[Bug][KVM][Crystal Ridge] Lack of data persistence guarantee of QEMU writes to host PMEM

[quanxian]

Description:
When the host persistent memory is used as the backend of vNVDIMM, QEMU itself may write to the host pmem, e.g., in vNVDIMM label emulation and live migration. In those cases, QEMU needs to take necessary operations (e.g., flush&fence after writes, or store with non-temporal hint) to ensure the data persistence of its own writes to PMEM. Therefore, data on vNVDIMM can be in a consistent status on abnormal termination of QEMU (e.g., QEMU/host kernel crash, host power failure, etc.). Unfortunately, such operations are missed in current QEMU implementation.

Target QEMU Release:3.0.0
Target Release: 19.04

[Christian Ehrhardt  (paelzer)]

Hi,
there is not even a release plan for Qemu 2.12 yet - if following the ususal schedule it might be April - but given the industry effort around recent major security issues it might slightly postpone. That said 2.13 is even after that maybe around August as usual?

With that in mind, I fail to see how to integrate that in time for 18.04.
Are there patches already in a -next tree or do I miss something else that would make that more reasonable for 18.04?

Otherwise I'd think we pick it up for 18.04 via the Ubuntu Cloud Archive that gets released with Qemu 2.13 and made available for 18.04 then - but I didn't plan which release that might be - gut feeling "R" might even be that release if 2.13 makes an August release.

[quanxian (quanxian-wang)]

It is a KVM feature.

[Joseph Salisbury (jsalisbury)]

Can we move this bug to the "Linux" package and make it public?

[quanxian (quanxian-wang)]

it is a bug. We can make it public. Thanks

[quanxian (quanxian-wang)]

All patches merged in qemu v3.0.0:

56eb90a migration/ram: ensure write persistence on loading all data to PMEM.
469dd51 migration/ram: Add check and info message to nvdimm post copy.
faf8a13 mem/nvdimm: ensure write persistence to PMEM in label emulation
a4de855 hostmem-file: add the 'pmem' option
1782440 configure: add libpmem support
cbfc017 memory, exec: switch file ram allocation functions to 'flags' parameters
b0e5de9 memory, exec: Expose all memory block related flags.

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Christian Ehrhardt  (paelzer)]

FYI: I dupped the other bug that came up by Paul onto this.

[Christian Ehrhardt  (paelzer)]

I evaluated the changes here and must postpone them to Ubuntu 19.04
To some extend it is that the changes are actually noisy to backport, but that would appear to be doable.
To some other that the info on the actual changes was made available 1 day after feature freeze and this clearly is a feature to me.
But the biggest reason is that this will make it qemu rely on libpmem which is rather new and not in main.

I must beg your pardon, but we can't shove all of that in right now due to the reasons above.
We plan for >=qemu 3.0 in Ubuntu 19.04 anyway.
Lets use this opportunity to start driving a MIR request for pmdk to be able to do that in the 19.04 cycle.

[Christian Ehrhardt  (paelzer)]

FYI - filed bug 1790856 for the MIR of the pmem library.

[Christian Ehrhardt  (paelzer)]

MIR still blocked, so not including it on the current merge

[quanxian (quanxian-wang)]

1782440 469dd51 56eb90a a4de855 b0e5de9 cbfc017 faf8a13

[Christian Ehrhardt  (paelzer)]

Hi,
The qemu code is in Ubuntu 19.04 as we are on qemu 3.1 (in proposed atm, but soon in proper 19.04.
What is missing is to pass the MIR for the pmem library in bug 1790856 that I referenced in comment #9.

If you want/need this feature for qemu you'd want to track and help driving that MIR bug.
Once that is aaccepteble to be in Main [1]. Last time I checked there it still ahd a lot of churn for the upstream project still having a lot of experimental bits and huge changes each release, as well as some differences between Debian/Ubuntu that made it loose some packaging improvements we have.

TL;DR - Qemu is ok in terms of code, to enable it the MIR for libpmem must pass.

[1]: https://help.ubuntu.com/community/Repositories/Ubuntu

[quanxian (quanxian-wang)]

3.1 is used in 19.04

[quanxian (quanxian-wang)]

close this, and track the bug separately. bug 1790856

[guo xuelian (belindaguo)]

not reproduced on Ubuntu19.04 Beta