1096e95c0SMatthew Dillon /* 28e7c4729SMatthew Dillon * (MPSAFE) 38e7c4729SMatthew Dillon * 4096e95c0SMatthew Dillon * Copyright (c) 2010 The DragonFly Project. All rights reserved. 5096e95c0SMatthew Dillon * 6096e95c0SMatthew Dillon * This code is derived from software contributed to The DragonFly Project 7096e95c0SMatthew Dillon * by Matthew Dillon <dillon@backplane.com> 8096e95c0SMatthew Dillon * 9096e95c0SMatthew Dillon * Redistribution and use in source and binary forms, with or without 10096e95c0SMatthew Dillon * modification, are permitted provided that the following conditions 11096e95c0SMatthew Dillon * are met: 12096e95c0SMatthew Dillon * 13096e95c0SMatthew Dillon * 1. Redistributions of source code must retain the above copyright 14096e95c0SMatthew Dillon * notice, this list of conditions and the following disclaimer. 15096e95c0SMatthew Dillon * 2. Redistributions in binary form must reproduce the above copyright 16096e95c0SMatthew Dillon * notice, this list of conditions and the following disclaimer in 17096e95c0SMatthew Dillon * the documentation and/or other materials provided with the 18096e95c0SMatthew Dillon * distribution. 19096e95c0SMatthew Dillon * 3. Neither the name of The DragonFly Project nor the names of its 20096e95c0SMatthew Dillon * contributors may be used to endorse or promote products derived 21096e95c0SMatthew Dillon * from this software without specific, prior written permission. 22096e95c0SMatthew Dillon * 23096e95c0SMatthew Dillon * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 24096e95c0SMatthew Dillon * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 25096e95c0SMatthew Dillon * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 26096e95c0SMatthew Dillon * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 27096e95c0SMatthew Dillon * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 28096e95c0SMatthew Dillon * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 29096e95c0SMatthew Dillon * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 30096e95c0SMatthew Dillon * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 31096e95c0SMatthew Dillon * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 32096e95c0SMatthew Dillon * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 33096e95c0SMatthew Dillon * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34096e95c0SMatthew Dillon * SUCH DAMAGE. 35096e95c0SMatthew Dillon */ 36096e95c0SMatthew Dillon 37096e95c0SMatthew Dillon /* 38096e95c0SMatthew Dillon * Implement the swapcache daemon. When enabled swap is assumed to be 39096e95c0SMatthew Dillon * configured on a fast storage device such as a SSD. Swap is assigned 40096e95c0SMatthew Dillon * to clean vnode-backed pages in the inactive queue, clustered by object 41096e95c0SMatthew Dillon * if possible, and written out. The swap assignment sticks around even 42096e95c0SMatthew Dillon * after the underlying pages have been recycled. 43096e95c0SMatthew Dillon * 44096e95c0SMatthew Dillon * The daemon manages write bandwidth based on sysctl settings to control 45096e95c0SMatthew Dillon * wear on the SSD. 46096e95c0SMatthew Dillon * 47096e95c0SMatthew Dillon * The vnode strategy code will check for the swap assignments and divert 483ffc7051SMatthew Dillon * reads to the swap device when the data is present in the swapcache. 49096e95c0SMatthew Dillon * 50096e95c0SMatthew Dillon * This operates on both regular files and the block device vnodes used by 51096e95c0SMatthew Dillon * filesystems to manage meta-data. 52096e95c0SMatthew Dillon */ 53096e95c0SMatthew Dillon 54096e95c0SMatthew Dillon #include "opt_vm.h" 55096e95c0SMatthew Dillon #include <sys/param.h> 56096e95c0SMatthew Dillon #include <sys/systm.h> 57096e95c0SMatthew Dillon #include <sys/kernel.h> 58096e95c0SMatthew Dillon #include <sys/proc.h> 59096e95c0SMatthew Dillon #include <sys/kthread.h> 60096e95c0SMatthew Dillon #include <sys/resourcevar.h> 61096e95c0SMatthew Dillon #include <sys/signalvar.h> 62096e95c0SMatthew Dillon #include <sys/vnode.h> 63096e95c0SMatthew Dillon #include <sys/vmmeter.h> 64096e95c0SMatthew Dillon #include <sys/sysctl.h> 65497524bfSMatthew Dillon #include <sys/eventhandler.h> 66096e95c0SMatthew Dillon 67096e95c0SMatthew Dillon #include <vm/vm.h> 68096e95c0SMatthew Dillon #include <vm/vm_param.h> 69096e95c0SMatthew Dillon #include <sys/lock.h> 70096e95c0SMatthew Dillon #include <vm/vm_object.h> 71096e95c0SMatthew Dillon #include <vm/vm_page.h> 72096e95c0SMatthew Dillon #include <vm/vm_map.h> 73096e95c0SMatthew Dillon #include <vm/vm_pageout.h> 74096e95c0SMatthew Dillon #include <vm/vm_pager.h> 75096e95c0SMatthew Dillon #include <vm/swap_pager.h> 76096e95c0SMatthew Dillon #include <vm/vm_extern.h> 77096e95c0SMatthew Dillon 78096e95c0SMatthew Dillon #include <sys/thread2.h> 79b12defdcSMatthew Dillon #include <sys/spinlock2.h> 80096e95c0SMatthew Dillon #include <vm/vm_page2.h> 81096e95c0SMatthew Dillon 82096e95c0SMatthew Dillon /* the kernel process "vm_pageout"*/ 83aabd5ce8SMatthew Dillon static int vm_swapcached_flush (vm_page_t m, int isblkdev); 843ffc7051SMatthew Dillon static int vm_swapcache_test(vm_page_t m); 85*64949baaSMatthew Dillon static int vm_swapcache_writing_heuristic(void); 86*64949baaSMatthew Dillon static int vm_swapcache_writing(vm_page_t marker, int count, int scount); 8700a3fdcaSMatthew Dillon static void