1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1998-2010 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 * Copyright (c) 1994 John S. Dyson 37 * Copyright (c) 1990 University of Utah. 38 * Copyright (c) 1991, 1993 39 * The Regents of the University of California. All rights reserved. 40 * 41 * This code is derived from software contributed to Berkeley by 42 * the Systems Programming Group of the University of Utah Computer 43 * Science Department. 44 * 45 * Redistribution and use in source and binary forms, with or without 46 * modification, are permitted provided that the following conditions 47 * are met: 48 * 1. Redistributions of source code must retain the above copyright 49 * notice, this list of conditions and the following disclaimer. 50 * 2. Redistributions in binary form must reproduce the above copyright 51 * notice, this list of conditions and the following disclaimer in the 52 * documentation and/or other materials provided with the distribution. 53 * 3. Neither the name of the University nor the names of its contributors 54 * may be used to endorse or promote products derived from this software 55 * without specific prior written permission. 56 * 57 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 58 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 59 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 60 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 61 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 62 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 63 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 64 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 65 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 66 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 67 * SUCH DAMAGE. 68 * 69 * New Swap System 70 * Matthew Dillon 71 * 72 * Radix Bitmap 'blists'. 73 * 74 * - The new swapper uses the new radix bitmap code. This should scale 75 * to arbitrarily small or arbitrarily large swap spaces and an almost 76 * arbitrary degree of fragmentation. 77 * 78 * Features: 79 * 80 * - on the fly reallocation of swap during putpages. The new system 81 * does not try to keep previously allocated swap blocks for dirty 82 * pages. 83 * 84 * - on the fly deallocation of swap 85 * 86 * - No more garbage collection required. Unnecessarily allocated swap 87 * blocks only exist for dirty vm_page_t's now and these are already 88 * cycled (in a high-load system) by the pager. We also do on-the-fly 89 * removal of invalidated swap blocks when a page is destroyed 90 * or renamed. 91 * 92 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$ 93 * @(#)swap_pager.c 8.9 (Berkeley) 3/21/94 94 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $ 95 */ 96 97 #include "opt_swap.h" 98 #include <sys/param.h> 99 #include <sys/systm.h> 100 #include <sys/conf.h> 101 #include <sys/kernel.h> 102 #include <sys/proc.h> 103 #include <sys/buf.h> 104 #include <sys/vnode.h> 105 #include <sys/malloc.h> 106 #include <sys/vmmeter.h> 107 #include <sys/sysctl.h> 108 #include <sys/blist.h> 109 #include <sys/lock.h> 110 #include <sys/kcollect.h> 111 112 #include <vm/vm.h> 113 #include <vm/vm_object.h> 114 #include <vm/vm_page.h> 115 #include <vm/vm_pager.h> 116 #include <vm/vm_pageout.h> 117 #include <vm/swap_pager.h> 118 #include <vm/vm_extern.h> 119 #include <vm/vm_zone.h> 120 #include <vm/vnode_pager.h> 121 122 #include <sys/buf2.h> 123 #include <vm/vm_page2.h> 124 125 #ifndef MAX_PAGEOUT_CLUSTER 126 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES 127 #endif 128 129 #define SWM_FREE 0x02 /* free, period */ 130 #define SWM_POP 0x04 /* pop out */ 131 132 #define SWBIO_READ 0x01 133 #define SWBIO_WRITE 0x02 134 #define SWBIO_SYNC 0x04 135 #define SWBIO_TTC 0x08 /* for VM_PAGER_TRY_TO_CACHE */ 136 137 struct swfreeinfo { 138 vm_object_t object; 139 vm_pindex_t basei; 140 vm_pindex_t begi; 141 vm_pindex_t endi; /* inclusive */ 142 }; 143 144 struct swswapoffinfo { 145 vm_object_t object; 146 int devidx; 147 int shared; 148 }; 149 150 /* 151 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks 152 * in the old system. 153 */ 154 155 int swap_pager_full; /* swap space exhaustion (task killing) */ 156 int swap_fail_ticks; /* when we became exhausted */ 157 int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/ 158 swblk_t vm_swap_cache_use; 159 swblk_t vm_swap_anon_use; 160 static int vm_report_swap_allocs; 161 162 static struct krate kswaprate = { 1 }; 163 static int nsw_rcount; /* free read buffers */ 164 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 165 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 166 static int nsw_wcount_async_max;/* assigned maximum */ 167 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 168 169 struct blist *swapblist; 170 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 171 static int swap_burst_read = 0; /* allow burst reading */ 172 static swblk_t swapiterator; /* linearize allocations */ 173 int swap_user_async = 0; /* user swap pager operation can be async */ 174 175 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin"); 176 177 /* from vm_swap.c */ 178 extern struct vnode *swapdev_vp; 179 extern struct swdevt *swdevt; 180 extern int nswdev; 181 182 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / SWB_DMMAX % nswdev : 0) 183 184 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 185 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 186 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read, 187 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins"); 188 SYSCTL_INT(_vm, OID_AUTO, swap_user_async, 189 CTLFLAG_RW, &swap_user_async, 0, "Allow async uuser swap write I/O"); 190 191 #if SWBLK_BITS == 64 192 SYSCTL_LONG(_vm, OID_AUTO, swap_cache_use, 193 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 194 SYSCTL_LONG(_vm, OID_AUTO, swap_anon_use, 195 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 196 SYSCTL_LONG(_vm, OID_AUTO, swap_free, 197 CTLFLAG_RD, &vm_swap_size, 0, ""); 198 SYSCTL_LONG(_vm, OID_AUTO, swap_size, 199 CTLFLAG_RD, &vm_swap_max, 0, ""); 200 #else 201 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use, 202 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 203 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use, 204 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 205 SYSCTL_INT(_vm, OID_AUTO, swap_free, 206 CTLFLAG_RD, &vm_swap_size, 0, ""); 207 SYSCTL_INT(_vm, OID_AUTO, swap_size, 208 CTLFLAG_RD, &vm_swap_max, 0, ""); 209 #endif 210 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs, 211 CTLFLAG_RW, &vm_report_swap_allocs, 0, ""); 212 213 __read_mostly vm_zone_t swap_zone; 214 215 /* 216 * Red-Black tree for swblock entries 217 * 218 * The caller must hold vm_token 219 */ 220 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare, 221 vm_pindex_t, swb_index); 222 223 int 224 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2) 225 { 226 if (swb1->swb_index < swb2->swb_index) 227 return(-1); 228 if (swb1->swb_index > swb2->swb_index) 229 return(1); 230 return(0); 231 } 232 233 static 234 int 235 rb_swblock_scancmp(struct swblock *swb, void *data) 236 { 237 struct swfreeinfo *info = data; 238 239 if (swb->swb_index < info->basei) 240 return(-1); 241 if (swb->swb_index > info->endi) 242 return(1); 243 return(0); 244 } 245 246 static 247 int 248 rb_swblock_condcmp(struct swblock *swb, void *data) 249 { 250 struct swfreeinfo *info = data; 251 252 if (swb->swb_index < info->basei) 253 return(-1); 254 return(0); 255 } 256 257 /* 258 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 259 * calls hooked from other parts of the VM system and do not appear here. 260 * (see vm/swap_pager.h). 261 */ 262 263 static void swap_pager_dealloc (vm_object_t object); 264 static int swap_pager_getpage (vm_object_t, vm_page_t *, int); 265 static void swap_chain_iodone(struct bio *biox); 266 267 struct pagerops swappagerops = { 268 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 269 swap_pager_getpage, /* pagein */ 270 swap_pager_putpages, /* pageout */ 271 swap_pager_haspage /* get backing store status for page */ 272 }; 273 274 /* 275 * SWB_DMMAX is in page-sized chunks with the new swap system. It was 276 * dev-bsized chunks in the old. SWB_DMMAX is always a power of 2. 277 * 278 * swap_*() routines are externally accessible. swp_*() routines are 279 * internal. 280 */ 281 282 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 283 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 284 285 static __inline void swp_sizecheck (void); 286 static void swp_pager_async_iodone (struct bio *bio); 287 288 /* 289 * Swap bitmap functions 290 */ 291 292 static __inline void swp_pager_freeswapspace(vm_object_t object, 293 swblk_t blk, int npages); 294 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages); 295 296 /* 297 * Metadata functions 298 */ 299 300 static void swp_pager_meta_convert(vm_object_t); 301 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t); 302 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 303 static void swp_pager_meta_free_all(vm_object_t); 304 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 305 306 /* 307 * SWP_SIZECHECK() - update swap_pager_full indication 308 * 309 * update the swap_pager_almost_full indication and warn when we are 310 * about to run out of swap space, using lowat/hiwat hysteresis. 311 * 312 * Clear swap_pager_full ( task killing ) indication when lowat is met. 313 * 314 * No restrictions on call 315 * This routine may not block. 316 * SMP races are ok. 317 */ 318 static __inline void 319 swp_sizecheck(void) 320 { 321 if (vm_swap_size < nswap_lowat) { 322 if (swap_pager_almost_full == 0) { 323 kprintf("swap_pager: out of swap space\n"); 324 swap_pager_almost_full = 1; 325 swap_fail_ticks = ticks; 326 } 327 } else { 328 swap_pager_full = 0; 329 if (vm_swap_size > nswap_hiwat) 330 swap_pager_almost_full = 0; 331 } 332 } 333 334 /* 335 * Long-term data collection on 10-second interval. Return the value 336 * for KCOLLECT_SWAPPCT and set the values for SWAPANO and SWAPCCAC. 337 * 338 * Return total swap in the scale field. This can change if swap is 339 * regularly added or removed and may cause some historical confusion 340 * in that case, but SWAPPCT will always be historically accurate. 341 */ 342 343 #define PTOB(value) ((uint64_t)(value) << PAGE_SHIFT) 344 345 static uint64_t 346 collect_swap_callback(int n) 347 { 348 uint64_t total = vm_swap_max; 349 uint64_t anon = vm_swap_anon_use; 350 uint64_t cache = vm_swap_cache_use; 351 352 if (total == 0) /* avoid divide by zero */ 353 total = 1; 354 kcollect_setvalue(KCOLLECT_SWAPANO, PTOB(anon)); 355 kcollect_setvalue(KCOLLECT_SWAPCAC, PTOB(cache)); 356 kcollect_setscale(KCOLLECT_SWAPANO, 357 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, PTOB(total))); 358 kcollect_setscale(KCOLLECT_SWAPCAC, 359 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, PTOB(total))); 360 return (((anon + cache) * 10000 + (total >> 1)) / total); 361 } 362 363 /* 364 * SWAP_PAGER_INIT() - initialize the swap pager! 365 * 366 * Expected to be started from system init. NOTE: This code is run 367 * before much else so be careful what you depend on. Most of the VM 368 * system has yet to be initialized at this point. 369 * 370 * Called from the low level boot code only. 371 */ 372 static void 373 swap_pager_init(void *arg __unused) 374 { 375 kcollect_register(KCOLLECT_SWAPPCT, "swapuse", collect_swap_callback, 376 KCOLLECT_SCALE(KCOLLECT_SWAPPCT_FORMAT, 0)); 377 kcollect_register(KCOLLECT_SWAPANO, "swapano", NULL, 378 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, 0)); 379 kcollect_register(KCOLLECT_SWAPCAC, "swapcac", NULL, 380 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, 0)); 381 } 382 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL); 383 384 /* 385 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 386 * 387 * Expected to be started from pageout process once, prior to entering 388 * its main loop. 389 * 390 * Called from the low level boot code only. 391 */ 392 void 393 swap_pager_swap_init(void) 394 { 395 int n, n2; 396 397 /* 398 * Number of in-transit swap bp operations. Don't 399 * exhaust the pbufs completely. Make sure we 400 * initialize workable values (0 will work for hysteresis 401 * but it isn't very efficient). 402 * 403 * The nsw_cluster_max is constrained by the number of pages an XIO 404 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined 405 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 406 * constrained by the swap device interleave stripe size. 407 * 408 * Currently we hardwire nsw_wcount_async to 4. This limit is 409 * designed to prevent other I/O from having high latencies due to 410 * our pageout I/O. The value 4 works well for one or two active swap 411 * devices but is probably a little low if you have more. Even so, 412 * a higher value would probably generate only a limited improvement 413 * with three or four active swap devices since the system does not 414 * typically have to pageout at extreme bandwidths. We will want 415 * at least 2 per swap devices, and 4 is a pretty good value if you 416 * have one NFS swap device due to the command/ack latency over NFS. 417 * So it all works out pretty well. 418 */ 419 420 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 421 422 nsw_rcount = (nswbuf_kva + 1) / 2; 423 nsw_wcount_sync = (nswbuf_kva + 3) / 4; 424 nsw_wcount_async = 4; 425 nsw_wcount_async_max = nsw_wcount_async; 426 427 /* 428 * The zone is dynamically allocated so generally size it to 429 * maxswzone (32MB to 256GB of KVM). Set a minimum size based 430 * on physical memory of around 8x (each swblock can hold 16 pages). 431 * 432 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio 433 * has increased dramatically. 434 */ 435 n = vmstats.v_page_count / 2; 436 if (maxswzone && n < maxswzone / sizeof(struct swblock)) 437 n = maxswzone / sizeof(struct swblock); 438 n2 = n; 439 440 do { 441 swap_zone = zinit( 442 "SWAPMETA", 443 sizeof(struct swblock), 444 n, 445 ZONE_INTERRUPT); 446 if (swap_zone != NULL) 447 break; 448 /* 449 * if the allocation failed, try a zone two thirds the 450 * size of the previous attempt. 451 */ 452 n -= ((n + 2) / 3); 453 } while (n > 0); 454 455 if (swap_zone == NULL) 456 panic("swap_pager_swap_init: swap_zone == NULL"); 457 if (n2 != n) 458 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n); 459 } 460 461 /* 462 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 463 * its metadata structures. 464 * 465 * This routine is called from the mmap and fork code to create a new 466 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 467 * and then converting it with swp_pager_meta_convert(). 468 * 469 * We only support unnamed objects. 470 * 471 * No restrictions. 472 */ 473 vm_object_t 474 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 475 { 476 vm_object_t object; 477 478 KKASSERT(handle == NULL); 479 object = vm_object_allocate_hold(OBJT_DEFAULT, 480 OFF_TO_IDX(offset + PAGE_MASK + size)); 481 swp_pager_meta_convert(object); 482 vm_object_drop(object); 483 484 return (object); 485 } 486 487 /* 488 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 489 * 490 * The swap backing for the object is destroyed. The code is 491 * designed such that we can reinstantiate it later, but this 492 * routine is typically called only when the entire object is 493 * about to be destroyed. 494 * 495 * The object must be locked or unreferenceable. 496 * No other requirements. 497 */ 498 static void 499 swap_pager_dealloc(vm_object_t object) 500 { 501 vm_object_hold(object); 502 vm_object_pip_wait(object, "swpdea"); 503 504 /* 505 * Free all remaining metadata. We only bother to free it from 506 * the swap meta data. We do not attempt to free swapblk's still 507 * associated with vm_page_t's for this object. We do not care 508 * if paging is still in progress on some objects. 509 */ 510 swp_pager_meta_free_all(object); 511 vm_object_drop(object); 512 } 513 514 /************************************************************************ 515 * SWAP PAGER BITMAP ROUTINES * 516 ************************************************************************/ 517 518 /* 519 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 520 * 521 * Allocate swap for the requested number of pages. The starting 522 * swap block number (a page index) is returned or SWAPBLK_NONE 523 * if the allocation failed. 524 * 525 * Also has the side effect of advising that somebody made a mistake 526 * when they configured swap and didn't configure enough. 527 * 528 * The caller must hold the object. 529 * This routine may not block. 530 */ 531 static __inline swblk_t 532 swp_pager_getswapspace(vm_object_t object, int npages) 533 { 534 swblk_t blk; 535 536 lwkt_gettoken(&vm_token); 537 blk = blist_allocat(swapblist, npages, swapiterator); 538 if (blk == SWAPBLK_NONE) 539 blk = blist_allocat(swapblist, npages, 0); 540 if (blk == SWAPBLK_NONE) { 541 if (swap_pager_full != 2) { 542 if (vm_swap_max == 0) { 543 krateprintf(&kswaprate, 544 "Warning: The system would like to " 545 "page to swap but no swap space " 546 "is configured!\n"); 547 } else { 548 krateprintf(&kswaprate, 549 "swap_pager_getswapspace: " 550 "swap full allocating %d pages\n", 551 npages); 552 } 553 swap_pager_full = 2; 554 if (swap_pager_almost_full == 0) 555 swap_fail_ticks = ticks; 556 swap_pager_almost_full = 1; 557 } 558 } else { 559 /* swapiterator = blk; disable for now, doesn't work well */ 560 swapacctspace(blk, -npages); 561 if (object->type == OBJT_SWAP) 562 vm_swap_anon_use += npages; 563 else 564 vm_swap_cache_use += npages; 565 swp_sizecheck(); 566 } 567 lwkt_reltoken(&vm_token); 568 return(blk); 569 } 570 571 /* 572 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 573 * 574 * This routine returns the specified swap blocks back to the bitmap. 575 * 576 * Note: This routine may not block (it could in the old swap code), 577 * and through the use of the new blist routines it does not block. 578 * 579 * This routine may not block. 580 */ 581 582 static __inline void 583 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages) 584 { 585 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)]; 586 587 lwkt_gettoken(&vm_token); 588 sp->sw_nused -= npages; 589 if (object->type == OBJT_SWAP) 590 vm_swap_anon_use -= npages; 591 else 592 vm_swap_cache_use -= npages; 593 594 if (sp->sw_flags & SW_CLOSING) { 595 lwkt_reltoken(&vm_token); 596 return; 597 } 598 599 blist_free(swapblist, blk, npages); 600 vm_swap_size += npages; 601 swp_sizecheck(); 602 lwkt_reltoken(&vm_token); 603 } 604 605 /* 606 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 607 * range within an object. 608 * 609 * This is a globally accessible routine. 610 * 611 * This routine removes swapblk assignments from swap metadata. 612 * 613 * The external callers of this routine typically have already destroyed 614 * or renamed vm_page_t's associated with this range in the object so 615 * we should be ok. 616 * 617 * No requirements. 