On 2012年12月15日 04:47, Eric Blake wrote:
On 12/14/2012 01:33 PM, Eric Blake wrote:
On 12/13/2012 12:05 PM, Osier Yang wrote:
Since "rawio" and "cdbfilter" are only valid for "lun", this groups them together; And since both of them intend to allow the unprivledged user to use the SG_IO commands, they must be
s/unprivledged/unprivileged/
rawio cdbfilter result missing missing kernel prevents SG_IO missing no kernel prevents SG_IO missing yes SG_IO allowed for disk no missing kernel prevents SG_IO no no kernel prevents SG_IO no yes SG_IO allowed for disk yes missing SG_IO allowed for process yes no SG_IO allowed for process yes yes error
This table shows another reason why I don't like your naming - the defaults are screwy, when an omitted rawio means 'no', but an omitted cdbfilter means 'yes'. Corrected, the table looks like:
rawio cdbfilter result missing missing kernel prevents SG_IO missing no SG_IO allowed for disk missing yes kernel prevents SG_IO no missing kernel prevents SG_IO no no error? or SG_IO allowed for disk? no yes error? or kernel prevents SG_IO? yes missing SG_IO allowed for process yes no error yes yes error? or kernel prevents SG_IO?
Why not simplify things, and have a single attribute rawio, with multiple values?
rawio result missing kernel prevents SG_IO no kernel prevents SG_IO yes SG_IO allowed for process cdb SG_IO allowed for disk
where we document that 'yes' works on more kernel versions than 'cdb', but that 'cdb' (new to 1.0.1) is more secure.
Oh, another thing, what does 'cdb' mean? It happens to be an acronym that matches the kernel implementation,
Yes, it's abstraction of "Command Descriptor Block", see: http://en.wikipedia.org/wiki/SCSI_CDB
but it doesn't convey much meaning on its own. Given that our existing rawio='yes' merely turns on SG_IO via a capability for the entire process,
However, the "rawio" tag is designed for disk-basis. Though current qemu implementation turns on the capability for the whole process, it's just a qemu implementation. And it's still possible to hack 'rawio' to enable the SG_IO capabilty for single disk in future. This is Daniel's thought, and I agreed with it. See: https://www.redhat.com/archives/libvir-list/2012-November/msg01077.html
and the new filtering method enables per-disk whitelisting of which disks get to use SG_IO passthrough, would it be any better to document things as:
no|yes|disk
As explained above, "rawio" tag is already designed for disk-basic, and the underlying implementation for "rawio" and the new sysfs knob is complete different; And as far as I get from the kernel patches, I guess it could only filter sub-set of the commands in future, not simply filtering all or not filtering all. That means we probably will have to add new values in future. For above reasons, I'm afraid that mixing them up will lead us into a mess. So I'm inclined to use a new tag, and document it as a different approach with "rawio" to enable SG_IO. For the tag "name", I don't have better idea except "cdbfilter", though I agreed that it doesn't convey meaning much on its own.
and where we additionally allow 'process' as a synonym for 'yes' on input (for back-compat, we have to output 'yes', not 'process'). That is, I don't know that exposing the term 'cdb' buys us any understanding into what is really going on.
Thanks, Osier