lkml.org 
[lkml]   [2019]   [Feb]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC PATCH] docs/memory-barriers.txt: Rewrite "KERNEL I/O BARRIER EFFECTS" section
On Tue, Feb 19, 2019 at 02:01:50PM +0100, Arnd Bergmann wrote:
> On Tue, Feb 19, 2019 at 12:36 PM Will Deacon <will.deacon@arm.com> wrote:
> > On Tue, Feb 19, 2019 at 12:31:50PM +0100, Arnd Bergmann wrote:
> > > On Tue, Feb 19, 2019 at 11:27 AM Thomas Petazzoni
> > > <thomas.petazzoni@bootlin.com> wrote:
> > >
> > > I think an example of this would be a driver using outb() to disable
> > > an interrupt, and then relying on the the interrupt no longer happening
> > > after the outb().
> >
> > Isn't that racy already? i.e. the interrupt could fire just before you
> > disabled it, but not get delivered by the irq controller until after you'd
> > disabled it at the device?
>
> Probably, I had a hard enough time trying to come up with any example ;-)

You and me both!

> One reference to non-posted transaction in a comment is in
> drivers/net/ethernet/dec/tulip/de4x5.c:
>
> /*
> ** The DE4X5 interrupt handler.
> **
> ** I/O Read/Writes through intermediate PCI bridges are never 'posted',
> ** so that the asserted interrupt always has some real data to work with -
> ** if these I/O accesses are ever changed to memory accesses, ensure the
> ** STS write is read immediately to complete the transaction if the adapter
> ** is not on bus 0. Lost interrupts can still occur when the PCI bus load
> ** is high and descriptor status bits cannot be set before the associated
> ** interrupt is asserted and this routine entered.
> */

Thankfully, that driver is both old and non-portable:

depends on VIRT_TO_BUS || ALPHA || PPC || SPARC

so I'm not especially concerned about it. Judging by the comment, we'd
need to add something like:


diff --git a/drivers/net/ethernet/dec/tulip/de4x5.c b/drivers/net/ethernet/dec/tulip/de4x5.c
index 66535d1653f6..c85089f65b0e 100644
--- a/drivers/net/ethernet/dec/tulip/de4x5.c
+++ b/drivers/net/ethernet/dec/tulip/de4x5.c
@@ -1556,6 +1556,10 @@ de4x5_interrupt(int irq, void *dev_id)
sts = inl(DE4X5_STS); /* Read IRQ status */
outl(sts, DE4X5_STS); /* Reset the board interrupts */

+ /* Beware of PCI posted writes */
+ if (IS_ENABLED(CONFIG_ARM64) && lp->bus_num)
+ inl(DE4X5_STS);
+
if (!(sts & lp->irq_mask)) break;/* All done */
handled = 1;


if we wanted to get this working reliably on arm64. However, I'll be honest
and say we haven't had much demand for supporting DEC PCI devices yet :)

> I found another comment in the via-rhine driver:
>
> /* Beware of PCI posted writes */
> #define IOSYNC do { ioread8(ioaddr + StationAddr); } while (0)
>
> this one is used in the chip reset function, and in the transmit
> path.

Since this is doing a read-back, I take this comment as saying "posted
writes are a thing, so perform a read-back to force the prior write to
complete", which is fine.

Will

\
 
 \ /
  Last update: 2019-02-19 14:21    [W:0.060 / U:8.476 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site