1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Use DWARF Debug information to skip unnecessary callchain entries.
4  *
5  * Copyright (C) 2014 Sukadev Bhattiprolu, IBM Corporation.
6  * Copyright (C) 2014 Ulrich Weigand, IBM Corporation.
7  */
8 #include <inttypes.h>
9 #include <dwarf.h>
10 #include <elfutils/libdwfl.h>
11 
12 #include "util/thread.h"
13 #include "util/callchain.h"
14 #include "util/debug.h"
15 #include "util/dso.h"
16 #include "util/event.h" // struct ip_callchain
17 #include "util/map.h"
18 #include "util/symbol.h"
19 
20 /*
21  * When saving the callchain on Power, the kernel conservatively saves
22  * excess entries in the callchain. A few of these entries are needed
23  * in some cases but not others. If the unnecessary entries are not
24  * ignored, we end up with duplicate arcs in the call-graphs. Use
25  * DWARF debug information to skip over any unnecessary callchain
26  * entries.
27  *
28  * See function header for arch_adjust_callchain() below for more details.
29  *
30  * The libdwfl code in this file is based on code from elfutils
31  * (libdwfl/argp-std.c, libdwfl/tests/addrcfi.c, etc).
32  */
33 static char *debuginfo_path;
34 
35 static const Dwfl_Callbacks offline_callbacks = {
36 	.debuginfo_path = &debuginfo_path,
37 	.find_debuginfo = dwfl_standard_find_debuginfo,
38 	.section_address = dwfl_offline_section_address,
39 };
40 
41 
42 /*
43  * Use the DWARF expression for the Call-frame-address and determine
44  * if return address is in LR and if a new frame was allocated.
45  */
check_return_reg(int ra_regno,Dwarf_Frame * frame)46 static int check_return_reg(int ra_regno, Dwarf_Frame *frame)
47 {
48 	Dwarf_Op ops_mem[3];
49 	Dwarf_Op dummy;
50 	Dwarf_Op *ops = &dummy;
51 	size_t nops;
52 	int result;
53 
54 	result = dwarf_frame_register(frame, ra_regno, ops_mem, &ops, &nops);
55 	if (result < 0) {
56 		pr_debug("dwarf_frame_register() %s\n", dwarf_errmsg(-1));
57 		return -1;
58 	}
59 
60 	/*
61 	 * Check if return address is on the stack. If return address
62 	 * is in a register (typically R0), it is yet to be saved on
63 	 * the stack.
64 	 */
65 	if ((nops != 0 || ops != NULL) &&
66 		!(nops == 1 && ops[0].atom == DW_OP_regx &&
67 			ops[0].number2 == 0 && ops[0].offset == 0))
68 		return 0;
69 
70 	/*
71 	 * Return address is in LR. Check if a frame was allocated
72 	 * but not-yet used.
73 	 */
74 	result = dwarf_frame_cfa(frame, &ops, &nops);
75 	if (result < 0) {
76 		pr_debug("dwarf_frame_cfa() returns %d, %s\n", result,
77 					dwarf_errmsg(-1));
78 		return -1;
79 	}
80 
81 	/*
82 	 * If call frame address is in r1, no new frame was allocated.
83 	 */
84 	if (nops == 1 && ops[0].atom == DW_OP_bregx && ops[0].number == 1 &&
85 				ops[0].number2 == 0)
86 		return 1;
87 
88 	/*
89 	 * A new frame was allocated but has not yet been used.
90 	 */
91 	return 2;
92 }
93 
94 /*
95  * Get the DWARF frame from the .eh_frame section.
96  */
get_eh_frame(Dwfl_Module * mod,Dwarf_Addr pc)97 static Dwarf_Frame *get_eh_frame(Dwfl_Module *mod, Dwarf_Addr pc)
98 {
99 	int		result;
100 	Dwarf_Addr	bias;
101 	Dwarf_CFI	*cfi;
102 	Dwarf_Frame	*frame;
103 
104 	cfi = dwfl_module_eh_cfi(mod, &bias);
105 	if (!cfi) {
106 		pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
107 		return NULL;
108 	}
109 
110 	result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
111 	if (result) {
112 		pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
113 		return NULL;
114 	}
115 
116 	return frame;
117 }
118 
119 /*
120  * Get the DWARF frame from the .debug_frame section.
