Re: [PATCH v2 1/5] PCI: endpoint: Add RC-to-EP doorbell support using platform MSI controller

[Frank Li <Frank.li@nxp.com>]

On Fri, Sep 29, 2023 at 03:00:16PM +0530, Kishon Vijay Abraham I wrote:
> Hi Frank,
> 
> On 9/12/2023 3:39 AM, Frank Li wrote:
> > This commit introduces a common method for sending messages from the Root
> > Complex (RC) to the Endpoint (EP) by utilizing the platform MSI interrupt
> > controller, such as ARM GIC, as an EP doorbell. Maps the memory assigned
> > for the BAR region by the PCI host to the message address of the platform
> > MSI interrupt controller in the PCI EP. As a result, when the PCI RC writes
> > to the BAR region, it triggers an IRQ at the EP. This implementation serves
> > as a common method for all endpoint function drivers.
> 
> This would help avoid the polling used in current EPF drivers. Thanks!
> > 
> > However, it currently supports only one EP physical function due to
> > limitations in ARM MSI/IMS readiness.
> 
> Any such platform or architecture restrictions should not be handled in the
> endpoint core layer.
> > 
> > Signed-off-by: Frank Li <Frank.Li@nxp.com>
> > ---
> >   drivers/pci/endpoint/pci-epc-core.c | 192 ++++++++++++++++++++++++++++
> >   drivers/pci/endpoint/pci-epf-core.c |  44 +++++++
> >   include/linux/pci-epc.h             |   6 +
> >   include/linux/pci-epf.h             |   7 +
> >   4 files changed, 249 insertions(+)
> > 
> > diff --git a/drivers/pci/endpoint/pci-epc-core.c b/drivers/pci/endpoint/pci-epc-core.c
> > index 5a4a8b0be6262..d336a99c6a94f 100644
> > --- a/drivers/pci/endpoint/pci-epc-core.c
> > +++ b/drivers/pci/endpoint/pci-epc-core.c
> > @@ -10,6 +10,7 @@
> >   #include <linux/slab.h>
> >   #include <linux/module.h>
> > +#include <linux/msi.h>
> >   #include <linux/pci-epc.h>
> >   #include <linux/pci-epf.h>
> >   #include <linux/pci-ep-cfs.h>
> > @@ -783,6 +784,197 @@ void pci_epc_bme_notify(struct pci_epc *epc)
> >   }
> >   EXPORT_SYMBOL_GPL(pci_epc_bme_notify);
> > +/**
> > + * pci_epc_alloc_doorbell() - alloc an address space to let RC trigger EP side IRQ by write data to
> > + *			      the space.
> > + *
> > + * @epc: the EPC device that need doorbell address and data from RC.
> > + * @func_no: the physical endpoint function number in the EPC device.
> > + * @vfunc_no: the virtual endpoint function number in the physical function.
> > + * @num_msgs: the total number of doorbell messages
> > + *
> > + * Return: 0 success, other is failure
> > + */
> > +int pci_epc_alloc_doorbell(struct pci_epc *epc, u8 func_no, u8 vfunc_no, int num_msgs)
> > +{
> > +	int ret;
> > +
> > +	if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
> > +		return -EINVAL;
> > +
> > +	if (vfunc_no > 0 && (!epc->max_vfs || vfunc_no > epc->max_vfs[func_no]))
> > +		return -EINVAL;
> > +
> > +	if (!epc->ops->alloc_doorbell)
> > +		return 0;
> > +
> > +	mutex_lock(&epc->lock);
> > +	ret = epc->ops->alloc_doorbell(epc, func_no, vfunc_no, num_msgs);
> > +	mutex_unlock(&epc->lock);
> > +
> > +	return ret;
> > +}
> > +EXPORT_SYMBOL_GPL(pci_epc_alloc_doorbell);
> > +
> > +/**
> > + * pci_epc_free_doorbell() - free resource allocated by pci_epc_alloc_doorbell()
> > + *
> > + * @epc: the EPC device that need doorbell address and data from RC.
> > + * @func_no: the physical endpoint function number in the EPC device.
> > + * @vfunc_no: the virtual endpoint function number in the physical function.
> > + *
> > + * Return: 0 success, other is failure
> > + */
> > +void pci_epc_free_doorbell(struct pci_epc *epc, u8 func_no, u8 vfunc_no)
> > +{
> > +	if (IS_ERR_OR_NULL(epc) || func_no >= epc->max_functions)
> > +		return;
> > +
> > +	if (vfunc_no > 0 && (!epc->max_vfs || vfunc_no > epc->max_vfs[func_no]))
> > +		return;
> > +
> > +	if (!epc->ops->free_doorbell)
> > +		return;
> > +
> > +	mutex_lock(&epc->lock);
> > +	epc->ops->free_doorbell(epc, func_no, vfunc_no);
> > +	mutex_unlock(&epc->lock);
> > +}
> > +EXPORT_SYMBOL_GPL(pci_epc_free_doorbell);
> > +
> > +static irqreturn_t pci_epf_generic_doorbell_handler(int irq, void *data)
> > +{
> > +	struct pci_epf *epf = data;
> > +
> > +	if (epf->event_ops && epf->event_ops->doorbell)
> > +		epf->event_ops->doorbell(epf, irq - epf->virq_base);
> > +
> > +	return IRQ_HANDLED;
> > +}
> 
> IMO the handler should be directly implemented in the EPF drivers. There
> should be one API which returns the virq and the msi_msg to the EPF driver
> and the EPF driver should do request_irq.
> > +
> > +static void pci_epc_generic_write_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
> > +{
> > +	struct pci_epc *epc = NULL;
> > +	struct class_dev_iter iter;
> > +	struct pci_epf *epf;
> > +	struct device *dev;
> > +
> > +	class_dev_iter_init(&iter, pci_epc_class, NULL, NULL);
> > +	while ((dev = class_dev_iter_next(&iter))) {
> > +		if (dev->parent != desc->dev)
> > +			continue;
> 
> Ideally the msi_desc should be associated directly with the EPF device.

