1. © 2006 RightNow Technologies, Inc. Tribute to Synchronization Jeff Elser December 11 th, 2006

2. Page 2 Outline Why synchronize? –Synergy. Historical solutions Synchronization Primitives Brief word on concurrency Semaphores in depth –semid_ds –sem –Working with semaphores My project References

3. Page 3 Prehistoric Solutions

4. Page 4 Dekker's Algorithm boolean lockA = lockB = false; int turn = A; // or B //Thread A: lockA = true; while(lockB) { if (turn != A) { lockA = false; while(turn != A){} lockA = true; } lockA = false; turn = B; //Thread B: lockB = true; while(lockA) { if (turn != B) { lockB = false; while(turn != B){} lockB = true; } lockB = false; turn = A;

5. Page 5 Peterson's Algorithm boolean lockA = lockB = false; int turn = A; // or B //Thread A: lockA = true; turn = B; while( lockB && turn != A ) { } lockA = false; //Thread B: lockB = true; turn = A; while( lockA && turn != B ) { } lockB = false;

6. Page 6 Synchronization Primitives Completion variables Per-CPU variables Atomic operation Memory barrier Spin lock Seqlocks Read-copy-update Semaphore Local interrupt disabling Local softirq disabling

7. Page 7 Completion Variables Basically a simplified semaphore wait_for_completion complete complete_all Linux Kernel Development (2nd Edition) (Novell Press) (Paperback) by Robert Love http://www.amazon.com/Linux-Kernel-Development- Robert-Love/dp/0672327201http://www.amazon.com/Linux-Kernel-Development- Robert-Love/dp/0672327201 Couldn’t find where the Linux OS actually used these – they may just be for Linux developers

8. Page 8 Synchronization Primitives – cont. Per-CPU Variables –Array of data structures that can be accessed by only 1 CPU –Only useful when data can be split across CPUs –Fails if kernel preemption is enabled (race conditions) –How does this work with the O(1) scheduler? Atomic operation –Read-modify-write oatomic_read(v) oatomic_set(v, i) – (test & set) oatomic_add(i, v) –Another set operates on bit masks –Not all are hardware supported. Some are just locked sections of code.

9. Page 9 Synchronization Primitives – cont. Memory barrier –Avoids code optimization problems by creating a barrier that ensures that all code before the barrier gets executed before code after the barrier Spin locks –Busy wait while blocked –Useful for multi-processor systems when the critical code will execute very fast o1 = unlocked, 0 = locked –Read/Write Spin Locks

10. Page 10 Synchronization Primitives – cont. Seqlocks –Read/Write spin locks that give high priority to writers Read-Copy Update –Allows many readers and writers –No locks or counters –Writer just copies the old data, updates it and changes the pointer –Old data cannot be freed until all readers are finished

11. Page 11 Synchronization Primitives – cont. Semaphores –Kernel and System V IPC (User) –count oIf count > 0 the resource is free ocount = 0 means the resource is 100% utilized but the wait queue is empty oIf count < 0 the resource is 100% utilized and there are processes in the wait queue –wait – pointer to the wait queue –sleepers – flag that is set if there are processes that are waiting blocked –More to come!

12. Page 12 Synchronization Primitives – cont. Local Interrupt Disabling –Creates a critical section in the kernel so that even hardware interrupts won’t break the execution of a set of statements –CPU specific Local softirq disabling –Disable deferrable functions w/o disabling interrupts –Increment the softirq counter to disable; decrement the softirq counter to enable

13. Page 13 ++ Concurrency == ++ Synergy Maximize the number of I/O devices operating –Disable interrupts for very short periods Maximize the number of productive CPUs –Avoid using spin locks if possible

14. Page 14 Sema-what?

15. Page 15 More on Semaphores (semid_ds) /* One semid data structure for each set of semaphores in the system. */ struct semid_ds { struct ipc_perm sem_perm; /* permissions.. see ipc.h */ time_t sem_otime; /* last semop time */ time_t sem_ctime; /* last change time */ struct sem *sem_base; /* ptr to first semaphore in array */ struct wait_queue *eventn; struct wait_queue *eventz; struct sem_undo *undo; /* undo requests on this array */ ushort sem_nsems; /* no. of semaphores in array */ }; sem_perm –This is an instance of the ipc_perm structure, which is defined for us in linux/ipc.h. This holds the permission information for the semaphore set, including the access permissions, and information about the creator of the set (uid, etc). sem_otime –Time of the last semop() operation sem_ctime –Time of the last change to this structure (mode change, etc) sem_base –Pointer to the first semaphore in the array (see next structure) sem_undo –Number of undo requests in this array sem_nsems –Number of semaphores in the semaphore set (the array)

16. Page 16 More on Semaphores (sem) /* One semaphore structure for each semaphore in the system. */ struct sem { short sempid; /* pid of last operation */ ushort semval; /* current value */ ushort semncnt; /* num procs awaiting increase in semval */ ushort semzcnt; /* num procs awaiting semval = 0 */ }; sem_pid –The PID (process ID) that performed the last operation sem_semval –The current value of the semaphore sem_semncnt –Number of processes waiting for resources to become available sem_semzcnt –Number of processes waiting for 100% resource utilization

17. Page 17 Working with Semaphores int semget ( key_t key, int nsems, int semflg ); int semop ( int semid, struct sembuf *sops, unsigned nsops); struct sembuf { ushort sem_num; /* semaphore index in array */ short sem_op; /* semaphore operation */ short sem_flg; /* operation flags */ };

18. Page 18 Working with Semaphores – cont. int semctl ( int semid, int semnum, int cmd, union semun arg ); union semun { int val; /* value for SETVAL */ struct semid_ds *buf; /* buffer for IPC_STAT & IPC_SET */ ushort *array; /* array for GETALL & SETALL */ struct seminfo *__buf; /* buffer for IPC_INFO */ void *__pad; };

19. Page 19 Working with Semaphores – cont. static inline void down(struct semaphore * sem) { might_sleep(); __asm__ __volatile__( "# atomic down operation\n\t" LOCK_PREFIX "decl %0\n\t" /* --sem->count */ "js 2f\n" "1:\n" LOCK_SECTION_START("") "2:\tlea %0,%eax\n\t" "call __down_failed\n\t" "jmp 1b\n" LOCK_SECTION_END :"=m" (sem->count) : :"memory","ax"); }

20. Page 20 Working with Semaphores – cont. //Jump only for contention static inline void up(struct semaphore * sem) { __asm__ __volatile__( "# atomic up operation\n\t" LOCK_PREFIX "incl %0\n\t" /* ++sem->count */ "jle 2f\n" "1:\n" LOCK_SECTION_START("") "2:\tlea %0,%eax\n\t" "call __up_wakeup\n\t" "jmp 1b\n" LOCK_SECTION_END ".subsection 0\n" :"=m" (sem->count) : :"memory","ax"); }

21. Page 21 Semaphore Commands IPC_STAT –Retrieves the semid_ds structure for a set, and stores it in the address of the buf argument in the semun union. IPC_SET –Sets the value of the ipc_perm member of the semid_ds structure for a set. Takes the values from the buf argument of the semun union. IPC_RMID –Removes the set from the kernel. GETALL –Used to obtain the values of all semaphores in a set. The integer values are stored in an array of unsigned short integers pointed to by the array member of the union. GETNCNT –Returns the number of processes currently waiting for resources. GETPID –Returns the PID of the process which performed the last semop call. GETVAL –Returns the value of a single semaphore within the set. GETZCNT –Returns the number of processes currently waiting for 100% resource utilization. SETALL –Sets all semaphore values with a set to the matching values contained in the array member of the union. SETVAL –Sets the value of an individual semaphore within the set to the val member of the union.

22. Page 22 Wait queue Processes are added to the end of the queue –down –down_interuptable –down_trylock Processes are removed from the front of the queue –up oAdd the process to the run queue oNext time the process is scheduled it is removed from the wait queue Processes in the wait queue can be interrupted –Scheduler, when it runs next, will suspend waiting processes that are interrupted

23. Page 23 Project Outline The Big Pictures The changes I made The results

24. Page 24 My project – The Big Picture Releasing processes from the wait queue –Moved to run queue –Leave the rest to the scheduler Why release in FIFO? –Probably the result of a coin flip. Shouldn’t we use the dynamic priorities? –Seems like a good idea. Release All/Most/Some/More than one –Trust that the scheduler will pick the very best one and run it first. –Then the rest can be added back on the wait queue oIn order of their dynamic priorities –sem_queue

25. Page 25 If it Ain’t Broke…Lets find out how it works and tweak it. static inline void up(struct semaphore * sem) { int i; for(i=0; i<10; i++) { __asm__ __volatile__( "# atomic up operation\n\t" LOCK_PREFIX "incl %0\n\t" /* ++sem->count */ "jle 2f\n" "1:\n" LOCK_SECTION_START("") "2:\tlea %0,%eax\n\t" "call __up_wakeup\n\t" "jmp 1b\n" LOCK_SECTION_END ".subsection 0\n" :"=m" (sem->count) : :"memory","ax"); }

26. Page 26 Results The while loop broke it. Played with a for loop. –~10 booted fine –Larger numbers broke it. After doing Student’s t-Test we find that there is a 97% confidence that the before and after data sets came from the same population.

27. Page 27 Results – cont.

28. Page 28 SUM Primary goal: dig into the synchronization in Linux and learn lots about semaphores –Success Secondary goal: alter the semaphore in a way that would leave the computer operational –Success Tertiary goal: pass the class without losing my job, girlfriend, or sanity (with “sanity” being loosely defined) –Pending (I think it’ll turn out ok as long as I get her a nice Christmas present)

29. Page 29 Future work Examine how the scheduler interacts with the wait queue –Rather than release from the front of the queue would it be better to let the scheduler choose which process to release from the queue? oToo much overhead? (Maybe we could even have another O(1) scheduler.) oGain in performance with respect to priority levels?

30. Page 30 References D. Bovet and M. Cesati. Understanding the Linux Kernel, Third Edition.O'Reilly and Associates, Inc., 2002. The Linux Documentation Project: http://www.tldp.org/LDP/lpg/node46.html http://www.tldp.org/LDP/lpg/node46.html R. Kline. Peterson's Algorithm, 2006. http://www.cs.wcupa.edu/~rkline/OS/Peterson.html http://www.cs.wcupa.edu/~rkline/OS/Peterson.html The Linux Tutorial: http://www.linux-tutorial.infohttp://www.linux-tutorial.info http://lxr.linux.no/source/include/asm-i386/semaphore.h http://lxr.linux.no/source/ipc/sem.c Student’s t-Test: http://www.physics.csbsju.edu/stats/t- test.htmlhttp://www.physics.csbsju.edu/stats/t- test.html

31. Page 31 Special Thanks To: Everybody that is still awake The synchronized swimmers from University of Minnesota

32. Page 32 Questions?