2 * linux/fs/9p/trans_rdma.c
4 * RDMA transport layer based on the trans_fd.c implementation.
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 * @wc_op: The original WR op for when the CQE completes in error.
113 * @busa: Bus address to unmap when the WR completes
114 * @req: Keeps track of requests (send)
115 * @rc: Keepts track of replies (receive)
118 struct p9_rdma_context {
119 enum ib_wc_opcode wc_op;
122 struct p9_req_t *req;
128 * p9_rdma_opts - Collection of mount options
129 * @port: port of connection
130 * @sq_depth: The requested depth of the SQ. This really doesn't need
131 * to be any deeper than the number of threads used in the client
132 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
133 * @timeout: Time to wait in msecs for CM events
135 struct p9_rdma_opts {
144 * Option Parsing (code inspired by NFS code)
147 /* Options that take integer arguments */
148 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
149 /* Options that take no argument */
154 static match_table_t tokens = {
155 {Opt_port, "port=%u"},
156 {Opt_sq_depth, "sq=%u"},
157 {Opt_rq_depth, "rq=%u"},
158 {Opt_timeout, "timeout=%u"},
159 {Opt_privport, "privport"},
164 * parse_opts - parse mount options into rdma options structure
165 * @params: options string passed from mount
166 * @opts: rdma transport-specific structure to parse options into
168 * Returns 0 upon success, -ERRNO upon failure
170 static int parse_opts(char *params, struct p9_rdma_opts *opts)
173 substring_t args[MAX_OPT_ARGS];
175 char *options, *tmp_options;
177 opts->port = P9_PORT;
178 opts->sq_depth = P9_RDMA_SQ_DEPTH;
179 opts->rq_depth = P9_RDMA_RQ_DEPTH;
180 opts->timeout = P9_RDMA_TIMEOUT;
186 tmp_options = kstrdup(params, GFP_KERNEL);
188 p9_debug(P9_DEBUG_ERROR,
189 "failed to allocate copy of option string\n");
192 options = tmp_options;
194 while ((p = strsep(&options, ",")) != NULL) {
199 token = match_token(p, tokens, args);
200 if ((token != Opt_err) && (token != Opt_privport)) {
201 r = match_int(&args[0], &option);
203 p9_debug(P9_DEBUG_ERROR,
204 "integer field, but no integer?\n");
213 opts->sq_depth = option;
216 opts->rq_depth = option;
219 opts->timeout = option;
228 /* RQ must be at least as large as the SQ */
229 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
235 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
237 struct p9_client *c = id->context;
238 struct p9_trans_rdma *rdma = c->trans;
239 switch (event->event) {
240 case RDMA_CM_EVENT_ADDR_RESOLVED:
241 BUG_ON(rdma->state != P9_RDMA_INIT);
242 rdma->state = P9_RDMA_ADDR_RESOLVED;
245 case RDMA_CM_EVENT_ROUTE_RESOLVED:
246 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
247 rdma->state = P9_RDMA_ROUTE_RESOLVED;
250 case RDMA_CM_EVENT_ESTABLISHED:
251 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
252 rdma->state = P9_RDMA_CONNECTED;
255 case RDMA_CM_EVENT_DISCONNECTED:
257 rdma->state = P9_RDMA_CLOSED;
259 c->status = Disconnected;
262 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
265 case RDMA_CM_EVENT_ADDR_CHANGE:
266 case RDMA_CM_EVENT_ROUTE_ERROR:
267 case RDMA_CM_EVENT_DEVICE_REMOVAL:
268 case RDMA_CM_EVENT_MULTICAST_JOIN:
269 case RDMA_CM_EVENT_MULTICAST_ERROR:
270 case RDMA_CM_EVENT_REJECTED:
271 case RDMA_CM_EVENT_CONNECT_REQUEST:
272 case RDMA_CM_EVENT_CONNECT_RESPONSE:
273 case RDMA_CM_EVENT_CONNECT_ERROR:
274 case RDMA_CM_EVENT_ADDR_ERROR:
275 case RDMA_CM_EVENT_UNREACHABLE:
276 c->status = Disconnected;
277 rdma_disconnect(rdma->cm_id);
282 complete(&rdma->cm_done);
287 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
288 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
290 struct p9_req_t *req;
295 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
298 if (status != IB_WC_SUCCESS)
301 err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
305 req = p9_tag_lookup(client, tag);
309 /* Check that we have not yet received a reply for this request.
311 if (unlikely(req->rc)) {
312 pr_err("Duplicate reply for request %d", tag);
317 p9_client_cb(client, req, REQ_STATUS_RCVD);
322 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
323 rdma->state = P9_RDMA_FLUSHING;
324 client->status = Disconnected;
328 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
329 struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
331 ib_dma_unmap_single(rdma->cm_id->device,
332 c->busa, c->req->tc->size,
336 static void qp_event_handler(struct ib_event *event, void *context)
338 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
339 event->event, context);
342 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
344 struct p9_client *client = cq_context;
345 struct p9_trans_rdma *rdma = client->trans;
349 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
350 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
351 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
355 handle_recv(client, rdma, c, wc.status, wc.byte_len);
360 handle_send(client, rdma, c, wc.status, wc.byte_len);
365 pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
366 c->wc_op, wc.opcode, wc.status);
373 static void cq_event_handler(struct ib_event *e, void *v)
375 p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
378 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
383 if (rdma->qp && !IS_ERR(rdma->qp))
384 ib_destroy_qp(rdma->qp);
386 if (rdma->pd && !IS_ERR(rdma->pd))
387 ib_dealloc_pd(rdma->pd);
389 if (rdma->cq && !IS_ERR(rdma->cq))
390 ib_destroy_cq(rdma->cq);
392 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
393 rdma_destroy_id(rdma->cm_id);
399 post_recv(struct p9_client *client, struct p9_rdma_context *c)
401 struct p9_trans_rdma *rdma = client->trans;
402 struct ib_recv_wr wr, *bad_wr;
405 c->busa = ib_dma_map_single(rdma->cm_id->device,
406 c->rc->sdata, client->msize,
408 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
412 sge.length = client->msize;
413 sge.lkey = rdma->pd->local_dma_lkey;
416 c->wc_op = IB_WC_RECV;
417 wr.wr_id = (unsigned long) c;
420 return ib_post_recv(rdma->qp, &wr, &bad_wr);
423 p9_debug(P9_DEBUG_ERROR, "EIO\n");
427 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
429 struct p9_trans_rdma *rdma = client->trans;
430 struct ib_send_wr wr, *bad_wr;
434 struct p9_rdma_context *c = NULL;
435 struct p9_rdma_context *rpl_context = NULL;
437 /* When an error occurs between posting the recv and the send,
438 * there will be a receive context posted without a pending request.
439 * Since there is no way to "un-post" it, we remember it and skip
440 * post_recv() for the next request.
442 * see if we are this `next request' and need to absorb an excess rc.
443 * If yes, then drop and free our own, and do not recv_post().
445 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
446 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
450 goto dont_need_post_recv;
452 /* We raced and lost. */
453 atomic_inc(&rdma->excess_rc);
457 /* Allocate an fcall for the reply */
458 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
463 rpl_context->rc = req->rc;
466 * Post a receive buffer for this request. We need to ensure
467 * there is a reply buffer available for every outstanding
468 * request. A flushed request can result in no reply for an
469 * outstanding request, so we must keep a count to avoid
470 * overflowing the RQ.
472 if (down_interruptible(&rdma->rq_sem)) {
477 err = post_recv(client, rpl_context);
479 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
482 /* remove posted receive buffer from request structure */
486 /* Post the request */
487 c = kmalloc(sizeof *c, GFP_NOFS);
494 c->busa = ib_dma_map_single(rdma->cm_id->device,
495 c->req->tc->sdata, c->req->tc->size,
497 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
503 sge.length = c->req->tc->size;
504 sge.lkey = rdma->pd->local_dma_lkey;
507 c->wc_op = IB_WC_SEND;
508 wr.wr_id = (unsigned long) c;
509 wr.opcode = IB_WR_SEND;
510 wr.send_flags = IB_SEND_SIGNALED;
514 if (down_interruptible(&rdma->sq_sem)) {
519 /* Mark request as `sent' *before* we actually send it,
520 * because doing if after could erase the REQ_STATUS_RCVD
521 * status in case of a very fast reply.
523 req->status = REQ_STATUS_SENT;
524 err = ib_post_send(rdma->qp, &wr, &bad_wr);
531 /* Handle errors that happened during or while preparing the send: */
533 req->status = REQ_STATUS_ERROR;
535 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
538 * We did recv_post(), but not send. We have one recv_post in excess.
540 atomic_inc(&rdma->excess_rc);
543 /* Handle errors that happened during or while preparing post_recv(): */
546 spin_lock_irqsave(&rdma->req_lock, flags);
547 if (rdma->state < P9_RDMA_CLOSING) {
548 rdma->state = P9_RDMA_CLOSING;
549 spin_unlock_irqrestore(&rdma->req_lock, flags);
550 rdma_disconnect(rdma->cm_id);
552 spin_unlock_irqrestore(&rdma->req_lock, flags);
556 static void rdma_close(struct p9_client *client)
558 struct p9_trans_rdma *rdma;
563 rdma = client->trans;
567 client->status = Disconnected;
568 rdma_disconnect(rdma->cm_id);
569 rdma_destroy_trans(rdma);
573 * alloc_rdma - Allocate and initialize the rdma transport structure
574 * @opts: Mount options structure
576 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
578 struct p9_trans_rdma *rdma;
580 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
584 rdma->sq_depth = opts->sq_depth;
585 rdma->rq_depth = opts->rq_depth;
586 rdma->timeout = opts->timeout;
587 spin_lock_init(&rdma->req_lock);
588 init_completion(&rdma->cm_done);
589 sema_init(&rdma->sq_sem, rdma->sq_depth);
590 sema_init(&rdma->rq_sem, rdma->rq_depth);
591 atomic_set(&rdma->excess_rc, 0);
596 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
598 /* Nothing to do here.
599 * We will take care of it (if we have to) in rdma_cancelled()
604 /* A request has been fully flushed without a reply.
605 * That means we have posted one buffer in excess.
607 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
609 struct p9_trans_rdma *rdma = client->trans;
610 atomic_inc(&rdma->excess_rc);
614 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
616 struct sockaddr_in cl = {
617 .sin_family = AF_INET,
618 .sin_addr.s_addr = htonl(INADDR_ANY),
620 int port, err = -EINVAL;
622 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
623 cl.sin_port = htons((ushort)port);
624 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
625 if (err != -EADDRINUSE)
632 * trans_create_rdma - Transport method for creating atransport instance
633 * @client: client instance
634 * @addr: IP address string
635 * @args: Mount options string
638 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
641 struct p9_rdma_opts opts;
642 struct p9_trans_rdma *rdma;
643 struct rdma_conn_param conn_param;
644 struct ib_qp_init_attr qp_attr;
645 struct ib_cq_init_attr cq_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(p9_cm_event_handler, client, RDMA_PS_TCP,
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 cq_attr.cqe = opts.sq_depth + opts.rq_depth + 1;
699 rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
700 cq_event_handler, client,
702 if (IS_ERR(rdma->cq))
704 ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
706 /* Create the Protection Domain */
707 rdma->pd = ib_alloc_pd(rdma->cm_id->device);
708 if (IS_ERR(rdma->pd))
711 /* Create the Queue Pair */
712 memset(&qp_attr, 0, sizeof qp_attr);
713 qp_attr.event_handler = qp_event_handler;
714 qp_attr.qp_context = client;
715 qp_attr.cap.max_send_wr = opts.sq_depth;
716 qp_attr.cap.max_recv_wr = opts.rq_depth;
717 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
718 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
719 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
720 qp_attr.qp_type = IB_QPT_RC;
721 qp_attr.send_cq = rdma->cq;
722 qp_attr.recv_cq = rdma->cq;
723 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
726 rdma->qp = rdma->cm_id->qp;
728 /* Request a connection */
729 memset(&conn_param, 0, sizeof(conn_param));
730 conn_param.private_data = NULL;
731 conn_param.private_data_len = 0;
732 conn_param.responder_resources = P9_RDMA_IRD;
733 conn_param.initiator_depth = P9_RDMA_ORD;
734 err = rdma_connect(rdma->cm_id, &conn_param);
737 err = wait_for_completion_interruptible(&rdma->cm_done);
738 if (err || (rdma->state != P9_RDMA_CONNECTED))
741 client->status = Connected;
746 rdma_destroy_trans(rdma);
750 static struct p9_trans_module p9_rdma_trans = {
752 .maxsize = P9_RDMA_MAXSIZE,
754 .owner = THIS_MODULE,
755 .create = rdma_create_trans,
757 .request = rdma_request,
758 .cancel = rdma_cancel,
759 .cancelled = rdma_cancelled,
763 * p9_trans_rdma_init - Register the 9P RDMA transport driver
765 static int __init p9_trans_rdma_init(void)
767 v9fs_register_trans(&p9_rdma_trans);
771 static void __exit p9_trans_rdma_exit(void)
773 v9fs_unregister_trans(&p9_rdma_trans);
776 module_init(p9_trans_rdma_init);
777 module_exit(p9_trans_rdma_exit);
780 MODULE_DESCRIPTION("RDMA Transport for 9P");
781 MODULE_LICENSE("Dual BSD/GPL");