Today's Agenda:
	Block devices

Interlude:  Memory-mapping files?
	Use mmap, munmap, msync, etc
	Not for expanding files
		You can use write to append, and then unmap and mmap
		My demo didn't support this, but it'd be pretty easy
	read/write/lseek sort of supports random access
		Not so natural as this
	This is *probably* faster for a lot of things

Is memory mapping actually different?
	Caching:  The file isn't really left on the disk

Weird example:  Files with holes
	We can make one
	I *might* have done this in CS253

Block vs. character device
	"random" access
		- random depends on your perspective
	Block device divided into sectors
		Sector vs. block

Using dd

/dev conventions, hda, sda, nvme, etc
	And symbolic links 

Blocks are represented as a buffer
	Size of a block:  At least a sector, no more than a memory page
	sector:  Hardware division, property of the hard drive
	block:  Software division, property of the operating system
	struct buffer_head defines one
	Flags for state

bio structure
	"Block I/O"

I/O schedulers
	Minimize head movement
	merging and sorting
	Independent of process scheduler
		- Process makes request
		- I/O scheduler gets around to it
	Merge if requests are for very close sectors
	Sort by proximity to head
		- Without reversing head direction if possible
		- Like an elevator
		- Sometimes called I/O elevator
	The Linus Elevator:
		- Default in 2.4, replaced in 2.6
		- New requests are checked against all others for merging
		- Merge onto front OR back depending
		- If no merging, added to sorted point in queue
		- If no good sorted point, add to the end of the queue
		- If tail is old, new requests go at the end regardless
			+ To prevent infinite starvation
	The Deadline I/O scheduler:
		- Writes starving reads
		- Requests have expiration time
			+ 500 ms for read, 5 seconds for write
		- Extra queue for each, sorted by time
		- If a request expires in the extra queue, it is serviced
		- Generally services requests by expiration time
	The Anticipatory I/O Scheduler
		- Heavy write load may perform badly on Deadline scheduler
		- Waits a few ms after performing read before resuming write
			+ In case the application needs something else
		- Keeps track of process access patterns to determine wait
	Completely Fair Queuing I/O Scheduler
		- Each process has a queue
		- Merging is performed, across queues (duplicates, references)
		- Round-robin with selection of a few items from each queue
		- Generally works well with non-pathological cases
		- Intended for multimedia originally
		- Default I/O scheduler as of book publication
	Noop I/O Scheduler
		- Merge requests, but don't do ANY sorting
		- Simpler than the others
		- Intended for flash memory only (no head position)