xref: /titanic_50/usr/src/uts/sparc/dtrace/fbt.c (revision dfb96a4f56fb431b915bc67e5d9d5c8d4f4f6679)
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  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/errno.h>
29 #include <sys/stat.h>
30 #include <sys/modctl.h>
31 #include <sys/conf.h>
32 #include <sys/systm.h>
33 #include <sys/ddi.h>
34 #include <sys/sunddi.h>
35 #include <sys/cpuvar.h>
36 #include <sys/kmem.h>
37 #include <sys/strsubr.h>
38 #include <sys/dtrace.h>
39 #include <sys/kobj.h>
40 #include <sys/modctl.h>
41 #include <sys/atomic.h>
42 #include <vm/seg_kmem.h>
43 #include <sys/stack.h>
44 #include <sys/ctf_api.h>
45 #include <sys/sysmacros.h>
46 
47 static dev_info_t		*fbt_devi;
48 static dtrace_provider_id_t	fbt_id;
49 static uintptr_t		fbt_trampoline;
50 static caddr_t			fbt_trampoline_window;
51 static size_t			fbt_trampoline_size;
52 static int			fbt_verbose = 0;
53 
54 /*
55  * Various interesting bean counters.
56  */
57 static int			fbt_entry;
58 static int			fbt_ret;
59 static int			fbt_retl;
60 static int			fbt_retl_jmptab;
61 static int			fbt_retl_twoinstr;
62 static int			fbt_retl_tailcall;
63 static int			fbt_retl_tailjmpl;
64 static int			fbt_leaf_functions;
65 
66 extern char			stubs_base[];
67 extern char			stubs_end[];
68 
69 #define	FBT_REG_G0		0
70 #define	FBT_REG_G1		1
71 #define	FBT_REG_O0		8
72 #define	FBT_REG_O1		9
73 #define	FBT_REG_O2		10
74 #define	FBT_REG_O3		11
75 #define	FBT_REG_O4		12
76 #define	FBT_REG_O5		13
77 #define	FBT_REG_O6		14
78 #define	FBT_REG_O7		15
79 #define	FBT_REG_I0		24
80 #define	FBT_REG_I1		25
81 #define	FBT_REG_I2		26
82 #define	FBT_REG_I3		27
83 #define	FBT_REG_I4		28
84 #define	FBT_REG_I7		31
85 #define	FBT_REG_L0		16
86 #define	FBT_REG_L1		17
87 #define	FBT_REG_L2		18
88 #define	FBT_REG_L3		19
89 #define	FBT_REG_PC		5
90 
91 #define	FBT_REG_ISGLOBAL(r)	((r) < 8)
92 #define	FBT_REG_ISOUTPUT(r)	((r) >= 8 && (r) < 16)
93 #define	FBT_REG_ISLOCAL(r)	((r) >= 16 && (r) < 24)
94 #define	FBT_REG_ISVOLATILE(r)	\
95 	((FBT_REG_ISGLOBAL(r) || FBT_REG_ISOUTPUT(r)) && (r) != FBT_REG_G0)
96 #define	FBT_REG_NLOCALS		8
97 
98 #define	FBT_REG_MARKLOCAL(locals, r)	\
99 	if (FBT_REG_ISLOCAL(r)) \
100 		(locals)[(r) - FBT_REG_L0] = 1;
101 
102 #define	FBT_REG_INITLOCALS(local, locals)	\
103 	for ((local) = 0; (local) < FBT_REG_NLOCALS; (local)++)  \
104 		(locals)[(local)] = 0; \
105 	(local) = FBT_REG_L0
106 
107 #define	FBT_REG_ALLOCLOCAL(local, locals)	\
108 	while ((locals)[(local) - FBT_REG_L0]) \
109 		(local)++; \
110 	(locals)[(local) - FBT_REG_L0] = 1;
111 
112 #define	FBT_OP_MASK		0xc0000000
113 #define	FBT_OP_SHIFT		30
114 #define	FBT_OP(val)		((val) & FBT_FMT1_MASK)
115 
116 #define	FBT_SIMM13_MASK		0x1fff
117 #define	FBT_SIMM13_MAX		((int32_t)0xfff)
118 #define	FBT_IMM22_MASK		0x3fffff
119 #define	FBT_IMM22_SHIFT		10
120 #define	FBT_IMM10_MASK		0x3ff
121 
122 #define	FBT_DISP30_MASK		0x3fffffff
123 #define	FBT_DISP30(from, to)	\
124 	(((uintptr_t)(to) - (uintptr_t)(from) >> 2) & FBT_DISP30_MASK)
125 
126 #define	FBT_DISP22_MASK		0x3fffff
127 #define	FBT_DISP22(from, to)	\
128 	(((uintptr_t)(to) - (uintptr_t)(from) >> 2) & FBT_DISP22_MASK)
129 
130 #define	FBT_DISP19_MASK		0x7ffff
131 #define	FBT_DISP19(from, to)	\
132 	(((uintptr_t)(to) - (uintptr_t)(from) >> 2) & FBT_DISP19_MASK)
133 
134 #define	FBT_DISP16_HISHIFT	20
135 #define	FBT_DISP16_HIMASK	(0x3 << FBT_DISP16_HISHIFT)
136 #define	FBT_DISP16_LOMASK	(0x3fff)
137 #define	FBT_DISP16_MASK		(FBT_DISP16_HIMASK | FBT_DISP16_LOMASK)
138 #define	FBT_DISP16(val)	\
139 	((((val) & FBT_DISP16_HIMASK) >> 6) | ((val) & FBT_DISP16_LOMASK))
140 
141 #define	FBT_DISP14_MASK		0x3fff
142 #define	FBT_DISP14(from, to)	\
143 	(((uintptr_t)(to) - (uintptr_t)(from) >> 2) & FBT_DISP14_MASK)
144 
145 #define	FBT_OP0			(((uint32_t)0) << FBT_OP_SHIFT)
146 #define	FBT_OP1			(((uint32_t)1) << FBT_OP_SHIFT)
147 #define	FBT_OP2			(((uint32_t)2) << FBT_OP_SHIFT)
148 #define	FBT_ILLTRAP		0
149 
150 #define	FBT_ANNUL_SHIFT		29
151 #define	FBT_ANNUL		(1 << FBT_ANNUL_SHIFT)
152 
153 #define	FBT_FMT3_OP3_SHIFT	19
154 #define	FBT_FMT3_OP_MASK	0xc1f80000
155 #define	FBT_FMT3_OP(val)	((val) & FBT_FMT3_OP_MASK)
156 
157 #define	FBT_FMT3_RD_SHIFT	25
158 #define	FBT_FMT3_RD_MASK	(0x1f << FBT_FMT3_RD_SHIFT)
159 #define	FBT_FMT3_RD(val)	\
160 	(((val) & FBT_FMT3_RD_MASK) >> FBT_FMT3_RD_SHIFT)
161 
162 #define	FBT_FMT3_RS1_SHIFT	14
163 #define	FBT_FMT3_RS1_MASK	(0x1f << FBT_FMT3_RS1_SHIFT)
164 #define	FBT_FMT3_RS1(val)	\
165 	(((val) & FBT_FMT3_RS1_MASK) >> FBT_FMT3_RS1_SHIFT)
166 #define	FBT_FMT3_RS1_SET(val, rs1) \
167 	(val) = ((val) & ~FBT_FMT3_RS1_MASK) | ((rs1) << FBT_FMT3_RS1_SHIFT)
168 
169 #define	FBT_FMT3_RS2_SHIFT	0
170 #define	FBT_FMT3_RS2_MASK	(0x1f << FBT_FMT3_RS2_SHIFT)
171 #define	FBT_FMT3_RS2(val)	\
172 	(((val) & FBT_FMT3_RS2_MASK) >> FBT_FMT3_RS2_SHIFT)
173 #define	FBT_FMT3_RS2_SET(val, rs2) \
174 	(val) = ((val) & ~FBT_FMT3_RS2_MASK) | ((rs2) << FBT_FMT3_RS2_SHIFT)
175 
176 #define	FBT_FMT3_IMM_SHIFT	13
177 #define	FBT_FMT3_IMM		(1 << FBT_FMT3_IMM_SHIFT)
178 #define	FBT_FMT3_SIMM13_MASK	FBT_SIMM13_MASK
179 
180 #define	FBT_FMT3_ISIMM(val)	((val) & FBT_FMT3_IMM)
181 #define	FBT_FMT3_SIMM13(val)	((val) & FBT_FMT3_SIMM13_MASK)
182 
183 #define	FBT_FMT2_OP2_SHIFT	22
184 #define	FBT_FMT2_OP2_MASK	(0x7 << FBT_FMT2_OP2_SHIFT)
185 #define	FBT_FMT2_RD_SHIFT	25
186 
187 #define	FBT_FMT1_OP(val)	((val) & FBT_OP_MASK)
188 #define	FBT_FMT1_DISP30(val)	((val) & FBT_DISP30_MASK)
189 
190 #define	FBT_FMT2_OP2_BPCC	(0x01 << FBT_FMT2_OP2_SHIFT)
191 #define	FBT_FMT2_OP2_BCC	(0x02 << FBT_FMT2_OP2_SHIFT)
192 #define	FBT_FMT2_OP2_BPR	(0x03 << FBT_FMT2_OP2_SHIFT)
193 #define	FBT_FMT2_OP2_SETHI	(0x04 << FBT_FMT2_OP2_SHIFT)
194 
195 #define	FBT_FMT2_COND_SHIFT	25
196 #define	FBT_FMT2_COND_BA	(0x8 << FBT_FMT2_COND_SHIFT)
197 #define	FBT_FMT2_COND_BL	(0x3 << FBT_FMT2_COND_SHIFT)
198 #define	FBT_FMT2_COND_BGE	(0xb << FBT_FMT2_COND_SHIFT)
199 
200 #define	FBT_OP_RESTORE		(FBT_OP2 | (0x3d << FBT_FMT3_OP3_SHIFT))
201 #define	FBT_OP_SAVE		(FBT_OP2 | (0x3c << FBT_FMT3_OP3_SHIFT))
202 #define	FBT_OP_JMPL		(FBT_OP2 | (0x38 << FBT_FMT3_OP3_SHIFT))
203 #define	FBT_OP_RETURN		(FBT_OP2 | (0x39 << FBT_FMT3_OP3_SHIFT))
204 #define	FBT_OP_CALL		FBT_OP1
205 #define	FBT_OP_SETHI		(FBT_OP0 | FBT_FMT2_OP2_SETHI)
206 #define	FBT_OP_ADD		(FBT_OP2 | (0x00 << FBT_FMT3_OP3_SHIFT))
207 #define	FBT_OP_OR		(FBT_OP2 | (0x02 << FBT_FMT3_OP3_SHIFT))
208 #define	FBT_OP_SUB		(FBT_OP2 | (0x04 << FBT_FMT3_OP3_SHIFT))
209 #define	FBT_OP_CC		(FBT_OP2 | (0x10 << FBT_FMT3_OP3_SHIFT))
210 #define	FBT_OP_BA		(FBT_OP0 | FBT_FMT2_OP2_BCC | FBT_FMT2_COND_BA)
211 #define	FBT_OP_BL		(FBT_OP0 | FBT_FMT2_OP2_BCC | FBT_FMT2_COND_BL)
212 #define	FBT_OP_BGE		(FBT_OP0 | FBT_FMT2_OP2_BCC | FBT_FMT2_COND_BGE)
213 #define	FBT_OP_BAPCC		(FBT_OP0 | FBT_FMT2_OP2_BPCC | FBT_FMT2_COND_BA)
214 #define	FBT_OP_RD		(FBT_OP2 | (0x28 << FBT_FMT3_OP3_SHIFT))
215 
216 #define	FBT_ORLO(rs, val, rd) \
217 	(FBT_OP_OR | ((rs) << FBT_FMT3_RS1_SHIFT) | \
218 	((rd) << FBT_FMT3_RD_SHIFT) | FBT_FMT3_IMM | ((val) & FBT_IMM10_MASK))
219 
220 #define	FBT_ORSIMM13(rs, val, rd) \
221 	(FBT_OP_OR | ((rs) << FBT_FMT3_RS1_SHIFT) | \
222 	((rd) << FBT_FMT3_RD_SHIFT) | FBT_FMT3_IMM | ((val) & FBT_SIMM13_MASK))
223 
224 #define	FBT_ADDSIMM13(rs, val, rd) \
225 	(FBT_OP_ADD | ((rs) << FBT_FMT3_RS1_SHIFT) | \
226 	((rd) << FBT_FMT3_RD_SHIFT) | FBT_FMT3_IMM | ((val) & FBT_SIMM13_MASK))
227 
228 #define	FBT_ADD(rs1, rs2, rd) \
229 	(FBT_OP_ADD | ((rs1) << FBT_FMT3_RS1_SHIFT) | \
230 	((rs2) << FBT_FMT3_RS2_SHIFT) | ((rd) << FBT_FMT3_RD_SHIFT))
231 
232 #define	FBT_CMP(rs1, rs2) \
233 	(FBT_OP_SUB | FBT_OP_CC | ((rs1) << FBT_FMT3_RS1_SHIFT) | \
234 	((rs2) << FBT_FMT3_RS2_SHIFT) | (FBT_REG_G0 << FBT_FMT3_RD_SHIFT))
235 
236 #define	FBT_MOV(rs, rd) \
237 	(FBT_OP_OR | (FBT_REG_G0 << FBT_FMT3_RS1_SHIFT) | \
238 	((rs) << FBT_FMT3_RS2_SHIFT) | ((rd) << FBT_FMT3_RD_SHIFT))
239 
240 #define	FBT_SETHI(val, reg)	\
241 	(FBT_OP_SETHI | (reg << FBT_FMT2_RD_SHIFT) | \
242 	((val >> FBT_IMM22_SHIFT) & FBT_IMM22_MASK))
243 
244 #define	FBT_CALL(orig, dest)	(FBT_OP_CALL | FBT_DISP30(orig, dest))
245 
246 #define	FBT_RET \
247 	(FBT_OP_JMPL | (FBT_REG_I7 << FBT_FMT3_RS1_SHIFT) | \
248 	(FBT_REG_G0 << FBT_FMT3_RD_SHIFT) | FBT_FMT3_IMM | (sizeof (pc_t) << 1))
249 
250 #define	FBT_SAVEIMM(rd, val, rs1)	\
251 	(FBT_OP_SAVE | ((rs1) << FBT_FMT3_RS1_SHIFT) | \
252 	((rd) << FBT_FMT3_RD_SHIFT) | FBT_FMT3_IMM | ((val) & FBT_SIMM13_MASK))
253 
254 #define	FBT_RESTORE(rd, rs1, rs2)	\
255 	(FBT_OP_RESTORE | ((rs1) << FBT_FMT3_RS1_SHIFT) | \
256 	((rd) << FBT_FMT3_RD_SHIFT) | ((rs2) << FBT_FMT3_RS2_SHIFT))
257 
258 #define	FBT_RETURN(rs1, val)		\
259 	(FBT_OP_RETURN | ((rs1) << FBT_FMT3_RS1_SHIFT) | \
260 	FBT_FMT3_IMM | ((val) & FBT_SIMM13_MASK))
261 
262 #define	FBT_BA(orig, dest)	(FBT_OP_BA | FBT_DISP22(orig, dest))
263 #define	FBT_BAA(orig, dest)	(FBT_BA(orig, dest) | FBT_ANNUL)
264 #define	FBT_BL(orig, dest)	(FBT_OP_BL | FBT_DISP22(orig, dest))
265 #define	FBT_BGE(orig, dest)	(FBT_OP_BGE | FBT_DISP22(orig, dest))
266 #define	FBT_BDEST(va, instr)	((uintptr_t)(va) + \
267 	(((int32_t)(((instr) & FBT_DISP22_MASK) << 10)) >> 8))
268 #define	FBT_BPCCDEST(va, instr)	((uintptr_t)(va) + \
269 	(((int32_t)(((instr) & FBT_DISP19_MASK) << 13)) >> 11))
270 #define	FBT_BPRDEST(va, instr)	((uintptr_t)(va) + \
271 	(((int32_t)((FBT_DISP16(instr)) << 16)) >> 14))
272 
273 /*
274  * We're only going to treat a save as safe if (a) both rs1 and rd are
275  * %sp and (b) if the instruction has a simm, the value isn't 0.
276  */
277 #define	FBT_IS_SAVE(instr)	\
278 	(FBT_FMT3_OP(instr) == FBT_OP_SAVE && \
279 	FBT_FMT3_RD(instr) == FBT_REG_O6 && \
280 	FBT_FMT3_RS1(instr) == FBT_REG_O6 && \
281 	!(FBT_FMT3_ISIMM(instr) && FBT_FMT3_SIMM13(instr) == 0))
282 
283 #define	FBT_IS_BA(instr)	(((instr) & ~FBT_DISP22_MASK) == FBT_OP_BA)
284 #define	FBT_IS_BAPCC(instr)	(((instr) & ~FBT_DISP22_MASK) == FBT_OP_BAPCC)
285 
286 #define	FBT_IS_RDPC(instr)	((FBT_FMT3_OP(instr) == FBT_OP_RD) && \
287 	(FBT_FMT3_RD(instr) == FBT_REG_PC))
288 
289 #define	FBT_IS_PCRELATIVE(instr)	\
290 	((((instr) & FBT_OP_MASK) == FBT_OP0 && \
291 	((instr) & FBT_FMT2_OP2_MASK) != FBT_FMT2_OP2_SETHI) || \
292 	((instr) & FBT_OP_MASK) == FBT_OP1 || \
293 	FBT_IS_RDPC(instr))
294 
295 #define	FBT_IS_CTI(instr)	\
296 	((((instr) & FBT_OP_MASK) == FBT_OP0 && \
297 	((instr) & FBT_FMT2_OP2_MASK) != FBT_FMT2_OP2_SETHI) || \
298 	((instr) & FBT_OP_MASK) == FBT_OP1 || \
299 	(FBT_FMT3_OP(instr) == FBT_OP_JMPL) || \
300 	(FBT_FMT3_OP(instr) == FBT_OP_RETURN))
301 
302 #define	FBT_PROBENAME_ENTRY	"entry"
303 #define	FBT_PROBENAME_RETURN	"return"
304 #define	FBT_ESTIMATE_ID		(UINT32_MAX)
305 #define	FBT_COUNTER(id, count)	if ((id) != FBT_ESTIMATE_ID) (count)++
306 
307 #define	FBT_ENTENT_MAXSIZE	(16 * sizeof (uint32_t))
308 #define	FBT_RETENT_MAXSIZE	(11 * sizeof (uint32_t))
309 #define	FBT_RETLENT_MAXSIZE	(23 * sizeof (uint32_t))
310 #define	FBT_ENT_MAXSIZE		\
311 	MAX(MAX(FBT_ENTENT_MAXSIZE, FBT_RETENT_MAXSIZE), FBT_RETLENT_MAXSIZE)
312 
313 typedef struct fbt_probe {
314 	char		*fbtp_name;
315 	dtrace_id_t	fbtp_id;
316 	uintptr_t	fbtp_addr;
317 	struct modctl	*fbtp_ctl;
318 	int		fbtp_loadcnt;
319 	int		fbtp_symndx;
320 	int		fbtp_primary;
321 	int		fbtp_return;
322 	uint32_t	*fbtp_patchpoint;
323 	uint32_t	fbtp_patchval;
324 	uint32_t	fbtp_savedval;
325 	struct fbt_probe *fbtp_next;
326 } fbt_probe_t;
327 
328 typedef struct fbt_trampoline {
329 	uintptr_t	fbtt_va;
330 	uintptr_t	fbtt_limit;
331 	uintptr_t	fbtt_next;
332 } fbt_trampoline_t;
333 
334 static caddr_t
335 fbt_trampoline_map(uintptr_t tramp, size_t size)
336 {
337 	uintptr_t offs;
338 	page_t **ppl;
339 
340 	ASSERT(fbt_trampoline_window == NULL);
341 	ASSERT(fbt_trampoline_size == 0);
342 	ASSERT(fbt_trampoline == NULL);
343 
344 	size += tramp & PAGEOFFSET;
345 	fbt_trampoline = tramp & PAGEMASK;
346 	fbt_trampoline_size = (size + PAGESIZE - 1) & PAGEMASK;
347 	fbt_trampoline_window =
348 	    vmem_alloc(heap_arena, fbt_trampoline_size, VM_SLEEP);
349 
350 	(void) as_pagelock(&kas, &ppl, (caddr_t)fbt_trampoline,
351 	    fbt_trampoline_size, S_WRITE);
352 
353 	for (offs = 0; offs < fbt_trampoline_size; offs += PAGESIZE) {
354 		hat_devload(kas.a_hat, fbt_trampoline_window + offs, PAGESIZE,
355 		    hat_getpfnum(kas.a_hat, (caddr_t)fbt_trampoline + offs),
356 		    PROT_READ | PROT_WRITE,
357 		    HAT_LOAD_LOCK | HAT_LOAD_NOCONSIST);
358 	}
359 
360 	as_pageunlock(&kas, ppl, (caddr_t)fbt_trampoline, fbt_trampoline_size,
361 	    S_WRITE);
362 
363 	return (fbt_trampoline_window + (tramp & PAGEOFFSET));
364 }
365 
366 static void
367 fbt_trampoline_unmap()
368 {
369 	ASSERT(fbt_trampoline_window != NULL);
370 	ASSERT(fbt_trampoline_size != 0);
371 	ASSERT(fbt_trampoline != NULL);
372 
373 	membar_enter();
374 	sync_icache((caddr_t)fbt_trampoline, fbt_trampoline_size);
375 	sync_icache(fbt_trampoline_window, fbt_trampoline_size);
376 
377 	hat_unload(kas.a_hat, fbt_trampoline_window, fbt_trampoline_size,
378 	    HAT_UNLOAD_UNLOCK);
379 
380 	vmem_free(heap_arena, fbt_trampoline_window, fbt_trampoline_size);
381 
382 	fbt_trampoline_window = NULL;
383 	fbt_trampoline = NULL;
384 	fbt_trampoline_size = 0;
385 }
386 
387 static uintptr_t
388 fbt_patch_entry(uint32_t *instr, uint32_t id, fbt_trampoline_t *tramp,
389     int nargs)
390 {
391 	uint32_t *tinstr = (uint32_t *)tramp->fbtt_next;
392 	uint32_t first = *instr;
393 	uintptr_t va = tramp->fbtt_va;
394 	uintptr_t base = tramp->fbtt_next;
395 
396 	if (tramp->fbtt_next + FBT_ENTENT_MAXSIZE > tramp->fbtt_limit) {
397 		/*
398 		 * There isn't sufficient room for this entry; return failure.
399 		 */
400 		return (0);
401 	}
402 
403 	FBT_COUNTER(id, fbt_entry);
404 
405 	if (FBT_IS_SAVE(first)) {
406 		*tinstr++ = first;
407 	} else {
408 		*tinstr++ = FBT_SAVEIMM(FBT_REG_O6, -SA(MINFRAME), FBT_REG_O6);
409 	}
410 
411 	if (id > (uint32_t)FBT_SIMM13_MAX) {
412 		*tinstr++ = FBT_SETHI(id, FBT_REG_O0);
413 		*tinstr++ = FBT_ORLO(FBT_REG_O0, id, FBT_REG_O0);
414 	} else {
415 		*tinstr++ = FBT_ORSIMM13(FBT_REG_G0, id, FBT_REG_O0);
416 	}
417 
418 	if (nargs >= 1)
419 		*tinstr++ = FBT_MOV(FBT_REG_I0, FBT_REG_O1);
420 
421 	if (nargs >= 2)
422 		*tinstr++ = FBT_MOV(FBT_REG_I1, FBT_REG_O2);
423 
424 	if (nargs >= 3)
425 		*tinstr++ = FBT_MOV(FBT_REG_I2, FBT_REG_O3);
426 
427 	if (nargs >= 4)
428 		*tinstr++ = FBT_MOV(FBT_REG_I3, FBT_REG_O4);
429 
430 	if (nargs >= 5)
431 		*tinstr++ = FBT_MOV(FBT_REG_I4, FBT_REG_O5);
432 
433 	if (FBT_IS_SAVE(first)) {
434 		uintptr_t ret = (uintptr_t)instr - sizeof (uint32_t);
435 
436 		*tinstr++ = FBT_SETHI(ret, FBT_REG_G1);
437 		*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dtrace_probe);
438 		tinstr++;
439 		*tinstr++ = FBT_ORLO(FBT_REG_G1, ret, FBT_REG_O7);
440 	} else {
441 		uintptr_t slot = *--tinstr;
442 		uintptr_t ret = (uintptr_t)instr + sizeof (uint32_t);
443 		uint32_t delay = first;
444 
445 		*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dtrace_probe);
446 		tinstr++;
447 		*tinstr++ = slot;
448 		*tinstr++ = FBT_RESTORE(FBT_REG_G0, FBT_REG_G0, FBT_REG_G0);
449 
450 		if (FBT_IS_BA(first) || FBT_IS_BAPCC(first)) {
451 			/*
452 			 * This is a special case:  we are instrumenting a
453 			 * a non-annulled branch-always (or variant).  We'll
454 			 * return directly to the destination of the branch,
455 			 * copying the instruction in the delay slot here,
456 			 * and then executing it in the slot of a ba.
457 			 */
458 			if (FBT_IS_BA(first)) {
459 				ret = FBT_BDEST(instr, *instr);
460 			} else {
461 				ret = FBT_BPCCDEST(instr, *instr);
462 			}
463 
464 			delay = *(instr + 1);
465 		}
466 
467 		if ((first & FBT_OP_MASK) != FBT_OP0 ||
468 		    (first & FBT_FMT2_OP2_MASK) != FBT_FMT2_OP2_BPR) {
469 			*tinstr = FBT_BA((uintptr_t)tinstr - base + va, ret);
470 			tinstr++;
471 			*tinstr++ = delay;
472 		} else {
473 			/*
474 			 * If this is a branch-on-register, we have a little
475 			 * more work to do:  because the displacement is only
476 			 * sixteen bits, we're going to thunk the branch into
477 			 * the trampoline, and then ba,a to the appropriate
478 			 * destination in the branch targets.  That is, we're
479 			 * constructing this sequence in the trampoline:
480 			 *
481 			 *		br[cc]	%[rs], 1f
482 			 *		<delay-instruction>
483 			 *		ba,a	<not-taken-destination>
484 			 *	1:	ba,a	<taken-destination>
485 			 *
486 			 */
487 			uintptr_t targ = FBT_BPRDEST(instr, first);
488 
489 			*tinstr = first & ~(FBT_DISP16_MASK);
490 			*tinstr |= FBT_DISP14(tinstr, &tinstr[3]);
491 			tinstr++;
492 			*tinstr++ = *(instr + 1);
493 			*tinstr = FBT_BAA((uintptr_t)tinstr - base + va,
494 			    ret + sizeof (uint32_t));
495 			tinstr++;
496 			*tinstr = FBT_BAA((uintptr_t)tinstr - base + va, targ);
497 			tinstr++;
498 		}
499 	}
500 
501 	tramp->fbtt_va += (uintptr_t)tinstr - tramp->fbtt_next;
502 	tramp->fbtt_next = (uintptr_t)tinstr;
503 
504 	return (1);
505 }
506 
507 /*
508  * We are patching control-transfer/restore couplets.  There are three
509  * variants of couplet:
510  *
511  * (a)	return		rs1 + imm
512  *	delay
513  *
514  * (b)	jmpl		rs1 + (rs2 | offset), rd
515  *	restore		rs1, rs2 | imm, rd
516  *
517  * (c)	call		displacement
518  *	restore		rs1, rs2 | imm, rd
519  *
520  * If rs1 in (a) is anything other than %i7, or imm is anything other than 8,
521  * or delay is a DCTI, we fail.  If rd from the jmpl in (b) is something other
522  * than %g0 (a ret or a tail-call through a function pointer) or %o7 (a call
523  * through a register), we fail.
524  *
525  * Note that rs1 and rs2 in the restore instructions in (b) and (c) are
526  * potentially outputs and/or globals.  Because these registers cannot be
527  * relied upon across the call to dtrace_probe(), we move rs1 into an unused
528  * local, ls0, and rs2 into an unused local, ls1, and restructure the restore
529  * to be:
530  *
531  *	restore		ls0, ls1, rd
532  *
533  * Likewise, rs1 and rs2 in the jmpl of case (b) may be outputs and/or globals.
534  * If the jmpl uses outputs or globals, we restructure it to be:
535  *
536  * 	jmpl		ls2 + (ls3 | offset), (%g0 | %o7)
537  *
538  */
539 /*ARGSUSED*/
540 static int
541 fbt_canpatch_return(uint32_t *instr, int offset, const char *name)
542 {
543 	int rd;
544 
545 	if (FBT_FMT3_OP(*instr) == FBT_OP_RETURN) {
546 		uint32_t delay = *(instr + 1);
547 
548 		if (*instr != FBT_RETURN(FBT_REG_I7, 8)) {
549 			/*
550 			 * It's unclear if we should warn about this or not.
551 			 * We really wouldn't expect the compiler to generate
552 			 * return instructions with something other than %i7
553 			 * as rs1 and 8 as the simm13 -- it would just be
554 			 * mean-spirited.  That said, such a construct isn't
555 			 * necessarily incorrect.  Sill, we err on the side of
556 			 * caution and warn about it...
557 			 */
558 			cmn_err(CE_NOTE, "cannot instrument return of %s at "
559 			    "%p: non-canonical return instruction", name,
560 			    (void *)instr);
561 			return (0);
562 		}
563 
564 		if (FBT_IS_CTI(delay)) {
565 			/*
566 			 * This is even weirder -- a DCTI coupled with a
567 			 * return instruction.  Similar constructs are used to
568 			 * return from utraps, but these typically have the
569 			 * return in the slot -- and we wouldn't expect to see
570 			 * it in the kernel regardless.  At any rate, we don't
571 			 * want to try to instrument this construct, whatever
572 			 * it may be.
