]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
57319d80 QR |
2 | /* |
3 | * mpx.c - Memory Protection eXtensions | |
4 | * | |
5 | * Copyright (c) 2014, Intel Corporation. | |
6 | * Qiaowei Ren <qiaowei.ren@intel.com> | |
7 | * Dave Hansen <dave.hansen@intel.com> | |
8 | */ | |
9 | #include <linux/kernel.h> | |
fcc7ffd6 | 10 | #include <linux/slab.h> |
589ee628 | 11 | #include <linux/mm_types.h> |
57319d80 QR |
12 | #include <linux/syscalls.h> |
13 | #include <linux/sched/sysctl.h> | |
14 | ||
fe3d197f | 15 | #include <asm/insn.h> |
32542ee2 | 16 | #include <asm/insn-eval.h> |
57319d80 | 17 | #include <asm/mman.h> |
1de4fa14 | 18 | #include <asm/mmu_context.h> |
57319d80 | 19 | #include <asm/mpx.h> |
fe3d197f | 20 | #include <asm/processor.h> |
78f7f1e5 | 21 | #include <asm/fpu/internal.h> |
57319d80 | 22 | |
e7126cf5 DH |
23 | #define CREATE_TRACE_POINTS |
24 | #include <asm/trace/mpx.h> | |
25 | ||
613fcb7d DH |
26 | static inline unsigned long mpx_bd_size_bytes(struct mm_struct *mm) |
27 | { | |
28 | if (is_64bit_mm(mm)) | |
29 | return MPX_BD_SIZE_BYTES_64; | |
30 | else | |
31 | return MPX_BD_SIZE_BYTES_32; | |
32 | } | |
33 | ||
34 | static inline unsigned long mpx_bt_size_bytes(struct mm_struct *mm) | |
35 | { | |
36 | if (is_64bit_mm(mm)) | |
37 | return MPX_BT_SIZE_BYTES_64; | |
38 | else | |
39 | return MPX_BT_SIZE_BYTES_32; | |
40 | } | |
41 | ||
57319d80 QR |
42 | /* |
43 | * This is really a simplified "vm_mmap". it only handles MPX | |
44 | * bounds tables (the bounds directory is user-allocated). | |
57319d80 QR |
45 | */ |
46 | static unsigned long mpx_mmap(unsigned long len) | |
47 | { | |
57319d80 | 48 | struct mm_struct *mm = current->mm; |
1fcfd8db | 49 | unsigned long addr, populate; |
57319d80 | 50 | |
eb099e5b | 51 | /* Only bounds table can be allocated here */ |
613fcb7d | 52 | if (len != mpx_bt_size_bytes(mm)) |
57319d80 QR |
53 | return -EINVAL; |
54 | ||
55 | down_write(&mm->mmap_sem); | |
1fcfd8db | 56 | addr = do_mmap(NULL, 0, len, PROT_READ | PROT_WRITE, |
897ab3e0 | 57 | MAP_ANONYMOUS | MAP_PRIVATE, VM_MPX, 0, &populate, NULL); |
57319d80 | 58 | up_write(&mm->mmap_sem); |
1fcfd8db ON |
59 | if (populate) |
60 | mm_populate(addr, populate); | |
61 | ||
62 | return addr; | |
57319d80 | 63 | } |
fcc7ffd6 | 64 | |
fcc7ffd6 DH |
65 | static int mpx_insn_decode(struct insn *insn, |
66 | struct pt_regs *regs) | |
67 | { | |
68 | unsigned char buf[MAX_INSN_SIZE]; | |
69 | int x86_64 = !test_thread_flag(TIF_IA32); | |
70 | int not_copied; | |
71 | int nr_copied; | |
72 | ||
73 | not_copied = copy_from_user(buf, (void __user *)regs->ip, sizeof(buf)); | |
74 | nr_copied = sizeof(buf) - not_copied; | |
75 | /* | |
76 | * The decoder _should_ fail nicely if we pass it a short buffer. | |
77 | * But, let's not depend on that implementation detail. If we | |
78 | * did not get anything, just error out now. | |
79 | */ | |
80 | if (!nr_copied) | |
81 | return -EFAULT; | |
82 | insn_init(insn, buf, nr_copied, x86_64); | |
83 | insn_get_length(insn); | |
84 | /* | |
85 | * copy_from_user() tries to get as many bytes as we could see in | |
86 | * the largest possible instruction. If the instruction we are | |
87 | * after is shorter than that _and_ we attempt to copy from | |
88 | * something unreadable, we might get a short read. This is OK | |
89 | * as long as the read did not stop in the middle of the | |
90 | * instruction. Check to see if we got a partial instruction. | |
91 | */ | |
92 | if (nr_copied < insn->length) | |
93 | return -EFAULT; | |
94 | ||
95 | insn_get_opcode(insn); | |
96 | /* | |
97 | * We only _really_ need to decode bndcl/bndcn/bndcu | |
98 | * Error out on anything else. | |
99 | */ | |
100 | if (insn->opcode.bytes[0] != 0x0f) | |
101 | goto bad_opcode; | |
102 | if ((insn->opcode.bytes[1] != 0x1a) && | |
103 | (insn->opcode.bytes[1] != 0x1b)) | |
104 | goto bad_opcode; | |
105 | ||
106 | return 0; | |
107 | bad_opcode: | |
108 | return -EINVAL; | |
109 | } | |
110 | ||
111 | /* | |
112 | * If a bounds overflow occurs then a #BR is generated. This | |
113 | * function decodes MPX instructions to get violation address | |
114 | * and set this address into extended struct siginfo. | |
115 | * | |
116 | * Note that this is not a super precise way of doing this. | |
117 | * Userspace could have, by the time we get here, written | |
118 | * anything it wants in to the instructions. We can not | |
119 | * trust anything about it. They might not be valid | |
120 | * instructions or might encode invalid registers, etc... | |
