Ludovic Courtès writes: > Hi, > > Christopher Baines skribis: > >> In f50f5751fff4cfc6d5abba9681054569694b7a5c, the way fetch was called within >> process-substitution was changed. As with-cached-connection actually includes >> important error handling for the opening of a HTTP request (when using a >> cached connection), this change removed some error handling. >> >> This commit adds that error handling back, >> with-cached-connection/call-with-cached-connection is back, rebranded as >> call-with-fresh-connection-retry. >> >> * guix/scripts/substitute.scm (process-substitution): Retry once for some >> errors when making HTTP requests to fetch substitutes. > > [...] > >> + (define (call-with-fresh-connection-retry uri proc) >> + (define (get-port) >> + (open-connection-for-uri/cached uri >> + #:verify-certificate? #f)) >> + >> + (let ((port (get-port))) >> + (catch #t >> + (lambda () >> + (proc port)) >> + (lambda (key . args) >> + ;; If PORT was cached and the server closed the connection in the >> + ;; meantime, we get EPIPE. In that case, open a fresh connection >> + ;; and retry. We might also get 'bad-response or a similar >> + ;; exception from (web response) later on, once we've sent the >> + ;; request, or a ERROR/INVALID-SESSION from GnuTLS. >> + (if (or (and (eq? key 'system-error) >> + (= EPIPE (system-error-errno `(,key ,@args)))) >> + (and (eq? key 'gnutls-error) >> + (eq? (first args) error/invalid-session)) >> + (memq key '(bad-response bad-header bad-header-component))) >> + (begin >> + (close-port port) ; close the port to get a fresh one >> + (proc (get-port))) >> + (apply throw key args)))))) > > I think this should be at the top level for clarity. It used to have > ‘cached’ in its name because catching all these exceptions is something > you wouldn’t normally do; it only makes sense in the context of cached > connections. I initially tried to just put the error handling in just where it's needed, but that was difficult since the http-fetch bit needs to happen again when there's a relevant error. The two things: getting a port which maybe is a cached connection and handling some errors plus potentially re-running proc is difficult to capture in a name, but "call-with-cached-connection-and-error-handling" is an improvement over "with-cached-connection" I think. >> (define (fetch uri) >> (case (uri-scheme uri) >> ((file) >> @@ -424,11 +450,13 @@ the current output port." >> (call-with-connection-error-handling >> uri >> (lambda () >> - (http-fetch uri #:text? #f >> - #:open-connection open-connection-for-uri/cached >> - #:keep-alive? #t >> - #:buffered? #f >> - #:verify-certificate? #f)))))) >> + (call-with-fresh-connection-retry >> + uri >> + (lambda (port) >> + (http-fetch uri #:text? #f >> + #:port port >> + #:keep-alive? #t >> + #:buffered? #f)))))))) > > Does ‘call-with-connection-error-handling’ still make sense here? > There’s already ‘with-networking’ at the top level to do proper > networking error reporting. So, looking back, the call-with-connection-error-handling error handling was related to (call-)with-cached-connection, but it was only relevant inside of fetch-narinfos, as that's when open-connection-for-uri/maybe was passed in to call-with-cached-connection. Which means no, I think it can be removed, at least that's more consistent with the older behaviour. I'll send some updated patches. > Regarding , I would lean towards > perhaps reverting the connection/error-handling patch series and > starting anew from that “known state”. > > This area is unfortunately quite tedious to test and to get right so I’d > err on the path of conservative, incremental changes. > > Thought? My preference is still to try and move forward and to make the error handling easier to see in the code. Particularly with this change, I think the problem was introduced in this commit [1], but I think it's hard to tell from the diff, since the error handling and retrying is within with-cached-connection. 1: https://git.savannah.gnu.org/cgit/guix.git/commit/?id=f50f5751fff4cfc6d5abba9681054569694b7a5c That commit was one of the commits where I was making small incremental changes prior to actually getting to the changes I was looking at making, but a breakage was still introduced. What I was thinking about with this patch was how to make the error handling being added back here easier to see, and thus harder to break/remove.