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 * NOTE: This function supports the KVABIO API wherein bp->b_data might 950 * not be synchronized to the current cpu. 951 * 952 * No requirements. 953 */ 954 void 955 swap_pager_strategy(vm_object_t object, struct bio *bio) 956 { 957 struct buf *bp = bio->bio_buf; 958 struct bio *nbio; 959 vm_pindex_t start; 960 vm_pindex_t biox_blkno = 0; 961 int count; 962 char *data; 963 struct bio *biox; 964 struct buf *bufx; 965 #if 0 966 struct bio_track *track; 967 #endif 968 969 #if 0 970 /* 971 * tracking for swapdev vnode I/Os 972 */ 973 if (bp->b_cmd == BUF_CMD_READ) 974 track = &swapdev_vp->v_track_read; 975 else 976 track = &swapdev_vp->v_track_write; 977 #endif 978 979 if (bp->b_bcount & PAGE_MASK) { 980 bp->b_error = EINVAL; 981 bp->b_flags |= B_ERROR | B_INVAL; 982 biodone(bio); 983 kprintf("swap_pager_strategy: bp %p offset %lld size %d, " 984 "not page bounded\n", 985 bp, (long long)bio->bio_offset, (int)bp->b_bcount); 986 return; 987 } 988 989 /* 990 * Clear error indication, initialize page index, count, data pointer. 991 */ 992 bp->b_error = 0; 993 bp->b_flags &= ~B_ERROR; 994 bp->b_resid = bp->b_bcount; 995 996 start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT); 997 count = howmany(bp->b_bcount, PAGE_SIZE); 998 999 /* 1000 * WARNING! Do not dereference *data without issuing a bkvasync() 1001 */ 1002 data = bp->b_data; 1003 1004 /* 1005 * Deal with BUF_CMD_FREEBLKS 1006 */ 1007 if (bp->b_cmd == BUF_CMD_FREEBLKS) { 1008 /* 1009 * FREE PAGE(s) - destroy underlying swap that is no longer 1010 * needed. 1011 */ 1012 vm_object_hold(object); 1013 swp_pager_meta_free(object, start, count); 1014 vm_object_drop(object); 1015 bp->b_resid = 0; 1016 biodone(bio); 1017 return; 1018 } 1019 1020 /* 1021 * We need to be able to create a new cluster of I/O's. We cannot 1022 * use the caller fields of the passed bio so push a new one. 1023 * 1024 * Because nbio is just a placeholder for the cluster links, 1025 * we can biodone() the original bio instead of nbio to make 1026 * things a bit more efficient. 1027 */ 1028 nbio = push_bio(bio); 1029 nbio->bio_offset = bio->bio_offset; 1030 nbio->bio_caller_info1.cluster_head = NULL; 1031 nbio->bio_caller_info2.cluster_tail = NULL; 1032 1033 biox = NULL; 1034 bufx = NULL; 1035 1036 /* 1037 * Execute read or write 1038 */ 1039 vm_object_hold(object); 1040 1041 while (count > 0) { 1042 swblk_t blk; 1043 1044 /* 1045 * Obtain block. If block not found and writing, allocate a 1046 * new block and build it into the object. 1047 */ 1048 blk = swp_pager_meta_ctl(object, start, 0); 1049 if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) { 1050 blk = swp_pager_getswapspace(object, 1); 1051 if (blk == SWAPBLK_NONE) { 1052 bp->b_error = ENOMEM; 1053 bp->b_flags |= B_ERROR; 1054 break; 1055 } 1056 swp_pager_meta_build(object, start, blk); 1057 } 1058 1059 /* 1060 * Do we have to flush our current collection? Yes if: 1061 * 1062 * - no swap block at this index 1063 * - swap block is not contiguous 1064 * - we cross a physical disk boundry in the 1065 * stripe. 1066 */ 1067 if ( 1068 biox && (biox_blkno + btoc(bufx->b_bcount) != blk || 1069 ((biox_blkno ^ blk) & ~SWB_DMMASK) 1070 ) 1071 ) { 1072 if (bp->b_cmd == BUF_CMD_READ) { 1073 ++mycpu->gd_cnt.v_swapin; 1074 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1075 } else { 1076 ++mycpu->gd_cnt.v_swapout; 1077 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1078 bufx->b_dirtyend = bufx->b_bcount; 1079 } 1080 1081 /* 1082 * Finished with this buf. 1083 */ 1084 KKASSERT(bufx->b_bcount != 0); 1085 if (bufx->b_cmd != BUF_CMD_READ) 1086 bufx->b_dirtyend = bufx->b_bcount; 1087 biox = NULL; 1088 bufx = NULL; 1089 } 1090 1091 /* 1092 * Add new swapblk to biox, instantiating biox if necessary. 1093 * Zero-fill reads are able to take a shortcut. 1094 */ 1095 if (blk == SWAPBLK_NONE) { 1096 /* 1097 * We can only get here if we are reading. 1098 */ 1099 bkvasync(bp); 1100 bzero(data, PAGE_SIZE); 1101 bp->b_resid -= PAGE_SIZE; 1102 } else { 1103 if (biox == NULL) { 1104 /* XXX chain count > 4, wait to <= 4 */ 1105 1106 bufx = getpbuf(NULL); 1107 bufx->b_flags |= B_KVABIO; 1108 biox = &bufx->b_bio1; 1109 cluster_append(nbio, bufx); 1110 bufx->b_cmd = bp->b_cmd; 1111 biox->bio_done = swap_chain_iodone; 1112 biox->bio_offset = (off_t)blk << PAGE_SHIFT; 1113 biox->bio_caller_info1.cluster_parent = nbio; 1114 biox_blkno = blk; 1115 bufx->b_bcount = 0; 1116 bufx->b_data = data; 1117 } 1118 bufx->b_bcount += PAGE_SIZE; 1119 } 1120 --count; 1121 ++start; 1122 data += PAGE_SIZE; 1123 } 1124 1125 vm_object_drop(object); 1126 1127 /* 1128 * Flush out last buffer 1129 */ 1130 if (biox) { 1131 if (bufx->b_cmd == BUF_CMD_READ) { 1132 ++mycpu->gd_cnt.v_swapin; 1133 mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount); 1134 } else { 1135 ++mycpu->gd_cnt.v_swapout; 1136 mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount); 1137 bufx->b_dirtyend = bufx->b_bcount; 1138 } 1139 KKASSERT(bufx->b_bcount); 1140 if (bufx->b_cmd != BUF_CMD_READ) 1141 bufx->b_dirtyend = bufx->b_bcount; 1142 /* biox, bufx = NULL */ 1143 } 1144 1145 /* 1146 * Now initiate all the I/O. Be careful looping on our chain as 1147 * I/O's may complete while we are still initiating them. 1148 * 1149 * If the request is a 100% sparse read no bios will be present 1150 * and we just biodone() the buffer. 1151 */ 1152 nbio->bio_caller_info2.cluster_tail = NULL; 1153 bufx = nbio->bio_caller_info1.cluster_head; 1154 1155 if (bufx) { 1156 while (bufx) { 1157 biox = &bufx->b_bio1; 1158 BUF_KERNPROC(bufx); 1159 bufx = bufx->b_cluster_next; 1160 vn_strategy(swapdev_vp, biox); 1161 } 1162 } else { 1163 biodone(bio); 1164 } 1165 1166 /* 1167 * Completion of the cluster will also call biodone_chain(nbio). 