bhyve: check vm_map_gpa and paddr_guest2host returns in storage emulation

[ojanerif_amd.com]

Summary

pci_nvme: fail the command with NVME_SC_INTERNAL_DEVICE_ERROR when
vm_map_gpa() returns NULL in create_io_sq, create_io_cq and identify.

pci_ahci: check paddr_guest2host() in ahci_build_iov(), read_prdt() and
write_prdt(), and fail the command instead of submitting it.

[chuck]

Thank you for finding and proposing fixes for these issues!
In the future, submitting the changes with more context (e.g., git format-patch -U999999 --stdout other-branch > change.diff) helps speed things along.

usr.sbin/bhyve/pci_nvme.c

1282

This should be
pci_nvme_status_genc(&compl->status, NVME_SC_INTERNAL_DEVICE_ERROR);

here and below.

[ojanerif_amd.com]

Thank you for the review and the status code correction.

Updated all five sites to use with , including the create_io_sq and create_io_cq paths. Re-uploaded with full context.

[ojanerif_amd.com]

Correction to my previous comment: all five sites now use pci_nvme_status_genc() with NVME_SC_INTERNAL_DEVICE_ERROR.

[chuck]

usr.sbin/bhyve/pci_ahci.c

638

The NULL pointer check is obviously a good thing, but have you checked that just returning here is OK?

[ojanerif_amd.com]

No, it wasn't: aior->more stays set, so ahci_ioreq_cb() resubmits the same request forever. ahci_build_iov() now returns an error and the callers fail the command as if blockif had failed.

[chuck]

With the small change to read_prdt, this looks good to me. Thanks again!

usr.sbin/bhyve/pci_ahci.c

832

For consistency, I'd vote for changing the return to a break.

[ojanerif_amd.com]

Done — read_prdt now breaks out of the loop like write_prdt does, falling through to the final return (size - len). Also regenerated with full context (-U999999).