On Mon, Feb 17, 2014 at 3:46 PM, Russell King <[email protected]> wrote:
>
> One fix touches code outside of arch/arm, which is related to sorting
> out the DMA masks correctly. There is a long standing issue with the
> conversion from PFNs to addresses where people assume that shifting an
> unsigned long left by PAGE_SHIFT results in a correct address.
You should probably have used PFN_PHYS(), which does this correctly.
Your explicit u64 isn't exactly wrong, but phys_addr_t is really the
right type for the result.
That said, it's admittedly a disgusting name, and I wonder if we
should introduce a nicer-named "pfn_to_phys()" that matches the other
"xyz_to_abc()" functions we have (including "pfn_to_virt()")
Looking at it, the Xen people then do this disgusting thing:
"__va(PFN_PHYS(pfn))" which is both ugly and pointless (__va() isn't
going to work for a phys_addr_t anyway). And <linux/mm.h> has this
gem:
__va(PFN_PHYS(page_to_pfn(page)));
Ugh. The ugly - it burns. that really should be
"pfn_to_virt(page_to_pfn())", I think. Adding a few mailing lists in
the hope that some sucker^Whumanitarian person would want to take a
look.
Anyway, I pulled your change to scsi_lib.c, since it's certainly no
worse than what we used to have, but James and company cc'd too.
Linus
On Tue, Feb 18, 2014 at 03:49:03PM -0800, Linus Torvalds wrote:
> On Mon, Feb 17, 2014 at 3:46 PM, Russell King <[email protected]> wrote:
> >
> > One fix touches code outside of arch/arm, which is related to sorting
> > out the DMA masks correctly. There is a long standing issue with the
> > conversion from PFNs to addresses where people assume that shifting an
> > unsigned long left by PAGE_SHIFT results in a correct address.
>
> You should probably have used PFN_PHYS(), which does this correctly.
> Your explicit u64 isn't exactly wrong, but phys_addr_t is really the
> right type for the result.
Almost, but not quite. If we're going to avoid u64, then dma_addr_t
woudl be the right type here because we're talking about DMA addresses.
We could also switch to keeping this as PFNs - block internally converts
it to a PFN anyway:
void blk_queue_bounce_limit(struct request_queue *q, u64 max_addr)
{
unsigned long b_pfn = max_addr >> PAGE_SHIFT;
...
and that is ultimately assigned to q->limits.bounce_pfn. So, if we
arranged for blk_queue_bounce_limit() to take a PFN, and then modified
it's two callers, then we don't need to care about converting between
phys and pfns.
Maybe blk_queue_bounce_pfn_limit() so we ensure all users get caught?
> That said, it's admittedly a disgusting name, and I wonder if we
> should introduce a nicer-named "pfn_to_phys()" that matches the other
> "xyz_to_abc()" functions we have (including "pfn_to_virt()")
We have these on ARM:
arch/arm/include/asm/memory.h:#define __pfn_to_phys(pfn) ((phys_addr_t)(pfn) << PAGE_SHIFT)
arch/arm/include/asm/memory.h:#define __phys_to_pfn(paddr) ((unsigned long)((paddr) >> PAGE_SHIFT))
it probably makes sense to pick those right out, maybe losing the
__ prefix on them.
> Looking at it, the Xen people then do this disgusting thing:
> "__va(PFN_PHYS(pfn))" which is both ugly and pointless (__va() isn't
> going to work for a phys_addr_t anyway). And <linux/mm.h> has this
> gem:
>
> __va(PFN_PHYS(page_to_pfn(page)));
Wow. Two things spring to mind there... highmem pages, and don't we
already have page_address() for that?
> Anyway, I pulled your change to scsi_lib.c, since it's certainly no
> worse than what we used to have, but James and company cc'd too.
Thanks. I do worry about all the other places which I also found -
but the first step is getting concensus on what the macro should be.
--
FTTC broadband for 0.8mile line: 5.8Mbps down 500kbps up. Estimation
in database were 13.1 to 19Mbit for a good line, about 7.5+ for a bad.
Estimate before purchase was "up to 13.2Mbit".
On Tue, Feb 18, 2014 at 4:03 PM, Russell King - ARM Linux
<[email protected]> wrote:
>
> Almost, but not quite. If we're going to avoid u64, then dma_addr_t
> woudl be the right type here because we're talking about DMA addresses.
Well, phys_addr_t had better be as big as dma_addr_t, because that's
what the resource management handles.
> We could also switch to keeping this as PFNs - block internally converts
> it to a PFN anyway:
Yeah, that definitely sounds like it would be a good idea.
> Maybe blk_queue_bounce_pfn_limit() so we ensure all users get caught?
>
>> That said, it's admittedly a disgusting name, and I wonder if we
>> should introduce a nicer-named "pfn_to_phys()" that matches the other
>> "xyz_to_abc()" functions we have (including "pfn_to_virt()")
>
> We have these on ARM:
>
> arch/arm/include/asm/memory.h:#define __pfn_to_phys(pfn) ((phys_addr_t)(pfn) << PAGE_SHIFT)
> arch/arm/include/asm/memory.h:#define __phys_to_pfn(paddr) ((unsigned long)((paddr) >> PAGE_SHIFT))
>
> it probably makes sense to pick those right out, maybe losing the
> __ prefix on them.
Yup.
>> __va(PFN_PHYS(page_to_pfn(page)));
>
> Wow. Two things spring to mind there... highmem pages, and don't we
> already have page_address() for that?
Well, that code clearly cannot handle highmem anyway, but yes, it
really smells like xen should use page_address().
Adding Xen people who I didn't add the last time around.
Linus