Re: next Feb 13 drivers/scsi/aha1542.ko build break

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Fri, 2009-02-13 at 10:22 -0800, Randy Dunlap wrote:
> Sachin P. Sant wrote:
> > Feb 13 next x86 allmodconfig build breaks with
> > 
> > ERROR: "__udivdi3" [drivers/scsi/aha1542.ko] undefined!
> > 
> > .config attached.
> 
> 
> I'm seeing that also.  I think that it's this line in aha1542.c:
> 
> 495:		mbo = (scsi2int(mb[mbi].ccbptr) - (SCSI_BUF_PA(&ccb[0]))) / sizeof(struct ccb);
> 
> since I can comment out the "/ sizeof(struct ccb)" and the driver builds successfully.
> 
> This may need a patch to use one of the div() macros (?).

It shouldn't ... the two inputs should be longs and the sizeof should be
dividing by a constant size_t.  __udivdi3 is used for 64 bit division on
a 32 bit platform, so where's it getting the idea that something in the
division is 64 bit?

The culprit looks to be

commit 976e8f677e42757e5586ea04a9ac8bb8ddaa037e
Author: Jeremy Fitzhardinge <jeremy@goop.org>
Date:   Fri Feb 6 13:29:44 2009 -0800

    x86: asm/io.h: unify virt_to_phys/phys_to_virt

Because it changed the return of virt_to_phys from long to phys_addr_t
which is unsigned long long on a PAE platform, which this is.

So, I could suggest a fix below since isa addresses may never be above
32 bits, but I think this issue has more ramifications for the other
legacy users of virt_to_phys.

James

---

diff --git a/arch/x86/include/asm/io.h b/arch/x86/include/asm/io.h
index e5a2ab4..4f8e820 100644
--- a/arch/x86/include/asm/io.h
+++ b/arch/x86/include/asm/io.h
@@ -125,7 +125,7 @@ static inline void *phys_to_virt(phys_addr_t address)
 /*
  * ISA I/O bus memory addresses are 1:1 with the physical address.
  */
-#define isa_virt_to_bus virt_to_phys
+#define isa_virt_to_bus (unsigned long)virt_to_phys
 #define isa_page_to_bus page_to_phys
 #define isa_bus_to_virt phys_to_virt