blob: af0a8a6cd3fd8f7aa280ca70b81e3c0ab861568d [file] [log] [blame]
Thomas Gleixner1f327612019-05-28 09:57:16 -07001// SPDX-License-Identifier: GPL-2.0-only
Tom Tuckerfc79d4b2008-10-22 18:47:39 -05002/*
3 * linux/fs/9p/trans_rdma.c
4 *
5 * RDMA transport layer based on the trans_fd.c implementation.
6 *
7 * Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
8 * Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
9 * Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
10 * Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
11 * Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050012 */
13
Joe Perches5d385152011-11-28 10:40:46 -080014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050016#include <linux/in.h>
17#include <linux/module.h>
18#include <linux/net.h>
19#include <linux/ipv6.h>
20#include <linux/kthread.h>
21#include <linux/errno.h>
22#include <linux/kernel.h>
23#include <linux/un.h>
24#include <linux/uaccess.h>
25#include <linux/inet.h>
26#include <linux/idr.h>
27#include <linux/file.h>
28#include <linux/parser.h>
29#include <linux/semaphore.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090030#include <linux/slab.h>
David Howellsc4fac912017-07-05 16:25:37 +010031#include <linux/seq_file.h>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050032#include <net/9p/9p.h>
33#include <net/9p/client.h>
34#include <net/9p/transport.h>
35#include <rdma/ib_verbs.h>
36#include <rdma/rdma_cm.h>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050037
38#define P9_PORT 5640
39#define P9_RDMA_SQ_DEPTH 32
40#define P9_RDMA_RQ_DEPTH 32
41#define P9_RDMA_SEND_SGE 4
42#define P9_RDMA_RECV_SGE 4
43#define P9_RDMA_IRD 0
44#define P9_RDMA_ORD 0
45#define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
Simon Derr3fcc62f2013-06-21 15:32:37 +020046#define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050047
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050048/**
49 * struct p9_trans_rdma - RDMA transport instance
50 *
51 * @state: tracks the transport state machine for connection setup and tear down
52 * @cm_id: The RDMA CM ID
53 * @pd: Protection Domain pointer
54 * @qp: Queue Pair pointer
55 * @cq: Completion Queue pointer
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050056 * @timeout: Number of uSecs to wait for connection management events
David Howellsc4fac912017-07-05 16:25:37 +010057 * @privport: Whether a privileged port may be used
58 * @port: The port to use
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050059 * @sq_depth: The depth of the Send Queue
60 * @sq_sem: Semaphore for the SQ
61 * @rq_depth: The depth of the Receive Queue.
Simon Derrfd453d02013-06-21 15:32:39 +020062 * @rq_sem: Semaphore for the RQ
Simon Derr1cff3302013-06-21 15:32:42 +020063 * @excess_rc : Amount of posted Receive Contexts without a pending request.
64 * See rdma_request()
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050065 * @addr: The remote peer's address
66 * @req_lock: Protects the active request list
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050067 * @cm_done: Completion event for connection management tracking
68 */
69struct p9_trans_rdma {
70 enum {
71 P9_RDMA_INIT,
72 P9_RDMA_ADDR_RESOLVED,
73 P9_RDMA_ROUTE_RESOLVED,
74 P9_RDMA_CONNECTED,
75 P9_RDMA_FLUSHING,
76 P9_RDMA_CLOSING,
77 P9_RDMA_CLOSED,
78 } state;
79 struct rdma_cm_id *cm_id;
80 struct ib_pd *pd;
81 struct ib_qp *qp;
82 struct ib_cq *cq;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050083 long timeout;
David Howellsc4fac912017-07-05 16:25:37 +010084 bool privport;
85 u16 port;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050086 int sq_depth;
87 struct semaphore sq_sem;
88 int rq_depth;
Simon Derrfd453d02013-06-21 15:32:39 +020089 struct semaphore rq_sem;
Simon Derr1cff3302013-06-21 15:32:42 +020090 atomic_t excess_rc;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050091 struct sockaddr_in addr;
92 spinlock_t req_lock;
93
94 struct completion cm_done;
95};
96
Andrew Lunn15e522a2020-07-13 01:14:57 +020097struct p9_rdma_req;
98
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050099/**
Andrew Lunn15e522a2020-07-13 01:14:57 +0200100 * struct p9_rdma_context - Keeps track of in-process WR
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500101 *
Andrew Lunn760b3d62020-10-31 19:26:55 +0100102 * @cqe: completion queue entry
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500103 * @busa: Bus address to unmap when the WR completes
104 * @req: Keeps track of requests (send)
105 * @rc: Keepts track of replies (receive)
106 */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500107struct p9_rdma_context {
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100108 struct ib_cqe cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500109 dma_addr_t busa;
110 union {
111 struct p9_req_t *req;
Dominique Martinet523adb62018-07-30 05:55:19 +0000112 struct p9_fcall rc;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500113 };
114};
115
116/**
Andrew Lunn15e522a2020-07-13 01:14:57 +0200117 * struct p9_rdma_opts - Collection of mount options
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500118 * @port: port of connection
Andrew Lunn760b3d62020-10-31 19:26:55 +0100119 * @privport: Whether a privileged port may be used
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500120 * @sq_depth: The requested depth of the SQ. This really doesn't need
121 * to be any deeper than the number of threads used in the client
122 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
123 * @timeout: Time to wait in msecs for CM events
124 */
125struct p9_rdma_opts {
126 short port;
David Howellsc4fac912017-07-05 16:25:37 +0100127 bool privport;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500128 int sq_depth;
129 int rq_depth;
130 long timeout;
131};
132
133/*
134 * Option Parsing (code inspired by NFS code)
135 */
136enum {
137 /* Options that take integer arguments */
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100138 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
139 /* Options that take no argument */
140 Opt_privport,
141 Opt_err,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500142};
143
144static match_table_t tokens = {
145 {Opt_port, "port=%u"},
146 {Opt_sq_depth, "sq=%u"},
147 {Opt_rq_depth, "rq=%u"},
148 {Opt_timeout, "timeout=%u"},
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100149 {Opt_privport, "privport"},
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500150 {Opt_err, NULL},
151};
152
David Howellsc4fac912017-07-05 16:25:37 +0100153static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
154{
155 struct p9_trans_rdma *rdma = clnt->trans;
156
157 if (rdma->port != P9_PORT)
158 seq_printf(m, ",port=%u", rdma->port);
159 if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
160 seq_printf(m, ",sq=%u", rdma->sq_depth);
161 if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
162 seq_printf(m, ",rq=%u", rdma->rq_depth);
163 if (rdma->timeout != P9_RDMA_TIMEOUT)
164 seq_printf(m, ",timeout=%lu", rdma->timeout);
165 if (rdma->privport)
166 seq_puts(m, ",privport");
167 return 0;
168}
169
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500170/**
Abhishek Kulkarni0e155972009-07-19 13:41:55 -0600171 * parse_opts - parse mount options into rdma options structure
172 * @params: options string passed from mount
173 * @opts: rdma transport-specific structure to parse options into
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500174 *
175 * Returns 0 upon success, -ERRNO upon failure
176 */
177static int parse_opts(char *params, struct p9_rdma_opts *opts)
178{
179 char *p;
180 substring_t args[MAX_OPT_ARGS];
181 int option;
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600182 char *options, *tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500183
184 opts->port = P9_PORT;
185 opts->sq_depth = P9_RDMA_SQ_DEPTH;
186 opts->rq_depth = P9_RDMA_RQ_DEPTH;
187 opts->timeout = P9_RDMA_TIMEOUT;
David Howellsc4fac912017-07-05 16:25:37 +0100188 opts->privport = false;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500189
190 if (!params)
191 return 0;
192
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600193 tmp_options = kstrdup(params, GFP_KERNEL);
194 if (!tmp_options) {
Joe Perches5d385152011-11-28 10:40:46 -0800195 p9_debug(P9_DEBUG_ERROR,
