be/src/common/phdr_cache.cpp
Line | Count | Source |
1 | | // Licensed to the Apache Software Foundation (ASF) under one |
2 | | // or more contributor license agreements. See the NOTICE file |
3 | | // distributed with this work for additional information |
4 | | // regarding copyright ownership. The ASF licenses this file |
5 | | // to you under the Apache License, Version 2.0 (the |
6 | | // "License"); you may not use this file except in compliance |
7 | | // with the License. You may obtain a copy of the License at |
8 | | // |
9 | | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | | // |
11 | | // Unless required by applicable law or agreed to in writing, |
12 | | // software distributed under the License is distributed on an |
13 | | // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
14 | | // KIND, either express or implied. See the License for the |
15 | | // specific language governing permissions and limitations |
16 | | // under the License. |
17 | | // This file is copied from |
18 | | // https://github.com/ClickHouse/ClickHouse/blob/master/src/base/phdr_cache.cpp |
19 | | // and modified by Doris |
20 | | |
21 | | #include "common/phdr_cache.h" |
22 | | |
23 | | #if defined(__clang__) |
24 | | #pragma clang diagnostic ignored "-Wreserved-identifier" |
25 | | #endif |
26 | | |
27 | | /// This code was based on the code by Fedor Korotkiy https://www.linkedin.com/in/fedor-korotkiy-659a1838/ |
28 | | |
29 | | #if defined(__linux__) && !defined(THREAD_SANITIZER) && !defined(USE_MUSL) |
30 | | #define USE_PHDR_CACHE 1 |
31 | | #endif |
32 | | |
33 | | /// Thread Sanitizer uses dl_iterate_phdr function on initialization and fails if we provide our own. |
34 | | #ifdef USE_PHDR_CACHE |
35 | | |
36 | | #if defined(__clang__) |
37 | | #pragma clang diagnostic ignored "-Wreserved-id-macro" |
38 | | #pragma clang diagnostic ignored "-Wunused-macros" |
39 | | #endif |
40 | | |
41 | | #define __msan_unpoison(X, Y) // NOLINT |
42 | | #if defined(__clang__) && defined(__has_feature) |
43 | | #if __has_feature(memory_sanitizer) |
44 | | #undef __msan_unpoison |
45 | | #include <sanitizer/msan_interface.h> |
46 | | #endif |
47 | | #endif |
48 | | |
49 | | #include <dlfcn.h> |
50 | | #include <link.h> |
51 | | |
52 | | #include <algorithm> |
53 | | #include <array> |
54 | | #include <atomic> |
55 | | #include <cstddef> |
56 | | #include <cstring> |
57 | | #include <limits> |
58 | | #include <memory> |
59 | | #include <mutex> |
60 | | #include <stdexcept> |
61 | | #include <string> |
62 | | #include <utility> |
63 | | #include <vector> |
64 | | |
65 | | #if defined(USE_UNWIND) && USE_UNWIND |
66 | | #ifndef UNW_LOCAL_ONLY |
67 | | #define UNW_LOCAL_ONLY |
68 | | #endif |
69 | | #include <libunwind.h> |
70 | | #endif |
71 | | |
72 | | namespace { |
73 | | |
74 | | // This is adapted from |
75 | | // https://github.com/scylladb/seastar/blob/master/core/exception_hacks.hh |
76 | | // https://github.com/scylladb/seastar/blob/master/core/exception_hacks.cc |
77 | | |
78 | | constexpr size_t MAX_PHDR_CACHE_LOADED_OBJECTS = 4096; |
79 | | constexpr size_t MAX_PHDR_CACHE_PROGRAM_HEADERS = 128; |
80 | | constexpr size_t MAX_PHDR_CACHE_OBJECT_NAME = 4096; |
81 | | |
82 | | using DLIterateFunction = int (*)(int (*callback)(dl_phdr_info* info, size_t size, void* data), |
83 | | void* data); |
84 | | |
85 | 69.4k | DLIterateFunction getOriginalDLIteratePHDR() { |
86 | 69.4k | void* func = dlsym(RTLD_NEXT, "dl_iterate_phdr"); |
87 | 69.4k | if (!func) { |
88 | 0 | throw std::runtime_error("Cannot find dl_iterate_phdr function with dlsym"); |
89 | 0 | } |
90 | 69.4k | return reinterpret_cast<DLIterateFunction>(func); |
91 | 69.4k | } |
92 | | |
93 | | struct RawPHDRCacheEntry { |
94 | | dl_phdr_info info {}; |
95 | | std::array<ElfW(Phdr), MAX_PHDR_CACHE_PROGRAM_HEADERS> phdrs {}; |
96 | | size_t phdr_count = 0; |
97 | | std::array<char, MAX_PHDR_CACHE_OBJECT_NAME> name {}; |
98 | | uintptr_t address_begin = std::numeric_limits<uintptr_t>::max(); |
99 | | uintptr_t address_end = 0; |
100 | | }; |
101 | | |
102 | | struct RawPHDRCacheSnapshot { |
103 | | std::array<RawPHDRCacheEntry, MAX_PHDR_CACHE_LOADED_OBJECTS> entries {}; |
104 | | size_t size = 0; |
105 | | bool overflow = false; |
106 | | bool phdr_truncated = false; |
107 | | bool name_truncated = false; |
108 | | }; |
109 | | |
110 | | struct PHDRCacheEntry { |
111 | | dl_phdr_info info {}; |
112 | | std::vector<ElfW(Phdr)> phdrs; |
113 | | std::string name; |
114 | | uintptr_t address_begin = std::numeric_limits<uintptr_t>::max(); |
115 | | uintptr_t address_end = 0; |
116 | | |
117 | 1.18k | bool contains(uintptr_t ip) const { |
118 | 1.18k | return ip == 0 || (address_begin <= ip && ip < address_end); |
119 | 1.18k | } |
120 | | }; |
121 | | |
122 | | using PHDRCache = std::vector<PHDRCacheEntry>; |
123 | | std::atomic<PHDRCache*> phdr_cache {}; |
124 | | // This flag is flipped inside the stack-trace signal handler. Force a static TLS access model so |
125 | | // reading it from our dl_iterate_phdr interposer does not call into the dynamic loader's TLS path. |
126 | | __thread bool use_phdr_cache __attribute__((tls_model("initial-exec"))) = false; |
127 | | |
128 | 331 | uintptr_t saturated_segment_end(uintptr_t begin, uintptr_t size) { |
129 | 331 | const uintptr_t max_address = std::numeric_limits<uintptr_t>::max(); |
130 | 331 | return begin > max_address - size ? max_address : begin + size; |
131 | 331 | } |
132 | | |
133 | | void copy_object_name(const char* source, RawPHDRCacheEntry* entry, |
134 | 151 | RawPHDRCacheSnapshot* snapshot) { |
135 | 151 | if (source == nullptr) { |
136 | 0 | entry->name[0] = '\0'; |
137 | 0 | return; |
138 | 0 | } |
139 | | |
140 | 151 | size_t length = 0; |
141 | 3.34k | while (length + 1 < entry->name.size() && source[length] != '\0') { |
142 | 3.19k | entry->name[length] = source[length]; |
143 | 3.19k | ++length; |
144 | 3.19k | } |
145 | 151 | entry->name[length] = '\0'; |
146 | 151 | if (source[length] != '\0') { |
147 | 0 | snapshot->name_truncated = true; |
148 | 0 | } |
149 | 151 | } |
150 | | |
151 | 151 | int collectPHDRCacheEntry(dl_phdr_info* info, size_t /*size*/, void* data) { |
152 | 151 | auto* snapshot = reinterpret_cast<RawPHDRCacheSnapshot*>(data); |
153 | 151 | if (snapshot->size >= snapshot->entries.size()) { |
154 | 0 | snapshot->overflow = true; |
155 | 0 | return 0; |
156 | 0 | } |
157 | | |
158 | 151 | auto& entry = snapshot->entries[snapshot->size++]; |
159 | 151 | entry.info = *info; |
160 | 151 | copy_object_name(info->dlpi_name, &entry, snapshot); |
161 | | |
