xref: /netbsd-src/external/cddl/osnet/dev/dtrace/amd64/dtrace_subr.c (revision 6a493d6bc668897c91594964a732d38505b70cbb)
1 /*	$NetBSD: dtrace_subr.c,v 1.7 2012/06/16 17:31:47 chs Exp $	*/
2 
3 /*
4  * CDDL HEADER START
5  *
6  * The contents of this file are subject to the terms of the
7  * Common Development and Distribution License, Version 1.0 only
8  * (the "License").  You may not use this file except in compliance
9  * with the License.
10  *
11  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
12  * or http://www.opensolaris.org/os/licensing.
13  * See the License for the specific language governing permissions
14  * and limitations under the License.
15  *
16  * When distributing Covered Code, include this CDDL HEADER in each
17  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
18  * If applicable, add the following below this CDDL HEADER, with the
19  * fields enclosed by brackets "[]" replaced with your own identifying
20  * information: Portions Copyright [yyyy] [name of copyright owner]
21  *
22  * CDDL HEADER END
23  *
24  * $FreeBSD: src/sys/cddl/dev/dtrace/amd64/dtrace_subr.c,v 1.3.2.1 2009/08/03 08:13:06 kensmith Exp $
25  *
26  */
27 /*
28  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
29  * Use is subject to license terms.
30  */
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/types.h>
35 #include <sys/kernel.h>
36 #include <sys/malloc.h>
37 #include <sys/kmem.h>
38 #include <sys/xcall.h>
39 #include <sys/cpu.h>
40 #include <sys/cpuvar.h>
41 //#include <sys/smp.h>
42 #include <sys/dtrace_impl.h>
43 #include <sys/dtrace_bsd.h>
44 #include <machine/frame.h>
45 #include <machine/cpu_counter.h>
46 #include <uvm/uvm_pglist.h>
47 #include <uvm/uvm_prot.h>
48 #include <uvm/uvm_pmap.h>
49 
50 extern uintptr_t 	kernelbase;
51 extern uintptr_t 	dtrace_in_probe_addr;
52 extern int		dtrace_in_probe;
53 
54 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
55 
56 typedef struct dtrace_invop_hdlr {
57 	int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
58 	struct dtrace_invop_hdlr *dtih_next;
59 } dtrace_invop_hdlr_t;
60 
61 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
62 void dtrace_gethrtime_init(void *);
63 
64 
65 int
66 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
67 {
68 	dtrace_invop_hdlr_t *hdlr;
69 	int rval;
70 
71 	for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
72 		if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
73 			return (rval);
74 
75 	return (0);
76 }
77 
78 void
79 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
80 {
81 	dtrace_invop_hdlr_t *hdlr;
82 
83 	hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
84 	hdlr->dtih_func = func;
85 	hdlr->dtih_next = dtrace_invop_hdlr;
86 	dtrace_invop_hdlr = hdlr;
87 }
88 
89 void
90 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
91 {
92 	dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
93 
94 	for (;;) {
95 		if (hdlr == NULL)
96 			panic("attempt to remove non-existent invop handler");
97 
98 		if (hdlr->dtih_func == func)
99 			break;
100 
101 		prev = hdlr;
102 		hdlr = hdlr->dtih_next;
103 	}
104 
105 	if (prev == NULL) {
106 		ASSERT(dtrace_invop_hdlr == hdlr);
107 		dtrace_invop_hdlr = hdlr->dtih_next;
108 	} else {
109 		ASSERT(dtrace_invop_hdlr != hdlr);
110 		prev->dtih_next = hdlr->dtih_next;
111 	}
112 
113 	kmem_free(hdlr, sizeof (dtrace_invop_hdlr_t));
114 }
115 
116 /*ARGSUSED*/
117 void
118 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
119 {
120 	(*func)(0, kernelbase);
121 }
122 
123 static void
124 xcall_func(void *arg0, void *arg1)
125 {
126     	dtrace_xcall_t func = arg0;
127 
128     	(*func)(arg1);
129 }
130 
131 void
132 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
133 {
134 	uint64_t where;
135 
136 	if (cpu == DTRACE_CPUALL) {
137 		where = xc_broadcast(0, xcall_func, func, arg);
138 	} else {
139 		struct cpu_info *cinfo = cpu_lookup(cpu);
140 
141 		KASSERT(cinfo != NULL);
142 		where = xc_unicast(0, xcall_func, func, arg, cinfo);
143 	}
144 	xc_wait(where);
145 
146 	/* XXX Q. Do we really need the other cpus to wait also?
