1/*
2 * Use DWARF Debug information to skip unnecessary callchain entries.
3 *
4 * Copyright (C) 2014 Sukadev Bhattiprolu, IBM Corporation.
5 * Copyright (C) 2014 Ulrich Weigand, IBM Corporation.
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; either version
10 * 2 of the License, or (at your option) any later version.
11 */
12#include <inttypes.h>
13#include <dwarf.h>
14#include <elfutils/libdwfl.h>
15
16#include "util/thread.h"
17#include "util/callchain.h"
18#include "util/debug.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 */
33static char *debuginfo_path;
34
35static 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 */
46static int check_return_reg(int ra_regno, Dwarf_Frame *frame)
47{
48	Dwarf_Op ops_mem[2];
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.
62	 */
63	if (nops != 0 || ops != NULL)
64		return 0;
65
66	/*
67	 * Return address is in LR. Check if a frame was allocated
68	 * but not-yet used.
69	 */
70	result = dwarf_frame_cfa(frame, &ops, &nops);
71	if (result < 0) {
72		pr_debug("dwarf_frame_cfa() returns %d, %s\n", result,
73					dwarf_errmsg(-1));
74		return -1;
75	}
76
77	/*
78	 * If call frame address is in r1, no new frame was allocated.
79	 */
80	if (nops == 1 && ops[0].atom == DW_OP_bregx && ops[0].number == 1 &&
81				ops[0].number2 == 0)
82		return 1;
83
84	/*
85	 * A new frame was allocated but has not yet been used.
86	 */
87	return 2;
88}
89
90/*
91 * Get the DWARF frame from the .eh_frame section.
92 */
93static Dwarf_Frame *get_eh_frame(Dwfl_Module *mod, Dwarf_Addr pc)
94{
95	int		result;
96	Dwarf_Addr	bias;
97	Dwarf_CFI	*cfi;
98	Dwarf_Frame	*frame;
99
100	cfi = dwfl_module_eh_cfi(mod, &bias);
101	if (!cfi) {
102		pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
103		return NULL;
104	}
105
106	result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
107	if (result) {
108		pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
109		return NULL;
110	}
111
112	return frame;
113}
114
115/*
116 * Get the DWARF frame from the .debug_frame section.
117 */
118static Dwarf_Frame *get_dwarf_frame(Dwfl_Module *mod, Dwarf_Addr pc)
119{
120	Dwarf_CFI       *cfi;
121	Dwarf_Addr      bias;
122	Dwarf_Frame     *frame;
123	int             result;
124
125	cfi = dwfl_module_dwarf_cfi(mod, &bias);
126	if (!cfi) {
127		pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
128		return NULL;
129	}
130
131	result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
132	if (result) {
133		pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
134		return NULL;
135	}
136
137	return frame;
138}
139
140/*
141 * Return:
142 *	0 if return address for the program counter @pc is on stack
143 *	1 if return address is in LR and no new stack frame was allocated
144 *	2 if return address is in LR and a new frame was allocated (but not
145 *		yet used)
146 *	-1 in case of errors
147 */
148static int check_return_addr(struct dso *dso, u64 map_start, Dwarf_Addr pc)
149{
150	int		rc = -1;
151	Dwfl		*dwfl;
152	Dwfl_Module	*mod;
153	Dwarf_Frame	*frame;
154	int		ra_regno;
155	Dwarf_Addr	start = pc;
156	Dwarf_Addr	end = pc;
157	bool		signalp;
158	const char	*exec_file = dso->long_name;
159
160	dwfl = dso->dwfl;
161
162	if (!dwfl) {
163		dwfl = dwfl_begin(&offline_callbacks);
164		if (!dwfl) {
165			pr_debug("dwfl_begin() failed: %s\n", dwarf_errmsg(-1));
166			return -1;
167		}
168
169		mod = dwfl_report_elf(dwfl, exec_file, exec_file, -1,
170						map_start, false);
171		if (!mod) {
172			pr_debug("dwfl_report_elf() failed %s\n",
173						dwarf_errmsg(-1));
174			/*
175			 * We normally cache the DWARF debug info and never
176			 * call dwfl_end(). But to prevent fd leak, free in
177			 * case of error.
178			 */
179			dwfl_end(dwfl);
180			goto out;
181		}
182		dso->dwfl = dwfl;
183	}
184
185	mod = dwfl_addrmodule(dwfl, pc);
186	if (!mod) {
187		pr_debug("dwfl_addrmodule() failed, %s\n", dwarf_errmsg(-1));
188		goto out;
189	}
190
191	/*
192	 * To work with split debug info files (eg: glibc), check both
193	 * .eh_frame and .debug_frame sections of the ELF header.
194	 */
195	frame = get_eh_frame(mod, pc);
196	if (!frame) {
197		frame = get_dwarf_frame(mod, pc);
198		if (!frame)
199			goto out;
200	}
201
202	ra_regno = dwarf_frame_info(frame, &start, &end, &signalp);
203	if (ra_regno < 0) {
204		pr_debug("Return address register unavailable: %s\n",
205				dwarf_errmsg(-1));
206		goto out;
207	}
208
209	rc = check_return_reg(ra_regno, frame);
210
211out:
212	return rc;
213}
214
215/*
216 * The callchain saved by the kernel always includes the link register (LR).
217 *
218 *	0:	PERF_CONTEXT_USER
219 *	1:	Program counter (Next instruction pointer)
220 *	2:	LR value
221 *	3:	Caller's caller
222 *	4:	...
223 *
224 * The value in LR is only needed when it holds a return address. If the
225 * return address is on the stack, we should ignore the LR value.
226 *
227 * Further, when the return address is in the LR, if a new frame was just
228 * allocated but the LR was not saved into it, then the LR contains the
229 * caller, slot 4: contains the caller's caller and the contents of slot 3:
230 * (chain->ips[3]) is undefined and must be ignored.
231 *
232 * Use DWARF debug information to determine if any entries need to be skipped.
233 *
234 * Return:
235 *	index:	of callchain entry that needs to be ignored (if any)
236 *	-1	if no entry needs to be ignored or in case of errors
237 */
238int arch_skip_callchain_idx(struct thread *thread, struct ip_callchain *chain)
239{
240	struct addr_location al;
241	struct dso *dso = NULL;
242	int rc;
243	u64 ip;
244	u64 skip_slot = -1;
245
246	if (chain->nr < 3)
247		return skip_slot;
248
249	ip = chain->ips[2];
250
251	thread__find_addr_location(thread, PERF_RECORD_MISC_USER,
252			MAP__FUNCTION, ip, &al);
253
254	if (al.map)
255		dso = al.map->dso;
256
257	if (!dso) {
258		pr_debug("%" PRIx64 " dso is NULL\n", ip);
259		return skip_slot;
260	}
261
262	rc = check_return_addr(dso, al.map->start, ip);
263
264	pr_debug("[DSO %s, sym %s, ip 0x%" PRIx64 "] rc %d\n",
265				dso->long_name, al.sym->name, ip, rc);
266
267	if (rc == 0) {
268		/*
269		 * Return address on stack. Ignore LR value in callchain
270		 */
271		skip_slot = 2;
272	} else if (rc == 2) {
273		/*
274		 * New frame allocated but return address still in LR.
275		 * Ignore the caller's caller entry in callchain.
276		 */
277		skip_slot = 3;
278	}
279	return skip_slot;
280}
281