aboutsummaryrefslogtreecommitdiffstats
path: root/net
diff options
context:
space:
mode:
authorTom Tucker <tom@opengridcomputing.com>2007-12-12 17:13:21 -0500
committerJ. Bruce Fields <bfields@citi.umich.edu>2008-02-01 16:42:14 -0500
commit377f9b2f4529e0ac702fd7b91e216afd0adc959e (patch)
treeb81b97c041dcc3f3030c4df350ecf734ea8186d4 /net
parentef7fbf59e6f780a0fa03536839e3c42e9ce40ad1 (diff)
rdma: SVCRDMA Core Transport Services
This file implements the core transport data management and I/O path. The I/O path for RDMA involves receiving callbacks on interrupt context. Since all the svc transport locks are _bh locks we enqueue the transport on a list, schedule a tasklet to dequeue data indications from the RDMA completion queue. The tasklet in turn takes _bh locks to enqueue receive data indications on a list for the transport. The svc_rdma_recvfrom transport function dequeues data from this list in an NFSD thread context. Signed-off-by: Tom Tucker <tom@opengridcomputing.com> Acked-by: Neil Brown <neilb@suse.de> Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Diffstat (limited to 'net')
-rw-r--r--net/sunrpc/xprtrdma/svc_rdma_transport.c1080
1 files changed, 1080 insertions, 0 deletions
diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c
new file mode 100644
index 000000000000..f09444c451bc
--- /dev/null
+++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c
@@ -0,0 +1,1080 @@
1/*
2 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
8 * license below:
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 *
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 *
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
21 *
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
25 * permission.
26 *
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38 *
39 * Author: Tom Tucker <tom@opengridcomputing.com>
40 */
41
42#include <linux/sunrpc/svc_xprt.h>
43#include <linux/sunrpc/debug.h>
44#include <linux/sunrpc/rpc_rdma.h>
45#include <linux/spinlock.h>
46#include <rdma/ib_verbs.h>
47#include <rdma/rdma_cm.h>
48#include <linux/sunrpc/svc_rdma.h>
49
50#define RPCDBG_FACILITY RPCDBG_SVCXPRT
51
52static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
53 struct sockaddr *sa, int salen,
54 int flags);
55static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
56static void svc_rdma_release_rqst(struct svc_rqst *);
57static void rdma_destroy_xprt(struct svcxprt_rdma *xprt);
58static void dto_tasklet_func(unsigned long data);
59static void svc_rdma_detach(struct svc_xprt *xprt);
60static void svc_rdma_free(struct svc_xprt *xprt);
61static int svc_rdma_has_wspace(struct svc_xprt *xprt);
62static void rq_cq_reap(struct svcxprt_rdma *xprt);
63static void sq_cq_reap(struct svcxprt_rdma *xprt);
64
65DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
66static DEFINE_SPINLOCK(dto_lock);
67static LIST_HEAD(dto_xprt_q);
68
69static struct svc_xprt_ops svc_rdma_ops = {
70 .xpo_create = svc_rdma_create,
71 .xpo_recvfrom = svc_rdma_recvfrom,
72 .xpo_sendto = svc_rdma_sendto,
73 .xpo_release_rqst = svc_rdma_release_rqst,
74 .xpo_detach = svc_rdma_detach,
75 .xpo_free = svc_rdma_free,
76 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
77 .xpo_has_wspace = svc_rdma_has_wspace,
78 .xpo_accept = svc_rdma_accept,
79};
80
81struct svc_xprt_class svc_rdma_class = {
82 .xcl_name = "rdma",
83 .xcl_owner = THIS_MODULE,
84 .xcl_ops = &svc_rdma_ops,
85 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
86};
87
88static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
89{
90 int target;
91 int at_least_one = 0;
92 struct svc_rdma_op_ctxt *ctxt;
93
94 target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
95 xprt->sc_ctxt_max);
96
97 spin_lock_bh(&xprt->sc_ctxt_lock);
98 while (xprt->sc_ctxt_cnt < target) {
99 xprt->sc_ctxt_cnt++;
100 spin_unlock_bh(&xprt->sc_ctxt_lock);
101
102 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
103
104 spin_lock_bh(&xprt->sc_ctxt_lock);
