trans_rdma.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * 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>
 10 */
 11
 12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 13
 14#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/file.h>
 25#include <linux/fs_context.h>
 26#include <linux/semaphore.h>
 27#include <linux/slab.h>
 28#include <linux/seq_file.h>
 29#include <net/9p/9p.h>
 30#include <net/9p/client.h>
 31#include <net/9p/transport.h>
 32#include <rdma/ib_verbs.h>
 33#include <rdma/rdma_cm.h>
 34
 35#define P9_RDMA_SEND_SGE	4
 36#define P9_RDMA_RECV_SGE	4
 37#define P9_RDMA_IRD		0
 38#define P9_RDMA_ORD		0
 39#define P9_RDMA_MAXSIZE		(1024*1024)	/* 1MB */
 40
 41/**
 42 * struct p9_trans_rdma - RDMA transport instance
 43 *
 44 * @state: tracks the transport state machine for connection setup and tear down
 45 * @cm_id: The RDMA CM ID
 46 * @pd: Protection Domain pointer
 47 * @qp: Queue Pair pointer
 48 * @cq: Completion Queue pointer
 49 * @timeout: Number of uSecs to wait for connection management events
 50 * @privport: Whether a privileged port may be used
 51 * @port: The port to use
 52 * @sq_depth: The depth of the Send Queue
 53 * @sq_sem: Semaphore for the SQ
 54 * @rq_depth: The depth of the Receive Queue.
 55 * @rq_sem: Semaphore for the RQ
 56 * @excess_rc : Amount of posted Receive Contexts without a pending request.
 57 *		See rdma_request()
 58 * @addr: The remote peer's address
 59 * @req_lock: Protects the active request list
 60 * @cm_done: Completion event for connection management tracking
 61 */
 62struct p9_trans_rdma {
 63	enum {
 64		P9_RDMA_INIT,
 65		P9_RDMA_ADDR_RESOLVED,
 66		P9_RDMA_ROUTE_RESOLVED,
 67		P9_RDMA_CONNECTED,
 68		P9_RDMA_FLUSHING,
 69		P9_RDMA_CLOSING,
 70		P9_RDMA_CLOSED,
 71	} state;
 72	struct rdma_cm_id *cm_id;
 73	struct ib_pd *pd;
 74	struct ib_qp *qp;
 75	struct ib_cq *cq;
 76	long timeout;
 77	bool privport;
 78	u16 port;
 79	int sq_depth;
 80	struct semaphore sq_sem;
 81	int rq_depth;
 82	struct semaphore rq_sem;
 83	atomic_t excess_rc;
 84	struct sockaddr_in addr;
 85	spinlock_t req_lock;
 86
 87	struct completion cm_done;
 88};
 89
 90struct p9_rdma_req;
 91
 92/**
 93 * struct p9_rdma_context - Keeps track of in-process WR
 94 *
 95 * @cqe: completion queue entry
 96 * @busa: Bus address to unmap when the WR completes
 97 * @req: Keeps track of requests (send)
 98 * @rc: Keepts track of replies (receive)
 99 */
100struct p9_rdma_context {
101	struct ib_cqe cqe;
102	dma_addr_t busa;
103	union {
104		struct p9_req_t *req;
105		struct p9_fcall rc;
106	};
107};
108
109static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
110{
111	struct p9_trans_rdma *rdma = clnt->trans;
112
113	if (rdma->port != P9_RDMA_PORT)
114		seq_printf(m, ",port=%u", rdma->port);
115	if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
116		seq_printf(m, ",sq=%u", rdma->sq_depth);
117	if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
118		seq_printf(m, ",rq=%u", rdma->rq_depth);
119	if (rdma->timeout != P9_RDMA_TIMEOUT)
120		seq_printf(m, ",timeout=%lu", rdma->timeout);
121	if (rdma->privport)
122		seq_puts(m, ",privport");
123	return 0;
124}
125
126static int
127p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
128{
129	struct p9_client *c = id->context;
130	struct p9_trans_rdma *rdma = c->trans;
131	switch (event->event) {
132	case RDMA_CM_EVENT_ADDR_RESOLVED:
133		BUG_ON(rdma->state != P9_RDMA_INIT);
134		rdma->state = P9_RDMA_ADDR_RESOLVED;
135		break;
136
137	case RDMA_CM_EVENT_ROUTE_RESOLVED:
