4 * Copyright (c) 2010 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * Implement the swapcache daemon. When enabled swap is assumed to be
39 * configured on a fast storage device such as a SSD. Swap is assigned
40 * to clean vnode-backed pages in the inactive queue, clustered by object
41 * if possible, and written out. The swap assignment sticks around even
42 * after the underlying pages have been recycled.
44 * The daemon manages write bandwidth based on sysctl settings to control
47 * The vnode strategy code will check for the swap assignments and divert
48 * reads to the swap device when the data is present in the swapcache.
50 * This operates on both regular files and the block device vnodes used by
51 * filesystems to manage meta-data.
55 #include <sys/param.h>
56 #include <sys/systm.h>
57 #include <sys/kernel.h>
59 #include <sys/kthread.h>
60 #include <sys/resourcevar.h>
61 #include <sys/signalvar.h>
62 #include <sys/vnode.h>
63 #include <sys/vmmeter.h>
64 #include <sys/sysctl.h>
65 #include <sys/eventhandler.h>
68 #include <vm/vm_param.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_pageout.h>
74 #include <vm/vm_pager.h>
75 #include <vm/swap_pager.h>
76 #include <vm/vm_extern.h>
78 #include <sys/thread2.h>
79 #include <sys/spinlock2.h>
80 #include <vm/vm_page2.h>
82 #define INACTIVE_LIST (&vm_page_queues[PQ_INACTIVE].pl)
84 /* the kernel process "vm_pageout"*/
85 static int vm_swapcached_flush (vm_page_t m, int isblkdev);
86 static int vm_swapcache_test(vm_page_t m);
87 static void vm_swapcache_writing(vm_page_t marker);
88 static void vm_swapcache_cleaning(vm_object_t marker);
89 struct thread *swapcached_thread;
91 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
93 int vm_swapcache_read_enable;
94 int vm_swapcache_inactive_heuristic;
95 static int vm_swapcache_sleep;
96 static int vm_swapcache_maxlaunder = 256;
97 static int vm_swapcache_data_enable = 0;
98 static int vm_swapcache_meta_enable = 0;
99 static int vm_swapcache_maxswappct = 75;
100 static int vm_swapcache_hysteresis;
101 int vm_swapcache_use_chflags = 1; /* require chflags cache */
102 static int64_t vm_swapcache_minburst = 10000000LL; /* 10MB */
103 static int64_t vm_swapcache_curburst = 4000000000LL; /* 4G after boot */
104 static int64_t vm_swapcache_maxburst = 2000000000LL; /* 2G nominal max */
105 static int64_t vm_swapcache_accrate = 100000LL; /* 100K/s */
106 static int64_t vm_swapcache_write_count;
107 static int64_t vm_swapcache_maxfilesize;
109 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
110 CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
112 SYSCTL_INT(_vm_swapcache, OID_AUTO, data_enable,
113 CTLFLAG_RW, &vm_swapcache_data_enable, 0, "");
114 SYSCTL_INT(_vm_swapcache, OID_AUTO, meta_enable,
115 CTLFLAG_RW, &vm_swapcache_meta_enable, 0, "");
116 SYSCTL_INT(_vm_swapcache, OID_AUTO, read_enable,
117 CTLFLAG_RW, &vm_swapcache_read_enable, 0, "");
118 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxswappct,
119 CTLFLAG_RW, &vm_swapcache_maxswappct, 0, "");
120 SYSCTL_INT(_vm_swapcache, OID_AUTO, hysteresis,
121 CTLFLAG_RW, &vm_swapcache_hysteresis, 0, "");
122 SYSCTL_INT(_vm_swapcache, OID_AUTO, use_chflags,
123 CTLFLAG_RW, &vm_swapcache_use_chflags, 0, "");
125 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, minburst,
126 CTLFLAG_RW, &vm_swapcache_minburst, 0, "");
127 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, curburst,
128 CTLFLAG_RW, &vm_swapcache_curburst, 0, "");
129 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxburst,
130 CTLFLAG_RW, &vm_swapcache_maxburst, 0, "");
131 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, maxfilesize,
132 CTLFLAG_RW, &vm_swapcache_maxfilesize, 0, "");
133 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, accrate,
134 CTLFLAG_RW, &vm_swapcache_accrate, 0, "");
135 SYSCTL_QUAD(_vm_swapcache, OID_AUTO, write_count,
136 CTLFLAG_RW, &vm_swapcache_write_count, 0, "");
138 #define SWAPMAX(adj) \
139 ((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
142 * When shutting down the machine we want to stop swapcache operation
143 * immediately so swap is not accessed after devices have been shuttered.
