1 /* $NetBSD: uvm_swap.c,v 1.168 2014/03/16 05:20:30 dholland Exp $ */ 2 3 /* 4 * Copyright (c) 1995, 1996, 1997, 2009 Matthew R. Green 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 17 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 18 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 19 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 22 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp 29 * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp 30 */ 31 32 #include <sys/cdefs.h> 33 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.168 2014/03/16 05:20:30 dholland Exp $"); 34 35 #include "opt_uvmhist.h" 36 #include "opt_compat_netbsd.h" 37 #include "opt_ddb.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/buf.h> 42 #include <sys/bufq.h> 43 #include <sys/conf.h> 44 #include <sys/proc.h> 45 #include <sys/namei.h> 46 #include <sys/disklabel.h> 47 #include <sys/errno.h> 48 #include <sys/kernel.h> 49 #include <sys/vnode.h> 50 #include <sys/file.h> 51 #include <sys/vmem.h> 52 #include <sys/blist.h> 53 #include <sys/mount.h> 54 #include <sys/pool.h> 55 #include <sys/kmem.h> 56 #include <sys/syscallargs.h> 57 #include <sys/swap.h> 58 #include <sys/kauth.h> 59 #include <sys/sysctl.h> 60 #include <sys/workqueue.h> 61 62 #include <uvm/uvm.h> 63 64 #include <miscfs/specfs/specdev.h> 65 66 /* 67 * uvm_swap.c: manage configuration and i/o to swap space. 68 */ 69 70 /* 71 * swap space is managed in the following way: 72 * 73 * each swap partition or file is described by a "swapdev" structure. 74 * each "swapdev" structure contains a "swapent" structure which contains 75 * information that is passed up to the user (via system calls). 76 * 77 * each swap partition is assigned a "priority" (int) which controls 78 * swap parition usage. 79 * 80 * the system maintains a global data structure describing all swap 81 * partitions/files. there is a sorted LIST of "swappri" structures 82 * which describe "swapdev"'s at that priority. this LIST is headed 83 * by the "swap_priority" global var. each "swappri" contains a 84 * TAILQ of "swapdev" structures at that priority. 85 * 86 * locking: 87 * - swap_syscall_lock (krwlock_t): this lock serializes the swapctl 88 * system call and prevents the swap priority list from changing 89 * while we are in the middle of a system call (e.g. SWAP_STATS). 90 * - uvm_swap_data_lock (kmutex_t): this lock protects all swap data 91 * structures including the priority list, the swapdev structures, 92 * and the swapmap arena. 93 * 94 * each swap device has the following info: 95 * - swap device in use (could be disabled, preventing future use) 96 * - swap enabled (allows new allocations on swap) 97 * - map info in /dev/drum 98 * - vnode pointer 99 * for swap files only: 100 * - block size 101 * - max byte count in buffer 102 * - buffer 103 * 104 * userland controls and configures swap with the swapctl(2) system call. 105 * the sys_swapctl performs the following operations: 106 * [1] SWAP_NSWAP: returns the number of swap devices currently configured 107 * [2] SWAP_STATS: given a pointer to an array of swapent structures 108 * (passed in via "arg") of a size passed in via "misc" ... we load 109 * the current swap config into the array. The actual work is done 110 * in the uvm_swap_stats() function. 111 * [3] SWAP_ON: given a pathname in arg (could be device or file) and a 112 * priority in "misc", start swapping on it. 113 * [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device 114 * [5] SWAP_CTL: changes the priority of a swap device (new priority in 115 * "misc") 116 */ 117 118 /* 119 * swapdev: describes a single swap partition/file 120 * 121 * note the following should be true: 122 * swd_inuse <= swd_nblks [number of blocks in use is <= total blocks] 123 * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel] 124 */ 125 struct swapdev { 126 dev_t swd_dev; /* device id */ 127 int swd_flags; /* flags:inuse/enable/fake */ 128 int swd_priority; /* our priority */ 129 int swd_nblks; /* blocks in this device */ 130 char *swd_path; /* saved pathname of device */ 131 int swd_pathlen; /* length of pathname */ 132 int swd_npages; /* #pages we can use */ 133 int swd_npginuse; /* #pages in use */ 134 int swd_npgbad; /* #pages bad */ 135 int swd_drumoffset; /* page0 offset in drum */ 136 int swd_drumsize; /* #pages in drum */ 137 blist_t swd_blist; /* blist for this swapdev */ 138 struct vnode *swd_vp; /* backing vnode */ 139 TAILQ_ENTRY(swapdev) swd_next; /* priority tailq */ 140 141 int swd_bsize; /* blocksize (bytes) */ 142 int swd_maxactive; /* max active i/o reqs */ 143 struct bufq_state *swd_tab; /* buffer list */ 144 int swd_active; /* number of active buffers */ 145 }; 146 147 /* 148 * swap device priority entry; the list is kept sorted on `spi_priority'. 149 */ 150 struct swappri { 151 int spi_priority; /* priority */ 152 TAILQ_HEAD(spi_swapdev, swapdev) spi_swapdev; 153 /* tailq of swapdevs at this priority */ 154 LIST_ENTRY(swappri) spi_swappri; /* global list of pri's */ 155 }; 156 157 /* 158 * The following two structures are used to keep track of data transfers 159 * on swap devices associated with regular files. 160 * NOTE: this code is more or less a copy of vnd.c; we use the same 161 * structure names here to ease porting.. 162 */ 163 struct vndxfer { 164 struct buf *vx_bp; /* Pointer to parent buffer */ 165 struct swapdev *vx_sdp; 166 int vx_error; 167 int vx_pending; /* # of pending aux buffers */ 168 int vx_flags; 169 #define VX_BUSY 1 170 #define VX_DEAD 2 171 }; 172 173 struct vndbuf { 174 struct buf vb_buf; 175 struct vndxfer *vb_xfer; 176 }; 177 178 /* 179 * NetBSD 1.3 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit 180 * dev_t and has no se_path[] member. 181 */ 182 struct swapent13 { 183 int32_t se13_dev; /* device id */ 184 int se13_flags; /* flags */ 185 int se13_nblks; /* total blocks */ 186 int se13_inuse; /* blocks in use */ 187 int se13_priority; /* priority of this device */ 188 }; 189 190 /* 191 * NetBSD 5.0 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit 192 * dev_t. 193 */ 194 struct swapent50 { 195 int32_t se50_dev; /* device id */ 196 int se50_flags; /* flags */ 197 int se50_nblks; /* total blocks */ 198 int se50_inuse; /* blocks in use */ 199 int se50_priority; /* priority of this device */ 200 char se50_path[PATH_MAX+1]; /* path name */ 201 }; 202 203 /* 204 * We keep a of pool vndbuf's and vndxfer structures. 