problem reported here but is part of the solution):
xen_kpm_sync() is not working as expected,
leading to races between CPUs.
1 the check (xpq_cpu != &x86_curcpu) is always false because we
have different x86_curcpu symbols with different addresses in the kernel.
Fortunably, all addresses dissaemble to the same code.
Because of this we always use the code intended for bootstrap, which doesn't
use cross-calls or lock.
2 once 1 above is fixed, xen_kpm_sync() will use xcalls to sync other CPUs,
which cause it to sleep and pmap.c doesn't like that. It triggers this
KASSERT() in pmap_unmap_ptes():
KASSERT(pmap->pm_ncsw == curlwp->l_ncsw);
3 pmap->pm_cpus is not safe for the purpose of xen_kpm_sync(), which
needs to know on which CPU a pmap is loaded *now*:
pmap->pm_cpus is cleared before cpu_load_pmap() is called to switch
to a new pmap, leaving a window where a pmap is still in a CPU's
ci_kpm_pdir but not in pm_cpus. As a virtual CPU may be preempted
by the hypervisor at any time, it can be large enough to let another
CPU free the PTP and reuse it as a normal page.
To fix 2), avoid cross-calls and IPIs completely, and instead
use a mutex to update all CPU's ci_kpm_pdir from the local CPU.
It's safe because we just need to update the table page, a tlbflush IPI will
happen later. As a side effect, we don't need a different code for bootstrap,
fixing 1). The mutex added to struct cpu needs a small headers reorganisation.
to fix 3), introduce a pm_xen_ptp_cpus which is updated from
cpu_pmap_load(), whith the ci_kpm_mtx mutex held. Checking it with
ci_kpm_mtx held will avoid overwriting the wrong pmap's ci_kpm_pdir.
While there I removed the unused pmap_is_active() function;
and added some more details to DIAGNOSTIC panics.
d3ccea851c