Message ID | 1373356564-21668-1-git-send-email-wangyijing@huawei.com (mailing list archive) |
---|---|
State | New, archived |
Delegated to: | Bjorn Helgaas |
Headers | show |
On Tue, 2013-07-09 at 15:56 +0800, Yijing Wang wrote: > Currently, pciehp_resume will call pciehp_enable_slot() to add > device if there is a device in the slot. But if the device was > present before suspend, it's no necessary to add again. Now in > such case, there is some uncomfortable message like > > pciehp 0000:00:1c.1:pcie04: Device 0000:03:00.0 already exists at 0000:03:00, cannot hot-add > pciehp 0000:00:1c.1:pcie04: Cannot add device at 0000:03:00 > > This problem was reported by Paul Bolle <pebolle@tiscali.nl> > The discussion link: http://comments.gmane.org/gmane.linux.kernel.pci/19876 > > We can use PCIe Device Serial Number to identify the device if > device support DSN. > > currently: > 1. slot is empty before suspend, insert card during suspend. > In this case, is safe, pciehp will add device by check adapter > status register in pciehp_resume. > > 2. slot is non empty before suspend, remove card during suspend. > Also be safe, pciehp will remove device by check adapter > status register in pciehp_resume. > > 3. slot is non empty before suspend, remove card during suspend > and insert a new card. > Now pciehp just call pciehp_enable_slot() roughly. We should > remove the old card firstly, then add the new card. > > 4. slot is non empty before suspend, no action during suspend. > We should do nothing in pciehp_resume, but we call > pciehp_enable_slot(), so some uncomfortable messages show like above. > In this case, we can improve it a little by add a guard > if (!list_empty(bus->devices)). Great! I'm currently trying to bisect another problem, but hope to test this patch (and the preceding patch it apparently needs) in a few days. Please feel free to prod me if you think testing is needed but I'm taking to long to report back. Thanks! Paul Bolle -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
>> In this case, we can improve it a little by add a guard >> if (!list_empty(bus->devices)). > > Great! > > I'm currently trying to bisect another problem, but hope to test this > patch (and the preceding patch it apparently needs) in a few days. > Please feel free to prod me if you think testing is needed but I'm > taking to long to report back. Hi Paul, Thank you for helping test. Because I have no machine to test this case, so your test report is very important, it can help us going in the right way to solve this problem. Thanks! Yijing. > > > . >
On Tue, Jul 9, 2013 at 1:56 AM, Yijing Wang <wangyijing@huawei.com> wrote: > Currently, pciehp_resume will call pciehp_enable_slot() to add > device if there is a device in the slot. But if the device was > present before suspend, it's no necessary to add again. Now in > such case, there is some uncomfortable message like > > pciehp 0000:00:1c.1:pcie04: Device 0000:03:00.0 already exists at 0000:03:00, cannot hot-add > pciehp 0000:00:1c.1:pcie04: Cannot add device at 0000:03:00 > > This problem was reported by Paul Bolle <pebolle@tiscali.nl> > The discussion link: http://comments.gmane.org/gmane.linux.kernel.pci/19876 > > We can use PCIe Device Serial Number to identify the device if > device support DSN. I think I like the idea of this, especially because the Microsoft PCI Hardware Compliance Test apparently requires DSN for hot-pluggable PCIe devices [1], so it should be pretty universal. [1] http://www.techtalkz.com/microsoft-device-drivers-dtm/341362-dtm-pcihct-test-violates-pci-express-base-specification-revision-1-a.html > currently: > 1. slot is empty before suspend, insert card during suspend. > In this case, is safe, pciehp will add device by check adapter > status register in pciehp_resume. Your patch doesn't change anything here. > 2. slot is non empty before suspend, remove card during suspend. > Also be safe, pciehp will remove device by check adapter > status register in pciehp_resume. Your patch doesn't change anything here. (But I think the driver .remove() method will try to poke at the non-existent device; see below.) > 3. slot is non empty before suspend, remove card during suspend > and insert a new card. > Now pciehp just call pciehp_enable_slot() roughly. We should > remove the old card firstly, then add the new card. With your patch, I think we'll call the old driver's .remove() method on the new device. This seems bad; see below. With your patch, if we remove and reinsert the same device while suspended, we do nothing because the DSN didn't change. Previously we called pciehp_enable_slot(). I don't know if we need to do anything here or not. > 4. slot is non