Byungchul Park | ef0758d | 2017-08-07 16:13:01 +0900 | [diff] [blame] | 1 | Crossrelease |
| 2 | ============ |
| 3 | |
| 4 | Started by Byungchul Park <byungchul.park@lge.com> |
| 5 | |
| 6 | Contents: |
| 7 | |
| 8 | (*) Background |
| 9 | |
| 10 | - What causes deadlock |
| 11 | - How lockdep works |
| 12 | |
| 13 | (*) Limitation |
| 14 | |
| 15 | - Limit lockdep |
| 16 | - Pros from the limitation |
| 17 | - Cons from the limitation |
| 18 | - Relax the limitation |
| 19 | |
| 20 | (*) Crossrelease |
| 21 | |
| 22 | - Introduce crossrelease |
| 23 | - Introduce commit |
| 24 | |
| 25 | (*) Implementation |
| 26 | |
| 27 | - Data structures |
| 28 | - How crossrelease works |
| 29 | |
| 30 | (*) Optimizations |
| 31 | |
| 32 | - Avoid duplication |
| 33 | - Lockless for hot paths |
| 34 | |
| 35 | (*) APPENDIX A: What lockdep does to work aggresively |
| 36 | |
| 37 | (*) APPENDIX B: How to avoid adding false dependencies |
| 38 | |
| 39 | |
| 40 | ========== |
| 41 | Background |
| 42 | ========== |
| 43 | |
| 44 | What causes deadlock |
| 45 | -------------------- |
| 46 | |
| 47 | A deadlock occurs when a context is waiting for an event to happen, |
| 48 | which is impossible because another (or the) context who can trigger the |
| 49 | event is also waiting for another (or the) event to happen, which is |
| 50 | also impossible due to the same reason. |
| 51 | |
| 52 | For example: |
| 53 | |
| 54 | A context going to trigger event C is waiting for event A to happen. |
| 55 | A context going to trigger event A is waiting for event B to happen. |
| 56 | A context going to trigger event B is waiting for event C to happen. |
| 57 | |
| 58 | A deadlock occurs when these three wait operations run at the same time, |
| 59 | because event C cannot be triggered if event A does not happen, which in |
| 60 | turn cannot be triggered if event B does not happen, which in turn |
| 61 | cannot be triggered if event C does not happen. After all, no event can |
| 62 | be triggered since any of them never meets its condition to wake up. |
| 63 | |
| 64 | A dependency might exist between two waiters and a deadlock might happen |
| 65 | due to an incorrect releationship between dependencies. Thus, we must |
| 66 | define what a dependency is first. A dependency exists between them if: |
| 67 | |
| 68 | 1. There are two waiters waiting for each event at a given time. |
| 69 | 2. The only way to wake up each waiter is to trigger its event. |
| 70 | 3. Whether one can be woken up depends on whether the other can. |
| 71 | |
| 72 | Each wait in the example creates its dependency like: |
| 73 | |
| 74 | Event C depends on event A. |
| 75 | Event A depends on event B. |
| 76 | Event B depends on event C. |
| 77 | |
| 78 | NOTE: Precisely speaking, a dependency is one between whether a |
| 79 | waiter for an event can be woken up and whether another waiter for |
| 80 | another event can be woken up. However from now on, we will describe |
| 81 | a dependency as if it's one between an event and another event for |
| 82 | simplicity. |
| 83 | |
| 84 | And they form circular dependencies like: |
| 85 | |
| 86 | -> C -> A -> B - |
| 87 | / \ |
| 88 | \ / |
| 89 | ---------------- |
| 90 | |
| 91 | where 'A -> B' means that event A depends on event B. |
| 92 | |
| 93 | Such circular dependencies lead to a deadlock since no waiter can meet |
| 94 | its condition to wake up as described. |
| 95 | |
| 96 | CONCLUSION |
| 97 | |
| 98 | Circular dependencies cause a deadlock. |
| 99 | |
| 100 | |
| 101 | How lockdep works |
| 102 | ----------------- |
| 103 | |
| 104 | Lockdep tries to detect a deadlock by checking dependencies created by |
| 105 | lock operations, acquire and release. Waiting for a lock corresponds to |
| 106 | waiting for an event, and releasing a lock corresponds to triggering an |
| 107 | event in the previous section. |
| 108 | |
| 109 | In short, lockdep does: |
| 110 | |
| 111 | 1. Detect a new dependency. |
| 112 | 2. Add the dependency into a global graph. |
| 113 | 3. Check if that makes dependencies circular. |
| 114 | 4. Report a deadlock or its possibility if so. |
