Patchwork [v2,1/4] tpm: ignore burstcount to improve tpm_tis send() performance.

login
register
mail settings
Submitter Nayna Jain
Date Sept. 6, 2017, 12:56 p.m.
Message ID <20170906125643.5070-2-nayna@linux.vnet.ibm.com>
Download mbox | patch
Permalink /patch/332965/
State New
Headers show

Comments

Nayna Jain - Sept. 6, 2017, 12:56 p.m.
The TPM burstcount status indicates the number of bytes that can
be sent to the TPM without causing bus wait states.  Effectively,
it is the number of empty bytes in the command FIFO. Further,
some TPMs have a static burstcount, when the value remains zero
until the entire FIFO is empty.

This patch adds an optimization to check for burstcount only once.
And if it is valid, it writes all the bytes at once, permitting
wait states. The performance of a 34 byte extend on a TPM 1.2 with
an 8 byte burstcount improved from 41 msec to 14 msec.

This functionality is enabled only by passing module
parameter ignore_burst_count=1. By default, this parameter
is disabled.

After this change, performance on a TPM 1.2 with an 8 byte
burstcount for 1000 extends improved from ~41sec to ~14sec.

Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
conjunction with the TPM Device Driver work group.
Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
---
 Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
 drivers/char/tpm/tpm_tis_core.c                 | 24 +++++++++++++++++++++---
 2 files changed, 29 insertions(+), 3 deletions(-)
Jason Gunthorpe - Sept. 6, 2017, 4:12 p.m.
On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:

> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index 4e303be83df6..3c59bb91e1ee 100644
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -1465,6 +1465,14 @@
>  			mode generally follows that for the NaN encoding,
>  			except where unsupported by hardware.
>  
> +	ignore_burst_count [TPM_TIS_CORE]
> +			tpm_tis_core driver queries for the burstcount before
> +			every send call in a loop. However, it causes delay to
> +			the send command for TPMs with low burstcount value.
> +			Setting this value to 1, will make driver to query for
> +			burstcount only once in the loop to improve the
> +			performance. By default, its value is set to 0.

Really don't want to see a kernel command line parameter for this..

Please figure out a different approach or at least a better name..

> +		/*
> +		 * Get the initial burstcount to ensure TPM is ready to
> +		 * accept data, even when waiting for burstcount is disabled.
> +		 */
>  		burstcnt = get_burstcount(chip);
>  		if (burstcnt < 0) {
>  			dev_err(&chip->dev, "Unable to read burstcount\n");
>  			rc = burstcnt;
>  			goto out_err;
>  		}
> -		burstcnt = min_t(int, burstcnt, len - count - 1);
> +
> +		if (ignore_burst_count)
> +			sendcnt = len - 1;
> +		else
> +			sendcnt = min_t(int, burstcnt, len - count - 1);
> +
>  		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
> -					 burstcnt, buf + count);
> +					 sendcnt, buf + count);

The problem with this approach is that the TPM could totally block the CPU for
very long periods of time.

It seems very risky to enable..

