xref: /linux/mm/damon/vaddr.c (revision 97ef3b7f4fdf8ad6818aa2c8201c3b72cc635e16)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * DAMON Primitives for Virtual Address Spaces
4  *
5  * Author: SeongJae Park <sjpark@amazon.de>
6  */
7 
8 #define pr_fmt(fmt) "damon-va: " fmt
9 
10 #include <linux/damon.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mm.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/highmem.h>
15 #include <linux/page_idle.h>
16 #include <linux/pagewalk.h>
17 #include <linux/random.h>
18 #include <linux/sched/mm.h>
19 #include <linux/slab.h>
20 
21 #ifdef CONFIG_DAMON_VADDR_KUNIT_TEST
22 #undef DAMON_MIN_REGION
23 #define DAMON_MIN_REGION 1
24 #endif
25 
26 /* Get a random number in [l, r) */
27 #define damon_rand(l, r) (l + prandom_u32_max(r - l))
28 
29 /*
30  * 't->id' should be the pointer to the relevant 'struct pid' having reference
31  * count.  Caller must put the returned task, unless it is NULL.
32  */
33 #define damon_get_task_struct(t) \
34 	(get_pid_task((struct pid *)t->id, PIDTYPE_PID))
35 
36 /*
37  * Get the mm_struct of the given target
38  *
39  * Caller _must_ put the mm_struct after use, unless it is NULL.
40  *
41  * Returns the mm_struct of the target on success, NULL on failure
42  */
43 static struct mm_struct *damon_get_mm(struct damon_target *t)
44 {
45 	struct task_struct *task;
46 	struct mm_struct *mm;
47 
48 	task = damon_get_task_struct(t);
49 	if (!task)
50 		return NULL;
51 
52 	mm = get_task_mm(task);
53 	put_task_struct(task);
54 	return mm;
55 }
56 
57 /*
58  * Functions for the initial monitoring target regions construction
59  */
60 
61 /*
62  * Size-evenly split a region into 'nr_pieces' small regions
63  *
64  * Returns 0 on success, or negative error code otherwise.
65  */
66 static int damon_va_evenly_split_region(struct damon_target *t,
67 		struct damon_region *r, unsigned int nr_pieces)
68 {
69 	unsigned long sz_orig, sz_piece, orig_end;
70 	struct damon_region *n = NULL, *next;
71 	unsigned long start;
72 
73 	if (!r || !nr_pieces)
74 		return -EINVAL;
75 
76 	orig_end = r->ar.end;
77 	sz_orig = r->ar.end - r->ar.start;
78 	sz_piece = ALIGN_DOWN(sz_orig / nr_pieces, DAMON_MIN_REGION);
79 
80 	if (!sz_piece)
81 		return -EINVAL;
82 
83 	r->ar.end = r->ar.start + sz_piece;
84 	next = damon_next_region(r);
85 	for (start = r->ar.end; start + sz_piece <= orig_end;
86 			start += sz_piece) {
87 		n = damon_new_region(start, start + sz_piece);
88 		if (!n)
89 			return -ENOMEM;
90 		damon_insert_region(n, r, next, t);
91 		r = n;
92 	}
93 	/* complement last region for possible rounding error */
94 	if (n)
95 		n->ar.end = orig_end;
96 
97 	return 0;
98 }
99 
100 static unsigned long sz_range(struct damon_addr_range *r)
101 {
102 	return r->end - r->start;
103 }
104 
105 static void swap_ranges(struct damon_addr_range *r1,
106 			struct damon_addr_range *r2)
107 {
108 	struct damon_addr_range tmp;
109 
110 	tmp = *r1;
111 	*r1 = *r2;
112 	*r2 = tmp;
113 }
114 
115 /*
116  * Find three regions separated by two biggest unmapped regions
117  *
118  * vma		the head vma of the target address space
119  * regions	an array of three address ranges that results will be saved
120  *
121  * This function receives an address space and finds three regions in it which
122  * separated by the two biggest unmapped regions in the space.  Please refer to
123  * below comments of '__damon_va_init_regions()' function to know why this is
124  * necessary.
125  *
126  * Returns 0 if success, or negative error code otherwise.
127  */
128 static int __damon_va_three_regions(struct vm_area_struct *vma,
129 				       struct damon_addr_range regions[3])
130 {
131 	struct damon_addr_range gap = {0}, first_gap = {0}, second_gap = {0};
132 	struct vm_area_struct *last_vma = NULL;
133 	unsigned long start = 0;
134 	struct rb_root rbroot;
135 
136 	/* Find two biggest gaps so that first_gap > second_gap > others */
137 	for (; vma; vma = vma->vm_next) {
138 		if (!last_vma) {
139 			start = vma->vm_start;
140 			goto next;
141 		}
142 
143 		if (vma->rb_subtree_gap <= sz_range(&second_gap)) {
144 			rbroot.rb_node = &vma->vm_rb;
145 			vma = rb_entry(rb_last(&rbroot),
146 					struct vm_area_struct, vm_rb);
147 			goto next;
148 		}
149 
150 		gap.start = last_vma->vm_end;
151 		gap.end = vma->vm_start;
152 		if (sz_range(&gap) > sz_range(&second_gap)) {
153 			swap_ranges(&gap, &second_gap);
154 			if (sz_range(&second_gap) > sz_range(&first_gap))
155 				swap_ranges(&second_gap, &first_gap);
156 		}
157 next:
158 		last_vma = vma;
159 	}
160 
161 	if (!sz_range(&second_gap) || !sz_range(&first_gap))
162 		return -EINVAL;
163 
164 	/* Sort the two biggest gaps by address */
165 	if (first_gap.start > second_gap.start)
166 		swap_ranges(&first_gap, &second_gap);
167 
168 	/* Store the result */
169 	regions[0].start = ALIGN(start, DAMON_MIN_REGION);
170 	regions[0].end = ALIGN(first_gap.start, DAMON_MIN_REGION);
171 	regions[1].start = ALIGN(first_gap.end, DAMON_MIN_REGION);
172 	regions[1].end = ALIGN(second_gap.start, DAMON_MIN_REGION);
173 	regions[2].start = ALIGN(second_gap.end, DAMON_MIN_REGION);
174 	regions[2].end = ALIGN(last_vma->vm_end, DAMON_MIN_REGION);
175 
176 	return 0;
177 }
178 
179 /*
180  * Get the three regions in the given target (task)
181  *
182  * Returns 0 on success, negative error code otherwise.
183  */
184 static int damon_va_three_regions(struct damon_target *t,
185 				struct damon_addr_range regions[3])
186 {
187 	struct mm_struct *mm;
188 	int rc;
189 
190 	mm = damon_get_mm(t);
191 	if (!mm)
192 		return -EINVAL;
193 
194 	mmap_read_lock(mm);
195 	rc = __damon_va_three_regions(mm->mmap, regions);
196 	mmap_read_unlock(mm);
197 
198 	mmput(mm);
199 	return rc;
200 }
201 
202 /*
203  * Initialize the monitoring target regions for the given target (task)
204  *
205  * t	the given target
206  *
207  * Because only a number of small portions of the entire address space
208  * is actually mapped to the memory and accessed, monitoring the unmapped
209  * regions is wasteful.  That said, because we can deal with small noises,
210  * tracking every mapping is not strictly required but could even incur a high
211  * overhead if the mapping frequently changes or the number of mappings is
212  * high.  The adaptive regions adjustment mechanism will further help to deal
213  * with the noise by simply identifying the unmapped areas as a region that
214  * has no access.  Moreover, applying the real mappings that would have many
215  * unmapped areas inside will make the adaptive mechanism quite complex.  That
216  * said, too huge unmapped areas inside the monitoring target should be removed
217  * to not take the time for the adaptive mechanism.
218  *
219  * For the reason, we convert the complex mappings to three distinct regions
220  * that cover every mapped area of the address space.  Also the two gaps
221  * between the three regions are the two biggest unmapped areas in the given
222  * address space.  In detail, this function first identifies the start and the
223  * end of the mappings and the two biggest unmapped areas of the address space.
224  * Then, it constructs the three regions as below:
225  *
226  *     [mappings[0]->start, big_two_unmapped_areas[0]->start)
227  *     [big_two_unmapped_areas[0]->end, big_two_unmapped_areas[1]->start)
228  *     [big_two_unmapped_areas[1]->end, mappings[nr_mappings - 1]->end)
229  *
230  * As usual memory map of processes is as below, the gap between the heap and
231  * the uppermost mmap()-ed region, and the gap between the lowermost mmap()-ed
232  * region and the stack will be two biggest unmapped regions.  Because these
233  * gaps are exceptionally huge areas in usual address space, excluding these
234  * two biggest unmapped regions will be sufficient to make a trade-off.