fcc7ffd6 | 121 | */ |
8d68fa0e | 122 | int mpx_fault_info(struct mpx_fault_info *info, struct pt_regs *regs) |
fcc7ffd6 | 123 | { |
1126cb45 DH |
124 | const struct mpx_bndreg_state *bndregs; |
125 | const struct mpx_bndreg *bndreg; | |
fcc7ffd6 DH |
126 | struct insn insn; |
127 | uint8_t bndregno; | |
128 | int err; | |
fcc7ffd6 DH |
129 | |
130 | err = mpx_insn_decode(&insn, regs); | |
131 | if (err) | |
132 | goto err_out; | |
133 | ||
134 | /* | |
135 | * We know at this point that we are only dealing with | |
136 | * MPX instructions. | |
137 | */ | |
138 | insn_get_modrm(&insn); | |
139 | bndregno = X86_MODRM_REG(insn.modrm.value); | |
140 | if (bndregno > 3) { | |
141 | err = -EINVAL; | |
142 | goto err_out; | |
143 | } | |
a84eeaa9 | 144 | /* get bndregs field from current task's xsave area */ |
d91cab78 | 145 | bndregs = get_xsave_field_ptr(XFEATURE_MASK_BNDREGS); |
fe3d197f DH |
146 | if (!bndregs) { |
147 | err = -EINVAL; | |
148 | goto err_out; | |
149 | } | |
150 | /* now go select the individual register in the set of 4 */ | |
1126cb45 | 151 | bndreg = &bndregs->bndreg[bndregno]; |
fe3d197f | 152 | |
fcc7ffd6 DH |
153 | /* |
154 | * The registers are always 64-bit, but the upper 32 | |
155 | * bits are ignored in 32-bit mode. Also, note that the | |
156 | * upper bounds are architecturally represented in 1's | |
157 | * complement form. | |
158 | * | |
159 | * The 'unsigned long' cast is because the compiler | |
160 | * complains when casting from integers to different-size | |
161 | * pointers. | |
162 | */ | |
8d68fa0e EB |
163 | info->lower = (void __user *)(unsigned long)bndreg->lower_bound; |
164 | info->upper = (void __user *)(unsigned long)~bndreg->upper_bound; | |
165 | info->addr = insn_get_addr_ref(&insn, regs); | |
166 | ||
fcc7ffd6 DH |
167 | /* |
168 | * We were not able to extract an address from the instruction, | |
169 | * probably because there was something invalid in it. | |
170 | */ | |
8d68fa0e | 171 | if (info->addr == (void __user *)-1) { |
fcc7ffd6 DH |
172 | err = -EINVAL; |
173 | goto err_out; | |
174 | } | |
8d68fa0e EB |
175 | trace_mpx_bounds_register_exception(info->addr, bndreg); |
176 | return 0; | |
fcc7ffd6 | 177 | err_out: |
fe3d197f | 178 | /* info might be NULL, but kfree() handles that */ |
8d68fa0e | 179 | return err; |
fcc7ffd6 | 180 | } |
fe3d197f | 181 | |
46a6e0cf | 182 | static __user void *mpx_get_bounds_dir(void) |
fe3d197f | 183 | { |
1126cb45 | 184 | const struct mpx_bndcsr *bndcsr; |
fe3d197f DH |
185 | |
186 | if (!cpu_feature_enabled(X86_FEATURE_MPX)) | |
187 | return MPX_INVALID_BOUNDS_DIR; | |
188 | ||
189 | /* | |
190 | * The bounds directory pointer is stored in a register | |
191 | * only accessible if we first do an xsave. | |
192 | */ | |
d91cab78 | 193 | bndcsr = get_xsave_field_ptr(XFEATURE_MASK_BNDCSR); |
fe3d197f DH |
194 | if (!bndcsr) |
195 | return MPX_INVALID_BOUNDS_DIR; | |
196 | ||
197 | /* | |
198 | * Make sure the register looks valid by checking the | |
199 | * enable bit. | |
200 | */ | |
201 | if (!(bndcsr->bndcfgu & MPX_BNDCFG_ENABLE_FLAG)) | |
202 | return MPX_INVALID_BOUNDS_DIR; | |
203 | ||
204 | /* | |
205 | * Lastly, mask off the low bits used for configuration | |
206 | * flags, and return the address of the bounds table. | |
207 | */ | |
208 | return (void __user *)(unsigned long) | |
209 | (bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK); | |
210 | } | |
211 | ||
46a6e0cf | 212 | int mpx_enable_management(void) |
fe3d197f DH |
213 | { |
214 | void __user *bd_base = MPX_INVALID_BOUNDS_DIR; | |
46a6e0cf | 215 | struct mm_struct *mm = current->mm; |
fe3d197f DH |
216 | int ret = 0; |
217 | ||
218 | /* | |
219 | * runtime in the userspace will be responsible for allocation of | |
220 | * the bounds directory. Then, it will save the base of the bounds | |
221 | * directory into XSAVE/XRSTOR Save Area and enable MPX through | |
222 | * XRSTOR instruction. | |
223 | * | |
a84eeaa9 DH |
224 | * The copy_xregs_to_kernel() beneath get_xsave_field_ptr() is |
225 | * expected to be relatively expensive. Storing the bounds | |
226 | * directory here means that we do not have to do xsave in the | |
cb02de96 | 227 | * unmap path; we can just use mm->context.bd_addr instead. |
fe3d197f | 228 | */ |
46a6e0cf | 229 | bd_base = mpx_get_bounds_dir(); |
fe3d197f | 230 | down_write(&mm->mmap_sem); |
44b04912 KS |
231 | |
232 | /* MPX doesn't support addresses above 47 bits yet. */ | |
233 | if (find_vma(mm, DEFAULT_MAP_WINDOW)) { | |
