[Top] [All Lists]

Re: [RFC 00/17] RFC parent inode pointers.

To: Geoffrey Wehrman <gwehrman@xxxxxxx>
Subject: Re: [RFC 00/17] RFC parent inode pointers.
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Tue, 4 Feb 2014 11:09:44 +1100
Cc: Mark Tinguely <tinguely@xxxxxxx>, xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20140128220230.GF13953@xxxxxxx>
References: <20140115220012.624438534@xxxxxxx> <20140116055607.GR3431@dastard> <52D99FD2.6000601@xxxxxxx> <20140118031247.GE18112@dastard> <52E6B67B.6070001@xxxxxxx> <20140128030040.GG2212@dastard> <20140128220230.GF13953@xxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Tue, Jan 28, 2014 at 04:02:30PM -0600, Geoffrey Wehrman wrote:
> On Tue, Jan 28, 2014 at 02:00:40PM +1100, Dave Chinner wrote:
> | On Mon, Jan 27, 2014 at 01:41:47PM -0600, Mark Tinguely wrote:
> | > 2) Add the filename to EA. Not a fan, but I will ask but if DMF needs it
> | >    for performance then it has to be done. My point was this assumes
> | >    that we can keep all the links' EA entries inline in the inode. A
> | >    couple 255 character files or several links of modest sized filenames
> | >    would negate that assumption. I tried to minimize the EA entries to
> | >    keep them inline in the inode. I will talk to the DMF group.
> | 
> | Actually, I made the point about DMF needing them inline performance
> | because that's an argument SGI might find persuasive. What I didn't
> | say just then is that *I* need them inline, too, as online directory
> | tree scrubbing needs to be able to do bulks scans, as does
> | xfs_repair. However, I have idefinitely said this before in previous
> | parent poitner discussions, so it should be no surprise here...
> I appologize in advance for my ignorance.  What is "online directory
> tree scrubbing" and how does it benefit from parent inode pointers?

It will be a kernel thread that walks the directory tree
periodically verifying that all inodes are reachable and that the
directory tree is intact. It's the same principle as RAID array
media scrubbing - proactive detection of errors that could cause
failures and repairing them before an error is delivered to the
application. The reason for wanting parent pointers is that it
enables immediate repair of most corruptions as the scrubbing
detects them without requiring duplicate copies of the metadata.

IOWs, parent pointers are a fundamental part of bringing on-line
repair functionality to XFS. While parent pointers are not directly
mentioned in this document:


There is a section on reverse mapping for used blocks and how to use
that for online repair of detected damage. Parent pointers provide
reverse mappings for the directory structure, and hence provide the
same functionality. Parent pointers also means that if we get an IO
error in a block in a file, we can identify the owner of the block
by full path and offset into the file where the error has

The v5 metadata format has an owner field in all metadata that has a
single parent object for keeping such parent information, and it is
intended for improving scrub-based validation and for identifying
the owner of lost metadata blocks to improve recover from errors.
i.e. reverse mapping is a one of the fundamental architectural
requirements underlying the v5 metadata changes.

The two missing pieces are the parent pointers for directory
structure reverse mapping, and an AGF reverse mapping btree to
enable arbitrary block->owner lookups. I have prototype code for the
AGF rmap btree, and the parent pointers provide that functionality
for the directory structure. With both of these implemented, we will
be able to do fully-online filesystem metadata validity checking
equivalent to 'xfs_repair -n' or better.....

Reverse block mapping and directory structure parent pointers have
been considered necessary for robust exception handling and online
repair since we first started thinking about the CRC metadata format


Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>