On Fri, 2020-12-25 at 19:22 +0000, Marc Zyngier wrote:
> Hi Jianjun,
>
> On Fri, 25 Dec 2020 10:03:07 +0000,
> Jianjun Wang <jianjun.wang@xxxxxxxxxxxx> wrote:
> >
> > MediaTek's PCIe host controller has three generation HWs, the new
> > generation HW is an individual bridge, it supports Gen3 speed and
> > up to 256 MSI interrupt numbers for multi-function devices.
> >
> > Add support for new Gen3 controller which can be found on MT8192.
> >
> > Signed-off-by: Jianjun Wang <jianjun.wang@xxxxxxxxxxxx>
> > Acked-by: Ryder Lee <ryder.lee@xxxxxxxxxxxx>
> > ---
> > drivers/pci/controller/Kconfig | 13 +
> > drivers/pci/controller/Makefile | 1 +
> > drivers/pci/controller/pcie-mediatek-gen3.c | 1084 +++++++++++++++++++
> > 3 files changed, 1098 insertions(+)
> > create mode 100644 drivers/pci/controller/pcie-mediatek-gen3.c
>
> This is a pretty large patch, and it'd be great if you would split it
> into at least 4 parts (core PCIe, PM, MSI, INTx).
Hi Marc,
Thanks for your review, I will try to split it to those parts in the
next version.
>
> >
> > diff --git a/drivers/pci/controller/Kconfig b/drivers/pci/controller/Kconfig
> > index 64e2f5e379aa..b242b17025b3 100644
> > --- a/drivers/pci/controller/Kconfig
> > +++ b/drivers/pci/controller/Kconfig
> > @@ -242,6 +242,19 @@ config PCIE_MEDIATEK
> > Say Y here if you want to enable PCIe controller support on
> > MediaTek SoCs.
> >
> > +config PCIE_MEDIATEK_GEN3
> > + tristate "MediaTek Gen3 PCIe controller"
> > + depends on ARCH_MEDIATEK || COMPILE_TEST
> > + depends on PCI_MSI_IRQ_DOMAIN
> > + help
> > + Adds support for PCIe Gen3 MAC controller for MediaTek SoCs.
> > + This PCIe controller is compatible with Gen3, Gen2 and Gen1 speed,
> > + and support up to 256 MSI interrupt numbers for
> > + multi-function devices.
> > +
> > + Say Y here if you want to enable Gen3 PCIe controller support on
> > + MediaTek SoCs.
> > +
> > config PCIE_TANGO_SMP8759
> > bool "Tango SMP8759 PCIe controller (DANGEROUS)"
> > depends on ARCH_TANGO && PCI_MSI && OF
> > diff --git a/drivers/pci/controller/Makefile b/drivers/pci/controller/Makefile
> > index 04c6edc285c5..df5d77d72a9d 100644
> > --- a/drivers/pci/controller/Makefile
> > +++ b/drivers/pci/controller/Makefile
> > @@ -27,6 +27,7 @@ obj-$(CONFIG_PCIE_ROCKCHIP) += pcie-rockchip.o
> > obj-$(CONFIG_PCIE_ROCKCHIP_EP) += pcie-rockchip-ep.o
> > obj-$(CONFIG_PCIE_ROCKCHIP_HOST) += pcie-rockchip-host.o
> > obj-$(CONFIG_PCIE_MEDIATEK) += pcie-mediatek.o
> > +obj-$(CONFIG_PCIE_MEDIATEK_GEN3) += pcie-mediatek-gen3.o
> > obj-$(CONFIG_PCIE_TANGO_SMP8759) += pcie-tango.o
> > obj-$(CONFIG_VMD) += vmd.o
> > obj-$(CONFIG_PCIE_BRCMSTB) += pcie-brcmstb.o
> > diff --git a/drivers/pci/controller/pcie-mediatek-gen3.c b/drivers/pci/controller/pcie-mediatek-gen3.c
> > new file mode 100644
> > index 000000000000..00cdb598a9f5
> > --- /dev/null
> > +++ b/drivers/pci/controller/pcie-mediatek-gen3.c
> > @@ -0,0 +1,1084 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * MediaTek PCIe host controller driver.
> > + *
> > + * Copyright (c) 2020 MediaTek Inc.
