2017-07-24 17:29:44

by Robin Murphy

Subject: [PATCH] dma-mapping: Reduce dma_mapping_error() inline bloat

Thanks to the nested inlining, all drivers correctly calling
dma_mapping_error() after a mapping a page or single buffer generate two
calls to get_arch_dma_ops() per callsite, which all adds up to a fair
old chunk of useless code, e.g. ~3KB for an arm64 defconfig plus extras:

text data bss dec hex filename
13051391 1503898 327768 14883057 e318f1 vmlinux.o.old
13050751 1503898 327768 14882417 e31671 vmlinux.o.new

Give the compiler a hand by making it clear we want the same ops.

Signed-off-by: Robin Murphy <[email protected]>
---
include/linux/dma-mapping.h | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
index 843ab866e0f4..239e53d12ee8 100644
--- a/include/linux/dma-mapping.h
+++ b/include/linux/dma-mapping.h
@@ -541,10 +541,11 @@ static inline void dma_free_noncoherent(struct device *dev, size_t size,

static inline int dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
{
- debug_dma_mapping_error(dev, dma_addr);
+ const struct dma_map_ops *ops = get_dma_ops(dev);

- if (get_dma_ops(dev)->mapping_error)
- return get_dma_ops(dev)->mapping_error(dev, dma_addr);
+ debug_dma_mapping_error(dev, dma_addr);
+ if (ops->mapping_error)
+ return ops->mapping_error(dev, dma_addr);
return 0;
}

--
2.12.2.dirty

2017-08-01 07:58:14

by Marek Szyprowski

Subject: Re: [PATCH] dma-mapping: Reduce dma_mapping_error() inline bloat

Hi Robin,

On 2017-07-24 19:29, Robin Murphy wrote:
> Thanks to the nested inlining, all drivers correctly calling
> dma_mapping_error() after a mapping a page or single buffer generate two
> calls to get_arch_dma_ops() per callsite, which all adds up to a fair
> old chunk of useless code, e.g. ~3KB for an arm64 defconfig plus extras:
>
> text data bss dec hex filename
> 13051391 1503898 327768 14883057 e318f1 vmlinux.o.old
> 13050751 1503898 327768 14882417 e31671 vmlinux.o.new
>
> Give the compiler a hand by making it clear we want the same ops.

Reviewed-by: Marek Szyprowski <[email protected]>

Similar pattern is used in drivers/xen/swiotlb-xen.c for mmap and
get_sgtable. This could be also fixed, although those are not used so
frequently as dma_mapping_error.

> Signed-off-by: Robin Murphy <[email protected]>
> ---
> include/linux/dma-mapping.h | 7 ++++---
> 1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
> index 843ab866e0f4..239e53d12ee8 100644
> --- a/include/linux/dma-mapping.h
> +++ b/include/linux/dma-mapping.h
> @@ -541,10 +541,11 @@ static inline void dma_free_noncoherent(struct device *dev, size_t size,
>
> static inline int dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
> {
> - debug_dma_mapping_error(dev, dma_addr);
> + const struct dma_map_ops *ops = get_dma_ops(dev);
>
> - if (get_dma_ops(dev)->mapping_error)
> - return get_dma_ops(dev)->mapping_error(dev, dma_addr);
> + debug_dma_mapping_error(dev, dma_addr);
> + if (ops->mapping_error)
> + return ops->mapping_error(dev, dma_addr);
> return 0;
> }
>

Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland

2017-08-25 14:56:34

by Christoph Hellwig

Subject: Re: [PATCH] dma-mapping: Reduce dma_mapping_error() inline bloat

Thanks,

applied to the dma-mapping-4.14 tree.