Hi Jon,
I think we can do a little more about inconsistent mps problem found in boot path(BIOS configure bug for mps)
or after hotplug device.As shown in your comment on line 1451, it's unsafe to modifying the MPS on the running devices.
Since bus child devices is not running when pcie_bus_configure_settings be called. So we can try to configure child device's
mps according to the mps of bus.
there are two situations:
1、now current running bus mps is larger than child devices mpss can support;
2、now cuurent running bus mps is smaller than child devices mpss can support;
We cannot modifying the MPS for 1, it's not safe; but we can show message to user to add boot parameter pci=pcie_bus_safe.
The second situation we can modify all child devices mps as the mps value of running bus, it's safe and can correct mps problem automatically
for users.
I wrote a draft patch about this solution, if there is some thing wrong with my understanding, let me know.
Thanks very much!
Yijing
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 5485883..59036a8 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -78,7 +78,7 @@ unsigned long pci_cardbus_mem_size = DEFAULT_CARDBUS_MEM_SIZE;
unsigned long pci_hotplug_io_size = DEFAULT_HOTPLUG_IO_SIZE;
unsigned long pci_hotplug_mem_size = DEFAULT_HOTPLUG_MEM_SIZE;
-enum pcie_bus_config_types pcie_bus_config = PCIE_BUS_TUNE_OFF;
+enum pcie_bus_config_types pcie_bus_config = PCIE_BUS_AUTO;
/*
* The default CLS is used if arch didn't set CLS explicitly and not
diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
index e8b7d5e..6cbfbe1 100644
--- a/drivers/pci/probe.c
+++ b/drivers/pci/probe.c
@@ -1467,6 +1467,19 @@ static int pcie_find_smpss(struct pci_dev *dev, void *data)
return 0;
}
+static int pcie_find_min_mpss(struct pci_dev *dev, void *data)
+{
+ u8 *mpss = data;
+
+ if (!pci_is_pcie(dev))
+ return 0;
+
+ if (*mpss > dev->pcie_mpss)
+ *mpss = dev->pcie_mpss;
+
+ return 0;
+}
+
static void pcie_write_mps(struct pci_dev *dev, int mps)
{
int rc;
@@ -1560,6 +1573,7 @@ static int pcie_bus_configure_set(struct pci_dev *dev, void *data)
void pcie_bus_configure_settings(struct pci_bus *bus, u8 mpss)
{
u8 smpss;
+ u8 mps = pcie_get_mps(bus->self) >> 8;
if (!pci_is_pcie(bus->self))
return;
@@ -1581,7 +1595,19 @@ void pcie_bus_configure_settings(struct pci_bus *bus, u8 mpss)
case PCIE_BUS_PEER2PEER:
smpss = 0;
break;
-
+ case PCIE_BUS_AUTO:
+ smpss = bus->self->pcie_mpss;
+ pci_walk_bus(bus, pcie_find_min_mpss, &smpss);
+ if (mps > smpss) {
+ dev_info(&bus->dev,
+ "Current mps %d used in bus 0x%02x is larger than children devices mpss %d support\n"
+ "Please use the pci=pcie_bus_safe boot parameter for safe\n",
+ 128 << mpss, bus->number, 128 << smpss);
+ return;
+ }
+ else
+ pci_walk_bus(bus, pcie_bus_configure_set, &mps);
+ return;
case PCIE_BUS_SAFE:
smpss = mpss;
@@ -1594,8 +1620,6 @@ void pcie_bus_configure_settings(struct pci_bus *bus, u8 mpss)
return;
}
- }
-
pcie_bus_configure_set(bus->self, &smpss);
pci_walk_bus(bus, pcie_bus_configure_set, &smpss);
}
diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
index 7a451ff..539b7e4 100644
--- a/drivers/pci/quirks.c
+++ b/drivers/pci/quirks.c
@@ -2787,7 +2787,8 @@ static void __devinit quirk_intel_mc_errata(struct pci_dev *dev)
int err;
u16 rcc;
- if (pcie_bus_config == PCIE_BUS_TUNE_OFF)
+ if (pcie_bus_config == PCIE_BUS_TUNE_OFF ||
+ pcie_bus_config == PCIE_AUTO_AUTO)
return;
/* Intel errata specifies bits to change but does not say what they are.
