xref: /dflybsd-src/sys/vm/vm_swapcache.c (revision d83c779ab2c938232fa7b53777cd18cc9c4fc8e4)
1 /*
2  * Copyright (c) 2010 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
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
16  *    distribution.
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.
20  *
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
32  * SUCH DAMAGE.
33  */
34 
35 /*
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.
41  *
42  * The daemon manages write bandwidth based on sysctl settings to control
43  * wear on the SSD.
44  *
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.
47  *
48  * This operates on both regular files and the block device vnodes used by
49  * filesystems to manage meta-data.
50  */
51 
52 #include "opt_vm.h"
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
56 #include <sys/proc.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>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_param.h>
66 #include <sys/lock.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>
74 
75 #include <sys/thread2.h>
76 #include <vm/vm_page2.h>
77 
78 #define INACTIVE_LIST	(&vm_page_queues[PQ_INACTIVE].pl)
79 
80 /* the kernel process "vm_pageout"*/
81 static void vm_swapcached (void);
82 static int vm_swapcached_flush (vm_page_t m, int isblkdev);
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;
87 
88 static struct kproc_desc swpc_kp = {
89 	"swapcached",
90 	vm_swapcached,
91 	&swapcached_thread
92 };
93 SYSINIT(swapcached, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start, &swpc_kp)
94 
95 SYSCTL_NODE(_vm, OID_AUTO, swapcache, CTLFLAG_RW, NULL, NULL);
96 
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;
112 
113 SYSCTL_INT(_vm_swapcache, OID_AUTO, maxlaunder,
114 	CTLFLAG_RW, &vm_swapcache_maxlaunder, 0, "");
115 
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, "");
128 
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, "");
141 
142 #define SWAPMAX(adj)	\
143 	((int64_t)vm_swap_max * (vm_swapcache_maxswappct + (adj)) / 100)
144 
145 /*
146  * vm_swapcached is the high level pageout daemon.
147  */
148 static void
149 vm_swapcached(void)
150 {
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;
155 
156 	/*
157 	 * Thread setup
158 	 */
159 	curthread->td_flags |= TDF_SYSTHREAD;
160 	crit_enter();
161 
162 	/*
163 	 * Initialize our marker for the inactive scan (SWAPC_WRITING)
164 	 */
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;
172 
173 	/*
174 	 * Initialize our marker for the vm_object scan (SWAPC_CLEANING)
175 	 */
176 	bzero(&object_marker, sizeof(object_marker));
177 	object_marker.type = OBJT_MARKER;
178 	TAILQ_INSERT_HEAD(&vm_object_list, &object_marker, object_list);
179 
180 	for (;;) {
181 		/*
182 		 * Check every 5 seconds when not enabled or if no swap
183 		 * is present.
184 		 */
185 		if ((vm_swapcache_data_enable == 0 &&
186 		     vm_swapcache_meta_enable == 0) ||
187 		    vm_swap_max == 0) {
188 			tsleep(&vm_swapcache_sleep, 0, "csleep", hz * 5);
189 			continue;
190 		}
191 
192 		/*
193 		 * Polling rate when enabled is approximately 10 hz.
194 		 */
195 		tsleep(&vm_swapcache_sleep, 0, "csleep", hz / 10);
196 
197 		/*
198 		 * State hysteresis.  Generate write activity up to 75% of
199 		 * swap, then clean out swap assignments down to 70%, then
200 		 * repeat.
201 		 */
202 		if (state == SWAPC_WRITING) {
203 			if (vm_swap_cache_use > SWAPMAX(0))
204 				state = SWAPC_CLEANING;
205 		} else {
206 			if (vm_swap_cache_use < SWAPMAX(-5))
207 				state = SWAPC_WRITING;
208 		}
209 
210 		/*
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.
214 		 */
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;
219 		}
220 
221 		/*
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.
225 		 */
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;
236 				}
237 			}
238 		} else {
239 			vm_swapcache_cleaning(&object_marker);
240 		}
241 	}
242 	TAILQ_REMOVE(INACTIVE_LIST, &page_marker, pageq);
243 	TAILQ_REMOVE(&vm_object_list, &object_marker, object_list);
244 	crit_exit();
245 }
246 
247 static void
248 vm_swapcache_writing(vm_page_t marker)
249 {
250 	vm_object_t object;
251 	struct vnode *vp;
252 	vm_page_t m;
253 	int count;
254 	int isblkdev;
255 
256 	/*
257 	 * Deal with an overflow of the heuristic counter or if the user
258 	 * manually changes the hysteresis.
259 	 *
260 	 * Try to avoid small incremental pageouts by waiting for enough
261 	 * pages to buildup in the inactive queue to hopefully get a good
262 	 * burst in.  This heuristic is bumped by the VM system and reset
263 	 * when our scan hits the end of the queue.
264 	 */
265 	if (vm_swapcache_inactive_heuristic < -vm_swapcache_hysteresis)
266 		vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
267 	if (vm_swapcache_inactive_heuristic < 0)
268 		return;
269 
270 	/*
271 	 * Scan the inactive queue from our marker to locate
272 	 * suitable pages to push to the swap cache.