146 shutdown_swapcache(void *arg __unused)
148 vm_swapcache_read_enable = 0;
149 vm_swapcache_data_enable = 0;
150 vm_swapcache_meta_enable = 0;
151 wakeup(&vm_swapcache_sleep); /* shortcut 5-second wait */
155 * vm_swapcached is the high level pageout daemon.
160 vm_swapcached_thread(void)
162 enum { SWAPC_WRITING, SWAPC_CLEANING } state = SWAPC_WRITING;
163 enum { SWAPB_BURSTING, SWAPB_RECOVERING } burst = SWAPB_BURSTING;
164 struct vm_page page_marker;
165 struct vm_object object_marker;
170 curthread->td_flags |= TDF_SYSTHREAD;
171 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
172 swapcached_thread, SHUTDOWN_PRI_FIRST);
173 EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_swapcache,
174 NULL, SHUTDOWN_PRI_SECOND);
177 * Initialize our marker for the inactive scan (SWAPC_WRITING)
179 bzero(&page_marker, sizeof(page_marker));
180 page_marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
181 page_marker.queue = PQ_INACTIVE;
182 page_marker.wire_count = 1;
184 vm_page_queues_spin_lock(PQ_INACTIVE);
185 TAILQ_INSERT_HEAD(INACTIVE_LIST, &page_marker, pageq);
186 vm_page_queues_spin_unlock(PQ_INACTIVE);
188 vm_swapcache_hysteresis = vmstats.v_inactive_target / 2;
189 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
192 * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
194 bzero(&object_marker, sizeof(object_marker));
195 object_marker.type = OBJT_MARKER;
196 lwkt_gettoken(&vmobj_token);
197 TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
198 lwkt_reltoken(&vmobj_token);
204 kproc_suspend_loop();
207 * Check every 5 seconds when not enabled or if no swap
210 if ((vm_swapcache_data_enable == 0 &&
211 vm_swapcache_meta_enable == 0) ||
213 tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
218 * Polling rate when enabled is approximately 10 hz.
220 tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
223 * State hysteresis. Generate write activity up to 75% of
224 * swap, then clean out swap assignments down to 70%, then
227 if (state == SWAPC_WRITING) {
228 if (vm_swap_cache_use > SWAPMAX(0))
229 state = SWAPC_CLEANING;
231 if (vm_swap_cache_use < SWAPMAX(-5))
232 state = SWAPC_WRITING;
236 * We are allowed to continue accumulating burst value
237 * in either state. Allow the user to set curburst > maxburst
238 * for the initial load-in.
240 if (vm_swapcache_curburst < vm_swapcache_maxburst) {
241 vm_swapcache_curburst += vm_swapcache_accrate / 10;
242 if (vm_swapcache_curburst > vm_swapcache_maxburst)
243 vm_swapcache_curburst = vm_swapcache_maxburst;
247 * We don't want to nickle-and-dime the scan as that will
248 * create unnecessary fragmentation. The minimum burst
249 * is one-seconds worth of accumulation.
251 if (state == SWAPC_WRITING) {
252 if (vm_swapcache_curburst >= vm_swapcache_accrate) {
253 if (burst == SWAPB_BURSTING) {
254 vm_swapcache_writing(&page_marker);
255 if (vm_swapcache_curburst <= 0)
256 burst = SWAPB_RECOVERING;
257 } else if (vm_swapcache_curburst >
258 vm_swapcache_minburst) {
259 vm_swapcache_writing(&page_marker);
260 burst = SWAPB_BURSTING;
264 vm_swapcache_cleaning(&object_marker);
269 * Cleanup (NOT REACHED)
271 vm_page_queues_spin_lock(PQ_INACTIVE);
272 TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq);
273 vm_page_queues_spin_unlock(PQ_INACTIVE);
275 lwkt_gettoken(&vmobj_token);
276 TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
277 lwkt_reltoken(&vmobj_token);
280 static struct kproc_desc swpc_kp = {
282 vm_swapcached_thread,
285 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
288 vm_swapcache_writing(vm_page_t marker)
297 * Deal with an overflow of the heuristic counter or if the user
298 * manually changes the hysteresis.
300 * Try to avoid small incremental pageouts by waiting for enough
301 * pages to buildup in the inactive queue to hopefully get a good
302 * burst in. This heuristic is bumped by the VM system and reset
303 * when our scan hits the end of the queue.
