xref: /dflybsd-src/sys/vm/vm_page.c (revision 8d1ea4cc63e8cedbe8f670822d4606adf2cf7d45)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 2003-2011 The DragonFly Project.  All rights reserved.
5  *
6  * This code is derived from software contributed to Berkeley by
7  * The Mach Operating System project at Carnegie-Mellon University.
8  *
9  * This code is derived from software contributed to The DragonFly Project
10  * by Matthew Dillon <dillon@backplane.com>
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_page.c	7.4 (Berkeley) 5/7/91
37  * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
38  */
39 
40 /*
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  */
66 /*
67  * Resident memory management module.  The module manipulates 'VM pages'.
68  * A VM page is the core building block for memory management.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/malloc.h>
74 #include <sys/proc.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 #include <sys/kernel.h>
78 #include <sys/alist.h>
79 #include <sys/sysctl.h>
80 #include <sys/cpu_topology.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_param.h>
84 #include <sys/lock.h>
85 #include <vm/vm_kern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
94 
95 #include <machine/inttypes.h>
96 #include <machine/md_var.h>
97 #include <machine/specialreg.h>
98 #include <machine/bus_dma.h>
99 
100 #include <vm/vm_page2.h>
101 #include <sys/spinlock2.h>
102 
103 /*
104  * SET - Minimum required set associative size, must be a power of 2.  We
105  *	 want this to match or exceed the set-associativeness of the cpu.
106  *
107  * GRP - A larger set that allows bleed-over into the domains of other
108  *	 nearby cpus.  Also must be a power of 2.  Used by the page zeroing
109  *	 code to smooth things out a bit.
110  */
111 #define PQ_SET_ASSOC		16
112 #define PQ_SET_ASSOC_MASK	(PQ_SET_ASSOC - 1)
113 
114 #define PQ_GRP_ASSOC		(PQ_SET_ASSOC * 2)
115 #define PQ_GRP_ASSOC_MASK	(PQ_GRP_ASSOC - 1)
116 
117 static void vm_page_queue_init(void);
118 static void vm_page_free_wakeup(void);
119 static vm_page_t vm_page_select_cache(u_short pg_color);
120 static vm_page_t _vm_page_list_find2(int basequeue, int index);
121 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
122 static void vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes);
123 
124 /*
125  * Array of tailq lists
126  */
127 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
128 
129 static volatile int vm_pages_waiting;
130 static struct alist vm_contig_alist;
131 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
132 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
133 
134 static u_long vm_dma_reserved = 0;
135 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
136 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
137 	    "Memory reserved for DMA");
138 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
139 	    &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
140 
141 static int vm_contig_verbose = 0;
142 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
143 
144 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
145 	     vm_pindex_t, pindex);
146 
147 static void
148 vm_page_queue_init(void)
149 {
150 	int i;
151 
152 	for (i = 0; i < PQ_L2_SIZE; i++)
153 		vm_page_queues[PQ_FREE+i].cnt_offset =
154 			offsetof(struct vmstats, v_free_count);
155 	for (i = 0; i < PQ_L2_SIZE; i++)
156 		vm_page_queues[PQ_CACHE+i].cnt_offset =
157 			offsetof(struct vmstats, v_cache_count);
158 	for (i = 0; i < PQ_L2_SIZE; i++)
159 		vm_page_queues[PQ_INACTIVE+i].cnt_offset =
160 			offsetof(struct vmstats, v_inactive_count);
161 	for (i = 0; i < PQ_L2_SIZE; i++)
162 		vm_page_queues[PQ_ACTIVE+i].cnt_offset =
163 			offsetof(struct vmstats, v_active_count);
164 	for (i = 0; i < PQ_L2_SIZE; i++)
165 		vm_page_queues[PQ_HOLD+i].cnt_offset =
166 			offsetof(struct vmstats, v_active_count);
167 	/* PQ_NONE has no queue */
168 
169 	for (i = 0; i < PQ_COUNT; i++) {
170 		TAILQ_INIT(&vm_page_queues[i].pl);
171 		spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
172 	}
173 }
174 
175 /*
176  * note: place in initialized data section?  Is this necessary?
177  */
178 vm_pindex_t first_page = 0;
179 vm_pindex_t vm_page_array_size = 0;
180 vm_page_t vm_page_array = NULL;
181 vm_paddr_t vm_low_phys_reserved;
182 
183 /*
184  * (low level boot)
185  *
186  * Sets the page size, perhaps based upon the memory size.
187  * Must be called before any use of page-size dependent functions.
188  */
189 void
190 vm_set_page_size(void)
191 {
192 	if (vmstats.v_page_size == 0)
193 		vmstats.v_page_size = PAGE_SIZE;
194 	if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
195 		panic("vm_set_page_size: page size not a power of two");
196 }
197 
198 /*
199  * (low level boot)
200  *
201  * Add a new page to the freelist for use by the system.  New pages
202  * are added to both the head and tail of the associated free page
203  * queue in a bottom-up fashion, so both zero'd and non-zero'd page
204  * requests pull 'recent' adds (higher physical addresses) first.
205  *
206  * Beware that the page zeroing daemon will also be running soon after
207  * boot, moving pages from the head to the tail of the PQ_FREE queues.
208  *
209  * Must be called in a critical section.
210  */
211 static void
212 vm_add_new_page(vm_paddr_t pa)
213 {
214 	struct vpgqueues *vpq;
215 	vm_page_t m;
216 
217 	m = PHYS_TO_VM_PAGE(pa);
218 	m->phys_addr = pa;
219 	m->flags = 0;
220 	m->pat_mode = PAT_WRITE_BACK;
221 	m->pc = (pa >> PAGE_SHIFT);
222 
223 	/*
224 	 * Twist for cpu localization in addition to page coloring, so
225 	 * different cpus selecting by m->queue get different page colors.
226 	 */
227 	m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE);
228 	m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE));
229 	m->pc &= PQ_L2_MASK;
230 
231 	/*
232 	 * Reserve a certain number of contiguous low memory pages for
233 	 * contigmalloc() to use.
234 	 */
235 	if (pa < vm_low_phys_reserved) {
236 		atomic_add_long(&vmstats.v_page_count, 1);
237 		atomic_add_long(&vmstats.v_dma_pages, 1);
238 		m->queue = PQ_NONE;
239 		m->wire_count = 1;
240 		atomic_add_long(&vmstats.v_wire_count, 1);
241 		alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
242 		return;
243 	}
244 
245 	/*
246 	 * General page
247 	 */
248 	m->queue = m->pc + PQ_FREE;
249 	KKASSERT(m->dirty == 0);
250 
251 	atomic_add_long(&vmstats.v_page_count, 1);
252 	atomic_add_long(&vmstats.v_free_count, 1);
253 	vpq = &vm_page_queues[m->queue];
254 	TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
255 	++vpq->lcnt;
256 }
257 
258 /*
259  * (low level boot)
260  *
261  * Initializes the resident memory module.
262  *
263  * Preallocates memory for critical VM structures and arrays prior to
264  * kernel_map becoming available.
265  *
266  * Memory is allocated from (virtual2_start, virtual2_end) if available,
267  * otherwise memory is allocated from (virtual_start, virtual_end).
268  *
269  * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
270  * large enough to hold vm_page_array & other structures for machines with
271  * large amounts of ram, so we want to use virtual2* when available.
272  */
273 void
274 vm_page_startup(void)
275 {
276 	vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
277 	vm_offset_t mapped;
278 	vm_pindex_t npages;
279 	vm_paddr_t page_range;
280 	vm_paddr_t new_end;
281 	int i;
282 	vm_paddr_t pa;
283 	vm_paddr_t last_pa;
284 	vm_paddr_t end;
285 	vm_paddr_t biggestone, biggestsize;
286 	vm_paddr_t total;
287 	vm_page_t m;
288 
289 	total = 0;
290 	biggestsize = 0;
291 	biggestone = 0;
292 	vaddr = round_page(vaddr);
293 
294 	/*
295 	 * Make sure ranges are page-aligned.
296 	 */
297 	for (i = 0; phys_avail[i].phys_end; ++i) {
298 		phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
299 		phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
300 		if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
301 			phys_avail[i].phys_end = phys_avail[i].phys_beg;
302 	}
303 
304 	/*
305 	 * Locate largest block
306 	 */
307 	for (i = 0; phys_avail[i].phys_end; ++i) {
308 		vm_paddr_t size = phys_avail[i].phys_end -
309 				  phys_avail[i].phys_beg;
310 
311 		if (size > biggestsize) {
312 			biggestone = i;
313 			biggestsize = size;
314 		}
315 		total += size;
316 	}
317 	--i;	/* adjust to last entry for use down below */
318 
319 	end = phys_avail[biggestone].phys_end;
320 	end = trunc_page(end);
321 
322 	/*
323 	 * Initialize the queue headers for the free queue, the active queue
324 	 * and the inactive queue.
325 	 */
326 	vm_page_queue_init();
327 
328 #if !defined(_KERNEL_VIRTUAL)
329 	/*
330 	 * VKERNELs don't support minidumps and as such don't need
331 	 * vm_page_dump
332 	 *
333 	 * Allocate a bitmap to indicate that a random physical page
334 	 * needs to be included in a minidump.
335 	 *
336 	 * The amd64 port needs this to indicate which direct map pages
337 	 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
338 	 *
339 	 * However, x86 still needs this workspace internally within the
340 	 * minidump code.  In theory, they are not needed on x86, but are
341 	 * included should the sf_buf code decide to use them.
342 	 */
343 	page_range = phys_avail[i].phys_end / PAGE_SIZE;
344 	vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
345 	end -= vm_page_dump_size;
346 	vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
347 					VM_PROT_READ | VM_PROT_WRITE);
348 	bzero((void *)vm_page_dump, vm_page_dump_size);
349 #endif
350 	/*
351 	 * Compute the number of pages of memory that will be available for
352 	 * use (taking into account the overhead of a page structure per
353 	 * page).
354 	 */
355 	first_page = phys_avail[0].phys_beg / PAGE_SIZE;
356 	page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page;
357 	npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
358 
359 #ifndef _KERNEL_VIRTUAL
360 	/*
361 	 * (only applies to real kernels)
362 	 *
363 	 * Reserve a large amount of low memory for potential 32-bit DMA
364 	 * space allocations.  Once device initialization is complete we
365 	 * release most of it, but keep (vm_dma_reserved) memory reserved
366 	 * for later use.  Typically for X / graphics.  Through trial and
367 	 * error we find that GPUs usually requires ~60-100MB or so.
368 	 *
369 	 * By default, 128M is left in reserve on machines with 2G+ of ram.
370 	 */
371 	vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
372 	if (vm_low_phys_reserved > total / 4)
373 		vm_low_phys_reserved = total / 4;
374 	if (vm_dma_reserved == 0) {
375 		vm_dma_reserved = 128 * 1024 * 1024;	/* 128MB */
376 		if (vm_dma_reserved > total / 16)
377 			vm_dma_reserved = total / 16;
378 	}
379 #endif
380 	alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
381 		   ALIST_RECORDS_65536);
382 
383 	/*
384 	 * Initialize the mem entry structures now, and put them in the free
385 	 * queue.
386 	 */
387 	if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
388 		kprintf("initializing vm_page_array ");
389 	new_end = trunc_page(end - page_range * sizeof(struct vm_page));
390 	mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
391 	vm_page_array = (vm_page_t)mapped;
392 
393 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
394 	/*
395 	 * since pmap_map on amd64 returns stuff out of a direct-map region,
396 	 * we have to manually add these pages to the minidump tracking so
397 	 * that they can be dumped, including the vm_page_array.
398 	 */
399 	for (pa = new_end;
400 	     pa < phys_avail[biggestone].phys_end;
401 	     pa += PAGE_SIZE) {
402 		dump_add_page(pa);
403 	}
404 #endif
405 
406 	/*
407 	 * Clear all of the page structures, run basic initialization so
408 	 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
409 	 * map.
410 	 */
411 	bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
412 	vm_page_array_size = page_range;
413 	if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
414 		kprintf("size = 0x%zx\n", vm_page_array_size);
415 
416 	m = &vm_page_array[0];
417 	pa = ptoa(first_page);
418 	for (i = 0; i < page_range; ++i) {
419 		spin_init(&m->spin, "vm_page");
420 		m->phys_addr = pa;
421 		pa += PAGE_SIZE;
422 		++m;
423 	}
424 
425 	/*
426 	 * Construct the free queue(s) in ascending order (by physical
427 	 * address) so that the first 16MB of physical memory is allocated
428 	 * last rather than first.  On large-memory machines, this avoids
429 	 * the exhaustion of low physical memory before isa_dma_init has run.
430 	 */
431 	vmstats.v_page_count = 0;
432 	vmstats.v_free_count = 0;
433 	for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
434 		pa = phys_avail[i].phys_beg;
435 		if (i == biggestone)
436 			last_pa = new_end;
437 		else
438 			last_pa = phys_avail[i].phys_end;
439 		while (pa < last_pa && npages-- > 0) {
440 			vm_add_new_page(pa);
441 			pa += PAGE_SIZE;
442 		}
443 	}
444 	if (virtual2_start)
445 		virtual2_start = vaddr;
446 	else
447 		virtual_start = vaddr;
448 	mycpu->gd_vmstats = vmstats;
449 }
450 
451 /*
452  * (called from early boot only)
453  *
454  * Reorganize VM pages based on numa data.  May be called as many times as
455  * necessary.  Will reorganize the vm_page_t page color and related queue(s)
456  * to allow vm_page_alloc() to choose pages based on socket affinity.
457  *
458  * NOTE: This function is only called while we are still in UP mode, so
459  *	 we only need a critical section to protect the queues (which
460  *	 saves a lot of time, there are likely a ton of pages).
461  */
462 void
463 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
464 {
465 	vm_paddr_t scan_beg;
466 	vm_paddr_t scan_end;
467 	vm_paddr_t ran_end;
468 	struct vpgqueues *vpq;
469 	vm_page_t m;
470 	vm_page_t mend;
471 	int socket_mod;
472 	int socket_value;
473 	int i;
474 
475 	/*
476 	 * Check if no physical information, or there was only one socket
477 	 * (so don't waste time doing nothing!).
478 	 */
479 	if (cpu_topology_phys_ids <= 1 ||
480 	    cpu_topology_core_ids == 0) {
481 		return;
482 	}
483 
484 	/*
485 	 * Setup for our iteration.  Note that ACPI may iterate CPU
486 	 * sockets starting at 0 or 1 or some other number.  The
487 	 * cpu_topology code mod's it against the socket count.
488 	 */
489 	ran_end = ran_beg + bytes;
490 
491 	socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
492 	socket_value = (physid % cpu_topology_phys_ids) * socket_mod;
493 	mend = &vm_page_array[vm_page_array_size];
494 
495 	crit_enter();
496 
497 	/*
498 	 * Adjust cpu_topology's phys_mem parameter
499 	 */
500 	if (root_cpu_node)
501 		vm_numa_add_topology_mem(root_cpu_node, physid, (long)bytes);
502 
503 	/*
504 	 * Adjust vm_page->pc and requeue all affected pages.  The
505 	 * allocator will then be able to localize memory allocations
506 	 * to some degree.
