Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 1 | /* |
Maciej Żenczykowski | 283c25a | 2023-10-02 19:43:30 -0700 | [diff] [blame] | 2 | * Copyright (C) 2018-2023 The Android Open Source Project |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 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 | |
Maciej Żenczykowski | 283c25a | 2023-10-02 19:43:30 -0700 | [diff] [blame] | 17 | #define LOG_TAG "NetBpfLoader" |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 18 | |
| 19 | #include <errno.h> |
| 20 | #include <fcntl.h> |
| 21 | #include <linux/bpf.h> |
| 22 | #include <linux/elf.h> |
| 23 | #include <log/log.h> |
| 24 | #include <stdint.h> |
| 25 | #include <stdio.h> |
| 26 | #include <stdlib.h> |
| 27 | #include <string.h> |
| 28 | #include <sysexits.h> |
| 29 | #include <sys/stat.h> |
| 30 | #include <sys/utsname.h> |
| 31 | #include <sys/wait.h> |
| 32 | #include <unistd.h> |
| 33 | |
Maciej Żenczykowski | 59fb77f | 2024-03-08 08:00:29 +0000 | [diff] [blame] | 34 | // This is Network BpfLoader v0.42 |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 35 | // WARNING: If you ever hit cherrypick conflicts here you're doing it wrong: |
| 36 | // You are NOT allowed to cherrypick bpfloader related patches out of order. |
| 37 | // (indeed: cherrypicking is probably a bad idea and you should merge instead) |
| 38 | // Mainline supports ONLY the published versions of the bpfloader for each Android release. |
| 39 | #define BPFLOADER_VERSION_MAJOR 0u |
Maciej Żenczykowski | 59fb77f | 2024-03-08 08:00:29 +0000 | [diff] [blame] | 40 | #define BPFLOADER_VERSION_MINOR 42u |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 41 | #define BPFLOADER_VERSION ((BPFLOADER_VERSION_MAJOR << 16) | BPFLOADER_VERSION_MINOR) |
| 42 | |
| 43 | #include "BpfSyscallWrappers.h" |
| 44 | #include "bpf/BpfUtils.h" |
| 45 | #include "bpf/bpf_map_def.h" |
Maciej Żenczykowski | 40dfe53 | 2023-10-08 20:21:11 -0700 | [diff] [blame] | 46 | #include "loader.h" |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 47 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 48 | #include <cstdlib> |
| 49 | #include <fstream> |
| 50 | #include <iostream> |
| 51 | #include <optional> |
| 52 | #include <string> |
| 53 | #include <unordered_map> |
| 54 | #include <vector> |
| 55 | |
| 56 | #include <android-base/cmsg.h> |
| 57 | #include <android-base/file.h> |
Elliott Hughes | 497e3e0 | 2023-10-09 22:18:08 +0000 | [diff] [blame] | 58 | #include <android-base/properties.h> |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 59 | #include <android-base/strings.h> |
| 60 | #include <android-base/unique_fd.h> |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 61 | |
| 62 | #define BPF_FS_PATH "/sys/fs/bpf/" |
| 63 | |
| 64 | // Size of the BPF log buffer for verifier logging |
| 65 | #define BPF_LOAD_LOG_SZ 0xfffff |
| 66 | |
| 67 | // Unspecified attach type is 0 which is BPF_CGROUP_INET_INGRESS. |
| 68 | #define BPF_ATTACH_TYPE_UNSPEC BPF_CGROUP_INET_INGRESS |
| 69 | |
| 70 | using android::base::StartsWith; |
| 71 | using android::base::unique_fd; |
| 72 | using std::ifstream; |
| 73 | using std::ios; |
| 74 | using std::optional; |
| 75 | using std::string; |
| 76 | using std::vector; |
| 77 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 78 | namespace android { |
| 79 | namespace bpf { |
| 80 | |
| 81 | const std::string& getBuildType() { |
Elliott Hughes | 497e3e0 | 2023-10-09 22:18:08 +0000 | [diff] [blame] | 82 | static std::string t = android::base::GetProperty("ro.build.type", "unknown"); |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 83 | return t; |
| 84 | } |
| 85 | |
| 86 | static unsigned int page_size = static_cast<unsigned int>(getpagesize()); |
| 87 | |
| 88 | constexpr const char* lookupSelinuxContext(const domain d, const char* const unspecified = "") { |
| 89 | switch (d) { |
| 90 | case domain::unspecified: return unspecified; |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 91 | case domain::tethering: return "fs_bpf_tethering"; |
| 92 | case domain::net_private: return "fs_bpf_net_private"; |
| 93 | case domain::net_shared: return "fs_bpf_net_shared"; |
| 94 | case domain::netd_readonly: return "fs_bpf_netd_readonly"; |
| 95 | case domain::netd_shared: return "fs_bpf_netd_shared"; |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 96 | default: return "(unrecognized)"; |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | domain getDomainFromSelinuxContext(const char s[BPF_SELINUX_CONTEXT_CHAR_ARRAY_SIZE]) { |
| 101 | for (domain d : AllDomains) { |
| 102 | // Not sure how to enforce this at compile time, so abort() bpfloader at boot instead |
| 103 | if (strlen(lookupSelinuxContext(d)) >= BPF_SELINUX_CONTEXT_CHAR_ARRAY_SIZE) abort(); |
| 104 | if (!strncmp(s, lookupSelinuxContext(d), BPF_SELINUX_CONTEXT_CHAR_ARRAY_SIZE)) return d; |
| 105 | } |
| 106 | ALOGW("ignoring unrecognized selinux_context '%-32s'", s); |
| 107 | // We should return 'unrecognized' here, however: returning unspecified will |
| 108 | // result in the system simply using the default context, which in turn |
| 109 | // will allow future expansion by adding more restrictive selinux types. |
| 110 | // Older bpfloader will simply ignore that, and use the less restrictive default. |
| 111 | // This does mean you CANNOT later add a *less* restrictive type than the default. |
| 112 | // |
| 113 | // Note: we cannot just abort() here as this might be a mainline module shipped optional update |
| 114 | return domain::unspecified; |
| 115 | } |
| 116 | |
| 117 | constexpr const char* lookupPinSubdir(const domain d, const char* const unspecified = "") { |
| 118 | switch (d) { |
| 119 | case domain::unspecified: return unspecified; |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 120 | case domain::tethering: return "tethering/"; |
| 121 | case domain::net_private: return "net_private/"; |
| 122 | case domain::net_shared: return "net_shared/"; |
| 123 | case domain::netd_readonly: return "netd_readonly/"; |
| 124 | case domain::netd_shared: return "netd_shared/"; |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 125 | default: return "(unrecognized)"; |
| 126 | } |
| 127 | }; |
| 128 | |
| 129 | domain getDomainFromPinSubdir(const char s[BPF_PIN_SUBDIR_CHAR_ARRAY_SIZE]) { |
| 130 | for (domain d : AllDomains) { |
| 131 | // Not sure how to enforce this at compile time, so abort() bpfloader at boot instead |
| 132 | if (strlen(lookupPinSubdir(d)) >= BPF_PIN_SUBDIR_CHAR_ARRAY_SIZE) abort(); |
| 133 | if (!strncmp(s, lookupPinSubdir(d), BPF_PIN_SUBDIR_CHAR_ARRAY_SIZE)) return d; |
| 134 | } |
| 135 | ALOGE("unrecognized pin_subdir '%-32s'", s); |
