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         if (!swapbkva_mem)
 237                 panic("Not enough pager_map VM space for physical buffers");
 238         swapbkva_kva = kmem_alloc_pageable(&pager_map, nswbuf_kva * MAXPHYS);
 239         if (!swapbkva_kva)
 240                 panic("Not enough pager_map VM space for physical buffers");
 241
 242         /*
 243          * Initial pbuf setup.
 244          *
 245          * mem - These pbufs have permanently allocated memory
 246          * kva - These pbufs have unallocated kva reservations
 247          * raw - These pbufs have no kva reservations
 248          */
 249
 250         /*
 251          * Buffers with pre-allocated kernel memory can be convenient for
 252          * copyin/copyout because no SMP page invalidation or other pmap
 253          * operations are needed.
 254          */
 255 #if 1
 256         bp = swbuf_mem;
 257         for (i = 0; i < nswbuf_mem; ++i, ++bp) {
 258                 vm_page_t m;
 259                 vm_pindex_t pg;
 260                 int j;
 261
 262                 bp->b_kvabase = (caddr_t)((intptr_t)i * MAXPHYS) + swapbkva_mem;
 263                 bp->b_kvasize = MAXPHYS;
 264                 bp->b_swindex = i & BSWHMASK;
 265                 BUF_LOCKINIT(bp);
 266                 buf_dep_init(bp);
 267                 TAILQ_INSERT_HEAD(&bswlist_mem[i & BSWHMASK], bp, b_freelist);
 268                 atomic_add_int(&pbuf_mem_count, 1);
 269                 bp->b_data = bp->b_kvabase;
 270                 bp->b_bcount = MAXPHYS;
 271                 bp->b_xio.xio_pages = bp->b_xio.xio_internal_pages;
 272
 273                 pg = (vm_offset_t)bp->b_kvabase >> PAGE_SHIFT;
 274                 vm_object_hold(&kernel_object);
 275                 for (j = 0; j < MAXPHYS / PAGE_SIZE; ++j) {
 276                         m = vm_page_alloc(&kernel_object, pg, VM_ALLOC_NORMAL |
 277                                                               VM_ALLOC_SYSTEM);
 278                         KKASSERT(m != NULL);
 279                         bp->b_xio.xio_internal_pages[j] = m;
 280                         vm_page_wire(m);
 281                         /* early boot, no other cpus running yet */
 282                         pmap_kenter_noinval(pg * PAGE_SIZE, VM_PAGE_TO_PHYS(m));
 283                         cpu_invlpg((void *)(pg * PAGE_SIZE));
 284                         vm_page_wakeup(m);
 285                         ++pg;
 286                 }
 287                 vm_object_drop(&kernel_object);
 288                 bp->b_xio.xio_npages = j;
 289         }
 290 #endif
 291
 292         /*
 293          * Buffers with pre-assigned KVA bases.  The KVA has no memory pages
 294          * assigned to it.  Saves the caller from having to reserve KVA for
 295          * the page map.
 296          */
 297         bp = swbuf_kva;
 298         for (i = 0; i < nswbuf_kva; ++i, ++bp) {
 299                 bp->b_kvabase = (caddr_t)((intptr_t)i * MAXPHYS) + swapbkva_kva;
 300                 bp->b_kvasize = MAXPHYS;
 301                 bp->b_swindex = i & BSWHMASK;
 302                 BUF_LOCKINIT(bp);
 303                 buf_dep_init(bp);
 304                 TAILQ_INSERT_HEAD(&bswlist_kva[i & BSWHMASK], bp, b_freelist);
 305                 atomic_add_int(&pbuf_kva_count, 1);
 306         }
 307
 308         /*
 309          * RAW buffers with no KVA mappings.
 310          *
 311          * NOTE: We use KM_NOTLBSYNC here to reduce unnecessary IPIs
 312          *       during startup, which can really slow down emulated
 313          *       systems.
