On Tue, Feb 15, 2022 at 8:13 AM Al Viro [off-list ref] wrote:

On Tue, Feb 15, 2022 at 07:29:42AM +0100, Christoph Hellwig wrote:

quoted

On Tue, Feb 15, 2022 at 12:37:41AM +0000, Al Viro wrote:

quoted

Perhaps simply wrap that sucker into #ifdef CONFIG_CPU_HAS_ADDRESS_SPACES
(and trim the comment down to "coldfire and 68000 will pick generic
variant")?

I wonder if we should invert CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE,
select the separate address space config for s390, sparc64, non-coldfire
m68k and mips with EVA and then just have one single access_ok for
overlapping address space (as added by Arnd) and non-overlapping ones
(always return true).

parisc is also such...  How about

        select ALTERNATE_SPACE_USERLAND

for that bunch?

Either of those works for me. My current version has this keyed off
TASK_SIZE_MAX==ULONG_MAX, but a CONFIG_ symbol does
look more descriptive.

 While we are at it, how many unusual access_ok() instances are
left after this series?  arm64, itanic, um, anything else?

x86 adds a WARN_ON_IN_IRQ() check in there. This could be
made generic, but it's not obvious what exactly the exceptions are
that other architectures need. The arm64 tagged pointers could
probably also get integrated into the generic version.

FWIW, sparc32 has a slightly unusual instance (see uaccess_32.h there); it's
obviously cheaper than generic and I wonder if the trick is legitimate (and
applicable elsewhere, perhaps)...

Right, a few others have the same, but I wasn't convinced that this
is actually safe for call possible cases: it's trivial to construct a caller
that works on other architectures but not this one, if you pass a large
enough size value and don't access the contents in sequence.

Also, like the ((addr | (addr + size)) & MASK) check on some other
architectures, it is less portable because it makes assumptions about
the actual layout beyond a fixed address limit.

        Arnd