sys/vm/vm_swapcache.c

   1 /*
   2  * (MPSAFE)
   3  *
   4  * Copyright (c) 2010,2019 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 struct swmarker {
  82         struct vm_object dummy_obj;
  83         struct vm_object *save_obj;
  84         vm_ooffset_t save_off;
  85 };
  86
  87 typedef struct swmarker swmarker_t;
  88
  89 /* the kernel process "vm_pageout"*/
  90 static int vm_swapcached_flush (vm_page_t m, int isblkdev);
  91 static int vm_swapcache_test(vm_page_t m);
  92 static int vm_swapcache_writing_heuristic(void);
  93 static int vm_swapcache_writing(vm_page_t marker, int count, int scount);
  94 static void vm_swapcache_cleaning(swmarker_t *marker,
  95                         struct vm_object_hash **swindexp);
  96 static void vm_swapcache_movemarker(swmarker_t *marker,
  97                         struct vm_object_hash *swindex, vm_object_t object);
  98 struct thread *swapcached_thread;
  99
 100 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
 101
 102 int vm_swapcache_read_enable;
 103 static long vm_swapcache_wtrigger;
 104 static int vm_swapcache_sleep;
 105 static int vm_swapcache_maxscan = PQ_L2_SIZE * 8;
 106 static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4;
 107 static int vm_swapcache_data_enable = 0;
 108 static int vm_swapcache_meta_enable = 0;
 109 static int vm_swapcache_maxswappct = 75;
 110 static int vm_swapcache_hysteresis;
 111 static int vm_swapcache_min_hysteresis;
 112 int vm_swapcache_use_chflags = 0;       /* require chflags cache */
 113 static int64_t vm_swapcache_minburst = 10000000LL;      /* 10MB */
 114 static int64_t vm_swapcache_curburst = 4000000000LL;    /* 4G after boot */
 115 static int64_t vm_swapcache_maxburst = 2000000000LL;    /* 2G nominal max */
 116 static int64_t vm_swapcache_accrate = 100000LL;         /* 100K/s */
 117 static int64_t vm_swapcache_write_count;
 118 static int64_t vm_swapcache_maxfilesize;
 119 static int64_t vm_swapcache_cleanperobj = 16*1024*1024;
 120
 121 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 122         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 123 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan,
 124         CTLFLAG_RW, &vm_swapcache_maxscan, 0, "");
 125
 126 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 127         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 128 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 129         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 130 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 131         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 132 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 133         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 134 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 135         CTLFLAG_RD, &vm_swapcache_hysteresis, 0, "");
 136 SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis,
 137         CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, "");
 138 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 139         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 140
 141 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 142         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 143 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 144         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 145 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 146         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 147 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 148         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 149 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 150         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 151 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 152         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 153 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj,
 154         CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, "");
 155
 156 #define SWAPMAX(adj)    \
 157         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 158
 159 /*
 160  * When shutting down the machine we want to stop swapcache operation
 161  * immediately so swap is not accessed after devices have been shuttered.
 162  */
 163 static void
 164 shutdown_swapcache(void *arg __unused)
 165 {
 166         vm_swapcache_read_enable = 0;
 167         vm_swapcache_data_enable = 0;
 168         vm_swapcache_meta_enable = 0;
 169         wakeup(&vm_swapcache_sleep);    /* shortcut 5-second wait */
 170 }
 171
 172 /*
 173  * vm_swapcached is the high level pageout daemon.
 174  *
 175  * No requirements.