| 115 | |
| 116 | For example, consider a graph built by lockdep that looks like: |
| 117 | |
| 118 | A -> B - |
| 119 | \ |
| 120 | -> E |
| 121 | / |
| 122 | C -> D - |
| 123 | |
| 124 | where A, B,..., E are different lock classes. |
| 125 | |
| 126 | Lockdep will add a dependency into the graph on detection of a new |
| 127 | dependency. For example, it will add a dependency 'E -> C' when a new |
| 128 | dependency between lock E and lock C is detected. Then the graph will be: |
| 129 | |
| 130 | A -> B - |
| 131 | \ |
| 132 | -> E - |
| 133 | / \ |
| 134 | -> C -> D - \ |
| 135 | / / |
| 136 | \ / |
| 137 | ------------------ |
| 138 | |
| 139 | where A, B,..., E are different lock classes. |
| 140 | |
| 141 | This graph contains a subgraph which demonstrates circular dependencies: |
| 142 | |
| 143 | -> E - |
| 144 | / \ |
| 145 | -> C -> D - \ |
| 146 | / / |
| 147 | \ / |
| 148 | ------------------ |
| 149 | |
| 150 | where C, D and E are different lock classes. |
| 151 | |
| 152 | This is the condition under which a deadlock might occur. Lockdep |
| 153 | reports it on detection after adding a new dependency. This is the way |
| 154 | how lockdep works. |
| 155 | |
| 156 | CONCLUSION |
| 157 | |
| 158 | Lockdep detects a deadlock or its possibility by checking if circular |
| 159 | dependencies were created after adding each new dependency. |
| 160 | |
| 161 | |
| 162 | ========== |
| 163 | Limitation |
| 164 | ========== |
| 165 | |
| 166 | Limit lockdep |
| 167 | ------------- |
| 168 | |
| 169 | Limiting lockdep to work on only typical locks e.g. spin locks and |
| 170 | mutexes, which are released within the acquire context, the |
| 171 | implementation becomes simple but its capacity for detection becomes |
| 172 | limited. Let's check pros and cons in next section. |
| 173 | |
| 174 | |
| 175 | Pros from the limitation |
| 176 | ------------------------ |
| 177 | |
| 178 | Given the limitation, when acquiring a lock, locks in a held_locks |
| 179 | cannot be released if the context cannot acquire it so has to wait to |
| 180 | acquire it, which means all waiters for the locks in the held_locks are |
| 181 | stuck. It's an exact case to create dependencies between each lock in |
| 182 | the held_locks and the lock to acquire. |
| 183 | |
| 184 | For example: |
| 185 | |
| 186 | CONTEXT X |
| 187 | --------- |
| 188 | acquire A |
| 189 | acquire B /* Add a dependency 'A -> B' */ |
| 190 | release B |
| 191 | release A |
| 192 | |
| 193 | where A and B are different lock classes. |
| 194 | |
| 195 | When acquiring lock A, the held_locks of CONTEXT X is empty thus no |
| 196 | dependency is added. But when acquiring lock B, lockdep detects and adds |
| 197 | a new dependency 'A -> B' between lock A in the held_locks and lock B. |
| 198 | They can be simply added whenever acquiring each lock. |
| 199 | |
| 200 | And data required by lockdep exists in a local structure, held_locks |
| 201 | embedded in task_struct. Forcing to access the data within the context, |
| 202 | lockdep can avoid racy problems without explicit locks while handling |
| 203 | the local data. |
| 204 | |
| 205 | Lastly, lockdep only needs to keep locks currently being held, to build |
| 206 | a dependency graph. However, relaxing the limitation, it needs to keep |
| 207 | even locks already released, because a decision whether they created |
| 208 | dependencies might be long-deferred. |
| 209 | |
| 210 | To sum up, we can expect several advantages from the limitation: |
| 211 | |
| 212 | 1. Lockdep can easily identify a dependency when acquiring a lock. |
| 213 | 2. Races are avoidable while accessing local locks in a held_locks. |
| 214 | 3. Lockdep only needs to keep locks currently being held. |
| 215 | |
| 216 | CONCLUSION |
| 217 | |
| 218 | Given the limitation, the implementation becomes simple and efficient. |
| 219 | |
| 220 | |
| 221 | Cons from the limitation |
| 222 | ------------------------ |
| 223 | |
| 224 | Given the limitation, lockdep is applicable only to typical locks. For |
| 225 | example, page locks for page access or completions for synchronization |