235  *
236  *   <heap>
237  *   <BIG UNMAPPED REGION 1>
238  *   <uppermost mmap()-ed region>
239  *   (other mmap()-ed regions and small unmapped regions)
240  *   <lowermost mmap()-ed region>
241  *   <BIG UNMAPPED REGION 2>
242  *   <stack>
243  */
244 static void __damon_va_init_regions(struct damon_ctx *ctx,
245 				     struct damon_target *t)
246 {
247 	struct damon_region *r;
248 	struct damon_addr_range regions[3];
249 	unsigned long sz = 0, nr_pieces;
250 	int i;
251 
252 	if (damon_va_three_regions(t, regions)) {
253 		pr_err("Failed to get three regions of target %lu\n", t->id);
254 		return;
255 	}
256 
257 	for (i = 0; i < 3; i++)
258 		sz += regions[i].end - regions[i].start;
259 	if (ctx->min_nr_regions)
260 		sz /= ctx->min_nr_regions;
261 	if (sz < DAMON_MIN_REGION)
262 		sz = DAMON_MIN_REGION;
263 
264 	/* Set the initial three regions of the target */
265 	for (i = 0; i < 3; i++) {
266 		r = damon_new_region(regions[i].start, regions[i].end);
267 		if (!r) {
268 			pr_err("%d'th init region creation failed\n", i);
269 			return;
270 		}
271 		damon_add_region(r, t);
272 
273 		nr_pieces = (regions[i].end - regions[i].start) / sz;
274 		damon_va_evenly_split_region(t, r, nr_pieces);
275 	}
276 }
277 
278 /* Initialize '->regions_list' of every target (task) */
279 void damon_va_init(struct damon_ctx *ctx)
280 {
281 	struct damon_target *t;
282 
283 	damon_for_each_target(t, ctx) {
284 		/* the user may set the target regions as they want */
285 		if (!damon_nr_regions(t))
286 			__damon_va_init_regions(ctx, t);
287 	}
288 }
289 
290 /*
291  * Functions for the dynamic monitoring target regions update
292  */
293 
294 /*
295  * Check whether a region is intersecting an address range
296  *
297  * Returns true if it is.
298  */
299 static bool damon_intersect(struct damon_region *r, struct damon_addr_range *re)
300 {
301 	return !(r->ar.end <= re->start || re->end <= r->ar.start);
302 }
303 
304 /*
305  * Update damon regions for the three big regions of the given target
306  *
307  * t		the given target
308  * bregions	the three big regions of the target
309  */
310 static void damon_va_apply_three_regions(struct damon_target *t,
311 		struct damon_addr_range bregions[3])
312 {
313 	struct damon_region *r, *next;
314 	unsigned int i = 0;
315 
316 	/* Remove regions which are not in the three big regions now */
317 	damon_for_each_region_safe(r, next, t) {
318 		for (i = 0; i < 3; i++) {
319 			if (damon_intersect(r, &bregions[i]))
320 				break;
321 		}
322 		if (i == 3)
323 			damon_destroy_region(r, t);
324 	}
325 
326 	/* Adjust intersecting regions to fit with the three big regions */
327 	for (i = 0; i < 3; i++) {
328 		struct damon_region *first = NULL, *last;
329 		struct damon_region *newr;
330 		struct damon_addr_range *br;
331 
332 		br = &bregions[i];
333 		/* Get the first and last regions which intersects with br */
334 		damon_for_each_region(r, t) {
335 			if (damon_intersect(r, br)) {
336 				if (!first)
337 					first = r;
338 				last = r;
339 			}
340 			if (r->ar.start >= br->end)
341 				break;
342 		}
343 		if (!first) {
344 			/* no damon_region intersects with this big region */
345 			newr = damon_new_region(
346 					ALIGN_DOWN(br->start,
347 						DAMON_MIN_REGION),
348 					ALIGN(br->end, DAMON_MIN_REGION));
349 			if (!newr)
350 				continue;
351 			damon_insert_region(newr, damon_prev_region(r), r, t);
352 		} else {
353 			first->ar.start = ALIGN_DOWN(br->start,
354 					DAMON_MIN_REGION);
355 			last->ar.end = ALIGN(br->end, DAMON_MIN_REGION);
356 		}
357 	}
358 }
359 
360 /*
361  * Update regions for current memory mappings
362  */
363 void damon_va_update(struct damon_ctx *ctx)
364 {
365 	struct damon_addr_range three_regions[3];
366 	struct damon_target *t;
367 
368 	damon_for_each_target(t, ctx) {
369 		if (damon_va_three_regions(t, three_regions))
370 			continue;
371 		damon_va_apply_three_regions(t, three_regions);
372 	}
373 }
374 
375 /*
376  * Get an online page for a pfn if it's in the LRU list.  Otherwise, returns
377  * NULL.
378  *
379  * The body of this function is stolen from the 'page_idle_get_page()'.  We
380  * steal rather than reuse it because the code is quite simple.
