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_cache_use <= SWAPMAX(0)) ||
 236                     vm_swap_max == 0) {
 237                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 238                         continue;
 239                 }
 240
 241                 /*
 242                  * Polling rate when enabled is approximately 10 hz.
 243                  */
 244                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 245
 246                 /*
 247                  * State hysteresis.  Generate write activity up to 75% of
 248                  * swap, then clean out swap assignments down to 70%, then
 249                  * repeat.
 250                  */
 251                 if (state == SWAPC_WRITING) {
 252                         if (vm_swap_cache_use > SWAPMAX(0))
 253                                 state = SWAPC_CLEANING;
 254                 } else {
 255                         if (vm_swap_cache_use < SWAPMAX(-10))
 256                                 state = SWAPC_WRITING;
 257                 }
 258
 259                 /*
 260                  * We are allowed to continue accumulating burst value
 261                  * in either state.  Allow the user to set curburst > maxburst
 262                  * for the initial load-in.
 263                  */
 264                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 265                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 266                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 267                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 268                 }
 269
 270                 /*
 271                  * We don't want to nickle-and-dime the scan as that will
 272                  * create unnecessary fragmentation.  The minimum burst
 273                  * is one-seconds worth of accumulation.
 274                  */
 275                 if (state != SWAPC_WRITING) {
 276                         vm_swapcache_cleaning(&swmarker, &swindex);
 277                         continue;
 278                 }
 279                 if (vm_swapcache_curburst < vm_swapcache_accrate)
 280                         continue;
 281
 282                 reached_end = 0;
 283                 count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2;
 284                 scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2;
 285
 286                 if (burst == SWAPB_BURSTING) {
 287                         if (vm_swapcache_writing_heuristic()) {
 288                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 289                                         reached_end +=
 290                                                 vm_swapcache_writing(
 291                                                         &page_marker[q],
 292                                                         count,
 293                                                         scount);
 294                                 }
 295                         }
 296                         if (vm_swapcache_curburst <= 0)
 297                                 burst = SWAPB_RECOVERING;
 298                 } else if (vm_swapcache_curburst > vm_swapcache_minburst) {
 299                         if (vm_swapcache_writing_heuristic()) {
 300                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 301                                         reached_end +=
 302                                                 vm_swapcache_writing(
 303                                                         &page_marker[q],
 304                                                         count,
 305                                                         scount);
 306                                 }
 307                         }
 308                         burst = SWAPB_BURSTING;
 309                 }
 310                 if (reached_end == PQ_L2_SIZE) {
 311                         vm_swapcache_inactive_heuristic =
 312                                 -vm_swapcache_hysteresis;
 313                 }
 314         }
 315
 316         /*
 317          * Cleanup (NOT REACHED)
 318          */
 319         for (q = 0; q < PQ_L2_SIZE; ++q) {
 320                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 321                 TAILQ_REMOVE(
 322                         &vm_page_queues[PQ_INACTIVE + q].pl,
 323                         &page_marker[q], pageq);
 324                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 325         }
 326
 327         lwkt_gettoken(&vmobj_tokens[swindex]);
 328         TAILQ_REMOVE(&vm_object_lists[swindex], &swmarker, object_list);
 329         lwkt_reltoken(&vmobj_tokens[swindex]);
 330 }
 331
 332 static struct kproc_desc swpc_kp = {
 333         "swapcached",
 334         vm_swapcached_thread,
 335         &swapcached_thread
 336 };
 337 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
 338
 339 /*
 340  * Deal with an overflow of the heuristic counter or if the user
 341  * manually changes the hysteresis.
 342  *
 343  * Try to avoid small incremental pageouts by waiting for enough
 344  * pages to buildup in the inactive queue to hopefully get a good
 345  * burst in.  This heuristic is bumped by the VM system and reset
 346  * when our scan hits the end of the queue.
 347  *
 348  * Return TRUE if we need to take a writing pass.
