| Message ID | 20230531154836.1366225-1-catalin.marinas@arm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm, dma, arm64: Reduce ARCH_KMALLOC_MINALIGN to 8 | expand |
On Wed, May 31, 2023 at 8:48 AM Catalin Marinas <catalin.marinas@arm.com> wrote: > Here's version 6 of the series reducing the kmalloc() minimum alignment > on arm64 to 8 (from 128). There are patches already to do the same for > riscv (pretty straight-forward after this series). Thanks, Catalin for getting these patches out. Please add my "Tested-by:" tag for the series: Tested-by: Isaac J. Manjarres <isaacmanjarres@google.com> With the first 11 patches, I observed a reduction of 18.4 MB in the slab memory footprint on my Pixel 6 device. After applying the rest of the patches in the series, I observed a total reduction of 26.5 MB in the slab memory footprint on my device. These are great results! --Isaac
On Thu, 8 Jun 2023 at 07:45, Isaac Manjarres <isaacmanjarres@google.com> wrote: > > On Wed, May 31, 2023 at 8:48 AM Catalin Marinas <catalin.marinas@arm.com> wrote: > > Here's version 6 of the series reducing the kmalloc() minimum alignment > > on arm64 to 8 (from 128). There are patches already to do the same for > > riscv (pretty straight-forward after this series). > Thanks, Catalin for getting these patches out. Please add my "Tested-by:" tag > for the series: > > Tested-by: Isaac J. Manjarres <isaacmanjarres@google.com> > > With the first 11 patches, I observed a reduction of 18.4 MB > in the slab memory footprint on my Pixel 6 device. After applying the > rest of the patches in the series, I observed a total reduction of > 26.5 MB in the > slab memory footprint on my device. These are great results! > It would also be good to get an insight into how much bouncing is going on in this case, given that (AFAIK) Pixel 6 uses non-cache coherent DMA.
On Thu, Jun 08, 2023 at 10:05:58AM +0200, Ard Biesheuvel wrote: > On Thu, 8 Jun 2023 at 07:45, Isaac Manjarres <isaacmanjarres@google.com> wrote: > > > > On Wed, May 31, 2023 at 8:48 AM Catalin Marinas <catalin.marinas@arm.com> wrote: > > > Here's version 6 of the series reducing the kmalloc() minimum alignment > > > on arm64 to 8 (from 128). There are patches already to do the same for > > > riscv (pretty straight-forward after this series). > > Thanks, Catalin for getting these patches out. Please add my "Tested-by:" tag > > for the series: > > > > Tested-by: Isaac J. Manjarres <isaacmanjarres@google.com> > > > > With the first 11 patches, I observed a reduction of 18.4 MB > > in the slab memory footprint on my Pixel 6 device. After applying the > > rest of the patches in the series, I observed a total reduction of > > 26.5 MB in the > > slab memory footprint on my device. These are great results! > > > > It would also be good to get an insight into how much bouncing is > going on in this case, given that (AFAIK) Pixel 6 uses non-cache > coherent DMA. I enabled the "swiotlb_bounced" trace event from the kernel command line to see if anything was being bounced. It turns out that for Pixel 6 there are non-coherent DMA transfers occurring, but none of the transfers that are in either the DMA_FROM_DEVICE or DMA_BIDIRECTIONAL directions are small enough to require bouncing. --Isaac P.S. I noticed that the trace_swiotlb_bounced() tracepoint may not be invoked even though bouncing occurs. For example, in the dma-iommu path, swiotlb_tbl_map_single() is called when bouncing, instead of swiotlb_map(), which is what ends up calling trace_swiotlb_bounced(). Would it make sense to move the call to trace_swiotlb_bounced() to swiotlb_tbl_map_single() since that function is always invoked?
