>-----Original Message----- >From: Andrew Lunn <[email protected]> >Sent: Tuesday, November 17, 2020 4:45 AM >To: Claudiu Manoil <[email protected]> >Cc: [email protected]; Jakub Kicinski <[email protected]>; David S . >Miller <[email protected]>; Alexandru Marginean ><[email protected]>; Vladimir Oltean ><[email protected]> >Subject: Re: [PATCH net] enetc: Workaround for MDIO register access issue > >> +static inline void enetc_lock_mdio(void) >> +{ >> + read_lock(&enetc_mdio_lock); >> +} >> + > >> +static inline u32 _enetc_rd_mdio_reg_wa(void __iomem *reg) >> +{ >> + unsigned long flags; >> + u32 val; >> + >> + write_lock_irqsave(&enetc_mdio_lock, flags); >> + val = ioread32(reg); >> + write_unlock_irqrestore(&enetc_mdio_lock, flags); >> + >> + return val; >> +} > >Can you mix read_lock() with write_lock_irqsave()? Normal locks you >should not mix, so i assume read/writes also cannot be mixed? >
Not sure I understand your concerns, but this is the readers-writers locking scheme. The readers (read_lock) are "lightweight", they get the most calls, can be taken from any context including interrupt context, and compete only with the writers (write_lock). The writers can take the lock only when there are no readers holding it, and the writer must insure that it doesn't get preempted (by interrupts etc.) when holding the lock (irqsave). The good part is that mdio operations are not frequent. Also, we had this code out of the tree for quite some time, it's well exercised.
