1 // SPDX-License-Identifier: GPL-2.0+ 2 /* 3 * ipmi_watchdog.c 4 * 5 * A watchdog timer based upon the IPMI interface. 6 * 7 * Author: MontaVista Software, Inc. 8 * Corey Minyard <minyard@mvista.com> 9 * source@mvista.com 10 * 11 * Copyright 2002 MontaVista Software Inc. 12 */ 13 14 #include <linux/module.h> 15 #include <linux/moduleparam.h> 16 #include <linux/ipmi.h> 17 #include <linux/ipmi_smi.h> 18 #include <linux/mutex.h> 19 #include <linux/watchdog.h> 20 #include <linux/miscdevice.h> 21 #include <linux/init.h> 22 #include <linux/completion.h> 23 #include <linux/kdebug.h> 24 #include <linux/rwsem.h> 25 #include <linux/errno.h> 26 #include <linux/uaccess.h> 27 #include <linux/notifier.h> 28 #include <linux/nmi.h> 29 #include <linux/reboot.h> 30 #include <linux/wait.h> 31 #include <linux/poll.h> 32 #include <linux/string.h> 33 #include <linux/ctype.h> 34 #include <linux/delay.h> 35 #include <linux/atomic.h> 36 #include <linux/sched/signal.h> 37 38 #ifdef CONFIG_X86 39 /* 40 * This is ugly, but I've determined that x86 is the only architecture 41 * that can reasonably support the IPMI NMI watchdog timeout at this 42 * time. If another architecture adds this capability somehow, it 43 * will have to be a somewhat different mechanism and I have no idea 44 * how it will work. So in the unlikely event that another 45 * architecture supports this, we can figure out a good generic 46 * mechanism for it at that time. 47 */ 48 #include <asm/kdebug.h> 49 #include <asm/nmi.h> 50 #define HAVE_DIE_NMI 51 #endif 52 53 #define PFX "IPMI Watchdog: " 54 55 /* 56 * The IPMI command/response information for the watchdog timer. 57 */ 58 59 /* values for byte 1 of the set command, byte 2 of the get response. */ 60 #define WDOG_DONT_LOG (1 << 7) 61 #define WDOG_DONT_STOP_ON_SET (1 << 6) 62 #define WDOG_SET_TIMER_USE(byte, use) \ 63 byte = ((byte) & 0xf8) | ((use) & 0x7) 64 #define WDOG_GET_TIMER_USE(byte) ((byte) & 0x7) 65 #define WDOG_TIMER_USE_BIOS_FRB2 1 66 #define WDOG_TIMER_USE_BIOS_POST 2 67 #define WDOG_TIMER_USE_OS_LOAD 3 68 #define WDOG_TIMER_USE_SMS_OS 4 69 #define WDOG_TIMER_USE_OEM 5 70 71 /* values for byte 2 of the set command, byte 3 of the get response. */ 72 #define WDOG_SET_PRETIMEOUT_ACT(byte, use) \ 73 byte = ((byte) & 0x8f) | (((use) & 0x7) << 4) 74 #define WDOG_GET_PRETIMEOUT_ACT(byte) (((byte) >> 4) & 0x7) 75 #define WDOG_PRETIMEOUT_NONE 0 76 #define WDOG_PRETIMEOUT_SMI 1 77 #define WDOG_PRETIMEOUT_NMI 2 78 #define WDOG_PRETIMEOUT_MSG_INT 3 79 80 /* Operations that can be performed on a pretimout. */ 81 #define WDOG_PREOP_NONE 0 82 #define WDOG_PREOP_PANIC 1 83 /* Cause data to be available to read. Doesn't work in NMI mode. */ 84 #define WDOG_PREOP_GIVE_DATA 2 85 86 /* Actions to perform on a full timeout. */ 87 #define WDOG_SET_TIMEOUT_ACT(byte, use) \ 88 byte = ((byte) & 0xf8) | ((use) & 0x7) 89 #define WDOG_GET_TIMEOUT_ACT(byte) ((byte) & 0x7) 90 #define WDOG_TIMEOUT_NONE 0 91 #define WDOG_TIMEOUT_RESET 1 92 #define WDOG_TIMEOUT_POWER_DOWN 2 93 #define WDOG_TIMEOUT_POWER_CYCLE 3 94 95 /* 96 * Byte 3 of the get command, byte 4 of the get response is the 97 * pre-timeout in seconds. 98 */ 99 100 /* Bits for setting byte 4 of the set command, byte 5 of the get response. */ 101 #define WDOG_EXPIRE_CLEAR_BIOS_FRB2 (1 << 1) 102 #define WDOG_EXPIRE_CLEAR_BIOS_POST (1 << 2) 103 #define WDOG_EXPIRE_CLEAR_OS_LOAD (1 << 3) 104 #define WDOG_EXPIRE_CLEAR_SMS_OS (1 << 4) 105 #define WDOG_EXPIRE_CLEAR_OEM (1 << 5) 106 107 /* 108 * Setting/getting the watchdog timer value. This is for bytes 5 and 109 * 6 (the timeout time) of the set command, and bytes 6 and 7 (the 110 * timeout time) and 8 and 9 (the current countdown value) of the 111 * response. The timeout value is given in seconds (in the command it 112 * is 100ms intervals). 113 */ 114 #define WDOG_SET_TIMEOUT(byte1, byte2, val) \ 115 (byte1) = (((val) * 10) & 0xff), (byte2) = (((val) * 10) >> 8) 116 #define WDOG_GET_TIMEOUT(byte1, byte2) \ 117 (((byte1) | ((byte2) << 8)) / 10) 118 119 #define IPMI_WDOG_RESET_TIMER 0x22 120 #define IPMI_WDOG_SET_TIMER 0x24 121 #define IPMI_WDOG_GET_TIMER 0x25 122 123 #define IPMI_WDOG_TIMER_NOT_INIT_RESP 0x80 124 125 static DEFINE_MUTEX(ipmi_watchdog_mutex); 126 static bool nowayout = WATCHDOG_NOWAYOUT; 127 128 static ipmi_user_t watchdog_user; 129 static int watchdog_ifnum; 130 131 /* Default the timeout to 10 seconds. */ 132 static int timeout = 10; 133 134 /* The pre-timeout is disabled by default. */ 135 static int pretimeout; 136 137 /* Default timeout to set on panic */ 138 static int panic_wdt_timeout = 255; 139 140 /* Default action is to reset the board on a timeout. */ 141 static unsigned char action_val = WDOG_TIMEOUT_RESET; 142 143 static char action[16] = "reset"; 144 145 static unsigned char preaction_val = WDOG_PRETIMEOUT_NONE; 146 147 static char preaction[16] = "pre_none"; 148 149 static unsigned char preop_val = WDOG_PREOP_NONE; 150 151 static char preop[16] = "preop_none"; 152 static DEFINE_SPINLOCK(ipmi_read_lock); 153 static char data_to_read; 154 static DECLARE_WAIT_QUEUE_HEAD(read_q); 155 static struct fasync_struct *fasync_q; 156 static char pretimeout_since_last_heartbeat; 157 static char expect_close; 158 159 static int ifnum_to_use = -1; 160 161 /* Parameters to ipmi_set_timeout */ 162 #define IPMI_SET_TIMEOUT_NO_HB 0 163 #define IPMI_SET_TIMEOUT_HB_IF_NECESSARY 1 164 #define IPMI_SET_TIMEOUT_FORCE_HB 2 165 166 static int ipmi_set_timeout(int do_heartbeat); 167 static void ipmi_register_watchdog(int ipmi_intf); 168 static void ipmi_unregister_watchdog(int ipmi_intf); 169 170 /* 171 * If true, the driver will start running as soon as it is configured 172 * and ready. 173 */ 174 static int start_now; 175 176 static int set_param_timeout(const char *val, const struct kernel_param *kp) 177 { 178 char *endp; 179 int l; 180 int rv = 0; 181 182 if (!val) 183 return -EINVAL; 184 l = simple_strtoul(val, &endp, 0); 185 if (endp == val) 186 return -EINVAL; 187 188 *((int *)kp->arg) = l; 189 if (watchdog_user) 190 rv = ipmi_set_timeout(IPMI_SET_TIMEOUT_HB_IF_NECESSARY); 191 192 return rv; 193 } 194 195 static const struct kernel_param_ops param_ops_timeout = { 196 .set = set_param_timeout, 197 .get = param_get_int, 198 }; 199 #define param_check_timeout param_check_int 200 201 typedef int (*action_fn)(const char *intval, char *outval); 202 203 static int action_op(const char *inval, char *outval); 204 static int preaction_op(const char *inval, char *outval); 205 static int preop_op(const char *inval, char *outval); 206 static void check_parms(void); 207 208 static int set_param_str(const char *val, const struct kernel_param *kp) 209 { 210 action_fn fn = (action_fn) kp->arg; 211 int rv = 0; 212 char valcp[16]; 213 char *s; 214 215 strncpy(valcp, val, 15); 216 valcp[15] = '\0'; 217 218 s = strstrip(valcp); 219 220 rv = fn(s, NULL); 221 if (rv) 222 goto out; 223 224 check_parms(); 225 if (watchdog_user) 226 rv = ipmi_set_timeout(IPMI_SET_TIMEOUT_HB_IF_NECESSARY); 227 228 out: 229 return rv; 230 } 231 232 static int get_param_str(char *buffer, const struct kernel_param *kp) 233 { 234 action_fn fn = (action_fn) kp->arg; 235 int rv; 236 237 rv = fn(NULL, buffer); 238 if (rv) 239 return rv; 240 return strlen(buffer); 241 } 242 243 244 static int set_param_wdog_ifnum(const char *val, const struct kernel_param *kp) 245 { 246 int rv = param_set_int(val, kp); 247 if (rv) 248 return rv; 249 if ((ifnum_to_use < 0) || (ifnum_to_use == watchdog_ifnum)) 250 return 0; 251 252 ipmi_unregister_watchdog(watchdog_ifnum); 253 ipmi_register_watchdog(ifnum_to_use); 254 return 0; 255 } 256 257 static const struct kernel_param_ops param_ops_wdog_ifnum = { 258 .set = set_param_wdog_ifnum, 259 .get = param_get_int, 260 }; 261 262 #define param_check_wdog_ifnum param_check_int 263 264 static const struct kernel_param_ops param_ops_str = { 265 .set = set_param_str, 266 .get = get_param_str, 267 }; 268 269 module_param(ifnum_to_use, wdog_ifnum, 0644); 270 MODULE_PARM_DESC(ifnum_to_use, "The interface number to use for the watchdog " 271 "timer. Setting to -1 defaults to the first registered " 272 "interface"); 273 274 module_param(timeout, timeout, 0644); 275 MODULE_PARM_DESC(timeout, "Timeout value in seconds."); 276 277 module_param(pretimeout, timeout, 0644); 278 MODULE_PARM_DESC(pretimeout, "Pretimeout value in seconds."); 279 280 module_param(panic_wdt_timeout, timeout, 0644); 281 MODULE_PARM_DESC(panic_wdt_timeout, "Timeout value on kernel panic in seconds."); 282 283 module_param_cb(action, ¶m_ops_str, action_op, 0644); 284 MODULE_PARM_DESC(action, "Timeout action. One of: " 285 "reset, none, power_cycle, power_off."); 286 287 module_param_cb(preaction, ¶m_ops_str, preaction_op, 0644); 288 MODULE_PARM_DESC(preaction, "Pretimeout action. One of: " 289 "pre_none, pre_smi, pre_nmi, pre_int."); 290 291 module_param_cb(preop, ¶m_ops_str, preop_op, 0644); 292 MODULE_PARM_DESC(preop, "Pretimeout driver operation. One of: " 293 "preop_none, preop_panic, preop_give_data."); 294 295 module_param(start_now, int, 0444); 296 MODULE_PARM_DESC(start_now, "Set to 1 to start the watchdog as" 297 "soon as the driver is loaded."); 298 299 module_param(nowayout, bool, 0644); 300 MODULE_PARM_DESC(nowayout, "Watchdog cannot be stopped once started " 301 "(default=CONFIG_WATCHDOG_NOWAYOUT)"); 302 303 /* Default state of the timer. */ 304 static unsigned char ipmi_watchdog_state = WDOG_TIMEOUT_NONE; 305 306 /* If shutting down via