618 */ 619 void 620 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 621 { 622 vm_object_hold(object); 623 swp_pager_meta_free(object, start, size); 624 vm_object_drop(object); 625 } 626 627 /* 628 * No requirements. 629 */ 630 void 631 swap_pager_freespace_all(vm_object_t object) 632 { 633 vm_object_hold(object); 634 swp_pager_meta_free_all(object); 635 vm_object_drop(object); 636 } 637 638 /* 639 * This function conditionally frees swap cache swap starting at 640 * (*basei) in the object. (count) swap blocks will be nominally freed. 641 * The actual number of blocks freed can be more or less than the 642 * requested number. 643 * 644 * This function nominally returns the number of blocks freed. However, 645 * the actual number of blocks freed may be less then the returned value. 646 * If the function is unable to exhaust the object or if it is able to 647 * free (approximately) the requested number of blocks it returns 648 * a value n > count. 649 * 650 * If we exhaust the object we will return a value n <= count. 651 * 652 * The caller must hold the object. 653 * 654 * WARNING! If count == 0 then -1 can be returned as a degenerate case, 655 * callers should always pass a count value > 0. 656 */ 657 static int swap_pager_condfree_callback(struct swblock *swap, void *data); 658 659 int 660 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count) 661 { 662 struct swfreeinfo info; 663 int n; 664 int t; 665 666 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 667 668 info.object = object; 669 info.basei = *basei; /* skip up to this page index */ 670 info.begi = count; /* max swap pages to destroy */ 671 info.endi = count * 8; /* max swblocks to scan */ 672 673 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp, 674 swap_pager_condfree_callback, &info); 675 *basei = info.basei; 676 677 /* 678 * Take the higher difference swblocks vs pages 679 */ 680 n = count - (int)info.begi; 681 t = count * 8 - (int)info.endi; 682 if (n < t) 683 n = t; 684 if (n < 1) 685 n = 1; 686 return(n); 687 } 688 689 /* 690 * The idea is to free whole meta-block to avoid fragmenting 691 * the swap space or disk I/O. We only do this if NO VM pages 692 * are present. 693 * 694 * We do not have to deal with clearing PG_SWAPPED in related VM 695 * pages because there are no related VM pages. 696 * 697 * The caller must hold the object. 698 */ 699 static int 700 swap_pager_condfree_callback(struct swblock *swap, void *data) 701 { 702 struct swfreeinfo *info = data; 703 vm_object_t object = info->object; 704 int i; 705 706 for (i = 0; i < SWAP_META_PAGES; ++i) { 707 if (vm_page_lookup(object, swap->swb_index + i)) 708 break; 709 } 710 info->basei = swap->swb_index + SWAP_META_PAGES; 711 if (i == SWAP_META_PAGES) { 712 info->begi -= swap->swb_count; 713 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES); 714 } 715 --info->endi; 716 if ((int)info->begi < 0 || (int)info->endi < 0) 717 return(-1); 718 lwkt_yield(); 719 return(0); 720 } 721 722 /* 723 * Called by vm_page_alloc() when a new VM page is inserted 724 * into a VM object. Checks whether swap has been assigned to 725 * the page and sets PG_SWAPPED as necessary. 726 * 727 * (m) must be busied by caller and remains busied on return. 728 */ 729 void 730 swap_pager_page_inserted(vm_page_t m) 731 { 732 if (m->object->swblock_count) { 733 vm_object_hold(m->object); 734 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE) 735 vm_page_flag_set(m, PG_SWAPPED); 736 vm_object_drop(m->object); 737 } 738 } 739 740 /* 741 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 742 * 743 * Assigns swap blocks to the specified range within the object. The 744 * swap blocks are not zerod. Any previous swap assignment is destroyed. 745 * 746 * Returns 0 on success, -1 on failure. 747 * 748 * The caller is responsible for avoiding races in the specified range. 749 * No other requirements. 750 */ 751 int 752 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 753 { 754 int n = 0; 755 swblk_t blk = SWAPBLK_NONE; 756 vm_pindex_t beg = start; /* save start index */ 757 758 vm_object_hold(object); 759 760 while (size) { 761 if (n == 0) { 762 n = BLIST_MAX_ALLOC; 763 while ((blk = swp_pager_getswapspace(object, n)) == 764 SWAPBLK_NONE) 765 { 766 n >>= 1; 767 if (n == 0) { 768 swp_pager_meta_free(object, beg, 769 start - beg); 770 vm_object_drop(object); 771 return(-1); 772 } 773 } 774 } 775 swp_pager_meta_build(object, start, blk); 776 --size; 777 ++start; 778 ++blk; 779 --n; 780 } 781 swp_pager_meta_free(object, start, n); 782 vm_object_drop(object); 783 return(0); 784 } 785 786 /* 787 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 788 * and destroy the source. 789 * 790 * Copy any valid swapblks from the source to the destination. In 791 * cases where both the source and destination have a valid swapblk, 792 * we keep the destination's. 793 * 794 * This routine is allowed to block. It may block allocating metadata 795 * indirectly through swp_pager_meta_build() or if paging is still in 796 * progress on the source. 797 * 798 * XXX vm_page_collapse() kinda expects us not to block because we 799 * supposedly do not need to allocate memory, but for the moment we 800 * *may* have to get a little memory from the zone allocator, but 801 * it is taken from the interrupt memory. We should be ok. 802 * 803 * The source object contains no vm_page_t's (which is just as well) 804 * The source object is of type OBJT_SWAP. 805 * 806 * The source and destination objects must be held by the caller. 807 */ 808 void 809 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 810 vm_pindex_t base_index, int destroysource) 811 { 812 vm_pindex_t i; 813 814 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject)); 815 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject)); 816 817 /* 818 * transfer source to destination. 819 */ 820 for (i = 0; i < dstobject->size; ++i) { 821 swblk_t dstaddr; 822 823 /* 824 * Locate (without changing) the swapblk on the destination, 825 * unless it is invalid in which case free it silently, or 826 * if the destination is a resident page, in which case the 827 * source is thrown away. 828 */ 829 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 830 831 if (dstaddr == SWAPBLK_NONE) { 832 /* 833 * Destination has no swapblk and is not resident, 834 * copy source. 835 */ 836 swblk_t srcaddr; 837 838 srcaddr = swp_pager_meta_ctl(srcobject, 839 base_index + i, SWM_POP); 840 841 if (srcaddr != SWAPBLK_NONE) 842 swp_pager_meta_build(dstobject, i, srcaddr); 843 } else { 844 /* 845 * Destination has valid swapblk or it is represented 846 * by a resident page. We destroy the sourceblock. 847 */ 848 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE); 849 } 850 } 851 852 /* 853 * Free left over swap blocks in source. 854 * 855 * We have to revert the type to OBJT_DEFAULT so we do not accidently 856 * double-remove the object from the swap queues. 857 */ 858 if (destroysource) { 859 /* 860 * Reverting the type is not necessary, the caller is going 861 * to destroy srcobject directly, but I'm doing it here 862 * for consistency since we've removed the object from its 863 * queues. 864 */ 865 swp_pager_meta_free_all(srcobject); 866 if (srcobject->type == OBJT_SWAP) 867 srcobject->type = OBJT_DEFAULT; 868 } 869 } 870 871 /* 872 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 873 * the requested page. 874 * 875 * We determine whether good backing store exists for the requested 876 * page and return TRUE if it does, FALSE if it doesn't. 877 * 878 * If TRUE, we also try to determine how much valid, contiguous backing 879 * store exists before and after the requested page within a reasonable 880 * distance. We do not try to restrict it to the swap device stripe 881 * (that is handled in getpages/putpages). It probably isn't worth 882 * doing here. 883 * 884 * No requirements. 885 */ 886 boolean_t 887 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex) 888 { 889 swblk_t blk0; 890 891 /* 892 * do we have good backing store at the requested index ? 893 */ 894 vm_object_hold(object); 895 blk0 = swp_pager_meta_ctl(object, pindex, 0); 896 897 if (blk0 == SWAPBLK_NONE) { 898 vm_object_drop(object); 899 return (FALSE); 900 } 901 vm_object_drop(object); 902 return (TRUE); 903 } 904 905 /* 906 * Object must be held exclusive or shared by the caller. 907 */ 908 boolean_t 909 swap_pager_haspage_locked(vm_object_t object, vm_pindex_t pindex) 910 { 911 if (swp_pager_meta_ctl(object, pindex, 0) == SWAPBLK_NONE) 912 return FALSE; 913 return TRUE; 914 } 915 916 /* 917 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 918 * 919 * This removes any associated swap backing store, whether valid or 920 * not, from the page. This operates on any VM object, not just OBJT_SWAP 921 * objects. 922 * 923 * This routine is typically called when a page is made dirty, at 924 * which point any associated swap can be freed. MADV_FREE also 925 * calls us in a special-case situation 926 * 927 * NOTE!!! If the page is clean and the swap was valid, the caller 928 * should make the page dirty before calling this routine. 929 * This routine does NOT change the m->dirty status of the page. 930 * Also: MADV_FREE depends on it. 931 * 932 * The page must be busied. 933 * The caller can hold the object to avoid blocking, else we might block. 934 * No other requirements. 