sys/vm/vm_swapcache.c

   1 /*
   2  * Copyright (c) 2010 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34
  35 /*
  36  * Implement the swapcache daemon.  When enabled swap is assumed to be
  37  * configured on a fast storage device such as a SSD.  Swap is assigned
  38  * to clean vnode-backed pages in the inactive queue, clustered by object
  39  * if possible, and written out.  The swap assignment sticks around even
  40  * after the underlying pages have been recycled.
  41  *
  42  * The daemon manages write bandwidth based on sysctl settings to control
  43  * wear on the SSD.
  44  *
  45  * The vnode strategy code will check for the swap assignments and divert
  46  * reads to the swap device when the data is present in the swapcache.
  47  *
  48  * This operates on both regular files and the block device vnodes used by
  49  * filesystems to manage meta-data.
  50  */
  51
  52 #include "opt_vm.h"
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/kernel.h>
  56 #include <sys/proc.h>
  57 #include <sys/kthread.h>
  58 #include <sys/resourcevar.h>
  59 #include <sys/signalvar.h>
  60 #include <sys/vnode.h>
  61 #include <sys/vmmeter.h>
  62 #include <sys/sysctl.h>
  63
  64 #include <vm/vm.h>
  65 #include <vm/vm_param.h>
  66 #include <sys/lock.h>
  67 #include <vm/vm_object.h>
  68 #include <vm/vm_page.h>
  69 #include <vm/vm_map.h>
  70 #include <vm/vm_pageout.h>
  71 #include <vm/vm_pager.h>
  72 #include <vm/swap_pager.h>
  73 #include <vm/vm_extern.h>
  74
  75 #include <sys/thread2.h>
  76 #include <vm/vm_page2.h>
  77
  78 #define INACTIVE_LIST   (&vm_page_queues[PQ_INACTIVE].pl)
  79
  80 /* the kernel process "vm_pageout"*/
  81 static void vm_swapcached (void);
  82 static int vm_swapcached_flush (vm_page_t m);
  83 static int vm_swapcache_test(vm_page_t m);
  84 static void vm_swapcache_writing(vm_page_t marker);
  85 static void vm_swapcache_cleaning(vm_object_t marker);
  86 struct thread *swapcached_thread;
  87
  88 static struct kproc_desc swpc_kp = {
  89         "swapcached",
  90         vm_swapcached,
  91         &swapcached_thread
  92 };
  93 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
  94
  95 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  96
  97 int vm_swapcache_read_enable;
  98 int vm_swapcache_inactive_heuristic;
  99 static int vm_swapcache_sleep;
 100 static int vm_swapcache_maxlaunder = 256;
 101 static int vm_swapcache_data_enable = 0;
 102 static int vm_swapcache_meta_enable = 0;
 103 static int vm_swapcache_maxswappct = 75;
 104 static int vm_swapcache_hysteresis;
 105 static 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
 113 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 114         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 115
 116 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 117         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 118 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 119         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 120 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 121         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 122 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
 123         CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
 124 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
 125         CTLFLAG_RW, &vm_swapcache_hysteresis, 0, "");
 126 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
 127         CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
 128
 129 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
 130         CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
 131 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 132         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 134         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 135 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
 136         CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
 137 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 138         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 139 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 140         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 141
 142 #define SWAPMAX(adj)    \
 143         ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
 144
 145 /*
 146  * vm_swapcached is the high level pageout daemon.
 147  */
 148 static void
 149 vm_swapcached(void)
 150 {
 151         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 152         enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
 153         struct vm_page page_marker;
 154         struct vm_object object_marker;
 155
 156         /*
 157          * Thread setup
 158          */
 159         curthread->td_flags |= TDF_SYSTHREAD;
 160         crit_enter();
 161
 162         /*
 163          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 164          */
 165         bzero(&page_marker, sizeof(page_marker));
 166         page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 167         page_marker.queue = PQ_INACTIVE;
 168         page_marker.wire_count = 1;
 169         TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq);
 170         vm_swapcache_hysteresis = vmstats.v_inactive_target / 2;
 171         vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 172
 173         /*
 174          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 175          */
 176         bzero(&object_marker, sizeof(object_marker));
 177         object_marker.type = OBJT_MARKER;
 178         TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
 179
 180         for (;;) {
 181                 /*
 182                  * Check every 5 seconds when not enabled or if no swap
 183                  * is present.
