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.
  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 void vm_swapcached_flush (vm_page_t m);
  83 static void vm_swapcache_writing(vm_page_t marker);
  84 static void vm_swapcache_cleaning(vm_object_t marker);
  85 struct thread *swapcached_thread;
  86
  87 static struct kproc_desc swpc_kp = {
  88         "swapcached",
  89         vm_swapcached,
  90         &swapcached_thread
  91 };
  92 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
  93
  94 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
  95
  96 int vm_swapcache_read_enable;
  97 static int vm_swapcache_sleep;
  98 static int vm_swapcache_maxlaunder = 256;
  99 static int vm_swapcache_data_enable = 0;
 100 static int vm_swapcache_meta_enable = 0;
 101 static int64_t vm_swapcache_curburst = 1000000000LL;
 102 static int64_t vm_swapcache_maxburst = 1000000000LL;
 103 static int64_t vm_swapcache_accrate = 1000000LL;
 104 static int64_t vm_swapcache_write_count;
 105
 106 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
 107         CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
 108
 109 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
 110         CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
 111 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
 112         CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
 113 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
 114         CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
 115
 116 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
 117         CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
 118 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
 119         CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
 120 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
 121         CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
 122 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
 123         CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
 124
 125 /*
 126  * vm_swapcached is the high level pageout daemon.
 127  */
 128 static void
 129 vm_swapcached(void)
 130 {
 131         enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
 132         struct vm_page page_marker;
 133         struct vm_object object_marker;
 134
 135         /*
 136          * Thread setup
 137          */
 138         curthread->td_flags |= TDF_SYSTHREAD;
 139         crit_enter();
 140
 141         /*
 142          * Initialize our marker for the inactive scan (SWAPC_WRITING)
 143          */
 144         bzero(&page_marker, sizeof(page_marker));
 145         page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
 146         page_marker.queue = PQ_INACTIVE;
 147         page_marker.wire_count = 1;
 148         TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq);
 149
 150         /*
 151          * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
 152          */
 153         bzero(&object_marker, sizeof(object_marker));
 154         object_marker.type = OBJT_MARKER;
 155         TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
 156
 157         for (;;) {
 158                 /*
 159                  * Loop once a second or so looking for work when enabled.
 160                  */
 161                 if (vm_swapcache_data_enable == 0 &&
 162                     vm_swapcache_meta_enable == 0) {
 163                         tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
 164                         continue;
 165                 }
 166
 167                 /*
 168                  * Polling rate when enabled is 10 hz.
 169                  */
 170                 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
 171
 172                 /*
 173                  * State hysteresis.  Generate write activity up to 75% of
 174                  * swap, then clean out swap assignments down to 70%, then
 175                  * repeat.
 176                  */
 177                 if (state == SWAPC_WRITING) {
 178                         if (vm_swap_cache_use > (int64_t)vm_swap_max * 75 / 100)
 179                                 state = SWAPC_CLEANING;
 180                 } else {
 181                         if (vm_swap_cache_use < (int64_t)vm_swap_max * 70 / 100)
 182                                 state = SWAPC_WRITING;
 183                 }
 184
 185                 /*
 186                  * We are allowed to continue accumulating burst value
 187                  * in either state.
 188                  */
 189                 vm_swapcache_curburst += vm_swapcache_accrate / 10;
 190                 if (vm_swapcache_curburst > vm_swapcache_maxburst)
 191                         vm_swapcache_curburst = vm_swapcache_maxburst;
 192
 193                 /*
 194                  * We don't want to nickle-and-dime the scan as that will
 195                  * create unnecessary fragmentation.  The minimum burst
 196                  * is one-seconds worth of accumulation.
 197                  */
 198                 if (state == SWAPC_WRITING) {
 199                         if (vm_swapcache_curburst >= vm_swapcache_accrate)
 200                                 vm_swapcache_writing(&page_marker);
 201                 } else {
 202                         vm_swapcache_cleaning(&object_marker);
 203                 }
 204         }
 205         TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq);
 206         TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
 207         crit_exit();
 208 }
 209
 210 static void
 211 vm_swapcache_writing(vm_page_t marker)
 212 {
 213         vm_object_t object;
 214         struct vnode *vp;
 215         vm_page_t m;
 216         int count;
 217
 218         /*
 219          * Scan the inactive queue from our marker to locate
 220          * suitable pages to push to the swap cache.
 221          *
 222          * We are looking for clean vnode-backed pages.
 223          *
 224          * NOTE: PG_SWAPPED pages in particular are not part of
 225          *       our count because once the cache stabilizes we
 226          *       can end up with a very high datarate of VM pages
 227          *       cycling from it.
 228          */
 229         m = marker;
 230         count = vm_swapcache_maxlaunder;
 231
 232         while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) {
 233                 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
 234                         ++count;
 235                         continue;
 236                 }
 237                 if (vm_swapcache_curburst < 0)
 238                         break;
 239                 if (m->flags & (PG_BUSY | PG_UNMANAGED))
 240                         continue;
 241                 if (m->busy || m->hold_count || m->wire_count)
 242                         continue;
 243                 if (m->valid != VM_PAGE_BITS_ALL)
 244                         continue;
 245                 if (m->dirty & m->valid)
 246                         continue;
 247                 if ((object = m->object) == NULL)
 248                         continue;
 249                 if (object->type != OBJT_VNODE ||
 250                     (object->flags & OBJ_DEAD)) {
 251                         continue;
 252                 }
 253                 vm_page_test_dirty(m);
 254                 if (m->dirty & m->valid)
 255                         continue;
 256                 vp = object->handle;
 257                 if (vp == NULL)
 258                         continue;
 259                 switch(vp->v_type) {
 260                 case VREG:
 261                         if (vm_swapcache_data_enable == 0)
 262                                 continue;
 263                         break;
 264                 case VCHR:
 265                         if (vm_swapcache_meta_enable == 0)
 266                                 continue;
 267                         break;
 268                 default:
 269                         continue;
 270                 }
 271
 272                 /*
 273                  * Ok, move the marker and soft-busy the page.
 274                  */
 275                 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
 276                 TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq);
 277
 278                 /*
 279                  * Assign swap and initiate I/O
 280                  */
 281                 vm_swapcached_flush(m);
 282
 283                 /*
 284                  * Setup for next loop using marker.
 285                  */
 286                 m = marker;
 287         }
 288
 289         /*
 290          * Cleanup marker position.  If we hit the end of the
 291          * list the marker is placed at the tail.  Newly deactivated
 292          * pages will be placed after it.
 293          *
 294          * Earlier inactive pages that were dirty and become clean
 295          * are typically moved to the end of PQ_INACTIVE by virtue
 296          * of vfs_vmio_release() when they become unwired from the
 297          * buffer cache.
 298          */
 299         TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
 300         if (m)
 301                 TAILQ_INSERT_BEFORE(m, marker, pageq);
 302         else
 303                 TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq);
 304 }
 305
 306 /*
 307  * Flush the specified page using the swap_pager.
 308  */
 309 static
 310 void
 311 vm_swapcached_flush(vm_page_t m)
 312 {
 313         vm_object_t object;
 314         int rtvals;
 315
 316         vm_page_io_start(m);
 317         vm_page_protect(m, VM_PROT_READ);
 318
 319         object = m->object;
 320         vm_object_pip_add(object, 1);
 321         swap_pager_putpages(object, &m, 1, FALSE, &rtvals);
 322         vm_swapcache_write_count += PAGE_SIZE;
 323         vm_swapcache_curburst -= PAGE_SIZE;
 324
 325         if (rtvals != VM_PAGER_PEND) {
 326                 vm_object_pip_wakeup(object);
 327                 vm_page_io_finish(m);
 328         }
 329 }
 330
 331 static
 332 void
 333 vm_swapcache_cleaning(vm_object_t marker)
 334 {
 335         vm_object_t object;
 336         struct vnode *vp;
 337         int count;
 338         int n;
 339
 340         object = marker;
 341         count = vm_swapcache_maxlaunder;
 342
 343         /*
 344          * Look for vnode objects
 345          */
 346         while ((object = TAILQ_NEXT(object, object_list)) != NULL && count--) {
 347                 if (object->type != OBJT_VNODE)
 348                         continue;
 349                 if ((object->flags & OBJ_DEAD) || object->swblock_count == 0)
 350                         continue;
 351                 if ((vp = object->handle) == NULL)
 352                         continue;
 353                 if (vp->v_type != VREG && vp->v_type != VCHR)
 354                         continue;
 355
 356                 /*
 357                  * Adjust iterator.
 358                  */
 359                 if (marker->backing_object != object)
 360                         marker->size = 0;
 361
 362                 /*
 363                  * Move the marker so we can work on the VM object
 364                  */
 365                 TAILQ_REMOVE(&vm_object_list, marker, object_list);
 366                 TAILQ_INSERT_AFTER(&vm_object_list, object,
 367                                    marker, object_list);
 368
 369                 /*
 370                  * Look for swblocks starting at our iterator.
 371                  *
 372                  * The swap_pager_condfree() function attempts to free
 373                  * swap space starting at the specified index.  The index
 374                  * will be updated on return.  The function will return
 375                  * a scan factor (NOT the number of blocks freed).
 376                  *
 377                  * If it must cut its scan of the object short due to an
 378                  * excessive number of swblocks, or is able to free the
 379                  * requested number of blocks, it will return n >= count
 380                  * and we break and pick it back up on a future attempt.
 381                  */
 382                 n = swap_pager_condfree(object, &marker->size, count);
 383                 count -= n;
 384                 if (count < 0)
 385                         break;
 386
 387                 /*
 388                  * Setup for loop.
 389                  */
 390                 marker->size = 0;
 391                 object = marker;
 392         }
 393
 394         /*
 395          * Adjust marker so we continue the scan from where we left off.
 396          * When we reach the end we start back at the beginning.
 397          */
 398         TAILQ_REMOVE(&vm_object_list, marker, object_list);
 399         if (object)
 400                 TAILQ_INSERT_BEFORE(object, marker, object_list);
 401         else
 402                 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
 403         marker->backing_object = object;
 404 }