aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorSukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>2014-06-25 11:49:03 -0400
committerJiri Olsa <jolsa@kernel.org>2014-06-27 05:14:51 -0400
commita60335ba32981db5bc057b35782644e9e2436407 (patch)
treee6d63c2427f4c15a628cbbfd955d1b9b91c22704
parent21da83fb6c3ff67e969b8770bbb33138c9ede88e (diff)
perf tools powerpc: Adjust callchain based on DWARF debug info
When saving the callchain on Power, the kernel conservatively saves excess entries in the callchain. A few of these entries are needed in some cases but not others. We should use the DWARF debug information to determine when the entries are needed. Eg: the value in the link register (LR) is needed only when it holds the return address of a function. At other times it must be ignored. If the unnecessary entries are not ignored, we end up with duplicate arcs in the call-graphs. Use the DWARF debug information to determine if any callchain entries should be ignored when building call-graphs. Callgraph before the patch: 14.67% 2234 sprintft libc-2.18.so [.] __random | --- __random | |--61.12%-- __random | | | |--97.15%-- rand | | do_my_sprintf | | main | | generic_start_main.isra.0 | | __libc_start_main | | 0x0 | | | --2.85%-- do_my_sprintf | main | generic_start_main.isra.0 | __libc_start_main | 0x0 | --38.88%-- rand | |--94.01%-- rand | do_my_sprintf | main | generic_start_main.isra.0 | __libc_start_main | 0x0 | --5.99%-- do_my_sprintf main generic_start_main.isra.0 __libc_start_main 0x0 Callgraph after the patch: 14.67% 2234 sprintft libc-2.18.so [.] __random | --- __random | |--95.93%-- rand | do_my_sprintf | main | generic_start_main.isra.0 | __libc_start_main | 0x0 | --4.07%-- do_my_sprintf main generic_start_main.isra.0 __libc_start_main 0x0 TODO: For split-debug info objects like glibc, we can only determine the call-frame-address only when both .eh_frame and .debug_info sections are available. We should be able to determin the CFA even without the .eh_frame section. Fix suggested by Anton Blanchard. Thanks to valuable input on DWARF debug information from Ulrich Weigand. Reported-by: Maynard Johnson <maynard@us.ibm.com> Tested-by: Maynard Johnson <maynard@us.ibm.com> Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> Link: http://lkml.kernel.org/r/20140625154903.GA29607@us.ibm.com Signed-off-by: Jiri Olsa <jolsa@kernel.org>
-rw-r--r--tools/perf/arch/powerpc/Makefile1
-rw-r--r--tools/perf/arch/powerpc/util/skip-callchain-idx.c266
-rw-r--r--tools/perf/config/Makefile4
-rw-r--r--tools/perf/util/callchain.h13
-rw-r--r--tools/perf/util/machine.c18
5 files changed, 300 insertions, 2 deletions
diff --git a/tools/perf/arch/powerpc/Makefile b/tools/perf/arch/powerpc/Makefile
index 744e629797be..b92219b1900d 100644
--- a/tools/perf/arch/powerpc/Makefile
+++ b/tools/perf/arch/powerpc/Makefile
@@ -3,3 +3,4 @@ PERF_HAVE_DWARF_REGS := 1
3LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o 3LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
4endif 4endif
5LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/header.o 5LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/header.o
6LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/skip-callchain-idx.o
diff --git a/tools/perf/arch/powerpc/util/skip-callchain-idx.c b/tools/perf/arch/powerpc/util/skip-callchain-idx.c
new file mode 100644
index 000000000000..a7c23a4b3778
--- /dev/null
+++ b/tools/perf/arch/powerpc/util/skip-callchain-idx.c
@@ -0,0 +1,266 @@
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
19/*
20 * When saving the callchain on Power, the kernel conservatively saves
21 * excess entries in the callchain. A few of these entries are needed
22 * in some cases but not others. If the unnecessary entries are not
23 * ignored, we end up with duplicate arcs in the call-graphs. Use
24 * DWARF debug information to skip over any unnecessary callchain
25 * entries.
