d60f152009-04-19Martin Stjernholm Multi-cpu support in Pike ------------------------- This is a draft spec for how to implement multi-cpu support in Pike. The intention is that it gets extended along the way as more issues gets ironed out. Discussions take place in "Pike dev" in LysKOM or pike-devel@lists.lysator.liu.se. Initial draft created 8 Nov 2008 by Martin Stjernholm. Background and goals Pike supports multiple threads, but like many other high-level languages it only allows one thread at a time to access the data structures. This means that the utilization of multi-cpu and multi-core systems remains low, even though there are some modules that can do isolated computational tasks in parallell (e.g. the Image module). It is the so-called "interpreter lock" that must be locked to access any reference variable (i.e. everything except floats and native integers). This lock is held by default in essentially all C code and is explicitly unlocked in a region by the THREADS_ALLOW/ THREADS_DISALLOW macros. On the pike level, the lock is always held - no pike variable can be accessed and no pike function can be called otherwise. The purpose of the multi-cpu support is to rectify this. The design goals are, in order of importance: 1. Pike threads should be able to execute pike code concurrently on multiple cpus as long as they only modify thread local pike data and read a shared pool of static data (i.e. the pike programs, modules and constants). 2. There should be as few internal hot spots as possible (preferably none) when pike code is executed concurrently. Care must be taken to avoid internal synchronization, or updates of shared data that would cause "cache line ping-pong" between cpus. 3. The concurrency should be transparent on the pike level. Pike code should still be able to access shared data without locking and without risking low-level inconsistencies. (So Thread.Mutex etc would still be necessary to achieve higher level synchronization.) 4. There should be tools on the pike level to allow further performance tuning, e.g. lock-free queues, concurrent access hash tables, and the possibility to lock different regions of shared data separately. These tools should be designed so that they are easy to slot into existing code with few changes. 5. There should be tools to monitor and debug concurrency. It should be possible to make assertions that certain objects aren't shared, and that certain access patterns don't cause thread synchronization. This is especially important if goal (3) is realized, since the pike code by itself won't show what is shared and what is thread local. 6. C modules should continue to work without source level modification (but likely without allowing any kind of concurrency). Note that even if goal (3) is accomplished, this is no miracle cure that would make all multithreaded pike programs run with optimal efficiency on multiple cpus. One could expect better concurrency in old code without adaptions, but it could still be hampered considerably by e.g. frequent updates to shared data. Concurrency is a problem that must be taken into account on all levels. Other languages Perl: All data is thread local by default. Data can be explicitly shared, in which case Perl ensures internal consistency. Every shared variable is apparently locked individually. Referencing a thread local variable from a shared one causes the thread to die. See perthrtut(1). Python: Afaik it's the same state of affairs as Pike. Solution overview The basic approach is to divide all data into thread local and shared: o Thread local data is everything that is accessible to one thread only, i.e. there are no references to anything in it from shared data or from any other thread. This is typically data that the current thread has created itself and only reference from the stack. The thread can access its local data without locking. o Shared data is everything that is accessible from more than one thread. Access to it is synchronized using a global read/write lock, the so-called "global lock". I.e. this lock can either be locked for reading by many threads, or be locked by a single thread for writing. Locking the global lock for writing is the same as locking the interpreter lock in current pikes. (This single lock is refined later - see issue "Lock spaces".) o There is also a special case where data can be "disowned", i.e. not shared and not local in any thread. This is used in e.g. Thread.Queue for the objects that are in transit between threads. Disowned data cannot have arbitrary references to it - it must always be under the control of some object that in some way ensures consistency. (Garbage could be made disowned since it by definition no longer is accessible from anywhere, but of course it is always better to clean it up instead.) +--------+ +---------------------+ Direct +--------+ | |<-- refs --| Thread 1 local data |<- - access - -| | | | +---------------------+ | Thread | | | | 1 | | |<- - - - Access through global lock only - - - -| | | Shared | +--------+ | | | data | +---------------------+ Direct +--------+ | |<-- refs --| Thread 2 local data |<- - access - -| | | | +---------------------+ | Thread | | | | 2 | | |<- - - - Access through global lock