Re: Hello, I have a question about XFS File System

[Jeff Liu <jeff.liu@oracle.com>]

On 03/07 2014 12:19 PM, Dave Chinner wrote:
> On Fri, Mar 07, 2014 at 10:23:38AM +0800, Jeff Liu wrote:
>>
<snip>
>>>
>>> There is some documentation about some of the logging concepts and
>>> design. eg:
>>>
>>> http://oss.sgi.com/cgi-bin/gitweb.cgi?p=xfs/xfs-documentation.git;a=blob;f=design/xfs-delayed-logging-design.asciidoc
>>
>> Not sure if someone else also think that XFS journal design is the stumbling-block
>> to get involved into the development...but I once heard of "I'm really confused by
>> the design of delayed logging, I have to give up after reading the document for about
>> 2 or 3 weeks..." from 2 Chinese developers in the past year, though nothing can help
>> someone out without taking infinite patience.
> 
> Let me put it this way: it took me *5 years* of working deep in the
> XFS code to really understand how the XFS transaction and
> journalling subsystems are supposed to function. Delayed logging
> took me 3 failed design attempts over 3 years before I had learnt
> enough to come up with a design that worked.  It's by far the most
> complex part of XFS - expecting to understand how it works by
> spending a couple of weeks reading the code is unrealistic.

Hah, this would help me relax a lot when I felt frustrating to understand
something in XFS :-P.


> Fundamentally, understanding delayed logging means you have to first
> understand why relogging is necessary in the XFS journal. To
> understand why relogging is necessary, you first need to understand
> the transaction subsystem, the log space reservation subsystem, log
> recovery constraints, how tail pushing works, the physical log
> interface code, the on-disk log format, etc andhow they all
> interact.
> 
> IOWs, delayed logging is the last thing in the journalling layer
> that anyone should try to understand because understanding it fully
> requires a high level of knoweldge about the XFS metadata and
> logging subsystem architecture and fundamental principles....

Thanks for the nice guidance.

> 
>>> But the only way to learn about the actual structure of the log is to
>>> read the code and use xfs_logprint to study the contents of the log.
>>
>> To Yongmin,
>>
>> For your information only.
>>
>> I'm trying to understand XFS journal via the following steps:
>>
>> 1) Download Linux-2.6.34 source, read the journal code.
>> Understand the original design as there is no delayed-logging support at that time.
> 
> Delayed logging changes neither the journal nor the transaction
> layer code or design. If you can't understand the fundamental
> principles behind those subsystems from the current code, then
> looking at the older code won't make it any clearer because it is
> exactly the same...
> 
>> FYI, two obsoleted documents could be found at,
>> http://oss.sgi.com/projects/xfs/design_docs/xfsdocs93_pdf/log_mgr-overview.pdf
>> http://oss.sgi.com/projects/xfs/design_docs/xfsdocs93_pdf/log_mgr.pdf
> 
> The first of those really doesn't contain much useful information.
> 
> The second really only documents the physical log format. That might
> be useful as a first step, but it doesn't document any of the
> alogorithms that the log uses, and that is where all the complexity
> lies.

Actually, both documents are only a little useful to me when I began to
understand the semantics of in-core logs.

> 
> Reading code will only get you so far - the only way to continue the
> learning process is by trying to modify the code and fixing what you
> break, along with asking questions about things you don't understand
> on the list so that people who do understand them can teach you
> things that aren't obvious from the code and aren't documented
> anywhere other than the code.


Thanks,
-Jeff

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs