GitLab

Jisheng Zhang <jszhang@kernel.org> says:

Currently, riscv defines ARCH_DMA_MINALIGN as L1_CACHE_BYTES, I.E
64Bytes, if CONFIG_RISCV_DMA_NONCOHERENT=y. To support unified kernel
Image, usually we have to enable CONFIG_RISCV_DMA_NONCOHERENT, thus
it brings some bad effects to coherent platforms:

Firstly, it wastes memory, kmalloc-96, kmalloc-32, kmalloc-16 and
kmalloc-8 slab caches don't exist any more, they are replaced with
either kmalloc-128 or kmalloc-64.

Secondly, larger than necessary kmalloc aligned allocations results
in unnecessary cache/TLB pressure.

This issue also exists on arm64 platforms. From last year, Catalin
tried to solve this issue by decoupling ARCH_KMALLOC_MINALIGN from
ARCH_DMA_MINALIGN, limiting kmalloc() minimum alignment to
dma_get_cache_alignment() and replacing ARCH_KMALLOC_MINALIGN usage
in various drivers with ARCH_DMA_MINALIGN etc.[1]

One fact we can make use of for riscv: if the CPU doesn't support
ZICBOM or T-HEAD CMO, we know the platform is coherent. Based on
Catalin's work and above fact, we can easily solve the kmalloc align
issue for riscv: we can override dma_get_cache_alignment(), then let
it return ARCH_DMA_MINALIGN at the beginning and return 1 once we know
the underlying HW neither supports ZICBOM nor supports T-HEAD CMO.

So what about if the CPU supports ZICBOM or T-HEAD CMO, but all the
devices are dma coherent? Well, we use ARCH_DMA_MINALIGN as the
kmalloc minimum alignment, nothing changed in this case. This case
can be improved in the future once we see such platforms in mainline.

After this patch, a simple test of booting to a small buildroot rootfs
on qemu shows:

kmalloc-96           5041    5041     96  ...
kmalloc-64           9606    9606     64  ...
kmalloc-32           5128    5128     32  ...
kmalloc-16           7682    7682     16  ...
kmalloc-8           10246   10246      8  ...

So we save about 1268KB memory. The saving will be much larger in normal
OS env on real HW platforms.

patch1 allows kmalloc() caches aligned to the smallest value.
patch2 enables DMA_BOUNCE_UNALIGNED_KMALLOC.

After this series:

As for coherent platforms, kmalloc-{8,16,32,96} caches come back on
coherent both RV32 and RV64 platforms, I.E !ZICBOM and !THEAD_CMO.

As for noncoherent RV32 platforms, nothing changed.

As for noncoherent RV64 platforms, I.E either ZICBOM or THEAD_CMO, the
above kmalloc caches also come back if > 4GB memory or users pass
"swiotlb=mmnn,force" to force swiotlb creation if <= 4GB memory. How
much mmnn should be depends on the specific platform, it needs to be
tried and tested all possible usage case on the specific hardware. For
example, I can use the minimal I/O TLB slabs on Sipeed M1S Dock.

* b4-shazam-merge:
  riscv: enable DMA_BOUNCE_UNALIGNED_KMALLOC for !dma_coherent
  riscv: allow kmalloc() caches aligned to the smallest value

Link: https://lore.kernel.org/linux-arm-kernel/20230524171904.3967031-1-catalin.marinas@arm.com/ [1]
Link: https://lore.kernel.org/r/20230718152214.2907-1-jszhang@kernel.org

Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>