507 	 */
508 	for (i = 0; phys_avail[i].phys_end; ++i) {
509 		scan_beg = phys_avail[i].phys_beg;
510 		scan_end = phys_avail[i].phys_end;
511 		if (scan_end <= ran_beg)
512 			continue;
513 		if (scan_beg >= ran_end)
514 			continue;
515 		if (scan_beg < ran_beg)
516 			scan_beg = ran_beg;
517 		if (scan_end > ran_end)
518 			scan_end = ran_end;
519 		if (atop(scan_end) > first_page + vm_page_array_size)
520 			scan_end = ptoa(first_page + vm_page_array_size);
521 
522 		m = PHYS_TO_VM_PAGE(scan_beg);
523 		while (scan_beg < scan_end) {
524 			KKASSERT(m < mend);
525 			if (m->queue != PQ_NONE) {
526 				vpq = &vm_page_queues[m->queue];
527 				TAILQ_REMOVE(&vpq->pl, m, pageq);
528 				--vpq->lcnt;
529 				/* queue doesn't change, no need to adj cnt */
530 				m->queue -= m->pc;
531 				m->pc %= socket_mod;
532 				m->pc += socket_value;
533 				m->pc &= PQ_L2_MASK;
534 				m->queue += m->pc;
535 				vpq = &vm_page_queues[m->queue];
536 				TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
537 				++vpq->lcnt;
538 				/* queue doesn't change, no need to adj cnt */
539 			} else {
540 				m->pc %= socket_mod;
541 				m->pc += socket_value;
542 				m->pc &= PQ_L2_MASK;
543 			}
544 			scan_beg += PAGE_SIZE;
545 			++m;
546 		}
547 	}
548 
549 	crit_exit();
550 }
551 
552 /*
553  * (called from early boot only)
554  *
555  * Don't allow the NUMA organization to leave vm_page_queues[] nodes
556  * completely empty for a logical cpu.  Doing so would force allocations
557  * on that cpu to always borrow from a nearby cpu, create unnecessary
558  * contention, and cause vm_page_alloc() to iterate more queues and run more
559  * slowly.
560  *
561  * This situation can occur when memory sticks are not entirely populated,
562  * populated at different densities, or in naturally assymetric systems
563  * such as the 2990WX.  There could very well be many vm_page_queues[]
564  * entries with *NO* pages assigned to them.
565  *
566  * Fixing this up ensures that each logical CPU has roughly the same
567  * sized memory pool, and more importantly ensures that logical CPUs
568  * do not wind up with an empty memory pool.
569  *
570  * At them moment we just iterate the other queues and borrow pages,
571  * moving them into the queues for cpus with severe deficits even though
572  * the memory might not be local to those cpus.  I am not doing this in
573  * a 'smart' way, its effectively UMA style (sorta, since its page-by-page
574  * whereas real UMA typically exchanges address bits 8-10 with high address
575  * bits).  But it works extremely well and gives us fairly good deterministic
576  * results on the cpu cores associated with these secondary nodes.
577  */
578 void
579 vm_numa_organize_finalize(void)
580 {
581 	struct vpgqueues *vpq;
582 	vm_page_t m;
583 	long lcnt_lo;
584 	long lcnt_hi;
585 	int iter;
586 	int i;
587 	int scale_lim;
588 
589 	crit_enter();
590 
591 	/*
592 	 * Machines might not use an exact power of 2 for phys_ids,
593 	 * core_ids, ht_ids, etc.  This can slightly reduce the actual
594 	 * range of indices in vm_page_queues[] that are nominally used.
595 	 */
596 	if (cpu_topology_ht_ids) {
597 		scale_lim = PQ_L2_SIZE / cpu_topology_phys_ids;
598 		scale_lim = scale_lim / cpu_topology_core_ids;
599 		scale_lim = scale_lim / cpu_topology_ht_ids;
600 		scale_lim = scale_lim * cpu_topology_ht_ids;
601 		scale_lim = scale_lim * cpu_topology_core_ids;
602 		scale_lim = scale_lim * cpu_topology_phys_ids;
603 	} else {
604 		scale_lim = PQ_L2_SIZE;
605 	}
606 
607 	/*
608 	 * Calculate an average, set hysteresis for balancing from
609 	 * 10% below the average to the average.
610 	 */
611 	lcnt_hi = 0;
612 	for (i = 0; i < scale_lim; ++i) {
613 		lcnt_hi += vm_page_queues[i].lcnt;
614 	}
615 	lcnt_hi /= scale_lim;
616 	lcnt_lo = lcnt_hi - lcnt_hi / 10;
617 
618 	kprintf("vm_page: avg %ld pages per queue, %d queues\n",
619 		lcnt_hi, scale_lim);
620 
621 	iter = 0;
622 	for (i = 0; i < scale_lim; ++i) {
623 		vpq = &vm_page_queues[PQ_FREE + i];
624 		while (vpq->lcnt < lcnt_lo) {
625 			struct vpgqueues *vptmp;
626 
627 			iter = (iter + 1) & PQ_L2_MASK;
628 			vptmp = &vm_page_queues[PQ_FREE + iter];
629 			if (vptmp->lcnt < lcnt_hi)
630 				continue;
631 			m = TAILQ_FIRST(&vptmp->pl);
632 			KKASSERT(m->queue == PQ_FREE + iter);
633 			TAILQ_REMOVE(&vptmp->pl, m, pageq);
634 			--vptmp->lcnt;
635 			/* queue doesn't change, no need to adj cnt */
636 			m->queue -= m->pc;
637 			m->pc = i;
638 			m->queue += m->pc;
639 			TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
640 			++vpq->lcnt;
641 		}
642 	}
643 	crit_exit();
644 }
645 
646 static
647 void
648 vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes)
649 {
650 	int cpuid;
651 	int i;
652 
653 	switch(cpup->type) {
654 	case PACKAGE_LEVEL:
655 		cpup->phys_mem += bytes;
656 		break;
657 	case CHIP_LEVEL:
658 		/*
659 		 * All members should have the same chipid, so we only need
660 		 * to pull out one member.
661 		 */
662 		if (CPUMASK_TESTNZERO(cpup->members)) {
663 			cpuid = BSFCPUMASK(cpup->members);
664 			if (physid ==
665 			    get_chip_ID_from_APICID(CPUID_TO_APICID(cpuid))) {
666 				cpup->phys_mem += bytes;
667 			}
668 		}
669 		break;
670 	case CORE_LEVEL:
671 	case THREAD_LEVEL:
672 		/*
673 		 * Just inherit from the parent node
674 		 */
675 		cpup->phys_mem = cpup->parent_node->phys_mem;
676 		break;
677 	}
678 	for (i = 0; i < MAXCPU && cpup->child_node[i]; ++i)
679 		vm_numa_add_topology_mem(cpup->child_node[i], physid, bytes);
680 }
681 
682 /*
683  * We tended to reserve a ton of memory for contigmalloc().  Now that most
684  * drivers have initialized we want to return most the remaining free
685  * reserve back to the VM page queues so they can be used for normal
686  * allocations.
687  *
688  * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
689  */
690 static void
691 vm_page_startup_finish(void *dummy __unused)
692 {
693 	alist_blk_t blk;
694 	alist_blk_t rblk;
695 	alist_blk_t count;
696 	alist_blk_t xcount;
697 	alist_blk_t bfree;
698 	vm_page_t m;
699 
700 	spin_lock(&vm_contig_spin);
701 	for (;;) {
702 		bfree = alist_free_info(&vm_contig_alist, &blk, &count);
703 		if (bfree <= vm_dma_reserved / PAGE_SIZE)
704 			break;
705 		if (count == 0)
706 			break;
707 
708 		/*
709 		 * Figure out how much of the initial reserve we have to
710 		 * free in order to reach our target.
711 		 */
712 		bfree -= vm_dma_reserved / PAGE_SIZE;
713 		if (count > bfree) {
714 			blk += count - bfree;
715 			count = bfree;
716 		}
717 
718 		/*
719 		 * Calculate the nearest power of 2 <= count.
720 		 */
721 		for (xcount = 1; xcount <= count; xcount <<= 1)
722 			;
723 		xcount >>= 1;
724 		blk += count - xcount;
725 		count = xcount;
726 
727 		/*
728 		 * Allocate the pages from the alist, then free them to
729 		 * the normal VM page queues.
730 		 *
731 		 * Pages allocated from the alist are wired.  We have to
732 		 * busy, unwire, and free them.  We must also adjust
733 		 * vm_low_phys_reserved before freeing any pages to prevent
734 		 * confusion.
735 		 */
736 		rblk = alist_alloc(&vm_contig_alist, blk, count);
737 		if (rblk != blk) {
738 			kprintf("vm_page_startup_finish: Unable to return "
739 				"dma space @0x%08x/%d -> 0x%08x\n",
740 				blk, count, rblk);
741 			break;
742 		}
743 		atomic_add_long(&vmstats.v_dma_pages, -(long)count);
744 		spin_unlock(&vm_contig_spin);
745 
746 		m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
747 		vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
748 		while (count) {
749 			vm_page_busy_wait(m, FALSE, "cpgfr");
750 			vm_page_unwire(m, 0);
751 			vm_page_free(m);
752 			--count;
753 			++m;
754 		}
755 		spin_lock(&vm_contig_spin);
756 	}
757 	spin_unlock(&vm_contig_spin);
758 
759 	/*
760 	 * Print out how much DMA space drivers have already allocated and
761 	 * how much is left over.
762 	 */
763 	kprintf("DMA space used: %jdk, remaining available: %jdk\n",
764 		(intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
765 		(PAGE_SIZE / 1024),
766 		(intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
767 }
768 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
769 	vm_page_startup_finish, NULL);
770 
771 
772 /*
773  * Scan comparison function for Red-Black tree scans.  An inclusive
774  * (start,end) is expected.  Other fields are not used.
775  */
776 int
777 rb_vm_page_scancmp(struct vm_page *p, void *data)
778 {
779 	struct rb_vm_page_scan_info *info = data;
780 
781 	if (p->pindex < info->start_pindex)
782 		return(-1);
783 	if (p->pindex > info->end_pindex)
784 		return(1);
785 	return(0);
786 }
787 
788 int
789 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
790 {
791 	if (p1->pindex < p2->pindex)
792 		return(-1);
793 	if (p1->pindex > p2->pindex)
794 		return(1);
795 	return(0);
796 }
797 
798 void
799 vm_page_init(vm_page_t m)
800 {
801 	/* do nothing for now.  Called from pmap_page_init() */
802 }
803 
804 /*
805  * Each page queue has its own spin lock, which is fairly optimal for
806  * allocating and freeing pages at least.
807  *
808  * The caller must hold the vm_page_spin_lock() before locking a vm_page's
809  * queue spinlock via this function.  Also note that m->queue cannot change
810  * unless both the page and queue are locked.
811  */
812 static __inline
813 void
814 _vm_page_queue_spin_lock(vm_page_t m)
815 {
816 	u_short queue;
817 
818 	queue = m->queue;
819 	if (queue != PQ_NONE) {
820 		spin_lock(&vm_page_queues[queue].spin);
821 		KKASSERT(queue == m->queue);
822 	}
823 }
824 
825 static __inline
826 void
827 _vm_page_queue_spin_unlock(vm_page_t m)
828 {
829 	u_short queue;
830 
831 	queue = m->queue;
832 	cpu_ccfence();
833 	if (queue != PQ_NONE)
834 		spin_unlock(&vm_page_queues[queue].spin);
835 }
836 
837 static __inline
838 void
839 _vm_page_queues_spin_lock(u_short queue)
840 {
841 	cpu_ccfence();
842 	if (queue != PQ_NONE)
843 		spin_lock(&vm_page_queues[queue].spin);
844 }
845 
846 
847 static __inline
848 void
849 _vm_page_queues_spin_unlock(u_short queue)
850 {
851 	cpu_ccfence();
852 	if (queue != PQ_NONE)
853 		spin_unlock(&vm_page_queues[queue].spin);
854 }
855 
856 void
857 vm_page_queue_spin_lock(vm_page_t m)
858 {
859 	_vm_page_queue_spin_lock(m);
860 }
861 
862 void
863 vm_page_queues_spin_lock(u_short queue)
864 {
865 	_vm_page_queues_spin_lock(queue);
866 }
867 
868 void
869 vm_page_queue_spin_unlock(vm_page_t m)
870 {
871 	_vm_page_queue_spin_unlock(m);
872 }
873 
874 void
875 vm_page_queues_spin_unlock(u_short queue)
876 {
877 	_vm_page_queues_spin_unlock(queue);
878 }
879 
880 /*
881  * This locks the specified vm_page and its queue in the proper order
882  * (page first, then queue).  The queue may change so the caller must
883  * recheck on return.
884  */
885 static __inline
886 void
887 _vm_page_and_queue_spin_lock(vm_page_t m)
888 {
889 	vm_page_spin_lock(m);
890 	_vm_page_queue_spin_lock(m);
891 }
892 
893 static __inline
894 void
895 _vm_page_and_queue_spin_unlock(vm_page_t m)
896 {
897 	_vm_page_queues_spin_unlock(m->queue);
898 	vm_page_spin_unlock(m);
899 }
900 
901 void
902 vm_page_and_queue_spin_unlock(vm_page_t m)
903 {
904 	_vm_page_and_queue_spin_unlock(m);
905 }
906 
907 void
908 vm_page_and_queue_spin_lock(vm_page_t m)
909 {
910 	_vm_page_and_queue_spin_lock(m);
911 }
912 
913 /*
914  * Helper function removes vm_page from its current queue.
915  * Returns the base queue the page used to be on.
916  *
917  * The vm_page and the queue must be spinlocked.
918  * This function will unlock the queue but leave the page spinlocked.
919  */
920 static __inline u_short
921 _vm_page_rem_queue_spinlocked(vm_page_t m)
922 {
923 	struct vpgqueues *pq;
924 	u_short queue;
925 	u_short oqueue;
926 	long *cnt;
927 
928 	queue = m->queue;
929 	if (queue != PQ_NONE) {
930 		pq = &vm_page_queues[queue];
931 		TAILQ_REMOVE(&pq->pl, m, pageq);
932 
933 		/*
934 		 * Adjust our pcpu stats.  In order for the nominal low-memory
935 		 * algorithms to work properly we don't let any pcpu stat get
936 		 * too negative before we force it to be rolled-up into the
937 		 * global stats.  Otherwise our pageout and vm_wait tests
938 		 * will fail badly.
939 		 *
940 		 * The idea here is to reduce unnecessary SMP cache
941 		 * mastership changes in the global vmstats, which can be
942 		 * particularly bad in multi-socket systems.