 314          */
 315         nswbuf_raw = nbuf * 2;
 316         swbuf_raw = (void *)kmem_alloc3(&kernel_map,
 317                                 round_page(nswbuf_raw * sizeof(struct buf)),
 318                                 KM_NOTLBSYNC);
 319         smp_invltlb();
 320         bp = swbuf_raw;
 321         for (i = 0; i < nswbuf_raw; ++i, ++bp) {
 322                 bp->b_swindex = i & BSWHMASK;
 323                 BUF_LOCKINIT(bp);
 324                 buf_dep_init(bp);
 325                 TAILQ_INSERT_HEAD(&bswlist_raw[i & BSWHMASK], bp, b_freelist);
 326                 atomic_add_int(&pbuf_raw_count, 1);
 327         }
 328 }
 329
 330 SYSINIT(do_vmpg, SI_BOOT2_MACHDEP, SI_ORDER_FIRST, vm_pager_bufferinit, NULL);
 331
 332 /*
 333  * No requirements.
 334  */
 335 void
 336 vm_pager_deallocate(vm_object_t object)
 337 {
 338         (*pagertab[object->type]->pgo_dealloc) (object);
 339 }
 340
 341 /*
 342  * vm_pager_get_pages() - inline, see vm/vm_pager.h
 343  * vm_pager_put_pages() - inline, see vm/vm_pager.h
 344  * vm_pager_has_page() - inline, see vm/vm_pager.h
 345  * vm_pager_page_inserted() - inline, see vm/vm_pager.h
 346  * vm_pager_page_removed() - inline, see vm/vm_pager.h
 347  */
 348
 349 /*
 350  * Search the specified pager object list for an object with the
 351  * specified handle.  If an object with the specified handle is found,
 352  * increase its reference count and return it.  Otherwise, return NULL.
 353  *
 354  * The pager object list must be locked.
 355  */
 356 vm_object_t
 357 vm_pager_object_lookup(struct pagerlst *pg_list, void *handle)
 358 {
 359         vm_object_t object;
 360
 361         TAILQ_FOREACH(object, pg_list, pager_object_list) {
 362                 if (object->handle == handle) {
 363                         VM_OBJECT_LOCK(object);
 364                         if ((object->flags & OBJ_DEAD) == 0) {
 365                                 vm_object_reference_locked(object);
 366                                 VM_OBJECT_UNLOCK(object);
 367                                 break;
 368                         }
 369                         VM_OBJECT_UNLOCK(object);
 370                 }
 371         }
 372         return (object);
 373 }
 374
 375 /*
 376  * Initialize a physical buffer.
 377  *
 378  * No requirements.
 379  */
 380 static void
 381 initpbuf(struct buf *bp)
 382 {
 383         bp->b_qindex = 0;               /* BQUEUE_NONE */
 384         bp->b_data = bp->b_kvabase;     /* NULL if pbuf sans kva */
 385         bp->b_flags = B_PAGING;
 386         bp->b_cmd = BUF_CMD_DONE;
 387         bp->b_error = 0;
 388         bp->b_bcount = 0;
 389         bp->b_bufsize = MAXPHYS;
 390         initbufbio(bp);
 391         xio_init(&bp->b_xio);
 392         BUF_LOCK(bp, LK_EXCLUSIVE);
 393 }
 394
 395 /*
 396  * Allocate a physical buffer
 397  *
 398  * If (pfreecnt != NULL) then *pfreecnt will be decremented on return and
 399  * the function will block while it is <= 0.
 400  *
 401  * Physical buffers can be with or without KVA space reserved.  There
 402  * are severe limitations on the ones with KVA reserved, and fewer
 403  * limitations on the ones without.  getpbuf() gets one without,
 404  * getpbuf_kva() gets one with.
 405  *
 406  * No requirements.
