| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Copyright (C) 2015 Anton Ivanov (aivanov@{brocade.com,kot-begemot.co.uk}) |
| 4 | * Copyright (C) 2015 Thomas Meyer (thomas@m3y3r.de) |
| 5 | * Copyright (C) 2004 PathScale, Inc |
| 6 | * Copyright (C) 2004 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com) |
| 7 | */ |
| 8 | |
| 9 | #include <stdlib.h> |
| 10 | #include <stdarg.h> |
| 11 | #include <stdbool.h> |
| 12 | #include <errno.h> |
| 13 | #include <signal.h> |
| 14 | #include <string.h> |
| 15 | #include <strings.h> |
| 16 | #include <as-layout.h> |
| 17 | #include <kern_util.h> |
| 18 | #include <os.h> |
| 19 | #include <sysdep/mcontext.h> |
| 20 | #include <um_malloc.h> |
| 21 | #include <sys/ucontext.h> |
| 22 | #include <timetravel.h> |
| 23 | #include "internal.h" |
| 24 | |
| 25 | void (*sig_info[NSIG])(int, struct siginfo *, struct uml_pt_regs *, void *mc) = { |
| 26 | [SIGTRAP] = relay_signal, |
| 27 | [SIGFPE] = relay_signal, |
| 28 | [SIGILL] = relay_signal, |
| 29 | [SIGWINCH] = winch, |
| 30 | [SIGBUS] = relay_signal, |
| 31 | [SIGSEGV] = segv_handler, |
| 32 | [SIGIO] = sigio_handler, |
| 33 | [SIGCHLD] = sigchld_handler, |
| 34 | }; |
| 35 | |
| 36 | static void sig_handler_common(int sig, struct siginfo *si, mcontext_t *mc) |
| 37 | { |
| 38 | struct uml_pt_regs r; |
| 39 | int save_errno = errno; |
| 40 | |
| 41 | r.is_user = 0; |
| 42 | if (sig == SIGSEGV) { |
| 43 | /* For segfaults, we want the data from the sigcontext. */ |
| 44 | get_regs_from_mc(&r, mc); |
| 45 | GET_FAULTINFO_FROM_MC(r.faultinfo, mc); |
| 46 | } |
| 47 | |
| 48 | /* enable signals if sig isn't IRQ signal */ |
| 49 | if ((sig != SIGIO) && (sig != SIGWINCH) && (sig != SIGCHLD)) |
| 50 | unblock_signals_trace(); |
| 51 | |
| 52 | (*sig_info[sig])(sig, si, &r, mc); |
| 53 | |
| 54 | errno = save_errno; |
| 55 | } |
| 56 | |
| 57 | /* |
| 58 | * These are the asynchronous signals. SIGPROF is excluded because we want to |
| 59 | * be able to profile all of UML, not just the non-critical sections. If |
| 60 | * profiling is not thread-safe, then that is not my problem. We can disable |
| 61 | * profiling when SMP is enabled in that case. |
| 62 | */ |
| 63 | #define SIGIO_BIT 0 |
| 64 | #define SIGIO_MASK (1 << SIGIO_BIT) |
| 65 | |
| 66 | #define SIGALRM_BIT 1 |
| 67 | #define SIGALRM_MASK (1 << SIGALRM_BIT) |
| 68 | |
| 69 | #define SIGCHLD_BIT 2 |
| 70 | #define SIGCHLD_MASK (1 << SIGCHLD_BIT) |
| 71 | |
| 72 | __thread int signals_enabled; |
| 73 | #if IS_ENABLED(CONFIG_UML_TIME_TRAVEL_SUPPORT) |
| 74 | static int signals_blocked, signals_blocked_pending; |
| 75 | #endif |
| 76 | static __thread unsigned int signals_pending; |
| 77 | static __thread unsigned int signals_active; |
| 78 | |
| 79 | static void sig_handler(int sig, struct siginfo *si, mcontext_t *mc) |
| 80 | { |
| 81 | int enabled = signals_enabled; |
| 82 | |
| 83 | #if IS_ENABLED(CONFIG_UML_TIME_TRAVEL_SUPPORT) |
| 84 | if ((signals_blocked || |
| 85 | __atomic_load_n(&signals_blocked_pending, __ATOMIC_SEQ_CST)) && |
| 86 | (sig == SIGIO)) { |
| 87 | /* increment so unblock will do another round */ |
| 88 | __atomic_add_fetch(&signals_blocked_pending, 1, |
| 89 | __ATOMIC_SEQ_CST); |
| 90 | return; |
| 91 | } |
| 92 | #endif |
| 93 | |
