[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <da6bc7cb-ef13-08cd-f162-663b4a66043b@de.ibm.com>
Date: Tue, 5 May 2020 16:31:47 +0200
From: Christian Borntraeger <borntraeger@...ibm.com>
To: Dave Hansen <dave.hansen@...el.com>,
Ulrich Weigand <uweigand@...ibm.com>
Cc: Claudio Imbrenda <imbrenda@...ux.ibm.com>, viro@...iv.linux.org.uk,
david@...hat.com, akpm@...ux-foundation.org, aarcange@...hat.com,
linux-mm@...ck.org, frankja@...ux.ibm.com, sfr@...b.auug.org.au,
jhubbard@...dia.com, linux-kernel@...r.kernel.org,
linux-s390@...r.kernel.org, jack@...e.cz, kirill@...temov.name,
peterz@...radead.org, sean.j.christopherson@...el.com,
Ulrich.Weigand@...ibm.com
Subject: Re: [PATCH v2 1/1] fs/splice: add missing callback for inaccessible
pages
On 05.05.20 16:24, Dave Hansen wrote:
> On 5/5/20 7:00 AM, Christian Borntraeger wrote:
>> We are certainly not married to our approach. I would happily extend/change
>> this to anything that works for your case and the s390 case. So can you outline
>> your requirements a bit more?
>
> For SEV, the guest define which pages are encrypted or not. You could
> theoretically do DMA to them or have the CPU access their contents, but
> you'd get either get ciphertext for reads, or data corruption and loss
> of cache coherency for writes. That's not so cool.
>
> Ideally, we would stop the CPU from ever accessing those pages by
> unmapping them. But, the pages go in and out of the encrypted state and
> the host really needs to be *sure* about what's going on before it
> restores its mapping and messes with the page. That includes situations
> where someone does a gup, starts an I/O to an unencrypted page, then the
> guest tries to convert that page over to being encrypted.
>
> So, the requirements are:
>
> 1. Allow host-side DMA and CPU access to shared pages
> 2. Stop host-side DMA and CPU access to encrypted pages
> 3. Allow pages to be converted between the states at the request of the
> guest
>
> Stopping the DMA is pretty easy, even across the gazillions of drivers
> in the tree because even random ethernet drivers do stuff like:
>
> txdr->buffer_info[i].dma =
> dma_map_single(&pdev->dev, skb->data, skb->len,
> DMA_TO_DEVICE);
>
> So the DMA can be stopped at the mapping layer. It's a *LOT* easier to
> catch there since the IOMMUs already provide isolation between the I/O
> and CPU address spaces.
And your problem is that the guest could convert this after the dma_map?
So you looked into our code if this would help?
Powered by blists - more mailing lists