xref: /illumos-gate/usr/src/lib/libproc/common/Pstack.c (revision 45ede40b2394db7967e59f19288fae9b62efd4aa)
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 /*
23  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Pstack.c
29  *
30  * Common helper functions for stack walking.  The ISA-specific code is found in
31  * Pstack_iter() in Pisadep.c.
32  */
33 
34 #include <stdlib.h>
35 #include <unistd.h>
36 #include <string.h>
37 #include <errno.h>
38 
39 #include "libproc.h"
40 #include "Pcontrol.h"
41 #include "P32ton.h"
42 #include "Pstack.h"
43 
44 /*
45  * Utility function to prevent stack loops from running on forever by
46  * detecting when there is a stack loop (the %fp has been seen before).
47  */
48 int
49 stack_loop(prgreg_t fp, prgreg_t **prevfpp, int *nfpp, uint_t *pfpsizep)
50 {
51 	prgreg_t *prevfp = *prevfpp;
52 	uint_t pfpsize = *pfpsizep;
53 	int nfp = *nfpp;
54 	int i;
55 
56 	for (i = 0; i < nfp; i++) {
57 		if (fp == prevfp[i])
58 			return (1); /* stack loop detected */
59 	}
60 
61 	if (nfp == pfpsize) {
62 		pfpsize = pfpsize ? pfpsize * 2 : 16;
63 		prevfp = realloc(prevfp, pfpsize * sizeof (prgreg_t));
64 		/*
65 		 * Just assume there is no loop in the face of allocation
66 		 * failure; the caller still has the original prevfp pointer.
67 		 */
68 		if (prevfp == NULL)
69 			return (0);
70 	}
71 
72 	prevfp[nfp++] = fp;
73 	*prevfpp = prevfp;
74 	*pfpsizep = pfpsize;
75 	*nfpp = nfp;
76 
77 	return (0);
78 }
79 
80 /*
81  * Signal Frame Detection
82  *
83  * In order to facilitate detection and processing of signal handler frames
84  * during a stack backtrace, we define a set of utility routines to operate on
85  * a uclist (ucontext address list), and then use these routines in the various
86  * implementations of Pstack_iter below.  Certain source-level debuggers and
87  * virtual machines that shall remain nameless believe that in order to detect
88  * signal handler frames, one must hard-code checks for symbol names defined
89  * in libc and libthread and knowledge of their implementation.  We make no
90  * such assumptions, allowing us to operate on programs that manipulate their
91  * underlying kernel signal handlers (i.e. use __sigaction) and to not require
92  * changes in the face of future library modifications.
93  *
94  * A signal handler frame is essentially a set of data pushed on to the user
95  * stack by the kernel prior to returning to the user program in one of the
96  * pre-defined signal handlers.  The signal handler itself receives the signal
97  * number, an optional pointer to a siginfo_t, and a pointer to the interrupted
98  * ucontext as arguments.  When performing a stack backtrace, we would like to
99  * detect these frames so that we can correctly return the interrupted program
100  * counter and frame pointer as a separate frame.  When a signal handler frame
101  * is constructed on the stack by the kernel, the signalled LWP has its
102  * lwp_oldcontext member (exported through /proc as lwpstatus.pr_oldcontext)
103  * set to the user address at which the ucontext_t was placed on the LWP's
104  * stack.  The ucontext_t's uc_link member is set to the previous value of
105  * lwp_oldcontext.  Thus when signal handlers are active, pr_oldcontext will
106  * point to the first element of a linked list of ucontext_t addresses.
107  *
108  * The stack layout for a signal handler frame is as follows:
109  *
110  * SPARC v7/v9:                           Intel ia32:
111  * +--------------+ -        high         +--------------+ -
112  * |  struct fq   | ^        addrs        |  siginfo_t   | optional
113  * +--------------+ |          ^          +--------------+ -
114  * |  gwindows_t  |            |          |  ucontext_t  | ^
115  * +--------------+ optional              +--------------+ |
116  * |  siginfo_t   |                       | ucontext_t * | |
117  * +--------------+ |          |          +--------------+
118  * |  xregs data  | v          v          |  siginfo_t * | mandatory
119  * +--------------+ -         low         +--------------+
120  * |  ucontext_t  | ^        addrs        |  int (signo) | |
121  * +--------------+ mandatory             +--------------+ |
122  * | struct frame | v                     | struct frame | v
123  * +--------------+ - <- %sp on resume    +--------------+ - <- %esp on resume
124  *
125  * amd64 (64-bit):
126  * +--------------+ -
127  * |  siginfo_t   | optional
128  * +--------------+ -
129  * |  ucontext_t  | ^
130  * +--------------+ |
131  * |  siginfo_t * |
132  * +--------------+ mandatory
133  * |  int (signo) |
134  * +--------------+ |
135  * | struct frame | v
136  * +--------------+ - <- %rsp on resume
137  *
138  * The bottom-most struct frame is actually constructed by the kernel by
139  * copying the previous stack frame, allowing naive backtrace code to simply
140  * skip over the interrupted frame.  The copied frame is never really used,
141  * since it is presumed the libc or libthread signal handler wrapper function
142  * will explicitly setcontext(2) to the interrupted context if the user
143  * program's handler returns.  If we detect a signal handler frame, we simply
144  * read the interrupted context structure from the stack, use its embedded
145  * gregs to construct the register set for the interrupted frame, and then
146  * continue our backtrace.  Detecting the frame itself is easy according to
147  * the diagram ("oldcontext" represents any element in the uc_link chain):
148  *
149  * On SPARC v7 or v9:
150  * %fp + sizeof (struct frame) == oldcontext
151  *
152  * On Intel ia32:
153  * %ebp + sizeof (struct frame) + (3 * regsize) == oldcontext
154  *
155  * On amd64:
156  * %rbp + sizeof (struct frame) + (2 * regsize) == oldcontext
157  *
158  * A final complication is that we want libproc to support backtraces from
159  * arbitrary addresses without the caller passing in an LWP id.  To do this,
160  * we must first determine all the known oldcontexts by iterating over all
161  * LWPs and following their pr_oldcontext pointers.  We optimize our search
162  * by discarding NULL pointers and pointers whose value is less than that
163  * of the initial stack pointer (since stacks grow down from high memory),
164  * and then sort the resulting list by virtual address so we can binary search.
