On 2024-01-29 2:07 pm, Alexander Lobakin wrote:
From: Alexander Lobakin <aleksander.lobakin@xxxxxxxxx>
Date: Fri, 26 Jan 2024 17:45:11 +0100
From: Robin Murphy <robin.murphy@xxxxxxx>
Date: Fri, 26 Jan 2024 15:48:54 +0000
On 26/01/2024 1:54 pm, Alexander Lobakin wrote:
From: Eric Dumazet <edumazet@xxxxxxxxxx>
Quite often, NIC 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->skip_dma_sync boolean which is set during the device
initialization depending on the setup: dev_is_dma_coherent() for direct
DMA, !(sync_single_for_device || sync_single_for_cpu) or positive result
from the new callback, dma_map_ops::can_skip_sync for non-NULL DMA ops.
Then later, if/when swiotlb is used for the first time, the flag
is turned off, from swiotlb_tbl_map_single().
I think you could probably just promote the dma_uses_io_tlb flag from
SWIOTLB_DYNAMIC to a general SWIOTLB thing to serve this purpose now.
Nice catch!
BTW, this implies such hotpath check:
if (dev->dma_skip_sync && !READ_ONCE(dev->dma_uses_io_tlb))
// ...
This seems less effective than just resetting dma_skip_sync on first
allocation.
Well, my point is not to have a dma_skip_sync at all; I'm suggesting the
check would be:
if (dev_is_dma_coherent(dev) && dev_uses_io_tlb(dev))
...
where on the platform which cares about this most, that first condition
is a compile-time constant (and as implied, the second would want to be
similarly wrapped for !SWIOTLB configs).
Thanks,
Robin.
