From: Robin Murphy <robin.murphy@xxxxxxx> Date: Thu, 9 May 2024 15:33:13 +0100 > On 09/05/2024 2:43 pm, Steven Price wrote: >> On 07/05/2024 12:20, Alexander Lobakin wrote: >>> 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@xxxxxx> # direct DMA shortcut >>> Co-developed-by: Eric Dumazet <edumazet@xxxxxxxxxx> >>> Signed-off-by: Eric Dumazet <edumazet@xxxxxxxxxx> >>> Signed-off-by: Alexander Lobakin <aleksander.lobakin@xxxxxxxxx> >> >> I've bisected a boot failure (on a Firefly RK3288) to this commit. >> AFAICT the problem is that I have (at least) two drivers which don't >> call dma_set_mask() and therefore never initialise the new dma_need_sync >> variable. >> >> The specific drivers are "rockchip-drm" and "rk_gmac-dwmac". Is it a >> requirement that all drivers engaging in DMA should call dma_set_mask() >> - and therefore this has uncovered a bug in those drivers. Or is the >> assumption that all drivers call dma_set_mask() faulty? > > Historically it's long been documented (at least in DMA-API-HOWTO) that > a 32-bit DMA mask is assumed by default, so as much as we would prefer > to shift expectations, there are still going to be a great many drivers > relying on that :( > > Perhaps its time for dma-debug to start warning about implicit mask > usage, maybe that might help push the agenda a bit? I also thought of this, but currently don't know how to detect whether a driver has called dma_set_mask*(). The fix will arrive in several minutes. > > Thanks, > Robin. Thanks, Olek
