| Age | Commit message (Collapse) | Author |
|---|
|
Merge v2.0.23 and v2.0.24 into master.
|
|
The write counters were close to useless because they were only
incremented on a write "event" which is only triggered when writing
would block.
Read handling was a little more useful in that all reads must happen
through the socket engine, so these were happening at the correct time,
but we can clean this up by doing it in the SE itself rather than each
platform port.
This means that both read and write events are now easily and usefully
defined as "a syscall of either read or write was attempted".
We also count empty read and write events as being an event, because
they still were an attempt to poll a socket in some way. This may help
to identify "bad" code which is repeatedly trying to read a socket for
some reason.
Lastly, we check for failed read/write calls, and log them as an error
event. A lot of the time, this is how sockets are determined as being
disconnected (ie. at read/write time).
While we're at it, split Update() in two to make the calls more
self-describing. This has no real impact since only one call is made at
a time anyway.
|
|
|
|
|
|
|
|
|
|
|
|
Remove enum EventType
|
|
|
|
Do not exit if we can't determine it
|
|
on GetMaxFds()
From man epoll:
Since Linux 2.6.8, the size argument is unused, but must be greater than zero. (The kernel dynamically sizes the required data structures without needing this initial hint.)
|
|
We can cope with fds greater than SocketEngine::GetMaxFds() since 3752b3f59d5216d7dc6221a361efc76b9ad2273d
|
|
|
|
|
|
select()
|
|
This removes the need to call GetRef() for every socket in DispatchEvents()
|
|
|
|
|
|
|
|
|
|
|
|
Move timespec to DispatchEvents() in kqueue, there is no reason for it to be a member variable
|
|
|
|
|
|
removing an fd
|
|
|
|
|
|
|
|
Use vectors that grow as necessary instead of mass allocating everything at once
Rework poll engine logic to make sense
|
|
correct events
|
|
|
|
|
|
|
|
When getting a write event for an EventHandler that wanted a single write only, remove it from the WriteSet (socketengine_select)
Fixes #556 reported by @WindowsUser
|
|
|
|
|
|
* 'const char*' to 'const std::string&'.
* snprintf to std::string concatenation.
* Replace duplicated OneOfMatches with InspIRCd::MatchMask.
|
|
|
|
|
|
|
|
|
|
Thanks to fraggeln for reporting and testing
|
|
|
|
|
|
-Fix x64 builds for Windows. Now all configurations compile.
-Remove the non-working rebase stuff.
-Remove the Windows fork hack and instead use FreeConsole() to emulate the behavior. This directly allows us to compile with ASLR, which is turned on now.
-Remove the old IPC mechanism for the removed GUI. This is not needed anymore as the GUI wasn't ever supported on anything newer than 1.2
-Remove the WIN32/WINDOWS macros. _WIN32 is supported on all x86-based VC++ targets, so that's what we need.
-Enable optimizations for release builds.
-De-duplicate printf_c(), it was previously copy-pasted into colors.h for configure
-Add the VC++ specific bad files in .gitignore
-Disable PID writing on Windows. This is only making sense for *nix builds.
-Replace the CPU usage retrieval with an algorithm analogous to the *nix behavior. Also supports separated now/total values. (Tested with a dummy busy loop - seems working)
-Removed certain unused functions and variables
-Remove stdint defines from the windows wrapper
-Remove CRT debug alloc. This is a bad idea as it would define a macro to replace free which breaks builds.
-Re-evaluated the warnings list, commented it.
-Moved inspircd_config/_version to include/ to match *nix
-Removed the creation of inspircd_se_config, as it isn't used at all.
-Made non-git builds show as "r0" instead of "r" (thanks to @SaberUK for pointing this out)
-Fixed up m_spanningtree's project paths. Now all configurations (debug/release x86/x64) have been tested and build properly.
-Moved FindDNS out of the wrapper and matched its log behavior with *nix. (It's pointless having it in the wrapper after the recent slimming down)
-Replaced random/srandom wrappers with a mechanism that tries to use Windows' Random API first is no SSL module is loaded.
-Removed more old junk from support for compilers older than VC++ 2010 (we don't have project files for these, so compiling them would be hard anyways)
-Removed the unused ClearConsole()
-Removed unused includes from the wrapper. Also, do not include psapi.h here if we don't link psapi.lib. This should be done where appropriate.
-Made inet_aton an inline function for increased performance
-C4800, performance warning about bool forcing, resolved at all occurrences.
-C4701, uninitialized variable 'cached', resolved at all occurrences.
-dlerror() was migrated out of the wrapper for more thread safety (no global buffer being shared) and increased performance.
-Removed the wrong CRT debug flags. This drains a lot of performance.
-Removed the clock_gettime/gettimeofday wrappers
-Replaced all TCHAR/ANSI mix-ups of functions with the correct respective function.
-Added a block of C4355 for < VS2012
-Update project files for c870714
|
|
|
|
EPOLLET is the highest bit, so storing EPOLLET in a signed value leads to
undefined behaviour.
|
|
|
|
|
|
The maximum number of file descriptors was incorrectly using the OS maximum,
not the process maximum. Thanks to Liath for finding & testing.
|
