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