26 *
27 * See function header for arch_adjust_callchain() below for more details.
28 *
29 * The libdwfl code in this file is based on code from elfutils
30 * (libdwfl/argp-std.c, libdwfl/tests/addrcfi.c, etc).
31 */
32static char *debuginfo_path;
33
34static const Dwfl_Callbacks offline_callbacks = {
35 .debuginfo_path = &debuginfo_path,
36 .find_debuginfo = dwfl_standard_find_debuginfo,
37 .section_address = dwfl_offline_section_address,
38};
39
40
41/*
42 * Use the DWARF expression for the Call-frame-address and determine
43 * if return address is in LR and if a new frame was allocated.
44 */
45static int check_return_reg(int ra_regno, Dwarf_Frame *frame)
46{
47 Dwarf_Op ops_mem[2];
48 Dwarf_Op dummy;
49 Dwarf_Op *ops = &dummy;
50 size_t nops;
51 int result;
52
53 result = dwarf_frame_register(frame, ra_regno, ops_mem, &ops, &nops);
54 if (result < 0) {
55 pr_debug("dwarf_frame_register() %s\n", dwarf_errmsg(-1));
56 return -1;
57 }
58
59 /*
60 * Check if return address is on the stack.
61 */
62 if (nops != 0 || ops != NULL)
63 return 0;
64
65 /*
66 * Return address is in LR. Check if a frame was allocated
67 * but not-yet used.
68 */
69 result = dwarf_frame_cfa(frame, &ops, &nops);
70 if (result < 0) {
71 pr_debug("dwarf_frame_cfa() returns %d, %s\n", result,
72 dwarf_errmsg(-1));
73 return -1;
74 }
75
76 /*
77 * If call frame address is in r1, no new frame was allocated.
78 */
79 if (nops == 1 && ops[0].atom == DW_OP_bregx && ops[0].number == 1 &&
80 ops[0].number2 == 0)
81 return 1;
82
83 /*
84 * A new frame was allocated but has not yet been used.
85 */
86 return 2;
87}
88
89/*
90 * Get the DWARF frame from the .eh_frame section.
91 */
92static Dwarf_Frame *get_eh_frame(Dwfl_Module *mod, Dwarf_Addr pc)
93{
94 int result;
95 Dwarf_Addr bias;
96 Dwarf_CFI *cfi;
97 Dwarf_Frame *frame;
98
99 cfi = dwfl_module_eh_cfi(mod, &bias);
100 if (!cfi) {
101 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
102 return NULL;
103 }
104
105 result = dwarf_cfi_addrframe(cfi, pc, &frame);
106 if (result) {
107 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
108 return NULL;
109 }
110
111 return frame;
112}
113
114/*
115 * Get the DWARF frame from the .debug_frame section.
116 */
117static Dwarf_Frame *get_dwarf_frame(Dwfl_Module *mod, Dwarf_Addr pc)
118{
119 Dwarf_CFI *cfi;
120 Dwarf_Addr bias;
121 Dwarf_Frame *frame;
122 int result;
123
124 cfi = dwfl_module_dwarf_cfi(mod, &bias);
125 if (!cfi) {
126 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
127 return NULL;
128 }
129
130 result = dwarf_cfi_addrframe(cfi, pc, &frame);
131 if (result) {
132 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
133 return NULL;
134 }
135
136 return frame;
137}
138
139/*
140 * Return:
141 * 0 if return address for the program counter @pc is on stack
142 * 1 if return address is in LR and no new stack frame was allocated
143 * 2 if return address is in LR and a new frame was allocated (but not
144 * yet used)
145 * -1 in case of errors
146 */
147static int check_return_addr(const char *exec_file, Dwarf_Addr pc)
148{
149 int rc = -1;
150 Dwfl *dwfl;
151 Dwfl_Module *mod;
152 Dwarf_Frame *frame;
153 int ra_regno;
154 Dwarf_Addr start = pc;
155 Dwarf_Addr end = pc;
156 bool signalp;
157
158 dwfl = dwfl_begin(&offline_callbacks);
159 if (!dwfl) {
160 pr_debug("dwfl_begin() failed: %s\n", dwarf_errmsg(-1));
161 return -1;
162 }
163
164 if (dwfl_report_offline(dwfl, "", exec_file, -1) == NULL) {
165 pr_debug("dwfl_report_offline() failed %s\n", dwarf_errmsg(-1));
166 goto out;
167 }
168
169 mod = dwfl_addrmodule(dwfl, pc);
170 if (!mod) {
171 pr_debug("dwfl_addrmodule() failed, %s\n", dwarf_errmsg(-1));
172 goto out;
173 }
174
175 /*
176 * To work with split debug info files (eg: glibc), check both
177 * .eh_frame and .debug_frame sections of the ELF header.
