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 int vm_swapcache_writing_heuristic(void);
  86 static int vm_swapcache_writing(vm_page_t marker, int count, int scount);
  87 static void vm_swapcache_cleaning(vm_object_t marker, int *swindexp);
  88 static void vm_swapcache_movemarker(vm_object_t marker, int swindex,
  89                                 vm_object_t object);
  90 struct thread *swapcached_thread;
  91
  92 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  93
  94 int vm_swapcache_read_enable;
  95 int vm_swapcache_inactive_heuristic;
  96 static int vm_swapcache_sleep;
  97 static int vm_swapcache_maxscan = PQ_L2_SIZE * 8;
  98 static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4;
  99 static int vm_swapcache_data_enable = 0;
 100 static int vm_swapcache_meta_enable = 0;
 101 static int vm_swapcache_maxswappct = 75;
 102 static int vm_swapcache_hysteresis;
 103 static int vm_swapcache_min_hysteresis;
 104 int vm_swapcache_use_chflags = 1;       /* require chflags cache */
 105 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 106 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 107 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 108 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 109 static int64_t vm_swapcache_write_count;
 110 static int64_t vm_swapcache_maxfilesize;
 111 static int64_t vm_swapcache_cleanperobj = 16*1024*1024;
 112
 113 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 114         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 115 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan,
 116         CTLFLAG_RW, &vm_swapcache_maxscan, 0, "");
 117
 118 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 119         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 120 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 121         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 122 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 123         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 124 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 125         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 126 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 127         CTLFLAG_RD, &vm_swapcache_hysteresis, 0, "");
 128 SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis,
 129         CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, "");
 130 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 131         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 132
 133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 134         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 135 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 136         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 137 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 138         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 139 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 140         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 141 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 142         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 143 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 144         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 145 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj,
 146         CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, "");
 147
 148 #define SWAPMAX(adj)    \
 149         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 150
 151 /*
 152  * When shutting down the machine we want to stop swapcache operation
 153  * immediately so swap is not accessed after devices have been shuttered.
 154  */
 155 static void
 156 shutdown_swapcache(void *arg __unused)
 157 {
 158         vm_swapcache_read_enable = 0;
 159         vm_swapcache_data_enable = 0;
 160         vm_swapcache_meta_enable = 0;
 161         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 162 }
 163
 164 /*
 165  * vm_swapcached is the high level pageout daemon.
 166  *
 167  * No requirements.
 168  */
 169 static void
 170 vm_swapcached_thread(void)
 171 {
 172         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 173         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 174         static struct vm_page page_marker[PQ_L2_SIZE];
 175         static struct vm_object swmarker;
 176         static int swindex;
 177         int q;
 178
 179         /*
 180          * Thread setup
 181          */
 182         curthread->td_flags |= TDF_SYSTHREAD;
 183         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 184                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 185         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 186                               NULL, SHUTDOWN_PRI_SECOND);
 187
 188         /*
 189          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 190          */
 191         bzero(&page_marker, sizeof(page_marker));
 192         for (q = 0; q < PQ_L2_SIZE; ++q) {
 193                 page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 194                 page_marker[q].queue = PQ_INACTIVE + q;
 195                 page_marker[q].pc = q;
 196                 page_marker[q].wire_count = 1;
 197                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 198                 TAILQ_INSERT_HEAD(
 199                         &vm_page_queues[PQ_INACTIVE + q].pl,
 200                         &page_marker[q], pageq);
 201                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 202         }
 203
 204         vm_swapcache_min_hysteresis = 1024;
 205         vm_swapcache_hysteresis = vm_swapcache_min_hysteresis;
 206         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 207
 208         /*
 209          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 210          */
 211         bzero(&swmarker, sizeof(swmarker));
 212         swmarker.type = OBJT_MARKER;
 213         swindex = 0;
 214         lwkt_gettoken(&vmobj_tokens[swindex]);
 215         TAILQ_INSERT_HEAD(&vm_object_lists[swindex],
 216                           &swmarker, object_list);
 217         lwkt_reltoken(&vmobj_tokens[swindex]);
 218
 219         for (;;) {
 220                 int reached_end;
 221                 int scount;
 222                 int count;
 223
 224                 /*
 225                  * Handle shutdown
 226                  */
 227                 kproc_suspend_loop();
 228
 229                 /*
 230                  * Check every 5 seconds when not enabled or if no swap
 231                  * is present.
