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);
  88 static void vm_swapcache_movemarker(vm_object_t marker, vm_object_t object);
  89 struct thread *swapcached_thread;
  90
  91 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  92
  93 int vm_swapcache_read_enable;
  94 int vm_swapcache_inactive_heuristic;
  95 static int vm_swapcache_sleep;
  96 static int vm_swapcache_maxscan = PQ_L2_SIZE * 8;
  97 static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4;
  98 static int vm_swapcache_data_enable = 0;
  99 static int vm_swapcache_meta_enable = 0;
 100 static int vm_swapcache_maxswappct = 75;
 101 static int vm_swapcache_hysteresis;
 102 static int vm_swapcache_min_hysteresis;
 103 int vm_swapcache_use_chflags = 1;       /* require chflags cache */
 104 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 105 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 106 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 107 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 108 static int64_t vm_swapcache_write_count;
 109 static int64_t vm_swapcache_maxfilesize;
 110 static int64_t vm_swapcache_cleanperobj = 16*1024*1024;
 111
 112 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 113         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 114 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan,
 115         CTLFLAG_RW, &vm_swapcache_maxscan, 0, "");
 116
 117 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 118         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 119 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 120         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 121 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 122         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 123 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 124         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 125 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 126         CTLFLAG_RD, &vm_swapcache_hysteresis, 0, "");
 127 SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis,
 128         CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, "");
 129 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 130         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 131
 132 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 133         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 134 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 135         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 136 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 137         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 138 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 139         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 140 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 141         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 142 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 143         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 144 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj,
 145         CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, "");
 146
 147 #define SWAPMAX(adj)    \
 148         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 149
 150 /*
 151  * When shutting down the machine we want to stop swapcache operation
 152  * immediately so swap is not accessed after devices have been shuttered.
 153  */
 154 static void
 155 shutdown_swapcache(void *arg __unused)
 156 {
 157         vm_swapcache_read_enable = 0;
 158         vm_swapcache_data_enable = 0;
 159         vm_swapcache_meta_enable = 0;
 160         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 161 }
 162
 163 /*
 164  * vm_swapcached is the high level pageout daemon.
 165  *
 166  * No requirements.
 167  */
 168 static void
 169 vm_swapcached_thread(void)
 170 {
 171         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 172         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 173         static struct vm_page page_marker[PQ_L2_SIZE];
 174         static struct vm_object object_marker;
 175         int q;
 176
 177         /*
 178          * Thread setup
 179          */
 180         curthread->td_flags |= TDF_SYSTHREAD;
 181         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 182                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 183         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 184                               NULL, SHUTDOWN_PRI_SECOND);
 185
 186         /*
 187          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 188          */
 189         bzero(&page_marker, sizeof(page_marker));
 190         for (q = 0; q < PQ_L2_SIZE; ++q) {
 191                 page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 192                 page_marker[q].queue = PQ_INACTIVE + q;
 193                 page_marker[q].pc = q;
 194                 page_marker[q].wire_count = 1;
 195                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 196                 TAILQ_INSERT_HEAD(
 197                         &vm_page_queues[PQ_INACTIVE + q].pl,
 198                         &page_marker[q], pageq);
 199                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 200         }
 201
 202         vm_swapcache_min_hysteresis = 1024;
 203         vm_swapcache_hysteresis = vm_swapcache_min_hysteresis;
 204         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 205
 206         /*
 207          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 208          */
 209         bzero(&object_marker, sizeof(object_marker));
 210         object_marker.type = OBJT_MARKER;
 211         lwkt_gettoken(&vmobj_token);
 212         TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
 213         lwkt_reltoken(&vmobj_token);
 214
 215         for (;;) {
 216                 int reached_end;
 217                 int scount;
 218                 int count;
 219
 220                 /*
 221                  * Handle shutdown
 222                  */
 223                 kproc_suspend_loop();
 224
 225                 /*
 226                  * Check every 5 seconds when not enabled or if no swap
 227                  * is present.
