| 1 |
On Tuesday 25 March 2008, Enrico Weigelt wrote: |
| 2 |
> * Alan McKinnon <alan.mckinnon@×××××.com> wrote: |
| 3 |
> > If SMART (or something conceptually similar) detects that a drive |
| 4 |
> > might be failing and be beyond the range of the drive's ability to |
| 5 |
> > cope, it could raise an event and move the blocks used to another |
| 6 |
> > disk. |
| 7 |
> |
| 8 |
> And it even would get funnier if the drive's relocation table |
| 9 |
> could be accessed (no idea if this is possible): |
| 10 |
> The LVM would notice if the drive has relocated an (LBA) block, |
| 11 |
> move it out of the way (somewhere else in the LV) and then |
| 12 |
> remove the relocation (never access that LBA block anymore). |
| 13 |
> This way an slowly dying disk can be used for quite a long time. |
| 14 |
> Think of boxes with very limited physical access (eg. outoor field |
| 15 |
> systems) or huge archives w/ non-critical/regeneratable data |
| 16 |
> (eg. media collections w/ originals available, mirrors, etc). |
| 17 |
> |
| 18 |
> The idea of using even old and damaged disks at really low costs |
| 19 |
> (not counting the power consumption ;-P) is seems quite fascinating |
| 20 |
> to me :) |
| 21 |
|
| 22 |
The tricky part is to figure out where a concept like this fits in the |
| 23 |
various abstraction layers. You couldn't build it into LVM, as LVM |
| 24 |
doesn't really know about individual blocks on the disk, it only works |
| 25 |
with it's own segments. The disk driver itself only really understands |
| 26 |
it's own IDE/SCSI commands and manipulates the data at the sector |
| 27 |
level. The file system can work at the block level. |
| 28 |
|
| 29 |
So there are at least 3 different allocation unit sizes and no obvious |
| 30 |
way to write one driver that spans all three. The one idea that keeps |
| 31 |
coming back to me is a new function in the kernel where a driver for a |
| 32 |
physical storage device can fire an event if it detects a failing unit, |
| 33 |
and other drivers can register to receive those events. When it gets |
| 34 |
one, the high level driver can decide if it should move it's own blocks |
| 35 |
of data around. In a way, conceptually similar to a GUI event |
| 36 |
framework. |
| 37 |
|
| 38 |
Could be very useful to a niche market - Linux already has 1000s of |
| 39 |
those :-) - but would require very careful design to not break |
| 40 |
everything else in the system |
| 41 |
-- |
| 42 |
Alan McKinnon |
| 43 |
alan dot mckinnon at gmail dot com |
| 44 |
|
| 45 |
-- |
| 46 |
gentoo-user@l.g.o mailing list |
Archives