xref: /netbsd-src/sys/rump/librump/rumpkern/vm.c (revision c2f76ff004a2cb67efe5b12d97bd3ef7fe89e18d)
1 /*	$NetBSD: vm.c,v 1.108 2011/01/22 13:13:46 pooka Exp $	*/
2 
3 /*
4  * Copyright (c) 2007-2010 Antti Kantee.  All Rights Reserved.
5  *
6  * Development of this software was supported by
7  * The Finnish Cultural Foundation and the Research Foundation of
8  * The Helsinki University of Technology.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
20  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
25  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * Virtual memory emulation routines.
34  */
35 
36 /*
37  * XXX: we abuse pg->uanon for the virtual address of the storage
38  * for each page.  phys_addr would fit the job description better,
39  * except that it will create unnecessary lossage on some platforms
40  * due to not being a pointer type.
41  */
42 
43 #include <sys/cdefs.h>
44 __KERNEL_RCSID(0, "$NetBSD: vm.c,v 1.108 2011/01/22 13:13:46 pooka Exp $");
45 
46 #include <sys/param.h>
47 #include <sys/atomic.h>
48 #include <sys/buf.h>
49 #include <sys/kernel.h>
50 #include <sys/kmem.h>
51 #include <sys/mman.h>
52 #include <sys/null.h>
53 #include <sys/vnode.h>
54 
55 #include <machine/pmap.h>
56 
57 #include <rump/rumpuser.h>
58 
59 #include <uvm/uvm.h>
60 #include <uvm/uvm_ddb.h>
61 #include <uvm/uvm_pdpolicy.h>
62 #include <uvm/uvm_prot.h>
63 #include <uvm/uvm_readahead.h>
64 
65 #include "rump_private.h"
66 #include "rump_vfs_private.h"
67 
68 kmutex_t uvm_pageqlock;
69 kmutex_t uvm_swap_data_lock;
70 
71 struct uvmexp uvmexp;
72 int *uvmexp_pagesize;
73 int *uvmexp_pagemask;
74 int *uvmexp_pageshift;
75 struct uvm uvm;
76 
77 struct vm_map rump_vmmap;
78 static struct vm_map_kernel kmem_map_store;
79 struct vm_map *kmem_map = &kmem_map_store.vmk_map;
80 
81 static struct vm_map_kernel kernel_map_store;
82 struct vm_map *kernel_map = &kernel_map_store.vmk_map;
83 
84 static unsigned int pdaemon_waiters;
85 static kmutex_t pdaemonmtx;
86 static kcondvar_t pdaemoncv, oomwait;
87 
88 unsigned long rump_physmemlimit = RUMPMEM_UNLIMITED;
89 static unsigned long curphysmem;
90 static unsigned long dddlim;		/* 90% of memory limit used */
91 #define NEED_PAGEDAEMON() \
92     (rump_physmemlimit != RUMPMEM_UNLIMITED && curphysmem > dddlim)
93 
94 /*
95  * Try to free two pages worth of pages from objects.
96  * If this succesfully frees a full page cache page, we'll
97  * free the released page plus PAGE_SIZE�/sizeof(vm_page).
98  */
99 #define PAGEDAEMON_OBJCHUNK (2*PAGE_SIZE / sizeof(struct vm_page))
100 
101 /*
102  * Keep a list of least recently used pages.  Since the only way a
103  * rump kernel can "access" a page is via lookup, we put the page
104  * at the back of queue every time a lookup for it is done.  If the
105  * page is in front of this global queue and we're short of memory,
106  * it's a candidate for pageout.
107  */
108 static struct pglist vmpage_lruqueue;
109 static unsigned vmpage_onqueue;
110 
111 static int
112 pg_compare_key(void *ctx, const void *n, const void *key)
113 {
114 	voff_t a = ((const struct vm_page *)n)->offset;
115 	voff_t b = *(const voff_t *)key;
116 
117 	if (a < b)
118 		return -1;
119 	else if (a > b)
120 		return 1;
121 	else
122 		return 0;
123 }
124 
125 static int
126 pg_compare_nodes(void *ctx, const void *n1, const void *n2)
127 {
128 
129 	return pg_compare_key(ctx, n1, &((const struct vm_page *)n2)->offset);
130 }
131 
132 const rb_tree_ops_t uvm_page_tree_ops = {
133 	.rbto_compare_nodes = pg_compare_nodes,
134 	.rbto_compare_key = pg_compare_key,
135 	.rbto_node_offset = offsetof(struct vm_page, rb_node),
136 	.rbto_context = NULL
137 };
138 
139 /*
140  * vm pages
141  */
142 
143 static int
144 pgctor(void *arg, void *obj, int flags)
145 {
146 	struct vm_page *pg = obj;
147 
148 	memset(pg, 0, sizeof(*pg));
149 	pg->uanon = rump_hypermalloc(PAGE_SIZE, PAGE_SIZE,
150 	    (flags & PR_WAITOK) == PR_WAITOK, "pgalloc");
151 	return pg->uanon == NULL;
152 }
153 
154 static void
155 pgdtor(void *arg, void *obj)
156 {
157 	struct vm_page *pg = obj;
158 
159 	rump_hyperfree(pg->uanon, PAGE_SIZE);
160 }
161 
162 static struct pool_cache pagecache;
163 
164 /*
165  * Called with the object locked.  We don't support anons.
