[RFC] interconnect: qcom: add functions to query addr/cmds for a path
diff mbox series

Message ID 20200701042528.12321-1-jonathan@marek.ca
State New
Headers show
Series
  • [RFC] interconnect: qcom: add functions to query addr/cmds for a path
Related show

Commit Message

Jonathan Marek July 1, 2020, 4:25 a.m. UTC
The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
to query the interconnect driver for bcm addresses and commands.

I'm not sure what is the best way to go about implementing this, this is
what I came up with.

I included a quick example of how this can be used by the a6xx driver to
fill out the GMU bw_table (two ddr bandwidth levels in this example, note
this would be using the frequency table in dts and not hardcoded values).

Signed-off-by: Jonathan Marek <jonathan@marek.ca>
---
 drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
 drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
 include/soc/qcom/icc.h                | 11 ++++++
 3 files changed, 68 insertions(+), 13 deletions(-)
 create mode 100644 include/soc/qcom/icc.h

Comments

Jordan Crouse July 1, 2020, 4:56 p.m. UTC | #1
On Wed, Jul 01, 2020 at 12:25:25AM -0400, Jonathan Marek wrote:
> The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
> to query the interconnect driver for bcm addresses and commands.
> 
> I'm not sure what is the best way to go about implementing this, this is
> what I came up with.
> 
> I included a quick example of how this can be used by the a6xx driver to
> fill out the GMU bw_table (two ddr bandwidth levels in this example, note
> this would be using the frequency table in dts and not hardcoded values).

I would like to add my enthusiasm for this idea but I'm not much of an
interconnect or RPMh expert so I would defer to them to be sure that the APIs
are robust enough to cover all the corner cases.

> Signed-off-by: Jonathan Marek <jonathan@marek.ca>
> ---
>  drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
>  drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
>  include/soc/qcom/icc.h                | 11 ++++++
>  3 files changed, 68 insertions(+), 13 deletions(-)
>  create mode 100644 include/soc/qcom/icc.h
> 
> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> index ccd44d0418f8..1fb8f0480be3 100644
> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> @@ -4,6 +4,7 @@
>  #include <linux/completion.h>
>  #include <linux/circ_buf.h>
>  #include <linux/list.h>
> +#include <soc/qcom/icc.h>
>  
>  #include "a6xx_gmu.h"
>  #include "a6xx_gmu.xml.h"
> @@ -320,24 +321,18 @@ static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
>  	msg->cnoc_cmds_data[1][2] =  0x60000001;
>  }
>  
> -static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
> +static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg, struct icc_path *path)
>  {
>  	/*
>  	 * Send a single "off" entry just to get things running
>  	 * TODO: bus scaling
>  	 */
> -	msg->bw_level_num = 1;
> -
> -	msg->ddr_cmds_num = 3;
> +	msg->bw_level_num = 2;
>  	msg->ddr_wait_bitmask = 0x01;

We're going to need a API function for the wait bitmask too.
 
