[PATCH for-next v3 0/2] RDMA/rxe: Prefetching pages with explicit ODP

[PATCH for-next v3 0/2] RDMA/rxe: Prefetching pages with explicit ODP

[Daisuke Matsuda]

There is ibv_advise_mr(3) that can be used by applications to optimize
memory access. This series enables the feature on rxe driver, which has
already been available in mlx5.

There is a tiny change on the rdma-core util.
cf. https://github.com/linux-rdma/rdma-core/pull/1605

Daisuke Matsuda (2):
  RDMA/rxe: Implement synchronous prefetch for ODP MRs
  RDMA/rxe: Enable asynchronous prefetch for ODP MRs

 drivers/infiniband/sw/rxe/rxe.c     |   7 ++
 drivers/infiniband/sw/rxe/rxe_loc.h |  10 ++
 drivers/infiniband/sw/rxe/rxe_odp.c | 162 ++++++++++++++++++++++++++++
 3 files changed, 179 insertions(+)

-- 
2.43.0

[PATCH for-next v3 1/2] RDMA/rxe: Implement synchronous prefetch for ODP MRs

[Daisuke Matsuda]

Minimal implementation of ibv_advise_mr(3) requires synchronous calls being
successful with the IBV_ADVISE_MR_FLAG_FLUSH flag. Asynchronous requests,
which are best-effort, will be supported subsequently.

Signed-off-by: Daisuke Matsuda <dskmtsd@gmail.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
---
 drivers/infiniband/sw/rxe/rxe.c     |  7 +++
 drivers/infiniband/sw/rxe/rxe_loc.h | 10 ++++
 drivers/infiniband/sw/rxe/rxe_odp.c | 82 +++++++++++++++++++++++++++++
 3 files changed, 99 insertions(+)

diff --git a/drivers/infiniband/sw/rxe/rxe.c b/drivers/infiniband/sw/rxe/rxe.c
index 3a77d6db1720..e891199cbdef 100644
--- a/drivers/infiniband/sw/rxe/rxe.c
+++ b/drivers/infiniband/sw/rxe/rxe.c
@@ -34,6 +34,10 @@ void rxe_dealloc(struct ib_device *ib_dev)
 	mutex_destroy(&rxe->usdev_lock);
 }
 
+static const struct ib_device_ops rxe_ib_dev_odp_ops = {
+	.advise_mr = rxe_ib_advise_mr,
+};
+
 /* initialize rxe device parameters */
 static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev)
 {
@@ -103,6 +107,9 @@ static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev)
 		rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
 		rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH;
 		rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE;
+
+		/* set handler for ODP prefetching API - ibv_advise_mr(3) */
+		ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops);
 	}
 }
 
diff --git a/drivers/infiniband/sw/rxe/rxe_loc.h b/drivers/infiniband/sw/rxe/rxe_loc.h
index f7dbb9cddd12..21b070f3dbb8 100644
--- a/drivers/infiniband/sw/rxe/rxe_loc.h
+++ b/drivers/infiniband/sw/rxe/rxe_loc.h
@@ -197,6 +197,9 @@ enum resp_states rxe_odp_atomic_op(struct rxe_mr *mr, u64 iova, int opcode,
 int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova,
 			    unsigned int length);
 enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value);
+int rxe_ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
+		     u32 flags, struct ib_sge *sg_list, u32 num_sge,
+		     struct uverbs_attr_bundle *attrs);
 #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 static inline int
 rxe_odp_mr_init_user(struct rxe_dev *rxe, u64 start, u64 length, u64 iova,
@@ -225,6 +228,13 @@ static inline enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr,
 {
 	return RESPST_ERR_UNSUPPORTED_OPCODE;
 }
+static inline int rxe_ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
+				   u32 flags, struct ib_sge *sg_list, u32 num_sge,
+				   struct uverbs_attr_bundle *attrs)
+{
+	return -EOPNOTSUPP;
+}
+
 #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 
 #endif /* RXE_LOC_H */
diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
index 6149d9ffe7f7..4c98a02d572c 100644
--- a/drivers/infiniband/sw/rxe/rxe_odp.c
+++ b/drivers/infiniband/sw/rxe/rxe_odp.c
@@ -424,3 +424,85 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
 
 	return RESPST_NONE;
 }
+
+static int rxe_ib_prefetch_sg_list(struct ib_pd *ibpd,
+				   enum ib_uverbs_advise_mr_advice advice,
+				   u32 pf_flags, struct ib_sge *sg_list,
+				   u32 num_sge)
+{
+	struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd);
+	int ret = 0;
+	u32 i;
+
+	for (i = 0; i < num_sge; ++i) {
+		struct rxe_mr *mr;
+		struct ib_umem_odp *umem_odp;
+
+		mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE,
+			       sg_list[i].lkey, RXE_LOOKUP_LOCAL);
+
+		if (IS_ERR(mr)) {
+			rxe_dbg_pd(pd, "mr with lkey %x not found\n", sg_list[i].lkey);
+			return PTR_ERR(mr);
+		}
+
+		if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
+		    !mr->umem->writable) {
+			rxe_dbg_mr(mr, "missing write permission\n");
+			rxe_put(mr);
+			return -EPERM;
+		}
+
+		ret = rxe_odp_do_pagefault_and_lock(mr, sg_list[i].addr,
+						    sg_list[i].length, pf_flags);
+		if (ret < 0) {
+			rxe_dbg_mr(mr, "failed to prefetch the mr\n");
+			rxe_put(mr);
+			return ret;
+		}
+
+		umem_odp = to_ib_umem_odp(mr->umem);
+		mutex_unlock(&umem_odp->umem_mutex);
+
+		rxe_put(mr);
+	}
+
+	return 0;
+}
+
+static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,
+				     enum ib_uverbs_advise_mr_advice advice,
+				     u32 flags, struct ib_sge *sg_list, u32 num_sge)
+{
+	u32 pf_flags = RXE_PAGEFAULT_DEFAULT;
+
+	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
+		pf_flags |= RXE_PAGEFAULT_RDONLY;
+
+	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
+		pf_flags |= RXE_PAGEFAULT_SNAPSHOT;
+
+	/* Synchronous call */
+	if (flags & IB_UVERBS_ADVISE_MR_FLAG_FLUSH)
+		return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list,
+					       num_sge);
+
+	/* Asynchronous call is to be added in the next patch */
+	return -EOPNOTSUPP;
+}
+
+int rxe_ib_advise_mr(struct ib_pd *ibpd,
+		     enum ib_uverbs_advise_mr_advice advice,
+		     u32 flags,
+		     struct ib_sge *sg_list,
+		     u32 num_sge,
+		     struct uverbs_attr_bundle *attrs)
+{
+	if (advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH &&
+	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE &&
+	    advice != IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_NO_FAULT)
+		return -EOPNOTSUPP;
+
+	return rxe_ib_advise_mr_prefetch(ibpd, advice, flags,
+					 sg_list, num_sge);
+}
-- 
2.43.0

[PATCH for-next v3 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs

[Daisuke Matsuda]

Calling ibv_advise_mr(3) with flags other than IBV_ADVISE_MR_FLAG_FLUSH
invokes asynchronous request. It is best-effort, and thus can safely be
deferred to the system-wide workqueue.

Signed-off-by: Daisuke Matsuda <dskmtsd@gmail.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
---
 drivers/infiniband/sw/rxe/rxe_odp.c | 84 ++++++++++++++++++++++++++++-
 1 file changed, 82 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
index 4c98a02d572c..0f3b281a265f 100644
--- a/drivers/infiniband/sw/rxe/rxe_odp.c
+++ b/drivers/infiniband/sw/rxe/rxe_odp.c
@@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
 	return RESPST_NONE;
 }
 