Jason

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Jarkko Sakkinen - Sept. 13, 2017, 12:45 a.m.
On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:
> The TPM burstcount status indicates the number of bytes that can
> be sent to the TPM without causing bus wait states.  Effectively,
> it is the number of empty bytes in the command FIFO. Further,
> some TPMs have a static burstcount, when the value remains zero
> until the entire FIFO is empty.
> 
> This patch adds an optimization to check for burstcount only once.
> And if it is valid, it writes all the bytes at once, permitting
> wait states. The performance of a 34 byte extend on a TPM 1.2 with
> an 8 byte burstcount improved from 41 msec to 14 msec.
> 
> This functionality is enabled only by passing module
> parameter ignore_burst_count=1. By default, this parameter
> is disabled.
> 
> After this change, performance on a TPM 1.2 with an 8 byte
> burstcount for 1000 extends improved from ~41sec to ~14sec.
> 
> Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
> conjunction with the TPM Device Driver work group.
> Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
> Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
> ---
>  Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
>  drivers/char/tpm/tpm_tis_core.c                 | 24 +++++++++++++++++++++---
>  2 files changed, 29 insertions(+), 3 deletions(-)
> 
> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> index 4e303be83df6..3c59bb91e1ee 100644
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -1465,6 +1465,14 @@
>  			mode generally follows that for the NaN encoding,
>  			except where unsupported by hardware.
>  
> +	ignore_burst_count [TPM_TIS_CORE]
> +			tpm_tis_core driver queries for the burstcount before
> +			every send call in a loop. However, it causes delay to
> +			the send command for TPMs with low burstcount value.
> +			Setting this value to 1, will make driver to query for
> +			burstcount only once in the loop to improve the
> +			performance. By default, its value is set to 0.
> +
>  	ignore_loglevel	[KNL]
>  			Ignore loglevel setting - this will print /all/
>  			kernel messages to the console. Useful for debugging.
> diff --git a/drivers/char/tpm/tpm_tis_core.c b/drivers/char/tpm/tpm_tis_core.c
> index 63bc6c3b949e..6b9bf4c4d434 100644
> --- a/drivers/char/tpm/tpm_tis_core.c
> +++ b/drivers/char/tpm/tpm_tis_core.c
> @@ -31,6 +31,11 @@
>  #include "tpm.h"
>  #include "tpm_tis_core.h"
>  
> +static bool ignore_burst_count = false;
> +module_param(ignore_burst_count, bool, 0444);
> +MODULE_PARM_DESC(ignore_burst_count,
> +		"Ignore burstcount value while writing data");
> +
>  /* Before we attempt to access the TPM we must see that the valid bit is set.
>   * The specification says that this bit is 0 at reset and remains 0 until the
>   * 'TPM has gone through its self test and initialization and has established
> @@ -256,6 +261,7 @@ static int tpm_tis_send_data(struct tpm_chip *chip, u8 *buf, size_t len)
>  {
>  	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
>  	int rc, status, burstcnt;
> +	int sendcnt;
>  	size_t count = 0;
>  	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
>  
> @@ -271,19 +277,31 @@ static int tpm_tis_send_data(struct tpm_chip *chip, u8 *buf, size_t len)
>  	}
>  
>  	while (count < len - 1) {
> +
> +		/*
> +		 * Get the initial burstcount to ensure TPM is ready to
> +		 * accept data, even when waiting for burstcount is disabled.
> +		 */
>  		burstcnt = get_burstcount(chip);
>  		if (burstcnt < 0) {
>  			dev_err(&chip->dev, "Unable to read burstcount\n");
>  			rc = burstcnt;
>  			goto out_err;
>  		}
> -		burstcnt = min_t(int, burstcnt, len - count - 1);
> +
> +		if (ignore_burst_count)
> +			sendcnt = len - 1;
> +		else
> +			sendcnt = min_t(int, burstcnt, len - count - 1);
> +
>  		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
> -					 burstcnt, buf + count);
> +					 sendcnt, buf + count);
>  		if (rc < 0)
>  			goto out_err;
>  
> -		count += burstcnt;
> +		count += sendcnt;
> +		if (ignore_burst_count)
> +			continue;
>  
>  		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
>  					&priv->int_queue, false) < 0) {
> -- 
> 2.13.3
> 

Makes sense to discuss whether to have the kernel command-line
parameter or not before applying this.

To fuel the discussion, alternative to this would be:

