xref: /freebsd/sys/vm/vm_pagequeue.h (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
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 the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_page.h	8.2 (Berkeley) 12/13/93
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  *
62  * $FreeBSD$
63  */
64 
65 #ifndef	_VM_PAGEQUEUE_
66 #define	_VM_PAGEQUEUE_
67 
68 #ifdef _KERNEL
69 struct vm_pagequeue {
70 	struct mtx	pq_mutex;
71 	struct pglist	pq_pl;
72 	int		pq_cnt;
73 	const char	* const pq_name;
74 	uint64_t	pq_pdpages;
75 } __aligned(CACHE_LINE_SIZE);
76 
77 #if __SIZEOF_LONG__ == 8
78 #define	VM_BATCHQUEUE_SIZE	63
79 #else
80 #define	VM_BATCHQUEUE_SIZE	15
81 #endif
82 
83 struct vm_batchqueue {
84 	vm_page_t	bq_pa[VM_BATCHQUEUE_SIZE];
85 	int		bq_cnt;
86 } __aligned(CACHE_LINE_SIZE);
87 
88 #include <vm/uma.h>
89 #include <sys/_blockcount.h>
90 #include <sys/pidctrl.h>
91 struct sysctl_oid;
92 
93 /*
94  * One vm_domain per NUMA domain.  Contains pagequeues, free page structures,
95  * and accounting.
96  *
97  * Lock Key:
98  * f	vmd_free_mtx
99  * p	vmd_pageout_mtx
100  * d	vm_domainset_lock
101  * a	atomic
102  * c	const after boot
103  * q	page queue lock
104  *
105  * A unique page daemon thread manages each vm_domain structure and is
106  * responsible for ensuring that some free memory is available by freeing
107  * inactive pages and aging active pages.  To decide how many pages to process,
108  * it uses thresholds derived from the number of pages in the domain:
109  *
110  *  vmd_page_count
111  *       ---
112  *        |
113  *        |-> vmd_inactive_target (~3%)
114  *        |   - The active queue scan target is given by
115  *        |     (vmd_inactive_target + vmd_free_target - vmd_free_count).
116  *        |
117  *        |
118  *        |-> vmd_free_target (~2%)
119  *        |   - Target for page reclamation.
120  *        |
121  *        |-> vmd_pageout_wakeup_thresh (~1.8%)
122  *        |   - Threshold for waking up the page daemon.
123  *        |
124  *        |
125  *        |-> vmd_free_min (~0.5%)
126  *        |   - First low memory threshold.
127  *        |   - Causes per-CPU caching to be lazily disabled in UMA.
128  *        |   - vm_wait() sleeps below this threshold.
129  *        |
130  *        |-> vmd_free_severe (~0.25%)
131  *        |   - Second low memory threshold.
132  *        |   - Triggers aggressive UMA reclamation, disables delayed buffer
133  *        |     writes.
134  *        |
135  *        |-> vmd_free_reserved (~0.13%)
136  *        |   - Minimum for VM_ALLOC_NORMAL page allocations.
137  *        |-> vmd_pageout_free_min (32 + 2 pages)
138  *        |   - Minimum for waking a page daemon thread sleeping in vm_wait().
139  *        |-> vmd_interrupt_free_min (2 pages)
140  *        |   - Minimum for VM_ALLOC_SYSTEM page allocations.
141  *       ---
142  *
143  *--
144  * Free page count regulation:
145  *
146  * The page daemon attempts to ensure that the free page count is above the free
147  * target.  It wakes up periodically (every 100ms) to input the current free
148  * page shortage (free_target - free_count) to a PID controller, which in
149  * response outputs the number of pages to attempt to reclaim.  The shortage's
150  * current magnitude, rate of change, and cumulative value are together used to
151  * determine the controller's output.  The page daemon target thus adapts
152  * dynamically to the system's demand for free pages, resulting in less
153  * burstiness than a simple hysteresis loop.
