xref: /netbsd-src/external/cddl/osnet/dev/dtrace/amd64/dtrace_subr.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /*	$NetBSD: dtrace_subr.c,v 1.10 2018/06/05 12:04:19 christos 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: head/sys/cddl/dev/dtrace/amd64/dtrace_subr.c 313850 2017-02-17 03:27:20Z markj $
25  *
26  */
27 /*
28  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
29  * Use is subject to license terms.
30  */
31 
32 /*
33  * Copyright (c) 2011, Joyent, Inc. All rights reserved.
34  */
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/types.h>
39 #include <sys/kernel.h>
40 #include <sys/malloc.h>
41 #include <sys/kmem.h>
42 #include <sys/xcall.h>
43 #include <sys/cpu.h>
44 #include <sys/cpuvar.h>
45 #include <sys/dtrace_impl.h>
46 #include <sys/dtrace_bsd.h>
47 #include <machine/frame.h>
48 #include <machine/cpu_counter.h>
49 #include <machine/cpufunc.h>
50 
51 extern uintptr_t 	kernelbase;
52 
53 int dtrace_invop(uintptr_t, struct trapframe *, uintptr_t);
54 
55 typedef struct dtrace_invop_hdlr {
56 	int (*dtih_func)(uintptr_t, struct trapframe *, uintptr_t);
57 	struct dtrace_invop_hdlr *dtih_next;
58 } dtrace_invop_hdlr_t;
59 
60 dtrace_invop_hdlr_t *dtrace_invop_hdlr;
61 
62 void dtrace_gethrtime_init(void *);
63 void dtrace_getnanotime(struct timespec *);
64 
65 int
66 dtrace_invop(uintptr_t addr, struct trapframe *frame, 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, frame, eax)) != 0)
73 			return (rval);
74 
75 	return (0);
76 }
77 
78 void
79 dtrace_invop_add(int (*func)(uintptr_t, struct trapframe *, 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, struct trapframe *, 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 void
164 dtrace_safe_synchronous_signal(void)
165 {
166 	kthread_t *t = curthread;
167 	struct regs *rp = lwptoregs(ttolwp(t));
168 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
169 
170 	ASSERT(t->t_dtrace_on);
171 
172 	/*
173 	 * If we're not in the range of scratch addresses, we're not actually
174 	 * tracing user instructions so turn off the flags. If the instruction
175 	 * we copied out caused a synchonous trap, reset the pc back to its
176 	 * original value and turn off the flags.
177 	 */
178 	if (rp->r_pc < t->t_dtrace_scrpc ||
179 	    rp->r_pc > t->t_dtrace_astpc + isz) {
180 		t->t_dtrace_ft = 0;
181 	} else if (rp->r_pc == t->t_dtrace_scrpc ||
182 	    rp->r_pc == t->t_dtrace_astpc) {
183 		rp->r_pc = t->t_dtrace_pc;
184 		t->t_dtrace_ft = 0;
185 	}
186 }
187 
188 int
189 dtrace_safe_defer_signal(void)
190 {
191 	kthread_t *t = curthread;
192 	struct regs *rp = lwptoregs(ttolwp(t));
193 	size_t isz = t->t_dtrace_npc - t->t_dtrace_pc;
194 
195 	ASSERT(t->t_dtrace_on);
196 
197 	/*
198 	 * If we're not in the range of scratch addresses, we're not actually
199 	 * tracing user instructions so turn off the flags.
200 	 */
201 	if (rp->r_pc < t->t_dtrace_scrpc ||
202 	    rp->r_pc > t->t_dtrace_astpc + isz) {
203 		t->t_dtrace_ft = 0;
204 		return (0);
205 	}
206 
207 	/*
208 	 * If we have executed the original instruction, but we have performed
209 	 * neither the jmp back to t->t_dtrace_npc nor the clean up of any
210 	 * registers used to emulate %rip-relative instructions in 64-bit mode,
211 	 * we'll save ourselves some effort by doing that here and taking the
212 	 * signal right away.  We detect this condition by seeing if the program
213 	 * counter is the range [scrpc + isz, astpc).
214 	 */
215 	if (rp->r_pc >= t->t_dtrace_scrpc + isz &&
216 	    rp->r_pc < t->t_dtrace_astpc) {
217 #ifdef __amd64
218 		/*
219 		 * If there is a scratch register and we're on the
220 		 * instruction immediately after the modified instruction,
221 		 * restore the value of that scratch register.