| 226 | cannot work with lockdep. |
| 227 | |
| 228 | Can we detect deadlocks below, under the limitation? |
| 229 | |
| 230 | Example 1: |
| 231 | |
| 232 | CONTEXT X CONTEXT Y CONTEXT Z |
| 233 | --------- --------- ---------- |
| 234 | mutex_lock A |
| 235 | lock_page B |
| 236 | lock_page B |
| 237 | mutex_lock A /* DEADLOCK */ |
| 238 | unlock_page B held by X |
| 239 | unlock_page B |
| 240 | mutex_unlock A |
| 241 | mutex_unlock A |
| 242 | |
| 243 | where A and B are different lock classes. |
| 244 | |
| 245 | No, we cannot. |
| 246 | |
| 247 | Example 2: |
| 248 | |
| 249 | CONTEXT X CONTEXT Y |
| 250 | --------- --------- |
| 251 | mutex_lock A |
| 252 | mutex_lock A |
| 253 | wait_for_complete B /* DEADLOCK */ |
| 254 | complete B |
| 255 | mutex_unlock A |
| 256 | mutex_unlock A |
| 257 | |
| 258 | where A is a lock class and B is a completion variable. |
| 259 | |
| 260 | No, we cannot. |
| 261 | |
| 262 | CONCLUSION |
| 263 | |
| 264 | Given the limitation, lockdep cannot detect a deadlock or its |
| 265 | possibility caused by page locks or completions. |
| 266 | |
| 267 | |
| 268 | Relax the limitation |
| 269 | -------------------- |
| 270 | |
| 271 | Under the limitation, things to create dependencies are limited to |
| 272 | typical locks. However, synchronization primitives like page locks and |
| 273 | completions, which are allowed to be released in any context, also |
| 274 | create dependencies and can cause a deadlock. So lockdep should track |
| 275 | these locks to do a better job. We have to relax the limitation for |
| 276 | these locks to work with lockdep. |
| 277 | |
| 278 | Detecting dependencies is very important for lockdep to work because |
| 279 | adding a dependency means adding an opportunity to check whether it |
| 280 | causes a deadlock. The more lockdep adds dependencies, the more it |
| 281 | thoroughly works. Thus Lockdep has to do its best to detect and add as |
| 282 | many true dependencies into a graph as possible. |
| 283 | |
| 284 | For example, considering only typical locks, lockdep builds a graph like: |
| 285 | |
| 286 | A -> B - |
| 287 | \ |
| 288 | -> E |
| 289 | / |
| 290 | C -> D - |
| 291 | |
| 292 | where A, B,..., E are different lock classes. |
| 293 | |
| 294 | On the other hand, under the relaxation, additional dependencies might |
| 295 | be created and added. Assuming additional 'FX -> C' and 'E -> GX' are |
| 296 | added thanks to the relaxation, the graph will be: |
| 297 | |
| 298 | A -> B - |
| 299 | \ |
| 300 | -> E -> GX |
| 301 | / |
| 302 | FX -> C -> D - |
| 303 | |
| 304 | where A, B,..., E, FX and GX are different lock classes, and a suffix |
| 305 | 'X' is added on non-typical locks. |
| 306 | |
| 307 | The latter graph gives us more chances to check circular dependencies |
| 308 | than the former. However, it might suffer performance degradation since |
| 309 | relaxing the limitation, with which design and implementation of lockdep |
| 310 | can be efficient, might introduce inefficiency inevitably. So lockdep |
| 311 | should provide two options, strong detection and efficient detection. |
| 312 | |
| 313 | Choosing efficient detection: |
| 314 | |
| 315 | Lockdep works with only locks restricted to be released within the |
| 316 | acquire context. However, lockdep works efficiently. |
| 317 | |
| 318 | Choosing strong detection: |
| 319 | |
| 320 | Lockdep works with all synchronization primitives. However, lockdep |
| 321 | suffers performance degradation. |
| 322 | |
| 323 | CONCLUSION |
| 324 | |
| 325 | Relaxing the limitation, lockdep can add additional dependencies giving |
| 326 | additional opportunities to check circular dependencies. |
| 327 | |
| 328 | |
| 329 | ============ |
| 330 | Crossrelease |
| 331 | ============ |
| 332 | |
| 333 | Introduce crossrelease |
| 334 | ---------------------- |
| 335 | |
| 336 | In order to allow lockdep to handle additional dependencies by what |
| 337 | might be released in any context, namely 'crosslock', we have to be able |
| 338 | to identify those created by crosslocks. The proposed 'crossrelease' |
| 339 | feature provoides a way to do that. |