1. Have this always on i.e. no command-line parameter.
2. If someone yells, we add the command-line parameter later on.

/Jarkko

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Peter Huewe - Sept. 13, 2017, 6:39 p.m.
Am 12. September 2017 17:45:08 GMT-07:00 schrieb Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>:
>On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:
>> The TPM burstcount status indicates the number of bytes that can
>> be sent to the TPM without causing bus wait states.  Effectively,
>> it is the number of empty bytes in the command FIFO. Further,
>> some TPMs have a static burstcount, when the value remains zero
>> until the entire FIFO is empty.
>> 
>> This patch adds an optimization to check for burstcount only once.
>> And if it is valid, it writes all the bytes at once, permitting
>> wait states. The performance of a 34 byte extend on a TPM 1.2 with
>> an 8 byte burstcount improved from 41 msec to 14 msec.
>> 
>> This functionality is enabled only by passing module
>> parameter ignore_burst_count=1. By default, this parameter
>> is disabled.
>> 
>> After this change, performance on a TPM 1.2 with an 8 byte
>> burstcount for 1000 extends improved from ~41sec to ~14sec.
>> 
>> Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
>> conjunction with the TPM Device Driver work group.
>> Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
>> Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
>> ---
>>  Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
>>  drivers/char/tpm/tpm_tis_core.c                 | 24
>+++++++++++++++++++++---
>>  2 files changed, 29 insertions(+), 3 deletions(-)
>> 
>> diff --git a/Documentation/admin-guide/kernel-parameters.txt
>b/Documentation/admin-guide/kernel-parameters.txt
>> index 4e303be83df6..3c59bb91e1ee 100644
>> --- a/Documentation/admin-guide/kernel-parameters.txt
>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>> @@ -1465,6 +1465,14 @@
>>  			mode generally follows that for the NaN encoding,
>>  			except where unsupported by hardware.
>>  
>> +	ignore_burst_count [TPM_TIS_CORE]
>> +			tpm_tis_core driver queries for the burstcount before
>> +			every send call in a loop. However, it causes delay to
>> +			the send command for TPMs with low burstcount value.
>> +			Setting this value to 1, will make driver to query for
>> +			burstcount only once in the loop to improve the
>> +			performance. By default, its value is set to 0.
>> +
>>  	ignore_loglevel	[KNL]
>>  			Ignore loglevel setting - this will print /all/
>>  			kernel messages to the console. Useful for debugging.
>> diff --git a/drivers/char/tpm/tpm_tis_core.c
>b/drivers/char/tpm/tpm_tis_core.c
>> index 63bc6c3b949e..6b9bf4c4d434 100644
>> --- a/drivers/char/tpm/tpm_tis_core.c
>> +++ b/drivers/char/tpm/tpm_tis_core.c
>> @@ -31,6 +31,11 @@
>>  #include "tpm.h"
>>  #include "tpm_tis_core.h"
>>  
>> +static bool ignore_burst_count = false;
>> +module_param(ignore_burst_count, bool, 0444);
>> +MODULE_PARM_DESC(ignore_burst_count,
>> +		"Ignore burstcount value while writing data");
>> +
>>  /* Before we attempt to access the TPM we must see that the valid
>bit is set.
>>   * The specification says that this bit is 0 at reset and remains 0
>until the
>>   * 'TPM has gone through its self test and initialization and has
>established
>> @@ -256,6 +261,7 @@ static int tpm_tis_send_data(struct tpm_chip
>*chip, u8 *buf, size_t len)
>>  {
>>  	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
>>  	int rc, status, burstcnt;
>> +	int sendcnt;
>>  	size_t count = 0;
>>  	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
>>  
>> @@ -271,19 +277,31 @@ static int tpm_tis_send_data(struct tpm_chip
>*chip, u8 *buf, size_t len)
>>  	}
>>  
>>  	while (count < len - 1) {
>> +
>> +		/*
>> +		 * Get the initial burstcount to ensure TPM is ready to
>> +		 * accept data, even when waiting for burstcount is disabled.
>> +		 */
>>  		burstcnt = get_burstcount(chip);
>>  		if (burstcnt < 0) {
>>  			dev_err(&chip->dev, "Unable to read burstcount\n");
>>  			rc = burstcnt;
>>  			goto out_err;
>>  		}
>> -		burstcnt = min_t(int, burstcnt, len - count - 1);
>> +
>> +		if (ignore_burst_count)
>> +			sendcnt = len - 1;
>> +		else
>> +			sendcnt = min_t(int, burstcnt, len - count - 1);
>> +
>>  		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
>> -					 burstcnt, buf + count);
>> +					 sendcnt, buf + count);
>>  		if (rc < 0)
>>  			goto out_err;
>>  
>> -		count += burstcnt;
>> +		count += sendcnt;
>> +		if (ignore_burst_count)
>> +			continue;
>>  
>>  		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
>>  					&priv->int_queue, false) < 0) {
>> -- 
>> 2.13.3
>> 
>
>Makes sense to discuss whether to have the kernel command-line
>parameter or not before applying this.
>
>To fuel the discussion, alternative to this would be:
>
>1. Have this always on i.e. no command-line parameter.
>2. If someone yells, we add the command-line parameter later on.
>
According to what I've read in the tcg ddwg group this patch should not cause problems on _sane_ tpms.

I'm not 100%convinced that all tpms are sane all the time, but I think we do not want yet another cmdline parameter.

So if we want to pull it in (and ddwg does not see an issue, so yes) it should be on by default, without a kernel parameter.

If there is a kernel parameter, then it should only be one called "failsafe" - which includes the force behavior and maybe the "broken" tpm path.

But I agree with Alex, every additonal code path reduces testing coverage.


We would be happy to test a "default on" patch.

Peter

>/Jarkko
Ken Goldman - Sept. 13, 2017, 6:52 p.m.
On 9/6/2017 12:12 PM, Jason Gunthorpe wrote:
> 
> The problem with this approach is that the TPM could totally block
> the CPU for very long periods of time.
> 
> It seems very risky to enable..
> 

How would you characterize "very long"?

The TPM vendors confirm that they empty the FIFO at internal speeds that
are comparable to the bus speed.  Thus, any stall will be sub-usec.  Is
that an issue?

In addition, new TPMs have ever larger FIFO's, making stalls less likely
going forward.


------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Peter Huewe - Sept. 13, 2017, 7:01 p.m.
Am 13. September 2017 11:52:12 GMT-07:00 schrieb Ken Goldman <kgold@linux.vnet.ibm.com>:
>On 9/6/2017 12:12 PM, Jason Gunthorpe wrote:
>> 
>> The problem with this approach is that the TPM could totally block
>> the CPU for very long periods of time.
>> 
>> It seems very risky to enable..
>> 
>
>How would you characterize "very long"?
>
>The TPM vendors confirm that they empty the FIFO at internal speeds
>that
>are comparable to the bus speed.  Thus, any stall will be sub-usec.  Is
>that an issue?

If the tpm does behave correctly, this is fine. 
If the tpm hangs for whatever reason, your machine is frozen and you will never figure out why.

That's my concern there.
However ddwg seems fine.

