2 * Use DWARF Debug information to skip unnecessary callchain entries.
4 * Copyright (C) 2014 Sukadev Bhattiprolu, IBM Corporation.
5 * Copyright (C) 2014 Ulrich Weigand, IBM Corporation.
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.
14 #include <elfutils/libdwfl.h>
16 #include "util/thread.h"
17 #include "util/callchain.h"
18 #include "util/debug.h"
21 #include "util/symbol.h"
24 * When saving the callchain on Power, the kernel conservatively saves
25 * excess entries in the callchain. A few of these entries are needed
26 * in some cases but not others. If the unnecessary entries are not
27 * ignored, we end up with duplicate arcs in the call-graphs. Use
28 * DWARF debug information to skip over any unnecessary callchain
31 * See function header for arch_adjust_callchain() below for more details.
33 * The libdwfl code in this file is based on code from elfutils
34 * (libdwfl/argp-std.c, libdwfl/tests/addrcfi.c, etc).
36 static char *debuginfo_path;
38 static const Dwfl_Callbacks offline_callbacks = {
39 .debuginfo_path = &debuginfo_path,
40 .find_debuginfo = dwfl_standard_find_debuginfo,
41 .section_address = dwfl_offline_section_address,
46 * Use the DWARF expression for the Call-frame-address and determine
47 * if return address is in LR and if a new frame was allocated.
49 static int check_return_reg(int ra_regno, Dwarf_Frame *frame)
53 Dwarf_Op *ops = &dummy;
57 result = dwarf_frame_register(frame, ra_regno, ops_mem, &ops, &nops);
59 pr_debug("dwarf_frame_register() %s\n", dwarf_errmsg(-1));
64 * Check if return address is on the stack. If return address
65 * is in a register (typically R0), it is yet to be saved on
68 if ((nops != 0 || ops != NULL) &&
69 !(nops == 1 && ops[0].atom == DW_OP_regx &&
70 ops[0].number2 == 0 && ops[0].offset == 0))
74 * Return address is in LR. Check if a frame was allocated
77 result = dwarf_frame_cfa(frame, &ops, &nops);
79 pr_debug("dwarf_frame_cfa() returns %d, %s\n", result,
85 * If call frame address is in r1, no new frame was allocated.
87 if (nops == 1 && ops[0].atom == DW_OP_bregx && ops[0].number == 1 &&
92 * A new frame was allocated but has not yet been used.
98 * Get the DWARF frame from the .eh_frame section.
100 static Dwarf_Frame *get_eh_frame(Dwfl_Module *mod, Dwarf_Addr pc)
107 cfi = dwfl_module_eh_cfi(mod, &bias);
109 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
113 result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
115 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
123 * Get the DWARF frame from the .debug_frame section.
125 static Dwarf_Frame *get_dwarf_frame(Dwfl_Module *mod, Dwarf_Addr pc)
132 cfi = dwfl_module_dwarf_cfi(mod, &bias);
134 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1));
138 result = dwarf_cfi_addrframe(cfi, pc-bias, &frame);
140 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1));
149 * 0 if return address for the program counter @pc is on stack
150 * 1 if return address is in LR and no new stack frame was allocated
151 * 2 if return address is in LR and a new frame was allocated (but not
153 * -1 in case of errors
155 static int check_return_addr(struct dso *dso, u64 map_start, Dwarf_Addr pc)
162 Dwarf_Addr start = pc;
165 const char *exec_file = dso->long_name;
170 dwfl = dwfl_begin(&offline_callbacks);
172 pr_debug("dwfl_begin() failed: %s\n", dwarf_errmsg(-1));
176 mod = dwfl_report_elf(dwfl, exec_file, exec_file, -1,
179 pr_debug("dwfl_report_elf() failed %s\n",
182 * We normally cache the DWARF debug info and never
183 * call dwfl_end(). But to prevent fd leak, free in
192 mod = dwfl_addrmodule(dwfl, pc);
194 pr_debug("dwfl_addrmodule() failed, %s\n", dwarf_errmsg(-1));
199 * To work with split debug info files (eg: glibc), check both
200 * .eh_frame and .debug_frame sections of the ELF header.
202 frame = get_eh_frame(mod, pc);
204 frame = get_dwarf_frame(mod, pc);
209 ra_regno = dwarf_frame_info(frame, &start, &end, &signalp);
211 pr_debug("Return address register unavailable: %s\n",
216 rc = check_return_reg(ra_regno, frame);
223 * The callchain saved by the kernel always includes the link register (LR).
225 * 0: PERF_CONTEXT_USER
226 * 1: Program counter (Next instruction pointer)
231 * The value in LR is only needed when it holds a return address. If the
232 * return address is on the stack, we should ignore the LR value.
234 * Further, when the return address is in the LR, if a new frame was just
235 * allocated but the LR was not saved into it, then the LR contains the
236 * caller, slot 4: contains the caller's caller and the contents of slot 3:
237 * (chain->ips[3]) is undefined and must be ignored.
239 * Use DWARF debug information to determine if any entries need to be skipped.
242 * index: of callchain entry that needs to be ignored (if any)
243 * -1 if no entry needs to be ignored or in case of errors
245 int arch_skip_callchain_idx(struct thread *thread, struct ip_callchain *chain)
247 struct addr_location al;
248 struct dso *dso = NULL;
253 if (!chain || chain->nr < 3)
258 thread__find_symbol(thread, PERF_RECORD_MISC_USER, ip, &al);
264 pr_debug("%" PRIx64 " dso is NULL\n", ip);
268 rc = check_return_addr(dso, al.map->start, ip);
270 pr_debug("[DSO %s, sym %s, ip 0x%" PRIx64 "] rc %d\n",
271 dso->long_name, al.sym->name, ip, rc);
275 * Return address on stack. Ignore LR value in callchain
278 } else if (rc == 2) {
280 * New frame allocated but return address still in LR.
281 * Ignore the caller's caller entry in callchain.