You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
pmgr-misc: idle_fabric_pstate=1 crashes the internal DCP on M2 Max (t6021) #341
Drafted with an AI assistant (Claude Code) at my request.
On an M2 Max (MacBook Pro 14-inch, J414c, t6021), turning on the idle fabric governor crashes the DCP within about a second. The display goes black and the machine stops responding (no TTY), so it needs a hard reboot. The governor is off by default, and as far as we can tell nothing in our packages or command lines turns it on, so we don't think anyone is hitting this today.
What we ran
Kernel: integration/aurora 396f2d6 plus #336 and #337 (only to read the clock-generator words through debugfs), built with the stock aurora config.
echo 1 | sudo tee /sys/module/apple_pmgr_misc/parameters/idle_fabric_pstate
Then we read fabric_ps and dcs_ps from /sys/kernel/debug/apple-pmgr-misc-28e20c000.power-management/ four times a second.
Within the first second, fabric_ps showed 0x00000111 (fabric lowered to p-state 1) between readings of 0x00000555, and at the same second the kernel logged:
The readings went on for about 14 more seconds, then the machine hung. An external display was attached on the other DCP (289c00000.dcp) during the run.
With the governor at its default of 0, the same reads under CPU, memory and GPU load ran clean, so this is the governor lowering the fabric, not the debugfs reads.
Why we think it happens
apple_pmgr_fabric_voters is gfx, avd_sys and isp_sys. The comment says external displays vote through apple_pmgr_fabric_hold() instead. Nothing seems to keep the fabric up for the internal display's DCP, so once the CPUs are at their minimum the governor lowers the fabric while that DCP is still scanning out. The ~0.5 W figure in the parameter's comment was measured on t6000. We haven't checked whether t6000 tolerates this and t602x doesn't, or whether t6000 just wasn't tested with the panel on.
Suggestion
Until it's qualified, the governor could refuse idle_fabric_pstate on t602x (or on anything except the measured t6000 boards). The longer-term fix is to count the internal DCP as a bandwidth voter, the same way the external one holds the fabric.
We haven't retried it, since each run takes the display down.
Activity