147 	 * (see solaris:xc_sync())
148 	 */
149 }
150 
151 static void
152 dtrace_sync_func(void)
153 {
154 }
155 
156 void
157 dtrace_sync(void)
158 {
159         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
160 }
161 
162 #ifdef notyet
163 int (*dtrace_fasttrap_probe_ptr)(struct regs *);
164 int (*dtrace_pid_probe_ptr)(struct regs *);
165 int (*dtrace_return_probe_ptr)(struct regs *);
166 
167 void
168 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
169 {
170 	krwlock_t *rwp;
171 	proc_t *p = curproc;
172 	extern void trap(struct regs *, caddr_t, processorid_t);
173 
174 	if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
175 		if (curthread->t_cred != p->p_cred) {
176 			cred_t *oldcred = curthread->t_cred;
177 			/*
178 			 * DTrace accesses t_cred in probe context.  t_cred
179 			 * must always be either NULL, or point to a valid,
180 			 * allocated cred structure.
181 			 */
182 			curthread->t_cred = crgetcred();
183 			crfree(oldcred);
184 		}
185 	}
186 
187 	if (rp->r_trapno == T_DTRACE_RET) {
188 		uint8_t step = curthread->t_dtrace_step;
189 		uint8_t ret = curthread->t_dtrace_ret;
190 		uintptr_t npc = curthread->t_dtrace_npc;
191 
192 		if (curthread->t_dtrace_ast) {
193 			aston(curthread);
194 			curthread->t_sig_check = 1;
195 		}
196 
197 		/*
198 		 * Clear all user tracing flags.
199 		 */
200 		curthread->t_dtrace_ft = 0;
201 
202 		/*
203 		 * If we weren't expecting to take a return probe trap, kill
204 		 * the process as though it had just executed an unassigned
205 		 * trap instruction.
206 		 */
207 		if (step == 0) {
208 			tsignal(curthread, SIGILL);
209 			return;
210 		}
211 
212 		/*
213 		 * If we hit this trap unrelated to a return probe, we're
214 		 * just here to reset the AST flag since we deferred a signal
215 		 * until after we logically single-stepped the instruction we
216 		 * copied out.
217 		 */
218 		if (ret == 0) {
219 			rp->r_pc = npc;
220 			return;
221 		}
222 
223 		/*
224 		 * We need to wait until after we've called the
225 		 * dtrace_return_probe_ptr function pointer to set %pc.
226 		 */
227 		rwp = &CPU->cpu_ft_lock;
228 		rw_enter(rwp, RW_READER);
229 		if (dtrace_return_probe_ptr != NULL)
230 			(void) (*dtrace_return_probe_ptr)(rp);
231 		rw_exit(rwp);
232 		rp->r_pc = npc;
233 
234 	} else if (rp->r_trapno == T_DTRACE_PROBE) {
235 		rwp = &CPU->cpu_ft_lock;
236 		rw_enter(rwp, RW_READER);
237 		if (dtrace_fasttrap_probe_ptr != NULL)
238 			(void) (*dtrace_fasttrap_probe_ptr)(rp);
239 		rw_exit(rwp);
240 
241 	} else if (rp->r_trapno == T_BPTFLT) {
242 		uint8_t instr;
243 		rwp = &CPU->cpu_ft_lock;
244 
245 		/*
246 		 * The DTrace fasttrap provider uses the breakpoint trap
247 		 * (int 3). We let DTrace take the first crack at handling
248 		 * this trap; if it's not a probe that DTrace knowns about,
249 		 * we call into the trap() routine to handle it like a
250 		 * breakpoint placed by a conventional debugger.