205 */ 206 static struct pool vndxfer_pool, vndbuf_pool; 207 208 /* 209 * local variables 210 */ 211 static vmem_t *swapmap; /* controls the mapping of /dev/drum */ 212 213 /* list of all active swap devices [by priority] */ 214 LIST_HEAD(swap_priority, swappri); 215 static struct swap_priority swap_priority; 216 217 /* locks */ 218 static krwlock_t swap_syscall_lock; 219 220 /* workqueue and use counter for swap to regular files */ 221 static int sw_reg_count = 0; 222 static struct workqueue *sw_reg_workqueue; 223 224 /* tuneables */ 225 u_int uvm_swapisfull_factor = 99; 226 227 /* 228 * prototypes 229 */ 230 static struct swapdev *swapdrum_getsdp(int); 231 232 static struct swapdev *swaplist_find(struct vnode *, bool); 233 static void swaplist_insert(struct swapdev *, 234 struct swappri *, int); 235 static void swaplist_trim(void); 236 237 static int swap_on(struct lwp *, struct swapdev *); 238 static int swap_off(struct lwp *, struct swapdev *); 239 240 static void sw_reg_strategy(struct swapdev *, struct buf *, int); 241 static void sw_reg_biodone(struct buf *); 242 static void sw_reg_iodone(struct work *wk, void *dummy); 243 static void sw_reg_start(struct swapdev *); 244 245 static int uvm_swap_io(struct vm_page **, int, int, int); 246 247 /* 248 * uvm_swap_init: init the swap system data structures and locks 249 * 250 * => called at boot time from init_main.c after the filesystems 251 * are brought up (which happens after uvm_init()) 252 */ 253 void 254 uvm_swap_init(void) 255 { 256 UVMHIST_FUNC("uvm_swap_init"); 257 258 UVMHIST_CALLED(pdhist); 259 /* 260 * first, init the swap list, its counter, and its lock. 261 * then get a handle on the vnode for /dev/drum by using 262 * the its dev_t number ("swapdev", from MD conf.c). 263 */ 264 265 LIST_INIT(&swap_priority); 266 uvmexp.nswapdev = 0; 267 rw_init(&swap_syscall_lock); 268 mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE); 269 270 if (bdevvp(swapdev, &swapdev_vp)) 271 panic("%s: can't get vnode for swap device", __func__); 272 if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY)) 273 panic("%s: can't lock swap device", __func__); 274 if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED)) 275 panic("%s: can't open swap device", __func__); 276 VOP_UNLOCK(swapdev_vp); 277 278 /* 279 * create swap block resource map to map /dev/drum. the range 280 * from 1 to INT_MAX allows 2 gigablocks of swap space. note 281 * that block 0 is reserved (used to indicate an allocation 282 * failure, or no allocation). 283 */ 284 swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0, 285 VM_NOSLEEP, IPL_NONE); 286 if (swapmap == 0) { 287 panic("%s: vmem_create failed", __func__); 288 } 289 290 pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx", 291 NULL, IPL_BIO); 292 pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd", 293 NULL, IPL_BIO); 294 295 UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0); 296 } 297 298 /* 299 * swaplist functions: functions that operate on the list of swap 300 * devices on the system. 301 */ 302 303 /* 304 * swaplist_insert: insert swap device "sdp" into the global list 305 * 306 * => caller must hold both swap_syscall_lock and uvm_swap_data_lock 307 * => caller must provide a newly allocated swappri structure (we will 308 * FREE it if we don't need it... this it to prevent allocation 309 * blocking here while adding swap) 310 */ 311 static void 312 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority) 313 { 314 struct swappri *spp, *pspp; 315 UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist); 316 317 /* 318 * find entry at or after which to insert the new device. 319 */ 320 pspp = NULL; 321 LIST_FOREACH(spp, &swap_priority, spi_swappri) { 322 if (priority <= spp->spi_priority) 323 break; 324 pspp = spp; 325 } 326 327 /* 328 * new priority? 329 */ 330 if (spp == NULL || spp->spi_priority != priority) { 331 spp = newspp; /* use newspp! */ 332 UVMHIST_LOG(pdhist, "created new swappri = %d", 333 priority, 0, 0, 0); 334 335 spp->spi_priority = priority; 336 TAILQ_INIT(&spp->spi_swapdev); 337 338 if (pspp) 339 LIST_INSERT_AFTER(pspp, spp, spi_swappri); 340 else 341 LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri); 342 } else { 343 /* we don't need a new priority structure, free it */ 344 kmem_free(newspp, sizeof(*newspp)); 345 } 346 347 /* 348 * priority found (or created). now insert on the priority's 349 * tailq list and bump the total number of swapdevs. 350 */ 351 sdp->swd_priority = priority; 352 TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next); 353 uvmexp.nswapdev++; 354 } 355 356 /* 357 * swaplist_find: find and optionally remove a swap device from the 358 * global list. 359 * 360 * => caller must hold both swap_syscall_lock and uvm_swap_data_lock 361 * => we return the swapdev we found (and removed) 362 */ 363 static struct swapdev * 364 swaplist_find(struct vnode *vp, bool remove) 365 { 366 struct swapdev *sdp; 367 struct swappri *spp; 368 369 /* 370 * search the lists for the requested vp 371 */ 372 373 LIST_FOREACH(spp, &swap_priority, spi_swappri) { 374 TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) { 375 if (sdp->swd_vp == vp) { 376 if (remove) { 377 TAILQ_REMOVE(&spp->spi_swapdev, 378 sdp, swd_next); 379 uvmexp.nswapdev--; 380 } 381 return(sdp); 382 } 383 } 384 } 385 return (NULL); 386 } 387 388 /* 389 * swaplist_trim: scan priority list for empty priority entries and kill 390 * them. 391 * 392 * => caller must hold both swap_syscall_lock and uvm_swap_data_lock 393 */ 394 static void 395 swaplist_trim(void) 396 { 397 struct swappri *spp, *nextspp; 398 399 LIST_FOREACH_SAFE(spp, &swap_priority, spi_swappri, nextspp) { 400 if (!TAILQ_EMPTY(&spp->spi_swapdev)) 401 continue; 402 LIST_REMOVE(spp, spi_swappri); 403 kmem_free(spp, sizeof(*spp)); 404 } 405 } 406 407 /* 408 * swapdrum_getsdp: given a page offset in /dev/drum, convert it back 409 * to the "swapdev" that maps that section of the drum. 410 * 411 * => each swapdev takes one big contig chunk of the drum 412 * => caller must hold uvm_swap_data_lock 413 */ 414 static struct swapdev * 415 swapdrum_getsdp(int pgno) 416 { 417 struct swapdev *sdp; 418 struct swappri *spp; 419 420 LIST_FOREACH(spp, &swap_priority, spi_swappri) { 421 TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) { 422 if (sdp->swd_flags & SWF_FAKE) 423 continue; 424 if (pgno >= sdp->swd_drumoffset && 425 pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) { 426 return sdp; 427 } 428 } 429 } 430 return NULL; 431 } 432 433 434 /* 435 * sys_swapctl: main entry point for swapctl(2) system call 436 * [with two helper functions: swap_on and swap_off] 437 */ 438 int 439 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval) 440 { 441 /* { 442 syscallarg(int) cmd; 443 syscallarg(void *) arg; 444 syscallarg(int) misc; 445 } */ 446 struct vnode *vp; 447 struct nameidata nd; 448 struct swappri *spp; 449 struct swapdev *sdp; 450 struct swapent *sep; 451 #define SWAP_PATH_MAX (PATH_MAX + 1) 452 char *userpath; 453 size_t len = 0; 454 int error, misc; 455 int priority; 456 UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist); 457 458 /* 459 * we handle the non-priv NSWAP and STATS request first. 