Date:	Fri, 28 May 2010 19:05:05 -0500
From:	"Moffett, Kyle D" <Kyle.D.Moffett@...ing.com>
To:	Robert Hancock <hancockrwd@...il.com>,
	Torsten Kaiser <just.for.lkml@...glemail.com>
CC:	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Vivek Mahajan <vivek.mahajan@...escale.com>,
	Jeff Garzik <jgarzik@...ox.com>,
	"linux-ide@...r.kernel.org" <linux-ide@...r.kernel.org>,
	Kyle Moffett <kyle@...fetthome.net>,
	"Livingston, John A" <john.a.livingston@...ing.com>
Subject: Re: New MSI support in sata_sil24 still broken in 2.6.33-rc3

My advance apologies if this email gets badly MIME-mangled...

On 2010/01/06 20:59, "Robert Hancock" <hancockrwd@...il.com> wrote:
> On 01/06/2010 03:37 AM, Torsten Kaiser wrote:
>> After activating the MSI support by adding sata_sil24.msi=1 to the
>> kernel command line, the first write to a drive attached to the SiI
>> 3132 controller results in the following errors:
>> 
>> [  138.950074] ata2.00: exception Emask 0x0 SAct 0xf SErr 0x0 action 0x6
>> frozen
>> [  138.961023] ata2.00: failed command: WRITE FPDMA QUEUED
>> [  138.970034] ata2.00: cmd 61/00:00:a5:95:4a/04:00:01:00:00/40 tag 0
>> ncq 524288 out
>> [  138.970037]          res 40/00:00:00:00:00/00:00:00:00:00/00 Emask
>> 0x4 (timeout)
> 
> Looking at the code in sata_sil24 and the SiI3132 datasheet, there's a
> control bit which doesn't seem to be handled in the driver, global
> control register bit 30: "MSI Acknowledge (W). Writing a one to this bit
> acknowledges a Message Signaled Interrupt and permits generation of
> another MSI. This bit is cleared immediately after the acknowledgement
> is recognized by the control logic, hence the bit will always be read as
> a zero. If all interrupt conditions are removed subsequent to an MSI, it
> is not necessary to assert this Acknowledge; another MSI will be
> generated when an interrupt condition occurs."
> 
> The way the interrupt handler for this driver works is that we check the
> global IRQ status register, and then based on what ports indicated an
> interrupt in that register, we check the individual port command
> completion registers. The issue would seem to be that if a port got an
> interrupt condition in between these two operations, we'd miss it, and
> the MSI logic described above then wouldn't generate any more interrupts
> since we didn't remove all interrupt conditions.
> 
> Can you try this patch and see if it helps? (Might be whitespace damaged
> but hopefully you can apply manually in that case.)

I've got this custom board that uses the sata_sil24 driver (off a P2020
processor).  My current kernel is a slightly patched 2.6.32 kernel
(including the sata_sil24 enable-MSI patch).

Unfortunately when I turn MSI on, I get the exact same hang described here,
boot log included as dmesg1.txt.

With this patch applied, it seems to get a little further (dmesg2.txt), but
still dies miserably.

I'm relatively sure that MSI works on this chipset as I also have an e1000e
controller off an adjacent PCI-E bus which works correctly with MSI.

It's relatively critical for me to get MSI working, because the legacy-PCI
INTx interrupt for that PCI-E port happens to share an IRQ line with a
device that is very unfriendly to shared IRQs (it has no internal IRQ
disable register).  I'd rather not have to go in there with a soldering iron
and some scraps of wire to make it work. :-D

Cheers,
Kyle Moffett


Download attachment "dmesg1.txt" of type "application/octet-stream" (84028 bytes)

Download attachment "dmesg2.txt" of type "application/octet-stream" (143621 bytes)

Powered by blists - more mailing lists