Re: TR: SiI2112 + Seagate + nFroce2: no DMA!

From: Emmanuel Hislen
Date: Sun Jan 25 2004 - 15:55:11 EST



Hi again,

Here's a more complete report on this issue.

-1-

Going to Fedora Core 12.4.22-1.2115.nptl fixed the DMA issue. The Seagate SATA drive (ST3160023AS) came up in DMA mode.
The speed however (hdparm -t) was 25MB/s, better but still unacceptable.

-2-

I got the latest Fedora Core: 2.4.22-1.2149.nptl

Interestingly enough I found out that this release comes with a sata_sil module that is not even in 2.4.24, it was probably ported from 2.6.1 by the Fedora team.
Anyway I believe this is just a RAID driver, it did not change anything (same as -1-).

-3-

Now I decided to live dangerously and I tried tuning max_kb_per_request, the default being 15:

/echo "max_kb_per_request:128" > /proc/ide/hde/settings/

On my first hdparm -t I got 55MB/s !!!


BUT the drive was still making a lot of noise after I got the results, and every time I tried the command again (only waiting a couple seconds before retrying after I got the results), the throughput went lower each time. In 5 retries I was at 7 MB/s. The drive was very noisy, still busy minutes after I was done torturing him.

Unfortunately this did not come as a surprise as I had found the following thread on Google:

Andre Hedrick wrote:

/>/
/>Seagate and Silicon Image are the only two player (well intel now) who did/
/>their own PHY. They did not use the Marvel pairs./
/>/
/>It is a function of possible ECC on the wire and the relation to the/
/>segments in the PIO or SG operations. It is a FIFO issue based on 512byte/
/>boundaries being breached on corner cases./
/>/
/>The data on the wire is in 8K units./
/>/
/>It is a 7.5K + 0.5K corner case./
/>/
/>max_kb_per_request:15 == 7.5K/
/>/
/>This prevents this corner case until I can code the proper special case SG/
/>table./
/>/
/>drive->id->hwconfig |= 0x6000;/
/>/
/>Is needed to fake the driver for device side cable detect./
/>There are several issues and I have not had time to keep up./


-3-

The same machine is now dual boot with Windows 2000 SP4. I installed Sandra 2004, a software for benchmnarking, and I got the following results regarding the SATA drive:

Buffered Read: 83 MB/s
Sequential Read: 48 MB/s
Random Read: 8 MB/s
Buffered write: 68 MB/s
Sequential write: 48 MB/s
Random write: 9 MB/s

I know, comparing results obtained with 2 different benchmarking products is almost irrelevant, but it is enough for the big picture:

There is nothing wrong with the harsware (Seagate or Silicon Image), it works just fine with the windows drivers/libraries ... Oops did I just praise microsoft?


-4-

I compiled/installed a new linux kernel to replace the one from the Fedora distrib with 2.4.24

Absolutely no difference, same as -1-: 25 MB/s


-5-

Now I tried kernel 2.6.1: it is worse, it dropped to 14 MB/s !?!?!
I've seem several similar reports about 2.6.


-6-

So at this point I am ready to give up, and to simply buy a PATA drive.

However, after doing all the above (which is reported in chronological order), I booted the latest Fedora Core: 2.4.22-1.2149.nptl again, and decided to give another try to:

echo "max_kb_per_request:128" > /proc/ide/hde/settings


Now I have consistent 55 MB/s and no horrible noise from the drive, and I let this run for 4 hours without any issue. Even typing the "hdparm -Tt" commands one after another results are consistent.

What the heck is going on here (not that I don't appreciate things working, but I'd like to understand)??? I can see only the following possibilities:

- I was lucky, and it all depends on when you change "max_kb_per_request" due to some race condition.
- going to 2.6.1, I upgraded 2 packages:
quota-tools: went from 3.06 to 3.10
procps: went from 2.0.17 to 3.1.15
I really don't see how this would change anything, but who knows...
- running the Win2000 benchmark with Sandra somehow did something to the drive and/or the SATA controller, and they are now friends. Fishy...


Does anybody knows where the default 15kb is coming from for this drive?
How can I change this permanently? I mean other than than a hack rc script running at init time...

I understand 2.6.1 is not using /proc/ide/hde/settings, how do I change max_kb_per_request with this kernel?

Anyway I will make several more attempts with max_kb_per_request set to 128, to make sure it is really OK before enforcing it with a script (anyway this script would have to be kernel-version specific, to keep a back door).


Hope I haven't been too long and boring. Comments/suggestions are welcome.


Thanks for your help!


Emmanuel.


Emmanuel Hislen wrote:


-----Message d'origine-----
De : Emmanuel Hislen [mailto:hislen@xxxxxxxxxxxxxx]
Envoye : Saturday, January 24, 2004 1:27 AM
A : Hugo Mills
Cc : linux-kernel@xxxxxxxxxxxxxxx
Objet : RE: SiI2112 + Seagate + nFroce2: no DMA!



Hi Hugo,

I have re-installed my machine with Fedora Core 1 (thinking a newer version
than RH9 would make the jump to 2.4.24 or 2.6 easier), and this fixed the
DMA issue :-)

Now my SATA drive is running stable in UDMA6.

However, performance is still way below expectations.

I got a huge improvement: my disc read speed (hdparm -t) went from 1.3 to 25
MB/sec.
This is still slower than my PATA drive on my 3 years old AMD 900 PC running
Redhat 9.0 (around 35 MB/sec).

Could you please let me know what you are getting so I know what to expect?


My next step is to try 2.4.24 or 2.6.1 as you suggested, but googling around
a little bit before I do so I found out a few worrying things:

- people have reported a drop in performance on SATA in some 2.6 based
kernel (2.6.0-test9), with reported speeds around 20MB/sec. Apparently there
is no more way to tune max_kb_per_req in 2.6??
- I have found reports that both ide and libdata libraries are limiting
max_kb_per_req to 15 Kb specifically for Seagate drives. So It looks like I
can't even set it to 128 (I did not even try as I saw reports of memory
corruption).

So basically since you've got it to work I'd like to know:

* what speed you get, and what is the RPM of your Seagate
* what is your max_kb_per_req setting (I have 15K)
* what is your accoustic management setting (I have 0)


I could not fix the time on linux so I am sending this mail from WinXP (just
kidding I lost my mails on the linux machine after re-installing :-).


Thanks,


Emmanuel.



-----Message d'origine-----
De : Hugo Ranger Mills [mailto:hrm@xxxxxxxxxxxxx]De la part de Hugo
Mills
Envoye : Tuesday, January 20, 2004 12:58 AM
A : manu
Cc : linux-kernel@xxxxxxxxxxxxxxx
Objet : Re: SiI2112 + Seagate + nFroce2: no DMA!


On Tue, Dec 31, 2002 at 09:55:59PM -0800, manu wrote:

Incidentally, did you know that the date on your computer is very,
very wrong?

I'm about to give up on my SATA drive as I can't get it to work properly.
So I thought I may try asking the experts before falling back to PATA.

I have seen many mails reporting the same issue, some of them 6-month old:

- SATA drive comes up in pio mode, not in dma
- trying to turn on dma with hdparm is a nightmare: I/O errors, crash
with data corruption... I tried both:

hddarm -d1 /dev/hde

and:

hdparm -u1 -c3 -d1 -X66 /dev/hde

crash in both cases :-((


Here's my equipment:


ABIT AN7 motherboard (nForce2 chipset, SiI3112 SATA controller)
AMD Athlon XP 2600+ (+ 512 DDR / 400 MHz)
SATA HD Seagate Barracuda 160 Gb

The SATA HD is my only drive. The only thing connected to my IDE
controllers is a DVD/CD combo.

Running Linux Redhat 9.0
kernel 2.4.20-28.9

^^^^^^^^^^^^^^^^^^
This is your problem. There have been a number of bug-fixes to the
SiI drivers since 2.4.20. Try it again with a newer kernel -- such as
2.4.24.

I've been googling for days now and could not come accross a solution,
on the contrary I came under the impression that the combination of
SiI3112 +and Seagate was doomed.


Not so. I have a SiI3112 controller and a 120GiB Seagate drive, and
they work very well together. I'm using 2.6.1, although 2.4.23 also
worked well for me.

[snip]

Isn't there a solution??

I am willing to try patches of experimental code. At this point I am
looking at reinstalling everything on a PATA drive anyway, so I have
nothing to loose.


Try using 2.4.24 or 2.6.1.

Hugo.

--
=== Hugo Mills: hugo@... carfax.org.uk | darksatanic.net | lug.org.uk ===
PGP key: 1C335860 from wwwkeys.eu.pgp.net or http://www.carfax.org.uk
--- All hope abandon, Ye who press Enter here. ---




-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/