943 		 */
944 		cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
945 		atomic_add_long(cnt, -1);
946 		if (*cnt < -VMMETER_SLOP_COUNT) {
947 			u_long copy = atomic_swap_long(cnt, 0);
948 			cnt = (long *)((char *)&vmstats + pq->cnt_offset);
949 			atomic_add_long(cnt, copy);
950 			cnt = (long *)((char *)&mycpu->gd_vmstats +
951 				      pq->cnt_offset);
952 			atomic_add_long(cnt, copy);
953 		}
954 		pq->lcnt--;
955 		m->queue = PQ_NONE;
956 		oqueue = queue;
957 		queue -= m->pc;
958 		vm_page_queues_spin_unlock(oqueue);	/* intended */
959 	}
960 	return queue;
961 }
962 
963 /*
964  * Helper function places the vm_page on the specified queue.  Generally
965  * speaking only PQ_FREE pages are placed at the head, to allow them to
966  * be allocated sooner rather than later on the assumption that they
967  * are cache-hot.
968  *
969  * The vm_page must be spinlocked.
970  * This function will return with both the page and the queue locked.
971  */
972 static __inline void
973 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
974 {
975 	struct vpgqueues *pq;
976 	u_long *cnt;
977 
978 	KKASSERT(m->queue == PQ_NONE);
979 
980 	if (queue != PQ_NONE) {
981 		vm_page_queues_spin_lock(queue);
982 		pq = &vm_page_queues[queue];
983 		++pq->lcnt;
984 
985 		/*
986 		 * Adjust our pcpu stats.  If a system entity really needs
987 		 * to incorporate the count it will call vmstats_rollup()
988 		 * to roll it all up into the global vmstats strufture.
989 		 */
990 		cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
991 		atomic_add_long(cnt, 1);
992 
993 		/*
994 		 * PQ_FREE is always handled LIFO style to try to provide
995 		 * cache-hot pages to programs.
996 		 */
997 		m->queue = queue;
998 		if (queue - m->pc == PQ_FREE) {
999 			TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1000 		} else if (athead) {
1001 			TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1002 		} else {
1003 			TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1004 		}
1005 		/* leave the queue spinlocked */
1006 	}
1007 }
1008 
1009 /*
1010  * Wait until page is no longer BUSY.  If also_m_busy is TRUE we wait
1011  * until the page is no longer BUSY or SBUSY (busy_count field is 0).
1012  *
1013  * Returns TRUE if it had to sleep, FALSE if we did not.  Only one sleep
1014  * call will be made before returning.
1015  *
1016  * This function does NOT busy the page and on return the page is not
1017  * guaranteed to be available.
1018  */
1019 void
1020 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
1021 {
1022 	u_int32_t busy_count;
1023 
1024 	for (;;) {
1025 		busy_count = m->busy_count;
1026 		cpu_ccfence();
1027 
1028 		if ((busy_count & PBUSY_LOCKED) == 0 &&
1029 		    (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) {
1030 			break;
1031 		}
1032 		tsleep_interlock(m, 0);
1033 		if (atomic_cmpset_int(&m->busy_count, busy_count,
1034 				      busy_count | PBUSY_WANTED)) {
1035 			atomic_set_int(&m->flags, PG_REFERENCED);
1036 			tsleep(m, PINTERLOCKED, msg, 0);
1037 			break;
1038 		}
1039 	}
1040 }
1041 
1042 /*
1043  * This calculates and returns a page color given an optional VM object and
1044  * either a pindex or an iterator.  We attempt to return a cpu-localized
1045  * pg_color that is still roughly 16-way set-associative.  The CPU topology
1046  * is used if it was probed.
1047  *
1048  * The caller may use the returned value to index into e.g. PQ_FREE when
1049  * allocating a page in order to nominally obtain pages that are hopefully
1050  * already localized to the requesting cpu.  This function is not able to
1051  * provide any sort of guarantee of this, but does its best to improve
1052  * hardware cache management performance.
1053  *
1054  * WARNING! The caller must mask the returned value with PQ_L2_MASK.
1055  */
1056 u_short
1057 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
1058 {
1059 	u_short pg_color;
1060 	int object_pg_color;
1061 
1062 	/*
1063 	 * WARNING! cpu_topology_core_ids might not be a power of two.
1064 	 *	    We also shouldn't make assumptions about
1065 	 *	    cpu_topology_phys_ids either.
1066 	 *
1067 	 * WARNING! ncpus might not be known at this time (during early
1068 	 *	    boot), and might be set to 1.
1069 	 *
1070 	 * General format: [phys_id][core_id][cpuid][set-associativity]
1071 	 * (but uses modulo, so not necessarily precise bit masks)
1072 	 */
1073 	object_pg_color = object ? object->pg_color : 0;
1074 
1075 	if (cpu_topology_ht_ids) {
1076 		int phys_id;
1077 		int core_id;
1078 		int ht_id;
1079 		int physcale;
1080 		int grpscale;
1081 		int cpuscale;
1082 
1083 		/*
1084 		 * Translate cpuid to socket, core, and hyperthread id.
1085 		 */
1086 		phys_id = get_cpu_phys_id(cpuid);
1087 		core_id = get_cpu_core_id(cpuid);
1088 		ht_id = get_cpu_ht_id(cpuid);
1089 
1090 		/*
1091 		 * Calculate pg_color for our array index.
1092 		 *
1093 		 * physcale - socket multiplier.
1094 		 * grpscale - core multiplier (cores per socket)
1095 		 * cpu*	    - cpus per core
1096 		 *
1097 		 * WARNING! In early boot, ncpus has not yet been
1098 		 *	    initialized and may be set to (1).
1099 		 *
1100 		 * WARNING! physcale must match the organization that
1101 		 *	    vm_numa_organize() creates to ensure that
1102 		 *	    we properly localize allocations to the
1103 		 *	    requested cpuid.
1104 		 */
1105 		physcale = PQ_L2_SIZE / cpu_topology_phys_ids;
1106 		grpscale = physcale / cpu_topology_core_ids;
1107 		cpuscale = grpscale / cpu_topology_ht_ids;
1108 
1109 		pg_color = phys_id * physcale;
1110 		pg_color += core_id * grpscale;
1111 		pg_color += ht_id * cpuscale;
1112 		pg_color += (pindex + object_pg_color) % cpuscale;
1113 
1114 #if 0
1115 		if (grpsize >= 8) {
1116 			pg_color += (pindex + object_pg_color) % grpsize;
1117 		} else {
1118 			if (grpsize <= 2) {
1119 				grpsize = 8;
1120 			} else {
1121 				/* 3->9, 4->8, 5->10, 6->12, 7->14 */
1122 				grpsize += grpsize;
1123 				if (grpsize < 8)
1124 					grpsize += grpsize;
1125 			}
1126 			pg_color += (pindex + object_pg_color) % grpsize;
1127 		}
1128 #endif
1129 	} else {
1130 		/*
1131 		 * Unknown topology, distribute things evenly.
1132 		 *
1133 		 * WARNING! In early boot, ncpus has not yet been
1134 		 *	    initialized and may be set to (1).
1135 		 */
1136 		int cpuscale;
1137 
1138 		cpuscale = PQ_L2_SIZE / ncpus;
1139 
1140 		pg_color = cpuid * cpuscale;
1141 		pg_color += (pindex + object_pg_color) % cpuscale;
1142 	}
1143 	return (pg_color & PQ_L2_MASK);
1144 }
1145 
1146 /*
1147  * Wait until BUSY can be set, then set it.  If also_m_busy is TRUE we
1148  * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED.
1149  */
1150 void
1151 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
1152 				     int also_m_busy, const char *msg
1153 				     VM_PAGE_DEBUG_ARGS)
1154 {
1155 	u_int32_t busy_count;
1156 
1157 	for (;;) {
1158 		busy_count = m->busy_count;
1159 		cpu_ccfence();
1160 		if (busy_count & PBUSY_LOCKED) {
1161 			tsleep_interlock(m, 0);
1162 			if (atomic_cmpset_int(&m->busy_count, busy_count,
1163 					  busy_count | PBUSY_WANTED)) {
1164 				atomic_set_int(&m->flags, PG_REFERENCED);
1165 				tsleep(m, PINTERLOCKED, msg, 0);
1166 			}
1167 		} else if (also_m_busy && busy_count) {
1168 			tsleep_interlock(m, 0);
1169 			if (atomic_cmpset_int(&m->busy_count, busy_count,
1170 					  busy_count | PBUSY_WANTED)) {
1171 				atomic_set_int(&m->flags, PG_REFERENCED);
1172 				tsleep(m, PINTERLOCKED, msg, 0);
1173 			}
1174 		} else {
1175 			if (atomic_cmpset_int(&m->busy_count, busy_count,
1176 					      busy_count | PBUSY_LOCKED)) {
1177 #ifdef VM_PAGE_DEBUG
1178 				m->busy_func = func;
1179 				m->busy_line = lineno;
1180 #endif
1181 				break;
1182 			}
1183 		}
1184 	}
1185 }
1186 
1187 /*
1188  * Attempt to set BUSY.  If also_m_busy is TRUE we only succeed if
1189  * m->busy_count is also 0.
1190  *
1191  * Returns non-zero on failure.
1192  */
1193 int
1194 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1195 				    VM_PAGE_DEBUG_ARGS)
1196 {
1197 	u_int32_t busy_count;
1198 
1199 	for (;;) {
1200 		busy_count = m->busy_count;
1201 		cpu_ccfence();
1202 		if (busy_count & PBUSY_LOCKED)
1203 			return TRUE;
1204 		if (also_m_busy && (busy_count & PBUSY_MASK) != 0)
1205 			return TRUE;
1206 		if (atomic_cmpset_int(&m->busy_count, busy_count,
1207 				      busy_count | PBUSY_LOCKED)) {
1208 #ifdef VM_PAGE_DEBUG
1209 				m->busy_func = func;
1210 				m->busy_line = lineno;
1211 #endif
1212 			return FALSE;
1213 		}
1214 	}
1215 }
1216 
1217 /*
1218  * Clear the BUSY flag and return non-zero to indicate to the caller
1219  * that a wakeup() should be performed.
1220  *
1221  * The vm_page must be spinlocked and will remain spinlocked on return.
1222  * The related queue must NOT be spinlocked (which could deadlock us).
1223  *
1224  * (inline version)
1225  */
1226 static __inline
1227 int
1228 _vm_page_wakeup(vm_page_t m)
1229 {
1230 	u_int32_t busy_count;
1231 
1232 	for (;;) {
1233 		busy_count = m->busy_count;
1234 		cpu_ccfence();
1235 		if (atomic_cmpset_int(&m->busy_count, busy_count,
1236 				      busy_count &
1237 				      ~(PBUSY_LOCKED | PBUSY_WANTED))) {
1238 			break;
1239 		}
1240 	}
1241 	return((int)(busy_count & PBUSY_WANTED));
1242 }
1243 
1244 /*
1245  * Clear the BUSY flag and wakeup anyone waiting for the page.  This
1246  * is typically the last call you make on a page before moving onto
1247  * other things.
1248  */
1249 void
1250 vm_page_wakeup(vm_page_t m)
1251 {
1252         KASSERT(m->busy_count & PBUSY_LOCKED,
1253 		("vm_page_wakeup: page not busy!!!"));
1254 	vm_page_spin_lock(m);
1255 	if (_vm_page_wakeup(m)) {
1256 		vm_page_spin_unlock(m);
1257 		wakeup(m);
1258 	} else {
1259 		vm_page_spin_unlock(m);
1260 	}
1261 }
1262 
1263 /*
1264  * Holding a page keeps it from being reused.  Other parts of the system
1265  * can still disassociate the page from its current object and free it, or
1266  * perform read or write I/O on it and/or otherwise manipulate the page,
1267  * but if the page is held the VM system will leave the page and its data
1268  * intact and not reuse the page for other purposes until the last hold
1269  * reference is released.  (see vm_page_wire() if you want to prevent the
1270  * page from being disassociated from its object too).
1271  *
1272  * The caller must still validate the contents of the page and, if necessary,
1273  * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1274  * before manipulating the page.
1275  *
1276  * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1277  */
1278 void
1279 vm_page_hold(vm_page_t m)
1280 {
1281 	vm_page_spin_lock(m);
1282 	atomic_add_int(&m->hold_count, 1);
1283 	if (m->queue - m->pc == PQ_FREE) {
1284 		_vm_page_queue_spin_lock(m);
1285 		_vm_page_rem_queue_spinlocked(m);
1286 		_vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1287 		_vm_page_queue_spin_unlock(m);
1288 	}
1289 	vm_page_spin_unlock(m);
1290 }
1291 
1292 /*
1293  * The opposite of vm_page_hold().  If the page is on the HOLD queue
1294  * it was freed while held and must be moved back to the FREE queue.
1295  */
1296 void
1297 vm_page_unhold(vm_page_t m)
1298 {
1299 	KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1300 		("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1301 		 m, m->hold_count, m->queue - m->pc));
1302 	vm_page_spin_lock(m);
1303 	atomic_add_int(&m->hold_count, -1);
1304 	if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1305 		_vm_page_queue_spin_lock(m);
1306 		_vm_page_rem_queue_spinlocked(m);
1307 		_vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1308 		_vm_page_queue_spin_unlock(m);
1309 	}
1310 	vm_page_spin_unlock(m);
1311 }
1312 
1313 /*
1314  *	vm_page_getfake:
1315  *
1316  *	Create a fictitious page with the specified physical address and
1317  *	memory attribute.  The memory attribute is the only the machine-
1318  *	dependent aspect of a fictitious page that must be initialized.
1319  */
1320 
1321 void
1322 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1323 {
1324 
1325 	if ((m->flags & PG_FICTITIOUS) != 0) {
1326 		/*
1327 		 * The page's memattr might have changed since the
1328 		 * previous initialization.  Update the pmap to the
1329 		 * new memattr.
1330 		 */
1331 		goto memattr;
1332 	}
1333 	m->phys_addr = paddr;
1334 	m->queue = PQ_NONE;
1335 	/* Fictitious pages don't use "segind". */
1336 	/* Fictitious pages don't use "order" or "pool". */
1337 	m->flags = PG_FICTITIOUS | PG_UNMANAGED;
1338 	m->busy_count = PBUSY_LOCKED;
1339 	m->wire_count = 1;
1340 	spin_init(&m->spin, "fake_page");
1341 	pmap_page_init(m);
1342 memattr:
1343 	pmap_page_set_memattr(m, memattr);
1344 }
1345 
1346 /*
1347  * Inserts the given vm_page into the object and object list.
1348  *
1349  * The pagetables are not updated but will presumably fault the page
1350  * in if necessary, or if a kernel page the caller will at some point
1351  * enter the page into the kernel's pmap.  We are not allowed to block
1352  * here so we *can't* do this anyway.
1353  *
1354  * This routine may not block.
1355  * This routine must be called with the vm_object held.
1356  * This routine must be called with a critical section held.
