 | 462 | folio = damon_get_folio(pmd_pfn(pmde)); |
dc1b7866 | 463 | if (!folio) |
3f49584b | 464 | goto huge_out; |
e7ee3f97 | 465 | if (pmd_young(pmde) || !folio_test_idle(folio) || |
3f49584b | 466 | mmu_notifier_test_young(walk->mm, |
7477d756 | 467 | addr)) |
3f49584b | 468 | priv->young = true; |
7477d756 | 469 | *priv->folio_sz = HPAGE_PMD_SIZE; |
dc1b7866 | 470 | folio_put(folio); |
3f49584b SP |
471 | huge_out: |
472 | spin_unlock(ptl); | |
473 | return 0; | |
474 | } | |
475 | ||
476 | regular_page: | |
477 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
478 | ||
3f49584b | 479 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
7780d040 HD |
480 | if (!pte) { |
481 | walk->action = ACTION_AGAIN; | |
482 | return 0; | |
483 | } | |
c33c7948 RR |
484 | ptent = ptep_get(pte); |
485 | if (!pte_present(ptent)) | |
3f49584b | 486 | goto out; |
c33c7948 | 487 | folio = damon_get_folio(pte_pfn(ptent)); |
dc1b7866 | 488 | if (!folio) |
3f49584b | 489 | goto out; |
c33c7948 | 490 | if (pte_young(ptent) || !folio_test_idle(folio) || |
7477d756 | 491 | mmu_notifier_test_young(walk->mm, addr)) |
3f49584b | 492 | priv->young = true; |
7477d756 | 493 | *priv->folio_sz = folio_size(folio); |
dc1b7866 | 494 | folio_put(folio); |
3f49584b SP |
495 | out: |
496 | pte_unmap_unlock(pte, ptl); | |
497 | return 0; | |
498 | } | |
499 | ||
49f4203a BW |
500 | #ifdef CONFIG_HUGETLB_PAGE |
501 | static int damon_young_hugetlb_entry(pte_t *pte, unsigned long hmask, | |
502 | unsigned long addr, unsigned long end, | |
503 | struct mm_walk *walk) | |
504 | { | |
505 | struct damon_young_walk_private *priv = walk->private; | |
506 | struct hstate *h = hstate_vma(walk->vma); | |
6b7cea90 | 507 | struct folio *folio; |
49f4203a BW |
508 | spinlock_t *ptl; |
509 | pte_t entry; | |
510 | ||
511 | ptl = huge_pte_lock(h, walk->mm, pte); | |
512 | entry = huge_ptep_get(pte); | |
513 | if (!pte_present(entry)) | |
514 | goto out; | |
515 | ||
6b7cea90 KW |
516 | folio = pfn_folio(pte_pfn(entry)); |
517 | folio_get(folio); | |
49f4203a | 518 | |
6b7cea90 | 519 | if (pte_young(entry) || !folio_test_idle(folio) || |
7477d756 | 520 | mmu_notifier_test_young(walk->mm, addr)) |
49f4203a | 521 | priv->young = true; |
7477d756 | 522 | *priv->folio_sz = huge_page_size(h); |
49f4203a | 523 | |
6b7cea90 | 524 | folio_put(folio); |
49f4203a BW |
525 | |
526 | out: | |
527 | spin_unlock(ptl); | |
528 | return 0; | |
529 | } | |
530 | #else | |
531 | #define damon_young_hugetlb_entry NULL | |
532 | #endif /* CONFIG_HUGETLB_PAGE */ | |
533 | ||
199b50f4 | 534 | static const struct mm_walk_ops damon_young_ops = { |
3f49584b | 535 | .pmd_entry = damon_young_pmd_entry, |
49f4203a | 536 | .hugetlb_entry = damon_young_hugetlb_entry, |
49b06385 | 537 | .walk_lock = PGWALK_RDLOCK, |
3f49584b SP |
538 | }; |
539 | ||
540 | static bool damon_va_young(struct mm_struct *mm, unsigned long addr, | |
fc8c7d23 | 541 | unsigned long *folio_sz) |
3f49584b SP |
542 | { |
543 | struct damon_young_walk_private arg = { | |
fc8c7d23 | 544 | .folio_sz = folio_sz, |
3f49584b SP |
545 | .young = false, |
546 | }; | |
547 | ||
548 | mmap_read_lock(mm); | |
549 | walk_page_range(mm, addr, addr + 1, &damon_young_ops, &arg); | |
