Am 08.08.2012 03:45, schrieb David Gibson: > On Wed, Aug 08, 2012 at 12:32:39AM +0200, Andreas Färber wrote: >> Am 08.08.2012 00:02, schrieb Benjamin Herrenschmidt: >>> On Fri, 2012-08-03 at 17:01 +0200, Andreas Färber wrote: >>>> >>>> I have posted a suggestion where CPU reset is triggered by "the >>>> machine >>>> as an abstract concept" (needs a bit of tweaking still, but the >>>> general >>>> idea is there). >>>> Based on that, shouldn't it be rather easy to add a Notifier similar >>>> to >>>> "machine init done" that lets individual machines do post-reset setup? >>>> I.e. not have QEMUMachine trigger and control the reset. >>>> >>> >>> Note that we really want pre and post reset vs the device reset. >>> >>> That's why the machine should be the one in charge. The top level of the >>> reset sequencing is -not- the CPU, it's the machine. All machines (or >>> SoCs) have some kind of reset controller and provide facilities for >>> resetting individual devices, busses, processor cores.... the global >>> "system" reset (when it exists) itself might have interesting ordering >>> or sequencing requirements. >>> >>> Now, to fix our immediate problem on ppc for 1.2 the hook proposed by >>> Anthony for which David sent a patch does the job just fine, it allows >>> us to clean out all our iommu tables before the device-reset, meaning >>> that in-flights DMA cannot overwrite the various "files" (SLOF image >>> etc.... that are auto-loaded via reset handlers implicitely created by >>> load_image_targphys), and we can then do some post-initializations as >>> well to get things ready for a restart (rebuild the device-tree, etc...) >> >> That's all good, except for embedded machines without such implicit >> reset handling. It does contradict the "a machine is just a config file, >> setting up QOM objects" concept, but I was not the one to push that! :) >> >> What I was thinking about however were those mentioned individual cores >> being reset using cpu_reset(). If we want to piggy-back some >> machine-specific register initialization for individual CPUStates then >> QEMUMachine::reset is not going to be enough because it only gets >> triggered for complete system reset. My suggestion was thus to just call >> cpu_reset() in your QEMUMachine::reset and have cpu_reset() take care of >> its initialization wherever called from. Any of these solutions are easy >> to implement for 1.2 if agreement is reached what people want. > > So, I more or less reaslied that myself and my new version of the > reset patch (which I expect to send out later today) kind of does > that. I no longer do the machine specific CPU state setup from the > QEMUMachine::reset, it's done from the per-cpu reset handler. The > QEMUMachine::reset just does the special setup that's only for the > CPU0 entry conditions, which *is* specific to a full system reset (not > that I think we can get an individual CPU reset on pseries, anyway). > >> What I am missing from Anthony's side is some communication to machine >> maintainers on the course to adopt before applying random patches. Right >> now x86 and ppc are moving into opposite directions and arm, mips, etc. >> maintainers may not even be aware of ongoing changes, and there's a >> pending uc32 machine that should be reviewed in this light. > > So.. having the CPU reset at the top of the tree definitely makes no > sense - if nothing else, *which* cpu when there's more than one.
Maybe let me restate clearly what I am looking for in this discussion: I would like a clear definition of * what is the "normal" case, and * what is the special case. The special case sPAPR seems uncontroversial. So, a bonus would be if we can have a default implementation (of QEMUMachine::reset or whatever we end up doing) so that the average machine does not need to fiddle with reset callbacks in QEMUMachine::init. For example, have a machine_default_reset() as fallback for QEMUMachine::reset == NULL that resets all CPUs (in order of the singly linked list) and then does qemu_devices_reset()? sPAPR would then override that default implementation by specifying its own implementation and we could get rid of reset callbacks in an estimated 70% of QEMUMachine::init. (The less people fiddle at that level the easier to refactor for me.) That could well be a later follow-up to your v2, which looked okay on brief sight. Cheers, Andreas -- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg
