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                         struct vm_object_hash **swindexp);
  89 static void vm_swapcache_movemarker(vm_object_t marker,
  90                         struct vm_object_hash *swindex, vm_object_t object);
  91 struct thread *swapcached_thread;
  92
  93 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  94
  95 int vm_swapcache_read_enable;
  96 int vm_swapcache_inactive_heuristic;
  97 static int vm_swapcache_sleep;
  98 static int vm_swapcache_maxscan = PQ_L2_SIZE * 8;
  99 static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4;
 100 static int vm_swapcache_data_enable = 0;
 101 static int vm_swapcache_meta_enable = 0;
 102 static int vm_swapcache_maxswappct = 75;
 103 static int vm_swapcache_hysteresis;
 104 static int vm_swapcache_min_hysteresis;
 105 int vm_swapcache_use_chflags = 1;       /* require chflags cache */
 106 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 107 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 108 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 109 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 110 static int64_t vm_swapcache_write_count;
 111 static int64_t vm_swapcache_maxfilesize;
 112 static int64_t vm_swapcache_cleanperobj = 16*1024*1024;
 113
 114 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 115         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 116 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan,
 117         CTLFLAG_RW, &vm_swapcache_maxscan, 0, "");
 118
 119 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 120         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 121 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 122         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 123 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 124         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 125 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 126         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 127 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 128         CTLFLAG_RD, &vm_swapcache_hysteresis, 0, "");
 129 SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis,
 130         CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, "");
 131 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 132         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 133
 134 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 135         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 136 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 137         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 138 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 139         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 140 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 141         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 142 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 143         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 144 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 145         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 146 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj,
 147         CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, "");
 148
 149 #define SWAPMAX(adj)    \
 150         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 151
 152 /*
 153  * When shutting down the machine we want to stop swapcache operation
 154  * immediately so swap is not accessed after devices have been shuttered.
 155  */
 156 static void
 157 shutdown_swapcache(void *arg __unused)
 158 {
 159         vm_swapcache_read_enable = 0;
 160         vm_swapcache_data_enable = 0;
 161         vm_swapcache_meta_enable = 0;
 162         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 163 }
 164
 165 /*
 166  * vm_swapcached is the high level pageout daemon.
 167  *
 168  * No requirements.
 169  */
 170 static void
 171 vm_swapcached_thread(void)
 172 {
 173         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 174         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 175         static struct vm_page page_marker[PQ_L2_SIZE];
 176         static struct vm_object swmarker;
 177         static struct vm_object_hash *swindex;
 178         int q;
 179
 180         /*
 181          * Thread setup
 182          */
 183         curthread->td_flags |= TDF_SYSTHREAD;
 184         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 185                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 186         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 187                               NULL, SHUTDOWN_PRI_SECOND);
 188
 189         /*
 190          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 191          */
 192         bzero(&page_marker, sizeof(page_marker));
 193         for (q = 0; q < PQ_L2_SIZE; ++q) {
 194                 page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 195                 page_marker[q].queue = PQ_INACTIVE + q;
 196                 page_marker[q].pc = q;
 197                 page_marker[q].wire_count = 1;
 198                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 199                 TAILQ_INSERT_HEAD(
 200                         &vm_page_queues[PQ_INACTIVE + q].pl,
 201                         &page_marker[q], pageq);
 202                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 203         }
 204
 205         vm_swapcache_min_hysteresis = 1024;
 206         vm_swapcache_hysteresis = vm_swapcache_min_hysteresis;
 207         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 208
 209         /*
 210          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 211          */
 212         bzero(&swmarker, sizeof(swmarker));
 213         swmarker.type = OBJT_MARKER;
 214         swindex = &vm_object_hash[0];
 215         lwkt_gettoken(&swindex->token);
 216         TAILQ_INSERT_HEAD(&swindex->list, &swmarker, object_list);
 217         lwkt_reltoken(&swindex->token);
 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(&swindex->token);
 328         TAILQ_REMOVE(&swindex->list, &swmarker, object_list);
 329         lwkt_reltoken(&swindex->token);
 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                 /*
 391                  * Stop using swap if paniced, dumping, or dumped.
 392                  * Don't try to write if our curburst has been exhausted.
 393                  */
 394                 if (panicstr || dumping)
 395                         break;
 396                 if (vm_swapcache_curburst < 0)
 397                         break;
 398
 399                 /*
 400                  * Move marker
 401                  */
 402                 TAILQ_REMOVE(
 403                         &vm_page_queues[marker->queue].pl, marker, pageq);
 404                 TAILQ_INSERT_AFTER(
 405                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 406
 407                 /*
 408                  * Ignore markers and ignore pages that already have a swap
 409                  * assignment.
 410                  */
 411                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 412                         continue;
 413                 if (vm_page_busy_try(m, TRUE))
 414                         continue;
 415                 vm_page_queues_spin_unlock(marker->queue);
 416
 417                 if ((object = m->object) == NULL) {
 418                         vm_page_wakeup(m);
 419                         vm_page_queues_spin_lock(marker->queue);
 420                         continue;
 421                 }
 422                 vm_object_hold(object);
 423                 if (m->object != object) {
 424                         vm_object_drop(object);
 425                         vm_page_wakeup(m);
 426                         vm_page_queues_spin_lock(marker->queue);
 427                         continue;
 428                 }
 429                 if (vm_swapcache_test(m)) {
 430                         vm_object_drop(object);
 431                         vm_page_wakeup(m);
 432                         vm_page_queues_spin_lock(marker->queue);
 433                         continue;
 434                 }
 435
 436                 vp = object->handle;
 437                 if (vp == NULL) {
 438                         vm_object_drop(object);
 439                         vm_page_wakeup(m);
 440                         vm_page_queues_spin_lock(marker->queue);
 441                         continue;
 442                 }
 443
 444                 switch(vp->v_type) {
 445                 case VREG:
 446                         /*
 447                          * PG_NOTMETA generically means 'don't swapcache this',
 448                          * and HAMMER will set this for regular data buffers
 449                          * (and leave it unset for meta-data buffers) as
 450                          * appropriate when double buffering is enabled.
 451                          */
 452                         if (m->flags & PG_NOTMETA) {
 453                                 vm_object_drop(object);
 454                                 vm_page_wakeup(m);
 455                                 vm_page_queues_spin_lock(marker->queue);
 456                                 continue;
 457                         }
 458
 459                         /*
 460                          * If data_enable is 0 do not try to swapcache data.
 461                          * If use_chflags is set then only swapcache data for
 462                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 463                          */
 464                         if (vm_swapcache_data_enable == 0 ||
 465                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 466                              vm_swapcache_use_chflags)) {
 467                                 vm_object_drop(object);
 468                                 vm_page_wakeup(m);
 469                                 vm_page_queues_spin_lock(marker->queue);
 470                                 continue;
 471                         }
 472                         if (vm_swapcache_maxfilesize &&
 473                             object->size >
 474                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 475                                 vm_object_drop(object);
 476                                 vm_page_wakeup(m);
 477                                 vm_page_queues_spin_lock(marker->queue);
 478                                 continue;
 479                         }
 480                         isblkdev = 0;
 481                         break;
 482                 case VCHR:
 483                         /*
 484                          * PG_NOTMETA generically means 'don't swapcache this',
 485                          * and HAMMER will set this for regular data buffers
 486                          * (and leave it unset for meta-data buffers) as
 487                          * appropriate when double buffering is enabled.
 488                          */
 489                         if (m->flags & PG_NOTMETA) {
 490                                 vm_object_drop(object);
 491                                 vm_page_wakeup(m);
 492                                 vm_page_queues_spin_lock(marker->queue);
 493                                 continue;
 494                         }
 495                         if (vm_swapcache_meta_enable == 0) {
 496                                 vm_object_drop(object);
 497                                 vm_page_wakeup(m);
 498                                 vm_page_queues_spin_lock(marker->queue);
 499                                 continue;
 500                         }
 501                         isblkdev = 1;
 502                         break;
 503                 default:
 504                         vm_object_drop(object);
 505                         vm_page_wakeup(m);
 506                         vm_page_queues_spin_lock(marker->queue);
 507                         continue;
 508                 }
 509
 510
 511                 /*
 512                  * Assign swap and initiate I/O.
 513                  *
 514                  * (adjust for the --count which also occurs in the loop)
 515                  */
 516                 count -= vm_swapcached_flush(m, isblkdev);
 517
 518                 /*
 519                  * Setup for next loop using marker.
 520                  */
 521                 vm_object_drop(object);
 522                 vm_page_queues_spin_lock(marker->queue);
 523         }
 524
 525         /*
 526          * The marker could wind up at the end, which is ok.  If we hit the
 527          * end of the list adjust the heuristic.
 528          *
 529          * Earlier inactive pages that were dirty and become clean
 530          * are typically moved to the end of PQ_INACTIVE by virtue
 531          * of vfs_vmio_release() when they become unwired from the
 532          * buffer cache.
 533          */
 534         vm_page_queues_spin_unlock(marker->queue);
 535
 536         /*
 537          * m invalid but can be used to test for NULL
 538          */
 539         return (m == NULL);
 540 }
 541
 542 /*
 543  * Flush the specified page using the swap_pager.  The page
 544  * must be busied by the caller and its disposition will become
 545  * the responsibility of this function.
 546  *
 547  * Try to collect surrounding pages, including pages which may
 548  * have already been assigned swap.  Try to cluster within a
 549  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 550  * to match what swap_pager_putpages() can do.
