Re: 2.6.18-rc5-mm1: drivers/infiniband/hw/amso1100/c2.c compileerror

[Andrew Morton]

On Fri, 01 Sep 2006 14:05:36 -0700
Roland Dreier <rdreier@xxxxxxxxx> wrote:

>     Andrew> If the hardware/driver absolutely requires that the 64-bit
>     Andrew> write be atomic on-the-bus then sure, the fix is to
>     Andrew> disable that driver on that architecture in Kconfig.
> 
>     Andrew> If, however, the atomicity requirement is a software thing
>     Andrew> (we need to be atomic against other CPU reads and writes)
>     Andrew> then that can be solved with locking, and we can design
>     Andrew> APIs for this which can be implemented efficiently on all
>     Andrew> architectures.
> 
> It seems that there are cases of both.  ipath needs actual 64-bit bus
> transactions to work properly.

If we define __raw_writeq() to be 64-bit-atomic-on-the-bus then an
appropriate solution for ipath would be to call __raw_writeq() directly. 
If the arch cannot implement __raw_write() then build error -> Kconfig fix.

>  mthca needs to make sure that if
> doorbell writes are split into two 32-bit halves, then no other writes
> go to the same MMIO page in between the halves.
> 
> What do you think the API would look like?  Something along the lines
> of mthca_doorbell.h, where we have macros for
> 
> DECLARE_WRITEQ_LOCK()
> INIT_WRITEQ_LOCK()
> GET_WRITEQ_LOCK()
> 
> which get stubbed out on architectures where writeq is already atomic,
> and then pass the lock into writeq()?
> 
> But then you probably need some Kconfig symbol to say if writeq() is
> really atomic or just software atomic (for ipath et al to depend on).
> 

It depends on how many other devices have (or are expected to have)
mthca-like requirements.  If the answer is "very few, maybe none" then
perhaps we don't need to go designing generic interfaces to support such
things.

As for interfaces, umm, something like

#ifdef CONFIG_ARCH_HAS_64BIT_ATOMIC_MMIO_WRITES

struct be64_port {
	void __iomem *addr;
};

static inline void atomic_be64_mmio_write(u64 v, struct be64_port *port)
{
	__raw_writeq(v, port->addr);
}

#define be64_port_init(port, addr)
	port->addr = addr;

#define be64_port_init_external_locking(port, addr, lockp)
	be64_port_init(port, addr)


#else


struct be64_port {
	void __iomem *addr;
	spinlock_t lock;
	spinlock_t *lockp;
};

static inline void atomic_be64_mmio_write(u64 v, struct be64_port *port)
{
	unsigned long flags;
	
	spin_lock_irqsave(port->lockp, flags);
	__raw_writel(...);
	__raw_writel(...);
	spin_unlock_irqrestore(port->lockp, flags);
}

#define be64_port_init(port, addr)
	spin_lock_init(&port->lock);
	port->lockp = &port->lock;
	port->addr = addr;

#define be64_port_init_external_locking(port, addr, lockp)
	port->lockp = lockp;
	port->addr = addr;

#endif

perhaps?

btw, 32-bit mthca_write64() is downright scary from an endianness POV.  I
guess it's right, but I wouldn't label it "obviously correct" ;)


-- 
VGER BF report: H 0
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/