1 /*
2 * Copyright 2013-2014 Andrew Turner.
3 * Copyright 2013-2014 Ian Lepore.
4 * Copyright 2013-2014 Rui Paulo.
5 * Copyright 2013 Eitan Adler.
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 are
10 * met:
11 *
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following 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 AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
21 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE
22 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
23 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
24 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
25 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
26 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
27 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
28 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 */
30
31 #include <sys/cdefs.h>
32 #include <sys/param.h>
33 #include <sys/kernel.h>
34 #include <sys/linker.h>
35 #include <sys/malloc.h>
36 #include <sys/proc.h>
37 #include <sys/queue.h>
38 #include <sys/systm.h>
39
40 #include <machine/machdep.h>
41 #include <machine/stack.h>
42
43 #include "linker_if.h"
44
45 /*
46 * Definitions for the instruction interpreter.
47 *
48 * The ARM EABI specifies how to perform the frame unwinding in the
49 * Exception Handling ABI for the ARM Architecture document. To perform
50 * the unwind we need to know the initial frame pointer, stack pointer,
51 * link register and program counter. We then find the entry within the
52 * index table that points to the function the program counter is within.
53 * This gives us either a list of three instructions to process, a 31-bit
54 * relative offset to a table of instructions, or a value telling us
55 * we can't unwind any further.
56 *
57 * When we have the instructions to process we need to decode them
58 * following table 4 in section 9.3. This describes a collection of bit
59 * patterns to encode that steps to take to update the stack pointer and
60 * link register to the correct values at the start of the function.
61 */
62
63 /* A special case when we are unable to unwind past this function */
64 #define EXIDX_CANTUNWIND 1
65
66 /*
67 * Entry types.
68 * These are the only entry types that have been seen in the kernel.
69 */
70 #define ENTRY_MASK 0xff000000
71 #define ENTRY_ARM_SU16 0x80000000
72 #define ENTRY_ARM_LU16 0x81000000
73
74 /* Instruction masks. */
75 #define INSN_VSP_MASK 0xc0
76 #define INSN_VSP_SIZE_MASK 0x3f
77 #define INSN_STD_MASK 0xf0
78 #define INSN_STD_DATA_MASK 0x0f
79 #define INSN_POP_TYPE_MASK 0x08
80 #define INSN_POP_COUNT_MASK 0x07
81 #define INSN_VSP_LARGE_INC_MASK 0xff
82
83 /* Instruction definitions */
84 #define INSN_VSP_INC 0x00
85 #define INSN_VSP_DEC 0x40
86 #define INSN_POP_MASKED 0x80
87 #define INSN_VSP_REG 0x90
88 #define INSN_POP_COUNT 0xa0
89 #define INSN_FINISH 0xb0
90 #define INSN_POP_REGS 0xb1
91 #define INSN_VSP_LARGE_INC 0xb2
92
93 /* An item in the exception index table */
94 struct unwind_idx {
95 uint32_t offset;
96 uint32_t insn;
97 };
98
99 /*
100 * Local cache of unwind info for loaded modules.
101 *
102 * To unwind the stack through the code in a loaded module, we need to access
103 * the module's exidx unwind data. To locate that data, one must search the
104 * elf section headers for the SHT_ARM_EXIDX section. Those headers are
105 * available at the time the module is being loaded, but are discarded by time
106 * the load process has completed. Code in kern/link_elf.c locates the data we
107 * need and stores it into the linker_file structure before calling the arm
108 * machdep routine for handling loaded modules (in arm/elf_machdep.c). That
109 * function calls into this code to pass along the unwind info, which we save
110 * into one of these module_info structures.
111 *
112 * Because we have to help stack(9) gather stack info at any time, including in
113 * contexts where sleeping is not allowed, we cannot use linker_file_foreach()
114 * to walk the kernel's list of linker_file structs, because doing so requires
115 * acquiring an exclusive sx_lock. So instead, we keep a local list of these
116 * structures, one for each loaded module (and one for the kernel itself that we
117 * synthesize at init time). New entries are added to the end of this list as
118 * needed, but entries are never deleted from the list. Instead, they are
119 * cleared out in-place to mark them as unused. That means the code doing stack
120 * unwinding can always safely walk the list without locking, because the
121 * structure of the list never changes in a way that would cause the walker to
122 * follow a bad link.
