1 /* $NetBSD: uvm_pdaemon.c,v 1.123 2020/01/15 17:55:45 ad Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_pageout.c 8.5 (Berkeley) 2/14/94 37 * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 /* 65 * uvm_pdaemon.c: the page daemon 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: uvm_pdaemon.c,v 1.123 2020/01/15 17:55:45 ad Exp $"); 70 71 #include "opt_uvmhist.h" 72 #include "opt_readahead.h" 73 74 #include <sys/param.h> 75 #include <sys/proc.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/pool.h> 79 #include <sys/buf.h> 80 #include <sys/module.h> 81 #include <sys/atomic.h> 82 #include <sys/kthread.h> 83 84 #include <uvm/uvm.h> 85 #include <uvm/uvm_pdpolicy.h> 86 #include <uvm/uvm_pgflcache.h> 87 88 #ifdef UVMHIST 89 UVMHIST_DEFINE(pdhist); 90 #endif 91 92 /* 93 * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate 94 * in a pass thru the inactive list when swap is full. the value should be 95 * "small"... if it's too large we'll cycle the active pages thru the inactive 96 * queue too quickly to for them to be referenced and avoid being freed. 97 */ 98 99 #define UVMPD_NUMDIRTYREACTS 16 100 101 #define UVMPD_NUMTRYLOCKOWNER 128 102 103 /* 104 * local prototypes 105 */ 106 107 static void uvmpd_scan(void); 108 static void uvmpd_scan_queue(void); 109 static void uvmpd_tune(void); 110 static void uvmpd_pool_drain_thread(void *); 111 static void uvmpd_pool_drain_wakeup(void); 112 113 static unsigned int uvm_pagedaemon_waiters; 114 115 /* State for the pool drainer thread */ 116 static kmutex_t uvmpd_lock __cacheline_aligned; 117 static kcondvar_t uvmpd_pool_drain_cv; 118 static bool uvmpd_pool_drain_run = false; 119 120 /* 121 * XXX hack to avoid hangs when large processes fork. 122 */ 123 u_int uvm_extrapages; 124 125 /* 126 * uvm_wait: wait (sleep) for the page daemon to free some pages 127 * 128 * => should be called with all locks released 129 * => should _not_ be called by the page daemon (to avoid deadlock) 130 */ 131 132 void 133 uvm_wait(const char *wmsg) 134 { 135 int timo = 0; 136 137 if (uvm.pagedaemon_lwp == NULL) 138 panic("out of memory before the pagedaemon thread exists"); 139 140 mutex_spin_enter(&uvmpd_lock); 141 142 /* 143 * check for page daemon going to sleep (waiting for itself) 144 */ 145 146 if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) { 147 /* 148 * now we have a problem: the pagedaemon wants to go to 149 * sleep until it frees more memory. but how can it 150 * free more memory if it is asleep? that is a deadlock. 151 * we have two options: 152 * [1] panic now 153 * [2] put a timeout on the sleep, thus causing the 154 * pagedaemon to only pause (rather than sleep forever) 155 * 156 * note that option [2] will only help us if we get lucky 157 * and some other process on the system breaks the deadlock 158 * by exiting or freeing memory (thus allowing the pagedaemon 159 * to continue). for now we panic if DEBUG is defined, 160 * otherwise we hope for the best with option [2] (better 161 * yet, this should never happen in the first place!). 162 */ 163 164 printf("pagedaemon: deadlock detected!