 228                  */
 229                 if ((vm_swapcache_data_enable == 0 &&
 230                      vm_swapcache_meta_enable == 0) ||
 231                     vm_swap_max == 0) {
 232                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 233                         continue;
 234                 }
 235
 236                 /*
 237                  * Polling rate when enabled is approximately 10 hz.
 238                  */
 239                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 240
 241                 /*
 242                  * State hysteresis.  Generate write activity up to 75% of
 243                  * swap, then clean out swap assignments down to 70%, then
 244                  * repeat.
 245                  */
 246                 if (state == SWAPC_WRITING) {
 247                         if (vm_swap_cache_use > SWAPMAX(0))
 248                                 state = SWAPC_CLEANING;
 249                 } else {
 250                         if (vm_swap_cache_use < SWAPMAX(-10))
 251                                 state = SWAPC_WRITING;
 252                 }
 253
 254                 /*
 255                  * We are allowed to continue accumulating burst value
 256                  * in either state.  Allow the user to set curburst > maxburst
 257                  * for the initial load-in.
 258                  */
 259                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 260                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 261                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 262                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 263                 }
 264
 265                 /*
 266                  * We don't want to nickle-and-dime the scan as that will
 267                  * create unnecessary fragmentation.  The minimum burst
 268                  * is one-seconds worth of accumulation.
 269                  */
 270                 if (state != SWAPC_WRITING) {
 271                         vm_swapcache_cleaning(&object_marker);
 272                         continue;
 273                 }
 274                 if (vm_swapcache_curburst < vm_swapcache_accrate)
 275                         continue;
 276
 277                 reached_end = 0;
 278                 count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2;
 279                 scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2;
 280
 281                 if (burst == SWAPB_BURSTING) {
 282                         if (vm_swapcache_writing_heuristic()) {
 283                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 284                                         reached_end +=
 285                                                 vm_swapcache_writing(
 286                                                         &page_marker[q],
 287                                                         count,
 288                                                         scount);
 289                                 }
 290                         }
 291                         if (vm_swapcache_curburst <= 0)
 292                                 burst = SWAPB_RECOVERING;
 293                 } else if (vm_swapcache_curburst > vm_swapcache_minburst) {
 294                         if (vm_swapcache_writing_heuristic()) {
 295                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 296                                         reached_end +=
 297                                                 vm_swapcache_writing(
 298                                                         &page_marker[q],
 299                                                         count,
 300                                                         scount);
 301                                 }
 302                         }
 303                         burst = SWAPB_BURSTING;
 304                 }
 305                 if (reached_end == PQ_L2_SIZE) {
 306                         vm_swapcache_inactive_heuristic =
 307                                 -vm_swapcache_hysteresis;
 308                 }
 309         }
 310
 311         /*
 312          * Cleanup (NOT REACHED)
 313          */
 314         for (q = 0; q < PQ_L2_SIZE; ++q) {
 315                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 316                 TAILQ_REMOVE(
 317                         &vm_page_queues[PQ_INACTIVE + q].pl,
 318                         &page_marker[q], pageq);
 319                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 320         }
 321
 322         lwkt_gettoken(&vmobj_token);
 323         TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
 324         lwkt_reltoken(&vmobj_token);
 325 }
 326
 327 static struct kproc_desc swpc_kp = {
 328         "swapcached",
 329         vm_swapcached_thread,
 330         &swapcached_thread
 331 };
 332 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
 333
 334 /*
 335  * Deal with an overflow of the heuristic counter or if the user
 336  * manually changes the hysteresis.
 337  *
 338  * Try to avoid small incremental pageouts by waiting for enough
 339  * pages to buildup in the inactive queue to hopefully get a good
 340  * burst in.  This heuristic is bumped by the VM system and reset
 341  * when our scan hits the end of the queue.
 342  *
 343  * Return TRUE if we need to take a writing pass.
 344  */
 345 static int
 346 vm_swapcache_writing_heuristic(void)
 347 {
 348         int hyst;
 349
 350         hyst = vmstats.v_inactive_count / 4;
 351         if (hyst < vm_swapcache_min_hysteresis)
 352                 hyst = vm_swapcache_min_hysteresis;
 353         cpu_ccfence();
 354         vm_swapcache_hysteresis = hyst;
 355
 356         if (vm_swapcache_inactive_heuristic < -hyst)
 357                 vm_swapcache_inactive_heuristic = -hyst;
 358
 359         return (vm_swapcache_inactive_heuristic >= 0);
 360 }
 361
 362 /*
 363  * Take a writing pass on one of the inactive queues, return non-zero if
 364  * we hit the end of the queue.
 365  */
 366 static int
 367 vm_swapcache_writing(vm_page_t marker, int count, int scount)
 368 {
 369         vm_object_t object;
 370         struct vnode *vp;
 371         vm_page_t m;
 372         int isblkdev;
 373
 374         /*
 375          * Scan the inactive queue from our marker to locate
 376          * suitable pages to push to the swap cache.
 377          *
 378          * We are looking for clean vnode-backed pages.
 379          */
 380         vm_page_queues_spin_lock(marker->queue);
 381         while ((m = TAILQ_NEXT(marker, pageq)) != NULL &&
 382                count > 0 && scount-- > 0) {
 383                 KKASSERT(m->queue == marker->queue);
 384
 385                 if (vm_swapcache_curburst < 0)
 386                         break;
 387                 TAILQ_REMOVE(
 388                         &vm_page_queues[marker->queue].pl, marker, pageq);
 389                 TAILQ_INSERT_AFTER(
 390                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 391
 392                 /*
 393                  * Ignore markers and ignore pages that already have a swap
 394                  * assignment.
 395                  */
 396                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 397                         continue;
 398                 if (vm_page_busy_try(m, TRUE))
 399                         continue;
 400                 vm_page_queues_spin_unlock(marker->queue);
 401
 402                 if ((object = m->object) == NULL) {
 403                         vm_page_wakeup(m);
 404                         vm_page_queues_spin_lock(marker->queue);
 405                         continue;
 406                 }
 407                 vm_object_hold(object);
 408                 if (m->object != object) {
 409                         vm_object_drop(object);
 410                         vm_page_wakeup(m);
 411                         vm_page_queues_spin_lock(marker->queue);
 412                         continue;
 413                 }
 414                 if (vm_swapcache_test(m)) {
 415                         vm_object_drop(object);
 416                         vm_page_wakeup(m);
 417                         vm_page_queues_spin_lock(marker->queue);
 418                         continue;
 419                 }
 420
 421                 vp = object->handle;
 422                 if (vp == NULL) {
 423                         vm_object_drop(object);
 424                         vm_page_wakeup(m);
 425                         vm_page_queues_spin_lock(marker->queue);
 426                         continue;
 427                 }
 428
 429                 switch(vp->v_type) {
 430                 case VREG:
 431                         /*
 432                          * PG_NOTMETA generically means 'don't swapcache this',
 433                          * and HAMMER will set this for regular data buffers
 434                          * (and leave it unset for meta-data buffers) as
 435                          * appropriate when double buffering is enabled.
 436                          */
 437                         if (m->flags & PG_NOTMETA) {
 438                                 vm_object_drop(object);
 439                                 vm_page_wakeup(m);
 440                                 vm_page_queues_spin_lock(marker->queue);
 441                                 continue;
 442                         }
 443
 444                         /*
 445                          * If data_enable is 0 do not try to swapcache data.
 446                          * If use_chflags is set then only swapcache data for
 447                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 448                          */
 449                         if (vm_swapcache_data_enable == 0 ||
 450                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 451                              vm_swapcache_use_chflags)) {
 452                                 vm_object_drop(object);
 453                                 vm_page_wakeup(m);
 454                                 vm_page_queues_spin_lock(marker->queue);
 455                                 continue;
 456                         }
 457                         if (vm_swapcache_maxfilesize &&
 458                             object->size >
 459                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 460                                 vm_object_drop(object);
 461                                 vm_page_wakeup(m);
 462                                 vm_page_queues_spin_lock(marker->queue);
 463                                 continue;
 464                         }
 465                         isblkdev = 0;
 466                         break;
 467                 case VCHR:
 468                         /*
 469                          * PG_NOTMETA generically means 'don't swapcache this',
 470                          * and HAMMER will set this for regular data buffers
 471                          * (and leave it unset for meta-data buffers) as
 472                          * appropriate when double buffering is enabled.
 473                          */
 474                         if (m->flags & PG_NOTMETA) {
 475                                 vm_object_drop(object);
 476                                 vm_page_wakeup(m);
 477                                 vm_page_queues_spin_lock(marker->queue);
 478                                 continue;
 479                         }
 480                         if (vm_swapcache_meta_enable == 0) {
 481                                 vm_object_drop(object);
 482                                 vm_page_wakeup(m);
 483                                 vm_page_queues_spin_lock(marker->queue);
 484                                 continue;
 485                         }
 486                         isblkdev = 1;
 487                         break;
 488                 default:
 489                         vm_object_drop(object);
 490                         vm_page_wakeup(m);
 491                         vm_page_queues_spin_lock(marker->queue);
 492                         continue;
 493                 }
 494
 495
 496                 /*
 497                  * Assign swap and initiate I/O.
 498                  *
 499                  * (adjust for the --count which also occurs in the loop)
 500                  */
 501                 count -= vm_swapcached_flush(m, isblkdev);
 502
 503                 /*
 504                  * Setup for next loop using marker.
 505                  */
 506                 vm_object_drop(object);
 507                 vm_page_queues_spin_lock(marker->queue);
 508         }
 509
 510         /*
 511          * The marker could wind up at the end, which is ok.  If we hit the
 512          * end of the list adjust the heuristic.
 513          *
 514          * Earlier inactive pages that were dirty and become clean
 515          * are typically moved to the end of PQ_INACTIVE by virtue
 516          * of vfs_vmio_release() when they become unwired from the
 517          * buffer cache.
 518          */
 519         vm_page_queues_spin_unlock(marker->queue);
 520
 521         /*
 522          * m invalid but can be used to test for NULL
 523          */
 524         return (m == NULL);
 525 }
 526
 527 /*
 528  * Flush the specified page using the swap_pager.  The page
 529  * must be busied by the caller and its disposition will become
 530  * the responsibility of this function.
 531  *
 532  * Try to collect surrounding pages, including pages which may
 533  * have already been assigned swap.  Try to cluster within a
 534  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 535  * to match what swap_pager_putpages() can do.
 536  *
 537  * We also want to try to match against the buffer cache blocksize
 538  * but we don't really know what it is here.  Since the buffer cache
 539  * wires and unwires pages in groups the fact that we skip wired pages
 540  * should be sufficient.
 541  *
 542  * Returns a count of pages we might have flushed (minimum 1)
 543  */
 544 static
 545 int
 546 vm_swapcached_flush(vm_page_t m, int isblkdev)
 547 {
 548         vm_object_t object;
 549         vm_page_t marray[SWAP_META_PAGES];
 550         vm_pindex_t basei;
 551         int rtvals[SWAP_META_PAGES];
 552         int x;
 553         int i;
 554         int j;
 555         int count;
 556         int error;
 557
 558         vm_page_io_start(m);
 559         vm_page_protect(m, VM_PROT_READ);
 560         object = m->object;
 561         vm_object_hold(object);
 562
 563         /*
 564          * Try to cluster around (m), keeping in mind that the swap pager
 565          * can only do SMAP_META_PAGES worth of continguous write.
