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;
101 struct semaphore sq_sem;
103 struct semaphore rq_sem;
105 struct sockaddr_in addr;
108 struct completion cm_done;
112 * p9_rdma_context - Keeps track of in-process WR
114 * @wc_op: The original WR op for when the CQE completes in error.
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 {
121 enum ib_wc_opcode wc_op;
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 {
145 * Option Parsing (code inspired by NFS code)
148 /* Options that take integer arguments */
149 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
152 static match_table_t tokens = {
153 {Opt_port, "port=%u"},
154 {Opt_sq_depth, "sq=%u"},
155 {Opt_rq_depth, "rq=%u"},
156 {Opt_timeout, "timeout=%u"},
161 * parse_opts - parse mount options into rdma options structure
162 * @params: options string passed from mount
163 * @opts: rdma transport-specific structure to parse options into
165 * Returns 0 upon success, -ERRNO upon failure
167 static int parse_opts(char *params, struct p9_rdma_opts *opts)
170 substring_t args[MAX_OPT_ARGS];
172 char *options, *tmp_options;
174 opts->port = P9_PORT;
175 opts->sq_depth = P9_RDMA_SQ_DEPTH;
176 opts->rq_depth = P9_RDMA_RQ_DEPTH;
177 opts->timeout = P9_RDMA_TIMEOUT;
182 tmp_options = kstrdup(params, GFP_KERNEL);
184 p9_debug(P9_DEBUG_ERROR,
185 "failed to allocate copy of option string\n");
188 options = tmp_options;
190 while ((p = strsep(&options, ",")) != NULL) {
195 token = match_token(p, tokens, args);
196 r = match_int(&args[0], &option);
198 p9_debug(P9_DEBUG_ERROR,
199 "integer field, but no integer?\n");
207 opts->sq_depth = option;
210 opts->rq_depth = option;
213 opts->timeout = option;
219 /* RQ must be at least as large as the SQ */
220 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
226 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
228 struct p9_client *c = id->context;
229 struct p9_trans_rdma *rdma = c->trans;
230 switch (event->event) {
231 case RDMA_CM_EVENT_ADDR_RESOLVED:
232 BUG_ON(rdma->state != P9_RDMA_INIT);
233 rdma->state = P9_RDMA_ADDR_RESOLVED;
236 case RDMA_CM_EVENT_ROUTE_RESOLVED:
237 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
238 rdma->state = P9_RDMA_ROUTE_RESOLVED;
241 case RDMA_CM_EVENT_ESTABLISHED:
242 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
243 rdma->state = P9_RDMA_CONNECTED;
246 case RDMA_CM_EVENT_DISCONNECTED:
248 rdma->state = P9_RDMA_CLOSED;
250 c->status = Disconnected;
253 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
256 case RDMA_CM_EVENT_ADDR_CHANGE:
257 case RDMA_CM_EVENT_ROUTE_ERROR:
258 case RDMA_CM_EVENT_DEVICE_REMOVAL:
259 case RDMA_CM_EVENT_MULTICAST_JOIN:
260 case RDMA_CM_EVENT_MULTICAST_ERROR:
261 case RDMA_CM_EVENT_REJECTED:
262 case RDMA_CM_EVENT_CONNECT_REQUEST:
263 case RDMA_CM_EVENT_CONNECT_RESPONSE:
264 case RDMA_CM_EVENT_CONNECT_ERROR:
265 case RDMA_CM_EVENT_ADDR_ERROR:
266 case RDMA_CM_EVENT_UNREACHABLE:
267 c->status = Disconnected;
268 rdma_disconnect(rdma->cm_id);
273 complete(&rdma->cm_done);
278 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
279 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
281 struct p9_req_t *req;
286 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
289 if (status != IB_WC_SUCCESS)
292 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
296 req = p9_tag_lookup(client, tag);
300 /* Check that we have not yet received a reply for this request.
