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, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 * 22 * $FreeBSD$ 23 * 24 */ 25 /* 26 * Copyright 2005 Sun Microsystems, Inc. All rights reserved. 27 * Use is subject to license terms. 28 */ 29 30 /* 31 * Copyright (c) 2011, Joyent, Inc. All rights reserved. 32 */ 33 34 #include <sys/param.h> 35 #include <sys/systm.h> 36 #include <sys/types.h> 37 #include <sys/cpuset.h> 38 #include <sys/kernel.h> 39 #include <sys/malloc.h> 40 #include <sys/kmem.h> 41 #include <sys/smp.h> 42 #include <sys/dtrace_impl.h> 43 #include <sys/dtrace_bsd.h> 44 #include <machine/clock.h> 45 #include <machine/cpufunc.h> 46 #include <machine/frame.h> 47 #include <machine/psl.h> 48 #include <machine/trap.h> 49 #include <vm/pmap.h> 50 51 extern uintptr_t kernelbase; 52 53 extern void dtrace_getnanotime(struct timespec *tsp); 54 extern int (*dtrace_invop_jump_addr)(struct trapframe *); 55 56 int dtrace_invop(uintptr_t, struct trapframe *, uintptr_t); 57 int dtrace_invop_start(struct trapframe *frame); 58 void dtrace_invop_init(void); 59 void dtrace_invop_uninit(void); 60 61 typedef struct dtrace_invop_hdlr { 62 int (*dtih_func)(uintptr_t, struct trapframe *, uintptr_t); 63 struct dtrace_invop_hdlr *dtih_next; 64 } dtrace_invop_hdlr_t; 65 66 dtrace_invop_hdlr_t *dtrace_invop_hdlr; 67 68 int 69 dtrace_invop(uintptr_t addr, struct trapframe *frame, uintptr_t eax) 70 { 71 dtrace_invop_hdlr_t *hdlr; 72 int rval; 73 74 for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next) 75 if ((rval = hdlr->dtih_func(addr, frame, eax)) != 0) 76 return (rval); 77 78 return (0); 79 } 80 81 void 82 dtrace_invop_add(int (*func)(uintptr_t, struct trapframe *, uintptr_t)) 83 { 84 dtrace_invop_hdlr_t *hdlr; 85 86 hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP); 87 hdlr->dtih_func = func; 88 hdlr->dtih_next = dtrace_invop_hdlr; 89 dtrace_invop_hdlr = hdlr; 90 } 91 92 void 93 dtrace_invop_remove(int (*func)(uintptr_t, struct trapframe *, uintptr_t)) 94 { 95 dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL; 96 97 for (;;) { 98 if (hdlr == NULL) 99 panic("attempt to remove non-existent invop handler"); 100 101 if (hdlr->dtih_func == func) 102 break; 103 104 prev = hdlr; 105 hdlr = hdlr->dtih_next; 106 } 107 108 if (prev == NULL) { 109 ASSERT(dtrace_invop_hdlr == hdlr); 110 dtrace_invop_hdlr = hdlr->dtih_next; 111 } else { 112 ASSERT(dtrace_invop_hdlr != hdlr); 113 prev->dtih_next = hdlr->dtih_next; 114 } 115 116 kmem_free(hdlr, 0); 117 } 118 119 void 120 dtrace_invop_init(void) 121 { 122 123 dtrace_invop_jump_addr = dtrace_invop_start; 124 } 125 126 void 127 dtrace_invop_uninit(void) 128 { 129 130 dtrace_invop_jump_addr = NULL; 131 } 132 133 void 134 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit)) 135 { 136 (*func)(0, kernelbase); 137 } 138 139 void 140 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg) 141 { 142 cpuset_t cpus; 143 144 if (cpu == DTRACE_CPUALL) 145 cpus = all_cpus; 146 else 147 CPU_SETOF(cpu, &cpus); 148 149 smp_rendezvous_cpus(cpus, smp_no_rendezvous_barrier, func, 150 smp_no_rendezvous_barrier, arg); 151 } 152 153 static void 154 dtrace_sync_func(void) 155 { 156 } 157 158 void 159 dtrace_sync(void) 160 { 161 dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL); 162 } 163 164 #ifdef notyet 165 void 166 dtrace_safe_synchronous_signal(void) 167 { 168 kthread_t *t = curthread; 169 struct regs *rp = lwptoregs(ttolwp(t)); 170 size_t isz = t->t_dtrace_npc - t->t_dtrace_pc; 171 172 ASSERT(t->t_dtrace_on); 173 174 /* 175 * If we're not in the range of scratch addresses, we're not actually 176 * tracing user instructions so turn off the flags. If the instruction 177 * we copied out caused a synchonous trap, reset the pc back to its 178 * original value and turn off the flags. 