]> CyberLeo.Net >> Repos - FreeBSD/releng/9.0.git/blob - sys/cddl/dev/dtrace/i386/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 / i386 / 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/cpuset.h>
34 #include <sys/kernel.h>
35 #include <sys/malloc.h>
36 #include <sys/kmem.h>
37 #include <sys/smp.h>
38 #include <sys/dtrace_impl.h>
39 #include <sys/dtrace_bsd.h>
40 #include <machine/clock.h>
41 #include <machine/frame.h>
42 #include <vm/pmap.h>
43
44 extern uintptr_t        kernelbase;
45 extern uintptr_t        dtrace_in_probe_addr;
46 extern int              dtrace_in_probe;
47
48 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
49
50 typedef struct dtrace_invop_hdlr {
51         int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
52         struct dtrace_invop_hdlr *dtih_next;
53 } dtrace_invop_hdlr_t;
54
55 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
56
57 int
58 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
59 {
60         dtrace_invop_hdlr_t *hdlr;
61         int rval;
62
63         for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
64                 if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
65                         return (rval);
66
67         return (0);
68 }
69
70 void
71 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
72 {
73         dtrace_invop_hdlr_t *hdlr;
74
75         hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
76         hdlr->dtih_func = func;
77         hdlr->dtih_next = dtrace_invop_hdlr;
78         dtrace_invop_hdlr = hdlr;
79 }
80
81 void
82 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
83 {
84         dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
85
86         for (;;) {
87                 if (hdlr == NULL)
88                         panic("attempt to remove non-existent invop handler");
89
90                 if (hdlr->dtih_func == func)
91                         break;
92
93                 prev = hdlr;
94                 hdlr = hdlr->dtih_next;
95         }
96
97         if (prev == NULL) {
98                 ASSERT(dtrace_invop_hdlr == hdlr);
99                 dtrace_invop_hdlr = hdlr->dtih_next;
100         } else {
101                 ASSERT(dtrace_invop_hdlr != hdlr);
102                 prev->dtih_next = hdlr->dtih_next;
103         }
104
105         kmem_free(hdlr, 0);
106 }
107
108 void
109 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
110 {
111         (*func)(0, kernelbase);
112 }
113
114 void
115 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
116 {
117         cpuset_t cpus;
118
119         if (cpu == DTRACE_CPUALL)
120                 cpus = all_cpus;
121         else
122                 CPU_SETOF(cpu, &cpus);
123
124         smp_rendezvous_cpus(cpus, smp_no_rendevous_barrier, func,
125             smp_no_rendevous_barrier, arg);
126 }
127
128 static void
129 dtrace_sync_func(void)
130 {
131 }
132
133 void
134 dtrace_sync(void)
135 {
136         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
137 }
138
139 #ifdef notyet
140 int (*dtrace_fasttrap_probe_ptr)(struct regs *);
141 int (*dtrace_pid_probe_ptr)(struct regs *);
142 int (*dtrace_return_probe_ptr)(struct regs *);
143
144 void
145 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
146 {
147         krwlock_t *rwp;
148         proc_t *p = curproc;
149         extern void trap(struct regs *, caddr_t, processorid_t);
150
151         if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
152                 if (curthread->t_cred != p->p_cred) {
153                         cred_t *oldcred = curthread->t_cred;
154                         /*
155                          * DTrace accesses t_cred in probe context.  t_cred
156                          * must always be either NULL, or point to a valid,
157                          * allocated cred structure.
158                          */
159                         curthread->t_cred = crgetcred();
160                         crfree(oldcred);
161                 }
162         }
163
164         if (rp->r_trapno == T_DTRACE_RET) {
165                 uint8_t step = curthread->t_dtrace_step;
166                 uint8_t ret = curthread->t_dtrace_ret;
167                 uintptr_t npc = curthread->t_dtrace_npc;
168
169                 if (curthread->t_dtrace_ast) {
170                         aston(curthread);
171                         curthread->t_sig_check = 1;
172                 }
173
174                 /*
175                  * Clear all user tracing flags.
176                  */
177                 curthread->t_dtrace_ft = 0;
178
179                 /*
180                  * If we weren't expecting to take a return probe trap, kill
181                  * the process as though it had just executed an unassigned
182                  * trap instruction.
183                  */
184                 if (step == 0) {
185                         tsignal(curthread, SIGILL);
186                         return;
187                 }
188
189                 /*
190                  * If we hit this trap unrelated to a return probe, we're
191                  * just here to reset the AST flag since we deferred a signal
192                  * until after we logically single-stepped the instruction we
193                  * copied out.
194                  */
195                 if (ret == 0) {
196                         rp->r_pc = npc;
197                         return;
198                 }
199
200                 /*
201                  * We need to wait until after we've called the
202                  * dtrace_return_probe_ptr function pointer to set %pc.
203                  */
204                 rwp = &CPU->cpu_ft_lock;
205                 rw_enter(rwp, RW_READER);
206                 if (dtrace_return_probe_ptr != NULL)
207                         (void) (*dtrace_return_probe_ptr)(rp);
208                 rw_exit(rwp);
209                 rp->r_pc = npc;
210
211         } else if (rp->r_trapno == T_DTRACE_PROBE) {
212                 rwp = &CPU->cpu_ft_lock;
213                 rw_enter(rwp, RW_READER);
214                 if (dtrace_fasttrap_probe_ptr != NULL)
215                         (void) (*dtrace_fasttrap_probe_ptr)(rp);
216                 rw_exit(rwp);
217
218         } else if (rp->r_trapno == T_BPTFLT) {
219                 uint8_t instr;
220                 rwp = &CPU->cpu_ft_lock;
221
222                 /*
223                  * The DTrace fasttrap provider uses the breakpoint trap
224                  * (int 3). We let DTrace take the first crack at handling
225                  * this trap; if it's not a probe that DTrace knowns about,
226                  * we call into the trap() routine to handle it like a
227                  * breakpoint placed by a conventional debugger.
228                  */
229                 rw_enter(rwp, RW_READER);
230                 if (dtrace_pid_probe_ptr != NULL &&
231                     (*dtrace_pid_probe_ptr)(rp) == 0) {
232                         rw_exit(rwp);
233                         return;
234                 }
235                 rw_exit(rwp);
236
237                 /*
238                  * If the instruction that caused the breakpoint trap doesn't
239                  * look like an int 3 anymore, it may be that this tracepoint
240                  * was removed just after the user thread executed it. In
241                  * that case, return to user land to retry the instuction.
242                  */
243                 if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
244                     instr != FASTTRAP_INSTR) {
245                         rp->r_pc--;
246                         return;
247                 }
248
249                 trap(rp, addr, cpuid);
250
251         } else {
252                 trap(rp, addr, cpuid);
253         }
254 }
255
256 void
257 dtrace_safe_synchronous_signal(void)
258 {
259         kthread_t *t = curthread;
260         struct regs *rp = lwptoregs(ttolwp(t));
261         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
262
263         ASSERT(t->t_dtrace_on);
264
265         /*
266          * If we're not in the range of scratch addresses, we're not actually
267          * tracing user instructions so turn off the flags. If the instruction
268          * we copied out caused a synchonous trap, reset the pc back to its
269          * original value and turn off the flags.
270          */
271         if (rp->r_pc < t->t_dtrace_scrpc ||
272             rp->r_pc > t->t_dtrace_astpc + isz) {
273                 t->t_dtrace_ft = 0;
274         } else if (rp->r_pc == t->t_dtrace_scrpc ||
275             rp->r_pc == t->t_dtrace_astpc) {
276                 rp->r_pc = t->t_dtrace_pc;
277                 t->t_dtrace_ft = 0;
278         }
279 }
280
281 int
282 dtrace_safe_defer_signal(void)
283 {
284         kthread_t *t = curthread;
285         struct regs *rp = lwptoregs(ttolwp(t));
286         size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
287
288         ASSERT(t->t_dtrace_on);
289
290         /*
291          * If we're not in the range of scratch addresses, we're not actually
292          * tracing user instructions so turn off the flags.
293          */
294         if (rp->r_pc < t->t_dtrace_scrpc ||
295             rp->r_pc > t->t_dtrace_astpc + isz) {
296                 t->t_dtrace_ft = 0;
297                 return (0);
298         }
299
300         /*
301          * If we've executed the original instruction, but haven't performed
302          * the jmp back to t->t_dtrace_npc or the clean up of any registers
303          * used to emulate %rip-relative instructions in 64-bit mode, do that
304          * here and take the signal right away. We detect this condition by
305          * seeing if the program counter is the range [scrpc + isz, astpc).
306          */
307         if (t->t_dtrace_astpc - rp->r_pc <
308             t->t_dtrace_astpc - t->t_dtrace_scrpc - isz) {
309 #ifdef __amd64
310                 /*
311                  * If there is a scratch register and we're on the
312                  * instruction immediately after the modified instruction,
313                  * restore the value of that scratch register.
314                  */
315                 if (t->t_dtrace_reg != 0 &&
316                     rp->r_pc == t->t_dtrace_scrpc + isz) {
317                         switch (t->t_dtrace_reg) {
318                         case REG_RAX:
319                                 rp->r_rax = t->t_dtrace_regv;
320                                 break;
321                         case REG_RCX:
322                                 rp->r_rcx = t->t_dtrace_regv;
323                                 break;
324                         case REG_R8:
325                                 rp->r_r8 = t->t_dtrace_regv;
326                                 break;
327                         case REG_R9:
328                                 rp->r_r9 = t->t_dtrace_regv;
329                                 break;
330                         }
331                 }
332 #endif
333                 rp->r_pc = t->t_dtrace_npc;
334                 t->t_dtrace_ft = 0;
335                 return (0);
336         }
337
338         /*
339          * Otherwise, make sure we'll return to the kernel after executing
340          * the copied out instruction and defer the signal.
