On Tue, 15 Sep 2020 at 15:45, Horia Geantă <horia.gea...@nxp.com> wrote: > > On 9/15/2020 1:26 PM, Ard Biesheuvel wrote: > > On Tue, 15 Sep 2020 at 13:02, Horia Geantă <horia.gea...@nxp.com> wrote: > >> > >> On 9/14/2020 9:20 PM, Ard Biesheuvel wrote: > >>> On Mon, 14 Sep 2020 at 20:12, Horia Geantă <horia.gea...@nxp.com> wrote: > >>>> > >>>> On 9/14/2020 7:28 PM, Ard Biesheuvel wrote: > >>>>> On Mon, 14 Sep 2020 at 19:24, Horia Geantă <horia.gea...@nxp.com> wrote: > >>>>>> > >>>>>> On 9/9/2020 1:10 AM, Herbert Xu wrote: > >>>>>>> On Tue, Sep 08, 2020 at 01:35:04PM +0300, Horia Geantă wrote: > >>>>>>>> > >>>>>>>>> Just go with the get_unaligned unconditionally. > >>>>>>>> > >>>>>>>> Won't this lead to sub-optimal code for ARMv7 > >>>>>>>> in case the IV is aligned? > >>>>>>> > >>>>>>> If this should be optimised in ARMv7 then that should be done > >>>>>>> in get_unaligned itself and not open-coded. > >>>>>>> > >>>>>> I am not sure what's wrong with avoiding using the unaligned accessors > >>>>>> in case data is aligned. > >>>>>> > >>>>>> Documentation/core-api/unaligned-memory-access.rst clearly states: > >>>>>> These macros work for memory accesses of any length (not just 32 bits > >>>>>> as > >>>>>> in the examples above). Be aware that when compared to standard access > >>>>>> of > >>>>>> aligned memory, using these macros to access unaligned memory can be > >>>>>> costly in > >>>>>> terms of performance. > >>>>>> > >>>>>> So IMO it makes sense to use get_unaligned() only when needed. > >>>>>> There are several cases of users doing this, e.g. siphash. > >>>>>> > >>>>> > >>>>> For ARMv7 code, using the unaligned accessors unconditionally is fine, > >>>>> and it will not affect performance. > >>>>> > >>>>> In general, when CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS is defined, > >>>>> you can use the unaligned accessors. If it is not, it helps to have > >>>>> different code paths. > >>>>> > >>>> arch/arm/include/asm/unaligned.h doesn't make use of > >>>> linux/unaligned/access_ok.h, even if > >>>> CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS > >>>> is set. > >>>> > >>>> I understand the comment in the file, however using get_unaligned() > >>>> unconditionally takes away the opportunity to generate optimized code > >>>> (using ldrd/ldm) when data is aligned. > >>>> > >>> > >>> But the minimal optimization that is possible here (one ldrd/ldm > >>> instruction vs two ldr instructions) is defeated by the fact that you > >>> are using a conditional branch to select between the two. And this is > >>> not even a hot path to begin with, > >>> > >> This is actually on the hot path (encrypt/decrypt callbacks), > >> but you're probably right that the conditional branching is going to offset > >> the optimized code. > >> > > > > This is called once per XTS request, right? And you are saying the > > extra cycle makes a difference? > > > Yes, once per request and no, not super-important. > > >> To avoid branching, code could be rewritten as: > >> > >> #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS > >> size = *(u64 *)(req->iv + (ivsize / 2)); > >> #else > >> size = get_unaligned((u64 *)(req->iv + (ivsize / 2))); > >> #endif > >> > >> however in this case ARMv7 would suffer since > >> CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS=y and > >> ldrd/ldm for accesses not word-aligned are inefficient - lead to traps. > >> > > > > CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS means 'just use the unaligned > > accessors as they are basically free'. Casting a potentially > > misaligned u8* to a u64* is not permitted by the C standard. > > > Seems that I misunderstood CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS. >
You're not the only one :-) I have been intending to get the discussion going with the networking folks, who rely heavily on this as well. > Looking at its usage, e.g. ether_addr_equal() or __crypto_memneq_*(), > I see similar casts of pointers possibly misaligned. > Yes, that is the confusion. CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS should indicate whether using the unaligned accessors is fine in all cases, or whether you should find other ways to load the data more efficiently (compare NET_IP_ALIGN, which shifts the entire IP header so the 32-bit address field appear aligned in memory) CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS does *not* mean that you can simply cast any pointer to any type and dereference it, but the meaning appears to have shifted this way over the years (and the Documentation/ was even updated to this effect) Pre-v6 ARM (and MIPS as well, IIRC) require byte sized accesses and shift/or sequences to do unaligned accesses, whereas v6 and up simply allows ldr from a misaligned address. So in the former case, you could use cra_alignmask to align the data in memory, while the latter case can ignore it. (arch/arm/crypto/aes-cipher-glue.c uses this as well) > >> Would it be ok to use: > >> #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && !defined(CONFIG_ARM) > >> to workaround the ARMv7 inconsistency? > >> > > > > No, please just use the get_unaligned() accessor. > > > Ok. > > Thanks, > Horia
