On Tue, Dec 10, 2024 at 09:55:41AM -0500, James Bottomley wrote: > On Tue, 2024-12-10 at 10:40 -0400, Jason Gunthorpe wrote: > > On Tue, Dec 10, 2024 at 03:34:23PM +0100, Stefano Garzarella wrote: > > > > > + if (platform_device_add_data(&tpm_device, &pops, > > > sizeof(pops))) > > > + return -ENODEV; > > > + if (platform_device_register(&tpm_device)) > > > + return -ENODEV; > > > > This seems like an old fashioned way to instantiate a device. Why do > > this? Just put the TPM driver here and forget about pops? Simple tpm > > drivers are not very complex. > > This driver may be for the AMD SEV SVSM vTPM module, but there are > other platforms where there's an internal vTPM which might be contacted > via a platform specific enlightenment (Intel SNP and Microsoft > OpenHCL). Sure, that's what TPM drivers are for, give those platforms TPM drivers too. Why put a mini driver hidden under an already mini driver? > This separation of the platform device from the contact > mechanism is designed to eliminate the duplication of having a platform > device within each implementation and to make any bugs in the mssim > protocol centrally fixable (every vTPM currently speaks this). That makes sense, but that isn't really what I see in this series? Patch one just has tpm_class_ops send() invoke pops sendrcv() after re-arranging the arguments? It looks to me like there would be mert in adding a new op to tpm_class_ops for the send/recv type operating mode and have the core code manage the buffer singleton (is a global static even *correct*??) After that, there is no meaningful shared code here, and maybe the TPM_CHIP_FLAG_IRQ hack can be avoided too. Simply call tpm_chip_alloc/register from the sev code directly and provide an op that does the send/recv. Let the tpm core code deal with everything else. It is much cleaner than platform devices and driver data.. Jason
