xref: /openbmc/linux/mm/damon/paddr.c (revision e9adcfec)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * DAMON Primitives for The Physical Address Space
4  *
5  * Author: SeongJae Park <sj@kernel.org>
6  */
7 
8 #define pr_fmt(fmt) "damon-pa: " fmt
9 
10 #include <linux/mmu_notifier.h>
11 #include <linux/page_idle.h>
12 #include <linux/pagemap.h>
13 #include <linux/rmap.h>
14 #include <linux/swap.h>
15 
16 #include "../internal.h"
17 #include "ops-common.h"
18 
19 static bool __damon_pa_mkold(struct folio *folio, struct vm_area_struct *vma,
20 		unsigned long addr, void *arg)
21 {
22 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
23 
24 	while (page_vma_mapped_walk(&pvmw)) {
25 		addr = pvmw.address;
26 		if (pvmw.pte)
27 			damon_ptep_mkold(pvmw.pte, vma->vm_mm, addr);
28 		else
29 			damon_pmdp_mkold(pvmw.pmd, vma->vm_mm, addr);
30 	}
31 	return true;
32 }
33 
34 static void damon_pa_mkold(unsigned long paddr)
35 {
36 	struct folio *folio = damon_get_folio(PHYS_PFN(paddr));
37 	struct rmap_walk_control rwc = {
38 		.rmap_one = __damon_pa_mkold,
39 		.anon_lock = folio_lock_anon_vma_read,
40 	};
41 	bool need_lock;
42 
43 	if (!folio)
44 		return;
45 
46 	if (!folio_mapped(folio) || !folio_raw_mapping(folio)) {
47 		folio_set_idle(folio);
48 		goto out;
49 	}
50 
51 	need_lock = !folio_test_anon(folio) || folio_test_ksm(folio);
52 	if (need_lock && !folio_trylock(folio))
53 		goto out;
54 
55 	rmap_walk(folio, &rwc);
56 
57 	if (need_lock)
58 		folio_unlock(folio);
59 
60 out:
61 	folio_put(folio);
62 }
63 
64 static void __damon_pa_prepare_access_check(struct damon_region *r)
65 {
66 	r->sampling_addr = damon_rand(r->ar.start, r->ar.end);
67 
68 	damon_pa_mkold(r->sampling_addr);
69 }
70 
71 static void damon_pa_prepare_access_checks(struct damon_ctx *ctx)
72 {
73 	struct damon_target *t;
74 	struct damon_region *r;
75 
76 	damon_for_each_target(t, ctx) {
77 		damon_for_each_region(r, t)
78 			__damon_pa_prepare_access_check(r);
79 	}
80 }
81 
82 struct damon_pa_access_chk_result {
83 	unsigned long page_sz;
84 	bool accessed;
85 };
86 
87 static bool __damon_pa_young(struct folio *folio, struct vm_area_struct *vma,
88 		unsigned long addr, void *arg)
89 {
90 	struct damon_pa_access_chk_result *result = arg;
91 	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, addr, 0);
92 
93 	result->accessed = false;
94 	result->page_sz = PAGE_SIZE;
95 	while (page_vma_mapped_walk(&pvmw)) {
96 		addr = pvmw.address;
97 		if (pvmw.pte) {
98 			result->accessed = pte_young(*pvmw.pte) ||
99 				!folio_test_idle(folio) ||
100 				mmu_notifier_test_young(vma->vm_mm, addr);
101 		} else {
102 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
103 			result->accessed = pmd_young(*pvmw.pmd) ||
104 				!folio_test_idle(folio) ||
105 				mmu_notifier_test_young(vma->vm_mm, addr);
106 			result->page_sz = HPAGE_PMD_SIZE;
107 #else
108 			WARN_ON_ONCE(1);
109 #endif	/* CONFIG_TRANSPARENT_HUGEPAGE */
110 		}