105 if (ctxt) {
106 at_least_one = 1;
107 ctxt->next = xprt->sc_ctxt_head;
108 xprt->sc_ctxt_head = ctxt;
109 } else {
110 /* kmalloc failed...give up for now */
111 xprt->sc_ctxt_cnt--;
112 break;
113 }
114 }
115 spin_unlock_bh(&xprt->sc_ctxt_lock);
116 dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
117 xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
118 return at_least_one;
119}
120
121struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
122{
123 struct svc_rdma_op_ctxt *ctxt;
124
125 while (1) {
126 spin_lock_bh(&xprt->sc_ctxt_lock);
127 if (unlikely(xprt->sc_ctxt_head == NULL)) {
128 /* Try to bump my cache. */
129 spin_unlock_bh(&xprt->sc_ctxt_lock);
130
131 if (rdma_bump_context_cache(xprt))
132 continue;
133
134 printk(KERN_INFO "svcrdma: sleeping waiting for "
135 "context memory on xprt=%p\n",
136 xprt);
137 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
138 continue;
139 }
140 ctxt = xprt->sc_ctxt_head;
141 xprt->sc_ctxt_head = ctxt->next;
142 spin_unlock_bh(&xprt->sc_ctxt_lock);
143 ctxt->xprt = xprt;
144 INIT_LIST_HEAD(&ctxt->dto_q);
145 ctxt->count = 0;
146 break;
147 }
148 return ctxt;
149}
150
151void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
152{
153 struct svcxprt_rdma *xprt;
154 int i;
155
156 BUG_ON(!ctxt);
157 xprt = ctxt->xprt;
158 if (free_pages)
159 for (i = 0; i < ctxt->count; i++)
160 put_page(ctxt->pages[i]);
161
162 for (i = 0; i < ctxt->count; i++)
163 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
164 ctxt->sge[i].addr,
165 ctxt->sge[i].length,
166 ctxt->direction);
167 spin_lock_bh(&xprt->sc_ctxt_lock);
168 ctxt->next = xprt->sc_ctxt_head;
169 xprt->sc_ctxt_head = ctxt;
170 spin_unlock_bh(&xprt->sc_ctxt_lock);
171}
172
173/* ib_cq event handler */
174static void cq_event_handler(struct ib_event *event, void *context)
175{
176 struct svc_xprt *xprt = context;
177 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
178 event->event, context);
179 set_bit(XPT_CLOSE, &xprt->xpt_flags);
180}
181
182/* QP event handler */
183static void qp_event_handler(struct ib_event *event, void *context)
184{
185 struct svc_xprt *xprt = context;
186
187 switch (event->event) {
188 /* These are considered benign events */
189 case IB_EVENT_PATH_MIG:
190 case IB_EVENT_COMM_EST:
191 case IB_EVENT_SQ_DRAINED:
192 case IB_EVENT_QP_LAST_WQE_REACHED:
193 dprintk("svcrdma: QP event %d received for QP=%p\n",
194 event->event, event->element.qp);
195 break;
196 /* These are considered fatal events */
197 case IB_EVENT_PATH_MIG_ERR:
198 case IB_EVENT_QP_FATAL:
199 case IB_EVENT_QP_REQ_ERR:
200 case IB_EVENT_QP_ACCESS_ERR:
201 case IB_EVENT_DEVICE_FATAL:
202 default:
203 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
204 "closing transport\n",
205 event->event, event->element.qp);
206 set_bit(XPT_CLOSE, &xprt->xpt_flags);
207 break;
208 }
209}
210
211/*
212 * Data Transfer Operation Tasklet
213 *
214 * Walks a list of transports with I/O pending, removing entries as
215 * they are added to the server's I/O pending list. Two bits indicate
216 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
217 * spinlock that serializes access to the transport list with the RQ
218 * and SQ interrupt handlers.
219 */
220static void dto_tasklet_func(unsigned long data)
221{
222 struct svcxprt_rdma *xprt;
223 unsigned long flags;
224
225 spin_lock_irqsave(&dto_lock, flags);
226 while (!list_empty(&dto_xprt_q)) {
227 xprt = list_entry(dto_xprt_q.next,
228 struct svcxprt_rdma, sc_dto_q);
229 list_del_init(&xprt->sc_dto_q);
230 spin_unlock_irqrestore(&dto_lock, flags);
231
232 if (test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) {
233 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
234 rq_cq_reap(xprt);
235 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
236 /*
237 * If data arrived before established event,
238 * don't enqueue. This defers RPC I/O until the
239 * RDMA connection is complete.
