Assorted Project 4 Notes

Joshua Wise

Quick overview

• P4 this semester
• How it all fits together
• P4 and the real world

Project 4

• "What's Project 4 this semester?"
• A small extension to your Project 3 kernel...
• Sound Blaster 16 driver
• Kernel components
• Driver and syscall interface
• Shouldn't be too difficult concurrencywise...
• ...but the hardware...

Why SB16?

• Hugely popular card
• About 2347845923854 clones
• Everybody advertised being "Sound Blaster Compatible"
• Emulated by everybody, too
• ...but Simics...
• ISA
• Saves you the hassle of learning about PCI
• Easily programmable
• Designed to be programmed in assembly

Overview of the SB16

Command interface distinct from data interface DMA interface off-chip Cannot talk directly to memory! You get to program each of the bits

The SB16

• Where is it?
• You get to go looking for it
• Has 'command queues'
• Playback
• "Wake me up after n samples"
• Reads samples from the DMA controller, who knows nothing about this number n
• n carefully chosen so that you can refill the buffer

The DMA Talk

• Who here knows what DMA stands for?
• Mechanism for a device to get data from the host
• Without intervention on each byte, word, qword, ...
• Basic idea
• Kernel says "Here's how to get to some memory, have at"
• Some time later, device comes back

A Device's Worth

• DMA engine usually tightly coupled to each device
• How about 'on the same chip'?
• Bus-master DMA
• Mechanism:
• Device asks for a word
• DMA engine says either "OK!", or "Uh-oh"
• Host sets up DMA engine, then tells DMA engine and device to go ahead.

DMA buffer strategies

• Simplest?
• Buffer, length, done.
• A little better?
• List of buffers and lengths
• What's better than a list?
• Especially if the hardware guys are stuck with implementing it?
• Linked list of buffers and lengths!
• Used on many network cards
• Some buffers can have flags associated with them

Typical scatter-gather setup

ISA DMA

• Like everything else on the PC, DMA is sad
• Logic very expensive in the late 1980s
• DMA controller goes with the bus, not the card
• What you will implement is a crazy part of PC history, and not "how DMA works" in the real world.

Your buffer

• Back to the SB16!
• DMA controller will be programmed
• You tell it:
• Location of buffer (over the course of many writes)
• Size of buffer
• It feeds samples and loops

Buffer strategy, part 1

• Naive strategy?
• Fill it all up, then tell the sound card to interrupt you after reading the whole thing
• When woken up, you go fill it up again

Buffer strategy, part 1.5

• What went wrong?
• Well, the interrupt didn't really get there in time, maybe
• Or even if it did, by the time you started copying data...
• Old data got played, and it sounds pretty bad.

Buffer strategy, part 2

• We really need two buffers
• While we're filling one, the card is playing from the other
• We're guaranteed not to intersect the card's play pointer!

Buffer strategy, part 2.5

• ...except our DMA controller sucks.
• We only get one buffer. :(
• So let's split the buffer in the middle.
• Tell the card the buffer is half as big as it is

Not just 2

• Since the buffer loops, it might as well be one long buffer
• Let's consider a different strategy:
• Interrupt every N
• For N <<<< the buffer size, we can see it like this:

Pick a N, any N

• If we get some interrupt, what can we tell?
• We know where the card just played from... ...so we know what's not in use... ...so we know what we can fill.
• This is another possible strategy.
• Which you choose is a design decision.
• (Sorry.)

In Summary

Sound is fun Buffers are a pain DMA controllers are weird You're lucky you're not using PCI Have fun with this!