196 "failed to allocate copy of option string\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500197 return -ENOMEM;
198 }
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600199 options = tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500200
201 while ((p = strsep(&options, ",")) != NULL) {
202 int token;
203 int r;
204 if (!*p)
205 continue;
206 token = match_token(p, tokens, args);
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100207 if ((token != Opt_err) && (token != Opt_privport)) {
208 r = match_int(&args[0], &option);
209 if (r < 0) {
210 p9_debug(P9_DEBUG_ERROR,
211 "integer field, but no integer?\n");
212 continue;
213 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500214 }
215 switch (token) {
216 case Opt_port:
217 opts->port = option;
218 break;
219 case Opt_sq_depth:
220 opts->sq_depth = option;
221 break;
222 case Opt_rq_depth:
223 opts->rq_depth = option;
224 break;
225 case Opt_timeout:
226 opts->timeout = option;
227 break;
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100228 case Opt_privport:
David Howellsc4fac912017-07-05 16:25:37 +0100229 opts->privport = true;
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100230 break;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500231 default:
232 continue;
233 }
234 }
235 /* RQ must be at least as large as the SQ */
236 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600237 kfree(tmp_options);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500238 return 0;
239}
240
241static int
242p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
243{
244 struct p9_client *c = id->context;
245 struct p9_trans_rdma *rdma = c->trans;
246 switch (event->event) {
247 case RDMA_CM_EVENT_ADDR_RESOLVED:
248 BUG_ON(rdma->state != P9_RDMA_INIT);
249 rdma->state = P9_RDMA_ADDR_RESOLVED;
250 break;
251
252 case RDMA_CM_EVENT_ROUTE_RESOLVED:
253 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
254 rdma->state = P9_RDMA_ROUTE_RESOLVED;
255 break;
256
257 case RDMA_CM_EVENT_ESTABLISHED:
258 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
259 rdma->state = P9_RDMA_CONNECTED;
260 break;
261
262 case RDMA_CM_EVENT_DISCONNECTED:
263 if (rdma)
264 rdma->state = P9_RDMA_CLOSED;
Dominique Martinet473c7dd2018-09-08 00:26:50 +0900265 c->status = Disconnected;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500266 break;
267
268 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
269 break;
270
271 case RDMA_CM_EVENT_ADDR_CHANGE:
272 case RDMA_CM_EVENT_ROUTE_ERROR:
273 case RDMA_CM_EVENT_DEVICE_REMOVAL:
274 case RDMA_CM_EVENT_MULTICAST_JOIN:
275 case RDMA_CM_EVENT_MULTICAST_ERROR:
276 case RDMA_CM_EVENT_REJECTED:
277 case RDMA_CM_EVENT_CONNECT_REQUEST:
278 case RDMA_CM_EVENT_CONNECT_RESPONSE:
279 case RDMA_CM_EVENT_CONNECT_ERROR:
280 case RDMA_CM_EVENT_ADDR_ERROR:
281 case RDMA_CM_EVENT_UNREACHABLE:
282 c->status = Disconnected;
283 rdma_disconnect(rdma->cm_id);
284 break;
285 default:
286 BUG();
287 }
288 complete(&rdma->cm_done);
289 return 0;
290}
291
292static void
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100293recv_done(struct ib_cq *cq, struct ib_wc *wc)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500294{
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100295 struct p9_client *client = cq->cq_context;
296 struct p9_trans_rdma *rdma = client->trans;
297 struct p9_rdma_context *c =
298 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500299 struct p9_req_t *req;
300 int err = 0;
301 int16_t tag;
302
303 req = NULL;
304 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
305 DMA_FROM_DEVICE);
306
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100307 if (wc->status != IB_WC_SUCCESS)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500308 goto err_out;
309
Dominique Martinet523adb62018-07-30 05:55:19 +0000310 c->rc.size = wc->byte_len;
311 err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500312 if (err)
313 goto err_out;
314
315 req = p9_tag_lookup(client, tag);
316 if (!req)
317 goto err_out;
318
Simon Derr47229ff2013-06-21 15:32:38 +0200319 /* Check that we have not yet received a reply for this request.
