xref: /freebsd/sys/vm/memguard.c (revision 26a222dc0c048fc071b548eadad7b80405a1b126)
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 #include <sys/vmem.h>
52 
53 #include <vm/vm.h>
54 #include <vm/uma.h>
55 #include <vm/vm_param.h>
56 #include <vm/vm_page.h>
57 #include <vm/vm_map.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_kern.h>
60 #include <vm/vm_extern.h>
61 #include <vm/uma_int.h>
62 #include <vm/memguard.h>
63 
64 static SYSCTL_NODE(_vm, OID_AUTO, memguard, CTLFLAG_RW, NULL, "MemGuard data");
65 /*
66  * The vm_memguard_divisor variable controls how much of kmem_map should be
67  * reserved for MemGuard.
68  */
69 static u_int vm_memguard_divisor;
70 SYSCTL_UINT(_vm_memguard, OID_AUTO, divisor, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
71     &vm_memguard_divisor,
72     0, "(kmem_size/memguard_divisor) == memguard submap size");
73 
74 /*
75  * Short description (ks_shortdesc) of memory type to monitor.
76  */
77 static char vm_memguard_desc[128] = "";
78 static struct malloc_type *vm_memguard_mtype = NULL;
79 TUNABLE_STR("vm.memguard.desc", vm_memguard_desc, sizeof(vm_memguard_desc));
80 static int
81 memguard_sysctl_desc(SYSCTL_HANDLER_ARGS)
82 {
83 	char desc[sizeof(vm_memguard_desc)];
84 	int error;
85 
86 	strlcpy(desc, vm_memguard_desc, sizeof(desc));
87 	error = sysctl_handle_string(oidp, desc, sizeof(desc), req);
88 	if (error != 0 || req->newptr == NULL)
89 		return (error);
90 
91 	mtx_lock(&malloc_mtx);
92 	/* If mtp is NULL, it will be initialized in memguard_cmp() */
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_offset_t memguard_cursor;
103 static vm_offset_t memguard_base;
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 arena 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 SYSCTL_INT(_vm_memguard, OID_AUTO, options, CTLFLAG_RWTUN,
134     &memguard_options, 0,
135     "MemGuard options:\n"
136     "\t0x001 - add guard pages around each allocation\n"
137     "\t0x002 - always use MemGuard for allocations over a page\n"
138     "\t0x004 - guard uma(9) zones with UMA_ZONE_NOFREE flag");
139 
140 static u_int memguard_minsize;
141 static u_long memguard_minsize_reject;
142 SYSCTL_UINT(_vm_memguard, OID_AUTO, minsize, CTLFLAG_RW,
143     &memguard_minsize, 0, "Minimum size for page promotion");
144 SYSCTL_ULONG(_vm_memguard, OID_AUTO, minsize_reject, CTLFLAG_RD,
145     &memguard_minsize_reject, 0, "# times rejected for size");
146 
147 static u_int memguard_frequency;
148 static u_long memguard_frequency_hits;
149 SYSCTL_UINT(_vm_memguard, OID_AUTO, frequency, CTLFLAG_RWTUN,
150     &memguard_frequency, 0, "Times in 100000 that MemGuard will randomly run");
151 SYSCTL_ULONG(_vm_memguard, OID_AUTO, frequency_hits, CTLFLAG_RD,
152     &memguard_frequency_hits, 0, "# times MemGuard randomly chose");
153 
154 
155 /*
156  * Return a fudged value to be used for vm_kmem_size for allocating
157  * the kmem_map.  The memguard memory will be a submap.
158  */
159 unsigned long
160 memguard_fudge(unsigned long km_size, const struct vm_map *parent_map)
161 {
162 	u_long mem_pgs, parent_size;
163 
164 	vm_memguard_divisor = 10;
165 	TUNABLE_INT_FETCH("vm.memguard.divisor", &vm_memguard_divisor);
166 
167 	parent_size = vm_map_max(parent_map) - vm_map_min(parent_map) +
168 	    PAGE_SIZE;
169 	/* Pick a conservative value if provided value sucks. */
170 	if ((vm_memguard_divisor <= 0) ||
171 	    ((parent_size / vm_memguard_divisor) == 0))
172 		vm_memguard_divisor = 10;
173 	/*
174 	 * Limit consumption of physical pages to
175 	 * 1/vm_memguard_divisor of system memory.  If the KVA is
176 	 * smaller than this then the KVA limit comes into play first.
