xref: /freebsd/sys/cddl/dev/dtrace/i386/dtrace_subr.c (revision b3aaa0cc21c63d388230c7ef2a80abd631ff20d5)
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 	kernelbase;
44 extern uintptr_t 	dtrace_in_probe_addr;
45 extern int		dtrace_in_probe;
46 
47 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
48 
49 typedef struct dtrace_invop_hdlr {
50 	int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
51 	struct dtrace_invop_hdlr *dtih_next;
52 } dtrace_invop_hdlr_t;
53 
54 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
55 
56 int
57 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
58 {
59 	dtrace_invop_hdlr_t *hdlr;
60 	int rval;
61 
62 	for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
63 		if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
64 			return (rval);
65 
66 	return (0);
67 }
68 
69 void
70 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
71 {
72 	dtrace_invop_hdlr_t *hdlr;
73 
74 	hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
75 	hdlr->dtih_func = func;
76 	hdlr->dtih_next = dtrace_invop_hdlr;
77 	dtrace_invop_hdlr = hdlr;
78 }
79 
80 void
81 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
82 {
83 	dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
84 
85 	for (;;) {
86 		if (hdlr == NULL)
87 			panic("attempt to remove non-existent invop handler");
88 
89 		if (hdlr->dtih_func == func)
90 			break;
91 
92 		prev = hdlr;
93 		hdlr = hdlr->dtih_next;
94 	}
95 
96 	if (prev == NULL) {
97 		ASSERT(dtrace_invop_hdlr == hdlr);
98 		dtrace_invop_hdlr = hdlr->dtih_next;
99 	} else {
100 		ASSERT(dtrace_invop_hdlr != hdlr);
101 		prev->dtih_next = hdlr->dtih_next;
102 	}
103 
104 	kmem_free(hdlr, 0);
105 }
106 
107 void
108 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
109 {
110 	(*func)(0, kernelbase);
111 }
112 
113 void
114 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
115 {
116 	cpumask_t cpus;
117 
118 	critical_enter();
119 
120 	if (cpu == DTRACE_CPUALL)
121 		cpus = all_cpus;
122 	else
123 		cpus = (cpumask_t) (1 << cpu);
124 
125 	/* If the current CPU is in the set, call the function directly: */
126 	if ((cpus & (1 << curcpu)) != 0) {
127 		(*func)(arg);
128 
129 		/* Mask the current CPU from the set */
130 		cpus &= ~(1 << curcpu);
131 	}
132 
133 	/* If there are any CPUs in the set, cross-call to those CPUs */
134 	if (cpus != 0)
135 		smp_rendezvous_cpus(cpus, NULL, func, smp_no_rendevous_barrier, arg);
136 
137 	critical_exit();
138 }
139 
140 static void
141 dtrace_sync_func(void)
142 {
143 }
144 
145 void
146 dtrace_sync(void)
147 {
148         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
149 }
150 
151 #ifdef notyet
152 int (*dtrace_fasttrap_probe_ptr)(struct regs *);
153 int (*dtrace_pid_probe_ptr)(struct regs *);
154 int (*dtrace_return_probe_ptr)(struct regs *);
155 
156 void
157 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
158 {
159 	krwlock_t *rwp;
160 	proc_t *p = curproc;
161 	extern void trap(struct regs *, caddr_t, processorid_t);
162 
163 	if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
164 		if (curthread->t_cred != p->p_cred) {
165 			cred_t *oldcred = curthread->t_cred;
166 			/*
167 			 * DTrace accesses t_cred in probe context.  t_cred
168 			 * must always be either NULL, or point to a valid,
169 			 * allocated cred structure.
170 			 */
171 			curthread->t_cred = crgetcred();
172 			crfree(oldcred);
173 		}
174 	}
175 
176 	if (rp->r_trapno == T_DTRACE_RET) {
177 		uint8_t step = curthread->t_dtrace_step;
178 		uint8_t ret = curthread->t_dtrace_ret;
179 		uintptr_t npc = curthread->t_dtrace_npc;
180 
181 		if (curthread->t_dtrace_ast) {
182 			aston(curthread);
183 			curthread->t_sig_check = 1;
184 		}
185 
186 		/*
187 		 * Clear all user tracing flags.
188 		 */
189 		curthread->t_dtrace_ft = 0;
190 
191 		/*
192 		 * If we weren't expecting to take a return probe trap, kill
193 		 * the process as though it had just executed an unassigned
194 		 * trap instruction.
195 		 */
196 		if (step == 0) {
197 			tsignal(curthread, SIGILL);
198 			return;
199 		}
200 
201 		/*
202 		 * If we hit this trap unrelated to a return probe, we're
203 		 * just here to reset the AST flag since we deferred a signal
204 		 * until after we logically single-stepped the instruction we
205 		 * copied out.
206 		 */
207 		if (ret == 0) {
208 			rp->r_pc = npc;
209 			return;
210 		}
211 
212 		/*
213 		 * We need to wait until after we've called the
214 		 * dtrace_return_probe_ptr function pointer to set %pc.
