| Message ID | 20231006153252.3162299-1-catalin.marinas@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | arm64: swiotlb: Reduce the default size if no ZONE_DMA bouncing needed | expand |
On 2023-10-06 16:32, Catalin Marinas wrote: > With CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC enabled, the arm64 kernel still > allocates the default SWIOTLB buffer (64MB) even if ZONE_DMA is disabled > or all the RAM fits into this zone. However, this potentially wastes a > non-negligible amount of memory on platforms with little RAM. > > Reduce the SWIOTLB size to 1MB per 1GB of RAM if only needed for > kmalloc() buffer bouncing. > > Signed-off-by: Catalin Marinas <catalin.marinas@arm.com> > Suggested-by: Ross Burton <ross.burton@arm.com> > Cc: Ross Burton <ross.burton@arm.com> > Cc: Robin Murphy <robin.murphy@arm.com> > Cc: Will Deacon <will@kernel.org> > --- > arch/arm64/mm/init.c | 10 +++++++++- > 1 file changed, 9 insertions(+), 1 deletion(-) > > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c > index 8a0f8604348b..54ee1a4868c2 100644 > --- a/arch/arm64/mm/init.c > +++ b/arch/arm64/mm/init.c > @@ -493,8 +493,16 @@ void __init mem_init(void) > { > bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit); > > - if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC)) > + if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) { > + /* > + * If no bouncing needed for ZONE_DMA, reduce the swiotlb > + * buffer for kmalloc() bouncing to 1MB per 1GB of RAM. > + */ > + unsigned long size = > + ALIGN(memblock_phys_mem_size(), SZ_1G) >> 10; Hmm, I wondered if DIV_ROUND_UP(memblock_phys_mem_size(), 1024) might be any easier, but by the time I've typed it out it's still just as long :) Either way then, Reviewed-by Robin Murphy <robin.murphy@arm.com> > + swiotlb_adjust_size(min(swiotlb_size_or_default(), size)); > swiotlb = true; > + } > > swiotlb_init(swiotlb, SWIOTLB_VERBOSE); >
On Fri, Oct 06, 2023 at 05:25:03PM +0100, Robin Murphy wrote: > On 2023-10-06 16:32, Catalin Marinas wrote: > > With CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC enabled, the arm64 kernel still > > allocates the default SWIOTLB buffer (64MB) even if ZONE_DMA is disabled > > or all the RAM fits into this zone. However, this potentially wastes a > > non-negligible amount of memory on platforms with little RAM. > > > > Reduce the SWIOTLB size to 1MB per 1GB of RAM if only needed for > > kmalloc() buffer bouncing. > > > > Signed-off-by: Catalin Marinas <catalin.marinas@arm.com> > > Suggested-by: Ross Burton <ross.burton@arm.com> > > Cc: Ross Burton <ross.burton@arm.com> > > Cc: Robin Murphy <robin.murphy@arm.com> > > Cc: Will Deacon <will@kernel.org> > > --- > > arch/arm64/mm/init.c | 10 +++++++++- > > 1 file changed, 9 insertions(+), 1 deletion(-) > > > > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c > > index 8a0f8604348b..54ee1a4868c2 100644 > > --- a/arch/arm64/mm/init.c > > +++ b/arch/arm64/mm/init.c > > @@ -493,8 +493,16 @@ void __init mem_init(void) > > { > > bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit); > > - if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC)) > > + if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) { > > + /* > > + * If no bouncing needed for ZONE_DMA, reduce the swiotlb > > + * buffer for kmalloc() bouncing to 1MB per 1GB of RAM. > > + */ > > + unsigned long size = > > + ALIGN(memblock_phys_mem_size(), SZ_1G) >> 10; > > Hmm, I wondered if DIV_ROUND_UP(memblock_phys_mem_size(), 1024) might be any > easier, but by the time I've typed it out it's still just as long :) It's not the same result as it's rounded up to the nearest KB rather than MB. But this may work as well. I just made up this number really.
On Fri, 06 Oct 2023 16:32:52 +0100, Catalin Marinas wrote: > With CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC enabled, the arm64 kernel still > allocates the default SWIOTLB buffer (64MB) even if ZONE_DMA is disabled > or all the RAM fits into this zone. However, this potentially wastes a > non-negligible amount of memory on platforms with little RAM. > > Reduce the SWIOTLB size to 1MB per 1GB of RAM if only needed for > kmalloc() buffer bouncing. > > [...] Applied to arm64 (for-next/misc), thanks! Also changed the division to DIV_ROUND_UP() as per Robin's suggestion. [1/1] arm64: swiotlb: Reduce the default size if no ZONE_DMA bouncing needed https://git.kernel.org/arm64/c/65033574ade9
diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c index 8a0f8604348b..54ee1a4868c2 100644 --- a/arch/arm64/mm/init.c +++ b/arch/arm64/mm/init.c @@ -493,8 +493,16 @@ void __init mem_init(void) { bool swiotlb = max_pfn > PFN_DOWN(arm64_dma_phys_limit); - if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC)) + if (IS_ENABLED(CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC) && !swiotlb) { + /* + * If no bouncing needed for ZONE_DMA, reduce the swiotlb + * buffer for kmalloc() bouncing to 1MB per 1GB of RAM. + */ + unsigned long size = + ALIGN(memblock_phys_mem_size(), SZ_1G) >> 10; + swiotlb_adjust_size(min(swiotlb_size_or_default(), size)); swiotlb = true; + } swiotlb_init(swiotlb, SWIOTLB_VERBOSE);
With CONFIG_DMA_BOUNCE_UNALIGNED_KMALLOC enabled, the arm64 kernel still allocates the default SWIOTLB buffer (64MB) even if ZONE_DMA is disabled or all the RAM fits into this zone. However, this potentially wastes a non-negligible amount of memory on platforms with little RAM. Reduce the SWIOTLB size to 1MB per 1GB of RAM if only needed for kmalloc() buffer bouncing. Signed-off-by: Catalin Marinas <catalin.marinas@arm.com> Suggested-by: Ross Burton <ross.burton@arm.com> Cc: Ross Burton <ross.burton@arm.com> Cc: Robin Murphy <robin.murphy@arm.com> Cc: Will Deacon <will@kernel.org> --- arch/arm64/mm/init.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)