Instead of allocating a whole set of temp fd per `listen' directive, just use a global set per "sockets" process. This is fine since gotwebd doesn't keep multiple connections open at the same time, once it enters gotwebd_process_request() it won't exit until all the page has been processed. (this because we don't have async APIs for got operations.) ok stsp@

gotd and gotwebd are using almost the same log.c file, so unify. This only slightly changes gotwebd logs to include the process name, like gotd. ok stsp@

- don't assume an implicit "listen on socket" when another listen was specified in a server block - allow servers to share the sockets they're listening on - fix unix_socket_name; now it's actually used - remove `listen on socket off' and `unix_socket off' options that are now no longer useful - unify the internal handling of the sockets, i.e. don't consider unix socket special. ok tracey@

During `rcctl stop gotwebd' the sockets process gets a SIGTERM and dies, and this makes the parent process assume it exited abnormally. Instead, catch SIGINT and SIGTERM and exit gracefully. Issue reported by xs on IRC. ok jamsek

While here, remove completely the libevent handling of SIGPIPE in favour of the already existing signal(3) call to ignore it; there's no point in logging a SIGPIPE.

It only had 4 slots so was never quite useful, and sharing of sock->pack_fds across cached repositories seems problematic. with help from + ok op@

proc.c really shines when there's a network of different types of processes, potentially with a various number of instances each, that needs to exchange messages. Gotwebd instead has a much simpler design, and using proc.c causes more overhead (/headaches) than it solves. So, this attempts to provide the same functionalities but with a much simpler implementation that fits gotwebd better. ok stsp@

Instead of hardcoding the flags, keep ai_{family,socktype,protocol} from the getaddrinfo() call and apply them here. ok stsp@

It was added due to gotwebd weird structure sockaddr_storage handling. Instead, save the size reported by getaddrinfo() and not reach into the struct sockaddr_storage at all (except for extracting the port number for diagnostics purposes.) sockets_conf_new_socket_fcgi() gets an hardcoded ipproto to zero (which is the only value it can get in practice, and keeps for the moment the hardcoded SOCK_STREAM. It'll be cleaned in a follow-up. ok stsp@

The main process doesn't need write access anywhere and /tmp is not needed. The sockets process didn't use unveil at all so far, but is happy with just "x" for libexec helpers and "r" for each server's repository path. Input from op@ Tested by myself and Kyle Ackerman who also reviewed the diffs. Prompted by questions from mlarkin@ ok mlarkin@, op@

Reduces the indirection in fcgi.c, starts to make the struct template opaque, simplifies the template usage. All with a net negative :-) reads fine to stsp@ (thanks!)

ok jamsek, stsp

ok stsp@

simplify gotweb_process_request more, handling all the pages inside the big switch. There's only one entrypoint for rendering the templates gotweb_render_page() that takes the page' content as argument. The only real difference is that gotweb_render_index() now skips directory entries which fails to handle. ok tracey@

If a short read happens, or if all the fastcgi record to read don't fit in the buffer, gotwebd fails to continue reading and hits the timeout. If this happens before gotweb_process_request is called, it will crash in gotweb_free_transport since c->t will be NULL. This register the event with EV_PERSIST so fcgi_request is called again when there's more to read and guards gotweb_free_transport. It also makes spaces for the records as soon as they're successfully parsed. With lots of help from stsp ok stsp@

ok tracey@

ok op@

ok tracey

ok tracey

With the previous patch the listen port was correct, but the debug log output was still displaying the swapped port number. Now both listen behaviour and debug log output agree.

ok tracey

ok tracey

ok millert@