1168 * We never call biodone(nbio) so we don't have to worry about 1169 * setting up a bio_done callback. It's handled in the sub-IO. 1170 */ 1171 /**/ 1172 } 1173 1174 /* 1175 * biodone callback 1176 * 1177 * No requirements. 1178 */ 1179 static void 1180 swap_chain_iodone(struct bio *biox) 1181 { 1182 struct buf **nextp; 1183 struct buf *bufx; /* chained sub-buffer */ 1184 struct bio *nbio; /* parent nbio with chain glue */ 1185 struct buf *bp; /* original bp associated with nbio */ 1186 int chain_empty; 1187 1188 bufx = biox->bio_buf; 1189 nbio = biox->bio_caller_info1.cluster_parent; 1190 bp = nbio->bio_buf; 1191 1192 /* 1193 * Update the original buffer 1194 */ 1195 KKASSERT(bp != NULL); 1196 if (bufx->b_flags & B_ERROR) { 1197 atomic_set_int(&bufx->b_flags, B_ERROR); 1198 bp->b_error = bufx->b_error; /* race ok */ 1199 } else if (bufx->b_resid != 0) { 1200 atomic_set_int(&bufx->b_flags, B_ERROR); 1201 bp->b_error = EINVAL; /* race ok */ 1202 } else { 1203 atomic_subtract_int(&bp->b_resid, bufx->b_bcount); 1204 } 1205 1206 /* 1207 * Remove us from the chain. 1208 */ 1209 spin_lock(&swapbp_spin); 1210 nextp = &nbio->bio_caller_info1.cluster_head; 1211 while (*nextp != bufx) { 1212 KKASSERT(*nextp != NULL); 1213 nextp = &(*nextp)->b_cluster_next; 1214 } 1215 *nextp = bufx->b_cluster_next; 1216 chain_empty = (nbio->bio_caller_info1.cluster_head == NULL); 1217 spin_unlock(&swapbp_spin); 1218 1219 /* 1220 * Clean up bufx. If the chain is now empty we finish out 1221 * the parent. Note that we may be racing other completions 1222 * so we must use the chain_empty status from above. 1223 */ 1224 if (chain_empty) { 1225 if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) { 1226 atomic_set_int(&bp->b_flags, B_ERROR); 1227 bp->b_error = EINVAL; 1228 } 1229 biodone_chain(nbio); 1230 } 1231 relpbuf(bufx, NULL); 1232 } 1233 1234 /* 1235 * SWAP_PAGER_GETPAGES() - bring page in from swap 1236 * 1237 * The requested page may have to be brought in from swap. Calculate the 1238 * swap block and bring in additional pages if possible. All pages must 1239 * have contiguous swap block assignments and reside in the same object. 1240 * 1241 * The caller has a single vm_object_pip_add() reference prior to 1242 * calling us and we should return with the same. 1243 * 1244 * The caller has BUSY'd the page. We should return with (*mpp) left busy, 1245 * and any additinal pages unbusied. 1246 * 1247 * If the caller encounters a PG_RAM page it will pass it to us even though 1248 * it may be valid and dirty. We cannot overwrite the page in this case! 1249 * The case is used to allow us to issue pure read-aheads. 1250 * 1251 * NOTE! XXX This code does not entirely pipeline yet due to the fact that 1252 * the PG_RAM page is validated at the same time as mreq. What we 1253 * really need to do is issue a separate read-ahead pbuf. 1254 * 1255 * No requirements. 1256 */ 1257 static int 1258 swap_pager_getpage(vm_object_t object, vm_page_t *mpp, int seqaccess) 1259 { 1260 struct buf *bp; 1261 struct bio *bio; 1262 vm_page_t mreq; 1263 vm_page_t m; 1264 vm_offset_t kva; 1265 swblk_t blk; 1266 int i; 1267 int j; 1268 int raonly; 1269 int error; 1270 u_int32_t busy_count; 1271 vm_page_t marray[XIO_INTERNAL_PAGES]; 1272 1273 mreq = *mpp; 1274 1275 vm_object_hold(object); 1276 if (mreq->object != object) { 1277 panic("swap_pager_getpages: object mismatch %p/%p", 1278 object, 1279 mreq->object 1280 ); 1281 } 1282 1283 /* 1284 * We don't want to overwrite a fully valid page as it might be 1285 * dirty. This case can occur when e.g. vm_fault hits a perfectly 1286 * valid page with PG_RAM set. 1287 * 1288 * In this case we see if the next page is a suitable page-in 1289 * candidate and if it is we issue read-ahead. PG_RAM will be 1290 * set on the last page of the read-ahead to continue the pipeline. 1291 */ 1292 if (mreq->valid == VM_PAGE_BITS_ALL) { 1293 if (swap_burst_read == 0 || mreq->pindex + 1 >= object->size) { 1294 vm_object_drop(object); 1295 return(VM_PAGER_OK); 1296 } 1297 blk = swp_pager_meta_ctl(object, mreq->pindex + 1, 0); 1298 if (blk == SWAPBLK_NONE) { 1299 vm_object_drop(object); 1300 return(VM_PAGER_OK); 1301 } 1302 m = vm_page_lookup_busy_try(object, mreq->pindex + 1, 1303 TRUE, &error); 1304 if (error) { 1305 vm_object_drop(object); 1306 return(VM_PAGER_OK); 1307 } else if (m == NULL) { 1308 /* 1309 * Use VM_ALLOC_QUICK to avoid blocking on cache 1310 * page reuse. 1311 */ 1312 m = vm_page_alloc(object, mreq->pindex + 1, 1313 VM_ALLOC_QUICK); 1314 if (m == NULL) { 1315 vm_object_drop(object); 1316 return(VM_PAGER_OK); 1317 } 1318 } else { 1319 if (m->valid) { 1320 vm_page_wakeup(m); 1321 vm_object_drop(object); 1322 return(VM_PAGER_OK); 1323 } 1324 vm_page_unqueue_nowakeup(m); 1325 } 1326 /* page is busy */ 1327 mreq = m; 1328 raonly = 1; 1329 } else { 1330 raonly = 0; 1331 } 1332 1333 /* 1334 * Try to block-read contiguous pages from swap if sequential, 1335 * otherwise just read one page. Contiguous pages from swap must 1336 * reside within a single device stripe because the I/O cannot be 1337 * broken up across multiple stripes. 1338 * 1339 * Note that blk and iblk can be SWAPBLK_NONE but the loop is 1340 * set up such that the case(s) are handled implicitly. 1341 */ 1342 blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0); 1343 marray[0] = mreq; 1344 1345 for (i = 1; i <= swap_burst_read && 1346 i < XIO_INTERNAL_PAGES && 1347 mreq->pindex + i < object->size; ++i) { 1348 swblk_t iblk; 1349 1350 iblk = swp_pager_meta_ctl(object, mreq->pindex + i, 0); 1351 if (iblk != blk + i) 1352 break; 1353 if ((blk ^ iblk) & ~SWB_DMMASK) 1354 break; 1355 m = vm_page_lookup_busy_try(object, mreq->pindex + i, 1356 TRUE, &error); 1357 if (error) { 1358 break; 1359 } else if (m == NULL) { 1360 /* 1361 * Use VM_ALLOC_QUICK to avoid blocking on cache 1362 * page reuse. 1363 */ 1364 m = vm_page_alloc(object, mreq->pindex + i, 1365 VM_ALLOC_QUICK); 1366 if (m == NULL) 1367 break; 1368 } else { 1369 if (m->valid) { 1370 vm_page_wakeup(m); 1371 break; 1372 } 1373 vm_page_unqueue_nowakeup(m); 1374 } 1375 /* page is busy */ 1376 marray[i] = m; 1377 } 1378 if (i > 1) 1379 vm_page_flag_set(marray[i - 1], PG_RAM); 1380 1381 /* 1382 * If mreq is the requested page and we have nothing to do return 1383 * VM_PAGER_FAIL. If raonly is set mreq is just another read-ahead 1384 * page and must be cleaned up. 1385 */ 1386 if (blk == SWAPBLK_NONE) { 1387 KKASSERT(i == 1); 1388 if (raonly) { 1389 vnode_pager_freepage(mreq); 1390 vm_object_drop(object); 1391 return(VM_PAGER_OK); 1392 } else { 1393 vm_object_drop(object); 1394 return(VM_PAGER_FAIL); 1395 } 1396 } 1397 1398 /* 1399 * Map our page(s) into kva for input 1400 * 1401 * Use the KVABIO API to avoid synchronizing the pmap. 1402 */ 1403 bp = getpbuf_kva(&nsw_rcount); 1404 bio = &bp->b_bio1; 1405 kva = (vm_offset_t) bp->b_kvabase; 1406 bcopy(marray, bp->b_xio.xio_pages, i * sizeof(vm_page_t)); 1407 pmap_qenter_noinval(kva, bp->b_xio.xio_pages, i); 1408 1409 bp->b_data = (caddr_t)kva; 1410 bp->b_bcount = PAGE_SIZE * i; 1411 bp->b_xio.xio_npages = i; 1412 bp->b_flags |= B_KVABIO; 1413 bio->bio_done = swp_pager_async_iodone; 1414 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1415 bio->bio_caller_info1.index = SWBIO_READ; 1416 1417 /* 1418 * Set index. If raonly set the index beyond the array so all 1419 * the pages are treated the same, otherwise the original mreq is 1420 * at index 0. 1421 */ 1422 if (raonly) 1423 bio->bio_driver_info = (void *)(intptr_t)i; 1424 else 1425 bio->bio_driver_info = (void *)(intptr_t)0; 1426 1427 for (j = 0; j < i; ++j) { 1428 atomic_set_int(&bp->b_xio.xio_pages[j]->busy_count, 1429 PBUSY_SWAPINPROG); 1430 } 1431 1432 mycpu->gd_cnt.v_swapin++; 1433 mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages; 1434 1435 /* 1436 * We still hold the lock on mreq, and our automatic completion routine 1437 * does not remove it. 1438 */ 1439 vm_object_pip_add(object, bp->b_xio.xio_npages); 1440 1441 /* 1442 * perform the I/O. NOTE!!! bp cannot be considered valid after 1443 * this point because we automatically release it on completion. 1444 * Instead, we look at the one page we are interested in which we 1445 * still hold a lock on even through the I/O completion. 1446 * 1447 * The other pages in our m[] array are also released on completion, 1448 * so we cannot assume they are valid anymore either. 1449 */ 1450 bp->b_cmd = BUF_CMD_READ; 1451 BUF_KERNPROC(bp); 1452 vn_strategy(swapdev_vp, bio); 1453 1454 /* 1455 * Wait for the page we want to complete. PBUSY_SWAPINPROG is always 1456 * cleared on completion. If an I/O error occurs, SWAPBLK_NONE 1457 * is set in the meta-data. 1458 * 1459 * If this is a read-ahead only we return immediately without 1460 * waiting for I/O. 1461 */ 1462 if (raonly) { 1463 vm_object_drop(object); 1464 return(VM_PAGER_OK); 1465 } 1466 1467 /* 1468 * Read-ahead includes originally requested page case. 1469 */ 1470 for (;;) { 1471 busy_count = mreq->busy_count; 1472 cpu_ccfence(); 1473 if ((busy_count & PBUSY_SWAPINPROG) == 0) 1474 break; 1475 tsleep_interlock(mreq, 0); 1476 if (!atomic_cmpset_int(&mreq->busy_count, busy_count, 1477 busy_count | 1478 PBUSY_SWAPINPROG | PBUSY_WANTED)) { 1479 continue; 1480 } 1481 atomic_set_int(&mreq->flags, PG_REFERENCED); 1482 mycpu->gd_cnt.v_intrans++; 1483 if (tsleep(mreq, PINTERLOCKED, "swread", hz*20)) { 1484 kprintf( 1485 "swap_pager: indefinite wait buffer: " 1486 " bp %p offset: %lld, size: %ld\n", 1487 bp, 1488 (long long)bio->bio_offset, 1489 (long)bp->b_bcount 1490 ); 1491 } 1492 } 1493 1494 /* 1495 * Disallow speculative reads prior to the SWAPINPROG test. 1496 */ 1497 cpu_lfence(); 1498 1499 /* 1500 * mreq is left busied after completion, but all the other pages 1501 * are freed. If we had an unrecoverable read error the page will 1502 * not be valid. 1503 */ 1504 vm_object_drop(object); 1505 if (mreq->valid != VM_PAGE_BITS_ALL) 1506 return(VM_PAGER_ERROR); 1507 else 1508 return(VM_PAGER_OK); 1509 1510 /* 1511 * A final note: in a low swap situation, we cannot deallocate swap 1512 * and mark a page dirty here because the caller is likely to mark 1513 * the page clean when we return, causing the page to possibly revert 1514 * to all-zero's later. 1515 */ 1516 } 1517 1518 /* 1519 * swap_pager_putpages: 1520 * 1521 * Assign swap (if necessary) and initiate I/O on the specified pages. 1522 * 1523 * We support both OBJT_DEFAULT and OBJT_SWAP objects. DEFAULT objects 1524 * are automatically converted to SWAP objects. 1525 * 1526 * In a low memory situation we may block in vn_strategy(), but the new 1527 * vm_page reservation system coupled with properly written VFS devices 1528 * should ensure that no low-memory deadlock occurs. This is an area 1529 * which needs work. 1530 * 1531 * The parent has N vm_object_pip_add() references prior to 1532 * calling us and will remove references for rtvals[] that are 1533 * not set to VM_PAGER_PEND. We need to remove the rest on I/O 1534 * completion. 1535 * 1536 * The parent has soft-busy'd the pages it passes us and will unbusy 1537 * those whos rtvals[] entry is not set to VM_PAGER_PEND on return. 1538 * We need to unbusy the rest on I/O completion. 1539 * 1540 * No requirements. 1541 */ 1542 void 1543 swap_pager_putpages(vm_object_t object, vm_page_t *m, int count, 1544 int flags, int *rtvals) 1545 { 1546 int i; 1547 int n = 0; 1548 1549 vm_object_hold(object); 1550 1551 if (count && m[0]->object != object) { 1552 panic("swap_pager_getpages: object mismatch %p/%p", 1553 object, 1554 m[0]->object 1555 ); 1556 } 1557 1558 /* 1559 * Step 1 1560 * 1561 * Turn object into OBJT_SWAP 1562 * Check for bogus sysops 1563 * 1564 * Force sync if not pageout process, we don't want any single 1565 * non-pageout process to be able to hog the I/O subsystem! This 1566 * can be overridden by setting. 