sys/vm/vm_swapcache.c

   1 /*
   2  * (MPSAFE)
   3  *
   4  * Copyright (c) 2010 The DragonFly Project.  All rights reserved.
   5  *
   6  * This code is derived from software contributed to The DragonFly Project
   7  * by Matthew Dillon <dillon@backplane.com>
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  *
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice, this list of conditions and the following disclaimer in
  17  *    the documentation and/or other materials provided with the
  18  *    distribution.
  19  * 3. Neither the name of The DragonFly Project nor the names of its
  20  *    contributors may be used to endorse or promote products derived
  21  *    from this software without specific, prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  26  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  27  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  28  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  29  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  30  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  31  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  32  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  33  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  34  * SUCH DAMAGE.
  35  */
  36
  37 /*
  38  * Implement the swapcache daemon.  When enabled swap is assumed to be
  39  * configured on a fast storage device such as a SSD.  Swap is assigned
  40  * to clean vnode-backed pages in the inactive queue, clustered by object
  41  * if possible, and written out.  The swap assignment sticks around even
  42  * after the underlying pages have been recycled.
  43  *
  44  * The daemon manages write bandwidth based on sysctl settings to control
  45  * wear on the SSD.
  46  *
  47  * The vnode strategy code will check for the swap assignments and divert
  48  * reads to the swap device when the data is present in the swapcache.
  49  *
  50  * This operates on both regular files and the block device vnodes used by
  51  * filesystems to manage meta-data.
  52  */
  53
  54 #include "opt_vm.h"
  55 #include <sys/param.h>
  56 #include <sys/systm.h>
  57 #include <sys/kernel.h>
  58 #include <sys/proc.h>
  59 #include <sys/kthread.h>
  60 #include <sys/resourcevar.h>
  61 #include <sys/signalvar.h>
  62 #include <sys/vnode.h>
  63 #include <sys/vmmeter.h>
  64 #include <sys/sysctl.h>
  65 #include <sys/eventhandler.h>
  66
  67 #include <vm/vm.h>
  68 #include <vm/vm_param.h>
  69 #include <sys/lock.h>
  70 #include <vm/vm_object.h>
  71 #include <vm/vm_page.h>
  72 #include <vm/vm_map.h>
  73 #include <vm/vm_pageout.h>
  74 #include <vm/vm_pager.h>
  75 #include <vm/swap_pager.h>
  76 #include <vm/vm_extern.h>
  77
  78 #include <sys/thread2.h>
  79 #include <sys/spinlock2.h>
  80 #include <vm/vm_page2.h>
  81
  82 /* the kernel process "vm_pageout"*/
  83 static int vm_swapcached_flush (vm_page_t m, int isblkdev);
  84 static int vm_swapcache_test(vm_page_t m);
  85 static void vm_swapcache_writing(vm_page_t marker);
  86 static void vm_swapcache_cleaning(vm_object_t marker);
  87 struct thread *swapcached_thread;
  88
  89 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  90
  91 int vm_swapcache_read_enable;
  92 int vm_swapcache_inactive_heuristic;
  93 static int vm_swapcache_sleep;
  94 static int vm_swapcache_maxlaunder = 256;
  95 static int vm_swapcache_data_enable = 0;
  96 static int vm_swapcache_meta_enable = 0;
  97 static int vm_swapcache_maxswappct = 75;
  98 static int vm_swapcache_hysteresis;
  99 int vm_swapcache_use_chflags = 1;       /* require chflags cache */
 100 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 101 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 102 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 103 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 104 static int64_t vm_swapcache_write_count;
 105 static int64_t vm_swapcache_maxfilesize;
 106
 107 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 108         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 109
 110 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 111         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 112 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 113         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 114 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 115         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 116 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 117         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 118 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 119         CTLFLAG_RW, &vm_swapcache_hysteresis, 0, "");
 120 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 121         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 122
 123 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 124         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 125 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 126         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 127 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 128         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 129 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 130         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 131 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 132         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 134         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 135
 136 #define SWAPMAX(adj)    \
 137         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 138
 139 /*
 140  * When shutting down the machine we want to stop swapcache operation
 141  * immediately so swap is not accessed after devices have been shuttered.
