xref: /freebsd/sys/vm/memguard.c (revision ab2043b81eaba0d7d7769b4a58b2b6d17bc464a3)
1 /*-
2  * Copyright (c) 2005, Bosko Milekic <bmilekic@FreeBSD.org>.
3  * Copyright (c) 2010 Isilon Systems, Inc. (http://www.isilon.com/)
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice unmodified, this list of conditions, and the following
11  *    disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
21  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
22  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
23  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
24  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
25  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
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 
52 #include <vm/vm.h>
53 #include <vm/uma.h>
54 #include <vm/vm_param.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_map.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_extern.h>
59 #include <vm/uma_int.h>
60 #include <vm/memguard.h>
61 
62 static SYSCTL_NODE(_vm, OID_AUTO, memguard, CTLFLAG_RW, NULL, "MemGuard data");
63 /*
64  * The vm_memguard_divisor variable controls how much of kmem_map should be
65  * reserved for MemGuard.
66  */
67 static u_int vm_memguard_divisor;
68 SYSCTL_UINT(_vm_memguard, OID_AUTO, divisor, CTLFLAG_RDTUN,
69     &vm_memguard_divisor,
70     0, "(kmem_size/memguard_divisor) == memguard submap size");
71 
72 /*
73  * Short description (ks_shortdesc) of memory type to monitor.
74  */
75 static char vm_memguard_desc[128] = "";
76 static struct malloc_type *vm_memguard_mtype = NULL;
77 TUNABLE_STR("vm.memguard.desc", vm_memguard_desc, sizeof(vm_memguard_desc));
78 static int
79 memguard_sysctl_desc(SYSCTL_HANDLER_ARGS)
80 {
81 	char desc[sizeof(vm_memguard_desc)];
82 	int error;
83 
84 	strlcpy(desc, vm_memguard_desc, sizeof(desc));
85 	error = sysctl_handle_string(oidp, desc, sizeof(desc), req);
86 	if (error != 0 || req->newptr == NULL)
87 		return (error);
88 
89 	mtx_lock(&malloc_mtx);
90 	/*
91 	 * If mtp is NULL, it will be initialized in memguard_cmp().
92 	 */
93 	vm_memguard_mtype = malloc_desc2type(desc);
94 	strlcpy(vm_memguard_desc, desc, sizeof(vm_memguard_desc));
95 	mtx_unlock(&malloc_mtx);
96 	return (error);
97 }
98 SYSCTL_PROC(_vm_memguard, OID_AUTO, desc,
99     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0,
100     memguard_sysctl_desc, "A", "Short description of memory type to monitor");
101 
102 static vm_map_t memguard_map = NULL;
103 static vm_offset_t memguard_cursor;
104 static vm_size_t memguard_mapsize;
105 static vm_size_t memguard_physlimit;
106 static u_long memguard_wasted;
107 static u_long memguard_wrap;
108 static u_long memguard_succ;
109 static u_long memguard_fail_kva;
110 static u_long memguard_fail_pgs;
111 
112 SYSCTL_ULONG(_vm_memguard, OID_AUTO, cursor, CTLFLAG_RD,
113     &memguard_cursor, 0, "MemGuard cursor");
114 SYSCTL_ULONG(_vm_memguard, OID_AUTO, mapsize, CTLFLAG_RD,
115     &memguard_mapsize, 0, "MemGuard private vm_map size");
116 SYSCTL_ULONG(_vm_memguard, OID_AUTO, phys_limit, CTLFLAG_RD,
117     &memguard_physlimit, 0, "Limit on MemGuard memory consumption");
118 SYSCTL_ULONG(_vm_memguard, OID_AUTO, wasted, CTLFLAG_RD,
119     &memguard_wasted, 0, "Excess memory used through page promotion");
120 SYSCTL_ULONG(_vm_memguard, OID_AUTO, wrapcnt, CTLFLAG_RD,
121     &memguard_wrap, 0, "MemGuard cursor wrap count");
122 SYSCTL_ULONG(_vm_memguard, OID_AUTO, numalloc, CTLFLAG_RD,
123     &memguard_succ, 0, "Count of successful MemGuard allocations");
124 SYSCTL_ULONG(_vm_memguard, OID_AUTO, fail_kva, CTLFLAG_RD,
125     &memguard_fail_kva, 0, "MemGuard failures due to lack of KVA");
126 SYSCTL_ULONG(_vm_memguard, OID_AUTO, fail_pgs, CTLFLAG_RD,
127     &memguard_fail_pgs, 0, "MemGuard failures due to lack of pages");
128 
129 #define MG_GUARD_AROUND		0x001
130 #define MG_GUARD_ALLLARGE	0x002
131 #define MG_GUARD_NOFREE		0x004
132 static int memguard_options = MG_GUARD_AROUND;
133 TUNABLE_INT("vm.memguard.options", &memguard_options);
134 SYSCTL_INT(_vm_memguard, OID_AUTO, options, CTLFLAG_RW,
135     &memguard_options, 0,
136     "MemGuard options:\n"
137     "\t0x001 - add guard pages around each allocation\n"
138     "\t0x002 - always use MemGuard for allocations over a page\n"
139     "\t0x004 - guard uma(9) zones with UMA_ZONE_NOFREE flag");
140 
141 static u_int memguard_minsize;
142 static u_long memguard_minsize_reject;
143 SYSCTL_UINT(_vm_memguard, OID_AUTO, minsize, CTLFLAG_RW,
144     &memguard_minsize, 0, "Minimum size for page promotion");
145 SYSCTL_ULONG(_vm_memguard, OID_AUTO, minsize_reject, CTLFLAG_RD,
146     &memguard_minsize_reject, 0, "# times rejected for size");
147 
148 static u_int memguard_frequency;
149 static u_long memguard_frequency_hits;
150 TUNABLE_INT("vm.memguard.frequency", &memguard_frequency);
151 SYSCTL_UINT(_vm_memguard, OID_AUTO, frequency, CTLFLAG_RW,
152     &memguard_frequency, 0, "Times in 100000 that MemGuard will randomly run");
153 SYSCTL_ULONG(_vm_memguard, OID_AUTO, frequency_hits, CTLFLAG_RD,
154     &memguard_frequency_hits, 0, "# times MemGuard randomly chose");
155 
156 
157 /*
158  * Return a fudged value to be used for vm_kmem_size for allocating
159  * the kmem_map.  The memguard memory will be a submap.
160  */
161 unsigned long
162 memguard_fudge(unsigned long km_size, const struct vm_map *parent_map)
163 {
164 	u_long mem_pgs, parent_size;
165 
166 	vm_memguard_divisor = 10;
167 	TUNABLE_INT_FETCH("vm.memguard.divisor", &vm_memguard_divisor);
168 
169 	parent_size = vm_map_max(parent_map) - vm_map_min(parent_map) +
170 	    PAGE_SIZE;
171 	/* Pick a conservative value if provided value sucks. */
172 	if ((vm_memguard_divisor <= 0) ||
173 	    ((parent_size / vm_memguard_divisor) == 0))
174 		vm_memguard_divisor = 10;
175 	/*
176 	 * Limit consumption of physical pages to
177 	 * 1/vm_memguard_divisor of system memory.  If the KVA is
178 	 * smaller than this then the KVA limit comes into play first.
179 	 * This prevents memguard's page promotions from completely
180 	 * using up memory, since most malloc(9) calls are sub-page.
181 	 */
182 	mem_pgs = cnt.v_page_count;
183 	memguard_physlimit = (mem_pgs / vm_memguard_divisor) * PAGE_SIZE;
184 	/*
185 	 * We want as much KVA as we can take safely.  Use at most our
186 	 * allotted fraction of the parent map's size.  Limit this to
187 	 * twice the physical memory to avoid using too much memory as
188 	 * pagetable pages (size must be multiple of PAGE_SIZE).
189 	 */
190 	memguard_mapsize = round_page(parent_size / vm_memguard_divisor);
191 	if (memguard_mapsize / (2 * PAGE_SIZE) > mem_pgs)
192 		memguard_mapsize = mem_pgs * 2 * PAGE_SIZE;
193 	if (km_size + memguard_mapsize > parent_size)
194 		memguard_mapsize = 0;
195 	return (km_size + memguard_mapsize);
196 }
197 
198 /*
199  * Initialize the MemGuard mock allocator.  All objects from MemGuard come
200  * out of a single VM map (contiguous chunk of address space).
201  */
202 void
203 memguard_init(vm_map_t parent_map)
204 {
205 	vm_offset_t base, limit;
206 
207 	memguard_map = kmem_suballoc(parent_map, &base, &limit,
208 	    memguard_mapsize, FALSE);
209 	memguard_map->system_map = 1;
210 	KASSERT(memguard_mapsize == limit - base,
211 	    ("Expected %lu, got %lu", (u_long)memguard_mapsize,
212 	     (u_long)(limit - base)));
213 	memguard_cursor = base;
214 
215 	printf("MEMGUARD DEBUGGING ALLOCATOR INITIALIZED:\n");
216 	printf("\tMEMGUARD map base: 0x%lx\n", (u_long)base);
217 	printf("\tMEMGUARD map limit: 0x%lx\n", (u_long)limit);
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 
232 	SYSCTL_ADD_ULONG(NULL, parent, OID_AUTO, "mapstart", CTLFLAG_RD,
233 	    &memguard_map->min_offset, "MemGuard KVA base");
234 	SYSCTL_ADD_ULONG(NULL, parent, OID_AUTO, "maplimit", CTLFLAG_RD,
235 	    &memguard_map->max_offset, "MemGuard KVA end");
236 	SYSCTL_ADD_ULONG(NULL, parent, OID_AUTO, "mapused", CTLFLAG_RD,
237 	    &memguard_map->size, "MemGuard KVA used");
238 }
239 SYSINIT(memguard, SI_SUB_KLD, SI_ORDER_ANY, memguard_sysinit, NULL);
240 
241 /*
242  * v2sizep() converts a virtual address of the first page allocated for
243  * an item to a pointer to u_long recording the size of the original
244  * allocation request.