\n"); 165 timo = hz >> 3; /* set timeout */ 166 #if defined(DEBUG) 167 /* DEBUG: panic so we can debug it */ 168 panic("pagedaemon deadlock"); 169 #endif 170 } 171 172 uvm_pagedaemon_waiters++; 173 wakeup(&uvm.pagedaemon); /* wake the daemon! */ 174 UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvmpd_lock, false, wmsg, timo); 175 } 176 177 /* 178 * uvm_kick_pdaemon: perform checks to determine if we need to 179 * give the pagedaemon a nudge, and do so if necessary. 180 */ 181 182 void 183 uvm_kick_pdaemon(void) 184 { 185 int fpages = uvm_availmem(); 186 187 if (fpages + uvmexp.paging < uvmexp.freemin || 188 (fpages + uvmexp.paging < uvmexp.freetarg && 189 uvmpdpol_needsscan_p()) || 190 uvm_km_va_starved_p()) { 191 mutex_spin_enter(&uvmpd_lock); 192 wakeup(&uvm.pagedaemon); 193 mutex_spin_exit(&uvmpd_lock); 194 } 195 } 196 197 /* 198 * uvmpd_tune: tune paging parameters 199 * 200 * => called when ever memory is added (or removed?) to the system 201 */ 202 203 static void 204 uvmpd_tune(void) 205 { 206 int val; 207 208 UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist); 209 210 /* 211 * try to keep 0.5% of available RAM free, but limit to between 212 * 128k and 1024k per-CPU. XXX: what are these values good for? 213 */ 214 val = uvmexp.npages / 200; 215 val = MAX(val, (128*1024) >> PAGE_SHIFT); 216 val = MIN(val, (1024*1024) >> PAGE_SHIFT); 217 val *= ncpu; 218 219 /* Make sure there's always a user page free. */ 220 if (val < uvmexp.reserve_kernel + 1) 221 val = uvmexp.reserve_kernel + 1; 222 uvmexp.freemin = val; 223 224 /* Calculate free target. */ 225 val = (uvmexp.freemin * 4) / 3; 226 if (val <= uvmexp.freemin) 227 val = uvmexp.freemin + 1; 228 uvmexp.freetarg = val + atomic_swap_uint(&uvm_extrapages, 0); 229 230 uvmexp.wiredmax = uvmexp.npages / 3; 231 UVMHIST_LOG(pdhist, "<- done, freemin=%jd, freetarg=%jd, wiredmax=%jd", 232 uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0); 233 } 234 235 /* 236 * uvm_pageout: the main loop for the pagedaemon 237 */ 238 239 void 240 uvm_pageout(void *arg) 241 { 242 int npages = 0; 243 int extrapages = 0; 244 int fpages; 245 246 UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist); 247 248 UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0); 249 250 mutex_init(&uvmpd_lock, MUTEX_DEFAULT, IPL_VM); 251 cv_init(&uvmpd_pool_drain_cv, "pooldrain"); 252 253 /* Create the pool drainer kernel thread. */ 254 if (kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL, 255 uvmpd_pool_drain_thread, NULL, NULL, "pooldrain")) 256 panic("fork pooldrain"); 257 258 /* 259 * ensure correct priority and set paging parameters... 260 */ 261 262 uvm.pagedaemon_lwp = curlwp; 263 npages = uvmexp.npages; 264 uvmpd_tune(); 265 266 /* 267 * main loop 268 */ 269 270 for (;;) { 271 bool needsscan, needsfree, kmem_va_starved; 272 273 kmem_va_starved = uvm_km_va_starved_p(); 274 275 mutex_spin_enter(&uvmpd_lock); 276 if ((uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) && 277 !kmem_va_starved) { 278 UVMHIST_LOG(pdhist," <<SLEEPING>>",0,0,0,0); 279 UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon, 280 &uvmpd_lock, false, "pgdaemon", 0); 281 uvmexp.pdwoke++; 282 UVMHIST_LOG(pdhist," <<WOKE UP>>",0,0,0,0); 283 } else { 284 mutex_spin_exit(&uvmpd_lock); 285 } 286 287 /* 288 * now recompute inactive count 289 */ 290 291 if (npages != uvmexp.npages || extrapages != uvm_extrapages) { 292 npages = uvmexp.npages; 293 extrapages = uvm_extrapages; 294 uvmpd_tune(); 295 } 296 297 uvmpdpol_tune(); 298 299 /* 300 * Estimate a hint. Note that bufmem are returned to 301 * system only when entire pool page is empty. 