| 340 | |
| 341 | Crossrelease feature has to do: |
| 342 | |
| 343 | 1. Identify dependencies created by crosslocks. |
| 344 | 2. Add the dependencies into a dependency graph. |
| 345 | |
| 346 | That's all. Once a meaningful dependency is added into graph, then |
| 347 | lockdep would work with the graph as it did. The most important thing |
| 348 | crossrelease feature has to do is to correctly identify and add true |
| 349 | dependencies into the global graph. |
| 350 | |
| 351 | A dependency e.g. 'A -> B' can be identified only in the A's release |
| 352 | context because a decision required to identify the dependency can be |
| 353 | made only in the release context. That is to decide whether A can be |
| 354 | released so that a waiter for A can be woken up. It cannot be made in |
| 355 | other than the A's release context. |
| 356 | |
| 357 | It's no matter for typical locks because each acquire context is same as |
| 358 | its release context, thus lockdep can decide whether a lock can be |
| 359 | released in the acquire context. However for crosslocks, lockdep cannot |
| 360 | make the decision in the acquire context but has to wait until the |
| 361 | release context is identified. |
| 362 | |
| 363 | Therefore, deadlocks by crosslocks cannot be detected just when it |
| 364 | happens, because those cannot be identified until the crosslocks are |
| 365 | released. However, deadlock possibilities can be detected and it's very |
| 366 | worth. See 'APPENDIX A' section to check why. |
| 367 | |
| 368 | CONCLUSION |
| 369 | |
| 370 | Using crossrelease feature, lockdep can work with what might be released |
| 371 | in any context, namely crosslock. |
| 372 | |
| 373 | |
| 374 | Introduce commit |
| 375 | ---------------- |
| 376 | |
| 377 | Since crossrelease defers the work adding true dependencies of |
| 378 | crosslocks until they are actually released, crossrelease has to queue |
| 379 | all acquisitions which might create dependencies with the crosslocks. |
| 380 | Then it identifies dependencies using the queued data in batches at a |
| 381 | proper time. We call it 'commit'. |
| 382 | |
| 383 | There are four types of dependencies: |
| 384 | |
| 385 | 1. TT type: 'typical lock A -> typical lock B' |
| 386 | |
| 387 | Just when acquiring B, lockdep can see it's in the A's release |
| 388 | context. So the dependency between A and B can be identified |
| 389 | immediately. Commit is unnecessary. |
| 390 | |
| 391 | 2. TC type: 'typical lock A -> crosslock BX' |
| 392 | |
| 393 | Just when acquiring BX, lockdep can see it's in the A's release |
| 394 | context. So the dependency between A and BX can be identified |
| 395 | immediately. Commit is unnecessary, too. |
| 396 | |
| 397 | 3. CT type: 'crosslock AX -> typical lock B' |
| 398 | |
| 399 | When acquiring B, lockdep cannot identify the dependency because |
| 400 | there's no way to know if it's in the AX's release context. It has |
| 401 | to wait until the decision can be made. Commit is necessary. |
| 402 | |
| 403 | 4. CC type: 'crosslock AX -> crosslock BX' |
| 404 | |
| 405 | When acquiring BX, lockdep cannot identify the dependency because |
| 406 | there's no way to know if it's in the AX's release context. It has |
| 407 | to wait until the decision can be made. Commit is necessary. |
| 408 | But, handling CC type is not implemented yet. It's a future work. |
| 409 | |
| 410 | Lockdep can work without commit for typical locks, but commit step is |
| 411 | necessary once crosslocks are involved. Introducing commit, lockdep |
| 412 | performs three steps. What lockdep does in each step is: |
| 413 | |
| 414 | 1. Acquisition: For typical locks, lockdep does what it originally did |
| 415 | and queues the lock so that CT type dependencies can be checked using |
| 416 | it at the commit step. For crosslocks, it saves data which will be |
| 417 | used at the commit step and increases a reference count for it. |
| 418 | |
| 419 | 2. Commit: No action is reauired for typical locks. For crosslocks, |
| 420 | lockdep adds CT type dependencies using the data saved at the |
| 421 | acquisition step. |