only - - - -| | | | +--------+ +--------+ ... etc ... The principal use case for this model is that threads can do most of their work with local data and read access to the shared data, and comparatively seldom require the global write lock to update the shared data. Every shared thing does not have its own lock since that would cause excessive lock overhead. Note that the shared data is typically the same as the data referenced from the common environment (i.e. the "global data"). Also note that the current object (this) always is shared in pike modules, so a thread cannot assume free access to it. In other pike classes it would often be shared too, but it is still important to utilize the situation when it is thread local. See issue "Function calls". A thread local thing, and all the things it references directly or indirectly, automatically becomes shared whenever it gets referenced from a shared thing. A shared thing never automatically becomes thread local, but there is a function to explicitly "take" it. It would first have to make sure there are no references to it from shared or other thread local things. Thread.Queue has a special case so that if a thread local thing with no other refs is enqueued, it is disowned by the current thread, and later becomes thread local in the thread that dequeues it. Issue: Lock spaces Having a single global read/write lock for all shared data could become a bottleneck. Thus there is a need for shared data with locks separate from the global lock. Things that share a common lock is called a "lock space", and it is always possible to look up the lock that governs any given thing (see issue "Memory object structure"). A special global lock space, which corresponds to the shared data discussed above, is created on startup. All others have to be created explicitly. The intended use case for lock spaces is a "moderately large" collection of things: Too large and you get outlocking problems, too small and the lock overhead (both execution- and memorywise) gets prohibiting. A typical lock space could be a RAM cache consisting of a mapping and all its content. Many different varieties of lock space locks can be considered, e.g. a simple exclusive access mutex lock or a read/write lock, priority locks, locks that ensure fairness, etc. Therefore different (C-level) implementations should be allowed. One important characteristic of lock space locks is whether they are implicit or explicit: Implicit locks are locked internally, without intervention on the pike level. The lock duration is unspecified; locks are only acquired to ensure internal consistency. All low level data access functions check whether the lock space for the accessed thing is locked already. If it isn't then the lock is acquired automatically. All implicit locks have a well defined lock order (by pointer comparison), and since they only are taken to guarantee internal consistency, an access function can always release a lock to ensure correct order (see also issue "Lock space locking"). Explicit locks are exposed to the pike level and must be locked in a similar way to Thread.Mutex. If a low level data access function encounters an explicit lock that isn't locked, it throws an error. Thus it is left to the pike programmer to avoid deadlocks, but the pike core won't cause any by itself. Since the pike core keeps track which lock governs which thing it ensures that no lock violating access occurs, which is a valuable aid to ensure correctness. One can also consider a variant with a read/write lock space lock that is implicit for read but explicit for write, thus combining atomic pike-level updates with the convenience of implicit locking for read access. The scope of a lock space lock is (at least) the state inside all the things it contains, but not the set of things itself, i.e. things might be added to a lock space without holding a write lock (provided the memory structure allows it). Removing a thing from a lock space always requires the write lock since that is necessary to ensure that a lock actually governs a thing for as long as it is held (regardless it's for reading or writing). FIXME: Allow removing garbage from a lock space without the write lock? See also issues "Memory object structure" and "Lock space locking" for more details. Issue: Memory object structure Of concern are the refcounted memory objects known to the gc. They are called "things", to avoid confusion with "objects" which are the structs for pike objects. There are three types of things: o First class things with ref counter, lock space pointer, and double-linked list pointers (to be able to visit all things in memory, regardless of other references). Most pike visible types are first class things. The exceptions are ints and floats, which are passed by value, and strings and types. o Second class things with ref counter and lock space pointer but no double-linked list pointers. These are always reached through pointers from one or more first class things. It's the job of the visit functions for those first class things to ensure that the gc visits these, thus they don't need the double-linked list pointers. Only strings and types are likely to be of this type. o Third class things contain