 142  */
 143 static void
 144 shutdown_swapcache(void *arg __unused)
 145 {
 146         vm_swapcache_read_enable = 0;
 147         vm_swapcache_data_enable = 0;
 148         vm_swapcache_meta_enable = 0;
 149         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 150 }
 151
 152 /*
 153  * vm_swapcached is the high level pageout daemon.
 154  *
 155  * No requirements.
 156  */
 157 static void
 158 vm_swapcached_thread(void)
 159 {
 160         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 161         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 162         static struct vm_page page_marker[PQ_L2_SIZE];
 163         static struct vm_object object_marker;
 164         int q;
 165
 166         /*
 167          * Thread setup
 168          */
 169         curthread->td_flags |= TDF_SYSTHREAD;
 170         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 171                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 172         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 173                               NULL, SHUTDOWN_PRI_SECOND);
 174
 175         /*
 176          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 177          */
 178         bzero(&page_marker, sizeof(page_marker));
 179         for (q = 0; q < PQ_L2_SIZE; ++q) {
 180                 page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 181                 page_marker[q].queue = PQ_INACTIVE + q;
 182                 page_marker[q].pc = q;
 183                 page_marker[q].wire_count = 1;
 184                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 185                 TAILQ_INSERT_HEAD(
 186                         &vm_page_queues[PQ_INACTIVE + q].pl,
 187                         &page_marker[q], pageq);
 188                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 189         }
 190
 191         vm_swapcache_hysteresis = vmstats.v_inactive_target / 2;
 192         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 193
 194         /*
 195          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 196          */
 197         bzero(&object_marker, sizeof(object_marker));
 198         object_marker.type = OBJT_MARKER;
 199         lwkt_gettoken(&vmobj_token);
 200         TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
 201         lwkt_reltoken(&vmobj_token);
 202
 203         for (;;) {
 204                 /*
 205                  * Handle shutdown
 206                  */
 207                 kproc_suspend_loop();
 208
 209                 /*
 210                  * Check every 5 seconds when not enabled or if no swap
 211                  * is present.
 212                  */
 213                 if ((vm_swapcache_data_enable == 0 &&
 214                      vm_swapcache_meta_enable == 0) ||
 215                     vm_swap_max == 0) {
 216                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 217                         continue;
 218                 }
 219
 220                 /*
 221                  * Polling rate when enabled is approximately 10 hz.
 222                  */
 223                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 224
 225                 /*
 226                  * State hysteresis.  Generate write activity up to 75% of
 227                  * swap, then clean out swap assignments down to 70%, then
 228                  * repeat.
 229                  */
 230                 if (state == SWAPC_WRITING) {
 231                         if (vm_swap_cache_use > SWAPMAX(0))
 232                                 state = SWAPC_CLEANING;
 233                 } else {
 234                         if (vm_swap_cache_use < SWAPMAX(-10))
 235                                 state = SWAPC_WRITING;
 236                 }
 237
 238                 /*
 239                  * We are allowed to continue accumulating burst value
 240                  * in either state.  Allow the user to set curburst > maxburst
 241                  * for the initial load-in.
 242                  */
 243                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 244                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 245                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 246                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 247                 }
 248
 249                 /*
 250                  * We don't want to nickle-and-dime the scan as that will
 251                  * create unnecessary fragmentation.  The minimum burst
 252                  * is one-seconds worth of accumulation.