154  *
155  * When the free page count drops below the wakeup threshold,
156  * vm_domain_allocate() proactively wakes up the page daemon.  This helps ensure
157  * that the system responds promptly to a large instantaneous free page
158  * shortage.
159  *
160  * The page daemon also attempts to ensure that some fraction of the system's
161  * memory is present in the inactive (I) and laundry (L) page queues, so that it
162  * can respond promptly to a sudden free page shortage.  In particular, the page
163  * daemon thread aggressively scans active pages so long as the following
164  * condition holds:
165  *
166  *         len(I) + len(L) + free_target - free_count < inactive_target
167  *
168  * Otherwise, when the inactive target is met, the page daemon periodically
169  * scans a small portion of the active queue in order to maintain up-to-date
170  * per-page access history.  Unreferenced pages in the active queue thus
171  * eventually migrate to the inactive queue.
172  *
173  * The per-domain laundry thread periodically launders dirty pages based on the
174  * number of clean pages freed by the page daemon since the last laundering.  If
175  * the page daemon fails to meet its scan target (i.e., the PID controller
176  * output) because of a shortage of clean inactive pages, the laundry thread
177  * attempts to launder enough pages to meet the free page target.
178  *
179  *--
180  * Page allocation priorities:
181  *
182  * The system defines three page allocation priorities: VM_ALLOC_NORMAL,
183  * VM_ALLOC_SYSTEM and VM_ALLOC_INTERRUPT.  An interrupt-priority allocation can
184  * claim any free page.  This priority is used in the pmap layer when attempting
185  * to allocate a page for the kernel page tables; in such cases an allocation
186  * failure will usually result in a kernel panic.  The system priority is used
187  * for most other kernel memory allocations, for instance by UMA's slab
188  * allocator or the buffer cache.  Such allocations will fail if the free count
189  * is below interrupt_free_min.  All other allocations occur at the normal
190  * priority, which is typically used for allocation of user pages, for instance
191  * in the page fault handler or when allocating page table pages or pv_entry
192  * structures for user pmaps.  Such allocations fail if the free count is below
193  * the free_reserved threshold.
194  *
195  *--
196  * Free memory shortages:
197  *
198  * The system uses the free_min and free_severe thresholds to apply
199  * back-pressure and give the page daemon a chance to recover.  When a page
200  * allocation fails due to a shortage and the allocating thread cannot handle
201  * failure, it may call vm_wait() to sleep until free pages are available.
202  * vm_domain_freecnt_inc() wakes sleeping threads once the free page count rises
203  * above the free_min threshold; the page daemon and laundry threads are given
204  * priority and will wake up once free_count reaches the (much smaller)
205  * pageout_free_min threshold.
206  *
207  * On NUMA systems, the domainset iterators always prefer NUMA domains where the
208  * free page count is above the free_min threshold.  This means that given the
209  * choice between two NUMA domains, one above the free_min threshold and one
210  * below, the former will be used to satisfy the allocation request regardless
211  * of the domain selection policy.
212  *
213  * In addition to reclaiming memory from the page queues, the vm_lowmem event
214  * fires every ten seconds so long as the system is under memory pressure (i.e.,
215  * vmd_free_count < vmd_free_target).  This allows kernel subsystems to register
216  * for notifications of free page shortages, upon which they may shrink their
217  * caches.  Following a vm_lowmem event, UMA's caches are pruned to ensure that
218  * they do not contain an excess of unused memory.  When a domain is below the
219  * free_min threshold, UMA limits the population of per-CPU caches.  When a
220  * domain falls below the free_severe threshold, UMA's caches are completely
221  * drained.
222  *
223  * If the system encounters a global memory shortage, it may resort to the
224  * out-of-memory (OOM) killer, which selects a process and delivers SIGKILL in a
225  * last-ditch attempt to free up some pages.  Either of the two following
226  * conditions will activate the OOM killer:
227  *
228  *  1. The page daemons collectively fail to reclaim any pages during their
229  *     inactive queue scans.  After vm_pageout_oom_seq consecutive scans fail,
230  *     the page daemon thread votes for an OOM kill, and an OOM kill is
231  *     triggered when all page daemons have voted.  This heuristic is strict and
232  *     may fail to trigger even when the system is effectively deadlocked.
233  *
234  *  2. Threads in the user fault handler are repeatedly unable to make progress
235  *     while allocating a page to satisfy the fault.  After
236  *     vm_pfault_oom_attempts page allocation failures with intervening
237  *     vm_wait() calls, the faulting thread will trigger an OOM kill.
238  */
239 struct vm_domain {
240 	struct vm_pagequeue vmd_pagequeues[PQ_COUNT];
241 	struct mtx_padalign vmd_free_mtx;
242 	struct mtx_padalign vmd_pageout_mtx;
243 	struct vm_pgcache {
244 		int domain;
245 		int pool;
246 		uma_zone_t zone;
247 	} vmd_pgcache[VM_NFREEPOOL];
248 	struct vmem *vmd_kernel_arena;	/* (c) per-domain kva R/W arena. */
249 	struct vmem *vmd_kernel_rwx_arena; /* (c) per-domain kva R/W/X arena. */
250 	u_int vmd_domain;		/* (c) Domain number. */
251 	u_int vmd_page_count;		/* (c) Total page count. */
252 	long vmd_segs;			/* (c) bitmask of the segments */
253 	u_int __aligned(CACHE_LINE_SIZE) vmd_free_count; /* (a,f) free page count */
254 	u_int vmd_pageout_deficit;	/* (a) Estimated number of pages deficit */
255 	uint8_t vmd_pad[CACHE_LINE_SIZE - (sizeof(u_int) * 2)];
256 
257 	/* Paging control variables, used within single threaded page daemon. */
258 	struct pidctrl vmd_pid;		/* Pageout controller. */
259 	boolean_t vmd_oom;
260 	u_int vmd_inactive_threads;
261 	u_int vmd_inactive_shortage;		/* Per-thread shortage. */
262 	blockcount_t vmd_inactive_running;	/* Number of inactive threads. */
263 	blockcount_t vmd_inactive_starting;	/* Number of threads started. */
264 	volatile u_int vmd_addl_shortage;	/* Shortage accumulator. */
265 	volatile u_int vmd_inactive_freed;	/* Successful inactive frees. */
266 	volatile u_int vmd_inactive_us;		/* Microseconds for above. */
267 	u_int vmd_inactive_pps;		/* Exponential decay frees/second. */
268 	int vmd_oom_seq;
269 	int vmd_last_active_scan;
270 	struct vm_page vmd_markers[PQ_COUNT]; /* (q) markers for queue scans */
271 	struct vm_page vmd_inacthead; /* marker for LRU-defeating insertions */
272 	struct vm_page vmd_clock[2]; /* markers for active queue scan */
273 
274 	int vmd_pageout_wanted;		/* (a, p) pageout daemon wait channel */
275 	int vmd_pageout_pages_needed;	/* (d) page daemon waiting for pages? */
276 	bool vmd_minset;		/* (d) Are we in vm_min_domains? */
277 	bool vmd_severeset;		/* (d) Are we in vm_severe_domains? */
