Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 1 | /* |
| 2 | * mpx.c - Memory Protection eXtensions |
| 3 | * |
| 4 | * Copyright (c) 2014, Intel Corporation. |
| 5 | * Qiaowei Ren <qiaowei.ren@intel.com> |
| 6 | * Dave Hansen <dave.hansen@intel.com> |
| 7 | */ |
| 8 | #include <linux/kernel.h> |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 9 | #include <linux/slab.h> |
Ingo Molnar | 589ee62 | 2017-02-04 00:16:44 +0100 | [diff] [blame] | 10 | #include <linux/mm_types.h> |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 11 | #include <linux/syscalls.h> |
| 12 | #include <linux/sched/sysctl.h> |
| 13 | |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 14 | #include <asm/insn.h> |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 15 | #include <asm/mman.h> |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 16 | #include <asm/mmu_context.h> |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 17 | #include <asm/mpx.h> |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 18 | #include <asm/processor.h> |
Ingo Molnar | 78f7f1e | 2015-04-24 02:54:44 +0200 | [diff] [blame] | 19 | #include <asm/fpu/internal.h> |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 20 | |
Dave Hansen | e7126cf | 2015-06-07 11:37:03 -0700 | [diff] [blame] | 21 | #define CREATE_TRACE_POINTS |
| 22 | #include <asm/trace/mpx.h> |
| 23 | |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 24 | static inline unsigned long mpx_bd_size_bytes(struct mm_struct *mm) |
| 25 | { |
| 26 | if (is_64bit_mm(mm)) |
| 27 | return MPX_BD_SIZE_BYTES_64; |
| 28 | else |
| 29 | return MPX_BD_SIZE_BYTES_32; |
| 30 | } |
| 31 | |
| 32 | static inline unsigned long mpx_bt_size_bytes(struct mm_struct *mm) |
| 33 | { |
| 34 | if (is_64bit_mm(mm)) |
| 35 | return MPX_BT_SIZE_BYTES_64; |
| 36 | else |
| 37 | return MPX_BT_SIZE_BYTES_32; |
| 38 | } |
| 39 | |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 40 | /* |
| 41 | * This is really a simplified "vm_mmap". it only handles MPX |
| 42 | * bounds tables (the bounds directory is user-allocated). |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 43 | */ |
| 44 | static unsigned long mpx_mmap(unsigned long len) |
| 45 | { |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 46 | struct mm_struct *mm = current->mm; |
Oleg Nesterov | 1fcfd8d | 2015-09-09 15:39:29 -0700 | [diff] [blame] | 47 | unsigned long addr, populate; |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 48 | |
Dave Hansen | eb099e5 | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 49 | /* Only bounds table can be allocated here */ |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 50 | if (len != mpx_bt_size_bytes(mm)) |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 51 | return -EINVAL; |
| 52 | |
| 53 | down_write(&mm->mmap_sem); |
Oleg Nesterov | 1fcfd8d | 2015-09-09 15:39:29 -0700 | [diff] [blame] | 54 | addr = do_mmap(NULL, 0, len, PROT_READ | PROT_WRITE, |
Mike Rapoport | 897ab3e | 2017-02-24 14:58:22 -0800 | [diff] [blame] | 55 | MAP_ANONYMOUS | MAP_PRIVATE, VM_MPX, 0, &populate, NULL); |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 56 | up_write(&mm->mmap_sem); |
Oleg Nesterov | 1fcfd8d | 2015-09-09 15:39:29 -0700 | [diff] [blame] | 57 | if (populate) |
| 58 | mm_populate(addr, populate); |
| 59 | |
| 60 | return addr; |
Qiaowei Ren | 57319d8 | 2014-11-14 07:18:27 -0800 | [diff] [blame] | 61 | } |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 62 | |
| 63 | enum reg_type { |
| 64 | REG_TYPE_RM = 0, |
| 65 | REG_TYPE_INDEX, |
| 66 | REG_TYPE_BASE, |
| 67 | }; |
| 68 | |
Dave Hansen | 68c009c | 2014-11-18 10:23:43 -0800 | [diff] [blame] | 69 | static int get_reg_offset(struct insn *insn, struct pt_regs *regs, |
| 70 | enum reg_type type) |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 71 | { |
| 72 | int regno = 0; |
| 73 | |
| 74 | static const int regoff[] = { |
| 75 | offsetof(struct pt_regs, ax), |
| 76 | offsetof(struct pt_regs, cx), |
| 77 | offsetof(struct pt_regs, dx), |
| 78 | offsetof(struct pt_regs, bx), |
| 79 | offsetof(struct pt_regs, sp), |
| 80 | offsetof(struct pt_regs, bp), |
| 81 | offsetof(struct pt_regs, si), |
| 82 | offsetof(struct pt_regs, di), |
| 83 | #ifdef CONFIG_X86_64 |
| 84 | offsetof(struct pt_regs, r8), |
| 85 | offsetof(struct pt_regs, r9), |
| 86 | offsetof(struct pt_regs, r10), |
| 87 | offsetof(struct pt_regs, r11), |
| 88 | offsetof(struct pt_regs, r12), |
| 89 | offsetof(struct pt_regs, r13), |
| 90 | offsetof(struct pt_regs, r14), |
| 91 | offsetof(struct pt_regs, r15), |
| 92 | #endif |
| 93 | }; |
| 94 | int nr_registers = ARRAY_SIZE(regoff); |
| 95 | /* |
| 96 | * Don't possibly decode a 32-bit instructions as |
| 97 | * reading a 64-bit-only register. |
| 98 | */ |
| 99 | if (IS_ENABLED(CONFIG_X86_64) && !insn->x86_64) |
| 100 | nr_registers -= 8; |
| 101 | |
| 102 | switch (type) { |
| 103 | case REG_TYPE_RM: |
| 104 | regno = X86_MODRM_RM(insn->modrm.value); |
Dave Hansen | 8e8efe0 | 2015-11-30 16:31:13 -0800 | [diff] [blame] | 105 | if (X86_REX_B(insn->rex_prefix.value)) |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 106 | regno += 8; |
| 107 | break; |
| 108 | |
| 109 | case REG_TYPE_INDEX: |
| 110 | regno = X86_SIB_INDEX(insn->sib.value); |
Dave Hansen | 8e8efe0 | 2015-11-30 16:31:13 -0800 | [diff] [blame] | 111 | if (X86_REX_X(insn->rex_prefix.value)) |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 112 | regno += 8; |
| 113 | break; |
| 114 | |
| 115 | case REG_TYPE_BASE: |
| 116 | regno = X86_SIB_BASE(insn->sib.value); |
Dave Hansen | 8e8efe0 | 2015-11-30 16:31:13 -0800 | [diff] [blame] | 117 | if (X86_REX_B(insn->rex_prefix.value)) |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 118 | regno += 8; |
| 119 | break; |
| 120 | |
| 121 | default: |
| 122 | pr_err("invalid register type"); |
| 123 | BUG(); |
| 124 | break; |
| 125 | } |
| 126 | |
Colin Ian King | 9bf148cb | 2016-02-26 18:55:31 +0000 | [diff] [blame] | 127 | if (regno >= nr_registers) { |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 128 | WARN_ONCE(1, "decoded an instruction with an invalid register"); |