 232                  */
 233                 if ((vm_swapcache_data_enable == 0 &&
 234                      vm_swapcache_meta_enable == 0) ||
 235                     vm_swap_max == 0) {
 236                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 237                         continue;
 238                 }
 239
 240                 /*
 241                  * Polling rate when enabled is approximately 10 hz.
 242                  */
 243                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 244
 245                 /*
 246                  * State hysteresis.  Generate write activity up to 75% of
 247                  * swap, then clean out swap assignments down to 70%, then
 248                  * repeat.
 249                  */
 250                 if (state == SWAPC_WRITING) {
 251                         if (vm_swap_cache_use > SWAPMAX(0))
 252                                 state = SWAPC_CLEANING;
 253                 } else {
 254                         if (vm_swap_cache_use < SWAPMAX(-10))
 255                                 state = SWAPC_WRITING;
 256                 }
 257
 258                 /*
 259                  * We are allowed to continue accumulating burst value
 260                  * in either state.  Allow the user to set curburst > maxburst
 261                  * for the initial load-in.
 262                  */
 263                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 264                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 265                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 266                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 267                 }
 268
 269                 /*
 270                  * We don't want to nickle-and-dime the scan as that will
 271                  * create unnecessary fragmentation.  The minimum burst
 272                  * is one-seconds worth of accumulation.
 273                  */
 274                 if (state != SWAPC_WRITING) {
 275                         vm_swapcache_cleaning(&swmarker, &swindex);
 276                         continue;
 277                 }
 278                 if (vm_swapcache_curburst < vm_swapcache_accrate)
 279                         continue;
 280
 281                 reached_end = 0;
 282                 count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2;
 283                 scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2;
 284
 285                 if (burst == SWAPB_BURSTING) {
 286                         if (vm_swapcache_writing_heuristic()) {
 287                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 288                                         reached_end +=
 289                                                 vm_swapcache_writing(
 290                                                         &page_marker[q],
 291                                                         count,
 292                                                         scount);
 293                                 }
 294                         }
 295                         if (vm_swapcache_curburst <= 0)
 296                                 burst = SWAPB_RECOVERING;
 297                 } else if (vm_swapcache_curburst > vm_swapcache_minburst) {
 298                         if (vm_swapcache_writing_heuristic()) {
 299                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 300                                         reached_end +=
 301                                                 vm_swapcache_writing(
 302                                                         &page_marker[q],
 303                                                         count,
 304                                                         scount);
 305                                 }
 306                         }
 307                         burst = SWAPB_BURSTING;
 308                 }
 309                 if (reached_end == PQ_L2_SIZE) {
 310                         vm_swapcache_inactive_heuristic =
 311                                 -vm_swapcache_hysteresis;
 312                 }
 313         }
 314
 315         /*
 316          * Cleanup (NOT REACHED)
 317          */
 318         for (q = 0; q < PQ_L2_SIZE; ++q) {
 319                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 320                 TAILQ_REMOVE(
 321                         &vm_page_queues[PQ_INACTIVE + q].pl,
 322                         &page_marker[q], pageq);
 323                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 324         }
 325
 326         lwkt_gettoken(&vmobj_tokens[swindex]);
 327         TAILQ_REMOVE(&vm_object_lists[swindex], &swmarker, object_list);
 328         lwkt_reltoken(&vmobj_tokens[swindex]);
 329 }
 330
 331 static struct kproc_desc swpc_kp = {
 332         "swapcached",
 333         vm_swapcached_thread,
 334         &swapcached_thread
 335 };
 336 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
 337
 338 /*
 339  * Deal with an overflow of the heuristic counter or if the user
 340  * manually changes the hysteresis.
 341  *
 342  * Try to avoid small incremental pageouts by waiting for enough
 343  * pages to buildup in the inactive queue to hopefully get a good
 344  * burst in.  This heuristic is bumped by the VM system and reset
 345  * when our scan hits the end of the queue.