 176  */
 177 static void
 178 vm_swapcached_thread(void)
 179 {
 180         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 181         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 182         static struct vm_page page_marker[PQ_L2_SIZE];
 183         static swmarker_t swmarker;
 184         static struct vm_object_hash *swindex;
 185         int q;
 186
 187         /*
 188          * Thread setup
 189          */
 190         curthread->td_flags |= TDF_SYSTHREAD;
 191         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
 192                               swapcached_thread, SHUTDOWN_PRI_FIRST);
 193         EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
 194                               NULL, SHUTDOWN_PRI_SECOND);
 195
 196         /*
 197          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 198          */
 199         bzero(&page_marker, sizeof(page_marker));
 200         for (q = 0; q < PQ_L2_SIZE; ++q) {
 201                 page_marker[q].flags = PG_FICTITIOUS | PG_MARKER;
 202                 page_marker[q].busy_count = PBUSY_LOCKED;
 203                 page_marker[q].queue = PQ_INACTIVE + q;
 204                 page_marker[q].pc = q;
 205                 page_marker[q].wire_count = 1;
 206                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 207                 TAILQ_INSERT_HEAD(
 208                         &vm_page_queues[PQ_INACTIVE + q].pl,
 209                         &page_marker[q], pageq);
 210                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 211         }
 212
 213         vm_swapcache_min_hysteresis = 1024;
 214         vm_swapcache_hysteresis = vm_swapcache_min_hysteresis;
 215         vm_swapcache_wtrigger = -vm_swapcache_hysteresis;
 216
 217         /*
 218          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 219          */
 220         bzero(&swmarker, sizeof(swmarker));
 221         swmarker.dummy_obj.type = OBJT_MARKER;
 222         swindex = &vm_object_hash[0];
 223         lwkt_gettoken(&swindex->token);
 224         TAILQ_INSERT_HEAD(&swindex->list, &swmarker.dummy_obj, object_entry);
 225         lwkt_reltoken(&swindex->token);
 226
 227         for (;;) {
 228                 int reached_end;
 229                 int scount;
 230                 int count;
 231
 232                 /*
 233                  * Handle shutdown
 234                  */
 235                 kproc_suspend_loop();
 236
 237                 /*
 238                  * Check every 5 seconds when not enabled or if no swap
 239                  * is present.
 240                  */
 241                 if ((vm_swapcache_data_enable == 0 &&
 242                      vm_swapcache_meta_enable == 0 &&
 243                      vm_swap_cache_use <= SWAPMAX(0)) ||
 244                     vm_swap_max == 0) {
 245                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 246                         continue;
 247                 }
 248
 249                 /*
 250                  * Polling rate when enabled is approximately 10 hz.
 251                  */
 252                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 253
 254                 /*
 255                  * State hysteresis.  Generate write activity up to 75% of
 256                  * swap, then clean out swap assignments down to 70%, then
 257                  * repeat.
 258                  */
 259                 if (state == SWAPC_WRITING) {
 260                         if (vm_swap_cache_use > SWAPMAX(0))
 261                                 state = SWAPC_CLEANING;
 262                 } else {
 263                         if (vm_swap_cache_use < SWAPMAX(-10))
 264                                 state = SWAPC_WRITING;
 265                 }
 266
 267                 /*
 268                  * We are allowed to continue accumulating burst value
 269                  * in either state.  Allow the user to set curburst > maxburst
 270                  * for the initial load-in.
 271                  */
 272                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 273                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 274                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 275                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 276                 }
 277
 278                 /*
 279                  * We don't want to nickle-and-dime the scan as that will
 280                  * create unnecessary fragmentation.  The minimum burst
 281                  * is one-seconds worth of accumulation.
