260 likes | 368 Vues
Read-Copy Update P. E. McKenney, J. Appavoo, A. Kleen, O. Krieger, R. Russell, D. Saram, M. Soni Ottawa Linux Symposium 2001 Presented by Bogdan Simion. Motivation. Locking can be expensive Overhead of locking code Cache bouncing
E N D
Read-Copy Update P. E. McKenney, J. Appavoo, A. Kleen, O. Krieger, R. Russell, D. Saram, M. Soni Ottawa Linux Symposium 2001 Presented by Bogdan Simion
Motivation • Locking can be expensive • Overhead of locking code • Cache bouncing • Linux uses locking to protect against infrequent destructive modifications • e.g., racy accesses to unloaded modules • Want to avoid locking expense for reads of data that are infrequently modified
Key Idea • Example: module unloading • Give ongoing operations a grace period to finish
Grace Periods • Starts when new operations see new state • e.g., remove pointer to module from a list • First phase of the update • No new references made once period starts • Extends until after all operations that started before the grace period finish • Operations with outstanding references finish safely • When period ends, system may cleanup • Second phase of the update, e.g., free module data
Grace Period Duration • Safe to end the grace period when all CPUs have finished prior operations • A non-preemptive operating system finishes all operations when it context switches • Thus, grace period ends after all CPUs have context switched at least once • Zero reference count deduced without using any shared data!
RCU So Far • RCU performs updates in two phases: • Update enough so new operations see new state but old operations can proceed using old state • Complete the update after the grace period • RCU works well when • Updates can be done in two phases • Operations still work with stale state • Destructive updates are infrequent • Let's look at an example of how it's used
Example: Reference Counted Search • Simple circular doubly linked-list • Compare a reference-counting locking algorithm taken from Linux with its read-copy-update equivalent
Reference Counted Search • For each algorithm: • search() • delete() • search(): returns a pointer to an element in the list given its addr, and ensures that element is not being freed up • delete(): arranges for the specified element to eventually be freed up • delete() may not be able to free the element immediately due to concurrent searches
Reference-Counted Usage • Read-only and update (including delete) operation
Search / Delete Discussion • Searching scales perfectly • No locks – scales well • No cache line bouncing • Clear advantage over reference counting • Search can return stale data • There is a race between search and delete • Reference counting + locks does not have this problem • Delete is similar – global lock • Good speedups only if many more searches than del • kfree_rcu is neither trivial nor inexpensive
Read-Copy Deletion Scenario • To delete element B, the updater task acquires list lock to exclude other list manipulation, unlinks element B from the list and releases list lock
Read-Copy Deletion Scenario • The updater task passes a pointer to B to the kfree_rcu() primitive, which adds the memory to a list waiting to be freed. • Safe to return B to the freelist at the end of the grace period (when all pre-existing ops complete) List After Element B Returned to Freelist List After Grace Period
Implementing kfree_rcu • Basic idea:
Implementing kfree_rcu • Execute updater on each CPU:
Implementing kfree_rcu • Delay deletion until the end of the grace period: • wait_for_rcu() { • ... current->cpus_allowed = (1 << num_cpus) - 1; • while(true) { • current->cpus_allowed &= ~(1 << cpu_index()); • if (current->cpus_allowed == 0) break; schedule(); • } /* Grace period now over. Now it's safe to delete. */ ... • }
Implementing kfree_rcu • Delay deletion until the end of the grace period: • wait_for_rcu() { • ... current->cpus_allowed = (1 << num_cpus) - 1; • while(true) { • current->cpus_allowed &= ~(1 << cpu_index()); • if (current->cpus_allowed == 0) break; schedule(); • } /* Grace period now over. Now it's safe to delete. */ ... • } • Doesn't work with preemptible kernels. Why? • Can't be called from an interrupt handler or while a spin lock is held. Why? • Can be relatively slow. Why?
Deferring wait_for_rcu struct rcu_head { tq_struct task; }; • void* kmalloc_rcu(size_t size, int flags) { • rcu_head* ret = kmalloc(size + sizeof(*ret), flags); • return ret + 1; • } • void sync_and_destroy(void* head) { • wait_for_rcu(); • kfree(head); • } • kfree_rcu(void* obj) { • rcu_head* head = ((rcu_head*) obj) – 1; • head->task.routine = &sync_and_destroy; • head->task.data = head; • schedule_task(&head->task); • }
Deferring wait_for_rcu struct rcu_head { tq_struct task; }; • void* kmalloc_rcu(size_t size, int flags) { • rcu_head* ret = kmalloc(size + sizeof(*ret), flags); • return ret + 1; • } • void sync_and_destroy(void* head) { • wait_for_rcu(); • kfree(head); • } • kfree_rcu(void* obj) { • rcu_head* head = ((rcu_head*) obj) – 1; • head->task.routine = &sync_and_destroy; • head->task.data = head; • schedule_task(&head->task); • }
Deferring wait_for_rcu struct rcu_head { tq_struct task; }; • void* kmalloc_rcu(size_t size, int flags) { • rcu_head* ret = kmalloc(size + sizeof(*ret), flags); • return ret + 1; • } • void sync_and_destroy(void* head) { • wait_for_rcu(); • kfree(head); • } • kfree_rcu(void* obj) { • rcu_head* head = ((rcu_head*) obj) – 1; • head->task.routine = &sync_and_destroy; • head->task.data = head; • schedule_task(&head->task); • } Why is kmalloc_rcu necessary?
RCU Application: File Descriptors • Kernel maintains mapping of file descriptors to instances of struct file with an array • Expansion of the array is a destructive update: • Copies the old elements into a new array • Updates pointers and deletes the old array • RCU employed: • Phase 1: Create new arrays and update pointers • Phase 2: Delete the old arrays
RCU Performance: File Descriptors • Chat benchmark, 2.4.2 SMP Kernel • Why does R/W lock incur so much overhead?
RCU Performance Improvements • A number of improvements to the basic mechanism • Batch grace period measurements • wait_for_rcu is expensive • A single measurement satisfies multiple deferred free requests • Maintain per-CPU request lists • Faster grace period algorithm • See the paper for details
Comparing RCU to other Locking Algorithms • Data locking • Does not avoid reader locks • Also prone to deadlocks • Although list elements can be manipulated in parallel, searches cannot be done in parallel • Can be used to prevent stale reads in RCU • brlock • Effectively lock-free reads • Not clear how its performance differs from RCU • i.e., Can't brlock be used for the file descriptor arrays?
Conclusions • RCU is an effective approach for avoiding locking for read-mostly data structures • An elegant method for implicit reference counting • Main advantage: readers need not acquire locks, perform any atomic ops, write to shared memory or use barriers. • The destructive update is delayed until the grace period finishes – until all CPUs context switch (if non-preemptible) • Since 2001, it has been used in hundreds of places in the Linux kernel