xref: /netbsd-src/sys/kern/subr_kmem.c (revision f3cfa6f6ce31685c6c4a758bc430e69eb99f50a4)
1 /*	$NetBSD: subr_kmem.c,v 1.75 2019/04/07 09:20:04 maxv Exp $	*/
2 
3 /*-
4  * Copyright (c) 2009-2015 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran and Maxime Villard.
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 NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*-
33  * Copyright (c)2006 YAMAMOTO Takashi,
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  *
45  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  */
57 
58 /*
59  * Allocator of kernel wired memory. This allocator has some debug features
60  * enabled with "option DIAGNOSTIC" and "option DEBUG".
61  */
62 
63 /*
64  * KMEM_SIZE: detect alloc/free size mismatch bugs.
65  *	Prefix each allocations with a fixed-sized, aligned header and record
66  *	the exact user-requested allocation size in it. When freeing, compare
67  *	it with kmem_free's "size" argument.
68  *
69  * This option enabled on DIAGNOSTIC.
70  *
71  *  |CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|CHUNK|
72  *  +-----+-----+-----+-----+-----+-----+-----+-----+-----+---+-+
73  *  |/////|     |     |     |     |     |     |     |     |   |U|
74  *  |/HSZ/|     |     |     |     |     |     |     |     |   |U|
75  *  |/////|     |     |     |     |     |     |     |     |   |U|
76  *  +-----+-----+-----+-----+-----+-----+-----+-----+-----+---+-+
77  *  |Size |    Buffer usable by the caller (requested size)   |Unused\
78  */
79 
80 /*
81  * KMEM_GUARD
82  *	A kernel with "option DEBUG" has "kmem_guard" debugging feature compiled
83  *	in. See the comment below for what kind of bugs it tries to detect. Even
84  *	if compiled in, it's disabled by default because it's very expensive.
85  *	You can enable it on boot by:
86  *		boot -d
87  *		db> w kmem_guard_depth 0t30000
88  *		db> c
89  *
90  *	The default value of kmem_guard_depth is 0, which means disabled.
91  *	It can be changed by KMEM_GUARD_DEPTH kernel config option.
92  */
93 
94 #include <sys/cdefs.h>
95 __KERNEL_RCSID(0, "$NetBSD: subr_kmem.c,v 1.75 2019/04/07 09:20:04 maxv Exp $");
96 
97 #ifdef _KERNEL_OPT
98 #include "opt_kmem.h"
99 #endif
100 
101 #include <sys/param.h>
102 #include <sys/callback.h>
103 #include <sys/kmem.h>
104 #include <sys/pool.h>
105 #include <sys/debug.h>
106 #include <sys/lockdebug.h>
107 #include <sys/cpu.h>
108 #include <sys/asan.h>
109 
110 #include <uvm/uvm_extern.h>
111 #include <uvm/uvm_map.h>
112 
113 #include <lib/libkern/libkern.h>
114 
115 struct kmem_cache_info {
116 	size_t		kc_size;
117 	const char *	kc_name;
118 };
119 
120 static const struct kmem_cache_info kmem_cache_sizes[] = {
121 	{  8, "kmem-8" },
122 	{ 16, "kmem-16" },
123 	{ 24, "kmem-24" },
124 	{ 32, "kmem-32" },
125 	{ 40, "kmem-40" },
126 	{ 48, "kmem-48" },
127 	{ 56, "kmem-56" },
128 	{ 64, "kmem-64" },
129 	{ 80, "kmem-80" },
130 	{ 96, "kmem-96" },
131 	{ 112, "kmem-112" },
132 	{ 128, "kmem-128" },
133 	{ 160, "kmem-160" },
134 	{ 192, "kmem-192" },
135 	{ 224, "kmem-224" },
136 	{ 256, "kmem-256" },
137 	{ 320, "kmem-320" },
138 	{ 384, "kmem-384" },
139 	{ 448, "kmem-448" },
140 	{ 512, "kmem-512" },
141 	{ 768, "kmem-768" },
142 	{ 1024, "kmem-1024" },
143 	{ 0, NULL }
144 };
145 
146 static const struct kmem_cache_info kmem_cache_big_sizes[] = {
147 	{ 2048, "kmem-2048" },
148 	{ 4096, "kmem-4096" },
149 	{ 8192, "kmem-8192" },
150 	{ 16384, "kmem-16384" },
151 	{ 0, NULL }
152 };
153 
154 /*
155  * KMEM_ALIGN is the smallest guaranteed alignment and also the
156  * smallest allocateable quantum.