251 		 */
252 		rw_enter(rwp, RW_READER);
253 		if (dtrace_pid_probe_ptr != NULL &&
254 		    (*dtrace_pid_probe_ptr)(rp) == 0) {
255 			rw_exit(rwp);
256 			return;
257 		}
258 		rw_exit(rwp);
259 
260 		/*
261 		 * If the instruction that caused the breakpoint trap doesn't
262 		 * look like an int 3 anymore, it may be that this tracepoint
263 		 * was removed just after the user thread executed it. In
264 		 * that case, return to user land to retry the instuction.
265 		 */
266 		if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
267 		    instr != FASTTRAP_INSTR) {
268 			rp->r_pc--;
269 			return;
270 		}
271 
272 		trap(rp, addr, cpuid);
273 
274 	} else {
275 		trap(rp, addr, cpuid);
276 	}
277 }
278 
279 void
280 dtrace_safe_synchronous_signal(void)
281 {
282 	kthread_t *t = curthread;
283 	struct regs *rp = lwptoregs(ttolwp(t));
284 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
285 
286 	ASSERT(t->t_dtrace_on);
287 
288 	/*
289 	 * If we're not in the range of scratch addresses, we're not actually
290 	 * tracing user instructions so turn off the flags. If the instruction
291 	 * we copied out caused a synchonous trap, reset the pc back to its
292 	 * original value and 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 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
298 	    rp->r_pc == t->t_dtrace_astpc) {
299 		rp->r_pc = t->t_dtrace_pc;
300 		t->t_dtrace_ft = 0;
301 	}
302 }
303 
304 int
305 dtrace_safe_defer_signal(void)
306 {
307 	kthread_t *t = curthread;
308 	struct regs *rp = lwptoregs(ttolwp(t));
309 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
310 
311 	ASSERT(t->t_dtrace_on);
312 
313 	/*
314 	 * If we're not in the range of scratch addresses, we're not actually
315 	 * tracing user instructions so turn off the flags.
316 	 */
317 	if (rp->r_pc < t->t_dtrace_scrpc ||
318 	    rp->r_pc > t->t_dtrace_astpc + isz) {
319 		t->t_dtrace_ft = 0;
320 		return (0);
321 	}
322 
323 	/*
324 	 * If we've executed the original instruction, but haven't performed
325 	 * the jmp back to t->t_dtrace_npc or the clean up of any registers
326 	 * used to emulate %rip-relative instructions in 64-bit mode, do that
327 	 * here and take the signal right away. We detect this condition by
328 	 * seeing if the program counter is the range [scrpc + isz, astpc).
329 	 */
330 	if (t->t_dtrace_astpc - rp->r_pc <
331 	    t->t_dtrace_astpc - t->t_dtrace_scrpc - isz) {
332 #ifdef __amd64
333 		/*
334 		 * If there is a scratch register and we're on the
335 		 * instruction immediately after the modified instruction,
336 		 * restore the value of that scratch register.
337 		 */
338 		if (t->t_dtrace_reg != 0 &&
339 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
340 			switch (t->t_dtrace_reg) {
341 			case REG_RAX:
342 				rp->r_rax = t->t_dtrace_regv;
343 				break;
344 			case REG_RCX:
345 				rp->r_rcx = t->t_dtrace_regv;
346 				break;
347 			case REG_R8:
348 				rp->r_r8 = t->t_dtrace_regv;
349 				break;
350 			case REG_R9:
351 				rp->r_r9 = t->t_dtrace_regv;
352 				break;
353 			}
354 		}
355 #endif
356 		rp->r_pc = t->t_dtrace_npc;
357 		t->t_dtrace_ft = 0;
358 		return (0);
359 	}
360 
361 	/*
362 	 * Otherwise, make sure we'll return to the kernel after executing
363 	 * the copied out instruction and defer the signal.