240 */
241 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
242 svc_xprt_enqueue(&xprt->sc_xprt);
243 }
244
245 if (test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) {
246 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
247 sq_cq_reap(xprt);
248 }
249
250 spin_lock_irqsave(&dto_lock, flags);
251 }
252 spin_unlock_irqrestore(&dto_lock, flags);
253}
254
255/*
256 * Receive Queue Completion Handler
257 *
258 * Since an RQ completion handler is called on interrupt context, we
259 * need to defer the handling of the I/O to a tasklet
260 */
261static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
262{
263 struct svcxprt_rdma *xprt = cq_context;
264 unsigned long flags;
265
266 /*
267 * Set the bit regardless of whether or not it's on the list
268 * because it may be on the list already due to an SQ
269 * completion.
270 */
271 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
272
273 /*
274 * If this transport is not already on the DTO transport queue,
275 * add it
276 */
277 spin_lock_irqsave(&dto_lock, flags);
278 if (list_empty(&xprt->sc_dto_q))
279 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
280 spin_unlock_irqrestore(&dto_lock, flags);
281
282 /* Tasklet does all the work to avoid irqsave locks. */
283 tasklet_schedule(&dto_tasklet);
284}
285
286/*
287 * rq_cq_reap - Process the RQ CQ.
288 *
289 * Take all completing WC off the CQE and enqueue the associated DTO
290 * context on the dto_q for the transport.
291 */
292static void rq_cq_reap(struct svcxprt_rdma *xprt)
293{
294 int ret;
295 struct ib_wc wc;
296 struct svc_rdma_op_ctxt *ctxt = NULL;
297
298 atomic_inc(&rdma_stat_rq_poll);
299
300 spin_lock_bh(&xprt->sc_rq_dto_lock);
301 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
302 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
303 ctxt->wc_status = wc.status;
304 ctxt->byte_len = wc.byte_len;
305 if (wc.status != IB_WC_SUCCESS) {
306 /* Close the transport */
307 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
308 svc_rdma_put_context(ctxt, 1);
309 continue;
310 }
311 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
312 }
313 spin_unlock_bh(&xprt->sc_rq_dto_lock);
314
315 if (ctxt)
316 atomic_inc(&rdma_stat_rq_prod);
317}
318
319/*
320 * Send Queue Completion Handler - potentially called on interrupt context.
321 */
322static void sq_cq_reap(struct svcxprt_rdma *xprt)
323{
324 struct svc_rdma_op_ctxt *ctxt = NULL;
325 struct ib_wc wc;
326 struct ib_cq *cq = xprt->sc_sq_cq;
327 int ret;
328
329 atomic_inc(&rdma_stat_sq_poll);
330 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
331 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
332 xprt = ctxt->xprt;
333
334 if (wc.status != IB_WC_SUCCESS)
335 /* Close the transport */
336 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
337
338 /* Decrement used SQ WR count */
339 atomic_dec(&xprt->sc_sq_count);
340 wake_up(&xprt->sc_send_wait);
341
342 switch (ctxt->wr_op) {
343 case IB_WR_SEND:
344 case IB_WR_RDMA_WRITE:
345 svc_rdma_put_context(ctxt, 1);
346 break;
347
348 case IB_WR_RDMA_READ:
349 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
350 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
351 set_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
352 spin_lock_bh(&xprt->sc_read_complete_lock);
353 list_add_tail(&ctxt->dto_q,
354 &xprt->sc_read_complete_q);
355 spin_unlock_bh(&xprt->sc_read_complete_lock);
356 svc_xprt_enqueue(&xprt->sc_xprt);
357 }
358 break;
359
360 default:
361 printk(KERN_ERR "svcrdma: unexpected completion type, "
362 "opcode=%d, status=%d\n",
363 wc.opcode, wc.status);
364 break;
365 }
366 }
367
368 if (ctxt)
369 atomic_inc(&rdma_stat_sq_prod);
370}
371
372static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
373{
374 struct svcxprt_rdma *xprt = cq_context;
375 unsigned long flags;
376
377 /*
378 * Set the bit regardless of whether or not it's on the list
379 * because it may be on the list already due to an RQ
380 * completion.
