On Tue, 2019-09-03 at 18:56 +0900, Seunghun Han wrote: > Thank you for your notification. I am sorry. I missed it and > misunderstood Jarkko's idea. So, I would like to invite Matthew > Garrett to this thread and attach my opinion on that. The problem is > that command and response buffers are in ACPI NVS area. ACPI NVS area > is saved and restored by drivers/acpi/nvs.c during hibernation, so > command and response buffers in ACPI NVS are also handled by nvs.c > file. However, TPM CRB driver uses the buffers to control a TPM > device, therefore, something may break. > > I agree on that point. To remove uncertainty and find the solution, > I read the threads we discussed and did research about two points, 1) > the race condition and 2) the unexpected behavior of the TPM device. > > 1) The race condition concern comes from unknowing buffer access order > while hibernation. > If nvs.c and TPM CRB driver access the buffers concurrently, the race > condition occurs. Then, we can't know the contents of the buffers > deterministically, and it may occur the failure of TPM device. > However, hibernation_snapshot() function calls dpm_suspend() and > suspend_nvs_save() in order when the system enters into hibernation. > It also calls suspend_nvs_restore() and dpm_resume() in order when the > system exits from hibernation. So, no race condition occurs while > hibernation, and we always guarantee the contents of buffers as we > expect. > > 2) The unexpected behavior of the TPM device. > If nvs.c saves and restores the contents of the TPM CRB buffers while > hibernation, it may occur the unexpected behavior of the TPM device > because the buffers are used to control the TPM device. When the > system entered into hibernation, suspend_nvs_save() saved the command > and response buffers, and they had the last command and response data. > After exiting from hibernation, suspend_nvs_restore() restored the > last command and response data into the buffers and nothing happened. > I realized that they were just buffers. If we want to send a command > to the TPM device, we have to set the CRB_START_INVOKE bit to a > control_start register of a control area. The control area was not in > the ACPI NVS area, so it was not affected by nvs.c file. We can > guarantee the behavior of the TPM device. > > Because of these two reasons, I agreed on Jarkko's idea in > https://lkml.org/lkml/2019/8/29/962 . It seems that removing or > changing regions described in the ACPI table is not natural after > setup. In my view, saving and restoring buffers was OK like other NVS > areas were expected because the buffers were in ACPI NVS area. > > So, I made and sent this patch series. I would like to solve this > AMD's fTPM problem because I have been doing research on TPM and this > problem is critical for me (as you know fTPM doesn't work). If you > have any other concern or advice on the patch I made, please let me > know. Please take time to edit your responses. Nobody will read that properly because it is way too exhausting. A long prose only indicates unclear thoughts in the end. If you know what you are doing, you can put things into nutshell only in few senteces. /Jarkko
