xref: /illumos-gate/usr/src/uts/i86pc/os/dtrace_subr.c (revision 49ef7e0638c8b771d8a136eae78b1c0f99acc8e0)
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 #ifdef __amd64
111 	extern uintptr_t toxic_addr;
112 	extern size_t toxic_size;
113 
114 	(*func)(0, _userlimit);
115 
116 	if (hole_end > hole_start)
117 		(*func)(hole_start, hole_end);
118 	(*func)(toxic_addr, toxic_addr + toxic_size);
119 #else
120 	extern void *device_arena_contains(void *, size_t, size_t *);
121 	caddr_t	vaddr;
122 	size_t	len;
123 
124 	for (vaddr = (caddr_t)kernelbase; vaddr < (caddr_t)KERNEL_TEXT;
125 	    vaddr += len) {
126 		len = (caddr_t)KERNEL_TEXT - vaddr;
127 		vaddr = device_arena_contains(vaddr, len, &len);
128 		if (vaddr == NULL)
129 			break;
130 		(*func)((uintptr_t)vaddr, (uintptr_t)vaddr + len);
131 	}
132 #endif
133 	(*func)(0, _userlimit);
134 }
135 
136 static int
137 dtrace_xcall_func(dtrace_xcall_t func, void *arg)
138 {
139 	(*func)(arg);
140 
141 	return (0);
142 }
143 
144 /*ARGSUSED*/
145 void
146 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
147 {
148 	cpuset_t set;
149 
150 	CPUSET_ZERO(set);
151 
152 	if (cpu == DTRACE_CPUALL) {
153 		CPUSET_ALL(set);
154 	} else {
155 		CPUSET_ADD(set, cpu);
156 	}
157 
158 	kpreempt_disable();
159 	xc_sync((xc_arg_t)func, (xc_arg_t)arg, 0, CPUSET2BV(set),
160 	    (xc_func_t)dtrace_xcall_func);
161 	kpreempt_enable();
162 }
163 
164 void
165 dtrace_sync_func(void)
166 {}
167 
168 void
169 dtrace_sync(void)
170 {
171 	dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
172 }
173 
174 int (*dtrace_pid_probe_ptr)(struct regs *);
175 int (*dtrace_return_probe_ptr)(struct regs *);
176 
177 void
178 dtrace_user_probe(struct regs *rp, caddr_t addr, processorid_t cpuid)
179 {
180 	krwlock_t *rwp;
181 	proc_t *p = curproc;
182 	extern void trap(struct regs *, caddr_t, processorid_t);
183 
184 	if (USERMODE(rp->r_cs) || (rp->r_ps & PS_VM)) {
185 		if (curthread->t_cred != p->p_cred) {
186 			cred_t *oldcred = curthread->t_cred;
187 			/*
188 			 * DTrace accesses t_cred in probe context.  t_cred
189 			 * must always be either NULL, or point to a valid,
190 			 * allocated cred structure.
191 			 */
192 			curthread->t_cred = crgetcred();
193 			crfree(oldcred);
194 		}
195 	}
196 
197 	if (rp->r_trapno == T_DTRACE_RET) {
198 		uint8_t step = curthread->t_dtrace_step;
199 		uint8_t ret = curthread->t_dtrace_ret;
200 		uintptr_t npc = curthread->t_dtrace_npc;
201 
202 		if (curthread->t_dtrace_ast) {
203 			aston(curthread);
204 			curthread->t_sig_check = 1;
205 		}
206 
207 		/*
208 		 * Clear all user tracing flags.
209 		 */
210 		curthread->t_dtrace_ft = 0;
211 
212 		/*
213 		 * If we weren't expecting to take a return probe trap, kill
214 		 * the process as though it had just executed an unassigned
215 		 * trap instruction.
216 		 */
217 		if (step == 0) {
218 			tsignal(curthread, SIGILL);
219 			return;
220 		}
221 
222 		/*
223 		 * If we hit this trap unrelated to a return probe, we're
224 		 * just here to reset the AST flag since we deferred a signal
225 		 * until after we logically single-stepped the instruction we
226 		 * copied out.
227 		 */
228 		if (ret == 0) {
229 			rp->r_pc = npc;
230 			return;
231 		}
232 
233 		/*
234 		 * We need to wait until after we've called the
235 		 * dtrace_return_probe_ptr function pointer to set %pc.
236 		 */
237 		rwp = &CPU->cpu_ft_lock;
238 		rw_enter(rwp, RW_READER);
239 		if (dtrace_return_probe_ptr != NULL)
240 			(void) (*dtrace_return_probe_ptr)(rp);
241 		rw_exit(rwp);
242 		rp->r_pc = npc;
243 
244 	} else if (rp->r_trapno == T_BPTFLT) {
245 		uint8_t instr, instr2;
246 		caddr_t linearpc;
247 		rwp = &CPU->cpu_ft_lock;
248 
249 		/*
250 		 * The DTrace fasttrap provider uses the breakpoint trap
251 		 * (int 3). We let DTrace take the first crack at handling
252 		 * this trap; if it's not a probe that DTrace knowns about,
253 		 * we call into the trap() routine to handle it like a
254 		 * breakpoint placed by a conventional debugger.