460 * 461 * SWAP_NSWAP: return number of config'd swap devices 462 * [can also be obtained with uvmexp sysctl] 463 */ 464 if (SCARG(uap, cmd) == SWAP_NSWAP) { 465 const int nswapdev = uvmexp.nswapdev; 466 UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", nswapdev, 0, 0, 0); 467 *retval = nswapdev; 468 return 0; 469 } 470 471 misc = SCARG(uap, misc); 472 userpath = kmem_alloc(SWAP_PATH_MAX, KM_SLEEP); 473 474 /* 475 * ensure serialized syscall access by grabbing the swap_syscall_lock 476 */ 477 rw_enter(&swap_syscall_lock, RW_WRITER); 478 479 /* 480 * SWAP_STATS: get stats on current # of configured swap devs 481 * 482 * note that the swap_priority list can't change as long 483 * as we are holding the swap_syscall_lock. we don't want 484 * to grab the uvm_swap_data_lock because we may fault&sleep during 485 * copyout() and we don't want to be holding that lock then! 486 */ 487 if (SCARG(uap, cmd) == SWAP_STATS 488 #if defined(COMPAT_50) 489 || SCARG(uap, cmd) == SWAP_STATS50 490 #endif 491 #if defined(COMPAT_13) 492 || SCARG(uap, cmd) == SWAP_STATS13 493 #endif 494 ) { 495 if ((size_t)misc > (size_t)uvmexp.nswapdev) 496 misc = uvmexp.nswapdev; 497 498 if (misc == 0) { 499 error = EINVAL; 500 goto out; 501 } 502 KASSERT(misc > 0); 503 #if defined(COMPAT_13) 504 if (SCARG(uap, cmd) == SWAP_STATS13) 505 len = sizeof(struct swapent13) * misc; 506 else 507 #endif 508 #if defined(COMPAT_50) 509 if (SCARG(uap, cmd) == SWAP_STATS50) 510 len = sizeof(struct swapent50) * misc; 511 else 512 #endif 513 len = sizeof(struct swapent) * misc; 514 sep = (struct swapent *)kmem_alloc(len, KM_SLEEP); 515 516 uvm_swap_stats(SCARG(uap, cmd), sep, misc, retval); 517 error = copyout(sep, SCARG(uap, arg), len); 518 519 kmem_free(sep, len); 520 UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0); 521 goto out; 522 } 523 if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) { 524 dev_t *devp = (dev_t *)SCARG(uap, arg); 525 526 error = copyout(&dumpdev, devp, sizeof(dumpdev)); 527 goto out; 528 } 529 530 /* 531 * all other requests require superuser privs. verify. 532 */ 533 if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL, 534 0, NULL, NULL, NULL))) 535 goto out; 536 537 if (SCARG(uap, cmd) == SWAP_DUMPOFF) { 538 /* drop the current dump device */ 539 dumpdev = NODEV; 540 dumpcdev = NODEV; 541 cpu_dumpconf(); 542 goto out; 543 } 544 545 /* 546 * at this point we expect a path name in arg. we will 547 * use namei() to gain a vnode reference (vref), and lock 548 * the vnode (VOP_LOCK). 549 * 550 * XXX: a NULL arg means use the root vnode pointer (e.g. for 551 * miniroot) 552 */ 553 if (SCARG(uap, arg) == NULL) { 554 vp = rootvp; /* miniroot */ 555 vref(vp); 556 if (vn_lock(vp, LK_EXCLUSIVE)) { 557 vrele(vp); 558 error = EBUSY; 559 goto out; 560 } 561 if (SCARG(uap, cmd) == SWAP_ON && 562 copystr("miniroot", userpath, SWAP_PATH_MAX, &len)) 563 panic("swapctl: miniroot copy failed"); 564 KASSERT(len > 0); 565 } else { 566 struct pathbuf *pb; 567 568 /* 569 * This used to allow copying in one extra byte 570 * (SWAP_PATH_MAX instead of PATH_MAX) for SWAP_ON. 571 * This was completely pointless because if anyone 572 * used that extra byte namei would fail with 573 * ENAMETOOLONG anyway, so I've removed the excess 574 * logic. - dholland 20100215 575 */ 576 577 error = pathbuf_copyin(SCARG(uap, arg), &pb); 578 if (error) { 579 goto out; 580 } 581 if (SCARG(uap, cmd) == SWAP_ON) { 582 /* get a copy of the string */ 583 pathbuf_copystring(pb, userpath, SWAP_PATH_MAX); 584 len = strlen(userpath) + 1; 585 } 586 NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT, pb); 587 if ((error = namei(&nd))) { 588 pathbuf_destroy(pb); 589 goto out; 590 } 591 vp = nd.ni_vp; 592 pathbuf_destroy(pb); 593 } 594 /* note: "vp" is referenced and locked */ 595 596 error = 0; /* assume no error */ 597 switch(SCARG(uap, cmd)) { 598 599 case SWAP_DUMPDEV: 600 if (vp->v_type != VBLK) { 601 error = ENOTBLK; 602 break; 603 } 604 if (bdevsw_lookup(vp->v_rdev)) { 605 dumpdev = vp->v_rdev; 606 dumpcdev = devsw_blk2chr(dumpdev); 607 } else 608 dumpdev = NODEV; 609 cpu_dumpconf(); 610 break; 611 612 case SWAP_CTL: 613 /* 614 * get new priority, remove old entry (if any) and then 615 * reinsert it in the correct place. finally, prune out 616 * any empty priority structures. 617 */ 618 priority = SCARG(uap, misc); 619 spp = kmem_alloc(sizeof(*spp), KM_SLEEP); 620 mutex_enter(&uvm_swap_data_lock); 621 if ((sdp = swaplist_find(vp, true)) == NULL) { 622 error = ENOENT; 623 } else { 624 swaplist_insert(sdp, spp, priority); 625 swaplist_trim(); 626 } 627 mutex_exit(&uvm_swap_data_lock); 628 if (error) 629 kmem_free(spp, sizeof(*spp)); 630 break; 631 632 case SWAP_ON: 633 634 /* 635 * check for duplicates. if none found, then insert a 636 * dummy entry on the list to prevent someone else from 637 * trying to enable this device while we are working on 638 * it. 639 */ 640 641 priority = SCARG(uap, misc); 642 sdp = kmem_zalloc(sizeof(*sdp), KM_SLEEP); 643 spp = kmem_alloc(sizeof(*spp), KM_SLEEP); 644 sdp->swd_flags = SWF_FAKE; 645 sdp->swd_vp = vp; 646 sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV; 647 bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK); 648 mutex_enter(&uvm_swap_data_lock); 649 if (swaplist_find(vp, false) != NULL) { 650 error = EBUSY; 651 mutex_exit(&uvm_swap_data_lock); 652 bufq_free(sdp->swd_tab); 653 kmem_free(sdp, sizeof(*sdp)); 654 kmem_free(spp, sizeof(*spp)); 655 break; 656 } 657 swaplist_insert(sdp, spp, priority); 658 mutex_exit(&uvm_swap_data_lock); 659 660 KASSERT(len > 0); 661 sdp->swd_pathlen = len; 662 sdp->swd_path = kmem_alloc(len, KM_SLEEP); 663 if (copystr(userpath, sdp->swd_path, len, 0) != 0) 664 panic("swapctl: copystr"); 665 666 /* 667 * we've now got a FAKE placeholder in the swap list. 668 * now attempt to enable swap on it. if we fail, undo 669 * what we've done and kill the fake entry we just inserted. 670 * if swap_on is a success, it will clear the SWF_FAKE flag 671 */ 672 673 if ((error = swap_on(l, sdp)) != 0) { 674 mutex_enter(&uvm_swap_data_lock); 675 (void) swaplist_find(vp, true); /* kill fake entry */ 676 swaplist_trim(); 677 mutex_exit(&uvm_swap_data_lock); 678 bufq_free(sdp->swd_tab); 679 kmem_free(sdp->swd_path, sdp->swd_pathlen); 680 kmem_free(sdp, sizeof(*sdp)); 681 break; 682 } 683 break; 684 685 case SWAP_OFF: 686 mutex_enter(&uvm_swap_data_lock); 687 if ((sdp = swaplist_find(vp, false)) == NULL) { 688 mutex_exit(&uvm_swap_data_lock); 689 error = ENXIO; 690 break; 691 } 692 693 /* 694 * If a device isn't in use or enabled, we 695 * can't stop swapping from it (again). 