only a ref counter. They are similar to second class except that their lock spaces are implicit from the referencing things, which means all those things must always be in the same lock space. Thread local things could have NULL as lock space pointer, but as a debug measure they could also point to the thread object so that it's possible to detect bugs with a thread accessing things local to another thread. Before the multi-cpu architecture, all first class things are linked into the same global double-linked lists (one for each type: array, mapping, multiset, object, and program). This gets split into one set of double-linked lists for each thread and for each lock space. That allows things to be added and removed to a thread or lock space without requiring other locks (a lock-free double-linked list is apparently difficult to accomplish). It also allows the gc to do garbage collection locally in each thread and in each lock space (although cyclic structures over several lock spaces won't be freed that way). A global lock-free hash table (see issue "Lock-free hash table") is used to keep track of all lock space lock objects, and hence all things they contain in their double-linked lists. +----------+ +----------+ | Thread 1 | | Thread 2 | +----------+ +----------+ // \\ // \\ // \\ // \\ ,--- O O O O ,------------- O O O O ---. | \\ // \\ // | \\ // \\ // | ref | O O -. | ref O O | ref | | | | v refs ref | v v O <----- O `--> O O O O // \\ // \\ // \\ // \\ refs // \\ // \\ O O -> O O O O O O <----> O O O O \\ // \\ // \\ // \\ // \\ // \\ // +--------------+ +--------------+ +--------------+ | Lock space 1 | | Lock space 2 | | Lock space 3 | +--------------+ +--------------+ +--------------+ ^________ ^ ____^ | | | +-----------------------+-|-+-----+-|-+-------+-|-+----------------- | | X | | X | | X | ... +-----------------------+---+-----+---+-------+---+----------------- Figure 2: "Space Invaders". The O's represent things, and the \\ and // represent the double-linked lists. Some examples of references between things are included, and at the bottom is the global hash table with pointers to all lock spaces. Accessing a lock space lock structure from the global hash table requires a hazard pointer (c.f. issue "Hazard pointers"). Accessing it from a thing is safe if the thread controls at least one ref to the thing, because a lock space has to be empty to delete the lock space lock struct. Issue: Lock space lock semantics There are three types of locks: o A read-safe lock ensures only that the data is consistent, not that it stays constant. This allows lock-free updates in things where possible (which could include arrays, mappings, and maybe even multisets and objects of selected classes). o A read-constant lock ensures both consistency and constantness (i.e. what usually is assumed for a read-only lock). o A write lock ensures complete exclusive access. The owning thread can modify the data, and it can assume no other changes occur to it (barring refcounters - see below). The owning thread can also under limited time leave the data in inconsistent state. This is however still limited by the calls to check_threads(), which means that the state must be consistent again every time the evaluator callbacks are run. See issue "Emulating the interpreter lock". Allowing lock-free updates is attractive, so the standard read/write lock that governs the global lock space will probably be multiple read-safe/single write. An exception to the lock semantics above are the reference counters in refcounted things (c.f. issue "Refcounting and shared data"). A ref to a thing can always be added or removed if it is certain that the thing cannot asynchronously disappear. That means: o Refcount changes must always be atomic, even when a write lock is held. o The refcount may be incremented or decremented when any kind of read lock is held. o The refcount may be incremented or decremented without any kind of lock at all, provided the same thread already holds at least one other ref to the same thing. This means another thread might hold a write lock, but it still won't free the thing since the refcount never can reach zero. o A thing may be freed if its refcount is zero and a write lock is held. FIXME: Whether or not to free a thing if its refcount is zero and only some kind of read lock is held is tricky. To allow that it's necessary to have an atomic-decrement-and-get instruction (can be emulated with CAS, though) to ensure no other thread is decrementing it and reaching zero at the same time. Lock-free linked lists are also necessary to make unlinking possible. Barring that, we need to figure out a policy for scheduling frees of things reaching refcount zero during read locks. Issue: Lock space locking Assuming that a thread already controls at least one ref to a thing (so it won't be freed asynchronously), this is the locking process before accessing it: 1. Read the lock space pointer. If it's NULL then the thing is thread local and nothing more needs to be done. 2. Address an array containing the pointers to the lock spaces that are already locked by the thread. 