1. Concurrency and Race Conditions Linux Kernel Programming CIS 4930/COP 5641

2. MOTIVATION: EXAMPLE PITFALL IN SCULL

3. Pitfalls in scull Race condition: result of uncontrolled access to shared data if (!dptr->data[s_pos]) { dptr->data[s_pos] = kzalloc(quantum, GFP_KERNEL); if (!dptr->data[s_pos]) { goto out; }

4. Pitfalls in scull Race condition: result of uncontrolled access to shared data if (!dptr->data[s_pos]) { dptr->data[s_pos] = kzalloc(quantum, GFP_KERNEL); if (!dptr->data[s_pos]) { goto out; }

5. Pitfalls in scull Race condition: result of uncontrolled access to shared data if (!dptr->data[s_pos]) { dptr->data[s_pos] = kmalloc(quantum, GFP_KERNEL); if (!dptr->data[s_pos]) { goto out; }

6. Pitfalls in scull Race condition: result of uncontrolled access to shared data if (!dptr->data[s_pos]) { dptr->data[s_pos] = kmalloc(quantum, GFP_KERNEL); if (!dptr->data[s_pos]) { goto out; } Memory leak

7. MANAGING CONCURRENCY

8. Concurrency and Its Management Sources of concurrency Multiple user-space processes Multiple CPUs Device interrupts Timers

9. Some guiding principles Try to avoid concurrent access entirely Global variables Apply locking and mutual exclusion principles Implications to device drivers Use sufficient concurrency mechanisms (depending on context) No object can be made available to the kernel until it can function properly References to such objects must be tracked for proper removal Avoid “roll your own” solutions

10. Managing Concurrency Atomic operation: all or nothing from the perspective of other threads Critical section: code executed by only one thread at a time Not all critical sections are the same Access from interrupt handlers Latency constraints

11. Lock Design Considerations Context Can another thread be scheduled on the current processor? Assumptions of kernel operation Breaking assumptions will break code that relies on them Time expected to wait for lock Considerations Amount of time lock is expected to be held Amount of expected contention Long Other threads can make better use of the processor Short Time to switch to another thread will be longer than just waiting a short amount of time

12. Kernel Locking Implementations mutex Sleep if lock cannot be acquired immediately Allow other threads to use the processor spinlock Continuously try to grab the lock Generally do not allow sleeping Why?

13. MUTEX

14. Mutex Implementation Architecture-dependent code Optimizations Initialization DEFINE_MUTEX(name) void mutex_init(struct mutex *lock); Various routines void mutex_lock(struct mutex *lock); int mutex_lock_interruptible(struct mutex *lock); int mutex_lock_killable(struct mutex *lock); void mutex_unlock(struct mutex *lock);

15. Using mutexes in scull scull_dev structure revisited struct scull_dev { struct scull_qset *data; /* Pointer to first quantum set */ int quantum; /* the current quantum size */ int qset; /* the current array size */ unsigned long size; /* amount of data stored here */ unsigned int access_key; /* used by sculluid & scullpriv */ struct mutex mutex; /* mutual exclusion */ struct cdev cdev; /* Char device structure */ };

16. Using mutexes in scull scull_dev initialization for (i = 0; i < scull_nr_devs; i++) { scull_devices[i].quantum = scull_quantum; scull_devices[i].qset = scull_qset; mutex_init(&scull_devices[i].mutex); /* before cdev_add */ scull_setup_cdev(&scull_devices[i], i); }

17. Using mutexes in scull scull_write() if (mutex_lock_interruptible(&dev->mutex)) return -ERESTARTSYS; scull_write ends with out: mutex_unlock(&dev->mutex); return retval;

18. mutex_lock_interruptible() returns nonzero If can be resubmitted Undo visible changes if any and restart Otherwise return ‑ EINTR E.g., could not undo changes

19. mutex_lock_interruptible() (returns non-zero) If can be resubmitted Undo visible changes if any and restart Otherwise return ‑ EINTR E.g., could not undo changes

20. Restartable system call Automatic restarting of certain interrupted system calls Retry with same arguments (values) Simplifies user-space programming for dealing with "interrupted system call“ POSIX permits an implementation to restart system calls, but it is not required. SUS defines the SA_RESTART flag to provide a means by which an application can request that an interrupted system calls be restarted. http://pubs.opengroup.org/onlinepubs/009604499/function s/sigaction.html return ‑ ERESTARTSYS

21. Restartable system call Arguments may need to be modified return ‑ ERESTARTSYS_RESTARTBLOCK Specify callback function to modify arguments http://lwn.net/Articles/17744/

22. Userspace write() and kernelspace *_interruptible() From POSIX man page If write() is interrupted by a signal before it writes any data, it shall return -1 with errno set to [EINTR]. If write() is interrupted by a signal after it successfully writes some data, it shall return the number of bytes written. http://pubs.opengroup.org/onlinepubs/009604 499/functions/sigaction.html

23. mutex_lock_killable() mutex_lock() Process assumes that it cannot be interrupted by a signal Breaking assumption breaks user-kernel space interface If process receives fatal signal and mutex_lock() never returns Results in an immortal process Assumptions/expectations do not apply if process receives fatal signal Process that called system call will never return Does not break assumption since process does not continue http://lwn.net/Articles/288056/

24. MUTEX USAGE AS COMPLETION (ERROR) HTTPS://LKML.ORG/LKML/2013/12/2/997

25. General Pattern refcount variable for deciding which thread to perform cleanup Usage Initialize shared object Set refcount to number of concurrent threads Start multiple threads Last thread cleans up mutex_lock(obj->lock); dead = !--obj->refcount; mutex_unlock(obj->lock); if (dead) free(obj);

26. fs/pipe.c __pipe_lock(pipe); … spin_lock(&inode->i_lock); if (!--pipe->files) { inode->i_pipe = NULL; kill = 1; } spin_unlock(&inode->i_lock); __pipe_unlock(pipe); if (kill) free_pipe_info(pipe);

27. mutex_unlock(obj->lock); __mutex_fastpath_unlock() fastpath fails (because mutex is nonpositive __mutex_unlock_slowpath: if (__mutex_slowpath_needs_to_unlock()) atomic_set(&lock->count, 1); if ((atomic_read(&lock->count) == 1) && (atomic_cmpxchg(&lock->count, 1, 0) == 1)) {.. and now CPU2 owns the mutex, and goes on dead = !--obj->refcount; // refcount was 1, is now 0, dead = 1. mutex_unlock(obj->lock); if (dead) free(obj); mutex_lock(obj->lock); dead = !--obj->refcount; // refcount was 2, is now 1, dead = 0. mutex_lock(obj->lock); // blocks on obj->lock, goes to slowpath // mutex is negative, CPU2 is in optimistic // spinning mode in __mutex_lock_common CPU 1 CPU 2 but in the meantime, CPU1 is busy still unlocking: if (!list_empty(&lock->wait_list)) {

28. Conclusion Mutex serializes what is inside the mutex, but not necessarily the lock ITSELF Use spinlocks and/or atomic ref counts "don't use mutexes to implement completions"

29. COMPLETIONS

30. Completions Start and wait for operation to complete (outside current thread) Common pattern in kernel programming E.g., wait for initialization to complete Reasons to use instead of mutexes Wake up multiple threads More efficient More meaningful syntax Subtle races with mutex implementation code Cleanup of mutex itself http://lkml.iu.edu//hypermail/linux/kernel/0107.3/0674.html https://lkml.org/lkml/2008/4/11/323 completions #include

31. Completions To create a completion DECLARE_COMPLETION(my_completion); Or struct completion my_completion; init_completion(&my_completion); To wait for the completion, call void wait_for_completion(struct completion *c); void wait_for_completion_interruptible(struct completion *c); void wait_for_completion_timeout(struct completion *c, unsigned long timeout);

32. Completions To signal a completion event, call one of the following /* wake up one waiting thread */ void complete(struct completion *c); /* wake up multiple waiting threads */ /* need to call INIT_COMPLETION(struct completion c) to reuse the completion structure */ void complete_all(struct completion *c);