IPMI, we ignore the heartbeat. */ 307 static int ipmi_ignore_heartbeat; 308 309 /* Is someone using the watchdog? Only one user is allowed. */ 310 static unsigned long ipmi_wdog_open; 311 312 /* 313 * If set to 1, the heartbeat command will set the state to reset and 314 * start the timer. The timer doesn't normally run when the driver is 315 * first opened until the heartbeat is set the first time, this 316 * variable is used to accomplish this. 317 */ 318 static int ipmi_start_timer_on_heartbeat; 319 320 /* IPMI version of the BMC. */ 321 static unsigned char ipmi_version_major; 322 static unsigned char ipmi_version_minor; 323 324 /* If a pretimeout occurs, this is used to allow only one panic to happen. */ 325 static atomic_t preop_panic_excl = ATOMIC_INIT(-1); 326 327 #ifdef HAVE_DIE_NMI 328 static int testing_nmi; 329 static int nmi_handler_registered; 330 #endif 331 332 static int ipmi_heartbeat(void); 333 334 /* 335 * We use a mutex to make sure that only one thing can send a set 336 * timeout at one time, because we only have one copy of the data. 337 * The mutex is claimed when the set_timeout is sent and freed 338 * when both messages are free. 339 */ 340 static atomic_t set_timeout_tofree = ATOMIC_INIT(0); 341 static DEFINE_MUTEX(set_timeout_lock); 342 static DECLARE_COMPLETION(set_timeout_wait); 343 static void set_timeout_free_smi(struct ipmi_smi_msg *msg) 344 { 345 if (atomic_dec_and_test(&set_timeout_tofree)) 346 complete(&set_timeout_wait); 347 } 348 static void set_timeout_free_recv(struct ipmi_recv_msg *msg) 349 { 350 if (atomic_dec_and_test(&set_timeout_tofree)) 351 complete(&set_timeout_wait); 352 } 353 static struct ipmi_smi_msg set_timeout_smi_msg = { 354 .done = set_timeout_free_smi 355 }; 356 static struct ipmi_recv_msg set_timeout_recv_msg = { 357 .done = set_timeout_free_recv 358 }; 359 360 static int i_ipmi_set_timeout(struct ipmi_smi_msg *smi_msg, 361 struct ipmi_recv_msg *recv_msg, 362 int *send_heartbeat_now) 363 { 364 struct kernel_ipmi_msg msg; 365 unsigned char data[6]; 366 int rv; 367 struct ipmi_system_interface_addr addr; 368 int hbnow = 0; 369 370 371 /* These can be cleared as we are setting the timeout. */ 372 pretimeout_since_last_heartbeat = 0; 373 374 data[0] = 0; 375 WDOG_SET_TIMER_USE(data[0], WDOG_TIMER_USE_SMS_OS); 376 377 if ((ipmi_version_major > 1) 378 || ((ipmi_version_major == 1) && (ipmi_version_minor >= 5))) { 379 /* This is an IPMI 1.5-only feature. */ 380 data[0] |= WDOG_DONT_STOP_ON_SET; 381 } else if (ipmi_watchdog_state != WDOG_TIMEOUT_NONE) { 382 /* 383 * In ipmi 1.0, setting the timer stops the watchdog, we 384 * need to start it back up again. 385 */ 386 hbnow = 1; 387 } 388 389 data[1] = 0; 390 WDOG_SET_TIMEOUT_ACT(data[1], ipmi_watchdog_state); 391 if ((pretimeout > 0) && (ipmi_watchdog_state != WDOG_TIMEOUT_NONE)) { 392 WDOG_SET_PRETIMEOUT_ACT(data[1], preaction_val); 393 data[2] = pretimeout; 394 } else { 395 WDOG_SET_PRETIMEOUT_ACT(data[1], WDOG_PRETIMEOUT_NONE); 396 data[2] = 0; /* No pretimeout. */ 397 } 398 data[3] = 0; 399 WDOG_SET_TIMEOUT(data[4], data[5], timeout); 400 401 addr.addr_type = IPMI_SYSTEM_INTERFACE_ADDR_TYPE; 402 addr.channel = IPMI_BMC_CHANNEL; 403 addr.lun = 0; 404 405 msg.netfn = 0x06; 406 msg.cmd = IPMI_WDOG_SET_TIMER; 407 msg.data = data; 408 msg.data_len = sizeof(data); 409 rv = ipmi_request_supply_msgs(watchdog_user, 410 (struct ipmi_addr *) &addr, 411 0, 412 &msg, 413 NULL, 414 smi_msg, 415 recv_msg, 416 1); 417 if (rv) { 418 printk(KERN_WARNING PFX "set timeout error: %d\n", 419 rv); 420 } 421 422 if (send_heartbeat_now) 423 *send_heartbeat_now = hbnow; 424 425 return rv; 426 } 427 428 static int ipmi_set_timeout(int do_heartbeat) 429 { 430 int send_heartbeat_now; 431 int rv; 432 433 434 /* We can only send one of these at a time. */ 435 mutex_lock(&set_timeout_lock); 436 437 atomic_set(&set_timeout_tofree, 2); 438 439 rv = i_ipmi_set_timeout(&set_timeout_smi_msg, 440 &set_timeout_recv_msg, 441 &send_heartbeat_now); 442 if (rv) { 443 mutex_unlock(&set_timeout_lock); 444 goto out; 445 } 446 447 wait_for_completion(&set_timeout_wait); 448 449 mutex_unlock(&set_timeout_lock); 450 451 if ((do_heartbeat == IPMI_SET_TIMEOUT_FORCE_HB) 452 || ((send_heartbeat_now) 453 && (do_heartbeat == IPMI_SET_TIMEOUT_HB_IF_NECESSARY))) 454 rv = ipmi_heartbeat(); 455 456 out: 457 return rv; 458 } 459 460 static atomic_t panic_done_count = ATOMIC_INIT(0); 461 462 static void panic_smi_free(struct ipmi_smi_msg *msg) 463 { 464 atomic_dec(&panic_done_count); 465 } 466 static void panic_recv_free(struct ipmi_recv_msg *msg) 467 { 468 atomic_dec(&panic_done_count); 469 } 470 471 static struct ipmi_smi_msg panic_halt_heartbeat_smi_msg = { 472 .done = panic_smi_free 473 }; 474 static struct ipmi_recv_msg panic_halt_heartbeat_recv_msg = { 475 .done = panic_recv_free 476 }; 477 478 static void panic_halt_ipmi_heartbeat(void) 479 { 480 struct kernel_ipmi_msg msg; 481 struct ipmi_system_interface_addr addr; 482 int rv; 483 484 /* 485 * Don't reset the timer if we have the timer turned off, that 486 * re-enables the watchdog. 