1567 */ 1568 if (object->type == OBJT_DEFAULT) { 1569 if (object->type == OBJT_DEFAULT) 1570 swp_pager_meta_convert(object); 1571 } 1572 1573 /* 1574 * Normally we force synchronous swap I/O if this is not the 1575 * pageout daemon to prevent any single user process limited 1576 * via RLIMIT_RSS from hogging swap write bandwidth. 1577 */ 1578 if (curthread != pagethread && 1579 curthread != emergpager && 1580 swap_user_async == 0) { 1581 flags |= VM_PAGER_PUT_SYNC; 1582 } 1583 1584 /* 1585 * Step 2 1586 * 1587 * Update nsw parameters from swap_async_max sysctl values. 1588 * Do not let the sysop crash the machine with bogus numbers. 1589 */ 1590 if (swap_async_max != nsw_wcount_async_max) { 1591 int n; 1592 1593 /* 1594 * limit range 1595 */ 1596 if ((n = swap_async_max) > nswbuf_kva / 2) 1597 n = nswbuf_kva / 2; 1598 if (n < 1) 1599 n = 1; 1600 swap_async_max = n; 1601 1602 /* 1603 * Adjust difference ( if possible ). If the current async 1604 * count is too low, we may not be able to make the adjustment 1605 * at this time. 1606 * 1607 * vm_token needed for nsw_wcount sleep interlock 1608 */ 1609 lwkt_gettoken(&vm_token); 1610 n -= nsw_wcount_async_max; 1611 if (nsw_wcount_async + n >= 0) { 1612 nsw_wcount_async_max += n; 1613 pbuf_adjcount(&nsw_wcount_async, n); 1614 } 1615 lwkt_reltoken(&vm_token); 1616 } 1617 1618 /* 1619 * Step 3 1620 * 1621 * Assign swap blocks and issue I/O. We reallocate swap on the fly. 1622 * The page is left dirty until the pageout operation completes 1623 * successfully. 1624 */ 1625 1626 for (i = 0; i < count; i += n) { 1627 struct buf *bp; 1628 struct bio *bio; 1629 swblk_t blk; 1630 int j; 1631 1632 /* 1633 * Maximum I/O size is limited by a number of factors. 1634 */ 1635 1636 n = min(BLIST_MAX_ALLOC, count - i); 1637 n = min(n, nsw_cluster_max); 1638 1639 lwkt_gettoken(&vm_token); 1640 1641 /* 1642 * Get biggest block of swap we can. If we fail, fall 1643 * back and try to allocate a smaller block. Don't go 1644 * overboard trying to allocate space if it would overly 1645 * fragment swap. 1646 */ 1647 while ( 1648 (blk = swp_pager_getswapspace(object, n)) == SWAPBLK_NONE && 1649 n > 4 1650 ) { 1651 n >>= 1; 1652 } 1653 if (blk == SWAPBLK_NONE) { 1654 for (j = 0; j < n; ++j) 1655 rtvals[i+j] = VM_PAGER_FAIL; 1656 lwkt_reltoken(&vm_token); 1657 continue; 1658 } 1659 if (vm_report_swap_allocs > 0) { 1660 kprintf("swap_alloc %08jx,%d\n", (intmax_t)blk, n); 1661 --vm_report_swap_allocs; 1662 } 1663 1664 /* 1665 * The I/O we are constructing cannot cross a physical 1666 * disk boundry in the swap stripe. 1667 */ 1668 if ((blk ^ (blk + n)) & ~SWB_DMMASK) { 1669 j = ((blk + SWB_DMMAX) & ~SWB_DMMASK) - blk; 1670 swp_pager_freeswapspace(object, blk + j, n - j); 1671 n = j; 1672 } 1673 1674 /* 1675 * All I/O parameters have been satisfied, build the I/O 1676 * request and assign the swap space. 1677 * 1678 * Use the KVABIO API to avoid synchronizing the pmap. 1679 */ 1680 if ((flags & VM_PAGER_PUT_SYNC)) 1681 bp = getpbuf_kva(&nsw_wcount_sync); 1682 else 1683 bp = getpbuf_kva(&nsw_wcount_async); 1684 bio = &bp->b_bio1; 1685 1686 lwkt_reltoken(&vm_token); 1687 1688 pmap_qenter_noinval((vm_offset_t)bp->b_data, &m[i], n); 1689 1690 bp->b_flags |= B_KVABIO; 1691 bp->b_bcount = PAGE_SIZE * n; 1692 bio->bio_offset = (off_t)blk << PAGE_SHIFT; 1693 1694 for (j = 0; j < n; ++j) { 1695 vm_page_t mreq = m[i+j]; 1696 1697 swp_pager_meta_build(mreq->object, mreq->pindex, 1698 blk + j); 1699 if (object->type == OBJT_SWAP) 1700 vm_page_dirty(mreq); 1701 rtvals[i+j] = VM_PAGER_OK; 1702 1703 atomic_set_int(&mreq->busy_count, PBUSY_SWAPINPROG); 1704 bp->b_xio.xio_pages[j] = mreq; 1705 } 1706 bp->b_xio.xio_npages = n; 1707 1708 mycpu->gd_cnt.v_swapout++; 1709 mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages; 1710 1711 bp->b_dirtyoff = 0; /* req'd for NFS */ 1712 bp->b_dirtyend = bp->b_bcount; /* req'd for NFS */ 1713 bp->b_cmd = BUF_CMD_WRITE; 1714 bio->bio_caller_info1.index = SWBIO_WRITE; 1715 1716 /* 1717 * asynchronous 1718 */ 1719 if ((flags & VM_PAGER_PUT_SYNC) == 0) { 1720 bio->bio_done = swp_pager_async_iodone; 1721 BUF_KERNPROC(bp); 1722 vn_strategy(swapdev_vp, bio); 1723 1724 for (j = 0; j < n; ++j) 1725 rtvals[i+j] = VM_PAGER_PEND; 1726 continue; 1727 } 1728 1729 /* 1730 * Issue synchrnously. 1731 * 1732 * Wait for the sync I/O to complete, then update rtvals. 1733 * We just set the rtvals[] to VM_PAGER_PEND so we can call 1734 * our async completion routine at the end, thus avoiding a 1735 * double-free. 1736 */ 1737 bio->bio_caller_info1.index |= SWBIO_SYNC; 1738 if (flags & VM_PAGER_TRY_TO_CACHE) 1739 bio->bio_caller_info1.index |= SWBIO_TTC; 1740 bio->bio_done = biodone_sync; 1741 bio->bio_flags |= BIO_SYNC; 1742 vn_strategy(swapdev_vp, bio); 1743 biowait(bio, "swwrt"); 1744 1745 for (j = 0; j < n; ++j) 1746 rtvals[i+j] = VM_PAGER_PEND; 1747 1748 /* 1749 * Now that we are through with the bp, we can call the 1750 * normal async completion, which frees everything up. 1751 */ 1752 swp_pager_async_iodone(bio); 1753 } 1754 vm_object_drop(object); 1755 } 1756 1757 /* 1758 * No requirements. 1759 * 1760 * Recalculate the low and high-water marks. 1761 */ 1762 void 1763 swap_pager_newswap(void) 1764 { 1765 /* 1766 * NOTE: vm_swap_max cannot exceed 1 billion blocks, which is the 1767 * limitation imposed by the blist code. Remember that this 1768 * will be divided by NSWAP_MAX (4), so each swap device is 1769 * limited to around a terrabyte. 