vm.c revision 1.84 1 /* $NetBSD: vm.c,v 1.84 2010/06/14 21:04:56 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. Contents:
34 * + anon objects & pager
35 * + misc support routines
36 */
37
38 /*
39 * XXX: we abuse pg->uanon for the virtual address of the storage
40 * for each page. phys_addr would fit the job description better,
41 * except that it will create unnecessary lossage on some platforms
42 * due to not being a pointer type.
43 */
44
45 #include <sys/cdefs.h>
46 __KERNEL_RCSID(0, "$NetBSD: vm.c,v 1.84 2010/06/14 21:04:56 pooka Exp $");
47
48 #include <sys/param.h>
49 #include <sys/atomic.h>
50 #include <sys/buf.h>
51 #include <sys/kernel.h>
52 #include <sys/kmem.h>
53 #include <sys/mman.h>
54 #include <sys/null.h>
55 #include <sys/vnode.h>
56
57 #include <machine/pmap.h>
58
59 #include <rump/rumpuser.h>
60
61 #include <uvm/uvm.h>
62 #include <uvm/uvm_ddb.h>
63 #include <uvm/uvm_prot.h>
64 #include <uvm/uvm_readahead.h>
65
66 #include "rump_private.h"
67
68 static int ao_get(struct uvm_object *, voff_t, struct vm_page **,
69 int *, int, vm_prot_t, int, int);
70 static int ao_put(struct uvm_object *, voff_t, voff_t, int);
71
72 const struct uvm_pagerops aobj_pager = {
73 .pgo_get = ao_get,
74 .pgo_put = ao_put,
75 };
76
77 kmutex_t uvm_pageqlock;
78
79 struct uvmexp uvmexp;
80 struct uvm uvm;
81
82 struct vm_map rump_vmmap;
83 static struct vm_map_kernel kmem_map_store;
84 struct vm_map *kmem_map = &kmem_map_store.vmk_map;
85 const struct rb_tree_ops uvm_page_tree_ops;
86
87 static struct vm_map_kernel kernel_map_store;
88 struct vm_map *kernel_map = &kernel_map_store.vmk_map;
89
90 static unsigned int pdaemon_waiters;
91 static kmutex_t pdaemonmtx;
92 static kcondvar_t pdaemoncv, oomwait;
93
94 #define RUMPMEM_UNLIMITED ((unsigned long)-1)
95 static unsigned long physmemlimit = RUMPMEM_UNLIMITED;
96 static unsigned long curphysmem;
97
98 /*
99 * vm pages
100 */
101
102 /* called with the object locked */
103 struct vm_page *
104 uvm_pagealloc_strat(struct uvm_object *uobj, voff_t off, struct vm_anon *anon,
105 int flags, int strat, int free_list)
106 {
107 struct vm_page *pg;
108
109 pg = kmem_zalloc(sizeof(struct vm_page), KM_SLEEP);
110 pg->offset = off;
111 pg->uobject = uobj;
112
113 pg->uanon = (void *)kmem_alloc(PAGE_SIZE, KM_SLEEP);
114 if (flags & UVM_PGA_ZERO)
115 memset(pg->uanon, 0, PAGE_SIZE);
116 pg->flags = PG_CLEAN|PG_BUSY|PG_FAKE;
117
118 TAILQ_INSERT_TAIL(&uobj->memq, pg, listq.queue);
119 uobj->uo_npages++;
120
121 return pg;
122 }
123
124 /*
125 * Release a page.
126 *
127 * Called with the vm object locked.
128 */
129 void
130 uvm_pagefree(struct vm_page *pg)
131 {
132 struct uvm_object *uobj = pg->uobject;
133
134 if (pg->flags & PG_WANTED)
135 wakeup(pg);
136
137 uobj->uo_npages--;
138 TAILQ_REMOVE(&uobj->memq, pg, listq.queue);
139 kmem_free((void *)pg->uanon, PAGE_SIZE);
140 kmem_free(pg, sizeof(*pg));
141 }
142
143 void
144 uvm_pagezero(struct vm_page *pg)
145 {
146
147 pg->flags &= ~PG_CLEAN;
148 memset((void *)pg->uanon, 0, PAGE_SIZE);
149 }
150
151 /*
152 * Anon object stuff
153 */
154
155 static int
156 ao_get(struct uvm_object *uobj, voff_t off, struct vm_page **pgs,
157 int *npages, int centeridx, vm_prot_t access_type,
158 int advice, int flags)
159 {
160 struct vm_page *pg;
161 int i;
162
163 if (centeridx)
164 panic("%s: centeridx != 0 not supported", __func__);
165
166 /* loop over pages */
167 off = trunc_page(off);
168 for (i = 0; i < *npages; i++) {
169 retrylookup:
170 pg = uvm_pagelookup(uobj, off + (i << PAGE_SHIFT));
171 if (pg) {
172 if (pg->flags & PG_BUSY) {
173 pg->flags |= PG_WANTED;
174 UVM_UNLOCK_AND_WAIT(pg, &uobj->vmobjlock, 0,
175 "aogetpg", 0);
176 goto retrylookup;
177 }
178 pg->flags |= PG_BUSY;
179 pgs[i] = pg;
180 } else {
181 pg = uvm_pagealloc(uobj,
182 off + (i << PAGE_SHIFT), NULL, UVM_PGA_ZERO);
183 pgs[i] = pg;
184 }
185 }
186 mutex_exit(&uobj->vmobjlock);
187
188 return 0;
189
190 }
191
192 static int
193 ao_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
194 {
195 struct vm_page *pg;
196
197 /* we only free all pages for now */
198 if ((flags & PGO_FREE) == 0 || (flags & PGO_ALLPAGES) == 0) {
199 mutex_exit(&uobj->vmobjlock);
200 return 0;
201 }
202
203 while ((pg = TAILQ_FIRST(&uobj->memq)) != NULL)
204 uvm_pagefree(pg);
205 mutex_exit(&uobj->vmobjlock);
206
207 return 0;
208 }
209
210 struct uvm_object *
211 uao_create(vsize_t size, int flags)
212 {
213 struct uvm_object *uobj;
214
215 uobj = kmem_zalloc(sizeof(struct uvm_object), KM_SLEEP);
216 uobj->pgops = &aobj_pager;
217 TAILQ_INIT(&uobj->memq);
218 mutex_init(&uobj->vmobjlock, MUTEX_DEFAULT, IPL_NONE);
219
220 return uobj;
221 }
222
223 void
224 uao_detach(struct uvm_object *uobj)
225 {
226
227 mutex_enter(&uobj->vmobjlock);
228 ao_put(uobj, 0, 0, PGO_ALLPAGES | PGO_FREE);
229 mutex_destroy(&uobj->vmobjlock);
230 kmem_free(uobj, sizeof(*uobj));
231 }
232
233 /*
234 * Misc routines
235 */
236
237 static kmutex_t pagermtx;
238
239 void
240 uvm_init(void)
241 {
242 char buf[64];
243 int error;
244
245 if (rumpuser_getenv("RUMP_MEMLIMIT", buf, sizeof(buf), &error) == 0) {
246 physmemlimit = strtoll(buf, NULL, 10);
247 /* it's not like we'd get far with, say, 1 byte, but ... */
248 if (physmemlimit == 0)
249 panic("uvm_init: no memory available");
250 #define HUMANIZE_BYTES 9
251 CTASSERT(sizeof(buf) >= HUMANIZE_BYTES);
252 format_bytes(buf, HUMANIZE_BYTES, physmemlimit);
