[PATCH v1 1/2] phy: keystone: serdes driver for gbe 10gbe and pcie

[Kwok, WingMan]

Hello,

> -----Original Message-----
> From: Arnd Bergmann [mailto:arnd at arndb.de]
> Sent: Thursday, October 15, 2015 4:53 PM
> To: Kwok, WingMan
> Cc: linux-arm-kernel at lists.infradead.org; robh+dt at kernel.org;
> pawel.moll at arm.com; mark.rutland at arm.com; ijc+devicetree at hellion.org.uk;
> galak at codeaurora.org; KISHON VIJAY ABRAHAM; Quadros, Roger; Karicheri,
> Muralidharan; bhelgaas at google.com; ssantosh at kernel.org;
> linux at arm.linux.org.uk; devicetree at vger.kernel.org; linux-
> kernel at vger.kernel.org; linux-pci at vger.kernel.org
> Subject: Re: [PATCH v1 1/2] phy: keystone: serdes driver for gbe 10gbe and
> pcie
> 
> On Thursday 15 October 2015 20:08:32 Kwok, WingMan wrote:
> >
> > > > +#define reg_rmw(addr, value, mask) \
> > > > +   __raw_writel(((__raw_readl(addr) & (~(mask))) | \
> > > > +                   (value & (mask))), (addr))
> > >
> > > not endian safe, and potentially racy.
> > >
> >
> > will change to
> >
> > #define reg_rmw(addr, value, mask) \
> >         writel(((readl(addr) & (~(mask))) | \
> >                         (value & (mask))), (addr))
> 
> Ok, sounds good. Note that if any of this is performance critical,
> better use readl_relaxed(), but as long as this is just for setup
> code and not for data transfers, staying with readl() as you
> suggest is better.
> 

since this is only for initialization, I'll probably stay with readl().

> > > > +static inline void _kserdes_reset_cdr(void __iomem *sregs, int lane)
> > > > +{
> > > > +   /* toggle signal detect */
> > > > +   _kserdes_force_signal_detect_low(sregs, lane);
> > > > +   mdelay(1);
> > > > +   _kserdes_force_signal_detect_high(sregs, lane);
> > > > +}
> > >
> > > Can you change the code so you can use msleep(1) here?
> > >
> >
> > will replace delays with usleep_range()
> 
> Ok.
> 
> > > > +
> > > > +   do {
> > > > +           mdelay(10);
> > > > +           memset(lane_down, 0, sizeof(lane_down));
> > > > +
> > > > +           link_up = _kserdes_check_link_status(dev, sregs,
> > > > +                                                pcsr_regmap, lanes,
> > > > +                                                lanes_enable,
> > > > +                                                current_state,
> lane_down);
> > > > +
> > > > +           /* if we did not get link up then wait 100ms
> > > > +            * before calling it again
> > > > +            */
> > > > +           if (link_up)
> > > > +                   break;
> > > > +
> > > > +           for (i = 0; i < lanes; i++) {
> > > > +                   if ((lanes_enable & (1 << i)) && lane_down[i])
> > > > +                           dev_dbg(dev,
> > > > +                                   "XGE: detected lane down on lane
> %d\n",
> > > > +                                   i);
> > > > +           }
> > > > +
> > > > +           if (++retries > 100)
> > > > +                   return -ETIMEDOUT;
> > > > +
> > > > +   } while (!link_up);
> > >
> > > an more importantly here. Blocking the CPU for over one second is not
> good.
> > >
> > > Any use of mdelay() should have a comment explaining why you cannot use
> > > msleep() in that instance.
> > >
> >
> > will replace delays with usleep_range()
> 
> Here you have to be careful with the total runtime. Using usleep_range()
> is a good idea, and you can have a particularly wide range, but then you
> should changen the timeout condition from number of retries to total
> elapsed time like
> 
> 	unsigned long timeout = jiffies + HZ; /* 1 second maximum */
> 	do {
> 		...
> 
> 		if (link_up)
> 			break;
> 
> 		if (time_after(jiffies, timeout)
> 			return -ETIMEOUT;
> 
> 		usleep_range(1000, 50000);
> 	} while (1);
> 

will do.  

> 	Arnd

Thanks so much for your comments.
WingMan