1770 */ 1771 if (vm_swap_max) { 1772 nswap_lowat = (int64_t)vm_swap_max * 4 / 100; /* 4% left */ 1773 nswap_hiwat = (int64_t)vm_swap_max * 6 / 100; /* 6% left */ 1774 kprintf("swap low/high-water marks set to %d/%d\n", 1775 nswap_lowat, nswap_hiwat); 1776 } else { 1777 nswap_lowat = 128; 1778 nswap_hiwat = 512; 1779 } 1780 swp_sizecheck(); 1781 } 1782 1783 /* 1784 * swp_pager_async_iodone: 1785 * 1786 * Completion routine for asynchronous reads and writes from/to swap. 1787 * Also called manually by synchronous code to finish up a bp. 1788 * 1789 * For READ operations, the pages are BUSY'd. For WRITE operations, 1790 * the pages are vm_page_t->busy'd. For READ operations, we BUSY 1791 * unbusy all pages except the 'main' request page. For WRITE 1792 * operations, we vm_page_t->busy'd unbusy all pages ( we can do this 1793 * because we marked them all VM_PAGER_PEND on return from putpages ). 1794 * 1795 * This routine may not block. 1796 * 1797 * No requirements. 1798 */ 1799 static void 1800 swp_pager_async_iodone(struct bio *bio) 1801 { 1802 struct buf *bp = bio->bio_buf; 1803 vm_object_t object = NULL; 1804 int i; 1805 int *nswptr; 1806 1807 /* 1808 * report error 1809 */ 1810 if (bp->b_flags & B_ERROR) { 1811 kprintf( 1812 "swap_pager: I/O error - %s failed; offset %lld," 1813 "size %ld, error %d\n", 1814 ((bio->bio_caller_info1.index & SWBIO_READ) ? 1815 "pagein" : "pageout"), 1816 (long long)bio->bio_offset, 1817 (long)bp->b_bcount, 1818 bp->b_error 1819 ); 1820 } 1821 1822 /* 1823 * set object. 1824 */ 1825 if (bp->b_xio.xio_npages) 1826 object = bp->b_xio.xio_pages[0]->object; 1827 1828 #if 0 1829 /* PMAP TESTING CODE (useful, keep it in but #if 0'd) */ 1830 if (bio->bio_caller_info1.index & SWBIO_WRITE) { 1831 if (bio->bio_crc != iscsi_crc32(bp->b_data, bp->b_bcount)) { 1832 kprintf("SWAPOUT: BADCRC %08x %08x\n", 1833 bio->bio_crc, 1834 iscsi_crc32(bp->b_data, bp->b_bcount)); 1835 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1836 vm_page_t m = bp->b_xio.xio_pages[i]; 1837 if (m->flags & PG_WRITEABLE) 1838 kprintf("SWAPOUT: " 1839 "%d/%d %p writable\n", 1840 i, bp->b_xio.xio_npages, m); 1841 } 1842 } 1843 } 1844 #endif 1845 1846 /* 1847 * remove the mapping for kernel virtual 1848 */ 1849 pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages); 1850 1851 /* 1852 * cleanup pages. If an error occurs writing to swap, we are in 1853 * very serious trouble. If it happens to be a disk error, though, 1854 * we may be able to recover by reassigning the swap later on. So 1855 * in this case we remove the m->swapblk assignment for the page 1856 * but do not free it in the rlist. The errornous block(s) are thus 1857 * never reallocated as swap. Redirty the page and continue. 1858 */ 1859 for (i = 0; i < bp->b_xio.xio_npages; ++i) { 1860 vm_page_t m = bp->b_xio.xio_pages[i]; 1861 1862 if (bp->b_flags & B_ERROR) { 1863 /* 1864 * If an error occurs I'd love to throw the swapblk 1865 * away without freeing it back to swapspace, so it 1866 * can never be used again. But I can't from an 1867 * interrupt. 1868 */ 1869 1870 if (bio->bio_caller_info1.index & SWBIO_READ) { 1871 /* 1872 * When reading, reqpage needs to stay 1873 * locked for the parent, but all other 1874 * pages can be freed. We still want to 1875 * wakeup the parent waiting on the page, 1876 * though. ( also: pg_reqpage can be -1 and 1877 * not match anything ). 1878 * 1879 * We have to wake specifically requested pages 1880 * up too because we cleared SWAPINPROG and 1881 * someone may be waiting for that. 1882 * 1883 * NOTE: For reads, m->dirty will probably 1884 * be overridden by the original caller 1885 * of getpages so don't play cute tricks 1886 * here. 1887 * 1888 * NOTE: We can't actually free the page from 1889 * here, because this is an interrupt. 1890 * It is not legal to mess with 1891 * object->memq from an interrupt. 1892 * Deactivate the page instead. 1893 * 1894 * WARNING! The instant SWAPINPROG is 1895 * cleared another cpu may start 1896 * using the mreq page (it will 1897 * check m->valid immediately). 1898 */ 1899 1900 m->valid = 0; 1901 atomic_clear_int(&m->busy_count, 1902 PBUSY_SWAPINPROG); 1903 1904 /* 1905 * bio_driver_info holds the requested page 1906 * index. 1907 */ 1908 if (i != (int)(intptr_t)bio->bio_driver_info) { 1909 vm_page_deactivate(m); 1910 vm_page_wakeup(m); 1911 } else { 1912 vm_page_flash(m); 1913 } 1914 /* 1915 * If i == bp->b_pager.pg_reqpage, do not wake 1916 * the page up. The caller needs to. 1917 */ 1918 } else { 1919 /* 1920 * If a write error occurs remove the swap 1921 * assignment (note that PG_SWAPPED may or 1922 * may not be set depending on prior activity). 1923 * 1924 * Re-dirty OBJT_SWAP pages as there is no 1925 * other backing store, we can't throw the 1926 * page away. 1927 * 1928 * Non-OBJT_SWAP pages (aka swapcache) must 1929 * not be dirtied since they may not have 1930 * been dirty in the first place, and they 1931 * do have backing store (the vnode). 1932 */ 1933 vm_page_busy_wait(m, FALSE, "swadpg"); 1934 vm_object_hold(m->object); 1935 swp_pager_meta_ctl(m->object, m->pindex, 1936 SWM_FREE); 1937 vm_page_flag_clear(m, PG_SWAPPED); 1938 vm_object_drop(m->object); 1939 if (m->object->type == OBJT_SWAP) { 1940 vm_page_dirty(m); 1941 vm_page_activate(m); 1942 } 1943 vm_page_io_finish(m); 1944 atomic_clear_int(&m->busy_count, 1945 PBUSY_SWAPINPROG); 1946 vm_page_wakeup(m); 1947 } 1948 } else if (bio->bio_caller_info1.index & SWBIO_READ) { 1949 /* 1950 * NOTE: for reads, m->dirty will probably be 1951 * overridden by the original caller of getpages so 1952 * we cannot set them in order to free the underlying 1953 * swap in a low-swap situation. I don't think we'd 1954 * want to do that anyway, but it was an optimization 1955 * that existed in the old swapper for a time before 1956 * it got ripped out due to precisely this problem. 