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 2007 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 /* 29 * When the operating system detects that it is in an invalid state, a panic 30 * is initiated in order to minimize potential damage to user data and to 31 * facilitate debugging. There are three major tasks to be performed in 32 * a system panic: recording information about the panic in memory (and thus 33 * making it part of the crash dump), synchronizing the file systems to 34 * preserve user file data, and generating the crash dump. We define the 35 * system to be in one of four states with respect to the panic code: 36 * 37 * CALM - the state of the system prior to any thread initiating a panic 38 * 39 * QUIESCE - the state of the system when the first thread to initiate 40 * a system panic records information about the cause of the panic 41 * and renders the system quiescent by stopping other processors 42 * 43 * SYNC - the state of the system when we synchronize the file systems 44 * DUMP - the state when we generate the crash dump. 45 * 46 * The transitions between these states are irreversible: once we begin 47 * panicking, we only make one attempt to perform the actions associated with 48 * each state. 49 * 50 * The panic code itself must be re-entrant because actions taken during any 51 * state may lead to another system panic. Additionally, any Solaris 52 * thread may initiate a panic at any time, and so we must have synchronization 53 * between threads which attempt to initiate a state transition simultaneously. 54 * The panic code makes use of a special locking primitive, a trigger, to 55 * perform this synchronization. A trigger is simply a word which is set 56 * atomically and can only be set once. We declare three triggers, one for 57 * each transition between the four states. When a thread enters the panic 58 * code it attempts to set each trigger; if it fails it moves on to the 59 * next trigger. A special case is the first trigger: if two threads race 60 * to perform the transition to QUIESCE, the losing thread may execute before 61 * the winner has a chance to stop its CPU. To solve this problem, we have 62 * the loser look ahead to see if any other triggers are set; if not, it 63 * presumes a panic is underway and simply spins. Unfortunately, since we 64 * are panicking, it is not possible to know this with absolute certainty. 65 * 66 * There are two common reasons for re-entering the panic code once a panic 67 * has been initiated: (1) after we debug_enter() at the end of QUIESCE, 68 * the operator may type "sync" instead of "go", and the PROM's sync callback 69 * routine will invoke panic(); (2) if the clock routine decides that sync 70 * or dump is not making progress, it will invoke panic() to force a timeout. 71 * The design assumes that a third possibility, another thread causing an 72 * unrelated panic while sync or dump is still underway, is extremely unlikely. 73 * If this situation occurs, we may end up triggering dump while sync is 74 * still in progress. This third case is considered extremely unlikely because 75 * all other CPUs are stopped and low-level interrupts have been blocked. 76 * 77 * The panic code is entered via a call directly to the vpanic() function, 78 * or its varargs wrappers panic() and cmn_err(9F). The vpanic routine 79 * is implemented in assembly language to record the current machine 80 * registers, attempt to set the trigger for the QUIESCE state, and 81 * if successful, switch stacks on to the panic_stack before calling into 82 * the common panicsys() routine. The first thread to initiate a panic 83 * is allowed to make use of the reserved panic_stack so that executing 84 * the panic code itself does not overwrite valuable data on that thread's 85 * stack *ahead* of the current stack pointer. This data will be preserved 86 * in the crash dump and may prove invaluable in determining what this 87 * thread has previously been doing. The first thread, saved in panic_thread, 88 * is also responsible for stopping the other CPUs as quickly as possible, 89 * and then setting the various panic_* variables. Most important among 90 * these is panicstr, which allows threads to subsequently bypass held 91 * locks so that we can proceed without ever blocking. We must stop the 92 * other CPUs *prior* to setting panicstr in case threads running there are 93 * currently spinning to acquire a lock; we want that state to be preserved. 94 * Every thread which initiates a panic has its T_PANIC flag set so we can 95 * identify all such threads in the crash dump. 96 * 97 * The panic_thread is also allowed to make use of the special memory buffer 98 * panicbuf, which on machines with appropriate hardware is preserved across 99 * reboots. We allow the panic_thread to store its register set and panic 100 * message in this buffer, so even if we fail to obtain a crash dump we will 101 * be able to examine the machine after reboot and determine some of the 102 * state at the time of the panic. If we do get a dump, the panic buffer 103 * data is structured so that a debugger can easily consume the information 104 * therein (see <sys/panic.h>). 105 * 106 * Each platform or architecture is required to implement the functions 107 * panic_savetrap() to record trap-specific information to panicbuf, 108 * panic_saveregs() to record a register set to panicbuf, panic_stopcpus() 109 * to halt all CPUs but the panicking CPU, panic_quiesce_hw() to perform 110 * miscellaneous platform-specific tasks *after* panicstr is set, 111 * panic_showtrap() to print trap-specific information to the console, 112 * and panic_dump_hw() to perform platform tasks prior to calling dumpsys(). 113 * 114 * A Note on Word Formation, courtesy of the Oxford Guide to English Usage: 115 * 116 * Words ending in -c interpose k before suffixes which otherwise would 117 * indicate a soft c, and thus the verb and adjective forms of 'panic' are 118 * spelled "panicked", "panicking", and "panicky" respectively. Use of 119 * the ill-conceived "panicing" and "panic'd" is discouraged. 120 */ 121 122 #include <sys/types.h> 123 #include <sys/varargs.h> 124 #include <sys/sysmacros.h> 125 #include <sys/cmn_err.h> 126 #include <sys/cpuvar.h> 127 #include <sys/thread.h> 128 #include <sys/t_lock.h> 129 #include <sys/cred.h> 130 #include <sys/systm.h> 131 #include <sys/archsystm.h> 132 #include <sys/uadmin.h> 133 #include <sys/callb.h> 134 #include <sys/vfs.h> 135 #include <sys/log.h> 136 #include <sys/disp.h> 137 #include <sys/param.h> 138 #include <sys/dumphdr.h> 139 #include <sys/ftrace.h> 140 #include <sys/reboot.h> 141 #include <sys/debug.h> 142 #include <sys/stack.h> 143 #include <sys/spl.h> 144 #include <sys/errorq.h> 145 #include <sys/panic.h> 146 #include <sys/fm/util.h> 147 148 /* 149 * Panic variables which are set once during the QUIESCE state by the 150 * first thread to initiate a panic. These are examined by post-mortem 151 * debugging tools; the inconsistent use of 'panic' versus 'panic_' in 152 * the variable naming is historical and allows legacy tools to work. 153 */ 154 #pragma align STACK_ALIGN(panic_stack) 155 char panic_stack[PANICSTKSIZE]; /* reserved stack for panic_thread */ 156 kthread_t *panic_thread; /* first thread to call panicsys() */ 157 cpu_t panic_cpu; /* cpu from first call to panicsys() */ 158 label_t panic_regs; /* setjmp label from panic_thread */ 159 struct regs *panic_reg; /* regs struct