| 1 | /* |
| 2 | * Copyright (c) 2006-2021 Apple Inc. All rights reserved. |
| 3 | * |
| 4 | * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ |
| 5 | * |
| 6 | * This file contains Original Code and/or Modifications of Original Code |
| 7 | * as defined in and that are subject to the Apple Public Source License |
| 8 | * Version 2.0 (the 'License'). You may not use this file except in |
| 9 | * compliance with the License. The rights granted to you under the License |
| 10 | * may not be used to create, or enable the creation or redistribution of, |
| 11 | * unlawful or unlicensed copies of an Apple operating system, or to |
| 12 | * circumvent, violate, or enable the circumvention or violation of, any |
| 13 | * terms of an Apple operating system software license agreement. |
| 14 | * |
| 15 | * Please obtain a copy of the License at |
| 16 | * http://www.opensource.apple.com/apsl/ and read it before using this file. |
| 17 | * |
| 18 | * The Original Code and all software distributed under the License are |
| 19 | * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER |
| 20 | * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, |
| 21 | * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, |
| 22 | * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. |
| 23 | * Please see the License for the specific language governing rights and |
| 24 | * limitations under the License. |
| 25 | * |
| 26 | * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ |
| 27 | * |
| 28 | */ |
| 29 | |
| 30 | #include <kern/task.h> |
| 31 | #include <libkern/libkern.h> |
| 32 | #include <machine/atomic.h> |
| 33 | #include <mach/coalition.h> |
| 34 | #include <os/log.h> |
| 35 | #include <sys/coalition.h> |
| 36 | #include <sys/proc.h> |
| 37 | #include <sys/proc_internal.h> |
| 38 | #include <sys/kdebug.h> |
| 39 | #include <sys/kern_memorystatus.h> |
| 40 | #include <vm/vm_protos.h> |
| 41 | |
| 42 | #include <kern/kern_memorystatus_internal.h> |
| 43 | |
| 44 | /* |
| 45 | * All memory pressure policy decisions should live here, and there should be |
| 46 | * as little mechanism as possible. This file prioritizes readability. |
| 47 | */ |
| 48 | |
| 49 | #pragma mark Policy Function Declarations |
| 50 | |
| 51 | #if CONFIG_JETSAM |
| 52 | static bool memorystatus_check_aggressive_jetsam_needed(int *jld_idle_kills); |
| 53 | #endif /* CONFIG_JETSAM */ |
| 54 | |
| 55 | #pragma mark Memorystatus Health Check |
| 56 | |
| 57 | /* |
| 58 | * Each subsystem that relies on the memorystatus thread |
| 59 | * for resource exhaustion should put a health check in this section. |
| 60 | * The memorystatus thread runs all of the health checks |
| 61 | * to determine if the system is healthy. If the system is unhealthy |
| 62 | * it picks an action based on the system health status. See the |
| 63 | * Memorystatus Thread Actions section below. |
| 64 | */ |
| 65 | |
| 66 | extern bool vm_compressor_needs_to_swap(bool wake_memorystatus_thread); |
| 67 | extern boolean_t vm_compressor_low_on_space(void); |
| 68 | extern bool vm_compressor_compressed_pages_nearing_limit(void); |
| 69 | extern bool vm_compressor_is_thrashing(void); |
| 70 | extern bool vm_compressor_swapout_is_ripe(void); |
| 71 | |
| 72 | #if XNU_TARGET_OS_WATCH |
| 73 | #define FREEZE_PREVENT_REFREEZE_OF_LAST_THAWED true |
| 74 | #define FREEZE_PREVENT_REFREEZE_OF_LAST_THAWED_TIMEOUT_SECONDS (60 * 15) |
| 75 | #else |
| 76 | #define FREEZE_PREVENT_REFREEZE_OF_LAST_THAWED false |
| 77 | #endif |
| 78 | extern pid_t memorystatus_freeze_last_pid_thawed; |
| 79 | extern uint64_t memorystatus_freeze_last_pid_thawed_ts; |
| 80 | |
| 81 | static void |
| 82 | memorystatus_health_check(memorystatus_system_health_t *status) |
| 83 | { |
| 84 | memset(s: status, c: 0, n: sizeof(memorystatus_system_health_t)); |
| 85 | #if CONFIG_JETSAM |
| 86 | status->msh_available_pages_below_pressure = memorystatus_avail_pages_below_pressure(); |