696 */ 697 if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) { 698 mutex_exit(&uvm_swap_data_lock); 699 error = EBUSY; 700 break; 701 } 702 703 /* 704 * do the real work. 705 */ 706 error = swap_off(l, sdp); 707 break; 708 709 default: 710 error = EINVAL; 711 } 712 713 /* 714 * done! release the ref gained by namei() and unlock. 715 */ 716 vput(vp); 717 out: 718 rw_exit(&swap_syscall_lock); 719 kmem_free(userpath, SWAP_PATH_MAX); 720 721 UVMHIST_LOG(pdhist, "<- done! error=%d", error, 0, 0, 0); 722 return (error); 723 } 724 725 /* 726 * uvm_swap_stats: implements swapctl(SWAP_STATS). The function is kept 727 * away from sys_swapctl() in order to allow COMPAT_* swapctl() 728 * emulation to use it directly without going through sys_swapctl(). 729 * The problem with using sys_swapctl() there is that it involves 730 * copying the swapent array to the stackgap, and this array's size 731 * is not known at build time. Hence it would not be possible to 732 * ensure it would fit in the stackgap in any case. 733 */ 734 void 735 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval) 736 { 737 struct swappri *spp; 738 struct swapdev *sdp; 739 int count = 0; 740 741 LIST_FOREACH(spp, &swap_priority, spi_swappri) { 742 TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) { 743 int inuse; 744 745 if (sec-- <= 0) 746 break; 747 748 /* 749 * backwards compatibility for system call. 750 * For NetBSD 1.3 and 5.0, we have to use 751 * the 32 bit dev_t. For 5.0 and -current 752 * we have to add the path. 753 */ 754 inuse = btodb((uint64_t)sdp->swd_npginuse << 755 PAGE_SHIFT); 756 757 #if defined(COMPAT_13) || defined(COMPAT_50) 758 if (cmd == SWAP_STATS) { 759 #endif 760 sep->se_dev = sdp->swd_dev; 761 sep->se_flags = sdp->swd_flags; 762 sep->se_nblks = sdp->swd_nblks; 763 sep->se_inuse = inuse; 764 sep->se_priority = sdp->swd_priority; 765 KASSERT(sdp->swd_pathlen < 766 sizeof(sep->se_path)); 767 strcpy(sep->se_path, sdp->swd_path); 768 sep++; 769 #if defined(COMPAT_13) 770 } else if (cmd == SWAP_STATS13) { 771 struct swapent13 *sep13 = 772 (struct swapent13 *)sep; 773 774 sep13->se13_dev = sdp->swd_dev; 775 sep13->se13_flags = sdp->swd_flags; 776 sep13->se13_nblks = sdp->swd_nblks; 777 sep13->se13_inuse = inuse; 778 sep13->se13_priority = sdp->swd_priority; 779 sep = (struct swapent *)(sep13 + 1); 780 #endif 781 #if defined(COMPAT_50) 782 } else if (cmd == SWAP_STATS50) { 783 struct swapent50 *sep50 = 784 (struct swapent50 *)sep; 785 786 sep50->se50_dev = sdp->swd_dev; 787 sep50->se50_flags = sdp->swd_flags; 788 sep50->se50_nblks = sdp->swd_nblks; 789 sep50->se50_inuse = inuse; 790 sep50->se50_priority = sdp->swd_priority; 791 KASSERT(sdp->swd_pathlen < 792 sizeof(sep50->se50_path)); 793 strcpy(sep50->se50_path, sdp->swd_path); 794 sep = (struct swapent *)(sep50 + 1); 795 #endif 796 #if defined(COMPAT_13) || defined(COMPAT_50) 797 } 798 #endif 799 count++; 800 } 801 } 802 *retval = count; 803 } 804 805 /* 806 * swap_on: attempt to enable a swapdev for swapping. note that the 807 * swapdev is already on the global list, but disabled (marked 808 * SWF_FAKE). 809 * 810 * => we avoid the start of the disk (to protect disk labels) 811 * => we also avoid the miniroot, if we are swapping to root. 812 * => caller should leave uvm_swap_data_lock unlocked, we may lock it 813 * if needed. 814 */ 815 static int 816 swap_on(struct lwp *l, struct swapdev *sdp) 817 { 818 struct vnode *vp; 819 int error, npages, nblocks, size; 820 long addr; 821 vmem_addr_t result; 822 struct vattr va; 823 dev_t dev; 824 UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist); 825 826 /* 827 * we want to enable swapping on sdp. the swd_vp contains 828 * the vnode we want (locked and ref'd), and the swd_dev 829 * contains the dev_t of the file, if it a block device. 830 */ 831 832 vp = sdp->swd_vp; 833 dev = sdp->swd_dev; 834 835 /* 836 * open the swap file (mostly useful for block device files to 837 * let device driver know what is up). 838 * 839 * we skip the open/close for root on swap because the root 840 * has already been opened when root was mounted (mountroot). 841 */ 842 if (vp != rootvp) { 843 if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred))) 844 return (error); 845 } 846 847 /* XXX this only works for block devices */ 848 UVMHIST_LOG(pdhist, " dev=%d, major(dev)=%d", dev, major(dev), 0,0); 849 850 /* 851 * we now need to determine the size of the swap area. for 852 * block specials we can call the d_psize function. 853 * for normal files, we must stat [get attrs]. 854 * 855 * we put the result in nblks. 856 * for normal files, we also want the filesystem block size 857 * (which we get with statfs). 858 */ 859 switch (vp->v_type) { 860 case VBLK: 861 if ((nblocks = bdev_size(dev)) == -1) { 862 error = ENXIO; 863 goto bad; 864 } 865 break; 866 867 case VREG: 868 if ((error = VOP_GETATTR(vp, &va, l->l_cred))) 869 goto bad; 870 nblocks = (int)btodb(va.va_size); 871 sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift; 872 /* 873 * limit the max # of outstanding I/O requests we issue 874 * at any one time. take it easy on NFS servers. 875 */ 876 if (vp->v_tag == VT_NFS) 877 sdp->swd_maxactive = 2; /* XXX */ 878 else 879 sdp->swd_maxactive = 8; /* XXX */ 880 break; 881 882 default: 883 error = ENXIO; 884 goto bad; 885 } 886 887 /* 888 * save nblocks in a safe place and convert to pages. 889 */ 890 891 sdp->swd_nblks = nblocks; 892 npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT; 893 894 /* 895 * for block special files, we want to make sure that leave 896 * the disklabel and bootblocks alone, so we arrange to skip 897 * over them (arbitrarily choosing to skip PAGE_SIZE bytes). 898 * note that because of this the "size" can be less than the 899 * actual number of blocks on the device. 900 */ 901 if (vp->v_type == VBLK) { 902 /* we use pages 1 to (size - 1) [inclusive] */ 903 size = npages - 1; 904 addr = 1; 905 } else { 906 /* we use pages 0 to (size - 1) [inclusive] */ 907 size = npages; 908 addr = 0; 909 } 910 911 /* 912 * make sure we have enough blocks for a reasonable sized swap 913 * area. we want at least one page. 914 */ 915 916 if (size < 1) { 917 UVMHIST_LOG(pdhist, " size <= 1!!", 0, 0, 0, 0); 918 error = EINVAL; 919 goto bad; 920 } 921 922 UVMHIST_LOG(pdhist, " dev=%x: size=%d addr=%ld\n", dev, size, addr, 0); 923 924 /* 925 * now we need to allocate an extent to manage this swap device 926 */ 927 928 sdp->swd_blist = blist_create(npages); 929 /* mark all expect the `saved' region free. */ 930 blist_free(sdp->swd_blist, addr, size); 931 932 /* 933 * if the vnode we are swapping to is the root vnode 934 * (i.e. we are swapping to the miniroot) then we want 935 * to make sure we don't overwrite it. do a statfs to 936 * find its size and skip over it. 937 */ 938 if (vp == rootvp) { 939 struct mount *mp; 940 struct statvfs *sp; 941 int rootblocks, rootpages; 942 943 mp = rootvnode->v_mount; 944 sp = &mp->mnt_stat; 945 rootblocks = sp->f_blocks * btodb(sp->f_frsize); 946 /* 947 * XXX: sp->f_blocks isn't the total number of 948 * blocks in the filesystem, it's the number of 949 * data blocks. so, our rootblocks almost 950 * definitely underestimates the total size 951 * of the filesystem - how badly depends on the 952 * details of the filesystem type. there isn't 953 * an obvious way to deal with this cleanly 954 * and perfectly, so for now we just pad our 955 * rootblocks estimate with an extra 5 percent. 