245  *
246  * This routine is very similar to those defined by UMA in uma_int.h.
247  * The difference is that this routine stores the originally allocated
248  * size in one of the page's fields that is unused when the page is
249  * wired rather than the object field, which is used.
250  */
251 static u_long *
252 v2sizep(vm_offset_t va)
253 {
254 	vm_paddr_t pa;
255 	struct vm_page *p;
256 
257 	pa = pmap_kextract(va);
258 	if (pa == 0)
259 		panic("MemGuard detected double-free of %p", (void *)va);
260 	p = PHYS_TO_VM_PAGE(pa);
261 	KASSERT(p->wire_count != 0 && p->queue == PQ_NONE,
262 	    ("MEMGUARD: Expected wired page %p in vtomgfifo!", p));
263 	return ((u_long *)&p->pageq.tqe_next);
264 }
265 
266 /*
267  * Allocate a single object of specified size with specified flags
268  * (either M_WAITOK or M_NOWAIT).
269  */
270 void *
271 memguard_alloc(unsigned long req_size, int flags)
272 {
273 	vm_offset_t addr;
274 	u_long size_p, size_v;
275 	int do_guard, rv;
276 
277 	size_p = round_page(req_size);
278 	if (size_p == 0)
279 		return (NULL);
280 	/*
281 	 * To ensure there are holes on both sides of the allocation,
282 	 * request 2 extra pages of KVA.  We will only actually add a
283 	 * vm_map_entry and get pages for the original request.  Save
284 	 * the value of memguard_options so we have a consistent
285 	 * value.
286 	 */
287 	size_v = size_p;
288 	do_guard = (memguard_options & MG_GUARD_AROUND) != 0;
289 	if (do_guard)
290 		size_v += 2 * PAGE_SIZE;
291 
292 	vm_map_lock(memguard_map);
293 	/*
294 	 * When we pass our memory limit, reject sub-page allocations.
295 	 * Page-size and larger allocations will use the same amount
296 	 * of physical memory whether we allocate or hand off to
297 	 * uma_large_alloc(), so keep those.
298 	 */
299 	if (memguard_map->size >= memguard_physlimit &&
300 	    req_size < PAGE_SIZE) {
301 		addr = (vm_offset_t)NULL;
302 		memguard_fail_pgs++;
303 		goto out;
304 	}
305 	/*
306 	 * Keep a moving cursor so we don't recycle KVA as long as
307 	 * possible.  It's not perfect, since we don't know in what
308 	 * order previous allocations will be free'd, but it's simple
309 	 * and fast, and requires O(1) additional storage if guard
310 	 * pages are not used.
311 	 *
312 	 * XXX This scheme will lead to greater fragmentation of the
313 	 * map, unless vm_map_findspace() is tweaked.
314 	 */
315 	for (;;) {
316 		rv = vm_map_findspace(memguard_map, memguard_cursor,
317 		    size_v, &addr);
318 		if (rv == KERN_SUCCESS)
319 			break;
320 		/*
321 		 * The map has no space.  This may be due to
322 		 * fragmentation, or because the cursor is near the
323 		 * end of the map.
324 		 */
325 		if (memguard_cursor == vm_map_min(memguard_map)) {
326 			memguard_fail_kva++;
327 			addr = (vm_offset_t)NULL;
328 			goto out;
329 		}
330 		memguard_wrap++;
331 		memguard_cursor = vm_map_min(memguard_map);
332 	}
333 	if (do_guard)
334 		addr += PAGE_SIZE;
335 	rv = kmem_back(memguard_map, addr, size_p, flags);
336 	if (rv != KERN_SUCCESS) {
337 		memguard_fail_pgs++;
338 		addr = (vm_offset_t)NULL;
339 		goto out;
340 	}
341 	memguard_cursor = addr + size_p;
342 	*v2sizep(trunc_page(addr)) = req_size;
343 	memguard_succ++;
344 	if (req_size < PAGE_SIZE) {
345 		memguard_wasted += (PAGE_SIZE - req_size);
346 		if (do_guard) {
347 			/*
348 			 * Align the request to 16 bytes, and return
349 			 * an address near the end of the page, to
350 			 * better detect array overrun.