573 			 */
574 			cmn_err(CE_NOTE, "cannot instrument return of %s at "
575 			    "%p: CTI in delay slot of return instruction",
576 			    name, (void *)instr);
577 			return (0);
578 		}
579 
580 		if (FBT_IS_PCRELATIVE(delay)) {
581 			/*
582 			 * This is also very weird, but might be correct code
583 			 * if the function is (for example) returning the
584 			 * address of the delay instruction of the return as
585 			 * its return value (e.g. "rd %pc, %o0" in the slot).
586 			 * Perhaps correct, but still too weird to not warn
587 			 * about it...
588 			 */
589 			cmn_err(CE_NOTE, "cannot instrument return of %s at "
590 			    "%p: PC-relative instruction in delay slot of "
591 			    "return instruction", name, (void *)instr);
592 			return (0);
593 		}
594 
595 		return (1);
596 	}
597 
598 	if (FBT_FMT3_OP(*(instr + 1)) != FBT_OP_RESTORE)
599 		return (0);
600 
601 	if (FBT_FMT1_OP(*instr) == FBT_OP_CALL)
602 		return (1);
603 
604 	if (FBT_FMT3_OP(*instr) != FBT_OP_JMPL)
605 		return (0);
606 
607 	rd = FBT_FMT3_RD(*instr);
608 
609 	if (rd == FBT_REG_I7 || rd == FBT_REG_O7 || rd == FBT_REG_G0)
610 		return (1);
611 
612 	/*
613 	 * We have encountered a jmpl that is storing the calling %pc in
614 	 * some register besides %i7, %o7 or %g0.  This is strange; emit
615 	 * a warning and fail.
616 	 */
617 	cmn_err(CE_NOTE, "cannot instrument return of %s at %p: unexpected "
618 	    "jmpl destination register", name, (void *)instr);
619 	return (0);
620 }
621 
622 static int
623 fbt_canpatch_retl(uint32_t *instr, int offset, const char *name)
624 {
625 	if (FBT_FMT1_OP(*instr) == FBT_OP_CALL ||
626 	    (FBT_FMT3_OP(*instr) == FBT_OP_JMPL &&
627 	    FBT_FMT3_RD(*instr) == FBT_REG_O7)) {
628 		/*
629 		 * If this is a call (or a jmpl that links into %o7), we can
630 		 * patch it iff the next instruction uses %o7 as a destination
631 		 * register.  Because there is an ABI responsibility to
632 		 * restore %o7 to the value before the call/jmpl, we don't
633 		 * particularly care how this routine is managing to restore
634 		 * it (mov, add, ld or divx for all we care).  If it doesn't
635 		 * seem to be restoring it at all, however, we'll refuse
636 		 * to patch it.
637 		 */
638 		uint32_t delay = *(instr + 1);
639 		uint32_t op, rd;
640 
641 		op = FBT_FMT1_OP(delay);
642 		rd = FBT_FMT3_RD(delay);
643 
644 		if (op != FBT_OP2 || rd != FBT_REG_O7) {
645 			/*
646 			 * This is odd.  Before we assume that we're looking
647 			 * at something bizarre (and warn accordingly), we'll
648 			 * check to see if it's obviously a jump table entry.
649 			 */
650 			if (*instr < (uintptr_t)instr &&
651 			    *instr >= (uintptr_t)instr - offset)
652 				return (0);
653 
654 			cmn_err(CE_NOTE, "cannot instrument return of %s at "
655 			    "%p: leaf jmpl/call delay isn't restoring %%o7",
656 			    name, (void *)instr);
657 			return (0);
658 		}
659 
660 		return (1);
661 	}
662 
663 	if (offset == sizeof (uint32_t)) {
664 		/*
665 		 * If this is the second instruction in the function, we're
666 		 * going to allow it to be patched if the first instruction
667 		 * is a patchable return-from-leaf instruction.
668 		 */
669 		if (fbt_canpatch_retl(instr - 1, 0, name))
670 			return (1);
671 	}
672 
673 	if (FBT_FMT3_OP(*instr) != FBT_OP_JMPL)
674 		return (0);
675 
676 	if (FBT_FMT3_RD(*instr) != FBT_REG_G0)
677 		return (0);
678 
679 	return (1);
680 }
681 
682 /*ARGSUSED*/
683 static uint32_t
684 fbt_patch_return(uint32_t *instr, uint32_t *funcbase, uint32_t *funclim,
685     int offset, uint32_t id, fbt_trampoline_t *tramp, const char *name)
686 {
687 	uint32_t *tinstr = (uint32_t *)tramp->fbtt_next;
688 	uint32_t cti = *instr, restore = *(instr + 1), rs1, dest;
689 	uintptr_t va = tramp->fbtt_va;
690 	uintptr_t base = tramp->fbtt_next;
691 	uint32_t locals[FBT_REG_NLOCALS], local;
692 
693 	if (tramp->fbtt_next + FBT_RETENT_MAXSIZE > tramp->fbtt_limit) {
694 		/*
695 		 * There isn't sufficient room for this entry; return failure.
696 		 */
697 		return (FBT_ILLTRAP);
698 	}
699 
700 	FBT_COUNTER(id, fbt_ret);
701 
702 	if (FBT_FMT3_OP(*instr) == FBT_OP_RETURN) {
703 		/*
704 		 * To handle the case of the return instruction, we'll emit a
705 		 * restore, followed by the instruction in the slot (which
706 		 * we'll transplant here), and then another save.  While it
707 		 * may seem intellectually unsatisfying to emit the additional
708 		 * restore/save couplet, one can take solace in the fact that
709 		 * we don't do this if the instruction in the return delay
710 		 * slot is a nop -- which it is nearly 90% of the time with
711 		 * gcc.  (And besides, this couplet can't induce unnecessary
712 		 * spill/fill traps; rewriting the delay instruction to be
713 		 * in terms of the current window hardly seems worth the
714 		 * trouble -- let alone the risk.)
715 		 */
716 		uint32_t delay = *(instr + 1);
717 		ASSERT(*instr == FBT_RETURN(FBT_REG_I7, 8));
718 
719 		cti = FBT_RET;
720 		restore = FBT_RESTORE(FBT_REG_G0, FBT_REG_G0, FBT_REG_G0);
721 
722 		if (delay != FBT_SETHI(0, FBT_REG_G0)) {
723 			*tinstr++ = restore;
724 			*tinstr++ = delay;
725 			*tinstr++ = FBT_SAVEIMM(FBT_REG_O6,
726 			    -SA(MINFRAME), FBT_REG_O6);
727 		}
728 	}
729 
730 	FBT_REG_INITLOCALS(local, locals);
731 
732 	/*
733 	 * Mark the locals used in the jmpl.
734 	 */
735 	if (FBT_FMT3_OP(cti) == FBT_OP_JMPL) {
736 		uint32_t rs1 = FBT_FMT3_RS1(cti);
737 		FBT_REG_MARKLOCAL(locals, rs1);
738 
739 		if (!FBT_FMT3_ISIMM(cti)) {
740 			uint32_t rs2 = FBT_FMT3_RS2(cti);
741 			FBT_REG_MARKLOCAL(locals, rs2);
742 		}
743 	}
744 
745 	/*
746 	 * And mark the locals used in the restore.
747 	 */
748 	rs1 = FBT_FMT3_RS1(restore);
749 	FBT_REG_MARKLOCAL(locals, rs1);
750 
751 	if (!FBT_FMT3_ISIMM(restore)) {
752 		uint32_t rs2 = FBT_FMT3_RS2(restore);
753 		FBT_REG_MARKLOCAL(locals, rs2);
754 	}
755 
756 	if (FBT_FMT3_OP(cti) == FBT_OP_JMPL) {
757 		uint32_t rs1 = FBT_FMT3_RS1(cti);
758 
759 		if (FBT_REG_ISVOLATILE(rs1)) {
760 			FBT_REG_ALLOCLOCAL(local, locals);
761 			FBT_FMT3_RS1_SET(cti, local);
762 			*tinstr++ = FBT_MOV(rs1, local);
763 		}
764 
765 		if (!FBT_FMT3_ISIMM(cti)) {
766 			uint32_t rs2 = FBT_FMT3_RS2(cti);
767 
768 			if (FBT_REG_ISVOLATILE(rs2)) {
769 				FBT_REG_ALLOCLOCAL(local, locals);
770 				FBT_FMT3_RS2_SET(cti, local);
771 				*tinstr++ = FBT_MOV(rs2, local);
772 			}
773 		}
774 	}
775 
776 	rs1 = FBT_FMT3_RS1(restore);
777 
778 	if (FBT_REG_ISVOLATILE(rs1)) {
779 		FBT_REG_ALLOCLOCAL(local, locals);
780 		FBT_FMT3_RS1_SET(restore, local);
781 		*tinstr++ = FBT_MOV(rs1, local);
782 	}
783 
784 	if (!FBT_FMT3_ISIMM(restore)) {
785 		uint32_t rs2 = FBT_FMT3_RS2(restore);
786 
787 		if (FBT_REG_ISVOLATILE(rs2)) {
788 			FBT_REG_ALLOCLOCAL(local, locals);
789 			FBT_FMT3_RS2_SET(restore, local);
790 			*tinstr++ = FBT_MOV(rs2, local);
791 		}
792 	}
793 
794 	if (id > (uint32_t)FBT_SIMM13_MAX) {
795 		*tinstr++ = FBT_SETHI(id, FBT_REG_O0);
796 		*tinstr++ = FBT_ORLO(FBT_REG_O0, id, FBT_REG_O0);
797 	} else {
798 		*tinstr++ = FBT_ORSIMM13(FBT_REG_G0, id, FBT_REG_O0);
799 	}
800 
801 	if (offset > (uint32_t)FBT_SIMM13_MAX) {
802 		*tinstr++ = FBT_SETHI(offset, FBT_REG_O1);
803 		*tinstr++ = FBT_ORLO(FBT_REG_O1, offset, FBT_REG_O1);
804 	} else {
805 		*tinstr++ = FBT_ORSIMM13(FBT_REG_G0, offset, FBT_REG_O1);
806 	}
807 
808 	*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dtrace_probe);
809 	tinstr++;
810 
811 	if (FBT_FMT3_RD(restore) == FBT_REG_O0) {
812 		/*
813 		 * If the destination register of the restore is %o0, we
814 		 * need to perform the implied calculation to derive the
815 		 * return value.
