blob: 88e5638266743e45fe703a9916127d5757704b44 [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/*
Tom Tuckerfc79d4b2008-10-22 18:47:39 -05003 * RDMA transport layer based on the trans_fd.c implementation.
4 *
5 * Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
6 * Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
7 * Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
8 * Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
9 * Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050010 */
11
Joe Perches5d385152011-11-28 10:40:46 -080012#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050014#include <linux/in.h>
15#include <linux/module.h>
16#include <linux/net.h>
17#include <linux/ipv6.h>
18#include <linux/kthread.h>
19#include <linux/errno.h>
20#include <linux/kernel.h>
21#include <linux/un.h>
22#include <linux/uaccess.h>
23#include <linux/inet.h>
24#include <linux/idr.h>
25#include <linux/file.h>
26#include <linux/parser.h>
27#include <linux/semaphore.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090028#include <linux/slab.h>
David Howellsc4fac912017-07-05 16:25:37 +010029#include <linux/seq_file.h>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050030#include <net/9p/9p.h>
31#include <net/9p/client.h>
32#include <net/9p/transport.h>
33#include <rdma/ib_verbs.h>
34#include <rdma/rdma_cm.h>
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050035
36#define P9_PORT 5640
37#define P9_RDMA_SQ_DEPTH 32
38#define P9_RDMA_RQ_DEPTH 32
39#define P9_RDMA_SEND_SGE 4
40#define P9_RDMA_RECV_SGE 4
41#define P9_RDMA_IRD 0
42#define P9_RDMA_ORD 0
43#define P9_RDMA_TIMEOUT 30000 /* 30 seconds */
Simon Derr3fcc62f2013-06-21 15:32:37 +020044#define P9_RDMA_MAXSIZE (1024*1024) /* 1MB */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050045
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050046/**
47 * struct p9_trans_rdma - RDMA transport instance
48 *
49 * @state: tracks the transport state machine for connection setup and tear down
50 * @cm_id: The RDMA CM ID
51 * @pd: Protection Domain pointer
52 * @qp: Queue Pair pointer
53 * @cq: Completion Queue pointer
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050054 * @timeout: Number of uSecs to wait for connection management events
David Howellsc4fac912017-07-05 16:25:37 +010055 * @privport: Whether a privileged port may be used
56 * @port: The port to use
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050057 * @sq_depth: The depth of the Send Queue
58 * @sq_sem: Semaphore for the SQ
59 * @rq_depth: The depth of the Receive Queue.
Simon Derrfd453d02013-06-21 15:32:39 +020060 * @rq_sem: Semaphore for the RQ
Simon Derr1cff3302013-06-21 15:32:42 +020061 * @excess_rc : Amount of posted Receive Contexts without a pending request.
62 * See rdma_request()
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050063 * @addr: The remote peer's address
64 * @req_lock: Protects the active request list
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050065 * @cm_done: Completion event for connection management tracking
66 */
67struct p9_trans_rdma {
68 enum {
69 P9_RDMA_INIT,
70 P9_RDMA_ADDR_RESOLVED,
71 P9_RDMA_ROUTE_RESOLVED,
72 P9_RDMA_CONNECTED,
73 P9_RDMA_FLUSHING,
74 P9_RDMA_CLOSING,
75 P9_RDMA_CLOSED,
76 } state;
77 struct rdma_cm_id *cm_id;
78 struct ib_pd *pd;
79 struct ib_qp *qp;
80 struct ib_cq *cq;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050081 long timeout;
David Howellsc4fac912017-07-05 16:25:37 +010082 bool privport;
83 u16 port;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050084 int sq_depth;
85 struct semaphore sq_sem;
86 int rq_depth;
Simon Derrfd453d02013-06-21 15:32:39 +020087 struct semaphore rq_sem;
Simon Derr1cff3302013-06-21 15:32:42 +020088 atomic_t excess_rc;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050089 struct sockaddr_in addr;
90 spinlock_t req_lock;
91
92 struct completion cm_done;
93};
94
Andrew Lunn15e522a2020-07-13 01:14:57 +020095struct p9_rdma_req;
96
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050097/**
Andrew Lunn15e522a2020-07-13 01:14:57 +020098 * struct p9_rdma_context - Keeps track of in-process WR