166  */
167 struct vm_page *
168 uvm_pagealloc_strat(struct uvm_object *uobj, voff_t off, struct vm_anon *anon,
169 	int flags, int strat, int free_list)
170 {
171 	struct vm_page *pg;
172 
173 	KASSERT(uobj && mutex_owned(&uobj->vmobjlock));
174 	KASSERT(anon == NULL);
175 
176 	pg = pool_cache_get(&pagecache, PR_NOWAIT);
177 	if (__predict_false(pg == NULL)) {
178 		return NULL;
179 	}
180 
181 	pg->offset = off;
182 	pg->uobject = uobj;
183 
184 	pg->flags = PG_CLEAN|PG_BUSY|PG_FAKE;
185 	if (flags & UVM_PGA_ZERO) {
186 		uvm_pagezero(pg);
187 	}
188 
189 	TAILQ_INSERT_TAIL(&uobj->memq, pg, listq.queue);
190 	(void)rb_tree_insert_node(&uobj->rb_tree, pg);
191 
192 	/*
193 	 * Don't put anons on the LRU page queue.  We can't flush them
194 	 * (there's no concept of swap in a rump kernel), so no reason
195 	 * to bother with them.
196 	 */
197 	if (!UVM_OBJ_IS_AOBJ(uobj)) {
198 		atomic_inc_uint(&vmpage_onqueue);
199 		mutex_enter(&uvm_pageqlock);
200 		TAILQ_INSERT_TAIL(&vmpage_lruqueue, pg, pageq.queue);
201 		mutex_exit(&uvm_pageqlock);
202 	}
203 
204 	uobj->uo_npages++;
205 
206 	return pg;
207 }
208 
209 /*
210  * Release a page.
211  *
212  * Called with the vm object locked.
213  */
214 void
215 uvm_pagefree(struct vm_page *pg)
216 {
217 	struct uvm_object *uobj = pg->uobject;
218 
219 	KASSERT(mutex_owned(&uvm_pageqlock));
220 	KASSERT(mutex_owned(&uobj->vmobjlock));
221 
222 	if (pg->flags & PG_WANTED)
223 		wakeup(pg);
224 
225 	TAILQ_REMOVE(&uobj->memq, pg, listq.queue);
226 
227 	uobj->uo_npages--;
228 	rb_tree_remove_node(&uobj->rb_tree, pg);
229 
230 	if (!UVM_OBJ_IS_AOBJ(uobj)) {
231 		TAILQ_REMOVE(&vmpage_lruqueue, pg, pageq.queue);
232 		atomic_dec_uint(&vmpage_onqueue);
233 	}
234 
235 	pool_cache_put(&pagecache, pg);
236 }
237 
238 void
239 uvm_pagezero(struct vm_page *pg)
240 {
241 
242 	pg->flags &= ~PG_CLEAN;
243 	memset((void *)pg->uanon, 0, PAGE_SIZE);
244 }
245 
246 /*
247  * Misc routines
248  */
249 
250 static kmutex_t pagermtx;
251 
252 void
253 uvm_init(void)
254 {
255 	char buf[64];
256 	int error;
257 
258 	if (rumpuser_getenv("RUMP_MEMLIMIT", buf, sizeof(buf), &error) == 0) {
259 		unsigned long tmp;
260 		char *ep;
261 		int mult;
262 
263 		tmp = strtoll(buf, &ep, 10);
264 		if (strlen(ep) > 1)
265 			panic("uvm_init: invalid RUMP_MEMLIMIT: %s", buf);
266 
267 		/* mini-dehumanize-number */
268 		mult = 1;
269 		switch (*ep) {
270 		case 'k':
271 			mult = 1024;
272 			break;
273 		case 'm':
274 			mult = 1024*1024;
275 			break;
276 		case 'g':
277 			mult = 1024*1024*1024;
278 			break;
279 		case 0:
280 			break;
281 		default:
282 			panic("uvm_init: invalid RUMP_MEMLIMIT: %s", buf);
283 		}
284 		rump_physmemlimit = tmp * mult;
285 
286 		if (rump_physmemlimit / mult != tmp)
287 			panic("uvm_init: RUMP_MEMLIMIT overflow: %s", buf);
288 		/* it's not like we'd get far with, say, 1 byte, but ... */
289 		if (rump_physmemlimit == 0)
290 			panic("uvm_init: no memory");
291 
292 #define HUMANIZE_BYTES 9
293 		CTASSERT(sizeof(buf) >= HUMANIZE_BYTES);
294 		format_bytes(buf, HUMANIZE_BYTES, rump_physmemlimit);
295 #undef HUMANIZE_BYTES
296 		dddlim = 9 * (rump_physmemlimit / 10);
297 	} else {
298 		strlcpy(buf, "unlimited (host limit)", sizeof(buf));
299 	}
300 	aprint_verbose("total memory = %s\n", buf);
301 
302 	TAILQ_INIT(&vmpage_lruqueue);
303 
304 	uvmexp.free = 1024*1024; /* XXX: arbitrary & not updated */
305 
306 	mutex_init(&pagermtx, MUTEX_DEFAULT, 0);
307 	mutex_init(&uvm_pageqlock, MUTEX_DEFAULT, 0);
308 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, 0);
309 
310 	mutex_init(&pdaemonmtx, MUTEX_DEFAULT, 0);
311 	cv_init(&pdaemoncv, "pdaemon");
312 	cv_init(&oomwait, "oomwait");
313 
314 	kernel_map->pmap = pmap_kernel();
315 	callback_head_init(&kernel_map_store.vmk_reclaim_callback, IPL_VM);
316 	kmem_map->pmap = pmap_kernel();
317 	callback_head_init(&kmem_map_store.vmk_reclaim_callback, IPL_VM);
318 
319 	pool_cache_bootstrap(&pagecache, sizeof(struct vm_page), 0, 0, 0,
320 	    "page$", NULL, IPL_NONE, pgctor, pgdtor, NULL);
321 }
322 
323 void
324 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax)
325 {
326 
327 	vm->vm_map.pmap = pmap_kernel();
328 	vm->vm_refcnt = 1;
329 }
330 
331 void
332 uvm_pagewire(struct vm_page *pg)
333 {
334 
335 	/* nada */
336 }
337 
338 void
339 uvm_pageunwire(struct vm_page *pg)
340 {
341 
342 	/* nada */
343 }
344 
345 /*
346  * The uvm reclaim hook is not currently necessary because it is
347  * used only by ZFS and implements exactly the same functionality
348  * as the kva reclaim hook which we already run in the pagedaemon
349  * (rump vm does not have a concept of uvm_map(), so we cannot
350  * reclaim kva it when a mapping operation fails due to insufficient
351  * available kva).
352  */
353 void
354 uvm_reclaim_hook_add(struct uvm_reclaim_hook *hook_entry)
355 {
356 
357 }
358 __strong_alias(uvm_reclaim_hook_del,uvm_reclaim_hook_add);
359 
360 /* where's your schmonz now? */
361 #define PUNLIMIT(a)	\
362 p->p_rlimit[a].rlim_cur = p->p_rlimit[a].rlim_max = RLIM_INFINITY;
363 void
364 uvm_init_limits(struct proc *p)
365 {
366 
367 	PUNLIMIT(RLIMIT_STACK);
368 	PUNLIMIT(RLIMIT_DATA);
369 	PUNLIMIT(RLIMIT_RSS);
370 	PUNLIMIT(RLIMIT_AS);
371 	/* nice, cascade */
372 }
373 #undef PUNLIMIT
374 
375 /*
376  * This satisfies the "disgusting mmap hack" used by proplib.
377  * We probably should grow some more assertables to make sure we're
378  * not satisfying anything we shouldn't be satisfying.
379  */
380 int
381 uvm_mmap(struct vm_map *map, vaddr_t *addr, vsize_t size, vm_prot_t prot,
382 	vm_prot_t maxprot, int flags, void *handle, voff_t off, vsize_t locklim)
383 {
384 	void *uaddr;
385 	int error;
386 
387 	if (prot != (VM_PROT_READ | VM_PROT_WRITE))
388 		panic("uvm_mmap() variant unsupported");
389 	if (flags != (MAP_PRIVATE | MAP_ANON))
390 		panic("uvm_mmap() variant unsupported");
391 
392 	/* no reason in particular, but cf. uvm_default_mapaddr() */
393 	if (*addr != 0)
394 		panic("uvm_mmap() variant unsupported");
395 
396 	if (RUMP_LOCALPROC_P(curproc)) {
397 		uaddr = rumpuser_anonmmap(NULL, size, 0, 0, &error);
398 	} else {
399 		error = rumpuser_sp_anonmmap(curproc->p_vmspace->vm_map.pmap,
400 		    size, &uaddr);
401 	}
402 	if (uaddr == NULL)
403 		return error;
404 
405 	*addr = (vaddr_t)uaddr;
406 	return 0;
407 }
408 
409 struct pagerinfo {
410 	vaddr_t pgr_kva;
411 	int pgr_npages;
412 	struct vm_page **pgr_pgs;
413 	bool pgr_read;
414 
415 	LIST_ENTRY(pagerinfo) pgr_entries;
416 };
417 static LIST_HEAD(, pagerinfo) pagerlist = LIST_HEAD_INITIALIZER(pagerlist);
418 
419 /*
420  * Pager "map" in routine.  Instead of mapping, we allocate memory
421  * and copy page contents there.  Not optimal or even strictly
422  * correct (the caller might modify the page contents after mapping
423  * them in), but what the heck.  Assumes UVMPAGER_MAPIN_WAITOK.