> > + * Author: Jianjun Wang <jianjun.wang@xxxxxxxxxxxx>
> > + */
> > +
> > +#include <linux/clk.h>
> > +#include <linux/delay.h>
> > +#include <linux/iopoll.h>
> > +#include <linux/irq.h>
> > +#include <linux/irqchip/chained_irq.h>
> > +#include <linux/irqdomain.h>
> > +#include <linux/kernel.h>
> > +#include <linux/module.h>
> > +#include <linux/msi.h>
> > +#include <linux/of_address.h>
> > +#include <linux/of_clk.h>
> > +#include <linux/of_pci.h>
> > +#include <linux/of_platform.h>
> > +#include <linux/pci.h>
> > +#include <linux/phy/phy.h>
> > +#include <linux/platform_device.h>
> > +#include <linux/pm_domain.h>
> > +#include <linux/pm_runtime.h>
> > +#include <linux/reset.h>
> > +
> > +#include "../pci.h"
> > +
> > +#define PCIE_SETTING_REG 0x80
> > +#define PCIE_PCI_IDS_1 0x9c
> > +#define PCI_CLASS(class) (class << 8)
> > +#define PCIE_RC_MODE BIT(0)
> > +
> > +#define PCIE_CFGNUM_REG 0x140
> > +#define PCIE_CFG_DEVFN(devfn) ((devfn) & GENMASK(7, 0))
> > +#define PCIE_CFG_BUS(bus) (((bus) << 8) & GENMASK(15, 8))
> > +#define PCIE_CFG_BYTE_EN(bytes) (((bytes) << 16) & GENMASK(19, 16))
> > +#define PCIE_CFG_FORCE_BYTE_EN BIT(20)
> > +#define PCIE_CFG_OFFSET_ADDR 0x1000
> > +#define PCIE_CFG_HEADER(bus, devfn) \
> > + (PCIE_CFG_BUS(bus) | PCIE_CFG_DEVFN(devfn))
> > +
> > +#define PCIE_RST_CTRL_REG 0x148
> > +#define PCIE_MAC_RSTB BIT(0)
> > +#define PCIE_PHY_RSTB BIT(1)
> > +#define PCIE_BRG_RSTB BIT(2)
> > +#define PCIE_PE_RSTB BIT(3)
> > +
> > +#define PCIE_LTSSM_STATUS_REG 0x150
> > +#define PCIE_LTSSM_STATE_MASK GENMASK(28, 24)
> > +#define PCIE_LTSSM_STATE(val) ((val & PCIE_LTSSM_STATE_MASK) >> 24)
> > +#define PCIE_LTSSM_STATE_L2_IDLE 0x14
> > +
> > +#define PCIE_LINK_STATUS_REG 0x154
> > +#define PCIE_PORT_LINKUP BIT(8)
> > +
> > +#define PCIE_MSI_SET_NUM 8
> > +#define PCIE_MSI_IRQS_PER_SET 32
> > +#define PCIE_MSI_IRQS_NUM \
> > + (PCIE_MSI_IRQS_PER_SET * (PCIE_MSI_SET_NUM))
> > +
> > +#define PCIE_INT_ENABLE_REG 0x180
> > +#define PCIE_MSI_MASK GENMASK(PCIE_MSI_SET_NUM + 8 - 1, 8)
> > +#define PCIE_MSI_SHIFT 8
> > +#define PCIE_INTX_SHIFT 24
> > +#define PCIE_INTX_MASK GENMASK(27, 24)
> > +
> > +#define PCIE_INT_STATUS_REG 0x184
> > +#define PCIE_MSI_SET_ENABLE_REG 0x190
> > +
> > +#define PCIE_ICMD_PM_REG 0x198
> > +#define PCIE_TURN_OFF_LINK BIT(4)
> > +
> > +#define PCIE_MSI_ADDR_BASE_REG 0xc00
> > +#define PCIE_MSI_SET_OFFSET 0x10
> > +#define PCIE_MSI_STATUS_OFFSET 0x04
> > +#define PCIE_MSI_ENABLE_OFFSET 0x08
> > +
> > +#define PCIE_TRANS_TABLE_BASE_REG 0x800
> > +#define PCIE_ATR_SRC_ADDR_MSB_OFFSET 0x4
> > +#define PCIE_ATR_TRSL_ADDR_LSB_OFFSET 0x8
> > +#define PCIE_ATR_TRSL_ADDR_MSB_OFFSET 0xc
> > +#define PCIE_ATR_TRSL_PARAM_OFFSET 0x10
> > +#define PCIE_ATR_TLB_SET_OFFSET 0x20
> > +
> > +#define PCIE_MAX_TRANS_TABLES 8
> > +#define PCIE_ATR_EN BIT(0)
> > +#define PCIE_ATR_SIZE(size) \
> > + (((((size) - 1) << 1) & GENMASK(6, 1)) | PCIE_ATR_EN)
> > +#define PCIE_ATR_ID(id) ((id) & GENMASK(3, 0))
> > +#define PCIE_ATR_TYPE_MEM PCIE_ATR_ID(0)
> > +#define PCIE_ATR_TYPE_IO PCIE_ATR_ID(1)
> > +#define PCIE_ATR_TLP_TYPE(type) (((type) << 16) & GENMASK(18, 16))
> > +#define PCIE_ATR_TLP_TYPE_MEM PCIE_ATR_TLP_TYPE(0)
> > +#define PCIE_ATR_TLP_TYPE_IO PCIE_ATR_TLP_TYPE(2)
> > +
> > +/**
> > + * struct mtk_pcie_msi - MSI information for each set
> > + * @base: IO mapped register base
> > + * @irq: MSI set Interrupt number
> > + * @index: MSI set number
> > + * @msg_addr: MSI message address
> > + * @domain: IRQ domain
> > + */
> > +struct mtk_pcie_msi {
> > + void __iomem *base;
> > + unsigned int irq;
> > + int index;
> > + phys_addr_t msg_addr;
> > + struct irq_domain *domain;
> > +};
>
> This looks odd. You seem to say that this covers a set if MSIs, and
> yet the irq field here clearly isn't part of a set. Is that per MSI
> instead? Either way, something is not quite as it should be.