 282                  */
 283                 if (state != SWAPC_WRITING) {
 284                         vm_swapcache_cleaning(&swmarker, &swindex);
 285                         continue;
 286                 }
 287                 if (vm_swapcache_curburst < vm_swapcache_accrate)
 288                         continue;
 289
 290                 reached_end = 0;
 291                 count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2;
 292                 scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2;
 293
 294                 if (burst == SWAPB_BURSTING) {
 295                         if (vm_swapcache_writing_heuristic()) {
 296                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 297                                         reached_end +=
 298                                                 vm_swapcache_writing(
 299                                                         &page_marker[q],
 300                                                         count,
 301                                                         scount);
 302                                 }
 303                         }
 304                         if (vm_swapcache_curburst <= 0)
 305                                 burst = SWAPB_RECOVERING;
 306                 } else if (vm_swapcache_curburst > vm_swapcache_minburst) {
 307                         if (vm_swapcache_writing_heuristic()) {
 308                                 for (q = 0; q < PQ_L2_SIZE; ++q) {
 309                                         reached_end +=
 310                                                 vm_swapcache_writing(
 311                                                         &page_marker[q],
 312                                                         count,
 313                                                         scount);
 314                                 }
 315                         }
 316                         burst = SWAPB_BURSTING;
 317                 }
 318                 if (reached_end == PQ_L2_SIZE) {
 319                         vm_swapcache_wtrigger = -vm_swapcache_hysteresis;
 320                 }
 321         }
 322
 323         /*
 324          * Cleanup (NOT REACHED)
 325          */
 326         for (q = 0; q < PQ_L2_SIZE; ++q) {
 327                 vm_page_queues_spin_lock(PQ_INACTIVE + q);
 328                 TAILQ_REMOVE(
 329                         &vm_page_queues[PQ_INACTIVE + q].pl,
 330                         &page_marker[q], pageq);
 331                 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
 332         }
 333
 334         lwkt_gettoken(&swindex->token);
 335         TAILQ_REMOVE(&swindex->list, &swmarker.dummy_obj, object_entry);
 336         lwkt_reltoken(&swindex->token);
 337 }
 338
 339 static struct kproc_desc swpc_kp = {
 340         "swapcached",
 341         vm_swapcached_thread,
 342         &swapcached_thread
 343 };
 344 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp);
 345
 346 /*
 347  * Deal with an overflow of the heuristic counter or if the user
 348  * manually changes the hysteresis.
 349  *
 350  * Try to avoid small incremental pageouts by waiting for enough
 351  * pages to buildup in the inactive queue to hopefully get a good
 352  * burst in.  This heuristic is bumped by the VM system and reset
 353  * when our scan hits the end of the queue.
 354  *
 355  * Return TRUE if we need to take a writing pass.
 356  */
 357 static int
 358 vm_swapcache_writing_heuristic(void)
 359 {
 360         int hyst;
 361         int q;
 362         long adds;
 363
 364         hyst = vmstats.v_inactive_count / 4;
 365         if (hyst < vm_swapcache_min_hysteresis)
 366                 hyst = vm_swapcache_min_hysteresis;
 367         cpu_ccfence();
 368         vm_swapcache_hysteresis = hyst;
 369
 370         adds = 0;
 371         for (q = PQ_INACTIVE; q < PQ_INACTIVE + PQ_L2_SIZE; ++q) {
 372                 adds += atomic_swap_long(&vm_page_queues[q].adds, 0);
 373         }
 374         vm_swapcache_wtrigger += adds;
 375         if (vm_swapcache_wtrigger < -hyst)
 376                 vm_swapcache_wtrigger = -hyst;
 377         return (vm_swapcache_wtrigger >= 0);
 378 }
 379
 380 /*
 381  * Take a writing pass on one of the inactive queues, return non-zero if
 382  * we hit the end of the queue.
 383  */
 384 static int
 385 vm_swapcache_writing(vm_page_t marker, int count, int scount)
 386 {
 387         vm_object_t object;
 388         struct vnode *vp;
 389         vm_page_t m;
 390         int isblkdev;
 391
 392         /*
 393          * Scan the inactive queue from our marker to locate
 394          * suitable pages to push to the swap cache.
 395          *
 396          * We are looking for clean vnode-backed pages.
 397          */
 398         vm_page_queues_spin_lock(marker->queue);
 399         while ((m = TAILQ_NEXT(marker, pageq)) != NULL &&
 400                count > 0 && scount-- > 0) {
 401                 KKASSERT(m->queue == marker->queue);
 402
 403                 /*
 404                  * Stop using swap if paniced, dumping, or dumped.
 405                  * Don't try to write if our curburst has been exhausted.