The key problem is platform_msi_domain_alloc_irqs only support put epc
driver's dev into desc.

IMS may resolve this problem, but ARM IMS progress is quite slow.

> > +
> > +		epc = to_pci_epc(dev);
> > +
> > +		class_dev_iter_exit(&iter);
> > +		break;
> > +	}
> > +
> > +	if (!epc)
> > +		return;
> > +
> > +	/* Only support one EPF for doorbell */
> > +	epf = list_first_entry_or_null(&epc->pci_epf, struct pci_epf, list);
> > +
> > +	if (!epf)
> > +		return;
> 
> If this is a platform restriction, this should be moved elsewhere.

where is good place? 

> > +
> > +	if (epf->msg && desc->msi_index < epf->num_msgs)
> > +		epf->msg[desc->msi_index] = *msg;
> > +}
> > +
> > +
> > +/**
> > + * pci_epc_generic_alloc_doorbell() - Common help function. Allocate address space from MSI
> > + *                                    controller
> > + *
> > + * @epc: the EPC device that need doorbell address and data from RC.
> > + * @func_no: the physical endpoint function number in the EPC device.
> > + * @vfunc_no: the virtual endpoint function number in the physical function.
> > + * @num_msgs: the total number of doorbell messages
> > + *
> > + * Remark: use this function only if EPC driver just register one EPC device.
> > + *
> > + * Return: 0 success, other is failure
> > + */
> > +int pci_epc_generic_alloc_doorbell(struct pci_epc *epc, u8 func_no, u8 vfunc_no, int num_msgs)
> > +{
> > +	struct pci_epf *epf;
> > +	struct device *dev;
> > +	int virq, last;
> > +	int ret;
> > +	int i;
> > +
> > +	if (IS_ERR_OR_NULL(epc))
> > +		return -EINVAL;
> > +
> > +	/* Currently only support one func and one vfunc for doorbell */
> > +	if (func_no || vfunc_no)
> > +		return -EINVAL;
> > +
> > +	epf = list_first_entry_or_null(&epc->pci_epf, struct pci_epf, list);
> > +	if (!epf)
> > +		return -EINVAL;
> > +
> > +	dev = epc->dev.parent;
> > +	ret = platform_msi_domain_alloc_irqs(dev, num_msgs, pci_epc_generic_write_msi_msg);
> > +	if (ret) {
> > +		dev_err(dev, "Failed to allocate MSI\n");
> > +		return -ENOMEM;
> > +	}
> 
> The alloc_irqs should be for a EPF device IMO.

The GIC ITS driver need of_node information to alloc msi for such devices.
But EPF device have not of_node information.

> > +
> > +	last = -1;
> > +	for (i = 0; i < num_msgs; i++) {
> > +		virq = msi_get_virq(dev, i);
> > +		if (i == 0)
> > +			epf->virq_base = virq;
> > +
> > +		ret = request_irq(virq, pci_epf_generic_doorbell_handler, 0,
> > +				  kasprintf(GFP_KERNEL, "pci-epc-doorbell%d", i), epf);
> > +
> > +		if (ret) {
> > +			dev_err(dev, "Failed to request doorbell\n");
> > +			goto err_free_irq;
> > +		}
> > +		last = i;
> > +	}
> > +
> > +	return 0;
> > +
> > +err_free_irq:
> > +	for (i = 0; i < last; i++)
> > +		kfree(free_irq(epf->virq_base + i, epf));
> > +	platform_msi_domain_free_irqs(dev);
> > +
> > +	return -EINVAL;
> > +}
> 
> Thanks,
> Kishon

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel