Weaken three asserts in preparation for lock pushdown in fault.
ClosedPublic
Actions

Authored by jeff on Dec 2 2019, 11:47 PM.

Details

Reviewers

alc
doug_freebsd.con.com
kib
markj

Commits

rS355516: It is safe to wire a page while the object is busy.
rS355515: It is now safe to rename a page that is still on a queue. Allowing this
rS355514: Do not assert that the object lock is held in vm_object_set_writeable_dirty.

Summary

I believe vfs_set_writeable_dirty is safe without an object lock at all. I originally had it required to keep the type stable but it does not matter if we race with dead and that would be a reference bug.

Mark tells me that I can now vm_page_replace() while a page is in queue. This allows me to optimize fault in two ways. Firstly, we allocate and free a page in first_object every time we discover a busy backing object page. We only need to do this once and leave it invalid but in a paging queue. the cow optimization replace runs into this. Secondly, this allows me to skip requiring the first_object write lock in cases where we want to free the page.

Lastly, vm_object_busy() is a standin for xbusy. Allow wired refs while the object is busy.

Diff Detail

Repository

rS FreeBSD src repository - subversion

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

jeff created this revision.Dec 2 2019, 11:47 PM

Harbormaster completed remote builds in B27898: Diff 65145.Dec 2 2019, 11:47 PM

jeff edited the summary of this revision. (Show Details)Dec 2 2019, 11:51 PM

jeff added reviewers: alc, doug_freebsd.con.com, kib, markj.

jeff set the repository for this revision to rS FreeBSD src repository - subversion.

Herald added a subscriber: imp. · View Herald TranscriptDec 2 2019, 11:51 PM

markj added inline comments.Dec 3 2019, 3:32 PM

sys/vm/vm_object.c
2228 ↗	(On Diff #65145)	Can we instead assert that object->ref_count > 0?
sys/vm/vm_page.c
1692 ↗	(On Diff #65145)	As you noted, I think this is fine. The assertion was added in r331128 but the reasoning given in the commit message no longer applies: the page daemon now handles pages without an object (by ignoring them).
3774 ↗	(On Diff #65145)	This should be fine now that xbusy is treated as a proper lock.