Re: [PATCH vhost 7/7] vdpa/mlx5: Postpone MR deletion

[Dragos Tatulea <dtatulea@nvidia.com>]

On 29.08.24 17:07, Eugenio Perez Martin wrote:
> On Wed, Aug 21, 2024 at 1:42 PM Dragos Tatulea <dtatulea@nvidia.com> wrote:
>>
>> Currently, when a new MR is set up, the old MR is deleted. MR deletion
>> is about 30-40% the time of MR creation. As deleting the old MR is not
>> important for the process of setting up the new MR, this operation
>> can be postponed.
>>
>> This series adds a workqueue that does MR garbage collection at a later
>> point. If the MR lock is taken, the handler will back off and
>> reschedule. The exception during shutdown: then the handler must
>> not postpone the work.
>>
>> Note that this is only a speculative optimization: if there is some
>> mapping operation that is triggered while the garbage collector handler
>> has the lock taken, this operation it will have to wait for the handler
>> to finish.
>>
>> Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
>> Reviewed-by: Cosmin Ratiu <cratiu@nvidia.com>
>> ---
>>  drivers/vdpa/mlx5/core/mlx5_vdpa.h | 10 ++++++
>>  drivers/vdpa/mlx5/core/mr.c        | 51 ++++++++++++++++++++++++++++--
>>  drivers/vdpa/mlx5/net/mlx5_vnet.c  |  3 +-
>>  3 files changed, 60 insertions(+), 4 deletions(-)
>>
>> diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>> index c3e17bc888e8..2cedf7e2dbc4 100644
>> --- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>> +++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
>> @@ -86,8 +86,18 @@ enum {
>>  struct mlx5_vdpa_mr_resources {
>>         struct mlx5_vdpa_mr *mr[MLX5_VDPA_NUM_AS];
>>         unsigned int group2asid[MLX5_VDPA_NUMVQ_GROUPS];
>> +
>> +       /* Pre-deletion mr list */
>>         struct list_head mr_list_head;
>> +
>> +       /* Deferred mr list */
>> +       struct list_head mr_gc_list_head;
>> +       struct workqueue_struct *wq_gc;
>> +       struct delayed_work gc_dwork_ent;
>> +
>>         struct mutex lock;
>> +
>> +       atomic_t shutdown;
>>  };
>>
>>  struct mlx5_vdpa_dev {
>> diff --git a/drivers/vdpa/mlx5/core/mr.c b/drivers/vdpa/mlx5/core/mr.c
>> index ec75f165f832..43fce6b39cf2 100644
>> --- a/drivers/vdpa/mlx5/core/mr.c
>> +++ b/drivers/vdpa/mlx5/core/mr.c
>> @@ -653,14 +653,46 @@ static void _mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev *mvdev, struct mlx5_vdpa_
>>         kfree(mr);
>>  }
>>
>> +#define MLX5_VDPA_MR_GC_TRIGGER_MS 2000
>> +
>> +static void mlx5_vdpa_mr_gc_handler(struct work_struct *work)
>> +{
>> +       struct mlx5_vdpa_mr_resources *mres;
>> +       struct mlx5_vdpa_mr *mr, *tmp;
>> +       struct mlx5_vdpa_dev *mvdev;
>> +
>> +       mres = container_of(work, struct mlx5_vdpa_mr_resources, gc_dwork_ent.work);
>> +
>> +       if (atomic_read(&mres->shutdown)) {
>> +               mutex_lock(&mres->lock);
>> +       } else if (!mutex_trylock(&mres->lock)) {
> 
> Is the trylock worth it? My understanding is that mutex_lock will add
> the kthread to the waitqueue anyway if it is not able to acquire the
> lock.
> 
I want to believe it is :). I noticed during testing that this can
interfere with the case where there are several .set_map() operations
in quick succession. That's why the work is delayed by such a long
time.

It's not a perfect heuristic but I found that it's better than not
having it.

