mirror of
https://github.com/MariaDB/server.git
synced 2025-01-19 05:22:25 +01:00
60 lines
2.6 KiB
Text
60 lines
2.6 KiB
Text
|
|
||
|
|
Here are some notes on the internals of the implementation.
|
||
|
|
|
||
|
|
LIBC routines.
|
||
|
|
|
||
|
|
Unfortuanately many of the libc routine return a pointer to static data.
|
||
|
|
There are two methods to deal with this. One write a new routine where the
|
||
|
|
arguments are different, and have one argument be a pointer to some space
|
||
|
|
to store the data, or two use thread specific data and warn the user that
|
||
|
|
data isn't valid if the calling thread is terminated.
|
||
|
|
|
||
|
|
INTERNAL LOCKING
|
||
|
|
To prevent deadlocks the following rules were used for locks.
|
||
|
|
|
||
|
|
1. Local locks for mutex queues and other like things are only locked
|
||
|
|
by running threads, at NO time will a local lock be held by
|
||
|
|
a thread in a non running state.
|
||
|
|
2. Only threads that are in a run state can attempt to lock another thread,
|
||
|
|
this way, we can assume that the lock will be released shortly, and don't
|
||
|
|
have to unlock the local lock.
|
||
|
|
3. The only time a thread will have a pthread->lock and is not in a run
|
||
|
|
state is when it is in the reschedule routine.
|
||
|
|
4. The reschedule routine assumes all local locks have been released,
|
||
|
|
there is a lock on the currently running thread (pthread_run),
|
||
|
|
and that this thread is being rescheduled to a non running state.
|
||
|
|
It is safe to unlock the currently running threads lock after it
|
||
|
|
has been rescheduled.
|
||
|
|
5. The reschedule routine locks the kernel, sets the state of the currently
|
||
|
|
running thread, unlocks the currently running thread, calls the
|
||
|
|
context switch routines.
|
||
|
|
6 the kernel lock is used only ...
|
||
|
|
|
||
|
|
|
||
|
|
7. The order of locking is ...
|
||
|
|
|
||
|
|
1 local locks
|
||
|
|
2 pthread->lock /* Assumes it will get it soon */
|
||
|
|
3 pthread_run->lock /* Assumes it will get it soon, but must release 2 */
|
||
|
|
4 kernel lock /* Currently assumes it will ALWAYS get it. */
|
||
|
|
|
||
|
|
8. The kernel lock will be changed to a spin lock for systems that
|
||
|
|
already support kernel threads, this way we can mutiplex threads onto
|
||
|
|
kernel threads.
|
||
|
|
9. There are points where the kernel is locked and it needs to get
|
||
|
|
either a local lock or a pthread lock, if at these points the code
|
||
|
|
fails to get the lock the kernel gives up and sets a flag which will
|
||
|
|
be checked at a later point.
|
||
|
|
10. Interrupts are dissabled while the kernel is locked, the interrupt
|
||
|
|
mask must be checked afterwards or cleared in some way, after interrputs
|
||
|
|
have been reenabled, this allows back to back interrupts, but should always
|
||
|
|
avoid missing one.
|
||
|
|
|
||
|
|
------------------------------------------------------------------------------
|
||
|
|
Copyright (c) 1994 Chris Provenzano. All rights reserved.
|
||
|
|
This product includes software developed by the Univeristy of California,
|
||
|
|
Berkeley and its contributors.
|
||
|
|
|
||
|
|
For further licencing and distribution restrictions see the file COPYRIGHT
|
||
|
|
included in this directory.
|