uvm_km.c revision 1.158
1/*	$NetBSD: uvm_km.c,v 1.158 2020/07/08 13:26:22 skrll Exp $	*/
2
3/*
4 * Copyright (c) 1997 Charles D. Cranor and Washington University.
5 * Copyright (c) 1991, 1993, The Regents of the University of California.
6 *
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
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 *	@(#)vm_kern.c   8.3 (Berkeley) 1/12/94
37 * from: Id: uvm_km.c,v 1.1.2.14 1998/02/06 05:19:27 chs Exp
38 *
39 *
40 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
41 * All rights reserved.
42 *
43 * Permission to use, copy, modify and distribute this software and
44 * its documentation is hereby granted, provided that both the copyright
45 * notice and this permission notice appear in all copies of the
46 * software, derivative works or modified versions, and any portions
47 * thereof, and that both notices appear in supporting documentation.
48 *
49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
50 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 *
53 * Carnegie Mellon requests users of this software to return to
54 *
55 *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
56 *  School of Computer Science
57 *  Carnegie Mellon University
58 *  Pittsburgh PA 15213-3890
59 *
60 * any improvements or extensions that they make and grant Carnegie the
61 * rights to redistribute these changes.
62 */
63
64/*
65 * uvm_km.c: handle kernel memory allocation and management
66 */
67
68/*
69 * overview of kernel memory management:
70 *
71 * the kernel virtual address space is mapped by "kernel_map."   kernel_map
72 * starts at VM_MIN_KERNEL_ADDRESS and goes to VM_MAX_KERNEL_ADDRESS.
73 * note that VM_MIN_KERNEL_ADDRESS is equal to vm_map_min(kernel_map).
74 *
75 * the kernel_map has several "submaps."   submaps can only appear in
76 * the kernel_map (user processes can't use them).   submaps "take over"
77 * the management of a sub-range of the kernel's address space.  submaps
78 * are typically allocated at boot time and are never released.   kernel
79 * virtual address space that is mapped by a submap is locked by the
80 * submap's lock -- not the kernel_map's lock.
81 *
82 * thus, the useful feature of submaps is that they allow us to break
83 * up the locking and protection of the kernel address space into smaller
84 * chunks.
85 *
86 * the vm system has several standard kernel submaps/arenas, including:
87 *   kmem_arena => used for kmem/pool (memoryallocators(9))
88 *   pager_map => used to map "buf" structures into kernel space
89 *   exec_map => used during exec to handle exec args
90 *   etc...
91 *
92 * The kmem_arena is a "special submap", as it lives in a fixed map entry
93 * within the kernel_map and is controlled by vmem(9).
94 *
95 * the kernel allocates its private memory out of special uvm_objects whose
96 * reference count is set to UVM_OBJ_KERN (thus indicating that the objects
97 * are "special" and never die).   all kernel objects should be thought of
98 * as large, fixed-sized, sparsely populated uvm_objects.   each kernel
99 * object is equal to the size of kernel virtual address space (i.e. the
100 * value "VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS").
101 *
102 * note that just because a kernel object spans the entire kernel virtual
103 * address space doesn't mean that it has to be mapped into the entire space.
104 * large chunks of a kernel object's space go unused either because
105 * that area of kernel VM is unmapped, or there is some other type of
106 * object mapped into that range (e.g. a vnode).    for submap's kernel
107 * objects, the only part of the object that can ever be populated is the
108 * offsets that are managed by the submap.
109 *
110 * note that the "offset" in a kernel object is always the kernel virtual
111 * address minus the VM_MIN_KERNEL_ADDRESS (aka vm_map_min(kernel_map)).
112 * example:
113 *   suppose VM_MIN_KERNEL_ADDRESS is 0xf8000000 and the kernel does a
114 *   uvm_km_alloc(kernel_map, PAGE_SIZE) [allocate 1 wired down page in the
115 *   kernel map].    if uvm_km_alloc returns virtual address 0xf8235000,
116 *   then that means that the page at offset 0x235000 in kernel_object is
117 *   mapped at 0xf8235000.
118 *
119 * kernel object have one other special property: when the kernel virtual
120 * memory mapping them is unmapped, the backing memory in the object is
121 * freed right away.   this is done with the uvm_km_pgremove() function.
122 * this has to be done because there is no backing store for kernel pages
123 * and no need to save them after they are no longer referenced.
124 *
125 * Generic arenas:
126 *
127 * kmem_arena:
128 *	Main arena controlling the kernel KVA used by other arenas.
129 *
130 * kmem_va_arena:
131 *	Implements quantum caching in order to speedup allocations and
132 *	reduce fragmentation.  The pool(9), unless created with a custom
133 *	meta-data allocator, and kmem(9) subsystems use this arena.
134 *
135 * Arenas for meta-data allocations are used by vmem(9) and pool(9).
136 * These arenas cannot use quantum cache.  However, kmem_va_meta_arena
137 * compensates this by importing larger chunks from kmem_arena.
138 *
139 * kmem_va_meta_arena:
140 *	Space for meta-data.
141 *
142 * kmem_meta_arena:
143 *	Imports from kmem_va_meta_arena.  Allocations from this arena are
144 *	backed with the pages.
145 *
146 * Arena stacking:
147 *
148 *	kmem_arena
149 *		kmem_va_arena
150 *		kmem_va_meta_arena
151 *			kmem_meta_arena
152 */
153
154#include <sys/cdefs.h>
155__KERNEL_RCSID(0, "$NetBSD: uvm_km.c,v 1.158 2020/07/08 13:26:22 skrll Exp $");
156
157#include "opt_uvmhist.h"
158
159#include "opt_kmempages.h"
160
161#ifndef NKMEMPAGES
162#define NKMEMPAGES 0
163#endif
164
165/*
166 * Defaults for lower and upper-bounds for the kmem_arena page count.
167 * Can be overridden by kernel config options.
168 */
169#ifndef NKMEMPAGES_MIN
170#define NKMEMPAGES_MIN NKMEMPAGES_MIN_DEFAULT
171#endif
172
173#ifndef NKMEMPAGES_MAX
174#define NKMEMPAGES_MAX NKMEMPAGES_MAX_DEFAULT
175#endif
176
177
178#include <sys/param.h>
179#include <sys/systm.h>
180#include <sys/atomic.h>
181#include <sys/proc.h>
182#include <sys/pool.h>
183#include <sys/vmem.h>
184#include <sys/vmem_impl.h>
185#include <sys/kmem.h>
186#include <sys/msan.h>
187
188#include <uvm/uvm.h>
189
190/*
191 * global data structures
192 */
193
194struct vm_map *kernel_map = NULL;
195
196/*
197 * local data structues
198 */
199
200static struct vm_map		kernel_map_store;
201static struct vm_map_entry	kernel_image_mapent_store;
202static struct vm_map_entry	kernel_kmem_mapent_store;
203
204int nkmempages = 0;
205vaddr_t kmembase;
206vsize_t kmemsize;
207
208static struct vmem kmem_arena_store;
209vmem_t *kmem_arena = NULL;
210static struct vmem kmem_va_arena_store;
211vmem_t *kmem_va_arena;
212
213/*
214 * kmeminit_nkmempages: calculate the size of kmem_arena.
215 */
216void
217kmeminit_nkmempages(void)
218{
219	int npages;
220
221	if (nkmempages != 0) {
222		/*
223		 * It's already been set (by us being here before)
224		 * bail out now;
225		 */
226		return;
227	}
228
229#if defined(KMSAN)
230	npages = (physmem / 8);
231#elif defined(PMAP_MAP_POOLPAGE)
232	npages = (physmem / 4);
233#else
234	npages = (physmem / 3) * 2;
235#endif /* defined(PMAP_MAP_POOLPAGE) */
236
237#ifndef NKMEMPAGES_MAX_UNLIMITED
238	if (npages > NKMEMPAGES_MAX)
239		npages = NKMEMPAGES_MAX;
240#endif
241
242	if (npages < NKMEMPAGES_MIN)
243		npages = NKMEMPAGES_MIN;
244
245	nkmempages = npages;
246}
247
248/*
249 * uvm_km_bootstrap: init kernel maps and objects to reflect reality (i.e.
