Skip to contentSkip to conversation

PECI patchset tweaks

PATCH Applied Zev Weiss 3 participants 18 messages Sep 26, 2020 – Oct 7, 2020

Subsystems (2)
lore
Share & tools
lore

Zev Weiss

author
#0 lore raw

cover PECI patchset tweaks

[Re-sending as I clumsily typoed the mailing list's address on the first attempt; apologies for the duplication.]

Hello,

These patches address a few small things I noticed with the PECI patch set currently in the OpenBMC kernel tree. Assuming they're deemed acceptable, I'd of course hope they get folded in to the version of the PECI code that gets upstreamed into the mainline kernel.

(Please let me know if there's some other way I should send this -- I looked for but couldn't find a gerrit repo for the obmc kernel or any kernel-specific patch submission instructions in openbmc/docs/CONTRIBUTING.md, so 'git send-email' is my best guess.)

Thanks,
Zev

author
#1 lore raw

PATCH 1/2 peci: fix error-handling in peci_dev_ioctl()

peci_get_xfer_msg() returns NULL on failure, not an ERR_PTR. Also avoid calling kfree() on an ERR_PTR.

---
 drivers/peci/peci-dev.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/peci/peci-dev.c +4 -3
index e0fe09467a80..84e90af81ccc 100644
@@ -122,8 +122,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
 		}
 
 		xmsg = peci_get_xfer_msg(uxmsg.tx_len, uxmsg.rx_len);
-		if (IS_ERR(xmsg)) {
-			ret = PTR_ERR(xmsg);
+		if (!xmsg) {
+			ret = -ENOMEM;
 			break;
 		}
 
@@ -162,7 +162,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
 	}
 
 	peci_put_xfer_msg(xmsg);
-	kfree(msg);
+	if (!IS_ERR(msg))
+		kfree(msg);
 
 	return (long)ret;
 }
Signature · 1 line
2.28.0
#2 lore raw

Re: [PATCH 1/2] peci: fix error-handling in peci_dev_ioctl()

Hello Zev,

Quoted 13 lines · Zev Weiss

peci_get_xfer_msg() returns NULL on failure, not an ERR_PTR. Also avoid calling kfree() on an ERR_PTR.

---
  drivers/peci/peci-dev.c | 7 ++++---
  1 file changed, 4 insertions(+), 3 deletions(-)
drivers/peci/peci-dev.c +0 -0
index e0fe09467a80..84e90af81ccc 100644

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg) · line 126

Quoted 7 lines
drivers/peci/peci-dev.c +2 -2
@@ -122,8 +122,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
  		}
 
  		xmsg = peci_get_xfer_msg(uxmsg.tx_len, uxmsg.rx_len);
-		if (IS_ERR(xmsg)) {
-			ret = PTR_ERR(xmsg);
+		if (!xmsg) {
+			ret = -ENOMEM;

Yes, it's a right fix. Thanks!

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg) · line 166

Quoted 10 lines
drivers/peci/peci-dev.c +2 -1
  			break;
  		}
 
@@ -162,7 +162,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
  	}
 
  	peci_put_xfer_msg(xmsg);
-	kfree(msg);
+	if (!IS_ERR(msg))
+		kfree(msg);

Not needed. kfree itself has null pointer checking inside.

Quoted 4 lines
drivers/peci/peci-dev.c +0 -0
  	return (long)ret;
  }
 
author
#3 lore raw

Re: [PATCH 1/2] peci: fix error-handling in peci_dev_ioctl()

Quoted 25 lines · Jae Hyun Yoo

Hello Zev,

Quoted 13 lines · Zev Weiss

peci_get_xfer_msg() returns NULL on failure, not an ERR_PTR. Also avoid calling kfree() on an ERR_PTR.

---
 drivers/peci/peci-dev.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/peci/peci-dev.c +0 -0
index e0fe09467a80..84e90af81ccc 100644

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg) · line 125

Quoted 6 lines
drivers/peci/peci-dev.c +2 -2
@@ -122,8 +122,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
 		}
 		xmsg = peci_get_xfer_msg(uxmsg.tx_len, uxmsg.rx_len);
-		if (IS_ERR(xmsg)) {
-			ret = PTR_ERR(xmsg);
+		if (!xmsg) {
+			ret = -ENOMEM;

Yes, it's a right fix. Thanks!

Quoted 11 lines

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg) · line 165

Quoted 8 lines
drivers/peci/peci-dev.c +2 -1
 			break;
 		}
@@ -162,7 +162,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
 	}
 	peci_put_xfer_msg(xmsg);
-	kfree(msg);
+	if (!IS_ERR(msg))
+		kfree(msg);

Not needed. kfree itself has null pointer checking inside.

Certainly, but the condition in question here isn't whether it's NULL, but whether it's an ERR_PTR (which as far as I can tell kfree() does not check for). As is, there's an error path that leads to passing a non-NULL but also non-kfree-safe ERR_PTR (the memdup_user() return value).

Zev

Reviewed-by
#4 lore raw

Re: [PATCH 1/2] peci: fix error-handling in peci_dev_ioctl()

Quoted 40 lines · Zev Weiss
Quoted 26 lines · Jae Hyun Yoo

Hello Zev,

Quoted 13 lines · Zev Weiss

peci_get_xfer_msg() returns NULL on failure, not an ERR_PTR.  Also avoid calling kfree() on an ERR_PTR.

---
 drivers/peci/peci-dev.c | 7 ++++---

 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/peci/peci-dev.c +0 -0
index e0fe09467a80..84e90af81ccc 100644

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, · line 126

Quoted 7 lines
drivers/peci/peci-dev.c +2 -2
@@ -122,8 +122,8 @@ static long peci_dev_ioctl(struct file *file, 
 uint iocmd, ulong arg)
          }
          xmsg = peci_get_xfer_msg(uxmsg.tx_len, uxmsg.rx_len);
-        if (IS_ERR(xmsg)) {
-            ret = PTR_ERR(xmsg);
+        if (!xmsg) {
+            ret = -ENOMEM;

Yes, it's a right fix. Thanks!

Quoted 12 lines

drivers/peci/peci-dev.c static long peci_dev_ioctl(struct file *file, · line 166

Quoted 9 lines
drivers/peci/peci-dev.c +2 -1
              break;
          }
@@ -162,7 +162,8 @@ static long peci_dev_ioctl(struct file *file, 
 uint iocmd, ulong arg)
      }
      peci_put_xfer_msg(xmsg);
-    kfree(msg);
+    if (!IS_ERR(msg))
+        kfree(msg);

Not needed. kfree itself has null pointer checking inside.

Quoted 1 line

Certainly, but the condition in question here isn't whether it's NULL, but whether it's an ERR_PTR (which as far as I can tell kfree() does not check for).  As is, there's an error path that leads to passing a non-NULL but also non-kfree-safe ERR_PTR (the memdup_user() return value).

Yeah, checked that memdup_user can also return ERR_PTR(-ENOMEM) or ERR_PTR(-EFAULT) in error cases so null checking isn't enough for calling free(). I'll add this change into PECI patch set. Thanks!

Quoted 4 lines

Zev

Reviewed-by on v1 1/2 Applied
#5 lore raw

Re: [PATCH 1/2] peci: fix error-handling in peci_dev_ioctl()

In reply to Zev Weiss
Quoted 5 lines · Zev Weiss

peci_get_xfer_msg() returns NULL on failure, not an ERR_PTR. Also avoid calling kfree() on an ERR_PTR.

Fixes 90ddc4e972b5 ("peci: Add support for PECI bus driver core")

Applied to dev-5.8.

Cheers,

Joel

Quoted 32 lines
---
 drivers/peci/peci-dev.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/peci/peci-dev.c +4 -4
index e0fe09467a80..84e90af81ccc 100644
@@ -122,8 +122,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
                }
 
                xmsg = peci_get_xfer_msg(uxmsg.tx_len, uxmsg.rx_len);
-               if (IS_ERR(xmsg)) {
-                       ret = PTR_ERR(xmsg);
+               if (!xmsg) {
+                       ret = -ENOMEM;
                        break;
                }
 
@@ -162,7 +162,8 @@ static long peci_dev_ioctl(struct file *file, uint iocmd, ulong arg)
        }
 
        peci_put_xfer_msg(xmsg);
-       kfree(msg);
+       if (!IS_ERR(msg))
+               kfree(msg);
 
        return (long)ret;
 }
--

2.28.0

author
#6 lore raw

PATCH 2/2 peci-cputemp: label CPU cores from zero instead of one

In reply to Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
 drivers/hwmon/peci-cputemp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +1 -1
index b9fe91281d58..78e442f433a7 100644
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct peci_cputemp *priv, int idx)
 	if (!priv->coretemp_label[idx])
 		return -ENOMEM;
 
-	sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+	sprintf(priv->coretemp_label[idx], "Core %d", idx);
 
 	return 0;
 }
Signature · 1 line
2.28.0
#7 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 13 lines · Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
  drivers/hwmon/peci-cputemp.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +0 -0
index b9fe91281d58..78e442f433a7 100644

drivers/hwmon/peci-cputemp.c static int create_core_temp_label(struct peci_cputemp *priv, int idx) · line 366

Quoted 5 lines
drivers/hwmon/peci-cputemp.c +1 -1
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct peci_cputemp *priv, int idx)
  	if (!priv->coretemp_label[idx])
  		return -ENOMEM;
 
-	sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+	sprintf(priv->coretemp_label[idx], "Core %d", idx);

Differently from low level indexing, it's labeling for users and it should be synced with other temp or ADC sensors such as

PVCCIN CPU1
PVDQ ABC CPU1
CPU1 P12V PVCCIN
CPU1 VR Mem ABCD
CPU1 VR P1V8

These are using indexes starting from '1'.

Thanks,
Jae

Quoted 3 lines
drivers/hwmon/peci-cputemp.c +0 -0
  	return 0;
  }
 
author
#8 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 24 lines · Jae Hyun Yoo
Quoted 13 lines · Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
 drivers/hwmon/peci-cputemp.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +0 -0
index b9fe91281d58..78e442f433a7 100644

drivers/hwmon/peci-cputemp.c static int create_core_temp_label(struct peci_cputemp *priv, int idx) · line 365

Quoted 4 lines
drivers/hwmon/peci-cputemp.c +1 -1
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct peci_cputemp *priv, int idx)
 	if (!priv->coretemp_label[idx])
 		return -ENOMEM;
-	sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+	sprintf(priv->coretemp_label[idx], "Core %d", idx);

Differently from low level indexing, it's labeling for users and it should be synced with other temp or ADC sensors such as

Quoted 8 lines

PVCCIN CPU1
PVDQ ABC CPU1
CPU1 P12V PVCCIN
CPU1 VR Mem ABCD
CPU1 VR P1V8

These are using indexes starting from '1'.

OK, if it's for consistency with other existing drivers I suppose that's reasonable, though for my own reference, could you point me to where those are implemented? Some rough grepping around the source tree didn't appear to turn up anything relevant.

Thanks,
Zev

#9 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 32 lines · Zev Weiss
Quoted 31 lines · Jae Hyun Yoo
Quoted 13 lines · Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
 drivers/hwmon/peci-cputemp.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +0 -0
index b9fe91281d58..78e442f433a7 100644

drivers/hwmon/peci-cputemp.c static int create_core_temp_label(struct · line 366

Quoted 5 lines
drivers/hwmon/peci-cputemp.c +1 -1
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct 
 peci_cputemp *priv, int idx)
      if (!priv->coretemp_label[idx])
          return -ENOMEM;
-    sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+    sprintf(priv->coretemp_label[idx], "Core %d", idx);

Differently from low level indexing, it's labeling for users and it should be synced with other temp or ADC sensors such as

PVCCIN CPU1
PVDQ ABC CPU1
CPU1 P12V PVCCIN
CPU1 VR Mem ABCD
CPU1 VR P1V8

Quoted 4 lines
Quoted 2 lines

These are using indexes starting from '1'.