302 if (unlikely(req->rc)) {
303 pr_err("Duplicate reply for request %d", tag);
308 p9_client_cb(client, req, REQ_STATUS_RCVD);
313 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
314 rdma->state = P9_RDMA_FLUSHING;
315 client->status = Disconnected;
319 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
320 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
322 ib_dma_unmap_single(rdma->cm_id->device,
323 c->busa, c->req->tc->size,
327 static void qp_event_handler(struct ib_event *event, void *context)
329 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
330 event->event, context);
333 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
335 struct p9_client *client = cq_context;
336 struct p9_trans_rdma *rdma = client->trans;
340 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
341 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
342 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
346 handle_recv(client, rdma, c, wc.status, wc.byte_len);
351 handle_send(client, rdma, c, wc.status, wc.byte_len);
356 pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
357 c->wc_op, wc.opcode, wc.status);
364 static void cq_event_handler(struct ib_event *e, void *v)
366 p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
369 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
374 if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
375 ib_dereg_mr(rdma->dma_mr);
377 if (rdma->qp && !IS_ERR(rdma->qp))
378 ib_destroy_qp(rdma->qp);
380 if (rdma->pd && !IS_ERR(rdma->pd))
381 ib_dealloc_pd(rdma->pd);
383 if (rdma->cq && !IS_ERR(rdma->cq))
384 ib_destroy_cq(rdma->cq);
386 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
387 rdma_destroy_id(rdma->cm_id);
393 post_recv(struct p9_client *client, struct p9_rdma_context *c)
395 struct p9_trans_rdma *rdma = client->trans;
396 struct ib_recv_wr wr, *bad_wr;
399 c->busa = ib_dma_map_single(rdma->cm_id->device,
400 c->rc->sdata, client->msize,
402 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
406 sge.length = client->msize;
407 sge.lkey = rdma->lkey;
410 c->wc_op = IB_WC_RECV;
411 wr.wr_id = (unsigned long) c;
414 return ib_post_recv(rdma->qp, &wr, &bad_wr);
417 p9_debug(P9_DEBUG_ERROR, "EIO\n");
421 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
423 struct p9_trans_rdma *rdma = client->trans;
424 struct ib_send_wr wr, *bad_wr;
428 struct p9_rdma_context *c = NULL;
429 struct p9_rdma_context *rpl_context = NULL;
431 /* When an error occurs between posting the recv and the send,
432 * there will be a receive context posted without a pending request.
433 * Since there is no way to "un-post" it, we remember it and skip
434 * post_recv() for the next request.
436 * see if we are this `next request' and need to absorb an excess rc.
437 * If yes, then drop and free our own, and do not recv_post().
439 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
440 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
444 goto dont_need_post_recv;
446 /* We raced and lost. */
447 atomic_inc(&rdma->excess_rc);
451 /* Allocate an fcall for the reply */
452 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
457 rpl_context->rc = req->rc;
460 * Post a receive buffer for this request. We need to ensure
461 * there is a reply buffer available for every outstanding
462 * request. A flushed request can result in no reply for an
463 * outstanding request, so we must keep a count to avoid
464 * overflowing the RQ.
466 if (down_interruptible(&rdma->rq_sem)) {
471 err = post_recv(client, rpl_context);
473 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
476 /* remove posted receive buffer from request structure */
480 /* Post the request */
481 c = kmalloc(sizeof *c, GFP_NOFS);
488 c->busa = ib_dma_map_single(rdma->cm_id->device,
489 c->req->tc->sdata, c->req->tc->size,
491 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
497 sge.length = c->req->tc->size;
498 sge.lkey = rdma->lkey;
501 c->wc_op = IB_WC_SEND;
502 wr.wr_id = (unsigned long) c;
503 wr.opcode = IB_WR_SEND;
504 wr.send_flags = IB_SEND_SIGNALED;
508 if (down_interruptible(&rdma->sq_sem)) {
513 err = ib_post_send(rdma->qp, &wr, &bad_wr);
520 /* Handle errors that happened during or while preparing the send: */
523 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
526 * We did recv_post(), but not send. We have one recv_post in excess.