816 		 */
817 		uint32_t add = (restore & ~FBT_FMT3_OP_MASK) | FBT_OP_ADD;
818 		add &= ~FBT_FMT3_RD_MASK;
819 		*tinstr++ = add | (FBT_REG_O2 << FBT_FMT3_RD_SHIFT);
820 	} else {
821 		*tinstr++ = FBT_MOV(FBT_REG_I0, FBT_REG_O2);
822 	}
823 
824 	/*
825 	 * If the control transfer instruction is %pc-relative (i.e. a
826 	 * call), we need to reset it appropriately.
827 	 */
828 	if (FBT_FMT1_OP(cti) == FBT_OP_CALL) {
829 		dest = (uintptr_t)instr + (FBT_FMT1_DISP30(cti) << 2);
830 		*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dest);
831 		tinstr++;
832 	} else {
833 		*tinstr++ = cti;
834 	}
835 
836 	*tinstr++ = restore;
837 	tramp->fbtt_va += (uintptr_t)tinstr - tramp->fbtt_next;
838 	tramp->fbtt_next = (uintptr_t)tinstr;
839 
840 	return (FBT_BAA(instr, va));
841 }
842 
843 static uint32_t
844 fbt_patch_retl(uint32_t *instr, uint32_t *funcbase, uint32_t *funclim,
845     int offset, uint32_t id, fbt_trampoline_t *tramp, const char *name)
846 {
847 	uint32_t *tinstr = (uint32_t *)tramp->fbtt_next;
848 	uintptr_t va = tramp->fbtt_va;
849 	uintptr_t base = tramp->fbtt_next;
850 	uint32_t cti = *instr, dest;
851 	int annul = 0;
852 
853 	FBT_COUNTER(id, fbt_retl);
854 
855 	if (tramp->fbtt_next + FBT_RETLENT_MAXSIZE > tramp->fbtt_limit) {
856 		/*
857 		 * There isn't sufficient room for this entry; return failure.
858 		 */
859 		return (FBT_ILLTRAP);
860 	}
861 
862 	if (offset == sizeof (uint32_t) &&
863 	    fbt_canpatch_retl(instr - 1, 0, name)) {
864 		*tinstr++ = *instr;
865 		annul = 1;
866 		FBT_COUNTER(id, fbt_retl_twoinstr);
867 	} else {
868 		if (FBT_FMT3_OP(cti) == FBT_OP_JMPL &&
869 		    FBT_FMT3_RD(cti) != FBT_REG_O7 &&
870 		    FBT_FMT3_RS1(cti) != FBT_REG_O7) {
871 			annul = 1;
872 			*tinstr++ = *(instr + 1);
873 		}
874 	}
875 
876 	*tinstr++ = FBT_SAVEIMM(FBT_REG_O6, -SA(MINFRAME), FBT_REG_O6);
877 
878 	if (FBT_FMT3_OP(cti) == FBT_OP_JMPL) {
879 		uint32_t rs1, rs2, o2i = FBT_REG_I0 - FBT_REG_O0;
880 
881 		/*
882 		 * If we have a jmpl and it's in terms of output registers, we
883 		 * need to rewrite it to be in terms of the corresponding input
884 		 * registers.  If it's in terms of the globals, we'll rewrite
885 		 * it to be in terms of locals.
886 		 */
887 		rs1 = FBT_FMT3_RS1(cti);
888 
889 		if (FBT_REG_ISOUTPUT(rs1))
890 			rs1 += o2i;
891 
892 		if (FBT_REG_ISGLOBAL(rs1)) {
893 			*tinstr++ = FBT_MOV(rs1, FBT_REG_L0);
894 			rs1 = FBT_REG_L0;
895 		}
896 
897 		FBT_FMT3_RS1_SET(cti, rs1);
898 
899 		if (!FBT_FMT3_ISIMM(cti)) {
900 			rs2 = FBT_FMT3_RS2(cti);
901 
902 			if (FBT_REG_ISOUTPUT(rs2))
903 				rs2 += o2i;
904 
905 			if (FBT_REG_ISGLOBAL(rs2)) {
906 				*tinstr++ = FBT_MOV(rs2, FBT_REG_L1);
907 				rs2 = FBT_REG_L1;
908 			}
909 
910 			FBT_FMT3_RS2_SET(cti, rs2);
911 		}
912 
913 		/*
914 		 * Now we need to check the rd and source register for the jmpl;
915 		 * If neither rd nor the source register is %o7, then we might
916 		 * have a jmp that is actually part of a jump table.  We need
917 		 * to generate the code to compare it to the base and limit of
918 		 * the function.
919 		 */
920 		if (FBT_FMT3_RD(cti) != FBT_REG_O7 && rs1 != FBT_REG_I7) {
921 			uintptr_t base = (uintptr_t)funcbase;
922 			uintptr_t limit = (uintptr_t)funclim;
923 
924 			FBT_COUNTER(id, fbt_retl_jmptab);
925 
926 			if (FBT_FMT3_ISIMM(cti)) {
927 				*tinstr++ = FBT_ADDSIMM13(rs1,
928 				    FBT_FMT3_SIMM13(cti), FBT_REG_L2);
929 			} else {
930 				*tinstr++ = FBT_ADD(rs1, rs2, FBT_REG_L2);
931 			}
932 
933 			*tinstr++ = FBT_SETHI(base, FBT_REG_L3);
934 			*tinstr++ = FBT_ORLO(FBT_REG_L3, base, FBT_REG_L3);
935 			*tinstr++ = FBT_CMP(FBT_REG_L2, FBT_REG_L3);
936 			*tinstr++ = FBT_BL(0, 8 * sizeof (uint32_t));
937 			*tinstr++ = FBT_SETHI(limit, FBT_REG_L3);
938 			*tinstr++ = FBT_ORLO(FBT_REG_L3, limit, FBT_REG_L3);
939 			*tinstr++ = FBT_CMP(FBT_REG_L2, FBT_REG_L3);
940 			*tinstr++ = FBT_BGE(0, 4 * sizeof (uint32_t));
941 			*tinstr++ = FBT_SETHI(0, FBT_REG_G0);
942 			*tinstr++ = cti;
943 			*tinstr++ = FBT_RESTORE(FBT_REG_G0,
944 			    FBT_REG_G0, FBT_REG_G0);
945 		}
946 	}
947 
948 	if (id > (uint32_t)FBT_SIMM13_MAX) {
949 		*tinstr++ = FBT_SETHI(id, FBT_REG_O0);
950 		*tinstr++ = FBT_ORLO(FBT_REG_O0, id, FBT_REG_O0);
951 	} else {
952 		*tinstr++ = FBT_ORSIMM13(FBT_REG_G0, id, FBT_REG_O0);
953 	}
954 
955 	if (offset > (uint32_t)FBT_SIMM13_MAX) {
956 		*tinstr++ = FBT_SETHI(offset, FBT_REG_O1);
957 		*tinstr++ = FBT_ORLO(FBT_REG_O1, offset, FBT_REG_O1);
958 	} else {
959 		*tinstr++ = FBT_ORSIMM13(FBT_REG_G0, offset, FBT_REG_O1);
960 	}
961 
962 	*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dtrace_probe);
963 	tinstr++;
964 	*tinstr++ = FBT_MOV(FBT_REG_I0, FBT_REG_O2);
965 
966 	/*
967 	 * If the control transfer instruction is %pc-relative (i.e. a
968 	 * call), we need to reset it appropriately.
969 	 */
970 	if (FBT_FMT1_OP(cti) == FBT_OP_CALL) {
971 		FBT_COUNTER(id, fbt_retl_tailcall);
972 		dest = (uintptr_t)instr + (FBT_FMT1_DISP30(cti) << 2);
973 		*tinstr = FBT_CALL((uintptr_t)tinstr - base + va, dest);
974 		tinstr++;
975 		annul = 1;
976 	} else {
977 		if (FBT_FMT3_OP(cti) == FBT_OP_JMPL) {
978 			*tinstr++ = cti;
979 
980 			if (FBT_FMT3_RD(cti) == FBT_REG_O7) {
981 				FBT_COUNTER(id, fbt_retl_tailjmpl);
982 				annul = 1;
983 			}
984 		} else {
985 			*tinstr++ = FBT_RET;
986 		}
987 	}
988 
989 	*tinstr++ = FBT_RESTORE(FBT_REG_G0, FBT_REG_G0, FBT_REG_G0);
990 
991 	tramp->fbtt_va += (uintptr_t)tinstr - tramp->fbtt_next;
992 	tramp->fbtt_next = (uintptr_t)tinstr;
993 
994 	return (annul ? FBT_BAA(instr, va) : FBT_BA(instr, va));
995 }
996 
997 /*ARGSUSED*/
998 static void
999 fbt_provide_module(void *arg, struct modctl *ctl)
1000 {
1001 	struct module *mp = ctl->mod_mp;
1002 	char *modname = ctl->mod_modname;
1003 	char *str = mp->strings;
1004 	int nsyms = mp->nsyms;
1005 	Shdr *symhdr = mp->symhdr;
1006 	size_t symsize;
1007 	char *name;
1008 	int i;
1009 	fbt_probe_t *fbt, *retfbt;
1010 	fbt_trampoline_t tramp;
1011 	uintptr_t offset;
1012 	int primary = 0;
1013 	ctf_file_t *fp = NULL;
1014 	int error;
1015 	int estimate = 1;
1016 	uint32_t faketramp[50];
1017 	size_t fbt_size = 0;
1018 
1019 	/*
1020 	 * Employees of dtrace and their families are ineligible.  Void
1021 	 * where prohibited.
1022 	 */
1023 	if (strcmp(modname, "dtrace") == 0)
1024 		return;
1025 
1026 	if (ctl->mod_requisites != NULL) {
1027 		struct modctl_list *list;
1028 
1029 		list = (struct modctl_list *)ctl->mod_requisites;
1030 
1031 		for (; list != NULL; list = list->modl_next) {
1032 			if (strcmp(list->modl_modp->mod_modname, "dtrace") == 0)
1033 				return;
1034 		}
1035 	}
1036 
1037 	/*
1038 	 * KMDB is ineligible for instrumentation -- it may execute in
1039 	 * any context, including probe context.
