Re: [PATCH v8 6/9] drivers: perf: hisi: Add support for Hisilicon Djtag driver

[John Garry]

On 14/06/2017 12:40, Will Deacon wrote:
> On Wed, Jun 14, 2017 at 12:35:07PM +0100, John Garry wrote:
> > On 14/06/2017 12:01, Will Deacon wrote:
> > > On Wed, Jun 14, 2017 at 11:42:30AM +0100, Mark Rutland wrote:
> > > > On Wed, Jun 14, 2017 at 11:06:58AM +0100, Will Deacon wrote:
> > > > > Apologies, I misunderstood your algorithm (I thought step (a) was on one CPU
> > > > > and step (b) was on another). Still, I don't understand the need for the
> > > > > timeout. If you instead read back the flag immediately, wouldn't it still
> > > > > work? e.g.
> > > > >
> > > > >
> > > > > lock:
> > > > >  Readl_relaxed flag
> > > > >  if (locked)
> > > > >    goto lock;
> > > > >
> > > > >  Writel_relaxed unique ID to flag
> > > > >  Readl flag
> > > > >  if (locked by somebody else)
> > > > >    goto lock;
> > > > >
> > > > > <critical section>
> > > > >
> > > > > unlock:
> > > > >  Writel unlocked value to flag
> > > > >
> > > > >
> > > > > Given that we're dealing with iomem, I think it will work, but I could be
> > > > > missing something obvious.
> > > >
> > > > Don't we have the race below where both threads can enter the critical
> > > > section?
> > > >
> > > > 	// flag f initial zero (unlocked)
> > > >
> > > > 	// t1, flag 1			// t2, flag 2
> > > > 	readl(f); // reads 0		l = readl(f); // reads 0
> > > >
> > > > 	<thinks lock is free>		<thinks lock is free>
> > > >
> > > > 	writel(1, f);
> > > > 	readl(f); // reads 1
> > > > 	<thinks lock owned>
> > > > 					writel(2, f);
> > > > 					readl(f) // reads 2
> > > > 					<thinks lock owned>
> > > >
> > > > 	<crticial section>		<critical section>
> > >
> > > Urgh, yeah, of course and *that's* what the udelay is trying to avoid,
> > > by "ensuring" that the <thinks lock is free> time and subsequent write
> > > propagation is all over before we re-read the flag.
> > >
> > > John -- how much space do you have on this device? Do you have, e.g. a byte
> > > for each CPU?
> >
> > Hi Will,
> >
> > To be clear, the agents in our case are the kernel and UEFI. Within the
> > kernel, we use a kernel spinlock to lock the same djtag between threads, for
> > these reasons:
> > - kernel has a native spinlock
>
> If we only have to effectively deal with two threads, then we might be able
> to use something like Dekker's.
>
> > - we are limited in locking values, as the lock flag is only a 8b field in
> > v2 hw (called module select)
>
> By 8b do you mean 8 bits or 8 bytes? If the latter, does it support sub-word
> accesses?

8 bits

So the size depends: on v1 hw is a 6-bit field in a 32-bit register 
(recent news to me), and on v2 hw it is a 8-bit field in a 32-bit register.

So for reading and writing the flag, we use readl/writel and also 
necessary shifts+masks. Obviously this is not atomic, but the whole 
process of write-and-check is not atomic - hence the delay.

I am not sure if sub-word access is required.

Thanks,
John

> Will
>
> .