2 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
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
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
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.
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
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.
39 * Author: Tom Tucker <tom@opengridcomputing.com>
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>
50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
52 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
53 struct sockaddr *sa, int salen,
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);
64 DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
65 static DEFINE_SPINLOCK(dto_lock);
66 static LIST_HEAD(dto_xprt_q);
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,
80 struct svc_xprt_class svc_rdma_class = {
82 .xcl_owner = THIS_MODULE,
83 .xcl_ops = &svc_rdma_ops,
84 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
87 /* WR context cache. Created in svc_rdma.c */
88 extern struct kmem_cache *svc_rdma_ctxt_cachep;
90 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
92 struct svc_rdma_op_ctxt *ctxt;
95 ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL);
98 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
101 INIT_LIST_HEAD(&ctxt->dto_q);
103 atomic_inc(&xprt->sc_ctxt_used);
107 static void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
109 struct svcxprt_rdma *xprt = ctxt->xprt;
111 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
112 atomic_dec(&xprt->sc_dma_used);
113 ib_dma_unmap_single(xprt->sc_cm_id->device,
120 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
122 struct svcxprt_rdma *xprt;
128 for (i = 0; i < ctxt->count; i++)
129 put_page(ctxt->pages[i]);
131 kmem_cache_free(svc_rdma_ctxt_cachep, ctxt);
132 atomic_dec(&xprt->sc_ctxt_used);
135 /* Temporary NFS request map cache. Created in svc_rdma.c */
136 extern struct kmem_cache *svc_rdma_map_cachep;
139 * Temporary NFS req mappings are shared across all transport
140 * instances. These are short lived and should be bounded by the number
141 * of concurrent server threads * depth of the SQ.
143 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
145 struct svc_rdma_req_map *map;
147 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
150 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
156 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
158 kmem_cache_free(svc_rdma_map_cachep, map);
161 /* ib_cq event handler */
162 static void cq_event_handler(struct ib_event *event, void *context)
164 struct svc_xprt *xprt = context;
165 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
166 event->event, context);
167 set_bit(XPT_CLOSE, &xprt->xpt_flags);
170 /* QP event handler */
171 static void qp_event_handler(struct ib_event *event, void *context)
173 struct svc_xprt *xprt = context;
175 switch (event->event) {
176 /* These are considered benign events */
177 case IB_EVENT_PATH_MIG:
178 case IB_EVENT_COMM_EST:
179 case IB_EVENT_SQ_DRAINED:
180 case IB_EVENT_QP_LAST_WQE_REACHED:
181 dprintk("svcrdma: QP event %d received for QP=%p\n",
182 event->event, event->element.qp);
184 /* These are considered fatal events */
185 case IB_EVENT_PATH_MIG_ERR:
186 case IB_EVENT_QP_FATAL:
187 case IB_EVENT_QP_REQ_ERR:
188 case IB_EVENT_QP_ACCESS_ERR:
189 case IB_EVENT_DEVICE_FATAL:
191 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
192 "closing transport\n",
193 event->event, event->element.qp);
194 set_bit(XPT_CLOSE, &xprt->xpt_flags);
200 * Data Transfer Operation Tasklet
202 * Walks a list of transports with I/O pending, removing entries as
203 * they are added to the server's I/O pending list. Two bits indicate
204 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
205 * spinlock that serializes access to the transport list with the RQ
206 * and SQ interrupt handlers.
208 static void dto_tasklet_func(unsigned long data)
210 struct svcxprt_rdma *xprt;
213 spin_lock_irqsave(&dto_lock, flags);
214 while (!list_empty(&dto_xprt_q)) {
215 xprt = list_entry(dto_xprt_q.next,
216 struct svcxprt_rdma, sc_dto_q);
217 list_del_init(&xprt->sc_dto_q);
218 spin_unlock_irqrestore(&dto_lock, flags);
223 svc_xprt_put(&xprt->sc_xprt);
224 spin_lock_irqsave(&dto_lock, flags);
226 spin_unlock_irqrestore(&dto_lock, flags);
230 * Receive Queue Completion Handler
232 * Since an RQ completion handler is called on interrupt context, we
233 * need to defer the handling of the I/O to a tasklet
235 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
237 struct svcxprt_rdma *xprt = cq_context;
240 /* Guard against unconditional flush call for destroyed QP */
241 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
245 * Set the bit regardless of whether or not it's on the list
246 * because it may be on the list already due to an SQ
249 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
252 * If this transport is not already on the DTO transport queue,
255 spin_lock_irqsave(&dto_lock, flags);
256 if (list_empty(&xprt->sc_dto_q)) {
257 svc_xprt_get(&xprt->sc_xprt);
258 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
260 spin_unlock_irqrestore(&dto_lock, flags);
262 /* Tasklet does all the work to avoid irqsave locks. */
263 tasklet_schedule(&dto_tasklet);
267 * rq_cq_reap - Process the RQ CQ.