Tom Tuckerfc79d4b2008-10-22 18:47:39 -050099 *
Andrew Lunn760b3d62020-10-31 19:26:55 +0100100 * @cqe: completion queue entry
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500101 * @busa: Bus address to unmap when the WR completes
102 * @req: Keeps track of requests (send)
103 * @rc: Keepts track of replies (receive)
104 */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500105struct p9_rdma_context {
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100106 struct ib_cqe cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500107 dma_addr_t busa;
108 union {
109 struct p9_req_t *req;
Dominique Martinet523adb62018-07-30 05:55:19 +0000110 struct p9_fcall rc;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500111 };
112};
113
114/**
Andrew Lunn15e522a2020-07-13 01:14:57 +0200115 * struct p9_rdma_opts - Collection of mount options
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500116 * @port: port of connection
Andrew Lunn760b3d62020-10-31 19:26:55 +0100117 * @privport: Whether a privileged port may be used
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500118 * @sq_depth: The requested depth of the SQ. This really doesn't need
119 * to be any deeper than the number of threads used in the client
120 * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
121 * @timeout: Time to wait in msecs for CM events
122 */
123struct p9_rdma_opts {
124 short port;
David Howellsc4fac912017-07-05 16:25:37 +0100125 bool privport;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500126 int sq_depth;
127 int rq_depth;
128 long timeout;
129};
130
131/*
132 * Option Parsing (code inspired by NFS code)
133 */
134enum {
135 /* Options that take integer arguments */
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100136 Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
137 /* Options that take no argument */
138 Opt_privport,
139 Opt_err,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500140};
141
142static match_table_t tokens = {
143 {Opt_port, "port=%u"},
144 {Opt_sq_depth, "sq=%u"},
145 {Opt_rq_depth, "rq=%u"},
146 {Opt_timeout, "timeout=%u"},
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100147 {Opt_privport, "privport"},
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500148 {Opt_err, NULL},
149};
150
David Howellsc4fac912017-07-05 16:25:37 +0100151static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
152{
153 struct p9_trans_rdma *rdma = clnt->trans;
154
155 if (rdma->port != P9_PORT)
156 seq_printf(m, ",port=%u", rdma->port);
157 if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
158 seq_printf(m, ",sq=%u", rdma->sq_depth);
159 if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
160 seq_printf(m, ",rq=%u", rdma->rq_depth);
161 if (rdma->timeout != P9_RDMA_TIMEOUT)
162 seq_printf(m, ",timeout=%lu", rdma->timeout);
163 if (rdma->privport)
164 seq_puts(m, ",privport");
165 return 0;
166}
167
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500168/**
Abhishek Kulkarni0e155972009-07-19 13:41:55 -0600169 * parse_opts - parse mount options into rdma options structure
170 * @params: options string passed from mount
171 * @opts: rdma transport-specific structure to parse options into
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500172 *
173 * Returns 0 upon success, -ERRNO upon failure
174 */
175static int parse_opts(char *params, struct p9_rdma_opts *opts)
176{
177 char *p;
178 substring_t args[MAX_OPT_ARGS];
179 int option;
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600180 char *options, *tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500181
182 opts->port = P9_PORT;
183 opts->sq_depth = P9_RDMA_SQ_DEPTH;
184 opts->rq_depth = P9_RDMA_RQ_DEPTH;
185 opts->timeout = P9_RDMA_TIMEOUT;
David Howellsc4fac912017-07-05 16:25:37 +0100186 opts->privport = false;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500187
188 if (!params)
189 return 0;
190
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600191 tmp_options = kstrdup(params, GFP_KERNEL);
192 if (!tmp_options) {
Joe Perches5d385152011-11-28 10:40:46 -0800193 p9_debug(P9_DEBUG_ERROR,
