Skip to contentSkip to conversation

cpufreq: ti: Fix probe ordering with k3-socinfo

PATCH Applied Akashdeep Kaur 3 participants 6 messages May 28 – Jun 8 CPU frequency scaling framework

v4 of 5 Applied Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v4 cover cpufreq: ti: Fix probe ordering with k3-socinfo

For K3 SoCs, ti-cpufreq depends on k3-socinfo to provide SoC revision information via soc_device_match(). If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, causing incorrect revision detection and OPP table initialization failures.

Add a needs_k3_socinfo flag to ti_cpufreq_soc_data to properly handle probe deferral when k3-socinfo hasn't registered the SoC device yet. This follows the existing multi_regulator pattern in the driver.

Changes in v4
- Use needs_k3_socinfo flag in soc_data
- Link to v3: https://lore.kernel.org/all/[email protected]/

Changes in v3
- Simplified to single patch (EPROBE_DEFER only)
- Changed the logic to detect whether the probe is for am62 family SoCs
- Dropped device link patch after analysis as revision information is cached after successful probe. Module reload also works as k3-socinfo stays bound during normal operation
- Link to v2: https://lore.kernel.org/all/[email protected]/

Changes in v2
- Added DT bindings documentation for ti,soc-info property
- Reordered patches: bindings first, then driver changes, then DTS
- Link to v1: https://lore.kernel.org/all/[email protected]/

Testing
- Verified correct probe ordering on AM625, AM62A7, AM62P5 platforms

---

Akashdeep Kaur (1):
cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

 drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
Signature · 1 line
2.34.1
#1 lore raw

PATCH v4 1/1 cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

On K3 SoCs, ti-cpufreq relies on k3-socinfo to register the SoC device before soc_device_match() can return valid revision information. If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, leading to incorrect CPU frequency scaling behavior.
Defer probe when k3-socinfo hasn't registered the SoC device yet.

---
 drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/cpufreq/ti-cpufreq.c +12 -0
index a01abc1622eb..8219751da175 100644
@@ -99,6 +99,7 @@ struct ti_cpufreq_soc_data {
 	unsigned long efuse_shift;
 	unsigned long rev_offset;
 	bool multi_regulator;
+	bool needs_k3_socinfo;
 /* Backward compatibility hack: Might have missing syscon */
 #define TI_QUIRK_SYSCON_MAY_BE_MISSING	0x1
 /* Backward compatibility hack: new syscon size is 1 register wide */
@@ -347,6 +348,7 @@ static struct ti_cpufreq_soc_data am625_soc_data = {
 	.efuse_mask = 0x07c0,
 	.efuse_shift = 0x6,
 	.multi_regulator = false,
+	.needs_k3_socinfo = true,
 	.quirks = TI_QUIRK_SYSCON_IS_SINGLE_REG,
 };
 
@@ -356,6 +358,7 @@ static struct ti_cpufreq_soc_data am62a7_soc_data = {
 	.efuse_mask = 0x07c0,
 	.efuse_shift = 0x6,
 	.multi_regulator = false,
+	.needs_k3_socinfo = true,
 };
 
 static struct ti_cpufreq_soc_data am62l3_soc_data = {
@@ -364,6 +367,7 @@ static struct ti_cpufreq_soc_data am62l3_soc_data = {
 	.efuse_mask = 0x07c0,
 	.efuse_shift = 0x6,
 	.multi_regulator = false,
+	.needs_k3_socinfo = true,
 };
 
 static struct ti_cpufreq_soc_data am62p5_soc_data = {
@@ -372,6 +376,7 @@ static struct ti_cpufreq_soc_data am62p5_soc_data = {
 	.efuse_mask = 0x07c0,
 	.efuse_shift = 0x6,
 	.multi_regulator = false,
+	.needs_k3_socinfo = true,
 };
 
 /**
@@ -443,6 +448,12 @@ static int ti_cpufreq_get_rev(struct ti_cpufreq_data *opp_data,
 		goto done;
 	}
 
