--- old/src/solaris/native/java/lang/UNIXProcess_md.c Thu Nov 22 21:04:35 2012 +++ new/src/solaris/native/java/lang/UNIXProcess_md.c Thu Nov 22 21:04:34 2012 @@ -49,6 +49,12 @@ #include #include +#if defined(__solaris__) || defined(__APPLE__) +#include +#endif + +#include "unixprocess_md.h" + #ifdef __APPLE__ #include #define environ (*_NSGetEnviron()) @@ -55,7 +61,7 @@ #endif /* - * There are 3 possible strategies we might use to "fork": + * There are 4 possible strategies we might use to "fork": * * - fork(2). Very portable and reliable but subject to * failure due to overcommit (see the documentation on @@ -90,37 +96,23 @@ * http://sources.redhat.com/bugzilla/show_bug.cgi?id=10311 * but the glibc maintainers closed it as WONTFIX. * + * - posix_spawn(). While posix_spawn() is a fairly elaborate and + * complicated system call, it can't quite do everything that the old + * fork()/exec() combination can do, so the only feasible way to do + * this, is to use posix_spawn to launch a new helper executable + * "jprochelper", which in turn execs the target (after cleaning + * up file-descriptors etc.) The end result is the same as before, + * a child process linked to the parent in the same way, but it + * avoids the problem of duplicating the parent (VM) process + * address space temporarily, before launching the target command. + * * Based on the above analysis, we are currently using vfork() on - * Linux and fork() on other Unix systems, but the code to use clone() - * remains. + * Linux and spawn() on other Unix systems, but the code to use clone() + * and fork() remains. */ -#define START_CHILD_USE_CLONE 0 /* clone() currently disabled; see above. */ - -#ifndef START_CHILD_USE_CLONE - #ifdef __linux__ - #define START_CHILD_USE_CLONE 1 - #else - #define START_CHILD_USE_CLONE 0 - #endif -#endif - -/* By default, use vfork() on Linux. */ -#ifndef START_CHILD_USE_VFORK - #ifdef __linux__ - #define START_CHILD_USE_VFORK 1 - #else - #define START_CHILD_USE_VFORK 0 - #endif -#endif - -#if START_CHILD_USE_CLONE +#ifdef __linux__ #include -#define START_CHILD_SYSTEM_CALL "clone" -#elif START_CHILD_USE_VFORK -#define START_CHILD_SYSTEM_CALL "vfork" -#else -#define START_CHILD_SYSTEM_CALL "fork" #endif #ifndef STDIN_FILENO @@ -214,6 +206,16 @@ static const char* defaultPath(void) { +#ifdef _CS_PATH + char *pathbuf; + size_t n; + n = confstr(_CS_PATH,NULL,(size_t) 0); + pathbuf = malloc(n); + if (pathbuf == NULL) + abort(); + confstr(_CS_PATH, pathbuf, n); + return pathbuf; +#else #ifdef __solaris__ /* These really are the Solaris defaults! */ return (geteuid() == 0 || getuid() == 0) ? @@ -222,6 +224,7 @@ #else return ":/bin:/usr/bin"; /* glibc */ #endif +#endif } static const char* @@ -271,23 +274,25 @@ * Cached value of JVM's effective PATH. * (We don't support putenv("PATH=...") in native code) */ -static const char *parentPath; +const char *xx_parentPath; /** - * Split, canonicalized version of parentPath + * Split, canonicalized version of xx_parentPath */ -static const char * const *parentPathv; +const char * const *xx_parentPathv; static jfieldID field_exitcode; +static jfieldID helperfield; JNIEXPORT void JNICALL Java_java_lang_UNIXProcess_initIDs(JNIEnv *env, jclass clazz) { field_exitcode = (*env)->GetFieldID(env, clazz, "exitcode", "I"); + helperfield = (*env)->GetStaticFieldID(env, clazz, + "helperpath", "Ljava/lang/String;"); + xx_parentPath = effectivePath(); + xx_parentPathv = splitPath(env, xx_parentPath); - parentPath = effectivePath(); - parentPathv = splitPath(env, parentPath); - setSIGCHLDHandler(env); } @@ -318,6 +323,7 @@ /* We used to use waitid() on Solaris, waitpid() on Linux, but * waitpid() is more standard, so use it on all POSIX platforms. */ int status; + /* Wait for the child process to exit. This returns immediately if the child has already exited. */ while (waitpid(pid, &status, 0) < 0) { @@ -438,8 +444,9 @@ { if (fd_from != fd_to) { if ((restartableDup2(fd_from, fd_to) == -1) || - (restartableClose(fd_from) == -1)) + (restartableClose(fd_from) == -1)) { return -1; + } } return 0; } @@ -458,14 +465,16 @@ (*env)->ReleaseByteArrayElements(env, arr, (jbyte*) parr, JNI_ABORT); } -static void -initVectorFromBlock(const char**vector, const char* block, int count) +void +xx_initVectorFromBlock(const char**vector, const char* block, int count) { int i; const char *p; + for (i = 0, p = block; i < count; i++) { /* Invariant: p always points to the start of a C string. */ vector[i] = p; + while (*(p++)); } vector[count] = NULL; @@ -541,20 +550,20 @@ * be a shell script and the system default shell is invoked to run it. */ static void -execve_with_shell_fallback(const char *file, +execve_with_shell_fallback(int mode, const char *file, const char *argv[], const char *const envp[]) { -#if START_CHILD_USE_CLONE || START_CHILD_USE_VFORK - /* shared address space; be very careful. */ - execve(file, (char **) argv, (char **) envp); - if (errno == ENOEXEC) - execve_as_traditional_shell_script(file, argv, envp); -#else - /* unshared address space; we can mutate environ. */ - environ = (char **) envp; - execvp(file, (char **) argv); -#endif + if (mode == MODE_CLONE || mode == MODE_VFORK) { + /* shared address space; be very careful. */ + execve(file, (char **) argv, (char **) envp); + if (errno == ENOEXEC) + execve_as_traditional_shell_script(file, argv, envp); + } else { + /* unshared address space; we can mutate environ. */ + environ = (char **) envp; + execvp(file, (char **) argv); + } } /** @@ -565,7 +574,7 @@ * specified via the 3rd argument instead of being inherited from environ. */ static void -JDK_execvpe(const char *file, +JDK_execvpe(int mode, const char *file, const char *argv[], const char *const envp[]) { @@ -580,7 +589,7 @@ } if (strchr(file, '/') != NULL) { - execve_with_shell_fallback(file, argv, envp); + execve_with_shell_fallback(mode, file, argv, envp); } else { /* We must search PATH (parent's, not child's) */ char expanded_file[PATH_MAX]; @@ -587,7 +596,8 @@ int filelen = strlen(file); int sticky_errno = 0; const char * const * dirs; - for (dirs = parentPathv; *dirs; dirs++) { + + for (dirs = xx_parentPathv; *dirs; dirs++) { const char * dir = *dirs; int dirlen = strlen(dir); if (filelen + dirlen + 1 >= PATH_MAX) { @@ -597,7 +607,7 @@ memcpy(expanded_file, dir, dirlen); memcpy(expanded_file + dirlen, file, filelen); expanded_file[dirlen + filelen] = '\0'; - execve_with_shell_fallback(expanded_file, argv, envp); + execve_with_shell_fallback(mode, expanded_file, argv, envp); /* There are 3 responses to various classes of errno: * return immediately, continue (especially for ENOENT), * or continue with "sticky" errno. @@ -639,14 +649,14 @@ } /* - * Reads nbyte bytes from file descriptor fd into buf, + * Reads nbyte bytes from file descriptor fd into buf. * The read operation is retried in case of EINTR or partial reads. * * Returns number of bytes read (normally nbyte, but may be less in * case of EOF). In case of read errors, returns -1 and sets errno. */ -static ssize_t -readFully(int fd, void *buf, size_t nbyte) +ssize_t +xx_readFully(int fd, void *buf, size_t nbyte) { ssize_t remaining = nbyte; for (;;) { @@ -669,22 +679,6 @@ } } -typedef struct _ChildStuff -{ - int in[2]; - int out[2]; - int err[2]; - int fail[2]; - int fds[3]; - const char **argv; - const char **envv; - const char *pdir; - jboolean redirectErrorStream; -#if START_CHILD_USE_CLONE - void *clone_stack; -#endif -} ChildStuff; - static void copyPipe(int from[2], int to[2]) { @@ -693,32 +687,41 @@ } /** - * Child process after a successful fork() or clone(). + * Child process after a successful v/fork(), clone() or posix_spawn(). * This function must not return, and must be prepared for either all * of its address space to be shared with its parent, or to be a copy. * It must not modify global variables such as "environ". */ -static int -childProcess(void *arg) +int +xx_childProcess(void *arg) { - const ChildStuff* p = (const ChildStuff*) arg; + ChildStuff* p = (ChildStuff*) arg; /* Close the parent sides of the pipes. Closing pipe fds here is redundant, since closeDescriptors() - would do it anyways, but a little paranoia is a good thing. */ + would do it anyways, but a little paranoia is a good thing. + We close both sides of the childenv pipe since in spawn + mode we are finished with it now, and it's not used at all + in the other modes */ if ((closeSafely(p->in[1]) == -1) || (closeSafely(p->out[0]) == -1) || (closeSafely(p->err[0]) == -1) || + (closeSafely(p->childenv[0]) == -1) || + (closeSafely(p->childenv[1]) == -1) || (closeSafely(p->fail[0]) == -1)) goto WhyCantJohnnyExec; + p->in[1] = -1; p->out[0] = -1; p->err[0] = -1; + p->childenv[0] = -1; p->childenv[1] = -1; p->fail[0] = -1; + /* Give the child sides of the pipes the right fileno's. */ /* Note: it is possible for in[0] == 0 */ if ((moveDescriptor(p->in[0] != -1 ? p->in[0] : p->fds[0], STDIN_FILENO) == -1) || (moveDescriptor(p->out[1]!= -1 ? p->out[1] : p->fds[1], - STDOUT_FILENO) == -1)) + STDOUT_FILENO) == -1)) { goto WhyCantJohnnyExec; + } if (p->redirectErrorStream) { if ((closeSafely(p->err[1]) == -1) || @@ -749,7 +752,7 @@ if (fcntl(FAIL_FILENO, F_SETFD, FD_CLOEXEC) == -1) goto WhyCantJohnnyExec; - JDK_execvpe(p->argv[0], p->argv, p->envv); + JDK_execvpe(p->mode, p->argv[0], p->argv, p->envv); WhyCantJohnnyExec: /* We used to go to an awful lot of trouble to predict whether the @@ -771,8 +774,55 @@ return 0; /* Suppress warning "no return value from function" */ } +/* arg is an array of pointers to 0 terminated strings. array is terminated + * by a null element. + * + * *nelems and *nbytes receive the number of elements of array (incl 0) + * and total number of bytes (incl. 0) + * Note. An empty array will have one null element + * But if arg is null, then *nelems set to 0, and *nbytes to 0 + */ +static void arraysize (const char * const *arg, int *nelems, int *nbytes) +{ + int i, bytes, count; + const char * const *a = arg; + char *p; + int *q; + if (arg == 0) { + *nelems = 0; + *nbytes = 0; + return; + } + /* count the array elements and number of bytes */ + for (count=0, bytes=0; *a != 0; count++, a++) { + bytes += strlen (*a)+1; + } + *nbytes = bytes; + *nelems = count+1; +} + +/* copy the strings from arg[] into buf, starting at given offset + * return new offset to next free byte + */ +static int copystrings (char *buf, int offset, const char * const *arg) { + char *p; + const char * const *a; + int count=0; + + if (arg == 0) { + return offset; + } + for (p=buf+offset, a=arg; *a != 0; a++) { + int len = strlen (*a) +1; + memcpy (p, *a, len); + p += len; + count += len; + } + return offset+count; +} + /** - * Start a child process running function childProcess. + * Start a child process running function xx_childProcess. * This function only returns in the parent. * We are unusually paranoid; use of clone/vfork is * especially likely to tickle gcc/glibc bugs. @@ -780,10 +830,12 @@ #ifdef __attribute_noinline__ /* See: sys/cdefs.h */ __attribute_noinline__ #endif + static pid_t -startChild(ChildStuff *c) { -#if START_CHILD_USE_CLONE +cloneChild(ChildStuff *c) { +#ifdef __linux__ #define START_CHILD_CLONE_STACK_SIZE (64 * 1024) + /* * See clone(2). * Instead of worrying about which direction the stack grows, just @@ -792,11 +844,19 @@ if ((c->clone_stack = malloc(2 * START_CHILD_CLONE_STACK_SIZE)) == NULL) /* errno will be set to ENOMEM */ return -1; - return clone(childProcess, + return clone(xx_childProcess, c->clone_stack + START_CHILD_CLONE_STACK_SIZE, CLONE_VFORK | CLONE_VM | SIGCHLD, c); #else - #if START_CHILD_USE_VFORK + /* not available on Solaris / Mac */ + assert(0); +#endif +} + +static pid_t +vforkChild(ChildStuff *c) { + volatile pid_t resultPid; + /* * We separate the call to vfork into a separate