 406                  */
 407                 if (panicstr || dumping)
 408                         break;
 409                 if (vm_swapcache_curburst < 0)
 410                         break;
 411
 412                 /*
 413                  * Move marker
 414                  */
 415                 TAILQ_REMOVE(
 416                         &vm_page_queues[marker->queue].pl, marker, pageq);
 417                 TAILQ_INSERT_AFTER(
 418                         &vm_page_queues[marker->queue].pl, m, marker, pageq);
 419
 420                 /*
 421                  * Ignore markers and ignore pages that already have a swap
 422                  * assignment.
 423                  */
 424                 if (m->flags & (PG_MARKER | PG_SWAPPED))
 425                         continue;
 426                 if (vm_page_busy_try(m, TRUE))
 427                         continue;
 428                 vm_page_queues_spin_unlock(marker->queue);
 429
 430                 if ((object = m->object) == NULL) {
 431                         vm_page_wakeup(m);
 432                         vm_page_queues_spin_lock(marker->queue);
 433                         continue;
 434                 }
 435                 vm_object_hold(object);
 436                 if (m->object != object) {
 437                         vm_object_drop(object);
 438                         vm_page_wakeup(m);
 439                         vm_page_queues_spin_lock(marker->queue);
 440                         continue;
 441                 }
 442                 if (vm_swapcache_test(m)) {
 443                         vm_object_drop(object);
 444                         vm_page_wakeup(m);
 445                         vm_page_queues_spin_lock(marker->queue);
 446                         continue;
 447                 }
 448
 449                 vp = object->handle;
 450                 if (vp == NULL) {
 451                         vm_object_drop(object);
 452                         vm_page_wakeup(m);
 453                         vm_page_queues_spin_lock(marker->queue);
 454                         continue;
 455                 }
 456
 457                 switch(vp->v_type) {
 458                 case VREG:
 459                         /*
 460                          * PG_NOTMETA generically means 'don't swapcache this',
 461                          * and HAMMER will set this for regular data buffers
 462                          * (and leave it unset for meta-data buffers) as
 463                          * appropriate when double buffering is enabled.
 464                          */
 465                         if (m->flags & PG_NOTMETA) {
 466                                 vm_object_drop(object);
 467                                 vm_page_wakeup(m);
 468                                 vm_page_queues_spin_lock(marker->queue);
 469                                 continue;
 470                         }
 471
 472                         /*
 473                          * If data_enable is 0 do not try to swapcache data.
 474                          * If use_chflags is set then only swapcache data for
 475                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 476                          */
 477                         if (vm_swapcache_data_enable == 0 ||
 478                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 479                              vm_swapcache_use_chflags)) {
 480                                 vm_object_drop(object);
 481                                 vm_page_wakeup(m);
 482                                 vm_page_queues_spin_lock(marker->queue);
 483                                 continue;
 484                         }
 485                         if (vm_swapcache_maxfilesize &&
 486                             object->size >
 487                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 488                                 vm_object_drop(object);
 489                                 vm_page_wakeup(m);
 490                                 vm_page_queues_spin_lock(marker->queue);
 491                                 continue;
 492                         }
 493                         isblkdev = 0;
 494                         break;
 495                 case VCHR:
 496                         /*
 497                          * PG_NOTMETA generically means 'don't swapcache this',
 498                          * and HAMMER will set this for regular data buffers
 499                          * (and leave it unset for meta-data buffers) as
 500                          * appropriate when double buffering is enabled.
 501                          */
 502                         if (m->flags & PG_NOTMETA) {
 503                                 vm_object_drop(object);
 504                                 vm_page_wakeup(m);
 505                                 vm_page_queues_spin_lock(marker->queue);
 506                                 continue;
 507                         }
 508                         if (vm_swapcache_meta_enable == 0) {
 509                                 vm_object_drop(object);
 510                                 vm_page_wakeup(m);
 511                                 vm_page_queues_spin_lock(marker->queue);
 512                                 continue;
 513                         }
 514                         isblkdev = 1;
 515                         break;
 516                 default:
 517                         vm_object_drop(object);
 518                         vm_page_wakeup(m);
 519                         vm_page_queues_spin_lock(marker->queue);
 520                         continue;
 521                 }
 522
 523
 524                 /*
 525                  * Assign swap and initiate I/O.
