On 11/11/2012 05:16 PM, Eric Sandeen wrote:
On 11/8/12 11:02 AM, Theodore Ts'o wrote:
On Thu, Nov 01, 2012 at 05:32:11PM -0200, Carlos Maiolino wrote:
I have not added all mount options I believe need to raise a
warning, just those with a flag set on superblock, but I expect to
generate a discussion here about which mount options should be
warned and get suggestions about how to deal with argumented options
(ex: commit= opt), to be included in a V2 of this patch.
Let's take a step back, and see if we can be explicit about why it
would be useful to warn when a user uses mount options that we were
not present with the implementation of the file system drivers found
in fs/ext2 and fs/ext3. While we're at it, we can also examine the
same question with respect to file system features --- i.e., if
someone mounts a file system with an extent feature enabled using
mount -t ext2, what if anything should we do. What are we trying to
achieve, or conversely what are we trying to prevent?
Suppose the user does something like this:
mount -t ext3 -o delalloc /dev/sdb /u1
Yes, the traditional ext3 driver code doesn't understand the delalloc
mount option, but what's the concern that is leading us to want to
print a warning message to the log (that the user may or may not even
see). Are we worried that what should happen is not sufficiently
well-defined? Are we worried that while it might work with a
particular kernel which is compiled a certain way, it might not work
with another kernel? Is it part of the larger question of wanting to
warn if the user is using a set of not-well-tested combination of
mount options and/or file system features?
If it is the latter, what is the right approach? At Barcelona, I was
chatting with Ric Wheeler and Ralf Flaxa, and they had differing
opinions about what the right thing to do for features which are not
supported. Ralf felt that warning in the syslog might be sufficient.
I sugested setting a new kernel taint flag, to make it easier for the
supper desk to twig to the fact that the user was doing something
unusual. Ric offerred his opinion that it was better to hard-fail the
mount. And of course, this is with distro kernels; for the upstream
kernel, I think our goal is to (a) warn the user that they are doing
something unusual, and (b) ask them to tell us (at linux-ext4) what
they are doing and why, so we have a better understanding of what
users want, so we can either add it to our test matrix, or perhaps
warn the user off.
The overarching reason is to cut down the test matrix if possible, I think.
This has advantages for both distros & upstream, and might even lead
to code simplification if we can exclude some corner-case behaviors.
This is my key goal - make sure that we get tested and sane mount options out to
the users. And we normally only test a handful of things, certainly it is not
reasonable to test all combinations of things you could do.
But the other question that might need answering is this:
If a user has asked ext4.ko to mount an ext3 device, should it behave
as closely as possible to what the ext3.ko driver used to do? And I'd
say yes. Therefore things like -t ext3 -o delalloc should be hard-failed.
If you really want your old ext3 filesystem to be mounted with delalloc,
then I think the right way would be to use mount -t ext4 /dev/ext3-dev.
Then you get the runtime behaviors such as delalloc, and ext4 should
not write anything which is not ext3-compatible.
Perhaps we need to go through in fine detail, but I think mount -t
ext3 should reject any options not applicable to ext3.ko. If you want
the newer ext4 behaviors, then use mount -t ext4.
I agree with that. At some point, that is probably what the user actually meant
(just did not know it at the time :)).
So not only do we need to decide which mount options and file system
features we want to support, or warn against etc., it's also useful to
think about what we want to do when the user does something a little
out of the norm.
And since I very much doubt that upstream, Red Hat, SuSE, Canonical,
etc., will ever agree on the right thing is, this is probably one of
those things where we want to have a scheme where it is relatively
easy for distributors to set their own policy, which may differ for a
community versus enterprise distro kernel.
Perhaps, but a well-thought-out rationale for what combinations make
sense might just fly everywhere.
A complex scheme of configuration policy sounds like potentially another
layer of knobs on an already vast layer of knobs. ;)
(That said, distros may well want to just nuke things like data=writeback
from orbit)
(And that said, I'm of the opinion that upstream should too ;))
Anyway, back to my main point: As a guiding principle I think I would
say that mount -t ext3 with ext4.ko should hard-reject any option not
understood by ext3.ko. It's clear and predictable, and should make for
a decent first cut.
-Eric
Agreed (sounds almost like we had coordinated our answers before I spoke to Ted
in Barcelona!).
Ric
But in order to make sure we don't end up talking past each other, I
think it's useful to be very explicit about why we want to do this,
before we try to figure what's and the how's.
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
