On Fri, Nov 25, 2022 at 01:58:55PM +0900, Junio C Hamano wrote: > Jeff King <peff@xxxxxxxx> writes: > > > One thing I'd worry about is buffering. One of the nice things about > > "-v" is that there is nothing between you and the running programs, so > > you are much less likely to be fooled about the order of events in the > > output. Or wondering why nothing is happening because real-time output > > seems to have stalled. But piping through "cat" may end up with weird > > pauses while it fills up a 4k buffer. Using stdbuf could help, but > > that's far from portable. > > We could pipe to "dd bs=1 conv=fsync" (tongue-in-cheek---I think > conv=fsync is a GNU thing). We don't need the fsync; we are just worried about in-process buffering, not kernel-level flushing. So bs=1 is sufficient, in that it would use syscalls to read/write single bytes. It would just be horribly inefficient. We really just want immediate partial read()/write() but with sensible buffer sizes. It would be easy to write a 5-line C program that did this if we really wanted to. I'm not entirely convinced it's worth worrying too much about, though. -Peff