On Thu, 8 Jun 2023 14:29:45 -0700 Isaac Manjarres <isaacmanjarres@google.com> wrote: >[...] > P.S. I noticed that the trace_swiotlb_bounced() tracepoint may not be > invoked even though bouncing occurs. For example, in the dma-iommu path, > swiotlb_tbl_map_single() is called when bouncing, instead of > swiotlb_map(), which is what ends up calling trace_swiotlb_bounced(). > > Would it make sense to move the call to trace_swiotlb_bounced() to > swiotlb_tbl_map_single() since that function is always invoked? Definitely, if you ask me. I believe the change was merely forgotten in commit eb605a5754d0 ("swiotlb: add swiotlb_tbl_map_single library function"). Let me take the author into Cc. Plus Konrad, who built further on that commit, may also have an opinion. Petr T
On Mon, Jun 12, 2023 at 09:47:55AM +0200, Christoph Hellwig wrote: > On Mon, Jun 12, 2023 at 07:44:46AM +0000, Tomonori Fujita wrote: > > I cannot recall the patch but from quick look, moving trace_swiotlb_bounced() to > > swiotlb_tbl_map_single() makes sense. > > Agreed. There's actually two call-sites for trace_swiotlb_bounced(): swiotlb_map() and xen_swiotlb_map_page(). Both those functions also invoke swiotlb_tbl_map_single(), so moving the call to trace_swiotlb_bounced() to swiotlb_tbl_map_single() means that there will be 2 traces per bounce buffering event. The difference between the two call-sites of trace_swiotlb_bounced() is that the call in swiotlb_map() uses phys_to_dma() for the device address, while xen_swiotlb_map_page() uses xen_phys_to_dma(). Would it make sense to move the trace_swiotlb_bounced() call to swiotlb_tbl_map_single() and then introduce a swiotlb_tbl_map_single_notrace() function which doesn't do the tracing, and xen_swiotlb_map_page() can call this? --Isaac
Hi, Here's version 6 of the series reducing the kmalloc() minimum alignment on arm64 to 8 (from 128). There are patches already to do the same for riscv (pretty straight-forward after this series). The first 11 patches decouple ARCH_KMALLOC_MINALIGN from ARCH_DMA_MINALIGN and, for arm64, limit the kmalloc() caches to those aligned to the run-time probed cache_line_size(). On arm64 we gain the kmalloc-{64,192} caches. The subsequent patches (11 to 17) further reduce the kmalloc() caches to kmalloc-{8,16,32,96} if the default swiotlb is present by bouncing small buffers in the DMA API. Changes since v5: - Renaming of the sg_* accessors for consistency. - IIO_DMA_MINALIGN defined to ARCH_DMA_MINALIGN (missed it in previous versions). - Modified Robin's patch 11 to use #ifdef CONFIG_NEED_SG_DMA_FLAGS instead of CONFIG_PCI_P2PDMA in scatterlist.h. - Added the new sg_dma_*_swiotlb() under the same #ifdef as above. The updated patches are also available on this branch: git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux devel/kmalloc-minalign Thanks. Catalin Marinas (15): mm/slab: Decouple ARCH_KMALLOC_MINALIGN from ARCH_DMA_MINALIGN dma: Allow dma_get_cache_alignment() to be overridden by the arch code mm/slab: Simplify create_kmalloc_cache() args and make it static mm/slab: Limit kmalloc() minimum alignment to dma_get_cache_alignment() drivers/base: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN drivers/gpu: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN drivers/usb: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN drivers/spi: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN dm-crypt: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN iio: core: Use ARCH_DMA_MINALIGN instead of ARCH_KMALLOC_MINALIGN arm64: Allow kmalloc() caches aligned to the smaller cache_line_size() dma-mapping: Force bouncing if the kmalloc() size is not cache-line-aligned iommu/dma: Force bouncing if the size is not cacheline-aligned mm: slab: Reduce the kmalloc() minimum alignment if DMA bouncing possible arm64: Enable ARCH_WANT_KMALLOC_DMA_BOUNCE for arm64 Robin Murphy (2): scatterlist: Add dedicated config for DMA flags dma-mapping: Name SG DMA flag helpers consistently arch/arm64/Kconfig | 1 + arch/arm64/include/asm/cache.h | 3 ++ arch/arm64/mm/init.c | 7 +++- drivers/base/devres.c | 6 ++-- drivers/gpu/drm/drm_managed.c | 6 ++-- drivers/iommu/Kconfig | 1 + drivers/iommu/dma-iommu.c | 58 ++++++++++++++++++++++++-------- drivers/iommu/iommu.c | 2 +- drivers/md/dm-crypt.c | 2 +- drivers/pci/Kconfig | 1 + drivers/spi/spidev.c | 2 +- drivers/usb/core/buffer.c | 8 ++--- include/linux/dma-map-ops.h | 61 ++++++++++++++++++++++++++++++++++ include/linux/dma-mapping.h | 4 ++- include/linux/iio/iio.h | 2 +- include/linux/scatterlist.h | 60 ++++++++++++++++++++++++++------- include/linux/slab.h | 14 ++++++-- kernel/dma/Kconfig | 7 ++++ kernel/dma/direct.c | 2 +- kernel/dma/direct.h | 3 +- mm/slab.c | 6 +--- mm/slab.h | 5 ++- mm/slab_common.c | 46 +++++++++++++++++++------ 23 files changed, 243 insertions(+), 64 deletions(-)