364 	 */
365 	if (!t->t_dtrace_step) {
366 		ASSERT(rp->r_pc < t->t_dtrace_astpc);
367 		rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
368 		t->t_dtrace_step = 1;
369 	}
370 
371 	t->t_dtrace_ast = 1;
372 
373 	return (1);
374 }
375 #endif
376 
377 #ifdef __NetBSD__
378 static __inline uint64_t
379 dtrace_rdtsc(void)
380 {
381 	uint32_t hi, lo;
382 
383 	__asm volatile("rdtsc" : "=d" (hi), "=a" (lo));
384 	return (((uint64_t)hi << 32) | (uint64_t) lo);
385 }
386 #define rdtsc dtrace_rdtsc
387 #endif
388 
389 #ifdef notyet
390 static int64_t	tgt_cpu_tsc;
391 static int64_t	hst_cpu_tsc;
392 #endif
393 static int64_t	tsc_skew[MAXCPUS];
394 static uint64_t	nsec_scale;
395 
396 /* See below for the explanation of this macro. */
397 #define SCALE_SHIFT	28
398 
399 #ifdef notyet
400 static void
401 dtrace_gethrtime_init_sync(void *arg)
402 {
403 #ifdef CHECK_SYNC
404 	/*
405 	 * Delay this function from returning on one
406 	 * of the CPUs to check that the synchronisation
407 	 * works.
408 	 */
409 	uintptr_t cpu = (uintptr_t) arg;
410 
411 	if (cpu == cpu_number()) {
412 		int i;
413 		for (i = 0; i < 1000000000; i++)
414 			tgt_cpu_tsc = rdtsc();
415 		tgt_cpu_tsc = 0;
416 	}
417 #endif
418 }
419 
420 static void
421 dtrace_gethrtime_init_cpu(void *arg)
422 {
423 	uintptr_t cpu = (uintptr_t) arg;
424 
425 	if (cpu == cpu_number())
426 		tgt_cpu_tsc = rdtsc();
427 	else
428 		hst_cpu_tsc = rdtsc();
429 }
430 #endif
431 
432 void
433 dtrace_gethrtime_init(void *arg)
434 {
435 	uint64_t tsc_f;
436 	CPU_INFO_ITERATOR cpuind;
437 	struct cpu_info *cinfo = curcpu();
438 	cpuid_t cur_cpuid = cpu_number();	/* current cpu id */
439 
440 	/*
441 	 * Get TSC frequency known at this moment.
442 	 * This should be constant if TSC is invariant.
443 	 * Otherwise tick->time conversion will be inaccurate, but
444 	 * will preserve monotonic property of TSC.
445 	 */
446 	tsc_f = cpu_frequency(cinfo);
447 
448 	/*
449 	 * The following line checks that nsec_scale calculated below
450 	 * doesn't overflow 32-bit unsigned integer, so that it can multiply
451 	 * another 32-bit integer without overflowing 64-bit.
452 	 * Thus minimum supported TSC frequency is 62.5MHz.
453 	 */
454 	//KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)), ("TSC frequency is too low"));
455 	KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)));
456 
457 	/*
458 	 * We scale up NANOSEC/tsc_f ratio to preserve as much precision
459 	 * as possible.
