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/map.h" 17 #include "util/symbol.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 */ 32 static char *debuginfo_path; 33 34 static 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 */ 45 static 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. If return address 61 * is in a register (typically R0), it is yet to be saved on 62 * the stack. 63 */ 64 if ((nops != 0 || ops != NULL) && 65 !(nops == 1 && ops[0].atom == DW_OP_regx && 66 ops[0].number2 == 0 && ops[0].offset == 0)) 67 return 0; 68 69 /* 70 * Return address is in LR. Check if a frame was allocated 71 * but not-yet used. 72 */ 73 result = dwarf_frame_cfa(frame, &ops, &nops); 74 if (result < 0) { 75 pr_debug("dwarf_frame_cfa() returns %d, %s\n", result, 76 dwarf_errmsg(-1)); 77 return -1; 78 } 79 80 /* 81 * If call frame address is in r1, no new frame was allocated. 82 */ 83 if (nops == 1 && ops[0].atom == DW_OP_bregx && ops[0].number == 1 && 84 ops[0].number2 == 0) 85 return 1; 86 87 /* 88 * A new frame was allocated but has not yet been used. 89 */ 90 return 2; 91 } 92 93 /* 94 * Get the DWARF frame from the .eh_frame section. 95 */ 96 static Dwarf_Frame *get_eh_frame(Dwfl_Module *mod, Dwarf_Addr pc) 97 { 98 int result; 99 Dwarf_Addr bias; 100 Dwarf_CFI *cfi; 101 Dwarf_Frame *frame; 102 103 cfi = dwfl_module_eh_cfi(mod, &bias); 104 if (!cfi) { 105 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1)); 106 return NULL; 107 } 108 109 result = dwarf_cfi_addrframe(cfi, pc-bias, &frame); 110 if (result) { 111 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1)); 112 return NULL; 113 } 114 115 return frame; 116 } 117 118 /* 119 * Get the DWARF frame from the .debug_frame section. 120 */ 121 static Dwarf_Frame *get_dwarf_frame(Dwfl_Module *mod, Dwarf_Addr pc) 122 { 123 Dwarf_CFI *cfi; 124 Dwarf_Addr bias; 125 Dwarf_Frame *frame; 126 int result; 127 128 cfi = dwfl_module_dwarf_cfi(mod, &bias); 129 if (!cfi) { 130 pr_debug("%s(): no CFI - %s\n", __func__, dwfl_errmsg(-1)); 131 return NULL; 132 } 133 134 result = dwarf_cfi_addrframe(cfi, pc-bias, &frame); 135 if (result) { 136 pr_debug("%s(): %s\n", __func__, dwfl_errmsg(-1)); 137 return NULL; 138 } 139 140 return frame; 141 } 142 143 /* 144 * Return: 145 * 0 if return address for the program counter @pc is on stack 146 * 1 if return address is in LR and no new stack frame was allocated 147 * 2 if return address is in LR and a new frame was allocated (but not 148 * yet used) 149 * -1 in case of errors 150 */ 151 static int check_return_addr(struct dso *dso, u64 map_start, Dwarf_Addr pc) 152 { 153 int rc = -1; 154 Dwfl *dwfl; 155 Dwfl_Module *mod; 156 Dwarf_Frame *frame; 157 int ra_regno; 158 Dwarf_Addr start = pc; 159 Dwarf_Addr end = pc; 160 bool signalp; 161 const char *exec_file = dso->long_name; 162 163 dwfl = dso->dwfl; 164 165 if (!dwfl) { 166 dwfl = dwfl_begin(&offline_callbacks); 167 if (!dwfl) { 168 pr_debug("dwfl_begin() failed: %s\n", dwarf_errmsg(-1)); 169 return -1; 170 } 171 172 mod = dwfl_report_elf(dwfl, exec_file, exec_file, -1, 173 map_start, false); 174 if (!mod) { 175 pr_debug("dwfl_report_elf() failed %s\n", 176 dwarf_errmsg(-1)); 177 /* 178 * We normally cache the DWARF debug info and never 179 * call dwfl_end(). But to prevent fd leak, free in 180 * case of error. 181 */ 182 dwfl_end(dwfl); 183 goto out; 184 } 185 dso->dwfl = dwfl; 186 } 187 188 mod = dwfl_addrmodule(dwfl, pc); 189 if (!mod) { 190 pr_debug("dwfl_addrmodule() failed, %s\n", dwarf_errmsg(-1)); 191 goto out; 192 } 193 194 /* 195 * To work with split debug info files (eg: glibc), check both 196 * .eh_frame and .debug_frame sections of the ELF header. 197 */ 198 frame = get_eh_frame(mod, pc); 199 if (!frame) { 200 frame = get_dwarf_frame(mod, pc); 201 if (!frame) 202 goto out; 203 } 204 205 ra_regno = dwarf_frame_info(frame, &start, &end, &signalp); 206 if (ra_regno < 0) { 207 pr_debug("Return address register unavailable: %s\n", 208 dwarf_errmsg(-1)); 209 goto out; 210 } 211 212 rc = check_return_reg(ra_regno, frame); 213 214 out: 215 return rc; 216 } 217 218 /* 219 * The callchain saved by the kernel always includes the link register (LR). 220 * 221 * 0: PERF_CONTEXT_USER 222 * 1: Program counter (Next instruction pointer) 223 * 2: LR value 224 * 3: Caller's caller 225 * 4: ... 226 * 227 * The value in LR is only needed when it holds a return address. If the 228 * return address is on the stack, we should ignore the LR value. 229 * 230 * Further, when the return address is in the LR, if a new frame was just 231 * allocated but the LR was not saved into it, then the LR contains the 232 * caller, slot 4: contains the caller's caller and the contents of slot 3: 233 * (chain->ips[3]) is undefined and must be ignored. 234 * 235 * Use DWARF debug information to determine if any entries need to be skipped. 236 * 237 * Return: 238 * index: of callchain entry that needs to be ignored (if any) 239 * -1 if no entry needs to be ignored or in case of errors 240 */ 241 int arch_skip_callchain_idx(struct thread *thread, struct ip_callchain *chain) 242 { 243 struct addr_location al; 244 struct dso *dso = NULL; 245 int rc; 246 u64 ip; 247 u64 skip_slot = -1; 248 249 if (!chain || chain->nr < 3) 250 return skip_slot; 251 252 ip = chain->ips[1]; 253 254 thread__find_symbol(thread, PERF_RECORD_MISC_USER, ip, &al); 255 256 if (al.map) 257 dso = al.map->dso; 258 259 if (!dso) { 260 pr_debug("%" PRIx64 " dso is NULL\n", ip); 261 return skip_slot; 262 } 263 264 rc = check_return_addr(dso, al.map->start, ip); 265 266 pr_debug("[DSO %s, sym %s, ip 0x%" PRIx64 "] rc %d\n", 267 dso->long_name, al.sym->name, ip, rc); 268 269 if (rc == 0) { 270 /* 271 * Return address on stack. Ignore LR value in callchain 272 */ 273 skip_slot = 2; 274 } else if (rc == 2) { 275 /* 276 * New frame allocated but return address still in LR. 277 * Ignore the caller's caller entry in callchain. 278 */ 279 skip_slot = 3; 280 } 281 return skip_slot; 282 } 283