+struct prefetch_mr_work {
+	struct work_struct work;
+	u32 pf_flags;
+	u32 num_sge;
+	struct {
+		u64 io_virt;
+		struct rxe_mr *mr;
+		size_t length;
+	} frags[];
+};
+
+static void rxe_ib_prefetch_mr_work(struct work_struct *w)
+{
+	struct prefetch_mr_work *work =
+		container_of(w, struct prefetch_mr_work, work);
+	int ret;
+	u32 i;
+
+	/* We rely on IB/core that work is executed if we have num_sge != 0 only. */
+	WARN_ON(!work->num_sge);
+	for (i = 0; i < work->num_sge; ++i) {
+		struct ib_umem_odp *umem_odp;
+
+		ret = rxe_odp_do_pagefault_and_lock(work->frags[i].mr, work->frags[i].io_virt,
+						    work->frags[i].length, work->pf_flags);
+		if (ret < 0) {
+			rxe_dbg_mr(work->frags[i].mr, "failed to prefetch the mr\n");
+			continue;
+		}
+
+		umem_odp = to_ib_umem_odp(work->frags[i].mr->umem);
+		mutex_unlock(&umem_odp->umem_mutex);
+	}
+
+	kvfree(work);
+}
+
+static int rxe_init_prefetch_work(struct ib_pd *ibpd,
+				  enum ib_uverbs_advise_mr_advice advice,
+				  u32 pf_flags, struct prefetch_mr_work *work,
+				  struct ib_sge *sg_list, u32 num_sge)
+{
+	struct rxe_pd *pd = container_of(ibpd, struct rxe_pd, ibpd);
+	u32 i;
+
+	INIT_WORK(&work->work, rxe_ib_prefetch_mr_work);
+	work->pf_flags = pf_flags;
+
+	for (i = 0; i < num_sge; ++i) {
+		struct rxe_mr *mr;
+
+		mr = lookup_mr(pd, IB_ACCESS_LOCAL_WRITE,
+			       sg_list[i].lkey, RXE_LOOKUP_LOCAL);
+		if (IS_ERR(mr)) {
+			work->num_sge = i;
+			return PTR_ERR(mr);
+		}
+		work->frags[i].io_virt = sg_list[i].addr;
+		work->frags[i].length = sg_list[i].length;
+		work->frags[i].mr = mr;
+
+		rxe_put(mr);
+	}
+	work->num_sge = num_sge;
+	return 0;
+}
+
 static int rxe_ib_prefetch_sg_list(struct ib_pd *ibpd,
 				   enum ib_uverbs_advise_mr_advice advice,
 				   u32 pf_flags, struct ib_sge *sg_list,
@@ -475,6 +542,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,
 				     u32 flags, struct ib_sge *sg_list, u32 num_sge)
 {
 	u32 pf_flags = RXE_PAGEFAULT_DEFAULT;
+	struct prefetch_mr_work *work;
+	int rc;
 
 	if (advice == IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH)
 		pf_flags |= RXE_PAGEFAULT_RDONLY;
@@ -487,8 +556,19 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,
 		return rxe_ib_prefetch_sg_list(ibpd, advice, pf_flags, sg_list,
 					       num_sge);
 
-	/* Asynchronous call is to be added in the next patch */
-	return -EOPNOTSUPP;
+	/* Asynchronous call is "best-effort" and allowed to fail */
+	work = kvzalloc(struct_size(work, frags, num_sge), GFP_KERNEL);
+	if (!work)
+		return -ENOMEM;
+
+	rc = rxe_init_prefetch_work(ibpd, advice, pf_flags, work, sg_list, num_sge);
+	if (rc) {
+		kvfree(work);
+		return rc;
+	}
+	queue_work(system_unbound_wq, &work->work);
+
+	return 0;
 }
 
 int rxe_ib_advise_mr(struct ib_pd *ibpd,
-- 
2.43.0

Re: [PATCH for-next v3 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs

[Leon Romanovsky]