 526                  *
 527                  * (adjust for the --count which also occurs in the loop)
 528                  */
 529                 count -= vm_swapcached_flush(m, isblkdev);
 530
 531                 /*
 532                  * Setup for next loop using marker.
 533                  */
 534                 vm_object_drop(object);
 535                 vm_page_queues_spin_lock(marker->queue);
 536         }
 537
 538         /*
 539          * The marker could wind up at the end, which is ok.  If we hit the
 540          * end of the list adjust the heuristic.
 541          *
 542          * Earlier inactive pages that were dirty and become clean
 543          * are typically moved to the end of PQ_INACTIVE by virtue
 544          * of vfs_vmio_release() when they become unwired from the
 545          * buffer cache.
 546          */
 547         vm_page_queues_spin_unlock(marker->queue);
 548
 549         /*
 550          * m invalid but can be used to test for NULL
 551          */
 552         return (m == NULL);
 553 }
 554
 555 /*
 556  * Flush the specified page using the swap_pager.  The page
 557  * must be busied by the caller and its disposition will become
 558  * the responsibility of this function.
 559  *
 560  * Try to collect surrounding pages, including pages which may
 561  * have already been assigned swap.  Try to cluster within a
 562  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 563  * to match what swap_pager_putpages() can do.
 564  *
 565  * We also want to try to match against the buffer cache blocksize
 566  * but we don't really know what it is here.  Since the buffer cache
 567  * wires and unwires pages in groups the fact that we skip wired pages
 568  * should be sufficient.
 569  *
 570  * Returns a count of pages we might have flushed (minimum 1)
 571  */
 572 static
 573 int
 574 vm_swapcached_flush(vm_page_t m, int isblkdev)
 575 {
 576         vm_object_t object;
 577         vm_page_t marray[SWAP_META_PAGES];
 578         vm_pindex_t basei;
 579         int rtvals[SWAP_META_PAGES];
 580         int x;
 581         int i;
 582         int j;
 583         int count;
 584         int error;
 585
 586         vm_page_io_start(m);
 587         vm_page_protect(m, VM_PROT_READ);
 588         object = m->object;
 589         vm_object_hold(object);
 590
 591         /*
 592          * Try to cluster around (m), keeping in mind that the swap pager
 593          * can only do SMAP_META_PAGES worth of continguous write.
 594          */
 595         x = (int)m->pindex & SWAP_META_MASK;
 596         marray[x] = m;
 597         basei = m->pindex;
 598         vm_page_wakeup(m);
 599
 600         for (i = x - 1; i >= 0; --i) {
 601                 m = vm_page_lookup_busy_try(object, basei - x + i,
 602                                             TRUE, &error);
 603                 if (error || m == NULL)
 604                         break;
 605                 if (vm_swapcache_test(m)) {
 606                         vm_page_wakeup(m);
 607                         break;
 608                 }
 609                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 610                         vm_page_wakeup(m);
 611                         break;
 612                 }
 613                 vm_page_io_start(m);
 614                 vm_page_protect(m, VM_PROT_READ);
 615                 if (m->queue - m->pc == PQ_CACHE) {
 616                         vm_page_unqueue_nowakeup(m);
 617                         vm_page_deactivate(m);
 618                 }
 619                 marray[i] = m;
 620                 vm_page_wakeup(m);
 621         }
 622         ++i;
 623
 624         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 625                 m = vm_page_lookup_busy_try(object, basei - x + j,
 626                                             TRUE, &error);
 627                 if (error || m == NULL)
 628                         break;
 629                 if (vm_swapcache_test(m)) {
 630                         vm_page_wakeup(m);
 631                         break;
 632                 }
 633                 if (isblkdev && (m->flags & PG_NOTMETA)) {
 634                         vm_page_wakeup(m);
 635                         break;
 636                 }
 637                 vm_page_io_start(m);
 638                 vm_page_protect(m, VM_PROT_READ);
 639                 if (m->queue - m->pc == PQ_CACHE) {
 640                         vm_page_unqueue_nowakeup(m);
 641                         vm_page_deactivate(m);
 642                 }
 643                 marray[j] = m;
 644                 vm_page_wakeup(m);
 645         }
 646
 647         count = j - i;
 648         vm_object_pip_add(object, count);
 649         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 650         vm_swapcache_write_count += count * PAGE_SIZE;
 651         vm_swapcache_curburst -= count * PAGE_SIZE;
 652
 653         while (i < j) {
 654                 if (rtvals[i] != VM_PAGER_PEND) {
 655                         vm_page_busy_wait(marray[i], FALSE, "swppgfd");
 656                         vm_page_io_finish(marray[i]);
 657                         vm_page_wakeup(marray[i]);
 658                         vm_object_pip_wakeup(object);
 659                 }
 660                 ++i;
 661         }
 662         vm_object_drop(object);
 663         return(count);
 664 }
 665
 666 /*
 667  * Test whether a VM page is suitable for writing to the swapcache.
