Commit a6251a43 changed the serial device to check open_flag
instead of flag. So we should change the rt_kprintf counterpart as well.

ARMCC will define __GNUC__ in "GNU mode"(with --gnu command line
option). So iff __GNUC__ && !__CC_ARM is GCC.

Use signed int as array index is unsecure.

The rt_list_isempty has prototype of "int rt_list_isempty(const rt_list_t *l)".
So the compiler has a good reason that the rt_thread_defunct list does
not change within rt_thread_idle_excute thus optimize the "while" loop
into a "if".

So add the volatile qualifier when test against the rt_thread_defunc list.

relation=>relocation

When thread wake up from waiting for memory, there is a chance that
there is no memory available in high pressure. So use a loop to check
again. Otherwise, there will be a NULL reference.

[Kernel] The rt_object_find routine can support to find an object inside a module. The usage is: rt_object_find("module_name/object_name").

Time out result and other errors is recorded in thread->error. Dirty
error will screw up the error handling code after the thread has been
wake up.

Out side world is difficult to tell whethere there is dead thread
remaining. If rt_thread_idle_excute only do one cleanup, it's hard to
finish cleanups outside the idle thread. So let is loop and do all the
cleanups in one call.

Out side world may be interested in whether the scheduler is locked or
not. Provide an API to tell that.

Mutex has the idea of ownership, only the thread which owns the mutex
can release it. So rt_mutex_release could only be called in thread
context.  Add a debug guard to it.

Skip list is a "random" data structure that in high possibilities it
would get O(log(N)) time complexity in inserting while the old list get
O(N). Forthermore, when set RT_TIMER_SKIP_LIST_LEVEL to 1, it will just
the same as the old double linked list, both in time and space
complexity.

Benchmarks shows that when RT_TIMER_SKIP_LIST_LEVEL is 3, the average
time of random insertion of new timer is about 2 times faster than the
old timer when there are 100 timers and 3 times faster when there are
200 timers.

However, it restores the deprecated funcion rt_system_timer_init. BSPs
must invoke it upon system startup.

In thread context means: 1) the scheduler has been started; 2) not in
interrupt context. It is more stronger than RT_DEBUG_NOT_IN_INTERRUPT.
With this commit, you will catch the error on situations like taking
mutex before scheduling instead of crashing on NULL pointer reference.

vsnprintf is a common string function that could be used in many places.
Using both vsnprintf in libc and vsnprintf in the RTT could make a
bigger image. Moreover, if newlib is not enabled when compiling with
GCC, referencing vsnprintf will lead to link error:

    .../arm-none-eabi/lib/armv7-ar/thumb/softfp/libc.a(lib_a-sbrkr.o):
    In function `_sbrk_r':
    sbrkr.c:(.text._sbrk_r+0xc): undefined reference to `_sbrk'
    collect2: error: ld returned 1 exit status

Using rt_vsnprintf could avoid such problem.

Either RT_EVENT_FLAG_AND or RT_EVENT_FLAG_OR should be set in the option
flag. If none of them is present, the waiting thread will never be
resumed.

When print the thread names, we should use "%.*s" instead of bare "%s".

When realloc a bigger space and the next node is free and big enough, we
should directly relocate the next node instead of doing alloc/memcpy.
The new method not only faster in this cases, it would avoid memory
fragment as well.

This is a simple work around to the current device stack design. A
ref_count could let different modules to open/close the same device
independently without interfere others in some degree.

But there is still some data shared between the modules, like flag,
open_flag and user_data. Moreover, it won't yield an error if A open a
device, and B read from it before open it in B. Maybe alloc a new handle
in rt_device_open will be the ultimate solution. But that is much bigger
change and we may leave it to future development.

convert Tabs to spaces