Skip to contentSkip to conversation

Replace -EINVAL with PCIBIOS_BAD_REGISTER_NUMBER

PATCH In review Bolarinwa Olayemi Saheed 3 participants 7 messages Apr 9, 2020 – Apr 24, 2020 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Replace -EINVAL with PCIBIOS_BAD_REGISTER_NUMBER

---
 drivers/pci/access.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +2 -2
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
 	*val = 0;
 	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
 	*val = 0;
 	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);
Signature · 1 line
2.17.1
#1 lore raw

Hi Bolarinwa,

I notice some drivers use these functions and if there is an error, pass the error code directly to the userspace. As it's our private error code, is it appropriate to pass or should we call pcibios_err_to_errno()(include/linux/pci.h, line 672) to do the conversion?

Regards,
Yicong

Quoted 28 lines · Bolarinwa Olayemi Saheed
---
 drivers/pci/access.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +2 -2
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
 	*val = 0;
 	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
 	*val = 0;
 	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);
#2 lore raw
Quoted 5 lines · Yicong Yang

Hi Bolarinwa,

I notice some drivers use these functions and if there is an error, pass the error code directly to the userspace. As it's our private error code, is it appropriate to pass or should we call pcibios_err_to_errno()(include/linux/pci.h, line 672) to do the conversion?

Hello Yicong,

Thank you for pointing out that function to me. I have resent the patch.

Saheed

Quoted 32 lines

Regards,
Yicong

Quoted 28 lines · Bolarinwa Olayemi Saheed
---
  drivers/pci/access.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +2 -2
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
  	*val = 0;
  	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
  	if (pcie_capability_reg_implemented(dev, pos)) {
  		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
  	*val = 0;
  	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
  	if (pcie_capability_reg_implemented(dev, pos)) {
  		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);
#3 lore raw
In reply to Yicong Yang
Quoted 7 lines · Yicong Yang

Hi Bolarinwa,

I notice some drivers use these functions and if there is an error, pass the error code directly to the userspace. As it's our private error code, is it appropriate to pass or should we call pcibios_err_to_errno()(include/linux/pci.h, line 672) to do the conversion?

The whole point of this is to make the return values of the pcie_capability_{read,write,etc}*() functions work the same as the pci_{read,write}_config*() functions.

The latter return PCIBIOS_* error codes, so the former should as well.

When we do this, we do need to audit every caller of the pcie_capability_{read,write}*() functions to make sure we don't break them. If some callers pass the error code directly to userspace, they may need some change.

Yicong, can you point to the ones you noticed so Saheed can check them out?

Bjorn

Quoted 30 lines · Bolarinwa Olayemi Saheed
Quoted 28 lines
---
 drivers/pci/access.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +2 -2
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
 	*val = 0;
 	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
 	*val = 0;
 	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);
#4 lore raw

Hi Bjorn and Saheed,

Callers use return value(most callers even don't check) of pcie_capability_{read,write}_*() I found lists below. some may directly print them to dmesg, others return the error codes to its caller. I think we should do the conversion in both condition.

- pcie_speeds() in drivers/infiniband/hw/hfi1/pcie.c, line 306
- amd_ntb_get_link_status() in drivers/ntb/hw/amd/ntb_hw_amd.c, line 216, line 233

the probably change may look like:

    ret = pcie_capability_{read, write}_*();
    if (ret)
        return pcibios_err_to_errno(ret);

However, pci_{read, write}_config*() also have such problem, as they are also used widely outside pci driver and these drivers don't do the conversion. for example in arch/x86/platform/intel/iosf_mbi.c, iosf_mbi_pci_read_mdr() at line 39:

    result = pci_read_config_dword();
    if (result < 0)
        goto fail_read;

Seems it'll nevet get a failure result. Perhaps another patch is needed to solve these issues.

AS PCIBIOS_* error code canbe *equivalent* to generic error code, why can't we directly use the generic ones? Considering of compatibility, maybe possible change will be like:

- #define PCIBIOS_FUNC_NOT_SUPPORTED 0X81
+ #define PCIBIOS_FUNC_NOT_SUPPORTED -ENOENT
......

and pcibios_err_to_errno() is not neccessary any more.

I don't know why we didn't use generic error code and define positive private errors. Please tell me if there is any background.

Regards,
Yicong

Quoted 55 lines · Bjorn Helgaas
Quoted 7 lines · Yicong Yang

Hi Bolarinwa,

I notice some drivers use these functions and if there is an error, pass the error code directly to the userspace. As it's our private error code, is it appropriate to pass or should we call pcibios_err_to_errno()(include/linux/pci.h, line 672) to do the conversion?

The whole point of this is to make the return values of the pcie_capability_{read,write,etc}*() functions work the same as the pci_{read,write}_config*() functions.

The latter return PCIBIOS_* error codes, so the former should as well.

When we do this, we do need to audit every caller of the pcie_capability_{read,write}*() functions to make sure we don't break them. If some callers pass the error code directly to userspace, they may need some change.

Yicong, can you point to the ones you noticed so Saheed can check them out?

Bjorn

Quoted 22 lines · Bolarinwa Olayemi Saheed
Quoted 21 lines
---
 drivers/pci/access.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +1 -1
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
 	*val = 0;
 	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
Quoted 7 lines

drivers/pci/access.c int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val) · line 450

Quoted 7 lines
drivers/pci/access.c +1 -1
 	*val = 0;
 	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
 	if (pcie_capability_reg_implemented(dev, pos)) {
 		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);

.

#5 lore raw

Hello Bjorn and Yicong

Quoted 4 lines · Yicong Yang

Hi Bjorn and Saheed,

Callers use return value(most callers even don't check) of pcie_capability_{read,write}_*() I found lists below. some

In this patch I have focused only on pcie_capability_read_{word & dword}()

Do you think the consistency issue is of concern also with pcie_capability_write_*()?

Quoted 6 lines

may directly print them to dmesg, others return the error codes to its caller. I think we should do the conversion in both condition.

- pcie_speeds() in drivers/infiniband/hw/hfi1/pcie.c, line 306
- amd_ntb_get_link_status() in drivers/ntb/hw/amd/ntb_hw_amd.c, line 216, line 233

Thank you Yicong for helping me with this. I have also search for functions that

either persist the return value of pcie_capability_read_*() or return it directly.

I have sent a RFC Patch with a report of my audit of the functions I found. I will

appreciate your comment.

- Saheed

Quoted 89 lines

the probably change may look like:

     ret = pcie_capability_{read, write}_*();
     if (ret)
         return pcibios_err_to_errno(ret);

However, pci_{read, write}_config*() also have such problem, as they are also used widely outside pci driver and these drivers don't do the conversion. for example in arch/x86/platform/intel/iosf_mbi.c, iosf_mbi_pci_read_mdr() at line 39:

     result = pci_read_config_dword();
     if (result < 0)
         goto fail_read;

Seems it'll nevet get a failure result. Perhaps another patch is needed to solve these issues.

AS PCIBIOS_* error code canbe *equivalent* to generic error code, why can't we directly use the generic ones? Considering of compatibility, maybe possible change will be like:

- #define PCIBIOS_FUNC_NOT_SUPPORTED 0X81
+ #define PCIBIOS_FUNC_NOT_SUPPORTED -ENOENT
......

and pcibios_err_to_errno() is not neccessary any more.

I don't know why we didn't use generic error code and define positive private errors. Please tell me if there is any background.

Regards,
Yicong

Quoted 55 lines · Bjorn Helgaas
Quoted 7 lines · Yicong Yang

Hi Bolarinwa,

I notice some drivers use these functions and if there is an error, pass the error code directly to the userspace. As it's our private error code, is it appropriate to pass or should we call pcibios_err_to_errno()(include/linux/pci.h, line 672) to do the conversion?

The whole point of this is to make the return values of the pcie_capability_{read,write,etc}*() functions work the same as the pci_{read,write}_config*() functions.

The latter return PCIBIOS_* error codes, so the former should as well.

When we do this, we do need to audit every caller of the pcie_capability_{read,write}*() functions to make sure we don't break them. If some callers pass the error code directly to userspace, they may need some change.

Yicong, can you point to the ones you noticed so Saheed can check them out?

Bjorn

Quoted 22 lines · Bolarinwa Olayemi Saheed
Quoted 21 lines
---
  drivers/pci/access.c | 4 ++--
  1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/access.c +1 -1
index 79c4a2ef269a..451f2b8b2b3c 100644
@@ -409,7 +409,7 @@ int pcie_capability_read_word(struct pci_dev *dev, int pos, u16 *val)
 
  	*val = 0;
  	if (pos & 1)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
  	if (pcie_capability_reg_implemented(dev, pos)) {
  		ret = pci_read_config_word(dev, pci_pcie_cap(dev) + pos, val);
@@ -444,7 +444,7 @@ int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val)
 
Quoted 7 lines

drivers/pci/access.c int pcie_capability_read_dword(struct pci_dev *dev, int pos, u32 *val) · line 450

Quoted 7 lines
drivers/pci/access.c +1 -1
  	*val = 0;
  	if (pos & 3)
-		return -EINVAL;
+		return PCIBIOS_BAD_REGISTER_NUMBER;
 
  	if (pcie_capability_reg_implemented(dev, pos)) {
  		ret = pci_read_config_dword(dev, pci_pcie_cap(dev) + pos, val);

.

#6 lore raw
Quoted 5 lines · Saheed Bolarinwa

Hello Bjorn and Yicong

Quoted 2 lines · Yicong Yang

Hi Bjorn and Saheed,

Quoted 7 lines
Quoted 2 lines

Callers use return value(most callers even don't check) of pcie_capability_{read,write}_*() I found lists below. some

In this patch I have focused only on pcie_capability_read_{word & dword}()

Do you think the consistency issue is of concern also with pcie_capability_write_*()?

Yes. pcie_capability_write_*() can return either -EINVAL or PCIBIOS_*, so it has the same consistency problem. I'd like to fix all of pcie_capability_read_*() and pcie_capability_write_*() in a single patch because it's logically the same change and they can all be reviewed together.

Thanks for the audit of the callers. I'd include a brief summary in the commit log, but not the entire thing because it's a little too detailed for the log. But it could be useful in a cover letter (which will be acccessible from the commit via the "Link:
https://lore.kernel.org/r/..." URL).

Bjorn