 566          */
 567         x = (int)m->pindex & SWAP_META_MASK;
 568         marray[x] = m;
 569         basei = m->pindex;
 570         vm_page_wakeup(m);
 571
 572         for (i = x - 1; i >= 0; --i) {
 573                 m = vm_page_lookup_busy_try(object, basei - x + i,
 574                                             TRUE, &error);
 575                 if (error || m == NULL)
 576                         break;
 577                 if (vm_swapcache_test(m)) {
 578                         vm_page_wakeup(m);
 579                         break;
 580                 }
 581                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 582                         vm_page_wakeup(m);
 583                         break;
 584                 }
 585                 vm_page_io_start(m);
 586                 vm_page_protect(m, VM_PROT_READ);
 587                 if (m->queue - m->pc == PQ_CACHE) {
 588                         vm_page_unqueue_nowakeup(m);
 589                         vm_page_deactivate(m);
 590                 }
 591                 marray[i] = m;
 592                 vm_page_wakeup(m);
 593         }
 594         ++i;
 595
 596         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 597                 m = vm_page_lookup_busy_try(object, basei - x + j,
 598                                             TRUE, &error);
 599                 if (error || m == NULL)
 600                         break;
 601                 if (vm_swapcache_test(m)) {
 602                         vm_page_wakeup(m);
 603                         break;
 604                 }
 605                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 606                         vm_page_wakeup(m);
 607                         break;
 608                 }
 609                 vm_page_io_start(m);
 610                 vm_page_protect(m, VM_PROT_READ);
 611                 if (m->queue - m->pc == PQ_CACHE) {
 612                         vm_page_unqueue_nowakeup(m);
 613                         vm_page_deactivate(m);
 614                 }
 615                 marray[j] = m;
 616                 vm_page_wakeup(m);
 617         }
 618
 619         count = j - i;
 620         vm_object_pip_add(object, count);
 621         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 622         vm_swapcache_write_count += count * PAGE_SIZE;
 623         vm_swapcache_curburst -= count * PAGE_SIZE;
 624
 625         while (i < j) {
 626                 if (rtvals[i] != VM_PAGER_PEND) {
 627                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 628                         vm_page_io_finish(marray[i]);
 629                         vm_page_wakeup(marray[i]);
 630                         vm_object_pip_wakeup(object);
 631                 }
 632                 ++i;
 633         }
 634         vm_object_drop(object);
 635         return(count);
 636 }
 637
 638 /*
 639  * Test whether a VM page is suitable for writing to the swapcache.
 640  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 641  *
 642  * Returns 0 on success, 1 on failure
 643  */
 644 static int
 645 vm_swapcache_test(vm_page_t m)
 646 {
 647         vm_object_t object;
 648
 649         if (m->flags & PG_UNMANAGED)
 650                 return(1);
 651         if (m->hold_count || m->wire_count)
 652                 return(1);
 653         if (m->valid != VM_PAGE_BITS_ALL)
 654                 return(1);
 655         if (m->dirty & m->valid)
 656                 return(1);
 657         if ((object = m->object) == NULL)
 658                 return(1);
 659         if (object->type != OBJT_VNODE ||
 660             (object->flags & OBJ_DEAD)) {
 661                 return(1);
 662         }
 663         vm_page_test_dirty(m);
 664         if (m->dirty & m->valid)
 665                 return(1);
 666         return(0);
 667 }
 668
 669 /*
 670  * Cleaning pass.
 671  *
 672  * We clean whole objects up to 16MB
 673  */
 674 static
 675 void
 676 vm_swapcache_cleaning(vm_object_t marker)
 677 {
 678         vm_object_t object;
 679         struct vnode *vp;
 680         int count;
 681         int scount;
 682         int n;
 683
 684         count = vm_swapcache_maxlaunder;
 685         scount = vm_swapcache_maxscan;
 686
 687         /*
 688          * Look for vnode objects
 689          */
 690         lwkt_gettoken(&vmobj_token);
 691
 692         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 693                 /*
 694                  * We have to skip markers.  We cannot hold/drop marker
 695                  * objects!
