sys/vm/vm_pager.c

   1 /*
   2  * (MPSAFE)
   3  *
   4  * Copyright (c) 1991, 1993
   5  *      The Regents of the University of California.  All rights reserved.
   6  *
   7  * This code is derived from software contributed to Berkeley by
   8  * The Mach Operating System project at Carnegie-Mellon University.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  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 the
  17  *    documentation and/or other materials provided with the distribution.
  18  * 3. Neither the name of the University nor the names of its contributors
  19  *    may be used to endorse or promote products derived from this software
  20  *    without specific prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  *      from: @(#)vm_pager.c    8.6 (Berkeley) 1/12/94
  35  *
  36  *
  37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
  38  * All rights reserved.
  39  *
  40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
  41  *
  42  * Permission to use, copy, modify and distribute this software and
  43  * its documentation is hereby granted, provided that both the copyright
  44  * notice and this permission notice appear in all copies of the
  45  * software, derivative works or modified versions, and any portions
  46  * thereof, and that both notices appear in supporting documentation.
  47  *
  48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
  49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
  50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
  51  *
  52  * Carnegie Mellon requests users of this software to return to
  53  *
  54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
  55  *  School of Computer Science
  56  *  Carnegie Mellon University
  57  *  Pittsburgh PA 15213-3890
  58  *
  59  * any improvements or extensions that they make and grant Carnegie the
  60  * rights to redistribute these changes.
  61  *
  62  * $FreeBSD: src/sys/vm/vm_pager.c,v 1.54.2.2 2001/11/18 07:11:00 dillon Exp $
  63  */
  64
  65 /*
  66  *      Paging space routine stubs.  Emulates a matchmaker-like interface
  67  *      for builtin pagers.
  68  */
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/kernel.h>
  73 #include <sys/vnode.h>
  74 #include <sys/buf.h>
  75 #include <sys/ucred.h>
  76 #include <sys/dsched.h>
  77 #include <sys/proc.h>
  78 #include <sys/sysctl.h>
  79 #include <sys/thread2.h>
  80
  81 #include <vm/vm.h>
  82 #include <vm/vm_param.h>
  83 #include <vm/vm_kern.h>
  84 #include <vm/vm_object.h>
  85 #include <vm/vm_page.h>
  86 #include <vm/vm_pager.h>
  87 #include <vm/vm_extern.h>
  88
  89 #include <sys/buf2.h>
  90 #include <vm/vm_page2.h>
  91
  92 extern struct pagerops defaultpagerops;
  93 extern struct pagerops swappagerops;
  94 extern struct pagerops vnodepagerops;
  95 extern struct pagerops devicepagerops;
  96 extern struct pagerops physpagerops;
  97
  98 static int dead_pager_getpage (vm_object_t, vm_page_t *, int);
  99 static void dead_pager_putpages (vm_object_t, vm_page_t *, int, int, int *);
 100 static boolean_t dead_pager_haspage (vm_object_t, vm_pindex_t);
 101 static void dead_pager_dealloc (vm_object_t);
 102
 103 /*
 104  * No requirements.
 105  */
 106 static int
 107 dead_pager_getpage(vm_object_t obj, vm_page_t *mpp, int seqaccess)
 108 {
 109         return VM_PAGER_FAIL;
 110 }
 111
 112 /*
 113  * No requirements.
 114  */
 115 static void
 116 dead_pager_putpages(vm_object_t object, vm_page_t *m, int count, int flags,
 117                     int *rtvals)
 118 {
 119         int i;
 120
 121         for (i = 0; i < count; i++) {
 122                 rtvals[i] = VM_PAGER_AGAIN;
 123         }
 124 }
 125
 126 /*
 127  * No requirements.
 128  */
 129 static boolean_t
 130 dead_pager_haspage(vm_object_t object, vm_pindex_t pindex)
 131 {
 132         return FALSE;
 133 }
 134
 135 /*
 136  * No requirements.
 137  */
 138 static void
 139 dead_pager_dealloc(vm_object_t object)
 140 {
 141         KKASSERT(object->swblock_count == 0);
 142         return;
 143 }
 144
 145 static struct pagerops deadpagerops = {
 146         dead_pager_dealloc,
 147         dead_pager_getpage,
 148         dead_pager_putpages,
 149         dead_pager_haspage
 150 };
 151
 152 struct pagerops *pagertab[] = {
 153         &defaultpagerops,       /* OBJT_DEFAULT */
 154         &swappagerops,          /* OBJT_SWAP */
 155         &vnodepagerops,         /* OBJT_VNODE */
 156         &devicepagerops,        /* OBJT_DEVICE */
 157         &devicepagerops,        /* OBJT_MGTDEVICE */
 158         &physpagerops,          /* OBJT_PHYS */
 159         &deadpagerops           /* OBJT_DEAD */
 160 };
 161
 162 int npagers = NELEM(pagertab);
 163
 164 /*
 165  * Kernel address space for mapping pages.
 166  * Used by pagers where KVAs are needed for IO.
 167  *
 168  * XXX needs to be large enough to support the number of pending async
 169  * cleaning requests (NPENDINGIO == 64) * the maximum swap cluster size
 170  * (MAXPHYS == 64k) if you want to get the most efficiency.
 171  */
 172 #define PAGER_MAP_SIZE  (8 * 1024 * 1024)
 173
 174 #define BSWHSIZE        16
 175 #define BSWHMASK        (BSWHSIZE - 1)
 176
 177 TAILQ_HEAD(swqueue, buf);
 178
 179 int pager_map_size = PAGER_MAP_SIZE;
 180 struct vm_map pager_map;
 181
 182 static vm_offset_t swapbkva_mem;        /* swap buffers kva */
 183 static vm_offset_t swapbkva_kva;        /* swap buffers kva */
 184 static struct swqueue bswlist_mem[BSWHSIZE];    /* with preallocated memory */
 185 static struct swqueue bswlist_kva[BSWHSIZE];    /* with kva */
 186 static struct swqueue bswlist_raw[BSWHSIZE];    /* without kva */
 187 static struct spinlock bswspin_mem[BSWHSIZE];
 188 static struct spinlock bswspin_kva[BSWHSIZE];
 189 static struct spinlock bswspin_raw[BSWHSIZE];
 190 static int pbuf_raw_count;
 191 static int pbuf_kva_count;
 192 static int pbuf_mem_count;
 193
 194 SYSCTL_INT(_vfs, OID_AUTO, pbuf_raw_count, CTLFLAG_RD, &pbuf_raw_count, 0,
 195     "Kernel pbuf raw reservations");
 196 SYSCTL_INT(_vfs, OID_AUTO, pbuf_kva_count, CTLFLAG_RD, &pbuf_kva_count, 0,
 197     "Kernel pbuf kva reservations");
 198 SYSCTL_INT(_vfs, OID_AUTO, pbuf_mem_count, CTLFLAG_RD, &pbuf_mem_count, 0,
 199     "Kernel pbuf mem reservations");
 200
 201 /*
 202  * Initialize the swap buffer list.
 203  *
 204  * Called from the low level boot code only.
 205  */
 206 static void
 207 vm_pager_init(void *arg __unused)
 208 {
 209         int i;
 210
 211         for (i = 0; i < BSWHSIZE; ++i) {
 212                 TAILQ_INIT(&bswlist_mem[i]);
 213                 TAILQ_INIT(&bswlist_kva[i]);
 214                 TAILQ_INIT(&bswlist_raw[i]);
 215                 spin_init(&bswspin_mem[i], "bswmem");
 216                 spin_init(&bswspin_kva[i], "bswkva");
 217                 spin_init(&bswspin_raw[i], "bswraw");
 218         }
 219 }
 220 SYSINIT(vm_mem, SI_BOOT1_VM, SI_ORDER_SECOND, vm_pager_init, NULL);
 221
 222 /*
 223  * Called from the low level boot code only.
 224  */
 225 static
 226 void
 227 vm_pager_bufferinit(void *dummy __unused)
 228 {
 229         struct buf *bp;
 230         long i;
 231
 232         /*
 233          * Reserve KVM space for pbuf data.
 234          */
 235         swapbkva_mem = kmem_alloc_pageable(&pager_map, nswbuf_mem * MAXPHYS,
 236                                            VM_SUBSYS_BUFDATA);
 237         if (!swapbkva_mem)
 238                 panic("Not enough pager_map VM space for physical buffers");
 239         swapbkva_kva = kmem_alloc_pageable(&pager_map, nswbuf_kva * MAXPHYS,
 240                                            VM_SUBSYS_BUFDATA);
 241         if (!swapbkva_kva)
 242                 panic("Not enough pager_map VM space for physical buffers");
 243
 244         /*
 245          * Initial pbuf setup.
 246          *
 247          * mem - These pbufs have permanently allocated memory
 248          * kva - These pbufs have unallocated kva reservations
 249          * raw - These pbufs have no kva reservations
 250          */
 251
 252         /*
 253          * Buffers with pre-allocated kernel memory can be convenient for
 254          * copyin/copyout because no SMP page invalidation or other pmap
 255          * operations are needed.
 256          */
 257         bp = swbuf_mem;
 258         for (i = 0; i < nswbuf_mem; ++i, ++bp) {
 259                 vm_page_t m;
 260                 vm_pindex_t pg;
 261                 int j;
 262
 263                 bp->b_kvabase = (caddr_t)((intptr_t)i * MAXPHYS) + swapbkva_mem;
 264                 bp->b_kvasize = MAXPHYS;
 265                 bp->b_swindex = i & BSWHMASK;
 266                 bp->b_cpumask = smp_active_mask;
 267                 BUF_LOCKINIT(bp);
 268                 buf_dep_init(bp);
 269                 TAILQ_INSERT_HEAD(&bswlist_mem[i & BSWHMASK], bp, b_freelist);
 270                 atomic_add_int(&pbuf_mem_count, 1);
 271                 bp->b_data = bp->b_kvabase;
 272                 bp->b_bcount = MAXPHYS;
 273                 bp->b_xio.xio_pages = bp->b_xio.xio_internal_pages;
 274
 275                 pg = (vm_offset_t)bp->b_kvabase >> PAGE_SHIFT;
 276                 vm_object_hold(&kernel_object);
 277                 for (j = 0; j < MAXPHYS / PAGE_SIZE; ++j) {
 278                         m = vm_page_alloc(&kernel_object, pg, VM_ALLOC_NORMAL |
 279                                                               VM_ALLOC_SYSTEM);
 280                         KKASSERT(m != NULL);
 281                         bp->b_xio.xio_internal_pages[j] = m;
 282                         vm_page_wire(m);
 283                         /* early boot, no other cpus running yet */
 284                         pmap_kenter_noinval(pg * PAGE_SIZE, VM_PAGE_TO_PHYS(m));
 285                         cpu_invlpg((void *)(pg * PAGE_SIZE));
 286                         vm_page_wakeup(m);
 287                         ++pg;
 288                 }
 289                 vm_object_drop(&kernel_object);
 290                 bp->b_xio.xio_npages = j;
 291         }
 292
 293         /*
 294          * Buffers with pre-assigned KVA bases.  The KVA has no memory pages
 295          * assigned to it.  Saves the caller from having to reserve KVA for
 296          * the page map.
 297          */
 298         bp = swbuf_kva;
 299         for (i = 0; i < nswbuf_kva; ++i, ++bp) {
 300                 bp->b_kvabase = (caddr_t)((intptr_t)i * MAXPHYS) + swapbkva_kva;
 301                 bp->b_kvasize = MAXPHYS;
 302                 bp->b_swindex = i & BSWHMASK;
 303                 BUF_LOCKINIT(bp);
 304                 buf_dep_init(bp);
 305                 TAILQ_INSERT_HEAD(&bswlist_kva[i & BSWHMASK], bp, b_freelist);
 306                 atomic_add_int(&pbuf_kva_count, 1);
 307         }
 308
 309         /*
 310          * RAW buffers with no KVA mappings.
 311          *
 312          * NOTE: We use KM_NOTLBSYNC here to reduce unnecessary IPIs
 313          *       during startup, which can really slow down emulated
 314          *       systems.
 315          */
 316         nswbuf_raw = nbuf * 2;
 317         swbuf_raw = (void *)kmem_alloc3(&kernel_map,
 318                                 round_page(nswbuf_raw * sizeof(struct buf)),
 319                                 VM_SUBSYS_BUFDATA,
 320                                 KM_NOTLBSYNC);
 321         smp_invltlb();
 322         bp = swbuf_raw;
 323         for (i = 0; i < nswbuf_raw; ++i, ++bp) {
 324                 bp->b_swindex = i & BSWHMASK;
 325                 BUF_LOCKINIT(bp);
 326                 buf_dep_init(bp);
 327                 TAILQ_INSERT_HEAD(&bswlist_raw[i & BSWHMASK], bp, b_freelist);
 328                 atomic_add_int(&pbuf_raw_count, 1);
 329         }
 330 }
 331
 332 SYSINIT(do_vmpg, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, vm_pager_bufferinit, NULL);
 333
 334 /*
 335  * No requirements.
 336  */
 337 void
 338 vm_pager_deallocate(vm_object_t object)
 339 {
 340         (*pagertab[object->type]->pgo_dealloc) (object);
 341 }
 342
 343 /*
 344  * vm_pager_get_pages() - inline, see vm/vm_pager.h
 345  * vm_pager_put_pages() - inline, see vm/vm_pager.h
 346  * vm_pager_has_page() - inline, see vm/vm_pager.h
 347  * vm_pager_page_inserted() - inline, see vm/vm_pager.h
 348  * vm_pager_page_removed() - inline, see vm/vm_pager.h
 349  */
 350
 351 /*
 352  * Search the specified pager object list for an object with the
 353  * specified handle.  If an object with the specified handle is found,
 354  * increase its reference count and return it.  Otherwise, return NULL.
 355  *
 356  * The pager object list must be locked.
 357  */
 358 vm_object_t
 359 vm_pager_object_lookup(struct pagerlst *pg_list, void *handle)
 360 {
 361         vm_object_t object;
 362
 363         TAILQ_FOREACH(object, pg_list, pager_object_list) {
 364                 if (object->handle == handle) {
 365                         VM_OBJECT_LOCK(object);
 366                         if ((object->flags & OBJ_DEAD) == 0) {
 367                                 vm_object_reference_locked(object);
 368                                 VM_OBJECT_UNLOCK(object);
 369                                 break;
 370                         }
 371                         VM_OBJECT_UNLOCK(object);
 372                 }
 373         }
 374         return (object);
 375 }
 376
 377 /*
 378  * Initialize a physical buffer.
 379  *
 380  * No requirements.
 381  */
 382 static void
 383 initpbuf(struct buf *bp)
 384 {
 385         bp->b_qindex = 0;               /* BQUEUE_NONE */
 386         bp->b_data = bp->b_kvabase;     /* NULL if pbuf sans kva */
 387         bp->b_flags = B_PAGING;
 388         bp->b_cmd = BUF_CMD_DONE;
 389         bp->b_error = 0;
 390         bp->b_bcount = 0;
 391         bp->b_bufsize = MAXPHYS;
 392         initbufbio(bp);
 393         xio_init(&bp->b_xio);
 394         BUF_LOCK(bp, LK_EXCLUSIVE);
 395 }
 396
 397 /*
 398  * Allocate a physical buffer
 399  *
 400  * If (pfreecnt != NULL) then *pfreecnt will be decremented on return and
 401  * the function will block while it is <= 0.
 402  *
 403  * Physical buffers can be with or without KVA space reserved.  There
 404  * are severe limitations on the ones with KVA reserved, and fewer
 405  * limitations on the ones without.  getpbuf() gets one without,
 406  * getpbuf_kva() gets one with.