302 */ 303 fpages = uvm_availmem(); 304 UVMHIST_LOG(pdhist," free/ftarg=%jd/%jd", 305 fpages, uvmexp.freetarg, 0,0); 306 307 needsfree = fpages + uvmexp.paging < uvmexp.freetarg; 308 needsscan = needsfree || uvmpdpol_needsscan_p(); 309 310 /* 311 * scan if needed 312 */ 313 if (needsscan) { 314 uvmpd_scan(); 315 } 316 317 /* 318 * if there's any free memory to be had, 319 * wake up any waiters. 320 */ 321 if (uvm_availmem() > uvmexp.reserve_kernel || 322 uvmexp.paging == 0) { 323 mutex_spin_enter(&uvmpd_lock); 324 wakeup(&uvmexp.free); 325 uvm_pagedaemon_waiters = 0; 326 mutex_spin_exit(&uvmpd_lock); 327 } 328 329 /* 330 * scan done. if we don't need free memory, we're done. 331 */ 332 333 if (!needsfree && !kmem_va_starved) 334 continue; 335 336 /* 337 * kick the pool drainer thread. 338 */ 339 340 uvmpd_pool_drain_wakeup(); 341 } 342 /*NOTREACHED*/ 343 } 344 345 346 /* 347 * uvm_aiodone_worker: a workqueue callback for the aiodone daemon. 348 */ 349 350 void 351 uvm_aiodone_worker(struct work *wk, void *dummy) 352 { 353 struct buf *bp = (void *)wk; 354 355 KASSERT(&bp->b_work == wk); 356 357 /* 358 * process an i/o that's done. 359 */ 360 361 (*bp->b_iodone)(bp); 362 } 363 364 void 365 uvm_pageout_start(int npages) 366 { 367 368 atomic_add_int(&uvmexp.paging, npages); 369 } 370 371 void 372 uvm_pageout_done(int npages) 373 { 374 375 KASSERT(uvmexp.paging >= npages); 376 atomic_add_int(&uvmexp.paging, -npages); 377 378 /* 379 * wake up either of pagedaemon or LWPs waiting for it. 380 */ 381 382 mutex_spin_enter(&uvmpd_lock); 383 if (uvm_availmem() <= uvmexp.reserve_kernel) { 384 wakeup(&uvm.pagedaemon); 385 } else if (uvm_pagedaemon_waiters != 0) { 386 wakeup(&uvmexp.free); 387 uvm_pagedaemon_waiters = 0; 388 } 389 mutex_spin_exit(&uvmpd_lock); 390 } 391 392 /* 393 * uvmpd_trylockowner: trylock the page's owner. 394 * 395 * => called with page interlock held. 396 * => resolve orphaned O->A loaned page. 397 * => return the locked mutex on success. otherwise, return NULL. 398 */ 399 400 kmutex_t * 401 uvmpd_trylockowner(struct vm_page *pg) 402 { 403 struct uvm_object *uobj = pg->uobject; 404 struct vm_anon *anon = pg->uanon; 405 int tries, count; 406 bool running; 407 kmutex_t *slock; 408 409 KASSERT(mutex_owned(&pg->interlock)); 410 411 if (uobj != NULL) { 412 slock = uobj->vmobjlock; 413 KASSERTMSG(slock != NULL, "pg %p uobj %p, NULL lock", pg, uobj); 414 } else if (anon != NULL) { 415 slock = anon->an_lock; 416 KASSERTMSG(slock != NULL, "pg %p anon %p, NULL lock", pg, anon); 417 } else { 418 /* Page may be in state of flux - ignore. */ 419 mutex_exit(&pg->interlock); 420 return NULL; 421 } 422 423 /* 424 * Now try to lock the objects. We'll try hard, but don't really 425 * plan on spending more than a millisecond or so here. 426 */ 427 tries = (curlwp == uvm.pagedaemon_lwp ? UVMPD_NUMTRYLOCKOWNER : 1); 428 for (;;) { 429 if (mutex_tryenter(slock)) { 430 if (uobj == NULL) { 431 /* 432 * set PG_ANON if it isn't set already. 