1096e95c0SMatthew Dillon /* 2096e95c0SMatthew Dillon * Copyright (c) 2010 The DragonFly Project. All rights reserved. 3096e95c0SMatthew Dillon * 4096e95c0SMatthew Dillon * This code is derived from software contributed to The DragonFly Project 5096e95c0SMatthew Dillon * by Matthew Dillon <dillon@backplane.com> 6096e95c0SMatthew Dillon * 7096e95c0SMatthew Dillon * Redistribution and use in source and binary forms, with or without 8096e95c0SMatthew Dillon * modification, are permitted provided that the following conditions 9096e95c0SMatthew Dillon * are met: 10096e95c0SMatthew Dillon * 11096e95c0SMatthew Dillon * 1. Redistributions of source code must retain the above copyright 12096e95c0SMatthew Dillon * notice, this list of conditions and the following disclaimer. 13096e95c0SMatthew Dillon * 2. Redistributions in binary form must reproduce the above copyright 14096e95c0SMatthew Dillon * notice, this list of conditions and the following disclaimer in 15096e95c0SMatthew Dillon * the documentation and/or other materials provided with the 16096e95c0SMatthew Dillon * distribution. 17096e95c0SMatthew Dillon * 3. Neither the name of The DragonFly Project nor the names of its 18096e95c0SMatthew Dillon * contributors may be used to endorse or promote products derived 19096e95c0SMatthew Dillon * from this software without specific, prior written permission. 20096e95c0SMatthew Dillon * 21096e95c0SMatthew Dillon * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22096e95c0SMatthew Dillon * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23096e95c0SMatthew Dillon * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24096e95c0SMatthew Dillon * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25096e95c0SMatthew Dillon * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26096e95c0SMatthew Dillon * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27096e95c0SMatthew Dillon * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28096e95c0SMatthew Dillon * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29096e95c0SMatthew Dillon * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30096e95c0SMatthew Dillon * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31096e95c0SMatthew Dillon * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32096e95c0SMatthew Dillon * SUCH DAMAGE. 33096e95c0SMatthew Dillon */ 34096e95c0SMatthew Dillon 35096e95c0SMatthew Dillon /* 36096e95c0SMatthew Dillon * Implement the swapcache daemon. When enabled swap is assumed to be 37096e95c0SMatthew Dillon * configured on a fast storage device such as a SSD. Swap is assigned 38096e95c0SMatthew Dillon * to clean vnode-backed pages in the inactive queue, clustered by object 39096e95c0SMatthew Dillon * if possible, and written out. The swap assignment sticks around even 40096e95c0SMatthew Dillon * after the underlying pages have been recycled. 41096e95c0SMatthew Dillon * 42096e95c0SMatthew Dillon * The daemon manages write bandwidth based on sysctl settings to control 43096e95c0SMatthew Dillon * wear on the SSD. 44096e95c0SMatthew Dillon * 45096e95c0SMatthew Dillon * The vnode strategy code will check for the swap assignments and divert 463ffc7051SMatthew Dillon * reads to the swap device when the data is present in the swapcache. 47096e95c0SMatthew Dillon * 48096e95c0SMatthew Dillon * This operates on both regular files and the block device vnodes used by 49096e95c0SMatthew Dillon * filesystems to manage meta-data. 50096e95c0SMatthew Dillon */ 51096e95c0SMatthew Dillon 52096e95c0SMatthew Dillon #include "opt_vm.h" 53096e95c0SMatthew Dillon #include <sys/param.h> 54096e95c0SMatthew Dillon #include <sys/systm.h> 55096e95c0SMatthew Dillon #include <sys/kernel.h> 56096e95c0SMatthew Dillon #include <sys/proc.h> 57096e95c0SMatthew Dillon #include <sys/kthread.h> 58096e95c0SMatthew Dillon #include <sys/resourcevar.h> 59096e95c0SMatthew Dillon #include <sys/signalvar.h> 60096e95c0SMatthew Dillon #include <sys/vnode.h> 61096e95c0SMatthew Dillon #include <sys/vmmeter.h> 62096e95c0SMatthew Dillon #include <sys/sysctl.h> 63096e95c0SMatthew Dillon 64096e95c0SMatthew Dillon #include <vm/vm.h> 65096e95c0SMatthew Dillon #include <vm/vm_param.h> 66096e95c0SMatthew Dillon #include <sys/lock.h> 67096e95c0SMatthew Dillon #include <vm/vm_object.h> 68096e95c0SMatthew Dillon #include <vm/vm_page.h> 69096e95c0SMatthew Dillon #include <vm/vm_map.h> 70096e95c0SMatthew Dillon #include <vm/vm_pageout.h> 71096e95c0SMatthew Dillon #include <vm/vm_pager.h> 72096e95c0SMatthew Dillon #include <vm/swap_pager.h> 73096e95c0SMatthew Dillon #include <vm/vm_extern.h> 74096e95c0SMatthew Dillon 75096e95c0SMatthew Dillon #include <sys/thread2.h> 76096e95c0SMatthew Dillon #include <vm/vm_page2.h> 77096e95c0SMatthew Dillon 78096e95c0SMatthew Dillon #define INACTIVE_LIST (&vm_page_queues[PQ_INACTIVE].pl) 79096e95c0SMatthew Dillon 80096e95c0SMatthew Dillon /* the kernel process "vm_pageout"*/ 81096e95c0SMatthew Dillon static void vm_swapcached (void); 823ffc7051SMatthew Dillon static int vm_swapcached_flush (vm_page_t m); 833ffc7051SMatthew Dillon static int vm_swapcache_test(vm_page_t m); 8400a3fdcaSMatthew Dillon static void vm_swapcache_writing(vm_page_t marker); 8500a3fdcaSMatthew Dillon static void vm_swapcache_cleaning(vm_object_t marker); 86096e95c0SMatthew Dillon struct thread *swapcached_thread; 87096e95c0SMatthew Dillon 88096e95c0SMatthew Dillon static struct kproc_desc swpc_kp = { 89096e95c0SMatthew Dillon "swapcached", 90096e95c0SMatthew Dillon vm_swapcached, 91096e95c0SMatthew Dillon &swapcached_thread 92096e95c0SMatthew Dillon }; 93096e95c0SMatthew Dillon SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp) 94096e95c0SMatthew Dillon 95096e95c0SMatthew Dillon SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL); 96096e95c0SMatthew Dillon 97c504e38eSMatthew Dillon int vm_swapcache_read_enable; 98*e527fb6bSMatthew Dillon int vm_swapcache_inactive_heuristic; 99096e95c0SMatthew Dillon static int vm_swapcache_sleep; 1001e5196f0SMatthew Dillon static int vm_swapcache_maxlaunder = 256; 101096e95c0SMatthew Dillon static int vm_swapcache_data_enable = 0; 102096e95c0SMatthew Dillon static int vm_swapcache_meta_enable = 0; 103e9b56058SMatthew Dillon static int vm_swapcache_maxswappct = 75; 104*e527fb6bSMatthew Dillon static int vm_swapcache_hysteresis; 105e9b56058SMatthew Dillon static int vm_swapcache_use_chflags = 1; /* require chflags cache */ 1063ffc7051SMatthew Dillon static int64_t vm_swapcache_minburst = 10000000LL; /* 10MB */ 1073ffc7051SMatthew Dillon static int64_t vm_swapcache_curburst = 4000000000LL; /* 4G after boot */ 1083ffc7051SMatthew Dillon static int64_t vm_swapcache_maxburst = 2000000000LL; /* 2G nominal max */ 1093ffc7051SMatthew Dillon static int64_t vm_swapcache_accrate = 100000LL; /* 100K/s */ 110096e95c0SMatthew Dillon static int64_t vm_swapcache_write_count; 1113ffc7051SMatthew Dillon static int64_t vm_swapcache_maxfilesize; 112096e95c0SMatthew Dillon 113096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder, 114096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, ""); 115c504e38eSMatthew Dillon 116096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable, 117096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_data_enable, 0, ""); 118096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable, 