[源码分析文章](https://zhuanlan.zhihu.com/c_1094189343643652096) Linuxthreads - POSIX 1003.1c kernel threads for Linux Copyright 1996 Xavier Leroy (Xavier.Leroy@inria.fr) DESCRIPTION: This is release 0.5 (BETA) of LinuxThreads, a BiCapitalized implementation of the Posix 1003.1c "pthread" interface for Linux. Several implementations of pthread are already available for Linux (e.g. Chris Provenzano's implementation, which is part of recent releases of the Linux C library), but they are all user-level: all threads live within a single Linux process, which schedules the various threads itself. In contrast, LinuxThreads provides kernel-level threads: each thread is a separate Unix process, sharing its address space with the other threads through the new system call clone(). Scheduling between threads is handled by the kernel scheduler, just like scheduling between Unix processes. The advantages of kernel-level threads over user-level threads are: - supports multiprocessors; - possibly smoother scheduling, since the same scheduler (the kernel's) handles both threads and other Unix processes; - no special action needed to make blocking system call suspend only the calling thread, not all threads of the process (user-level threads usually rely on a combination of select() and jacketing of C library functions for system calls); - as a consequence of the above, more efficient I/O and timer-based operations (no need for extra select() and gettimeofday()). The disadvantages are: - more expensive context switches between threads, at least when blocking on conditions and mutexes; - thread creation is more expensive. REQUIREMENTS: - Linux version 2.0 and up (requires the new clone() system call and the new realtime scheduler). - For Intel platforms: libc 5.2.18 or later is required. 5.2.18 or 5.4.12 are recommended; 5.3.12 and 5.4.7 have problems (see below). - Also supports glibc (a.k.a. libc 6), which actually comes with a specially-adapted version of this library. - Currently supports Intel, Alpha, Sparc, Motorola 68k and MIPS platforms. - Multiprocessors are supported. INSTALLATION: - Edit the Makefile, set the variables in the "Configuration" section. - Do "make". - Do "make install". USING LINUXTHREADS: gcc -D_REENTRANT ... -lpthread A complete set of manual pages is included. Also see the subdirectory Examples/ for some sample programs. STATUS: - All functions in the Posix 1003.1c base interface implemented. Partial support for priority scheduling also available. - For users of libc 5 (H.J.Lu's libc), a number of C library functions are reimplemented or wrapped to make them thread-safe, including: * malloc functions * stdio functions (define _REENTRANT before including <stdio.h>) * per-thread errno variable (define _REENTRANT before including <errno.h>) * directory reading functions (opendir(), etc) * sleep() * gmtime(), localtime() New library functions provided: * flockfile(), funlockfile(), ftrylockfile() * reentrant versions of network database functions (gethostbyname_r(), etc) and password functions (getpwnam_r(), etc). All versions of libc between 5.2.18 and 5.4.12 exclusive have problems with the per-thread errno variable. Symptoms include inability to do fdopen() on a pipe or socket. Either downgrade to 5.2.18 or upgrade to 5.4.12 or above. - libc 6 (glibc) provides much better thread support than libc 5, and comes with a specially-adapted version of LinuxThreads. For serious multithreaded programming, you should consider switching to libc 6. WARNING: Many existing libraries are not compatible with LinuxThreads, either because they are not inherently thread-safe, or because they have not been compiled with the -D_REENTRANT. A prime example of the latter is Xlib. If you link it with LinuxThreads, you'll probably get an "unknown 0 error" very early. This is just a consequence of the Xlib binaries using the global variable "errno" to fetch error codes, while LinuxThreads and the C library use the per-thread "errno" location. See the file README.Xfree3.2 for info on how to compile the Xfree 3.2 libraries to make them compatible with LinuxThreads. KNOWN BUGS AND LIMITATIONS: - Threads share pretty much everything they should share according to the standard: memory space, file descriptors, signal handlers, current working directory, etc. One thing that they do not share is their pid's and parent pid's. According to the standard, they should have the same, but that's one thing we cannot achieve in this implementation (until the CLONE_PID flag to clone() becomes usable). - The current implementation uses the two signals SIGUSR1 and SIGUSR2, so user-level code cannot employ them. Ideally, there should be two signals reserved for this library. One signal is used for restarting threads blocked on mutexes or conditions; the other is for thread cancellation. - The stacks for the threads are allocated high in the memory space, below the stack of the initial process, and spaced 2M apart. Stacks are allocated with the "grow on demand" flag, so they don't use much virtual space initially (4k, currently), but can grow up to 2M if needed. Reserving such a large address space for each thread means that, on a 32-bit architecture, no more than about 1000 threads can coexist (assuming a 2Gb address space for user processes), but this is reasonable, since each thread uses up one entry in the kernel's process table, which is usually limited to 512 processes. Another potential problem of the "grow on demand" scheme is that nothing prevents the user from mmap'ing something in the 2M address window reserved for a thread stack, possibly causing later extensions of that stack to fail. Mapping at fixed addresses should be avoided when using this library. - Signal handling does not fully conform to the Posix standard, due to the fact that threads are here distinct processes that can be sent signals individually, so there's no notion of sending a signal to "the" process (the collection of all threads). More precisely, here is a summary of the standard requirements and how they are met by the implementation: 1- Synchronous signals (generated by the thread execution, e.g. SIGFPE) are delivered to the thread that raised them. (OK.) 2- A fatal asynchronous signal terminates all threads in the process. (OK. The thread manager notices when a thread dies on a signal and kills all other threads with the same signal.) 3- If one or several threads are waiting on a signal using sigwait(), then that signal will be delivered to one of those threads. (No, it will be delivered to the thread it's been sent to, based on the pid.) 4- An asynchronous signal will be delivered to one of the threads of the program which does not block the signal (it is unspecified which). (No, the signal is delivered to the thread it's been sent to, as in 3- above. 5- The signal will be delivered to at most one thread. (OK, except for signals generated from the terminal or sent to the process group, which will be delivered to all threads.) - The realtime extensions (scheduling control) have not been tested extensively. A known discrepancy with the POSIX standard is that multiple threads waiting on a mutex are not restarted in priority order when the mutex is unlocked: we just do first-suspended, first-restarted. - The signal-safe semaphores (sem_post, etc) are not available on the 386 nor on the Sparc because these architectures do not provide atomic compare-and-swap, neither load-reservation/store-conditional. Various fallback, less efficient solutions are being considered. For the time being, semaphores should not be used in programs that are intended to run on all the Intel x86 line, including the 386. - The current implementations of the MIPS support assumes a MIPS ISA II processor or better. These processors support atomic operations by ll/sc instructions. Older R2000/R3000 series processors are not supported yet; support for these will have higher overhead. - Mutexes, conditions, semaphores cannot be shared between processes. This optional component of the Posix standard is not obvious to implement efficiently (i.e. without reverting to SystemV IPCs) and does not seem terribly useful to me. If you need to synchronize processes (with distinct address spaces), you'd be better off with traditional Unix synchronization tools (pipes, sockets, SystemV IPCs, etc.)