1357  *
1358  * This routine returns TRUE if the page was inserted into the object
1359  * successfully, and FALSE if the page already exists in the object.
1360  */
1361 int
1362 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1363 {
1364 	ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1365 	if (m->object != NULL)
1366 		panic("vm_page_insert: already inserted");
1367 
1368 	atomic_add_int(&object->generation, 1);
1369 
1370 	/*
1371 	 * Record the object/offset pair in this page and add the
1372 	 * pv_list_count of the page to the object.
1373 	 *
1374 	 * The vm_page spin lock is required for interactions with the pmap.
1375 	 */
1376 	vm_page_spin_lock(m);
1377 	m->object = object;
1378 	m->pindex = pindex;
1379 	if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1380 		m->object = NULL;
1381 		m->pindex = 0;
1382 		vm_page_spin_unlock(m);
1383 		return FALSE;
1384 	}
1385 	++object->resident_page_count;
1386 	++mycpu->gd_vmtotal.t_rm;
1387 	vm_page_spin_unlock(m);
1388 
1389 	/*
1390 	 * Since we are inserting a new and possibly dirty page,
1391 	 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1392 	 */
1393 	if ((m->valid & m->dirty) ||
1394 	    (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1395 		vm_object_set_writeable_dirty(object);
1396 
1397 	/*
1398 	 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1399 	 */
1400 	swap_pager_page_inserted(m);
1401 	return TRUE;
1402 }
1403 
1404 /*
1405  * Removes the given vm_page_t from the (object,index) table
1406  *
1407  * The underlying pmap entry (if any) is NOT removed here.
1408  * This routine may not block.
1409  *
1410  * The page must be BUSY and will remain BUSY on return.
1411  * No other requirements.
1412  *
1413  * NOTE: FreeBSD side effect was to unbusy the page on return.  We leave
1414  *	 it busy.
1415  */
1416 void
1417 vm_page_remove(vm_page_t m)
1418 {
1419 	vm_object_t object;
1420 
1421 	if (m->object == NULL) {
1422 		return;
1423 	}
1424 
1425 	if ((m->busy_count & PBUSY_LOCKED) == 0)
1426 		panic("vm_page_remove: page not busy");
1427 
1428 	object = m->object;
1429 
1430 	vm_object_hold(object);
1431 
1432 	/*
1433 	 * Remove the page from the object and update the object.
1434 	 *
1435 	 * The vm_page spin lock is required for interactions with the pmap.
1436 	 */
1437 	vm_page_spin_lock(m);
1438 	vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1439 	--object->resident_page_count;
1440 	--mycpu->gd_vmtotal.t_rm;
1441 	m->object = NULL;
1442 	atomic_add_int(&object->generation, 1);
1443 	vm_page_spin_unlock(m);
1444 
1445 	vm_object_drop(object);
1446 }
1447 
1448 /*
1449  * Locate and return the page at (object, pindex), or NULL if the
1450  * page could not be found.
1451  *
1452  * The caller must hold the vm_object token.
1453  */
1454 vm_page_t
1455 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1456 {
1457 	vm_page_t m;
1458 
1459 	/*
1460 	 * Search the hash table for this object/offset pair
1461 	 */
1462 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1463 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1464 	KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1465 	return(m);
1466 }
1467 
1468 vm_page_t
1469 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1470 					    vm_pindex_t pindex,
1471 					    int also_m_busy, const char *msg
1472 					    VM_PAGE_DEBUG_ARGS)
1473 {
1474 	u_int32_t busy_count;
1475 	vm_page_t m;
1476 
1477 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1478 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1479 	while (m) {
1480 		KKASSERT(m->object == object && m->pindex == pindex);
1481 		busy_count = m->busy_count;
1482 		cpu_ccfence();
1483 		if (busy_count & PBUSY_LOCKED) {
1484 			tsleep_interlock(m, 0);
1485 			if (atomic_cmpset_int(&m->busy_count, busy_count,
1486 					  busy_count | PBUSY_WANTED)) {
1487 				atomic_set_int(&m->flags, PG_REFERENCED);
1488 				tsleep(m, PINTERLOCKED, msg, 0);
1489 				m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1490 							      pindex);
1491 			}
1492 		} else if (also_m_busy && busy_count) {
1493 			tsleep_interlock(m, 0);
1494 			if (atomic_cmpset_int(&m->busy_count, busy_count,
1495 					  busy_count | PBUSY_WANTED)) {
1496 				atomic_set_int(&m->flags, PG_REFERENCED);
1497 				tsleep(m, PINTERLOCKED, msg, 0);
1498 				m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1499 							      pindex);
1500 			}
1501 		} else if (atomic_cmpset_int(&m->busy_count, busy_count,
1502 					     busy_count | PBUSY_LOCKED)) {
1503 #ifdef VM_PAGE_DEBUG
1504 			m->busy_func = func;
1505 			m->busy_line = lineno;
1506 #endif
1507 			break;
1508 		}
1509 	}
1510 	return m;
1511 }
1512 
1513 /*
1514  * Attempt to lookup and busy a page.
1515  *
1516  * Returns NULL if the page could not be found
1517  *
1518  * Returns a vm_page and error == TRUE if the page exists but could not
1519  * be busied.
1520  *
1521  * Returns a vm_page and error == FALSE on success.
1522  */
1523 vm_page_t
1524 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1525 					   vm_pindex_t pindex,
1526 					   int also_m_busy, int *errorp
1527 					   VM_PAGE_DEBUG_ARGS)
1528 {
1529 	u_int32_t busy_count;
1530 	vm_page_t m;
1531 
1532 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1533 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1534 	*errorp = FALSE;
1535 	while (m) {
1536 		KKASSERT(m->object == object && m->pindex == pindex);
1537 		busy_count = m->busy_count;
1538 		cpu_ccfence();
1539 		if (busy_count & PBUSY_LOCKED) {
1540 			*errorp = TRUE;
1541 			break;
1542 		}
1543 		if (also_m_busy && busy_count) {
1544 			*errorp = TRUE;
1545 			break;
1546 		}
1547 		if (atomic_cmpset_int(&m->busy_count, busy_count,
1548 				      busy_count | PBUSY_LOCKED)) {
1549 #ifdef VM_PAGE_DEBUG
1550 			m->busy_func = func;
1551 			m->busy_line = lineno;
1552 #endif
1553 			break;
1554 		}
1555 	}
1556 	return m;
1557 }
1558 
1559 /*
1560  * Returns a page that is only soft-busied for use by the caller in
1561  * a read-only fashion.  Returns NULL if the page could not be found,
1562  * the soft busy could not be obtained, or the page data is invalid.
1563  */
1564 vm_page_t
1565 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex,
1566 			 int pgoff, int pgbytes)
1567 {
1568 	vm_page_t m;
1569 
1570 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1571 	m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1572 	if (m) {
1573 		if ((m->valid != VM_PAGE_BITS_ALL &&
1574 		     !vm_page_is_valid(m, pgoff, pgbytes)) ||
1575 		    (m->flags & PG_FICTITIOUS)) {
1576 			m = NULL;
1577 		} else if (vm_page_sbusy_try(m)) {
1578 			m = NULL;
1579 		} else if ((m->valid != VM_PAGE_BITS_ALL &&
1580 			    !vm_page_is_valid(m, pgoff, pgbytes)) ||
1581 			   (m->flags & PG_FICTITIOUS)) {
1582 			vm_page_sbusy_drop(m);
1583 			m = NULL;
1584 		}
1585 	}
1586 	return m;
1587 }
1588 
1589 /*
1590  * Caller must hold the related vm_object
1591  */
1592 vm_page_t
1593 vm_page_next(vm_page_t m)
1594 {
1595 	vm_page_t next;
1596 
1597 	next = vm_page_rb_tree_RB_NEXT(m);
1598 	if (next && next->pindex != m->pindex + 1)
1599 		next = NULL;
1600 	return (next);
1601 }
1602 
1603 /*
1604  * vm_page_rename()
1605  *
1606  * Move the given vm_page from its current object to the specified
1607  * target object/offset.  The page must be busy and will remain so
1608  * on return.
1609  *
1610  * new_object must be held.
1611  * This routine might block. XXX ?
1612  *
1613  * NOTE: Swap associated with the page must be invalidated by the move.  We
1614  *       have to do this for several reasons:  (1) we aren't freeing the
1615  *       page, (2) we are dirtying the page, (3) the VM system is probably
1616  *       moving the page from object A to B, and will then later move
1617  *       the backing store from A to B and we can't have a conflict.
1618  *
1619  * NOTE: We *always* dirty the page.  It is necessary both for the
1620  *       fact that we moved it, and because we may be invalidating
1621  *	 swap.  If the page is on the cache, we have to deactivate it
1622  *	 or vm_page_dirty() will panic.  Dirty pages are not allowed
1623  *	 on the cache.
1624  */
1625 void
1626 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1627 {
1628 	KKASSERT(m->busy_count & PBUSY_LOCKED);
1629 	ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1630 	if (m->object) {
1631 		ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1632 		vm_page_remove(m);
1633 	}
1634 	if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1635 		panic("vm_page_rename: target exists (%p,%"PRIu64")",
1636 		      new_object, new_pindex);
1637 	}
1638 	if (m->queue - m->pc == PQ_CACHE)
1639 		vm_page_deactivate(m);
1640 	vm_page_dirty(m);
1641 }
1642 
1643 /*
1644  * vm_page_unqueue() without any wakeup.  This routine is used when a page
1645  * is to remain BUSYied by the caller.
1646  *
1647  * This routine may not block.
1648  */
1649 void
1650 vm_page_unqueue_nowakeup(vm_page_t m)
1651 {
1652 	vm_page_and_queue_spin_lock(m);
1653 	(void)_vm_page_rem_queue_spinlocked(m);
1654 	vm_page_spin_unlock(m);
1655 }
1656 
1657 /*
1658  * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1659  * if necessary.
1660  *
1661  * This routine may not block.
1662  */
1663 void
1664 vm_page_unqueue(vm_page_t m)
1665 {
1666 	u_short queue;
1667 
1668 	vm_page_and_queue_spin_lock(m);
1669 	queue = _vm_page_rem_queue_spinlocked(m);
1670 	if (queue == PQ_FREE || queue == PQ_CACHE) {
1671 		vm_page_spin_unlock(m);
1672 		pagedaemon_wakeup();
1673 	} else {
1674 		vm_page_spin_unlock(m);
1675 	}
1676 }
1677 
1678 /*
1679  * vm_page_list_find()
1680  *
1681  * Find a page on the specified queue with color optimization.
1682  *
1683  * The page coloring optimization attempts to locate a page that does
1684  * not overload other nearby pages in the object in the cpu's L1 or L2
1685  * caches.  We need this optimization because cpu caches tend to be
1686  * physical caches, while object spaces tend to be virtual.
1687  *
1688  * The page coloring optimization also, very importantly, tries to localize
1689  * memory to cpus and physical sockets.
1690  *
1691  * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the
1692  * algorithm is adjusted to localize allocations on a per-core basis.
1693  * This is done by 'twisting' the colors.
1694  *
1695  * The page is returned spinlocked and removed from its queue (it will
1696  * be on PQ_NONE), or NULL. The page is not BUSY'd.  The caller
1697  * is responsible for dealing with the busy-page case (usually by
1698  * deactivating the page and looping).
1699  *
1700  * NOTE:  This routine is carefully inlined.  A non-inlined version
1701  *	  is available for outside callers but the only critical path is
1702  *	  from within this source file.
1703  *
1704  * NOTE:  This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1705  *	  represent stable storage, allowing us to order our locks vm_page
1706  *	  first, then queue.
1707  */
1708 static __inline
1709 vm_page_t
1710 _vm_page_list_find(int basequeue, int index)
1711 {
1712 	struct vpgqueues *pq;
1713 	vm_page_t m;
1714 
1715 	index &= PQ_L2_MASK;
1716 	pq = &vm_page_queues[basequeue + index];
1717 
1718 	/*
1719 	 * Try this cpu's colored queue first.  Test for a page unlocked,
1720 	 * then lock the queue and locate a page.  Note that the lock order
1721 	 * is reversed, but we do not want to dwadle on the page spinlock
1722 	 * anyway as it is held significantly longer than the queue spinlock.
1723 	 */
1724 	if (TAILQ_FIRST(&pq->pl)) {
1725 		spin_lock(&pq->spin);
1726 		TAILQ_FOREACH(m, &pq->pl, pageq) {
1727 			if (spin_trylock(&m->spin) == 0)
1728 				continue;
1729 			KKASSERT(m->queue == basequeue + index);
1730 			_vm_page_rem_queue_spinlocked(m);
1731 			return(m);
1732 		}
1733 		spin_unlock(&pq->spin);
1734 	}
1735 
1736 	/*
1737 	 * If we are unable to get a page, do a more involved NUMA-aware
1738 	 * search.
1739 	 */
1740 	m = _vm_page_list_find2(basequeue, index);
1741 	return(m);
1742 }
1743 
1744 /*
1745  * If we could not find the page in the desired queue try to find it in
1746  * a nearby (NUMA-aware) queue.
1747  */
1748 static vm_page_t
1749 _vm_page_list_find2(int basequeue, int index)
1750 {
1751 	struct vpgqueues *pq;
1752 	vm_page_t m = NULL;
1753 	int pqmask = PQ_SET_ASSOC_MASK >> 1;
1754 	int pqi;
1755 	int i;
1756 
1757 	index &= PQ_L2_MASK;
1758 	pq = &vm_page_queues[basequeue];
1759 
1760 	/*
1761 	 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1762 	 * else fails (PQ_L2_MASK which is 255).
1763 	 *
1764 	 * Test each queue unlocked first, then lock the queue and locate
1765 	 * a page.  Note that the lock order is reversed, but we do not want
1766 	 * to dwadle on the page spinlock anyway as it is held significantly
1767 	 * longer than the queue spinlock.
1768 	 */
1769 	do {
1770 		pqmask = (pqmask << 1) | 1;
1771 		for (i = 0; i <= pqmask; ++i) {
1772 			pqi = (index & ~pqmask) | ((index + i) & pqmask);
1773 			if (TAILQ_FIRST(&pq[pqi].pl)) {
1774 				spin_lock(&pq[pqi].spin);
1775 				TAILQ_FOREACH(m, &pq[pqi].pl, pageq) {
1776 					if (spin_trylock(&m->spin) == 0)
1777 						continue;
1778 					KKASSERT(m->queue == basequeue + pqi);
1779 					_vm_page_rem_queue_spinlocked(m);
1780 					return(m);
1781 				}
1782 				spin_unlock(&pq[pqi].spin);
1783 			}
1784 		}
1785 	} while (pqmask != PQ_L2_MASK);
1786 
1787 	return(m);
1788 }
1789 
1790 /*
1791  * Returns a vm_page candidate for allocation.  The page is not busied so
1792  * it can move around.  The caller must busy the page (and typically
1793  * deactivate it if it cannot be busied!)
