One of the biggest remaining performance problems in the httpd is
the code that scans HTTP requests. In the current implementation,
read_request_line() and ap_get_mime_headers() call ap_rgetline_core(),
which has to:
- create a temporary brigade
- call ap_get_brigade() to get the next line of input from the
core input filter, which in turn has to scan for LF and split
the bucket
- copy the content into a buffer,
- destroy the temp brigade
- call itself recursively in the (rare) folding case
And all this happens for every line of the request header.
We're creating a ton of temporary brigades and temporary buckets
in this process, plus registering and unregistering pool cleanup
callbacks and doing at least one memcpy operation per line.
I'd like to streamline this processing so that ap_read_request()
can do something like this:
- get the input brigade from the input filter stack
- loop through the brigade and parse the HTTP header
- Note: in the common case, the whole header will
be in the first bucket
- split the bucket where the request header ends,
remove the bucket(s) containing the header from
the brigade, and hand the remainder of the brigade
back to the filter stack
- Then use the bytes in the header bucket(s) as a
read-write copy of the request. (In the case where
the whole request header is in one bucket, we could
achieve zero-copy input for all the header fields
by null-terminating them in place and storing those
strings in r->headers_in, as long as we ensure that
the data lasts for the lifetime of the request.)
What's the best way to do this? Turn the reading of the
request header into an input filter? Or something more like
the current implementation, but with an AP_MODE_SPECULATIVE
read from the input filter stack to let ap_read_request()
peek at the entire request? Or something else entirely?
Thanks,
--Brian
