Augment vm_page_grab to support more cases that want to acquire busy locks.

[jeff]

Summary

This patch adds two new flags, VM_ALLOC_NOCREAT & VM_ALLOC_VALID, so that we can move more consumers to grab() and eliminate hand-rolled loops with sensitive synchronization.

NOCREAT does the lookup/busy/wire in a loop but fails if it has to allocate a page.
VALID only returns existing valid pages.

There are other callers which can be refactored into this paradigm later.

The second part of the patch implements vm_page_busy_acquire() which does a looping busy acquire so callers are not coding their own sleep_if_busy() loops or blindly calling vm_page_xbusy() and expecting it to succeed.

[kib]

sys/vm/vm_page.c

3895

Why do you changed the setting of PGA_REFERENCED ?

3921

It is somewhat strange semantic. If only ALLOC_VALID was specified, without ALLOC_NOCREAT, why zeroed page cannot satisfy the caller ?

[jeff]

sys/vm/vm_page.c

3895

I flipped this the wrong way. It should be NOCREAT == 0.

Truncate, for example, wants to grab the last page to zero it, but only if it already exists. If not the pager will do it properly when it's created and you just waste time paging in an unnecessary page.

3921

I overloaded the VALID name. I think we could use a different name for VM_ALLOC_VALID.

It basically means, only return an existing valid page, but do no work to make it valid.

There are some other weirdnesses. For example, vm_page_grab_pages() will set the page bits valid when it zeros a page. vm_page_grab() does not.

I like this api layer that allows us to centralize synchronization but I feel at some point it could be simplified and split into a few smaller functions. The swiss army knife of flags always gets confusing.

[jeff]

sys/vm/vm_page.c

3895

How about ALLOC_IFVALID

[kib]

sys/vm/vm_page.c

3895

I still do not see why would we remove protection against pagedaemon during the sleep. We did found the page, so why NOCREAT/IFVALID should make it more likely for the page to be reclaimed under us, while we sleep for busy ?

3921

I understand this, my point is that allocating a page is much more work than validate it by zeroing.