[libvirt] [PATCH 0/2] Make a wrapper for fork() - Take 2

This is an update to / deprecates the patchset I sent last night: https://www.redhat.com/archives/libvir-list/2010-February/msg00580.html I have corrected the problem found by Dan Berrange (neglecting to restore the signal mask when fork() fails). Aside from that, and corresponding comments in the commit logs, it is unchanged. Here's the original intro email: This was partly prompted by DV's suggestion last week. The first of these patches creates a new function called virFork() which behaves (almost) like fork() but takes care of some important details that pretty much any call to fork() should be doing. The 2nd switches three fork-calling functions in util.c over to using virFork() instead of fork(). In the future, except for odd circumstances, code that needs to fork should call virFork() instead, and if there is anything determined to be universally necessary at fork-time, it should be added to virFork() rather than to the callers of virFork(); hopefully this will ease maintenance and reduce replicated bugs. (Note that, while this is just an overall "code health" patch, a couple bug fix patches I'll be submitting either tomorrow or Thursday will assume it as a prerequisite).

virFork() contains bookkeeping that must be done any time a process forks. Currently this includes: 1) Call virLogLock() prior to fork() and virLogUnlock() just after, to avoid a deadlock if some other thread happens to hold that lock during the fork. 2) Reset the logging hooks and send all child process log messages to stderr. 3) Block all signals prior to fork(), then either a) reset the signal mask for the parent process, or b) clear the signal mask for the child process. Note that the signal mask handling in __virExec erroneously fails to restore the signal mask when fork() fails. virFork() fixes this problem. Other than this, it attempts to behave as closely to fork() as possible (including preserving errno for the caller), with a couple exceptions: 1) The return value is 0 (success) or -1 (failure), while the pid is returned via the pid_t* argument. Like fork(), if pid < 0 there is no child process, otherwise both the child and the parent will return to the caller, and both should look at the return value, which will indicate if some of the extra processing outlined above encountered an error. 2) If virFork() returns with pid < 0 or with a return value < 0 indicating an error condition, the error has already been reported. You can log an additional message if you like, but it isn't necessary, and may be awkwardly extraneous. Note that virFork()'s child process will *never* call _exit() - if a child process is created, it will return to the caller. util.c, util.h: add virFork() function, based on what is currently done in __virExec(). --- src/util/util.c | 118 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ src/util/util.h | 1 + 2 files changed, 119 insertions(+), 0 deletions(-) diff --git a/src/util/util.c b/src/util/util.c index cdab300..d9beddf 100644 --- a/src/util/util.c +++ b/src/util/util.c @@ -296,6 +296,124 @@ static int virClearCapabilities(void) } #endif + +/* virFork() - fork a new process while avoiding various race/deadlock conditions + + @pid - a pointer to a pid_t that will receive the return value from + fork() + + on return from virFork(), if *pid < 0, the fork failed and there is + no new process. Otherwise, just like fork(), if *pid == 0, it is the + child process returning, and if *pid > 0, it is the parent. + + Even if *pid >= 0, if the return value from virFork() is < 0, it + indicates a failure that occurred in the parent or child process + after the fork. In this case, the child process should call + _exit(1) after doing any additional error reporting. + + */ +int virFork(pid_t *pid) { + sigset_t oldmask, newmask; + struct sigaction sig_action; + int saved_errno, ret = -1; + + *pid = -1; + + /* + * Need to block signals now, so that child process can safely + * kill off caller's signal handlers without a race. + */ + sigfillset(&newmask); + if (pthread_sigmask(SIG_SETMASK, &newmask, &oldmask) != 0) { + saved_errno = errno; + virReportSystemError(errno, + "%s", _("cannot block signals")); + goto cleanup; + } + + /* Ensure we hold the logging lock, to protect child processes + * from deadlocking on another thread's inherited mutex state */ + virLogLock(); + + *pid = fork(); + saved_errno = errno; /* save for caller */ + + /* Unlock for both parent and child process */ + virLogUnlock(); + + if (*pid < 0) { + /* attempt to restore signal mask, but ignore failure, to + avoid obscuring the fork failure */ + pthread_sigmask(SIG_SETMASK, &oldmask, NULL); + virReportSystemError(saved_errno, + "%s", _("cannot fork child process")); + goto cleanup; + } + + if (*pid) { + + /* parent process */ + + /* Restore our original signal mask now that the child is + safely running */ + if (pthread_sigmask(SIG_SETMASK, &oldmask, NULL) != 0) { + saved_errno = errno; /* save for caller */ + virReportSystemError(errno, "%s", _("cannot unblock signals")); + goto cleanup; + } + ret = 0; + + } else { + + /* child process */ + + int logprio; + int i; + + /* Remove any error callback so errors in child now + get sent to stderr where they stand a fighting chance + of being seen / logged */ + virSetErrorFunc(NULL, NULL); + + /* Make sure any hook logging is sent to stderr, since child + * process may close the logfile FDs */ + logprio = virLogGetDefaultPriority(); + virLogReset(); + virLogSetDefaultPriority(logprio); + + /* Clear out all signal handlers from parent so nothing + unexpected can happen in our child once we unblock + signals */ + sig_action.sa_handler = SIG_DFL; + sig_action.sa_flags = 0; + sigemptyset(&sig_action.sa_mask); + + for (i = 1; i < NSIG; i++) { + /* Only possible errors are EFAULT or EINVAL + The former wont happen, the latter we + expect, so no need to check return value */ + + sigaction(i, &sig_action, NULL); + } + + /* Unmask all signals in child, since we've no idea + what the caller's done with their signal mask + and don't want to propagate that to children */ + sigemptyset(&newmask); + if (pthread_sigmask(SIG_SETMASK, &newmask, NULL) != 0) { + saved_errno = errno; /* save for caller */ + virReportSystemError(errno, "%s", _("cannot unblock signals")); + goto cleanup; + } + ret = 0; + } + +cleanup: + if (ret < 0) + errno = saved_errno; + return ret; +} + /* * @argv argv to exec * @envp optional environment to use for exec diff --git a/src/util/util.h b/src/util/util.h index 4207508..8460100 100644 --- a/src/util/util.h +++ b/src/util/util.h @@ -81,6 +81,7 @@ int virRun(const char *const*argv, int *status) ATTRIBUTE_RETURN_CHECK; int virRunWithHook(const char *const*argv, virExecHook hook, void *data, int *status) ATTRIBUTE_RETURN_CHECK; +int virFork(pid_t *pid); int virFileReadLimFD(int fd, int maxlen, char **buf) ATTRIBUTE_RETURN_CHECK; -- 1.6.6.1

For __virExec() this is a semantic NOP except for when fork() fails. __virExec() would previously forget to restore the signal mask in this case; virFork() corrects this behavior. virFileCreate() and virDirCreate() gain the code to reset the logging and properly deal with the signal handling race condition. This also removes a log message that had a typo ("cannot fork o create file '%s'") - this error is now logged in a more generic manner in virFork() (more generic, but really just as informative, since the fact that it's forking to create a file is immaterial to the fact that it simply can't fork) --- src/util/util.c | 100 ++++++++++++++---------------------------------------- 1 files changed, 26 insertions(+), 74 deletions(-) diff --git a/src/util/util.c b/src/util/util.c index d9beddf..c3b4084 100644 --- a/src/util/util.c +++ b/src/util/util.c @@ -451,20 +451,6 @@ __virExec(const char *const*argv, int pipeerr[2] = {-1,-1}; int childout = -1; int childerr = -1; - int logprio; - sigset_t oldmask, newmask; - struct sigaction sig_action; - - /* - * Need to block signals now, so that child process can safely - * kill off caller's signal handlers without a race. - */ - sigfillset(&newmask); - if (pthread_sigmask(SIG_SETMASK, &newmask, &oldmask) != 0) { - virReportSystemError(errno, - "%s", _("cannot block signals")); - return -1; - } if ((null = open("/dev/null", O_RDONLY)) < 0) { virReportSystemError(errno, @@ -531,17 +517,9 @@ __virExec(const char *const*argv, childerr = null; } - /* Ensure we hold the logging lock, to protect child processes - * from deadlocking on another threads inheirited mutex state */ - virLogLock(); - pid = fork(); - - /* Unlock for both parent and child process */ - virLogUnlock(); + int forkRet = virFork(&pid); if (pid < 0) { - virReportSystemError(errno, - "%s", _("cannot fork child process")); goto cleanup; } @@ -556,12 +534,8 @@ __virExec(const char *const*argv, *errfd = pipeerr[0]; } - /* Restore our original signal mask now child is safely - running */ - if (pthread_sigmask(SIG_SETMASK, &oldmask, NULL) != 0) { - virReportSystemError(errno, - "%s", _("cannot