DF-2877 / verdict.json
{ "finding_id": "DF-2877", "status": "reproduced", "reproduced": 1, "impact": "dos", "confidence": "certain", "verdict": "kdmsg_iocom_reconnect (kern_dmsg.c:139-144) waits for the old reader/writer iocom threads with a non-PCATCH lksleep loop and, unlike kdmsg_iocom_uninit (:283-294), has NO fp_shutdown escape. The old reader parks inside fp_read() on the old peer socket and only re-checks KILLRX at loop top, so a silent-but-open peer pins it forever; the DIOCRECLUSTER caller then blocks in an UNKILLABLE kernel sleep (verified live: process in D state, wchan clstrkl, survives kill -9 - twice, once accidentally and once scripted) and the disk's iocom stays wedged for all future attaches. On a stock boot the 'hammer2 service' daemon DIOCRECLUSTERs every disk with a pipe (cmd_service.c:898), so any second connect to a connected-but-idle disk wedges. Fix kernel: attach #2 waits 10.49s, then completes normally.", "exploit_chain": "privileged local (needs disk open = SYSCAP_RESTRICTEDROOT): attach#1 holds peer socket open idle -> attach#2 wedges unkillably + disk iocom dead; release only by closing attach#1's fds", "evidence": [ "run.log (pid in D2/clstrkl alive after kill -9: STILL_ALIVE_UNKILLABLE)", "VERDICT.md", "run.fix.log (10.49s then success)" ], "kernel_refs": [ "sys/kern/kern_dmsg.c:139", "sys/kern/kern_dmsg.c:141", "sys/kern/kern_dmsg.c:143", "sys/kern/kern_dmsg.c:287", "sbin/hammer2/cmd_service.c:898" ], "poc_changes": "No seed; added dfpeer 'hold' mode (attach + park)", "attempts": 2, "guest_uname": "DragonFly dfbsd 6.5-DEVELOPMENT DragonFly 6.5-DEVELOPMENT #0: Thu Jul 2 06:02:54 UTC 2026 root@dfbsd:/usr/obj/usr/src/sys/X86_64_GENERIC x86_64", "runtime_sec": 300, "guest_dirty": 1, "build_cmd": "cc -O -Wall -o /tmp/dfpeer dfpeer.c", "run_cmd": "sh run_f4.sh (root)", "code_hash": "e748ca2203c63d4fc7c412513eefcdf465800329f3e8ff3c36dbd9f5d0583722", "notes": "First observed accidentally when the boot daemon's pipes wedged two test attaches; then reproduced deterministically twice.", "recommended_fix": "Mirror kdmsg_iocom_uninit: 10s retry counter then fp_shutdown of the old msg_fp so the kill-wait always terminates", "fix_status": "fixed", "fix_kernel_uname": "DragonFly dfbsd 6.5-DEVELOPMENT DragonFly 6.5-DEVELOPMENT #1: Wed Sep 2 18:43:16 UTC 2026 root@dfbsd:/usr/obj/usr/src/sys/X86_64_GENERIC x86_64 (fix.diff applied, make nativekernel)", "fix_baseline_reproduced": 1, "fix_patched_reproduced": 0, "fix_verdict": "Fix kernel: second attach blocks 10.49s (retry+shutdown escape), then completes (BLK_OPEN reply error=0); no D-state process remains.", "fix_evidence": [ "run.fix.log" ] } |