119096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_meta_enable, 0, ""); 120c504e38eSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable, 121c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_read_enable, 0, ""); 122e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct, 123e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxswappct, 0, ""); 124*e527fb6bSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis, 125*e527fb6bSMatthew Dillon CTLFLAG_RW, &vm_swapcache_hysteresis, 0, ""); 126e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags, 127e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_use_chflags, 0, ""); 128c504e38eSMatthew Dillon 1293ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst, 1303ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_minburst, 0, ""); 131c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst, 132c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_curburst, 0, ""); 133c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst, 134c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxburst, 0, ""); 1353ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize, 1363ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, ""); 137c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate, 138c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_accrate, 0, ""); 139096e95c0SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count, 140096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_write_count, 0, ""); 141096e95c0SMatthew Dillon 142e9b56058SMatthew Dillon #define SWAPMAX(adj) \ 143e9b56058SMatthew Dillon ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100) 144e9b56058SMatthew Dillon 145096e95c0SMatthew Dillon /* 146096e95c0SMatthew Dillon * vm_swapcached is the high level pageout daemon. 147096e95c0SMatthew Dillon */ 148096e95c0SMatthew Dillon static void 149096e95c0SMatthew Dillon vm_swapcached(void) 150096e95c0SMatthew Dillon { 15100a3fdcaSMatthew Dillon enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING; 1523ffc7051SMatthew Dillon enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING; 15300a3fdcaSMatthew Dillon struct vm_page page_marker; 15400a3fdcaSMatthew Dillon struct vm_object object_marker; 155096e95c0SMatthew Dillon 156096e95c0SMatthew Dillon /* 157096e95c0SMatthew Dillon * Thread setup 158096e95c0SMatthew Dillon */ 159096e95c0SMatthew Dillon curthread->td_flags |= TDF_SYSTHREAD; 16000a3fdcaSMatthew Dillon crit_enter(); 161096e95c0SMatthew Dillon 162096e95c0SMatthew Dillon /* 16300a3fdcaSMatthew Dillon * Initialize our marker for the inactive scan (SWAPC_WRITING) 164096e95c0SMatthew Dillon */ 16500a3fdcaSMatthew Dillon bzero(&page_marker, sizeof(page_marker)); 16600a3fdcaSMatthew Dillon page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 16700a3fdcaSMatthew Dillon page_marker.queue = PQ_INACTIVE; 16800a3fdcaSMatthew Dillon page_marker.wire_count = 1; 16900a3fdcaSMatthew Dillon TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq); 170*e527fb6bSMatthew Dillon vm_swapcache_hysteresis = vmstats.v_inactive_target / 2; 171*e527fb6bSMatthew Dillon vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 172096e95c0SMatthew Dillon 17300a3fdcaSMatthew Dillon /* 17400a3fdcaSMatthew Dillon * Initialize our marker for the vm_object scan (SWAPC_CLEANING) 17500a3fdcaSMatthew Dillon */ 17600a3fdcaSMatthew Dillon bzero(&object_marker, sizeof(object_marker)); 17700a3fdcaSMatthew Dillon object_marker.type = OBJT_MARKER; 17800a3fdcaSMatthew Dillon TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list); 179096e95c0SMatthew Dillon 180096e95c0SMatthew Dillon for (;;) { 181096e95c0SMatthew Dillon /* 1823da46bd7SMatthew Dillon * Check every 5 seconds when not enabled or if no swap 1833da46bd7SMatthew Dillon * is present. 184096e95c0SMatthew Dillon */ 1853da46bd7SMatthew Dillon if ((vm_swapcache_data_enable == 0 && 1863da46bd7SMatthew Dillon vm_swapcache_meta_enable == 0) || 1873da46bd7SMatthew Dillon vm_swap_max == 0) { 188096e95c0SMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5); 189096e95c0SMatthew Dillon continue; 190096e95c0SMatthew Dillon } 191c504e38eSMatthew Dillon 192c504e38eSMatthew Dillon /* 1933da46bd7SMatthew Dillon * Polling rate when enabled is approximately 10 hz. 194c504e38eSMatthew Dillon */ 195c504e38eSMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10); 19600a3fdcaSMatthew Dillon 19700a3fdcaSMatthew Dillon /* 19800a3fdcaSMatthew Dillon * State hysteresis. Generate write activity up to 75% of 19900a3fdcaSMatthew Dillon * swap, then clean out swap assignments down to 70%, then 20000a3fdcaSMatthew Dillon * repeat. 20100a3fdcaSMatthew Dillon */ 20200a3fdcaSMatthew Dillon if (state == SWAPC_WRITING) { 203e9b56058SMatthew Dillon if (vm_swap_cache_use > SWAPMAX(0)) 20400a3fdcaSMatthew Dillon state = SWAPC_CLEANING; 20500a3fdcaSMatthew Dillon } else { 206e9b56058SMatthew Dillon if (vm_swap_cache_use < SWAPMAX(-5)) 20700a3fdcaSMatthew Dillon state = SWAPC_WRITING; 20800a3fdcaSMatthew Dillon } 20900a3fdcaSMatthew Dillon 21000a3fdcaSMatthew Dillon /* 21100a3fdcaSMatthew Dillon * We are allowed to continue accumulating burst value 2123ffc7051SMatthew Dillon * in either state. Allow the user to set curburst > maxburst 2133ffc7051SMatthew Dillon * for the initial load-in. 21400a3fdcaSMatthew Dillon */ 2153ffc7051SMatthew Dillon if (vm_swapcache_curburst < vm_swapcache_maxburst) { 216c504e38eSMatthew Dillon vm_swapcache_curburst += vm_swapcache_accrate / 10; 217c504e38eSMatthew Dillon if (vm_swapcache_curburst > vm_swapcache_maxburst) 218c504e38eSMatthew Dillon vm_swapcache_curburst = vm_swapcache_maxburst; 2193ffc7051SMatthew Dillon } 220c504e38eSMatthew Dillon 221c504e38eSMatthew Dillon /* 22200a3fdcaSMatthew Dillon * We don't want to nickle-and-dime the scan as that will 22300a3fdcaSMatthew Dillon * create unnecessary fragmentation. The minimum burst 22400a3fdcaSMatthew Dillon * is one-seconds worth of accumulation. 225c504e38eSMatthew Dillon */ 22600a3fdcaSMatthew Dillon if (state == SWAPC_WRITING) { 2273ffc7051SMatthew Dillon if (vm_swapcache_curburst >= vm_swapcache_accrate) { 2283ffc7051SMatthew Dillon if (burst == SWAPB_BURSTING) { 22900a3fdcaSMatthew Dillon vm_swapcache_writing(&page_marker); 2303ffc7051SMatthew Dillon if (vm_swapcache_curburst <= 0) 2313ffc7051SMatthew Dillon burst = SWAPB_RECOVERING; 2323ffc7051SMatthew Dillon } else if (vm_swapcache_curburst > 2333ffc7051SMatthew Dillon vm_swapcache_minburst) { 2343ffc7051SMatthew Dillon vm_swapcache_writing(&page_marker); 2353ffc7051SMatthew Dillon burst = SWAPB_BURSTING; 2363ffc7051SMatthew Dillon } 2373ffc7051SMatthew Dillon } 23800a3fdcaSMatthew Dillon } else { 23900a3fdcaSMatthew Dillon vm_swapcache_cleaning(&object_marker); 24000a3fdcaSMatthew Dillon } 24100a3fdcaSMatthew Dillon } 24200a3fdcaSMatthew Dillon TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq); 24300a3fdcaSMatthew Dillon TAILQ_REMOVE(&vm_object_list, &object_marker, object_list); 24400a3fdcaSMatthew Dillon crit_exit(); 24500a3fdcaSMatthew Dillon } 246096e95c0SMatthew Dillon 24700a3fdcaSMatthew Dillon static void 24800a3fdcaSMatthew Dillon vm_swapcache_writing(vm_page_t marker) 24900a3fdcaSMatthew Dillon { 25000a3fdcaSMatthew Dillon vm_object_t object; 25100a3fdcaSMatthew Dillon struct vnode *vp; 25200a3fdcaSMatthew Dillon vm_page_t m; 25300a3fdcaSMatthew Dillon int count; 254096e95c0SMatthew Dillon 255096e95c0SMatthew Dillon /* 256*e527fb6bSMatthew Dillon * Try to avoid small incremental pageouts by waiting for enough 257*e527fb6bSMatthew Dillon * pages to buildup in the inactive queue to hopefully get a good 258*e527fb6bSMatthew Dillon * burst in. This heuristic is bumped by the VM system and reset 259*e527fb6bSMatthew Dillon * when our scan hits the end of the queue. 