Page MenuHomeFreeBSD

Synchronize page laundering with pmap_extract_and_hold()
ClosedPublic

Authored by markj on Aug 19 2017, 9:33 PM.
Tags
None
Referenced Files
Unknown Object (File)
Oct 31 2024, 5:43 PM
Unknown Object (File)
Sep 16 2024, 10:25 PM
Unknown Object (File)
Sep 15 2024, 12:07 PM
Unknown Object (File)
Sep 15 2024, 12:06 PM
Unknown Object (File)
Sep 5 2024, 2:05 PM
Unknown Object (File)
Sep 4 2024, 6:35 PM
Unknown Object (File)
Aug 17 2024, 6:10 PM
Unknown Object (File)
Aug 7 2024, 5:00 AM
Subscribers
None

Details

Summary

The hold_count check in vm_pageout_cluster() is susceptible to a race
with pmap_extract_and_hold() since the latter may be called without the
object lock or vnode lock held. In particular, there is a window between
the hold_count check and the later pmap_remove_write() call during which
a concurrent pmap_extract_and_hold() caller may obtain write access to
pages in the cluster. This appears to be a regression from r207410.

Attempt to handle this by atomically checking hold_count and removing
managed write mappings when adding a page to a pageout cluster. Note
that vm_object_page_collect_flush() ensures that the clustered pages
are read-only.

Diff Detail

Lint
Lint Passed
Unit
No Test Coverage
Build Status
Buildable 11172
Build 11553: arc lint + arc unit