Thanks Ray. In fact, I used CURLM APIs to multiplex multiple requests. So my problem is that some events that were sent during GOAWAY graceful shutdown will get 400 from the server. With GOAWAY callback, I could proactively open a new connection and migrate new quests to use it. Does that make sense?
On Fri, Jan 5, 2024 at 12:35 PM Ray Satiro via curl-library < curl-library@lists.haxx.se> wrote: > On 1/5/2024 11:28 AM, Cao Duc Quan via curl-library wrote: > > 1. Open a new HTTP2 connection by creating an HTTP GET to an url path. > > I created a CURLM object and added the easy handle for HTTP GET > > Request to it. > > 2. Next HTTP request will be sent in the HTTP2 connection created by > > step 1. I created a new easy handle for this request and added it to > > CURLM object created in step 1. > > 3. After a few hours or less, the connection at step 1 will be closed > > by server side (due to idle timeout or TTL). The server will send a > > GOAWAY frame to start Graceful shutdown before explicitly closing it > > after 60s so that both Client and Server could finish their in-flight > > transferring. > > 4. My library needs to detect the connection is closed and reopen a > > new one (step 1). > > > > So I have a problem at step 2: > > - If the connection created by step 1 received a GOAWAY frame, cURL > > will open a new HTTP2 to send the next HTTP request. The new HTTP2 > > request will fail because there is some logic check at server side to > > ensure only connection created by step 1 is allowed. I believe this is > > odd checking but that is the logic from the server side and I could > > not change it. > > > If the first connection has idled approximately 2 minutes (118 seconds) > [1] then curl should close it before before processing the next transfer > on the new connection. If a GOAWAY frame is received on the first > connection before then (either bundled with the previous request or due > to a timeout in the interim) it *should* do the same thing: > > * received GOAWAY, error=0, last_stream=1 > * Connection 0 seems to be dead > * The cache now contains 0 members > * Curl_disconnect(conn #0, dead=1) > * Closing connection > * TLSv1.2 (IN), TLS alert, close notify (256): > * TLSv1.2 (OUT), TLS alert, close notify (256): > > The connection cache logic for that is in function extract_if_dead [2]. > > nginx test #1: > - set `http2_max_requests 1;` > - `sudo systemctl reload nginx.service` to reload the config > - first curl_easy_perform > - first transfer reply and curl receives GOAWAY > - second curl_easy_perform > - curl opens a new connection for the second transfer > - second transfer reply > > nginx test #2: > - unset `http2_max_requests 1;` > - set `http2_idle_timeout 30s;` > - `sudo systemctl reload nginx.service` to reload the config > - first curl_easy_perform > - first transfer reply and curl does not receive GOAWAY > - wait 1 minute (GOAWAY packet sent by nginx during this time according > to wireshark) > - second curl_easy_perform > - curl receives GOAWAY before initiating the second transfer > - curl opens a new connection for the second transfer > - second transfer reply > > If curl is working as intended then I do not understand why your > application needs to handle GOAWAY on its own. Are you using the latest > curl? > > > [1]: https://curl.se/libcurl/c/CURLOPT_MAXAGE_CONN.html > [2]: https://github.com/curl/curl/blob/curl-8_5_0/lib/url.c#L767-L777 > > -- > Unsubscribe: https://lists.haxx.se/mailman/listinfo/curl-library > Etiquette: https://curl.se/mail/etiquette.html > -- -------------------------------- Watson Cao VN: (+84) 0976574864 CA: (+1) 2368658864
-- Unsubscribe: https://lists.haxx.se/mailman/listinfo/curl-library Etiquette: https://curl.se/mail/etiquette.html