From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932157AbbA0Ad5 (ORCPT ); Mon, 26 Jan 2015 19:33:57 -0500 Received: from mga02.intel.com ([134.134.136.20]:45147 "EHLO mga02.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756659AbbA0Ade (ORCPT ); Mon, 26 Jan 2015 19:33:34 -0500 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.09,471,1418112000"; d="scan'208";a="517961019" From: Sudeep Dutt To: Greg Kroah-Hartman Cc: Arnd Bergmann , Jonathan Corbet , linux-kernel@vger.kernel.org, linux-doc@vger.kernel.org, Dave Jiang , Nikhil Rao , Ashutosh Dixit , Sudeep Dutt Subject: [PATCH char-misc-next v3 10/13] misc: mic: SCIF messaging and node enumeration APIs Date: Mon, 26 Jan 2015 16:32:51 -0800 Message-Id: <7cd08be109ef782eb6798d1d790d126e4d085218.1422317099.git.sudeep.dutt@intel.com> X-Mailer: git-send-email 1.8.2.1 In-Reply-To: References: In-Reply-To: References: Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org SCIF messaging APIs which allow sending messages between the SCIF endpoints via a byte stream based ring buffer which has been optimized to avoid reads across PCIe. The SCIF messaging APIs are typically used for short < 1024 byte messages for best performance while the RDMA APIs which will be submitted in a future patch series is recommended for larger transfers. The node enumeration API enables a user to query for the number of nodes online in the SCIF network and their node ids. Reviewed-by: Ashutosh Dixit Signed-off-by: Nikhil Rao Signed-off-by: Sudeep Dutt --- drivers/misc/mic/scif/scif_epd.h | 4 + drivers/misc/mic/scif/scif_api.c | 402 ++++++++++++++++++++++++++++++++++++ drivers/misc/mic/scif/scif_epd.c | 32 +++ drivers/misc/mic/scif/scif_fd.c | 93 +++++++++ drivers/misc/mic/scif/scif_nm.c | 17 ++ drivers/misc/mic/scif/scif_nodeqp.c | 36 +++- 6 files changed, 583 insertions(+), 1 deletion(-) diff --git a/drivers/misc/mic/scif/scif_epd.h b/drivers/misc/mic/scif/scif_epd.h index 65c628e..331322a 100644 --- a/drivers/misc/mic/scif/scif_epd.h +++ b/drivers/misc/mic/scif/scif_epd.h @@ -144,6 +144,8 @@ int scif_rsrv_port(u16 port); void scif_get_port(u16 port); int scif_get_new_port(void); void scif_put_port(u16 port); +int scif_user_send(scif_epd_t epd, void __user *msg, int len, int flags); +int scif_user_recv(scif_epd_t epd, void __user *msg, int len, int flags); void scif_cnctreq(struct scif_dev *scifdev, struct scifmsg *msg); void scif_cnctgnt(struct scif_dev *scifdev, struct scifmsg *msg); void scif_cnctgnt_ack(struct scif_dev *scifdev, struct scifmsg *msg); @@ -151,6 +153,8 @@ void scif_cnctgnt_nack(struct scif_dev *scifdev, struct scifmsg *msg); void scif_cnctrej(struct scif_dev *scifdev, struct scifmsg *msg); void scif_discnct(struct scif_dev *scifdev, struct scifmsg *msg); void scif_discnt_ack(struct scif_dev *scifdev, struct scifmsg *msg); +void scif_clientsend(struct scif_dev *scifdev, struct scifmsg *msg); +void scif_clientrcvd(struct scif_dev *scifdev, struct scifmsg *msg); int __scif_connect(scif_epd_t epd, struct scif_port_id *dst, bool non_block); int __scif_flush(scif_epd_t epd); #endif /* SCIF_EPD_H */ diff --git a/drivers/misc/mic/scif/scif_api.c b/drivers/misc/mic/scif/scif_api.c index 8b55b59..41fba06 100644 --- a/drivers/misc/mic/scif/scif_api.c +++ b/drivers/misc/mic/scif/scif_api.c @@ -867,3 +867,405 @@ scif_accept_error_epalloc: return err; } EXPORT_SYMBOL_GPL(scif_accept); + +/* + * scif_msg_param_check: + * @epd: The end point returned from scif_open() + * @len: Length to receive + * @flags: blocking or non blocking + * + * Validate parameters for messaging APIs scif_send(..)/scif_recv(..). + */ +static inline int scif_msg_param_check(scif_epd_t epd, int len, int flags) +{ + int ret = -EINVAL; + + if (len < 0) + goto err_ret; + if (flags && (!(flags & SCIF_RECV_BLOCK))) + goto err_ret; + ret = 0; +err_ret: + return ret; +} + +static int _scif_send(scif_epd_t epd, void *msg, int len, int flags) +{ + struct scif_endpt *ep = (struct scif_endpt *)epd; + struct scifmsg notif_msg; + int curr_xfer_len = 0, sent_len = 0, write_count; + int ret = 0; + struct scif_qp *qp = ep->qp_info.qp; + + if (flags & SCIF_SEND_BLOCK) + might_sleep(); + + spin_lock(&ep->lock); + while (sent_len != len && SCIFEP_CONNECTED == ep->state) { + write_count = scif_rb_space(&qp->outbound_q); + if (write_count) { + /* Best effort to send as much data as possible */ + curr_xfer_len = min(len - sent_len, write_count); + ret = scif_rb_write(&qp->outbound_q, msg, + curr_xfer_len); + if (ret < 0) + break; + /* Success. Update write pointer */ + scif_rb_commit(&qp->outbound_q); + /* + * Send a notification to the peer about the + * produced data message. + */ + notif_msg.src = ep->port; + notif_msg.uop = SCIF_CLIENT_SENT; + notif_msg.payload[0] = ep->remote_ep; + ret = _scif_nodeqp_send(ep->remote_dev, ¬if_msg); + if (ret) + break; + sent_len += curr_xfer_len; + msg = msg + curr_xfer_len; + continue; + } + curr_xfer_len = min(len - sent_len, SCIF_ENDPT_QP_SIZE - 1); + /* Not enough RB space. return for the Non Blocking case */ + if (!(flags & SCIF_SEND_BLOCK)) + break; + + spin_unlock(&ep->lock); + /* Wait for a SCIF_CLIENT_RCVD message in the Blocking case */ + ret = + wait_event_interruptible(ep->sendwq, + (SCIFEP_CONNECTED != ep->state) || + (scif_rb_space(&qp->outbound_q) >= + curr_xfer_len)); + spin_lock(&ep->lock); + if (ret) + break; + } + if (sent_len) + ret = sent_len; + else if (!ret && SCIFEP_CONNECTED != ep->state) + ret = SCIFEP_DISCONNECTED == ep->state ? + -ECONNRESET : -ENOTCONN; + spin_unlock(&ep->lock); + return ret; +} + +static int _scif_recv(scif_epd_t epd, void *msg, int len, int flags) +{ + int read_size; + struct scif_endpt *ep = (struct scif_endpt *)epd; + struct scifmsg notif_msg; + int curr_recv_len = 0, remaining_len = len, read_count; + int ret = 0; + struct scif_qp *qp = ep->qp_info.qp; + + if (flags & SCIF_RECV_BLOCK) + might_sleep(); + spin_lock(&ep->lock); + while (remaining_len && (SCIFEP_CONNECTED == ep->state || + SCIFEP_DISCONNECTED == ep->state)) { + read_count = scif_rb_count(&qp->inbound_q, remaining_len); + if (read_count) { + /* + * Best effort to recv as much data as there + * are bytes to read in the RB particularly + * important for the Non Blocking case. + */ + curr_recv_len = min(remaining_len, read_count); + read_size = scif_rb_get_next(&qp->inbound_q, + msg, curr_recv_len); + if (ep->state == SCIFEP_CONNECTED) { + /* + * Update the read pointer only if the endpoint + * is still connected else the read pointer + * might no longer exist since the peer has + * freed resources! + */ + scif_rb_update_read_ptr(&qp->inbound_q); + /* + * Send a notification to the peer about the + * consumed data message only if the EP is in + * SCIFEP_CONNECTED state. + */ + notif_msg.src = ep->port; + notif_msg.uop = SCIF_CLIENT_RCVD; + notif_msg.payload[0] = ep->remote_ep; + ret = _scif_nodeqp_send(ep->remote_dev, + ¬if_msg); + if (ret) + break; + } + remaining_len -= curr_recv_len; + msg = msg + curr_recv_len; + continue; + } + /* + * Bail out now if the EP is in SCIFEP_DISCONNECTED state else + * we will keep looping forever. + */ + if (ep->state == SCIFEP_DISCONNECTED) + break; + /* + * Return in the Non Blocking case if there is no data + * to read in this iteration. + */ + if (!(flags & SCIF_RECV_BLOCK)) + break; + curr_recv_len = min(remaining_len, SCIF_ENDPT_QP_SIZE - 1); + spin_unlock(&ep->lock); + /* + * Wait for a SCIF_CLIENT_SEND message in the blocking case + * or until other side disconnects. + */ + ret = + wait_event_interruptible(ep->recvwq, + SCIFEP_CONNECTED != ep->state || + scif_rb_count(&qp->inbound_q, + curr_recv_len) + >= curr_recv_len); + spin_lock(&ep->lock); + if (ret) + break; + } + if (len - remaining_len) + ret = len - remaining_len; + else if (!ret && ep->state != SCIFEP_CONNECTED) + ret = ep->state == SCIFEP_DISCONNECTED ? + -ECONNRESET : -ENOTCONN; + spin_unlock(&ep->lock); + return ret; +} + +/** + * scif_user_send() - Send data to connection queue + * @epd: The end point returned from scif_open() + * @msg: Address to place data + * @len: Length to receive + * @flags: blocking or non blocking + * + * This function is called from the driver IOCTL entry point + * only and is a wrapper for _scif_send(). + */ +int scif_user_send(scif_epd_t epd, void __user *msg, int len, int flags) +{ + struct scif_endpt *ep = (struct scif_endpt *)epd; + int err = 0; + int sent_len = 0; + char *tmp; + int loop_len; + int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1))); + + dev_dbg(scif_info.mdev.this_device, + "SCIFAPI send (U): ep %p %s\n", ep, scif_ep_states[ep->state]); + if (!len) + return 0; + + err = scif_msg_param_check(epd, len, flags); + if (err) + goto send_err; + + tmp = kmalloc(chunk_len, GFP_KERNEL); + if (!tmp) { + err = -ENOMEM; + goto send_err; + } + /* + * Grabbing the lock before breaking up the transfer in + * multiple chunks is required to ensure that messages do + * not get fragmented and reordered. + */ + mutex_lock(&ep->sendlock); + while (sent_len != len) { + loop_len = len - sent_len; + loop_len = min(chunk_len, loop_len); + if (copy_from_user(tmp, msg, loop_len)) { + err = -EFAULT; + goto send_free_err; + } + err = _scif_send(epd, tmp, loop_len, flags); + if (err < 0) + goto send_free_err; + sent_len += err; + msg += err; + if (err != loop_len) + goto send_free_err; + } +send_free_err: + mutex_unlock(&ep->sendlock); + kfree(tmp); +send_err: + return err < 0 ? err : sent_len; +} + +/** + * scif_user_recv() - Receive data from connection queue + * @epd: The end point returned from scif_open() + * @msg: Address to place data + * @len: Length to receive + * @flags: blocking or non blocking + * + * This function is called from the driver IOCTL entry point + * only and is a wrapper for _scif_recv(). + */ +int scif_user_recv(scif_epd_t epd, void __user *msg, int len, int flags) +{ + struct scif_endpt *ep = (struct scif_endpt *)epd; + int err = 0; + int recv_len = 0; + char *tmp; + int loop_len; + int chunk_len = min(len, (1 << (MAX_ORDER + PAGE_SHIFT - 1))); + + dev_dbg(scif_info.mdev.this_device, + "SCIFAPI recv (U): ep %p %s\n", ep, scif_ep_states[ep->state]); + if (!len) + return 0; + + err = scif_msg_param_check(epd, len, flags); + if (err) + goto recv_err; + + tmp = kmalloc(chunk_len, GFP_KERNEL); + if (!tmp) { + err = -ENOMEM; + goto recv_err; + } + /* + * Grabbing the lock before breaking up the transfer in + * multiple chunks is required to ensure that messages do + * not get fragmented and reordered. + */ + mutex_lock(&ep->recvlock); + while (recv_len != len) { + loop_len = len - recv_len; + loop_len = min(chunk_len, loop_len); + err = _scif_recv(epd, tmp, loop_len, flags); + if (err < 0) + goto recv_free_err; + if (copy_to_user(msg, tmp, err)) { + err = -EFAULT; + goto recv_free_err; + } + recv_len += err; + msg += err; + if (err != loop_len) + goto recv_free_err; + } +recv_free_err: + mutex_unlock(&ep->recvlock); + kfree(tmp); +recv_err: + return err < 0 ? err : recv_len; +} + +/** + * scif_send() - Send data to connection queue + * @epd: The end point returned from scif_open() + * @msg: Address to place data + * @len: Length to receive + * @flags: blocking or non blocking + * + * This function is called from the kernel mode only and is + * a wrapper for _scif_send(). + */ +int scif_send(scif_epd_t epd, void *msg, int len, int flags) +{ + struct scif_endpt *ep = (struct scif_endpt *)epd; + int ret; + + dev_dbg(scif_info.mdev.this_device, + "SCIFAPI send (K): ep %p %s\n", ep, scif_ep_states[ep->state]); + if (!len) + return 0; + + ret = scif_msg_param_check(epd, len, flags); + if (ret) + return ret; + if (!ep->remote_dev) + return -ENOTCONN; + /* + * Grab the mutex lock in the blocking case only + * to ensure messages do not get fragmented/reordered. + * The non blocking mode is protected using spin locks + * in _scif_send(). + */ + if (flags & SCIF_SEND_BLOCK) + mutex_lock(&ep->sendlock); + + ret = _scif_send(epd, msg, len, flags); + + if (flags & SCIF_SEND_BLOCK) + mutex_unlock(&ep->sendlock); + return ret; +} +EXPORT_SYMBOL_GPL(scif_send); + +/** + * scif_recv() - Receive data from connection queue + * @epd: The end point returned from scif_open() + * @msg: Address to place data + * @len: Length to receive + * @flags: blocking or non blocking + * + * This function is called from the kernel mode only and is + * a wrapper for _scif_recv(). + */ +int scif_recv(scif_epd_t epd, void *msg, int len, int flags) +{ + struct scif_endpt *ep = (struct scif_endpt *)epd; + int ret; + + dev_dbg(scif_info.mdev.this_device, + "SCIFAPI recv (K): ep %p %s\n", ep, scif_ep_states[ep->state]); + if (!len) + return 0; + + ret = scif_msg_param_check(epd, len, flags); + if (ret) + return ret; + /* + * Grab the mutex lock in the blocking case only + * to ensure messages do not get fragmented/reordered. + * The non blocking mode is protected using spin locks + * in _scif_send(). + */ + if (flags & SCIF_RECV_BLOCK) + mutex_lock(&ep->recvlock); + + ret = _scif_recv(epd, msg, len, flags); + + if (flags & SCIF_RECV_BLOCK) + mutex_unlock(&ep->recvlock); + + return ret; +} +EXPORT_SYMBOL_GPL(scif_recv); + +int scif_get_node_ids(u16 *nodes, int len, u16 *self) +{ + int online = 0; + int offset = 0; + int node; + + if (!scif_is_mgmt_node()) + scif_get_node_info(); + + *self = scif_info.nodeid; + mutex_lock(&scif_info.conflock); + len = min_t(int, len, scif_info.total); + for (node = 0; node <= scif_info.maxid; node++) { + if (_scifdev_alive(&scif_dev[node])) { + online++; + if (offset < len) + nodes[offset++] = node; + } + } + dev_dbg(scif_info.mdev.this_device, + "SCIFAPI get_node_ids total %d online %d filled in %d nodes\n", + scif_info.total, online, offset); + mutex_unlock(&scif_info.conflock); + + return online; +} +EXPORT_SYMBOL_GPL(scif_get_node_ids); diff --git a/drivers/misc/mic/scif/scif_epd.c b/drivers/misc/mic/scif/scif_epd.c index 685ee5c..b4bfbb0 100644 --- a/drivers/misc/mic/scif/scif_epd.c +++ b/drivers/misc/mic/scif/scif_epd.c @@ -319,3 +319,35 @@ void scif_discnt_ack(struct scif_dev *scifdev, struct scifmsg *msg) spin_unlock(&ep->lock); complete(&ep->discon); } + +/** + * scif_clientsend() - Respond to SCIF_CLIENT_SEND interrupt message + * @msg: Interrupt message + * + * Remote side is confirming send or receive interrupt handling is complete. + */ +void scif_clientsend(struct scif_dev *scifdev, struct scifmsg *msg) +{ + struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0]; + + spin_lock(&ep->lock); + if (SCIFEP_CONNECTED == ep->state) + wake_up_interruptible(&ep->recvwq); + spin_unlock(&ep->lock); +} + +/** + * scif_clientrcvd() - Respond to SCIF_CLIENT_RCVD interrupt message + * @msg: Interrupt message + * + * Remote side is confirming send or receive interrupt handling is complete. + */ +void scif_clientrcvd(struct scif_dev *scifdev, struct scifmsg *msg) +{ + struct scif_endpt *ep = (struct scif_endpt *)msg->payload[0]; + + spin_lock(&ep->lock); + if (SCIFEP_CONNECTED == ep->state) + wake_up_interruptible(&ep->sendwq); + spin_unlock(&ep->lock); +} diff --git a/drivers/misc/mic/scif/scif_fd.c b/drivers/misc/mic/scif/scif_fd.c index c1d6bc3..b1678fe 100644 --- a/drivers/misc/mic/scif/scif_fd.c +++ b/drivers/misc/mic/scif/scif_fd.c @@ -69,6 +69,7 @@ static long scif_fdioctl(struct file *f, unsigned int cmd, unsigned long arg) struct scif_endpt *priv = f->private_data; void __user *argp = (void __user *)arg; int err = 0; + struct scifioctl_msg request; bool non_block = false; non_block = !!(f->f_flags & O_NONBLOCK); @@ -197,6 +198,98 @@ static long scif_fdioctl(struct file *f, unsigned int cmd, unsigned long arg) f->private_data = newep; return 0; } + case SCIF_SEND: + { + struct scif_endpt *priv = f->private_data; + + if (copy_from_user(&request, argp, + sizeof(struct scifioctl_msg))) { + err = -EFAULT; + goto send_err; + } + err = scif_user_send(priv, request.msg, + request.len, request.flags); + if (err < 0) + goto send_err; + if (copy_to_user(& + ((struct scifioctl_msg __user *)argp)->out_len, + &err, sizeof(err))) { + err = -EFAULT; + goto send_err; + } + err = 0; +send_err: + scif_err_debug(err, "scif_send"); + return err; + } + case SCIF_RECV: + { + struct scif_endpt *priv = f->private_data; + + if (copy_from_user(&request, argp, + sizeof(struct scifioctl_msg))) { + err = -EFAULT; + goto recv_err; + } + + err = scif_user_recv(priv, request.msg, + request.len, request.flags); + if (err < 0) + goto recv_err; + + if (copy_to_user(& + ((struct scifioctl_msg __user *)argp)->out_len, + &err, sizeof(err))) { + err = -EFAULT; + goto recv_err; + } + err = 0; +recv_err: + scif_err_debug(err, "scif_recv"); + return err; + } + case SCIF_GET_NODEIDS: + { + struct scifioctl_node_ids node_ids; + int entries; + u16 *nodes; + void __user *unodes, *uself; + u16 self; + + if (copy_from_user(&node_ids, argp, sizeof(node_ids))) { + err = -EFAULT; + goto getnodes_err2; + } + + entries = min_t(int, scif_info.maxid, node_ids.len); + nodes = kmalloc_array(entries, sizeof(u16), GFP_KERNEL); + if (entries && !nodes) { + err = -ENOMEM; + goto getnodes_err2; + } + node_ids.len = scif_get_node_ids(nodes, entries, &self); + + unodes = node_ids.nodes; + if (copy_to_user(unodes, nodes, sizeof(u16) * entries)) { + err = -EFAULT; + goto getnodes_err1; + } + + uself = node_ids.self; + if (copy_to_user(uself, &self, sizeof(u16))) { + err = -EFAULT; + goto getnodes_err1; + } + + if (copy_to_user(argp, &node_ids, sizeof(node_ids))) { + err = -EFAULT; + goto getnodes_err1; + } +getnodes_err1: + kfree(nodes); +getnodes_err2: + return err; + } } return -EINVAL; } diff --git a/drivers/misc/mic/scif/scif_nm.c b/drivers/misc/mic/scif/scif_nm.c index 0469aad..9b4c538 100644 --- a/drivers/misc/mic/scif/scif_nm.c +++ b/drivers/misc/mic/scif/scif_nm.c @@ -218,3 +218,20 @@ void scif_disconnect_node(u32 node_id, bool mgmt_initiated) (atomic_read(&scifdev->disconn_rescnt) == 1), SCIF_NODE_ALIVE_TIMEOUT); } + +void scif_get_node_info(void) +{ + struct scifmsg msg; + DECLARE_COMPLETION_ONSTACK(node_info); + + msg.uop = SCIF_GET_NODE_INFO; + msg.src.node = scif_info.nodeid; + msg.dst.node = SCIF_MGMT_NODE; + msg.payload[3] = (u64)&node_info; + + if ((scif_nodeqp_send(&scif_dev[SCIF_MGMT_NODE], &msg))) + return; + + /* Wait for a response with SCIF_GET_NODE_INFO */ + wait_for_completion(&node_info); +} diff --git a/drivers/misc/mic/scif/scif_nodeqp.c b/drivers/misc/mic/scif/scif_nodeqp.c index aee9025..141923c 100644 --- a/drivers/misc/mic/scif/scif_nodeqp.c +++ b/drivers/misc/mic/scif/scif_nodeqp.c @@ -576,7 +576,10 @@ static char *message_types[] = {"BAD", "CNCT_GNTNACK", "CNCT_REJ", "DISCNCT", - "DISCNT_ACK"}; + "DISCNT_ACK", + "CLIENT_SENT", + "CLIENT_RCVD", + "SCIF_GET_NODE_INFO"}; static void scif_display_message(struct scif_dev *scifdev, struct scifmsg *msg, @@ -946,6 +949,34 @@ scif_node_remove_ack(struct scif_dev *scifdev, struct scifmsg *msg) wake_up(&sdev->disconn_wq); } +/** + * scif_get_node_info: Respond to SCIF_GET_NODE_INFO interrupt message + * @msg: Interrupt message + * + * Retrieve node info i.e maxid and total from the mgmt node. + */ +static __always_inline void +scif_get_node_info_resp(struct scif_dev *scifdev, struct scifmsg *msg) +{ + if (scif_is_mgmt_node()) { + swap(msg->dst.node, msg->src.node); + mutex_lock(&scif_info.conflock); + msg->payload[1] = scif_info.maxid; + msg->payload[2] = scif_info.total; + mutex_unlock(&scif_info.conflock); + scif_nodeqp_send(scifdev, msg); + } else { + struct completion *node_info = + (struct completion *)msg->payload[3]; + + mutex_lock(&scif_info.conflock); + scif_info.maxid = msg->payload[1]; + scif_info.total = msg->payload[2]; + complete_all(node_info); + mutex_unlock(&scif_info.conflock); + } +} + static void scif_msg_unknown(struct scif_dev *scifdev, struct scifmsg *msg) { @@ -973,6 +1004,9 @@ static void (*scif_intr_func[SCIF_MAX_MSG + 1]) scif_cnctrej, /* SCIF_CNCT_REJ */ scif_discnct, /* SCIF_DISCNCT */ scif_discnt_ack, /* SCIF_DISCNT_ACK */ + scif_clientsend, /* SCIF_CLIENT_SENT */ + scif_clientrcvd, /* SCIF_CLIENT_RCVD */ + scif_get_node_info_resp,/* SCIF_GET_NODE_INFO */ }; /** -- 1.8.2.1