Re: [Libguestfs] [PATCH v3 12/14] nbd/client: Request extended headers during negotiation

On Wed, May 31, 2023 at 09:33:20PM +0300, Vladimir Sementsov-Ogievskiy wrote: > On 31.05.23 20:54, Eric Blake wrote: >> On Wed, May 31, 2023 at 08:39:53PM +0300, Vladimir Sementsov-Ogievskiy wrote: >>> On 15.05.23 22:53, Eric Blake wrote: >>>> All the pieces are in place for a client to finally request extended >>>> headers. Note that we must not request extended headers when qemu-nbd >>> >>> why must not? It should gracefully report ENOTSUP? Or not? >> >> The kernel code does not yet know how to send extended requests; once >> extended mode is negotiated, sending a simple request requires the > > but how it could be negotiated if kernel doesn't support it? That's the problem. The kernel doesn't do the negotiation, userspace

does. There is an ioctl for the userspace to tell the kernel what flags were advertised as part of the negotiation, but that does not include a flag for extended operation. The kernel ONLY takes care of NBD_CMD_ operations, it does not do NBD_OPT_ operations. So when qemu-nbd is preparing to connect to /dev/nbdN, it first has to negotiate in userspace, avoiding any attempt to use extended headers, then call the right ioctls for the kernel to take over command phase using the older compact headers.

> > I mean if we request extended headers during negotiation with kernel, the kernel will just say "unsupported option", isn't it? If we request extended headers in userspace before calling the ioctl to tell the kernel to start transmission, then the kernel's first command will use the compact style, which the server is not expecting, and while we can hope the server will hang up on the kernel, I didn't test what actually happens. > > Or, in other words, I understand that kernel doesn't support it, I don't understand why you note it here. Is kernel different from other NBD server implementations which doesn't support extended requests at the moment? The kernel is an NBD client, not a server. But if we are about to connect an NBD server over to the kernel for /dev/nbdN, we better make sure the server is not using any features the kernel doesn't support.