Re: [PATCH v5 3/6] accel/kvm: Report the loss of a large memory page

[David Hildenbrand <david@redhat.com>]

> Yes, we can collect the information from the block associated to this
> ram_addr. But instead of duplicating the necessary code into both i386
> and ARM, I came back to adding the change into the
> kvm_hwpoison_page_add() function called from both i386 and ARM specific
> code.
> 
> I also needed a new possibility to retrieve the information while we are
> dealing with the SIGBUS signal, and created a new function to gather the
> information from the RAMBlock:
> qemu_ram_block_location_info_from_addr(ram_addr_t ram_addr,
>                                          struct RAMBlockInfo *b_info)
> with the associated struct.
> 
> So that we can use the RCU_READ_LOCK_GUARD() and retrieve all the data.

Makes sense.

> 
> 
> Note about ARM failing on large pages:
> ----------=====----------------------
> I could test that ARM VMs impacted by memory errors on a large
> underlying memory page, can end up looping on reporting the error:
> The VM encountering an error has a high probability to crash and can try
> to save a vmcore with a kdump phase.

Yeah, that's what I thought. If you rip out 1 GiB of memory, your VM is 
going to have a bad time :/

> 
> This fix introduces qemu messages reporting errors when they are relayed
> to the VM.
> A large page being poisoned by an error on ARM can make a VM loop on the
> vmcore collection phase and the console would show messages like that
> appearing every 10 seconds (before the change):
> 
>    vvv
>            Starting Kdump Vmcore Save Service...
> [    3.095399] kdump[445]: Kdump is using the default log level(3).
> [    3.173998] kdump[481]: saving to
> /sysroot/var/crash/127.0.0.1-2025-01-27-20:17:40/
> [    3.189683] kdump[486]: saving vmcore-dmesg.txt to
> /sysroot/var/crash/127.0.0.1-2025-01-27-20:17:40/
> [    3.213584] kdump[492]: saving vmcore-dmesg.txt complete
> [    3.220295] kdump[494]: saving vmcore
> [   10.029515] EDAC MC0: 1 UE unknown on unknown memory ( page:0x116c60
> offset:0x0 grain:1 - APEI location: )
> [   10.033647] [Firmware Warn]: GHES: Invalid address in generic error
> data: 0x116c60000
> [   10.036974] {2}[Hardware Error]: Hardware error from APEI Generic
> Hardware Error Source: 0
> [   10.040514] {2}[Hardware Error]: event severity: recoverable
> [   10.042911] {2}[Hardware Error]:  Error 0, type: recoverable
> [   10.045310] {2}[Hardware Error]:   section_type: memory error
> [   10.047666] {2}[Hardware Error]:   physical_address: 0x0000000116c60000
> [   10.050486] {2}[Hardware Error]:   error_type: 0, unknown
> [   20.053205] EDAC MC0: 1 UE unknown on unknown memory ( page:0x116c60
> offset:0x0 grain:1 - APEI location: )
> [   20.057416] [Firmware Warn]: GHES: Invalid address in generic error
> data: 0x116c60000
> [   20.060781] {3}[Hardware Error]: Hardware error from APEI Generic
> Hardware Error Source: 0
> [   20.065472] {3}[Hardware Error]: event severity: recoverable
> [   20.067878] {3}[Hardware Error]:  Error 0, type: recoverable
> [   20.070273] {3}[Hardware Error]:   section_type: memory error
> [   20.072686] {3}[Hardware Error]:   physical_address: 0x0000000116c60000
> [   20.075590] {3}[Hardware Error]:   error_type: 0, unknown
>    ^^^
> 
> with the fix, we now have a flood of messages like:
> 
>    vvv
> qemu-system-aarch64: Memory Error on large page from
> ram-node1:d5e00000+0 +200000
> qemu-system-aarch64: Guest Memory Error at QEMU addr 0xffff35c79000 and
> GUEST addr 0x115e79000 of type BUS_MCEERR_AR injected
> qemu-system-aarch64: Memory Error on large page from
> ram-node1:d5e00000+0 +200000
> qemu-system-aarch64: Guest Memory Error at QEMU addr 0xffff35c79000 and
> GUEST addr 0x115e79000 of type BUS_MCEERR_AR injected
> qemu-system-aarch64: Memory Error on large page from
> ram-node1:d5e00000+0 +200000
> qemu-system-aarch64: Guest Memory Error at QEMU addr 0xffff35c79000 and
> GUEST addr 0x115e79000 of type BUS_MCEERR_AR injected
>    ^^^
> 
> 
> In both cases, this situation loops indefinitely !
> 
> I'm just informing of a change of behavior, fixing this issue would most
> probably require VM kernel modifications  or a work-around in qemu when
> errors are reported too often, but is out of the scope of this current
> qemu fix.

Agreed. I think one problem is that kdump cannot really cope with new 
memory errors (it tries to not touch pages that had a memory error in 
the old kernel).

Maybe this is also due to the fact that we inform the kernel only about 
a single page vanishing, whereby actually a whole 1 GiB is vanishing.

-- 
Cheers,

David / dhildenb