bhyve: Treat BAR registration failures as non-fatal

[markj]

Summary

Some guest operating systems (at least 9front) will leave device I/O
space decoding enabled while probing BAR sizes. In this case, a
sufficiently large 32-bit BAR can overlap with the HPET MMIO region
fixed at 0xfed00000. This results in an assertion failure in
modify_bar_registration(), even though the collision is transient, i.e.,
the guest will promptly map the BAR elsewhere.

Tolerate such overlaps. Keep track of a BAR's registration state and
make sure that modify_bar_registration() doesn't try to unregister an
I/O space if registration failed.

This fixes a problem where a 32MB framebuffer ends up overlapping with
the HPET MMIO region, causing 9front guests to crash.

PR: 290098

[wanpengqian_gmail.com]

We hit the same assertion in a similar scenario: SeaBIOS sizing the 256 MB BAR of a passed-through GPU with memory decoding left on. The BAR landed at 0xf0000000, over the LAPIC/IOAPIC/HPET, and register_mem() failed.

We handled it the same way in our tree: remember that the registration failed, skip the unregistration later and do not call pe_baraddr for that BAR, plus a log line so that it shows up in the log. It has been fine since.

Your approach looks right to us; the per-BAR flag is cleaner than our side table.