>
>In addition, new TPMs have ever larger FIFO's, making stalls less
>likely
>going forward.
But also reduced the polling loops that introduce the performance penalty ;)

>
>
>------------------------------------------------------------------------------
>Check out the vibrant tech community on one of the world's most
>engaging tech sites, Slashdot.org! http://sdm.link/slashdot
>_______________________________________________
>tpmdd-devel mailing list
>tpmdd-devel@lists.sourceforge.net
>https://lists.sourceforge.net/lists/listinfo/tpmdd-devel
Jarkko Sakkinen - Sept. 13, 2017, 11:10 p.m.
On Wed, Sep 13, 2017 at 11:39:03AM -0700, Peter Huewe wrote:
> 
> 
> Am 12. September 2017 17:45:08 GMT-07:00 schrieb Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>:
> >On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:
> >> The TPM burstcount status indicates the number of bytes that can
> >> be sent to the TPM without causing bus wait states.  Effectively,
> >> it is the number of empty bytes in the command FIFO. Further,
> >> some TPMs have a static burstcount, when the value remains zero
> >> until the entire FIFO is empty.
> >> 
> >> This patch adds an optimization to check for burstcount only once.
> >> And if it is valid, it writes all the bytes at once, permitting
> >> wait states. The performance of a 34 byte extend on a TPM 1.2 with
> >> an 8 byte burstcount improved from 41 msec to 14 msec.
> >> 
> >> This functionality is enabled only by passing module
> >> parameter ignore_burst_count=1. By default, this parameter
> >> is disabled.
> >> 
> >> After this change, performance on a TPM 1.2 with an 8 byte
> >> burstcount for 1000 extends improved from ~41sec to ~14sec.
> >> 
> >> Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
> >> conjunction with the TPM Device Driver work group.
> >> Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
> >> Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
> >> ---
> >>  Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
> >>  drivers/char/tpm/tpm_tis_core.c                 | 24
> >+++++++++++++++++++++---
> >>  2 files changed, 29 insertions(+), 3 deletions(-)
> >> 
> >> diff --git a/Documentation/admin-guide/kernel-parameters.txt
> >b/Documentation/admin-guide/kernel-parameters.txt
> >> index 4e303be83df6..3c59bb91e1ee 100644
> >> --- a/Documentation/admin-guide/kernel-parameters.txt
> >> +++ b/Documentation/admin-guide/kernel-parameters.txt
> >> @@ -1465,6 +1465,14 @@
> >>  			mode generally follows that for the NaN encoding,
> >>  			except where unsupported by hardware.
> >>  
> >> +	ignore_burst_count [TPM_TIS_CORE]
> >> +			tpm_tis_core driver queries for the burstcount before
> >> +			every send call in a loop. However, it causes delay to
> >> +			the send command for TPMs with low burstcount value.
> >> +			Setting this value to 1, will make driver to query for
> >> +			burstcount only once in the loop to improve the
> >> +			performance. By default, its value is set to 0.
> >> +
> >>  	ignore_loglevel	[KNL]
> >>  			Ignore loglevel setting - this will print /all/
> >>  			kernel messages to the console. Useful for debugging.
> >> diff --git a/drivers/char/tpm/tpm_tis_core.c
> >b/drivers/char/tpm/tpm_tis_core.c
> >> index 63bc6c3b949e..6b9bf4c4d434 100644
> >> --- a/drivers/char/tpm/tpm_tis_core.c
> >> +++ b/drivers/char/tpm/tpm_tis_core.c
> >> @@ -31,6 +31,11 @@
> >>  #include "tpm.h"
> >>  #include "tpm_tis_core.h"
> >>  
> >> +static bool ignore_burst_count = false;
> >> +module_param(ignore_burst_count, bool, 0444);
> >> +MODULE_PARM_DESC(ignore_burst_count,
> >> +		"Ignore burstcount value while writing data");
> >> +
> >>  /* Before we attempt to access the TPM we must see that the valid
> >bit is set.
> >>   * The specification says that this bit is 0 at reset and remains 0
> >until the
> >>   * 'TPM has gone through its self test and initialization and has
> >established
> >> @@ -256,6 +261,7 @@ static int tpm_tis_send_data(struct tpm_chip
> >*chip, u8 *buf, size_t len)
> >>  {
> >>  	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> >>  	int rc, status, burstcnt;
> >> +	int sendcnt;
> >>  	size_t count = 0;
> >>  	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
> >>  
> >> @@ -271,19 +277,31 @@ static int tpm_tis_send_data(struct tpm_chip
> >*chip, u8 *buf, size_t len)
> >>  	}
> >>  
> >>  	while (count < len - 1) {
> >> +
> >> +		/*
> >> +		 * Get the initial burstcount to ensure TPM is ready to
> >> +		 * accept data, even when waiting for burstcount is disabled.
> >> +		 */
> >>  		burstcnt = get_burstcount(chip);
> >>  		if (burstcnt < 0) {
> >>  			dev_err(&chip->dev, "Unable to read burstcount\n");
> >>  			rc = burstcnt;
> >>  			goto out_err;
> >>  		}
> >> -		burstcnt = min_t(int, burstcnt, len - count - 1);
> >> +
> >> +		if (ignore_burst_count)
> >> +			sendcnt = len - 1;
> >> +		else
> >> +			sendcnt = min_t(int, burstcnt, len - count - 1);
> >> +
> >>  		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
> >> -					 burstcnt, buf + count);
> >> +					 sendcnt, buf + count);
> >>  		if (rc < 0)
> >>  			goto out_err;
> >>  
> >> -		count += burstcnt;
> >> +		count += sendcnt;
> >> +		if (ignore_burst_count)
> >> +			continue;
> >>  
> >>  		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
> >>  					&priv->int_queue, false) < 0) {
> >> -- 
> >> 2.13.3
> >> 
> >
> >Makes sense to discuss whether to have the kernel command-line
> >parameter or not before applying this.
> >
> >To fuel the discussion, alternative to this would be:
> >
> >1. Have this always on i.e. no command-line parameter.
> >2. If someone yells, we add the command-line parameter later on.
> >
> According to what I've read in the tcg ddwg group this patch should
> not cause problems on _sane_ tpms.
> 
> I'm not 100%convinced that all tpms are sane all the time, but I think
> we do not want yet another cmdline parameter.
> 
> So if we want to pull it in (and ddwg does not see an issue, so yes)
> it should be on by default, without a kernel parameter.
> 
> If there is a kernel parameter, then it should only be one called
> "failsafe" - which includes the force behavior and maybe the "broken"
> tpm path.
> 
> But I agree with Alex, every additonal code path reduces testing coverage.
> 
> 
> We would be happy to test a "default on" patch.
> 
> Peter
> 
> >/Jarkko