| 136 | // pin_subdir affects the object's full pathname, |
| 137 | // and thus using the default would change the location and thus our code's ability to find it, |
| 138 | // hence this seems worth treating as a true error condition. |
| 139 | // |
| 140 | // Note: we cannot just abort() here as this might be a mainline module shipped optional update |
| 141 | // However, our callers will treat this as an error, and stop loading the specific .o, |
| 142 | // which will fail bpfloader if the .o is marked critical. |
| 143 | return domain::unrecognized; |
| 144 | } |
| 145 | |
| 146 | static string pathToObjName(const string& path) { |
| 147 | // extract everything after the final slash, ie. this is the filename 'foo@1.o' or 'bar.o' |
| 148 | string filename = android::base::Split(path, "/").back(); |
| 149 | // strip off everything from the final period onwards (strip '.o' suffix), ie. 'foo@1' or 'bar' |
| 150 | string name = filename.substr(0, filename.find_last_of('.')); |
| 151 | // strip any potential @1 suffix, this will leave us with just 'foo' or 'bar' |
| 152 | // this can be used to provide duplicate programs (mux based on the bpfloader version) |
| 153 | return name.substr(0, name.find_last_of('@')); |
| 154 | } |
| 155 | |
| 156 | typedef struct { |
| 157 | const char* name; |
| 158 | enum bpf_prog_type type; |
| 159 | enum bpf_attach_type expected_attach_type; |
| 160 | } sectionType; |
| 161 | |
| 162 | /* |
| 163 | * Map section name prefixes to program types, the section name will be: |
| 164 | * SECTION(<prefix>/<name-of-program>) |
| 165 | * For example: |
| 166 | * SECTION("tracepoint/sched_switch_func") where sched_switch_funcs |
| 167 | * is the name of the program, and tracepoint is the type. |
| 168 | * |
| 169 | * However, be aware that you should not be directly using the SECTION() macro. |
| 170 | * Instead use the DEFINE_(BPF|XDP)_(PROG|MAP)... & LICENSE/CRITICAL macros. |
Maciej Żenczykowski | fcea70f | 2023-10-08 23:35:00 -0700 | [diff] [blame] | 171 | * |
| 172 | * Programs shipped inside the tethering apex should be limited to networking stuff, |
| 173 | * as KPROBE, PERF_EVENT, TRACEPOINT are dangerous to use from mainline updatable code, |
| 174 | * since they are less stable abi/api and may conflict with platform uses of bpf. |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 175 | */ |
| 176 | sectionType sectionNameTypes[] = { |
| 177 | {"bind4/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_INET4_BIND}, |
| 178 | {"bind6/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_INET6_BIND}, |
| 179 | {"cgroupskb/", BPF_PROG_TYPE_CGROUP_SKB, BPF_ATTACH_TYPE_UNSPEC}, |
| 180 | {"cgroupsock/", BPF_PROG_TYPE_CGROUP_SOCK, BPF_ATTACH_TYPE_UNSPEC}, |
| 181 | {"connect4/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_INET4_CONNECT}, |
| 182 | {"connect6/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_INET6_CONNECT}, |
| 183 | {"egress/", BPF_PROG_TYPE_CGROUP_SKB, BPF_CGROUP_INET_EGRESS}, |
| 184 | {"getsockopt/", BPF_PROG_TYPE_CGROUP_SOCKOPT, BPF_CGROUP_GETSOCKOPT}, |
| 185 | {"ingress/", BPF_PROG_TYPE_CGROUP_SKB, BPF_CGROUP_INET_INGRESS}, |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 186 | {"lwt_in/", BPF_PROG_TYPE_LWT_IN, BPF_ATTACH_TYPE_UNSPEC}, |
| 187 | {"lwt_out/", BPF_PROG_TYPE_LWT_OUT, BPF_ATTACH_TYPE_UNSPEC}, |
| 188 | {"lwt_seg6local/", BPF_PROG_TYPE_LWT_SEG6LOCAL, BPF_ATTACH_TYPE_UNSPEC}, |
| 189 | {"lwt_xmit/", BPF_PROG_TYPE_LWT_XMIT, BPF_ATTACH_TYPE_UNSPEC}, |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 190 | {"postbind4/", BPF_PROG_TYPE_CGROUP_SOCK, BPF_CGROUP_INET4_POST_BIND}, |
| 191 | {"postbind6/", BPF_PROG_TYPE_CGROUP_SOCK, BPF_CGROUP_INET6_POST_BIND}, |
| 192 | {"recvmsg4/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_UDP4_RECVMSG}, |
| 193 | {"recvmsg6/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_UDP6_RECVMSG}, |
| 194 | {"schedact/", BPF_PROG_TYPE_SCHED_ACT, BPF_ATTACH_TYPE_UNSPEC}, |
| 195 | {"schedcls/", BPF_PROG_TYPE_SCHED_CLS, BPF_ATTACH_TYPE_UNSPEC}, |
| 196 | {"sendmsg4/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_UDP4_SENDMSG}, |
| 197 | {"sendmsg6/", BPF_PROG_TYPE_CGROUP_SOCK_ADDR, BPF_CGROUP_UDP6_SENDMSG}, |
| 198 | {"setsockopt/", BPF_PROG_TYPE_CGROUP_SOCKOPT, BPF_CGROUP_SETSOCKOPT}, |
| 199 | {"skfilter/", BPF_PROG_TYPE_SOCKET_FILTER, BPF_ATTACH_TYPE_UNSPEC}, |
| 200 | {"sockops/", BPF_PROG_TYPE_SOCK_OPS, BPF_CGROUP_SOCK_OPS}, |
| 201 | {"sysctl", BPF_PROG_TYPE_CGROUP_SYSCTL, BPF_CGROUP_SYSCTL}, |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 202 | {"xdp/", BPF_PROG_TYPE_XDP, BPF_ATTACH_TYPE_UNSPEC}, |
| 203 | }; |
| 204 | |
| 205 | typedef struct { |
| 206 | enum bpf_prog_type type; |
| 207 | enum bpf_attach_type expected_attach_type; |
| 208 | string name; |
| 209 | vector<char> data; |
| 210 | vector<char> rel_data; |
| 211 | optional<struct bpf_prog_def> prog_def; |
| 212 | |
| 213 | unique_fd prog_fd; /* fd after loading */ |
| 214 | } codeSection; |
| 215 | |
| 216 | static int readElfHeader(ifstream& elfFile, Elf64_Ehdr* eh) { |
| 217 | elfFile.seekg(0); |
| 218 | if (elfFile.fail()) return -1; |
| 219 | |
| 220 | if (!elfFile.read((char*)eh, sizeof(*eh))) return -1; |
| 221 | |
| 222 | return 0; |
| 223 | } |
| 224 | |
| 225 | /* Reads all section header tables into an Shdr array */ |
| 226 | static int readSectionHeadersAll(ifstream& elfFile, vector<Elf64_Shdr>& shTable) { |
| 227 | Elf64_Ehdr eh; |
| 228 | int ret = 0; |
| 229 | |
| 230 | ret = readElfHeader(elfFile, &eh); |
| 231 | if (ret) return ret; |
| 232 | |
| 233 | elfFile.seekg(eh.e_shoff); |
| 234 | if (elfFile.fail()) return -1; |
| 235 | |
| 236 | /* Read shdr table entries */ |
| 237 | shTable.resize(eh.e_shnum); |
| 238 | |
| 239 | if (!elfFile.read((char*)shTable.data(), (eh.e_shnum * eh.e_shentsize))) return -ENOMEM; |
| 240 | |
| 241 | return 0; |
| 242 | } |
| 243 | |
| 244 | /* Read a section by its index - for ex to get sec hdr strtab blob */ |
| 245 | static int readSectionByIdx(ifstream& elfFile, int id, vector<char>& sec) { |
| 246 | vector<Elf64_Shdr> shTable; |
| 247 | int ret = readSectionHeadersAll(elfFile, shTable); |
| 248 | if (ret) return ret; |
| 249 | |
| 250 | elfFile.seekg(shTable[id].sh_offset); |
| 251 | if (elfFile.fail()) return -1; |
| 252 | |
| 253 | sec.resize(shTable[id].sh_size); |
| 254 | if (!elfFile.read(sec.data(), shTable[id].sh_size)) return -1; |
| 255 | |
| 256 | return 0; |
| 257 | } |
| 258 | |
| 259 | /* Read whole section header string table */ |
| 260 | static int readSectionHeaderStrtab(ifstream& elfFile, vector<char>& strtab) { |
| 261 | Elf64_Ehdr eh; |
| 262 | int ret = readElfHeader(elfFile, &eh); |
| 263 | if (ret) return ret; |
| 264 | |