157  * Every cache size >= CACHE_LINE_SIZE gets CACHE_LINE_SIZE alignment.
158  */
159 #define	KMEM_ALIGN		8
160 #define	KMEM_SHIFT		3
161 #define	KMEM_MAXSIZE		1024
162 #define	KMEM_CACHE_COUNT	(KMEM_MAXSIZE >> KMEM_SHIFT)
163 
164 static pool_cache_t kmem_cache[KMEM_CACHE_COUNT] __cacheline_aligned;
165 static size_t kmem_cache_maxidx __read_mostly;
166 
167 #define	KMEM_BIG_ALIGN		2048
168 #define	KMEM_BIG_SHIFT		11
169 #define	KMEM_BIG_MAXSIZE	16384
170 #define	KMEM_CACHE_BIG_COUNT	(KMEM_BIG_MAXSIZE >> KMEM_BIG_SHIFT)
171 
172 static pool_cache_t kmem_cache_big[KMEM_CACHE_BIG_COUNT] __cacheline_aligned;
173 static size_t kmem_cache_big_maxidx __read_mostly;
174 
175 #if defined(DIAGNOSTIC) && defined(_HARDKERNEL)
176 #define	KMEM_SIZE
177 #endif
178 
179 #if defined(DEBUG) && defined(_HARDKERNEL)
180 #define	KMEM_SIZE
181 #define	KMEM_GUARD
182 static void *kmem_freecheck;
183 #endif
184 
185 #if defined(KMEM_SIZE)
186 struct kmem_header {
187 	size_t		size;
188 } __aligned(KMEM_ALIGN);
189 #define	SIZE_SIZE	sizeof(struct kmem_header)
190 static void kmem_size_set(void *, size_t);
191 static void kmem_size_check(void *, size_t);
192 #else
193 #define	SIZE_SIZE	0
194 #define	kmem_size_set(p, sz)	/* nothing */
195 #define	kmem_size_check(p, sz)	/* nothing */
196 #endif
197 
198 #if defined(KMEM_GUARD)
199 #ifndef KMEM_GUARD_DEPTH
200 #define KMEM_GUARD_DEPTH 0
201 #endif
202 struct kmem_guard {
203 	u_int		kg_depth;
204 	intptr_t *	kg_fifo;
205 	u_int		kg_rotor;
206 	vmem_t *	kg_vmem;
207 };
208 static bool kmem_guard_init(struct kmem_guard *, u_int, vmem_t *);
209 static void *kmem_guard_alloc(struct kmem_guard *, size_t, bool);
210 static void kmem_guard_free(struct kmem_guard *, size_t, void *);
211 int kmem_guard_depth = KMEM_GUARD_DEPTH;
212 static bool kmem_guard_enabled;
213 static struct kmem_guard kmem_guard;
214 #endif /* defined(KMEM_GUARD) */
215 
216 CTASSERT(KM_SLEEP == PR_WAITOK);
217 CTASSERT(KM_NOSLEEP == PR_NOWAIT);
218 
219 /*
220  * kmem_intr_alloc: allocate wired memory.