| 129 | return -EINVAL; |
| 130 | } |
| 131 | return regoff[regno]; |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * return the address being referenced be instruction |
| 136 | * for rm=3 returning the content of the rm reg |
| 137 | * for rm!=3 calculates the address using SIB and Disp |
| 138 | */ |
| 139 | static void __user *mpx_get_addr_ref(struct insn *insn, struct pt_regs *regs) |
| 140 | { |
Dave Hansen | 68c009c | 2014-11-18 10:23:43 -0800 | [diff] [blame] | 141 | unsigned long addr, base, indx; |
| 142 | int addr_offset, base_offset, indx_offset; |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 143 | insn_byte_t sib; |
| 144 | |
| 145 | insn_get_modrm(insn); |
| 146 | insn_get_sib(insn); |
| 147 | sib = insn->sib.value; |
| 148 | |
| 149 | if (X86_MODRM_MOD(insn->modrm.value) == 3) { |
| 150 | addr_offset = get_reg_offset(insn, regs, REG_TYPE_RM); |
| 151 | if (addr_offset < 0) |
| 152 | goto out_err; |
| 153 | addr = regs_get_register(regs, addr_offset); |
| 154 | } else { |
| 155 | if (insn->sib.nbytes) { |
| 156 | base_offset = get_reg_offset(insn, regs, REG_TYPE_BASE); |
| 157 | if (base_offset < 0) |
| 158 | goto out_err; |
| 159 | |
| 160 | indx_offset = get_reg_offset(insn, regs, REG_TYPE_INDEX); |
| 161 | if (indx_offset < 0) |
| 162 | goto out_err; |
| 163 | |
| 164 | base = regs_get_register(regs, base_offset); |
| 165 | indx = regs_get_register(regs, indx_offset); |
| 166 | addr = base + indx * (1 << X86_SIB_SCALE(sib)); |
| 167 | } else { |
| 168 | addr_offset = get_reg_offset(insn, regs, REG_TYPE_RM); |
| 169 | if (addr_offset < 0) |
| 170 | goto out_err; |
| 171 | addr = regs_get_register(regs, addr_offset); |
| 172 | } |
| 173 | addr += insn->displacement.value; |
| 174 | } |
| 175 | return (void __user *)addr; |
| 176 | out_err: |
| 177 | return (void __user *)-1; |
| 178 | } |
| 179 | |
| 180 | static int mpx_insn_decode(struct insn *insn, |
| 181 | struct pt_regs *regs) |
| 182 | { |
| 183 | unsigned char buf[MAX_INSN_SIZE]; |
| 184 | int x86_64 = !test_thread_flag(TIF_IA32); |
| 185 | int not_copied; |
| 186 | int nr_copied; |
| 187 | |
| 188 | not_copied = copy_from_user(buf, (void __user *)regs->ip, sizeof(buf)); |
| 189 | nr_copied = sizeof(buf) - not_copied; |
| 190 | /* |
| 191 | * The decoder _should_ fail nicely if we pass it a short buffer. |
| 192 | * But, let's not depend on that implementation detail. If we |
| 193 | * did not get anything, just error out now. |
| 194 | */ |
| 195 | if (!nr_copied) |
| 196 | return -EFAULT; |
| 197 | insn_init(insn, buf, nr_copied, x86_64); |
| 198 | insn_get_length(insn); |
| 199 | /* |
| 200 | * copy_from_user() tries to get as many bytes as we could see in |
| 201 | * the largest possible instruction. If the instruction we are |
| 202 | * after is shorter than that _and_ we attempt to copy from |
| 203 | * something unreadable, we might get a short read. This is OK |
| 204 | * as long as the read did not stop in the middle of the |
| 205 | * instruction. Check to see if we got a partial instruction. |
| 206 | */ |
| 207 | if (nr_copied < insn->length) |
| 208 | return -EFAULT; |
| 209 | |
| 210 | insn_get_opcode(insn); |
| 211 | /* |
| 212 | * We only _really_ need to decode bndcl/bndcn/bndcu |
| 213 | * Error out on anything else. |
| 214 | */ |
| 215 | if (insn->opcode.bytes[0] != 0x0f) |
| 216 | goto bad_opcode; |
| 217 | if ((insn->opcode.bytes[1] != 0x1a) && |
| 218 | (insn->opcode.bytes[1] != 0x1b)) |
| 219 | goto bad_opcode; |
| 220 | |
| 221 | return 0; |
| 222 | bad_opcode: |
| 223 | return -EINVAL; |
| 224 | } |
| 225 | |
| 226 | /* |
| 227 | * If a bounds overflow occurs then a #BR is generated. This |
| 228 | * function decodes MPX instructions to get violation address |
| 229 | * and set this address into extended struct siginfo. |
| 230 | * |
| 231 | * Note that this is not a super precise way of doing this. |
| 232 | * Userspace could have, by the time we get here, written |
| 233 | * anything it wants in to the instructions. We can not |
| 234 | * trust anything about it. They might not be valid |
| 235 | * instructions or might encode invalid registers, etc... |
| 236 | * |
| 237 | * The caller is expected to kfree() the returned siginfo_t. |
| 238 | */ |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 239 | siginfo_t *mpx_generate_siginfo(struct pt_regs *regs) |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 240 | { |
Dave Hansen | 1126cb45 | 2015-09-02 16:31:29 -0700 | [diff] [blame] | 241 | const struct mpx_bndreg_state *bndregs; |
| 242 | const struct mpx_bndreg *bndreg; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 243 | siginfo_t *info = NULL; |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 244 | struct insn insn; |
| 245 | uint8_t bndregno; |
| 246 | int err; |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 247 | |
| 248 | err = mpx_insn_decode(&insn, regs); |
| 249 | if (err) |
| 250 | goto err_out; |
| 251 | |
| 252 | /* |
| 253 | * We know at this point that we are only dealing with |
| 254 | * MPX instructions. |
| 255 | */ |
| 256 | insn_get_modrm(&insn); |
| 257 | bndregno = X86_MODRM_REG(insn.modrm.value); |
| 258 | if (bndregno > 3) { |
| 259 | err = -EINVAL; |
| 260 | goto err_out; |
| 261 | } |
Dave Hansen | a84eeaa | 2015-06-07 11:37:01 -0700 | [diff] [blame] | 262 | /* get bndregs field from current task's xsave area */ |
Dave Hansen | d91cab7 | 2015-09-02 16:31:26 -0700 | [diff] [blame] | 263 | bndregs = get_xsave_field_ptr(XFEATURE_MASK_BNDREGS); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 264 | if (!bndregs) { |
| 265 | err = -EINVAL; |
| 266 | goto err_out; |
| 267 | } |
| 268 | /* now go select the individual register in the set of 4 */ |
Dave Hansen | 1126cb45 | 2015-09-02 16:31:29 -0700 | [diff] [blame] | 269 | bndreg = &bndregs->bndreg[bndregno]; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 270 | |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 271 | info = kzalloc(sizeof(*info), GFP_KERNEL); |
| 272 | if (!info) { |
| 273 | err = -ENOMEM; |
| 274 | goto err_out; |
| 275 | } |
| 276 | /* |
| 277 | * The registers are always 64-bit, but the upper 32 |
| 278 | * bits are ignored in 32-bit mode. Also, note that the |
| 279 | * upper bounds are architecturally represented in 1's |
| 280 | * complement form. |
| 281 | * |
| 282 | * The 'unsigned long' cast is because the compiler |
| 283 | * complains when casting from integers to different-size |
| 284 | * pointers. |
| 285 | */ |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 286 | info->si_lower = (void __user *)(unsigned long)bndreg->lower_bound; |
| 287 | info->si_upper = (void __user *)(unsigned long)~bndreg->upper_bound; |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 288 | info->si_addr_lsb = 0; |
| 289 | info->si_signo = SIGSEGV; |
| 290 | info->si_errno = 0; |
| 291 | info->si_code = SEGV_BNDERR; |
| 292 | info->si_addr = mpx_get_addr_ref(&insn, regs); |
| 293 | /* |
| 294 | * We were not able to extract an address from the instruction, |
| 295 | * probably because there was something invalid in it. |
| 296 | */ |
Tobias Klauser | 4538286 | 2017-01-12 16:53:11 +0100 | [diff] [blame] | 297 | if (info->si_addr == (void __user *)-1) { |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 298 | err = -EINVAL; |
| 299 | goto err_out; |
| 300 | } |
Dave Hansen | 97efebf | 2015-06-07 11:37:03 -0700 | [diff] [blame] | 301 | trace_mpx_bounds_register_exception(info->si_addr, bndreg); |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 302 | return info; |
| 303 | err_out: |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 304 | /* info might be NULL, but kfree() handles that */ |
| 305 | kfree(info); |
Dave Hansen | fcc7ffd | 2014-11-14 07:18:28 -0800 | [diff] [blame] | 306 | return ERR_PTR(err); |
| 307 | } |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 308 | |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 309 | static __user void *mpx_get_bounds_dir(void) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 310 | { |
Dave Hansen | 1126cb45 | 2015-09-02 16:31:29 -0700 | [diff] [blame] | 311 | const struct mpx_bndcsr *bndcsr; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 312 | |
| 313 | if (!cpu_feature_enabled(X86_FEATURE_MPX)) |
| 314 | return MPX_INVALID_BOUNDS_DIR; |
| 315 | |
| 316 | /* |
| 317 | * The bounds directory pointer is stored in a register |
| 318 | * only accessible if we first do an xsave. |
| 319 | */ |
Dave Hansen | d91cab7 | 2015-09-02 16:31:26 -0700 | [diff] [blame] | 320 | bndcsr = get_xsave_field_ptr(XFEATURE_MASK_BNDCSR); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 321 | if (!bndcsr) |
| 322 | return MPX_INVALID_BOUNDS_DIR; |
| 323 | |
| 324 | /* |
| 325 | * Make sure the register looks valid by checking the |
| 326 | * enable bit. |
| 327 | */ |
| 328 | if (!(bndcsr->bndcfgu & MPX_BNDCFG_ENABLE_FLAG)) |
| 329 | return MPX_INVALID_BOUNDS_DIR; |
| 330 | |
| 331 | /* |
| 332 | * Lastly, mask off the low bits used for configuration |
| 333 | * flags, and return the address of the bounds table. |
| 334 | */ |
| 335 | return (void __user *)(unsigned long) |
| 336 | (bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK); |
| 337 | } |
| 338 | |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 339 | int mpx_enable_management(void) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 340 | { |
| 341 | void __user *bd_base = MPX_INVALID_BOUNDS_DIR; |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 342 | struct mm_struct *mm = current->mm; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 343 | int ret = 0; |
| 344 | |
| 345 | /* |
| 346 | * runtime in the userspace will be responsible for allocation of |
| 347 | * the bounds directory. Then, it will save the base of the bounds |
| 348 | * directory into XSAVE/XRSTOR Save Area and enable MPX through |
| 349 | * XRSTOR instruction. |
| 350 | * |
Dave Hansen | a84eeaa | 2015-06-07 11:37:01 -0700 | [diff] [blame] | 351 | * The copy_xregs_to_kernel() beneath get_xsave_field_ptr() is |
| 352 | * expected to be relatively expensive. Storing the bounds |
| 353 | * directory here means that we do not have to do xsave in the |
Mark Rutland | cb02de9 | 2016-12-16 12:40:55 +0000 | [diff] [blame] | 354 | * unmap path; we can just use mm->context.bd_addr instead. |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 355 | */ |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 356 | bd_base = mpx_get_bounds_dir(); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 357 | down_write(&mm->mmap_sem); |
Mark Rutland | cb02de9 | 2016-12-16 12:40:55 +0000 | [diff] [blame] | 358 | mm->context.bd_addr = bd_base; |
| 359 | if (mm->context.bd_addr == MPX_INVALID_BOUNDS_DIR) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 360 | ret = -ENXIO; |
| 361 | |
| 362 | up_write(&mm->mmap_sem); |
| 363 | return ret; |
| 364 | } |
| 365 | |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 366 | int mpx_disable_management(void) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 367 | { |
| 368 | struct mm_struct *mm = current->mm; |
| 369 | |
| 370 | if (!cpu_feature_enabled(X86_FEATURE_MPX)) |
| 371 | return -ENXIO; |
| 372 | |
| 373 | down_write(&mm->mmap_sem); |
Mark Rutland | cb02de9 | 2016-12-16 12:40:55 +0000 | [diff] [blame] | 374 | mm->context.bd_addr = MPX_INVALID_BOUNDS_DIR; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 375 | up_write(&mm->mmap_sem); |
| 376 | return 0; |
| 377 | } |
| 378 | |
Dave Hansen | 6ac52bb | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 379 | static int mpx_cmpxchg_bd_entry(struct mm_struct *mm, |
| 380 | unsigned long *curval, |
| 381 | unsigned long __user *addr, |
| 382 | unsigned long old_val, unsigned long new_val) |
| 383 | { |
| 384 | int ret; |
| 385 | /* |
| 386 | * user_atomic_cmpxchg_inatomic() actually uses sizeof() |
| 387 | * the pointer that we pass to it to figure out how much |
| 388 | * data to cmpxchg. We have to be careful here not to |
| 389 | * pass a pointer to a 64-bit data type when we only want |
| 390 | * a 32-bit copy. |
