Double vm_object reference in cdev_pager_allocate leaks device pager objects forever
Summary
vm_pager_object_lookup at device_pager.c:131 already increments ref_count (vm_pager.c:367). Found branch at :155-164 then calls vm_object_reference_locked(object) AGAIN at :160 -> ref_count += 2 per call instead of +1. Object never returns to 0 after all munmap -> dev_pager_dealloc never runs -> object never removed from dev_pager_object_list -> si_object never cleared -> permanent leak. Each D_MMAP device mmap-ed twice leaks one vm_object + dev_pager_object_list node. List lookups are O(N) so sustained leak degrades mmap perf for all devices. Most D_MMAP devices require privilege but some world-readable. Fix: drop vm_object_reference_locked; vm_pager_object_lookup already took our ref.
Discussion (0)
PoC verification
Evidence pack
findings/poc/DF-0958 · 2 files| File | Type | Description | Size | |
|---|---|---|---|---|
| fix.diff | suggested-fix | git-apply-able fix for the cited path | 512 B | view raw |
| VERDICT.md | verdict | source-confirmation narrative | 968 B | ↓ raw |
DF-0958 source-confirmation
Verdict: REPRODUCED (source-confirmed)
Impact: dos Confidence: likely
Kernel ref: sys/vm/device_pager.c:160
Mechanism
double vm_object_reference_locked -> refcount leak, object never freed
Confirmation method
Source-trace confirmed the cited code path matches the finding (exact line/condition
verified against sys/). Runtime PoC not exercised for this source-only Low-severity
item; confirmation is by code inspection.
Recommended fix
See fix.diff in this folder (git-apply-able unified diff).
Phase 8 (combined build)
All 64 fixes were batched into one combined patch
(../_batch/combined_64.patch) and applied to in-guest /usr/src. A single
make -j6 nativekernel KERNCONF=X86_64_GENERIC completed rc=0 with 0 errors under
-Werror (../_batch/fix_build.log, 35374 lines). The GENERIC kernel + all modules
(drm, firewire, usb, netgraph, smbfs, fuse, crypto, vm, pmap) compiled clean.
Fix verification
not_testablenot_testable (source-only): no runtime PoC exercised; fix.diff applies cleanly and the combined single nativekernel build of all 64 fixes completed rc=0 with 0 errors under -Werror (findings/poc/_batch/fix_build.log). Bug source-confirmed in baseline tree. Compile-validation of the fix is the requested Phase-8 deliverable for this batch.
combined build: '=== NK_DONE rc=0 ===' and '>>> Kernel build for X86_64_GENERIC completed on Thu Jul 23 09:17:55 UTC 2026' (0 'error:' lines in 35374-line log).
Confirmed kernel references
- s
- y
- s
- /
- v
- m
- /
- d
- e
- v
- i
- c
- e
- _
- p
- a
- g
- e
- r
- .
- c
- :
- 1
- 6
- 0
Detail
Exploit chain
none (non-corruption / source-only confirmation; no memory-corruption escalation chain developed for this Low-severity item)
Evidence (decisive lines)
source-only: double vm_object_reference_locked -> refcount leak, object never freed @ sys/vm/device_pager.c:160. Combined fix build: '=== NK_DONE rc=0 ===' / 'Kernel build for X86_64_GENERIC completed', 0 errors under -Werror (findings/poc/_batch/fix_build.log).
PoC changes
authored git-apply-able fix.diff targeting the cited line; source-trace verified the vulnerable path. Evidence pack (VERDICT.md, manifest.json, env.txt, fix.diff) in findings/poc/DF-0958/.
Verified recommended fix
drop the redundant vm_object_reference_locked (lookup already took the ref). matches finding proposal. Full diff in findings/poc/DF-0958/fix.diff.
Verdict
REPRODUCED (source-confirmed). Source-trace confirmed double vm_object_reference_locked -> refcount leak, object never freed at sys/vm/device_pager.c:160 against the audited sys/ tree; exact line/condition verified.
No comments yet.