273 	 *
274 	 * We are looking for clean vnode-backed pages.
275 	 *
276 	 * NOTE: PG_SWAPPED pages in particular are not part of
277 	 *	 our count because once the cache stabilizes we
278 	 *	 can end up with a very high datarate of VM pages
279 	 *	 cycling from it.
280 	 */
281 	m = marker;
282 	count = vm_swapcache_maxlaunder;
283 
284 	while ((m = TAILQ_NEXT(m, pageq)) != NULL && count--) {
285 		if (m->flags & (PG_MARKER | PG_SWAPPED)) {
286 			++count;
287 			continue;
288 		}
289 		if (vm_swapcache_curburst < 0)
290 			break;
291 		if (vm_swapcache_test(m))
292 			continue;
293 		object = m->object;
294 		vp = object->handle;
295 		if (vp == NULL)
296 			continue;
297 
298 		switch(vp->v_type) {
299 		case VREG:
300 			/*
301 			 * If data_enable is 0 do not try to swapcache data.
302 			 * If use_chflags is set then only swapcache data for
303 			 * VSWAPCACHE marked vnodes, otherwise any vnode.
304 			 */
305 			if (vm_swapcache_data_enable == 0 ||
306 			    ((vp->v_flag & VSWAPCACHE) == 0 &&
307 			     vm_swapcache_use_chflags)) {
308 				continue;
309 			}
310 			if (vm_swapcache_maxfilesize &&
311 			    object->size >
312 			    (vm_swapcache_maxfilesize >> PAGE_SHIFT)) {
313 				continue;
314 			}
315 			isblkdev = 0;
316 			break;
317 		case VCHR:
318 			/*
319 			 * The PG_NOTMETA flag only applies to pages
320 			 * associated with block devices.
321 			 */
322 			if (m->flags & PG_NOTMETA)
323 				continue;
324 			if (vm_swapcache_meta_enable == 0)
325 				continue;
326 			isblkdev = 1;
327 			break;
328 		default:
329 			continue;
330 		}
331 
332 		/*
333 		 * Ok, move the marker and soft-busy the page.
334 		 */
335 		TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
336 		TAILQ_INSERT_AFTER(INACTIVE_LIST, m, marker, pageq);
337 
338 		/*
339 		 * Assign swap and initiate I/O.
340 		 *
341 		 * (adjust for the --count which also occurs in the loop)
342 		 */
343 		count -= vm_swapcached_flush(m, isblkdev) - 1;
344 
345 		/*
346 		 * Setup for next loop using marker.
347 		 */
348 		m = marker;
349 	}
350 
351 	/*
352 	 * Cleanup marker position.  If we hit the end of the
353 	 * list the marker is placed at the tail.  Newly deactivated
354 	 * pages will be placed after it.
355 	 *
356 	 * Earlier inactive pages that were dirty and become clean
357 	 * are typically moved to the end of PQ_INACTIVE by virtue
358 	 * of vfs_vmio_release() when they become unwired from the
359 	 * buffer cache.
360 	 */
361 	TAILQ_REMOVE(INACTIVE_LIST, marker, pageq);
362 	if (m) {
363 		TAILQ_INSERT_BEFORE(m, marker, pageq);
364 	} else {
365 		TAILQ_INSERT_TAIL(INACTIVE_LIST, marker, pageq);
366 		vm_swapcache_inactive_heuristic = -vm_swapcache_hysteresis;
367 	}
368 }
369 
370 /*
371  * Flush the specified page using the swap_pager.
372  *
373  * Try to collect surrounding pages, including pages which may
374  * have already been assigned swap.  Try to cluster within a
375  * contiguous aligned SMAP_META_PAGES (typ 16 x PAGE_SIZE) block
376  * to match what swap_pager_putpages() can do.
377  *
378  * We also want to try to match against the buffer cache blocksize
379  * but we don't really know what it is here.  Since the buffer cache
380  * wires and unwires pages in groups the fact that we skip wired pages
381  * should be sufficient.
382  *
383  * Returns a count of pages we might have flushed (minimum 1)
384  */
385 static
386 int
387 vm_swapcached_flush(vm_page_t m, int isblkdev)
388 {
389 	vm_object_t object;
390 	vm_page_t marray[SWAP_META_PAGES];
391 	vm_pindex_t basei;
392 	int rtvals[SWAP_META_PAGES];
393 	int x;
394 	int i;
395 	int j;
396 	int count;
397 
398 	vm_page_io_start(m);
399 	vm_page_protect(m, VM_PROT_READ);
400 	object = m->object;
401 
402 	/*
403 	 * Try to cluster around (m), keeping in mind that the swap pager
404 	 * can only do SMAP_META_PAGES worth of continguous write.
