lkml.org 
[lkml]   [2018]   [Sep]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/5] lib/ioremap: Ensure phys_addr actually corresponds to a physical address
Hi Sean,

Thanks for looking at the patch.

On Wed, Sep 12, 2018 at 08:09:39AM -0700, Sean Christopherson wrote:
> On Wed, Sep 12, 2018 at 11:26:13AM +0100, Will Deacon wrote:
> > The current ioremap() code uses a phys_addr variable at each level of
> > page table, which is confusingly offset by subtracting the base virtual
> > address being mapped so that adding the current virtual address back on
> > when iterating through the page table entries gives back the corresponding
> > physical address.
> >
> > This is fairly confusing and results in all users of phys_addr having to
> > add the current virtual address back on. Instead, this patch just updates
> > phys_addr when iterating over the page table entries, ensuring that it's
> > always up-to-date and doesn't require explicit offsetting.
> >
> > Cc: Chintan Pandya <cpandya@codeaurora.org>
> > Cc: Toshi Kani <toshi.kani@hpe.com>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > Cc: Michal Hocko <mhocko@suse.com>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Signed-off-by: Will Deacon <will.deacon@arm.com>
> > ---
> > lib/ioremap.c | 28 ++++++++++++----------------
> > 1 file changed, 12 insertions(+), 16 deletions(-)
> >
> > diff --git a/lib/ioremap.c b/lib/ioremap.c
> > index 6c72764af19c..fc834a59c90c 100644
> > --- a/lib/ioremap.c
> > +++ b/lib/ioremap.c
> > @@ -101,19 +101,18 @@ static inline int ioremap_pmd_range(pud_t *pud, unsigned long addr,
> > pmd_t *pmd;
> > unsigned long next;
> >
> > - phys_addr -= addr;
> > pmd = pmd_alloc(&init_mm, pud, addr);
> > if (!pmd)
> > return -ENOMEM;
> > do {
> > next = pmd_addr_end(addr, end);
> >
> > - if (ioremap_try_huge_pmd(pmd, addr, next, phys_addr + addr, prot))
> > + if (ioremap_try_huge_pmd(pmd, addr, next, phys_addr, prot))
> > continue;
> >
> > - if (ioremap_pte_range(pmd, addr, next, phys_addr + addr, prot))
> > + if (ioremap_pte_range(pmd, addr, next, phys_addr, prot))
> > return -ENOMEM;
> > - } while (pmd++, addr = next, addr != end);
> > + } while (pmd++, addr = next, phys_addr += PMD_SIZE, addr != end);
>
> I think bumping phys_addr by PXX_SIZE is wrong if phys_addr and addr
> start unaligned with respect to PXX_SIZE. The addresses must be
> PAGE_ALIGNED, which lets ioremap_pte_range() do a simple calculation,
> but that doesn't hold true for the upper levels, i.e. phys_addr needs
> to be adjusted using an algorithm similar to pxx_addr_end().
>
> Using a 2mb page as an example (lower 32 bits only):
>
> pxx_size = 0x00020000
> pxx_mask = 0xfffe0000
> addr = 0x1000
> end = 0x00040000
> phys_addr = 0x1000
>
> Loop 1:
> addr = 0x1000
> phys = 0x1000
>
> Loop 2:
> addr = 0x20000
> phys = 0x21000

Yes, I think you're completely right, however I also don't think this
can happen with the current code (and I've failed to trigger it in my
testing). The virtual addresses allocated for VM_IOREMAP allocations
are aligned to the order of the allocation, which means that the virtual
address at the start of the mapping is aligned such that when we hit the
end of a pXd, we know we've mapped the previous PXD_SIZE bytes.

Having said that, this is clearly a change from the current code and I
haven't audited architectures other than arm64 (where IOREMAP_MAX_ORDER
corresponds to the maximum size of our huge mappings), so it would be
much better not to introduce this funny behaviour in a patch that aims
to reduce confusion in the first place!

Fixing this using the pxx_addr_end() macros is a bit strange, since we
don't have a physical end variable (nor do we need one), so perhaps
something like changing the while condition to be:

do {
...
} while (pmd++, phys_addr += (next - addr), addr = next, addr != end);

would do the trick. What do you reckon?

Will

\
 
 \ /
  Last update: 2018-09-12 18:39    [W:0.172 / U:0.232 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site