+	/* Defer if k3-socinfo hasn't registered the SoC device yet */
+	if (opp_data->soc_data->needs_k3_socinfo) {
+		dev_dbg(opp_data->cpu_dev,
+			"SoC info not ready yet, deferring probe\n");
+		return -EPROBE_DEFER;
+	}
+
 	ret = regmap_read(opp_data->syscon, opp_data->soc_data->rev_offset,
 			  &revision);

if (opp_data->soc_data->quirks & TI_QUIRK_SYSCON_MAY_BE_MISSING && ret == -EIO) {

Signature · 1 line
2.34.1
#2 lore raw

Re: [PATCH v4 1/1] cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

Quoted 6 lines · Akashdeep Kaur

On K3 SoCs, ti-cpufreq relies on k3-socinfo to register the SoC device before soc_device_match() can return valid revision information. If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, leading to incorrect CPU frequency scaling behavior.
Defer probe when k3-socinfo hasn't registered the SoC device yet.

Hi Akashdeep,

Thanks for the update.

Quoted 43 lines
---
  drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++
  1 file changed, 12 insertions(+)
drivers/cpufreq/ti-cpufreq.c +4 -0
index a01abc1622eb..8219751da175 100644
@@ -99,6 +99,7 @@ struct ti_cpufreq_soc_data {
  	unsigned long efuse_shift;
  	unsigned long rev_offset;
  	bool multi_regulator;
+	bool needs_k3_socinfo;
  /* Backward compatibility hack: Might have missing syscon */
  #define TI_QUIRK_SYSCON_MAY_BE_MISSING	0x1
  /* Backward compatibility hack: new syscon size is 1 register wide */
@@ -347,6 +348,7 @@ static struct ti_cpufreq_soc_data am625_soc_data = {
  	.efuse_mask = 0x07c0,
  	.efuse_shift = 0x6,
  	.multi_regulator = false,
+	.needs_k3_socinfo = true,
  	.quirks = TI_QUIRK_SYSCON_IS_SINGLE_REG,
  };
 
@@ -356,6 +358,7 @@ static struct ti_cpufreq_soc_data am62a7_soc_data = {
  	.efuse_mask = 0x07c0,
  	.efuse_shift = 0x6,
  	.multi_regulator = false,
+	.needs_k3_socinfo = true,
  };
 
  static struct ti_cpufreq_soc_data am62l3_soc_data = {
@@ -364,6 +367,7 @@ static struct ti_cpufreq_soc_data am62l3_soc_data = {
  	.efuse_mask = 0x07c0,
  	.efuse_shift = 0x6,
  	.multi_regulator = false,
+	.needs_k3_socinfo = true,
  };
 
  static struct ti_cpufreq_soc_data am62p5_soc_data = {

drivers/cpufreq/ti-cpufreq.c static struct ti_cpufreq_soc_data am62p5_soc_data = { · line 379

Quoted 4 lines
drivers/cpufreq/ti-cpufreq.c +1 -0
@@ -372,6 +376,7 @@ static struct ti_cpufreq_soc_data am62p5_soc_data = {
  	.efuse_mask = 0x07c0,
  	.efuse_shift = 0x6,
  	.multi_regulator = false,
+	.needs_k3_socinfo = true,

It seems that ti_cpufreq_soc_data is a hardware description–only structure.

After further consideration, may I know would it be reasonable to consider using of_machine_get_match() instead, for example:

static const struct of_device_id ti_k3_cpufreq_of_match[] = {
             { .compatible = "ti,am625"  },
             { .compatible = "ti,am62xx" },
               ......
             {}
     };
if (soc_device_match(k3_cpufreq_soc)) {
    *revision_value = 0x1;
    goto done;
} else if (of_machine_get_match(ti_k3_of_match)) {
    return dev_err_probe(xxx);
}

Just a thought — you might want to consider this if it makes sense. Sorry for not pointing this out earlier, I may have missed it before.

drivers/cpufreq/ti-cpufreq.c static int ti_cpufreq_get_rev(struct ti_cpufreq_data *opp_data, · line 455

Quoted 12 lines
drivers/cpufreq/ti-cpufreq.c +5 -0
  };
 
  /**
@@ -443,6 +448,12 @@ static int ti_cpufreq_get_rev(struct ti_cpufreq_data *opp_data,
  		goto done;
  	}
 
+	/* Defer if k3-socinfo hasn't registered the SoC device yet */
+	if (opp_data->soc_data->needs_k3_socinfo) {
+		dev_dbg(opp_data->cpu_dev,
+			"SoC info not ready yet, deferring probe\n");
+		return -EPROBE_DEFER;

It would be better to use dev_err_probe() here and return the original error, instead of using dev_dbg() with -EPROBE_DEFER.

Also, the message "SoC info not ready yet" seems a bit too generic and not very descriptive. Which SoC is this referring to?

Quoted 5 lines
drivers/cpufreq/ti-cpufreq.c +2 -0
+	}
+
  	ret = regmap_read(opp_data->syscon, opp_data->soc_data->rev_offset,
  			  &revision);
  	if (opp_data->soc_data->quirks & TI_QUIRK_SYSCON_MAY_BE_MISSING && ret == -EIO) {
Signature · 2 lines
Thx and BRs,
Zhongqiu Han
#3 lore raw

Re: [PATCH v4 1/1] cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

Hi Zhongqiu,

Thanks for the feedback!

Quoted 80 lines · Zhongqiu Han
Quoted 6 lines · Akashdeep Kaur

On K3 SoCs, ti-cpufreq relies on k3-socinfo to register the SoC device before soc_device_match() can return valid revision information. If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, leading to incorrect CPU frequency scaling behavior.
Defer probe when k3-socinfo hasn't registered the SoC device yet.

Hi Akashdeep,

Thanks for the update.

Quoted 40 lines

---
  drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++

  1 file changed, 12 insertions(+)
drivers/cpufreq/ti-cpufreq.c +4 -0
index a01abc1622eb..8219751da175 100644
@@ -99,6 +99,7 @@ struct ti_cpufreq_soc_data {
       unsigned long efuse_shift;
       unsigned long rev_offset;
       bool multi_regulator;
+    bool needs_k3_socinfo;
   /* Backward compatibility hack: Might have missing syscon */
   #define TI_QUIRK_SYSCON_MAY_BE_MISSING    0x1
   /* Backward compatibility hack: new syscon size is 1 register wide */
@@ -347,6 +348,7 @@ static struct ti_cpufreq_soc_data am625_soc_data = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
       .quirks = TI_QUIRK_SYSCON_IS_SINGLE_REG,
   };
@@ -356,6 +358,7 @@ static struct ti_cpufreq_soc_data am62a7_soc_data = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
   };
   static struct ti_cpufreq_soc_data am62l3_soc_data = {
@@ -364,6 +367,7 @@ static struct ti_cpufreq_soc_data am62l3_soc_data = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
   };
   static struct ti_cpufreq_soc_data am62p5_soc_data = {

drivers/cpufreq/ti-cpufreq.c static struct ti_cpufreq_soc_data am62p5_soc_data = { · line 379

Quoted 4 lines
drivers/cpufreq/ti-cpufreq.c +1 -0
@@ -372,6 +376,7 @@ static struct ti_cpufreq_soc_data am62p5_soc_data = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,

It seems that ti_cpufreq_soc_data is a hardware description–only structure.

After further consideration, may I know would it be reasonable to consider using of_machine_get_match() instead, for example:

static const struct of_device_id ti_k3_cpufreq_of_match[] = {
             { .compatible = "ti,am625"  },
             { .compatible = "ti,am62xx" },
               ......
             {}
     };

if (soc_device_match(k3_cpufreq_soc)) {
    *revision_value = 0x1;
    goto done;
} else if (of_machine_get_match(ti_k3_of_match)) {
    return dev_err_probe(xxx);
}

Quoted 2 lines

Just a thought — you might want to consider this if it makes sense. Sorry for not pointing this out earlier, I may have missed it befor

The structure already contains driver behavior flags alongside hardware descriptions:

1. multi_regulator (bool) - Controls whether the driver uses multiple
regulators (driver behavior, not HW description)
2. quirks (u8) - Explicitly documented as "Backward compatibility
hack" flags that control driver behavior:
- TI_QUIRK_SYSCON_MAY_BE_MISSING
- TI_QUIRK_SYSCON_IS_SINGLE_REG

The needs_k3_socinfo flag follows this existing pattern - it's a
driver behavior flag that controls probe deferral logic, similar to how multi_regulator controls regulator configuration and quirks control backward compatibility handling.

Quoted 18 lines

drivers/cpufreq/ti-cpufreq.c static int ti_cpufreq_get_rev(struct · line 456

Quoted 13 lines
drivers/cpufreq/ti-cpufreq.c +5 -0
   };
   /**
@@ -443,6 +448,12 @@ static int ti_cpufreq_get_rev(struct 
 ti_cpufreq_data *opp_data,
           goto done;
       }
 
+    /* Defer if k3-socinfo hasn't registered the SoC device yet */
+    if (opp_data->soc_data->needs_k3_socinfo) {
+        dev_dbg(opp_data->cpu_dev,
+            "SoC info not ready yet, deferring probe\n");
+        return -EPROBE_DEFER;

It would be better to use dev_err_probe() here and return the original error, instead of using dev_dbg() with -EPROBE_DEFER.

Quoted 2 lines

Also, the message "SoC info not ready yet" seems a bit too generic and not very descriptive. Which SoC is this referring to?

Thats good suggestion, will update both in next version.

Thanks,
Akashdeep Kaur

Quoted 9 lines
Quoted 7 lines
drivers/cpufreq/ti-cpufreq.c +2 -0
+    }
+
       ret = regmap_read(opp_data->syscon, opp_data->soc_data->rev_offset,

                &revision);
      if (opp_data->soc_data->quirks & TI_QUIRK_SYSCON_MAY_BE_MISSING && ret == -EIO) {

#4 lore raw

Re: [PATCH v4 1/1] cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

Quoted 100 lines · Akashdeep Kaur

Hi Zhongqiu,

Thanks for the feedback!

Quoted 83 lines · Zhongqiu Han
Quoted 6 lines · Akashdeep Kaur

On K3 SoCs, ti-cpufreq relies on k3-socinfo to register the SoC device before soc_device_match() can return valid revision information. If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, leading to incorrect CPU frequency scaling behavior.
Defer probe when k3-socinfo hasn't registered the SoC device yet.

Hi Akashdeep,

Thanks for the update.

Quoted 42 lines

---
  drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++

  1 file changed, 12 insertions(+)
drivers/cpufreq/ti-cpufreq.c +4 -0
index a01abc1622eb..8219751da175 100644
@@ -99,6 +99,7 @@ struct ti_cpufreq_soc_data {
       unsigned long efuse_shift;
       unsigned long rev_offset;
       bool multi_regulator;
+    bool needs_k3_socinfo;
   /* Backward compatibility hack: Might have missing syscon */
   #define TI_QUIRK_SYSCON_MAY_BE_MISSING    0x1
   /* Backward compatibility hack: new syscon size is 1 register wide */
@@ -347,6 +348,7 @@ static struct ti_cpufreq_soc_data am625_soc_data = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
       .quirks = TI_QUIRK_SYSCON_IS_SINGLE_REG,
   };
@@ -356,6 +358,7 @@ static struct ti_cpufreq_soc_data am62a7_soc_data 
 = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
   };
   static struct ti_cpufreq_soc_data am62l3_soc_data = {
@@ -364,6 +367,7 @@ static struct ti_cpufreq_soc_data am62l3_soc_data 
 = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,
   };
   static struct ti_cpufreq_soc_data am62p5_soc_data = {

drivers/cpufreq/ti-cpufreq.c static struct ti_cpufreq_soc_data am62p5_soc_data · line 380

Quoted 5 lines
drivers/cpufreq/ti-cpufreq.c +1 -0
@@ -372,6 +376,7 @@ static struct ti_cpufreq_soc_data am62p5_soc_data 
 = {
       .efuse_mask = 0x07c0,
       .efuse_shift = 0x6,
       .multi_regulator = false,
+    .needs_k3_socinfo = true,

It seems that ti_cpufreq_soc_data is a hardware description–only structure.

