svcrdma: Fix return value in svc_rdma_send
[safe/jmp/linux-2.6] / net / sunrpc / xprtrdma / svc_rdma_transport.c
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
52 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
53                                         struct sockaddr *sa, int salen,
54                                         int flags);
55 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
56 static void svc_rdma_release_rqst(struct svc_rqst *);
57 static void dto_tasklet_func(unsigned long data);
58 static void svc_rdma_detach(struct svc_xprt *xprt);
59 static void svc_rdma_free(struct svc_xprt *xprt);
60 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
61 static void rq_cq_reap(struct svcxprt_rdma *xprt);
62 static void sq_cq_reap(struct svcxprt_rdma *xprt);
63
64 DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
65 static DEFINE_SPINLOCK(dto_lock);
66 static LIST_HEAD(dto_xprt_q);
67
68 static struct svc_xprt_ops svc_rdma_ops = {
69         .xpo_create = svc_rdma_create,
70         .xpo_recvfrom = svc_rdma_recvfrom,
71         .xpo_sendto = svc_rdma_sendto,
72         .xpo_release_rqst = svc_rdma_release_rqst,
73         .xpo_detach = svc_rdma_detach,
74         .xpo_free = svc_rdma_free,
75         .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
76         .xpo_has_wspace = svc_rdma_has_wspace,
77         .xpo_accept = svc_rdma_accept,
78 };
79
80 struct svc_xprt_class svc_rdma_class = {
81         .xcl_name = "rdma",
82         .xcl_owner = THIS_MODULE,
83         .xcl_ops = &svc_rdma_ops,
84         .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
85 };
86
87 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
88 {
89         int target;
90         int at_least_one = 0;
91         struct svc_rdma_op_ctxt *ctxt;
92
93         target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
94                      xprt->sc_ctxt_max);
95
96         spin_lock_bh(&xprt->sc_ctxt_lock);
97         while (xprt->sc_ctxt_cnt < target) {
98                 xprt->sc_ctxt_cnt++;
99                 spin_unlock_bh(&xprt->sc_ctxt_lock);
100
101                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
102
103                 spin_lock_bh(&xprt->sc_ctxt_lock);
104                 if (ctxt) {
105                         at_least_one = 1;
106                         ctxt->next = xprt->sc_ctxt_head;
107                         xprt->sc_ctxt_head = ctxt;
108                 } else {
109                         /* kmalloc failed...give up for now */
110                         xprt->sc_ctxt_cnt--;
111                         break;
112                 }
113         }
114         spin_unlock_bh(&xprt->sc_ctxt_lock);
115         dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
116                 xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
117         return at_least_one;
118 }
119
120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
121 {
122         struct svc_rdma_op_ctxt *ctxt;
123
124         while (1) {
125                 spin_lock_bh(&xprt->sc_ctxt_lock);
126                 if (unlikely(xprt->sc_ctxt_head == NULL)) {
127                         /* Try to bump my cache. */
128                         spin_unlock_bh(&xprt->sc_ctxt_lock);
129
130                         if (rdma_bump_context_cache(xprt))
131                                 continue;
132
133                         printk(KERN_INFO "svcrdma: sleeping waiting for "
134                                "context memory on xprt=%p\n",
135                                xprt);
136                         schedule_timeout_uninterruptible(msecs_to_jiffies(500));
137                         continue;
138                 }
139                 ctxt = xprt->sc_ctxt_head;
140                 xprt->sc_ctxt_head = ctxt->next;
141                 spin_unlock_bh(&xprt->sc_ctxt_lock);
142                 ctxt->xprt = xprt;
143                 INIT_LIST_HEAD(&ctxt->dto_q);
144                 ctxt->count = 0;
145                 break;
146         }
147         return ctxt;
148 }
149
150 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
151 {
152         struct svcxprt_rdma *xprt;
153         int i;
154
155         BUG_ON(!ctxt);
156         xprt = ctxt->xprt;
157         if (free_pages)
158                 for (i = 0; i < ctxt->count; i++)
159                         put_page(ctxt->pages[i]);
160
161         for (i = 0; i < ctxt->count; i++)
162                 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
163                                  ctxt->sge[i].addr,
164                                  ctxt->sge[i].length,
165                                  ctxt->direction);
166         spin_lock_bh(&xprt->sc_ctxt_lock);
167         ctxt->next = xprt->sc_ctxt_head;
168         xprt->sc_ctxt_head = ctxt;
169         spin_unlock_bh(&xprt->sc_ctxt_lock);
170 }
171
172 /* ib_cq event handler */
173 static void cq_event_handler(struct ib_event *event, void *context)
174 {
175         struct svc_xprt *xprt = context;
176         dprintk("svcrdma: received CQ event id=%d, context=%p\n",
177                 event->event, context);
178         set_bit(XPT_CLOSE, &xprt->xpt_flags);
179 }
180
181 /* QP event handler */
182 static void qp_event_handler(struct ib_event *event, void *context)
183 {
184         struct svc_xprt *xprt = context;
185
186         switch (event->event) {
187         /* These are considered benign events */
188         case IB_EVENT_PATH_MIG:
189         case IB_EVENT_COMM_EST:
190         case IB_EVENT_SQ_DRAINED:
191         case IB_EVENT_QP_LAST_WQE_REACHED:
192                 dprintk("svcrdma: QP event %d received for QP=%p\n",
193                         event->event, event->element.qp);
194                 break;
195         /* These are considered fatal events */
196         case IB_EVENT_PATH_MIG_ERR:
197         case IB_EVENT_QP_FATAL:
198         case IB_EVENT_QP_REQ_ERR:
199         case IB_EVENT_QP_ACCESS_ERR:
200         case IB_EVENT_DEVICE_FATAL:
201         default:
202                 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
203                         "closing transport\n",
204                         event->event, event->element.qp);
205                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
206                 break;
207         }
208 }
209
210 /*
211  * Data Transfer Operation Tasklet
212  *
213  * Walks a list of transports with I/O pending, removing entries as
214  * they are added to the server's I/O pending list. Two bits indicate
215  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
216  * spinlock that serializes access to the transport list with the RQ
217  * and SQ interrupt handlers.
218  */
219 static void dto_tasklet_func(unsigned long data)
220 {
221         struct svcxprt_rdma *xprt;
222         unsigned long flags;
223
224         spin_lock_irqsave(&dto_lock, flags);
225         while (!list_empty(&dto_xprt_q)) {
226                 xprt = list_entry(dto_xprt_q.next,
227                                   struct svcxprt_rdma, sc_dto_q);
228                 list_del_init(&xprt->sc_dto_q);
229                 spin_unlock_irqrestore(&dto_lock, flags);
230
231                 rq_cq_reap(xprt);
232                 sq_cq_reap(xprt);
233
234                 svc_xprt_put(&xprt->sc_xprt);
235                 spin_lock_irqsave(&dto_lock, flags);
236         }
237         spin_unlock_irqrestore(&dto_lock, flags);
238 }
239
240 /*
241  * Receive Queue Completion Handler
242  *
243  * Since an RQ completion handler is called on interrupt context, we
244  * need to defer the handling of the I/O to a tasklet
245  */
246 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
247 {
248         struct svcxprt_rdma *xprt = cq_context;
249         unsigned long flags;
250
251         /*
252          * Set the bit regardless of whether or not it's on the list
253          * because it may be on the list already due to an SQ
254          * completion.
255         */
256         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
257
258         /*
259          * If this transport is not already on the DTO transport queue,
260          * add it
261          */
262         spin_lock_irqsave(&dto_lock, flags);
263         if (list_empty(&xprt->sc_dto_q)) {
264                 svc_xprt_get(&xprt->sc_xprt);
265                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
266         }
267         spin_unlock_irqrestore(&dto_lock, flags);
268
269         /* Tasklet does all the work to avoid irqsave locks. */
270         tasklet_schedule(&dto_tasklet);
271 }
272
273 /*
274  * rq_cq_reap - Process the RQ CQ.
275  *
276  * Take all completing WC off the CQE and enqueue the associated DTO
277  * context on the dto_q for the transport.
278  */
279 static void rq_cq_reap(struct svcxprt_rdma *xprt)
280 {
281         int ret;
282         struct ib_wc wc;
283         struct svc_rdma_op_ctxt *ctxt = NULL;
284
285         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
286                 return;
287
288         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
289         atomic_inc(&rdma_stat_rq_poll);
290
291         spin_lock_bh(&xprt->sc_rq_dto_lock);
292         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
293                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
294                 ctxt->wc_status = wc.status;
295                 ctxt->byte_len = wc.byte_len;
296                 if (wc.status != IB_WC_SUCCESS) {
297                         /* Close the transport */
298                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
299                         svc_rdma_put_context(ctxt, 1);
300                         continue;
301                 }
302                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
303         }
304         spin_unlock_bh(&xprt->sc_rq_dto_lock);
305
306         if (ctxt)
307                 atomic_inc(&rdma_stat_rq_prod);
308
309         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
310         /*
311          * If data arrived before established event,
312          * don't enqueue. This defers RPC I/O until the
313          * RDMA connection is complete.
314          */
315         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
316                 svc_xprt_enqueue(&xprt->sc_xprt);
317 }
318
319 /*
320  * Send Queue Completion Handler - potentially called on interrupt context.
321  */
322 static 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
330         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
331                 return;
332
333         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
334         atomic_inc(&rdma_stat_sq_poll);
335         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
336                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
337                 xprt = ctxt->xprt;
338
339                 if (wc.status != IB_WC_SUCCESS)
340                         /* Close the transport */
341                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
342
343                 /* Decrement used SQ WR count */
344                 atomic_dec(&xprt->sc_sq_count);
345                 wake_up(&xprt->sc_send_wait);
346
347                 switch (ctxt->wr_op) {
348                 case IB_WR_SEND:
349                 case IB_WR_RDMA_WRITE:
350                         svc_rdma_put_context(ctxt, 1);
351                         break;
352
353                 case IB_WR_RDMA_READ:
354                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
355                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
356                                 set_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
357                                 spin_lock_bh(&xprt->sc_read_complete_lock);
358                                 list_add_tail(&ctxt->dto_q,
359                                               &xprt->sc_read_complete_q);
360                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
361                                 svc_xprt_enqueue(&xprt->sc_xprt);
362                         }
363                         break;
364
365                 default:
366                         printk(KERN_ERR "svcrdma: unexpected completion type, "
367                                "opcode=%d, status=%d\n",
368                                wc.opcode, wc.status);
369                         break;
370                 }
371         }
372
373         if (ctxt)
374                 atomic_inc(&rdma_stat_sq_prod);
375 }
376
377 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
378 {
379         struct svcxprt_rdma *xprt = cq_context;
380         unsigned long flags;
381
382         /*
383          * Set the bit regardless of whether or not it's on the list
384          * because it may be on the list already due to an RQ
385          * completion.