424  */
425 vaddr_t
426 uvm_pagermapin(struct vm_page **pgs, int npages, int flags)
427 {
428 	struct pagerinfo *pgri;
429 	vaddr_t curkva;
430 	int i;
431 
432 	/* allocate structures */
433 	pgri = kmem_alloc(sizeof(*pgri), KM_SLEEP);
434 	pgri->pgr_kva = (vaddr_t)kmem_alloc(npages * PAGE_SIZE, KM_SLEEP);
435 	pgri->pgr_npages = npages;
436 	pgri->pgr_pgs = kmem_alloc(sizeof(struct vm_page *) * npages, KM_SLEEP);
437 	pgri->pgr_read = (flags & UVMPAGER_MAPIN_READ) != 0;
438 
439 	/* copy contents to "mapped" memory */
440 	for (i = 0, curkva = pgri->pgr_kva;
441 	    i < npages;
442 	    i++, curkva += PAGE_SIZE) {
443 		/*
444 		 * We need to copy the previous contents of the pages to
445 		 * the window even if we are reading from the
446 		 * device, since the device might not fill the contents of
447 		 * the full mapped range and we will end up corrupting
448 		 * data when we unmap the window.
449 		 */
450 		memcpy((void*)curkva, pgs[i]->uanon, PAGE_SIZE);
451 		pgri->pgr_pgs[i] = pgs[i];
452 	}
453 
454 	mutex_enter(&pagermtx);
455 	LIST_INSERT_HEAD(&pagerlist, pgri, pgr_entries);
456 	mutex_exit(&pagermtx);
457 
458 	return pgri->pgr_kva;
459 }
460 
461 /*
462  * map out the pager window.  return contents from VA to page storage
463  * and free structures.
464  *
465  * Note: does not currently support partial frees
466  */
467 void
468 uvm_pagermapout(vaddr_t kva, int npages)
469 {
470 	struct pagerinfo *pgri;
471 	vaddr_t curkva;
472 	int i;
473 
474 	mutex_enter(&pagermtx);
475 	LIST_FOREACH(pgri, &pagerlist, pgr_entries) {
476 		if (pgri->pgr_kva == kva)
477 			break;
478 	}
479 	KASSERT(pgri);
480 	if (pgri->pgr_npages != npages)
481 		panic("uvm_pagermapout: partial unmapping not supported");
482 	LIST_REMOVE(pgri, pgr_entries);
483 	mutex_exit(&pagermtx);
484 
485 	if (pgri->pgr_read) {
486 		for (i = 0, curkva = pgri->pgr_kva;
487 		    i < pgri->pgr_npages;
488 		    i++, curkva += PAGE_SIZE) {
489 			memcpy(pgri->pgr_pgs[i]->uanon,(void*)curkva,PAGE_SIZE);
490 		}
491 	}
492 
493 	kmem_free(pgri->pgr_pgs, npages * sizeof(struct vm_page *));
494 	kmem_free((void*)pgri->pgr_kva, npages * PAGE_SIZE);
495 	kmem_free(pgri, sizeof(*pgri));
496 }
497 
498 /*
499  * convert va in pager window to page structure.
500  * XXX: how expensive is this (global lock, list traversal)?
501  */
502 struct vm_page *
503 uvm_pageratop(vaddr_t va)
504 {
505 	struct pagerinfo *pgri;
506 	struct vm_page *pg = NULL;
507 	int i;
508 
509 	mutex_enter(&pagermtx);
510 	LIST_FOREACH(pgri, &pagerlist, pgr_entries) {
511 		if (pgri->pgr_kva <= va
512 		    && va < pgri->pgr_kva + pgri->pgr_npages*PAGE_SIZE)
513 			break;
514 	}
515 	if (pgri) {
516 		i = (va - pgri->pgr_kva) >> PAGE_SHIFT;
517 		pg = pgri->pgr_pgs[i];
518 	}
519 	mutex_exit(&pagermtx);
520 
521 	return pg;
522 }
523 
524 /*
525  * Called with the vm object locked.
526  *
527  * Put vnode object pages at the end of the access queue to indicate
528  * they have been recently accessed and should not be immediate
529  * candidates for pageout.  Do not do this for lookups done by
530  * the pagedaemon to mimic pmap_kentered mappings which don't track
531  * access information.
