On 8/9/2023 8:10 PM, Jason Wang wrote:> On Thu, Aug 10, 2023 at 8:40?AM Si-Wei Liu <si-wei.liu at oracle.com>
wrote:
>>
>>
>> On 8/8/2023 11:52 PM, Jason Wang wrote:
>>> On Wed, Aug 9, 2023 at 6:58?AM Si-Wei Liu <si-wei.liu at
oracle.com> wrote:
>>>>
>>>> On 8/7/2023 8:00 PM, Jason Wang wrote:
>>>>> On Fri, Aug 4, 2023 at 1:58?AM Si-Wei Liu <si-wei.liu at
oracle.com> wrote:
>>>>>> On 8/3/2023 1:03 AM, Jason Wang wrote:
>>>>>>> On Thu, Aug 3, 2023 at 1:13?AM Dragos Tatulea
<dtatulea at nvidia.com> wrote:
>>>>>>>> The mr->initialized flag is shared between
the control vq and data vq
>>>>>>>> part of the mr init/uninit. But if the control
vq and data vq get placed
>>>>>>>> in different ASIDs, it can happen that
initializing the control vq will
>>>>>>>> prevent the data vq mr from being initialized.
>>>>>>>>
>>>>>>>> This patch consolidates the control and data vq
init parts into their
>>>>>>>> own init functions. The mr->initialized will
now be used for the data vq
>>>>>>>> only. The control vq currently doesn't need
a flag.
>>>>>>>>
>>>>>>>> The uninitializing part is also taken care of:
mlx5_vdpa_destroy_mr got
>>>>>>>> split into data and control vq functions which
are now also ASID aware.
>>>>>>>>
>>>>>>>> Fixes: 8fcd20c30704 ("vdpa/mlx5: Support
different address spaces for control and data")
>>>>>>>> Signed-off-by: Dragos Tatulea <dtatulea at
nvidia.com>
>>>>>>>> Reviewed-by: Eugenio P?rez <eperezma at
redhat.com>
>>>>>>>> Reviewed-by: Gal Pressman <gal at
nvidia.com>
>>>>>>>> ---
>>>>>>>> drivers/vdpa/mlx5/core/mlx5_vdpa.h | 1 +
>>>>>>>> drivers/vdpa/mlx5/core/mr.c | 97
+++++++++++++++++++++---------
>>>>>>>> 2 files changed, 71 insertions(+), 27
deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>>>> index 25fc4120b618..a0420be5059f 100644
>>>>>>>> --- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>>>> +++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>>>>>>>> @@ -31,6 +31,7 @@ struct mlx5_vdpa_mr {
>>>>>>>> struct list_head head;
>>>>>>>> unsigned long num_directs;
>>>>>>>> unsigned long num_klms;
>>>>>>>> + /* state of dvq mr */
>>>>>>>> bool initialized;
>>>>>>>>
>>>>>>>> /* serialize mkey creation and
destruction */
>>>>>>>> diff --git a/drivers/vdpa/mlx5/core/mr.c
b/drivers/vdpa/mlx5/core/mr.c
>>>>>>>> index 03e543229791..4ae14a248a4b 100644
>>>>>>>> --- a/drivers/vdpa/mlx5/core/mr.c
>>>>>>>> +++ b/drivers/vdpa/mlx5/core/mr.c
>>>>>>>> @@ -489,60 +489,103 @@ static void
destroy_user_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_mr *mr
>>>>>>>> }
>>>>>>>> }
>>>>>>>>
>>>>>>>> -void mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev
*mvdev)
>>>>>>>> +static void _mlx5_vdpa_destroy_cvq_mr(struct
mlx5_vdpa_dev *mvdev, unsigned int asid)
>>>>>>>> +{
>>>>>>>> + if
(mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] != asid)
>>>>>>>> + return;
>>>>>>>> +
>>>>>>>> + prune_iotlb(mvdev);
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static void _mlx5_vdpa_destroy_dvq_mr(struct
