]> CyberLeo.Net >> Repos - FreeBSD/releng/10.0.git/blob - sys/cddl/dev/dtrace/amd64/dtrace_subr.c
- Copy stable/10 (r259064) to releng/10.0 as part of the
[FreeBSD/releng/10.0.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_fasttrap_probe_ptr)(struct regs *);
146 int (*dtrace_pid_probe_ptr)(struct regs *);
147 int (*dtrace_return_probe_ptr)(struct regs *);
148
149 void
150 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
151 {
152         krwlock_t *rwp;
153         proc_t *p = curproc;
154         extern void trap(struct regs *, caddr_t, processorid_t);
155
156         if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
157                 if (curthread->t_cred != p->p_cred) {
158                         cred_t *oldcred = curthread->t_cred;
159                         /*
160                          * DTrace accesses t_cred in probe context.  t_cred
161                          * must always be either NULL, or point to a valid,
162                          * allocated cred structure.
163                          */
164                         curthread->t_cred = crgetcred();
165                         crfree(oldcred);
166                 }
167         }
168
169         if (rp->r_trapno == T_DTRACE_RET) {
170                 uint8_t step = curthread->t_dtrace_step;
171                 uint8_t ret = curthread->t_dtrace_ret;
172                 uintptr_t npc = curthread->t_dtrace_npc;
173
174                 if (curthread->t_dtrace_ast) {
175                         aston(curthread);
176                         curthread->t_sig_check = 1;
177                 }
178
179                 /*
180                  * Clear all user tracing flags.
181                  */
182                 curthread->t_dtrace_ft = 0;
183
184                 /*
185                  * If we weren't expecting to take a return probe trap, kill
186                  * the process as though it had just executed an unassigned
187                  * trap instruction.
188                  */
189                 if (step == 0) {
190                         tsignal(curthread, SIGILL);
191                         return;
192                 }
193
194                 /*
195                  * If we hit this trap unrelated to a return probe, we're
196                  * just here to reset the AST flag since we deferred a signal
197                  * until after we logically single-stepped the instruction we
198                  * copied out.
199                  */
200                 if (ret == 0) {
201                         rp->r_pc = npc;
202                         return;
203                 }
204
205                 /*
206                  * We need to wait until after we've called the
207                  * dtrace_return_probe_ptr function pointer to set %pc.
208                  */
209                 rwp = &CPU->cpu_ft_lock;
210                 rw_enter(rwp, RW_READER);
211                 if (dtrace_return_probe_ptr != NULL)
212                         (void) (*dtrace_return_probe_ptr)(rp);
213                 rw_exit(rwp);
214                 rp->r_pc = npc;
215
216         } else if (rp->r_trapno == T_DTRACE_PROBE) {
217                 rwp = &CPU->cpu_ft_lock;
218                 rw_enter(rwp, RW_READER);
219                 if (dtrace_fasttrap_probe_ptr != NULL)
220                         (void) (*dtrace_fasttrap_probe_ptr)(rp);
221                 rw_exit(rwp);
222
223         } else if (rp->r_trapno == T_BPTFLT) {
224                 uint8_t instr;
225                 rwp = &CPU->cpu_ft_lock;
226
227                 /*
228                  * The DTrace fasttrap provider uses the breakpoint trap
229                  * (int 3). We let DTrace take the first crack at handling
230                  * this trap; if it's not a probe that DTrace knowns about,
231                  * we call into the trap() routine to handle it like a
232                  * breakpoint placed by a conventional debugger.
233                  */
234                 rw_enter(rwp, RW_READER);
235                 if (dtrace_pid_probe_ptr != NULL &&
236                     (*dtrace_pid_probe_ptr)(rp) == 0) {
237                         rw_exit(rwp);
238                         return;
239                 }
240                 rw_exit(rwp);
241
242                 /*
243                  * If the instruction that caused the breakpoint trap doesn't
244                  * look like an int 3 anymore, it may be that this tracepoint
245                  * was removed just after the user thread executed it. In
246                  * that case, return to user land to retry the instuction.
247                  */
248                 if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
249                     instr != FASTTRAP_INSTR) {
250                         rp->r_pc--;
251                         return;
252                 }
253
254                 trap(rp, addr, cpuid);
255
256         } else {
257                 trap(rp, addr, cpuid);
258         }
259 }
260
261 void
262 dtrace_safe_synchronous_signal(void)
263 {
264         kthread_t *t = curthread;
265         struct regs *rp = lwptoregs(ttolwp(t));
266         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
267
268         ASSERT(t->t_dtrace_on);
269
270         /*
271          * If we're not in the range of scratch addresses, we're not actually
272          * tracing user instructions so turn off the flags. If the instruction
273          * we copied out caused a synchonous trap, reset the pc back to its
274          * original value and turn off the flags.
275          */
276         if (rp->r_pc < t->t_dtrace_scrpc ||
277             rp->r_pc > t->t_dtrace_astpc + isz) {
278                 t->t_dtrace_ft = 0;
279         } else if (rp->r_pc == t->t_dtrace_scrpc ||
280             rp->r_pc == t->t_dtrace_astpc) {
281                 rp->r_pc = t->t_dtrace_pc;
282                 t->t_dtrace_ft = 0;
283         }
284 }
285
286 int
287 dtrace_safe_defer_signal(void)
288 {
289         kthread_t *t = curthread;
290         struct regs *rp = lwptoregs(ttolwp(t));
291         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
292
293         ASSERT(t->t_dtrace_on);
294
295         /*
296          * If we're not in the range of scratch addresses, we're not actually
297          * tracing user instructions so turn off the flags.
298          */
299         if (rp->r_pc < t->t_dtrace_scrpc ||
300             rp->r_pc > t->t_dtrace_astpc + isz) {
301                 t->t_dtrace_ft = 0;
302                 return (0);
303         }
304
305         /*
306          * If we have executed the original instruction, but we have performed
307          * neither the jmp back to t->t_dtrace_npc nor the clean up of any
308          * registers used to emulate %rip-relative instructions in 64-bit mode,
309          * we'll save ourselves some effort by doing that here and taking the
310          * signal right away.  We detect this condition by seeing if the program
311          * counter is the range [scrpc + isz, astpc).
312          */
313         if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
314             rp->r_pc < t->t_dtrace_astpc) {
315 #ifdef __amd64
316                 /*
317                  * If there is a scratch register and we're on the
318                  * instruction immediately after the modified instruction,
319                  * restore the value of that scratch register.
320                  */
321                 if (t->t_dtrace_reg != 0 &&
322                     rp->r_pc == t->t_dtrace_scrpc + isz) {
323                         switch (t->t_dtrace_reg) {
324                         case REG_RAX:
325                                 rp->r_rax = t->t_dtrace_regv;
326                                 break;
327                         case REG_RCX:
328                                 rp->r_rcx = t->t_dtrace_regv;
329                                 break;
330                         case REG_R8:
331                                 rp->r_r8 = t->t_dtrace_regv;
332                                 break;
333                         case REG_R9:
334                                 rp->r_r9 = t->t_dtrace_regv;
335                                 break;
336                         }
337                 }
338 #endif
339                 rp->r_pc = t->t_dtrace_npc;
340                 t->t_dtrace_ft = 0;
341                 return (0);
342         }
343
344         /*
345          * Otherwise, make sure we'll return to the kernel after executing
346          * the copied out instruction and defer the signal.
347          */
348         if (!t->t_dtrace_step) {
349                 ASSERT(rp->r_pc < t->t_dtrace_astpc);
350                 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
351                 t->t_dtrace_step = 1;
352         }
353
354         t->t_dtrace_ast = 1;
355
356         return (1);
357 }
358 #endif
359
360 static int64_t  tgt_cpu_tsc;
361 static int64_t  hst_cpu_tsc;
362 static int64_t  tsc_skew[MAXCPU];
363 static uint64_t nsec_scale;
364
365 /* See below for the explanation of this macro. */
366 #define SCALE_SHIFT     28
367
368 static void
369 dtrace_gethrtime_init_cpu(void *arg)
370 {
371         uintptr_t cpu = (uintptr_t) arg;
372
373         if (cpu == curcpu)
374                 tgt_cpu_tsc = rdtsc();
375         else
376                 hst_cpu_tsc = rdtsc();
377 }
378
379 static void
380 dtrace_gethrtime_init(void *arg)
381 {
382         struct pcpu *pc;
383         uint64_t tsc_f;
384         cpuset_t map;
385         int i;
386
387         /*
388          * Get TSC frequency known at this moment.
