On Thu, Aug 18, 2022 at 04:36:13PM +0100, Richard W.M. Jones wrote:
This is what I pushed - how does it look? I ignored the cases we cannot deal with, except O_WRONLY where I emit a debug message but continue:
https://gitlab.com/nbdkit/nbdkit/-/blob/master/plugins/file/file.c#L572
There's also the case where r == O_WRONLY which the plugin (and NBD) cannot deal with. Not sure what to do about that - error?
Or allow it, but with the caveat that every NBD_CMD_READ will fail. The only reason to special case h->can_write=false for O_RDONLY is because then we don't advertise it to the client; to save them from getting failures on NBD_CMD_WRITE - but that's because it is an easy thing to advertise. Advertising that NBD_CMD_READ will fail is not easy (and less likely to happen in practice), so failing to serve the file is just as viable as serving it and letting every NBD_CMD_READ fail.
That's basically what I did, plus a debug message :-)
What you have committed looks okay to me for handling the common case of O_RDONLY without worrying about the other cases. And if someone tries to play with passing in an fd opened under O_EXEC, they deserve whatever behavior happens naturally from that explicitly unusual corner-case. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org