269 * Take all completing WC off the CQE and enqueue the associated DTO
270 * context on the dto_q for the transport.
272 * Note that caller must hold a transport reference.
274 static void rq_cq_reap(struct svcxprt_rdma *xprt)
278 struct svc_rdma_op_ctxt *ctxt = NULL;
280 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
283 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
284 atomic_inc(&rdma_stat_rq_poll);
286 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
287 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
288 ctxt->wc_status = wc.status;
289 ctxt->byte_len = wc.byte_len;
290 svc_rdma_unmap_dma(ctxt);
291 if (wc.status != IB_WC_SUCCESS) {
292 /* Close the transport */
293 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
294 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
295 svc_rdma_put_context(ctxt, 1);
296 svc_xprt_put(&xprt->sc_xprt);
299 spin_lock_bh(&xprt->sc_rq_dto_lock);
300 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
301 spin_unlock_bh(&xprt->sc_rq_dto_lock);
302 svc_xprt_put(&xprt->sc_xprt);
306 atomic_inc(&rdma_stat_rq_prod);
308 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
310 * If data arrived before established event,
311 * don't enqueue. This defers RPC I/O until the
312 * RDMA connection is complete.
314 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
315 svc_xprt_enqueue(&xprt->sc_xprt);
319 * Send Queue Completion Handler - potentially called on interrupt context.
321 * Note that caller must hold a transport reference.
323 static void sq_cq_reap(struct svcxprt_rdma *xprt)
325 struct svc_rdma_op_ctxt *ctxt = NULL;
327 struct ib_cq *cq = xprt->sc_sq_cq;
331 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
334 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
335 atomic_inc(&rdma_stat_sq_poll);
336 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
337 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
340 svc_rdma_unmap_dma(ctxt);
341 if (wc.status != IB_WC_SUCCESS)
342 /* Close the transport */
343 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
345 /* Decrement used SQ WR count */
346 atomic_dec(&xprt->sc_sq_count);
347 wake_up(&xprt->sc_send_wait);
349 switch (ctxt->wr_op) {
351 svc_rdma_put_context(ctxt, 1);
354 case IB_WR_RDMA_WRITE:
355 svc_rdma_put_context(ctxt, 0);
358 case IB_WR_RDMA_READ:
359 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
360 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
362 spin_lock_bh(&xprt->sc_rq_dto_lock);
363 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
364 list_add_tail(&read_hdr->dto_q,
365 &xprt->sc_read_complete_q);
366 spin_unlock_bh(&xprt->sc_rq_dto_lock);
367 svc_xprt_enqueue(&xprt->sc_xprt);
369 svc_rdma_put_context(ctxt, 0);
373 printk(KERN_ERR "svcrdma: unexpected completion type, "
374 "opcode=%d, status=%d\n",
375 wc.opcode, wc.status);
378 svc_xprt_put(&xprt->sc_xprt);
382 atomic_inc(&rdma_stat_sq_prod);
385 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
387 struct svcxprt_rdma *xprt = cq_context;
390 /* Guard against unconditional flush call for destroyed QP */
391 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
395 * Set the bit regardless of whether or not it's on the list
396 * because it may be on the list already due to an RQ
399 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
402 * If this transport is not already on the DTO transport queue,
405 spin_lock_irqsave(&dto_lock, flags);
406 if (list_empty(&xprt->sc_dto_q)) {
407 svc_xprt_get(&xprt->sc_xprt);
408 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
410 spin_unlock_irqrestore(&dto_lock, flags);
412 /* Tasklet does all the work to avoid irqsave locks. */
413 tasklet_schedule(&dto_tasklet);
416 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
419 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