 346  *
 347  * Return TRUE if we need to take a writing pass.
 348  */
 349 static int
 350 vm_swapcache_writing_heuristic(void)
 351 {
 352         int hyst;
 353
 354         hyst = vmstats.v_inactive_count / 4;
 355         if (hyst < vm_swapcache_min_hysteresis)
 356                 hyst = vm_swapcache_min_hysteresis;
 357         cpu_ccfence();
 358         vm_swapcache_hysteresis = hyst;
 359
 360         if (vm_swapcache_inactive_heuristic < -hyst)
 361                 vm_swapcache_inactive_heuristic = -hyst;
 362
 363         return (vm_swapcache_inactive_heuristic >= 0);
 364 }
 365
 366 /*
 367  * Take a writing pass on one of the inactive queues, return non-zero if
 368  * we hit the end of the queue.
 369  */
 370 static int
 371 vm_swapcache_writing(vm_page_t marker, int count, int scount)
 372 {
 373         vm_object_t object;
 374         struct vnode *vp;
 375         vm_page_t m;
 376         int isblkdev;
 377
 378         /*
 379          * Scan the inactive queue from our marker to locate
 380          * suitable pages to push to the swap cache.
 381          *
 382          * We are looking for clean vnode-backed pages.
 383          */
 384         vm_page_queues_spin_lock(marker->queue);
 385         while ((m = TAILQ_NEXT(marker, pageq)) != NULL &&
 386                count > 0 && scount-- > 0) {
 387                 KKASSERT(m->queue == marker->queue);
 388
 389                 if (vm_swapcache_curburst < 0)
 390                         break;
 391                 TAILQ_REMOVE(
 392                         &vm_page_queues[marker->queue].pl, marker, pageq);
 393                 TAILQ_INSERT_AFTER(
 394                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 395
 396                 /*
 397                  * Ignore markers and ignore pages that already have a swap
 398                  * assignment.
 399                  */
 400                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 401                         continue;
 402                 if (vm_page_busy_try(m, TRUE))
 403                         continue;
 404                 vm_page_queues_spin_unlock(marker->queue);
 405
 406                 if ((object = m->object) == NULL) {
 407                         vm_page_wakeup(m);
 408                         vm_page_queues_spin_lock(marker->queue);
 409                         continue;
 410                 }
 411                 vm_object_hold(object);
 412                 if (m->object != object) {
 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_test(m)) {
 419                         vm_object_drop(object);
 420                         vm_page_wakeup(m);
 421                         vm_page_queues_spin_lock(marker->queue);
 422                         continue;
 423                 }
 424
 425                 vp = object->handle;
 426                 if (vp == NULL) {
 427                         vm_object_drop(object);
 428                         vm_page_wakeup(m);
 429                         vm_page_queues_spin_lock(marker->queue);
 430                         continue;
 431                 }
 432
 433                 switch(vp->v_type) {
 434                 case VREG:
 435                         /*
 436                          * PG_NOTMETA generically means 'don't swapcache this',
 437                          * and HAMMER will set this for regular data buffers
 438                          * (and leave it unset for meta-data buffers) as
 439                          * appropriate when double buffering is enabled.
 440                          */
 441                         if (m->flags & PG_NOTMETA) {
 442                                 vm_object_drop(object);
 443                                 vm_page_wakeup(m);
 444                                 vm_page_queues_spin_lock(marker->queue);
 445                                 continue;
 446                         }
 447
 448                         /*
 449                          * If data_enable is 0 do not try to swapcache data.
 450                          * If use_chflags is set then only swapcache data for
 451                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 452                          */
 453                         if (vm_swapcache_data_enable == 0 ||
 454                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 455                              vm_swapcache_use_chflags)) {
 456                                 vm_object_drop(object);
 457                                 vm_page_wakeup(m);
 458                                 vm_page_queues_spin_lock(marker->queue);
 459                                 continue;
 460                         }
 461                         if (vm_swapcache_maxfilesize &&
 462                             object->size >
 463                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 464                                 vm_object_drop(object);
 465                                 vm_page_wakeup(m);
 466                                 vm_page_queues_spin_lock(marker->queue);
 467                                 continue;
 468                         }
 469                         isblkdev = 0;
 470                         break;
 471                 case VCHR:
 472                         /*
 473                          * PG_NOTMETA generically means 'don't swapcache this',
 474                          * and HAMMER will set this for regular data buffers
 475                          * (and leave it unset for meta-data buffers) as
 476                          * appropriate when double buffering is enabled.
