On Sat, Jul 25, 2020 at 04:01:25PM -0700, Junio C Hamano wrote: > Jeff King <peff@xxxxxxxx> writes: > > > I think it was this one: > > > > https://lore.kernel.org/git/20160927191955.mympqgylrxhkp24n@xxxxxxxxxxxxxxxxxxxxx/ > > > > I still think it's a good idea, but the hard part is lib-ifying all of > > the functions that call die() to instead return an error up the stack > > (and handle cleanup, etc). Which is why I never really pushed it > > further. But if we're going to be lib-ifying such calls anyway, I think > > it's nice to do this flexible thing (from their perspective it's no more > > work to trigger the callback than it is to call error() and return). > > Yeah, I almost 80%+ agree. > > The remainder of 20% is that I am not so convinced that (fmt_string > + va_list) that isn't pretty much usable for anything but generating > human readable error messages is enough. It is certainly a good > enough replacement for (and much better than) the approach to > prepare an error string in a "struct strbuf err" that was passeed by > the caller, but I am not sure if that is a good bar to aim to pass > to begin with ;-). Yeah, I certainly agree that just passing around strings does not provide as much information as some callers would like. But I think introducing a system of machine-readable error tokens would be cumbersome for a lot of cases, to the point that it may tip this from "seems like an easy enough change" to something that nobody wants to take on. One in-between step that would be pretty easy is to record errno when available (i.e., if error_errno() becomes error_ctx_errno(), then we just stuff that errno value into the error_context struct). That would let us reliably propagate errno up the stack, or even give intermediate callers the opportunity to munge it (e.g., if a ref lookup failed with EISDIR while opening a file, it might convert that to ENOENT as it propagates up the stack). -Peff
