1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 /* 27 * When the operating system detects that it is in an invalid state, a panic 28 * is initiated in order to minimize potential damage to user data and to 29 * facilitate debugging. There are three major tasks to be performed in 30 * a system panic: recording information about the panic in memory (and thus 31 * making it part of the crash dump), synchronizing the file systems to 32 * preserve user file data, and generating the crash dump. We define the 33 * system to be in one of four states with respect to the panic code: 34 * 35 * CALM - the state of the system prior to any thread initiating a panic 36 * 37 * QUIESCE - the state of the system when the first thread to initiate 38 * a system panic records information about the cause of the panic 39 * and renders the system quiescent by stopping other processors 40 * 41 * SYNC - the state of the system when we synchronize the file systems 42 * DUMP - the state when we generate the crash dump. 43 * 44 * The transitions between these states are irreversible: once we begin 45 * panicking, we only make one attempt to perform the actions associated with 46 * each state. 47 * 48 * The panic code itself must be re-entrant because actions taken during any 49 * state may lead to another system panic. Additionally, any Solaris 50 * thread may initiate a panic at any time, and so we must have synchronization 51 * between threads which attempt to initiate a state transition simultaneously. 52 * The panic code makes use of a special locking primitive, a trigger, to 53 * perform this synchronization. A trigger is simply a word which is set 54 * atomically and can only be set once. We declare three triggers, one for 55 * each transition between the four states. When a thread enters the panic 56 * code it attempts to set each trigger; if it fails it moves on to the 57 * next trigger. A special case is the first trigger: if two threads race 58 * to perform the transition to QUIESCE, the losing thread may execute before 59 * the winner has a chance to stop its CPU. To solve this problem, we have 60 * the loser look ahead to see if any other triggers are set; if not, it 61 * presumes a panic is underway and simply spins. Unfortunately, since we 62 * are panicking, it is not possible to know this with absolute certainty. 63 * 64 * There are two common reasons for re-entering the panic code once a panic 65 * has been initiated: (1) after we debug_enter() at the end of QUIESCE, 66 * the operator may type "sync" instead of "go", and the PROM's sync callback 67 * routine will invoke panic(); (2) if the clock routine decides that sync 68 * or dump is not making progress, it will invoke panic() to force a timeout. 69 * The design assumes that a third possibility, another thread causing an 70 * unrelated panic while sync or dump is still underway, is extremely unlikely. 71 * If this situation occurs, we may end up triggering dump while sync is 72 * still in progress. This third case is considered extremely unlikely because 73 * all other CPUs are stopped and low-level interrupts have been blocked. 74 * 75 * The panic code is entered via a call directly to the vpanic() function, 76 * or its varargs wrappers panic() and cmn_err(9F). The vpanic routine 77 * is implemented in assembly language to record the current machine 78 * registers, attempt to set the trigger for the QUIESCE state, and 79 * if successful, switch stacks on to the panic_stack before calling into 80 * the common panicsys() routine. The first thread to initiate a panic 81 * is allowed to make use of the reserved panic_stack so that executing 82 * the panic code itself does not overwrite valuable data on that thread's 83 * stack *ahead* of the current stack pointer. This data will be preserved 84 * in the crash dump and may prove invaluable in determining what this 85 * thread has previously been doing. The first thread, saved in panic_thread, 86 * is also responsible for stopping the other CPUs as quickly as possible, 87 * and then setting the various panic_* variables. Most important among 88 * these is panicstr, which allows threads to subsequently bypass held 89 * locks so that we can proceed without ever blocking. We must stop the 90 * other CPUs *prior* to setting panicstr in case threads running there are 91 * currently spinning to acquire a lock; we want that state to be preserved. 92 * Every thread which initiates a panic has its T_PANIC flag set so we can 93 * identify all such threads in the crash dump. 94 * 95 * The panic_thread is also allowed to make use of the special memory buffer 96 * panicbuf, which on machines with appropriate hardware is preserved across 97 * reboots. We allow the panic_thread to store its register set and panic 98 * message in this buffer, so even if we fail to obtain a crash dump we will 99 * be able to examine the machine after reboot and determine some of the 100 * state at the time of the panic. If we do get a dump, the panic buffer 101 * data is structured so that a debugger can easily consume the information 102 * therein (see <sys/panic.h>). 103 * 104 * Each platform or architecture is required to implement the functions 105 * panic_savetrap() to record trap-specific information to panicbuf, 106 * panic_saveregs() to record a register set to panicbuf, panic_stopcpus() 107 * to halt all CPUs but the panicking CPU, panic_quiesce_hw() to perform 108 * miscellaneous platform-specific tasks *after* panicstr is set, 109 * panic_showtrap() to print trap-specific information to the console, 110 * and panic_dump_hw() to perform platform tasks prior to calling dumpsys(). 111 * 112 * A Note on Word Formation, courtesy of the Oxford Guide to English Usage: 113 * 114 * Words ending in -c interpose k before suffixes which otherwise would 115 * indicate a soft c, and thus the verb and adjective forms of 'panic' are 116 * spelled "panicked", "panicking", and "panicky" respectively. Use of 117 * the ill-conceived "panicing" and "panic'd" is discouraged. 118 */ 119 120 #include <sys/types.h> 121 #include <sys/varargs.h> 122 #include <sys/sysmacros.h> 123 #include <sys/cmn_err.h> 124 #include <sys/cpuvar.h> 125 #include <sys/thread.h> 126 #include <sys/t_lock.h> 127 #include <sys/cred.h> 128 #include <sys/systm.h> 129 #include <sys/archsystm.h> 130 #include <sys/uadmin.h> 131 #include <sys/callb.h> 132 #include <sys/vfs.h> 133 #include <sys/log.h> 134 #include <sys/disp.h> 135 #include <sys/param.h> 136 #include <sys/dumphdr.h> 137 #include <sys/ftrace.h> 138 #include <sys/reboot.h> 139 #include <sys/debug.h> 140 #include <sys/stack.h> 141 #include <sys/spl.h> 142 #include <sys/errorq.h> 143 #include <sys/panic.h> 144 #include <sys/fm/util.h> 145 #include <sys/clock_impl.h> 146 147 /* 148 * Panic variables which are set once during the QUIESCE state by the 149 * first thread to initiate a panic. These are examined by post-mortem 150 * debugging tools; the inconsistent use of 'panic' versus 'panic_' in 151 * the variable naming is historical and allows legacy tools to work. 152 */ 153 #pragma align STACK_ALIGN(panic_stack) 154 char panic_stack[PANICSTKSIZE]; /* reserved stack for panic_thread */ 155 kthread_t *panic_thread; /* first thread to call panicsys() */ 156 cpu_t panic_cpu; /* cpu from first call to panicsys() */ 157 label_t panic_regs; /* setjmp label from panic_thread */ 158 struct regs *panic_reg; /* regs struct from first panicsys() */ 159 char *volatile panicstr; /* format string to first panicsys() */ 160 va_list panicargs; /* arguments to first panicsys() */ 161 clock_t panic_lbolt; /* lbolt at time of panic */ 162 int64_t panic_lbolt64; /* lbolt64 at time of panic */ 163 hrtime_t panic_hrtime; /* hrtime at time of panic */ 164 timespec_t panic_hrestime; /* hrestime at time of panic */ 165 int panic_ipl; /* ipl on panic_cpu at time of panic */ 166 ushort_t panic_schedflag; /* t_schedflag for panic_thread */ 167 cpu_t *panic_bound_cpu; /* t_bound_cpu for panic_thread */ 168 char panic_preempt; /* t_preempt for panic_thread */ 169 170 /* 171 * Panic variables which can be set via /etc/system or patched while 172 * the system is in operation. Again, the stupid names are historic. 