381  */
382 static struct page *damon_get_page(unsigned long pfn)
383 {
384 	struct page *page = pfn_to_online_page(pfn);
385 
386 	if (!page || !PageLRU(page) || !get_page_unless_zero(page))
387 		return NULL;
388 
389 	if (unlikely(!PageLRU(page))) {
390 		put_page(page);
391 		page = NULL;
392 	}
393 	return page;
394 }
395 
396 static void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm,
397 			     unsigned long addr)
398 {
399 	bool referenced = false;
400 	struct page *page = damon_get_page(pte_pfn(*pte));
401 
402 	if (!page)
403 		return;
404 
405 	if (pte_young(*pte)) {
406 		referenced = true;
407 		*pte = pte_mkold(*pte);
408 	}
409 
410 #ifdef CONFIG_MMU_NOTIFIER
411 	if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
412 		referenced = true;
413 #endif /* CONFIG_MMU_NOTIFIER */
414 
415 	if (referenced)
416 		set_page_young(page);
417 
418 	set_page_idle(page);
419 	put_page(page);
420 }
421 
422 static void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm,
423 			     unsigned long addr)
424 {
425 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
426 	bool referenced = false;
427 	struct page *page = damon_get_page(pmd_pfn(*pmd));
428 
429 	if (!page)
430 		return;
431 
432 	if (pmd_young(*pmd)) {
433 		referenced = true;
434 		*pmd = pmd_mkold(*pmd);
435 	}
436 
437 #ifdef CONFIG_MMU_NOTIFIER
438 	if (mmu_notifier_clear_young(mm, addr,
439 				addr + ((1UL) << HPAGE_PMD_SHIFT)))
440 		referenced = true;
441 #endif /* CONFIG_MMU_NOTIFIER */
442 
443 	if (referenced)
444 		set_page_young(page);
445 
446 	set_page_idle(page);
447 	put_page(page);
448 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
449 }
450 
451 static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
452 		unsigned long next, struct mm_walk *walk)
453 {
454 	pte_t *pte;
455 	spinlock_t *ptl;
456 
457 	if (pmd_huge(*pmd)) {
458 		ptl = pmd_lock(walk->mm, pmd);
459 		if (pmd_huge(*pmd)) {
460 			damon_pmdp_mkold(pmd, walk->mm, addr);
461 			spin_unlock(ptl);
462 			return 0;
463 		}
464 		spin_unlock(ptl);
465 	}
466 
467 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
468 		return 0;
469 	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
470 	if (!pte_present(*pte))
471 		goto out;
472 	damon_ptep_mkold(pte, walk->mm, addr);
473 out:
474 	pte_unmap_unlock(pte, ptl);
475 	return 0;
476 }
477 
478 static struct mm_walk_ops damon_mkold_ops = {
479 	.pmd_entry = damon_mkold_pmd_entry,
480 };
481 
482 static void damon_va_mkold(struct mm_struct *mm, unsigned long addr)
483 {
484 	mmap_read_lock(mm);
485 	walk_page_range(mm, addr, addr + 1, &damon_mkold_ops, NULL);
486 	mmap_read_unlock(mm);
487 }
488 
489 /*
490  * Functions for the access checking of the regions
491  */
492 
493 static void damon_va_prepare_access_check(struct damon_ctx *ctx,
494 			struct mm_struct *mm, struct damon_region *r)
495 {
496 	r->sampling_addr = damon_rand(r->ar.start, r->ar.end);
497 
498 	damon_va_mkold(mm, r->sampling_addr);
499 }
500 
501 void damon_va_prepare_access_checks(struct damon_ctx *ctx)
502 {
503 	struct damon_target *t;
504 	struct mm_struct *mm;
505 	struct damon_region *r;
506 
507 	damon_for_each_target(t, ctx) {
508 		mm = damon_get_mm(t);
509 		if (!mm)
510 			continue;
511 		damon_for_each_region(r, t)
512 			damon_va_prepare_access_check(ctx, mm, r);
513 		mmput(mm);
514 	}
515 }
516 
517 struct damon_young_walk_private {
518 	unsigned long *page_sz;
519 	bool young;
520 };
521 
522 static int damon_young_pmd_entry(pmd_t *pmd, unsigned long addr,
523 		unsigned long next, struct mm_walk *walk)
524 {
525 	pte_t *pte;
526 	spinlock_t *ptl;
527 	struct page *page;
528 	struct damon_young_walk_private *priv = walk->private;
529 
530 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
531 	if (pmd_huge(*pmd)) {
