Replace C HTTP parser with JS HTTP parser #1457
Commits on Jun 21, 2015
-
-
-
-
http: improve parser performance
* Relax header value and status reason parsing * Avoid using RegExp for folded line detection * Use faster: string trimming, case-insensitive string comparing, CRLF searching, and case-insensitive header value searching
-
-
-
http: fix max chunk size string len
Anything more than 14 hex characters definitely has a larger value than our max chunk size.
-
-
-
benchmark: fix V8 warmup for http parser benchmarks
I think it's more accurate to "warm up" the one parser instance instead of creating many new throw-away instances since instances are typically reused in core.
-
-
benchmark: improve http parser benchmark more
This increases the loop count for the timed region and removes a couple of unlikely inputs.
-
-
http: remove chunk size string length checking
This was initially designed to verify we didn't have more digits than JavaScript can handle, but it's erroneously reporting errors when there are chunk extensions on the same line. The check's usefulness is semi-questionable anyway, so let's remove it for now.
-
-
http: allow double quotes and non-ASCII bytes in request URI
As the comments explain, these are not technically allowed, but since joyent/http-parser previously allowed them, we support them for better backwards compatibility.
-
-
-
-
-
http: fix whitespace preservation if folding is encountered
Addtitionally this re-checks special headers in case the values we check for were on folded lines.
-
-
This should now directly reflect how joyent/http-parser does keepalive checking.
-
-
http: ensure first byte is printable ASCII
This helps detect bad input early to prevent a "hung" state where the parser keeps looking for a CRLF. Generally this shouldn't be an issue, but could happen if TLS handshaking data is mistakenly written to the parser for example.
-
-
Typically the functions that return Error objects already have set the internal error state, so there is no need to do it twice.
-
-
-
-
-
-
-
-
http: use max js int value - 1 for content/chunk lengths
Before this it was possible that a Content-Length or chunk length of 9007199254740993 to be interpreted as 9007199254740992, making it impossible to know that the value was larger than our maximum. Using maximum js int value - 1 allows us to reliably detect when the max length has been exceeded.
-
http: remove unnecessary internal state variable
The folded state is no longer needed since we are now always deferring trimming of header values until the next new header or the end of the headers, whichever comes first.
-
-
-
This mainly fixes parsing of successive similar message types when there is a fixed-length body.
-
-
-
-
-
http: move keepalive boolean to flag
This improves performance some by not having to (re)set another instance property.
-
The ternary expression is faster, but may not be noticeable.
-
-
-
-
-
-
test: move http parser durability test to pummel
These tests take significantly longer to execute. Also the http 0.9 test has been corrected to reflect the capability of the current js parser.
-
-
http: switch to regexp-less, LF-compatible js parser
See this[1] comment for more details and benchmark results. Notable changes: * Removed all regexps. Only request URLs are not validated currently. * Added globally configurable start line length limits * Chunk size parsing no longer line buffers * Lines can end in just LF too * Faster than the original regexp-based implementation [1] nodejs#1457 (comment)
-
-
-
-
-
-
test: port URL tests from joyent/http-parser
This only tests basic URL validation because the JS HTTP parser does not actually parse the URL into its individual parts.
-
http: move property initialization to prototype
This bumps the large-alternate parser benchmark to ~19k+, making it match the result from joyent/http-parser 2.5.0.
-
-
http: switch to regexp search for longer field names
Somewhere between 10 and 14 characters, the equalsLower() algorithm becomes slower than searching with a simple, case-insensitive, non-anchored regexp. This at least gives a consistently higher bump in the large-alternate parser benchmark.
-
http: remove parser pause()/resume() usage
The js parser pause()/resume() are no-ops anyway.
-
-
