vm_pageout_scans: correct detection of active object

[kib]

Summary

For non-anonymous swap objects, there is always a reference from the
owner to the object to keep it from recycling.  Account for it when
deciding should be query pmap for hardware active references for the
page.

[markj]

Am I right that this is effectively an optimization? For pages belonging to a named swap object we will call pmap_ts_referenced() unnecessarily, but the returned value will be correct (i.e., 0 if the page is unmapped).

sys/vm/vm_pageout.c

730

This comment is out of date, the page lock is no longer used. If the page is deallocated, the references are generally lost. I can fix this in a follow-up.

745

Extra newline.

863–864

This should presumably check whether ref_count > 1 for non-anonymous object.

1572–1573

Presumably this check should be updated too.

[kib]

In D33924#767299, @markj wrote:

Am I right that this is effectively an optimization? For pages belonging to a named swap object we will call pmap_ts_referenced() unnecessarily, but the returned value will be correct (i.e., 0 if the page is unmapped).

Now that you formulate it this way, I think yes. Then it is somewhat mistery what I saw on my machine, where active queue stayed infinitely long with large unmapped tmpfs objects. Anyway, this patch indeed saves at least one lock/unlock of pv lock per scanned page.

sys/vm/vm_pageout.c

730

I do not think that references of the page being deallocated serve any use.

[kib]

Plug two more places where object reference is used.

[markj]

In D33924#767364, @kib wrote:

In D33924#767299, @markj wrote:

Am I right that this is effectively an optimization? For pages belonging to a named swap object we will call pmap_ts_referenced() unnecessarily, but the returned value will be correct (i.e., 0 if the page is unmapped).

Now that you formulate it this way, I think yes. Then it is somewhat mistery what I saw on my machine, where active queue stayed infinitely long with large unmapped tmpfs objects.

Hmm. I haven't seen such a problem before. Was the system otherwise idle? Which kernel revision was it running?

Anyway, this patch indeed saves at least one lock/unlock of pv lock per scanned page.

Yes, I think it's a worthy change.

sys/vm/vm_pageout.c

1573

I think the code would be more clear if we instead had bool vm_pageout_object_mapped(vm_object) which returned true if there is a mapping reference, and false otherwise.

1611

This should be updated as well.

[kib]

Change vm_pageout_object_act() to do the ref comparision.
Move the big comment again (I did not touched the page lock part).
Plug one more ref_count != 0 check.

[alc]

sys/vm/vm_pageout.c

717

[kib]

Fix grammar in the comment.