956 */ 957 rootblocks += (rootblocks >> 5) + 958 (rootblocks >> 6) + 959 (rootblocks >> 7); 960 rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT; 961 if (rootpages > size) 962 panic("swap_on: miniroot larger than swap?"); 963 964 if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) { 965 panic("swap_on: unable to preserve miniroot"); 966 } 967 968 size -= rootpages; 969 printf("Preserved %d pages of miniroot ", rootpages); 970 printf("leaving %d pages of swap\n", size); 971 } 972 973 /* 974 * add a ref to vp to reflect usage as a swap device. 975 */ 976 vref(vp); 977 978 /* 979 * now add the new swapdev to the drum and enable. 980 */ 981 error = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP, &result); 982 if (error != 0) 983 panic("swapdrum_add"); 984 /* 985 * If this is the first regular swap create the workqueue. 986 * => Protected by swap_syscall_lock. 987 */ 988 if (vp->v_type != VBLK) { 989 if (sw_reg_count++ == 0) { 990 KASSERT(sw_reg_workqueue == NULL); 991 if (workqueue_create(&sw_reg_workqueue, "swapiod", 992 sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0) 993 panic("%s: workqueue_create failed", __func__); 994 } 995 } 996 997 sdp->swd_drumoffset = (int)result; 998 sdp->swd_drumsize = npages; 999 sdp->swd_npages = size; 1000 mutex_enter(&uvm_swap_data_lock); 1001 sdp->swd_flags &= ~SWF_FAKE; /* going live */ 1002 sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE); 1003 uvmexp.swpages += size; 1004 uvmexp.swpgavail += size; 1005 mutex_exit(&uvm_swap_data_lock); 1006 return (0); 1007 1008 /* 1009 * failure: clean up and return error. 1010 */ 1011 1012 bad: 1013 if (sdp->swd_blist) { 1014 blist_destroy(sdp->swd_blist); 1015 } 1016 if (vp != rootvp) { 1017 (void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred); 1018 } 1019 return (error); 1020 } 1021 1022 /* 1023 * swap_off: stop swapping on swapdev 1024 * 1025 * => swap data should be locked, we will unlock. 1026 */ 1027 static int 1028 swap_off(struct lwp *l, struct swapdev *sdp) 1029 { 1030 int npages = sdp->swd_npages; 1031 int error = 0; 1032 1033 UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist); 1034 UVMHIST_LOG(pdhist, " dev=%x, npages=%d", sdp->swd_dev,npages,0,0); 1035 1036 /* disable the swap area being removed */ 1037 sdp->swd_flags &= ~SWF_ENABLE; 1038 uvmexp.swpgavail -= npages; 1039 mutex_exit(&uvm_swap_data_lock); 1040 1041 /* 1042 * the idea is to find all the pages that are paged out to this 1043 * device, and page them all in. in uvm, swap-backed pageable 1044 * memory can take two forms: aobjs and anons. call the 1045 * swapoff hook for each subsystem to bring in pages. 1046 */ 1047 1048 if (uao_swap_off(sdp->swd_drumoffset, 1049 sdp->swd_drumoffset + sdp->swd_drumsize) || 1050 amap_swap_off(sdp->swd_drumoffset, 1051 sdp->swd_drumoffset + sdp->swd_drumsize)) { 1052 error = ENOMEM; 1053 } else if (sdp->swd_npginuse > sdp->swd_npgbad) { 1054 error = EBUSY; 1055 } 1056 1057 if (error) { 1058 mutex_enter(&uvm_swap_data_lock); 1059 sdp->swd_flags |= SWF_ENABLE; 1060 uvmexp.swpgavail += npages; 1061 mutex_exit(&uvm_swap_data_lock); 1062 1063 return error; 1064 } 1065 1066 /* 1067 * If this is the last regular swap destroy the workqueue. 1068 * => Protected by swap_syscall_lock. 1069 */ 1070 if (sdp->swd_vp->v_type != VBLK) { 1071 KASSERT(sw_reg_count > 0); 1072 KASSERT(sw_reg_workqueue != NULL); 1073 if (--sw_reg_count == 0) { 1074 workqueue_destroy(sw_reg_workqueue); 1075 sw_reg_workqueue = NULL; 1076 } 1077 } 1078 1079 /* 1080 * done with the vnode. 1081 * drop our ref on the vnode before calling VOP_CLOSE() 1082 * so that spec_close() can tell if this is the last close. 1083 */ 1084 vrele(sdp->swd_vp); 1085 if (sdp->swd_vp != rootvp) { 1086 (void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred); 1087 } 1088 1089 mutex_enter(&uvm_swap_data_lock); 1090 uvmexp.swpages -= npages; 1091 uvmexp.swpginuse -= sdp->swd_npgbad; 1092 1093 if (swaplist_find(sdp->swd_vp, true) == NULL) 1094 panic("%s: swapdev not in list", __func__); 1095 swaplist_trim(); 1096 mutex_exit(&uvm_swap_data_lock); 1097 1098 /* 1099 * free all resources! 1100 */ 1101 vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize); 1102 blist_destroy(sdp->swd_blist); 1103 bufq_free(sdp->swd_tab); 1104 kmem_free(sdp, sizeof(*sdp)); 1105 return (0); 1106 } 1107 1108 void 1109 uvm_swap_shutdown(struct lwp *l) 1110 { 1111 struct swapdev *sdp; 1112 struct swappri *spp; 1113 struct vnode *vp; 1114 int error; 1115 1116 printf("turning of swap..."); 1117 rw_enter(&swap_syscall_lock, RW_WRITER); 1118 mutex_enter(&uvm_swap_data_lock); 1119 again: 1120 LIST_FOREACH(spp, &swap_priority, spi_swappri) 1121 TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) { 1122 if (sdp->swd_flags & SWF_FAKE) 1123 continue; 1124 if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) 1125 continue; 1126 #ifdef DEBUG 1127 printf("\nturning off swap on %s...", 1128 sdp->swd_path); 1129 #endif 1130 if (vn_lock(vp = sdp->swd_vp, LK_EXCLUSIVE)) { 1131 error = EBUSY; 1132 vp = NULL; 1133 } else 1134 error = 0; 1135 if (!error) { 1136 error = swap_off(l, sdp); 1137 mutex_enter(&uvm_swap_data_lock); 1138 } 1139 if (error) { 1140 printf("stopping swap on %s failed " 1141 "with error %d\n", sdp->swd_path, error); 1142 TAILQ_REMOVE(&spp->spi_swapdev, sdp, 1143 swd_next); 1144 uvmexp.nswapdev--; 1145 swaplist_trim(); 1146 if (vp) 1147 vput(vp); 1148 } 1149 goto again; 1150 } 1151 printf(" done\n"); 1152 mutex_exit(&uvm_swap_data_lock); 1153 rw_exit(&swap_syscall_lock); 1154 } 1155 1156 1157 /* 1158 * /dev/drum interface and i/o functions 1159 */ 1160 1161 /* 1162 * swstrategy: perform I/O on the drum 1163 * 1164 * => we must map the i/o request from the drum to the correct swapdev. 1165 */ 1166 static void 1167 swstrategy(struct buf *bp) 1168 { 1169 struct swapdev *sdp; 1170 struct vnode *vp; 1171 int pageno, bn; 1172 UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist); 1173 1174 /* 1175 * convert block number to swapdev. note that swapdev can't 1176 * be yanked out from under us because we are holding resources 1177 * in it (i.e. the blocks we are doing I/O on). 