1040 	 */
1041 	if (strcmp(modname, "kmdbmod") == 0)
1042 		return;
1043 
1044 	if (str == NULL || symhdr == NULL || symhdr->sh_addr == NULL) {
1045 		/*
1046 		 * If this module doesn't (yet) have its string or symbol
1047 		 * table allocated, clear out.
1048 		 */
1049 		return;
1050 	}
1051 
1052 	symsize = symhdr->sh_entsize;
1053 
1054 	if (mp->fbt_nentries) {
1055 		/*
1056 		 * This module has some FBT entries allocated; we're afraid
1057 		 * to screw with it.
1058 		 */
1059 		return;
1060 	}
1061 
1062 	if (mp->fbt_tab != NULL)
1063 		estimate = 0;
1064 
1065 	/*
1066 	 * This is a hack for unix/genunix/krtld.
1067 	 */
1068 	primary = vmem_contains(heap_arena, (void *)ctl,
1069 	    sizeof (struct modctl)) == 0;
1070 	kobj_textwin_alloc(mp);
1071 
1072 	/*
1073 	 * Open the CTF data for the module.  We'll use this to determine the
1074 	 * functions that can be instrumented.  Note that this call can fail,
1075 	 * in which case we'll use heuristics to determine the functions that
1076 	 * can be instrumented.  (But in particular, leaf functions will not be
1077 	 * instrumented.)
1078 	 */
1079 	fp = ctf_modopen(mp, &error);
1080 
1081 forreal:
1082 	if (!estimate) {
1083 		tramp.fbtt_next =
1084 		    (uintptr_t)fbt_trampoline_map((uintptr_t)mp->fbt_tab,
1085 		    mp->fbt_size);
1086 		tramp.fbtt_limit = tramp.fbtt_next + mp->fbt_size;
1087 		tramp.fbtt_va = (uintptr_t)mp->fbt_tab;
1088 	}
1089 
1090 	for (i = 1; i < nsyms; i++) {
1091 		ctf_funcinfo_t f;
1092 		uint32_t *instr, *base, *limit;
1093 		Sym *sym = (Sym *)(symhdr->sh_addr + i * symsize);
1094 		int have_ctf = 0, is_leaf = 0, nargs, cti = 0;
1095 		int (*canpatch)(uint32_t *, int, const char *);
1096 		uint32_t (*patch)(uint32_t *, uint32_t *, uint32_t *, int,
1097 		    uint32_t, fbt_trampoline_t *, const char *);
1098 
1099 		if (ELF_ST_TYPE(sym->st_info) != STT_FUNC)
1100 			continue;
1101 
1102 		/*
1103 		 * Weak symbols are not candidates.  This could be made to
1104 		 * work (where weak functions and their underlying function
1105 		 * appear as two disjoint probes), but it's not simple.
1106 		 */
1107 		if (ELF_ST_BIND(sym->st_info) == STB_WEAK)
1108 			continue;
1109 
1110 		name = str + sym->st_name;
1111 
1112 		if (strstr(name, "dtrace_") == name &&
1113 		    strstr(name, "dtrace_safe_") != name) {
1114 			/*
1115 			 * Anything beginning with "dtrace_" may be called
1116 			 * from probe context unless it explitly indicates
1117 			 * that it won't be called from probe context by
1118 			 * using the prefix "dtrace_safe_".
1119 			 */
1120 			continue;
1121 		}
1122 
1123 		if (strstr(name, "kdi_") == name ||
1124 		    strstr(name, "_kdi_") != NULL) {
1125 			/*
1126 			 * Any function name beginning with "kdi_" or
1127 			 * containing the string "_kdi_" is a part of the
1128 			 * kernel debugger interface and may be called in
1129 			 * arbitrary context -- including probe context.
1130 			 */
1131 			continue;
1132 		}
1133 
1134 		if (strstr(name, "__relocatable") != NULL) {
1135 			/*
1136 			 * Anything with the string "__relocatable" anywhere
1137 			 * in the function name is considered to be a function
1138 			 * that may be manually relocated before execution.
1139 			 * Because FBT uses a PC-relative technique for
1140 			 * instrumentation, these functions cannot safely
1141 			 * be instrumented by us.
1142 			 */
1143 			continue;
1144 		}
1145 
1146 		if (strstr(name, "ip_ocsum") == name) {
1147 			/*
1148 			 * The ip_ocsum_* family of routines are all ABI
1149 			 * violators.  (They expect incoming arguments in the
1150 			 * globals!)  Break the ABI?  No soup for you!
1151 			 */
1152 			continue;
1153 		}
1154 
1155 		/*
1156 		 * We want to scan the function for one (and only one) save.
1157 		 * Any more indicates that something fancy is going on.
1158 		 */
1159 		base = (uint32_t *)sym->st_value;
1160 		limit = (uint32_t *)(sym->st_value + sym->st_size);
1161 
1162 		/*
1163 		 * We don't want to interpose on the module stubs.
1164 		 */
1165 		if (base >= (uint32_t *)stubs_base &&
1166 		    base <= (uint32_t *)stubs_end)
1167 			continue;
1168 
1169 		/*
1170 		 * We can't safely trace a zero-length function...
1171 		 */
1172 		if (base == limit)
1173 			continue;
1174 
1175 		/*
1176 		 * Due to 4524008, _init and _fini may have a bloated st_size.
1177 		 * While this bug was fixed quite some time ago, old drivers
1178 		 * may be lurking.  We need to develop a better solution to
1179 		 * this problem, such that correct _init and _fini functions
1180 		 * (the vast majority) may be correctly traced.  One solution
1181 		 * may be to scan through the entire symbol table to see if
1182 		 * any symbol overlaps with _init.  If none does, set a bit in
1183 		 * the module structure that this module has correct _init and
1184 		 * _fini sizes.  This will cause some pain the first time a
1185 		 * module is scanned, but at least it would be O(N) instead of
1186 		 * O(N log N)...
1187 		 */
1188 		if (strcmp(name, "_init") == 0)
1189 			continue;
1190 
1191 		if (strcmp(name, "_fini") == 0)
1192 			continue;
1193 
1194 		instr = base;
1195 
1196 		/*
1197 		 * While we try hard to only trace safe functions (that is,
1198 		 * functions at TL=0), one unsafe function manages to otherwise
1199 		 * appear safe:  prom_trap().  We could discover prom_trap()
1200 		 * if we added an additional rule:  in order to trace a
1201 		 * function, we must either (a) discover a restore or (b)
1202 		 * determine that the function does not have any unlinked
1203 		 * control transfers to another function (i.e., the function
1204 		 * never returns).  Unfortunately, as of this writing, one
1205 		 * legitimate function (resume_from_zombie()) transfers
1206 		 * control to a different function (_resume_from_idle())
1207 		 * without executing a restore.  Barring a rule to figure out
1208 		 * that resume_from_zombie() is safe while prom_trap() is not,
1209 		 * we resort to hard-coding prom_trap() here.
1210 		 */
1211 		if (strcmp(name, "prom_trap") == 0)
1212 			continue;
1213 
1214 		if (fp != NULL && ctf_func_info(fp, i, &f) != CTF_ERR) {
1215 			nargs = f.ctc_argc;
1216 			have_ctf = 1;
1217 		} else {
1218 			nargs = 32;
1219 		}
1220 
1221 		/*
1222 		 * If the first instruction of the function is a branch and
1223 		 * it's not a branch-always-not-annulled, we're going to refuse
1224 		 * to patch it.
1225 		 */
1226 		if ((*instr & FBT_OP_MASK) == FBT_OP0 &&
1227 		    (*instr & FBT_FMT2_OP2_MASK) != FBT_FMT2_OP2_SETHI &&
1228 		    (*instr & FBT_FMT2_OP2_MASK) != FBT_FMT2_OP2_BPR) {
1229 			if (!FBT_IS_BA(*instr) && !FBT_IS_BAPCC(*instr)) {
1230 				if (have_ctf) {
1231 					cmn_err(CE_NOTE, "cannot instrument %s:"
1232 					    " begins with non-ba, "
1233 					    "non-br CTI", name);
1234 				}
1235 				continue;
1236 			}
1237 		}
1238 
1239 		while (!FBT_IS_SAVE(*instr)) {
1240 			/*
1241 			 * Before we assume that this is a leaf routine, check
1242 			 * forward in the basic block for a save.
1243 			 */
1244 			int op = *instr & FBT_OP_MASK;
1245 			int op2 = *instr & FBT_FMT2_OP2_MASK;
1246 
1247 			if (op == FBT_OP0 && op2 != FBT_FMT2_OP2_SETHI) {
1248 				/*
1249 				 * This is a CTI.  If we see a subsequent
1250 				 * save, we will refuse to process this
1251 				 * routine unless both of the following are
1252 				 * true:
1253 				 *
1254 				 *  (a)	The branch is not annulled
1255 				 *
1256 				 *  (b)	The subsequent save is in the delay
1257 				 *	slot of the branch
1258 				 */
1259 				if ((*instr & FBT_ANNUL) ||
1260 				    !FBT_IS_SAVE(*(instr + 1))) {
1261 					cti = 1;
1262 				} else {
1263 					instr++;
1264 					break;
1265 				}
1266 			}
1267 
1268 			if (op == FBT_OP1)
1269 				cti = 1;
1270 
1271 			if (++instr == limit)
1272 				break;
1273 		}
1274 
1275 		if (instr < limit && cti) {
1276 			/*
1277 			 * If we found a CTI before the save, we need to not
1278 			 * do anything.  But if we have CTF information, this
1279 			 * is weird enough that it merits a message.
1280 			 */
1281 			if (!have_ctf)
1282 				continue;
1283 
1284 			cmn_err(CE_NOTE, "cannot instrument %s: "
1285 			    "save not in first basic block", name);
1286 			continue;
1287 		}
1288 
1289 		if (instr == limit) {
1290 			if (!have_ctf)
1291 				continue;
1292 			is_leaf = 1;
1293 
1294 			if (!estimate)
1295 				fbt_leaf_functions++;
1296 
1297 			canpatch = fbt_canpatch_retl;
1298 			patch = fbt_patch_retl;
1299 		} else {
1300 			canpatch = fbt_canpatch_return;
1301 			patch = fbt_patch_return;
1302 		}
1303 
1304 		if (!have_ctf && !is_leaf) {
1305 			/*
1306 			 * Before we assume that this isn't something tricky,
1307 			 * look for other saves.  If we find them, there are
1308 			 * multiple entry points here (or something), and we'll
1309 			 * leave it alone.
1310 			 */
1311 			while (++instr < limit) {
1312 				if (FBT_IS_SAVE(*instr))
1313 					break;
1314 			}
1315 
1316 			if (instr != limit)
1317 				continue;
1318 		}
1319 
1320 		instr = base;
1321 
1322 		if (FBT_IS_CTI(*instr)) {
1323 			/*
1324 			 * If we have a CTI, we want to be sure that we don't
1325 			 * have a CTI or a PC-relative instruction in the
1326 			 * delay slot -- we want to be able to thunk the
1327 			 * instruction into the trampoline without worrying
1328 			 * about either DCTIs or relocations.  It would be
1329 			 * very odd for the compiler to generate this kind of
1330 			 * code, so we warn about it if we have CTF
1331 			 * information.
1332 			 */
1333 			if (FBT_IS_CTI(*(instr + 1))) {
1334 				if (!have_ctf)
1335 					continue;
1336 
1337 				cmn_err(CE_NOTE, "cannot instrument %s: "
1338 				    "CTI in delay slot of first instruction",
1339 				    name);
1340 				continue;
1341 			}
1342 
1343 			if (FBT_IS_PCRELATIVE(*(instr + 1))) {
1344 				if (!have_ctf)
1345 					continue;
1346 
1347 				cmn_err(CE_NOTE, "cannot instrument %s: "
1348 				    "PC-relative instruction in delay slot of"
1349 				    " first instruction", name);
1350 				continue;
1351 			}
1352 		}
1353 
1354 		if (estimate) {
1355 			tramp.fbtt_next = (uintptr_t)faketramp;
1356 			tramp.fbtt_limit = tramp.fbtt_next + sizeof (faketramp);
1357 			(void) fbt_patch_entry(instr, FBT_ESTIMATE_ID,
1358 			    &tramp, nargs);
1359 			fbt_size += tramp.fbtt_next - (uintptr_t)faketramp;
1360 		} else {
1361 			fbt = kmem_zalloc(sizeof (fbt_probe_t), KM_SLEEP);
1362 			fbt->fbtp_name = name;
1363 			fbt->fbtp_ctl = ctl;
1364 			fbt->fbtp_id = dtrace_probe_create(fbt_id, modname,
1365 			    name, FBT_PROBENAME_ENTRY, 1, fbt);
1366 			fbt->fbtp_patchval = FBT_BAA(instr, tramp.fbtt_va);
1367 
1368 			if (!fbt_patch_entry(instr, fbt->fbtp_id,
1369 			    &tramp, nargs)) {
1370 				cmn_err(CE_WARN, "unexpectedly short FBT table "
1371 				    "in module %s (sym %d of %d)", modname,
1372 				    i, nsyms);
1373 				break;
1374 			}
1375 
1376 			fbt->fbtp_patchpoint =
1377 			    (uint32_t *)((uintptr_t)mp->textwin +
1378 			    ((uintptr_t)instr - (uintptr_t)mp->text));
1379 			fbt->fbtp_savedval = *instr;
1380 
1381 			fbt->fbtp_loadcnt = ctl->mod_loadcnt;
1382 			fbt->fbtp_primary = primary;
1383 			fbt->fbtp_symndx = i;
1384 			mp->fbt_nentries++;
1385 		}
1386 
1387 		retfbt = NULL;
1388 again:
1389 		if (++instr == limit)
1390 			continue;
1391 
1392 		offset = (uintptr_t)instr - (uintptr_t)base;
1393 
1394 		if (!(*canpatch)(instr, offset, name))
1395 			goto again;
1396 
1397 		if (estimate) {
1398 			tramp.fbtt_next = (uintptr_t)faketramp;
1399 			tramp.fbtt_limit = tramp.fbtt_next + sizeof (faketramp);
1400 			(void) (*patch)(instr, base, limit,
1401 			    offset, FBT_ESTIMATE_ID, &tramp, name);
1402 			fbt_size += tramp.fbtt_next - (uintptr_t)faketramp;
1403 
1404 			goto again;
1405 		}
1406 
1407 		fbt = kmem_zalloc(sizeof (fbt_probe_t), KM_SLEEP);
1408 		fbt->fbtp_name = name;
1409 		fbt->fbtp_ctl = ctl;
1410 
1411 		if (retfbt == NULL) {
1412 			fbt->fbtp_id = dtrace_probe_create(fbt_id, modname,
1413 			    name, FBT_PROBENAME_RETURN, 1, fbt);
1414 		} else {
1415 			retfbt->fbtp_next = fbt;
1416 			fbt->fbtp_id = retfbt->fbtp_id;
1417 		}
1418 
1419 		fbt->fbtp_return = 1;
1420 		retfbt = fbt;
1421 
1422 		if ((fbt->fbtp_patchval = (*patch)(instr, base, limit, offset,
1423 		    fbt->fbtp_id, &tramp, name)) == FBT_ILLTRAP) {
1424 			cmn_err(CE_WARN, "unexpectedly short FBT table "
1425 			    "in module %s (sym %d of %d)", modname, i, nsyms);
1426 			break;
1427 		}
1428 
1429 		fbt->fbtp_patchpoint = (uint32_t *)((uintptr_t)mp->textwin +
1430 		    ((uintptr_t)instr - (uintptr_t)mp->text));
1431 		fbt->fbtp_savedval = *instr;
1432 		fbt->fbtp_loadcnt = ctl->mod_loadcnt;
1433 		fbt->fbtp_primary = primary;
1434 		fbt->fbtp_symndx = i;
1435 		mp->fbt_nentries++;
1436 
1437 		goto again;
1438 	}
1439 
1440 	if (estimate) {
1441 		/*
1442 		 * Slosh on another entry's worth...
1443 		 */
1444 		fbt_size += FBT_ENT_MAXSIZE;
1445 		mp->fbt_size = fbt_size;
1446 		mp->fbt_tab = kobj_texthole_alloc(mp->text, fbt_size);
1447 
1448 		if (mp->fbt_tab == NULL) {
1449 			cmn_err(CE_WARN, "couldn't allocate FBT table "
1450 			    "for module %s", modname);
1451 		} else {
1452 			estimate = 0;
1453 			goto forreal;
1454 		}
1455 	} else {
1456 		fbt_trampoline_unmap();
1457 	}
1458 
1459 error:
1460 	if (fp != NULL)
1461 		ctf_close(fp);
1462 }
1463 
1464 /*ARGSUSED*/
1465 static void
1466 fbt_destroy(void *arg, dtrace_id_t id, void *parg)
1467 {
1468 	fbt_probe_t *fbt = parg, *next;
1469 	struct modctl *ctl = fbt->fbtp_ctl;
1470 
1471 	do {
1472 		if (ctl != NULL && ctl->mod_loadcnt == fbt->fbtp_loadcnt) {
1473 			if ((ctl->mod_loadcnt == fbt->fbtp_loadcnt &&
1474 			    ctl->mod_loaded) || fbt->fbtp_primary) {
1475 				((struct module *)
1476 				    (ctl->mod_mp))->fbt_nentries--;
1477 			}
1478 		}
1479 
1480 		next = fbt->fbtp_next;
1481 		kmem_free(fbt, sizeof (fbt_probe_t));
1482 		fbt = next;
1483 	} while (fbt != NULL);
1484 }
1485 
1486 /*ARGSUSED*/
1487 static void
1488 fbt_enable(void *arg, dtrace_id_t id, void *parg)
1489 {
1490 	fbt_probe_t *fbt = parg, *f;
1491 	struct modctl *ctl = fbt->fbtp_ctl;
1492 
1493 	ctl->mod_nenabled++;
1494 
1495 	for (f = fbt; f != NULL; f = f->fbtp_next) {
1496 		if (f->fbtp_patchpoint == NULL) {
1497 			/*
1498 			 * Due to a shortened FBT table, this entry was never
1499 			 * completed; refuse to enable it.
1500 			 */
1501 			if (fbt_verbose) {
1502 				cmn_err(CE_NOTE, "fbt is failing for probe %s "
1503 				    "(short FBT table in %s)",
1504 				    fbt->fbtp_name, ctl->mod_modname);
1505 			}
1506 
1507 			return;
1508 		}
1509 	}
1510 
1511 	/*
1512 	 * If this module has disappeared since we discovered its probes,
1513 	 * refuse to enable it.
1514 	 */
1515 	if (!fbt->fbtp_primary && !ctl->mod_loaded) {
1516 		if (fbt_verbose) {
1517 			cmn_err(CE_NOTE, "fbt is failing for probe %s "
1518 			    "(module %s unloaded)",
1519 			    fbt->fbtp_name, ctl->mod_modname);
1520 		}
1521 
1522 		return;
1523 	}
1524 
1525 	/*
1526 	 * Now check that our modctl has the expected load count.  If it
1527 	 * doesn't, this module must have been unloaded and reloaded -- and
1528 	 * we're not going to touch it.
