Using RCU (Read-Copy-Update) for synchronization ================================================ Read-copy update (RCU) is a synchronization mechanism that is used to protect read-mostly data structures. RCU is very efficient and scalable on the read side (it is wait-free), and thus can make the read paths extremely fast. RCU supports concurrency between a single writer and multiple readers, thus it is not used alone. Typically, the write-side will use a lock to serialize multiple updates, but other approaches are possible (e.g., restricting updates to a single task). In QEMU, when a lock is used, this will often be the "iothread mutex", also known as the "big QEMU lock" (BQL). Also, restricting updates to a single task is done in QEMU using the "bottom half" API. RCU is fundamentally a "wait-to-finish" mechanism. The read side marks sections of code with "critical sections", and the update side will wait for the execution of all *currently running* critical sections before proceeding, or before asynchronously executing a callback. The key point here is that only the currently running critical sections are waited for; critical sections that are started _after_ the beginning of the wait do not extend the wait, despite running concurrently with the updater. This is the reason why RCU is more scalable than, for example, reader-writer locks. It is so much more scalable that the system will have a single instance of the RCU mechanism; a single mechanism can be used for an arbitrary number of "things", without having to worry about things such as contention or deadlocks. How is this possible? The basic idea is to split updates in two phases, "removal" and "reclamation". During removal, we ensure that subsequent readers will not be able to get a reference to the old data. After removal has completed, a critical section will not be able to access the old data. Therefore, critical sections that begin after removal do not matter; as soon as all previous critical sections have finished, there cannot be any readers who hold references to the data structure, and these can now be safely reclaimed (e.g., freed or unref'ed). Here is a picutre: thread 1 thread 2 thread 3 ------------------- ------------------------ ------------------- enter RCU crit.sec. | finish removal phase | begin wait | | enter RCU crit.sec. exit RCU crit.sec | | complete wait | begin reclamation phase | exit RCU crit.sec. Note how thread 3 is still executing its critical section when thread 2 starts reclaiming data. This is possible, because the old version of the data structure was not accessible at the time thread 3 began executing that critical section. RCU API ======= The core RCU API is small: void rcu_read_lock(void); Used by a reader to inform the reclaimer that the reader is entering an RCU read-side critical section. void rcu_read_unlock(void); Used by a reader to inform the reclaimer that the reader is exiting an RCU read-side critical section. Note that RCU read-side critical sections may be nested and/or overlapping. void synchronize_rcu(void); Blocks until all pre-existing RCU read-side critical sections on all threads have completed. This marks the end of the removal phase and the beginning of reclamation phase. Note that it would be valid for another update to come while synchronize_rcu is running. Because of this, it is better that the updater releases any locks it may hold before calling synchronize_rcu. If this is not possible (for example, because the updater is protected by the BQL), you can use call_rcu. void call_rcu1(struct rcu_head * head, void (*func)(struct rcu_head *head)); This function invokes func(head) after all pre-existing RCU read-side critical sections on all threads have completed. This marks the end of the removal phase, with func taking care asynchronously of the reclamation phase. The foo struct needs to have an rcu_head structure added, perhaps as follows: struct foo { struct rcu_head rcu; int a; char b; long c; }; so that the reclaimer function can fetch the struct foo address and free it: call_rcu1(&foo.rcu, foo_reclaim); void foo_reclaim(struct rcu_head *rp) { struct foo *fp = container_of(rp, struct foo, rcu); g_free(fp); } For the common case where the rcu_head member is the first of the struct, you can use the following macro. void call_rcu(T *p, void (*func)(T *p), field-name); void g_free_rcu(T *p, field-name); call_rcu1 is typically used through these macro, in the common case where the "struct rcu_head" is the first field in the struct. If the callback function is g_free, in particular, g_free_rcu can be used. In the above case, one could have written simply: g_free_rcu(&foo, rcu); typeof(*p) atomic_rcu_read(p); atomic_rcu_read() is similar to atomic_mb_read(), but it makes some assumptions on the code that calls it. This allows a more optimized implementation. atomic_rcu_read assumes that whenever a single RCU critical section reads multiple shared data, these reads are either data-dependent or need no ordering. This is almost always the case when using RCU, because read-side critical sections typically navigate one or more pointers (the pointers that are changed on every update) until reaching a data structure of interest, and then read from there. RCU read-side critical sections must use atomic_rcu_read() to read data, unless concurrent writes are presented by another synchronization mechanism. Furthermore, RCU read-side critical sections should traverse the data structure in a single direction, opposite to the direction in which the updater initializes it. void atomic_rcu_set(p, typeof(*p) v); atomic_rcu_set() is also similar to atomic_mb_set(), and it also makes assumptions on the code that calls it in order to allow a more optimized implementation. In particular, atomic_rcu_set() suffices for synchronization with readers, if the updater never mutates a field within a data item that is already accessible to readers. This is the case when initializing a new copy of the RCU-protected data structure; just ensure that initialization of *p is carried out before atomic_rcu_set() makes the data item visible to readers. If this rule is observed, writes will happen in the opposite order as reads in the RCU read-side critical sections (or if there is just one update), and there will be no need for other synchronization mechanism to coordinate the accesses. The following APIs must be used before RCU is used in a thread: void rcu_register_thread(void); Mark a thread as taking part in the RCU mechanism. Such a thread will have to report quiescent points regularly, either manually or through the QemuCond/QemuSemaphore/QemuEvent APIs. void rcu_unregister_thread(void); Mark a thread as not taking part anymore in the RCU mechanism. It is not a problem if such a thread reports quiescent points, either manually or by using the QemuCond/QemuSemaphore/QemuEvent APIs. Note that these APIs are relatively heavyweight, and should _not_ be nested. DIFFERENCES WITH LINUX ====================== - Waiting on a mutex is possible, though discouraged, within an RCU critical section. This is because spinlocks are rarely (if ever) used in userspace programming; not allowing this would prevent upgrading an RCU read-side critical section to become an updater. - atomic_rcu_read and atomic_rcu_set replace rcu_dereference and rcu_assign_pointer. They take a _pointer_ to the variable being accessed. - call_rcu is a macro that has an extra argument (the name of the first field in the struct, which must be a struct rcu_head), and expects the type of the callback's argument to be the type of the first argument. call_rcu1 is the same as Linux's call_rcu. RCU PATTERNS ============ Many patterns using read-writer locks translate directly to RCU, with the advantages of higher scalability and deadlock immunity. In general, RCU can be used whenever it is possible to create a new "version" of a data structure every time the updater runs. This may sound like a very strict restriction, however: - the updater does not mean "everything that writes to a data structure", but rather "everything that involves a reclamation step". See the array example below - in some cases, creating a new version of a data structure may actually be very cheap. For example, modifying the "next" pointer of a singly linked list is effectively creating a new version of the list. Here are some frequently-used RCU idioms that are worth noting. RCU list processing ------------------- TBD (not yet used in QEMU) RCU reference counting ---------------------- Because grace periods are not allowed to complete while there is an RCU read-side critical section in progress, the RCU read-side primitives may be used as a restricted reference-counting mechanism. For example, consider the following code fragment: rcu_read_lock(); p = atomic_rcu_read(&foo); /* do something with p. */ rcu_read_unlock(); The RCU read-side critical section ensures that the value of "p" remains valid until after the rcu_read_unlock(). In some sense, it is acquiring a reference to p that is later released when the critical section ends. The write side looks simply like this (with appropriate locking): qemu_mutex_lock(&foo_mutex); old = foo; atomic_rcu_set(&foo, new); qemu_mutex_unlock(&foo_mutex); synchronize_rcu(); free(old); If the processing cannot be done purely within the critical section, it is possible to combine this idiom with a "real" reference count: rcu_read_lock(); p = atomic_rcu_read(&foo); foo_ref(p); rcu_read_unlock(); /* do something with p. */ foo_unref(p); The write side can be like this: qemu_mutex_lock(&foo_mutex); old = foo; atomic_rcu_set(&foo, new); qemu_mutex_unlock(&foo_mutex); synchronize_rcu(); foo_unref(old); or with call_rcu: qemu_mutex_lock(&foo_mutex); old = foo; atomic_rcu_set(&foo, new); qemu_mutex_unlock(&foo_mutex); call_rcu(foo_unref, old, rcu); In both cases, the write side only performs removal. Reclamation happens when the last reference to a "foo" object is dropped. Using synchronize_rcu() is undesirably expensive, because the last reference may be dropped on the read side. Hence you can use call_rcu() instead: foo_unref(struct foo *p) { if (atomic_fetch_dec(&p->refcount) == 1) { call_rcu(foo_destroy, p, rcu); } } Note that the same idioms would be possible with reader/writer locks: read_lock(&foo_rwlock); write_mutex_lock(&foo_rwlock); p = foo; p = foo; /* do something with p. */ foo = new; read_unlock(&foo_rwlock); free(p); write_mutex_unlock(&foo_rwlock); free(p); ------------------------------------------------------------------ read_lock(&foo_rwlock); write_mutex_lock(&foo_rwlock); p = foo; old = foo; foo_ref(p); foo = new; read_unlock(&foo_rwlock); foo_unref(old); /* do something with p. */ write_mutex_unlock(&foo_rwlock); read_lock(&foo_rwlock); foo_unref(p); read_unlock(&foo_rwlock); foo_unref could use a mechanism such as bottom halves to move deallocation out of the write-side critical section. RCU resizable arrays -------------------- Resizable arrays can be used with RCU. The expensive RCU synchronization (or call_rcu) only needs to take place when the array is resized. The two items to take care of are: - ensuring that the old version of the array is available between removal and reclamation; - avoiding mismatches in the read side between the array data and the array size. The first problem is avoided simply by not using realloc. Instead, each resize will allocate a new array and copy the old data into it. The second problem would arise if the size and the data pointers were two members of a larger struct: struct mystuff { ... int data_size; int data_alloc; T *data; ... }; Instead, we store the size of the array with the array itself: struct arr { int size; int alloc; T data[]; }; struct arr *global_array; read side: rcu_read_lock(); struct arr *array = atomic_rcu_read(&global_array); x = i < array->size ? array->data[i] : -1; rcu_read_unlock(); return x; write side (running under a lock): if (global_array->size == global_array->alloc) { /* Creating a new version. */ new_array = g_malloc(sizeof(struct arr) + global_array->alloc * 2 * sizeof(T)); new_array->size = global_array->size; new_array->alloc = global_array->alloc * 2; memcpy(new_array->data, global_array->data, global_array->alloc * sizeof(T)); /* Removal phase. */ old_array = global_array; atomic_rcu_set(&new_array->data, new_array); synchronize_rcu(); /* Reclamation phase. */ free(old_array); } SOURCES ======= * Documentation/RCU/ from the Linux kernel