On 2/24/2013 7:29 PM, Al Viro wrote: > Unless I'm seriously misreading your code, you have all arguments of > compat syscall sign-extended by asm glue. If that's the case, consider > what will happen to 32bit binary doing > llseek(fd, 0, 0xffffffff, &pos, SEEK_CUR) > on 32bit and 64bit hosts resp. The former will move the current position > by 4Gb forward; the latter - by one byte backwards... Without testing this (I'm in an airport), I think you might be right. But the fix seems like it might just be changing the parameters of the llseek syscall in read_write.c to be "unsigned int" for offset_high and offset_low. That would let the usual syscall-wrappers code properly zero-extend the high bits. > sys_llseek() will do the right thing for compat on architectures that > zero-extend the arguments of compat syscall; x86/sparc/ppc/arm64 fall into > that category. mips and (AFAICS) tile do not (and s390 is just plain weird > and needs wrappers for just about everything). mips has sys_32_llseek() for > a good reason; we probably want to take that into fs/read_write.c, but > on such architectures something *is* needed - plain sys_llseek() won't work... > > FWIW, I suspect that we want __ARCH_COMPAT_ZERO_EXTENDS or > __ARCH_COMPAT_SIGN_EXTENDS defined in asm/compat.h; the need of sys_llseek() > would be "it's 32bit or it has compat and it's zero-extending", while > "has compat and it's not zero-extending" would pick compat_sys_llseek() > instead. __ARCH_WANT_SYS_LLSEEK shouldn't exist... > > And things like "do we need compat_sys_lseek(), compat_sys_truncate() > and their ilk" could be expressed via those as well - compat && !sign_extends. > > Comments? Cleanups like these sound generally good. For tilegx32, for good or bad, we emulated the register allocation conventions of our 32-bit tilepro architecture, so that for example truncate64 has r0 = filename, r1 = dummy, r2+r3 = offset. This doesn't match what the compiler would generate for a simple "char *, long long" argument pair, since for tilegx32 we (obviously) represent long long as a 64-bit register value in general. But for new architectures it might be good to look to make changes like those you are suggesting. -- Chris Metcalf, Tilera Corp. http://www.tilera.com -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