 407  *
 408  * No requirements.
 409  */
 410 struct buf *
 411 getpbuf(int *pfreecnt)
 412 {
 413         struct buf *bp;
 414         int iter;
 415         int loops;
 416
 417         for (;;) {
 418                 while (pfreecnt && *pfreecnt <= 0) {
 419                         tsleep_interlock(pfreecnt, 0);
 420                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 421                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf0", 0);
 422                 }
 423                 if (pbuf_raw_count <= 0) {
 424                         tsleep_interlock(&pbuf_raw_count, 0);
 425                         if ((int)atomic_fetchadd_int(&pbuf_raw_count, 0) <= 0)
 426                                 tsleep(&pbuf_raw_count, PINTERLOCKED,
 427                                        "wswbuf1", 0);
 428                         continue;
 429                 }
 430                 iter = mycpuid & BSWHMASK;
 431                 for (loops = BSWHSIZE; loops; --loops) {
 432                         if (TAILQ_FIRST(&bswlist_raw[iter]) == NULL) {
 433                                 iter = (iter + 1) & BSWHMASK;
 434                                 continue;
 435                         }
 436                         spin_lock(&bswspin_raw[iter]);
 437                         if ((bp = TAILQ_FIRST(&bswlist_raw[iter])) == NULL) {
 438                                 spin_unlock(&bswspin_raw[iter]);
 439                                 iter = (iter + 1) & BSWHMASK;
 440                                 continue;
 441                         }
 442                         TAILQ_REMOVE(&bswlist_raw[iter], bp, b_freelist);
 443                         atomic_add_int(&pbuf_raw_count, -1);
 444                         if (pfreecnt)
 445                                 atomic_add_int(pfreecnt, -1);
 446                         spin_unlock(&bswspin_raw[iter]);
 447                         initpbuf(bp);
 448
 449                         return bp;
 450                 }
 451         }
 452         /* not reached */
 453 }
 454
 455 struct buf *
 456 getpbuf_kva(int *pfreecnt)
 457 {
 458         struct buf *bp;
 459         int iter;
 460         int loops;
 461
 462         for (;;) {
 463                 while (pfreecnt && *pfreecnt <= 0) {
 464                         tsleep_interlock(pfreecnt, 0);
 465                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 466                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf2", 0);
 467                 }
 468                 if (pbuf_kva_count <= 0) {
 469                         tsleep_interlock(&pbuf_kva_count, 0);
 470                         if ((int)atomic_fetchadd_int(&pbuf_kva_count, 0) <= 0)
 471                                 tsleep(&pbuf_kva_count, PINTERLOCKED,
 472                                        "wswbuf3", 0);
 473                         continue;
 474                 }
 475                 iter = mycpuid & BSWHMASK;
 476                 for (loops = BSWHSIZE; loops; --loops) {
 477                         if (TAILQ_FIRST(&bswlist_kva[iter]) == NULL) {
 478                                 iter = (iter + 1) & BSWHMASK;
 479                                 continue;
 480                         }
 481                         spin_lock(&bswspin_kva[iter]);
 482                         if ((bp = TAILQ_FIRST(&bswlist_kva[iter])) == NULL) {
 483                                 spin_unlock(&bswspin_kva[iter]);
 484                                 iter = (iter + 1) & BSWHMASK;
 485                                 continue;
 486                         }
 487                         TAILQ_REMOVE(&bswlist_kva[iter], bp, b_freelist);
 488                         atomic_add_int(&pbuf_kva_count, -1);
 489                         if (pfreecnt)
 490                                 atomic_add_int(pfreecnt, -1);
 491                         spin_unlock(&bswspin_kva[iter]);
 492                         initpbuf(bp);
 493
 494                         return bp;
 495                 }
 496         }
 497         /* not reached */
 498 }
 499
 500 /*
 501  * Allocate a pbuf with kernel memory already preallocated.  Caller must
 502  * not change the mapping.
