| 1 |
//===-- sanitizer_deadlock_detector.h ---------------------------*- C++ -*-===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
// |
| 9 |
// This file is a part of Sanitizer runtime. |
| 10 |
// The deadlock detector maintains a directed graph of lock acquisitions. |
| 11 |
// When a lock event happens, the detector checks if the locks already held by |
| 12 |
// the current thread are reachable from the newly acquired lock. |
| 13 |
// |
| 14 |
// The detector can handle only a fixed amount of simultaneously live locks |
| 15 |
// (a lock is alive if it has been locked at least once and has not been |
| 16 |
// destroyed). When the maximal number of locks is reached the entire graph |
| 17 |
// is flushed and the new lock epoch is started. The node ids from the old |
| 18 |
// epochs can not be used with any of the detector methods except for |
| 19 |
// nodeBelongsToCurrentEpoch(). |
| 20 |
// |
| 21 |
// FIXME: this is work in progress, nothing really works yet. |
| 22 |
// |
| 23 |
//===----------------------------------------------------------------------===// |
| 24 |
|
| 25 |
#ifndef SANITIZER_DEADLOCK_DETECTOR_H |
| 26 |
#define SANITIZER_DEADLOCK_DETECTOR_H |
| 27 |
|
| 28 |
#include "sanitizer_bvgraph.h" |
| 29 |
#include "sanitizer_common.h" |
| 30 |
|
| 31 |
namespace __sanitizer { |
| 32 |
|
| 33 |
// Thread-local state for DeadlockDetector. |
| 34 |
// It contains the locks currently held by the owning thread. |
| 35 |
template <class BV> |
| 36 |
class DeadlockDetectorTLS { |
| 37 |
public: |
| 38 |
// No CTOR. |
| 39 |
void clear() { |
| 40 |
bv_.clear(); |
| 41 |
epoch_ = 0; |
| 42 |
n_recursive_locks = 0; |
| 43 |
n_all_locks_ = 0; |
| 44 |
} |
| 45 |
|
| 46 |
bool empty() const { return bv_.empty(); } |
| 47 |
|
| 48 |
void ensureCurrentEpoch(uptr current_epoch) { |
| 49 |
if (epoch_ == current_epoch) return; |
| 50 |
bv_.clear(); |
| 51 |
epoch_ = current_epoch; |
| 52 |
n_recursive_locks = 0; |
| 53 |
n_all_locks_ = 0; |
| 54 |
} |
| 55 |
|
| 56 |
uptr getEpoch() const { return epoch_; } |
| 57 |
|
| 58 |
// Returns true if this is the first (non-recursive) acquisition of this lock. |
| 59 |
bool addLock(uptr lock_id, uptr current_epoch, u32 stk) { |
| 60 |
CHECK_EQ(epoch_, current_epoch); |
| 61 |
if (!bv_.setBit(lock_id)) { |
| 62 |
// The lock is already held by this thread, it must be recursive. |
| 63 |
CHECK_LT(n_recursive_locks, ARRAY_SIZE(recursive_locks)); |
| 64 |
recursive_locks[n_recursive_locks++] = lock_id; |
| 65 |
return false; |
| 66 |
} |
| 67 |
CHECK_LT(n_all_locks_, ARRAY_SIZE(all_locks_with_contexts_)); |
| 68 |
// lock_id < BV::kSize, can cast to a smaller int. |
| 69 |
u32 lock_id_short = static_cast<u32>(lock_id); |
| 70 |
LockWithContext l = {lock_id_short, stk}; |
| 71 |
all_locks_with_contexts_[n_all_locks_++] = l; |
| 72 |
return true; |
| 73 |
} |
| 74 |
|
| 75 |
void removeLock(uptr lock_id) { |
| 76 |
if (n_recursive_locks) { |
| 77 |
for (sptr i = n_recursive_locks - 1; i >= 0; i--) { |
| 78 |
if (recursive_locks[i] == lock_id) { |
| 79 |
n_recursive_locks--; |
| 80 |
Swap(recursive_locks[i], recursive_locks[n_recursive_locks]); |
| 81 |
return; |
| 82 |
} |
| 83 |
} |
| 84 |
} |
| 85 |
if (!bv_.clearBit(lock_id)) |
| 86 |
return; // probably addLock happened before flush |
| 87 |
if (n_all_locks_) { |
| 88 |
for (sptr i = n_all_locks_ - 1; i >= 0; i--) { |
| 89 |
if (all_locks_with_contexts_[i].lock == static_cast<u32>(lock_id)) { |
| 90 |
Swap(all_locks_with_contexts_[i], |
| 91 |
all_locks_with_contexts_[n_all_locks_ - 1]); |
| 92 |
n_all_locks_--; |
| 93 |
break; |
| 94 |
} |
| 95 |
} |
| 96 |
} |
| 97 |
} |
| 98 |
|
| 99 |
u32 findLockContext(uptr lock_id) { |
| 100 |
for (uptr i = 0; i < n_all_locks_; i++) |
| 101 |
if (all_locks_with_contexts_[i].lock == static_cast<u32>(lock_id)) |
| 102 |
return all_locks_with_contexts_[i].stk; |
| 103 |
return 0; |
| 104 |
} |
| 105 |
|
| 106 |
const BV &getLocks(uptr current_epoch) const { |
| 107 |
CHECK_EQ(epoch_, current_epoch); |
| 108 |
return bv_; |
| 109 |
} |
| 110 |
|
| 111 |
uptr getNumLocks() const { return n_all_locks_; } |
| 112 |
uptr getLock(uptr idx) const { return all_locks_with_contexts_[idx].lock; } |
| 113 |
|
| 114 |
private: |
| 115 |
BV bv_; |
| 116 |
uptr epoch_; |
| 117 |
uptr recursive_locks[64]; |
| 118 |
uptr n_recursive_locks; |
| 119 |
struct LockWithContext { |
| 120 |
u32 lock; |
| 121 |
u32 stk; |
| 122 |
}; |
| 123 |
LockWithContext all_locks_with_contexts_[64]; |
| 124 |
uptr n_all_locks_; |
| 125 |
}; |
| 126 |
|
| 127 |
// DeadlockDetector. |
| 128 |
// For deadlock detection to work we need one global DeadlockDetector object |
| 129 |
// and one DeadlockDetectorTLS object per evey thread. |
| 130 |
// This class is not thread safe, all concurrent accesses should be guarded |
| 131 |
// by an external lock. |
| 132 |
// Most of the methods of this class are not thread-safe (i.e. should |
| 133 |
// be protected by an external lock) unless explicitly told otherwise. |
| 134 |
template <class BV> |
| 135 |
class DeadlockDetector { |
| 136 |
public: |
| 137 |
typedef BV BitVector; |
| 138 |
|
| 139 |
uptr size() const { return g_.size(); } |
| 140 |
|
| 141 |
// No CTOR. |
| 142 |
void clear() { |
| 143 |
current_epoch_ = 0; |
| 144 |
available_nodes_.clear(); |
| 145 |
recycled_nodes_.clear(); |
| 146 |
g_.clear(); |
| 147 |
n_edges_ = 0; |
| 148 |
} |
| 149 |
|
| 150 |
// Allocate new deadlock detector node. |
| 151 |
// If we are out of available nodes first try to recycle some. |
| 152 |
// If there is nothing to recycle, flush the graph and increment the epoch. |
| 153 |
// Associate 'data' (opaque user's object) with the new node. |
| 154 |
uptr newNode(uptr data) { |
| 155 |
if (!available_nodes_.empty()) |
| 156 |
return getAvailableNode(data); |
| 157 |
if (!recycled_nodes_.empty()) { |
| 158 |
for (sptr i = n_edges_ - 1; i >= 0; i--) { |
| 159 |
if (recycled_nodes_.getBit(edges_[i].from) || |
| 160 |
recycled_nodes_.getBit(edges_[i].to)) { |
| 161 |
Swap(edges_[i], edges_[n_edges_ - 1]); |
| 162 |
n_edges_--; |
| 163 |
} |
| 164 |
} |
| 165 |
CHECK(available_nodes_.empty()); |
| 166 |
// removeEdgesFrom was called in removeNode. |
| 167 |
g_.removeEdgesTo(recycled_nodes_); |
| 168 |
available_nodes_.setUnion(recycled_nodes_); |
| 169 |
recycled_nodes_.clear(); |
| 170 |
return getAvailableNode(data); |