138		BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
139		rdma->state = P9_RDMA_ROUTE_RESOLVED;
140		break;
141
142	case RDMA_CM_EVENT_ESTABLISHED:
143		BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
144		rdma->state = P9_RDMA_CONNECTED;
145		break;
146
147	case RDMA_CM_EVENT_DISCONNECTED:
148		if (rdma)
149			rdma->state = P9_RDMA_CLOSED;
150		c->status = Disconnected;
151		break;
152
153	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
154		break;
155
156	case RDMA_CM_EVENT_ADDR_CHANGE:
157	case RDMA_CM_EVENT_ROUTE_ERROR:
158	case RDMA_CM_EVENT_DEVICE_REMOVAL:
159	case RDMA_CM_EVENT_MULTICAST_JOIN:
160	case RDMA_CM_EVENT_MULTICAST_ERROR:
161	case RDMA_CM_EVENT_REJECTED:
162	case RDMA_CM_EVENT_CONNECT_REQUEST:
163	case RDMA_CM_EVENT_CONNECT_RESPONSE:
164	case RDMA_CM_EVENT_CONNECT_ERROR:
165	case RDMA_CM_EVENT_ADDR_ERROR:
166	case RDMA_CM_EVENT_UNREACHABLE:
167		c->status = Disconnected;
168		rdma_disconnect(rdma->cm_id);
169		break;
170	default:
171		BUG();
172	}
173	complete(&rdma->cm_done);
174	return 0;
175}
176
177static void
178recv_done(struct ib_cq *cq, struct ib_wc *wc)
179{
180	struct p9_client *client = cq->cq_context;
181	struct p9_trans_rdma *rdma = client->trans;
182	struct p9_rdma_context *c =
183		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
184	struct p9_req_t *req;
185	int err = 0;
186	int16_t tag;
187
188	req = NULL;
189	ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
190							 DMA_FROM_DEVICE);
191
192	if (wc->status != IB_WC_SUCCESS)
193		goto err_out;
194
195	c->rc.size = wc->byte_len;
196	err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
197	if (err)
198		goto err_out;
199
200	req = p9_tag_lookup(client, tag);
201	if (!req)
202		goto err_out;
203
204	/* Check that we have not yet received a reply for this request.
205	 */
206	if (unlikely(req->rc.sdata)) {
207		pr_err("Duplicate reply for request %d", tag);
208		goto err_out;
209	}
210
211	req->rc.size = c->rc.size;
212	req->rc.sdata = c->rc.sdata;
213	p9_client_cb(client, req, REQ_STATUS_RCVD);
214
215 out:
216	up(&rdma->rq_sem);
217	kfree(c);
218	return;
219
220 err_out:
221	p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
222			req, err, wc->status);
223	rdma->state = P9_RDMA_FLUSHING;
224	client->status = Disconnected;
225	goto out;
226}
227
228static void
229send_done(struct ib_cq *cq, struct ib_wc *wc)
230{
231	struct p9_client *client = cq->cq_context;
232	struct p9_trans_rdma *rdma = client->trans;
233	struct p9_rdma_context *c =
234		container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
235
236	ib_dma_unmap_single(rdma->cm_id->device,
237			    c->busa, c->req->tc.size,
238			    DMA_TO_DEVICE);
239	up(&rdma->sq_sem);
240	p9_req_put(client, c->req);
241	kfree(c);
242}
243
244static void qp_event_handler(struct ib_event *event, void *context)
245{
246	p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
247		 event->event, context);
248}
249
250static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
251{
252	if (!rdma)
253		return;
254
255	if (rdma->qp && !IS_ERR(rdma->qp))
256		ib_destroy_qp(rdma->qp);
257
258	if (rdma->pd && !IS_ERR(rdma->pd))
259		ib_dealloc_pd(rdma->pd);
260
261	if (rdma->cq && !IS_ERR(rdma->cq))
262		ib_free_cq(rdma->cq);
263
264	if (rdma->cm_id && !IS_ERR(rdma->cm_id))
265		rdma_destroy_id(rdma->cm_id);
266
267	kfree(rdma);
268}
269
270static int
271post_recv(struct p9_client *client, struct p9_rdma_context *c)
272{
273	struct p9_trans_rdma *rdma = client->trans;
274	struct ib_recv_wr wr;
275	struct ib_sge sge;
276	int ret;
277