from first panicsys() */ 160 char *volatile panicstr; /* format string to first panicsys() */ 161 va_list panicargs; /* arguments to first panicsys() */ 162 clock_t panic_lbolt; /* lbolt at time of panic */ 163 int64_t panic_lbolt64; /* lbolt64 at time of panic */ 164 hrtime_t panic_hrtime; /* hrtime at time of panic */ 165 timespec_t panic_hrestime; /* hrestime at time of panic */ 166 int panic_ipl; /* ipl on panic_cpu at time of panic */ 167 ushort_t panic_schedflag; /* t_schedflag for panic_thread */ 168 cpu_t *panic_bound_cpu; /* t_bound_cpu for panic_thread */ 169 char panic_preempt; /* t_preempt for panic_thread */ 170 171 /* 172 * Panic variables which can be set via /etc/system or patched while 173 * the system is in operation. Again, the stupid names are historic. 174 */ 175 char *panic_bootstr = NULL; /* mdboot string to use after panic */ 176 int panic_bootfcn = AD_BOOT; /* mdboot function to use after panic */ 177 int halt_on_panic = 0; /* halt after dump instead of reboot? */ 178 int nopanicdebug = 0; /* reboot instead of call debugger? */ 179 int in_sync = 0; /* skip vfs_syncall() and just dump? */ 180 181 /* 182 * The do_polled_io flag is set by the panic code to inform the SCSI subsystem 183 * to use polled mode instead of interrupt-driven i/o. 184 */ 185 int do_polled_io = 0; 186 187 /* 188 * The panic_forced flag is set by the uadmin A_DUMP code to inform the 189 * panic subsystem that it should not attempt an initial debug_enter. 190 */ 191 int panic_forced = 0; 192 193 /* 194 * Triggers for panic state transitions: 195 */ 196 int panic_quiesce; /* trigger for CALM -> QUIESCE */ 197 int panic_sync; /* trigger for QUIESCE -> SYNC */ 198 int panic_dump; /* trigger for SYNC -> DUMP */ 199 200 void 201 panicsys(const char *format, va_list alist, struct regs *rp, int on_panic_stack) 202 { 203 int s = spl8(); 204 kthread_t *t = curthread; 205 cpu_t *cp = CPU; 206 207 caddr_t intr_stack = NULL; 208 uint_t intr_actv; 209 210 ushort_t schedflag = t->t_schedflag; 211 cpu_t *bound_cpu = t->t_bound_cpu; 212 char preempt = t->t_preempt; 213 214 (void) setjmp(&t->t_pcb); 215 t->t_flag |= T_PANIC; 216 217 t->t_schedflag |= TS_DONT_SWAP; 218 t->t_bound_cpu = cp; 219 t->t_preempt++; 220 221 panic_enter_hw(s); 222 223 /* 224 * If we're on the interrupt stack and an interrupt thread is available 225 * in this CPU's pool, preserve the interrupt stack by detaching an 226 * interrupt thread and making its stack the intr_stack. 227 */ 228 if (CPU_ON_INTR(cp) && cp->cpu_intr_thread != NULL) { 229 kthread_t *it = cp->cpu_intr_thread; 230 231 intr_stack = cp->cpu_intr_stack; 232 intr_actv = cp->cpu_intr_actv; 233 234 cp->cpu_intr_stack = thread_stk_init(it->t_stk); 235 cp->cpu_intr_thread = it->t_link; 236 237 /* 238 * Clear only the high level bits of cpu_intr_actv. 239 * We want to indicate that high-level interrupts are 240 * not active without destroying the low-level interrupt 241 * information stored there. 242 */ 243 cp->cpu_intr_actv &= ((1 << (LOCK_LEVEL + 1)) - 1); 244 } 245 246 /* 247 * Record one-time panic information and quiesce the other CPUs. 248 * Then print out the panic message and stack trace. 249 */ 250 if (on_panic_stack) { 251 panic_data_t *pdp = (panic_data_t *)panicbuf; 252 253 pdp->pd_version = PANICBUFVERS; 254 pdp->pd_msgoff = sizeof (panic_data_t) - sizeof (panic_nv_t); 255 256 if (t->t_panic_trap != NULL) 257 panic_savetrap(pdp, t->t_panic_trap); 258 else 259 panic_saveregs(pdp, rp); 260 261 (void) vsnprintf(&panicbuf[pdp->pd_msgoff], 262 PANICBUFSIZE - pdp->pd_msgoff, format, alist); 263 264 /* 265 * Call into the platform code to stop the other CPUs. 266 * We currently have all interrupts blocked, and expect that 267 * the platform code will lower ipl only as far as needed to 268 * perform cross-calls, and will acquire as *few* locks as is 269 * possible -- panicstr is not set so we can still deadlock. 