empty before suspend, no action during suspend. > We should do nothing in pciehp_resume, but we call > pciehp_enable_slot(), so some uncomfortable messages show like above. > In this case, we can improve it a little by add a guard > if (!list_empty(bus->devices)). This is the common case. Previously we called pciehp_enable_slot(), and with your patch we do nothing. I think that seems sensible, but this part should be split into a separate patch. That way we can keep the benefit of this change even if we trip over something with the other changes. > Reported-by: Paul Bolle <pebolle@tiscali.nl> > Signed-off-by: Yijing Wang <wangyijing@huawei.com> > Cc: Paul Bolle <pebolle@tiscali.nl> > Cc: "Rafael J. Wysocki" <rjw@sisk.pl> > Cc: Oliver Neukum <oneukum@suse.de> > Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> > Cc: linux-pci@vger.kernel.org > --- > drivers/pci/hotplug/pciehp_core.c | 38 +++++++++++++++++++++++++++++++++--- > 1 files changed, 34 insertions(+), 4 deletions(-) > > diff --git a/drivers/pci/hotplug/pciehp_core.c b/drivers/pci/hotplug/pciehp_core.c > index 7d72c5e..d01e093 100644 > --- a/drivers/pci/hotplug/pciehp_core.c > +++ b/drivers/pci/hotplug/pciehp_core.c > @@ -291,6 +291,28 @@ static void pciehp_remove(struct pcie_device *dev) > } > > #ifdef CONFIG_PM > + > +/* If device support Device Serial Numner, use DSN s/support/supports/ s/Numner/Number/ Use conventional comment style: /* * If device ... */ > + * to identify the device > + */ > +static bool device_in_slot_is_changed(struct pci_bus *pbus) > +{ > + u64 old_dsn, new_dsn; > + struct pci_dev *pdev; > + > + pdev = pci_get_slot(pbus, PCI_DEVFN(0, 0)); pci_get_slot() can fail. > + old_dsn = pdev->sn; > + > + /* get func 0 device serial number */ > + pci_get_dsn(pdev, &new_dsn); > + pci_dev_put(pdev); > + > + if (old_dsn != new_dsn) > + return true; > + > + return false; > +} > + > static int pciehp_suspend (struct pcie_device *dev) > { > return 0; > @@ -300,6 +322,7 @@ static int pciehp_resume (struct pcie_device *dev) > { > struct controller *ctrl; > struct slot *slot; > + struct pci_bus *pbus = dev->port->subordinate; > u8 status; > > ctrl = get_service_data(dev); > @@ -311,10 +334,17 @@ static int pciehp_resume (struct pcie_device *dev) > > /* Check if slot is occupied */ > pciehp_get_adapter_status(slot, &status); > - if (status) > - pciehp_enable_slot(slot); > - else > - pciehp_disable_slot(slot); > + if (status) { > + if (list_empty(&pbus->devices)) > + pciehp_enable_slot(slot); > + else if (device_in_slot_is_changed(pbus)) { > + pciehp_disable_slot(slot); pciehp_disable_slot() ultimately calls the .remove() method for the device that has already been removed. That method is likely to poke at the device, which will do something unexpected because the new device is likely to be completely different. I think the call path is this: pciehp_resume pciehp_disable_slot remove_board pciehp_unconfigure_device pci_stop_and_remove_bus_device pci_stop_bus_device pci_stop_dev device_del bus_remove_device device_release_driver __device_release_driver pci_device_remove # pci_bus_type.remove dev->driver->remove I think we already had this problem for case 2 (card removed while suspended), but at least in that case, the driver .remove() method doesn't have a device to poke at, so it's less likely to do something bad. > + pciehp_enable_slot(slot); > + } > + } else { > + if (!list_empty(&pbus->devices)) > + pciehp_disable_slot(slot); > + } > return 0; > } > #endif /* PM */ > -- > 1.7.1 > > -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi Bjorn, Thanks for your review and comments! >> We can use PCIe Device Serial Number to identify the device if >> device support DSN. > > I think I like the idea of this, especially because the Microsoft PCI > Hardware Compliance Test apparently requires DSN for hot-pluggable > PCIe devices [1], so it should be pretty universal. > > [1] http://www.techtalkz.com/microsoft-device-drivers-dtm/341362-dtm-pcihct-test-violates-pci-express-base-specification-revision-1-a.html > >> currently: >> 1. slot is empty before suspend, insert card during suspend. >> In this case, is safe, pciehp will add device by check adapter >> status register in pciehp_resume. > > Your patch doesn't change anything here. Yes, I only to make some changes for case 3/4. > >> 2. slot is non empty before suspend, remove card during suspend. >> Also be safe, pciehp will remove device by check adapter >> status register in pciehp_resume. > > Your patch doesn't change anything here. (But I think the driver > .remove() method will try to poke at the non-existent device; see > below.) I'm not sure the result of driver .remove() method to poke at the