[PATCH] nvme: introduce passthrough ioctl for multipath

[Keith Busch]

On Tue, Feb 16, 2021 at 06:51:47PM +0900, Minwoo Im wrote:
> On 21-02-15 09:02:33, Keith Busch wrote:
> > On Sun, Feb 14, 2021 at 08:01:25PM +0900, Minwoo Im wrote:
> > > We don't allow NVME_IOCTL_IO_CMD ioctl in case that a controller has
> > > multiple namespaces attached.  Also, I/O request to the controller
> > > character device has not been recommended and deprecated because we have
> > > block device to I/O with where the multipath consideration is taken.
> > > 
> > > Once kernel decided a path to I/O for a namespace based on the I/O
> > > policy of a NVMe subsystem, userspace is not allowed to choose a path to
> > > I/O.  If a path is broken(inaccessible state in ANA), then it will not
> > > try to I/O to that path.
> > > 
> > > This patch introduced NVME_IOCTL_MPATH_IO command for controller
> > > device(e.g., /dev/nvme0) to support multipath I/O passthrough for
> > > userspace.  Regardless driver's path decision, userspace can target a
> > > namespace to I/O.  In this case, `cmd.nsid` will be used to find out the
> > > namespace instance target which is hidden(e.g., nvmeXcYnZ).
> > 
> > IO commands are not allowed through the character handle with the
> > existing ioctls. A new ioctl doesn't make it okay. If it was okay, then
> > we could just remove the limitation in the current ones.
> 
> Thanks for your feedback, Keith.  If you don't mind, may I ask why it's
> been entirely unsafe and deprecated in the exsiting
> ioctl(NVME_IOCTL_IO_CMD)?  I've seen a patch bfd8947194b2 ("nvme: fixes
> for NVME_IOCTL_IO_CMD on the char device), but have no idea why it's
> been really depreacted :)

A container could read data from namespaces assigned to a different
container.