| 87 | status->msh_available_pages_below_critical = memorystatus_avail_pages_below_critical(); |
| 88 | status->msh_compressor_is_low_on_space = (vm_compressor_low_on_space() == TRUE); |
| 89 | status->msh_compressed_pages_nearing_limit = vm_compressor_compressed_pages_nearing_limit(); |
| 90 | status->msh_compressor_is_thrashing = !memorystatus_swap_all_apps && vm_compressor_is_thrashing(); |
| 91 | #if CONFIG_PHANTOM_CACHE |
| 92 | status->msh_phantom_cache_pressure = os_atomic_load(&memorystatus_phantom_cache_pressure, acquire); |
| 93 | #else |
| 94 | status->msh_phantom_cache_pressure = false; |
| 95 | #endif /* CONFIG_PHANTOM_CACHE */ |
| 96 | if (!memorystatus_swap_all_apps && |
| 97 | status->msh_phantom_cache_pressure && |
| 98 | !(status->msh_compressor_is_thrashing && status->msh_compressor_is_low_on_space)) { |
| 99 | status->msh_filecache_is_thrashing = true; |
| 100 | } |
| 101 | status->msh_compressor_is_low_on_space = os_atomic_load(&memorystatus_compressor_space_shortage, acquire); |
| 102 | status->msh_pageout_starved = os_atomic_load(&memorystatus_pageout_starved, acquire); |
| 103 | status->msh_swappable_compressor_segments_over_limit = memorystatus_swap_over_trigger(100); |
| 104 | status->msh_swapin_queue_over_limit = memorystatus_swapin_over_trigger(); |
| 105 | status->msh_swap_low_on_space = vm_swap_low_on_space(); |
| 106 | status->msh_swap_out_of_space = vm_swap_out_of_space(); |
| 107 | #endif /* CONFIG_JETSAM */ |
| 108 | status->msh_zone_map_is_exhausted = os_atomic_load(&memorystatus_zone_map_is_exhausted, relaxed); |
| 109 | } |
| 110 | |
| 111 | bool |
| 112 | memorystatus_is_system_healthy(const memorystatus_system_health_t *status) |
| 113 | { |
| 114 | #if CONFIG_JETSAM |
| 115 | return !(status->msh_available_pages_below_critical || |
| 116 | status->msh_compressor_is_low_on_space || |
| 117 | status->msh_compressor_is_thrashing || |
| 118 | status->msh_filecache_is_thrashing || |
| 119 | status->msh_zone_map_is_exhausted || |
| 120 | status->msh_pageout_starved); |
| 121 | #else /* CONFIG_JETSAM */ |
| 122 | return !status->msh_zone_map_is_exhausted; |
| 123 | #endif /* CONFIG_JETSAM */ |
| 124 | } |
| 125 | |
| 126 | |
| 127 | #pragma mark Memorystatus Thread Actions |
| 128 | |
| 129 | /* |
| 130 | * This section picks the appropriate memorystatus_action & deploys it. |
| 131 | */ |
| 132 | |
| 133 | /* |
| 134 | * Inspects the state of various resources in the system to see if |
| 135 | * the system is healthy. If the system is not healthy, picks a |
| 136 | * memorystatus_action_t to recover the system. |
| 137 | * |
| 138 | * Every time the memorystatus thread wakes up it calls into here |
| 139 | * to pick an action. It will continue performing memorystatus actions until this |
| 140 | * function returns MEMORYSTATUS_KILL_NONE. At that point the thread will block. |
| 141 | */ |
| 142 | memorystatus_action_t |
| 143 | memorystatus_pick_action(struct jetsam_thread_state *jetsam_thread, |
| 144 | uint32_t *kill_cause, |
| 145 | bool highwater_remaining, |
| 146 | bool suspended_swappable_apps_remaining, |
| 147 | bool swappable_apps_remaining, |
| 148 | int *jld_idle_kills) |
| 149 | { |
| 150 | memorystatus_system_health_t status; |
| 151 | memorystatus_health_check(status: &status); |
| 152 | memorystatus_log_system_health(health: &status); |
| 153 | bool is_system_healthy = memorystatus_is_system_healthy(status: &status); |
| 154 | |
| 155 | #if CONFIG_JETSAM |
| 156 | if (status.msh_available_pages_below_pressure || !is_system_healthy) { |
| 157 | /* |
| 158 | * If swap is enabled, first check if we're running low or are out of swap space. |
| 159 | */ |
| 160 | if (memorystatus_swap_all_apps && jetsam_kill_on_low_swap) { |
| 161 | if (swappable_apps_remaining && status.msh_swap_out_of_space) { |
| 162 | *kill_cause = kMemorystatusKilledLowSwap; |
| 163 | return MEMORYSTATUS_KILL_SWAPPABLE; |
| 164 | } else if (suspended_swappable_apps_remaining && status.msh_swap_low_on_space) { |
| 165 | *kill_cause = kMemorystatusKilledLowSwap; |
| 166 | return MEMORYSTATUS_KILL_SUSPENDED_SWAPPABLE; |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | /* |
| 171 | * We're below the pressure level or the system is unhealthy, |
| 172 | * regardless of the system health let's check if we should be swapping |
| 173 | * and if there are high watermark kills left to do. |
| 174 | */ |
| 175 | if (memorystatus_swap_all_apps) { |
| 176 | if (status.msh_swappable_compressor_segments_over_limit && !vm_swapout_thread_running && !os_atomic_load(&vm_swapout_wake_pending, relaxed)) { |
| 177 | /* |
| 178 | * TODO: The swapper will keep running until it has drained the entire early swapout queue. |
| 179 | * That might be overly aggressive & we should look into tuning it. |
| 180 | * See rdar://84102304. |
| 181 | */ |
| 182 | return MEMORYSTATUS_WAKE_SWAPPER; |
| 183 | } else if (status.msh_swapin_queue_over_limit) { |
| 184 | return MEMORYSTATUS_PROCESS_SWAPIN_QUEUE; |
| 185 | } else if (status.msh_swappable_compressor_segments_over_limit) { |
| 186 | memorystatus_log_info( |
| 187 | "memorystatus: Skipping swap wakeup because the swap thread is already running. vm_swapout_thread_running=%d, vm_swapout_wake_pending=%d\n" , |
| 188 | vm_swapout_thread_running, os_atomic_load(&vm_swapout_wake_pending, relaxed)); |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | if (highwater_remaining) { |
| 193 | *kill_cause = kMemorystatusKilledHiwat; |
| 194 | memorystatus_log("memorystatus: Looking for highwatermark kills.\n" ); |
| 195 | return MEMORYSTATUS_KILL_HIWATER; |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | if (is_system_healthy) { |
| 200 | *kill_cause = 0; |
| 201 | return MEMORYSTATUS_KILL_NONE; |
| 202 | } |
| 203 | |
| 204 | /* |
| 205 | * At this point the system is unhealthy and there are no |
| 206 | * more highwatermark processes to kill. |
| 207 | */ |
| 208 | |
| 209 | if (!jetsam_thread->limit_to_low_bands) { |
| 210 | if (memorystatus_check_aggressive_jetsam_needed(jld_idle_kills)) { |
| 211 | memorystatus_log("memorystatus: Starting aggressive jetsam.\n" ); |
| 212 | *kill_cause = kMemorystatusKilledProcThrashing; |
| 213 | return MEMORYSTATUS_KILL_AGGRESSIVE; |
| 214 | } |
| 215 | } |
| 216 | /* |
| 217 | * The system is unhealthy and we either don't need aggressive jetsam |
| 218 | * or are not allowed to deploy it. |
| 219 | * Kill in priority order. We'll use LRU within every band except the |
| 220 | * FG (which will be sorted by coalition role). |
| 221 | */ |
| 222 | *kill_cause = memorystatus_pick_kill_cause(&status); |
| 223 | return MEMORYSTATUS_KILL_TOP_PROCESS; |
| 224 | #else /* CONFIG_JETSAM */ |
| 225 | (void) jetsam_thread; |
| 226 | (void) jld_idle_kills; |
| 227 | (void) suspended_swappable_apps_remaining; |
| 228 | (void) swappable_apps_remaining; |
| 229 | /* |
| 230 | * Without CONFIG_JETSAM, we only kill if the system is unhealthy. |
| 231 | * There is no aggressive jetsam and no |
| 232 | * early highwatermark killing. |
| 233 | */ |
| 234 | if (is_system_healthy) { |
| 235 | *kill_cause = 0; |
| 236 | return MEMORYSTATUS_KILL_NONE; |
| 237 | } |
| 238 | if (highwater_remaining) { |
| 239 | *kill_cause = kMemorystatusKilledHiwat; |
| 240 | return MEMORYSTATUS_KILL_HIWATER; |
| 241 | } else { |
| 242 | *kill_cause = memorystatus_pick_kill_cause(status: &status); |
| 243 | return MEMORYSTATUS_KILL_TOP_PROCESS; |
| 244 | } |
| 245 | #endif /* CONFIG_JETSAM */ |
| 246 | } |
| 247 | |
| 248 | #pragma mark Aggressive Jetsam |
| 249 | /* |
| 250 | * This section defines when we deploy aggressive jetsam. |
| 251 | * Aggressive jetsam kills everything up to the jld_priority_band_max band. |
| 252 | */ |
| 253 | |
| 254 | #if CONFIG_JETSAM |
| 255 | |
| 256 | static bool |
| 257 | memorystatus_aggressive_jetsam_needed_sysproc_aging(__unused int jld_eval_aggressive_count, __unused int *jld_idle_kills, __unused int jld_idle_kill_candidates, int *total_candidates); |
| 258 | |
| 259 | /* |
| 260 | * kJetsamHighRelaunchCandidatesThreshold defines the percentage of candidates |
| 261 | * in the idle & deferred bands that need to be bad candidates in order to trigger |
| 262 | * aggressive jetsam. |
| 263 | */ |
| 264 | #define kJetsamHighRelaunchCandidatesThreshold (100) |
| 265 | |
| 266 | /* kJetsamMinCandidatesThreshold defines the minimum number of candidates in the |
| 267 | * idle/deferred bands to trigger aggressive jetsam. This value basically decides |
| 268 | * how much memory the system is ready to hold in the lower bands without triggering |
| 269 | * aggressive jetsam. This number should ideally be tuned based on the memory config |
| 270 | * of the device. |
| 271 | */ |
| 272 | #define kJetsamMinCandidatesThreshold (5) |
| 273 | |
| 274 | static bool |
| 275 | memorystatus_check_aggressive_jetsam_needed(int *jld_idle_kills) |
| 276 | { |
| 277 | bool aggressive_jetsam_needed = false; |
| 278 | int total_candidates = 0; |
| 279 | /* |
| 280 | * The aggressive jetsam logic looks at the number of times it has been in the |
| 281 | * aggressive loop to determine the max priority band it should kill upto. The |
| 282 | * static variables below are used to track that property. |
| 283 | * |
| 284 | * To reset those values, the implementation checks if it has been |
| 285 | * memorystatus_jld_eval_period_msecs since the parameters were reset. |
| 286 | */ |
| 287 | |
| 288 | if (memorystatus_jld_enabled == FALSE) { |
| 289 | /* If aggressive jetsam is disabled, nothing to do here */ |
| 290 | return FALSE; |
| 291 | } |
| 292 | |
| 293 | /* Get current timestamp (msecs only) */ |
| 294 | struct timeval jld_now_tstamp = {0, 0}; |
| 295 | uint64_t jld_now_msecs = 0; |
| 296 | microuptime(&jld_now_tstamp); |
| 297 | jld_now_msecs = (jld_now_tstamp.tv_sec * 1000); |
| 298 | |
| 299 | /* |
| 300 | * Look at the number of candidates in the idle and deferred band and |
| 301 | * how many out of them are marked as high relaunch probability. |
| 302 | */ |
| 303 | aggressive_jetsam_needed = memorystatus_aggressive_jetsam_needed_sysproc_aging(jld_eval_aggressive_count, |
| 304 | jld_idle_kills, jld_idle_kill_candidates, &total_candidates); |
| 305 | |
| 306 | /* |
| 307 | * Check if its been really long since the aggressive jetsam evaluation |
| 308 | * parameters have been refreshed. This logic also resets the jld_eval_aggressive_count |
| 309 | * counter to make sure we reset the aggressive jetsam severity. |
| 310 | */ |
| 311 | boolean_t param_reval = false; |
| 312 | |
| 313 | if ((total_candidates == 0) || |
| 314 | (jld_now_msecs > (jld_timestamp_msecs + memorystatus_jld_eval_period_msecs))) { |
| 315 | jld_timestamp_msecs = jld_now_msecs; |
| 316 | jld_idle_kill_candidates = total_candidates; |
| 317 | *jld_idle_kills = 0; |
| 318 | jld_eval_aggressive_count = 0; |
| 319 | jld_priority_band_max = JETSAM_PRIORITY_UI_SUPPORT; |
| 320 | param_reval = true; |
| 321 | } |
| 322 | |
| 323 | /* |
| 324 | * It is also possible that the system is down to a very small number of processes in the candidate |
| 325 | * bands. In that case, the decisions made by the memorystatus_aggressive_jetsam_needed_* routines |
| 326 | * would not be useful. In that case, do not trigger aggressive jetsam. |
| 327 | */ |
| 328 | if (total_candidates < kJetsamMinCandidatesThreshold) { |
| 329 | #if DEVELOPMENT || DEBUG |
| 330 | memorystatus_log_info( |
| 331 | "memorystatus: aggressive: [FAILED] Low Candidate Count (current: %d, threshold: %d)\n" , total_candidates, kJetsamMinCandidatesThreshold); |
| 332 | #endif /* DEVELOPMENT || DEBUG */ |
| 333 | aggressive_jetsam_needed = false; |
| 334 | } |
| 335 | return aggressive_jetsam_needed; |
| 336 | } |
| 337 | |
| 338 | static bool |
| 339 | memorystatus_aggressive_jetsam_needed_sysproc_aging(__unused int eval_aggressive_count, __unused int *idle_kills, __unused int idle_kill_candidates, int *total_candidates) |
| 340 | { |
| 341 | bool aggressive_jetsam_needed = false; |
| 342 | |
| 343 | /* |
| 344 | * For the kJetsamAgingPolicySysProcsReclaimedFirst aging policy, we maintain the jetsam |
| 345 | * relaunch behavior for all daemons. Also, daemons and apps are aged in deferred bands on |
| 346 | * every dirty->clean transition. For this aging policy, the best way to determine if |
| 347 | * aggressive jetsam is needed, is to see if the kill candidates are mostly bad candidates. |
| 348 | * If yes, then we need to go to higher bands to reclaim memory. |
| 349 | */ |
| 350 | proc_list_lock(); |
| 351 | /* Get total candidate counts for idle and idle deferred bands */ |
| 352 | *total_candidates = memstat_bucket[JETSAM_PRIORITY_IDLE].count + memstat_bucket[system_procs_aging_band].count; |
| 353 | /* Get counts of bad kill candidates in idle and idle deferred bands */ |
| 354 | int bad_candidates = memstat_bucket[JETSAM_PRIORITY_IDLE].relaunch_high_count + memstat_bucket[system_procs_aging_band].relaunch_high_count; |
| 355 | |
| 356 | proc_list_unlock(); |
| 357 | |
| 358 | /* Check if the number of bad candidates is greater than kJetsamHighRelaunchCandidatesThreshold % */ |
| 359 | aggressive_jetsam_needed = (((bad_candidates * 100) / *total_candidates) >= kJetsamHighRelaunchCandidatesThreshold); |
| 360 | |
| 361 | /* |
| 362 | * Since the new aging policy bases the aggressive jetsam trigger on percentage of |
| 363 | * bad candidates, it is prone to being overly aggressive. In order to mitigate that, |
| 364 | * make sure the system is really under memory pressure before triggering aggressive |
| 365 | * jetsam. |
| 366 | */ |
| 367 | if (memorystatus_available_pages > memorystatus_sysproc_aging_aggr_pages) { |
| 368 | aggressive_jetsam_needed = false; |
| 369 | } |
| 370 | |
| 371 | #if DEVELOPMENT || DEBUG |
| 372 | memorystatus_log_info( |
| 373 | "memorystatus: aggressive%d: [%s] Bad Candidate Threshold Check (total: %d, bad: %d, threshold: %d %%); Memory Pressure Check (available_pgs: %llu, threshold_pgs: %llu)\n" , |
| 374 | eval_aggressive_count, aggressive_jetsam_needed ? "PASSED" : "FAILED" , *total_candidates, bad_candidates, |
| 375 | kJetsamHighRelaunchCandidatesThreshold, (uint64_t)MEMORYSTATUS_LOG_AVAILABLE_PAGES, (uint64_t)memorystatus_sysproc_aging_aggr_pages); |
| 376 | #endif /* DEVELOPMENT || DEBUG */ |
| 377 | return aggressive_jetsam_needed; |
| 378 | } |
| 379 | |
| 380 | #endif /* CONFIG_JETSAM */ |
| 381 | |
| 382 | #pragma mark Freezer |
| 383 | #if CONFIG_FREEZE |
| 384 | /* |
| 385 | * Freezer policies |
| 386 | */ |
| 387 | |
| 388 | /* |
| 389 | * These functions determine what is eligible for the freezer |
| 390 | * and the order that we consider freezing them |
| 391 | */ |
| 392 | |
| 393 | /* |
| 394 | * Checks if the given process is eligible for the freezer. |
| 395 | * Processes can only be frozen if this returns true. |
| 396 | */ |
| 397 | bool |
| 398 | memorystatus_is_process_eligible_for_freeze(proc_t p) |
| 399 | { |
| 400 | /* |
| 401 | * Called with proc_list_lock held. |
| 402 | */ |
| 403 | |
| 404 | LCK_MTX_ASSERT(&proc_list_mlock, LCK_MTX_ASSERT_OWNED); |
| 405 | |
| 406 | bool should_freeze = false; |
| 407 | uint32_t state = 0, pages = 0; |
| 408 | bool first_consideration = true; |
| 409 | task_t task; |
| 410 | |
| 411 | state = p->p_memstat_state; |
| 412 | |
| 413 | if (state & (P_MEMSTAT_TERMINATED | P_MEMSTAT_LOCKED | P_MEMSTAT_FREEZE_DISABLED | P_MEMSTAT_FREEZE_IGNORE)) { |
| 414 | if (state & P_MEMSTAT_FREEZE_DISABLED) { |
| 415 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonDisabled; |
| 416 | } |
| 417 | goto out; |
| 418 | } |
| 419 | |
| 420 | task = proc_task(p); |
| 421 | |
| 422 | if (isSysProc(p)) { |
| 423 | /* |
| 424 | * Daemon:- We consider freezing it if: |
| 425 | * - it belongs to a coalition and the leader is frozen, and, |
| 426 | * - its role in the coalition is XPC service. |
| 427 | * |
| 428 | * We skip memory size requirements in this case. |
| 429 | */ |
| 430 | int task_role_in_coalition = 0; |