| 171 |
} |
| 172 |
// We are out of vacant nodes. Flush and increment the current_epoch_. |
| 173 |
current_epoch_ += size(); |
| 174 |
recycled_nodes_.clear(); |
| 175 |
available_nodes_.setAll(); |
| 176 |
g_.clear(); |
| 177 |
n_edges_ = 0; |
| 178 |
return getAvailableNode(data); |
| 179 |
} |
| 180 |
|
| 181 |
// Get data associated with the node created by newNode(). |
| 182 |
uptr getData(uptr node) const { return data_[nodeToIndex(node)]; } |
| 183 |
|
| 184 |
bool nodeBelongsToCurrentEpoch(uptr node) { |
| 185 |
return node && (node / size() * size()) == current_epoch_; |
| 186 |
} |
| 187 |
|
| 188 |
void removeNode(uptr node) { |
| 189 |
uptr idx = nodeToIndex(node); |
| 190 |
CHECK(!available_nodes_.getBit(idx)); |
| 191 |
CHECK(recycled_nodes_.setBit(idx)); |
| 192 |
g_.removeEdgesFrom(idx); |
| 193 |
} |
| 194 |
|
| 195 |
void ensureCurrentEpoch(DeadlockDetectorTLS<BV> *dtls) { |
| 196 |
dtls->ensureCurrentEpoch(current_epoch_); |
| 197 |
} |
| 198 |
|
| 199 |
// Returns true if there is a cycle in the graph after this lock event. |
| 200 |
// Ideally should be called before the lock is acquired so that we can |
| 201 |
// report a deadlock before a real deadlock happens. |
| 202 |
bool onLockBefore(DeadlockDetectorTLS<BV> *dtls, uptr cur_node) { |
| 203 |
ensureCurrentEpoch(dtls); |
| 204 |
uptr cur_idx = nodeToIndex(cur_node); |
| 205 |
return g_.isReachable(cur_idx, dtls->getLocks(current_epoch_)); |
| 206 |
} |
| 207 |
|
| 208 |
u32 findLockContext(DeadlockDetectorTLS<BV> *dtls, uptr node) { |
| 209 |
return dtls->findLockContext(nodeToIndex(node)); |
| 210 |
} |
| 211 |
|
| 212 |
// Add cur_node to the set of locks held currently by dtls. |
| 213 |
void onLockAfter(DeadlockDetectorTLS<BV> *dtls, uptr cur_node, u32 stk = 0) { |
| 214 |
ensureCurrentEpoch(dtls); |
| 215 |
uptr cur_idx = nodeToIndex(cur_node); |
| 216 |
dtls->addLock(cur_idx, current_epoch_, stk); |
| 217 |
} |
| 218 |
|
| 219 |
// Experimental *racy* fast path function. |
| 220 |
// Returns true if all edges from the currently held locks to cur_node exist. |
| 221 |
bool hasAllEdges(DeadlockDetectorTLS<BV> *dtls, uptr cur_node) { |
| 222 |
uptr local_epoch = dtls->getEpoch(); |
| 223 |
// Read from current_epoch_ is racy. |
| 224 |
if (cur_node && local_epoch == current_epoch_ && |
| 225 |
local_epoch == nodeToEpoch(cur_node)) { |
| 226 |
uptr cur_idx = nodeToIndexUnchecked(cur_node); |
| 227 |
for (uptr i = 0, n = dtls->getNumLocks(); i < n; i++) { |
| 228 |
if (!g_.hasEdge(dtls->getLock(i), cur_idx)) |
| 229 |
return false; |
| 230 |
} |
| 231 |
return true; |
| 232 |
} |
| 233 |
return false; |
| 234 |
} |
| 235 |
|
| 236 |
// Adds edges from currently held locks to cur_node, |
| 237 |
// returns the number of added edges, and puts the sources of added edges |
| 238 |
// into added_edges[]. |
| 239 |
// Should be called before onLockAfter. |
| 240 |
uptr addEdges(DeadlockDetectorTLS<BV> *dtls, uptr cur_node, u32 stk, |
| 241 |
int unique_tid) { |
| 242 |
ensureCurrentEpoch(dtls); |
| 243 |
uptr cur_idx = nodeToIndex(cur_node); |
| 244 |
uptr added_edges[40]; |
| 245 |
uptr n_added_edges = g_.addEdges(dtls->getLocks(current_epoch_), cur_idx, |