>> +               queue_delayed_work(mres->wq_gc, &mres->gc_dwork_ent,
>> +                                  msecs_to_jiffies(MLX5_VDPA_MR_GC_TRIGGER_MS));
>> +               return;
>> +       }
>> +
>> +       mvdev = container_of(mres, struct mlx5_vdpa_dev, mres);
>> +
>> +       list_for_each_entry_safe(mr, tmp, &mres->mr_gc_list_head, mr_list) {
>> +               _mlx5_vdpa_destroy_mr(mvdev, mr);
>> +       }
>> +
>> +       mutex_unlock(&mres->lock);
>> +}
>> +
>>  static void _mlx5_vdpa_put_mr(struct mlx5_vdpa_dev *mvdev,
>>                               struct mlx5_vdpa_mr *mr)
>>  {
>> +       struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
>> +
>>         if (!mr)
>>                 return;
>>
>> -       if (refcount_dec_and_test(&mr->refcount))
>> -               _mlx5_vdpa_destroy_mr(mvdev, mr);
>> +       if (refcount_dec_and_test(&mr->refcount)) {
>> +               list_move_tail(&mr->mr_list, &mres->mr_gc_list_head);
>> +               queue_delayed_work(mres->wq_gc, &mres->gc_dwork_ent,
>> +                                  msecs_to_jiffies(MLX5_VDPA_MR_GC_TRIGGER_MS));
> 
> Why the delay?
> 
See above.

>> +       }
>>  }
>>
>>  void mlx5_vdpa_put_mr(struct mlx5_vdpa_dev *mvdev,
>> @@ -848,9 +880,17 @@ int mlx5_vdpa_init_mr_resources(struct mlx5_vdpa_dev *mvdev)
>>  {
>>         struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
>>
>> -       INIT_LIST_HEAD(&mres->mr_list_head);
>> +       mres->wq_gc = create_singlethread_workqueue("mlx5_vdpa_mr_gc");
>> +       if (!mres->wq_gc)
>> +               return -ENOMEM;
>> +
>> +       INIT_DELAYED_WORK(&mres->gc_dwork_ent, mlx5_vdpa_mr_gc_handler);
>> +
>>         mutex_init(&mres->lock);
>>
>> +       INIT_LIST_HEAD(&mres->mr_list_head);
>> +       INIT_LIST_HEAD(&mres->mr_gc_list_head);
>> +
>>         return 0;
>>  }
>>
>> @@ -858,5 +898,10 @@ void mlx5_vdpa_destroy_mr_resources(struct mlx5_vdpa_dev *mvdev)
>>  {
>>         struct mlx5_vdpa_mr_resources *mres = &mvdev->mres;
>>
>> +       atomic_set(&mres->shutdown, 1);
>> +
>> +       flush_delayed_work(&mres->gc_dwork_ent);
>> +       destroy_workqueue(mres->wq_gc);
>> +       mres->wq_gc = NULL;
>>         mutex_destroy(&mres->lock);
>>  }
>> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
>> index 1cadcb05a5c7..ee9482ef51e6 100644
>> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
>> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
>> @@ -3435,6 +3435,8 @@ static void mlx5_vdpa_free(struct vdpa_device *vdev)
>>         free_fixed_resources(ndev);
>>         mlx5_vdpa_clean_mrs(mvdev);
>>         mlx5_vdpa_destroy_mr_resources(&ndev->mvdev);
>> +       mlx5_cmd_cleanup_async_ctx(&mvdev->async_ctx);
>> +
>>         if (!is_zero_ether_addr(ndev->config.mac)) {
>>                 pfmdev = pci_get_drvdata(pci_physfn(mvdev->mdev->pdev));
>>                 mlx5_mpfs_del_mac(pfmdev, ndev->config.mac);
>> @@ -4044,7 +4046,6 @@ static void mlx5_vdpa_dev_del(struct vdpa_mgmt_dev *v_mdev, struct vdpa_device *
>>         destroy_workqueue(wq);
>>         mgtdev->ndev = NULL;
>>
> 
> Extra newline here.
Ack.
> 
>> -       mlx5_cmd_cleanup_async_ctx(&mvdev->async_ctx);
>>  }
>>
>>  static const struct vdpa_mgmtdev_ops mdev_ops = {
>> --
>> 2.45.1
>>
>