| Message ID | 20200628111251.19108-1-song.bao.hua@hisilicon.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | make dma_alloc_coherent NUMA-aware by per-NUMA CMA | expand |
> -----Original Message----- > From: Song Bao Hua (Barry Song) > Sent: Sunday, June 28, 2020 11:13 PM > To: hch@lst.de; m.szyprowski@samsung.com; robin.murphy@arm.com; > will@kernel.org; ganapatrao.kulkarni@cavium.com; > catalin.marinas@arm.com > Cc: iommu@lists.linux-foundation.org; Linuxarm <linuxarm@huawei.com>; > linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org; Song Bao > Hua (Barry Song) <song.bao.hua@hisilicon.com> > Subject: [PATCH v3 0/2] make dma_alloc_coherent NUMA-aware by > per-NUMA CMA > > Ganapatrao Kulkarni has put some effort on making arm-smmu-v3 use local > memory to save command queues[1]. I also did similar job in patch > "iommu/arm-smmu-v3: allocate the memory of queues in local numa node" > [2] while not realizing Ganapatrao has done that before. > > But it seems it is much better to make dma_alloc_coherent() to be inherently > NUMA-aware on NUMA-capable systems. > > Right now, smmu is using dma_alloc_coherent() to get memory to save queues > and tables. Typically, on ARM64 server, there is a default CMA located at > node0, which could be far away from node2, node3 etc. > Saving queues and tables remotely will increase the latency of ARM SMMU > significantly. For example, when SMMU is at node2 and the default global > CMA is at node0, after sending a CMD_SYNC in an empty command queue, we > have to wait more than 550ns for the completion of the command > CMD_SYNC. > However, if we save them locally, we only need to wait for 240ns. > > with per-numa CMA, smmu will get memory from local numa node to save > command queues and page tables. that means dma_unmap latency will be > shrunk much. > > Meanwhile, when iommu.passthrough is on, device drivers which call dma_ > alloc_coherent() will also get local memory and avoid the travel between > numa nodes. > > [1] > https://lists.linuxfoundation.org/pipermail/iommu/2017-October/024455.htm > l > [2] https://www.spinics.net/lists/iommu/msg44767.html > > -v3: > * move to use page_to_nid() while freeing cma with respect to Robin's > comment, but this will only work after applying my below patch: > "mm/cma.c: use exact_nid true to fix possible per-numa cma leak" > https://marc.info/?l=linux-mm&m=159333034726647&w=2 > > * handle the case count <= 1 more properly according to Robin's > comment; > > * add pernuma_cma parameter to support dynamic setting of per-numa > cma size; > ideally we can leverage the CMA_SIZE_MBYTES, CMA_SIZE_PERCENTAGE and > "cma=" kernel parameter and avoid a new paramter separately for per- > numa cma. Practically, it is really too complicated considering the > below problems: > (1) if we leverage the size of default numa for per-numa, we have to > avoid creating two cma with same size in node0 since default cma is > probably on node0. > (2) default cma can consider the address limitation for old devices > while per-numa cma doesn't support GFP_DMA and GFP_DMA32. all > allocations with limitation flags will fallback to default one. > (3) hard to apply CMA_SIZE_PERCENTAGE to per-numa. it is hard to > decide if the percentage should apply to the whole memory size > or only apply to the memory size of a specific numa node. > (4) default cma size has CMA_SIZE_SEL_MIN and CMA_SIZE_SEL_MAX, it > makes things even more complicated to per-numa cma. > > I haven't figured out a good way to leverage the size of default cma > for per-numa cma. it seems a separate parameter for per-numa could > make life easier. > > * move dma_pernuma_cma_reserve() after hugetlb_cma_reserve() to > reuse the comment before hugetlb_cma_reserve() with respect to > Robin's comment > > -v2: > * fix some issues reported by kernel test robot > * fallback to default cma while allocation fails in per-numa cma > free memory properly > > Barry Song (2): > dma-direct: provide the ability to reserve per-numa CMA > arm64: mm: reserve per-numa CMA to localize coherent dma buffers > > .../admin-guide/kernel-parameters.txt | 9 ++ > arch/arm64/mm/init.c | 2 + > include/linux/dma-contiguous.h | 4 + > kernel/dma/Kconfig | 10 ++ > kernel/dma/contiguous.c | 98 > +++++++++++++++++-- > 5 files changed, 114 insertions(+), 9 deletions(-) Gentle ping :-) Thanks Barry
