Revision 2 as of 2007-10-28 18:16:02

Clear message


Modern power management is handled via ACPI, a spec designed by various companies (Microsoft, Intel, Phoenix, HP, etc). It contains information about the given system's hardware, including details on how to suspend/resume (and hibernate).

To use suspend, an operating system must first configure wake-up events (things like the power button, lid-button, etc). These are listed in /proc/acpi/wakeup (though it may not always list the power button). To enable a given device, one can echo it to the file:

  • echo "MODM" > /proc/acpi/wakeup

To enter suspend, the OS writes the desired state to /sys/power/state. Assuming the "mem" suspend, the kernel does the following:

  • logically ejects all CPUs except boot processor
  • disables devices
    • call suspend() functions (saving configuration registers)
    • enter D3 state (consuming 0 or nearly 0 power)
  • read ACPI FADT (fixed address descriptor table) magic values
  • write ACPI registers with magic values
  • BIOS enters "system management mode", directly interacting with the hardware
    • PTS: prepare to sleep
    • GTS: go to sleep
      • may cause BIOS to cut power to various power planes
  • wake-up event occurs
  • BIOS checks for resume vs new boot
  • BIOS runs kernel's wake-up code and re-enters the kernel
  • enable devices
    • enter D0 state (working)
    • restore configuration registers
  • logically re-inject all CPUs

Above the kernel layer, user-space must do work before handing off control to the above kernel procedures. Currently Ubuntu uses /etc/acpi/sleep.sh to enter/leave suspend mode. HAL makes calls to this script when reacting to various key presses or power management events. The goal is to move to pmutils in Hardy.

The call stack:

  • gnome-power-manager
  • hal
  • /etc/acpi/sleep.sh
  • /sys/power/state
  • kernel
  • BIOS

Bad packages (swsusp, hibernate) can interfere with /etc/acpi/sleep.sh, so make sure they are purged.


  • S1 = "standby" (stop processor and keep power on to everything)
  • S2 (never implemented, seen as the same as S3)
  • S3 = "mem" (save and restore processor state via memory, keeping some things powered -- suspend)
  • S4 = "disk" (save and restore processor state via disk, keeping nothing powered -- hiberation)
  • S5 = active running processor


  • Biggest problem is graphics hardware
    • kernel doesn't know how to handle graphical devices
    • BIOS knows how to restore graphics state
      • via 16 bit segmented mode, C000:xxxx contains the visible 64k video ROM.
      • starting execution at C000:0003, normally re-POSTs the video BIOS (/usr/sbin/vbetool post)

        • more difficult in 64bit mode, since 16bit calls need to be emulated.
        • some memory is in 3-4G range, which requires remapping when emulating to avoid hitting the kernel which is mapped in the same space.
      • video BIOS may have paged POST code out of C000 window
      • nvidia BIOS rewrites ROM to just return to stop re-POSTing
  • try suspend from console (via /etc/acpi/sleep.sh)

    • make sure you're logged out of Xorg (or run sleep.sh with "force" argument)
    • if the video BIOS isn't left in a sane state, returning to Xorg may hang the hardware
    • tests capslock on resume (if no capslock, kernel hung)
    • if backlight doesn't come back on, video BIOS probably didn't reinitialize
    • if screen is blank, but has a backlight, try hitting enter or switching between virtual terminals
    • try in single-user mode (via appending "single" to the grub kernel boot options)
    • for details on actions, try bash -x /etc/acpi/sleep.sh >/root/sleep.log 2>&1

    • look at dmidecode information that matches settings in /usr/share/acpi-support/*.config

  • if single-user mode console suspend fails
    • enable PM trace (echo "1" > /sys/power/pm_trace) which will write device hashes to the system timer

    • attempt to suspend
    • after the failure, on reboot, examine the dmesg output for "device hash" entries to track down the device that hung the system during resume.
    • be aware that this will reset the system clock, and fsck will freak out ("has gone without a fsck for 31337 days"). consider tune2fs -c 0 /dev/your/filesystems.

  • can you resume multiple times?