433 */ 434 if ((pg->flags & PG_ANON) == 0) { 435 KASSERT(pg->loan_count > 0); 436 pg->loan_count--; 437 pg->flags |= PG_ANON; 438 /* anon now owns it */ 439 } 440 } 441 mutex_exit(&pg->interlock); 442 return slock; 443 } 444 running = mutex_owner_running(slock); 445 if (!running || --tries <= 0) { 446 break; 447 } 448 count = SPINLOCK_BACKOFF_MAX; 449 SPINLOCK_BACKOFF(count); 450 } 451 452 /* 453 * We didn't get the lock; chances are the very next page on the 454 * queue also has the same lock, so if the lock owner is not running 455 * take a breather and allow them to make progress. There could be 456 * only 1 CPU in the system, or the pagedaemon could have preempted 457 * the owner in kernel, or any number of other things could be going 458 * on. 459 */ 460 mutex_exit(&pg->interlock); 461 if (curlwp == uvm.pagedaemon_lwp) { 462 if (!running) { 463 (void)kpause("pdpglock", false, 1, NULL); 464 } 465 uvmexp.pdbusy++; 466 } 467 return NULL; 468 } 469 470 #if defined(VMSWAP) 471 struct swapcluster { 472 int swc_slot; 473 int swc_nallocated; 474 int swc_nused; 475 struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)]; 476 }; 477 478 static void 479 swapcluster_init(struct swapcluster *swc) 480 { 481 482 swc->swc_slot = 0; 483 swc->swc_nused = 0; 484 } 485 486 static int 487 swapcluster_allocslots(struct swapcluster *swc) 488 { 489 int slot; 490 int npages; 491 492 if (swc->swc_slot != 0) { 493 return 0; 494 } 495 496 /* Even with strange MAXPHYS, the shift 497 implicitly rounds down to a page. */ 498 npages = MAXPHYS >> PAGE_SHIFT; 499 slot = uvm_swap_alloc(&npages, true); 500 if (slot == 0) { 501 return ENOMEM; 502 } 503 swc->swc_slot = slot; 504 swc->swc_nallocated = npages; 505 swc->swc_nused = 0; 506 507 return 0; 508 } 509 510 static int 511 swapcluster_add(struct swapcluster *swc, struct vm_page *pg) 512 { 513 int slot; 514 struct uvm_object *uobj; 515 516 KASSERT(swc->swc_slot != 0); 517 KASSERT(swc->swc_nused < swc->swc_nallocated); 518 KASSERT((pg->flags & PG_SWAPBACKED) != 0); 519 520 slot = swc->swc_slot + swc->swc_nused; 521 uobj = pg->uobject; 522 if (uobj == NULL) { 523 KASSERT(mutex_owned(pg->uanon->an_lock)); 524 pg->uanon->an_swslot = slot; 525 } else { 526 int result; 527 528 KASSERT(mutex_owned(uobj->vmobjlock)); 529 result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot); 530 if (result == -1) { 531 return ENOMEM; 532 } 533 } 534 swc->swc_pages[swc->swc_nused] = pg; 535 swc->swc_nused++; 536 537 return 0; 538 } 539 540 static void 541 swapcluster_flush(struct swapcluster *swc, bool now) 542 { 543 int slot; 544 int nused; 545 int nallocated; 546 int error __diagused; 547 548 if (swc->swc_slot == 0) { 549 return; 550 } 551 KASSERT(swc->swc_nused <= swc->swc_nallocated); 552 553 slot = swc->swc_slot; 554 nused = swc->swc_nused; 555 nallocated = swc->swc_nallocated; 556 557 /* 558 * if this is the final pageout we could have a few 559 * unused swap blocks. if so, free them now. 560 */ 561 562 if (nused < nallocated) { 563 if (!now) { 564 return; 565 } 566 uvm_swap_free(slot + nused, nallocated - nused); 567 } 568 569 /* 570 * now start the pageout. 