From mboxrd@z Thu Jan 1 00:00:00 1970 From: Arnd Bergmann Subject: Re: [PATCH v9 00/24] ILP32 for ARM64 Date: Sun, 14 Oct 2018 21:49:01 +0200 Message-ID: References: <20180516081910.10067-1-ynorov@caviumnetworks.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Return-path: In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org To: Andy Lutomirski Cc: Yury Norov , Catalin Marinas , Linux ARM , Linux Kernel Mailing List , "open list:DOCUMENTATION" , linux-arch , Linux API , Adam Borowski , Alexander Graf , Alexey Klimov , Andreas Schwab , Andrew Pinski , Bamvor Zhangjian , Chris Metcalf , Christoph Muellner , Dave Martin , David Miller , Florian Weimer List-Id: linux-api@vger.kernel.org On Sat, Oct 13, 2018 at 9:36 PM Andy Lutomirski wrote: > > On Wed, May 16, 2018 at 1:19 AM Yury Norov wrote: > > > > This series enables AARCH64 with ILP32 mode. > > > > As supporting work, it introduces ARCH_32BIT_OFF_T configuration > > option that is enabled for existing 32-bit architectures but disabled > > for new arches (so 64-bit off_t userspace type is used by new userspace). > > Also it deprecates getrlimit and setrlimit syscalls prior to prlimit64. > > Second, ILP32 user code is highly unlikely > to end up with the same struct layout as ILP64 code. The latter seems > like it should be solved entirely in userspace by adding a way to > annotate a structure as being a kernel ABI structure and getting the > toolchain to lay it out as if it were ILP64 even though the target is > ILP32. The syscall ABI could be almost completely abstracted in glibc, the main issue is ioctl and a couple of related interfaces that pass data structures (read() on /dev/input/*, mmap on /dev/snd/* or raw sockets, fcntl). The question whether a data type is laid out like a 64-bit architecture would cannot be a property of the type in most of those cases, because the same types are used elsewhere. Many ioctls just take a pointer to a 'long' or similar, and then you have structures like 'timespec' that are used both in syscall/ioctl ABI and in normal user space code, but are required to be laid out differently there. (timespec is a bad example because y2038 of course, but it illustrates the point). > 2. I think you should make a conscious decision as to whether the > ILP32-ness of a syscall is a property of the task or of the syscall. > On x86, x32-ness is a property of the syscall, but historically it > also got rather entangled with the state of the task, and the result > was a mess. It looks like you're making it be a property of the task, > which is fine, but you're making it impossible for very clever ILP32 > libraries to include little ILP64 stubs that do fancy things with full > 64-bit syscalls. > > 3. Make very certain that you aren't exploitable by malicious > processes that set the high bits in ILP32 syscall args. x86 compat > has issues like that in the past. This point was actually the most important one for keeping the aarch64 ilp32 interface as restricted as it is: it doesn't allow anything that the normal aarch32/armv7 emulation doesn't already provide. Arnd