 503  */
 504 struct buf *
 505 getpbuf_mem(int *pfreecnt)
 506 {
 507         struct buf *bp;
 508         int iter;
 509         int loops;
 510
 511         for (;;) {
 512                 while (pfreecnt && *pfreecnt <= 0) {
 513                         tsleep_interlock(pfreecnt, 0);
 514                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 515                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf4", 0);
 516                 }
 517                 if (pbuf_mem_count <= 0) {
 518                         tsleep_interlock(&pbuf_mem_count, 0);
 519                         if ((int)atomic_fetchadd_int(&pbuf_mem_count, 0) <= 0)
 520                                 tsleep(&pbuf_mem_count, PINTERLOCKED,
 521                                        "wswbuf5", 0);
 522                         continue;
 523                 }
 524                 iter = mycpuid & BSWHMASK;
 525                 for (loops = BSWHSIZE; loops; --loops) {
 526                         if (TAILQ_FIRST(&bswlist_mem[iter]) == NULL) {
 527                                 iter = (iter + 1) & BSWHMASK;
 528                                 continue;
 529                         }
 530                         spin_lock(&bswspin_mem[iter]);
 531                         if ((bp = TAILQ_FIRST(&bswlist_mem[iter])) == NULL) {
 532                                 spin_unlock(&bswspin_mem[iter]);
 533                                 iter = (iter + 1) & BSWHMASK;
 534                                 continue;
 535                         }
 536                         TAILQ_REMOVE(&bswlist_mem[iter], bp, b_freelist);
 537                         atomic_add_int(&pbuf_mem_count, -1);
 538                         if (pfreecnt)
 539                                 atomic_add_int(pfreecnt, -1);
 540                         spin_unlock(&bswspin_mem[iter]);
 541                         initpbuf(bp);
 542
 543                         return bp;
 544                 }
 545         }
 546         /* not reached */
 547 }
 548
 549 /*
 550  * Allocate a physical buffer, if one is available.
 551  *
 552  * Note that there is no NULL hack here - all subsystems using this
 553  * call are required to use a non-NULL pfreecnt.
 554  *
 555  * No requirements.
 556  */
 557 struct buf *
 558 trypbuf(int *pfreecnt)
 559 {
 560         struct buf *bp;
 561         int iter = mycpuid & BSWHMASK;
 562         int loops;
 563
 564         for (loops = BSWHSIZE; loops; --loops) {
 565                 if (*pfreecnt <= 0 || TAILQ_FIRST(&bswlist_raw[iter]) == NULL) {
 566                         iter = (iter + 1) & BSWHMASK;
 567                         continue;
 568                 }
 569                 spin_lock(&bswspin_raw[iter]);
 570                 if (*pfreecnt <= 0 ||
 571                     (bp = TAILQ_FIRST(&bswlist_raw[iter])) == NULL) {
 572                         spin_unlock(&bswspin_raw[iter]);
 573                         iter = (iter + 1) & BSWHMASK;
 574                         continue;
 575                 }
 576                 TAILQ_REMOVE(&bswlist_raw[iter], bp, b_freelist);
 577                 atomic_add_int(&pbuf_raw_count, -1);
 578                 atomic_add_int(pfreecnt, -1);
 579
 580                 spin_unlock(&bswspin_raw[iter]);
 581
 582                 initpbuf(bp);
 583
 584                 return bp;
 585         }
 586         return NULL;
 587 }
 588
 589 struct buf *
 590 trypbuf_kva(int *pfreecnt)
 591 {
 592         struct buf *bp;
 593         int iter = mycpuid & BSWHMASK;
 594         int loops;
 595
 596         for (loops = BSWHSIZE; loops; --loops) {
 597                 if (*pfreecnt <= 0 || TAILQ_FIRST(&bswlist_kva[iter]) == NULL) {
 598                         iter = (iter + 1) & BSWHMASK;
 599                         continue;
 600                 }
 601                 spin_lock(&bswspin_kva[iter]);
 602                 if (*pfreecnt <= 0 ||
 603                     (bp = TAILQ_FIRST(&bswlist_kva[iter])) == NULL) {
 604                         spin_unlock(&bswspin_kva[iter]);
 605                         iter = (iter + 1) & BSWHMASK;
 606                         continue;
 607                 }
 608                 TAILQ_REMOVE(&bswlist_kva[iter], bp, b_freelist);
 609                 atomic_add_int(&pbuf_kva_count, -1);
 610                 atomic_add_int(pfreecnt, -1);
 611
 612                 spin_unlock(&bswspin_kva[iter]);
 613
 614                 initpbuf(bp);
 615
 616                 return bp;
 617         }
 618         return NULL;
 619 }
 620
 621 /*
 622  * Release a physical buffer
 623  *
 624  *      NOTE: pfreecnt can be NULL, but this 'feature' will be removed
 625  *      relatively soon when the rest of the subsystems get smart about it. XXX
 626  *
 627  * No requirements.