| 246 |
added_edges, ARRAY_SIZE(added_edges)); |
| 247 |
for (uptr i = 0; i < n_added_edges; i++) { |
| 248 |
if (n_edges_ < ARRAY_SIZE(edges_)) { |
| 249 |
Edge e = {(u16)added_edges[i], (u16)cur_idx, |
| 250 |
dtls->findLockContext(added_edges[i]), stk, |
| 251 |
unique_tid}; |
| 252 |
edges_[n_edges_++] = e; |
| 253 |
} |
| 254 |
} |
| 255 |
return n_added_edges; |
| 256 |
} |
| 257 |
|
| 258 |
bool findEdge(uptr from_node, uptr to_node, u32 *stk_from, u32 *stk_to, |
| 259 |
int *unique_tid) { |
| 260 |
uptr from_idx = nodeToIndex(from_node); |
| 261 |
uptr to_idx = nodeToIndex(to_node); |
| 262 |
for (uptr i = 0; i < n_edges_; i++) { |
| 263 |
if (edges_[i].from == from_idx && edges_[i].to == to_idx) { |
| 264 |
*stk_from = edges_[i].stk_from; |
| 265 |
*stk_to = edges_[i].stk_to; |
| 266 |
*unique_tid = edges_[i].unique_tid; |
| 267 |
return true; |
| 268 |
} |
| 269 |
} |
| 270 |
return false; |
| 271 |
} |
| 272 |
|
| 273 |
// Test-only function. Handles the before/after lock events, |
| 274 |
// returns true if there is a cycle. |
| 275 |
bool onLock(DeadlockDetectorTLS<BV> *dtls, uptr cur_node, u32 stk = 0) { |
| 276 |
ensureCurrentEpoch(dtls); |
| 277 |
bool is_reachable = !isHeld(dtls, cur_node) && onLockBefore(dtls, cur_node); |
| 278 |
addEdges(dtls, cur_node, stk, 0); |
| 279 |
onLockAfter(dtls, cur_node, stk); |
| 280 |
return is_reachable; |
| 281 |
} |
| 282 |
|
| 283 |
// Handles the try_lock event, returns false. |
| 284 |
// When a try_lock event happens (i.e. a try_lock call succeeds) we need |
| 285 |
// to add this lock to the currently held locks, but we should not try to |
| 286 |
// change the lock graph or to detect a cycle. We may want to investigate |
| 287 |
// whether a more aggressive strategy is possible for try_lock. |
| 288 |
bool onTryLock(DeadlockDetectorTLS<BV> *dtls, uptr cur_node, u32 stk = 0) { |
| 289 |
ensureCurrentEpoch(dtls); |
| 290 |
uptr cur_idx = nodeToIndex(cur_node); |
| 291 |
dtls->addLock(cur_idx, current_epoch_, stk); |
| 292 |
return false; |
| 293 |
} |
| 294 |
|
| 295 |
// Returns true iff dtls is empty (no locks are currently held) and we can |
| 296 |
// add the node to the currently held locks w/o changing the global state. |
| 297 |
// This operation is thread-safe as it only touches the dtls. |
| 298 |
bool onFirstLock(DeadlockDetectorTLS<BV> *dtls, uptr node, u32 stk = 0) { |
| 299 |
if (!dtls->empty()) return false; |
| 300 |
if (dtls->getEpoch() && dtls->getEpoch() == nodeToEpoch(node)) { |
| 301 |
dtls->addLock(nodeToIndexUnchecked(node), nodeToEpoch(node), stk); |
| 302 |
return true; |
| 303 |
} |
| 304 |
return false; |
| 305 |
} |
| 306 |
|
| 307 |
// Finds a path between the lock 'cur_node' (currently not held in dtls) |
| 308 |
// and some currently held lock, returns the length of the path |
| 309 |
// or 0 on failure. |
| 310 |
uptr findPathToLock(DeadlockDetectorTLS<BV> *dtls, uptr cur_node, uptr *path, |
| 311 |
uptr path_size) { |
| 312 |
tmp_bv_.copyFrom(dtls->getLocks(current_epoch_)); |
| 313 |
uptr idx = nodeToIndex(cur_node); |
| 314 |
CHECK(!tmp_bv_.getBit(idx)); |
| 315 |
uptr res = g_.findShortestPath(idx, tmp_bv_, path, path_size); |
| 316 |