1957 * 1958 * If not the requested page then deactivate it. 1959 * 1960 * Note that the requested page, reqpage, is left 1961 * busied, but we still have to wake it up. The 1962 * other pages are released (unbusied) by 1963 * vm_page_wakeup(). We do not set reqpage's 1964 * valid bits here, it is up to the caller. 1965 */ 1966 1967 /* 1968 * NOTE: Can't call pmap_clear_modify(m) from an 1969 * interrupt thread, the pmap code may have to 1970 * map non-kernel pmaps and currently asserts 1971 * the case. 1972 * 1973 * WARNING! The instant SWAPINPROG is 1974 * cleared another cpu may start 1975 * using the mreq page (it will 1976 * check m->valid immediately). 1977 */ 1978 /*pmap_clear_modify(m);*/ 1979 m->valid = VM_PAGE_BITS_ALL; 1980 vm_page_undirty(m); 1981 vm_page_flag_set(m, PG_SWAPPED); 1982 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG); 1983 1984 /* 1985 * We have to wake specifically requested pages 1986 * up too because we cleared SWAPINPROG and 1987 * could be waiting for it in getpages. However, 1988 * be sure to not unbusy getpages specifically 1989 * requested page - getpages expects it to be 1990 * left busy. 1991 * 1992 * bio_driver_info holds the requested page 1993 */ 1994 if (i != (int)(intptr_t)bio->bio_driver_info) { 1995 vm_page_deactivate(m); 1996 vm_page_wakeup(m); 1997 } else { 1998 vm_page_flash(m); 1999 } 2000 } else { 2001 /* 2002 * Mark the page clean but do not mess with the 2003 * pmap-layer's modified state. That state should 2004 * also be clear since the caller protected the 2005 * page VM_PROT_READ, but allow the case. 2006 * 2007 * We are in an interrupt, avoid pmap operations. 2008 * 2009 * If we have a severe page deficit, deactivate the 2010 * page. Do not try to cache it (which would also 2011 * involve a pmap op), because the page might still 2012 * be read-heavy. 2013 * 2014 * When using the swap to cache clean vnode pages 2015 * we do not mess with the page dirty bits. 2016 * 2017 * NOTE! Nobody is waiting for the key mreq page 2018 * on write completion. 2019 */ 2020 vm_page_busy_wait(m, FALSE, "swadpg"); 2021 if (m->object->type == OBJT_SWAP) 2022 vm_page_undirty(m); 2023 vm_page_flag_set(m, PG_SWAPPED); 2024 atomic_clear_int(&m->busy_count, PBUSY_SWAPINPROG); 2025 if (vm_page_count_severe()) 2026 vm_page_deactivate(m); 2027 vm_page_io_finish(m); 2028 if (bio->bio_caller_info1.index & SWBIO_TTC) 2029 vm_page_try_to_cache(m); 2030 else 2031 vm_page_wakeup(m); 2032 } 2033 } 2034 2035 /* 2036 * adjust pip. NOTE: the original parent may still have its own 2037 * pip refs on the object. 2038 */ 2039 2040 if (object) 2041 vm_object_pip_wakeup_n(object, bp->b_xio.xio_npages); 2042 2043 /* 2044 * Release the physical I/O buffer. 2045 * 2046 * NOTE: Due to synchronous operations in the write case b_cmd may 2047 * already be set to BUF_CMD_DONE and BIO_SYNC may have already 2048 * been cleared. 2049 * 2050 * Use vm_token to interlock nsw_rcount/wcount wakeup? 2051 */ 2052 lwkt_gettoken(&vm_token); 2053 if (bio->bio_caller_info1.index & SWBIO_READ) 2054 nswptr = &nsw_rcount; 2055 else if (bio->bio_caller_info1.index & SWBIO_SYNC) 2056 nswptr = &nsw_wcount_sync; 2057 else 2058 nswptr = &nsw_wcount_async; 2059 bp->b_cmd = BUF_CMD_DONE; 2060 relpbuf(bp, nswptr); 2061 lwkt_reltoken(&vm_token); 2062 } 2063 2064 /* 2065 * Fault-in a potentially swapped page and remove the swap reference. 2066 * (used by swapoff code) 2067 * 2068 * object must be held. 2069 */ 2070 static __inline void 2071 swp_pager_fault_page(vm_object_t object, int *sharedp, vm_pindex_t pindex) 2072 { 2073 struct vnode *vp; 2074 vm_page_t m; 2075 int error; 2076 2077 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2078 2079 if (object->type == OBJT_VNODE) { 2080 /* 2081 * Any swap related to a vnode is due to swapcache. We must 2082 * vget() the vnode in case it is not active (otherwise 2083 * vref() will panic). Calling vm_object_page_remove() will 2084 * ensure that any swap ref is removed interlocked with the 2085 * page. clean_only is set to TRUE so we don't throw away 2086 * dirty pages. 2087 */ 2088 vp = object->handle; 2089 error = vget(vp, LK_SHARED | LK_RETRY | LK_CANRECURSE); 2090 if (error == 0) { 2091 vm_object_page_remove(object, pindex, pindex + 1, TRUE); 2092 vput(vp); 2093 } 2094 } else { 2095 /* 2096 * Otherwise it is a normal OBJT_SWAP object and we can 2097 * fault the page in and remove the swap. 2098 */ 2099 m = vm_fault_object_page(object, IDX_TO_OFF(pindex), 2100 VM_PROT_NONE, 2101 VM_FAULT_DIRTY | VM_FAULT_UNSWAP, 2102 sharedp, &error); 2103 if (m) 2104 vm_page_unhold(m); 2105 } 2106 } 2107 2108 /* 2109 * This removes all swap blocks related to a particular device. We have 2110 * to be careful of ripups during the scan. 2111 */ 2112 static int swp_pager_swapoff_callback(struct swblock *swap, void *data); 2113 2114 int 2115 swap_pager_swapoff(int devidx) 2116 { 2117 struct vm_object_hash *hash; 2118 struct swswapoffinfo info; 2119 struct vm_object marker; 2120 vm_object_t object; 2121 int n; 2122 2123 bzero(&marker, sizeof(marker)); 2124 marker.type = OBJT_MARKER; 2125 2126 for (n = 0; n < VMOBJ_HSIZE; ++n) { 2127 hash = &vm_object_hash[n]; 2128 2129 lwkt_gettoken(&hash->token); 2130 TAILQ_INSERT_HEAD(&hash->list, &marker, object_list); 2131 2132 while ((object = TAILQ_NEXT(&marker, object_list)) != NULL) { 2133 if (object->type == OBJT_MARKER) 2134 goto skip; 2135 if (object->type != OBJT_SWAP && 2136 object->type != OBJT_VNODE) 2137 goto skip; 2138 vm_object_hold(object); 2139 if (object->type != OBJT_SWAP && 2140 object->type != OBJT_VNODE) { 2141 vm_object_drop(object); 2142 goto skip; 2143 } 2144 2145 /* 2146 * Object is special in that we can't just pagein 2147 * into vm_page's in it (tmpfs, vn). 