> -	msg->ddr_cmds_addrs[0] = 0x50000;
> -	msg->ddr_cmds_addrs[1] = 0x50004;
> -	msg->ddr_cmds_addrs[2] = 0x5007c;
> -
> -	msg->ddr_cmds_data[0][0] =  0x40000000;
> -	msg->ddr_cmds_data[0][1] =  0x40000000;
> -	msg->ddr_cmds_data[0][2] =  0x40000000;
> +	msg->ddr_cmds_num = qcom_icc_query_addr(path, msg->ddr_cmds_addrs);
> +	qcom_icc_query_cmd(path, msg->ddr_cmds_data[0], 0, 0);
> +	qcom_icc_query_cmd(path, msg->ddr_cmds_data[1], 0, 7216000);
>  
>  	/*
>  	 * These are the CX (CNOC) votes - these are used by the GMU but the
> @@ -388,7 +383,6 @@ static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
>  	msg->cnoc_cmds_data[1][2] =  0x60000001;
>  }
>  
> -
>  static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
>  {
>  	struct a6xx_hfi_msg_bw_table msg = { 0 };
> @@ -400,7 +394,7 @@ static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
>  	else if (adreno_is_a640(adreno_gpu))
>  		a640_build_bw_table(&msg);
>  	else if (adreno_is_a650(adreno_gpu))
> -		a650_build_bw_table(&msg);
> +		a650_build_bw_table(&msg, adreno_gpu->base.icc_path);
>  	else
>  		a6xx_build_bw_table(&msg);
>  
> diff --git a/drivers/interconnect/qcom/icc-rpmh.c b/drivers/interconnect/qcom/icc-rpmh.c
> index 3ac5182c9ab2..3ce2920330f9 100644
> --- a/drivers/interconnect/qcom/icc-rpmh.c
> +++ b/drivers/interconnect/qcom/icc-rpmh.c
> @@ -9,6 +9,7 @@
>  
>  #include "bcm-voter.h"
>  #include "icc-rpmh.h"
> +#include "../internal.h"
>  
>  /**
>   * qcom_icc_pre_aggregate - cleans up stale values from prior icc_set
> @@ -92,6 +93,55 @@ int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
>  }
>  EXPORT_SYMBOL_GPL(qcom_icc_set);
>  
> +static u32 bcm_query(struct qcom_icc_bcm *bcm, u64 sum_avg, u64 max_peak)
> +{
> +	u64 temp, agg_peak = 0;
> +	int i;
> +
> +	for (i = 0; i < bcm->num_nodes; i++) {
> +		temp = max_peak * bcm->aux_data.width;
> +		do_div(temp, bcm->nodes[i]->buswidth);
> +		agg_peak = max(agg_peak, temp);
> +	}
> +
> +	temp = agg_peak * 1000ULL;
> +	do_div(temp, bcm->aux_data.unit);
> +
> +	// TODO vote_x
> +
> +	return BCM_TCS_CMD(true, temp != 0, 0, temp);
> +}
> +
> +int qcom_icc_query_addr(struct icc_path *path, u32 *addr)

The leaf driver won't know the size of the path, so we'll likely need to kmalloc
and return the array or allow addr to be NULL and have the leaf driver do the
allocation itself once it knows what k is.

> +{
> +	struct qcom_icc_node *qn;
> +	int i, j, k = 0;
> +
> +	for (i = 0; i < path->num_nodes; i++) {
> +		qn = path->reqs[i].node->data;
> +		for (j = 0; j < qn->num_bcms; j++, k++)
> +			addr[k] = qn->bcms[j]->addr;
> +	}
> +
> +	return k;
> +}
> +EXPORT_SYMBOL_GPL(qcom_icc_query_addr);
> +
> +int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max)
> +{
> +	struct qcom_icc_node *qn;
> +	int i, j, k = 0;
> +
> +	for (i = 0; i < path->num_nodes; i++) {
> +		qn = path->reqs[i].node->data;
> +		for (j = 0; j < qn->num_bcms; j++, k++)
> +			cmd[k] = bcm_query(qn->bcms[j], avg, max);
> +	}
> +
> +	return 0;
> +}

Same as above.  When downstream did this for their old bespoke bus API they had
one function returns a struct with addrs / commands / wait bitmask.

I don't mind splitting up the function, but either way something is going to
have to query the number of commands in the path and allocate the buffers.