 628  */
 629 void
 630 relpbuf(struct buf *bp, int *pfreecnt)
 631 {
 632         int wake = 0;
 633         int wake_free = 0;
 634         int iter = bp->b_swindex;
 635
 636         KKASSERT(bp->b_flags & B_PAGING);
 637         dsched_buf_exit(bp);
 638
 639         BUF_UNLOCK(bp);
 640
 641         if (bp >= swbuf_mem && bp < &swbuf_mem[nswbuf_mem]) {
 642                 KKASSERT(bp->b_kvabase);
 643                 spin_lock(&bswspin_mem[iter]);
 644                 TAILQ_INSERT_HEAD(&bswlist_mem[iter], bp, b_freelist);
 645                 if (atomic_fetchadd_int(&pbuf_mem_count, 1) == nswbuf_mem / 4)
 646                         wake = 1;
 647                 if (pfreecnt) {
 648                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 649                                 wake_free = 1;
 650                 }
 651                 spin_unlock(&bswspin_mem[iter]);
 652                 if (wake)
 653                         wakeup(&pbuf_mem_count);
 654         } else if (bp >= swbuf_kva && bp < &swbuf_kva[nswbuf_kva]) {
 655                 KKASSERT(bp->b_kvabase);
 656                 CPUMASK_ASSZERO(bp->b_cpumask);
 657                 spin_lock(&bswspin_kva[iter]);
 658                 TAILQ_INSERT_HEAD(&bswlist_kva[iter], bp, b_freelist);
 659                 if (atomic_fetchadd_int(&pbuf_kva_count, 1) == nswbuf_kva / 4)
 660                         wake = 1;
 661                 if (pfreecnt) {
 662                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 663                                 wake_free = 1;
 664                 }
 665                 spin_unlock(&bswspin_kva[iter]);
 666                 if (wake)
 667                         wakeup(&pbuf_kva_count);
 668         } else {
 669                 KKASSERT(bp->b_kvabase == NULL);
 670                 KKASSERT(bp >= swbuf_raw && bp < &swbuf_raw[nswbuf_raw]);
 671                 CPUMASK_ASSZERO(bp->b_cpumask);
 672                 spin_lock(&bswspin_raw[iter]);
 673                 TAILQ_INSERT_HEAD(&bswlist_raw[iter], bp, b_freelist);
 674                 if (atomic_fetchadd_int(&pbuf_raw_count, 1) == nswbuf_raw / 4)
 675                         wake = 1;
 676                 if (pfreecnt) {
 677                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 678                                 wake_free = 1;
 679                 }
 680                 spin_unlock(&bswspin_raw[iter]);
 681                 if (wake)
 682                         wakeup(&pbuf_raw_count);
 683         }
 684         if (wake_free)
 685                 wakeup(pfreecnt);
 686 }
 687
 688 void
 689 pbuf_adjcount(int *pfreecnt, int n)
 690 {
 691         if (n) {
 692                 atomic_add_int(pfreecnt, n);
 693                 wakeup(pfreecnt);
 694         }
 695 }