Avoid some out-of-range DMA accesses with PCI intel ethernet adapters.

[kib]

Summary

This change sit in my local tree for very long time. It caused a conflict with recent netmap update, so I think it could be useful for me to get rid of it, either commit or revert locally.

The issue is that pre-PCIe, i.e. plain parallel PCI adapters from Intel, managed by the lem(4) drivers, apparently have very serious hardware bugs in the DMA engines. Both descriptor ring accesses, and packet buffer accesses, read past the end of the regions. I have a haswell machine with PCIe/PCI bridge, and 82541PI two-port card. If VT-d is enabled, card generates a fault for almost any access.

The patch in the review extends the descriptor rings by one element each, which fixes overruns there and rather cleanly demonstrates the bug. Unfortunately, I was not able to develop any form of sane fix for the packet memory accesses. So the patch is only partial.

Test Plan

Intel did not responded to the inquiries about the problem. This issue prevented me from having a working environment to integrate ACPI_DMAR and vmm.ko.

[sbruno]

Please commit this at your convenience.

[emaste]

sys/dev/e1000/if_lem.c

546–556

There should be a comment explaining the + 1 I think, it does not have to be overly detailed but should at least provide a "bread crumb" to suggest this is intentional.

[erj]

I agree with @emaste's request for a comment.

[kib]

Add comment.