Re: FW: ICH10R PMP support

Re: FW: ICH10R PMP support

[Mark Lord]

Mike P Moore wrote:
> Hi Mark,
> 
>   Sorry to hijack this thread a little, but I was wondering how is the 
> support for the 88SX6081?
> Does it have hot-plug, PMP w/FIS switching, etc. yet?
..

The 6081 are PCI(-X) only chips, and so cannot be used with PCIe.
These older chips had many issues with various things, and most of
the errata workarounds required are now in Linux.
A few more are coming for 2.6.27, though.

The "C0" revision (PCI "Rev.9") is the best of the lot;
I wouldn't really use the older ones if I could avoid them,
as they do not support PMP hotplug.

No FIS-based switching on any 6081, but they all do support PMPs.

There is a newer 6042 chip for PCI(-X) slots,
which is more or less identical to the 7042 internally
and featurewise on Linux.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

FW: ICH10R PMP support

[Milo Medin]

Mark, Tejun suggested I query you about your thoughts re: the Marvell 88SX7042 or the 88SE6145 SATA controllers when connected with FIS based port multipliers? 

Based on Intel's announcements last year for the ICH10R where they claimed they were going to implement FIS support for port multipliers, I had planned on rebuilding a Linux based NAS system using a P45 board and ICH10R, but it seems they didn't end up implementing FIS support in the ICH10 series after all (nice of them to neglect to mention that to the community after announcing that feature in the roadmap).  

So now I'm looking for a different solution.  I currently use SI3132 based SATA controllers, but they seem to have some performance issues, and folks seem to be reporting good performance with at least the 88SX7042 part under windows, hence the question.

Thanks in advance,
Milo


-----Original Message-----
From: Tejun Heo [mailto:tj@kernel.org] 
Sent: Sunday, June 22, 2008 4:51 PM
To: Milo Medin
Subject: Re: ICH10R PMP support

Milo Medin wrote:
> Tejun, what's your opinion of the Marvell 88SX7042 and 88SE6145
> controllers for feeding PMP's?  Do they work OK with FIS based PMP's
> in your experience?  Are they faster than the SI 3132's?

Dunno much about that chip.  You should ask Mark Lord <mlord@pobox.com>.
 Also please cc linux-ide@vger.kernel.org.

-- 
tejun

Re: FW: ICH10R PMP support

[Mark Lord]

Milo Medin wrote:
> Mark, Tejun suggested I query you about your thoughts re: the Marvell 88SX7042 or the 88SE6145 SATA controllers when connected with FIS based port multipliers? 
> 
> Based on Intel's announcements last year for the ICH10R where they claimed they were going to implement FIS support for port multipliers, I had planned on rebuilding a Linux based NAS system using a P45 board and ICH10R, but it seems they didn't end up implementing FIS support in the ICH10 series after all (nice of them to neglect to mention that to the community after announcing that feature in the roadmap).  
> 
> So now I'm looking for a different solution.  I currently use SI3132 based SATA controllers, but they seem to have some performance issues, and folks seem to be reporting good performance with at least the 88SX7042 part under windows, hence the question.
..

I don't know anything about the 88SE6145 (isn't that an AHCI clone?),
but the 7042 chip under Linux is *now* a very safe (and good) bet.

It has full kernel support for hot-plug, PMP with FIS-based switching,
and employs a decent host command queue mechansim.  I believe all of the
errata workarounds are now in place for this chip.

Performance is about as good as it gets under Linux,
though there's still another microsecond or so to be shaved
away from the IRQ handling in the next cycle.

Message-Switched-Interrupts (MSI) is there, but not fully supported yet.
This also will likely get cleaned up in the next cycle.

The 7042 is the best of the Marvell chips, for Linux.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

Re: FW: ICH10R PMP support

[Tejun Heo]

Mark Lord wrote:
> I don't know anything about the 88SE6145 (isn't that an AHCI clone?),
> but the 7042 chip under Linux is *now* a very safe (and good) bet.
> 
> It has full kernel support for hot-plug, PMP with FIS-based switching,
> and employs a decent host command queue mechansim.  I believe all of the
> errata workarounds are now in place for this chip.
> 
> Performance is about as good as it gets under Linux,
> though there's still another microsecond or so to be shaved
> away from the IRQ handling in the next cycle.
> 
> Message-Switched-Interrupts (MSI) is there, but not fully supported yet.
> This also will likely get cleaned up in the next cycle.
> 
> The 7042 is the best of the Marvell chips, for Linux.

How high does it go bandwidth-wise?

Thanks.

-- 
tejun

Re: FW: ICH10R PMP support

[Mark Lord]

Tejun Heo wrote:
> Mark Lord wrote:
..
>> The 7042 is the best of the Marvell chips, for Linux.
> 
> How high does it go bandwidth-wise?
..

The limiting factor will be the bus width of the card
and the slot into which it is plugged, I suspect.

There are 4-lane PCIe cards out there, and one of those
in a 4-lane ("4X") or wider PCIe slot ought to have no
real limitations due to the 7042 side of things.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

Re: FW: ICH10R PMP support

[Tejun Heo]

Mark Lord wrote:
> Tejun Heo wrote:
>> Mark Lord wrote:
> ..
>>> The 7042 is the best of the Marvell chips, for Linux.
>>
>> How high does it go bandwidth-wise?
> ..
> 
> The limiting factor will be the bus width of the card
> and the slot into which it is plugged, I suspect.
> 
> There are 4-lane PCIe cards out there, and one of those
> in a 4-lane ("4X") or wider PCIe slot ought to have no
> real limitations due to the 7042 side of things.

Cool.  I was asking because sil3124 and 32 seem to have inherent
bandwidth limit regardless of the bus it lives on.  I'm planning on
checking out 3132-2 chips to see whether they improved on that front.

Thanks.

-- 
tejun

Re: FW: ICH10R PMP support

[Mark Lord]

Tejun Heo wrote:
> Mark Lord wrote:
>> Tejun Heo wrote:
>>> Mark Lord wrote:
>> ..
>>>> The 7042 is the best of the Marvell chips, for Linux.
>>> How high does it go bandwidth-wise?
>> ..
>>
>> The limiting factor will be the bus width of the card
>> and the slot into which it is plugged, I suspect.
..

Note that this has yet to be confirmed in practice,
though I don't know if anyone has really tried hard yet.

Perhaps Grant Grundler ? 

..
> Cool.  I was asking because sil3124 and 32 seem to have inherent
> bandwidth limit regardless of the bus it lives on.  I'm planning on
> checking out 3132-2 chips to see whether they improved on that front.
..

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

Re: FW: ICH10R PMP support

[Brad Campbell]

Mark Lord wrote:
> Tejun Heo wrote:
>> Mark Lord wrote:
>>> Tejun Heo wrote:
>>>> Mark Lord wrote:
>>> ..
>>>>> The 7042 is the best of the Marvell chips, for Linux.
>>>> How high does it go bandwidth-wise?
>>> ..
>>>
>>> The limiting factor will be the bus width of the card
>>> and the slot into which it is plugged, I suspect.
> ..
> 
> Note that this has yet to be confirmed in practice,
> though I don't know if anyone has really tried hard yet.

I recently read a very nice review of the 7042 on a Sonnet card in a Mac and they were pulling over 
700MB/s using SIL port multipliers in a 4 lane slot. They managed over 1GB/s with 20 drives across 2 
of the cards! I've heard some unkind things about the SIL cards bandwidth, but these Marvell jobbies 
look very nice.

http://www.amug.org/amug-web/html/amug/reviews/articles/sonnet/mac-pro/

I've half a mind to order a couple of these and some PM's to play with. Shame the only PM's I can 
seem to find that do what I want are $85 each from addonics (I need 5 of them)

Brad
-- 
"Human beings, who are almost unique in having the ability
to learn from the experience of others, are also remarkable
for their apparent disinclination to do so." -- Douglas Adams

RE: FW: ICH10R PMP support

[Milo Medin]

Mark, thanks very much for the reply.  Sounds like the 88SX7042 is a good solution.  A couple more questions:

1) There seems to be talk on the mailing list about a bug where on highpoint controllers using this chip the controller corrupts some data on non-raid disk members.  I assume this is caused by the specific firmware on highpoint controllers using this chip, and that if I use a non highpoint version of the 88SX7042 that I won't see this problem.

2) You said this is the best marvell solution for Linux at this time.  Is there a better SATA controller that I should be thinking about other than the 88SX7042?  I know the ICH9R's had good performance, but no FIS switching for PMP's.

Thanks again...