I'm starting to dilate to this direction.

It is hard to believe that any such TPM would be in active use anywhere
assuming that there exist a TPM where this causes issues. This combined
to the assumption that you would run the latest mainline on it makes it
a pretty insignificant scenario.

/Jarkko

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Nayna Jain - Sept. 15, 2017, 12:29 p.m.
On 09/14/2017 04:40 AM, Jarkko Sakkinen wrote:
> On Wed, Sep 13, 2017 at 11:39:03AM -0700, Peter Huewe wrote:
>>
>> Am 12. September 2017 17:45:08 GMT-07:00 schrieb Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>:
>>> On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:
>>>> The TPM burstcount status indicates the number of bytes that can
>>>> be sent to the TPM without causing bus wait states.  Effectively,
>>>> it is the number of empty bytes in the command FIFO. Further,
>>>> some TPMs have a static burstcount, when the value remains zero
>>>> until the entire FIFO is empty.
>>>>
>>>> This patch adds an optimization to check for burstcount only once.
>>>> And if it is valid, it writes all the bytes at once, permitting
>>>> wait states. The performance of a 34 byte extend on a TPM 1.2 with
>>>> an 8 byte burstcount improved from 41 msec to 14 msec.
>>>>
>>>> This functionality is enabled only by passing module
>>>> parameter ignore_burst_count=1. By default, this parameter
>>>> is disabled.
>>>>
>>>> After this change, performance on a TPM 1.2 with an 8 byte
>>>> burstcount for 1000 extends improved from ~41sec to ~14sec.
>>>>
>>>> Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
>>>> conjunction with the TPM Device Driver work group.
>>>> Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
>>>> Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
>>>> ---
>>>>   Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
>>>>   drivers/char/tpm/tpm_tis_core.c                 | 24
>>> +++++++++++++++++++++---
>>>>   2 files changed, 29 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/Documentation/admin-guide/kernel-parameters.txt
>>> b/Documentation/admin-guide/kernel-parameters.txt
>>>> index 4e303be83df6..3c59bb91e1ee 100644
>>>> --- a/Documentation/admin-guide/kernel-parameters.txt
>>>> +++ b/Documentation/admin-guide/kernel-parameters.txt
>>>> @@ -1465,6 +1465,14 @@
>>>>   			mode generally follows that for the NaN encoding,
>>>>   			except where unsupported by hardware.
>>>>   
>>>> +	ignore_burst_count [TPM_TIS_CORE]
>>>> +			tpm_tis_core driver queries for the burstcount before
>>>> +			every send call in a loop. However, it causes delay to
>>>> +			the send command for TPMs with low burstcount value.
>>>> +			Setting this value to 1, will make driver to query for
>>>> +			burstcount only once in the loop to improve the
>>>> +			performance. By default, its value is set to 0.
>>>> +
>>>>   	ignore_loglevel	[KNL]
>>>>   			Ignore loglevel setting - this will print /all/
>>>>   			kernel messages to the console. Useful for debugging.
>>>> diff --git a/drivers/char/tpm/tpm_tis_core.c
>>> b/drivers/char/tpm/tpm_tis_core.c
>>>> index 63bc6c3b949e..6b9bf4c4d434 100644
>>>> --- a/drivers/char/tpm/tpm_tis_core.c
>>>> +++ b/drivers/char/tpm/tpm_tis_core.c
>>>> @@ -31,6 +31,11 @@
>>>>   #include "tpm.h"
>>>>   #include "tpm_tis_core.h"
>>>>   
>>>> +static bool ignore_burst_count = false;
>>>> +module_param(ignore_burst_count, bool, 0444);
>>>> +MODULE_PARM_DESC(ignore_burst_count,
>>>> +		"Ignore burstcount value while writing data");
>>>> +
>>>>   /* Before we attempt to access the TPM we must see that the valid
>>> bit is set.
>>>>    * The specification says that this bit is 0 at reset and remains 0
>>> until the
>>>>    * 'TPM has gone through its self test and initialization and has
>>> established
>>>> @@ -256,6 +261,7 @@ static int tpm_tis_send_data(struct tpm_chip
>>> *chip, u8 *buf, size_t len)
>>>>   {
>>>>   	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
>>>>   	int rc, status, burstcnt;
>>>> +	int sendcnt;
>>>>   	size_t count = 0;
>>>>   	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
>>>>   
>>>> @@ -271,19 +277,31 @@ static int tpm_tis_send_data(struct tpm_chip
>>> *chip, u8 *buf, size_t len)
>>>>   	}
>>>>   
>>>>   	while (count < len - 1) {
>>>> +
>>>> +		/*
>>>> +		 * Get the initial burstcount to ensure TPM is ready to
>>>> +		 * accept data, even when waiting for burstcount is disabled.
>>>> +		 */
>>>>   		burstcnt = get_burstcount(chip);
>>>>   		if (burstcnt < 0) {
>>>>   			dev_err(&chip->dev, "Unable to read burstcount\n");
>>>>   			rc = burstcnt;
>>>>   			goto out_err;
>>>>   		}
>>>> -		burstcnt = min_t(int, burstcnt, len - count - 1);
>>>> +
>>>> +		if (ignore_burst_count)
>>>> +			sendcnt = len - 1;
>>>> +		else
>>>> +			sendcnt = min_t(int, burstcnt, len - count - 1);
>>>> +
>>>>   		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
>>>> -					 burstcnt, buf + count);
>>>> +					 sendcnt, buf + count);
>>>>   		if (rc < 0)
>>>>   			goto out_err;
>>>>   
>>>> -		count += burstcnt;
>>>> +		count += sendcnt;
>>>> +		if (ignore_burst_count)
>>>> +			continue;
>>>>   
>>>>   		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
>>>>   					&priv->int_queue, false) < 0) {
>>>> -- 
>>>> 2.13.3
>>>>
>>> Makes sense to discuss whether to have the kernel command-line
>>> parameter or not before applying this.
>>>
>>> To fuel the discussion, alternative to this would be:
>>>
>>> 1. Have this always on i.e. no command-line parameter.
>>> 2. If someone yells, we add the command-line parameter later on.
>>>
>> According to what I've read in the tcg ddwg group this patch should
>> not cause problems on _sane_ tpms.
>>
>> I'm not 100%convinced that all tpms are sane all the time, but I think
>> we do not want yet another cmdline parameter.
>>
>> So if we want to pull it in (and ddwg does not see an issue, so yes)
>> it should be on by default, without a kernel parameter.
>>
>> If there is a kernel parameter, then it should only be one called
>> "failsafe" - which includes the force behavior and maybe the "broken"
>> tpm path.
>>
>> But I agree with Alex, every additonal code path reduces testing coverage.
>>
>>
>> We would be happy to test a "default on" patch.
>>
>> Peter
>>
>>> /Jarkko
> I'm starting to dilate to this direction.
>
> It is hard to believe that any such TPM would be in active use anywhere
> assuming that there exist a TPM where this causes issues. This combined
> to the assumption that you would run the latest mainline on it makes it
> a pretty insignificant scenario.