 477                          */
 478                         if (m->flags & PG_NOTMETA) {
 479                                 vm_object_drop(object);
 480                                 vm_page_wakeup(m);
 481                                 vm_page_queues_spin_lock(marker->queue);
 482                                 continue;
 483                         }
 484                         if (vm_swapcache_meta_enable == 0) {
 485                                 vm_object_drop(object);
 486                                 vm_page_wakeup(m);
 487                                 vm_page_queues_spin_lock(marker->queue);
 488                                 continue;
 489                         }
 490                         isblkdev = 1;
 491                         break;
 492                 default:
 493                         vm_object_drop(object);
 494                         vm_page_wakeup(m);
 495                         vm_page_queues_spin_lock(marker->queue);
 496                         continue;
 497                 }
 498
 499
 500                 /*
 501                  * Assign swap and initiate I/O.
 502                  *
 503                  * (adjust for the --count which also occurs in the loop)
 504                  */
 505                 count -= vm_swapcached_flush(m, isblkdev);
 506
 507                 /*
 508                  * Setup for next loop using marker.
 509                  */
 510                 vm_object_drop(object);
 511                 vm_page_queues_spin_lock(marker->queue);
 512         }
 513
 514         /*
 515          * The marker could wind up at the end, which is ok.  If we hit the
 516          * end of the list adjust the heuristic.
 517          *
 518          * Earlier inactive pages that were dirty and become clean
 519          * are typically moved to the end of PQ_INACTIVE by virtue
 520          * of vfs_vmio_release() when they become unwired from the
 521          * buffer cache.
 522          */
 523         vm_page_queues_spin_unlock(marker->queue);
 524
 525         /*
 526          * m invalid but can be used to test for NULL
 527          */
 528         return (m == NULL);
 529 }
 530
 531 /*
 532  * Flush the specified page using the swap_pager.  The page
 533  * must be busied by the caller and its disposition will become
 534  * the responsibility of this function.
 535  *
 536  * Try to collect surrounding pages, including pages which may
 537  * have already been assigned swap.  Try to cluster within a
 538  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 539  * to match what swap_pager_putpages() can do.
 540  *
 541  * We also want to try to match against the buffer cache blocksize
 542  * but we don't really know what it is here.  Since the buffer cache
 543  * wires and unwires pages in groups the fact that we skip wired pages
 544  * should be sufficient.
 545  *
 546  * Returns a count of pages we might have flushed (minimum 1)
 547  */
 548 static
 549 int
 550 vm_swapcached_flush(vm_page_t m, int isblkdev)
 551 {
 552         vm_object_t object;
 553         vm_page_t marray[SWAP_META_PAGES];
 554         vm_pindex_t basei;
 555         int rtvals[SWAP_META_PAGES];
 556         int x;
 557         int i;
 558         int j;
 559         int count;
 560         int error;
 561
 562         vm_page_io_start(m);
 563         vm_page_protect(m, VM_PROT_READ);
 564         object = m->object;
 565         vm_object_hold(object);
 566
 567         /*
 568          * Try to cluster around (m), keeping in mind that the swap pager
 569          * can only do SMAP_META_PAGES worth of continguous write.