173 */ 174 char *panic_bootstr = NULL; /* mdboot string to use after panic */ 175 int panic_bootfcn = AD_BOOT; /* mdboot function to use after panic */ 176 int halt_on_panic = 0; /* halt after dump instead of reboot? */ 177 int nopanicdebug = 0; /* reboot instead of call debugger? */ 178 int in_sync = 0; /* skip vfs_syncall() and just dump? */ 179 180 /* 181 * The do_polled_io flag is set by the panic code to inform the SCSI subsystem 182 * to use polled mode instead of interrupt-driven i/o. 183 */ 184 int do_polled_io = 0; 185 186 /* 187 * The panic_forced flag is set by the uadmin A_DUMP code to inform the 188 * panic subsystem that it should not attempt an initial debug_enter. 189 */ 190 int panic_forced = 0; 191 192 /* 193 * Triggers for panic state transitions: 194 */ 195 int panic_quiesce; /* trigger for CALM -> QUIESCE */ 196 int panic_sync; /* trigger for QUIESCE -> SYNC */ 197 int panic_dump; /* trigger for SYNC -> DUMP */ 198 199 /* 200 * Variable signifying quiesce(9E) is in progress. 201 */ 202 volatile int quiesce_active = 0; 203 204 void 205 panicsys(const char *format, va_list alist, struct regs *rp, int on_panic_stack) 206 { 207 int s = spl8(); 208 kthread_t *t = curthread; 209 cpu_t *cp = CPU; 210 211 caddr_t intr_stack = NULL; 212 uint_t intr_actv; 213 214 ushort_t schedflag = t->t_schedflag; 215 cpu_t *bound_cpu = t->t_bound_cpu; 216 char preempt = t->t_preempt; 217 218 (void) setjmp(&t->t_pcb); 219 t->t_flag |= T_PANIC; 220 221 t->t_schedflag |= TS_DONT_SWAP; 222 t->t_bound_cpu = cp; 223 t->t_preempt++; 224 225 /* 226 * Switch lbolt to event driven mode. 227 */ 228 lbolt_hybrid = lbolt_event_driven; 229 230 panic_enter_hw(s); 231 232 /* 233 * If we're on the interrupt stack and an interrupt thread is available 234 * in this CPU's pool, preserve the interrupt stack by detaching an 235 * interrupt thread and making its stack the intr_stack. 236 */ 237 if (CPU_ON_INTR(cp) && cp->cpu_intr_thread != NULL) { 238 kthread_t *it = cp->cpu_intr_thread; 239 240 intr_stack = cp->cpu_intr_stack; 241 intr_actv = cp->cpu_intr_actv; 242 243 cp->cpu_intr_stack = thread_stk_init(it->t_stk); 244 cp->cpu_intr_thread = it->t_link; 245 246 /* 247 * Clear only the high level bits of cpu_intr_actv. 248 * We want to indicate that high-level interrupts are 249 * not active without destroying the low-level interrupt 250 * information stored there. 251 */ 252 cp->cpu_intr_actv &= ((1 << (LOCK_LEVEL + 1)) - 1); 253 } 254 255 /* 256 * Record one-time panic information and quiesce the other CPUs. 257 * Then print out the panic message and stack trace. 258 */ 259 if (on_panic_stack) { 260 panic_data_t *pdp = (panic_data_t *)panicbuf; 261 262 pdp->pd_version = PANICBUFVERS; 263 pdp->pd_msgoff = sizeof (panic_data_t) - sizeof (panic_nv_t); 264 265 if (t->t_panic_trap != NULL) 266 panic_savetrap(pdp, t->t_panic_trap); 267 else 268 panic_saveregs(pdp, rp); 269 270 (void) vsnprintf(&panicbuf[pdp->pd_msgoff], 271 PANICBUFSIZE - pdp->pd_msgoff, format, alist); 272 273 /* 274 * Call into the platform code to stop the other CPUs. 275 * We currently have all interrupts blocked, and expect that 276 * the platform code will lower ipl only as far as needed to 277 * perform cross-calls, and will acquire as *few* locks as is 278 * possible -- panicstr is not set so we can still deadlock. 