234 | pr_warn_once("%s (%d): MPX cannot handle addresses " | |
235 | "above 47-bits. Disabling.", | |
236 | current->comm, current->pid); | |
237 | ret = -ENXIO; | |
238 | goto out; | |
239 | } | |
cb02de96 MR |
240 | mm->context.bd_addr = bd_base; |
241 | if (mm->context.bd_addr == MPX_INVALID_BOUNDS_DIR) | |
fe3d197f | 242 | ret = -ENXIO; |
44b04912 | 243 | out: |
fe3d197f DH |
244 | up_write(&mm->mmap_sem); |
245 | return ret; | |
246 | } | |
247 | ||
46a6e0cf | 248 | int mpx_disable_management(void) |
fe3d197f DH |
249 | { |
250 | struct mm_struct *mm = current->mm; | |
251 | ||
252 | if (!cpu_feature_enabled(X86_FEATURE_MPX)) | |
253 | return -ENXIO; | |
254 | ||
255 | down_write(&mm->mmap_sem); | |
cb02de96 | 256 | mm->context.bd_addr = MPX_INVALID_BOUNDS_DIR; |
fe3d197f DH |
257 | up_write(&mm->mmap_sem); |
258 | return 0; | |
259 | } | |
260 | ||
6ac52bb4 DH |
261 | static int mpx_cmpxchg_bd_entry(struct mm_struct *mm, |
262 | unsigned long *curval, | |
263 | unsigned long __user *addr, | |
264 | unsigned long old_val, unsigned long new_val) | |
265 | { | |
266 | int ret; | |
267 | /* | |
268 | * user_atomic_cmpxchg_inatomic() actually uses sizeof() | |
269 | * the pointer that we pass to it to figure out how much | |
270 | * data to cmpxchg. We have to be careful here not to | |
271 | * pass a pointer to a 64-bit data type when we only want | |
272 | * a 32-bit copy. | |
273 | */ | |
274 | if (is_64bit_mm(mm)) { | |
275 | ret = user_atomic_cmpxchg_inatomic(curval, | |
276 | addr, old_val, new_val); | |
277 | } else { | |
278 | u32 uninitialized_var(curval_32); | |
279 | u32 old_val_32 = old_val; | |
280 | u32 new_val_32 = new_val; | |
281 | u32 __user *addr_32 = (u32 __user *)addr; | |
282 | ||
283 | ret = user_atomic_cmpxchg_inatomic(&curval_32, | |
284 | addr_32, old_val_32, new_val_32); | |
285 | *curval = curval_32; | |
286 | } | |
287 | return ret; | |
288 | } | |
289 | ||
fe3d197f | 290 | /* |
613fcb7d DH |
291 | * With 32-bit mode, a bounds directory is 4MB, and the size of each |
292 | * bounds table is 16KB. With 64-bit mode, a bounds directory is 2GB, | |
fe3d197f DH |
293 | * and the size of each bounds table is 4MB. |
294 | */ | |
613fcb7d | 295 | static int allocate_bt(struct mm_struct *mm, long __user *bd_entry) |
fe3d197f DH |
296 | { |
297 | unsigned long expected_old_val = 0; | |
298 | unsigned long actual_old_val = 0; | |
299 | unsigned long bt_addr; | |
a1149fc8 | 300 | unsigned long bd_new_entry; |
fe3d197f DH |
301 | int ret = 0; |
302 | ||
303 | /* | |
304 | * Carve the virtual space out of userspace for the new | |
305 | * bounds table: | |
306 | */ | |
613fcb7d | 307 | bt_addr = mpx_mmap(mpx_bt_size_bytes(mm)); |
fe3d197f DH |
308 | if (IS_ERR((void *)bt_addr)) |
309 | return PTR_ERR((void *)bt_addr); | |
310 | /* | |
311 | * Set the valid flag (kinda like _PAGE_PRESENT in a pte) | |
312 | */ | |
a1149fc8 | 313 | bd_new_entry = bt_addr | MPX_BD_ENTRY_VALID_FLAG; |
fe3d197f DH |
314 | |
315 | /* | |
316 | * Go poke the address of the new bounds table in to the | |
317 | * bounds directory entry out in userspace memory. Note: | |
318 | * we may race with another CPU instantiating the same table. | |
319 | * In that case the cmpxchg will see an unexpected | |
320 | * 'actual_old_val'. | |
321 | * | |
322 | * This can fault, but that's OK because we do not hold | |
323 | * mmap_sem at this point, unlike some of the other part | |
324 | * of the MPX code that have to pagefault_disable(). | |
325 | */ | |
6ac52bb4 DH |
326 | ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val, bd_entry, |
327 | expected_old_val, bd_new_entry); | |
fe3d197f DH |
328 | if (ret) |
329 | goto out_unmap; | |
330 | ||
331 | /* | |
332 | * The user_atomic_cmpxchg_inatomic() will only return nonzero | |
333 | * for faults, *not* if the cmpxchg itself fails. Now we must | |
334 | * verify that the cmpxchg itself completed successfully. | |
335 | */ | |
336 | /* | |
337 | * We expected an empty 'expected_old_val', but instead found | |
338 | * an apparently valid entry. Assume we raced with another | |
339 | * thread to instantiate this table and desclare succecss. | |
340 | */ | |
341 | if (actual_old_val & MPX_BD_ENTRY_VALID_FLAG) { | |
342 | ret = 0; | |
343 | goto out_unmap; | |
344 | } | |
345 | /* | |
346 | * We found a non-empty bd_entry but it did not have the | |
347 | * VALID_FLAG set. Return an error which will result in | |
348 | * a SEGV since this probably means that somebody scribbled | |
349 | * some invalid data in to a bounds table. | |
350 | */ | |
351 | if (expected_old_val != actual_old_val) { | |
352 | ret = -EINVAL; | |
353 | goto out_unmap; | |
354 | } | |
cd4996dc | 355 | trace_mpx_new_bounds_table(bt_addr); |
fe3d197f DH |
356 | return 0; |
357 | out_unmap: | |
613fcb7d | 358 | vm_munmap(bt_addr, mpx_bt_size_bytes(mm)); |
fe3d197f DH |
359 | return ret; |
360 | } | |
361 | ||
362 | /* | |
363 | * When a BNDSTX instruction attempts to save bounds to a bounds | |
364 | * table, it will first attempt to look up the table in the | |
365 | * first-level bounds directory. If it does not find a table in | |
366 | * the directory, a #BR is generated and we get here in order to | |
367 | * allocate a new table. | |
368 | * | |
369 | * With 32-bit mode, the size of BD is 4MB, and the size of each | |
370 | * bound table is 16KB. With 64-bit mode, the size of BD is 2GB, | |
371 | * and the size of each bound table is 4MB. | |
372 | */ | |
46a6e0cf | 373 | static int do_mpx_bt_fault(void) |
fe3d197f DH |
374 | { |
375 | unsigned long bd_entry, bd_base; | |
1126cb45 | 376 | const struct mpx_bndcsr *bndcsr; |
613fcb7d | 377 | struct mm_struct *mm = current->mm; |
fe3d197f | 378 | |
d91cab78 | 379 | bndcsr = get_xsave_field_ptr(XFEATURE_MASK_BNDCSR); |
fe3d197f DH |
380 | if (!bndcsr) |
381 | return -EINVAL; | |
382 | /* | |
383 | * Mask off the preserve and enable bits | |
384 | */ | |
385 | bd_base = bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK; | |
386 | /* | |
387 | * The hardware provides the address of the missing or invalid | |
388 | * entry via BNDSTATUS, so we don't have to go look it up. | |
389 | */ | |
390 | bd_entry = bndcsr->bndstatus & MPX_BNDSTA_ADDR_MASK; | |
391 | /* | |
392 | * Make sure the directory entry is within where we think | |
393 | * the directory is. | |
394 | */ | |
395 | if ((bd_entry < bd_base) || | |
613fcb7d | 396 | (bd_entry >= bd_base + mpx_bd_size_bytes(mm))) |
fe3d197f DH |
397 | return -EINVAL; |
398 | ||
613fcb7d | 399 | return allocate_bt(mm, (long __user *)bd_entry); |
fe3d197f DH |
400 | } |
401 | ||
46a6e0cf | 402 | int mpx_handle_bd_fault(void) |
fe3d197f DH |
403 | { |
404 | /* | |
405 | * Userspace never asked us to manage the bounds tables, | |
406 | * so refuse to help. | |
407 | */ | |
408 | if (!kernel_managing_mpx_tables(current->mm)) | |
409 | return -EINVAL; | |
410 | ||
5ed386ec | 411 | return do_mpx_bt_fault(); |
fe3d197f | 412 | } |
1de4fa14 DH |
413 | |
414 | /* | |
415 | * A thin wrapper around get_user_pages(). Returns 0 if the | |
416 | * fault was resolved or -errno if not. | |
417 | */ | |
418 | static int mpx_resolve_fault(long __user *addr, int write) | |
419 | { | |
420 | long gup_ret; | |
421 | int nr_pages = 1; | |
1de4fa14 | 422 | |
768ae309 LS |
423 | gup_ret = get_user_pages((unsigned long)addr, nr_pages, |
424 | write ? FOLL_WRITE : 0, NULL, NULL); | |
1de4fa14 DH |
425 | /* |
426 | * get_user_pages() returns number of pages gotten. | |
427 | * 0 means we failed to fault in and get anything, | |
428 | * probably because 'addr' is bad. | |
429 | */ | |
430 | if (!gup_ret) | |
431 | return -EFAULT; | |
432 | /* Other error, return it */ | |
433 | if (gup_ret < 0) | |
434 | return gup_ret; | |
435 | /* must have gup'd a page and gup_ret>0, success */ | |
436 | return 0; | |
437 | } | |
438 | ||
54587653 DH |
439 | static unsigned long mpx_bd_entry_to_bt_addr(struct mm_struct *mm, |
440 | unsigned long bd_entry) | |
441 | { | |
442 | unsigned long bt_addr = bd_entry; | |
443 | int align_to_bytes; | |
444 | /* | |
445 | * Bit 0 in a bt_entry is always the valid bit. | |
446 | */ | |
447 | bt_addr &= ~MPX_BD_ENTRY_VALID_FLAG; | |
448 | /* | |
449 | * Tables are naturally aligned at 8-byte boundaries | |
450 | * on 64-bit and 4-byte boundaries on 32-bit. The | |
451 | * documentation makes it appear that the low bits | |
452 | * are ignored by the hardware, so we do the same. | |
453 | */ | |
454 | if (is_64bit_mm(mm)) | |
455 | align_to_bytes = 8; | |
456 | else | |
457 | align_to_bytes = 4; | |
458 | bt_addr &= ~(align_to_bytes-1); | |
459 | return bt_addr; | |
460 | } | |
461 | ||
46561c39 DH |
462 | /* |
463 | * We only want to do a 4-byte get_user() on 32-bit. Otherwise, | |
464 | * we might run off the end of the bounds table if we are on | |
465 | * a 64-bit kernel and try to get 8 bytes. | |
466 | */ | |
6bce725a | 467 | static int get_user_bd_entry(struct mm_struct *mm, unsigned long *bd_entry_ret, |
46561c39 DH |
468 | long __user *bd_entry_ptr) |
469 | { | |