250 * KVM already allocated for text, data, bss, and static data structures).
251 *
252 * => KVM is defined by VM_MIN_KERNEL_ADDRESS/VM_MAX_KERNEL_ADDRESS.
253 *    we assume that [vmin -> start] has already been allocated and that
254 *    "end" is the end.
255 */
256
257void
258uvm_km_bootstrap(vaddr_t start, vaddr_t end)
259{
260	bool kmem_arena_small;
261	vaddr_t base = VM_MIN_KERNEL_ADDRESS;
262	struct uvm_map_args args;
263	int error;
264
265	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
266	UVMHIST_LOG(maphist, "start=%#jx end=%#jx", start, end, 0,0);
267
268	kmeminit_nkmempages();
269	kmemsize = (vsize_t)nkmempages * PAGE_SIZE;
270	kmem_arena_small = kmemsize < 64 * 1024 * 1024;
271
272	UVMHIST_LOG(maphist, "kmemsize=%#jx", kmemsize, 0,0,0);
273
274	/*
275	 * next, init kernel memory objects.
276	 */
277
278	/* kernel_object: for pageable anonymous kernel memory */
279	uvm_kernel_object = uao_create(VM_MAX_KERNEL_ADDRESS -
280				VM_MIN_KERNEL_ADDRESS, UAO_FLAG_KERNOBJ);
281
282	/*
283	 * init the map and reserve any space that might already
284	 * have been allocated kernel space before installing.
285	 */
286
287	uvm_map_setup(&kernel_map_store, base, end, VM_MAP_PAGEABLE);
288	kernel_map_store.pmap = pmap_kernel();
289	if (start != base) {
290		error = uvm_map_prepare(&kernel_map_store,
291		    base, start - base,
292		    NULL, UVM_UNKNOWN_OFFSET, 0,
293		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
294		    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
295		if (!error) {
296			kernel_image_mapent_store.flags =
297			    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
298			error = uvm_map_enter(&kernel_map_store, &args,
299			    &kernel_image_mapent_store);
300		}
301
302		if (error)
303			panic(
304			    "uvm_km_bootstrap: could not reserve space for kernel");
305
306		kmembase = args.uma_start + args.uma_size;
307	} else {
308		kmembase = base;
309	}
310
311	error = uvm_map_prepare(&kernel_map_store,
312	    kmembase, kmemsize,
313	    NULL, UVM_UNKNOWN_OFFSET, 0,
314	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
315	    		UVM_ADV_RANDOM, UVM_FLAG_FIXED), &args);
316	if (!error) {
317		kernel_kmem_mapent_store.flags =
318		    UVM_MAP_KERNEL | UVM_MAP_STATIC | UVM_MAP_NOMERGE;
319		error = uvm_map_enter(&kernel_map_store, &args,
320		    &kernel_kmem_mapent_store);
321	}
322
323	if (error)
324		panic("uvm_km_bootstrap: could not reserve kernel kmem");
325
326	/*
327	 * install!
328	 */
329
330	kernel_map = &kernel_map_store;
331
332	pool_subsystem_init();
333
334	kmem_arena = vmem_init(&kmem_arena_store, "kmem",
335	    kmembase, kmemsize, PAGE_SIZE, NULL, NULL, NULL,
336	    0, VM_NOSLEEP | VM_BOOTSTRAP, IPL_VM);
337#ifdef PMAP_GROWKERNEL
338	/*
339	 * kmem_arena VA allocations happen independently of uvm_map.
340	 * grow kernel to accommodate the kmem_arena.