It sounds like we are getting in direction to have this change by default.
Before removing the ignore_burst_count parameter, I will post a test version of this
patch which enables ignore_burst_count by default, for testing purposes only.

Thanks Peter and Alex for testing.

Thanks & Regards,
    - Nayna


>
> /Jarkko
>


------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
Jarkko Sakkinen - Sept. 15, 2017, 3:19 p.m.
On Fri, Sep 15, 2017 at 05:59:21PM +0530, Nayna Jain wrote:
> 
> 
> On 09/14/2017 04:40 AM, Jarkko Sakkinen wrote:
> > On Wed, Sep 13, 2017 at 11:39:03AM -0700, Peter Huewe wrote:
> > > 
> > > Am 12. September 2017 17:45:08 GMT-07:00 schrieb Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>:
> > > > On Wed, Sep 06, 2017 at 08:56:36AM -0400, Nayna Jain wrote:
> > > > > The TPM burstcount status indicates the number of bytes that can
> > > > > be sent to the TPM without causing bus wait states.  Effectively,
> > > > > it is the number of empty bytes in the command FIFO. Further,
> > > > > some TPMs have a static burstcount, when the value remains zero
> > > > > until the entire FIFO is empty.
> > > > > 
> > > > > This patch adds an optimization to check for burstcount only once.
> > > > > And if it is valid, it writes all the bytes at once, permitting
> > > > > wait states. The performance of a 34 byte extend on a TPM 1.2 with
> > > > > an 8 byte burstcount improved from 41 msec to 14 msec.
> > > > > 
> > > > > This functionality is enabled only by passing module
> > > > > parameter ignore_burst_count=1. By default, this parameter
> > > > > is disabled.
> > > > > 
> > > > > After this change, performance on a TPM 1.2 with an 8 byte
> > > > > burstcount for 1000 extends improved from ~41sec to ~14sec.
> > > > > 
> > > > > Suggested-by: Ken Goldman <kgold@linux.vnet.ibm.com> in
> > > > > conjunction with the TPM Device Driver work group.
> > > > > Signed-off-by: Nayna Jain <nayna@linux.vnet.ibm.com>
> > > > > Acked-by: Mimi Zohar <zohar@linux.vnet.ibm.com>
> > > > > ---
> > > > >   Documentation/admin-guide/kernel-parameters.txt |  8 ++++++++
> > > > >   drivers/char/tpm/tpm_tis_core.c                 | 24
> > > > +++++++++++++++++++++---
> > > > >   2 files changed, 29 insertions(+), 3 deletions(-)
> > > > > 
> > > > > diff --git a/Documentation/admin-guide/kernel-parameters.txt
> > > > b/Documentation/admin-guide/kernel-parameters.txt
> > > > > index 4e303be83df6..3c59bb91e1ee 100644
> > > > > --- a/Documentation/admin-guide/kernel-parameters.txt
> > > > > +++ b/Documentation/admin-guide/kernel-parameters.txt
> > > > > @@ -1465,6 +1465,14 @@
> > > > >   			mode generally follows that for the NaN encoding,
> > > > >   			except where unsupported by hardware.
> > > > > +	ignore_burst_count [TPM_TIS_CORE]
> > > > > +			tpm_tis_core driver queries for the burstcount before
> > > > > +			every send call in a loop. However, it causes delay to
> > > > > +			the send command for TPMs with low burstcount value.
> > > > > +			Setting this value to 1, will make driver to query for
> > > > > +			burstcount only once in the loop to improve the
> > > > > +			performance. By default, its value is set to 0.
> > > > > +
> > > > >   	ignore_loglevel	[KNL]
> > > > >   			Ignore loglevel setting - this will print /all/
> > > > >   			kernel messages to the console. Useful for debugging.
> > > > > diff --git a/drivers/char/tpm/tpm_tis_core.c
> > > > b/drivers/char/tpm/tpm_tis_core.c
> > > > > index 63bc6c3b949e..6b9bf4c4d434 100644
> > > > > --- a/drivers/char/tpm/tpm_tis_core.c
> > > > > +++ b/drivers/char/tpm/tpm_tis_core.c
> > > > > @@ -31,6 +31,11 @@
> > > > >   #include "tpm.h"
> > > > >   #include "tpm_tis_core.h"
> > > > > +static bool ignore_burst_count = false;
> > > > > +module_param(ignore_burst_count, bool, 0444);
