From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mathieu Desnoyers Subject: Re: [RFC PATCH for 4.18] rseq: use __u64 for rseq_cs fields, validate user inputs Date: Mon, 2 Jul 2018 22:01:31 -0400 (EDT) Message-ID: <858886246.10882.1530583291379.JavaMail.zimbra@efficios.com> References: <20180702223143.4663-1-mathieu.desnoyers@efficios.com> <415287289.10831.1530572418907.JavaMail.zimbra@efficios.com> <825871008.10839.1530573419561.JavaMail.zimbra@efficios.com> <1959930320.10843.1530573742647.JavaMail.zimbra@efficios.com> <8B2E4CEB-3080-4602-8B62-774E400892EB@amacapital.net> <459661281.10865.1530580742205.JavaMail.zimbra@efficios.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: <459661281.10865.1530580742205.JavaMail.zimbra@efficios.com> Sender: linux-kernel-owner@vger.kernel.org To: Andy Lutomirski Cc: Linus Torvalds , Thomas Gleixner , linux-kernel , linux-api , Peter Zijlstra , "Paul E. McKenney" , Boqun Feng , Dave Watson , Paul Turner , Andrew Morton , Russell King , Ingo Molnar , "H. Peter Anvin" , Andi Kleen , Chris Lameter , Ben Maurer , rostedt , Josh Triplett , Catalin Marinas , Will Deacon , Michael List-Id: linux-api@vger.kernel.org ----- On Jul 2, 2018, at 9:19 PM, Mathieu Desnoyers mathieu.desnoyers@effic= ios.com wrote: > ----- On Jul 2, 2018, at 7:37 PM, Andy Lutomirski luto@amacapital.net wro= te: >=20 >>> On Jul 2, 2018, at 4:22 PM, Mathieu Desnoyers >>> wrote: >>>=20 >>> ----- On Jul 2, 2018, at 7:16 PM, Mathieu Desnoyers >>> mathieu.desnoyers@efficios.com wrote: >>>=20 >>>> ----- On Jul 2, 2018, at 7:06 PM, Linus Torvalds torvalds@linux-founda= tion.org >>>> wrote: >>>>=20 >>>>> On Mon, Jul 2, 2018 at 4:00 PM Mathieu Desnoyers >>>>> wrote: >>>>>>=20 >>>>>> Unfortunately, that rseq->rseq_cs field needs to be updated by user-= space >>>>>> with single-copy atomicity. Therefore, we want 32-bit user-space to = initialize >>>>>> the padding with 0, and only update the low bits with single-copy at= omicity. >>>>>=20 >>>>> Well... It's actually still single-copy atomicity as a 64-bit value. >>>>>=20 >>>>> Why? Because it doesn't matter how you write the upper bits. You'll b= e >>>>> writing the same value to them (zero) anyway. >>>>>=20 >>>>> So who cares if the write ends up being two instructions, because the >>>>> write to the upper bits doesn't actually *do* anything. >>>>>=20 >>>>> Hmm? >>>>=20 >>>> Are there any kind of guarantees that a __u64 update on a 32-bit archi= tecture >>>> won't be torn into something daft like byte-per-byte stores when perfo= rmed >>>> from C code ? >>>>=20 >>>> I don't worry whether the upper bits get updated or how, but I really = care >>>> about not having store tearing of the low bits update. >>>=20 >>> For the records, most updates of those low bits are done in assembly >>> from critical sections, for which we control exactly how the update is >>> performed. >>>=20 >>> However, there is one helper function in user-space that updates that v= alue >>> from C through a volatile store, e.g.: >>>=20 >>> static inline void rseq_prepare_unload(void) >>> { >>> __rseq_abi.rseq_cs =3D 0; >>> } >>=20 >> How about making the field be: >>=20 >> union { >> __u64 rseq_cs; >> struct { >> __u32 rseq_cs_low; >> __u32 rseq_cs_high; >> }; >> }; >>=20 >> 32-bit user code that cares about performance can just write to rseq_cs_= low >> because it already knows that rseq_cs_high =3D=3D 0. >>=20 >> The header could even supply a static inline helper write_rseq_cs() that >> atomically writes a pointer and just does the right thing for 64-bit, fo= r >> 32-bit BE, and for 32-bit LE. >>=20 >> I think the union really is needed because we can=E2=80=99t rely on user= code being >> built with -fno-strict-aliasing. Or the helper could use inline asm. >>=20 >> Anyway, the point is that we get optimal code generation (a single instr= uction >> write of the correct number of bits) without any compat magic in the ker= nel. >=20 > That works for me! Any objection from anyone else for this approach ? One thing to consider is how we will implement the load of that pointer on the kernel side. Strictly-speaking, the rseq uapi talks about single-cop= y atomicity, and does not specify _which_ thread is expected to update that pointer. So arguably, the common case is that the current thread is updatin= g it, which would allow the kernel to read it piece-wise. However, nothing prevents user-space from updating it from another thread with single-copy atomicity. So in order to be on the safe side, I prefer to guarantee single-copy atomicity of the get_user() load from the kernel that reads this pointer. This means a 32-bit kernel would have to perform two independent loads: one for low bits, one for high bits. So it does look like we need some __LP64__ ifdefery even with the union trick. Therefore, I'm not convinced the union is useful at all. Thoughts ? Thanks, Mathieu --=20 Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com