 696                  */
 697                 if (object->type == OBJT_MARKER) {
 698                         vm_swapcache_movemarker(marker, object);
 699                         continue;
 700                 }
 701
 702                 /*
 703                  * Safety, or in case there are millions of VM objects
 704                  * without swapcache backing.
 705                  */
 706                 if (--scount <= 0)
 707                         break;
 708
 709                 /*
 710                  * We must hold the object before potentially yielding.
 711                  */
 712                 vm_object_hold(object);
 713                 lwkt_yield();
 714
 715                 /*
 716                  * Only operate on live VNODE objects that are either
 717                  * VREG or VCHR (VCHR for meta-data).
 718                  */
 719                 if ((object->type != OBJT_VNODE) ||
 720                     ((object->flags & OBJ_DEAD) ||
 721                      object->swblock_count == 0) ||
 722                     ((vp = object->handle) == NULL) ||
 723                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 724                         vm_object_drop(object);
 725                         /* object may be invalid now */
 726                         vm_swapcache_movemarker(marker, object);
 727                         continue;
 728                 }
 729
 730                 /*
 731                  * Reset the object pindex stored in the marker if the
 732                  * working object has changed.
 733                  */
 734                 if (marker->backing_object != object) {
 735                         marker->size = 0;
 736                         marker->backing_object_offset = 0;
 737                         marker->backing_object = object;
 738                 }
 739
 740                 /*
 741                  * Look for swblocks starting at our iterator.
 742                  *
 743                  * The swap_pager_condfree() function attempts to free
 744                  * swap space starting at the specified index.  The index
 745                  * will be updated on return.  The function will return
 746                  * a scan factor (NOT the number of blocks freed).
 747                  *
 748                  * If it must cut its scan of the object short due to an
 749                  * excessive number of swblocks, or is able to free the
 750                  * requested number of blocks, it will return n >= count
 751                  * and we break and pick it back up on a future attempt.
 752                  *
 753                  * Scan the object linearly and try to batch large sets of
 754                  * blocks that are likely to clean out entire swap radix
 755                  * tree leafs.
 756                  */
 757                 lwkt_token_swap();
 758                 lwkt_reltoken(&vmobj_token);
 759
 760                 n = swap_pager_condfree(object, &marker->size,
 761                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 762
 763                 vm_object_drop(object);         /* object may be invalid now */
 764                 lwkt_gettoken(&vmobj_token);
 765
 766                 /*
 767                  * If we have exhausted the object or deleted our per-pass
 768                  * page limit then move us to the next object.  Note that
 769                  * the current object may no longer be on the vm_object_list.
 770                  */
 771                 if (n <= 0 ||
 772                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 773                         vm_swapcache_movemarker(marker, object);
 774                 }
 775
 776                 /*
 777                  * If we have exhausted our max-launder stop for now.
 778                  */
 779                 count -= n;
 780                 marker->backing_object_offset += n * PAGE_SIZE;
 781                 if (count < 0)
 782                         break;
 783         }
 784
 785         /*
 786          * If we wound up at the end of the list this will move the
 787          * marker back to the beginning.
 788          */
 789         if (object == NULL)
 790                 vm_swapcache_movemarker(marker, NULL);
 791
 792         lwkt_reltoken(&vmobj_token);
 793 }
 794
 795 /*
 796  * Move the marker past the current object.  Object can be stale, but we
 797  * still need it to determine if the marker has to be moved.  If the object
 798  * is still the 'current object' (object after the marker), we hop-scotch
 799  * the marker past it.
 800  */
 801 static void
 802 vm_swapcache_movemarker(vm_object_t marker, vm_object_t object)
 803 {
 804         if (TAILQ_NEXT(marker, object_list) == object) {
 805                 TAILQ_REMOVE(&vm_object_list, marker, object_list);
 806                 if (object) {
 807                         TAILQ_INSERT_AFTER(&vm_object_list, object,
 808                                            marker, object_list);
 809                 } else {
 810                         TAILQ_INSERT_HEAD(&vm_object_list,
 811                                           marker, object_list);
 812                 }
 813         }
 814 }