On Thu, Jul 9, 2020 at 8:39 AM Jason Gunthorpe <jgg@xxxxxxxxxxxx> wrote: > > On Thu, Jul 09, 2020 at 04:00:51PM +0100, Christoph Hellwig wrote: > > On Mon, Jul 06, 2020 at 06:59:32PM -0700, Dan Williams wrote: > > > The runtime firmware activation capability of Intel NVDIMM devices > > > requires memory transactions to be disabled for 100s of microseconds. > > > This timeout is large enough to cause in-flight DMA to fail and other > > > application detectable timeouts. Arrange for firmware activation to be > > > executed while the system is "quiesced", all processes and device-DMA > > > frozen. > > > > > > It is already required that invoking device ->freeze() callbacks is > > > sufficient to cease DMA. A device that continues memory writes outside > > > of user-direction violates expectations of the PM core to be to > > > establish a coherent hibernation image. > > > > > > That said, RDMA devices are an example of a device that access memory > > > outside of user process direction. > > Are you saying freeze doesn't work for some RDMA drivers? That would > be a driver bug, I think. Right, it's more my hunch than a known bug at this point, but in my experience with testing server class hardware when I've reported a power management bugs I've sometimes got the incredulous response "who suspends / hibernates servers!?". I can drop that comment. Are there protocol timeouts that might need to be adjusted for a 100s of microseconds blip in memory controller response? > The consequences of doing freeze are pretty serious, but it should > still stop DMA. Ok, and there is still the option to race the quiesce if the effects of the freeze are worse than a potential timeout from the quiesce.
