On Sun, Jun 28, 2020 at 10:23 AM Rafael J. Wysocki <raf...@kernel.org> wrote: > > On Fri, Jun 26, 2020 at 8:43 PM Dan Williams <dan.j.willi...@intel.com> wrote: > > > > On Fri, Jun 26, 2020 at 7:22 AM Rafael J. Wysocki <raf...@kernel.org> wrote: > > > > > > On Fri, Jun 26, 2020 at 2:06 AM Dan Williams <dan.j.willi...@intel.com> > > > wrote: > > > > > > > > Quoting the documentation: > > > > > > > > Some persistent memory devices run a firmware locally on the device > > > > / > > > > "DIMM" to perform tasks like media management, capacity > > > > provisioning, > > > > and health monitoring. The process of updating that firmware > > > > typically > > > > involves a reboot because it has implications for in-flight memory > > > > transactions. However, reboots are disruptive and at least the Intel > > > > persistent memory platform implementation, described by the Intel > > > > ACPI > > > > DSM specification [1], has added support for activating firmware at > > > > runtime. > > > > > > > > [1]: https://docs.pmem.io/persistent-memory/ > > > > > > > > The approach taken is to abstract the Intel platform specific mechanism > > > > behind a libnvdimm-generic sysfs interface. The interface could support > > > > runtime-firmware-activation on another architecture without need to > > > > change userspace tooling. > > > > > > > > The ACPI NFIT implementation involves a set of device-specific-methods > > > > (DSMs) to 'arm' individual devices for activation and bus-level > > > > 'trigger' method to execute the activation. Informational / enumeration > > > > methods are also provided at the bus and device level. > > > > > > > > One complicating aspect of the memory device firmware activation is that > > > > the memory controller may need to be quiesced, no memory cycles, during > > > > the activation. While the platform has mechanisms to support holding off > > > > in-flight DMA during the activation, the device response to that delay > > > > is potentially undefined. The platform may reject a runtime firmware > > > > update if, for example a PCI-E device does not support its completion > > > > timeout value being increased to meet the activation time. Outside of > > > > device timeouts the quiesce period may also violate application > > > > timeouts. > > > > > > > > Given the above device and application timeout considerations the > > > > implementation defaults to hooking into the suspend path to trigger the > > > > activation, i.e. that a suspend-resume cycle (at least up to the syscore > > > > suspend point) is required. > > > > > > Well, that doesn't work if the suspend method for the system is set to > > > suspend-to-idle (for example, via /sys/power/mem_sleep), because the > > > syscore callbacks are not invoked in that case. > > > > > > Also you probably don't need the device power state toggling that > > > happens during regular suspend/resume (you may not want it even for > > > some devices). > > > > > > The hibernation freeze/thaw may be a better match and there is some > > > test support in there already that may be kind of co-opted for your > > > use case. > > > > Hmm, yes I guess freeze should be sufficient to quiesce most > > device-DMA in the general case as applications will stop sending > > requests. > > It is expected to be sufficient to quiesce all of them. > > If that is not the case, the integrity of the hibernation image cannot > be guaranteed on the system in question. >
Ah, indeed, I was overlooking that property. > > I do expect some RDMA devices will happily keep on > > transmitting, but that likely will need explicit mitigation. It also > > appears the suspend callback for at least one RDMA device > > mlx5_suspend() is rather violent as it appears to fully teardown the > > device context, not just suspend operations. > > > > To be clear, what debug interface were you thinking I could glom onto > > to just trigger firmware-activate at the end of the freeze phase? > > Functionally, the same as for suspend, but using the hibernation > interface, so "echo platform > /sys/power/pm_test" followed by "echo > disk > /sys/power/state". > > But it might be cleaner to introduce a special "hibernation mode", ie. > is one more item in /sys/power/disk, that will trigger what you need > (in analogy with "test_resume"). I'll move the trigger to be after process freeze, but I'll keep it tied to suspend-debug vs hibernate-debug. It appears the hibernate debug path still goes through the exercise of allocating memory for the hibernation image which is unnecessary if the goal is just to 'freeze', 'activate', and 'thaw'.