3. Search for the lock space pointer in the array. If present then nothing more needs to be done. 4. Lock the lock space lock as appropriate. Note that this can imply other implicit locks that are held are unlocked to ensure correct lock order (see issue "Lock spaces"). Then it's added to the array. A thread typically won't hold more than a few locks at any time (less than ten or so), so a plain array and linear search should perform well. For quickest possible access the array should be a static thread local variable (c.f. issue "Thread local storage"). If the array gets full, implicit locks in it can be released automatically to make space. Still, a system where more arrays can be allocated and chained on would perhaps be prudent to avoid the theoretical possibility of running out of space for locked locks. "Controlling" a ref means either to add one "for the stack", or ensuring a lock on a thing that holds a ref. Note that implicit locks might be released in step 4, so unless the thread controls a ref to the referring thing too, it might no longer exist afterwards, and hence the thing itself might be gone. Since implicit locks can be released (almost) at will, they are open for performance tuning: Too long lock durations and they'll outlock other threads, too short and the locking overhead becomes more significant. As a starting point, it seems reasonable to release them at every evaluator callback call (i.e. at approximately every pike function call and return). Issue: Refcounting and shared data Using the traditional refcounting on shared data could easily produce hotspots: Some strings, shared constants, and the object instances for pike modules are often accessed from many threads, so their refcounts would be changed frequently from different processors. E.g. making a single function call in a pike module requires the refcount of the module object to be increased during the call since there is a new reference from a pike_frame. The refcounters in the module objects for commonly used modules like Stdio.pmod/module.pmod could easily become hotspots. Atomic increments and decrements are not enough to overcome this - the memory must not be changed at all to avoid slow synchronizations between cpu local caches. Observation: Refcounters become hotspots primarily in globally accessible shared data, which for the most part has a long lifetime (i.e. programs, module objects, and constants). Otoh, they are most valuable in short-lived data (shared or not), which would produce lots of garbage if they were to be reaped by the gc instead. Following this observation, the problem with refcounter hotspots can to a large degree be mitigated by simply turning off refcounting in the large body of practically static data in the shared runtime environment. A good way to do that is to extend the resolver in the master to mark all programs it compiles, their constants, and the module objects, so that refcounting of them is disabled. To do this, there has to be a function similar to Pike.count_memory that can walk through a structure recursively and mark everything in it. When those things lose their refs, they will always become garbage that only is freed by the gc. Question: Is there data that is missed with this approach? A disabled refcounter is recognized by a negative value and flagged by setting the topmost two bits to one and the rest to zero, i.e. a value in the middle of the negative range. That way, in case there is code that steps the refcounter then it stays negative. (Such code is still bad for performance and should be fixed, though.) Disabling refcounting requires the gc to operate differently; see issue "Garbage collection and external references". Issue: Strings Strings are unique in Pike. This property is hard to keep if threads have local string pools, since a thread local string might become shared at any moment, and thus would need to be moved. Therefore the string hash table remains global, and lock congestion is avoided with some concurrent access hash table implementation. See issue "Lock-free hash table". Lock-free is a good start, but the hash function must also provide a good even distribution to avoid hotspots. Pike currently uses an in-house algorithm (DO_HASHMEM in pike_memory.h). Replacing it with a more widespread and better studied alternative should be considered. There seems to be few that are below O(n) (which DO_HASHMEM is), though. Issue: Types Like strings, types are globally unique and always shared in Pike. That means lock-free access to them is desirable, and it should also be doable fairly easily since they are constant (except for the refcounts which can be updated atomically). Otoh it's probably not as vital as for strings since types typically only are built during compilation. Types are more or less always part of global shared data. That suggests they should have their refcounts disabled most of the time (see issue "Refcounting and shared data"). But again, since types typically only get built during compilation, their refcounts probably won't become hotspots anyway. So it looks like they could be exempt from that rule. Issue: Shared mapping and multiset data blocks An interesting issue is if things like mapping/multiset data blocks should be second or third class things (c.f. issue "Memory object structure"). If they're third class