vm_swapcache_cleaning(vm_object_t marker); 88f5f6d247SMatthew Dillon static void vm_swapcache_movemarker(vm_object_t marker, vm_object_t object); 89096e95c0SMatthew Dillon struct thread *swapcached_thread; 90096e95c0SMatthew Dillon 91096e95c0SMatthew Dillon SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL); 92096e95c0SMatthew Dillon 93c504e38eSMatthew Dillon int vm_swapcache_read_enable; 94e527fb6bSMatthew Dillon int vm_swapcache_inactive_heuristic; 95096e95c0SMatthew Dillon static int vm_swapcache_sleep; 96*64949baaSMatthew Dillon static int vm_swapcache_maxscan = PQ_L2_SIZE * 8; 97*64949baaSMatthew Dillon static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4; 98096e95c0SMatthew Dillon static int vm_swapcache_data_enable = 0; 99096e95c0SMatthew Dillon static int vm_swapcache_meta_enable = 0; 100e9b56058SMatthew Dillon static int vm_swapcache_maxswappct = 75; 101e527fb6bSMatthew Dillon static int vm_swapcache_hysteresis; 102*64949baaSMatthew Dillon static int vm_swapcache_min_hysteresis; 103bfa86281SMatthew Dillon int vm_swapcache_use_chflags = 1; /* require chflags cache */ 1043ffc7051SMatthew Dillon static int64_t vm_swapcache_minburst = 10000000LL; /* 10MB */ 1053ffc7051SMatthew Dillon static int64_t vm_swapcache_curburst = 4000000000LL; /* 4G after boot */ 1063ffc7051SMatthew Dillon static int64_t vm_swapcache_maxburst = 2000000000LL; /* 2G nominal max */ 1073ffc7051SMatthew Dillon static int64_t vm_swapcache_accrate = 100000LL; /* 100K/s */ 108096e95c0SMatthew Dillon static int64_t vm_swapcache_write_count; 1093ffc7051SMatthew Dillon static int64_t vm_swapcache_maxfilesize; 110f5f6d247SMatthew Dillon static int64_t vm_swapcache_cleanperobj = 16*1024*1024; 111096e95c0SMatthew Dillon 112096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder, 113096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, ""); 1140bf81261SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan, 1150bf81261SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxscan, 0, ""); 116c504e38eSMatthew Dillon 117096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable, 118096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_data_enable, 0, ""); 119096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable, 120096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_meta_enable, 0, ""); 121c504e38eSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable, 122c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_read_enable, 0, ""); 123e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct, 124e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxswappct, 0, ""); 125e527fb6bSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis, 126*64949baaSMatthew Dillon CTLFLAG_RD, &vm_swapcache_hysteresis, 0, ""); 127*64949baaSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis, 128*64949baaSMatthew Dillon CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, ""); 129e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags, 130e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_use_chflags, 0, ""); 131c504e38eSMatthew Dillon 1323ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst, 1333ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_minburst, 0, ""); 134c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst, 135c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_curburst, 0, ""); 136c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst, 137c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxburst, 0, ""); 1383ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize, 1393ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, ""); 140c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate, 141c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_accrate, 0, ""); 142096e95c0SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count, 143096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_write_count, 0, ""); 144f5f6d247SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj, 145f5f6d247SMatthew Dillon CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, ""); 146096e95c0SMatthew Dillon 147e9b56058SMatthew Dillon #define SWAPMAX(adj) \ 148e9b56058SMatthew Dillon ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100) 149e9b56058SMatthew Dillon 150096e95c0SMatthew Dillon /* 151497524bfSMatthew Dillon * When shutting down the machine we want to stop swapcache operation 152497524bfSMatthew Dillon * immediately so swap is not accessed after devices have been shuttered. 153497524bfSMatthew Dillon */ 154497524bfSMatthew Dillon static void 155497524bfSMatthew Dillon shutdown_swapcache(void *arg __unused) 156497524bfSMatthew Dillon { 157497524bfSMatthew Dillon vm_swapcache_read_enable = 0; 158497524bfSMatthew Dillon vm_swapcache_data_enable = 0; 159497524bfSMatthew Dillon vm_swapcache_meta_enable = 0; 160497524bfSMatthew Dillon wakeup(&vm_swapcache_sleep); /* shortcut 5-second wait */ 161497524bfSMatthew Dillon } 162497524bfSMatthew Dillon 163497524bfSMatthew Dillon /* 164096e95c0SMatthew Dillon * vm_swapcached is the high level pageout daemon. 1658e7c4729SMatthew Dillon * 1668e7c4729SMatthew Dillon * No requirements. 167096e95c0SMatthew Dillon */ 168096e95c0SMatthew Dillon static void 169cd8ab232SMatthew Dillon vm_swapcached_thread(void) 170096e95c0SMatthew Dillon { 17100a3fdcaSMatthew Dillon enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING; 1723ffc7051SMatthew Dillon enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING; 17351c99c61SMatthew Dillon static struct vm_page page_marker[PQ_L2_SIZE]; 174027193ebSMatthew Dillon static struct vm_object object_marker; 175027193ebSMatthew Dillon int q; 176096e95c0SMatthew Dillon 177096e95c0SMatthew Dillon /* 178096e95c0SMatthew Dillon * Thread setup 179096e95c0SMatthew Dillon */ 180096e95c0SMatthew Dillon curthread->td_flags |= TDF_SYSTHREAD; 181497524bfSMatthew Dillon EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 182497524bfSMatthew Dillon swapcached_thread, SHUTDOWN_PRI_FIRST); 183497524bfSMatthew Dillon EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache, 184497524bfSMatthew Dillon NULL, SHUTDOWN_PRI_SECOND); 185096e95c0SMatthew Dillon 186096e95c0SMatthew Dillon /* 18700a3fdcaSMatthew Dillon * Initialize our marker for the inactive scan (SWAPC_WRITING) 188096e95c0SMatthew Dillon */ 18900a3fdcaSMatthew Dillon bzero(&page_marker, sizeof(page_marker)); 19051c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 191027193ebSMatthew Dillon page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 192027193ebSMatthew Dillon page_marker[q].queue = PQ_INACTIVE + q; 193027193ebSMatthew Dillon page_marker[q].pc = q; 194027193ebSMatthew Dillon page_marker[q].wire_count = 1; 195027193ebSMatthew Dillon vm_page_queues_spin_lock(PQ_INACTIVE + q); 196027193ebSMatthew Dillon TAILQ_INSERT_HEAD( 197027193ebSMatthew Dillon &vm_page_queues[PQ_INACTIVE + q].pl, 198027193ebSMatthew Dillon &page_marker[q], pageq); 199027193ebSMatthew Dillon vm_page_queues_spin_unlock(PQ_INACTIVE + q); 200027193ebSMatthew Dillon } 201b12defdcSMatthew Dillon 202*64949baaSMatthew Dillon vm_swapcache_min_hysteresis = 1024; 203*64949baaSMatthew Dillon vm_swapcache_hysteresis = vm_swapcache_min_hysteresis; 204e527fb6bSMatthew Dillon vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 205096e95c0SMatthew Dillon 20600a3fdcaSMatthew Dillon /* 20700a3fdcaSMatthew Dillon * Initialize our marker for the vm_object scan (SWAPC_CLEANING) 20800a3fdcaSMatthew Dillon */ 20900a3fdcaSMatthew Dillon bzero(&object_marker, sizeof(object_marker)); 21000a3fdcaSMatthew Dillon object_marker.type = OBJT_MARKER; 2112de4f77eSMatthew Dillon lwkt_gettoken(&vmobj_token); 21200a3fdcaSMatthew Dillon TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list); 2132de4f77eSMatthew Dillon lwkt_reltoken(&vmobj_token); 214096e95c0SMatthew Dillon 215096e95c0SMatthew Dillon for (;;) { 216*64949baaSMatthew Dillon int reached_end; 217*64949baaSMatthew Dillon int scount; 218*64949baaSMatthew Dillon int count; 219*64949baaSMatthew Dillon 220096e95c0SMatthew Dillon /* 221497524bfSMatthew Dillon * Handle shutdown 222497524bfSMatthew Dillon */ 223497524bfSMatthew Dillon kproc_suspend_loop(); 224497524bfSMatthew Dillon 225497524bfSMatthew Dillon /* 2263da46bd7SMatthew Dillon * Check every 5 seconds when not enabled or if no swap 2273da46bd7SMatthew Dillon * is present. 228096e95c0SMatthew Dillon */ 2293da46bd7SMatthew Dillon if ((vm_swapcache_data_enable == 0 && 2303da46bd7SMatthew Dillon vm_swapcache_meta_enable == 0) || 2313da46bd7SMatthew Dillon vm_swap_max == 0) { 232096e95c0SMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5); 233096e95c0SMatthew Dillon continue; 234096e95c0SMatthew Dillon } 235c504e38eSMatthew Dillon 236c504e38eSMatthew Dillon /* 2373da46bd7SMatthew Dillon * Polling rate when enabled is approximately 10 hz. 238c504e38eSMatthew Dillon */ 239c504e38eSMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10); 24000a3fdcaSMatthew Dillon 24100a3fdcaSMatthew Dillon /* 24200a3fdcaSMatthew Dillon * State hysteresis. Generate write activity up to 75% of 24300a3fdcaSMatthew Dillon * swap, then clean out swap assignments down to 70%, then 24400a3fdcaSMatthew Dillon * repeat. 24500a3fdcaSMatthew Dillon */ 24600a3fdcaSMatthew Dillon if (state == SWAPC_WRITING) { 247e9b56058SMatthew Dillon if (vm_swap_cache_use > SWAPMAX(0)) 24800a3fdcaSMatthew Dillon state = SWAPC_CLEANING; 24900a3fdcaSMatthew Dillon } else { 25008fb7a9dSMatthew Dillon if (vm_swap_cache_use < SWAPMAX(-10)) 25100a3fdcaSMatthew Dillon state = SWAPC_WRITING; 25200a3fdcaSMatthew Dillon } 25300a3fdcaSMatthew Dillon 25400a3fdcaSMatthew Dillon /* 25500a3fdcaSMatthew Dillon * We are allowed to continue accumulating burst value 2563ffc7051SMatthew Dillon * in either state. Allow the user to set curburst > maxburst 2573ffc7051SMatthew Dillon * for the initial load-in. 25800a3fdcaSMatthew Dillon */ 2593ffc7051SMatthew Dillon if (vm_swapcache_curburst < vm_swapcache_maxburst) { 260c504e38eSMatthew Dillon vm_swapcache_curburst += vm_swapcache_accrate / 10; 261c504e38eSMatthew Dillon if (vm_swapcache_curburst > vm_swapcache_maxburst) 262c504e38eSMatthew Dillon vm_swapcache_curburst = vm_swapcache_maxburst; 2633ffc7051SMatthew Dillon } 264c504e38eSMatthew Dillon 265c504e38eSMatthew Dillon /* 26600a3fdcaSMatthew Dillon * We don't want to nickle-and-dime the scan as that will 26700a3fdcaSMatthew Dillon * create unnecessary fragmentation. The minimum burst 26800a3fdcaSMatthew Dillon * is one-seconds worth of accumulation. 