487 */ 488 if (ipmi_watchdog_state == WDOG_TIMEOUT_NONE) 489 return; 490 491 addr.addr_type = IPMI_SYSTEM_INTERFACE_ADDR_TYPE; 492 addr.channel = IPMI_BMC_CHANNEL; 493 addr.lun = 0; 494 495 msg.netfn = 0x06; 496 msg.cmd = IPMI_WDOG_RESET_TIMER; 497 msg.data = NULL; 498 msg.data_len = 0; 499 atomic_add(1, &panic_done_count); 500 rv = ipmi_request_supply_msgs(watchdog_user, 501 (struct ipmi_addr *) &addr, 502 0, 503 &msg, 504 NULL, 505 &panic_halt_heartbeat_smi_msg, 506 &panic_halt_heartbeat_recv_msg, 507 1); 508 if (rv) 509 atomic_sub(1, &panic_done_count); 510 } 511 512 static struct ipmi_smi_msg panic_halt_smi_msg = { 513 .done = panic_smi_free 514 }; 515 static struct ipmi_recv_msg panic_halt_recv_msg = { 516 .done = panic_recv_free 517 }; 518 519 /* 520 * Special call, doesn't claim any locks. This is only to be called 521 * at panic or halt time, in run-to-completion mode, when the caller 522 * is the only CPU and the only thing that will be going is these IPMI 523 * calls. 524 */ 525 static void panic_halt_ipmi_set_timeout(void) 526 { 527 int send_heartbeat_now; 528 int rv; 529 530 /* Wait for the messages to be free. */ 531 while (atomic_read(&panic_done_count) != 0) 532 ipmi_poll_interface(watchdog_user); 533 atomic_add(1, &panic_done_count); 534 rv = i_ipmi_set_timeout(&panic_halt_smi_msg, 535 &panic_halt_recv_msg, 536 &send_heartbeat_now); 537 if (rv) { 538 atomic_sub(1, &panic_done_count); 539 printk(KERN_WARNING PFX 540 "Unable to extend the watchdog timeout."); 541 } else { 542 if (send_heartbeat_now) 543 panic_halt_ipmi_heartbeat(); 544 } 545 while (atomic_read(&panic_done_count) != 0) 546 ipmi_poll_interface(watchdog_user); 547 } 548 549 /* 550 * We use a mutex to make sure that only one thing can send a 551 * heartbeat at one time, because we only have one copy of the data. 552 * The semaphore is claimed when the set_timeout is sent and freed 553 * when both messages are free. 554 */ 555 static atomic_t heartbeat_tofree = ATOMIC_INIT(0); 556 static DEFINE_MUTEX(heartbeat_lock); 557 static DECLARE_COMPLETION(heartbeat_wait); 558 static void heartbeat_free_smi(struct ipmi_smi_msg *msg) 559 { 560 if (atomic_dec_and_test(&heartbeat_tofree)) 561 complete(&heartbeat_wait); 562 } 563 static void heartbeat_free_recv(struct ipmi_recv_msg *msg) 564 { 565 if (atomic_dec_and_test(&heartbeat_tofree)) 566 complete(&heartbeat_wait); 567 } 568 static struct ipmi_smi_msg heartbeat_smi_msg = { 569 .done = heartbeat_free_smi 570 }; 571 static struct ipmi_recv_msg heartbeat_recv_msg = { 572 .done = heartbeat_free_recv 573 }; 574 575 static int ipmi_heartbeat(void) 576 { 577 struct kernel_ipmi_msg msg; 578 int rv; 579 struct ipmi_system_interface_addr addr; 580 int timeout_retries = 0; 581 582 if (ipmi_ignore_heartbeat) 583 return 0; 584 585 if (ipmi_start_timer_on_heartbeat) { 586 ipmi_start_timer_on_heartbeat = 0; 587 ipmi_watchdog_state = action_val; 588 return ipmi_set_timeout(IPMI_SET_TIMEOUT_FORCE_HB); 589 } else if (pretimeout_since_last_heartbeat) { 590 /* 591 * A pretimeout occurred, make sure we set the timeout. 592 * We don't want to set the action, though, we want to 593 * leave that alone (thus it can't be combined with the 594 * above operation. 595 */ 596 return ipmi_set_timeout(IPMI_SET_TIMEOUT_HB_IF_NECESSARY); 597 } 598 599 mutex_lock(&heartbeat_lock); 600 601 restart: 602 atomic_set(&heartbeat_tofree, 2); 603 604 /* 605 * Don't reset the timer if we have the timer turned off, that 606 * re-enables the watchdog. 607 */ 608 if (ipmi_watchdog_state == WDOG_TIMEOUT_NONE) { 609 mutex_unlock(&heartbeat_lock); 610 return 0; 611 } 612 613 addr.addr_type = IPMI_SYSTEM_INTERFACE_ADDR_TYPE; 614 addr.channel = IPMI_BMC_CHANNEL; 615 addr.lun = 0; 616 617 msg.netfn = 0x06; 618 msg.cmd = IPMI_WDOG_RESET_TIMER; 619 msg.data = NULL; 620 msg.data_len = 0; 621 rv = ipmi_request_supply_msgs(watchdog_user, 622 (struct ipmi_addr *) &addr, 623 0, 624 &msg, 625 NULL, 626 &heartbeat_smi_msg, 627 &heartbeat_recv_msg, 628 1); 629 if (rv) { 630 mutex_unlock(&heartbeat_lock); 631 printk(KERN_WARNING PFX "heartbeat failure: %d\n", 632 rv); 633 return rv; 634 } 635 636 /* Wait for the heartbeat to be sent. */ 637 wait_for_completion(&heartbeat_wait); 638 639 if (heartbeat_recv_msg.msg.data[0] == IPMI_WDOG_TIMER_NOT_INIT_RESP) { 640 timeout_retries++; 641 if (timeout_retries > 3) { 642 printk(KERN_ERR PFX ": Unable to restore the IPMI" 643 " watchdog's settings, giving up.