it means copy-on-write behavior doesn't work across lock spaces. If they're second class it means additional overhead handling the lock spaces of the mapping data blocks, and if a mapping data is shared between lock spaces then it has to be in some third lock space of its own, or in the global lock space, neither of which would be very good. So it doesn't look like there's a better way than to botch copy-on-write in this case. Issue: Emulating the interpreter lock For compatibility with old C modules, and for the _disable_threads function, it is necessary to retain a complete lock like the current interpretator lock. It has to lock the global area for writing, and also stop all access to all lock spaces, since the thread local data might refer to any lock space. This lock is implemented as a read/write lock, which normally is held permanently for reading by all threads. Only when a thread is waiting to acquire the compat interpreter lock is it released as each thread goes into check_threads(). This lock cannot wait for explicit lock space locks to be released. Thus it can override the assumption that a lock space is safe from tampering by holding a write lock on it. Still, it's only available from the C level (with the exception of _disable_threads) so the situation is not any different from the way the interpreter lock overrides Thread.Mutex today. Issue: Function calls A lock on an object is almost always necessary before calling a function in it. Therefore the central apply function (mega_apply) must ensure an appropriate lock is taken. Which kind of lock (read-safe/read-constant/write - see issue "Lock space lock semantics") depends on what the function wants to do. Therefore all object functions are extended with flags for this. The best default is probably read-safe. Flags for no locking (for the few special cases where the implementations actually are completely lock-free) and for compat-interpreter-lock-locking should probably exist as well. A compat-interpreter-lock flag is also necessary for global functions that don't have a "this" object (aka efuns). Having the required locking declared this way also alleviates each function from the burden of doing the locking to access the current storage, and it allows future compiler optimizations to minimize lock operations. Issue: Exceptions "Forgotten" locks after exceptions shouldn't be a problem: Explicit locks are handled just like today (i.e. it's up to the pike programmer), and implicit locks can safely be released when an exception is thrown. One case requires attention: An old-style function that requires the compat interpreter lock might catch an error. In that case the error system has to ensure that lock is reacquired. Issue: C module interface A new add_function variant is probably added for new-style functions. It takes bits for the flags discussed for issue "Function calls". New-style functions can only assume free access to the current storage according to those flags; everything else must be locked (through a new set of macros/functions). Accessor functions for data types (e.g. add_shared_strings, mapping_lookup, and object_index_no_free) handles the necessary locking internally. They will only assume that the thing is safe, i.e. that the caller ensures the current thread controls at least one ref. THREADS_ALLOW/THREADS_DISALLOW and their likes are not used in new-style functions. There will be new GC callbacks for walking module global pointers to things (see issue "Garbage collection and external references"). Issue: C module compatibility Ref issue "Emulating the interpreter lock". Ref issue "Garbage collection and external references". Issue: Garbage collection and external references The current gc design is that there is an initial "check" pass that determines external references by counting all internal references, and then for each thing subtract it from its refcount. If the result isn't zero then there are external references (e.g. from global C variables or from the C stack) and the thing is not garbage. Since refcounting can be disabled in some objects (see issue "Refcounting and shared data"), this approach no longer work; the gc has to be changed to find external references some other way: References from global C variables are few, so they can be dealt with by requiring C modules and the core parts to provide callbacks that lets the gc walk through them (see issue "C module interface"). This is however not compatible with old C modules. References from C stacks are common, and it is infeasible to require callbacks that keep track of them. The gc instead has to scan the C stacks for the threads and treat any aligned machine word containing an apparently valid pointer to a known thing as an external reference. This is the common approach used by standalone gc libraries that don't require application support. For reference, here is one such garbage collector, written in C++: http://developer.apple.com/DOCUMENTATION/Cocoa/Conceptual/GarbageCollection/Introduction.html#//apple_ref/doc/uid/TP40002427 Its source is here: http://www.opensource.apple.com/darwinsource/10.5.5/autozone-77.1/ The same approach is also necessary to cope with old C modules (see issue "C module compatibility"), but