 253                  */
 254                 if (state == SWAPC_WRITING) {
 255                         if (vm_swapcache_curburst >= vm_swapcache_accrate) {
 256                                 if (burst == SWAPB_BURSTING) {
 257                                         for (q = 0; q < PQ_L2_SIZE; ++q) {
 258                                                 vm_swapcache_writing(
 259                                                         &page_marker[q]);
 260                                         }
 261                                         if (vm_swapcache_curburst <= 0)
 262                                                 burst = SWAPB_RECOVERING;
 263                                 } else if (vm_swapcache_curburst >
 264                                            vm_swapcache_minburst) {
 265                                         for (q = 0; q < PQ_L2_SIZE; ++q) {
 266                                                 vm_swapcache_writing(
 267                                                         &page_marker[q]);
 268                                         }
 269                                         burst = SWAPB_BURSTING;
 270                                 }
 271                         }
 272                 } else {
 273                         vm_swapcache_cleaning(&object_marker);
 274                 }
 275         }
 276
 277         /*
 278          * Cleanup (NOT REACHED)
 279          */
 280         for (q = 0; q < PQ_L2_SIZE; ++q) {
 281                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 282                 TAILQ_REMOVE(
 283                         &vm_page_queues[PQ_INACTIVE + q].pl,
 284                         &page_marker[q], pageq);
 285                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 286         }
 287
 288         lwkt_gettoken(&vmobj_token);
 289         TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
 290         lwkt_reltoken(&vmobj_token);
 291 }
 292
 293 static struct kproc_desc swpc_kp = {
 294         "swapcached",
 295         vm_swapcached_thread,
 296         &swapcached_thread
 297 };
 298 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
 299
 300 static void
 301 vm_swapcache_writing(vm_page_t marker)
 302 {
 303         vm_object_t object;
 304         struct vnode *vp;
 305         vm_page_t m;
 306         int count;
 307         int isblkdev;
 308
 309         /*
 310          * Deal with an overflow of the heuristic counter or if the user
 311          * manually changes the hysteresis.
 312          *
 313          * Try to avoid small incremental pageouts by waiting for enough
 314          * pages to buildup in the inactive queue to hopefully get a good
 315          * burst in.  This heuristic is bumped by the VM system and reset
 316          * when our scan hits the end of the queue.
 317          */
 318         if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis)
 319                 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 320         if (vm_swapcache_inactive_heuristic < 0)
 321                 return;
 322
 323         /*
 324          * Scan the inactive queue from our marker to locate
 325          * suitable pages to push to the swap cache.
 326          *
 327          * We are looking for clean vnode-backed pages.
 328          *
 329          * NOTE: PG_SWAPPED pages in particular are not part of
 330          *       our count because once the cache stabilizes we
 331          *       can end up with a very high datarate of VM pages
 332          *       cycling from it.
 333          */
 334         count = vm_swapcache_maxlaunder;
 335
 336         vm_page_queues_spin_lock(marker->queue);
 337         while ((m = TAILQ_NEXT(marker, pageq)) != NULL && count-- > 0) {
 338                 KKASSERT(m->queue == marker->queue);
 339
 340                 if (vm_swapcache_curburst < 0)
 341                         break;
 342                 TAILQ_REMOVE(
 343                         &vm_page_queues[marker->queue].pl, marker, pageq);
 344                 TAILQ_INSERT_AFTER(
 345                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 346                 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
 347                         ++count;
 348                         continue;
 349                 }
 350                 if (vm_page_busy_try(m, TRUE))
 351                         continue;
 352                 vm_page_queues_spin_unlock(marker->queue);
 353
 354                 if ((object = m->object) == NULL) {
 355                         vm_page_wakeup(m);
 356                         vm_page_queues_spin_lock(marker->queue);
 357                         continue;
 358                 }
 359                 vm_object_hold(object);
 360                 if (m->object != object) {
 361                         vm_object_drop(object);
 362                         vm_page_wakeup(m);
 363                         vm_page_queues_spin_lock(marker->queue);
 364                         continue;
 365                 }
 366                 if (vm_swapcache_test(m)) {
 367                         vm_object_drop(object);
 368                         vm_page_wakeup(m);
 369                         vm_page_queues_spin_lock(marker->queue);
 370                         continue;
 371                 }
 372
 373                 vp = object->handle;
 374                 if (vp == NULL) {
 375                         vm_object_drop(object);
 376                         vm_page_wakeup(m);
 377                         vm_page_queues_spin_lock(marker->queue);
 378                         continue;
 379                 }
 380
 381                 switch(vp->v_type) {
 382                 case VREG:
 383                         /*
 384                          * PG_NOTMETA generically means 'don't swapcache this',
 385                          * and HAMMER will set this for regular data buffers
 386                          * (and leave it unset for meta-data buffers) as
 387                          * appropriate when double buffering is enabled.