mlx5_vdpa_dev *mvdev, unsigned int asid)
>>>>>>>> {
>>>>>>>> struct mlx5_vdpa_mr *mr =
&mvdev->mr;
>>>>>>>>
>>>>>>>> - mutex_lock(&mr->mkey_mtx);
>>>>>>>> + if
(mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] != asid)
>>>>>>>> + return;
>>>>>>>> +
>>>>>>>> if (!mr->initialized)
>>>>>>>> - goto out;
>>>>>>>> + return;
>>>>>>>>
>>>>>>>> - prune_iotlb(mvdev);
>>>>>>>> if (mr->user_mr)
>>>>>>>> destroy_user_mr(mvdev, mr);
>>>>>>>> else
>>>>>>>> destroy_dma_mr(mvdev, mr);
>>>>>>>>
>>>>>>>> mr->initialized = false;
>>>>>>>> -out:
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static void mlx5_vdpa_destroy_mr_asid(struct
mlx5_vdpa_dev *mvdev, unsigned int asid)
>>>>>>>> +{
>>>>>>>> + struct mlx5_vdpa_mr *mr =
&mvdev->mr;
>>>>>>>> +
>>>>>>>> + mutex_lock(&mr->mkey_mtx);
>>>>>>>> +
>>>>>>>> + _mlx5_vdpa_destroy_dvq_mr(mvdev, asid);
>>>>>>>> + _mlx5_vdpa_destroy_cvq_mr(mvdev, asid);
>>>>>>>> +
>>>>>>>> mutex_unlock(&mr->mkey_mtx);
>>>>>>>> }
>>>>>>>>
>>>>>>>> -static int _mlx5_vdpa_create_mr(struct
mlx5_vdpa_dev *mvdev,
>>>>>>>> - struct
vhost_iotlb *iotlb, unsigned int asid)
>>>>>>>> +void mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev
*mvdev)
>>>>>>>> +{
>>>>>>>> + mlx5_vdpa_destroy_mr_asid(mvdev,
mvdev->group2asid[MLX5_VDPA_CVQ_GROUP]);
>>>>>>>> + mlx5_vdpa_destroy_mr_asid(mvdev,
mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP]);
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static int _mlx5_vdpa_create_cvq_mr(struct
mlx5_vdpa_dev *mvdev,
>>>>>>>> + struct
vhost_iotlb *iotlb,
>>>>>>>> + unsigned
int asid)
>>>>>>>> +{
>>>>>>>> + if
(mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] != asid)
>>>>>>>> + return 0;
>>>>>>>> +
>>>>>>>> + return dup_iotlb(mvdev, iotlb);
>>>>>>> This worries me as conceptually, there should be no
difference between
>>>>>>> dvq mr and cvq mr. The virtqueue should be loosely
coupled with mr.
>>>>>>>
>>>>>>> One example is that, if we only do dup_iotlb() but
not try to create
>>>>>>> dma mr here, we will break virtio-vdpa:
>>>>>> For this case, I guess we may need another way to
support virtio-vdpa
>>>>>> 1:1 mapping rather than overloading virtio device reset
semantics, see:
>>>>>>
>>>>>> https://www.mail-archive.com/qemu-devel at
nongnu.org/msg953755.html
>>>>>>
>>>>>> > Conceptually, the address mapping is not a
part of the abstraction for
>>>>>> > a virtio device now. So resetting the memory
mapping during virtio
>>>>>> > device reset seems wrong.
>>>>>>
>>>>>> where we want to keep memory mapping intact across
virtio device reset
>>>>>> for best live migration latency/downtime. I wonder
would it work to
>>>>>> reset the mapping in vhost-vdpa life cycle out of
virtio reset, say
>>>>>> introduce a .reset_map() op to restore 1:1 mapping
within
>>>>>> vhost_vdpa_remove_as() right after
vhost_vdpa_iotlb_unmap()? Then we can
>>>>>> move the iotlb reset logic to there without worry
breaking virtio-vdpa.
