1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 2010 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to The DragonFly Project 7 * by Matthew Dillon <dillon@backplane.com> 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in 17 * the documentation and/or other materials provided with the 18 * distribution. 19 * 3. Neither the name of The DragonFly Project nor the names of its 20 * contributors may be used to endorse or promote products derived 21 * from this software without specific, prior written permission. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 */ 36 37 /* 38 * Implement the swapcache daemon. When enabled swap is assumed to be 39 * configured on a fast storage device such as a SSD. Swap is assigned 40 * to clean vnode-backed pages in the inactive queue, clustered by object 41 * if possible, and written out. The swap assignment sticks around even 42 * after the underlying pages have been recycled. 43 * 44 * The daemon manages write bandwidth based on sysctl settings to control 45 * wear on the SSD. 46 * 47 * The vnode strategy code will check for the swap assignments and divert 48 * reads to the swap device when the data is present in the swapcache. 49 * 50 * This operates on both regular files and the block device vnodes used by 51 * filesystems to manage meta-data. 52 */ 53 54 #include "opt_vm.h" 55 #include <sys/param.h> 56 #include <sys/systm.h> 57 #include <sys/kernel.h> 58 #include <sys/proc.h> 59 #include <sys/kthread.h> 60 #include <sys/resourcevar.h> 61 #include <sys/signalvar.h> 62 #include <sys/vnode.h> 63 #include <sys/vmmeter.h> 64 #include <sys/sysctl.h> 65 #include <sys/eventhandler.h> 66 67 #include <vm/vm.h> 68 #include <vm/vm_param.h> 69 #include <sys/lock.h> 70 #include <vm/vm_object.h> 71 #include <vm/vm_page.h> 72 #include <vm/vm_map.h> 73 #include <vm/vm_pageout.h> 74 #include <vm/vm_pager.h> 75 #include <vm/swap_pager.h> 76 #include <vm/vm_extern.h> 77 78 #include <sys/thread2.h> 79 #include <vm/vm_page2.h> 80 81 #define INACTIVE_LIST (&vm_page_queues[PQ_INACTIVE].pl) 82 83 /* the kernel process "vm_pageout"*/ 84 static int vm_swapcached_flush (vm_page_t m, int isblkdev); 85 static int vm_swapcache_test(vm_page_t m); 86 static void vm_swapcache_writing(vm_page_t marker); 87 static void vm_swapcache_cleaning(vm_object_t marker); 88 struct thread *swapcached_thread; 89 90 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL); 91 92 int vm_swapcache_read_enable; 93 int vm_swapcache_inactive_heuristic; 94 static int vm_swapcache_sleep; 95 static int vm_swapcache_maxlaunder = 256; 96 static int vm_swapcache_data_enable = 0; 97 static int vm_swapcache_meta_enable = 0; 98 static int vm_swapcache_maxswappct = 75; 99 static int vm_swapcache_hysteresis; 100 int vm_swapcache_use_chflags = 1; /* require chflags cache */ 101 static int64_t vm_swapcache_minburst = 10000000LL; /* 10MB */ 102 static int64_t vm_swapcache_curburst = 4000000000LL; /* 4G after boot */ 103 static int64_t vm_swapcache_maxburst = 2000000000LL; /* 2G nominal max */ 104 static int64_t vm_swapcache_accrate = 100000LL; /* 100K/s */ 105 static int64_t vm_swapcache_write_count; 106 static int64_t vm_swapcache_maxfilesize; 107 108 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder, 109 CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, ""); 110 111 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable, 112 CTLFLAG_RW, &vm_swapcache_data_enable, 0, ""); 113 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable, 114 CTLFLAG_RW, &vm_swapcache_meta_enable, 0, ""); 115 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable, 116 CTLFLAG_RW, &vm_swapcache_read_enable, 