221  */
222 
223 void *
224 kmem_intr_alloc(size_t requested_size, km_flag_t kmflags)
225 {
226 #ifdef KASAN
227 	const size_t origsize = requested_size;
228 #endif
229 	size_t allocsz, index;
230 	size_t size;
231 	pool_cache_t pc;
232 	uint8_t *p;
233 
234 	KASSERT(requested_size > 0);
235 
236 	KASSERT((kmflags & KM_SLEEP) || (kmflags & KM_NOSLEEP));
237 	KASSERT(!(kmflags & KM_SLEEP) || !(kmflags & KM_NOSLEEP));
238 
239 #ifdef KMEM_GUARD
240 	if (kmem_guard_enabled) {
241 		return kmem_guard_alloc(&kmem_guard, requested_size,
242 		    (kmflags & KM_SLEEP) != 0);
243 	}
244 #endif
245 
246 	kasan_add_redzone(&requested_size);
247 	size = kmem_roundup_size(requested_size);
248 	allocsz = size + SIZE_SIZE;
249 
250 	if ((index = ((allocsz -1) >> KMEM_SHIFT))
251 	    < kmem_cache_maxidx) {
252 		pc = kmem_cache[index];
253 	} else if ((index = ((allocsz - 1) >> KMEM_BIG_SHIFT))
254 	    < kmem_cache_big_maxidx) {
255 		pc = kmem_cache_big[index];
256 	} else {
257 		int ret = uvm_km_kmem_alloc(kmem_va_arena,
258 		    (vsize_t)round_page(size),
259 		    ((kmflags & KM_SLEEP) ? VM_SLEEP : VM_NOSLEEP)
260 		     | VM_INSTANTFIT, (vmem_addr_t *)&p);
261 		if (ret) {
262 			return NULL;
263 		}
264 		FREECHECK_OUT(&kmem_freecheck, p);
265 		return p;
266 	}
267 
268 	p = pool_cache_get(pc, kmflags);
269 
270 	if (__predict_true(p != NULL)) {
271 		FREECHECK_OUT(&kmem_freecheck, p);
272 		kmem_size_set(p, requested_size);
273 		p += SIZE_SIZE;
274 		kasan_mark(p, origsize, size, KASAN_KMEM_REDZONE);
275 		return p;
276 	}
277 	return p;
278 }
279 
280 /*
281  * kmem_intr_zalloc: allocate zeroed wired memory.
282  */
283 
284 void *
285 kmem_intr_zalloc(size_t size, km_flag_t kmflags)
286 {
287 	void *p;
288 
289 	p = kmem_intr_alloc(size, kmflags);
290 	if (p != NULL) {
291 		memset(p, 0, size);
292 	}
293 	return p;
294 }
295 
296 /*
297  * kmem_intr_free: free wired memory allocated by kmem_alloc.
298  */
299 
300 void
301 kmem_intr_free(void *p, size_t requested_size)
302 {
303 	size_t allocsz, index;
304 	size_t size;
305 	pool_cache_t pc;
306 
307 	KASSERT(p != NULL);
308 	KASSERT(requested_size > 0);
309 
310 #ifdef KMEM_GUARD
311 	if (kmem_guard_enabled) {
312 		kmem_guard_free(&kmem_guard, requested_size, p);
313 		return;
314 	}
315 #endif
316 
317 	kasan_add_redzone(&requested_size);
318 	size = kmem_roundup_size(requested_size);
319 	allocsz = size + SIZE_SIZE;
320 
321 	if ((index = ((allocsz -1) >> KMEM_SHIFT))
322 	    < kmem_cache_maxidx) {
323 		pc = kmem_cache[index];
324 	} else if ((index = ((allocsz - 1) >> KMEM_BIG_SHIFT))
325 	    < kmem_cache_big_maxidx) {
326 		pc = kmem_cache_big[index];
327 	} else {
328 		FREECHECK_IN(&kmem_freecheck, p);
329 		uvm_km_kmem_free(kmem_va_arena, (vaddr_t)p,
330 		    round_page(size));
331 		return;
332 	}
333 
334 	kasan_mark(p, size, size, 0);
335 
336 	p = (uint8_t *)p - SIZE_SIZE;
337 	kmem_size_check(p, requested_size);
338 	FREECHECK_IN(&kmem_freecheck, p);
339 	LOCKDEBUG_MEM_CHECK(p, size);
340 
341 	pool_cache_put(pc, p);
342 }
343 
344 /* ---- kmem API */
345 
346 /*
347  * kmem_alloc: allocate wired memory.
348  * => must not be called from interrupt context.
349  */
350 
351 void *
352 kmem_alloc(size_t size, km_flag_t kmflags)
353 {
354 	void *v;
355 
356 	KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()),
357 	    "kmem(9) should not be used from the interrupt context");
358 	v = kmem_intr_alloc(size, kmflags);
359 	KASSERT(v || (kmflags & KM_NOSLEEP) != 0);
360 	return v;
361 }
362 
363 /*
364  * kmem_zalloc: allocate zeroed wired memory.
365  * => must not be called from interrupt context.
366  */
367 
368 void *
369 kmem_zalloc(size_t size, km_flag_t kmflags)
370 {
371 	void *v;
372 
373 	KASSERTMSG((!cpu_intr_p() && !cpu_softintr_p()),
374 	    "kmem(9) should not be used from the interrupt context");
375 	v = kmem_intr_zalloc(size, kmflags);
376 	KASSERT(v || (kmflags & KM_NOSLEEP) != 0);
377 	return v;
378 }
379 
380 /*
381  * kmem_free: free wired memory allocated by kmem_alloc.
382  * => must not be called from interrupt context.
383  */
384 
385 void
386 kmem_free(void *p, size_t size)
387 {
388 	KASSERT(!cpu_intr_p());
389 	KASSERT(!cpu_softintr_p());
390 	kmem_intr_free(p, size);
391 }
392 
393 static size_t
394 kmem_create_caches(const struct kmem_cache_info *array,
395     pool_cache_t alloc_table[], size_t maxsize, int shift, int ipl)
396 {
397 	size_t maxidx = 0;
398 	size_t table_unit = (1 << shift);
399 	size_t size = table_unit;
400 	int i;
401 
402 	for (i = 0; array[i].kc_size != 0 ; i++) {
403 		const char *name = array[i].kc_name;
404 		size_t cache_size = array[i].kc_size;
405 		struct pool_allocator *pa;
406 		int flags = 0;
407 		pool_cache_t pc;
408 		size_t align;
409 
410 		if ((cache_size & (CACHE_LINE_SIZE - 1)) == 0)
411 			align = CACHE_LINE_SIZE;
412 		else if ((cache_size & (PAGE_SIZE - 1)) == 0)
413 			align = PAGE_SIZE;
414 		else
415 			align = KMEM_ALIGN;
416 
417 		if (cache_size < CACHE_LINE_SIZE)
418 			flags |= PR_NOTOUCH;
419 
420 		/* check if we reached the requested size */
421 		if (cache_size > maxsize || cache_size > PAGE_SIZE) {
422 			break;
423 		}
424 		if ((cache_size >> shift) > maxidx) {
425 			maxidx = cache_size >> shift;
426 		}
427 
428 		if ((cache_size >> shift) > maxidx) {
429 			maxidx = cache_size >> shift;
430 		}
431 
432 		pa = &pool_allocator_kmem;
433 		pc = pool_cache_init(cache_size, align, 0, flags,
434 		    name, pa, ipl, NULL, NULL, NULL);
435 
436 		while (size <= cache_size) {
437 			alloc_table[(size - 1) >> shift] = pc;
438 			size += table_unit;
439 		}
440 	}
441 	return maxidx;
442 }
443 
444 void
445 kmem_init(void)
446 {
447 #ifdef KMEM_GUARD
448 	kmem_guard_enabled = kmem_guard_init(&kmem_guard, kmem_guard_depth,
449 	    kmem_va_arena);
450 #endif
451 	kmem_cache_maxidx = kmem_create_caches(kmem_cache_sizes,
452 	    kmem_cache, KMEM_MAXSIZE, KMEM_SHIFT, IPL_VM);
453 	kmem_cache_big_maxidx = kmem_create_caches(kmem_cache_big_sizes,
454 	    kmem_cache_big, PAGE_SIZE, KMEM_BIG_SHIFT, IPL_VM);
455 }
456 
457 size_t
458 kmem_roundup_size(size_t size)
459 {
460 	return (size + (KMEM_ALIGN - 1)) & ~(KMEM_ALIGN - 1);
461 }
462 
463 /*
464  * Used to dynamically allocate string with kmem accordingly to format.
465  */
466 char *
467 kmem_asprintf(const char *fmt, ...)