935 */ 936 void 937 swap_pager_unswapped(vm_page_t m) 938 { 939 if (m->flags & PG_SWAPPED) { 940 vm_object_hold(m->object); 941 KKASSERT(m->flags & PG_SWAPPED); 942 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 943 vm_page_flag_clear(m, PG_SWAPPED); 944 vm_object_drop(m->object); 945 } 946 } 947 948 /* 949 * SWAP_PAGER_STRATEGY() - read, write, free blocks 950 * 951 * This implements a VM OBJECT strategy function using swap backing store. 952 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP 953 * types. Only BUF_CMD_{READ,WRITE,FREEBLKS} is supported, any other 954 * requests will return EINVAL. 955 * 956 * This is intended to be a cacheless interface (i.e. caching occurs at 957 * higher levels), and is also used as a swap-based SSD cache for vnode 958 * and device objects. 959 * 960 * All I/O goes directly to and from the swap device. 961 * 962 * We currently attempt to run I/O synchronously or asynchronously as 963 * the caller requests. This isn't perfect because we loose error 964 * sequencing when we run multiple ops in parallel to satisfy a request. 965 * But this is swap, so we let it all hang out. 966 * 967 * NOTE: This function supports the KVABIO API wherein bp->b_data might 968 * not be synchronized to the current cpu. 969 * 970 * No requirements. 971 */ 972 void 973 swap_pager_strategy(vm_object_t object, struct bio *bio) 974 { 975 struct buf *bp = bio->bio_buf; 976 struct bio *nbio; 977 vm_pindex_t start; 978 vm_pindex_t biox_blkno = 0; 979 int count; 980 char *data; 981 struct bio *biox; 982 struct buf *bufx; 983 #if 0 984 struct bio_track *track; 985 #endif 986 987 #if 0 988 /* 989 * tracking for swapdev vnode I/Os 990 */ 991 if (bp->b_cmd == BUF_CMD_READ) 992 track = &swapdev_vp->v_track_read; 993 else 994 track = &swapdev_vp->v_track_write; 995 #endif 996 997 /* 998 * Only supported commands 999 */ 1000 if (bp->b_cmd != BUF_CMD_FREEBLKS && 1001 bp->b_cmd != BUF_CMD_READ && 1002 bp->b_cmd != BUF_CMD_WRITE) { 1003 bp->b_error = EINVAL; 1004 bp->b_flags |= B_ERROR | B_INVAL; 1005 biodone(bio); 1006 return; 1007 } 1008 1009 /* 1010 * bcount must be an integral number of pages. 1011 */ 1012 if (bp->b_bcount & PAGE_MASK) { 1013 bp->b_error = EINVAL; 1014 bp->b_flags |= B_ERROR | B_INVAL; 1015 biodone(bio); 1016 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 1017 "not page bounded\n", 1018 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 1019 return; 1020 } 1021 1022 /* 1023 * Clear error indication, initialize page index, count, data pointer. 1024 */ 1025 bp->b_error = 0; 1026 bp->b_flags &= ~B_ERROR; 1027 bp->b_resid = bp->b_bcount; 1028 1029 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 1030 count = howmany(bp->b_bcount, PAGE_SIZE); 1031 1032 /* 1033 * WARNING! Do not dereference *data without issuing a bkvasync() 1034 */ 1035 data = bp->b_data; 1036 1037 /* 1038 * Deal with BUF_CMD_FREEBLKS 1039 */ 1040 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 1041 /* 1042 * FREE PAGE(s) - destroy underlying swap that is no longer 1043 * needed. 1044 */ 1045 vm_object_hold(object); 1046 swp_pager_meta_free(object, start, count); 1047 vm_object_drop(object); 1048 bp->b_resid = 0; 1049 biodone(bio); 1050 return; 1051 } 1052 1053 /* 1054 * We need to be able to create a new cluster of I/O's. We cannot 1055 * use the caller fields of the passed bio so push a new one. 1056 * 1057 * Because nbio is just a placeholder for the cluster links, 1058 * we can biodone() the original bio instead of nbio to make 1059 * things a bit more efficient. 1060 */ 1061 nbio = push_bio(bio); 1062 nbio->bio_offset = bio->bio_offset; 1063 nbio->bio_caller_info1.cluster_head = NULL; 1064 nbio->bio_caller_info2.cluster_tail = NULL; 1065 1066 biox = NULL; 1067 bufx = NULL; 1068 1069 /* 1070 * Execute read or write 1071 */ 1072 vm_object_hold(object); 1073 1074 while (count > 0) { 1075 swblk_t blk; 1076 1077 /* 1078 * Obtain block. If block not found and writing, allocate a 1079 * new block and build it into the object. 1080 */ 1081 blk = swp_pager_meta_ctl(object, start, 0); 1082 if ((blk == SWAPBLK_NONE) && bp->b_cmd == BUF_CMD_WRITE) { 1083 blk = swp_pager_getswapspace(object, 1); 1084 if (blk == SWAPBLK_NONE) { 1085 bp->b_error = ENOMEM; 1086 bp->b_flags |= B_ERROR; 1087 break; 1088 } 1089 swp_pager_meta_build(object, start, blk); 1090 } 1091 1092 /* 1093 * Do we have to flush our current collection? Yes if: 1094 * 1095 * - no swap block at this index 1096 * - swap block is not contiguous 1097 * - we cross a physical disk boundry in the 1098 * stripe. 1099 */ 1100 if (biox && 1101 (biox_blkno + btoc(bufx->b_bcount) != blk || 1102 ((biox_blkno ^ blk) & ~SWB_DMMASK))) { 1103 switch(bp->b_cmd) { 1104 case BUF_CMD_READ: 1105 ++mycpu->gd_cnt.v_swapin; 1106 mycpu->gd_cnt.v_swappgsin += 1107 btoc(bufx->b_bcount); 1108 break; 1109 case BUF_CMD_WRITE: 1110 ++mycpu->gd_cnt.v_swapout; 1111 mycpu->gd_cnt.v_swappgsout += 1112 btoc(bufx->b_bcount); 1113 bufx->b_dirtyend = bufx->b_bcount; 1114 break; 1115 default: 1116 /* NOT REACHED */ 1117 break; 1118 } 1119 1120 /* 1121 * Finished with this buf. 1122 */ 1123 KKASSERT(bufx->b_bcount != 0); 1124 if (bufx->b_cmd != BUF_CMD_READ) 1125 bufx->b_dirtyend = bufx->b_bcount; 1126 biox = NULL; 1127 bufx = NULL; 1128 } 1129 1130 /* 1131 * Add new swapblk to biox, instantiating biox if necessary. 1132 * Zero-fill reads are able to take a shortcut. 1133 */ 1134 if (blk == SWAPBLK_NONE) { 1135 /* 1136 * We can only get here if we are reading. 1137 */ 1138 bkvasync(bp); 1139 bzero(data, PAGE_SIZE); 1140 bp->b_resid -= PAGE_SIZE; 1141 } else { 1142 if (biox == NULL) { 1143 /* XXX chain count > 4, wait to <= 4 */ 1144 1145 bufx = getpbuf(NULL); 1146 bufx->b_flags |= B_KVABIO; 1147 biox = &bufx->b_bio1; 1148 cluster_append(nbio, bufx); 1149 bufx->b_cmd = bp->b_cmd; 1150 biox->bio_done = swap_chain_iodone; 1151 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1152 biox->bio_caller_info1.cluster_parent = nbio; 1153 biox_blkno = blk; 1154 bufx->b_bcount = 0; 1155 bufx->b_data = data; 1156 } 1157 bufx->b_bcount += PAGE_SIZE; 1158 } 1159 --count; 1160 ++start; 1161 data += PAGE_SIZE; 1162 } 1163 1164 vm_object_drop(object); 1165 1166 /* 1167 * Flush out last buffer 1168 */ 1169 if (biox) { 1170 if (bufx->b_cmd == BUF_CMD_READ) { 1171 ++mycpu->gd_cnt.v_swapin; 1172 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1173 } else { 1174 ++mycpu->gd_cnt.v_swapout; 1175 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1176 bufx->b_dirtyend = bufx->b_bcount; 1177 } 1178 KKASSERT(bufx->b_bcount); 1179 if (bufx->b_cmd != BUF_CMD_READ) 1180 bufx->b_dirtyend = bufx->b_bcount; 1181 /* biox, bufx = NULL */ 1182 } 1183 1184 /* 1185 * Now initiate all the I/O. Be careful looping on our chain as 1186 * I/O's may complete while we are still initiating them. 1187 * 1188 * If the request is a 100% sparse read no bios will be present 1189 * and we just biodone() the buffer. 1190 */ 1191 nbio->bio_caller_info2.cluster_tail = NULL; 1192 bufx = nbio->bio_caller_info1.cluster_head; 1193 1194 if (bufx) { 1195 while (bufx) { 1196 biox = &bufx->b_bio1; 1197 BUF_KERNPROC(bufx); 1198 bufx = bufx->b_cluster_next; 1199 vn_strategy(swapdev_vp, biox); 1200 } 1201 } else { 1202 biodone(bio); 1203 } 1204 1205 /* 1206 * Completion of the cluster will also call biodone_chain(nbio). 1207 * We never call biodone(nbio) so we don't have to worry about 1208 * setting up a bio_done callback. It's handled in the sub-IO. 1209 */ 1210 /**/ 1211 } 1212 1213 /* 1214 * biodone callback 1215 * 1216 * No requirements. 1217 */ 1218 static void 1219 swap_chain_iodone(struct bio *biox) 1220 { 1221 struct buf **nextp; 1222 struct buf *bufx; /* chained sub-buffer */ 1223 struct bio *nbio; /* parent nbio with chain glue */ 1224 struct buf *bp; /* original bp associated with nbio */ 1225 int chain_empty; 1226 1227 bufx = biox->bio_buf; 1228 nbio = biox->bio_caller_info1.cluster_parent; 1229 bp = nbio->bio_buf; 1230 1231 /* 1232 * Update the original buffer 1233 */ 1234 KKASSERT(bp != NULL); 1235 if (bufx->b_flags & B_ERROR) { 1236 atomic_set_int(&bufx->b_flags, B_ERROR); 1237 bp->b_error = bufx->b_error; /* race ok */ 1238 } else if (bufx->b_resid != 0) { 1239 atomic_set_int(&bufx->b_flags, B_ERROR); 1240 bp->b_error = EINVAL; /* race ok */ 1241 } else { 1242 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1243 } 1244 1245 /* 1246 * Remove us from the chain. 1247 */ 1248 spin_lock(&swapbp_spin); 1249 nextp = &nbio->bio_caller_info1.cluster_head; 1250 while (*nextp != bufx) { 1251 KKASSERT(*nextp != NULL); 1252 nextp = &(*nextp)->b_cluster_next; 1253 } 1254 *nextp = bufx->b_cluster_next; 1255 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1256 spin_unlock(&swapbp_spin); 1257 1258 /* 1259 * Clean up bufx. If the chain is now empty we finish out 1260 * the parent. Note that we may be racing other completions 1261 * so we must use the chain_empty status from above. 1262 */ 1263 if (chain_empty) { 1264 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1265 atomic_set_int(&bp->b_flags, B_ERROR); 1266 bp->b_error = EINVAL; 1267 } 1268 biodone_chain(nbio); 1269 } 1270 relpbuf(bufx, NULL); 1271 } 1272 1273 /* 1274 * SWAP_PAGER_GETPAGES() - bring page in from swap 1275 * 1276 * The requested page may have to be brought in from swap. Calculate the 1277 * swap block and bring in additional pages if possible. All pages must 1278 * have contiguous swap block assignments and reside in the same object. 1279 * 1280 * The caller has a single vm_object_pip_add() reference prior to 1281 * calling us and we should return with the same. 1282 * 1283 * The caller has BUSY'd the page. We should return with (*mpp) left busy, 1284 * and any additinal pages unbusied. 1285 * 1286 * If the caller encounters a PG_RAM page it will pass it to us even though 1287 * it may be valid and dirty. We cannot overwrite the page in this case! 1288 * The case is used to allow us to issue pure read-aheads. 1289 * 1290 * NOTE! XXX This code does not entirely pipeline yet due to the fact that 1291 * the PG_RAM page is validated at the same time as mreq. What we 1292 * really need to do is issue a separate read-ahead pbuf. 1293 * 1294 * No requirements. 1295 */ 1296 static int 1297 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 1298 { 1299 struct buf *bp; 1300 struct bio *bio; 1301 vm_page_t mreq; 1302 vm_page_t m; 1303 vm_offset_t kva; 1304 swblk_t blk; 1305 int i; 1306 int j; 1307 int raonly; 1308 int error; 1309 u_int32_t busy_count; 1310 vm_page_t marray[XIO_INTERNAL_PAGES]; 1311 1312 mreq = *mpp; 1313 1314 vm_object_hold(object); 1315 if (mreq->object != object) { 1316 panic("swap_pager_getpages: object mismatch %p/%p", 1317 object, 1318 mreq->object 1319 ); 1320 } 1321 1322 /* 1323 * We don't want to overwrite a fully valid page as it might be 1324 * dirty. This case can occur when e.g. vm_fault hits a perfectly 1325 * valid page with PG_RAM set. 1326 * 1327 * In this case we see if the next page is a suitable page-in 1328 * candidate and if it is we issue read-ahead. PG_RAM will be 1329 * set on the last page of the read-ahead to continue the pipeline. 1330 */ 1331 if (mreq->valid == VM_PAGE_BITS_ALL) { 1332 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) { 1333 vm_object_drop(object); 1334 return(VM_PAGER_OK); 1335 } 1336 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0); 1337 if (blk == SWAPBLK_NONE) { 1338 vm_object_drop(object); 1339 return(VM_PAGER_OK); 1340 } 1341 m = vm_page_lookup_busy_try(object, mreq->pindex + 1, 1342 TRUE, &error); 1343 if (error) { 1344 vm_object_drop(object); 1345 return(VM_PAGER_OK); 1346 } else if (m == NULL) { 1347 /* 1348 * Use VM_ALLOC_QUICK to avoid blocking on cache 1349 * page reuse. 1350 */ 1351 m = vm_page_alloc(object, mreq->pindex + 1, 1352 VM_ALLOC_QUICK); 1353 if (m == NULL) { 1354 vm_object_drop(object); 1355 return(VM_PAGER_OK); 1356 } 1357 } else { 1358 if (m->valid) { 1359 vm_page_wakeup(m); 1360 vm_object_drop(object); 1361 return(VM_PAGER_OK); 1362 } 1363 vm_page_unqueue_nowakeup(m); 1364 } 1365 /* page is busy */ 1366 mreq = m; 1367 raonly = 1; 1368 } else { 1369 raonly = 0; 1370 } 1371 1372 /* 1373 * Try to block-read contiguous pages from swap if sequential, 1374 * otherwise just read one page. Contiguous pages from swap must 1375 * reside within a single device stripe because the I/O cannot be 1376 * broken up across multiple stripes. 1377 * 1378 * Note that blk and iblk can be SWAPBLK_NONE but the loop is 1379 * set up such that the case(s) are handled implicitly. 1380 */ 1381 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1382 marray[0] = mreq; 1383 1384 for (i = 1; i <= swap_burst_read && 1385 i < XIO_INTERNAL_PAGES && 1386 mreq->pindex + i < object->size; ++i) { 1387 swblk_t iblk; 1388 1389 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0); 1390 if (iblk != blk + i) 1391 break; 1392 if ((blk ^ iblk) & ~SWB_DMMASK) 1393 break; 1394 m = vm_page_lookup_busy_try(object, mreq->pindex + i, 1395 TRUE, &error); 1396 if (error) { 1397 break; 1398 } else if (m == NULL) { 1399 /* 1400 * Use VM_ALLOC_QUICK to avoid blocking on cache 1401 * page reuse. 1402 */ 1403 m = vm_page_alloc(object, mreq->pindex + i, 1404 VM_ALLOC_QUICK); 1405 if (m == NULL) 1406 break; 1407 } else { 1408 if (m->valid) { 1409 vm_page_wakeup(m); 1410 break; 1411 } 1412 vm_page_unqueue_nowakeup(m); 1413 } 1414 /* page is busy */ 1415 marray[i] = m; 1416 } 1417 if (i > 1) 1418 vm_page_flag_set(marray[i - 1], PG_RAM); 1419 1420 /* 1421 * If mreq is the requested page and we have nothing to do return 1422 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead 1423 * page and must be cleaned up. 1424 */ 1425 if (blk == SWAPBLK_NONE) { 1426 KKASSERT(i == 1); 1427 if (raonly) { 1428 vnode_pager_freepage(mreq); 1429 vm_object_drop(object); 1430 return(VM_PAGER_OK); 1431 } else { 1432 vm_object_drop(object); 1433 return(VM_PAGER_FAIL); 1434 } 1435 } 1436 1437 /* 1438 * Map our page(s) into kva for input 1439 * 1440 * Use the KVABIO API to avoid synchronizing the pmap. 1441 */ 1442 bp = getpbuf_kva(&nsw_rcount); 1443 bio = &bp->b_bio1; 1444 kva = (vm_offset_t) bp->b_kvabase; 1445 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t)); 1446 pmap_qenter_noinval(kva, bp->b_xio.xio_pages, i); 1447 1448 bp->b_data = (caddr_t)kva; 1449 bp->b_bcount = PAGE_SIZE * i; 1450 bp->b_xio.xio_npages = i; 1451 bp->b_flags |= B_KVABIO; 1452 bio->bio_done = swp_pager_async_iodone; 1453 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1454 bio->bio_caller_info1.index = SWBIO_READ; 1455 1456 /* 1457 * Set index. If raonly set the index beyond the array so all 1458 * the pages are treated the same, otherwise the original mreq is 1459 * at index 0. 1460 */ 1461 if (raonly) 1462 bio->bio_driver_info = (void *)(intptr_t)i; 1463 else 1464 bio->bio_driver_info = (void *)(intptr_t)0; 1465 1466 for (j = 0; j < i; ++j) { 1467 atomic_set_int(&bp->b_xio.xio_pages[j]->busy_count, 1468 PBUSY_SWAPINPROG); 1469 } 1470 1471 mycpu->gd_cnt.v_swapin++; 1472 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1473 1474 /* 1475 * We still hold the lock on mreq, and our automatic completion routine 1476 * does not remove it. 1477 */ 1478 vm_object_pip_add(object, bp->b_xio.xio_npages); 1479 1480 /* 1481 * perform the I/O. NOTE!!! bp cannot be considered valid after 1482 * this point because we automatically release it on completion. 1483 * Instead, we look at the one page we are interested in which we 1484 * still hold a lock on even through the I/O completion. 1485 * 1486 * The other pages in our m[] array are also released on completion, 1487 * so we cannot assume they are valid anymore either. 1488 */ 1489 bp->b_cmd = BUF_CMD_READ; 1490 BUF_KERNPROC(bp); 1491 vn_strategy(swapdev_vp, bio); 1492 1493 /* 1494 * Wait for the page we want to complete. PBUSY_SWAPINPROG is always 1495 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1496 * is set in the meta-data. 1497 * 1498 * If this is a read-ahead only we return immediately without 1499 * waiting for I/O. 1500 */ 1501 if (raonly) { 1502 vm_object_drop(object); 1503 return(VM_PAGER_OK); 1504 } 1505 1506 /* 1507 * Read-ahead includes originally requested page case. 1508 */ 1509 for (;;) { 1510 busy_count = mreq->busy_count; 1511 cpu_ccfence(); 1512 if ((busy_count & PBUSY_SWAPINPROG) == 0) 1513 break; 1514 tsleep_interlock(mreq, 0); 1515 if (!atomic_cmpset_int(&mreq->busy_count, busy_count, 1516 busy_count | 1517 PBUSY_SWAPINPROG | PBUSY_WANTED)) { 1518 continue; 1519 } 1520 atomic_set_int(&mreq->flags, PG_REFERENCED); 1521 mycpu->gd_cnt.v_intrans++; 1522 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) { 1523 kprintf( 1524 "swap_pager: indefinite wait buffer: " 1525 " bp %p offset: %lld, size: %ld " 1526 " m=%p busy=%08x flags=%08x\n", 1527 bp, 1528 (long long)bio->bio_offset, 1529 (long)bp->b_bcount, 1530 mreq, mreq->busy_count, mreq->flags); 1531 } 1532 } 1533 1534 /* 1535 * Disallow speculative reads prior to the SWAPINPROG test. 1536 */ 1537 cpu_lfence(); 1538 1539 /* 1540 * mreq is left busied after completion, but all the other pages 1541 * are freed. If we had an unrecoverable read error the page will 1542 * not be valid. 1543 */ 1544 vm_object_drop(object); 1545 if (mreq->valid != VM_PAGE_BITS_ALL) 1546 return(VM_PAGER_ERROR); 1547 else 1548 return(VM_PAGER_OK); 1549 1550 /* 1551 * A final note: in a low swap situation, we cannot deallocate swap 1552 * and mark a page dirty here because the caller is likely to mark 1553 * the page clean when we return, causing the page to possibly revert 1554 * to all-zero's later. 1555 */ 1556 } 1557 1558 /* 1559 * swap_pager_putpages: 1560 * 1561 * Assign swap (if necessary) and initiate I/O on the specified pages. 1562 * 1563 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1564 * are automatically converted to SWAP objects. 1565 * 1566 * In a low memory situation we may block in vn_strategy(), but the new 1567 * vm_page reservation system coupled with properly written VFS devices 1568 * should ensure that no low-memory deadlock occurs. This is an area 1569 * which needs work. 1570 * 1571 * The parent has N vm_object_pip_add() references prior to 1572 * calling us and will remove references for rtvals[] that are 1573 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1574 * completion. 1575 * 1576 * The parent has soft-busy'd the pages it passes us and will unbusy 1577 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1578 * We need to unbusy the rest on I/O completion. 1579 * 1580 * No requirements. 1581 */ 1582 void 1583 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1584 int flags, int *rtvals) 1585 { 1586 int i; 1587 int n = 0; 1588 1589 vm_object_hold(object); 1590 1591 if (count && m[0]->object != object) { 1592 panic("swap_pager_getpages: object mismatch %p/%p", 1593 object, 1594 m[0]->object 1595 ); 1596 } 1597 1598 /* 1599 * Step 1 1600 * 1601 * Turn object into OBJT_SWAP 1602 * Check for bogus sysops 1603 * 1604 * Force sync if not pageout process, we don't want any single 1605 * non-pageout process to be able to hog the I/O subsystem! This 1606 * can be overridden by setting. 