| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * DAMON Code 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 | #include <linux/memory-tiers.h> |
| 16 | #include <linux/mm_inline.h> |
| 17 | |
| 18 | #include "../internal.h" |
| 19 | #include "ops-common.h" |
| 20 | |
| 21 | static phys_addr_t damon_pa_phys_addr( |
| 22 | unsigned long addr, unsigned long addr_unit) |
| 23 | { |
| 24 | return (phys_addr_t)addr * addr_unit; |
| 25 | } |
| 26 | |
| 27 | static unsigned long damon_pa_core_addr( |
| 28 | phys_addr_t pa, unsigned long addr_unit) |
| 29 | { |
| 30 | /* |
| 31 | * Use div_u64() for avoiding linking errors related with __udivdi3, |
| 32 | * __aeabi_uldivmod, or similar problems. This should also improve the |
| 33 | * performance optimization (read div_u64() comment for the detail). |
| 34 | */ |
| 35 | if (sizeof(pa) == 8 && sizeof(addr_unit) == 4) |
| 36 | return div_u64(dividend: pa, divisor: addr_unit); |
| 37 | return pa / addr_unit; |
| 38 | } |
| 39 | |
| 40 | static void damon_pa_mkold(phys_addr_t paddr) |
| 41 | { |
| 42 | struct folio *folio = damon_get_folio(PHYS_PFN(paddr)); |
| 43 | |
| 44 | if (!folio) |
| 45 | return; |
| 46 | |
| 47 | damon_folio_mkold(folio); |
| 48 | folio_put(folio); |
| 49 | } |
| 50 | |
| 51 | static void __damon_pa_prepare_access_check(struct damon_region *r, |
| 52 | unsigned long addr_unit) |
| 53 | { |
| 54 | r->sampling_addr = damon_rand(l: r->ar.start, r: r->ar.end); |
| 55 | |
| 56 | damon_pa_mkold(paddr: damon_pa_phys_addr(addr: r->sampling_addr, addr_unit)); |
| 57 | } |
| 58 | |
| 59 | static void damon_pa_prepare_access_checks(struct damon_ctx *ctx) |
| 60 | { |
| 61 | struct damon_target *t; |
| 62 | struct damon_region *r; |
| 63 | |
| 64 | damon_for_each_target(t, ctx) { |
| 65 | damon_for_each_region(r, t) |
| 66 | __damon_pa_prepare_access_check(r, addr_unit: ctx->addr_unit); |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | static bool damon_pa_young(phys_addr_t paddr, unsigned long *folio_sz) |
| 71 | { |
| 72 | struct folio *folio = damon_get_folio(PHYS_PFN(paddr)); |
| 73 | bool accessed; |
| 74 | |
| 75 | if (!folio) |
| 76 | return false; |
| 77 | |
| 78 | accessed = damon_folio_young(folio); |
| 79 | *folio_sz = folio_size(folio); |
| 80 | folio_put(folio); |
| 81 | return accessed; |
| 82 | } |
| 83 | |
| 84 | static void __damon_pa_check_access(struct damon_region *r, |
| 85 | struct damon_attrs *attrs, unsigned long addr_unit) |
| 86 | { |
| 87 | static phys_addr_t last_addr; |
| 88 | static unsigned long last_folio_sz = PAGE_SIZE; |
| 89 | static bool last_accessed; |
| 90 | phys_addr_t sampling_addr = damon_pa_phys_addr( |
| 91 | addr: r->sampling_addr, addr_unit); |
| 92 | |
| 93 | /* If the region is in the last checked page, reuse the result */ |
| 94 | if (ALIGN_DOWN(last_addr, last_folio_sz) == |
| 95 | ALIGN_DOWN(sampling_addr, last_folio_sz)) { |
| 96 | damon_update_region_access_rate(r, accessed: last_accessed, attrs); |
| 97 | return; |
| 98 | } |
| 99 | |
| 100 | last_accessed = damon_pa_young(paddr: sampling_addr, folio_sz: &last_folio_sz); |
| 101 | damon_update_region_access_rate(r, accessed: last_accessed, attrs); |
| 102 | |
| 103 | last_addr = sampling_addr; |
| 104 | } |
| 105 | |
| 106 | static unsigned int damon_pa_check_accesses(struct damon_ctx *ctx) |
| 107 | { |
| 108 | struct damon_target *t; |
| 109 | struct damon_region *r; |
| 110 | unsigned int max_nr_accesses = 0; |
| 111 | |
| 112 | damon_for_each_target(t, ctx) { |
| 113 | damon_for_each_region(r, t) { |
| 114 | __damon_pa_check_access( |
| 115 | r, attrs: &ctx->attrs, addr_unit: ctx->addr_unit); |
| 116 | max_nr_accesses = max(r->nr_accesses, max_nr_accesses); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | return max_nr_accesses; |
| 121 | } |
| 122 | |
| 123 | /* |
| 124 | * damos_pa_filter_out - Return true if the page should be filtered out. |
| 125 | */ |
| 126 | static bool damos_pa_filter_out(struct damos *scheme, struct folio *folio) |
| 127 | { |
| 128 | struct damos_filter *filter; |
| 129 | |
| 130 | if (scheme->core_filters_allowed) |
| 131 | return false; |
| 132 | |
| 133 | damos_for_each_ops_filter(filter, scheme) { |
| 134 | if (damos_folio_filter_match(filter, folio)) |
| 135 | return !filter->allow; |
| 136 | } |
| 137 | return scheme->ops_filters_default_reject; |
| 138 | } |
| 139 | |
| 140 | static bool damon_pa_invalid_damos_folio(struct folio *folio, struct damos *s) |
| 141 | { |
| 142 | if (!folio) |
| 143 | return true; |
| 144 | if (folio == s->last_applied) { |
| 145 | folio_put(folio); |
| 146 | return true; |
| 147 | } |
| 148 | return false; |
| 149 | } |
| 150 | |
| 151 | static unsigned long damon_pa_pageout(struct damon_region *r, |
| 152 | unsigned long addr_unit, struct damos *s, |
| 153 | unsigned long *sz_filter_passed) |
| 154 | { |
| 155 | phys_addr_t addr, applied; |
| 156 | LIST_HEAD(folio_list); |
| 157 | bool install_young_filter = true; |
| 158 | struct damos_filter *filter; |
| 159 | struct folio *folio; |
| 160 | |
| 161 | /* check access in page level again by default */ |
| 162 | damos_for_each_ops_filter(filter, s) { |
| 163 | if (filter->type == DAMOS_FILTER_TYPE_YOUNG) { |
| 164 | install_young_filter = false; |
| 165 | break; |
| 166 | } |
| 167 | } |
| 168 | if (install_young_filter) { |
| 169 | filter = damos_new_filter( |
| 170 | type: DAMOS_FILTER_TYPE_YOUNG, matching: true, allow: false); |
| 171 | if (!filter) |
| 172 | return 0; |
| 173 | damos_add_filter(s, f: filter); |
| 174 | } |
| 175 | |
| 176 | addr = damon_pa_phys_addr(addr: r->ar.start, addr_unit); |
| 177 | while (addr < damon_pa_phys_addr(addr: r->ar.end, addr_unit)) { |
| 178 | folio = damon_get_folio(PHYS_PFN(addr)); |
| 179 | if (damon_pa_invalid_damos_folio(folio, s)) { |
| 180 | addr += PAGE_SIZE; |
| 181 | continue; |
| 182 | } |
| 183 | |
| 184 | if (damos_pa_filter_out(scheme: s, folio)) |
| 185 | goto put_folio; |
| 186 | else |
| 187 | *sz_filter_passed += folio_size(folio) / addr_unit; |
| 188 | |
| 189 | folio_clear_referenced(folio); |
| 190 | folio_test_clear_young(folio); |
| 191 | if (!folio_isolate_lru(folio)) |
| 192 | goto put_folio; |
| 193 | if (folio_test_unevictable(folio)) |
| 194 | folio_putback_lru(folio); |
| 195 | else |
| 196 | list_add(new: &folio->lru, head: &folio_list); |
| 197 | put_folio: |
| 198 | addr += folio_size(folio); |
| 199 | folio_put(folio); |
| 200 | } |
| 201 | if (install_young_filter) |
| 202 | damos_destroy_filter(f: filter); |
| 203 | applied = reclaim_pages(folio_list: &folio_list); |
| 204 | cond_resched(); |
| 205 | s->last_applied = folio; |