177 	 * This prevents memguard's page promotions from completely
178 	 * using up memory, since most malloc(9) calls are sub-page.
179 	 */
180 	mem_pgs = vm_cnt.v_page_count;
181 	memguard_physlimit = (mem_pgs / vm_memguard_divisor) * PAGE_SIZE;
182 	/*
183 	 * We want as much KVA as we can take safely.  Use at most our
184 	 * allotted fraction of the parent map's size.  Limit this to
185 	 * twice the physical memory to avoid using too much memory as
186 	 * pagetable pages (size must be multiple of PAGE_SIZE).
187 	 */
188 	memguard_mapsize = round_page(parent_size / vm_memguard_divisor);
189 	if (memguard_mapsize / (2 * PAGE_SIZE) > mem_pgs)
190 		memguard_mapsize = mem_pgs * 2 * PAGE_SIZE;
191 	if (km_size + memguard_mapsize > parent_size)
192 		memguard_mapsize = 0;
193 	return (km_size + memguard_mapsize);
194 }
195 
196 /*
197  * Initialize the MemGuard mock allocator.  All objects from MemGuard come
198  * out of a single VM map (contiguous chunk of address space).
199  */
200 void
201 memguard_init(vmem_t *parent)
202 {
203 	vm_offset_t base;
204 
205 	vmem_alloc(parent, memguard_mapsize, M_BESTFIT | M_WAITOK, &base);
206 	vmem_init(memguard_arena, "memguard arena", base, memguard_mapsize,
207 	    PAGE_SIZE, 0, M_WAITOK);
208 	memguard_cursor = base;
209 	memguard_base = base;
210 
211 	printf("MEMGUARD DEBUGGING ALLOCATOR INITIALIZED:\n");
212 	printf("\tMEMGUARD map base: 0x%lx\n", (u_long)base);
213 	printf("\tMEMGUARD map size: %jd KBytes\n",
214 	    (uintmax_t)memguard_mapsize >> 10);
215 }
216 
217 /*
218  * Run things that can't be done as early as memguard_init().
219  */
220 static void
221 memguard_sysinit(void)
222 {
223 	struct sysctl_oid_list *parent;
224 
225 	parent = SYSCTL_STATIC_CHILDREN(_vm_memguard);
226 
227 	SYSCTL_ADD_UAUTO(NULL, parent, OID_AUTO, "mapstart", CTLFLAG_RD,
228 	    &memguard_base, "MemGuard KVA base");
229 	SYSCTL_ADD_UAUTO(NULL, parent, OID_AUTO, "maplimit", CTLFLAG_RD,
230 	    &memguard_mapsize, "MemGuard KVA size");
231 #if 0
232 	SYSCTL_ADD_ULONG(NULL, parent, OID_AUTO, "mapused", CTLFLAG_RD,
233 	    &memguard_map->size, "MemGuard KVA used");
234 #endif
235 }
236 SYSINIT(memguard, SI_SUB_KLD, SI_ORDER_ANY, memguard_sysinit, NULL);
237 
238 /*
239  * v2sizep() converts a virtual address of the first page allocated for
240  * an item to a pointer to u_long recording the size of the original
241  * allocation request.
242  *
243  * This routine is very similar to those defined by UMA in uma_int.h.
244  * The difference is that this routine stores the originally allocated
245  * size in one of the page's fields that is unused when the page is
246  * wired rather than the object field, which is used.
247  */
248 static u_long *
249 v2sizep(vm_offset_t va)
250 {
251 	vm_paddr_t pa;
252 	struct vm_page *p;
253 
254 	pa = pmap_kextract(va);
255 	if (pa == 0)
256 		panic("MemGuard detected double-free of %p", (void *)va);
257 	p = PHYS_TO_VM_PAGE(pa);
258 	KASSERT(p->wire_count != 0 && p->queue == PQ_NONE,
259 	    ("MEMGUARD: Expected wired page %p in vtomgfifo!", p));
260 	return (&p->plinks.memguard.p);
261 }
262 
263 static u_long *
264 v2sizev(vm_offset_t va)
265 {
266 	vm_paddr_t pa;
267 	struct vm_page *p;
268 
269 	pa = pmap_kextract(va);
270 	if (pa == 0)
271 		panic("MemGuard detected double-free of %p", (void *)va);
272 	p = PHYS_TO_VM_PAGE(pa);
273 	KASSERT(p->wire_count != 0 && p->queue == PQ_NONE,
274 	    ("MEMGUARD: Expected wired page %p in vtomgfifo!", p));
275 	return (&p->plinks.memguard.v);
276 }
277 
278 /*
279  * Allocate a single object of specified size with specified flags
280  * (either M_WAITOK or M_NOWAIT).
