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 <sys/param.h> 98 #include <sys/systm.h> 99 #include <sys/conf.h> 100 #include <sys/kernel.h> 101 #include <sys/proc.h> 102 #include <sys/buf.h> 103 #include <sys/vnode.h> 104 #include <sys/malloc.h> 105 #include <sys/vmmeter.h> 106 #include <sys/sysctl.h> 107 #include <sys/blist.h> 108 #include <sys/lock.h> 109 #include <sys/kcollect.h> 110 111 #include <unistd.h> 112 #include "opt_swap.h" 113 #include <vm/vm.h> 114 #include <vm/vm_object.h> 115 #include <vm/vm_page.h> 116 #include <vm/vm_pager.h> 117 #include <vm/vm_pageout.h> 118 #include <vm/swap_pager.h> 119 #include <vm/vm_extern.h> 120 #include <vm/vm_zone.h> 121 #include <vm/vnode_pager.h> 122 123 #include <sys/thread2.h> 124 #include <sys/buf2.h> 125 #include <vm/vm_page2.h> 126 127 #ifndef MAX_PAGEOUT_CLUSTER 128 #define MAX_PAGEOUT_CLUSTER SWB_NPAGES 129 #endif 130 131 #define SWM_FREE 0x02 /* free, period */ 132 #define SWM_POP 0x04 /* pop out */ 133 134 #define SWBIO_READ 0x01 135 #define SWBIO_WRITE 0x02 136 #define SWBIO_SYNC 0x04 137 #define SWBIO_TTC 0x08 /* for VM_PAGER_TRY_TO_CACHE */ 138 139 struct swfreeinfo { 140 vm_object_t object; 141 vm_pindex_t basei; 142 vm_pindex_t begi; 143 vm_pindex_t endi; /* inclusive */ 144 }; 145 146 struct swswapoffinfo { 147 vm_object_t object; 148 int devidx; 149 int shared; 150 }; 151 152 /* 153 * vm_swap_size is in page-sized chunks now. It was DEV_BSIZE'd chunks 154 * in the old system. 155 */ 156 157 int swap_pager_full; /* swap space exhaustion (task killing) */ 158 int swap_fail_ticks; /* when we became exhausted */ 159 int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/ 160 swblk_t vm_swap_cache_use; 161 swblk_t vm_swap_anon_use; 162 static int vm_report_swap_allocs; 163 164 static int nsw_rcount; /* free read buffers */ 165 static int nsw_wcount_sync; /* limit write buffers / synchronous */ 166 static int nsw_wcount_async; /* limit write buffers / asynchronous */ 167 static int nsw_wcount_async_max;/* assigned maximum */ 168 static int nsw_cluster_max; /* maximum VOP I/O allowed */ 169 170 struct blist *swapblist; 171 static int swap_async_max = 4; /* maximum in-progress async I/O's */ 172 static int swap_burst_read = 0; /* allow burst reading */ 173 static swblk_t swapiterator; /* linearize allocations */ 174 int swap_user_async = 0; /* user swap pager operation can be async */ 175 176 static struct spinlock swapbp_spin = SPINLOCK_INITIALIZER(&swapbp_spin, "swapbp_spin"); 177 178 /* from vm_swap.c */ 179 extern struct vnode *swapdev_vp; 180 extern struct swdevt *swdevt; 181 extern int nswdev; 182 183 #define BLK2DEVIDX(blk) (nswdev > 1 ? blk / SWB_DMMAX % nswdev : 0) 184 185 SYSCTL_INT(_vm, OID_AUTO, swap_async_max, 186 CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops"); 187 SYSCTL_INT(_vm, OID_AUTO, swap_burst_read, 188 CTLFLAG_RW, &swap_burst_read, 0, "Allow burst reads for pageins"); 189 SYSCTL_INT(_vm, OID_AUTO, swap_user_async, 190 CTLFLAG_RW, &swap_user_async, 0, "Allow async uuser swap write I/O"); 191 192 #if SWBLK_BITS == 64 193 SYSCTL_LONG(_vm, OID_AUTO, swap_cache_use, 194 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 195 SYSCTL_LONG(_vm, OID_AUTO, swap_anon_use, 196 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 197 SYSCTL_LONG(_vm, OID_AUTO, swap_size, 198 CTLFLAG_RD, &vm_swap_size, 0, ""); 199 #else 200 SYSCTL_INT(_vm, OID_AUTO, swap_cache_use, 201 CTLFLAG_RD, &vm_swap_cache_use, 0, ""); 202 SYSCTL_INT(_vm, OID_AUTO, swap_anon_use, 203 CTLFLAG_RD, &vm_swap_anon_use, 0, ""); 204 SYSCTL_INT(_vm, OID_AUTO, swap_size, 205 CTLFLAG_RD, &vm_swap_size, 0, ""); 206 #endif 207 SYSCTL_INT(_vm, OID_AUTO, report_swap_allocs, 208 CTLFLAG_RW, &vm_report_swap_allocs, 0, ""); 209 210 vm_zone_t swap_zone; 211 212 /* 213 * Red-Black tree for swblock entries 214 * 215 * The caller must hold vm_token 216 */ 217 RB_GENERATE2(swblock_rb_tree, swblock, swb_entry, rb_swblock_compare, 218 vm_pindex_t, swb_index); 219 220 int 221 rb_swblock_compare(struct swblock *swb1, struct swblock *swb2) 222 { 223 if (swb1->swb_index < swb2->swb_index) 224 return(-1); 225 if (swb1->swb_index > swb2->swb_index) 226 return(1); 227 return(0); 228 } 229 230 static 231 int 232 rb_swblock_scancmp(struct swblock *swb, void *data) 233 { 234 struct swfreeinfo *info = data; 235 236 if (swb->swb_index < info->basei) 237 return(-1); 238 if (swb->swb_index > info->endi) 239 return(1); 240 return(0); 241 } 242 243 static 244 int 245 rb_swblock_condcmp(struct swblock *swb, void *data) 246 { 247 struct swfreeinfo *info = data; 248 249 if (swb->swb_index < info->basei) 250 return(-1); 251 return(0); 252 } 253 254 /* 255 * pagerops for OBJT_SWAP - "swap pager". Some ops are also global procedure 256 * calls hooked from other parts of the VM system and do not appear here. 257 * (see vm/swap_pager.h). 258 */ 259 260 static void swap_pager_dealloc (vm_object_t object); 261 static int swap_pager_getpage (vm_object_t, vm_page_t *, int); 262 static void swap_chain_iodone(struct bio *biox); 263 264 struct pagerops swappagerops = { 265 swap_pager_dealloc, /* deallocate an OBJT_SWAP object */ 266 swap_pager_getpage, /* pagein */ 267 swap_pager_putpages, /* pageout */ 268 swap_pager_haspage /* get backing store status for page */ 269 }; 270 271 /* 272 * SWB_DMMAX is in page-sized chunks with the new swap system. It was 273 * dev-bsized chunks in the old. SWB_DMMAX is always a power of 2. 274 * 275 * swap_*() routines are externally accessible. swp_*() routines are 276 * internal. 277 */ 278 279 int nswap_lowat = 128; /* in pages, swap_pager_almost_full warn */ 280 int nswap_hiwat = 512; /* in pages, swap_pager_almost_full warn */ 281 282 static __inline void swp_sizecheck (void); 283 static void swp_pager_async_iodone (struct bio *bio); 284 285 /* 286 * Swap bitmap functions 287 */ 288 289 static __inline void swp_pager_freeswapspace(vm_object_t object, 290 swblk_t blk, int npages); 291 static __inline swblk_t swp_pager_getswapspace(vm_object_t object, int npages); 292 293 /* 294 * Metadata functions 295 */ 296 297 static void swp_pager_meta_convert(vm_object_t); 298 static void swp_pager_meta_build(vm_object_t, vm_pindex_t, swblk_t); 299 static void swp_pager_meta_free(vm_object_t, vm_pindex_t, vm_pindex_t); 300 static void swp_pager_meta_free_all(vm_object_t); 301 static swblk_t swp_pager_meta_ctl(vm_object_t, vm_pindex_t, int); 302 303 /* 304 * SWP_SIZECHECK() - update swap_pager_full indication 305 * 306 * update the swap_pager_almost_full indication and warn when we are 307 * about to run out of swap space, using lowat/hiwat hysteresis. 308 * 309 * Clear swap_pager_full ( task killing ) indication when lowat is met. 310 * 311 * No restrictions on call 312 * This routine may not block. 313 * SMP races are ok. 314 */ 315 static __inline void 316 swp_sizecheck(void) 317 { 318 if (vm_swap_size < nswap_lowat) { 319 if (swap_pager_almost_full == 0) { 320 kprintf("swap_pager: out of swap space\n"); 321 swap_pager_almost_full = 1; 322 swap_fail_ticks = ticks; 323 } 324 } else { 325 swap_pager_full = 0; 326 if (vm_swap_size > nswap_hiwat) 327 swap_pager_almost_full = 0; 328 } 329 } 330 331 /* 332 * Long-term data collection on 10-second interval. Return the value 333 * for KCOLLECT_SWAPPCT and set the values for SWAPANO and SWAPCCAC. 334 * 335 * Return total swap in the scale field. This can change if swap is 336 * regularly added or removed and may cause some historical confusion 337 * in that case, but SWAPPCT will always be historically accurate. 338 */ 339 340 #define PTOB(value) ((uint64_t)(value) << PAGE_SHIFT) 341 342 static uint64_t 343 collect_swap_callback(int n) 344 { 345 uint64_t total = vm_swap_max; 346 uint64_t anon = vm_swap_anon_use; 347 uint64_t cache = vm_swap_cache_use; 348 349 if (total == 0) /* avoid divide by zero */ 350 total = 1; 351 kcollect_setvalue(KCOLLECT_SWAPANO, PTOB(anon)); 352 kcollect_setvalue(KCOLLECT_SWAPCAC, PTOB(cache)); 353 kcollect_setscale(KCOLLECT_SWAPANO, 354 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, PTOB(total))); 355 kcollect_setscale(KCOLLECT_SWAPCAC, 356 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, PTOB(total))); 357 return (((anon + cache) * 10000 + (total >> 1)) / total); 358 } 359 360 /* 361 * SWAP_PAGER_INIT() - initialize the swap pager! 362 * 363 * Expected to be started from system init. NOTE: This code is run 364 * before much else so be careful what you depend on. Most of the VM 365 * system has yet to be initialized at this point. 366 * 367 * Called from the low level boot code only. 368 */ 369 static void 370 swap_pager_init(void *arg __unused) 371 { 372 kcollect_register(KCOLLECT_SWAPPCT, "swapuse", collect_swap_callback, 373 KCOLLECT_SCALE(KCOLLECT_SWAPPCT_FORMAT, 0)); 374 kcollect_register(KCOLLECT_SWAPANO, "swapano", NULL, 375 KCOLLECT_SCALE(KCOLLECT_SWAPANO_FORMAT, 0)); 376 kcollect_register(KCOLLECT_SWAPCAC, "swapcac", NULL, 377 KCOLLECT_SCALE(KCOLLECT_SWAPCAC_FORMAT, 0)); 378 } 379 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_THIRD, swap_pager_init, NULL); 380 381 /* 382 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process 383 * 384 * Expected to be started from pageout process once, prior to entering 385 * its main loop. 386 * 387 * Called from the low level boot code only. 388 */ 389 void 390 swap_pager_swap_init(void) 391 { 392 int n, n2; 393 394 /* 395 * Number of in-transit swap bp operations. Don't 396 * exhaust the pbufs completely. Make sure we 397 * initialize workable values (0 will work for hysteresis 398 * but it isn't very efficient). 