550 | mmap_read_unlock(mm); | |
551 | return arg.young; | |
552 | } | |
553 | ||
554 | /* | |
555 | * Check whether the region was accessed after the last preparation | |
556 | * | |
557 | * mm 'mm_struct' for the given virtual address space | |
558 | * r the region to be checked | |
559 | */ | |
09876ae7 | 560 | static void __damon_va_check_access(struct mm_struct *mm, |
ace30fb2 SP |
561 | struct damon_region *r, bool same_target, |
562 | struct damon_attrs *attrs) | |
3f49584b | 563 | { |
3f49584b | 564 | static unsigned long last_addr; |
fc8c7d23 | 565 | static unsigned long last_folio_sz = PAGE_SIZE; |
3f49584b SP |
566 | static bool last_accessed; |
567 | ||
22a77880 | 568 | if (!mm) { |
ace30fb2 | 569 | damon_update_region_access_rate(r, false, attrs); |
22a77880 SP |
570 | return; |
571 | } | |
572 | ||
3f49584b | 573 | /* If the region is in the last checked page, reuse the result */ |
fc8c7d23 SP |
574 | if (same_target && (ALIGN_DOWN(last_addr, last_folio_sz) == |
575 | ALIGN_DOWN(r->sampling_addr, last_folio_sz))) { | |
ace30fb2 | 576 | damon_update_region_access_rate(r, last_accessed, attrs); |
3f49584b SP |
577 | return; |
578 | } | |
579 | ||
fc8c7d23 | 580 | last_accessed = damon_va_young(mm, r->sampling_addr, &last_folio_sz); |
ace30fb2 | 581 | damon_update_region_access_rate(r, last_accessed, attrs); |
3f49584b | 582 | |
3f49584b SP |
583 | last_addr = r->sampling_addr; |
584 | } | |
585 | ||
cdeed009 | 586 | static unsigned int damon_va_check_accesses(struct damon_ctx *ctx) |
3f49584b SP |
587 | { |
588 | struct damon_target *t; | |
589 | struct mm_struct *mm; | |
590 | struct damon_region *r; | |
591 | unsigned int max_nr_accesses = 0; | |
95cd2522 | 592 | bool same_target; |
3f49584b SP |
593 | |
594 | damon_for_each_target(t, ctx) { | |
595 | mm = damon_get_mm(t); | |
95cd2522 | 596 | same_target = false; |
3f49584b | 597 | damon_for_each_region(r, t) { |
ace30fb2 SP |
598 | __damon_va_check_access(mm, r, same_target, |
599 | &ctx->attrs); | |
3f49584b | 600 | max_nr_accesses = max(r->nr_accesses, max_nr_accesses); |
95cd2522 | 601 | same_target = true; |
3f49584b | 602 | } |
22a77880 SP |
603 | if (mm) |
604 | mmput(mm); | |
3f49584b SP |
605 | } |
606 | ||
607 | return max_nr_accesses; | |
608 | } | |
609 | ||
610 | /* | |
611 | * Functions for the target validity check and cleanup | |
612 | */ | |
613 | ||
16bc1b0f | 614 | static bool damon_va_target_valid(struct damon_target *t) |
3f49584b | 615 | { |
3f49584b SP |
616 | struct task_struct *task; |
617 | ||
618 | task = damon_get_task_struct(t); | |
619 | if (task) { | |
620 | put_task_struct(task); | |
621 | return true; | |
622 | } | |
623 | ||
624 | return false; | |
625 | } | |
626 | ||
6dea8add | 627 | #ifndef CONFIG_ADVISE_SYSCALLS |
0e92c2ee SP |
628 | static unsigned long damos_madvise(struct damon_target *target, |
629 | struct damon_region *r, int behavior) | |
6dea8add | 630 | { |
0e92c2ee | 631 | return 0; |
6dea8add SP |
632 | } |
633 | #else | |
0e92c2ee SP |
634 | static unsigned long damos_madvise(struct damon_target *target, |
635 | struct damon_region *r, int behavior) | |
6dea8add SP |
636 | { |
637 | struct mm_struct *mm; | |