| 265 | ret = readSectionByIdx(elfFile, eh.e_shstrndx, strtab); |
| 266 | if (ret) return ret; |
| 267 | |
| 268 | return 0; |
| 269 | } |
| 270 | |
| 271 | /* Get name from offset in strtab */ |
| 272 | static int getSymName(ifstream& elfFile, int nameOff, string& name) { |
| 273 | int ret; |
| 274 | vector<char> secStrTab; |
| 275 | |
| 276 | ret = readSectionHeaderStrtab(elfFile, secStrTab); |
| 277 | if (ret) return ret; |
| 278 | |
| 279 | if (nameOff >= (int)secStrTab.size()) return -1; |
| 280 | |
| 281 | name = string((char*)secStrTab.data() + nameOff); |
| 282 | return 0; |
| 283 | } |
| 284 | |
| 285 | /* Reads a full section by name - example to get the GPL license */ |
| 286 | static int readSectionByName(const char* name, ifstream& elfFile, vector<char>& data) { |
| 287 | vector<char> secStrTab; |
| 288 | vector<Elf64_Shdr> shTable; |
| 289 | int ret; |
| 290 | |
| 291 | ret = readSectionHeadersAll(elfFile, shTable); |
| 292 | if (ret) return ret; |
| 293 | |
| 294 | ret = readSectionHeaderStrtab(elfFile, secStrTab); |
| 295 | if (ret) return ret; |
| 296 | |
| 297 | for (int i = 0; i < (int)shTable.size(); i++) { |
| 298 | char* secname = secStrTab.data() + shTable[i].sh_name; |
| 299 | if (!secname) continue; |
| 300 | |
| 301 | if (!strcmp(secname, name)) { |
| 302 | vector<char> dataTmp; |
| 303 | dataTmp.resize(shTable[i].sh_size); |
| 304 | |
| 305 | elfFile.seekg(shTable[i].sh_offset); |
| 306 | if (elfFile.fail()) return -1; |
| 307 | |
| 308 | if (!elfFile.read((char*)dataTmp.data(), shTable[i].sh_size)) return -1; |
| 309 | |
| 310 | data = dataTmp; |
| 311 | return 0; |
| 312 | } |
| 313 | } |
| 314 | return -2; |
| 315 | } |
| 316 | |
| 317 | unsigned int readSectionUint(const char* name, ifstream& elfFile, unsigned int defVal) { |
| 318 | vector<char> theBytes; |
| 319 | int ret = readSectionByName(name, elfFile, theBytes); |
| 320 | if (ret) { |
| 321 | ALOGD("Couldn't find section %s (defaulting to %u [0x%x]).", name, defVal, defVal); |
| 322 | return defVal; |
| 323 | } else if (theBytes.size() < sizeof(unsigned int)) { |
| 324 | ALOGE("Section %s too short (defaulting to %u [0x%x]).", name, defVal, defVal); |
| 325 | return defVal; |
| 326 | } else { |
| 327 | // decode first 4 bytes as LE32 uint, there will likely be more bytes due to alignment. |
| 328 | unsigned int value = static_cast<unsigned char>(theBytes[3]); |
| 329 | value <<= 8; |
| 330 | value += static_cast<unsigned char>(theBytes[2]); |
| 331 | value <<= 8; |
| 332 | value += static_cast<unsigned char>(theBytes[1]); |
| 333 | value <<= 8; |
| 334 | value += static_cast<unsigned char>(theBytes[0]); |
| 335 | ALOGI("Section %s value is %u [0x%x]", name, value, value); |
| 336 | return value; |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | static int readSectionByType(ifstream& elfFile, int type, vector<char>& data) { |
| 341 | int ret; |
| 342 | vector<Elf64_Shdr> shTable; |
| 343 | |
| 344 | ret = readSectionHeadersAll(elfFile, shTable); |
| 345 | if (ret) return ret; |
| 346 | |
| 347 | for (int i = 0; i < (int)shTable.size(); i++) { |
| 348 | if ((int)shTable[i].sh_type != type) continue; |
| 349 | |
| 350 | vector<char> dataTmp; |
| 351 | dataTmp.resize(shTable[i].sh_size); |
| 352 | |
| 353 | elfFile.seekg(shTable[i].sh_offset); |
| 354 | if (elfFile.fail()) return -1; |
| 355 | |
| 356 | if (!elfFile.read((char*)dataTmp.data(), shTable[i].sh_size)) return -1; |
| 357 | |
| 358 | data = dataTmp; |
| 359 | return 0; |
| 360 | } |
| 361 | return -2; |
| 362 | } |
| 363 | |
| 364 | static bool symCompare(Elf64_Sym a, Elf64_Sym b) { |
| 365 | return (a.st_value < b.st_value); |
| 366 | } |
| 367 | |
| 368 | static int readSymTab(ifstream& elfFile, int sort, vector<Elf64_Sym>& data) { |
| 369 | int ret, numElems; |
| 370 | Elf64_Sym* buf; |
| 371 | vector<char> secData; |
| 372 | |
| 373 | ret = readSectionByType(elfFile, SHT_SYMTAB, secData); |
| 374 | if (ret) return ret; |
| 375 | |
| 376 | buf = (Elf64_Sym*)secData.data(); |
| 377 | numElems = (secData.size() / sizeof(Elf64_Sym)); |
| 378 | data.assign(buf, buf + numElems); |
| 379 | |
| 380 | if (sort) std::sort(data.begin(), data.end(), symCompare); |
| 381 | return 0; |
| 382 | } |
| 383 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 384 | static enum bpf_prog_type getSectionType(string& name) { |
| 385 | for (auto& snt : sectionNameTypes) |
| 386 | if (StartsWith(name, snt.name)) return snt.type; |
| 387 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 388 | return BPF_PROG_TYPE_UNSPEC; |
| 389 | } |
| 390 | |
| 391 | static enum bpf_attach_type getExpectedAttachType(string& name) { |
| 392 | for (auto& snt : sectionNameTypes) |
| 393 | if (StartsWith(name, snt.name)) return snt.expected_attach_type; |
| 394 | return BPF_ATTACH_TYPE_UNSPEC; |
| 395 | } |
| 396 | |
Maciej Żenczykowski | fcea70f | 2023-10-08 23:35:00 -0700 | [diff] [blame] | 397 | /* |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 398 | static string getSectionName(enum bpf_prog_type type) |
| 399 | { |
| 400 | for (auto& snt : sectionNameTypes) |
| 401 | if (snt.type == type) |
| 402 | return string(snt.name); |
| 403 | |
| 404 | return "UNKNOWN SECTION NAME " + std::to_string(type); |
| 405 | } |
Maciej Żenczykowski | fcea70f | 2023-10-08 23:35:00 -0700 | [diff] [blame] | 406 | */ |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 407 | |
| 408 | static int readProgDefs(ifstream& elfFile, vector<struct bpf_prog_def>& pd, |
| 409 | size_t sizeOfBpfProgDef) { |
| 410 | vector<char> pdData; |
| 411 | int ret = readSectionByName("progs", elfFile, pdData); |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 412 | if (ret) return ret; |
| 413 | |
| 414 | if (pdData.size() % sizeOfBpfProgDef) { |
| 415 | ALOGE("readProgDefs failed due to improper sized progs section, %zu %% %zu != 0", |
| 416 | pdData.size(), sizeOfBpfProgDef); |
| 417 | return -1; |
| 418 | }; |
| 419 | |
| 420 | int progCount = pdData.size() / sizeOfBpfProgDef; |
| 421 | pd.resize(progCount); |
| 422 | size_t trimmedSize = std::min(sizeOfBpfProgDef, sizeof(struct bpf_prog_def)); |
| 423 | |
| 424 | const char* dataPtr = pdData.data(); |
| 425 | for (auto& p : pd) { |
| 426 | // First we zero initialize |
| 427 | memset(&p, 0, sizeof(p)); |
| 428 | // Then we set non-zero defaults |
| 429 | p.bpfloader_max_ver = DEFAULT_BPFLOADER_MAX_VER; // v1.0 |
| 430 | // Then we copy over the structure prefix from the ELF file. |
| 431 | memcpy(&p, dataPtr, trimmedSize); |
| 432 | // Move to next struct in the ELF file |
| 433 | dataPtr += sizeOfBpfProgDef; |
| 434 | } |
| 435 | return 0; |
| 436 | } |
| 437 | |
| 438 | static int getSectionSymNames(ifstream& elfFile, const string& sectionName, vector<string>& names, |
