Replace OBJ_MIGHTBEDIRTY with a system using atomics. Remove theTMPFS_DIRTY flag to simplify set_writeable_dirty().
ClosedPublic
Actions

Authored by jeff on Oct 22 2019, 10:02 PM.

Details

Reviewers

kib
markj
alc
dougm

Commits

rS354158: Replace OBJ_MIGHTBEDIRTY with a system using atomics. Remove the TMPFS_DIRTY

Summary

This replaces the MIGHTBEDIRTYFLAG with a 'cleangeneration' counter. This allows us to use an atomic to update 'generation' without touching the flag field. If cleangeneration != generation the object might be dirty.

Remove the object flag for TMPFS_DIRTY and re-use this mechanism directly to avoid the vnode type checks. object_page_clean() won't do anything for tmpfs so it will not reset cleangeneration.

With this in place more cases of vm_fault_dirty can work with a read object lock. I'm least certain about my changes to the vm_fault_soft_fast() callsite and vm_fault_dirty().

Diff Detail

Lint

Lint Passed

Unit

No Test Coverage

Build Status

Buildable 27159
Build 25431: arc lint + arc unit

Event Timeline

jeff created this revision.Oct 22 2019, 10:02 PM

Harbormaster completed remote builds in B27159: Diff 63556.Oct 22 2019, 10:02 PM

jeff retitled this revision from Replace OBJ_MIGHTBEDIRTY with a system using atomics. Remove the TMPFS_DIRTY flag to simplify set_writeable_dirty(). to Replace OBJ_MIGHTBEDIRTY with a system using atomics. Remove theTMPFS_DIRTY flag to simplify set_writeable_dirty()..Oct 22 2019, 10:08 PM

jeff edited the summary of this revision. (Show Details)

jeff added reviewers: kib, markj, alc, dougm.

One last point; it would be trivial to turn generation into a timestamp so that it could be used for more precise mtime without needing periodic polling.

kib added inline comments.Oct 23 2019, 1:24 PM

sys/vm/vm_fault.c
726	This was a microoptimization to avoid taking vm object lock shared when we cannot do a useful work with that lock.
sys/vm/vm_object.c
821	Why do you check the object type ?

jeff added inline comments.Oct 23 2019, 6:10 PM

sys/vm/vm_fault.c
726	I'm not sure I understand then. I made it possible to set mightbedirty with the shared lock. We should now be able to proceed here.
sys/vm/vm_object.c
821	This is leftover from another version of the patch that did not check it in vm_object_mightbedirty(). I will fix that.

kib added inline comments.Oct 24 2019, 4:01 PM

sys/vm/vm_fault.c
726	I meant that this was the reason for recheck under the lock, or rather, a reason to first do the checks without the lock held. I did not requested a change there.