194 "failed to allocate copy of option string\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500195 return -ENOMEM;
196 }
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600197 options = tmp_options;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500198
199 while ((p = strsep(&options, ",")) != NULL) {
200 int token;
201 int r;
202 if (!*p)
203 continue;
204 token = match_token(p, tokens, args);
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100205 if ((token != Opt_err) && (token != Opt_privport)) {
206 r = match_int(&args[0], &option);
207 if (r < 0) {
208 p9_debug(P9_DEBUG_ERROR,
209 "integer field, but no integer?\n");
210 continue;
211 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500212 }
213 switch (token) {
214 case Opt_port:
215 opts->port = option;
216 break;
217 case Opt_sq_depth:
218 opts->sq_depth = option;
219 break;
220 case Opt_rq_depth:
221 opts->rq_depth = option;
222 break;
223 case Opt_timeout:
224 opts->timeout = option;
225 break;
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100226 case Opt_privport:
David Howellsc4fac912017-07-05 16:25:37 +0100227 opts->privport = true;
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100228 break;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500229 default:
230 continue;
231 }
232 }
233 /* RQ must be at least as large as the SQ */
234 opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
Eric Van Hensbergend8c8a9e2010-02-08 16:23:23 -0600235 kfree(tmp_options);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500236 return 0;
237}
238
239static int
240p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
241{
242 struct p9_client *c = id->context;
243 struct p9_trans_rdma *rdma = c->trans;
244 switch (event->event) {
245 case RDMA_CM_EVENT_ADDR_RESOLVED:
246 BUG_ON(rdma->state != P9_RDMA_INIT);
247 rdma->state = P9_RDMA_ADDR_RESOLVED;
248 break;
249
250 case RDMA_CM_EVENT_ROUTE_RESOLVED:
251 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
252 rdma->state = P9_RDMA_ROUTE_RESOLVED;
253 break;
254
255 case RDMA_CM_EVENT_ESTABLISHED:
256 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
257 rdma->state = P9_RDMA_CONNECTED;
258 break;
259
260 case RDMA_CM_EVENT_DISCONNECTED:
261 if (rdma)
262 rdma->state = P9_RDMA_CLOSED;
Dominique Martinet473c7dd2018-09-08 00:26:50 +0900263 c->status = Disconnected;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500264 break;
265
266 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
267 break;
268
269 case RDMA_CM_EVENT_ADDR_CHANGE:
270 case RDMA_CM_EVENT_ROUTE_ERROR:
271 case RDMA_CM_EVENT_DEVICE_REMOVAL:
272 case RDMA_CM_EVENT_MULTICAST_JOIN:
273 case RDMA_CM_EVENT_MULTICAST_ERROR:
274 case RDMA_CM_EVENT_REJECTED:
275 case RDMA_CM_EVENT_CONNECT_REQUEST:
276 case RDMA_CM_EVENT_CONNECT_RESPONSE:
277 case RDMA_CM_EVENT_CONNECT_ERROR:
278 case RDMA_CM_EVENT_ADDR_ERROR:
279 case RDMA_CM_EVENT_UNREACHABLE:
280 c->status = Disconnected;
281 rdma_disconnect(rdma->cm_id);
282 break;
283 default:
284 BUG();
285 }
286 complete(&rdma->cm_done);
287 return 0;
288}
289
290static void
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100291recv_done(struct ib_cq *cq, struct ib_wc *wc)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500292{
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100293 struct p9_client *client = cq->cq_context;
294 struct p9_trans_rdma *rdma = client->trans;
295 struct p9_rdma_context *c =
296 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500297 struct p9_req_t *req;
298 int err = 0;
299 int16_t tag;
300
301 req = NULL;
302 ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
303 DMA_FROM_DEVICE);
304
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100305 if (wc->status != IB_WC_SUCCESS)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500306 goto err_out;
307
Dominique Martinet523adb62018-07-30 05:55:19 +0000308 c->rc.size = wc->byte_len;
309 err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500310 if (err)
311 goto err_out;
312
313 req = p9_tag_lookup(client, tag);
314 if (!req)
315 goto err_out;
316
Simon Derr47229ff2013-06-21 15:32:38 +0200317 /* Check that we have not yet received a reply for this request.