| 94 | if (!enabled && (sig == SIGIO)) { |
| 95 | /* |
| 96 | * In TT_MODE_EXTERNAL, need to still call time-travel |
| 97 | * handlers. This will mark signals_pending by itself |
| 98 | * (only if necessary.) |
| 99 | * Note we won't get here if signals are hard-blocked |
| 100 | * (which is handled above), in that case the hard- |
| 101 | * unblock will handle things. |
| 102 | */ |
| 103 | if (time_travel_mode == TT_MODE_EXTERNAL) |
| 104 | sigio_run_timetravel_handlers(); |
| 105 | else |
| 106 | signals_pending |= SIGIO_MASK; |
| 107 | return; |
| 108 | } |
| 109 | |
| 110 | if (!enabled && (sig == SIGCHLD)) { |
| 111 | signals_pending |= SIGCHLD_MASK; |
| 112 | return; |
| 113 | } |
| 114 | |
| 115 | block_signals_trace(); |
| 116 | |
| 117 | sig_handler_common(sig, si, mc); |
| 118 | |
| 119 | um_set_signals_trace(enabled); |
| 120 | } |
| 121 | |
| 122 | static void timer_real_alarm_handler(mcontext_t *mc) |
| 123 | { |
| 124 | struct uml_pt_regs regs; |
| 125 | |
| 126 | if (mc != NULL) |
| 127 | get_regs_from_mc(®s, mc); |
| 128 | else |
| 129 | memset(®s, 0, sizeof(regs)); |
| 130 | timer_handler(SIGALRM, NULL, ®s); |
| 131 | } |
| 132 | |
| 133 | static void timer_alarm_handler(int sig, struct siginfo *unused_si, mcontext_t *mc) |
| 134 | { |
| 135 | int enabled; |
| 136 | |
| 137 | enabled = signals_enabled; |
| 138 | if (!signals_enabled) { |
| 139 | signals_pending |= SIGALRM_MASK; |
| 140 | return; |
| 141 | } |
| 142 | |
| 143 | block_signals_trace(); |
| 144 | |
| 145 | signals_active |= SIGALRM_MASK; |
| 146 | |
| 147 | timer_real_alarm_handler(mc); |
| 148 | |
| 149 | signals_active &= ~SIGALRM_MASK; |
| 150 | |
| 151 | um_set_signals_trace(enabled); |
| 152 | } |
| 153 | |
| 154 | void deliver_alarm(void) { |
| 155 | timer_alarm_handler(SIGALRM, NULL, NULL); |
| 156 | } |
| 157 | |
| 158 | void timer_set_signal_handler(void) |
| 159 | { |
| 160 | set_handler(SIGALRM); |
| 161 | } |
| 162 | |
| 163 | int timer_alarm_pending(void) |
| 164 | { |
| 165 | return !!(signals_pending & SIGALRM_MASK); |
| 166 | } |
| 167 | |
| 168 | void set_sigstack(void *sig_stack, int size) |
| 169 | { |
| 170 | stack_t stack = { |
| 171 | .ss_flags = 0, |
| 172 | .ss_sp = sig_stack, |
| 173 | .ss_size = size |
| 174 | }; |
| 175 | |
| 176 | if (sigaltstack(&stack, NULL) != 0) |
| 177 | panic("enabling signal stack failed, errno = %d\n" , errno); |
| 178 | } |
| 179 | |
| 180 | static void sigusr1_handler(int sig, struct siginfo *unused_si, mcontext_t *mc) |
| 181 | { |
| 182 | uml_pm_wake(); |
| 183 | } |
| 184 | |
| 185 | void register_pm_wake_signal(void) |
| 186 | { |
| 187 | set_handler(SIGUSR1); |
| 188 | } |
| 189 | |
| 190 | static void (*handlers[_NSIG])(int sig, struct siginfo *si, mcontext_t *mc) = { |
| 191 | [SIGSEGV] = sig_handler, |
| 192 | [SIGBUS] = sig_handler, |
| 193 | [SIGILL] = sig_handler, |
| 194 | [SIGFPE] = sig_handler, |
| 195 | [SIGTRAP] = sig_handler, |
| 196 | |
| 197 | [SIGIO] = sig_handler, |
| 198 | [SIGWINCH] = sig_handler, |
| 199 | /* SIGCHLD is only actually registered in seccomp mode. */ |
| 200 | [SIGCHLD] = sig_handler, |
| 201 | [SIGALRM] = timer_alarm_handler, |
| 202 | |
| 203 | [SIGUSR1] = sigusr1_handler, |
| 204 | }; |
| 205 | |
| 206 | static void hard_handler(int sig, siginfo_t *si, void *p) |
| 207 | { |
| 208 | ucontext_t *uc = p; |
| 209 | mcontext_t *mc = &uc->uc_mcontext; |
| 210 | |
| 211 | (*handlers[sig])(sig, (struct siginfo *)si, mc); |
| 212 | } |
| 213 | |
| 214 | void set_handler(int sig) |
| 215 | { |
| 216 | struct sigaction action; |
| 217 | int flags = SA_SIGINFO | SA_ONSTACK; |
| 218 | sigset_t sig_mask; |
| 219 | |
| 220 | action.sa_sigaction = hard_handler; |
| 221 | |
| 222 | /* block irq ones */ |
| 223 | sigemptyset(&action.sa_mask); |
| 224 | sigaddset(&action.sa_mask, SIGIO); |
| 225 | sigaddset(&action.sa_mask, SIGWINCH); |
| 226 | sigaddset(&action.sa_mask, SIGALRM); |
| 227 | |
| 228 | if (sig == SIGSEGV) |
| 229 | flags |= SA_NODEFER; |
| 230 | |
| 231 | if (sigismember(&action.sa_mask, sig)) |
| 232 | flags |= SA_RESTART; /* if it's an irq signal */ |
| 233 | |
| 234 | action.sa_flags = flags; |
| 235 | action.sa_restorer = NULL; |
| 236 | if (sigaction(sig, &action, NULL) < 0) |
| 237 | panic("sigaction failed - errno = %d\n" , errno); |
| 238 | |
| 239 | sigemptyset(&sig_mask); |
| 240 | sigaddset(&sig_mask, sig); |
| 241 | if (sigprocmask(SIG_UNBLOCK, &sig_mask, NULL) < 0) |
| 242 | panic("sigprocmask failed - errno = %d\n" , errno); |
| 243 | } |
| 244 | |
| 245 | void send_sigio_to_self(void) |
| 246 | { |
| 247 | kill(os_getpid(), SIGIO); |
| 248 | } |
| 249 | |
| 250 | int change_sig(int signal, int on) |
| 251 | { |
| 252 | sigset_t sigset; |
| 253 | |
| 254 | sigemptyset(&sigset); |
| 255 | sigaddset(&sigset, signal); |
| 256 | if (sigprocmask(on ? SIG_UNBLOCK : SIG_BLOCK, &sigset, NULL) < 0) |
| 257 | return -errno; |
| 258 | |
| 259 | return 0; |
| 260 | } |
| 261 | |
| 262 | static inline void __block_signals(void) |
| 263 | { |
| 264 | if (!signals_enabled) |
| 265 | return; |
| 266 | |
| 267 | os_local_ipi_disable(); |
| 268 | barrier(); |
| 269 | signals_enabled = 0; |
| 270 | } |
| 271 | |
| 272 | static inline void __unblock_signals(void) |
| 273 | { |
| 274 | if (signals_enabled) |
| 275 | return; |
| 276 | |
| 277 | signals_enabled = 1; |
| 278 | barrier(); |
| 279 | os_local_ipi_enable(); |
| 280 | } |
| 281 | |
| 282 | void block_signals(void) |
| 283 | { |
| 284 | __block_signals(); |
| 285 | /* |
| 286 | * This must return with signals disabled, so this barrier |
| 287 | * ensures that writes are flushed out before the return. |
| 288 | * This might matter if gcc figures out how to inline this and |
| 289 | * decides to shuffle this code into the caller. |
| 290 | */ |
| 291 | barrier(); |
| 292 | } |
| 293 | |
| 294 | void unblock_signals(void) |
| 295 | { |
| 296 | int save_pending; |
| 297 | |
| 298 | if (signals_enabled == 1) |
| 299 | return; |
| 300 | |
| 301 | __unblock_signals(); |
| 302 | |
| 303 | #if IS_ENABLED(CONFIG_UML_TIME_TRAVEL_SUPPORT) |
| 304 | deliver_time_travel_irqs(); |
| 305 | #endif |
| 306 | |
| 307 | /* |
| 308 | * We loop because the IRQ handler returns with interrupts off. So, |
| 309 | * interrupts may have arrived and we need to re-enable them and |
| 310 | * recheck signals_pending. |
| 311 | */ |
| 312 | while (1) { |
| 313 | /* |
| 314 | * Save and reset save_pending after enabling signals. This |
| 315 | * way, signals_pending won't be changed while we're reading it. |
| 316 | * |