162 | 151 | const size_t phdr_count = std::min<size_t>(info->dlpi_phnum, entry.phdrs.size()); |
163 | 151 | if (phdr_count < info->dlpi_phnum) { |
164 | 0 | snapshot->phdr_truncated = true; |
165 | 0 | } |
166 | 151 | entry.phdr_count = phdr_count; |
167 | 151 | entry.info.dlpi_phnum = static_cast<ElfW(Half)>(phdr_count); |
168 | 151 | entry.info.dlpi_name = nullptr; |
169 | 151 | entry.info.dlpi_phdr = nullptr; |
170 | | |
171 | 151 | if (info->dlpi_phdr == nullptr) { |
172 | 0 | return 0; |
173 | 0 | } |
174 | 151 | std::memcpy(entry.phdrs.data(), info->dlpi_phdr, phdr_count * sizeof(ElfW(Phdr))); |
175 | | |
176 | 1.32k | for (size_t i = 0; i < phdr_count; ++i) { |
177 | 1.17k | const auto& phdr = entry.phdrs[i]; |
178 | 1.17k | if (phdr.p_type != PT_LOAD || phdr.p_memsz == 0) { |
179 | 842 | continue; |
180 | 842 | } |
181 | 331 | const auto begin = static_cast<uintptr_t>(info->dlpi_addr + phdr.p_vaddr); |
182 | 331 | const auto end = saturated_segment_end(begin, static_cast<uintptr_t>(phdr.p_memsz)); |
183 | 331 | entry.address_begin = std::min(entry.address_begin, begin); |
184 | 331 | entry.address_end = std::max(entry.address_end, end); |
185 | 331 | } |
186 | 151 | return 0; |
187 | 151 | } |
188 | | |
189 | 11 | PHDRCache* buildPHDRCache(const RawPHDRCacheSnapshot& snapshot) { |
190 | 11 | auto* cache = new PHDRCache; |
191 | 11 | cache->reserve(snapshot.size); |
192 | 162 | for (size_t i = 0; i < snapshot.size; ++i) { |
193 | 151 | const auto& raw_entry = snapshot.entries[i]; |
194 | 151 | PHDRCacheEntry entry; |
195 | 151 | entry.info = raw_entry.info; |
196 | 151 | entry.phdrs.assign(raw_entry.phdrs.begin(), raw_entry.phdrs.begin() + raw_entry.phdr_count); |
197 | 151 | entry.name = raw_entry.name.data(); |
198 | 151 | entry.address_begin = raw_entry.address_begin; |
199 | 151 | entry.address_end = raw_entry.address_end; |
200 | 151 | cache->emplace_back(std::move(entry)); |
201 | 151 | } |
202 | 151 | for (auto& entry : *cache) { |
203 | 151 | entry.info.dlpi_phdr = entry.phdrs.data(); |
204 | 151 | entry.info.dlpi_name = entry.name.c_str(); |
205 | 151 | } |
206 | 11 | return cache; |
207 | 11 | } |
208 | | |
209 | | int iteratePHDRCache(int (*callback)(dl_phdr_info* info, size_t size, void* data), void* data, |
210 | 1.02k | uintptr_t ip) { |
211 | 1.02k | auto* current_phdr_cache = phdr_cache.load(std::memory_order_acquire); |
212 | 1.02k | if (current_phdr_cache == nullptr) { |
213 | 0 | return 0; |
214 | 0 | } |
215 | | |
216 | 1.02k | int result = 0; |
217 | 1.19k | for (auto& entry : *current_phdr_cache) { |
218 | 1.19k | if (!entry.contains(ip)) { |
219 | 137 | continue; |
220 | 137 | } |
221 | 1.05k | result = callback(&entry.info, sizeof(dl_phdr_info), data); |
222 | 1.05k | if (result != 0) { |
223 | 1.02k | break; |
224 | 1.02k | } |
225 | 1.05k | } |
226 | 1.02k | return result; |
227 | 1.02k | } |
228 | | |
229 | | } // namespace |
230 | | |
231 | | extern "C" |
232 | | #ifndef __clang__ |
233 | | [[gnu::visibility("default")]] [[gnu::externally_visible]] |
234 | | #endif |
235 | | int |
236 | 69.4k | dl_iterate_phdr(int (*callback)(dl_phdr_info* info, size_t size, void* data), void* data) { |
237 | 69.4k | if (!use_phdr_cache) { |
238 | 69.4k | return getOriginalDLIteratePHDR()(callback, data); |
239 | 69.4k | } |
240 | | |
241 | 2 | return iteratePHDRCache(callback, data, 0); |
242 | 69.4k | } |
243 | | |
244 | | extern "C" |
245 | | #ifndef __clang__ |
246 | | [[gnu::visibility("default")]] [[gnu::externally_visible]] |
247 | | #endif |
248 | | int |
249 | | doris_unwind_iterate_phdr(int (*callback)(dl_phdr_info* info, size_t size, void* data), |
250 | 1.02k | void* data, uintptr_t ip) { |
251 | 1.02k | return iteratePHDRCache(callback, data, ip); |
252 | 1.02k | } |
253 | | |
254 | | #include "util/debug/leak_annotations.h" |
255 | | |
256 | 11 | void updatePHDRCache() { |
257 | | // Fill out ELF header cache for access without locking. |
258 | | // Old snapshots are intentionally kept alive because another thread may already be unwinding |
259 | | // through the previous cache when a Doris-controlled dlopen/dlclose refreshes this one. |
260 | | |
261 | 11 | auto raw_snapshot = std::make_unique<RawPHDRCacheSnapshot>(); |
262 | 11 | getOriginalDLIteratePHDR()( |
263 | 151 | [](dl_phdr_info* info, size_t size, void* data) { |
264 | | // `info` is created by dl_iterate_phdr, which is a non-instrumented |
265 | | // libc function, so we have to unpoison it manually. |
266 | 151 | __msan_unpoison(info, sizeof(*info)); |
267 | | |
268 | 151 | return collectPHDRCacheEntry(info, size, data); |
269 | 151 | }, |
270 | 11 | raw_snapshot.get()); |
271 | | |
272 | 11 | PHDRCache* new_phdr_cache = buildPHDRCache(*raw_snapshot); |
273 | 11 | phdr_cache.store(new_phdr_cache, std::memory_order_release); |
274 | | |
275 | | /// Memory is intentionally leaked. |
276 | 11 | ANNOTATE_LEAKING_OBJECT_PTR(new_phdr_cache); |
277 | 11 | } |
278 | | |
279 | 21 | bool hasPHDRCache() { |
280 | 21 | return phdr_cache.load(std::memory_order_acquire) != nullptr; |
281 | 21 | } |
282 | | |
283 | 1 | void configureLibunwindPHDRCache() { |
284 | 1 | #if defined(USE_UNWIND) && USE_UNWIND |
285 | 1 | static std::once_flag once; |
286 | 1 | std::call_once(once, [] { |
287 | 1 | unw_context_t context; |
288 | 1 | unw_cursor_t cursor; |
289 | 1 | (void)unw_getcontext(&context); |
290 | 1 | (void)unw_init_local(&cursor, &context); |
291 | | // Doris-patched libunwind gets FDEs from the PHDR snapshot. Disable the global DWARF |
292 | | // register-state cache so a signal handler cannot self-deadlock on libunwind's cache mutex |
293 | | // after interrupting a thread that was already unwinding. |
294 | 1 | (void)unw_set_caching_policy(unw_local_addr_space, UNW_CACHE_NONE); |
295 | 1 | }); |
296 | 1 | #endif |
297 | 1 | } |
298 | | |
299 | 7 | ScopedPHDRCacheRead::ScopedPHDRCacheRead() : _previous(use_phdr_cache) { |
300 | 7 | use_phdr_cache = true; |
301 | 7 | } |
302 | | |
303 | 7 | ScopedPHDRCacheRead::~ScopedPHDRCacheRead() { |
304 | 7 | use_phdr_cache = _previous; |
305 | 7 | } |
306 | | |
307 | | #else |
308 | | |
309 | | void updatePHDRCache() {} |
310 | | |
311 | | void configureLibunwindPHDRCache() {} |
312 | | |
313 | | #if defined(USE_MUSL) |
314 | | /// With statically linked with musl, dl_iterate_phdr is immutable. |
315 | | bool hasPHDRCache() { |
316 | | return true; |
317 | | } |
318 | | #else |
319 | | bool hasPHDRCache() { |
320 | | return false; |
321 | | } |
322 | | #endif |
323 | | |
324 | | ScopedPHDRCacheRead::ScopedPHDRCacheRead() = default; |
325 | | |
326 | | ScopedPHDRCacheRead::~ScopedPHDRCacheRead() = default; |
327 | | |
328 | | #endif |