>
Appreciate all these comments, please allow me to explain the MSI
interrupt design in this HW.
The HW design of MSI interrupts will be like the following:
+-----+
| GIC |
+-----+
^
|
|[port->irq]
|
+-+-+-+-+-+-+-+-+
|0|1|2|3|4|5|6|7|[PCIe intc]
+-+-+-+-+-+-+-+-+
^ ^ ^
| | ... |[msi_info->irq]
+-------+ +------+ +-----------+
| | |
+-+-+---+--+--+ +-+-+---+--+--+ +-+-+---+--+--+
|0|1|...|30|31| |0|1|...|30|31| |0|1|...|30|31|[MSI sets]
+-+-+---+--+--+ +-+-+---+--+--+ +-+-+---+--+--+
^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^
| | | | | | | | | | | | [MSI irq]
| | | | | | | | | | | |
(MSI SET0) (MSI SET1) ... (MSI SET7)
In software parts, the port->msi_top_domain is created to maintains 8
MSI IRQs from the PCIe intc layer, its hardware IRQ will be mapped to
msi_info->irq by irq_create_mapping.
The port->msi_domain contains 256 MSI IRQs in total, it consist of 8 MSI
sets, and each MSI set contains 32 MSI IRQs.
The structure of mtk_pcie_msi is used to describe the MSI set, I think
it will be more convenient and comply with the HW design when use this
structure, we can get the information of MSI set directly, instead of
calculated by port->base.
When a MSI interrupt is received, the interrupt handle flow will be like
the following:
mtk_pcie_irq_handler (port->irq)
|
|(find mapping in msi_top_domain)
|
v
mtk_pcie_msi_handler (msi_info->irq)
|
|(find mapping in msi_domain)
|
v
handle_edge_irq (MSI irq)
|
|
v
(dispatch to device handler)
Yes, I had to admit that it's not a quite good solution of irqdomains,
since the local irq domain is partial coupled with the standard PCI MSI
irqdomain.
Should I need to create another irqdomain to maintain the MSI sets
layer, and set it as the parent domain of PCI MSI domain? I really need
your suggestions, thanks a lot.
> > +
> > +/**
> > + * struct mtk_pcie_port - PCIe port information
> > + * @dev: PCIe device
> > + * @base: IO mapped register base
> > + * @reg_base: Physical register base
> > + * @mac_reset: mac reset control
> > + * @phy_reset: phy reset control
> > + * @phy: PHY controller block
> > + * @clks: PCIe clocks
> > + * @num_clks: PCIe clocks count for this port
> > + * @irq: PCIe controller interrupt number
> > + * @intx_domain: legacy INTx IRQ domain
> > + * @msi_domain: MSI IRQ domain
> > + * @msi_top_domain: MSI IRQ top domain
> > + * @msi_info: MSI sets information
> > + * @lock: lock protecting IRQ bit map
> > + * @msi_irq_in_use: bit map for assigned MSI IRQ
> > + */
> > +struct mtk_pcie_port {
> > + struct device *dev;
> > + void __iomem *base;
> > + phys_addr_t reg_base;
> > + struct reset_control *mac_reset;
> > + struct reset_control *phy_reset;
> > + struct phy *phy;
> > + struct clk_bulk_data *clks;
> > + int num_clks;
> > +
> > + int irq;
> > + struct irq_domain *intx_domain;
> > + struct irq_domain *msi_domain;
> > + struct irq_domain *msi_top_domain;
> > + struct mtk_pcie_msi **msi_info;
> > + struct mutex lock;
> > + DECLARE_BITMAP(msi_irq_in_use, PCIE_MSI_IRQS_NUM);
> > +};
> > +
> > +/**
> > + * mtk_pcie_config_tlp_header
> > + * @bus: PCI bus to query
> > + * @devfn: device/function number
> > + * @where: offset in config space
> > + * @size: data size in TLP header
> > + *
> > + * Set byte enable field and device information in configuration TLP header.
> > + */
> > +static void mtk_pcie_config_tlp_header(struct pci_bus *bus, unsigned int devfn,
> > + int where, int size)
> > +{
> > + struct mtk_pcie_port *port = bus->sysdata;
> > + int bytes;
> > + u32 val;
> > +
> > + bytes = (GENMASK(size - 1, 0) & 0xf) << (where & 0x3);
> > +
> > + val = PCIE_CFG_FORCE_BYTE_EN | PCIE_CFG_BYTE_EN(bytes) |
> > + PCIE_CFG_HEADER(bus->number, devfn);
> > +
> > + writel(val, port->base + PCIE_CFGNUM_REG);
>
> Please convert all instances of writel/readl to their relaxed version,
> unless you can show that you need the extra memory ordering
> enforcement. This is specially valuable on an arm64 system (which
> MT8192 seems to be).
Sure, I will fix it in the next version.