178 */
179 frame = get_eh_frame(mod, pc);
180 if (!frame) {
181 frame = get_dwarf_frame(mod, pc);
182 if (!frame)
183 goto out;
184 }
185
186 ra_regno = dwarf_frame_info(frame, &start, &end, &signalp);
187 if (ra_regno < 0) {
188 pr_debug("Return address register unavailable: %s\n",
189 dwarf_errmsg(-1));
190 goto out;
191 }
192
193 rc = check_return_reg(ra_regno, frame);
194
195out:
196 dwfl_end(dwfl);
197 return rc;
198}
199
200/*
201 * The callchain saved by the kernel always includes the link register (LR).
202 *
203 * 0: PERF_CONTEXT_USER
204 * 1: Program counter (Next instruction pointer)
205 * 2: LR value
206 * 3: Caller's caller
207 * 4: ...
208 *
209 * The value in LR is only needed when it holds a return address. If the
210 * return address is on the stack, we should ignore the LR value.
211 *
212 * Further, when the return address is in the LR, if a new frame was just
213 * allocated but the LR was not saved into it, then the LR contains the
214 * caller, slot 4: contains the caller's caller and the contents of slot 3:
215 * (chain->ips[3]) is undefined and must be ignored.
216 *
217 * Use DWARF debug information to determine if any entries need to be skipped.
218 *
219 * Return:
220 * index: of callchain entry that needs to be ignored (if any)
221 * -1 if no entry needs to be ignored or in case of errors
222 */
223int arch_skip_callchain_idx(struct machine *machine, struct thread *thread,
224 struct ip_callchain *chain)
225{
226 struct addr_location al;
227 struct dso *dso = NULL;
228 int rc;
229 u64 ip;
230 u64 skip_slot = -1;
231
232 if (chain->nr < 3)
233 return skip_slot;
234
235 ip = chain->ips[2];
236
237 thread__find_addr_location(thread, machine, PERF_RECORD_MISC_USER,
238 MAP__FUNCTION, ip, &al);
239
240 if (al.map)
241 dso = al.map->dso;
242
243 if (!dso) {
244 pr_debug("%" PRIx64 " dso is NULL\n", ip);
245 return skip_slot;
246 }
247
248 rc = check_return_addr(dso->long_name, ip);
249
250 pr_debug("DSO %s, nr %" PRIx64 ", ip 0x%" PRIx64 "rc %d\n",
251 dso->long_name, chain->nr, ip, rc);
252
253 if (rc == 0) {
254 /*
255 * Return address on stack. Ignore LR value in callchain
256 */
257 skip_slot = 2;
258 } else if (rc == 2) {
259 /*
260 * New frame allocated but return address still in LR.
261 * Ignore the caller's caller entry in callchain.