non-existent device. If driver .remove() method cannot detect the real device, remove action will be block ? If the slot support surprise hot remove, this action maybe safe. right? If the slot does not support surprise hot remove, but the device was already removed, we seem to have no other way to clean the stale data related to the old device. Now if we check adapter status in slot and found adapter is non existent, pciehp resume call pciehp_disable_slot() , in pciehp_disable_slot() function, we will check latch status, I guess this case latch is open(because slot is empty), this action will abort. But I have no platform to test it. > >> 3. slot is non empty before suspend, remove card during suspend >> and insert a new card. >> Now pciehp just call pciehp_enable_slot() roughly. We should >> remove the old card firstly, then add the new card. > > With your patch, I think we'll call the old driver's .remove() method > on the new device. This seems bad; see below. Ah, this is issue. What about power off slot first, then call the old driver's remove() method will not touch the new physical device. After the old driver's remove() cleanup, we call pciehp_enable_slot() to power on and enable the new device. > > With your patch, if we remove and reinsert the same device while > suspended, we do nothing because the DSN didn't change. Previously we > called pciehp_enable_slot(). I don't know if we need to do anything > here or not. Mainly to avoid the redundant device add, the same like the changes for case 4 > >> 4. slot is non empty before suspend, no action during suspend. >> We should do nothing in pciehp_resume, but we call >> pciehp_enable_slot(), so some uncomfortable messages show like above. >> In this case, we can improve it a little by add a guard >> if (!list_empty(bus->devices)). > > This is the common case. Previously we called pciehp_enable_slot(), > and with your patch we do nothing. I think that seems sensible, but > this part should be split into a separate patch. That way we can keep > the benefit of this change even if we trip over something with the > other changes. OK, I will split this changes into a new patch. > >> Reported-by: Paul Bolle <pebolle@tiscali.nl> >> Signed-off-by: Yijing Wang <wangyijing@huawei.com> >> Cc: Paul Bolle <pebolle@tiscali.nl> >> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> >> Cc: Oliver Neukum <oneukum@suse.de> >> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> >> Cc: linux-pci@vger.kernel.org >> --- >> drivers/pci/hotplug/pciehp_core.c | 38 +++++++++++++++++++++++++++++++++--- >> 1 files changed, 34 insertions(+), 4 deletions(-) >> >> diff --git a/drivers/pci/hotplug/pciehp_core.c b/drivers/pci/hotplug/pciehp_core.c >> index 7d72c5e..d01e093 100644 >> --- a/drivers/pci/hotplug/pciehp_core.c >> +++ b/drivers/pci/hotplug/pciehp_core.c >> @@ -291,6 +291,28 @@ static void pciehp_remove(struct pcie_device *dev) >> } >> >> #ifdef CONFIG_PM >> + >> +/* If device support Device Serial Numner, use DSN > > s/support/supports/ > s/Numner/Number/ > Use conventional comment style: > /* > * If device ... > */ > Will update, thanks. >> + * to identify the device >> + */ >> +static bool device_in_slot_is_changed(struct pci_bus *pbus) >> +{ >> + u64 old_dsn, new_dsn; >> + struct pci_dev *pdev; >> + >> + pdev = pci_get_slot(pbus, PCI_DEVFN(0, 0)); > > pci_get_slot() can fail. Will add failure return check, thanks. > >> + old_dsn = pdev->sn; >> + >> + /* get func 0 device serial number */ >> + pci_get_dsn(pdev, &new_dsn); >> + if (status) { >> + if (list_empty(&pbus->devices)) >> + pciehp_enable_slot(slot); >> + else if (device_in_slot_is_changed(pbus)) { >> + pciehp_disable_slot(slot); > > pciehp_disable_slot() ultimately calls the .remove() method for the > device that has already been removed. That method is likely to poke > at the device, which will do something unexpected because the new > device is likely to be completely different. Yes, this is a really issue, so what about firstly power off the slot before call .remove() method for the old device. > > I think the call path is this: > > pciehp_resume > pciehp_disable_slot > remove_board > pciehp_unconfigure_device > pci_stop_and_remove_bus_device > pci_stop_bus_device > pci_stop_dev > device_del > bus_remove_device > device_release_driver > __device_release_driver > pci_device_remove # pci_bus_type.remove > dev->driver->remove > > I think we already had this problem for case 2 (card removed while > suspended), but at least in that case, the driver .remove() method > doesn't have a device to poke at, so it's less likely to do something > bad. Agree. > >> + pciehp_enable_slot(slot); >> + } >> + } else { >> + if (!list_empty(&pbus->devices)) >> + pciehp_disable_slot(slot); >> + } >> return 0; >> } >> #endif /* PM */ >> -- >> 1.7.1 >> >> > > . >