111 		if (result->accessed) {
112 			page_vma_mapped_walk_done(&pvmw);
113 			break;
114 		}
115 	}
116 
117 	/* If accessed, stop walking */
118 	return !result->accessed;
119 }
120 
121 static bool damon_pa_young(unsigned long paddr, unsigned long *page_sz)
122 {
123 	struct folio *folio = damon_get_folio(PHYS_PFN(paddr));
124 	struct damon_pa_access_chk_result result = {
125 		.page_sz = PAGE_SIZE,
126 		.accessed = false,
127 	};
128 	struct rmap_walk_control rwc = {
129 		.arg = &result,
130 		.rmap_one = __damon_pa_young,
131 		.anon_lock = folio_lock_anon_vma_read,
132 	};
133 	bool need_lock;
134 
135 	if (!folio)
136 		return false;
137 
138 	if (!folio_mapped(folio) || !folio_raw_mapping(folio)) {
139 		if (folio_test_idle(folio))
140 			result.accessed = false;
141 		else
142 			result.accessed = true;
143 		folio_put(folio);
144 		goto out;
145 	}
146 
147 	need_lock = !folio_test_anon(folio) || folio_test_ksm(folio);
148 	if (need_lock && !folio_trylock(folio)) {
149 		folio_put(folio);
150 		return false;
151 	}
152 
153 	rmap_walk(folio, &rwc);
154 
155 	if (need_lock)
156 		folio_unlock(folio);
157 	folio_put(folio);
158 
159 out:
160 	*page_sz = result.page_sz;
161 	return result.accessed;
162 }
163 
164 static void __damon_pa_check_access(struct damon_region *r)
165 {
166 	static unsigned long last_addr;
167 	static unsigned long last_page_sz = PAGE_SIZE;
168 	static bool last_accessed;
169 
170 	/* If the region is in the last checked page, reuse the result */
171 	if (ALIGN_DOWN(last_addr, last_page_sz) ==
172 				ALIGN_DOWN(r->sampling_addr, last_page_sz)) {
173 		if (last_accessed)
174 			r->nr_accesses++;
175 		return;
176 	}
177 
178 	last_accessed = damon_pa_young(r->sampling_addr, &last_page_sz);
179 	if (last_accessed)
180 		r->nr_accesses++;
181 
182 	last_addr = r->sampling_addr;
183 }
184 
185 static unsigned int damon_pa_check_accesses(struct damon_ctx *ctx)
186 {
187 	struct damon_target *t;
188 	struct damon_region *r;
189 	unsigned int max_nr_accesses = 0;
190 
191 	damon_for_each_target(t, ctx) {
192 		damon_for_each_region(r, t) {
193 			__damon_pa_check_access(r);
194 			max_nr_accesses = max(r->nr_accesses, max_nr_accesses);
195 		}
196 	}
197 
198 	return max_nr_accesses;
199 }
200 
201 static bool __damos_pa_filter_out(struct damos_filter *filter,
202 		struct folio *folio)
203 {
204 	bool matched = false;
205 	struct mem_cgroup *memcg;
206 
207 	switch (filter->type) {
208 	case DAMOS_FILTER_TYPE_ANON:
209 		matched = folio_test_anon(folio);
210 		break;
211 	case DAMOS_FILTER_TYPE_MEMCG:
212 		rcu_read_lock();
213 		memcg = folio_memcg_check(folio);
214 		if (!memcg)
215 			matched = false;
216 		else
217 			matched = filter->memcg_id == mem_cgroup_id(memcg);
218 		rcu_read_unlock();
219 		break;
220 	default:
221 		break;
222 	}
223 
224 	return matched == filter->matching;
225 }
226 
227 /*
228  * damos_pa_filter_out - Return true if the page should be filtered out.