532  */
533 struct vm_page *
534 uvm_pagelookup(struct uvm_object *uobj, voff_t off)
535 {
536 	struct vm_page *pg;
537 	bool ispagedaemon = curlwp == uvm.pagedaemon_lwp;
538 
539 	pg = rb_tree_find_node(&uobj->rb_tree, &off);
540 	if (pg && !UVM_OBJ_IS_AOBJ(pg->uobject) && !ispagedaemon) {
541 		mutex_enter(&uvm_pageqlock);
542 		TAILQ_REMOVE(&vmpage_lruqueue, pg, pageq.queue);
543 		TAILQ_INSERT_TAIL(&vmpage_lruqueue, pg, pageq.queue);
544 		mutex_exit(&uvm_pageqlock);
545 	}
546 
547 	return pg;
548 }
549 
550 void
551 uvm_page_unbusy(struct vm_page **pgs, int npgs)
552 {
553 	struct vm_page *pg;
554 	int i;
555 
556 	KASSERT(npgs > 0);
557 	KASSERT(mutex_owned(&pgs[0]->uobject->vmobjlock));
558 
559 	for (i = 0; i < npgs; i++) {
560 		pg = pgs[i];
561 		if (pg == NULL)
562 			continue;
563 
564 		KASSERT(pg->flags & PG_BUSY);
565 		if (pg->flags & PG_WANTED)
566 			wakeup(pg);
567 		if (pg->flags & PG_RELEASED)
568 			uvm_pagefree(pg);
569 		else
570 			pg->flags &= ~(PG_WANTED|PG_BUSY);
571 	}
572 }
573 
574 void
575 uvm_estimatepageable(int *active, int *inactive)
576 {
577 
578 	/* XXX: guessing game */
579 	*active = 1024;
580 	*inactive = 1024;
581 }
582 
583 struct vm_map_kernel *
584 vm_map_to_kernel(struct vm_map *map)
585 {
586 
587 	return (struct vm_map_kernel *)map;
588 }
589 
590 bool
591 vm_map_starved_p(struct vm_map *map)
592 {
593 
594 	if (map->flags & VM_MAP_WANTVA)
595 		return true;
596 
597 	return false;
598 }
599 
600 int
601 uvm_loan(struct vm_map *map, vaddr_t start, vsize_t len, void *v, int flags)
602 {
603 
604 	panic("%s: unimplemented", __func__);
605 }
606 
607 void
608 uvm_unloan(void *v, int npages, int flags)
609 {
610 
611 	panic("%s: unimplemented", __func__);
612 }
613 
614 int
615 uvm_loanuobjpages(struct uvm_object *uobj, voff_t pgoff, int orignpages,
616 	struct vm_page **opp)
617 {
618 
619 	return EBUSY;
620 }
621 
622 #ifdef DEBUGPRINT
623 void
624 uvm_object_printit(struct uvm_object *uobj, bool full,
625 	void (*pr)(const char *, ...))
626 {
627 
628 	pr("VM OBJECT at %p, refs %d", uobj, uobj->uo_refs);
629 }
630 #endif
631 
632 vaddr_t
633 uvm_default_mapaddr(struct proc *p, vaddr_t base, vsize_t sz)
634 {
635 
636 	return 0;
637 }
638 
639 int
640 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end,
641 	vm_prot_t prot, bool set_max)
642 {
643 
644 	return EOPNOTSUPP;
645 }
646 
647 /*
648  * UVM km
649  */
650 
651 vaddr_t
652 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
653 {
654 	void *rv, *desired = NULL;
655 	int alignbit, error;
656 
657 #ifdef __x86_64__
658 	/*
659 	 * On amd64, allocate all module memory from the lowest 2GB.
660 	 * This is because NetBSD kernel modules are compiled
661 	 * with -mcmodel=kernel and reserve only 4 bytes for
662 	 * offsets.  If we load code compiled with -mcmodel=kernel
663 	 * anywhere except the lowest or highest 2GB, it will not
664 	 * work.  Since userspace does not have access to the highest
665 	 * 2GB, use the lowest 2GB.
666 	 *
667 	 * Note: this assumes the rump kernel resides in
668 	 * the lowest 2GB as well.
669 	 *
670 	 * Note2: yes, it's a quick hack, but since this the only
671 	 * place where we care about the map we're allocating from,
672 	 * just use a simple "if" instead of coming up with a fancy
673 	 * generic solution.
674 	 */
675 	extern struct vm_map *module_map;
676 	if (map == module_map) {
677 		desired = (void *)(0x80000000 - size);
678 	}
679 #endif
680 
681 	alignbit = 0;
682 	if (align) {
683 		alignbit = ffs(align)-1;
684 	}
685 
686 	rv = rumpuser_anonmmap(desired, size, alignbit, flags & UVM_KMF_EXEC,
687 	    &error);
688 	if (rv == NULL) {
689 		if (flags & (UVM_KMF_CANFAIL | UVM_KMF_NOWAIT))
690 			return 0;
691 		else
692 			panic("uvm_km_alloc failed");
693 	}
694 
695 	if (flags & UVM_KMF_ZERO)
696 		memset(rv, 0, size);
697 
698 	return (vaddr_t)rv;
699 }
700 
701 void
702 uvm_km_free(struct vm_map *map, vaddr_t vaddr, vsize_t size, uvm_flag_t flags)
703 {
704 
705 	rumpuser_unmap((void *)vaddr, size);
706 }
707 
708 struct vm_map *
709 uvm_km_suballoc(struct vm_map *map, vaddr_t *minaddr, vaddr_t *maxaddr,
710 	vsize_t size, int pageable, bool fixed, struct vm_map_kernel *submap)
711 {
712 
713 	return (struct vm_map *)417416;
714 }
715 
716 vaddr_t
717 uvm_km_alloc_poolpage(struct vm_map *map, bool waitok)
718 {
719 
720 	return (vaddr_t)rump_hypermalloc(PAGE_SIZE, PAGE_SIZE,
721 	    waitok, "kmalloc");
722 }
723 
724 void
725 uvm_km_free_poolpage(struct vm_map *map, vaddr_t addr)
726 {
727 
728 	rump_hyperfree((void *)addr, PAGE_SIZE);
729 }
730 
731 vaddr_t
732 uvm_km_alloc_poolpage_cache(struct vm_map *map, bool waitok)
733 {
734 
735 	return uvm_km_alloc_poolpage(map, waitok);
736 }
737 
738 void
739 uvm_km_free_poolpage_cache(struct vm_map *map, vaddr_t vaddr)
740 {
741 
742 	uvm_km_free_poolpage(map, vaddr);
743 }
744 
745 void
746 uvm_km_va_drain(struct vm_map *map, uvm_flag_t flags)
747 {
748 
749 	/* we eventually maybe want some model for available memory */
750 }
751 
752 /*
753  * VM space locking routines.  We don't really have to do anything,
754  * since the pages are always "wired" (both local and remote processes).
755  */
756 int
757 uvm_vslock(struct vmspace *vs, void *addr, size_t len, vm_prot_t access)
758 {
759 
760 	return 0;
761 }
762 
763 void
764 uvm_vsunlock(struct vmspace *vs, void *addr, size_t len)
765 {
766 
767 }
768 
769 /*
770  * For the local case the buffer mappers don't need to do anything.
771  * For the remote case we need to reserve space and copy data in or
772  * out, depending on B_READ/B_WRITE.
773  */
774 void
775 vmapbuf(struct buf *bp, vsize_t len)
776 {
777 
778 	bp->b_saveaddr = bp->b_data;
779 
780 	/* remote case */
781 	if (!RUMP_LOCALPROC_P(curproc)) {
782 		bp->b_data = rump_hypermalloc(len, 0, true, "vmapbuf");
783 		if (BUF_ISWRITE(bp)) {
784 			copyin(bp->b_saveaddr, bp->b_data, len);
785 		}
786 	}
787 }
788 
789 void
790 vunmapbuf(struct buf *bp, vsize_t len)
791 {
792 
793 	/* remote case */
794 	if (!RUMP_LOCALPROC_P(bp->b_proc)) {
795 		if (BUF_ISREAD(bp)) {
796 			copyout_proc(bp->b_proc,
797 			    bp->b_data, bp->b_saveaddr, len);
798 		}
799 		rump_hyperfree(bp->b_data, len);
800 	}
801 
802 	bp->b_data = bp->b_saveaddr;
803 	bp->b_saveaddr = 0;
804 }
805 
806 void
807 uvmspace_addref(struct vmspace *vm)
808 {
809 
810 	/*
811 	 * No dynamically allocated vmspaces exist.
812 	 */
813 }
814 
815 void
816 uvmspace_free(struct vmspace *vm)
817 {
818 
819 	/* nothing for now */
820 }
821 
822 /*
823  * page life cycle stuff.  it really doesn't exist, so just stubs.
824  */
825 
826 void
827 uvm_pageactivate(struct vm_page *pg)
828 {
829 
830 	/* nada */
831 }
832 
833 void
834 uvm_pagedeactivate(struct vm_page *pg)
835 {
836 
837 	/* nada */
838 }
839 
840 void
841 uvm_pagedequeue(struct vm_page *pg)
842 {
843 
844 	/* nada*/
845 }
846 
847 void
848 uvm_pageenqueue(struct vm_page *pg)
849 {
850 
851 	/* nada */
852 }
853 
854 void
855 uvmpdpol_anfree(struct vm_anon *an)
856 {
857 
858 	/* nada */
859 }
860 
861 /*
862  * Physical address accessors.
863  */
864 
865 struct vm_page *
866 uvm_phys_to_vm_page(paddr_t pa)
867 {
868 
869 	return NULL;
870 }
871 
872 paddr_t
873 uvm_vm_page_to_phys(const struct vm_page *pg)
874 {
875 
876 	return 0;
877 }
878 
879 /*
880  * Routines related to the Page Baroness.