278 	enum {
279 		VM_LAUNDRY_IDLE = 0,
280 		VM_LAUNDRY_BACKGROUND,
281 		VM_LAUNDRY_SHORTFALL
282 	} vmd_laundry_request;
283 
284 	/* Paging thresholds and targets. */
285 	u_int vmd_clean_pages_freed;	/* (q) accumulator for laundry thread */
286 	u_int vmd_background_launder_target; /* (c) */
287 	u_int vmd_free_reserved;	/* (c) pages reserved for deadlock */
288 	u_int vmd_free_target;		/* (c) pages desired free */
289 	u_int vmd_free_min;		/* (c) pages desired free */
290 	u_int vmd_inactive_target;	/* (c) pages desired inactive */
291 	u_int vmd_pageout_free_min;	/* (c) min pages reserved for kernel */
292 	u_int vmd_pageout_wakeup_thresh;/* (c) min pages to wake pagedaemon */
293 	u_int vmd_interrupt_free_min;	/* (c) reserved pages for int code */
294 	u_int vmd_free_severe;		/* (c) severe page depletion point */
295 
296 	/* Name for sysctl etc. */
297 	struct sysctl_oid *vmd_oid;
298 	char vmd_name[sizeof(__XSTRING(MAXMEMDOM))];
299 } __aligned(CACHE_LINE_SIZE);
300 
301 extern struct vm_domain vm_dom[MAXMEMDOM];
302 
303 #define	VM_DOMAIN(n)		(&vm_dom[(n)])
304 #define	VM_DOMAIN_EMPTY(n)	(vm_dom[(n)].vmd_page_count == 0)
305 
306 #define	vm_pagequeue_assert_locked(pq)	mtx_assert(&(pq)->pq_mutex, MA_OWNED)
307 #define	vm_pagequeue_lock(pq)		mtx_lock(&(pq)->pq_mutex)
308 #define	vm_pagequeue_lockptr(pq)	(&(pq)->pq_mutex)
309 #define	vm_pagequeue_trylock(pq)	mtx_trylock(&(pq)->pq_mutex)
310 #define	vm_pagequeue_unlock(pq)		mtx_unlock(&(pq)->pq_mutex)
311 
312 #define	vm_domain_free_assert_locked(n)					\
313 	    mtx_assert(vm_domain_free_lockptr((n)), MA_OWNED)
314 #define	vm_domain_free_assert_unlocked(n)				\
315 	    mtx_assert(vm_domain_free_lockptr((n)), MA_NOTOWNED)
316 #define	vm_domain_free_lock(d)						\
317 	    mtx_lock(vm_domain_free_lockptr((d)))
318 #define	vm_domain_free_lockptr(d)					\
319 	    (&(d)->vmd_free_mtx)
320 #define	vm_domain_free_trylock(d)					\
321 	    mtx_trylock(vm_domain_free_lockptr((d)))
322 #define	vm_domain_free_unlock(d)					\
323 	    mtx_unlock(vm_domain_free_lockptr((d)))
324 
325 #define	vm_domain_pageout_lockptr(d)					\
326 	    (&(d)->vmd_pageout_mtx)
327 #define	vm_domain_pageout_assert_locked(n)				\
328 	    mtx_assert(vm_domain_pageout_lockptr((n)), MA_OWNED)
329 #define	vm_domain_pageout_assert_unlocked(n)				\
330 	    mtx_assert(vm_domain_pageout_lockptr((n)), MA_NOTOWNED)
331 #define	vm_domain_pageout_lock(d)					\
332 	    mtx_lock(vm_domain_pageout_lockptr((d)))
333 #define	vm_domain_pageout_unlock(d)					\
334 	    mtx_unlock(vm_domain_pageout_lockptr((d)))
335 
336 static __inline void
337 vm_pagequeue_cnt_add(struct vm_pagequeue *pq, int addend)
338 {
339 
340 	vm_pagequeue_assert_locked(pq);
341 	pq->pq_cnt += addend;
342 }
343 #define	vm_pagequeue_cnt_inc(pq)	vm_pagequeue_cnt_add((pq), 1)
344 #define	vm_pagequeue_cnt_dec(pq)	vm_pagequeue_cnt_add((pq), -1)
345 
346 static inline void
347 vm_pagequeue_remove(struct vm_pagequeue *pq, vm_page_t m)
348 {
349 
350 	TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