222 		 */
223 		if (t->t_dtrace_reg != 0 &&
224 		    rp->r_pc == t->t_dtrace_scrpc + isz) {
225 			switch (t->t_dtrace_reg) {
226 			case REG_RAX:
227 				rp->r_rax = t->t_dtrace_regv;
228 				break;
229 			case REG_RCX:
230 				rp->r_rcx = t->t_dtrace_regv;
231 				break;
232 			case REG_R8:
233 				rp->r_r8 = t->t_dtrace_regv;
234 				break;
235 			case REG_R9:
236 				rp->r_r9 = t->t_dtrace_regv;
237 				break;
238 			}
239 		}
240 #endif
241 		rp->r_pc = t->t_dtrace_npc;
242 		t->t_dtrace_ft = 0;
243 		return (0);
244 	}
245 
246 	/*
247 	 * Otherwise, make sure we'll return to the kernel after executing
248 	 * the copied out instruction and defer the signal.
249 	 */
250 	if (!t->t_dtrace_step) {
251 		ASSERT(rp->r_pc < t->t_dtrace_astpc);
252 		rp->r_pc += t->t_dtrace_astpc - t->t_dtrace_scrpc;
253 		t->t_dtrace_step = 1;
254 	}
255 
256 	t->t_dtrace_ast = 1;
257 
258 	return (1);
259 }
260 #endif
261 
262 #ifdef __NetBSD__
263 static __inline uint64_t
264 dtrace_rdtsc(void)
265 {
266 	uint32_t hi, lo;
267 
268 	__asm volatile("rdtsc" : "=d" (hi), "=a" (lo));
269 	return (((uint64_t)hi << 32) | (uint64_t) lo);
270 }
271 #define rdtsc dtrace_rdtsc
272 #endif
273 
274 #ifdef notyet
275 static int64_t	tgt_cpu_tsc;
276 static int64_t	hst_cpu_tsc;
277 #endif
278 static int64_t	tsc_skew[MAXCPUS];
279 static uint64_t	nsec_scale;
280 
281 /* See below for the explanation of this macro. */
282 #define SCALE_SHIFT	28
283 
284 #ifdef notyet
285 static void
286 dtrace_gethrtime_init_sync(void *arg)
287 {
288 #ifdef CHECK_SYNC
289 	/*
290 	 * Delay this function from returning on one
291 	 * of the CPUs to check that the synchronisation
292 	 * works.
293 	 */
294 	uintptr_t cpu = (uintptr_t) arg;
295 
296 	if (cpu == cpu_number()) {
297 		int i;
298 		for (i = 0; i < 1000000000; i++)
299 			tgt_cpu_tsc = rdtsc();
300 		tgt_cpu_tsc = 0;
301 	}
302 #endif
303 }
304 
305 static void
306 dtrace_gethrtime_init_cpu(void *arg)
307 {
308 	uintptr_t cpu = (uintptr_t) arg;
309 
310 	if (cpu == cpu_number())
311 		tgt_cpu_tsc = rdtsc();
312 	else
313 		hst_cpu_tsc = rdtsc();
314 }
315 #endif
316 
317 void
318 dtrace_gethrtime_init(void *arg)
319 {
320 	uint64_t tsc_f;
321 	CPU_INFO_ITERATOR cpuind;
322 	struct cpu_info *cinfo = curcpu();
323 	cpuid_t cur_cpuid = cpu_number();	/* current cpu id */
324 
325 	/*
326 	 * Get TSC frequency known at this moment.
327 	 * This should be constant if TSC is invariant.
328 	 * Otherwise tick->time conversion will be inaccurate, but
329 	 * will preserve monotonic property of TSC.
330 	 */
331 	tsc_f = cpu_frequency(cinfo);
332 
333 	/*
334 	 * The following line checks that nsec_scale calculated below
335 	 * doesn't overflow 32-bit unsigned integer, so that it can multiply
336 	 * another 32-bit integer without overflowing 64-bit.
337 	 * Thus minimum supported TSC frequency is 62.5MHz.
338 	 */
339 	KASSERT(tsc_f > (NANOSEC >> (32 - SCALE_SHIFT)));
340 
341 	/*
342 	 * We scale up NANOSEC/tsc_f ratio to preserve as much precision
343 	 * as possible.