881  */
882 
883 void
884 uvm_wait(const char *msg)
885 {
886 
887 	if (__predict_false(curlwp == uvm.pagedaemon_lwp))
888 		panic("pagedaemon out of memory");
889 	if (__predict_false(rump_threads == 0))
890 		panic("pagedaemon missing (RUMP_THREADS = 0)");
891 
892 	mutex_enter(&pdaemonmtx);
893 	pdaemon_waiters++;
894 	cv_signal(&pdaemoncv);
895 	cv_wait(&oomwait, &pdaemonmtx);
896 	mutex_exit(&pdaemonmtx);
897 }
898 
899 void
900 uvm_pageout_start(int npages)
901 {
902 
903 	/* we don't have the heuristics */
904 }
905 
906 void
907 uvm_pageout_done(int npages)
908 {
909 
910 	/* could wakeup waiters, but just let the pagedaemon do it */
911 }
912 
913 static bool
914 processpage(struct vm_page *pg, bool *lockrunning)
915 {
916 	struct uvm_object *uobj;
917 
918 	uobj = pg->uobject;
919 	if (mutex_tryenter(&uobj->vmobjlock)) {
920 		if ((pg->flags & PG_BUSY) == 0) {
921 			mutex_exit(&uvm_pageqlock);
922 			uobj->pgops->pgo_put(uobj, pg->offset,
923 			    pg->offset + PAGE_SIZE,
924 			    PGO_CLEANIT|PGO_FREE);
925 			KASSERT(!mutex_owned(&uobj->vmobjlock));
926 			return true;
927 		} else {
928 			mutex_exit(&uobj->vmobjlock);
929 		}
930 	} else if (*lockrunning == false && ncpu > 1) {
931 		CPU_INFO_ITERATOR cii;
932 		struct cpu_info *ci;
933 		struct lwp *l;
934 
935 		l = mutex_owner(&uobj->vmobjlock);
936 		for (CPU_INFO_FOREACH(cii, ci)) {
937 			if (ci->ci_curlwp == l) {
938 				*lockrunning = true;
939 				break;
940 			}
941 		}
942 	}
943 
944 	return false;
945 }
946 
947 /*
948  * The Diabolical pageDaemon Director (DDD).
949  */
950 void
951 uvm_pageout(void *arg)
952 {
953 	struct vm_page *pg;
954 	struct pool *pp, *pp_first;
955 	uint64_t where;
956 	int timo = 0;
957 	int cleaned, skip, skipped;
958 	bool succ = false;
959 	bool lockrunning;
960 
961 	mutex_enter(&pdaemonmtx);
962 	for (;;) {
963 		if (succ) {
964 			kernel_map->flags &= ~VM_MAP_WANTVA;
965 			kmem_map->flags &= ~VM_MAP_WANTVA;
966 			timo = 0;
967 			if (pdaemon_waiters) {
968 				pdaemon_waiters = 0;
969 				cv_broadcast(&oomwait);
970 			}
971 		}
972 		succ = false;
973 
974 		if (pdaemon_waiters == 0) {
975 			cv_timedwait(&pdaemoncv, &pdaemonmtx, timo);
976 			uvmexp.pdwoke++;
977 		}
978 
979 		/* tell the world that we are hungry */
980 		kernel_map->flags |= VM_MAP_WANTVA;
981 		kmem_map->flags |= VM_MAP_WANTVA;
982 
983 		if (pdaemon_waiters == 0 && !NEED_PAGEDAEMON())
984 			continue;
985 		mutex_exit(&pdaemonmtx);
986 
987 		/*
988 		 * step one: reclaim the page cache.  this should give
989 		 * us the biggest earnings since whole pages are released
990 		 * into backing memory.
991 		 */
992 		pool_cache_reclaim(&pagecache);
993 		if (!NEED_PAGEDAEMON()) {
994 			succ = true;
995 			mutex_enter(&pdaemonmtx);
996 			continue;
997 		}
998 
999 		/*
1000 		 * Ok, so that didn't help.  Next, try to hunt memory
1001 		 * by pushing out vnode pages.  The pages might contain
1002 		 * useful cached data, but we need the memory.
1003 		 */
1004 		cleaned = 0;
1005 		skip = 0;
1006 		lockrunning = false;
1007  again:
1008 		mutex_enter(&uvm_pageqlock);
1009 		while (cleaned < PAGEDAEMON_OBJCHUNK) {
1010 			skipped = 0;
1011 			TAILQ_FOREACH(pg, &vmpage_lruqueue, pageq.queue) {
1012 
1013 				/*
1014 				 * skip over pages we _might_ have tried
1015 				 * to handle earlier.  they might not be
1016 				 * exactly the same ones, but I'm not too
1017 				 * concerned.
1018 				 */
1019 				while (skipped++ < skip)
1020 					continue;
1021 
1022 				if (processpage(pg, &lockrunning)) {
1023 					cleaned++;
1024 					goto again;
1025 				}
1026 
1027 				skip++;
1028 			}
1029 			break;
1030 		}
1031 		mutex_exit(&uvm_pageqlock);
1032 
1033 		/*
1034 		 * Ok, someone is running with an object lock held.