423 svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
424 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
425 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
426 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
427 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
428 init_waitqueue_head(&cma_xprt->sc_send_wait);
430 spin_lock_init(&cma_xprt->sc_lock);
431 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
433 cma_xprt->sc_ord = svcrdma_ord;
435 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
436 cma_xprt->sc_max_requests = svcrdma_max_requests;
437 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
438 atomic_set(&cma_xprt->sc_sq_count, 0);
439 atomic_set(&cma_xprt->sc_ctxt_used, 0);
442 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
447 struct page *svc_rdma_get_page(void)
451 while ((page = alloc_page(GFP_KERNEL)) == NULL) {
452 /* If we can't get memory, wait a bit and try again */
453 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
455 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
460 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
462 struct ib_recv_wr recv_wr, *bad_recv_wr;
463 struct svc_rdma_op_ctxt *ctxt;
470 ctxt = svc_rdma_get_context(xprt);
472 ctxt->direction = DMA_FROM_DEVICE;
473 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
474 BUG_ON(sge_no >= xprt->sc_max_sge);
475 page = svc_rdma_get_page();
476 ctxt->pages[sge_no] = page;
477 atomic_inc(&xprt->sc_dma_used);
478 pa = ib_dma_map_page(xprt->sc_cm_id->device,
481 ctxt->sge[sge_no].addr = pa;
482 ctxt->sge[sge_no].length = PAGE_SIZE;
483 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
486 ctxt->count = sge_no;
488 recv_wr.sg_list = &ctxt->sge[0];
489 recv_wr.num_sge = ctxt->count;
490 recv_wr.wr_id = (u64)(unsigned long)ctxt;
492 svc_xprt_get(&xprt->sc_xprt);
493 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
495 svc_xprt_put(&xprt->sc_xprt);
496 svc_rdma_put_context(ctxt, 1);
502 * This function handles the CONNECT_REQUEST event on a listening
503 * endpoint. It is passed the cma_id for the _new_ connection. The context in
504 * this cma_id is inherited from the listening cma_id and is the svc_xprt
505 * structure for the listening endpoint.
507 * This function creates a new xprt for the new connection and enqueues it on
508 * the accept queue for the listent xprt. When the listen thread is kicked, it
509 * will call the recvfrom method on the listen xprt which will accept the new
512 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird)
514 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
515 struct svcxprt_rdma *newxprt;
518 /* Create a new transport */
519 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
521 dprintk("svcrdma: failed to create new transport\n");
524 newxprt->sc_cm_id = new_cma_id;
525 new_cma_id->context = newxprt;
526 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
527 newxprt, newxprt->sc_cm_id, listen_xprt);
529 /* Save client advertised inbound read limit for use later in accept. */
530 newxprt->sc_ord = client_ird;
532 /* Set the local and remote addresses in the transport */
533 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
534 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
535 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
536 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
539 * Enqueue the new transport on the accept queue of the listening
542 spin_lock_bh(&listen_xprt->sc_lock);
543 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
544 spin_unlock_bh(&listen_xprt->sc_lock);
547 * Can't use svc_xprt_received here because we are not on a
550 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
551 svc_xprt_enqueue(&listen_xprt->sc_xprt);
555 * Handles events generated on the listening endpoint. These events will be
556 * either be incoming connect requests or adapter removal events.
558 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
559 struct rdma_cm_event *event)
561 struct svcxprt_rdma *xprt = cma_id->context;
564 switch (event->event) {
565 case RDMA_CM_EVENT_CONNECT_REQUEST:
566 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
567 "event=%d\n", cma_id, cma_id->context, event->event);
568 handle_connect_req(cma_id,
569 event->param.conn.responder_resources);
572 case RDMA_CM_EVENT_ESTABLISHED:
573 /* Accept complete */
574 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
575 "cm_id=%p\n", xprt, cma_id);
578 case RDMA_CM_EVENT_DEVICE_REMOVAL:
579 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
582 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
586 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
587 "event=%d\n", cma_id, event->event);
594 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
595 struct rdma_cm_event *event)
597 struct svc_xprt *xprt = cma_id->context;
598 struct svcxprt_rdma *rdma =
599 container_of(xprt, struct svcxprt_rdma, sc_xprt);
600 switch (event->event) {
601 case RDMA_CM_EVENT_ESTABLISHED:
602 /* Accept complete */
604 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
605 "cm_id=%p\n", xprt, cma_id);
606 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
607 svc_xprt_enqueue(xprt);
609 case RDMA_CM_EVENT_DISCONNECTED:
610 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
613 set_bit(XPT_CLOSE, &xprt->xpt_flags);
614 svc_xprt_enqueue(xprt);
618 case RDMA_CM_EVENT_DEVICE_REMOVAL:
619 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
620 "event=%d\n", cma_id, xprt, event->event);
622 set_bit(XPT_CLOSE, &xprt->xpt_flags);
623 svc_xprt_enqueue(xprt);
627 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
628 "event=%d\n", cma_id, event->event);
635 * Create a listening RDMA service endpoint.
637 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
638 struct sockaddr *sa, int salen,
641 struct rdma_cm_id *listen_id;
642 struct svcxprt_rdma *cma_xprt;
643 struct svc_xprt *xprt;
646 dprintk("svcrdma: Creating RDMA socket\n");
648 cma_xprt = rdma_create_xprt(serv, 1);
650 return ERR_PTR(-ENOMEM);
651 xprt = &cma_xprt->sc_xprt;
653 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
654 if (IS_ERR(listen_id)) {
655 ret = PTR_ERR(listen_id);
656 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
660 ret = rdma_bind_addr(listen_id, sa);
662 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
665 cma_xprt->sc_cm_id = listen_id;
667 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
669 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
674 * We need to use the address from the cm_id in case the
675 * caller specified 0 for the port number.
677 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
678 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
680 return &cma_xprt->sc_xprt;
683 rdma_destroy_id(listen_id);
690 * This is the xpo_recvfrom function for listening endpoints. Its
691 * purpose is to accept incoming connections. The CMA callback handler
692 * has already created a new transport and attached it to the new CMA
695 * There is a queue of pending connections hung on the listening
696 * transport. This queue contains the new svc_xprt structure. This
697 * function takes svc_xprt structures off the accept_q and completes
700 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
702 struct svcxprt_rdma *listen_rdma;
703 struct svcxprt_rdma *newxprt = NULL;
704 struct rdma_conn_param conn_param;
705 struct ib_qp_init_attr qp_attr;
706 struct ib_device_attr devattr;
710 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
711 clear_bit(XPT_CONN, &xprt->xpt_flags);
712 /* Get the next entry off the accept list */
713 spin_lock_bh(&listen_rdma->sc_lock);
714 if (!list_empty(&listen_rdma->sc_accept_q)) {
715 newxprt = list_entry(listen_rdma->sc_accept_q.next,
716 struct svcxprt_rdma, sc_accept_q);
717 list_del_init(&newxprt->sc_accept_q);
719 if (!list_empty(&listen_rdma->sc_accept_q))
720 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
721 spin_unlock_bh(&listen_rdma->sc_lock);
725 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
726 newxprt, newxprt->sc_cm_id);
728 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
730 dprintk("svcrdma: could not query device attributes on "
731 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
735 /* Qualify the transport resource defaults with the
736 * capabilities of this particular device */
737 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
738 (size_t)RPCSVC_MAXPAGES);
739 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
740 (size_t)svcrdma_max_requests);
741 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
744 * Limit ORD based on client limit, local device limit, and
745 * configured svcrdma limit.
