On Thu, 2024-07-25 at 17:04 -0400, Michael S. Tsirkin wrote: > On Thu, Jul 25, 2024 at 10:00:24PM +0100, David Woodhouse wrote: > > On Thu, 2024-07-25 at 16:50 -0400, Michael S. Tsirkin wrote: > > > On Thu, Jul 25, 2024 at 08:35:40PM +0100, David Woodhouse wrote: > > > > On Thu, 2024-07-25 at 12:38 -0400, Michael S. Tsirkin wrote: > > > > > On Thu, Jul 25, 2024 at 04:18:43PM +0100, David Woodhouse wrote: > > > > > > The use case isn't necessarily for all users of gettimeofday(), of > > > > > > course; this is for those applications which *need* precision time. > > > > > > Like distributed databases which rely on timestamps for coherency, and > > > > > > users who get fined millions of dollars when LM messes up their clocks > > > > > > and they put wrong timestamps on financial transactions. > > > > > > > > > > I would however worry that with all this pass through, > > > > > applications have to be coded to each hypervisor or even > > > > > version of the hypervisor. > > > > > > > > Yes, that would be a problem. Which is why I feel it's so important to > > > > harmonise the contents of the shared memory, and I'm implementing it > > > > both QEMU and $DAYJOB, as well as aligning with virtio-rtc. > > > > > > > > > Writing an actual spec for this would be another thing that might help. Potentially, although working over it with our internal clock team and with Peter on virtio-rtc has put us in good shape. I'm confident now that we have something that's viable and extensible enough. > > > > > > > > virtio has been developed with the painful experience that we keep > > > > > making mistakes, or coming up with new needed features, > > > > > and that maintaining forward and backward compatibility > > > > > becomes a whole lot harder than it seems in the beginning. > > > > > > > > Yes. But as you note, this shared memory structure is a userspace ABI > > > > all of its own, so we get to make a completely *different* kind of > > > > mistake :) > > > > > > > > > > > > > So, something I still don't completely understand. > > > Can't the VDSO thing be written to by kernel? > > > Let's say on LM, an interrupt triggers and kernel copies > > > data from a specific device to the VDSO. > > > > > > Is that problematic somehow? I imagine there is a race where > > > userspace reads vdso after lm but before kernel updated > > > vdso - is that the concern? Yes. > > > Then can't we fix it by interrupting all CPUs right after LM? > > > > > > To me that seems like a cleaner approach - we then compartmentalize > > > the ABI issue - kernel has its own ABI against userspace, > > > devices have their own ABI against kernel. > > > It'd mean we need a way to detect that interrupt was sent, > > > maybe yet another counter inside that structure. > > > > > > WDYT? > > > > > > By the way the same idea would work for snapshots - > > > some people wanted to expose that info to userspace, too. Those people included me. I wanted to interrupt all the vCPUs, even the ones which were in userspace at the moment of migration, and have the kernel deal with passing it on to userspace via a different ABI. It ends up being complex and intricate, and requiring a lot of new kernel and userspace support. I gave up on it in the end for snapshots, and didn't go there again for this. By contrast, a driver which merely exposes a page of MMIO space identified by an ACPI device (without even the in-kernel PTP support) could probably be fewer than a hundred lines of code. In an externally- buildable module that goes back as far as RHEL8 or even further, allowing users to just build and use it from their application. > was there supposed to be text here, or did you just like this > so much you decided to repost my mail ;) Hm, weirdness. I've known Evolution get into a state where it sends completely *empty* messages, but I've never seen it eat only my own part before. I had definitely typed responses (along the lines of the above) last time.
<<attachment: smime.p7s>>