351 	vm_pagequeue_cnt_dec(pq);
352 }
353 
354 static inline void
355 vm_batchqueue_init(struct vm_batchqueue *bq)
356 {
357 
358 	bq->bq_cnt = 0;
359 }
360 
361 static inline int
362 vm_batchqueue_insert(struct vm_batchqueue *bq, vm_page_t m)
363 {
364 	int slots_free;
365 
366 	slots_free = nitems(bq->bq_pa) - bq->bq_cnt;
367 	if (slots_free > 0) {
368 		bq->bq_pa[bq->bq_cnt++] = m;
369 		return (slots_free);
370 	}
371 	return (slots_free);
372 }
373 
374 static inline vm_page_t
375 vm_batchqueue_pop(struct vm_batchqueue *bq)
376 {
377 
378 	if (bq->bq_cnt == 0)
379 		return (NULL);
380 	return (bq->bq_pa[--bq->bq_cnt]);
381 }
382 
383 void vm_domain_set(struct vm_domain *vmd);
384 void vm_domain_clear(struct vm_domain *vmd);
385 int vm_domain_allocate(struct vm_domain *vmd, int req, int npages);
386 
387 /*
388  *      vm_pagequeue_domain:
389  *
390  *      Return the memory domain the page belongs to.
391  */
392 static inline struct vm_domain *
393 vm_pagequeue_domain(vm_page_t m)
394 {
395 
396 	return (VM_DOMAIN(vm_page_domain(m)));
397 }
398 
399 /*
400  * Return the number of pages we need to free-up or cache
401  * A positive number indicates that we do not have enough free pages.
402  */
403 static inline int
404 vm_paging_target(struct vm_domain *vmd)
405 {
406 
407 	return (vmd->vmd_free_target - vmd->vmd_free_count);
408 }
409 
410 /*
411  * Returns TRUE if the pagedaemon needs to be woken up.
412  */
413 static inline int
414 vm_paging_needed(struct vm_domain *vmd, u_int free_count)
415 {
416 
417 	return (free_count < vmd->vmd_pageout_wakeup_thresh);
418 }
419 
420 /*
421  * Returns TRUE if the domain is below the min paging target.
422  */
423 static inline int
424 vm_paging_min(struct vm_domain *vmd)
425 {
426 
427         return (vmd->vmd_free_min > vmd->vmd_free_count);
428 }
429 
430 /*
431  * Returns TRUE if the domain is below the severe paging target.
432  */
433 static inline int
434 vm_paging_severe(struct vm_domain *vmd)
435 {
436 
437         return (vmd->vmd_free_severe > vmd->vmd_free_count);
438 }
439 
440 /*
441  * Return the number of pages we need to launder.
442  * A positive number indicates that we have a shortfall of clean pages.
443  */
444 static inline int
445 vm_laundry_target(struct vm_domain *vmd)
446 {
447 
448 	return (vm_paging_target(vmd));
449 }
450 
451 void pagedaemon_wakeup(int domain);
452 
453 static inline void
454 vm_domain_freecnt_inc(struct vm_domain *vmd, int adj)
455 {
456 	u_int old, new;
457 
458 	old = atomic_fetchadd_int(&vmd->vmd_free_count, adj);
459 	new = old + adj;
460 	/*
461 	 * Only update bitsets on transitions.  Notice we short-circuit the
462 	 * rest of the checks if we're above min already.
463 	 */
464 	if (old < vmd->vmd_free_min && (new >= vmd->vmd_free_min ||
465 	    (old < vmd->vmd_free_severe && new >= vmd->vmd_free_severe) ||
466 	    (old < vmd->vmd_pageout_free_min &&
467 	    new >= vmd->vmd_pageout_free_min)))
468 		vm_domain_clear(vmd);
469 }
470 
471 #endif	/* _KERNEL */
472 #endif				/* !_VM_PAGEQUEUE_ */
473