Christoph Hellwig wrote:
On Tue, Nov 14, 2000 at 12:04:21PM +0100, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is not (struct inode *) - are there
any?
I can't find one. This change was requested some
Alexander Viro wrote:
On Tue, 14 Nov 2000, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is not (struct inode *) - are there
any?
Currently - none. As soon as block devices go to pagecache we'll have
On Wed, 15 Nov 2000, Daniel Phillips wrote:
OK, now I'm confused.
Will there be a sure way of knowing what kind of object -host points
to? An efficient way would be even nicer. Why would it wrong for
*every* address_space to have an inode and a superblock? Or at least,
Huh? What
Currently - none. As soon as block devices go to pagecache we'll have such
beasts (you need one address_space for all inodes with given major:minor,
so -host can't point to any of them). Possibly earlier, if somebody will
need address_space in some other situation where we need some
On Wed, 15 Nov 2000, Steve Lord wrote:
I would expect that once there was more than one type of object hiding
behind the address_space the only way to access it would be via the
address_space ops vector, just like the private inode data. As for the
Wrong tense ;-) s/was/will be/. Block
On Wed, 15 Nov 2000, Steve Lord wrote:
I would expect that once there was more than one type of object hiding
behind the address_space the only way to access it would be via the
address_space ops vector, just like the private inode data. As for the
Wrong tense ;-) s/was/will be/.
Daniel Phillips wrote:
Even if I can find the filesystem that belongs to a given page, how does
the filesystem find a (struct file *) for the page so it can do
address_space-writepage?
Never mind, I know what to do here - don't use the address_space
function, write my own. Not as clean as
On Wed, 15 Nov 2000, Steve Lord wrote:
On Wed, 15 Nov 2000, Steve Lord wrote:
I would expect that once there was more than one type of object hiding
behind the address_space the only way to access it would be via the
address_space ops vector, just like the private inode
On Wed, 15 Nov 2000, Daniel Phillips wrote:
Daniel Phillips wrote:
Even if I can find the filesystem that belongs to a given page, how does
the filesystem find a (struct file *) for the page so it can do
address_space-writepage?
Never mind, I know what to do here - don't use the
Steve Lord wrote:
Currently - none. As soon as block devices go to pagecache we'll have such
beasts (you need one address_space for all inodes with given major:minor,
so -host can't point to any of them). Possibly earlier, if somebody will
need address_space in some other
On Wed, 15 Nov 2000, Daniel Phillips wrote:
It's used in nfs and smbfs - that's as far as I've checked so far - I
wouldn't be surprised to find it's used in all the network filesystems.
My question is: why? Why can't the network filesystems make the
accomodation here, instead of all the
On Tue, 14 Nov 2000, Alexander Viro wrote:
On Tue, 14 Nov 2000, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is not (struct inode *) - are there
any?
Currently - none. As soon as block devices go to
On Wed, 15 Nov 2000, Rik van Riel wrote:
On Tue, 14 Nov 2000, Alexander Viro wrote:
On Tue, 14 Nov 2000, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is not (struct inode *) - are there
any?
On Wed, 15 Nov 2000, Alexander Viro wrote:
On Wed, 15 Nov 2000, Rik van Riel wrote:
On Tue, 14 Nov 2000, Alexander Viro wrote:
On Tue, 14 Nov 2000, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is
On Wed, 15 Nov 2000, Alexander Viro wrote:
But we _never_ had cache consistency between the fs and raw device
accesses.
Alexander,
do I understand you correctly, i.e. do you really mean "raw" device or do
you mean block device? If we don't have the cache consistency between
(pagecache-aware)
On Wed, 15 Nov 2000, Tigran Aivazian wrote:
On Wed, 15 Nov 2000, Alexander Viro wrote:
But we _never_ had cache consistency between the fs and raw device
accesses.
Alexander,
do I understand you correctly, i.e. do you really mean "raw" device or do
you mean block device? If we
Alexander Viro wrote:
On Wed, 15 Nov 2000, Rik van Riel wrote:
On Tue, 14 Nov 2000, Alexander Viro wrote:
On Tue, 14 Nov 2000, Daniel Phillips wrote:
After quite a lot of grepping in 2.4.0-test10 I was unable to find any
places where address_mapping-host is not (struct inode
Daniel Phillips wrote:
...At this
point we need to be to get the page's inode efficiently - it's sitting
there in the -host field, no problem, except now you're talking about
putting other things there with no way to know what they are. Do you
have another suggestion on how to find the
Alexander Viro wrote:
Anyway, yes, fs-wide metadata is an obvious user for that stuff. E.g.
for FFS-like beasts inode table just begs for such treatment. Waiting
for 2.5... Patch treats cylinder group descriptors as analogs of indirect
blocks - that allows to play with alternative allocation
On Thu, 16 Nov 2000, Daniel Phillips wrote:
The conversion from page back to inode needs the -host field to be an
inode, reliably. Why would we want to do such a conversion? It's when
the mm runs low on memory and wants to flush a dirty, inactive page.
When it finds one, it looks at the
On Thu, 16 Nov 2000, Daniel Phillips wrote:
Daniel Phillips wrote:
...At this
point we need to be to get the page's inode efficiently - it's sitting
there in the -host field, no problem, except now you're talking about
putting other things there with no way to know what they are. Do
On Thu, 16 Nov 2000, Daniel Phillips wrote:
For 4K blocksize on 4K page_size systems there isn't any wastage at all,
and 1K filesystems *should* be small anyway - the metadata even
smaller. It will be a little bit wasteful on 4K blocksize/8K page_size
Hell knows. There is only one way to
22 matches
Mail list logo
