Re: [PATCH v2 1/7] exec: Fix for qemu_ram_resize() callback

[David Hildenbrand <david@redhat.com>]

On 06.02.20 12:28, Shameerali Kolothum Thodi wrote:
> 
> 
>> -----Original Message-----
>> From: David Hildenbrand [mailto:david@redhat.com]
>> Sent: 06 February 2020 10:56
>> To: Shameerali Kolothum Thodi <shameerali.kolothum.thodi@huawei.com>;
>> Igor Mammedov <imammedo@redhat.com>
>> Cc: peter.maydell@linaro.org; xiaoguangrong.eric@gmail.com;
>> mst@redhat.com; shannon.zhaosl@gmail.com; qemu-devel@nongnu.org;
>> xuwei (O) <xuwei5@huawei.com>; Linuxarm <linuxarm@huawei.com>;
>> eric.auger@redhat.com; qemu-arm@nongnu.org; lersek@redhat.com
>> Subject: Re: [PATCH v2 1/7] exec: Fix for qemu_ram_resize() callback
> 
> [...]
>  
>>> root@ubuntu:/# cat /dev/pmem
>>> pmem0  pmem1
>>>
>>> From the logs, it looks like the ram_load_precopy() --> qemu_ram_resize() is
>> not
>>> called as length == used_length and both seems to be page aligned values.
>>> And from
>> https://github.com/qemu/qemu/blob/master/migration/ram.c#L3421
>>> qemu_ram_resize() is called with length if length != used_length.
>>
>> Assume on your source, the old size is 12345 bytes. So 16384 aligned up
>> (4 pages).
>>
>> Assume on your target, the new size is 123456 bytes, so 126976 aligned
>> up (31 pages).
>>
>> If you migrate from source to destination, the migration code would
>> resize to 16384, although the "actual size" is 12345. The callback will
>> be called with the aligned size, not the actual size. Same the other way
>> around. That's what's inconsistent IMHO.
> 
> Thanks. You are right. I didn’t consider the case where the target can be
> configured with a larger number of devices than the source. I can replicate
> the scenario now,
> 
> Source:
> 
> fw_cfg_add_file_callback: filename etc/boot-fail-wait size 0x4
> fw_cfg_add_file_callback: filename etc/acpi/nvdimm-mem size 0x1000
> fw_cfg_add_file_callback: filename etc/acpi/tables size 0x6210
> 
> Target:
> ram_load_precopy: Ram blk mem1 length 0x40000000 used_length 0x40000000
> ram_load_precopy: Ram blk virt.flash0 length 0x4000000 used_length 0x4000000
> ram_load_precopy: Ram blk virt.flash1 length 0x4000000 used_length 0x4000000
> ram_load_precopy: Ram blk /rom@etc/acpi/tables length 0x7000 used_length 0x8000
> fw_cfg_modify_file: filename etc/acpi/tables len 0x7000
> 
> Target updates FWCfgEntry with a page aligned size :(. I will look into this and see how
> we can solve this. Any pointers welcome.

Can you look the original value up somehow and us the resize callback
only as a notification that something changed? (that value would have to
be stored somewhere and migrated I assume - maybe that's already being done)

-- 
Thanks,

David / dhildenb