399 * 400 * The nsw_cluster_max is constrained by the number of pages an XIO 401 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined 402 * MAX_PAGEOUT_CLUSTER. Also be aware that swap ops are 403 * constrained by the swap device interleave stripe size. 404 * 405 * Currently we hardwire nsw_wcount_async to 4. This limit is 406 * designed to prevent other I/O from having high latencies due to 407 * our pageout I/O. The value 4 works well for one or two active swap 408 * devices but is probably a little low if you have more. Even so, 409 * a higher value would probably generate only a limited improvement 410 * with three or four active swap devices since the system does not 411 * typically have to pageout at extreme bandwidths. We will want 412 * at least 2 per swap devices, and 4 is a pretty good value if you 413 * have one NFS swap device due to the command/ack latency over NFS. 414 * So it all works out pretty well. 415 */ 416 417 nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER); 418 419 nsw_rcount = (nswbuf_kva + 1) / 2; 420 nsw_wcount_sync = (nswbuf_kva + 3) / 4; 421 nsw_wcount_async = 4; 422 nsw_wcount_async_max = nsw_wcount_async; 423 424 /* 425 * The zone is dynamically allocated so generally size it to 426 * maxswzone (32MB to 256GB of KVM). Set a minimum size based 427 * on physical memory of around 8x (each swblock can hold 16 pages). 428 * 429 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio 430 * has increased dramatically. 431 */ 432 n = vmstats.v_page_count / 2; 433 if (maxswzone && n < maxswzone / sizeof(struct swblock)) 434 n = maxswzone / sizeof(struct swblock); 435 n2 = n; 436 437 do { 438 swap_zone = zinit( 439 "SWAPMETA", 440 sizeof(struct swblock), 441 n, 442 ZONE_INTERRUPT); 443 if (swap_zone != NULL) 444 break; 445 /* 446 * if the allocation failed, try a zone two thirds the 447 * size of the previous attempt. 448 */ 449 n -= ((n + 2) / 3); 450 } while (n > 0); 451 452 if (swap_zone == NULL) 453 panic("swap_pager_swap_init: swap_zone == NULL"); 454 if (n2 != n) 455 kprintf("Swap zone entries reduced from %d to %d.\n", n2, n); 456 } 457 458 /* 459 * SWAP_PAGER_ALLOC() - allocate a new OBJT_SWAP VM object and instantiate 460 * its metadata structures. 461 * 462 * This routine is called from the mmap and fork code to create a new 463 * OBJT_SWAP object. We do this by creating an OBJT_DEFAULT object 464 * and then converting it with swp_pager_meta_convert(). 465 * 466 * We only support unnamed objects. 467 * 468 * No restrictions. 469 */ 470 vm_object_t 471 swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset) 472 { 473 vm_object_t object; 474 475 KKASSERT(handle == NULL); 476 object = vm_object_allocate_hold(OBJT_DEFAULT, 477 OFF_TO_IDX(offset + PAGE_MASK + size)); 478 swp_pager_meta_convert(object); 479 vm_object_drop(object); 480 481 return (object); 482 } 483 484 /* 485 * SWAP_PAGER_DEALLOC() - remove swap metadata from object 486 * 487 * The swap backing for the object is destroyed. The code is 488 * designed such that we can reinstantiate it later, but this 489 * routine is typically called only when the entire object is 490 * about to be destroyed. 491 * 492 * The object must be locked or unreferenceable. 493 * No other requirements. 494 */ 495 static void 496 swap_pager_dealloc(vm_object_t object) 497 { 498 vm_object_hold(object); 499 vm_object_pip_wait(object, "swpdea"); 500 501 /* 502 * Free all remaining metadata. We only bother to free it from 503 * the swap meta data. We do not attempt to free swapblk's still 504 * associated with vm_page_t's for this object. We do not care 505 * if paging is still in progress on some objects. 506 */ 507 swp_pager_meta_free_all(object); 508 vm_object_drop(object); 509 } 510 511 /************************************************************************ 512 * SWAP PAGER BITMAP ROUTINES * 513 ************************************************************************/ 514 515 /* 516 * SWP_PAGER_GETSWAPSPACE() - allocate raw swap space 517 * 518 * Allocate swap for the requested number of pages. The starting 519 * swap block number (a page index) is returned or SWAPBLK_NONE 520 * if the allocation failed. 521 * 522 * Also has the side effect of advising that somebody made a mistake 523 * when they configured swap and didn't configure enough. 524 * 525 * The caller must hold the object. 526 * This routine may not block. 527 */ 528 static __inline swblk_t 529 swp_pager_getswapspace(vm_object_t object, int npages) 530 { 531 swblk_t blk; 532 533 lwkt_gettoken(&vm_token); 534 blk = blist_allocat(swapblist, npages, swapiterator); 535 if (blk == SWAPBLK_NONE) 536 blk = blist_allocat(swapblist, npages, 0); 537 if (blk == SWAPBLK_NONE) { 538 if (swap_pager_full != 2) { 539 if (vm_swap_max == 0) 540 kprintf("Warning: The system would like to " 541 "page to swap but no swap space " 542 "is configured!\n"); 543 else 544 kprintf("swap_pager_getswapspace: " 545 "swap full allocating %d pages\n", 546 npages); 547 swap_pager_full = 2; 548 if (swap_pager_almost_full == 0) 549 swap_fail_ticks = ticks; 550 swap_pager_almost_full = 1; 551 } 552 } else { 553 /* swapiterator = blk; disable for now, doesn't work well */ 554 swapacctspace(blk, -npages); 555 if (object->type == OBJT_SWAP) 556 vm_swap_anon_use += npages; 557 else 558 vm_swap_cache_use += npages; 559 swp_sizecheck(); 560 } 561 lwkt_reltoken(&vm_token); 562 return(blk); 563 } 564 565 /* 566 * SWP_PAGER_FREESWAPSPACE() - free raw swap space 567 * 568 * This routine returns the specified swap blocks back to the bitmap. 569 * 570 * Note: This routine may not block (it could in the old swap code), 571 * and through the use of the new blist routines it does not block. 572 * 573 * This routine may not block. 574 */ 575 576 static __inline void 577 swp_pager_freeswapspace(vm_object_t object, swblk_t blk, int npages) 578 { 579 struct swdevt *sp = &swdevt[BLK2DEVIDX(blk)]; 580 581 lwkt_gettoken(&vm_token); 582 sp->sw_nused -= npages; 583 if (object->type == OBJT_SWAP) 584 vm_swap_anon_use -= npages; 585 else 586 vm_swap_cache_use -= npages; 587 588 if (sp->sw_flags & SW_CLOSING) { 589 lwkt_reltoken(&vm_token); 590 return; 591 } 592 593 blist_free(swapblist, blk, npages); 594 vm_swap_size += npages; 595 swp_sizecheck(); 596 lwkt_reltoken(&vm_token); 597 } 598 599 /* 600 * SWAP_PAGER_FREESPACE() - frees swap blocks associated with a page 601 * range within an object. 602 * 603 * This is a globally accessible routine. 604 * 605 * This routine removes swapblk assignments from swap metadata. 606 * 607 * The external callers of this routine typically have already destroyed 608 * or renamed vm_page_t's associated with this range in the object so 609 * we should be ok. 610 * 611 * No requirements. 612 */ 613 void 614 swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_pindex_t size) 615 { 616 vm_object_hold(object); 617 swp_pager_meta_free(object, start, size); 618 vm_object_drop(object); 619 } 620 621 /* 622 * No requirements. 623 */ 624 void 625 swap_pager_freespace_all(vm_object_t object) 626 { 627 vm_object_hold(object); 628 swp_pager_meta_free_all(object); 629 vm_object_drop(object); 630 } 631 632 /* 633 * This function conditionally frees swap cache swap starting at 634 * (*basei) in the object. (count) swap blocks will be nominally freed. 635 * The actual number of blocks freed can be more or less than the 636 * requested number. 637 * 638 * This function nominally returns the number of blocks freed. However, 639 * the actual number of blocks freed may be less then the returned value. 640 * If the function is unable to exhaust the object or if it is able to 641 * free (approximately) the requested number of blocks it returns 642 * a value n > count. 643 * 644 * If we exhaust the object we will return a value n <= count. 645 * 646 * The caller must hold the object. 647 * 648 * WARNING! If count == 0 then -1 can be returned as a degenerate case, 649 * callers should always pass a count value > 0. 650 */ 651 static int swap_pager_condfree_callback(struct swblock *swap, void *data); 652 653 int 654 swap_pager_condfree(vm_object_t object, vm_pindex_t *basei, int count) 655 { 656 struct swfreeinfo info; 657 int n; 658 int t; 659 660 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 661 662 info.object = object; 663 info.basei = *basei; /* skip up to this page index */ 664 info.begi = count; /* max swap pages to destroy */ 665 info.endi = count * 8; /* max swblocks to scan */ 666 667 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_condcmp, 668 swap_pager_condfree_callback, &info); 669 *basei = info.basei; 670 671 /* 672 * Take the higher difference swblocks vs pages 673 */ 674 n = count - (int)info.begi; 675 t = count * 8 - (int)info.endi; 676 if (n < t) 677 n = t; 678 if (n < 1) 679 n = 1; 680 return(n); 681 } 682 683 /* 684 * The idea is to free whole meta-block to avoid fragmenting 685 * the swap space or disk I/O. We only do this if NO VM pages 686 * are present. 687 * 688 * We do not have to deal with clearing PG_SWAPPED in related VM 689 * pages because there are no related VM pages. 690 * 691 * The caller must hold the object. 692 */ 693 static int 694 swap_pager_condfree_callback(struct swblock *swap, void *data) 695 { 696 struct swfreeinfo *info = data; 697 vm_object_t object = info->object; 698 int i; 699 700 for (i = 0; i < SWAP_META_PAGES; ++i) { 701 if (vm_page_lookup(object, swap->swb_index + i)) 702 break; 703 } 704 info->basei = swap->swb_index + SWAP_META_PAGES; 705 if (i == SWAP_META_PAGES) { 706 info->begi -= swap->swb_count; 707 swap_pager_freespace(object, swap->swb_index, SWAP_META_PAGES); 708 } 709 --info->endi; 710 if ((int)info->begi < 0 || (int)info->endi < 0) 711 return(-1); 712 lwkt_yield(); 713 return(0); 714 } 715 716 /* 717 * Called by vm_page_alloc() when a new VM page is inserted 718 * into a VM object. Checks whether swap has been assigned to 719 * the page and sets PG_SWAPPED as necessary. 720 * 721 * (m) must be busied by caller and remains busied on return. 722 */ 723 void 724 swap_pager_page_inserted(vm_page_t m) 725 { 726 if (m->object->swblock_count) { 727 vm_object_hold(m->object); 728 if (swp_pager_meta_ctl(m->object, m->pindex, 0) != SWAPBLK_NONE) 729 vm_page_flag_set(m, PG_SWAPPED); 730 vm_object_drop(m->object); 731 } 732 } 733 734 /* 735 * SWAP_PAGER_RESERVE() - reserve swap blocks in object 736 * 737 * Assigns swap blocks to the specified range within the object. The 738 * swap blocks are not zerod. Any previous swap assignment is destroyed. 739 * 740 * Returns 0 on success, -1 on failure. 741 * 742 * The caller is responsible for avoiding races in the specified range. 743 * No other requirements. 744 */ 745 int 746 swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size) 747 { 748 int n = 0; 749 swblk_t blk = SWAPBLK_NONE; 750 vm_pindex_t beg = start; /* save start index */ 751 752 vm_object_hold(object); 753 754 while (size) { 755 if (n == 0) { 756 n = BLIST_MAX_ALLOC; 757 while ((blk = swp_pager_getswapspace(object, n)) == 758 SWAPBLK_NONE) 759 { 760 n >>= 1; 761 if (n == 0) { 762 swp_pager_meta_free(object, beg, 763 start - beg); 764 vm_object_drop(object); 765 return(-1); 766 } 767 } 768 } 769 swp_pager_meta_build(object, start, blk); 770 --size; 771 ++start; 772 ++blk; 773 --n; 774 } 775 swp_pager_meta_free(object, start, n); 776 vm_object_drop(object); 777 return(0); 778 } 779 780 /* 781 * SWAP_PAGER_COPY() - copy blocks from source pager to destination pager 782 * and destroy the source. 783 * 784 * Copy any valid swapblks from the source to the destination. In 785 * cases where both the source and destination have a valid swapblk, 786 * we keep the destination's. 787 * 788 * This routine is allowed to block. It may block allocating metadata 789 * indirectly through swp_pager_meta_build() or if paging is still in 790 * progress on the source. 791 * 792 * XXX vm_page_collapse() kinda expects us not to block because we 793 * supposedly do not need to allocate memory, but for the moment we 794 * *may* have to get a little memory from the zone allocator, but 795 * it is taken from the interrupt memory. We should be ok. 796 * 797 * The source object contains no vm_page_t's (which is just as well) 798 * The source object is of type OBJT_SWAP. 799 * 800 * The source and destination objects must be held by the caller. 801 */ 802 void 803 swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject, 804 vm_pindex_t base_index, int destroysource) 805 { 806 vm_pindex_t i; 807 808 ASSERT_LWKT_TOKEN_HELD(vm_object_token(srcobject)); 809 ASSERT_LWKT_TOKEN_HELD(vm_object_token(dstobject)); 810 811 /* 812 * transfer source to destination. 813 */ 814 for (i = 0; i < dstobject->size; ++i) { 815 swblk_t dstaddr; 816 817 /* 818 * Locate (without changing) the swapblk on the destination, 819 * unless it is invalid in which case free it silently, or 820 * if the destination is a resident page, in which case the 821 * source is thrown away. 822 */ 823 dstaddr = swp_pager_meta_ctl(dstobject, i, 0); 824 825 if (dstaddr == SWAPBLK_NONE) { 826 /* 827 * Destination has no swapblk and is not resident, 828 * copy source. 829 */ 830 swblk_t srcaddr; 831 832 srcaddr = swp_pager_meta_ctl(srcobject, 833 base_index + i, SWM_POP); 834 835 if (srcaddr != SWAPBLK_NONE) 836 swp_pager_meta_build(dstobject, i, srcaddr); 837 } else { 838 /* 839 * Destination has valid swapblk or it is represented 840 * by a resident page. We destroy the sourceblock. 841 */ 842 swp_pager_meta_ctl(srcobject, base_index + i, SWM_FREE); 843 } 844 } 845 846 /* 847 * Free left over swap blocks in source. 848 * 849 * We have to revert the type to OBJT_DEFAULT so we do not accidently 850 * double-remove the object from the swap queues. 851 */ 852 if (destroysource) { 853 /* 854 * Reverting the type is not necessary, the caller is going 855 * to destroy srcobject directly, but I'm doing it here 856 * for consistency since we've removed the object from its 857 * queues. 858 */ 859 swp_pager_meta_free_all(srcobject); 860 if (srcobject->type == OBJT_SWAP) 861 srcobject->type = OBJT_DEFAULT; 862 } 863 } 864 865 /* 866 * SWAP_PAGER_HASPAGE() - determine if we have good backing store for 867 * the requested page. 868 * 869 * We determine whether good backing store exists for the requested 870 * page and return TRUE if it does, FALSE if it doesn't. 871 * 872 * If TRUE, we also try to determine how much valid, contiguous backing 873 * store exists before and after the requested page within a reasonable 874 * distance. We do not try to restrict it to the swap device stripe 875 * (that is handled in getpages/putpages). It probably isn't worth 876 * doing here. 877 * 878 * No requirements. 879 */ 880 boolean_t 881 swap_pager_haspage(vm_object_t object, vm_pindex_t pindex) 882 { 883 swblk_t blk0; 884 885 /* 886 * do we have good backing store at the requested index ? 887 */ 888 vm_object_hold(object); 889 blk0 = swp_pager_meta_ctl(object, pindex, 0); 890 891 if (blk0 == SWAPBLK_NONE) { 892 vm_object_drop(object); 893 return (FALSE); 894 } 895 vm_object_drop(object); 896 return (TRUE); 897 } 898 899 /* 900 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page 901 * 902 * This removes any associated swap backing store, whether valid or 903 * not, from the page. This operates on any VM object, not just OBJT_SWAP 904 * objects. 905 * 906 * This routine is typically called when a page is made dirty, at 907 * which point any associated swap can be freed. MADV_FREE also 908 * calls us in a special-case situation 909 * 910 * NOTE!!! If the page is clean and the swap was valid, the caller 911 * should make the page dirty before calling this routine. 912 * This routine does NOT change the m->dirty status of the page. 913 * Also: MADV_FREE depends on it. 914 * 915 * The page must be busied. 916 * The caller can hold the object to avoid blocking, else we might block. 917 * No other requirements. 918 */ 919 void 920 swap_pager_unswapped(vm_page_t m) 921 { 922 if (m->flags & PG_SWAPPED) { 923 vm_object_hold(m->object); 924 KKASSERT(m->flags & PG_SWAPPED); 925 swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE); 926 vm_page_flag_clear(m, PG_SWAPPED); 927 vm_object_drop(m->object); 928 } 929 } 930 931 /* 932 * SWAP_PAGER_STRATEGY() - read, write, free blocks 933 * 934 * This implements a VM OBJECT strategy function using swap backing store. 935 * This can operate on any VM OBJECT type, not necessarily just OBJT_SWAP 936 * types. 937 * 938 * This is intended to be a cacheless interface (i.e. caching occurs at 939 * higher levels), and is also used as a swap-based SSD cache for vnode 940 * and device objects. 941 * 942 * All I/O goes directly to and from the swap device. 943 * 944 * We currently attempt to run I/O synchronously or asynchronously as 945 * the caller requests. This isn't perfect because we loose error 946 * sequencing when we run multiple ops in parallel to satisfy a request. 947 * But this is swap, so we let it all hang out. 948 * 949 * No requirements. 950 */ 951 void 952 swap_pager_strategy(vm_object_t object, struct bio *bio) 953 { 954 struct buf *bp = bio->bio_buf; 955 struct bio *nbio; 956 vm_pindex_t start; 957 vm_pindex_t biox_blkno = 0; 958 int count; 959 char *data; 960 struct bio *biox; 961 struct buf *bufx; 962 #if 0 963 struct bio_track *track; 964 #endif 965 966 #if 0 967 /* 968 * tracking for swapdev vnode I/Os 969 */ 970 if (bp->b_cmd == BUF_CMD_READ) 971 track = &swapdev_vp->v_track_read; 972 else 973 track = &swapdev_vp->v_track_write; 974 #endif 975 976 if (bp->b_bcount & PAGE_MASK) { 977 bp->b_error = EINVAL; 978 bp->b_flags |= B_ERROR | B_INVAL; 979 biodone(bio); 980 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 981 "not page bounded\n", 982 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 983 return; 984 } 985 986 /* 987 * Clear error indication, initialize page index, count, data pointer. 