Jordan

> +EXPORT_SYMBOL_GPL(qcom_icc_query_cmd);
> +
>  /**
>   * qcom_icc_bcm_init - populates bcm aux data and connect qnodes
>   * @bcm: bcm to be initialized
> diff --git a/include/soc/qcom/icc.h b/include/soc/qcom/icc.h
> new file mode 100644
> index 000000000000..8d0ddde49739
> --- /dev/null
> +++ b/include/soc/qcom/icc.h
> @@ -0,0 +1,11 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +
> +#ifndef __SOC_QCOM_ICC_H__
> +#define __SOC_QCOM_ICC_H__
> +
> +#include <linux/interconnect.h>
> +
> +int qcom_icc_query_addr(struct icc_path *path, u32 *addr);
> +int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max);
> +
> +#endif /* __SOC_QCOM_ICC_H__ */
> -- 
> 2.26.1
>
Jonathan Marek July 1, 2020, 5:03 p.m. UTC | #2
On 7/1/20 12:56 PM, Jordan Crouse wrote:
> On Wed, Jul 01, 2020 at 12:25:25AM -0400, Jonathan Marek wrote:
>> The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
>> to query the interconnect driver for bcm addresses and commands.
>>
>> I'm not sure what is the best way to go about implementing this, this is
>> what I came up with.
>>
>> I included a quick example of how this can be used by the a6xx driver to
>> fill out the GMU bw_table (two ddr bandwidth levels in this example, note
>> this would be using the frequency table in dts and not hardcoded values).
> 
> I would like to add my enthusiasm for this idea but I'm not much of an
> interconnect or RPMh expert so I would defer to them to be sure that the APIs
> are robust enough to cover all the corner cases.
> 
>> Signed-off-by: Jonathan Marek <jonathan@marek.ca>
>> ---
>>   drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
>>   drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
>>   include/soc/qcom/icc.h                | 11 ++++++
>>   3 files changed, 68 insertions(+), 13 deletions(-)
>>   create mode 100644 include/soc/qcom/icc.h
>>
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
>> index ccd44d0418f8..1fb8f0480be3 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
>> @@ -4,6 +4,7 @@
>>   #include <linux/completion.h>
>>   #include <linux/circ_buf.h>
>>   #include <linux/list.h>
>> +#include <soc/qcom/icc.h>
>>   
>>   #include "a6xx_gmu.h"
>>   #include "a6xx_gmu.xml.h"
>> @@ -320,24 +321,18 @@ static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
>>   	msg->cnoc_cmds_data[1][2] =  0x60000001;
>>   }
>>   
>> -static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
>> +static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg, struct icc_path *path)
>>   {
>>   	/*
>>   	 * Send a single "off" entry just to get things running
>>   	 * TODO: bus scaling
>>   	 */
>> -	msg->bw_level_num = 1;
>> -
>> -	msg->ddr_cmds_num = 3;
>> +	msg->bw_level_num = 2;
>>   	msg->ddr_wait_bitmask = 0x01;
> 
> We're going to need a API function for the wait bitmask too.
>   
>> -	msg->ddr_cmds_addrs[0] = 0x50000;
>> -	msg->ddr_cmds_addrs[1] = 0x50004;
>> -	msg->ddr_cmds_addrs[2] = 0x5007c;
>> -
>> -	msg->ddr_cmds_data[0][0] =  0x40000000;
>> -	msg->ddr_cmds_data[0][1] =  0x40000000;
>> -	msg->ddr_cmds_data[0][2] =  0x40000000;
>> +	msg->ddr_cmds_num = qcom_icc_query_addr(path, msg->ddr_cmds_addrs);
>> +	qcom_icc_query_cmd(path, msg->ddr_cmds_data[0], 0, 0);
>> +	qcom_icc_query_cmd(path, msg->ddr_cmds_data[1], 0, 7216000);
>>   
>>   	/*
>>   	 * These are the CX (CNOC) votes - these are used by the GMU but the
>> @@ -388,7 +383,6 @@ static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
>>   	msg->cnoc_cmds_data[1][2] =  0x60000001;
>>   }
>>   
>> -
>>   static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
>>   {
>>   	struct a6xx_hfi_msg_bw_table msg = { 0 };
>> @@ -400,7 +394,7 @@ static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
>>   	else if (adreno_is_a640(adreno_gpu))
>>   		a640_build_bw_table(&msg);
>>   	else if (adreno_is_a650(adreno_gpu))
>> -		a650_build_bw_table(&msg);
>> +		a650_build_bw_table(&msg, adreno_gpu->base.icc_path);
>>   	else
>>   		a6xx_build_bw_table(&msg);
>>   
>> diff --git a/drivers/interconnect/qcom/icc-rpmh.c b/drivers/interconnect/qcom/icc-rpmh.c
>> index 3ac5182c9ab2..3ce2920330f9 100644
>> --- a/drivers/interconnect/qcom/icc-rpmh.c
>> +++ b/drivers/interconnect/qcom/icc-rpmh.c
>> @@ -9,6 +9,7 @@
>>   
>>   #include "bcm-voter.h"
>>   #include "icc-rpmh.h"
>> +#include "../internal.h"
>>   
>>   /**
>>    * qcom_icc_pre_aggregate - cleans up stale values from prior icc_set
>> @@ -92,6 +93,55 @@ int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
>>   }
>>   EXPORT_SYMBOL_GPL(qcom_icc_set);
>>   
>> +static u32 bcm_query(struct qcom_icc_bcm *bcm, u64 sum_avg, u64 max_peak)
>> +{
>> +	u64 temp, agg_peak = 0;
>> +	int i;
>> +
>> +	for (i = 0; i < bcm->num_nodes; i++) {
>> +		temp = max_peak * bcm->aux_data.width;
>> +		do_div(temp, bcm->nodes[i]->buswidth);
>> +		agg_peak = max(agg_peak, temp);
>> +	}
>> +
>> +	temp = agg_peak * 1000ULL;
>> +	do_div(temp, bcm->aux_data.unit);
>> +
>> +	// TODO vote_x
>> +
>> +	return BCM_TCS_CMD(true, temp != 0, 0, temp);
>> +}
>> +
>> +int qcom_icc_query_addr(struct icc_path *path, u32 *addr)
> 
> The leaf driver won't know the size of the path, so we'll likely need to kmalloc
> and return the array or allow addr to be NULL and have the leaf driver do the
> allocation itself once it knows what k is.
> 

