On 20 June 2017 at 00:33, Mike Frysinger <vapier@xxxxxxxxxxxx> wrote:
> On Mon, Jun 19, 2017 at 2:43 AM, PrasannaKumar Muralidharan wrote:
>> On 19 June 2017 at 11:51, Herbert Xu wrote:
>>> On Sun, Jun 18, 2017 at 10:00:17PM -0700, Mike Frysinger wrote:
>>>>
>>>> in order to make tpm-rng react in the way you're implying, the TPM
>>>> subsystem would need to add a notification chain for transitions from
>>>> none<->some devices, then tpm-rng could subscribe to that, and during
>>>> those transition points, it would call hwrng_register/hwrng_unregister
>>>> to make itself visible accordingly to the hwrng subsystem. maybe
>>>> someone on the TPM side would be interested in writing all that logic,
>>>> but it sounds excessive for this minor usage. the current tpm-rng
>>>> driver is *extremely* simple -- it's 3 funcs, each of which are 1
>>>> line.
>>>
>>> It's simple and it's broken, as far as the way it hooks into the
>>> hwrng is concerned.
>>
>> *********************************************************************************************
>> diff --git a/drivers/char/hw_random/tpm-rng.c b/drivers/char/hw_random/tpm-rng.c
>> index d6d4482..4861b35 100644
>> --- a/drivers/char/hw_random/tpm-rng.c
>> +++ b/drivers/char/hw_random/tpm-rng.c
>> @@ -22,6 +22,10 @@
>> #include <linux/tpm.h>
>>
>> #define MODULE_NAME "tpm-rng"
>> +#define MAX_RETRIES 30
>> +
>> +static struct delayed_work check_tpm_work;
>> +static int retry_count;
>>
>> static int tpm_rng_read(struct hwrng *rng, void *data, size_t max, bool wait)
>> {
>> @@ -33,9 +37,27 @@ static struct hwrng tpm_rng = {
>> .read = tpm_rng_read,
>> };
>>
>> +static void check_tpm_presence(struct work_struct *work)
>> +{
>> + u8 data = 0;
>> + if (tpm_get_random(TPM_ANY_NUM, &data, 1) > 0) {
>> + hwrng_register(&tpm_rng);
>> + } else {
>> + if (retry_count < MAX_RETRIES) {
>> + retry_count++;
>> + schedule_delayed_work(&check_tpm_work, HZ * 10);
>> + } else {
>> + pr_err("Could not find any TPM chip, not
>> registering rng");
>> + }
>> + }
>> +}
>> +
>> static int __init rng_init(void)
>> {
>> - return hwrng_register(&tpm_rng);
>> + INIT_DELAYED_WORK(&check_tpm_work, check_tpm_presence);
>> + check_tpm_presence(NULL);
>> +
>> + return 0;
>> }
>> module_init(rng_init);
>> *********************************************************************************************
>>
>> Why not something like this? Patch is completely untested. If this
>> idea seems useful I can clean the code but would require help in
>> testing.
>
> first, that's not how deferred device probing works in the kernel.
> drivers shouldn't be doing their own sleeping. but we can ignore that
> because no amount of delay/retries will work -- TPMs can come & go at
> anytime via hotplugging or module loading/unloading. so the only way
> to pull it off would be to do something like what i described --
> extending the tpm framework so that it can signal children to come
> up/go down.
If TPM can come and go then notification or callback is the correct
way to handle this case.
> imo, standing all of that up is over-engineering and not worth the
> effort, so i'm not going to do it. but maybe you can convince some of
> the TPM maintainers it's worthwhile.
Okay.
Thanks,
PrasannaKumar