389          * This should be constant if TSC is invariant.
390          * Otherwise tick->time conversion will be inaccurate, but
391          * will preserve monotonic property of TSC.
392          */
393         tsc_f = atomic_load_acq_64(&tsc_freq);
394
395         /*
396          * The following line checks that nsec_scale calculated below
397          * doesn't overflow 32-bit unsigned integer, so that it can multiply
398          * another 32-bit integer without overflowing 64-bit.
399          * Thus minimum supported TSC frequency is 62.5MHz.
400          */
401         KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), ("TSC frequency is too low"));
402
403         /*
404          * We scale up NANOSEC/tsc_f ratio to preserve as much precision
405          * as possible.
406          * 2^28 factor was chosen quite arbitrarily from practical
407          * considerations:
408          * - it supports TSC frequencies as low as 62.5MHz (see above);
409          * - it provides quite good precision (e < 0.01%) up to THz
410          *   (terahertz) values;
411          */
412         nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
413
414         /* The current CPU is the reference one. */
415         sched_pin();
416         tsc_skew[curcpu] = 0;
417         CPU_FOREACH(i) {
418                 if (i == curcpu)
419                         continue;
420
421                 pc = pcpu_find(i);
422                 CPU_SETOF(PCPU_GET(cpuid), &map);
423                 CPU_SET(pc->pc_cpuid, &map);
424
425                 smp_rendezvous_cpus(map, NULL,
426                     dtrace_gethrtime_init_cpu,
427                     smp_no_rendevous_barrier, (void *)(uintptr_t) i);
428
429                 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
430         }
431         sched_unpin();
432 }
433
434 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
435
436 /*
437  * DTrace needs a high resolution time function which can
438  * be called from a probe context and guaranteed not to have
439  * instrumented with probes itself.
440  *
441  * Returns nanoseconds since boot.
442  */
443 uint64_t
444 dtrace_gethrtime()
445 {
446         uint64_t tsc;
447         uint32_t lo;
448         uint32_t hi;
449
450         /*
451          * We split TSC value into lower and higher 32-bit halves and separately
452          * scale them with nsec_scale, then we scale them down by 2^28
453          * (see nsec_scale calculations) taking into account 32-bit shift of
454          * the higher half and finally add.
455          */
456         tsc = rdtsc() - tsc_skew[curcpu];
457         lo = tsc;
458         hi = tsc >> 32;
459         return (((lo * nsec_scale) >> SCALE_SHIFT) +
460             ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
461 }
462
463 uint64_t
464 dtrace_gethrestime(void)
465 {
466         struct timespec current_time;
467
468         dtrace_getnanotime(&current_time);
469
470         return (current_time.tv_sec * 1000000000ULL + current_time.tv_nsec);
471 }
472
473 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
474 int
475 dtrace_trap(struct trapframe *frame, u_int type)
476 {
477         /*
478          * A trap can occur while DTrace executes a probe. Before
479          * executing the probe, DTrace blocks re-scheduling and sets
480          * a flag in it's per-cpu flags to indicate that it doesn't
481          * want to fault. On returning from the probe, the no-fault
482          * flag is cleared and finally re-scheduling is enabled.
483          *
484          * Check if DTrace has enabled 'no-fault' mode:
485          *
486          */
487         if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
488                 /*
489                  * There are only a couple of trap types that are expected.
490                  * All the rest will be handled in the usual way.
491                  */
492                 switch (type) {
493                 /* Privilieged instruction fault. */
494                 case T_PRIVINFLT:
495                         break;
496                 /* General protection fault. */
497                 case T_PROTFLT:
498                         /* Flag an illegal operation. */
499                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
500
501                         /*
502                          * Offset the instruction pointer to the instruction
503                          * following the one causing the fault.
504                          */
505                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
506                         return (1);
507                 /* Page fault. */
508                 case T_PAGEFLT:
509                         /* Flag a bad address. */
510                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
511                         cpu_core[curcpu].cpuc_dtrace_illval = frame->tf_addr;
512
513                         /*
514                          * Offset the instruction pointer to the instruction
515                          * following the one causing the fault.
516                          */
517                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
518                         return (1);
519                 default:
520                         /* Handle all other traps in the usual way. */
521                         break;
522                 }
523         }
524
525         /* Handle the trap in the usual way. */
526         return (0);
527 }