PCI: pci_restore_state() is returning 0 when it fails
Actually pci_restore_state() is returning 0 if the restore process fails, instead of a error value.
If it fails, I believe that it should return -EPERM, once that it is an invalid operation and probably pci_save_state() wasn't called.
--- drivers/pci/pci.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)
drivers/pci/pci.c +1 -1
@@ -857,7 +857,7 @@ pci_restore_state(struct pci_dev *dev)u32 val;if (!dev->state_saved) -return 0 ; +return -EPERM ;/* PCI Express register must be restored first */ pci_restore_pcie_state (dev);
Signature · 1 line
1.6.0.4
Quoted 6 lines · Breno Leitao
Actually
pci_restore_state()is returning 0 if the restore process fails, instead of a error value.If it fails, I believe that it should return
-EPERM, once that it is an invalid operation and probablypci_save_state()wasn't called.
I believe this patch will break a number of things.
Does it actually fix any problem you have observed?
Rafael
Quoted 19 lines
Signed-off-by Breno Leitao <[email protected]>--- drivers/pci/pci.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)drivers/pci/pci.c +1
-1@@ -857,7 +857,7 @@ pci_restore_state(struct pci_dev *dev)u32 val;if (!dev->state_saved) -return 0 ; +return -EPERM ;/* PCI Express register must be restored first */ pci_restore_pcie_state (dev);
Hi Rafael,
Quoted 9 lines · Rafael J. Wysocki
Quoted 6 lines · Breno Leitao
Actually
pci_restore_state()is returning 0 if the restore process fails, instead of a error value.If it fails, I believe that it should return
-EPERM, once that it is an invalid operation and probablypci_save_state()wasn't called.I believe this patch will break a number of things.
Well, I checked it, and found that there are around 10 places that really verify the return value for this function, and almost all of them do the correct thing, and the patch doesn't seem to break any of them except a specific case in the drivers/net/sfc/falcon.c file, that contains:
if (FALCON_IS_DUAL_FUNC(efx)) {
rc = pci_restore_state(nic_data->pci_dev2);
if (rc) {
EFX_ERR(efx, "failed to restore PCI config for "
"the secondary function\n");
goto fail3;
}
}
rc = pci_restore_state(efx->pci_dev);
if (rc) {
EFX_ERR(efx, "failed to restore PCI config for the "
"primary function\n");
goto fail4;In this case, if FALCON_IS_DUAL_FUNC(efx) returns true, then the next pci_restore_state(efx->pci_dev) will return -1, and cause the "failed to restore PCI config for the primary function" error.
That's because the code is calling pci_restore_state() twice without calling pci_save_state() in the middle.
Since this seems to be the only place that will be broken, and the fix is trivial, I believe that the patch can be applied smoothly.
Quoted 1 line
Does it actually fix any problem you have observed?
No, but we use this function to resume drivers after PPC machines detects errors (EEH). And before your patch, we basically save the state once (in the init function), and then call pci_restore_state() every time the machine detects an error. After your patch, it's not possible anymore, because pci_restore_state() will not restore the state after the first restore. So, I'd like to instrument some drivers to check if it's possible to recover or not.
I also believe that returning 0 for a failed function isn't a good plan.
Thanks for the review,
Breno
Quoted 16 lines · Breno Leitao
Hi Rafael,
Quoted 9 lines · Rafael J. Wysocki
Quoted 6 lines · Breno Leitao
Actually
pci_restore_state()is returning 0 if the restore process fails, instead of a error value.If it fails, I believe that it should return
-EPERM, once that it is an invalid operation and probablypci_save_state()wasn't called.I believe this patch will break a number of things.
Well, I checked it, and found that there are around 10 places that really verify the return value for this function, and almost all of them do the correct thing, and the patch doesn't seem to break any of them except a specific case in the
drivers/net/sfc/falcon.cfile, that contains:
[...]
Quoted 4 lines
That's because the code is calling
pci_restore_state()twice without callingpci_save_state()in the middle.
Since this seems to be the only place that will be broken, and the fix is trivial, I believe that the patch can be applied smoothly.
[...]
This code supports two similar PCI devices, one of which has a second function that is not truly independent. For that chip it saves and restores both functions' config space. So far as I know, there are no cases where it fails to match save and restore.
Ben.
Signature · 3 lines
Ben Hutchings, Senior Software Engineer, Solarflare Communications Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked.
Hi Rafael,
I didn't hear back after the analysis that there is no regression after this patch. Did you have a chance to think about this patch ?
Thanks
Breno
Quoted 30 lines · Ben Hutchings
Quoted 15 lines · Breno Leitao
Hi Rafael,
Quoted 8 lines · Rafael J. Wysocki
Quoted 6 lines · Breno Leitao
Actually
pci_restore_state()is returning 0 if the restore process fails, instead of a error value.If it fails, I believe that it should return
-EPERM, once that it is an invalid operation and probablypci_save_state()wasn't called.I believe this patch will break a number of things.
Well, I checked it, and found that there are around 10 places that really verify the return value for this function, and almost all of them do the correct thing, and the patch doesn't seem to break any of them except a specific case in the
drivers/net/sfc/falcon.cfile, that contains:[...]
Quoted 4 lines
That's because the code is calling
pci_restore_state()twice without callingpci_save_state()in the middle.
Since this seems to be the only place that will be broken, and the fix is trivial, I believe that the patch can be applied smoothly.[...]
This code supports two similar PCI devices, one of which has a second function that is not truly independent. For that chip it saves and restores both functions' config space. So far as I know, there are no cases where it fails to match save and restore.
Ben.
Quoted 4 lines · Breno Leitao
Hi Rafael,
I didn't hear back after the analysis that there is no regression after this patch. Did you have a chance to think about this patch ?
I'm not really sure why you insist that we take it. Does it solve any problem that you're observing in practice? If so, what's your test case.
Rafael