528 atomic_inc(&rdma->excess_rc);
531 /* Handle errors that happened during or while preparing post_recv(): */
534 spin_lock_irqsave(&rdma->req_lock, flags);
535 if (rdma->state < P9_RDMA_CLOSING) {
536 rdma->state = P9_RDMA_CLOSING;
537 spin_unlock_irqrestore(&rdma->req_lock, flags);
538 rdma_disconnect(rdma->cm_id);
540 spin_unlock_irqrestore(&rdma->req_lock, flags);
544 static void rdma_close(struct p9_client *client)
546 struct p9_trans_rdma *rdma;
551 rdma = client->trans;
555 client->status = Disconnected;
556 rdma_disconnect(rdma->cm_id);
557 rdma_destroy_trans(rdma);
561 * alloc_rdma - Allocate and initialize the rdma transport structure
562 * @opts: Mount options structure
564 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
566 struct p9_trans_rdma *rdma;
568 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
572 rdma->sq_depth = opts->sq_depth;
573 rdma->rq_depth = opts->rq_depth;
574 rdma->timeout = opts->timeout;
575 spin_lock_init(&rdma->req_lock);
576 init_completion(&rdma->cm_done);
577 sema_init(&rdma->sq_sem, rdma->sq_depth);
578 sema_init(&rdma->rq_sem, rdma->rq_depth);
579 atomic_set(&rdma->excess_rc, 0);
584 /* its not clear to me we can do anything after send has been posted */
585 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
591 * trans_create_rdma - Transport method for creating atransport instance
592 * @client: client instance
593 * @addr: IP address string
594 * @args: Mount options string
597 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
600 struct p9_rdma_opts opts;
601 struct p9_trans_rdma *rdma;
602 struct rdma_conn_param conn_param;
603 struct ib_qp_init_attr qp_attr;
604 struct ib_device_attr devattr;
606 /* Parse the transport specific mount options */
607 err = parse_opts(args, &opts);
611 /* Create and initialize the RDMA transport structure */
612 rdma = alloc_rdma(&opts);
616 /* Create the RDMA CM ID */
617 rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
619 if (IS_ERR(rdma->cm_id))
622 /* Associate the client with the transport */
623 client->trans = rdma;
625 /* Resolve the server's address */
626 rdma->addr.sin_family = AF_INET;
627 rdma->addr.sin_addr.s_addr = in_aton(addr);
628 rdma->addr.sin_port = htons(opts.port);
629 err = rdma_resolve_addr(rdma->cm_id, NULL,
630 (struct sockaddr *)&rdma->addr,
634 err = wait_for_completion_interruptible(&rdma->cm_done);
635 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
638 /* Resolve the route to the server */
639 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
642 err = wait_for_completion_interruptible(&rdma->cm_done);
643 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
646 /* Query the device attributes */
647 err = ib_query_device(rdma->cm_id->device, &devattr);
651 /* Create the Completion Queue */
652 rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
653 cq_event_handler, client,
654 opts.sq_depth + opts.rq_depth + 1, 0);
655 if (IS_ERR(rdma->cq))
657 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
659 /* Create the Protection Domain */
660 rdma->pd = ib_alloc_pd(rdma->cm_id->device);
661 if (IS_ERR(rdma->pd))
664 /* Cache the DMA lkey in the transport */
666 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
667 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
669 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
670 if (IS_ERR(rdma->dma_mr))
672 rdma->lkey = rdma->dma_mr->lkey;
675 /* Create the Queue Pair */
676 memset(&qp_attr, 0, sizeof qp_attr);
677 qp_attr.event_handler = qp_event_handler;
678 qp_attr.qp_context = client;
679 qp_attr.cap.max_send_wr = opts.sq_depth;
680 qp_attr.cap.max_recv_wr = opts.rq_depth;
681 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
682 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
683 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
684 qp_attr.qp_type = IB_QPT_RC;
685 qp_attr.send_cq = rdma->cq;
686 qp_attr.recv_cq = rdma->cq;
687 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
690 rdma->qp = rdma->cm_id->qp;
692 /* Request a connection */
693 memset(&conn_param, 0, sizeof(conn_param));
694 conn_param.private_data = NULL;
695 conn_param.private_data_len = 0;
696 conn_param.responder_resources = P9_RDMA_IRD;
697 conn_param.initiator_depth = P9_RDMA_ORD;
698 err = rdma_connect(rdma->cm_id, &conn_param);
701 err = wait_for_completion_interruptible(&rdma->cm_done);
702 if (err || (rdma->state != P9_RDMA_CONNECTED))
705 client->status = Connected;
710 rdma_destroy_trans(rdma);
714 static struct p9_trans_module p9_rdma_trans = {
716 .maxsize = P9_RDMA_MAXSIZE,
718 .owner = THIS_MODULE,
719 .create = rdma_create_trans,
721 .request = rdma_request,
722 .cancel = rdma_cancel,
726 * p9_trans_rdma_init - Register the 9P RDMA transport driver
728 static int __init p9_trans_rdma_init(void)
730 v9fs_register_trans(&p9_rdma_trans);
734 static void __exit p9_trans_rdma_exit(void)
736 v9fs_unregister_trans(&p9_rdma_trans);
739 module_init(p9_trans_rdma_init);
740 module_exit(p9_trans_rdma_exit);
742 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
743 MODULE_DESCRIPTION("RDMA Transport for 9P");
744 MODULE_LICENSE("Dual BSD/GPL");