Re: [PATCH v6 06/18] PCI: designware-ep: Add generic function for raising MSI irq

From: Kishon Vijay Abraham I
Date: Thu Dec 28 2017 - 09:40:09 EST


Hi Niklas,

On Thursday 28 December 2017 01:36 PM, Kishon Vijay Abraham I wrote:
> Hi Niklas,
>
> On Thursday 28 December 2017 03:59 AM, Niklas Cassel wrote:
>> On Tue, Dec 26, 2017 at 06:20:54PM +0530, Kishon Vijay Abraham I wrote:
>>> Hi Niklas,
>>
>> Hello Kishon
>>
>>>
>>> On Wednesday 20 December 2017 04:59 AM, Niklas Cassel wrote:
>>>> Add a generic function for raising MSI irqs that can be used by all
>>>> DWC based controllers.
>>>>
>>>> Note that certain controllers, like DRA7xx, have a special convenience
>>>> register for raising MSI irqs that doesn't require you to explicitly map
>>>> the MSI address. Therefore, it is likely that certain drivers will
>>>> not use this generic function, even if they can.
>>>>
>>>> Signed-off-by: Niklas Cassel <niklas.cassel@xxxxxxxx>
>>>> ---
>>>> drivers/pci/dwc/pcie-designware-ep.c | 35 +++++++++++++++++++++++++++++++++++
>>>> drivers/pci/dwc/pcie-designware.h | 9 +++++++++
>>>> 2 files changed, 44 insertions(+)
>>>>
>>>> diff --git a/drivers/pci/dwc/pcie-designware-ep.c b/drivers/pci/dwc/pcie-designware-ep.c
>>>> index 700ed2f4becf..c5aa1cac5041 100644
>>>> --- a/drivers/pci/dwc/pcie-designware-ep.c
>>>> +++ b/drivers/pci/dwc/pcie-designware-ep.c
>>>> @@ -282,6 +282,41 @@ static const struct pci_epc_ops epc_ops = {
>>>> .stop = dw_pcie_ep_stop,
>>>> };
>>>>
>>>> +int dw_pcie_ep_raise_msi_irq(struct dw_pcie_ep *ep,
>>>> + u8 interrupt_num)
>>>> +{
>>>> + struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
>>>> + struct pci_epc *epc = ep->epc;
>>>> + u16 msg_ctrl, msg_data;
>>>> + u32 msg_addr_lower, msg_addr_upper;
>>>> + u64 msg_addr;
>>>> + bool has_upper;
>>>> + int ret;
>>>> +
>>>> + /* Raise MSI per the PCI Local Bus Specification Revision 3.0, 6.8.1. */
>>>> + msg_ctrl = dw_pcie_readw_dbi(pci, MSI_MESSAGE_CONTROL);
>>>> + has_upper = !!(msg_ctrl & PCI_MSI_FLAGS_64BIT);
>>>> + msg_addr_lower = dw_pcie_readl_dbi(pci, MSI_MESSAGE_ADDR_L32);
>>>> + if (has_upper) {
>>>> + msg_addr_upper = dw_pcie_readl_dbi(pci, MSI_MESSAGE_ADDR_U32);
>>>> + msg_data = dw_pcie_readw_dbi(pci, MSI_MESSAGE_DATA_64);
>>>> + } else {
>>>> + msg_addr_upper = 0;
>>>> + msg_data = dw_pcie_readw_dbi(pci, MSI_MESSAGE_DATA_32);
>>>> + }
>>>> + msg_addr = ((u64) msg_addr_upper) << 32 | msg_addr_lower;
>>>> + ret = dw_pcie_ep_map_addr(epc, ep->msi_mem_phys, msg_addr,
>>>> + epc->mem->page_size);
>>>> + if (ret)
>>>> + return ret;
>>>> +
>>>> + writel(msg_data | (interrupt_num - 1), ep->msi_mem);
>>>
>>> Shouldn't this be msg_data + (interrupt_num - 1)?
>>
>> I'm not quite sure about this, but if there is a pending irq,