| 206 | return damon_pa_core_addr(pa: applied * PAGE_SIZE, addr_unit); |
| 207 | } |
| 208 | |
| 209 | static inline unsigned long damon_pa_mark_accessed_or_deactivate( |
| 210 | struct damon_region *r, unsigned long addr_unit, |
| 211 | struct damos *s, bool mark_accessed, |
| 212 | unsigned long *sz_filter_passed) |
| 213 | { |
| 214 | phys_addr_t addr, applied = 0; |
| 215 | struct folio *folio; |
| 216 | |
| 217 | addr = damon_pa_phys_addr(addr: r->ar.start, addr_unit); |
| 218 | while (addr < damon_pa_phys_addr(addr: r->ar.end, addr_unit)) { |
| 219 | folio = damon_get_folio(PHYS_PFN(addr)); |
| 220 | if (damon_pa_invalid_damos_folio(folio, s)) { |
| 221 | addr += PAGE_SIZE; |
| 222 | continue; |
| 223 | } |
| 224 | |
| 225 | if (damos_pa_filter_out(scheme: s, folio)) |
| 226 | goto put_folio; |
| 227 | else |
| 228 | *sz_filter_passed += folio_size(folio) / addr_unit; |
| 229 | |
| 230 | if (mark_accessed) |
| 231 | folio_mark_accessed(folio); |
| 232 | else |
| 233 | folio_deactivate(folio); |
| 234 | applied += folio_nr_pages(folio); |
| 235 | put_folio: |
| 236 | addr += folio_size(folio); |
| 237 | folio_put(folio); |
| 238 | } |
| 239 | s->last_applied = folio; |
| 240 | return damon_pa_core_addr(pa: applied * PAGE_SIZE, addr_unit); |
| 241 | } |
| 242 | |
| 243 | static unsigned long damon_pa_mark_accessed(struct damon_region *r, |
| 244 | unsigned long addr_unit, struct damos *s, |
| 245 | unsigned long *sz_filter_passed) |
| 246 | { |
| 247 | return damon_pa_mark_accessed_or_deactivate(r, addr_unit, s, mark_accessed: true, |
| 248 | sz_filter_passed); |
| 249 | } |
| 250 | |
| 251 | static unsigned long damon_pa_deactivate_pages(struct damon_region *r, |
| 252 | unsigned long addr_unit, struct damos *s, |
| 253 | unsigned long *sz_filter_passed) |
| 254 | { |
| 255 | return damon_pa_mark_accessed_or_deactivate(r, addr_unit, s, mark_accessed: false, |
| 256 | sz_filter_passed); |
| 257 | } |
| 258 | |
| 259 | static unsigned long damon_pa_migrate(struct damon_region *r, |
| 260 | unsigned long addr_unit, struct damos *s, |
| 261 | unsigned long *sz_filter_passed) |
| 262 | { |
| 263 | phys_addr_t addr, applied; |
| 264 | LIST_HEAD(folio_list); |
| 265 | struct folio *folio; |
| 266 | |
| 267 | addr = damon_pa_phys_addr(addr: r->ar.start, addr_unit); |
| 268 | while (addr < damon_pa_phys_addr(addr: r->ar.end, addr_unit)) { |
| 269 | folio = damon_get_folio(PHYS_PFN(addr)); |
| 270 | if (damon_pa_invalid_damos_folio(folio, s)) { |
| 271 | addr += PAGE_SIZE; |
| 272 | continue; |
| 273 | } |
| 274 | |
| 275 | if (damos_pa_filter_out(scheme: s, folio)) |
| 276 | goto put_folio; |
| 277 | else |
| 278 | *sz_filter_passed += folio_size(folio) / addr_unit; |
| 279 | |
| 280 | if (!folio_isolate_lru(folio)) |
| 281 | goto put_folio; |
| 282 | list_add(new: &folio->lru, head: &folio_list); |
| 283 | put_folio: |
| 284 | addr += folio_size(folio); |
| 285 | folio_put(folio); |
| 286 | } |
| 287 | applied = damon_migrate_pages(folio_list: &folio_list, target_nid: s->target_nid); |
| 288 | cond_resched(); |
| 289 | s->last_applied = folio; |
| 290 | return damon_pa_core_addr(pa: applied * PAGE_SIZE, addr_unit); |
| 291 | } |
| 292 | |
| 293 | static unsigned long damon_pa_stat(struct damon_region *r, |
| 294 | unsigned long addr_unit, struct damos *s, |
| 295 | unsigned long *sz_filter_passed) |
| 296 | { |
| 297 | phys_addr_t addr; |
| 298 | struct folio *folio; |
| 299 | |
| 300 | if (!damos_ops_has_filter(s)) |
| 301 | return 0; |
| 302 | |
| 303 | addr = damon_pa_phys_addr(addr: r->ar.start, addr_unit); |
| 304 | while (addr < damon_pa_phys_addr(addr: r->ar.end, addr_unit)) { |
| 305 | folio = damon_get_folio(PHYS_PFN(addr)); |
| 306 | if (damon_pa_invalid_damos_folio(folio, s)) { |
| 307 | addr += PAGE_SIZE; |
| 308 | continue; |
| 309 | } |
| 310 | |
| 311 | if (!damos_pa_filter_out(scheme: s, folio)) |
| 312 | *sz_filter_passed += folio_size(folio) / addr_unit; |
| 313 | addr += folio_size(folio); |
| 314 | folio_put(folio); |
| 315 | } |
| 316 | s->last_applied = folio; |
| 317 | return 0; |
| 318 | } |
| 319 | |
| 320 | static unsigned long damon_pa_apply_scheme(struct damon_ctx *ctx, |
| 321 | struct damon_target *t, struct damon_region *r, |
| 322 | struct damos *scheme, unsigned long *sz_filter_passed) |
| 323 | { |
| 324 | unsigned long aunit = ctx->addr_unit; |
| 325 | |
| 326 | switch (scheme->action) { |
| 327 | case DAMOS_PAGEOUT: |
| 328 | return damon_pa_pageout(r, addr_unit: aunit, s: scheme, sz_filter_passed); |
| 329 | case DAMOS_LRU_PRIO: |
| 330 | return damon_pa_mark_accessed(r, addr_unit: aunit, s: scheme, |
| 331 | sz_filter_passed); |
| 332 | case DAMOS_LRU_DEPRIO: |
| 333 | return damon_pa_deactivate_pages(r, addr_unit: aunit, s: scheme, |
| 334 | sz_filter_passed); |
| 335 | case DAMOS_MIGRATE_HOT: |
| 336 | case DAMOS_MIGRATE_COLD: |
| 337 | return damon_pa_migrate(r, addr_unit: aunit, s: scheme, sz_filter_passed); |
| 338 | case DAMOS_STAT: |
| 339 | return damon_pa_stat(r, addr_unit: aunit, s: scheme, sz_filter_passed); |
| 340 | default: |
| 341 | /* DAMOS actions that not yet supported by 'paddr'. */ |
| 342 | break; |
| 343 | } |
| 344 | return 0; |
| 345 | } |
| 346 | |
| 347 | static int damon_pa_scheme_score(struct damon_ctx *context, |
| 348 | struct damon_target *t, struct damon_region *r, |
| 349 | struct damos *scheme) |
| 350 | { |
| 351 | switch (scheme->action) { |
| 352 | case DAMOS_PAGEOUT: |
| 353 | return damon_cold_score(c: context, r, s: scheme); |
| 354 | case DAMOS_LRU_PRIO: |
| 355 | return damon_hot_score(c: context, r, s: scheme); |
| 356 | case DAMOS_LRU_DEPRIO: |
| 357 | return damon_cold_score(c: context, r, s: scheme); |
| 358 | case DAMOS_MIGRATE_HOT: |
| 359 | return damon_hot_score(c: context, r, s: scheme); |
| 360 | case DAMOS_MIGRATE_COLD: |
| 361 | return damon_cold_score(c: context, r, s: scheme); |
| 362 | default: |
| 363 | break; |
| 364 | } |
| 365 | |
| 366 | return DAMOS_MAX_SCORE; |
| 367 | } |
| 368 | |
| 369 | static int __init damon_pa_initcall(void) |
| 370 | { |
| 371 | struct damon_operations ops = { |
| 372 | .id = DAMON_OPS_PADDR, |
| 373 | .init = NULL, |
| 374 | .update = NULL, |
| 375 | .prepare_access_checks = damon_pa_prepare_access_checks, |
| 376 | .check_accesses = damon_pa_check_accesses, |
| 377 | .target_valid = NULL, |
| 378 | .cleanup = NULL, |
| 379 | .apply_scheme = damon_pa_apply_scheme, |
| 380 | .get_scheme_score = damon_pa_scheme_score, |
| 381 | }; |
| 382 | |
| 383 | return damon_register_ops(ops: &ops); |
| 384 | }; |
| 385 | |
| 386 | subsys_initcall(damon_pa_initcall); |
| 387 | |