Here are the signal changes from the 5.3 branch that optimizes signal
handler registration and switches from longjmp to siglongjmp in order
to make signal mask handling consistent across different UNIX operating
systems.
This adds proto_num to request_info. It is defaulted to HTTP 1.0 (1000)
such that it has a valid value even if the underlying sapi doesn't set it
correctly. It is then used to determine if a 302 or a 303 should be sent
on a Location redirect. Any non GET/HEAD HTTP 1.1 redirect will get a 303
instead of a 302 to be compatible with the HTTP spec.
may have ways of getting the request time without the extra syscall, but
for now let's just make sure we don't crash and people will eventually
fill these in where applicable.
- periodic update of date strings
- caching of Last-Modified values
- use of smart_str'ings for crafting HTTP header (static files)
and inside make_log_entry
- buffered log writing
- httpd_ntoa is about 8 times faster
Overall improvement: Around 50% faster now
uploads of up to 2GB on 32 bit platforms.
Uploads >16KB are put into a file-backed mmap area.
SG(request_info).content_type got corrupted somewhere. As a workaround,
we provide SAPI with a duplicate of the original string.
IRCG. These could cause thttpd to start a second request in the same
connection context, and thereby causing real damage.
Mozilla 1.0.1 is buggy in that context: When HTTP/1.1 pipelining is
enabled (defaults to off), it will send any number of requests over
a persistent connection (which is fine), even after it has received
a "Connection: close" header field in a subsequent response header.
This blatantly violates RFC 2616, section 8.1.2. Because it cannot
receive any response on the dead connection, the download manager
pops up and tries to download a file (which never arrives).
Also, we don't try to send a 400 message anymore, if the connection
dies.
