On Thu, 2018-07-19 at 12:05 -0700, Tadeusz Struk wrote: > On 07/19/2018 11:47 AM, James Bottomley wrote: > > On Thu, 2018-07-19 at 10:54 -0700, Tadeusz Struk wrote: > > > On 07/19/2018 10:19 AM, James Bottomley wrote: > > > > That's just an implementation, though, what's the use case? > > > > > > Hi James, > > > The use case is described in the TCTI spec [1] in the > > > "3.2.5.2 receive" section. > > > > Well, yes, but I think we've all agreed that the /dev/tpm and > > /dev/tpmrmX aren't TCTI interfaces, although you can layer TCTI on > > top of them, so why not simply do fragmentation on top if you need > > it? > > > > The reason for not doing it in the interface is that it alters the > > ABI. Before this patch we had a hard packet boundary: one packet > > per read, one per write and a -EFAULT if you fail to provide a > > correctly sized buffer. Now if you provide a buffer too small but > > don't know about the fragmentation you're going to misprocess a > > packet (because you think you got a whole reply but you didn't) and > > then you get a -EBUSY on your next command which you don't know how > > to handle. The only way out is to update the applications to > > handle the new behaviour, which is a no-no in Linux ABI terms. > > Don't all the existing applications that read a response in one go > do a 4K read now? So nothing will change for them. They will work > exactly the same with this change as they do without it. > This doesn't break the ABI.
The ABI break is the error case as I outlined above. We can't assume everyone uses the current interface without getting an error and one error and your hosed is a nasty failure case to change the interface to. Plus, if you assume everyone is passing 4k buffers, why would you even need the fragmentation case? > > It might be possible to layer the behaviour you want compatibly > > into the current device structure (say an ioctl to switch to the > > fragment behaviour) but I've got to ask why we'd go to the > > complexity without a use case? > > New IOCTL would add extra complexity, which isn't necessary. So what's wrong with fragmenting in the layer above the device driver (in userspace) and not actually changing the kernel? James