1 /* 2 * Octeon Watchdog driver 3 * 4 * Copyright (C) 2007, 2008, 2009, 2010 Cavium Networks 5 * 6 * Converted to use WATCHDOG_CORE by Aaro Koskinen <aaro.koskinen@iki.fi>. 7 * 8 * Some parts derived from wdt.c 9 * 10 * (c) Copyright 1996-1997 Alan Cox <alan@lxorguk.ukuu.org.uk>, 11 * All Rights Reserved. 12 * 13 * This program is free software; you can redistribute it and/or 14 * modify it under the terms of the GNU General Public License 15 * as published by the Free Software Foundation; either version 16 * 2 of the License, or (at your option) any later version. 17 * 18 * Neither Alan Cox nor CymruNet Ltd. admit liability nor provide 19 * warranty for any of this software. This material is provided 20 * "AS-IS" and at no charge. 21 * 22 * (c) Copyright 1995 Alan Cox <alan@lxorguk.ukuu.org.uk> 23 * 24 * This file is subject to the terms and conditions of the GNU General Public 25 * License. See the file "COPYING" in the main directory of this archive 26 * for more details. 27 * 28 * 29 * The OCTEON watchdog has a maximum timeout of 2^32 * io_clock. 30 * For most systems this is less than 10 seconds, so to allow for 31 * software to request longer watchdog heartbeats, we maintain software 32 * counters to count multiples of the base rate. If the system locks 33 * up in such a manner that we can not run the software counters, the 34 * only result is a watchdog reset sooner than was requested. But 35 * that is OK, because in this case userspace would likely not be able 36 * to do anything anyhow. 37 * 38 * The hardware watchdog interval we call the period. The OCTEON 39 * watchdog goes through several stages, after the first period an 40 * irq is asserted, then if it is not reset, after the next period NMI 41 * is asserted, then after an additional period a chip wide soft reset. 42 * So for the software counters, we reset watchdog after each period 43 * and decrement the counter. But for the last two periods we need to 44 * let the watchdog progress to the NMI stage so we disable the irq 45 * and let it proceed. Once in the NMI, we print the register state 46 * to the serial port and then wait for the reset. 47 * 48 * A watchdog is maintained for each CPU in the system, that way if 49 * one CPU suffers a lockup, we also get a register dump and reset. 50 * The userspace ping resets the watchdog on all CPUs. 51 * 52 * Before userspace opens the watchdog device, we still run the 53 * watchdogs to catch any lockups that may be kernel related. 54 * 55 */ 56 57 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 58 59 #include <linux/miscdevice.h> 60 #include <linux/interrupt.h> 61 #include <linux/watchdog.h> 62 #include <linux/cpumask.h> 63 #include <linux/bitops.h> 64 #include <linux/kernel.h> 65 #include <linux/module.h> 66 #include <linux/string.h> 67 #include <linux/delay.h> 68 #include <linux/cpu.h> 69 #include <linux/smp.h> 70 #include <linux/fs.h> 71 #include <linux/irq.h> 72 73 #include <asm/mipsregs.h> 74 #include <asm/uasm.h> 75 76 #include <asm/octeon/octeon.h> 77 78 /* The count needed to achieve timeout_sec. */ 79 static unsigned int timeout_cnt; 80 81 /* The maximum period supported. */ 82 static unsigned int max_timeout_sec; 83 84 /* The current period. */ 85 static unsigned int timeout_sec; 86 87 /* Set to non-zero when userspace countdown mode active */ 88 static int do_coundown; 89 static unsigned int countdown_reset; 90 static unsigned int per_cpu_countdown[NR_CPUS]; 91 92 static cpumask_t irq_enabled_cpus; 93 94 #define WD_TIMO 60 /* Default heartbeat = 60 seconds */ 95 96 static int heartbeat = WD_TIMO; 97 module_param(heartbeat, int, S_IRUGO); 98 MODULE_PARM_DESC(heartbeat, 99 "Watchdog heartbeat in seconds. (0 < heartbeat, default=" 100 __MODULE_STRING(WD_TIMO) ")"); 101 102 static bool nowayout = WATCHDOG_NOWAYOUT; 103 module_param(nowayout, bool, S_IRUGO); 104 MODULE_PARM_DESC(nowayout, 105 "Watchdog cannot be stopped once started (default=" 106 __MODULE_STRING(WATCHDOG_NOWAYOUT) ")"); 107 108 static u32 nmi_stage1_insns[64] __initdata; 109 /* We need one branch and therefore one relocation per target label. */ 110 static struct uasm_label labels[5] __initdata; 111 static struct uasm_reloc relocs[5] __initdata; 112 113 enum lable_id { 114 label_enter_bootloader = 1 115 }; 116 117 /* Some CP0 registers */ 118 #define K0 26 119 #define C0_CVMMEMCTL 11, 7 120 #define C0_STATUS 12, 0 121 #define C0_EBASE 15, 1 122 #define C0_DESAVE 31, 0 123 124 void octeon_wdt_nmi_stage2(void); 125 126 static void __init octeon_wdt_build_stage1(void) 127 { 128 int i; 129 int len; 130 u32 *p = nmi_stage1_insns; 131 #ifdef CONFIG_HOTPLUG_CPU 132 struct uasm_label *l = labels; 133 struct uasm_reloc *r = relocs; 134 #endif 135 136 /* 137 * For the next few instructions running the debugger may 138 * cause corruption of k0 in the saved registers. Since we're 139 * about to crash, nobody probably cares. 140 * 141 * Save K0 into the debug scratch register 142 */ 143 uasm_i_dmtc0(&p, K0, C0_DESAVE); 144 145 uasm_i_mfc0(&p, K0, C0_STATUS); 146 #ifdef CONFIG_HOTPLUG_CPU 147 if (octeon_bootloader_entry_addr) 148 uasm_il_bbit0(&p, &r, K0, ilog2(ST0_NMI), 149 label_enter_bootloader); 150 #endif 151 /* Force 64-bit addressing enabled */ 152 uasm_i_ori(&p, K0, K0, ST0_UX | ST0_SX | ST0_KX); 153 uasm_i_mtc0(&p, K0, C0_STATUS); 154 155 #ifdef CONFIG_HOTPLUG_CPU 156 if (octeon_bootloader_entry_addr) { 157 uasm_i_mfc0(&p, K0, C0_EBASE); 158 /* Coreid number in K0 */ 159 uasm_i_andi(&p, K0, K0, 0xf); 160 /* 8 * coreid in bits 16-31 */ 161 uasm_i_dsll_safe(&p, K0, K0, 3 + 16); 162 uasm_i_ori(&p, K0, K0, 0x8001); 163 uasm_i_dsll_safe(&p, K0, K0, 16); 164 uasm_i_ori(&p, K0, K0, 0x0700); 165 uasm_i_drotr_safe(&p, K0, K0, 32); 166 /* 167 * Should result in: 0x8001,0700,0000,8*coreid which is 168 * CVMX_CIU_WDOGX(coreid) - 0x0500 169 * 170 * Now ld K0, CVMX_CIU_WDOGX(coreid) 171 */ 172 uasm_i_ld(&p, K0, 0x500, K0); 173 /* 174 * If bit one set handle the NMI as a watchdog event. 175 * otherwise transfer control to bootloader. 