On Mon, Feb 21, 2022 at 5:25 AM Thomas Bogendoerfer
[off-list ref] wrote:

With this patch

[ .. snip snip ..]

I at least get my simple test cases fixed, but I'm not sure this is
correct.

I think you really want to do that anyway, just to get things like
wild kernel pointers right (ie think get_kernel_nofault() and friends
for ftrace etc).

They shouldn't happen in any normal situation, but those kinds of
unverified pointers is why we _have_ get_kernel_nofault() in the first
place.

On x86-64, the roughly equivalent situation is that addresses that
aren't in canonical format do not take a #PF (page fault), they take a
#GP (general protection) fault.

So I think you want to do that fixup_exception() for any possible addresses.

Is there a reason to not also #define TASK_SIZE_MAX   __UA_LIMIT like
for the 32bit case ?

I would suggest against using a non-constant TASK_SIZE_MAX. Being
constant is literally one reason why it exists, when TASK_SIZE itself
has often been about other things (ie "32-bit process").

Having to load variables for things like get_user() is annoying, if
you could do it with a simple constant instead (where that "simple"
part is to avoid having to load big values from a constant pool -
often constants like "high bit set" can be loaded and compared against
more efficiently).

               Linus