| 391 | */ |
| 392 | if (is_64bit_mm(mm)) { |
| 393 | ret = user_atomic_cmpxchg_inatomic(curval, |
| 394 | addr, old_val, new_val); |
| 395 | } else { |
| 396 | u32 uninitialized_var(curval_32); |
| 397 | u32 old_val_32 = old_val; |
| 398 | u32 new_val_32 = new_val; |
| 399 | u32 __user *addr_32 = (u32 __user *)addr; |
| 400 | |
| 401 | ret = user_atomic_cmpxchg_inatomic(&curval_32, |
| 402 | addr_32, old_val_32, new_val_32); |
| 403 | *curval = curval_32; |
| 404 | } |
| 405 | return ret; |
| 406 | } |
| 407 | |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 408 | /* |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 409 | * With 32-bit mode, a bounds directory is 4MB, and the size of each |
| 410 | * bounds table is 16KB. With 64-bit mode, a bounds directory is 2GB, |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 411 | * and the size of each bounds table is 4MB. |
| 412 | */ |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 413 | static int allocate_bt(struct mm_struct *mm, long __user *bd_entry) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 414 | { |
| 415 | unsigned long expected_old_val = 0; |
| 416 | unsigned long actual_old_val = 0; |
| 417 | unsigned long bt_addr; |
Dave Hansen | a1149fc | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 418 | unsigned long bd_new_entry; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 419 | int ret = 0; |
| 420 | |
| 421 | /* |
| 422 | * Carve the virtual space out of userspace for the new |
| 423 | * bounds table: |
| 424 | */ |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 425 | bt_addr = mpx_mmap(mpx_bt_size_bytes(mm)); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 426 | if (IS_ERR((void *)bt_addr)) |
| 427 | return PTR_ERR((void *)bt_addr); |
| 428 | /* |
| 429 | * Set the valid flag (kinda like _PAGE_PRESENT in a pte) |
| 430 | */ |
Dave Hansen | a1149fc | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 431 | bd_new_entry = bt_addr | MPX_BD_ENTRY_VALID_FLAG; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 432 | |
| 433 | /* |
| 434 | * Go poke the address of the new bounds table in to the |
| 435 | * bounds directory entry out in userspace memory. Note: |
| 436 | * we may race with another CPU instantiating the same table. |
| 437 | * In that case the cmpxchg will see an unexpected |
| 438 | * 'actual_old_val'. |
| 439 | * |
| 440 | * This can fault, but that's OK because we do not hold |
| 441 | * mmap_sem at this point, unlike some of the other part |
| 442 | * of the MPX code that have to pagefault_disable(). |
| 443 | */ |
Dave Hansen | 6ac52bb | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 444 | ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val, bd_entry, |
| 445 | expected_old_val, bd_new_entry); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 446 | if (ret) |
| 447 | goto out_unmap; |
| 448 | |
| 449 | /* |
| 450 | * The user_atomic_cmpxchg_inatomic() will only return nonzero |
| 451 | * for faults, *not* if the cmpxchg itself fails. Now we must |
| 452 | * verify that the cmpxchg itself completed successfully. |
| 453 | */ |
| 454 | /* |
| 455 | * We expected an empty 'expected_old_val', but instead found |
| 456 | * an apparently valid entry. Assume we raced with another |
| 457 | * thread to instantiate this table and desclare succecss. |
| 458 | */ |
| 459 | if (actual_old_val & MPX_BD_ENTRY_VALID_FLAG) { |
| 460 | ret = 0; |
| 461 | goto out_unmap; |
| 462 | } |
| 463 | /* |
| 464 | * We found a non-empty bd_entry but it did not have the |
| 465 | * VALID_FLAG set. Return an error which will result in |
| 466 | * a SEGV since this probably means that somebody scribbled |
| 467 | * some invalid data in to a bounds table. |
| 468 | */ |
| 469 | if (expected_old_val != actual_old_val) { |
| 470 | ret = -EINVAL; |
| 471 | goto out_unmap; |
| 472 | } |
Dave Hansen | cd4996d | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 473 | trace_mpx_new_bounds_table(bt_addr); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 474 | return 0; |
| 475 | out_unmap: |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 476 | vm_munmap(bt_addr, mpx_bt_size_bytes(mm)); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 477 | return ret; |
| 478 | } |
| 479 | |
| 480 | /* |
| 481 | * When a BNDSTX instruction attempts to save bounds to a bounds |
| 482 | * table, it will first attempt to look up the table in the |
| 483 | * first-level bounds directory. If it does not find a table in |
| 484 | * the directory, a #BR is generated and we get here in order to |
| 485 | * allocate a new table. |
| 486 | * |
| 487 | * With 32-bit mode, the size of BD is 4MB, and the size of each |
| 488 | * bound table is 16KB. With 64-bit mode, the size of BD is 2GB, |
| 489 | * and the size of each bound table is 4MB. |
| 490 | */ |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 491 | static int do_mpx_bt_fault(void) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 492 | { |
| 493 | unsigned long bd_entry, bd_base; |
Dave Hansen | 1126cb45 | 2015-09-02 16:31:29 -0700 | [diff] [blame] | 494 | const struct mpx_bndcsr *bndcsr; |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 495 | struct mm_struct *mm = current->mm; |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 496 | |
Dave Hansen | d91cab7 | 2015-09-02 16:31:26 -0700 | [diff] [blame] | 497 | bndcsr = get_xsave_field_ptr(XFEATURE_MASK_BNDCSR); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 498 | if (!bndcsr) |
| 499 | return -EINVAL; |
| 500 | /* |
| 501 | * Mask off the preserve and enable bits |
| 502 | */ |
| 503 | bd_base = bndcsr->bndcfgu & MPX_BNDCFG_ADDR_MASK; |
| 504 | /* |
| 505 | * The hardware provides the address of the missing or invalid |
| 506 | * entry via BNDSTATUS, so we don't have to go look it up. |
| 507 | */ |
| 508 | bd_entry = bndcsr->bndstatus & MPX_BNDSTA_ADDR_MASK; |
| 509 | /* |
| 510 | * Make sure the directory entry is within where we think |
| 511 | * the directory is. |
| 512 | */ |
