While this issue is visible in mysqli_poll() functions, the cause
lays deeper in the stream to socket casting API. On Win x64 the
SOCKET datatype is a 64 or 32 bit unsigned, while on Linux/Unix-like
it's 32 bit signed integer. The game of casting 32 bit var to/from
64 bit pointer back and forth is the best way to break it.
Further more, while socket and file descriptors are always integers
on Linux, those are different things using different APIs on Windows.
Even though using integer instead of SOCKET might work on Windows, this
issue might need to be revamped more carefully later. By this time
this patch is tested well with phpt and apps and shows no regressions,
neither in mysqli_poll() nor in any other parts.
stream when 0 is given as the value.
- PHP_STREAM_OPTION_WRITE_BUFFER no longer changes the chunk size in socket
streams.
- Added stream_set_chunk_size() function.
- Some signedness fixes.
- Test for commit r308474, now that it's possible to actually test it.
We avoid the problem by using poll(2).
On systems without poll(2) (older bsd-ish systems, and win32), we emulate
poll(2) using select(2) and check for valid descriptors before attempting
to access them via the descriptor sets.
If an out-of-range descriptor is detected, an E_WARNING is raised suggesting
that PHP should be recompiled with a larger FD_SETSIZE (and also with a
suggested value).
Most uses of select(2) in the source are to poll a single descriptor, so
a couple of handy wrapper functions have been added to make this easier.
A configure option --enable-fd-setsize has been added to both the unix and
win32 builds; on unix we default to 16384 and on windows we default to 256.
Windows FD_SETSIZE imposes a limit on the maximum number of descriptors that
can be select()ed at once, whereas the unix FD_SETSIZE limit is based on the
highest numbered descriptor; 256 should be plenty for PHP scripts under windows
(the default OS setting is 64).
The win32 specific parts are untested; will do that now.
using the default socket timeout of 60 seconds before returning the socket
to the calling script. The reason they were using that value is that the
same code is used for feof(), so the fix is allowing the caller to
indicate the timeout value for liveness checks.
A possible remaining issue now is that 0 second timeout[1] for pfsockopen
is possibly too short; it's impossible to specify a sane value for all
possible uses, so maybe we need a stream context or an .ini option to
control this, or maybe use the timeout value that was passed to
pfsockopen().
# [1] by timeout, I mean the time that PHP will wait for data on a
# persistent socket before deciding if a new connection should be made;
# NOT the timeout while waiting for a new connection to be established.