1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * DAMON-based page reclamation 4 * 5 * Author: SeongJae Park <sj@kernel.org> 6 */ 7 8 #define pr_fmt(fmt) "damon-reclaim: " fmt 9 10 #include <linux/damon.h> 11 #include <linux/ioport.h> 12 #include <linux/module.h> 13 #include <linux/sched.h> 14 #include <linux/workqueue.h> 15 16 #ifdef MODULE_PARAM_PREFIX 17 #undef MODULE_PARAM_PREFIX 18 #endif 19 #define MODULE_PARAM_PREFIX "damon_reclaim." 20 21 /* 22 * Enable or disable DAMON_RECLAIM. 23 * 24 * You can enable DAMON_RCLAIM by setting the value of this parameter as ``Y``. 25 * Setting it as ``N`` disables DAMON_RECLAIM. Note that DAMON_RECLAIM could 26 * do no real monitoring and reclamation due to the watermarks-based activation 27 * condition. Refer to below descriptions for the watermarks parameter for 28 * this. 29 */ 30 static bool enabled __read_mostly; 31 module_param(enabled, bool, 0600); 32 33 /* 34 * Time threshold for cold memory regions identification in microseconds. 35 * 36 * If a memory region is not accessed for this or longer time, DAMON_RECLAIM 37 * identifies the region as cold, and reclaims. 120 seconds by default. 38 */ 39 static unsigned long min_age __read_mostly = 120000000; 40 module_param(min_age, ulong, 0600); 41 42 /* 43 * Limit of time for trying the reclamation in milliseconds. 44 * 45 * DAMON_RECLAIM tries to use only up to this time within a time window 46 * (quota_reset_interval_ms) for trying reclamation of cold pages. This can be 47 * used for limiting CPU consumption of DAMON_RECLAIM. If the value is zero, 48 * the limit is disabled. 49 * 50 * 10 ms by default. 51 */ 52 static unsigned long quota_ms __read_mostly = 10; 53 module_param(quota_ms, ulong, 0600); 54 55 /* 56 * Limit of size of memory for the reclamation in bytes. 57 * 58 * DAMON_RECLAIM charges amount of memory which it tried to reclaim within a 59 * time window (quota_reset_interval_ms) and makes no more than this limit is 60 * tried. This can be used for limiting consumption of CPU and IO. If this 61 * value is zero, the limit is disabled. 62 * 63 * 128 MiB by default. 64 */ 65 static unsigned long quota_sz __read_mostly = 128 * 1024 * 1024; 66 module_param(quota_sz, ulong, 0600); 67 68 /* 69 * The time/size quota charge reset interval in milliseconds. 70 * 71 * The charge reset interval for the quota of time (quota_ms) and size 72 * (quota_sz). That is, DAMON_RECLAIM does not try reclamation for more than 73 * quota_ms milliseconds or quota_sz bytes within quota_reset_interval_ms 74 * milliseconds. 75 * 76 * 1 second by default. 77 */ 78 static unsigned long quota_reset_interval_ms __read_mostly = 1000; 79 module_param(quota_reset_interval_ms, ulong, 0600); 80 81 /* 82 * The watermarks check time interval in microseconds. 83 * 84 * Minimal time to wait before checking the watermarks, when DAMON_RECLAIM is 85 * enabled but inactive due to its watermarks rule. 5 seconds by default. 86 */ 87 static unsigned long wmarks_interval __read_mostly = 5000000; 88 module_param(wmarks_interval, ulong, 0600); 89 90 /* 91 * Free memory rate (per thousand) for the high watermark. 92 * 93 * If free memory of the system in bytes per thousand bytes is higher than 94 * this, DAMON_RECLAIM becomes inactive, so it does nothing but periodically 95 * checks the watermarks. 500 (50%) by default. 96 */ 97 static unsigned long wmarks_high __read_mostly = 500; 98 module_param(wmarks_high, ulong, 0600); 99 100 /* 101 * Free memory rate (per thousand) for the middle watermark. 102 * 103 * If free memory of the system in bytes per thousand bytes is between this and 104 * the low watermark, DAMON_RECLAIM becomes active, so starts the monitoring 105 * and the reclaiming. 400 (40%) by default. 106 */ 107 static unsigned long wmarks_mid __read_mostly = 400; 108 module_param(wmarks_mid, ulong, 0600); 109 110 /* 111 * Free memory rate (per thousand) for the low watermark. 112 * 113 * If free memory of the system in bytes per thousand bytes is lower than this, 114 * DAMON_RECLAIM becomes inactive, so it does nothing but periodically checks 115 * the watermarks. In the case, the system falls back to the LRU-based page 116 * granularity reclamation logic. 200 (20%) by default. 117 */ 118 static unsigned long wmarks_low __read_mostly = 200; 119 module_param(wmarks_low, ulong, 0600); 120 121 /* 122 * Sampling interval for the monitoring in microseconds. 123 * 124 * The sampling interval of DAMON for the cold memory monitoring. Please refer 125 * to the DAMON documentation for more detail. 5 ms by default. 126 */ 127 static unsigned long sample_interval __read_mostly = 5000; 128 module_param(sample_interval, ulong, 0600); 129 130 /* 131 * Aggregation interval for the monitoring in microseconds. 132 * 133 * The aggregation interval of DAMON for the cold memory monitoring. Please 134 * refer to the DAMON documentation for more detail. 100 ms by default. 135 */ 136 static unsigned long aggr_interval __read_mostly = 100000; 137 module_param(aggr_interval, ulong, 0600); 138 139 /* 140 * Minimum number of monitoring regions. 141 * 142 * The minimal number of monitoring regions of DAMON for the cold memory 143 * monitoring. This can be used to set lower-bound of the monitoring quality. 144 * But, setting this too high could result in increased monitoring overhead. 145 * Please refer to the DAMON documentation for more detail. 10 by default. 146 */ 147 static unsigned long min_nr_regions __read_mostly = 10; 148 module_param(min_nr_regions, ulong, 0600); 149 150 /* 151 * Maximum number of monitoring regions. 152 * 153 * The maximum number of monitoring regions of DAMON for the cold memory 154 * monitoring. This can be used to set upper-bound of the monitoring overhead. 155 * However, setting this too low could result in bad monitoring quality. 156 * Please refer to the DAMON documentation for more detail. 1000 by default. 157 */ 158 static unsigned long max_nr_regions __read_mostly = 1000; 159 module_param(max_nr_regions, ulong, 0600); 160 161 /* 162 * Start of the target memory region in physical address. 163 * 164 * The start physical address of memory region that DAMON_RECLAIM will do work 165 * against. By default, biggest System RAM is used as the region. 166 */ 167 static unsigned long monitor_region_start __read_mostly; 168 module_param(monitor_region_start, ulong, 0600); 169 170 /* 171 * End of the target memory region in physical address. 172 * 173 * The end physical address of memory region that DAMON_RECLAIM will do work 174 * against. By default, biggest System RAM is used as the region. 175 */ 176 static unsigned long monitor_region_end __read_mostly; 177 module_param(monitor_region_end, ulong, 0600); 178 179 /* 180 * PID of the DAMON thread 181 * 182 * If DAMON_RECLAIM is enabled, this becomes the PID of the worker thread. 183 * Else, -1. 184 */ 185 static int kdamond_pid __read_mostly = -1; 186 module_param(kdamond_pid, int, 0400); 187 188 /* 189 * Number of memory regions that tried to be reclaimed. 190 */ 191 static unsigned long nr_reclaim_tried_regions __read_mostly; 192 module_param(nr_reclaim_tried_regions, ulong, 0400); 193 194 /* 195 * Total bytes of memory regions that tried to be reclaimed. 196 */ 197 static unsigned long bytes_reclaim_tried_regions __read_mostly; 198 module_param(bytes_reclaim_tried_regions, ulong, 0400); 199 200 /* 201 * Number of memory regions that successfully be reclaimed. 202 */ 203 static unsigned long nr_reclaimed_regions __read_mostly; 204 module_param(nr_reclaimed_regions, ulong, 0400); 205 206 /* 207 * Total bytes of memory regions that successfully be reclaimed. 208 */ 209 static unsigned long bytes_reclaimed_regions __read_mostly; 210 module_param(bytes_reclaimed_regions, ulong, 0400); 211 212 /* 213 * Number of times that the time/space quota limits have exceeded 214 */ 215 static unsigned long nr_quota_exceeds __read_mostly; 216 module_param(nr_quota_exceeds, ulong, 0400); 217 218 static struct damon_ctx *ctx; 219 static struct damon_target *target; 220 221 struct damon_reclaim_ram_walk_arg { 222 unsigned long start; 223 unsigned long end; 224 }; 225 226 static int walk_system_ram(struct resource *res, void *arg) 227 { 228 struct damon_reclaim_ram_walk_arg *a = arg; 229 230 if (a->end - a->start < res->end - res->start) { 231 a->start = res->start; 232 a->end = res->end; 233 } 234 return 0; 235 } 236 237 /* 238 * Find biggest 'System RAM' resource and store its start and end address in 239 * @start and @end, respectively. If no System RAM is found, returns false. 240 */ 241 static bool get_monitoring_region(unsigned long *start, unsigned long *end) 242 { 243 struct damon_reclaim_ram_walk_arg arg = {}; 244 245 walk_system_ram_res(0, ULONG_MAX, &arg, walk_system_ram); 246 if (arg.end <= arg.start) 247 return false; 248 249 *start = arg.start; 250 *end = arg.end; 251 return true; 252 } 253 254 static struct damos *damon_reclaim_new_scheme(void) 255 { 256 struct damos_watermarks wmarks = { 257 .metric = DAMOS_WMARK_FREE_MEM_RATE, 258 .interval = wmarks_interval, 259 .high = wmarks_high, 260 .mid = wmarks_mid, 261 .low = wmarks_low, 262 }; 263 struct damos_quota quota = { 264 /* 265 * Do not try reclamation for more than quota_ms milliseconds 266 * or quota_sz bytes within quota_reset_interval_ms. 267 */ 268 .ms = quota_ms, 269 .sz = quota_sz, 270 .reset_interval = quota_reset_interval_ms, 271 /* Within the quota, page out older regions first. */ 272 .weight_sz = 0, 273 .weight_nr_accesses = 0, 274 .weight_age = 1 275 }; 276 struct damos *scheme = damon_new_scheme( 277 /* Find regions having PAGE_SIZE or larger size */ 278 PAGE_SIZE, ULONG_MAX, 279 /* and not accessed at all */ 280 0, 0, 281 /* for min_age or more micro-seconds, and */ 282 min_age / aggr_interval, UINT_MAX, 283 /* page out those, as soon as found */ 284 DAMOS_PAGEOUT, 285 /* under the quota. */ 286 "a, 287 /* (De)activate this according to the watermarks. */ 288 &wmarks); 289 290 return scheme; 291 } 292 293 static int damon_reclaim_turn(bool on) 294 { 295 struct damon_region *region; 296 struct damos *scheme; 297 int err; 298 299 if (!on) { 300 err = damon_stop(&ctx, 1); 301 if (!err) 302 kdamond_pid = -1; 303 return err; 304 } 305 306 err = damon_set_attrs(ctx, sample_interval, aggr_interval, 0, 307 min_nr_regions, max_nr_regions); 308 if (err) 309 return err; 310 311 if (monitor_region_start > monitor_region_end) 312 return -EINVAL; 313 if (!monitor_region_start && !monitor_region_end && 314 !get_monitoring_region(&monitor_region_start, 315 &monitor_region_end)) 316 return -EINVAL; 317 /* DAMON will free this on its own when finish monitoring */ 318 region = damon_new_region(monitor_region_start, monitor_region_end); 319 if (!region) 320 return -ENOMEM; 321 damon_add_region(region, target); 322 323 /* Will be freed by 'damon_set_schemes()' below */ 324 scheme = damon_reclaim_new_scheme(); 325 if (!scheme) { 326 err = -ENOMEM; 327 goto free_region_out; 328 } 329 err = damon_set_schemes(ctx, &scheme, 1); 330 if (err) 331 goto free_scheme_out; 332 333 err = damon_start(&ctx, 1); 334 if (!err) { 335 kdamond_pid = ctx->kdamond->pid; 336 return 0; 337 } 338 339 free_scheme_out: 340 damon_destroy_scheme(scheme); 341 free_region_out: 342 damon_destroy_region(region, target); 343 return err; 344 } 345 346 #define ENABLE_CHECK_INTERVAL_MS 1000 347 static struct delayed_work damon_reclaim_timer; 348 static void damon_reclaim_timer_fn(struct work_struct *work) 349 { 350 static bool last_enabled; 351 bool now_enabled; 352 353 now_enabled = enabled; 354 if (last_enabled != now_enabled) { 355 if (!damon_reclaim_turn(now_enabled)) 356 last_enabled = now_enabled; 357 else 358 enabled = last_enabled; 359 } 360 361 schedule_delayed_work(&damon_reclaim_timer, 362 msecs_to_jiffies(ENABLE_CHECK_INTERVAL_MS)); 363 } 364 static DECLARE_DELAYED_WORK(damon_reclaim_timer, damon_reclaim_timer_fn); 365 366 static int damon_reclaim_after_aggregation(struct damon_ctx *c) 367 { 368 struct damos *s; 369 370 /* update the stats parameter */ 371 damon_for_each_scheme(s, c) { 372 nr_reclaim_tried_regions = s->stat.nr_tried; 373 bytes_reclaim_tried_regions = s->stat.sz_tried; 374 nr_reclaimed_regions = s->stat.nr_applied; 375 bytes_reclaimed_regions = s->stat.sz_applied; 376 nr_quota_exceeds = s->stat.qt_exceeds; 377 } 378 return 0; 379 } 380 381 static int __init damon_reclaim_init(void) 382 { 383 ctx = damon_new_ctx(); 384 if (!ctx) 385 return -ENOMEM; 386 387 if (damon_select_ops(ctx, DAMON_OPS_PADDR)) 388 return -EINVAL; 389 390 ctx->callback.after_aggregation = damon_reclaim_after_aggregation; 391 392 target = damon_new_target(); 393 if (!target) { 394 damon_destroy_ctx(ctx); 395 return -ENOMEM; 396 } 397 damon_add_target(ctx, target); 398 399 schedule_delayed_work(&damon_reclaim_timer, 0); 400 return 0; 401 } 402 403 module_init(damon_reclaim_init); 404