229  */
230 static bool damos_pa_filter_out(struct damos *scheme, struct folio *folio)
231 {
232 	struct damos_filter *filter;
233 
234 	damos_for_each_filter(filter, scheme) {
235 		if (__damos_pa_filter_out(filter, folio))
236 			return true;
237 	}
238 	return false;
239 }
240 
241 static unsigned long damon_pa_pageout(struct damon_region *r, struct damos *s)
242 {
243 	unsigned long addr, applied;
244 	LIST_HEAD(folio_list);
245 
246 	for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
247 		struct folio *folio = damon_get_folio(PHYS_PFN(addr));
248 
249 		if (!folio)
250 			continue;
251 
252 		if (damos_pa_filter_out(s, folio)) {
253 			folio_put(folio);
254 			continue;
255 		}
256 
257 		folio_clear_referenced(folio);
258 		folio_test_clear_young(folio);
259 		if (folio_isolate_lru(folio)) {
260 			folio_put(folio);
261 			continue;
262 		}
263 		if (folio_test_unevictable(folio)) {
264 			folio_putback_lru(folio);
265 		} else {
266 			list_add(&folio->lru, &folio_list);
267 			folio_put(folio);
268 		}
269 	}
270 	applied = reclaim_pages(&folio_list);
271 	cond_resched();
272 	return applied * PAGE_SIZE;
273 }
274 
275 static inline unsigned long damon_pa_mark_accessed_or_deactivate(
276 		struct damon_region *r, struct damos *s, bool mark_accessed)
277 {
278 	unsigned long addr, applied = 0;
279 
280 	for (addr = r->ar.start; addr < r->ar.end; addr += PAGE_SIZE) {
281 		struct folio *folio = damon_get_folio(PHYS_PFN(addr));
282 
283 		if (!folio)
284 			continue;
285 
286 		if (damos_pa_filter_out(s, folio)) {
287 			folio_put(folio);
288 			continue;
289 		}
290 
291 		if (mark_accessed)
292 			folio_mark_accessed(folio);
293 		else
294 			folio_deactivate(folio);
295 		folio_put(folio);
296 		applied += folio_nr_pages(folio);
297 	}
298 	return applied * PAGE_SIZE;
299 }
300 
301 static unsigned long damon_pa_mark_accessed(struct damon_region *r,
302 	struct damos *s)
303 {
304 	return damon_pa_mark_accessed_or_deactivate(r, s, true);
305 }
306 
307 static unsigned long damon_pa_deactivate_pages(struct damon_region *r,
308 	struct damos *s)
309 {
310 	return damon_pa_mark_accessed_or_deactivate(r, s, false);
311 }
312 
313 static unsigned long damon_pa_apply_scheme(struct damon_ctx *ctx,
314 		struct damon_target *t, struct damon_region *r,
315 		struct damos *scheme)
316 {
317 	switch (scheme->action) {
318 	case DAMOS_PAGEOUT:
319 		return damon_pa_pageout(r, scheme);
320 	case DAMOS_LRU_PRIO:
321 		return damon_pa_mark_accessed(r, scheme);
322 	case DAMOS_LRU_DEPRIO:
323 		return damon_pa_deactivate_pages(r, scheme);
324 	case DAMOS_STAT:
325 		break;
326 	default:
327 		/* DAMOS actions that not yet supported by 'paddr'. */
328 		break;
329 	}
330 	return 0;
331 }
332 
333 static int damon_pa_scheme_score(struct damon_ctx *context,
334 		struct damon_target *t, struct damon_region *r,
335 		struct damos *scheme)
336 {
337 	switch (scheme->action) {
338 	case DAMOS_PAGEOUT:
339 		return damon_cold_score(context, r, scheme);
340 	case DAMOS_LRU_PRIO:
341 		return damon_hot_score(context, r, scheme);
342 	case DAMOS_LRU_DEPRIO:
343 		return damon_cold_score(context, r, scheme);
344 	default:
345 		break;
346 	}
347 
348 	return DAMOS_MAX_SCORE;
349 }
350 
351 static int __init damon_pa_initcall(void)
352 {
353 	struct damon_operations ops = {
354 		.id = DAMON_OPS_PADDR,
355 		.init = NULL,
356 		.update = NULL,
357 		.prepare_access_checks = damon_pa_prepare_access_checks,
358 		.check_accesses = damon_pa_check_accesses,
359 		.reset_aggregated = NULL,
360 		.target_valid = NULL,
361 		.cleanup = NULL,
362 		.apply_scheme = damon_pa_apply_scheme,
363 		.get_scheme_score = damon_pa_scheme_score,
364 	};
365 
366 	return damon_register_ops(&ops);
367 };
368 
369 subsys_initcall(damon_pa_initcall);
370