> > > > > +MODULE_PARM_DESC(ignore_burst_count,
> > > > > +		"Ignore burstcount value while writing data");
> > > > > +
> > > > >   /* Before we attempt to access the TPM we must see that the valid
> > > > bit is set.
> > > > >    * The specification says that this bit is 0 at reset and remains 0
> > > > until the
> > > > >    * 'TPM has gone through its self test and initialization and has
> > > > established
> > > > > @@ -256,6 +261,7 @@ static int tpm_tis_send_data(struct tpm_chip
> > > > *chip, u8 *buf, size_t len)
> > > > >   {
> > > > >   	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
> > > > >   	int rc, status, burstcnt;
> > > > > +	int sendcnt;
> > > > >   	size_t count = 0;
> > > > >   	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
> > > > > @@ -271,19 +277,31 @@ static int tpm_tis_send_data(struct tpm_chip
> > > > *chip, u8 *buf, size_t len)
> > > > >   	}
> > > > >   	while (count < len - 1) {
> > > > > +
> > > > > +		/*
> > > > > +		 * Get the initial burstcount to ensure TPM is ready to
> > > > > +		 * accept data, even when waiting for burstcount is disabled.
> > > > > +		 */
> > > > >   		burstcnt = get_burstcount(chip);
> > > > >   		if (burstcnt < 0) {
> > > > >   			dev_err(&chip->dev, "Unable to read burstcount\n");
> > > > >   			rc = burstcnt;
> > > > >   			goto out_err;
> > > > >   		}
> > > > > -		burstcnt = min_t(int, burstcnt, len - count - 1);
> > > > > +
> > > > > +		if (ignore_burst_count)
> > > > > +			sendcnt = len - 1;
> > > > > +		else
> > > > > +			sendcnt = min_t(int, burstcnt, len - count - 1);
> > > > > +
> > > > >   		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
> > > > > -					 burstcnt, buf + count);
> > > > > +					 sendcnt, buf + count);
> > > > >   		if (rc < 0)
> > > > >   			goto out_err;
> > > > > -		count += burstcnt;
> > > > > +		count += sendcnt;
> > > > > +		if (ignore_burst_count)
> > > > > +			continue;
> > > > >   		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
> > > > >   					&priv->int_queue, false) < 0) {
> > > > > -- 
> > > > > 2.13.3
> > > > > 
> > > > Makes sense to discuss whether to have the kernel command-line
> > > > parameter or not before applying this.
> > > > 
> > > > To fuel the discussion, alternative to this would be:
> > > > 
> > > > 1. Have this always on i.e. no command-line parameter.
> > > > 2. If someone yells, we add the command-line parameter later on.
> > > > 
> > > According to what I've read in the tcg ddwg group this patch should
> > > not cause problems on _sane_ tpms.
> > > 
> > > I'm not 100%convinced that all tpms are sane all the time, but I think
> > > we do not want yet another cmdline parameter.
> > > 
> > > So if we want to pull it in (and ddwg does not see an issue, so yes)
> > > it should be on by default, without a kernel parameter.
> > > 
> > > If there is a kernel parameter, then it should only be one called
> > > "failsafe" - which includes the force behavior and maybe the "broken"
> > > tpm path.
> > > 
> > > But I agree with Alex, every additonal code path reduces testing coverage.
> > > 
> > > 
> > > We would be happy to test a "default on" patch.
> > > 
> > > Peter
> > > 
> > > > /Jarkko
> > I'm starting to dilate to this direction.
> > 
> > It is hard to believe that any such TPM would be in active use anywhere
> > assuming that there exist a TPM where this causes issues. This combined
> > to the assumption that you would run the latest mainline on it makes it
> > a pretty insignificant scenario.
> 
> It sounds like we are getting in direction to have this change by default.
> Before removing the ignore_burst_count parameter, I will post a test version of this
> patch which enables ignore_burst_count by default, for testing purposes only.
> 
> Thanks Peter and Alex for testing.
> 
> Thanks & Regards,
>    - Nayna

