diff options
Diffstat (limited to 'scripts')
| -rwxr-xr-x | scripts/zz-bluetooth-resume | 23 |
1 files changed, 16 insertions, 7 deletions
diff --git a/scripts/zz-bluetooth-resume b/scripts/zz-bluetooth-resume index 4273339..4b51dae 100755 --- a/scripts/zz-bluetooth-resume +++ b/scripts/zz-bluetooth-resume @@ -5,13 +5,16 @@ # A systemd-sleep hook. Two things break bluetooth across sleep on a TLP # laptop, and nothing else on the machine fixes either one. # -# 1. The rfkill soft-block is not restored. systemd-rfkill would do it, and +# 1. The rfkill soft-block is not cleared. systemd-rfkill would do it, and # it is masked here deliberately -- it fights TLP's radio handling, so # configure_tlp_power masks it and TLP owns radios instead. TLP's own -# sleep hook runs `tlp resume`, but its setting is -# DEVICES_TO_ENABLE_ON_STARTUP: startup, not resume. TLP has no ON_RESUME -# at all, so the resume edge has no owner. WiFi survives only because -# NetworkManager unblocks itself; bluetooth has no equivalent. +# sleep hook runs `tlp resume`, but that only puts bluetooth back where +# it was at suspend and applies no policy: its enable setting is +# DEVICES_TO_ENABLE_ON_STARTUP, startup not resume, and TLP has no +# ON_RESUME at all. So a block present at suspend stays, and so does one +# that a controller re-probed after that replay comes up with. WiFi +# survives only because NetworkManager unblocks itself; bluetooth has no +# equivalent. # # 2. The controller comes back wedged from a hibernate. It reports powered # and unblocked while scanning finds nothing whatever -- zero devices @@ -28,8 +31,14 @@ # A machine whose TLP config does not ask for bluetooth keeps it off, which is # what stops this from overriding a deliberate block at every wakeup. # -# The zz- prefix orders it after TLP's own hook, so `tlp resume` has finished -# before this runs. +# The zz- prefix buys no ordering: systemd-sleep runs every hook here in +# parallel (man systemd-sleep), so `tlp resume` may still be running when +# this starts, and its first act is to replay the rfkill state it saved at +# suspend. On a hibernate-class wake the settle between unload and load below +# usually lands the final unblock after that replay. A plain suspend has no +# settle and unblocks at once, so that race is open. The tell is `rfkill list +# bluetooth` reading "Soft blocked: yes" after a suspend cycle that was entered +# with bluetooth already blocked (a hook that never ran reads the same way). # # Test seams: BTR_RFKILL, BTR_MODPROBE, BTR_TLP_CONF, BTR_TLP_CONF_DIR, # BTR_SETTLE (seconds to wait between driver unload and load). |