468 {
469 	int size __diagused, len;
470 	va_list va;
471 	char *str;
472 
473 	va_start(va, fmt);
474 	len = vsnprintf(NULL, 0, fmt, va);
475 	va_end(va);
476 
477 	str = kmem_alloc(len + 1, KM_SLEEP);
478 
479 	va_start(va, fmt);
480 	size = vsnprintf(str, len + 1, fmt, va);
481 	va_end(va);
482 
483 	KASSERT(size == len);
484 
485 	return str;
486 }
487 
488 char *
489 kmem_strdupsize(const char *str, size_t *lenp, km_flag_t flags)
490 {
491 	size_t len = strlen(str) + 1;
492 	char *ptr = kmem_alloc(len, flags);
493 	if (ptr == NULL)
494 		return NULL;
495 
496 	if (lenp)
497 		*lenp = len;
498 	memcpy(ptr, str, len);
499 	return ptr;
500 }
501 
502 char *
503 kmem_strndup(const char *str, size_t maxlen, km_flag_t flags)
504 {
505 	KASSERT(str != NULL);
506 	KASSERT(maxlen != 0);
507 
508 	size_t len = strnlen(str, maxlen);
509 	char *ptr = kmem_alloc(len + 1, flags);
510 	if (ptr == NULL)
511 		return NULL;
512 
513 	memcpy(ptr, str, len);
514 	ptr[len] = '\0';
515 
516 	return ptr;
517 }
518 
519 void
520 kmem_strfree(char *str)
521 {
522 	if (str == NULL)
523 		return;
524 
525 	kmem_free(str, strlen(str) + 1);
526 }
527 
528 /* ------------------ DEBUG / DIAGNOSTIC ------------------ */
529 
530 #if defined(KMEM_SIZE)
531 static void
532 kmem_size_set(void *p, size_t sz)
533 {
534 	struct kmem_header *hd;
535 	hd = (struct kmem_header *)p;
536 	hd->size = sz;
537 }
538 
539 static void
540 kmem_size_check(void *p, size_t sz)
541 {
542 	struct kmem_header *hd;
543 	size_t hsz;
544 
545 	hd = (struct kmem_header *)p;
546 	hsz = hd->size;
547 
548 	if (hsz != sz) {
549 		panic("kmem_free(%p, %zu) != allocated size %zu",
550 		    (const uint8_t *)p + SIZE_SIZE, sz, hsz);
551 	}
552 
553 	hd->size = -1;
554 }
555 #endif /* defined(KMEM_SIZE) */
556 
557 #if defined(KMEM_GUARD)
558 /*
559  * The ultimate memory allocator for debugging, baby.  It tries to catch:
560  *
561  * 1. Overflow, in realtime. A guard page sits immediately after the
562  *    requested area; a read/write overflow therefore triggers a page
563  *    fault.
564  * 2. Invalid pointer/size passed, at free. A kmem_header structure sits
565  *    just before the requested area, and holds the allocated size. Any
566  *    difference with what is given at free triggers a panic.
567  * 3. Underflow, at free. If an underflow occurs, the kmem header will be
568  *    modified, and 2. will trigger a panic.
569  * 4. Use-after-free. When freeing, the memory is unmapped, and depending
570  *    on the value of kmem_guard_depth, the kernel will more or less delay
571  *    the recycling of that memory. Which means that any ulterior read/write
572  *    access to the memory will trigger a page fault, given it hasn't been
573  *    recycled yet.
574  */
575 
576 #include <sys/atomic.h>
577 #include <uvm/uvm.h>
578 
579 static bool
580 kmem_guard_init(struct kmem_guard *kg, u_int depth, vmem_t *vm)
581 {
582 	vaddr_t va;
583 
584 	/* If not enabled, we have nothing to do. */
585 	if (depth == 0) {
586 		return false;
587 	}
588 	depth = roundup(depth, PAGE_SIZE / sizeof(void *));
589 	KASSERT(depth != 0);
590 
591 	/*
592 	 * Allocate fifo.
593 	 */
594 	va = uvm_km_alloc(kernel_map, depth * sizeof(void *), PAGE_SIZE,
595 	    UVM_KMF_WIRED | UVM_KMF_ZERO);
596 	if (va == 0) {
597 		return false;
598 	}
599 
600 	/*
601 	 * Init object.