diff --git a/include/linux/pci.h b/include/linux/pci.h
index be1de01..84c4ab1 100644
--- a/include/linux/pci.h
+++ b/include/linux/pci.h
@@ -661,6 +661,7 @@ extern void pcie_bus_configure_settings(struct pci_bus *bus, u8 smpss);
enum pcie_bus_config_types {
PCIE_BUS_TUNE_OFF,
+ PCIE_BUS_AUTO,
PCIE_BUS_SAFE,
PCIE_BUS_PERFORMANCE,
PCIE_BUS_PEER2PEER,
--
1.7.1
On 2012/10/24 6:57, Jon Mason wrote:
>> pcieport 0000:47:09.0 device mps current is 256 and mpss is 1024;
>> the newly hot added igb device 0000:4b:00.0 and 0000:4b:00.1 mps are 128 and mpss is 512;
>> 1、"pci_bus 0000:4b: Non-optimal PCI-E Bus MPS value of 128 being used instead of 1024." should ".......instead of 256(bridge mps)"？
>
> The print out shows a bug in the code (which I will push in a second
> version of the patch shortly), but having 128 instead of 256 is a
> separate issue. That is an existing limitation due to the hotplug
> slot not being connected to the root port. See the comment on line
> 1454. Since PCIE_BUS_SAFE/PERFORMANCE is not being used, it is not
> ratcheting down the MPS on the bus like it should (per the comment).
>
>> 2、“use the pci=pcie_bus_safe boot parameter for better performance” should "....use pci=pcie_bus_safe for safe"?
>
> The reason for the user to add the boot parameter is to get better
> performance than they would by default. For theoretically best
> performance, they would want to use "pcie_bus_performance". With this
> in mind, I believe "better" is the correct language.
>
>> 3、above logs is duplicate.
>
> The duplicate log would only be caused by pcie_bus_configure_settings
> being called twice. Is this only being seen on hotplug devices or on
> every device?
>
>>
>> igb 0000:4b:00.0: enabling device (0100 -> 0102)
>> igb 0000:4b:00.0: Intel(R) Gigabit Ethernet Network Connection
>> igb 0000:4b:00.0: eth4: (PCIe:2.5Gb/s:Width x4) 00:0e:0c:ff:ff:ff
>> igb 0000:4b:00.0: eth4: PBA No: FFFFFF-0FF
>> igb 0000:4b:00.0: Using MSI-X interrupts. 8 rx queue(s), 8 tx queue(s)
>> igb 0000:4b:00.1: enabling device (0100 -> 0102)
>> GSI 63 (level, low) -> CPU 27 (0x3300) vector 137
>> igb 0000:4b:00.1: Intel(R) Gigabit Ethernet Network Connection
>> igb 0000:4b:00.1: eth4: (PCIe:2.5Gb/s:Width x4) 00:0e:0c:ff:ff:fe
>> igb 0000:4b:00.1: eth4: PBA No: FFFFFF-0FF
>> igb 0000:4b:00.1: Using MSI-X interrupts. 8 rx queue(s), 8 tx queue(s)
>>
>>
>> Thanks!
>> Yijing
>>
>>
>>> + return;
>>> }
>>>
>>> pcie_bus_configure_set(bus->self, &smpss);
>>> diff --git a/drivers/pci/quirks.c b/drivers/pci/quirks.c
>>> index 5155317..e4eede0 100644
>>> --- a/drivers/pci/quirks.c
>>> +++ b/drivers/pci/quirks.c
>>> @@ -2787,7 +2787,8 @@ static void __devinit quirk_intel_mc_errata(struct pci_dev *dev)
>>> int err;
>>> u16 rcc;
>>>
>>> - if (pcie_bus_config == PCIE_BUS_TUNE_OFF)
>>> + if (pcie_bus_config == PCIE_BUS_TUNE_OFF ||
>>> + pcie_bus_config == PCIE_BUS_WARN)
>>> return;
>>>
>>> /* Intel errata specifies bits to change but does not say what they are.
>>> diff --git a/include/linux/pci.h b/include/linux/pci.h
>>> index 5faa831..410eaf9 100644
>>> --- a/include/linux/pci.h
>>> +++ b/include/linux/pci.h
>>> @@ -662,6 +662,7 @@ extern void pcie_bus_configure_settings(struct pci_bus *bus, u8 smpss);
>>>
>>> enum pcie_bus_config_types {
>>> PCIE_BUS_TUNE_OFF,
>>> + PCIE_BUS_WARN,
>>> PCIE_BUS_SAFE,
>>> PCIE_BUS_PERFORMANCE,
>>> PCIE_BUS_PEER2PEER,
>>>
>>
>>
>> --
>> Thanks!
>> Yijing
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
>> the body of a message to majordomo@xxxxxxxxxxxxxxx
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
> .
>
--
Thanks!
Yijing
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
