It turns out no browsers implement HTTP/2 GOAWAY handling properly, and large enough number of resources on a page results in failures to load some resources. In particular, Chrome seems to experience errors if loading of all resources requires more than 1 connection (while it is usually able to retry requests at least once, even with 2 connections there are occasional failures for some reason), Safari if loading requires more than 3 connections, and Firefox if loading requires more than 10 connections (can be configured with network.http.request.max-attempts, defaults to 10). It does not seem to be possible to resolve this on nginx side, even strict limiting of maximum concurrency does not help, and loading issues seems to be triggered by merely queueing of a request for a particular connection. The only available mitigation seems to use higher keepalive_requests value. The new default is 1000 and matches previously used default for http2_max_requests. It is expected to be enough for 99.98% of the pages (https://httparchive.org/reports/state-of-the-web?start=latest#reqTotal) even in Chrome.
Similar to lingering_time, it limits total connection lifetime before keepalive is switched off. The default is 1 hour, which is close to the total maximum connection lifetime possible with default keepalive_requests and keepalive_timeout.
Firefox uses several idle streams for PRIORITY frames[1], and "http2_max_concurrent_streams 1;" results in "client sent too many PRIORITY frames" errors when a connection is established by Firefox. Fix is to relax the PRIORITY frames limit to use at least 100 as the initial value (which is the recommended by the HTTP/2 protocol minimum limit on the number of concurrent streams, so it is not unreasonable for clients to assume that similar number of idle streams can be used for prioritization). [1] https://hg.mozilla.org/mozilla-central/file/32a9e6e145d6e3071c3993a20bb603a2f388722b/netwerk/protocol/http/Http2Stream.cpp#l1270
In current versions (all versions based on zlib 1.2.11, at least since 2018) it no longer uses 64K hash and does not force window bits to 13 if it is less than 13. That is, it needs just 16 bytes more memory than normal zlib, so these bytes are simply added to the normal size calculation.
In limit_req, auth_delay, and upstream code to check for broken connections, tests for possible connection close by the client did not work if the connection was already closed when relevant event handler was set. This happened because there were no additional events in case of edge-triggered event methods, and read events were disabled in case of level-triggered ones. Fix is to explicitly post a read event if the c->read->ready flag is set.
For connection close to be reported with eventport on Solaris, ngx_handle_read_event() needs to be called.
For new data to be reported with eventport on Solaris, ngx_handle_read_event() needs to be called after reading response headers. To do so, ngx_http_upstream_process_non_buffered_upstream() now called unconditionally if there are no prepread data. This won't cause any read() syscalls as long as upstream connection is not ready for reading (c->read->ready is not set), but will result in proper handling of all events.
Without explicit handling, a zero timer was actually added, leading to multiple unneeded syscalls. Further, sending GOAWAY frame early might be beneficial for clients. Reported by Sergey Kandaurov.