h2 perfromance regression for HTTP/1 cache

[krizhanovsky]

Scope

Caching

• The original report is Non-keepalive HTTP requests handling #1415 (comment) : there are perf profile symbols for hpack in pure HTTP/1 traffic. We sacrificed HTTP/1 performance to make HTTP/2 faster, but still need to explore the the profile and improve all the things around.

$ ./wrk --latency -c 1000 -d 1000 -t 2 https://debian:443 and Tempesta FW caching 2-bytes file produce following top profile:

     3.01%  [tempesta_fw]   [k] tfw_hpack_cache_decode_expand
     2.83%  [tempesta_fw]   [k] tfw_http_msg_expand_data

HTTPS connections handling

• tfw_cli_cache() allocates HTTP/2 context even for HTTP/1 connections. Done in HTTP/2 vs HTTPS handling #1643

Testing

Just make a performance test with profiling. Probably, keep-alive connections case will show the problem better as it emphasises the cache functionality and reduces VM overhead.