279 */ 280 panic_stopcpus(cp, t, s); 281 282 panicstr = (char *)format; 283 va_copy(panicargs, alist); 284 panic_lbolt = (clock_t)LBOLT_NO_ACCOUNT; 285 panic_lbolt64 = LBOLT_NO_ACCOUNT; 286 panic_hrestime = hrestime; 287 panic_hrtime = gethrtime_waitfree(); 288 panic_thread = t; 289 panic_regs = t->t_pcb; 290 panic_reg = rp; 291 panic_cpu = *cp; 292 panic_ipl = spltoipl(s); 293 panic_schedflag = schedflag; 294 panic_bound_cpu = bound_cpu; 295 panic_preempt = preempt; 296 297 if (intr_stack != NULL) { 298 panic_cpu.cpu_intr_stack = intr_stack; 299 panic_cpu.cpu_intr_actv = intr_actv; 300 } 301 302 /* 303 * Lower ipl to 10 to keep clock() from running, but allow 304 * keyboard interrupts to enter the debugger. These callbacks 305 * are executed with panicstr set so they can bypass locks. 306 */ 307 splx(ipltospl(CLOCK_LEVEL)); 308 panic_quiesce_hw(pdp); 309 (void) FTRACE_STOP(); 310 (void) callb_execute_class(CB_CL_PANIC, NULL); 311 312 if (log_intrq != NULL) 313 log_flushq(log_intrq); 314 315 /* 316 * If log_consq has been initialized and syslogd has started, 317 * print any messages in log_consq that haven't been consumed. 318 */ 319 if (log_consq != NULL && log_consq != log_backlogq) 320 log_printq(log_consq); 321 322 fm_banner(); 323 324 #if defined(__x86) 325 /* 326 * A hypervisor panic originates outside of Solaris, so we 327 * don't want to prepend the panic message with misleading 328 * pointers from within Solaris. 329 */ 330 if (!IN_XPV_PANIC()) 331 #endif 332 printf("\n\rpanic[cpu%d]/thread=%p: ", cp->cpu_id, 333 (void *)t); 334 vprintf(format, alist); 335 printf("\n\n"); 336 337 if (t->t_panic_trap != NULL) { 338 panic_showtrap(t->t_panic_trap); 339 printf("\n"); 340 } 341 342 traceregs(rp); 343 printf("\n"); 344 345 if (((boothowto & RB_DEBUG) || obpdebug) && 346 !nopanicdebug && !panic_forced) { 347 if (dumpvp != NULL) { 348 debug_enter("panic: entering debugger " 349 "(continue to save dump)"); 350 } else { 351 debug_enter("panic: entering debugger " 352 "(no dump device, continue to reboot)"); 353 } 354 } 355 356 } else if (panic_dump != 0 || panic_sync != 0 || panicstr != NULL) { 357 printf("\n\rpanic[cpu%d]/thread=%p: ", cp->cpu_id, (void *)t); 358 vprintf(format, alist); 359 printf("\n"); 360 } else 361 goto spin; 362 363 /* 364 * Prior to performing sync or dump, we make sure that do_polled_io is 365 * set, but we'll leave ipl at 10; deadman(), a CY_HIGH_LEVEL cyclic, 366 * will re-enter panic if we are not making progress with sync or dump. 367 */ 368 369 /* 370 * Sync the filesystems. Reset t_cred if not set because much of 371 * the filesystem code depends on CRED() being valid. 372 */ 373 if (!in_sync && panic_trigger(&panic_sync)) { 374 if (t->t_cred == NULL) 375 t->t_cred = kcred; 376 splx(ipltospl(CLOCK_LEVEL)); 377 do_polled_io = 1; 378 vfs_syncall(); 379 } 380 381 /* 382 * Take the crash dump. If the dump trigger is already set, try to 383 * enter the debugger again before rebooting the system. 384 */ 385 if (panic_trigger(&panic_dump)) { 386 panic_dump_hw(s); 387 splx(ipltospl(CLOCK_LEVEL)); 388 errorq_panic(); 389 do_polled_io = 1; 390 dumpsys(); 391 } else if (((boothowto & RB_DEBUG) || obpdebug) && !nopanicdebug) { 392 debug_enter("panic: entering debugger (continue to reboot)"); 393 } else 394 printf("dump aborted: please record the above information!\n"); 395 396 if (halt_on_panic) 397 mdboot(A_REBOOT, AD_HALT, NULL, B_FALSE); 398 else 399 mdboot(A_REBOOT, panic_bootfcn, panic_bootstr, B_FALSE); 400 spin: 401 /* 402 * Restore ipl to at most CLOCK_LEVEL so we don't end up spinning 403 * and unable to jump into the debugger. 404 */ 405 splx(MIN(s, ipltospl(CLOCK_LEVEL))); 406 for (;;) 407 ; 408 } 409 410 void 411 panic(const char *format, ...) 412 { 413 va_list alist; 414 415 va_start(alist, format); 416 vpanic(format, alist); 417 va_end(alist); 418 } 419