xref: /freebsd/sys/vm/memguard.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2005, Bosko Milekic <bmilekic@FreeBSD.org>.
5  * Copyright (c) 2010 Isilon Systems, Inc. (http://www.isilon.com/)
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice unmodified, this list of conditions, and the following
13  *    disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 /*
32  * MemGuard is a simple replacement allocator for debugging only
33  * which provides ElectricFence-style memory barrier protection on
34  * objects being allocated, and is used to detect tampering-after-free
35  * scenarios.
36  *
37  * See the memguard(9) man page for more information on using MemGuard.
38  */
39 
40 #include "opt_vm.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/types.h>
46 #include <sys/queue.h>
47 #include <sys/lock.h>
48 #include <sys/mutex.h>
49 #include <sys/malloc.h>
50 #include <sys/sysctl.h>
51 #include <sys/vmem.h>
52 #include <sys/vmmeter.h>
53 
54 #include <vm/vm.h>
55 #include <vm/uma.h>
56 #include <vm/vm_param.h>
57 #include <vm/vm_page.h>
58 #include <vm/vm_map.h>
59 #include <vm/vm_object.h>
60 #include <vm/vm_kern.h>
61 #include <vm/vm_extern.h>
62 #include <vm/uma_int.h>
63 #include <vm/memguard.h>
64 
65 static SYSCTL_NODE(_vm, OID_AUTO, memguard, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
66     "MemGuard data");
67 /*
68  * The vm_memguard_divisor variable controls how much of kernel_arena should be
69  * reserved for MemGuard.
70  */
71 static u_int vm_memguard_divisor;
72 SYSCTL_UINT(_vm_memguard, OID_AUTO, divisor, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
73     &vm_memguard_divisor,
74     0, "(kmem_size/memguard_divisor) == memguard submap size");
75 
76 /*
77  * Short description (ks_shortdesc) of memory type to monitor.
78  */
79 static char vm_memguard_desc[128] = "";
80 static struct malloc_type *vm_memguard_mtype = NULL;
81 TUNABLE_STR("vm.memguard.desc", vm_memguard_desc, sizeof(vm_memguard_desc));
82 static int
83 memguard_sysctl_desc(SYSCTL_HANDLER_ARGS)
84 {
85 	char desc[sizeof(vm_memguard_desc)];
86 	int error;
87 
88 	strlcpy(desc, vm_memguard_desc, sizeof(desc));
89 	error = sysctl_handle_string(oidp, desc, sizeof(desc), req);
90 	if (error != 0 || req->newptr == NULL)
91 		return (error);
92 
93 	mtx_lock(&malloc_mtx);
94 	/* If mtp is NULL, it will be initialized in memguard_cmp() */
95 	vm_memguard_mtype = malloc_desc2type(desc);
96 	strlcpy(vm_memguard_desc, desc, sizeof(vm_memguard_desc));
97 	mtx_unlock(&malloc_mtx);
98 	return (error);
99 }
100 SYSCTL_PROC(_vm_memguard, OID_AUTO, desc,
101     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
102     memguard_sysctl_desc, "A", "Short description of memory type to monitor");
103 
104 static int
105 memguard_sysctl_mapused(SYSCTL_HANDLER_ARGS)
106 {
107 	vmem_size_t size;
108 
109 	size = vmem_size(memguard_arena, VMEM_ALLOC);
110 	return (sysctl_handle_long(oidp, &size, sizeof(size), req));
111 }
112 
113 static vm_offset_t memguard_base;
114 static vm_size_t memguard_mapsize;
115 static vm_size_t memguard_physlimit;
116 static u_long memguard_wasted;
117 static u_long memguard_succ;
118 static u_long memguard_fail_kva;
119 static u_long memguard_fail_pgs;
120 
121 SYSCTL_ULONG(_vm_memguard, OID_AUTO, mapsize, CTLFLAG_RD,