386         */
387         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
388
389         /*
390          * If this transport is not already on the DTO transport queue,
391          * add it
392          */
393         spin_lock_irqsave(&dto_lock, flags);
394         if (list_empty(&xprt->sc_dto_q)) {
395                 svc_xprt_get(&xprt->sc_xprt);
396                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
397         }
398         spin_unlock_irqrestore(&dto_lock, flags);
399
400         /* Tasklet does all the work to avoid irqsave locks. */
401         tasklet_schedule(&dto_tasklet);
402 }
403
404 static void create_context_cache(struct svcxprt_rdma *xprt,
405                                  int ctxt_count, int ctxt_bump, int ctxt_max)
406 {
407         struct svc_rdma_op_ctxt *ctxt;
408         int i;
409
410         xprt->sc_ctxt_max = ctxt_max;
411         xprt->sc_ctxt_bump = ctxt_bump;
412         xprt->sc_ctxt_cnt = 0;
413         xprt->sc_ctxt_head = NULL;
414         for (i = 0; i < ctxt_count; i++) {
415                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
416                 if (ctxt) {
417                         ctxt->next = xprt->sc_ctxt_head;
418                         xprt->sc_ctxt_head = ctxt;
419                         xprt->sc_ctxt_cnt++;
420                 }
421         }
422 }
423
424 static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
425 {
426         struct svc_rdma_op_ctxt *next;
427         if (!ctxt)
428                 return;
429
430         do {
431                 next = ctxt->next;
432                 kfree(ctxt);
433                 ctxt = next;
434         } while (next);
435 }
436
437 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
438                                              int listener)
439 {
440         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
441
442         if (!cma_xprt)
443                 return NULL;
444         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
445         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
446         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
447         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
448         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
449         init_waitqueue_head(&cma_xprt->sc_send_wait);
450
451         spin_lock_init(&cma_xprt->sc_lock);
452         spin_lock_init(&cma_xprt->sc_read_complete_lock);
453         spin_lock_init(&cma_xprt->sc_ctxt_lock);
454         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
455
456         cma_xprt->sc_ord = svcrdma_ord;
457
458         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
459         cma_xprt->sc_max_requests = svcrdma_max_requests;
460         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
461         atomic_set(&cma_xprt->sc_sq_count, 0);
462
463         if (!listener) {
464                 int reqs = cma_xprt->sc_max_requests;
465                 create_context_cache(cma_xprt,
466                                      reqs << 1, /* starting size */
467                                      reqs,      /* bump amount */
468                                      reqs +
469                                      cma_xprt->sc_sq_depth +
470                                      RPCRDMA_MAX_THREADS + 1); /* max */
471                 if (!cma_xprt->sc_ctxt_head) {
472                         kfree(cma_xprt);
473                         return NULL;
474                 }
475                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
476         } else
477                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
478
479         return cma_xprt;
480 }
481
482 struct page *svc_rdma_get_page(void)
483 {
484         struct page *page;
485
486         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
487                 /* If we can't get memory, wait a bit and try again */
488                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
489                        "jiffies.\n");
490                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
491         }
492         return page;
493 }
494
495 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
496 {
497         struct ib_recv_wr recv_wr, *bad_recv_wr;
498         struct svc_rdma_op_ctxt *ctxt;
499         struct page *page;
500         unsigned long pa;
501         int sge_no;
502         int buflen;
503         int ret;
504
505         ctxt = svc_rdma_get_context(xprt);
506         buflen = 0;
507         ctxt->direction = DMA_FROM_DEVICE;
508         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
509                 BUG_ON(sge_no >= xprt->sc_max_sge);
510                 page = svc_rdma_get_page();
511                 ctxt->pages[sge_no] = page;
512                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
513                                      page, 0, PAGE_SIZE,
514                                      DMA_FROM_DEVICE);
515                 ctxt->sge[sge_no].addr = pa;
516                 ctxt->sge[sge_no].length = PAGE_SIZE;
517                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
518                 buflen += PAGE_SIZE;
519         }
520         ctxt->count = sge_no;
521         recv_wr.next = NULL;
522         recv_wr.sg_list = &ctxt->sge[0];
523         recv_wr.num_sge = ctxt->count;
524         recv_wr.wr_id = (u64)(unsigned long)ctxt;
525
526         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
527         return ret;
528 }
529
530 /*
531  * This function handles the CONNECT_REQUEST event on a listening
532  * endpoint. It is passed the cma_id for the _new_ connection. The context in
533  * this cma_id is inherited from the listening cma_id and is the svc_xprt
534  * structure for the listening endpoint.
535  *
536  * This function creates a new xprt for the new connection and enqueues it on
537  * the accept queue for the listent xprt. When the listen thread is kicked, it
538  * will call the recvfrom method on the listen xprt which will accept the new
539  * connection.
