From: Aron Griffis <agriffis@redhat.com>
Date: Wed, 17 Oct 2007 13:58:39 -0400
Subject: [scsi] cciss: disable refetch on P600
Message-id: 20071017175839.GA9269@redhat.com
O-Subject: [RHEL5.2 PATCH] BZ 251563 fix cciss mca (2nd try)
Bugzilla: 251563

https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=251563

Description
-----------
The P600 Smart Array adapter sometimes DMA prefetches too far.  This
is a bug in the adapter which can cause an MCA on systems with an
iommu, for example HP ia64 platforms.

This bug rarely shows on bare metal Linux because the driver allocates
physically contiguous regions for DMA and the iommu isn't involved.
However under Xen, dom0's pseudo-physical allocations aren't machine
contiguous, so the iommu is almost always used.  On bare metal, we've
observed the MCA in the rare condition that the overrun fetches
a non-populated physical address.

The workaround is to disable "refetch" on the adapter.  Refetch refers
to retrying a failed prefetch, as I understand it, and disabling
refetch prevents the bad read.

Test Status
-----------
At HP we've been pounding on this for a couple months.  I have a test
case that exposes the bug very quickly (forcing a PV domain to swap
seems to be a reliable repeater).  With this driver change, my test
case survives weekend-long runs where previously the machine would
crash within minutes.

Upstream Status
---------------
Presently in Jens Axboe's for-linux branch:
http://git.kernel.org/?p=linux/kernel/git/axboe/linux-2.6-block.git;a=commit;h=8bf50f71cbfc7d043f0f135da72b3feefeaa0eb8

Proposed Patch
--------------
Please review and ACK for 5.2.  (If only I could retroactively get
this into 5.1!)

----------------------------------------------------------------------

This patch disables DMA refetch in the PCI bridge. We have disabled DMA
prefetch for quite some time. Testing with XEN revealed another ASIC bug. If
dom0 resides on a P600 the board can can an MCA bi accessing invalid memory
addresses. Apparently, we need to disable both prefetch and refetch.
My understanding is a refetch operation should not occur but it is a valid
thing to do if prefetched data is no longer available for whatever reason.
Please consider this patch for inclusion.

Signed-off-by: Mike Miller <mike.miller@hp.com>
Signed-off-by: Alex Chiang <achiang@hp.com>

--------------------------------------------------------------------------------

Acked-by: Prarit Bhargava <prarit@redhat.com>
Acked-by: Jarod Wilson <jwilson@redhat.com>
Acked-by: Don Dutile <ddutile@redhat.com>
Acked-by: Pete Zaitcev <zaitcev@redhat.com>

diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
index b51ba7c..785078f 100644
--- a/drivers/block/cciss.c
+++ b/drivers/block/cciss.c
@@ -2968,15 +2968,20 @@ static int cciss_pci_init(ctlr_info_t *c, struct pci_dev *pdev)
 	}
 #endif
 
-	/* Disabling DMA prefetch for the P600
-	 * An ASIC bug may result in a prefetch beyond
-	 * physical memory.
+	/* Disabling DMA prefetch and refetch for the P600.
+	 * An ASIC bug may result in accesses to invalid memory addresses.
+	 * We've disabled prefetch for some time now. Testing with XEN
+	 * kernels revealed a bug in the refetch if dom0 resides on a P600.
 	 */
 	if(board_id == 0x3225103C) {
 		__u32 dma_prefetch;
+		__u32 dma_refetch;
 		dma_prefetch = readl(c->vaddr + I2O_DMA1_CFG);
 		dma_prefetch |= 0x8000;
 		writel(dma_prefetch, c->vaddr + I2O_DMA1_CFG);
+		pci_read_config_dword(pdev, PCI_COMMAND_PARITY, &dma_refetch);
+		dma_refetch |= 0x1;
+		pci_write_config_dword(pdev, PCI_COMMAND_PARITY, dma_refetch);
 	}
 
 #ifdef CCISS_DEBUG