278	c->busa = ib_dma_map_single(rdma->cm_id->device,
279				    c->rc.sdata, client->msize,
280				    DMA_FROM_DEVICE);
281	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
282		goto error;
283
284	c->cqe.done = recv_done;
285
286	sge.addr = c->busa;
287	sge.length = client->msize;
288	sge.lkey = rdma->pd->local_dma_lkey;
289
290	wr.next = NULL;
291	wr.wr_cqe = &c->cqe;
292	wr.sg_list = &sge;
293	wr.num_sge = 1;
294
295	ret = ib_post_recv(rdma->qp, &wr, NULL);
296	if (ret)
297		ib_dma_unmap_single(rdma->cm_id->device, c->busa,
298				    client->msize, DMA_FROM_DEVICE);
299	return ret;
300
301 error:
302	p9_debug(P9_DEBUG_ERROR, "EIO\n");
303	return -EIO;
304}
305
306static int rdma_request(struct p9_client *client, struct p9_req_t *req)
307{
308	struct p9_trans_rdma *rdma = client->trans;
309	struct ib_send_wr wr;
310	struct ib_sge sge;
311	int err = 0;
312	unsigned long flags;
313	struct p9_rdma_context *c = NULL;
314	struct p9_rdma_context *rpl_context = NULL;
315
316	/* When an error occurs between posting the recv and the send,
317	 * there will be a receive context posted without a pending request.
318	 * Since there is no way to "un-post" it, we remember it and skip
319	 * post_recv() for the next request.
320	 * So here,
321	 * see if we are this `next request' and need to absorb an excess rc.
322	 * If yes, then drop and free our own, and do not recv_post().
323	 **/
324	if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
325		if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
326			/* Got one! */
327			p9_fcall_fini(&req->rc);
328			req->rc.sdata = NULL;
329			goto dont_need_post_recv;
330		} else {
331			/* We raced and lost. */
332			atomic_inc(&rdma->excess_rc);
333		}
334	}
335
336	/* Allocate an fcall for the reply */
337	rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
338	if (!rpl_context) {
339		err = -ENOMEM;
340		goto recv_error;
341	}
342	rpl_context->rc.sdata = req->rc.sdata;
343
344	/*
345	 * Post a receive buffer for this request. We need to ensure
346	 * there is a reply buffer available for every outstanding
347	 * request. A flushed request can result in no reply for an
348	 * outstanding request, so we must keep a count to avoid
349	 * overflowing the RQ.
350	 */
351	if (down_interruptible(&rdma->rq_sem)) {
352		err = -EINTR;
353		goto recv_error;
354	}
355
356	err = post_recv(client, rpl_context);
357	if (err) {
358		p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
359		goto recv_error;
360	}
361	/* remove posted receive buffer from request structure */
362	req->rc.sdata = NULL;
363
364dont_need_post_recv:
365	/* Post the request */
366	c = kmalloc(sizeof *c, GFP_NOFS);
367	if (!c) {
368		err = -ENOMEM;
369		goto send_error;
370	}
371	c->req = req;
372
373	c->busa = ib_dma_map_single(rdma->cm_id->device,
374				    c->req->tc.sdata, c->req->tc.size,
375				    DMA_TO_DEVICE);
376	if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
377		err = -EIO;
378		goto send_error;
379	}
380
381	c->cqe.done = send_done;
382
383	sge.addr = c->busa;
384	sge.length = c->req->tc.size;
385	sge.lkey = rdma->pd->local_dma_lkey;
386
387	wr.next = NULL;
388	wr.wr_cqe = &c->cqe;
389	wr.opcode = IB_WR_SEND;
390	wr.send_flags = IB_SEND_SIGNALED;
391	wr.sg_list = &sge;
392	wr.num_sge = 1;
393
394	if (down_interruptible(&rdma->sq_sem)) {
395		err = -EINTR;
396		goto dma_unmap;
397	}
398
399	/* Mark request as `sent' *before* we actually send it,
400	 * because doing if after could erase the REQ_STATUS_RCVD
401	 * status in case of a very fast reply.