165  */
166 
167 int
168 load_uclist(uclist_t *ucl, const lwpstatus_t *psp)
169 {
170 	struct ps_prochandle *P = ucl->uc_proc;
171 	uintptr_t addr = psp->pr_oldcontext;
172 
173 	uintptr_t *new_addrs;
174 	uint_t new_size, i;
175 	ucontext_t uc;
176 
177 	if (addr == (uintptr_t)NULL)
178 		return (0);
179 
180 	for (;;) {
181 		if (ucl->uc_nelems == ucl->uc_size) {
182 			new_size = ucl->uc_size ? ucl->uc_size * 2 : 16;
183 			new_addrs = realloc(ucl->uc_addrs,
184 			    new_size * sizeof (uintptr_t));
185 
186 			if (new_addrs != NULL) {
187 				ucl->uc_addrs = new_addrs;
188 				ucl->uc_size = new_size;
189 			} else
190 				break; /* abort if allocation failure */
191 		}
192 #ifdef _LP64
193 		if (P->status.pr_dmodel == PR_MODEL_ILP32) {
194 			ucontext32_t u32;
195 
196 			if (Pread(P, &u32, sizeof (u32), addr) != sizeof (u32))
197 				break; /* abort if we fail to read ucontext */
198 			uc.uc_link = (ucontext_t *)(uintptr_t)u32.uc_link;
199 		} else
200 #endif
201 		if (Pread(P, &uc, sizeof (uc), addr) != sizeof (uc))
202 			break; /* abort if we fail to read ucontext */
203 
204 		dprintf("detected lwp %d signal context at %p\n",
205 		    (int)psp->pr_lwpid, (void *)addr);
206 		ucl->uc_addrs[ucl->uc_nelems++] = addr;
207 
208 		addr = (uintptr_t)uc.uc_link;
209 
210 		/*
211 		 * Abort if we find a NULL uc_link pointer or a duplicate
212 		 * entry which could indicate a cycle or a very peculiar
213 		 * interference pattern between threads.
214 		 */
215 		if (addr == (uintptr_t)NULL)
216 			break;
217 
218 		for (i = 0; i < ucl->uc_nelems - 1; i++) {
219 			if (ucl->uc_addrs[i] == addr)
220 				return (0);
221 		}
222 	}
223 
224 	return (0);
225 }
226 
227 int
228 sort_uclist(const void *lhp, const void *rhp)
229 {
230 	uintptr_t lhs = *((const uintptr_t *)lhp);
231 	uintptr_t rhs = *((const uintptr_t *)rhp);
232 
233 	if (lhs < rhs)
234 		return (-1);
235 	if (lhs > rhs)
236 		return (+1);
237 	return (0);
238 }
239 
240 void
241 init_uclist(uclist_t *ucl, struct ps_prochandle *P)
242 {
243 	if ((P->state == PS_STOP || P->state == PS_DEAD) &&
244 	    P->ucaddrs != NULL) {
245 		ucl->uc_proc = P;
246 		ucl->uc_addrs = P->ucaddrs;
247 		ucl->uc_nelems = P->ucnelems;
248 		ucl->uc_size = P->ucnelems;
249 		ucl->uc_cached = 1;
250 		return;
251 	}
252 
253 	ucl->uc_proc = P;
254 	ucl->uc_addrs = NULL;
255 	ucl->uc_nelems = 0;
256 	ucl->uc_size = 0;
257 
258 	(void) Plwp_iter(P, (proc_lwp_f *)load_uclist, ucl);
259 	qsort(ucl->uc_addrs, ucl->uc_nelems, sizeof (uintptr_t), sort_uclist);
260 
261 	if (P->state == PS_STOP || P->state == PS_DEAD) {
262 		P->ucaddrs = ucl->uc_addrs;
263 		P->ucnelems = ucl->uc_nelems;
264 		ucl->uc_cached = 1;
265 	} else {
266 		ucl->uc_cached = 0;
267 	}
268 }
269 
270 void
271 free_uclist(uclist_t *ucl)
272 {
273 	if (!ucl->uc_cached && ucl->uc_addrs != NULL)
274 		free(ucl->uc_addrs);
275 }
276 
277 int
278 find_uclink(uclist_t *ucl, uintptr_t addr)
279 {
280 	if (ucl->uc_nelems != 0) {
281 		return (bsearch(&addr, ucl->uc_addrs, ucl->uc_nelems,
282 		    sizeof (uintptr_t), sort_uclist) != NULL);
283 	}
284 
285 	return (0);
286 }
287