for (uptr i = 0; i < res; i++) |
| 317 |
path[i] = indexToNode(path[i]); |
| 318 |
if (res) |
| 319 |
CHECK_EQ(path[0], cur_node); |
| 320 |
return res; |
| 321 |
} |
| 322 |
|
| 323 |
// Handle the unlock event. |
| 324 |
// This operation is thread-safe as it only touches the dtls. |
| 325 |
void onUnlock(DeadlockDetectorTLS<BV> *dtls, uptr node) { |
| 326 |
if (dtls->getEpoch() == nodeToEpoch(node)) |
| 327 |
dtls->removeLock(nodeToIndexUnchecked(node)); |
| 328 |
} |
| 329 |
|
| 330 |
// Tries to handle the lock event w/o writing to global state. |
| 331 |
// Returns true on success. |
| 332 |
// This operation is thread-safe as it only touches the dtls |
| 333 |
// (modulo racy nature of hasAllEdges). |
| 334 |
bool onLockFast(DeadlockDetectorTLS<BV> *dtls, uptr node, u32 stk = 0) { |
| 335 |
if (hasAllEdges(dtls, node)) { |
| 336 |
dtls->addLock(nodeToIndexUnchecked(node), nodeToEpoch(node), stk); |
| 337 |
return true; |
| 338 |
} |
| 339 |
return false; |
| 340 |
} |
| 341 |
|
| 342 |
bool isHeld(DeadlockDetectorTLS<BV> *dtls, uptr node) const { |
| 343 |
return dtls->getLocks(current_epoch_).getBit(nodeToIndex(node)); |
| 344 |
} |
| 345 |
|
| 346 |
uptr testOnlyGetEpoch() const { return current_epoch_; } |
| 347 |
bool testOnlyHasEdge(uptr l1, uptr l2) { |
| 348 |
return g_.hasEdge(nodeToIndex(l1), nodeToIndex(l2)); |
| 349 |
} |
| 350 |
// idx1 and idx2 are raw indices to g_, not lock IDs. |
| 351 |
bool testOnlyHasEdgeRaw(uptr idx1, uptr idx2) { |
| 352 |
return g_.hasEdge(idx1, idx2); |
| 353 |
} |
| 354 |
|
| 355 |
void Print() { |
| 356 |
for (uptr from = 0; from < size(); from++) |
| 357 |
for (uptr to = 0; to < size(); to++) |
| 358 |
if (g_.hasEdge(from, to)) |
| 359 |
Printf(" %zx => %zx\n", from, to); |
| 360 |
} |
| 361 |
|
| 362 |
private: |
| 363 |
void check_idx(uptr idx) const { CHECK_LT(idx, size()); } |
| 364 |
|
| 365 |
void check_node(uptr node) const { |
| 366 |
CHECK_GE(node, size()); |
| 367 |
CHECK_EQ(current_epoch_, nodeToEpoch(node)); |
| 368 |
} |
| 369 |
|
| 370 |
uptr indexToNode(uptr idx) const { |
| 371 |
check_idx(idx); |
| 372 |
return idx + current_epoch_; |
| 373 |
} |
| 374 |
|
| 375 |
uptr nodeToIndexUnchecked(uptr node) const { return node % size(); } |
| 376 |
|
| 377 |
uptr nodeToIndex(uptr node) const { |
| 378 |
check_node(node); |
| 379 |
return nodeToIndexUnchecked(node); |
| 380 |
} |
| 381 |
|
| 382 |
uptr nodeToEpoch(uptr node) const { return node / size() * size(); } |
| 383 |
|
| 384 |
uptr getAvailableNode(uptr data) { |
| 385 |
uptr idx = available_nodes_.getAndClearFirstOne(); |
| 386 |
data_[idx] = data; |
| 387 |
return indexToNode(idx); |
| 388 |
} |
| 389 |
|
| 390 |
struct Edge { |
| 391 |
u16 from; |
| 392 |
u16 to; |
| 393 |
u32 stk_from; |
| 394 |
u32 stk_to; |
| 395 |
int unique_tid; |
| 396 |
}; |
| 397 |
|
| 398 |
uptr current_epoch_; |
| 399 |
BV available_nodes_; |
| 400 |
BV recycled_nodes_; |
| 401 |
BV tmp_bv_; |
| 402 |
BVGraph<BV> g_; |
| 403 |
uptr data_[BV::kSize]; |
| 404 |
Edge edges_[BV::kSize * 32]; |
| 405 |
uptr n_edges_; |
| 406 |
}; |
| 407 |
|
| 408 |
} // namespace __sanitizer |
| 409 |
|
| 410 |
#endif // SANITIZER_DEADLOCK_DETECTOR_H |