>> not yet processed by the RC, the msg_data we read out in this
>> function should have a bit set, matching the pending irq.
>
> IIUC, the msg_data that we read here should not depend on the pending irq on
> the RC side. msg_data should have the starting MSI vector number assigned by RC
> for that EP device. (msg.data = pos; in dw_msi_setup_msg() also seem to suggest
> the same).
>>
>> If that irq is the same as the irq we are trying to raise,
>> doing an addition will produce a bogus vector number,
>> but a bitwise or should work.
>
> if msg_data has the starting MSI vector, doing an addition should get to the
> correct MSI vector.
>>
>> For that reason, I think that doing bitwise or seems safer.
>> However, other than this case, I don't see why it should
>> matter if we do an addition or a bitwise or.
>>
>> Are you having some problem with the code?
>> It seems to be working fine on ARTPEC-6:
>>
>> # ./pcitest -m 1
>> MSI1: OKAY
>> # ./pcitest -m 2
>> MSI2: OKAY
>> # ./pcitest -m 3
>> MSI3: OKAY
>> # ./pcitest -m 4
>> MSI4: OKAY
>> # ./pcitest -m 5
>> MSI5: OKAY
>> # ./pcitest -m 6
>> MSI6: OKAY
>> # ./pcitest -m 7
>> MSI7: OKAY
>> # ./pcitest -m 8
>> MSI8: OKAY
>> # ./pcitest -m 9
>> MSI9: OKAY
>> # cat /proc/interrupts | grep -i msi
>> 82: 9 0 GIC-0 180 Level artpec6-pcie-msi
>> 188: 1 0 PCI-MSI 16 Edge pci-endpoint-test
>> 189: 1 0 PCI-MSI 17 Edge pci-endpoint-test
>> 190: 1 0 PCI-MSI 18 Edge pci-endpoint-test
>> 191: 1 0 PCI-MSI 19 Edge pci-endpoint-test
>> 192: 1 0 PCI-MSI 20 Edge pci-endpoint-test
>> 193: 1 0 PCI-MSI 21 Edge pci-endpoint-test
>> 194: 1 0 PCI-MSI 22 Edge pci-endpoint-test
>> 195: 1 0 PCI-MSI 23 Edge pci-endpoint-test
>> 196: 1 0 PCI-MSI 24 Edge pci-endpoint-test
>> 197: 0 0 PCI-MSI 25 Edge pci-endpoint-test
>> 198: 0 0 PCI-MSI 26 Edge pci-endpoint-test
>> 199: 0 0 PCI-MSI 27 Edge pci-endpoint-test
>> 200: 0 0 PCI-MSI 28 Edge pci-endpoint-test
>> 201: 0 0 PCI-MSI 29 Edge pci-endpoint-test
>> 202: 0 0 PCI-MSI 30 Edge pci-endpoint-test
>> 203: 0 0 PCI-MSI 31 Edge pci-endpoint-test
>>
>> From EP:
>> irq: 1 read msg_data: 0x10 writing: 0x10
>> irq: 2 read msg_data: 0x10 writing: 0x11
>> irq: 3 read msg_data: 0x10 writing: 0x12
>> irq: 4 read msg_data: 0x10 writing: 0x13
>> irq: 5 read msg_data: 0x10 writing: 0x14
>> irq: 6 read msg_data: 0x10 writing: 0x15
>> irq: 7 read msg_data: 0x10 writing: 0x16
>> irq: 8 read msg_data: 0x10 writing: 0x17
>> irq: 9 read msg_data: 0x10 writing: 0x18
>
> since your msg_data is 0x10, you are not facing the issue. What if it's 0x1? In
> my case If I have Gustavo's patch series applied, msg_data has a value of 0x1
> and I don't get certain MSI interrupts.

I think Gustavo's series doesn't align the MSI vector properly. That's why I
get 0x1. So your patch is fine as such.

Thanks
Kishon