2013-09-25 08:31 keltezéssel, Robert Hancock írta:

quoted

On Sun, Sep 22, 2013 at 1:13 AM, Levente Kurusa [off-list ref] wrote:

quoted

2013-09-21 19:04 keltezéssel, Robert Hancock írta:

quoted

On Sat, Sep 21, 2013 at 1:35 AM, Levente Kurusa [off-list ref] wrote:

quoted

quoted

quoted

quoted

quoted

quoted

quoted

The following dmesg is stuck in an infinite loop.
dmesg:
ata3: lost interrupt (Status 0x50)
ata3.00: exception Emask 0x0 SAct 0x0 SErr 0x0 action 0x6 frozen
ata3.00: failed command: READ DMA
ata3.00: cmd c8/00:08:00:00:00/00:00:00:00:00/e0 tag 0 dma 4096
in
                   res 40/00:00:00:00:00/00:00:00:00:00/00 Emask
0x4
(timeout)
ata3.00: status: { DRDY }
ata3: soft resetting link
ata3.00: configured for UDMA/33 (no error)
ata3.00: device reported invalid CHS sector 0
ata3: EH complete

Patch that fixes the infinite loop:

diff --git a/drivers/ata/libata-eh.c b/drivers/ata/libata-eh.c
index f9476fb..eeedf80 100644
--- a/drivers/ata/libata-eh.c
+++ b/drivers/ata/libata-eh.c

@@ -2437,6 +2437,14 @@ static void ata_eh_link_report(struct

ata_link
*link)
                                 ehc->i.action, frozen,
tries_buf);
                     if (desc)
                             ata_dev_err(ehc->i.dev, "%s\n",
desc);
+               ehc->i.dev->exce_cnt ++;
+               ata_dev_warn(ehc->i.dev, "Number of exceptions:
%d\n",
ehc->i.dev->exce_cnt);
+               /**
+                  * The device is failing terribly,
+                 * disable it to prevent damage.
+                 */
+               if(ehc->i.dev->exce_cnt > 2)
+                       ata_dev_disable(ehc->i.dev);
             } else {
                     ata_link_err(link, "exception Emask 0x%x "
                                  "SAct 0x%x SErr 0x%x action
0x%x%s%s\n",

diff --git a/include/linux/libata.h b/include/linux/libata.h
index eae7a05..fa52ee6 100644
--- a/include/linux/libata.h
+++ b/include/linux/libata.h

@@ -660,7 +660,8 @@ struct ata_device {
             u8

devslp_timing[ATA_LOG_DEVSLP_SIZE];

             /* error history */
-       int                     spdn_cnt;
+       int                     spdn_cnt; /* Number of
speed_downs
*/
+       int                     exce_cnt; /* Number of exceptions
that
happenned */
             /* ering is CLEAR_END, read comment above CLEAR_END
*/
             struct ata_ering        ering;
      };

This doesn't seem like a very good fix. It may prevent the
apparent
infinite loop but will just prevent that device from functioning
at
all.
It would be better if we could figure out what was actually going
wrong.

I have tested the problem with three different computers, all
switched
to legacy/IDE/compatibility mode, and they didn't have this
problem.
Of
course, they could have been set to AHCI mode, and there the kernel
would
boot normally. Feels strange, but so far I was only able to
reproduce
the
problem with a Toshiba MK8052GSX. On the topic of my patch, I still
don't
see why a device which fails so terribly that it reports 3
exceptions
shouldn't be disabled. Like in this case, it could cause infinite
loops.

The problem is that this could happen in some cases when you
wouldn't
want to disable the device, like an error that just happens
sporadically and works on retry, or a device you're trying to
recover
data from.

What do you think if I edit the patch in a way, that when an
operation
successfully completes, it resets exce_cnt to zero. Might as well add
a
module_param, which can set the maximum value of exce_cnt, while
having
zero
as an option to never disable the device. Please don't think me
wrong,
I
don't want to force this patch, I just want to learn how all this
works,
and
in the process try to make it better. :-)

That would be better, but I think you're still going to have an issue
with what magic number to pick to avoid disabling devices
inappropriately.

Conceptually, disabling the device doesn't really make sense anyway.
If someone in userspace wants to keep trying to read from that device,
why would you stop them because of some arbitrary judgement? The
kernel itself isn't "locked up" during this process, anything not
blocked on I/O to that device should be able to continue running, so
that process is only hurting itself. If the system fails to boot from
another device due to this, this would likely point out some kind of
problem in userspace or the distro boot process being overly
serialized.

I have been booting up with the initramfs from ubuntu 13.04,
and I have also tried to boot with the ubuntu install cd. They couldn't
continue the boot process. I'm gonna spend the weekend trying to figure
out where and why the interrupts don't happen. Whether it be a routing
or a hardware issue, which I highly doubt due to the fact that Windows
XP SP2 was able to boot up without errors.

Are you able to get out full dmesg output from a boot attempt and the
contents of /proc/interrupts?

As I said before, I am not able to get to the shell, without my 'symptom
cure'. With my patch I get the following dmesg output, with
some of my debug messages turned off:
http://pastebin.com/5eb5G3Dx
/proc/interrupts is here:
http://pastebin.com/84CJey2D
After yesterday's research, I have come to ata_piix.c . That file looks
like
the real culprit, as my netbook's controller is an Intel ICH7M one,
The values I am getting from the device are very different than those
that are expected.

Things I have noticed, but ignored in dmesg:
There is a stack dump, because nobody cared about IRQ#20. I have ignored
this because it is the EHCI IRQ, and I suppose it has nothing to do with
ata. The problem is with ata3 or /dev/sdc, while the IRQ happens
with /dev/sda, which works fine.

I think it is likely related to the problem. The kernel thinks this
controller is on IRQ 16, but apparently something is raising
un-acknowledged interrupts on IRQ 20 and nothing is coming in on IRQ
16. It seems quite likely that this is actually the ATA controller.

You mentioned that Windows XP was able to work in this mode. I wonder
if it was using the IOAPIC, as if not then the IRQ routing is
different which might mask the problem. Do you know what IRQ Device
Manager reported for this controller in Windows? And was it using any
IRQs over 15 (which would indicate the IOAPIC was in use)?

Hmm, according to WinXP's Device manager for this controller,
it listens to IRQ# 20, and therefore it is using the I/O APIC.
Now, one question remains where is the error that mismaps
controller?
I have created a simple patch which seems to fix this:
---

@@ -1704,6 +1767,8 @@ static int piix_init_one(struct pci_dev *pdev, const

struct pci_device_id *ent)
                hpriv->map = piix_init_sata_map(pdev, port_info,

piix_map_db_table[ent->driver_data]);

+       if(pdev->vendor == 0x8086 && pdev->device == 0x27C4)
+               pdev->irq = 20;
        rc = ata_pci_bmdma_prepare_host(pdev, ppi, &host);
        if (rc)
                return rc;

However, I am more than sure that this is not the way
to solve this problem. Do you have any idea on where
the ideal place would be to implement a fix?
According to specs of ICH7M, which is essentially the
same as ICH6M, we need to check on what interrupt pin
is the SATA controller, and after that check which IRQ line
is connected to the I/O APIC and decide the IRQ's number
on those findings.

Specs of ICH7:
http://www.intel.com/content/dam/doc/datasheet/i-o-controller-hub-7-datasheet.pdf
Device 31 Interrupt Route Register: Chapter 7.1.46
Device 31 Interrupt Pin Register: Chapter 7.1.41

The SATA controller is always Device 31.