On Tue, Jul 9, 2013 at 9:00 PM, Yijing Wang <wangyijing@huawei.com> wrote: > Hi Bjorn, > Thanks for your review and comments! > >>> We can use PCIe Device Serial Number to identify the device if >>> device support DSN. >> >> I think I like the idea of this, especially because the Microsoft PCI >> Hardware Compliance Test apparently requires DSN for hot-pluggable >> PCIe devices [1], so it should be pretty universal. >> >> [1] http://www.techtalkz.com/microsoft-device-drivers-dtm/341362-dtm-pcihct-test-violates-pci-express-base-specification-revision-1-a.html >> >>> currently: >>> 1. slot is empty before suspend, insert card during suspend. >>> In this case, is safe, pciehp will add device by check adapter >>> status register in pciehp_resume. >> >> Your patch doesn't change anything here. > > Yes, I only to make some changes for case 3/4. > >> >>> 2. slot is non empty before suspend, remove card during suspend. >>> Also be safe, pciehp will remove device by check adapter >>> status register in pciehp_resume. >> >> Your patch doesn't change anything here. (But I think the driver >> .remove() method will try to poke at the non-existent device; see >> below.) > > I'm not sure the result of driver .remove() method to poke at the non-existent device. > If driver .remove() method cannot detect the real device, remove action will be block ? > If the slot support surprise hot remove, this action maybe safe. right? If there's no device, config space accesses performed by .remove() will fail (reads will return -1 data or error; writes will be dropped). MMIO or I/O port accesses may fail with machine checks or similar bad things. But I don't see a way around that except by fixing drivers as we encounter issues like that. Since you're not changing anything here, I don't think we should worry about it for now. > If the slot does not support surprise hot remove, but the device was already removed, > we seem to have no other way to clean the stale data related to the old device. > > Now if we check adapter status in slot and found adapter is non existent, pciehp resume > call pciehp_disable_slot() , in pciehp_disable_slot() function, we will check latch status, > I guess this case latch is open(because slot is empty), this action will abort. > But I have no platform to test it. > >> >>> 3. slot is non empty before suspend, remove card during suspend >>> and insert a new card. >>> Now pciehp just call pciehp_enable_slot() roughly. We should >>> remove the old card firstly, then add the new card. >> >> With your patch, I think we'll call the old driver's .remove() method >> on the new device. This seems bad; see below. > > Ah, this is issue. > What about power off slot first, then call the old driver's remove() method > will not touch the new physical device. After the old driver's remove() cleanup, > we call pciehp_enable_slot() to power on and enable the new device. Turning off power to the slot seems like a reasonable approach. Then we can run the old .remove() method in basically the same way we would in case 2. >> With your patch, if we remove and reinsert the same device while >> suspended, we do nothing because the DSN didn't change. Previously we >> called pciehp_enable_slot(). I don't know if we need to do anything >> here or not. > > Mainly to avoid the redundant device add, the same like the changes for case 4 I don't know whether it's redundant or not. Obviously if we remove and reinsert a device, we lose *all* state that was in the device. If we lose everything even if the card stayed inserted the whole time we were suspended, we must already deal with that and the "add" would be redundant. But if the state of the card is different if it got pulled and reinserted, the "add" would be necessary. >>> 4. slot is non empty before suspend, no action during suspend. >>> We should do nothing in pciehp_resume, but we call >>> pciehp_enable_slot(), so some uncomfortable messages show like above. >>> In this case, we can improve it a little by add a guard >>> if (!list_empty(bus->devices)). >> >> This is the common case. Previously we called pciehp_enable_slot(), >> and with your patch we do nothing. I think that seems sensible, but >> this part should be split into a separate patch. That way we can keep >> the benefit of this change even if we trip over something with the >> other changes. > > OK, I will split this changes into a new patch. Actually, without your DSN changes, I don't think we can distinguish this from case 3. So I doubt it really could be split out. -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
>> If the slot support surprise hot remove, this action maybe safe. right? > > If there's no device, config space accesses performed by .remove() > will fail (reads will return -1 data or error; writes will be > dropped). MMIO or I/O port accesses may fail with machine checks or > similar bad things. But I don't see a way around that except by > fixing drivers as we encounter issues like that. > > Since you're not changing anything here, I don't think we should worry > about it for now. OK. > >>>> remove the old card firstly, then add the new card. >>> >>> With your patch, I think we'll call the old driver's .remove() method >>> on the new device. This seems bad; see below. >> >> Ah, this is issue. >> What about power off slot first, then call the old driver's remove() method >> will not touch the new physical device. After the old driver's remove() cleanup, >> we call pciehp_enable_slot() to power on and enable the new device. > > Turning off power to the slot seems like a reasonable approach. Then > we can run the old .remove() method in basically the same way we would > in case 2. Hmmm, I will follow this method to rework this patch in next version. > >>> With your patch, if we remove and reinsert the same device while >>> suspended, we do nothing because the DSN didn't change. Previously we >>> called pciehp_enable_slot(). I don't know if we need to do anything >>> here or not. >> >> Mainly to avoid the redundant device add, the same like the changes for case 4 > > I don't know whether it's redundant or not. Obviously if we remove > and reinsert a device, we lose *all* state that was in the device. If > we lose everything even if the card stayed inserted the whole time we > were suspended, we must already deal with that and the "add" would be > redundant. But if the state of the card is different if it got pulled > and reinserted, the "add" would be necessary. This is a key issue, sorry, I'm not familiar with PM :( In my opinion, the device driver .suspend() method will be called regardless of system enter in suspend to RAM(S3) or suspend to Disk(S4). Driver will save the pci/pcie/pci-x state in .suspend() method. So once device driver .resume() method be called, the pci/pcie/pci-x state willl be restored. Because suspend to disk will power off whole system, so I thought if the device was removed and inserted same device(DSN) again, maybe .resume will enable this device ok regardless of the pci config space state has been changed. If I have any thing above understanding wrong, please correct me, thanks! > >>>> 4. slot is non empty before suspend, no action during suspend. >>>> We should do nothing in pciehp_resume, but we call >>>> pciehp_enable_slot(), so some uncomfortable messages show like above. >>>> In this case, we can improve it a little by add a guard >>>> if (!list_empty(bus->devices)). >>> >>> This is the common case. Previously we called pciehp_enable_slot(), >>> and with your patch we do nothing. I think that seems sensible, but >>> this part should be split into a separate patch. That way we can keep >>> the benefit of this change even if we trip over something with the >>> other changes. >> >> OK, I will split this changes into a new patch. > > Actually, without your DSN changes, I don't think we can distinguish > this from case 3. So I doubt it really could be split out. I will try, but I think this is not a big deal :) > > . >
>> We should do nothing in pciehp_resume, but we call >> pciehp_enable_slot(), so some uncomfortable messages show like above. >> In this case, we can improve it a little by add a guard >> if (!list_empty(bus->devices)). > > Great! > > I'm currently trying to bisect another problem, but hope to test this > patch (and the preceding patch it apparently needs) in a few days. > Please feel free to prod me if you think testing is needed but I'm > taking to long to report back. Hi Paul, Can you provide the lspci -vvv and lspci -xxxx info messages ? I want to confirm your hardware information which cause your resume error. You can get these messages in any kernel version, that's ok. When you suspend and resume the system, the wireless card is always present in pcie slot, right? Thanks! Yijing. > > > . >