\n"); 644 rv = -EIO; 645 goto out_unlock; 646 } 647 648 /* 649 * The timer was not initialized, that means the BMC was 650 * probably reset and lost the watchdog information. Attempt 651 * to restore the timer's info. Note that we still hold 652 * the heartbeat lock, to keep a heartbeat from happening 653 * in this process, so must say no heartbeat to avoid a 654 * deadlock on this mutex. 655 */ 656 rv = ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 657 if (rv) { 658 printk(KERN_ERR PFX ": Unable to send the command to" 659 " set the watchdog's settings, giving up.\n"); 660 goto out_unlock; 661 } 662 663 /* We might need a new heartbeat, so do it now */ 664 goto restart; 665 } else if (heartbeat_recv_msg.msg.data[0] != 0) { 666 /* 667 * Got an error in the heartbeat response. It was already 668 * reported in ipmi_wdog_msg_handler, but we should return 669 * an error here. 670 */ 671 rv = -EINVAL; 672 } 673 674 out_unlock: 675 mutex_unlock(&heartbeat_lock); 676 677 return rv; 678 } 679 680 static struct watchdog_info ident = { 681 .options = 0, /* WDIOF_SETTIMEOUT, */ 682 .firmware_version = 1, 683 .identity = "IPMI" 684 }; 685 686 static int ipmi_ioctl(struct file *file, 687 unsigned int cmd, unsigned long arg) 688 { 689 void __user *argp = (void __user *)arg; 690 int i; 691 int val; 692 693 switch (cmd) { 694 case WDIOC_GETSUPPORT: 695 i = copy_to_user(argp, &ident, sizeof(ident)); 696 return i ? -EFAULT : 0; 697 698 case WDIOC_SETTIMEOUT: 699 i = copy_from_user(&val, argp, sizeof(int)); 700 if (i) 701 return -EFAULT; 702 timeout = val; 703 return ipmi_set_timeout(IPMI_SET_TIMEOUT_HB_IF_NECESSARY); 704 705 case WDIOC_GETTIMEOUT: 706 i = copy_to_user(argp, &timeout, sizeof(timeout)); 707 if (i) 708 return -EFAULT; 709 return 0; 710 711 case WDIOC_SETPRETIMEOUT: 712 i = copy_from_user(&val, argp, sizeof(int)); 713 if (i) 714 return -EFAULT; 715 pretimeout = val; 716 return ipmi_set_timeout(IPMI_SET_TIMEOUT_HB_IF_NECESSARY); 717 718 case WDIOC_GETPRETIMEOUT: 719 i = copy_to_user(argp, &pretimeout, sizeof(pretimeout)); 720 if (i) 721 return -EFAULT; 722 return 0; 723 724 case WDIOC_KEEPALIVE: 725 return ipmi_heartbeat(); 726 727 case WDIOC_SETOPTIONS: 728 i = copy_from_user(&val, argp, sizeof(int)); 729 if (i) 730 return -EFAULT; 731 if (val & WDIOS_DISABLECARD) { 732 ipmi_watchdog_state = WDOG_TIMEOUT_NONE; 733 ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 734 ipmi_start_timer_on_heartbeat = 0; 735 } 736 737 if (val & WDIOS_ENABLECARD) { 738 ipmi_watchdog_state = action_val; 739 ipmi_set_timeout(IPMI_SET_TIMEOUT_FORCE_HB); 740 } 741 return 0; 742 743 case WDIOC_GETSTATUS: 744 val = 0; 745 i = copy_to_user(argp, &val, sizeof(val)); 746 if (i) 747 return -EFAULT; 748 return 0; 749 750 default: 751 return -ENOIOCTLCMD; 752 } 753 } 754 755 static long ipmi_unlocked_ioctl(struct file *file, 756 unsigned int cmd, 757 unsigned long arg) 758 { 759 int ret; 760 761 mutex_lock(&ipmi_watchdog_mutex); 762 ret = ipmi_ioctl(file, cmd, arg); 763 mutex_unlock(&ipmi_watchdog_mutex); 764 765 return ret; 766 } 767 768 static ssize_t ipmi_write(struct file *file, 769 const char __user *buf, 770 size_t len, 771 loff_t *ppos) 772 { 773 int rv; 774 775 if (len) { 776 if (!nowayout) { 777 size_t i; 778 779 /* In case it was set long ago */ 780 expect_close = 0; 781 782 for (i = 0; i != len; i++) { 783 char c; 784 785 if (get_user(c, buf + i)) 786 return -EFAULT; 787 if (c == 'V') 788 expect_close = 42; 789 } 790 } 791 rv = ipmi_heartbeat(); 792 if (rv) 793 return rv; 794 } 795 return len; 796 } 797 798 static ssize_t ipmi_read(struct file *file, 799 char __user *buf, 800 size_t count, 801 loff_t *ppos) 802 { 803 int rv = 0; 804 wait_queue_entry_t wait; 805 806 if (count <= 0) 807 return 0; 808 809 /* 810 * Reading returns if the pretimeout has gone off, and it only does 811 * it once per pretimeout. 812 */ 813 spin_lock(&ipmi_read_lock); 814 if (!data_to_read) { 815 if (file->f_flags & O_NONBLOCK) { 816 rv = -EAGAIN; 817 goto out; 818 } 819 820 init_waitqueue_entry(&wait, current); 821 add_wait_queue(&read_q, &wait); 822 while (!data_to_read) { 823 set_current_state(TASK_INTERRUPTIBLE); 824 spin_unlock(&ipmi_read_lock); 825 schedule(); 826 spin_lock(&ipmi_read_lock); 827 } 828 remove_wait_queue(&read_q, &wait); 829 830 if (signal_pending(current)) { 831 rv = -ERESTARTSYS; 832 goto out; 833 } 834 } 835 data_to_read = 0; 836 837 out: 838 spin_unlock(&ipmi_read_lock); 839 840 if (rv == 0) { 841 if (copy_to_user(buf, &data_to_read, 1)) 842 rv = -EFAULT; 843 else 844 rv = 1; 845 } 846 847 return rv; 848 } 849 850 static int ipmi_open(struct inode *ino, struct file *filep) 851 { 852 switch (iminor(ino)) { 853 case WATCHDOG_MINOR: 854 if (test_and_set_bit(0, &ipmi_wdog_open)) 855 return -EBUSY; 856 857 858 /* 859 * Don't start the timer now, let it start on the 860 * first heartbeat. 