0, ""); 117 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct, 118 CTLFLAG_RW, &vm_swapcache_maxswappct, 0, ""); 119 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis, 120 CTLFLAG_RW, &vm_swapcache_hysteresis, 0, ""); 121 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags, 122 CTLFLAG_RW, &vm_swapcache_use_chflags, 0, ""); 123 124 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst, 125 CTLFLAG_RW, &vm_swapcache_minburst, 0, ""); 126 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst, 127 CTLFLAG_RW, &vm_swapcache_curburst, 0, ""); 128 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst, 129 CTLFLAG_RW, &vm_swapcache_maxburst, 0, ""); 130 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize, 131 CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, ""); 132 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate, 133 CTLFLAG_RW, &vm_swapcache_accrate, 0, ""); 134 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count, 135 CTLFLAG_RW, &vm_swapcache_write_count, 0, ""); 136 137 #define SWAPMAX(adj) \ 138 ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100) 139 140 /* 141 * When shutting down the machine we want to stop swapcache operation 142 * immediately so swap is not accessed after devices have been shuttered. 143 */ 144 static void 145 shutdown_swapcache(void *arg __unused) 146 { 147 vm_swapcache_read_enable = 0; 148 vm_swapcache_data_enable = 0; 149 vm_swapcache_meta_enable = 0; 150 wakeup(&vm_swapcache_sleep); /* shortcut 5-second wait */ 151 } 152 153 /* 154 * vm_swapcached is the high level pageout daemon. 155 * 156 * No requirements. 157 */ 158 static void 159 vm_swapcached_thread(void) 160 { 161 enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING; 162 enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING; 163 struct vm_page page_marker; 164 struct vm_object object_marker; 165 166 /* 167 * Thread setup 168 */ 169 curthread->td_flags |= TDF_SYSTHREAD; 170 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 171 swapcached_thread, SHUTDOWN_PRI_FIRST); 172 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache, 173 NULL, SHUTDOWN_PRI_SECOND); 174 lwkt_gettoken(&vm_token); 175 crit_enter(); 176 177 /* 178 * Initialize our marker for the inactive scan (SWAPC_WRITING) 179 */ 180 bzero(&page_marker, sizeof(page_marker)); 181 page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 182 page_marker.queue = PQ_INACTIVE; 183 page_marker.wire_count = 1; 184 TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq); 185 vm_swapcache_hysteresis = vmstats.v_inactive_target / 2; 186 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 187 188 /* 189 * Initialize our marker for the vm_object scan (SWAPC_CLEANING) 190 */ 191 bzero(&object_marker, sizeof(object_marker)); 192 object_marker.type = OBJT_MARKER; 193 lwkt_gettoken(&vmobj_token); 194 TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list); 195 lwkt_reltoken(&vmobj_token); 196 197 for (;;) { 198 /* 199 * Handle shutdown 200 */ 201 kproc_suspend_loop(); 202 203 /* 204 * Check every 5 seconds when not enabled or if no swap 205 * is present. 206 */ 207 if ((vm_swapcache_data_enable == 0 && 208 vm_swapcache_meta_enable == 0) || 209 vm_swap_max == 0) { 210 tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5); 211 continue; 212 } 213 214 /* 215 * Polling rate when enabled is approximately 10 hz. 216 */ 217 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10); 218 219 /* 220 * State hysteresis. Generate write activity up to 75% of 221 * swap, then clean out swap assignments down to 70%, then 222 * repeat. 223 */ 224 if (state == SWAPC_WRITING) { 225 if (vm_swap_cache_use > SWAPMAX(0)) 226 state = SWAPC_CLEANING; 227 } else { 228 if (vm_swap_cache_use < SWAPMAX(-5)) 229 state = SWAPC_WRITING; 230 } 231 232 /* 233 * We are allowed to continue accumulating burst value 234 * in either state. Allow the user to set curburst > maxburst 235 * for the initial load-in. 236 */ 237 if (vm_swapcache_curburst < vm_swapcache_maxburst) { 238 vm_swapcache_curburst += vm_swapcache_accrate / 10; 239 if (vm_swapcache_curburst > vm_swapcache_maxburst) 240 vm_swapcache_curburst = vm_swapcache_maxburst; 241 } 242 243 /* 244 * We don't want to nickle-and-dime the scan as that will 245 * create unnecessary fragmentation. The minimum burst 246 * is one-seconds worth of accumulation. 247 */ 248 if (state == SWAPC_WRITING) { 249 if (vm_swapcache_curburst >= vm_swapcache_accrate) { 250 if (burst == SWAPB_BURSTING) { 251 vm_swapcache_writing(&page_marker); 252 if (vm_swapcache_curburst <= 0) 253 burst = SWAPB_RECOVERING; 254 } else if (vm_swapcache_curburst > 255 vm_swapcache_minburst) { 256 vm_swapcache_writing(&page_marker); 257 burst = SWAPB_BURSTING; 258 } 259 } 260 } else { 261 vm_swapcache_cleaning(&object_marker); 262 } 263 } 264 265 /* 266 * Cleanup (NOT REACHED) 267 */ 268 TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq); 269 crit_exit(); 270 lwkt_reltoken(&vm_token); 271 272 lwkt_gettoken(&vmobj_token); 273 TAILQ_REMOVE(&vm_object_list, &object_marker, object_list); 274 lwkt_reltoken(&vmobj_token); 275 } 276 277 static struct kproc_desc swpc_kp = { 278 "swapcached", 279 vm_swapcached_thread, 280 &swapcached_thread 281 }; 282 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp) 283 284 /* 285 * The caller must hold vm_token. 286 */ 287 static void 288 vm_swapcache_writing(vm_page_t marker) 289 { 290 vm_object_t object; 291 struct vnode *vp; 292 vm_page_t m; 293 int count; 294 int isblkdev; 295 296 /* 297 * Deal with an overflow of the heuristic counter or if the user 298 * manually changes the hysteresis. 299 * 300 * Try to avoid small incremental pageouts by waiting for enough 301 * pages to buildup in the inactive queue to hopefully get a good 302 * burst in. This heuristic is bumped by the VM system and reset 303 * when our scan hits the end of the queue. 304 */ 305 if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis) 306 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 307 if (vm_swapcache_inactive_heuristic < 0) 308 return; 309 310 /* 311 * Scan the inactive queue from our marker to locate 312 * suitable pages to push to the swap cache. 313 * 314 * We are looking for clean vnode-backed pages. 315 * 316 * NOTE: PG_SWAPPED pages in particular are not part of 317 * our count because once the cache stabilizes we 318 * can end up with a very high datarate of VM pages 319 * cycling from it. 320 */ 321 m = marker; 322 count = vm_swapcache_maxlaunder; 323 324 while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) { 325 if (m->flags & (PG_MARKER | PG_SWAPPED)) { 326 ++count; 327 continue; 328 } 329 if (vm_swapcache_curburst < 0) 330 break; 331 if (vm_swapcache_test(m)) 332 continue; 333 object = m->object; 334 vp = object->handle; 335 if (vp == NULL) 336 continue; 337 338 switch(vp->v_type) { 339 case VREG: 340 /* 341 * PG_NOTMETA generically means 'don't swapcache this', 342 * and HAMMER will set this for regular data buffers 343 * (and leave it unset for meta-data buffers) as 344 * appropriate when double buffering is enabled. 345 */ 346 if (m->flags & PG_NOTMETA) 347 continue; 348 349 /* 350 * If data_enable is 0 do not try to swapcache data. 351 * If use_chflags is set then only swapcache data for 352 * VSWAPCACHE marked vnodes, otherwise any vnode. 