| 513 | if ((bd_entry < bd_base) || |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 514 | (bd_entry >= bd_base + mpx_bd_size_bytes(mm))) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 515 | return -EINVAL; |
| 516 | |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 517 | return allocate_bt(mm, (long __user *)bd_entry); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 518 | } |
| 519 | |
Dave Hansen | 46a6e0c | 2015-06-07 11:37:02 -0700 | [diff] [blame] | 520 | int mpx_handle_bd_fault(void) |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 521 | { |
| 522 | /* |
| 523 | * Userspace never asked us to manage the bounds tables, |
| 524 | * so refuse to help. |
| 525 | */ |
| 526 | if (!kernel_managing_mpx_tables(current->mm)) |
| 527 | return -EINVAL; |
| 528 | |
Joerg Roedel | 5ed386e | 2017-04-06 16:19:22 +0200 | [diff] [blame] | 529 | return do_mpx_bt_fault(); |
Dave Hansen | fe3d197 | 2014-11-14 07:18:29 -0800 | [diff] [blame] | 530 | } |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 531 | |
| 532 | /* |
| 533 | * A thin wrapper around get_user_pages(). Returns 0 if the |
| 534 | * fault was resolved or -errno if not. |
| 535 | */ |
| 536 | static int mpx_resolve_fault(long __user *addr, int write) |
| 537 | { |
| 538 | long gup_ret; |
| 539 | int nr_pages = 1; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 540 | |
Lorenzo Stoakes | 768ae30 | 2016-10-13 01:20:16 +0100 | [diff] [blame] | 541 | gup_ret = get_user_pages((unsigned long)addr, nr_pages, |
| 542 | write ? FOLL_WRITE : 0, NULL, NULL); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 543 | /* |
| 544 | * get_user_pages() returns number of pages gotten. |
| 545 | * 0 means we failed to fault in and get anything, |
| 546 | * probably because 'addr' is bad. |
| 547 | */ |
| 548 | if (!gup_ret) |
| 549 | return -EFAULT; |
| 550 | /* Other error, return it */ |
| 551 | if (gup_ret < 0) |
| 552 | return gup_ret; |
| 553 | /* must have gup'd a page and gup_ret>0, success */ |
| 554 | return 0; |
| 555 | } |
| 556 | |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 557 | static unsigned long mpx_bd_entry_to_bt_addr(struct mm_struct *mm, |
| 558 | unsigned long bd_entry) |
| 559 | { |
| 560 | unsigned long bt_addr = bd_entry; |
| 561 | int align_to_bytes; |
| 562 | /* |
| 563 | * Bit 0 in a bt_entry is always the valid bit. |
| 564 | */ |
| 565 | bt_addr &= ~MPX_BD_ENTRY_VALID_FLAG; |
| 566 | /* |
| 567 | * Tables are naturally aligned at 8-byte boundaries |
| 568 | * on 64-bit and 4-byte boundaries on 32-bit. The |
| 569 | * documentation makes it appear that the low bits |
| 570 | * are ignored by the hardware, so we do the same. |
| 571 | */ |
| 572 | if (is_64bit_mm(mm)) |
| 573 | align_to_bytes = 8; |
| 574 | else |
| 575 | align_to_bytes = 4; |
| 576 | bt_addr &= ~(align_to_bytes-1); |
| 577 | return bt_addr; |
| 578 | } |
| 579 | |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 580 | /* |
Dave Hansen | 46561c3 | 2015-11-11 10:19:31 -0800 | [diff] [blame] | 581 | * We only want to do a 4-byte get_user() on 32-bit. Otherwise, |
| 582 | * we might run off the end of the bounds table if we are on |
| 583 | * a 64-bit kernel and try to get 8 bytes. |
| 584 | */ |
Tobias Klauser | 6bce725 | 2017-03-08 14:30:34 +0100 | [diff] [blame] | 585 | static int get_user_bd_entry(struct mm_struct *mm, unsigned long *bd_entry_ret, |
Dave Hansen | 46561c3 | 2015-11-11 10:19:31 -0800 | [diff] [blame] | 586 | long __user *bd_entry_ptr) |
| 587 | { |
| 588 | u32 bd_entry_32; |
| 589 | int ret; |
| 590 | |
| 591 | if (is_64bit_mm(mm)) |
| 592 | return get_user(*bd_entry_ret, bd_entry_ptr); |
| 593 | |
| 594 | /* |
| 595 | * Note that get_user() uses the type of the *pointer* to |
| 596 | * establish the size of the get, not the destination. |
| 597 | */ |
| 598 | ret = get_user(bd_entry_32, (u32 __user *)bd_entry_ptr); |
| 599 | *bd_entry_ret = bd_entry_32; |
| 600 | return ret; |
| 601 | } |
| 602 | |
| 603 | /* |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 604 | * Get the base of bounds tables pointed by specific bounds |
| 605 | * directory entry. |
| 606 | */ |
| 607 | static int get_bt_addr(struct mm_struct *mm, |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 608 | long __user *bd_entry_ptr, |
| 609 | unsigned long *bt_addr_result) |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 610 | { |
| 611 | int ret; |
| 612 | int valid_bit; |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 613 | unsigned long bd_entry; |
| 614 | unsigned long bt_addr; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 615 | |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 616 | if (!access_ok(VERIFY_READ, (bd_entry_ptr), sizeof(*bd_entry_ptr))) |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 617 | return -EFAULT; |
| 618 | |
| 619 | while (1) { |
| 620 | int need_write = 0; |
| 621 | |
| 622 | pagefault_disable(); |
Dave Hansen | 46561c3 | 2015-11-11 10:19:31 -0800 | [diff] [blame] | 623 | ret = get_user_bd_entry(mm, &bd_entry, bd_entry_ptr); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 624 | pagefault_enable(); |
| 625 | if (!ret) |
| 626 | break; |
| 627 | if (ret == -EFAULT) |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 628 | ret = mpx_resolve_fault(bd_entry_ptr, need_write); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 629 | /* |
| 630 | * If we could not resolve the fault, consider it |
| 631 | * userspace's fault and error out. |
| 632 | */ |
| 633 | if (ret) |
| 634 | return ret; |
| 635 | } |
| 636 | |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 637 | valid_bit = bd_entry & MPX_BD_ENTRY_VALID_FLAG; |
| 638 | bt_addr = mpx_bd_entry_to_bt_addr(mm, bd_entry); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 639 | |
| 640 | /* |
| 641 | * When the kernel is managing bounds tables, a bounds directory |
| 642 | * entry will either have a valid address (plus the valid bit) |
| 643 | * *OR* be completely empty. If we see a !valid entry *and* some |
| 644 | * data in the address field, we know something is wrong. This |
