Page MenuHomeFreeBSD
Differential D23034

Reduce object locking in vm_fault. The object is only needed until we have an exclusive busy.
ClosedPublic
Actions

Authored by jeff on Jan 4 2020, 8:26 PM.
Tags
None
Referenced Files
Unknown Object (File)
Wed, Dec 25, 7:59 AM
Unknown Object (File)
Mon, Dec 16, 12:12 PM
Unknown Object (File)
Fri, Dec 13, 1:37 PM
Unknown Object (File)
Dec 9 2024, 2:30 PM
Unknown Object (File)
Nov 24 2024, 10:30 PM
Unknown Object (File)
Nov 10 2024, 8:28 PM
Unknown Object (File)
Nov 10 2024, 8:22 PM
Unknown Object (File)
Oct 17 2024, 2:06 PM
View All 58 Files
Subscribers
imp

Details

Reviewers
alc
dougm
kib
markj
Commits
rS356933: Reduce object locking in vm_fault. Once we have an exclusively busied page we
Summary

This drastically shortens maximum hold times on object locks in fault. We will basically only hold it through page allocation. It also eliminates some ugly lock/unlock/relock style code.

I intend to continue to refactor fault into smaller functions but I wanted to introduce the locking change first since it has the largest functional impact.

Diff Detail

Lint
Lint Passed
Unit
No Test Coverage
Build Status
Buildable 28627
Build 26659: arc lint + arc unit

Event Timeline

jeff created this revision.Jan 4 2020, 8:26 PM
Harbormaster completed remote builds in B28481: Diff 66350.Jan 4 2020, 8:26 PM
jeff mentioned this in D23033: Don't hold the object lock across getpages..Jan 4 2020, 8:38 PM
jeff edited the summary of this revision. (Show Details)Jan 4 2020, 8:45 PM
jeff added reviewers: alc, dougm, kib, markj.
jeff set the repository for this revision to rS FreeBSD src repository - subversion.
Herald added a subscriber: imp. · View Herald TranscriptJan 4 2020, 8:45 PM
jeff added inline comments.Jan 4 2020, 9:11 PM
sys/vm/vm_fault.c
337

We could almost drop the lock before pmap_enter() but we are relying on it to prevent invalidation of pages for read-only faults.

635

I don't really like this but my only other option is to perhaps make this bool part of the fault_state.

964

Default objects hold the lock for the full duration.

1123

I don't like the extra label but I am continuing to break this function up and simplify control flow in later patches.

1219–1234

Fixed indentation while adding the extra TRYWLOCK.

1471

This is actually a fairly substantial win.

kib added inline comments.Jan 6 2020, 1:48 AM
sys/vm/vm_fault.c
337

So we potentially dirty the object even despite the pmap_enter() could fail, same for prefault.

964

This explanation should be added to the comment, perhaps ?

1191–1192

'A valid page has been found, busied by us' (modulo my english).

1219–1234

And remove useless () around conditions while there, please.

jeff added inline comments.Jan 6 2020, 2:23 AM
sys/vm/vm_fault.c
337

Yeah it's safe to move this down. I can do so.

1191–1192

A valid page has been found and exclusively busied.

1219–1234

will do.

jeff updated this revision to Diff 66661.Jan 12 2020, 9:07 PM

Review feedback. Mostly style and comments. Re-shuffle vm_fault_soft_fast() to properly handle the error case.

Harbormaster completed remote builds in B28627: Diff 66661.Jan 12 2020, 9:07 PM
kib accepted this revision.Jan 13 2020, 8:39 PM
This revision is now accepted and ready to land.Jan 13 2020, 8:39 PM
markj accepted this revision.Jan 13 2020, 8:46 PM
Closed by commit rS356933: Reduce object locking in vm_fault. Once we have an exclusively busied page we (authored by jeff). · Explain WhyJan 20 2020, 10:50 PM
This revision was automatically updated to reflect the committed changes.
jeff added a commit: rS356933: Reduce object locking in vm_fault. Once we have an exclusively busied page we.

Revision Contents
Changeset List

PathSize
sys/
vm/
138 lines

Diff 66661

View Options

sys/vm/vm_fault.c

Loading...