1794  *
1795  * Returns a spinlocked vm_page that has been removed from its queue.
1796  */
1797 vm_page_t
1798 vm_page_list_find(int basequeue, int index)
1799 {
1800 	return(_vm_page_list_find(basequeue, index));
1801 }
1802 
1803 /*
1804  * Find a page on the cache queue with color optimization, remove it
1805  * from the queue, and busy it.  The returned page will not be spinlocked.
1806  *
1807  * A candidate failure will be deactivated.  Candidates can fail due to
1808  * being busied by someone else, in which case they will be deactivated.
1809  *
1810  * This routine may not block.
1811  *
1812  */
1813 static vm_page_t
1814 vm_page_select_cache(u_short pg_color)
1815 {
1816 	vm_page_t m;
1817 
1818 	for (;;) {
1819 		m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1820 		if (m == NULL)
1821 			break;
1822 		/*
1823 		 * (m) has been removed from its queue and spinlocked
1824 		 */
1825 		if (vm_page_busy_try(m, TRUE)) {
1826 			_vm_page_deactivate_locked(m, 0);
1827 			vm_page_spin_unlock(m);
1828 		} else {
1829 			/*
1830 			 * We successfully busied the page
1831 			 */
1832 			if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1833 			    m->hold_count == 0 &&
1834 			    m->wire_count == 0 &&
1835 			    (m->dirty & m->valid) == 0) {
1836 				vm_page_spin_unlock(m);
1837 				pagedaemon_wakeup();
1838 				return(m);
1839 			}
1840 
1841 			/*
1842 			 * The page cannot be recycled, deactivate it.
1843 			 */
1844 			_vm_page_deactivate_locked(m, 0);
1845 			if (_vm_page_wakeup(m)) {
1846 				vm_page_spin_unlock(m);
1847 				wakeup(m);
1848 			} else {
1849 				vm_page_spin_unlock(m);
1850 			}
1851 		}
1852 	}
1853 	return (m);
1854 }
1855 
1856 /*
1857  * Find a free page.  We attempt to inline the nominal case and fall back
1858  * to _vm_page_select_free() otherwise.  A busied page is removed from
1859  * the queue and returned.
1860  *
1861  * This routine may not block.
1862  */
1863 static __inline vm_page_t
1864 vm_page_select_free(u_short pg_color)
1865 {
1866 	vm_page_t m;
1867 
1868 	for (;;) {
1869 		m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1870 		if (m == NULL)
1871 			break;
1872 		if (vm_page_busy_try(m, TRUE)) {
1873 			/*
1874 			 * Various mechanisms such as a pmap_collect can
1875 			 * result in a busy page on the free queue.  We
1876 			 * have to move the page out of the way so we can
1877 			 * retry the allocation.  If the other thread is not
1878 			 * allocating the page then m->valid will remain 0 and
1879 			 * the pageout daemon will free the page later on.
1880 			 *
1881 			 * Since we could not busy the page, however, we
1882 			 * cannot make assumptions as to whether the page
1883 			 * will be allocated by the other thread or not,
1884 			 * so all we can do is deactivate it to move it out
1885 			 * of the way.  In particular, if the other thread
1886 			 * wires the page it may wind up on the inactive
1887 			 * queue and the pageout daemon will have to deal
1888 			 * with that case too.
1889 			 */
1890 			_vm_page_deactivate_locked(m, 0);
1891 			vm_page_spin_unlock(m);
1892 		} else {
1893 			/*
1894 			 * Theoretically if we are able to busy the page
1895 			 * atomic with the queue removal (using the vm_page
1896 			 * lock) nobody else should be able to mess with the
1897 			 * page before us.
1898 			 */
1899 			KKASSERT((m->flags & (PG_UNMANAGED |
1900 					      PG_NEED_COMMIT)) == 0);
1901 			KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1902 						     "pg %p q=%d flags=%08x hold=%d wire=%d",
1903 						     m, m->queue, m->flags, m->hold_count, m->wire_count));
1904 			KKASSERT(m->wire_count == 0);
1905 			vm_page_spin_unlock(m);
1906 			pagedaemon_wakeup();
1907 
1908 			/* return busied and removed page */
1909 			return(m);
1910 		}
1911 	}
1912 	return(m);
1913 }
1914 
1915 /*
1916  * vm_page_alloc()
1917  *
1918  * Allocate and return a memory cell associated with this VM object/offset
1919  * pair.  If object is NULL an unassociated page will be allocated.
1920  *
1921  * The returned page will be busied and removed from its queues.  This
1922  * routine can block and may return NULL if a race occurs and the page
1923  * is found to already exist at the specified (object, pindex).
1924  *
1925  *	VM_ALLOC_NORMAL		allow use of cache pages, nominal free drain
1926  *	VM_ALLOC_QUICK		like normal but cannot use cache
1927  *	VM_ALLOC_SYSTEM		greater free drain
1928  *	VM_ALLOC_INTERRUPT	allow free list to be completely drained
1929  *	VM_ALLOC_ZERO		advisory request for pre-zero'd page only
1930  *	VM_ALLOC_FORCE_ZERO	advisory request for pre-zero'd page only
1931  *	VM_ALLOC_NULL_OK	ok to return NULL on insertion collision
1932  *				(see vm_page_grab())
1933  *	VM_ALLOC_USE_GD		ok to use per-gd cache
1934  *
1935  *	VM_ALLOC_CPU(n)		allocate using specified cpu localization
1936  *
1937  * The object must be held if not NULL
1938  * This routine may not block
1939  *
1940  * Additional special handling is required when called from an interrupt
1941  * (VM_ALLOC_INTERRUPT).  We are not allowed to mess with the page cache
1942  * in this case.
1943  */
1944 vm_page_t
1945 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1946 {
1947 	globaldata_t gd;
1948 	vm_object_t obj;
1949 	vm_page_t m;
1950 	u_short pg_color;
1951 	int cpuid_local;
1952 
1953 #if 0
1954 	/*
1955 	 * Special per-cpu free VM page cache.  The pages are pre-busied
1956 	 * and pre-zerod for us.
1957 	 */
1958 	if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1959 		crit_enter_gd(gd);
1960 		if (gd->gd_vmpg_count) {
1961 			m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1962 			crit_exit_gd(gd);
1963 			goto done;
1964                 }
1965 		crit_exit_gd(gd);
1966         }
1967 #endif
1968 	m = NULL;
1969 
1970 	/*
1971 	 * CPU LOCALIZATION
1972 	 *
1973 	 * CPU localization algorithm.  Break the page queues up by physical
1974 	 * id and core id (note that two cpu threads will have the same core
1975 	 * id, and core_id != gd_cpuid).
1976 	 *
1977 	 * This is nowhere near perfect, for example the last pindex in a
1978 	 * subgroup will overflow into the next cpu or package.  But this
1979 	 * should get us good page reuse locality in heavy mixed loads.
1980 	 *
1981 	 * (may be executed before the APs are started, so other GDs might
1982 	 *  not exist!)
1983 	 */
1984 	if (page_req & VM_ALLOC_CPU_SPEC)
1985 		cpuid_local = VM_ALLOC_GETCPU(page_req);
1986 	else
1987 		cpuid_local = mycpu->gd_cpuid;
1988 
1989 	pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1990 
1991 	KKASSERT(page_req &
1992 		(VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1993 		 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1994 
1995 	/*
1996 	 * Certain system threads (pageout daemon, buf_daemon's) are
1997 	 * allowed to eat deeper into the free page list.
1998 	 */
1999 	if (curthread->td_flags & TDF_SYSTHREAD)
2000 		page_req |= VM_ALLOC_SYSTEM;
2001 
2002 	/*
2003 	 * Impose various limitations.  Note that the v_free_reserved test
2004 	 * must match the opposite of vm_page_count_target() to avoid
2005 	 * livelocks, be careful.
2006 	 */
2007 loop:
2008 	gd = mycpu;
2009 	if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
2010 	    ((page_req & VM_ALLOC_INTERRUPT) &&
2011 	     gd->gd_vmstats.v_free_count > 0) ||
2012 	    ((page_req & VM_ALLOC_SYSTEM) &&
2013 	     gd->gd_vmstats.v_cache_count == 0 &&
2014 		gd->gd_vmstats.v_free_count >
2015 		gd->gd_vmstats.v_interrupt_free_min)
2016 	) {
2017 		/*
2018 		 * The free queue has sufficient free pages to take one out.
2019 		 */
2020 		m = vm_page_select_free(pg_color);
2021 	} else if (page_req & VM_ALLOC_NORMAL) {
2022 		/*
2023 		 * Allocatable from the cache (non-interrupt only).  On
2024 		 * success, we must free the page and try again, thus
2025 		 * ensuring that vmstats.v_*_free_min counters are replenished.
2026 		 */
2027 #ifdef INVARIANTS
2028 		if (curthread->td_preempted) {
2029 			kprintf("vm_page_alloc(): warning, attempt to allocate"
2030 				" cache page from preempting interrupt\n");
2031 			m = NULL;
2032 		} else {
2033 			m = vm_page_select_cache(pg_color);
2034 		}
2035 #else
2036 		m = vm_page_select_cache(pg_color);
2037 #endif
2038 		/*
2039 		 * On success move the page into the free queue and loop.
2040 		 *
2041 		 * Only do this if we can safely acquire the vm_object lock,
2042 		 * because this is effectively a random page and the caller
2043 		 * might be holding the lock shared, we don't want to
2044 		 * deadlock.
2045 		 */
2046 		if (m != NULL) {
2047 			KASSERT(m->dirty == 0,
2048 				("Found dirty cache page %p", m));
2049 			if ((obj = m->object) != NULL) {
2050 				if (vm_object_hold_try(obj)) {
2051 					vm_page_protect(m, VM_PROT_NONE);
2052 					vm_page_free(m);
2053 					/* m->object NULL here */
2054 					vm_object_drop(obj);
2055 				} else {
2056 					vm_page_deactivate(m);
2057 					vm_page_wakeup(m);
2058 				}
2059 			} else {
2060 				vm_page_protect(m, VM_PROT_NONE);
2061 				vm_page_free(m);
2062 			}
2063 			goto loop;
2064 		}
2065 
2066 		/*
2067 		 * On failure return NULL
2068 		 */
2069 		atomic_add_int(&vm_pageout_deficit, 1);
2070 		pagedaemon_wakeup();
2071 		return (NULL);
2072 	} else {
2073 		/*
2074 		 * No pages available, wakeup the pageout daemon and give up.
2075 		 */
2076 		atomic_add_int(&vm_pageout_deficit, 1);
2077 		pagedaemon_wakeup();
2078 		return (NULL);
2079 	}
2080 
2081 	/*
2082 	 * v_free_count can race so loop if we don't find the expected
2083 	 * page.
2084 	 */
2085 	if (m == NULL) {
2086 		vmstats_rollup();
2087 		goto loop;
2088 	}
2089 
2090 	/*
2091 	 * Good page found.  The page has already been busied for us and
2092 	 * removed from its queues.
2093 	 */
2094 	KASSERT(m->dirty == 0,
2095 		("vm_page_alloc: free/cache page %p was dirty", m));
2096 	KKASSERT(m->queue == PQ_NONE);
2097 
2098 #if 0
2099 done:
2100 #endif
2101 	/*
2102 	 * Initialize the structure, inheriting some flags but clearing
2103 	 * all the rest.  The page has already been busied for us.
2104 	 */
2105 	vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
2106 
2107 	KKASSERT(m->wire_count == 0);
2108 	KKASSERT((m->busy_count & PBUSY_MASK) == 0);
2109 	m->act_count = 0;
2110 	m->valid = 0;
2111 
2112 	/*
2113 	 * Caller must be holding the object lock (asserted by
2114 	 * vm_page_insert()).
2115 	 *
2116 	 * NOTE: Inserting a page here does not insert it into any pmaps
2117 	 *	 (which could cause us to block allocating memory).
2118 	 *
2119 	 * NOTE: If no object an unassociated page is allocated, m->pindex
2120 	 *	 can be used by the caller for any purpose.
2121 	 */
2122 	if (object) {
2123 		if (vm_page_insert(m, object, pindex) == FALSE) {
2124 			vm_page_free(m);
2125 			if ((page_req & VM_ALLOC_NULL_OK) == 0)
2126 				panic("PAGE RACE %p[%ld]/%p",
2127 				      object, (long)pindex, m);
2128 			m = NULL;
2129 		}
2130 	} else {
2131 		m->pindex = pindex;
2132 	}
2133 
2134 	/*
2135 	 * Don't wakeup too often - wakeup the pageout daemon when
2136 	 * we would be nearly out of memory.
2137 	 */
2138 	pagedaemon_wakeup();
2139 
2140 	/*
2141 	 * A BUSY page is returned.
2142 	 */
2143 	return (m);
2144 }
2145 
2146 /*
2147  * Returns number of pages available in our DMA memory reserve
2148  * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2149  */
2150 vm_size_t
2151 vm_contig_avail_pages(void)
2152 {
2153 	alist_blk_t blk;
2154 	alist_blk_t count;
2155 	alist_blk_t bfree;
2156 	spin_lock(&vm_contig_spin);
2157 	bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2158 	spin_unlock(&vm_contig_spin);
2159 
2160 	return bfree;
2161 }
2162 
2163 /*
2164  * Attempt to allocate contiguous physical memory with the specified
2165  * requirements.
2166  */
2167 vm_page_t
2168 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2169 		     unsigned long alignment, unsigned long boundary,
2170 		     unsigned long size, vm_memattr_t memattr)
2171 {
2172 	alist_blk_t blk;
2173 	vm_page_t m;
2174 	vm_pindex_t i;
2175 #if 0
2176 	static vm_pindex_t contig_rover;
2177 #endif
2178 
2179 	alignment >>= PAGE_SHIFT;
2180 	if (alignment == 0)
2181 		alignment = 1;
2182 	boundary >>= PAGE_SHIFT;
2183 	if (boundary == 0)
2184 		boundary = 1;
2185 	size = (size + PAGE_MASK) >> PAGE_SHIFT;
2186 
2187 #if 0
2188 	/*
2189 	 * Disabled temporarily until we find a solution for DRM (a flag
2190 	 * to always use the free space reserve, for performance).
