On Fri, 2014-05-09 at 14:19 -0600, Bjorn Helgaas wrote: > On Fri, May 9, 2014 at 1:52 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote: > > On Friday 09 May 2014 09:32:26 Bjorn Helgaas wrote: > >> > There are people who care deeply about the performance of IOMMU API > >> > map/unmap. It isn't used *just* for virtual machines any more. See > >> > drivers/infiniband/hw/usnic/usnic_uiom.c for example. > >> > >> Of course we should care about IOMMU API performance. We should also > >> care about interface consistency, and it seems there's a tradeoff in > >> this case. I said "relatively" because I expect map/unmap to be less > >> frequent than read/write operations that use the mapping. I don't > >> know anything about infiniband, so maybe that assumption is false > >> there. > > > > In most drivers using the streaming DMA API, every mapping is used > > exactly once. Think of network or block drivers: they rarely send > > the same data twice to the device, and it usually comes from or > > goes to some user space buffer. > > Oh, good point. I don't work that high up in the stack, so thanks for > reminding me of that. To round this out, for most devices we have two types of mappings: the mailbox ones, which designate regions of communication memory between the kernel and the device which are usually permanent mappings, and the transmission mappings: every bit of data we send to the device is mapped, sent/received and then unmapped. The setup and teardown costs factor into the throughput. Some high iops devices (like SSD or high speed net) are peculiarly sensitive to this. James -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
