diff mbox series

[v2,5/7] vdpa/mlx5: cache and reuse hardware resources

Message ID 20220224155101.1991626-6-xuemingl@nvidia.com (mailing list archive)
State Superseded
Delegated to: Maxime Coquelin
Headers show
Series vdpa/mlx5: improve device shutdown time | expand

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

Xueming Li Feb. 24, 2022, 3:50 p.m. UTC
During device suspend and resume, resources are not changed normally.
When huge resources allocated to VM, like huge memory size or lots of
queues, time spent on release and recreate became significant.

To speed up, this patch reuse resoruces like VM MR and VirtQ memory if
not changed.

Signed-off-by: Xueming Li <xuemingl@nvidia.com>
---
 drivers/vdpa/mlx5/mlx5_vdpa.c       | 11 ++++-
 drivers/vdpa/mlx5/mlx5_vdpa.h       | 12 ++++-
 drivers/vdpa/mlx5/mlx5_vdpa_mem.c   | 27 ++++++++++-
 drivers/vdpa/mlx5/mlx5_vdpa_virtq.c | 73 +++++++++++++++++++++--------
 4 files changed, 99 insertions(+), 24 deletions(-)

Comments

Maxime Coquelin April 20, 2022, 3:03 p.m. UTC | #1
On 2/24/22 16:50, Xueming Li wrote:
> During device suspend and resume, resources are not changed normally.
> When huge resources allocated to VM, like huge memory size or lots of

"When huge resources were allocated"

> queues, time spent on release and recreate became significant.
> 
> To speed up, this patch reuse resoruces like VM MR and VirtQ memory if

"reuses resources"

> not changed.
> 
> Signed-off-by: Xueming Li <xuemingl@nvidia.com>
> ---
>   drivers/vdpa/mlx5/mlx5_vdpa.c       | 11 ++++-
>   drivers/vdpa/mlx5/mlx5_vdpa.h       | 12 ++++-
>   drivers/vdpa/mlx5/mlx5_vdpa_mem.c   | 27 ++++++++++-
>   drivers/vdpa/mlx5/mlx5_vdpa_virtq.c | 73 +++++++++++++++++++++--------
>   4 files changed, 99 insertions(+), 24 deletions(-)
> 

Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>

Thanks,
Maxime
Xueming Li April 25, 2022, 1:28 p.m. UTC | #2
Hi Maxime,

Thanks for the suggestion, I'll send out a new version.

Regards,
Xueming Li

> -----Original Message-----
> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> Sent: Wednesday, April 20, 2022 11:03 PM
> To: Xueming(Steven) Li <xuemingl@nvidia.com>; dev@dpdk.org
> Cc: Matan Azrad <matan@nvidia.com>; Slava Ovsiienko <viacheslavo@nvidia.com>
> Subject: Re: [PATCH v2 5/7] vdpa/mlx5: cache and reuse hardware resources
> 
> 
> 
> On 2/24/22 16:50, Xueming Li wrote:
> > During device suspend and resume, resources are not changed normally.
> > When huge resources allocated to VM, like huge memory size or lots of
> 
> "When huge resources were allocated"
> 
> > queues, time spent on release and recreate became significant.
> >
> > To speed up, this patch reuse resoruces like VM MR and VirtQ memory if
> 
> "reuses resources"
> 
> > not changed.
> >
> > Signed-off-by: Xueming Li <xuemingl@nvidia.com>
> > ---
> >   drivers/vdpa/mlx5/mlx5_vdpa.c       | 11 ++++-
> >   drivers/vdpa/mlx5/mlx5_vdpa.h       | 12 ++++-
> >   drivers/vdpa/mlx5/mlx5_vdpa_mem.c   | 27 ++++++++++-
> >   drivers/vdpa/mlx5/mlx5_vdpa_virtq.c | 73 +++++++++++++++++++++--------
> >   4 files changed, 99 insertions(+), 24 deletions(-)
> >
> 
> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> 
> Thanks,
> Maxime
Maxime Coquelin May 5, 2022, 8:01 p.m. UTC | #3
Hi Xueming,

On 4/25/22 15:28, Xueming(Steven) Li wrote:
> Hi Maxime,
> 
> Thanks for the suggestion, I'll send out a new version.

Ok, if you sent it early next week, it could be part of next week's PR.

Thanks,
Maxime

> Regards,
> Xueming Li
> 
>> -----Original Message-----
>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>> Sent: Wednesday, April 20, 2022 11:03 PM
>> To: Xueming(Steven) Li <xuemingl@nvidia.com>; dev@dpdk.org
>> Cc: Matan Azrad <matan@nvidia.com>; Slava Ovsiienko <viacheslavo@nvidia.com>
>> Subject: Re: [PATCH v2 5/7] vdpa/mlx5: cache and reuse hardware resources
>>
>>
>>
>> On 2/24/22 16:50, Xueming Li wrote:
>>> During device suspend and resume, resources are not changed normally.
>>> When huge resources allocated to VM, like huge memory size or lots of
>>
>> "When huge resources were allocated"
>>
>>> queues, time spent on release and recreate became significant.
>>>
>>> To speed up, this patch reuse resoruces like VM MR and VirtQ memory if
>>
>> "reuses resources"
>>
>>> not changed.
>>>
>>> Signed-off-by: Xueming Li <xuemingl@nvidia.com>
>>> ---
>>>    drivers/vdpa/mlx5/mlx5_vdpa.c       | 11 ++++-
>>>    drivers/vdpa/mlx5/mlx5_vdpa.h       | 12 ++++-
>>>    drivers/vdpa/mlx5/mlx5_vdpa_mem.c   | 27 ++++++++++-
>>>    drivers/vdpa/mlx5/mlx5_vdpa_virtq.c | 73 +++++++++++++++++++++--------
>>>    4 files changed, 99 insertions(+), 24 deletions(-)
>>>
>>
>> Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>>
>> Thanks,
>> Maxime
>
diff mbox series

Patch

diff --git a/drivers/vdpa/mlx5/mlx5_vdpa.c b/drivers/vdpa/mlx5/mlx5_vdpa.c
index 9862141497b..38ed45f95f7 100644
--- a/drivers/vdpa/mlx5/mlx5_vdpa.c
+++ b/drivers/vdpa/mlx5/mlx5_vdpa.c
@@ -241,6 +241,13 @@  mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 	return kern_mtu == vhost_mtu ? 0 : -1;
 }
 
+static void
+mlx5_vdpa_dev_cache_clean(struct mlx5_vdpa_priv *priv)
+{
+	mlx5_vdpa_virtqs_cleanup(priv);
+	mlx5_vdpa_mem_dereg(priv);
+}
+
 static int
 mlx5_vdpa_dev_close(int vid)
 {
@@ -260,7 +267,8 @@  mlx5_vdpa_dev_close(int vid)
 	}
 	mlx5_vdpa_steer_unset(priv);
 	mlx5_vdpa_virtqs_release(priv);
-	mlx5_vdpa_mem_dereg(priv);
+	if (priv->lm_mr.addr)
+		mlx5_os_wrapped_mkey_destroy(&priv->lm_mr);
 	priv->state = MLX5_VDPA_STATE_PROBED;
 	priv->vid = 0;
 	/* The mutex may stay locked after event thread cancel - initiate it. */
@@ -659,6 +667,7 @@  mlx5_vdpa_release_dev_resources(struct mlx5_vdpa_priv *priv)
 {
 	uint32_t i;
 
+	mlx5_vdpa_dev_cache_clean(priv);
 	mlx5_vdpa_event_qp_global_release(priv);
 	mlx5_vdpa_err_event_unset(priv);
 	if (priv->steer.tbl)
diff --git a/drivers/vdpa/mlx5/mlx5_vdpa.h b/drivers/vdpa/mlx5/mlx5_vdpa.h
index e0ba20b953c..540bf87a352 100644
--- a/drivers/vdpa/mlx5/mlx5_vdpa.h
+++ b/drivers/vdpa/mlx5/mlx5_vdpa.h
@@ -289,13 +289,21 @@  int mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv);
 void mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv);
 
 /**
- * Release a virtq and all its related resources.
+ * Release virtqs and resources except that to be reused.
  *
  * @param[in] priv
  *   The vdpa driver private structure.
  */
 void mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv);
 
+/**
+ * Cleanup cached resources of all virtqs.
+ *
+ * @param[in] priv
+ *   The vdpa driver private structure.
+ */
+void mlx5_vdpa_virtqs_cleanup(struct mlx5_vdpa_priv *priv);
+
 /**
  * Create all the HW virtqs resources and all their related resources.
  *
@@ -323,7 +331,7 @@  int mlx5_vdpa_virtqs_prepare(struct mlx5_vdpa_priv *priv);
 int mlx5_vdpa_virtq_enable(struct mlx5_vdpa_priv *priv, int index, int enable);
 
 /**
- * Unset steering and release all its related resources- stop traffic.
+ * Unset steering - stop traffic.
  *
  * @param[in] priv
  *   The vdpa driver private structure.
diff --git a/drivers/vdpa/mlx5/mlx5_vdpa_mem.c b/drivers/vdpa/mlx5/mlx5_vdpa_mem.c
index 62f5530e91d..d6e3dd664b5 100644
--- a/drivers/vdpa/mlx5/mlx5_vdpa_mem.c
+++ b/drivers/vdpa/mlx5/mlx5_vdpa_mem.c
@@ -32,8 +32,6 @@  mlx5_vdpa_mem_dereg(struct mlx5_vdpa_priv *priv)
 		entry = next;
 	}
 	SLIST_INIT(&priv->mr_list);
-	if (priv->lm_mr.addr)
-		mlx5_os_wrapped_mkey_destroy(&priv->lm_mr);
 	if (priv->vmem) {
 		free(priv->vmem);
 		priv->vmem = NULL;
@@ -149,6 +147,23 @@  mlx5_vdpa_vhost_mem_regions_prepare(int vid, uint8_t *mode, uint64_t *mem_size,
 	return mem;
 }
 
+static int
+mlx5_vdpa_mem_cmp(struct rte_vhost_memory *mem1, struct rte_vhost_memory *mem2)
+{
+	uint32_t i;
+
+	if (mem1->nregions != mem2->nregions)
+		return -1;
+	for (i = 0; i < mem1->nregions; i++) {
+		if (mem1->regions[i].guest_phys_addr !=
+		    mem2->regions[i].guest_phys_addr)
+			return -1;
+		if (mem1->regions[i].size != mem2->regions[i].size)
+			return -1;
+	}
+	return 0;
+}
+
 #define KLM_SIZE_MAX_ALIGN(sz) ((sz) > MLX5_MAX_KLM_BYTE_COUNT ? \
 				MLX5_MAX_KLM_BYTE_COUNT : (sz))
 
@@ -191,6 +206,14 @@  mlx5_vdpa_mem_register(struct mlx5_vdpa_priv *priv)
 
 	if (!mem)
 		return -rte_errno;
+	if (priv->vmem != NULL) {
+		if (mlx5_vdpa_mem_cmp(mem, priv->vmem) == 0) {
+			/* VM memory not changed, reuse resources. */
+			free(mem);
+			return 0;
+		}
+		mlx5_vdpa_mem_dereg(priv);
+	}
 	priv->vmem = mem;
 	for (i = 0; i < mem->nregions; i++) {
 		reg = &mem->regions[i];
diff --git a/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c b/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c
index 6bda9f1814a..c42846ecb3c 100644
--- a/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c
+++ b/drivers/vdpa/mlx5/mlx5_vdpa_virtq.c
@@ -66,10 +66,33 @@  mlx5_vdpa_virtq_kick_handler(void *cb_arg)
 	DRV_LOG(DEBUG, "Ring virtq %u doorbell.", virtq->index);
 }
 
+/* Release cached VQ resources. */
+void
+mlx5_vdpa_virtqs_cleanup(struct mlx5_vdpa_priv *priv)
+{
+	unsigned int i, j;
+
+	for (i = 0; i < priv->caps.max_num_virtio_queues * 2; i++) {
+		struct mlx5_vdpa_virtq *virtq = &priv->virtqs[i];
+
+		for (j = 0; j < RTE_DIM(virtq->umems); ++j) {
+			if (virtq->umems[j].obj) {
+				claim_zero(mlx5_glue->devx_umem_dereg
+							(virtq->umems[j].obj));
+				virtq->umems[j].obj = NULL;
+			}
+			if (virtq->umems[j].buf) {
+				rte_free(virtq->umems[j].buf);
+				virtq->umems[j].buf = NULL;
+			}
+			virtq->umems[j].size = 0;
+		}
+	}
+}
+
 static int
 mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
 {
-	unsigned int i;
 	int ret = -EAGAIN;
 
