]> CyberLeo.Net >> Repos - FreeBSD/releng/10.2.git/blob - sys/cddl/dev/dtrace/amd64/dtrace_subr.c
- Copy stable/10@285827 to releng/10.2 in preparation for 10.2-RC1
[FreeBSD/releng/10.2.git] / 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 /*
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/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/kmem.h>
40 #include <sys/smp.h>
41 #include <sys/dtrace_impl.h>
42 #include <sys/dtrace_bsd.h>
43 #include <machine/clock.h>
44 #include <machine/frame.h>
45 #include <vm/pmap.h>
46
47 extern uintptr_t        dtrace_in_probe_addr;
48 extern int              dtrace_in_probe;
49
50 extern void dtrace_getnanotime(struct timespec *tsp);
51
52 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
53
54 typedef struct dtrace_invop_hdlr {
55         int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
56         struct dtrace_invop_hdlr *dtih_next;
57 } dtrace_invop_hdlr_t;
58
59 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
60
61 int
62 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
63 {
64         dtrace_invop_hdlr_t *hdlr;
65         int rval;
66
67         for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
68                 if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
69                         return (rval);
70
71         return (0);
72 }
73
74 void
75 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
76 {
77         dtrace_invop_hdlr_t *hdlr;
78
79         hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
80         hdlr->dtih_func = func;
81         hdlr->dtih_next = dtrace_invop_hdlr;
82         dtrace_invop_hdlr = hdlr;
83 }
84
85 void
86 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
87 {
88         dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
89
90         for (;;) {
91                 if (hdlr == NULL)
92                         panic("attempt to remove non-existent invop handler");
93
94                 if (hdlr->dtih_func == func)
95                         break;
96
97                 prev = hdlr;
98                 hdlr = hdlr->dtih_next;
99         }
100
101         if (prev == NULL) {
102                 ASSERT(dtrace_invop_hdlr == hdlr);
103                 dtrace_invop_hdlr = hdlr->dtih_next;
104         } else {
105                 ASSERT(dtrace_invop_hdlr != hdlr);
106                 prev->dtih_next = hdlr->dtih_next;
107         }
108
109         kmem_free(hdlr, 0);
110 }
111
112 /*ARGSUSED*/
113 void
114 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
115 {
116         (*func)(0, (uintptr_t) addr_PTmap);
117 }
118
119 void
120 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
121 {
122         cpuset_t cpus;
123
124         if (cpu == DTRACE_CPUALL)
125                 cpus = all_cpus;
126         else
127                 CPU_SETOF(cpu, &cpus);
128
129         smp_rendezvous_cpus(cpus, smp_no_rendevous_barrier, func,
130             smp_no_rendevous_barrier, arg);
131 }
132
133 static void
134 dtrace_sync_func(void)
135 {
136 }
137
138 void
139 dtrace_sync(void)
140 {
141         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
142 }
143
144 #ifdef notyet
145 int (*dtrace_pid_probe_ptr)(struct regs *);
146 int (*dtrace_return_probe_ptr)(struct regs *);
147
148 void
149 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
150 {
151         krwlock_t *rwp;
152         proc_t *p = curproc;
153         extern void trap(struct regs *, caddr_t, processorid_t);
154
155         if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
156                 if (curthread->t_cred != p->p_cred) {
157                         cred_t *oldcred = curthread->t_cred;
158                         /*
159                          * DTrace accesses t_cred in probe context.  t_cred
160                          * must always be either NULL, or point to a valid,
161                          * allocated cred structure.
162                          */
163                         curthread->t_cred = crgetcred();
164                         crfree(oldcred);
165                 }
166         }
167
168         if (rp->r_trapno == T_DTRACE_RET) {
169                 uint8_t step = curthread->t_dtrace_step;
170                 uint8_t ret = curthread->t_dtrace_ret;
171                 uintptr_t npc = curthread->t_dtrace_npc;
172
173                 if (curthread->t_dtrace_ast) {
174                         aston(curthread);
175                         curthread->t_sig_check = 1;
176                 }
177
178                 /*
179                  * Clear all user tracing flags.
180                  */
181                 curthread->t_dtrace_ft = 0;
182
183                 /*
184                  * If we weren't expecting to take a return probe trap, kill
185                  * the process as though it had just executed an unassigned
186                  * trap instruction.
187                  */
188                 if (step == 0) {
189                         tsignal(curthread, SIGILL);
190                         return;
191                 }
192
193                 /*
194                  * If we hit this trap unrelated to a return probe, we're
195                  * just here to reset the AST flag since we deferred a signal
196                  * until after we logically single-stepped the instruction we
197                  * copied out.
198                  */
199                 if (ret == 0) {
200                         rp->r_pc = npc;
201                         return;
202                 }
203
204                 /*
205                  * We need to wait until after we've called the
206                  * dtrace_return_probe_ptr function pointer to set %pc.
207                  */
208                 rwp = &CPU->cpu_ft_lock;
209                 rw_enter(rwp, RW_READER);
210                 if (dtrace_return_probe_ptr != NULL)
211                         (void) (*dtrace_return_probe_ptr)(rp);
212                 rw_exit(rwp);
213                 rp->r_pc = npc;
214
215         } else if (rp->r_trapno == T_BPTFLT) {
216                 uint8_t instr;
217                 rwp = &CPU->cpu_ft_lock;
218
219                 /*
220                  * The DTrace fasttrap provider uses the breakpoint trap
221                  * (int 3). We let DTrace take the first crack at handling
222                  * this trap; if it's not a probe that DTrace knowns about,
223                  * we call into the trap() routine to handle it like a
224                  * breakpoint placed by a conventional debugger.
225                  */
226                 rw_enter(rwp, RW_READER);
227                 if (dtrace_pid_probe_ptr != NULL &&
228                     (*dtrace_pid_probe_ptr)(rp) == 0) {
229                         rw_exit(rwp);
230                         return;
231                 }
232                 rw_exit(rwp);
233
234                 /*
235                  * If the instruction that caused the breakpoint trap doesn't
236                  * look like an int 3 anymore, it may be that this tracepoint
237                  * was removed just after the user thread executed it. In
238                  * that case, return to user land to retry the instuction.
239                  */
240                 if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
241                     instr != FASTTRAP_INSTR) {
242                         rp->r_pc--;
243                         return;
244                 }
245
246                 trap(rp, addr, cpuid);
247
248         } else {
249                 trap(rp, addr, cpuid);
250         }
251 }
252
253 void
254 dtrace_safe_synchronous_signal(void)
255 {
256         kthread_t *t = curthread;
257         struct regs *rp = lwptoregs(ttolwp(t));
258         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
259
260         ASSERT(t->t_dtrace_on);
261
262         /*
263          * If we're not in the range of scratch addresses, we're not actually
264          * tracing user instructions so turn off the flags. If the instruction
265          * we copied out caused a synchonous trap, reset the pc back to its
266          * original value and turn off the flags.
267          */
268         if (rp->r_pc < t->t_dtrace_scrpc ||
269             rp->r_pc > t->t_dtrace_astpc + isz) {
270                 t->t_dtrace_ft = 0;
271         } else if (rp->r_pc == t->t_dtrace_scrpc ||
272             rp->r_pc == t->t_dtrace_astpc) {
273                 rp->r_pc = t->t_dtrace_pc;
274                 t->t_dtrace_ft = 0;
275         }
276 }
277
278 int
279 dtrace_safe_defer_signal(void)
280 {
281         kthread_t *t = curthread;
282         struct regs *rp = lwptoregs(ttolwp(t));
283         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
284
285         ASSERT(t->t_dtrace_on);
286
287         /*
288          * If we're not in the range of scratch addresses, we're not actually
289          * tracing user instructions so turn off the flags.
290          */
291         if (rp->r_pc < t->t_dtrace_scrpc ||
292             rp->r_pc > t->t_dtrace_astpc + isz) {
293                 t->t_dtrace_ft = 0;
294                 return (0);
295         }
296
297         /*
298          * If we have executed the original instruction, but we have performed
299          * neither the jmp back to t->t_dtrace_npc nor the clean up of any
300          * registers used to emulate %rip-relative instructions in 64-bit mode,
301          * we'll save ourselves some effort by doing that here and taking the
302          * signal right away.  We detect this condition by seeing if the program
303          * counter is the range [scrpc + isz, astpc).
304          */
305         if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
306             rp->r_pc < t->t_dtrace_astpc) {
307 #ifdef __amd64
308                 /*
309                  * If there is a scratch register and we're on the
310                  * instruction immediately after the modified instruction,
311                  * restore the value of that scratch register.
312                  */
313                 if (t->t_dtrace_reg != 0 &&
314                     rp->r_pc == t->t_dtrace_scrpc + isz) {
315                         switch (t->t_dtrace_reg) {
316                         case REG_RAX:
317                                 rp->r_rax = t->t_dtrace_regv;
318                                 break;
319                         case REG_RCX:
320                                 rp->r_rcx = t->t_dtrace_regv;
321                                 break;
322                         case REG_R8:
323                                 rp->r_r8 = t->t_dtrace_regv;
324                                 break;
325                         case REG_R9:
326                                 rp->r_r9 = t->t_dtrace_regv;
327                                 break;
328                         }
329                 }
330 #endif
331                 rp->r_pc = t->t_dtrace_npc;
332                 t->t_dtrace_ft = 0;
333                 return (0);
334         }
335
336         /*
337          * Otherwise, make sure we'll return to the kernel after executing
338          * the copied out instruction and defer the signal.
