Hello Ilya, On Tue, Apr 13, 2021 at 9:22 AM Ilya Lipnitskiy <ilya.lipnitskiy@xxxxxxxxx> wrote: > Mirror pci-rt3883 fix from commit e5067c718b3a ("MIPS: pci-rt3883: > Remove odd locking in PCI config space access code"). pci-rt2880 shares > the driver layout with pci-rt3883 and the same reasons apply. > > Caller (generic PCI code) already does proper locking, so no need to add > another one here. Local PCI read/write functions are never called > simultaneously, also they do not require synchronization with the PCI > controller ops, since they are used before the controller registration. > > Suggested-by: Sergey Ryazanov <ryazanov.s.a@xxxxxxxxx> > Signed-off-by: Ilya Lipnitskiy <ilya.lipnitskiy@xxxxxxxxx> > --- > arch/mips/pci/pci-rt2880.c | 13 ------------- > 1 file changed, 13 deletions(-) > > diff --git a/arch/mips/pci/pci-rt2880.c b/arch/mips/pci/pci-rt2880.c > index 19f7860fb28b..b4ee07cbcf2a 100644 > --- a/arch/mips/pci/pci-rt2880.c > +++ b/arch/mips/pci/pci-rt2880.c > @@ -41,7 +41,6 @@ > #define RT2880_PCI_REG_ARBCTL 0x80 > > static void __iomem *rt2880_pci_base; > -static DEFINE_SPINLOCK(rt2880_pci_lock); > > static u32 rt2880_pci_reg_read(u32 reg) > { > @@ -63,7 +62,6 @@ static inline u32 rt2880_pci_get_cfgaddr(unsigned int bus, unsigned int slot, > static int rt2880_pci_config_read(struct pci_bus *bus, unsigned int devfn, > int where, int size, u32 *val) > { > - unsigned long flags; > u32 address; > u32 data; > int busn = 0; > @@ -74,10 +72,8 @@ static int rt2880_pci_config_read(struct pci_bus *bus, unsigned int devfn, > address = rt2880_pci_get_cfgaddr(busn, PCI_SLOT(devfn), PCI_FUNC(devfn), > where); > > - spin_lock_irqsave(&rt2880_pci_lock, flags); > rt2880_pci_reg_write(address, RT2880_PCI_REG_CONFIG_ADDR); > data = rt2880_pci_reg_read(RT2880_PCI_REG_CONFIG_DATA); > - spin_unlock_irqrestore(&rt2880_pci_lock, flags); > > switch (size) { > case 1: > @@ -97,7 +93,6 @@ static int rt2880_pci_config_read(struct pci_bus *bus, unsigned int devfn, > static int rt2880_pci_config_write(struct pci_bus *bus, unsigned int devfn, > int where, int size, u32 val) > { > - unsigned long flags; > u32 address; > u32 data; > int busn = 0; > @@ -108,7 +103,6 @@ static int rt2880_pci_config_write(struct pci_bus *bus, unsigned int devfn, > address = rt2880_pci_get_cfgaddr(busn, PCI_SLOT(devfn), PCI_FUNC(devfn), > where); > > - spin_lock_irqsave(&rt2880_pci_lock, flags); > rt2880_pci_reg_write(address, RT2880_PCI_REG_CONFIG_ADDR); > data = rt2880_pci_reg_read(RT2880_PCI_REG_CONFIG_DATA); > > @@ -127,7 +121,6 @@ static int rt2880_pci_config_write(struct pci_bus *bus, unsigned int devfn, > } > > rt2880_pci_reg_write(data, RT2880_PCI_REG_CONFIG_DATA); > - spin_unlock_irqrestore(&rt2880_pci_lock, flags); > > return PCIBIOS_SUCCESSFUL; > } > @@ -159,31 +152,25 @@ static struct pci_controller rt2880_pci_controller = { > > static inline u32 rt2880_pci_read_u32(unsigned long reg) > { > - unsigned long flags; > u32 address; > u32 ret; > > address = rt2880_pci_get_cfgaddr(0, 0, 0, reg); > > - spin_lock_irqsave(&rt2880_pci_lock, flags); > rt2880_pci_reg_write(address, RT2880_PCI_REG_CONFIG_ADDR); > ret = rt2880_pci_reg_read(RT2880_PCI_REG_CONFIG_DATA); > - spin_unlock_irqrestore(&rt2880_pci_lock, flags); > > return ret; > } > > static inline void rt2880_pci_write_u32(unsigned long reg, u32 val) > { > - unsigned long flags; > u32 address; > > address = rt2880_pci_get_cfgaddr(0, 0, 0, reg); > > - spin_lock_irqsave(&rt2880_pci_lock, flags); > rt2880_pci_reg_write(address, RT2880_PCI_REG_CONFIG_ADDR); > rt2880_pci_reg_write(val, RT2880_PCI_REG_CONFIG_DATA); > - spin_unlock_irqrestore(&rt2880_pci_lock, flags); > } > > int pcibios_map_irq(const struct pci_dev *dev, u8 slot, u8 pin) RT2880 PCI driver calls rt2880_pci_write_u32()/rt2880_pci_read_u32() from pcibios_map_irq(), which is called outside the scope of the pci_lock spinlock that is defined in /drivers/pci/access.c. So it looks like this could lead to a race between rt2880_pci_write_u32()/rt2880_pci_read_u32() and rt2880_pci_config_read()/rt2880_pci_config_write() functions. The code that uses rt2880_pci_write_u32()/rt2880_pci_read_u32() in the pcibios_map_irq() duplicates a BAR initialization procedure, which is already performed by the rt288x_pci_probe(). Maybe we should remove duplicated code in the pcibios_map_irq() to reduce duplication and to avoid possible race in configuration space access? If you fix this possible race, feel free to add in the next patch version my Reviewed-by: Sergey Ryazanov <ryazanov.s.a@xxxxxxxxx> -- Sergey