2 * Copyright (c) 2010 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Implement the swapcache daemon. When enabled swap is assumed to be
37 * configured on a fast storage device such as a SSD. Swap is assigned
38 * to clean vnode-backed pages in the inactive queue, clustered by object
39 * if possible, and written out. The swap assignment sticks around even
40 * after the underlying pages have been recycled.
42 * The daemon manages write bandwidth based on sysctl settings to control
45 * The vnode strategy code will check for the swap assignments and divert
46 * reads to the swap device when the data is present in the swapcache.
48 * This operates on both regular files and the block device vnodes used by
49 * filesystems to manage meta-data.
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
57 #include <sys/kthread.h>
58 #include <sys/resourcevar.h>
59 #include <sys/signalvar.h>
60 #include <sys/vnode.h>
61 #include <sys/vmmeter.h>
62 #include <sys/sysctl.h>
65 #include <vm/vm_param.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_map.h>
70 #include <vm/vm_pageout.h>
71 #include <vm/vm_pager.h>
72 #include <vm/swap_pager.h>
73 #include <vm/vm_extern.h>
75 #include <sys/thread2.h>
76 #include <vm/vm_page2.h>
78 #define INACTIVE_LIST (&vm_page_queues[PQ_INACTIVE].pl)
80 /* the kernel process "vm_pageout"*/
81 static void vm_swapcached (void);
82 static int vm_swapcached_flush (vm_page_t m);
83 static int vm_swapcache_test(vm_page_t m);
84 static void vm_swapcache_writing(vm_page_t marker);
85 static void vm_swapcache_cleaning(vm_object_t marker);
86 struct thread *swapcached_thread;
88 static struct kproc_desc swpc_kp = {
93 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
95 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
97 int vm_swapcache_read_enable;
98 int vm_swapcache_inactive_heuristic;
99 static int vm_swapcache_sleep;
100 static int vm_swapcache_maxlaunder = 256;
101 static int vm_swapcache_data_enable = 0;
102 static int vm_swapcache_meta_enable = 0;
103 static int vm_swapcache_maxswappct = 75;
104 static int vm_swapcache_hysteresis;
105 static int vm_swapcache_use_chflags = 1; /* require chflags cache */
106 static int64_t vm_swapcache_minburst = 10000000LL; /* 10MB */
107 static int64_t vm_swapcache_curburst = 4000000000LL; /* 4G after boot */
108 static int64_t vm_swapcache_maxburst = 2000000000LL; /* 2G nominal max */
109 static int64_t vm_swapcache_accrate = 100000LL; /* 100K/s */
110 static int64_t vm_swapcache_write_count;
111 static int64_t vm_swapcache_maxfilesize;
113 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
114 CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
116 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
117 CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
118 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
119 CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
120 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
121 CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
122 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
123 CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
124 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
125 CTLFLAG_RW, &vm_swapcache_hysteresis, 0, "");
126 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
127 CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
129 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
130 CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
131 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
132 CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
134 CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
135 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
136 CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
137 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
138 CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
139 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
140 CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
142 #define SWAPMAX(adj) \
143 ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
146 * vm_swapcached is the high level pageout daemon.
151 enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
152 enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
153 struct vm_page page_marker;
154 struct vm_object object_marker;
159 curthread->td_flags |= TDF_SYSTHREAD;
163 * Initialize our marker for the inactive scan (SWAPC_WRITING)
165 bzero(&page_marker, sizeof(page_marker));
166 page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
167 page_marker.queue = PQ_INACTIVE;
168 page_marker.wire_count = 1;
169 TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq);
170 vm_swapcache_hysteresis = vmstats.v_inactive_target / 2;
171 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
174 * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
176 bzero(&object_marker, sizeof(object_marker));
177 object_marker.type = OBJT_MARKER;
178 TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
182 * Check every 5 seconds when not enabled or if no swap
185 if ((vm_swapcache_data_enable == 0 &&
186 vm_swapcache_meta_enable == 0) ||
188 tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
193 * Polling rate when enabled is approximately 10 hz.
195 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
198 * State hysteresis. Generate write activity up to 75% of
199 * swap, then clean out swap assignments down to 70%, then
202 if (state == SWAPC_WRITING) {
203 if (vm_swap_cache_use > SWAPMAX(0))
204 state = SWAPC_CLEANING;
206 if (vm_swap_cache_use < SWAPMAX(-5))
207 state = SWAPC_WRITING;
211 * We are allowed to continue accumulating burst value
212 * in either state. Allow the user to set curburst > maxburst
213 * for the initial load-in.
215 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
216 vm_swapcache_curburst += vm_swapcache_accrate / 10;
217 if (vm_swapcache_curburst > vm_swapcache_maxburst)
218 vm_swapcache_curburst = vm_swapcache_maxburst;
222 * We don't want to nickle-and-dime the scan as that will
223 * create unnecessary fragmentation. The minimum burst
224 * is one-seconds worth of accumulation.
226 if (state == SWAPC_WRITING) {
227 if (vm_swapcache_curburst >= vm_swapcache_accrate) {
228 if (burst == SWAPB_BURSTING) {
229 vm_swapcache_writing(&page_marker);
230 if (vm_swapcache_curburst <= 0)
231 burst = SWAPB_RECOVERING;
232 } else if (vm_swapcache_curburst >
233 vm_swapcache_minburst) {
234 vm_swapcache_writing(&page_marker);
235 burst = SWAPB_BURSTING;
239 vm_swapcache_cleaning(&object_marker);
242 TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq);
243 TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
248 vm_swapcache_writing(vm_page_t marker)
256 * Deal with an overflow of the heuristic counter or if the user
257 * manually changes the hysteresis.
259 * Try to avoid small incremental pageouts by waiting for enough
260 * pages to buildup in the inactive queue to hopefully get a good
261 * burst in. This heuristic is bumped by the VM system and reset
262 * when our scan hits the end of the queue.
264 if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis)
265 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
266 if (vm_swapcache_inactive_heuristic < 0)
270 * Scan the inactive queue from our marker to locate
271 * suitable pages to push to the swap cache.
273 * We are looking for clean vnode-backed pages.
275 * NOTE: PG_SWAPPED pages in particular are not part of
276 * our count because once the cache stabilizes we
277 * can end up with a very high datarate of VM pages
281 count = vm_swapcache_maxlaunder;
283 while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) {
284 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
288 if (vm_swapcache_curburst < 0)
290 if (vm_swapcache_test(m))
300 * If data_enable is 0 do not try to swapcache data.
301 * If use_chflags is set then only swapcache data for
302 * VSWAPCACHE marked vnodes, otherwise any vnode.
304 if (vm_swapcache_data_enable == 0 ||
305 ((vp->v_flag & VSWAPCACHE) == 0 &&
306 vm_swapcache_use_chflags)) {
309 if (vm_swapcache_maxfilesize &&
311 (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
316 if (vm_swapcache_meta_enable == 0)
324 * Ok, move the marker and soft-busy the page.
326 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
327 TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq);
330 * Assign swap and initiate I/O.
332 * (adjust for the --count which also occurs in the loop)
334 count -= vm_swapcached_flush(m) - 1;
337 * Setup for next loop using marker.
343 * Cleanup marker position. If we hit the end of the
344 * list the marker is placed at the tail. Newly deactivated
345 * pages will be placed after it.
347 * Earlier inactive pages that were dirty and become clean
348 * are typically moved to the end of PQ_INACTIVE by virtue
349 * of vfs_vmio_release() when they become unwired from the
352 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
354 TAILQ_INSERT_BEFORE(m, marker, pageq);
356 TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq);
357 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
362 * Flush the specified page using the swap_pager.
364 * Try to collect surrounding pages, including pages which may
365 * have already been assigned swap. Try to cluster within a
366 * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
367 * to match what swap_pager_putpages() can do.
369 * We also want to try to match against the buffer cache blocksize
370 * but we don't really know what it is here. Since the buffer cache
371 * wires and unwires pages in groups the fact that we skip wired pages
372 * should be sufficient.
374 * Returns a count of pages we might have flushed (minimum 1)
378 vm_swapcached_flush(vm_page_t m)
381 vm_page_t marray[SWAP_META_PAGES];
383 int rtvals[SWAP_META_PAGES];
390 vm_page_protect(m, VM_PROT_READ);
394 * Try to cluster around (m), keeping in mind that the swap pager
395 * can only do SMAP_META_PAGES worth of continguous write.
397 x = (int)m->pindex & SWAP_META_MASK;
401 for (i = x - 1; i >= 0; --i) {
402 m = vm_page_lookup(object, basei - x + i);
405 if (vm_swapcache_test(m))
408 vm_page_protect(m, VM_PROT_READ);
409 if (m->queue - m->pc == PQ_CACHE) {
410 vm_page_unqueue_nowakeup(m);
411 vm_page_deactivate(m);
417 for (j = x + 1; j < SWAP_META_PAGES; ++j) {
418 m = vm_page_lookup(object, basei - x + j);
421 if (vm_swapcache_test(m))
424 vm_page_protect(m, VM_PROT_READ);
425 if (m->queue - m->pc == PQ_CACHE) {
426 vm_page_unqueue_nowakeup(m);
427 vm_page_deactivate(m);
433 vm_object_pip_add(object, count);
434 swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
435 vm_swapcache_write_count += count * PAGE_SIZE;
436 vm_swapcache_curburst -= count * PAGE_SIZE;
439 if (rtvals[i] != VM_PAGER_PEND) {
440 vm_page_io_finish(marray[i]);
441 vm_object_pip_wakeup(object);
449 * Test whether a VM page is suitable for writing to the swapcache.
450 * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
452 * Returns 0 on success, 1 on failure
455 vm_swapcache_test(vm_page_t m)
459 if (m->flags & (PG_BUSY | PG_UNMANAGED | PG_NOTMETA))
461 if (m->busy || m->hold_count || m->wire_count)
463 if (m->valid != VM_PAGE_BITS_ALL)
465 if (m->dirty & m->valid)
467 if ((object = m->object) == NULL)
469 if (object->type != OBJT_VNODE ||
470 (object->flags & OBJ_DEAD)) {
473 vm_page_test_dirty(m);
474 if (m->dirty & m->valid)
484 vm_swapcache_cleaning(vm_object_t marker)
492 count = vm_swapcache_maxlaunder;
495 * Look for vnode objects
497 while ((object = TAILQ_NEXT(object, object_list)) != NULL && count--) {
498 if (object->type != OBJT_VNODE)
500 if ((object->flags & OBJ_DEAD) || object->swblock_count == 0)
502 if ((vp = object->handle) == NULL)
504 if (vp->v_type != VREG && vp->v_type != VCHR)
510 if (marker->backing_object != object)
514 * Move the marker so we can work on the VM object
516 TAILQ_REMOVE(&vm_object_list, marker, object_list);
517 TAILQ_INSERT_AFTER(&vm_object_list, object,
518 marker, object_list);
521 * Look for swblocks starting at our iterator.
523 * The swap_pager_condfree() function attempts to free
524 * swap space starting at the specified index. The index
525 * will be updated on return. The function will return
526 * a scan factor (NOT the number of blocks freed).
528 * If it must cut its scan of the object short due to an
529 * excessive number of swblocks, or is able to free the
530 * requested number of blocks, it will return n >= count
531 * and we break and pick it back up on a future attempt.
533 n = swap_pager_condfree(object, &marker->size, count);
546 * Adjust marker so we continue the scan from where we left off.
547 * When we reach the end we start back at the beginning.
549 TAILQ_REMOVE(&vm_object_list, marker, object_list);
551 TAILQ_INSERT_BEFORE(object, marker, object_list);
553 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
554 marker->backing_object = object;