1178 */ 1179 pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT; 1180 mutex_enter(&uvm_swap_data_lock); 1181 sdp = swapdrum_getsdp(pageno); 1182 mutex_exit(&uvm_swap_data_lock); 1183 if (sdp == NULL) { 1184 bp->b_error = EINVAL; 1185 bp->b_resid = bp->b_bcount; 1186 biodone(bp); 1187 UVMHIST_LOG(pdhist, " failed to get swap device", 0, 0, 0, 0); 1188 return; 1189 } 1190 1191 /* 1192 * convert drum page number to block number on this swapdev. 1193 */ 1194 1195 pageno -= sdp->swd_drumoffset; /* page # on swapdev */ 1196 bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */ 1197 1198 UVMHIST_LOG(pdhist, " %s: mapoff=%x bn=%x bcount=%ld", 1199 ((bp->b_flags & B_READ) == 0) ? "write" : "read", 1200 sdp->swd_drumoffset, bn, bp->b_bcount); 1201 1202 /* 1203 * for block devices we finish up here. 1204 * for regular files we have to do more work which we delegate 1205 * to sw_reg_strategy(). 1206 */ 1207 1208 vp = sdp->swd_vp; /* swapdev vnode pointer */ 1209 switch (vp->v_type) { 1210 default: 1211 panic("%s: vnode type 0x%x", __func__, vp->v_type); 1212 1213 case VBLK: 1214 1215 /* 1216 * must convert "bp" from an I/O on /dev/drum to an I/O 1217 * on the swapdev (sdp). 1218 */ 1219 bp->b_blkno = bn; /* swapdev block number */ 1220 bp->b_dev = sdp->swd_dev; /* swapdev dev_t */ 1221 1222 /* 1223 * if we are doing a write, we have to redirect the i/o on 1224 * drum's v_numoutput counter to the swapdevs. 1225 */ 1226 if ((bp->b_flags & B_READ) == 0) { 1227 mutex_enter(bp->b_objlock); 1228 vwakeup(bp); /* kills one 'v_numoutput' on drum */ 1229 mutex_exit(bp->b_objlock); 1230 mutex_enter(vp->v_interlock); 1231 vp->v_numoutput++; /* put it on swapdev */ 1232 mutex_exit(vp->v_interlock); 1233 } 1234 1235 /* 1236 * finally plug in swapdev vnode and start I/O 1237 */ 1238 bp->b_vp = vp; 1239 bp->b_objlock = vp->v_interlock; 1240 VOP_STRATEGY(vp, bp); 1241 return; 1242 1243 case VREG: 1244 /* 1245 * delegate to sw_reg_strategy function. 1246 */ 1247 sw_reg_strategy(sdp, bp, bn); 1248 return; 1249 } 1250 /* NOTREACHED */ 1251 } 1252 1253 /* 1254 * swread: the read function for the drum (just a call to physio) 1255 */ 1256 /*ARGSUSED*/ 1257 static int 1258 swread(dev_t dev, struct uio *uio, int ioflag) 1259 { 1260 UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist); 1261 1262 UVMHIST_LOG(pdhist, " dev=%x offset=%qx", dev, uio->uio_offset, 0, 0); 1263 return (physio(swstrategy, NULL, dev, B_READ, minphys, uio)); 1264 } 1265 1266 /* 1267 * swwrite: the write function for the drum (just a call to physio) 1268 */ 1269 /*ARGSUSED*/ 1270 static int 1271 swwrite(dev_t dev, struct uio *uio, int ioflag) 1272 { 1273 UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist); 1274 1275 UVMHIST_LOG(pdhist, " dev=%x offset=%qx", dev, uio->uio_offset, 0, 0); 1276 return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio)); 1277 } 1278 1279 const struct bdevsw swap_bdevsw = { 1280 .d_open = nullopen, 1281 .d_close = nullclose, 1282 .d_strategy = swstrategy, 1283 .d_ioctl = noioctl, 1284 .d_dump = nodump, 1285 .d_psize = nosize, 1286 .d_flag = D_OTHER 1287 }; 1288 1289 const struct cdevsw swap_cdevsw = { 1290 .d_open = nullopen, 1291 .d_close = nullclose, 1292 .d_read = swread, 1293 .d_write = swwrite, 1294 .d_ioctl = noioctl, 1295 .d_stop = nostop, 1296 .d_tty = notty, 1297 .d_poll = nopoll, 1298 .d_mmap = nommap, 1299 .d_kqfilter = nokqfilter, 1300 .d_flag = D_OTHER, 1301 }; 1302 1303 /* 1304 * sw_reg_strategy: handle swap i/o to regular files 1305 */ 1306 static void 1307 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn) 1308 { 1309 struct vnode *vp; 1310 struct vndxfer *vnx; 1311 daddr_t nbn; 1312 char *addr; 1313 off_t byteoff; 1314 int s, off, nra, error, sz, resid; 1315 UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist); 1316 1317 /* 1318 * allocate a vndxfer head for this transfer and point it to 1319 * our buffer. 1320 */ 1321 vnx = pool_get(&vndxfer_pool, PR_WAITOK); 1322 vnx->vx_flags = VX_BUSY; 1323 vnx->vx_error = 0; 1324 vnx->vx_pending = 0; 1325 vnx->vx_bp = bp; 1326 vnx->vx_sdp = sdp; 1327 1328 /* 1329 * setup for main loop where we read filesystem blocks into 1330 * our buffer. 1331 */ 1332 error = 0; 1333 bp->b_resid = bp->b_bcount; /* nothing transfered yet! */ 1334 addr = bp->b_data; /* current position in buffer */ 1335 byteoff = dbtob((uint64_t)bn); 1336 1337 for (resid = bp->b_resid; resid; resid -= sz) { 1338 struct vndbuf *nbp; 1339 1340 /* 1341 * translate byteoffset into block number. return values: 1342 * vp = vnode of underlying device 1343 * nbn = new block number (on underlying vnode dev) 1344 * nra = num blocks we can read-ahead (excludes requested 1345 * block) 1346 */ 1347 nra = 0; 1348 error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize, 1349 &vp, &nbn, &nra); 1350 1351 if (error == 0 && nbn == (daddr_t)-1) { 1352 /* 1353 * this used to just set error, but that doesn't 1354 * do the right thing. Instead, it causes random 1355 * memory errors. The panic() should remain until 1356 * this condition doesn't destabilize the system. 1357 */ 1358 #if 1 1359 panic("%s: swap to sparse file", __func__); 1360 #else 1361 error = EIO; /* failure */ 1362 #endif 1363 } 1364 1365 /* 1366 * punt if there was an error or a hole in the file. 1367 * we must wait for any i/o ops we have already started 1368 * to finish before returning. 1369 * 1370 * XXX we could deal with holes here but it would be 1371 * a hassle (in the write case). 1372 */ 1373 if (error) { 1374 s = splbio(); 1375 vnx->vx_error = error; /* pass error up */ 1376 goto out; 1377 } 1378 1379 /* 1380 * compute the size ("sz") of this transfer (in bytes). 1381 */ 1382 off = byteoff % sdp->swd_bsize; 1383 sz = (1 + nra) * sdp->swd_bsize - off; 1384 if (sz > resid) 1385 sz = resid; 1386 1387 UVMHIST_LOG(pdhist, "sw_reg_strategy: " 1388 "vp %p/%p offset 0x%x/0x%x", 1389 sdp->swd_vp, vp, byteoff, nbn); 1390 1391 /* 1392 * now get a buf structure. note that the vb_buf is 1393 * at the front of the nbp structure so that you can 1394 * cast pointers between the two structure easily. 