xref: /freebsd/sys/cddl/dev/dtrace/i386/dtrace_subr.c (revision d9f0ce31900a48d1a2bfc1c8c86f79d1e831451a)
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 /*
31  * Copyright (c) 2011, Joyent, Inc. All rights reserved.
32  */
33 
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/types.h>
37 #include <sys/cpuset.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/kmem.h>
41 #include <sys/smp.h>
42 #include <sys/dtrace_impl.h>
43 #include <sys/dtrace_bsd.h>
44 #include <machine/clock.h>
45 #include <machine/frame.h>
46 #include <vm/pmap.h>
47 
48 extern uintptr_t 	kernelbase;
49 
50 extern void dtrace_getnanotime(struct timespec *tsp);
51 
52 int dtrace_invop(uintptr_t, uintptr_t *, uintptr_t);
53 
54 typedef struct dtrace_invop_hdlr {
55 	int (*dtih_func)(uintptr_t, uintptr_t *, uintptr_t);
56 	struct dtrace_invop_hdlr *dtih_next;
57 } dtrace_invop_hdlr_t;
58 
59 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
60 
61 int
62 dtrace_invop(uintptr_t addr, uintptr_t *stack, uintptr_t eax)
63 {
64 	dtrace_invop_hdlr_t *hdlr;
65 	int rval;
66 
67 	for (hdlr = dtrace_invop_hdlr; hdlr != NULL; hdlr = hdlr->dtih_next)
68 		if ((rval = hdlr->dtih_func(addr, stack, eax)) != 0)
69 			return (rval);
70 
71 	return (0);
72 }
73 
74 void
75 dtrace_invop_add(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
76 {
77 	dtrace_invop_hdlr_t *hdlr;
78 
79 	hdlr = kmem_alloc(sizeof (dtrace_invop_hdlr_t), KM_SLEEP);
80 	hdlr->dtih_func = func;
81 	hdlr->dtih_next = dtrace_invop_hdlr;
82 	dtrace_invop_hdlr = hdlr;
83 }
84 
85 void
86 dtrace_invop_remove(int (*func)(uintptr_t, uintptr_t *, uintptr_t))
87 {
88 	dtrace_invop_hdlr_t *hdlr = dtrace_invop_hdlr, *prev = NULL;
89 
90 	for (;;) {
91 		if (hdlr == NULL)
92 			panic("attempt to remove non-existent invop handler");
93 
94 		if (hdlr->dtih_func == func)
95 			break;
96 
97 		prev = hdlr;
98 		hdlr = hdlr->dtih_next;
99 	}
100 
101 	if (prev == NULL) {
102 		ASSERT(dtrace_invop_hdlr == hdlr);
103 		dtrace_invop_hdlr = hdlr->dtih_next;
104 	} else {
105 		ASSERT(dtrace_invop_hdlr != hdlr);
106 		prev->dtih_next = hdlr->dtih_next;
107 	}
108 
109 	kmem_free(hdlr, 0);
110 }
111 
112 void
113 dtrace_toxic_ranges(void (*func)(uintptr_t base, uintptr_t limit))
114 {
115 	(*func)(0, kernelbase);
116 }
117 
118 void
119 dtrace_xcall(processorid_t cpu, dtrace_xcall_t func, void *arg)
120 {
121 	cpuset_t cpus;
122 
123 	if (cpu == DTRACE_CPUALL)
124 		cpus = all_cpus;
125 	else
126 		CPU_SETOF(cpu, &cpus);
127 
128 	smp_rendezvous_cpus(cpus, smp_no_rendevous_barrier, func,
129 	    smp_no_rendevous_barrier, arg);
130 }
131 
132 static void
133 dtrace_sync_func(void)
134 {
135 }
136 
137 void
138 dtrace_sync(void)
139 {
140         dtrace_xcall(DTRACE_CPUALL, (dtrace_xcall_t)dtrace_sync_func, NULL);
141 }
142 
143 #ifdef notyet
144 void
145 dtrace_safe_synchronous_signal(void)
146 {
147 	kthread_t *t = curthread;
148 	struct regs *rp = lwptoregs(ttolwp(t));
149 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
150 
151 	ASSERT(t->t_dtrace_on);
152 
153 	/*
154 	 * If we're not in the range of scratch addresses, we're not actually
155 	 * tracing user instructions so turn off the flags. If the instruction
156 	 * we copied out caused a synchonous trap, reset the pc back to its
157 	 * original value and turn off the flags.
