2009-09-04 17:33:00

by Chaitanya Lala

[permalink] [raw]
Subject: Disk failure behavior

Hi,

I am using a back-port of libata from ~ 2.6.20 on a 2.6.9
Red Hat kernel. I have SATA disks (using AHCI) in the
system which are hot-pluggable. The problem I am facing
is that, certain disk failures bring the system into a
weird state. The system tries to reset the disk but fails.
Finally it prints a message "reset failed, giving up."

At this point the port is left in a frozen state and
the interrupts from the port are masked. If now, this disk is
pulled out and a healthy disk is inserted, the new disk's
insertion does not raise any event/notification/interrupt.
In fact, the only way at this point to get the disk to work is
to reboot.

Below is a snippet of the code, I am referring to, from v2.6.20.
File - drivers/ata/libata-eh.c & function-name - ata_eh_recover

/* reset */
if (ehc->i.action & ATA_EH_RESET_MASK) {
ata_eh_freeze_port(ap);

rc = ata_eh_reset(ap, ata_port_nr_vacant(ap), prereset,
softreset, hardreset, postreset);
if (rc) {
ata_port_printk(ap, KERN_ERR,
"reset failed, giving up\n");
goto out;
}

ata_eh_thaw_port(ap);
}

A possible work-around is to thaw the port before going to "out".
That would enable the interrupts again before going to "out".
I understand that would enable future interrupts from the old disk as well,
but I am willing to live with that, if it helps to detect the new device.

/* reset */
if (ehc->i.action & ATA_EH_RESET_MASK) {
ata_eh_freeze_port(ap);

rc = ata_eh_reset(ap, ata_port_nr_vacant(ap), prereset,
softreset, hardreset, postreset);
if (rc) {
ata_port_printk(ap, KERN_ERR,
"reset failed, giving up\n");
+ ata_eh_thaw_port(ap);
goto out;
}

ata_eh_thaw_port(ap);
}

I have tested this successfully. But I would like to ask you if this would
possibly "break" some other functionality ? I am new to the kernel ata stuff
and want to be sure before I use this.

Thanks,
Chaitanya


2009-09-05 00:31:45

by Tejun Heo

[permalink] [raw]
Subject: Re: Disk failure behavior

Hello,

Chaitanya Lala wrote:
> I am using a back-port of libata from ~ 2.6.20 on a 2.6.9
> Red Hat kernel. I have SATA disks (using AHCI) in the
> system which are hot-pluggable. The problem I am facing
> is that, certain disk failures bring the system into a
> weird state. The system tries to reset the disk but fails.
> Finally it prints a message "reset failed, giving up."
>
> At this point the port is left in a frozen state and
> the interrupts from the port are masked. If now, this disk is
> pulled out and a healthy disk is inserted, the new disk's
> insertion does not raise any event/notification/interrupt.
> In fact, the only way at this point to get the disk to work is
> to reboot.

# echo - - - /sys/class/scsi_host/hostX/scan

should revive it too.

> Below is a snippet of the code, I am referring to, from v2.6.20.
> File - drivers/ata/libata-eh.c & function-name - ata_eh_recover
>
> /* reset */
> if (ehc->i.action & ATA_EH_RESET_MASK) {
> ata_eh_freeze_port(ap);
>
> rc = ata_eh_reset(ap, ata_port_nr_vacant(ap), prereset,
> softreset, hardreset, postreset);
> if (rc) {
> ata_port_printk(ap, KERN_ERR,
> "reset failed, giving up\n");
> goto out;
> }
>
> ata_eh_thaw_port(ap);
> }
>
> A possible work-around is to thaw the port before going to "out".
> That would enable the interrupts again before going to "out".
> I understand that would enable future interrupts from the old disk as well,
> but I am willing to live with that, if it helps to detect the new device.
>
> /* reset */
> if (ehc->i.action & ATA_EH_RESET_MASK) {
> ata_eh_freeze_port(ap);
>
> rc = ata_eh_reset(ap, ata_port_nr_vacant(ap), prereset,
> softreset, hardreset, postreset);
> if (rc) {
> ata_port_printk(ap, KERN_ERR,
> "reset failed, giving up\n");
> + ata_eh_thaw_port(ap);
> goto out;
> }
>
> ata_eh_thaw_port(ap);
> }
>
> I have tested this successfully. But I would like to ask you if this would
> possibly "break" some other functionality ? I am new to the kernel ata stuff
> and want to be sure before I use this.

Unless your controller causes IRQ storm bringing down the controller,
the above change shouldn't be dangerous.

--
tejun