341	 */
342	if (uvm_maxkaddr < kmembase + kmemsize) {
343		uvm_maxkaddr = pmap_growkernel(kmembase + kmemsize);
344		KASSERTMSG(uvm_maxkaddr >= kmembase + kmemsize,
345		    "%#"PRIxVADDR" %#"PRIxVADDR" %#"PRIxVSIZE,
346		    uvm_maxkaddr, kmembase, kmemsize);
347	}
348#endif
349
350	vmem_subsystem_init(kmem_arena);
351
352	UVMHIST_LOG(maphist, "kmem vmem created (base=%#jx, size=%#jx",
353	    kmembase, kmemsize, 0,0);
354
355	kmem_va_arena = vmem_init(&kmem_va_arena_store, "kva",
356	    0, 0, PAGE_SIZE, vmem_alloc, vmem_free, kmem_arena,
357	    (kmem_arena_small ? 4 : VMEM_QCACHE_IDX_MAX) * PAGE_SIZE,
358	    VM_NOSLEEP, IPL_VM);
359
360	UVMHIST_LOG(maphist, "<- done", 0,0,0,0);
361}
362
363/*
364 * uvm_km_init: init the kernel maps virtual memory caches
365 * and start the pool/kmem allocator.
366 */
367void
368uvm_km_init(void)
369{
370	kmem_init();
371}
372
373/*
374 * uvm_km_suballoc: allocate a submap in the kernel map.   once a submap
375 * is allocated all references to that area of VM must go through it.  this
376 * allows the locking of VAs in kernel_map to be broken up into regions.
377 *
378 * => if `fixed' is true, *vmin specifies where the region described
379 *   pager_map => used to map "buf" structures into kernel space
380 *      by the submap must start
381 * => if submap is non NULL we use that as the submap, otherwise we
382 *	alloc a new map
383 */
384
385struct vm_map *
386uvm_km_suballoc(struct vm_map *map, vaddr_t *vmin /* IN/OUT */,
387    vaddr_t *vmax /* OUT */, vsize_t size, int flags, bool fixed,
388    struct vm_map *submap)
389{
390	int mapflags = UVM_FLAG_NOMERGE | (fixed ? UVM_FLAG_FIXED : 0);
391	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
392
393	KASSERT(vm_map_pmap(map) == pmap_kernel());
394
395	size = round_page(size);	/* round up to pagesize */
396
397	/*
398	 * first allocate a blank spot in the parent map
399	 */
400
401	if (uvm_map(map, vmin, size, NULL, UVM_UNKNOWN_OFFSET, 0,
402	    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
403	    UVM_ADV_RANDOM, mapflags)) != 0) {
404		panic("%s: unable to allocate space in parent map", __func__);
405	}
406
407	/*
408	 * set VM bounds (vmin is filled in by uvm_map)
409	 */
410
411	*vmax = *vmin + size;
412
413	/*
414	 * add references to pmap and create or init the submap
415	 */
416
417	pmap_reference(vm_map_pmap(map));
418	if (submap == NULL) {
419		submap = kmem_alloc(sizeof(*submap), KM_SLEEP);
420	}
421	uvm_map_setup(submap, *vmin, *vmax, flags);
422	submap->pmap = vm_map_pmap(map);
423
424	/*
425	 * now let uvm_map_submap plug in it...
426	 */
427
428	if (uvm_map_submap(map, *vmin, *vmax, submap) != 0)
429		panic("uvm_km_suballoc: submap allocation failed");
430
431	return(submap);
432}
433
434/*
435 * uvm_km_pgremove: remove pages from a kernel uvm_object and KVA.
436 */
437
438void
439uvm_km_pgremove(vaddr_t startva, vaddr_t endva)
440{
441	struct uvm_object * const uobj = uvm_kernel_object;
442	const voff_t start = startva - vm_map_min(kernel_map);
443	const voff_t end = endva - vm_map_min(kernel_map);
444	struct vm_page *pg;
445	voff_t curoff, nextoff;
446	int swpgonlydelta = 0;
447	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
448
449	KASSERT(VM_MIN_KERNEL_ADDRESS <= startva);
450	KASSERT(startva < endva);
451	KASSERT(endva <= VM_MAX_KERNEL_ADDRESS);
452
453	rw_enter(uobj->vmobjlock, RW_WRITER);
454	pmap_remove(pmap_kernel(), startva, endva);
455	for (curoff = start; curoff < end; curoff = nextoff) {
456		nextoff = curoff + PAGE_SIZE;
457		pg = uvm_pagelookup(uobj, curoff);
458		if (pg != NULL && pg->flags & PG_BUSY) {
459			uvm_pagewait(pg, uobj->vmobjlock, "km_pgrm");
460			rw_enter(uobj->vmobjlock, RW_WRITER);
461			nextoff = curoff;
462			continue;
463		}
464
465		/*
466		 * free the swap slot, then the page.
467		 */
468
469		if (pg == NULL &&
470		    uao_find_swslot(uobj, curoff >> PAGE_SHIFT) > 0) {
471			swpgonlydelta++;
472		}
473		uao_dropswap(uobj, curoff >> PAGE_SHIFT);
474		if (pg != NULL) {
475			uvm_pagefree(pg);
476		}
477	}
478	rw_exit(uobj->vmobjlock);
479
480	if (swpgonlydelta > 0) {
481		KASSERT(uvmexp.swpgonly >= swpgonlydelta);
482		atomic_add_int(&uvmexp.swpgonly, -swpgonlydelta);
483	}
484}
485
486
487/*
488 * uvm_km_pgremove_intrsafe: like uvm_km_pgremove(), but for non object backed
489 *    regions.
490 *
491 * => when you unmap a part of anonymous kernel memory you want to toss
492 *    the pages right away.    (this is called from uvm_unmap_...).
493 * => none of the pages will ever be busy, and none of them will ever
494 *    be on the active or inactive queues (because they have no object).
495 */
496
497void
498uvm_km_pgremove_intrsafe(struct vm_map *map, vaddr_t start, vaddr_t end)
499{
500#define __PGRM_BATCH 16
501	struct vm_page *pg;
502	paddr_t pa[__PGRM_BATCH];
503	int npgrm, i;
504	vaddr_t va, batch_vastart;
505
506	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
507
508	KASSERT(VM_MAP_IS_KERNEL(map));
509	KASSERTMSG(vm_map_min(map) <= start,
510	    "vm_map_min(map) [%#"PRIxVADDR"] <= start [%#"PRIxVADDR"]"
511	    " (size=%#"PRIxVSIZE")",
512	    vm_map_min(map), start, end - start);
513	KASSERT(start < end);
514	KASSERT(end <= vm_map_max(map));
515
516	for (va = start; va < end;) {
517		batch_vastart = va;
518		/* create a batch of at most __PGRM_BATCH pages to free */
519		for (i = 0;
520		     i < __PGRM_BATCH && va < end;
521		     va += PAGE_SIZE) {
522			if (!pmap_extract(pmap_kernel(), va, &pa[i])) {
523				continue;
524			}
525			i++;
526		}
527		npgrm = i;
528		/* now remove the mappings */
529		pmap_kremove(batch_vastart, va - batch_vastart);
530		/* and free the pages */
531		for (i = 0; i < npgrm; i++) {
532			pg = PHYS_TO_VM_PAGE(pa[i]);
533			KASSERT(pg);
534			KASSERT(pg->uobject == NULL && pg->uanon == NULL);
535			KASSERT((pg->flags & PG_BUSY) == 0);
536			uvm_pagefree(pg);
537		}
538	}
539#undef __PGRM_BATCH
540}
541
542#if defined(DEBUG)
543void
544uvm_km_check_empty(struct vm_map *map, vaddr_t start, vaddr_t end)
545{
546	vaddr_t va;
547	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
548
549	KDASSERT(VM_MAP_IS_KERNEL(map));
550	KDASSERT(vm_map_min(map) <= start);
551	KDASSERT(start < end);
552	KDASSERT(end <= vm_map_max(map));
553
554	for (va = start; va < end; va += PAGE_SIZE) {
555		paddr_t pa;
556
557		if (pmap_extract(pmap_kernel(), va, &pa)) {
558			panic("uvm_km_check_empty: va %p has pa %#llx",
559			    (void *)va, (long long)pa);
560		}
561		/*
562		 * kernel_object should not have pages for the corresponding
563		 * region.  check it.
564		 *
565		 * why trylock?  because:
566		 * - caller might not want to block.
567		 * - we can recurse when allocating radix_node for
568		 *   kernel_object.
569		 */
570		if (rw_tryenter(uvm_kernel_object->vmobjlock, RW_READER)) {
571			struct vm_page *pg;
572
573			pg = uvm_pagelookup(uvm_kernel_object,
574			    va - vm_map_min(kernel_map));
575			rw_exit(uvm_kernel_object->vmobjlock);
576			if (pg) {
577				panic("uvm_km_check_empty: "
578				    "has page hashed at %p",
579				    (const void *)va);
580			}
581		}
582	}
583}
584#endif /* defined(DEBUG) */
585
586/*
587 * uvm_km_alloc: allocate an area of kernel memory.
588 *
589 * => NOTE: we can return 0 even if we can wait if there is not enough
590 *	free VM space in the map... caller should be prepared to handle
591 *	this case.
592 * => we return KVA of memory allocated
593 */
594
595vaddr_t
596uvm_km_alloc(struct vm_map *map, vsize_t size, vsize_t align, uvm_flag_t flags)
597{
598	vaddr_t kva, loopva;
599	vaddr_t offset;
600	vsize_t loopsize;
601	struct vm_page *pg;
602	struct uvm_object *obj;
603	int pgaflags;
604	vm_prot_t prot, vaprot;
605	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
606
607	KASSERT(vm_map_pmap(map) == pmap_kernel());
608	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
609		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
610		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
611	KASSERT((flags & UVM_KMF_VAONLY) != 0 || (flags & UVM_KMF_COLORMATCH) == 0);
612	KASSERT((flags & UVM_KMF_COLORMATCH) == 0 || (flags & UVM_KMF_VAONLY) != 0);
613
614	/*
615	 * setup for call
616	 */
617
618	kva = vm_map_min(map);	/* hint */
619	size = round_page(size);
620	obj = (flags & UVM_KMF_PAGEABLE) ? uvm_kernel_object : NULL;
621	UVMHIST_LOG(maphist,"  (map=%#jx, obj=%#jx, size=%#jx, flags=%jd)",
622	    (uintptr_t)map, (uintptr_t)obj, size, flags);
623
624	/*
625	 * allocate some virtual space
626	 */
627
628	vaprot = (flags & UVM_KMF_EXEC) ? UVM_PROT_ALL : UVM_PROT_RW;
629	if (__predict_false(uvm_map(map, &kva, size, obj, UVM_UNKNOWN_OFFSET,
630	    align, UVM_MAPFLAG(vaprot, UVM_PROT_ALL, UVM_INH_NONE,
631	    UVM_ADV_RANDOM,
632	    (flags & (UVM_KMF_TRYLOCK | UVM_KMF_NOWAIT | UVM_KMF_WAITVA
633	     | UVM_KMF_COLORMATCH)))) != 0)) {
634		UVMHIST_LOG(maphist, "<- done (no VM)",0,0,0,0);
635		return(0);
636	}
637
638	/*
639	 * if all we wanted was VA, return now
640	 */
641
642	if (flags & (UVM_KMF_VAONLY | UVM_KMF_PAGEABLE)) {
643		UVMHIST_LOG(maphist,"<- done valloc (kva=%#jx)", kva,0,0,0);
644		return(kva);
645	}
646
647	/*
648	 * recover object offset from virtual address
649	 */
650
651	offset = kva - vm_map_min(kernel_map);
652	UVMHIST_LOG(maphist, "  kva=%#jx, offset=%#jx", kva, offset,0,0);
653
654	/*
655	 * now allocate and map in the memory... note that we are the only ones
656	 * whom should ever get a handle on this area of VM.
657	 */
658
659	loopva = kva;
660	loopsize = size;
661
662	pgaflags = UVM_FLAG_COLORMATCH;
663	if (flags & UVM_KMF_NOWAIT)
664		pgaflags |= UVM_PGA_USERESERVE;
665	if (flags & UVM_KMF_ZERO)
666		pgaflags |= UVM_PGA_ZERO;
667	prot = VM_PROT_READ | VM_PROT_WRITE;
668	if (flags & UVM_KMF_EXEC)
669		prot |= VM_PROT_EXECUTE;
670	while (loopsize) {
671		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, NULL),
672		    "loopva=%#"PRIxVADDR, loopva);
673
674		pg = uvm_pagealloc_strat(NULL, offset, NULL, pgaflags,
675#ifdef UVM_KM_VMFREELIST
676		   UVM_PGA_STRAT_ONLY, UVM_KM_VMFREELIST
677#else
678		   UVM_PGA_STRAT_NORMAL, 0
679#endif
680		   );
681
682		/*
683		 * out of memory?