121  */
get_dwarf_frame(Dwfl_Module * mod,Dwarf_Addr pc)122 static Dwarf_Frame *get_dwarf_frame(Dwfl_Module *mod, Dwarf_Addr pc)
123 {
124 	Dwarf_CFI       *cfi;
125 	Dwarf_Addr      bias;
126 	Dwarf_Frame     *frame;
127 	int             result;
128 
129 	cfi = dwfl_module_dwarf_cfi(mod, &bias);
130 	if (!cfi) {
131 		pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
132 		return NULL;
133 	}
134 
135 	result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
136 	if (result) {
137 		pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
138 		return NULL;
139 	}
140 
141 	return frame;
142 }
143 
144 /*
145  * Return:
146  *	0 if return address for the program counter @pc is on stack
147  *	1 if return address is in LR and no new stack frame was allocated
148  *	2 if return address is in LR and a new frame was allocated (but not
149  *		yet used)
150  *	-1 in case of errors
151  */
check_return_addr(struct dso * dso,u64 map_start,Dwarf_Addr pc)152 static int check_return_addr(struct dso *dso, u64 map_start, Dwarf_Addr pc)
153 {
154 	int		rc = -1;
155 	Dwfl		*dwfl;
156 	Dwfl_Module	*mod;
157 	Dwarf_Frame	*frame;
158 	int		ra_regno;
159 	Dwarf_Addr	start = pc;
160 	Dwarf_Addr	end = pc;
161 	bool		signalp;
162 	const char	*exec_file = dso->long_name;
163 
164 	dwfl = dso->dwfl;
165 
166 	if (!dwfl) {
167 		dwfl = dwfl_begin(&offline_callbacks);
168 		if (!dwfl) {
169 			pr_debug("dwfl_begin() failed: %s\n", dwarf_errmsg(-1));
170 			return -1;
171 		}
172 
173 		mod = dwfl_report_elf(dwfl, exec_file, exec_file, -1,
174 						map_start, false);
175 		if (!mod) {
176 			pr_debug("dwfl_report_elf() failed %s\n",
177 						dwarf_errmsg(-1));
178 			/*
179 			 * We normally cache the DWARF debug info and never
180 			 * call dwfl_end(). But to prevent fd leak, free in
181 			 * case of error.
182 			 */
183 			dwfl_end(dwfl);
184 			goto out;
185 		}
186 		dso->dwfl = dwfl;
187 	}
188 
189 	mod = dwfl_addrmodule(dwfl, pc);
190 	if (!mod) {
191 		pr_debug("dwfl_addrmodule() failed, %s\n", dwarf_errmsg(-1));
192 		goto out;
193 	}
194 
195 	/*
196 	 * To work with split debug info files (eg: glibc), check both
197 	 * .eh_frame and .debug_frame sections of the ELF header.
198 	 */
199 	frame = get_eh_frame(mod, pc);
200 	if (!frame) {
201 		frame = get_dwarf_frame(mod, pc);
202 		if (!frame)
203 			goto out;
204 	}
205 
206 	ra_regno = dwarf_frame_info(frame, &start, &end, &signalp);
207 	if (ra_regno < 0) {
208 		pr_debug("Return address register unavailable: %s\n",
209 				dwarf_errmsg(-1));
210 		goto out;
211 	}
212 
213 	rc = check_return_reg(ra_regno, frame);
214 
215 out:
216 	return rc;
217 }
218 
219 /*
220  * The callchain saved by the kernel always includes the link register (LR).
221  *
222  *	0:	PERF_CONTEXT_USER
223  *	1:	Program counter (Next instruction pointer)
224  *	2:	LR value
225  *	3:	Caller's caller
226  *	4:	...
227  *
228  * The value in LR is only needed when it holds a return address. If the
229  * return address is on the stack, we should ignore the LR value.
230  *
231  * Further, when the return address is in the LR, if a new frame was just
232  * allocated but the LR was not saved into it, then the LR contains the
233  * caller, slot 4: contains the caller's caller and the contents of slot 3:
234  * (chain->ips[3]) is undefined and must be ignored.
235  *
236  * Use DWARF debug information to determine if any entries need to be skipped.
237  *
238  * Return:
239  *	index:	of callchain entry that needs to be ignored (if any)
240  *	-1	if no entry needs to be ignored or in case of errors
241  */
arch_skip_callchain_idx(struct thread * thread,struct ip_callchain * chain)242 int arch_skip_callchain_idx(struct thread *thread, struct ip_callchain *chain)
243 {
244 	struct addr_location al;
245 	struct dso *dso = NULL;
246 	int rc;
247 	u64 ip;
248 	u64 skip_slot = -1;
249 
250 	if (!chain || chain->nr < 3)
251 		return skip_slot;
252 
253 	addr_location__init(&al);
254 	ip = chain->ips[1];
255 
256 	thread__find_symbol(thread, PERF_RECORD_MISC_USER, ip, &al);
257 
258 	if (al.map)
259 		dso = map__dso(al.map);
260 
261 	if (!dso) {
262 		pr_debug("%" PRIx64 " dso is NULL\n", ip);
263 		addr_location__exit(&al);
264 		return skip_slot;
265 	}
266 
267 	rc = check_return_addr(dso, map__start(al.map), ip);
268 
269 	pr_debug("[DSO %s, sym %s, ip 0x%" PRIx64 "] rc %d\n",
270 				dso->long_name, al.sym->name, ip, rc);
271 
272 	if (rc == 0) {
273 		/*
274 		 * Return address on stack. Ignore LR value in callchain
275 		 */
276 		skip_slot = 2;
277 	} else if (rc == 2) {
278 		/*
279 		 * New frame allocated but return address still in LR.
280 		 * Ignore the caller's caller entry in callchain.
281 		 */
282 		skip_slot = 3;
283 	}
284 
285 	addr_location__exit(&al);
286 	return skip_slot;
287 }
288