Re: [PATCH rdma-next 5/5] RDMA/core: Convert RWQ table logic to ib_core allocation scheme

[Leon Romanovsky <leon@kernel.org>]

On Sun, Jun 28, 2020 at 02:55:31PM +0300, Yishai Hadas wrote:
> On 6/28/2020 1:33 PM, Leon Romanovsky wrote:
> > On Sun, Jun 28, 2020 at 01:08:24PM +0300, Yishai Hadas wrote:
> > > On 6/28/2020 12:41 PM, Leon Romanovsky wrote:
> > > > On Sun, Jun 28, 2020 at 12:11:41PM +0300, Yishai Hadas wrote:
> > > > > On 6/24/2020 1:54 PM, Leon Romanovsky wrote:
> > > > > > From: Leon Romanovsky <leonro@mellanox.com>
> > > > > >
> > > > > > Move struct ib_rwq_ind_table allocation to ib_core.
> > > > > >
> > > > > > Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
> > > > > > ---
> > > > > >     drivers/infiniband/core/device.c           |  1 +
> > > > > >     drivers/infiniband/core/uverbs_cmd.c       | 37 ++++++++++++-------
> > > > > >     drivers/infiniband/core/uverbs_std_types.c | 13 ++++++-
> > > > > >     drivers/infiniband/core/verbs.c            | 25 +------------
> > > > > >     drivers/infiniband/hw/mlx4/main.c          |  4 +-
> > > > > >     drivers/infiniband/hw/mlx4/mlx4_ib.h       | 12 +++---
> > > > > >     drivers/infiniband/hw/mlx4/qp.c            | 40 ++++++--------------
> > > > > >     drivers/infiniband/hw/mlx5/main.c          |  3 ++
> > > > > >     drivers/infiniband/hw/mlx5/mlx5_ib.h       |  8 ++--
> > > > > >     drivers/infiniband/hw/mlx5/qp.c            | 43 +++++++++-------------
> > > > > >     include/rdma/ib_verbs.h                    | 11 +++---
> > > > > >     11 files changed, 86 insertions(+), 111 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c
> > > > > > index b15fa3fa09ac..85d921c8e2b5 100644
> > > > > > --- a/drivers/infiniband/core/device.c
> > > > > > +++ b/drivers/infiniband/core/device.c
> > > > > > @@ -2686,6 +2686,7 @@ void ib_set_device_ops(struct ib_device *dev, const struct ib_device_ops *ops)
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_cq);
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_mw);
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_pd);
> > > > > > +	SET_OBJ_SIZE(dev_ops, ib_rwq_ind_table);
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_srq);
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_ucontext);
> > > > > >     	SET_OBJ_SIZE(dev_ops, ib_xrcd);
> > > > > > diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
> > > > > > index a5301f3d3059..a83d11d1e3ee 100644
> > > > > > --- a/drivers/infiniband/core/uverbs_cmd.c
> > > > > > +++ b/drivers/infiniband/core/uverbs_cmd.c
> > > > > > @@ -3090,13 +3090,13 @@ static int ib_uverbs_ex_create_rwq_ind_table(struct uverbs_attr_bundle *attrs)
> > > > > >     {
> > > > > >     	struct ib_uverbs_ex_create_rwq_ind_table cmd;
> > > > > >     	struct ib_uverbs_ex_create_rwq_ind_table_resp  resp = {};
> > > > > > -	struct ib_uobject		  *uobj;
> > > > > > +	struct ib_uobject *uobj;
> > > > > >     	int err;
> > > > > >     	struct ib_rwq_ind_table_init_attr init_attr = {};
> > > > > >     	struct ib_rwq_ind_table *rwq_ind_tbl;
> > > > > > -	struct ib_wq	**wqs = NULL;
> > > > > > +	struct ib_wq **wqs = NULL;
> > > > > >     	u32 *wqs_handles = NULL;
> > > > > > -	struct ib_wq	*wq = NULL;
> > > > > > +	struct ib_wq *wq = NULL;
> > > > > >     	int i, j, num_read_wqs;
> > > > > >     	u32 num_wq_handles;
> > > > > >     	struct uverbs_req_iter iter;
> > > > > > @@ -3151,17 +3151,15 @@ static int ib_uverbs_ex_create_rwq_ind_table(struct uverbs_attr_bundle *attrs)
> > > > > >     		goto put_wqs;
> > > > > >     	}
> > > > > > -	init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
> > > > > > -	init_attr.ind_tbl = wqs;
> > > > > > -
> > > > > > -	rwq_ind_tbl = ib_dev->ops.create_rwq_ind_table(ib_dev, &init_attr,
> > > > > > -						       &attrs->driver_udata);
> > > > > > -
> > > > > > -	if (IS_ERR(rwq_ind_tbl)) {
> > > > > > -		err = PTR_ERR(rwq_ind_tbl);
> > > > > > +	rwq_ind_tbl = rdma_zalloc_drv_obj(ib_dev, ib_rwq_ind_table);
> > > > > > +	if (!rwq_ind_tbl) {
> > > > > > +		err = -ENOMEM;
> > > > > >     		goto err_uobj;