269c504e38eSMatthew Dillon */ 270*64949baaSMatthew Dillon if (state != SWAPC_WRITING) { 271*64949baaSMatthew Dillon vm_swapcache_cleaning(&object_marker); 272*64949baaSMatthew Dillon continue; 273*64949baaSMatthew Dillon } 274*64949baaSMatthew Dillon if (vm_swapcache_curburst < vm_swapcache_accrate) 275*64949baaSMatthew Dillon continue; 276*64949baaSMatthew Dillon 277*64949baaSMatthew Dillon reached_end = 0; 278*64949baaSMatthew Dillon count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2; 279*64949baaSMatthew Dillon scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2; 280*64949baaSMatthew Dillon 2813ffc7051SMatthew Dillon if (burst == SWAPB_BURSTING) { 282*64949baaSMatthew Dillon if (vm_swapcache_writing_heuristic()) { 28351c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 284*64949baaSMatthew Dillon reached_end += 285027193ebSMatthew Dillon vm_swapcache_writing( 286*64949baaSMatthew Dillon &page_marker[q], 287*64949baaSMatthew Dillon count, 288*64949baaSMatthew Dillon scount); 289*64949baaSMatthew Dillon } 290027193ebSMatthew Dillon } 2913ffc7051SMatthew Dillon if (vm_swapcache_curburst <= 0) 2923ffc7051SMatthew Dillon burst = SWAPB_RECOVERING; 293*64949baaSMatthew Dillon } else if (vm_swapcache_curburst > vm_swapcache_minburst) { 294*64949baaSMatthew Dillon if (vm_swapcache_writing_heuristic()) { 29551c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 296*64949baaSMatthew Dillon reached_end += 297027193ebSMatthew Dillon vm_swapcache_writing( 298*64949baaSMatthew Dillon &page_marker[q], 299*64949baaSMatthew Dillon count, 300*64949baaSMatthew Dillon scount); 301*64949baaSMatthew Dillon } 302027193ebSMatthew Dillon } 3033ffc7051SMatthew Dillon burst = SWAPB_BURSTING; 3043ffc7051SMatthew Dillon } 305*64949baaSMatthew Dillon if (reached_end == PQ_L2_SIZE) { 306*64949baaSMatthew Dillon vm_swapcache_inactive_heuristic = 307*64949baaSMatthew Dillon -vm_swapcache_hysteresis; 30800a3fdcaSMatthew Dillon } 30900a3fdcaSMatthew Dillon } 310eccc8ca1SMatthew Dillon 311eccc8ca1SMatthew Dillon /* 312eccc8ca1SMatthew Dillon * Cleanup (NOT REACHED) 313eccc8ca1SMatthew Dillon */ 31451c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 315027193ebSMatthew Dillon vm_page_queues_spin_lock(PQ_INACTIVE + q); 316027193ebSMatthew Dillon TAILQ_REMOVE( 317027193ebSMatthew Dillon &vm_page_queues[PQ_INACTIVE + q].pl, 318027193ebSMatthew Dillon &page_marker[q], pageq); 319027193ebSMatthew Dillon vm_page_queues_spin_unlock(PQ_INACTIVE + q); 320027193ebSMatthew Dillon } 321eccc8ca1SMatthew Dillon 3222de4f77eSMatthew Dillon lwkt_gettoken(&vmobj_token); 32300a3fdcaSMatthew Dillon TAILQ_REMOVE(&vm_object_list, &object_marker, object_list); 3242de4f77eSMatthew Dillon lwkt_reltoken(&vmobj_token); 32500a3fdcaSMatthew Dillon } 326096e95c0SMatthew Dillon 327cd8ab232SMatthew Dillon static struct kproc_desc swpc_kp = { 328cd8ab232SMatthew Dillon "swapcached", 329cd8ab232SMatthew Dillon vm_swapcached_thread, 330cd8ab232SMatthew Dillon &swapcached_thread 331cd8ab232SMatthew Dillon }; 332cd8ab232SMatthew Dillon SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp) 333cd8ab232SMatthew Dillon 334096e95c0SMatthew Dillon /* 335fdc53cc7SMatthew Dillon * Deal with an overflow of the heuristic counter or if the user 336fdc53cc7SMatthew Dillon * manually changes the hysteresis. 337fdc53cc7SMatthew Dillon * 338e527fb6bSMatthew Dillon * Try to avoid small incremental pageouts by waiting for enough 339e527fb6bSMatthew Dillon * pages to buildup in the inactive queue to hopefully get a good 340e527fb6bSMatthew Dillon * burst in. This heuristic is bumped by the VM system and reset 341e527fb6bSMatthew Dillon * when our scan hits the end of the queue. 342*64949baaSMatthew Dillon * 343*64949baaSMatthew Dillon * Return TRUE if we need to take a writing pass. 344e527fb6bSMatthew Dillon */ 345*64949baaSMatthew Dillon static int 346*64949baaSMatthew Dillon vm_swapcache_writing_heuristic(void) 347*64949baaSMatthew Dillon { 348*64949baaSMatthew Dillon int hyst; 349*64949baaSMatthew Dillon 350*64949baaSMatthew Dillon hyst = vmstats.v_inactive_count / 4; 351*64949baaSMatthew Dillon if (hyst < vm_swapcache_min_hysteresis) 352*64949baaSMatthew Dillon hyst = vm_swapcache_min_hysteresis; 353*64949baaSMatthew Dillon cpu_ccfence(); 354*64949baaSMatthew Dillon vm_swapcache_hysteresis = hyst; 355*64949baaSMatthew Dillon 356*64949baaSMatthew Dillon if (vm_swapcache_inactive_heuristic < -hyst) 357*64949baaSMatthew Dillon vm_swapcache_inactive_heuristic = -hyst; 358*64949baaSMatthew Dillon 359*64949baaSMatthew Dillon return (vm_swapcache_inactive_heuristic >= 0); 360*64949baaSMatthew Dillon } 361*64949baaSMatthew Dillon 362*64949baaSMatthew Dillon /* 363*64949baaSMatthew Dillon * Take a writing pass on one of the inactive queues, return non-zero if 364*64949baaSMatthew Dillon * we hit the end of the queue. 365*64949baaSMatthew Dillon */ 366*64949baaSMatthew Dillon static int 367*64949baaSMatthew Dillon vm_swapcache_writing(vm_page_t marker, int count, int scount) 368*64949baaSMatthew Dillon { 369*64949baaSMatthew Dillon vm_object_t object; 370*64949baaSMatthew Dillon struct vnode *vp; 371*64949baaSMatthew Dillon vm_page_t m; 372*64949baaSMatthew Dillon int isblkdev; 373e527fb6bSMatthew Dillon 374e527fb6bSMatthew Dillon /* 375096e95c0SMatthew Dillon * Scan the inactive queue from our marker to locate 376096e95c0SMatthew Dillon * suitable pages to push to the swap cache. 377096e95c0SMatthew Dillon * 378096e95c0SMatthew Dillon * We are looking for clean vnode-backed pages. 