xref: /illumos-gate/usr/src/uts/i86pc/os/dtrace_subr.c (revision ce7a304d880a0cf033d8b45bfaf65dfdec0d3faf)
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 (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Copyright (c) 2011, Joyent, Inc. All rights reserved.
29  */
30 
31 #include <sys/dtrace.h>
32 #include <sys/fasttrap.h>
33 #include <sys/x_call.h>
34 #include <sys/cmn_err.h>
35 #include <sys/trap.h>
36 #include <sys/psw.h>
37 #include <sys/privregs.h>
38 #include <sys/machsystm.h>
39 #include <vm/seg_kmem.h>
40 
41 typedef struct dtrace_invop_hdlr {
42 	int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
43 	struct dtrace_invop_hdlr *dtih_next;
44 } dtrace_invop_hdlr_t;
45 
46 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
47 
48 int
49 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
50 {
51 	dtrace_invop_hdlr_t *hdlr;
52 	int rval;
53 
54 	for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next) {
55 		if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
56 			return (rval);
57 	}
58 
59 	return (0);
60 }
61 
62 void
63 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
64 {
65 	dtrace_invop_hdlr_t *hdlr;
66 
67 	hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
68 	hdlr->dtih_func = func;
69 	hdlr->dtih_next = dtrace_invop_hdlr;
70 	dtrace_invop_hdlr = hdlr;
71 }
72 
73 void
74 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
75 {
76 	dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
77 
78 	for (;;) {
79 		if (hdlr == NULL)
80 			panic("attempt to remove non-existent invop handler");
81 
82 		if (hdlr->dtih_func == func)
83 			break;
84 
85 		prev = hdlr;
86 		hdlr = hdlr->dtih_next;
87 	}
88 
89 	if (prev == NULL) {
90 		ASSERT(dtrace_invop_hdlr == hdlr);
91 		dtrace_invop_hdlr = hdlr->dtih_next;
92 	} else {
93 		ASSERT(dtrace_invop_hdlr != hdlr);
94 		prev->dtih_next = hdlr->dtih_next;
95 	}
96 
97 	kmem_free(hdlr, sizeof (dtrace_invop_hdlr_t));
98 }
99 
100 int
101 dtrace_getipl(void)
102 {
103 	return (CPU->cpu_pri);
104 }
105 
106 /*ARGSUSED*/
107 void
108 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
109 {
110 	extern uintptr_t toxic_addr;
111 	extern size_t toxic_size;
112 
113 	(*func)(0, _userlimit);
114 
115 	if (hole_end > hole_start)
116 		(*func)(hole_start, hole_end);
117 	(*func)(toxic_addr, toxic_addr + toxic_size);
118 	(*func)(0, _userlimit);
119 }
120 
121 static int
122 dtrace_xcall_func(xc_arg_t arg1, xc_arg_t arg2, xc_arg_t arg3 __unused)
123 {
124 	dtrace_xcall_t func = (dtrace_xcall_t)arg1;
125 	(*func)((void*)arg2);
126 
127 	return (0);
128 }
129 
130 /*ARGSUSED*/
131 void
132 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
133 {
134 	cpuset_t set;
135 
136 	CPUSET_ZERO(set);
137 
138 	if (cpu == DTRACE_CPUALL) {
139 		CPUSET_ALL(set);
140 	} else {
141 		CPUSET_ADD(set, cpu);
142 	}
143 
144 	kpreempt_disable();
145 	xc_sync((xc_arg_t)func, (xc_arg_t)arg, 0, CPUSET2BV(set),
146 	    dtrace_xcall_func);
147 	kpreempt_enable();
148 }
149 
150 void
151 dtrace_sync_func(void)
152 {}
153 
154 void
155 dtrace_sync(void)
156 {
157 	dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
158 }
159 
160 int (*dtrace_pid_probe_ptr)(struct regs *);
161 int (*dtrace_return_probe_ptr)(struct regs *);
162 
163 void
164 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
165 {
166 	krwlock_t *rwp;
167 	proc_t *p = curproc;
168 	extern void trap(struct regs *, caddr_t, processorid_t);
169 
170 	if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
171 		if (curthread->t_cred != p->p_cred) {
172 			cred_t *oldcred = curthread->t_cred;
173 			/*
174 			 * DTrace accesses t_cred in probe context.  t_cred
175 			 * must always be either NULL, or point to a valid,
176 			 * allocated cred structure.
177 			 */
178 			curthread->t_cred = crgetcred();
179 			crfree(oldcred);
180 		}
181 	}
182 
183 	if (rp->r_trapno == T_DTRACE_RET) {
184 		uint8_t step = curthread->t_dtrace_step;
185 		uint8_t ret = curthread->t_dtrace_ret;
186 		uintptr_t npc = curthread->t_dtrace_npc;
187 
188 		if (curthread->t_dtrace_ast) {
189 			aston(curthread);
190 			curthread->t_sig_check = 1;
191 		}
192 
193 		/*
194 		 * Clear all user tracing flags.
195 		 */
196 		curthread->t_dtrace_ft = 0;
197 
198 		/*
199 		 * If we weren't expecting to take a return probe trap, kill
200 		 * the process as though it had just executed an unassigned
201 		 * trap instruction.
202 		 */
203 		if (step == 0) {
204 			tsignal(curthread, SIGILL);
205 			return;
206 		}
207 
208 		/*
209 		 * If we hit this trap unrelated to a return probe, we're
210 		 * just here to reset the AST flag since we deferred a signal
211 		 * until after we logically single-stepped the instruction we
212 		 * copied out.