| 645 | * -EINVAL return will cause a SIGSEGV. |
| 646 | */ |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 647 | if (!valid_bit && bt_addr) |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 648 | return -EINVAL; |
| 649 | /* |
| 650 | * Do we have an completely zeroed bt entry? That is OK. It |
| 651 | * just means there was no bounds table for this memory. Make |
| 652 | * sure to distinguish this from -EINVAL, which will cause |
| 653 | * a SEGV. |
| 654 | */ |
| 655 | if (!valid_bit) |
| 656 | return -ENOENT; |
| 657 | |
Dave Hansen | 5458765 | 2015-06-07 11:37:04 -0700 | [diff] [blame] | 658 | *bt_addr_result = bt_addr; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 659 | return 0; |
| 660 | } |
| 661 | |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 662 | static inline int bt_entry_size_bytes(struct mm_struct *mm) |
| 663 | { |
| 664 | if (is_64bit_mm(mm)) |
| 665 | return MPX_BT_ENTRY_BYTES_64; |
| 666 | else |
| 667 | return MPX_BT_ENTRY_BYTES_32; |
| 668 | } |
| 669 | |
| 670 | /* |
| 671 | * Take a virtual address and turns it in to the offset in bytes |
| 672 | * inside of the bounds table where the bounds table entry |
| 673 | * controlling 'addr' can be found. |
| 674 | */ |
| 675 | static unsigned long mpx_get_bt_entry_offset_bytes(struct mm_struct *mm, |
| 676 | unsigned long addr) |
| 677 | { |
| 678 | unsigned long bt_table_nr_entries; |
| 679 | unsigned long offset = addr; |
| 680 | |
| 681 | if (is_64bit_mm(mm)) { |
| 682 | /* Bottom 3 bits are ignored on 64-bit */ |
| 683 | offset >>= 3; |
| 684 | bt_table_nr_entries = MPX_BT_NR_ENTRIES_64; |
| 685 | } else { |
| 686 | /* Bottom 2 bits are ignored on 32-bit */ |
| 687 | offset >>= 2; |
| 688 | bt_table_nr_entries = MPX_BT_NR_ENTRIES_32; |
| 689 | } |
| 690 | /* |
| 691 | * We know the size of the table in to which we are |
| 692 | * indexing, and we have eliminated all the low bits |
| 693 | * which are ignored for indexing. |
| 694 | * |
| 695 | * Mask out all the high bits which we do not need |
| 696 | * to index in to the table. Note that the tables |
| 697 | * are always powers of two so this gives us a proper |
| 698 | * mask. |
| 699 | */ |
| 700 | offset &= (bt_table_nr_entries-1); |
| 701 | /* |
| 702 | * We now have an entry offset in terms of *entries* in |
| 703 | * the table. We need to scale it back up to bytes. |
| 704 | */ |
| 705 | offset *= bt_entry_size_bytes(mm); |
| 706 | return offset; |
| 707 | } |
| 708 | |
| 709 | /* |
| 710 | * How much virtual address space does a single bounds |
| 711 | * directory entry cover? |
| 712 | * |
| 713 | * Note, we need a long long because 4GB doesn't fit in |
| 714 | * to a long on 32-bit. |
| 715 | */ |
| 716 | static inline unsigned long bd_entry_virt_space(struct mm_struct *mm) |
| 717 | { |
Dave Hansen | f3119b8 | 2015-11-11 10:19:34 -0800 | [diff] [blame] | 718 | unsigned long long virt_space; |
| 719 | unsigned long long GB = (1ULL << 30); |
| 720 | |
| 721 | /* |
| 722 | * This covers 32-bit emulation as well as 32-bit kernels |
Adam Buchbinder | 6a6256f | 2016-02-23 15:34:30 -0800 | [diff] [blame] | 723 | * running on 64-bit hardware. |
Dave Hansen | f3119b8 | 2015-11-11 10:19:34 -0800 | [diff] [blame] | 724 | */ |
| 725 | if (!is_64bit_mm(mm)) |
| 726 | return (4ULL * GB) / MPX_BD_NR_ENTRIES_32; |
| 727 | |
| 728 | /* |
| 729 | * 'x86_virt_bits' returns what the hardware is capable |
Adam Buchbinder | 6a6256f | 2016-02-23 15:34:30 -0800 | [diff] [blame] | 730 | * of, and returns the full >32-bit address space when |
Dave Hansen | f3119b8 | 2015-11-11 10:19:34 -0800 | [diff] [blame] | 731 | * running 32-bit kernels on 64-bit hardware. |
| 732 | */ |
| 733 | virt_space = (1ULL << boot_cpu_data.x86_virt_bits); |
| 734 | return virt_space / MPX_BD_NR_ENTRIES_64; |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 735 | } |
| 736 | |
| 737 | /* |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 738 | * Free the backing physical pages of bounds table 'bt_addr'. |
| 739 | * Assume start...end is within that bounds table. |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 740 | */ |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 741 | static noinline int zap_bt_entries_mapping(struct mm_struct *mm, |
| 742 | unsigned long bt_addr, |
| 743 | unsigned long start_mapping, unsigned long end_mapping) |
| 744 | { |
| 745 | struct vm_area_struct *vma; |
| 746 | unsigned long addr, len; |
| 747 | unsigned long start; |
| 748 | unsigned long end; |
| 749 | |
| 750 | /* |
| 751 | * if we 'end' on a boundary, the offset will be 0 which |
| 752 | * is not what we want. Back it up a byte to get the |
| 753 | * last bt entry. Then once we have the entry itself, |
| 754 | * move 'end' back up by the table entry size. |
| 755 | */ |
| 756 | start = bt_addr + mpx_get_bt_entry_offset_bytes(mm, start_mapping); |
| 757 | end = bt_addr + mpx_get_bt_entry_offset_bytes(mm, end_mapping - 1); |
| 758 | /* |
| 759 | * Move end back up by one entry. Among other things |
| 760 | * this ensures that it remains page-aligned and does |
| 761 | * not screw up zap_page_range() |
| 762 | */ |
| 763 | end += bt_entry_size_bytes(mm); |
| 764 | |
| 765 | /* |
| 766 | * Find the first overlapping vma. If vma->vm_start > start, there |
| 767 | * will be a hole in the bounds table. This -EINVAL return will |
| 768 | * cause a SIGSEGV. |
| 769 | */ |
| 770 | vma = find_vma(mm, start); |
| 771 | if (!vma || vma->vm_start > start) |
| 772 | return -EINVAL; |
| 773 | |
| 774 | /* |
| 775 | * A NUMA policy on a VM_MPX VMA could cause this bounds table to |
| 776 | * be split. So we need to look across the entire 'start -> end' |
| 777 | * range of this bounds table, find all of the VM_MPX VMAs, and |
| 778 | * zap only those. |
| 779 | */ |
| 780 | addr = start; |
| 781 | while (vma && vma->vm_start < end) { |
| 782 | /* |
| 783 | * We followed a bounds directory entry down |
| 784 | * here. If we find a non-MPX VMA, that's bad, |
| 785 | * so stop immediately and return an error. This |
| 786 | * probably results in a SIGSEGV. |
| 787 | */ |