861 */ 862 ipmi_start_timer_on_heartbeat = 1; 863 return nonseekable_open(ino, filep); 864 865 default: 866 return (-ENODEV); 867 } 868 } 869 870 static __poll_t ipmi_poll(struct file *file, poll_table *wait) 871 { 872 __poll_t mask = 0; 873 874 poll_wait(file, &read_q, wait); 875 876 spin_lock(&ipmi_read_lock); 877 if (data_to_read) 878 mask |= (EPOLLIN | EPOLLRDNORM); 879 spin_unlock(&ipmi_read_lock); 880 881 return mask; 882 } 883 884 static int ipmi_fasync(int fd, struct file *file, int on) 885 { 886 int result; 887 888 result = fasync_helper(fd, file, on, &fasync_q); 889 890 return (result); 891 } 892 893 static int ipmi_close(struct inode *ino, struct file *filep) 894 { 895 if (iminor(ino) == WATCHDOG_MINOR) { 896 if (expect_close == 42) { 897 ipmi_watchdog_state = WDOG_TIMEOUT_NONE; 898 ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 899 } else { 900 printk(KERN_CRIT PFX 901 "Unexpected close, not stopping watchdog!\n"); 902 ipmi_heartbeat(); 903 } 904 clear_bit(0, &ipmi_wdog_open); 905 } 906 907 expect_close = 0; 908 909 return 0; 910 } 911 912 static const struct file_operations ipmi_wdog_fops = { 913 .owner = THIS_MODULE, 914 .read = ipmi_read, 915 .poll = ipmi_poll, 916 .write = ipmi_write, 917 .unlocked_ioctl = ipmi_unlocked_ioctl, 918 .open = ipmi_open, 919 .release = ipmi_close, 920 .fasync = ipmi_fasync, 921 .llseek = no_llseek, 922 }; 923 924 static struct miscdevice ipmi_wdog_miscdev = { 925 .minor = WATCHDOG_MINOR, 926 .name = "watchdog", 927 .fops = &ipmi_wdog_fops 928 }; 929 930 static void ipmi_wdog_msg_handler(struct ipmi_recv_msg *msg, 931 void *handler_data) 932 { 933 if (msg->msg.cmd == IPMI_WDOG_RESET_TIMER && 934 msg->msg.data[0] == IPMI_WDOG_TIMER_NOT_INIT_RESP) 935 printk(KERN_INFO PFX "response: The IPMI controller appears" 936 " to have been reset, will attempt to reinitialize" 937 " the watchdog timer\n"); 938 else if (msg->msg.data[0] != 0) 939 printk(KERN_ERR PFX "response: Error %x on cmd %x\n", 940 msg->msg.data[0], 941 msg->msg.cmd); 942 943 ipmi_free_recv_msg(msg); 944 } 945 946 static void ipmi_wdog_pretimeout_handler(void *handler_data) 947 { 948 if (preaction_val != WDOG_PRETIMEOUT_NONE) { 949 if (preop_val == WDOG_PREOP_PANIC) { 950 if (atomic_inc_and_test(&preop_panic_excl)) 951 panic("Watchdog pre-timeout"); 952 } else if (preop_val == WDOG_PREOP_GIVE_DATA) { 953 spin_lock(&ipmi_read_lock); 954 data_to_read = 1; 955 wake_up_interruptible(&read_q); 956 kill_fasync(&fasync_q, SIGIO, POLL_IN); 957 958 spin_unlock(&ipmi_read_lock); 959 } 960 } 961 962 /* 963 * On some machines, the heartbeat will give an error and not 964 * work unless we re-enable the timer. So do so. 965 */ 966 pretimeout_since_last_heartbeat = 1; 967 } 968 969 static const struct ipmi_user_hndl ipmi_hndlrs = { 970 .ipmi_recv_hndl = ipmi_wdog_msg_handler, 971 .ipmi_watchdog_pretimeout = ipmi_wdog_pretimeout_handler 972 }; 973 974 static void ipmi_register_watchdog(int ipmi_intf) 975 { 976 int rv = -EBUSY; 977 978 if (watchdog_user) 979 goto out; 980 981 if ((ifnum_to_use >= 0) && (ifnum_to_use != ipmi_intf)) 982 goto out; 983 984 watchdog_ifnum = ipmi_intf; 985 986 rv = ipmi_create_user(ipmi_intf, &ipmi_hndlrs, NULL, &watchdog_user); 987 if (rv < 0) { 988 printk(KERN_CRIT PFX "Unable to register with ipmi\n"); 989 goto out; 990 } 991 992 rv = ipmi_get_version(watchdog_user, 993 &ipmi_version_major, 994 &ipmi_version_minor); 995 if (rv) { 996 pr_warn(PFX "Unable to get IPMI version, assuming 1.0\n"); 997 ipmi_version_major = 1; 998 ipmi_version_minor = 0; 999 } 1000 1001 rv = misc_register(&ipmi_wdog_miscdev); 1002 if (rv < 0) { 1003 ipmi_destroy_user(watchdog_user); 1004 watchdog_user = NULL; 1005 printk(KERN_CRIT PFX "Unable to register misc device\n"); 1006 } 1007 1008 #ifdef HAVE_DIE_NMI 1009 if (nmi_handler_registered) { 1010 int old_pretimeout = pretimeout; 1011 int old_timeout = timeout; 1012 int old_preop_val = preop_val; 1013 1014 /* 1015 * Set the pretimeout to go off in a second and give 1016 * ourselves plenty of time to stop the timer. 