747 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord);
748 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
750 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
751 if (IS_ERR(newxprt->sc_pd)) {
752 dprintk("svcrdma: error creating PD for connect request\n");
755 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
759 newxprt->sc_sq_depth,
761 if (IS_ERR(newxprt->sc_sq_cq)) {
762 dprintk("svcrdma: error creating SQ CQ for connect request\n");
765 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
769 newxprt->sc_max_requests,
771 if (IS_ERR(newxprt->sc_rq_cq)) {
772 dprintk("svcrdma: error creating RQ CQ for connect request\n");
776 memset(&qp_attr, 0, sizeof qp_attr);
777 qp_attr.event_handler = qp_event_handler;
778 qp_attr.qp_context = &newxprt->sc_xprt;
779 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
780 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
781 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
782 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
783 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
784 qp_attr.qp_type = IB_QPT_RC;
785 qp_attr.send_cq = newxprt->sc_sq_cq;
786 qp_attr.recv_cq = newxprt->sc_rq_cq;
787 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
788 " cm_id->device=%p, sc_pd->device=%p\n"
789 " cap.max_send_wr = %d\n"
790 " cap.max_recv_wr = %d\n"
791 " cap.max_send_sge = %d\n"
792 " cap.max_recv_sge = %d\n",
793 newxprt->sc_cm_id, newxprt->sc_pd,
794 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
795 qp_attr.cap.max_send_wr,
796 qp_attr.cap.max_recv_wr,
797 qp_attr.cap.max_send_sge,
798 qp_attr.cap.max_recv_sge);
800 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
803 * XXX: This is a hack. We need a xx_request_qp interface
804 * that will adjust the qp_attr's with a best-effort
807 qp_attr.cap.max_send_sge -= 2;
808 qp_attr.cap.max_recv_sge -= 2;
809 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
812 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
815 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
816 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
817 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
818 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
820 newxprt->sc_qp = newxprt->sc_cm_id->qp;
822 /* Register all of physical memory */
823 newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
824 IB_ACCESS_LOCAL_WRITE |
825 IB_ACCESS_REMOTE_WRITE);
826 if (IS_ERR(newxprt->sc_phys_mr)) {
827 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
831 /* Post receive buffers */
832 for (i = 0; i < newxprt->sc_max_requests; i++) {
833 ret = svc_rdma_post_recv(newxprt);
835 dprintk("svcrdma: failure posting receive buffers\n");
840 /* Swap out the handler */
841 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
844 * Arm the CQs for the SQ and RQ before accepting so we can't
845 * miss the first message
847 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
848 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
850 /* Accept Connection */
851 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
852 memset(&conn_param, 0, sizeof conn_param);
853 conn_param.responder_resources = 0;
854 conn_param.initiator_depth = newxprt->sc_ord;
855 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
857 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
862 dprintk("svcrdma: new connection %p accepted with the following "
864 " local_ip : %d.%d.%d.%d\n"
866 " remote_ip : %d.%d.%d.%d\n"
867 " remote_port : %d\n"
870 " max_requests : %d\n"
873 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
874 route.addr.src_addr)->sin_addr.s_addr),
875 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
876 route.addr.src_addr)->sin_port),
877 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
878 route.addr.dst_addr)->sin_addr.s_addr),
879 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
880 route.addr.dst_addr)->sin_port),
882 newxprt->sc_sq_depth,
883 newxprt->sc_max_requests,
886 return &newxprt->sc_xprt;
889 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
890 /* Take a reference in case the DTO handler runs */
891 svc_xprt_get(&newxprt->sc_xprt);
892 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
893 ib_destroy_qp(newxprt->sc_qp);
894 rdma_destroy_id(newxprt->sc_cm_id);
895 /* This call to put will destroy the transport */
896 svc_xprt_put(&newxprt->sc_xprt);
900 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
905 * When connected, an svc_xprt has at least two references:
907 * - A reference held by the cm_id between the ESTABLISHED and
908 * DISCONNECTED events. If the remote peer disconnected first, this
909 * reference could be gone.
911 * - A reference held by the svc_recv code that called this function
912 * as part of close processing.
914 * At a minimum one references should still be held.
916 static void svc_rdma_detach(struct svc_xprt *xprt)
918 struct svcxprt_rdma *rdma =
919 container_of(xprt, struct svcxprt_rdma, sc_xprt);
920 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
922 /* Disconnect and flush posted WQE */
923 rdma_disconnect(rdma->sc_cm_id);
926 static void __svc_rdma_free(struct work_struct *work)
928 struct svcxprt_rdma *rdma =
929 container_of(work, struct svcxprt_rdma, sc_work);
930 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
932 /* We should only be called from kref_put */
933 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
936 * Destroy queued, but not processed read completions. Note
937 * that this cleanup has to be done before destroying the
938 * cm_id because the device ptr is needed to unmap the dma in
939 * svc_rdma_put_context.