262 */
263 skip_slot = 3;
264 }
265 return skip_slot;
266}
diff --git a/tools/perf/config/Makefile b/tools/perf/config/Makefile
index f30ac5e5d271..346bdb617544 100644
--- a/tools/perf/config/Makefile
+++ b/tools/perf/config/Makefile
@@ -48,6 +48,10 @@ ifneq ($(ARCH),$(filter $(ARCH),x86 arm))
48 NO_LIBDW_DWARF_UNWIND := 1 48 NO_LIBDW_DWARF_UNWIND := 1
49endif 49endif
50 50
51ifeq ($(ARCH),powerpc)
52 CFLAGS += -DHAVE_SKIP_CALLCHAIN_IDX
53endif
54
51ifeq ($(LIBUNWIND_LIBS),) 55ifeq ($(LIBUNWIND_LIBS),)
52 NO_LIBUNWIND := 1 56 NO_LIBUNWIND := 1
53else 57else
diff --git a/tools/perf/util/callchain.h b/tools/perf/util/callchain.h
index 8f84423a75da..da43619d6173 100644
--- a/tools/perf/util/callchain.h
+++ b/tools/perf/util/callchain.h
@@ -176,4 +176,17 @@ static inline void callchain_cursor_snapshot(struct callchain_cursor *dest,
176 dest->first = src->curr; 176 dest->first = src->curr;
177 dest->nr -= src->pos; 177 dest->nr -= src->pos;
178} 178}
179
180#ifdef HAVE_SKIP_CALLCHAIN_IDX
181extern int arch_skip_callchain_idx(struct machine *machine,
182 struct thread *thread, struct ip_callchain *chain);
183#else
184static inline int arch_skip_callchain_idx(struct machine *machine __maybe_unused,
185 struct thread *thread __maybe_unused,
186 struct ip_callchain *chain __maybe_unused)
187{
188 return -1;
189}
190#endif
191
179#endif /* __PERF_CALLCHAIN_H */ 192#endif /* __PERF_CALLCHAIN_H */
diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
index c73e1fc12e53..e9b943acaa5e 100644
--- a/tools/perf/util/machine.c
+++ b/tools/perf/util/machine.c
@@ -1281,7 +1281,9 @@ static int machine__resolve_callchain_sample(struct machine *machine,
1281 u8 cpumode = PERF_RECORD_MISC_USER; 1281 u8 cpumode = PERF_RECORD_MISC_USER;
1282 int chain_nr = min(max_stack, (int)chain->nr); 1282 int chain_nr = min(max_stack, (int)chain->nr);
1283 int i; 1283 int i;
1284 int j;
1284 int err; 1285 int err;
1286 int skip_idx __maybe_unused;
1285 1287
1286 callchain_cursor_reset(&callchain_cursor); 1288 callchain_cursor_reset(&callchain_cursor);
1287 1289
@@ -1290,14 +1292,26 @@ static int machine__resolve_callchain_sample(struct machine *machine,
1290 return 0; 1292 return 0;
1291 } 1293 }
1292 1294
1295 /*
1296 * Based on DWARF debug information, some architectures skip
1297 * a callchain entry saved by the kernel.
1298 */
1299 skip_idx = arch_skip_callchain_idx(machine, thread, chain);
1300
1293 for (i = 0; i < chain_nr; i++) { 1301 for (i = 0; i < chain_nr; i++) {
1294 u64 ip; 1302 u64 ip;
1295 struct addr_location al; 1303 struct addr_location al;
1296 1304
1297 if (callchain_param.order == ORDER_CALLEE) 1305 if (callchain_param.order == ORDER_CALLEE)
1298 ip = chain->ips[i]; 1306 j = i;
1299 else 1307 else
1300 ip = chain->ips[chain->nr - i - 1]; 1308 j = chain->nr - i - 1;
1309
1310#ifdef HAVE_SKIP_CALLCHAIN_IDX
1311 if (j == skip_idx)
1312 continue;
1313#endif
1314 ip = chain->ips[j];
1301 1315
1302 if (ip >= PERF_CONTEXT_MAX) { 1316 if (ip >= PERF_CONTEXT_MAX) {
1303 switch (ip) { 1317 switch (ip) {