320 */
Dominique Martinet523adb62018-07-30 05:55:19 +0000321 if (unlikely(req->rc.sdata)) {
Simon Derr47229ff2013-06-21 15:32:38 +0200322 pr_err("Duplicate reply for request %d", tag);
323 goto err_out;
324 }
325
Dominique Martinet523adb62018-07-30 05:55:19 +0000326 req->rc.size = c->rc.size;
327 req->rc.sdata = c->rc.sdata;
Dominique Martinet2b6e72e2014-01-17 18:31:00 +0100328 p9_client_cb(client, req, REQ_STATUS_RCVD);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500329
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100330 out:
331 up(&rdma->rq_sem);
332 kfree(c);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500333 return;
334
335 err_out:
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100336 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
337 req, err, wc->status);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500338 rdma->state = P9_RDMA_FLUSHING;
339 client->status = Disconnected;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100340 goto out;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500341}
342
343static void
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100344send_done(struct ib_cq *cq, struct ib_wc *wc)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500345{
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100346 struct p9_client *client = cq->cq_context;
347 struct p9_trans_rdma *rdma = client->trans;
348 struct p9_rdma_context *c =
349 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
350
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500351 ib_dma_unmap_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000352 c->busa, c->req->tc.size,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500353 DMA_TO_DEVICE);
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100354 up(&rdma->sq_sem);
Tomas Bortoli728356d2018-08-14 19:43:42 +0200355 p9_req_put(c->req);
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100356 kfree(c);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500357}
358
359static void qp_event_handler(struct ib_event *event, void *context)
360{
Joe Perches5d385152011-11-28 10:40:46 -0800361 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
362 event->event, context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500363}
364
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500365static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
366{
367 if (!rdma)
368 return;
369
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500370 if (rdma->qp && !IS_ERR(rdma->qp))
371 ib_destroy_qp(rdma->qp);
372
373 if (rdma->pd && !IS_ERR(rdma->pd))
374 ib_dealloc_pd(rdma->pd);
375
376 if (rdma->cq && !IS_ERR(rdma->cq))
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100377 ib_free_cq(rdma->cq);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500378
379 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
380 rdma_destroy_id(rdma->cm_id);
381
382 kfree(rdma);
383}
384
385static int
386post_recv(struct p9_client *client, struct p9_rdma_context *c)
387{
388 struct p9_trans_rdma *rdma = client->trans;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700389 struct ib_recv_wr wr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500390 struct ib_sge sge;
391
392 c->busa = ib_dma_map_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000393 c->rc.sdata, client->msize,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500394 DMA_FROM_DEVICE);
395 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
396 goto error;
397
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100398 c->cqe.done = recv_done;
399
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500400 sge.addr = c->busa;
401 sge.length = client->msize;
Jason Gunthorpe2f31fa82015-07-30 17:22:25 -0600402 sge.lkey = rdma->pd->local_dma_lkey;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500403
404 wr.next = NULL;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100405 wr.wr_cqe = &c->cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500406 wr.sg_list = &sge;
407 wr.num_sge = 1;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700408 return ib_post_recv(rdma->qp, &wr, NULL);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500409
410 error:
Joe Perches5d385152011-11-28 10:40:46 -0800411 p9_debug(P9_DEBUG_ERROR, "EIO\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500412 return -EIO;
413}
414
415static int rdma_request(struct p9_client *client, struct p9_req_t *req)
416{
417 struct p9_trans_rdma *rdma = client->trans;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700418 struct ib_send_wr wr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500419 struct ib_sge sge;
420 int err = 0;
421 unsigned long flags;
422 struct p9_rdma_context *c = NULL;
423 struct p9_rdma_context *rpl_context = NULL;
424
Simon Derr1cff3302013-06-21 15:32:42 +0200425 /* When an error occurs between posting the recv and the send,
426 * there will be a receive context posted without a pending request.
427 * Since there is no way to "un-post" it, we remember it and skip
428 * post_recv() for the next request.
429 * So here,
430 * see if we are this `next request' and need to absorb an excess rc.
431 * If yes, then drop and free our own, and do not recv_post().