 	if (rte_intr_fd_get(virtq->intr_handle) >= 0) {
@@ -94,13 +117,6 @@  mlx5_vdpa_virtq_unset(struct mlx5_vdpa_virtq *virtq)
 		claim_zero(mlx5_devx_cmd_destroy(virtq->virtq));
 	}
 	virtq->virtq = NULL;
-	for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
-		if (virtq->umems[i].obj)
-			claim_zero(mlx5_glue->devx_umem_dereg
-							 (virtq->umems[i].obj));
-		rte_free(virtq->umems[i].buf);
-	}
-	memset(&virtq->umems, 0, sizeof(virtq->umems));
 	if (virtq->eqp.fw_qp)
 		mlx5_vdpa_event_qp_destroy(&virtq->eqp);
 	virtq->notifier_state = MLX5_VDPA_NOTIFIER_STATE_DISABLED;
@@ -120,7 +136,6 @@  mlx5_vdpa_virtqs_release(struct mlx5_vdpa_priv *priv)
 			claim_zero(mlx5_devx_cmd_destroy(virtq->counters));
 	}
 	priv->features = 0;
-	memset(priv->virtqs, 0, sizeof(*virtq) * priv->nr_virtqs);
 	priv->nr_virtqs = 0;
 }
 
@@ -215,6 +230,8 @@  mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv, int index)
 	ret = rte_vhost_get_vhost_vring(priv->vid, index, &vq);
 	if (ret)
 		return -1;
+	if (vq.size == 0)
+		return 0;
 	virtq->index = index;
 	virtq->vq_size = vq.size;
 	attr.tso_ipv4 = !!(priv->features & (1ULL << VIRTIO_NET_F_HOST_TSO4));
@@ -259,24 +276,42 @@  mlx5_vdpa_virtq_setup(struct mlx5_vdpa_priv *priv, int index)
 	}
 	/* Setup 3 UMEMs for each virtq. */
 	for (i = 0; i < RTE_DIM(virtq->umems); ++i) {
-		virtq->umems[i].size = priv->caps.umems[i].a * vq.size +
-							  priv->caps.umems[i].b;
-		virtq->umems[i].buf = rte_zmalloc(__func__,
-						  virtq->umems[i].size, 4096);
-		if (!virtq->umems[i].buf) {
+		uint32_t size;
+		void *buf;
+		struct mlx5dv_devx_umem *obj;
+
+		size = priv->caps.umems[i].a * vq.size + priv->caps.umems[i].b;
+		if (virtq->umems[i].size == size &&
+		    virtq->umems[i].obj != NULL) {
+			/* Reuse registered memory. */
+			memset(virtq->umems[i].buf, 0, size);
+			goto reuse;
+		}
+		if (virtq->umems[i].obj)
+			claim_zero(mlx5_glue->devx_umem_dereg
+				   (virtq->umems[i].obj));
+		if (virtq->umems[i].buf)
+			rte_free(virtq->umems[i].buf);
+		virtq->umems[i].size = 0;
+		virtq->umems[i].obj = NULL;
+		virtq->umems[i].buf = NULL;
+		buf = rte_zmalloc(__func__, size, 4096);
+		if (buf == NULL) {
 			DRV_LOG(ERR, "Cannot allocate umem %d memory for virtq"
 				" %u.", i, index);
 			goto error;
 		}
-		virtq->umems[i].obj = mlx5_glue->devx_umem_reg(priv->cdev->ctx,
-							virtq->umems[i].buf,
-							virtq->umems[i].size,
-							IBV_ACCESS_LOCAL_WRITE);
-		if (!virtq->umems[i].obj) {
+		obj = mlx5_glue->devx_umem_reg(priv->cdev->ctx, buf, size,
+					       IBV_ACCESS_LOCAL_WRITE);
+		if (obj == NULL) {
 			DRV_LOG(ERR, "Failed to register umem %d for virtq %u.",
 				i, index);
 			goto error;
 		}
+		virtq->umems[i].size = size;
+		virtq->umems[i].buf = buf;
+		virtq->umems[i].obj = obj;
+reuse:
 		attr.umems[i].id = virtq->umems[i].obj->umem_id;
 		attr.umems[i].offset = 0;
 		attr.umems[i].size = virtq->umems[i].size;