 184                  */
 185                 if ((vm_swapcache_data_enable == 0 &&
 186                      vm_swapcache_meta_enable == 0) ||
 187                     vm_swap_max == 0) {
 188                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 189                         continue;
 190                 }
 191
 192                 /*
 193                  * Polling rate when enabled is approximately 10 hz.
 194                  */
 195                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 196
 197                 /*
 198                  * State hysteresis.  Generate write activity up to 75% of
 199                  * swap, then clean out swap assignments down to 70%, then
 200                  * repeat.
 201                  */
 202                 if (state == SWAPC_WRITING) {
 203                         if (vm_swap_cache_use > SWAPMAX(0))
 204                                 state = SWAPC_CLEANING;
 205                 } else {
 206                         if (vm_swap_cache_use < SWAPMAX(-5))
 207                                 state = SWAPC_WRITING;
 208                 }
 209
 210                 /*
 211                  * We are allowed to continue accumulating burst value
 212                  * in either state.  Allow the user to set curburst > maxburst
 213                  * for the initial load-in.
 214                  */
 215                 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
 216                         vm_swapcache_curburst += vm_swapcache_accrate / 10;
 217                         if (vm_swapcache_curburst > vm_swapcache_maxburst)
 218                                 vm_swapcache_curburst = vm_swapcache_maxburst;
 219                 }
 220
 221                 /*
 222                  * We don't want to nickle-and-dime the scan as that will
 223                  * create unnecessary fragmentation.  The minimum burst
 224                  * is one-seconds worth of accumulation.
 225                  */
 226                 if (state == SWAPC_WRITING) {
 227                         if (vm_swapcache_curburst >= vm_swapcache_accrate) {
 228                                 if (burst == SWAPB_BURSTING) {
 229                                         vm_swapcache_writing(&page_marker);
 230                                         if (vm_swapcache_curburst <= 0)
 231                                                 burst = SWAPB_RECOVERING;
 232                                 } else if (vm_swapcache_curburst >
 233                                            vm_swapcache_minburst) {
 234                                         vm_swapcache_writing(&page_marker);
 235                                         burst = SWAPB_BURSTING;
 236                                 }
 237                         }
 238                 } else {
 239                         vm_swapcache_cleaning(&object_marker);
 240                 }
 241         }
 242         TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq);
 243         TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
 244         crit_exit();
 245 }
 246
 247 static void
 248 vm_swapcache_writing(vm_page_t marker)
 249 {
 250         vm_object_t object;
 251         struct vnode *vp;
 252         vm_page_t m;
 253         int count;
 254
 255         /*
 256          * Try to avoid small incremental pageouts by waiting for enough
 257          * pages to buildup in the inactive queue to hopefully get a good
 258          * burst in.  This heuristic is bumped by the VM system and reset
 259          * when our scan hits the end of the queue.
 260          */
 261         if (vm_swapcache_inactive_heuristic < 0)
 262                 return;
 263
 264         /*
 265          * Scan the inactive queue from our marker to locate
 266          * suitable pages to push to the swap cache.
 267          *
 268          * We are looking for clean vnode-backed pages.
 269          *
 270          * NOTE: PG_SWAPPED pages in particular are not part of
 271          *       our count because once the cache stabilizes we
 272          *       can end up with a very high datarate of VM pages
 273          *       cycling from it.
 274          */
 275         m = marker;
 276         count = vm_swapcache_maxlaunder;
 277
 278         while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) {
 279                 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
 280                         ++count;
 281                         continue;
 282                 }
 283                 if (vm_swapcache_curburst < 0)
 284                         break;
 285                 if (vm_swapcache_test(m))
 286                         continue;
 287                 object = m->object;
 288                 vp = object->handle;
 289                 if (vp == NULL)
 290                         continue;
 291
 292                 switch(vp->v_type) {
 293                 case VREG:
 294                         /*
 295                          * If data_enable is 0 do not try to swapcache data.
 296                          * If use_chflags is set then only swapcache data for
 297                          * VSWAPCACHE marked vnodes, otherwise any vnode.
 298                          */
 299                         if (vm_swapcache_data_enable == 0 ||
 300                             ((vp->v_flag & VSWAPCACHE) == 0 &&
 301                              vm_swapcache_use_chflags)) {
 302                                 continue;
 303                         }
 304                         if (vm_swapcache_maxfilesize &&
 305                             object->size >
 306                             (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
 307                                 continue;
 308                         }
 309                         break;
 310                 case VCHR:
 311                         if (vm_swapcache_meta_enable == 0)
 312                                 continue;
 313                         break;
 314                 default:
 315                         continue;
 316                 }
 317
 318                 /*
 319                  * Ok, move the marker and soft-busy the page.
 320                  */
 321                 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
 322                 TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq);
 323
 324                 /*
 325                  * Assign swap and initiate I/O.
 326                  *
 327                  * (adjust for the --count which also occurs in the loop)
 328                  */
 329                 count -= vm_swapcached_flush(m) - 1;
 330
 331                 /*
 332                  * Setup for next loop using marker.
 333                  */
 334                 m = marker;
 335         }
 336
 337         /*
 338          * Cleanup marker position.  If we hit the end of the
 339          * list the marker is placed at the tail.  Newly deactivated
 340          * pages will be placed after it.
 341          *
 342          * Earlier inactive pages that were dirty and become clean
 343          * are typically moved to the end of PQ_INACTIVE by virtue
 344          * of vfs_vmio_release() when they become unwired from the
 345          * buffer cache.
 346          */
 347         TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
 348         if (m) {
 349                 TAILQ_INSERT_BEFORE(m, marker, pageq);
 350         } else {
 351                 TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq);
 352                 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
 353         }
 354 }
 355
 356 /*
 357  * Flush the specified page using the swap_pager.
 358  *
 359  * Try to collect surrounding pages, including pages which may
 360  * have already been assigned swap.  Try to cluster within a
 361  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
 362  * to match what swap_pager_putpages() can do.
 363  *
 364  * We also want to try to match against the buffer cache blocksize
 365  * but we don't really know what it is here.  Since the buffer cache
 366  * wires and unwires pages in groups the fact that we skip wired pages
 367  * should be sufficient.
 368  *
 369  * Returns a count of pages we might have flushed (minimum 1)
 370  */
 371 static
 372 int
 373 vm_swapcached_flush(vm_page_t m)
 374 {
 375         vm_object_t object;
 376         vm_page_t marray[SWAP_META_PAGES];
 377         vm_pindex_t basei;
 378         int rtvals[SWAP_META_PAGES];
 379         int x;
 380         int i;
 381         int j;
 382         int count;
 383
 384         vm_page_io_start(m);
 385         vm_page_protect(m, VM_PROT_READ);
 386         object = m->object;
 387
 388         /*
 389          * Try to cluster around (m), keeping in mind that the swap pager
 390          * can only do SMAP_META_PAGES worth of continguous write.
 391          */
 392         x = (int)m->pindex & SWAP_META_MASK;
 393         marray[x] = m;
 394         basei = m->pindex;
 395
 396         for (i = x - 1; i >= 0; --i) {
 397                 m = vm_page_lookup(object, basei - x + i);
 398                 if (m == NULL)
 399                         break;
 400                 if (vm_swapcache_test(m))
 401                         break;
 402                 vm_page_io_start(m);
 403                 vm_page_protect(m, VM_PROT_READ);
 404                 if (m->queue - m->pc == PQ_CACHE) {
 405                         vm_page_unqueue_nowakeup(m);
 406                         vm_page_deactivate(m);
 407                 }
 408                 marray[i] = m;
 409         }
 410         ++i;
 411
 412         for (j = x + 1; j < SWAP_META_PAGES; ++j) {
 413                 m = vm_page_lookup(object, basei - x + j);
 414                 if (m == NULL)
 415                         break;
 416                 if (vm_swapcache_test(m))
 417                         break;
 418                 vm_page_io_start(m);
 419                 vm_page_protect(m, VM_PROT_READ);
 420                 if (m->queue - m->pc == PQ_CACHE) {
 421                         vm_page_unqueue_nowakeup(m);
 422                         vm_page_deactivate(m);
 423                 }
 424                 marray[j] = m;
 425         }
 426
 427         count = j - i;
 428         vm_object_pip_add(object, count);
 429         swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
 430         vm_swapcache_write_count += count * PAGE_SIZE;
 431         vm_swapcache_curburst -= count * PAGE_SIZE;
 432
 433         while (i < j) {
 434                 if (rtvals[i] != VM_PAGER_PEND) {
 435                         vm_page_io_finish(marray[i]);
 436                         vm_object_pip_wakeup(object);
 437                 }
 438                 ++i;
 439         }
 440         return(count);
 441 }
 442
 443 /*
 444  * Test whether a VM page is suitable for writing to the swapcache.
 445  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
 446  *
 447  * Returns 0 on success, 1 on failure
 448  */
 449 static int
 450 vm_swapcache_test(vm_page_t m)
 451 {
 452         vm_object_t object;
 453
 454         if (m->flags & (PG_BUSY | PG_UNMANAGED | PG_NOTMETA))
 455                 return(1);
 456         if (m->busy || m->hold_count || m->wire_count)
 457                 return(1);
 458         if (m->valid != VM_PAGE_BITS_ALL)
 459                 return(1);
 460         if (m->dirty & m->valid)
 461                 return(1);
 462         if ((object = m->object) == NULL)
 463                 return(1);
 464         if (object->type != OBJT_VNODE ||
 465             (object->flags & OBJ_DEAD)) {
 466                 return(1);
 467         }
 468         vm_page_test_dirty(m);
 469         if (m->dirty & m->valid)
 470                 return(1);
 471         return(0);
 472 }
 473
 474 /*
 475  * Cleaning pass
 476  */
 477 static
 478 void
 479 vm_swapcache_cleaning(vm_object_t marker)
 480 {
 481         vm_object_t object;
 482         struct vnode *vp;
 483         int count;
 484         int n;
 485
 486         object = marker;
 487         count = vm_swapcache_maxlaunder;
 488
 489         /*
 490          * Look for vnode objects
 491          */
 492         while ((object = TAILQ_NEXT(object, object_list)) != NULL && count--) {
 493                 if (object->type != OBJT_VNODE)
 494                         continue;
 495                 if ((object->flags & OBJ_DEAD) || object->swblock_count == 0)
 496                         continue;
 497                 if ((vp = object->handle) == NULL)
 498                         continue;
 499                 if (vp->v_type != VREG && vp->v_type != VCHR)
 500                         continue;
 501
 502                 /*
 503                  * Adjust iterator.
 504                  */
 505                 if (marker->backing_object != object)
 506                         marker->size = 0;
 507
 508                 /*
 509                  * Move the marker so we can work on the VM object
 510                  */
 511                 TAILQ_REMOVE(&vm_object_list, marker, object_list);
 512                 TAILQ_INSERT_AFTER(&vm_object_list, object,
 513                                    marker, object_list);
 514
 515                 /*
 516                  * Look for swblocks starting at our iterator.
 517                  *
 518                  * The swap_pager_condfree() function attempts to free
 519                  * swap space starting at the specified index.  The index
 520                  * will be updated on return.  The function will return
 521                  * a scan factor (NOT the number of blocks freed).
 522                  *
 523                  * If it must cut its scan of the object short due to an
 524                  * excessive number of swblocks, or is able to free the
 525                  * requested number of blocks, it will return n >= count
 526                  * and we break and pick it back up on a future attempt.
 527                  */
 528                 n = swap_pager_condfree(object, &marker->size, count);
 529                 count -= n;
 530                 if (count < 0)
 531                         break;
 532
 533                 /*
 534                  * Setup for loop.
 535                  */
 536                 marker->size = 0;
 537                 object = marker;
 538         }
 539
 540         /*
 541          * Adjust marker so we continue the scan from where we left off.
 542          * When we reach the end we start back at the beginning.
 543          */
 544         TAILQ_REMOVE(&vm_object_list, marker, object_list);
 545         if (object)
 546                 TAILQ_INSERT_BEFORE(object, marker, object_list);
 547         else
 548                 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
 549         marker->backing_object = object;
 550 }