402	 */
403	WRITE_ONCE(req->status, REQ_STATUS_SENT);
404	err = ib_post_send(rdma->qp, &wr, NULL);
405	if (err)
406		goto dma_unmap;
407
408	/* Success */
409	return 0;
410
411dma_unmap:
412	ib_dma_unmap_single(rdma->cm_id->device, c->busa,
413			    c->req->tc.size, DMA_TO_DEVICE);
414 /* Handle errors that happened during or while preparing the send: */
415 send_error:
416	WRITE_ONCE(req->status, REQ_STATUS_ERROR);
417	kfree(c);
418	p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
419
420	/* Ach.
421	 *  We did recv_post(), but not send. We have one recv_post in excess.
422	 */
423	atomic_inc(&rdma->excess_rc);
424	return err;
425
426 /* Handle errors that happened during or while preparing post_recv(): */
427 recv_error:
428	kfree(rpl_context);
429	spin_lock_irqsave(&rdma->req_lock, flags);
430	if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
431		rdma->state = P9_RDMA_CLOSING;
432		spin_unlock_irqrestore(&rdma->req_lock, flags);
433		rdma_disconnect(rdma->cm_id);
434	} else
435		spin_unlock_irqrestore(&rdma->req_lock, flags);
436	return err;
437}
438
439static void rdma_close(struct p9_client *client)
440{
441	struct p9_trans_rdma *rdma;
442
443	if (!client)
444		return;
445
446	rdma = client->trans;
447	if (!rdma)
448		return;
449
450	client->status = Disconnected;
451	rdma_disconnect(rdma->cm_id);
452	rdma_destroy_trans(rdma);
453}
454
455/**
456 * alloc_rdma - Allocate and initialize the rdma transport structure
457 * @opts: Mount options structure
458 */
459static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
460{
461	struct p9_trans_rdma *rdma;
462
463	rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
464	if (!rdma)
465		return NULL;
466
467	rdma->port = opts->port;
468	rdma->privport = opts->privport;
469	rdma->sq_depth = opts->sq_depth;
470	rdma->rq_depth = opts->rq_depth;
471	rdma->timeout = opts->timeout;
472	spin_lock_init(&rdma->req_lock);
473	init_completion(&rdma->cm_done);
474	sema_init(&rdma->sq_sem, rdma->sq_depth);
475	sema_init(&rdma->rq_sem, rdma->rq_depth);
476	atomic_set(&rdma->excess_rc, 0);
477
478	return rdma;
479}
480
481static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
482{
483	/* Nothing to do here.
484	 * We will take care of it (if we have to) in rdma_cancelled()
485	 */
486	return 1;
487}
488
489/* A request has been fully flushed without a reply.
490 * That means we have posted one buffer in excess.
491 */
492static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
493{
494	struct p9_trans_rdma *rdma = client->trans;
495	atomic_inc(&rdma->excess_rc);
496	return 0;
497}
498
499static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
500{
501	struct sockaddr_in cl = {
502		.sin_family = AF_INET,
503		.sin_addr.s_addr = htonl(INADDR_ANY),
504	};
505	int port, err = -EINVAL;
506
507	for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
508		cl.sin_port = htons((ushort)port);
509		err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
510		if (err != -EADDRINUSE)
511			break;
512	}
513	return err;
514}
515
516/**
517 * rdma_create_trans - Transport method for creating a transport instance
518 * @client: client instance
519 * @fc: The filesystem context
520 */
521static int
522rdma_create_trans(struct p9_client *client, struct fs_context *fc)
523{
524	const char *addr = fc->source;
525	struct v9fs_context *ctx = fc->fs_private;
526	struct p9_rdma_opts opts = ctx->rdma_opts;
527	int err;
528	struct p9_trans_rdma *rdma;
529	struct rdma_conn_param conn_param;
530	struct ib_qp_init_attr qp_attr;
531
532	if (addr == NULL)
533		return -EINVAL;
534
535	/* options are already parsed, in the fs context */
536	opts = ctx->rdma_opts;
537
538	/* Create and initialize the RDMA transport structure */
539	rdma = alloc_rdma(&opts);
540	if (!rdma)
541		return -ENOMEM;
542
543	/* Create the RDMA CM ID */
544	rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
545				     RDMA_PS_TCP, IB_QPT_RC);
546	if (IS_ERR(rdma->cm_id))
547		goto error;
548
549	/* Associate the client with the transport */
550	client->trans = rdma;
551
552	/* Bind to a privileged port if we need to */
553	if (opts.privport) {
554		err = p9_rdma_bind_privport(rdma);
555		if (err < 0) {
556			pr_err("%s (%d): problem binding to privport: %d\n",
557			       __func__, task_pid_nr(current), -err);
558			goto error;
559		}
560	}
561
562	/* Resolve the server's address */
563	rdma->addr.sin_family = AF_INET;
564	rdma->addr.sin_addr.s_addr = in_aton(addr);
565	rdma->addr.sin_port = htons(opts.port);
566	err = rdma_resolve_addr(rdma->cm_id, NULL,
567				(struct sockaddr *)&rdma->addr,
568				rdma->timeout);
569	if (err)
570		goto error;
571	err = wait_for_completion_interruptible(&rdma->cm_done);
572	if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
573		goto error;
574
575	/* Resolve the route to the server */
576	err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
577	if (err)
578		goto error;
579	err = wait_for_completion_interruptible(&rdma->cm_done);
580	if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
581		goto error;
582
583	/* Create the Completion Queue */
584	rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
585				   opts.sq_depth + opts.rq_depth + 1,
586				   IB_POLL_SOFTIRQ);
587	if (IS_ERR(rdma->cq))
588		goto error;
589
590	/* Create the Protection Domain */
591	rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
592	if (IS_ERR(rdma->pd))
593		goto error;
594
595	/* Create the Queue Pair */
596	memset(&qp_attr, 0, sizeof qp_attr);
597	qp_attr.event_handler = qp_event_handler;
598	qp_attr.qp_context = client;
599	qp_attr.cap.max_send_wr = opts.sq_depth;
600	qp_attr.cap.max_recv_wr = opts.rq_depth;
601	qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
602	qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
603	qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
604	qp_attr.qp_type = IB_QPT_RC;
605	qp_attr.send_cq = rdma->cq;
606	qp_attr.recv_cq = rdma->cq;
607	err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
608	if (err)
609		goto error;
610	rdma->qp = rdma->cm_id->qp;
611
612	/* Request a connection */
613	memset(&conn_param, 0, sizeof(conn_param));
614	conn_param.private_data = NULL;
615	conn_param.private_data_len = 0;
616	conn_param.responder_resources = P9_RDMA_IRD;
617	conn_param.initiator_depth = P9_RDMA_ORD;
618	err = rdma_connect(rdma->cm_id, &conn_param);
619	if (err)
620		goto error;
621	err = wait_for_completion_interruptible(&rdma->cm_done);
622	if (err || (rdma->state != P9_RDMA_CONNECTED))
623		goto error;
624
625	client->status = Connected;
626
627	return 0;
628
629error:
630	rdma_destroy_trans(rdma);
631	return -ENOTCONN;
632}
633
634static struct p9_trans_module p9_rdma_trans = {
635	.name = "rdma",
636	.maxsize = P9_RDMA_MAXSIZE,
637	.pooled_rbuffers = true,
638	.def = false,
639	.supports_vmalloc = false,
640	.owner = THIS_MODULE,
641	.create = rdma_create_trans,
642	.close = rdma_close,
643	.request = rdma_request,
644	.cancel = rdma_cancel,
645	.cancelled = rdma_cancelled,
646	.show_options = p9_rdma_show_options,
647};
648
649/**
650 * p9_trans_rdma_init - Register the 9P RDMA transport driver
651 */
652static int __init p9_trans_rdma_init(void)
653{
654	v9fs_register_trans(&p9_rdma_trans);
655	return 0;
656}
657
658static void __exit p9_trans_rdma_exit(void)
659{
660	v9fs_unregister_trans(&p9_rdma_trans);
661}
662
663module_init(p9_trans_rdma_init);
664module_exit(p9_trans_rdma_exit);
665MODULE_ALIAS_9P("rdma");
666
667MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
668MODULE_DESCRIPTION("RDMA Transport for 9P");
669MODULE_LICENSE("Dual BSD/GPL");