Re: [PATCH rdma-next v3 0/4] Improve ODP by using HMM API

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Sep 30, 2020 at 07:38:24PM +0300, Leon Romanovsky wrote:

> This series improves ODP performance by moving to use the HMM API as of below.
> 
> The get_user_pages_remote() functionality was replaced by HMM:
> - No need anymore to allocate and free memory to hold its output per call.
> - No need anymore to use the put_page() to unpin the pages.
> - The logic to detect contiguous pages is done based on the returned order
>   from HMM, no need to run per page, and evaluate.
> 
> Moving to use the HMM enables to reduce page faults in the system by using the
> snapshot mode. This mode allows existing pages in the CPU to become presented
> to the device without faulting.
> 
> This non-faulting mode may be used explicitly by an application with some new
> option of advice MR (i.e. PREFETCH_NO_FAULT) and is used upon ODP MR
> registration internally as part of initiating the device page table.
> 
> To achieve the above, internal changes in the ODP data structures were done
> and some flows were cleaned-up/adapted accordingly.
>
> Thanks
> 
> Yishai Hadas (4):
>   IB/core: Improve ODP to use hmm_range_fault()
>   IB/core: Enable ODP sync without faulting
>   RDMA/mlx5: Extend advice MR to support non faulting mode
>   RDMA/mlx5: Sync device with CPU pages upon ODP MR registration

Applied to for-next, thanks

Jason