2191 	 */
2192 	if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE &&
2193 	    boundary <= PAGE_SIZE && size == 1 &&
2194 	    memattr == VM_MEMATTR_DEFAULT) {
2195 		/*
2196 		 * Any page will work, use vm_page_alloc()
2197 		 * (e.g. when used from kmem_alloc_attr())
2198 		 */
2199 		m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF,
2200 				  VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2201 				  VM_ALLOC_INTERRUPT);
2202 		m->valid = VM_PAGE_BITS_ALL;
2203 		vm_page_wire(m);
2204 		vm_page_wakeup(m);
2205 	} else
2206 #endif
2207 	{
2208 		/*
2209 		 * Use the low-memory dma reserve
2210 		 */
2211 		spin_lock(&vm_contig_spin);
2212 		blk = alist_alloc(&vm_contig_alist, 0, size);
2213 		if (blk == ALIST_BLOCK_NONE) {
2214 			spin_unlock(&vm_contig_spin);
2215 			if (bootverbose) {
2216 				kprintf("vm_page_alloc_contig: %ldk nospace\n",
2217 					(size << PAGE_SHIFT) / 1024);
2218 				print_backtrace(5);
2219 			}
2220 			return(NULL);
2221 		}
2222 		if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2223 			alist_free(&vm_contig_alist, blk, size);
2224 			spin_unlock(&vm_contig_spin);
2225 			if (bootverbose) {
2226 				kprintf("vm_page_alloc_contig: %ldk high "
2227 					"%016jx failed\n",
2228 					(size << PAGE_SHIFT) / 1024,
2229 					(intmax_t)high);
2230 			}
2231 			return(NULL);
2232 		}
2233 		spin_unlock(&vm_contig_spin);
2234 		m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2235 	}
2236 	if (vm_contig_verbose) {
2237 		kprintf("vm_page_alloc_contig: %016jx/%ldk "
2238 			"(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n",
2239 			(intmax_t)m->phys_addr,
2240 			(size << PAGE_SHIFT) / 1024,
2241 			low, high, alignment, boundary, size, memattr);
2242 	}
2243 	if (memattr != VM_MEMATTR_DEFAULT) {
2244 		for (i = 0;i < size; i++)
2245 			pmap_page_set_memattr(&m[i], memattr);
2246 	}
2247 	return m;
2248 }
2249 
2250 /*
2251  * Free contiguously allocated pages.  The pages will be wired but not busy.
2252  * When freeing to the alist we leave them wired and not busy.
2253  */
2254 void
2255 vm_page_free_contig(vm_page_t m, unsigned long size)
2256 {
2257 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2258 	vm_pindex_t start = pa >> PAGE_SHIFT;
2259 	vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2260 
2261 	if (vm_contig_verbose) {
2262 		kprintf("vm_page_free_contig:  %016jx/%ldk\n",
2263 			(intmax_t)pa, size / 1024);
2264 	}
2265 	if (pa < vm_low_phys_reserved) {
2266 		KKASSERT(pa + size <= vm_low_phys_reserved);
2267 		spin_lock(&vm_contig_spin);
2268 		alist_free(&vm_contig_alist, start, pages);
2269 		spin_unlock(&vm_contig_spin);
2270 	} else {
2271 		while (pages) {
2272 			vm_page_busy_wait(m, FALSE, "cpgfr");
2273 			vm_page_unwire(m, 0);
2274 			vm_page_free(m);
2275 			--pages;
2276 			++m;
2277 		}
2278 
2279 	}
2280 }
2281 
2282 
2283 /*
2284  * Wait for sufficient free memory for nominal heavy memory use kernel
2285  * operations.
2286  *
2287  * WARNING!  Be sure never to call this in any vm_pageout code path, which
2288  *	     will trivially deadlock the system.
2289  */
2290 void
2291 vm_wait_nominal(void)
2292 {
2293 	while (vm_page_count_min(0))
2294 		vm_wait(0);
2295 }
2296 
2297 /*
2298  * Test if vm_wait_nominal() would block.
2299  */
2300 int
2301 vm_test_nominal(void)
2302 {
2303 	if (vm_page_count_min(0))
2304 		return(1);
2305 	return(0);
2306 }
2307 
2308 /*
2309  * Block until free pages are available for allocation, called in various
2310  * places before memory allocations.
2311  *
2312  * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2313  * more generous then that.
2314  */
2315 void
2316 vm_wait(int timo)
2317 {
2318 	/*
2319 	 * never wait forever
2320 	 */
2321 	if (timo == 0)
2322 		timo = hz;
2323 	lwkt_gettoken(&vm_token);
2324 
2325 	if (curthread == pagethread ||
2326 	    curthread == emergpager) {
2327 		/*
2328 		 * The pageout daemon itself needs pages, this is bad.
2329 		 */
2330 		if (vm_page_count_min(0)) {
2331 			vm_pageout_pages_needed = 1;
2332 			tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2333 		}
2334 	} else {
2335 		/*
2336 		 * Wakeup the pageout daemon if necessary and wait.
2337 		 *
2338 		 * Do not wait indefinitely for the target to be reached,
2339 		 * as load might prevent it from being reached any time soon.
2340 		 * But wait a little to try to slow down page allocations
2341 		 * and to give more important threads (the pagedaemon)
2342 		 * allocation priority.
2343 		 */
2344 		if (vm_page_count_target()) {
2345 			if (vm_pages_needed == 0) {
2346 				vm_pages_needed = 1;
2347 				wakeup(&vm_pages_needed);
2348 			}
2349 			++vm_pages_waiting;	/* SMP race ok */
2350 			tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2351 		}
2352 	}
2353 	lwkt_reltoken(&vm_token);
2354 }
2355 
2356 /*
2357  * Block until free pages are available for allocation
2358  *
2359  * Called only from vm_fault so that processes page faulting can be
2360  * easily tracked.
2361  */
2362 void
2363 vm_wait_pfault(void)
2364 {
2365 	/*
2366 	 * Wakeup the pageout daemon if necessary and wait.
2367 	 *
2368 	 * Do not wait indefinitely for the target to be reached,
2369 	 * as load might prevent it from being reached any time soon.
2370 	 * But wait a little to try to slow down page allocations
2371 	 * and to give more important threads (the pagedaemon)
2372 	 * allocation priority.
2373 	 */
2374 	if (vm_page_count_min(0)) {
2375 		lwkt_gettoken(&vm_token);
2376 		while (vm_page_count_severe()) {
2377 			if (vm_page_count_target()) {
2378 				thread_t td;
2379 
2380 				if (vm_pages_needed == 0) {
2381 					vm_pages_needed = 1;
2382 					wakeup(&vm_pages_needed);
2383 				}
2384 				++vm_pages_waiting;	/* SMP race ok */
2385 				tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2386 
2387 				/*
2388 				 * Do not stay stuck in the loop if the system is trying
2389 				 * to kill the process.
2390 				 */
2391 				td = curthread;
2392 				if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2393 					break;
2394 			}
2395 		}
2396 		lwkt_reltoken(&vm_token);
2397 	}
2398 }
2399 
2400 /*
2401  * Put the specified page on the active list (if appropriate).  Ensure
2402  * that act_count is at least ACT_INIT but do not otherwise mess with it.
2403  *
2404  * The caller should be holding the page busied ? XXX
2405  * This routine may not block.
2406  */
2407 void
2408 vm_page_activate(vm_page_t m)
2409 {
2410 	u_short oqueue;
2411 
2412 	vm_page_spin_lock(m);
2413 	if (m->queue - m->pc != PQ_ACTIVE) {
2414 		_vm_page_queue_spin_lock(m);
2415 		oqueue = _vm_page_rem_queue_spinlocked(m);
2416 		/* page is left spinlocked, queue is unlocked */
2417 
2418 		if (oqueue == PQ_CACHE)
2419 			mycpu->gd_cnt.v_reactivated++;
2420 		if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2421 			if (m->act_count < ACT_INIT)
2422 				m->act_count = ACT_INIT;
2423 			_vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2424 		}
2425 		_vm_page_and_queue_spin_unlock(m);
2426 		if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2427 			pagedaemon_wakeup();
2428 	} else {
2429 		if (m->act_count < ACT_INIT)
2430 			m->act_count = ACT_INIT;
2431 		vm_page_spin_unlock(m);
2432 	}
2433 }
2434 
2435 /*
2436  * Helper routine for vm_page_free_toq() and vm_page_cache().  This
2437  * routine is called when a page has been added to the cache or free
2438  * queues.
2439  *
2440  * This routine may not block.
2441  */
2442 static __inline void
2443 vm_page_free_wakeup(void)
2444 {
2445 	globaldata_t gd = mycpu;
2446 
2447 	/*
2448 	 * If the pageout daemon itself needs pages, then tell it that
2449 	 * there are some free.
2450 	 */
2451 	if (vm_pageout_pages_needed &&
2452 	    gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2453 	    gd->gd_vmstats.v_pageout_free_min
2454 	) {
2455 		vm_pageout_pages_needed = 0;
2456 		wakeup(&vm_pageout_pages_needed);
2457 	}
2458 
2459 	/*
2460 	 * Wakeup processes that are waiting on memory.
2461 	 *
2462 	 * Generally speaking we want to wakeup stuck processes as soon as
2463 	 * possible.  !vm_page_count_min(0) is the absolute minimum point
2464 	 * where we can do this.  Wait a bit longer to reduce degenerate
2465 	 * re-blocking (vm_page_free_hysteresis).  The target check is just
2466 	 * to make sure the min-check w/hysteresis does not exceed the
2467 	 * normal target.
2468 	 */
2469 	if (vm_pages_waiting) {
2470 		if (!vm_page_count_min(vm_page_free_hysteresis) ||
2471 		    !vm_page_count_target()) {
2472 			vm_pages_waiting = 0;
2473 			wakeup(&vmstats.v_free_count);
2474 			++mycpu->gd_cnt.v_ppwakeups;
2475 		}
2476 #if 0
2477 		if (!vm_page_count_target()) {
2478 			/*
2479 			 * Plenty of pages are free, wakeup everyone.
2480 			 */
2481 			vm_pages_waiting = 0;
2482 			wakeup(&vmstats.v_free_count);
2483 			++mycpu->gd_cnt.v_ppwakeups;
2484 		} else if (!vm_page_count_min(0)) {
2485 			/*
2486 			 * Some pages are free, wakeup someone.
2487 			 */
2488 			int wcount = vm_pages_waiting;
2489 			if (wcount > 0)
2490 				--wcount;
2491 			vm_pages_waiting = wcount;
2492 			wakeup_one(&vmstats.v_free_count);
2493 			++mycpu->gd_cnt.v_ppwakeups;
2494 		}
2495 #endif
2496 	}
2497 }
2498 
2499 /*
2500  * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2501  * it from its VM object.
2502  *
2503  * The vm_page must be BUSY on entry.  BUSY will be released on
2504  * return (the page will have been freed).
2505  */
2506 void
2507 vm_page_free_toq(vm_page_t m)
2508 {
2509 	mycpu->gd_cnt.v_tfree++;
2510 	KKASSERT((m->flags & PG_MAPPED) == 0);
2511 	KKASSERT(m->busy_count & PBUSY_LOCKED);
2512 
2513 	if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) {
2514 		kprintf("vm_page_free: pindex(%lu), busy %08x, "
2515 			"hold(%d)\n",
2516 			(u_long)m->pindex, m->busy_count, m->hold_count);
2517 		if ((m->queue - m->pc) == PQ_FREE)
2518 			panic("vm_page_free: freeing free page");
2519 		else
2520 			panic("vm_page_free: freeing busy page");
2521 	}
2522 
2523 	/*
2524 	 * Remove from object, spinlock the page and its queues and
2525 	 * remove from any queue.  No queue spinlock will be held
2526 	 * after this section (because the page was removed from any
2527 	 * queue).
2528 	 */
2529 	vm_page_remove(m);
2530 	vm_page_and_queue_spin_lock(m);
2531 	_vm_page_rem_queue_spinlocked(m);
2532 
2533 	/*
2534 	 * No further management of fictitious pages occurs beyond object
2535 	 * and queue removal.
2536 	 */
2537 	if ((m->flags & PG_FICTITIOUS) != 0) {
2538 		vm_page_spin_unlock(m);
2539 		vm_page_wakeup(m);
2540 		return;
2541 	}
2542 
2543 	m->valid = 0;
2544 	vm_page_undirty(m);
2545 
2546 	if (m->wire_count != 0) {
2547 		if (m->wire_count > 1) {
2548 		    panic(
2549 			"vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2550 			m->wire_count, (long)m->pindex);
2551 		}
2552 		panic("vm_page_free: freeing wired page");
2553 	}
2554 
2555 	/*
2556 	 * Clear the UNMANAGED flag when freeing an unmanaged page.
2557 	 * Clear the NEED_COMMIT flag
2558 	 */
2559 	if (m->flags & PG_UNMANAGED)
2560 		vm_page_flag_clear(m, PG_UNMANAGED);
2561 	if (m->flags & PG_NEED_COMMIT)
2562 		vm_page_flag_clear(m, PG_NEED_COMMIT);
2563 
2564 	if (m->hold_count != 0) {
2565 		_vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2566 	} else {
2567 		_vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2568 	}
2569 
2570 	/*
2571 	 * This sequence allows us to clear BUSY while still holding
2572 	 * its spin lock, which reduces contention vs allocators.  We
2573 	 * must not leave the queue locked or _vm_page_wakeup() may
2574 	 * deadlock.
2575 	 */
2576 	_vm_page_queue_spin_unlock(m);
2577 	if (_vm_page_wakeup(m)) {
2578 		vm_page_spin_unlock(m);
2579 		wakeup(m);
2580 	} else {
2581 		vm_page_spin_unlock(m);
2582 	}
2583 	vm_page_free_wakeup();
2584 }
2585 
2586 /*
2587  * vm_page_unmanage()
2588  *
2589  * Prevent PV management from being done on the page.  The page is
2590  * removed from the paging queues as if it were wired, and as a
2591  * consequence of no longer being managed the pageout daemon will not
2592  * touch it (since there is no way to locate the pte mappings for the
2593  * page).  madvise() calls that mess with the pmap will also no longer
2594  * operate on the page.
2595  *
2596  * Beyond that the page is still reasonably 'normal'.  Freeing the page
2597  * will clear the flag.
2598  *
2599  * This routine is used by OBJT_PHYS objects - objects using unswappable
2600  * physical memory as backing store rather then swap-backed memory and
2601  * will eventually be extended to support 4MB unmanaged physical
2602  * mappings.
2603  *
2604  * Caller must be holding the page busy.
2605  */
2606 void
2607 vm_page_unmanage(vm_page_t m)
2608 {
2609 	KKASSERT(m->busy_count & PBUSY_LOCKED);
2610 	if ((m->flags & PG_UNMANAGED) == 0) {
2611 		if (m->wire_count == 0)
2612 			vm_page_unqueue(m);
2613 	}
2614 	vm_page_flag_set(m, PG_UNMANAGED);
2615 }
2616 
2617 /*
2618  * Mark this page as wired down by yet another map, removing it from
2619  * paging queues as necessary.
2620  *
2621  * Caller must be holding the page busy.
2622  */
2623 void
2624 vm_page_wire(vm_page_t m)
2625 {
2626 	/*
2627 	 * Only bump the wire statistics if the page is not already wired,
2628 	 * and only unqueue the page if it is on some queue (if it is unmanaged
2629 	 * it is already off the queues).  Don't do anything with fictitious
2630 	 * pages because they are always wired.