176 */ 177 uasm_il_bbit0(&p, &r, K0, 1, label_enter_bootloader); 178 uasm_i_nop(&p); 179 } 180 #endif 181 182 /* Clear Dcache so cvmseg works right. */ 183 uasm_i_cache(&p, 1, 0, 0); 184 185 /* Use K0 to do a read/modify/write of CVMMEMCTL */ 186 uasm_i_dmfc0(&p, K0, C0_CVMMEMCTL); 187 /* Clear out the size of CVMSEG */ 188 uasm_i_dins(&p, K0, 0, 0, 6); 189 /* Set CVMSEG to its largest value */ 190 uasm_i_ori(&p, K0, K0, 0x1c0 | 54); 191 /* Store the CVMMEMCTL value */ 192 uasm_i_dmtc0(&p, K0, C0_CVMMEMCTL); 193 194 /* Load the address of the second stage handler */ 195 UASM_i_LA(&p, K0, (long)octeon_wdt_nmi_stage2); 196 uasm_i_jr(&p, K0); 197 uasm_i_dmfc0(&p, K0, C0_DESAVE); 198 199 #ifdef CONFIG_HOTPLUG_CPU 200 if (octeon_bootloader_entry_addr) { 201 uasm_build_label(&l, p, label_enter_bootloader); 202 /* Jump to the bootloader and restore K0 */ 203 UASM_i_LA(&p, K0, (long)octeon_bootloader_entry_addr); 204 uasm_i_jr(&p, K0); 205 uasm_i_dmfc0(&p, K0, C0_DESAVE); 206 } 207 #endif 208 uasm_resolve_relocs(relocs, labels); 209 210 len = (int)(p - nmi_stage1_insns); 211 pr_debug("Synthesized NMI stage 1 handler (%d instructions)\n", len); 212 213 pr_debug("\t.set push\n"); 214 pr_debug("\t.set noreorder\n"); 215 for (i = 0; i < len; i++) 216 pr_debug("\t.word 0x%08x\n", nmi_stage1_insns[i]); 217 pr_debug("\t.set pop\n"); 218 219 if (len > 32) 220 panic("NMI stage 1 handler exceeds 32 instructions, was %d\n", 221 len); 222 } 223 224 static int cpu2core(int cpu) 225 { 226 #ifdef CONFIG_SMP 227 return cpu_logical_map(cpu); 228 #else 229 return cvmx_get_core_num(); 230 #endif 231 } 232 233 static int core2cpu(int coreid) 234 { 235 #ifdef CONFIG_SMP 236 return cpu_number_map(coreid); 237 #else 238 return 0; 239 #endif 240 } 241 242 /** 243 * Poke the watchdog when an interrupt is received 244 * 245 * @cpl: 246 * @dev_id: 247 * 248 * Returns 249 */ 250 static irqreturn_t octeon_wdt_poke_irq(int cpl, void *dev_id) 251 { 252 unsigned int core = cvmx_get_core_num(); 253 int cpu = core2cpu(core); 254 255 if (do_coundown) { 256 if (per_cpu_countdown[cpu] > 0) { 257 /* We're alive, poke the watchdog */ 258 cvmx_write_csr(CVMX_CIU_PP_POKEX(core), 1); 259 per_cpu_countdown[cpu]--; 260 } else { 261 /* Bad news, you are about to reboot. */ 262 disable_irq_nosync(cpl); 263 cpumask_clear_cpu(cpu, &irq_enabled_cpus); 264 } 265 } else { 266 /* Not open, just ping away... */ 267 cvmx_write_csr(CVMX_CIU_PP_POKEX(core), 1); 268 } 269 return IRQ_HANDLED; 270 } 271 272 /* From setup.c */ 273 extern int prom_putchar(char c); 274 275 /** 276 * Write a string to the uart 277 * 278 * @str: String to write 279 */ 280 static void octeon_wdt_write_string(const char *str) 281 { 282 /* Just loop writing one byte at a time */ 283 while (*str) 284 prom_putchar(*str++); 285 } 286 287 /** 288 * Write a hex number out of the uart 289 * 290 * @value: Number to display 291 * @digits: Number of digits to print (1 to 16) 292 */ 293 static void octeon_wdt_write_hex(u64 value, int digits) 294 { 295 int d; 296 int v; 297 298 for (d = 0; d < digits; d++) { 299 v = (value >> ((digits - d - 1) * 4)) & 0xf; 300 if (v >= 10) 301 prom_putchar('a' + v - 10); 302 else 303 prom_putchar('0' + v); 304 } 305 } 306 307 static const char reg_name[][3] = { 308 "$0", "at", "v0", "v1", "a0", "a1", "a2", "a3", 309 "a4", "a5", "a6", "a7", "t0", "t1", "t2", "t3", 310 "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7", 311 "t8", "t9", "k0", "k1", "gp", "sp", "s8", "ra" 312 }; 313 314 /** 315 * NMI stage 3 handler. NMIs are handled in the following manner: 316 * 1) The first NMI handler enables CVMSEG and transfers from 317 * the bootbus region into normal memory. It is careful to not 318 * destroy any registers. 