432 **/
433 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
434 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
Dominique Martinet523adb62018-07-30 05:55:19 +0000435 /* Got one! */
436 p9_fcall_fini(&req->rc);
437 req->rc.sdata = NULL;
Simon Derr1cff3302013-06-21 15:32:42 +0200438 goto dont_need_post_recv;
439 } else {
440 /* We raced and lost. */
441 atomic_inc(&rdma->excess_rc);
442 }
443 }
444
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500445 /* Allocate an fcall for the reply */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530446 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000447 if (!rpl_context) {
448 err = -ENOMEM;
Simon Derr2f52d072013-06-21 15:32:41 +0200449 goto recv_error;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000450 }
Dominique Martinet523adb62018-07-30 05:55:19 +0000451 rpl_context->rc.sdata = req->rc.sdata;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500452
453 /*
454 * Post a receive buffer for this request. We need to ensure
455 * there is a reply buffer available for every outstanding
456 * request. A flushed request can result in no reply for an
457 * outstanding request, so we must keep a count to avoid
458 * overflowing the RQ.
459 */
Simon Derr2f52d072013-06-21 15:32:41 +0200460 if (down_interruptible(&rdma->rq_sem)) {
461 err = -EINTR;
462 goto recv_error;
463 }
Simon Derrfd453d02013-06-21 15:32:39 +0200464
465 err = post_recv(client, rpl_context);
466 if (err) {
Dominique Martinet8b894ad2018-08-30 19:29:36 +0900467 p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
Simon Derr2f52d072013-06-21 15:32:41 +0200468 goto recv_error;
Simon Derrfd453d02013-06-21 15:32:39 +0200469 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500470 /* remove posted receive buffer from request structure */
Dominique Martinet523adb62018-07-30 05:55:19 +0000471 req->rc.sdata = NULL;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500472
Simon Derr1cff3302013-06-21 15:32:42 +0200473dont_need_post_recv:
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500474 /* Post the request */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530475 c = kmalloc(sizeof *c, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000476 if (!c) {
477 err = -ENOMEM;
Simon Derr2f52d072013-06-21 15:32:41 +0200478 goto send_error;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000479 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500480 c->req = req;
481
482 c->busa = ib_dma_map_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000483 c->req->tc.sdata, c->req->tc.size,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500484 DMA_TO_DEVICE);
Simon Derr2f52d072013-06-21 15:32:41 +0200485 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
486 err = -EIO;
487 goto send_error;
488 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500489
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100490 c->cqe.done = send_done;
491
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500492 sge.addr = c->busa;
Dominique Martinet523adb62018-07-30 05:55:19 +0000493 sge.length = c->req->tc.size;
Jason Gunthorpe2f31fa82015-07-30 17:22:25 -0600494 sge.lkey = rdma->pd->local_dma_lkey;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500495
496 wr.next = NULL;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100497 wr.wr_cqe = &c->cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500498 wr.opcode = IB_WR_SEND;
499 wr.send_flags = IB_SEND_SIGNALED;
500 wr.sg_list = &sge;
501 wr.num_sge = 1;
502
Simon Derr2f52d072013-06-21 15:32:41 +0200503 if (down_interruptible(&rdma->sq_sem)) {
504 err = -EINTR;
505 goto send_error;
506 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500507
Simon Derr3f9d5b82014-03-10 16:38:50 +0100508 /* Mark request as `sent' *before* we actually send it,
509 * because doing if after could erase the REQ_STATUS_RCVD
510 * status in case of a very fast reply.
511 */
512 req->status = REQ_STATUS_SENT;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700513 err = ib_post_send(rdma->qp, &wr, NULL);
Simon Derr2f52d072013-06-21 15:32:41 +0200514 if (err)
515 goto send_error;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500516
Simon Derr2f52d072013-06-21 15:32:41 +0200517 /* Success */
518 return 0;
519
520 /* Handle errors that happened during or while preparing the send: */
521 send_error:
Simon Derr3f9d5b82014-03-10 16:38:50 +0100522 req->status = REQ_STATUS_ERROR;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000523 kfree(c);
Simon Derr2f52d072013-06-21 15:32:41 +0200524 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
Simon Derr1cff3302013-06-21 15:32:42 +0200525
526 /* Ach.
527 * We did recv_post(), but not send. We have one recv_post in excess.