1607 */ 1608 if (object->type == OBJT_DEFAULT) { 1609 if (object->type == OBJT_DEFAULT) 1610 swp_pager_meta_convert(object); 1611 } 1612 1613 /* 1614 * Normally we force synchronous swap I/O if this is not the 1615 * pageout daemon to prevent any single user process limited 1616 * via RLIMIT_RSS from hogging swap write bandwidth. 1617 */ 1618 if (curthread != pagethread && 1619 curthread != emergpager && 1620 swap_user_async == 0) { 1621 flags |= VM_PAGER_PUT_SYNC; 1622 } 1623 1624 /* 1625 * Step 2 1626 * 1627 * Update nsw parameters from swap_async_max sysctl values. 1628 * Do not let the sysop crash the machine with bogus numbers. 1629 */ 1630 if (swap_async_max != nsw_wcount_async_max) { 1631 int n; 1632 1633 /* 1634 * limit range 1635 */ 1636 if ((n = swap_async_max) > nswbuf_kva / 2) 1637 n = nswbuf_kva / 2; 1638 if (n < 1) 1639 n = 1; 1640 swap_async_max = n; 1641 1642 /* 1643 * Adjust difference ( if possible ). If the current async 1644 * count is too low, we may not be able to make the adjustment 1645 * at this time. 1646 * 1647 * vm_token needed for nsw_wcount sleep interlock 1648 */ 1649 lwkt_gettoken(&vm_token); 1650 n -= nsw_wcount_async_max; 1651 if (nsw_wcount_async + n >= 0) { 1652 nsw_wcount_async_max += n; 1653 pbuf_adjcount(&nsw_wcount_async, n); 1654 } 1655 lwkt_reltoken(&vm_token); 1656 } 1657 1658 /* 1659 * Step 3 1660 * 1661 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1662 * The page is left dirty until the pageout operation completes 1663 * successfully. 1664 */ 1665 1666 for (i = 0; i < count; i += n) { 1667 struct buf *bp; 1668 struct bio *bio; 1669 swblk_t blk; 1670 int j; 1671 1672 /* 1673 * Maximum I/O size is limited by a number of factors. 1674 */ 1675 1676 n = min(BLIST_MAX_ALLOC, count - i); 1677 n = min(n, nsw_cluster_max); 1678 1679 lwkt_gettoken(&vm_token); 1680 1681 /* 1682 * Get biggest block of swap we can. If we fail, fall 1683 * back and try to allocate a smaller block. Don't go 1684 * overboard trying to allocate space if it would overly 1685 * fragment swap. 1686 */ 1687 while ( 1688 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE && 1689 n > 4 1690 ) { 1691 n >>= 1; 1692 } 1693 if (blk == SWAPBLK_NONE) { 1694 for (j = 0; j < n; ++j) 1695 rtvals[i+j] = VM_PAGER_FAIL; 1696 lwkt_reltoken(&vm_token); 1697 continue; 1698 } 1699 if (vm_report_swap_allocs > 0) { 1700 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n); 1701 --vm_report_swap_allocs; 1702 } 1703 1704 /* 1705 * The I/O we are constructing cannot cross a physical 1706 * disk boundry in the swap stripe. 1707 */ 1708 if ((blk ^ (blk + n)) & ~SWB_DMMASK) { 1709 j = ((blk + SWB_DMMAX) & ~SWB_DMMASK) - blk; 1710 swp_pager_freeswapspace(object, blk + j, n - j); 1711 n = j; 1712 } 1713 1714 /* 1715 * All I/O parameters have been satisfied, build the I/O 1716 * request and assign the swap space. 1717 * 1718 * Use the KVABIO API to avoid synchronizing the pmap. 1719 */ 1720 if ((flags & VM_PAGER_PUT_SYNC)) 1721 bp = getpbuf_kva(&nsw_wcount_sync); 1722 else 1723 bp = getpbuf_kva(&nsw_wcount_async); 1724 bio = &bp->b_bio1; 1725 1726 lwkt_reltoken(&vm_token); 1727 1728 pmap_qenter_noinval((vm_offset_t)bp->b_data, &m[i], n); 1729 1730 bp->b_flags |= B_KVABIO; 1731 bp->b_bcount = PAGE_SIZE * n; 1732 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1733 1734 for (j = 0; j < n; ++j) { 1735 vm_page_t mreq = m[i+j]; 1736 1737 swp_pager_meta_build(mreq->object, mreq->pindex, 1738 blk + j); 1739 if (object->type == OBJT_SWAP) 1740 vm_page_dirty(mreq); 1741 rtvals[i+j] = VM_PAGER_OK; 1742 1743 atomic_set_int(&mreq->busy_count, PBUSY_SWAPINPROG); 1744 bp->b_xio.xio_pages[j] = mreq; 1745 } 1746 bp->b_xio.xio_npages = n; 1747 1748 mycpu->gd_cnt.v_swapout++; 1749 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1750 1751 bp->b_dirtyoff = 0; /* req'd for NFS */ 1752 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1753 bp->b_cmd = BUF_CMD_WRITE; 1754 bio->bio_caller_info1.index = SWBIO_WRITE; 1755 1756 /* 1757 * asynchronous 1758 */ 1759 if ((flags & VM_PAGER_PUT_SYNC) == 0) { 1760 bio->bio_done = swp_pager_async_iodone; 1761 BUF_KERNPROC(bp); 1762 vn_strategy(swapdev_vp, bio); 1763 1764 for (j = 0; j < n; ++j) 1765 rtvals[i+j] = VM_PAGER_PEND; 1766 continue; 1767 } 1768 1769 /* 1770 * Issue synchrnously. 1771 * 1772 * Wait for the sync I/O to complete, then update rtvals. 1773 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1774 * our async completion routine at the end, thus avoiding a 1775 * double-free. 1776 */ 1777 bio->bio_caller_info1.index |= SWBIO_SYNC; 1778 if (flags & VM_PAGER_TRY_TO_CACHE) 1779 bio->bio_caller_info1.index |= SWBIO_TTC; 1780 bio->bio_done = biodone_sync; 1781 bio->bio_flags |= BIO_SYNC; 1782 vn_strategy(swapdev_vp, bio); 1783 biowait(bio, "swwrt"); 1784 1785 for (j = 0; j < n; ++j) 1786 rtvals[i+j] = VM_PAGER_PEND; 1787 1788 /* 1789 * Now that we are through with the bp, we can call the 1790 * normal async completion, which frees everything up. 1791 */ 1792 swp_pager_async_iodone(bio); 1793 } 1794 vm_object_drop(object); 1795 } 1796 1797 /* 1798 * No requirements. 1799 * 1800 * Recalculate the low and high-water marks. 1801 */ 1802 void 1803 swap_pager_newswap(void) 1804 { 1805 /* 1806 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the 1807 * limitation imposed by the blist code. Remember that this 1808 * will be divided by NSWAP_MAX (4), so each swap device is 1809 * limited to around a terrabyte. 1810 */ 1811 if (vm_swap_max) { 1812 nswap_lowat = (int64_t)vm_swap_max * 4 / 100; /* 4% left */ 1813 nswap_hiwat = (int64_t)vm_swap_max * 6 / 100; /* 6% left */ 1814 kprintf("swap low/high-water marks set to %d/%d\n", 1815 nswap_lowat, nswap_hiwat); 1816 } else { 1817 nswap_lowat = 128; 1818 nswap_hiwat = 512; 1819 } 1820 swp_sizecheck(); 1821 } 1822 1823 /* 1824 * swp_pager_async_iodone: 1825 * 1826 * Completion routine for asynchronous reads and writes from/to swap. 1827 * Also called manually by synchronous code to finish up a bp. 1828 * 1829 * For READ operations, the pages are BUSY'd. For WRITE operations, 1830 * the pages are vm_page_t->busy'd. For READ operations, we BUSY 1831 * unbusy all pages except the 'main' request page. For WRITE 1832 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1833 * because we marked them all VM_PAGER_PEND on return from putpages ). 1834 * 1835 * This routine may not block. 1836 * 1837 * No requirements. 1838 */ 1839 static void 1840 swp_pager_async_iodone(struct bio *bio) 1841 { 1842 struct buf *bp = bio->bio_buf; 1843 vm_object_t object = NULL; 1844 int i; 1845 int *nswptr; 1846 1847 /* 1848 * report error 1849 */ 1850 if (bp->b_flags & B_ERROR) { 1851 kprintf( 1852 "swap_pager: I/O error - %s failed; offset %lld," 1853 "size %ld, error %d\n", 1854 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1855 "pagein" : "pageout"), 1856 (long long)bio->bio_offset, 1857 (long)bp->b_bcount, 1858 bp->b_error 1859 ); 1860 } 1861 1862 /* 1863 * set object. 1864 */ 1865 if (bp->b_xio.xio_npages) 1866 object = bp->b_xio.xio_pages[0]->object; 1867 1868 #if 0 1869 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */ 1870 if (bio->bio_caller_info1.index & SWBIO_WRITE) { 1871 if (bio->bio_crc != iscsi_crc32(bp->b_data, bp->b_bcount)) { 1872 kprintf("SWAPOUT: BADCRC %08x %08x\n", 1873 bio->bio_crc, 1874 iscsi_crc32(bp->b_data, bp->b_bcount)); 1875 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1876 vm_page_t m = bp->b_xio.xio_pages[i]; 1877 if ((m->flags & PG_WRITEABLE) && 1878 (pmap_mapped_sync(m) & PG_WRITEABLE)) { 1879 kprintf("SWAPOUT: " 1880 "%d/%d %p writable\n", 1881 i, bp->b_xio.xio_npages, m); 1882 } 1883 } 1884 } 1885 } 1886 #endif 1887 1888 /* 1889 * remove the mapping for kernel virtual 1890 */ 1891 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1892 1893 /* 1894 * cleanup pages. If an error occurs writing to swap, we are in 1895 * very serious trouble. If it happens to be a disk error, though, 1896 * we may be able to recover by reassigning the swap later on. So 1897 * in this case we remove the m->swapblk assignment for the page 1898 * but do not free it in the rlist. The errornous block(s) are thus 1899 * never reallocated as swap. Redirty the page and continue. 1900 */ 1901 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1902 vm_page_t m = bp->b_xio.xio_pages[i]; 1903 1904 if (bp->b_flags & B_ERROR) { 1905 /* 1906 * If an error occurs I'd love to throw the swapblk 1907 * away without freeing it back to swapspace, so it 1908 * can never be used again. But I can't from an 1909 * interrupt. 1910 */ 1911 1912 if (bio->bio_caller_info1.index & SWBIO_READ) { 1913 /* 1914 * When reading, reqpage needs to stay 1915 * locked for the parent, but all other 1916 * pages can be freed. We still want to 1917 * wakeup the parent waiting on the page, 1918 * though. ( also: pg_reqpage can be -1 and 1919 * not match anything ). 1920 * 1921 * We have to wake specifically requested pages 1922 * up too because we cleared SWAPINPROG and 1923 * someone may be waiting for that. 1924 * 1925 * NOTE: For reads, m->dirty will probably 1926 * be overridden by the original caller 1927 * of getpages so don't play cute tricks 1928 * here. 1929 * 1930 * NOTE: We can't actually free the page from 1931 * here, because this is an interrupt. 1932 * It is not legal to mess with 1933 * object->memq from an interrupt. 1934 * Deactivate the page instead. 1935 * 1936 * WARNING! The instant SWAPINPROG is 1937 * cleared another cpu may start 1938 * using the mreq page (it will 1939 * check m->valid immediately). 