[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
Re: aarch64 pmap tweaks for review
On Tue, Jun 02, 2020 at 05:03:39PM +0900, Ryo Shimizu wrote:
> That's excelent! And thank you.
> At a cursory glance, I think there is no problem.
> it's okay to remove the pv_pa (#ifdef XXXAD) part :)
vtophys() would be nice to use there but it looks quite tricky.
> >I made some tweaks to the aarch64 pmap based on lessons learned in the x86
> >pmap recently. They reduce memory consumption and speed up things like
> >fork/exec/exit/UBC a little:
> > http://www.netbsd.org/~ad/2020/aarch64.diff
> >Approximate times for kernel build on RK3399 with all 6 cores running at
> > before 1354.07s real 6092.55s user 1591.35s system
> > after 1307.90s real 6026.60s user 1432.83s system
> >Description below. Comments welcome.
> >- Fix a lock order reversal via pmap_page_protect().
> >- Align struct pmap to a cache line boundary.
> >- Move wired/resident count update out from PMAPCOUNTERS ifdef in one place.
> > It shouldn't depend on it.
> >- Make sure pmap is always locked when updating stats. Then atomics are no
> > longer needed to update stats.
> >- Remove unneeded traversal of PV lists in pmap_enter_pv().
> >- Shrink struct vm_page from 136 to 128 bytes (cache line sized - reduce
> > cache misses).
> >- Shrink struct pv_entry from 48 to 32 bytes (power of 2 sized - reduce cache
> > misses).
> >- Embed a pv_entry in each vm_page. That means PV entries don't need to be
> > allocated for pages that are mapped only once, for example private
> > anonymous memory / COW pages / most UBC mappings. Dynamic PV entries are
> > then used only for stuff like shared libraries and shared memory.
> >- Comment out PMAPCOUNTERS option because global counters are costly on MP
> > due to cache coherency overhead. The problem gets exponentially worse as
> > more CPUs are added.
> >- Use the pmap as a source of pre-zeroed pages for the VM system.
> >- Do unlocked checks in pmap_page_protect() and pmap_clear_modify(): avoid
> > taking the lock if the page has no mappings.
> ryo shimizu
Main Index |
Thread Index |