213 		 */
214 		if (ret == 0) {
215 			rp->r_pc = npc;
216 			return;
217 		}
218 
219 		/*
220 		 * We need to wait until after we've called the
221 		 * dtrace_return_probe_ptr function pointer to set %pc.
222 		 */
223 		rwp = &CPU->cpu_ft_lock;
224 		rw_enter(rwp, RW_READER);
225 		if (dtrace_return_probe_ptr != NULL)
226 			(void) (*dtrace_return_probe_ptr)(rp);
227 		rw_exit(rwp);
228 		rp->r_pc = npc;
229 
230 	} else if (rp->r_trapno == T_BPTFLT) {
231 		uint8_t instr, instr2;
232 		caddr_t linearpc;
233 		rwp = &CPU->cpu_ft_lock;
234 
235 		/*
236 		 * The DTrace fasttrap provider uses the breakpoint trap
237 		 * (int 3). We let DTrace take the first crack at handling
238 		 * this trap; if it's not a probe that DTrace knowns about,
239 		 * we call into the trap() routine to handle it like a
240 		 * breakpoint placed by a conventional debugger.
241 		 */
242 		rw_enter(rwp, RW_READER);
243 		if (dtrace_pid_probe_ptr != NULL &&
244 		    (*dtrace_pid_probe_ptr)(rp) == 0) {
245 			rw_exit(rwp);
246 			return;
247 		}
248 		rw_exit(rwp);
249 
250 		if (dtrace_linear_pc(rp, p, &linearpc) != 0) {
251 			trap(rp, addr, cpuid);
252 			return;
253 		}
254 
255 		/*
256 		 * If the instruction that caused the breakpoint trap doesn't
257 		 * look like an int 3 anymore, it may be that this tracepoint
258 		 * was removed just after the user thread executed it. In
259 		 * that case, return to user land to retry the instuction.
260 		 * Note that we assume the length of the instruction to retry
261 		 * is 1 byte because that's the length of FASTTRAP_INSTR.
262 		 * We check for r_pc > 0 and > 2 so that we don't have to
263 		 * deal with segment wraparound.
264 		 */
265 		if (rp->r_pc > 0 && fuword8(linearpc - 1, &instr) == 0 &&
266 		    instr != FASTTRAP_INSTR &&
267 		    (instr != 3 || (rp->r_pc >= 2 &&
268 		    (fuword8(linearpc - 2, &instr2) != 0 || instr2 != 0xCD)))) {
269 			rp->r_pc--;
270 			return;
271 		}
272 
273 		trap(rp, addr, cpuid);
274 
275 	} else {
276 		trap(rp, addr, cpuid);
277 	}
278 }
279 
280 void
281 dtrace_safe_synchronous_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. If the instruction
292 	 * we copied out caused a synchonous trap, reset the pc back to its
293 	 * original value and turn off the flags.
294 	 */
295 	if (rp->r_pc < t->t_dtrace_scrpc ||
296 	    rp->r_pc > t->t_dtrace_astpc + isz) {
297 		t->t_dtrace_ft = 0;
298 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
299 	    rp->r_pc == t->t_dtrace_astpc) {
300 		rp->r_pc = t->t_dtrace_pc;
301 		t->t_dtrace_ft = 0;
302 	}
303 }
304 
305 int
306 dtrace_safe_defer_signal(void)
307 {
308 	kthread_t *t = curthread;
309 	struct regs *rp = lwptoregs(ttolwp(t));
310 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
311 
312 	ASSERT(t->t_dtrace_on);
313 
314 	/*
315 	 * If we're not in the range of scratch addresses, we're not actually
316 	 * tracing user instructions so turn off the flags.
317 	 */
318 	if (rp->r_pc < t->t_dtrace_scrpc ||
319 	    rp->r_pc > t->t_dtrace_astpc + isz) {
320 		t->t_dtrace_ft = 0;
321 		return (0);
322 	}
323 
324 	/*
325 	 * If we have executed the original instruction, but we have performed
326 	 * neither the jmp back to t->t_dtrace_npc nor the clean up of any
327 	 * registers used to emulate %rip-relative instructions in 64-bit mode,
328 	 * we'll save ourselves some effort by doing that here and taking the
329 	 * signal right away.  We detect this condition by seeing if the program
330 	 * counter is the range [scrpc + isz, astpc).
331 	 */
332 	if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
333 	    rp->r_pc < t->t_dtrace_astpc) {
334 		/*
335 		 * If there is a scratch register and we're on the
336 		 * instruction immediately after the modified instruction,
337 		 * restore the value of that scratch register.
338 		 */
339 		if (t->t_dtrace_reg != 0 &&
340 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
341 			switch (t->t_dtrace_reg) {
342 			case REG_RAX:
343 				rp->r_rax = t->t_dtrace_regv;
344 				break;
345 			case REG_RCX:
346 				rp->r_rcx = t->t_dtrace_regv;
347 				break;
348 			case REG_R8:
349 				rp->r_r8 = t->t_dtrace_regv;
350 				break;
351 			case REG_R9:
352 				rp->r_r9 = t->t_dtrace_regv;
353 				break;
354 			}
355 		}
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 
376 /*
377  * Additional artificial frames for the machine type. For i86pc, we're already
378  * accounted for, so return 0. On the hypervisor, we have an additional frame
379  * (xen_callback_handler).
380  */
381 int
382 dtrace_mach_aframes(void)
383 {
384 #ifdef __xpv
385 	return (1);
386 #else
387 	return (0);
388 #endif
389 }
390