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 static void vm_swapcache_movemarker(vm_object_t marker, vm_object_t object);
  88 struct thread *swapcached_thread;
  89
  90 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  91
  92 int vm_swapcache_read_enable;
  93 int vm_swapcache_inactive_heuristic;
  94 static int vm_swapcache_sleep;
  95 static int vm_swapcache_maxlaunder = 256;
  96 static int vm_swapcache_data_enable = 0;
  97 static int vm_swapcache_meta_enable = 0;
  98 static int vm_swapcache_maxswappct = 75;
  99 static int vm_swapcache_hysteresis;
 100 int vm_swapcache_use_chflags = 1;       /* require chflags cache */
 101 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 102 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 103 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 104 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 105 static int64_t vm_swapcache_write_count;
 106 static int64_t vm_swapcache_maxfilesize;
 107 static int64_t vm_swapcache_cleanperobj = 16*1024*1024;
 108
 109 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 110         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 111
 112 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 113         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 114 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 115         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 116 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 117         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 118 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 119         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 120 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 121         CTLFLAG_RW, &vm_swapcache_hysteresis, 0, "");
 122 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 123         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 124
 125 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 126         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 127 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 128         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 129 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 130         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 131 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 132         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 134         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 135 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 136         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 137 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj,
 138         CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, "");
 139
 140 #define SWAPMAX(adj)    \
 141         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 142
 143 /*
 144  * When shutting down the machine we want to stop swapcache operation
 145  * immediately so swap is not accessed after devices have been shuttered.
 146  */
 147 static void
 148 shutdown_swapcache(void *arg __unused)
 149 {
 150         vm_swapcache_read_enable = 0;
 151         vm_swapcache_data_enable = 0;
 152         vm_swapcache_meta_enable = 0;
 153         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 154 }
 155
 156 /*
 157  * vm_swapcached is the high level pageout daemon.
 158  *
 159  * No requirements.
 160  */
 161 static void
 162 vm_swapcached_thread(void)
 163 {
 164         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 165         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 166         static struct vm_page page_marker[PQ_L2_SIZE];
 167         static struct vm_object object_marker;
 168         int q;
 169
 170         /*
 171          * Thread setup
 172          */
 173         curthread->td_flags |= TDF_SYSTHREAD;
 174         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 175                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 176         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 177                               NULL, SHUTDOWN_PRI_SECOND);
 178
 179         /*
 180          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 181          */
 182         bzero(&page_marker, sizeof(page_marker));
 183         for (q = 0; q < PQ_L2_SIZE; ++q) {
 184                 page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 185                 page_marker[q].queue = PQ_INACTIVE + q;
 186                 page_marker[q].pc = q;
 187                 page_marker[q].wire_count = 1;
 188                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 189                 TAILQ_INSERT_HEAD(
 190                         &vm_page_queues[PQ_INACTIVE + q].pl,
 191                         &page_marker[q], pageq);
 192                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 193         }
 194
 195         vm_swapcache_hysteresis = vmstats.v_inactive_target / 2;
 196         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 197
 198         /*
 199          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 200          */
 201         bzero(&object_marker, sizeof(object_marker));
 202         object_marker.type = OBJT_MARKER;
 203         lwkt_gettoken(&vmobj_token);
 204         TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
 205         lwkt_reltoken(&vmobj_token);
 206
 207         for (;;) {
 208                 /*
 209                  * Handle shutdown
 210                  */
 211                 kproc_suspend_loop();
 212
 213                 /*
 214                  * Check every 5 seconds when not enabled or if no swap
 215                  * is present.
 216                  */
 217                 if ((vm_swapcache_data_enable == 0 &&
 218                      vm_swapcache_meta_enable == 0) ||
 219                     vm_swap_max == 0) {
 220                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 221                         continue;
 222                 }
 223
 224                 /*
 225                  * Polling rate when enabled is approximately 10 hz.
 226                  */
 227                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 228
 229                 /*
 230                  * State hysteresis.  Generate write activity up to 75% of
 231                  * swap, then clean out swap assignments down to 70%, then
 232                  * repeat.
 233                  */
 234                 if (state == SWAPC_WRITING) {
 235                         if (vm_swap_cache_use > SWAPMAX(0))
 236                                 state = SWAPC_CLEANING;
 237                 } else {
 238                         if (vm_swap_cache_use < SWAPMAX(-10))
 239                                 state = SWAPC_WRITING;
 240                 }
 241
 242                 /*
 243                  * We are allowed to continue accumulating burst value
 244                  * in either state.  Allow the user to set curburst > maxburst
 245                  * for the initial load-in.
 246                  */
 247                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 248                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 249                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 250                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 251                 }
 252
 253                 /*
 254                  * We don't want to nickle-and-dime the scan as that will
 255                  * create unnecessary fragmentation.  The minimum burst
 256                  * is one-seconds worth of accumulation.
 257                  */
 258                 if (state == SWAPC_WRITING) {
 259                         if (vm_swapcache_curburst >= vm_swapcache_accrate) {
 260                                 if (burst == SWAPB_BURSTING) {
 261                                         for (q = 0; q < PQ_L2_SIZE; ++q) {
 262                                                 vm_swapcache_writing(
 263                                                         &page_marker[q]);
 264                                         }
 265                                         if (vm_swapcache_curburst <= 0)
 266                                                 burst = SWAPB_RECOVERING;
 267                                 } else if (vm_swapcache_curburst >
 268                                            vm_swapcache_minburst) {
 269                                         for (q = 0; q < PQ_L2_SIZE; ++q) {
 270                                                 vm_swapcache_writing(
 271                                                         &page_marker[q]);
 272                                         }
 273                                         burst = SWAPB_BURSTING;
 274                                 }
 275                         }
 276                 } else {
 277                         vm_swapcache_cleaning(&object_marker);
 278                 }
 279         }
 280
 281         /*
 282          * Cleanup (NOT REACHED)
 283          */
 284         for (q = 0; q < PQ_L2_SIZE; ++q) {
 285                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 286                 TAILQ_REMOVE(
 287                         &vm_page_queues[PQ_INACTIVE + q].pl,
 288                         &page_marker[q], pageq);
 289                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 290         }
 291
 292         lwkt_gettoken(&vmobj_token);
 293         TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
 294         lwkt_reltoken(&vmobj_token);
 295 }
 296
 297 static struct kproc_desc swpc_kp = {
 298         "swapcached",
 299         vm_swapcached_thread,
 300         &swapcached_thread
 301 };
 302 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
 303
 304 static void
 305 vm_swapcache_writing(vm_page_t marker)
 306 {
 307         vm_object_t object;
 308         struct vnode *vp;
 309         vm_page_t m;
 310         int count;
 311         int isblkdev;
 312
 313         /*
 314          * Deal with an overflow of the heuristic counter or if the user
 315          * manually changes the hysteresis.
 316          *
 317          * Try to avoid small incremental pageouts by waiting for enough
 318          * pages to buildup in the inactive queue to hopefully get a good
 319          * burst in.  This heuristic is bumped by the VM system and reset
 320          * when our scan hits the end of the queue.
 321          */
 322         if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis)
 323                 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 324         if (vm_swapcache_inactive_heuristic < 0)
 325                 return;
 326
 327         /*
 328          * Scan the inactive queue from our marker to locate
 329          * suitable pages to push to the swap cache.
 330          *
 331          * We are looking for clean vnode-backed pages.
 332          *
 333          * NOTE: PG_SWAPPED pages in particular are not part of
 334          *       our count because once the cache stabilizes we
 335          *       can end up with a very high datarate of VM pages
 336          *       cycling from it.
 337          */
 338         count = vm_swapcache_maxlaunder;
 339
 340         vm_page_queues_spin_lock(marker->queue);
 341         while ((m = TAILQ_NEXT(marker, pageq)) != NULL && count-- > 0) {
 342                 KKASSERT(m->queue == marker->queue);
 343
 344                 if (vm_swapcache_curburst < 0)
 345                         break;
 346                 TAILQ_REMOVE(
 347                         &vm_page_queues[marker->queue].pl, marker, pageq);
 348                 TAILQ_INSERT_AFTER(
 349                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 350
 351                 /*
 352                  * Ignore markers and ignore pages that already have a swap
 353                  * assignment.
 354                  */
 355                 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
 356                         ++count;
 357                         continue;
 358                 }
 359                 if (vm_page_busy_try(m, TRUE))
 360                         continue;
 361                 vm_page_queues_spin_unlock(marker->queue);
 362
 363                 if ((object = m->object) == NULL) {
 364                         vm_page_wakeup(m);
 365                         vm_page_queues_spin_lock(marker->queue);
 366                         continue;
 367                 }
 368                 vm_object_hold(object);
 369                 if (m->object != object) {
 370                         vm_object_drop(object);
 371                         vm_page_wakeup(m);
 372                         vm_page_queues_spin_lock(marker->queue);
 373                         continue;
 374                 }
 375                 if (vm_swapcache_test(m)) {
 376                         vm_object_drop(object);
 377                         vm_page_wakeup(m);
 378                         vm_page_queues_spin_lock(marker->queue);
 379                         continue;
 380                 }
 381
 382                 vp = object->handle;
 383                 if (vp == NULL) {
 384                         vm_object_drop(object);
 385                         vm_page_wakeup(m);
 386                         vm_page_queues_spin_lock(marker->queue);
 387                         continue;
 388                 }
 389
 390                 switch(vp->v_type) {
 391                 case VREG:
 392                         /*
 393                          * PG_NOTMETA generically means 'don't swapcache this',
 394                          * and HAMMER will set this for regular data buffers
 395                          * (and leave it unset for meta-data buffers) as
 396                          * appropriate when double buffering is enabled.
 397                          */
 398                         if (m->flags & PG_NOTMETA) {
 399                                 vm_object_drop(object);
 400                                 vm_page_wakeup(m);
 401                                 vm_page_queues_spin_lock(marker->queue);
 402                                 continue;
 403                         }
 404
 405                         /*
 406                          * If data_enable is 0 do not try to swapcache data.
 407                          * If use_chflags is set then only swapcache data for
 408                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 409                          */
 410                         if (vm_swapcache_data_enable == 0 ||
 411                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 412                              vm_swapcache_use_chflags)) {
 413                                 vm_object_drop(object);
 414                                 vm_page_wakeup(m);
 415                                 vm_page_queues_spin_lock(marker->queue);
 416                                 continue;
 417                         }
 418                         if (vm_swapcache_maxfilesize &&
 419                             object->size >
 420                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 421                                 vm_object_drop(object);
 422                                 vm_page_wakeup(m);
 423                                 vm_page_queues_spin_lock(marker->queue);
 424                                 continue;
 425                         }
 426                         isblkdev = 0;
 427                         break;
 428                 case VCHR:
 429                         /*
 430                          * PG_NOTMETA generically means 'don't swapcache this',
 431                          * and HAMMER will set this for regular data buffers
 432                          * (and leave it unset for meta-data buffers) as
 433                          * appropriate when double buffering is enabled.
 434                          */
 435                         if (m->flags & PG_NOTMETA) {
 436                                 vm_object_drop(object);
 437                                 vm_page_wakeup(m);
 438                                 vm_page_queues_spin_lock(marker->queue);
 439                                 continue;
 440                         }
 441                         if (vm_swapcache_meta_enable == 0) {
 442                                 vm_object_drop(object);
 443                                 vm_page_wakeup(m);
 444                                 vm_page_queues_spin_lock(marker->queue);
 445                                 continue;
 446                         }
 447                         isblkdev = 1;
 448                         break;
 449                 default:
 450                         vm_object_drop(object);
 451                         vm_page_wakeup(m);
 452                         vm_page_queues_spin_lock(marker->queue);
 453                         continue;
 454                 }
 455
 456
 457                 /*
 458                  * Assign swap and initiate I/O.
 459                  *
 460                  * (adjust for the --count which also occurs in the loop)
 461                  */
 462                 count -= vm_swapcached_flush(m, isblkdev) - 1;
 463
 464                 /*
 465                  * Setup for next loop using marker.
 466                  */
 467                 vm_object_drop(object);
 468                 vm_page_queues_spin_lock(marker->queue);
 469         }
 470
 471         /*
 472          * The marker could wind up at the end, which is ok.  If we hit the
 473          * end of the list adjust the heuristic.
 474          *
 475          * Earlier inactive pages that were dirty and become clean
 476          * are typically moved to the end of PQ_INACTIVE by virtue
 477          * of vfs_vmio_release() when they become unwired from the
 478          * buffer cache.
 479          */
 480         if (m == NULL)
 481                 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 482         vm_page_queues_spin_unlock(marker->queue);
 483 }
 484
 485 /*
 486  * Flush the specified page using the swap_pager.  The page
 487  * must be busied by the caller and its disposition will become
 488  * the responsibility of this function.
 489  *
 490  * Try to collect surrounding pages, including pages which may
 491  * have already been assigned swap.  Try to cluster within a
 492  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 493  * to match what swap_pager_putpages() can do.
 494  *
 495  * We also want to try to match against the buffer cache blocksize
 496  * but we don't really know what it is here.  Since the buffer cache
 497  * wires and unwires pages in groups the fact that we skip wired pages
 498  * should be sufficient.
 499  *
 500  * Returns a count of pages we might have flushed (minimum 1)
 501  */
 502 static
 503 int
 504 vm_swapcached_flush(vm_page_t m, int isblkdev)
 505 {
 506         vm_object_t object;
 507         vm_page_t marray[SWAP_META_PAGES];
 508         vm_pindex_t basei;
 509         int rtvals[SWAP_META_PAGES];
 510         int x;
 511         int i;
 512         int j;
 513         int count;
 514         int error;
 515
 516         vm_page_io_start(m);
 517         vm_page_protect(m, VM_PROT_READ);
 518         object = m->object;
 519         vm_object_hold(object);
 520
 521         /*
 522          * Try to cluster around (m), keeping in mind that the swap pager
 523          * can only do SMAP_META_PAGES worth of continguous write.
 524          */
 525         x = (int)m->pindex & SWAP_META_MASK;
 526         marray[x] = m;
 527         basei = m->pindex;
 528         vm_page_wakeup(m);
 529
 530         for (i = x - 1; i >= 0; --i) {
 531                 m = vm_page_lookup_busy_try(object, basei - x + i,
 532                                             TRUE, &error);
 533                 if (error || m == NULL)
 534                         break;
 535                 if (vm_swapcache_test(m)) {
 536                         vm_page_wakeup(m);
 537                         break;
 538                 }
 539                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 540                         vm_page_wakeup(m);
 541                         break;
 542                 }
 543                 vm_page_io_start(m);
 544                 vm_page_protect(m, VM_PROT_READ);
 545                 if (m->queue - m->pc == PQ_CACHE) {
 546                         vm_page_unqueue_nowakeup(m);
 547                         vm_page_deactivate(m);
 548                 }
 549                 marray[i] = m;
 550                 vm_page_wakeup(m);
 551         }
 552         ++i;
 553
 554         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 555                 m = vm_page_lookup_busy_try(object, basei - x + j,
 556                                             TRUE, &error);
 557                 if (error || m == NULL)
 558                         break;
 559                 if (vm_swapcache_test(m)) {
 560                         vm_page_wakeup(m);
 561                         break;
 562                 }
 563                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 564                         vm_page_wakeup(m);
 565                         break;
 566                 }
 567                 vm_page_io_start(m);
 568                 vm_page_protect(m, VM_PROT_READ);
 569                 if (m->queue - m->pc == PQ_CACHE) {
 570                         vm_page_unqueue_nowakeup(m);
 571                         vm_page_deactivate(m);
 572                 }
 573                 marray[j] = m;
 574                 vm_page_wakeup(m);
 575         }
 576
 577         count = j - i;
 578         vm_object_pip_add(object, count);
 579         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 580         vm_swapcache_write_count += count * PAGE_SIZE;
 581         vm_swapcache_curburst -= count * PAGE_SIZE;
 582
 583         while (i < j) {
 584                 if (rtvals[i] != VM_PAGER_PEND) {
 585                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 586                         vm_page_io_finish(marray[i]);
 587                         vm_page_wakeup(marray[i]);
 588                         vm_object_pip_wakeup(object);
 589                 }
 590                 ++i;
 591         }
 592         vm_object_drop(object);
 593         return(count);
 594 }
 595
 596 /*
 597  * Test whether a VM page is suitable for writing to the swapcache.
 598  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 599  *
 600  * Returns 0 on success, 1 on failure
 601  */
 602 static int
 603 vm_swapcache_test(vm_page_t m)
 604 {
 605         vm_object_t object;
 606
 607         if (m->flags & PG_UNMANAGED)
 608                 return(1);
 609         if (m->hold_count || m->wire_count)
 610                 return(1);
 611         if (m->valid != VM_PAGE_BITS_ALL)
 612                 return(1);
 613         if (m->dirty & m->valid)
 614                 return(1);
 615         if ((object = m->object) == NULL)
 616                 return(1);
 617         if (object->type != OBJT_VNODE ||
 618             (object->flags & OBJ_DEAD)) {
 619                 return(1);
 620         }
 621         vm_page_test_dirty(m);
 622         if (m->dirty & m->valid)
 623                 return(1);
 624         return(0);
 625 }
 626
 627 /*
 628  * Cleaning pass.
 629  *
 630  * We clean whole objects up to 16MB
 631  */
 632 static
 633 void
 634 vm_swapcache_cleaning(vm_object_t marker)
 635 {
 636         vm_object_t object;
 637         struct vnode *vp;
 638         int count;
 639         int n;
 640
 641         count = vm_swapcache_maxlaunder;
 642
 643         /*
 644          * Look for vnode objects
 645          */
 646         lwkt_gettoken(&vmobj_token);
 647
 648         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 649                 /*
 650                  * We have to skip markers.  We cannot hold/drop marker
 651                  * objects!
 652                  */
 653                 if (object->type == OBJT_MARKER) {
 654                         vm_swapcache_movemarker(marker, object);
 655                         continue;
 656                 }
 657
 658                 /*
 659                  * Safety, or in case there are millions of VM objects
 660                  * without swapcache backing.
 661                  */
 662                 if (--count <= 0)
 663                         break;
 664
 665                 /*
 666                  * We must hold the object before potentially yielding.
 667                  */
 668                 vm_object_hold(object);
 669                 lwkt_yield();
 670
 671                 /*
 672                  * Only operate on live VNODE objects that are either
 673                  * VREG or VCHR (VCHR for meta-data).
 674                  */
 675                 if ((object->type != OBJT_VNODE) ||
 676                     ((object->flags & OBJ_DEAD) ||
 677                      object->swblock_count == 0) ||
 678                     ((vp = object->handle) == NULL) ||
 679                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 680                         vm_object_drop(object);
 681                         /* object may be invalid now */
 682                         vm_swapcache_movemarker(marker, object);
 683                         continue;
 684                 }
 685
 686                 /*
 687                  * Reset the object pindex stored in the marker if the
 688                  * working object has changed.
 689                  */
 690                 if (marker->backing_object != object) {
 691                         marker->size = 0;
 692                         marker->backing_object_offset = 0;
 693                         marker->backing_object = object;
 694                 }
 695
 696                 /*
 697                  * Look for swblocks starting at our iterator.
 698                  *
 699                  * The swap_pager_condfree() function attempts to free
 700                  * swap space starting at the specified index.  The index
 701                  * will be updated on return.  The function will return
 702                  * a scan factor (NOT the number of blocks freed).
 703                  *
 704                  * If it must cut its scan of the object short due to an
 705                  * excessive number of swblocks, or is able to free the
 706                  * requested number of blocks, it will return n >= count
 707                  * and we break and pick it back up on a future attempt.
 708                  *
 709                  * Scan the object linearly and try to batch large sets of
 710                  * blocks that are likely to clean out entire swap radix
 711                  * tree leafs.
 712                  */
 713                 lwkt_token_swap();
 714                 lwkt_reltoken(&vmobj_token);
 715
 716                 n = swap_pager_condfree(object, &marker->size,
 717                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 718
 719                 vm_object_drop(object);         /* object may be invalid now */
 720                 lwkt_gettoken(&vmobj_token);
 721
 722                 /*
 723                  * If we have exhausted the object or deleted our per-pass
 724                  * page limit then move us to the next object.  Note that
 725                  * the current object may no longer be on the vm_object_list.
 726                  */
 727                 if (n <= 0 ||
 728                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 729                         vm_swapcache_movemarker(marker, object);
 730                 }
 731
 732                 /*
 733                  * If we have exhausted our max-launder stop for now.
 734                  */
 735                 count -= n;
 736                 marker->backing_object_offset += n * PAGE_SIZE;
 737                 if (count < 0)
 738                         break;
 739         }
 740
 741         /*
 742          * If we wound up at the end of the list this will move the
 743          * marker back to the beginning.
 744          */
 745         if (object == NULL)
 746                 vm_swapcache_movemarker(marker, NULL);
 747
 748         lwkt_reltoken(&vmobj_token);
 749 }
 750
 751 /*
 752  * Move the marker past the current object.  Object can be stale, but we
 753  * still need it to determine if the marker has to be moved.  If the object
 754  * is still the 'current object' (object after the marker), we hop-scotch
 755  * the marker past it.
 756  */
 757 static void
 758 vm_swapcache_movemarker(vm_object_t marker, vm_object_t object)
 759 {
 760         if (TAILQ_NEXT(marker, object_list) == object) {
 761                 TAILQ_REMOVE(&vm_object_list, marker, object_list);
 762                 if (object) {
 763                         TAILQ_INSERT_AFTER(&vm_object_list, object,
 764                                            marker, object_list);
 765                 } else {
 766                         TAILQ_INSERT_HEAD(&vm_object_list,
 767                                           marker, object_list);
 768                 }
 769         }
 770 }