function to make * very sure to keep stack of child from corrupting stack of parent, @@ -803,8 +863,19 @@ * as suggested by the scary gcc warning: * warning: variable 'foo' might be clobbered by 'longjmp' or 'vfork' */ - volatile pid_t resultPid = vfork(); - #else + resultPid = vfork(); + + if (resultPid == 0) { + xx_childProcess(c); + } + assert(resultPid != 0); /* xx_childProcess never returns */ + return resultPid; +} + +static pid_t +forkChild(ChildStuff *c) { + pid_t resultPid; + /* * From Solaris fork(2): In Solaris 10, a call to fork() is * identical to a call to fork1(); only the calling thread is @@ -811,18 +882,127 @@ * replicated in the child process. This is the POSIX-specified * behavior for fork(). */ - pid_t resultPid = fork(); - #endif - if (resultPid == 0) - childProcess(c); - assert(resultPid != 0); /* childProcess never returns */ + resultPid = fork(); + + if (resultPid == 0) { + xx_childProcess(c); + } + assert(resultPid != 0); /* xx_childProcess never returns */ return resultPid; -#endif /* ! START_CHILD_USE_CLONE */ } +static pid_t +spawnChild(JNIEnv *env, jobject process, ChildStuff *c) { + pid_t resultPid; + const char *helperpath; + jobject helperobj; + jboolean isCopy; + int i, offset, rval, bufsize; + char *buf, buf1[16]; + char *hlpargs[2]; + SpawnInfo sp; + + helperobj = (*env)->GetStaticObjectField( + env, (*env)->GetObjectClass (env, process), + helperfield + ); + if (helperobj == NULL) { + return -1; + } + helperpath = GetStringPlatformChars (env, helperobj, &isCopy); + /* need to tell helper which fd is for receiving the childstuff + * and which fd to send response back on + */ + sprintf (buf1, "%d:%d", c->childenv[0], c->fail[1]); + /* put the fd string as argument to the helper cmd */ + hlpargs[0] = buf1; + hlpargs[1] = 0; + + /* Following items are sent down the pipe to the helper + * after it is spawned. + * All strings are null terminated. All arrays of strings + * have an empty string for termination. + * - the ChildStuff struct + * - the SpawnInfo struct + * - the argv strings array + * - the envv strings array + * - the home directory string + * - the parentPath string + * - the parentPathv array + */ + /* First calculate the sizes */ + arraysize (c->argv, &sp.nargv, &sp.argvBytes); + bufsize = sp.argvBytes; + arraysize (c->envv, &sp.nenvv, &sp.envvBytes); + bufsize += sp.envvBytes; + sp.dirlen = c->pdir == 0 ? 0 : strlen (c->pdir)+1; + bufsize += sp.dirlen; + sp.parentPathLen = xx_parentPath == 0 ? 0 : strlen (xx_parentPath)+1; + bufsize += sp.parentPathLen; + arraysize (xx_parentPathv, &sp.nparentPathv, &sp.parentPathvBytes); + bufsize += sp.parentPathvBytes; + /* We need to clear FD_CLOEXEC if set in the fds[]. + * Files are created FD_CLOEXEC in Java. + * Otherwise, they will be closed when the target gets exec'd */ + for (i=0; i<3; i++) { + if (c->fds[i] != -1) { + int flags = fcntl(c->fds[i], F_GETFD); + if (flags & FD_CLOEXEC) { + fcntl (c->fds[i], F_SETFD, flags & (~1)); + } + } + } + + rval = posix_spawn (&resultPid, helperpath, 0, 0, (char * const *) hlpargs, environ); + + if (rval != 0) { + return -1; + } + + /* now the lengths are known, copy the data */ + buf = NEW (char, bufsize); + if (buf == 0) { + return -1; + } + offset = copystrings (buf, 0, &c->argv[0]); + offset = copystrings (buf, offset, &c->envv[0]); + memcpy (buf+offset, c->pdir, sp.dirlen); + offset += sp.dirlen; + memcpy (buf+offset, xx_parentPath, sp.parentPathLen); + offset += sp.parentPathLen; + offset = copystrings (buf, offset, xx_parentPathv); + assert (offset == bufsize); + + /* write the two structs and the data buffer */ + write (c->childenv[1], (char *)c, sizeof(*c)); + write (c->childenv[1], (char *)&sp, sizeof(sp)); + write (c->childenv[1], buf, bufsize); + free (buf); + + /* In this mode an external main() in invoked which calls back into + * xx_childProcess() in this file, rather than directly + * via the statement below */ + return resultPid; +} + +static pid_t +startChild(JNIEnv *env, jobject process, ChildStuff *c) { + switch (c->mode) { + case MODE_CLONE: + return cloneChild(c); + case MODE_VFORK: + return vforkChild(c); + case MODE_FORK: + return forkChild(c); + case MODE_SPAWN: + return spawnChild(env, process, c); + } +} + JNIEXPORT jint JNICALL Java_java_lang_UNIXProcess_forkAndExec(JNIEnv *env, jobject process, + jint mode, jbyteArray prog, jbyteArray argBlock, jint argc, jbyteArray envBlock, jint envc, @@ -832,7 +1012,7 @@ { int errnum; int resultPid = -1; - int in[2], out[2], err[2], fail[2]; + int in[2], out[2], err[2], fail[2], childenv[2]; jint *fds = NULL; const char *pprog = NULL; const char *pargBlock = NULL; @@ -840,18 +1020,18 @@ ChildStuff *c; in[0] = in[1] = out[0] = out[1] = err[0] = err[1] = fail[0] = fail[1] = -1; + childenv[0] = childenv[1] = -1; if ((c = NEW(ChildStuff, 1)) == NULL) return -1; c->argv = NULL; c->envv = NULL; c->pdir = NULL; -#if START_CHILD_USE_CLONE c->clone_stack = NULL; -#endif /* Convert prog + argBlock into a char ** argv. * Add one word room for expansion of argv for use by * execve_as_traditional_shell_script. + * This word also used when using spawn mode */ assert(prog != NULL && argBlock != NULL); if ((pprog = getBytes(env, prog)) == NULL) goto Catch; @@ -858,13 +1038,14 @@ if ((pargBlock = getBytes(env, argBlock)) == NULL) goto Catch; if ((c->argv = NEW(const char *, argc + 3)) == NULL) goto Catch; c->argv[0] = pprog; - initVectorFromBlock(c->argv+1, pargBlock, argc); + c->argc = argc + 2; + xx_initVectorFromBlock(c->argv+1, pargBlock, argc); if (envBlock != NULL) { /* Convert envBlock into a char ** envv */ if ((penvBlock = getBytes(env, envBlock)) == NULL) goto Catch; if ((c->envv = NEW(const char *, envc + 1)) == NULL) goto Catch; - initVectorFromBlock(c->envv, penvBlock, envc); + xx_initVectorFromBlock(c->envv, penvBlock, envc); } if (dir != NULL) { @@ -878,6 +1059,7 @@ if ((fds[0] == -1 && pipe(in) < 0) || (fds[1] == -1 && pipe(out) < 0) || (fds[2] == -1 && pipe(err) < 0) || + (pipe(childenv) < 0) || (pipe(fail) < 0)) { throwIOException(env, errno, "Bad file descriptor"); goto Catch; @@ -890,20 +1072,30 @@ copyPipe(out, c->out); copyPipe(err, c->err); copyPipe(fail, c->fail); + copyPipe(childenv, c->childenv); c->redirectErrorStream = redirectErrorStream; + c->mode = mode; - resultPid = startChild(c); + resultPid = startChild(env, process, c); assert(resultPid != 0); if (resultPid < 0) { - throwIOException(env, errno, START_CHILD_SYSTEM_CALL " failed"); + switch (c->mode) { + case MODE_CLONE: + throwIOException(env, errno, "clone failed"); + case MODE_VFORK: + throwIOException(env, errno, "vfork failed"); + case MODE_FORK: + throwIOException(env, errno, "fork failed"); + case MODE_SPAWN: + throwIOException(env, errno, "spawn failed"); + } goto Catch; } - restartableClose(fail[1]); fail[1] = -1; /* See: WhyCantJohnnyExec */ - switch (readFully(fail[0], &errnum, sizeof(errnum))) { + switch (xx_readFully(fail[0], &errnum, sizeof(errnum))) { case 0: break; /* Exec succeeded */ case sizeof(errnum): waitpid(resultPid, NULL, 0); @@ -919,9 +1111,7 @@ fds[2] = (err[0] != -1) ? err[0] : -1; Finally: -#if START_CHILD_USE_CLONE free(c->clone_stack); -#endif /* Always clean up the child's side of the pipes */ closeSafely(in [0]); @@ -928,9 +1118,11 @@ closeSafely(out[1]); closeSafely(err[1]); - /* Always clean up fail descriptors */ + /* Always clean up fail and childEnv descriptors */ closeSafely(fail[0]); closeSafely(fail[1]); + closeSafely(childenv[0]); + closeSafely(childenv[1]); releaseBytes(env, prog, pprog); releaseBytes(env, argBlock, pargBlock); @@ -948,9 +1140,9 @@ Catch: /* Clean up the parent's side of the pipes in case of failure only */ - closeSafely(in [1]); - closeSafely(out[0]); - closeSafely(err[0]); + closeSafely(in [1]); in[1] = -1; + closeSafely(out[0]); out[0] = -1; + closeSafely(err[0]); err[0] = -1; goto Finally; }