sys/cddl/dev/dtrace/amd64/dtrace_subr.c

   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 #include <sys/param.h>
  31 #include <sys/systm.h>
  32 #include <sys/types.h>
  33 #include <sys/kernel.h>
  34 #include <sys/malloc.h>
  35 #include <sys/kmem.h>
  36 #include <sys/smp.h>
  37 #include <sys/dtrace_impl.h>
  38 #include <sys/dtrace_bsd.h>
  39 #include <machine/clock.h>
  40 #include <machine/frame.h>
  41 #include <vm/pmap.h>
  42
  43 extern uintptr_t        kernelbase;
  44 extern uintptr_t        dtrace_in_probe_addr;
  45 extern int              dtrace_in_probe;
  46
  47 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
  48
  49 typedef struct dtrace_invop_hdlr {
  50         int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
  51         struct dtrace_invop_hdlr *dtih_next;
  52 } dtrace_invop_hdlr_t;
  53
  54 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
  55
  56 int
  57 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
  58 {
  59         dtrace_invop_hdlr_t *hdlr;
  60         int rval;
  61
  62         for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
  63                 if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
  64                         return (rval);
  65
  66         return (0);
  67 }
  68
  69 void
  70 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
  71 {
  72         dtrace_invop_hdlr_t *hdlr;
  73
  74         hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
  75         hdlr->dtih_func = func;
  76         hdlr->dtih_next = dtrace_invop_hdlr;
  77         dtrace_invop_hdlr = hdlr;
  78 }
  79
  80 void
  81 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
  82 {
  83         dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
  84
  85         for (;;) {
  86                 if (hdlr == NULL)
  87                         panic("attempt to remove non-existent invop handler");
  88
  89                 if (hdlr->dtih_func == func)
  90                         break;
  91
  92                 prev = hdlr;
  93                 hdlr = hdlr->dtih_next;
  94         }
  95
  96         if (prev == NULL) {
  97                 ASSERT(dtrace_invop_hdlr == hdlr);
  98                 dtrace_invop_hdlr = hdlr->dtih_next;
  99         } else {
 100                 ASSERT(dtrace_invop_hdlr != hdlr);
 101                 prev->dtih_next = hdlr->dtih_next;
 102         }
 103
 104         kmem_free(hdlr, 0);
 105 }
 106
 107 /*ARGSUSED*/
 108 void
 109 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
 110 {
 111         (*func)(0, (uintptr_t) addr_PTmap);
 112 }
 113
 114 void
 115 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
 116 {
 117         cpumask_t cpus;
 118
 119         critical_enter();
 120
 121         if (cpu == DTRACE_CPUALL)
 122                 cpus = all_cpus;
 123         else
 124                 cpus = (cpumask_t) (1 << cpu);
 125
 126         /* If the current CPU is in the set, call the function directly: */
 127         if ((cpus & (1 << curcpu)) != 0) {
 128                 (*func)(arg);
 129
 130                 /* Mask the current CPU from the set */
 131                 cpus &= ~(1 << curcpu);
 132         }
 133
 134         /* If there are any CPUs in the set, cross-call to those CPUs */
 135         if (cpus != 0)
 136                 smp_rendezvous_cpus(cpus, NULL, func, smp_no_rendevous_barrier, arg);
 137
 138         critical_exit();
 139 }
 140
 141 static void
 142 dtrace_sync_func(void)
 143 {
 144 }
 145
 146 void
 147 dtrace_sync(void)
 148 {
 149         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
 150 }
 151
 152 #ifdef notyet
 153 int (*dtrace_fasttrap_probe_ptr)(struct regs *);
 154 int (*dtrace_pid_probe_ptr)(struct regs *);
 155 int (*dtrace_return_probe_ptr)(struct regs *);
 156
 157 void
 158 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
 159 {
 160         krwlock_t *rwp;
 161         proc_t *p = curproc;
 162         extern void trap(struct regs *, caddr_t, processorid_t);
 163
 164         if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
 165                 if (curthread->t_cred != p->p_cred) {
 166                         cred_t *oldcred = curthread->t_cred;
 167                         /*
 168                          * DTrace accesses t_cred in probe context.  t_cred
 169                          * must always be either NULL, or point to a valid,
 170                          * allocated cred structure.
 171                          */
 172                         curthread->t_cred = crgetcred();
 173                         crfree(oldcred);
 174                 }
 175         }
 176
 177         if (rp->r_trapno == T_DTRACE_RET) {
 178                 uint8_t step = curthread->t_dtrace_step;
 179                 uint8_t ret = curthread->t_dtrace_ret;
 180                 uintptr_t npc = curthread->t_dtrace_npc;
 181
 182                 if (curthread->t_dtrace_ast) {
 183                         aston(curthread);
 184                         curthread->t_sig_check = 1;
 185                 }
 186
 187                 /*
 188                  * Clear all user tracing flags.
 189                  */
 190                 curthread->t_dtrace_ft = 0;
 191
 192                 /*
 193                  * If we weren't expecting to take a return probe trap, kill
 194                  * the process as though it had just executed an unassigned
 195                  * trap instruction.
 196                  */
 197                 if (step == 0) {
 198                         tsignal(curthread, SIGILL);
 199                         return;
 200                 }
 201
 202                 /*
 203                  * If we hit this trap unrelated to a return probe, we're
 204                  * just here to reset the AST flag since we deferred a signal
 205                  * until after we logically single-stepped the instruction we
 206                  * copied out.
 207                  */
 208                 if (ret == 0) {
 209                         rp->r_pc = npc;
 210                         return;
 211                 }
 212
 213                 /*
 214                  * We need to wait until after we've called the
 215                  * dtrace_return_probe_ptr function pointer to set %pc.
 216                  */
 217                 rwp = &CPU->cpu_ft_lock;
 218                 rw_enter(rwp, RW_READER);
 219                 if (dtrace_return_probe_ptr != NULL)
 220                         (void) (*dtrace_return_probe_ptr)(rp);
 221                 rw_exit(rwp);
 222                 rp->r_pc = npc;
 223
 224         } else if (rp->r_trapno == T_DTRACE_PROBE) {
 225                 rwp = &CPU->cpu_ft_lock;
 226                 rw_enter(rwp, RW_READER);
 227                 if (dtrace_fasttrap_probe_ptr != NULL)
 228                         (void) (*dtrace_fasttrap_probe_ptr)(rp);
 229                 rw_exit(rwp);
 230
 231         } else if (rp->r_trapno == T_BPTFLT) {
 232                 uint8_t instr;
 233                 rwp = &CPU->cpu_ft_lock;
 234
 235                 /*
 236                  * The DTrace fasttrap provider uses the breakpoint trap
 237                  * (int 3). We let DTrace take the first crack at handling
 238                  * this trap; if it's not a probe that DTrace knowns about,
 239                  * we call into the trap() routine to handle it like a
 240                  * breakpoint placed by a conventional debugger.
 241                  */
 242                 rw_enter(rwp, RW_READER);
 243                 if (dtrace_pid_probe_ptr != NULL &&
 244                     (*dtrace_pid_probe_ptr)(rp) == 0) {
 245                         rw_exit(rwp);
 246                         return;
 247                 }
 248                 rw_exit(rwp);
 249
 250                 /*
 251                  * If the instruction that caused the breakpoint trap doesn't
 252                  * look like an int 3 anymore, it may be that this tracepoint
 253                  * was removed just after the user thread executed it. In
 254                  * that case, return to user land to retry the instuction.
 255                  */
 256                 if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
 257                     instr != FASTTRAP_INSTR) {
 258                         rp->r_pc--;
 259                         return;
 260                 }
 261
 262                 trap(rp, addr, cpuid);
 263
 264         } else {
 265                 trap(rp, addr, cpuid);
 266         }
 267 }
 268
 269 void
 270 dtrace_safe_synchronous_signal(void)
 271 {
 272         kthread_t *t = curthread;
 273         struct regs *rp = lwptoregs(ttolwp(t));
 274         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
 275
 276         ASSERT(t->t_dtrace_on);
 277
 278         /*
 279          * If we're not in the range of scratch addresses, we're not actually
 280          * tracing user instructions so turn off the flags. If the instruction
 281          * we copied out caused a synchonous trap, reset the pc back to its
 282          * original value and turn off the flags.
 283          */
 284         if (rp->r_pc < t->t_dtrace_scrpc ||
 285             rp->r_pc > t->t_dtrace_astpc + isz) {
 286                 t->t_dtrace_ft = 0;
 287         } else if (rp->r_pc == t->t_dtrace_scrpc ||
 288             rp->r_pc == t->t_dtrace_astpc) {
 289                 rp->r_pc = t->t_dtrace_pc;
 290                 t->t_dtrace_ft = 0;
 291         }
 292 }
 293
 294 int
 295 dtrace_safe_defer_signal(void)
 296 {
 297         kthread_t *t = curthread;
 298         struct regs *rp = lwptoregs(ttolwp(t));
 299         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
 300
 301         ASSERT(t->t_dtrace_on);
 302
 303         /*
 304          * If we're not in the range of scratch addresses, we're not actually
 305          * tracing user instructions so turn off the flags.
 306          */
 307         if (rp->r_pc < t->t_dtrace_scrpc ||
 308             rp->r_pc > t->t_dtrace_astpc + isz) {
 309                 t->t_dtrace_ft = 0;
 310                 return (0);
 311         }
 312
 313         /*
 314          * If we've executed the original instruction, but haven't performed
 315          * the jmp back to t->t_dtrace_npc or the clean up of any registers
 316          * used to emulate %rip-relative instructions in 64-bit mode, do that
 317          * here and take the signal right away. We detect this condition by
 318          * seeing if the program counter is the range [scrpc + isz, astpc).
 319          */
 320         if (t->t_dtrace_astpc - rp->r_pc <
 321             t->t_dtrace_astpc - t->t_dtrace_scrpc - isz) {
 322 #ifdef __amd64
 323                 /*
 324                  * If there is a scratch register and we're on the
 325                  * instruction immediately after the modified instruction,
 326                  * restore the value of that scratch register.
 327                  */
 328                 if (t->t_dtrace_reg != 0 &&
 329                     rp->r_pc == t->t_dtrace_scrpc + isz) {
 330                         switch (t->t_dtrace_reg) {
 331                         case REG_RAX:
 332                                 rp->r_rax = t->t_dtrace_regv;
 333                                 break;
 334                         case REG_RCX:
 335                                 rp->r_rcx = t->t_dtrace_regv;
 336                                 break;
 337                         case REG_R8:
 338                                 rp->r_r8 = t->t_dtrace_regv;
 339                                 break;
 340                         case REG_R9:
 341                                 rp->r_r9 = t->t_dtrace_regv;
 342                                 break;
 343                         }
 344                 }
 345 #endif
 346                 rp->r_pc = t->t_dtrace_npc;
 347                 t->t_dtrace_ft = 0;
 348                 return (0);
 349         }
 350
 351         /*
 352          * Otherwise, make sure we'll return to the kernel after executing
 353          * the copied out instruction and defer the signal.
 354          */
 355         if (!t->t_dtrace_step) {
 356                 ASSERT(rp->r_pc < t->t_dtrace_astpc);
 357                 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
 358                 t->t_dtrace_step = 1;
 359         }
 360
 361         t->t_dtrace_ast = 1;
 362
 363         return (1);
 364 }
 365 #endif
 366
 367 static int64_t  tgt_cpu_tsc;
 368 static int64_t  hst_cpu_tsc;
 369 static int64_t  tsc_skew[MAXCPU];
 370
 371 static void
 372 dtrace_gethrtime_init_sync(void *arg)
 373 {
 374 #ifdef CHECK_SYNC
 375         /*
 376          * Delay this function from returning on one
 377          * of the CPUs to check that the synchronisation
 378          * works.
