From: Michael Holzheu <[email protected]>
[S390] cio: fix stsch_reset.
Copy inline assembly of stsch and add "memory" to clobber list in order
to prevent gcc from optimizing away the checking of the global variable
"pgm_check_occured".
Signed-off-by: Michael Holzheu <[email protected]>
Signed-off-by: Martin Schwidefsky <[email protected]>
---
drivers/s390/cio/cio.c | 13 ++++++++++---
1 files changed, 10 insertions(+), 3 deletions(-)
diff -urpN linux-2.6/drivers/s390/cio/cio.c linux-2.6-patched/drivers/s390/cio/cio.c
--- linux-2.6/drivers/s390/cio/cio.c 2006-12-28 00:33:22.000000000 +0100
+++ linux-2.6-patched/drivers/s390/cio/cio.c 2006-12-28 00:33:36.000000000 +0100
@@ -2,8 +2,7 @@
* drivers/s390/cio/cio.c
* S/390 common I/O routines -- low level i/o calls
*
- * Copyright (C) 1999-2002 IBM Deutschland Entwicklung GmbH,
- * IBM Corporation
+ * Copyright (C) IBM Corp. 1999,2006
* Author(s): Ingo Adlung ([email protected])
* Cornelia Huck ([email protected])
* Arnd Bergmann ([email protected])
@@ -881,10 +880,18 @@ static void cio_reset_pgm_check_handler(
static int stsch_reset(struct subchannel_id schid, volatile struct schib *addr)
{
int rc;
+ register struct subchannel_id reg1 asm ("1") = schid;
pgm_check_occured = 0;
s390_reset_pgm_handler = cio_reset_pgm_check_handler;
- rc = stsch(schid, addr);
+
+ asm volatile(
+ " stsch 0(%2)\n"
+ " ipm %0\n"
+ " srl %0,28"
+ : "=d" (rc)
+ : "d" (reg1), "a" (addr), "m" (*addr) : "memory", "cc");
+
s390_reset_pgm_handler = NULL;
if (pgm_check_occured)
return -EIO;
In-Reply-To: <20061228103925.GB6270@skybase>
On Thu, 28 Dec 2006 11:39:25 +0100, Martin Schwidefsky wrote:
> @@ -881,10 +880,18 @@ static void cio_reset_pgm_check_handler(
> static int stsch_reset(struct subchannel_id schid, volatile struct schib *addr)
> {
> int rc;
> + register struct subchannel_id reg1 asm ("1") = schid;
>
> pgm_check_occured = 0;
> s390_reset_pgm_handler = cio_reset_pgm_check_handler;
> - rc = stsch(schid, addr);
> +
> + asm volatile(
> + " stsch 0(%2)\n"
> + " ipm %0\n"
> + " srl %0,28"
> + : "=d" (rc)
> + : "d" (reg1), "a" (addr), "m" (*addr) : "memory", "cc");
> +
> s390_reset_pgm_handler = NULL;
> if (pgm_check_occured)
> return -EIO;
Can't you just put a barrier() before the stsch() call?
--
MBTI: IXTP
On Sun, 2006-12-31 at 01:31 -0500, Chuck Ebbert wrote:
> > @@ -881,10 +880,18 @@ static void cio_reset_pgm_check_handler(
> > static int stsch_reset(struct subchannel_id schid, volatile struct schib *addr)
> > {
> > int rc;
> > + register struct subchannel_id reg1 asm ("1") = schid;
> >
> > pgm_check_occured = 0;
> > s390_reset_pgm_handler = cio_reset_pgm_check_handler;
> > - rc = stsch(schid, addr);
> > +
> > + asm volatile(
> > + " stsch 0(%2)\n"
> > + " ipm %0\n"
> > + " srl %0,28"
> > + : "=d" (rc)
> > + : "d" (reg1), "a" (addr), "m" (*addr) : "memory", "cc");
> > +
> > s390_reset_pgm_handler = NULL;
> > if (pgm_check_occured)
> > return -EIO;
>
>
> Can't you just put a barrier() before the stsch() call?
You mean after. We have to prevent the sequence
1) load pgm_check_occured,
2) stsch and the execution of cio_reset_pgm_check_handler
3) use of the the value loaded in 1)
A barrier before 2) forces a reload of pgm_check_occured but it does not
force the reload to be before the stsch call.
But the basic idea is valid. The standard stsch() inline followed by a
barrier() is equivalent to the new inline assembly.
--
blue skies,
Martin.
Martin Schwidefsky
Linux for zSeries Development & Services
IBM Deutschland Entwicklung GmbH
"Reality continues to ruin my life." - Calvin.
On Sun, Dec 31, 2006 at 01:31:43AM -0500, Chuck Ebbert wrote:
> In-Reply-To: <20061228103925.GB6270@skybase>
>
> On Thu, 28 Dec 2006 11:39:25 +0100, Martin Schwidefsky wrote:
>
> > @@ -881,10 +880,18 @@ static void cio_reset_pgm_check_handler(
> > static int stsch_reset(struct subchannel_id schid, volatile struct schib *addr)
> > {
> > int rc;
> > + register struct subchannel_id reg1 asm ("1") = schid;
> >
> > pgm_check_occured = 0;
> > s390_reset_pgm_handler = cio_reset_pgm_check_handler;
> > - rc = stsch(schid, addr);
> > +
> > + asm volatile(
> > + " stsch 0(%2)\n"
> > + " ipm %0\n"
> > + " srl %0,28"
> > + : "=d" (rc)
> > + : "d" (reg1), "a" (addr), "m" (*addr) : "memory", "cc");
> > +
> > s390_reset_pgm_handler = NULL;
> > if (pgm_check_occured)
> > return -EIO;
>
>
> Can't you just put a barrier() before the stsch() call?
Yes, that would work too and would look much nicer.
I think we should change the reset program check handler, so that it searches
the exception tables. At least for in-kernel addresses that should work.
For addresses within modules this might cause deadlocks, since the module
code takes spinlocks and we are in a context where we just killed all cpus
not knowing what they executed... Hmm..
[email protected] wrote on 12/31/2006 01:22:03 PM:
> On Sun, Dec 31, 2006 at 01:31:43AM -0500, Chuck Ebbert wrote:
> > In-Reply-To: <20061228103925.GB6270@skybase>
> >
> > On Thu, 28 Dec 2006 11:39:25 +0100, Martin Schwidefsky wrote:
> >
> > > @@ -881,10 +880,18 @@ static void cio_reset_pgm_check_handler(
> > > static int stsch_reset(struct subchannel_id schid, volatile
> struct schib *addr)
> > > {
> > > int rc;
> > > + register struct subchannel_id reg1 asm ("1") = schid;
> > >
> > > pgm_check_occured = 0;
> > > s390_reset_pgm_handler = cio_reset_pgm_check_handler;
> > > - rc = stsch(schid, addr);
> > > +
> > > + asm volatile(
> > > + " stsch 0(%2)\n"
> > > + " ipm %0\n"
> > > + " srl %0,28"
> > > + : "=d" (rc)
> > > + : "d" (reg1), "a" (addr), "m" (*addr) : "memory",
"cc");
> > > +
> > > s390_reset_pgm_handler = NULL;
> > > if (pgm_check_occured)
> > > return -EIO;
> >
> >
> > Can't you just put a barrier() before the stsch() call?
>
> Yes, that would work too and would look much nicer.
>
> I think we should change the reset program check handler, so that it
searches
> the exception tables.
I think that this is really overkill, since having program checks in the
reset case will be VERY rare and stsch will probably the only case.
I would do that only, if we have a component, which needs that.
And I assume, that this will never happen!
Michael