158 	 */
159 	if (rp->r_pc < t->t_dtrace_scrpc ||
160 	    rp->r_pc > t->t_dtrace_astpc + isz) {
161 		t->t_dtrace_ft = 0;
162 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
163 	    rp->r_pc == t->t_dtrace_astpc) {
164 		rp->r_pc = t->t_dtrace_pc;
165 		t->t_dtrace_ft = 0;
166 	}
167 }
168 
169 int
170 dtrace_safe_defer_signal(void)
171 {
172 	kthread_t *t = curthread;
173 	struct regs *rp = lwptoregs(ttolwp(t));
174 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
175 
176 	ASSERT(t->t_dtrace_on);
177 
178 	/*
179 	 * If we're not in the range of scratch addresses, we're not actually
180 	 * tracing user instructions so turn off the flags.
181 	 */
182 	if (rp->r_pc < t->t_dtrace_scrpc ||
183 	    rp->r_pc > t->t_dtrace_astpc + isz) {
184 		t->t_dtrace_ft = 0;
185 		return (0);
186 	}
187 
188 	/*
189 	 * If we have executed the original instruction, but we have performed
190 	 * neither the jmp back to t->t_dtrace_npc nor the clean up of any
191 	 * registers used to emulate %rip-relative instructions in 64-bit mode,
192 	 * we'll save ourselves some effort by doing that here and taking the
193 	 * signal right away.  We detect this condition by seeing if the program
194 	 * counter is the range [scrpc + isz, astpc).
195 	 */
196 	if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
197 	    rp->r_pc < t->t_dtrace_astpc) {
198 #ifdef __amd64
199 		/*
200 		 * If there is a scratch register and we're on the
201 		 * instruction immediately after the modified instruction,
202 		 * restore the value of that scratch register.
203 		 */
204 		if (t->t_dtrace_reg != 0 &&
205 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
206 			switch (t->t_dtrace_reg) {
207 			case REG_RAX:
208 				rp->r_rax = t->t_dtrace_regv;
209 				break;
210 			case REG_RCX:
211 				rp->r_rcx = t->t_dtrace_regv;
212 				break;
213 			case REG_R8:
214 				rp->r_r8 = t->t_dtrace_regv;
215 				break;
216 			case REG_R9:
217 				rp->r_r9 = t->t_dtrace_regv;
218 				break;
219 			}
220 		}
221 #endif
222 		rp->r_pc = t->t_dtrace_npc;
223 		t->t_dtrace_ft = 0;
224 		return (0);
225 	}
226 
227 	/*
228 	 * Otherwise, make sure we'll return to the kernel after executing
229 	 * the copied out instruction and defer the signal.
230 	 */
231 	if (!t->t_dtrace_step) {
232 		ASSERT(rp->r_pc < t->t_dtrace_astpc);
233 		rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
234 		t->t_dtrace_step = 1;
235 	}
236 
237 	t->t_dtrace_ast = 1;
238 
239 	return (1);
240 }
241 #endif
242 
243 static int64_t	tgt_cpu_tsc;
244 static int64_t	hst_cpu_tsc;
245 static int64_t	tsc_skew[MAXCPU];
246 static uint64_t	nsec_scale;
247 
248 /* See below for the explanation of this macro. */
249 #define SCALE_SHIFT	28
250 
251 /*
252  * Get the frequency and scale factor as early as possible so that they can be
253  * used for boot-time tracing.
254  */
255 static void
256 dtrace_gethrtime_init_early(void *arg)
257 {
258 	uint64_t tsc_f;
259 
260 	/*
261 	 * Get TSC frequency known at this moment.
262 	 * This should be constant if TSC is invariant.
263 	 * Otherwise tick->time conversion will be inaccurate, but
264 	 * will preserve monotonic property of TSC.
265 	 */
266 	tsc_f = atomic_load_acq_64(&tsc_freq);
267 
268 	/*
269 	 * The following line checks that nsec_scale calculated below
270 	 * doesn't overflow 32-bit unsigned integer, so that it can multiply
271 	 * another 32-bit integer without overflowing 64-bit.
272 	 * Thus minimum supported TSC frequency is 62.5MHz.
273 	 */
274 	KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)),
275 	    ("TSC frequency is too low"));
276 
277 	/*
278 	 * We scale up NANOSEC/tsc_f ratio to preserve as much precision
279 	 * as possible.