405 	 */
406 	x = (int)m->pindex & SWAP_META_MASK;
407 	marray[x] = m;
408 	basei = m->pindex;
409 
410 	for (i = x - 1; i >= 0; --i) {
411 		m = vm_page_lookup(object, basei - x + i);
412 		if (m == NULL)
413 			break;
414 		if (vm_swapcache_test(m))
415 			break;
416 		if (isblkdev && (m->flags & PG_NOTMETA))
417 			break;
418 		vm_page_io_start(m);
419 		vm_page_protect(m, VM_PROT_READ);
420 		if (m->queue - m->pc == PQ_CACHE) {
421 			vm_page_unqueue_nowakeup(m);
422 			vm_page_deactivate(m);
423 		}
424 		marray[i] = m;
425 	}
426 	++i;
427 
428 	for (j = x + 1; j < SWAP_META_PAGES; ++j) {
429 		m = vm_page_lookup(object, basei - x + j);
430 		if (m == NULL)
431 			break;
432 		if (vm_swapcache_test(m))
433 			break;
434 		if (isblkdev && (m->flags & PG_NOTMETA))
435 			break;
436 		vm_page_io_start(m);
437 		vm_page_protect(m, VM_PROT_READ);
438 		if (m->queue - m->pc == PQ_CACHE) {
439 			vm_page_unqueue_nowakeup(m);
440 			vm_page_deactivate(m);
441 		}
442 		marray[j] = m;
443 	}
444 
445 	count = j - i;
446 	vm_object_pip_add(object, count);
447 	swap_pager_putpages(object, marray + i, count, FALSE, rtvals + i);
448 	vm_swapcache_write_count += count * PAGE_SIZE;
449 	vm_swapcache_curburst -= count * PAGE_SIZE;
450 
451 	while (i < j) {
452 		if (rtvals[i] != VM_PAGER_PEND) {
453 			vm_page_io_finish(marray[i]);
454 			vm_object_pip_wakeup(object);
455 		}
456 		++i;
457 	}
458 	return(count);
459 }
460 
461 /*
462  * Test whether a VM page is suitable for writing to the swapcache.
463  * Does not test m->queue, PG_MARKER, or PG_SWAPPED.
464  *
465  * Returns 0 on success, 1 on failure
466  */
467 static int
468 vm_swapcache_test(vm_page_t m)
469 {
470 	vm_object_t object;
471 
472 	if (m->flags & (PG_BUSY | PG_UNMANAGED))
473 		return(1);
474 	if (m->busy || m->hold_count || m->wire_count)
475 		return(1);
476 	if (m->valid != VM_PAGE_BITS_ALL)
477 		return(1);
478 	if (m->dirty & m->valid)
479 		return(1);
480 	if ((object = m->object) == NULL)
481 		return(1);
482 	if (object->type != OBJT_VNODE ||
483 	    (object->flags & OBJ_DEAD)) {
484 		return(1);
485 	}
486 	vm_page_test_dirty(m);
487 	if (m->dirty & m->valid)
488 		return(1);
489 	return(0);
490 }
491 
492 /*
493  * Cleaning pass
494  */
495 static
496 void
497 vm_swapcache_cleaning(vm_object_t marker)
498 {
499 	vm_object_t object;
500 	struct vnode *vp;
501 	int count;
502 	int n;
503 
504 	object = marker;
505 	count = vm_swapcache_maxlaunder;
506 
507 	/*
508 	 * Look for vnode objects
509 	 */
510 	while ((object = TAILQ_NEXT(object, object_list)) != NULL && count--) {
511 		if (object->type != OBJT_VNODE)
512 			continue;
513 		if ((object->flags & OBJ_DEAD) || object->swblock_count == 0)
514 			continue;
515 		if ((vp = object->handle) == NULL)
516 			continue;
517 		if (vp->v_type != VREG && vp->v_type != VCHR)
518 			continue;
519 
520 		/*
521 		 * Adjust iterator.
522 		 */
523 		if (marker->backing_object != object)
524 			marker->size = 0;
525 
526 		/*
527 		 * Move the marker so we can work on the VM object
528 		 */
529 		TAILQ_REMOVE(&vm_object_list, marker, object_list);
530 		TAILQ_INSERT_AFTER(&vm_object_list, object,
531 				   marker, object_list);
532 
533 		/*
534 		 * Look for swblocks starting at our iterator.
535 		 *
536 		 * The swap_pager_condfree() function attempts to free
537 		 * swap space starting at the specified index.  The index
538 		 * will be updated on return.  The function will return
539 		 * a scan factor (NOT the number of blocks freed).
540 		 *
541 		 * If it must cut its scan of the object short due to an
542 		 * excessive number of swblocks, or is able to free the
543 		 * requested number of blocks, it will return n >= count
544 		 * and we break and pick it back up on a future attempt.
545 		 */
546 		n = swap_pager_condfree(object, &marker->size, count);
547 		count -= n;
548 		if (count < 0)
549 			break;
550 
551 		/*
552 		 * Setup for loop.
553 		 */
554 		marker->size = 0;
555 		object = marker;
556 	}
557 
558 	/*
559 	 * Adjust marker so we continue the scan from where we left off.
560 	 * When we reach the end we start back at the beginning.
561 	 */
562 	TAILQ_REMOVE(&vm_object_list, marker, object_list);
563 	if (object)
564 		TAILQ_INSERT_BEFORE(object, marker, object_list);
565 	else
566 		TAILQ_INSERT_HEAD(&vm_object_list, marker, object_list);
567 	marker->backing_object = object;
568 }
569