528 */
529 atomic_inc(&rdma->excess_rc);
Simon Derr2f52d072013-06-21 15:32:41 +0200530 return err;
531
532 /* Handle errors that happened during or while preparing post_recv(): */
533 recv_error:
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000534 kfree(rpl_context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500535 spin_lock_irqsave(&rdma->req_lock, flags);
Dominique Martinet8b894ad2018-08-30 19:29:36 +0900536 if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500537 rdma->state = P9_RDMA_CLOSING;
538 spin_unlock_irqrestore(&rdma->req_lock, flags);
539 rdma_disconnect(rdma->cm_id);
540 } else
541 spin_unlock_irqrestore(&rdma->req_lock, flags);
542 return err;
543}
544
545static void rdma_close(struct p9_client *client)
546{
547 struct p9_trans_rdma *rdma;
548
549 if (!client)
550 return;
551
552 rdma = client->trans;
553 if (!rdma)
554 return;
555
556 client->status = Disconnected;
557 rdma_disconnect(rdma->cm_id);
558 rdma_destroy_trans(rdma);
559}
560
561/**
562 * alloc_rdma - Allocate and initialize the rdma transport structure
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500563 * @opts: Mount options structure
564 */
565static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
566{
567 struct p9_trans_rdma *rdma;
568
569 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
570 if (!rdma)
571 return NULL;
572
David Howellsc4fac912017-07-05 16:25:37 +0100573 rdma->port = opts->port;
574 rdma->privport = opts->privport;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500575 rdma->sq_depth = opts->sq_depth;
576 rdma->rq_depth = opts->rq_depth;
577 rdma->timeout = opts->timeout;
578 spin_lock_init(&rdma->req_lock);
579 init_completion(&rdma->cm_done);
580 sema_init(&rdma->sq_sem, rdma->sq_depth);
Simon Derrfd453d02013-06-21 15:32:39 +0200581 sema_init(&rdma->rq_sem, rdma->rq_depth);
Simon Derr1cff3302013-06-21 15:32:42 +0200582 atomic_set(&rdma->excess_rc, 0);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500583
584 return rdma;
585}
586
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500587static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
588{
Simon Derr931700d2014-03-10 16:38:51 +0100589 /* Nothing to do here.
590 * We will take care of it (if we have to) in rdma_cancelled()
591 */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500592 return 1;
593}
594
Simon Derr931700d2014-03-10 16:38:51 +0100595/* A request has been fully flushed without a reply.
596 * That means we have posted one buffer in excess.
597 */
598static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
599{
600 struct p9_trans_rdma *rdma = client->trans;
601 atomic_inc(&rdma->excess_rc);
602 return 0;
603}
604
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100605static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
606{
607 struct sockaddr_in cl = {
608 .sin_family = AF_INET,
609 .sin_addr.s_addr = htonl(INADDR_ANY),
610 };
611 int port, err = -EINVAL;
612
613 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
614 cl.sin_port = htons((ushort)port);
615 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
616 if (err != -EADDRINUSE)
617 break;
618 }
619 return err;
620}
621
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500622/**
Sun Lianwen4a026da2018-05-08 09:49:38 +0800623 * rdma_create_trans - Transport method for creating a transport instance
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500624 * @client: client instance
625 * @addr: IP address string
626 * @args: Mount options string
627 */
628static int
629rdma_create_trans(struct p9_client *client, const char *addr, char *args)
630{
631 int err;
632 struct p9_rdma_opts opts;
633 struct p9_trans_rdma *rdma;
634 struct rdma_conn_param conn_param;
635 struct ib_qp_init_attr qp_attr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500636
Tomas Bortoli10aa1452018-07-27 13:05:58 +0200637 if (addr == NULL)
638 return -EINVAL;
639
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500640 /* Parse the transport specific mount options */
641 err = parse_opts(args, &opts);
642 if (err < 0)
643 return err;
644
645 /* Create and initialize the RDMA transport structure */
646 rdma = alloc_rdma(&opts);
647 if (!rdma)
648 return -ENOMEM;
649
650 /* Create the RDMA CM ID */
Guy Shapirofa201052015-10-22 15:20:10 +0300651 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
652 RDMA_PS_TCP, IB_QPT_RC);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500653 if (IS_ERR(rdma->cm_id))
654 goto error;
655
Tom Tucker517ac452008-10-23 16:30:13 -0500656 /* Associate the client with the transport */
657 client->trans = rdma;
658
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100659 /* Bind to a privileged port if we need to */
660 if (opts.privport) {
661 err = p9_rdma_bind_privport(rdma);
662 if (err < 0) {
663 pr_err("%s (%d): problem binding to privport: %d\n",
664 __func__, task_pid_nr(current), -err);
665 goto error;
666 }
667 }
668
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500669 /* Resolve the server's address */
670 rdma->addr.sin_family = AF_INET;
671 rdma->addr.sin_addr.s_addr = in_aton(addr);
672 rdma->addr.sin_port = htons(opts.port);
673 err = rdma_resolve_addr(rdma->cm_id, NULL,
674 (struct sockaddr *)&rdma->addr,
675 rdma->timeout);
676 if (err)
677 goto error;
678 err = wait_for_completion_interruptible(&rdma->cm_done);
679 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
680 goto error;
681
682 /* Resolve the route to the server */
683 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
684 if (err)
685 goto error;
686 err = wait_for_completion_interruptible(&rdma->cm_done);
687 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
688 goto error;
689
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500690 /* Create the Completion Queue */
Chuck Lever20cf4e02019-07-29 13:22:09 -0400691 rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
692 opts.sq_depth + opts.rq_depth + 1,
693 IB_POLL_SOFTIRQ);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500694 if (IS_ERR(rdma->cq))
695 goto error;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500696
697 /* Create the Protection Domain */
Christoph Hellwiged082d32016-09-05 12:56:17 +0200698 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500699 if (IS_ERR(rdma->pd))
700 goto error;
701
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500702 /* Create the Queue Pair */
703 memset(&qp_attr, 0, sizeof qp_attr);
704 qp_attr.event_handler = qp_event_handler;
705 qp_attr.qp_context = client;
706 qp_attr.cap.max_send_wr = opts.sq_depth;
707 qp_attr.cap.max_recv_wr = opts.rq_depth;
708 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
709 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
710 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
711 qp_attr.qp_type = IB_QPT_RC;
712 qp_attr.send_cq = rdma->cq;
713 qp_attr.recv_cq = rdma->cq;
714 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
715 if (err)
716 goto error;
717 rdma->qp = rdma->cm_id->qp;
718
719 /* Request a connection */
720 memset(&conn_param, 0, sizeof(conn_param));
721 conn_param.private_data = NULL;
722 conn_param.private_data_len = 0;
723 conn_param.responder_resources = P9_RDMA_IRD;
724 conn_param.initiator_depth = P9_RDMA_ORD;
725 err = rdma_connect(rdma->cm_id, &conn_param);
726 if (err)
727 goto error;
728 err = wait_for_completion_interruptible(&rdma->cm_done);
729 if (err || (rdma->state != P9_RDMA_CONNECTED))
730 goto error;
731
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500732 client->status = Connected;
733
734 return 0;
735
736error:
737 rdma_destroy_trans(rdma);
738 return -ENOTCONN;
739}
740
741static struct p9_trans_module p9_rdma_trans = {
742 .name = "rdma",
743 .maxsize = P9_RDMA_MAXSIZE,
744 .def = 0,
745 .owner = THIS_MODULE,
746 .create = rdma_create_trans,
747 .close = rdma_close,
748 .request = rdma_request,
749 .cancel = rdma_cancel,
Simon Derr931700d2014-03-10 16:38:51 +0100750 .cancelled = rdma_cancelled,
David Howellsc4fac912017-07-05 16:25:37 +0100751 .show_options = p9_rdma_show_options,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500752};
753
754/**
755 * p9_trans_rdma_init - Register the 9P RDMA transport driver
756 */
757static int __init p9_trans_rdma_init(void)
758{
759 v9fs_register_trans(&p9_rdma_trans);
760 return 0;
761}
762
763static void __exit p9_trans_rdma_exit(void)
764{
765 v9fs_unregister_trans(&p9_rdma_trans);
766}
767
768module_init(p9_trans_rdma_init);
769module_exit(p9_trans_rdma_exit);
770
771MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
772MODULE_DESCRIPTION("RDMA Transport for 9P");
773MODULE_LICENSE("Dual BSD/GPL");