381 */
382 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
383
384 /*
385 * If this transport is not already on the DTO transport queue,
386 * add it
387 */
388 spin_lock_irqsave(&dto_lock, flags);
389 if (list_empty(&xprt->sc_dto_q))
390 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
391 spin_unlock_irqrestore(&dto_lock, flags);
392
393 /* Tasklet does all the work to avoid irqsave locks. */
394 tasklet_schedule(&dto_tasklet);
395}
396
397static void create_context_cache(struct svcxprt_rdma *xprt,
398 int ctxt_count, int ctxt_bump, int ctxt_max)
399{
400 struct svc_rdma_op_ctxt *ctxt;
401 int i;
402
403 xprt->sc_ctxt_max = ctxt_max;
404 xprt->sc_ctxt_bump = ctxt_bump;
405 xprt->sc_ctxt_cnt = 0;
406 xprt->sc_ctxt_head = NULL;
407 for (i = 0; i < ctxt_count; i++) {
408 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
409 if (ctxt) {
410 ctxt->next = xprt->sc_ctxt_head;
411 xprt->sc_ctxt_head = ctxt;
412 xprt->sc_ctxt_cnt++;
413 }
414 }
415}
416
417static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
418{
419 struct svc_rdma_op_ctxt *next;
420 if (!ctxt)
421 return;
422
423 do {
424 next = ctxt->next;
425 kfree(ctxt);
426 ctxt = next;
427 } while (next);
428}
429
430static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
431 int listener)
432{
433 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
434
435 if (!cma_xprt)
436 return NULL;
437 svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
438 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
439 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
440 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
441 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
442 init_waitqueue_head(&cma_xprt->sc_send_wait);
443
444 spin_lock_init(&cma_xprt->sc_lock);
445 spin_lock_init(&cma_xprt->sc_read_complete_lock);
446 spin_lock_init(&cma_xprt->sc_ctxt_lock);
447 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
448
449 cma_xprt->sc_ord = svcrdma_ord;
450
451 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
452 cma_xprt->sc_max_requests = svcrdma_max_requests;
453 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
454 atomic_set(&cma_xprt->sc_sq_count, 0);
455
456 if (!listener) {
457 int reqs = cma_xprt->sc_max_requests;
458 create_context_cache(cma_xprt,
459 reqs << 1, /* starting size */
460 reqs, /* bump amount */
461 reqs +
462 cma_xprt->sc_sq_depth +
463 RPCRDMA_MAX_THREADS + 1); /* max */
464 if (!cma_xprt->sc_ctxt_head) {
465 kfree(cma_xprt);
466 return NULL;
467 }
468 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
469 } else
470 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
471
472 return cma_xprt;
473}
474
475struct page *svc_rdma_get_page(void)
476{
477 struct page *page;
478
479 while ((page = alloc_page(GFP_KERNEL)) == NULL) {
480 /* If we can't get memory, wait a bit and try again */
481 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
482 "jiffies.\n");
483 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
484 }
485 return page;
486}
487
488int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
489{
490 struct ib_recv_wr recv_wr, *bad_recv_wr;
491 struct svc_rdma_op_ctxt *ctxt;
492 struct page *page;
493 unsigned long pa;
494 int sge_no;
495 int buflen;
496 int ret;
497
498 ctxt = svc_rdma_get_context(xprt);
499 buflen = 0;
500 ctxt->direction = DMA_FROM_DEVICE;
501 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
502 BUG_ON(sge_no >= xprt->sc_max_sge);
503 page = svc_rdma_get_page();
504 ctxt->pages[sge_no] = page;
505 pa = ib_dma_map_page(xprt->sc_cm_id->device,
506 page, 0, PAGE_SIZE,
507 DMA_FROM_DEVICE);
508 ctxt->sge[sge_no].addr = pa;
509 ctxt->sge[sge_no].length = PAGE_SIZE;
510 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
511 buflen += PAGE_SIZE;
512 }
513 ctxt->count = sge_no;
514 recv_wr.next = NULL;
515 recv_wr.sg_list = &ctxt->sge[0];
516 recv_wr.num_sge = ctxt->count;
517 recv_wr.wr_id = (u64)(unsigned long)ctxt;
518
519 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
520 return ret;
521}
522
523/*
524 * This function handles the CONNECT_REQUEST event on a listening
525 * endpoint. It is passed the cma_id for the _new_ connection. The context in
526 * this cma_id is inherited from the listening cma_id and is the svc_xprt
527 * structure for the listening endpoint.
528 *
529 * This function creates a new xprt for the new connection and enqueues it on
530 * the accept queue for the listent xprt. When the listen thread is kicked, it
531 * will call the recvfrom method on the listen xprt which will accept the new
532 * connection.