 349  */
 350 static int
 351 vm_swapcache_writing_heuristic(void)
 352 {
 353         int hyst;
 354
 355         hyst = vmstats.v_inactive_count / 4;
 356         if (hyst < vm_swapcache_min_hysteresis)
 357                 hyst = vm_swapcache_min_hysteresis;
 358         cpu_ccfence();
 359         vm_swapcache_hysteresis = hyst;
 360
 361         if (vm_swapcache_inactive_heuristic < -hyst)
 362                 vm_swapcache_inactive_heuristic = -hyst;
 363
 364         return (vm_swapcache_inactive_heuristic >= 0);
 365 }
 366
 367 /*
 368  * Take a writing pass on one of the inactive queues, return non-zero if
 369  * we hit the end of the queue.
 370  */
 371 static int
 372 vm_swapcache_writing(vm_page_t marker, int count, int scount)
 373 {
 374         vm_object_t object;
 375         struct vnode *vp;
 376         vm_page_t m;
 377         int isblkdev;
 378
 379         /*
 380          * Scan the inactive queue from our marker to locate
 381          * suitable pages to push to the swap cache.
 382          *
 383          * We are looking for clean vnode-backed pages.
 384          */
 385         vm_page_queues_spin_lock(marker->queue);
 386         while ((m = TAILQ_NEXT(marker, pageq)) != NULL &&
 387                count > 0 && scount-- > 0) {
 388                 KKASSERT(m->queue == marker->queue);
 389
 390                 if (vm_swapcache_curburst < 0)
 391                         break;
 392                 TAILQ_REMOVE(
 393                         &vm_page_queues[marker->queue].pl, marker, pageq);
 394                 TAILQ_INSERT_AFTER(
 395                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 396
 397                 /*
 398                  * Ignore markers and ignore pages that already have a swap
 399                  * assignment.
 400                  */
 401                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 402                         continue;
 403                 if (vm_page_busy_try(m, TRUE))
 404                         continue;
 405                 vm_page_queues_spin_unlock(marker->queue);
 406
 407                 if ((object = m->object) == NULL) {
 408                         vm_page_wakeup(m);
 409                         vm_page_queues_spin_lock(marker->queue);
 410                         continue;
 411                 }
 412                 vm_object_hold(object);
 413                 if (m->object != object) {
 414                         vm_object_drop(object);
 415                         vm_page_wakeup(m);
 416                         vm_page_queues_spin_lock(marker->queue);
 417                         continue;
 418                 }
 419                 if (vm_swapcache_test(m)) {
 420                         vm_object_drop(object);
 421                         vm_page_wakeup(m);
 422                         vm_page_queues_spin_lock(marker->queue);
 423                         continue;
 424                 }
 425
 426                 vp = object->handle;
 427                 if (vp == NULL) {
 428                         vm_object_drop(object);
 429                         vm_page_wakeup(m);
 430                         vm_page_queues_spin_lock(marker->queue);
 431                         continue;
 432                 }
 433
 434                 switch(vp->v_type) {
 435                 case VREG:
 436                         /*
 437                          * PG_NOTMETA generically means 'don't swapcache this',
 438                          * and HAMMER will set this for regular data buffers
 439                          * (and leave it unset for meta-data buffers) as
 440                          * appropriate when double buffering is enabled.
 441                          */
 442                         if (m->flags & PG_NOTMETA) {
 443                                 vm_object_drop(object);
 444                                 vm_page_wakeup(m);
 445                                 vm_page_queues_spin_lock(marker->queue);
 446                                 continue;
 447                         }
 448
 449                         /*
 450                          * If data_enable is 0 do not try to swapcache data.
 451                          * If use_chflags is set then only swapcache data for
 452                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 453                          */
 454                         if (vm_swapcache_data_enable == 0 ||
 455                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 456                              vm_swapcache_use_chflags)) {
 457                                 vm_object_drop(object);
 458                                 vm_page_wakeup(m);
 459                                 vm_page_queues_spin_lock(marker->queue);
 460                                 continue;
 461                         }
 462                         if (vm_swapcache_maxfilesize &&
 463                             object->size >
 464                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 465                                 vm_object_drop(object);
 466                                 vm_page_wakeup(m);
 467                                 vm_page_queues_spin_lock(marker->queue);
 468                                 continue;
 469                         }
 470                         isblkdev = 0;
 471                         break;
 472                 case VCHR:
 473                         /*
 474                          * PG_NOTMETA generically means 'don't swapcache this',
 475                          * and HAMMER will set this for regular data buffers
 476                          * (and leave it unset for meta-data buffers) as
 477                          * appropriate when double buffering is enabled.