318 */
Dominique Martinet523adb62018-07-30 05:55:19 +0000319 if (unlikely(req->rc.sdata)) {
Simon Derr47229ff2013-06-21 15:32:38 +0200320 pr_err("Duplicate reply for request %d", tag);
321 goto err_out;
322 }
323
Dominique Martinet523adb62018-07-30 05:55:19 +0000324 req->rc.size = c->rc.size;
325 req->rc.sdata = c->rc.sdata;
Dominique Martinet2b6e72e2014-01-17 18:31:00 +0100326 p9_client_cb(client, req, REQ_STATUS_RCVD);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500327
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100328 out:
329 up(&rdma->rq_sem);
330 kfree(c);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500331 return;
332
333 err_out:
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100334 p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
335 req, err, wc->status);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500336 rdma->state = P9_RDMA_FLUSHING;
337 client->status = Disconnected;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100338 goto out;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500339}
340
341static void
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100342send_done(struct ib_cq *cq, struct ib_wc *wc)
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500343{
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100344 struct p9_client *client = cq->cq_context;
345 struct p9_trans_rdma *rdma = client->trans;
346 struct p9_rdma_context *c =
347 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
348
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500349 ib_dma_unmap_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000350 c->busa, c->req->tc.size,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500351 DMA_TO_DEVICE);
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100352 up(&rdma->sq_sem);
Tomas Bortoli728356d2018-08-14 19:43:42 +0200353 p9_req_put(c->req);
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100354 kfree(c);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500355}
356
357static void qp_event_handler(struct ib_event *event, void *context)
358{
Joe Perches5d385152011-11-28 10:40:46 -0800359 p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
360 event->event, context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500361}
362
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500363static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
364{
365 if (!rdma)
366 return;
367
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500368 if (rdma->qp && !IS_ERR(rdma->qp))
369 ib_destroy_qp(rdma->qp);
370
371 if (rdma->pd && !IS_ERR(rdma->pd))
372 ib_dealloc_pd(rdma->pd);
373
374 if (rdma->cq && !IS_ERR(rdma->cq))
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100375 ib_free_cq(rdma->cq);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500376
377 if (rdma->cm_id && !IS_ERR(rdma->cm_id))
378 rdma_destroy_id(rdma->cm_id);
379
380 kfree(rdma);
381}
382
383static int
384post_recv(struct p9_client *client, struct p9_rdma_context *c)
385{
386 struct p9_trans_rdma *rdma = client->trans;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700387 struct ib_recv_wr wr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500388 struct ib_sge sge;
389
390 c->busa = ib_dma_map_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000391 c->rc.sdata, client->msize,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500392 DMA_FROM_DEVICE);
393 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
394 goto error;
395
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100396 c->cqe.done = recv_done;
397
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500398 sge.addr = c->busa;
399 sge.length = client->msize;
Jason Gunthorpe2f31fa82015-07-30 17:22:25 -0600400 sge.lkey = rdma->pd->local_dma_lkey;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500401
402 wr.next = NULL;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100403 wr.wr_cqe = &c->cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500404 wr.sg_list = &sge;
405 wr.num_sge = 1;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700406 return ib_post_recv(rdma->qp, &wr, NULL);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500407
408 error:
Joe Perches5d385152011-11-28 10:40:46 -0800409 p9_debug(P9_DEBUG_ERROR, "EIO\n");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500410 return -EIO;
411}
412
413static int rdma_request(struct p9_client *client, struct p9_req_t *req)
414{
415 struct p9_trans_rdma *rdma = client->trans;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700416 struct ib_send_wr wr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500417 struct ib_sge sge;
418 int err = 0;
419 unsigned long flags;
420 struct p9_rdma_context *c = NULL;
421 struct p9_rdma_context *rpl_context = NULL;
422
Simon Derr1cff3302013-06-21 15:32:42 +0200423 /* When an error occurs between posting the recv and the send,
424 * there will be a receive context posted without a pending request.
425 * Since there is no way to "un-post" it, we remember it and skip
426 * post_recv() for the next request.
427 * So here,
428 * see if we are this `next request' and need to absorb an excess rc.
429 * If yes, then drop and free our own, and do not recv_post().
430 **/
431 if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
432 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
Dominique Martinet523adb62018-07-30 05:55:19 +0000433 /* Got one! */
434 p9_fcall_fini(&req->rc);
435 req->rc.sdata = NULL;
Simon Derr1cff3302013-06-21 15:32:42 +0200436 goto dont_need_post_recv;
437 } else {
438 /* We raced and lost. */
439 atomic_inc(&rdma->excess_rc);
440 }
441 }
442
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500443 /* Allocate an fcall for the reply */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530444 rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000445 if (!rpl_context) {
446 err = -ENOMEM;
Simon Derr2f52d072013-06-21 15:32:41 +0200447 goto recv_error;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000448 }
Dominique Martinet523adb62018-07-30 05:55:19 +0000449 rpl_context->rc.sdata = req->rc.sdata;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500450
451 /*
452 * Post a receive buffer for this request. We need to ensure
453 * there is a reply buffer available for every outstanding
454 * request. A flushed request can result in no reply for an
455 * outstanding request, so we must keep a count to avoid
456 * overflowing the RQ.
457 */
Simon Derr2f52d072013-06-21 15:32:41 +0200458 if (down_interruptible(&rdma->rq_sem)) {
459 err = -EINTR;
460 goto recv_error;
461 }
Simon Derrfd453d02013-06-21 15:32:39 +0200462
463 err = post_recv(client, rpl_context);
464 if (err) {
Dominique Martinet8b894ad2018-08-30 19:29:36 +0900465 p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
Simon Derr2f52d072013-06-21 15:32:41 +0200466 goto recv_error;
Simon Derrfd453d02013-06-21 15:32:39 +0200467 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500468 /* remove posted receive buffer from request structure */
Dominique Martinet523adb62018-07-30 05:55:19 +0000469 req->rc.sdata = NULL;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500470
Simon Derr1cff3302013-06-21 15:32:42 +0200471dont_need_post_recv:
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500472 /* Post the request */
Aneesh Kumar K.Veeff66e2011-03-08 16:39:47 +0530473 c = kmalloc(sizeof *c, GFP_NOFS);
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000474 if (!c) {
475 err = -ENOMEM;
Simon Derr2f52d072013-06-21 15:32:41 +0200476 goto send_error;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000477 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500478 c->req = req;
479
480 c->busa = ib_dma_map_single(rdma->cm_id->device,
Dominique Martinet523adb62018-07-30 05:55:19 +0000481 c->req->tc.sdata, c->req->tc.size,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500482 DMA_TO_DEVICE);
Simon Derr2f52d072013-06-21 15:32:41 +0200483 if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
484 err = -EIO;
485 goto send_error;
486 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500487
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100488 c->cqe.done = send_done;
489
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500490 sge.addr = c->busa;
Dominique Martinet523adb62018-07-30 05:55:19 +0000491 sge.length = c->req->tc.size;
Jason Gunthorpe2f31fa82015-07-30 17:22:25 -0600492 sge.lkey = rdma->pd->local_dma_lkey;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500493
494 wr.next = NULL;
Christoph Hellwig7cf20fc2016-03-03 09:36:06 +0100495 wr.wr_cqe = &c->cqe;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500496 wr.opcode = IB_WR_SEND;
497 wr.send_flags = IB_SEND_SIGNALED;
498 wr.sg_list = &sge;
499 wr.num_sge = 1;
500
Simon Derr2f52d072013-06-21 15:32:41 +0200501 if (down_interruptible(&rdma->sq_sem)) {
502 err = -EINTR;
503 goto send_error;
504 }
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500505
Simon Derr3f9d5b82014-03-10 16:38:50 +0100506 /* Mark request as `sent' *before* we actually send it,
507 * because doing if after could erase the REQ_STATUS_RCVD
508 * status in case of a very fast reply.
509 */
510 req->status = REQ_STATUS_SENT;
Bart Van Assche72bc4d32018-07-18 09:25:26 -0700511 err = ib_post_send(rdma->qp, &wr, NULL);
Simon Derr2f52d072013-06-21 15:32:41 +0200512 if (err)
513 goto send_error;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500514
Simon Derr2f52d072013-06-21 15:32:41 +0200515 /* Success */
516 return 0;
517
518 /* Handle errors that happened during or while preparing the send: */
519 send_error:
Simon Derr3f9d5b82014-03-10 16:38:50 +0100520 req->status = REQ_STATUS_ERROR;
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000521 kfree(c);
Simon Derr2f52d072013-06-21 15:32:41 +0200522 p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
Simon Derr1cff3302013-06-21 15:32:42 +0200523
524 /* Ach.
525 * We did recv_post(), but not send. We have one recv_post in excess.
526 */
527 atomic_inc(&rdma->excess_rc);
Simon Derr2f52d072013-06-21 15:32:41 +0200528 return err;
529
530 /* Handle errors that happened during or while preparing post_recv(): */
531 recv_error:
Davidlohr Bueso1d6400c2010-09-13 15:53:18 +0000532 kfree(rpl_context);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500533 spin_lock_irqsave(&rdma->req_lock, flags);
Dominique Martinet8b894ad2018-08-30 19:29:36 +0900534 if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500535 rdma->state = P9_RDMA_CLOSING;
536 spin_unlock_irqrestore(&rdma->req_lock, flags);
537 rdma_disconnect(rdma->cm_id);
538 } else
539 spin_unlock_irqrestore(&rdma->req_lock, flags);
540 return err;
541}
542
543static void rdma_close(struct p9_client *client)
544{
545 struct p9_trans_rdma *rdma;
546
547 if (!client)
548 return;
549
550 rdma = client->trans;
551 if (!rdma)
552 return;
553
554 client->status = Disconnected;
555 rdma_disconnect(rdma->cm_id);
556 rdma_destroy_trans(rdma);
557}
558
559/**
560 * alloc_rdma - Allocate and initialize the rdma transport structure
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500561 * @opts: Mount options structure
562 */
563static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
564{
565 struct p9_trans_rdma *rdma;
566
567 rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
568 if (!rdma)
569 return NULL;
570
David Howellsc4fac912017-07-05 16:25:37 +0100571 rdma->port = opts->port;
572 rdma->privport = opts->privport;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500573 rdma->sq_depth = opts->sq_depth;
574 rdma->rq_depth = opts->rq_depth;
575 rdma->timeout = opts->timeout;
576 spin_lock_init(&rdma->req_lock);
577 init_completion(&rdma->cm_done);
578 sema_init(&rdma->sq_sem, rdma->sq_depth);
Simon Derrfd453d02013-06-21 15:32:39 +0200579 sema_init(&rdma->rq_sem, rdma->rq_depth);
Simon Derr1cff3302013-06-21 15:32:42 +0200580 atomic_set(&rdma->excess_rc, 0);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500581
582 return rdma;
583}
584
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500585static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
586{
Simon Derr931700d2014-03-10 16:38:51 +0100587 /* Nothing to do here.
588 * We will take care of it (if we have to) in rdma_cancelled()
589 */
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500590 return 1;
591}
592
Simon Derr931700d2014-03-10 16:38:51 +0100593/* A request has been fully flushed without a reply.
594 * That means we have posted one buffer in excess.
595 */
596static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
597{
598 struct p9_trans_rdma *rdma = client->trans;
599 atomic_inc(&rdma->excess_rc);
600 return 0;
601}
602
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100603static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
604{
605 struct sockaddr_in cl = {
606 .sin_family = AF_INET,
607 .sin_addr.s_addr = htonl(INADDR_ANY),
608 };
609 int port, err = -EINVAL;
610
611 for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
612 cl.sin_port = htons((ushort)port);
613 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
614 if (err != -EADDRINUSE)
615 break;
616 }
617 return err;
618}
619
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500620/**
Sun Lianwen4a026da2018-05-08 09:49:38 +0800621 * rdma_create_trans - Transport method for creating a transport instance
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500622 * @client: client instance
623 * @addr: IP address string
624 * @args: Mount options string
625 */
626static int
627rdma_create_trans(struct p9_client *client, const char *addr, char *args)
628{
629 int err;
630 struct p9_rdma_opts opts;
631 struct p9_trans_rdma *rdma;
632 struct rdma_conn_param conn_param;
633 struct ib_qp_init_attr qp_attr;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500634
Tomas Bortoli10aa1452018-07-27 13:05:58 +0200635 if (addr == NULL)
636 return -EINVAL;
637
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500638 /* Parse the transport specific mount options */
639 err = parse_opts(args, &opts);
640 if (err < 0)
641 return err;
642
643 /* Create and initialize the RDMA transport structure */
644 rdma = alloc_rdma(&opts);
645 if (!rdma)
646 return -ENOMEM;
647
648 /* Create the RDMA CM ID */
Guy Shapirofa201052015-10-22 15:20:10 +0300649 rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