Milo


-----Original Message-----
From: Mark Lord [mailto:mlord@pobox.com] 
Sent: Monday, June 23, 2008 2:09 AM
To: Milo Medin
Cc: linux-ide@vger.kernel.org
Subject: Re: FW: ICH10R PMP support

Milo Medin wrote:
> Mark, Tejun suggested I query you about your thoughts re: the Marvell 88SX7042 or the 88SE6145 SATA controllers when connected with FIS based port multipliers? 
> 
> Based on Intel's announcements last year for the ICH10R where they claimed they were going to implement FIS support for port multipliers, I had planned on rebuilding a Linux based NAS system using a P45 board and ICH10R, but it seems they didn't end up implementing FIS support in the ICH10 series after all (nice of them to neglect to mention that to the community after announcing that feature in the roadmap).  
> 
> So now I'm looking for a different solution.  I currently use SI3132 based SATA controllers, but they seem to have some performance issues, and folks seem to be reporting good performance with at least the 88SX7042 part under windows, hence the question.
..

I don't know anything about the 88SE6145 (isn't that an AHCI clone?),
but the 7042 chip under Linux is *now* a very safe (and good) bet.

It has full kernel support for hot-plug, PMP with FIS-based switching,
and employs a decent host command queue mechansim.  I believe all of the
errata workarounds are now in place for this chip.

Performance is about as good as it gets under Linux,
though there's still another microsecond or so to be shaved
away from the IRQ handling in the next cycle.

Message-Switched-Interrupts (MSI) is there, but not fully supported yet.
This also will likely get cleaned up in the next cycle.

The 7042 is the best of the Marvell chips, for Linux.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

Re: FW: ICH10R PMP support

[Mark Lord]

Milo Medin wrote:
> Mark, thanks very much for the reply.  Sounds like the 88SX7042 is a good solution.  A couple more questions:
> 
> 1) There seems to be talk on the mailing list about a bug where on highpoint controllers using this chip the controller corrupts some data on non-raid disk members.  I assume this is caused by the specific firmware on highpoint controllers using this chip, and that if I use a non highpoint version of the 88SX7042 that I won't see this problem.
..

The firmware on the Highpoint RAID cards unconditionally overwrites
sectors on any attached drives, before booting any O/S.

Never plug a drive that already has data on it into any Highpoint RAID card.
But for new drives, just partition around the firmware areas and you'll be fine.
There are notes (and a WARNING) in the driver source code regarding those.
Avoid sectors 1..63 (cannot boot from this controller as a result),
and avoid the final (full!) GB of each drive.

> 2) You said this is the best marvell solution for Linux at this time.  Is there a better SATA controller that I should be thinking about other than the 88SX7042?  I know the ICH9R's had good performance, but no FIS switching for PMP's.
..

Dunno.  I think everyone in the biz is going to AHCI style controllers now.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com

RE: FW: ICH10R PMP support

[Milo Medin]

Thanks.  Just to confirm, non highpoint 88SX7042 based cards don't have this issue right?  This seems silly for them to do such a thing, but some people like their management tools...

Thanks
Milo


-----Original Message-----
From: Mark Lord [mailto:mlord@pobox.com] 
Sent: Tuesday, June 24, 2008 12:12 AM
To: Milo Medin
Cc: linux-ide@vger.kernel.org
Subject: Re: FW: ICH10R PMP support

Milo Medin wrote:
> Mark, thanks very much for the reply.  Sounds like the 88SX7042 is a good solution.  A couple more questions:
> 
> 1) There seems to be talk on the mailing list about a bug where on highpoint controllers using this chip the controller corrupts some data on non-raid disk members.  I assume this is caused by the specific firmware on highpoint controllers using this chip, and that if I use a non highpoint version of the 88SX7042 that I won't see this problem.
..

The firmware on the Highpoint RAID cards unconditionally overwrites
sectors on any attached drives, before booting any O/S.

Never plug a drive that already has data on it into any Highpoint RAID card.
But for new drives, just partition around the firmware areas and you'll be fine.
There are notes (and a WARNING) in the driver source code regarding those.
Avoid sectors 1..63 (cannot boot from this controller as a result),
and avoid the final (full!) GB of each drive.

> 2) You said this is the best marvell solution for Linux at this time.  Is there a better SATA controller that I should be thinking about other than the 88SX7042?  I know the ICH9R's had good performance, but no FIS switching for PMP's.
..

Dunno.  I think everyone in the biz is going to AHCI style controllers now.

Cheers
-- 
Mark Lord
Real-Time Remedies Inc.
mlord@pobox.com