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 <linux/seq_file.h>
47 #include <net/9p/9p.h>
48 #include <net/9p/client.h>
49 #include <net/9p/transport.h>
50 #include <rdma/ib_verbs.h>
51 #include <rdma/rdma_cm.h>
54 #define P9_RDMA_SQ_DEPTH 32
55 #define P9_RDMA_RQ_DEPTH 32
56 #define P9_RDMA_SEND_SGE 4
57 #define P9_RDMA_RECV_SGE 4
60 #define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
61 #define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */
64 * struct p9_trans_rdma - RDMA transport instance
66 * @state: tracks the transport state machine for connection setup and tear down
67 * @cm_id: The RDMA CM ID
68 * @pd: Protection Domain pointer
69 * @qp: Queue Pair pointer
70 * @cq: Completion Queue pointer
71 * @timeout: Number of uSecs to wait for connection management events
72 * @privport: Whether a privileged port may be used
73 * @port: The port to use
74 * @sq_depth: The depth of the Send Queue
75 * @sq_sem: Semaphore for the SQ
76 * @rq_depth: The depth of the Receive Queue.
77 * @rq_sem: Semaphore for the RQ
78 * @excess_rc : Amount of posted Receive Contexts without a pending request.
80 * @addr: The remote peer's address
81 * @req_lock: Protects the active request list
82 * @cm_done: Completion event for connection management tracking
84 struct p9_trans_rdma {
87 P9_RDMA_ADDR_RESOLVED,
88 P9_RDMA_ROUTE_RESOLVED,
94 struct rdma_cm_id *cm_id;
102 struct semaphore sq_sem;
104 struct semaphore rq_sem;
106 struct sockaddr_in addr;
109 struct completion cm_done;
113 * p9_rdma_context - Keeps track of in-process WR
115 * @busa: Bus address to unmap when the WR completes
116 * @req: Keeps track of requests (send)
117 * @rc: Keepts track of replies (receive)
120 struct p9_rdma_context {
124 struct p9_req_t *req;
130 * p9_rdma_opts - Collection of mount options
131 * @port: port of connection
132 * @sq_depth: The requested depth of the SQ. This really doesn't need
133 * to be any deeper than the number of threads used in the client
134 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
135 * @timeout: Time to wait in msecs for CM events
137 struct p9_rdma_opts {
146 * Option Parsing (code inspired by NFS code)
149 /* Options that take integer arguments */
150 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
151 /* Options that take no argument */
156 static match_table_t tokens = {
157 {Opt_port, "port=%u"},
158 {Opt_sq_depth, "sq=%u"},
159 {Opt_rq_depth, "rq=%u"},
160 {Opt_timeout, "timeout=%u"},
161 {Opt_privport, "privport"},
165 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
167 struct p9_trans_rdma *rdma = clnt->trans;
169 if (rdma->port != P9_PORT)
170 seq_printf(m, ",port=%u", rdma->port);
171 if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
172 seq_printf(m, ",sq=%u", rdma->sq_depth);
173 if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
174 seq_printf(m, ",rq=%u", rdma->rq_depth);
175 if (rdma->timeout != P9_RDMA_TIMEOUT)
176 seq_printf(m, ",timeout=%lu", rdma->timeout);
178 seq_puts(m, ",privport");
183 * parse_opts - parse mount options into rdma options structure
184 * @params: options string passed from mount
185 * @opts: rdma transport-specific structure to parse options into
187 * Returns 0 upon success, -ERRNO upon failure
189 static int parse_opts(char *params, struct p9_rdma_opts *opts)
192 substring_t args[MAX_OPT_ARGS];
194 char *options, *tmp_options;
196 opts->port = P9_PORT;
197 opts->sq_depth = P9_RDMA_SQ_DEPTH;
198 opts->rq_depth = P9_RDMA_RQ_DEPTH;
199 opts->timeout = P9_RDMA_TIMEOUT;
200 opts->privport = false;
205 tmp_options = kstrdup(params, GFP_KERNEL);
207 p9_debug(P9_DEBUG_ERROR,
208 "failed to allocate copy of option string\n");
211 options = tmp_options;
213 while ((p = strsep(&options, ",")) != NULL) {
218 token = match_token(p, tokens, args);
219 if ((token != Opt_err) && (token != Opt_privport)) {
220 r = match_int(&args[0], &option);
222 p9_debug(P9_DEBUG_ERROR,
223 "integer field, but no integer?\n");
232 opts->sq_depth = option;
235 opts->rq_depth = option;
238 opts->timeout = option;
241 opts->privport = true;
247 /* RQ must be at least as large as the SQ */
248 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
254 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
256 struct p9_client *c = id->context;
257 struct p9_trans_rdma *rdma = c->trans;
258 switch (event->event) {
259 case RDMA_CM_EVENT_ADDR_RESOLVED:
260 BUG_ON(rdma->state != P9_RDMA_INIT);
261 rdma->state = P9_RDMA_ADDR_RESOLVED;
264 case RDMA_CM_EVENT_ROUTE_RESOLVED:
265 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
266 rdma->state = P9_RDMA_ROUTE_RESOLVED;
269 case RDMA_CM_EVENT_ESTABLISHED:
270 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
271 rdma->state = P9_RDMA_CONNECTED;
274 case RDMA_CM_EVENT_DISCONNECTED:
276 rdma->state = P9_RDMA_CLOSED;
278 c->status = Disconnected;
281 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
284 case RDMA_CM_EVENT_ADDR_CHANGE:
285 case RDMA_CM_EVENT_ROUTE_ERROR:
286 case RDMA_CM_EVENT_DEVICE_REMOVAL:
287 case RDMA_CM_EVENT_MULTICAST_JOIN:
288 case RDMA_CM_EVENT_MULTICAST_ERROR:
289 case RDMA_CM_EVENT_REJECTED:
290 case RDMA_CM_EVENT_CONNECT_REQUEST:
291 case RDMA_CM_EVENT_CONNECT_RESPONSE:
292 case RDMA_CM_EVENT_CONNECT_ERROR:
293 case RDMA_CM_EVENT_ADDR_ERROR:
294 case RDMA_CM_EVENT_UNREACHABLE:
295 c->status = Disconnected;
296 rdma_disconnect(rdma->cm_id);
301 complete(&rdma->cm_done);
306 recv_done(struct ib_cq *cq, struct ib_wc *wc)
308 struct p9_client *client = cq->cq_context;
309 struct p9_trans_rdma *rdma = client->trans;
310 struct p9_rdma_context *c =
311 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
312 struct p9_req_t *req;
317 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
320 if (wc->status != IB_WC_SUCCESS)
323 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
327 req = p9_tag_lookup(client, tag);
331 /* Check that we have not yet received a reply for this request.
333 if (unlikely(req->rc)) {
334 pr_err("Duplicate reply for request %d", tag);
339 p9_client_cb(client, req, REQ_STATUS_RCVD);
347 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
348 req, err, wc->status);
349 rdma->state = P9_RDMA_FLUSHING;
350 client->status = Disconnected;
355 send_done(struct ib_cq *cq, struct ib_wc *wc)
357 struct p9_client *client = cq->cq_context;
358 struct p9_trans_rdma *rdma = client->trans;
359 struct p9_rdma_context *c =
360 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
362 ib_dma_unmap_single(rdma->cm_id->device,
363 c->busa, c->req->tc->size,
369 static void qp_event_handler(struct ib_event *event, void *context)
371 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
372 event->event, context);
375 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
380 if (rdma->qp && !IS_ERR(rdma->qp))
381 ib_destroy_qp(rdma->qp);
383 if (rdma->pd && !IS_ERR(rdma->pd))
384 ib_dealloc_pd(rdma->pd);
386 if (rdma->cq && !IS_ERR(rdma->cq))
387 ib_free_cq(rdma->cq);
389 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
390 rdma_destroy_id(rdma->cm_id);
396 post_recv(struct p9_client *client, struct p9_rdma_context *c)
398 struct p9_trans_rdma *rdma = client->trans;
399 struct ib_recv_wr wr, *bad_wr;
402 c->busa = ib_dma_map_single(rdma->cm_id->device,
403 c->rc->sdata, client->msize,
405 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
408 c->cqe.done = recv_done;
411 sge.length = client->msize;
412 sge.lkey = rdma->pd->local_dma_lkey;
418 return ib_post_recv(rdma->qp, &wr, &bad_wr);
421 p9_debug(P9_DEBUG_ERROR, "EIO\n");
425 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
427 struct p9_trans_rdma *rdma = client->trans;
428 struct ib_send_wr wr, *bad_wr;
432 struct p9_rdma_context *c = NULL;
433 struct p9_rdma_context *rpl_context = NULL;
435 /* When an error occurs between posting the recv and the send,
436 * there will be a receive context posted without a pending request.
437 * Since there is no way to "un-post" it, we remember it and skip
438 * post_recv() for the next request.
440 * see if we are this `next request' and need to absorb an excess rc.
441 * If yes, then drop and free our own, and do not recv_post().
443 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
444 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
448 goto dont_need_post_recv;
450 /* We raced and lost. */
451 atomic_inc(&rdma->excess_rc);
455 /* Allocate an fcall for the reply */
456 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
461 rpl_context->rc = req->rc;
464 * Post a receive buffer for this request. We need to ensure
465 * there is a reply buffer available for every outstanding
466 * request. A flushed request can result in no reply for an
467 * outstanding request, so we must keep a count to avoid
468 * overflowing the RQ.
470 if (down_interruptible(&rdma->rq_sem)) {
475 err = post_recv(client, rpl_context);
477 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
480 /* remove posted receive buffer from request structure */
484 /* Post the request */
485 c = kmalloc(sizeof *c, GFP_NOFS);
492 c->busa = ib_dma_map_single(rdma->cm_id->device,
493 c->req->tc->sdata, c->req->tc->size,
495 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
500 c->cqe.done = send_done;
503 sge.length = c->req->tc->size;
504 sge.lkey = rdma->pd->local_dma_lkey;
508 wr.opcode = IB_WR_SEND;
509 wr.send_flags = IB_SEND_SIGNALED;
513 if (down_interruptible(&rdma->sq_sem)) {
518 /* Mark request as `sent' *before* we actually send it,
519 * because doing if after could erase the REQ_STATUS_RCVD
520 * status in case of a very fast reply.
522 req->status = REQ_STATUS_SENT;
523 err = ib_post_send(rdma->qp, &wr, &bad_wr);
530 /* Handle errors that happened during or while preparing the send: */
532 req->status = REQ_STATUS_ERROR;
534 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
537 * We did recv_post(), but not send. We have one recv_post in excess.
539 atomic_inc(&rdma->excess_rc);
542 /* Handle errors that happened during or while preparing post_recv(): */
545 spin_lock_irqsave(&rdma->req_lock, flags);
546 if (rdma->state < P9_RDMA_CLOSING) {
547 rdma->state = P9_RDMA_CLOSING;
548 spin_unlock_irqrestore(&rdma->req_lock, flags);
549 rdma_disconnect(rdma->cm_id);
551 spin_unlock_irqrestore(&rdma->req_lock, flags);
555 static void rdma_close(struct p9_client *client)
557 struct p9_trans_rdma *rdma;
562 rdma = client->trans;
566 client->status = Disconnected;
567 rdma_disconnect(rdma->cm_id);
568 rdma_destroy_trans(rdma);
572 * alloc_rdma - Allocate and initialize the rdma transport structure
573 * @opts: Mount options structure
575 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
577 struct p9_trans_rdma *rdma;
579 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
583 rdma->port = opts->port;
584 rdma->privport = opts->privport;
585 rdma->sq_depth = opts->sq_depth;
586 rdma->rq_depth = opts->rq_depth;
587 rdma->timeout = opts->timeout;
588 spin_lock_init(&rdma->req_lock);
589 init_completion(&rdma->cm_done);
590 sema_init(&rdma->sq_sem, rdma->sq_depth);
591 sema_init(&rdma->rq_sem, rdma->rq_depth);
592 atomic_set(&rdma->excess_rc, 0);
597 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
599 /* Nothing to do here.
600 * We will take care of it (if we have to) in rdma_cancelled()
605 /* A request has been fully flushed without a reply.
606 * That means we have posted one buffer in excess.