1017 */ 1018 ipmi_watchdog_state = WDOG_TIMEOUT_RESET; 1019 preop_val = WDOG_PREOP_NONE; /* Make sure nothing happens */ 1020 pretimeout = 99; 1021 timeout = 100; 1022 1023 testing_nmi = 1; 1024 1025 rv = ipmi_set_timeout(IPMI_SET_TIMEOUT_FORCE_HB); 1026 if (rv) { 1027 printk(KERN_WARNING PFX "Error starting timer to" 1028 " test NMI: 0x%x. The NMI pretimeout will" 1029 " likely not work\n", rv); 1030 rv = 0; 1031 goto out_restore; 1032 } 1033 1034 msleep(1500); 1035 1036 if (testing_nmi != 2) { 1037 printk(KERN_WARNING PFX "IPMI NMI didn't seem to" 1038 " occur. The NMI pretimeout will" 1039 " likely not work\n"); 1040 } 1041 out_restore: 1042 testing_nmi = 0; 1043 preop_val = old_preop_val; 1044 pretimeout = old_pretimeout; 1045 timeout = old_timeout; 1046 } 1047 #endif 1048 1049 out: 1050 if ((start_now) && (rv == 0)) { 1051 /* Run from startup, so start the timer now. */ 1052 start_now = 0; /* Disable this function after first startup. */ 1053 ipmi_watchdog_state = action_val; 1054 ipmi_set_timeout(IPMI_SET_TIMEOUT_FORCE_HB); 1055 printk(KERN_INFO PFX "Starting now!\n"); 1056 } else { 1057 /* Stop the timer now. */ 1058 ipmi_watchdog_state = WDOG_TIMEOUT_NONE; 1059 ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 1060 } 1061 } 1062 1063 static void ipmi_unregister_watchdog(int ipmi_intf) 1064 { 1065 int rv; 1066 1067 if (!watchdog_user) 1068 goto out; 1069 1070 if (watchdog_ifnum != ipmi_intf) 1071 goto out; 1072 1073 /* Make sure no one can call us any more. */ 1074 misc_deregister(&ipmi_wdog_miscdev); 1075 1076 /* 1077 * Wait to make sure the message makes it out. The lower layer has 1078 * pointers to our buffers, we want to make sure they are done before 1079 * we release our memory. 1080 */ 1081 while (atomic_read(&set_timeout_tofree)) 1082 schedule_timeout_uninterruptible(1); 1083 1084 /* Disconnect from IPMI. */ 1085 rv = ipmi_destroy_user(watchdog_user); 1086 if (rv) { 1087 printk(KERN_WARNING PFX "error unlinking from IPMI: %d\n", 1088 rv); 1089 } 1090 watchdog_user = NULL; 1091 1092 out: 1093 return; 1094 } 1095 1096 #ifdef HAVE_DIE_NMI 1097 static int 1098 ipmi_nmi(unsigned int val, struct pt_regs *regs) 1099 { 1100 /* 1101 * If we get here, it's an NMI that's not a memory or I/O 1102 * error. We can't truly tell if it's from IPMI or not 1103 * without sending a message, and sending a message is almost 1104 * impossible because of locking. 1105 */ 1106 1107 if (testing_nmi) { 1108 testing_nmi = 2; 1109 return NMI_HANDLED; 1110 } 1111 1112 /* If we are not expecting a timeout, ignore it. */ 1113 if (ipmi_watchdog_state == WDOG_TIMEOUT_NONE) 1114 return NMI_DONE; 1115 1116 if (preaction_val != WDOG_PRETIMEOUT_NMI) 1117 return NMI_DONE; 1118 1119 /* 1120 * If no one else handled the NMI, we assume it was the IPMI 1121 * watchdog. 1122 */ 1123 if (preop_val == WDOG_PREOP_PANIC) { 1124 /* On some machines, the heartbeat will give 1125 an error and not work unless we re-enable 1126 the timer. So do so. */ 1127 pretimeout_since_last_heartbeat = 1; 1128 if (atomic_inc_and_test(&preop_panic_excl)) 1129 nmi_panic(regs, PFX "pre-timeout"); 1130 } 1131 1132 return NMI_HANDLED; 1133 } 1134 #endif 1135 1136 static int wdog_reboot_handler(struct notifier_block *this, 1137 unsigned long code, 1138 void *unused) 1139 { 1140 static int reboot_event_handled; 1141 1142 if ((watchdog_user) && (!reboot_event_handled)) { 1143 /* Make sure we only do this once. */ 1144 reboot_event_handled = 1; 1145 1146 if (code == SYS_POWER_OFF || code == SYS_HALT) { 1147 /* Disable the WDT if we are shutting down. */ 1148 ipmi_watchdog_state = WDOG_TIMEOUT_NONE; 1149 ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 1150 } else if (ipmi_watchdog_state != WDOG_TIMEOUT_NONE) { 1151 /* Set a long timer to let the reboot happen or 1152 reset if it hangs, but only if the watchdog 1153 timer was already running. */ 1154 if (timeout < 120) 1155 timeout = 120; 1156 pretimeout = 0; 1157 ipmi_watchdog_state = WDOG_TIMEOUT_RESET; 1158 ipmi_set_timeout(IPMI_SET_TIMEOUT_NO_HB); 1159 } 1160 } 1161 return NOTIFY_OK; 1162 } 1163 1164 static struct notifier_block wdog_reboot_notifier = { 1165 .notifier_call = wdog_reboot_handler, 1166 .next = NULL, 1167 .priority = 0 1168 }; 1169 1170 static int wdog_panic_handler(struct notifier_block *this, 1171 unsigned long event, 1172 void *unused) 1173 { 1174 static int panic_event_handled; 1175 1176 /* On a panic, if we have a panic timeout, make sure to extend 1177 the watchdog timer to a reasonable value to complete the 1178 panic, if the watchdog timer is running. Plus the 1179 pretimeout is meaningless at panic time. */ 1180 if (watchdog_user && !panic_event_handled && 1181 ipmi_watchdog_state != WDOG_TIMEOUT_NONE) { 1182 /* Make sure we do this only once. */ 1183 panic_event_handled = 1; 1184 1185 timeout = panic_wdt_timeout; 1186 pretimeout = 0; 1187 panic_halt_ipmi_set_timeout(); 1188 } 1189 1190 return NOTIFY_OK; 1191 } 1192 1193 static struct notifier_block wdog_panic_notifier = { 1194 .notifier_call = wdog_panic_handler, 1195 .next = NULL, 1196 .priority = 150 /* priority: INT_MAX >= x >= 0 */ 1197 }; 1198 1199 1200 static void ipmi_new_smi(int if_num, struct device *device) 1201 { 1202 ipmi_register_watchdog(if_num); 1203 } 1204 