With iSCSI root on x86_64 DomU I was seeing an oops in scsi_calculate_bounce_limit() where PCI_DMA_BUS_IS_PHYS() was called. Issue is that dma_ops is NULL. This changeset fixes it: http://xenbits.xensource.com/ext/linux-2.6.tip-xen.hg?cs=e5a7f30e1db3
Background is unless no_iommu_init() is called, dma_ops on x86_64 is NULL causing the scsi subsystem to oops when trying to use iSCSI root. The patch basically fixes a merge error. While porting the xen stuff to 2.6.28, the iommu init stuff from x86/kernel/pci-dma.c was copied to i386/kernel/pci-dma-xen.c and the pci_iommu_init() function appears to have been dropped. Patch was worked out between cdub and myself and just brings us back in line with upstream kernel.org. I'm running both my x86_64 dom0 and domU with it. There doesn't appear to be any regressions. What I'm not clear on is why the problem didn't result in more oops than just with iSCSI.
*** Bug 215463 has been marked as a duplicate of this bug. ***
Created attachment 141168 [details] iscsi-x86_64-no_iommu_init.patch
I am having the same problem with kernel 2.6.19-1.2895.fc6xen with iscsi-initiator-utils-6.2.0.747-0.0.fc6 on a FC6 XenU guest.
I updated my kernel to 2.6.19-1.2911.fc6xen using yum. I also had to run "ethtool -K eth0 tx off" in my Xen guest to get rid of some other TCP/IP crc problems I was having in combination with my nating Xen0. After these two changes I am unable to reproduce the XenU kernel iscsi oopses I had before. (by the way: I did not explicitly use the above mentioned patch, but maybe it has been integrated in the newer kernel I am using now) So.. The problem has been solved for me!
We still need this patch in rawhide
What's the ETA for getting this in? We're in final freeze stages and trying to get everything locked and ready to go for Fedora 7...
This doesn't look to be in AFAICT. If this is going to get in for F7, it needs to get built by about noon Eastern tomorrow as we're hoping to get an RC going tomorrow.
F-7 build task submitted: http://koji.fedoraproject.org/koji/taskinfo?taskID=14007
Fix was committed to 'devel' branch on CVS, also.