>>>>> It looks to me we don't need a new ops. We can simply
do set_map()
>>>>> twice
>>>> What does it mean, first set_map(0, -1ULL) with zero iotlb
entry passed
>>>> in to destroy all iotlb mappings previously added, and second
set_map(0,
>>>> -1ULL) to restore 1:1 DMA MR? But userspace (maybe a buggy one
but
>>>> doesn't do harm) apart from vhost-vdpa itself can do unmap
twice anyway,
>>>> this is supported today I think. Why there'll be such
obscure
>>>> distinction, or what's the benefit to treat second
.set_map() as
>>>> recreating 1:1 mapping?
>>> Ok, I think I miss some context. I agree that it's better to
decouple
>>> memory mappings from the virtio reset. It helps to reduce the
>>> unnecessary memory transactions. It might require a new feature
flag.
>> This I agreed. AFAICT QEMU would need to check this new feature flag to
>> make sure memory mappings are kept intact across reset, otherwise for
>> the sake of avoid breaking older kernels it has to recreate all the
>> mappings after reset like how it is done today.
>>
>>> Regarding the method of restoring to 1:1 DMA MR, it might be
dangerous
>>> for (buggy) vhost-vDPA devices. Since its userspace doesn't set
up any
>>> mapping it can explore the kernel with that via CVQ?
>> Not sure I understand this proposal. The 1:1 DMA MR is first created at
>> vdpa device add, and gets destroyed implicitly when the first .set_map
>> or .dma_map call is made, which is only possible after the vhost-vdpa
>> module is loaded and bound to vdpa devices.
> So what happens if there's a buggy userspace that doesn't do any
IOTLB setup?
Then parent driver doesn't do anything in .reset_map() - as the DMA MR
is still there. Parent driver should be able to tell apart if DMA MR has
been destroyed or not by checking the internal state.
-Siwei
>
> Thanks
>
>> Naturally the DMA MR should
>> be restored to how it was before when vhost-vdpa module is unloaded, or
>> if anything the 1:1 DMA MR creation can be deferred to until
virtio-vdpa
>> is probed and bound to devices. Today vhost_vdpa_remove_as() as part of
>> the vhost-vdpa unload code path already gets all mappings purged
through
>> vhost_vdpa_iotlb_unmap(0, -1ULL), and it should be pretty safe to
>> restore DMA MR via .reset_map() right after. Not sure what's the
concern
>> here with buggy vhost-vdpa device?
>>
>> Noted when vhost-vdpa is being unloaded there's even no chance to
probe
>> kernel through CVQ, as the virtio feature is not even negotiated at
that
>> point. And it is even trickier to wait for CVQ response from device
>> indefinitely when trying to unload a module.
>>
>> Regards,
>> -Siwei
>>> Thanks
>>>
>>>>> or do you mean it would be faster?
>>>> I think with .reset_map() we at least can avoid indefinite
latency
>>>> hiccup from destroying and recreating 1:1 mapping with the
unwarranted
>>>> 2rd unmap call. And .reset_map() should work with both
.dma_map() and
>>>> .set_map() APIs with clear semantics.
>>>>
>>>> Regards,
>>>> -Siwei
>>>>> Thanks
>>>>>
>>>>>> Thanks,
>>>>>> -Siwei
>>>>>>
>>>>>>> commit 6f5312f801836e6af9bcbb0bdb44dc423e129206
>>>>>>> Author: Eli Cohen <elic at nvidia.com>
>>>>>>> Date: Wed Jun 2 11:58:54 2021 +0300
>>>>>>>
>>>>>>> vdpa/mlx5: Add support for running with
virtio_vdpa
>>>>>>>
>>>>>>> In order to support running vdpa using
vritio_vdpa driver, we need to
>>>>>>> create a different kind of MR, one that has
1:1 mapping, since the
>>>>>>> addresses referring to virtqueues are dma
addresses.
>>>>>>>
>>>>>>> We create the 1:1 MR in mlx5_vdpa_dev_add()
only in case firmware
>>>>>>> supports the general capability umem_uid_0.