Yijing, On Thu, 2013-07-11 at 11:55 +0800, Yijing Wang wrote: > Can you provide the lspci -vvv and lspci -xxxx info messages ? > I want to confirm your hardware information which cause your resume error. > You can get these messages in any kernel version, that's ok. Would it be sufficient to send that information just for the two pcie ports on my laptop, and the wireless card that is present in one of those two ports? And would it be OK to send that information off list? > When you suspend and resume the system, the wireless card is always > present in pcie slot, right? Yes. I actually wouldn't even know how to remove it. I only discovered my wireless card was in a pcie slot because of the error messages at resume (which started a few releases ago, I think v3.7). Paul Bolle -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 2013/7/11 18:19, Paul Bolle wrote: > Yijing, > > On Thu, 2013-07-11 at 11:55 +0800, Yijing Wang wrote: >> Can you provide the lspci -vvv and lspci -xxxx info messages ? >> I want to confirm your hardware information which cause your resume error. >> You can get these messages in any kernel version, that's ok. > > Would it be sufficient to send that information just for the two pcie > ports on my laptop, and the wireless card that is present in one of > those two ports? > If you can provide both the lspci -vvv and lspci -xxxx info before suspend and after resume, it's better. > And would it be OK to send that information off list? OK~ Thanks! Yijing. > >> When you suspend and resume the system, the wireless card is always >> present in pcie slot, right? > > Yes. I actually wouldn't even know how to remove it. I only discovered > my wireless card was in a pcie slot because of the error messages at > resume (which started a few releases ago, I think v3.7). > > > Paul Bolle > > > . >
diff --git a/drivers/pci/hotplug/pciehp_core.c b/drivers/pci/hotplug/pciehp_core.c index 7d72c5e..d01e093 100644 --- a/drivers/pci/hotplug/pciehp_core.c +++ b/drivers/pci/hotplug/pciehp_core.c @@ -291,6 +291,28 @@ static void pciehp_remove(struct pcie_device *dev) } #ifdef CONFIG_PM + +/* If device support Device Serial Numner, use DSN + * to identify the device + */ +static bool device_in_slot_is_changed(struct pci_bus *pbus) +{ + u64 old_dsn, new_dsn; + struct pci_dev *pdev; + + pdev = pci_get_slot(pbus, PCI_DEVFN(0, 0)); + old_dsn = pdev->sn; + + /* get func 0 device serial number */ + pci_get_dsn(pdev, &new_dsn); + pci_dev_put(pdev); + + if (old_dsn != new_dsn) + return true; + + return false; +} + static int pciehp_suspend (struct pcie_device *dev) { return 0; @@ -300,6 +322,7 @@ static int pciehp_resume (struct pcie_device *dev) { struct controller *ctrl; struct slot *slot; + struct pci_bus *pbus = dev->port->subordinate; u8 status; ctrl = get_service_data(dev); @@ -311,10 +334,17 @@ static int pciehp_resume (struct pcie_device *dev) /* Check if slot is occupied */ pciehp_get_adapter_status(slot, &status); - if (status) - pciehp_enable_slot(slot); - else - pciehp_disable_slot(slot); + if (status) { + if (list_empty(&pbus->devices)) + pciehp_enable_slot(slot); + else if (device_in_slot_is_changed(pbus)) { + pciehp_disable_slot(slot); + pciehp_enable_slot(slot); + } + } else { + if (!list_empty(&pbus->devices)) + pciehp_disable_slot(slot); + } return 0; } #endif /* PM */
Currently, pciehp_resume will call pciehp_enable_slot() to add device if there is a device in the slot. But if the device was present before suspend, it's no necessary to add again. Now in such case, there is some uncomfortable message like pciehp 0000:00:1c.1:pcie04: Device 0000:03:00.0 already exists at 0000:03:00, cannot hot-add pciehp 0000:00:1c.1:pcie04: Cannot add device at 0000:03:00 This problem was reported by Paul Bolle <pebolle@tiscali.nl> The discussion link: http://comments.gmane.org/gmane.linux.kernel.pci/19876 We can use PCIe Device Serial Number to identify the device if device support DSN. currently: 1. slot is empty before suspend, insert card during suspend. In this case, is safe, pciehp will add device by check adapter status register in pciehp_resume. 2. slot is non empty before suspend, remove card during suspend. Also be safe, pciehp will remove device by check adapter status register in pciehp_resume. 3. slot is non empty before suspend, remove card during suspend and insert a new card. Now pciehp just call pciehp_enable_slot() roughly. We should remove the old card firstly, then add the new card. 4. slot is non empty before suspend, no action during suspend. We should do nothing in pciehp_resume, but we call pciehp_enable_slot(), so some uncomfortable messages show like above. In this case, we can improve it a little by add a guard if (!list_empty(bus->devices)). Reported-by: Paul Bolle <pebolle@tiscali.nl> Signed-off-by: Yijing Wang <wangyijing@huawei.com> Cc: Paul Bolle <pebolle@tiscali.nl> Cc: "Rafael J. Wysocki" <rjw@sisk.pl> Cc: Oliver Neukum <oneukum@suse.de> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: linux-pci@vger.kernel.org --- drivers/pci/hotplug/pciehp_core.c | 38 +++++++++++++++++++++++++++++++++--- 1 files changed, 34 insertions(+), 4 deletions(-)