281  */
282 void *
283 memguard_alloc(unsigned long req_size, int flags)
284 {
285 	vm_offset_t addr;
286 	u_long size_p, size_v;
287 	int do_guard, rv;
288 
289 	size_p = round_page(req_size);
290 	if (size_p == 0)
291 		return (NULL);
292 	/*
293 	 * To ensure there are holes on both sides of the allocation,
294 	 * request 2 extra pages of KVA.  We will only actually add a
295 	 * vm_map_entry and get pages for the original request.  Save
296 	 * the value of memguard_options so we have a consistent
297 	 * value.
298 	 */
299 	size_v = size_p;
300 	do_guard = (memguard_options & MG_GUARD_AROUND) != 0;
301 	if (do_guard)
302 		size_v += 2 * PAGE_SIZE;
303 
304 	/*
305 	 * When we pass our memory limit, reject sub-page allocations.
306 	 * Page-size and larger allocations will use the same amount
307 	 * of physical memory whether we allocate or hand off to
308 	 * uma_large_alloc(), so keep those.
309 	 */
310 	if (vmem_size(memguard_arena, VMEM_ALLOC) >= memguard_physlimit &&
311 	    req_size < PAGE_SIZE) {
312 		addr = (vm_offset_t)NULL;
313 		memguard_fail_pgs++;
314 		goto out;
315 	}
316 	/*
317 	 * Keep a moving cursor so we don't recycle KVA as long as
318 	 * possible.  It's not perfect, since we don't know in what
319 	 * order previous allocations will be free'd, but it's simple
320 	 * and fast, and requires O(1) additional storage if guard
321 	 * pages are not used.
322 	 *
323 	 * XXX This scheme will lead to greater fragmentation of the
324 	 * map, unless vm_map_findspace() is tweaked.
325 	 */
326 	for (;;) {
327 		if (vmem_xalloc(memguard_arena, size_v, 0, 0, 0,
328 		    memguard_cursor, VMEM_ADDR_MAX,
329 		    M_BESTFIT | M_NOWAIT, &addr) == 0)
330 			break;
331 		/*
332 		 * The map has no space.  This may be due to
333 		 * fragmentation, or because the cursor is near the
334 		 * end of the map.
335 		 */
336 		if (memguard_cursor == memguard_base) {
337 			memguard_fail_kva++;
338 			addr = (vm_offset_t)NULL;
339 			goto out;
340 		}
341 		memguard_wrap++;
342 		memguard_cursor = memguard_base;
343 	}
344 	if (do_guard)
345 		addr += PAGE_SIZE;
346 	rv = kmem_back(kmem_object, addr, size_p, flags);
347 	if (rv != KERN_SUCCESS) {
348 		vmem_xfree(memguard_arena, addr, size_v);
349 		memguard_fail_pgs++;
350 		addr = (vm_offset_t)NULL;
351 		goto out;
352 	}
353 	memguard_cursor = addr + size_v;
354 	*v2sizep(trunc_page(addr)) = req_size;
355 	*v2sizev(trunc_page(addr)) = size_v;
356 	memguard_succ++;
357 	if (req_size < PAGE_SIZE) {
358 		memguard_wasted += (PAGE_SIZE - req_size);
359 		if (do_guard) {
360 			/*
361 			 * Align the request to 16 bytes, and return
362 			 * an address near the end of the page, to
363 			 * better detect array overrun.
364 			 */
365 			req_size = roundup2(req_size, 16);
366 			addr += (PAGE_SIZE - req_size);
367 		}
368 	}
369 out:
370 	return ((void *)addr);
371 }
372 
373 int
374 is_memguard_addr(void *addr)
375 {
376 	vm_offset_t a = (vm_offset_t)(uintptr_t)addr;
377 
378 	return (a >= memguard_base && a < memguard_base + memguard_mapsize);
379 }
380 
381 /*
382  * Free specified single object.