305 if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis)
306 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
307 if (vm_swapcache_inactive_heuristic < 0)
311 * Scan the inactive queue from our marker to locate
312 * suitable pages to push to the swap cache.
314 * We are looking for clean vnode-backed pages.
316 * NOTE: PG_SWAPPED pages in particular are not part of
317 * our count because once the cache stabilizes we
318 * can end up with a very high datarate of VM pages
321 count = vm_swapcache_maxlaunder;
323 vm_page_queues_spin_lock(PQ_INACTIVE);
324 while ((m = TAILQ_NEXT(marker, pageq)) != NULL && count-- > 0) {
325 KKASSERT(m->queue == PQ_INACTIVE);
327 if (vm_swapcache_curburst < 0)
329 TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
330 TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq);
331 if (m->flags & (PG_MARKER | PG_SWAPPED)) {
335 if (vm_page_busy_try(m, TRUE))
337 vm_page_queues_spin_unlock(PQ_INACTIVE);
339 if ((object = m->object) == NULL) {
341 vm_page_queues_spin_lock(PQ_INACTIVE);
344 vm_object_hold(object);
345 if (m->object != object) {
346 vm_object_drop(object);
348 vm_page_queues_spin_lock(PQ_INACTIVE);
351 if (vm_swapcache_test(m)) {
352 vm_object_drop(object);
354 vm_page_queues_spin_lock(PQ_INACTIVE);
360 vm_object_drop(object);
362 vm_page_queues_spin_lock(PQ_INACTIVE);
369 * PG_NOTMETA generically means 'don't swapcache this',
370 * and HAMMER will set this for regular data buffers
371 * (and leave it unset for meta-data buffers) as
372 * appropriate when double buffering is enabled.
374 if (m->flags & PG_NOTMETA) {
375 vm_object_drop(object);
377 vm_page_queues_spin_lock(PQ_INACTIVE);
382 * If data_enable is 0 do not try to swapcache data.
383 * If use_chflags is set then only swapcache data for
384 * VSWAPCACHE marked vnodes, otherwise any vnode.
386 if (vm_swapcache_data_enable == 0 ||
387 ((vp->v_flag & VSWAPCACHE) == 0 &&
388 vm_swapcache_use_chflags)) {
389 vm_object_drop(object);
391 vm_page_queues_spin_lock(PQ_INACTIVE);
394 if (vm_swapcache_maxfilesize &&
396 (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
397 vm_object_drop(object);
399 vm_page_queues_spin_lock(PQ_INACTIVE);
406 * PG_NOTMETA generically means 'don't swapcache this',
407 * and HAMMER will set this for regular data buffers
408 * (and leave it unset for meta-data buffers) as
409 * appropriate when double buffering is enabled.
411 if (m->flags & PG_NOTMETA) {
412 vm_object_drop(object);
414 vm_page_queues_spin_lock(PQ_INACTIVE);
417 if (vm_swapcache_meta_enable == 0) {
418 vm_object_drop(object);
420 vm_page_queues_spin_lock(PQ_INACTIVE);
426 vm_object_drop(object);
428 vm_page_queues_spin_lock(PQ_INACTIVE);
434 * Assign swap and initiate I/O.
436 * (adjust for the --count which also occurs in the loop)
438 count -= vm_swapcached_flush(m, isblkdev) - 1;
441 * Setup for next loop using marker.
443 vm_object_drop(object);
444 vm_page_queues_spin_lock(PQ_INACTIVE);
448 * The marker could wind up at the end, which is ok. If we hit the
449 * end of the list adjust the heuristic.
451 * Earlier inactive pages that were dirty and become clean
452 * are typically moved to the end of PQ_INACTIVE by virtue
453 * of vfs_vmio_release() when they become unwired from the
457 vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
458 vm_page_queues_spin_unlock(PQ_INACTIVE);
462 * Flush the specified page using the swap_pager. The page
463 * must be busied by the caller and its disposition will become
464 * the responsibility of this function.
466 * Try to collect surrounding pages, including pages which may
467 * have already been assigned swap. Try to cluster within a
468 * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
469 * to match what swap_pager_putpages() can do.
471 * We also want to try to match against the buffer cache blocksize
472 * but we don't really know what it is here. Since the buffer cache
473 * wires and unwires pages in groups the fact that we skip wired pages
474 * should be sufficient.