I could apply it immediately after some testing to my next branch where
it gets pulled to linux-next. There's still a lot of time before next
pull request so many people would get exposed. If it cause problems,
we reconsider.

/Jarkko

------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot

Patch

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index 4e303be83df6..3c59bb91e1ee 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -1465,6 +1465,14 @@ 
 			mode generally follows that for the NaN encoding,
 			except where unsupported by hardware.
 
+	ignore_burst_count [TPM_TIS_CORE]
+			tpm_tis_core driver queries for the burstcount before
+			every send call in a loop. However, it causes delay to
+			the send command for TPMs with low burstcount value.
+			Setting this value to 1, will make driver to query for
+			burstcount only once in the loop to improve the
+			performance. By default, its value is set to 0.
+
 	ignore_loglevel	[KNL]
 			Ignore loglevel setting - this will print /all/
 			kernel messages to the console. Useful for debugging.
diff --git a/drivers/char/tpm/tpm_tis_core.c b/drivers/char/tpm/tpm_tis_core.c
index 63bc6c3b949e..6b9bf4c4d434 100644
--- a/drivers/char/tpm/tpm_tis_core.c
+++ b/drivers/char/tpm/tpm_tis_core.c
@@ -31,6 +31,11 @@ 
 #include "tpm.h"
 #include "tpm_tis_core.h"
 
+static bool ignore_burst_count = false;
+module_param(ignore_burst_count, bool, 0444);
+MODULE_PARM_DESC(ignore_burst_count,
+		"Ignore burstcount value while writing data");
+
 /* Before we attempt to access the TPM we must see that the valid bit is set.
  * The specification says that this bit is 0 at reset and remains 0 until the
  * 'TPM has gone through its self test and initialization and has established
@@ -256,6 +261,7 @@  static int tpm_tis_send_data(struct tpm_chip *chip, u8 *buf, size_t len)
 {
 	struct tpm_tis_data *priv = dev_get_drvdata(&chip->dev);
 	int rc, status, burstcnt;
+	int sendcnt;
 	size_t count = 0;
 	bool itpm = priv->flags & TPM_TIS_ITPM_WORKAROUND;
 
@@ -271,19 +277,31 @@  static int tpm_tis_send_data(struct tpm_chip *chip, u8 *buf, size_t len)
 	}
 
 	while (count < len - 1) {
+
+		/*
+		 * Get the initial burstcount to ensure TPM is ready to
+		 * accept data, even when waiting for burstcount is disabled.
+		 */
 		burstcnt = get_burstcount(chip);
 		if (burstcnt < 0) {
 			dev_err(&chip->dev, "Unable to read burstcount\n");
 			rc = burstcnt;
 			goto out_err;
 		}
-		burstcnt = min_t(int, burstcnt, len - count - 1);
+
+		if (ignore_burst_count)
+			sendcnt = len - 1;
+		else
+			sendcnt = min_t(int, burstcnt, len - count - 1);
+
 		rc = tpm_tis_write_bytes(priv, TPM_DATA_FIFO(priv->locality),
-					 burstcnt, buf + count);
+					 sendcnt, buf + count);
 		if (rc < 0)
 			goto out_err;
 
-		count += burstcnt;
+		count += sendcnt;
+		if (ignore_burst_count)
+			continue;
 
 		if (wait_for_tpm_stat(chip, TPM_STS_VALID, chip->timeout_c,
 					&priv->int_queue, false) < 0) {