280 	 * 2^28 factor was chosen quite arbitrarily from practical
281 	 * considerations:
282 	 * - it supports TSC frequencies as low as 62.5MHz (see above);
283 	 * - it provides quite good precision (e < 0.01%) up to THz
284 	 *   (terahertz) values;
285 	 */
286 	nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
287 }
288 SYSINIT(dtrace_gethrtime_init_early, SI_SUB_CPU, SI_ORDER_ANY,
289     dtrace_gethrtime_init_early, NULL);
290 
291 static void
292 dtrace_gethrtime_init_cpu(void *arg)
293 {
294 	uintptr_t cpu = (uintptr_t) arg;
295 
296 	if (cpu == curcpu)
297 		tgt_cpu_tsc = rdtsc();
298 	else
299 		hst_cpu_tsc = rdtsc();
300 }
301 
302 static void
303 dtrace_gethrtime_init(void *arg)
304 {
305 	cpuset_t map;
306 	struct pcpu *pc;
307 	int i;
308 
309 	/* The current CPU is the reference one. */
310 	sched_pin();
311 	tsc_skew[curcpu] = 0;
312 	CPU_FOREACH(i) {
313 		if (i == curcpu)
314 			continue;
315 
316 		pc = pcpu_find(i);
317 		CPU_SETOF(PCPU_GET(cpuid), &map);
318 		CPU_SET(pc->pc_cpuid, &map);
319 
320 		smp_rendezvous_cpus(map, NULL,
321 		    dtrace_gethrtime_init_cpu,
322 		    smp_no_rendevous_barrier, (void *)(uintptr_t) i);
323 
324 		tsc_skew[i] = tgt_cpu_tsc - hst_cpu_tsc;
325 	}
326 	sched_unpin();
327 }
328 SYSINIT(dtrace_gethrtime_init, SI_SUB_SMP, SI_ORDER_ANY, dtrace_gethrtime_init,
329     NULL);
330 
331 /*
332  * DTrace needs a high resolution time function which can
333  * be called from a probe context and guaranteed not to have
334  * instrumented with probes itself.
335  *
336  * Returns nanoseconds since boot.
337  */
338 uint64_t
339 dtrace_gethrtime()
340 {
341 	uint64_t tsc;
342 	uint32_t lo;
343 	uint32_t hi;
344 
345 	/*
346 	 * We split TSC value into lower and higher 32-bit halves and separately
347 	 * scale them with nsec_scale, then we scale them down by 2^28
348 	 * (see nsec_scale calculations) taking into account 32-bit shift of
349 	 * the higher half and finally add.
350 	 */
351 	tsc = rdtsc() - tsc_skew[curcpu];
352 	lo = tsc;
353 	hi = tsc >> 32;
354 	return (((lo * nsec_scale) >> SCALE_SHIFT) +
355 	    ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
356 }
357 
358 uint64_t
359 dtrace_gethrestime(void)
360 {
361 	struct timespec current_time;
362 
363 	dtrace_getnanotime(&current_time);
364 
365 	return (current_time.tv_sec * 1000000000ULL + current_time.tv_nsec);
366 }
367 
368 /* Function to handle DTrace traps during probes. See i386/i386/trap.c */
369 int
370 dtrace_trap(struct trapframe *frame, u_int type)
371 {
372 	/*
373 	 * A trap can occur while DTrace executes a probe. Before
374 	 * executing the probe, DTrace blocks re-scheduling and sets
375 	 * a flag in its per-cpu flags to indicate that it doesn't
376 	 * want to fault. On returning from the probe, the no-fault
377 	 * flag is cleared and finally re-scheduling is enabled.
378 	 *
379 	 * Check if DTrace has enabled 'no-fault' mode:
380 	 */
381 	if ((cpu_core[curcpu].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0) {
382 		/*
383 		 * There are only a couple of trap types that are expected.
384 		 * All the rest will be handled in the usual way.
385 		 */
386 		switch (type) {
387 		/* General protection fault. */
388 		case T_PROTFLT:
389 			/* Flag an illegal operation. */
390 			cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
391 
392 			/*
393 			 * Offset the instruction pointer to the instruction
394 			 * following the one causing the fault.
395 			 */
396 			frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
397 			return (1);
398 		/* Page fault. */
399 		case T_PAGEFLT:
400 			/* Flag a bad address. */
401 			cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
402 			cpu_core[curcpu].cpuc_dtrace_illval = rcr2();
403 
404 			/*
405 			 * Offset the instruction pointer to the instruction
406 			 * following the one causing the fault.
407 			 */
408 			frame->tf_eip += dtrace_instr_size((u_char *) frame->tf_eip);
409 			return (1);
410 		default:
411 			/* Handle all other traps in the usual way. */
412 			break;
413 		}
414 	}
415 
416 	/* Handle the trap in the usual way. */
417 	return (0);
418 }
419