Comment 30 for bug 1853306

------- Comment From <email address hidden> 2023-07-20 13:50 EDT-------
Verification of this kernel feature is complete. All tests were run using the kernel in jammy-proposed (5.15.0-79-generic #86-Ubuntu) on the host as well as in the KVM guest (except where otherwise noted, see below).
Because the necessary qemu changes are not yet in jammy-proposed, I used a custom qemu build based on the launchpad qemu repository + applied the QEMU feature backport. I will do a second verification of the QEMU feature once that package is available.

As for the testing itself: NVMe and RoCE testing went just as expected.

During ISM testing, I did run into one issue that is not caused by this feature (therefore I am marking verification complete) but it is something that will need to be fixed separately once an upstream patch is available. Specifically, when using the jammy-proposed kernel in the KVM guest I noted occasional UBSAN array-index-out-of-bounds warnings from net/smc/af_smc.c -- This is not directly due to the code added by this feature or even due to being run in a KVM guest, but rather a bug in net/smc code (specifically the SMC_STAT_PAYLOAD_SUB macro) that is much more obvious when ubsan is enabled. I've already identified the root cause, verified that it also exists upstream, and reported the issue to the SMC maintainers along with a suggested fix - without UBSAN enabled no warning is issued but an incorrect performance status counter is quietly incremented. I had panic_on_warn enabled in the KVM guest, therefore the warning caused the guest to crash; otherwise I likely would not have noticed this issue at all.

In order to complete all verification tests related to ISM devices, I proceeded with 2 different KVM guest kernel configurations and ran all ISM tests using both:
1) a custom built (upstream 6.4.0) guest kernel that did not include UBSAN
2) the jammy-proposed kernel that includes UBSAN but with panic_on_warn=0

All test scenarios passed with these guest kernels.