In the a6xx gpu case, the a6xx_hfi_msg_bw_table has a fixed array size 
(allows up to 8 commands for ddr and 6 for cnoc), so there shouldn't be 
a need for any allocation.

Allowing addr to be NULL to get the # of addrs/cmds (so the a6xx driver 
can bail out if it can't fit, although that should never happen) would 
be OK (or having an array size parameter so the function can return an 
error), but IMO not needed for the "qcom_icc_query_cmd" function below, 
since it returns the same number of commands the "qcom_icc_query_addr" 
returns addresses.

>> +{
>> +	struct qcom_icc_node *qn;
>> +	int i, j, k = 0;
>> +
>> +	for (i = 0; i < path->num_nodes; i++) {
>> +		qn = path->reqs[i].node->data;
>> +		for (j = 0; j < qn->num_bcms; j++, k++)
>> +			addr[k] = qn->bcms[j]->addr;
>> +	}
>> +
>> +	return k;
>> +}
>> +EXPORT_SYMBOL_GPL(qcom_icc_query_addr);
>> +
>> +int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max)
>> +{
>> +	struct qcom_icc_node *qn;
>> +	int i, j, k = 0;
>> +
>> +	for (i = 0; i < path->num_nodes; i++) {
>> +		qn = path->reqs[i].node->data;
>> +		for (j = 0; j < qn->num_bcms; j++, k++)
>> +			cmd[k] = bcm_query(qn->bcms[j], avg, max);
>> +	}
>> +
>> +	return 0;
>> +}
> 
> Same as above.  When downstream did this for their old bespoke bus API they had
> one function returns a struct with addrs / commands / wait bitmask.
> 
> I don't mind splitting up the function, but either way something is going to
> have to query the number of commands in the path and allocate the buffers.
> 
> Jordan
> 
>> +EXPORT_SYMBOL_GPL(qcom_icc_query_cmd);
>> +
>>   /**
>>    * qcom_icc_bcm_init - populates bcm aux data and connect qnodes
>>    * @bcm: bcm to be initialized
>> diff --git a/include/soc/qcom/icc.h b/include/soc/qcom/icc.h
>> new file mode 100644
>> index 000000000000..8d0ddde49739
>> --- /dev/null
>> +++ b/include/soc/qcom/icc.h
>> @@ -0,0 +1,11 @@
>> +/* SPDX-License-Identifier: GPL-2.0-only */
>> +
>> +#ifndef __SOC_QCOM_ICC_H__
>> +#define __SOC_QCOM_ICC_H__
>> +
>> +#include <linux/interconnect.h>
>> +
>> +int qcom_icc_query_addr(struct icc_path *path, u32 *addr);
>> +int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max);
>> +
>> +#endif /* __SOC_QCOM_ICC_H__ */
>> -- 
>> 2.26.1
>>
>
Georgi Djakov July 13, 2020, 3:24 p.m. UTC | #3
On 7/1/20 07:25, Jonathan Marek wrote:
> The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
> to query the interconnect driver for bcm addresses and commands.

It's not very clear to me how the GMU firmware would be dealing with this? Does
anyone have an idea whether the GMU makes any bandwidth decisions? Or is it just
a static configuration and it just enables/disables a TCS?

I think that we can query the address from the cmd-db, but we have to know the
bcm names and the path. All the BCM/TCS information looks to be very low-level
and implementation specific, so exposing it through an API is not very good,
but hard-coding all this information is not good either.

Thanks,
Georgi

> 
> I'm not sure what is the best way to go about implementing this, this is
> what I came up with.
> 
> I included a quick example of how this can be used by the a6xx driver to
> fill out the GMU bw_table (two ddr bandwidth levels in this example, note
> this would be using the frequency table in dts and not hardcoded values).
> 
> Signed-off-by: Jonathan Marek <jonathan@marek.ca>
> ---
>  drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
>  drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
>  include/soc/qcom/icc.h                | 11 ++++++
>  3 files changed, 68 insertions(+), 13 deletions(-)
Jonathan Marek July 13, 2020, 3:33 p.m. UTC | #4
On 7/13/20 11:24 AM, Georgi Djakov wrote:
> On 7/1/20 07:25, Jonathan Marek wrote:
>> The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
>> to query the interconnect driver for bcm addresses and commands.
> 
> It's not very clear to me how the GMU firmware would be dealing with this? Does
> anyone have an idea whether the GMU makes any bandwidth decisions? Or is it just
> a static configuration and it just enables/disables a TCS?
> 

The GMU is able to vote for bandwidth, and presumably RPMh is 
aggregating those votes with the votes from the interconnect driver. We 
need to be able to fill out the "struct a6xx_hfi_msg_bw_table" in 
drivers/gpu/drm/msm/adreno/a6xx_hfi.h, which has a table of DDR 
bandwidth cmds which can be selected when changing the GPU frequency 
(the bw field in a6xx_hfi_gx_bw_perf_vote_cmd).

 From Jordan: "the GMU registers a bus vote directly with the hardware. 
It should handle aggregation for us."

> I think that we can query the address from the cmd-db, but we have to know the
> bcm names and the path. All the BCM/TCS information looks to be very low-level
> and implementation specific, so exposing it through an API is not very good,
> but hard-coding all this information is not good either.
> 
> Thanks,
> Georgi
> 
>>
>> I'm not sure what is the best way to go about implementing this, this is
>> what I came up with.
>>
>> I included a quick example of how this can be used by the a6xx driver to
>> fill out the GMU bw_table (two ddr bandwidth levels in this example, note
>> this would be using the frequency table in dts and not hardcoded values).
>>
>> Signed-off-by: Jonathan Marek <jonathan@marek.ca>
>> ---
>>   drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
>>   drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
>>   include/soc/qcom/icc.h                | 11 ++++++
>>   3 files changed, 68 insertions(+), 13 deletions(-)
Jordan Crouse July 13, 2020, 5:22 p.m. UTC | #5
On Mon, Jul 13, 2020 at 06:24:26PM +0300, Georgi Djakov wrote:
> On 7/1/20 07:25, Jonathan Marek wrote:
> > The a6xx GMU can vote for ddr and cnoc bandwidth, but it needs to be able
> > to query the interconnect driver for bcm addresses and commands.
> 
> It's not very clear to me how the GMU firmware would be dealing with this? Does
> anyone have an idea whether the GMU makes any bandwidth decisions? Or is it just
> a static configuration and it just enables/disables a TCS?

The GMU can perform a direct vote to the hardware. For now it is a static
configuration with pre-determined bandwidths generated from the OPP table.

> I think that we can query the address from the cmd-db, but we have to know the
> bcm names and the path. All the BCM/TCS information looks to be very low-level
> and implementation specific, so exposing it through an API is not very good,
> but hard-coding all this information is not good either.

Exactly my concern. The BCM information in particular is going to end up being
extremely target specific.

Jordan

> Thanks,
> Georgi
> 
> > 
> > I'm not sure what is the best way to go about implementing this, this is
> > what I came up with.
> > 
> > I included a quick example of how this can be used by the a6xx driver to
> > fill out the GMU bw_table (two ddr bandwidth levels in this example, note
> > this would be using the frequency table in dts and not hardcoded values).
> > 
> > Signed-off-by: Jonathan Marek <jonathan@marek.ca>
> > ---
> >  drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 20 ++++-------
> >  drivers/interconnect/qcom/icc-rpmh.c  | 50 +++++++++++++++++++++++++++
> >  include/soc/qcom/icc.h                | 11 ++++++
> >  3 files changed, 68 insertions(+), 13 deletions(-)

Patch
diff mbox series

diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
index ccd44d0418f8..1fb8f0480be3 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
@@ -4,6 +4,7 @@ 
 #include <linux/completion.h>
 #include <linux/circ_buf.h>
 #include <linux/list.h>
+#include <soc/qcom/icc.h>
 
 #include "a6xx_gmu.h"
 #include "a6xx_gmu.xml.h"
@@ -320,24 +321,18 @@  static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
 	msg->cnoc_cmds_data[1][2] =  0x60000001;
 }
 
-static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
+static void a650_build_bw_table(struct a6xx_hfi_msg_bw_table *msg, struct icc_path *path)
 {
 	/*
 	 * Send a single "off" entry just to get things running
 	 * TODO: bus scaling
 	 */
-	msg->bw_level_num = 1;
-
-	msg->ddr_cmds_num = 3;
+	msg->bw_level_num = 2;
 	msg->ddr_wait_bitmask = 0x01;
 
-	msg->ddr_cmds_addrs[0] = 0x50000;
-	msg->ddr_cmds_addrs[1] = 0x50004;
-	msg->ddr_cmds_addrs[2] = 0x5007c;
-
-	msg->ddr_cmds_data[0][0] =  0x40000000;
-	msg->ddr_cmds_data[0][1] =  0x40000000;
-	msg->ddr_cmds_data[0][2] =  0x40000000;
+	msg->ddr_cmds_num = qcom_icc_query_addr(path, msg->ddr_cmds_addrs);
+	qcom_icc_query_cmd(path, msg->ddr_cmds_data[0], 0, 0);
+	qcom_icc_query_cmd(path, msg->ddr_cmds_data[1], 0, 7216000);
 
 	/*
 	 * These are the CX (CNOC) votes - these are used by the GMU but the
@@ -388,7 +383,6 @@  static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
 	msg->cnoc_cmds_data[1][2] =  0x60000001;
 }
 
-
 static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
 {
 	struct a6xx_hfi_msg_bw_table msg = { 0 };
@@ -400,7 +394,7 @@  static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
 	else if (adreno_is_a640(adreno_gpu))
 		a640_build_bw_table(&msg);
 	else if (adreno_is_a650(adreno_gpu))
-		a650_build_bw_table(&msg);
+		a650_build_bw_table(&msg, adreno_gpu->base.icc_path);
 	else
 		a6xx_build_bw_table(&msg);
 
diff --git a/drivers/interconnect/qcom/icc-rpmh.c b/drivers/interconnect/qcom/icc-rpmh.c
index 3ac5182c9ab2..3ce2920330f9 100644
--- a/drivers/interconnect/qcom/icc-rpmh.c
+++ b/drivers/interconnect/qcom/icc-rpmh.c
@@ -9,6 +9,7 @@ 
 
 #include "bcm-voter.h"
 #include "icc-rpmh.h"
+#include "../internal.h"
 
 /**
  * qcom_icc_pre_aggregate - cleans up stale values from prior icc_set
@@ -92,6 +93,55 @@  int qcom_icc_set(struct icc_node *src, struct icc_node *dst)
 }
 EXPORT_SYMBOL_GPL(qcom_icc_set);
 
+static u32 bcm_query(struct qcom_icc_bcm *bcm, u64 sum_avg, u64 max_peak)
+{
+	u64 temp, agg_peak = 0;
+	int i;
+
+	for (i = 0; i < bcm->num_nodes; i++) {
+		temp = max_peak * bcm->aux_data.width;
+		do_div(temp, bcm->nodes[i]->buswidth);
+		agg_peak = max(agg_peak, temp);
+	}
+
+	temp = agg_peak * 1000ULL;
+	do_div(temp, bcm->aux_data.unit);
+
+	// TODO vote_x
+
+	return BCM_TCS_CMD(true, temp != 0, 0, temp);
+}
+
+int qcom_icc_query_addr(struct icc_path *path, u32 *addr)
+{
+	struct qcom_icc_node *qn;
+	int i, j, k = 0;
+
+	for (i = 0; i < path->num_nodes; i++) {
+		qn = path->reqs[i].node->data;
+		for (j = 0; j < qn->num_bcms; j++, k++)
+			addr[k] = qn->bcms[j]->addr;
+	}
+
+	return k;
+}
+EXPORT_SYMBOL_GPL(qcom_icc_query_addr);
+
+int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max)
+{
+	struct qcom_icc_node *qn;
+	int i, j, k = 0;
+
+	for (i = 0; i < path->num_nodes; i++) {
+		qn = path->reqs[i].node->data;
+		for (j = 0; j < qn->num_bcms; j++, k++)
+			cmd[k] = bcm_query(qn->bcms[j], avg, max);
+	}
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(qcom_icc_query_cmd);
+
 /**
  * qcom_icc_bcm_init - populates bcm aux data and connect qnodes
  * @bcm: bcm to be initialized
diff --git a/include/soc/qcom/icc.h b/include/soc/qcom/icc.h
new file mode 100644
index 000000000000..8d0ddde49739
--- /dev/null
+++ b/include/soc/qcom/icc.h
@@ -0,0 +1,11 @@ 
+/* SPDX-License-Identifier: GPL-2.0-only */
+
+#ifndef __SOC_QCOM_ICC_H__
+#define __SOC_QCOM_ICC_H__
+
+#include <linux/interconnect.h>
+
+int qcom_icc_query_addr(struct icc_path *path, u32 *addr);
+int qcom_icc_query_cmd(struct icc_path *path, u32 *cmd, u64 avg, u64 max);
+
+#endif /* __SOC_QCOM_ICC_H__ */