383  */
384 void
385 memguard_free(void *ptr)
386 {
387 	vm_offset_t addr;
388 	u_long req_size, size, sizev;
389 	char *temp;
390 	int i;
391 
392 	addr = trunc_page((uintptr_t)ptr);
393 	req_size = *v2sizep(addr);
394 	sizev = *v2sizev(addr);
395 	size = round_page(req_size);
396 
397 	/*
398 	 * Page should not be guarded right now, so force a write.
399 	 * The purpose of this is to increase the likelihood of
400 	 * catching a double-free, but not necessarily a
401 	 * tamper-after-free (the second thread freeing might not
402 	 * write before freeing, so this forces it to and,
403 	 * subsequently, trigger a fault).
404 	 */
405 	temp = ptr;
406 	for (i = 0; i < size; i += PAGE_SIZE)
407 		temp[i] = 'M';
408 
409 	/*
410 	 * This requires carnal knowledge of the implementation of
411 	 * kmem_free(), but since we've already replaced kmem_malloc()
412 	 * above, it's not really any worse.  We want to use the
413 	 * vm_map lock to serialize updates to memguard_wasted, since
414 	 * we had the lock at increment.
415 	 */
416 	kmem_unback(kmem_object, addr, size);
417 	if (sizev > size)
418 		addr -= PAGE_SIZE;
419 	vmem_xfree(memguard_arena, addr, sizev);
420 	if (req_size < PAGE_SIZE)
421 		memguard_wasted -= (PAGE_SIZE - req_size);
422 }
423 
424 /*
425  * Re-allocate an allocation that was originally guarded.
426  */
427 void *
428 memguard_realloc(void *addr, unsigned long size, struct malloc_type *mtp,
429     int flags)
430 {
431 	void *newaddr;
432 	u_long old_size;
433 
434 	/*
435 	 * Allocate the new block.  Force the allocation to be guarded
436 	 * as the original may have been guarded through random
437 	 * chance, and that should be preserved.
438 	 */
439 	if ((newaddr = memguard_alloc(size, flags)) == NULL)
440 		return (NULL);
441 
442 	/* Copy over original contents. */
443 	old_size = *v2sizep(trunc_page((uintptr_t)addr));
444 	bcopy(addr, newaddr, min(size, old_size));
445 	memguard_free(addr);
446 	return (newaddr);
447 }
448 
449 static int
450 memguard_cmp(unsigned long size)
451 {
452 
453 	if (size < memguard_minsize) {
454 		memguard_minsize_reject++;
455 		return (0);
456 	}
457 	if ((memguard_options & MG_GUARD_ALLLARGE) != 0 && size >= PAGE_SIZE)
458 		return (1);
459 	if (memguard_frequency > 0 &&
460 	    (random() % 100000) < memguard_frequency) {
461 		memguard_frequency_hits++;
462 		return (1);
463 	}
464 
465 	return (0);
466 }
467 
468 int
469 memguard_cmp_mtp(struct malloc_type *mtp, unsigned long size)
470 {
471 
472 	if (memguard_cmp(size))
473 		return(1);
474 
475 #if 1
476 	/*
477 	 * The safest way of comparsion is to always compare short description
478 	 * string of memory type, but it is also the slowest way.
479 	 */
480 	return (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0);
481 #else
482 	/*
483 	 * If we compare pointers, there are two possible problems:
484 	 * 1. Memory type was unloaded and new memory type was allocated at the
485 	 *    same address.
486 	 * 2. Memory type was unloaded and loaded again, but allocated at a
487 	 *    different address.
488 	 */
489 	if (vm_memguard_mtype != NULL)
490 		return (mtp == vm_memguard_mtype);
491 	if (strcmp(mtp->ks_shortdesc, vm_memguard_desc) == 0) {
492 		vm_memguard_mtype = mtp;
493 		return (1);
494 	}
495 	return (0);
496 #endif
497 }
498 
499 int
500 memguard_cmp_zone(uma_zone_t zone)
501 {
502 
503 	if ((memguard_options & MG_GUARD_NOFREE) == 0 &&
504 	    zone->uz_flags & UMA_ZONE_NOFREE)
505 		return (0);
506 
507 	if (memguard_cmp(zone->uz_size))
508 		return (1);
509 
510 	/*
511 	 * The safest way of comparsion is to always compare zone name,
512 	 * but it is also the slowest way.
513 	 */
514 	return (strcmp(zone->uz_name, vm_memguard_desc) == 0);
515 }
516