2631 	 */
2632 	KKASSERT(m->busy_count & PBUSY_LOCKED);
2633 	if ((m->flags & PG_FICTITIOUS) == 0) {
2634 		if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2635 			if ((m->flags & PG_UNMANAGED) == 0)
2636 				vm_page_unqueue(m);
2637 			atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2638 		}
2639 		KASSERT(m->wire_count != 0,
2640 			("vm_page_wire: wire_count overflow m=%p", m));
2641 	}
2642 }
2643 
2644 /*
2645  * Release one wiring of this page, potentially enabling it to be paged again.
2646  *
2647  * Many pages placed on the inactive queue should actually go
2648  * into the cache, but it is difficult to figure out which.  What
2649  * we do instead, if the inactive target is well met, is to put
2650  * clean pages at the head of the inactive queue instead of the tail.
2651  * This will cause them to be moved to the cache more quickly and
2652  * if not actively re-referenced, freed more quickly.  If we just
2653  * stick these pages at the end of the inactive queue, heavy filesystem
2654  * meta-data accesses can cause an unnecessary paging load on memory bound
2655  * processes.  This optimization causes one-time-use metadata to be
2656  * reused more quickly.
2657  *
2658  * Pages marked PG_NEED_COMMIT are always activated and never placed on
2659  * the inactive queue.  This helps the pageout daemon determine memory
2660  * pressure and act on out-of-memory situations more quickly.
2661  *
2662  * BUT, if we are in a low-memory situation we have no choice but to
2663  * put clean pages on the cache queue.
2664  *
2665  * A number of routines use vm_page_unwire() to guarantee that the page
2666  * will go into either the inactive or active queues, and will NEVER
2667  * be placed in the cache - for example, just after dirtying a page.
2668  * dirty pages in the cache are not allowed.
2669  *
2670  * This routine may not block.
2671  */
2672 void
2673 vm_page_unwire(vm_page_t m, int activate)
2674 {
2675 	KKASSERT(m->busy_count & PBUSY_LOCKED);
2676 	if (m->flags & PG_FICTITIOUS) {
2677 		/* do nothing */
2678 	} else if (m->wire_count <= 0) {
2679 		panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2680 	} else {
2681 		if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2682 			atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1);
2683 			if (m->flags & PG_UNMANAGED) {
2684 				;
2685 			} else if (activate || (m->flags & PG_NEED_COMMIT)) {
2686 				vm_page_spin_lock(m);
2687 				_vm_page_add_queue_spinlocked(m,
2688 							PQ_ACTIVE + m->pc, 0);
2689 				_vm_page_and_queue_spin_unlock(m);
2690 			} else {
2691 				vm_page_spin_lock(m);
2692 				vm_page_flag_clear(m, PG_WINATCFLS);
2693 				_vm_page_add_queue_spinlocked(m,
2694 							PQ_INACTIVE + m->pc, 0);
2695 				++vm_swapcache_inactive_heuristic;
2696 				_vm_page_and_queue_spin_unlock(m);
2697 			}
2698 		}
2699 	}
2700 }
2701 
2702 /*
2703  * Move the specified page to the inactive queue.  If the page has
2704  * any associated swap, the swap is deallocated.
2705  *
2706  * Normally athead is 0 resulting in LRU operation.  athead is set
2707  * to 1 if we want this page to be 'as if it were placed in the cache',
2708  * except without unmapping it from the process address space.
2709  *
2710  * vm_page's spinlock must be held on entry and will remain held on return.
2711  * This routine may not block.
2712  */
2713 static void
2714 _vm_page_deactivate_locked(vm_page_t m, int athead)
2715 {
2716 	u_short oqueue;
2717 
2718 	/*
2719 	 * Ignore if already inactive.
2720 	 */
2721 	if (m->queue - m->pc == PQ_INACTIVE)
2722 		return;
2723 	_vm_page_queue_spin_lock(m);
2724 	oqueue = _vm_page_rem_queue_spinlocked(m);
2725 
2726 	if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2727 		if (oqueue == PQ_CACHE)
2728 			mycpu->gd_cnt.v_reactivated++;
2729 		vm_page_flag_clear(m, PG_WINATCFLS);
2730 		_vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2731 		if (athead == 0)
2732 			++vm_swapcache_inactive_heuristic;
2733 	}
2734 	/* NOTE: PQ_NONE if condition not taken */
2735 	_vm_page_queue_spin_unlock(m);
2736 	/* leaves vm_page spinlocked */
2737 }
2738 
2739 /*
2740  * Attempt to deactivate a page.
2741  *
2742  * No requirements.
2743  */
2744 void
2745 vm_page_deactivate(vm_page_t m)
2746 {
2747 	vm_page_spin_lock(m);
2748 	_vm_page_deactivate_locked(m, 0);
2749 	vm_page_spin_unlock(m);
2750 }
2751 
2752 void
2753 vm_page_deactivate_locked(vm_page_t m)
2754 {
2755 	_vm_page_deactivate_locked(m, 0);
2756 }
2757 
2758 /*
2759  * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2760  *
2761  * This function returns non-zero if it successfully moved the page to
2762  * PQ_CACHE.
2763  *
2764  * This function unconditionally unbusies the page on return.
2765  */
2766 int
2767 vm_page_try_to_cache(vm_page_t m)
2768 {
2769 	vm_page_spin_lock(m);
2770 	if (m->dirty || m->hold_count || m->wire_count ||
2771 	    (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2772 		if (_vm_page_wakeup(m)) {
2773 			vm_page_spin_unlock(m);
2774 			wakeup(m);
2775 		} else {
2776 			vm_page_spin_unlock(m);
2777 		}
2778 		return(0);
2779 	}
2780 	vm_page_spin_unlock(m);
2781 
2782 	/*
2783 	 * Page busied by us and no longer spinlocked.  Dirty pages cannot
2784 	 * be moved to the cache.
2785 	 */
2786 	vm_page_test_dirty(m);
2787 	if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2788 		vm_page_wakeup(m);
2789 		return(0);
2790 	}
2791 	vm_page_cache(m);
2792 	return(1);
2793 }
2794 
2795 /*
2796  * Attempt to free the page.  If we cannot free it, we do nothing.
2797  * 1 is returned on success, 0 on failure.
2798  *
2799  * No requirements.
2800  */
2801 int
2802 vm_page_try_to_free(vm_page_t m)
2803 {
2804 	vm_page_spin_lock(m);
2805 	if (vm_page_busy_try(m, TRUE)) {
2806 		vm_page_spin_unlock(m);
2807 		return(0);
2808 	}
2809 
2810 	/*
2811 	 * The page can be in any state, including already being on the free
2812 	 * queue.  Check to see if it really can be freed.
2813 	 */
2814 	if (m->dirty ||				/* can't free if it is dirty */
2815 	    m->hold_count ||			/* or held (XXX may be wrong) */
2816 	    m->wire_count ||			/* or wired */
2817 	    (m->flags & (PG_UNMANAGED |		/* or unmanaged */
2818 			 PG_NEED_COMMIT)) ||	/* or needs a commit */
2819 	    m->queue - m->pc == PQ_FREE ||	/* already on PQ_FREE */
2820 	    m->queue - m->pc == PQ_HOLD) {	/* already on PQ_HOLD */
2821 		if (_vm_page_wakeup(m)) {
2822 			vm_page_spin_unlock(m);
2823 			wakeup(m);
2824 		} else {
2825 			vm_page_spin_unlock(m);
2826 		}
2827 		return(0);
2828 	}
2829 	vm_page_spin_unlock(m);
2830 
2831 	/*
2832 	 * We can probably free the page.
2833 	 *
2834 	 * Page busied by us and no longer spinlocked.  Dirty pages will
2835 	 * not be freed by this function.    We have to re-test the
2836 	 * dirty bit after cleaning out the pmaps.
2837 	 */
2838 	vm_page_test_dirty(m);
2839 	if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2840 		vm_page_wakeup(m);
2841 		return(0);
2842 	}
2843 	vm_page_protect(m, VM_PROT_NONE);
2844 	if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2845 		vm_page_wakeup(m);
2846 		return(0);
2847 	}
2848 	vm_page_free(m);
2849 	return(1);
2850 }
2851 
2852 /*
2853  * vm_page_cache
2854  *
2855  * Put the specified page onto the page cache queue (if appropriate).
2856  *
2857  * The page must be busy, and this routine will release the busy and
2858  * possibly even free the page.
2859  */
2860 void
2861 vm_page_cache(vm_page_t m)
2862 {
2863 	/*
2864 	 * Not suitable for the cache
2865 	 */
2866 	if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2867 	    (m->busy_count & PBUSY_MASK) ||
2868 	    m->wire_count || m->hold_count) {
2869 		vm_page_wakeup(m);
2870 		return;
2871 	}
2872 
2873 	/*
2874 	 * Already in the cache (and thus not mapped)
2875 	 */
2876 	if ((m->queue - m->pc) == PQ_CACHE) {
2877 		KKASSERT((m->flags & PG_MAPPED) == 0);
2878 		vm_page_wakeup(m);
2879 		return;
2880 	}
2881 
2882 	/*
2883 	 * Caller is required to test m->dirty, but note that the act of
2884 	 * removing the page from its maps can cause it to become dirty
2885 	 * on an SMP system due to another cpu running in usermode.
2886 	 */
2887 	if (m->dirty) {
2888 		panic("vm_page_cache: caching a dirty page, pindex: %ld",
2889 			(long)m->pindex);
2890 	}
2891 
2892 	/*
2893 	 * Remove all pmaps and indicate that the page is not
2894 	 * writeable or mapped.  Our vm_page_protect() call may
2895 	 * have blocked (especially w/ VM_PROT_NONE), so recheck
2896 	 * everything.
2897 	 */
2898 	vm_page_protect(m, VM_PROT_NONE);
2899 	if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2900 	    (m->busy_count & PBUSY_MASK) ||
2901 	    m->wire_count || m->hold_count) {
2902 		vm_page_wakeup(m);
2903 	} else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2904 		vm_page_deactivate(m);
2905 		vm_page_wakeup(m);
2906 	} else {
2907 		_vm_page_and_queue_spin_lock(m);
2908 		_vm_page_rem_queue_spinlocked(m);
2909 		_vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2910 		_vm_page_queue_spin_unlock(m);
2911 		if (_vm_page_wakeup(m)) {
2912 			vm_page_spin_unlock(m);
2913 			wakeup(m);
2914 		} else {
2915 			vm_page_spin_unlock(m);
2916 		}
2917 		vm_page_free_wakeup();
2918 	}
2919 }
2920 
2921 /*
2922  * vm_page_dontneed()
2923  *
2924  * Cache, deactivate, or do nothing as appropriate.  This routine
2925  * is typically used by madvise() MADV_DONTNEED.
2926  *
2927  * Generally speaking we want to move the page into the cache so
2928  * it gets reused quickly.  However, this can result in a silly syndrome
2929  * due to the page recycling too quickly.  Small objects will not be
2930  * fully cached.  On the otherhand, if we move the page to the inactive
2931  * queue we wind up with a problem whereby very large objects
2932  * unnecessarily blow away our inactive and cache queues.
2933  *
2934  * The solution is to move the pages based on a fixed weighting.  We
2935  * either leave them alone, deactivate them, or move them to the cache,
2936  * where moving them to the cache has the highest weighting.
2937  * By forcing some pages into other queues we eventually force the
2938  * system to balance the queues, potentially recovering other unrelated
2939  * space from active.  The idea is to not force this to happen too
2940  * often.
2941  *
2942  * The page must be busied.
2943  */
2944 void
2945 vm_page_dontneed(vm_page_t m)
2946 {
2947 	static int dnweight;
2948 	int dnw;
2949 	int head;
2950 
2951 	dnw = ++dnweight;
2952 
2953 	/*
2954 	 * occassionally leave the page alone
2955 	 */
2956 	if ((dnw & 0x01F0) == 0 ||
2957 	    m->queue - m->pc == PQ_INACTIVE ||
2958 	    m->queue - m->pc == PQ_CACHE
2959 	) {
2960 		if (m->act_count >= ACT_INIT)
2961 			--m->act_count;
2962 		return;
2963 	}
2964 
2965 	/*
2966 	 * If vm_page_dontneed() is inactivating a page, it must clear
2967 	 * the referenced flag; otherwise the pagedaemon will see references
2968 	 * on the page in the inactive queue and reactivate it. Until the
2969 	 * page can move to the cache queue, madvise's job is not done.
2970 	 */
2971 	vm_page_flag_clear(m, PG_REFERENCED);
2972 	pmap_clear_reference(m);
2973 
2974 	if (m->dirty == 0)
2975 		vm_page_test_dirty(m);
2976 
2977 	if (m->dirty || (dnw & 0x0070) == 0) {
2978 		/*
2979 		 * Deactivate the page 3 times out of 32.
2980 		 */
2981 		head = 0;
2982 	} else {
2983 		/*
2984 		 * Cache the page 28 times out of every 32.  Note that
2985 		 * the page is deactivated instead of cached, but placed
2986 		 * at the head of the queue instead of the tail.
2987 		 */
2988 		head = 1;
2989 	}
2990 	vm_page_spin_lock(m);
2991 	_vm_page_deactivate_locked(m, head);
2992 	vm_page_spin_unlock(m);
2993 }
2994 
2995 /*
2996  * These routines manipulate the 'soft busy' count for a page.  A soft busy
2997  * is almost like a hard BUSY except that it allows certain compatible
2998  * operations to occur on the page while it is busy.  For example, a page
2999  * undergoing a write can still be mapped read-only.
3000  *
3001  * We also use soft-busy to quickly pmap_enter shared read-only pages
3002  * without having to hold the page locked.
3003  *
3004  * The soft-busy count can be > 1 in situations where multiple threads
3005  * are pmap_enter()ing the same page simultaneously, or when two buffer
3006  * cache buffers overlap the same page.
3007  *
3008  * The caller must hold the page BUSY when making these two calls.
3009  */
3010 void
3011 vm_page_io_start(vm_page_t m)
3012 {
3013 	uint32_t ocount;
3014 
3015 	ocount = atomic_fetchadd_int(&m->busy_count, 1);
3016 	KKASSERT(ocount & PBUSY_LOCKED);
3017 }
3018 
3019 void
3020 vm_page_io_finish(vm_page_t m)
3021 {
3022 	uint32_t ocount;
3023 
3024 	ocount = atomic_fetchadd_int(&m->busy_count, -1);
3025 	KKASSERT(ocount & PBUSY_MASK);
3026 #if 0
3027 	if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0)
3028 		wakeup(m);
3029 #endif
3030 }
3031 
3032 /*
3033  * Attempt to soft-busy a page.  The page must not be PBUSY_LOCKED.
3034  *
3035  * We can't use fetchadd here because we might race a hard-busy and the
3036  * page freeing code asserts on a non-zero soft-busy count (even if only
3037  * temporary).
3038  *
3039  * Returns 0 on success, non-zero on failure.
3040  */
3041 int
3042 vm_page_sbusy_try(vm_page_t m)
3043 {
3044 	uint32_t ocount;
3045 
3046 	for (;;) {
3047 		ocount = m->busy_count;
3048 		cpu_ccfence();
3049 		if (ocount & PBUSY_LOCKED)
3050 			return 1;
3051 		if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1))
3052 			break;
3053 	}
3054 	return 0;
3055 #if 0
3056 	if (m->busy_count & PBUSY_LOCKED)
3057 		return 1;
3058 	ocount = atomic_fetchadd_int(&m->busy_count, 1);
3059 	if (ocount & PBUSY_LOCKED) {
3060 		vm_page_sbusy_drop(m);
3061 		return 1;
3062 	}
3063 	return 0;
3064 #endif
3065 }
3066 
3067 /*
3068  * Indicate that a clean VM page requires a filesystem commit and cannot
3069  * be reused.  Used by tmpfs.
3070  */
3071 void
3072 vm_page_need_commit(vm_page_t m)
3073 {
3074 	vm_page_flag_set(m, PG_NEED_COMMIT);
3075 	vm_object_set_writeable_dirty(m->object);
3076 }
3077 
3078 void
3079 vm_page_clear_commit(vm_page_t m)
3080 {
3081 	vm_page_flag_clear(m, PG_NEED_COMMIT);
3082 }
3083 
3084 /*
3085  * Grab a page, blocking if it is busy and allocating a page if necessary.
3086  * A busy page is returned or NULL.  The page may or may not be valid and
3087  * might not be on a queue (the caller is responsible for the disposition of
3088  * the page).
3089  *
3090  * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
3091  * page will be zero'd and marked valid.
3092  *
3093  * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
3094  * valid even if it already exists.
3095  *
3096  * If VM_ALLOC_RETRY is specified this routine will never return NULL.  Also
3097  * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
3098  * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
3099  *
3100  * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
3101  * always returned if we had blocked.
3102  *
3103  * This routine may not be called from an interrupt.
3104  *
3105  * No other requirements.
3106  */
3107 vm_page_t
3108 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3109 {
3110 	vm_page_t m;
3111 	int error;
3112 	int shared = 1;
3113 
3114 	KKASSERT(allocflags &
3115 		(VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
3116 	vm_object_hold_shared(object);
3117 	for (;;) {
3118 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
3119 		if (error) {
3120 			vm_page_sleep_busy(m, TRUE, "pgrbwt");
3121 			if ((allocflags & VM_ALLOC_RETRY) == 0) {
3122 				m = NULL;
3123 				break;
3124 			}
3125 			/* retry */
3126 		} else if (m == NULL) {
3127 			if (shared) {
3128 				vm_object_upgrade(object);
3129 				shared = 0;
3130 			}
3131 			if (allocflags & VM_ALLOC_RETRY)
3132 				allocflags |= VM_ALLOC_NULL_OK;
3133 			m = vm_page_alloc(object, pindex,
3134 					  allocflags & ~VM_ALLOC_RETRY);
3135 			if (m)
3136 				break;
3137 			vm_wait(0);
3138 			if ((allocflags & VM_ALLOC_RETRY) == 0)
3139 				goto failed;
3140 		} else {
3141 			/* m found */
3142 			break;
3143 		}
3144 	}
3145 
3146 	/*
3147 	 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
3148 	 *
3149 	 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
3150 	 * valid even if already valid.
3151 	 *
3152 	 * NOTE!  We have removed all of the PG_ZERO optimizations and also
3153 	 *	  removed the idle zeroing code.  These optimizations actually
3154 	 *	  slow things down on modern cpus because the zerod area is
3155 	 *	  likely uncached, placing a memory-access burden on the
3156 	 *	  accesors taking the fault.
3157 	 *
3158 	 *	  By always zeroing the page in-line with the fault, no
3159 	 *	  dynamic ram reads are needed and the caches are hot, ready
3160 	 *	  for userland to access the memory.
3161 	 */
3162 	if (m->valid == 0) {
3163 		if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
3164 			pmap_zero_page(VM_PAGE_TO_PHYS(m));
3165 			m->valid = VM_PAGE_BITS_ALL;
3166 		}
3167 	} else if (allocflags & VM_ALLOC_FORCE_ZERO) {
3168 		pmap_zero_page(VM_PAGE_TO_PHYS(m));
3169 		m->valid = VM_PAGE_BITS_ALL;
3170 	}
3171 failed:
3172 	vm_object_drop(object);
3173 	return(m);
3174 }
3175 
3176 /*
3177  * Mapping function for valid bits or for dirty bits in
3178  * a page.  May not block.
3179  *
3180  * Inputs are required to range within a page.
3181  *
3182  * No requirements.
3183  * Non blocking.
3184  */
3185 int
3186 vm_page_bits(int base, int size)
3187 {
3188 	int first_bit;
3189 	int last_bit;
3190 
3191 	KASSERT(
3192 	    base + size <= PAGE_SIZE,
3193 	    ("vm_page_bits: illegal base/size %d/%d", base, size)
3194 	);
3195 
3196 	if (size == 0)		/* handle degenerate case */
3197 		return(0);
3198 
3199 	first_bit = base >> DEV_BSHIFT;
3200 	last_bit = (base + size - 1) >> DEV_BSHIFT;
3201 
3202 	return ((2 << last_bit) - (1 << first_bit));
3203 }
3204 
3205 /*
3206  * Sets portions of a page valid and clean.  The arguments are expected
3207  * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3208  * of any partial chunks touched by the range.  The invalid portion of
3209  * such chunks will be zero'd.
3210  *
3211  * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
3212  *	 align base to DEV_BSIZE so as not to mark clean a partially
3213  *	 truncated device block.  Otherwise the dirty page status might be
3214  *	 lost.
3215  *
3216  * This routine may not block.
3217  *
3218  * (base + size) must be less then or equal to PAGE_SIZE.
3219  */
3220 static void
3221 _vm_page_zero_valid(vm_page_t m, int base, int size)
3222 {
3223 	int frag;
3224 	int endoff;
3225 
3226 	if (size == 0)	/* handle degenerate case */
3227 		return;
3228 
3229 	/*
3230 	 * If the base is not DEV_BSIZE aligned and the valid
3231 	 * bit is clear, we have to zero out a portion of the
3232 	 * first block.
3233 	 */
3234 
3235 	if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3236 	    (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3237 	) {
3238 		pmap_zero_page_area(
3239 		    VM_PAGE_TO_PHYS(m),
3240 		    frag,
3241 		    base - frag
3242 		);
3243 	}
3244 
3245 	/*
3246 	 * If the ending offset is not DEV_BSIZE aligned and the
3247 	 * valid bit is clear, we have to zero out a portion of
3248 	 * the last block.
3249 	 */
3250 
3251 	endoff = base + size;
3252 
3253 	if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3254 	    (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3255 	) {
3256 		pmap_zero_page_area(
3257 		    VM_PAGE_TO_PHYS(m),
3258 		    endoff,
3259 		    DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3260 		);
3261 	}
3262 }
3263 
3264 /*
3265  * Set valid, clear dirty bits.  If validating the entire
3266  * page we can safely clear the pmap modify bit.  We also
3267  * use this opportunity to clear the PG_NOSYNC flag.  If a process
3268  * takes a write fault on a MAP_NOSYNC memory area the flag will
3269  * be set again.
3270  *
3271  * We set valid bits inclusive of any overlap, but we can only
3272  * clear dirty bits for DEV_BSIZE chunks that are fully within
3273  * the range.
3274  *
3275  * Page must be busied?
3276  * No other requirements.
3277  */
3278 void
3279 vm_page_set_valid(vm_page_t m, int base, int size)
3280 {
3281 	_vm_page_zero_valid(m, base, size);
3282 	m->valid |= vm_page_bits(base, size);
3283 }
3284 
3285 
3286 /*
3287  * Set valid bits and clear dirty bits.
3288  *
3289  * Page must be busied by caller.
3290  *
3291  * NOTE: This function does not clear the pmap modified bit.
3292  *	 Also note that e.g. NFS may use a byte-granular base
3293  *	 and size.
3294  *
3295  * No other requirements.
3296  */
3297 void
3298 vm_page_set_validclean(vm_page_t m, int base, int size)
3299 {
3300 	int pagebits;
3301 
3302 	_vm_page_zero_valid(m, base, size);
3303 	pagebits = vm_page_bits(base, size);
3304 	m->valid |= pagebits;
3305 	m->dirty &= ~pagebits;
3306 	if (base == 0 && size == PAGE_SIZE) {
3307 		/*pmap_clear_modify(m);*/
3308 		vm_page_flag_clear(m, PG_NOSYNC);
3309 	}
3310 }
3311 
3312 /*
3313  * Set valid & dirty.  Used by buwrite()
3314  *
3315  * Page must be busied by caller.
3316  */
3317 void
3318 vm_page_set_validdirty(vm_page_t m, int base, int size)
3319 {
3320 	int pagebits;
3321 
3322 	pagebits = vm_page_bits(base, size);
3323 	m->valid |= pagebits;
3324 	m->dirty |= pagebits;
3325 	if (m->object)
3326 	       vm_object_set_writeable_dirty(m->object);
3327 }
3328 
3329 /*
3330  * Clear dirty bits.
3331  *
3332  * NOTE: This function does not clear the pmap modified bit.
3333  *	 Also note that e.g. NFS may use a byte-granular base
3334  *	 and size.
3335  *
3336  * Page must be busied?
3337  * No other requirements.
3338  */
3339 void
3340 vm_page_clear_dirty(vm_page_t m, int base, int size)
3341 {
3342 	m->dirty &= ~vm_page_bits(base, size);
3343 	if (base == 0 && size == PAGE_SIZE) {
3344 		/*pmap_clear_modify(m);*/
3345 		vm_page_flag_clear(m, PG_NOSYNC);
3346 	}
3347 }
3348 
3349 /*
3350  * Make the page all-dirty.
3351  *
3352  * Also make sure the related object and vnode reflect the fact that the
3353  * object may now contain a dirty page.
3354  *
3355  * Page must be busied?
3356  * No other requirements.
3357  */
3358 void
3359 vm_page_dirty(vm_page_t m)
3360 {
3361 #ifdef INVARIANTS
3362         int pqtype = m->queue - m->pc;
3363 #endif
3364         KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3365                 ("vm_page_dirty: page in free/cache queue!"));
3366 	if (m->dirty != VM_PAGE_BITS_ALL) {
3367 		m->dirty = VM_PAGE_BITS_ALL;
3368 		if (m->object)
3369 			vm_object_set_writeable_dirty(m->object);
3370 	}
3371 }
3372 
3373 /*
3374  * Invalidates DEV_BSIZE'd chunks within a page.  Both the
3375  * valid and dirty bits for the effected areas are cleared.
3376  *
3377  * Page must be busied?
3378  * Does not block.
3379  * No other requirements.
3380  */
3381 void
3382 vm_page_set_invalid(vm_page_t m, int base, int size)
3383 {
3384 	int bits;
3385 
3386 	bits = vm_page_bits(base, size);
3387 	m->valid &= ~bits;
3388 	m->dirty &= ~bits;
3389 	atomic_add_int(&m->object->generation, 1);
3390 }
3391 
3392 /*
3393  * The kernel assumes that the invalid portions of a page contain
3394  * garbage, but such pages can be mapped into memory by user code.
3395  * When this occurs, we must zero out the non-valid portions of the
3396  * page so user code sees what it expects.
3397  *
3398  * Pages are most often semi-valid when the end of a file is mapped
3399  * into memory and the file's size is not page aligned.
3400  *
3401  * Page must be busied?
3402  * No other requirements.
3403  */
3404 void
3405 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3406 {
3407 	int b;
3408 	int i;
3409 
3410 	/*
3411 	 * Scan the valid bits looking for invalid sections that
3412 	 * must be zerod.  Invalid sub-DEV_BSIZE'd areas ( where the
3413 	 * valid bit may be set ) have already been zerod by
3414 	 * vm_page_set_validclean().
3415 	 */
3416 	for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3417 		if (i == (PAGE_SIZE / DEV_BSIZE) ||
3418 		    (m->valid & (1 << i))
3419 		) {
3420 			if (i > b) {
3421 				pmap_zero_page_area(
3422 				    VM_PAGE_TO_PHYS(m),
3423 				    b << DEV_BSHIFT,
3424 				    (i - b) << DEV_BSHIFT
3425 				);
3426 			}
3427 			b = i + 1;
3428 		}
3429 	}
3430 
3431 	/*
3432 	 * setvalid is TRUE when we can safely set the zero'd areas
3433 	 * as being valid.  We can do this if there are no cache consistency
3434 	 * issues.  e.g. it is ok to do with UFS, but not ok to do with NFS.
3435 	 */
3436 	if (setvalid)
3437 		m->valid = VM_PAGE_BITS_ALL;
3438 }
3439 
3440 /*
3441  * Is a (partial) page valid?  Note that the case where size == 0
3442  * will return FALSE in the degenerate case where the page is entirely
3443  * invalid, and TRUE otherwise.
3444  *
3445  * Does not block.
3446  * No other requirements.
3447  */
3448 int
3449 vm_page_is_valid(vm_page_t m, int base, int size)
3450 {
3451 	int bits = vm_page_bits(base, size);
3452 
3453 	if (m->valid && ((m->valid & bits) == bits))
3454 		return 1;
3455 	else
3456 		return 0;
3457 }
3458 
3459 /*
3460  * update dirty bits from pmap/mmu.  May not block.
3461  *
3462  * Caller must hold the page busy
3463  */
3464 void
3465 vm_page_test_dirty(vm_page_t m)
3466 {
3467 	if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3468 		vm_page_dirty(m);
3469 	}
3470 }
3471 
3472 #include "opt_ddb.h"
3473 #ifdef DDB
3474 #include <ddb/ddb.h>
3475 
3476 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3477 {
3478 	db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count);
3479 	db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count);
3480 	db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count);
3481 	db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count);
3482 	db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count);
3483 	db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved);
3484 	db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min);
3485 	db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target);
3486 	db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min);
3487 	db_printf("vmstats.v_inactive_target: %ld\n",
3488 		  vmstats.v_inactive_target);
3489 }
3490 
3491 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3492 {
3493 	int i;
3494 	db_printf("PQ_FREE:");
3495 	for (i = 0; i < PQ_L2_SIZE; i++) {
3496 		db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3497 	}
3498 	db_printf("\n");
3499 
3500 	db_printf("PQ_CACHE:");
3501 	for(i = 0; i < PQ_L2_SIZE; i++) {
3502 		db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3503 	}
3504 	db_printf("\n");
3505 
3506 	db_printf("PQ_ACTIVE:");
3507 	for(i = 0; i < PQ_L2_SIZE; i++) {
3508 		db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3509 	}
3510 	db_printf("\n");
3511 
3512 	db_printf("PQ_INACTIVE:");
3513 	for(i = 0; i < PQ_L2_SIZE; i++) {
3514 		db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);
3515 	}
3516 	db_printf("\n");
3517 }
3518 #endif /* DDB */
3519