123 *
124 * A cleared-out entry on the list has module start=UINTPTR_MAX and end=0, so
125 * start <= addr < end cannot be true for any value of addr being searched for.
126 * We also don't have to worry about races where we look up the unwind info just
127 * before a module is unloaded and try to access it concurrently with or just
128 * after the unloading happens in another thread, because that means the path of
129 * execution leads through a now-unloaded module, and that's already well into
130 * undefined-behavior territory.
131 *
132 * List entries marked as unused get reused when new modules are loaded. We
133 * don't worry about holding a few unused bytes of memory in the list after
134 * unloading a module.
135 */
136 struct module_info {
137 uintptr_t module_start; /* Start of loaded module */
138 uintptr_t module_end; /* End of loaded module */
139 uintptr_t exidx_start; /* Start of unwind data */
140 uintptr_t exidx_end; /* End of unwind data */
141 STAILQ_ENTRY(module_info)
142 link; /* Link to next entry */
143 };
144 static STAILQ_HEAD(, module_info) module_list;
145
146 /*
147 * Hide ugly casting in somewhat-less-ugly macros.
148 * CADDR - cast a pointer or number to caddr_t.
149 * UADDR - cast a pointer or number to uintptr_t.
150 */
151 #define CADDR(addr) ((caddr_t)(void*)(uintptr_t)(addr))
152 #define UADDR(addr) ((uintptr_t)(addr))
153
154 /*
155 * Clear the info in an existing module_info entry on the list. The
156 * module_start/end addresses are set to values that cannot match any real
157 * memory address. The entry remains on the list, but will be ignored until it
158 * is populated with new data.
159 */
160 static void
clear_module_info(struct module_info * info)161 clear_module_info(struct module_info *info)
162 {
163 info->module_start = UINTPTR_MAX;
164 info->module_end = 0;
165 }
166
167 /*
168 * Populate an existing module_info entry (which is already on the list) with
169 * the info for a new module.
170 */
171 static void
populate_module_info(struct module_info * info,linker_file_t lf)172 populate_module_info(struct module_info *info, linker_file_t lf)
173 {
174
175 /*
176 * Careful! The module_start and module_end fields must not be set
177 * until all other data in the structure is valid.
178 */
179 info->exidx_start = UADDR(lf->exidx_addr);
180 info->exidx_end = UADDR(lf->exidx_addr) + lf->exidx_size;
181 info->module_start = UADDR(lf->address);
182 info->module_end = UADDR(lf->address) + lf->size;
183 }
184
185 /*
186 * Create a new empty module_info entry and add it to the tail of the list.
187 */
188 static struct module_info *
create_module_info(void)189 create_module_info(void)
190 {
191 struct module_info *info;
192
193 info = malloc(sizeof(*info), M_CACHE, M_WAITOK | M_ZERO);
194 clear_module_info(info);
195 STAILQ_INSERT_TAIL(&module_list, info, link);
196 return (info);
197 }
198
199 /*
200 * Search for a module_info entry on the list whose address range contains the
201 * given address. If the search address is zero (no module will be loaded at
202 * zero), then we're looking for an empty item to reuse, which is indicated by
203 * module_start being set to UINTPTR_MAX in the entry.
204 */
205 static struct module_info *
find_module_info(uintptr_t addr)206 find_module_info(uintptr_t addr)
207 {
208 struct module_info *info;
209
210 STAILQ_FOREACH(info, &module_list, link) {
211 if ((addr >= info->module_start && addr < info->module_end) ||
212 (addr == 0 && info->module_start == UINTPTR_MAX))
213 return (info);
214 }
215 return (NULL);
216 }
217
218 /*
219 * Handle the loading of a new module by populating a module_info for it. This
220 * is called for both preloaded and dynamically loaded modules.
221 */
222 void
unwind_module_loaded(struct linker_file * lf)223 unwind_module_loaded(struct linker_file *lf)
224 {
225 struct module_info *info;
226
227 /*
228 * A module that contains only data may have no unwind info; don't
229 * create any module info for it.
230 */
231 if (lf->exidx_size == 0)
232 return;
233
234 /*
235 * Find an unused entry in the existing list to reuse. If we don't find
236 * one, create a new one and link it into the list. This is the only
237 * place the module_list is modified. Adding a new entry to the list
238 * will not perturb any other threads currently walking the list. This
239 * function is invoked while kern_linker is still holding its lock
240 * to prevent its module list from being modified, so we don't have to
241 * worry about racing other threads doing an insert concurrently.
242 */
243 if ((info = find_module_info(0)) == NULL) {
244 info = create_module_info();
245 }
246 populate_module_info(info, lf);
247 }
248
249 /* Handle the unloading of a module. */
250 void
unwind_module_unloaded(struct linker_file * lf)251 unwind_module_unloaded(struct linker_file *lf)
252 {
253 struct module_info *info;
254
255 /*
256 * A module that contains only data may have no unwind info and there
257 * won't be a list entry for it.
258 */
259 if (lf->exidx_size == 0)
260 return;
261
262 /*
263 * When a module is unloaded, we clear the info out of its entry in the
264 * module list, making that entry available for later reuse.
265 */
266 if ((info = find_module_info(UADDR(lf->address))) == NULL) {
267 printf("arm unwind: module '%s' not on list at unload time\n",
268 lf->filename);
269 return;
270 }
271 clear_module_info(info);
272 }
273
274 /*
275 * Initialization must run fairly early, as soon as malloc(9) is available, and
276 * definitely before witness, which uses stack(9). We synthesize a module_info
277 * entry for the kernel, because unwind_module_loaded() doesn't get called for
278 * it. Also, it is unlike other modules in that the elf metadata for locating
279 * the unwind tables might be stripped, so instead we have to use the
280 * _exidx_start/end symbols created by ldscript.arm.
281 */
282 static int
module_info_init(void * arg __unused)283 module_info_init(void *arg __unused)
284 {
285 struct linker_file thekernel;
286
287 STAILQ_INIT(&module_list);
288
289 thekernel.filename = "kernel";
290 thekernel.address = CADDR(&_start);
291 thekernel.size = UADDR(&_end) - UADDR(&_start);
292 thekernel.exidx_addr = CADDR(&_exidx_start);
293 thekernel.exidx_size = UADDR(&_exidx_end) - UADDR(&_exidx_start);
294 populate_module_info(create_module_info(), &thekernel);
295
296 return (0);
297 }
298 SYSINIT(unwind_init, SI_SUB_KMEM, SI_ORDER_ANY, module_info_init, NULL);
299
300 /* Expand a 31-bit signed value to a 32-bit signed value */
301 static __inline int32_t
expand_prel31(uint32_t prel31)302 expand_prel31(uint32_t prel31)
303 {
304
305 return ((int32_t)(prel31 & 0x7fffffffu) << 1) / 2;
306 }
307
308 /*
309 * Perform a binary search of the index table to find the function
310 * with the largest address that doesn't exceed addr.
