Re: [PATCH v2 28/29] migration: Add direct-io parameter

[Fabiano Rosas]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Tue, Oct 31, 2023 at 04:05:46PM -0300, Fabiano Rosas wrote:
>> Daniel P. Berrangé <berrange@redhat.com> writes:
>> 
>> > On Tue, Oct 31, 2023 at 12:52:41PM -0300, Fabiano Rosas wrote:
>> >> Daniel P. Berrangé <berrange@redhat.com> writes:
>> >> >
>> >> > I guess I'm not seeing the problem still.  A single FD is passed across
>> >> > from libvirt, but QEMU is free to turn that into *many* FDs for its
>> >> > internal use, using dup() and then setting O_DIRECT on as many/few of
>> >> > the dup()d FDs as its wants to.
>> >> 
>> >> The problem is that duplicated FDs share the file status flags. If we
>> >> set O_DIRECT on the multifd channels and the main thread happens to do
>> >> an unaligned write with qemu_file_put* then the filesystem will fail
>> >> that write.
>> >
>> > Doh, I had forgotten that sharing.
>> >
>> > Do we have any synchronization between multifd  channels and the main
>> > thread ?  eg does the main thread wait for RAM sending completion
>> > before carrying on writing other non-RAM data ?
>> 
>> We do have, but the issue with that approach is that there are no rules
>> for adding data into the stream. Anyone could add a qemu_put_* call
>> right in the middle of the section for whatever reason.
>> 
>> That is almost a separate matter due to our current compatibility model
>> being based on capabilities rather than resilience of the stream
>> format. So extraneous data in the stream always causes the migration to
>> break.
>> 
>> But with the O_DIRECT situation we'd be adding another aspect to
>> this. Not only changing the code requires syncing capabilities (as it
>> does today), but it would also require knowing which parts of the stream
>> can be interrupted by new data and which cannot.
>> 
>> So while it would probably work, it's also a little fragile. If QEMU
>> were given 2 FDs or given access to the file, then only the multifd
>> channels would get O_DIRECT and they are guaranteed to not have
>> extraneous unaligned data showing up.
>
> So the problem with add-fd is that when requesting a FD, the monitor
> code masks flags with O_ACCMODE.  What if we extended it such that
> the monitor masked with O_ACCMODE | O_DIRECT.
>
> That would let us pass 1 plain FD and one O_DIRECT fd, and be able
> to ask for each separately by setting O_DIRECT or not.

That would likely work. The usage gets a little more complicated, but
we'd be using fdset as it was intended.

Should we keep the direct-io capability? If the user now needs to set
O_DIRECT and also set the cap, that seems a little redundant. I could
keep O_DIRECT in the flags (when supported) and test after open if we
got the flag set. If it's not set, then we remove O_DIRECT from the
flags and retry.

> Existing users of add-fd are not likely to be affected since none of
> them will be using O_DIRECT.

I had thought of passing a comparison function into
monitor_fdset_dup_fd_add() to avoid affecting existing users. That would
require plumbing it through qemu_open_internal() or moving
monitor_fdset_dup_fd_add() earlier in the stack (probably more
sensible). I'll not worry about that for now though, let's first make
sure the approach you suggested works.