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); 8564949baaSMatthew Dillon static int vm_swapcache_writing_heuristic(void); 8664949baaSMatthew Dillon static int vm_swapcache_writing(vm_page_t marker, int count, int scount); 87fde6be6aSMatthew Dillon static void vm_swapcache_cleaning(vm_object_t marker, 88fde6be6aSMatthew Dillon struct vm_object_hash **swindexp); 89fde6be6aSMatthew Dillon static void vm_swapcache_movemarker(vm_object_t marker, 90fde6be6aSMatthew Dillon struct vm_object_hash *swindex, vm_object_t object); 91096e95c0SMatthew Dillon struct thread *swapcached_thread; 92096e95c0SMatthew Dillon 93096e95c0SMatthew Dillon SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL); 94096e95c0SMatthew Dillon 95c504e38eSMatthew Dillon int vm_swapcache_read_enable; 96e527fb6bSMatthew Dillon int vm_swapcache_inactive_heuristic; 97096e95c0SMatthew Dillon static int vm_swapcache_sleep; 9864949baaSMatthew Dillon static int vm_swapcache_maxscan = PQ_L2_SIZE * 8; 9964949baaSMatthew Dillon static int vm_swapcache_maxlaunder = PQ_L2_SIZE * 4; 100096e95c0SMatthew Dillon static int vm_swapcache_data_enable = 0; 101096e95c0SMatthew Dillon static int vm_swapcache_meta_enable = 0; 102e9b56058SMatthew Dillon static int vm_swapcache_maxswappct = 75; 103e527fb6bSMatthew Dillon static int vm_swapcache_hysteresis; 10464949baaSMatthew Dillon static int vm_swapcache_min_hysteresis; 105*d4d9221dSMatthew Dillon int vm_swapcache_use_chflags = 0; /* 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; 112f5f6d247SMatthew Dillon static int64_t vm_swapcache_cleanperobj = 16*1024*1024; 113096e95c0SMatthew Dillon 114096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder, 115096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, ""); 1160bf81261SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxscan, 1170bf81261SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxscan, 0, ""); 118c504e38eSMatthew Dillon 119096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable, 120096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_data_enable, 0, ""); 121096e95c0SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable, 122096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_meta_enable, 0, ""); 123c504e38eSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable, 124c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_read_enable, 0, ""); 125e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct, 126e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxswappct, 0, ""); 127e527fb6bSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis, 12864949baaSMatthew Dillon CTLFLAG_RD, &vm_swapcache_hysteresis, 0, ""); 12964949baaSMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, min_hysteresis, 13064949baaSMatthew Dillon CTLFLAG_RW, &vm_swapcache_min_hysteresis, 0, ""); 131e9b56058SMatthew Dillon SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags, 132e9b56058SMatthew Dillon CTLFLAG_RW, &vm_swapcache_use_chflags, 0, ""); 133c504e38eSMatthew Dillon 1343ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst, 1353ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_minburst, 0, ""); 136c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst, 137c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_curburst, 0, ""); 138c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst, 139c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxburst, 0, ""); 1403ffc7051SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize, 1413ffc7051SMatthew Dillon CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, ""); 142c504e38eSMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate, 143c504e38eSMatthew Dillon CTLFLAG_RW, &vm_swapcache_accrate, 0, ""); 144096e95c0SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count, 145096e95c0SMatthew Dillon CTLFLAG_RW, &vm_swapcache_write_count, 0, ""); 146f5f6d247SMatthew Dillon SYSCTL_QUAD(_vm_swapcache, OID_AUTO, cleanperobj, 147f5f6d247SMatthew Dillon CTLFLAG_RW, &vm_swapcache_cleanperobj, 0, ""); 148096e95c0SMatthew Dillon 149e9b56058SMatthew Dillon #define SWAPMAX(adj) \ 150e9b56058SMatthew Dillon ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100) 151e9b56058SMatthew Dillon 152096e95c0SMatthew Dillon /* 153497524bfSMatthew Dillon * When shutting down the machine we want to stop swapcache operation 154497524bfSMatthew Dillon * immediately so swap is not accessed after devices have been shuttered. 155497524bfSMatthew Dillon */ 156497524bfSMatthew Dillon static void 157497524bfSMatthew Dillon shutdown_swapcache(void *arg __unused) 158497524bfSMatthew Dillon { 159497524bfSMatthew Dillon vm_swapcache_read_enable = 0; 160497524bfSMatthew Dillon vm_swapcache_data_enable = 0; 161497524bfSMatthew Dillon vm_swapcache_meta_enable = 0; 162497524bfSMatthew Dillon wakeup(&vm_swapcache_sleep); /* shortcut 5-second wait */ 163497524bfSMatthew Dillon } 164497524bfSMatthew Dillon 165497524bfSMatthew Dillon /* 166096e95c0SMatthew Dillon * vm_swapcached is the high level pageout daemon. 