Ganapatrao Kulkarni has put some effort on making arm-smmu-v3 use local memory to save command queues[1]. I also did similar job in patch "iommu/arm-smmu-v3: allocate the memory of queues in local numa node" [2] while not realizing Ganapatrao has done that before. But it seems it is much better to make dma_alloc_coherent() to be inherently NUMA-aware on NUMA-capable systems. Right now, smmu is using dma_alloc_coherent() to get memory to save queues and tables. Typically, on ARM64 server, there is a default CMA located at node0, which could be far away from node2, node3 etc. Saving queues and tables remotely will increase the latency of ARM SMMU significantly. For example, when SMMU is at node2 and the default global CMA is at node0, after sending a CMD_SYNC in an empty command queue, we have to wait more than 550ns for the completion of the command CMD_SYNC. However, if we save them locally, we only need to wait for 240ns. with per-numa CMA, smmu will get memory from local numa node to save command queues and page tables. that means dma_unmap latency will be shrunk much. Meanwhile, when iommu.passthrough is on, device drivers which call dma_ alloc_coherent() will also get local memory and avoid the travel between numa nodes. [1] https://lists.linuxfoundation.org/pipermail/iommu/2017-October/024455.html [2] https://www.spinics.net/lists/iommu/msg44767.html -v3: * move to use page_to_nid() while freeing cma with respect to Robin's comment, but this will only work after applying my below patch: "mm/cma.c: use exact_nid true to fix possible per-numa cma leak" https://marc.info/?l=linux-mm&m=159333034726647&w=2 * handle the case count <= 1 more properly according to Robin's comment; * add pernuma_cma parameter to support dynamic setting of per-numa cma size; ideally we can leverage the CMA_SIZE_MBYTES, CMA_SIZE_PERCENTAGE and "cma=" kernel parameter and avoid a new paramter separately for per- numa cma. Practically, it is really too complicated considering the below problems: (1) if we leverage the size of default numa for per-numa, we have to avoid creating two cma with same size in node0 since default cma is probably on node0. (2) default cma can consider the address limitation for old devices while per-numa cma doesn't support GFP_DMA and GFP_DMA32. all allocations with limitation flags will fallback to default one. (3) hard to apply CMA_SIZE_PERCENTAGE to per-numa. it is hard to decide if the percentage should apply to the whole memory size or only apply to the memory size of a specific numa node. (4) default cma size has CMA_SIZE_SEL_MIN and CMA_SIZE_SEL_MAX, it makes things even more complicated to per-numa cma. I haven't figured out a good way to leverage the size of default cma for per-numa cma. it seems a separate parameter for per-numa could make life easier. * move dma_pernuma_cma_reserve() after hugetlb_cma_reserve() to reuse the comment before hugetlb_cma_reserve() with respect to Robin's comment -v2: * fix some issues reported by kernel test robot * fallback to default cma while allocation fails in per-numa cma free memory properly Barry Song (2): dma-direct: provide the ability to reserve per-numa CMA arm64: mm: reserve per-numa CMA to localize coherent dma buffers .../admin-guide/kernel-parameters.txt | 9 ++ arch/arm64/mm/init.c | 2 + include/linux/dma-contiguous.h | 4 + kernel/dma/Kconfig | 10 ++ kernel/dma/contiguous.c | 98 +++++++++++++++++-- 5 files changed, 114 insertions(+), 9 deletions(-)