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/spinlock2.h>
  79 #include <vm/vm_page2.h>
  80
  81 /* the kernel process "vm_pageout"*/
  82 static int vm_swapcached_flush (vm_page_t m, int isblkdev);
  83 static int vm_swapcache_test(vm_page_t m);
  84 static int vm_swapcache_writing_heuristic(void);
  85 static int vm_swapcache_writing(vm_page_t marker, int count, int scount);
  86 static void vm_swapcache_cleaning(vm_object_t marker,
  87                         struct vm_object_hash **swindexp);
  88 static void vm_swapcache_movemarker(vm_object_t marker,
  89                         struct vm_object_hash *swindex, 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 static long vm_swapcache_wtrigger;
  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 = 0;       /* 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 struct vm_object_hash *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_FICTITIOUS | PG_MARKER;
 194                 page_marker[q].busy_count = PBUSY_LOCKED;
 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_wtrigger = -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_wtrigger = -vm_swapcache_hysteresis;
 312                 }
 313         }
 314
 315         /*
 316          * Cleanup (NOT REACHED)
 317          */
 318         for (q = 0; q < PQ_L2_SIZE; ++q) {
 319                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 320                 TAILQ_REMOVE(
 321                         &vm_page_queues[PQ_INACTIVE + q].pl,
 322                         &page_marker[q], pageq);
 323                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 324         }
 325
 326         lwkt_gettoken(&swindex->token);
 327         TAILQ_REMOVE(&swindex->list, &swmarker, object_list);
 328         lwkt_reltoken(&swindex->token);
 329 }
 330
 331 static struct kproc_desc swpc_kp = {
 332         "swapcached",
 333         vm_swapcached_thread,
 334         &swapcached_thread
 335 };
 336 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp);
 337
 338 /*
 339  * Deal with an overflow of the heuristic counter or if the user
 340  * manually changes the hysteresis.
 341  *
 342  * Try to avoid small incremental pageouts by waiting for enough
 343  * pages to buildup in the inactive queue to hopefully get a good
 344  * burst in.  This heuristic is bumped by the VM system and reset
 345  * when our scan hits the end of the queue.
 346  *
 347  * Return TRUE if we need to take a writing pass.
 348  */
 349 static int
 350 vm_swapcache_writing_heuristic(void)
 351 {
 352         int hyst;
 353         int q;
 354         long adds;
 355
 356         hyst = vmstats.v_inactive_count / 4;
 357         if (hyst < vm_swapcache_min_hysteresis)
 358                 hyst = vm_swapcache_min_hysteresis;
 359         cpu_ccfence();
 360         vm_swapcache_hysteresis = hyst;
 361
 362         adds = 0;
 363         for (q = PQ_INACTIVE; q < PQ_INACTIVE + PQ_L2_SIZE; ++q) {
 364                 adds += atomic_swap_long(&vm_page_queues[q].adds, 0);
 365         }
 366         vm_swapcache_wtrigger += adds;
 367         if (vm_swapcache_wtrigger < -hyst)
 368                 vm_swapcache_wtrigger = -hyst;
 369         return (vm_swapcache_wtrigger >= 0);
 370 }
 371
 372 /*
 373  * Take a writing pass on one of the inactive queues, return non-zero if
 374  * we hit the end of the queue.
 375  */
 376 static int
 377 vm_swapcache_writing(vm_page_t marker, int count, int scount)
 378 {
 379         vm_object_t object;
 380         struct vnode *vp;
 381         vm_page_t m;
 382         int isblkdev;
 383
 384         /*
 385          * Scan the inactive queue from our marker to locate
 386          * suitable pages to push to the swap cache.
 387          *
 388          * We are looking for clean vnode-backed pages.
 389          */
 390         vm_page_queues_spin_lock(marker->queue);
 391         while ((m = TAILQ_NEXT(marker, pageq)) != NULL &&
 392                count > 0 && scount-- > 0) {
 393                 KKASSERT(m->queue == marker->queue);
 394
 395                 /*
 396                  * Stop using swap if paniced, dumping, or dumped.