The reason for that is that
>>>>>>> 1:1 MRs must be created with uid == 0 while
virtqueue objects can be
>>>>>>> created with uid == 0 only when the
firmware capability is on.
>>>>>>>
>>>>>>> If the set_map() callback is called with
new translations provided
>>>>>>> through iotlb, the driver will destroy the
1:1 MR and create a regular
>>>>>>> one.
>>>>>>>
>>>>>>> Signed-off-by: Eli Cohen <elic at
nvidia.com>
>>>>>>> Link:
https://lore.kernel.org/r/20210602085854.62690-1-elic at nvidia.com
>>>>>>> Signed-off-by: Michael S. Tsirkin <mst
at redhat.com>
>>>>>>> Acked-by: Jason Wang <jasowang at
redhat.com>
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>>
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static int _mlx5_vdpa_create_dvq_mr(struct
mlx5_vdpa_dev *mvdev,
>>>>>>>> + struct
vhost_iotlb *iotlb,
>>>>>>>> + unsigned
int asid)
>>>>>>>> {
>>>>>>>> struct mlx5_vdpa_mr *mr =
&mvdev->mr;
>>>>>>>> int err;
>>>>>>>>
>>>>>>>> - if (mr->initialized)
>>>>>>>> + if
(mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] != asid)
>>>>>>>> return 0;
>>>>>>>>
>>>>>>>> - if
(mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] == asid) {
>>>>>>>> - if (iotlb)
>>>>>>>> - err =
create_user_mr(mvdev, iotlb);
>>>>>>>> - else
>>>>>>>> - err =
create_dma_mr(mvdev, mr);
>>>>>>>> + if (mr->initialized)
>>>>>>>> + return 0;
>>>>>>>>
>>>>>>>> - if (err)
>>>>>>>> - return err;
>>>>>>>> - }
>>>>>>>> + if (iotlb)
>>>>>>>> + err = create_user_mr(mvdev,
iotlb);
>>>>>>>> + else
>>>>>>>> + err = create_dma_mr(mvdev, mr);
>>>>>>>>
>>>>>>>> - if
(mvdev->group2asid[MLX5_VDPA_CVQ_GROUP] == asid) {
>>>>>>>> - err = dup_iotlb(mvdev, iotlb);
>>>>>>>> - if (err)
>>>>>>>> - goto out_err;
>>>>>>>> - }
>>>>>>>> + if (err)
>>>>>>>> + return err;
>>>>>>>>
>>>>>>>> mr->initialized = true;
>>>>>>>> +
>>>>>>>> + return 0;
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static int _mlx5_vdpa_create_mr(struct
mlx5_vdpa_dev *mvdev,
>>>>>>>> + struct
vhost_iotlb *iotlb, unsigned int asid)
>>>>>>>> +{
>>>>>>>> + int err;
>>>>>>>> +
>>>>>>>> + err = _mlx5_vdpa_create_dvq_mr(mvdev,
iotlb, asid);
>>>>>>>> + if (err)
>>>>>>>> + return err;
>>>>>>>> +
>>>>>>>> + err = _mlx5_vdpa_create_cvq_mr(mvdev,
iotlb, asid);
>>>>>>>> + if (err)
>>>>>>>> + goto out_err;
>>>>>>>> +
>>>>>>>> return 0;
>>>>>>>>
>>>>>>>> out_err:
>>>>>>>> - if
(mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP] == asid) {
>>>>>>>> - if (iotlb)
>>>>>>>> - destroy_user_mr(mvdev,
mr);
>>>>>>>> - else
>>>>>>>> - destroy_dma_mr(mvdev,
mr);
>>>>>>>> - }
>>>>>>>> + _mlx5_vdpa_destroy_dvq_mr(mvdev, asid);
>>>>>>>>
>>>>>>>> return err;
>>>>>>>> }
>>>>>>>> --
>>>>>>>> 2.41.0
>>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> Virtualization mailing list
>>>>>>> Virtualization at lists.linux-foundation.org
>>>>>>>
https://lists.linuxfoundation.org/mailman/listinfo/virtualization