540  */
541 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
542 {
543         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
544         struct svcxprt_rdma *newxprt;
545
546         /* Create a new transport */
547         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
548         if (!newxprt) {
549                 dprintk("svcrdma: failed to create new transport\n");
550                 return;
551         }
552         newxprt->sc_cm_id = new_cma_id;
553         new_cma_id->context = newxprt;
554         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
555                 newxprt, newxprt->sc_cm_id, listen_xprt);
556
557         /*
558          * Enqueue the new transport on the accept queue of the listening
559          * transport
560          */
561         spin_lock_bh(&listen_xprt->sc_lock);
562         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
563         spin_unlock_bh(&listen_xprt->sc_lock);
564
565         /*
566          * Can't use svc_xprt_received here because we are not on a
567          * rqstp thread
568         */
569         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
570         svc_xprt_enqueue(&listen_xprt->sc_xprt);
571 }
572
573 /*
574  * Handles events generated on the listening endpoint. These events will be
575  * either be incoming connect requests or adapter removal  events.
576  */
577 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
578                                struct rdma_cm_event *event)
579 {
580         struct svcxprt_rdma *xprt = cma_id->context;
581         int ret = 0;
582
583         switch (event->event) {
584         case RDMA_CM_EVENT_CONNECT_REQUEST:
585                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
586                         "event=%d\n", cma_id, cma_id->context, event->event);
587                 handle_connect_req(cma_id);
588                 break;
589
590         case RDMA_CM_EVENT_ESTABLISHED:
591                 /* Accept complete */
592                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
593                         "cm_id=%p\n", xprt, cma_id);
594                 break;
595
596         case RDMA_CM_EVENT_DEVICE_REMOVAL:
597                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
598                         xprt, cma_id);
599                 if (xprt)
600                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
601                 break;
602
603         default:
604                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
605                         "event=%d\n", cma_id, event->event);
606                 break;
607         }
608
609         return ret;
610 }
611
612 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
613                             struct rdma_cm_event *event)
614 {
615         struct svc_xprt *xprt = cma_id->context;
616         struct svcxprt_rdma *rdma =
617                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
618         switch (event->event) {
619         case RDMA_CM_EVENT_ESTABLISHED:
620                 /* Accept complete */
621                 svc_xprt_get(xprt);
622                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
623                         "cm_id=%p\n", xprt, cma_id);
624                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
625                 svc_xprt_enqueue(xprt);
626                 break;
627         case RDMA_CM_EVENT_DISCONNECTED:
628                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
629                         xprt, cma_id);
630                 if (xprt) {
631                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
632                         svc_xprt_enqueue(xprt);
633                 }
634                 break;
635         case RDMA_CM_EVENT_DEVICE_REMOVAL:
636                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
637                         "event=%d\n", cma_id, xprt, event->event);
638                 if (xprt) {
639                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
640                         svc_xprt_enqueue(xprt);
641                 }
642                 break;
643         default:
644                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
645                         "event=%d\n", cma_id, event->event);
646                 break;
647         }
648         return 0;
649 }
650
651 /*
652  * Create a listening RDMA service endpoint.
653  */
654 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
655                                         struct sockaddr *sa, int salen,
656                                         int flags)
657 {
658         struct rdma_cm_id *listen_id;
659         struct svcxprt_rdma *cma_xprt;
660         struct svc_xprt *xprt;
661         int ret;
662
663         dprintk("svcrdma: Creating RDMA socket\n");
664
665         cma_xprt = rdma_create_xprt(serv, 1);
666         if (!cma_xprt)
667                 return ERR_PTR(ENOMEM);
668         xprt = &cma_xprt->sc_xprt;
669
670         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
671         if (IS_ERR(listen_id)) {
672                 svc_xprt_put(&cma_xprt->sc_xprt);
673                 dprintk("svcrdma: rdma_create_id failed = %ld\n",
674                         PTR_ERR(listen_id));
675                 return (void *)listen_id;
676         }
677         ret = rdma_bind_addr(listen_id, sa);
678         if (ret) {
679                 rdma_destroy_id(listen_id);
680                 svc_xprt_put(&cma_xprt->sc_xprt);
681                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
682                 return ERR_PTR(ret);
683         }
684         cma_xprt->sc_cm_id = listen_id;
685
686         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
687         if (ret) {
688                 rdma_destroy_id(listen_id);
689                 svc_xprt_put(&cma_xprt->sc_xprt);
690                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
691                 return ERR_PTR(ret);
692         }
693
694         /*
695          * We need to use the address from the cm_id in case the
696          * caller specified 0 for the port number.
697          */
698         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
699         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
700
701         return &cma_xprt->sc_xprt;
702 }
703
704 /*
705  * This is the xpo_recvfrom function for listening endpoints. Its
706  * purpose is to accept incoming connections. The CMA callback handler
707  * has already created a new transport and attached it to the new CMA
708  * ID.
709  *
710  * There is a queue of pending connections hung on the listening
711  * transport. This queue contains the new svc_xprt structure. This
712  * function takes svc_xprt structures off the accept_q and completes
713  * the connection.
714  */
715 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
716 {
717         struct svcxprt_rdma *listen_rdma;
718         struct svcxprt_rdma *newxprt = NULL;
719         struct rdma_conn_param conn_param;
720         struct ib_qp_init_attr qp_attr;
721         struct ib_device_attr devattr;
722         struct sockaddr *sa;
723         int ret;
724         int i;
725
726         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
727         clear_bit(XPT_CONN, &xprt->xpt_flags);
728         /* Get the next entry off the accept list */
729         spin_lock_bh(&listen_rdma->sc_lock);
730         if (!list_empty(&listen_rdma->sc_accept_q)) {
731                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
732                                      struct svcxprt_rdma, sc_accept_q);
733                 list_del_init(&newxprt->sc_accept_q);
734         }
735         if (!list_empty(&listen_rdma->sc_accept_q))
736                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
737         spin_unlock_bh(&listen_rdma->sc_lock);
738         if (!newxprt)
739                 return NULL;
740
741         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
742                 newxprt, newxprt->sc_cm_id);
743
744         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
745         if (ret) {
746                 dprintk("svcrdma: could not query device attributes on "
747                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
748                 goto errout;
749         }
750
751         /* Qualify the transport resource defaults with the
752          * capabilities of this particular device */
753         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
754                                   (size_t)RPCSVC_MAXPAGES);
755         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
756                                    (size_t)svcrdma_max_requests);
757         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
758
759         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
760                                (size_t)svcrdma_ord);
761
762         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
763         if (IS_ERR(newxprt->sc_pd)) {
764                 dprintk("svcrdma: error creating PD for connect request\n");
765                 goto errout;
766         }
767         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
768                                          sq_comp_handler,
769                                          cq_event_handler,
770                                          newxprt,
771                                          newxprt->sc_sq_depth,
772                                          0);
773         if (IS_ERR(newxprt->sc_sq_cq)) {
774                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
775                 goto errout;
776         }
777         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
778                                          rq_comp_handler,
779                                          cq_event_handler,
780                                          newxprt,
781                                          newxprt->sc_max_requests,
782                                          0);
783         if (IS_ERR(newxprt->sc_rq_cq)) {
784                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
785                 goto errout;
786         }
787
788         memset(&qp_attr, 0, sizeof qp_attr);
789         qp_attr.event_handler = qp_event_handler;
790         qp_attr.qp_context = &newxprt->sc_xprt;
791         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
792         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
793         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
794         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
795         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
796         qp_attr.qp_type = IB_QPT_RC;
797         qp_attr.send_cq = newxprt->sc_sq_cq;
798         qp_attr.recv_cq = newxprt->sc_rq_cq;
799         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
800                 "    cm_id->device=%p, sc_pd->device=%p\n"
801                 "    cap.max_send_wr = %d\n"
802                 "    cap.max_recv_wr = %d\n"
803                 "    cap.max_send_sge = %d\n"
804                 "    cap.max_recv_sge = %d\n",
805                 newxprt->sc_cm_id, newxprt->sc_pd,
806                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
807                 qp_attr.cap.max_send_wr,
808                 qp_attr.cap.max_recv_wr,
809                 qp_attr.cap.max_send_sge,
810                 qp_attr.cap.max_recv_sge);
811
812         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
813         if (ret) {
814                 /*
815                  * XXX: This is a hack. We need a xx_request_qp interface
816                  * that will adjust the qp_attr's with a best-effort
817                  * number
818                  */
819                 qp_attr.cap.max_send_sge -= 2;
820                 qp_attr.cap.max_recv_sge -= 2;
821                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
822                                      &qp_attr);
823                 if (ret) {
824                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
825                         goto errout;
826                 }
827                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
828                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
829                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
830                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
831         }
832         svc_xprt_get(&newxprt->sc_xprt);
833         newxprt->sc_qp = newxprt->sc_cm_id->qp;
834
835         /* Register all of physical memory */
836         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
837                                             IB_ACCESS_LOCAL_WRITE |
838                                             IB_ACCESS_REMOTE_WRITE);
839         if (IS_ERR(newxprt->sc_phys_mr)) {
840                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
841                 goto errout;
842         }
843
844         /* Post receive buffers */
845         for (i = 0; i < newxprt->sc_max_requests; i++) {
846                 ret = svc_rdma_post_recv(newxprt);
847                 if (ret) {
848                         dprintk("svcrdma: failure posting receive buffers\n");
849                         goto errout;
850                 }
851         }
852
853         /* Swap out the handler */
854         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
855
856         /* Accept Connection */
857         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
858         memset(&conn_param, 0, sizeof conn_param);
859         conn_param.responder_resources = 0;
860         conn_param.initiator_depth = newxprt->sc_ord;
861         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
862         if (ret) {
863                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
864                        ret);
865                 goto errout;
866         }
867
868         dprintk("svcrdma: new connection %p accepted with the following "
869                 "attributes:\n"
870                 "    local_ip        : %d.%d.%d.%d\n"
871                 "    local_port      : %d\n"
872                 "    remote_ip       : %d.%d.%d.%d\n"
873                 "    remote_port     : %d\n"
874                 "    max_sge         : %d\n"
875                 "    sq_depth        : %d\n"
876                 "    max_requests    : %d\n"
877                 "    ord             : %d\n",
878                 newxprt,
879                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
880                          route.addr.src_addr)->sin_addr.s_addr),
881                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
882                        route.addr.src_addr)->sin_port),
883                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
884                          route.addr.dst_addr)->sin_addr.s_addr),
885                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
886                        route.addr.dst_addr)->sin_port),
887                 newxprt->sc_max_sge,
888                 newxprt->sc_sq_depth,
889                 newxprt->sc_max_requests,
890                 newxprt->sc_ord);
891
892         /* Set the local and remote addresses in the transport */
893         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
894         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
895         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
896         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
897
898         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
899         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
900         return &newxprt->sc_xprt;
901
902  errout:
903         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
904         /* Take a reference in case the DTO handler runs */
905         svc_xprt_get(&newxprt->sc_xprt);
906         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
907                 ib_destroy_qp(newxprt->sc_qp);
908                 svc_xprt_put(&newxprt->sc_xprt);
909         }
910         rdma_destroy_id(newxprt->sc_cm_id);
911         /* This call to put will destroy the transport */
912         svc_xprt_put(&newxprt->sc_xprt);
913         return NULL;
914 }
915
916 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
917 {
918 }
919
920 /*
921  * When connected, an svc_xprt has at least three references:
922  *
923  * - A reference held by the QP. We still hold that here because this
924  *   code deletes the QP and puts the reference.
925  *
926  * - A reference held by the cm_id between the ESTABLISHED and
927  *   DISCONNECTED events. If the remote peer disconnected first, this
928  *   reference could be gone.
929  *
930  * - A reference held by the svc_recv code that called this function
931  *   as part of close processing.
932  *
933  * At a minimum two references should still be held.