 397                  * Don't try to write if our curburst has been exhausted.
 398                  */
 399                 if (panicstr || dumping)
 400                         break;
 401                 if (vm_swapcache_curburst < 0)
 402                         break;
 403
 404                 /*
 405                  * Move marker
 406                  */
 407                 TAILQ_REMOVE(
 408                         &vm_page_queues[marker->queue].pl, marker, pageq);
 409                 TAILQ_INSERT_AFTER(
 410                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 411
 412                 /*
 413                  * Ignore markers and ignore pages that already have a swap
 414                  * assignment.
 415                  */
 416                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 417                         continue;
 418                 if (vm_page_busy_try(m, TRUE))
 419                         continue;
 420                 vm_page_queues_spin_unlock(marker->queue);
 421
 422                 if ((object = m->object) == NULL) {
 423                         vm_page_wakeup(m);
 424                         vm_page_queues_spin_lock(marker->queue);
 425                         continue;
 426                 }
 427                 vm_object_hold(object);
 428                 if (m->object != object) {
 429                         vm_object_drop(object);
 430                         vm_page_wakeup(m);
 431                         vm_page_queues_spin_lock(marker->queue);
 432                         continue;
 433                 }
 434                 if (vm_swapcache_test(m)) {
 435                         vm_object_drop(object);
 436                         vm_page_wakeup(m);
 437                         vm_page_queues_spin_lock(marker->queue);
 438                         continue;
 439                 }
 440
 441                 vp = object->handle;
 442                 if (vp == NULL) {
 443                         vm_object_drop(object);
 444                         vm_page_wakeup(m);
 445                         vm_page_queues_spin_lock(marker->queue);
 446                         continue;
 447                 }
 448
 449                 switch(vp->v_type) {
 450                 case VREG:
 451                         /*
 452                          * PG_NOTMETA generically means 'don't swapcache this',
 453                          * and HAMMER will set this for regular data buffers
 454                          * (and leave it unset for meta-data buffers) as
 455                          * appropriate when double buffering is enabled.
 456                          */
 457                         if (m->flags & PG_NOTMETA) {
 458                                 vm_object_drop(object);
 459                                 vm_page_wakeup(m);
 460                                 vm_page_queues_spin_lock(marker->queue);
 461                                 continue;
 462                         }
 463
 464                         /*
 465                          * If data_enable is 0 do not try to swapcache data.
 466                          * If use_chflags is set then only swapcache data for
 467                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 468                          */
 469                         if (vm_swapcache_data_enable == 0 ||
 470                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 471                              vm_swapcache_use_chflags)) {
 472                                 vm_object_drop(object);
 473                                 vm_page_wakeup(m);
 474                                 vm_page_queues_spin_lock(marker->queue);
 475                                 continue;
 476                         }
 477                         if (vm_swapcache_maxfilesize &&
 478                             object->size >
 479                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 480                                 vm_object_drop(object);
 481                                 vm_page_wakeup(m);
 482                                 vm_page_queues_spin_lock(marker->queue);
 483                                 continue;
 484                         }
 485                         isblkdev = 0;
 486                         break;
 487                 case VCHR:
 488                         /*
 489                          * PG_NOTMETA generically means 'don't swapcache this',
 490                          * and HAMMER will set this for regular data buffers
 491                          * (and leave it unset for meta-data buffers) as
 492                          * appropriate when double buffering is enabled.
 493                          */
 494                         if (m->flags & PG_NOTMETA) {
 495                                 vm_object_drop(object);
 496                                 vm_page_wakeup(m);
 497                                 vm_page_queues_spin_lock(marker->queue);
 498                                 continue;
 499                         }
 500                         if (vm_swapcache_meta_enable == 0) {
 501                                 vm_object_drop(object);
 502                                 vm_page_wakeup(m);
 503                                 vm_page_queues_spin_lock(marker->queue);
 504                                 continue;
 505                         }
 506                         isblkdev = 1;
 507                         break;
 508                 default:
 509                         vm_object_drop(object);
 510                         vm_page_wakeup(m);
 511                         vm_page_queues_spin_lock(marker->queue);
 512                         continue;
 513                 }
 514
 515
 516                 /*
 517                  * Assign swap and initiate I/O.
 518                  *
 519                  * (adjust for the --count which also occurs in the loop)
 520                  */
 521                 count -= vm_swapcached_flush(m, isblkdev);
 522
 523                 /*
 524                  * Setup for next loop using marker.
 525                  */
 526                 vm_object_drop(object);
 527                 vm_page_queues_spin_lock(marker->queue);
 528         }
 529
 530         /*
 531          * The marker could wind up at the end, which is ok.  If we hit the
 532          * end of the list adjust the heuristic.
 533          *
 534          * Earlier inactive pages that were dirty and become clean
 535          * are typically moved to the end of PQ_INACTIVE by virtue
 536          * of vfs_vmio_release() when they become unwired from the
 537          * buffer cache.
 538          */
 539         vm_page_queues_spin_unlock(marker->queue);
 540
 541         /*
 542          * m invalid but can be used to test for NULL
 543          */
 544         return (m == NULL);
 545 }
 546
 547 /*
 548  * Flush the specified page using the swap_pager.  The page
 549  * must be busied by the caller and its disposition will become
 550  * the responsibility of this function.
 551  *
 552  * Try to collect surrounding pages, including pages which may
 553  * have already been assigned swap.  Try to cluster within a
 554  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 555  * to match what swap_pager_putpages() can do.
 556  *
 557  * We also want to try to match against the buffer cache blocksize
 558  * but we don't really know what it is here.  Since the buffer cache
 559  * wires and unwires pages in groups the fact that we skip wired pages
 560  * should be sufficient.
 561  *
 562  * Returns a count of pages we might have flushed (minimum 1)
 563  */
 564 static
 565 int
 566 vm_swapcached_flush(vm_page_t m, int isblkdev)
 567 {
 568         vm_object_t object;
 569         vm_page_t marray[SWAP_META_PAGES];
 570         vm_pindex_t basei;
 571         int rtvals[SWAP_META_PAGES];
 572         int x;
 573         int i;
 574         int j;
 575         int count;
 576         int error;
 577
 578         vm_page_io_start(m);
 579         vm_page_protect(m, VM_PROT_READ);
 580         object = m->object;
 581         vm_object_hold(object);
 582
 583         /*
 584          * Try to cluster around (m), keeping in mind that the swap pager
 585          * can only do SMAP_META_PAGES worth of continguous write.