OK, if it's for consistency with other existing drivers I suppose that's reasonable, though for my own reference, could you point me to where those are implemented?  Some rough grepping around the source tree didn't appear to turn up anything relevant.

Sensor names get assigned through these services
https://github.com/openbmc/entity-manager
https://github.com/openbmc/dbus-sensors

and it depends on board configuration of each machine.

Quoted 3 lines

Thanks,
Zev

author
#10 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 36 lines · Jae Hyun Yoo
Quoted 35 lines · Zev Weiss
Quoted 33 lines · Jae Hyun Yoo
Quoted 13 lines · Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
 drivers/hwmon/peci-cputemp.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +0 -0
index b9fe91281d58..78e442f433a7 100644

drivers/hwmon/peci-cputemp.c static int create_core_temp_label(struct · line 366

Quoted 5 lines
drivers/hwmon/peci-cputemp.c +1 -1
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct 
 peci_cputemp *priv, int idx)
      if (!priv->coretemp_label[idx])
          return -ENOMEM;
-    sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+    sprintf(priv->coretemp_label[idx], "Core %d", idx);

Differently from low level indexing, it's labeling for users and it should be synced with other temp or ADC sensors such as

PVCCIN CPU1
PVDQ ABC CPU1
CPU1 P12V PVCCIN
CPU1 VR Mem ABCD
CPU1 VR P1V8

These are using indexes starting from '1'.

Quoted 8 lines
Quoted 1 line

OK, if it's for consistency with other existing drivers I suppose that's reasonable, though for my own reference, could you point me to where those are implemented?  Some rough grepping around the source tree didn't appear to turn up anything relevant.

Sensor names get assigned through these services
https://github.com/openbmc/entity-manager
https://github.com/openbmc/dbus-sensors

and it depends on board configuration of each machine.

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling. But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs. So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

Thanks,
Zev

#11 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 43 lines · Zev Weiss
Quoted 42 lines · Jae Hyun Yoo
Quoted 32 lines · Zev Weiss
Quoted 31 lines · Jae Hyun Yoo
Quoted 13 lines · Zev Weiss

Zero-based numbering is more consistent with all other cpu/core numbering I'm aware of (including the PECI spec).

---
 drivers/hwmon/peci-cputemp.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/hwmon/peci-cputemp.c +0 -0
index b9fe91281d58..78e442f433a7 100644

drivers/hwmon/peci-cputemp.c static int create_core_temp_label(struct · line 366

Quoted 5 lines
drivers/hwmon/peci-cputemp.c +1 -1
@@ -363,7 +363,7 @@ static int create_core_temp_label(struct 
 peci_cputemp *priv, int idx)
      if (!priv->coretemp_label[idx])
          return -ENOMEM;
-    sprintf(priv->coretemp_label[idx], "Core %d", idx + 1);
+    sprintf(priv->coretemp_label[idx], "Core %d", idx);

Differently from low level indexing, it's labeling for users and it should be synced with other temp or ADC sensors such as

PVCCIN CPU1
PVDQ ABC CPU1
CPU1 P12V PVCCIN
CPU1 VR Mem ABCD
CPU1 VR P1V8

Quoted 4 lines
Quoted 2 lines

These are using indexes starting from '1'.

OK, if it's for consistency with other existing drivers I suppose that's reasonable, though for my own reference, could you point me to where those are implemented?  Some rough grepping around the source tree didn't appear to turn up anything relevant.

Sensor names get assigned through these services
https://github.com/openbmc/entity-manager
https://github.com/openbmc/dbus-sensors

Quoted 6 lines
Quoted 2 lines

and it depends on board configuration of each machine.

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling.  But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs.  So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

Thanks,
Jae

author
#12 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 11 lines · Jae Hyun Yoo
Quoted 3 lines

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling.  But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs.  So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

[Re-adding CCs...]

