vm_fault.c: Introduce 'vm_fault_wire'
Needs ReviewPublic
Actions

Authored by bnovkov on Thu, Oct 15, 6:21 PM.

Project Tags

None

Referenced Files

	F177248855: D60751.1793998441.diff
	Thu, Nov 5, 8:54 PM

Subscribers

imp

Details

Reviewers

markj
kib

Summary

This change introduces 'vm_fault_wire', a memory wiring routine with
a fast path for objects implementing the own 'pgo_populate' pager
methods. The fast path (vm_fault_wire_quick) uses 'vm_pager_populate'
to preallocate pages for the entire map entry instead of using 'vm_page_grab'
for each 0-order pager. Unlike the slow path, it is superpage-aware and
will avoid extraneous page lookup when processing pages whose psind
is larger than zero. This change is mostly aimed at speeding up memory
wiring in the bhyve hypervisor.

Test Plan

See the test plan in D60752.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Skipped

Unit

Tests Skipped

Build Status

Buildable 78470
Build 75353: arc lint + arc unit

Event Timeline

bnovkov created this revision.Thu, Oct 15, 6:21 PM

Herald added a subscriber: imp. · View Herald TranscriptThu, Oct 15, 6:21 PM

bnovkov requested review of this revision.Thu, Oct 15, 6:21 PM

Harbormaster completed remote builds in B78245: Diff 189985.Thu, Oct 15, 6:21 PM

bnovkov added a parent revision: D60750: vm_reserv.c: Introduce 'vm_reserv_pindex_aligned'.Thu, Oct 15, 6:21 PM

bnovkov added a child revision: D60752: vmm/vmm_mem.c: Use a custom 'pgo_populate' operation for sysmem segments.

bnovkov edited the test plan for this revision. (Show Details)Thu, Oct 15, 6:22 PM

kib added inline comments.Fri, Oct 16, 8:07 PM

sys/vm/vm_fault.c
2538	Perhaps assert that the OBJ_POPULATE flag is set.
2541	Why do you need to take a reference on the object? If you pip_add-ed it, the object cannot be terminated.
2589	Why not do the cleanup from 'we broke out of the loop...' there? If you set rv = KERN_SUCCESS before the loop, that if() become unneeded I think.
2632	Why do you filter out MAP_ENTRY_COW? The indent on this and next line should be just +4 spaces.

Address @kib's comments

Harbormaster completed remote builds in B78470: Diff 190516.Wed, Oct 21, 7:36 PM

bnovkov added inline comments.Wed, Oct 21, 7:36 PM

sys/vm/vm_fault.c
2541	Thanks for pointing this out, the object reference was me being overly careful.
2632	I wanted to avoid dealing with cow-related things in the fast path, but after looking at the rest of `sys/vm` I see that filtering only MAP_ENTRY_NEEDS_COPY would suffice.

kib added inline comments.Fri, Oct 30, 6:23 PM

sys/vm/vm_fault.c
2538	And existing code calling the object populate method is careful to ensure that the map and the vnode are unlocked around it. E.g. it could end up in vm_page_grab sleeping and waiting for pagedaemon, with the vnode locked.
2569	I think that the term 'remove' is confusing. It is more like 'unbusy and clip the unneeded tail of the run' then removal of pages.
2586	Suppose that several calls to pmap_enter() succeeded, but then it failed, due to PMAP_ENTER_NOSLEEP. In this case, we have wired PTEs which point to non-wired pages. Even if this state is transient and cleaned later by the slow path, I think that unbusying these pages while PTEs are installed is not safe.