379096e95c0SMatthew Dillon */ 380027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 3810bf81261SMatthew Dillon while ((m = TAILQ_NEXT(marker, pageq)) != NULL && 3820bf81261SMatthew Dillon count > 0 && scount-- > 0) { 383027193ebSMatthew Dillon KKASSERT(m->queue == marker->queue); 384b12defdcSMatthew Dillon 385b12defdcSMatthew Dillon if (vm_swapcache_curburst < 0) 386b12defdcSMatthew Dillon break; 387027193ebSMatthew Dillon TAILQ_REMOVE( 388027193ebSMatthew Dillon &vm_page_queues[marker->queue].pl, marker, pageq); 389027193ebSMatthew Dillon TAILQ_INSERT_AFTER( 390027193ebSMatthew Dillon &vm_page_queues[marker->queue].pl, m, marker, pageq); 391f5f6d247SMatthew Dillon 392f5f6d247SMatthew Dillon /* 393f5f6d247SMatthew Dillon * Ignore markers and ignore pages that already have a swap 394f5f6d247SMatthew Dillon * assignment. 395f5f6d247SMatthew Dillon */ 3960bf81261SMatthew Dillon if (m->flags & (PG_MARKER | PG_SWAPPED)) 397096e95c0SMatthew Dillon continue; 398b12defdcSMatthew Dillon if (vm_page_busy_try(m, TRUE)) 399096e95c0SMatthew Dillon continue; 400027193ebSMatthew Dillon vm_page_queues_spin_unlock(marker->queue); 401b12defdcSMatthew Dillon 402b12defdcSMatthew Dillon if ((object = m->object) == NULL) { 403b12defdcSMatthew Dillon vm_page_wakeup(m); 404027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 405b12defdcSMatthew Dillon continue; 406b12defdcSMatthew Dillon } 407b12defdcSMatthew Dillon vm_object_hold(object); 408b12defdcSMatthew Dillon if (m->object != object) { 409b12defdcSMatthew Dillon vm_object_drop(object); 410b12defdcSMatthew Dillon vm_page_wakeup(m); 411027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 412b12defdcSMatthew Dillon continue; 413b12defdcSMatthew Dillon } 414b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 415b12defdcSMatthew Dillon vm_object_drop(object); 416b12defdcSMatthew Dillon vm_page_wakeup(m); 417027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 418b12defdcSMatthew Dillon continue; 419b12defdcSMatthew Dillon } 420b12defdcSMatthew Dillon 421c504e38eSMatthew Dillon vp = object->handle; 422b12defdcSMatthew Dillon if (vp == NULL) { 423b12defdcSMatthew Dillon vm_object_drop(object); 424b12defdcSMatthew Dillon vm_page_wakeup(m); 425027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 426c504e38eSMatthew Dillon continue; 427b12defdcSMatthew Dillon } 428d3070b8dSMatthew Dillon 429c504e38eSMatthew Dillon switch(vp->v_type) { 430c504e38eSMatthew Dillon case VREG: 431e9b56058SMatthew Dillon /* 432bfa86281SMatthew Dillon * PG_NOTMETA generically means 'don't swapcache this', 433bfa86281SMatthew Dillon * and HAMMER will set this for regular data buffers 434bfa86281SMatthew Dillon * (and leave it unset for meta-data buffers) as 435bfa86281SMatthew Dillon * appropriate when double buffering is enabled. 436bfa86281SMatthew Dillon */ 437b12defdcSMatthew Dillon if (m->flags & PG_NOTMETA) { 438b12defdcSMatthew Dillon vm_object_drop(object); 439b12defdcSMatthew Dillon vm_page_wakeup(m); 440027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 441bfa86281SMatthew Dillon continue; 442b12defdcSMatthew Dillon } 443bfa86281SMatthew Dillon 444bfa86281SMatthew Dillon /* 445e9b56058SMatthew Dillon * If data_enable is 0 do not try to swapcache data. 446e9b56058SMatthew Dillon * If use_chflags is set then only swapcache data for 447e9b56058SMatthew Dillon * VSWAPCACHE marked vnodes, otherwise any vnode. 448e9b56058SMatthew Dillon */ 449e9b56058SMatthew Dillon if (vm_swapcache_data_enable == 0 || 450e9b56058SMatthew Dillon ((vp->v_flag & VSWAPCACHE) == 0 && 451e9b56058SMatthew Dillon vm_swapcache_use_chflags)) { 452b12defdcSMatthew Dillon vm_object_drop(object); 453b12defdcSMatthew Dillon vm_page_wakeup(m); 454027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 455c504e38eSMatthew Dillon continue; 456e9b56058SMatthew Dillon } 457d3070b8dSMatthew Dillon if (vm_swapcache_maxfilesize && 458d3070b8dSMatthew Dillon object->size > 459d3070b8dSMatthew Dillon (vm_swapcache_maxfilesize >> PAGE_SHIFT)) { 460b12defdcSMatthew Dillon vm_object_drop(object); 461b12defdcSMatthew Dillon vm_page_wakeup(m); 462027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 463d3070b8dSMatthew Dillon continue; 464d3070b8dSMatthew Dillon } 465aabd5ce8SMatthew Dillon isblkdev = 0; 466c504e38eSMatthew Dillon break; 467c504e38eSMatthew Dillon case VCHR: 468aabd5ce8SMatthew Dillon /* 469bfa86281SMatthew Dillon * PG_NOTMETA generically means 'don't swapcache this', 470bfa86281SMatthew Dillon * and HAMMER will set this for regular data buffers 471bfa86281SMatthew Dillon * (and leave it unset for meta-data buffers) as 472bfa86281SMatthew Dillon * appropriate when double buffering is enabled. 473aabd5ce8SMatthew Dillon */ 474b12defdcSMatthew Dillon if (m->flags & PG_NOTMETA) { 475b12defdcSMatthew Dillon vm_object_drop(object); 476b12defdcSMatthew Dillon vm_page_wakeup(m); 477027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 478aabd5ce8SMatthew Dillon continue; 479b12defdcSMatthew Dillon } 480b12defdcSMatthew Dillon if (vm_swapcache_meta_enable == 0) { 481b12defdcSMatthew Dillon vm_object_drop(object); 482b12defdcSMatthew Dillon vm_page_wakeup(m); 483027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 484c504e38eSMatthew Dillon continue; 485b12defdcSMatthew Dillon } 486aabd5ce8SMatthew Dillon isblkdev = 1; 487c504e38eSMatthew Dillon break; 488c504e38eSMatthew Dillon default: 489b12defdcSMatthew Dillon vm_object_drop(object); 490b12defdcSMatthew Dillon vm_page_wakeup(m); 491027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 492c504e38eSMatthew Dillon continue; 493c504e38eSMatthew Dillon } 494096e95c0SMatthew Dillon 495096e95c0SMatthew Dillon 496096e95c0SMatthew Dillon /* 4973ffc7051SMatthew Dillon * Assign swap and initiate I/O. 4983ffc7051SMatthew Dillon * 4993ffc7051SMatthew Dillon * (adjust for the --count which also occurs in the loop) 500096e95c0SMatthew Dillon */ 5010bf81261SMatthew Dillon count -= vm_swapcached_flush(m, isblkdev); 502096e95c0SMatthew Dillon 503096e95c0SMatthew Dillon /* 504096e95c0SMatthew Dillon * Setup for next loop using marker. 505096e95c0SMatthew Dillon */ 506b12defdcSMatthew Dillon vm_object_drop(object); 507027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 508096e95c0SMatthew Dillon } 5091e5196f0SMatthew Dillon 5101e5196f0SMatthew Dillon /* 511b12defdcSMatthew Dillon * The marker could wind up at the end, which is ok. If we hit the 512b12defdcSMatthew Dillon * end of the list adjust the heuristic. 5131e5196f0SMatthew Dillon * 5141e5196f0SMatthew Dillon * Earlier inactive pages that were dirty and become clean 5151e5196f0SMatthew Dillon * are typically moved to the end of PQ_INACTIVE by virtue 5161e5196f0SMatthew Dillon * of vfs_vmio_release() when they become unwired from the 5171e5196f0SMatthew Dillon * buffer cache. 5181e5196f0SMatthew Dillon */ 519027193ebSMatthew Dillon vm_page_queues_spin_unlock(marker->queue); 520*64949baaSMatthew Dillon 521*64949baaSMatthew Dillon /* 522*64949baaSMatthew Dillon * m invalid but can be used to test for NULL 523*64949baaSMatthew Dillon */ 524*64949baaSMatthew Dillon return (m == NULL); 525096e95c0SMatthew Dillon } 526096e95c0SMatthew Dillon 527096e95c0SMatthew Dillon /* 528b12defdcSMatthew Dillon * Flush the specified page using the swap_pager. The page 529b12defdcSMatthew Dillon * must be busied by the caller and its disposition will become 530b12defdcSMatthew Dillon * the responsibility of this function. 5313ffc7051SMatthew Dillon * 5323ffc7051SMatthew Dillon * Try to collect surrounding pages, including pages which may 5333ffc7051SMatthew Dillon * have already been assigned swap. Try to cluster within a 5343ffc7051SMatthew Dillon * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block 5353ffc7051SMatthew Dillon * to match what swap_pager_putpages() can do. 5363ffc7051SMatthew Dillon * 5373ffc7051SMatthew Dillon * We also want to try to match against the buffer cache blocksize 5383ffc7051SMatthew Dillon * but we don't really know what it is here. Since the buffer cache 5393ffc7051SMatthew Dillon * wires and unwires pages in groups the fact that we skip wired pages 5403ffc7051SMatthew Dillon * should be sufficient. 5413ffc7051SMatthew Dillon * 5423ffc7051SMatthew Dillon * Returns a count of pages we might have flushed (minimum 1) 543096e95c0SMatthew Dillon */ 544096e95c0SMatthew Dillon static 5453ffc7051SMatthew Dillon int 546aabd5ce8SMatthew Dillon vm_swapcached_flush(vm_page_t m, int isblkdev) 547096e95c0SMatthew Dillon { 548096e95c0SMatthew Dillon vm_object_t object; 5493ffc7051SMatthew Dillon vm_page_t marray[SWAP_META_PAGES]; 5503ffc7051SMatthew Dillon vm_pindex_t basei; 5513ffc7051SMatthew Dillon int rtvals[SWAP_META_PAGES]; 5523ffc7051SMatthew Dillon int x; 5533ffc7051SMatthew Dillon int i; 5543ffc7051SMatthew Dillon int j; 5553ffc7051SMatthew Dillon int count; 556b12defdcSMatthew Dillon int error; 557096e95c0SMatthew Dillon 558096e95c0SMatthew Dillon vm_page_io_start(m); 559096e95c0SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 560096e95c0SMatthew Dillon object = m->object; 561b12defdcSMatthew Dillon vm_object_hold(object); 562096e95c0SMatthew Dillon 5633ffc7051SMatthew Dillon /* 5643ffc7051SMatthew Dillon * Try to cluster around (m), keeping in mind that the swap pager 5653ffc7051SMatthew Dillon * can only do SMAP_META_PAGES worth of continguous write. 5663ffc7051SMatthew Dillon */ 5673ffc7051SMatthew Dillon x = (int)m->pindex & SWAP_META_MASK; 5683ffc7051SMatthew Dillon marray[x] = m; 5693ffc7051SMatthew Dillon basei = m->pindex; 570b12defdcSMatthew Dillon vm_page_wakeup(m); 5713ffc7051SMatthew Dillon 5723ffc7051SMatthew Dillon for (i = x - 1; i >= 0; --i) { 573b12defdcSMatthew Dillon m = vm_page_lookup_busy_try(object, basei - x + i, 574b12defdcSMatthew Dillon TRUE, &error); 575b12defdcSMatthew Dillon if (error || m == NULL) 5763ffc7051SMatthew Dillon break; 577b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 578b12defdcSMatthew Dillon vm_page_wakeup(m); 5793ffc7051SMatthew Dillon break; 580b12defdcSMatthew Dillon } 581b12defdcSMatthew Dillon if (isblkdev && (m->flags & PG_NOTMETA)) { 582b12defdcSMatthew Dillon vm_page_wakeup(m); 583aabd5ce8SMatthew Dillon break; 584b12defdcSMatthew Dillon } 5853ffc7051SMatthew Dillon vm_page_io_start(m); 5863ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 5873ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 5883ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 5893ffc7051SMatthew Dillon vm_page_deactivate(m); 5903ffc7051SMatthew Dillon } 5913ffc7051SMatthew Dillon marray[i] = m; 592b12defdcSMatthew Dillon vm_page_wakeup(m); 5933ffc7051SMatthew Dillon } 5943ffc7051SMatthew Dillon ++i; 5953ffc7051SMatthew Dillon 5963ffc7051SMatthew Dillon for (j = x + 1; j < SWAP_META_PAGES; ++j) { 597b12defdcSMatthew Dillon m = vm_page_lookup_busy_try(object, basei - x + j, 598b12defdcSMatthew Dillon TRUE, &error); 599b12defdcSMatthew Dillon if (error || m == NULL) 6003ffc7051SMatthew Dillon break; 601b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 602b12defdcSMatthew Dillon vm_page_wakeup(m); 6033ffc7051SMatthew Dillon break; 604b12defdcSMatthew Dillon } 605b12defdcSMatthew Dillon if (isblkdev && (m->flags & PG_NOTMETA)) { 606b12defdcSMatthew Dillon vm_page_wakeup(m); 607aabd5ce8SMatthew Dillon break; 608b12defdcSMatthew Dillon } 6093ffc7051SMatthew Dillon vm_page_io_start(m); 6103ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 6113ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 6123ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 6133ffc7051SMatthew Dillon vm_page_deactivate(m); 6143ffc7051SMatthew Dillon } 6153ffc7051SMatthew Dillon marray[j] = m; 616b12defdcSMatthew Dillon vm_page_wakeup(m); 6173ffc7051SMatthew Dillon } 6183ffc7051SMatthew Dillon 6193ffc7051SMatthew Dillon count = j - i; 6203ffc7051SMatthew Dillon vm_object_pip_add(object, count); 6213ffc7051SMatthew Dillon swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i); 6223ffc7051SMatthew Dillon vm_swapcache_write_count += count * PAGE_SIZE; 6233ffc7051SMatthew Dillon vm_swapcache_curburst -= count * PAGE_SIZE; 6243ffc7051SMatthew Dillon 6253ffc7051SMatthew Dillon while (i < j) { 6263ffc7051SMatthew Dillon if (rtvals[i] != VM_PAGER_PEND) { 627b12defdcSMatthew Dillon vm_page_busy_wait(marray[i], FALSE, "swppgfd"); 6283ffc7051SMatthew Dillon vm_page_io_finish(marray[i]); 629b12defdcSMatthew Dillon vm_page_wakeup(marray[i]); 630096e95c0SMatthew Dillon vm_object_pip_wakeup(object); 631096e95c0SMatthew Dillon } 6323ffc7051SMatthew Dillon ++i; 6333ffc7051SMatthew Dillon } 634b12defdcSMatthew Dillon vm_object_drop(object); 6353ffc7051SMatthew Dillon return(count); 636096e95c0SMatthew Dillon } 63700a3fdcaSMatthew Dillon 6383ffc7051SMatthew Dillon /* 6393ffc7051SMatthew Dillon * Test whether a VM page is suitable for writing to the swapcache. 6403ffc7051SMatthew Dillon * Does not test m->queue, PG_MARKER, or PG_SWAPPED. 6413ffc7051SMatthew Dillon * 6423ffc7051SMatthew Dillon * Returns 0 on success, 1 on failure 6433ffc7051SMatthew Dillon */ 6443ffc7051SMatthew Dillon static int 6453ffc7051SMatthew Dillon vm_swapcache_test(vm_page_t m) 6463ffc7051SMatthew Dillon { 6473ffc7051SMatthew Dillon vm_object_t object; 6483ffc7051SMatthew Dillon 649b12defdcSMatthew Dillon if (m->flags & PG_UNMANAGED) 6503ffc7051SMatthew Dillon return(1); 651b12defdcSMatthew Dillon if (m->hold_count || m->wire_count) 6523ffc7051SMatthew Dillon return(1); 6533ffc7051SMatthew Dillon if (m->valid != VM_PAGE_BITS_ALL) 6543ffc7051SMatthew Dillon return(1); 6553ffc7051SMatthew Dillon if (m->dirty & m->valid) 6563ffc7051SMatthew Dillon return(1); 6573ffc7051SMatthew Dillon if ((object = m->object) == NULL) 6583ffc7051SMatthew Dillon return(1); 6593ffc7051SMatthew Dillon if (object->type != OBJT_VNODE || 6603ffc7051SMatthew Dillon (object->flags & OBJ_DEAD)) { 6613ffc7051SMatthew Dillon return(1); 6623ffc7051SMatthew Dillon } 6633ffc7051SMatthew Dillon vm_page_test_dirty(m); 6643ffc7051SMatthew Dillon if (m->dirty & m->valid) 6653ffc7051SMatthew Dillon return(1); 6663ffc7051SMatthew Dillon return(0); 6673ffc7051SMatthew Dillon } 6683ffc7051SMatthew Dillon 6693ffc7051SMatthew Dillon /* 670f5f6d247SMatthew Dillon * Cleaning pass. 671f5f6d247SMatthew Dillon * 672f5f6d247SMatthew Dillon * We clean whole objects up to 16MB 6733ffc7051SMatthew Dillon */ 67400a3fdcaSMatthew Dillon static 67500a3fdcaSMatthew Dillon void 67600a3fdcaSMatthew Dillon vm_swapcache_cleaning(vm_object_t marker) 67700a3fdcaSMatthew Dillon { 67800a3fdcaSMatthew Dillon vm_object_t object; 67900a3fdcaSMatthew Dillon struct vnode *vp; 68000a3fdcaSMatthew Dillon int count; 6810bf81261SMatthew Dillon int scount; 68200a3fdcaSMatthew Dillon int n; 68300a3fdcaSMatthew Dillon 68400a3fdcaSMatthew Dillon count = vm_swapcache_maxlaunder; 6850bf81261SMatthew Dillon scount = vm_swapcache_maxscan; 68600a3fdcaSMatthew Dillon 68700a3fdcaSMatthew Dillon /* 68800a3fdcaSMatthew Dillon * Look for vnode objects 68900a3fdcaSMatthew Dillon */ 6902de4f77eSMatthew Dillon lwkt_gettoken(&vmobj_token); 6912de4f77eSMatthew Dillon 692f5f6d247SMatthew Dillon while ((object = TAILQ_NEXT(marker, object_list)) != NULL) { 6932f2d9e58SVenkatesh Srinivas /* 694f5f6d247SMatthew Dillon * We have to skip markers. We cannot hold/drop marker 695f5f6d247SMatthew Dillon * objects! 6962f2d9e58SVenkatesh Srinivas */ 697f5f6d247SMatthew Dillon if (object->type == OBJT_MARKER) { 698f5f6d247SMatthew Dillon vm_swapcache_movemarker(marker, object); 69900a3fdcaSMatthew Dillon continue; 7002f2d9e58SVenkatesh Srinivas } 70100a3fdcaSMatthew Dillon 70200a3fdcaSMatthew Dillon /* 703f5f6d247SMatthew Dillon * Safety, or in case there are millions of VM objects 704f5f6d247SMatthew Dillon * without swapcache backing. 70500a3fdcaSMatthew Dillon */ 7060bf81261SMatthew Dillon if (--scount <= 0) 707f5f6d247SMatthew Dillon break; 70800a3fdcaSMatthew Dillon 70900a3fdcaSMatthew Dillon /* 710f5f6d247SMatthew Dillon * We must hold the object before potentially yielding. 71100a3fdcaSMatthew Dillon */ 712f5f6d247SMatthew Dillon vm_object_hold(object); 713f5f6d247SMatthew Dillon lwkt_yield(); 714f5f6d247SMatthew Dillon 715f5f6d247SMatthew Dillon /* 716f5f6d247SMatthew Dillon * Only operate on live VNODE objects that are either 717f5f6d247SMatthew Dillon * VREG or VCHR (VCHR for meta-data). 718f5f6d247SMatthew Dillon */ 719f5f6d247SMatthew Dillon if ((object->type != OBJT_VNODE) || 720f5f6d247SMatthew Dillon ((object->flags & OBJ_DEAD) || 721f5f6d247SMatthew Dillon object->swblock_count == 0) || 722f5f6d247SMatthew Dillon ((vp = object->handle) == NULL) || 723f5f6d247SMatthew Dillon (vp->v_type != VREG && vp->v_type != VCHR)) { 724f5f6d247SMatthew Dillon vm_object_drop(object); 725f5f6d247SMatthew Dillon /* object may be invalid now */ 726f5f6d247SMatthew Dillon vm_swapcache_movemarker(marker, object); 727f5f6d247SMatthew Dillon continue; 728f5f6d247SMatthew Dillon } 729f5f6d247SMatthew Dillon 730f5f6d247SMatthew Dillon /* 731f5f6d247SMatthew Dillon * Reset the object pindex stored in the marker if the 732f5f6d247SMatthew Dillon * working object has changed. 733f5f6d247SMatthew Dillon */ 734f5f6d247SMatthew Dillon if (marker->backing_object != object) { 735f5f6d247SMatthew Dillon marker->size = 0; 736f5f6d247SMatthew Dillon marker->backing_object_offset = 0; 737f5f6d247SMatthew Dillon marker->backing_object = object; 738f5f6d247SMatthew Dillon } 73900a3fdcaSMatthew Dillon 74000a3fdcaSMatthew Dillon /* 74100a3fdcaSMatthew Dillon * Look for swblocks starting at our iterator. 74200a3fdcaSMatthew Dillon * 74300a3fdcaSMatthew Dillon * The swap_pager_condfree() function attempts to free 74400a3fdcaSMatthew Dillon * swap space starting at the specified index. The index 74500a3fdcaSMatthew Dillon * will be updated on return. The function will return 74600a3fdcaSMatthew Dillon * a scan factor (NOT the number of blocks freed). 74700a3fdcaSMatthew Dillon * 74800a3fdcaSMatthew Dillon * If it must cut its scan of the object short due to an 74900a3fdcaSMatthew Dillon * excessive number of swblocks, or is able to free the 75000a3fdcaSMatthew Dillon * requested number of blocks, it will return n >= count 75100a3fdcaSMatthew Dillon * and we break and pick it back up on a future attempt. 752f5f6d247SMatthew Dillon * 753f5f6d247SMatthew Dillon * Scan the object linearly and try to batch large sets of 754f5f6d247SMatthew Dillon * blocks that are likely to clean out entire swap radix 755f5f6d247SMatthew Dillon * tree leafs. 75600a3fdcaSMatthew Dillon */ 757739be60bSMatthew Dillon lwkt_token_swap(); 75808fb7a9dSMatthew Dillon lwkt_reltoken(&vmobj_token); 75927b6ee03SMatthew Dillon 760f5f6d247SMatthew Dillon n = swap_pager_condfree(object, &marker->size, 761f5f6d247SMatthew Dillon (count + SWAP_META_MASK) & ~SWAP_META_MASK); 7622f2d9e58SVenkatesh Srinivas 763f5f6d247SMatthew Dillon vm_object_drop(object); /* object may be invalid now */ 76427b6ee03SMatthew Dillon lwkt_gettoken(&vmobj_token); 7652f2d9e58SVenkatesh Srinivas 76600a3fdcaSMatthew Dillon /* 767f5f6d247SMatthew Dillon * If we have exhausted the object or deleted our per-pass 768f5f6d247SMatthew Dillon * page limit then move us to the next object. Note that 769f5f6d247SMatthew Dillon * the current object may no longer be on the vm_object_list. 77000a3fdcaSMatthew Dillon */ 771f5f6d247SMatthew Dillon if (n <= 0 || 772f5f6d247SMatthew Dillon marker->backing_object_offset > vm_swapcache_cleanperobj) { 773f5f6d247SMatthew Dillon vm_swapcache_movemarker(marker, object); 77400a3fdcaSMatthew Dillon } 77500a3fdcaSMatthew Dillon 77600a3fdcaSMatthew Dillon /* 777f5f6d247SMatthew Dillon * If we have exhausted our max-launder stop for now. 77800a3fdcaSMatthew Dillon */ 779f5f6d247SMatthew Dillon count -= n; 780f5f6d247SMatthew Dillon marker->backing_object_offset += n * PAGE_SIZE; 781f5f6d247SMatthew Dillon if (count < 0) 782f5f6d247SMatthew Dillon break; 783f5f6d247SMatthew Dillon } 7847a175765SMatthew Dillon 7857a175765SMatthew Dillon /* 7867a175765SMatthew Dillon * If we wound up at the end of the list this will move the 7877a175765SMatthew Dillon * marker back to the beginning. 7887a175765SMatthew Dillon */ 7897a175765SMatthew Dillon if (object == NULL) 7907a175765SMatthew Dillon vm_swapcache_movemarker(marker, NULL); 7917a175765SMatthew Dillon 7922de4f77eSMatthew Dillon lwkt_reltoken(&vmobj_token); 79300a3fdcaSMatthew Dillon } 794f5f6d247SMatthew Dillon 795f5f6d247SMatthew Dillon /* 796f5f6d247SMatthew Dillon * Move the marker past the current object. Object can be stale, but we 797f5f6d247SMatthew Dillon * still need it to determine if the marker has to be moved. If the object 798f5f6d247SMatthew Dillon * is still the 'current object' (object after the marker), we hop-scotch 799f5f6d247SMatthew Dillon * the marker past it. 800f5f6d247SMatthew Dillon */ 801f5f6d247SMatthew Dillon static void 802f5f6d247SMatthew Dillon vm_swapcache_movemarker(vm_object_t marker, vm_object_t object) 803f5f6d247SMatthew Dillon { 804f5f6d247SMatthew Dillon if (TAILQ_NEXT(marker, object_list) == object) { 805f5f6d247SMatthew Dillon TAILQ_REMOVE(&vm_object_list, marker, object_list); 806f5f6d247SMatthew Dillon if (object) { 807f5f6d247SMatthew Dillon TAILQ_INSERT_AFTER(&vm_object_list, object, 808f5f6d247SMatthew Dillon marker, object_list); 809f5f6d247SMatthew Dillon } else { 810f5f6d247SMatthew Dillon TAILQ_INSERT_HEAD(&vm_object_list, 811f5f6d247SMatthew Dillon marker, object_list); 812f5f6d247SMatthew Dillon } 813f5f6d247SMatthew Dillon } 814f5f6d247SMatthew Dillon } 815