]> CyberLeo.Net >> Repos - FreeBSD/releng/9.0.git/blob - sys/cddl/dev/dtrace/amd64/dtrace_subr.c
Copy stable/9 to releng/9.0 as part of the FreeBSD 9.0-RELEASE release
[FreeBSD/releng/9.0.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 #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        dtrace_in_probe_addr;
44 extern int              dtrace_in_probe;
45
46 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
47
48 typedef struct dtrace_invop_hdlr {
49         int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
50         struct dtrace_invop_hdlr *dtih_next;
51 } dtrace_invop_hdlr_t;
52
53 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
54
55 int
56 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
57 {
58         dtrace_invop_hdlr_t *hdlr;
59         int rval;
60
61         for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
62                 if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
63                         return (rval);
64
65         return (0);
66 }
67
68 void
69 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
70 {
71         dtrace_invop_hdlr_t *hdlr;
72
73         hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
74         hdlr->dtih_func = func;
75         hdlr->dtih_next = dtrace_invop_hdlr;
76         dtrace_invop_hdlr = hdlr;
77 }
78
79 void
80 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
81 {
82         dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
83
84         for (;;) {
85                 if (hdlr == NULL)
86                         panic("attempt to remove non-existent invop handler");
87
88                 if (hdlr->dtih_func == func)
89                         break;
90
91                 prev = hdlr;
92                 hdlr = hdlr->dtih_next;
93         }
94
95         if (prev == NULL) {
96                 ASSERT(dtrace_invop_hdlr == hdlr);
97                 dtrace_invop_hdlr = hdlr->dtih_next;
98         } else {
99                 ASSERT(dtrace_invop_hdlr != hdlr);
100                 prev->dtih_next = hdlr->dtih_next;
101         }
102
103         kmem_free(hdlr, 0);
104 }
105
106 /*ARGSUSED*/
107 void
108 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
109 {
110         (*func)(0, (uintptr_t) addr_PTmap);
111 }
112
113 void
114 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
115 {
116         cpuset_t cpus;
117
118         if (cpu == DTRACE_CPUALL)
119                 cpus = all_cpus;
120         else
121                 CPU_SETOF(cpu, &cpus);
122
123         smp_rendezvous_cpus(cpus, smp_no_rendevous_barrier, func,
124             smp_no_rendevous_barrier, arg);
125 }
126
127 static void
128 dtrace_sync_func(void)
129 {
130 }
131
132 void
133 dtrace_sync(void)
134 {
135         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
136 }
137
138 #ifdef notyet
139 int (*dtrace_fasttrap_probe_ptr)(struct regs *);
140 int (*dtrace_pid_probe_ptr)(struct regs *);
141 int (*dtrace_return_probe_ptr)(struct regs *);
142
143 void
144 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
145 {
146         krwlock_t *rwp;
147         proc_t *p = curproc;
148         extern void trap(struct regs *, caddr_t, processorid_t);
149
150         if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
151                 if (curthread->t_cred != p->p_cred) {
152                         cred_t *oldcred = curthread->t_cred;
153                         /*
154                          * DTrace accesses t_cred in probe context.  t_cred
155                          * must always be either NULL, or point to a valid,
156                          * allocated cred structure.
157                          */
158                         curthread->t_cred = crgetcred();
159                         crfree(oldcred);
160                 }
161         }
162
163         if (rp->r_trapno == T_DTRACE_RET) {
164                 uint8_t step = curthread->t_dtrace_step;
165                 uint8_t ret = curthread->t_dtrace_ret;
166                 uintptr_t npc = curthread->t_dtrace_npc;
167
168                 if (curthread->t_dtrace_ast) {
169                         aston(curthread);
170                         curthread->t_sig_check = 1;
171                 }
172
173                 /*
174                  * Clear all user tracing flags.
175                  */
176                 curthread->t_dtrace_ft = 0;
177
178                 /*
179                  * If we weren't expecting to take a return probe trap, kill
180                  * the process as though it had just executed an unassigned
181                  * trap instruction.
182                  */
183                 if (step == 0) {
184                         tsignal(curthread, SIGILL);
185                         return;
186                 }
187
188                 /*
189                  * If we hit this trap unrelated to a return probe, we're
190                  * just here to reset the AST flag since we deferred a signal
191                  * until after we logically single-stepped the instruction we
192                  * copied out.
193                  */
194                 if (ret == 0) {
195                         rp->r_pc = npc;
196                         return;
197                 }
198
199                 /*
200                  * We need to wait until after we've called the
201                  * dtrace_return_probe_ptr function pointer to set %pc.