| 431 | proc_t leader_proc = memorystatus_get_coalition_leader_and_role(p, &task_role_in_coalition); |
| 432 | if (leader_proc == PROC_NULL || leader_proc == p) { |
| 433 | /* |
| 434 | * Jetsam coalition is leaderless or the leader is not an app. |
| 435 | * Either way, don't freeze this proc. |
| 436 | */ |
| 437 | goto out; |
| 438 | } |
| 439 | |
| 440 | /* Leader must be frozen */ |
| 441 | if (!(leader_proc->p_memstat_state & P_MEMSTAT_FROZEN)) { |
| 442 | goto out; |
| 443 | } |
| 444 | /* Only freeze XPC services */ |
| 445 | if (task_role_in_coalition == COALITION_TASKROLE_XPC) { |
| 446 | should_freeze = true; |
| 447 | } |
| 448 | |
| 449 | goto out; |
| 450 | } else { |
| 451 | /* |
| 452 | * Application. Only freeze if it's suspended. |
| 453 | */ |
| 454 | if (!(state & P_MEMSTAT_SUSPENDED)) { |
| 455 | goto out; |
| 456 | } |
| 457 | } |
| 458 | |
| 459 | /* |
| 460 | * We're interested in tracking what percentage of |
| 461 | * eligible apps actually get frozen. |
| 462 | * To avoid skewing the metrics towards processes which |
| 463 | * are considered more frequently, we only track failures once |
| 464 | * per process. |
| 465 | */ |
| 466 | first_consideration = !(state & P_MEMSTAT_FREEZE_CONSIDERED); |
| 467 | |
| 468 | if (first_consideration) { |
| 469 | memorystatus_freezer_stats.mfs_process_considered_count++; |
| 470 | p->p_memstat_state |= P_MEMSTAT_FREEZE_CONSIDERED; |
| 471 | } |
| 472 | |
| 473 | /* Only freeze applications meeting our minimum resident page criteria */ |
| 474 | memorystatus_get_task_page_counts(proc_task(p), &pages, NULL, NULL); |
| 475 | if (pages < memorystatus_freeze_pages_min) { |
| 476 | if (first_consideration) { |
| 477 | memorystatus_freezer_stats.mfs_error_below_min_pages_count++; |
| 478 | } |
| 479 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonBelowMinPages; |
| 480 | goto out; |
| 481 | } |
| 482 | |
| 483 | /* Don't freeze processes that are already exiting on core. It may have started exiting |
| 484 | * after we chose it for freeze, but before we obtained the proc_list_lock. |
| 485 | * NB: This is only possible if we're coming in from memorystatus_freeze_process_sync. |
| 486 | * memorystatus_freeze_top_process holds the proc_list_lock while it traverses the bands. |
| 487 | */ |
| 488 | if (proc_list_exited(p)) { |
| 489 | if (first_consideration) { |
| 490 | memorystatus_freezer_stats.mfs_error_other_count++; |
| 491 | } |
| 492 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonOther; |
| 493 | goto out; |
| 494 | } |
| 495 | |
| 496 | if (!memorystatus_freezer_use_ordered_list) { |
| 497 | /* |
| 498 | * We're not using the ordered list so we need to check |
| 499 | * that dasd recommended the process. Note that the ordered list |
| 500 | * algorithm only considers processes on the list in the first place |
| 501 | * so there's no need to double check here. |
| 502 | */ |
| 503 | if (!memorystatus_freeze_process_is_recommended(p)) { |
| 504 | if (first_consideration) { |
| 505 | memorystatus_freezer_stats.mfs_error_low_probability_of_use_count++; |
| 506 | } |
| 507 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonLowProbOfUse; |
| 508 | goto out; |
| 509 | } |
| 510 | } |
| 511 | |
| 512 | if (!(state & P_MEMSTAT_FROZEN) && p->p_memstat_effectivepriority > memorystatus_freeze_max_candidate_band) { |
| 513 | /* |
| 514 | * Proc has been elevated by something else. |
| 515 | * Don't freeze it. |
| 516 | */ |
| 517 | if (first_consideration) { |
| 518 | memorystatus_freezer_stats.mfs_error_elevated_count++; |
| 519 | } |
| 520 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonElevated; |
| 521 | goto out; |
| 522 | } |
| 523 | |
| 524 | should_freeze = true; |
| 525 | out: |
| 526 | if (should_freeze && !(state & P_MEMSTAT_FROZEN)) { |
| 527 | /* |
| 528 | * Reset the skip reason. If it's killed before we manage to actually freeze it |
| 529 | * we failed to consider it early enough. |
| 530 | */ |
| 531 | p->p_memstat_freeze_skip_reason = kMemorystatusFreezeSkipReasonNone; |
| 532 | if (!first_consideration) { |
| 533 | /* |
| 534 | * We're freezing this for the first time and we previously considered it ineligible. |
| 535 | * Bump the considered count so that we track this as 1 failure |
| 536 | * and 1 success. |
| 537 | */ |
| 538 | memorystatus_freezer_stats.mfs_process_considered_count++; |
| 539 | } |
| 540 | } |
| 541 | return should_freeze; |
| 542 | } |
| 543 | |
| 544 | bool |
| 545 | memorystatus_freeze_proc_is_refreeze_eligible(proc_t p) |
| 546 | { |
| 547 | return (p->p_memstat_state & P_MEMSTAT_REFREEZE_ELIGIBLE) != 0; |
| 548 | } |
| 549 | |
| 550 | |
| 551 | static proc_t |
| 552 | memorystatus_freeze_pick_refreeze_process(proc_t last_p) |
| 553 | { |
| 554 | proc_t p = PROC_NULL, next_p = PROC_NULL; |
| 555 | unsigned int band = (unsigned int) memorystatus_freeze_jetsam_band; |
| 556 | if (last_p == PROC_NULL) { |
| 557 | next_p = memorystatus_get_first_proc_locked(&band, FALSE); |
| 558 | } else { |
| 559 | next_p = memorystatus_get_next_proc_locked(&band, last_p, FALSE); |
| 560 | } |
| 561 | while (next_p) { |
| 562 | p = next_p; |
| 563 | next_p = memorystatus_get_next_proc_locked(&band, p, FALSE); |
| 564 | if ((p->p_memstat_state & P_MEMSTAT_FROZEN) && !memorystatus_freeze_proc_is_refreeze_eligible(p)) { |
| 565 | /* Process is already frozen & hasn't been thawed. */ |
| 566 | continue; |
| 567 | } |
| 568 | /* |
| 569 | * Has to have been frozen once before. |
| 570 | */ |
| 571 | if (!(p->p_memstat_state & P_MEMSTAT_FROZEN)) { |
| 572 | continue; |
| 573 | } |
| 574 | |
| 575 | /* |
| 576 | * Not currently being looked at for something. |
| 577 | */ |
| 578 | if (p->p_memstat_state & P_MEMSTAT_LOCKED) { |
| 579 | continue; |
| 580 | } |
| 581 | |
| 582 | #if FREEZE_PREVENT_REFREEZE_OF_LAST_THAWED |
| 583 | /* |
| 584 | * Don't refreeze the last process we just thawed if still within the timeout window |
| 585 | */ |
| 586 | if (p->p_pid == memorystatus_freeze_last_pid_thawed) { |
| 587 | uint64_t timeout_delta_abs; |
| 588 | nanoseconds_to_absolutetime(FREEZE_PREVENT_REFREEZE_OF_LAST_THAWED_TIMEOUT_SECONDS * NSEC_PER_SEC, &timeout_delta_abs); |
| 589 | if (mach_absolute_time() < (memorystatus_freeze_last_pid_thawed_ts + timeout_delta_abs)) { |
| 590 | continue; |
| 591 | } |
| 592 | } |
| 593 | #endif |
| 594 | |
| 595 | /* |
| 596 | * Found it |
| 597 | */ |
| 598 | return p; |
| 599 | } |
| 600 | return PROC_NULL; |
| 601 | } |
| 602 | |
| 603 | proc_t |
| 604 | memorystatus_freeze_pick_process(struct memorystatus_freeze_list_iterator *iterator) |
| 605 | { |
| 606 | proc_t p = PROC_NULL, next_p = PROC_NULL; |
| 607 | unsigned int band = JETSAM_PRIORITY_IDLE; |
| 608 | |
| 609 | LCK_MTX_ASSERT(&proc_list_mlock, LCK_MTX_ASSERT_OWNED); |
| 610 | /* |
| 611 | * If the freezer is full, only consider refreezes. |
| 612 | */ |
| 613 | if (iterator->refreeze_only || memorystatus_frozen_count >= memorystatus_frozen_processes_max) { |
| 614 | if (!iterator->refreeze_only) { |
| 615 | /* |
| 616 | * The first time the iterator starts to return refreeze |
| 617 | * candidates, we need to reset the last pointer b/c it's pointing into the wrong band. |
| 618 | */ |
| 619 | iterator->last_p = PROC_NULL; |
| 620 | iterator->refreeze_only = true; |
| 621 | } |
| 622 | iterator->last_p = memorystatus_freeze_pick_refreeze_process(iterator->last_p); |
| 623 | return iterator->last_p; |
| 624 | } |
| 625 | |
| 626 | /* |
| 627 | * Search for the next freezer candidate. |
| 628 | */ |
| 629 | if (memorystatus_freezer_use_ordered_list) { |
| 630 | while (iterator->global_freeze_list_index < memorystatus_global_freeze_list.mfcl_length) { |
| 631 | p = memorystatus_freezer_candidate_list_get_proc( |
| 632 | &memorystatus_global_freeze_list, |
| 633 | (iterator->global_freeze_list_index)++, |
| 634 | &memorystatus_freezer_stats.mfs_freeze_pid_mismatches); |
| 635 | |
| 636 | if (p != PROC_NULL && memorystatus_is_process_eligible_for_freeze(p)) { |
