On Wednesday 19 September 2007, Stuart Anderson wrote: > On Wed, 19 Sep 2007, J. Mayer wrote: > > Then, the changes you've done, changing long arguments (which should be > > target_long to be correct, you can take a look at the last patch I sent > > on the list) to pointers, for example in function prototypes, are > > incorrect. > > I just went, and looked at the linux code again for 32 on 64 for x86_64 and > powerpc. In both of these cases (and I suspect the others as well), the > parameters which are passed via registers are 0 extended from 32 bits to > 64 bit in the syscall entry asm code. This way, once the C code is > called via the sys_call_table, everything is dealt with as 64 bits. This > actually keeps the rest of the code simpler as the rest of the kernel > doesn't have to be extending & truncating pointers everywhere else. > > On x86_64 and powerpc, it appears that both user (ie target) and kernel > pointers co-exist and that the code that maps structures assume that the > __get_user()/__put_user() and copy_*_user() routines can handle any > special situation. The pointers passed into functions like > cp_compat_stat() are 64-bits for both the structure located in the > kernel, and the one located in user space. > > My understanding is that we want to do as the kernel does as much as > possible. In light of this, wouldn't we want to be decreasing the use > of target_long where pointers may be involved instead of increasing it?
No. We're doing more than most 32-64 syscall thunks. To a first approximation the syscall thunks can bindly zero extend all values. In qemu we need to know whether something is a pointer or a value. Kernel and userspace addresses are not interchangeable in the kernel. Any place that does so is probably a bug. Paul
