Here is a refresh of partition hibernation support for pSeries. It includes a fix for a regression in the partition migration support. Barring any objections, I think these patches should be ready to merge.
Overview --------- Partition Hibernation on pSeries is a new platform feature that allows for long term suspension of a logical partition much like suspending a laptop to disk. The primary difference is that writing the memory image out to disk is driven by system firmware and the Virtual I/O Server rather than from the LPAR itself. Partition hibernation on Power is initiated from the Hardware Management Console. The user selects the partition, then selects the suspend function. This results in a command (drmgr) getting sent to the Linux partition, indicating it should prepare for suspension. A "stream id" is sent to the Linux LPAR which is used by the OS to correlate with firmware. In the Linux LPAR, the drmgr command then writes this stream id to a new sysfs file: /sys/devices/system/power/hibernate. The kernel then takes over, calling H_VASI_ENABLE with the stream id as long as firmware indicates it is suspending but not ready for the client LPAR to enter the final phase of the suspension. Once H_VASI_ENABLE returns a state of H_VASI_SUSPENDING, the client OS is expected to enter the final phase of hibernation. To do this, we then simply invoke the pm_suspend code and mimic suspend to ram. We mimic suspend to ram rather than suspend to disk, since firmware and the VIOS takes care of writing everything out to disk. We are then able to leverage all the existing suspend code in the kernel. Once we enter the prepare_late phase of suspend, we set a flag which we check when disable_nonboot_cpus gets called. When the nonboot CPU gets offlined and placed into the inactive state, we hook into pseries_mach_cpu_die in order to call H_JOIN, since the nonboot CPUs need to be in H_JOIN state when we finally suspend. Once all the nonboot CPUs have been offlined and are in H_JOIN, and we get to the "enter" state, we make the ibm,suspend-me RTAS call on the remaining CPU which then completes the hibernation. When we resume, there is very little platform code required to execute. enable_nonboot_cpus already sends an H_PROD as part of bringing up the nonboot cpus, so this will kick the CPU out of H_JOIN. I've already added resume handlers to the virtual I/O drivers to check for any dropped interrupts. drmgr then handles updating the device tree just like it does today for live partition migration. These patches have been tested with repeated suspend/resume cycles. Partition migration has also been regression tested since the first patch touches that path. -- Brian King Linux on Power Virtualization IBM Linux Technology Center _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