650 RDMA_PS_TCP, IB_QPT_RC);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500651 if (IS_ERR(rdma->cm_id))
652 goto error;
653
Tom Tucker517ac452008-10-23 16:30:13 -0500654 /* Associate the client with the transport */
655 client->trans = rdma;
656
Dominique Martinetf569d3e2015-01-09 13:07:00 +0100657 /* Bind to a privileged port if we need to */
658 if (opts.privport) {
659 err = p9_rdma_bind_privport(rdma);
660 if (err < 0) {
661 pr_err("%s (%d): problem binding to privport: %d\n",
662 __func__, task_pid_nr(current), -err);
663 goto error;
664 }
665 }
666
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500667 /* Resolve the server's address */
668 rdma->addr.sin_family = AF_INET;
669 rdma->addr.sin_addr.s_addr = in_aton(addr);
670 rdma->addr.sin_port = htons(opts.port);
671 err = rdma_resolve_addr(rdma->cm_id, NULL,
672 (struct sockaddr *)&rdma->addr,
673 rdma->timeout);
674 if (err)
675 goto error;
676 err = wait_for_completion_interruptible(&rdma->cm_done);
677 if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
678 goto error;
679
680 /* Resolve the route to the server */
681 err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
682 if (err)
683 goto error;
684 err = wait_for_completion_interruptible(&rdma->cm_done);
685 if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
686 goto error;
687
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500688 /* Create the Completion Queue */
Chuck Lever20cf4e02019-07-29 13:22:09 -0400689 rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
690 opts.sq_depth + opts.rq_depth + 1,
691 IB_POLL_SOFTIRQ);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500692 if (IS_ERR(rdma->cq))
693 goto error;
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500694
695 /* Create the Protection Domain */
Christoph Hellwiged082d32016-09-05 12:56:17 +0200696 rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500697 if (IS_ERR(rdma->pd))
698 goto error;
699
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500700 /* Create the Queue Pair */
701 memset(&qp_attr, 0, sizeof qp_attr);
702 qp_attr.event_handler = qp_event_handler;
703 qp_attr.qp_context = client;
704 qp_attr.cap.max_send_wr = opts.sq_depth;
705 qp_attr.cap.max_recv_wr = opts.rq_depth;
706 qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
707 qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
708 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
709 qp_attr.qp_type = IB_QPT_RC;
710 qp_attr.send_cq = rdma->cq;
711 qp_attr.recv_cq = rdma->cq;
712 err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
713 if (err)
714 goto error;
715 rdma->qp = rdma->cm_id->qp;
716
717 /* Request a connection */
718 memset(&conn_param, 0, sizeof(conn_param));
719 conn_param.private_data = NULL;
720 conn_param.private_data_len = 0;
721 conn_param.responder_resources = P9_RDMA_IRD;
722 conn_param.initiator_depth = P9_RDMA_ORD;
723 err = rdma_connect(rdma->cm_id, &conn_param);
724 if (err)
725 goto error;
726 err = wait_for_completion_interruptible(&rdma->cm_done);
727 if (err || (rdma->state != P9_RDMA_CONNECTED))
728 goto error;
729
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500730 client->status = Connected;
731
732 return 0;
733
734error:
735 rdma_destroy_trans(rdma);
736 return -ENOTCONN;
737}
738
739static struct p9_trans_module p9_rdma_trans = {
740 .name = "rdma",
741 .maxsize = P9_RDMA_MAXSIZE,
742 .def = 0,
743 .owner = THIS_MODULE,
744 .create = rdma_create_trans,
745 .close = rdma_close,
746 .request = rdma_request,
747 .cancel = rdma_cancel,
Simon Derr931700d2014-03-10 16:38:51 +0100748 .cancelled = rdma_cancelled,
David Howellsc4fac912017-07-05 16:25:37 +0100749 .show_options = p9_rdma_show_options,
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500750};
751
752/**
753 * p9_trans_rdma_init - Register the 9P RDMA transport driver
754 */
755static int __init p9_trans_rdma_init(void)
756{
757 v9fs_register_trans(&p9_rdma_trans);
758 return 0;
759}
760
761static void __exit p9_trans_rdma_exit(void)
762{
763 v9fs_unregister_trans(&p9_rdma_trans);
764}
765
766module_init(p9_trans_rdma_init);
767module_exit(p9_trans_rdma_exit);
Thomas Weißschuh4cd82a52021-10-17 15:46:11 +0200768MODULE_ALIAS_9P("rdma");
Tom Tuckerfc79d4b2008-10-22 18:47:39 -0500769
770MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
771MODULE_DESCRIPTION("RDMA Transport for 9P");
772MODULE_LICENSE("Dual BSD/GPL");