RFC for ideas to clean up PCI MSI code
Hi Bjorn, Yijing and Thomas,
Recently Yijing is trying to kill some weak functions indrivers/pci/msi.c by using msi_chip mechanism. And it inspired me that we could go one step further to kill those weak functions at all and move common PCI MSI code into PCI core by using the new hierarch irqdomain framework.
The way to achieve the goal is:
1) arch code creates irqdomain for PCI MSI controllers. By default, there's only one global PCI MSI irqdomain. But there's may be multiple for some architecture like x86. So arch also need to implement a weak function (I know you don't like weak functions):
struct irq_domain *arch_get_pci_msi_irq_domain(struct pci_dev *);
2) replace arch_setup_msi_irqs()/arch_setup_msi_irqs()/arch_teardown_msi_irqs()/arch_teardown_msi_irq() withirq_domain interfaces.
3) kill arch_msi_mask_irq()/arch_msi_mask_irq() by directly setting irq_chip.irq_mask()/irq_unmask().
4) we could also find a way to kill arch_restore_msi_irqs() by using new irqdomain interfaces.
For arch code to support the new MSI, it needs to implement callbacks in msi_domain_ops and irq_chip. All other PCI MSI code becomes common and will be moved into pci/msi.c. Proposed interfaces as below:
struct irq_chip {
void (*irq_compose_msi_msg)(struct irq_data *data,
struct msi_msg *msg);
};
struct msi_domain_ops {
void (*calc_hwirq)(struct msi_domain_info *info, void *arg,
struct msi_desc *desc);
irq_hw_number_t (*get_hwirq)(struct msi_domain_info *info, void
*arg);
int (*msi_init)(struct irq_domain *domain, structmsi_domain_info *info,
unsigned int virq, irq_hw_number_t hwirq, void
*arg);
void (*msi_free)(struct irq_domain *domain,
struct msi_domain_info *info, unsigned int virq);
};
struct msi_domain_info {
struct msi_domain_ops *ops;
struct irq_chip *chip;
void *data;
};struct irq_domain *msi_create_irq_domain(struct device_node *of_node,
struct msi_domain_info *info, struct irq_domain *parent);
So, what's your thoughts about the proposal? If now objection, I may send out a draft version for x86 within about one week:) Thanks!
Gerry
Quoted 11 lines · Jiang Liu
Hi Bjorn, Yijing and Thomas,
Recently Yijing is trying to kill some weak functions indrivers/pci/msi.cby usingmsi_chipmechanism. And it inspired me that we could go one step further to kill those weak functions at all and move common PCI MSI code into PCI core by using the new hierarch irqdomain framework.
The way to achieve the goal is:
1) arch code creates irqdomain for PCI MSI controllers. By default, there's only one global PCI MSI irqdomain. But there's may be multiple for some architecture like x86. So arch also need to implement a weak function (I know you don't like weak functions):
Quoted 1 line
struct irq_domain *arch_get_pci_msi_irq_domain(struct pci_dev *);
Arm also may have multiple msi_chips/msi irq domains, currently, corresponding msi_chip is stored in bus->msi. But it's not a good idea, I'm refactoring generic pci host bridge, we could save it in generic pci_host_bridge.
Quoted 5 lines
2) replace arch_setup_msi_irqs()/arch_setup_msi_irqs()/arch_teardown_msi_irqs()/arch_teardown_msi_irq() with
irq_domaininterfaces.
Agree.
Quoted 3 lines
3) kill arch_msi_mask_irq()/arch_msi_mask_irq() by directly setting irq_chip.irq_mask()/irq_unmask().
arch_msi_mask_irq() is only called in pci_msi_shutdown, in this case, the msi mask status will not be overrode, so we can restore it later. If use irq_chip.irq_mask(), we need another place to save the mask status.
Quoted 3 lines
4) we could also find a way to kill
arch_restore_msi_irqs()by using new irqdomain interfaces.
My personal opinion, I think place arch_retore_msi_irqs() stuffs in irq_chip is better. :)
Quoted 29 lines
For arch code to support the new MSI, it needs to implement callbacks in
msi_domain_opsandirq_chip. All other PCI MSI code becomes common and will be moved intopci/msi.c. Proposed interfaces as below:struct
irq_chip{
void (*irq_compose_msi_msg)(structirq_data*data,
structmsi_msg*msg);
};struct
msi_domain_ops{
void (*calc_hwirq)(structmsi_domain_info*info, void *arg,
structmsi_desc*desc);
irq_hw_number_t(*get_hwirq)(structmsi_domain_info*info, void
*arg);
int (*msi_init)(structirq_domain*domain, structmsi_domain_info*info,
unsigned int virq,irq_hw_number_thwirq, void
*arg);
void (*msi_free)(structirq_domain*domain,
structmsi_domain_info*info, unsigned int virq);
};struct msi_domain_info { struct msi_domain_ops *ops; struct irq_chip *chip; void *data; };
Quoted 6 lines
struct
irq_domain*msi_create_irq_domain(structdevice_node*of_node,
structmsi_domain_info*info, structirq_domain*parent);So, what's your thoughts about the proposal? If now objection, I may send out a draft version for x86 within about one week:)
I think the proposal is better than original msi_chip framework. I like this idea. :)
Thanks!
Yijing.
Quoted 5 lines
Thanks!
Gerry.
Signature · 2 lines
Thanks! Yijing