122     &memguard_mapsize, 0, "MemGuard private arena size");
123 SYSCTL_ULONG(_vm_memguard, OID_AUTO, phys_limit, CTLFLAG_RD,
124     &memguard_physlimit, 0, "Limit on MemGuard memory consumption");
125 SYSCTL_ULONG(_vm_memguard, OID_AUTO, wasted, CTLFLAG_RD,
126     &memguard_wasted, 0, "Excess memory used through page promotion");
127 SYSCTL_ULONG(_vm_memguard, OID_AUTO, numalloc, CTLFLAG_RD,
128     &memguard_succ, 0, "Count of successful MemGuard allocations");
129 SYSCTL_ULONG(_vm_memguard, OID_AUTO, fail_kva, CTLFLAG_RD,
130     &memguard_fail_kva, 0, "MemGuard failures due to lack of KVA");
131 SYSCTL_ULONG(_vm_memguard, OID_AUTO, fail_pgs, CTLFLAG_RD,
132     &memguard_fail_pgs, 0, "MemGuard failures due to lack of pages");
133 
134 #define MG_GUARD_AROUND		0x001
135 #define MG_GUARD_ALLLARGE	0x002
136 #define MG_GUARD_NOFREE		0x004
137 static int memguard_options = MG_GUARD_AROUND;
138 SYSCTL_INT(_vm_memguard, OID_AUTO, options, CTLFLAG_RWTUN,
139     &memguard_options, 0,
140     "MemGuard options:\n"
141     "\t0x001 - add guard pages around each allocation\n"
142     "\t0x002 - always use MemGuard for allocations over a page\n"
143     "\t0x004 - guard uma(9) zones with UMA_ZONE_NOFREE flag");
144 
145 static u_int memguard_minsize;
146 static u_long memguard_minsize_reject;
147 SYSCTL_UINT(_vm_memguard, OID_AUTO, minsize, CTLFLAG_RW,
148     &memguard_minsize, 0, "Minimum size for page promotion");
149 SYSCTL_ULONG(_vm_memguard, OID_AUTO, minsize_reject, CTLFLAG_RD,
150     &memguard_minsize_reject, 0, "# times rejected for size");
151 
152 static u_int memguard_frequency;
153 static u_long memguard_frequency_hits;
154 SYSCTL_UINT(_vm_memguard, OID_AUTO, frequency, CTLFLAG_RWTUN,
155     &memguard_frequency, 0, "Times in 100000 that MemGuard will randomly run");
156 SYSCTL_ULONG(_vm_memguard, OID_AUTO, frequency_hits, CTLFLAG_RD,
157     &memguard_frequency_hits, 0, "# times MemGuard randomly chose");
158 
159 /*
160  * Return a fudged value to be used for vm_kmem_size for allocating
161  * the kernel_arena.
162  */
163 unsigned long
164 memguard_fudge(unsigned long km_size, const struct vm_map *parent_map)
165 {
166 	u_long mem_pgs, parent_size;
167 
168 	vm_memguard_divisor = 10;
169 	/* CTFLAG_RDTUN doesn't work during the early boot process. */
170 	TUNABLE_INT_FETCH("vm.memguard.divisor", &vm_memguard_divisor);
171 
172 	parent_size = vm_map_max(parent_map) - vm_map_min(parent_map) +
173 	    PAGE_SIZE;
174 	/* Pick a conservative value if provided value sucks. */
175 	if ((vm_memguard_divisor <= 0) ||
176 	    ((parent_size / vm_memguard_divisor) == 0))
177 		vm_memguard_divisor = 10;
178 	/*
179 	 * Limit consumption of physical pages to
180 	 * 1/vm_memguard_divisor of system memory.  If the KVA is
181 	 * smaller than this then the KVA limit comes into play first.
182 	 * This prevents memguard's page promotions from completely
183 	 * using up memory, since most malloc(9) calls are sub-page.
184 	 */
185 	mem_pgs = vm_cnt.v_page_count;
186 	memguard_physlimit = (mem_pgs / vm_memguard_divisor) * PAGE_SIZE;
187 	/*
188 	 * We want as much KVA as we can take safely.  Use at most our
189 	 * allotted fraction of the parent map's size.  Limit this to
190 	 * twice the physical memory to avoid using too much memory as
191 	 * pagetable pages (size must be multiple of PAGE_SIZE).
192 	 */
193 	memguard_mapsize = round_page(parent_size / vm_memguard_divisor);
194 	if (memguard_mapsize / (2 * PAGE_SIZE) > mem_pgs)
195 		memguard_mapsize = mem_pgs * 2 * PAGE_SIZE;
196 	if (km_size + memguard_mapsize > parent_size)
197 		memguard_mapsize = 0;
198 	return (km_size + memguard_mapsize);
199 }
200 
201 /*
202  * Initialize the MemGuard mock allocator.  All objects from MemGuard come
203  * out of a single contiguous chunk of kernel address space that is managed
204  * by a vmem arena.
205  */
206 void
207 memguard_init(vmem_t *parent)
208 {
209 	vm_offset_t base;
210 
211 	vmem_alloc(parent, memguard_mapsize, M_BESTFIT | M_WAITOK, &base);
212 	vmem_init(memguard_arena, "memguard arena", base, memguard_mapsize,
213 	    PAGE_SIZE, 0, M_WAITOK);
214 	memguard_base = base;
215 
216 	printf("MEMGUARD DEBUGGING ALLOCATOR INITIALIZED:\n");
217 	printf("\tMEMGUARD map base: 0x%lx\n", (u_long)base);
218 	printf("\tMEMGUARD map size: %jd KBytes\n",
219 	    (uintmax_t)memguard_mapsize >> 10);
220 }
221 
222 /*
223  * Run things that can't be done as early as memguard_init().
224  */
225 static void
226 memguard_sysinit(void)
227 {
228 	struct sysctl_oid_list *parent;
229 
230 	parent = SYSCTL_STATIC_CHILDREN(_vm_memguard);
231 	SYSCTL_ADD_UAUTO(NULL, parent, OID_AUTO, "mapstart",
232 	    CTLFLAG_RD, &memguard_base,
233 	    "MemGuard KVA base");
234 	SYSCTL_ADD_UAUTO(NULL, parent, OID_AUTO, "maplimit",
235 	    CTLFLAG_RD, &memguard_mapsize,
236 	    "MemGuard KVA size");
237 	SYSCTL_ADD_PROC(NULL, parent, OID_AUTO, "mapused",
238 	    CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_ULONG, NULL, 0, memguard_sysctl_mapused, "LU",
239 	    "MemGuard KVA used");
240 }
241 SYSINIT(memguard, SI_SUB_KLD, SI_ORDER_ANY, memguard_sysinit, NULL);
242 
243 /*
244  * v2sizep() converts a virtual address of the first page allocated for
245  * an item to a pointer to u_long recording the size of the original
246  * allocation request.
247  *
248  * This routine is very similar to those defined by UMA in uma_int.h.
249  * The difference is that this routine stores the originally allocated
250  * size in one of the page's fields that is unused when the page is
251  * wired rather than the object field, which is used.
252  */
253 static u_long *
254 v2sizep(vm_offset_t va)
255 {
256 	vm_paddr_t pa;
257 	struct vm_page *p;
258 
259 	pa = pmap_kextract(va);
260 	if (pa == 0)
261 		panic("MemGuard detected double-free of %p", (void *)va);
262 	p = PHYS_TO_VM_PAGE(pa);
263 	KASSERT(vm_page_wired(p) && p->a.queue == PQ_NONE,
264 	    ("MEMGUARD: Expected wired page %p in vtomgfifo!", p));
265 	return (&p->plinks.memguard.p);
266 }
267 
268 static u_long *
269 v2sizev(vm_offset_t va)
270 {
271 	vm_paddr_t pa;
272 	struct vm_page *p;
273 
274 	pa = pmap_kextract(va);
275 	if (pa == 0)
276 		panic("MemGuard detected double-free of %p", (void *)va);
277 	p = PHYS_TO_VM_PAGE(pa);
278 	KASSERT(vm_page_wired(p) && p->a.queue == PQ_NONE,
279 	    ("MEMGUARD: Expected wired page %p in vtomgfifo!", p));
280 	return (&p->plinks.memguard.v);
281 }
282 
283 /*
284  * Allocate a single object of specified size with specified flags
285  * (either M_WAITOK or M_NOWAIT).
286  */
287 void *
288 memguard_alloc(unsigned long req_size, int flags)
289 {
290 	vm_offset_t addr, origaddr;
291 	u_long size_p, size_v;
292 	int do_guard, error, rv;
293 
294 	size_p = round_page(req_size);
295 	if (size_p == 0)
296 		return (NULL);
297 
298 	/*
299 	 * To ensure there are holes on both sides of the allocation,
300 	 * request 2 extra pages of KVA.  Save the value of memguard_options
301 	 * so that we use a consistent value throughout this function.
302 	 */
303 	size_v = size_p;
304 	do_guard = (memguard_options & MG_GUARD_AROUND) != 0;
305 	if (do_guard)
306 		size_v += 2 * PAGE_SIZE;
307 
308 	/*
309 	 * When we pass our memory limit, reject sub-page allocations.
310 	 * Page-size and larger allocations will use the same amount
311 	 * of physical memory whether we allocate or hand off to
312 	 * malloc_large(), so keep those.
313 	 */
314 	if (vmem_size(memguard_arena, VMEM_ALLOC) >= memguard_physlimit &&
315 	    req_size < PAGE_SIZE) {
316 		addr = (vm_offset_t)NULL;
317 		memguard_fail_pgs++;
318 		goto out;
319 	}
320 
321 	/*
322 	 * Attempt to avoid address reuse for as long as possible, to increase
323 	 * the likelihood of catching a use-after-free.
324 	 */
325 	error = vmem_alloc(memguard_arena, size_v, M_NEXTFIT | M_NOWAIT,
326 	    &origaddr);
327 	if (error != 0) {
328 		memguard_fail_kva++;
329 		addr = (vm_offset_t)NULL;
330 		goto out;
331 	}
332 	addr = origaddr;
333 	if (do_guard)
334 		addr += PAGE_SIZE;
335 	rv = kmem_back(kernel_object, addr, size_p, flags);
336 	if (rv != KERN_SUCCESS) {
337 		vmem_xfree(memguard_arena, origaddr, size_v);
338 		memguard_fail_pgs++;
339 		addr = (vm_offset_t)NULL;
340 		goto out;
341 	}
342 	*v2sizep(trunc_page(addr)) = req_size;
343 	*v2sizev(trunc_page(addr)) = size_v;
344 	memguard_succ++;
345 	if (req_size < PAGE_SIZE) {
346 		memguard_wasted += (PAGE_SIZE - req_size);
347 		if (do_guard) {
348 			/*
349 			 * Align the request to 16 bytes, and return
350 			 * an address near the end of the page, to
351 			 * better detect array overrun.
352 			 */
353 			req_size = roundup2(req_size, 16);
354 			addr += (PAGE_SIZE - req_size);
355 		}
356 	}
357 out:
358 	return ((void *)addr);
359 }
360 
361 int
362 is_memguard_addr(void *addr)
363 {
364 	vm_offset_t a = (vm_offset_t)(uintptr_t)addr;
365 
366 	return (a >= memguard_base && a < memguard_base + memguard_mapsize);
367 }
368 
369 /*
370  * Free specified single object.
371  */
372 void
373 memguard_free(void *ptr)
374 {
375 	vm_offset_t addr;
376 	u_long req_size, size, sizev;
377 	char *temp;
378 	int i;
379 
380 	addr = trunc_page((uintptr_t)ptr);
381 	req_size = *v2sizep(addr);
382 	sizev = *v2sizev(addr);
383 	size = round_page(req_size);
384 
385 	/*
386 	 * Page should not be guarded right now, so force a write.