351 			 */
352 			req_size = roundup2(req_size, 16);
353 			addr += (PAGE_SIZE - req_size);
354 		}
355 	}
356 out:
357 	vm_map_unlock(memguard_map);
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_map->min_offset && a < memguard_map->max_offset);
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;
377 	char *temp;
378 	int i;
379 
380 	addr = trunc_page((uintptr_t)ptr);
381 	req_size = *v2sizep(addr);
382 	size = round_page(req_size);
383 
384 	/*
385 	 * Page should not be guarded right now, so force a write.
386 	 * The purpose of this is to increase the likelihood of
387 	 * catching a double-free, but not necessarily a
388 	 * tamper-after-free (the second thread freeing might not
389 	 * write before freeing, so this forces it to and,
390 	 * subsequently, trigger a fault).
391 	 */
392 	temp = ptr;
393 	for (i = 0; i < size; i += PAGE_SIZE)
394 		temp[i] = 'M';
395 
396 	/*
397 	 * This requires carnal knowledge of the implementation of
398 	 * kmem_free(), but since we've already replaced kmem_malloc()
399 	 * above, it's not really any worse.  We want to use the
400 	 * vm_map lock to serialize updates to memguard_wasted, since
401 	 * we had the lock at increment.
402 	 */
403 	vm_map_lock(memguard_map);
404 	if (req_size < PAGE_SIZE)
405 		memguard_wasted -= (PAGE_SIZE - req_size);
406 	(void)vm_map_delete(memguard_map, addr, addr + size);
407 	vm_map_unlock(memguard_map);
408 }
409 
410 /*
411  * Re-allocate an allocation that was originally guarded.
412  */
413 void *
414 memguard_realloc(void *addr, unsigned long size, struct malloc_type *mtp,
415     int flags)
416 {
417 	void *newaddr;
418 	u_long old_size;
419 
420 	/*
421 	 * Allocate the new block.  Force the allocation to be guarded
422 	 * as the original may have been guarded through random
423 	 * chance, and that should be preserved.
424 	 */
425 	if ((newaddr = memguard_alloc(size, flags)) == NULL)
426 		return (NULL);
427 
428 	/* Copy over original contents. */
429 	old_size = *v2sizep(trunc_page((uintptr_t)addr));
430 	bcopy(addr, newaddr, min(size, old_size));
431 	memguard_free(addr);
432 	return (newaddr);
433 }
434 
435 static int
436 memguard_cmp(unsigned long size)
437 {
438 
439 	if (size < memguard_minsize) {
440 		memguard_minsize_reject++;
441 		return (0);
442 	}
443 	if ((memguard_options & MG_GUARD_ALLLARGE) != 0 && size >= PAGE_SIZE)
444 		return (1);
445 	if (memguard_frequency > 0 &&
446 	    (random() % 100000) < memguard_frequency) {
447 		memguard_frequency_hits++;
448 		return (1);
449 	}
450 
451 	return (0);
452 }
453 
454 int
455 memguard_cmp_mtp(struct malloc_type *mtp, unsigned long size)
456 {
457 
458 	if (memguard_cmp(size))
459 		return(1);
460 
461 #if 1
462 	/*
463 	 * The safest way of comparsion is to always compare short description
464 	 * string of memory type, but it is also the slowest way.
465 	 */
466 	return (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0);
467 #else
468 	/*
469 	 * If we compare pointers, there are two possible problems:
470 	 * 1. Memory type was unloaded and new memory type was allocated at the
471 	 *    same address.
472 	 * 2. Memory type was unloaded and loaded again, but allocated at a
473 	 *    different address.
474 	 */
475 	if (vm_memguard_mtype != NULL)
476 		return (mtp == vm_memguard_mtype);
477 	if (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0) {
478 		vm_memguard_mtype = mtp;
479 		return (1);
480 	}
481 	return (0);
482 #endif
483 }
484 
485 int
486 memguard_cmp_zone(uma_zone_t zone)
487 {
488 
489 	 if ((memguard_options & MG_GUARD_NOFREE) == 0 &&
490 	    zone->uz_flags & UMA_ZONE_NOFREE)
491 		return (0);
492 
493 	if (memguard_cmp(zone->uz_size))
494 		return (1);
495 
496 	/*
497 	 * The safest way of comparsion is to always compare zone name,
498 	 * but it is also the slowest way.
499 	 */
500 	return (strcmp(zone->uz_name, vm_memguard_desc) == 0);
501 }
502