179 */ 180 if (rp->r_pc < t->t_dtrace_scrpc || 181 rp->r_pc > t->t_dtrace_astpc + isz) { 182 t->t_dtrace_ft = 0; 183 } else if (rp->r_pc == t->t_dtrace_scrpc || 184 rp->r_pc == t->t_dtrace_astpc) { 185 rp->r_pc = t->t_dtrace_pc; 186 t->t_dtrace_ft = 0; 187 } 188 } 189 190 int 191 dtrace_safe_defer_signal(void) 192 { 193 kthread_t *t = curthread; 194 struct regs *rp = lwptoregs(ttolwp(t)); 195 size_t isz = t->t_dtrace_npc - t->t_dtrace_pc; 196 197 ASSERT(t->t_dtrace_on); 198 199 /* 200 * If we're not in the range of scratch addresses, we're not actually 201 * tracing user instructions so turn off the flags. 202 */ 203 if (rp->r_pc < t->t_dtrace_scrpc || 204 rp->r_pc > t->t_dtrace_astpc + isz) { 205 t->t_dtrace_ft = 0; 206 return (0); 207 } 208 209 /* 210 * If we have executed the original instruction, but we have performed 211 * neither the jmp back to t->t_dtrace_npc nor the clean up of any 212 * registers used to emulate %rip-relative instructions in 64-bit mode, 213 * we'll save ourselves some effort by doing that here and taking the 214 * signal right away. We detect this condition by seeing if the program 215 * counter is the range [scrpc + isz, astpc). 216 */ 217 if (rp->r_pc >= t->t_dtrace_scrpc + isz && 218 rp->r_pc < t->t_dtrace_astpc) { 219 #ifdef __amd64 220 /* 221 * If there is a scratch register and we're on the 222 * instruction immediately after the modified instruction, 223 * restore the value of that scratch register. 224 */ 225 if (t->t_dtrace_reg != 0 && 226 rp->r_pc == t->t_dtrace_scrpc + isz) { 227 switch (t->t_dtrace_reg) { 228 case REG_RAX: 229 rp->r_rax = t->t_dtrace_regv; 230 break; 231 case REG_RCX: 232 rp->r_rcx = t->t_dtrace_regv; 233 break; 234 case REG_R8: 235 rp->r_r8 = t->t_dtrace_regv; 236 break; 237 case REG_R9: 238 rp->r_r9 = t->t_dtrace_regv; 239 break; 240 } 241 } 242 #endif 243 rp->r_pc = t->t_dtrace_npc; 244 t->t_dtrace_ft = 0; 245 return (0); 246 } 247 248 /* 249 * Otherwise, make sure we'll return to the kernel after executing 250 * the copied out instruction and defer the signal. 251 */ 252 if (!t->t_dtrace_step) { 253 ASSERT(rp->r_pc < t->t_dtrace_astpc); 254 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc; 255 t->t_dtrace_step = 1; 256 } 257 258 t->t_dtrace_ast = 1; 259 260 return (1); 261 } 262 #endif 263 264 static int64_t tgt_cpu_tsc; 265 static int64_t hst_cpu_tsc; 266 static int64_t tsc_skew[MAXCPU]; 267 static uint64_t nsec_scale; 268 269 /* See below for the explanation of this macro. */ 270 #define SCALE_SHIFT 28 271 272 static void 273 dtrace_gethrtime_init_cpu(void *arg) 274 { 275 uintptr_t cpu = (uintptr_t) arg; 276 277 if (cpu == curcpu) 278 tgt_cpu_tsc = rdtsc(); 279 else 280 hst_cpu_tsc = rdtsc(); 281 } 282 283 #ifdef EARLY_AP_STARTUP 284 static void 285 dtrace_gethrtime_init(void *arg) 286 { 287 struct pcpu *pc; 288 uint64_t tsc_f; 289 cpuset_t map; 290 int i; 291 #else 292 /* 293 * Get the frequency and scale factor as early as possible so that they can be 294 * used for boot-time tracing. 295 */ 296 static void 297 dtrace_gethrtime_init_early(void *arg) 298 { 299 uint64_t tsc_f; 300 #endif 301 302 /* 303 * Get TSC frequency known at this moment. 304 * This should be constant if TSC is invariant. 305 * Otherwise tick->time conversion will be inaccurate, but 306 * will preserve monotonic property of TSC. 307 */ 308 tsc_f = atomic_load_acq_64(&tsc_freq); 309 310 /* 311 * The following line checks that nsec_scale calculated below 312 * doesn't overflow 32-bit unsigned integer, so that it can multiply 313 * another 32-bit integer without overflowing 64-bit. 314 * Thus minimum supported TSC frequency is 62.5MHz. 315 */ 316 KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), 317 ("TSC frequency is too low")); 318 319 /* 320 * We scale up NANOSEC/tsc_f ratio to preserve as much precision 321 * as possible. 322 * 2^28 factor was chosen quite arbitrarily from practical 323 * considerations: 324 * - it supports TSC frequencies as low as 62.5MHz (see above); 325 * - it provides quite good precision (e < 0.01%) up to THz 326 * (terahertz) values; 327 */ 328 nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f; 329 #ifndef EARLY_AP_STARTUP 330 } 331 SYSINIT(dtrace_gethrtime_init_early, SI_SUB_CPU, SI_ORDER_ANY, 332 dtrace_gethrtime_init_early, NULL); 333 334 static void 335 dtrace_gethrtime_init(void *arg) 336 { 337 cpuset_t map; 338 struct pcpu *pc; 339 int i; 340 #endif 341 342 if (vm_guest != VM_GUEST_NO) 343 return; 344 345 /* The current CPU is the reference one. */ 346 sched_pin(); 347 tsc_skew[curcpu] = 0; 348 CPU_FOREACH(i) { 349 if (i == curcpu) 350 continue; 351 352 pc = pcpu_find(i); 353 CPU_SETOF(PCPU_GET(cpuid), &map); 354 CPU_SET(pc->pc_cpuid, &map); 355 356 smp_rendezvous_cpus(map, NULL, 357 dtrace_gethrtime_init_cpu, 358 smp_no_rendezvous_barrier, (void *)(uintptr_t) i); 359 360 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc; 361 } 362 sched_unpin(); 363 } 364 #ifdef EARLY_AP_STARTUP 365 SYSINIT(dtrace_gethrtime_init, SI_SUB_DTRACE, SI_ORDER_ANY, 366 dtrace_gethrtime_init, NULL); 367 #else 368 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, 369 NULL); 370 #endif 371 372 /* 373 * DTrace needs a high resolution time function which can 374 * be called from a probe context and guaranteed not to have 375 * instrumented with probes itself. 376 * 377 * Returns nanoseconds since boot. 378 */ 379 uint64_t 380 dtrace_gethrtime(void) 381 { 382 uint64_t tsc; 383 uint32_t lo, hi; 384 register_t eflags; 385 386 /* 387 * We split TSC value into lower and higher 32-bit halves and separately 388 * scale them with nsec_scale, then we scale them down by 2^28 389 * (see nsec_scale calculations) taking into account 32-bit shift of 390 * the higher half and finally add. 391 */ 392 eflags = intr_disable(); 393 tsc = rdtsc() - tsc_skew[curcpu]; 394 intr_restore(eflags); 395 396 lo = tsc; 397 hi = tsc >> 32; 398 return (((lo * nsec_scale) >> SCALE_SHIFT) + 399 ((hi * nsec_scale) << (32 - SCALE_SHIFT))); 400 } 401 402 uint64_t 403 dtrace_gethrestime(void) 404 { 405 struct timespec current_time; 406 407 dtrace_getnanotime(¤t_time); 408 409 return (current_time.tv_sec * 1000000000ULL + current_time.tv_nsec); 410 } 411 412 /* Function to handle DTrace traps during probes. See i386/i386/trap.c */ 413 int 414 dtrace_trap(struct trapframe *frame, u_int type) 415 { 416 uint16_t nofault; 417 418 /* 419 * A trap can occur while DTrace executes a probe. Before 420 * executing the probe, DTrace blocks re-scheduling and sets 421 * a flag in its per-cpu flags to indicate that it doesn't 422 * want to fault. On returning from the probe, the no-fault 423 * flag is cleared and finally re-scheduling is enabled. 424 * 425 * Check if DTrace has enabled 'no-fault' mode: 426 */ 427 sched_pin(); 428 nofault = cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT; 429 sched_unpin(); 430 if (nofault) { 431 KASSERT((read_eflags() & PSL_I) == 0, ("interrupts enabled")); 432 433 /* 434 * There are only a couple of trap types that are expected. 435 * All the rest will be handled in the usual way. 436 */ 437 switch (type) { 438 /* General protection fault. */ 439 case T_PROTFLT: 440 /* Flag an illegal operation. */ 441 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP; 442 443 /* 444 * Offset the instruction pointer to the instruction 445 * following the one causing the fault. 446 */ 447 frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip); 448 return (1); 449 /* Page fault. */ 450 case T_PAGEFLT: 451 /* Flag a bad address. */ 452 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR; 453 cpu_core[curcpu].cpuc_dtrace_illval = rcr2(); 454 455 /* 456 * Offset the instruction pointer to the instruction 457 * following the one causing the fault. 458 */ 459 frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip); 460 return (1); 461 default: 462 /* Handle all other traps in the usual way. */ 463 break; 464 } 465 } 466 467 /* Handle the trap in the usual way. */ 468 return (0); 469 } 470