255 		 */
256 		rw_enter(rwp, RW_READER);
257 		if (dtrace_pid_probe_ptr != NULL &&
258 		    (*dtrace_pid_probe_ptr)(rp) == 0) {
259 			rw_exit(rwp);
260 			return;
261 		}
262 		rw_exit(rwp);
263 
264 		if (dtrace_linear_pc(rp, p, &linearpc) != 0) {
265 			trap(rp, addr, cpuid);
266 			return;
267 		}
268 
269 		/*
270 		 * If the instruction that caused the breakpoint trap doesn't
271 		 * look like an int 3 anymore, it may be that this tracepoint
272 		 * was removed just after the user thread executed it. In
273 		 * that case, return to user land to retry the instuction.
274 		 * Note that we assume the length of the instruction to retry
275 		 * is 1 byte because that's the length of FASTTRAP_INSTR.
276 		 * We check for r_pc > 0 and > 2 so that we don't have to
277 		 * deal with segment wraparound.
278 		 */
279 		if (rp->r_pc > 0 && fuword8(linearpc - 1, &instr) == 0 &&
280 		    instr != FASTTRAP_INSTR &&
281 		    (instr != 3 || (rp->r_pc >= 2 &&
282 		    (fuword8(linearpc - 2, &instr2) != 0 || instr2 != 0xCD)))) {
283 			rp->r_pc--;
284 			return;
285 		}
286 
287 		trap(rp, addr, cpuid);
288 
289 	} else {
290 		trap(rp, addr, cpuid);
291 	}
292 }
293 
294 void
295 dtrace_safe_synchronous_signal(void)
296 {
297 	kthread_t *t = curthread;
298 	struct regs *rp = lwptoregs(ttolwp(t));
299 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
300 
301 	ASSERT(t->t_dtrace_on);
302 
303 	/*
304 	 * If we're not in the range of scratch addresses, we're not actually
305 	 * tracing user instructions so turn off the flags. If the instruction
306 	 * we copied out caused a synchonous trap, reset the pc back to its
307 	 * original value and turn off the flags.
308 	 */
309 	if (rp->r_pc < t->t_dtrace_scrpc ||
310 	    rp->r_pc > t->t_dtrace_astpc + isz) {
311 		t->t_dtrace_ft = 0;
312 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
313 	    rp->r_pc == t->t_dtrace_astpc) {
314 		rp->r_pc = t->t_dtrace_pc;
315 		t->t_dtrace_ft = 0;
316 	}
317 }
318 
319 int
320 dtrace_safe_defer_signal(void)
321 {
322 	kthread_t *t = curthread;
323 	struct regs *rp = lwptoregs(ttolwp(t));
324 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
325 
326 	ASSERT(t->t_dtrace_on);
327 
328 	/*
329 	 * If we're not in the range of scratch addresses, we're not actually
330 	 * tracing user instructions so turn off the flags.
331 	 */
332 	if (rp->r_pc < t->t_dtrace_scrpc ||
333 	    rp->r_pc > t->t_dtrace_astpc + isz) {
334 		t->t_dtrace_ft = 0;
335 		return (0);
336 	}
337 
338 	/*
339 	 * If we have executed the original instruction, but we have performed
340 	 * neither the jmp back to t->t_dtrace_npc nor the clean up of any
341 	 * registers used to emulate %rip-relative instructions in 64-bit mode,
342 	 * we'll save ourselves some effort by doing that here and taking the
343 	 * signal right away.  We detect this condition by seeing if the program
344 	 * counter is the range [scrpc + isz, astpc).
345 	 */
346 	if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
347 	    rp->r_pc < t->t_dtrace_astpc) {
348 #ifdef __amd64
349 		/*
350 		 * If there is a scratch register and we're on the
351 		 * instruction immediately after the modified instruction,
352 		 * restore the value of that scratch register.
353 		 */
354 		if (t->t_dtrace_reg != 0 &&
355 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
356 			switch (t->t_dtrace_reg) {
357 			case REG_RAX:
358 				rp->r_rax = t->t_dtrace_regv;
359 				break;
360 			case REG_RCX:
361 				rp->r_rcx = t->t_dtrace_regv;
362 				break;
363 			case REG_R8:
364 				rp->r_r8 = t->t_dtrace_regv;
365 				break;
366 			case REG_R9:
367 				rp->r_r9 = t->t_dtrace_regv;
368 				break;
369 			}
370 		}
371 #endif
372 		rp->r_pc = t->t_dtrace_npc;
373 		t->t_dtrace_ft = 0;
374 		return (0);
375 	}
376 
377 	/*
378 	 * Otherwise, make sure we'll return to the kernel after executing
379 	 * the copied out instruction and defer the signal.
380 	 */
381 	if (!t->t_dtrace_step) {
382 		ASSERT(rp->r_pc < t->t_dtrace_astpc);
383 		rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
384 		t->t_dtrace_step = 1;
385 	}
386 
387 	t->t_dtrace_ast = 1;
388 
389 	return (1);
390 }
391 
392 /*
393  * Additional artificial frames for the machine type. For i86pc, we're already
394  * accounted for, so return 0. On the hypervisor, we have an additional frame
395  * (xen_callback_handler).
396  */
397 int
398 dtrace_mach_aframes(void)
399 {
400 #ifdef __xpv
401 	return (1);
402 #else
403 	return (0);
404 #endif
405 }
406