 668  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 669  *
 670  * Returns 0 on success, 1 on failure
 671  */
 672 static int
 673 vm_swapcache_test(vm_page_t m)
 674 {
 675         vm_object_t object;
 676
 677         if (m->flags & (PG_UNQUEUED | PG_FICTITIOUS))
 678                 return(1);
 679         if (m->hold_count || m->wire_count)
 680                 return(1);
 681         if (m->valid != VM_PAGE_BITS_ALL)
 682                 return(1);
 683         if (m->dirty & m->valid)
 684                 return(1);
 685         if ((object = m->object) == NULL)
 686                 return(1);
 687         if (object->type != OBJT_VNODE ||
 688             (object->flags & OBJ_DEAD)) {
 689                 return(1);
 690         }
 691         vm_page_test_dirty(m);
 692         if (m->dirty & m->valid)
 693                 return(1);
 694         return(0);
 695 }
 696
 697 /*
 698  * Cleaning pass.
 699  *
 700  * We clean whole objects up to 16MB
 701  */
 702 static
 703 void
 704 vm_swapcache_cleaning(swmarker_t *marker, struct vm_object_hash **swindexp)
 705 {
 706         vm_object_t object;
 707         struct vnode *vp;
 708         int count;
 709         int scount;
 710         int n;
 711         int didmove;
 712
 713         count = vm_swapcache_maxlaunder;
 714         scount = vm_swapcache_maxscan;
 715
 716         /*
 717          * Look for vnode objects
 718          */
 719         lwkt_gettoken(&(*swindexp)->token);
 720
 721         didmove = 0;
 722 outerloop:
 723         while ((object = TAILQ_NEXT(&marker->dummy_obj,
 724                                     object_entry)) != NULL) {
 725                 /*
 726                  * We have to skip markers.  We cannot hold/drop marker
 727                  * objects!
 728                  */
 729                 if (object->type == OBJT_MARKER) {
 730                         vm_swapcache_movemarker(marker, *swindexp, object);
 731                         didmove = 1;
 732                         continue;
 733                 }
 734
 735                 /*
 736                  * Safety, or in case there are millions of VM objects
 737                  * without swapcache backing.
 738                  */
 739                 if (--scount <= 0)
 740                         goto breakout;
 741
 742                 /*
 743                  * We must hold the object before potentially yielding.
 744                  */
 745                 vm_object_hold(object);
 746                 lwkt_yield();
 747
 748                 /*
 749                  * Only operate on live VNODE objects that are either
 750                  * VREG or VCHR (VCHR for meta-data).
 751                  */
 752                 if ((object->type != OBJT_VNODE) ||
 753                     ((object->flags & OBJ_DEAD) ||
 754                      object->swblock_count == 0) ||
 755                     ((vp = object->handle) == NULL) ||
 756                     (vp->v_type != VREG && vp->v_type != VCHR)) {
 757                         vm_object_drop(object);
 758                         /* object may be invalid now */
 759                         vm_swapcache_movemarker(marker, *swindexp, object);
 760                         didmove = 1;
 761                         continue;
 762                 }
 763
 764                 /*
 765                  * Reset the object pindex stored in the marker if the
 766                  * working object has changed.
 767                  */
 768                 if (marker->save_obj != object || didmove) {
 769                         marker->dummy_obj.size = 0;
 770                         marker->save_off = 0;
 771                         marker->save_obj = object;
 772                         didmove = 0;
 773                 }
 774
 775                 /*
 776                  * Look for swblocks starting at our iterator.
 777                  *
 778                  * The swap_pager_condfree() function attempts to free
 779                  * swap space starting at the specified index.  The index
 780                  * will be updated on return.  The function will return
 781                  * a scan factor (NOT the number of blocks freed).
 782                  *
 783                  * If it must cut its scan of the object short due to an
 784                  * excessive number of swblocks, or is able to free the
 785                  * requested number of blocks, it will return n >= count
 786                  * and we break and pick it back up on a future attempt.
 787                  *
 788                  * Scan the object linearly and try to batch large sets of
 789                  * blocks that are likely to clean out entire swap radix
 790                  * tree leafs.
 791                  */
 792                 lwkt_token_swap();
 793                 lwkt_reltoken(&(*swindexp)->token);
 794
 795                 n = swap_pager_condfree(object, &marker->dummy_obj.size,
 796                                     (count + SWAP_META_MASK) & ~SWAP_META_MASK);
 797
 798                 vm_object_drop(object);         /* object may be invalid now */
 799                 lwkt_gettoken(&(*swindexp)->token);
 800
 801                 /*
 802                  * If we have exhausted the object or deleted our per-pass
 803                  * page limit then move us to the next object.  Note that
 804                  * the current object may no longer be on the vm_object_entry.
 805                  */
 806                 if (n <= 0 ||
 807                     marker->save_off > vm_swapcache_cleanperobj) {
 808                         vm_swapcache_movemarker(marker, *swindexp, object);
 809                         didmove = 1;
 810                 }
 811
 812                 /*
 813                  * If we have exhausted our max-launder stop for now.
 814                  */
 815                 count -= n;
 816                 marker->save_off += n * PAGE_SIZE;
 817                 if (count < 0)
 818                         goto breakout;
 819         }
 820
 821         /*
 822          * Iterate vm_object_hash[] hash table
 823          */
 824         TAILQ_REMOVE(&(*swindexp)->list, &marker->dummy_obj, object_entry);
 825         lwkt_reltoken(&(*swindexp)->token);
 826         if (++*swindexp >= &vm_object_hash[VMOBJ_HSIZE])
 827                 *swindexp = &vm_object_hash[0];
 828         lwkt_gettoken(&(*swindexp)->token);
 829         TAILQ_INSERT_HEAD(&(*swindexp)->list, &marker->dummy_obj, object_entry);
 830
 831         if (*swindexp != &vm_object_hash[0])
 832                 goto outerloop;
 833
 834 breakout:
 835         lwkt_reltoken(&(*swindexp)->token);
 836 }
 837
 838 /*
 839  * Move the marker past the current object.  Object can be stale, but we
 840  * still need it to determine if the marker has to be moved.  If the object
 841  * is still the 'current object' (object after the marker), we hop-scotch
 842  * the marker past it.
 843  */
 844 static void
 845 vm_swapcache_movemarker(swmarker_t *marker, struct vm_object_hash *swindex,
 846                         vm_object_t object)
 847 {
 848         if (TAILQ_NEXT(&marker->dummy_obj, object_entry) == object) {
 849                 TAILQ_REMOVE(&swindex->list, &marker->dummy_obj, object_entry);
 850                 TAILQ_INSERT_AFTER(&swindex->list, object,
 851                                    &marker->dummy_obj, object_entry);
 852         }
 853 }