/* Multiplex file i/o (C) 2015-2017 Niall Douglas (9 commits) File Created: Dec 2015 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License in the accompanying file Licence.txt or at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. Distributed under the Boost Software License, Version 1.0. (See accompanying file Licence.txt or copy at http://www.boost.org/LICENSE_1_0.txt) */ #ifndef LLFIO_IO_SERVICE_H #define LLFIO_IO_SERVICE_H #include "handle.hpp" #include #include #include #undef _threadid // windows macro splosh sigh //! \file io_service.hpp Provides io_service. // Need to decide which kind of POSIX AIO to use #ifndef _WIN32 // Right now the only thing we support is POSIX AIO #if !defined(LLFIO_USE_POSIX_AIO) /*! \brief Undefined to autodetect, 1 to use POSIX AIO, 0 to not use \warning On FreeBSD the AIO kernel module needs to be loaded for POSIX AIO to work. Run as root 'kldload aio' or add 'aio_load=YES' in loader.conf. */ #define LLFIO_USE_POSIX_AIO 1 #endif // BSD kqueues not implemented yet //# if defined(__FreeBSD__) && !defined(LLFIO_COMPILE_KQUEUES) //# define LLFIO_COMPILE_KQUEUES 1 //# endif #if LLFIO_COMPILE_KQUEUES #if defined(LLFIO_USE_POSIX_AIO) && !LLFIO_USE_POSIX_AIO #error BSD kqueues must be combined with POSIX AIO! #endif #if !defined(LLFIO_USE_POSIX_AIO) #define LLFIO_USE_POSIX_AIO 1 #endif #endif #if DOXYGEN_SHOULD_SKIP_THIS //! Undefined to autodetect, 1 to compile in BSD kqueue support, 0 to leave it out #define LLFIO_COMPILE_KQUEUES 0 #endif #if LLFIO_USE_POSIX_AIO // We'll be using POSIX AIO and signal based interruption for post() #include // Do we have realtime signals? #if !defined(LLFIO_HAVE_REALTIME_SIGNALS) && defined(_POSIX_RTSIG_MAX) && defined(SIGRTMIN) #ifndef LLFIO_IO_POST_SIGNAL #define LLFIO_IO_POST_SIGNAL (-1) #endif #define LLFIO_HAVE_REALTIME_SIGNALS 1 #else #ifndef LLFIO_IO_POST_SIGNAL //! Undefined to autoset to first free SIGRTMIN if realtime signals available, else SIGUSR1. Only used if LLFIO_USE_KQUEUES=0. #define LLFIO_IO_POST_SIGNAL (SIGUSR1) #endif //! Undefined to autodetect. 0 to use non-realtime signals. Note performance in this use case is abysmal. #define LLFIO_HAVE_REALTIME_SIGNALS 0 #endif struct aiocb; #endif #endif #ifdef _MSC_VER #pragma warning(push) #pragma warning(disable : 4251) // dll interface #endif LLFIO_V2_NAMESPACE_EXPORT_BEGIN class io_service; class async_file_handle; /*! \class io_service \brief An asynchronous i/o multiplexer service. This service is used in conjunction with `async_file_handle` to multiplex initating i/o and completing it onto a single kernel thread. Unlike the `io_service` in ASIO or the Networking TS, this `io_service` is much simpler, in particular it is single threaded per instance only i.e. you must run a separate `io_service` instance one per kernel thread if you wish to run i/o processing across multiple threads. LLFIO does not do this for you (and for good reason, unlike socket i/o, it is generally unwise to distribute file i/o across kernel threads due to the much more code executable between user space and physical storage i.e. keeping processing per CPU core hot in cache delivers outsize benefits compared to socket i/o). Furthermore, you cannot use this i/o service in any way from any thread other than where it was created. You cannot call its `run()` from any thread other than where it was created. And you cannot initiate i/o on an `async_file_handle` from any thread other than where its owning i/o service was created. In other words, keep your i/o service and all associated file handles on their owning thread. The sole function you can call from another thread is `post()` which lets you execute some callback in the `run()` of the owning thread. This lets you schedule i/o from other threads if you really must do that. \snippet coroutines.cpp coroutines_example */ class LLFIO_DECL io_service { friend class async_file_handle; public: //! The file extent type used by this i/o service using extent_type = io_handle::extent_type; //! The memory extent type used by this i/o service using size_type = io_handle::size_type; //! The scatter buffer type used by this i/o service using buffer_type = io_handle::buffer_type; //! The gather buffer type used by this i/o service using const_buffer_type = io_handle::const_buffer_type; //! The scatter buffers type used by this i/o service using buffers_type = io_handle::buffers_type; //! The gather buffers type used by this i/o service using const_buffers_type = io_handle::const_buffers_type; //! The i/o request type used by this i/o service template using io_request = io_handle::io_request; //! The i/o result type used by this i/o service template using io_result = io_handle::io_result; private: #ifdef _WIN32 win::handle _threadh{}; win::dword _threadid; #else pthread_t _threadh; #endif std::mutex _posts_lock; struct post_info { io_service *service; detail::function_ptr f; post_info(io_service *s, detail::function_ptr _f) : service(s) , f(std::move(_f)) { } }; std::deque _posts; using shared_size_type = std::atomic; shared_size_type _work_queued; #if LLFIO_USE_POSIX_AIO bool _use_kqueues; #if LLFIO_COMPILE_KQUEUES int _kqueueh; #endif std::vector _aiocbsv; // for fast aio_suspend() #endif public: // LOCK MUST BE HELD ON ENTRY! void __post_done(post_info *pi) { // Find the post_info and remove it for(auto &i : _posts) { if(&i == pi) { i.f.reset(); i.service = nullptr; pi = nullptr; break; } } assert(!pi); if(pi != nullptr) { abort(); } _work_done(); while(!_posts.empty() && (_posts.front().service == nullptr)) { _posts.pop_front(); } } void _post_done(post_info *pi) { std::lock_guard g(_posts_lock); return __post_done(pi); } void _work_enqueued(size_type i = 1) { _work_queued += i; } void _work_done() { --_work_queued; } /*! Creates an i/o service for the calling thread, installing a global signal handler via set_interruption_signal() if not yet installed if on POSIX and BSD kqueues not in use. */ LLFIO_HEADERS_ONLY_MEMFUNC_SPEC io_service(); io_service(io_service &&) = delete; io_service(const io_service &) = delete; io_service &operator=(io_service &&) = delete; io_service &operator=(const io_service &) = delete; LLFIO_HEADERS_ONLY_VIRTUAL_SPEC ~io_service(); #ifdef LLFIO_IO_POST_SIGNAL private: int _blocked_interrupt_signal{0}; std::atomic _need_signal{false}; // false = signal not needed, true = signal needed LLFIO_HEADERS_ONLY_MEMFUNC_SPEC void _block_interruption() noexcept; LLFIO_HEADERS_ONLY_MEMFUNC_SPEC void _unblock_interruption() noexcept; public: /*! Returns the signal used for interrupting run_until(). Only used on POSIX when BSD kqueues are not used. Defaults to LLFIO_IO_POST_SIGNAL on platforms which use it. \note Only present if LLFIO_IO_POST_SIGNAL is defined. */ static LLFIO_HEADERS_ONLY_MEMFUNC_SPEC int interruption_signal() noexcept; /*! Sets the signal used for interrupting run_until(), returning the former signal setting. Only used on POSIX when BSD kqueues are not used. Special values are 0 for deinstall global signal handler, and -1 for install to first unused signal between SIGRTMIN and SIGRTMAX. Changing this while any io_service instances exist is a bad idea. \note Only present if LLFIO_IO_POST_SIGNAL is defined. */ static LLFIO_HEADERS_ONLY_MEMFUNC_SPEC int set_interruption_signal(int signo = LLFIO_IO_POST_SIGNAL); #endif #if LLFIO_USE_POSIX_AIO //! True if this i/o service is using BSD kqueues bool using_kqueues() const noexcept { return _use_kqueues; } //! Force disable any use of BSD kqueues LLFIO_HEADERS_ONLY_MEMFUNC_SPEC void disable_kqueues(); #endif /*! Runs the i/o service for the thread owning this i/o service. Returns true if more work remains and we just handled an i/o or post; false if there is no more work; `errc::timed_out` if the deadline passed; `errc::operation_not_supported` if you try to call it from a non-owning thread; `errc::invalid_argument` if deadline is invalid. */ LLFIO_HEADERS_ONLY_VIRTUAL_SPEC result run_until(deadline d) noexcept; //! \overload result run() noexcept { return run_until(deadline()); } private: LLFIO_HEADERS_ONLY_VIRTUAL_SPEC void _post(detail::function_ptr &&f); public: /*! Schedule the callable to be invoked by the thread owning this object and executing `run()` at its next available opportunity. Unlike any other function in this API layer, this function is thread safe. */ template void post(U &&f) { _post(detail::make_function_ptr(std::forward(f))); } #if defined(LLFIO_ENABLE_COROUTINES) || defined(DOXYGEN_IS_IN_THE_HOUSE) /*! An awaitable suspending execution of this coroutine on the current kernel thread, and resuming execution on the kernel thread running this i/o service. This is a convenience wrapper for `post()`. */ struct awaitable_post_to_self { io_service *service; //! Constructor, takes the i/o service whose kernel thread we are to reschedule onto explicit awaitable_post_to_self(io_service &_service) : service(&_service) { } bool await_ready() { return false; } void await_suspend(coroutine_handle<> co) { service->post([co = co](io_service * /*unused*/) mutable { co.resume(); }); } void await_resume() {} }; #endif }; LLFIO_V2_NAMESPACE_END #ifdef _MSC_VER #pragma warning(pop) #endif #endif