Skip to contentSkip to conversation

interconnect: qcom: add functions to query addr/cmds for a path

RFC PATCH In review Jonathan Marek 3 participants 6 messages Jul 1, 2020 – Jul 13, 2020

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 interconnect: qcom: add functions to query addr/cmds for a path

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).

---
 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
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +7 -13
index ccd44d0418f8..1fb8f0480be3 100644
@@ -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);
 
drivers/interconnect/qcom/icc-rpmh.c +50 -0

62 lines not on this page · raw message

include/soc/qcom/icc.h new file +11 -0
new file mode 100644
index 000000000000..8d0ddde49739
@@ -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__ */
Signature · 1 line
2.26.1
#1 lore raw
Quoted 9 lines · Jonathan Marek

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.

Quoted 26 lines
---
 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
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +2 -1
index ccd44d0418f8..1fb8f0480be3 100644
@@ -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)

drivers/gpu/drm/msm/adreno/a6xx_hfi.c static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg) · line 331

Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +1 -3
 {
 	/*
 	 * 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.

Quoted 56 lines
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +4 -9
-	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);
 
drivers/interconnect/qcom/icc-rpmh.c +11 -0
index 3ac5182c9ab2..3ce2920330f9 100644
@@ -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);
+	}

drivers/interconnect/qcom/icc-rpmh.c int qcom_icc_set(struct icc_node *src, struct icc_node *dst) · line 115

Quoted 10 lines
drivers/interconnect/qcom/icc-rpmh.c +10 -0
+
+	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.

Quoted 17 lines
drivers/interconnect/qcom/icc-rpmh.c +17 -0
+{
+	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;

drivers/interconnect/qcom/icc-rpmh.c int qcom_icc_set(struct icc_node *src, struct icc_node *dst) · line 142

Quoted 10 lines
drivers/interconnect/qcom/icc-rpmh.c +10 -0
+	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

Quoted 25 lines
drivers/interconnect/qcom/icc-rpmh.c +2 -0
+EXPORT_SYMBOL_GPL(qcom_icc_query_cmd);
+
 /**
  * qcom_icc_bcm_init - populates bcm aux data and connect qnodes
  * @bcm: bcm to be initialized
include/soc/qcom/icc.h new file +11 -0
new file mode 100644
index 000000000000..8d0ddde49739
@@ -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__ */
Signature · 1 line
2.26.1
Signature · 2 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
#2 lore raw
Quoted 116 lines · Jordan Crouse
Quoted 9 lines · Jonathan Marek

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.

Quoted 26 lines
---
  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

drivers/gpu/drm/msm/adreno/a6xx_hfi.c +2 -1
index ccd44d0418f8..1fb8f0480be3 100644
@@ -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)

drivers/gpu/drm/msm/adreno/a6xx_hfi.c static void a640_build_bw_table(struct a6xx_hfi_msg_bw_table *msg) · line 331

Quoted 10 lines
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +1 -3
  {
  	/*
  	 * 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.

Quoted 62 lines
drivers/gpu/drm/msm/adreno/a6xx_hfi.c +4 -9
-	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);
 
drivers/interconnect/qcom/icc-rpmh.c +17 -0
index 3ac5182c9ab2..3ce2920330f9 100644
@@ -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
+
Quoted 9 lines

drivers/interconnect/qcom/icc-rpmh.c int qcom_icc_set(struct icc_node *src, struct icc_node *dst) · line 115

Quoted 4 lines
drivers/interconnect/qcom/icc-rpmh.c +4 -0
+	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.

Quoted 62 lines
Quoted 17 lines
drivers/interconnect/qcom/icc-rpmh.c +17 -0
+{
+	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;

drivers/interconnect/qcom/icc-rpmh.c int qcom_icc_set(struct icc_node *src, struct icc_node *dst) · line 142

Quoted 10 lines
drivers/interconnect/qcom/icc-rpmh.c +10 -0
+	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

Quoted 25 lines
drivers/interconnect/qcom/icc-rpmh.c +2 -0
+EXPORT_SYMBOL_GPL(qcom_icc_query_cmd);
+
  /**
   * qcom_icc_bcm_init - populates bcm aux data and connect qnodes
   * @bcm: bcm to be initialized
include/soc/qcom/icc.h new file +11 -0
new file mode 100644
index 000000000000..8d0ddde49739
@@ -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__ */
Signature · 1 line
2.26.1
#3 lore raw
In reply to Jonathan Marek
Quoted 2 lines · Jonathan Marek

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

Quoted 14 lines

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).

---
 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(-)
#4 lore raw
Quoted 8 lines · Georgi Djakov
Quoted 2 lines · Jonathan Marek

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."

Quoted 22 lines

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

Quoted 14 lines

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).

---
  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(-)
#5 lore raw
In reply to Georgi Djakov
Quoted 7 lines · Georgi Djakov
Quoted 2 lines · Jonathan Marek

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.

Quoted 4 lines

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

Quoted 17 lines

Thanks,
Georgi

Quoted 14 lines

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).

---
 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(-)
Signature · 6 lines
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel