Implementation experience has shown that the original plan for async wait
fds was too simplistic. Originally the async logic created a pipe internally
and user/engine code could then get access to it via API calls. It is more
flexible if the engine is able to create its own fd and provide it to the
async code.
Another issue is that there can be a lot of churn in the fd value within
the context of (say) a single SSL connection leading to continually adding
and removing fds from (say) epoll. It is better if we can provide some
stability of the fd value across a whole SSL connection. This is
problematic because an engine has no concept of an SSL connection.
This commit refactors things to introduce an ASYNC_WAIT_CTX which acts as a
proxy for an SSL connection down at the engine layer.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Add no-async option to Configure that forces ASYNC_NULL.
Related to RT1979
An embedded system or replacement C library (e.g. musl or uClibc)
may not support the *context APIs that are needed for async operation.
Compiles with musl. Ran unit tests, async tests skipped as expected.
Signed-off-by: Rich Salz <rsalz@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
This was done by the following
find . -name '*.[ch]' | /tmp/pl
where /tmp/pl is the following three-line script:
print unless $. == 1 && m@/\* .*\.[ch] \*/@;
close ARGV if eof; # Close file to reset $.
And then some hand-editing of other files.
Reviewed-by: Viktor Dukhovni <viktor@openssl.org>
In the async code for MacOS/X define _XOPEN_SOURCE (if not already
defined) as early as possible. We must do this before including
any header files, because on MacOS/X <stlib.h> includes <signal.h>
which includes <ucontext.h>. If we delay defining _XOPEN_SOURCE
and include <ucontext.h> after various system headers are included,
we are very likely to end up with the wrong (truncated) definition
of ucontext_t.
Also, better error handling and some code cleanup in POSIX fibre
construction and destruction. We make sure that async_fibre_makecontext()
always initializes the fibre to a state that can be freed.
For all implementations, check for error returns from
async_fibre_makecontext().
Reviewed-by: Matt Caswell <matt@openssl.org>
Implements Thread Local Storage in the windows async port. This also has
some knock on effects to the posix and null implementations.
Reviewed-by: Rich Salz <rsalz@openssl.org>
In theory the pthreads approach for Thread Local Storage should be more
portable.
This also changes some APIs in order to accommodate this change. In
particular ASYNC_init_pool is renamed ASYNC_init_thread and
ASYNC_free_pool is renamed ASYNC_cleanup_thread. Also introduced ASYNC_init
and ASYNC_cleanup.
Reviewed-by: Rich Salz <rsalz@openssl.org>
A lot of the pool handling code was in the arch specific files, but was
actually boiler plate and the same across the implementations. This commit
moves as much code as possible out of the arch specific files.
Reviewed-by: Rich Salz <rsalz@openssl.org>
We were using _pipe to create a pipe on windows. This uses the "int" type
for its file descriptor for compatibility. However most windows functions
expect to use a "HANDLE". Probably we could get away with just casting but
it seems more robust to use the proper type and main stream windows
functions.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Even with _XOPEN_SOURCE defined OS-X still displays warnings that
makecontext and friends are deprecated. This isn't a problem until you
try and build with --strict-warnings, and the build fails. This change
suppresses the warnings. We know they are deprecated but there is no
alternative!
Reviewed-by: Rich Salz <rsalz@openssl.org>
async_fibre_makecontext was initialise the fibre first and then calling
getcontext(). It should be the other way around because the getcontext
call may overwrite some of the things we just initialised. This didn't
cause an issue on Linux and so the problem went unnoticed. On OS-X it
causes a crash.
Reviewed-by: Rich Salz <rsalz@openssl.org>
_longjmp/_setjmp do not manipulate the signal mask whilst
longjmp/setjmp may do. Online sources suggest this could result
in a significant speed up in the context switching.
Reviewed-by: Rich Salz <rsalz@openssl.org>
The ASYNC null implementation has not kept pace with the rest of the async
development and so was failing to compile.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Initial API implemented for notifying applications that an ASYNC_JOB
has completed. Currently only s_server is using this. The Dummy Async
engine "cheats" in that it notifies that it has completed *before* it
pauses the job. A normal async engine would not do that.
Only the posix version of this has been implemented so far, so it will
probably fail to compile on Windows at the moment.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Where we can we should use longjmp and setjmp in preference to swapcontext/
setcontext as they seem to be more performant.
Reviewed-by: Rich Salz <rsalz@openssl.org>
Create a "null" async implementation for platforms that lack support. This
just does nothing when called and therefore performs synchronously.
Reviewed-by: Rich Salz <rsalz@openssl.org>