470 | u32 bd_entry_32; | |
471 | int ret; | |
472 | ||
473 | if (is_64bit_mm(mm)) | |
474 | return get_user(*bd_entry_ret, bd_entry_ptr); | |
475 | ||
476 | /* | |
477 | * Note that get_user() uses the type of the *pointer* to | |
478 | * establish the size of the get, not the destination. | |
479 | */ | |
480 | ret = get_user(bd_entry_32, (u32 __user *)bd_entry_ptr); | |
481 | *bd_entry_ret = bd_entry_32; | |
482 | return ret; | |
483 | } | |
484 | ||
1de4fa14 DH |
485 | /* |
486 | * Get the base of bounds tables pointed by specific bounds | |
487 | * directory entry. | |
488 | */ | |
489 | static int get_bt_addr(struct mm_struct *mm, | |
54587653 DH |
490 | long __user *bd_entry_ptr, |
491 | unsigned long *bt_addr_result) | |
1de4fa14 DH |
492 | { |
493 | int ret; | |
494 | int valid_bit; | |
54587653 DH |
495 | unsigned long bd_entry; |
496 | unsigned long bt_addr; | |
1de4fa14 | 497 | |
96d4f267 | 498 | if (!access_ok((bd_entry_ptr), sizeof(*bd_entry_ptr))) |
1de4fa14 DH |
499 | return -EFAULT; |
500 | ||
501 | while (1) { | |
502 | int need_write = 0; | |
503 | ||
504 | pagefault_disable(); | |
46561c39 | 505 | ret = get_user_bd_entry(mm, &bd_entry, bd_entry_ptr); |
1de4fa14 DH |
506 | pagefault_enable(); |
507 | if (!ret) | |
508 | break; | |
509 | if (ret == -EFAULT) | |
54587653 | 510 | ret = mpx_resolve_fault(bd_entry_ptr, need_write); |
1de4fa14 DH |
511 | /* |
512 | * If we could not resolve the fault, consider it | |
513 | * userspace's fault and error out. | |
514 | */ | |
515 | if (ret) | |
516 | return ret; | |
517 | } | |
518 | ||
54587653 DH |
519 | valid_bit = bd_entry & MPX_BD_ENTRY_VALID_FLAG; |
520 | bt_addr = mpx_bd_entry_to_bt_addr(mm, bd_entry); | |
1de4fa14 DH |
521 | |
522 | /* | |
523 | * When the kernel is managing bounds tables, a bounds directory | |
524 | * entry will either have a valid address (plus the valid bit) | |
525 | * *OR* be completely empty. If we see a !valid entry *and* some | |
526 | * data in the address field, we know something is wrong. This | |
527 | * -EINVAL return will cause a SIGSEGV. | |
528 | */ | |
54587653 | 529 | if (!valid_bit && bt_addr) |
1de4fa14 DH |
530 | return -EINVAL; |
531 | /* | |
532 | * Do we have an completely zeroed bt entry? That is OK. It | |
533 | * just means there was no bounds table for this memory. Make | |
534 | * sure to distinguish this from -EINVAL, which will cause | |
535 | * a SEGV. | |
536 | */ | |
537 | if (!valid_bit) | |
538 | return -ENOENT; | |
539 | ||
54587653 | 540 | *bt_addr_result = bt_addr; |
1de4fa14 DH |
541 | return 0; |
542 | } | |
543 | ||
613fcb7d DH |
544 | static inline int bt_entry_size_bytes(struct mm_struct *mm) |
545 | { | |
546 | if (is_64bit_mm(mm)) | |
547 | return MPX_BT_ENTRY_BYTES_64; | |
548 | else | |
549 | return MPX_BT_ENTRY_BYTES_32; | |
550 | } | |
551 | ||
552 | /* | |
553 | * Take a virtual address and turns it in to the offset in bytes | |
554 | * inside of the bounds table where the bounds table entry | |
555 | * controlling 'addr' can be found. | |
556 | */ | |
557 | static unsigned long mpx_get_bt_entry_offset_bytes(struct mm_struct *mm, | |
558 | unsigned long addr) | |
559 | { | |
560 | unsigned long bt_table_nr_entries; | |
561 | unsigned long offset = addr; | |
562 | ||
563 | if (is_64bit_mm(mm)) { | |
564 | /* Bottom 3 bits are ignored on 64-bit */ | |
565 | offset >>= 3; | |
566 | bt_table_nr_entries = MPX_BT_NR_ENTRIES_64; | |
567 | } else { | |
568 | /* Bottom 2 bits are ignored on 32-bit */ | |
569 | offset >>= 2; | |
570 | bt_table_nr_entries = MPX_BT_NR_ENTRIES_32; | |
571 | } | |
572 | /* | |
573 | * We know the size of the table in to which we are | |
574 | * indexing, and we have eliminated all the low bits | |
575 | * which are ignored for indexing. | |
576 | * | |
577 | * Mask out all the high bits which we do not need | |
578 | * to index in to the table. Note that the tables | |
579 | * are always powers of two so this gives us a proper | |
580 | * mask. | |
581 | */ | |
582 | offset &= (bt_table_nr_entries-1); | |
583 | /* | |
584 | * We now have an entry offset in terms of *entries* in | |
585 | * the table. We need to scale it back up to bytes. | |
586 | */ | |
587 | offset *= bt_entry_size_bytes(mm); | |
588 | return offset; | |
589 | } | |
590 | ||
591 | /* | |
592 | * How much virtual address space does a single bounds | |
593 | * directory entry cover? | |
594 | * | |
595 | * Note, we need a long long because 4GB doesn't fit in | |
596 | * to a long on 32-bit. | |
597 | */ | |
598 | static inline unsigned long bd_entry_virt_space(struct mm_struct *mm) | |
599 | { | |
f3119b83 DH |
600 | unsigned long long virt_space; |
601 | unsigned long long GB = (1ULL << 30); | |
602 | ||
603 | /* | |
604 | * This covers 32-bit emulation as well as 32-bit kernels | |
6a6256f9 | 605 | * running on 64-bit hardware. |
f3119b83 DH |
606 | */ |
607 | if (!is_64bit_mm(mm)) | |
608 | return (4ULL * GB) / MPX_BD_NR_ENTRIES_32; | |
609 | ||
610 | /* | |
611 | * 'x86_virt_bits' returns what the hardware is capable | |
6a6256f9 | 612 | * of, and returns the full >32-bit address space when |
f3119b83 DH |
613 | * running 32-bit kernels on 64-bit hardware. |
614 | */ | |
615 | virt_space = (1ULL << boot_cpu_data.x86_virt_bits); | |
616 | return virt_space / MPX_BD_NR_ENTRIES_64; | |
613fcb7d DH |
617 | } |
618 | ||
619 | /* | |
3ceaccdf DH |
620 | * Free the backing physical pages of bounds table 'bt_addr'. |
621 | * Assume start...end is within that bounds table. | |
613fcb7d | 622 | */ |
3ceaccdf DH |
623 | static noinline int zap_bt_entries_mapping(struct mm_struct *mm, |
624 | unsigned long bt_addr, | |
625 | unsigned long start_mapping, unsigned long end_mapping) | |
626 | { | |
627 | struct vm_area_struct *vma; | |
628 | unsigned long addr, len; | |
629 | unsigned long start; | |
630 | unsigned long end; | |
631 | ||
632 | /* | |
633 | * if we 'end' on a boundary, the offset will be 0 which | |
634 | * is not what we want. Back it up a byte to get the | |
635 | * last bt entry. Then once we have the entry itself, | |
636 | * move 'end' back up by the table entry size. | |
637 | */ | |
638 | start = bt_addr + mpx_get_bt_entry_offset_bytes(mm, start_mapping); | |
639 | end = bt_addr + mpx_get_bt_entry_offset_bytes(mm, end_mapping - 1); | |
640 | /* | |
641 | * Move end back up by one entry. Among other things | |
642 | * this ensures that it remains page-aligned and does | |
643 | * not screw up zap_page_range() | |
644 | */ | |
645 | end += bt_entry_size_bytes(mm); | |
646 | ||
647 | /* | |
648 | * Find the first overlapping vma. If vma->vm_start > start, there | |
649 | * will be a hole in the bounds table. This -EINVAL return will | |
650 | * cause a SIGSEGV. | |
651 | */ | |
652 | vma = find_vma(mm, start); | |
653 | if (!vma || vma->vm_start > start) | |
654 | return -EINVAL; | |
655 | ||
656 | /* | |
657 | * A NUMA policy on a VM_MPX VMA could cause this bounds table to | |
658 | * be split. So we need to look across the entire 'start -> end' | |
659 | * range of this bounds table, find all of the VM_MPX VMAs, and | |
660 | * zap only those. | |
661 | */ | |
662 | addr = start; | |
663 | while (vma && vma->vm_start < end) { | |
664 | /* | |
665 | * We followed a bounds directory entry down | |
666 | * here. If we find a non-MPX VMA, that's bad, | |
667 | * so stop immediately and return an error. This | |
668 | * probably results in a SIGSEGV. | |
669 | */ | |
a8965276 | 670 | if (!(vma->vm_flags & VM_MPX)) |
3ceaccdf DH |
671 | return -EINVAL; |
672 | ||
673 | len = min(vma->vm_end, end) - addr; | |
ecf1385d | 674 | zap_page_range(vma, addr, len); |
3ceaccdf DH |
675 | trace_mpx_unmap_zap(addr, addr+len); |
676 | ||
677 | vma = vma->vm_next; | |
678 | addr = vma->vm_start; | |
679 | } | |
680 | return 0; | |
681 | } | |
682 | ||
613fcb7d DH |
683 | static unsigned long mpx_get_bd_entry_offset(struct mm_struct *mm, |
684 | unsigned long addr) | |
685 | { | |
686 | /* | |
687 | * There are several ways to derive the bd offsets. We | |
688 | * use the following approach here: | |
689 | * 1. We know the size of the virtual address space | |
690 | * 2. We know the number of entries in a bounds table | |
691 | * 3. We know that each entry covers a fixed amount of | |
692 | * virtual address space. | |
693 | * So, we can just divide the virtual address by the | |
694 | * virtual space used by one entry to determine which | |
695 | * entry "controls" the given virtual address. | |
696 | */ | |
697 | if (is_64bit_mm(mm)) { | |
698 | int bd_entry_size = 8; /* 64-bit pointer */ | |
699 | /* | |
700 | * Take the 64-bit addressing hole in to account. | |
701 | */ | |
702 | addr &= ((1UL << boot_cpu_data.x86_virt_bits) - 1); | |
703 | return (addr / bd_entry_virt_space(mm)) * bd_entry_size; | |
704 | } else { | |
705 | int bd_entry_size = 4; /* 32-bit pointer */ | |
706 | /* | |
707 | * 32-bit has no hole so this case needs no mask | |
708 | */ | |
709 | return (addr / bd_entry_virt_space(mm)) * bd_entry_size; | |
710 | } | |
711 | /* | |