311 */
312 static struct unwind_idx *
find_index(uint32_t addr)313 find_index(uint32_t addr)
314 {
315 struct module_info *info;
316 unsigned int min, mid, max;
317 struct unwind_idx *start;
318 struct unwind_idx *item;
319 int32_t prel31_addr;
320 uint32_t func_addr;
321
322 info = find_module_info(addr);
323 if (info == NULL)
324 return NULL;
325
326 min = 0;
327 max = (info->exidx_end - info->exidx_start) / sizeof(struct unwind_idx);
328 start = (struct unwind_idx *)CADDR(info->exidx_start);
329
330 while (min != max) {
331 mid = min + (max - min + 1) / 2;
332
333 item = &start[mid];
334
335 prel31_addr = expand_prel31(item->offset);
336 func_addr = (uint32_t)&item->offset + prel31_addr;
337
338 if (func_addr <= addr) {
339 min = mid;
340 } else {
341 max = mid - 1;
342 }
343 }
344
345 return &start[min];
346 }
347
348 /* Reads the next byte from the instruction list */
349 static uint8_t
unwind_exec_read_byte(struct unwind_state * state)350 unwind_exec_read_byte(struct unwind_state *state)
351 {
352 uint8_t insn;
353
354 /* Read the unwind instruction */
355 insn = (*state->insn) >> (state->byte * 8);
356
357 /* Update the location of the next instruction */
358 if (state->byte == 0) {
359 state->byte = 3;
360 state->insn++;
361 state->entries--;
362 } else
363 state->byte--;
364
365 return insn;
366 }
367
368 /* Executes the next instruction on the list */
369 static int
unwind_exec_insn(struct unwind_state * state)370 unwind_exec_insn(struct unwind_state *state)
371 {
372 struct thread *td = curthread;
373 unsigned int insn;
374 uint32_t *vsp = (uint32_t *)state->registers[SP];
375 int update_vsp = 0;
376
377 /* This should never happen */
378 if (state->entries == 0)
379 return 1;
380
381 /* Read the next instruction */
382 insn = unwind_exec_read_byte(state);
383
384 if ((insn & INSN_VSP_MASK) == INSN_VSP_INC) {
385 state->registers[SP] += ((insn & INSN_VSP_SIZE_MASK) << 2) + 4;
386
387 } else if ((insn & INSN_VSP_MASK) == INSN_VSP_DEC) {
388 state->registers[SP] -= ((insn & INSN_VSP_SIZE_MASK) << 2) + 4;
389
390 } else if ((insn & INSN_STD_MASK) == INSN_POP_MASKED) {
391 unsigned int mask, reg;
392
393 /* Load the mask */
394 mask = unwind_exec_read_byte(state);
395 mask |= (insn & INSN_STD_DATA_MASK) << 8;
396
397 /* We have a refuse to unwind instruction */
398 if (mask == 0)
399 return 1;
400
401 if (!__is_aligned(vsp, sizeof(register_t)))
402 return 1;
403
404 /* Update SP */
405 update_vsp = 1;
406
407 /* Load the registers */
408 for (reg = 4; mask && reg < 16; mask >>= 1, reg++) {
409 if (mask & 1) {
410 if (!kstack_contains(td, (uintptr_t)vsp,
411 sizeof(*vsp)))
412 return 1;
413
414 state->registers[reg] = *vsp++;
415 state->update_mask |= 1 << reg;
416
417 /* If we have updated SP kep its value */
418 if (reg == SP)
419 update_vsp = 0;
420 }
421 }
422
423 } else if ((insn & INSN_STD_MASK) == INSN_VSP_REG &&
424 ((insn & INSN_STD_DATA_MASK) != 13) &&
425 ((insn & INSN_STD_DATA_MASK) != 15)) {
426 /* sp = register */
427 state->registers[SP] =
428 state->registers[insn & INSN_STD_DATA_MASK];
429
430 } else if ((insn & INSN_STD_MASK) == INSN_POP_COUNT) {
431 unsigned int count, reg;
432
433 /* Read how many registers to load */
434 count = insn & INSN_POP_COUNT_MASK;
435
436 if (!__is_aligned(vsp, sizeof(register_t)))
437 return 1;
438
439 /* Update sp */
440 update_vsp = 1;
441
442 /* Pop the registers */
443 if (!kstack_contains(td, (uintptr_t)vsp,
444 sizeof(*vsp) * (4 + count)))
445 return 1;
446 for (reg = 4; reg <= 4 + count; reg++) {
447 state->registers[reg] = *vsp++;
448 state->update_mask |= 1 << reg;
449 }
450
451 /* Check if we are in the pop r14 version */
452 if ((insn & INSN_POP_TYPE_MASK) != 0) {
453 if (!kstack_contains(td, (uintptr_t)vsp, sizeof(*vsp)))
454 return 1;
455 state->registers[14] = *vsp++;
456 }
457
458 } else if (insn == INSN_FINISH) {
459 /* Stop processing */
460 state->entries = 0;
461
462 } else if (insn == INSN_POP_REGS) {
463 unsigned int mask, reg;
464
465 mask = unwind_exec_read_byte(state);
466 if (mask == 0 || (mask & 0xf0) != 0)
467 return 1;
468
469 if (!__is_aligned(vsp, sizeof(register_t)))
470 return 1;
471
472 /* Update SP */
473 update_vsp = 1;
474
475 /* Load the registers */
476 for (reg = 0; mask && reg < 4; mask >>= 1, reg++) {
477 if (mask & 1) {
478 if (!kstack_contains(td, (uintptr_t)vsp,
479 sizeof(*vsp)))
480 return 1;
481 state->registers[reg] = *vsp++;
482 state->update_mask |= 1 << reg;
483 }
484 }
485
486 } else if ((insn & INSN_VSP_LARGE_INC_MASK) == INSN_VSP_LARGE_INC) {
487 unsigned int uleb128;
488
489 /* Read the increment value */
490 uleb128 = unwind_exec_read_byte(state);
491
492 state->registers[SP] += 0x204 + (uleb128 << 2);
493
494 } else {
495 /* We hit a new instruction that needs to be implemented */
496 #if 0
497 db_printf("Unhandled instruction %.2x\n", insn);
498 #endif
499 return 1;
500 }
501
502 if (update_vsp) {
503 state->registers[SP] = (uint32_t)vsp;
504 }
505
506 #if 0
507 db_printf("fp = %08x, sp = %08x, lr = %08x, pc = %08x\n",
508 state->registers[FP], state->registers[SP], state->registers[LR],
509 state->registers[PC]);
510 #endif
511
512 return 0;
513 }
514
515 /* Performs the unwind of a function */
516 static int
unwind_tab(struct unwind_state * state)517 unwind_tab(struct unwind_state *state)
518 {
519 uint32_t entry;
520
521 /* Set PC to a known value */
522 state->registers[PC] = 0;
523
524 /* Read the personality */
525 entry = *state->insn & ENTRY_MASK;
526
527 if (entry == ENTRY_ARM_SU16) {
528 state->byte = 2;
529 state->entries = 1;
530 } else if (entry == ENTRY_ARM_LU16) {
531 state->byte = 1;
532 state->entries = ((*state->insn >> 16) & 0xFF) + 1;
533 } else {
534 #if 0
535 db_printf("Unknown entry: %x\n", entry);
536 #endif
537 return 1;
538 }
539
540 while (state->entries > 0) {
541 if (unwind_exec_insn(state) != 0)
542 return 1;
543 }
544
545 /*
546 * The program counter was not updated, load it from the link register.
547 */
548 if (state->registers[PC] == 0) {
549 state->registers[PC] = state->registers[LR];
550
551 /*
552 * If the program counter changed, flag it in the update mask.
553 */
554 if (state->start_pc != state->registers[PC])
555 state->update_mask |= 1 << PC;
556 }
557
558 return 0;
559 }
560
561 /*
562 * Unwind a single stack frame.
563 * Return 0 on success or 1 if the stack cannot be unwound any further.
564 *
565 * XXX The can_lock argument is no longer germane; a sweep of callers should be
566 * made to remove it after this new code has proven itself for a while.
567 */
568 int
unwind_stack_one(struct unwind_state * state,int can_lock __unused)569 unwind_stack_one(struct unwind_state *state, int can_lock __unused)
570 {
571 struct unwind_idx *index;
572
573 /* Reset the mask of updated registers */
574 state->update_mask = 0;
575
576 /* The pc value is correct and will be overwritten, save it */
577 state->start_pc = state->registers[PC];
578
579 /* Find the item to run */
580 index = find_index(state->start_pc);
581 if (index == NULL || index->insn == EXIDX_CANTUNWIND)
582 return 1;
583
584 if (index->insn & (1U << 31)) {
585 /* The data is within the instruction */
586 state->insn = &index->insn;
587 } else {
588 /* A prel31 offset to the unwind table */
589 state->insn = (uint32_t *)
590 ((uintptr_t)&index->insn +
591 expand_prel31(index->insn));
592 }
593
594 /* Run the unwind function, return its finished/not-finished status. */
595 return (unwind_tab(state));
596 }
597