| 317 | * Setting signals_enabled and reading signals_pending must |
| 318 | * happen in this order, so have the barrier here. |
| 319 | */ |
| 320 | barrier(); |
| 321 | |
| 322 | save_pending = signals_pending; |
| 323 | if (save_pending == 0) |
| 324 | return; |
| 325 | |
| 326 | signals_pending = 0; |
| 327 | |
| 328 | /* |
| 329 | * We have pending interrupts, so disable signals, as the |
| 330 | * handlers expect them off when they are called. They will |
| 331 | * be enabled again above. We need to trace this, as we're |
| 332 | * expected to be enabling interrupts already, but any more |
| 333 | * tracing that happens inside the handlers we call for the |
| 334 | * pending signals will mess up the tracing state. |
| 335 | */ |
| 336 | __block_signals(); |
| 337 | um_trace_signals_off(); |
| 338 | |
| 339 | /* |
| 340 | * Deal with SIGIO first because the alarm handler might |
| 341 | * schedule, leaving the pending SIGIO stranded until we come |
| 342 | * back here. |
| 343 | * |
| 344 | * SIGIO's handler doesn't use siginfo or mcontext, |
| 345 | * so they can be NULL. |
| 346 | */ |
| 347 | if (save_pending & SIGIO_MASK) |
| 348 | sig_handler_common(SIGIO, NULL, NULL); |
| 349 | |
| 350 | if (save_pending & SIGCHLD_MASK) { |
| 351 | struct uml_pt_regs regs = {}; |
| 352 | |
| 353 | sigchld_handler(SIGCHLD, NULL, ®s, NULL); |
| 354 | } |
| 355 | |
| 356 | /* Do not reenter the handler */ |
| 357 | |
| 358 | if ((save_pending & SIGALRM_MASK) && (!(signals_active & SIGALRM_MASK))) |
| 359 | timer_real_alarm_handler(NULL); |
| 360 | |
| 361 | /* Rerun the loop only if there is still pending SIGIO and not in TIMER handler */ |
| 362 | |
| 363 | if (!(signals_pending & SIGIO_MASK) && (signals_active & SIGALRM_MASK)) |
| 364 | return; |
| 365 | |
| 366 | /* Re-enable signals and trace that we're doing so. */ |
| 367 | um_trace_signals_on(); |
| 368 | __unblock_signals(); |
| 369 | } |
| 370 | } |
| 371 | |
| 372 | int um_get_signals(void) |
| 373 | { |
| 374 | return signals_enabled; |
| 375 | } |
| 376 | |
| 377 | int um_set_signals(int enable) |
| 378 | { |
| 379 | int ret; |
| 380 | if (signals_enabled == enable) |
| 381 | return enable; |
| 382 | |
| 383 | ret = signals_enabled; |
| 384 | if (enable) |
| 385 | unblock_signals(); |
| 386 | else block_signals(); |
| 387 | |
| 388 | return ret; |
| 389 | } |
| 390 | |
| 391 | int um_set_signals_trace(int enable) |
| 392 | { |
| 393 | int ret; |
| 394 | if (signals_enabled == enable) |
| 395 | return enable; |
| 396 | |
| 397 | ret = signals_enabled; |
| 398 | if (enable) |
| 399 | unblock_signals_trace(); |
| 400 | else |
| 401 | block_signals_trace(); |
| 402 | |
| 403 | return ret; |
| 404 | } |
| 405 | |
| 406 | #if IS_ENABLED(CONFIG_UML_TIME_TRAVEL_SUPPORT) |
| 407 | void mark_sigio_pending(void) |
| 408 | { |
| 409 | /* |
| 410 | * It would seem that this should be atomic so |
| 411 | * it isn't a read-modify-write with a signal |
| 412 | * that could happen in the middle, losing the |
| 413 | * value set by the signal. |
| 414 | * |
| 415 | * However, this function is only called when in |
| 416 | * time-travel=ext simulation mode, in which case |
| 417 | * the only signal ever pending is SIGIO, which |
| 418 | * is blocked while this can be called, and the |
