[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <45c23e30-8405-470b-825c-e5166cd8a313@linux.dev>
Date: Sat, 28 Oct 2023 10:48:00 +0800
From: Zhu Yanjun <yanjun.zhu@...ux.dev>
To: Bart Van Assche <bvanassche@....org>,
Li Zhijian <lizhijian@...itsu.com>, zyjzyj2000@...il.com,
jgg@...pe.ca, leon@...nel.org, linux-rdma@...r.kernel.org
Cc: linux-kernel@...r.kernel.org, rpearsonhpe@...il.com,
matsuda-daisuke@...itsu.com
Subject: Re: [PATCH RFC 1/2] RDMA/rxe: don't allow registering !PAGE_SIZE mr
在 2023/10/28 5:46, Bart Van Assche 写道:
> On 10/27/23 01:17, Zhu Yanjun wrote:
>> When ULP uses folio or compound page, ULP can not work well with RXE
>> after this commit is applied.
>
> Isn't it the responsibility of the DMA mapping code to build an sg-list
> for folios? Drivers like ib_srp see an sg-list, whether that comes from
A folio is a way of representing a set of physically contiguous base
pages. In current implementations of folio, it seems that sg-list is not
used.
In Folio, some huge pages whose size is not PAGE_SIZE is dma-mapped into
hardware.
So the page size of folio is not equal to PAGE_SIZE. If this commit is
applied, it causes potential risks to the future folio.
I have developed some folio work for some NIC and RDMA drivers. In
Folio, the page size of Folio is possibly not equal to PAGE_SIZE, it is
multiple PAGE_SIZE. And when folio is dma-mapped to HW, the page size is
equal to multiple PAGE_SIZE.
In this case, ULP with folio will not work well with current RXE after
this commit is applied.
Removing page_size in RXE seems a plan for this problem.
Zhu Yanjun
> a folio or from something else.
>
> Bart.
Powered by blists - more mailing lists