934  */
935 static void svc_rdma_detach(struct svc_xprt *xprt)
936 {
937         struct svcxprt_rdma *rdma =
938                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
939         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
940
941         /* Disconnect and flush posted WQE */
942         rdma_disconnect(rdma->sc_cm_id);
943
944         /* Destroy the QP if present (not a listener) */
945         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
946                 ib_destroy_qp(rdma->sc_qp);
947                 svc_xprt_put(xprt);
948         }
949
950         /* Destroy the CM ID */
951         rdma_destroy_id(rdma->sc_cm_id);
952 }
953
954 static void svc_rdma_free(struct svc_xprt *xprt)
955 {
956         struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
957         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
958         /* We should only be called from kref_put */
959         BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
960         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
961                 ib_destroy_cq(rdma->sc_sq_cq);
962
963         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
964                 ib_destroy_cq(rdma->sc_rq_cq);
965
966         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
967                 ib_dereg_mr(rdma->sc_phys_mr);
968
969         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
970                 ib_dealloc_pd(rdma->sc_pd);
971
972         destroy_context_cache(rdma->sc_ctxt_head);
973         kfree(rdma);
974 }
975
976 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
977 {
978         struct svcxprt_rdma *rdma =
979                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
980
981         /*
982          * If there are fewer SQ WR available than required to send a
983          * simple response, return false.
984          */
985         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
986                 return 0;
987
988         /*
989          * ...or there are already waiters on the SQ,
990          * return false.
991          */
992         if (waitqueue_active(&rdma->sc_send_wait))
993                 return 0;
994
995         /* Otherwise return true. */
996         return 1;
997 }
998
999 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1000 {
1001         struct ib_send_wr *bad_wr;
1002         int ret;
1003
1004         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1005                 return -ENOTCONN;
1006
1007         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1008         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1009                 wr->opcode);
1010         /* If the SQ is full, wait until an SQ entry is available */
1011         while (1) {
1012                 spin_lock_bh(&xprt->sc_lock);
1013                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1014                         spin_unlock_bh(&xprt->sc_lock);
1015                         atomic_inc(&rdma_stat_sq_starve);
1016
1017                         /* See if we can opportunistically reap SQ WR to make room */
1018                         sq_cq_reap(xprt);
1019
1020                         /* Wait until SQ WR available if SQ still full */
1021                         wait_event(xprt->sc_send_wait,
1022                                    atomic_read(&xprt->sc_sq_count) <
1023                                    xprt->sc_sq_depth);
1024                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1025                                 return 0;
1026                         continue;
1027                 }
1028                 /* Bumped used SQ WR count and post */
1029                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1030                 if (!ret)
1031                         atomic_inc(&xprt->sc_sq_count);
1032                 else
1033                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1034                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1035                                ret, atomic_read(&xprt->sc_sq_count),
1036                                xprt->sc_sq_depth);
1037                 spin_unlock_bh(&xprt->sc_lock);
1038                 break;
1039         }
1040         return ret;
1041 }
1042
1043 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1044                         enum rpcrdma_errcode err)
1045 {
1046         struct ib_send_wr err_wr;
1047         struct ib_sge sge;
1048         struct page *p;
1049         struct svc_rdma_op_ctxt *ctxt;
1050         u32 *va;
1051         int length;
1052         int ret;
1053
1054         p = svc_rdma_get_page();
1055         va = page_address(p);
1056
1057         /* XDR encode error */
1058         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1059
1060         /* Prepare SGE for local address */
1061         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1062                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1063         sge.lkey = xprt->sc_phys_mr->lkey;
1064         sge.length = length;
1065
1066         ctxt = svc_rdma_get_context(xprt);
1067         ctxt->count = 1;
1068         ctxt->pages[0] = p;
1069
1070         /* Prepare SEND WR */
1071         memset(&err_wr, 0, sizeof err_wr);
1072         ctxt->wr_op = IB_WR_SEND;
1073         err_wr.wr_id = (unsigned long)ctxt;
1074         err_wr.sg_list = &sge;
1075         err_wr.num_sge = 1;
1076         err_wr.opcode = IB_WR_SEND;
1077         err_wr.send_flags = IB_SEND_SIGNALED;
1078
1079         /* Post It */
1080         ret = svc_rdma_send(xprt, &err_wr);
1081         if (ret) {
1082                 dprintk("svcrdma: Error posting send = %d\n", ret);
1083                 svc_rdma_put_context(ctxt, 1);
1084         }
1085
1086         return ret;
1087 }