 478                          */
 479                         if (m->flags & PG_NOTMETA) {
 480                                 vm_object_drop(object);
 481                                 vm_page_wakeup(m);
 482                                 vm_page_queues_spin_lock(marker->queue);
 483                                 continue;
 484                         }
 485                         if (vm_swapcache_meta_enable == 0) {
 486                                 vm_object_drop(object);
 487                                 vm_page_wakeup(m);
 488                                 vm_page_queues_spin_lock(marker->queue);
 489                                 continue;
 490                         }
 491                         isblkdev = 1;
 492                         break;
 493                 default:
 494                         vm_object_drop(object);
 495                         vm_page_wakeup(m);
 496                         vm_page_queues_spin_lock(marker->queue);
 497                         continue;
 498                 }
 499
 500
 501                 /*
 502                  * Assign swap and initiate I/O.
 503                  *
 504                  * (adjust for the --count which also occurs in the loop)
 505                  */
 506                 count -= vm_swapcached_flush(m, isblkdev);
 507
 508                 /*
 509                  * Setup for next loop using marker.
 510                  */
 511                 vm_object_drop(object);
 512                 vm_page_queues_spin_lock(marker->queue);
 513         }
 514
 515         /*
 516          * The marker could wind up at the end, which is ok.  If we hit the
 517          * end of the list adjust the heuristic.
 518          *
 519          * Earlier inactive pages that were dirty and become clean
 520          * are typically moved to the end of PQ_INACTIVE by virtue
 521          * of vfs_vmio_release() when they become unwired from the
 522          * buffer cache.
 523          */
 524         vm_page_queues_spin_unlock(marker->queue);
 525
 526         /*
 527          * m invalid but can be used to test for NULL
 528          */
 529         return (m == NULL);
 530 }
 531
 532 /*
 533  * Flush the specified page using the swap_pager.  The page
 534  * must be busied by the caller and its disposition will become
 535  * the responsibility of this function.
 536  *
 537  * Try to collect surrounding pages, including pages which may
 538  * have already been assigned swap.  Try to cluster within a
 539  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 540  * to match what swap_pager_putpages() can do.
 541  *
 542  * We also want to try to match against the buffer cache blocksize
 543  * but we don't really know what it is here.  Since the buffer cache
 544  * wires and unwires pages in groups the fact that we skip wired pages
 545  * should be sufficient.
 546  *
 547  * Returns a count of pages we might have flushed (minimum 1)
 548  */
 549 static
 550 int
 551 vm_swapcached_flush(vm_page_t m, int isblkdev)
 552 {
 553         vm_object_t object;
 554         vm_page_t marray[SWAP_META_PAGES];
 555         vm_pindex_t basei;
 556         int rtvals[SWAP_META_PAGES];
 557         int x;
 558         int i;
 559         int j;
 560         int count;
 561         int error;
 562
 563         vm_page_io_start(m);
 564         vm_page_protect(m, VM_PROT_READ);
 565         object = m->object;
 566         vm_object_hold(object);
 567
 568         /*
 569          * Try to cluster around (m), keeping in mind that the swap pager
 570          * can only do SMAP_META_PAGES worth of continguous write.
 571          */
 572         x = (int)m->pindex & SWAP_META_MASK;
 573         marray[x] = m;
 574         basei = m->pindex;
 575         vm_page_wakeup(m);
 576
 577         for (i = x - 1; i >= 0; --i) {
 578                 m = vm_page_lookup_busy_try(object, basei - x + i,
 579                                             TRUE, &error);
 580                 if (error || m == NULL)
 581                         break;
 582                 if (vm_swapcache_test(m)) {
 583                         vm_page_wakeup(m);
 584                         break;
 585                 }
 586                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 587                         vm_page_wakeup(m);
 588                         break;
 589                 }
 590                 vm_page_io_start(m);
 591                 vm_page_protect(m, VM_PROT_READ);
 592                 if (m->queue - m->pc == PQ_CACHE) {
 593                         vm_page_unqueue_nowakeup(m);
 594                         vm_page_deactivate(m);
 595                 }
 596                 marray[i] = m;
 597                 vm_page_wakeup(m);
 598         }
 599         ++i;
 600
 601         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 602                 m = vm_page_lookup_busy_try(object, basei - x + j,
 603                                             TRUE, &error);
 604                 if (error || m == NULL)
 605                         break;
 606                 if (vm_swapcache_test(m)) {
 607                         vm_page_wakeup(m);
 608                         break;
 609                 }
 610                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 611                         vm_page_wakeup(m);
 612                         break;
 613                 }
 614                 vm_page_io_start(m);
 615                 vm_page_protect(m, VM_PROT_READ);
 616                 if (m->queue - m->pc == PQ_CACHE) {
 617                         vm_page_unqueue_nowakeup(m);
 618                         vm_page_deactivate(m);
 619                 }
 620                 marray[j] = m;
 621                 vm_page_wakeup(m);
 622         }
 623
 624         count = j - i;
 625         vm_object_pip_add(object, count);
 626         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 627         vm_swapcache_write_count += count * PAGE_SIZE;
 628         vm_swapcache_curburst -= count * PAGE_SIZE;
 629
 630         while (i < j) {
 631                 if (rtvals[i] != VM_PAGER_PEND) {
 632                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 633                         vm_page_io_finish(marray[i]);
 634                         vm_page_wakeup(marray[i]);
 635                         vm_object_pip_wakeup(object);
 636                 }
 637                 ++i;
 638         }
 639         vm_object_drop(object);
 640         return(count);
 641 }
 642
 643 /*
 644  * Test whether a VM page is suitable for writing to the swapcache.
 645  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 646  *
 647  * Returns 0 on success, 1 on failure
 648  */
 649 static int
 650 vm_swapcache_test(vm_page_t m)
 651 {
 652         vm_object_t object;
 653
 654         if (m->flags & PG_UNMANAGED)
 655                 return(1);
 656         if (m->hold_count || m->wire_count)
 657                 return(1);
 658         if (m->valid != VM_PAGE_BITS_ALL)
 659                 return(1);
 660         if (m->dirty & m->valid)
 661                 return(1);
 662         if ((object = m->object) == NULL)
 663                 return(1);
 664         if (object->type != OBJT_VNODE ||
 665             (object->flags & OBJ_DEAD)) {
 666                 return(1);
 667         }
 668         vm_page_test_dirty(m);
 669         if (m->dirty & m->valid)
 670                 return(1);
 671         return(0);
 672 }
 673
 674 /*
 675  * Cleaning pass.
 676  *
 677  * We clean whole objects up to 16MB
 678  */
 679 static
 680 void
 681 vm_swapcache_cleaning(vm_object_t marker, int *swindexp)
 682 {
 683         vm_object_t object;
 684         struct vnode *vp;
 685         int count;
 686         int scount;
 687         int n;
 688
 689         count = vm_swapcache_maxlaunder;
 690         scount = vm_swapcache_maxscan;
 691
 692         /*
 693          * Look for vnode objects
 694          */
 695         lwkt_gettoken(&vmobj_tokens[*swindexp]);
 696
 697 outerloop:
 698         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 699                 /*
 700                  * We have to skip markers.  We cannot hold/drop marker
 701                  * objects!
 702                  */
 703                 if (object->type == OBJT_MARKER) {
 704                         vm_swapcache_movemarker(marker, *swindexp, object);
 705                         continue;
 706                 }
 707
 708                 /*
 709                  * Safety, or in case there are millions of VM objects
 710                  * without swapcache backing.
 711                  */
 712                 if (--scount <= 0)
 713                         goto breakout;
 714
 715                 /*
 716                  * We must hold the object before potentially yielding.
 717                  */
 718                 vm_object_hold(object);
 719                 lwkt_yield();
 720
 721                 /*
 722                  * Only operate on live VNODE objects that are either
 723                  * VREG or VCHR (VCHR for meta-data).
 724                  */
 725                 if ((object->type != OBJT_VNODE) ||
 726                     ((object->flags & OBJ_DEAD) ||
 727                      object->swblock_count == 0) ||
 728                     ((vp = object->handle) == NULL) ||
 729                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 730                         vm_object_drop(object);
 731                         /* object may be invalid now */
 732                         vm_swapcache_movemarker(marker, *swindexp, object);
 733                         continue;
 734                 }
 735
 736                 /*
 737                  * Reset the object pindex stored in the marker if the
 738                  * working object has changed.
 739                  */
 740                 if (marker->backing_object != object) {
 741                         marker->size = 0;
 742                         marker->backing_object_offset = 0;
 743                         marker->backing_object = object;
 744                 }
 745
 746                 /*
 747                  * Look for swblocks starting at our iterator.
 748                  *
 749                  * The swap_pager_condfree() function attempts to free
 750                  * swap space starting at the specified index.  The index
 751                  * will be updated on return.  The function will return
 752                  * a scan factor (NOT the number of blocks freed).
 753                  *
 754                  * If it must cut its scan of the object short due to an
 755                  * excessive number of swblocks, or is able to free the
 756                  * requested number of blocks, it will return n >= count
 757                  * and we break and pick it back up on a future attempt.
 758                  *
 759                  * Scan the object linearly and try to batch large sets of
 760                  * blocks that are likely to clean out entire swap radix
 761                  * tree leafs.
 762                  */
 763                 lwkt_token_swap();
 764                 lwkt_reltoken(&vmobj_tokens[*swindexp]);
 765
 766                 n = swap_pager_condfree(object, &marker->size,
 767                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 768
 769                 vm_object_drop(object);         /* object may be invalid now */
 770                 lwkt_gettoken(&vmobj_tokens[*swindexp]);
 771
 772                 /*
 773                  * If we have exhausted the object or deleted our per-pass
 774                  * page limit then move us to the next object.  Note that
 775                  * the current object may no longer be on the vm_object_list.
 776                  */
 777                 if (n <= 0 ||
 778                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 779                         vm_swapcache_movemarker(marker, *swindexp, object);
 780                 }
 781
 782                 /*
 783                  * If we have exhausted our max-launder stop for now.
 784                  */
 785                 count -= n;
 786                 marker->backing_object_offset += n * PAGE_SIZE;
 787                 if (count < 0)
 788                         goto breakout;
 789         }
 790
 791         /*
 792          * Iterate vm_object_lists[] hash table
 793          */
 794         TAILQ_REMOVE(&vm_object_lists[*swindexp], marker, object_list);
 795         lwkt_reltoken(&vmobj_tokens[*swindexp]);
 796         if (++*swindexp >= VMOBJ_HSIZE)
 797                 *swindexp = 0;
 798         lwkt_gettoken(&vmobj_tokens[*swindexp]);
 799         TAILQ_INSERT_HEAD(&vm_object_lists[*swindexp], marker, object_list);
 800
 801         if (*swindexp != 0)
 802                 goto outerloop;
 803
 804 breakout:
 805         lwkt_reltoken(&vmobj_tokens[*swindexp]);
 806 }
 807
 808 /*
 809  * Move the marker past the current object.  Object can be stale, but we
 810  * still need it to determine if the marker has to be moved.  If the object
 811  * is still the 'current object' (object after the marker), we hop-scotch
 812  * the marker past it.
 813  */
 814 static void
 815 vm_swapcache_movemarker(vm_object_t marker, int swindex, vm_object_t object)
 816 {
 817         if (TAILQ_NEXT(marker, object_list) == object) {
 818                 TAILQ_REMOVE(&vm_object_lists[swindex], marker, object_list);
 819                 TAILQ_INSERT_AFTER(&vm_object_lists[swindex], object,
 820                                    marker, object_list);
 821         }
 822 }