988 */ 989 bp->b_error = 0; 990 bp->b_flags &= ~B_ERROR; 991 bp->b_resid = bp->b_bcount; 992 993 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 994 count = howmany(bp->b_bcount, PAGE_SIZE); 995 data = bp->b_data; 996 997 /* 998 * Deal with BUF_CMD_FREEBLKS 999 */ 1000 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 1001 /* 1002 * FREE PAGE(s) - destroy underlying swap that is no longer 1003 * needed. 1004 */ 1005 vm_object_hold(object); 1006 swp_pager_meta_free(object, start, count); 1007 vm_object_drop(object); 1008 bp->b_resid = 0; 1009 biodone(bio); 1010 return; 1011 } 1012 1013 /* 1014 * We need to be able to create a new cluster of I/O's. We cannot 1015 * use the caller fields of the passed bio so push a new one. 1016 * 1017 * Because nbio is just a placeholder for the cluster links, 1018 * we can biodone() the original bio instead of nbio to make 1019 * things a bit more efficient. 1020 */ 1021 nbio = push_bio(bio); 1022 nbio->bio_offset = bio->bio_offset; 1023 nbio->bio_caller_info1.cluster_head = NULL; 1024 nbio->bio_caller_info2.cluster_tail = NULL; 1025 1026 biox = NULL; 1027 bufx = NULL; 1028 1029 /* 1030 * Execute read or write 1031 */ 1032 vm_object_hold(object); 1033 1034 while (count > 0) { 1035 swblk_t blk; 1036 1037 /* 1038 * Obtain block. If block not found and writing, allocate a 1039 * new block and build it into the object. 1040 */ 1041 blk = swp_pager_meta_ctl(object, start, 0); 1042 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) { 1043 blk = swp_pager_getswapspace(object, 1); 1044 if (blk == SWAPBLK_NONE) { 1045 bp->b_error = ENOMEM; 1046 bp->b_flags |= B_ERROR; 1047 break; 1048 } 1049 swp_pager_meta_build(object, start, blk); 1050 } 1051 1052 /* 1053 * Do we have to flush our current collection? Yes if: 1054 * 1055 * - no swap block at this index 1056 * - swap block is not contiguous 1057 * - we cross a physical disk boundry in the 1058 * stripe. 1059 */ 1060 if ( 1061 biox && (biox_blkno + btoc(bufx->b_bcount) != blk || 1062 ((biox_blkno ^ blk) & ~SWB_DMMASK) 1063 ) 1064 ) { 1065 if (bp->b_cmd == BUF_CMD_READ) { 1066 ++mycpu->gd_cnt.v_swapin; 1067 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1068 } else { 1069 ++mycpu->gd_cnt.v_swapout; 1070 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1071 bufx->b_dirtyend = bufx->b_bcount; 1072 } 1073 1074 /* 1075 * Finished with this buf. 1076 */ 1077 KKASSERT(bufx->b_bcount != 0); 1078 if (bufx->b_cmd != BUF_CMD_READ) 1079 bufx->b_dirtyend = bufx->b_bcount; 1080 biox = NULL; 1081 bufx = NULL; 1082 } 1083 1084 /* 1085 * Add new swapblk to biox, instantiating biox if necessary. 1086 * Zero-fill reads are able to take a shortcut. 1087 */ 1088 if (blk == SWAPBLK_NONE) { 1089 /* 1090 * We can only get here if we are reading. 1091 */ 1092 bzero(data, PAGE_SIZE); 1093 bp->b_resid -= PAGE_SIZE; 1094 } else { 1095 if (biox == NULL) { 1096 /* XXX chain count > 4, wait to <= 4 */ 1097 1098 bufx = getpbuf(NULL); 1099 biox = &bufx->b_bio1; 1100 cluster_append(nbio, bufx); 1101 bufx->b_cmd = bp->b_cmd; 1102 biox->bio_done = swap_chain_iodone; 1103 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1104 biox->bio_caller_info1.cluster_parent = nbio; 1105 biox_blkno = blk; 1106 bufx->b_bcount = 0; 1107 bufx->b_data = data; 1108 } 1109 bufx->b_bcount += PAGE_SIZE; 1110 } 1111 --count; 1112 ++start; 1113 data += PAGE_SIZE; 1114 } 1115 1116 vm_object_drop(object); 1117 1118 /* 1119 * Flush out last buffer 1120 */ 1121 if (biox) { 1122 if (bufx->b_cmd == BUF_CMD_READ) { 1123 ++mycpu->gd_cnt.v_swapin; 1124 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1125 } else { 1126 ++mycpu->gd_cnt.v_swapout; 1127 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1128 bufx->b_dirtyend = bufx->b_bcount; 1129 } 1130 KKASSERT(bufx->b_bcount); 1131 if (bufx->b_cmd != BUF_CMD_READ) 1132 bufx->b_dirtyend = bufx->b_bcount; 1133 /* biox, bufx = NULL */ 1134 } 1135 1136 /* 1137 * Now initiate all the I/O. Be careful looping on our chain as 1138 * I/O's may complete while we are still initiating them. 1139 * 1140 * If the request is a 100% sparse read no bios will be present 1141 * and we just biodone() the buffer. 1142 */ 1143 nbio->bio_caller_info2.cluster_tail = NULL; 1144 bufx = nbio->bio_caller_info1.cluster_head; 1145 1146 if (bufx) { 1147 while (bufx) { 1148 biox = &bufx->b_bio1; 1149 BUF_KERNPROC(bufx); 1150 bufx = bufx->b_cluster_next; 1151 vn_strategy(swapdev_vp, biox); 1152 } 1153 } else { 1154 biodone(bio); 1155 } 1156 1157 /* 1158 * Completion of the cluster will also call biodone_chain(nbio). 1159 * We never call biodone(nbio) so we don't have to worry about 1160 * setting up a bio_done callback. It's handled in the sub-IO. 1161 */ 1162 /**/ 1163 } 1164 1165 /* 1166 * biodone callback 1167 * 1168 * No requirements. 1169 */ 1170 static void 1171 swap_chain_iodone(struct bio *biox) 1172 { 1173 struct buf **nextp; 1174 struct buf *bufx; /* chained sub-buffer */ 1175 struct bio *nbio; /* parent nbio with chain glue */ 1176 struct buf *bp; /* original bp associated with nbio */ 1177 int chain_empty; 1178 1179 bufx = biox->bio_buf; 1180 nbio = biox->bio_caller_info1.cluster_parent; 1181 bp = nbio->bio_buf; 1182 1183 /* 1184 * Update the original buffer 1185 */ 1186 KKASSERT(bp != NULL); 1187 if (bufx->b_flags & B_ERROR) { 1188 atomic_set_int(&bufx->b_flags, B_ERROR); 1189 bp->b_error = bufx->b_error; /* race ok */ 1190 } else if (bufx->b_resid != 0) { 1191 atomic_set_int(&bufx->b_flags, B_ERROR); 1192 bp->b_error = EINVAL; /* race ok */ 1193 } else { 1194 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1195 } 1196 1197 /* 1198 * Remove us from the chain. 1199 */ 1200 spin_lock(&swapbp_spin); 1201 nextp = &nbio->bio_caller_info1.cluster_head; 1202 while (*nextp != bufx) { 1203 KKASSERT(*nextp != NULL); 1204 nextp = &(*nextp)->b_cluster_next; 1205 } 1206 *nextp = bufx->b_cluster_next; 1207 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1208 spin_unlock(&swapbp_spin); 1209 1210 /* 1211 * Clean up bufx. If the chain is now empty we finish out 1212 * the parent. Note that we may be racing other completions 1213 * so we must use the chain_empty status from above. 1214 */ 1215 if (chain_empty) { 1216 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1217 atomic_set_int(&bp->b_flags, B_ERROR); 1218 bp->b_error = EINVAL; 1219 } 1220 biodone_chain(nbio); 1221 } 1222 relpbuf(bufx, NULL); 1223 } 1224 1225 /* 1226 * SWAP_PAGER_GETPAGES() - bring page in from swap 1227 * 1228 * The requested page may have to be brought in from swap. Calculate the 1229 * swap block and bring in additional pages if possible. All pages must 1230 * have contiguous swap block assignments and reside in the same object. 1231 * 1232 * The caller has a single vm_object_pip_add() reference prior to 1233 * calling us and we should return with the same. 1234 * 1235 * The caller has BUSY'd the page. We should return with (*mpp) left busy, 1236 * and any additinal pages unbusied. 1237 * 1238 * If the caller encounters a PG_RAM page it will pass it to us even though 1239 * it may be valid and dirty. We cannot overwrite the page in this case! 1240 * The case is used to allow us to issue pure read-aheads. 1241 * 1242 * NOTE! XXX This code does not entirely pipeline yet due to the fact that 1243 * the PG_RAM page is validated at the same time as mreq. What we 1244 * really need to do is issue a separate read-ahead pbuf. 1245 * 1246 * No requirements. 1247 */ 1248 static int 1249 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 1250 { 1251 struct buf *bp; 1252 struct bio *bio; 1253 vm_page_t mreq; 1254 vm_page_t m; 1255 vm_offset_t kva; 1256 swblk_t blk; 1257 int i; 1258 int j; 1259 int raonly; 1260 int error; 1261 u_int32_t flags; 1262 vm_page_t marray[XIO_INTERNAL_PAGES]; 1263 1264 mreq = *mpp; 1265 1266 vm_object_hold(object); 1267 if (mreq->object != object) { 1268 panic("swap_pager_getpages: object mismatch %p/%p", 1269 object, 1270 mreq->object 1271 ); 1272 } 1273 1274 /* 1275 * We don't want to overwrite a fully valid page as it might be 1276 * dirty. This case can occur when e.g. vm_fault hits a perfectly 1277 * valid page with PG_RAM set. 1278 * 1279 * In this case we see if the next page is a suitable page-in 1280 * candidate and if it is we issue read-ahead. PG_RAM will be 1281 * set on the last page of the read-ahead to continue the pipeline. 1282 */ 1283 if (mreq->valid == VM_PAGE_BITS_ALL) { 1284 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) { 1285 vm_object_drop(object); 1286 return(VM_PAGER_OK); 1287 } 1288 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0); 1289 if (blk == SWAPBLK_NONE) { 1290 vm_object_drop(object); 1291 return(VM_PAGER_OK); 1292 } 1293 m = vm_page_lookup_busy_try(object, mreq->pindex + 1, 1294 TRUE, &error); 1295 if (error) { 1296 vm_object_drop(object); 1297 return(VM_PAGER_OK); 1298 } else if (m == NULL) { 1299 /* 1300 * Use VM_ALLOC_QUICK to avoid blocking on cache 1301 * page reuse. 