 551  *
 552  * We also want to try to match against the buffer cache blocksize
 553  * but we don't really know what it is here.  Since the buffer cache
 554  * wires and unwires pages in groups the fact that we skip wired pages
 555  * should be sufficient.
 556  *
 557  * Returns a count of pages we might have flushed (minimum 1)
 558  */
 559 static
 560 int
 561 vm_swapcached_flush(vm_page_t m, int isblkdev)
 562 {
 563         vm_object_t object;
 564         vm_page_t marray[SWAP_META_PAGES];
 565         vm_pindex_t basei;
 566         int rtvals[SWAP_META_PAGES];
 567         int x;
 568         int i;
 569         int j;
 570         int count;
 571         int error;
 572
 573         vm_page_io_start(m);
 574         vm_page_protect(m, VM_PROT_READ);
 575         object = m->object;
 576         vm_object_hold(object);
 577
 578         /*
 579          * Try to cluster around (m), keeping in mind that the swap pager
 580          * can only do SMAP_META_PAGES worth of continguous write.
 581          */
 582         x = (int)m->pindex & SWAP_META_MASK;
 583         marray[x] = m;
 584         basei = m->pindex;
 585         vm_page_wakeup(m);
 586
 587         for (i = x - 1; i >= 0; --i) {
 588                 m = vm_page_lookup_busy_try(object, basei - x + i,
 589                                             TRUE, &error);
 590                 if (error || m == NULL)
 591                         break;
 592                 if (vm_swapcache_test(m)) {
 593                         vm_page_wakeup(m);
 594                         break;
 595                 }
 596                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 597                         vm_page_wakeup(m);
 598                         break;
 599                 }
 600                 vm_page_io_start(m);
 601                 vm_page_protect(m, VM_PROT_READ);
 602                 if (m->queue - m->pc == PQ_CACHE) {
 603                         vm_page_unqueue_nowakeup(m);
 604                         vm_page_deactivate(m);
 605                 }
 606                 marray[i] = m;
 607                 vm_page_wakeup(m);
 608         }
 609         ++i;
 610
 611         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 612                 m = vm_page_lookup_busy_try(object, basei - x + j,
 613                                             TRUE, &error);
 614                 if (error || m == NULL)
 615                         break;
 616                 if (vm_swapcache_test(m)) {
 617                         vm_page_wakeup(m);
 618                         break;
 619                 }
 620                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 621                         vm_page_wakeup(m);
 622                         break;
 623                 }
 624                 vm_page_io_start(m);
 625                 vm_page_protect(m, VM_PROT_READ);
 626                 if (m->queue - m->pc == PQ_CACHE) {
 627                         vm_page_unqueue_nowakeup(m);
 628                         vm_page_deactivate(m);
 629                 }
 630                 marray[j] = m;
 631                 vm_page_wakeup(m);
 632         }
 633
 634         count = j - i;
 635         vm_object_pip_add(object, count);
 636         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 637         vm_swapcache_write_count += count * PAGE_SIZE;
 638         vm_swapcache_curburst -= count * PAGE_SIZE;
 639
 640         while (i < j) {
 641                 if (rtvals[i] != VM_PAGER_PEND) {
 642                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 643                         vm_page_io_finish(marray[i]);
 644                         vm_page_wakeup(marray[i]);
 645                         vm_object_pip_wakeup(object);
 646                 }
 647                 ++i;
 648         }
 649         vm_object_drop(object);
 650         return(count);
 651 }
 652
 653 /*
 654  * Test whether a VM page is suitable for writing to the swapcache.
 655  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 656  *
 657  * Returns 0 on success, 1 on failure
 658  */
 659 static int
 660 vm_swapcache_test(vm_page_t m)
 661 {
 662         vm_object_t object;
 663
 664         if (m->flags & PG_UNMANAGED)
 665                 return(1);
 666         if (m->hold_count || m->wire_count)
 667                 return(1);
 668         if (m->valid != VM_PAGE_BITS_ALL)
 669                 return(1);
 670         if (m->dirty & m->valid)
 671                 return(1);
 672         if ((object = m->object) == NULL)
 673                 return(1);
 674         if (object->type != OBJT_VNODE ||
 675             (object->flags & OBJ_DEAD)) {
 676                 return(1);
 677         }
 678         vm_page_test_dirty(m);
 679         if (m->dirty & m->valid)
 680                 return(1);
 681         return(0);
 682 }
 683
 684 /*
 685  * Cleaning pass.
 686  *
 687  * We clean whole objects up to 16MB
 688  */
 689 static
 690 void
 691 vm_swapcache_cleaning(vm_object_t marker, struct vm_object_hash **swindexp)
 692 {
 693         vm_object_t object;
 694         struct vnode *vp;
 695         int count;
 696         int scount;
 697         int n;
 698
 699         count = vm_swapcache_maxlaunder;
 700         scount = vm_swapcache_maxscan;
 701
 702         /*
 703          * Look for vnode objects
 704          */
 705         lwkt_gettoken(&(*swindexp)->token);
 706
 707 outerloop:
 708         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 709                 /*
 710                  * We have to skip markers.  We cannot hold/drop marker
 711                  * objects!