339          */
340         if (!t->t_dtrace_step) {
341                 ASSERT(rp->r_pc < t->t_dtrace_astpc);
342                 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
343                 t->t_dtrace_step = 1;
344         }
345
346         t->t_dtrace_ast = 1;
347
348         return (1);
349 }
350 #endif
351
352 static int64_t  tgt_cpu_tsc;
353 static int64_t  hst_cpu_tsc;
354 static int64_t  tsc_skew[MAXCPU];
355 static uint64_t nsec_scale;
356
357 /* See below for the explanation of this macro. */
358 #define SCALE_SHIFT     28
359
360 static void
361 dtrace_gethrtime_init_cpu(void *arg)
362 {
363         uintptr_t cpu = (uintptr_t) arg;
364
365         if (cpu == curcpu)
366                 tgt_cpu_tsc = rdtsc();
367         else
368                 hst_cpu_tsc = rdtsc();
369 }
370
371 static void
372 dtrace_gethrtime_init(void *arg)
373 {
374         struct pcpu *pc;
375         uint64_t tsc_f;
376         cpuset_t map;
377         int i;
378
379         /*
380          * Get TSC frequency known at this moment.
381          * This should be constant if TSC is invariant.
382          * Otherwise tick->time conversion will be inaccurate, but
383          * will preserve monotonic property of TSC.
384          */
385         tsc_f = atomic_load_acq_64(&tsc_freq);
386
387         /*
388          * The following line checks that nsec_scale calculated below
389          * doesn't overflow 32-bit unsigned integer, so that it can multiply
390          * another 32-bit integer without overflowing 64-bit.
391          * Thus minimum supported TSC frequency is 62.5MHz.
392          */
393         KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), ("TSC frequency is too low"));
394
395         /*
396          * We scale up NANOSEC/tsc_f ratio to preserve as much precision
397          * as possible.
398          * 2^28 factor was chosen quite arbitrarily from practical
399          * considerations:
400          * - it supports TSC frequencies as low as 62.5MHz (see above);
401          * - it provides quite good precision (e < 0.01%) up to THz
402          *   (terahertz) values;
403          */
404         nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
405
406         /* The current CPU is the reference one. */
407         sched_pin();
408         tsc_skew[curcpu] = 0;
409         CPU_FOREACH(i) {
410                 if (i == curcpu)
411                         continue;
412
413                 pc = pcpu_find(i);
414                 CPU_SETOF(PCPU_GET(cpuid), &map);
415                 CPU_SET(pc->pc_cpuid, &map);
416
417                 smp_rendezvous_cpus(map, NULL,
418                     dtrace_gethrtime_init_cpu,
419                     smp_no_rendevous_barrier, (void *)(uintptr_t) i);
420
421                 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
422         }
423         sched_unpin();
424 }
425
426 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
427
428 /*
429  * DTrace needs a high resolution time function which can
430  * be called from a probe context and guaranteed not to have
431  * instrumented with probes itself.
432  *
433  * Returns nanoseconds since boot.
434  */
435 uint64_t
436 dtrace_gethrtime()
437 {
438         uint64_t tsc;
439         uint32_t lo;
440         uint32_t hi;
441
442         /*
443          * We split TSC value into lower and higher 32-bit halves and separately
444          * scale them with nsec_scale, then we scale them down by 2^28
445          * (see nsec_scale calculations) taking into account 32-bit shift of
446          * the higher half and finally add.
447          */
448         tsc = rdtsc() - tsc_skew[curcpu];
449         lo = tsc;
450         hi = tsc >> 32;
451         return (((lo * nsec_scale) >> SCALE_SHIFT) +
452             ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
453 }
454
455 uint64_t
456 dtrace_gethrestime(void)
457 {
458         struct timespec current_time;
459
460         dtrace_getnanotime(&current_time);
461
462         return (current_time.tv_sec * 1000000000ULL + current_time.tv_nsec);
463 }
464
465 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
466 int
467 dtrace_trap(struct trapframe *frame, u_int type)
468 {
469         /*
470          * A trap can occur while DTrace executes a probe. Before
471          * executing the probe, DTrace blocks re-scheduling and sets
472          * a flag in it's per-cpu flags to indicate that it doesn't
473          * want to fault. On returning from the probe, the no-fault
474          * flag is cleared and finally re-scheduling is enabled.
475          *
476          * Check if DTrace has enabled 'no-fault' mode:
477          *
478          */
479         if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
480                 /*
481                  * There are only a couple of trap types that are expected.
482                  * All the rest will be handled in the usual way.
483                  */
484                 switch (type) {
485                 /* Privilieged instruction fault. */
486                 case T_PRIVINFLT:
487                         break;
488                 /* General protection fault. */
489                 case T_PROTFLT:
490                         /* Flag an illegal operation. */
491                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
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                 /* Page fault. */
500                 case T_PAGEFLT:
501                         /* Flag a bad address. */
502                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
503                         cpu_core[curcpu].cpuc_dtrace_illval = frame->tf_addr;
504
505                         /*
506                          * Offset the instruction pointer to the instruction
507                          * following the one causing the fault.
508                          */
509                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
510                         return (1);
511                 default:
512                         /* Handle all other traps in the usual way. */
513                         break;
514                 }
515         }
516
517         /* Handle the trap in the usual way. */
518         return (0);
519 }