 586          */
 587         x = (int)m->pindex & SWAP_META_MASK;
 588         marray[x] = m;
 589         basei = m->pindex;
 590         vm_page_wakeup(m);
 591
 592         for (i = x - 1; i >= 0; --i) {
 593                 m = vm_page_lookup_busy_try(object, basei - x + i,
 594                                             TRUE, &error);
 595                 if (error || m == NULL)
 596                         break;
 597                 if (vm_swapcache_test(m)) {
 598                         vm_page_wakeup(m);
 599                         break;
 600                 }
 601                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 602                         vm_page_wakeup(m);
 603                         break;
 604                 }
 605                 vm_page_io_start(m);
 606                 vm_page_protect(m, VM_PROT_READ);
 607                 if (m->queue - m->pc == PQ_CACHE) {
 608                         vm_page_unqueue_nowakeup(m);
 609                         vm_page_deactivate(m);
 610                 }
 611                 marray[i] = m;
 612                 vm_page_wakeup(m);
 613         }
 614         ++i;
 615
 616         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 617                 m = vm_page_lookup_busy_try(object, basei - x + j,
 618                                             TRUE, &error);
 619                 if (error || m == NULL)
 620                         break;
 621                 if (vm_swapcache_test(m)) {
 622                         vm_page_wakeup(m);
 623                         break;
 624                 }
 625                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 626                         vm_page_wakeup(m);
 627                         break;
 628                 }
 629                 vm_page_io_start(m);
 630                 vm_page_protect(m, VM_PROT_READ);
 631                 if (m->queue - m->pc == PQ_CACHE) {
 632                         vm_page_unqueue_nowakeup(m);
 633                         vm_page_deactivate(m);
 634                 }
 635                 marray[j] = m;
 636                 vm_page_wakeup(m);
 637         }
 638
 639         count = j - i;
 640         vm_object_pip_add(object, count);
 641         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 642         vm_swapcache_write_count += count * PAGE_SIZE;
 643         vm_swapcache_curburst -= count * PAGE_SIZE;
 644
 645         while (i < j) {
 646                 if (rtvals[i] != VM_PAGER_PEND) {
 647                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 648                         vm_page_io_finish(marray[i]);
 649                         vm_page_wakeup(marray[i]);
 650                         vm_object_pip_wakeup(object);
 651                 }
 652                 ++i;
 653         }
 654         vm_object_drop(object);
 655         return(count);
 656 }
 657
 658 /*
 659  * Test whether a VM page is suitable for writing to the swapcache.
 660  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 661  *
 662  * Returns 0 on success, 1 on failure
 663  */
 664 static int
 665 vm_swapcache_test(vm_page_t m)
 666 {
 667         vm_object_t object;
 668
 669         if (m->flags & PG_UNMANAGED)
 670                 return(1);
 671         if (m->hold_count || m->wire_count)
 672                 return(1);
 673         if (m->valid != VM_PAGE_BITS_ALL)
 674                 return(1);
 675         if (m->dirty & m->valid)
 676                 return(1);
 677         if ((object = m->object) == NULL)
 678                 return(1);
 679         if (object->type != OBJT_VNODE ||
 680             (object->flags & OBJ_DEAD)) {
 681                 return(1);
 682         }
 683         vm_page_test_dirty(m);
 684         if (m->dirty & m->valid)
 685                 return(1);
 686         return(0);
 687 }
 688
 689 /*
 690  * Cleaning pass.
 691  *
 692  * We clean whole objects up to 16MB
 693  */
 694 static
 695 void
 696 vm_swapcache_cleaning(vm_object_t marker, struct vm_object_hash **swindexp)
 697 {
 698         vm_object_t object;
 699         struct vnode *vp;
 700         int count;
 701         int scount;
 702         int n;
 703         int didmove;
 704
 705         count = vm_swapcache_maxlaunder;
 706         scount = vm_swapcache_maxscan;
 707
 708         /*
 709          * Look for vnode objects
 710          */
 711         lwkt_gettoken(&(*swindexp)->token);
 712
 713         didmove = 0;
 714 outerloop:
 715         while ((object = TAILQ_NEXT(marker, object_list)) != NULL) {
 716                 /*
 717                  * We have to skip markers.  We cannot hold/drop marker
 718                  * objects!
 719                  */
 720                 if (object->type == OBJT_MARKER) {
 721                         vm_swapcache_movemarker(marker, *swindexp, object);
 722                         didmove = 1;
 723                         continue;
 724                 }
 725
 726                 /*
 727                  * Safety, or in case there are millions of VM objects
 728                  * without swapcache backing.
 729                  */
 730                 if (--scount <= 0)
 731                         goto breakout;
 732
 733                 /*
 734                  * We must hold the object before potentially yielding.
 735                  */
 736                 vm_object_hold(object);
 737                 lwkt_yield();
 738
 739                 /*
 740                  * Only operate on live VNODE objects that are either
 741                  * VREG or VCHR (VCHR for meta-data).
 742                  */
 743                 if ((object->type != OBJT_VNODE) ||
 744                     ((object->flags & OBJ_DEAD) ||
 745                      object->swblock_count == 0) ||
 746                     ((vp = object->handle) == NULL) ||
 747                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 748                         vm_object_drop(object);
 749                         /* object may be invalid now */
 750                         vm_swapcache_movemarker(marker, *swindexp, object);
 751                         didmove = 1;
 752                         continue;
 753                 }
 754
 755                 /*
 756                  * Reset the object pindex stored in the marker if the
 757                  * working object has changed.
 758                  */
 759                 if (marker->backing_object != object || didmove) {
 760                         marker->size = 0;
 761                         marker->backing_object_offset = 0;
 762                         marker->backing_object = object;
 763                         didmove = 0;
 764                 }
 765
 766                 /*
 767                  * Look for swblocks starting at our iterator.
 768                  *
 769                  * The swap_pager_condfree() function attempts to free
 770                  * swap space starting at the specified index.  The index
 771                  * will be updated on return.  The function will return
 772                  * a scan factor (NOT the number of blocks freed).
 773                  *
 774                  * If it must cut its scan of the object short due to an
 775                  * excessive number of swblocks, or is able to free the
 776                  * requested number of blocks, it will return n >= count
 777                  * and we break and pick it back up on a future attempt.
 778                  *
 779                  * Scan the object linearly and try to batch large sets of
 780                  * blocks that are likely to clean out entire swap radix
 781                  * tree leafs.
 782                  */
 783                 lwkt_token_swap();
 784                 lwkt_reltoken(&(*swindexp)->token);
 785
 786                 n = swap_pager_condfree(object, &marker->size,
 787                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 788
 789                 vm_object_drop(object);         /* object may be invalid now */
 790                 lwkt_gettoken(&(*swindexp)->token);
 791
 792                 /*
 793                  * If we have exhausted the object or deleted our per-pass
 794                  * page limit then move us to the next object.  Note that
 795                  * the current object may no longer be on the vm_object_list.
 796                  */
 797                 if (n <= 0 ||
 798                     marker->backing_object_offset > vm_swapcache_cleanperobj) {
 799                         vm_swapcache_movemarker(marker, *swindexp, object);
 800                         didmove = 1;
 801                 }
 802
 803                 /*
 804                  * If we have exhausted our max-launder stop for now.
 805                  */
 806                 count -= n;
 807                 marker->backing_object_offset += n * PAGE_SIZE;
 808                 if (count < 0)
 809                         goto breakout;
 810         }
 811
 812         /*
 813          * Iterate vm_object_lists[] hash table
 814          */
 815         TAILQ_REMOVE(&(*swindexp)->list, marker, object_list);
 816         lwkt_reltoken(&(*swindexp)->token);
 817         if (++*swindexp >= &vm_object_hash[VMOBJ_HSIZE])
 818                 *swindexp = &vm_object_hash[0];
 819         lwkt_gettoken(&(*swindexp)->token);
 820         TAILQ_INSERT_HEAD(&(*swindexp)->list, marker, object_list);
 821
 822         if (*swindexp != &vm_object_hash[0])
 823                 goto outerloop;
 824
 825 breakout:
 826         lwkt_reltoken(&(*swindexp)->token);
 827 }
 828
 829 /*
 830  * Move the marker past the current object.  Object can be stale, but we
 831  * still need it to determine if the marker has to be moved.  If the object
 832  * is still the 'current object' (object after the marker), we hop-scotch
 833  * the marker past it.
 834  */
 835 static void
 836 vm_swapcache_movemarker(vm_object_t marker, struct vm_object_hash *swindex,
 837                         vm_object_t object)
 838 {
 839         if (TAILQ_NEXT(marker, object_list) == object) {
 840                 TAILQ_REMOVE(&swindex->list, marker, object_list);
 841                 TAILQ_INSERT_AFTER(&swindex->list, object, marker, object_list);
 842         }
 843 }