]> git.dujemihanovic.xyz Git - linux.git/commit
dma: avoid redundant calls for sync operations
authorAlexander Lobakin <aleksander.lobakin@intel.com>
Tue, 7 May 2024 11:20:21 +0000 (13:20 +0200)
committerChristoph Hellwig <hch@lst.de>
Tue, 7 May 2024 11:29:53 +0000 (13:29 +0200)
commitf406c8e4b770ca3b0df84a17349e13f2b6b07d10
tree799929448e230d2e199fd5bd30c043700447eaee
parentfe7514b149e0a8a6f3031d286e52d40163b0b11a
dma: avoid redundant calls for sync operations

Quite often, devices do not need dma_sync operations on x86_64 at least.
Indeed, when dev_is_dma_coherent(dev) is true and
dev_use_swiotlb(dev) is false, iommu_dma_sync_single_for_cpu()
and friends do nothing.

However, indirectly calling them when CONFIG_RETPOLINE=y consumes about
10% of cycles on a cpu receiving packets from softirq at ~100Gbit rate.
Even if/when CONFIG_RETPOLINE is not set, there is a cost of about 3%.

Add dev->need_dma_sync boolean and turn it off during the device
initialization (dma_set_mask()) depending on the setup:
dev_is_dma_coherent() for the direct DMA, !(sync_single_for_device ||
sync_single_for_cpu) or the new dma_map_ops flag, %DMA_F_CAN_SKIP_SYNC,
advertised for non-NULL DMA ops.
Then later, if/when swiotlb is used for the first time, the flag
is reset back to on, from swiotlb_tbl_map_single().

On iavf, the UDP trafficgen with XDP_DROP in skb mode test shows
+3-5% increase for direct DMA.

Suggested-by: Christoph Hellwig <hch@lst.de> # direct DMA shortcut
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
include/linux/device.h
include/linux/dma-map-ops.h
include/linux/dma-mapping.h
kernel/dma/mapping.c
kernel/dma/swiotlb.c