| 439 | optional<unsigned> symbolType = std::nullopt) { |
| 440 | int ret; |
| 441 | string name; |
| 442 | vector<Elf64_Sym> symtab; |
| 443 | vector<Elf64_Shdr> shTable; |
| 444 | |
| 445 | ret = readSymTab(elfFile, 1 /* sort */, symtab); |
| 446 | if (ret) return ret; |
| 447 | |
| 448 | /* Get index of section */ |
| 449 | ret = readSectionHeadersAll(elfFile, shTable); |
| 450 | if (ret) return ret; |
| 451 | |
| 452 | int sec_idx = -1; |
| 453 | for (int i = 0; i < (int)shTable.size(); i++) { |
| 454 | ret = getSymName(elfFile, shTable[i].sh_name, name); |
| 455 | if (ret) return ret; |
| 456 | |
| 457 | if (!name.compare(sectionName)) { |
| 458 | sec_idx = i; |
| 459 | break; |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | /* No section found with matching name*/ |
| 464 | if (sec_idx == -1) { |
| 465 | ALOGW("No %s section could be found in elf object", sectionName.c_str()); |
| 466 | return -1; |
| 467 | } |
| 468 | |
| 469 | for (int i = 0; i < (int)symtab.size(); i++) { |
| 470 | if (symbolType.has_value() && ELF_ST_TYPE(symtab[i].st_info) != symbolType) continue; |
| 471 | |
| 472 | if (symtab[i].st_shndx == sec_idx) { |
| 473 | string s; |
| 474 | ret = getSymName(elfFile, symtab[i].st_name, s); |
| 475 | if (ret) return ret; |
| 476 | names.push_back(s); |
| 477 | } |
| 478 | } |
| 479 | |
| 480 | return 0; |
| 481 | } |
| 482 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 483 | /* Read a section by its index - for ex to get sec hdr strtab blob */ |
Maciej Żenczykowski | fcea70f | 2023-10-08 23:35:00 -0700 | [diff] [blame] | 484 | static int readCodeSections(ifstream& elfFile, vector<codeSection>& cs, size_t sizeOfBpfProgDef) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 485 | vector<Elf64_Shdr> shTable; |
| 486 | int entries, ret = 0; |
| 487 | |
| 488 | ret = readSectionHeadersAll(elfFile, shTable); |
| 489 | if (ret) return ret; |
| 490 | entries = shTable.size(); |
| 491 | |
| 492 | vector<struct bpf_prog_def> pd; |
| 493 | ret = readProgDefs(elfFile, pd, sizeOfBpfProgDef); |
| 494 | if (ret) return ret; |
| 495 | vector<string> progDefNames; |
| 496 | ret = getSectionSymNames(elfFile, "progs", progDefNames); |
| 497 | if (!pd.empty() && ret) return ret; |
| 498 | |
| 499 | for (int i = 0; i < entries; i++) { |
| 500 | string name; |
| 501 | codeSection cs_temp; |
| 502 | cs_temp.type = BPF_PROG_TYPE_UNSPEC; |
| 503 | |
| 504 | ret = getSymName(elfFile, shTable[i].sh_name, name); |
| 505 | if (ret) return ret; |
| 506 | |
| 507 | enum bpf_prog_type ptype = getSectionType(name); |
| 508 | |
| 509 | if (ptype == BPF_PROG_TYPE_UNSPEC) continue; |
| 510 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 511 | // This must be done before '/' is replaced with '_'. |
| 512 | cs_temp.expected_attach_type = getExpectedAttachType(name); |
| 513 | |
| 514 | string oldName = name; |
| 515 | |
| 516 | // convert all slashes to underscores |
| 517 | std::replace(name.begin(), name.end(), '/', '_'); |
| 518 | |
| 519 | cs_temp.type = ptype; |
| 520 | cs_temp.name = name; |
| 521 | |
| 522 | ret = readSectionByIdx(elfFile, i, cs_temp.data); |
| 523 | if (ret) return ret; |
| 524 | ALOGD("Loaded code section %d (%s)", i, name.c_str()); |
| 525 | |
| 526 | vector<string> csSymNames; |
| 527 | ret = getSectionSymNames(elfFile, oldName, csSymNames, STT_FUNC); |
| 528 | if (ret || !csSymNames.size()) return ret; |
| 529 | for (size_t i = 0; i < progDefNames.size(); ++i) { |
| 530 | if (!progDefNames[i].compare(csSymNames[0] + "_def")) { |
| 531 | cs_temp.prog_def = pd[i]; |
| 532 | break; |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | /* Check for rel section */ |
| 537 | if (cs_temp.data.size() > 0 && i < entries) { |
| 538 | ret = getSymName(elfFile, shTable[i + 1].sh_name, name); |
| 539 | if (ret) return ret; |
| 540 | |
| 541 | if (name == (".rel" + oldName)) { |
| 542 | ret = readSectionByIdx(elfFile, i + 1, cs_temp.rel_data); |
| 543 | if (ret) return ret; |
| 544 | ALOGD("Loaded relo section %d (%s)", i, name.c_str()); |
| 545 | } |
| 546 | } |
| 547 | |
| 548 | if (cs_temp.data.size() > 0) { |
| 549 | cs.push_back(std::move(cs_temp)); |
| 550 | ALOGD("Adding section %d to cs list", i); |
| 551 | } |
| 552 | } |
| 553 | return 0; |
| 554 | } |
| 555 | |
| 556 | static int getSymNameByIdx(ifstream& elfFile, int index, string& name) { |
| 557 | vector<Elf64_Sym> symtab; |
| 558 | int ret = 0; |
| 559 | |
| 560 | ret = readSymTab(elfFile, 0 /* !sort */, symtab); |
| 561 | if (ret) return ret; |
| 562 | |
| 563 | if (index >= (int)symtab.size()) return -1; |
| 564 | |
| 565 | return getSymName(elfFile, symtab[index].st_name, name); |
| 566 | } |
| 567 | |
| 568 | static bool mapMatchesExpectations(const unique_fd& fd, const string& mapName, |
| 569 | const struct bpf_map_def& mapDef, const enum bpf_map_type type) { |
Maciej Żenczykowski | 7e5b09a | 2024-03-11 12:40:06 +0000 | [diff] [blame] | 570 | // bpfGetFd... family of functions require at minimum a 4.14 kernel, |
| 571 | // so on 4.9-T kernels just pretend the map matches our expectations. |
| 572 | // Additionally we'll get almost equivalent test coverage on newer devices/kernels. |
| 573 | // This is because the primary failure mode we're trying to detect here |
| 574 | // is either a source code misconfiguration (which is likely kernel independent) |
| 575 | // or a newly introduced kernel feature/bug (which is unlikely to get backported to 4.9). |
| 576 | if (!isAtLeastKernelVersion(4, 14, 0)) return true; |
| 577 | |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 578 | // Assuming fd is a valid Bpf Map file descriptor then |
| 579 | // all the following should always succeed on a 4.14+ kernel. |
| 580 | // If they somehow do fail, they'll return -1 (and set errno), |
| 581 | // which should then cause (among others) a key_size mismatch. |
| 582 | int fd_type = bpfGetFdMapType(fd); |
| 583 | int fd_key_size = bpfGetFdKeySize(fd); |
| 584 | int fd_value_size = bpfGetFdValueSize(fd); |
| 585 | int fd_max_entries = bpfGetFdMaxEntries(fd); |
| 586 | int fd_map_flags = bpfGetFdMapFlags(fd); |
| 587 | |
| 588 | // DEVMAPs are readonly from the bpf program side's point of view, as such |
| 589 | // the kernel in kernel/bpf/devmap.c dev_map_init_map() will set the flag |
| 590 | int desired_map_flags = (int)mapDef.map_flags; |
| 591 | if (type == BPF_MAP_TYPE_DEVMAP || type == BPF_MAP_TYPE_DEVMAP_HASH) |
| 592 | desired_map_flags |= BPF_F_RDONLY_PROG; |
| 593 | |
| 594 | // The .h file enforces that this is a power of two, and page size will |
| 595 | // also always be a power of two, so this logic is actually enough to |
