Received-SPF: pass (google.com: domain of linux-kernel+bounces-17916-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Fri, 5 Jan 2024 10:18:49 -0400
From: Jason Gunthorpe <jgg@ziepe.ca>
To: Junxian Huang <huangjunxian6@hisilicon.com>
Cc: Leon Romanovsky <leon@kernel.org>, linux-rdma@vger.kernel.org,
	linuxarm@huawei.com, linux-kernel@vger.kernel.org
Subject: Re: [PATCH for-next 4/6] RDMA/hns: Support flexible pagesize
Message-ID: <20240105141849.GE50608@ziepe.ca>
References: <20231225075330.4116470-1-huangjunxian6@hisilicon.com>
 <20231225075330.4116470-5-huangjunxian6@hisilicon.com>
 <20231226085202.GA13350@unreal>
 <fbd65691-b0a2-0963-96fc-7e09a66cd203@hisilicon.com>
 <20240104202902.GD50608@ziepe.ca>
 <3d4cdc2e-0053-f486-7323-72127027886f@hisilicon.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <3d4cdc2e-0053-f486-7323-72127027886f@hisilicon.com>

On Fri, Jan 05, 2024 at 01:55:11PM +0800, Junxian Huang wrote:
> 
> 
> On 2024/1/5 4:29, Jason Gunthorpe wrote:
> > On Tue, Dec 26, 2023 at 05:16:33PM +0800, Junxian Huang wrote:
> >>
> >>
> >> On 2023/12/26 16:52, Leon Romanovsky wrote:
> >>> On Mon, Dec 25, 2023 at 03:53:28PM +0800, Junxian Huang wrote:
> >>>> From: Chengchang Tang <tangchengchang@huawei.com>
> >>>>
> >>>> In the current implementation, a fixed page size is used to
> >>>> configure the PBL, which is not flexible enough and is not
> >>>> conducive to the performance of the HW.
> >>>>
> >>>> Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
> >>>> Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
> >>>> ---
> >>>>  drivers/infiniband/hw/hns/hns_roce_alloc.c  |   6 -
> >>>>  drivers/infiniband/hw/hns/hns_roce_device.h |   9 ++
> >>>>  drivers/infiniband/hw/hns/hns_roce_mr.c     | 168 +++++++++++++++-----
> >>>>  3 files changed, 139 insertions(+), 44 deletions(-)
> >>>
> >>> I'm wonder if the ib_umem_find_best_pgsz() API should be used instead.
> >>> What is missing there?
> >>>
> >>> Thanks
> >>
> >> Actually this API is used for umem.
> >> For kmem, we add hns_roce_find_buf_best_pgsz() to do a similar job.
> > 
> > But why do you need to do something like this for kmem? It looked to
> > me like kmem knows its allocation size when it was allocated, how come
> > you need to iterate over all of it again?
> > 
> > Jason
> > 
> 
> kmem was split into multiple small pages for allocation to prevent allocation
> failure due to memory fragmentation.
> 
> And now we add this function to confirm whether these small pages have contiguous
> address. If so, they can be combined into one huge page for use, which is more
> likely when iommu/smmu is enabled.

That seems unncessary. The chances you get contiguous pages from
a lot of random allocations is really slim.

If you care about this optimization then you should have the allocator
explicitly request contiguous pages with high order allocations in a
manner that quickly fails and falls back to PAGE_SIZE.

Then you just use the size that the allocator was able to get, not try
to figure it out after the fact.

Jason