Re: RFC: ATA to CAM integration patch

:James R. Van Artsdalen wrote:
:> Has anyone found any PCI-X or PCI cards that do AHCI?  Or a PCI-Express,
:> PCI-X or PCI card that supports AHCI & FIS-based switching for port
:> multipliers?
:> 
...
:
:I still haven't found any AHCI card with FBS support. Ideas welcome.
:
:I am working on driver for SiI3124/3132. I haven't tested yet how fast 
:they really are (3132 is quite cheap), but they declare PM with FBS 
:support. First one is 4-port PCI-X, second - 2-port PCIe.
:
:-- 
:Alexander Motin

    I haven't had much luck on the AHCI front either re: FBSS.  No cards,
    and very few motherboards have the FBSS capability.  FBSS is a very
    new feature for AHCI, so the expectation is that it will work its way
    into more chipsets over time.  It should become common in less then a
    year.  There is huge demand for the feature.

    Alexander and I have been discussing the Sili chip.  The linux folks
    found one very serious hardware bug and there are multiple other
    issues when working with targets behind a PM.  Fortunately the worst
    hardware bug does seem to have workarounds which do not screw up
    performance too badly.  The bug is related to reading the LRAM (that's
    right, just doing a simple memory read of a mapped LRAM location(!)(!))
    on the chip while any command is active produces corruption.

    It is possible to operate the Sili chip with NCQ+FBSS (The chip does
    the FBSS internally).  The chip seems to be able to handle upwards of
    30,000 transactions per second but the main performance limitation
    appears to be physical port bandwidth limitations.  Each physical
    port on the 3132 seems to be limited to 120-150 MBytes/sec (at least
    on the 3132), even if the phy probes at 3 Gbit/sec.  And, of course,
    the 3132 uses only one PCI-E lane and that's something like 3 GBits/sec.

    I successfully ran a test on a 10-Terrabyte PM array with 5 2TB disks over
    2 days of continuous reading and writing with no errors so the Sili chip
    does work once the issues are worked out, but with 5 disks going at once
    bandwidth was limited to 20-30 MB/sec on each one due to the above
    mentioned limitations.

    The SAS chipsets can probably do a lot better.  The advantage of the Sili
    chipsets is that they are ridiculously cheap.  So cheap that many of
    the PM enclosures sold on the market also come with a freebie PCI-e 3132
    card to stuff into your computer.

    There are other issues with the Sili chipset and PM's related to hot
    insertions and firmware bugs... it isn't a walk in the park, but it
    seems to be possible to work around them and get something reasonably
    robust out of the boards.  Also, error handling requires command
    exclusivity (due to the LRAM bug) so a failing target behind the PM
    can potentially cause problems for the entire enclosure.

						-Matt