> > > > > >     	}
> > > > > > +	init_attr.log_ind_tbl_size = cmd.log_ind_tbl_size;
> > > > > > +	init_attr.ind_tbl = wqs;
> > > > > > +
> > > > > >     	rwq_ind_tbl->ind_tbl = wqs;
> > > > > >     	rwq_ind_tbl->log_ind_tbl_size = init_attr.log_ind_tbl_size;
> > > > > >     	rwq_ind_tbl->uobject = uobj;
> > > > > > @@ -3169,6 +3167,12 @@ static int ib_uverbs_ex_create_rwq_ind_table(struct uverbs_attr_bundle *attrs)
> > > > > >     	rwq_ind_tbl->device = ib_dev;
> > > > > >     	atomic_set(&rwq_ind_tbl->usecnt, 0);
> > > > > > +	err = ib_dev->ops.create_rwq_ind_table(rwq_ind_tbl, &init_attr,
> > > > > > +					       &rwq_ind_tbl->ind_tbl_num,
> > > > > > +					       &attrs->driver_udata);
> > > > > > +	if (err)
> > > > > > +		goto err_create;
> > > > > > +
> > > > > >     	for (i = 0; i < num_wq_handles; i++)
> > > > > >     		atomic_inc(&wqs[i]->usecnt);
> > > > > > @@ -3190,7 +3194,13 @@ static int ib_uverbs_ex_create_rwq_ind_table(struct uverbs_attr_bundle *attrs)
> > > > > >     	return 0;
> > > > > >     err_copy:
> > > > > > -	ib_destroy_rwq_ind_table(rwq_ind_tbl);
> > > > > > +	for (i = 0; i < num_wq_handles; i++)
> > > > > > +		atomic_dec(&wqs[i]->usecnt);
> > > > > > +
> > > > > > +	if (ib_dev->ops.destroy_rwq_ind_table)
> > > > > > +		ib_dev->ops.destroy_rwq_ind_table(rwq_ind_tbl);
> > > > > > +err_create:
> > > > > > +	kfree(rwq_ind_tbl);
> > > > > >     err_uobj:
> > > > > >     	uobj_alloc_abort(uobj, attrs);
> > > > > >     put_wqs:
> > > > > > @@ -4018,8 +4028,7 @@ const struct uapi_definition uverbs_def_write_intf[] = {
> > > > > >     			IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL,
> > > > > >     			ib_uverbs_ex_destroy_rwq_ind_table,
> > > > > >     			UAPI_DEF_WRITE_I(
> > > > > > -				struct ib_uverbs_ex_destroy_rwq_ind_table),
> > > > > > -			UAPI_DEF_METHOD_NEEDS_FN(destroy_rwq_ind_table))),
> > > > > > +				struct ib_uverbs_ex_destroy_rwq_ind_table))),
> > > > > >     	DECLARE_UVERBS_OBJECT(
> > > > > >     		UVERBS_OBJECT_WQ,
> > > > > > diff --git a/drivers/infiniband/core/uverbs_std_types.c b/drivers/infiniband/core/uverbs_std_types.c
> > > > > > index e4994cc4cc51..cd82a5a80bdf 100644
> > > > > > --- a/drivers/infiniband/core/uverbs_std_types.c
> > > > > > +++ b/drivers/infiniband/core/uverbs_std_types.c
> > > > > > @@ -82,12 +82,21 @@ static int uverbs_free_rwq_ind_tbl(struct ib_uobject *uobject,
> > > > > >     {
> > > > > >     	struct ib_rwq_ind_table *rwq_ind_tbl = uobject->object;
> > > > > >     	struct ib_wq **ind_tbl = rwq_ind_tbl->ind_tbl;
> > > > > > -	int ret;
> > > > > > +	u32 table_size = (1 << rwq_ind_tbl->log_ind_tbl_size);
> > > > > > +	int ret = 0, i;
> > > > > > +
> > > > > > +	if (atomic_read(&rwq_ind_tbl->usecnt))
> > > > > > +		ret = -EBUSY;
> > > > > > -	ret = ib_destroy_rwq_ind_table(rwq_ind_tbl);
> > > > > >     	if (ib_is_destroy_retryable(ret, why, uobject))
> > > > > >     		return ret;
> > > > > > +	for (i = 0; i < table_size; i++)
> > > > > > +		atomic_dec(&ind_tbl[i]->usecnt);
> > > > > > +
> > > > >
> > > > > Upon destroy we first expect to destroy the object that is pointing to other
> > > > > objetcs and only then descraese their ref count, this was the orignal logic
> > > > > here. (See also ib_destroy_qp_user(), ib_destroy_wq() and others).
> > > >
> > > > It is a bug, proper logic is to allocate HW object, elevate reference
> > > > counting, decrease reference counter and deallocate HW object.
> > > >
> > >
> > > I don't agree, only upon a succesfull destroy the we can go a ahead and
> > > safely decraese ref count from the pointed objects. This pevents any race
> > > from parallel desctrcution of the pointed object and consider any potenaial
> > > failure on the original object before decaresing ref counts as of in the QP
> > > case.
> >
> > There is no such thing successful destroy, from user perspective ALL
> > valid destroys should success.
>
> This is exactly the ref count purpsoe to prevent calling destoy on an object
> that may be pointed by other one, otherwise it will fail in the driver / FW.

It is hard to believe, at this point, the refcount decrease is done without
any relation to the real value, drivers doesn't check refcounts, or not supposed
too. The idea that FW will be always in sync with this value without locking
can't be true either.

More on that, it is not uobject refcount, but object usecnt which has slightly
different meaning.

>
> > Users don't know how to recover from
> > failure at this stage and they have one of two options: reboot the
> > server or leave memory/HW leak.
> >
> > There are not supposed to be parallel destructions if core code locked
> > everything properly and one of the threads decreased reference counter
> > to zero.
> >
>
> Right, this should be achived by a proer ref counting odering as done today.

I afraid that we are talking about different ref counters.

>
> > Can you please be more specific and give the scenario of your potential
> > flow?
> >
>
> The above clearly explain the motivation for current ref count usage.

Not really, at least to me.

>
> > >
> > > > Everything else should be fixed too.
> > >
> > > I beleive that this patch should be fixed and not all other places.
> > > By the way, see your previous patch from this series regarding MW that
> > > follows same logic, first dealloc_mw() and only then does atomic_dec() on
> > > the pointed PD.
> >
> > Of course, it was intended, it is completely different case and flow. PD holds
> > MW and this is why the reference counter is decreased only after MW was destroyed.
>
> PD doesn't hold MW, MW points to a PD upon its creation as of QP, MR and
> othres.

It doesn't matter as long as you understood my point that this is
different flow.

>
> >
> > >
> > > >
> > > > >
> > > > > > +	if (rwq_ind_tbl->device->ops.destroy_rwq_ind_table)
> > > > > > +		rwq_ind_tbl->device->ops.destroy_rwq_ind_table(rwq_ind_tbl);
> > > > >
> > > > > This doesn't really make sense, if no destroy function was set, the
> > > > > allocation should be prevented as well.
> > > >
> > > > Should we delete mlx4 RWQ implementation too?
> > > > After this refactoring, mlx4 won't have .destroy_rwq_ind_table.
> > > >
> > >
> > > We can expect from symetric reasons to have always alloc/destroy from
> > > drivers point of view. The IB layer will not check whether the destroy
> > > function exist but will always call it, the driver in its turn may do
> > > nothing.
> >
> > Maybe, I don't like empty functions in the drivers, it is silly.
> >
>
> We can consier that point, let's get other people feedback as well.

Sure

>
> Yishai