0e92c2ee | 638 | unsigned long start = PAGE_ALIGN(r->ar.start); |
ab63f63f | 639 | unsigned long len = PAGE_ALIGN(damon_sz_region(r)); |
0e92c2ee | 640 | unsigned long applied; |
6dea8add SP |
641 | |
642 | mm = damon_get_mm(target); | |
643 | if (!mm) | |
0e92c2ee | 644 | return 0; |
6dea8add | 645 | |
0e92c2ee | 646 | applied = do_madvise(mm, start, len, behavior) ? 0 : len; |
6dea8add | 647 | mmput(mm); |
0e92c2ee SP |
648 | |
649 | return applied; | |
6dea8add SP |
650 | } |
651 | #endif /* CONFIG_ADVISE_SYSCALLS */ | |
652 | ||
0e92c2ee SP |
653 | static unsigned long damon_va_apply_scheme(struct damon_ctx *ctx, |
654 | struct damon_target *t, struct damon_region *r, | |
655 | struct damos *scheme) | |
6dea8add SP |
656 | { |
657 | int madv_action; | |
658 | ||
659 | switch (scheme->action) { | |
660 | case DAMOS_WILLNEED: | |
661 | madv_action = MADV_WILLNEED; | |
662 | break; | |
663 | case DAMOS_COLD: | |
664 | madv_action = MADV_COLD; | |
665 | break; | |
666 | case DAMOS_PAGEOUT: | |
667 | madv_action = MADV_PAGEOUT; | |
668 | break; | |
669 | case DAMOS_HUGEPAGE: | |
670 | madv_action = MADV_HUGEPAGE; | |
671 | break; | |
672 | case DAMOS_NOHUGEPAGE: | |
673 | madv_action = MADV_NOHUGEPAGE; | |
674 | break; | |
2f0b548c SP |
675 | case DAMOS_STAT: |
676 | return 0; | |
6dea8add | 677 | default: |
5934ec13 KX |
678 | /* |
679 | * DAMOS actions that are not yet supported by 'vaddr'. | |
680 | */ | |
0e92c2ee | 681 | return 0; |
6dea8add SP |
682 | } |
683 | ||
684 | return damos_madvise(t, r, madv_action); | |
685 | } | |
686 | ||
cdeed009 XH |
687 | static int damon_va_scheme_score(struct damon_ctx *context, |
688 | struct damon_target *t, struct damon_region *r, | |
689 | struct damos *scheme) | |
198f0f4c SP |
690 | { |
691 | ||
692 | switch (scheme->action) { | |
693 | case DAMOS_PAGEOUT: | |
e3e486e6 | 694 | return damon_cold_score(context, r, scheme); |
198f0f4c SP |
695 | default: |
696 | break; | |
697 | } | |
698 | ||
699 | return DAMOS_MAX_SCORE; | |
700 | } | |
701 | ||
7752925f SP |
702 | static int __init damon_va_initcall(void) |
703 | { | |
704 | struct damon_operations ops = { | |
705 | .id = DAMON_OPS_VADDR, | |
706 | .init = damon_va_init, | |
707 | .update = damon_va_update, | |
708 | .prepare_access_checks = damon_va_prepare_access_checks, | |
709 | .check_accesses = damon_va_check_accesses, | |
710 | .reset_aggregated = NULL, | |
711 | .target_valid = damon_va_target_valid, | |
712 | .cleanup = NULL, | |
713 | .apply_scheme = damon_va_apply_scheme, | |
714 | .get_scheme_score = damon_va_scheme_score, | |
715 | }; | |
de6d0154 SP |
716 | /* ops for fixed virtual address ranges */ |
717 | struct damon_operations ops_fvaddr = ops; | |
718 | int err; | |
719 | ||
720 | /* Don't set the monitoring target regions for the entire mapping */ | |
721 | ops_fvaddr.id = DAMON_OPS_FVADDR; | |
722 | ops_fvaddr.init = NULL; | |
723 | ops_fvaddr.update = NULL; | |
724 | ||
725 | err = damon_register_ops(&ops); | |
726 | if (err) | |
727 | return err; | |
728 | return damon_register_ops(&ops_fvaddr); | |
7752925f SP |
729 | }; |
730 | ||
731 | subsys_initcall(damon_va_initcall); | |
732 | ||
17ccae8b | 733 | #include "vaddr-test.h" |