260*e527fb6bSMatthew Dillon */ 261*e527fb6bSMatthew Dillon if (vm_swapcache_inactive_heuristic < 0) 262*e527fb6bSMatthew Dillon return; 263*e527fb6bSMatthew Dillon 264*e527fb6bSMatthew Dillon /* 265096e95c0SMatthew Dillon * Scan the inactive queue from our marker to locate 266096e95c0SMatthew Dillon * suitable pages to push to the swap cache. 267096e95c0SMatthew Dillon * 268096e95c0SMatthew Dillon * We are looking for clean vnode-backed pages. 2695ac04117SMatthew Dillon * 2705ac04117SMatthew Dillon * NOTE: PG_SWAPPED pages in particular are not part of 2715ac04117SMatthew Dillon * our count because once the cache stabilizes we 2725ac04117SMatthew Dillon * can end up with a very high datarate of VM pages 2735ac04117SMatthew Dillon * cycling from it. 274096e95c0SMatthew Dillon */ 27500a3fdcaSMatthew Dillon m = marker; 27600a3fdcaSMatthew Dillon count = vm_swapcache_maxlaunder; 27700a3fdcaSMatthew Dillon 278096e95c0SMatthew Dillon while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) { 2795ac04117SMatthew Dillon if (m->flags & (PG_MARKER | PG_SWAPPED)) { 280096e95c0SMatthew Dillon ++count; 281096e95c0SMatthew Dillon continue; 282096e95c0SMatthew Dillon } 283c504e38eSMatthew Dillon if (vm_swapcache_curburst < 0) 284c504e38eSMatthew Dillon break; 2853ffc7051SMatthew Dillon if (vm_swapcache_test(m)) 286096e95c0SMatthew Dillon continue; 2873ffc7051SMatthew Dillon object = m->object; 288c504e38eSMatthew Dillon vp = object->handle; 289c504e38eSMatthew Dillon if (vp == NULL) 290c504e38eSMatthew Dillon continue; 291d3070b8dSMatthew Dillon 292c504e38eSMatthew Dillon switch(vp->v_type) { 293c504e38eSMatthew Dillon case VREG: 294e9b56058SMatthew Dillon /* 295e9b56058SMatthew Dillon * If data_enable is 0 do not try to swapcache data. 296e9b56058SMatthew Dillon * If use_chflags is set then only swapcache data for 297e9b56058SMatthew Dillon * VSWAPCACHE marked vnodes, otherwise any vnode. 298e9b56058SMatthew Dillon */ 299e9b56058SMatthew Dillon if (vm_swapcache_data_enable == 0 || 300e9b56058SMatthew Dillon ((vp->v_flag & VSWAPCACHE) == 0 && 301e9b56058SMatthew Dillon vm_swapcache_use_chflags)) { 302c504e38eSMatthew Dillon continue; 303e9b56058SMatthew Dillon } 304d3070b8dSMatthew Dillon if (vm_swapcache_maxfilesize && 305d3070b8dSMatthew Dillon object->size > 306d3070b8dSMatthew Dillon (vm_swapcache_maxfilesize >> PAGE_SHIFT)) { 307d3070b8dSMatthew Dillon continue; 308d3070b8dSMatthew Dillon } 309c504e38eSMatthew Dillon break; 310c504e38eSMatthew Dillon case VCHR: 311c504e38eSMatthew Dillon if (vm_swapcache_meta_enable == 0) 312c504e38eSMatthew Dillon continue; 313c504e38eSMatthew Dillon break; 314c504e38eSMatthew Dillon default: 315c504e38eSMatthew Dillon continue; 316c504e38eSMatthew Dillon } 317096e95c0SMatthew Dillon 318096e95c0SMatthew Dillon /* 319096e95c0SMatthew Dillon * Ok, move the marker and soft-busy the page. 320096e95c0SMatthew Dillon */ 32100a3fdcaSMatthew Dillon TAILQ_REMOVE(INACTIVE_LIST, marker, pageq); 32200a3fdcaSMatthew Dillon TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq); 323096e95c0SMatthew Dillon 324096e95c0SMatthew Dillon /* 3253ffc7051SMatthew Dillon * Assign swap and initiate I/O. 3263ffc7051SMatthew Dillon * 3273ffc7051SMatthew Dillon * (adjust for the --count which also occurs in the loop) 328096e95c0SMatthew Dillon */ 3293ffc7051SMatthew Dillon count -= vm_swapcached_flush(m) - 1; 330096e95c0SMatthew Dillon 331096e95c0SMatthew Dillon /* 332096e95c0SMatthew Dillon * Setup for next loop using marker. 