 407  */
 408 struct buf *
 409 getpbuf(int *pfreecnt)
 410 {
 411         struct buf *bp;
 412         int iter;
 413         int loops;
 414
 415         for (;;) {
 416                 while (pfreecnt && *pfreecnt <= 0) {
 417                         tsleep_interlock(pfreecnt, 0);
 418                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 419                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf0", 0);
 420                 }
 421                 if (pbuf_raw_count <= 0) {
 422                         tsleep_interlock(&pbuf_raw_count, 0);
 423                         if ((int)atomic_fetchadd_int(&pbuf_raw_count, 0) <= 0)
 424                                 tsleep(&pbuf_raw_count, PINTERLOCKED,
 425                                        "wswbuf1", 0);
 426                         continue;
 427                 }
 428                 iter = mycpuid & BSWHMASK;
 429                 for (loops = BSWHSIZE; loops; --loops) {
 430                         if (TAILQ_FIRST(&bswlist_raw[iter]) == NULL) {
 431                                 iter = (iter + 1) & BSWHMASK;
 432                                 continue;
 433                         }
 434                         spin_lock(&bswspin_raw[iter]);
 435                         if ((bp = TAILQ_FIRST(&bswlist_raw[iter])) == NULL) {
 436                                 spin_unlock(&bswspin_raw[iter]);
 437                                 iter = (iter + 1) & BSWHMASK;
 438                                 continue;
 439                         }
 440                         TAILQ_REMOVE(&bswlist_raw[iter], bp, b_freelist);
 441                         atomic_add_int(&pbuf_raw_count, -1);
 442                         if (pfreecnt)
 443                                 atomic_add_int(pfreecnt, -1);
 444                         spin_unlock(&bswspin_raw[iter]);
 445                         initpbuf(bp);
 446
 447                         return bp;
 448                 }
 449         }
 450         /* not reached */
 451 }
 452
 453 struct buf *
 454 getpbuf_kva(int *pfreecnt)
 455 {
 456         struct buf *bp;
 457         int iter;
 458         int loops;
 459
 460         for (;;) {
 461                 while (pfreecnt && *pfreecnt <= 0) {
 462                         tsleep_interlock(pfreecnt, 0);
 463                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 464                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf2", 0);
 465                 }
 466                 if (pbuf_kva_count <= 0) {
 467                         tsleep_interlock(&pbuf_kva_count, 0);
 468                         if ((int)atomic_fetchadd_int(&pbuf_kva_count, 0) <= 0)
 469                                 tsleep(&pbuf_kva_count, PINTERLOCKED,
 470                                        "wswbuf3", 0);
 471                         continue;
 472                 }
 473                 iter = mycpuid & BSWHMASK;
 474                 for (loops = BSWHSIZE; loops; --loops) {
 475                         if (TAILQ_FIRST(&bswlist_kva[iter]) == NULL) {
 476                                 iter = (iter + 1) & BSWHMASK;
 477                                 continue;
 478                         }
 479                         spin_lock(&bswspin_kva[iter]);
 480                         if ((bp = TAILQ_FIRST(&bswlist_kva[iter])) == NULL) {
 481                                 spin_unlock(&bswspin_kva[iter]);
 482                                 iter = (iter + 1) & BSWHMASK;
 483                                 continue;
 484                         }
 485                         TAILQ_REMOVE(&bswlist_kva[iter], bp, b_freelist);
 486                         atomic_add_int(&pbuf_kva_count, -1);
 487                         if (pfreecnt)
 488                                 atomic_add_int(pfreecnt, -1);
 489                         spin_unlock(&bswspin_kva[iter]);
 490                         initpbuf(bp);
 491
 492                         return bp;
 493                 }
 494         }
 495         /* not reached */
 496 }
 497
 498 /*
 499  * Allocate a pbuf with kernel memory already preallocated.  Caller must
 500  * not change the mapping.
 501  */
 502 struct buf *
 503 getpbuf_mem(int *pfreecnt)
 504 {
 505         struct buf *bp;
 506         int iter;
 507         int loops;
 508
 509         for (;;) {
 510                 while (pfreecnt && *pfreecnt <= 0) {
 511                         tsleep_interlock(pfreecnt, 0);
 512                         if ((int)atomic_fetchadd_int(pfreecnt, 0) <= 0)
 513                                 tsleep(pfreecnt, PINTERLOCKED, "wswbuf4", 0);
 514                 }
 515                 if (pbuf_mem_count <= 0) {
 516                         tsleep_interlock(&pbuf_mem_count, 0);
 517                         if ((int)atomic_fetchadd_int(&pbuf_mem_count, 0) <= 0)
 518                                 tsleep(&pbuf_mem_count, PINTERLOCKED,
 519                                        "wswbuf5", 0);
 520                         continue;
 521                 }
 522                 iter = mycpuid & BSWHMASK;
 523                 for (loops = BSWHSIZE; loops; --loops) {
 524                         if (TAILQ_FIRST(&bswlist_mem[iter]) == NULL) {
 525                                 iter = (iter + 1) & BSWHMASK;
 526                                 continue;
 527                         }
 528                         spin_lock(&bswspin_mem[iter]);
 529                         if ((bp = TAILQ_FIRST(&bswlist_mem[iter])) == NULL) {
 530                                 spin_unlock(&bswspin_mem[iter]);
 531                                 iter = (iter + 1) & BSWHMASK;
 532                                 continue;
 533                         }
 534                         TAILQ_REMOVE(&bswlist_mem[iter], bp, b_freelist);
 535                         atomic_add_int(&pbuf_mem_count, -1);
 536                         if (pfreecnt)
 537                                 atomic_add_int(pfreecnt, -1);
 538                         spin_unlock(&bswspin_mem[iter]);
 539                         initpbuf(bp);
 540
 541                         return bp;
 542                 }
 543         }
 544         /* not reached */
 545 }
 546
 547 /*
 548  * Allocate a physical buffer, if one is available.
 549  *
 550  * Note that there is no NULL hack here - all subsystems using this
 551  * call are required to use a non-NULL pfreecnt.
 552  *
 553  * No requirements.
 554  */
 555 struct buf *
 556 trypbuf(int *pfreecnt)
 557 {
 558         struct buf *bp;
 559         int iter = mycpuid & BSWHMASK;
 560         int loops;
 561
 562         for (loops = BSWHSIZE; loops; --loops) {
 563                 if (*pfreecnt <= 0 || TAILQ_FIRST(&bswlist_raw[iter]) == NULL) {
 564                         iter = (iter + 1) & BSWHMASK;
 565                         continue;
 566                 }
 567                 spin_lock(&bswspin_raw[iter]);
 568                 if (*pfreecnt <= 0 ||
 569                     (bp = TAILQ_FIRST(&bswlist_raw[iter])) == NULL) {
 570                         spin_unlock(&bswspin_raw[iter]);
 571                         iter = (iter + 1) & BSWHMASK;
 572                         continue;
 573                 }
 574                 TAILQ_REMOVE(&bswlist_raw[iter], bp, b_freelist);
 575                 atomic_add_int(&pbuf_raw_count, -1);
 576                 atomic_add_int(pfreecnt, -1);
 577
 578                 spin_unlock(&bswspin_raw[iter]);
 579
 580                 initpbuf(bp);
 581
 582                 return bp;
 583         }
 584         return NULL;
 585 }
 586
 587 struct buf *
 588 trypbuf_kva(int *pfreecnt)
 589 {
 590         struct buf *bp;
 591         int iter = mycpuid & BSWHMASK;
 592         int loops;
 593
 594         for (loops = BSWHSIZE; loops; --loops) {
 595                 if (*pfreecnt <= 0 || TAILQ_FIRST(&bswlist_kva[iter]) == NULL) {
 596                         iter = (iter + 1) & BSWHMASK;
 597                         continue;
 598                 }
 599                 spin_lock(&bswspin_kva[iter]);
 600                 if (*pfreecnt <= 0 ||
 601                     (bp = TAILQ_FIRST(&bswlist_kva[iter])) == NULL) {
 602                         spin_unlock(&bswspin_kva[iter]);
 603                         iter = (iter + 1) & BSWHMASK;
 604                         continue;
 605                 }
 606                 TAILQ_REMOVE(&bswlist_kva[iter], bp, b_freelist);
 607                 atomic_add_int(&pbuf_kva_count, -1);
 608                 atomic_add_int(pfreecnt, -1);
 609
 610                 spin_unlock(&bswspin_kva[iter]);
 611
 612                 initpbuf(bp);
 613
 614                 return bp;
 615         }
 616         return NULL;
 617 }
 618
 619 /*
 620  * Release a physical buffer
 621  *
 622  *      NOTE: pfreecnt can be NULL, but this 'feature' will be removed
 623  *      relatively soon when the rest of the subsystems get smart about it. XXX
 624  *
 625  * No requirements.
 626  */
 627 void
 628 relpbuf(struct buf *bp, int *pfreecnt)
 629 {
 630         int wake = 0;
 631         int wake_free = 0;
 632         int iter = bp->b_swindex;
 633
 634         KKASSERT(bp->b_flags & B_PAGING);
 635         dsched_buf_exit(bp);
 636
 637         BUF_UNLOCK(bp);
 638
 639         if (bp >= swbuf_mem && bp < &swbuf_mem[nswbuf_mem]) {
 640                 KKASSERT(bp->b_kvabase);
 641                 spin_lock(&bswspin_mem[iter]);
 642                 TAILQ_INSERT_HEAD(&bswlist_mem[iter], bp, b_freelist);
 643                 if (atomic_fetchadd_int(&pbuf_mem_count, 1) == nswbuf_mem / 4)
 644                         wake = 1;
 645                 if (pfreecnt) {
 646                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 647                                 wake_free = 1;
 648                 }
 649                 spin_unlock(&bswspin_mem[iter]);
 650                 if (wake)
 651                         wakeup(&pbuf_mem_count);
 652         } else if (bp >= swbuf_kva && bp < &swbuf_kva[nswbuf_kva]) {
 653                 KKASSERT(bp->b_kvabase);
 654                 spin_lock(&bswspin_kva[iter]);
 655                 TAILQ_INSERT_HEAD(&bswlist_kva[iter], bp, b_freelist);
 656                 if (atomic_fetchadd_int(&pbuf_kva_count, 1) == nswbuf_kva / 4)
 657                         wake = 1;
 658                 if (pfreecnt) {
 659                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 660                                 wake_free = 1;
 661                 }
 662                 spin_unlock(&bswspin_kva[iter]);
 663                 if (wake)
 664                         wakeup(&pbuf_kva_count);
 665         } else {
 666                 KKASSERT(bp->b_kvabase == NULL);
 667                 KKASSERT(bp >= swbuf_raw && bp < &swbuf_raw[nswbuf_raw]);
 668                 spin_lock(&bswspin_raw[iter]);
 669                 TAILQ_INSERT_HEAD(&bswlist_raw[iter], bp, b_freelist);
 670                 if (atomic_fetchadd_int(&pbuf_raw_count, 1) == nswbuf_raw / 4)
 671                         wake = 1;
 672                 if (pfreecnt) {
 673                         if (atomic_fetchadd_int(pfreecnt, 1) == 1)
 674                                 wake_free = 1;
 675                 }
 676                 spin_unlock(&bswspin_raw[iter]);
 677                 if (wake)
 678                         wakeup(&pbuf_raw_count);
 679         }
 680         if (wake_free)
 681                 wakeup(pfreecnt);
 682 }
 683
 684 void
 685 pbuf_adjcount(int *pfreecnt, int n)
 686 {
 687         if (n) {
 688                 atomic_add_int(pfreecnt, n);
 689                 wakeup(pfreecnt);
 690         }
 691 }