2 * linux/fs/9p/trans_rdma.c
4 * RDMA transport layer based on the trans_fd.c implementation.
6 * Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
7 * Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
8 * Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
9 * Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
10 * Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2
14 * as published by the Free Software Foundation.
16 * This program is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
19 * GNU General Public License for more details.
21 * You should have received a copy of the GNU General Public License
22 * along with this program; if not, write to:
23 * Free Software Foundation
24 * 51 Franklin Street, Fifth Floor
25 * Boston, MA 02111-1301 USA
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32 #include <linux/module.h>
33 #include <linux/net.h>
34 #include <linux/ipv6.h>
35 #include <linux/kthread.h>
36 #include <linux/errno.h>
37 #include <linux/kernel.h>
39 #include <linux/uaccess.h>
40 #include <linux/inet.h>
41 #include <linux/idr.h>
42 #include <linux/file.h>
43 #include <linux/parser.h>
44 #include <linux/semaphore.h>
45 #include <linux/slab.h>
46 #include <net/9p/9p.h>
47 #include <net/9p/client.h>
48 #include <net/9p/transport.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/rdma_cm.h>
53 #define P9_RDMA_SQ_DEPTH 32
54 #define P9_RDMA_RQ_DEPTH 32
55 #define P9_RDMA_SEND_SGE 4
56 #define P9_RDMA_RECV_SGE 4
59 #define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
60 #define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */
63 * struct p9_trans_rdma - RDMA transport instance
65 * @state: tracks the transport state machine for connection setup and tear down
66 * @cm_id: The RDMA CM ID
67 * @pd: Protection Domain pointer
68 * @qp: Queue Pair pointer
69 * @cq: Completion Queue pointer
70 * @dm_mr: DMA Memory Region pointer
71 * @lkey: The local access only memory region key
72 * @timeout: Number of uSecs to wait for connection management events
73 * @sq_depth: The depth of the Send Queue
74 * @sq_sem: Semaphore for the SQ
75 * @rq_depth: The depth of the Receive Queue.
76 * @rq_sem: Semaphore for the RQ
77 * @excess_rc : Amount of posted Receive Contexts without a pending request.
79 * @addr: The remote peer's address
80 * @req_lock: Protects the active request list
81 * @cm_done: Completion event for connection management tracking
83 struct p9_trans_rdma {
86 P9_RDMA_ADDR_RESOLVED,
87 P9_RDMA_ROUTE_RESOLVED,
93 struct rdma_cm_id *cm_id;
99 struct semaphore sq_sem;
101 struct semaphore rq_sem;
103 struct sockaddr_in addr;
106 struct completion cm_done;
110 * p9_rdma_context - Keeps track of in-process WR
112 * @busa: Bus address to unmap when the WR completes
113 * @req: Keeps track of requests (send)
114 * @rc: Keepts track of replies (receive)
117 struct p9_rdma_context {
121 struct p9_req_t *req;
127 * p9_rdma_opts - Collection of mount options
128 * @port: port of connection
129 * @sq_depth: The requested depth of the SQ. This really doesn't need
130 * to be any deeper than the number of threads used in the client
131 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
132 * @timeout: Time to wait in msecs for CM events
134 struct p9_rdma_opts {
143 * Option Parsing (code inspired by NFS code)
146 /* Options that take integer arguments */
147 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
148 /* Options that take no argument */
153 static match_table_t tokens = {
154 {Opt_port, "port=%u"},
155 {Opt_sq_depth, "sq=%u"},
156 {Opt_rq_depth, "rq=%u"},
157 {Opt_timeout, "timeout=%u"},
158 {Opt_privport, "privport"},
163 * parse_opts - parse mount options into rdma options structure
164 * @params: options string passed from mount
165 * @opts: rdma transport-specific structure to parse options into
167 * Returns 0 upon success, -ERRNO upon failure
169 static int parse_opts(char *params, struct p9_rdma_opts *opts)
172 substring_t args[MAX_OPT_ARGS];
174 char *options, *tmp_options;
176 opts->port = P9_PORT;
177 opts->sq_depth = P9_RDMA_SQ_DEPTH;
178 opts->rq_depth = P9_RDMA_RQ_DEPTH;
179 opts->timeout = P9_RDMA_TIMEOUT;
185 tmp_options = kstrdup(params, GFP_KERNEL);
187 p9_debug(P9_DEBUG_ERROR,
188 "failed to allocate copy of option string\n");
191 options = tmp_options;
193 while ((p = strsep(&options, ",")) != NULL) {
198 token = match_token(p, tokens, args);
199 if ((token != Opt_err) && (token != Opt_privport)) {
200 r = match_int(&args[0], &option);
202 p9_debug(P9_DEBUG_ERROR,
203 "integer field, but no integer?\n");
212 opts->sq_depth = option;
215 opts->rq_depth = option;
218 opts->timeout = option;
227 /* RQ must be at least as large as the SQ */
228 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
234 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
236 struct p9_client *c = id->context;
237 struct p9_trans_rdma *rdma = c->trans;
238 switch (event->event) {
239 case RDMA_CM_EVENT_ADDR_RESOLVED:
240 BUG_ON(rdma->state != P9_RDMA_INIT);
241 rdma->state = P9_RDMA_ADDR_RESOLVED;
244 case RDMA_CM_EVENT_ROUTE_RESOLVED:
245 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
246 rdma->state = P9_RDMA_ROUTE_RESOLVED;
249 case RDMA_CM_EVENT_ESTABLISHED:
250 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
251 rdma->state = P9_RDMA_CONNECTED;
254 case RDMA_CM_EVENT_DISCONNECTED:
256 rdma->state = P9_RDMA_CLOSED;
258 c->status = Disconnected;
261 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
264 case RDMA_CM_EVENT_ADDR_CHANGE:
265 case RDMA_CM_EVENT_ROUTE_ERROR:
266 case RDMA_CM_EVENT_DEVICE_REMOVAL:
267 case RDMA_CM_EVENT_MULTICAST_JOIN:
268 case RDMA_CM_EVENT_MULTICAST_ERROR:
269 case RDMA_CM_EVENT_REJECTED:
270 case RDMA_CM_EVENT_CONNECT_REQUEST:
271 case RDMA_CM_EVENT_CONNECT_RESPONSE:
272 case RDMA_CM_EVENT_CONNECT_ERROR:
273 case RDMA_CM_EVENT_ADDR_ERROR:
274 case RDMA_CM_EVENT_UNREACHABLE:
275 c->status = Disconnected;
276 rdma_disconnect(rdma->cm_id);
281 complete(&rdma->cm_done);
286 recv_done(struct ib_cq *cq, struct ib_wc *wc)
288 struct p9_client *client = cq->cq_context;
289 struct p9_trans_rdma *rdma = client->trans;
290 struct p9_rdma_context *c =
291 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
292 struct p9_req_t *req;
297 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
300 if (wc->status != IB_WC_SUCCESS)
303 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
307 req = p9_tag_lookup(client, tag);
311 /* Check that we have not yet received a reply for this request.
313 if (unlikely(req->rc)) {
314 pr_err("Duplicate reply for request %d", tag);
319 p9_client_cb(client, req, REQ_STATUS_RCVD);
327 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
328 req, err, wc->status);
329 rdma->state = P9_RDMA_FLUSHING;
330 client->status = Disconnected;
335 send_done(struct ib_cq *cq, struct ib_wc *wc)
337 struct p9_client *client = cq->cq_context;
338 struct p9_trans_rdma *rdma = client->trans;
339 struct p9_rdma_context *c =
340 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
342 ib_dma_unmap_single(rdma->cm_id->device,
343 c->busa, c->req->tc->size,
349 static void qp_event_handler(struct ib_event *event, void *context)
351 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
352 event->event, context);
355 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
360 if (rdma->qp && !IS_ERR(rdma->qp))
361 ib_destroy_qp(rdma->qp);
363 if (rdma->pd && !IS_ERR(rdma->pd))
364 ib_dealloc_pd(rdma->pd);
366 if (rdma->cq && !IS_ERR(rdma->cq))
367 ib_free_cq(rdma->cq);
369 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
370 rdma_destroy_id(rdma->cm_id);
376 post_recv(struct p9_client *client, struct p9_rdma_context *c)
378 struct p9_trans_rdma *rdma = client->trans;
379 struct ib_recv_wr wr, *bad_wr;
382 c->busa = ib_dma_map_single(rdma->cm_id->device,
383 c->rc->sdata, client->msize,
385 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
388 c->cqe.done = recv_done;
391 sge.length = client->msize;
392 sge.lkey = rdma->pd->local_dma_lkey;
398 return ib_post_recv(rdma->qp, &wr, &bad_wr);
401 p9_debug(P9_DEBUG_ERROR, "EIO\n");
405 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
407 struct p9_trans_rdma *rdma = client->trans;
408 struct ib_send_wr wr, *bad_wr;
412 struct p9_rdma_context *c = NULL;
413 struct p9_rdma_context *rpl_context = NULL;
415 /* When an error occurs between posting the recv and the send,
416 * there will be a receive context posted without a pending request.
417 * Since there is no way to "un-post" it, we remember it and skip
418 * post_recv() for the next request.
420 * see if we are this `next request' and need to absorb an excess rc.
421 * If yes, then drop and free our own, and do not recv_post().
423 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
424 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
428 goto dont_need_post_recv;
430 /* We raced and lost. */
431 atomic_inc(&rdma->excess_rc);
435 /* Allocate an fcall for the reply */
436 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
441 rpl_context->rc = req->rc;
444 * Post a receive buffer for this request. We need to ensure
445 * there is a reply buffer available for every outstanding
446 * request. A flushed request can result in no reply for an
447 * outstanding request, so we must keep a count to avoid
448 * overflowing the RQ.
450 if (down_interruptible(&rdma->rq_sem)) {
455 err = post_recv(client, rpl_context);
457 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
460 /* remove posted receive buffer from request structure */
464 /* Post the request */
465 c = kmalloc(sizeof *c, GFP_NOFS);
472 c->busa = ib_dma_map_single(rdma->cm_id->device,
473 c->req->tc->sdata, c->req->tc->size,
475 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
480 c->cqe.done = send_done;
483 sge.length = c->req->tc->size;
484 sge.lkey = rdma->pd->local_dma_lkey;
488 wr.opcode = IB_WR_SEND;
489 wr.send_flags = IB_SEND_SIGNALED;
493 if (down_interruptible(&rdma->sq_sem)) {
498 /* Mark request as `sent' *before* we actually send it,
499 * because doing if after could erase the REQ_STATUS_RCVD
500 * status in case of a very fast reply.
502 req->status = REQ_STATUS_SENT;
503 err = ib_post_send(rdma->qp, &wr, &bad_wr);
510 /* Handle errors that happened during or while preparing the send: */
512 req->status = REQ_STATUS_ERROR;
514 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
517 * We did recv_post(), but not send. We have one recv_post in excess.
519 atomic_inc(&rdma->excess_rc);
522 /* Handle errors that happened during or while preparing post_recv(): */
525 spin_lock_irqsave(&rdma->req_lock, flags);
526 if (rdma->state < P9_RDMA_CLOSING) {
527 rdma->state = P9_RDMA_CLOSING;
528 spin_unlock_irqrestore(&rdma->req_lock, flags);
529 rdma_disconnect(rdma->cm_id);
531 spin_unlock_irqrestore(&rdma->req_lock, flags);
535 static void rdma_close(struct p9_client *client)
537 struct p9_trans_rdma *rdma;
542 rdma = client->trans;
546 client->status = Disconnected;
547 rdma_disconnect(rdma->cm_id);
548 rdma_destroy_trans(rdma);
552 * alloc_rdma - Allocate and initialize the rdma transport structure
553 * @opts: Mount options structure
555 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
557 struct p9_trans_rdma *rdma;
559 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
563 rdma->sq_depth = opts->sq_depth;
564 rdma->rq_depth = opts->rq_depth;
565 rdma->timeout = opts->timeout;
566 spin_lock_init(&rdma->req_lock);
567 init_completion(&rdma->cm_done);
568 sema_init(&rdma->sq_sem, rdma->sq_depth);
569 sema_init(&rdma->rq_sem, rdma->rq_depth);
570 atomic_set(&rdma->excess_rc, 0);
575 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
577 /* Nothing to do here.
578 * We will take care of it (if we have to) in rdma_cancelled()
583 /* A request has been fully flushed without a reply.
584 * That means we have posted one buffer in excess.
586 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
588 struct p9_trans_rdma *rdma = client->trans;
589 atomic_inc(&rdma->excess_rc);
593 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
595 struct sockaddr_in cl = {
596 .sin_family = AF_INET,
597 .sin_addr.s_addr = htonl(INADDR_ANY),
599 int port, err = -EINVAL;
601 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
602 cl.sin_port = htons((ushort)port);
603 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
604 if (err != -EADDRINUSE)
611 * trans_create_rdma - Transport method for creating atransport instance
612 * @client: client instance
613 * @addr: IP address string
614 * @args: Mount options string
617 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
620 struct p9_rdma_opts opts;
621 struct p9_trans_rdma *rdma;
622 struct rdma_conn_param conn_param;
623 struct ib_qp_init_attr qp_attr;
625 /* Parse the transport specific mount options */
626 err = parse_opts(args, &opts);
630 /* Create and initialize the RDMA transport structure */
631 rdma = alloc_rdma(&opts);
635 /* Create the RDMA CM ID */
636 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
637 RDMA_PS_TCP, IB_QPT_RC);
638 if (IS_ERR(rdma->cm_id))
641 /* Associate the client with the transport */
642 client->trans = rdma;
644 /* Bind to a privileged port if we need to */
646 err = p9_rdma_bind_privport(rdma);
648 pr_err("%s (%d): problem binding to privport: %d\n",
649 __func__, task_pid_nr(current), -err);
654 /* Resolve the server's address */
655 rdma->addr.sin_family = AF_INET;
656 rdma->addr.sin_addr.s_addr = in_aton(addr);
657 rdma->addr.sin_port = htons(opts.port);
658 err = rdma_resolve_addr(rdma->cm_id, NULL,
659 (struct sockaddr *)&rdma->addr,
663 err = wait_for_completion_interruptible(&rdma->cm_done);
664 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
667 /* Resolve the route to the server */
668 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
671 err = wait_for_completion_interruptible(&rdma->cm_done);
672 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
675 /* Create the Completion Queue */
676 rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
677 opts.sq_depth + opts.rq_depth + 1,
679 if (IS_ERR(rdma->cq))
682 /* Create the Protection Domain */
683 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
684 if (IS_ERR(rdma->pd))
687 /* Create the Queue Pair */
688 memset(&qp_attr, 0, sizeof qp_attr);
689 qp_attr.event_handler = qp_event_handler;
690 qp_attr.qp_context = client;
691 qp_attr.cap.max_send_wr = opts.sq_depth;
692 qp_attr.cap.max_recv_wr = opts.rq_depth;
693 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
694 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
695 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
696 qp_attr.qp_type = IB_QPT_RC;
697 qp_attr.send_cq = rdma->cq;
698 qp_attr.recv_cq = rdma->cq;
699 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
702 rdma->qp = rdma->cm_id->qp;
704 /* Request a connection */
705 memset(&conn_param, 0, sizeof(conn_param));
706 conn_param.private_data = NULL;
707 conn_param.private_data_len = 0;
708 conn_param.responder_resources = P9_RDMA_IRD;
709 conn_param.initiator_depth = P9_RDMA_ORD;
710 err = rdma_connect(rdma->cm_id, &conn_param);
713 err = wait_for_completion_interruptible(&rdma->cm_done);
714 if (err || (rdma->state != P9_RDMA_CONNECTED))
717 client->status = Connected;
722 rdma_destroy_trans(rdma);
726 static struct p9_trans_module p9_rdma_trans = {
728 .maxsize = P9_RDMA_MAXSIZE,
730 .owner = THIS_MODULE,
731 .create = rdma_create_trans,
733 .request = rdma_request,
734 .cancel = rdma_cancel,
735 .cancelled = rdma_cancelled,
739 * p9_trans_rdma_init - Register the 9P RDMA transport driver
741 static int __init p9_trans_rdma_init(void)
743 v9fs_register_trans(&p9_rdma_trans);
747 static void __exit p9_trans_rdma_exit(void)
749 v9fs_unregister_trans(&p9_rdma_trans);
752 module_init(p9_trans_rdma_init);
753 module_exit(p9_trans_rdma_exit);
755 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
756 MODULE_DESCRIPTION("RDMA Transport for 9P");
757 MODULE_LICENSE("Dual BSD/GPL");