On Tue, Jan 17, 2017 at 12:27:28PM -0700, Jason Gunthorpe wrote: > On Tue, Jan 17, 2017 at 09:58:27AM -0800, Andrey Pronin wrote: > > > Yes, sorry, I should have mentioned that.. Maybe that is too much to > > > fix.. > > > > If we fix sysfs to go through tpm_try_get_ops, then all we can do for > > shutdown is indeed something like > > Maybe yes, I also had at one point a thought to push the read side of > the ops_sem all the way down to the transmit_cmd level... But that > complicates calling shutdown. > > > down_write(&chip->ops_sem); > > if (chip->ops && chip->flags & TPM_CHIP_FLAG_TPM2) > > tpm2_shutdown(chip, TPM2_SU_CLEAR); > > chip->ops = NULL; > > up_write(&chip->ops_sem); > > > > Does that sound like a good plan? > > If we don't want sysfs to increment/decrement the reference count for > > the device, we can still make it go through > > Grabbing the extra kref is harmless.. > > > > I'm confused - doesn't your system reset the TPM when it reboots? > > > Isn't that required so the firmware starts with known PCRs? Doesn't > > > reset trump unorderly shutdown? > > > > > > > That's right, the TPM is reset when the system reboots. However, for > > TPM 2.0, if it resets w/o Shutdown(CLEAR) first, it will detect it > > during Startup, and mark as unorderly shutdown. Shutdown(CLEAR) is > > the signal to the TPM to save its state to nvram and prepare to reset. > > If it was not done, it is possible that something was not saved (e.g. > > the DA counter), and the chip correctly marks it as a potential DA > > problem. > > Okay, that makes sense, and needs to go in a comment someplace! > > > > > All these things are handled by tpm_chip_unregister(). I thought about > > > > creating a tpm_chip_shutdown routine that could be called from shutdown > > > > handlers of the drivers that need it (and I'd do it for every driver, > > > > especially in 2.0 case). But decided that it's better to reuse the > > > > existing tpm_chip_unregister() that already does what's needed. > > > > > > If for some reason we need this for every driver then this is probably > > > a better approach - but that seems very, very strange to me. > > > > As described above, we can do a smaller tpm_chip_shutdown() that the > > drivers that need it (2.0 or susceptible to issues if reset in the > > middle of command) can call. > > I'll do it, if it sounds like the right plan to you. > > Yes please.. > > Is there some way we can have the TPM core do this without requiring > the driver to add a shutdown the struct driver? > > Maybe we could put something in chip->dev->driver? Not sure..
I can play more with it. We can check in tpm_chip_register() if chip->dev->driver->shutdown is NULL, and, if so, set it to a default handler. Or, do register_reboot_notifier() instead, to avoid messing with struct device_driver from tpm-chip.c. Not sure if that's a consideration at alli - any reason not to mess with those structures? In any case, driver->shutdown or register_reboot_notifier, if we still export that same common tpm_shutdown for those drivers that want to do their custom shutdown handlers and register them through module_driver(), we should be ok. Whatever we do, we should allow the drivers to still send (vendor-specific) commands from their shutdown handlers. At some point, we actually used to have a register_reboot_notifier() in the common tpm-chip.c code to make sure that it is done during shutdown. But it is called before .shutdown, so a driver can't do device-specific things with the device (or it can, but through re-implementing the common transfer routines). That's why I switched to a solution where a driver calls this common handler itself, when it is ready for it. Similarly to what's done for tpm_pm_suspend/resume(). But, yes, setting a default handler through chip->dev->driver might just be good enough. > > Jason