-
Simon McVittie authored
This avoids a coreutils limitation in versions [8.27, 8.29), which unfortunately includes Ubuntu 18.04: timeout(1) relied on SIGCHLD not being blocked, but did not unblock it themselves. The Steam client runs s-r-s-i from a thread that has SIGCHLD blocked, and SIGALRM, SIGPIPE and SIGTRAP ignored. This is contrary to frequent expectations of general-purpose libraries and command-line utilities: in particular, SIGPIPE is used by processes writing to a pipe to detect that the read end has died, SIGTRAP is frequently used to terminate processes on an assertion failure, and SIGCHLD is often necessary for subprocess management. Games (whether Steam or non-Steam) appear to be run with only SIGPIPE and SIGTRAP ignored, so strictly speaking, unignoring those two makes the behaviour of s-r-s-i unlike the behaviour of actual games. However, the effect of unblocking SIGPIPE and SIGTRAP is to make s-r-s-i more likely to crash (the default disposition of both those signals is process termination), which is probably what we want: if there is any difference, we want s-r-s-i's helpers to be more strict than games (crash more often), so that they can act as an early-warning system for issues that could manifest in games. In s-r-s-i, don't unblock signals until after we have finished command-line parsing and redirected stdout to print to the original stderr. Otherwise, g_debug() would go to the original stdout, resulting in it being invalid JSON. Signed-off-by:
Simon McVittie <smcv@collabora.com>
Simon McVittie authoredThis avoids a coreutils limitation in versions [8.27, 8.29), which unfortunately includes Ubuntu 18.04: timeout(1) relied on SIGCHLD not being blocked, but did not unblock it themselves. The Steam client runs s-r-s-i from a thread that has SIGCHLD blocked, and SIGALRM, SIGPIPE and SIGTRAP ignored. This is contrary to frequent expectations of general-purpose libraries and command-line utilities: in particular, SIGPIPE is used by processes writing to a pipe to detect that the read end has died, SIGTRAP is frequently used to terminate processes on an assertion failure, and SIGCHLD is often necessary for subprocess management. Games (whether Steam or non-Steam) appear to be run with only SIGPIPE and SIGTRAP ignored, so strictly speaking, unignoring those two makes the behaviour of s-r-s-i unlike the behaviour of actual games. However, the effect of unblocking SIGPIPE and SIGTRAP is to make s-r-s-i more likely to crash (the default disposition of both those signals is process termination), which is probably what we want: if there is any difference, we want s-r-s-i's helpers to be more strict than games (crash more often), so that they can act as an early-warning system for issues that could manifest in games. In s-r-s-i, don't unblock signals until after we have finished command-line parsing and redirected stdout to print to the original stderr. Otherwise, g_debug() would go to the original stdout, resulting in it being invalid JSON. Signed-off-by:
Simon McVittie <smcv@collabora.com>