253 #undef HUMANIZE_BYTES
254 } else {
255 strlcpy(buf, "unlimited (host limit)", sizeof(buf));
256 }
257 aprint_verbose("total memory = %s\n", buf);
258
259 uvmexp.free = 1024*1024; /* XXX: arbitrary & not updated */
260
261 mutex_init(&pagermtx, MUTEX_DEFAULT, 0);
262 mutex_init(&uvm_pageqlock, MUTEX_DEFAULT, 0);
263
264 mutex_init(&pdaemonmtx, MUTEX_DEFAULT, 0);
265 cv_init(&pdaemoncv, "pdaemon");
266 cv_init(&oomwait, "oomwait");
267
268 kernel_map->pmap = pmap_kernel();
269 callback_head_init(&kernel_map_store.vmk_reclaim_callback, IPL_VM);
270 kmem_map->pmap = pmap_kernel();
271 callback_head_init(&kmem_map_store.vmk_reclaim_callback, IPL_VM);
272 }
273
274 void
275 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax)
276 {
277
278 vm->vm_map.pmap = pmap_kernel();
279 vm->vm_refcnt = 1;
280 }
281
282 void
283 uvm_pagewire(struct vm_page *pg)
284 {
285
286 /* nada */
287 }
288
289 void
290 uvm_pageunwire(struct vm_page *pg)
291 {
292
293 /* nada */
294 }
295
296 /* where's your schmonz now? */
297 #define PUNLIMIT(a) \
298 p->p_rlimit[a].rlim_cur = p->p_rlimit[a].rlim_max = RLIM_INFINITY;
299 void
300 uvm_init_limits(struct proc *p)
301 {
302
303 PUNLIMIT(RLIMIT_STACK);
304 PUNLIMIT(RLIMIT_DATA);
305 PUNLIMIT(RLIMIT_RSS);
306 PUNLIMIT(RLIMIT_AS);
307 /* nice, cascade */
308 }
309 #undef PUNLIMIT
310
311 /*
312 * This satisfies the "disgusting mmap hack" used by proplib.
313 * We probably should grow some more assertables to make sure we're
314 * not satisfying anything we shouldn't be satisfying. At least we
315 * should make sure it's the local machine we're mmapping ...
316 */
317 int
318 uvm_mmap(struct vm_map *map, vaddr_t *addr, vsize_t size, vm_prot_t prot,
319 vm_prot_t maxprot, int flags, void *handle, voff_t off, vsize_t locklim)
320 {
321 void *uaddr;
322 int error;
323
324 if (prot != (VM_PROT_READ | VM_PROT_WRITE))
325 panic("uvm_mmap() variant unsupported");
326 if (flags != (MAP_PRIVATE | MAP_ANON))
327 panic("uvm_mmap() variant unsupported");
328 /* no reason in particular, but cf. uvm_default_mapaddr() */
329 if (*addr != 0)
330 panic("uvm_mmap() variant unsupported");
331
332 uaddr = rumpuser_anonmmap(NULL, size, 0, 0, &error);
333 if (uaddr == NULL)
334 return error;
335
336 *addr = (vaddr_t)uaddr;
337 return 0;
338 }
339
340 struct pagerinfo {
341 vaddr_t pgr_kva;
342 int pgr_npages;
343 struct vm_page **pgr_pgs;
344 bool pgr_read;
345
346 LIST_ENTRY(pagerinfo) pgr_entries;
347 };
348 static LIST_HEAD(, pagerinfo) pagerlist = LIST_HEAD_INITIALIZER(pagerlist);
349
350 /*
351 * Pager "map" in routine. Instead of mapping, we allocate memory
352 * and copy page contents there. Not optimal or even strictly
353 * correct (the caller might modify the page contents after mapping
354 * them in), but what the heck. Assumes UVMPAGER_MAPIN_WAITOK.
