From 4fb583cf545b120ac5f43570d02855b5299988be Mon Sep 17 00:00:00 2001
From: Sergei Poljanski <me@asxp.io>
Date: Sun, 8 Nov 2026 05:08:45 +0400
Subject: [PATCH] power: drop the disk swapfile and hibernate

Remove the 58 GiB swapfile on the LUKS root, along with what depended
on it: resumeDevice, resume_offset, the suspend-then-hibernate logind
actions and HibernateDelaySec. zram is now the only swap, capped at
8 GiB so it can't cut into memory needed for large LLM/GTT loads.

Lid, idle and the suspend key now do plain s2idle suspend; the power
key powers off. This loses the hibernate fallback that used to end a
spurious-wake loop; acpi.ec_no_wakeup=1 and the gsd-power tweaks are
now all that keeps s2idle stable.

Assisted-by: Claude:opus-5-5
---
 hardware-configuration.nix |  7 ++----
 modules/power.nix          | 48 ++++++++++++++------------------------
 2 files changed, 19 insertions(+), 36 deletions(-)

diff --git a/hardware-configuration.nix b/hardware-configuration.nix
index 63650d5..2d9d50b 100644
--- a/hardware-configuration.nix
+++ b/hardware-configuration.nix
@@ -40,11 +40,8 @@
       options = [ "nofail" ];
     };
 
-  # Encrypted swapfile on LUKS root, sized for hibernate (RAM + headroom).
-  # NixOS auto-creates the file and computes resume offset on activation.
-  swapDevices = [
-    { device = "/var/swapfile"; size = 59392; }  # 58 GiB, RAM is 54 GiB
-  ];
+  # No disk swap: zram only (modules/power.nix). Hibernate disabled.
+  swapDevices = [ ];
 
   nixpkgs.hostPlatform = lib.mkDefault "x86_64-linux";
   hardware.cpu.amd.updateMicrocode = lib.mkDefault config.hardware.enableRedistributableFirmware;
diff --git a/modules/power.nix b/modules/power.nix
index 7c8676a..431e5a8 100644
--- a/modules/power.nix
+++ b/modules/power.nix
@@ -3,55 +3,41 @@
 
 {
   # AMD P-State driver (active mode for full EPP control)
-  # resume_offset = first physical block of /var/swapfile (for hibernate from
-  # an ext4 swapfile on LUKS). Recompute if the swapfile is recreated:
-  #   sudo filefrag -v /var/swapfile | awk 'NR==4{gsub(/\.\./,""); print $4}'
   # acpi.ec_no_wakeup=1: prevents EC wakeups from draining battery in s2idle
   # on P14s G6 AMD. See https://lore.kernel.org/all/ZnFYpWHJ5Ml724Nv@ohnotp/
-  boot.kernelParams = [ "amd_pstate=active" "resume_offset=133797888" "acpi.ec_no_wakeup=1" ];
+  boot.kernelParams = [ "amd_pstate=active" "acpi.ec_no_wakeup=1" ];
 
   # Allow manual fan control via /proc/acpi/ibm/fan (level 0-7, auto,
   # full-speed). Firmware stays in charge while level is "auto".
   boot.extraModprobeConfig = "options thinkpad_acpi fan_control=1";
 
-  # Hibernate / resume from encrypted swapfile on LUKS root.
-  # P14s G6 AMD has no S3 (Modern Standby only); this enables
-  # suspend-then-hibernate so long sleeps drain to disk.
-  boot.resumeDevice = "/dev/mapper/luks-71c97ce7-dd29-4b07-8d2a-8cc3985a65bd";
+  # Hibernate disabled: the disk swapfile was removed (it was the resume
+  # target). To restore: re-add swapDevices, boot.resumeDevice,
+  # resume_offset (filefrag -v /var/swapfile) and the *-hibernate actions.
 
-  # zram ahead of the disk swapfile: when a large LLM load pins 20-30 GiB of
-  # GTT, evicted app pages compress into RAM instead of contending with model
-  # reads on the same NVMe (this contention caused 8-minute model loads).
-  # Hibernate is unaffected: the image goes to resumeDevice/resume_offset
-  # (the swapfile), which stays available as overflow swap at priority -2.
+  # zram only, capped at 8 GiB so it can't eat into memory needed for large
+  # LLM/GTT loads. memoryMax caps the 50% default (~30 GiB).
   zramSwap = {
     enable = true;
     memoryPercent = 50;
+    memoryMax = 8 * 1024 * 1024 * 1024;
     priority = 100;
   };
 
-  # On battery: suspend-then-hibernate (s2idle then hibernate after
-  # HibernateDelaySec). Previously this looped suspend->wake->suspend due
-  # to ACPI EC interrupts, but acpi.ec_no_wakeup=1 (above) plus disabling
-  # gsd-power's sleep handlers (in desktop.nix via dconf) makes it stable.
-  # On AC too: plain suspend let a spurious-wake loop run unbounded for 1.5
-  # days (logind re-suspends every 30s while the lid is closed); with
-  # suspend-then-hibernate it ends in hibernate after HibernateDelaySec.
+  # Without swap, hibernate is unavailable: plain s2idle suspend everywhere.
+  # Caveat: plain suspend once let a spurious-wake loop run for 1.5 days on
+  # AC (logind re-suspends every 30s with the lid closed); there's no longer
+  # a hibernate fallback to end it. acpi.ec_no_wakeup=1 + gsd-power dconf
+  # tweaks (desktop.nix) are what keep s2idle stable.
   services.logind.settings.Login = {
-    HandleLidSwitch = "suspend-then-hibernate";
-    HandleLidSwitchExternalPower = "suspend-then-hibernate";
-    HandleSuspendKey = "suspend-then-hibernate";
-    HandlePowerKey = "hibernate";
-    IdleAction = "suspend-then-hibernate";
+    HandleLidSwitch = "suspend";
+    HandleLidSwitchExternalPower = "suspend";
+    HandleSuspendKey = "suspend";
+    HandlePowerKey = "poweroff";
+    IdleAction = "suspend";
     IdleActionSec = "30min";
   };
 
-  # s2idle window before transitioning to hibernate. acpi.ec_no_wakeup=1
-  # (above) prevents the prior EC-driven s2idle thrash on this machine.
-  # Conservative value while bag-thermal behavior is being re-validated;
-  # raise to 30-45min once confirmed safe.
-  systemd.sleep.settings.Sleep.HibernateDelaySec = "15min";
-
   # Stop USB devices (e.g. Aerox 3 WL receiver) from waking the machine
   # from s2idle. Only usb_device nodes have power/wakeup; matching on the
   # devtype avoids "Invalid operator for ATTRS" on udev verification.