| 422 | |
| 423 | 3. Release: No changes are required for typical locks. When a crosslock |
| 424 | is released, it decreases a reference count for it. |
| 425 | |
| 426 | CONCLUSION |
| 427 | |
| 428 | Crossrelease introduces commit step to handle dependencies of crosslocks |
| 429 | in batches at a proper time. |
| 430 | |
| 431 | |
| 432 | ============== |
| 433 | Implementation |
| 434 | ============== |
| 435 | |
| 436 | Data structures |
| 437 | --------------- |
| 438 | |
| 439 | Crossrelease introduces two main data structures. |
| 440 | |
| 441 | 1. hist_lock |
| 442 | |
| 443 | This is an array embedded in task_struct, for keeping lock history so |
| 444 | that dependencies can be added using them at the commit step. Since |
| 445 | it's local data, it can be accessed locklessly in the owner context. |
| 446 | The array is filled at the acquisition step and consumed at the |
| 447 | commit step. And it's managed in circular manner. |
| 448 | |
| 449 | 2. cross_lock |
| 450 | |
| 451 | One per lockdep_map exists. This is for keeping data of crosslocks |
| 452 | and used at the commit step. |
| 453 | |
| 454 | |
| 455 | How crossrelease works |
| 456 | ---------------------- |
| 457 | |
| 458 | It's the key of how crossrelease works, to defer necessary works to an |
| 459 | appropriate point in time and perform in at once at the commit step. |
| 460 | Let's take a look with examples step by step, starting from how lockdep |
| 461 | works without crossrelease for typical locks. |
| 462 | |
| 463 | acquire A /* Push A onto held_locks */ |
| 464 | acquire B /* Push B onto held_locks and add 'A -> B' */ |
| 465 | acquire C /* Push C onto held_locks and add 'B -> C' */ |
| 466 | release C /* Pop C from held_locks */ |
| 467 | release B /* Pop B from held_locks */ |
| 468 | release A /* Pop A from held_locks */ |
| 469 | |
| 470 | where A, B and C are different lock classes. |
| 471 | |
| 472 | NOTE: This document assumes that readers already understand how |
| 473 | lockdep works without crossrelease thus omits details. But there's |
| 474 | one thing to note. Lockdep pretends to pop a lock from held_locks |
| 475 | when releasing it. But it's subtly different from the original pop |
| 476 | operation because lockdep allows other than the top to be poped. |
| 477 | |
| 478 | In this case, lockdep adds 'the top of held_locks -> the lock to acquire' |
| 479 | dependency every time acquiring a lock. |
| 480 | |
| 481 | After adding 'A -> B', a dependency graph will be: |
| 482 | |
| 483 | A -> B |
| 484 | |
| 485 | where A and B are different lock classes. |
| 486 | |
| 487 | And after adding 'B -> C', the graph will be: |
| 488 | |
| 489 | A -> B -> C |
| 490 | |
| 491 | where A, B and C are different lock classes. |
| 492 | |
| 493 | Let's performs commit step even for typical locks to add dependencies. |
| 494 | Of course, commit step is not necessary for them, however, it would work |
| 495 | well because this is a more general way. |
| 496 | |
| 497 | acquire A |
| 498 | /* |
| 499 | * Queue A into hist_locks |
| 500 | * |
| 501 | * In hist_locks: A |
| 502 | * In graph: Empty |
| 503 | */ |
| 504 | |
| 505 | acquire B |
| 506 | /* |
| 507 | * Queue B into hist_locks |
| 508 | * |
| 509 | * In hist_locks: A, B |
| 510 | * In graph: Empty |
| 511 | */ |
| 512 | |
| 513 | acquire C |
| 514 | /* |
| 515 | * Queue C into hist_locks |
| 516 | * |
| 517 | * In hist_locks: A, B, C |
| 518 | * In graph: Empty |
| 519 | */ |
| 520 | |
| 521 | commit C |
| 522 | /* |
| 523 | * Add 'C -> ?' |
| 524 | * Answer the following to decide '?' |
| 525 | * What has been queued since acquire C: Nothing |
| 526 | * |
| 527 | * In hist_locks: A, B, C |
| 528 | * In graph: Empty |
| 529 | */ |
| 530 | |
| 531 | release C |
| 532 | |
| 533 | commit B |
| 534 | /* |
| 535 | * Add 'B -> ?' |
| 536 | * Answer the following to decide '?' |
| 537 | * What has been queued since acquire B: C |
| 538 | * |
| 539 | * In hist_locks: A, B, C |
| 540 | * In graph: 'B -> C' |
| 541 | */ |
| 542 | |
| 543 | release B |
| 544 | |
| 545 | commit A |
| 546 | /* |
| 547 | * Add 'A -> ?' |
| 548 | * Answer the following to decide '?' |
| 549 | * What has been queued since acquire A: B, C |
| 550 | * |
| 551 | * In hist_locks: A, B, C |
| 552 | * In graph: 'B -> C', 'A -> B', 'A -> C' |
| 553 | */ |
| 554 | |
| 555 | release A |
| 556 | |
| 557 | where A, B and C are different lock classes. |
| 558 | |
| 559 | In this case, dependencies are added at the commit step as described. |
| 560 | |
| 561 | After commits for A, B and C, the graph will be: |
| 562 | |
| 563 | A -> B -> C |
| 564 | |
| 565 | where A, B and C are different lock classes. |
| 566 | |
| 567 | NOTE: A dependency 'A -> C' is optimized out. |
| 568 | |
| 569 | We can see the former graph built without commit step is same as the |
| 570 | latter graph built using commit steps. Of course the former way leads to |
| 571 | earlier finish for building the graph, which means we can detect a |
| 572 | deadlock or its possibility sooner. So the former way would be prefered |
| 573 | when possible. But we cannot avoid using the latter way for crosslocks. |
| 574 | |
| 575 | Let's look at how commit steps work for crosslocks. In this case, the |
| 576 | commit step is performed only on crosslock AX as real. And it assumes |
| 577 | that the AX release context is different from the AX acquire context. |
| 578 | |
| 579 | BX RELEASE CONTEXT BX ACQUIRE CONTEXT |
| 580 | ------------------ ------------------ |
| 581 | acquire A |
| 582 | /* |
| 583 | * Push A onto held_locks |
| 584 | * Queue A into hist_locks |
| 585 | * |
| 586 | * In held_locks: A |
| 587 | * In hist_locks: A |
| 588 | * In graph: Empty |
| 589 | */ |
| 590 | |
| 591 | acquire BX |
| 592 | /* |
| 593 | * Add 'the top of held_locks -> BX' |
| 594 | * |
| 595 | * In held_locks: A |
| 596 | * In hist_locks: A |
| 597 | * In graph: 'A -> BX' |
| 598 | */ |
| 599 | |
| 600 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 601 | It must be guaranteed that the following operations are seen after |
| 602 | acquiring BX globally. It can be done by things like barrier. |
| 603 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 604 | |
| 605 | acquire C |
| 606 | /* |
| 607 | * Push C onto held_locks |
| 608 | * Queue C into hist_locks |
| 609 | * |
| 610 | * In held_locks: C |
| 611 | * In hist_locks: C |
| 612 | * In graph: 'A -> BX' |
| 613 | */ |
| 614 | |
| 615 | release C |
| 616 | /* |
| 617 | * Pop C from held_locks |
| 618 | * |
| 619 | * In held_locks: Empty |
| 620 | * In hist_locks: C |
| 621 | * In graph: 'A -> BX' |
| 622 | */ |
| 623 | acquire D |
| 624 | /* |
| 625 | * Push D onto held_locks |
| 626 | * Queue D into hist_locks |
| 627 | * Add 'the top of held_locks -> D' |
| 628 | * |
| 629 | * In held_locks: A, D |
| 630 | * In hist_locks: A, D |
| 631 | * In graph: 'A -> BX', 'A -> D' |
| 632 | */ |
| 633 | acquire E |
| 634 | /* |
| 635 | * Push E onto held_locks |
| 636 | * Queue E into hist_locks |
| 637 | * |
| 638 | * In held_locks: E |
| 639 | * In hist_locks: C, E |
| 640 | * In graph: 'A -> BX', 'A -> D' |
| 641 | */ |
| 642 | |
| 643 | release E |
| 644 | /* |
| 645 | * Pop E from held_locks |
| 646 | * |
| 647 | * In held_locks: Empty |
| 648 | * In hist_locks: D, E |
| 649 | * In graph: 'A -> BX', 'A -> D' |
| 650 | */ |
| 651 | release D |
| 652 | /* |
| 653 | * Pop D from held_locks |
| 654 | * |
| 655 | * In held_locks: A |
| 656 | * In hist_locks: A, D |
| 657 | * In graph: 'A -> BX', 'A -> D' |
| 658 | */ |
| 659 | commit BX |
| 660 | /* |
| 661 | * Add 'BX -> ?' |
| 662 | * What has been queued since acquire BX: C, E |
| 663 | * |
| 664 | * In held_locks: Empty |
| 665 | * In hist_locks: D, E |
| 666 | * In graph: 'A -> BX', 'A -> D', |
| 667 | * 'BX -> C', 'BX -> E' |
| 668 | */ |
| 669 | |
| 670 | release BX |
| 671 | /* |
| 672 | * In held_locks: Empty |
| 673 | * In hist_locks: D, E |
| 674 | * In graph: 'A -> BX', 'A -> D', |
| 675 | * 'BX -> C', 'BX -> E' |
| 676 | */ |
| 677 | release A |
| 678 | /* |
| 679 | * Pop A from held_locks |
| 680 | * |
| 681 | * In held_locks: Empty |
| 682 | * In hist_locks: A, D |
| 683 | * In graph: 'A -> BX', 'A -> D', |
| 684 | * 'BX -> C', 'BX -> E' |
| 685 | */ |