215 		 */
216 		rwp = &CPU->cpu_ft_lock;
217 		rw_enter(rwp, RW_READER);
218 		if (dtrace_return_probe_ptr != NULL)
219 			(void) (*dtrace_return_probe_ptr)(rp);
220 		rw_exit(rwp);
221 		rp->r_pc = npc;
222 
223 	} else if (rp->r_trapno == T_DTRACE_PROBE) {
224 		rwp = &CPU->cpu_ft_lock;
225 		rw_enter(rwp, RW_READER);
226 		if (dtrace_fasttrap_probe_ptr != NULL)
227 			(void) (*dtrace_fasttrap_probe_ptr)(rp);
228 		rw_exit(rwp);
229 
230 	} else if (rp->r_trapno == T_BPTFLT) {
231 		uint8_t instr;
232 		rwp = &CPU->cpu_ft_lock;
233 
234 		/*
235 		 * The DTrace fasttrap provider uses the breakpoint trap
236 		 * (int 3). We let DTrace take the first crack at handling
237 		 * this trap; if it's not a probe that DTrace knowns about,
238 		 * we call into the trap() routine to handle it like a
239 		 * breakpoint placed by a conventional debugger.
240 		 */
241 		rw_enter(rwp, RW_READER);
242 		if (dtrace_pid_probe_ptr != NULL &&
243 		    (*dtrace_pid_probe_ptr)(rp) == 0) {
244 			rw_exit(rwp);
245 			return;
246 		}
247 		rw_exit(rwp);
248 
249 		/*
250 		 * If the instruction that caused the breakpoint trap doesn't
251 		 * look like an int 3 anymore, it may be that this tracepoint
252 		 * was removed just after the user thread executed it. In
253 		 * that case, return to user land to retry the instuction.
254 		 */
255 		if (fuword8((void *)(rp->r_pc - 1), &instr) == 0 &&
256 		    instr != FASTTRAP_INSTR) {
257 			rp->r_pc--;
258 			return;
259 		}
260 
261 		trap(rp, addr, cpuid);
262 
263 	} else {
264 		trap(rp, addr, cpuid);
265 	}
266 }
267 
268 void
269 dtrace_safe_synchronous_signal(void)
270 {
271 	kthread_t *t = curthread;
272 	struct regs *rp = lwptoregs(ttolwp(t));
273 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
274 
275 	ASSERT(t->t_dtrace_on);
276 
277 	/*
278 	 * If we're not in the range of scratch addresses, we're not actually
279 	 * tracing user instructions so turn off the flags. If the instruction
280 	 * we copied out caused a synchonous trap, reset the pc back to its
281 	 * original value and turn off the flags.
282 	 */
283 	if (rp->r_pc < t->t_dtrace_scrpc ||
284 	    rp->r_pc > t->t_dtrace_astpc + isz) {
285 		t->t_dtrace_ft = 0;
286 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
287 	    rp->r_pc == t->t_dtrace_astpc) {
288 		rp->r_pc = t->t_dtrace_pc;
289 		t->t_dtrace_ft = 0;
290 	}
291 }
292 
293 int
294 dtrace_safe_defer_signal(void)
295 {
296 	kthread_t *t = curthread;
297 	struct regs *rp = lwptoregs(ttolwp(t));
298 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
299 
300 	ASSERT(t->t_dtrace_on);
301 
302 	/*
303 	 * If we're not in the range of scratch addresses, we're not actually
304 	 * tracing user instructions so turn off the flags.
305 	 */
306 	if (rp->r_pc < t->t_dtrace_scrpc ||
307 	    rp->r_pc > t->t_dtrace_astpc + isz) {
308 		t->t_dtrace_ft = 0;
309 		return (0);
310 	}
311 
312 	/*
313 	 * If we've executed the original instruction, but haven't performed
314 	 * the jmp back to t->t_dtrace_npc or the clean up of any registers
315 	 * used to emulate %rip-relative instructions in 64-bit mode, do that
316 	 * here and take the signal right away. We detect this condition by
317 	 * seeing if the program counter is the range [scrpc + isz, astpc).
318 	 */
319 	if (t->t_dtrace_astpc - rp->r_pc <
320 	    t->t_dtrace_astpc - t->t_dtrace_scrpc - isz) {
321 #ifdef __amd64
322 		/*
323 		 * If there is a scratch register and we're on the
324 		 * instruction immediately after the modified instruction,
325 		 * restore the value of that scratch register.
326 		 */
327 		if (t->t_dtrace_reg != 0 &&
328 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
329 			switch (t->t_dtrace_reg) {
330 			case REG_RAX:
331 				rp->r_rax = t->t_dtrace_regv;
332 				break;
333 			case REG_RCX:
334 				rp->r_rcx = t->t_dtrace_regv;
335 				break;
336 			case REG_R8:
337 				rp->r_r8 = t->t_dtrace_regv;
338 				break;
339 			case REG_R9:
340 				rp->r_r9 = t->t_dtrace_regv;
341 				break;
342 			}
343 		}
344 #endif
345 		rp->r_pc = t->t_dtrace_npc;
346 		t->t_dtrace_ft = 0;
347 		return (0);
348 	}
349 
350 	/*
351 	 * Otherwise, make sure we'll return to the kernel after executing
352 	 * the copied out instruction and defer the signal.