1940 */ 1941 1942 m->valid = 0; 1943 atomic_clear_int(&m->busy_count, 1944 PBUSY_SWAPINPROG); 1945 1946 /* 1947 * bio_driver_info holds the requested page 1948 * index. 1949 */ 1950 if (i != (int)(intptr_t)bio->bio_driver_info) { 1951 vm_page_deactivate(m); 1952 vm_page_wakeup(m); 1953 } else { 1954 vm_page_flash(m); 1955 } 1956 /* 1957 * If i == bp->b_pager.pg_reqpage, do not wake 1958 * the page up. The caller needs to. 1959 */ 1960 } else { 1961 /* 1962 * If a write error occurs remove the swap 1963 * assignment (note that PG_SWAPPED may or 1964 * may not be set depending on prior activity). 1965 * 1966 * Re-dirty OBJT_SWAP pages as there is no 1967 * other backing store, we can't throw the 1968 * page away. 1969 * 1970 * Non-OBJT_SWAP pages (aka swapcache) must 1971 * not be dirtied since they may not have 1972 * been dirty in the first place, and they 1973 * do have backing store (the vnode). 1974 */ 1975 vm_page_busy_wait(m, FALSE, "swadpg"); 1976 vm_object_hold(m->object); 1977 swp_pager_meta_ctl(m->object, m->pindex, 1978 SWM_FREE); 1979 vm_page_flag_clear(m, PG_SWAPPED); 1980 vm_object_drop(m->object); 1981 if (m->object->type == OBJT_SWAP) { 1982 vm_page_dirty(m); 1983 vm_page_activate(m); 1984 } 1985 vm_page_io_finish(m); 1986 atomic_clear_int(&m->busy_count, 1987 PBUSY_SWAPINPROG); 1988 vm_page_wakeup(m); 1989 } 1990 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1991 /* 1992 * NOTE: for reads, m->dirty will probably be 1993 * overridden by the original caller of getpages so 1994 * we cannot set them in order to free the underlying 1995 * swap in a low-swap situation. I don't think we'd 1996 * want to do that anyway, but it was an optimization 1997 * that existed in the old swapper for a time before 1998 * it got ripped out due to precisely this problem. 1999 * 2000 * If not the requested page then deactivate it. 2001 * 2002 * Note that the requested page, reqpage, is left 2003 * busied, but we still have to wake it up. The 2004 * other pages are released (unbusied) by 2005 * vm_page_wakeup(). We do not set reqpage's 2006 * valid bits here, it is up to the caller. 2007 */ 2008 2009 /* 2010 * NOTE: Can't call pmap_clear_modify(m) from an 2011 * interrupt thread, the pmap code may have to 2012 * map non-kernel pmaps and currently asserts 2013 * the case. 2014 * 2015 * WARNING! The instant SWAPINPROG is 2016 * cleared another cpu may start 2017 * using the mreq page (it will 2018 * check m->valid immediately). 2019 */ 2020 /*pmap_clear_modify(m);*/ 2021 m->valid = VM_PAGE_BITS_ALL; 2022 vm_page_undirty(m); 2023 vm_page_flag_set(m, PG_SWAPPED); 2024 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG); 2025 2026 /* 2027 * We have to wake specifically requested pages 2028 * up too because we cleared SWAPINPROG and 2029 * could be waiting for it in getpages. However, 2030 * be sure to not unbusy getpages specifically 2031 * requested page - getpages expects it to be 2032 * left busy. 2033 * 2034 * bio_driver_info holds the requested page 2035 */ 2036 if (i != (int)(intptr_t)bio->bio_driver_info) { 2037 vm_page_deactivate(m); 2038 vm_page_wakeup(m); 2039 } else { 2040 vm_page_flash(m); 2041 } 2042 } else { 2043 /* 2044 * Mark the page clean but do not mess with the 2045 * pmap-layer's modified state. That state should 2046 * also be clear since the caller protected the 2047 * page VM_PROT_READ, but allow the case. 2048 * 2049 * We are in an interrupt, avoid pmap operations. 2050 * 2051 * If we have a severe page deficit, deactivate the 2052 * page. Do not try to cache it (which would also 2053 * involve a pmap op), because the page might still 2054 * be read-heavy. 2055 * 2056 * When using the swap to cache clean vnode pages 2057 * we do not mess with the page dirty bits. 2058 * 2059 * NOTE! Nobody is waiting for the key mreq page 2060 * on write completion. 2061 */ 2062 vm_page_busy_wait(m, FALSE, "swadpg"); 2063 if (m->object->type == OBJT_SWAP) 2064 vm_page_undirty(m); 2065 vm_page_flag_set(m, PG_SWAPPED); 2066 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG); 2067 if (vm_page_count_severe()) 2068 vm_page_deactivate(m); 2069 vm_page_io_finish(m); 2070 if (bio->bio_caller_info1.index & SWBIO_TTC) 2071 vm_page_try_to_cache(m); 2072 else 2073 vm_page_wakeup(m); 2074 } 2075 } 2076 2077 /* 2078 * adjust pip. NOTE: the original parent may still have its own 2079 * pip refs on the object. 2080 */ 2081 2082 if (object) 2083 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages); 2084 2085 /* 2086 * Release the physical I/O buffer. 2087 * 2088 * NOTE: Due to synchronous operations in the write case b_cmd may 2089 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 2090 * been cleared. 2091 * 2092 * Use vm_token to interlock nsw_rcount/wcount wakeup? 2093 */ 2094 lwkt_gettoken(&vm_token); 2095 if (bio->bio_caller_info1.index & SWBIO_READ) 2096 nswptr = &nsw_rcount; 2097 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 2098 nswptr = &nsw_wcount_sync; 2099 else 2100 nswptr = &nsw_wcount_async; 2101 bp->b_cmd = BUF_CMD_DONE; 2102 relpbuf(bp, nswptr); 2103 lwkt_reltoken(&vm_token); 2104 } 2105 2106 /* 2107 * Fault-in a potentially swapped page and remove the swap reference. 2108 * (used by swapoff code) 2109 * 2110 * object must be held. 2111 */ 2112 static __inline void 2113 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex) 2114 { 2115 struct vnode *vp; 2116 vm_page_t m; 2117 int error; 2118 2119 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2120 2121 if (object->type == OBJT_VNODE) { 2122 /* 2123 * Any swap related to a vnode is due to swapcache. We must 2124 * vget() the vnode in case it is not active (otherwise 2125 * vref() will panic). Calling vm_object_page_remove() will 2126 * ensure that any swap ref is removed interlocked with the 2127 * page. clean_only is set to TRUE so we don't throw away 2128 * dirty pages. 2129 */ 2130 vp = object->handle; 2131 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 2132 if (error == 0) { 2133 vm_object_page_remove(object, pindex, pindex + 1, TRUE); 2134 vput(vp); 2135 } 2136 } else { 2137 /* 2138 * Otherwise it is a normal OBJT_SWAP object and we can 2139 * fault the page in and remove the swap. 2140 */ 2141 m = vm_fault_object_page(object, IDX_TO_OFF(pindex), 2142 VM_PROT_NONE, 2143 VM_FAULT_DIRTY | VM_FAULT_UNSWAP, 2144 sharedp, &error); 2145 if (m) 2146 vm_page_unhold(m); 2147 } 2148 } 2149 2150 /* 2151 * This removes all swap blocks related to a particular device. We have 2152 * to be careful of ripups during the scan. 2153 */ 2154 static int swp_pager_swapoff_callback(struct swblock *swap, void *data); 2155 2156 int 2157 swap_pager_swapoff(int devidx) 2158 { 2159 struct vm_object_hash *hash; 2160 struct swswapoffinfo info; 2161 struct vm_object marker; 2162 vm_object_t object; 2163 int n; 2164 2165 bzero(&marker, sizeof(marker)); 2166 marker.type = OBJT_MARKER; 2167 2168 for (n = 0; n < VMOBJ_HSIZE; ++n) { 2169 hash = &vm_object_hash[n]; 2170 2171 lwkt_gettoken(&hash->token); 2172 TAILQ_INSERT_HEAD(&hash->list, &marker, object_entry); 2173 2174 while ((object = TAILQ_NEXT(&marker, object_entry)) != NULL) { 2175 if (object->type == OBJT_MARKER) 2176 goto skip; 2177 if (object->type != OBJT_SWAP && 2178 object->type != OBJT_VNODE) 2179 goto skip; 2180 vm_object_hold(object); 2181 if (object->type != OBJT_SWAP && 2182 object->type != OBJT_VNODE) { 2183 vm_object_drop(object); 2184 goto skip; 2185 } 2186 2187 /* 2188 * Object is special in that we can't just pagein 2189 * into vm_page's in it (tmpfs, vn). 2190 */ 2191 if ((object->flags & OBJ_NOPAGEIN) && 2192 RB_ROOT(&object->swblock_root)) { 2193 vm_object_drop(object); 2194 goto skip; 2195 } 2196 2197 info.object = object; 2198 info.shared = 0; 2199 info.devidx = devidx; 2200 swblock_rb_tree_RB_SCAN(&object->swblock_root, 2201 NULL, swp_pager_swapoff_callback, 2202 &info); 2203 vm_object_drop(object); 2204 skip: 2205 if (object == TAILQ_NEXT(&marker, object_entry)) { 2206 TAILQ_REMOVE(&hash->list, &marker, 2207 object_entry); 2208 TAILQ_INSERT_AFTER(&hash->list, object, 2209 &marker, object_entry); 2210 } 2211 } 2212 TAILQ_REMOVE(&hash->list, &marker, object_entry); 2213 lwkt_reltoken(&hash->token); 2214 } 2215 2216 /* 2217 * If we fail to locate all swblocks we just fail gracefully and 2218 * do not bother to restore paging on the swap device. If the 2219 * user wants to retry the user can retry. 2220 */ 2221 if (swdevt[devidx].sw_nused) 2222 return (1); 2223 else 2224 return (0); 2225 } 2226 2227 static 2228 int 2229 swp_pager_swapoff_callback(struct swblock *swap, void *data) 2230 { 2231 struct swswapoffinfo *info = data; 2232 vm_object_t object = info->object; 2233 vm_pindex_t index; 2234 swblk_t v; 2235 int i; 2236 2237 index = swap->swb_index; 2238 for (i = 0; i < SWAP_META_PAGES; ++i) { 2239 /* 2240 * Make sure we don't race a dying object. This will 2241 * kill the scan of the object's swap blocks entirely. 2242 */ 2243 if (object->flags & OBJ_DEAD) 2244 return(-1); 2245 2246 /* 2247 * Fault the page, which can obviously block. If the swap 2248 * structure disappears break out. 2249 */ 2250 v = swap->swb_pages[i]; 2251 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) { 2252 swp_pager_fault_page(object, &info->shared, 2253 swap->swb_index + i); 2254 /* swap ptr might go away */ 2255 if (RB_LOOKUP(swblock_rb_tree, 2256 &object->swblock_root, index) != swap) { 2257 break; 2258 } 2259 } 2260 } 2261 return(0); 2262 } 2263 2264 /************************************************************************ 2265 * SWAP META DATA * 2266 ************************************************************************ 2267 * 2268 * These routines manipulate the swap metadata stored in the 2269 * OBJT_SWAP object. 2270 * 2271 * Swap metadata is implemented with a global hash and not directly 2272 * linked into the object. Instead the object simply contains 2273 * appropriate tracking counters. 2274 */ 2275 2276 /* 2277 * Lookup the swblock containing the specified swap block index. 2278 * 2279 * The caller must hold the object. 2280 */ 2281 static __inline 2282 struct swblock * 2283 swp_pager_lookup(vm_object_t object, vm_pindex_t index) 2284 { 2285 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2286 index &= ~(vm_pindex_t)SWAP_META_MASK; 2287 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index)); 2288 } 2289 2290 /* 2291 * Remove a swblock from the RB tree. 2292 * 2293 * The caller must hold the object. 2294 */ 2295 static __inline 2296 void 2297 swp_pager_remove(vm_object_t object, struct swblock *swap) 2298 { 2299 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2300 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap); 2301 } 2302 2303 /* 2304 * Convert default object to swap object if necessary 2305 * 2306 * The caller must hold the object. 2307 */ 2308 static void 2309 swp_pager_meta_convert(vm_object_t object) 2310 { 2311 if (object->type == OBJT_DEFAULT) { 2312 object->type = OBJT_SWAP; 2313 KKASSERT(object->swblock_count == 0); 2314 } 2315 } 2316 2317 /* 2318 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2319 * 2320 * We first convert the object to a swap object if it is a default 2321 * object. Vnode objects do not need to be converted. 2322 * 2323 * The specified swapblk is added to the object's swap metadata. If 2324 * the swapblk is not valid, it is freed instead. Any previously 2325 * assigned swapblk is freed. 2326 * 2327 * The caller must hold the object. 2328 */ 2329 static void 2330 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk) 2331 { 2332 struct swblock *swap; 2333 struct swblock *oswap; 2334 vm_pindex_t v; 2335 2336 KKASSERT(swapblk != SWAPBLK_NONE); 2337 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2338 2339 /* 2340 * Convert object if necessary 2341 */ 2342 if (object->type == OBJT_DEFAULT) 2343 swp_pager_meta_convert(object); 2344 2345 /* 2346 * Locate swblock. If not found create, but if we aren't adding 2347 * anything just return. If we run out of space in the map we wait 2348 * and, since the hash table may have changed, retry. 2349 */ 2350 retry: 2351 swap = swp_pager_lookup(object, index); 2352 2353 if (swap == NULL) { 2354 int i; 2355 2356 swap = zalloc(swap_zone); 2357 if (swap == NULL) { 2358 vm_wait(0); 2359 goto retry; 2360 } 2361 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK; 2362 swap->swb_count = 0; 2363 2364 ++object->swblock_count; 2365 2366 for (i = 0; i < SWAP_META_PAGES; ++i) 2367 swap->swb_pages[i] = SWAPBLK_NONE; 2368 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap); 2369 KKASSERT(oswap == NULL); 2370 } 2371 2372 /* 2373 * Delete prior contents of metadata. 2374 * 2375 * NOTE: Decrement swb_count after the freeing operation (which 2376 * might block) to prevent racing destruction of the swblock. 2377 */ 2378 index &= SWAP_META_MASK; 2379 2380 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) { 2381 swap->swb_pages[index] = SWAPBLK_NONE; 2382 /* can block */ 2383 swp_pager_freeswapspace(object, v, 1); 2384 --swap->swb_count; 2385 --mycpu->gd_vmtotal.t_vm; 2386 } 2387 2388 /* 2389 * Enter block into metadata 2390 */ 2391 swap->swb_pages[index] = swapblk; 2392 if (swapblk != SWAPBLK_NONE) { 2393 ++swap->swb_count; 2394 ++mycpu->gd_vmtotal.t_vm; 2395 } 2396 } 2397 2398 /* 2399 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2400 * 2401 * The requested range of blocks is freed, with any associated swap 2402 * returned to the swap bitmap. 2403 * 2404 * This routine will free swap metadata structures as they are cleaned 2405 * out. This routine does *NOT* operate on swap metadata associated 2406 * with resident pages. 2407 * 2408 * The caller must hold the object. 2409 */ 2410 static int swp_pager_meta_free_callback(struct swblock *swb, void *data); 2411 2412 static void 2413 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 2414 { 2415 struct swfreeinfo info; 2416 2417 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2418 2419 /* 2420 * Nothing to do 2421 */ 2422 if (object->swblock_count == 0) { 2423 KKASSERT(RB_EMPTY(&object->swblock_root)); 2424 return; 2425 } 2426 if (count == 0) 2427 return; 2428 2429 /* 2430 * Setup for RB tree scan. Note that the pindex range can be huge 2431 * due to the 64 bit page index space so we cannot safely iterate. 2432 */ 2433 info.object = object; 2434 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK; 2435 info.begi = index; 2436 info.endi = index + count - 1; 2437 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp, 2438 swp_pager_meta_free_callback, &info); 2439 } 2440 2441 /* 2442 * The caller must hold the object. 2443 */ 2444 static 2445 int 2446 swp_pager_meta_free_callback(struct swblock *swap, void *data) 2447 { 2448 struct swfreeinfo *info = data; 2449 vm_object_t object = info->object; 2450 int index; 2451 int eindex; 2452 2453 /* 2454 * Figure out the range within the swblock. The wider scan may 2455 * return edge-case swap blocks when the start and/or end points 2456 * are in the middle of a block. 2457 */ 2458 if (swap->swb_index < info->begi) 2459 index = (int)info->begi & SWAP_META_MASK; 2460 else 2461 index = 0; 2462 2463 if (swap->swb_index + SWAP_META_PAGES > info->endi) 2464 eindex = (int)info->endi & SWAP_META_MASK; 2465 else 2466 eindex = SWAP_META_MASK; 2467 2468 /* 2469 * Scan and free the blocks. The loop terminates early 2470 * if (swap) runs out of blocks and could be freed. 2471 * 2472 * NOTE: Decrement swb_count after swp_pager_freeswapspace() 2473 * to deal with a zfree race. 2474 */ 2475 while (index <= eindex) { 2476 swblk_t v = swap->swb_pages[index]; 2477 2478 if (v != SWAPBLK_NONE) { 2479 swap->swb_pages[index] = SWAPBLK_NONE; 2480 /* can block */ 2481 swp_pager_freeswapspace(object, v, 1); 2482 --mycpu->gd_vmtotal.t_vm; 2483 if (--swap->swb_count == 0) { 2484 swp_pager_remove(object, swap); 2485 zfree(swap_zone, swap); 2486 --object->swblock_count; 2487 break; 2488 } 2489 } 2490 ++index; 2491 } 2492 2493 /* swap may be invalid here due to zfree above */ 2494 lwkt_yield(); 2495 2496 return(0); 2497 } 2498 2499 /* 2500 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2501 * 2502 * This routine locates and destroys all swap metadata associated with 2503 * an object. 2504 * 2505 * NOTE: Decrement swb_count after the freeing operation (which 2506 * might block) to prevent racing destruction of the swblock. 2507 * 2508 * The caller must hold the object. 2509 */ 2510 static void 2511 swp_pager_meta_free_all(vm_object_t object) 2512 { 2513 struct swblock *swap; 2514 int i; 2515 2516 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2517 2518 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) { 2519 swp_pager_remove(object, swap); 2520 for (i = 0; i < SWAP_META_PAGES; ++i) { 2521 swblk_t v = swap->swb_pages[i]; 2522 if (v != SWAPBLK_NONE) { 2523 /* can block */ 2524 swp_pager_freeswapspace(object, v, 1); 2525 --swap->swb_count; 2526 --mycpu->gd_vmtotal.t_vm; 2527 } 2528 } 2529 if (swap->swb_count != 0) 2530 panic("swap_pager_meta_free_all: swb_count != 0"); 2531 zfree(swap_zone, swap); 2532 --object->swblock_count; 2533 lwkt_yield(); 2534 } 2535 KKASSERT(object->swblock_count == 0); 2536 } 2537 2538 /* 2539 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2540 * 2541 * This routine is capable of looking up, popping, or freeing 2542 * swapblk assignments in the swap meta data or in the vm_page_t. 2543 * The routine typically returns the swapblk being looked-up, or popped, 2544 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2545 * was invalid. This routine will automatically free any invalid 2546 * meta-data swapblks. 2547 * 2548 * It is not possible to store invalid swapblks in the swap meta data 2549 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2550 * 2551 * When acting on a busy resident page and paging is in progress, we 2552 * have to wait until paging is complete but otherwise can act on the 2553 * busy page. 2554 * 2555 * SWM_FREE remove and free swap block from metadata 2556 * SWM_POP remove from meta data but do not free.. pop it out 2557 * 2558 * The caller must hold the object. 2559 */ 2560 static swblk_t 2561 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags) 2562 { 2563 struct swblock *swap; 2564 swblk_t r1; 2565 2566 if (object->swblock_count == 0) 2567 return(SWAPBLK_NONE); 2568 2569 r1 = SWAPBLK_NONE; 2570 swap = swp_pager_lookup(object, index); 2571 2572 if (swap != NULL) { 2573 index &= SWAP_META_MASK; 2574 r1 = swap->swb_pages[index]; 2575 2576 if (r1 != SWAPBLK_NONE) { 2577 if (flags & (SWM_FREE|SWM_POP)) { 2578 swap->swb_pages[index] = SWAPBLK_NONE; 2579 --mycpu->gd_vmtotal.t_vm; 2580 if (--swap->swb_count == 0) { 2581 swp_pager_remove(object, swap); 2582 zfree(swap_zone, swap); 2583 --object->swblock_count; 2584 } 2585 } 2586 /* swap ptr may be invalid */ 2587 if (flags & SWM_FREE) { 2588 swp_pager_freeswapspace(object, r1, 1); 2589 r1 = SWAPBLK_NONE; 2590 } 2591 } 2592 /* swap ptr may be invalid */ 2593 } 2594 return(r1); 2595 } 2596