533 */
534static void handle_connect_req(struct rdma_cm_id *new_cma_id)
535{
536 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
537 struct svcxprt_rdma *newxprt;
538
539 /* Create a new transport */
540 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
541 if (!newxprt) {
542 dprintk("svcrdma: failed to create new transport\n");
543 return;
544 }
545 newxprt->sc_cm_id = new_cma_id;
546 new_cma_id->context = newxprt;
547 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
548 newxprt, newxprt->sc_cm_id, listen_xprt);
549
550 /*
551 * Enqueue the new transport on the accept queue of the listening
552 * transport
553 */
554 spin_lock_bh(&listen_xprt->sc_lock);
555 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
556 spin_unlock_bh(&listen_xprt->sc_lock);
557
558 /*
559 * Can't use svc_xprt_received here because we are not on a
560 * rqstp thread
561 */
562 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
563 svc_xprt_enqueue(&listen_xprt->sc_xprt);
564}
565
566/*
567 * Handles events generated on the listening endpoint. These events will be
568 * either be incoming connect requests or adapter removal events.
569 */
570static int rdma_listen_handler(struct rdma_cm_id *cma_id,
571 struct rdma_cm_event *event)
572{
573 struct svcxprt_rdma *xprt = cma_id->context;
574 int ret = 0;
575
576 switch (event->event) {
577 case RDMA_CM_EVENT_CONNECT_REQUEST:
578 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
579 "event=%d\n", cma_id, cma_id->context, event->event);
580 handle_connect_req(cma_id);
581 break;
582
583 case RDMA_CM_EVENT_ESTABLISHED:
584 /* Accept complete */
585 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
586 "cm_id=%p\n", xprt, cma_id);
587 break;
588
589 case RDMA_CM_EVENT_DEVICE_REMOVAL:
590 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
591 xprt, cma_id);
592 if (xprt)
593 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
594 break;
595
596 default:
597 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
598 "event=%d\n", cma_id, event->event);
599 break;
600 }
601
602 return ret;
603}
604
605static int rdma_cma_handler(struct rdma_cm_id *cma_id,
606 struct rdma_cm_event *event)
607{
608 struct svc_xprt *xprt = cma_id->context;
609 struct svcxprt_rdma *rdma =
610 container_of(xprt, struct svcxprt_rdma, sc_xprt);
611 switch (event->event) {
612 case RDMA_CM_EVENT_ESTABLISHED:
613 /* Accept complete */
614 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
615 "cm_id=%p\n", xprt, cma_id);
616 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
617 svc_xprt_enqueue(xprt);
618 break;
619 case RDMA_CM_EVENT_DISCONNECTED:
620 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
621 xprt, cma_id);
622 if (xprt) {
623 set_bit(XPT_CLOSE, &xprt->xpt_flags);
624 svc_xprt_enqueue(xprt);
625 }
626 break;
627 case RDMA_CM_EVENT_DEVICE_REMOVAL:
628 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
629 "event=%d\n", cma_id, xprt, event->event);
630 if (xprt) {
631 set_bit(XPT_CLOSE, &xprt->xpt_flags);
632 svc_xprt_enqueue(xprt);
633 }
634 break;
635 default:
636 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
637 "event=%d\n", cma_id, event->event);
638 break;
639 }
640 return 0;
641}
642
643/*
644 * Create a listening RDMA service endpoint.
645 */
646static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
647 struct sockaddr *sa, int salen,
648 int flags)
649{
650 struct rdma_cm_id *listen_id;
651 struct svcxprt_rdma *cma_xprt;
652 struct svc_xprt *xprt;
653 int ret;
654
655 dprintk("svcrdma: Creating RDMA socket\n");
656
657 cma_xprt = rdma_create_xprt(serv, 1);
658 if (!cma_xprt)
659 return ERR_PTR(ENOMEM);
660 xprt = &cma_xprt->sc_xprt;
661
662 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
663 if (IS_ERR(listen_id)) {
664 rdma_destroy_xprt(cma_xprt);
665 dprintk("svcrdma: rdma_create_id failed = %ld\n",
666 PTR_ERR(listen_id));
667 return (void *)listen_id;
668 }
669 ret = rdma_bind_addr(listen_id, sa);
670 if (ret) {
671 rdma_destroy_xprt(cma_xprt);
672 rdma_destroy_id(listen_id);
673 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
674 return ERR_PTR(ret);
675 }
676 cma_xprt->sc_cm_id = listen_id;
677
678 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
679 if (ret) {
680 rdma_destroy_id(listen_id);
681 rdma_destroy_xprt(cma_xprt);
682 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
683 }
684
685 /*
686 * We need to use the address from the cm_id in case the
687 * caller specified 0 for the port number.