| 596 | // force it to be a multiple of the page size, as required by the kernel. |
| 597 | unsigned int desired_max_entries = mapDef.max_entries; |
| 598 | if (type == BPF_MAP_TYPE_RINGBUF) { |
| 599 | if (desired_max_entries < page_size) desired_max_entries = page_size; |
| 600 | } |
| 601 | |
| 602 | // The following checks should *never* trigger, if one of them somehow does, |
| 603 | // it probably means a bpf .o file has been changed/replaced at runtime |
| 604 | // and bpfloader was manually rerun (normally it should only run *once* |
| 605 | // early during the boot process). |
| 606 | // Another possibility is that something is misconfigured in the code: |
| 607 | // most likely a shared map is declared twice differently. |
| 608 | // But such a change should never be checked into the source tree... |
| 609 | if ((fd_type == type) && |
| 610 | (fd_key_size == (int)mapDef.key_size) && |
| 611 | (fd_value_size == (int)mapDef.value_size) && |
| 612 | (fd_max_entries == (int)desired_max_entries) && |
| 613 | (fd_map_flags == desired_map_flags)) { |
| 614 | return true; |
| 615 | } |
| 616 | |
| 617 | ALOGE("bpf map name %s mismatch: desired/found: " |
| 618 | "type:%d/%d key:%u/%d value:%u/%d entries:%u/%d flags:%u/%d", |
| 619 | mapName.c_str(), type, fd_type, mapDef.key_size, fd_key_size, mapDef.value_size, |
| 620 | fd_value_size, mapDef.max_entries, fd_max_entries, desired_map_flags, fd_map_flags); |
| 621 | return false; |
| 622 | } |
| 623 | |
| 624 | static int createMaps(const char* elfPath, ifstream& elfFile, vector<unique_fd>& mapFds, |
Maciej Żenczykowski | af07d6d | 2023-10-10 03:40:51 -0700 | [diff] [blame] | 625 | const char* prefix, const size_t sizeOfBpfMapDef) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 626 | int ret; |
| 627 | vector<char> mdData; |
| 628 | vector<struct bpf_map_def> md; |
| 629 | vector<string> mapNames; |
| 630 | string objName = pathToObjName(string(elfPath)); |
| 631 | |
| 632 | ret = readSectionByName("maps", elfFile, mdData); |
| 633 | if (ret == -2) return 0; // no maps to read |
| 634 | if (ret) return ret; |
| 635 | |
| 636 | if (mdData.size() % sizeOfBpfMapDef) { |
| 637 | ALOGE("createMaps failed due to improper sized maps section, %zu %% %zu != 0", |
| 638 | mdData.size(), sizeOfBpfMapDef); |
| 639 | return -1; |
| 640 | }; |
| 641 | |
| 642 | int mapCount = mdData.size() / sizeOfBpfMapDef; |
| 643 | md.resize(mapCount); |
| 644 | size_t trimmedSize = std::min(sizeOfBpfMapDef, sizeof(struct bpf_map_def)); |
| 645 | |
| 646 | const char* dataPtr = mdData.data(); |
| 647 | for (auto& m : md) { |
| 648 | // First we zero initialize |
| 649 | memset(&m, 0, sizeof(m)); |
| 650 | // Then we set non-zero defaults |
| 651 | m.bpfloader_max_ver = DEFAULT_BPFLOADER_MAX_VER; // v1.0 |
| 652 | m.max_kver = 0xFFFFFFFFu; // matches KVER_INF from bpf_helpers.h |
| 653 | // Then we copy over the structure prefix from the ELF file. |
| 654 | memcpy(&m, dataPtr, trimmedSize); |
| 655 | // Move to next struct in the ELF file |
| 656 | dataPtr += sizeOfBpfMapDef; |
| 657 | } |
| 658 | |
| 659 | ret = getSectionSymNames(elfFile, "maps", mapNames); |
| 660 | if (ret) return ret; |
| 661 | |
| 662 | unsigned kvers = kernelVersion(); |
| 663 | |
| 664 | for (int i = 0; i < (int)mapNames.size(); i++) { |
| 665 | if (md[i].zero != 0) abort(); |
| 666 | |
| 667 | if (BPFLOADER_VERSION < md[i].bpfloader_min_ver) { |
| 668 | ALOGI("skipping map %s which requires bpfloader min ver 0x%05x", mapNames[i].c_str(), |
| 669 | md[i].bpfloader_min_ver); |
| 670 | mapFds.push_back(unique_fd()); |
| 671 | continue; |
| 672 | } |
| 673 | |
| 674 | if (BPFLOADER_VERSION >= md[i].bpfloader_max_ver) { |
| 675 | ALOGI("skipping map %s which requires bpfloader max ver 0x%05x", mapNames[i].c_str(), |
| 676 | md[i].bpfloader_max_ver); |
| 677 | mapFds.push_back(unique_fd()); |
| 678 | continue; |
| 679 | } |
| 680 | |
| 681 | if (kvers < md[i].min_kver) { |
| 682 | ALOGI("skipping map %s which requires kernel version 0x%x >= 0x%x", |
| 683 | mapNames[i].c_str(), kvers, md[i].min_kver); |
| 684 | mapFds.push_back(unique_fd()); |
| 685 | continue; |
| 686 | } |
| 687 | |
| 688 | if (kvers >= md[i].max_kver) { |
| 689 | ALOGI("skipping map %s which requires kernel version 0x%x < 0x%x", |
| 690 | mapNames[i].c_str(), kvers, md[i].max_kver); |
| 691 | mapFds.push_back(unique_fd()); |
| 692 | continue; |
| 693 | } |
| 694 | |
| 695 | if ((md[i].ignore_on_eng && isEng()) || (md[i].ignore_on_user && isUser()) || |
| 696 | (md[i].ignore_on_userdebug && isUserdebug())) { |
| 697 | ALOGI("skipping map %s which is ignored on %s builds", mapNames[i].c_str(), |
| 698 | getBuildType().c_str()); |
| 699 | mapFds.push_back(unique_fd()); |
| 700 | continue; |
| 701 | } |
| 702 | |
| 703 | if ((isArm() && isKernel32Bit() && md[i].ignore_on_arm32) || |
| 704 | (isArm() && isKernel64Bit() && md[i].ignore_on_aarch64) || |
| 705 | (isX86() && isKernel32Bit() && md[i].ignore_on_x86_32) || |
| 706 | (isX86() && isKernel64Bit() && md[i].ignore_on_x86_64) || |
| 707 | (isRiscV() && md[i].ignore_on_riscv64)) { |
| 708 | ALOGI("skipping map %s which is ignored on %s", mapNames[i].c_str(), |
| 709 | describeArch()); |
| 710 | mapFds.push_back(unique_fd()); |
| 711 | continue; |
| 712 | } |
| 713 | |
| 714 | enum bpf_map_type type = md[i].type; |
Maciej Żenczykowski | 7e5b09a | 2024-03-11 12:40:06 +0000 | [diff] [blame] | 715 | if (type == BPF_MAP_TYPE_DEVMAP && !isAtLeastKernelVersion(4, 14, 0)) { |
| 716 | // On Linux Kernels older than 4.14 this map type doesn't exist, but it can kind |
| 717 | // of be approximated: ARRAY has the same userspace api, though it is not usable |
| 718 | // by the same ebpf programs. However, that's okay because the bpf_redirect_map() |
| 719 | // helper doesn't exist on 4.9-T anyway (so the bpf program would fail to load, |
| 720 | // and thus needs to be tagged as 4.14+ either way), so there's nothing useful you |
| 721 | // could do with a DEVMAP anyway (that isn't already provided by an ARRAY)... |
| 722 | // Hence using an ARRAY instead of a DEVMAP simply makes life easier for userspace. |
| 723 | type = BPF_MAP_TYPE_ARRAY; |
| 724 | } |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 725 | if (type == BPF_MAP_TYPE_DEVMAP_HASH && !isAtLeastKernelVersion(5, 4, 0)) { |
| 726 | // On Linux Kernels older than 5.4 this map type doesn't exist, but it can kind |
| 727 | // of be approximated: HASH has the same userspace visible api. |
| 728 | // However it cannot be used by ebpf programs in the same way. |
| 729 | // Since bpf_redirect_map() only requires 4.14, a program using a DEVMAP_HASH map |
| 730 | // would fail to load (due to trying to redirect to a HASH instead of DEVMAP_HASH). |
| 731 | // One must thus tag any BPF_MAP_TYPE_DEVMAP_HASH + bpf_redirect_map() using |
| 732 | // programs as being 5.4+... |
| 733 | type = BPF_MAP_TYPE_HASH; |
| 734 | } |
| 735 | |
| 736 | // The .h file enforces that this is a power of two, and page size will |
| 737 | // also always be a power of two, so this logic is actually enough to |
| 738 | // force it to be a multiple of the page size, as required by the kernel. |
| 739 | unsigned int max_entries = md[i].max_entries; |
| 740 | if (type == BPF_MAP_TYPE_RINGBUF) { |
| 741 | if (max_entries < page_size) max_entries = page_size; |
| 742 | } |
| 743 | |
| 744 | domain selinux_context = getDomainFromSelinuxContext(md[i].selinux_context); |
| 745 | if (specified(selinux_context)) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 746 | ALOGI("map %s selinux_context [%-32s] -> %d -> '%s' (%s)", mapNames[i].c_str(), |
| 747 | md[i].selinux_context, selinux_context, lookupSelinuxContext(selinux_context), |
| 748 | lookupPinSubdir(selinux_context)); |
| 749 | } |
| 750 | |
| 751 | domain pin_subdir = getDomainFromPinSubdir(md[i].pin_subdir); |
| 752 | if (unrecognized(pin_subdir)) return -ENOTDIR; |
| 753 | if (specified(pin_subdir)) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 754 | ALOGI("map %s pin_subdir [%-32s] -> %d -> '%s'", mapNames[i].c_str(), md[i].pin_subdir, |
| 755 | pin_subdir, lookupPinSubdir(pin_subdir)); |
| 756 | } |
| 757 | |
| 758 | // Format of pin location is /sys/fs/bpf/<pin_subdir|prefix>map_<objName>_<mapName> |
| 759 | // except that maps shared across .o's have empty <objName> |
| 760 | // Note: <objName> refers to the extension-less basename of the .o file (without @ suffix). |
| 761 | string mapPinLoc = string(BPF_FS_PATH) + lookupPinSubdir(pin_subdir, prefix) + "map_" + |
| 762 | (md[i].shared ? "" : objName) + "_" + mapNames[i]; |
| 763 | bool reuse = false; |
| 764 | unique_fd fd; |
| 765 | int saved_errno; |
| 766 | |
| 767 | if (access(mapPinLoc.c_str(), F_OK) == 0) { |
| 768 | fd.reset(mapRetrieveRO(mapPinLoc.c_str())); |
| 769 | saved_errno = errno; |
| 770 | ALOGD("bpf_create_map reusing map %s, ret: %d", mapNames[i].c_str(), fd.get()); |
| 771 | reuse = true; |
| 772 | } else { |
| 773 | union bpf_attr req = { |
| 774 | .map_type = type, |
| 775 | .key_size = md[i].key_size, |
| 776 | .value_size = md[i].value_size, |
| 777 | .max_entries = max_entries, |
| 778 | .map_flags = md[i].map_flags, |
| 779 | }; |
| 780 | strlcpy(req.map_name, mapNames[i].c_str(), sizeof(req.map_name)); |
| 781 | fd.reset(bpf(BPF_MAP_CREATE, req)); |
| 782 | saved_errno = errno; |
| 783 | ALOGD("bpf_create_map name %s, ret: %d", mapNames[i].c_str(), fd.get()); |
| 784 | } |
| 785 | |
| 786 | if (!fd.ok()) return -saved_errno; |
| 787 | |
| 788 | // When reusing a pinned map, we need to check the map type/sizes/etc match, but for |
| 789 | // safety (since reuse code path is rare) run these checks even if we just created it. |
| 790 | // We assume failure is due to pinned map mismatch, hence the 'NOT UNIQUE' return code. |
| 791 | if (!mapMatchesExpectations(fd, mapNames[i], md[i], type)) return -ENOTUNIQ; |
| 792 | |
| 793 | if (!reuse) { |
| 794 | if (specified(selinux_context)) { |
| 795 | string createLoc = string(BPF_FS_PATH) + lookupPinSubdir(selinux_context) + |
| 796 | "tmp_map_" + objName + "_" + mapNames[i]; |
| 797 | ret = bpfFdPin(fd, createLoc.c_str()); |
| 798 | if (ret) { |
| 799 | int err = errno; |
| 800 | ALOGE("create %s -> %d [%d:%s]", createLoc.c_str(), ret, err, strerror(err)); |
| 801 | return -err; |
| 802 | } |
| 803 | ret = renameat2(AT_FDCWD, createLoc.c_str(), |
| 804 | AT_FDCWD, mapPinLoc.c_str(), RENAME_NOREPLACE); |
| 805 | if (ret) { |
| 806 | int err = errno; |
| 807 | ALOGE("rename %s %s -> %d [%d:%s]", createLoc.c_str(), mapPinLoc.c_str(), ret, |
| 808 | err, strerror(err)); |
| 809 | return -err; |
| 810 | } |
| 811 | } else { |
| 812 | ret = bpfFdPin(fd, mapPinLoc.c_str()); |
| 813 | if (ret) { |
| 814 | int err = errno; |
| 815 | ALOGE("pin %s -> %d [%d:%s]", mapPinLoc.c_str(), ret, err, strerror(err)); |
| 816 | return -err; |
| 817 | } |
| 818 | } |
| 819 | ret = chmod(mapPinLoc.c_str(), md[i].mode); |
| 820 | if (ret) { |
| 821 | int err = errno; |
| 822 | ALOGE("chmod(%s, 0%o) = %d [%d:%s]", mapPinLoc.c_str(), md[i].mode, ret, err, |
| 823 | strerror(err)); |
| 824 | return -err; |
| 825 | } |
| 826 | ret = chown(mapPinLoc.c_str(), (uid_t)md[i].uid, (gid_t)md[i].gid); |
| 827 | if (ret) { |
| 828 | int err = errno; |
| 829 | ALOGE("chown(%s, %u, %u) = %d [%d:%s]", mapPinLoc.c_str(), md[i].uid, md[i].gid, |
| 830 | ret, err, strerror(err)); |
| 831 | return -err; |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | int mapId = bpfGetFdMapId(fd); |
| 836 | if (mapId == -1) { |
| 837 | ALOGE("bpfGetFdMapId failed, ret: %d [%d]", mapId, errno); |
| 838 | } else { |
| 839 | ALOGI("map %s id %d", mapPinLoc.c_str(), mapId); |
| 840 | } |
| 841 | |
| 842 | mapFds.push_back(std::move(fd)); |
| 843 | } |
| 844 | |
| 845 | return ret; |
| 846 | } |
| 847 | |
| 848 | /* For debugging, dump all instructions */ |
| 849 | static void dumpIns(char* ins, int size) { |
| 850 | for (int row = 0; row < size / 8; row++) { |
| 851 | ALOGE("%d: ", row); |
| 852 | for (int j = 0; j < 8; j++) { |
| 853 | ALOGE("%3x ", ins[(row * 8) + j]); |
| 854 | } |
| 855 | ALOGE("\n"); |
| 856 | } |
| 857 | } |
| 858 | |
| 859 | /* For debugging, dump all code sections from cs list */ |
| 860 | static void dumpAllCs(vector<codeSection>& cs) { |
| 861 | for (int i = 0; i < (int)cs.size(); i++) { |
| 862 | ALOGE("Dumping cs %d, name %s", int(i), cs[i].name.c_str()); |
| 863 | dumpIns((char*)cs[i].data.data(), cs[i].data.size()); |
| 864 | ALOGE("-----------"); |
| 865 | } |
| 866 | } |
| 867 | |
| 868 | static void applyRelo(void* insnsPtr, Elf64_Addr offset, int fd) { |
| 869 | int insnIndex; |
| 870 | struct bpf_insn *insn, *insns; |
| 871 | |
| 872 | insns = (struct bpf_insn*)(insnsPtr); |
| 873 | |
| 874 | insnIndex = offset / sizeof(struct bpf_insn); |
| 875 | insn = &insns[insnIndex]; |
| 876 | |
| 877 | // Occasionally might be useful for relocation debugging, but pretty spammy |
| 878 | if (0) { |
| 879 | ALOGD("applying relo to instruction at byte offset: %llu, " |
| 880 | "insn offset %d, insn %llx", |
| 881 | (unsigned long long)offset, insnIndex, *(unsigned long long*)insn); |
| 882 | } |
| 883 | |
| 884 | if (insn->code != (BPF_LD | BPF_IMM | BPF_DW)) { |
| 885 | ALOGE("Dumping all instructions till ins %d", insnIndex); |
| 886 | ALOGE("invalid relo for insn %d: code 0x%x", insnIndex, insn->code); |
| 887 | dumpIns((char*)insnsPtr, (insnIndex + 3) * 8); |