460 	 * 2^28 factor was chosen quite arbitrarily from practical
461 	 * considerations:
462 	 * - it supports TSC frequencies as low as 62.5MHz (see above);
463 	 * - it provides quite good precision (e < 0.01%) up to THz
464 	 *   (terahertz) values;
465 	 */
466 	nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
467 
468 	/* The current CPU is the reference one. */
469 	tsc_skew[cur_cpuid] = 0;
470 
471 	for (CPU_INFO_FOREACH(cpuind, cinfo)) {
472 		/* use skew relative to cpu 0 */
473 		tsc_skew[cpu_index(cinfo)] = cinfo->ci_data.cpu_cc_skew;
474 	}
475 
476 	/* Already handled in x86/tsc.c for ci_data.cpu_cc_skew */
477 #if 0
478 	for (i = 0; i <= mp_maxid; i++) {
479 		if (i == curcpu)
480 			continue;
481 
482 		if (pcpu_find(i) == NULL)
483 			continue;
484 
485 		map = 0;
486 		map |= (1 << curcpu);
487 		map |= (1 << i);
488 
489 		smp_rendezvous_cpus(map, dtrace_gethrtime_init_sync,
490 		    dtrace_gethrtime_init_cpu,
491 		    smp_no_rendevous_barrier, (void *)(uintptr_t) i);
492 
493 		tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
494 	}
495 #endif
496 }
497 
498 /*
499  * DTrace needs a high resolution time function which can
500  * be called from a probe context and guaranteed not to have
501  * instrumented with probes itself.
502  *
503  * Returns nanoseconds since boot.
504  */
505 uint64_t
506 dtrace_gethrtime()
507 {
508 	uint64_t tsc;
509 	uint32_t lo;
510 	uint32_t hi;
511 
512 	/*
513 	 * We split TSC value into lower and higher 32-bit halves and separately
514 	 * scale them with nsec_scale, then we scale them down by 2^28
515 	 * (see nsec_scale calculations) taking into account 32-bit shift of
516 	 * the higher half and finally add.
517 	 */
518 	tsc = rdtsc() + tsc_skew[cpu_number()];
519 	lo = tsc;
520 	hi = tsc >> 32;
521 	return (((lo * nsec_scale) >> SCALE_SHIFT) +
522 	    ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
523 }
524 
525 uint64_t
526 dtrace_gethrestime(void)
527 {
528 	printf("%s(%d): XXX\n",__func__,__LINE__);
529 	return (0);
530 }
531 
532 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
533 int
534 dtrace_trap(struct trapframe *frame, u_int type)
535 {
536 	/*
537 	 * A trap can occur while DTrace executes a probe. Before
538 	 * executing the probe, DTrace blocks re-scheduling and sets
539 	 * a flag in it's per-cpu flags to indicate that it doesn't
540 	 * want to fault. On returning from the the probe, the no-fault
541 	 * flag is cleared and finally re-scheduling is enabled.
542 	 *
543 	 * Check if DTrace has enabled 'no-fault' mode:
544 	 *
545 	 */
546 	if ((cpu_core[cpu_number()].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
547 		/*
548 		 * There are only a couple of trap types that are expected.
549 		 * All the rest will be handled in the usual way.
550 		 */
551 		switch (type) {
552 		/* Privilieged instruction fault. */
553 		case T_PRIVINFLT:
554 			break;
555 		/* General protection fault. */
556 		case T_PROTFLT:
557 			/* Flag an illegal operation. */
558 			cpu_core[cpu_number()].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
559 
560 			/*
561 			 * Offset the instruction pointer to the instruction
562 			 * following the one causing the fault.
563 			 */
564 			frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
565 			return (1);
566 		/* Page fault. */
567 		case T_PAGEFLT:
568 			/* Flag a bad address. */
569 			cpu_core[cpu_number()].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
570 			cpu_core[cpu_number()].cpuc_dtrace_illval = rcr2();
571 
572 			/*
573 			 * Offset the instruction pointer to the instruction
574 			 * following the one causing the fault.
575 			 */
576 			frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
577 			return (1);
578 		default:
579 			/* Handle all other traps in the usual way. */
580 			break;
581 		}
582 	}
583 
584 	/* Handle the trap in the usual way. */
585 	return (0);
586 }
587