Date: Thu, 16 Aug 2018 20:43:53 +0000 From: Matt Churchyard <matt.churchyard@userve.net> To: "araujo@freebsd.org" <araujo@freebsd.org> Cc: Allan Jude <allanjude@freebsd.org>, "freebsd-virtualization@freebsd.org" <freebsd-virtualization@freebsd.org>, "Rodney W. Grimes" <freebsd-rwg@pdx.rh.cn85.dnsmgr.net> Subject: Re: Checking bhyve supported features (sysctls) Message-ID: <12D64664-2135-4D60-B534-5DACCB839A08@userve.net> In-Reply-To: <CAOfEmZhFKbMtPC-6pHExSikNOpCgEHGB_TZqdwpKn=uJZ=L2nQ@mail.gmail.com> References: <201808161628.w7GGS52P054505@pdx.rh.CN85.dnsmgr.net> <BC22EE63-357B-47F5-9121-A73B59633FE9@FreeBSD.org>, <CAOfEmZhFKbMtPC-6pHExSikNOpCgEHGB_TZqdwpKn=uJZ=L2nQ@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On 16 Aug 2018, at 19:55, Marcelo Araujo <araujobsdport@gmail.com<mailto:ar= aujobsdport@gmail.com>> wrote: 2018-08-17 0:53 GMT+08:00 Allan Jude <allanjude@freebsd.org<mailto:allanjud= e@freebsd.org>>: On August 16, 2018 5:28:05 PM GMT+01:00, "Rodney W. Grimes" <freebsd-rwg@pd= x.rh.CN85.dnsmgr.net<mailto:freebsd-rwg@pdx.rh.CN85.dnsmgr.net>> wrote: >> >> Text manually wrapped to 80, any broken quoting is my fault - rwg >> >> > > Hello, >> > > >> > > I'm looking for better ways to check for bhyve support / >available >> > > features without trying to scan through dmesg output. >> > >> > >Yes, it would be very good to remove that, as it usually tries >> > >to grep a non-existent file /var/run/dmesg.boot that is not >> > >created until after vm_bhyve has been called from >/usr/local/etc/rc.d >> > >when you have things set to autostartup >in /etc/rc.conf >> > >> > >> > > >> > > I notice that the following 2 sysctl's appear to be set to 1 as >soon >> > > as the vmm module is loaded >> > > >> > > hw.vmm.vmx.initialized: 1 >> > > hw.vmm.vmx.cap.unrestricted_guest: 1 >> > > >> > > Will these be available on both Intel & AMD processors as a way >> > > to determine if the module has loaded successfully and can run >guests? >> > > >> > > I also see the below sysctl related to iommu. >> > > >> > > hw.vmm.iommu.initialized >> > > >> > > Again, will this be set to 1 as soon as the module is loaded if >> > > iommu is supported, or only when it is used? >> > > There also seems to be a vmm.amdvi.enable sysctl. >> > > Would both these need checking or is vmm.iommu enough to >> > > determine support on any processor. >> > >> > >Probalby the safest way for a shell script to decide if bhyve is >> > >up and running is to stat /dev/vmm, if that exists then the >modules >> > >have loaded and initialized and bhyve should be ready to process >guests. >> > >> > Hmm, I don't get /dev/vmm unless I actually have running guests. >> >> I'll investigate that, I was pretty sure that you should get this >> as soon as the vmm.ko module is finished initialzing, but you might >> be right in that it takes a first vm to cause its creation. >> Confirmed, /dev/vmm does not exist until the first vm >> is created. >> >> > >> > >sysctl's mentiond above would be a poor way to make this >determination. >> > >> > It would be nice if sysctls were better documented. >> >> Agreed. >> >> > If vmx.initialized is set once vmm has successfully loaded, I can't >see a better way of checking for bhyve support (assuming it's not Intel >specific). This entry definitely exists and is set to 0 if you load the >module on a non-supported system, and set to 1 as soon as vmm loads on >my Intel test system. >> >> Given its undocumented status you would be relying on an >> undocumented feature that could change in either name or >> behavior, and that is not desirable. >> >> Let me see if I can come up with something else. > >I looked at the code for bhyvectl, bhyveload and >byhve. They do not actually try to decide if vmm >is supported or not, they simply process the error >from a vm_create() or vm_open() call and exit >with an error code if they can not handle it >(some of the code can handle a vm_create failure >if infact we are trying to create a vm that >already exists). > >If you want to maintain full compatibility a similiar >stratergy may be in order. > >Why is it that vm-bhyve specifically needs to know >if the kernel has vmm support or not? >Cant it just be written to handle the errors returned >if the supported functions do not exist? I think the question vm-bhyve wants to answer is: does the CPU have the req= uired features to run a multicore VM. These or similar sysctls do seem to be the correct way to communicate that = support. You are correct! The question in case as I understood was about CPU feature supported, actua= lly vmm(8) knows all this information! Some examples such like CPU with VMX= unrestricted mode support (UG) that is necessary for guest VMs running wit= h multiple vCPU or like VT-d necessary for PCI device passthrough. I have a patch that exposes a sysctl saying what bhyve(8) is capable to run= , however it needs to be polished a bit more to be more informative. I think for third part software like vm-bhyve these information are crucial= as these software can get advantage of these information prior to run a ce= rtain set that will end up in a fail because of a partial CPU support. Best, As mentioned in my first email, it does seem like some of these exist alrea= dy in the way of vmm.vmx.cap.* sysctls. We could look at bhyve output and try to process that, but that seems more = messy if there are sysctls that expose support, especially as the vmm module does seem to = know what features the cpu/hardware supports. vm-bhyve has already forked into the ba= ckground by the time bhyve runs so can't easily provide feedback to the caller other= than through the log file. We do also try and take action in some cases, such as reducing cpu count to= 1 if UG support isn't found, rather than just having bhyve fail. (Of course you could argue= we should just exit with an error and let the user decide if they want to drop the cpu cou= nt to 1. We could just do nothing, let bhyve run and if it falls over people can use= debug mode and see the bhyve output themselves in the log. Just seems useful to be able to= tell users that their hardware doesn't support the features they are trying to use up = front. It still seems to me that vmx.initialised is a reasonable indicator that vm= m has loaded without issue, but it would be useful to have some documented way of checki= ng exactly what virt features the system supports, without just running someth= ing and seeing if it falls over. Matt -- Allan Jude _______________________________________________ freebsd-virtualization@freebsd.org<mailto:freebsd-virtualization@freebsd.or= g> mailing list https://lists.freebsd.org/mailman/listinfo/freebsd-virtualization To unsubscribe, send any mail to "freebsd-virtualization-unsubscribe@freebs= d.org<mailto:freebsd-virtualization-unsubscribe@freebsd.org>" -- -- Marcelo Araujo (__) araujo@FreeBSD.org<mailto:araujo@FreeBSD.org> \\\'',) http://www.FreeBSD.org<http://www.freebsd.org/>; \/ \ ^ Power To Server. .\. /_)
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?12D64664-2135-4D60-B534-5DACCB839A08>