PR 430/rtems
* include/rtems/score/watchdog.h: _Watchdog_Ticks_since_boot should
be a VOLATILE variable.
* src/watchdoginsert.c: 'restart' algorithm needs to enforce
reloading the list head in case a TICK interrupt during ISR_Flash()
modified the list. This is achieved by a proper VOLATILE cast.
Also _Watchdog_Sync_count++ should be protected by _ISR_Disable
(prevent corruption in case ISR calls watchdoginsert)
* src/watchdogadjust.c: ISR protection added.
* src/watchdogtickle.c: ISR protection added.
NOTE: PowerPC BSPs using the new exception processing MUST BE UPDATED
to maintain _ISR_Nest_level. See also PR288 which provides fixes
for the affected BSPs distributed with RTEMS.
PR 356/bsps
* src/threadhandlerc: This patch makes RTEMS/PowerPC eabi compliant.
Let Thread_Handler verify that _init/_main are non-NULL
before calling them (gives the linker script/user more
freedom to handle special cases).
* src/coretodset.c: Correct calculation of ticks until next section
boundary. It was incorrectly based upon current time not the
time that is being set.
* include/rtems/score/apimutex.h (_API_Mutex_Lock): Per PR253
add the missing _ISR_Disable. This fix was already applied to
the old location (c/src/exec/score/...).
* macros/rtems/score/object.inl: Corrected implementation of
_Objects_Open, _Objects_Close, and _Objects_Namespace_remove
to be consistent with the inline implementation.
* Mega patch merge to change the format of the object IDs to
loosen the dependency between the SCORE and the various APIs.
There was considerable work to simplify the object name management
and it appears that the name_table field is no longer needed.
This patch also includes the addition of the internal mutex
which is currently only used to protect some types of allocation
and deallocation. This significantly can reduce context
switch latency under certain circumstances. In particular,
some heap/region operations were O(n) and had dispatching
disabled. This should help enormously. With this merge,
the patch is not as clean as it should be. In particular,
the documentation has not been modified to reflect the new object
ID layout, the IDs in the test screens are not updated, and
_Objects_Get_information needs to be a real routine not inlined.
As part of this patch a lot of MP code for thread/proxy blocking
was made conditional and cleaned up.
* include/Makefile.am, include/rtems/score/coremsg.h,
include/rtems/score/coremutex.h, include/rtems/score/coresem.h,
include/rtems/score/object.h, include/rtems/score/threadq.h,
inline/rtems/score/object.inl, inline/rtems/score/thread.inl,
macros/rtems/score/object.inl, src/Makefile.am, src/coremsg.c,
src/coremutex.c, src/coresem.c, src/mpci.c,
src/objectcomparenameraw.c, src/objectextendinformation.c,
src/objectinitializeinformation.c, src/objectnametoid.c,
src/thread.c, src/threadclose.c, src/threadget.c, src/threadq.c,
src/threadqextractwithproxy.c: Modified as part of above.
* include/rtems/score/apimutex.h, src/objectgetnoprotection.c: New
files.
* include/rtems/score/thread.h, inline/rtems/score/thread.inl,
src/threaddispatch.c, src/threadinitialize.c:
Move the C library re-enterrant support directly into
the thread dispatch code. RTEMS needs libc and so requiring
libc to use a user extension with its overhead is not the best
solution. This patch lowers the overhead to 2 pointer moves.
* src/threaddispatch.c, src/threadhandler.c: Per PR211 fix
saving/restoring floating point context. The fpsave and fprestore
routines are only used in a executing context which _is_ fp and hence
has the FPU enabled. The current behavior required the FPU always to
be on which is very dangerous if lazy context switching is used.
[Joel Note: Some ports explicitly enabled the FPU in the FP save and
restore routines to avoid this.]
The patch also makes sure (on powerpc only) that the FPU is disabled
for integer tasks. Note that this is crucial if deferred fp context
switching is used. Otherwise, fp context corruption may go undetected!
Also note that even tasks which merely push/pop FP registers to/from
the stack without modifying them still MUST be FP tasks - otherwise
(if lazy FP context switching is used), FP register corruption (of
other, FP, tasks may occur)!
Furthermore, (on PPC) by default, lazy FP context save/restore
is _disabled_.
* include/rtems/score/object.h, inline/rtems/score/object.inl,
src/objectcomparenamestring.c: Address PR81 that
reworked POSIX message queues to add a descriptor separate from
the underlying message queue. This allows non-blocking to follow
the "open" not the underlying queue. As part of debugging this
it became clear that _Objects_Compare_name_string was broken
and a simple version using strncmp() was substituted.
* src/threadinitialize.c: Per PR181, clear the array of user extension
pointers. This lets user extensions that have hooked the switch handler
know if a task has been processed by the user extension before. If a
user extension is created after a task is started it may not know it.
* macros/rtems/score/userext.inl: Now works after merging patch for
functionality requested in PR174.
* inline/rtems/score/userext.inl: Added a comment explaining the
order in which routines appear since it is not the obvious order.
* Per PR141 and PR174, make task switch extension its own list and
fix all odd problems introduced by providing macro version.
* inline/rtems/score/userext.inl: Fix.
* macros/rtems/score/userext.inl: Updated to reflect modifications
to inline version from PR142.
* inline/rtems/score/userext.inl: Cleanup as side-effect of above.
* Per PR142, make task switch extension its own list.
* include/rtems/score/userext.h: Reflect above by adding
User_extensions_Switch_control and adding it to User_extenions_Control.
* inline/rtems/score/userext.inl: Allocate all memory in one chunk
to minimize overhead. Address processing dedicated switch chain.
* Per PR142, make task switch extension its own list.
* include/rtems/score/userext.h: Reflect above by adding
User_extensions_Switch_control and adding it to User_extenions_Control.
* inline/rtems/score/userext.inl: Allocate all memory in one chunk
to minimize overhead. Address processing dedicated switch chain.
* include/rtems/system.h: Only include cpuopts.h when building a
multilib configuration. Some ports still need targopts.h but this
small modification lets those ports work non-multilib while
fixing being fixed for multilib.
* inline/rtems/score/object.inl, macros/rtems/score/object.inl: Add
add casts to Objects_Id in _Objects_Build_ids to avoid implicit
typecasts from enum to int16 on bit16 targets (here: h8300).
This was tracked as PR91.
* include/rtems/score/isr.h, inline/rtems/score/isr.inl,
macros/rtems/score/isr.inl: Modified to allow any port to provide
its own implementation of the macro _ISR_Is_in_progress. If the
port overrides this macro, it must provide a non-inlined function
implementation.
* src/threadhandler.c: When __USE__MAIN__ is defined by the toolset,
invoke the global constructors via __main. Reported as tested by
Alexandra Kossovsky <sasha@oktet.ru> and Victor V. Vengerov
<vvv@oktet.ru> in conjunction with a new set of tool RPMs
(gcc2.95.3newlib1.9.0-3). This was tracked as GNATS PR tools/84.
Reported by Todor.Todorov@barco.com and tracked as PR36.
* include/rtems/score/object.h: Added prototype for
_Objects_Get_by_index().
* src/objectget.c, src/objectgetisr.c: Corrected procedure for
getting index from Id so it is correct and optimal for both single
and multiprocessor configurations.