532 		ptl = pmd_lock(walk->mm, pmd);
533 		if (!pmd_huge(*pmd)) {
534 			spin_unlock(ptl);
535 			goto regular_page;
536 		}
537 		page = damon_get_page(pmd_pfn(*pmd));
538 		if (!page)
539 			goto huge_out;
540 		if (pmd_young(*pmd) || !page_is_idle(page) ||
541 					mmu_notifier_test_young(walk->mm,
542 						addr)) {
543 			*priv->page_sz = ((1UL) << HPAGE_PMD_SHIFT);
544 			priv->young = true;
545 		}
546 		put_page(page);
547 huge_out:
548 		spin_unlock(ptl);
549 		return 0;
550 	}
551 
552 regular_page:
553 #endif	/* CONFIG_TRANSPARENT_HUGEPAGE */
554 
555 	if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
556 		return -EINVAL;
557 	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
558 	if (!pte_present(*pte))
559 		goto out;
560 	page = damon_get_page(pte_pfn(*pte));
561 	if (!page)
562 		goto out;
563 	if (pte_young(*pte) || !page_is_idle(page) ||
564 			mmu_notifier_test_young(walk->mm, addr)) {
565 		*priv->page_sz = PAGE_SIZE;
566 		priv->young = true;
567 	}
568 	put_page(page);
569 out:
570 	pte_unmap_unlock(pte, ptl);
571 	return 0;
572 }
573 
574 static struct mm_walk_ops damon_young_ops = {
575 	.pmd_entry = damon_young_pmd_entry,
576 };
577 
578 static bool damon_va_young(struct mm_struct *mm, unsigned long addr,
579 		unsigned long *page_sz)
580 {
581 	struct damon_young_walk_private arg = {
582 		.page_sz = page_sz,
583 		.young = false,
584 	};
585 
586 	mmap_read_lock(mm);
587 	walk_page_range(mm, addr, addr + 1, &damon_young_ops, &arg);
588 	mmap_read_unlock(mm);
589 	return arg.young;
590 }
591 
592 /*
593  * Check whether the region was accessed after the last preparation
594  *
595  * mm	'mm_struct' for the given virtual address space
596  * r	the region to be checked
597  */
598 static void damon_va_check_access(struct damon_ctx *ctx,
599 			       struct mm_struct *mm, struct damon_region *r)
600 {
601 	static struct mm_struct *last_mm;
602 	static unsigned long last_addr;
603 	static unsigned long last_page_sz = PAGE_SIZE;
604 	static bool last_accessed;
605 
606 	/* If the region is in the last checked page, reuse the result */
607 	if (mm == last_mm && (ALIGN_DOWN(last_addr, last_page_sz) ==
608 				ALIGN_DOWN(r->sampling_addr, last_page_sz))) {
609 		if (last_accessed)
610 			r->nr_accesses++;
611 		return;
612 	}
613 
614 	last_accessed = damon_va_young(mm, r->sampling_addr, &last_page_sz);
615 	if (last_accessed)
616 		r->nr_accesses++;
617 
618 	last_mm = mm;
619 	last_addr = r->sampling_addr;
620 }
621 
622 unsigned int damon_va_check_accesses(struct damon_ctx *ctx)
623 {
624 	struct damon_target *t;
625 	struct mm_struct *mm;
626 	struct damon_region *r;
627 	unsigned int max_nr_accesses = 0;
628 
629 	damon_for_each_target(t, ctx) {
630 		mm = damon_get_mm(t);
631 		if (!mm)
632 			continue;
633 		damon_for_each_region(r, t) {
634 			damon_va_check_access(ctx, mm, r);
635 			max_nr_accesses = max(r->nr_accesses, max_nr_accesses);
636 		}
637 		mmput(mm);
638 	}
639 
640 	return max_nr_accesses;
641 }
642 
643 /*
644  * Functions for the target validity check and cleanup
645  */
646 
647 bool damon_va_target_valid(void *target)
648 {
649 	struct damon_target *t = target;
650 	struct task_struct *task;
651 
652 	task = damon_get_task_struct(t);
653 	if (task) {
654 		put_task_struct(task);
655 		return true;
656 	}
657 
658 	return false;
659 }
660 
661 void damon_va_set_primitives(struct damon_ctx *ctx)
662 {
663 	ctx->primitive.init = damon_va_init;
664 	ctx->primitive.update = damon_va_update;
665 	ctx->primitive.prepare_access_checks = damon_va_prepare_access_checks;
666 	ctx->primitive.check_accesses = damon_va_check_accesses;
667 	ctx->primitive.reset_aggregated = NULL;
668 	ctx->primitive.target_valid = damon_va_target_valid;
669 	ctx->primitive.cleanup = NULL;
670 }
671 
672 #include "vaddr-test.h"
673