mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
From: Kirti Wankhede <kwankhede@nvidia.com>
To: Alex Williamson <alex.williamson@redhat.com>
Cc: <pbonzini@redhat.com>, <kraxel@redhat.com>, <cjia@nvidia.com>,
	<qemu-devel@nongnu.org>, <kvm@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 1/1] vfio iommu type1: set size to PAGE_SIZE while looking for iova in dma list
Date: Tue, 6 Dec 2016 09:26:36 +0530	[thread overview]
Message-ID: <130cf141-d9a0-4730-2a99-46eeb5dae664@nvidia.com> (raw)
In-Reply-To: <20161205155421.075bbec5@t450s.home>



On 12/6/2016 4:24 AM, Alex Williamson wrote:
> On Tue, 6 Dec 2016 02:38:01 +0530
> Kirti Wankhede <kwankhede@nvidia.com> wrote:
> 
>> In the functions of pin_pages/unpin_pages from mdev vendor driver,
>> vfio_find_dma() should be called with size as PAGE_SIZE instead of 0.
>> vfio_find_dma() searches for the range in dma_list.
>>
>> In vfio_dma_do_unmap(), vfio_find_dma() when used to look for start
>> address of unmap->iova, size should be 1, not 0. Otherwise vfio_find_dma()
>> returns NULL.
> 
> Hi Kirti,
> 
> I'd prefer to fix the preexisting case of this issue as a separate
> patch, we might want to backport it for stable trees separately from the
> new cases introduced with mdev.  It would also be great if we could go
> into some detail about why size=0 doesn't work in these cases, ie.
> vfio_find_dma() accounting for the end address of the range makes it
> fall through at the wrong branch point.  For the case of the
> preexisting issue, it would also be useful to mention that the other
> case of passing size=0 works correctly due to the -1 in the start
> address calculation.  Finally, I think I agree with your choice to use
> PAGE_SIZE in one case and 1 in the preexisting case, but perhaps we
> could spell that out as one is related to a page size interface and the
> other is trying to test boundary conditions.  Thanks,
> 

Thanks Alex,

Actual problem here is: vfio_dma keeps track of address range from
(dma->iova + 0) to (dma->iova + dma->size - 1), while vfio_find_dma()
search logic looks for range from (dma->iova + 1) to (dma->iova +
dma->size). So either we have to fix the size argument send to
vfio_find_dma() or we have to fix rb-tree creation and search logic.
I took the former path here. I'll add this detail in commit description
and split it as you mentioned above.

Thanks,
Kirti

> Alex
> 
>> Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
>> Signed-off-by: Neo Jia <cjia@nvidia.com>
>> Change-Id: Iee6fd45441c342b5e8626087046b2e0075d19a08
>> ---
>>  drivers/vfio/vfio_iommu_type1.c | 8 ++++----
>>  1 file changed, 4 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>> index a28fbddb505c..023fba7b8d5a 100644
>> --- a/drivers/vfio/vfio_iommu_type1.c
>> +++ b/drivers/vfio/vfio_iommu_type1.c
>> @@ -581,7 +581,7 @@ static int vfio_iommu_type1_pin_pages(void *iommu_data,
>>  		struct vfio_pfn *vpfn;
>>  
>>  		iova = user_pfn[i] << PAGE_SHIFT;
>> -		dma = vfio_find_dma(iommu, iova, 0);
>> +		dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
>>  		if (!dma) {
>>  			ret = -EINVAL;
>>  			goto pin_unwind;
>> @@ -622,7 +622,7 @@ pin_unwind:
>>  		dma_addr_t iova;
>>  
>>  		iova = user_pfn[j] << PAGE_SHIFT;
>> -		dma = vfio_find_dma(iommu, iova, 0);
>> +		dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
>>  		vfio_unpin_page_external(dma, iova, do_accounting);
>>  		phys_pfn[j] = 0;
>>  	}
>> @@ -659,7 +659,7 @@ static int vfio_iommu_type1_unpin_pages(void *iommu_data,
>>  		dma_addr_t iova;
>>  
>>  		iova = user_pfn[i] << PAGE_SHIFT;
>> -		dma = vfio_find_dma(iommu, iova, 0);
>> +		dma = vfio_find_dma(iommu, iova, PAGE_SIZE);
>>  		if (!dma)
>>  			goto unpin_exit;
>>  		vfio_unpin_page_external(dma, iova, do_accounting);
>> @@ -826,7 +826,7 @@ again:
>>  	 * mappings within the range.
>>  	 */
>>  	if (iommu->v2) {
>> -		dma = vfio_find_dma(iommu, unmap->iova, 0);
>> +		dma = vfio_find_dma(iommu, unmap->iova, 1);
>>  		if (dma && dma->iova != unmap->iova) {
>>  			ret = -EINVAL;
>>  			goto unlock;
> 

      reply	other threads:[~2016-12-06  4:04 UTC|newest]

Thread overview: 3+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-12-05 21:08 Kirti Wankhede
2016-12-05 22:54 ` Alex Williamson
2016-12-06  3:56   ` Kirti Wankhede [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=130cf141-d9a0-4730-2a99-46eeb5dae664@nvidia.com \
    --to=kwankhede@nvidia.com \
    --cc=alex.williamson@redhat.com \
    --cc=cjia@nvidia.com \
    --cc=kraxel@redhat.com \
    --cc=kvm@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=pbonzini@redhat.com \
    --cc=qemu-devel@nongnu.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®