On Fri, 8 Jan 2021 at 11:59, Arnd Bergmann <arnd@xxxxxxxxxx> wrote: > > On Fri, Jan 8, 2021 at 11:44 AM Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote: > > > > On Fri, Jan 08, 2021 at 11:42:53AM +0100, Arnd Bergmann wrote: > > > > > > How does this work for kernels with CONFIG_VMAP_STACK? > > > I remember some other subsystems (usb, hid) adding workarounds > > > for that, but I don't see those in drivers/crypto > > > > I'm referring to the situation in general and not the subject of > > this thread. > > > > For shash_desc of course it doesn't happen since it sync only. > > Ah, got it. > I can fold in the following --- a/include/linux/crypto.h +++ b/include/linux/crypto.h @@ -151,9 +151,12 @@ * The macro CRYPTO_MINALIGN_ATTR (along with the void * type in the actual * declaration) is used to ensure that the crypto_tfm context structure is * aligned correctly for the given architecture so that there are no alignment - * faults for C data types. In particular, this is required on platforms such - * as arm where pointers are 32-bit aligned but there are data types such as - * u64 which require 64-bit alignment. + * faults for C data types. On architectures that support non-cache coherent + * DMA, such as ARM or arm64, it also takes into account the minimal alignment + * that is required to ensure that the context struct member does not share any + * cachelines with the rest of the struct. This is needed to ensure that cache + * maintenance for non-coherent DMA does not affect data that may be accessed + * by the CPU concurrently. */ #define CRYPTO_MINALIGN ARCH_KMALLOC_MINALIGN (or send it as a separate patch) Note that the comment about 64-bit alignment for 64-bit types on ARM is incorrect.