| 888 | return; |
| 889 | } |
| 890 | |
| 891 | insn->imm = fd; |
| 892 | insn->src_reg = BPF_PSEUDO_MAP_FD; |
| 893 | } |
| 894 | |
| 895 | static void applyMapRelo(ifstream& elfFile, vector<unique_fd> &mapFds, vector<codeSection>& cs) { |
| 896 | vector<string> mapNames; |
| 897 | |
| 898 | int ret = getSectionSymNames(elfFile, "maps", mapNames); |
| 899 | if (ret) return; |
| 900 | |
| 901 | for (int k = 0; k != (int)cs.size(); k++) { |
| 902 | Elf64_Rel* rel = (Elf64_Rel*)(cs[k].rel_data.data()); |
| 903 | int n_rel = cs[k].rel_data.size() / sizeof(*rel); |
| 904 | |
| 905 | for (int i = 0; i < n_rel; i++) { |
| 906 | int symIndex = ELF64_R_SYM(rel[i].r_info); |
| 907 | string symName; |
| 908 | |
| 909 | ret = getSymNameByIdx(elfFile, symIndex, symName); |
| 910 | if (ret) return; |
| 911 | |
| 912 | /* Find the map fd and apply relo */ |
| 913 | for (int j = 0; j < (int)mapNames.size(); j++) { |
| 914 | if (!mapNames[j].compare(symName)) { |
| 915 | applyRelo(cs[k].data.data(), rel[i].r_offset, mapFds[j]); |
| 916 | break; |
| 917 | } |
| 918 | } |
| 919 | } |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | static int loadCodeSections(const char* elfPath, vector<codeSection>& cs, const string& license, |
Maciej Żenczykowski | af07d6d | 2023-10-10 03:40:51 -0700 | [diff] [blame] | 924 | const char* prefix) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 925 | unsigned kvers = kernelVersion(); |
| 926 | |
| 927 | if (!kvers) { |
| 928 | ALOGE("unable to get kernel version"); |
| 929 | return -EINVAL; |
| 930 | } |
| 931 | |
| 932 | string objName = pathToObjName(string(elfPath)); |
| 933 | |
| 934 | for (int i = 0; i < (int)cs.size(); i++) { |
| 935 | unique_fd& fd = cs[i].prog_fd; |
| 936 | int ret; |
| 937 | string name = cs[i].name; |
| 938 | |
| 939 | if (!cs[i].prog_def.has_value()) { |
| 940 | ALOGE("[%d] '%s' missing program definition! bad bpf.o build?", i, name.c_str()); |
| 941 | return -EINVAL; |
| 942 | } |
| 943 | |
| 944 | unsigned min_kver = cs[i].prog_def->min_kver; |
| 945 | unsigned max_kver = cs[i].prog_def->max_kver; |
| 946 | ALOGD("cs[%d].name:%s min_kver:%x .max_kver:%x (kvers:%x)", i, name.c_str(), min_kver, |
| 947 | max_kver, kvers); |
| 948 | if (kvers < min_kver) continue; |
| 949 | if (kvers >= max_kver) continue; |
| 950 | |
| 951 | unsigned bpfMinVer = cs[i].prog_def->bpfloader_min_ver; |
| 952 | unsigned bpfMaxVer = cs[i].prog_def->bpfloader_max_ver; |
| 953 | domain selinux_context = getDomainFromSelinuxContext(cs[i].prog_def->selinux_context); |
| 954 | domain pin_subdir = getDomainFromPinSubdir(cs[i].prog_def->pin_subdir); |
| 955 | // Note: make sure to only check for unrecognized *after* verifying bpfloader |
| 956 | // version limits include this bpfloader's version. |
| 957 | |
| 958 | ALOGD("cs[%d].name:%s requires bpfloader version [0x%05x,0x%05x)", i, name.c_str(), |
| 959 | bpfMinVer, bpfMaxVer); |
| 960 | if (BPFLOADER_VERSION < bpfMinVer) continue; |
| 961 | if (BPFLOADER_VERSION >= bpfMaxVer) continue; |
| 962 | |
| 963 | if ((cs[i].prog_def->ignore_on_eng && isEng()) || |
| 964 | (cs[i].prog_def->ignore_on_user && isUser()) || |
| 965 | (cs[i].prog_def->ignore_on_userdebug && isUserdebug())) { |
| 966 | ALOGD("cs[%d].name:%s is ignored on %s builds", i, name.c_str(), |
| 967 | getBuildType().c_str()); |
| 968 | continue; |
| 969 | } |
| 970 | |
| 971 | if ((isArm() && isKernel32Bit() && cs[i].prog_def->ignore_on_arm32) || |
| 972 | (isArm() && isKernel64Bit() && cs[i].prog_def->ignore_on_aarch64) || |
| 973 | (isX86() && isKernel32Bit() && cs[i].prog_def->ignore_on_x86_32) || |
| 974 | (isX86() && isKernel64Bit() && cs[i].prog_def->ignore_on_x86_64) || |
| 975 | (isRiscV() && cs[i].prog_def->ignore_on_riscv64)) { |
| 976 | ALOGD("cs[%d].name:%s is ignored on %s", i, name.c_str(), describeArch()); |
| 977 | continue; |
| 978 | } |
| 979 | |
| 980 | if (unrecognized(pin_subdir)) return -ENOTDIR; |
| 981 | |
| 982 | if (specified(selinux_context)) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 983 | ALOGI("prog %s selinux_context [%-32s] -> %d -> '%s' (%s)", name.c_str(), |
| 984 | cs[i].prog_def->selinux_context, selinux_context, |
| 985 | lookupSelinuxContext(selinux_context), lookupPinSubdir(selinux_context)); |
| 986 | } |
| 987 | |
| 988 | if (specified(pin_subdir)) { |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 989 | ALOGI("prog %s pin_subdir [%-32s] -> %d -> '%s'", name.c_str(), |
| 990 | cs[i].prog_def->pin_subdir, pin_subdir, lookupPinSubdir(pin_subdir)); |
| 991 | } |
| 992 | |
| 993 | // strip any potential $foo suffix |
| 994 | // this can be used to provide duplicate programs |
| 995 | // conditionally loaded based on running kernel version |
| 996 | name = name.substr(0, name.find_last_of('$')); |
| 997 | |
| 998 | bool reuse = false; |
| 999 | // Format of pin location is |
| 1000 | // /sys/fs/bpf/<prefix>prog_<objName>_<progName> |
| 1001 | string progPinLoc = string(BPF_FS_PATH) + lookupPinSubdir(pin_subdir, prefix) + "prog_" + |
| 1002 | objName + '_' + string(name); |
| 1003 | if (access(progPinLoc.c_str(), F_OK) == 0) { |
| 1004 | fd.reset(retrieveProgram(progPinLoc.c_str())); |
| 1005 | ALOGD("New bpf prog load reusing prog %s, ret: %d (%s)", progPinLoc.c_str(), fd.get(), |
| 1006 | (!fd.ok() ? std::strerror(errno) : "no error")); |
| 1007 | reuse = true; |
| 1008 | } else { |
| 1009 | vector<char> log_buf(BPF_LOAD_LOG_SZ, 0); |
| 1010 | |
| 1011 | union bpf_attr req = { |
| 1012 | .prog_type = cs[i].type, |
| 1013 | .kern_version = kvers, |
| 1014 | .license = ptr_to_u64(license.c_str()), |
| 1015 | .insns = ptr_to_u64(cs[i].data.data()), |
| 1016 | .insn_cnt = static_cast<__u32>(cs[i].data.size() / sizeof(struct bpf_insn)), |
| 1017 | .log_level = 1, |
| 1018 | .log_buf = ptr_to_u64(log_buf.data()), |
| 1019 | .log_size = static_cast<__u32>(log_buf.size()), |
| 1020 | .expected_attach_type = cs[i].expected_attach_type, |
| 1021 | }; |
| 1022 | strlcpy(req.prog_name, cs[i].name.c_str(), sizeof(req.prog_name)); |
| 1023 | fd.reset(bpf(BPF_PROG_LOAD, req)); |
| 1024 | |
| 1025 | ALOGD("BPF_PROG_LOAD call for %s (%s) returned fd: %d (%s)", elfPath, |
| 1026 | cs[i].name.c_str(), fd.get(), (!fd.ok() ? std::strerror(errno) : "no error")); |
| 1027 | |
| 1028 | if (!fd.ok()) { |
| 1029 | vector<string> lines = android::base::Split(log_buf.data(), "\n"); |
| 1030 | |
| 1031 | ALOGW("BPF_PROG_LOAD - BEGIN log_buf contents:"); |
| 1032 | for (const auto& line : lines) ALOGW("%s", line.c_str()); |
| 1033 | ALOGW("BPF_PROG_LOAD - END log_buf contents."); |
| 1034 | |
| 1035 | if (cs[i].prog_def->optional) { |
| 1036 | ALOGW("failed program is marked optional - continuing..."); |
