Add vm_page_grab_pages()

[alc]

Summary

I've made some minor changes from the version that we've discussed in email. I highlight the notable ones below.

[alc]

vm/vm_page.c

3202–3204

I changed the way that this function determines if it needs to sleep on a page. This snippet is copied from vm_page_grab(). In many cases, the old version would perform both vm_page_xbusied() and vm_page_busied(). Now, we always do one or the other, never both.

3213

I made the sleep channel message distinct from vm_page_grab().

[markj]

vm/vm_page.c

3174

The NOBUSY flag description needs to be updated.

[alc]

Should vm_page_grab_pages() also include this assertion from vm_page_grab():

KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));

[alc]

Update the description of VM_ALLOC_NOBUSY.

Update the block comment in vm_page.h to reflect the capabilities of vm_page_grab_pages().

[markj]

In D11926#247397, @alc wrote:

Should vm_page_grab_pages() also include this assertion from vm_page_grab():

KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
    (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
    ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));

I can't see any reason not to, especially if we're considering implementing vm_page_grab() in terms of vm_page_grab_pages(). It might be that there's some valid use-case for specifying ALLOC_SBUSY without ALLOC_IGN_SBUSY, but until that arises it seems reasonable to be strict.