 388                          */
 389                         if (m->flags & PG_NOTMETA) {
 390                                 vm_object_drop(object);
 391                                 vm_page_wakeup(m);
 392                                 vm_page_queues_spin_lock(marker->queue);
 393                                 continue;
 394                         }
 395
 396                         /*
 397                          * If data_enable is 0 do not try to swapcache data.
 398                          * If use_chflags is set then only swapcache data for
 399                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 400                          */
 401                         if (vm_swapcache_data_enable == 0 ||
 402                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 403                              vm_swapcache_use_chflags)) {
 404                                 vm_object_drop(object);
 405                                 vm_page_wakeup(m);
 406                                 vm_page_queues_spin_lock(marker->queue);
 407                                 continue;
 408                         }
 409                         if (vm_swapcache_maxfilesize &&
 410                             object->size >
 411                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 412                                 vm_object_drop(object);
 413                                 vm_page_wakeup(m);
 414                                 vm_page_queues_spin_lock(marker->queue);
 415                                 continue;
 416                         }
 417                         isblkdev = 0;
 418                         break;
 419                 case VCHR:
 420                         /*
 421                          * PG_NOTMETA generically means 'don't swapcache this',
 422                          * and HAMMER will set this for regular data buffers
 423                          * (and leave it unset for meta-data buffers) as
 424                          * appropriate when double buffering is enabled.
 425                          */
 426                         if (m->flags & PG_NOTMETA) {
 427                                 vm_object_drop(object);
 428                                 vm_page_wakeup(m);
 429                                 vm_page_queues_spin_lock(marker->queue);
 430                                 continue;
 431                         }
 432                         if (vm_swapcache_meta_enable == 0) {
 433                                 vm_object_drop(object);
 434                                 vm_page_wakeup(m);
 435                                 vm_page_queues_spin_lock(marker->queue);
 436                                 continue;
 437                         }
 438                         isblkdev = 1;
 439                         break;
 440                 default:
 441                         vm_object_drop(object);
 442                         vm_page_wakeup(m);
 443                         vm_page_queues_spin_lock(marker->queue);
 444                         continue;
 445                 }
 446
 447
 448                 /*
 449                  * Assign swap and initiate I/O.
 450                  *
 451                  * (adjust for the --count which also occurs in the loop)
 452                  */
 453                 count -= vm_swapcached_flush(m, isblkdev) - 1;
 454
 455                 /*
 456                  * Setup for next loop using marker.
 457                  */
 458                 vm_object_drop(object);
 459                 vm_page_queues_spin_lock(marker->queue);
 460         }
 461
 462         /*
 463          * The marker could wind up at the end, which is ok.  If we hit the
 464          * end of the list adjust the heuristic.
 465          *
 466          * Earlier inactive pages that were dirty and become clean
 467          * are typically moved to the end of PQ_INACTIVE by virtue
 468          * of vfs_vmio_release() when they become unwired from the
 469          * buffer cache.
 470          */
 471         if (m == NULL)
 472                 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 473         vm_page_queues_spin_unlock(marker->queue);
 474 }
 475
 476 /*
 477  * Flush the specified page using the swap_pager.  The page
 478  * must be busied by the caller and its disposition will become
 479  * the responsibility of this function.