Kirill A. Shutemov | a8965276 | 2015-07-20 14:29:58 -0700 | [diff] [blame] | 788 | if (!(vma->vm_flags & VM_MPX)) |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 789 | return -EINVAL; |
| 790 | |
| 791 | len = min(vma->vm_end, end) - addr; |
Kirill A. Shutemov | ecf1385 | 2017-02-22 15:46:37 -0800 | [diff] [blame] | 792 | zap_page_range(vma, addr, len); |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 793 | trace_mpx_unmap_zap(addr, addr+len); |
| 794 | |
| 795 | vma = vma->vm_next; |
| 796 | addr = vma->vm_start; |
| 797 | } |
| 798 | return 0; |
| 799 | } |
| 800 | |
Dave Hansen | 613fcb7 | 2015-06-07 11:37:05 -0700 | [diff] [blame] | 801 | static unsigned long mpx_get_bd_entry_offset(struct mm_struct *mm, |
| 802 | unsigned long addr) |
| 803 | { |
| 804 | /* |
| 805 | * There are several ways to derive the bd offsets. We |
| 806 | * use the following approach here: |
| 807 | * 1. We know the size of the virtual address space |
| 808 | * 2. We know the number of entries in a bounds table |
| 809 | * 3. We know that each entry covers a fixed amount of |
| 810 | * virtual address space. |
| 811 | * So, we can just divide the virtual address by the |
| 812 | * virtual space used by one entry to determine which |
| 813 | * entry "controls" the given virtual address. |
| 814 | */ |
| 815 | if (is_64bit_mm(mm)) { |
| 816 | int bd_entry_size = 8; /* 64-bit pointer */ |
| 817 | /* |
| 818 | * Take the 64-bit addressing hole in to account. |
| 819 | */ |
| 820 | addr &= ((1UL << boot_cpu_data.x86_virt_bits) - 1); |
| 821 | return (addr / bd_entry_virt_space(mm)) * bd_entry_size; |
| 822 | } else { |
| 823 | int bd_entry_size = 4; /* 32-bit pointer */ |
| 824 | /* |
| 825 | * 32-bit has no hole so this case needs no mask |
| 826 | */ |
| 827 | return (addr / bd_entry_virt_space(mm)) * bd_entry_size; |
| 828 | } |
| 829 | /* |
| 830 | * The two return calls above are exact copies. If we |
| 831 | * pull out a single copy and put it in here, gcc won't |
| 832 | * realize that we're doing a power-of-2 divide and use |
| 833 | * shifts. It uses a real divide. If we put them up |
| 834 | * there, it manages to figure it out (gcc 4.8.3). |
| 835 | */ |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 836 | } |
| 837 | |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 838 | static int unmap_entire_bt(struct mm_struct *mm, |
| 839 | long __user *bd_entry, unsigned long bt_addr) |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 840 | { |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 841 | unsigned long expected_old_val = bt_addr | MPX_BD_ENTRY_VALID_FLAG; |
| 842 | unsigned long uninitialized_var(actual_old_val); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 843 | int ret; |
| 844 | |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 845 | while (1) { |
| 846 | int need_write = 1; |
| 847 | unsigned long cleared_bd_entry = 0; |
| 848 | |
| 849 | pagefault_disable(); |
| 850 | ret = mpx_cmpxchg_bd_entry(mm, &actual_old_val, |
| 851 | bd_entry, expected_old_val, cleared_bd_entry); |
| 852 | pagefault_enable(); |
| 853 | if (!ret) |
| 854 | break; |
| 855 | if (ret == -EFAULT) |
| 856 | ret = mpx_resolve_fault(bd_entry, need_write); |
| 857 | /* |
| 858 | * If we could not resolve the fault, consider it |
| 859 | * userspace's fault and error out. |
| 860 | */ |
| 861 | if (ret) |
| 862 | return ret; |
| 863 | } |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 864 | /* |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 865 | * The cmpxchg was performed, check the results. |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 866 | */ |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 867 | if (actual_old_val != expected_old_val) { |
| 868 | /* |
| 869 | * Someone else raced with us to unmap the table. |
| 870 | * That is OK, since we were both trying to do |
| 871 | * the same thing. Declare success. |
| 872 | */ |
| 873 | if (!actual_old_val) |
| 874 | return 0; |
| 875 | /* |
| 876 | * Something messed with the bounds directory |
| 877 | * entry. We hold mmap_sem for read or write |
| 878 | * here, so it could not be a _new_ bounds table |
| 879 | * that someone just allocated. Something is |
| 880 | * wrong, so pass up the error and SIGSEGV. |
| 881 | */ |
| 882 | return -EINVAL; |
| 883 | } |
| 884 | /* |
| 885 | * Note, we are likely being called under do_munmap() already. To |
| 886 | * avoid recursion, do_munmap() will check whether it comes |
| 887 | * from one bounds table through VM_MPX flag. |
| 888 | */ |
Mike Rapoport | 897ab3e | 2017-02-24 14:58:22 -0800 | [diff] [blame] | 889 | return do_munmap(mm, bt_addr, mpx_bt_size_bytes(mm), NULL); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 890 | } |
| 891 | |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 892 | static int try_unmap_single_bt(struct mm_struct *mm, |
| 893 | unsigned long start, unsigned long end) |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 894 | { |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 895 | struct vm_area_struct *next; |
| 896 | struct vm_area_struct *prev; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 897 | /* |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 898 | * "bta" == Bounds Table Area: the area controlled by the |
| 899 | * bounds table that we are unmapping. |
| 900 | */ |
| 901 | unsigned long bta_start_vaddr = start & ~(bd_entry_virt_space(mm)-1); |
| 902 | unsigned long bta_end_vaddr = bta_start_vaddr + bd_entry_virt_space(mm); |
| 903 | unsigned long uninitialized_var(bt_addr); |
| 904 | void __user *bde_vaddr; |
| 905 | int ret; |
| 906 | /* |
Dave Hansen | bea03c5 | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 907 | * We already unlinked the VMAs from the mm's rbtree so 'start' |
| 908 | * is guaranteed to be in a hole. This gets us the first VMA |
| 909 | * before the hole in to 'prev' and the next VMA after the hole |
| 910 | * in to 'next'. |
| 911 | */ |
| 912 | next = find_vma_prev(mm, start, &prev); |
| 913 | /* |