 712                  */
 713                 if (object->type == OBJT_MARKER) {
 714                         vm_swapcache_movemarker(marker, *swindexp, object);
 715                         continue;
 716                 }
 717
 718                 /*
 719                  * Safety, or in case there are millions of VM objects
 720                  * without swapcache backing.
 721                  */
 722                 if (--scount <= 0)
 723                         goto breakout;
 724
 725                 /*
 726                  * We must hold the object before potentially yielding.
 727                  */
 728                 vm_object_hold(object);
 729                 lwkt_yield();
 730
 731                 /*
 732                  * Only operate on live VNODE objects that are either
 733                  * VREG or VCHR (VCHR for meta-data).
 734                  */
 735                 if ((object->type != OBJT_VNODE) ||
 736                     ((object->flags & OBJ_DEAD) ||
 737                      object->swblock_count == 0) ||
 738                     ((vp = object->handle) == NULL) ||
 739                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 740                         vm_object_drop(object);
 741                         /* object may be invalid now */
 742                         vm_swapcache_movemarker(marker, *swindexp, object);
 743                         continue;
 744                 }
 745
 746                 /*
 747                  * Reset the object pindex stored in the marker if the
 748                  * working object has changed.
 749                  */
 750                 if (marker->backing_object != object) {
 751                         marker->size = 0;
 752                         marker->backing_object_offset = 0;
 753                         marker->backing_object = object;
 754                 }
 755
 756                 /*
 757                  * Look for swblocks starting at our iterator.
 758                  *
 759                  * The swap_pager_condfree() function attempts to free
 760                  * swap space starting at the specified index.  The index
 761                  * will be updated on return.  The function will return
 762                  * a scan factor (NOT the number of blocks freed).
 763                  *
 764                  * If it must cut its scan of the object short due to an
 765                  * excessive number of swblocks, or is able to free the
 766                  * requested number of blocks, it will return n >= count
 767                  * and we break and pick it back up on a future attempt.
 768                  *
 769                  * Scan the object linearly and try to batch large sets of
 770                  * blocks that are likely to clean out entire swap radix
 771                  * tree leafs.
 772                  */
 773                 lwkt_token_swap();
 774                 lwkt_reltoken(&(*swindexp)->token);
 775
 776                 n = swap_pager_condfree(object, &marker->size,
 777                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 778
 779                 vm_object_drop(object);         /* object may be invalid now */
 780                 lwkt_gettoken(&(*swindexp)->token);
 781
 782                 /*
 783                  * If we have exhausted the object or deleted our per-pass
 784                  * page limit then move us to the next object.  Note that
 785                  * the current object may no longer be on the vm_object_list.
 786                  */
 787                 if (n <= 0 ||
 788                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 789                         vm_swapcache_movemarker(marker, *swindexp, object);
 790                 }
 791
 792                 /*
 793                  * If we have exhausted our max-launder stop for now.
 794                  */
 795                 count -= n;
 796                 marker->backing_object_offset += n * PAGE_SIZE;
 797                 if (count < 0)
 798                         goto breakout;
 799         }
 800
 801         /*
 802          * Iterate vm_object_lists[] hash table
 803          */
 804         TAILQ_REMOVE(&(*swindexp)->list, marker, object_list);
 805         lwkt_reltoken(&(*swindexp)->token);
 806         if (++*swindexp >= &vm_object_hash[VMOBJ_HSIZE])
 807                 *swindexp = &vm_object_hash[0];
 808         lwkt_gettoken(&(*swindexp)->token);
 809         TAILQ_INSERT_HEAD(&(*swindexp)->list, marker, object_list);
 810
 811         if (*swindexp != &vm_object_hash[0])
 812                 goto outerloop;
 813
 814 breakout:
 815         lwkt_reltoken(&(*swindexp)->token);
 816 }
 817
 818 /*
 819  * Move the marker past the current object.  Object can be stale, but we
 820  * still need it to determine if the marker has to be moved.  If the object
 821  * is still the 'current object' (object after the marker), we hop-scotch
 822  * the marker past it.
 823  */
 824 static void
 825 vm_swapcache_movemarker(vm_object_t marker, struct vm_object_hash *swindex,
 826                         vm_object_t object)
 827 {
 828         if (TAILQ_NEXT(marker, object_list) == object) {
 829                 TAILQ_REMOVE(&swindex->list, marker, object_list);
 830                 TAILQ_INSERT_AFTER(&swindex->list, object, marker, object_list);
 831         }
 832 }