1035 		 * We want to yield the host CPU to make sure the
1036 		 * thread is not parked on the host.  Since sched_yield()
1037 		 * doesn't appear to do anything on NetBSD, nanosleep
1038 		 * for the smallest possible time and hope we're back in
1039 		 * the game soon.
1040 		 */
1041 		if (cleaned == 0 && lockrunning) {
1042 			uint64_t sec, nsec;
1043 
1044 			sec = 0;
1045 			nsec = 1;
1046 			rumpuser_nanosleep(&sec, &nsec, NULL);
1047 
1048 			lockrunning = false;
1049 			skip = 0;
1050 
1051 			/* and here we go again */
1052 			goto again;
1053 		}
1054 
1055 		/*
1056 		 * And of course we need to reclaim the page cache
1057 		 * again to actually release memory.
1058 		 */
1059 		pool_cache_reclaim(&pagecache);
1060 		if (!NEED_PAGEDAEMON()) {
1061 			succ = true;
1062 			mutex_enter(&pdaemonmtx);
1063 			continue;
1064 		}
1065 
1066 		/*
1067 		 * Still not there?  sleeves come off right about now.
1068 		 * First: do reclaim on kernel/kmem map.
1069 		 */
1070 		callback_run_roundrobin(&kernel_map_store.vmk_reclaim_callback,
1071 		    NULL);
1072 		callback_run_roundrobin(&kmem_map_store.vmk_reclaim_callback,
1073 		    NULL);
1074 
1075 		/*
1076 		 * And then drain the pools.  Wipe them out ... all of them.
1077 		 */
1078 
1079 		pool_drain_start(&pp_first, &where);
1080 		pp = pp_first;
1081 		for (;;) {
1082 			rump_vfs_drainbufs(10 /* XXX: estimate better */);
1083 			succ = pool_drain_end(pp, where);
1084 			if (succ)
1085 				break;
1086 			pool_drain_start(&pp, &where);
1087 			if (pp == pp_first) {
1088 				succ = pool_drain_end(pp, where);
1089 				break;
1090 			}
1091 		}
1092 
1093 		/*
1094 		 * Need to use PYEC on our bag of tricks.
1095 		 * Unfortunately, the wife just borrowed it.
1096 		 */
1097 
1098 		if (!succ && cleaned == 0) {
1099 			rumpuser_dprintf("pagedaemoness: failed to reclaim "
1100 			    "memory ... sleeping (deadlock?)\n");
1101 			timo = hz;
1102 		}
1103 
1104 		mutex_enter(&pdaemonmtx);
1105 	}
1106 
1107 	panic("you can swap out any time you like, but you can never leave");
1108 }
1109 
1110 void
1111 uvm_kick_pdaemon()
1112 {
1113 
1114 	/*
1115 	 * Wake up the diabolical pagedaemon director if we are over
1116 	 * 90% of the memory limit.  This is a complete and utter
1117 	 * stetson-harrison decision which you are allowed to finetune.
1118 	 * Don't bother locking.  If we have some unflushed caches,
1119 	 * other waker-uppers will deal with the issue.
1120 	 */
1121 	if (NEED_PAGEDAEMON()) {
1122 		cv_signal(&pdaemoncv);
1123 	}
1124 }
1125 
1126 void *
1127 rump_hypermalloc(size_t howmuch, int alignment, bool waitok, const char *wmsg)
1128 {
1129 	unsigned long newmem;
1130 	void *rv;
1131 
1132 	uvm_kick_pdaemon(); /* ouch */
1133 
1134 	/* first we must be within the limit */
1135  limitagain:
1136 	if (rump_physmemlimit != RUMPMEM_UNLIMITED) {
1137 		newmem = atomic_add_long_nv(&curphysmem, howmuch);
1138 		if (newmem > rump_physmemlimit) {
1139 			newmem = atomic_add_long_nv(&curphysmem, -howmuch);
1140 			if (!waitok) {
1141 				return NULL;
1142 			}
1143 			uvm_wait(wmsg);
1144 			goto limitagain;
1145 		}
1146 	}
1147 
1148 	/* second, we must get something from the backend */
1149  again:
1150 	rv = rumpuser_malloc(howmuch, alignment);
1151 	if (__predict_false(rv == NULL && waitok)) {
1152 		uvm_wait(wmsg);
1153 		goto again;
1154 	}
1155 
1156 	return rv;
1157 }
1158 
1159 void
1160 rump_hyperfree(void *what, size_t size)
1161 {
1162 
1163 	if (rump_physmemlimit != RUMPMEM_UNLIMITED) {
1164 		atomic_add_long(&curphysmem, -size);
1165 	}
1166 	rumpuser_free(what);
1167 }
1168