602 	 */
603 	kg->kg_vmem = vm;
604 	kg->kg_fifo = (void *)va;
605 	kg->kg_depth = depth;
606 	kg->kg_rotor = 0;
607 
608 	printf("kmem_guard(%p): depth %d\n", kg, depth);
609 	return true;
610 }
611 
612 static void *
613 kmem_guard_alloc(struct kmem_guard *kg, size_t requested_size, bool waitok)
614 {
615 	struct vm_page *pg;
616 	vm_flag_t flags;
617 	vmem_addr_t va;
618 	vaddr_t loopva;
619 	vsize_t loopsize;
620 	size_t size;
621 	void **p;
622 
623 	/*
624 	 * Compute the size: take the kmem header into account, and add a guard
625 	 * page at the end.
626 	 */
627 	size = round_page(requested_size + SIZE_SIZE) + PAGE_SIZE;
628 
629 	/* Allocate pages of kernel VA, but do not map anything in yet. */
630 	flags = VM_BESTFIT | (waitok ? VM_SLEEP : VM_NOSLEEP);
631 	if (vmem_alloc(kg->kg_vmem, size, flags, &va) != 0) {
632 		return NULL;
633 	}
634 
635 	loopva = va;
636 	loopsize = size - PAGE_SIZE;
637 
638 	while (loopsize) {
639 		pg = uvm_pagealloc(NULL, loopva, NULL, 0);
640 		if (__predict_false(pg == NULL)) {
641 			if (waitok) {
642 				uvm_wait("kmem_guard");
643 				continue;
644 			} else {
645 				uvm_km_pgremove_intrsafe(kernel_map, va,
646 				    va + size);
647 				vmem_free(kg->kg_vmem, va, size);
648 				return NULL;
649 			}
650 		}
651 
652 		pg->flags &= ~PG_BUSY;	/* new page */
653 		UVM_PAGE_OWN(pg, NULL);
654 		pmap_kenter_pa(loopva, VM_PAGE_TO_PHYS(pg),
655 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
656 
657 		loopva += PAGE_SIZE;
658 		loopsize -= PAGE_SIZE;
659 	}
660 
661 	pmap_update(pmap_kernel());
662 
663 	/*
664 	 * Offset the returned pointer so that the unmapped guard page sits
665 	 * immediately after the returned object.
666 	 */
667 	p = (void **)((va + (size - PAGE_SIZE) - requested_size) & ~(uintptr_t)ALIGNBYTES);
668 	kmem_size_set((uint8_t *)p - SIZE_SIZE, requested_size);
669 	return (void *)p;
670 }
671 
672 static void
673 kmem_guard_free(struct kmem_guard *kg, size_t requested_size, void *p)
674 {
675 	vaddr_t va;
676 	u_int rotor;
677 	size_t size;
678 	uint8_t *ptr;
679 
680 	ptr = (uint8_t *)p - SIZE_SIZE;
681 	kmem_size_check(ptr, requested_size);
682 	va = trunc_page((vaddr_t)ptr);
683 	size = round_page(requested_size + SIZE_SIZE) + PAGE_SIZE;
684 
685 	KASSERT(pmap_extract(pmap_kernel(), va, NULL));
686 	KASSERT(!pmap_extract(pmap_kernel(), va + (size - PAGE_SIZE), NULL));
687 
688 	/*
689 	 * Unmap and free the pages. The last one is never allocated.
690 	 */
691 	uvm_km_pgremove_intrsafe(kernel_map, va, va + size);
692 	pmap_update(pmap_kernel());
693 
694 #if 0
695 	/*
696 	 * XXX: Here, we need to atomically register the va and its size in the
697 	 * fifo.
698 	 */
699 
700 	/*
701 	 * Put the VA allocation into the list and swap an old one out to free.
702 	 * This behaves mostly like a fifo.
703 	 */
704 	rotor = atomic_inc_uint_nv(&kg->kg_rotor) % kg->kg_depth;
705 	va = (vaddr_t)atomic_swap_ptr(&kg->kg_fifo[rotor], (void *)va);
706 	if (va != 0) {
707 		vmem_free(kg->kg_vmem, va, size);
708 	}
709 #else
710 	(void)rotor;
711 	vmem_free(kg->kg_vmem, va, size);
712 #endif
713 }
714 
715 #endif /* defined(KMEM_GUARD) */
716