1 /*
2 * Copyright (C) 2019 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "src/profiling/perf/regs_parsing.h"
18
19 #include <linux/perf_event.h>
20 #include <stdint.h>
21 #include <unistd.h>
22
23 #include <cinttypes>
24 #include <memory>
25
26 #include <unwindstack/Elf.h>
27 #include <unwindstack/MachineArm.h>
28 #include <unwindstack/MachineArm64.h>
29 #include <unwindstack/Regs.h>
30 #include <unwindstack/RegsArm.h>
31 #include <unwindstack/RegsArm64.h>
32 #include <unwindstack/RegsX86.h>
33 #include <unwindstack/RegsX86_64.h>
34 #include <unwindstack/UserArm.h>
35 #include <unwindstack/UserArm64.h>
36 #include <unwindstack/UserX86.h>
37 #include <unwindstack/UserX86_64.h>
38
39 // kernel uapi headers
40 #include <uapi/asm-arm/asm/perf_regs.h>
41 #include <uapi/asm-x86/asm/perf_regs.h>
42 #define perf_event_arm_regs perf_event_arm64_regs
43 #include <uapi/asm-arm64/asm/perf_regs.h>
44 #undef perf_event_arm_regs
45
46 namespace perfetto {
47 namespace profiling {
48
49 namespace {
50
constexpr_max(size_t x,size_t y)51 constexpr size_t constexpr_max(size_t x, size_t y) {
52 return x > y ? x : y;
53 }
54
55 template <typename T>
ReadValue(T * value_out,const char * ptr)56 const char* ReadValue(T* value_out, const char* ptr) {
57 memcpy(value_out, reinterpret_cast<const void*>(ptr), sizeof(T));
58 return ptr + sizeof(T);
59 }
60
61 // Supported configurations:
62 // * 32 bit daemon, 32 bit userspace
63 // * 64 bit daemon, mixed bitness userspace
64 // Therefore give the kernel the mask corresponding to our build architecture.
65 // Register parsing handles the mixed userspace ABI cases.
66 // For simplicity, we ask for as many registers as we can, even if not all of
67 // them will be used during unwinding.
68 // TODO(rsavitski): cleanly detect 32 bit builds being side-loaded onto a system
69 // with 64 bit userspace processes.
PerfUserRegsMask(unwindstack::ArchEnum arch)70 uint64_t PerfUserRegsMask(unwindstack::ArchEnum arch) {
71 switch (static_cast<uint8_t>(arch)) { // cast to please -Wswitch-enum
72 case unwindstack::ARCH_ARM64:
73 return (1ULL << PERF_REG_ARM64_MAX) - 1;
74 case unwindstack::ARCH_ARM:
75 return ((1ULL << PERF_REG_ARM_MAX) - 1);
76 // perf on x86_64 doesn't allow sampling ds/es/fs/gs registers. See
77 // arch/x86/kernel/perf_regs.c in the kernel.
78 case unwindstack::ARCH_X86_64:
79 return (((1ULL << PERF_REG_X86_64_MAX) - 1) & ~(1ULL << PERF_REG_X86_DS) &
80 ~(1ULL << PERF_REG_X86_ES) & ~(1ULL << PERF_REG_X86_FS) &
81 ~(1ULL << PERF_REG_X86_GS));
82 // Note: excluding these segment registers might not be necessary on x86,
83 // but they won't be used anyway (so follow x64).
84 case unwindstack::ARCH_X86:
85 return ((1ULL << PERF_REG_X86_32_MAX) - 1) & ~(1ULL << PERF_REG_X86_DS) &
86 ~(1ULL << PERF_REG_X86_ES) & ~(1ULL << PERF_REG_X86_FS) &
87 ~(1ULL << PERF_REG_X86_GS);
88 default:
89 PERFETTO_FATAL("Unsupported architecture");
90 }
91 }
92
93 // Adjusts the given architecture enum based on the ABI (as recorded in the perf
94 // sample). Note: we do not support 64 bit samples on a 32 bit daemon build, so
95 // this only converts from 64 bit to 32 bit architectures.
ArchForAbi(unwindstack::ArchEnum arch,uint64_t abi)96 unwindstack::ArchEnum ArchForAbi(unwindstack::ArchEnum arch, uint64_t abi) {
97 if (arch == unwindstack::ARCH_ARM64 && abi == PERF_SAMPLE_REGS_ABI_32) {
98 return unwindstack::ARCH_ARM;
99 }
100 if (arch == unwindstack::ARCH_X86_64 && abi == PERF_SAMPLE_REGS_ABI_32) {
101 return unwindstack::ARCH_X86;
102 }
103 return arch;
104 }
105
106 // Register values as an array, indexed using the kernel uapi perf_events.h enum
107 // values. Unsampled values will be left as zeroes.
108 struct RawRegisterData {
109 static constexpr uint64_t kMaxSize =
110 constexpr_max(PERF_REG_ARM64_MAX,
111 constexpr_max(PERF_REG_ARM_MAX, PERF_REG_X86_64_MAX));
112 uint64_t regs[kMaxSize] = {};
113 };
114
115 // First converts the |RawRegisterData| array to libunwindstack's "user"
116 // register structs (which match the ptrace/coredump format, also available at
117 // <sys/user.h>), then constructs the relevant unwindstack::Regs subclass out
118 // of the latter.
ToLibUnwindstackRegs(const RawRegisterData & raw_regs,unwindstack::ArchEnum arch)119 std::unique_ptr<unwindstack::Regs> ToLibUnwindstackRegs(
120 const RawRegisterData& raw_regs,
121 unwindstack::ArchEnum arch) {
122 if (arch == unwindstack::ARCH_ARM64) {
123 static_assert(static_cast<int>(unwindstack::ARM64_REG_R0) ==
124 static_cast<int>(PERF_REG_ARM64_X0) &&
125 static_cast<int>(unwindstack::ARM64_REG_R0) == 0,
126 "register layout mismatch");
127 static_assert(static_cast<int>(unwindstack::ARM64_REG_R30) ==
128 static_cast<int>(PERF_REG_ARM64_LR),
129 "register layout mismatch");
130 // Both the perf_event register order and the "user" format are derived from
131 // "struct pt_regs", so we can directly memcpy the first 31 regs (up to and
132 // including LR).
133 unwindstack::arm64_user_regs arm64_user_regs = {};
134 memcpy(&arm64_user_regs.regs[0], &raw_regs.regs[0],
135 sizeof(uint64_t) * (PERF_REG_ARM64_LR + 1));
136 arm64_user_regs.sp = raw_regs.regs[PERF_REG_ARM64_SP];
137 arm64_user_regs.pc = raw_regs.regs[PERF_REG_ARM64_PC];
138 return std::unique_ptr<unwindstack::Regs>(
139 unwindstack::RegsArm64::Read(&arm64_user_regs));
140 }
141
142 if (arch == unwindstack::ARCH_ARM) {
143 static_assert(static_cast<int>(unwindstack::ARM_REG_R0) ==
144 static_cast<int>(PERF_REG_ARM_R0) &&
145 static_cast<int>(unwindstack::ARM_REG_R0) == 0,
146 "register layout mismatch");
147 static_assert(static_cast<int>(unwindstack::ARM_REG_LAST) ==
148 static_cast<int>(PERF_REG_ARM_MAX),
149 "register layout mismatch");
150 // As with arm64, the layouts match, but we need to downcast to u32.
151 unwindstack::arm_user_regs arm_user_regs = {};
152 for (size_t i = 0; i < unwindstack::ARM_REG_LAST; i++) {
153 arm_user_regs.regs[i] = static_cast<uint32_t>(raw_regs.regs[i]);
154 }
155 return std::unique_ptr<unwindstack::Regs>(
156 unwindstack::RegsArm::Read(&arm_user_regs));
157 }
158
159 if (arch == unwindstack::ARCH_X86_64) {
160 // We've sampled more registers than what libunwindstack will use. Don't
161 // copy over cs/ss/flags.
162 unwindstack::x86_64_user_regs x86_64_user_regs = {};
163 x86_64_user_regs.rax = raw_regs.regs[PERF_REG_X86_AX];
164 x86_64_user_regs.rbx = raw_regs.regs[PERF_REG_X86_BX];
165 x86_64_user_regs.rcx = raw_regs.regs[PERF_REG_X86_CX];
166 x86_64_user_regs.rdx = raw_regs.regs[PERF_REG_X86_DX];
167 x86_64_user_regs.r8 = raw_regs.regs[PERF_REG_X86_R8];
168 x86_64_user_regs.r9 = raw_regs.regs[PERF_REG_X86_R9];
169 x86_64_user_regs.r10 = raw_regs.regs[PERF_REG_X86_R10];
170 x86_64_user_regs.r11 = raw_regs.regs[PERF_REG_X86_R11];
171 x86_64_user_regs.r12 = raw_regs.regs[PERF_REG_X86_R12];
172 x86_64_user_regs.r13 = raw_regs.regs[PERF_REG_X86_R13];
173 x86_64_user_regs.r14 = raw_regs.regs[PERF_REG_X86_R14];
174 x86_64_user_regs.r15 = raw_regs.regs[PERF_REG_X86_R15];
175 x86_64_user_regs.rdi = raw_regs.regs[PERF_REG_X86_DI];
176 x86_64_user_regs.rsi = raw_regs.regs[PERF_REG_X86_SI];
177 x86_64_user_regs.rbp = raw_regs.regs[PERF_REG_X86_BP];
178 x86_64_user_regs.rsp = raw_regs.regs[PERF_REG_X86_SP];
179 x86_64_user_regs.rip = raw_regs.regs[PERF_REG_X86_IP];
180 return std::unique_ptr<unwindstack::Regs>(
181 unwindstack::RegsX86_64::Read(&x86_64_user_regs));
182 }
183
184 if (arch == unwindstack::ARCH_X86) {
185 // We've sampled more registers than what libunwindstack will use. Don't
186 // copy over cs/ss/flags.
187 unwindstack::x86_user_regs x86_user_regs = {};
188 x86_user_regs.eax = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_AX]);
189 x86_user_regs.ebx = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_BX]);
190 x86_user_regs.ecx = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_CX]);
191 x86_user_regs.edx = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_DX]);
192 x86_user_regs.ebp = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_BP]);
193 x86_user_regs.edi = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_DI]);
194 x86_user_regs.esi = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_SI]);
195 x86_user_regs.esp = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_SP]);
196 x86_user_regs.eip = static_cast<uint32_t>(raw_regs.regs[PERF_REG_X86_IP]);
197 return std::unique_ptr<unwindstack::Regs>(
198 unwindstack::RegsX86::Read(&x86_user_regs));
199 }
200
201 PERFETTO_FATAL("Unsupported architecture");
202 }
203
204 } // namespace
205
PerfUserRegsMaskForArch(unwindstack::ArchEnum arch)206 uint64_t PerfUserRegsMaskForArch(unwindstack::ArchEnum arch) {
207 return PerfUserRegsMask(arch);
208 }
209
210 // Assumes that the sampling was configured with
211 // |PerfUserRegsMaskForArch(unwindstack::Regs::CurrentArch())|.
ReadPerfUserRegsData(const char ** data)212 std::unique_ptr<unwindstack::Regs> ReadPerfUserRegsData(const char** data) {
213 unwindstack::ArchEnum requested_arch = unwindstack::Regs::CurrentArch();
214
215 // Layout, assuming a sparse bitmask requesting r1 and r15:
216 // userspace thread: [u64 abi] [u64 r1] [u64 r15]
217 // kernel thread: [u64 abi]
218 const char* parse_pos = *data;
219 uint64_t sampled_abi;
220 parse_pos = ReadValue(&sampled_abi, parse_pos);
221
222 // ABI_NONE means there were no registers, as we've sampled a kernel thread,
223 // which doesn't have userspace registers.
224 if (sampled_abi == PERF_SAMPLE_REGS_ABI_NONE) {
225 *data = parse_pos; // adjust caller's parsing position
226 return nullptr;
227 }
228
229 // Unpack the densely-packed register values into |RawRegisterData|, which has
230 // a value for every register (unsampled registers will be left at zero).
231 RawRegisterData raw_regs{};
232 uint64_t regs_mask = PerfUserRegsMaskForArch(requested_arch);
233 for (size_t i = 0; regs_mask && (i < RawRegisterData::kMaxSize); i++) {
234 if (regs_mask & (1ULL << i)) {
235 parse_pos = ReadValue(&raw_regs.regs[i], parse_pos);
236 }
237 }
238
239 // Special case: we've requested arm64 registers from a 64 bit kernel, but
240 // ended up sampling a 32 bit arm userspace process. The 32 bit execution
241 // state of the target process was saved by the exception entry in an
242 // ISA-specific way. The userspace R0-R14 end up saved as arm64 W0-W14, but
243 // the program counter (R15 on arm32) is still in PERF_REG_ARM64_PC (the 33rd
244 // register). So we can take the kernel-dumped 64 bit register state, reassign
245 // the PC into the R15 slot, and treat the resulting RawRegisterData as an
246 // arm32 register bank. See "Fundamentals of ARMv8-A" (ARM DOC
247 // 100878_0100_en), page 28.
248 // x86-64 doesn't need any such fixups.
249 if (requested_arch == unwindstack::ARCH_ARM64 &&
250 sampled_abi == PERF_SAMPLE_REGS_ABI_32) {
251 raw_regs.regs[PERF_REG_ARM_PC] = raw_regs.regs[PERF_REG_ARM64_PC];
252 }
253
254 *data = parse_pos; // adjust caller's parsing position
255
256 unwindstack::ArchEnum sampled_arch = ArchForAbi(requested_arch, sampled_abi);
257 return ToLibUnwindstackRegs(raw_regs, sampled_arch);
258 }
259
260 } // namespace profiling
261 } // namespace perfetto
262