since global C level pointers are few, it might not be mandatory to get this working. Btw, using this approach to find external refs should be considerably more efficient than the old "check" pass, even if C stacks are scanned wholesale. Issue: Local garbage collection Each thread periodically invokes a gc that only looks for garbage in the local data of that thread. This can naturally be done without disturbing the other threads. It follows that this gc also can be disabled on a per-thread basis. This is a reason for keeping thread local data in separate double-linked lists (see issue "Memory object structure"). Similarly, if gc statistics are added to each lock space, they could also be gc'd for internal garbage at appropriate times when they get write locked by some thread. That might be interesting since known cyclic structures could then be put in lock spaces of their own and be gc'd efficiently without a global gc. Note that a global gc is still required to clean up cycles with things in more than one lock space. Issue: Global pike level caches Issue: Thread.Queue A lock-free implementation should be used. The things in the queue are typically disowned to allow them to become thread local in the reading thread. Issue: False sharing False sharing occurs when thread local things used frequently by different threads are next to each other so that they share the same cache line. Thus the cpu caches might force frequent resynchronization of the cache line even though there is no apparent hotspot problem on the C level. This can be a problem in particular for all the block_alloc pools containing small structs. Using thread local pools is seldom a workable solution since most thread local structs might become shared later on. One way to avoid it is to add padding (and alignment). Cache line sizes are usually 64 bytes or less (at least for Intel ia32). That should be small enough to make this viable in many cases. FIXME: Check cache line sizes on the other important architectures. Worth noting that the problem is greatest for the frequently changed ref counters at the start of each thing, so the most important thing is to keep ref counters separated. I.e. things larger than a cache line can probably be packed without padding. Another way is to move things when they get shared, but that is pretty complicated and slow. Issue: Malloc and block_alloc Standard OS mallocs are usually locking. Bundling a lock-free one could be important. FIXME: Survey free implementations. Block_alloc is a simple homebrew memory manager used in several different places to allocate fixed-size blocks. The block_alloc pools are often shared, so they must allow efficient concurrent access. With a modern malloc, it is possible that the need for block_alloc is gone, or perhaps the malloc lib has builtin support for fixed-size pools. Making a lock-free implementation is nontrivial, so the homebrew ought to be ditched in any case. A problem with ditching block_alloc is that there is some code that walks through all allocated blocks in a pool, and also avoids garbage by freeing the whole pool altogether. FIXME: Investigate alternatives here. See also issue "False sharing". Issue: The compiler Issue: Foreign thread visits JVM threads.. Issue: Pike security system It is possible that keeping the pike security system intact would complicate the implementation, and even if it was kept intact a lot of testing would be required before one can be confident that it really works (and there are currently very few tests for it in the test suite). Also, the security system isn't used at all to my (mast's) knowledge, and it is not even compiled in by default (has to be enabled with a configure flag). All this leads to the conclusion that it is easiest to ignore the security system altogether, and if possible leave it as it is with the option to get it working later. Issue: Contention-free counters There is probably a need for contention-free counters in several different areas. They should be possible to update from several threads in parallell without synchronization. Querying the current count is always approximate since it can be changing simultaneously in other threads. However, the thread's own local count is always accurate. They should be separated from the blocks they apply to, to avoid cache line invalidation of those blocks. To accomplish that, a generic tool somewhat similar to block_alloc is created that allocates one or more counter blocks for each thread. In these blocks indexes are allocated, so a counter is defined by the same index into all the thread local counter blocks. Each thread can then modify its own counters without locking, and it typically has its own counter blocks in the local cache while the corresponding main memory is marked invalid. To query a counter, a thread would need to read the blocks for all other threads. This means that these counters are efficient for updates but less so for queries. However, since queries always are approximate, it is possible to cache them for some time (e.g. 1 ms). Each thread would need its own cache though, since the local count cannot be cached. It should be lock-free for allocating and freeing counters, and preferably also for