unblock signals")); - return -1; + if (forkRet < 0) { + goto cleanup; } *retpid = pid; @@ -570,38 +544,10 @@ __virExec(const char *const*argv, /* child */ - /* Remove any error callback too, so errors in child now - get sent to stderr where they stand a fighting chance - of being seen / logged */ - virSetErrorFunc(NULL, NULL); - - /* Make sure any hook logging is sent to stderr, since virExec will - * close any unknown FDs (including logging handlers) before launching - * the new process */ - logprio = virLogGetDefaultPriority(); - virLogReset(); - virLogSetDefaultPriority(logprio); - - /* Clear out all signal handlers from parent so nothing - unexpected can happen in our child once we unblock - signals */ - sig_action.sa_handler = SIG_DFL; - sig_action.sa_flags = 0; - sigemptyset(&sig_action.sa_mask); - - for (i = 1 ; i < NSIG ; i++) - /* Only possible errors are EFAULT or EINVAL - The former wont happen, the latter we - expect, so no need to check return value */ - sigaction(i, &sig_action, NULL); - - /* Unmask all signals in child, since we've no idea - what the caller's done with their signal mask - and don't want to propagate that to children */ - sigemptyset(&newmask); - if (pthread_sigmask(SIG_SETMASK, &newmask, NULL) != 0) { - virReportSystemError(errno, - "%s", _("cannot unblock signals")); + if (forkRet < 0) { + /* The fork was sucessful, but after that there was an error + * in the child (which was already logged). + */ _exit(1); } @@ -1370,14 +1316,10 @@ int virFileCreate(const char *path, mode_t mode, * following dance avoids problems caused by root-squashing * NFS servers. */ - virLogLock(); - pid = fork(); - virLogUnlock(); + int forkRet = virFork(&pid); if (pid < 0) { ret = errno; - virReportSystemError(errno, - _("cannot fork o create file '%s'"), path); return ret; } @@ -1429,7 +1371,15 @@ parenterror: return ret; } - /* child - set desired uid/gid, then attempt to create the file */ + + /* child */ + + if (forkRet < 0) { + /* error encountered and logged in virFork() after the fork. */ + goto childerror; + } + + /* set desired uid/gid, then attempt to create the file */ if ((gid != 0) && (setgid(gid) != 0)) { ret = errno; @@ -1480,15 +1430,10 @@ int virDirCreate(const char *path, mode_t mode, return virDirCreateSimple(path, mode, uid, gid, flags); } - virLogLock(); - pid = fork(); - virLogUnlock(); + int forkRet = virFork(&pid); if (pid < 0) { ret = errno; - virReportSystemError(errno, - _("cannot fork to create directory '%s'"), - path); return ret; } @@ -1538,7 +1483,14 @@ parenterror: return ret; } - /* child - set desired uid/gid, then attempt to create the directory */ + /* child */ + + if (forkRet < 0) { + /* error encountered and logged in virFork() after the fork. */ + goto childerror; + } + + /* set desired uid/gid, then attempt to create the directory */ if ((gid != 0) && (setgid(gid) != 0)) { ret = errno; -- 1.6.6.1

On Thu, Feb 18, 2010 at 11:22:26AM -0500, Laine Stump wrote:
This is an update to / deprecates the patchset I sent last night:
https://www.redhat.com/archives/libvir-list/2010-February/msg00580.html
I have corrected the problem found by Dan Berrange (neglecting to restore the signal mask when fork() fails). Aside from that, and corresponding comments in the commit logs, it is unchanged.
Here's the original intro email:
This was partly prompted by DV's suggestion last week.
The first of these patches creates a new function called virFork() which behaves (almost) like fork() but takes care of some important details that pretty much any call to fork() should be doing. The 2nd switches three fork-calling functions in util.c over to using virFork() instead of fork().
In the future, except for odd circumstances, code that needs to fork should call virFork() instead, and if there is anything determined to be universally necessary at fork-time, it should be added to virFork() rather than to the callers of virFork(); hopefully this will ease maintenance and reduce replicated bugs.
(Note that, while this is just an overall "code health" patch, a couple bug fix patches I'll be submitting either tomorrow or Thursday will assume it as a prerequisite).
Okay, thanks, the new function looks fine, and the second patch is the logical next step. ACK, applied both ! Daniel -- Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/ daniel@veillard.com | Rpmfind RPM search engine http://rpmfind.net/ http://veillard.com/ | virtualization library http://libvirt.org/
participants (2)
-
Daniel Veillard
-
Laine Stump