Skip to contentSkip to conversation

XFS to main kernel source

DISCUSSION Steve Lord 3 participants 3 messages Sep 21, 2001 – Nov 6, 2001 1 message not in the archive

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Steve Lord

author
#1 lore raw
Quoted 11 lines · Steve Lord
Quoted 3 lines

XFS quotas are transactional, when space is added to a file the quota is adjusted in the same transaction. It is fairly hard to do this without your own quota code.

Actually not. The quotas in ext3 are transactional as well. It's just that the "ext3" journal layer allows nested transactions, so it is possible to start a write transaction, call into the journal code which calls back into the ext3 write code to start a nested transaction on the journal file (i.e. it is in the same transaction as the initial write), and then the initial write completes.

OK, good point, but doing a major rewrite of XFS to use a different transaction mechanism is not really on the cards, plus we have on disk compatibility with the Irix version to consider.

Steve

#2 lore raw
In reply to a message not in the archive

hi,

Quoted 12 lines · Steve Lord
Quoted 11 lines · Steve Lord
Quoted 3 lines

XFS quotas are transactional, when space is added to a file the quota is adjusted in the same transaction. It is fairly hard to do this without your own quota code.

Actually not. The quotas in ext3 are transactional as well. It's just that the "ext3" journal layer allows nested transactions, so it is possible to start a write transaction, call into the journal code which calls back into the ext3 write code to start a nested transaction on the journal file (i.e. it is in the same transaction as the initial write), and then the initial write completes.

Quoted 4 lines

OK, good point, but doing a major rewrite of XFS to use a different transaction mechanism is not really on the cards, plus we have on disk compatibility with the Irix version to consider.

XFS also uses quite a different _model_ of quota, and it is tightly integrated into XFS (by its very nature). There are a number of issues that it attempts to address, and in particular it works around the inherent problems of the traditional quotacheck/mount/quotaon/quotaoff model (ie. the BSD quota model which the Linux VFS also uses, and which was used in IRIX for the EFS filesystem).

Its simply different, there are advantages and disadvantages of each way, but XFS is particularly aimed at scalability, and the desire to never have to run quotacheck(8) on large filesystems was one of the issues which the original design aimed to address.

So XFS quota should be a non-issue. I have had discussions with Jan Kara and Alan in the past about how to most cleanly integrate it with their (new) VFS quota, and they seem happy with the design we collectively came up with (its slightly different to the one in both the XFS patch and Alan's patch at the moment).

cheers.

Signature · 1 line
Nathan
#3 lore raw

Re: [PATCH] lp.c, eexpress.c jiffies cleanup

In reply to a message not in the archive
Quoted 8 lines · Andreas Dilger
Quoted 3 lines · Tim Schmielau

In eexpress.c I also turned absolute jiffies number into multiples of HZ, yet the resulting timeout values still do not always seem reasonable to me.

I agree. It seems very ugly. I looked at a few drivers which loop 1 or 2 jiffies, but to busy-loop for 1/10th of a second, or even 20 seconds is terribly bad.

Those timeouts are only a last resort. If the card is working properly the loop will terminate much sooner.

p.