AlphaServer SC patch kit:
==========================

AlphaServer SC Version : SC V2.6 UK1
Kit Name:                SCV26UK13209964871
Release Date:            07/04/2005
QuIX:                    QXCM1000220347/QXCM1000225914
WFM Number:              1206623459/3209964871
Abstract: scfs: Use REMOTEREAD to avoid elan driver prefetch problem


Description of Patch:
=====================
The elan driver has a built-in prefetch algorithm that tries to predict what pages a remote
process will read next and speculatively prepares these pages for transfer. This prefetcher
has problems with DMAs that go right up to a 256k elan page boundary if the next page has
the permission "local_read" (which is the most secure and means "can only be read by local
processes"). The prefetcher can attempt to read ahead into a protected page where it will
get an "access denied" error. This error causes the prefetcher to place an entry in the
DMA retry queue that is very difficult for the ep_dma_retry thread to retransmit if the
destination node is very busy - there is a race condition that the sending node is likely
to lose over and over again for an extended period (maybe hours).

SCFS uses local_read pages for all data that it will only read locally. This is the correct
thing for SCFS to do, but the prefetcher will have problems if it happens to read ahead into
one of these pages. Quadrics believe that this is exactly what is happening in the
"scfs_buf_wait" stalls being experienced by several large customers. The larger and busier
the system the more likely this is to happen and the longer it will last.

The most elegant solution would be to rewrite this part of the elan prefetcher, however this
is a very complex task. Instead we are opting for a simpler and more reliable solution - we
can avoid the problem completely by having SCFS use a lower protection level for its pages
(remote_read) that allows the prefetcher to read them. That's exactly what this patch does.

Note: In order for this workaround to operate reliably, we must be sure that all DMAs are
correctly aligned on 256k boundaries. This is why the updated ep3.mod elan driver
SCV26UK11206623459C is a pre-requisite for this patch.


Kit location:
=============
The patch kit is SCV26UK13209964871.tar.gz and it is available from ITRC.


Prerequisites:
==============
Before installing this Patch kit, you should ensure the following:

1) You have all mandatory patches for this release installed

2) You have SCV26UK11206623459C installed. You can find this on ITRC 


Kit checksum:
=============
    # cksum SCV26UK13209964871.tar.gz
    3984345051 702080 SCV26UK13209964871.tar.gz

Updated files:
==============
A list of the files included in this patch is given below
along with the cksum values for each file:

    751783857   559290  /usr/opt/SCFS/sys/scfs.mod
    1554028406  661224  /usr/opt/SCFS/sys/scfs_client.mod


Instructions:
=============
This patch is provided as a setld installable kit.  Unpack it
into a directory that is NFS mounted on all domains e.g.
/nfs/SCV26UK13209964871 and install it as follows:

Patch required on Management Server (if used) : NO
Patch required on Domains                     : YES


On Domains:
-----------
    Install the patch as follows:

    # cd /nfs/SCV26UK13209964871
    # scrun -d all setld -l /nfs/SCV26UK13209964871 SCV26UK13209964871
    # 
    # scrun -d all BuildKernels
    # scrun -d all DeployKernels
    # sra shutdown -domains all
    # sra boot -domains all

Installation is complete at this point.


    To remove the patch use the following commands:

    # scrun -d all setld -d SCV26UK13209964871
    # 
    # scrun -d all BuildKernels
    # scrun -d all DeployKernels
    # sra shutdown -domains all
    # sra boot -domains all



Once the patch is installed and you are satisfied that it is working correctly
the GENERIC kernels should also be updated, using the following command:

    # scrun -d all DeployKernels -g
=======================================================================