(I know, changing variables names AND introducing changes is better
done in more commits, but…)
Added back an lseek that was missing. If TLS_WANT_POLL{IN,OUT}, we
need to re-send that block, but we need also to rewind the file, in
order to read(2) that chunk again. This doesn’t solve the corruption
in transferring big files, but reduces them. I still haven’t tracked
down the corruption :/
At the moment there is an hardcoded table that maps mime types to
extensions. For the time being this can be OK, as I don’t even
currently serve all those types of file, but in the future I’d like to
let user pass a file with the mapping, like /usr/share/misc/mime.types
on OpenBSD, to map. However, even in this case, we should hardcode
text/gemini IMHO, since most mime.types listing doesn’t have it yet.
We can handle up to MAX_USERS (64 by default) concurrently.
Now, given that we don’t support CGI, it’s not a big deal. Gemini
requests are small (up to 1024 bytes), and also the replies from the
server are small (one line plus the document — if any), all over TLS
obviously. (but even there, it’s lighter than HTTP because we don’t
need to send the whole chain for the certificate — see TOFU).
Given all the above, this doesn’t really improve the performance in
the real world, but it’s nice to have. The main use case for this is
to disallow slow clients to stop fast clients.