688 */
689 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
690 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
691
692 return &cma_xprt->sc_xprt;
693}
694
695/*
696 * This is the xpo_recvfrom function for listening endpoints. Its
697 * purpose is to accept incoming connections. The CMA callback handler
698 * has already created a new transport and attached it to the new CMA
699 * ID.
700 *
701 * There is a queue of pending connections hung on the listening
702 * transport. This queue contains the new svc_xprt structure. This
703 * function takes svc_xprt structures off the accept_q and completes
704 * the connection.
705 */
706static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
707{
708 struct svcxprt_rdma *listen_rdma;
709 struct svcxprt_rdma *newxprt = NULL;
710 struct rdma_conn_param conn_param;
711 struct ib_qp_init_attr qp_attr;
712 struct ib_device_attr devattr;
713 struct sockaddr *sa;
714 int ret;
715 int i;
716
717 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
718 clear_bit(XPT_CONN, &xprt->xpt_flags);
719 /* Get the next entry off the accept list */
720 spin_lock_bh(&listen_rdma->sc_lock);
721 if (!list_empty(&listen_rdma->sc_accept_q)) {
722 newxprt = list_entry(listen_rdma->sc_accept_q.next,
723 struct svcxprt_rdma, sc_accept_q);
724 list_del_init(&newxprt->sc_accept_q);
725 }
726 if (!list_empty(&listen_rdma->sc_accept_q))
727 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
728 spin_unlock_bh(&listen_rdma->sc_lock);
729 if (!newxprt)
730 return NULL;
731
732 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
733 newxprt, newxprt->sc_cm_id);
734
735 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
736 if (ret) {
737 dprintk("svcrdma: could not query device attributes on "
738 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
739 goto errout;
740 }
741
742 /* Qualify the transport resource defaults with the
743 * capabilities of this particular device */
744 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
745 (size_t)RPCSVC_MAXPAGES);
746 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
747 (size_t)svcrdma_max_requests);
748 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
749
750 newxprt->sc_ord = min((size_t)devattr.max_qp_rd_atom,
751 (size_t)svcrdma_ord);
752
753 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
754 if (IS_ERR(newxprt->sc_pd)) {
755 dprintk("svcrdma: error creating PD for connect request\n");
756 goto errout;
757 }
758 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
759 sq_comp_handler,
760 cq_event_handler,
761 newxprt,
762 newxprt->sc_sq_depth,
763 0);
764 if (IS_ERR(newxprt->sc_sq_cq)) {
765 dprintk("svcrdma: error creating SQ CQ for connect request\n");
766 goto errout;
767 }
768 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
769 rq_comp_handler,
770 cq_event_handler,
771 newxprt,
772 newxprt->sc_max_requests,
773 0);
774 if (IS_ERR(newxprt->sc_rq_cq)) {
775 dprintk("svcrdma: error creating RQ CQ for connect request\n");
776 goto errout;
777 }
778
779 memset(&qp_attr, 0, sizeof qp_attr);
780 qp_attr.event_handler = qp_event_handler;
781 qp_attr.qp_context = &newxprt->sc_xprt;
782 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
783 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
784 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
785 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
786 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
787 qp_attr.qp_type = IB_QPT_RC;
788 qp_attr.send_cq = newxprt->sc_sq_cq;
789 qp_attr.recv_cq = newxprt->sc_rq_cq;
790 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
791 " cm_id->device=%p, sc_pd->device=%p\n"
792 " cap.max_send_wr = %d\n"
793 " cap.max_recv_wr = %d\n"
794 " cap.max_send_sge = %d\n"
795 " cap.max_recv_sge = %d\n",
796 newxprt->sc_cm_id, newxprt->sc_pd,
797 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
798 qp_attr.cap.max_send_wr,
799 qp_attr.cap.max_recv_wr,
800 qp_attr.cap.max_send_sge,
801 qp_attr.cap.max_recv_sge);
802
803 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
804 if (ret) {
805 /*
806 * XXX: This is a hack. We need a xx_request_qp interface
807 * that will adjust the qp_attr's with a best-effort
808 * number
809 */
810 qp_attr.cap.max_send_sge -= 2;
811 qp_attr.cap.max_recv_sge -= 2;
812 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
813 &qp_attr);
814 if (ret) {
815 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
816 goto errout;
817 }
818 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
819 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
820 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
821 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
822 }
823 newxprt->sc_qp = newxprt->sc_cm_id->qp;
824
825 /* Register all of physical memory */
826 newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
827 IB_ACCESS_LOCAL_WRITE |
828 IB_ACCESS_REMOTE_WRITE);
829 if (IS_ERR(newxprt->sc_phys_mr)) {
830 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
831 goto errout;
832 }
833
834 /* Post receive buffers */
835 for (i = 0; i < newxprt->sc_max_requests; i++) {
836 ret = svc_rdma_post_recv(newxprt);
837 if (ret) {
838 dprintk("svcrdma: failure posting receive buffers\n");
839 goto errout;
840 }
841 }
842
843 /* Swap out the handler */
844 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
845
846 /* Accept Connection */
847 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
848 memset(&conn_param, 0, sizeof conn_param);
849 conn_param.responder_resources = 0;
850 conn_param.initiator_depth = newxprt->sc_ord;
851 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
852 if (ret) {
853 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
854 ret);
855 goto errout;
856 }
857
858 dprintk("svcrdma: new connection %p accepted with the following "
859 "attributes:\n"
860 " local_ip : %d.%d.%d.%d\n"
861 " local_port : %d\n"
862 " remote_ip : %d.%d.%d.%d\n"
863 " remote_port : %d\n"
864 " max_sge : %d\n"
865 " sq_depth : %d\n"
866 " max_requests : %d\n"
867 " ord : %d\n",
868 newxprt,
869 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
870 route.addr.src_addr)->sin_addr.s_addr),
871 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
872 route.addr.src_addr)->sin_port),
873 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
874 route.addr.dst_addr)->sin_addr.s_addr),
875 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
876 route.addr.dst_addr)->sin_port),
877 newxprt->sc_max_sge,
878 newxprt->sc_sq_depth,
879 newxprt->sc_max_requests,
880 newxprt->sc_ord);
881
882 /* Set the local and remote addresses in the transport */
883 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
884 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
885 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
886 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
887
888 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
889 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
890 return &newxprt->sc_xprt;
891
892 errout:
893 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
894 rdma_destroy_id(newxprt->sc_cm_id);
895 rdma_destroy_xprt(newxprt);
896 return NULL;
897}
898
899/*
900 * Post an RQ WQE to the RQ when the rqst is being released. This
901 * effectively returns an RQ credit to the client. The rq_xprt_ctxt
902 * will be null if the request is deferred due to an RDMA_READ or the
903 * transport had no data ready (EAGAIN). Note that an RPC deferred in
904 * svc_process will still return the credit, this is because the data
905 * is copied and no longer consume a WQE/WC.
906 */
907static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
908{
909 int err;
910 struct svcxprt_rdma *rdma =
911 container_of(rqstp->rq_xprt, struct svcxprt_rdma, sc_xprt);
912 if (rqstp->rq_xprt_ctxt) {
913 BUG_ON(rqstp->rq_xprt_ctxt != rdma);
914 err = svc_rdma_post_recv(rdma);
915 if (err)
916 dprintk("svcrdma: failed to post an RQ WQE error=%d\n",
917 err);
918 }
919 rqstp->rq_xprt_ctxt = NULL;
920}
921
922/* Disable data ready events for this connection */
923static void svc_rdma_detach(struct svc_xprt *xprt)
924{
925 struct svcxprt_rdma *rdma =
926 container_of(xprt, struct svcxprt_rdma, sc_xprt);
927 unsigned long flags;
928
929 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
930 /*
931 * Shutdown the connection. This will ensure we don't get any
932 * more events from the provider.
933 */
934 rdma_disconnect(rdma->sc_cm_id);
935 rdma_destroy_id(rdma->sc_cm_id);
936
937 /* We may already be on the DTO list */
938 spin_lock_irqsave(&dto_lock, flags);
939 if (!list_empty(&rdma->sc_dto_q))
940 list_del_init(&rdma->sc_dto_q);
941 spin_unlock_irqrestore(&dto_lock, flags);
942}
943
944static void svc_rdma_free(struct svc_xprt *xprt)
945{
946 struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
947 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
948 rdma_destroy_xprt(rdma);
949 kfree(rdma);
950}
951
952static void rdma_destroy_xprt(struct svcxprt_rdma *xprt)
953{
954 if (xprt->sc_qp && !IS_ERR(xprt->sc_qp))
955 ib_destroy_qp(xprt->sc_qp);
956
957 if (xprt->sc_sq_cq && !IS_ERR(xprt->sc_sq_cq))
958 ib_destroy_cq(xprt->sc_sq_cq);
959
960 if (xprt->sc_rq_cq && !IS_ERR(xprt->sc_rq_cq))
961 ib_destroy_cq(xprt->sc_rq_cq);
962
963 if (xprt->sc_phys_mr && !IS_ERR(xprt->sc_phys_mr))
964 ib_dereg_mr(xprt->sc_phys_mr);
965
966 if (xprt->sc_pd && !IS_ERR(xprt->sc_pd))
967 ib_dealloc_pd(xprt->sc_pd);
968
969 destroy_context_cache(xprt->sc_ctxt_head);
970}
971
972static int svc_rdma_has_wspace(struct svc_xprt *xprt)
973{
974 struct svcxprt_rdma *rdma =
975 container_of(xprt, struct svcxprt_rdma, sc_xprt);
976
977 /*
978 * If there are fewer SQ WR available than required to send a
979 * simple response, return false.
980 */
981 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
982 return 0;
983
984 /*
985 * ...or there are already waiters on the SQ,
986 * return false.
987 */
988 if (waitqueue_active(&rdma->sc_send_wait))
989 return 0;
990
991 /* Otherwise return true. */
992 return 1;
993}
994
995int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
996{
997 struct ib_send_wr *bad_wr;
998 int ret;
999
1000 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1001 return 0;
1002
1003 BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1004 BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1005 wr->opcode);
1006 /* If the SQ is full, wait until an SQ entry is available */
1007 while (1) {
1008 spin_lock_bh(&xprt->sc_lock);
1009 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1010 spin_unlock_bh(&xprt->sc_lock);
1011 atomic_inc(&rdma_stat_sq_starve);
1012 /* See if we can reap some SQ WR */
1013 sq_cq_reap(xprt);
1014
1015 /* Wait until SQ WR available if SQ still full */
1016 wait_event(xprt->sc_send_wait,
1017 atomic_read(&xprt->sc_sq_count) <
1018 xprt->sc_sq_depth);
1019 continue;
1020 }
1021 /* Bumped used SQ WR count and post */
1022 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1023 if (!ret)
1024 atomic_inc(&xprt->sc_sq_count);
1025 else
1026 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1027 "sc_sq_count=%d, sc_sq_depth=%d\n",
1028 ret, atomic_read(&xprt->sc_sq_count),
1029 xprt->sc_sq_depth);
1030 spin_unlock_bh(&xprt->sc_lock);
1031 break;
1032 }
1033 return ret;
1034}
1035
1036int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1037 enum rpcrdma_errcode err)
1038{
1039 struct ib_send_wr err_wr;
1040 struct ib_sge sge;
1041 struct page *p;
1042 struct svc_rdma_op_ctxt *ctxt;
1043 u32 *va;
1044 int length;
1045 int ret;
1046
1047 p = svc_rdma_get_page();
1048 va = page_address(p);
1049
1050 /* XDR encode error */
1051 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1052
1053 /* Prepare SGE for local address */
1054 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1055 p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1056 sge.lkey = xprt->sc_phys_mr->lkey;
1057 sge.length = length;
1058
1059 ctxt = svc_rdma_get_context(xprt);
1060 ctxt->count = 1;
1061 ctxt->pages[0] = p;
1062
1063 /* Prepare SEND WR */
1064 memset(&err_wr, 0, sizeof err_wr);
1065 ctxt->wr_op = IB_WR_SEND;
1066 err_wr.wr_id = (unsigned long)ctxt;
1067 err_wr.sg_list = &sge;
1068 err_wr.num_sge = 1;
1069 err_wr.opcode = IB_WR_SEND;
1070 err_wr.send_flags = IB_SEND_SIGNALED;
1071
1072 /* Post It */
1073 ret = svc_rdma_send(xprt, &err_wr);
1074 if (ret) {
1075 dprintk("svcrdma: Error posting send = %d\n", ret);
1076 svc_rdma_put_context(ctxt, 1);
1077 }
1078
1079 return ret;
1080}
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-16 07:55:41 -0500