Re: [Netatalk-devel] AFP 3.3

[Alex d. <ale...@gm...>]

On 2009-08-29, at 12:44 PM, Frank Lahm wrote:

>>
>
> What is TMNIS?

Have a look through:

http://developer.apple.com/mac/library/documentation/NetworkingInternetWeb/Conceptual/TimeMachineNetworkInterfaceSpecification/TimeMachineRequirements/TimeMachineRequirements.html#//apple_ref/doc/uid/TP40008951-CH100-SW1

It talks about various requirements of TM, which is nice given that I  
don't think they've ever been advertized before.  Maybe this will help  
with various people's problems not being able to use netatalk as a  
backup device with 10.6...


Here's what I mean about the problems with the replay cache....  
guaranteed write completion is a fundamental part of TM over AFP, and  
Linux's fflush() doesn't do that.

Let's say there's a flood of data coming from the client, and an  
FSyncFork comes in.  The expectation from the client is that this will  
block until the data is known to have been written to spinning media  
on the server.  Specifically on Linux (but not BSD), fflush() doesn't  
actually guarantee that the data is written, it may still be sitting  
around in memory.  If somebody does hard power the server between an  
FPSycnFork and when the data gets written to disk, data will be lost.   
With the growth of netatalk and time machine usage, this is inevitable.

I haven't completely groked the replay cache yet, but I think it may  
make it worse if it allows out-of-sequence DSI replies.  No matter  
what, you'll have a lot more going on and more risk of hitting the  
above.

But the replay cache can offer an increase in reliability; if the  
server goes down and hasn't completed a request, the client just  
replays it.

- A