| 686 | |
| 687 | where A, BX, C,..., E are different lock classes, and a suffix 'X' is |
| 688 | added on crosslocks. |
| 689 | |
| 690 | Crossrelease considers all acquisitions after acqiuring BX are |
| 691 | candidates which might create dependencies with BX. True dependencies |
| 692 | will be determined when identifying the release context of BX. Meanwhile, |
| 693 | all typical locks are queued so that they can be used at the commit step. |
| 694 | And then two dependencies 'BX -> C' and 'BX -> E' are added at the |
| 695 | commit step when identifying the release context. |
| 696 | |
| 697 | The final graph will be, with crossrelease: |
| 698 | |
| 699 | -> C |
| 700 | / |
| 701 | -> BX - |
| 702 | / \ |
| 703 | A - -> E |
| 704 | \ |
| 705 | -> D |
| 706 | |
| 707 | where A, BX, C,..., E are different lock classes, and a suffix 'X' is |
| 708 | added on crosslocks. |
| 709 | |
| 710 | However, the final graph will be, without crossrelease: |
| 711 | |
| 712 | A -> D |
| 713 | |
| 714 | where A and D are different lock classes. |
| 715 | |
| 716 | The former graph has three more dependencies, 'A -> BX', 'BX -> C' and |
| 717 | 'BX -> E' giving additional opportunities to check if they cause |
| 718 | deadlocks. This way lockdep can detect a deadlock or its possibility |
| 719 | caused by crosslocks. |
| 720 | |
| 721 | CONCLUSION |
| 722 | |
| 723 | We checked how crossrelease works with several examples. |
| 724 | |
| 725 | |
| 726 | ============= |
| 727 | Optimizations |
| 728 | ============= |
| 729 | |
| 730 | Avoid duplication |
| 731 | ----------------- |
| 732 | |
| 733 | Crossrelease feature uses a cache like what lockdep already uses for |
| 734 | dependency chains, but this time it's for caching CT type dependencies. |
| 735 | Once that dependency is cached, the same will never be added again. |
| 736 | |
| 737 | |
| 738 | Lockless for hot paths |
| 739 | ---------------------- |
| 740 | |
| 741 | To keep all locks for later use at the commit step, crossrelease adopts |
| 742 | a local array embedded in task_struct, which makes access to the data |
| 743 | lockless by forcing it to happen only within the owner context. It's |
| 744 | like how lockdep handles held_locks. Lockless implmentation is important |
| 745 | since typical locks are very frequently acquired and released. |
| 746 | |
| 747 | |
| 748 | ================================================= |
| 749 | APPENDIX A: What lockdep does to work aggresively |
| 750 | ================================================= |
| 751 | |
| 752 | A deadlock actually occurs when all wait operations creating circular |
| 753 | dependencies run at the same time. Even though they don't, a potential |
| 754 | deadlock exists if the problematic dependencies exist. Thus it's |
| 755 | meaningful to detect not only an actual deadlock but also its potential |
| 756 | possibility. The latter is rather valuable. When a deadlock occurs |
| 757 | actually, we can identify what happens in the system by some means or |
| 758 | other even without lockdep. However, there's no way to detect possiblity |
| 759 | without lockdep unless the whole code is parsed in head. It's terrible. |
| 760 | Lockdep does the both, and crossrelease only focuses on the latter. |
| 761 | |
| 762 | Whether or not a deadlock actually occurs depends on several factors. |
| 763 | For example, what order contexts are switched in is a factor. Assuming |
| 764 | circular dependencies exist, a deadlock would occur when contexts are |
| 765 | switched so that all wait operations creating the dependencies run |
| 766 | simultaneously. Thus to detect a deadlock possibility even in the case |
| 767 | that it has not occured yet, lockdep should consider all possible |
| 768 | combinations of dependencies, trying to: |
| 769 | |
| 770 | 1. Use a global dependency graph. |
| 771 | |
| 772 | Lockdep combines all dependencies into one global graph and uses them, |
| 773 | regardless of which context generates them or what order contexts are |
| 774 | switched in. Aggregated dependencies are only considered so they are |
| 775 | prone to be circular if a problem exists. |
