April 12th Lecture
=====================

Mapping of threads to kernel scheduling entities:
 * kernel:  "1:1 mapping"
 * user:    "N:1 mapping"
 * hybrid:  "N:M mapping"
    - one option: associate groups of threads with individual
      kernel processes

Scheduler activations:
 * kernel provides notifications ("upcalls") to notify a user-level scheduler
   of important events, but leaves their processing to user-space
 * no need for user-space libraries to "poll" the kernel with non-blocking IO
   system calls anymore
 * based on the idea of "virtual processors"
 * implementations:
    - FreeBSD: KSE ("Kernel Scheduled Entities") since 5.0
    - Linux: VPI ("Virtual Processor Interface") 3rd party patch independent
      from official kernel tree



Synchronization

Consider (assume x starts as 0):

T1              T2
==              ==
x = x + 1       x = x + 2

What is the final value of x?  3?  2?  1?
It could be any of the above.  Look at the assembly generated:

T1              T2
====            ====
load R0, X      load R1, X
add R0, #1      add R1, #2
store X, R0     store X, R1

Remember that a context switch could occur at any time.

T1              T2
====            ====
load R0, X 
add R0, #1 
                load R1, X
                add R1, #2
                store X, R1
store X, R0

All of T2's work is lost!