Re: [V9fs-developer] [Qemu-devel] d_off field in struct dirent and 32-on-64 emulation

[Dominique Martinet]

Theodore Y. Ts'o wrote on Fri, Dec 28, 2018:
> On Sat, Dec 29, 2018 at 03:37:21AM +0100, Dominique Martinet wrote:
> > > Are there going to be cases where a process or a thread will sometimes
> > > want the 64-bit interface, and sometimes want the 32-bit interface?
> > > Or is it always going to be one or the other?  I wonder if we could
> > > simply add a new flag to the process personality(2) flags.
> > 
> > That would likely work for qemu user, but the qemu system+9p case is
> > going to be more painful..
> > More precisely, the 9p protocol does not plan for anything other than
> > 64bit offset so if the vfs needs to hand out a 32bit offset we'll need
> > to make a correspondance table between the 32bit offsets we hand off and
> > the 64bit ones to use; unless some flag can be passed at lopen to tell
> > the server to always hand out 32bit offsets for this directory... And if
> > we do that then 9p servers will need a way to use both APIs in parallel
> > for both types of directories.
> 
> How about if we add a fcntl(2) mediated flag, which is tied to a
> struct file?  Would that be more or less painful for 9p and qemu
> system+9p?

Hmm. 9P2000.L doesn't have anything akin to fcntl either, the only two
obvious places where we could pass a flag is lopen (which already
handles a bunch of linux-specific flags, e.g. passing O_LARGEFILE
O_NOATIME etc will just forward these through for qemu/diod at least),
or adding a new parameter to the 9p readdir.

The former would let us get away without modifying the protocol as
servers will just ignore flags they don't handle on implementations I
checked, so it'd definitely be the least effort choice from what I can
tell.


On the other hand a fcntl would solve the server-side problem, it'd
allow the server to request appropriately-sized offsets per fd, so it's
a good start; we "just" need to figure how to translate that on the wire.

-- 
Dominique Martinet | Asmadeus