The query_classifier library is now only a wrapper that loads an
actual query classifier implementation. Currently it is hardwired
to load qc_mysqlembedded, which implements the query classifier
API using MySQL embedded.
This will be changed, so that the library to load is specified
when qc_init() is called. That will then allow the query classifier
to be specified in the config file.
Currently there seems to be a conflict between the mysql_library_end()
call made in qc_mysqlembedded and the mysql_library_end() call made in
gateway.c. The reason is that they both finalize a shared library.
For the time being mysql_library_end() is not called in gateway.c.
This problem is likely to go away by switching from the client
library to the connector-c library.
Earlier the same (or almost the same) code was duplicated in
several places.
A conflicting declaration was also removed. There was no
implementation for that declaration.
The polling statistics collection used atomic_add to increment values. This
is not an optimal way to update statistical values. Moved to per thread
values which are summed up when they are read.
Moved the functions used to gather polling statistics to their own file and
created a specific data type for statistics.
The THREAD type was not used everywhere and pthread_t was used instead.
The thread creation function also returned the address of a stack allocated
value which isn't guaranteed to be usable.
Only the query classifier needs the functionality of the embedded
server, while the rest of MaxScale is content with the client
library or Connector/C.
This have now been rearranged so that query-classifier links with
the embedded static library and then explicitly exports its own
functions using the query_classifier.map linker script. That way
query classifier will use the embedded library, while the rest of
maxscale use the client library, and this without conflicts.
Currently, query_classifier is not linked to maxscale-common,
but executables must link to maxscale-common and query_classifier.
Earlier thread initialization was done in poll_waitevents(),
which was somewhat confusing, since the main thread also calls
into poll_waitevents().
Now there is a specific thread "main" function, which first
performs thread specific initialization and then calls into
poll_waitevents().
This is the first change in an attempt to arrange the linking so that
more errors are detected at link-time.
- All files in server/core but for gateway.c are linked to one shared
library called maxscale-common.
- The files log_manager/log_manager.cc and utils/skygw_utils.cc are
built into maxscale-common as well.
- MaxScale itself consists now only of gateway.c and is linked with
maxscale-common.
- All plugins link with maxscale-common.
- All executables link in addition with {EMBEDDED_LIB}.
After this change, the change (MXS-517) where query_classifier is the
only component that uses ${EMBEDDED_LIB} and the rest mysqlclient can
be made much cleaner.
After a few additional steps, all shared libraries can be linked with
the linker flags "-Wl,-z,defs", which directs the linker to resolve
all symbols. That will require that all shared libraries list all the
libraries they need, but will at the same time ensure that any
missing symbols are detected at link-time and not at run-time.
The current implementation of idle connection timeouts is not safe. The sessions
are handled in a way which is not thread-safe and the checking is done from
a non-polling thread.
With this change, the checks for the session timeouts are done in one of the
polling threads in a thread-safe manner only if at least one service has enabled
the timing out of idle client connections.
Currently the initialization function only prepares PCRE2 patterns for use.
Added the call to the new utils_init() function to MaxScale and all relevant
tests.
The replace_quoted now uses a shared PCRE2 pattern which should remove some
of the overhead of the function.
The various global directory setter functions now process the input they receive
and remove redundant and trailing forward slashes from the directory paths.
Since the PCRE2 library was always going to be a part of MaxScale, there was
no real reason to have it as a shared library apart from smaller binaries.
Changed default number of threads to 1 instead of autoconfigured value and
added a new `auto` variable which enables autoconfiguration of thread count.
The number of threads used when autoconfiguratio fails was changed from 4 to 1.
The default value of using N threads where N is the number of CPU cores was
not optimal as the possibility of rescheduling was higher the more utility
threads there were. Due to this, N-1 is deemed to be the better autoconfigured
value for thread count.
When generating a fake hangup event, EPOLLRDHUP is used if available,
otherwise EPOLLHUP. process_pollq(int) does the same thing both in the
case of EPOLLRDHUP and EPOLLHUP, so it seems this should work.