 480  *
 481  * Try to collect surrounding pages, including pages which may
 482  * have already been assigned swap.  Try to cluster within a
 483  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 484  * to match what swap_pager_putpages() can do.
 485  *
 486  * We also want to try to match against the buffer cache blocksize
 487  * but we don't really know what it is here.  Since the buffer cache
 488  * wires and unwires pages in groups the fact that we skip wired pages
 489  * should be sufficient.
 490  *
 491  * Returns a count of pages we might have flushed (minimum 1)
 492  */
 493 static
 494 int
 495 vm_swapcached_flush(vm_page_t m, int isblkdev)
 496 {
 497         vm_object_t object;
 498         vm_page_t marray[SWAP_META_PAGES];
 499         vm_pindex_t basei;
 500         int rtvals[SWAP_META_PAGES];
 501         int x;
 502         int i;
 503         int j;
 504         int count;
 505         int error;
 506
 507         vm_page_io_start(m);
 508         vm_page_protect(m, VM_PROT_READ);
 509         object = m->object;
 510         vm_object_hold(object);
 511
 512         /*
 513          * Try to cluster around (m), keeping in mind that the swap pager
 514          * can only do SMAP_META_PAGES worth of continguous write.
 515          */
 516         x = (int)m->pindex & SWAP_META_MASK;
 517         marray[x] = m;
 518         basei = m->pindex;
 519         vm_page_wakeup(m);
 520
 521         for (i = x - 1; i >= 0; --i) {
 522                 m = vm_page_lookup_busy_try(object, basei - x + i,
 523                                             TRUE, &error);
 524                 if (error || m == NULL)
 525                         break;
 526                 if (vm_swapcache_test(m)) {
 527                         vm_page_wakeup(m);
 528                         break;
 529                 }
 530                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 531                         vm_page_wakeup(m);
 532                         break;
 533                 }
 534                 vm_page_io_start(m);
 535                 vm_page_protect(m, VM_PROT_READ);
 536                 if (m->queue - m->pc == PQ_CACHE) {
 537                         vm_page_unqueue_nowakeup(m);
 538                         vm_page_deactivate(m);
 539                 }
 540                 marray[i] = m;
 541                 vm_page_wakeup(m);
 542         }
 543         ++i;
 544
 545         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 546                 m = vm_page_lookup_busy_try(object, basei - x + j,
 547                                             TRUE, &error);
 548                 if (error || m == NULL)
 549                         break;
 550                 if (vm_swapcache_test(m)) {
 551                         vm_page_wakeup(m);
 552                         break;
 553                 }
 554                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 555                         vm_page_wakeup(m);
 556                         break;
 557                 }
 558                 vm_page_io_start(m);
 559                 vm_page_protect(m, VM_PROT_READ);
 560                 if (m->queue - m->pc == PQ_CACHE) {
 561                         vm_page_unqueue_nowakeup(m);
 562                         vm_page_deactivate(m);
 563                 }
 564                 marray[j] = m;
 565                 vm_page_wakeup(m);
 566         }
 567
 568         count = j - i;
 569         vm_object_pip_add(object, count);
 570         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 571         vm_swapcache_write_count += count * PAGE_SIZE;
 572         vm_swapcache_curburst -= count * PAGE_SIZE;
 573
 574         while (i < j) {
 575                 if (rtvals[i] != VM_PAGER_PEND) {
 576                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 577                         vm_page_io_finish(marray[i]);
 578                         vm_page_wakeup(marray[i]);
 579                         vm_object_pip_wakeup(object);
 580                 }
 581                 ++i;
 582         }
 583         vm_object_drop(object);
 584         return(count);
 585 }
 586
 587 /*
 588  * Test whether a VM page is suitable for writing to the swapcache.
