Re: libc.so's vmobjlock / v_interlock

[Andrew Doran <ad%netbsd.org@localhost>]

On Tue, Jan 21, 2020 at 09:46:03PM +0000, Andrew Doran wrote:

> - PV lists in the pmap need explicit locking (unless the pmap uses a global
>   lock, which is fine if it suits the machine).  I have done the PV list
>   locking for x86 and it's in the patch.

To follow up, the x86 changes are in and and I'm checking over the other
pmaps at the moment.  Most are in good shape.  Anyway this detangles the x86
pmap from UVM object locks.  It doesn't care any more *.  (Other than e.g. 
needing to hold a lock across pmap_remove() -> pmap_update(), to stop pages
gaining a new identity while still visible somewhere via stale TLB entry.)

It would be nice to to track all these mappings in batch, at the
uvm_map_entry level or something like that, cos the PV tracking scheme
causes uses loads of memory and is causes big cache contention.

Andrew