270 */ 271 panic_stopcpus(cp, t, s); 272 273 panicstr = (char *)format; 274 va_copy(panicargs, alist); 275 panic_lbolt = lbolt; 276 panic_lbolt64 = lbolt64; 277 panic_hrestime = hrestime; 278 panic_hrtime = gethrtime_waitfree(); 279 panic_thread = t; 280 panic_regs = t->t_pcb; 281 panic_reg = rp; 282 panic_cpu = *cp; 283 panic_ipl = spltoipl(s); 284 panic_schedflag = schedflag; 285 panic_bound_cpu = bound_cpu; 286 panic_preempt = preempt; 287 288 if (intr_stack != NULL) { 289 panic_cpu.cpu_intr_stack = intr_stack; 290 panic_cpu.cpu_intr_actv = intr_actv; 291 } 292 293 /* 294 * Lower ipl to 10 to keep clock() from running, but allow 295 * keyboard interrupts to enter the debugger. These callbacks 296 * are executed with panicstr set so they can bypass locks. 297 */ 298 splx(ipltospl(CLOCK_LEVEL)); 299 panic_quiesce_hw(pdp); 300 (void) FTRACE_STOP(); 301 (void) callb_execute_class(CB_CL_PANIC, NULL); 302 303 log_flushall(); 304 fm_banner(); 305 306 #if defined(__x86) 307 /* 308 * A hypervisor panic originates outside of Solaris, so we 309 * don't want to prepend the panic message with misleading 310 * pointers from within Solaris. 311 */ 312 if (!IN_XPV_PANIC()) 313 #endif 314 printf("\n\rpanic[cpu%d]/thread=%p: ", cp->cpu_id, 315 (void *)t); 316 vprintf(format, alist); 317 printf("\n\n"); 318 319 if (t->t_panic_trap != NULL) { 320 panic_showtrap(t->t_panic_trap); 321 printf("\n"); 322 } 323 324 traceregs(rp); 325 printf("\n"); 326 327 if (((boothowto & RB_DEBUG) || obpdebug) && 328 !nopanicdebug && !panic_forced) { 329 if (dumpvp != NULL) { 330 debug_enter("panic: entering debugger " 331 "(continue to save dump)"); 332 } else { 333 debug_enter("panic: entering debugger " 334 "(no dump device, continue to reboot)"); 335 } 336 } 337 338 } else if (panic_dump != 0 || panic_sync != 0 || panicstr != NULL) { 339 printf("\n\rpanic[cpu%d]/thread=%p: ", cp->cpu_id, (void *)t); 340 vprintf(format, alist); 341 printf("\n"); 342 } else 343 goto spin; 344 345 /* 346 * Prior to performing sync or dump, we make sure that do_polled_io is 347 * set, but we'll leave ipl at 10; deadman(), a CY_HIGH_LEVEL cyclic, 348 * will re-enter panic if we are not making progress with sync or dump. 349 */ 350 351 /* 352 * Sync the filesystems. Reset t_cred if not set because much of 353 * the filesystem code depends on CRED() being valid. 354 */ 355 if (!in_sync && panic_trigger(&panic_sync)) { 356 if (t->t_cred == NULL) 357 t->t_cred = kcred; 358 splx(ipltospl(CLOCK_LEVEL)); 359 do_polled_io = 1; 360 vfs_syncall(); 361 } 362 363 /* 364 * Take the crash dump. If the dump trigger is already set, try to 365 * enter the debugger again before rebooting the system. 366 */ 367 if (panic_trigger(&panic_dump)) { 368 panic_dump_hw(s); 369 splx(ipltospl(CLOCK_LEVEL)); 370 errorq_panic(); 371 do_polled_io = 1; 372 dumpsys(); 373 } else if (((boothowto & RB_DEBUG) || obpdebug) && !nopanicdebug) { 374 debug_enter("panic: entering debugger (continue to reboot)"); 375 } else 376 printf("dump aborted: please record the above information!\n"); 377 378 if (halt_on_panic) 379 mdboot(A_REBOOT, AD_HALT, NULL, B_FALSE); 380 else 381 mdboot(A_REBOOT, panic_bootfcn, panic_bootstr, B_FALSE); 382 spin: 383 /* 384 * Restore ipl to at most CLOCK_LEVEL so we don't end up spinning 385 * and unable to jump into the debugger. 386 */ 387 splx(MIN(s, ipltospl(CLOCK_LEVEL))); 388 for (;;) 389 ; 390 } 391 392 void 393 panic(const char *format, ...) 394 { 395 va_list alist; 396 397 va_start(alist, format); 398 vpanic(format, alist); 399 va_end(alist); 400 } 401