| 637 | iterator->last_p = p; |
| 638 | return iterator->last_p; |
| 639 | } |
| 640 | } |
| 641 | } else { |
| 642 | if (iterator->last_p == PROC_NULL) { |
| 643 | next_p = memorystatus_get_first_proc_locked(&band, FALSE); |
| 644 | } else { |
| 645 | next_p = memorystatus_get_next_proc_locked(&band, iterator->last_p, FALSE); |
| 646 | } |
| 647 | while (next_p) { |
| 648 | p = next_p; |
| 649 | if (memorystatus_is_process_eligible_for_freeze(p)) { |
| 650 | iterator->last_p = p; |
| 651 | return iterator->last_p; |
| 652 | } else { |
| 653 | next_p = memorystatus_get_next_proc_locked(&band, p, FALSE); |
| 654 | } |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | /* |
| 659 | * Failed to find a new freezer candidate. |
| 660 | * Try to re-freeze. |
| 661 | */ |
| 662 | if (memorystatus_refreeze_eligible_count >= memorystatus_min_thaw_refreeze_threshold) { |
| 663 | assert(!iterator->refreeze_only); |
| 664 | iterator->refreeze_only = true; |
| 665 | iterator->last_p = memorystatus_freeze_pick_refreeze_process(PROC_NULL); |
| 666 | return iterator->last_p; |
| 667 | } |
| 668 | return PROC_NULL; |
| 669 | } |
| 670 | |
| 671 | /* |
| 672 | * memorystatus_pages_update calls this function whenever the number |
| 673 | * of available pages changes. It wakes the freezer thread iff the function returns |
| 674 | * true. The freezer thread will try to freeze (or refreeze) up to 1 process |
| 675 | * before blocking again. |
| 676 | * |
| 677 | * Note the freezer thread is also woken up by memorystatus_on_inactivity. |
| 678 | */ |
| 679 | |
| 680 | bool |
| 681 | memorystatus_freeze_thread_should_run() |
| 682 | { |
| 683 | /* |
| 684 | * No freezer_mutex held here...see why near call-site |
| 685 | * within memorystatus_pages_update(). |
| 686 | */ |
| 687 | |
| 688 | if (memorystatus_freeze_enabled == false) { |
| 689 | return false; |
| 690 | } |
| 691 | |
| 692 | if (memorystatus_available_pages > memorystatus_freeze_threshold) { |
| 693 | return false; |
| 694 | } |
| 695 | |
| 696 | memorystatus_freezer_stats.mfs_below_threshold_count++; |
| 697 | |
| 698 | if ((memorystatus_frozen_count >= memorystatus_frozen_processes_max)) { |
| 699 | /* |
| 700 | * Consider this as a skip even if we wake up to refreeze because |
| 701 | * we won't freeze any new procs. |
| 702 | */ |
| 703 | memorystatus_freezer_stats.mfs_skipped_full_count++; |
| 704 | if (memorystatus_refreeze_eligible_count < memorystatus_min_thaw_refreeze_threshold) { |
| 705 | return false; |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | if (memorystatus_frozen_shared_mb_max && (memorystatus_frozen_shared_mb >= memorystatus_frozen_shared_mb_max)) { |
| 710 | memorystatus_freezer_stats.mfs_skipped_shared_mb_high_count++; |
| 711 | return false; |
| 712 | } |
| 713 | |
| 714 | uint64_t curr_time = mach_absolute_time(); |
| 715 | |
| 716 | if (curr_time < memorystatus_freezer_thread_next_run_ts) { |
| 717 | return false; |
| 718 | } |
| 719 | |
| 720 | return true; |
| 721 | } |
| 722 | |
| 723 | size_t |
| 724 | memorystatus_pick_freeze_count_for_wakeup() |
| 725 | { |
| 726 | size_t num_to_freeze = 0; |
| 727 | if (!memorystatus_swap_all_apps) { |
| 728 | num_to_freeze = 1; |
| 729 | } else { |
| 730 | /* |
| 731 | * When app swap is enabled, we want the freezer thread to aggressively freeze |
| 732 | * all candidates so we clear out space for the fg working set. |
| 733 | * But we still cap it to the current size of the candidate bands to avoid |
| 734 | * consuming excessive CPU if there's a lot of churn in the candidate band. |
| 735 | */ |
| 736 | proc_list_lock(); |
| 737 | for (unsigned int band = JETSAM_PRIORITY_IDLE; band <= memorystatus_freeze_max_candidate_band; band++) { |
| 738 | num_to_freeze += memstat_bucket[band].count; |
| 739 | } |
| 740 | proc_list_unlock(); |
| 741 | } |
| 742 | |
| 743 | return num_to_freeze; |
| 744 | } |
| 745 | |
| 746 | #endif /* CONFIG_FREEZE */ |
| 747 | |