353 	 */
354 	if (!t->t_dtrace_step) {
355 		ASSERT(rp->r_pc < t->t_dtrace_astpc);
356 		rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
357 		t->t_dtrace_step = 1;
358 	}
359 
360 	t->t_dtrace_ast = 1;
361 
362 	return (1);
363 }
364 #endif
365 
366 static int64_t	tgt_cpu_tsc;
367 static int64_t	hst_cpu_tsc;
368 static int64_t	tsc_skew[MAXCPU];
369 
370 static void
371 dtrace_gethrtime_init_sync(void *arg)
372 {
373 #ifdef CHECK_SYNC
374 	/*
375 	 * Delay this function from returning on one
376 	 * of the CPUs to check that the synchronisation
377 	 * works.
378 	 */
379 	uintptr_t cpu = (uintptr_t) arg;
380 
381 	if (cpu == curcpu) {
382 		int i;
383 		for (i = 0; i < 1000000000; i++)
384 			tgt_cpu_tsc = rdtsc();
385 		tgt_cpu_tsc = 0;
386 	}
387 #endif
388 }
389 
390 static void
391 dtrace_gethrtime_init_cpu(void *arg)
392 {
393 	uintptr_t cpu = (uintptr_t) arg;
394 
395 	if (cpu == curcpu)
396 		tgt_cpu_tsc = rdtsc();
397 	else
398 		hst_cpu_tsc = rdtsc();
399 }
400 
401 static void
402 dtrace_gethrtime_init(void *arg)
403 {
404 	cpumask_t map;
405 	int i;
406 
407 	/* The current CPU is the reference one. */
408 	tsc_skew[curcpu] = 0;
409 
410 	for (i = 0; i <= mp_maxid; i++) {
411 		if (i == curcpu)
412 			continue;
413 
414 		if (pcpu_find(i) == NULL)
415 			continue;
416 
417 		map = 0;
418 		map |= (1 << curcpu);
419 		map |= (1 << i);
420 
421 		smp_rendezvous_cpus(map, dtrace_gethrtime_init_sync,
422 		    dtrace_gethrtime_init_cpu,
423 		    smp_no_rendevous_barrier, (void *)(uintptr_t) i);
424 
425 		tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
426 	}
427 }
428 
429 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init, NULL);
430 
431 /*
432  * DTrace needs a high resolution time function which can
433  * be called from a probe context and guaranteed not to have
434  * instrumented with probes itself.
435  *
436  * Returns nanoseconds since boot.
437  */
438 uint64_t
439 dtrace_gethrtime()
440 {
441 	return ((rdtsc() + tsc_skew[curcpu]) * (int64_t) 1000000000 / tsc_freq);
442 }
443 
444 uint64_t
445 dtrace_gethrestime(void)
446 {
447 	printf("%s(%d): XXX\n",__func__,__LINE__);
448 	return (0);
449 }
450 
451 /* Function to handle DTrace traps during probes. See i386/i386/trap.c */
452 int
453 dtrace_trap(struct trapframe *frame, u_int type)
454 {
455 	/*
456 	 * A trap can occur while DTrace executes a probe. Before
457 	 * executing the probe, DTrace blocks re-scheduling and sets
458 	 * a flag in it's per-cpu flags to indicate that it doesn't
459 	 * want to fault. On returning from the the probe, the no-fault
460 	 * flag is cleared and finally re-scheduling is enabled.
461 	 *
462 	 * Check if DTrace has enabled 'no-fault' mode:
463 	 *
464 	 */
465 	if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
466 		/*
467 		 * There are only a couple of trap types that are expected.
468 		 * All the rest will be handled in the usual way.
469 		 */
470 		switch (type) {
471 		/* General protection fault. */
472 		case T_PROTFLT:
473 			/* Flag an illegal operation. */
474 			cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
475 
476 			/*
477 			 * Offset the instruction pointer to the instruction
478 			 * following the one causing the fault.
479 			 */
480 			frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
481 			return (1);
482 		/* Page fault. */
483 		case T_PAGEFLT:
484 			/* Flag a bad address. */
485 			cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
486 			cpu_core[curcpu].cpuc_dtrace_illval = rcr2();
487 
488 			/*
489 			 * Offset the instruction pointer to the instruction
490 			 * following the one causing the fault.
491 			 */
492 			frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
493 			return (1);
494 		default:
495 			/* Handle all other traps in the usual way. */
496 			break;
497 		}
498 	}
499 
500 	/* Handle the trap in the usual way. */
501 	return (0);
502 }
503