 589  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 590  *
 591  * Returns 0 on success, 1 on failure
 592  */
 593 static int
 594 vm_swapcache_test(vm_page_t m)
 595 {
 596         vm_object_t object;
 597
 598         if (m->flags & PG_UNMANAGED)
 599                 return(1);
 600         if (m->hold_count || m->wire_count)
 601                 return(1);
 602         if (m->valid != VM_PAGE_BITS_ALL)
 603                 return(1);
 604         if (m->dirty & m->valid)
 605                 return(1);
 606         if ((object = m->object) == NULL)
 607                 return(1);
 608         if (object->type != OBJT_VNODE ||
 609             (object->flags & OBJ_DEAD)) {
 610                 return(1);
 611         }
 612         vm_page_test_dirty(m);
 613         if (m->dirty & m->valid)
 614                 return(1);
 615         return(0);
 616 }
 617
 618 /*
 619  * Cleaning pass
 620  */
 621 static
 622 void
 623 vm_swapcache_cleaning(vm_object_t marker)
 624 {
 625         vm_object_t object;
 626         struct vnode *vp;
 627         int count;
 628         int n;
 629
 630         object = marker;
 631         count = vm_swapcache_maxlaunder;
 632
 633         /*
 634          * Look for vnode objects
 635          */
 636         lwkt_gettoken(&vmobj_token);
 637
 638         while ((object = TAILQ_NEXT(object, object_list)) != NULL) {
 639                 vm_object_hold(object);
 640
 641                 lwkt_yield();
 642                 if (--count <= 0) {
 643                         vm_object_drop(object);
 644                         break;
 645                 }
 646
 647                 /*
 648                  * Only operate on live VNODE objects with regular/chardev types
 649                  */
 650                 if ((object->type != OBJT_VNODE) ||
 651                     ((object->flags & OBJ_DEAD) || object->swblock_count == 0) ||
 652                     ((vp = object->handle) == NULL) ||
 653                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 654                         vm_object_drop(object);
 655                         continue;
 656                 }
 657
 658                 /*
 659                  * Adjust iterator.
 660                  */
 661                 if (marker->backing_object != object)
 662                         marker->size = 0;
 663
 664                 /*
 665                  * Move the marker so we can work on the VM object
 666                  */
 667                 TAILQ_REMOVE(&vm_object_list, marker, object_list);
 668                 TAILQ_INSERT_AFTER(&vm_object_list, object,
 669                                    marker, object_list);
 670
 671                 /*
 672                  * Look for swblocks starting at our iterator.
 673                  *
 674                  * The swap_pager_condfree() function attempts to free
 675                  * swap space starting at the specified index.  The index
 676                  * will be updated on return.  The function will return
 677                  * a scan factor (NOT the number of blocks freed).
 678                  *
 679                  * If it must cut its scan of the object short due to an
 680                  * excessive number of swblocks, or is able to free the
 681                  * requested number of blocks, it will return n >= count
 682                  * and we break and pick it back up on a future attempt.
 683                  */
 684                 lwkt_token_swap();
 685                 lwkt_reltoken(&vmobj_token);
 686
 687                 n = swap_pager_condfree(object, &marker->size, count);
 688
 689                 vm_object_drop(object);
 690                 lwkt_gettoken(&vmobj_token);
 691
 692                 count -= n;
 693                 if (count < 0)
 694                         break;
 695
 696                 /*
 697                  * Setup for loop.
 698                  */
 699                 marker->size = 0;
 700                 object = marker;
 701         }
 702
 703         /*
 704          * Adjust marker so we continue the scan from where we left off.
 705          * When we reach the end we start back at the beginning.
 706          */
 707         TAILQ_REMOVE(&vm_object_list, marker, object_list);
 708         if (object)
 709                 TAILQ_INSERT_BEFORE(object, marker, object_list);
 710         else
 711                 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
 712         marker->backing_object = object;
 713
 714         lwkt_reltoken(&vmobj_token);
 715 }