starting and stopping threads (c.f. issue "Foreign thread visits"). In both cases the freeing steps represents a race problem - see issue "Hazard pointers". To free counters, the counter index would constitute the hazard pointer. Issue: Lock-free hash table A good lock-free hash table implementation is necessary. A promising one is http://blogs.azulsystems.com/cliff/2007/03/a_nonblocking_h.html. It requires a CAS (Compare And Swap) instruction to work, but that shouldn't be a problem. The java implementation (http://sourceforge.net/projects/high-scale-lib) is Public Domain. In the comments there is talk about efforts to make a C version. It supports (through putIfAbsent) the uniqueness requirement for strings, i.e. if several threads try to add the same string (at different addresses) then all will end up with the same string pointer afterwards. The java implementation relies on the gc to free up the old hash tables after resize. We don't have that convenience, but the problem is still solvable; see issue "Hazard pointers". Issue: Hazard pointers A problem with most lock-free algorithms is how to know no other thread is accessing a block that is about to be freed. Another is the ABA problem which can occur when a block is freed and immediately allocated again (common for block_alloc). Hazard pointers are a good way to solve these problems without leaving the blocks to the garbage collector (see http://www.research.ibm.com/people/m/michael/ieeetpds-2004.pdf). So a generic hazard pointer tool is necessary. Note however that a more difficult variant of the ABA problem still can occur when the block cannot be freed after leaving the data structure. (In the canonical example with a lock-free stack - see e.g. "ABA problem" in Wikipedia - consider the case when A is a thing that continues to live on and actually gets pushed back.) The only reliable way to cope with that is probably to use wrappers. Issue: Thread local storage Implementation would be considerably simpler if working TLS can be assumed on the C level, through the __thread keyword (or __declspec(thread) in Visual C++). A survey of the support for TLS in common compilers and OS'es is needed to decide whether this is an workable assumption: o GCC: __thread is supported. Source: Wikipedia. FIXME: Check from which version. o Visual C++: __declspec(thread) is supported. Source: Wikipedia. FIXME: Check from which version. o Intel C compiler: Support exists. Source: Wikipedia. FIXME: Check from which version. o Sun C compiler: Support exists. Source: Wikipedia. FIXME: Check from which version. o Linux (i386, x86_64, sparc32, sparc64): TLS is supported and works for dynamic libs. C.f. http://people.redhat.com/drepper/tls.pdf. FIXME: Check from which version of glibc and kernel (if relevant). o Windows (i386, x86_64): TLS is supported but does not always work in dll's loaded using LoadLibrary (which means all dynamic modules in pike). C.f. http://msdn.microsoft.com/en-us/library/2s9wt68x.aspx. According to Wikipedia this is fixed in Vista and Server 2008 (FIXME: verify). In any case, TLS is still usable in the pike core. o MacOS X: FIXME: Check this. o Solaris: FIXME: Check this. o *BSD: FIXME: Check this. Issue: Platform specific primitives Some low-level primitives, such as CAS and fences, are necessary to build the various lock-free tools. A third-party library would be useful. An effort to make a standardized library is here: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2006/n2047.html (C level interface at the end). It apparently lacks implementation, though. Required operations: CAS(address, old_value, new_value) Compare-and-set: Atomically sets *address to new_value iff its current value is old_value. Needed for 32-bit variables, and on 64-bit systems also for 64-bit variables. ATOMIC_INC(address) ATOMIC_DEC(address) Increments/decrements *address atomically. Can be simulated with CAS. 32-bit version necessary, 64-bit version would be nice. LFENCE() Load fence: All memory reads in the thread before this point are guaranteed to be done (i.e. be globally visible) before any following it. SFENCE() Store fence: All memory writes in the thread before this point are guaranteed to be done before any following it. MFENCE() Memory fence: Both load and store fence at the same time. (On many architectures this is implied by CAS etc, but we shouldn't assume that.) The following operations are uncertain - still not known if they're useful and supported enough to be required, or if it's better to do without them: CASW(address, old_value_low, old_value_high, new_value_low, new_value_high) A compare-and-set that works on a double pointer size area. Supported on more modern x86 and x86_64 processors (c.f. http://en.wikipedia.org/wiki/Compare-and-swap#Extensions). FIXME: More.. Survey of platform support: o Windows/Visual Studio: Got "Interlocked Variable Access": http://msdn.microsoft.com/en-us/library/ms684122.aspx o FIXME: More.. Various links Pragmatic nonblocking synchronization for real-time systems http://www.usenix.org/publications/library/proceedings/usenix01/full_papers/hohmuth/hohmuth_html/index.html DCAS is not a silver bullet for nonblocking algorithm design http://portal.acm.org/citation.cfm?id=1007945