On 9/10/20 4:56 PM, Daniel P. Berrangé wrote:> On Thu, Sep 10, 2020 at 04:54:08PM +0200, Milan Zamazal wrote: >> Daniel P. Berrangé <berrange@redhat.com> writes: >> >>> On Thu, Sep 10, 2020 at 04:26:40PM +0200, Milan Zamazal wrote: >>>> Daniel P. Berrangé <berrange@redhat.com> writes: >>>> >>> >>>>> On Thu, Jul 02, 2020 at 01:21:15PM +0200, Milan Zamazal wrote: >>>>>> Hi, >>>>>> >>>>> >>>>>> I've met two situations with NVDIMM support in libvirt where I'm not >>>>>> sure all the parties (libvirt & I) do the things correctly. >>>>>> >>>>>> The first problem is with memory alignment and size changes. In >>>>>> addition to the size changes applied to NVDIMMs by QEMU, libvirt also >>>>>> makes some NVDIMM size changes for better alignments, in >>>>>> qemuDomainMemoryDeviceAlignSize. This can lead to the size being >>>>>> rounded up, exceeding the size of the backing device and QEMU failing to >>>>>> start the VM for that reason (I've experienced that actually). I work >>>>>> with emulated NVDIMM devices, not a bare metal hardware, so one might >>>>>> argue that in practice the device sizes should already be aligned, but >>>>>> I'm not sure it must be always the case considering labels or whatever >>>>>> else the user decides to set up. And I still don't feel very >>>>>> comfortable that I have to count with two internal size adjustments >>>>>> (libvirt & QEMU) to the `size' value I specify, with the ultimate goal >>>>>> of getting the VM started and having the NVDIMM aligned properly to make >>>>>> (non-NVDIMM) memory hot plug working. Is the size alignment performed >>>>>> by libvirt, especially rounding up, completely correct for NVDIMMs? >>>>> >>>>> The comment on the function says QEMU aligns to "page size", which >>>>> is something that can vary depending not only on architecture, and >>>>> also the build config options for the kernel on that architecture. >>>>> eg aarch64 has different page size in RHEL than other distros because >>>>> of different choice of page size in kernel config. >>>>> >>>>> Libvirt rounds up to 1 MB, essentially so that the size works no matter >>>>> what architecture or build options were used. I think this is quite >>>>> compelling as I don't think mgmt apps are likely to care enough about >>>>> non-x86 architectures to pick the right rounded sizes. >>>>> >>>>> If we're enforcing this 1 MB rounding though, we really should be >>>>> documenting it clearly, so that apps can pick the right backing file >>>>> size. I think we dropped the ball on docs. >>>> >>>> I still can't see it in the documentation, would it be possible to be >>>> clear about it in the docs, please? For first, it's not very intuitive >>>> to figure out that (if I've figured out it correctly) on POWER one >>>> *must* specify the NVDIMM size S as >>>> >>>> S == aligned_size + label_size >>>> >>>> and that size is used for the QEMU device; while on x86_64 one can >>>> specify any size S and >>>> >>>> align_up(S) >>>> >>>> will be used for the QEMU device (and label size doesn't influence the >>>> value). And additional alignment may be required for having any memory >>>> hot plug working. >>>> >>>> For second, and more importantly, I'm afraid that without documenting >>>> it, future changes may break the current behavior without warning. For >>>> example, the recent changes regarding POWER alignment in 6.7.0 are for >>>> good IMHO and one can use the same size with both 6.7 and 6.6 versions, >>>> but they could still cause pre-6.7 sizes stop working. >>> >>> I don't know what changes you are referring to here, but if they were >>> in libvirt I'd consider that a bug - we shouldn't break a previously >>> working configuration by increasing required alignment. >> >> I mean disabling the auto alignment in >> https://gitlab.com/libvirt/libvirt/-/commit/07de813924caf37e535855541c0c1183d9d382e2 >> and replacing it with validation in >> https://gitlab.com/libvirt/libvirt/-/commit/0ccceaa57c50e5ee528f7073fa8723afd62b88b7 >> >> That change can cause a VM fail to start but after (manually) adjusting >> the device size, all should work all right. Changes that would actually >> change sizes would be more dangerous. > > Sigh, that second commit even calls out the fact that it breaks > existing guests. This needs to be reverted, as that is not acceptable.Thing is, on PPC it was never working IIRC. I remember discussing this with Andrea. So from my POV, there wasn't really anything to break. Michal
On Thu, 2020-09-10 at 20:53 +0200, Michal Privoznik wrote:> On 9/10/20 4:56 PM, Daniel P. Berrangé wrote: > > On Thu, Sep 10, 2020 at 04:54:08PM +0200, Milan Zamazal wrote: > > > Daniel P. Berrangé <berrange@redhat.com> writes: > > > > > > If we're enforcing this 1 MB rounding though, we really should be > > > > > > documenting it clearly, so that apps can pick the right backing file > > > > > > size. I think we dropped the ball on docs. > > > > > > > > > > I still can't see it in the documentation, would it be possible to be > > > > > clear about it in the docs, please? For first, it's not very intuitive > > > > > to figure out that (if I've figured out it correctly) on POWER one > > > > > *must* specify the NVDIMM size S as > > > > > > > > > > S == aligned_size + label_size > > > > > > > > > > and that size is used for the QEMU device; while on x86_64 one can > > > > > specify any size S and > > > > > > > > > > align_up(S) > > > > > > > > > > will be used for the QEMU device (and label size doesn't influence the > > > > > value). And additional alignment may be required for having any memory > > > > > hot plug working.The ppc64-specific requirements were documented with commit 8f474ceea05aec349be19726e394a62e300efe77 Author: Daniel Henrique Barboza <danielhb413@gmail.com> Date: Mon Jul 20 13:51:46 2020 -0300 formatdomain.html.in: mention pSeries NVDIMM 'align down' mechanic The reason why we align down the guest area (total-size - label-size) is explained in the body of qemuDomainNVDimmAlignSizePseries(). This behavior must also be documented in the user docs. Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com> Reviewed-by: Andrea Bolognani <abologna@redhat.com> but later reverted. See below.> > > > > For second, and more importantly, I'm afraid that without documenting > > > > > it, future changes may break the current behavior without warning. For > > > > > example, the recent changes regarding POWER alignment in 6.7.0 are for > > > > > good IMHO and one can use the same size with both 6.7 and 6.6 versions, > > > > > but they could still cause pre-6.7 sizes stop working. > > > > > > > > I don't know what changes you are referring to here, but if they were > > > > in libvirt I'd consider that a bug - we shouldn't break a previously > > > > working configuration by increasing required alignment. > > > > > > I mean disabling the auto alignment in > > > https://gitlab.com/libvirt/libvirt/-/commit/07de813924caf37e535855541c0c1183d9d382e2 > > > and replacing it with validation in > > > https://gitlab.com/libvirt/libvirt/-/commit/0ccceaa57c50e5ee528f7073fa8723afd62b88b7 > > > > > > That change can cause a VM fail to start but after (manually) adjusting > > > the device size, all should work all right. Changes that would actually > > > change sizes would be more dangerous. > > > > Sigh, that second commit even calls out the fact that it breaks > > existing guests. This needs to be reverted, as that is not acceptable. > > Thing is, on PPC it was never working IIRC. I remember discussing this > with Andrea. So from my POV, there wasn't really anything to break.This is my fault for not keeping a close enough eye on the patch series when it was being posted and reviewed upstream. Sorry :( While it is true that QEMU will reject sizes that are not compliant with the ppc64-specific alignment requirements, libvirt was already aware of this and behaved accordingly: so a configuration like <memory model='nvdimm'> <source> <path>/var/lib/libvirt/images/nvdimm.raw</path> </source> <target> <size unit='KiB'>524288</size> <label> <size unit='KiB'>128</size> </label> </target> </memory> can't work because the usable area (512 MiB - 128 KiB) is not aligned to 256 MiB, however libvirt will align it down and generate a QEMU command line that looks like -object memory-backend-file,[...],size=268566528 -device nvdimm,[...],label-size=131072 where 131072 is 128 KiB and 268566528 is 265 MiB + 128 KiB. In this scenario, assuming the user has allocated a backing file that's 512 MiB in size, ~256 MiB will never be used and will go to waste, but other than that the sizing requirements will be satisfied from QEMU's point of view and the guest will happily start. So ultimately this change caused existing guests to stop working and I agree with Dan that it should be reverted. -- Andrea Bolognani / Red Hat / Virtualization
On 9/14/20 1:40 PM, Andrea Bolognani wrote:> On Thu, 2020-09-10 at 20:53 +0200, Michal Privoznik wrote: >> On 9/10/20 4:56 PM, Daniel P. Berrangé wrote: >>> On Thu, Sep 10, 2020 at 04:54:08PM +0200, Milan Zamazal wrote: >>>> Daniel P. Berrangé <berrange@redhat.com> writes: >>>>>>> If we're enforcing this 1 MB rounding though, we really should be >>>>>>> documenting it clearly, so that apps can pick the right backing file >>>>>>> size. I think we dropped the ball on docs. >>>>>> >>>>>> I still can't see it in the documentation, would it be possible to be >>>>>> clear about it in the docs, please? For first, it's not very intuitive >>>>>> to figure out that (if I've figured out it correctly) on POWER one >>>>>> *must* specify the NVDIMM size S as >>>>>> >>>>>> S == aligned_size + label_size >>>>>> >>>>>> and that size is used for the QEMU device; while on x86_64 one can >>>>>> specify any size S and >>>>>> >>>>>> align_up(S) >>>>>> >>>>>> will be used for the QEMU device (and label size doesn't influence the >>>>>> value). And additional alignment may be required for having any memory >>>>>> hot plug working. > > The ppc64-specific requirements were documented with > > commit 8f474ceea05aec349be19726e394a62e300efe77 > Author: Daniel Henrique Barboza <danielhb413@gmail.com> > Date: Mon Jul 20 13:51:46 2020 -0300 > > formatdomain.html.in: mention pSeries NVDIMM 'align down' mechanic > > The reason why we align down the guest area (total-size - label-size) is > explained in the body of qemuDomainNVDimmAlignSizePseries(). This > behavior must also be documented in the user docs. > > Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com> > Reviewed-by: Andrea Bolognani <abologna@redhat.com> > > but later reverted. See below. > >>>>>> For second, and more importantly, I'm afraid that without documenting >>>>>> it, future changes may break the current behavior without warning. For >>>>>> example, the recent changes regarding POWER alignment in 6.7.0 are for >>>>>> good IMHO and one can use the same size with both 6.7 and 6.6 versions, >>>>>> but they could still cause pre-6.7 sizes stop working. >>>>> >>>>> I don't know what changes you are referring to here, but if they were >>>>> in libvirt I'd consider that a bug - we shouldn't break a previously >>>>> working configuration by increasing required alignment. >>>> >>>> I mean disabling the auto alignment in >>>> https://gitlab.com/libvirt/libvirt/-/commit/07de813924caf37e535855541c0c1183d9d382e2 >>>> and replacing it with validation in >>>> https://gitlab.com/libvirt/libvirt/-/commit/0ccceaa57c50e5ee528f7073fa8723afd62b88b7 >>>> >>>> That change can cause a VM fail to start but after (manually) adjusting >>>> the device size, all should work all right. Changes that would actually >>>> change sizes would be more dangerous. >>> >>> Sigh, that second commit even calls out the fact that it breaks >>> existing guests. This needs to be reverted, as that is not acceptable. >> >> Thing is, on PPC it was never working IIRC. I remember discussing this >> with Andrea. So from my POV, there wasn't really anything to break.Yes, this is correct. However, this is a Libvirt design violation, regardless of whether there are existing guests to break or not, and this is why I have already agreed with the revert and 'll post patches soon.> > This is my fault for not keeping a close enough eye on the patch > series when it was being posted and reviewed upstream. Sorry :(Nah. I'm the one that posted the patches ignoring the fact that this was breaking the intended design. Let's not blame Brno for a Brazilian mess up :P We'ĺl get this reverted, tidy it up what was there before to make the size consistent between what QEMU and domain XML sees (without breaking guests) and get it all wrapped up for the next Libvirt release. Thanks, DHB