712 | * The two return calls above are exact copies. If we | |
713 | * pull out a single copy and put it in here, gcc won't | |
714 | * realize that we're doing a power-of-2 divide and use | |
715 | * shifts. It uses a real divide. If we put them up | |
716 | * there, it manages to figure it out (gcc 4.8.3). | |
717 | */ | |
1de4fa14 DH |
718 | } |
719 | ||
3ceaccdf DH |
720 | static int unmap_entire_bt(struct mm_struct *mm, |
721 | long __user *bd_entry, unsigned long bt_addr) | |
1de4fa14 | 722 | { |
3ceaccdf DH |
723 | unsigned long expected_old_val = bt_addr | MPX_BD_ENTRY_VALID_FLAG; |
724 | unsigned long uninitialized_var(actual_old_val); | |
1de4fa14 DH |
725 | int ret; |
726 | ||
3ceaccdf DH |
727 | while (1) { |
728 | int need_write = 1; | |
729 | unsigned long cleared_bd_entry = 0; | |
730 | ||
731 | pagefault_disable(); | |
732 | ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val, | |
733 | bd_entry, expected_old_val, cleared_bd_entry); | |
734 | pagefault_enable(); | |
735 | if (!ret) | |
736 | break; | |
737 | if (ret == -EFAULT) | |
738 | ret = mpx_resolve_fault(bd_entry, need_write); | |
739 | /* | |
740 | * If we could not resolve the fault, consider it | |
741 | * userspace's fault and error out. | |
742 | */ | |
743 | if (ret) | |
744 | return ret; | |
745 | } | |
1de4fa14 | 746 | /* |
3ceaccdf | 747 | * The cmpxchg was performed, check the results. |
1de4fa14 | 748 | */ |
3ceaccdf DH |
749 | if (actual_old_val != expected_old_val) { |
750 | /* | |
751 | * Someone else raced with us to unmap the table. | |
752 | * That is OK, since we were both trying to do | |
753 | * the same thing. Declare success. | |
754 | */ | |
755 | if (!actual_old_val) | |
756 | return 0; | |
757 | /* | |
758 | * Something messed with the bounds directory | |
759 | * entry. We hold mmap_sem for read or write | |
760 | * here, so it could not be a _new_ bounds table | |
761 | * that someone just allocated. Something is | |
762 | * wrong, so pass up the error and SIGSEGV. | |
763 | */ | |
764 | return -EINVAL; | |
765 | } | |
766 | /* | |
767 | * Note, we are likely being called under do_munmap() already. To | |
768 | * avoid recursion, do_munmap() will check whether it comes | |
769 | * from one bounds table through VM_MPX flag. | |
770 | */ | |
897ab3e0 | 771 | return do_munmap(mm, bt_addr, mpx_bt_size_bytes(mm), NULL); |
1de4fa14 DH |
772 | } |
773 | ||
3ceaccdf DH |
774 | static int try_unmap_single_bt(struct mm_struct *mm, |
775 | unsigned long start, unsigned long end) | |
1de4fa14 | 776 | { |
3ceaccdf DH |
777 | struct vm_area_struct *next; |
778 | struct vm_area_struct *prev; | |
779 | /* | |
780 | * "bta" == Bounds Table Area: the area controlled by the | |
781 | * bounds table that we are unmapping. | |
782 | */ | |
783 | unsigned long bta_start_vaddr = start & ~(bd_entry_virt_space(mm)-1); | |
784 | unsigned long bta_end_vaddr = bta_start_vaddr + bd_entry_virt_space(mm); | |
785 | unsigned long uninitialized_var(bt_addr); | |
786 | void __user *bde_vaddr; | |
1de4fa14 | 787 | int ret; |
bea03c50 DH |
788 | /* |
789 | * We already unlinked the VMAs from the mm's rbtree so 'start' | |
790 | * is guaranteed to be in a hole. This gets us the first VMA | |
791 | * before the hole in to 'prev' and the next VMA after the hole | |
792 | * in to 'next'. | |
793 | */ | |
794 | next = find_vma_prev(mm, start, &prev); | |
795 | /* | |
796 | * Do not count other MPX bounds table VMAs as neighbors. | |
797 | * Although theoretically possible, we do not allow bounds | |
798 | * tables for bounds tables so our heads do not explode. | |
799 | * If we count them as neighbors here, we may end up with | |
800 | * lots of tables even though we have no actual table | |
801 | * entries in use. | |
802 | */ | |
a8965276 | 803 | while (next && (next->vm_flags & VM_MPX)) |
bea03c50 | 804 | next = next->vm_next; |
a8965276 | 805 | while (prev && (prev->vm_flags & VM_MPX)) |
bea03c50 | 806 | prev = prev->vm_prev; |
1de4fa14 | 807 | /* |
3ceaccdf DH |
808 | * We know 'start' and 'end' lie within an area controlled |
809 | * by a single bounds table. See if there are any other | |
810 | * VMAs controlled by that bounds table. If there are not | |
811 | * then we can "expand" the are we are unmapping to possibly | |
812 | * cover the entire table. | |
1de4fa14 DH |
813 | */ |
814 | next = find_vma_prev(mm, start, &prev); | |
3ceaccdf DH |
815 | if ((!prev || prev->vm_end <= bta_start_vaddr) && |
816 | (!next || next->vm_start >= bta_end_vaddr)) { | |
817 | /* | |
818 | * No neighbor VMAs controlled by same bounds | |
819 | * table. Try to unmap the whole thing | |
820 | */ | |
821 | start = bta_start_vaddr; | |
822 | end = bta_end_vaddr; | |
1de4fa14 DH |
823 | } |
824 | ||
cb02de96 | 825 | bde_vaddr = mm->context.bd_addr + mpx_get_bd_entry_offset(mm, start); |
3ceaccdf | 826 | ret = get_bt_addr(mm, bde_vaddr, &bt_addr); |
1de4fa14 | 827 | /* |
3ceaccdf | 828 | * No bounds table there, so nothing to unmap. |
1de4fa14 | 829 | */ |
3ceaccdf DH |
830 | if (ret == -ENOENT) { |
831 | ret = 0; | |
832 | return 0; | |
833 | } | |
1de4fa14 DH |
834 | if (ret) |
835 | return ret; | |
3ceaccdf DH |
836 | /* |
837 | * We are unmapping an entire table. Either because the | |
838 | * unmap that started this whole process was large enough | |
839 | * to cover an entire table, or that the unmap was small | |
840 | * but was the area covered by a bounds table. | |
841 | */ | |
842 | if ((start == bta_start_vaddr) && | |
843 | (end == bta_end_vaddr)) | |
844 | return unmap_entire_bt(mm, bde_vaddr, bt_addr); | |
845 | return zap_bt_entries_mapping(mm, bt_addr, start, end); | |
1de4fa14 DH |
846 | } |
847 | ||
848 | static int mpx_unmap_tables(struct mm_struct *mm, | |
849 | unsigned long start, unsigned long end) | |
850 | { | |
3ceaccdf | 851 | unsigned long one_unmap_start; |
2a1dcb1f | 852 | trace_mpx_unmap_search(start, end); |
1de4fa14 | 853 | |
3ceaccdf DH |
854 | one_unmap_start = start; |
855 | while (one_unmap_start < end) { | |
856 | int ret; | |
857 | unsigned long next_unmap_start = ALIGN(one_unmap_start+1, | |
858 | bd_entry_virt_space(mm)); | |
859 | unsigned long one_unmap_end = end; | |
860 | /* | |
861 | * if the end is beyond the current bounds table, | |
862 | * move it back so we only deal with a single one | |
863 | * at a time | |
864 | */ | |
865 | if (one_unmap_end > next_unmap_start) | |
866 | one_unmap_end = next_unmap_start; | |
867 | ret = try_unmap_single_bt(mm, one_unmap_start, one_unmap_end); | |
1de4fa14 DH |
868 | if (ret) |
869 | return ret; | |
1de4fa14 | 870 | |
3ceaccdf DH |
871 | one_unmap_start = next_unmap_start; |
872 | } | |
1de4fa14 DH |
873 | return 0; |
874 | } | |
875 | ||
876 | /* | |
877 | * Free unused bounds tables covered in a virtual address region being | |
878 | * munmap()ed. Assume end > start. | |
879 | * | |
880 | * This function will be called by do_munmap(), and the VMAs covering | |
881 | * the virtual address region start...end have already been split if | |
882 | * necessary, and the 'vma' is the first vma in this range (start -> end). | |
883 | */ | |
884 | void mpx_notify_unmap(struct mm_struct *mm, struct vm_area_struct *vma, | |
885 | unsigned long start, unsigned long end) | |
886 | { | |
887 | int ret; | |
888 | ||
889 | /* | |
890 | * Refuse to do anything unless userspace has asked | |
891 | * the kernel to help manage the bounds tables, | |
892 | */ | |
893 | if (!kernel_managing_mpx_tables(current->mm)) | |
894 | return; | |
895 | /* | |
896 | * This will look across the entire 'start -> end' range, | |
897 | * and find all of the non-VM_MPX VMAs. | |
898 | * | |
899 | * To avoid recursion, if a VM_MPX vma is found in the range | |
900 | * (start->end), we will not continue follow-up work. This | |
901 | * recursion represents having bounds tables for bounds tables, | |
902 | * which should not occur normally. Being strict about it here | |
903 | * helps ensure that we do not have an exploitable stack overflow. | |
904 | */ | |
905 | do { | |
906 | if (vma->vm_flags & VM_MPX) | |
907 | return; | |
908 | vma = vma->vm_next; | |
909 | } while (vma && vma->vm_start < end); | |
910 | ||
911 | ret = mpx_unmap_tables(mm, start, end); | |
912 | if (ret) | |
913 | force_sig(SIGSEGV, current); | |
914 | } | |
44b04912 KS |
915 | |
916 | /* MPX cannot handle addresses above 47 bits yet. */ | |
917 | unsigned long mpx_unmapped_area_check(unsigned long addr, unsigned long len, | |
918 | unsigned long flags) | |
919 | { | |
920 | if (!kernel_managing_mpx_tables(current->mm)) | |
921 | return addr; | |
922 | if (addr + len <= DEFAULT_MAP_WINDOW) | |
923 | return addr; | |
924 | if (flags & MAP_FIXED) | |
925 | return -ENOMEM; | |
926 | ||
927 | /* | |
928 | * Requested len is larger than the whole area we're allowed to map in. | |
929 | * Resetting hinting address wouldn't do much good -- fail early. | |
930 | */ | |
931 | if (len > DEFAULT_MAP_WINDOW) | |
932 | return -ENOMEM; | |
933 | ||
934 | /* Look for unmap area within DEFAULT_MAP_WINDOW */ | |
935 | return 0; | |
936 | } |