On Thu, 2010-03-04 at 14:53 +0100, Pavel Machek wrote:
> Hi!
>
> > > journalling assumptions broken: commit block is there, but previous
> > > blocks are not intact. Data loss.
> > >
> > > ...and that was the first I could think about. Lets not do
> > > this. Barriers were invented for a reason.
> >
> > Very well. Then we still need a solution to the original problem:
> > Devices sometimes need to be unregistered during resume, but
> > del_gendisk() blocks on the writeback thread, which is frozen until
> > after the resume finishes. How do you suggest this be fixed?
>
> Avoid unregistering device during resume. Instead, return errors until
> resume is done and you can call del_gendisk?
This won't help ether. The same driver needs to unregister perfectly
working device on suspend, because the user might replace the card
during suspend and fool the os.
There is a setting, CONFIG_MMC_UNSAFE_RESUME and I use it, but it isn't
default.
Anyway to revive that old thread, how about introducing new
del_gendisk_no_sync?
A less safe version of del_gendisk, but which won't sync the filesystem.
Since driver knows that card is gone, there is no point of syncing it.
(the sync is done by invalidate_partition, so some flag should be
propagated to it).
Best regards,
Maxim Levitsky
On Fri, 4 Jun 2010, Maxim Levitsky wrote:
> On Thu, 2010-03-04 at 14:53 +0100, Pavel Machek wrote:
> > Hi!
> >
> > > > journalling assumptions broken: commit block is there, but previous
> > > > blocks are not intact. Data loss.
> > > >
> > > > ...and that was the first I could think about. Lets not do
> > > > this. Barriers were invented for a reason.
> > >
> > > Very well. Then we still need a solution to the original problem:
> > > Devices sometimes need to be unregistered during resume, but
> > > del_gendisk() blocks on the writeback thread, which is frozen until
> > > after the resume finishes. How do you suggest this be fixed?
> >
> > Avoid unregistering device during resume. Instead, return errors until
> > resume is done and you can call del_gendisk?
>
> This won't help ether. The same driver needs to unregister perfectly
> working device on suspend, because the user might replace the card
> during suspend and fool the os.
> There is a setting, CONFIG_MMC_UNSAFE_RESUME and I use it, but it isn't
> default.
People have generally agreed that the best answer is to have
del_gendisk always thaw the writeback thread.
> Anyway to revive that old thread, how about introducing new
> del_gendisk_no_sync?
>
> A less safe version of del_gendisk, but which won't sync the filesystem.
> Since driver knows that card is gone, there is no point of syncing it.
>
> (the sync is done by invalidate_partition, so some flag should be
> propagated to it).
That might work for mmc, but it wouldn't help other drivers subject to
the same problem.
Besides, it's subject to races. What if the card _isn't_ gone, but for
some other reason the driver wants to unregister the device at a time
when the writeback thread is frozen?
Alan Stern
On Fri, 2010-06-04 at 10:59 -0400, Alan Stern wrote:
> On Fri, 4 Jun 2010, Maxim Levitsky wrote:
>
> > On Thu, 2010-03-04 at 14:53 +0100, Pavel Machek wrote:
> > > Hi!
> > >
> > > > > journalling assumptions broken: commit block is there, but previous
> > > > > blocks are not intact. Data loss.
> > > > >
> > > > > ...and that was the first I could think about. Lets not do
> > > > > this. Barriers were invented for a reason.
> > > >
> > > > Very well. Then we still need a solution to the original problem:
> > > > Devices sometimes need to be unregistered during resume, but
> > > > del_gendisk() blocks on the writeback thread, which is frozen until
> > > > after the resume finishes. How do you suggest this be fixed?
> > >
> > > Avoid unregistering device during resume. Instead, return errors until
> > > resume is done and you can call del_gendisk?
> >
> > This won't help ether. The same driver needs to unregister perfectly
> > working device on suspend, because the user might replace the card
> > during suspend and fool the os.
> > There is a setting, CONFIG_MMC_UNSAFE_RESUME and I use it, but it isn't
> > default.
>
> People have generally agreed that the best answer is to have
> del_gendisk always thaw the writeback thread.
Now the question is how to do that? :-)
Best regards,
Maxim Levitsky
>
> > Anyway to revive that old thread, how about introducing new
> > del_gendisk_no_sync?
> >
> > A less safe version of del_gendisk, but which won't sync the filesystem.
> > Since driver knows that card is gone, there is no point of syncing it.
> >
> > (the sync is done by invalidate_partition, so some flag should be
> > propagated to it).
>
> That might work for mmc, but it wouldn't help other drivers subject to
> the same problem.
>
> Besides, it's subject to races. What if the card _isn't_ gone, but for
> some other reason the driver wants to unregister the device at a time
> when the writeback thread is frozen?
>
> Alan Stern
>
On Fri, 4 Jun 2010, Maxim Levitsky wrote:
> On Fri, 2010-06-04 at 10:59 -0400, Alan Stern wrote:
> > On Fri, 4 Jun 2010, Maxim Levitsky wrote:
> >
> > > On Thu, 2010-03-04 at 14:53 +0100, Pavel Machek wrote:
> > > > Hi!
> > > >
> > > > > > journalling assumptions broken: commit block is there, but previous
> > > > > > blocks are not intact. Data loss.
> > > > > >
> > > > > > ...and that was the first I could think about. Lets not do
> > > > > > this. Barriers were invented for a reason.
> > > > >
> > > > > Very well. Then we still need a solution to the original problem:
> > > > > Devices sometimes need to be unregistered during resume, but
> > > > > del_gendisk() blocks on the writeback thread, which is frozen until
> > > > > after the resume finishes. How do you suggest this be fixed?
> > > >
> > > > Avoid unregistering device during resume. Instead, return errors until
> > > > resume is done and you can call del_gendisk?
> > >
> > > This won't help ether. The same driver needs to unregister perfectly
> > > working device on suspend, because the user might replace the card
> > > during suspend and fool the os.
> > > There is a setting, CONFIG_MMC_UNSAFE_RESUME and I use it, but it isn't
> > > default.
> >
> > People have generally agreed that the best answer is to have
> > del_gendisk always thaw the writeback thread.
> Now the question is how to do that? :-)
Here's a start:
http://marc.info/?l=linux-kernel&m=127378922620074&w=2
It's not quite right, because it needs to make the writeback thread
unfreezable before thawing it.
Alan Stern