1302 */ 1303 m = vm_page_alloc(object, mreq->pindex + 1, 1304 VM_ALLOC_QUICK); 1305 if (m == NULL) { 1306 vm_object_drop(object); 1307 return(VM_PAGER_OK); 1308 } 1309 } else { 1310 if (m->valid) { 1311 vm_page_wakeup(m); 1312 vm_object_drop(object); 1313 return(VM_PAGER_OK); 1314 } 1315 vm_page_unqueue_nowakeup(m); 1316 } 1317 /* page is busy */ 1318 mreq = m; 1319 raonly = 1; 1320 } else { 1321 raonly = 0; 1322 } 1323 1324 /* 1325 * Try to block-read contiguous pages from swap if sequential, 1326 * otherwise just read one page. Contiguous pages from swap must 1327 * reside within a single device stripe because the I/O cannot be 1328 * broken up across multiple stripes. 1329 * 1330 * Note that blk and iblk can be SWAPBLK_NONE but the loop is 1331 * set up such that the case(s) are handled implicitly. 1332 */ 1333 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1334 marray[0] = mreq; 1335 1336 for (i = 1; i <= swap_burst_read && 1337 i < XIO_INTERNAL_PAGES && 1338 mreq->pindex + i < object->size; ++i) { 1339 swblk_t iblk; 1340 1341 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0); 1342 if (iblk != blk + i) 1343 break; 1344 if ((blk ^ iblk) & ~SWB_DMMASK) 1345 break; 1346 m = vm_page_lookup_busy_try(object, mreq->pindex + i, 1347 TRUE, &error); 1348 if (error) { 1349 break; 1350 } else if (m == NULL) { 1351 /* 1352 * Use VM_ALLOC_QUICK to avoid blocking on cache 1353 * page reuse. 1354 */ 1355 m = vm_page_alloc(object, mreq->pindex + i, 1356 VM_ALLOC_QUICK); 1357 if (m == NULL) 1358 break; 1359 } else { 1360 if (m->valid) { 1361 vm_page_wakeup(m); 1362 break; 1363 } 1364 vm_page_unqueue_nowakeup(m); 1365 } 1366 /* page is busy */ 1367 marray[i] = m; 1368 } 1369 if (i > 1) 1370 vm_page_flag_set(marray[i - 1], PG_RAM); 1371 1372 /* 1373 * If mreq is the requested page and we have nothing to do return 1374 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead 1375 * page and must be cleaned up. 1376 */ 1377 if (blk == SWAPBLK_NONE) { 1378 KKASSERT(i == 1); 1379 if (raonly) { 1380 vnode_pager_freepage(mreq); 1381 vm_object_drop(object); 1382 return(VM_PAGER_OK); 1383 } else { 1384 vm_object_drop(object); 1385 return(VM_PAGER_FAIL); 1386 } 1387 } 1388 1389 /* 1390 * map our page(s) into kva for input 1391 */ 1392 bp = getpbuf_kva(&nsw_rcount); 1393 bio = &bp->b_bio1; 1394 kva = (vm_offset_t) bp->b_kvabase; 1395 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t)); 1396 pmap_qenter(kva, bp->b_xio.xio_pages, i); 1397 1398 bp->b_data = (caddr_t)kva; 1399 bp->b_bcount = PAGE_SIZE * i; 1400 bp->b_xio.xio_npages = i; 1401 bio->bio_done = swp_pager_async_iodone; 1402 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1403 bio->bio_caller_info1.index = SWBIO_READ; 1404 1405 /* 1406 * Set index. If raonly set the index beyond the array so all 1407 * the pages are treated the same, otherwise the original mreq is 1408 * at index 0. 1409 */ 1410 if (raonly) 1411 bio->bio_driver_info = (void *)(intptr_t)i; 1412 else 1413 bio->bio_driver_info = (void *)(intptr_t)0; 1414 1415 for (j = 0; j < i; ++j) 1416 vm_page_flag_set(bp->b_xio.xio_pages[j], PG_SWAPINPROG); 1417 1418 mycpu->gd_cnt.v_swapin++; 1419 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1420 1421 /* 1422 * We still hold the lock on mreq, and our automatic completion routine 1423 * does not remove it. 1424 */ 1425 vm_object_pip_add(object, bp->b_xio.xio_npages); 1426 1427 /* 1428 * perform the I/O. NOTE!!! bp cannot be considered valid after 1429 * this point because we automatically release it on completion. 1430 * Instead, we look at the one page we are interested in which we 1431 * still hold a lock on even through the I/O completion. 1432 * 1433 * The other pages in our m[] array are also released on completion, 1434 * so we cannot assume they are valid anymore either. 1435 */ 1436 bp->b_cmd = BUF_CMD_READ; 1437 BUF_KERNPROC(bp); 1438 vn_strategy(swapdev_vp, bio); 1439 1440 /* 1441 * Wait for the page we want to complete. PG_SWAPINPROG is always 1442 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1443 * is set in the meta-data. 1444 * 1445 * If this is a read-ahead only we return immediately without 1446 * waiting for I/O. 1447 */ 1448 if (raonly) { 1449 vm_object_drop(object); 1450 return(VM_PAGER_OK); 1451 } 1452 1453 /* 1454 * Read-ahead includes originally requested page case. 1455 */ 1456 for (;;) { 1457 flags = mreq->flags; 1458 cpu_ccfence(); 1459 if ((flags & PG_SWAPINPROG) == 0) 1460 break; 1461 tsleep_interlock(mreq, 0); 1462 if (!atomic_cmpset_int(&mreq->flags, flags, 1463 flags | PG_WANTED | PG_REFERENCED)) { 1464 continue; 1465 } 1466 mycpu->gd_cnt.v_intrans++; 1467 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) { 1468 kprintf( 1469 "swap_pager: indefinite wait buffer: " 1470 " bp %p offset: %lld, size: %ld\n", 1471 bp, 1472 (long long)bio->bio_offset, 1473 (long)bp->b_bcount 1474 ); 1475 } 1476 } 1477 1478 /* 1479 * Disallow speculative reads prior to the PG_SWAPINPROG test. 1480 */ 1481 cpu_lfence(); 1482 1483 /* 1484 * mreq is left busied after completion, but all the other pages 1485 * are freed. If we had an unrecoverable read error the page will 1486 * not be valid. 1487 */ 1488 vm_object_drop(object); 1489 if (mreq->valid != VM_PAGE_BITS_ALL) 1490 return(VM_PAGER_ERROR); 1491 else 1492 return(VM_PAGER_OK); 1493 1494 /* 1495 * A final note: in a low swap situation, we cannot deallocate swap 1496 * and mark a page dirty here because the caller is likely to mark 1497 * the page clean when we return, causing the page to possibly revert 1498 * to all-zero's later. 1499 */ 1500 } 1501 1502 /* 1503 * swap_pager_putpages: 1504 * 1505 * Assign swap (if necessary) and initiate I/O on the specified pages. 1506 * 1507 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1508 * are automatically converted to SWAP objects. 1509 * 1510 * In a low memory situation we may block in vn_strategy(), but the new 1511 * vm_page reservation system coupled with properly written VFS devices 1512 * should ensure that no low-memory deadlock occurs. This is an area 1513 * which needs work. 1514 * 1515 * The parent has N vm_object_pip_add() references prior to 1516 * calling us and will remove references for rtvals[] that are 1517 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1518 * completion. 1519 * 1520 * The parent has soft-busy'd the pages it passes us and will unbusy 1521 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1522 * We need to unbusy the rest on I/O completion. 1523 * 1524 * No requirements. 1525 */ 1526 void 1527 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1528 int flags, int *rtvals) 1529 { 1530 int i; 1531 int n = 0; 1532 1533 vm_object_hold(object); 1534 1535 if (count && m[0]->object != object) { 1536 panic("swap_pager_getpages: object mismatch %p/%p", 1537 object, 1538 m[0]->object 1539 ); 1540 } 1541 1542 /* 1543 * Step 1 1544 * 1545 * Turn object into OBJT_SWAP 1546 * Check for bogus sysops 1547 * 1548 * Force sync if not pageout process, we don't want any single 1549 * non-pageout process to be able to hog the I/O subsystem! This 1550 * can be overridden by setting. 1551 */ 1552 if (object->type == OBJT_DEFAULT) { 1553 if (object->type == OBJT_DEFAULT) 1554 swp_pager_meta_convert(object); 1555 } 1556 1557 /* 1558 * Normally we force synchronous swap I/O if this is not the 1559 * pageout daemon to prevent any single user process limited 1560 * via RLIMIT_RSS from hogging swap write bandwidth. 1561 */ 1562 if (curthread != pagethread && 1563 curthread != emergpager && 1564 swap_user_async == 0) { 1565 flags |= VM_PAGER_PUT_SYNC; 1566 } 1567 1568 /* 1569 * Step 2 1570 * 1571 * Update nsw parameters from swap_async_max sysctl values. 1572 * Do not let the sysop crash the machine with bogus numbers. 1573 */ 1574 if (swap_async_max != nsw_wcount_async_max) { 1575 int n; 1576 1577 /* 1578 * limit range 1579 */ 1580 if ((n = swap_async_max) > nswbuf_kva / 2) 1581 n = nswbuf_kva / 2; 1582 if (n < 1) 1583 n = 1; 1584 swap_async_max = n; 1585 1586 /* 1587 * Adjust difference ( if possible ). If the current async 1588 * count is too low, we may not be able to make the adjustment 1589 * at this time. 1590 * 1591 * vm_token needed for nsw_wcount sleep interlock 1592 */ 1593 lwkt_gettoken(&vm_token); 1594 n -= nsw_wcount_async_max; 1595 if (nsw_wcount_async + n >= 0) { 1596 nsw_wcount_async_max += n; 1597 pbuf_adjcount(&nsw_wcount_async, n); 1598 } 1599 lwkt_reltoken(&vm_token); 1600 } 1601 1602 /* 1603 * Step 3 1604 * 1605 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1606 * The page is left dirty until the pageout operation completes 1607 * successfully. 1608 */ 1609 1610 for (i = 0; i < count; i += n) { 1611 struct buf *bp; 1612 struct bio *bio; 1613 swblk_t blk; 1614 int j; 1615 1616 /* 1617 * Maximum I/O size is limited by a number of factors. 1618 */ 1619 1620 n = min(BLIST_MAX_ALLOC, count - i); 1621 n = min(n, nsw_cluster_max); 1622 1623 lwkt_gettoken(&vm_token); 1624 1625 /* 1626 * Get biggest block of swap we can. If we fail, fall 1627 * back and try to allocate a smaller block. Don't go 1628 * overboard trying to allocate space if it would overly 1629 * fragment swap. 1630 */ 1631 while ( 1632 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE && 1633 n > 4 1634 ) { 1635 n >>= 1; 1636 } 1637 if (blk == SWAPBLK_NONE) { 1638 for (j = 0; j < n; ++j) 1639 rtvals[i+j] = VM_PAGER_FAIL; 1640 lwkt_reltoken(&vm_token); 1641 continue; 1642 } 1643 if (vm_report_swap_allocs > 0) { 1644 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n); 1645 --vm_report_swap_allocs; 1646 } 1647 1648 /* 1649 * The I/O we are constructing cannot cross a physical 1650 * disk boundry in the swap stripe. 1651 */ 1652 if ((blk ^ (blk + n)) & ~SWB_DMMASK) { 1653 j = ((blk + SWB_DMMAX) & ~SWB_DMMASK) - blk; 1654 swp_pager_freeswapspace(object, blk + j, n - j); 1655 n = j; 1656 } 1657 1658 /* 1659 * All I/O parameters have been satisfied, build the I/O 1660 * request and assign the swap space. 1661 */ 1662 if ((flags & VM_PAGER_PUT_SYNC)) 1663 bp = getpbuf_kva(&nsw_wcount_sync); 1664 else 1665 bp = getpbuf_kva(&nsw_wcount_async); 1666 bio = &bp->b_bio1; 1667 1668 lwkt_reltoken(&vm_token); 1669 1670 pmap_qenter((vm_offset_t)bp->b_data, &m[i], n); 1671 1672 bp->b_bcount = PAGE_SIZE * n; 1673 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1674 1675 for (j = 0; j < n; ++j) { 1676 vm_page_t mreq = m[i+j]; 1677 1678 swp_pager_meta_build(mreq->object, mreq->pindex, 1679 blk + j); 1680 if (object->type == OBJT_SWAP) 1681 vm_page_dirty(mreq); 1682 rtvals[i+j] = VM_PAGER_OK; 1683 1684 vm_page_flag_set(mreq, PG_SWAPINPROG); 1685 bp->b_xio.xio_pages[j] = mreq; 1686 } 1687 bp->b_xio.xio_npages = n; 1688 1689 mycpu->gd_cnt.v_swapout++; 1690 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1691 1692 bp->b_dirtyoff = 0; /* req'd for NFS */ 1693 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1694 bp->b_cmd = BUF_CMD_WRITE; 1695 bio->bio_caller_info1.index = SWBIO_WRITE; 1696 1697 #if 0 1698 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */ 1699 bio->bio_crc = iscsi_crc32(bp->b_data, bp->b_bcount); 1700 { 1701 uint32_t crc = 0; 1702 for (j = 0; j < n; ++j) { 1703 vm_page_t mm = bp->b_xio.xio_pages[j]; 1704 char *p = (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mm)); 1705 crc = iscsi_crc32_ext(p, PAGE_SIZE, crc); 1706 } 1707 if (bio->bio_crc != crc) { 1708 kprintf("PREWRITE MISMATCH-A " 1709 "bdata=%08x dmap=%08x bdata=%08x (%d)\n", 1710 bio->bio_crc, 1711 crc, 1712 iscsi_crc32(bp->b_data, bp->b_bcount), 1713 bp->b_bcount); 1714 #ifdef _KERNEL_VIRTUAL 1715 madvise(bp->b_data, bp->b_bcount, MADV_INVAL); 1716 #endif 1717 crc = 0; 1718 for (j = 0; j < n; ++j) { 1719 vm_page_t mm = bp->b_xio.xio_pages[j]; 1720 char *p = (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mm)); 1721 crc = iscsi_crc32_ext(p, PAGE_SIZE, crc); 1722 } 1723 kprintf("PREWRITE MISMATCH-B " 1724 "bdata=%08x dmap=%08x\n", 1725 iscsi_crc32(bp->b_data, bp->b_bcount), 1726 crc); 1727 } 1728 } 1729 #endif 1730 1731 /* 1732 * asynchronous 1733 */ 1734 if ((flags & VM_PAGER_PUT_SYNC) == 0) { 1735 bio->bio_done = swp_pager_async_iodone; 1736 BUF_KERNPROC(bp); 1737 vn_strategy(swapdev_vp, bio); 1738 1739 for (j = 0; j < n; ++j) 1740 rtvals[i+j] = VM_PAGER_PEND; 1741 continue; 1742 } 1743 1744 /* 1745 * Issue synchrnously. 1746 * 1747 * Wait for the sync I/O to complete, then update rtvals. 1748 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1749 * our async completion routine at the end, thus avoiding a 1750 * double-free. 1751 */ 1752 bio->bio_caller_info1.index |= SWBIO_SYNC; 1753 if (flags & VM_PAGER_TRY_TO_CACHE) 1754 bio->bio_caller_info1.index |= SWBIO_TTC; 1755 bio->bio_done = biodone_sync; 1756 bio->bio_flags |= BIO_SYNC; 1757 vn_strategy(swapdev_vp, bio); 1758 biowait(bio, "swwrt"); 1759 1760 for (j = 0; j < n; ++j) 1761 rtvals[i+j] = VM_PAGER_PEND; 1762 1763 /* 1764 * Now that we are through with the bp, we can call the 1765 * normal async completion, which frees everything up. 1766 */ 1767 swp_pager_async_iodone(bio); 1768 } 1769 vm_object_drop(object); 1770 } 1771 1772 /* 1773 * No requirements. 1774 * 1775 * Recalculate the low and high-water marks. 1776 */ 1777 void 1778 swap_pager_newswap(void) 1779 { 1780 /* 1781 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the 1782 * limitation imposed by the blist code. Remember that this 1783 * will be divided by NSWAP_MAX (4), so each swap device is 1784 * limited to around a terrabyte. 1785 */ 1786 if (vm_swap_max) { 1787 nswap_lowat = (int64_t)vm_swap_max * 4 / 100; /* 4% left */ 1788 nswap_hiwat = (int64_t)vm_swap_max * 6 / 100; /* 6% left */ 1789 kprintf("swap low/high-water marks set to %d/%d\n", 1790 nswap_lowat, nswap_hiwat); 1791 } else { 1792 nswap_lowat = 128; 1793 nswap_hiwat = 512; 1794 } 1795 swp_sizecheck(); 1796 } 1797 1798 /* 1799 * swp_pager_async_iodone: 1800 * 1801 * Completion routine for asynchronous reads and writes from/to swap. 1802 * Also called manually by synchronous code to finish up a bp. 1803 * 1804 * For READ operations, the pages are PG_BUSY'd. For WRITE operations, 1805 * the pages are vm_page_t->busy'd. For READ operations, we PG_BUSY 1806 * unbusy all pages except the 'main' request page. For WRITE 1807 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1808 * because we marked them all VM_PAGER_PEND on return from putpages ). 1809 * 1810 * This routine may not block. 1811 * 1812 * No requirements. 1813 */ 1814 static void 1815 swp_pager_async_iodone(struct bio *bio) 1816 { 1817 struct buf *bp = bio->bio_buf; 1818 vm_object_t object = NULL; 1819 int i; 1820 int *nswptr; 1821 1822 /* 1823 * report error 1824 */ 1825 if (bp->b_flags & B_ERROR) { 1826 kprintf( 1827 "swap_pager: I/O error - %s failed; offset %lld," 1828 "size %ld, error %d\n", 1829 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1830 "pagein" : "pageout"), 1831 (long long)bio->bio_offset, 1832 (long)bp->b_bcount, 1833 bp->b_error 1834 ); 1835 } 1836 1837 /* 1838 * set object. 1839 */ 1840 if (bp->b_xio.xio_npages) 1841 object = bp->b_xio.xio_pages[0]->object; 1842 1843 #if 0 1844 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */ 1845 if (bio->bio_caller_info1.index & SWBIO_WRITE) { 1846 if (bio->bio_crc != iscsi_crc32(bp->b_data, bp->b_bcount)) { 1847 kprintf("SWAPOUT: BADCRC %08x %08x\n", 1848 bio->bio_crc, 1849 iscsi_crc32(bp->b_data, bp->b_bcount)); 1850 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1851 vm_page_t m = bp->b_xio.xio_pages[i]; 1852 if (m->flags & PG_WRITEABLE) 1853 kprintf("SWAPOUT: " 1854 "%d/%d %p writable\n", 1855 i, bp->b_xio.xio_npages, m); 1856 } 1857 } 1858 } 1859 #endif 1860 1861 /* 1862 * remove the mapping for kernel virtual 1863 */ 1864 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1865 1866 /* 1867 * cleanup pages. If an error occurs writing to swap, we are in 1868 * very serious trouble. If it happens to be a disk error, though, 1869 * we may be able to recover by reassigning the swap later on. So 1870 * in this case we remove the m->swapblk assignment for the page 1871 * but do not free it in the rlist. The errornous block(s) are thus 1872 * never reallocated as swap. Redirty the page and continue. 1873 */ 1874 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1875 vm_page_t m = bp->b_xio.xio_pages[i]; 1876 1877 if (bp->b_flags & B_ERROR) { 1878 /* 1879 * If an error occurs I'd love to throw the swapblk 1880 * away without freeing it back to swapspace, so it 1881 * can never be used again. But I can't from an 1882 * interrupt. 1883 */ 1884 1885 if (bio->bio_caller_info1.index & SWBIO_READ) { 1886 /* 1887 * When reading, reqpage needs to stay 1888 * locked for the parent, but all other 1889 * pages can be freed. We still want to 1890 * wakeup the parent waiting on the page, 1891 * though. ( also: pg_reqpage can be -1 and 1892 * not match anything ). 1893 * 1894 * We have to wake specifically requested pages 1895 * up too because we cleared PG_SWAPINPROG and 1896 * someone may be waiting for that. 1897 * 1898 * NOTE: For reads, m->dirty will probably 1899 * be overridden by the original caller 1900 * of getpages so don't play cute tricks 1901 * here. 1902 * 1903 * NOTE: We can't actually free the page from 1904 * here, because this is an interrupt. 1905 * It is not legal to mess with 1906 * object->memq from an interrupt. 1907 * Deactivate the page instead. 1908 * 1909 * WARNING! The instant PG_SWAPINPROG is 1910 * cleared another cpu may start 1911 * using the mreq page (it will 1912 * check m->valid immediately). 1913 */ 1914 1915 m->valid = 0; 1916 vm_page_flag_clear(m, PG_SWAPINPROG); 1917 1918 /* 1919 * bio_driver_info holds the requested page 1920 * index. 1921 */ 1922 if (i != (int)(intptr_t)bio->bio_driver_info) { 1923 vm_page_deactivate(m); 1924 vm_page_wakeup(m); 1925 } else { 1926 vm_page_flash(m); 1927 } 1928 /* 1929 * If i == bp->b_pager.pg_reqpage, do not wake 1930 * the page up. The caller needs to. 1931 */ 1932 } else { 1933 /* 1934 * If a write error occurs remove the swap 1935 * assignment (note that PG_SWAPPED may or 1936 * may not be set depending on prior activity). 