Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH] vdpa/mlx5: Add support for doorbell bypassing
To:     Eli Cohen <elic@nvidia.com>
Cc:     mst@redhat.com, virtualization@lists.linux-foundation.org,
        linux-kernel@vger.kernel.org,
        "Zhu, Lingshan" <lingshan.zhu@intel.com>
References: <20210421104145.115907-1-elic@nvidia.com>
 <e1885255-34f2-9e90-6478-ff0850a5a3d4@redhat.com>
 <20210422060358.GA140698@mtl-vdi-166.wap.labs.mlnx>
 <20210422080725.GB140698@mtl-vdi-166.wap.labs.mlnx>
 <9d3d8976-800d-bb14-0a4a-c4b008f6872c@redhat.com>
 <20210422083902.GA146406@mtl-vdi-166.wap.labs.mlnx>
From:   Jason Wang <jasowang@redhat.com>
Message-ID: <bdf10e38-8746-51cf-b460-a904a133329c@redhat.com>
Date:   Thu, 22 Apr 2021 16:59:11 +0800
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:78.0)
 Gecko/20100101 Thunderbird/78.10.0
MIME-Version: 1.0
In-Reply-To: <20210422083902.GA146406@mtl-vdi-166.wap.labs.mlnx>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


在 2021/4/22 下午4:39, Eli Cohen 写道:
> On Thu, Apr 22, 2021 at 04:21:45PM +0800, Jason Wang wrote:
>> 在 2021/4/22 下午4:07, Eli Cohen 写道:
>>> On Thu, Apr 22, 2021 at 09:03:58AM +0300, Eli Cohen wrote:
>>>> On Thu, Apr 22, 2021 at 10:37:38AM +0800, Jason Wang wrote:
>>>>> 在 2021/4/21 下午6:41, Eli Cohen 写道:
>>>>>> Implement mlx5_get_vq_notification() to return the doorbell address.
>>>>>> Size is set to one system page as required.
>>>>>>
>>>>>> Signed-off-by: Eli Cohen <elic@nvidia.com>
>>>>>> ---
>>>>>>     drivers/vdpa/mlx5/core/mlx5_vdpa.h | 1 +
>>>>>>     drivers/vdpa/mlx5/core/resources.c | 1 +
>>>>>>     drivers/vdpa/mlx5/net/mlx5_vnet.c  | 6 ++++++
>>>>>>     3 files changed, 8 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>> index b6cc53ba980c..49de62cda598 100644
>>>>>> --- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>> +++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>> @@ -41,6 +41,7 @@ struct mlx5_vdpa_resources {
>>>>>>     	u32 pdn;
>>>>>>     	struct mlx5_uars_page *uar;
>>>>>>     	void __iomem *kick_addr;
>>>>>> +	u64 phys_kick_addr;
>>>>>>     	u16 uid;
>>>>>>     	u32 null_mkey;
>>>>>>     	bool valid;
>>>>>> diff --git a/drivers/vdpa/mlx5/core/resources.c b/drivers/vdpa/mlx5/core/resources.c
>>>>>> index 6521cbd0f5c2..665f8fc1710f 100644
>>>>>> --- a/drivers/vdpa/mlx5/core/resources.c
>>>>>> +++ b/drivers/vdpa/mlx5/core/resources.c
>>>>>> @@ -247,6 +247,7 @@ int mlx5_vdpa_alloc_resources(struct mlx5_vdpa_dev *mvdev)
>>>>>>     		goto err_key;
>>>>>>     	kick_addr = mdev->bar_addr + offset;
>>>>>> +	res->phys_kick_addr = kick_addr;
>>>>>>     	res->kick_addr = ioremap(kick_addr, PAGE_SIZE);
>>>>>>     	if (!res->kick_addr) {
>>>>>> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
>>>>>> index 10c5fef3c020..680751074d2a 100644
>>>>>> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
>>>>>> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
>>>>>> @@ -1865,8 +1865,14 @@ static void mlx5_vdpa_free(struct vdpa_device *vdev)
>>>>>>     static struct vdpa_notification_area mlx5_get_vq_notification(struct vdpa_device *vdev, u16 idx)
>>>>>>     {
>>>>>> +	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
>>>>>>     	struct vdpa_notification_area ret = {};
>>>>>> +	struct mlx5_vdpa_net *ndev;
>>>>>> +
>>>>>> +	ndev = to_mlx5_vdpa_ndev(mvdev);
>>>>>> +	ret.addr = (phys_addr_t)ndev->mvdev.res.phys_kick_addr;
>>>>>> +	ret.size = PAGE_SIZE;
>>>>> Note that the page will be mapped in to guest, so it's only safe if the
>>>>> doorbeel exclusively own the page. This means if there're other registers in
>>>>> the page, we can not let the doorbell bypass to work.
>>>>>
>>>>> So this is suspicious at least in the case of subfunction where we calculate
>>>>> the bar length in mlx5_sf_dev_table_create() as:
>>>>>
>>>>> table->sf_bar_length = 1 << (MLX5_CAP_GEN(dev, log_min_sf_size) + 12);
>>>>>
>>>>> It looks to me this can only work for the arch with PAGE_SIZE = 4096,
>>>>> otherwise we can map more into the userspace(guest).
>>>>>
>>>> Correct, so I guess I should return here 4096.
>>
>> I'm not quite sure but since the calculation of the sf_bar_length is doen
>> via a shift of 12, it might be correct.
>>
>> And please double check if the doorbell own the page exclusively.
> I am checking if it is safe to map the any part of the SF's BAR to
> userspace without harming other functions. If this is true, I will check
> if I can return PAGE_SIZE without compromising security.


It's usally not safe and a layer violation if other registers are placed 
at the same page.


>   I think we may
> need to extend struct vdpa_notification_area to contain another field
> offset which indicates the offset from addr where the actual doorbell
> resides.


The movitiaton of the current design is to be fit seamless into how Qemu 
model doorbell layouts currently:

1) page-per-vq, each vq has its own page aligned doorbell
2) 2 bytes doorbell, each vq has its own 2 byte aligend doorbell

Only 1) is support in vhost-vDPA (and vhost-user) since it's rather 
simple and secure (page aligned) to be modelled and implemented via mmap().

Exporting a complex layout is possbile but requires careful design.

Actually, we had antoher option

3) shared doorbell: all virtqueue shares a single page aligned doorbell

This is not yet supported by Qemu.

Thanks


>>
>>>> I also think that the check in vhost_vdpa_mmap() should verify that the
>>>> returned size is not smaller than PAGE_SIZE because the returned address
>>> Actually I think it's ok since you verify the size equals vma->vm_end -
>>> vma->vm_start which must be at least PAGE_SIZE.
>>
>> Yes.
>>
>> Thanks
>>
>>
>>>> might just be aligned to PAGE_SIZE. I think this should be enoght but
>>>> maybe also use the same logic in vhost_vdpa_fault().