On 16/10/2024 16:09, Jeff King wrote: > On Wed, Oct 16, 2024 at 10:55:40AM -0400, Jeff King wrote: > >>> The logic is broken on Cygwin though: when a client asks the daemon to >>> exit, they won't see the EOF and will instead get an error message: >>> >>> fatal: read error from cache daemon: Software caused connection abort >>> >>> This issue is known in Cygwin, see for example [1], but the exact root >>> cause is not known. >>> [...] >>> [1]: https://github.com/cygporter/git/issues/51 >> >> I don't see any details at that issue, but I'm not sure how it would fix >> things. From the client's perspective, they are going to see the >> descriptor either way. Unless there is some magic that fclose() does >> that a normal descriptor-close-on-exit does not do. >> >> That "Software caused connection abort" thing seems like a weird >> not-standard-Unix errno value. Searching for it mostly yields people >> complaining about getting it from ssh under cygwin. :) >> >> If the magic that cygwin needs is actually "fclose before unlink", then >> that is in opposition to other platforms (and I suspect would make t0301 >> racy there). > > This all seemed eerily familiar. Try this thread: > > https://lore.kernel.org/git/9dc3e85f-a532-6cff-de11-1dfb2e4bc6b6@xxxxxxxxxxxxxxxxxxxx/ > > It looks like the conclusion was that we should adjust errno handling on > the client side, but nobody ever followed up with an actual patch. Heh, our emails crossed. Yes, I was hoping that, given that Adam had identified the cygwin commit that caused the regression, some resolution on the cygwin side would fix things up. I waited ... and then put it on my TODO list! :) I did look at the cygwin commit and it wasn't at all obvious what happened. In fact there was a comment about making sure that errno values didn't change as a side-effect! Sorry for being tardy, again ... ATB, Ramsay Jones
