SATA hotswap and write caches

SATA hotswap and write caches

[Michel Lespinasse]

Hi,

I have a few questions about disk cache synchronization. I'm running a
setup with several SATA drives (seagate ST3500630AS), installed in
hot-swap bays. The host controller is an Intel ICH8 running in AHCI
mode, if that makes any difference. Also my kernel is from kernel.org,
version 2.6.20.12.

When hot-removing a drive, I get messages such as:
Jun  1 23:48:50 server kernel: ata3.00: detaching (SCSI 2:0:0:0)
Jun  1 23:48:50 server kernel: Synchronizing SCSI cache for disk sdb: 
Jun  1 23:48:50 server kernel: FAILED
Jun  1 23:48:50 server kernel:   status = 0, message = 00, host = 4, driver = 00
Jun  1 23:50:22 server kernel:   <3>ata3: exception Emask 0x10 SAct 0x0 SErr 0x4050002 action 0x2 frozen

When the drive is plugged it mentions about caching:
Jun  1 23:50:32 server kernel: SCSI device sdb: write cache: enabled, read cache: enabled, doesn't support DPO or FUA.

I tried to google for more information about caching but could not come out
with much. Mainly I found http://linux-ata.org/shutdown.html but this only
talks about server poweron/down and suspend operations, not hotswap issues.

I'm a bit confused about this caching and have two questions:

* Is there anything one needs to know before hot-removing a drive ?
  Maybe a file under /sys that asks the drive to flush its cache ?

* Does the kernel issue such cache flushing commands when executing
  sync() or umount() operations ?

Please CC me in replies - I have subscribed to the list but did not get
the majordomo confirmation message yet so who knows if that worked :)

Thanks,

-- 
Michel "Walken" Lespinasse
"Bill Gates is a monocle and a Persian cat away from being the villain
in a James Bond movie." -- Dennis Miller

Re: SATA hotswap and write caches

[Jeff Garzik]

Michel Lespinasse wrote:
> * Is there anything one needs to know before hot-removing a drive ?

Make sure all data is flushed to the drive before removing it.


> * Does the kernel issue such cache flushing commands when executing
>   sync() or umount() operations ?

No.  (it should, IMO...)

	Jeff

Re: SATA hotswap and write caches

[Michel Lespinasse]

On Sun, Jun 03, 2007 at 10:24:38AM -0400, Jeff Garzik wrote:
> Michel Lespinasse wrote:
> >* Is there anything one needs to know before hot-removing a drive ?
> 
> Make sure all data is flushed to the drive before removing it.

Sure. Is there a way to do that, though, other than using /sbin/halt
which seems to have, errr, undesirable side-effects :)

Ideally it'd be very nice if that was doable through /sys or /proc files.
It'd also be nice if disk standby was user-controllable in this way too.
(or maybe both as one step, i.e. a way to tell the kernel we'd like to
take the disk out in the safest way possible).

-- 
Michel "Walken" Lespinasse
"Bill Gates is a monocle and a Persian cat away from being the villain
in a James Bond movie." -- Dennis Miller

Re: SATA hotswap and write caches

[Tejun Heo]

Michel Lespinasse wrote:
> On Sun, Jun 03, 2007 at 10:24:38AM -0400, Jeff Garzik wrote:
>> Michel Lespinasse wrote:
>>> * Is there anything one needs to know before hot-removing a drive ?
>> Make sure all data is flushed to the drive before removing it.
> 
> Sure. Is there a way to do that, though, other than using /sbin/halt
> which seems to have, errr, undesirable side-effects :)
> 
> Ideally it'd be very nice if that was doable through /sys or /proc files.
> It'd also be nice if disk standby was user-controllable in this way too.
> (or maybe both as one step, i.e. a way to tell the kernel we'd like to
> take the disk out in the safest way possible).

If you're using barrier, you should be in the clear, I think.

-- 
tejun