 570          */
 571         x = (int)m->pindex & SWAP_META_MASK;
 572         marray[x] = m;
 573         basei = m->pindex;
 574         vm_page_wakeup(m);
 575
 576         for (i = x - 1; i >= 0; --i) {
 577                 m = vm_page_lookup_busy_try(object, basei - x + i,
 578                                             TRUE, &error);
 579                 if (error || m == NULL)
 580                         break;
 581                 if (vm_swapcache_test(m)) {
 582                         vm_page_wakeup(m);
 583                         break;
 584                 }
 585                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 586                         vm_page_wakeup(m);
 587                         break;
 588                 }
 589                 vm_page_io_start(m);
 590                 vm_page_protect(m, VM_PROT_READ);
 591                 if (m->queue - m->pc == PQ_CACHE) {
 592                         vm_page_unqueue_nowakeup(m);
 593                         vm_page_deactivate(m);
 594                 }
 595                 marray[i] = m;
 596                 vm_page_wakeup(m);
 597         }
 598         ++i;
 599
 600         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 601                 m = vm_page_lookup_busy_try(object, basei - x + j,
 602                                             TRUE, &error);
 603                 if (error || m == NULL)
 604                         break;
 605                 if (vm_swapcache_test(m)) {
 606                         vm_page_wakeup(m);
 607                         break;
 608                 }
 609                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 610                         vm_page_wakeup(m);
 611                         break;
 612                 }
 613                 vm_page_io_start(m);
 614                 vm_page_protect(m, VM_PROT_READ);
 615                 if (m->queue - m->pc == PQ_CACHE) {
 616                         vm_page_unqueue_nowakeup(m);
 617                         vm_page_deactivate(m);
 618                 }
 619                 marray[j] = m;
 620                 vm_page_wakeup(m);
 621         }
 622
 623         count = j - i;
 624         vm_object_pip_add(object, count);
 625         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 626         vm_swapcache_write_count += count * PAGE_SIZE;
 627         vm_swapcache_curburst -= count * PAGE_SIZE;
 628
 629         while (i < j) {
 630                 if (rtvals[i] != VM_PAGER_PEND) {
 631                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 632                         vm_page_io_finish(marray[i]);
 633                         vm_page_wakeup(marray[i]);
 634                         vm_object_pip_wakeup(object);
 635                 }
 636                 ++i;
 637         }
 638         vm_object_drop(object);
 639         return(count);
 640 }
 641
 642 /*
 643  * Test whether a VM page is suitable for writing to the swapcache.
 644  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 645  *
 646  * Returns 0 on success, 1 on failure
 647  */
 648 static int
 649 vm_swapcache_test(vm_page_t m)
 650 {
 651         vm_object_t object;
 652
 653         if (m->flags & PG_UNMANAGED)
 654                 return(1);
 655         if (m->hold_count || m->wire_count)
 656                 return(1);
 657         if (m->valid != VM_PAGE_BITS_ALL)
 658                 return(1);
 659         if (m->dirty & m->valid)
 660                 return(1);
 661         if ((object = m->object) == NULL)
 662                 return(1);
 663         if (object->type != OBJT_VNODE ||
 664             (object->flags & OBJ_DEAD)) {
 665                 return(1);
 666         }
 667         vm_page_test_dirty(m);
 668         if (m->dirty & m->valid)
 669                 return(1);
 670         return(0);
 671 }
 672
 673 /*
 674  * Cleaning pass.
 675  *
 676  * We clean whole objects up to 16MB
 677  */
 678 static
 679 void
 680 vm_swapcache_cleaning(vm_object_t marker, int *swindexp)
 681 {
 682         vm_object_t object;
 683         struct vnode *vp;
 684         int count;
 685         int scount;
 686         int n;
 687
 688         count = vm_swapcache_maxlaunder;
 689         scount = vm_swapcache_maxscan;
 690
 691         /*
 692          * Look for vnode objects
 693          */
 694         lwkt_gettoken(&vmobj_tokens[*swindexp]);
 695
 696 outerloop:
 697         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 698                 /*
 699                  * We have to skip markers.  We cannot hold/drop marker
 700                  * objects!
 701                  */
 702                 if (object->type == OBJT_MARKER) {
 703                         vm_swapcache_movemarker(marker, *swindexp, object);
 704                         continue;
 705                 }
 706
 707                 /*
 708                  * Safety, or in case there are millions of VM objects
 709                  * without swapcache backing.
 710                  */
 711                 if (--scount <= 0)
 712                         goto breakout;
 713
 714                 /*
 715                  * We must hold the object before potentially yielding.
