Rebase.

Fixes: 7becd87c988c ("vm_grab: use iterator for grab lookup")

D49741 has just the vm_fault.c changes added to the latest update here, so that it can be reviewed separately.

Add an iterator to the vm_fault.c code that leads to page allocation, so that the failed lookup is saved in the iterator and not repeated in the allocation call. Testing suggests that this reduces cycle count for page allocation.

Apply some suggestions from @bnovkov

Update comments.

Rename the function and de-inline it. Modify comments.

Update the correct variable.

Sorry, that was all to complicated. Simplify.

Don't break pindex.

Accept suggestions from @kib.

Apply suggestions from @kib.

Apply suggested comment edit.

Edit function header comment slightly.

Seek to improve the function name and header comment.

Add a *found parameter to vm_page_grab_lookup to distinguish found-but-sleep cases from not-found cases.

In vm_page_grab, if VM_ALLOC_WAITFAIL is set in allocflags and vm_page_grab_lookup returns NULL, return NULL.

Backing out the last update to the patch, since I've evidently completely failed to understand @kib's concerns, and so won't try to eliminate the WAITFAIL case from page grabbing.

Add/restore some comments.

In D49391#1128121, @markj wrote:

In D49391#1127942, @dougm wrote:

In D49391#1127890, @markj wrote:

The functional portion of the change seems fine, but I'm not sure why it's a problem for vm_page_grab* to support VM_ALLOC_WAITFAIL?

Because in comments on D49371, @kib suggested that VM_ALLOC_WAITFAIL was invalid here. So, to remove that obstacle to getting D49371 approved, I investigated and found that he might be right. Now, @alc has verbally placed a hold on this change because, when he is freed from another task, he plans to look for other illegal allocflag settings.

My reading of the comments is that kib is pointing out a bug in the new code.

In D49415#1128114, @markj wrote:

Doesn't this leave msucc uninitialized in vm_reserv_from_object()'s callers?

Only if the function returns a non-NULL value.

Apply @markj comments.

In D49391#1127890, @markj wrote:

The functional portion of the change seems fine, but I'm not sure why it's a problem for vm_page_grab* to support VM_ALLOC_WAITFAIL?

Thanks for the clarifications. Looks okay to me.

Update after (expected) checkin of D49391.

Further correct the page allocation parameter comments of vm_page.h, guided by @alc. Add a func parameter to vm_page_grab_check.

Correct a couple of vm allocflags annotations in readcli.h.

Apply @alc suggestion for func in KASSERTS in vm_page_grab_check().

Address @kib comments.

In vm_page_grab_lookup, use the iterator for the lookup. Duh.

Apply suggestions from @alc. I expect to commit it tonight, so further comments should be made promptly.

Undo a fatal line-swap from the last version.

Apply @alc suggestions.

Rewrite a comment. Take some suggestions from @kib, adapt some others. Rename the new function - though still need a better name.

Let vm_object_getpages() do the copying of the pages between the borders, even though it happens with a lock held when that's not strictly necessary.

Move the umz_zfree() in swap_pager_swapoff() to past the last backjumping goto.

The latest version has held up under a few hours of stress testing. Drop the handling of NULL parameters from vm_object_getpages().

Allocate the struct buf for swap_pager_getpages_locked early.

Fix comments.

Combine the two vm_object functions into one. Add assertions.

After some goto targets, assert that the iterators have been reset.

Rebase.