1529 	 */
1530 	if (ctl->mod_loadcnt != fbt->fbtp_loadcnt) {
1531 		if (fbt_verbose) {
1532 			cmn_err(CE_NOTE, "fbt is failing for probe %s "
1533 			    "(module %s reloaded)",
1534 			    fbt->fbtp_name, ctl->mod_modname);
1535 		}
1536 
1537 		return;
1538 	}
1539 
1540 	for (; fbt != NULL; fbt = fbt->fbtp_next)
1541 		*fbt->fbtp_patchpoint = fbt->fbtp_patchval;
1542 }
1543 
1544 /*ARGSUSED*/
1545 static void
1546 fbt_disable(void *arg, dtrace_id_t id, void *parg)
1547 {
1548 	fbt_probe_t *fbt = parg, *f;
1549 	struct modctl *ctl = fbt->fbtp_ctl;
1550 
1551 	ASSERT(ctl->mod_nenabled > 0);
1552 	ctl->mod_nenabled--;
1553 
1554 	for (f = fbt; f != NULL; f = f->fbtp_next) {
1555 		if (f->fbtp_patchpoint == NULL)
1556 			return;
1557 	}
1558 
1559 	if ((!fbt->fbtp_primary && !ctl->mod_loaded) ||
1560 	    (ctl->mod_loadcnt != fbt->fbtp_loadcnt))
1561 		return;
1562 
1563 	for (; fbt != NULL; fbt = fbt->fbtp_next)
1564 		*fbt->fbtp_patchpoint = fbt->fbtp_savedval;
1565 }
1566 
1567 /*ARGSUSED*/
1568 static void
1569 fbt_suspend(void *arg, dtrace_id_t id, void *parg)
1570 {
1571 	fbt_probe_t *fbt = parg;
1572 	struct modctl *ctl = fbt->fbtp_ctl;
1573 
1574 	if (!fbt->fbtp_primary && !ctl->mod_loaded)
1575 		return;
1576 
1577 	if (ctl->mod_loadcnt != fbt->fbtp_loadcnt)
1578 		return;
1579 
1580 	ASSERT(ctl->mod_nenabled > 0);
1581 
1582 	for (; fbt != NULL; fbt = fbt->fbtp_next)
1583 		*fbt->fbtp_patchpoint = fbt->fbtp_savedval;
1584 }
1585 
1586 /*ARGSUSED*/
1587 static void
1588 fbt_resume(void *arg, dtrace_id_t id, void *parg)
1589 {
1590 	fbt_probe_t *fbt = parg;
1591 	struct modctl *ctl = fbt->fbtp_ctl;
1592 
1593 	if (!fbt->fbtp_primary && !ctl->mod_loaded)
1594 		return;
1595 
1596 	if (ctl->mod_loadcnt != fbt->fbtp_loadcnt)
1597 		return;
1598 
1599 	ASSERT(ctl->mod_nenabled > 0);
1600 
1601 	for (; fbt != NULL; fbt = fbt->fbtp_next)
1602 		*fbt->fbtp_patchpoint = fbt->fbtp_patchval;
1603 }
1604 
1605 /*ARGSUSED*/
1606 static void
1607 fbt_getargdesc(void *arg, dtrace_id_t id, void *parg, dtrace_argdesc_t *desc)
1608 {
1609 	fbt_probe_t *fbt = parg;
1610 	struct modctl *ctl = fbt->fbtp_ctl;
1611 	struct module *mp = ctl->mod_mp;
1612 	ctf_file_t *fp = NULL, *pfp;
1613 	ctf_funcinfo_t f;
1614 	int error;
1615 	ctf_id_t argv[32], type;
1616 	int argc = sizeof (argv) / sizeof (ctf_id_t);
1617 	const char *parent;
1618 
1619 	if (!ctl->mod_loaded || (ctl->mod_loadcnt != fbt->fbtp_loadcnt))
1620 		goto err;
1621 
1622 	if (fbt->fbtp_return && desc->dtargd_ndx == 0) {
1623 		(void) strcpy(desc->dtargd_native, "int");
1624 		return;
1625 	}
1626 
1627 	if ((fp = ctf_modopen(mp, &error)) == NULL) {
1628 		/*
1629 		 * We have no CTF information for this module -- and therefore
1630 		 * no args[] information.
1631 		 */
1632 		goto err;
1633 	}
1634 
1635 	/*
1636 	 * If we have a parent container, we must manually import it.
1637 	 */
1638 	if ((parent = ctf_parent_name(fp)) != NULL) {
1639 		struct modctl *mp = &modules;
1640 		struct modctl *mod = NULL;
1641 
1642 		/*
1643 		 * We must iterate over all modules to find the module that
1644 		 * is our parent.
1645 		 */
1646 		do {
1647 			if (strcmp(mp->mod_modname, parent) == 0) {
1648 				mod = mp;
1649 				break;
1650 			}
1651 		} while ((mp = mp->mod_next) != &modules);
1652 
1653 		if (mod == NULL)
1654 			goto err;
1655 
1656 		if ((pfp = ctf_modopen(mod->mod_mp, &error)) == NULL)
1657 			goto err;
1658 
1659 		if (ctf_import(fp, pfp) != 0) {
1660 			ctf_close(pfp);
1661 			goto err;
1662 		}
1663 
1664 		ctf_close(pfp);
1665 	}
1666 
1667 	if (ctf_func_info(fp, fbt->fbtp_symndx, &f) == CTF_ERR)
1668 		goto err;
1669 
1670 	if (fbt->fbtp_return) {
1671 		if (desc->dtargd_ndx > 1)
1672 			goto err;
1673 
1674 		ASSERT(desc->dtargd_ndx == 1);
1675 		type = f.ctc_return;
1676 	} else {
1677 		if (desc->dtargd_ndx + 1 > f.ctc_argc)
1678 			goto err;
1679 
1680 		if (ctf_func_args(fp, fbt->fbtp_symndx, argc, argv) == CTF_ERR)
1681 			goto err;
1682 
1683 		type = argv[desc->dtargd_ndx];
1684 	}
1685 
1686 	if (ctf_type_name(fp, type, desc->dtargd_native,
1687 	    DTRACE_ARGTYPELEN) != NULL) {
1688 		ctf_close(fp);
1689 		return;
1690 	}
1691 err:
1692 	if (fp != NULL)
1693 		ctf_close(fp);
1694 
1695 	desc->dtargd_ndx = DTRACE_ARGNONE;
1696 }
1697 
1698 static dtrace_pattr_t fbt_attr = {
1699 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1700 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1701 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1702 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1703 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
1704 };
1705 
1706 static dtrace_pops_t fbt_pops = {
1707 	NULL,
1708 	fbt_provide_module,
1709 	fbt_enable,
1710 	fbt_disable,
1711 	fbt_suspend,
1712 	fbt_resume,
1713 	fbt_getargdesc,
1714 	NULL,
1715 	NULL,
1716 	fbt_destroy
1717 };
1718 
1719 static int
1720 fbt_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
1721 {
1722 	switch (cmd) {
1723 	case DDI_ATTACH:
1724 		break;
1725 	case DDI_RESUME:
1726 		return (DDI_SUCCESS);
1727 	default:
1728 		return (DDI_FAILURE);
1729 	}
1730 
1731 	if (ddi_create_minor_node(devi, "fbt", S_IFCHR, 0,
1732 	    DDI_PSEUDO, NULL) == DDI_FAILURE ||
1733 	    dtrace_register("fbt", &fbt_attr, DTRACE_PRIV_KERNEL, NULL,
1734 	    &fbt_pops, NULL, &fbt_id) != 0) {
1735 		ddi_remove_minor_node(devi, NULL);
1736 		return (DDI_FAILURE);
1737 	}
1738 
1739 	ddi_report_dev(devi);
1740 	fbt_devi = devi;
1741 	return (DDI_SUCCESS);
1742 }
1743 
1744 static int
1745 fbt_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
1746 {
1747 	switch (cmd) {
1748 	case DDI_DETACH:
1749 		break;
1750 	case DDI_SUSPEND:
1751 		return (DDI_SUCCESS);
1752 	default:
1753 		return (DDI_FAILURE);
1754 	}
1755 
1756 	if (dtrace_unregister(fbt_id) != 0)
1757 		return (DDI_FAILURE);
1758 
1759 	ddi_remove_minor_node(devi, NULL);
1760 	return (DDI_SUCCESS);
1761 }
1762 
1763 /*ARGSUSED*/
1764 static int
1765 fbt_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
1766 {
1767 	int error;
1768 
1769 	switch (infocmd) {
1770 	case DDI_INFO_DEVT2DEVINFO:
1771 		*result = (void *)fbt_devi;
1772 		error = DDI_SUCCESS;
1773 		break;
1774 	case DDI_INFO_DEVT2INSTANCE:
1775 		*result = (void *)0;
1776 		error = DDI_SUCCESS;
1777 		break;
1778 	default:
1779 		error = DDI_FAILURE;
1780 	}
1781 	return (error);
1782 }
1783 
1784 /*ARGSUSED*/
1785 static int
1786 fbt_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
1787 {
1788 	return (0);
1789 }
1790 
1791 static struct cb_ops fbt_cb_ops = {
1792 	fbt_open,		/* open */
1793 	nodev,			/* close */
1794 	nulldev,		/* strategy */
1795 	nulldev,		/* print */
1796 	nodev,			/* dump */
1797 	nodev,			/* read */
1798 	nodev,			/* write */
1799 	nodev,			/* ioctl */
1800 	nodev,			/* devmap */
1801 	nodev,			/* mmap */
1802 	nodev,			/* segmap */
1803 	nochpoll,		/* poll */
1804 	ddi_prop_op,		/* cb_prop_op */
1805 	0,			/* streamtab  */
1806 	D_NEW | D_MP		/* Driver compatibility flag */
1807 };
1808 
1809 static struct dev_ops fbt_ops = {
1810 	DEVO_REV,		/* devo_rev */
1811 	0,			/* refcnt */
1812 	fbt_info,		/* get_dev_info */
1813 	nulldev,		/* identify */
1814 	nulldev,		/* probe */
1815 	fbt_attach,		/* attach */
1816 	fbt_detach,		/* detach */
1817 	nodev,			/* reset */
1818 	&fbt_cb_ops,		/* driver operations */
1819 	NULL,			/* bus operations */
1820 	nodev			/* dev power */
1821 };
1822 
1823 /*
1824  * Module linkage information for the kernel.
1825  */
1826 static struct modldrv modldrv = {
1827 	&mod_driverops,		/* module type (this is a pseudo driver) */
1828 	"Function Boundary Tracing",	/* name of module */
1829 	&fbt_ops,		/* driver ops */
1830 };
1831 
1832 static struct modlinkage modlinkage = {
1833 	MODREV_1,
1834 	(void *)&modldrv,
1835 	NULL
1836 };
1837 
1838 int
1839 _init(void)
1840 {
1841 	return (mod_install(&modlinkage));
1842 }
1843 
1844 int
1845 _info(struct modinfo *modinfop)
1846 {
1847 	return (mod_info(&modlinkage, modinfop));
1848 }
1849 
1850 int
1851 _fini(void)
1852 {
1853 	return (mod_remove(&modlinkage));
1854 }
1855