>
> > +}
> > +
> > +static void __iomem *mtk_pcie_map_bus(struct pci_bus *bus, unsigned int devfn,
> > + int where)
> > +{
> > + struct mtk_pcie_port *port = bus->sysdata;
> > +
> > + return port->base + PCIE_CFG_OFFSET_ADDR + where;
> > +}
> > +
> > +static int mtk_pcie_config_read(struct pci_bus *bus, unsigned int devfn,
> > + int where, int size, u32 *val)
> > +{
> > + mtk_pcie_config_tlp_header(bus, devfn, where, size);
> > +
> > + return pci_generic_config_read32(bus, devfn, where, size, val);
> > +}
> > +
> > +static int mtk_pcie_config_write(struct pci_bus *bus, unsigned int devfn,
> > + int where, int size, u32 val)
> > +{
> > + mtk_pcie_config_tlp_header(bus, devfn, where, size);
> > +
> > + if (size <= 2)
> > + val <<= (where & 0x3) * 8;
> > +
> > + return pci_generic_config_write32(bus, devfn, where, 4, val);
> > +}
> > +
> > +static struct pci_ops mtk_pcie_ops = {
> > + .map_bus = mtk_pcie_map_bus,
> > + .read = mtk_pcie_config_read,
> > + .write = mtk_pcie_config_write,
> > +};
> > +
> > +static int mtk_pcie_set_trans_table(struct mtk_pcie_port *port,
> > + resource_size_t cpu_addr,
> > + resource_size_t pci_addr,
> > + resource_size_t size,
> > + unsigned long type, int num)
> > +{
> > + void __iomem *table;
> > + u32 val;
> > +
> > + if (num >= PCIE_MAX_TRANS_TABLES) {
> > + dev_err(port->dev, "not enough translate table[%d] for addr: %#llx, limited to [%d]\n",
> > + num, (unsigned long long) cpu_addr,
> > + PCIE_MAX_TRANS_TABLES);
> > + return -ENODEV;
> > + }
> > +
> > + table = port->base + PCIE_TRANS_TABLE_BASE_REG +
> > + num * PCIE_ATR_TLB_SET_OFFSET;
> > +
> > + writel(lower_32_bits(cpu_addr) | PCIE_ATR_SIZE(fls(size) - 1), table);
> > + writel(upper_32_bits(cpu_addr), table + PCIE_ATR_SRC_ADDR_MSB_OFFSET);
> > + writel(lower_32_bits(pci_addr), table + PCIE_ATR_TRSL_ADDR_LSB_OFFSET);
> > + writel(upper_32_bits(pci_addr), table + PCIE_ATR_TRSL_ADDR_MSB_OFFSET);
> > +
> > + if (type == IORESOURCE_IO)
> > + val = PCIE_ATR_TYPE_IO | PCIE_ATR_TLP_TYPE_IO;
> > + else
> > + val = PCIE_ATR_TYPE_MEM | PCIE_ATR_TLP_TYPE_MEM;
> > +
> > + writel(val, table + PCIE_ATR_TRSL_PARAM_OFFSET);
> > +
> > + return 0;
> > +}
> > +
> > +static int mtk_pcie_startup_port(struct mtk_pcie_port *port)
> > +{
> > + struct resource_entry *entry;
> > + struct pci_host_bridge *host = pci_host_bridge_from_priv(port);
> > + unsigned int table_index = 0;
> > + int err;
> > + u32 val;
> > +
> > + /* Set as RC mode */
> > + val = readl(port->base + PCIE_SETTING_REG);
> > + val |= PCIE_RC_MODE;
> > + writel(val, port->base + PCIE_SETTING_REG);
> > +
> > + /* Set class code */
> > + val = readl(port->base + PCIE_PCI_IDS_1);
> > + val &= ~GENMASK(31, 8);
> > + val |= PCI_CLASS(PCI_CLASS_BRIDGE_PCI << 8);
> > + writel(val, port->base + PCIE_PCI_IDS_1);
> > +
> > + /* Assert all reset signals */
> > + val = readl(port->base + PCIE_RST_CTRL_REG);
> > + val |= PCIE_MAC_RSTB | PCIE_PHY_RSTB | PCIE_BRG_RSTB | PCIE_PE_RSTB;
> > + writel(val, port->base + PCIE_RST_CTRL_REG);
> > +
> > + /* De-assert reset signals */
> > + val &= ~(PCIE_MAC_RSTB | PCIE_PHY_RSTB | PCIE_BRG_RSTB);
> > + writel(val, port->base + PCIE_RST_CTRL_REG);
> > +
> > + /* Delay 100ms to wait the reference clocks become stable */
> > + msleep(100);
> > +
> > + /* De-assert PERST# signal */
> > + val &= ~PCIE_PE_RSTB;
> > + writel(val, port->base + PCIE_RST_CTRL_REG);
> > +
> > + /* Check if the link is up or not */
> > + err = readl_poll_timeout(port->base + PCIE_LINK_STATUS_REG, val,
> > + !!(val & PCIE_PORT_LINKUP), 20,
> > + 50 * USEC_PER_MSEC);
> > + if (err) {
> > + val = readl(port->base + PCIE_LTSSM_STATUS_REG);
> > + dev_err(port->dev, "PCIe link down, ltssm reg val: %#x\n", val);
> > + return err;
> > + }
> > +
> > + /* Set PCIe translation windows */
> > + resource_list_for_each_entry(entry, &host->windows) {
> > + struct resource *res = entry->res;
> > + unsigned long type = resource_type(res);
> > + resource_size_t cpu_addr;
> > + resource_size_t pci_addr;
> > + resource_size_t size;
> > + const char *range_type;
> > +
> > + if (type == IORESOURCE_IO) {
