Re: [PATCH 2/5] net: qrtr: Implement outgoing flow control

From: Chris Lew
Date: Mon May 13 2019 - 17:28:16 EST



On 5/7/2019 11:06 PM, Bjorn Andersson wrote:
In order to prevent overconsumption of resources on the remote side QRTR
implements a flow control mechanism.

The mechanism works by the sender keeping track of the number of
outstanding unconfirmed messages that has been transmitted to a
particular node/port pair.

Upon count reaching a low watermark (L) the confirm_rx bit is set in the
outgoing message and when the count reaching a high watermark (H)
transmission will be blocked upon the reception of a resume_tx message
from the remote, that resets the counter to 0.

This guarantees that there will be at most 2H - L messages in flight.
Values chosen for L and H are 5 and 10 respectively.

Signed-off-by: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx>
---
net/qrtr/qrtr.c | 143 +++++++++++++++++++++++++++++++++++++++++++++---
1 file changed, 136 insertions(+), 7 deletions(-)

diff --git a/net/qrtr/qrtr.c b/net/qrtr/qrtr.c
index 07a35362fba2..62abd622618d 100644
--- a/net/qrtr/qrtr.c
+++ b/net/qrtr/qrtr.c
@@ -16,6 +16,7 @@
#include <linux/qrtr.h>
#include <linux/termios.h> /* For TIOCINQ/OUTQ */
#include <linux/numa.h>
+#include <linux/wait.h>
#include <net/sock.h>
@@ -121,6 +122,9 @@ static DEFINE_MUTEX(qrtr_port_lock);
* @ep: endpoint
* @ref: reference count for node
* @nid: node id
+ * @qrtr_tx_flow: tree with tx counts per flow
+ * @resume_tx: waiters for a resume tx from the remote
+ * @qrtr_tx_lock: lock for qrtr_tx_flow
* @rx_queue: receive queue
* @work: scheduled work struct for recv work
* @item: list item for broadcast list
@@ -131,11 +135,26 @@ struct qrtr_node {
struct kref ref;
unsigned int nid;
+ struct radix_tree_root qrtr_tx_flow;
+ struct wait_queue_head resume_tx;
+ struct mutex qrtr_tx_lock; /* for qrtr_tx_flow */
+
struct sk_buff_head rx_queue;
struct work_struct work;
struct list_head item;
};
+/**
+ * struct qrtr_tx_flow - tx flow control
+ * @pending: number of waiting senders
+ */
+struct qrtr_tx_flow {
+ atomic_t pending;
+};
+
+#define QRTR_TX_FLOW_HIGH 10
+#define QRTR_TX_FLOW_LOW 5
+
static int qrtr_local_enqueue(struct qrtr_node *node, struct sk_buff *skb,
int type, struct sockaddr_qrtr *from,
struct sockaddr_qrtr *to);
@@ -150,7 +169,9 @@ static int qrtr_bcast_enqueue(struct qrtr_node *node, struct sk_buff *skb,
*/
static void __qrtr_node_release(struct kref *kref)
{
+ struct radix_tree_iter iter;
struct qrtr_node *node = container_of(kref, struct qrtr_node, ref);
+ void __rcu **slot;
if (node->nid != QRTR_EP_NID_AUTO)
radix_tree_delete(&qrtr_nodes, node->nid);
@@ -158,6 +179,12 @@ static void __qrtr_node_release(struct kref *kref)
list_del(&node->item);
mutex_unlock(&qrtr_node_lock);
+ /* Free tx flow counters */
+ radix_tree_for_each_slot(slot, &node->qrtr_tx_flow, &iter, 0) {
+ radix_tree_iter_delete(&node->qrtr_tx_flow, &iter, slot);
+ kfree(*slot);
+ }
+
skb_queue_purge(&node->rx_queue);
kfree(node);
}
@@ -178,15 +205,106 @@ static void qrtr_node_release(struct qrtr_node *node)
kref_put_mutex(&node->ref, __qrtr_node_release, &qrtr_node_lock);
}
+/**
+ * qrtr_tx_resume() - reset flow control counter
+ * @node: qrtr_node that the QRTR_TYPE_RESUME_TX packet arrived on
+ * @skb: resume_tx packet
+ */
+static void qrtr_tx_resume(struct qrtr_node *node, struct sk_buff *skb)
+{
+ struct qrtr_ctrl_pkt *pkt = (struct qrtr_ctrl_pkt *)skb->data;
+ struct qrtr_tx_flow *flow;
+ unsigned long key;
+ u64 remote_node = le32_to_cpu(pkt->client.node);
+ u32 remote_port = le32_to_cpu(pkt->client.port);
+
+ key = remote_node << 32 | remote_port;
+
+ flow = radix_tree_lookup(&node->qrtr_tx_flow, key);
+ if (flow)
+ atomic_set(&flow->pending, 0);
+
+ wake_up_interruptible_all(&node->resume_tx);
+
+ consume_skb(skb);
+}
+
+/**
+ * qrtr_tx_wait() - flow control for outgoing packets
+ * @node: qrtr_node that the packet is to be send to
+ * @dest_node: node id of the destination
+ * @dest_port: port number of the destination
+ * @type: type of message
+ *
+ * The flow control scheme is based around the low and high "watermarks". When
+ * the low watermark is passed the confirm_rx flag is set on the outgoing
+ * message, which will trigger the remote to send a control message of the type
+ * QRTR_TYPE_RESUME_TX to reset the counter. If the high watermark is hit
+ * further transmision should be paused.
+ *
+ * Return: 1 if confirm_rx should be set, 0 otherwise or errno failure
+ */
+static int qrtr_tx_wait(struct qrtr_node *node, int dest_node, int dest_port,
+ int type)
+{
+ struct qrtr_tx_flow *flow;
+ unsigned long key = (u64)dest_node << 32 | dest_port;
+ int confirm_rx = 0;
+ int ret;
+
+ /* Never set confirm_rx on non-data packets */
+ if (type != QRTR_TYPE_DATA)
+ return 0;
+
+ mutex_lock(&node->qrtr_tx_lock);
+ flow = radix_tree_lookup(&node->qrtr_tx_flow, key);
+ if (!flow) {
+ flow = kzalloc(sizeof(*flow), GFP_KERNEL);
+ if (!flow)
+ confirm_rx = 1;
+ else
+ radix_tree_insert(&node->qrtr_tx_flow, key, flow);
+ }
+ mutex_unlock(&node->qrtr_tx_lock);
+
+ for (;;) {
+ ret = wait_event_interruptible(node->resume_tx,
+ atomic_read(&flow->pending) < QRTR_TX_FLOW_HIGH ||
+ !node->ep);
+ if (ret)
+ return ret;
+
+ if (!node->ep)
+ return -EPIPE;
+
+ mutex_lock(&node->qrtr_tx_lock);
+ if (atomic_read(&flow->pending) < QRTR_TX_FLOW_HIGH) {
+ confirm_rx = atomic_inc_return(&flow->pending) == QRTR_TX_FLOW_LOW;
+ mutex_unlock(&node->qrtr_tx_lock);
+ break;
+ }
+ mutex_unlock(&node->qrtr_tx_lock);
+ }
+
+ return confirm_rx;
+}
+
/* Pass an outgoing packet socket buffer to the endpoint driver. */
static int qrtr_node_enqueue(struct qrtr_node *node, struct sk_buff *skb,
int type, struct sockaddr_qrtr *from,
struct sockaddr_qrtr *to)
{
struct qrtr_hdr_v1 *hdr;
+ int confirm_rx;
size_t len = skb->len;
int rc = -ENODEV;
+ confirm_rx = qrtr_tx_wait(node, to->sq_node, to->sq_port, type);
+ if (confirm_rx < 0) {
+ kfree_skb(skb);
+ return confirm_rx;
+ }
+
hdr = skb_push(skb, sizeof(*hdr));
hdr->version = cpu_to_le32(QRTR_PROTO_VER_1);
hdr->type = cpu_to_le32(type);
@@ -201,7 +319,7 @@ static int qrtr_node_enqueue(struct qrtr_node *node, struct sk_buff *skb,
}
hdr->size = cpu_to_le32(len);
- hdr->confirm_rx = 0;
+ hdr->confirm_rx = !!confirm_rx;
skb_put_padto(skb, ALIGN(len, 4));

We had issues where the underlying transport layer failed to send the packet but flow control count wasn't adjusted. Eventually we couldn't
send to that remote port if the packet with the control flag bit was
dropped by the transport.

@@ -318,7 +436,8 @@ int qrtr_endpoint_post(struct qrtr_endpoint *ep, const void *data, size_t len)
if (len != ALIGN(size, 4) + hdrlen)
goto err;
- if (cb->dst_port != QRTR_PORT_CTRL && cb->type != QRTR_TYPE_DATA)
+ if (cb->dst_port != QRTR_PORT_CTRL && cb->type != QRTR_TYPE_DATA &&
+ cb->type != QRTR_TYPE_RESUME_TX)
goto err;
skb_put_data(skb, data + hdrlen, size);
@@ -377,14 +496,18 @@ static void qrtr_node_rx_work(struct work_struct *work)
qrtr_node_assign(node, cb->src_node);
- ipc = qrtr_port_lookup(cb->dst_port);
- if (!ipc) {
- kfree_skb(skb);
+ if (cb->type == QRTR_TYPE_RESUME_TX) {
+ qrtr_tx_resume(node, skb);
} else {
- if (sock_queue_rcv_skb(&ipc->sk, skb))
+ ipc = qrtr_port_lookup(cb->dst_port);
+ if (!ipc) {
kfree_skb(skb);
+ } else {
+ if (sock_queue_rcv_skb(&ipc->sk, skb))
+ kfree_skb(skb);
- qrtr_port_put(ipc);
+ qrtr_port_put(ipc);
+ }
}
}
}
@@ -415,6 +538,9 @@ int qrtr_endpoint_register(struct qrtr_endpoint *ep, unsigned int nid)
node->nid = QRTR_EP_NID_AUTO;
node->ep = ep;
+ INIT_RADIX_TREE(&node->qrtr_tx_flow, GFP_KERNEL);
+ init_waitqueue_head(&node->resume_tx);
+
qrtr_node_assign(node, nid);
mutex_lock(&qrtr_node_lock);
@@ -449,6 +575,9 @@ void qrtr_endpoint_unregister(struct qrtr_endpoint *ep)
qrtr_local_enqueue(NULL, skb, QRTR_TYPE_BYE, &src, &dst);
}
+ /* Wake up any transmitters waiting for resume-tx from the node */
+ wake_up_interruptible_all(&node->resume_tx);
+
qrtr_node_release(node);
ep->node = NULL;
}


--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project