33. Completions Example: misc-modules/complete.c DECLARE_COMPLETION(comp); ssize_t complete_read(struct file *filp, char __user *buf, size_t count, loff_t *pos) { printk(KERN_DEBUG "process %i (%s) going to sleep\n", current->pid, current->comm); wait_for_completion(&comp); printk(KERN_DEBUG "awoken %i (%s)\n", current->pid, current->comm); return 0; /* EOF */ }

34. Completions Example ssize_t complete_write(struct file *filp, const char __user *buf, size_t count, loff_t *pos) { printk(KERN_DEBUG "process %i (%s) awakening the readers...\n", current->pid, current->comm); complete(&comp); return count; /* succeed, to avoid retrial */ }

35. SPINLOCKS

36. Spinlocks Generally used in code that should not sleep (e.g., interrupt handlers) Usually implemented as a single bit If the lock is available, the bit is set and the code continues If the lock is taken, the code enters a tight loop Repeatedly checks the lock until it become available

37. Spinlocks Actual implementation varies for different architectures Protect a process from other CPUs and interrupts Usually does nothing on uniprocessor machines Exception: changing the IRQ masking status

38. Introduction to Spinlock API #include To initialize, declare spinlock_t my_lock = SPIN_LOCK_UNLOCKED; Or call void spin_lock_init(spinlock_t *lock); To acquire a lock, call void spin_lock(spinlock_t *lock); Spinlock waits are uninterruptible To release a lock, call void spin_unlock(spinlock_t *lock);

39. Spinlocks and Atomic Context While holding a spinlock, be atomic Do not sleep or relinquish the processor Examples of calls that can sleep  Copying data to or from user space  User-space page may need to be on disk…  Memory allocation  Memory might not be available Disable interrupts (on the local CPU) as needed Hold spinlocks for the minimum time possible

40. The Spinlock Functions Four functions to acquire a spinlock void spin_lock(spinlock_t *lock); /* disables interrupts on the local CPU */ void spin_lock_irqsave(spinlock_t *lock, unsigned long flags); /* only if no other code disabled interrupts */ void spin_lock_irq(spinlock_t *lock); /* disables software interrupts; leaves hardware interrupts enabled (e.g. tasklets)*/ void spin_lock_bh(spinlock_t *lock);

41. The Spinlock Functions Four functions to release a spinlock void spin_unlock(spinlock_t *lock); /* need to use the same flags variable for locking */ /* need to call spin_lock_irqsave and spin_unlock_irqrestore in the same function, or your code may break on some architectures */ void spin_unlock_irqrestore(spinlock_t *lock, unsigned long flags); void spin_unlock_irq(spinlock_t *lock); void spin_unlock_bh(spinlock_t *lock);

42. Locking Traps It is very hard to manage concurrency What can possibly go wrong?

43. Ambiguous Rules Shared data structure D, protected by lock L function A() { lock(&L); /* call function B() that accesses D */ unlock(&L); } If function B() calls lock(&L), we have a deadlock

44. Ambiguous Rules Solution Have clear entry points to access data structures Document assumptions about locking

45. Lock Ordering Rules function A() { lock(&L1); lock(&L2); /* access D */ unlock(&L2); unlock(&L1) } function B() { lock(&L2); lock(&L1); /* access D */ unlock(&L1); unlock(&L2) } - Multiple locks should always be acquired in the same order - Easier said than done

46. Lock Ordering Rules function A() { lock(&L1); X(); unlock(&L1) } function X() { lock(&L2); /* access D */ unlock(&L2); } function B() { lock(&L2); Y(); unlock(&L2) } function Y() { lock(&L1); /* access D */ unlock(&L1); }

47. Lock Ordering Rules of Thumb Choose a lock ordering that is local to your code before taking a lock belonging to a more central part of the kernel Lock of central kernel code likely has more users (more contention) Obtain the mutex first before taking the spinlock Grabbing a mutex (which can sleep) inside a spinlock can lead to deadlocks

48. Fine- Versus Coarse-Grained Locking Coarse-grained locking Poor concurrency Fine-grained locking Need to know which one to acquire And which order to acquire At the device driver level Start with coarse-grained locking Refine the granularity as contention arises Can enable lockstat to check lock holding time

49. BKL Kernel used to have “big kernel lock” Giant spinlock introduced in Linux 2.0 Only one CPU could be executing locked kernel code at any time BKL has been removed https://lwn.net/Articles/384855/ https://www.linux.com/learn/tutorials/447301:w hats-new-in-linux-2639-ding-dong-the-big- kernel-lock-is-dead https://www.linux.com/learn/tutorials/447301:w hats-new-in-linux-2639-ding-dong-the-big- kernel-lock-is-dead

50. Alternatives to Locking Lock-free algorithms Atomic variables Bit operations seqlocks Read-copy-update (RCU)

51. Lock-Free Algorithms Circular buffer Producer places data into one end of an array When the end of the array is reached, the producer wraps back Consumer removes data from the other end

52. Lock-Free Algorithms Producer and consumer can access buffer concurrently without race conditions Always store the value before updating the index into the array Need to make sure that producer/consumer indices do not overrun each other A generic circular buffer is available See

53. ATOMIC VARIABLES

54. Atomic Variables If the shared resource is an integer value Locking is overkill (if supported by processor) The kernel provides atomic types atomic_t - integer atomic64_t – long integer Both types must be accessed through special functions (See ) SMP safe

55. Atomic Variables Atomic operations atomic_sub(amount, &account1); atomic_add(amount, &account2); A higher level locking must be used

56. Bit Operations Atomic bit operations See SMP safe

57. OTHER SYNCHRONIZATION MECHANISMS

58. Read-Copy-Update (RCU) Assumptions Reads are common Writes are rare Resources accessed via pointers All references to those resources held by atomic code

59. Read-Copy-Update Basic idea The writing thread makes a copy Make changes to the copy Switch a few pointers to commit changes Deallocate the old version when all references to the old version are gone

60. EVEN MORE…

61. seqlocks sequential lock Designed to protect small, simple, and frequently accessed resource Write access is rare Must obtain an exclusive lock (spinlock) Allow readers free access to the resource Lockless Operation Check for collisions with writers Retry as needed Not for protecting pointers

62. seqlocks Expected non-blocking reader usage: do { seq = read_seqbegin(&foo);... } while (read_seqretry(&foo, seq));

63. lglock (local/global locks) Fast per-cpu access Allows access to other cpu data (slow) Implementation per-CPU array of spinlocks Can only be declared as global variables to avoid overhead and keep things simple http://lwn.net/Articles/401738/

64. brlocks Sat Oct 5 14:19:39 2013 -0400 no need to keep brlock macros anymore... 0f6ed63b170778b9c93fb0ae4017f110c9ee64 16

65. Reader/Writer Semaphores Allow multiple concurrent readers Single writer (for infrequent writes) Too many writers can lead to reader starvation (unbounded waiting) #include Do not follow the return value convention E.g., returns 1 if successful Not interruptible

66. Reader/Writer Spinlocks Analogous to the reader/writer semaphores Allow multiple readers to enter a critical section Provide exclusive access for writers #include

67. Reader/Writer Spinlocks To declare and initialize, there are two ways /* static way */ rwlock_t my_rwlock = RW_LOCK_UNLOCKED; /* dynamic way */ rwlock_t my_rwlock; rwlock_init(&my_rwlock);

68. Reader/Writer Spinlocks Similar functions are available void read_lock(rwlock_t *lock); void read_lock_irqsave(rwlock_t *lock, unsigned long flags); void read_lock_irq(rwlock_t *lock); void read_lock_bh(rwlock_t *lock); void read_unlock(rwlock_t *lock); void read_unlock_irqrestore(rwlock_t *lock, unsigned long flags); void read_unlock_irq(rwlock_t *lock); void read_unlock_bh(rwlock_t *lock);

69. Reader/Writer Spinlocks Similar functions are available void write_lock(rwlock_t *lock); void write_lock_irqsave(rwlock_t *lock, unsigned long flags); void write_lock_irq(rwlock_t *lock); void write_lock_bh(rwlock_t *lock); void write_unlock(rwlock_t *lock); void write_unlock_irqrestore(rwlock_t *lock, unsigned long flags); void write_unlock_irq(rwlock_t *lock); void write_unlock_bh(rwlock_t *lock);