1205 static void ipmi_smi_gone(int if_num) 1206 { 1207 ipmi_unregister_watchdog(if_num); 1208 } 1209 1210 static struct ipmi_smi_watcher smi_watcher = { 1211 .owner = THIS_MODULE, 1212 .new_smi = ipmi_new_smi, 1213 .smi_gone = ipmi_smi_gone 1214 }; 1215 1216 static int action_op(const char *inval, char *outval) 1217 { 1218 if (outval) 1219 strcpy(outval, action); 1220 1221 if (!inval) 1222 return 0; 1223 1224 if (strcmp(inval, "reset") == 0) 1225 action_val = WDOG_TIMEOUT_RESET; 1226 else if (strcmp(inval, "none") == 0) 1227 action_val = WDOG_TIMEOUT_NONE; 1228 else if (strcmp(inval, "power_cycle") == 0) 1229 action_val = WDOG_TIMEOUT_POWER_CYCLE; 1230 else if (strcmp(inval, "power_off") == 0) 1231 action_val = WDOG_TIMEOUT_POWER_DOWN; 1232 else 1233 return -EINVAL; 1234 strcpy(action, inval); 1235 return 0; 1236 } 1237 1238 static int preaction_op(const char *inval, char *outval) 1239 { 1240 if (outval) 1241 strcpy(outval, preaction); 1242 1243 if (!inval) 1244 return 0; 1245 1246 if (strcmp(inval, "pre_none") == 0) 1247 preaction_val = WDOG_PRETIMEOUT_NONE; 1248 else if (strcmp(inval, "pre_smi") == 0) 1249 preaction_val = WDOG_PRETIMEOUT_SMI; 1250 #ifdef HAVE_DIE_NMI 1251 else if (strcmp(inval, "pre_nmi") == 0) 1252 preaction_val = WDOG_PRETIMEOUT_NMI; 1253 #endif 1254 else if (strcmp(inval, "pre_int") == 0) 1255 preaction_val = WDOG_PRETIMEOUT_MSG_INT; 1256 else 1257 return -EINVAL; 1258 strcpy(preaction, inval); 1259 return 0; 1260 } 1261 1262 static int preop_op(const char *inval, char *outval) 1263 { 1264 if (outval) 1265 strcpy(outval, preop); 1266 1267 if (!inval) 1268 return 0; 1269 1270 if (strcmp(inval, "preop_none") == 0) 1271 preop_val = WDOG_PREOP_NONE; 1272 else if (strcmp(inval, "preop_panic") == 0) 1273 preop_val = WDOG_PREOP_PANIC; 1274 else if (strcmp(inval, "preop_give_data") == 0) 1275 preop_val = WDOG_PREOP_GIVE_DATA; 1276 else 1277 return -EINVAL; 1278 strcpy(preop, inval); 1279 return 0; 1280 } 1281 1282 static void check_parms(void) 1283 { 1284 #ifdef HAVE_DIE_NMI 1285 int do_nmi = 0; 1286 int rv; 1287 1288 if (preaction_val == WDOG_PRETIMEOUT_NMI) { 1289 do_nmi = 1; 1290 if (preop_val == WDOG_PREOP_GIVE_DATA) { 1291 printk(KERN_WARNING PFX "Pretimeout op is to give data" 1292 " but NMI pretimeout is enabled, setting" 1293 " pretimeout op to none\n"); 1294 preop_op("preop_none", NULL); 1295 do_nmi = 0; 1296 } 1297 } 1298 if (do_nmi && !nmi_handler_registered) { 1299 rv = register_nmi_handler(NMI_UNKNOWN, ipmi_nmi, 0, 1300 "ipmi"); 1301 if (rv) { 1302 printk(KERN_WARNING PFX 1303 "Can't register nmi handler\n"); 1304 return; 1305 } else 1306 nmi_handler_registered = 1; 1307 } else if (!do_nmi && nmi_handler_registered) { 1308 unregister_nmi_handler(NMI_UNKNOWN, "ipmi"); 1309 nmi_handler_registered = 0; 1310 } 1311 #endif 1312 } 1313 1314 static int __init ipmi_wdog_init(void) 1315 { 1316 int rv; 1317 1318 if (action_op(action, NULL)) { 1319 action_op("reset", NULL); 1320 printk(KERN_INFO PFX "Unknown action '%s', defaulting to" 1321 " reset\n", action); 1322 } 1323 1324 if (preaction_op(preaction, NULL)) { 1325 preaction_op("pre_none", NULL); 1326 printk(KERN_INFO PFX "Unknown preaction '%s', defaulting to" 1327 " none\n", preaction); 1328 } 1329 1330 if (preop_op(preop, NULL)) { 1331 preop_op("preop_none", NULL); 1332 printk(KERN_INFO PFX "Unknown preop '%s', defaulting to" 1333 " none\n", preop); 1334 } 1335 1336 check_parms(); 1337 1338 register_reboot_notifier(&wdog_reboot_notifier); 1339 atomic_notifier_chain_register(&panic_notifier_list, 1340 &wdog_panic_notifier); 1341 1342 rv = ipmi_smi_watcher_register(&smi_watcher); 1343 if (rv) { 1344 #ifdef HAVE_DIE_NMI 1345 if (nmi_handler_registered) 1346 unregister_nmi_handler(NMI_UNKNOWN, "ipmi"); 1347 #endif 1348 atomic_notifier_chain_unregister(&panic_notifier_list, 1349 &wdog_panic_notifier); 1350 unregister_reboot_notifier(&wdog_reboot_notifier); 1351 printk(KERN_WARNING PFX "can't register smi watcher\n"); 1352 return rv; 1353 } 1354 1355 printk(KERN_INFO PFX "driver initialized\n"); 1356 1357 return 0; 1358 } 1359 1360 static void __exit ipmi_wdog_exit(void) 1361 { 1362 ipmi_smi_watcher_unregister(&smi_watcher); 1363 ipmi_unregister_watchdog(watchdog_ifnum); 1364 1365 #ifdef HAVE_DIE_NMI 1366 if (nmi_handler_registered) 1367 unregister_nmi_handler(NMI_UNKNOWN, "ipmi"); 1368 #endif 1369 1370 atomic_notifier_chain_unregister(&panic_notifier_list, 1371 &wdog_panic_notifier); 1372 unregister_reboot_notifier(&wdog_reboot_notifier); 1373 } 1374 module_exit(ipmi_wdog_exit); 1375 module_init(ipmi_wdog_init); 1376 MODULE_LICENSE("GPL"); 1377 MODULE_AUTHOR("Corey Minyard <minyard@mvista.com>"); 1378 MODULE_DESCRIPTION("watchdog timer based upon the IPMI interface."); 1379