Projects Remaining:
	Project 4:  Mouse Moving
	Project 5:  Network Monitor

Memory, Part 1, Userspace and Kernel Response

When a process starts, memory allocation:
	Stack
	Code segment
	Initial (empty) heap

Later:
	Stack growth
		What if we overdo it with the stack?  What happens?
	malloc family (and new, like "new myClass()")
	heap allocations are not guaranteed!
		How much space can we malloc?  Let's try it!
		Might need more than one process
	Swap space, memory limits, swapping, separate partition vs. system partition
		speed of storage devices
		silly rule of thumb

Kernel only enforces a few rules
	valgrind can check your program for you
	not perfect, but good

Inside The kernel 

How the Kernel Manages Process Memory
	This is more described in chapter 15
	Basic stuff about OS memory management:
		- Each process is presented with a "clean" virtual memory space
		- Hardware translation of virtual to real addresses
		- Since processes have separate address space, they can't change each other
		- The OS handles allocation in an "optimistic" manner
	Page Table:  http://www.tldp.org/LDP/tlk/mm/memory.html
	Swap partition
		- You can disable this, on Linux at least

Ok, Chapter 12 memory management
Pages:
	Described by struct page
	4KB on 32-bit usually, 8KB on 64-bit (usually)
		- Book numbers predate AMD64
	Note use of atomic_t
	Page struct describes physical memory
	One of these per 8KB uses some memory
		- Book says 40 bytes per page_struct
		- Might be more for us!  Let's count
		- What fraction of our memory is used for page structs?

Zones:
	Various architectures have constraints on memory
	Areas without 16-bit or 32-bit access, no DMA, high vs. low
	AMD64 is pretty clean in this regard
		- But legacy devices are funny.
	struct zone describes them

Using pages:
	You can get and free pages directly in the kernel
	Without using kmalloc
	kmalloc is more typical
		- It relinquishes direct physical control
		- But it works a lot like malloc
		- Can request less than a full page, or multiple pages
		- There's a matching kfree
	vmalloc acts similar, but won't always return physically contiguous allocations
		- It seems to be less popular, for performance reasons

Can we just use the stack?
	It's smaller than in userspace
	Expect 4 to 16 KB
		- "small"
	Don't use a recursive algorithm unless it does very few recursive calls
		- Or alloca, although this is probably not available anyway
	Stack overflow is not managed.  May crash, or act funny
	Let's try that!  NOT on the instructor station

Slab allocation:
	General idea:  Don't allocate and deallocate so much
		- Re-use existing structures
	An example from game engines
	Book example:  inodes
		- We haven't really talked much about the VFS yet
	Take stuff in-house for greater speed

Per-cpu issues:
	Multi-core CPUs have individual caches for each core
		- And a shared cache for all of them
	L1 vs. L2 vs. L3
	percpu macros allow creation of a variable with one instance per CPU
		- You can modify the copy of another CPU
		- Allows a way to keep a variable in L1 or a register
		- Multi-threading possibilities
		- During access, kernel preemption must be disabled
		- Can't sleep in the middle of accessing it
	You could do this with pthreads and expect it to work out ok 
		- One variable per thread