1 Runtime locking correctness validator 2 ===================================== 3 4 started by Ingo Molnar <mingo@redhat.com> 5 6 additions by Arjan van de Ven <arjan@linux.inte 7 8 Lock-class 9 ---------- 10 11 The basic object the validator operates upon i 12 13 A class of locks is a group of locks that are 14 respect to locking rules, even if the locks ma 15 tens of thousands of) instantiations. For exam 16 struct is one class, while each inode has its 17 lock class. 18 19 The validator tracks the 'usage state' of lock 20 the dependencies between different lock-classe 21 how a lock is used with regard to its IRQ cont 22 dependency can be understood as lock order, wh 23 a task is attempting to acquire L2 while holdi 24 perspective, the two locks (L1 and L2) are not 25 dependency just means the order ever happened. 26 continuing effort to prove lock usages and dep 27 the validator will shoot a splat if incorrect. 28 29 A lock-class's behavior is constructed by its 30 when the first instance of a lock-class is use 31 gets registered, then all (subsequent) instanc 32 class and hence their usages and dependencies 33 the class. A lock-class does not go away when 34 it can be removed if the memory space of the l 35 dynamic) is reclaimed, this happens for exampl 36 unloaded or a workqueue is destroyed. 37 38 State 39 ----- 40 41 The validator tracks lock-class usage history 42 (4 usages * n STATEs + 1) categories: 43 44 where the 4 usages can be: 45 46 - 'ever held in STATE context' 47 - 'ever held as readlock in STATE context' 48 - 'ever held with STATE enabled' 49 - 'ever held as readlock with STATE enabled' 50 51 where the n STATEs are coded in kernel/locking 52 now they include: 53 54 - hardirq 55 - softirq 56 57 where the last 1 category is: 58 59 - 'ever used' 60 61 When locking rules are violated, these usage b 62 locking error messages, inside curlies, with a 63 A contrived example:: 64 65 modprobe/2287 is trying to acquire lock: 66 (&sio_locks[i].lock){-.-.}, at: [<c02867fd 67 68 but task is already holding lock: 69 (&sio_locks[i].lock){-.-.}, at: [<c02867fd 70 71 72 For a given lock, the bit positions from left 73 of the lock and readlock (if exists), for each 74 above respectively, and the character displaye 75 indicates: 76 77 === ====================================== 78 '.' acquired while irqs disabled and not i 79 '-' acquired in irq context 80 '+' acquired with irqs enabled 81 '?' acquired in irq context with irqs enab 82 === ====================================== 83 84 The bits are illustrated with an example:: 85 86 (&sio_locks[i].lock){-.-.}, at: [<c02867fd 87 |||| 88 ||| \-> softirq disab 89 || \--> acquired in s 90 | \---> hardirq disab 91 \----> acquired in h 92 93 94 For a given STATE, whether the lock is ever ac 95 context and whether that STATE is enabled yiel 96 shown in the table below. The bit character is 97 exact case is for the lock as of the reporting 98 99 +--------------+-------------+-------------- 100 | | irq enabled | irq disabled 101 +--------------+-------------+-------------- 102 | ever in irq | '?' | '-' 103 +--------------+-------------+-------------- 104 | never in irq | '+' | '.' 105 +--------------+-------------+-------------- 106 107 The character '-' suggests irq is disabled bec 108 character '?' would have been shown instead. S 109 applied for '+' too. 110 111 Unused locks (e.g., mutexes) cannot be part of 112 113 114 Single-lock state rules: 115 ------------------------ 116 117 A lock is irq-safe means it was ever used in a 118 is irq-unsafe means it was ever acquired with 119 120 A softirq-unsafe lock-class is automatically h 121 following states must be exclusive: only one o 122 for any lock-class based on its usage:: 123 124 <hardirq-safe> or <hardirq-unsafe> 125 <softirq-safe> or <softirq-unsafe> 126 127 This is because if a lock can be used in irq c 128 cannot be ever acquired with irq enabled (irq- 129 deadlock may happen. For example, in the scena 130 was acquired but before released, if the conte 131 lock will be attempted to acquire twice, which 132 referred to as lock recursion deadlock. 133 134 The validator detects and reports lock usage t 135 single-lock state rules. 136 137 Multi-lock dependency rules: 138 ---------------------------- 139 140 The same lock-class must not be acquired twice 141 to lock recursion deadlocks. 142 143 Furthermore, two locks can not be taken in inv 144 145 <L1> -> <L2> 146 <L2> -> <L1> 147 148 because this could lead to a deadlock - referr 149 deadlock - as attempts to acquire the two lock 150 could lead to the two contexts waiting for eac 151 validator will find such dependency circle in 152 i.e., there can be any other locking sequence 153 operations; the validator will still find whet 154 acquired in a circular fashion. 155 156 Furthermore, the following usage based lock de 157 between any two lock-classes:: 158 159 <hardirq-safe> -> <hardirq-unsafe> 160 <softirq-safe> -> <softirq-unsafe> 161 162 The first rule comes from the fact that a hard 163 taken by a hardirq context, interrupting a har 164 thus could result in a lock inversion deadlock 165 lock could be taken by an softirq context, int 166 lock. 167 168 The above rules are enforced for any locking s 169 kernel: when acquiring a new lock, the validat 170 any rule violation between the new lock and an 171 172 When a lock-class changes its state, the follo 173 dependency rules are enforced: 174 175 - if a new hardirq-safe lock is discovered, we 176 took any hardirq-unsafe lock in the past. 177 178 - if a new softirq-safe lock is discovered, we 179 any softirq-unsafe lock in the past. 180 181 - if a new hardirq-unsafe lock is discovered, 182 hardirq-safe lock took it in the past. 183 184 - if a new softirq-unsafe lock is discovered, 185 softirq-safe lock took it in the past. 186 187 (Again, we do these checks too on the basis th 188 could interrupt _any_ of the irq-unsafe or har 189 could lead to a lock inversion deadlock - even 190 not trigger in practice yet.) 191 192 Exception: Nested data dependencies leading to 193 ---------------------------------------------- 194 195 There are a few cases where the Linux kernel a 196 instance of the same lock-class. Such cases ty 197 is some sort of hierarchy within objects of th 198 cases there is an inherent "natural" ordering 199 (defined by the properties of the hierarchy), 200 locks in this fixed order on each of the objec 201 202 An example of such an object hierarchy that re 203 is that of a "whole disk" block-dev object and 204 object; the partition is "part of" the whole d 205 always takes the whole disk lock as a higher l 206 lock, the lock ordering is fully correct. The 207 automatically detect this natural ordering, as 208 the ordering is not static. 209 210 In order to teach the validator about this cor 211 versions of the various locking primitives wer 212 specify a "nesting level". An example call, fo 213 looks like this:: 214 215 enum bdev_bd_mutex_lock_class 216 { 217 BD_MUTEX_NORMAL, 218 BD_MUTEX_WHOLE, 219 BD_MUTEX_PARTITION 220 }; 221 222 mutex_lock_nested(&bdev->bd_contains->bd_mut 223 224 In this case the locking is done on a bdev obj 225 partition. 226 227 The validator treats a lock that is taken in s 228 separate (sub)class for the purposes of valida 229 230 Note: When changing code to use the _nested() 231 check really thoroughly that the hierarchy is 232 you can get false positives or false negatives 233 234 Annotations 235 ----------- 236 237 Two constructs can be used to annotate and che 238 must be held: lockdep_assert_held*(&lock) and 239 240 As the name suggests, lockdep_assert_held* fam 241 particular lock is held at a certain time (and 242 This annotation is largely used all over the k 243 core.c:: 244 245 void update_rq_clock(struct rq *rq) 246 { 247 s64 delta; 248 249 lockdep_assert_held(&rq->lock); 250 [...] 251 } 252 253 where holding rq->lock is required to safely u 254 255 The other family of macros is lockdep_*pin_loc 256 used for rq->lock ATM. Despite their limited a 257 generate a WARN() if the lock of interest is " 258 out to be especially helpful to debug code wit 259 layer assumes a lock remains taken, but a lowe 260 and reacquire the lock ("unwittingly" introduc 261 returns a 'struct pin_cookie' that is then use 262 that nobody tampered with the lock, e.g. kerne 263 264 static inline void rq_pin_lock(struct rq *rq 265 { 266 rf->cookie = lockdep_pin_lock(&rq->loc 267 [...] 268 } 269 270 static inline void rq_unpin_lock(struct rq * 271 { 272 [...] 273 lockdep_unpin_lock(&rq->lock, rf->cook 274 } 275 276 While comments about locking requirements migh 277 the runtime checks performed by annotations ar 278 locking problems and they carry the same level 279 code. Always prefer annotations when in doubt 280 281 Proof of 100% correctness: 282 -------------------------- 283 284 The validator achieves perfect, mathematical ' 285 correctness) in the sense that for every simpl 286 locking sequence that occurred at least once d 287 kernel, the validator proves it with a 100% ce 288 combination and timing of these locking sequen 289 lock related deadlock. [1]_ 290 291 I.e. complex multi-CPU and multi-task locking 292 occur in practice to prove a deadlock: only th 293 locking chains have to occur at least once (an 294 task/context) for the validator to be able to 295 example, complex deadlocks that would normally 296 a very unlikely constellation of tasks, irq-co 297 occur, can be detected on a plain, lightly loa 298 well!) 299 300 This radically decreases the complexity of loc 301 kernel: what has to be done during QA is to tr 302 single-task locking dependencies in the kernel 303 once, to prove locking correctness - instead o 304 possible combination of locking interaction be 305 every possible hardirq and softirq nesting sce 306 to do in practice). 307 308 .. [1] 309 310 assuming that the validator itself is 100% 311 part of the system corrupts the state of t 312 We also assume that all NMI/SMM paths [whi 313 even hardirq-disabled codepaths] are corre 314 with the validator. We also assume that th 315 value is unique for every lock-chain in th 316 recursion must not be higher than 20. 317 318 Performance: 319 ------------ 320 321 The above rules require **massive** amounts of 322 that for every lock taken and for every irqs-e 323 render the system practically unusably slow. T 324 is O(N^2), so even with just a few hundred loc 325 tens of thousands of checks for every event. 326 327 This problem is solved by checking any given ' 328 sequence of locks taken after each other) only 329 held locks is maintained, and a lightweight 64 330 calculated, which hash is unique for every loc 331 when the chain is validated for the first time 332 table, which hash-table can be checked in a lo 333 locking chain occurs again later on, the hash 334 don't have to validate the chain again. 335 336 Troubleshooting: 337 ---------------- 338 339 The validator tracks a maximum of MAX_LOCKDEP_ 340 Exceeding this number will trigger the followi 341 342 (DEBUG_LOCKS_WARN_ON(id >= MAX_LOCKDEP 343 344 By default, MAX_LOCKDEP_KEYS is currently set 345 desktop systems have less than 1,000 lock clas 346 normally results from lock-class leakage or fa 347 initialize locks. These two problems are illu 348 349 1. Repeated module loading and unloading 350 will result in lock-class leakage. Th 351 load of the module will create a new s 352 that module's locks, but module unload 353 classes (see below discussion of reuse 354 Therefore, if that module is loaded an 355 the number of lock classes will eventu 356 357 2. Using structures such as arrays that h 358 locks that are not explicitly initiali 359 a hash table with 8192 buckets where e 360 spinlock_t will consume 8192 lock clas 361 is explicitly initialized at runtime, 362 run-time spin_lock_init() as opposed t 363 such as __SPIN_LOCK_UNLOCKED(). Failu 364 the per-bucket spinlocks would guarant 365 In contrast, a loop that called spin_l 366 would place all 8192 locks into a sing 367 368 The moral of this story is that you sh 369 initialize your locks. 370 371 One might argue that the validator should be m 372 lock classes to be reused. However, if you ar 373 argument, first review the code and think thro 374 be required, keeping in mind that the lock cla 375 likely to be linked into the lock-dependency g 376 be harder to do than to say. 377 378 Of course, if you do run out of lock classes, 379 to find the offending lock classes. First, th 380 you the number of lock classes currently in us 381 382 grep "lock-classes" /proc/lockdep_stat 383 384 This command produces the following output on 385 386 lock-classes: 387 388 If the number allocated (748 above) increases 389 then there is likely a leak. The following co 390 identify the leaking lock classes:: 391 392 grep "BD" /proc/lockdep 393 394 Run the command and save the output, then comp 395 a later run of this command to identify the le 396 can also help you find situations where runtim 397 been omitted. 398 399 Recursive read locks: 400 --------------------- 401 The whole of the rest document tries to prove 402 to deadlock possibility. 403 404 There are three types of lockers: writers (i.e 405 spin_lock() or write_lock()), non-recursive re 406 down_read()) and recursive readers (recursive 407 And we use the following notations of those lo 408 409 W or E: stands for writers (exclusive 410 r: stands for non-recursive reade 411 R: stands for recursive readers. 412 S: stands for all readers (non-re 413 N: stands for writers and non-rec 414 415 Obviously, N is "r or W" and S is "r or R". 416 417 Recursive readers, as their name indicates, ar 418 even inside the critical section of another re 419 in other words, allowing nested read-side crit 420 421 While non-recursive readers will cause a self 422 the critical section of another reader of the 423 424 The difference between recursive readers and n 425 recursive readers get blocked only by a write 426 readers could get blocked by a write lock *wai 427 example:: 428 429 TASK A: TASK B: 430 431 read_lock(X); 432 write_lock(X); 433 read_lock_2(X); 434 435 Task A gets the reader (no matter whether recu 436 read_lock() first. And when task B tries to ac 437 and become a waiter for writer on X. Now if re 438 task A will make progress, because writer wait 439 and there is no deadlock. However, if read_loc 440 it will get blocked by writer waiter B, and ca 441 442 Block conditions on readers/writers of the sam 443 ---------------------------------------------- 444 There are simply four block conditions: 445 446 1. Writers block other writers. 447 2. Readers block writers. 448 3. Writers block both recursive readers a 449 4. And readers (recursive or not) don't b 450 may block non-recursive readers (becau 451 writer waiters) 452 453 Block condition matrix, Y means the row blocks 454 455 +---+---+---+---+ 456 | | W | r | R | 457 +---+---+---+---+ 458 | W | Y | Y | Y | 459 +---+---+---+---+ 460 | r | Y | Y | N | 461 +---+---+---+---+ 462 | R | Y | Y | N | 463 +---+---+---+---+ 464 465 (W: writers, r: non-recursive readers, 466 467 468 acquired recursively. Unlike non-recursive rea 469 only get blocked by current write lock *holder 470 *waiters*, for example:: 471 472 TASK A: TASK B: 473 474 read_lock(X); 475 476 write_lock(X); 477 478 read_lock(X); 479 480 is not a deadlock for recursive read locks, as 481 the lock X, the second read_lock() doesn't nee 482 read lock. However if the read_lock() is non-r 483 case is a deadlock, because even if the write_ 484 lock, but it can block the second read_lock() 485 486 Note that a lock can be a write lock (exclusiv 487 lock (non-recursive shared lock) or a recursiv 488 lock), depending on the lock operations used t 489 the value of the 'read' parameter for lock_acq 490 lock instance has three types of acquisition d 491 functions: exclusive, non-recursive read, and 492 493 To be concise, we call that write locks and no 494 "non-recursive" locks and recursive read locks 495 496 Recursive locks don't block each other, while 497 even true for two non-recursive read locks). A 498 corresponding recursive lock, and vice versa. 499 500 A deadlock case with recursive locks involved 501 502 TASK A: TASK B: 503 504 read_lock(X); 505 read_lock(Y); 506 write_lock(Y); 507 write_lock(X); 508 509 Task A is waiting for task B to read_unlock() 510 A to read_unlock() X. 511 512 Dependency types and strong dependency paths: 513 --------------------------------------------- 514 Lock dependencies record the orders of the acq 515 because there are 3 types for lockers, there a 516 dependencies, but we can show that 4 types of 517 deadlock detection. 518 519 For each lock dependency:: 520 521 L1 -> L2 522 523 , which means lockdep has seen L1 held before 524 And in deadlock detection, we care whether we 525 IOW, whether there is a locker L3 that L1 bloc 526 we only care about 1) what L1 blocks and 2) wh 527 recursive readers and non-recursive readers fo 528 we can combine writers and non-recursive reade 529 same types). 530 531 With the above combination for simplification, 532 in the lockdep graph: 533 534 1) -(ER)->: 535 exclusive writer to recursive read 536 X -> Y and X is a writer and Y is 537 538 2) -(EN)->: 539 exclusive writer to non-recursive 540 X -> Y and X is a writer and Y is 541 542 3) -(SR)->: 543 shared reader to recursive reader 544 X -> Y and X is a reader (recursiv 545 546 4) -(SN)->: 547 shared reader to non-recursive loc 548 X -> Y and X is a reader (recursiv 549 non-recursive reader. 550 551 Note that given two locks, they may have multi 552 for example:: 553 554 TASK A: 555 556 read_lock(X); 557 write_lock(Y); 558 ... 559 560 TASK B: 561 562 write_lock(X); 563 write_lock(Y); 564 565 , we have both X -(SN)-> Y and X -(EN)-> Y in 566 567 We use -(xN)-> to represent edges that are eit 568 similar for -(Ex)->, -(xR)-> and -(Sx)-> 569 570 A "path" is a series of conjunct dependency ed 571 "strong" path, which indicates the strong depe 572 in the path, as the path that doesn't have two 573 -(xR)-> and -(Sx)->. In other words, a "strong 574 walking to another through the lock dependenci 575 path (where X, Y, Z are locks), and the walk f 576 -(ER)-> dependency, the walk from Y to Z must 577 -(SR)-> dependency. 578 579 We will see why the path is called "strong" in 580 581 Recursive Read Deadlock Detection: 582 ---------------------------------- 583 584 We now prove two things: 585 586 Lemma 1: 587 588 If there is a closed strong path (i.e. a stron 589 combination of locking sequences that causes d 590 sufficient for deadlock detection. 591 592 Lemma 2: 593 594 If there is no closed strong path (i.e. strong 595 combination of locking sequences that could ca 596 circles are necessary for deadlock detection. 597 598 With these two Lemmas, we can easily say a clo 599 and necessary for deadlocks, therefore a close 600 deadlock possibility. As a closed strong path 601 could cause deadlocks, so we call it "strong", 602 circles that won't cause deadlocks. 603 604 Proof for sufficiency (Lemma 1): 605 606 Let's say we have a strong circle:: 607 608 L1 -> L2 ... -> Ln -> L1 609 610 , which means we have dependencies:: 611 612 L1 -> L2 613 L2 -> L3 614 ... 615 Ln-1 -> Ln 616 Ln -> L1 617 618 We now can construct a combination of locking 619 620 Firstly let's make one CPU/task get the L1 in 621 the L2 in L2 -> L3, and so on. After this, all 622 held by different CPU/tasks. 623 624 And then because we have L1 -> L2, so the hold 625 in L1 -> L2, however since L2 is already held 626 L2 and L2 -> L3 are not -(xR)-> and -(Sx)-> (t 627 means either L2 in L1 -> L2 is a non-recursive 628 the L2 in L2 -> L3, is writer (blocking anyone 629 cannot get L2, it has to wait L2's holder to r 630 631 Moreover, we can have a similar conclusion for 632 holder to release, and so on. We now can prove 633 Lx+1's holder to release, and note that Ln+1 i 634 waiting scenario and nobody can get progress, 635 636 Proof for necessary (Lemma 2): 637 638 Lemma 2 is equivalent to: If there is a deadlo 639 strong circle in the dependency graph. 640 641 According to Wikipedia[1], if there is a deadl 642 waiting scenario, means there are N CPU/tasks, 643 a lock held by P2, and P2 is waiting for a loc 644 for a lock held by P1. Let's name the lock Px 645 for L1 and holding Ln, so we will have Ln -> L 646 we have L1 -> L2, L2 -> L3, ..., Ln-1 -> Ln in 647 have a circle:: 648 649 Ln -> L1 -> L2 -> ... -> Ln 650 651 , and now let's prove the circle is strong: 652 653 For a lock Lx, Px contributes the dependency L 654 the dependency Lx -> Lx+1, and since Px is wai 655 so it's impossible that Lx on Px+1 is a reader 656 reader, because readers (no matter recursive o 657 readers, therefore Lx-1 -> Lx and Lx -> Lx+1 c 658 and this is true for any lock in the circle, t 659 660 References: 661 ----------- 662 [1]: https://en.wikipedia.org/wiki/Deadlock 663 [2]: Shibu, K. (2009). Intro To Embedded Syste
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.