 379          */
 380         uintptr_t cpu = (uintptr_t) arg;
 381
 382         if (cpu == curcpu) {
 383                 int i;
 384                 for (i = 0; i < 1000000000; i++)
 385                         tgt_cpu_tsc = rdtsc();
 386                 tgt_cpu_tsc = 0;
 387         }
 388 #endif
 389 }
 390
 391 static void
 392 dtrace_gethrtime_init_cpu(void *arg)
 393 {
 394         uintptr_t cpu = (uintptr_t) arg;
 395
 396         if (cpu == curcpu)
 397                 tgt_cpu_tsc = rdtsc();
 398         else
 399                 hst_cpu_tsc = rdtsc();
 400 }
 401
 402 static void
 403 dtrace_gethrtime_init(void *arg)
 404 {
 405         cpumask_t map;
 406         int i;
 407         struct pcpu *cp;
 408
 409         /* The current CPU is the reference one. */
 410         tsc_skew[curcpu] = 0;
 411
 412         for (i = 0; i <= mp_maxid; i++) {
 413                 if (i == curcpu)
 414                         continue;
 415
 416                 if ((cp = pcpu_find(i)) == NULL)
 417                         continue;
 418
 419                 map = 0;
 420                 map |= (1 << curcpu);
 421                 map |= (1 << i);
 422
 423                 smp_rendezvous_cpus(map, dtrace_gethrtime_init_sync,
 424                     dtrace_gethrtime_init_cpu,
 425                     smp_no_rendevous_barrier, (void *)(uintptr_t) i);
 426
 427                 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
 428         }
 429 }
 430
 431 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
 432
 433 /*
 434  * DTrace needs a high resolution time function which can
 435  * be called from a probe context and guaranteed not to have
 436  * instrumented with probes itself.
 437  *
 438  * Returns nanoseconds since boot.
 439  */
 440 uint64_t
 441 dtrace_gethrtime()
 442 {
 443         return ((rdtsc() + tsc_skew[curcpu]) * (int64_t) 1000000000 / tsc_freq);
 444 }
 445
 446 uint64_t
 447 dtrace_gethrestime(void)
 448 {
 449         printf("%s(%d): XXX\n",__func__,__LINE__);
 450         return (0);
 451 }
 452
 453 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
 454 int
 455 dtrace_trap(struct trapframe *frame, u_int type)
 456 {
 457         /*
 458          * A trap can occur while DTrace executes a probe. Before
 459          * executing the probe, DTrace blocks re-scheduling and sets
 460          * a flag in it's per-cpu flags to indicate that it doesn't
 461          * want to fault. On returning from the the probe, the no-fault
 462          * flag is cleared and finally re-scheduling is enabled.
 463          *
 464          * Check if DTrace has enabled 'no-fault' mode:
 465          *
 466          */
 467         if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
 468                 /*
 469                  * There are only a couple of trap types that are expected.
 470                  * All the rest will be handled in the usual way.
 471                  */
 472                 switch (type) {
 473                 /* Privilieged instruction fault. */
 474                 case T_PRIVINFLT:
 475                         break;
 476                 /* General protection fault. */
 477                 case T_PROTFLT:
 478                         /* Flag an illegal operation. */
 479                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
 480
 481                         /*
 482                          * Offset the instruction pointer to the instruction
 483                          * following the one causing the fault.
 484                          */
 485                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
 486                         return (1);
 487                 /* Page fault. */
 488                 case T_PAGEFLT:
 489                         /* Flag a bad address. */
 490                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
 491                         cpu_core[curcpu].cpuc_dtrace_illval = frame->tf_addr;
 492
 493                         /*
 494                          * Offset the instruction pointer to the instruction
 495                          * following the one causing the fault.
 496                          */
 497                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
 498                         return (1);
 499                 default:
 500                         /* Handle all other traps in the usual way. */
 501                         break;
 502                 }
 503         }
 504
 505         /* Handle the trap in the usual way. */
 506         return (0);
 507 }