On Tue, May 13, 2025 at 05:04:05AM +0000, Daisuke Matsuda wrote:
> Calling ibv_advise_mr(3) with flags other than IBV_ADVISE_MR_FLAG_FLUSH
> invokes asynchronous request. It is best-effort, and thus can safely be
> deferred to the system-wide workqueue.
> 
> Signed-off-by: Daisuke Matsuda <dskmtsd@gmail.com>
> Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
> ---
>  drivers/infiniband/sw/rxe/rxe_odp.c | 84 ++++++++++++++++++++++++++++-
>  1 file changed, 82 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
> index 4c98a02d572c..0f3b281a265f 100644
> --- a/drivers/infiniband/sw/rxe/rxe_odp.c
> +++ b/drivers/infiniband/sw/rxe/rxe_odp.c
> @@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
>  	return RESPST_NONE;
>  }

<...>

> +static int rxe_init_prefetch_work(struct ib_pd *ibpd,
> +				  enum ib_uverbs_advise_mr_advice advice,
> +				  u32 pf_flags, struct prefetch_mr_work *work,
> +				  struct ib_sge *sg_list, u32 num_sge)

There is no need one-time called function. It can be embedded into rxe_ib_advise_mr_prefetch().

> +{

<...>

> @@ -475,6 +542,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,
>  				     u32 flags, struct ib_sge *sg_list, u32 num_sge)
>  {

<...>

> +	queue_work(system_unbound_wq, &work->work);

How do you ensure that this work isn't running after RXE is destroyed?

Thanks

> +
> +	return 0;
>  }
>  
>  int rxe_ib_advise_mr(struct ib_pd *ibpd,
> -- 
> 2.43.0
> 

Re: [PATCH for-next v3 2/2] RDMA/rxe: Enable asynchronous prefetch for ODP MRs

[Daisuke Matsuda]

On 2025/05/18 14:54, Leon Romanovsky wrote:
> On Tue, May 13, 2025 at 05:04:05AM +0000, Daisuke Matsuda wrote:
>> Calling ibv_advise_mr(3) with flags other than IBV_ADVISE_MR_FLAG_FLUSH
>> invokes asynchronous request. It is best-effort, and thus can safely be
>> deferred to the system-wide workqueue.
>>
>> Signed-off-by: Daisuke Matsuda <dskmtsd@gmail.com>
>> Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
>> ---
>>   drivers/infiniband/sw/rxe/rxe_odp.c | 84 ++++++++++++++++++++++++++++-
>>   1 file changed, 82 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
>> index 4c98a02d572c..0f3b281a265f 100644
>> --- a/drivers/infiniband/sw/rxe/rxe_odp.c
>> +++ b/drivers/infiniband/sw/rxe/rxe_odp.c
>> @@ -425,6 +425,73 @@ enum resp_states rxe_odp_do_atomic_write(struct rxe_mr *mr, u64 iova, u64 value)
>>   	return RESPST_NONE;
>>   }
> 
> <...>
> 
>> +static int rxe_init_prefetch_work(struct ib_pd *ibpd,
>> +				  enum ib_uverbs_advise_mr_advice advice,
>> +				  u32 pf_flags, struct prefetch_mr_work *work,
>> +				  struct ib_sge *sg_list, u32 num_sge)
> 
> There is no need one-time called function. It can be embedded into rxe_ib_advise_mr_prefetch().

Certainly.

> 
>> +{
> 
> <...>
> 
>> @@ -475,6 +542,8 @@ static int rxe_ib_advise_mr_prefetch(struct ib_pd *ibpd,
>>   				     u32 flags, struct ib_sge *sg_list, u32 num_sge)
>>   {
> 
> <...>
> 
>> +	queue_work(system_unbound_wq, &work->work);
> 
> How do you ensure that this work isn't running after RXE is destroyed?

I think we can use per-pd reference counter in struct rxe_pd.
I will fix it in v4.

Thanks,
Daisuke

> 
> Thanks
> 
>> +
>> +	return 0;
>>   }
>>   
>>   int rxe_ib_advise_mr(struct ib_pd *ibpd,
>> -- 
>> 2.43.0
>>