| 419 | * timer signal (SIGALRM) cannot happen. |
| 420 | */ |
| 421 | signals_pending |= SIGIO_MASK; |
| 422 | } |
| 423 | |
| 424 | void block_signals_hard(void) |
| 425 | { |
| 426 | signals_blocked++; |
| 427 | barrier(); |
| 428 | } |
| 429 | |
| 430 | void unblock_signals_hard(void) |
| 431 | { |
| 432 | static bool unblocking; |
| 433 | |
| 434 | if (!signals_blocked) |
| 435 | panic("unblocking signals while not blocked" ); |
| 436 | |
| 437 | if (--signals_blocked) |
| 438 | return; |
| 439 | /* |
| 440 | * Must be set to 0 before we check pending so the |
| 441 | * SIGIO handler will run as normal unless we're still |
| 442 | * going to process signals_blocked_pending. |
| 443 | */ |
| 444 | barrier(); |
| 445 | |
| 446 | /* |
| 447 | * Note that block_signals_hard()/unblock_signals_hard() can be called |
| 448 | * within the unblock_signals()/sigio_run_timetravel_handlers() below. |
| 449 | * This would still be prone to race conditions since it's actually a |
| 450 | * call _within_ e.g. vu_req_read_message(), where we observed this |
| 451 | * issue, which loops. Thus, if the inner call handles the recorded |
| 452 | * pending signals, we can get out of the inner call with the real |
| 453 | * signal hander no longer blocked, and still have a race. Thus don't |
| 454 | * handle unblocking in the inner call, if it happens, but only in |
| 455 | * the outermost call - 'unblocking' serves as an ownership for the |
| 456 | * signals_blocked_pending decrement. |
| 457 | */ |
| 458 | if (unblocking) |
| 459 | return; |
| 460 | unblocking = true; |
| 461 | |
| 462 | while (__atomic_load_n(&signals_blocked_pending, __ATOMIC_SEQ_CST)) { |
| 463 | if (signals_enabled) { |
| 464 | /* signals are enabled so we can touch this */ |
| 465 | signals_pending |= SIGIO_MASK; |
| 466 | /* |
| 467 | * this is a bit inefficient, but that's |
| 468 | * not really important |
| 469 | */ |
| 470 | block_signals(); |
| 471 | unblock_signals(); |
| 472 | } else { |
| 473 | /* |
| 474 | * we need to run time-travel handlers even |
| 475 | * if not enabled |
| 476 | */ |
| 477 | sigio_run_timetravel_handlers(); |
| 478 | } |
| 479 | |
| 480 | /* |
| 481 | * The decrement of signals_blocked_pending must be atomic so |
| 482 | * that the signal handler will either happen before or after |
| 483 | * the decrement, not during a read-modify-write: |
| 484 | * - If it happens before, it can increment it and we'll |
| 485 | * decrement it and do another round in the loop. |
| 486 | * - If it happens after it'll see 0 for both signals_blocked |
| 487 | * and signals_blocked_pending and thus run the handler as |
| 488 | * usual (subject to signals_enabled, but that's unrelated.) |
| 489 | * |
| 490 | * Note that a call to unblock_signals_hard() within the calls |
| 491 | * to unblock_signals() or sigio_run_timetravel_handlers() above |
| 492 | * will do nothing due to the 'unblocking' state, so this cannot |
| 493 | * underflow as the only one decrementing will be the outermost |
| 494 | * one. |
| 495 | */ |
| 496 | if (__atomic_sub_fetch(&signals_blocked_pending, 1, |
| 497 | __ATOMIC_SEQ_CST) < 0) |
| 498 | panic("signals_blocked_pending underflow" ); |
| 499 | } |
| 500 | |
| 501 | unblocking = false; |
| 502 | } |
| 503 | #endif |
| 504 | |