684		 */
685
686		if (__predict_false(pg == NULL)) {
687			if ((flags & UVM_KMF_NOWAIT) ||
688			    ((flags & UVM_KMF_CANFAIL) && !uvm_reclaimable())) {
689				/* free everything! */
690				uvm_km_free(map, kva, size,
691				    flags & UVM_KMF_TYPEMASK);
692				return (0);
693			} else {
694				uvm_wait("km_getwait2");	/* sleep here */
695				continue;
696			}
697		}
698
699		pg->flags &= ~PG_BUSY;	/* new page */
700		UVM_PAGE_OWN(pg, NULL);
701
702		/*
703		 * map it in
704		 */
705
706		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
707		    prot, PMAP_KMPAGE);
708		loopva += PAGE_SIZE;
709		offset += PAGE_SIZE;
710		loopsize -= PAGE_SIZE;
711	}
712
713	pmap_update(pmap_kernel());
714
715	if ((flags & UVM_KMF_ZERO) == 0) {
716		kmsan_orig((void *)kva, size, KMSAN_TYPE_UVM, __RET_ADDR);
717		kmsan_mark((void *)kva, size, KMSAN_STATE_UNINIT);
718	}
719
720	UVMHIST_LOG(maphist,"<- done (kva=%#jx)", kva,0,0,0);
721	return(kva);
722}
723
724/*
725 * uvm_km_protect: change the protection of an allocated area
726 */
727
728int
729uvm_km_protect(struct vm_map *map, vaddr_t addr, vsize_t size, vm_prot_t prot)
730{
731	return uvm_map_protect(map, addr, addr + round_page(size), prot, false);
732}
733
734/*
735 * uvm_km_free: free an area of kernel memory
736 */
737
738void
739uvm_km_free(struct vm_map *map, vaddr_t addr, vsize_t size, uvm_flag_t flags)
740{
741	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
742
743	KASSERT((flags & UVM_KMF_TYPEMASK) == UVM_KMF_WIRED ||
744		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_PAGEABLE ||
745		(flags & UVM_KMF_TYPEMASK) == UVM_KMF_VAONLY);
746	KASSERT((addr & PAGE_MASK) == 0);
747	KASSERT(vm_map_pmap(map) == pmap_kernel());
748
749	size = round_page(size);
750
751	if (flags & UVM_KMF_PAGEABLE) {
752		uvm_km_pgremove(addr, addr + size);
753	} else if (flags & UVM_KMF_WIRED) {
754		/*
755		 * Note: uvm_km_pgremove_intrsafe() extracts mapping, thus
756		 * remove it after.  See comment below about KVA visibility.
757		 */
758		uvm_km_pgremove_intrsafe(map, addr, addr + size);
759	}
760
761	/*
762	 * Note: uvm_unmap_remove() calls pmap_update() for us, before
763	 * KVA becomes globally available.
764	 */
765
766	uvm_unmap1(map, addr, addr + size, UVM_FLAG_VAONLY);
767}
768
769/* Sanity; must specify both or none. */
770#if (defined(PMAP_MAP_POOLPAGE) || defined(PMAP_UNMAP_POOLPAGE)) && \
771    (!defined(PMAP_MAP_POOLPAGE) || !defined(PMAP_UNMAP_POOLPAGE))
772#error Must specify MAP and UNMAP together.
773#endif
774
775#if defined(PMAP_ALLOC_POOLPAGE) && \
776    !defined(PMAP_MAP_POOLPAGE) && !defined(PMAP_UNMAP_POOLPAGE)
777#error Must specify ALLOC with MAP and UNMAP
778#endif
779
780int
781uvm_km_kmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags,
782    vmem_addr_t *addr)
783{
784	struct vm_page *pg;
785	vmem_addr_t va;
786	int rc;
787	vaddr_t loopva;
788	vsize_t loopsize;
789
790	size = round_page(size);
791
792#if defined(PMAP_MAP_POOLPAGE)
793	if (size == PAGE_SIZE) {
794again:
795#ifdef PMAP_ALLOC_POOLPAGE
796		pg = PMAP_ALLOC_POOLPAGE((flags & VM_SLEEP) ?
797		   0 : UVM_PGA_USERESERVE);
798#else
799		pg = uvm_pagealloc(NULL, 0, NULL,
800		   (flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE);
801#endif /* PMAP_ALLOC_POOLPAGE */
802		if (__predict_false(pg == NULL)) {
803			if (flags & VM_SLEEP) {
804				uvm_wait("plpg");
805				goto again;
806			}
807			return ENOMEM;
808		}
809		va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg));
810		KASSERT(va != 0);
811		*addr = va;
812		return 0;
813	}
814#endif /* PMAP_MAP_POOLPAGE */
815
816	rc = vmem_alloc(vm, size, flags, &va);
817	if (rc != 0)
818		return rc;
819
820#ifdef PMAP_GROWKERNEL
821	/*
822	 * These VA allocations happen independently of uvm_map
823	 * so this allocation must not extend beyond the current limit.
824	 */
825	KASSERTMSG(uvm_maxkaddr >= va + size,
826	    "%#"PRIxVADDR" %#"PRIxPTR" %#zx",
827	    uvm_maxkaddr, va, size);
828#endif
829
830	loopva = va;
831	loopsize = size;
832
833	while (loopsize) {
834		paddr_t pa __diagused;
835		KASSERTMSG(!pmap_extract(pmap_kernel(), loopva, &pa),
836		    "loopva=%#"PRIxVADDR" loopsize=%#"PRIxVSIZE
837		    " pa=%#"PRIxPADDR" vmem=%p",
838		    loopva, loopsize, pa, vm);
839
840		pg = uvm_pagealloc(NULL, loopva, NULL,
841		    UVM_FLAG_COLORMATCH
842		    | ((flags & VM_SLEEP) ? 0 : UVM_PGA_USERESERVE));
843		if (__predict_false(pg == NULL)) {
844			if (flags & VM_SLEEP) {
845				uvm_wait("plpg");
846				continue;
847			} else {
848				uvm_km_pgremove_intrsafe(kernel_map, va,
849				    va + size);
850				vmem_free(vm, va, size);
851				return ENOMEM;
852			}
853		}
854
855		pg->flags &= ~PG_BUSY;	/* new page */
856		UVM_PAGE_OWN(pg, NULL);
857		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
858		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
859
860		loopva += PAGE_SIZE;
861		loopsize -= PAGE_SIZE;
862	}
863	pmap_update(pmap_kernel());
864
865	*addr = va;
866
867	return 0;
868}
869
870void
871uvm_km_kmem_free(vmem_t *vm, vmem_addr_t addr, size_t size)
872{
873
874	size = round_page(size);
875#if defined(PMAP_UNMAP_POOLPAGE)
876	if (size == PAGE_SIZE) {
877		paddr_t pa;
878
879		pa = PMAP_UNMAP_POOLPAGE(addr);
880		uvm_pagefree(PHYS_TO_VM_PAGE(pa));
881		return;
882	}
883#endif /* PMAP_UNMAP_POOLPAGE */
884	uvm_km_pgremove_intrsafe(kernel_map, addr, addr + size);
885	pmap_update(pmap_kernel());
886
887	vmem_free(vm, addr, size);
888}
889
890bool
891uvm_km_va_starved_p(void)
892{
893	vmem_size_t total;
894	vmem_size_t free;
895
896	if (kmem_arena == NULL)
897		return false;
898
899	total = vmem_size(kmem_arena, VMEM_ALLOC|VMEM_FREE);
900	free = vmem_size(kmem_arena, VMEM_FREE);
901
902	return (free < (total / 10));
903}
904