2148 */ 2149 if ((object->flags & OBJ_NOPAGEIN) && 2150 RB_ROOT(&object->swblock_root)) { 2151 vm_object_drop(object); 2152 goto skip; 2153 } 2154 2155 info.object = object; 2156 info.shared = 0; 2157 info.devidx = devidx; 2158 swblock_rb_tree_RB_SCAN(&object->swblock_root, 2159 NULL, swp_pager_swapoff_callback, 2160 &info); 2161 vm_object_drop(object); 2162 skip: 2163 if (object == TAILQ_NEXT(&marker, object_list)) { 2164 TAILQ_REMOVE(&hash->list, &marker, object_list); 2165 TAILQ_INSERT_AFTER(&hash->list, object, 2166 &marker, object_list); 2167 } 2168 } 2169 TAILQ_REMOVE(&hash->list, &marker, object_list); 2170 lwkt_reltoken(&hash->token); 2171 } 2172 2173 /* 2174 * If we fail to locate all swblocks we just fail gracefully and 2175 * do not bother to restore paging on the swap device. If the 2176 * user wants to retry the user can retry. 2177 */ 2178 if (swdevt[devidx].sw_nused) 2179 return (1); 2180 else 2181 return (0); 2182 } 2183 2184 static 2185 int 2186 swp_pager_swapoff_callback(struct swblock *swap, void *data) 2187 { 2188 struct swswapoffinfo *info = data; 2189 vm_object_t object = info->object; 2190 vm_pindex_t index; 2191 swblk_t v; 2192 int i; 2193 2194 index = swap->swb_index; 2195 for (i = 0; i < SWAP_META_PAGES; ++i) { 2196 /* 2197 * Make sure we don't race a dying object. This will 2198 * kill the scan of the object's swap blocks entirely. 2199 */ 2200 if (object->flags & OBJ_DEAD) 2201 return(-1); 2202 2203 /* 2204 * Fault the page, which can obviously block. If the swap 2205 * structure disappears break out. 2206 */ 2207 v = swap->swb_pages[i]; 2208 if (v != SWAPBLK_NONE && BLK2DEVIDX(v) == info->devidx) { 2209 swp_pager_fault_page(object, &info->shared, 2210 swap->swb_index + i); 2211 /* swap ptr might go away */ 2212 if (RB_LOOKUP(swblock_rb_tree, 2213 &object->swblock_root, index) != swap) { 2214 break; 2215 } 2216 } 2217 } 2218 return(0); 2219 } 2220 2221 /************************************************************************ 2222 * SWAP META DATA * 2223 ************************************************************************ 2224 * 2225 * These routines manipulate the swap metadata stored in the 2226 * OBJT_SWAP object. 2227 * 2228 * Swap metadata is implemented with a global hash and not directly 2229 * linked into the object. Instead the object simply contains 2230 * appropriate tracking counters. 2231 */ 2232 2233 /* 2234 * Lookup the swblock containing the specified swap block index. 2235 * 2236 * The caller must hold the object. 2237 */ 2238 static __inline 2239 struct swblock * 2240 swp_pager_lookup(vm_object_t object, vm_pindex_t index) 2241 { 2242 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2243 index &= ~(vm_pindex_t)SWAP_META_MASK; 2244 return (RB_LOOKUP(swblock_rb_tree, &object->swblock_root, index)); 2245 } 2246 2247 /* 2248 * Remove a swblock from the RB tree. 2249 * 2250 * The caller must hold the object. 2251 */ 2252 static __inline 2253 void 2254 swp_pager_remove(vm_object_t object, struct swblock *swap) 2255 { 2256 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2257 RB_REMOVE(swblock_rb_tree, &object->swblock_root, swap); 2258 } 2259 2260 /* 2261 * Convert default object to swap object if necessary 2262 * 2263 * The caller must hold the object. 2264 */ 2265 static void 2266 swp_pager_meta_convert(vm_object_t object) 2267 { 2268 if (object->type == OBJT_DEFAULT) { 2269 object->type = OBJT_SWAP; 2270 KKASSERT(object->swblock_count == 0); 2271 } 2272 } 2273 2274 /* 2275 * SWP_PAGER_META_BUILD() - add swap block to swap meta data for object 2276 * 2277 * We first convert the object to a swap object if it is a default 2278 * object. Vnode objects do not need to be converted. 2279 * 2280 * The specified swapblk is added to the object's swap metadata. If 2281 * the swapblk is not valid, it is freed instead. Any previously 2282 * assigned swapblk is freed. 2283 * 2284 * The caller must hold the object. 2285 */ 2286 static void 2287 swp_pager_meta_build(vm_object_t object, vm_pindex_t index, swblk_t swapblk) 2288 { 2289 struct swblock *swap; 2290 struct swblock *oswap; 2291 vm_pindex_t v; 2292 2293 KKASSERT(swapblk != SWAPBLK_NONE); 2294 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2295 2296 /* 2297 * Convert object if necessary 2298 */ 2299 if (object->type == OBJT_DEFAULT) 2300 swp_pager_meta_convert(object); 2301 2302 /* 2303 * Locate swblock. If not found create, but if we aren't adding 2304 * anything just return. If we run out of space in the map we wait 2305 * and, since the hash table may have changed, retry. 2306 */ 2307 retry: 2308 swap = swp_pager_lookup(object, index); 2309 2310 if (swap == NULL) { 2311 int i; 2312 2313 swap = zalloc(swap_zone); 2314 if (swap == NULL) { 2315 vm_wait(0); 2316 goto retry; 2317 } 2318 swap->swb_index = index & ~(vm_pindex_t)SWAP_META_MASK; 2319 swap->swb_count = 0; 2320 2321 ++object->swblock_count; 2322 2323 for (i = 0; i < SWAP_META_PAGES; ++i) 2324 swap->swb_pages[i] = SWAPBLK_NONE; 2325 oswap = RB_INSERT(swblock_rb_tree, &object->swblock_root, swap); 2326 KKASSERT(oswap == NULL); 2327 } 2328 2329 /* 2330 * Delete prior contents of metadata. 2331 * 2332 * NOTE: Decrement swb_count after the freeing operation (which 2333 * might block) to prevent racing destruction of the swblock. 2334 */ 2335 index &= SWAP_META_MASK; 2336 2337 while ((v = swap->swb_pages[index]) != SWAPBLK_NONE) { 2338 swap->swb_pages[index] = SWAPBLK_NONE; 2339 /* can block */ 2340 swp_pager_freeswapspace(object, v, 1); 2341 --swap->swb_count; 2342 --mycpu->gd_vmtotal.t_vm; 2343 } 2344 2345 /* 2346 * Enter block into metadata 2347 */ 2348 swap->swb_pages[index] = swapblk; 2349 if (swapblk != SWAPBLK_NONE) { 2350 ++swap->swb_count; 2351 ++mycpu->gd_vmtotal.t_vm; 2352 } 2353 } 2354 2355 /* 2356 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata 2357 * 2358 * The requested range of blocks is freed, with any associated swap 2359 * returned to the swap bitmap. 