571 */ 572 573 if (nused > 0) { 574 uvmexp.pdpageouts++; 575 uvm_pageout_start(nused); 576 error = uvm_swap_put(slot, swc->swc_pages, nused, 0); 577 KASSERT(error == 0 || error == ENOMEM); 578 } 579 580 /* 581 * zero swslot to indicate that we are 582 * no longer building a swap-backed cluster. 583 */ 584 585 swc->swc_slot = 0; 586 swc->swc_nused = 0; 587 } 588 589 static int 590 swapcluster_nused(struct swapcluster *swc) 591 { 592 593 return swc->swc_nused; 594 } 595 596 /* 597 * uvmpd_dropswap: free any swap allocated to this page. 598 * 599 * => called with owner locked. 600 * => return true if a page had an associated slot. 601 */ 602 603 bool 604 uvmpd_dropswap(struct vm_page *pg) 605 { 606 bool result = false; 607 struct vm_anon *anon = pg->uanon; 608 609 if ((pg->flags & PG_ANON) && anon->an_swslot) { 610 uvm_swap_free(anon->an_swslot, 1); 611 anon->an_swslot = 0; 612 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY); 613 result = true; 614 } else if (pg->flags & PG_AOBJ) { 615 int slot = uao_set_swslot(pg->uobject, 616 pg->offset >> PAGE_SHIFT, 0); 617 if (slot) { 618 uvm_swap_free(slot, 1); 619 uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY); 620 result = true; 621 } 622 } 623 624 return result; 625 } 626 627 #endif /* defined(VMSWAP) */ 628 629 /* 630 * uvmpd_scan_queue: scan an replace candidate list for pages 631 * to clean or free. 632 * 633 * => we work on meeting our free target by converting inactive pages 634 * into free pages. 635 * => we handle the building of swap-backed clusters 636 */ 637 638 static void 639 uvmpd_scan_queue(void) 640 { 641 struct vm_page *p; 642 struct uvm_object *uobj; 643 struct vm_anon *anon; 644 #if defined(VMSWAP) 645 struct swapcluster swc; 646 #endif /* defined(VMSWAP) */ 647 int dirtyreacts; 648 kmutex_t *slock; 649 UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist); 650 651 /* 652 * swslot is non-zero if we are building a swap cluster. we want 653 * to stay in the loop while we have a page to scan or we have 654 * a swap-cluster to build. 655 */ 656 657 #if defined(VMSWAP) 658 swapcluster_init(&swc); 659 #endif /* defined(VMSWAP) */ 660 661 dirtyreacts = 0; 662 uvmpdpol_scaninit(); 663 664 while (/* CONSTCOND */ 1) { 665 666 /* 667 * see if we've met the free target. 668 */ 669 670 if (uvm_availmem() + uvmexp.paging 671 #if defined(VMSWAP) 672 + swapcluster_nused(&swc) 673 #endif /* defined(VMSWAP) */ 674 >= uvmexp.freetarg << 2 || 675 dirtyreacts == UVMPD_NUMDIRTYREACTS) { 676 UVMHIST_LOG(pdhist," met free target: " 677 "exit loop", 0, 0, 0, 0); 678 break; 679 } 680 681 /* 682 * first we have the pdpolicy select a victim page 683 * and attempt to lock the object that the page 684 * belongs to. if our attempt fails we skip on to 685 * the next page (no harm done). it is important to 686 * "try" locking the object as we are locking in the 687 * wrong order (pageq -> object) and we don't want to 688 * deadlock. 689 * 690 * the only time we expect to see an ownerless page 691 * (i.e. a page with no uobject and !PG_ANON) is if an 692 * anon has loaned a page from a uvm_object and the 693 * uvm_object has dropped the ownership. in that 694 * case, the anon can "take over" the loaned page 695 * and make it its own. 696 */ 697 698 p = uvmpdpol_selectvictim(&slock); 699 if (p == NULL) { 700 break; 701 } 702 KASSERT(uvmpdpol_pageisqueued_p(p)); 703 KASSERT(uvm_page_owner_locked_p(p)); 704 KASSERT(p->wire_count == 0); 705 706 /* 707 * we are below target and have a new page to consider. 708 */ 709 710 anon = p->uanon; 711 uobj = p->uobject; 712 713 if (p->flags & PG_BUSY) { 714 mutex_exit(slock); 715 uvmexp.pdbusy++; 716 continue; 717 } 718 719 /* does the page belong to an object? */ 720 if (uobj != NULL) { 721 uvmexp.pdobscan++; 722 } else { 723 #if defined(VMSWAP) 724 KASSERT(anon != NULL); 725 uvmexp.pdanscan++; 726 #else /* defined(VMSWAP) */ 727 panic("%s: anon", __func__); 728 #endif /* defined(VMSWAP) */ 729 } 730 731 732 /* 733 * we now have the object locked. 734 * if the page is not swap-backed, call the object's 735 * pager to flush and free the page. 736 */ 737 738 #if defined(READAHEAD_STATS) 739 if ((p->flags & PG_READAHEAD) != 0) { 740 p->flags &= ~PG_READAHEAD; 741 uvm_ra_miss.ev_count++; 742 } 743 #endif /* defined(READAHEAD_STATS) */ 744 745 if ((p->flags & PG_SWAPBACKED) == 0) { 746 KASSERT(uobj != NULL); 747 (void) (uobj->pgops->pgo_put)(uobj, p->offset, 748 p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE); 749 continue; 750 } 751 752 /* 753 * the page is swap-backed. remove all the permissions 754 * from the page so we can sync the modified info 755 * without any race conditions. if the page is clean 756 * we can free it now and continue. 757 */ 758 759 pmap_page_protect(p, VM_PROT_NONE); 760 if (uvm_pagegetdirty(p) == UVM_PAGE_STATUS_UNKNOWN) { 761 if (pmap_clear_modify(p)) { 762 uvm_pagemarkdirty(p, UVM_PAGE_STATUS_DIRTY); 763 } else { 764 uvm_pagemarkdirty(p, UVM_PAGE_STATUS_CLEAN); 765 } 766 } 767 if (uvm_pagegetdirty(p) != UVM_PAGE_STATUS_DIRTY) { 768 int slot; 769 int pageidx; 770 771 pageidx = p->offset >> PAGE_SHIFT; 772 uvm_pagefree(p); 773 atomic_inc_uint(&uvmexp.pdfreed); 774 775 /* 776 * for anons, we need to remove the page 777 * from the anon ourselves. for aobjs, 778 * pagefree did that for us. 779 */ 780 781 if (anon) { 782 KASSERT(anon->an_swslot != 0); 783 anon->an_page = NULL; 784 slot = anon->an_swslot; 785 } else { 786 slot = uao_find_swslot(uobj, pageidx); 787 } 788 if (slot > 0) { 789 /* this page is now only in swap. */ 790 KASSERT(uvmexp.swpgonly < uvmexp.swpginuse); 791 atomic_inc_uint(&uvmexp.swpgonly); 792 } 793 mutex_exit(slock); 794 continue; 795 } 796 797 #if defined(VMSWAP) 798 /* 799 * this page is dirty, skip it if we'll have met our 800 * free target when all the current pageouts complete. 801 */ 802 803 if (uvm_availmem() + uvmexp.paging > uvmexp.freetarg << 2) { 804 mutex_exit(slock); 805 continue; 806 } 807 808 /* 809 * free any swap space allocated to the page since 810 * we'll have to write it again with its new data. 811 */ 812 813 uvmpd_dropswap(p); 814 815 /* 816 * start new swap pageout cluster (if necessary). 817 * 818 * if swap is full reactivate this page so that 819 * we eventually cycle all pages through the 820 * inactive queue. 821 */ 822 823 if (swapcluster_allocslots(&swc)) { 824 dirtyreacts++; 825 uvm_pagelock(p); 826 uvm_pageactivate(p); 827 uvm_pageunlock(p); 828 mutex_exit(slock); 829 continue; 830 } 831 832 /* 833 * at this point, we're definitely going reuse this 834 * page. mark the page busy and delayed-free. 835 * we should remove the page from the page queues 836 * so we don't ever look at it again. 837 * adjust counters and such. 838 */ 839 840 p->flags |= PG_BUSY; 841 UVM_PAGE_OWN(p, "scan_queue"); 842 p->flags |= PG_PAGEOUT; 843 uvmexp.pgswapout++; 844 845 uvm_pagelock(p); 846 uvm_pagedequeue(p); 847 uvm_pageunlock(p); 848 849 /* 850 * add the new page to the cluster. 851 */ 852 853 if (swapcluster_add(&swc, p)) { 854 p->flags &= ~(PG_BUSY|PG_PAGEOUT); 855 UVM_PAGE_OWN(p, NULL); 856 dirtyreacts++; 857 uvm_pagelock(p); 858 uvm_pageactivate(p); 859 uvm_pageunlock(p); 860 mutex_exit(slock); 861 continue; 862 } 863 mutex_exit(slock); 864 865 swapcluster_flush(&swc, false); 866 867 /* 868 * the pageout is in progress. bump counters and set up 869 * for the next loop. 870 */ 871 872 atomic_inc_uint(&uvmexp.pdpending); 873 874 #else /* defined(VMSWAP) */ 875 uvm_pagelock(p); 876 uvm_pageactivate(p); 877 uvm_pageunlock(p); 878 mutex_exit(slock); 879 #endif /* defined(VMSWAP) */ 880 } 881 882 uvmpdpol_scanfini(); 883 884 #if defined(VMSWAP) 885 swapcluster_flush(&swc, true); 886 #endif /* defined(VMSWAP) */ 887 } 888 889 /* 890 * uvmpd_scan: scan the page queues and attempt to meet our targets. 891 */ 892 893 static void 894 uvmpd_scan(void) 895 { 896 int swap_shortage, pages_freed, fpages; 897 UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist); 898 899 uvmexp.pdrevs++; 900 901 /* 902 * work on meeting our targets. first we work on our free target 903 * by converting inactive pages into free pages. then we work on 904 * meeting our inactive target by converting active pages to 905 * inactive ones. 906 */ 907 908 UVMHIST_LOG(pdhist, " starting 'free' loop",0,0,0,0); 909 910 pages_freed = uvmexp.pdfreed; 911 uvmpd_scan_queue(); 912 pages_freed = uvmexp.pdfreed - pages_freed; 913 914 /* 915 * detect if we're not going to be able to page anything out 916 * until we free some swap resources from active pages. 917 */ 918 919 swap_shortage = 0; 920 fpages = uvm_availmem(); 921 if (fpages < uvmexp.freetarg && 922 uvmexp.swpginuse >= uvmexp.swpgavail && 923 !uvm_swapisfull() && 924 pages_freed == 0) { 925 swap_shortage = uvmexp.freetarg - fpages; 926 } 927 928 uvmpdpol_balancequeue(swap_shortage); 929 930 /* 931 * if still below the minimum target, try unloading kernel 932 * modules. 933 */ 934 935 if (uvm_availmem() < uvmexp.freemin) { 936 module_thread_kick(); 937 } 938 } 939 940 /* 941 * uvm_reclaimable: decide whether to wait for pagedaemon. 942 * 943 * => return true if it seems to be worth to do uvm_wait. 944 * 945 * XXX should be tunable. 946 * XXX should consider pools, etc? 947 */ 948 949 bool 950 uvm_reclaimable(void) 951 { 952 int filepages; 953 int active, inactive; 954 955 /* 956 * if swap is not full, no problem. 957 */ 958 959 if (!uvm_swapisfull()) { 960 return true; 961 } 962 963 /* 964 * file-backed pages can be reclaimed even when swap is full. 965 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim. 966 * 967 * XXX assume the worst case, ie. all wired pages are file-backed. 968 * 969 * XXX should consider about other reclaimable memory. 970 * XXX ie. pools, traditional buffer cache. 971 */ 972 973 cpu_count_sync_all(); 974 filepages = (int)cpu_count_get(CPU_COUNT_FILEPAGES) + 975 (int)cpu_count_get(CPU_COUNT_EXECPAGES) - uvmexp.wired; 976 uvm_estimatepageable(&active, &inactive); 977 if (filepages >= MIN((active + inactive) >> 4, 978 5 * 1024 * 1024 >> PAGE_SHIFT)) { 979 return true; 980 } 981 982 /* 983 * kill the process, fail allocation, etc.. 984 */ 985 986 return false; 987 } 988 989 void 990 uvm_estimatepageable(int *active, int *inactive) 991 { 992 993 uvmpdpol_estimatepageable(active, inactive); 994 } 995 996 997 /* 998 * Use a separate thread for draining pools. 999 * This work can't done from the main pagedaemon thread because 1000 * some pool allocators need to take vm_map locks. 1001 */ 1002 1003 static void 1004 uvmpd_pool_drain_thread(void *arg) 1005 { 1006 struct pool *firstpool, *curpool; 1007 int bufcnt, lastslept; 1008 bool cycled; 1009 1010 firstpool = NULL; 1011 cycled = true; 1012 for (;;) { 1013 /* 1014 * sleep until awoken by the pagedaemon. 1015 */ 1016 mutex_enter(&uvmpd_lock); 1017 if (!uvmpd_pool_drain_run) { 1018 lastslept = hardclock_ticks; 1019 cv_wait(&uvmpd_pool_drain_cv, &uvmpd_lock); 1020 if (hardclock_ticks != lastslept) { 1021 cycled = false; 1022 firstpool = NULL; 1023 } 1024 } 1025 uvmpd_pool_drain_run = false; 1026 mutex_exit(&uvmpd_lock); 1027 1028 /* 1029 * rate limit draining, otherwise in desperate circumstances 1030 * this can totally saturate the system with xcall activity. 1031 */ 1032 if (cycled) { 1033 kpause("uvmpdlmt", false, 1, NULL); 1034 cycled = false; 1035 firstpool = NULL; 1036 } 1037 1038 /* 1039 * drain and temporarily disable the freelist cache. 1040 */ 1041 uvm_pgflcache_pause(); 1042 1043 /* 1044 * kill unused metadata buffers. 1045 */ 1046 bufcnt = uvmexp.freetarg - uvm_availmem(); 1047 if (bufcnt < 0) 1048 bufcnt = 0; 1049 1050 mutex_enter(&bufcache_lock); 1051 buf_drain(bufcnt << PAGE_SHIFT); 1052 mutex_exit(&bufcache_lock); 1053 1054 /* 1055 * drain a pool, and then re-enable the freelist cache. 1056 */ 1057 (void)pool_drain(&curpool); 1058 KASSERT(curpool != NULL); 1059 if (firstpool == NULL) { 1060 firstpool = curpool; 1061 } else if (firstpool == curpool) { 1062 cycled = true; 1063 } 1064 uvm_pgflcache_resume(); 1065 } 1066 /*NOTREACHED*/ 1067 } 1068 1069 static void 1070 uvmpd_pool_drain_wakeup(void) 1071 { 1072 1073 mutex_enter(&uvmpd_lock); 1074 uvmpd_pool_drain_run = true; 1075 cv_signal(&uvmpd_pool_drain_cv); 1076 mutex_exit(&uvmpd_lock); 1077 } 1078