1395 */ 1396 nbp = pool_get(&vndbuf_pool, PR_WAITOK); 1397 buf_init(&nbp->vb_buf); 1398 nbp->vb_buf.b_flags = bp->b_flags; 1399 nbp->vb_buf.b_cflags = bp->b_cflags; 1400 nbp->vb_buf.b_oflags = bp->b_oflags; 1401 nbp->vb_buf.b_bcount = sz; 1402 nbp->vb_buf.b_bufsize = sz; 1403 nbp->vb_buf.b_error = 0; 1404 nbp->vb_buf.b_data = addr; 1405 nbp->vb_buf.b_lblkno = 0; 1406 nbp->vb_buf.b_blkno = nbn + btodb(off); 1407 nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno; 1408 nbp->vb_buf.b_iodone = sw_reg_biodone; 1409 nbp->vb_buf.b_vp = vp; 1410 nbp->vb_buf.b_objlock = vp->v_interlock; 1411 if (vp->v_type == VBLK) { 1412 nbp->vb_buf.b_dev = vp->v_rdev; 1413 } 1414 1415 nbp->vb_xfer = vnx; /* patch it back in to vnx */ 1416 1417 /* 1418 * Just sort by block number 1419 */ 1420 s = splbio(); 1421 if (vnx->vx_error != 0) { 1422 buf_destroy(&nbp->vb_buf); 1423 pool_put(&vndbuf_pool, nbp); 1424 goto out; 1425 } 1426 vnx->vx_pending++; 1427 1428 /* sort it in and start I/O if we are not over our limit */ 1429 /* XXXAD locking */ 1430 bufq_put(sdp->swd_tab, &nbp->vb_buf); 1431 sw_reg_start(sdp); 1432 splx(s); 1433 1434 /* 1435 * advance to the next I/O 1436 */ 1437 byteoff += sz; 1438 addr += sz; 1439 } 1440 1441 s = splbio(); 1442 1443 out: /* Arrive here at splbio */ 1444 vnx->vx_flags &= ~VX_BUSY; 1445 if (vnx->vx_pending == 0) { 1446 error = vnx->vx_error; 1447 pool_put(&vndxfer_pool, vnx); 1448 bp->b_error = error; 1449 biodone(bp); 1450 } 1451 splx(s); 1452 } 1453 1454 /* 1455 * sw_reg_start: start an I/O request on the requested swapdev 1456 * 1457 * => reqs are sorted by b_rawblkno (above) 1458 */ 1459 static void 1460 sw_reg_start(struct swapdev *sdp) 1461 { 1462 struct buf *bp; 1463 struct vnode *vp; 1464 UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist); 1465 1466 /* recursion control */ 1467 if ((sdp->swd_flags & SWF_BUSY) != 0) 1468 return; 1469 1470 sdp->swd_flags |= SWF_BUSY; 1471 1472 while (sdp->swd_active < sdp->swd_maxactive) { 1473 bp = bufq_get(sdp->swd_tab); 1474 if (bp == NULL) 1475 break; 1476 sdp->swd_active++; 1477 1478 UVMHIST_LOG(pdhist, 1479 "sw_reg_start: bp %p vp %p blkno %p cnt %lx", 1480 bp, bp->b_vp, bp->b_blkno, bp->b_bcount); 1481 vp = bp->b_vp; 1482 KASSERT(bp->b_objlock == vp->v_interlock); 1483 if ((bp->b_flags & B_READ) == 0) { 1484 mutex_enter(vp->v_interlock); 1485 vp->v_numoutput++; 1486 mutex_exit(vp->v_interlock); 1487 } 1488 VOP_STRATEGY(vp, bp); 1489 } 1490 sdp->swd_flags &= ~SWF_BUSY; 1491 } 1492 1493 /* 1494 * sw_reg_biodone: one of our i/o's has completed 1495 */ 1496 static void 1497 sw_reg_biodone(struct buf *bp) 1498 { 1499 workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL); 1500 } 1501 1502 /* 1503 * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup 1504 * 1505 * => note that we can recover the vndbuf struct by casting the buf ptr 1506 */ 1507 static void 1508 sw_reg_iodone(struct work *wk, void *dummy) 1509 { 1510 struct vndbuf *vbp = (void *)wk; 1511 struct vndxfer *vnx = vbp->vb_xfer; 1512 struct buf *pbp = vnx->vx_bp; /* parent buffer */ 1513 struct swapdev *sdp = vnx->vx_sdp; 1514 int s, resid, error; 1515 KASSERT(&vbp->vb_buf.b_work == wk); 1516 UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist); 1517 1518 UVMHIST_LOG(pdhist, " vbp=%p vp=%p blkno=%x addr=%p", 1519 vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data); 1520 UVMHIST_LOG(pdhist, " cnt=%lx resid=%lx", 1521 vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0); 1522 1523 /* 1524 * protect vbp at splbio and update. 1525 */ 1526 1527 s = splbio(); 1528 resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid; 1529 pbp->b_resid -= resid; 1530 vnx->vx_pending--; 1531 1532 if (vbp->vb_buf.b_error != 0) { 1533 /* pass error upward */ 1534 error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO; 1535 UVMHIST_LOG(pdhist, " got error=%d !", error, 0, 0, 0); 1536 vnx->vx_error = error; 1537 } 1538 1539 /* 1540 * kill vbp structure 1541 */ 1542 buf_destroy(&vbp->vb_buf); 1543 pool_put(&vndbuf_pool, vbp); 1544 1545 /* 1546 * wrap up this transaction if it has run to completion or, in 1547 * case of an error, when all auxiliary buffers have returned. 1548 */ 1549 if (vnx->vx_error != 0) { 1550 /* pass error upward */ 1551 error = vnx->vx_error; 1552 if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) { 1553 pbp->b_error = error; 1554 biodone(pbp); 1555 pool_put(&vndxfer_pool, vnx); 1556 } 1557 } else if (pbp->b_resid == 0) { 1558 KASSERT(vnx->vx_pending == 0); 1559 if ((vnx->vx_flags & VX_BUSY) == 0) { 1560 UVMHIST_LOG(pdhist, " iodone error=%d !", 1561 pbp, vnx->vx_error, 0, 0); 1562 biodone(pbp); 1563 pool_put(&vndxfer_pool, vnx); 1564 } 1565 } 1566 1567 /* 1568 * done! start next swapdev I/O if one is pending 1569 */ 1570 sdp->swd_active--; 1571 sw_reg_start(sdp); 1572 splx(s); 1573 } 1574 1575 1576 /* 1577 * uvm_swap_alloc: allocate space on swap 1578 * 1579 * => allocation is done "round robin" down the priority list, as we 1580 * allocate in a priority we "rotate" the circle queue. 1581 * => space can be freed with uvm_swap_free 1582 * => we return the page slot number in /dev/drum (0 == invalid slot) 1583 * => we lock uvm_swap_data_lock 1584 * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM 1585 */ 1586 int 1587 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok) 1588 { 1589 struct swapdev *sdp; 1590 struct swappri *spp; 1591 UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist); 1592 1593 /* 1594 * no swap devices configured yet? definite failure. 1595 */ 1596 if (uvmexp.nswapdev < 1) 1597 return 0; 1598 1599 /* 1600 * XXXJAK: BEGIN HACK 1601 * 1602 * blist_alloc() in subr_blist.c will panic if we try to allocate 1603 * too many slots. 1604 */ 1605 if (*nslots > BLIST_MAX_ALLOC) { 1606 if (__predict_false(lessok == false)) 1607 return 0; 1608 *nslots = BLIST_MAX_ALLOC; 1609 } 1610 /* XXXJAK: END HACK */ 1611 1612 /* 1613 * lock data lock, convert slots into blocks, and enter loop 1614 */ 1615 mutex_enter(&uvm_swap_data_lock); 1616 1617 ReTry: /* XXXMRG */ 1618 LIST_FOREACH(spp, &swap_priority, spi_swappri) { 1619 TAILQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) { 1620 uint64_t result; 1621 1622 /* if it's not enabled, then we can't swap from it */ 1623 if ((sdp->swd_flags & SWF_ENABLE) == 0) 1624 continue; 1625 if (sdp->swd_npginuse + *nslots > sdp->swd_npages) 1626 continue; 1627 result = blist_alloc(sdp->swd_blist, *nslots); 1628 if (result == BLIST_NONE) { 1629 continue; 1630 } 1631 KASSERT(result < sdp->swd_drumsize); 1632 1633 /* 1634 * successful allocation! now rotate the tailq. 1635 */ 1636 TAILQ_REMOVE(&spp->spi_swapdev, sdp, swd_next); 1637 TAILQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next); 1638 sdp->swd_npginuse += *nslots; 1639 uvmexp.swpginuse += *nslots; 1640 mutex_exit(&uvm_swap_data_lock); 1641 /* done! return drum slot number */ 1642 UVMHIST_LOG(pdhist, 1643 "success! returning %d slots starting at %d", 1644 *nslots, result + sdp->swd_drumoffset, 0, 0); 1645 return (result + sdp->swd_drumoffset); 1646 } 1647 } 1648 1649 /* XXXMRG: BEGIN HACK */ 1650 if (*nslots > 1 && lessok) { 1651 *nslots = 1; 1652 /* XXXMRG: ugh! blist should support this for us */ 1653 goto ReTry; 1654 } 1655 /* XXXMRG: END HACK */ 1656 1657 mutex_exit(&uvm_swap_data_lock); 1658 return 0; 1659 } 1660 1661 /* 1662 * uvm_swapisfull: return true if most of available swap is allocated 1663 * and in use. we don't count some small portion as it may be inaccessible 1664 * to us at any given moment, for example if there is lock contention or if 1665 * pages are busy. 1666 */ 1667 bool 1668 uvm_swapisfull(void) 1669 { 1670 int swpgonly; 1671 bool rv; 1672 1673 mutex_enter(&uvm_swap_data_lock); 1674 KASSERT(uvmexp.swpgonly <= uvmexp.swpages); 1675 swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 / 1676 uvm_swapisfull_factor); 1677 rv = (swpgonly >= uvmexp.swpgavail); 1678 mutex_exit(&uvm_swap_data_lock); 1679 1680 return (rv); 1681 } 1682 1683 /* 1684 * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors 1685 * 1686 * => we lock uvm_swap_data_lock 1687 */ 1688 void 1689 uvm_swap_markbad(int startslot, int nslots) 1690 { 1691 struct swapdev *sdp; 1692 UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist); 1693 1694 mutex_enter(&uvm_swap_data_lock); 1695 sdp = swapdrum_getsdp(startslot); 1696 KASSERT(sdp != NULL); 1697 1698 /* 1699 * we just keep track of how many pages have been marked bad 1700 * in this device, to make everything add up in swap_off(). 1701 * we assume here that the range of slots will all be within 1702 * one swap device. 1703 */ 1704 1705 KASSERT(uvmexp.swpgonly >= nslots); 1706 uvmexp.swpgonly -= nslots; 1707 sdp->swd_npgbad += nslots; 1708 UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0); 1709 mutex_exit(&uvm_swap_data_lock); 1710 } 1711 1712 /* 1713 * uvm_swap_free: free swap slots 1714 * 1715 * => this can be all or part of an allocation made by uvm_swap_alloc 1716 * => we lock uvm_swap_data_lock 1717 */ 1718 void 1719 uvm_swap_free(int startslot, int nslots) 1720 { 1721 struct swapdev *sdp; 1722 UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist); 1723 1724 UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots, 1725 startslot, 0, 0); 1726 1727 /* 1728 * ignore attempts to free the "bad" slot. 1729 */ 1730 1731 if (startslot == SWSLOT_BAD) { 1732 return; 1733 } 1734 1735 /* 1736 * convert drum slot offset back to sdp, free the blocks 1737 * in the extent, and return. must hold pri lock to do 1738 * lookup and access the extent. 1739 */ 1740 1741 mutex_enter(&uvm_swap_data_lock); 1742 sdp = swapdrum_getsdp(startslot); 1743 KASSERT(uvmexp.nswapdev >= 1); 1744 KASSERT(sdp != NULL); 1745 KASSERT(sdp->swd_npginuse >= nslots); 1746 blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots); 1747 sdp->swd_npginuse -= nslots; 1748 uvmexp.swpginuse -= nslots; 1749 mutex_exit(&uvm_swap_data_lock); 1750 } 1751 1752 /* 1753 * uvm_swap_put: put any number of pages into a contig place on swap 1754 * 1755 * => can be sync or async 1756 */ 1757 1758 int 1759 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags) 1760 { 1761 int error; 1762 1763 error = uvm_swap_io(ppsp, swslot, npages, B_WRITE | 1764 ((flags & PGO_SYNCIO) ? 0 : B_ASYNC)); 1765 return error; 1766 } 1767 1768 /* 1769 * uvm_swap_get: get a single page from swap 1770 * 1771 * => usually a sync op (from fault) 1772 */ 1773 1774 int 1775 uvm_swap_get(struct vm_page *page, int swslot, int flags) 1776 { 1777 int error; 1778 1779 uvmexp.nswget++; 1780 KASSERT(flags & PGO_SYNCIO); 1781 if (swslot == SWSLOT_BAD) { 1782 return EIO; 1783 } 1784 1785 error = uvm_swap_io(&page, swslot, 1, B_READ | 1786 ((flags & PGO_SYNCIO) ? 0 : B_ASYNC)); 1787 if (error == 0) { 1788 1789 /* 1790 * this page is no longer only in swap. 1791 */ 1792 1793 mutex_enter(&uvm_swap_data_lock); 1794 KASSERT(uvmexp.swpgonly > 0); 1795 uvmexp.swpgonly--; 1796 mutex_exit(&uvm_swap_data_lock); 1797 } 1798 return error; 1799 } 1800 1801 /* 1802 * uvm_swap_io: do an i/o operation to swap 1803 */ 1804 1805 static int 1806 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags) 1807 { 1808 daddr_t startblk; 1809 struct buf *bp; 1810 vaddr_t kva; 1811 int error, mapinflags; 1812 bool write, async; 1813 UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist); 1814 1815 UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d", 1816 startslot, npages, flags, 0); 1817 1818 write = (flags & B_READ) == 0; 1819 async = (flags & B_ASYNC) != 0; 1820 1821 /* 1822 * allocate a buf for the i/o. 1823 */ 1824 1825 KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async)); 1826 bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp); 1827 if (bp == NULL) { 1828 uvm_aio_aiodone_pages(pps, npages, true, ENOMEM); 1829 return ENOMEM; 1830 } 1831 1832 /* 1833 * convert starting drum slot to block number 1834 */ 1835 1836 startblk = btodb((uint64_t)startslot << PAGE_SHIFT); 1837 1838 /* 1839 * first, map the pages into the kernel. 1840 */ 1841 1842 mapinflags = !write ? 1843 UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ : 1844 UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE; 1845 kva = uvm_pagermapin(pps, npages, mapinflags); 1846 1847 /* 1848 * fill in the bp/sbp. we currently route our i/o through 1849 * /dev/drum's vnode [swapdev_vp]. 1850 */ 1851 1852 bp->b_cflags = BC_BUSY | BC_NOCACHE; 1853 bp->b_flags = (flags & (B_READ|B_ASYNC)); 1854 bp->b_proc = &proc0; /* XXX */ 1855 bp->b_vnbufs.le_next = NOLIST; 1856 bp->b_data = (void *)kva; 1857 bp->b_blkno = startblk; 1858 bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT; 1859 1860 /* 1861 * bump v_numoutput (counter of number of active outputs). 1862 */ 1863 1864 if (write) { 1865 mutex_enter(swapdev_vp->v_interlock); 1866 swapdev_vp->v_numoutput++; 1867 mutex_exit(swapdev_vp->v_interlock); 1868 } 1869 1870 /* 1871 * for async ops we must set up the iodone handler. 1872 */ 1873 1874 if (async) { 1875 bp->b_iodone = uvm_aio_biodone; 1876 UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0); 1877 if (curlwp == uvm.pagedaemon_lwp) 1878 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 1879 else 1880 BIO_SETPRIO(bp, BPRIO_TIMELIMITED); 1881 } else { 1882 bp->b_iodone = NULL; 1883 BIO_SETPRIO(bp, BPRIO_TIMECRITICAL); 1884 } 1885 UVMHIST_LOG(pdhist, 1886 "about to start io: data = %p blkno = 0x%x, bcount = %ld", 1887 bp->b_data, bp->b_blkno, bp->b_bcount, 0); 1888 1889 /* 1890 * now we start the I/O, and if async, return. 1891 */ 1892 1893 VOP_STRATEGY(swapdev_vp, bp); 1894 if (async) 1895 return 0; 1896 1897 /* 1898 * must be sync i/o. wait for it to finish 1899 */ 1900 1901 error = biowait(bp); 1902 1903 /* 1904 * kill the pager mapping 1905 */ 1906 1907 uvm_pagermapout(kva, npages); 1908 1909 /* 1910 * now dispose of the buf and we're done. 1911 */ 1912 1913 if (write) { 1914 mutex_enter(swapdev_vp->v_interlock); 1915 vwakeup(bp); 1916 mutex_exit(swapdev_vp->v_interlock); 1917 } 1918 putiobuf(bp); 1919 UVMHIST_LOG(pdhist, "<- done (sync) error=%d", error, 0, 0, 0); 1920 1921 return (error); 1922 } 1923