387 	 * The purpose of this is to increase the likelihood of
388 	 * catching a double-free, but not necessarily a
389 	 * tamper-after-free (the second thread freeing might not
390 	 * write before freeing, so this forces it to and,
391 	 * subsequently, trigger a fault).
392 	 */
393 	temp = ptr;
394 	for (i = 0; i < size; i += PAGE_SIZE)
395 		temp[i] = 'M';
396 
397 	/*
398 	 * This requires carnal knowledge of the implementation of
399 	 * kmem_free(), but since we've already replaced kmem_malloc()
400 	 * above, it's not really any worse.  We want to use the
401 	 * vm_map lock to serialize updates to memguard_wasted, since
402 	 * we had the lock at increment.
403 	 */
404 	kmem_unback(kernel_object, addr, size);
405 	if (sizev > size)
406 		addr -= PAGE_SIZE;
407 	vmem_xfree(memguard_arena, addr, sizev);
408 	if (req_size < PAGE_SIZE)
409 		memguard_wasted -= (PAGE_SIZE - req_size);
410 }
411 
412 /*
413  * Re-allocate an allocation that was originally guarded.
414  */
415 void *
416 memguard_realloc(void *addr, unsigned long size, struct malloc_type *mtp,
417     int flags)
418 {
419 	void *newaddr;
420 	u_long old_size;
421 
422 	/*
423 	 * Allocate the new block.  Force the allocation to be guarded
424 	 * as the original may have been guarded through random
425 	 * chance, and that should be preserved.
426 	 */
427 	if ((newaddr = memguard_alloc(size, flags)) == NULL)
428 		return (NULL);
429 
430 	/* Copy over original contents. */
431 	old_size = *v2sizep(trunc_page((uintptr_t)addr));
432 	bcopy(addr, newaddr, min(size, old_size));
433 	memguard_free(addr);
434 	return (newaddr);
435 }
436 
437 static int
438 memguard_cmp(unsigned long size)
439 {
440 
441 	if (size < memguard_minsize) {
442 		memguard_minsize_reject++;
443 		return (0);
444 	}
445 	if ((memguard_options & MG_GUARD_ALLLARGE) != 0 && size >= PAGE_SIZE)
446 		return (1);
447 	if (memguard_frequency > 0 &&
448 	    (random() % 100000) < memguard_frequency) {
449 		memguard_frequency_hits++;
450 		return (1);
451 	}
452 
453 	return (0);
454 }
455 
456 int
457 memguard_cmp_mtp(struct malloc_type *mtp, unsigned long size)
458 {
459 
460 	if (memguard_cmp(size))
461 		return(1);
462 
463 #if 1
464 	/*
465 	 * The safest way of comparison is to always compare short description
466 	 * string of memory type, but it is also the slowest way.
467 	 */
468 	return (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0);
469 #else
470 	/*
471 	 * If we compare pointers, there are two possible problems:
472 	 * 1. Memory type was unloaded and new memory type was allocated at the
473 	 *    same address.
474 	 * 2. Memory type was unloaded and loaded again, but allocated at a
475 	 *    different address.
476 	 */
477 	if (vm_memguard_mtype != NULL)
478 		return (mtp == vm_memguard_mtype);
479 	if (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0) {
480 		vm_memguard_mtype = mtp;
481 		return (1);
482 	}
483 	return (0);
484 #endif
485 }
486 
487 int
488 memguard_cmp_zone(uma_zone_t zone)
489 {
490 
491 	if ((memguard_options & MG_GUARD_NOFREE) == 0 &&
492 	    zone->uz_flags & UMA_ZONE_NOFREE)
493 		return (0);
494 
495 	if (memguard_cmp(zone->uz_size))
496 		return (1);
497 
498 	/*
499 	 * The safest way of comparison is to always compare zone name,
500 	 * but it is also the slowest way.
501 	 */
502 	return (strcmp(zone->uz_name, vm_memguard_desc) == 0);
503 }
504 
505 unsigned long
506 memguard_get_req_size(const void *addr)
507 {
508 	return (*v2sizep(trunc_page((uintptr_t)addr)));
509 }
510