| 1037 | continue; |
| 1038 | } |
| 1039 | ALOGE("non-optional program failed to load."); |
| 1040 | } |
| 1041 | } |
| 1042 | |
| 1043 | if (!fd.ok()) return fd.get(); |
| 1044 | |
| 1045 | if (!reuse) { |
| 1046 | if (specified(selinux_context)) { |
| 1047 | string createLoc = string(BPF_FS_PATH) + lookupPinSubdir(selinux_context) + |
| 1048 | "tmp_prog_" + objName + '_' + string(name); |
| 1049 | ret = bpfFdPin(fd, createLoc.c_str()); |
| 1050 | if (ret) { |
| 1051 | int err = errno; |
| 1052 | ALOGE("create %s -> %d [%d:%s]", createLoc.c_str(), ret, err, strerror(err)); |
| 1053 | return -err; |
| 1054 | } |
| 1055 | ret = renameat2(AT_FDCWD, createLoc.c_str(), |
| 1056 | AT_FDCWD, progPinLoc.c_str(), RENAME_NOREPLACE); |
| 1057 | if (ret) { |
| 1058 | int err = errno; |
| 1059 | ALOGE("rename %s %s -> %d [%d:%s]", createLoc.c_str(), progPinLoc.c_str(), ret, |
| 1060 | err, strerror(err)); |
| 1061 | return -err; |
| 1062 | } |
| 1063 | } else { |
| 1064 | ret = bpfFdPin(fd, progPinLoc.c_str()); |
| 1065 | if (ret) { |
| 1066 | int err = errno; |
| 1067 | ALOGE("create %s -> %d [%d:%s]", progPinLoc.c_str(), ret, err, strerror(err)); |
| 1068 | return -err; |
| 1069 | } |
| 1070 | } |
| 1071 | if (chmod(progPinLoc.c_str(), 0440)) { |
| 1072 | int err = errno; |
| 1073 | ALOGE("chmod %s 0440 -> [%d:%s]", progPinLoc.c_str(), err, strerror(err)); |
| 1074 | return -err; |
| 1075 | } |
| 1076 | if (chown(progPinLoc.c_str(), (uid_t)cs[i].prog_def->uid, |
| 1077 | (gid_t)cs[i].prog_def->gid)) { |
| 1078 | int err = errno; |
| 1079 | ALOGE("chown %s %d %d -> [%d:%s]", progPinLoc.c_str(), cs[i].prog_def->uid, |
| 1080 | cs[i].prog_def->gid, err, strerror(err)); |
| 1081 | return -err; |
| 1082 | } |
| 1083 | } |
| 1084 | |
| 1085 | int progId = bpfGetFdProgId(fd); |
| 1086 | if (progId == -1) { |
| 1087 | ALOGE("bpfGetFdProgId failed, ret: %d [%d]", progId, errno); |
| 1088 | } else { |
| 1089 | ALOGI("prog %s id %d", progPinLoc.c_str(), progId); |
| 1090 | } |
| 1091 | } |
| 1092 | |
| 1093 | return 0; |
| 1094 | } |
| 1095 | |
| 1096 | int loadProg(const char* elfPath, bool* isCritical, const Location& location) { |
| 1097 | vector<char> license; |
| 1098 | vector<char> critical; |
| 1099 | vector<codeSection> cs; |
| 1100 | vector<unique_fd> mapFds; |
| 1101 | int ret; |
| 1102 | |
| 1103 | if (!isCritical) return -1; |
| 1104 | *isCritical = false; |
| 1105 | |
| 1106 | ifstream elfFile(elfPath, ios::in | ios::binary); |
| 1107 | if (!elfFile.is_open()) return -1; |
| 1108 | |
| 1109 | ret = readSectionByName("critical", elfFile, critical); |
| 1110 | *isCritical = !ret; |
| 1111 | |
| 1112 | ret = readSectionByName("license", elfFile, license); |
| 1113 | if (ret) { |
| 1114 | ALOGE("Couldn't find license in %s", elfPath); |
| 1115 | return ret; |
| 1116 | } else { |
| 1117 | ALOGD("Loading %s%s ELF object %s with license %s", |
| 1118 | *isCritical ? "critical for " : "optional", *isCritical ? (char*)critical.data() : "", |
| 1119 | elfPath, (char*)license.data()); |
| 1120 | } |
| 1121 | |
| 1122 | // the following default values are for bpfloader V0.0 format which does not include them |
| 1123 | unsigned int bpfLoaderMinVer = |
| 1124 | readSectionUint("bpfloader_min_ver", elfFile, DEFAULT_BPFLOADER_MIN_VER); |
| 1125 | unsigned int bpfLoaderMaxVer = |
| 1126 | readSectionUint("bpfloader_max_ver", elfFile, DEFAULT_BPFLOADER_MAX_VER); |
| 1127 | unsigned int bpfLoaderMinRequiredVer = |
| 1128 | readSectionUint("bpfloader_min_required_ver", elfFile, 0); |
| 1129 | size_t sizeOfBpfMapDef = |
| 1130 | readSectionUint("size_of_bpf_map_def", elfFile, DEFAULT_SIZEOF_BPF_MAP_DEF); |
| 1131 | size_t sizeOfBpfProgDef = |
| 1132 | readSectionUint("size_of_bpf_prog_def", elfFile, DEFAULT_SIZEOF_BPF_PROG_DEF); |
| 1133 | |
| 1134 | // inclusive lower bound check |
| 1135 | if (BPFLOADER_VERSION < bpfLoaderMinVer) { |
| 1136 | ALOGI("BpfLoader version 0x%05x ignoring ELF object %s with min ver 0x%05x", |
| 1137 | BPFLOADER_VERSION, elfPath, bpfLoaderMinVer); |
| 1138 | return 0; |
| 1139 | } |
| 1140 | |
| 1141 | // exclusive upper bound check |
| 1142 | if (BPFLOADER_VERSION >= bpfLoaderMaxVer) { |
| 1143 | ALOGI("BpfLoader version 0x%05x ignoring ELF object %s with max ver 0x%05x", |
| 1144 | BPFLOADER_VERSION, elfPath, bpfLoaderMaxVer); |
| 1145 | return 0; |
| 1146 | } |
| 1147 | |
| 1148 | if (BPFLOADER_VERSION < bpfLoaderMinRequiredVer) { |
| 1149 | ALOGI("BpfLoader version 0x%05x failing due to ELF object %s with required min ver 0x%05x", |
| 1150 | BPFLOADER_VERSION, elfPath, bpfLoaderMinRequiredVer); |
| 1151 | return -1; |
| 1152 | } |
| 1153 | |
| 1154 | ALOGI("BpfLoader version 0x%05x processing ELF object %s with ver [0x%05x,0x%05x)", |
| 1155 | BPFLOADER_VERSION, elfPath, bpfLoaderMinVer, bpfLoaderMaxVer); |
| 1156 | |
| 1157 | if (sizeOfBpfMapDef < DEFAULT_SIZEOF_BPF_MAP_DEF) { |
| 1158 | ALOGE("sizeof(bpf_map_def) of %zu is too small (< %d)", sizeOfBpfMapDef, |
| 1159 | DEFAULT_SIZEOF_BPF_MAP_DEF); |
| 1160 | return -1; |
| 1161 | } |
| 1162 | |
| 1163 | if (sizeOfBpfProgDef < DEFAULT_SIZEOF_BPF_PROG_DEF) { |
| 1164 | ALOGE("sizeof(bpf_prog_def) of %zu is too small (< %d)", sizeOfBpfProgDef, |
| 1165 | DEFAULT_SIZEOF_BPF_PROG_DEF); |
| 1166 | return -1; |
| 1167 | } |
| 1168 | |
Maciej Żenczykowski | fcea70f | 2023-10-08 23:35:00 -0700 | [diff] [blame] | 1169 | ret = readCodeSections(elfFile, cs, sizeOfBpfProgDef); |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 1170 | if (ret) { |
| 1171 | ALOGE("Couldn't read all code sections in %s", elfPath); |
| 1172 | return ret; |
| 1173 | } |
| 1174 | |
| 1175 | /* Just for future debugging */ |
| 1176 | if (0) dumpAllCs(cs); |
| 1177 | |
Maciej Żenczykowski | af07d6d | 2023-10-10 03:40:51 -0700 | [diff] [blame] | 1178 | ret = createMaps(elfPath, elfFile, mapFds, location.prefix, sizeOfBpfMapDef); |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 1179 | if (ret) { |
| 1180 | ALOGE("Failed to create maps: (ret=%d) in %s", ret, elfPath); |
| 1181 | return ret; |
| 1182 | } |
| 1183 | |
| 1184 | for (int i = 0; i < (int)mapFds.size(); i++) |
| 1185 | ALOGD("map_fd found at %d is %d in %s", i, mapFds[i].get(), elfPath); |
| 1186 | |
| 1187 | applyMapRelo(elfFile, mapFds, cs); |
| 1188 | |
Maciej Żenczykowski | af07d6d | 2023-10-10 03:40:51 -0700 | [diff] [blame] | 1189 | ret = loadCodeSections(elfPath, cs, string(license.data()), location.prefix); |
Maciej Żenczykowski | 60c159f | 2023-10-02 14:54:48 -0700 | [diff] [blame] | 1190 | if (ret) ALOGE("Failed to load programs, loadCodeSections ret=%d", ret); |
| 1191 | |
| 1192 | return ret; |
| 1193 | } |
| 1194 | |
| 1195 | } // namespace bpf |
| 1196 | } // namespace android |