1937 * 1938 * Re-dirty OBJT_SWAP pages as there is no 1939 * other backing store, we can't throw the 1940 * page away. 1941 * 1942 * Non-OBJT_SWAP pages (aka swapcache) must 1943 * not be dirtied since they may not have 1944 * been dirty in the first place, and they 1945 * do have backing store (the vnode). 1946 */ 1947 vm_page_busy_wait(m, FALSE, "swadpg"); 1948 vm_object_hold(m->object); 1949 swp_pager_meta_ctl(m->object, m->pindex, 1950 SWM_FREE); 1951 vm_page_flag_clear(m, PG_SWAPPED); 1952 vm_object_drop(m->object); 1953 if (m->object->type == OBJT_SWAP) { 1954 vm_page_dirty(m); 1955 vm_page_activate(m); 1956 } 1957 vm_page_io_finish(m); 1958 vm_page_flag_clear(m, PG_SWAPINPROG); 1959 vm_page_wakeup(m); 1960 } 1961 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1962 /* 1963 * NOTE: for reads, m->dirty will probably be 1964 * overridden by the original caller of getpages so 1965 * we cannot set them in order to free the underlying 1966 * swap in a low-swap situation. I don't think we'd 1967 * want to do that anyway, but it was an optimization 1968 * that existed in the old swapper for a time before 1969 * it got ripped out due to precisely this problem. 1970 * 1971 * If not the requested page then deactivate it. 1972 * 1973 * Note that the requested page, reqpage, is left 1974 * busied, but we still have to wake it up. The 1975 * other pages are released (unbusied) by 1976 * vm_page_wakeup(). We do not set reqpage's 1977 * valid bits here, it is up to the caller. 1978 */ 1979 1980 /* 1981 * NOTE: Can't call pmap_clear_modify(m) from an 1982 * interrupt thread, the pmap code may have to 1983 * map non-kernel pmaps and currently asserts 1984 * the case. 1985 * 1986 * WARNING! The instant PG_SWAPINPROG is 1987 * cleared another cpu may start 1988 * using the mreq page (it will 1989 * check m->valid immediately). 1990 */ 1991 /*pmap_clear_modify(m);*/ 1992 m->valid = VM_PAGE_BITS_ALL; 1993 vm_page_undirty(m); 1994 vm_page_flag_set(m, PG_SWAPPED); 1995 vm_page_flag_clear(m, PG_SWAPINPROG); 1996 1997 /* 1998 * We have to wake specifically requested pages 1999 * up too because we cleared PG_SWAPINPROG and 2000 * could be waiting for it in getpages. However, 2001 * be sure to not unbusy getpages specifically 2002 * requested page - getpages expects it to be 2003 * left busy. 2004 * 2005 * bio_driver_info holds the requested page 2006 */ 2007 if (i != (int)(intptr_t)bio->bio_driver_info) { 2008 vm_page_deactivate(m); 2009 vm_page_wakeup(m); 2010 } else { 2011 vm_page_flash(m); 2012 } 2013 } else { 2014 /* 2015 * Mark the page clean but do not mess with the 2016 * pmap-layer's modified state. That state should 2017 * also be clear since the caller protected the 2018 * page VM_PROT_READ, but allow the case. 2019 * 2020 * We are in an interrupt, avoid pmap operations. 2021 * 2022 * If we have a severe page deficit, deactivate the 2023 * page. Do not try to cache it (which would also 2024 * involve a pmap op), because the page might still 2025 * be read-heavy. 2026 * 2027 * When using the swap to cache clean vnode pages 2028 * we do not mess with the page dirty bits. 2029 * 2030 * NOTE! Nobody is waiting for the key mreq page 2031 * on write completion. 2032 */ 2033 vm_page_busy_wait(m, FALSE, "swadpg"); 2034 if (m->object->type == OBJT_SWAP) 2035 vm_page_undirty(m); 2036 vm_page_flag_set(m, PG_SWAPPED); 2037 vm_page_flag_clear(m, PG_SWAPINPROG); 2038 if (vm_page_count_severe()) 2039 vm_page_deactivate(m); 2040 vm_page_io_finish(m); 2041 if (bio->bio_caller_info1.index & SWBIO_TTC) 2042 vm_page_try_to_cache(m); 2043 else 2044 vm_page_wakeup(m); 2045 } 2046 } 2047 2048 /* 2049 * adjust pip. NOTE: the original parent may still have its own 2050 * pip refs on the object. 2051 */ 2052 2053 if (object) 2054 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages); 2055 2056 /* 2057 * Release the physical I/O buffer. 2058 * 2059 * NOTE: Due to synchronous operations in the write case b_cmd may 2060 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 2061 * been cleared. 2062 * 2063 * Use vm_token to interlock nsw_rcount/wcount wakeup? 2064 */ 2065 lwkt_gettoken(&vm_token); 2066 if (bio->bio_caller_info1.index & SWBIO_READ) 2067 nswptr = &nsw_rcount; 2068 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 2069 nswptr = &nsw_wcount_sync; 2070 else 2071 nswptr = &nsw_wcount_async; 2072 bp->b_cmd = BUF_CMD_DONE; 2073 relpbuf(bp, nswptr); 2074 lwkt_reltoken(&vm_token); 2075 } 2076 2077 /* 2078 * Fault-in a potentially swapped page and remove the swap reference. 2079 * (used by swapoff code) 2080 * 2081 * object must be held. 2082 */ 2083 static __inline void 2084 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex) 2085 { 2086 struct vnode *vp; 2087 vm_page_t m; 2088 int error; 2089 2090 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2091 2092 if (object->type == OBJT_VNODE) { 2093 /* 2094 * Any swap related to a vnode is due to swapcache. We must 2095 * vget() the vnode in case it is not active (otherwise 2096 * vref() will panic). Calling vm_object_page_remove() will 2097 * ensure that any swap ref is removed interlocked with the 2098 * page. clean_only is set to TRUE so we don't throw away 2099 * dirty pages. 2100 */ 2101 vp = object->handle; 2102 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 2103 if (error == 0) { 2104 vm_object_page_remove(object, pindex, pindex + 1, TRUE); 2105 vput(vp); 2106 } 2107 } else { 2108 /* 2109 * Otherwise it is a normal OBJT_SWAP object and we can 2110 * fault the page in and remove the swap. 2111 */ 2112 m = vm_fault_object_page(object, IDX_TO_OFF(pindex), 2113 VM_PROT_NONE, 2114 VM_FAULT_DIRTY | VM_FAULT_UNSWAP, 2115 sharedp, &error); 2116 if (m) 2117 vm_page_unhold(m); 2118 } 2119 } 2120 2121 /* 2122 * This removes all swap blocks related to a particular device. We have 2123 * to be careful of ripups during the scan. 2124 */ 2125 static int swp_pager_swapoff_callback(struct swblock *swap, void *data); 2126 2127 int 2128 swap_pager_swapoff(int devidx) 2129 { 2130 struct vm_object_hash *hash; 2131 struct swswapoffinfo info; 2132 struct vm_object marker; 2133 vm_object_t object; 2134 int n; 2135 2136 bzero(&marker, sizeof(marker)); 2137 marker.type = OBJT_MARKER; 2138 2139 for (n = 0; n < VMOBJ_HSIZE; ++n) { 2140 hash = &vm_object_hash[n]; 2141 2142 lwkt_gettoken(&hash->token); 2143 TAILQ_INSERT_HEAD(&hash->list, &marker, object_list); 2144 2145 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) { 2146 if (object->type == OBJT_MARKER) 2147 goto skip; 2148 if (object->type != OBJT_SWAP && 2149 object->type != OBJT_VNODE) 2150 goto skip; 2151 vm_object_hold(object); 2152 if (object->type != OBJT_SWAP && 2153 object->type != OBJT_VNODE) { 2154 vm_object_drop(object); 2155 goto skip; 2156 } 2157 info.object = object; 2158 info.shared = 0; 2159 info.devidx = devidx; 2160 swblock_rb_tree_RB_SCAN(&object->swblock_root, 2161 NULL, swp_pager_swapoff_callback, 2162 &info); 2163 vm_object_drop(object); 2164 skip: 2165 if (object == TAILQ_NEXT(&marker, object_list)) { 2166 TAILQ_REMOVE(&hash->list, &marker, object_list); 2167 TAILQ_INSERT_AFTER(&hash->list, object, 2168 &marker, object_list); 2169 } 2170 } 2171 TAILQ_REMOVE(&hash->list, &marker, object_list); 2172 lwkt_reltoken(&hash->token); 2173 } 2174 2175 /* 2176 * If we fail to locate all swblocks we just fail gracefully and 2177 * do not bother to restore paging on the swap device. If the 2178 * user wants to retry the user can retry. 2179 */ 2180 if (swdevt[devidx].sw_nused) 2181 return (1); 2182 else 2183 return (0); 2184 } 2185 2186 static 2187 int 2188 swp_pager_swapoff_callback(struct swblock *swap, void *data) 2189 { 2190 struct swswapoffinfo *info = data; 2191 vm_object_t object = info->object; 2192 vm_pindex_t index; 2193 swblk_t v; 2194 int i; 2195 2196 index = swap->swb_index; 2197 for (i = 0; i < SWAP_META_PAGES; ++i) { 2198 /* 2199 * Make sure we don't race a dying object. This will 2200 * kill the scan of the object's swap blocks entirely. 2201 */ 2202 if (object->flags & OBJ_DEAD) 2203 return(-1); 2204 2205 /* 2206 * Fault the page, which can obviously block. If the swap 2207 * structure disappears break out. 2208 */ 2209 v = swap->swb_pages[i]; 2210 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) { 2211 swp_pager_fault_page(object, &info->shared, 2212 swap->swb_index + i); 2213 /* swap ptr might go away */ 2214 if (RB_LOOKUP(swblock_rb_tree, 2215 &object->swblock_root, index) != swap) { 2216 break; 2217 } 2218 } 2219 } 2220 return(0); 2221 } 2222 2223 /************************************************************************ 2224 * SWAP META DATA * 2225 ************************************************************************ 2226 * 2227 * These routines manipulate the swap metadata stored in the 2228 * OBJT_SWAP object. 2229 * 2230 * Swap metadata is implemented with a global hash and not directly 2231 * linked into the object. Instead the object simply contains 2232 * appropriate tracking counters. 2233 */ 2234 2235 /* 2236 * Lookup the swblock containing the specified swap block index. 2237 * 2238 * The caller must hold the object. 2239 */ 2240 static __inline 2241 struct swblock * 2242 swp_pager_lookup(vm_object_t object, vm_pindex_t index) 2243 { 2244 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2245 index &= ~(vm_pindex_t)SWAP_META_MASK; 2246 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index)); 2247 } 2248 2249 /* 2250 * Remove a swblock from the RB tree. 2251 * 2252 * The caller must hold the object. 2253 */ 2254 static __inline 2255 void 2256 swp_pager_remove(vm_object_t object, struct swblock *swap) 2257 { 2258 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2259 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap); 2260 } 2261 2262 /* 2263 * Convert default object to swap object if necessary 2264 * 2265 * The caller must hold the object. 2266 */ 2267 static void 2268 swp_pager_meta_convert(vm_object_t object) 2269 { 2270 if (object->type == OBJT_DEFAULT) { 2271 object->type = OBJT_SWAP; 2272 KKASSERT(object->swblock_count == 0); 2273 } 2274 } 2275 2276 /* 2277 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2278 * 2279 * We first convert the object to a swap object if it is a default 2280 * object. Vnode objects do not need to be converted. 2281 * 2282 * The specified swapblk is added to the object's swap metadata. If 2283 * the swapblk is not valid, it is freed instead. Any previously 2284 * assigned swapblk is freed. 2285 * 2286 * The caller must hold the object. 2287 */ 2288 static void 2289 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk) 2290 { 2291 struct swblock *swap; 2292 struct swblock *oswap; 2293 vm_pindex_t v; 2294 2295 KKASSERT(swapblk != SWAPBLK_NONE); 2296 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2297 2298 /* 2299 * Convert object if necessary 2300 */ 2301 if (object->type == OBJT_DEFAULT) 2302 swp_pager_meta_convert(object); 2303 2304 /* 2305 * Locate swblock. If not found create, but if we aren't adding 2306 * anything just return. If we run out of space in the map we wait 2307 * and, since the hash table may have changed, retry. 2308 */ 2309 retry: 2310 swap = swp_pager_lookup(object, index); 2311 2312 if (swap == NULL) { 2313 int i; 2314 2315 swap = zalloc(swap_zone); 2316 if (swap == NULL) { 2317 vm_wait(0); 2318 goto retry; 2319 } 2320 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK; 2321 swap->swb_count = 0; 2322 2323 ++object->swblock_count; 2324 2325 for (i = 0; i < SWAP_META_PAGES; ++i) 2326 swap->swb_pages[i] = SWAPBLK_NONE; 2327 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap); 2328 KKASSERT(oswap == NULL); 2329 } 2330 2331 /* 2332 * Delete prior contents of metadata. 2333 * 2334 * NOTE: Decrement swb_count after the freeing operation (which 2335 * might block) to prevent racing destruction of the swblock. 2336 */ 2337 index &= SWAP_META_MASK; 2338 2339 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) { 2340 swap->swb_pages[index] = SWAPBLK_NONE; 2341 /* can block */ 2342 swp_pager_freeswapspace(object, v, 1); 2343 --swap->swb_count; 2344 --mycpu->gd_vmtotal.t_vm; 2345 } 2346 2347 /* 2348 * Enter block into metadata 2349 */ 2350 swap->swb_pages[index] = swapblk; 2351 if (swapblk != SWAPBLK_NONE) { 2352 ++swap->swb_count; 2353 ++mycpu->gd_vmtotal.t_vm; 2354 } 2355 } 2356 2357 /* 2358 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2359 * 2360 * The requested range of blocks is freed, with any associated swap 2361 * returned to the swap bitmap. 2362 * 2363 * This routine will free swap metadata structures as they are cleaned 2364 * out. This routine does *NOT* operate on swap metadata associated 2365 * with resident pages. 2366 * 2367 * The caller must hold the object. 2368 */ 2369 static int swp_pager_meta_free_callback(struct swblock *swb, void *data); 2370 2371 static void 2372 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 2373 { 2374 struct swfreeinfo info; 2375 2376 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2377 2378 /* 2379 * Nothing to do 2380 */ 2381 if (object->swblock_count == 0) { 2382 KKASSERT(RB_EMPTY(&object->swblock_root)); 2383 return; 2384 } 2385 if (count == 0) 2386 return; 2387 2388 /* 2389 * Setup for RB tree scan. Note that the pindex range can be huge 2390 * due to the 64 bit page index space so we cannot safely iterate. 2391 */ 2392 info.object = object; 2393 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK; 2394 info.begi = index; 2395 info.endi = index + count - 1; 2396 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp, 2397 swp_pager_meta_free_callback, &info); 2398 } 2399 2400 /* 2401 * The caller must hold the object. 2402 */ 2403 static 2404 int 2405 swp_pager_meta_free_callback(struct swblock *swap, void *data) 2406 { 2407 struct swfreeinfo *info = data; 2408 vm_object_t object = info->object; 2409 int index; 2410 int eindex; 2411 2412 /* 2413 * Figure out the range within the swblock. The wider scan may 2414 * return edge-case swap blocks when the start and/or end points 2415 * are in the middle of a block. 2416 */ 2417 if (swap->swb_index < info->begi) 2418 index = (int)info->begi & SWAP_META_MASK; 2419 else 2420 index = 0; 2421 2422 if (swap->swb_index + SWAP_META_PAGES > info->endi) 2423 eindex = (int)info->endi & SWAP_META_MASK; 2424 else 2425 eindex = SWAP_META_MASK; 2426 2427 /* 2428 * Scan and free the blocks. The loop terminates early 2429 * if (swap) runs out of blocks and could be freed. 2430 * 2431 * NOTE: Decrement swb_count after swp_pager_freeswapspace() 2432 * to deal with a zfree race. 2433 */ 2434 while (index <= eindex) { 2435 swblk_t v = swap->swb_pages[index]; 2436 2437 if (v != SWAPBLK_NONE) { 2438 swap->swb_pages[index] = SWAPBLK_NONE; 2439 /* can block */ 2440 swp_pager_freeswapspace(object, v, 1); 2441 --mycpu->gd_vmtotal.t_vm; 2442 if (--swap->swb_count == 0) { 2443 swp_pager_remove(object, swap); 2444 zfree(swap_zone, swap); 2445 --object->swblock_count; 2446 break; 2447 } 2448 } 2449 ++index; 2450 } 2451 2452 /* swap may be invalid here due to zfree above */ 2453 lwkt_yield(); 2454 2455 return(0); 2456 } 2457 2458 /* 2459 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2460 * 2461 * This routine locates and destroys all swap metadata associated with 2462 * an object. 2463 * 2464 * NOTE: Decrement swb_count after the freeing operation (which 2465 * might block) to prevent racing destruction of the swblock. 2466 * 2467 * The caller must hold the object. 2468 */ 2469 static void 2470 swp_pager_meta_free_all(vm_object_t object) 2471 { 2472 struct swblock *swap; 2473 int i; 2474 2475 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2476 2477 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) { 2478 swp_pager_remove(object, swap); 2479 for (i = 0; i < SWAP_META_PAGES; ++i) { 2480 swblk_t v = swap->swb_pages[i]; 2481 if (v != SWAPBLK_NONE) { 2482 /* can block */ 2483 swp_pager_freeswapspace(object, v, 1); 2484 --swap->swb_count; 2485 --mycpu->gd_vmtotal.t_vm; 2486 } 2487 } 2488 if (swap->swb_count != 0) 2489 panic("swap_pager_meta_free_all: swb_count != 0"); 2490 zfree(swap_zone, swap); 2491 --object->swblock_count; 2492 lwkt_yield(); 2493 } 2494 KKASSERT(object->swblock_count == 0); 2495 } 2496 2497 /* 2498 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2499 * 2500 * This routine is capable of looking up, popping, or freeing 2501 * swapblk assignments in the swap meta data or in the vm_page_t. 2502 * The routine typically returns the swapblk being looked-up, or popped, 2503 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2504 * was invalid. This routine will automatically free any invalid 2505 * meta-data swapblks. 2506 * 2507 * It is not possible to store invalid swapblks in the swap meta data 2508 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2509 * 2510 * When acting on a busy resident page and paging is in progress, we 2511 * have to wait until paging is complete but otherwise can act on the 2512 * busy page. 2513 * 2514 * SWM_FREE remove and free swap block from metadata 2515 * SWM_POP remove from meta data but do not free.. pop it out 2516 * 2517 * The caller must hold the object. 2518 */ 2519 static swblk_t 2520 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags) 2521 { 2522 struct swblock *swap; 2523 swblk_t r1; 2524 2525 if (object->swblock_count == 0) 2526 return(SWAPBLK_NONE); 2527 2528 r1 = SWAPBLK_NONE; 2529 swap = swp_pager_lookup(object, index); 2530 2531 if (swap != NULL) { 2532 index &= SWAP_META_MASK; 2533 r1 = swap->swb_pages[index]; 2534 2535 if (r1 != SWAPBLK_NONE) { 2536 if (flags & (SWM_FREE|SWM_POP)) { 2537 swap->swb_pages[index] = SWAPBLK_NONE; 2538 --mycpu->gd_vmtotal.t_vm; 2539 if (--swap->swb_count == 0) { 2540 swp_pager_remove(object, swap); 2541 zfree(swap_zone, swap); 2542 --object->swblock_count; 2543 } 2544 } 2545 /* swap ptr may be invalid */ 2546 if (flags & SWM_FREE) { 2547 swp_pager_freeswapspace(object, r1, 1); 2548 r1 = SWAPBLK_NONE; 2549 } 2550 } 2551 /* swap ptr may be invalid */ 2552 } 2553 return(r1); 2554 } 2555