355 */
356 vaddr_t
357 uvm_pagermapin(struct vm_page **pgs, int npages, int flags)
358 {
359 struct pagerinfo *pgri;
360 vaddr_t curkva;
361 int i;
362
363 /* allocate structures */
364 pgri = kmem_alloc(sizeof(*pgri), KM_SLEEP);
365 pgri->pgr_kva = (vaddr_t)kmem_alloc(npages * PAGE_SIZE, KM_SLEEP);
366 pgri->pgr_npages = npages;
367 pgri->pgr_pgs = kmem_alloc(sizeof(struct vm_page *) * npages, KM_SLEEP);
368 pgri->pgr_read = (flags & UVMPAGER_MAPIN_READ) != 0;
369
370 /* copy contents to "mapped" memory */
371 for (i = 0, curkva = pgri->pgr_kva;
372 i < npages;
373 i++, curkva += PAGE_SIZE) {
374 /*
375 * We need to copy the previous contents of the pages to
376 * the window even if we are reading from the
377 * device, since the device might not fill the contents of
378 * the full mapped range and we will end up corrupting
379 * data when we unmap the window.
380 */
381 memcpy((void*)curkva, pgs[i]->uanon, PAGE_SIZE);
382 pgri->pgr_pgs[i] = pgs[i];
383 }
384
385 mutex_enter(&pagermtx);
386 LIST_INSERT_HEAD(&pagerlist, pgri, pgr_entries);
387 mutex_exit(&pagermtx);
388
389 return pgri->pgr_kva;
390 }
391
392 /*
393 * map out the pager window. return contents from VA to page storage
394 * and free structures.
395 *
396 * Note: does not currently support partial frees
397 */
398 void
399 uvm_pagermapout(vaddr_t kva, int npages)
400 {
401 struct pagerinfo *pgri;
402 vaddr_t curkva;
403 int i;
404
405 mutex_enter(&pagermtx);
406 LIST_FOREACH(pgri, &pagerlist, pgr_entries) {
407 if (pgri->pgr_kva == kva)
408 break;
409 }
410 KASSERT(pgri);
411 if (pgri->pgr_npages != npages)
412 panic("uvm_pagermapout: partial unmapping not supported");
413 LIST_REMOVE(pgri, pgr_entries);
414 mutex_exit(&pagermtx);
415
416 if (pgri->pgr_read) {
417 for (i = 0, curkva = pgri->pgr_kva;
418 i < pgri->pgr_npages;
419 i++, curkva += PAGE_SIZE) {
420 memcpy(pgri->pgr_pgs[i]->uanon,(void*)curkva,PAGE_SIZE);
421 }
422 }
423
424 kmem_free(pgri->pgr_pgs, npages * sizeof(struct vm_page *));
425 kmem_free((void*)pgri->pgr_kva, npages * PAGE_SIZE);
426 kmem_free(pgri, sizeof(*pgri));
427 }
428
429 /*
430 * convert va in pager window to page structure.
431 * XXX: how expensive is this (global lock, list traversal)?
432 */
433 struct vm_page *
434 uvm_pageratop(vaddr_t va)
435 {
436 struct pagerinfo *pgri;
437 struct vm_page *pg = NULL;
438 int i;
439
440 mutex_enter(&pagermtx);
441 LIST_FOREACH(pgri, &pagerlist, pgr_entries) {
442 if (pgri->pgr_kva <= va
443 && va < pgri->pgr_kva + pgri->pgr_npages*PAGE_SIZE)
444 break;
445 }
446 if (pgri) {
447 i = (va - pgri->pgr_kva) >> PAGE_SHIFT;
448 pg = pgri->pgr_pgs[i];
449 }
450 mutex_exit(&pagermtx);
451
452 return pg;
453 }
454
455 /* Called with the vm object locked */
456 struct vm_page *
457 uvm_pagelookup(struct uvm_object *uobj, voff_t off)
458 {
459 struct vm_page *pg;
460
461 TAILQ_FOREACH(pg, &uobj->memq, listq.queue) {
462 if (pg->offset == off) {
463 return pg;
464 }
465 }
466
467 return NULL;
468 }
469
470 void
471 uvm_page_unbusy(struct vm_page **pgs, int npgs)
472 {
473 struct vm_page *pg;
474 int i;
475
476 for (i = 0; i < npgs; i++) {
477 pg = pgs[i];
478 if (pg == NULL)
479 continue;
480
481 KASSERT(pg->flags & PG_BUSY);
482 if (pg->flags & PG_WANTED)
483 wakeup(pg);
484 if (pg->flags & PG_RELEASED)
485 uvm_pagefree(pg);
486 else
487 pg->flags &= ~(PG_WANTED|PG_BUSY);
488 }
489 }
490
491 void
492 uvm_estimatepageable(int *active, int *inactive)
493 {
494
495 /* XXX: guessing game */
496 *active = 1024;
497 *inactive = 1024;
498 }
499
500 struct vm_map_kernel *
501 vm_map_to_kernel(struct vm_map *map)
502 {
503
504 return (struct vm_map_kernel *)map;
505 }
506
507 bool
508 vm_map_starved_p(struct vm_map *map)
509 {
510
511 if (map->flags & VM_MAP_WANTVA)
512 return true;
513
514 return false;
515 }
516
517 int
518 uvm_loan(struct vm_map *map, vaddr_t start, vsize_t len, void *v, int flags)
519 {
520
521 panic("%s: unimplemented", __func__);
522 }
523
524 void
525 uvm_unloan(void *v, int npages, int flags)
526 {
527
528 panic("%s: unimplemented", __func__);
529 }
530
531 int
532 uvm_loanuobjpages(struct uvm_object *uobj, voff_t pgoff, int orignpages,
533 struct vm_page **opp)
534 {
535
536 return EBUSY;
537 }
538
539 #ifdef DEBUGPRINT
540 void
541 uvm_object_printit(struct uvm_object *uobj, bool full,
542 void (*pr)(const char *, ...))
543 {
544
545 pr("VM OBJECT at %p, refs %d", uobj, uobj->uo_refs);
546 }
547 #endif
548
549 vaddr_t
550 uvm_default_mapaddr(struct proc *p, vaddr_t base, vsize_t sz)
551 {
552
553 return 0;
554 }
555
556 int
557 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end,
558 vm_prot_t prot, bool set_max)
559 {
560
561 return EOPNOTSUPP;
562 }
563
564 /*
565 * UVM km
566 */
567
568 vaddr_t
569 uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
570 {
571 void *rv, *desired = NULL;
572 int alignbit, error;
573
574 #ifdef __x86_64__
575 /*
576 * On amd64, allocate all module memory from the lowest 2GB.
577 * This is because NetBSD kernel modules are compiled
578 * with -mcmodel=kernel and reserve only 4 bytes for
579 * offsets. If we load code compiled with -mcmodel=kernel
580 * anywhere except the lowest or highest 2GB, it will not
581 * work. Since userspace does not have access to the highest
582 * 2GB, use the lowest 2GB.
583 *
584 * Note: this assumes the rump kernel resides in
585 * the lowest 2GB as well.
586 *
587 * Note2: yes, it's a quick hack, but since this the only
588 * place where we care about the map we're allocating from,
589 * just use a simple "if" instead of coming up with a fancy
590 * generic solution.