341          */
342         if (!t->t_dtrace_step) {
343                 ASSERT(rp->r_pc < t->t_dtrace_astpc);
344                 rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
345                 t->t_dtrace_step = 1;
346         }
347
348         t->t_dtrace_ast = 1;
349
350         return (1);
351 }
352 #endif
353
354 static int64_t  tgt_cpu_tsc;
355 static int64_t  hst_cpu_tsc;
356 static int64_t  tsc_skew[MAXCPU];
357 static uint64_t nsec_scale;
358
359 /* See below for the explanation of this macro. */
360 #define SCALE_SHIFT     28
361
362 static void
363 dtrace_gethrtime_init_cpu(void *arg)
364 {
365         uintptr_t cpu = (uintptr_t) arg;
366
367         if (cpu == curcpu)
368                 tgt_cpu_tsc = rdtsc();
369         else
370                 hst_cpu_tsc = rdtsc();
371 }
372
373 static void
374 dtrace_gethrtime_init(void *arg)
375 {
376         cpuset_t map;
377         struct pcpu *pc;
378         uint64_t tsc_f;
379         int i;
380
381         /*
382          * Get TSC frequency known at this moment.
383          * This should be constant if TSC is invariant.
384          * Otherwise tick->time conversion will be inaccurate, but
385          * will preserve monotonic property of TSC.
386          */
387         tsc_f = atomic_load_acq_64(&tsc_freq);
388
389         /*
390          * The following line checks that nsec_scale calculated below
391          * doesn't overflow 32-bit unsigned integer, so that it can multiply
392          * another 32-bit integer without overflowing 64-bit.
393          * Thus minimum supported TSC frequency is 62.5MHz.
394          */
395         KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), ("TSC frequency is too low"));
396
397         /*
398          * We scale up NANOSEC/tsc_f ratio to preserve as much precision
399          * as possible.
400          * 2^28 factor was chosen quite arbitrarily from practical
401          * considerations:
402          * - it supports TSC frequencies as low as 62.5MHz (see above);
403          * - it provides quite good precision (e < 0.01%) up to THz
404          *   (terahertz) values;
405          */
406         nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
407
408         /* The current CPU is the reference one. */
409         sched_pin();
410         tsc_skew[curcpu] = 0;
411         CPU_FOREACH(i) {
412                 if (i == curcpu)
413                         continue;
414
415                 pc = pcpu_find(i);
416                 CPU_SETOF(PCPU_GET(cpuid), &map);
417                 CPU_SET(pc->pc_cpuid, &map);
418
419                 smp_rendezvous_cpus(map, NULL,
420                     dtrace_gethrtime_init_cpu,
421                     smp_no_rendevous_barrier, (void *)(uintptr_t) i);
422
423                 tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
424         }
425         sched_unpin();
426 }
427
428 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
429
430 /*
431  * DTrace needs a high resolution time function which can
432  * be called from a probe context and guaranteed not to have
433  * instrumented with probes itself.
434  *
435  * Returns nanoseconds since boot.
436  */
437 uint64_t
438 dtrace_gethrtime()
439 {
440         uint64_t tsc;
441         uint32_t lo;
442         uint32_t hi;
443
444         /*
445          * We split TSC value into lower and higher 32-bit halves and separately
446          * scale them with nsec_scale, then we scale them down by 2^28
447          * (see nsec_scale calculations) taking into account 32-bit shift of
448          * the higher half and finally add.
449          */
450         tsc = rdtsc() + tsc_skew[curcpu];
451         lo = tsc;
452         hi = tsc >> 32;
453         return (((lo * nsec_scale) >> SCALE_SHIFT) +
454             ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
455 }
456
457 uint64_t
458 dtrace_gethrestime(void)
459 {
460         printf("%s(%d): XXX\n",__func__,__LINE__);
461         return (0);
462 }
463
464 /* Function to handle DTrace traps during probes. See i386/i386/trap.c */
465 int
466 dtrace_trap(struct trapframe *frame, u_int type)
467 {
468         /*
469          * A trap can occur while DTrace executes a probe. Before
470          * executing the probe, DTrace blocks re-scheduling and sets
471          * a flag in it's per-cpu flags to indicate that it doesn't
472          * want to fault. On returning from the probe, the no-fault
473          * flag is cleared and finally re-scheduling is enabled.
474          *
475          * Check if DTrace has enabled 'no-fault' mode:
476          *
477          */
478         if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
479                 /*
480                  * There are only a couple of trap types that are expected.
481                  * All the rest will be handled in the usual way.
482                  */
483                 switch (type) {
484                 /* General protection fault. */
485                 case T_PROTFLT:
486                         /* Flag an illegal operation. */
487                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
488
489                         /*
490                          * Offset the instruction pointer to the instruction
491                          * following the one causing the fault.
492                          */
493                         frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
494                         return (1);
495                 /* Page fault. */
496                 case T_PAGEFLT:
497                         /* Flag a bad address. */
498                         cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
499                         cpu_core[curcpu].cpuc_dtrace_illval = rcr2();
500
501                         /*
502                          * Offset the instruction pointer to the instruction
503                          * following the one causing the fault.
504                          */
505                         frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
506                         return (1);
507                 default:
508                         /* Handle all other traps in the usual way. */
509                         break;
510                 }
511         }
512
513         /* Handle the trap in the usual way. */
514         return (0);
515 }