After further consideration, may I know would it be reasonable to consider using of_machine_get_match() instead, for example:

static const struct of_device_id ti_k3_cpufreq_of_match[] = { { .compatible = "ti,am625"  },
              { .compatible = "ti,am62xx" },
                ......
              {}
      };

if (soc_device_match(k3_cpufreq_soc)) { *revision_value = 0x1;
     goto done;
} else if (of_machine_get_match(ti_k3_of_match)) { return dev_err_probe(xxx);
}

Quoted 2 lines

Just a thought — you might want to consider this if it makes sense. Sorry for not pointing this out earlier, I may have missed it befor

The structure already contains driver behavior flags alongside hardware descriptions:

  1. multi_regulator (bool) - Controls whether the driver uses multiple
regulators (driver behavior, not HW description)
  2. quirks (u8) - Explicitly documented as "Backward compatibility
hack" flags that control driver behavior:
    - TI_QUIRK_SYSCON_MAY_BE_MISSING
    - TI_QUIRK_SYSCON_IS_SINGLE_REG

Quoted 2 lines

  The needs_k3_socinfo flag follows this existing pattern - it's a
driver behavior flag that controls probe deferral logic, similar to how multi_regulator controls regulator configuration and quirks control backward compatibility handling.

Thanks for the clarification.

Quoted 36 lines
Quoted 18 lines

drivers/cpufreq/ti-cpufreq.c static int ti_cpufreq_get_rev(struct · line 456

Quoted 13 lines
drivers/cpufreq/ti-cpufreq.c +5 -0
   };
   /**
@@ -443,6 +448,12 @@ static int ti_cpufreq_get_rev(struct 
 ti_cpufreq_data *opp_data,
           goto done;
       }
 
+    /* Defer if k3-socinfo hasn't registered the SoC device yet */
+    if (opp_data->soc_data->needs_k3_socinfo) {
+        dev_dbg(opp_data->cpu_dev,
+            "SoC info not ready yet, deferring probe\n");
+        return -EPROBE_DEFER;

It would be better to use dev_err_probe() here and return the original error, instead of using dev_dbg() with -EPROBE_DEFER.

Quoted 2 lines

Also, the message "SoC info not ready yet" seems a bit too generic and not very descriptive. Which SoC is this referring to?

Thats good suggestion, will update both in next version.

Thanks,
Akashdeep Kaur

Quoted 10 lines
Quoted 8 lines
drivers/cpufreq/ti-cpufreq.c +2 -0
+    }
+

      ret = regmap_read(opp_data->syscon, opp_data->soc_data-

Quoted 1 line

rev_offset,

                &revision);
      if (opp_data->soc_data->quirks & TI_QUIRK_SYSCON_MAY_BE_MISSING && ret == -EIO) {

Signature · 2 lines
Thx and BRs,
Zhongqiu Han
maintainer Applied
#5 lore raw

Re: [PATCH v4 1/1] cpufreq: ti: Add EPROBE_DEFER for K3 SoCs

In reply to Akashdeep Kaur
Quoted 11 lines · Akashdeep Kaur

On K3 SoCs, ti-cpufreq relies on k3-socinfo to register the SoC device before soc_device_match() can return valid revision information. If ti-cpufreq probes before k3-socinfo, soc_device_match() returns NULL, leading to incorrect CPU frequency scaling behavior.
Defer probe when k3-socinfo hasn't registered the SoC device yet.

---
 drivers/cpufreq/ti-cpufreq.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

Applied. Thanks.

Signature · 1 line
viresh