353 */ 354 if (vm_swapcache_data_enable == 0 || 355 ((vp->v_flag & VSWAPCACHE) == 0 && 356 vm_swapcache_use_chflags)) { 357 continue; 358 } 359 if (vm_swapcache_maxfilesize && 360 object->size > 361 (vm_swapcache_maxfilesize >> PAGE_SHIFT)) { 362 continue; 363 } 364 isblkdev = 0; 365 break; 366 case VCHR: 367 /* 368 * PG_NOTMETA generically means 'don't swapcache this', 369 * and HAMMER will set this for regular data buffers 370 * (and leave it unset for meta-data buffers) as 371 * appropriate when double buffering is enabled. 372 */ 373 if (m->flags & PG_NOTMETA) 374 continue; 375 if (vm_swapcache_meta_enable == 0) 376 continue; 377 isblkdev = 1; 378 break; 379 default: 380 continue; 381 } 382 383 /* 384 * Ok, move the marker and soft-busy the page. 385 */ 386 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq); 387 TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq); 388 389 /* 390 * Assign swap and initiate I/O. 391 * 392 * (adjust for the --count which also occurs in the loop) 393 */ 394 count -= vm_swapcached_flush(m, isblkdev) - 1; 395 396 /* 397 * Setup for next loop using marker. 398 */ 399 m = marker; 400 } 401 402 /* 403 * Cleanup marker position. If we hit the end of the 404 * list the marker is placed at the tail. Newly deactivated 405 * pages will be placed after it. 406 * 407 * Earlier inactive pages that were dirty and become clean 408 * are typically moved to the end of PQ_INACTIVE by virtue 409 * of vfs_vmio_release() when they become unwired from the 410 * buffer cache. 411 */ 412 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq); 413 if (m) { 414 TAILQ_INSERT_BEFORE(m, marker, pageq); 415 } else { 416 TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq); 417 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 418 } 419 } 420 421 /* 422 * Flush the specified page using the swap_pager. 423 * 424 * Try to collect surrounding pages, including pages which may 425 * have already been assigned swap. Try to cluster within a 426 * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block 427 * to match what swap_pager_putpages() can do. 428 * 429 * We also want to try to match against the buffer cache blocksize 430 * but we don't really know what it is here. Since the buffer cache 431 * wires and unwires pages in groups the fact that we skip wired pages 432 * should be sufficient. 433 * 434 * Returns a count of pages we might have flushed (minimum 1) 435 * 436 * The caller must hold vm_token. 437 */ 438 static 439 int 440 vm_swapcached_flush(vm_page_t m, int isblkdev) 441 { 442 vm_object_t object; 443 vm_page_t marray[SWAP_META_PAGES]; 444 vm_pindex_t basei; 445 int rtvals[SWAP_META_PAGES]; 446 int x; 447 int i; 448 int j; 449 int count; 450 451 vm_page_io_start(m); 452 vm_page_protect(m, VM_PROT_READ); 453 object = m->object; 454 455 /* 456 * Try to cluster around (m), keeping in mind that the swap pager 457 * can only do SMAP_META_PAGES worth of continguous write. 458 */ 459 x = (int)m->pindex & SWAP_META_MASK; 460 marray[x] = m; 461 basei = m->pindex; 462 463 for (i = x - 1; i >= 0; --i) { 464 m = vm_page_lookup(object, basei - x + i); 465 if (m == NULL) 466 break; 467 if (vm_swapcache_test(m)) 468 break; 469 if (isblkdev && (m->flags & PG_NOTMETA)) 470 break; 471 vm_page_io_start(m); 472 vm_page_protect(m, VM_PROT_READ); 473 if (m->queue - m->pc == PQ_CACHE) { 474 vm_page_unqueue_nowakeup(m); 475 vm_page_deactivate(m); 476 } 477 marray[i] = m; 478 } 479 ++i; 480 481 for (j = x + 1; j < SWAP_META_PAGES; ++j) { 482 m = vm_page_lookup(object, basei - x + j); 483 if (m == NULL) 484 break; 485 if (vm_swapcache_test(m)) 486 break; 487 if (isblkdev && (m->flags & PG_NOTMETA)) 488 break; 489 vm_page_io_start(m); 490 vm_page_protect(m, VM_PROT_READ); 491 if (m->queue - m->pc == PQ_CACHE) { 492 vm_page_unqueue_nowakeup(m); 493 vm_page_deactivate(m); 494 } 495 marray[j] = m; 496 } 497 498 count = j - i; 499 vm_object_pip_add(object, count); 500 swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i); 501 vm_swapcache_write_count += count * PAGE_SIZE; 502 vm_swapcache_curburst -= count * PAGE_SIZE; 503 504 while (i < j) { 505 if (rtvals[i] != VM_PAGER_PEND) { 506 vm_page_io_finish(marray[i]); 507 vm_object_pip_wakeup(object); 508 } 509 ++i; 510 } 511 return(count); 512 } 513 514 /* 515 * Test whether a VM page is suitable for writing to the swapcache. 516 * Does not test m->queue, PG_MARKER, or PG_SWAPPED. 517 * 518 * Returns 0 on success, 1 on failure 519 * 520 * The caller must hold vm_token. 521 */ 522 static int 523 vm_swapcache_test(vm_page_t m) 524 { 525 vm_object_t object; 526 527 if (m->flags & (PG_BUSY | PG_UNMANAGED)) 528 return(1); 529 if (m->busy || m->hold_count || m->wire_count) 530 return(1); 531 if (m->valid != VM_PAGE_BITS_ALL) 532 return(1); 533 if (m->dirty & m->valid) 534 return(1); 535 if ((object = m->object) == NULL) 536 return(1); 537 if (object->type != OBJT_VNODE || 538 (object->flags & OBJ_DEAD)) { 539 return(1); 540 } 541 vm_page_test_dirty(m); 542 if (m->dirty & m->valid) 543 return(1); 544 return(0); 545 } 546 547 /* 548 * Cleaning pass 549 * 550 * The caller must hold vm_token. 551 */ 552 static 553 void 554 vm_swapcache_cleaning(vm_object_t marker) 555 { 556 vm_object_t object; 557 struct vnode *vp; 558 int count; 559 int n; 560 561 object = marker; 562 count = vm_swapcache_maxlaunder; 563 564 /* 565 * Look for vnode objects 566 */ 567 lwkt_gettoken(&vm_token); 568 lwkt_gettoken(&vmobj_token); 569 570 while ((object = TAILQ_NEXT(object, object_list)) != NULL) { 571 if (--count <= 0) 572 break; 573 if (object->type != OBJT_VNODE) 574 continue; 575 if ((object->flags & OBJ_DEAD) || object->swblock_count == 0) 576 continue; 577 if ((vp = object->handle) == NULL) 578 continue; 579 if (vp->v_type != VREG && vp->v_type != VCHR) 580 continue; 581 582 /* 583 * Adjust iterator. 584 */ 585 if (marker->backing_object != object) 586 marker->size = 0; 587 588 /* 589 * Move the marker so we can work on the VM object 590 */ 591 TAILQ_REMOVE(&vm_object_list, marker, object_list); 592 TAILQ_INSERT_AFTER(&vm_object_list, object, 593 marker, object_list); 594 595 /* 596 * Look for swblocks starting at our iterator. 597 * 598 * The swap_pager_condfree() function attempts to free 599 * swap space starting at the specified index. The index 600 * will be updated on return. The function will return 601 * a scan factor (NOT the number of blocks freed). 602 * 603 * If it must cut its scan of the object short due to an 604 * excessive number of swblocks, or is able to free the 605 * requested number of blocks, it will return n >= count 606 * and we break and pick it back up on a future attempt. 607 */ 608 n = swap_pager_condfree(object, &marker->size, count); 609 count -= n; 610 if (count < 0) 611 break; 612 613 /* 614 * Setup for loop. 615 */ 616 marker->size = 0; 617 object = marker; 618 } 619 620 /* 621 * Adjust marker so we continue the scan from where we left off. 622 * When we reach the end we start back at the beginning. 623 */ 624 TAILQ_REMOVE(&vm_object_list, marker, object_list); 625 if (object) 626 TAILQ_INSERT_BEFORE(object, marker, object_list); 627 else 628 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list); 629 marker->backing_object = object; 630 631 lwkt_reltoken(&vmobj_token); 632 lwkt_reltoken(&vm_token); 633 } 634