2360 * 2361 * This routine will free swap metadata structures as they are cleaned 2362 * out. This routine does *NOT* operate on swap metadata associated 2363 * with resident pages. 2364 * 2365 * The caller must hold the object. 2366 */ 2367 static int swp_pager_meta_free_callback(struct swblock *swb, void *data); 2368 2369 static void 2370 swp_pager_meta_free(vm_object_t object, vm_pindex_t index, vm_pindex_t count) 2371 { 2372 struct swfreeinfo info; 2373 2374 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2375 2376 /* 2377 * Nothing to do 2378 */ 2379 if (object->swblock_count == 0) { 2380 KKASSERT(RB_EMPTY(&object->swblock_root)); 2381 return; 2382 } 2383 if (count == 0) 2384 return; 2385 2386 /* 2387 * Setup for RB tree scan. Note that the pindex range can be huge 2388 * due to the 64 bit page index space so we cannot safely iterate. 2389 */ 2390 info.object = object; 2391 info.basei = index & ~(vm_pindex_t)SWAP_META_MASK; 2392 info.begi = index; 2393 info.endi = index + count - 1; 2394 swblock_rb_tree_RB_SCAN(&object->swblock_root, rb_swblock_scancmp, 2395 swp_pager_meta_free_callback, &info); 2396 } 2397 2398 /* 2399 * The caller must hold the object. 2400 */ 2401 static 2402 int 2403 swp_pager_meta_free_callback(struct swblock *swap, void *data) 2404 { 2405 struct swfreeinfo *info = data; 2406 vm_object_t object = info->object; 2407 int index; 2408 int eindex; 2409 2410 /* 2411 * Figure out the range within the swblock. The wider scan may 2412 * return edge-case swap blocks when the start and/or end points 2413 * are in the middle of a block. 2414 */ 2415 if (swap->swb_index < info->begi) 2416 index = (int)info->begi & SWAP_META_MASK; 2417 else 2418 index = 0; 2419 2420 if (swap->swb_index + SWAP_META_PAGES > info->endi) 2421 eindex = (int)info->endi & SWAP_META_MASK; 2422 else 2423 eindex = SWAP_META_MASK; 2424 2425 /* 2426 * Scan and free the blocks. The loop terminates early 2427 * if (swap) runs out of blocks and could be freed. 2428 * 2429 * NOTE: Decrement swb_count after swp_pager_freeswapspace() 2430 * to deal with a zfree race. 2431 */ 2432 while (index <= eindex) { 2433 swblk_t v = swap->swb_pages[index]; 2434 2435 if (v != SWAPBLK_NONE) { 2436 swap->swb_pages[index] = SWAPBLK_NONE; 2437 /* can block */ 2438 swp_pager_freeswapspace(object, v, 1); 2439 --mycpu->gd_vmtotal.t_vm; 2440 if (--swap->swb_count == 0) { 2441 swp_pager_remove(object, swap); 2442 zfree(swap_zone, swap); 2443 --object->swblock_count; 2444 break; 2445 } 2446 } 2447 ++index; 2448 } 2449 2450 /* swap may be invalid here due to zfree above */ 2451 lwkt_yield(); 2452 2453 return(0); 2454 } 2455 2456 /* 2457 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object 2458 * 2459 * This routine locates and destroys all swap metadata associated with 2460 * an object. 2461 * 2462 * NOTE: Decrement swb_count after the freeing operation (which 2463 * might block) to prevent racing destruction of the swblock. 2464 * 2465 * The caller must hold the object. 2466 */ 2467 static void 2468 swp_pager_meta_free_all(vm_object_t object) 2469 { 2470 struct swblock *swap; 2471 int i; 2472 2473 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2474 2475 while ((swap = RB_ROOT(&object->swblock_root)) != NULL) { 2476 swp_pager_remove(object, swap); 2477 for (i = 0; i < SWAP_META_PAGES; ++i) { 2478 swblk_t v = swap->swb_pages[i]; 2479 if (v != SWAPBLK_NONE) { 2480 /* can block */ 2481 swp_pager_freeswapspace(object, v, 1); 2482 --swap->swb_count; 2483 --mycpu->gd_vmtotal.t_vm; 2484 } 2485 } 2486 if (swap->swb_count != 0) 2487 panic("swap_pager_meta_free_all: swb_count != 0"); 2488 zfree(swap_zone, swap); 2489 --object->swblock_count; 2490 lwkt_yield(); 2491 } 2492 KKASSERT(object->swblock_count == 0); 2493 } 2494 2495 /* 2496 * SWP_PAGER_METACTL() - misc control of swap and vm_page_t meta data. 2497 * 2498 * This routine is capable of looking up, popping, or freeing 2499 * swapblk assignments in the swap meta data or in the vm_page_t. 2500 * The routine typically returns the swapblk being looked-up, or popped, 2501 * or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block 2502 * was invalid. This routine will automatically free any invalid 2503 * meta-data swapblks. 2504 * 2505 * It is not possible to store invalid swapblks in the swap meta data 2506 * (other then a literal 'SWAPBLK_NONE'), so we don't bother checking. 2507 * 2508 * When acting on a busy resident page and paging is in progress, we 2509 * have to wait until paging is complete but otherwise can act on the 2510 * busy page. 2511 * 2512 * SWM_FREE remove and free swap block from metadata 2513 * SWM_POP remove from meta data but do not free.. pop it out 2514 * 2515 * The caller must hold the object. 2516 */ 2517 static swblk_t 2518 swp_pager_meta_ctl(vm_object_t object, vm_pindex_t index, int flags) 2519 { 2520 struct swblock *swap; 2521 swblk_t r1; 2522 2523 if (object->swblock_count == 0) 2524 return(SWAPBLK_NONE); 2525 2526 r1 = SWAPBLK_NONE; 2527 swap = swp_pager_lookup(object, index); 2528 2529 if (swap != NULL) { 2530 index &= SWAP_META_MASK; 2531 r1 = swap->swb_pages[index]; 2532 2533 if (r1 != SWAPBLK_NONE) { 2534 if (flags & (SWM_FREE|SWM_POP)) { 2535 swap->swb_pages[index] = SWAPBLK_NONE; 2536 --mycpu->gd_vmtotal.t_vm; 2537 if (--swap->swb_count == 0) { 2538 swp_pager_remove(object, swap); 2539 zfree(swap_zone, swap); 2540 --object->swblock_count; 2541 } 2542 } 2543 /* swap ptr may be invalid */ 2544 if (flags & SWM_FREE) { 2545 swp_pager_freeswapspace(object, r1, 1); 2546 r1 = SWAPBLK_NONE; 2547 } 2548 } 2549 /* swap ptr may be invalid */ 2550 } 2551 return(r1); 2552 } 2553