> > + cpu_addr = pci_pio_to_address(res->start);
> > + range_type = "IO";
> > + } else if (type == IORESOURCE_MEM) {
> > + cpu_addr = res->start;
> > + range_type = "MEM";
> > + } else {
> > + continue;
> > + }
> > +
> > + pci_addr = res->start - entry->offset;
> > + size = resource_size(res);
> > + err = mtk_pcie_set_trans_table(port, cpu_addr, pci_addr, size,
> > + type, table_index);
> > + if (err)
> > + return err;
> > +
> > + dev_dbg(port->dev, "set %s trans window[%d]: cpu_addr = %#llx, pci_addr = %#llx, size = %#llx\n",
> > + range_type, table_index, (unsigned long long) cpu_addr,
> > + (unsigned long long) pci_addr,
> > + (unsigned long long) size);
> > +
> > + table_index++;
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +static inline struct mtk_pcie_msi *mtk_get_msi_info(struct mtk_pcie_port *port,
> > + unsigned long hwirq)
> > +{
> > + return port->msi_info[hwirq / PCIE_MSI_IRQS_PER_SET];
> > +}
> > +
> > +static int mtk_pcie_set_affinity(struct irq_data *data,
> > + const struct cpumask *mask, bool force)
> > +{
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + struct irq_data *port_data = irq_get_irq_data(port->irq);
>
> NAK. Either you are resolving an irqdata that is the same as the one
> as the one passed as a parameter (and it is thus pointless), or you
> are poking into some other interrupt, and it is a terrible layering
> violation. Either way, this must go.
>
> > + struct irq_chip *port_chip = irq_data_get_irq_chip(port_data);
> > + int ret;
> > +
> > + if (!port_chip || !port_chip->irq_set_affinity)
> > + return -EINVAL;
>
> In what circumstances can this fail?
>
> > +
> > + ret = port_chip->irq_set_affinity(port_data, mask, force);
>
> You really have to explain this indirection, and possibly remove it.
As the MSI HW designs described before, port->irq is the only interrupt
which connected with GIC, so the only way to set irq affinity is that to
use the irq_set_affinity callback of port_chip.
It seems there is no chance to implement this callback, I will replace
to "return -EINVAL" in the next version.
>
> > +
> > + irq_data_update_effective_affinity(data, mask);
> > +
> > + return ret;
> > +}
> > +
> > +static void mtk_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
> > +{
> > + struct mtk_pcie_msi *msi_info;
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long hwirq;
> > +
> > + msi_info = mtk_get_msi_info(port, data->hwirq);
>
> So msi_info *is* per interrupt, and yet contains information that is
> seemingly global. Please address this.
I will fix it in the next version.
>
> > + hwirq = data->hwirq % PCIE_MSI_IRQS_PER_SET;
> > +
> > + msg->address_hi = 0;
>
> What guarantees this?
It should not be set to 0 directly, I will replace to
upper_32_bits(msi_info->msi_addr) in the next version.
>
> > + msg->address_lo = lower_32_bits(msi_info->msg_addr);
> > + msg->data = hwirq;
> > + dev_dbg(port->dev, "msi#%#lx address_hi %#x address_lo %#x data %d\n",
> > + hwirq, msg->address_hi, msg->address_lo, msg->data);
>
> If address_hi is always 0, what is the point of printing it?
>
> > +}
> > +
> > +static void mtk_msi_irq_ack(struct irq_data *data)
> > +{
> > + struct mtk_pcie_msi *msi_info;
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long hwirq;
> > +
> > + msi_info = mtk_get_msi_info(port, data->hwirq);
> > + hwirq = data->hwirq % PCIE_MSI_IRQS_PER_SET;
> > +
> > + writel(BIT(hwirq), msi_info->base + PCIE_MSI_STATUS_OFFSET);
> > +}
> > +
> > +static void mtk_msi_irq_mask(struct irq_data *data)
> > +{
> > + struct mtk_pcie_msi *msi_info;
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long hwirq;
> > + u32 val;
> > +
> > + msi_info = mtk_get_msi_info(port, data->hwirq);
> > + hwirq = data->hwirq % PCIE_MSI_IRQS_PER_SET;
> > +
> > + val = readl(msi_info->base + PCIE_MSI_ENABLE_OFFSET);
> > + val &= ~BIT(hwirq);
> > + writel(val, msi_info->base + PCIE_MSI_ENABLE_OFFSET);
> > +
> > + pci_msi_mask_irq(data);
>
> This is at the wrong location. The common idiom is to split PCI and
> generic MSI. Please follow the same idiom in order to keep the code
> maintainable.