941 while (!list_empty(&rdma->sc_read_complete_q)) {
942 struct svc_rdma_op_ctxt *ctxt;
943 ctxt = list_entry(rdma->sc_read_complete_q.next,
944 struct svc_rdma_op_ctxt,
946 list_del_init(&ctxt->dto_q);
947 svc_rdma_put_context(ctxt, 1);
950 /* Destroy queued, but not processed recv completions */
951 while (!list_empty(&rdma->sc_rq_dto_q)) {
952 struct svc_rdma_op_ctxt *ctxt;
953 ctxt = list_entry(rdma->sc_rq_dto_q.next,
954 struct svc_rdma_op_ctxt,
956 list_del_init(&ctxt->dto_q);
957 svc_rdma_put_context(ctxt, 1);
960 /* Warn if we leaked a resource or under-referenced */
961 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
962 WARN_ON(atomic_read(&rdma->sc_dma_used) != 0);
964 /* Destroy the QP if present (not a listener) */
965 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
966 ib_destroy_qp(rdma->sc_qp);
968 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
969 ib_destroy_cq(rdma->sc_sq_cq);
971 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
972 ib_destroy_cq(rdma->sc_rq_cq);
974 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
975 ib_dereg_mr(rdma->sc_phys_mr);
977 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
978 ib_dealloc_pd(rdma->sc_pd);
980 /* Destroy the CM ID */
981 rdma_destroy_id(rdma->sc_cm_id);
986 static void svc_rdma_free(struct svc_xprt *xprt)
988 struct svcxprt_rdma *rdma =
989 container_of(xprt, struct svcxprt_rdma, sc_xprt);
990 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
991 schedule_work(&rdma->sc_work);
994 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
996 struct svcxprt_rdma *rdma =
997 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1000 * If there are fewer SQ WR available than required to send a
1001 * simple response, return false.
1003 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1007 * ...or there are already waiters on the SQ,
1010 if (waitqueue_active(&rdma->sc_send_wait))
1013 /* Otherwise return true. */
1017 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1019 struct ib_send_wr *bad_wr;
1022 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1025 BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1026 BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1028 /* If the SQ is full, wait until an SQ entry is available */
1030 spin_lock_bh(&xprt->sc_lock);
1031 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1032 spin_unlock_bh(&xprt->sc_lock);
1033 atomic_inc(&rdma_stat_sq_starve);
1035 /* See if we can opportunistically reap SQ WR to make room */
1038 /* Wait until SQ WR available if SQ still full */
1039 wait_event(xprt->sc_send_wait,
1040 atomic_read(&xprt->sc_sq_count) <
1042 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1046 /* Bumped used SQ WR count and post */
1047 svc_xprt_get(&xprt->sc_xprt);
1048 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1050 atomic_inc(&xprt->sc_sq_count);
1052 svc_xprt_put(&xprt->sc_xprt);
1053 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1054 "sc_sq_count=%d, sc_sq_depth=%d\n",
1055 ret, atomic_read(&xprt->sc_sq_count),
1058 spin_unlock_bh(&xprt->sc_lock);
1064 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1065 enum rpcrdma_errcode err)
1067 struct ib_send_wr err_wr;
1070 struct svc_rdma_op_ctxt *ctxt;
1075 p = svc_rdma_get_page();
1076 va = page_address(p);
1078 /* XDR encode error */
1079 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1081 /* Prepare SGE for local address */
1082 atomic_inc(&xprt->sc_dma_used);
1083 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1084 p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1085 sge.lkey = xprt->sc_phys_mr->lkey;
1086 sge.length = length;
1088 ctxt = svc_rdma_get_context(xprt);
1092 /* Prepare SEND WR */
1093 memset(&err_wr, 0, sizeof err_wr);
1094 ctxt->wr_op = IB_WR_SEND;
1095 err_wr.wr_id = (unsigned long)ctxt;
1096 err_wr.sg_list = &sge;
1098 err_wr.opcode = IB_WR_SEND;
1099 err_wr.send_flags = IB_SEND_SIGNALED;
1102 ret = svc_rdma_send(xprt, &err_wr);
1104 dprintk("svcrdma: Error %d posting send for protocol error\n",
1106 svc_rdma_put_context(ctxt, 1);