202                  */
203                 rwp = &CPU->cpu_ft_lock;
204                 rw_enter(rwp, RW_READER);
205                 if (dtrace_return_probe_ptr != NULL)
206                         (void) (*dtrace_return_probe_ptr)(rp);
207                 rw_exit(rwp);
208                 rp->r_pc = npc;
209
210         } else if (rp->r_trapno == T_DTRACE_PROBE) {
211                 rwp = &CPU->cpu_ft_lock;
212                 rw_enter(rwp, RW_READER);
213                 if (dtrace_fasttrap_probe_ptr != NULL)
214                         (void) (*dtrace_fasttrap_probe_ptr)(rp);
215                 rw_exit(rwp);
216
217         } else if (rp->r_trapno == T_BPTFLT) {
218                 uint8_t instr;
219                 rwp = &CPU->cpu_ft_lock;
220
221                 /*
222                  * The DTrace fasttrap provider uses the breakpoint trap
223                  * (int 3). We let DTrace take the first crack at handling
224                  * this trap; if it's not a probe that DTrace knowns about,
225                  * we call into the trap() routine to handle it like a
226                  * breakpoint placed by a conventional debugger.
227                  */
228                 rw_enter(rwp, RW_READER);
229                 if (dtrace_pid_probe_ptr != NULL &&
230                     (*dtrace_pid_probe_ptr)(rp) == 0) {
231                         rw_exit(rwp);
232                         return;
233                 }
234                 rw_exit(rwp);
235
236                 /*
237                  * If the instruction that caused the breakpoint trap doesn't
238                  * look like an int 3 anymore, it may be that this tracepoint
239                  * was removed just after the user thread executed it. In
240                  * that case, return to user land to retry the instuction.
241                  */
242                 if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
243                     instr != FASTTRAP_INSTR) {
244                         rp->r_pc--;
245                         return;
246                 }
247
248                 trap(rp, addr, cpuid);
249
250         } else {
251                 trap(rp, addr, cpuid);
252         }
253 }
254
255 void
256 dtrace_safe_synchronous_signal(void)
257 {
258         kthread_t *t = curthread;
259         struct regs *rp = lwptoregs(ttolwp(t));
260         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
261
262         ASSERT(t->t_dtrace_on);
263
264         /*
265          * If we're not in the range of scratch addresses, we're not actually
266          * tracing user instructions so turn off the flags. If the instruction
267          * we copied out caused a synchonous trap, reset the pc back to its
268          * original value and turn off the flags.
269          */
270         if (rp->r_pc < t->t_dtrace_scrpc ||
271             rp->r_pc > t->t_dtrace_astpc + isz) {
272                 t->t_dtrace_ft = 0;
273         } else if (rp->r_pc == t->t_dtrace_scrpc ||
274             rp->r_pc == t->t_dtrace_astpc) {
275                 rp->r_pc = t->t_dtrace_pc;
276                 t->t_dtrace_ft = 0;
277         }
278 }
279
280 int
281 dtrace_safe_defer_signal(void)
282 {
283         kthread_t *t = curthread;
284         struct regs *rp = lwptoregs(ttolwp(t));
285         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
286
287         ASSERT(t->t_dtrace_on);
288
289         /*
290          * If we're not in the range of scratch addresses, we're not actually
291          * tracing user instructions so turn off the flags.
292          */
293         if (rp->r_pc < t->t_dtrace_scrpc ||
294             rp->r_pc > t->t_dtrace_astpc + isz) {
295                 t->t_dtrace_ft = 0;
296                 return (0);
297         }
298
299         /*
300          * If we've executed the original instruction, but haven't performed
301          * the jmp back to t->t_dtrace_npc or the clean up of any registers
302          * used to emulate %rip-relative instructions in 64-bit mode, do that
303          * here and take the signal right away. We detect this condition by
304          * seeing if the program counter is the range [scrpc + isz, astpc).
305          */
306         if (t->t_dtrace_astpc - rp->r_pc <
307             t->t_dtrace_astpc - t->t_dtrace_scrpc - isz) {
308 #ifdef __amd64
309                 /*
310                  * If there is a scratch register and we're on the
311                  * instruction immediately after the modified instruction,
312                  * restore the value of that scratch register.
313                  */
314                 if (t->t_dtrace_reg != 0 &&
315                     rp->r_pc == t->t_dtrace_scrpc + isz) {
316                         switch (t->t_dtrace_reg) {
317                         case REG_RAX:
318                                 rp->r_rax = t->t_dtrace_regv;
319                                 break;
320                         case REG_RCX:
321                                 rp->r_rcx = t->t_dtrace_regv;
322                                 break;
323                         case REG_R8:
324                                 rp->r_r8 = t->t_dtrace_regv;
325                                 break;
326                         case REG_R9:
327                                 rp->r_r9 = t->t_dtrace_regv;
328                                 break;
329                         }
330                 }
331 #endif
332                 rp->r_pc = t->t_dtrace_npc;
333                 t->t_dtrace_ft = 0;
334                 return (0);
335         }
336
337         /*
338          * Otherwise, make sure we'll return to the kernel after executing
339          * the copied out instruction and defer the signal.
340          */
341         if (!t->t_dtrace_step) {
342                 ASSERT(rp->r_pc < t->t_dtrace_astpc);
343                 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
344                 t->t_dtrace_step = 1;
345         }
346
347         t->t_dtrace_ast = 1;
348
349         return (1);
350 }
351 #endif
352
353 static int64_t  tgt_cpu_tsc;
354 static int64_t  hst_cpu_tsc;
355 static int64_t  tsc_skew[MAXCPU];
356 static uint64_t nsec_scale;
357
358 /* See below for the explanation of this macro. */
359 #define SCALE_SHIFT     28
360
361 static void
362 dtrace_gethrtime_init_cpu(void *arg)
363 {
364         uintptr_t cpu = (uintptr_t) arg;
365
366         if (cpu == curcpu)
367                 tgt_cpu_tsc = rdtsc();
368         else
369                 hst_cpu_tsc = rdtsc();
370 }
371
372 static void
373 dtrace_gethrtime_init(void *arg)
374 {
375         struct pcpu *pc;
376         uint64_t tsc_f;
377         cpuset_t map;
378         int i;
379
380         /*
381          * Get TSC frequency known at this moment.
382          * This should be constant if TSC is invariant.
383          * Otherwise tick->time conversion will be inaccurate, but
384          * will preserve monotonic property of TSC.
385          */
386         tsc_f = atomic_load_acq_64(&tsc_freq);
387
388         /*
389          * The following line checks that nsec_scale calculated below
390          * doesn't overflow 32-bit unsigned integer, so that it can multiply
391          * another 32-bit integer without overflowing 64-bit.
392          * Thus minimum supported TSC frequency is 62.5MHz.
393          */
394         KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), ("TSC frequency is too low"));
395
396         /*
397          * We scale up NANOSEC/tsc_f ratio to preserve as much precision
398          * as possible.
399          * 2^28 factor was chosen quite arbitrarily from practical
400          * considerations:
401          * - it supports TSC frequencies as low as 62.5MHz (see above);
402          * - it provides quite good precision (e < 0.01%) up to THz
403          *   (terahertz) values;
404          */
405         nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
406
407         /* The current CPU is the reference one. */
408         sched_pin();
409         tsc_skew[curcpu] = 0;
410         CPU_FOREACH(i) {
411                 if (i == curcpu)
412                         continue;
413
414                 pc = pcpu_find(i);
415                 CPU_SETOF(PCPU_GET(cpuid), &map);
416                 CPU_SET(pc->pc_cpuid, &map);
417
418                 smp_rendezvous_cpus(map, NULL,
419                     dtrace_gethrtime_init_cpu,
420                     smp_no_rendevous_barrier, (void *)(uintptr_t) i);
421
422                 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
423         }
424         sched_unpin();
425 }
426
427 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
428
429 /*
430  * DTrace needs a high resolution time function which can
431  * be called from a probe context and guaranteed not to have
432  * instrumented with probes itself.
433  *
434  * Returns nanoseconds since boot.
435  */
436 uint64_t
437 dtrace_gethrtime()
438 {
439         uint64_t tsc;
440         uint32_t lo;
441         uint32_t hi;
442
443         /*
444          * We split TSC value into lower and higher 32-bit halves and separately
445          * scale them with nsec_scale, then we scale them down by 2^28
446          * (see nsec_scale calculations) taking into account 32-bit shift of
447          * the higher half and finally add.
448          */
449         tsc = rdtsc() + tsc_skew[curcpu];
450         lo = tsc;
451         hi = tsc >> 32;
452         return (((lo * nsec_scale) >> SCALE_SHIFT) +
453             ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
454 }
455
456 uint64_t
457 dtrace_gethrestime(void)
458 {
459         printf("%s(%d): XXX\n",__func__,__LINE__);
460         return (0);
461 }
462
463 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
464 int
465 dtrace_trap(struct trapframe *frame, u_int type)
466 {
467         /*
468          * A trap can occur while DTrace executes a probe. Before
469          * executing the probe, DTrace blocks re-scheduling and sets
470          * a flag in it's per-cpu flags to indicate that it doesn't
471          * want to fault. On returning from the probe, the no-fault
472          * flag is cleared and finally re-scheduling is enabled.
473          *
474          * Check if DTrace has enabled 'no-fault' mode:
475          *
476          */
477         if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
478                 /*
479                  * There are only a couple of trap types that are expected.
480                  * All the rest will be handled in the usual way.
481                  */
482                 switch (type) {
483                 /* Privilieged instruction fault. */
484                 case T_PRIVINFLT:
485                         break;
486                 /* General protection fault. */
487                 case T_PROTFLT:
488                         /* Flag an illegal operation. */
489                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
490
491                         /*
492                          * Offset the instruction pointer to the instruction
493                          * following the one causing the fault.
494                          */
495                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
496                         return (1);
497                 /* Page fault. */
498                 case T_PAGEFLT:
499                         /* Flag a bad address. */
500                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
501                         cpu_core[curcpu].cpuc_dtrace_illval = frame->tf_addr;
502
503                         /*
504                          * Offset the instruction pointer to the instruction
505                          * following the one causing the fault.
506                          */
507                         frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
508                         return (1);
509                 default:
510                         /* Handle all other traps in the usual way. */
511                         break;
512                 }
513         }
514
515         /* Handle the trap in the usual way. */
516         return (0);
517 }