333096e95c0SMatthew Dillon */ 33400a3fdcaSMatthew Dillon m = marker; 335096e95c0SMatthew Dillon } 3361e5196f0SMatthew Dillon 3371e5196f0SMatthew Dillon /* 3381e5196f0SMatthew Dillon * Cleanup marker position. If we hit the end of the 3391e5196f0SMatthew Dillon * list the marker is placed at the tail. Newly deactivated 3401e5196f0SMatthew Dillon * pages will be placed after it. 3411e5196f0SMatthew Dillon * 3421e5196f0SMatthew Dillon * Earlier inactive pages that were dirty and become clean 3431e5196f0SMatthew Dillon * are typically moved to the end of PQ_INACTIVE by virtue 3441e5196f0SMatthew Dillon * of vfs_vmio_release() when they become unwired from the 3451e5196f0SMatthew Dillon * buffer cache. 3461e5196f0SMatthew Dillon */ 34700a3fdcaSMatthew Dillon TAILQ_REMOVE(INACTIVE_LIST, marker, pageq); 348*e527fb6bSMatthew Dillon if (m) { 34900a3fdcaSMatthew Dillon TAILQ_INSERT_BEFORE(m, marker, pageq); 350*e527fb6bSMatthew Dillon } else { 35100a3fdcaSMatthew Dillon TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq); 352*e527fb6bSMatthew Dillon vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 353*e527fb6bSMatthew Dillon } 354096e95c0SMatthew Dillon } 355096e95c0SMatthew Dillon 356096e95c0SMatthew Dillon /* 357096e95c0SMatthew Dillon * Flush the specified page using the swap_pager. 3583ffc7051SMatthew Dillon * 3593ffc7051SMatthew Dillon * Try to collect surrounding pages, including pages which may 3603ffc7051SMatthew Dillon * have already been assigned swap. Try to cluster within a 3613ffc7051SMatthew Dillon * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block 3623ffc7051SMatthew Dillon * to match what swap_pager_putpages() can do. 3633ffc7051SMatthew Dillon * 3643ffc7051SMatthew Dillon * We also want to try to match against the buffer cache blocksize 3653ffc7051SMatthew Dillon * but we don't really know what it is here. Since the buffer cache 3663ffc7051SMatthew Dillon * wires and unwires pages in groups the fact that we skip wired pages 3673ffc7051SMatthew Dillon * should be sufficient. 3683ffc7051SMatthew Dillon * 3693ffc7051SMatthew Dillon * Returns a count of pages we might have flushed (minimum 1) 370096e95c0SMatthew Dillon */ 371096e95c0SMatthew Dillon static 3723ffc7051SMatthew Dillon int 373096e95c0SMatthew Dillon vm_swapcached_flush(vm_page_t m) 374096e95c0SMatthew Dillon { 375096e95c0SMatthew Dillon vm_object_t object; 3763ffc7051SMatthew Dillon vm_page_t marray[SWAP_META_PAGES]; 3773ffc7051SMatthew Dillon vm_pindex_t basei; 3783ffc7051SMatthew Dillon int rtvals[SWAP_META_PAGES]; 3793ffc7051SMatthew Dillon int x; 3803ffc7051SMatthew Dillon int i; 3813ffc7051SMatthew Dillon int j; 3823ffc7051SMatthew Dillon int count; 383096e95c0SMatthew Dillon 384096e95c0SMatthew Dillon vm_page_io_start(m); 385096e95c0SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 386096e95c0SMatthew Dillon object = m->object; 387096e95c0SMatthew Dillon 3883ffc7051SMatthew Dillon /* 3893ffc7051SMatthew Dillon * Try to cluster around (m), keeping in mind that the swap pager 3903ffc7051SMatthew Dillon * can only do SMAP_META_PAGES worth of continguous write. 