608 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
610 struct p9_trans_rdma *rdma = client->trans;
611 atomic_inc(&rdma->excess_rc);
615 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
617 struct sockaddr_in cl = {
618 .sin_family = AF_INET,
619 .sin_addr.s_addr = htonl(INADDR_ANY),
621 int port, err = -EINVAL;
623 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
624 cl.sin_port = htons((ushort)port);
625 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
626 if (err != -EADDRINUSE)
633 * rdma_create_trans - Transport method for creating a transport instance
634 * @client: client instance
635 * @addr: IP address string
636 * @args: Mount options string
639 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
642 struct p9_rdma_opts opts;
643 struct p9_trans_rdma *rdma;
644 struct rdma_conn_param conn_param;
645 struct ib_qp_init_attr qp_attr;
647 /* Parse the transport specific mount options */
648 err = parse_opts(args, &opts);
652 /* Create and initialize the RDMA transport structure */
653 rdma = alloc_rdma(&opts);
657 /* Create the RDMA CM ID */
658 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
659 RDMA_PS_TCP, IB_QPT_RC);
660 if (IS_ERR(rdma->cm_id))
663 /* Associate the client with the transport */
664 client->trans = rdma;
666 /* Bind to a privileged port if we need to */
668 err = p9_rdma_bind_privport(rdma);
670 pr_err("%s (%d): problem binding to privport: %d\n",
671 __func__, task_pid_nr(current), -err);
676 /* Resolve the server's address */
677 rdma->addr.sin_family = AF_INET;
678 rdma->addr.sin_addr.s_addr = in_aton(addr);
679 rdma->addr.sin_port = htons(opts.port);
680 err = rdma_resolve_addr(rdma->cm_id, NULL,
681 (struct sockaddr *)&rdma->addr,
685 err = wait_for_completion_interruptible(&rdma->cm_done);
686 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
689 /* Resolve the route to the server */
690 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
693 err = wait_for_completion_interruptible(&rdma->cm_done);
694 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
697 /* Create the Completion Queue */
698 rdma->cq = ib_alloc_cq(rdma->cm_id->device, client,
699 opts.sq_depth + opts.rq_depth + 1,
701 if (IS_ERR(rdma->cq))
704 /* Create the Protection Domain */
705 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
706 if (IS_ERR(rdma->pd))
709 /* Create the Queue Pair */
710 memset(&qp_attr, 0, sizeof qp_attr);
711 qp_attr.event_handler = qp_event_handler;
712 qp_attr.qp_context = client;
713 qp_attr.cap.max_send_wr = opts.sq_depth;
714 qp_attr.cap.max_recv_wr = opts.rq_depth;
715 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
716 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
717 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
718 qp_attr.qp_type = IB_QPT_RC;
719 qp_attr.send_cq = rdma->cq;
720 qp_attr.recv_cq = rdma->cq;
721 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
724 rdma->qp = rdma->cm_id->qp;
726 /* Request a connection */
727 memset(&conn_param, 0, sizeof(conn_param));
728 conn_param.private_data = NULL;
729 conn_param.private_data_len = 0;
730 conn_param.responder_resources = P9_RDMA_IRD;
731 conn_param.initiator_depth = P9_RDMA_ORD;
732 err = rdma_connect(rdma->cm_id, &conn_param);
735 err = wait_for_completion_interruptible(&rdma->cm_done);
736 if (err || (rdma->state != P9_RDMA_CONNECTED))
739 client->status = Connected;
744 rdma_destroy_trans(rdma);
748 static struct p9_trans_module p9_rdma_trans = {
750 .maxsize = P9_RDMA_MAXSIZE,
752 .owner = THIS_MODULE,
753 .create = rdma_create_trans,
755 .request = rdma_request,
756 .cancel = rdma_cancel,
757 .cancelled = rdma_cancelled,
758 .show_options = p9_rdma_show_options,
762 * p9_trans_rdma_init - Register the 9P RDMA transport driver
764 static int __init p9_trans_rdma_init(void)
766 v9fs_register_trans(&p9_rdma_trans);
770 static void __exit p9_trans_rdma_exit(void)
772 v9fs_unregister_trans(&p9_rdma_trans);
775 module_init(p9_trans_rdma_init);
776 module_exit(p9_trans_rdma_exit);
778 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
779 MODULE_DESCRIPTION("RDMA Transport for 9P");
780 MODULE_LICENSE("Dual BSD/GPL");