>
> > +}
> > +
> > +static void mtk_msi_irq_unmask(struct irq_data *data)
> > +{
> > + struct mtk_pcie_msi *msi_info;
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long hwirq;
> > + u32 val;
> > +
> > + msi_info = mtk_get_msi_info(port, data->hwirq);
> > + hwirq = data->hwirq % PCIE_MSI_IRQS_PER_SET;
> > +
> > + val = readl(msi_info->base + PCIE_MSI_ENABLE_OFFSET);
> > + val |= BIT(hwirq);
> > + writel(val, msi_info->base + PCIE_MSI_ENABLE_OFFSET);
> > +
> > + pci_msi_unmask_irq(data);
>
> Same thing here.
>
> > +}
> > +
> > +static struct irq_chip mtk_msi_irq_chip = {
> > + .irq_ack = mtk_msi_irq_ack,
> > + .irq_compose_msi_msg = mtk_compose_msi_msg,
> > + .irq_mask = mtk_msi_irq_mask,
> > + .irq_unmask = mtk_msi_irq_unmask,
> > + .irq_set_affinity = mtk_pcie_set_affinity,
> > + .name = "PCIe",
> > +};
> > +
> > +static irq_hw_number_t mtk_pcie_get_hwirq(struct msi_domain_info *info,
> > + msi_alloc_info_t *arg)
> > +{
> > + struct msi_desc *desc = arg->desc;
> > + irq_hw_number_t hwirq = arg->hwirq;
> > +
> > + arg->hwirq += desc->nvec_used;
> > +
> > + return hwirq;
> > +}
>
> Please follow the common MSI flow. There is absolutely no need to
> reinvent the wheel.
>
> > +
> > +static void mtk_pcie_msi_free(struct irq_domain *domain,
> > + struct msi_domain_info *info, unsigned int virq)
> > +{
> > + struct irq_data *data = irq_domain_get_irq_data(domain, virq);
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > +
> > + mutex_lock(&port->lock);
> > +
> > + bitmap_clear(port->msi_irq_in_use, data->hwirq, 1);
> > +
> > + mutex_unlock(&port->lock);
> > +}
> > +
> > +static int mtk_pcie_msi_prepare(struct irq_domain *domain, struct device *dev,
> > + int nvec, msi_alloc_info_t *arg)
> > +{
> > + struct msi_domain_info *info = domain->host_data;
> > + struct mtk_pcie_port *port = info->chip_data;
> > + struct msi_desc *desc = first_pci_msi_entry(to_pci_dev(dev));
> > + int hwirq, ret = 0;
> > +
> > + mutex_lock(&port->lock);
> > +
> > + if (desc->msi_attrib.is_msix) {
> > + int i;
> > + unsigned long bit;
> > +
> > + for (i = 0; i < nvec; i++) {
> > + bit = find_first_zero_bit(port->msi_irq_in_use,
> > + PCIE_MSI_IRQS_NUM);
> > + if (bit >= PCIE_MSI_IRQS_NUM) {
> > + ret = -ENOSPC;
> > + goto msi_prepare_out;
> > + } else {
> > + set_bit(bit, port->msi_irq_in_use);
> > + }
> > + }
> > +
> > + hwirq = bit - nvec;
> > + } else {
> > + hwirq = bitmap_find_free_region(port->msi_irq_in_use,
> > + PCIE_MSI_IRQS_NUM,
> > + order_base_2(nvec));
> > + if (hwirq < 0)
> > + ret = -ENOSPC;
> > + }
> > +
> > +msi_prepare_out:
> > + mutex_unlock(&port->lock);
> > +
> > + if (ret)
> > + return ret;
> > +
> > + memset(arg, 0, sizeof(*arg));
> > + arg->hwirq = hwirq;
> > +
> > + return 0;
> > +}
>
> NAK. This should take place in the irqdomain alloc path. .prepare is
> reserved for HW that require per-device setup, and this doesn't
> interact with the HW at all.
>
> Your distinction between MSI-X and non-MSI-X also becomes moot once
> you place the allocator at the place where allocations take place.
>
> > +
> > +static void mtk_pcie_msi_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
> > +{
> > + arg->desc = desc;
> > +}
>
> NAK. Please use the common PCI flow.
>
> > +
> > +static struct msi_domain_ops mtk_msi_domain_ops = {
> > + .get_hwirq = mtk_pcie_get_hwirq,
> > + .msi_free = mtk_pcie_msi_free,
> > + .msi_prepare = mtk_pcie_msi_prepare,
> > + .set_desc = mtk_pcie_msi_set_desc,
> > +};
> > +
> > +static struct msi_domain_info mtk_msi_domain_info = {
> > + .flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_PCI_MSIX |
> > + MSI_FLAG_USE_DEF_CHIP_OPS | MSI_FLAG_MULTI_PCI_MSI),
> > + .chip = &mtk_msi_irq_chip,
> > + .ops = &mtk_msi_domain_ops,
> > + .handler = handle_edge_irq,
> > + .handler_name = "MSI",
> > +};
> > +
> > +static void mtk_msi_top_irq_eoi(struct irq_data *data)
> > +{
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long msi_irq = data->hwirq + PCIE_MSI_SHIFT;
> > +
> > + writel(BIT(msi_irq), port->base + PCIE_INT_STATUS_REG);
> > +}
> > +
> > +static struct irq_chip mtk_msi_top_irq_chip = {
> > + .irq_eoi = mtk_msi_top_irq_eoi,
> > + .name = "PCIe",
> > +};
> > +
> > +static void mtk_pcie_msi_handler(struct irq_desc *desc)
> > +{
> > + struct mtk_pcie_msi *msi_info = irq_desc_get_handler_data(desc);
> > + struct irq_chip *irqchip = irq_desc_get_chip(desc);
> > + unsigned long msi_enable, msi_status;
> > + unsigned int virq;
> > + irq_hw_number_t bit, hwirq;
> > +
> > + chained_irq_enter(irqchip, desc);
> > +
> > + msi_enable = readl(msi_info->base + PCIE_MSI_ENABLE_OFFSET);
> > + while ((msi_status = readl(msi_info->base + PCIE_MSI_STATUS_OFFSET) &
> > + msi_enable)) {
> > + for_each_set_bit(bit, &msi_status, PCIE_MSI_IRQS_PER_SET) {
> > + hwirq = bit + msi_info->index * PCIE_MSI_IRQS_PER_SET;
> > + virq = irq_find_mapping(msi_info->domain, hwirq);
> > + generic_handle_irq(virq);
> > + }
> > + }
> > +
> > + chained_irq_exit(irqchip, desc);
>
> If your MSIs are behind a muxing interrupt you *cannot* implement an
> affinity setting callback, as this changes the affinity of *all* MSIs
> in one go, which is a userspace ABI violation. I can't believe that
> this is almost 2021, and people are still shipping HW that is that
> broken :-(.
Sorry, I haven't noticed the potential risk of userspace ABI violation,
thanks for pointing out.
>
> > +}
> > +
> > +static int mtk_msi_top_domain_map(struct irq_domain *domain,
> > + unsigned int virq, irq_hw_number_t hwirq)
> > +{
> > + struct mtk_pcie_port *port = domain->host_data;
> > + struct mtk_pcie_msi *msi_info = port->msi_info[hwirq];
> > +
> > + irq_domain_set_info(domain, virq, hwirq,
> > + &mtk_msi_top_irq_chip, domain->host_data,
> > + mtk_pcie_msi_handler, msi_info, NULL);
> > +
> > + return 0;
> > +}
> > +
> > +static const struct irq_domain_ops mtk_msi_top_domain_ops = {
> > + .map = mtk_msi_top_domain_map,
>
> NAK. MSIs do not need a mapping callback. This belongs to the
> irqdomain alloc callback, which you don't implement.
>
> > +};
> > +
> > +static void mtk_intx_mask(struct irq_data *data)
> > +{
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + u32 val;
> > +
> > + val = readl(port->base + PCIE_INT_ENABLE_REG);
> > + val &= ~BIT(data->hwirq + PCIE_INTX_SHIFT);
> > + writel(val, port->base + PCIE_INT_ENABLE_REG);
> > +}
> > +
> > +static void mtk_intx_unmask(struct irq_data *data)
> > +{
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + u32 val;
> > +
> > + val = readl(port->base + PCIE_INT_ENABLE_REG);
> > + val |= BIT(data->hwirq + PCIE_INTX_SHIFT);
> > + writel(val, port->base + PCIE_INT_ENABLE_REG);
> > +}
> > +
> > +static void mtk_intx_eoi(struct irq_data *data)
> > +{
> > + struct mtk_pcie_port *port = irq_data_get_irq_chip_data(data);
> > + unsigned long hwirq;
> > +
> > + /**
> > + * As an emulated level IRQ, its interrupt status will remain
> > + * until the corresponding de-assert message is received; hence that
> > + * the status can only be cleared when the interrupt has been serviced.
> > + */
> > + hwirq = data->hwirq + PCIE_INTX_SHIFT;
> > + writel(BIT(hwirq), port->base + PCIE_INT_STATUS_REG);
>
> All of this is the description of a level interrupt, so this is pretty
> much devoid of any information as to *why* you need to write to clear
> this bit. What happens if the interrupt is still asserted because
> nothing has handled it? Without any further information, this looks
> terribly wrong.
Sorry, this comment should be used to describe the mtk_intx_eoi
function, it cause misunderstandings at this place. I just want to add
the comment to explain that why this interrupt needs to be acked at the
end of interrupt. I will move it to the front of mtk_intx_eoi in the
next version.