3913ffc7051SMatthew Dillon */ 3923ffc7051SMatthew Dillon x = (int)m->pindex & SWAP_META_MASK; 3933ffc7051SMatthew Dillon marray[x] = m; 3943ffc7051SMatthew Dillon basei = m->pindex; 3953ffc7051SMatthew Dillon 3963ffc7051SMatthew Dillon for (i = x - 1; i >= 0; --i) { 3973ffc7051SMatthew Dillon m = vm_page_lookup(object, basei - x + i); 3983ffc7051SMatthew Dillon if (m == NULL) 3993ffc7051SMatthew Dillon break; 4003ffc7051SMatthew Dillon if (vm_swapcache_test(m)) 4013ffc7051SMatthew Dillon break; 4023ffc7051SMatthew Dillon vm_page_io_start(m); 4033ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 4043ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 4053ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 4063ffc7051SMatthew Dillon vm_page_deactivate(m); 4073ffc7051SMatthew Dillon } 4083ffc7051SMatthew Dillon marray[i] = m; 4093ffc7051SMatthew Dillon } 4103ffc7051SMatthew Dillon ++i; 4113ffc7051SMatthew Dillon 4123ffc7051SMatthew Dillon for (j = x + 1; j < SWAP_META_PAGES; ++j) { 4133ffc7051SMatthew Dillon m = vm_page_lookup(object, basei - x + j); 4143ffc7051SMatthew Dillon if (m == NULL) 4153ffc7051SMatthew Dillon break; 4163ffc7051SMatthew Dillon if (vm_swapcache_test(m)) 4173ffc7051SMatthew Dillon break; 4183ffc7051SMatthew Dillon vm_page_io_start(m); 4193ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 4203ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 4213ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 4223ffc7051SMatthew Dillon vm_page_deactivate(m); 4233ffc7051SMatthew Dillon } 4243ffc7051SMatthew Dillon marray[j] = m; 4253ffc7051SMatthew Dillon } 4263ffc7051SMatthew Dillon 4273ffc7051SMatthew Dillon count = j - i; 4283ffc7051SMatthew Dillon vm_object_pip_add(object, count); 4293ffc7051SMatthew Dillon swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i); 4303ffc7051SMatthew Dillon vm_swapcache_write_count += count * PAGE_SIZE; 4313ffc7051SMatthew Dillon vm_swapcache_curburst -= count * PAGE_SIZE; 4323ffc7051SMatthew Dillon 4333ffc7051SMatthew Dillon while (i < j) { 4343ffc7051SMatthew Dillon if (rtvals[i] != VM_PAGER_PEND) { 4353ffc7051SMatthew Dillon vm_page_io_finish(marray[i]); 436096e95c0SMatthew Dillon vm_object_pip_wakeup(object); 437096e95c0SMatthew Dillon } 4383ffc7051SMatthew Dillon ++i; 4393ffc7051SMatthew Dillon } 4403ffc7051SMatthew Dillon return(count); 441096e95c0SMatthew Dillon } 44200a3fdcaSMatthew Dillon 4433ffc7051SMatthew Dillon /* 4443ffc7051SMatthew Dillon * Test whether a VM page is suitable for writing to the swapcache. 4453ffc7051SMatthew Dillon * Does not test m->queue, PG_MARKER, or PG_SWAPPED. 4463ffc7051SMatthew Dillon * 4473ffc7051SMatthew Dillon * Returns 0 on success, 1 on failure 4483ffc7051SMatthew Dillon */ 4493ffc7051SMatthew Dillon static int 4503ffc7051SMatthew Dillon vm_swapcache_test(vm_page_t m) 4513ffc7051SMatthew Dillon { 4523ffc7051SMatthew Dillon vm_object_t object; 4533ffc7051SMatthew Dillon 454b8a41159SMatthew Dillon if (m->flags & (PG_BUSY | PG_UNMANAGED | PG_NOTMETA)) 4553ffc7051SMatthew Dillon return(1); 4563ffc7051SMatthew Dillon if (m->busy || m->hold_count || m->wire_count) 4573ffc7051SMatthew Dillon return(1); 4583ffc7051SMatthew Dillon if (m->valid != VM_PAGE_BITS_ALL) 4593ffc7051SMatthew Dillon return(1); 4603ffc7051SMatthew Dillon if (m->dirty & m->valid) 4613ffc7051SMatthew Dillon return(1); 4623ffc7051SMatthew Dillon if ((object = m->object) == NULL) 4633ffc7051SMatthew Dillon return(1); 4643ffc7051SMatthew Dillon if (object->type != OBJT_VNODE || 4653ffc7051SMatthew Dillon (object->flags & OBJ_DEAD)) { 4663ffc7051SMatthew Dillon return(1); 4673ffc7051SMatthew Dillon } 4683ffc7051SMatthew Dillon vm_page_test_dirty(m); 4693ffc7051SMatthew Dillon if (m->dirty & m->valid) 4703ffc7051SMatthew Dillon return(1); 4713ffc7051SMatthew Dillon return(0); 4723ffc7051SMatthew Dillon } 4733ffc7051SMatthew Dillon 4743ffc7051SMatthew Dillon /* 4753ffc7051SMatthew Dillon * Cleaning pass 4763ffc7051SMatthew Dillon */ 47700a3fdcaSMatthew Dillon static 47800a3fdcaSMatthew Dillon void 47900a3fdcaSMatthew Dillon vm_swapcache_cleaning(vm_object_t marker) 48000a3fdcaSMatthew Dillon { 48100a3fdcaSMatthew Dillon vm_object_t object; 48200a3fdcaSMatthew Dillon struct vnode *vp; 48300a3fdcaSMatthew Dillon int count; 48400a3fdcaSMatthew Dillon int n; 48500a3fdcaSMatthew Dillon 48600a3fdcaSMatthew Dillon object = marker; 48700a3fdcaSMatthew Dillon count = vm_swapcache_maxlaunder; 48800a3fdcaSMatthew Dillon 48900a3fdcaSMatthew Dillon /* 49000a3fdcaSMatthew Dillon * Look for vnode objects 49100a3fdcaSMatthew Dillon */ 49200a3fdcaSMatthew Dillon while ((object = TAILQ_NEXT(object, object_list)) != NULL && count--) { 49300a3fdcaSMatthew Dillon if (object->type != OBJT_VNODE) 49400a3fdcaSMatthew Dillon continue; 49500a3fdcaSMatthew Dillon if ((object->flags & OBJ_DEAD) || object->swblock_count == 0) 49600a3fdcaSMatthew Dillon continue; 49700a3fdcaSMatthew Dillon if ((vp = object->handle) == NULL) 49800a3fdcaSMatthew Dillon continue; 49900a3fdcaSMatthew Dillon if (vp->v_type != VREG && vp->v_type != VCHR) 50000a3fdcaSMatthew Dillon continue; 50100a3fdcaSMatthew Dillon 50200a3fdcaSMatthew Dillon /* 50300a3fdcaSMatthew Dillon * Adjust iterator. 50400a3fdcaSMatthew Dillon */ 50500a3fdcaSMatthew Dillon if (marker->backing_object != object) 50600a3fdcaSMatthew Dillon marker->size = 0; 50700a3fdcaSMatthew Dillon 50800a3fdcaSMatthew Dillon /* 50900a3fdcaSMatthew Dillon * Move the marker so we can work on the VM object 51000a3fdcaSMatthew Dillon */ 51100a3fdcaSMatthew Dillon TAILQ_REMOVE(&vm_object_list, marker, object_list); 51200a3fdcaSMatthew Dillon TAILQ_INSERT_AFTER(&vm_object_list, object, 51300a3fdcaSMatthew Dillon marker, object_list); 51400a3fdcaSMatthew Dillon 51500a3fdcaSMatthew Dillon /* 51600a3fdcaSMatthew Dillon * Look for swblocks starting at our iterator. 51700a3fdcaSMatthew Dillon * 51800a3fdcaSMatthew Dillon * The swap_pager_condfree() function attempts to free 51900a3fdcaSMatthew Dillon * swap space starting at the specified index. The index 52000a3fdcaSMatthew Dillon * will be updated on return. The function will return 52100a3fdcaSMatthew Dillon * a scan factor (NOT the number of blocks freed). 52200a3fdcaSMatthew Dillon * 52300a3fdcaSMatthew Dillon * If it must cut its scan of the object short due to an 52400a3fdcaSMatthew Dillon * excessive number of swblocks, or is able to free the 52500a3fdcaSMatthew Dillon * requested number of blocks, it will return n >= count 52600a3fdcaSMatthew Dillon * and we break and pick it back up on a future attempt. 52700a3fdcaSMatthew Dillon */ 52800a3fdcaSMatthew Dillon n = swap_pager_condfree(object, &marker->size, count); 52900a3fdcaSMatthew Dillon count -= n; 53000a3fdcaSMatthew Dillon if (count < 0) 53100a3fdcaSMatthew Dillon break; 53200a3fdcaSMatthew Dillon 53300a3fdcaSMatthew Dillon /* 53400a3fdcaSMatthew Dillon * Setup for loop. 53500a3fdcaSMatthew Dillon */ 53600a3fdcaSMatthew Dillon marker->size = 0; 53700a3fdcaSMatthew Dillon object = marker; 53800a3fdcaSMatthew Dillon } 53900a3fdcaSMatthew Dillon 54000a3fdcaSMatthew Dillon /* 54100a3fdcaSMatthew Dillon * Adjust marker so we continue the scan from where we left off. 54200a3fdcaSMatthew Dillon * When we reach the end we start back at the beginning. 54300a3fdcaSMatthew Dillon */ 54400a3fdcaSMatthew Dillon TAILQ_REMOVE(&vm_object_list, marker, object_list); 54500a3fdcaSMatthew Dillon if (object) 54600a3fdcaSMatthew Dillon TAILQ_INSERT_BEFORE(object, marker, object_list); 54700a3fdcaSMatthew Dillon else 54800a3fdcaSMatthew Dillon TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list); 54900a3fdcaSMatthew Dillon marker->backing_object = object; 55000a3fdcaSMatthew Dillon } 551