476 * Returns a count of pages we might have flushed (minimum 1)
480 vm_swapcached_flush(vm_page_t m, int isblkdev)
483 vm_page_t marray[SWAP_META_PAGES];
485 int rtvals[SWAP_META_PAGES];
493 vm_page_protect(m, VM_PROT_READ);
495 vm_object_hold(object);
498 * Try to cluster around (m), keeping in mind that the swap pager
499 * can only do SMAP_META_PAGES worth of continguous write.
501 x = (int)m->pindex & SWAP_META_MASK;
506 for (i = x - 1; i >= 0; --i) {
507 m = vm_page_lookup_busy_try(object, basei - x + i,
509 if (error || m == NULL)
511 if (vm_swapcache_test(m)) {
515 if (isblkdev && (m->flags & PG_NOTMETA)) {
520 vm_page_protect(m, VM_PROT_READ);
521 if (m->queue - m->pc == PQ_CACHE) {
522 vm_page_unqueue_nowakeup(m);
523 vm_page_deactivate(m);
530 for (j = x + 1; j < SWAP_META_PAGES; ++j) {
531 m = vm_page_lookup_busy_try(object, basei - x + j,
533 if (error || m == NULL)
535 if (vm_swapcache_test(m)) {
539 if (isblkdev && (m->flags & PG_NOTMETA)) {
544 vm_page_protect(m, VM_PROT_READ);
545 if (m->queue - m->pc == PQ_CACHE) {
546 vm_page_unqueue_nowakeup(m);
547 vm_page_deactivate(m);
554 vm_object_pip_add(object, count);
555 swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
556 vm_swapcache_write_count += count * PAGE_SIZE;
557 vm_swapcache_curburst -= count * PAGE_SIZE;
560 if (rtvals[i] != VM_PAGER_PEND) {
561 vm_page_busy_wait(marray[i], FALSE, "swppgfd");
562 vm_page_io_finish(marray[i]);
563 vm_page_wakeup(marray[i]);
564 vm_object_pip_wakeup(object);
568 vm_object_drop(object);
573 * Test whether a VM page is suitable for writing to the swapcache.
574 * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
576 * Returns 0 on success, 1 on failure
579 vm_swapcache_test(vm_page_t m)
583 if (m->flags & PG_UNMANAGED)
585 if (m->hold_count || m->wire_count)
587 if (m->valid != VM_PAGE_BITS_ALL)
589 if (m->dirty & m->valid)
591 if ((object = m->object) == NULL)
593 if (object->type != OBJT_VNODE ||
594 (object->flags & OBJ_DEAD)) {
597 vm_page_test_dirty(m);
598 if (m->dirty & m->valid)
608 vm_swapcache_cleaning(vm_object_t marker)
616 count = vm_swapcache_maxlaunder;
619 * Look for vnode objects
621 lwkt_gettoken(&vmobj_token);
623 while ((object = TAILQ_NEXT(object, object_list)) != NULL) {
627 vm_object_hold(object);
630 * Only operate on live VNODE objects with regular/chardev types
632 if ((object->type != OBJT_VNODE) ||
633 ((object->flags & OBJ_DEAD) || object->swblock_count == 0) ||
634 ((vp = object->handle) == NULL) ||
635 (vp->v_type != VREG && vp->v_type != VCHR)) {
636 vm_object_drop(object);
643 if (marker->backing_object != object)
647 * Move the marker so we can work on the VM object
649 TAILQ_REMOVE(&vm_object_list, marker, object_list);
650 TAILQ_INSERT_AFTER(&vm_object_list, object,
651 marker, object_list);
654 * Look for swblocks starting at our iterator.
656 * The swap_pager_condfree() function attempts to free
657 * swap space starting at the specified index. The index
658 * will be updated on return. The function will return
659 * a scan factor (NOT the number of blocks freed).
661 * If it must cut its scan of the object short due to an
662 * excessive number of swblocks, or is able to free the
663 * requested number of blocks, it will return n >= count
664 * and we break and pick it back up on a future attempt.
666 n = swap_pager_condfree(object, &marker->size, count);
668 vm_object_drop(object);
682 * Adjust marker so we continue the scan from where we left off.
683 * When we reach the end we start back at the beginning.
685 TAILQ_REMOVE(&vm_object_list, marker, object_list);
687 TAILQ_INSERT_BEFORE(object, marker, object_list);
689 TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
690 marker->backing_object = object;
692 lwkt_reltoken(&vmobj_token);