| 776 | |
| 777 | 2. Check dependencies between classes instead of instances. |
| 778 | |
| 779 | What actually causes a deadlock are instances of lock. However, |
| 780 | lockdep checks dependencies between classes instead of instances. |
| 781 | This way lockdep can detect a deadlock which has not happened but |
| 782 | might happen in future by others but the same class. |
| 783 | |
| 784 | 3. Assume all acquisitions lead to waiting. |
| 785 | |
| 786 | Although locks might be acquired without waiting which is essential |
| 787 | to create dependencies, lockdep assumes all acquisitions lead to |
| 788 | waiting since it might be true some time or another. |
| 789 | |
| 790 | CONCLUSION |
| 791 | |
| 792 | Lockdep detects not only an actual deadlock but also its possibility, |
| 793 | and the latter is more valuable. |
| 794 | |
| 795 | |
| 796 | ================================================== |
| 797 | APPENDIX B: How to avoid adding false dependencies |
| 798 | ================================================== |
| 799 | |
| 800 | Remind what a dependency is. A dependency exists if: |
| 801 | |
| 802 | 1. There are two waiters waiting for each event at a given time. |
| 803 | 2. The only way to wake up each waiter is to trigger its event. |
| 804 | 3. Whether one can be woken up depends on whether the other can. |
| 805 | |
| 806 | For example: |
| 807 | |
| 808 | acquire A |
| 809 | acquire B /* A dependency 'A -> B' exists */ |
| 810 | release B |
| 811 | release A |
| 812 | |
| 813 | where A and B are different lock classes. |
| 814 | |
| 815 | A depedency 'A -> B' exists since: |
| 816 | |
| 817 | 1. A waiter for A and a waiter for B might exist when acquiring B. |
| 818 | 2. Only way to wake up each is to release what it waits for. |
| 819 | 3. Whether the waiter for A can be woken up depends on whether the |
| 820 | other can. IOW, TASK X cannot release A if it fails to acquire B. |
| 821 | |
| 822 | For another example: |
| 823 | |
| 824 | TASK X TASK Y |
| 825 | ------ ------ |
| 826 | acquire AX |
| 827 | acquire B /* A dependency 'AX -> B' exists */ |
| 828 | release B |
| 829 | release AX held by Y |
| 830 | |
| 831 | where AX and B are different lock classes, and a suffix 'X' is added |
| 832 | on crosslocks. |
| 833 | |
| 834 | Even in this case involving crosslocks, the same rule can be applied. A |
| 835 | depedency 'AX -> B' exists since: |
| 836 | |
| 837 | 1. A waiter for AX and a waiter for B might exist when acquiring B. |
| 838 | 2. Only way to wake up each is to release what it waits for. |
| 839 | 3. Whether the waiter for AX can be woken up depends on whether the |
| 840 | other can. IOW, TASK X cannot release AX if it fails to acquire B. |
| 841 | |
| 842 | Let's take a look at more complicated example: |
| 843 | |
| 844 | TASK X TASK Y |
| 845 | ------ ------ |
| 846 | acquire B |
| 847 | release B |
| 848 | fork Y |
| 849 | acquire AX |
| 850 | acquire C /* A dependency 'AX -> C' exists */ |
| 851 | release C |
| 852 | release AX held by Y |
| 853 | |
| 854 | where AX, B and C are different lock classes, and a suffix 'X' is |
| 855 | added on crosslocks. |
| 856 | |
| 857 | Does a dependency 'AX -> B' exist? Nope. |
| 858 | |
| 859 | Two waiters are essential to create a dependency. However, waiters for |
| 860 | AX and B to create 'AX -> B' cannot exist at the same time in this |
| 861 | example. Thus the dependency 'AX -> B' cannot be created. |
| 862 | |
| 863 | It would be ideal if the full set of true ones can be considered. But |
| 864 | we can ensure nothing but what actually happened. Relying on what |
| 865 | actually happens at runtime, we can anyway add only true ones, though |
| 866 | they might be a subset of true ones. It's similar to how lockdep works |
| 867 | for typical locks. There might be more true dependencies than what |
| 868 | lockdep has detected in runtime. Lockdep has no choice but to rely on |
| 869 | what actually happens. Crossrelease also relies on it. |
| 870 | |
| 871 | CONCLUSION |
| 872 | |
| 873 | Relying on what actually happens, lockdep can avoid adding false |
| 874 | dependencies. |