Well, as I see it basically just consistency with a larger set of things. Most other related numbering schemes I'm aware of are zero-based -- userspace tools like 'taskset' and 'lscpu', system APIs like the <sched.h> CPU_SET() routines, and the kernel's own numbering (e.g. what's shown in /proc/cpuinfo) all number processors starting from zero, so dbus-sensors seems kind of like the odd one out there. (Personally I'd be fully in support of changing it to be zero-based as well, though I have no idea offhand about how distruptive a change that would be.)

It also seems pretty OpenBMC-specific, whereas I'd expect we want to aim for greater generality in things going into mainline.

Thanks,
Zev

#13 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 13 lines · Zev Weiss
Quoted 11 lines · Jae Hyun Yoo
Quoted 3 lines

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling. But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs.  So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

Quoted 5 lines

[Re-adding CCs...]

Well, as I see it basically just consistency with a larger set of things.  Most other related numbering schemes I'm aware of are zero-based -- userspace tools like 'taskset' and 'lscpu', system APIs like the <sched.h> CPU_SET() routines, and the kernel's own numbering (e.g. what's shown in /proc/cpuinfo) all number processors starting from zero, so dbus-sensors seems kind of like the odd one out there. (Personally I'd be fully in support of changing it to be zero-based as well, though I have no idea offhand about how distruptive a change that would be.)

It also seems pretty OpenBMC-specific, whereas I'd expect we want to aim for greater generality in things going into mainline.

First of all, it's for labeling of sysfs interfaces in hwmon subsystem which monitors target CPUs, not for local CPUs you mentioned. As you can see in hwmon subsystem, property indexing starts from 1. Also, we have used this natural number indexing traditionally for all Intel BMCs we made so far, and it's also for keeping product value against misreading of actual core numbers. If you don't have any critical or blocking issue, I need to keep the current indexing as is.

Thanks,
Jae

#14 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

In reply to Zev Weiss
Quoted 22 lines · Zev Weiss
Quoted 12 lines · Jae Hyun Yoo
Quoted 11 lines

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling. But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs. So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

Quoted 7 lines

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

Quoted 1 line

[Re-adding CCs...]

Thanks. Please keep the discussion on the list.

Quoted 11 lines

Well, as I see it basically just consistency with a larger set of things. Most other related numbering schemes I'm aware of are zero-based -- userspace tools like 'taskset' and 'lscpu', system APIs like the <sched.h> CPU_SET() routines, and the kernel's own numbering (e.g. what's shown in /proc/cpuinfo) all number processors starting from zero, so dbus-sensors seems kind of like the odd one out there. (Personally I'd be fully in support of changing it to be zero-based as well, though I have no idea offhand about how distruptive a change that would be.)

Quoted 2 lines

It also seems pretty OpenBMC-specific, whereas I'd expect we want to aim for greater generality in things going into mainline.

Agreed. The hwmon numbering varies; some attributes are zero indexed and some start at 1. More commonly we start counting from zero in the kernel, so I would expect PECI to do the same.

If there's some userspace that depends on the behaviour of these out of tree PECI patches, then that userspace will need to change. This reminds us why the project prefers patches exposing userspace ABI are merged to mainline first.

Cheers,

Joel

Reviewed-by
#15 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Hi Zev,

Quoted 41 lines · Joel Stanley
Quoted 22 lines · Zev Weiss
Quoted 12 lines · Jae Hyun Yoo
Quoted 11 lines

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling. But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs. So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

Quoted 7 lines

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

Quoted 1 line

[Re-adding CCs...]

Thanks. Please keep the discussion on the list.

Quoted 11 lines

Well, as I see it basically just consistency with a larger set of things. Most other related numbering schemes I'm aware of are zero-based -- userspace tools like 'taskset' and 'lscpu', system APIs like the <sched.h> CPU_SET() routines, and the kernel's own numbering (e.g. what's shown in /proc/cpuinfo) all number processors starting from zero, so dbus-sensors seems kind of like the odd one out there. (Personally I'd be fully in support of changing it to be zero-based as well, though I have no idea offhand about how distruptive a change that would be.)

Quoted 2 lines

It also seems pretty OpenBMC-specific, whereas I'd expect we want to aim for greater generality in things going into mainline.

Quoted 8 lines

Agreed. The hwmon numbering varies; some attributes are zero indexed and some start at 1. More commonly we start counting from zero in the kernel, so I would expect PECI to do the same.

If there's some userspace that depends on the behaviour of these out of tree PECI patches, then that userspace will need to change. This reminds us why the project prefers patches exposing userspace ABI are merged to mainline first.

Okay. Not a big deal. The coretemp module for local CPU also uses zero starting label index for core numbers so better match up. Thanks for your patch.

Applied
#16 lore raw

Re: [PATCH 2/2] peci-cputemp: label CPU cores from zero instead of one

Quoted 54 lines · Jae Hyun Yoo

Hi Zev,

Quoted 41 lines · Joel Stanley
Quoted 22 lines · Zev Weiss
Quoted 12 lines · Jae Hyun Yoo
Quoted 11 lines

Oh I see -- I had thought you were referring to other existing hwmon drivers in the kernel.

As far as I can tell, all those instances appear to be numbering CPU *sockets* though -- which as Jason mentioned in a call earlier today I gather is done to line up with motherboard silkscreen labeling. But in the code in question here we're labeling *cores* within a given socket, which I don't see arising anywhere in any existing entity-manager configs. So I'm still unclear on why we want to use one-based indexing here instead of zero-based -- I'd think we'd want the PECI driver to match the PECI spec?

Quoted 7 lines

PECI driver uses zero-based index for PECI command handling but label is user facing stuff which shouldn't make confusion to users. We can modify driver like you did in this patch and previous driver also used zero-based indexing but I changed it to natural number based indexing to avoid confusion between driver labels and dbus-sensors names. Any specific reason for the zero-based indexing? Any benefit?

Quoted 1 line

[Re-adding CCs...]

Thanks. Please keep the discussion on the list.

Quoted 11 lines

Well, as I see it basically just consistency with a larger set of things. Most other related numbering schemes I'm aware of are zero-based -- userspace tools like 'taskset' and 'lscpu', system APIs like the <sched.h> CPU_SET() routines, and the kernel's own numbering (e.g. what's shown in /proc/cpuinfo) all number processors starting from zero, so dbus-sensors seems kind of like the odd one out there. (Personally I'd be fully in support of changing it to be zero-based as well, though I have no idea offhand about how distruptive a change that would be.)

Quoted 2 lines

It also seems pretty OpenBMC-specific, whereas I'd expect we want to aim for greater generality in things going into mainline.

Quoted 8 lines

Agreed. The hwmon numbering varies; some attributes are zero indexed and some start at 1. More commonly we start counting from zero in the kernel, so I would expect PECI to do the same.

If there's some userspace that depends on the behaviour of these out of tree PECI patches, then that userspace will need to change. This reminds us why the project prefers patches exposing userspace ABI are merged to mainline first.

Quoted 5 lines

Okay. Not a big deal. The coretemp module for local CPU also uses zero starting label index for core numbers so better match up. Thanks for your patch.

Applied to dev-5.8.

Cheers,

Joel

#17 lore raw
In reply to Zev Weiss

Hello Zev,

Quoted 10 lines · Zev Weiss

[Re-sending as I clumsily typoed the mailing list's address on the first attempt; apologies for the duplication.]

Hello,

These patches address a few small things I noticed with the PECI patch set currently in the OpenBMC kernel tree. Assuming they're deemed acceptable, I'd of course hope they get folded in to the version of the PECI code that gets upstreamed into the mainline kernel.

Quoted 4 lines

(Please let me know if there's some other way I should send this -- I looked for but couldn't find a gerrit repo for the obmc kernel or any kernel-specific patch submission instructions in openbmc/docs/CONTRIBUTING.md, so 'git send-email' is my best guess.)

This mailing list would be a right place as long as the PECI patch set is in OpenBMC kernel tree but in case if it is removed from the tree later (probably from the 5.9 kernel), you may need to use pull request through https://github.com/Intel-BMC/openbmc when it replaces the kernel repo with forked kernel.

Thanks,
Jae

Quoted 5 lines

Thanks,
Zev