BACKGROUND:
We have learned that on some systems timeout timers are inaccurate and
might occasionally fire off too early. To make the multi_socket API work
with this, we made libcurl execute timeout actions a bit early too if
they are within our MULTI_TIMEOUT_INACCURACY. (added in commit
2c72732ebf, present since 7.21.0)
Switching everything to the multi API made this inaccuracy problem
slightly more notable as now everyone can be affected.
Recently (commit 21091549c02) we tweaked that inaccuracy value to make
timeouts more accurate and made it platform specific. We also figured
out that we have code at places that check for fixed timeout values so
they MUST NOT run too early as then they will not trigger at all (see
commit be28223f35 and a691e044705) - so there are definitately problems
with running timeouts before they're supposed to run. (We've handled
that so far by adding the inaccuracy margin to those specific timeouts.)
The libcurl multi_socket API tells the application with a callback that
a timeout expires in N milliseconds (and it explicitly will not tell it
again for the same timeout), and the application is then supposed to
call libcurl when that timeout expires. When libcurl subsequently gets
called with curl_multi_socket_action(...CURL_SOCKET_TIMEOUT...), it
knows that the application thinks the timeout expired - and alas, if it
is within the inaccuracy level libcurl will run code handling that
handle.
If the application says CURL_SOCKET_TIMEOUT to libcurl and _isn't_
within the inaccuracy level, libcurl will not consider the timeout
expired and it will not tell the application again since the timeout
value is still the same.
NOW:
This change introduces a modified behavior here. If the application says
CURL_SOCKET_TIMEOUT and libcurl finds no timeout code to run, it will
inform the application about the timeout value - *again* even if it is
the same timeout that it already told about before (although libcurl
will of course tell it the updated time so that it'll still get the
correct remaining time). This way, we will not risk that the application
believes it has done its job and libcurl thinks the time hasn't come yet
to run any code and both just sit waiting. This also allows us to
decrease the MULTI_TIMEOUT_INACCURACY margin, but that will be handled
in a separate commit.
A repeated timeout update to the application risk that the timeout will
then fire again immediately and we have what basically is a busy-loop
until the time is fine even for libcurl. If that becomes a problem, we
need to address it.
HTTP Pipelining with libcurl
============================
Background
Since pipelining implies that one or more requests are sent to a server before
the previous response(s) have been received, we only support it for multi
interface use.
Considerations
When using the multi interface, you create one easy handle for each transfer.
Bascially any number of handles can be created, added and used with the multi
interface - simultaneously. It is an interface designed to allow many
simultaneous transfers while still using a single thread. Pipelining does not
change any of these details.
API
We've added a new option to curl_multi_setopt() called CURLMOPT_PIPELINING
that enables "attempted pipelining" and then all easy handles used on that
handle will attempt to use an existing pipeline.
Details
- A pipeline is only created if a previous connection exists to the same IP
address that the new request is being made to use.
- Pipelines are only supported for HTTP(S) as no other currently supported
protocol has features resemembling this, but we still name this feature
plain 'pipelining' to possibly one day support it for other protocols as
well.
- HTTP Pipelining is for GET and HEAD requests only.
- When a pipeline is in use, we must take precautions so that when used easy
handles (i.e those who still wait for a response) are removed from the multi
handle, we must deal with the outstanding response nicely.
- Explicitly asking for pipelining handle X and handle Y won't be supported.
It isn't easy for an app to do this association. The lib should probably
still resolve the second one properly to make sure that they actually _can_
be considered for pipelining. Also, asking for explicit pipelining on handle
X may be tricky when handle X get a closed connection.
980659a2ca