591 */
592 extern struct vm_map *module_map;
593 if (map == module_map) {
594 desired = (void *)(0x80000000 - size);
595 }
596 #endif
597
598 alignbit = 0;
599 if (align) {
600 alignbit = ffs(align)-1;
601 }
602
603 rv = rumpuser_anonmmap(desired, size, alignbit, flags & UVM_KMF_EXEC,
604 &error);
605 if (rv == NULL) {
606 if (flags & (UVM_KMF_CANFAIL | UVM_KMF_NOWAIT))
607 return 0;
608 else
609 panic("uvm_km_alloc failed");
610 }
611
612 if (flags & UVM_KMF_ZERO)
613 memset(rv, 0, size);
614
615 return (vaddr_t)rv;
616 }
617
618 void
619 uvm_km_free(struct vm_map *map, vaddr_t vaddr, vsize_t size, uvm_flag_t flags)
620 {
621
622 rumpuser_unmap((void *)vaddr, size);
623 }
624
625 struct vm_map *
626 uvm_km_suballoc(struct vm_map *map, vaddr_t *minaddr, vaddr_t *maxaddr,
627 vsize_t size, int pageable, bool fixed, struct vm_map_kernel *submap)
628 {
629
630 return (struct vm_map *)417416;
631 }
632
633 vaddr_t
634 uvm_km_alloc_poolpage(struct vm_map *map, bool waitok)
635 {
636
637 return (vaddr_t)rump_hypermalloc(PAGE_SIZE, PAGE_SIZE,
638 waitok, "kmalloc");
639 }
640
641 void
642 uvm_km_free_poolpage(struct vm_map *map, vaddr_t addr)
643 {
644
645 rump_hyperfree((void *)addr, PAGE_SIZE);
646 }
647
648 vaddr_t
649 uvm_km_alloc_poolpage_cache(struct vm_map *map, bool waitok)
650 {
651
652 return uvm_km_alloc_poolpage(map, waitok);
653 }
654
655 void
656 uvm_km_free_poolpage_cache(struct vm_map *map, vaddr_t vaddr)
657 {
658
659 uvm_km_free_poolpage(map, vaddr);
660 }
661
662 void
663 uvm_km_va_drain(struct vm_map *map, uvm_flag_t flags)
664 {
665
666 /* we eventually maybe want some model for available memory */
667 }
668
669 /*
670 * Mapping and vm space locking routines.
671 * XXX: these don't work for non-local vmspaces
672 */
673 int
674 uvm_vslock(struct vmspace *vs, void *addr, size_t len, vm_prot_t access)
675 {
676
677 KASSERT(vs == &vmspace0);
678 return 0;
679 }
680
681 void
682 uvm_vsunlock(struct vmspace *vs, void *addr, size_t len)
683 {
684
685 KASSERT(vs == &vmspace0);
686 }
687
688 void
689 vmapbuf(struct buf *bp, vsize_t len)
690 {
691
692 bp->b_saveaddr = bp->b_data;
693 }
694
695 void
696 vunmapbuf(struct buf *bp, vsize_t len)
697 {
698
699 bp->b_data = bp->b_saveaddr;
700 bp->b_saveaddr = 0;
701 }
702
703 void
704 uvmspace_addref(struct vmspace *vm)
705 {
706
707 /*
708 * there is only vmspace0. we're not planning on
709 * feeding it to the fishes.
710 */
711 }
712
713 void
714 uvmspace_free(struct vmspace *vm)
715 {
716
717 /* nothing for now */
718 }
719
720 int
721 uvm_io(struct vm_map *map, struct uio *uio)
722 {
723
724 /*
725 * just do direct uio for now. but this needs some vmspace
726 * olympics for rump_sysproxy.
727 */
728 return uiomove((void *)(vaddr_t)uio->uio_offset, uio->uio_resid, uio);
729 }
730
731 /*
732 * page life cycle stuff. it really doesn't exist, so just stubs.
733 */
734
735 void
736 uvm_pageactivate(struct vm_page *pg)
737 {
738
739 /* nada */
740 }
741
742 void
743 uvm_pagedeactivate(struct vm_page *pg)
744 {
745
746 /* nada */
747 }
748
749 void
750 uvm_pagedequeue(struct vm_page *pg)
751 {
752
753 /* nada*/
754 }
755
756 void
757 uvm_pageenqueue(struct vm_page *pg)
758 {
759
760 /* nada */
761 }
762
763 /*
764 * Routines related to the Page Baroness.