1678e7c4729SMatthew Dillon * 1688e7c4729SMatthew Dillon * No requirements. 169096e95c0SMatthew Dillon */ 170096e95c0SMatthew Dillon static void 171cd8ab232SMatthew Dillon vm_swapcached_thread(void) 172096e95c0SMatthew Dillon { 17300a3fdcaSMatthew Dillon enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING; 1743ffc7051SMatthew Dillon enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING; 17551c99c61SMatthew Dillon static struct vm_page page_marker[PQ_L2_SIZE]; 1767b00fbb4SMatthew Dillon static struct vm_object swmarker; 177fde6be6aSMatthew Dillon static struct vm_object_hash *swindex; 178027193ebSMatthew Dillon int q; 179096e95c0SMatthew Dillon 180096e95c0SMatthew Dillon /* 181096e95c0SMatthew Dillon * Thread setup 182096e95c0SMatthew Dillon */ 183096e95c0SMatthew Dillon curthread->td_flags |= TDF_SYSTHREAD; 184497524bfSMatthew Dillon EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc, 185497524bfSMatthew Dillon swapcached_thread, SHUTDOWN_PRI_FIRST); 186497524bfSMatthew Dillon EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache, 187497524bfSMatthew Dillon NULL, SHUTDOWN_PRI_SECOND); 188096e95c0SMatthew Dillon 189096e95c0SMatthew Dillon /* 19000a3fdcaSMatthew Dillon * Initialize our marker for the inactive scan (SWAPC_WRITING) 191096e95c0SMatthew Dillon */ 19200a3fdcaSMatthew Dillon bzero(&page_marker, sizeof(page_marker)); 19351c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 194027193ebSMatthew Dillon page_marker[q].flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 195027193ebSMatthew Dillon page_marker[q].queue = PQ_INACTIVE + q; 196027193ebSMatthew Dillon page_marker[q].pc = q; 197027193ebSMatthew Dillon page_marker[q].wire_count = 1; 198027193ebSMatthew Dillon vm_page_queues_spin_lock(PQ_INACTIVE + q); 199027193ebSMatthew Dillon TAILQ_INSERT_HEAD( 200027193ebSMatthew Dillon &vm_page_queues[PQ_INACTIVE + q].pl, 201027193ebSMatthew Dillon &page_marker[q], pageq); 202027193ebSMatthew Dillon vm_page_queues_spin_unlock(PQ_INACTIVE + q); 203027193ebSMatthew Dillon } 204b12defdcSMatthew Dillon 20564949baaSMatthew Dillon vm_swapcache_min_hysteresis = 1024; 20664949baaSMatthew Dillon vm_swapcache_hysteresis = vm_swapcache_min_hysteresis; 207e527fb6bSMatthew Dillon vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis; 208096e95c0SMatthew Dillon 20900a3fdcaSMatthew Dillon /* 21000a3fdcaSMatthew Dillon * Initialize our marker for the vm_object scan (SWAPC_CLEANING) 21100a3fdcaSMatthew Dillon */ 2127b00fbb4SMatthew Dillon bzero(&swmarker, sizeof(swmarker)); 2137b00fbb4SMatthew Dillon swmarker.type = OBJT_MARKER; 214fde6be6aSMatthew Dillon swindex = &vm_object_hash[0]; 215fde6be6aSMatthew Dillon lwkt_gettoken(&swindex->token); 216fde6be6aSMatthew Dillon TAILQ_INSERT_HEAD(&swindex->list, &swmarker, object_list); 217fde6be6aSMatthew Dillon lwkt_reltoken(&swindex->token); 218096e95c0SMatthew Dillon 219096e95c0SMatthew Dillon for (;;) { 22064949baaSMatthew Dillon int reached_end; 22164949baaSMatthew Dillon int scount; 22264949baaSMatthew Dillon int count; 22364949baaSMatthew Dillon 224096e95c0SMatthew Dillon /* 225497524bfSMatthew Dillon * Handle shutdown 226497524bfSMatthew Dillon */ 227497524bfSMatthew Dillon kproc_suspend_loop(); 228497524bfSMatthew Dillon 229497524bfSMatthew Dillon /* 2303da46bd7SMatthew Dillon * Check every 5 seconds when not enabled or if no swap 2313da46bd7SMatthew Dillon * is present. 232096e95c0SMatthew Dillon */ 2333da46bd7SMatthew Dillon if ((vm_swapcache_data_enable == 0 && 234c8ddd5d8SMatthew Dillon vm_swapcache_meta_enable == 0 && 235c8ddd5d8SMatthew Dillon vm_swap_cache_use <= SWAPMAX(0)) || 2363da46bd7SMatthew Dillon vm_swap_max == 0) { 237096e95c0SMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5); 238096e95c0SMatthew Dillon continue; 239096e95c0SMatthew Dillon } 240c504e38eSMatthew Dillon 241c504e38eSMatthew Dillon /* 2423da46bd7SMatthew Dillon * Polling rate when enabled is approximately 10 hz. 243c504e38eSMatthew Dillon */ 244c504e38eSMatthew Dillon tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10); 24500a3fdcaSMatthew Dillon 24600a3fdcaSMatthew Dillon /* 24700a3fdcaSMatthew Dillon * State hysteresis. Generate write activity up to 75% of 24800a3fdcaSMatthew Dillon * swap, then clean out swap assignments down to 70%, then 24900a3fdcaSMatthew Dillon * repeat. 25000a3fdcaSMatthew Dillon */ 25100a3fdcaSMatthew Dillon if (state == SWAPC_WRITING) { 252e9b56058SMatthew Dillon if (vm_swap_cache_use > SWAPMAX(0)) 25300a3fdcaSMatthew Dillon state = SWAPC_CLEANING; 25400a3fdcaSMatthew Dillon } else { 25508fb7a9dSMatthew Dillon if (vm_swap_cache_use < SWAPMAX(-10)) 25600a3fdcaSMatthew Dillon state = SWAPC_WRITING; 25700a3fdcaSMatthew Dillon } 25800a3fdcaSMatthew Dillon 25900a3fdcaSMatthew Dillon /* 26000a3fdcaSMatthew Dillon * We are allowed to continue accumulating burst value 2613ffc7051SMatthew Dillon * in either state. Allow the user to set curburst > maxburst 2623ffc7051SMatthew Dillon * for the initial load-in. 26300a3fdcaSMatthew Dillon */ 2643ffc7051SMatthew Dillon if (vm_swapcache_curburst < vm_swapcache_maxburst) { 265c504e38eSMatthew Dillon vm_swapcache_curburst += vm_swapcache_accrate / 10; 266c504e38eSMatthew Dillon if (vm_swapcache_curburst > vm_swapcache_maxburst) 267c504e38eSMatthew Dillon vm_swapcache_curburst = vm_swapcache_maxburst; 2683ffc7051SMatthew Dillon } 269c504e38eSMatthew Dillon 270c504e38eSMatthew Dillon /* 27100a3fdcaSMatthew Dillon * We don't want to nickle-and-dime the scan as that will 27200a3fdcaSMatthew Dillon * create unnecessary fragmentation. The minimum burst 27300a3fdcaSMatthew Dillon * is one-seconds worth of accumulation. 274c504e38eSMatthew Dillon */ 27564949baaSMatthew Dillon if (state != SWAPC_WRITING) { 2767b00fbb4SMatthew Dillon vm_swapcache_cleaning(&swmarker, &swindex); 27764949baaSMatthew Dillon continue; 27864949baaSMatthew Dillon } 27964949baaSMatthew Dillon if (vm_swapcache_curburst < vm_swapcache_accrate) 28064949baaSMatthew Dillon continue; 28164949baaSMatthew Dillon 28264949baaSMatthew Dillon reached_end = 0; 28364949baaSMatthew Dillon count = vm_swapcache_maxlaunder / PQ_L2_SIZE + 2; 28464949baaSMatthew Dillon scount = vm_swapcache_maxscan / PQ_L2_SIZE + 2; 28564949baaSMatthew Dillon 2863ffc7051SMatthew Dillon if (burst == SWAPB_BURSTING) { 28764949baaSMatthew Dillon if (vm_swapcache_writing_heuristic()) { 28851c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 28964949baaSMatthew Dillon reached_end += 290027193ebSMatthew Dillon vm_swapcache_writing( 29164949baaSMatthew Dillon &page_marker[q], 29264949baaSMatthew Dillon count, 29364949baaSMatthew Dillon scount); 29464949baaSMatthew Dillon } 295027193ebSMatthew Dillon } 2963ffc7051SMatthew Dillon if (vm_swapcache_curburst <= 0) 2973ffc7051SMatthew Dillon burst = SWAPB_RECOVERING; 29864949baaSMatthew Dillon } else if (vm_swapcache_curburst > vm_swapcache_minburst) { 29964949baaSMatthew Dillon if (vm_swapcache_writing_heuristic()) { 30051c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 30164949baaSMatthew Dillon reached_end += 302027193ebSMatthew Dillon vm_swapcache_writing( 30364949baaSMatthew Dillon &page_marker[q], 30464949baaSMatthew Dillon count, 30564949baaSMatthew Dillon scount); 30664949baaSMatthew Dillon } 307027193ebSMatthew Dillon } 3083ffc7051SMatthew Dillon burst = SWAPB_BURSTING; 3093ffc7051SMatthew Dillon } 31064949baaSMatthew Dillon if (reached_end == PQ_L2_SIZE) { 31164949baaSMatthew Dillon vm_swapcache_inactive_heuristic = 31264949baaSMatthew Dillon -vm_swapcache_hysteresis; 31300a3fdcaSMatthew Dillon } 31400a3fdcaSMatthew Dillon } 315eccc8ca1SMatthew Dillon 316eccc8ca1SMatthew Dillon /* 317eccc8ca1SMatthew Dillon * Cleanup (NOT REACHED) 318eccc8ca1SMatthew Dillon */ 31951c99c61SMatthew Dillon for (q = 0; q < PQ_L2_SIZE; ++q) { 320027193ebSMatthew Dillon vm_page_queues_spin_lock(PQ_INACTIVE + q); 321027193ebSMatthew Dillon TAILQ_REMOVE( 322027193ebSMatthew Dillon &vm_page_queues[PQ_INACTIVE + q].pl, 323027193ebSMatthew Dillon &page_marker[q], pageq); 324027193ebSMatthew Dillon vm_page_queues_spin_unlock(PQ_INACTIVE + q); 325027193ebSMatthew Dillon } 326eccc8ca1SMatthew Dillon 327fde6be6aSMatthew Dillon lwkt_gettoken(&swindex->token); 328fde6be6aSMatthew Dillon TAILQ_REMOVE(&swindex->list, &swmarker, object_list); 329fde6be6aSMatthew Dillon lwkt_reltoken(&swindex->token); 33000a3fdcaSMatthew Dillon } 331096e95c0SMatthew Dillon 332cd8ab232SMatthew Dillon static struct kproc_desc swpc_kp = { 333cd8ab232SMatthew Dillon "swapcached", 334cd8ab232SMatthew Dillon vm_swapcached_thread, 335cd8ab232SMatthew Dillon &swapcached_thread 336cd8ab232SMatthew Dillon }; 337f3f3eadbSSascha Wildner SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp); 338cd8ab232SMatthew Dillon 339096e95c0SMatthew Dillon /* 340fdc53cc7SMatthew Dillon * Deal with an overflow of the heuristic counter or if the user 341fdc53cc7SMatthew Dillon * manually changes the hysteresis. 342fdc53cc7SMatthew Dillon * 343e527fb6bSMatthew Dillon * Try to avoid small incremental pageouts by waiting for enough 344e527fb6bSMatthew Dillon * pages to buildup in the inactive queue to hopefully get a good 345e527fb6bSMatthew Dillon * burst in. This heuristic is bumped by the VM system and reset 346e527fb6bSMatthew Dillon * when our scan hits the end of the queue. 34764949baaSMatthew Dillon * 34864949baaSMatthew Dillon * Return TRUE if we need to take a writing pass. 349e527fb6bSMatthew Dillon */ 35064949baaSMatthew Dillon static int 35164949baaSMatthew Dillon vm_swapcache_writing_heuristic(void) 35264949baaSMatthew Dillon { 35364949baaSMatthew Dillon int hyst; 35464949baaSMatthew Dillon 35564949baaSMatthew Dillon hyst = vmstats.v_inactive_count / 4; 35664949baaSMatthew Dillon if (hyst < vm_swapcache_min_hysteresis) 35764949baaSMatthew Dillon hyst = vm_swapcache_min_hysteresis; 35864949baaSMatthew Dillon cpu_ccfence(); 35964949baaSMatthew Dillon vm_swapcache_hysteresis = hyst; 36064949baaSMatthew Dillon 36164949baaSMatthew Dillon if (vm_swapcache_inactive_heuristic < -hyst) 36264949baaSMatthew Dillon vm_swapcache_inactive_heuristic = -hyst; 36364949baaSMatthew Dillon 36464949baaSMatthew Dillon return (vm_swapcache_inactive_heuristic >= 0); 36564949baaSMatthew Dillon } 36664949baaSMatthew Dillon 36764949baaSMatthew Dillon /* 36864949baaSMatthew Dillon * Take a writing pass on one of the inactive queues, return non-zero if 36964949baaSMatthew Dillon * we hit the end of the queue. 37064949baaSMatthew Dillon */ 37164949baaSMatthew Dillon static int 37264949baaSMatthew Dillon vm_swapcache_writing(vm_page_t marker, int count, int scount) 37364949baaSMatthew Dillon { 37464949baaSMatthew Dillon vm_object_t object; 37564949baaSMatthew Dillon struct vnode *vp; 37664949baaSMatthew Dillon vm_page_t m; 37764949baaSMatthew Dillon int isblkdev; 378e527fb6bSMatthew Dillon 379e527fb6bSMatthew Dillon /* 380096e95c0SMatthew Dillon * Scan the inactive queue from our marker to locate 381096e95c0SMatthew Dillon * suitable pages to push to the swap cache. 382096e95c0SMatthew Dillon * 383096e95c0SMatthew Dillon * We are looking for clean vnode-backed pages. 384096e95c0SMatthew Dillon */ 385027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 3860bf81261SMatthew Dillon while ((m = TAILQ_NEXT(marker, pageq)) != NULL && 3870bf81261SMatthew Dillon count > 0 && scount-- > 0) { 388027193ebSMatthew Dillon KKASSERT(m->queue == marker->queue); 389b12defdcSMatthew Dillon 3901de864f0SMatthew Dillon /* 3911de864f0SMatthew Dillon * Stop using swap if paniced, dumping, or dumped. 3921de864f0SMatthew Dillon * Don't try to write if our curburst has been exhausted. 3931de864f0SMatthew Dillon */ 3941de864f0SMatthew Dillon if (panicstr || dumping) 3951de864f0SMatthew Dillon break; 396b12defdcSMatthew Dillon if (vm_swapcache_curburst < 0) 397b12defdcSMatthew Dillon break; 3981de864f0SMatthew Dillon 3991de864f0SMatthew Dillon /* 4001de864f0SMatthew Dillon * Move marker 4011de864f0SMatthew Dillon */ 402027193ebSMatthew Dillon TAILQ_REMOVE( 403027193ebSMatthew Dillon &vm_page_queues[marker->queue].pl, marker, pageq); 404027193ebSMatthew Dillon TAILQ_INSERT_AFTER( 405027193ebSMatthew Dillon &vm_page_queues[marker->queue].pl, m, marker, pageq); 406f5f6d247SMatthew Dillon 407f5f6d247SMatthew Dillon /* 408f5f6d247SMatthew Dillon * Ignore markers and ignore pages that already have a swap 409f5f6d247SMatthew Dillon * assignment. 410f5f6d247SMatthew Dillon */ 4110bf81261SMatthew Dillon if (m->flags & (PG_MARKER | PG_SWAPPED)) 412096e95c0SMatthew Dillon continue; 413b12defdcSMatthew Dillon if (vm_page_busy_try(m, TRUE)) 414096e95c0SMatthew Dillon continue; 415027193ebSMatthew Dillon vm_page_queues_spin_unlock(marker->queue); 416b12defdcSMatthew Dillon 417b12defdcSMatthew Dillon if ((object = m->object) == NULL) { 418b12defdcSMatthew Dillon vm_page_wakeup(m); 419027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 420b12defdcSMatthew Dillon continue; 421b12defdcSMatthew Dillon } 422b12defdcSMatthew Dillon vm_object_hold(object); 423b12defdcSMatthew Dillon if (m->object != object) { 424b12defdcSMatthew Dillon vm_object_drop(object); 425b12defdcSMatthew Dillon vm_page_wakeup(m); 426027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 427b12defdcSMatthew Dillon continue; 428b12defdcSMatthew Dillon } 429b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 430b12defdcSMatthew Dillon vm_object_drop(object); 431b12defdcSMatthew Dillon vm_page_wakeup(m); 432027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 433b12defdcSMatthew Dillon continue; 434b12defdcSMatthew Dillon } 435b12defdcSMatthew Dillon 436c504e38eSMatthew Dillon vp = object->handle; 437b12defdcSMatthew Dillon if (vp == NULL) { 438b12defdcSMatthew Dillon vm_object_drop(object); 439b12defdcSMatthew Dillon vm_page_wakeup(m); 440027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 441c504e38eSMatthew Dillon continue; 442b12defdcSMatthew Dillon } 443d3070b8dSMatthew Dillon 444c504e38eSMatthew Dillon switch(vp->v_type) { 445c504e38eSMatthew Dillon case VREG: 446e9b56058SMatthew Dillon /* 447bfa86281SMatthew Dillon * PG_NOTMETA generically means 'don't swapcache this', 448bfa86281SMatthew Dillon * and HAMMER will set this for regular data buffers 449bfa86281SMatthew Dillon * (and leave it unset for meta-data buffers) as 450bfa86281SMatthew Dillon * appropriate when double buffering is enabled. 451bfa86281SMatthew Dillon */ 452b12defdcSMatthew Dillon if (m->flags & PG_NOTMETA) { 453b12defdcSMatthew Dillon vm_object_drop(object); 454b12defdcSMatthew Dillon vm_page_wakeup(m); 455027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 456bfa86281SMatthew Dillon continue; 457b12defdcSMatthew Dillon } 458bfa86281SMatthew Dillon 459bfa86281SMatthew Dillon /* 460e9b56058SMatthew Dillon * If data_enable is 0 do not try to swapcache data. 461e9b56058SMatthew Dillon * If use_chflags is set then only swapcache data for 462e9b56058SMatthew Dillon * VSWAPCACHE marked vnodes, otherwise any vnode. 463e9b56058SMatthew Dillon */ 464e9b56058SMatthew Dillon if (vm_swapcache_data_enable == 0 || 465e9b56058SMatthew Dillon ((vp->v_flag & VSWAPCACHE) == 0 && 466e9b56058SMatthew Dillon vm_swapcache_use_chflags)) { 467b12defdcSMatthew Dillon vm_object_drop(object); 468b12defdcSMatthew Dillon vm_page_wakeup(m); 469027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 470c504e38eSMatthew Dillon continue; 471e9b56058SMatthew Dillon } 472d3070b8dSMatthew Dillon if (vm_swapcache_maxfilesize && 473d3070b8dSMatthew Dillon object->size > 474d3070b8dSMatthew Dillon (vm_swapcache_maxfilesize >> PAGE_SHIFT)) { 475b12defdcSMatthew Dillon vm_object_drop(object); 476b12defdcSMatthew Dillon vm_page_wakeup(m); 477027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 478d3070b8dSMatthew Dillon continue; 479d3070b8dSMatthew Dillon } 480aabd5ce8SMatthew Dillon isblkdev = 0; 481c504e38eSMatthew Dillon break; 482c504e38eSMatthew Dillon case VCHR: 483aabd5ce8SMatthew Dillon /* 484bfa86281SMatthew Dillon * PG_NOTMETA generically means 'don't swapcache this', 485bfa86281SMatthew Dillon * and HAMMER will set this for regular data buffers 486bfa86281SMatthew Dillon * (and leave it unset for meta-data buffers) as 487bfa86281SMatthew Dillon * appropriate when double buffering is enabled. 488aabd5ce8SMatthew Dillon */ 489b12defdcSMatthew Dillon if (m->flags & PG_NOTMETA) { 490b12defdcSMatthew Dillon vm_object_drop(object); 491b12defdcSMatthew Dillon vm_page_wakeup(m); 492027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 493aabd5ce8SMatthew Dillon continue; 494b12defdcSMatthew Dillon } 495b12defdcSMatthew Dillon if (vm_swapcache_meta_enable == 0) { 496b12defdcSMatthew Dillon vm_object_drop(object); 497b12defdcSMatthew Dillon vm_page_wakeup(m); 498027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 499c504e38eSMatthew Dillon continue; 500b12defdcSMatthew Dillon } 501aabd5ce8SMatthew Dillon isblkdev = 1; 502c504e38eSMatthew Dillon break; 503c504e38eSMatthew Dillon default: 504b12defdcSMatthew Dillon vm_object_drop(object); 505b12defdcSMatthew Dillon vm_page_wakeup(m); 506027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 507c504e38eSMatthew Dillon continue; 508c504e38eSMatthew Dillon } 509096e95c0SMatthew Dillon 510096e95c0SMatthew Dillon 511096e95c0SMatthew Dillon /* 5123ffc7051SMatthew Dillon * Assign swap and initiate I/O. 5133ffc7051SMatthew Dillon * 5143ffc7051SMatthew Dillon * (adjust for the --count which also occurs in the loop) 515096e95c0SMatthew Dillon */ 5160bf81261SMatthew Dillon count -= vm_swapcached_flush(m, isblkdev); 517096e95c0SMatthew Dillon 518096e95c0SMatthew Dillon /* 519096e95c0SMatthew Dillon * Setup for next loop using marker. 520096e95c0SMatthew Dillon */ 521b12defdcSMatthew Dillon vm_object_drop(object); 522027193ebSMatthew Dillon vm_page_queues_spin_lock(marker->queue); 523096e95c0SMatthew Dillon } 5241e5196f0SMatthew Dillon 5251e5196f0SMatthew Dillon /* 526b12defdcSMatthew Dillon * The marker could wind up at the end, which is ok. If we hit the 527b12defdcSMatthew Dillon * end of the list adjust the heuristic. 5281e5196f0SMatthew Dillon * 5291e5196f0SMatthew Dillon * Earlier inactive pages that were dirty and become clean 5301e5196f0SMatthew Dillon * are typically moved to the end of PQ_INACTIVE by virtue 5311e5196f0SMatthew Dillon * of vfs_vmio_release() when they become unwired from the 5321e5196f0SMatthew Dillon * buffer cache. 5331e5196f0SMatthew Dillon */ 534027193ebSMatthew Dillon vm_page_queues_spin_unlock(marker->queue); 53564949baaSMatthew Dillon 53664949baaSMatthew Dillon /* 53764949baaSMatthew Dillon * m invalid but can be used to test for NULL 53864949baaSMatthew Dillon */ 53964949baaSMatthew Dillon return (m == NULL); 540096e95c0SMatthew Dillon } 541096e95c0SMatthew Dillon 542096e95c0SMatthew Dillon /* 543b12defdcSMatthew Dillon * Flush the specified page using the swap_pager. The page 544b12defdcSMatthew Dillon * must be busied by the caller and its disposition will become 545b12defdcSMatthew Dillon * the responsibility of this function. 5463ffc7051SMatthew Dillon * 5473ffc7051SMatthew Dillon * Try to collect surrounding pages, including pages which may 5483ffc7051SMatthew Dillon * have already been assigned swap. Try to cluster within a 5493ffc7051SMatthew Dillon * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block 5503ffc7051SMatthew Dillon * to match what swap_pager_putpages() can do. 5513ffc7051SMatthew Dillon * 5523ffc7051SMatthew Dillon * We also want to try to match against the buffer cache blocksize 5533ffc7051SMatthew Dillon * but we don't really know what it is here. Since the buffer cache 5543ffc7051SMatthew Dillon * wires and unwires pages in groups the fact that we skip wired pages 5553ffc7051SMatthew Dillon * should be sufficient. 5563ffc7051SMatthew Dillon * 5573ffc7051SMatthew Dillon * Returns a count of pages we might have flushed (minimum 1) 558096e95c0SMatthew Dillon */ 559096e95c0SMatthew Dillon static 5603ffc7051SMatthew Dillon int 561aabd5ce8SMatthew Dillon vm_swapcached_flush(vm_page_t m, int isblkdev) 562096e95c0SMatthew Dillon { 563096e95c0SMatthew Dillon vm_object_t object; 5643ffc7051SMatthew Dillon vm_page_t marray[SWAP_META_PAGES]; 5653ffc7051SMatthew Dillon vm_pindex_t basei; 5663ffc7051SMatthew Dillon int rtvals[SWAP_META_PAGES]; 5673ffc7051SMatthew Dillon int x; 5683ffc7051SMatthew Dillon int i; 5693ffc7051SMatthew Dillon int j; 5703ffc7051SMatthew Dillon int count; 571b12defdcSMatthew Dillon int error; 572096e95c0SMatthew Dillon 573096e95c0SMatthew Dillon vm_page_io_start(m); 574096e95c0SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 575096e95c0SMatthew Dillon object = m->object; 576b12defdcSMatthew Dillon vm_object_hold(object); 577096e95c0SMatthew Dillon 5783ffc7051SMatthew Dillon /* 5793ffc7051SMatthew Dillon * Try to cluster around (m), keeping in mind that the swap pager 5803ffc7051SMatthew Dillon * can only do SMAP_META_PAGES worth of continguous write. 5813ffc7051SMatthew Dillon */ 5823ffc7051SMatthew Dillon x = (int)m->pindex & SWAP_META_MASK; 5833ffc7051SMatthew Dillon marray[x] = m; 5843ffc7051SMatthew Dillon basei = m->pindex; 585b12defdcSMatthew Dillon vm_page_wakeup(m); 5863ffc7051SMatthew Dillon 5873ffc7051SMatthew Dillon for (i = x - 1; i >= 0; --i) { 588b12defdcSMatthew Dillon m = vm_page_lookup_busy_try(object, basei - x + i, 589b12defdcSMatthew Dillon TRUE, &error); 590b12defdcSMatthew Dillon if (error || m == NULL) 5913ffc7051SMatthew Dillon break; 592b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 593b12defdcSMatthew Dillon vm_page_wakeup(m); 5943ffc7051SMatthew Dillon break; 595b12defdcSMatthew Dillon } 596b12defdcSMatthew Dillon if (isblkdev && (m->flags & PG_NOTMETA)) { 597b12defdcSMatthew Dillon vm_page_wakeup(m); 598aabd5ce8SMatthew Dillon break; 599b12defdcSMatthew Dillon } 6003ffc7051SMatthew Dillon vm_page_io_start(m); 6013ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 6023ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 6033ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 6043ffc7051SMatthew Dillon vm_page_deactivate(m); 6053ffc7051SMatthew Dillon } 6063ffc7051SMatthew Dillon marray[i] = m; 607b12defdcSMatthew Dillon vm_page_wakeup(m); 6083ffc7051SMatthew Dillon } 6093ffc7051SMatthew Dillon ++i; 6103ffc7051SMatthew Dillon 6113ffc7051SMatthew Dillon for (j = x + 1; j < SWAP_META_PAGES; ++j) { 612b12defdcSMatthew Dillon m = vm_page_lookup_busy_try(object, basei - x + j, 613b12defdcSMatthew Dillon TRUE, &error); 614b12defdcSMatthew Dillon if (error || m == NULL) 6153ffc7051SMatthew Dillon break; 616b12defdcSMatthew Dillon if (vm_swapcache_test(m)) { 617b12defdcSMatthew Dillon vm_page_wakeup(m); 6183ffc7051SMatthew Dillon break; 619b12defdcSMatthew Dillon } 620b12defdcSMatthew Dillon if (isblkdev && (m->flags & PG_NOTMETA)) { 621b12defdcSMatthew Dillon vm_page_wakeup(m); 622aabd5ce8SMatthew Dillon break; 623b12defdcSMatthew Dillon } 6243ffc7051SMatthew Dillon vm_page_io_start(m); 6253ffc7051SMatthew Dillon vm_page_protect(m, VM_PROT_READ); 6263ffc7051SMatthew Dillon if (m->queue - m->pc == PQ_CACHE) { 6273ffc7051SMatthew Dillon vm_page_unqueue_nowakeup(m); 6283ffc7051SMatthew Dillon vm_page_deactivate(m); 6293ffc7051SMatthew Dillon } 6303ffc7051SMatthew Dillon marray[j] = m; 631b12defdcSMatthew Dillon vm_page_wakeup(m); 6323ffc7051SMatthew Dillon } 6333ffc7051SMatthew Dillon 6343ffc7051SMatthew Dillon count = j - i; 6353ffc7051SMatthew Dillon vm_object_pip_add(object, count); 6363ffc7051SMatthew Dillon swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i); 6373ffc7051SMatthew Dillon vm_swapcache_write_count += count * PAGE_SIZE; 6383ffc7051SMatthew Dillon vm_swapcache_curburst -= count * PAGE_SIZE; 6393ffc7051SMatthew Dillon 6403ffc7051SMatthew Dillon while (i < j) { 6413ffc7051SMatthew Dillon if (rtvals[i] != VM_PAGER_PEND) { 642b12defdcSMatthew Dillon vm_page_busy_wait(marray[i], FALSE, "swppgfd"); 6433ffc7051SMatthew Dillon vm_page_io_finish(marray[i]); 644b12defdcSMatthew Dillon vm_page_wakeup(marray[i]); 645096e95c0SMatthew Dillon vm_object_pip_wakeup(object); 646096e95c0SMatthew Dillon } 6473ffc7051SMatthew Dillon ++i; 6483ffc7051SMatthew Dillon } 649b12defdcSMatthew Dillon vm_object_drop(object); 6503ffc7051SMatthew Dillon return(count); 651096e95c0SMatthew Dillon } 65200a3fdcaSMatthew Dillon 6533ffc7051SMatthew Dillon /* 6543ffc7051SMatthew Dillon * Test whether a VM page is suitable for writing to the swapcache. 6553ffc7051SMatthew Dillon * Does not test m->queue, PG_MARKER, or PG_SWAPPED. 6563ffc7051SMatthew Dillon * 6573ffc7051SMatthew Dillon * Returns 0 on success, 1 on failure 6583ffc7051SMatthew Dillon */ 6593ffc7051SMatthew Dillon static int 6603ffc7051SMatthew Dillon vm_swapcache_test(vm_page_t m) 6613ffc7051SMatthew Dillon { 6623ffc7051SMatthew Dillon vm_object_t object; 6633ffc7051SMatthew Dillon 664b12defdcSMatthew Dillon if (m->flags & PG_UNMANAGED) 6653ffc7051SMatthew Dillon return(1); 666b12defdcSMatthew Dillon if (m->hold_count || m->wire_count) 6673ffc7051SMatthew Dillon return(1); 6683ffc7051SMatthew Dillon if (m->valid != VM_PAGE_BITS_ALL) 6693ffc7051SMatthew Dillon return(1); 6703ffc7051SMatthew Dillon if (m->dirty & m->valid) 6713ffc7051SMatthew Dillon return(1); 6723ffc7051SMatthew Dillon if ((object = m->object) == NULL) 6733ffc7051SMatthew Dillon return(1); 6743ffc7051SMatthew Dillon if (object->type != OBJT_VNODE || 6753ffc7051SMatthew Dillon (object->flags & OBJ_DEAD)) { 6763ffc7051SMatthew Dillon return(1); 6773ffc7051SMatthew Dillon } 6783ffc7051SMatthew Dillon vm_page_test_dirty(m); 6793ffc7051SMatthew Dillon if (m->dirty & m->valid) 6803ffc7051SMatthew Dillon return(1); 6813ffc7051SMatthew Dillon return(0); 6823ffc7051SMatthew Dillon } 6833ffc7051SMatthew Dillon 6843ffc7051SMatthew Dillon /* 685f5f6d247SMatthew Dillon * Cleaning pass. 686f5f6d247SMatthew Dillon * 687f5f6d247SMatthew Dillon * We clean whole objects up to 16MB 6883ffc7051SMatthew Dillon */ 68900a3fdcaSMatthew Dillon static 69000a3fdcaSMatthew Dillon void 691fde6be6aSMatthew Dillon vm_swapcache_cleaning(vm_object_t marker, struct vm_object_hash **swindexp) 69200a3fdcaSMatthew Dillon { 69300a3fdcaSMatthew Dillon vm_object_t object; 69400a3fdcaSMatthew Dillon struct vnode *vp; 69500a3fdcaSMatthew Dillon int count; 6960bf81261SMatthew Dillon int scount; 69700a3fdcaSMatthew Dillon int n; 69800a3fdcaSMatthew Dillon 69900a3fdcaSMatthew Dillon count = vm_swapcache_maxlaunder; 7000bf81261SMatthew Dillon scount = vm_swapcache_maxscan; 70100a3fdcaSMatthew Dillon 70200a3fdcaSMatthew Dillon /* 70300a3fdcaSMatthew Dillon * Look for vnode objects 70400a3fdcaSMatthew Dillon */ 705fde6be6aSMatthew Dillon lwkt_gettoken(&(*swindexp)->token); 7062de4f77eSMatthew Dillon 707b2286ce7SMatthew Dillon outerloop: 708f5f6d247SMatthew Dillon while ((object = TAILQ_NEXT(marker, object_list)) != NULL) { 7092f2d9e58SVenkatesh Srinivas /* 710f5f6d247SMatthew Dillon * We have to skip markers. We cannot hold/drop marker 711f5f6d247SMatthew Dillon * objects! 7122f2d9e58SVenkatesh Srinivas */ 713f5f6d247SMatthew Dillon if (object->type == OBJT_MARKER) { 7147b00fbb4SMatthew Dillon vm_swapcache_movemarker(marker, *swindexp, object); 71500a3fdcaSMatthew Dillon continue; 7162f2d9e58SVenkatesh Srinivas } 71700a3fdcaSMatthew Dillon 71800a3fdcaSMatthew Dillon /* 719f5f6d247SMatthew Dillon * Safety, or in case there are millions of VM objects 720f5f6d247SMatthew Dillon * without swapcache backing. 72100a3fdcaSMatthew Dillon */ 7220bf81261SMatthew Dillon if (--scount <= 0) 7237b00fbb4SMatthew Dillon goto breakout; 72400a3fdcaSMatthew Dillon 72500a3fdcaSMatthew Dillon /* 726f5f6d247SMatthew Dillon * We must hold the object before potentially yielding. 72700a3fdcaSMatthew Dillon */ 728f5f6d247SMatthew Dillon vm_object_hold(object); 729f5f6d247SMatthew Dillon lwkt_yield(); 730f5f6d247SMatthew Dillon 731f5f6d247SMatthew Dillon /* 732f5f6d247SMatthew Dillon * Only operate on live VNODE objects that are either 733f5f6d247SMatthew Dillon * VREG or VCHR (VCHR for meta-data). 734f5f6d247SMatthew Dillon */ 735f5f6d247SMatthew Dillon if ((object->type != OBJT_VNODE) || 736f5f6d247SMatthew Dillon ((object->flags & OBJ_DEAD) || 737f5f6d247SMatthew Dillon object->swblock_count == 0) || 738f5f6d247SMatthew Dillon ((vp = object->handle) == NULL) || 739f5f6d247SMatthew Dillon (vp->v_type != VREG && vp->v_type != VCHR)) { 740f5f6d247SMatthew Dillon vm_object_drop(object); 741f5f6d247SMatthew Dillon /* object may be invalid now */ 7427b00fbb4SMatthew Dillon vm_swapcache_movemarker(marker, *swindexp, object); 743f5f6d247SMatthew Dillon continue; 744f5f6d247SMatthew Dillon } 745f5f6d247SMatthew Dillon 746f5f6d247SMatthew Dillon /* 747f5f6d247SMatthew Dillon * Reset the object pindex stored in the marker if the 748f5f6d247SMatthew Dillon * working object has changed. 