[RFC,V3,0/6] vhost: accelerate metadata access

Message

Jason Wang April 23, 2019, 5:54 a.m. UTC

This series tries to access virtqueue metadata through kernel virtual
address instead of copy_user() friends since they had too much
overheads like checks, spec barriers or even hardware feature
toggling. This is done through setup kernel address through direct
mapping and co-opreate VM management with MMU notifiers.

Test shows about 23% improvement on TX PPS. TCP_STREAM doesn't see
obvious improvement.

Thanks

Changes from RFC V2:
- switch to use direct mapping instead of vmap()
- switch to use spinlock + RCU to synchronize MMU notifier and vhost
  data/control path
- set dirty pages in the invalidation callbacks
- always use copy_to/from_users() friends for the archs that may need
  flush_dcache_pages()
- various minor fixes
Changes from V4:
- use invalidate_range() instead of invalidate_range_start()
- track dirty pages
Changes from V3:
- don't try to use vmap for file backed pages
- rebase to master
Changes from V2:
- fix buggy range overlapping check
- tear down MMU notifier during vhost ioctl to make sure invalidation
  request can read metadata userspace address and vq size without
  holding vq mutex.
Changes from V1:
- instead of pinning pages, use MMU notifier to invalidate vmaps and
  remap duing metadata prefetch
- fix build warning on MIPS

Jason Wang (6):
  vhost: generalize adding used elem
  vhost: fine grain userspace memory accessors
  vhost: rename vq_iotlb_prefetch() to vq_meta_prefetch()
  vhost: introduce helpers to get the size of metadata area
  vhost: factor out setting vring addr and num
  vhost: access vq metadata through kernel virtual address

 drivers/vhost/net.c   |   4 +-
 drivers/vhost/vhost.c | 852 ++++++++++++++++++++++++++++++++++++------
 drivers/vhost/vhost.h |  34 +-
 3 files changed, 764 insertions(+), 126 deletions(-)

Comments

Jason Wang May 5, 2019, 9:20 a.m. UTC | #1

On 2019/4/23 下午1:54, Jason Wang wrote:
> This series tries to access virtqueue metadata through kernel virtual
> address instead of copy_user() friends since they had too much
> overheads like checks, spec barriers or even hardware feature
> toggling. This is done through setup kernel address through direct
> mapping and co-opreate VM management with MMU notifiers.
>
> Test shows about 23% improvement on TX PPS. TCP_STREAM doesn't see
> obvious improvement.


Ping. Comments are more than welcomed.

Thanks