>
> > +}
> > +
> > +static struct irq_chip mtk_intx_irq_chip = {
> > + .irq_mask = mtk_intx_mask,
> > + .irq_unmask = mtk_intx_unmask,
> > + .irq_eoi = mtk_intx_eoi,
> > + .irq_set_affinity = mtk_pcie_set_affinity,
> > + .name = "PCIe",
> > +};
> > +
> > +static int mtk_pcie_intx_map(struct irq_domain *domain, unsigned int irq,
> > + irq_hw_number_t hwirq)
> > +{
> > + irq_set_chip_and_handler_name(irq, &mtk_intx_irq_chip,
> > + handle_fasteoi_irq, "INTx");
> > + irq_set_chip_data(irq, domain->host_data);
> > +
> > + return 0;
> > +}
> > +
> > +static const struct irq_domain_ops intx_domain_ops = {
> > + .map = mtk_pcie_intx_map,
> > +};
> > +
> > +static int mtk_pcie_init_irq_domains(struct mtk_pcie_port *port,
> > + struct device_node *node)
> > +{
> > + struct device *dev = port->dev;
> > + struct device_node *intc_node;
> > + struct fwnode_handle *fwnode = of_node_to_fwnode(node);
> > + struct mtk_pcie_msi *msi_info;
> > + struct msi_domain_info *info;
> > + int i, ret;
> > +
> > + /* Setup INTx */
> > + intc_node = of_get_child_by_name(node, "interrupt-controller");
> > + if (!intc_node) {
> > + dev_err(dev, "missing PCIe Intc node\n");
> > + return -ENODEV;
> > + }
> > +
> > + port->intx_domain = irq_domain_add_linear(intc_node, PCI_NUM_INTX,
> > + &intx_domain_ops, port);
> > + if (!port->intx_domain) {
> > + dev_err(dev, "failed to get INTx IRQ domain\n");
> > + return -ENODEV;
> > + }
> > +
> > + /* Setup MSI */
> > + mutex_init(&port->lock);
> > +
> > + info = devm_kzalloc(dev, sizeof(*info), GFP_KERNEL);
> > + if (!info)
> > + return -ENOMEM;
> > +
> > + memcpy(info, &mtk_msi_domain_info, sizeof(*info));
> > + info->chip_data = port;
> > +
> > + port->msi_domain = pci_msi_create_irq_domain(fwnode, info, NULL);
> > + if (!port->msi_domain) {
> > + dev_info(dev, "failed to create MSI domain\n");
> > + ret = -ENODEV;
> > + goto err_msi_domain;
> > + }
> > +
> > + /* Enable MSI and setup PCIe domains */
> > + port->msi_top_domain = irq_domain_add_hierarchy(NULL, 0, 0, node,
> > + &mtk_msi_top_domain_ops,
> > + port);
> > + if (!port->msi_top_domain) {
> > + dev_info(dev, "failed to create MSI top domain\n");
> > + ret = -ENODEV;
> > + goto err_msi_top_domain;
> > + }
> > +
> > + port->msi_info = devm_kzalloc(dev, PCIE_MSI_SET_NUM, GFP_KERNEL);
> > + if (!port->msi_info) {
> > + ret = -ENOMEM;
> > + goto err_msi_info;
> > + }
> > +
> > + for (i = 0; i < PCIE_MSI_SET_NUM; i++) {
> > + int offset = i * PCIE_MSI_SET_OFFSET;
> > + u32 val;
> > +
> > + msi_info = devm_kzalloc(dev, sizeof(*msi_info), GFP_KERNEL);
> > + if (!msi_info) {
> > + ret = -ENOMEM;
> > + goto err_msi_set;
> > + }
> > +
> > + msi_info->base = port->base + PCIE_MSI_ADDR_BASE_REG + offset;
> > + msi_info->msg_addr = port->reg_base + PCIE_MSI_ADDR_BASE_REG +
> > + offset;
>
> All of this can be derived from i (and this hwirq) at runtime. Why do
> we need to cache this information?
I try to use msi_info structure to describe the MSI set, I think this
will be more comply with the HW design and more easier to understand the
interrupt handle flow.
>
> > +
> > + writel(lower_32_bits(msi_info->msg_addr), msi_info->base);
> > +
> > + msi_info->index = i;
> > + msi_info->domain = port->msi_domain;
>
> If this is common to all sets, why isn't it global?
I will fix it in the next version.
>
> > +
> > + port->msi_info[i] = msi_info;
> > +
> > + /* Alloc IRQ for each MSI set */
> > + msi_info->irq = irq_create_mapping(port->msi_top_domain, i);
> > + if (!msi_info->irq) {
> > + dev_info(dev, "allocate MSI top IRQ failed\n");
> > + ret = -ENOSPC;
> > + goto err_msi_set;
> > + }
> > +
> > + val = readl(port->base + PCIE_INT_ENABLE_REG);
> > + val |= BIT(i + PCIE_MSI_SHIFT);
> > + writel(val, port->base + PCIE_INT_ENABLE_REG);
> > +
> > + val = readl(port->base + PCIE_MSI_SET_ENABLE_REG);
> > + val |= BIT(i);
> > + writel(val, port->base + PCIE_MSI_SET_ENABLE_REG);
> > + }
> > +
> > + return 0;
> > +
> > +err_msi_set:
> > + while (i-- > 0) {
> > + msi_info = port->msi_info[i];
> > + irq_dispose_mapping(msi_info->irq);
> > + }
> > +err_msi_info:
> > + irq_domain_remove(port->msi_top_domain);
> > +err_msi_top_domain:
> > + irq_domain_remove(port->msi_domain);
> > +err_msi_domain:
> > + irq_domain_remove(port->intx_domain);
> > +
> > + return ret;
> > +}
>
> At this stage, I stopped, because there is really a lot to change in
> this driver before it is in an acceptable shape. Please use the common
> idioms established for existing MSI implementations.
>
> I'd appreciate if you kept me posted on the following versions of this
> driver.
>
> Thanks,
>
> M.
>