749f5f6d247SMatthew Dillon */ 750f5f6d247SMatthew Dillon if (marker->backing_object != object) { 751f5f6d247SMatthew Dillon marker->size = 0; 752f5f6d247SMatthew Dillon marker->backing_object_offset = 0; 753f5f6d247SMatthew Dillon marker->backing_object = object; 754f5f6d247SMatthew Dillon } 75500a3fdcaSMatthew Dillon 75600a3fdcaSMatthew Dillon /* 75700a3fdcaSMatthew Dillon * Look for swblocks starting at our iterator. 75800a3fdcaSMatthew Dillon * 75900a3fdcaSMatthew Dillon * The swap_pager_condfree() function attempts to free 76000a3fdcaSMatthew Dillon * swap space starting at the specified index. The index 76100a3fdcaSMatthew Dillon * will be updated on return. The function will return 76200a3fdcaSMatthew Dillon * a scan factor (NOT the number of blocks freed). 76300a3fdcaSMatthew Dillon * 76400a3fdcaSMatthew Dillon * If it must cut its scan of the object short due to an 76500a3fdcaSMatthew Dillon * excessive number of swblocks, or is able to free the 76600a3fdcaSMatthew Dillon * requested number of blocks, it will return n >= count 76700a3fdcaSMatthew Dillon * and we break and pick it back up on a future attempt. 768f5f6d247SMatthew Dillon * 769f5f6d247SMatthew Dillon * Scan the object linearly and try to batch large sets of 770f5f6d247SMatthew Dillon * blocks that are likely to clean out entire swap radix 771f5f6d247SMatthew Dillon * tree leafs. 77200a3fdcaSMatthew Dillon */ 773739be60bSMatthew Dillon lwkt_token_swap(); 774fde6be6aSMatthew Dillon lwkt_reltoken(&(*swindexp)->token); 77527b6ee03SMatthew Dillon 776f5f6d247SMatthew Dillon n = swap_pager_condfree(object, &marker->size, 777f5f6d247SMatthew Dillon (count + SWAP_META_MASK) & ~SWAP_META_MASK); 7782f2d9e58SVenkatesh Srinivas 779f5f6d247SMatthew Dillon vm_object_drop(object); /* object may be invalid now */ 780fde6be6aSMatthew Dillon lwkt_gettoken(&(*swindexp)->token); 7812f2d9e58SVenkatesh Srinivas 78200a3fdcaSMatthew Dillon /* 783f5f6d247SMatthew Dillon * If we have exhausted the object or deleted our per-pass 784f5f6d247SMatthew Dillon * page limit then move us to the next object. Note that 785f5f6d247SMatthew Dillon * the current object may no longer be on the vm_object_list. 78600a3fdcaSMatthew Dillon */ 787f5f6d247SMatthew Dillon if (n <= 0 || 788f5f6d247SMatthew Dillon marker->backing_object_offset > vm_swapcache_cleanperobj) { 7897b00fbb4SMatthew Dillon vm_swapcache_movemarker(marker, *swindexp, object); 79000a3fdcaSMatthew Dillon } 79100a3fdcaSMatthew Dillon 79200a3fdcaSMatthew Dillon /* 793f5f6d247SMatthew Dillon * If we have exhausted our max-launder stop for now. 79400a3fdcaSMatthew Dillon */ 795f5f6d247SMatthew Dillon count -= n; 796f5f6d247SMatthew Dillon marker->backing_object_offset += n * PAGE_SIZE; 797f5f6d247SMatthew Dillon if (count < 0) 7987b00fbb4SMatthew Dillon goto breakout; 799f5f6d247SMatthew Dillon } 8007a175765SMatthew Dillon 8017a175765SMatthew Dillon /* 8027b00fbb4SMatthew Dillon * Iterate vm_object_lists[] hash table 8037a175765SMatthew Dillon */ 804fde6be6aSMatthew Dillon TAILQ_REMOVE(&(*swindexp)->list, marker, object_list); 805fde6be6aSMatthew Dillon lwkt_reltoken(&(*swindexp)->token); 806fde6be6aSMatthew Dillon if (++*swindexp >= &vm_object_hash[VMOBJ_HSIZE]) 807fde6be6aSMatthew Dillon *swindexp = &vm_object_hash[0]; 808fde6be6aSMatthew Dillon lwkt_gettoken(&(*swindexp)->token); 809fde6be6aSMatthew Dillon TAILQ_INSERT_HEAD(&(*swindexp)->list, marker, object_list); 8107a175765SMatthew Dillon 811fde6be6aSMatthew Dillon if (*swindexp != &vm_object_hash[0]) 8127b00fbb4SMatthew Dillon goto outerloop; 8137b00fbb4SMatthew Dillon 8147b00fbb4SMatthew Dillon breakout: 815fde6be6aSMatthew Dillon lwkt_reltoken(&(*swindexp)->token); 81600a3fdcaSMatthew Dillon } 817f5f6d247SMatthew Dillon 818f5f6d247SMatthew Dillon /* 819f5f6d247SMatthew Dillon * Move the marker past the current object. Object can be stale, but we 820f5f6d247SMatthew Dillon * still need it to determine if the marker has to be moved. If the object 821f5f6d247SMatthew Dillon * is still the 'current object' (object after the marker), we hop-scotch 822f5f6d247SMatthew Dillon * the marker past it. 823f5f6d247SMatthew Dillon */ 824f5f6d247SMatthew Dillon static void 825fde6be6aSMatthew Dillon vm_swapcache_movemarker(vm_object_t marker, struct vm_object_hash *swindex, 826fde6be6aSMatthew Dillon vm_object_t object) 827f5f6d247SMatthew Dillon { 828f5f6d247SMatthew Dillon if (TAILQ_NEXT(marker, object_list) == object) { 829fde6be6aSMatthew Dillon TAILQ_REMOVE(&swindex->list, marker, object_list); 830fde6be6aSMatthew Dillon TAILQ_INSERT_AFTER(&swindex->list, object, marker, object_list); 831f5f6d247SMatthew Dillon } 832f5f6d247SMatthew Dillon } 833