[Cc +NVIDIA] On Thu, 12 Mar 2020 23:05:48 -0400 Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote: > It is better for a device model to use IOVAs to read/write memory to > perform some sort of virtual DMA on behalf of the device. > > patch 1 exports VFIO group to external user so that it can hold the group > reference until finishing using of it. It saves ~500 cycles that are spent > on VFIO group looking up, referencing and dereferencing. (this data is > measured with 1 VFIO user). > > patch 2 introduces interface vfio_dma_rw(). > > patch 3 introduces interfaces vfio_group_pin_pages() and > vfio_group_unpin_pages() to get rid of VFIO group looking-up in > vfio_pin_pages() and vfio_unpin_pages(). > > patch 4-5 let kvmgt switch from calling kvm_read/write_guest() to calling > vfio_dma_rw to rw IOVAs. > > patch 6 let kvmgt switch to use lighter version of vfio_pin/unpin_pages(), > i.e. vfio_group_pin/unpin_pages() > > patch 7 enables kvmgt to read/write IOVAs of size larger than PAGE_SIZE. This looks pretty good to me, hopefully Kirti and Neo can find some advantage with this series as well. Given that we're also trying to get the migration interface and dirty page tracking integrated for v5.7, would it make sense to merge the first 3 patches via my next branch? This is probably the easiest way to update the dirty tracking. Thanks, Alex > Performance: > > Comparison between vfio_dma_rw() and kvm_read/write_guest(): > > 1. avergage CPU cycles of each interface measured with 1 running VM: > -------------------------------------------------- > | rw | avg cycles of | > | size | (vfio_dma_rw - kvm_read/write_guest) | > |---------- ---------------------------------------| > | <= 1 page | +155 ~ +195 | > |--------------------------------------------------| > | 5 pages | -530 | > |--------------------------------------------------| > | 20 pages | -2005 ~ -2533 | > -------------------------------------------------- > > 2. average scores > > base: base code before applying code in this series. use > kvm_read/write_pages() to rw IOVAs > > base + this series: use vfio_dma_rw() to read IOVAs and use > vfio_group_pin/unpin_pages(), and kvmgt is able to rw several pages > at a time. > > Scores of benchmarks running in 1 VM each: > ----------------------------------------------------------------- > | | glmark2 | lightsmark | openarena | heavens | > |-----------------------------------------------------------------| > | base | 1248 | 219.70 | 114.9 | 560 | > |-----------------------------------------------------------------| > |base + this series | 1248 | 225.8 | 115 | 559 | > ----------------------------------------------------------------- > > Sum of scores of two benchmark instances running in 2 VMs each: > ------------------------------------------------------- > | | glmark2 | lightsmark | openarena | > |-------------------------------------------------------| > | base | 812 | 211.46 | 115.3 | > |-------------------------------------------------------| > |base + this series | 814 | 214.69 | 115.9 | > ------------------------------------------------------- > > > Changelogs: > v3 --> v4: > - rebased to 5.6.0-rc4+ > - adjust wrap position for vfio_group_get_external_user_from_dev() in > header file.(Alex) > - updated function description of vfio_group_get_external_user_from_dev() > (Alex) > - fixed Error path group reference leaking in > vfio_group_get_external_user_from_dev() (Alex) > - reurn 0 for success or errno in vfio_dma_rw_chunk(). (Alex) > - renamed iova to user_iova in interface vfio_dam_rw(). > - renamed vfio_pin_pages_from_group() and vfio_unpin_pages_from_group() to > vfio_group_pin_pages() and vfio_group_unpin_pages() > - renamed user_pfn to user_iova_pfn in vfio_group_pin_pages() and > vfio_group_unpin_pages() > > v2 --> v3: > - add vfio_group_get_external_user_from_dev() to improve performance (Alex) > - add vfio_pin/unpin_pages_from_group() to avoid repeated looking up of > VFIO group in vfio_pin/unpin_pages() (Alex) > - add a check for IOMMU_READ permission. (Alex) > - rename vfio_iommu_type1_rw_dma_nopin() to > vfio_iommu_type1_dma_rw_chunk(). (Alex) > - in kvmgt, change "write ? vfio_dma_rw(...,true) : > vfio_dma_rw(...,false)" to vfio_dma_rw(dev, gpa, buf, len, write) > (Alex and Paolo) > - in kvmgt, instead of read/write context pages 1:1, combining the > reads/writes of continuous IOVAs to take advantage of vfio_dma_rw() for > faster crossing page boundary accesses. > > v1 --> v2: > - rename vfio_iova_rw to vfio_dma_rw, vfio iommu driver ops .iova_rw > to .dma_rw. (Alex). > - change iova and len from unsigned long to dma_addr_t and size_t, > respectively. (Alex) > - fix possible overflow in dma->vaddr + iova - dma->iova + offset (Alex) > - split DMAs from on page boundary to on max available size to eliminate > redundant searching of vfio_dma and switching mm. (Alex) > - add a check for IOMMU_WRITE permission. > > > Yan Zhao (7): > vfio: allow external user to get vfio group from device > vfio: introduce vfio_dma_rw to read/write a range of IOVAs > vfio: avoid inefficient operations on VFIO group in > vfio_pin/unpin_pages > drm/i915/gvt: hold reference of VFIO group during opening of vgpu > drm/i915/gvt: subsitute kvm_read/write_guest with vfio_dma_rw > drm/i915/gvt: switch to user vfio_group_pin/upin_pages > drm/i915/gvt: rw more pages a time for shadow context > > drivers/gpu/drm/i915/gvt/kvmgt.c | 46 ++++--- > drivers/gpu/drm/i915/gvt/scheduler.c | 97 ++++++++++----- > drivers/vfio/vfio.c | 180 +++++++++++++++++++++++++++ > drivers/vfio/vfio_iommu_type1.c | 76 +++++++++++ > include/linux/vfio.h | 13 ++ > 5 files changed, 360 insertions(+), 52 deletions(-) >