765 */
766
767 void
768 uvm_wait(const char *msg)
769 {
770
771 if (__predict_false(curlwp == uvm.pagedaemon_lwp))
772 panic("pagedaemon out of memory");
773 if (__predict_false(rump_threads == 0))
774 panic("pagedaemon missing (RUMP_THREADS = 0)");
775
776 mutex_enter(&pdaemonmtx);
777 pdaemon_waiters++;
778 cv_signal(&pdaemoncv);
779 cv_wait(&oomwait, &pdaemonmtx);
780 mutex_exit(&pdaemonmtx);
781 }
782
783 void
784 uvm_pageout_start(int npages)
785 {
786
787 /* we don't have the heuristics */
788 }
789
790 void
791 uvm_pageout_done(int npages)
792 {
793
794 /* could wakeup waiters, but just let the pagedaemon do it */
795 }
796
797 /*
798 * Under-construction page mistress. This is lacking vfs support, namely:
799 *
800 * 1) draining vfs buffers
801 * 2) paging out pages in vm vnode objects
802 * (we will not page out anon memory on the basis that
803 * that's the task of the host)
804 */
805
806 void
807 uvm_pageout(void *arg)
808 {
809 struct pool *pp, *pp_first;
810 uint64_t where;
811 int timo = 0;
812 bool succ;
813
814 mutex_enter(&pdaemonmtx);
815 for (;;) {
816 cv_timedwait(&pdaemoncv, &pdaemonmtx, timo);
817 uvmexp.pdwoke++;
818 kernel_map->flags |= VM_MAP_WANTVA;
819 mutex_exit(&pdaemonmtx);
820
821 succ = false;
822 pool_drain_start(&pp_first, &where);
823 pp = pp_first;
824 for (;;) {
825 succ = pool_drain_end(pp, where);
826 if (succ)
827 break;
828 pool_drain_start(&pp, &where);
829 if (pp == pp_first) {
830 succ = pool_drain_end(pp, where);
831 break;
832 }
833 }
834 mutex_enter(&pdaemonmtx);
835
836 if (!succ) {
837 rumpuser_dprintf("pagedaemoness: failed to reclaim "
838 "memory ... sleeping (deadlock?)\n");
839 timo = hz;
840 continue;
841 }
842 kernel_map->flags &= ~VM_MAP_WANTVA;
843 timo = 0;
844
845 if (pdaemon_waiters) {
846 pdaemon_waiters = 0;
847 cv_broadcast(&oomwait);
848 }
849 }
850
851 panic("you can swap out any time you like, but you can never leave");
852 }
853
854 /*
855 * In a regular kernel the pagedaemon is activated when memory becomes
856 * low. In a virtual rump kernel we do not know exactly how much memory
857 * we have available -- it depends on the conditions on the host.
858 * Therefore, we cannot preemptively kick the pagedaemon. Rather, we
859 * wait until things we desperate and we're forced to uvm_wait().
860 *
861 * The alternative would be to allocate a huge chunk of memory at
862 * startup, but that solution has a number of problems including
863 * being a resource hog, failing anyway due to host memory overcommit
864 * and core dump size.
865 */
866
867 void
868 uvm_kick_pdaemon()
869 {
870
871 /* nada */
872 }
873
874 void *
875 rump_hypermalloc(size_t howmuch, int alignment, bool waitok, const char *wmsg)
876 {
877 unsigned long newmem;
878 void *rv;
879
880 /* first we must be within the limit */
881 limitagain:
882 if (physmemlimit != RUMPMEM_UNLIMITED) {
883 newmem = atomic_add_long_nv(&curphysmem, howmuch);
884 if (newmem > physmemlimit) {
885 newmem = atomic_add_long_nv(&curphysmem, -howmuch);
886 if (!waitok)
887 return NULL;
888 uvm_wait(wmsg);
889 goto limitagain;
890 }
891 }
892
893 /* second, we must get something from the backend */
894 again:
895 rv = rumpuser_malloc(howmuch, alignment);
896 if (__predict_false(rv == NULL && waitok)) {
897 uvm_wait(wmsg);
898 goto again;
899 }
900
901 return rv;
902 }
903
904 void
905 rump_hyperfree(void *what, size_t size)
906 {
907
908 if (physmemlimit != RUMPMEM_UNLIMITED) {
909 atomic_add_long(&curphysmem, -size);
910 }
911 rumpuser_free(what);
912 }
913