 716                  */
 717                 vm_object_hold(object);
 718                 lwkt_yield();
 719
 720                 /*
 721                  * Only operate on live VNODE objects that are either
 722                  * VREG or VCHR (VCHR for meta-data).
 723                  */
 724                 if ((object->type != OBJT_VNODE) ||
 725                     ((object->flags & OBJ_DEAD) ||
 726                      object->swblock_count == 0) ||
 727                     ((vp = object->handle) == NULL) ||
 728                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 729                         vm_object_drop(object);
 730                         /* object may be invalid now */
 731                         vm_swapcache_movemarker(marker, *swindexp, object);
 732                         continue;
 733                 }
 734
 735                 /*
 736                  * Reset the object pindex stored in the marker if the
 737                  * working object has changed.
 738                  */
 739                 if (marker->backing_object != object) {
 740                         marker->size = 0;
 741                         marker->backing_object_offset = 0;
 742                         marker->backing_object = object;
 743                 }
 744
 745                 /*
 746                  * Look for swblocks starting at our iterator.
 747                  *
 748                  * The swap_pager_condfree() function attempts to free
 749                  * swap space starting at the specified index.  The index
 750                  * will be updated on return.  The function will return
 751                  * a scan factor (NOT the number of blocks freed).
 752                  *
 753                  * If it must cut its scan of the object short due to an
 754                  * excessive number of swblocks, or is able to free the
 755                  * requested number of blocks, it will return n >= count
 756                  * and we break and pick it back up on a future attempt.
 757                  *
 758                  * Scan the object linearly and try to batch large sets of
 759                  * blocks that are likely to clean out entire swap radix
 760                  * tree leafs.
 761                  */
 762                 lwkt_token_swap();
 763                 lwkt_reltoken(&vmobj_tokens[*swindexp]);
 764
 765                 n = swap_pager_condfree(object, &marker->size,
 766                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 767
 768                 vm_object_drop(object);         /* object may be invalid now */
 769                 lwkt_gettoken(&vmobj_tokens[*swindexp]);
 770
 771                 /*
 772                  * If we have exhausted the object or deleted our per-pass
 773                  * page limit then move us to the next object.  Note that
 774                  * the current object may no longer be on the vm_object_list.
 775                  */
 776                 if (n <= 0 ||
 777                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 778                         vm_swapcache_movemarker(marker, *swindexp, object);
 779                 }
 780
 781                 /*
 782                  * If we have exhausted our max-launder stop for now.
 783                  */
 784                 count -= n;
 785                 marker->backing_object_offset += n * PAGE_SIZE;
 786                 if (count < 0)
 787                         goto breakout;
 788         }
 789
 790         /*
 791          * Iterate vm_object_lists[] hash table
 792          */
 793         TAILQ_REMOVE(&vm_object_lists[*swindexp], marker, object_list);
 794         lwkt_reltoken(&vmobj_tokens[*swindexp]);
 795         if (++*swindexp >= VMOBJ_HSIZE)
 796                 *swindexp = 0;
 797         lwkt_gettoken(&vmobj_tokens[*swindexp]);
 798         TAILQ_INSERT_HEAD(&vm_object_lists[*swindexp], marker, object_list);
 799
 800         if (*swindexp != 0)
 801                 goto outerloop;
 802
 803 breakout:
 804         lwkt_reltoken(&vmobj_tokens[*swindexp]);
 805 }
 806
 807 /*
 808  * Move the marker past the current object.  Object can be stale, but we
 809  * still need it to determine if the marker has to be moved.  If the object
 810  * is still the 'current object' (object after the marker), we hop-scotch
 811  * the marker past it.
 812  */
 813 static void
 814 vm_swapcache_movemarker(vm_object_t marker, int swindex, vm_object_t object)
 815 {
 816         if (TAILQ_NEXT(marker, object_list) == object) {
 817                 TAILQ_REMOVE(&vm_object_lists[swindex], marker, object_list);
 818                 TAILQ_INSERT_AFTER(&vm_object_lists[swindex], object,
 819                                    marker, object_list);
 820         }
 821 }