| 914 | * Do not count other MPX bounds table VMAs as neighbors. |
| 915 | * Although theoretically possible, we do not allow bounds |
| 916 | * tables for bounds tables so our heads do not explode. |
| 917 | * If we count them as neighbors here, we may end up with |
| 918 | * lots of tables even though we have no actual table |
| 919 | * entries in use. |
| 920 | */ |
Kirill A. Shutemov | a8965276 | 2015-07-20 14:29:58 -0700 | [diff] [blame] | 921 | while (next && (next->vm_flags & VM_MPX)) |
Dave Hansen | bea03c5 | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 922 | next = next->vm_next; |
Kirill A. Shutemov | a8965276 | 2015-07-20 14:29:58 -0700 | [diff] [blame] | 923 | while (prev && (prev->vm_flags & VM_MPX)) |
Dave Hansen | bea03c5 | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 924 | prev = prev->vm_prev; |
| 925 | /* |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 926 | * We know 'start' and 'end' lie within an area controlled |
| 927 | * by a single bounds table. See if there are any other |
| 928 | * VMAs controlled by that bounds table. If there are not |
| 929 | * then we can "expand" the are we are unmapping to possibly |
| 930 | * cover the entire table. |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 931 | */ |
| 932 | next = find_vma_prev(mm, start, &prev); |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 933 | if ((!prev || prev->vm_end <= bta_start_vaddr) && |
| 934 | (!next || next->vm_start >= bta_end_vaddr)) { |
| 935 | /* |
| 936 | * No neighbor VMAs controlled by same bounds |
| 937 | * table. Try to unmap the whole thing |
| 938 | */ |
| 939 | start = bta_start_vaddr; |
| 940 | end = bta_end_vaddr; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 941 | } |
| 942 | |
Mark Rutland | cb02de9 | 2016-12-16 12:40:55 +0000 | [diff] [blame] | 943 | bde_vaddr = mm->context.bd_addr + mpx_get_bd_entry_offset(mm, start); |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 944 | ret = get_bt_addr(mm, bde_vaddr, &bt_addr); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 945 | /* |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 946 | * No bounds table there, so nothing to unmap. |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 947 | */ |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 948 | if (ret == -ENOENT) { |
| 949 | ret = 0; |
| 950 | return 0; |
| 951 | } |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 952 | if (ret) |
| 953 | return ret; |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 954 | /* |
| 955 | * We are unmapping an entire table. Either because the |
| 956 | * unmap that started this whole process was large enough |
| 957 | * to cover an entire table, or that the unmap was small |
| 958 | * but was the area covered by a bounds table. |
| 959 | */ |
| 960 | if ((start == bta_start_vaddr) && |
| 961 | (end == bta_end_vaddr)) |
| 962 | return unmap_entire_bt(mm, bde_vaddr, bt_addr); |
| 963 | return zap_bt_entries_mapping(mm, bt_addr, start, end); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 964 | } |
| 965 | |
| 966 | static int mpx_unmap_tables(struct mm_struct *mm, |
| 967 | unsigned long start, unsigned long end) |
| 968 | { |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 969 | unsigned long one_unmap_start; |
Dave Hansen | 2a1dcb1 | 2015-06-07 11:37:03 -0700 | [diff] [blame] | 970 | trace_mpx_unmap_search(start, end); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 971 | |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 972 | one_unmap_start = start; |
| 973 | while (one_unmap_start < end) { |
| 974 | int ret; |
| 975 | unsigned long next_unmap_start = ALIGN(one_unmap_start+1, |
| 976 | bd_entry_virt_space(mm)); |
| 977 | unsigned long one_unmap_end = end; |
| 978 | /* |
| 979 | * if the end is beyond the current bounds table, |
| 980 | * move it back so we only deal with a single one |
| 981 | * at a time |
| 982 | */ |
| 983 | if (one_unmap_end > next_unmap_start) |
| 984 | one_unmap_end = next_unmap_start; |
| 985 | ret = try_unmap_single_bt(mm, one_unmap_start, one_unmap_end); |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 986 | if (ret) |
| 987 | return ret; |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 988 | |
Dave Hansen | 3ceaccd | 2015-06-07 11:37:06 -0700 | [diff] [blame] | 989 | one_unmap_start = next_unmap_start; |
| 990 | } |
Dave Hansen | 1de4fa1 | 2014-11-14 07:18:31 -0800 | [diff] [blame] | 991 | return 0; |
| 992 | } |
| 993 | |
| 994 | /* |
| 995 | * Free unused bounds tables covered in a virtual address region being |
| 996 | * munmap()ed. Assume end > start. |
| 997 | * |
| 998 | * This function will be called by do_munmap(), and the VMAs covering |
| 999 | * the virtual address region start...end have already been split if |
| 1000 | * necessary, and the 'vma' is the first vma in this range (start -> end). |
| 1001 | */ |
| 1002 | void mpx_notify_unmap(struct mm_struct *mm, struct vm_area_struct *vma, |
| 1003 | unsigned long start, unsigned long end) |
| 1004 | { |
| 1005 | int ret; |
| 1006 | |
| 1007 | /* |
| 1008 | * Refuse to do anything unless userspace has asked |
| 1009 | * the kernel to help manage the bounds tables, |
| 1010 | */ |
| 1011 | if (!kernel_managing_mpx_tables(current->mm)) |
| 1012 | return; |
| 1013 | /* |
| 1014 | * This will look across the entire 'start -> end' range, |
| 1015 | * and find all of the non-VM_MPX VMAs. |
| 1016 | * |
| 1017 | * To avoid recursion, if a VM_MPX vma is found in the range |
| 1018 | * (start->end), we will not continue follow-up work. This |
| 1019 | * recursion represents having bounds tables for bounds tables, |
| 1020 | * which should not occur normally. Being strict about it here |
| 1021 | * helps ensure that we do not have an exploitable stack overflow. |
| 1022 | */ |
| 1023 | do { |
| 1024 | if (vma->vm_flags & VM_MPX) |
| 1025 | return; |
| 1026 | vma = vma->vm_next; |
| 1027 | } while (vma && vma->vm_start < end); |
| 1028 | |
| 1029 | ret = mpx_unmap_tables(mm, start, end); |
| 1030 | if (ret) |
| 1031 | force_sig(SIGSEGV, current); |
| 1032 | } |