319 * 2) The second stage handler uses CVMSEG to save the registers 320 * and create a stack for C code. It then calls the third level 321 * handler with one argument, a pointer to the register values. 322 * 3) The third, and final, level handler is the following C 323 * function that prints out some useful infomration. 324 * 325 * @reg: Pointer to register state before the NMI 326 */ 327 void octeon_wdt_nmi_stage3(u64 reg[32]) 328 { 329 u64 i; 330 331 unsigned int coreid = cvmx_get_core_num(); 332 /* 333 * Save status and cause early to get them before any changes 334 * might happen. 335 */ 336 u64 cp0_cause = read_c0_cause(); 337 u64 cp0_status = read_c0_status(); 338 u64 cp0_error_epc = read_c0_errorepc(); 339 u64 cp0_epc = read_c0_epc(); 340 341 /* Delay so output from all cores output is not jumbled together. */ 342 __delay(100000000ull * coreid); 343 344 octeon_wdt_write_string("\r\n*** NMI Watchdog interrupt on Core 0x"); 345 octeon_wdt_write_hex(coreid, 1); 346 octeon_wdt_write_string(" ***\r\n"); 347 for (i = 0; i < 32; i++) { 348 octeon_wdt_write_string("\t"); 349 octeon_wdt_write_string(reg_name[i]); 350 octeon_wdt_write_string("\t0x"); 351 octeon_wdt_write_hex(reg[i], 16); 352 if (i & 1) 353 octeon_wdt_write_string("\r\n"); 354 } 355 octeon_wdt_write_string("\terr_epc\t0x"); 356 octeon_wdt_write_hex(cp0_error_epc, 16); 357 358 octeon_wdt_write_string("\tepc\t0x"); 359 octeon_wdt_write_hex(cp0_epc, 16); 360 octeon_wdt_write_string("\r\n"); 361 362 octeon_wdt_write_string("\tstatus\t0x"); 363 octeon_wdt_write_hex(cp0_status, 16); 364 octeon_wdt_write_string("\tcause\t0x"); 365 octeon_wdt_write_hex(cp0_cause, 16); 366 octeon_wdt_write_string("\r\n"); 367 368 octeon_wdt_write_string("\tsum0\t0x"); 369 octeon_wdt_write_hex(cvmx_read_csr(CVMX_CIU_INTX_SUM0(coreid * 2)), 16); 370 octeon_wdt_write_string("\ten0\t0x"); 371 octeon_wdt_write_hex(cvmx_read_csr(CVMX_CIU_INTX_EN0(coreid * 2)), 16); 372 octeon_wdt_write_string("\r\n"); 373 374 octeon_wdt_write_string("*** Chip soft reset soon ***\r\n"); 375 } 376 377 static void octeon_wdt_disable_interrupt(int cpu) 378 { 379 unsigned int core; 380 unsigned int irq; 381 union cvmx_ciu_wdogx ciu_wdog; 382 383 core = cpu2core(cpu); 384 385 irq = OCTEON_IRQ_WDOG0 + core; 386 387 /* Poke the watchdog to clear out its state */ 388 cvmx_write_csr(CVMX_CIU_PP_POKEX(core), 1); 389 390 /* Disable the hardware. */ 391 ciu_wdog.u64 = 0; 392 cvmx_write_csr(CVMX_CIU_WDOGX(core), ciu_wdog.u64); 393 394 free_irq(irq, octeon_wdt_poke_irq); 395 } 396 397 static void octeon_wdt_setup_interrupt(int cpu) 398 { 399 unsigned int core; 400 unsigned int irq; 401 union cvmx_ciu_wdogx ciu_wdog; 402 403 core = cpu2core(cpu); 404 405 /* Disable it before doing anything with the interrupts. */ 406 ciu_wdog.u64 = 0; 407 cvmx_write_csr(CVMX_CIU_WDOGX(core), ciu_wdog.u64); 408 409 per_cpu_countdown[cpu] = countdown_reset; 410 411 irq = OCTEON_IRQ_WDOG0 + core; 412 413 if (request_irq(irq, octeon_wdt_poke_irq, 414 IRQF_NO_THREAD, "octeon_wdt", octeon_wdt_poke_irq)) 415 panic("octeon_wdt: Couldn't obtain irq %d", irq); 416 417 cpumask_set_cpu(cpu, &irq_enabled_cpus); 418 419 /* Poke the watchdog to clear out its state */ 420 cvmx_write_csr(CVMX_CIU_PP_POKEX(core), 1); 421 422 /* Finally enable the watchdog now that all handlers are installed */ 423 ciu_wdog.u64 = 0; 424 ciu_wdog.s.len = timeout_cnt; 425 ciu_wdog.s.mode = 3; /* 3 = Interrupt + NMI + Soft-Reset */ 426 cvmx_write_csr(CVMX_CIU_WDOGX(core), ciu_wdog.u64); 427 } 428 429 static int octeon_wdt_cpu_callback(struct notifier_block *nfb, 430 unsigned long action, void *hcpu) 431 { 432 unsigned int cpu = (unsigned long)hcpu; 433 434 switch (action) { 435 case CPU_DOWN_PREPARE: 436 octeon_wdt_disable_interrupt(cpu); 437 break; 438 case CPU_ONLINE: 439 case CPU_DOWN_FAILED: 440 octeon_wdt_setup_interrupt(cpu); 441 break; 442 default: 443 break; 444 } 445 return NOTIFY_OK; 446 } 447 448 static int octeon_wdt_ping(struct watchdog_device __always_unused *wdog) 449 { 450 int cpu; 451 int coreid; 452 453 for_each_online_cpu(cpu) { 454 coreid = cpu2core(cpu); 455 cvmx_write_csr(CVMX_CIU_PP_POKEX(coreid), 1); 456 per_cpu_countdown[cpu] = countdown_reset; 457 if ((countdown_reset || !do_coundown) && 458 !cpumask_test_cpu(cpu, &irq_enabled_cpus)) { 459 /* We have to enable the irq */ 460 int irq = OCTEON_IRQ_WDOG0 + coreid; 461 462 enable_irq(irq); 463 cpumask_set_cpu(cpu, &irq_enabled_cpus); 464 } 465 } 466 return 0; 467 } 468 469 static void octeon_wdt_calc_parameters(int t) 470 { 471 unsigned int periods; 472 473 timeout_sec = max_timeout_sec; 474 475 476 /* 477 * Find the largest interrupt period, that can evenly divide 478 * the requested heartbeat time. 479 */ 480 while ((t % timeout_sec) != 0) 481 timeout_sec--; 482 483 periods = t / timeout_sec; 484 485 /* 486 * The last two periods are after the irq is disabled, and 487 * then to the nmi, so we subtract them off. 488 */ 489 490 countdown_reset = periods > 2 ? periods - 2 : 0; 491 heartbeat = t; 492 timeout_cnt = ((octeon_get_io_clock_rate() >> 8) * timeout_sec) >> 8; 493 } 494 495 static int octeon_wdt_set_timeout(struct watchdog_device *wdog, 496 unsigned int t) 497 { 498 int cpu; 499 int coreid; 500 union cvmx_ciu_wdogx ciu_wdog; 501 502 if (t <= 0) 503 return -1; 504 505 octeon_wdt_calc_parameters(t); 506 507 for_each_online_cpu(cpu) { 508 coreid = cpu2core(cpu); 509 cvmx_write_csr(CVMX_CIU_PP_POKEX(coreid), 1); 510 ciu_wdog.u64 = 0; 511 ciu_wdog.s.len = timeout_cnt; 512 ciu_wdog.s.mode = 3; /* 3 = Interrupt + NMI + Soft-Reset */ 513 cvmx_write_csr(CVMX_CIU_WDOGX(coreid), ciu_wdog.u64); 514 cvmx_write_csr(CVMX_CIU_PP_POKEX(coreid), 1); 515 } 516 octeon_wdt_ping(wdog); /* Get the irqs back on. */ 517 return 0; 518 } 519 520 static int octeon_wdt_start(struct watchdog_device *wdog) 521 { 522 octeon_wdt_ping(wdog); 523 do_coundown = 1; 524 return 0; 525 } 526 527 static int octeon_wdt_stop(struct watchdog_device *wdog) 528 { 529 do_coundown = 0; 530 octeon_wdt_ping(wdog); 531 return 0; 532 } 533 534 static struct notifier_block octeon_wdt_cpu_notifier = { 535 .notifier_call = octeon_wdt_cpu_callback, 536 }; 537 538 static const struct watchdog_info octeon_wdt_info = { 539 .options = WDIOF_SETTIMEOUT | WDIOF_MAGICCLOSE | WDIOF_KEEPALIVEPING, 540 .identity = "OCTEON", 541 }; 542 543 static const struct watchdog_ops octeon_wdt_ops = { 544 .owner = THIS_MODULE, 545 .start = octeon_wdt_start, 546 .stop = octeon_wdt_stop, 547 .ping = octeon_wdt_ping, 548 .set_timeout = octeon_wdt_set_timeout, 549 }; 550 551 static struct watchdog_device octeon_wdt = { 552 .info = &octeon_wdt_info, 553 .ops = &octeon_wdt_ops, 554 }; 555 556 /** 557 * Module/ driver initialization. 558 * 559 * Returns Zero on success 560 */ 561 static int __init octeon_wdt_init(void) 562 { 563 int i; 564 int ret; 565 int cpu; 566 u64 *ptr; 567 568 /* 569 * Watchdog time expiration length = The 16 bits of LEN 570 * represent the most significant bits of a 24 bit decrementer 571 * that decrements every 256 cycles. 572 * 573 * Try for a timeout of 5 sec, if that fails a smaller number 574 * of even seconds, 575 */ 576 max_timeout_sec = 6; 577 do { 578 max_timeout_sec--; 579 timeout_cnt = ((octeon_get_io_clock_rate() >> 8) * 580 max_timeout_sec) >> 8; 581 } while (timeout_cnt > 65535); 582 583 BUG_ON(timeout_cnt == 0); 584 585 octeon_wdt_calc_parameters(heartbeat); 586 587 pr_info("Initial granularity %d Sec\n", timeout_sec); 588 589 octeon_wdt.timeout = timeout_sec; 590 octeon_wdt.max_timeout = UINT_MAX; 591 592 watchdog_set_nowayout(&octeon_wdt, nowayout); 593 594 ret = watchdog_register_device(&octeon_wdt); 595 if (ret) { 596 pr_err("watchdog_register_device() failed: %d\n", ret); 597 return ret; 598 } 599 600 /* Build the NMI handler ... */ 601 octeon_wdt_build_stage1(); 602 603 /* ... and install it. */ 604 ptr = (u64 *) nmi_stage1_insns; 605 for (i = 0; i < 16; i++) { 606 cvmx_write_csr(CVMX_MIO_BOOT_LOC_ADR, i * 8); 607 cvmx_write_csr(CVMX_MIO_BOOT_LOC_DAT, ptr[i]); 608 } 609 cvmx_write_csr(CVMX_MIO_BOOT_LOC_CFGX(0), 0x81fc0000); 610 611 cpumask_clear(&irq_enabled_cpus); 612 613 cpu_notifier_register_begin(); 614 for_each_online_cpu(cpu) 615 octeon_wdt_setup_interrupt(cpu); 616 617 __register_hotcpu_notifier(&octeon_wdt_cpu_notifier); 618 cpu_notifier_register_done(); 619 620 return 0; 621 } 622 623 /** 624 * Module / driver shutdown 625 */ 626 static void __exit octeon_wdt_cleanup(void) 627 { 628 int cpu; 629 630 watchdog_unregister_device(&octeon_wdt); 631 632 cpu_notifier_register_begin(); 633 __unregister_hotcpu_notifier(&octeon_wdt_cpu_notifier); 634 635 for_each_online_cpu(cpu) { 636 int core = cpu2core(cpu); 637 /* Disable the watchdog */ 638 cvmx_write_csr(CVMX_CIU_WDOGX(core), 0); 639 /* Free the interrupt handler */ 640 free_irq(OCTEON_IRQ_WDOG0 + core, octeon_wdt_poke_irq); 641 } 642 643 cpu_notifier_register_done(); 644 645 /* 646 * Disable the boot-bus memory, the code it points to is soon 647 * to go missing. 648 */ 649 cvmx_write_csr(CVMX_MIO_BOOT_LOC_CFGX(0), 0); 650 } 651 652 MODULE_LICENSE("GPL"); 653 MODULE_AUTHOR("Cavium Networks <support@caviumnetworks.com>"); 654 MODULE_DESCRIPTION("Cavium Networks Octeon Watchdog driver."); 655 module_init(octeon_wdt_init); 656 module_exit(octeon_wdt_cleanup); 657