[Qemu-devel] [RFC] Replication agent requirements (was [RFC PATCH] replication agent module)

[Ori Mamluk <omamluk@zerto.com>]

Hi,
Following previous mails from Kevin and Dor, I'd like to specify the 
high level requirements of a replication agent as I see them.

1. Report each write to a protected volume to the rephub, at an IO 
transaction granularity
     * The reporting is not synchronous, i.e. the write completion is 
not delayed until the rephub received it.
     * The IOs have to be the raw guest IOs - i.e. not converted to any 
sparse format or another filter that alters the size/offset
2. Report failures to report an IO (e.g. socket disconnect or send 
timeout) or failed IOs (bad status from storage) to rephub
     * It is enough to disconnect the socket - that can be considered a 
'failure report'
3. Enable rephub to read arbitrary regions in the protected volume
     * Assume that rephub can identify IOs which were dropped by the 
replication system, and needs to re-read the data of these IOs.

We'd like to treat the following requirement as a second stage - not to 
implement it in the first version:
4. Synchronously report IO writes meta data (offset, size) to an 
external API
     * Synchronous meaning that it is reported (blocking) before the IO 
is processed by storage.
     * The goal is to maintain a dirty bitmap outside of the Qemu process
     * The tracking needs to be more persistent than the Qemu process. A 
good example for that is to expose an additional process API
         (yet another NBD??) that will be hold the bitmap by either the 
host RAM or by writing persistently to storage.

The emphasis to report single IO transactions is because high end 
replication (near synchronous) requires access to every IO shortly after 
it is written to the storage.

Thanks,
Ori