344 	 * 2^28 factor was chosen quite arbitrarily from practical
345 	 * considerations:
346 	 * - it supports TSC frequencies as low as 62.5MHz (see above);
347 	 * - it provides quite good precision (e < 0.01%) up to THz
348 	 *   (terahertz) values;
349 	 */
350 	nsec_scale = ((uint64_t)NANOSEC << SCALE_SHIFT) / tsc_f;
351 
352 	/* The current CPU is the reference one. */
353 	tsc_skew[cur_cpuid] = 0;
354 
355 	for (CPU_INFO_FOREACH(cpuind, cinfo)) {
356 		/* use skew relative to cpu 0 */
357 		tsc_skew[cpu_index(cinfo)] = cinfo->ci_data.cpu_cc_skew;
358 	}
359 }
360 
361 /*
362  * DTrace needs a high resolution time function which can
363  * be called from a probe context and guaranteed not to have
364  * instrumented with probes itself.
365  *
366  * Returns nanoseconds since boot.
367  */
368 uint64_t
369 dtrace_gethrtime()
370 {
371 	uint64_t tsc;
372 	uint32_t lo;
373 	uint32_t hi;
374 
375 	/*
376 	 * We split TSC value into lower and higher 32-bit halves and separately
377 	 * scale them with nsec_scale, then we scale them down by 2^28
378 	 * (see nsec_scale calculations) taking into account 32-bit shift of
379 	 * the higher half and finally add.
380 	 */
381 	tsc = rdtsc() + tsc_skew[cpu_number()];
382 	lo = tsc;
383 	hi = tsc >> 32;
384 	return (((lo * nsec_scale) >> SCALE_SHIFT) +
385 	    ((hi * nsec_scale) << (32 - SCALE_SHIFT)));
386 }
387 
388 uint64_t
389 dtrace_gethrestime(void)
390 {
391 	struct timespec current_time;
392 
393 	dtrace_getnanotime(&current_time);
394 
395 	return (current_time.tv_sec * 1000000000ULL + current_time.tv_nsec);
396 }
397 
398 /* Function to handle DTrace traps during probes. See amd64/amd64/trap.c */
399 int
400 dtrace_trap(struct trapframe *frame, u_int type)
401 {
402 	bool nofault;
403 	cpuid_t cpuid = cpu_number();	/* current cpu id */
404 
405 	/*
406 	 * A trap can occur while DTrace executes a probe. Before
407 	 * executing the probe, DTrace blocks re-scheduling and sets
408 	 * a flag in it's per-cpu flags to indicate that it doesn't
409 	 * want to fault. On returning from the the probe, the no-fault
410 	 * flag is cleared and finally re-scheduling is enabled.
411 	 *
412 	 * Check if DTrace has enabled 'no-fault' mode:
413 	 *
414 	 */
415 	nofault = (cpu_core[cpuid].cpuc_dtrace_flags & CPU_DTRACE_NOFAULT) != 0;
416 	if (nofault) {
417 		KASSERTMSG((read_rflags() & PSL_I) == 0, "interrupts enabled");
418 
419 		/*
420 		 * There are only a couple of trap types that are expected.
421 		 * All the rest will be handled in the usual way.
422 		 */
423 		switch (type) {
424 		/* General protection fault. */
425 		case T_PROTFLT:
426 			/* Flag an illegal operation. */
427 			cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
428 
429 			/*
430 			 * Offset the instruction pointer to the instruction
431 			 * following the one causing the fault.
432 			 */
433 			frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
434 			return (1);
435 		/* Page fault. */
436 		case T_PAGEFLT:
437 			/* Flag a bad address. */
438 			cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_BADADDR;
439 			cpu_core[cpuid].cpuc_dtrace_illval = rcr2();
440 
441 			/*
442 			 * Offset the instruction pointer to the instruction
443 			 * following the one causing the fault.
444 			 */
445 			frame->tf_rip += dtrace_instr_size((u_char *) frame->tf_rip);
446 			return (1);
447 		default:
448 			/* Handle all other traps in the usual way. */
449 			break;
450 		}
451 	}
452 
453 	/* Handle the trap in the usual way. */
454 	return (0);
455 }
456