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 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
91 struct svc_rdma_op_ctxt *ctxt;
93 target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
96 spin_lock_bh(&xprt->sc_ctxt_lock);
97 while (xprt->sc_ctxt_cnt < target) {
99 spin_unlock_bh(&xprt->sc_ctxt_lock);
101 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
103 spin_lock_bh(&xprt->sc_ctxt_lock);
106 INIT_LIST_HEAD(&ctxt->free_list);
107 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
109 /* kmalloc failed...give up for now */
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);
120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
122 struct svc_rdma_op_ctxt *ctxt;
125 spin_lock_bh(&xprt->sc_ctxt_lock);
126 if (unlikely(list_empty(&xprt->sc_ctxt_free))) {
127 /* Try to bump my cache. */
128 spin_unlock_bh(&xprt->sc_ctxt_lock);
130 if (rdma_bump_context_cache(xprt))
133 printk(KERN_INFO "svcrdma: sleeping waiting for "
134 "context memory on xprt=%p\n",
136 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
139 ctxt = list_entry(xprt->sc_ctxt_free.next,
140 struct svc_rdma_op_ctxt,
142 list_del_init(&ctxt->free_list);
143 spin_unlock_bh(&xprt->sc_ctxt_lock);
145 INIT_LIST_HEAD(&ctxt->dto_q);
147 atomic_inc(&xprt->sc_ctxt_used);
153 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
155 struct svcxprt_rdma *xprt;
161 for (i = 0; i < ctxt->count; i++)
162 put_page(ctxt->pages[i]);
164 for (i = 0; i < ctxt->count; i++)
165 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
170 spin_lock_bh(&xprt->sc_ctxt_lock);
171 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
172 spin_unlock_bh(&xprt->sc_ctxt_lock);
173 atomic_dec(&xprt->sc_ctxt_used);
176 /* ib_cq event handler */
177 static void cq_event_handler(struct ib_event *event, void *context)
179 struct svc_xprt *xprt = context;
180 dprintk("svcrdma: received CQ event id=%d, context=%p\n",
181 event->event, context);
182 set_bit(XPT_CLOSE, &xprt->xpt_flags);
185 /* QP event handler */
186 static void qp_event_handler(struct ib_event *event, void *context)
188 struct svc_xprt *xprt = context;
190 switch (event->event) {
191 /* These are considered benign events */
192 case IB_EVENT_PATH_MIG:
193 case IB_EVENT_COMM_EST:
194 case IB_EVENT_SQ_DRAINED:
195 case IB_EVENT_QP_LAST_WQE_REACHED:
196 dprintk("svcrdma: QP event %d received for QP=%p\n",
197 event->event, event->element.qp);
199 /* These are considered fatal events */
200 case IB_EVENT_PATH_MIG_ERR:
201 case IB_EVENT_QP_FATAL:
202 case IB_EVENT_QP_REQ_ERR:
203 case IB_EVENT_QP_ACCESS_ERR:
204 case IB_EVENT_DEVICE_FATAL:
206 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
207 "closing transport\n",
208 event->event, event->element.qp);
209 set_bit(XPT_CLOSE, &xprt->xpt_flags);
215 * Data Transfer Operation Tasklet
217 * Walks a list of transports with I/O pending, removing entries as
218 * they are added to the server's I/O pending list. Two bits indicate
219 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
220 * spinlock that serializes access to the transport list with the RQ
221 * and SQ interrupt handlers.
223 static void dto_tasklet_func(unsigned long data)
225 struct svcxprt_rdma *xprt;
228 spin_lock_irqsave(&dto_lock, flags);
229 while (!list_empty(&dto_xprt_q)) {
230 xprt = list_entry(dto_xprt_q.next,
231 struct svcxprt_rdma, sc_dto_q);
232 list_del_init(&xprt->sc_dto_q);
233 spin_unlock_irqrestore(&dto_lock, flags);
238 svc_xprt_put(&xprt->sc_xprt);
239 spin_lock_irqsave(&dto_lock, flags);
241 spin_unlock_irqrestore(&dto_lock, flags);
245 * Receive Queue Completion Handler
247 * Since an RQ completion handler is called on interrupt context, we
248 * need to defer the handling of the I/O to a tasklet
250 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
252 struct svcxprt_rdma *xprt = cq_context;
256 * Set the bit regardless of whether or not it's on the list
257 * because it may be on the list already due to an SQ
260 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
263 * If this transport is not already on the DTO transport queue,
266 spin_lock_irqsave(&dto_lock, flags);
267 if (list_empty(&xprt->sc_dto_q)) {
268 svc_xprt_get(&xprt->sc_xprt);
269 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
271 spin_unlock_irqrestore(&dto_lock, flags);
273 /* Tasklet does all the work to avoid irqsave locks. */
274 tasklet_schedule(&dto_tasklet);
278 * rq_cq_reap - Process the RQ CQ.
280 * Take all completing WC off the CQE and enqueue the associated DTO
281 * context on the dto_q for the transport.
283 * Note that caller must hold a transport reference.
285 static void rq_cq_reap(struct svcxprt_rdma *xprt)
289 struct svc_rdma_op_ctxt *ctxt = NULL;
291 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
294 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
295 atomic_inc(&rdma_stat_rq_poll);
297 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
298 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
299 ctxt->wc_status = wc.status;
300 ctxt->byte_len = wc.byte_len;
301 if (wc.status != IB_WC_SUCCESS) {
302 /* Close the transport */
303 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
304 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
305 svc_rdma_put_context(ctxt, 1);
306 svc_xprt_put(&xprt->sc_xprt);
309 spin_lock_bh(&xprt->sc_rq_dto_lock);
310 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
311 spin_unlock_bh(&xprt->sc_rq_dto_lock);
312 svc_xprt_put(&xprt->sc_xprt);
316 atomic_inc(&rdma_stat_rq_prod);
318 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
320 * If data arrived before established event,
321 * don't enqueue. This defers RPC I/O until the
322 * RDMA connection is complete.
324 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
325 svc_xprt_enqueue(&xprt->sc_xprt);
329 * Send Queue Completion Handler - potentially called on interrupt context.
331 * Note that caller must hold a transport reference.
333 static void sq_cq_reap(struct svcxprt_rdma *xprt)
335 struct svc_rdma_op_ctxt *ctxt = NULL;
337 struct ib_cq *cq = xprt->sc_sq_cq;
341 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
344 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
345 atomic_inc(&rdma_stat_sq_poll);
346 while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
347 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
350 if (wc.status != IB_WC_SUCCESS)
351 /* Close the transport */
352 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
354 /* Decrement used SQ WR count */
355 atomic_dec(&xprt->sc_sq_count);
356 wake_up(&xprt->sc_send_wait);
358 switch (ctxt->wr_op) {
360 case IB_WR_RDMA_WRITE:
361 svc_rdma_put_context(ctxt, 1);
364 case IB_WR_RDMA_READ:
365 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
366 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
368 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
369 spin_lock_bh(&xprt->sc_read_complete_lock);
370 list_add_tail(&read_hdr->dto_q,
371 &xprt->sc_read_complete_q);
372 spin_unlock_bh(&xprt->sc_read_complete_lock);
373 svc_xprt_enqueue(&xprt->sc_xprt);
375 svc_rdma_put_context(ctxt, 0);
379 printk(KERN_ERR "svcrdma: unexpected completion type, "
380 "opcode=%d, status=%d\n",
381 wc.opcode, wc.status);
384 svc_xprt_put(&xprt->sc_xprt);
388 atomic_inc(&rdma_stat_sq_prod);
391 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
393 struct svcxprt_rdma *xprt = cq_context;
397 * Set the bit regardless of whether or not it's on the list
398 * because it may be on the list already due to an RQ
401 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
404 * If this transport is not already on the DTO transport queue,
407 spin_lock_irqsave(&dto_lock, flags);
408 if (list_empty(&xprt->sc_dto_q)) {
409 svc_xprt_get(&xprt->sc_xprt);
410 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
412 spin_unlock_irqrestore(&dto_lock, flags);
414 /* Tasklet does all the work to avoid irqsave locks. */
415 tasklet_schedule(&dto_tasklet);
418 static void create_context_cache(struct svcxprt_rdma *xprt,
419 int ctxt_count, int ctxt_bump, int ctxt_max)
421 struct svc_rdma_op_ctxt *ctxt;
424 xprt->sc_ctxt_max = ctxt_max;
425 xprt->sc_ctxt_bump = ctxt_bump;
426 xprt->sc_ctxt_cnt = 0;
427 atomic_set(&xprt->sc_ctxt_used, 0);
429 INIT_LIST_HEAD(&xprt->sc_ctxt_free);
430 for (i = 0; i < ctxt_count; i++) {
431 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
433 INIT_LIST_HEAD(&ctxt->free_list);
434 list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
440 static void destroy_context_cache(struct svcxprt_rdma *xprt)
442 while (!list_empty(&xprt->sc_ctxt_free)) {
443 struct svc_rdma_op_ctxt *ctxt;
444 ctxt = list_entry(xprt->sc_ctxt_free.next,
445 struct svc_rdma_op_ctxt,
447 list_del_init(&ctxt->free_list);
452 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
455 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
459 svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
460 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
461 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
462 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
463 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
464 init_waitqueue_head(&cma_xprt->sc_send_wait);
466 spin_lock_init(&cma_xprt->sc_lock);
467 spin_lock_init(&cma_xprt->sc_read_complete_lock);
468 spin_lock_init(&cma_xprt->sc_ctxt_lock);
469 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
471 cma_xprt->sc_ord = svcrdma_ord;
473 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
474 cma_xprt->sc_max_requests = svcrdma_max_requests;
475 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
476 atomic_set(&cma_xprt->sc_sq_count, 0);
479 int reqs = cma_xprt->sc_max_requests;
480 create_context_cache(cma_xprt,
481 reqs << 1, /* starting size */
482 reqs, /* bump amount */
484 cma_xprt->sc_sq_depth +
485 RPCRDMA_MAX_THREADS + 1); /* max */
486 if (list_empty(&cma_xprt->sc_ctxt_free)) {
490 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
492 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
497 struct page *svc_rdma_get_page(void)
501 while ((page = alloc_page(GFP_KERNEL)) == NULL) {
502 /* If we can't get memory, wait a bit and try again */
503 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
505 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
510 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
512 struct ib_recv_wr recv_wr, *bad_recv_wr;
513 struct svc_rdma_op_ctxt *ctxt;
520 ctxt = svc_rdma_get_context(xprt);
522 ctxt->direction = DMA_FROM_DEVICE;
523 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
524 BUG_ON(sge_no >= xprt->sc_max_sge);
525 page = svc_rdma_get_page();
526 ctxt->pages[sge_no] = page;
527 pa = ib_dma_map_page(xprt->sc_cm_id->device,
530 ctxt->sge[sge_no].addr = pa;
531 ctxt->sge[sge_no].length = PAGE_SIZE;
532 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
535 ctxt->count = sge_no;
537 recv_wr.sg_list = &ctxt->sge[0];
538 recv_wr.num_sge = ctxt->count;
539 recv_wr.wr_id = (u64)(unsigned long)ctxt;
541 svc_xprt_get(&xprt->sc_xprt);
542 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
544 svc_xprt_put(&xprt->sc_xprt);
545 svc_rdma_put_context(ctxt, 1);
551 * This function handles the CONNECT_REQUEST event on a listening
552 * endpoint. It is passed the cma_id for the _new_ connection. The context in
553 * this cma_id is inherited from the listening cma_id and is the svc_xprt
554 * structure for the listening endpoint.
556 * This function creates a new xprt for the new connection and enqueues it on
557 * the accept queue for the listent xprt. When the listen thread is kicked, it
558 * will call the recvfrom method on the listen xprt which will accept the new
561 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
563 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
564 struct svcxprt_rdma *newxprt;
566 /* Create a new transport */
567 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
569 dprintk("svcrdma: failed to create new transport\n");
572 newxprt->sc_cm_id = new_cma_id;
573 new_cma_id->context = newxprt;
574 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
575 newxprt, newxprt->sc_cm_id, listen_xprt);
578 * Enqueue the new transport on the accept queue of the listening
581 spin_lock_bh(&listen_xprt->sc_lock);
582 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
583 spin_unlock_bh(&listen_xprt->sc_lock);
586 * Can't use svc_xprt_received here because we are not on a
589 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
590 svc_xprt_enqueue(&listen_xprt->sc_xprt);
594 * Handles events generated on the listening endpoint. These events will be
595 * either be incoming connect requests or adapter removal events.
597 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
598 struct rdma_cm_event *event)
600 struct svcxprt_rdma *xprt = cma_id->context;
603 switch (event->event) {
604 case RDMA_CM_EVENT_CONNECT_REQUEST:
605 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
606 "event=%d\n", cma_id, cma_id->context, event->event);
607 handle_connect_req(cma_id);
610 case RDMA_CM_EVENT_ESTABLISHED:
611 /* Accept complete */
612 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
613 "cm_id=%p\n", xprt, cma_id);
616 case RDMA_CM_EVENT_DEVICE_REMOVAL:
617 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
620 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
624 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
625 "event=%d\n", cma_id, event->event);
632 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
633 struct rdma_cm_event *event)
635 struct svc_xprt *xprt = cma_id->context;
636 struct svcxprt_rdma *rdma =
637 container_of(xprt, struct svcxprt_rdma, sc_xprt);
638 switch (event->event) {
639 case RDMA_CM_EVENT_ESTABLISHED:
640 /* Accept complete */
642 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
643 "cm_id=%p\n", xprt, cma_id);
644 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
645 svc_xprt_enqueue(xprt);
647 case RDMA_CM_EVENT_DISCONNECTED:
648 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
651 set_bit(XPT_CLOSE, &xprt->xpt_flags);
652 svc_xprt_enqueue(xprt);
656 case RDMA_CM_EVENT_DEVICE_REMOVAL:
657 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
658 "event=%d\n", cma_id, xprt, event->event);
660 set_bit(XPT_CLOSE, &xprt->xpt_flags);
661 svc_xprt_enqueue(xprt);
665 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
666 "event=%d\n", cma_id, event->event);
673 * Create a listening RDMA service endpoint.
675 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
676 struct sockaddr *sa, int salen,
679 struct rdma_cm_id *listen_id;
680 struct svcxprt_rdma *cma_xprt;
681 struct svc_xprt *xprt;
684 dprintk("svcrdma: Creating RDMA socket\n");
686 cma_xprt = rdma_create_xprt(serv, 1);
688 return ERR_PTR(-ENOMEM);
689 xprt = &cma_xprt->sc_xprt;
691 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
692 if (IS_ERR(listen_id)) {
693 ret = PTR_ERR(listen_id);
694 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
698 ret = rdma_bind_addr(listen_id, sa);
700 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
703 cma_xprt->sc_cm_id = listen_id;
705 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
707 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
712 * We need to use the address from the cm_id in case the
713 * caller specified 0 for the port number.
715 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
716 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
718 return &cma_xprt->sc_xprt;
721 rdma_destroy_id(listen_id);
728 * This is the xpo_recvfrom function for listening endpoints. Its
729 * purpose is to accept incoming connections. The CMA callback handler
730 * has already created a new transport and attached it to the new CMA
733 * There is a queue of pending connections hung on the listening
734 * transport. This queue contains the new svc_xprt structure. This
735 * function takes svc_xprt structures off the accept_q and completes
738 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
740 struct svcxprt_rdma *listen_rdma;
741 struct svcxprt_rdma *newxprt = NULL;
742 struct rdma_conn_param conn_param;
743 struct ib_qp_init_attr qp_attr;
744 struct ib_device_attr devattr;
749 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
750 clear_bit(XPT_CONN, &xprt->xpt_flags);
751 /* Get the next entry off the accept list */
752 spin_lock_bh(&listen_rdma->sc_lock);
753 if (!list_empty(&listen_rdma->sc_accept_q)) {
754 newxprt = list_entry(listen_rdma->sc_accept_q.next,
755 struct svcxprt_rdma, sc_accept_q);
756 list_del_init(&newxprt->sc_accept_q);
758 if (!list_empty(&listen_rdma->sc_accept_q))
759 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
760 spin_unlock_bh(&listen_rdma->sc_lock);
764 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
765 newxprt, newxprt->sc_cm_id);
767 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
769 dprintk("svcrdma: could not query device attributes on "
770 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
774 /* Qualify the transport resource defaults with the
775 * capabilities of this particular device */
776 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
777 (size_t)RPCSVC_MAXPAGES);
778 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
779 (size_t)svcrdma_max_requests);
780 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
782 newxprt->sc_ord = min((size_t)devattr.max_qp_rd_atom,
783 (size_t)svcrdma_ord);
785 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
786 if (IS_ERR(newxprt->sc_pd)) {
787 dprintk("svcrdma: error creating PD for connect request\n");
790 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
794 newxprt->sc_sq_depth,
796 if (IS_ERR(newxprt->sc_sq_cq)) {
797 dprintk("svcrdma: error creating SQ CQ for connect request\n");
800 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
804 newxprt->sc_max_requests,
806 if (IS_ERR(newxprt->sc_rq_cq)) {
807 dprintk("svcrdma: error creating RQ CQ for connect request\n");
811 memset(&qp_attr, 0, sizeof qp_attr);
812 qp_attr.event_handler = qp_event_handler;
813 qp_attr.qp_context = &newxprt->sc_xprt;
814 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
815 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
816 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
817 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
818 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
819 qp_attr.qp_type = IB_QPT_RC;
820 qp_attr.send_cq = newxprt->sc_sq_cq;
821 qp_attr.recv_cq = newxprt->sc_rq_cq;
822 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
823 " cm_id->device=%p, sc_pd->device=%p\n"
824 " cap.max_send_wr = %d\n"
825 " cap.max_recv_wr = %d\n"
826 " cap.max_send_sge = %d\n"
827 " cap.max_recv_sge = %d\n",
828 newxprt->sc_cm_id, newxprt->sc_pd,
829 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
830 qp_attr.cap.max_send_wr,
831 qp_attr.cap.max_recv_wr,
832 qp_attr.cap.max_send_sge,
833 qp_attr.cap.max_recv_sge);
835 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
838 * XXX: This is a hack. We need a xx_request_qp interface
839 * that will adjust the qp_attr's with a best-effort
842 qp_attr.cap.max_send_sge -= 2;
843 qp_attr.cap.max_recv_sge -= 2;
844 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
847 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
850 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
851 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
852 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
853 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
855 svc_xprt_get(&newxprt->sc_xprt);
856 newxprt->sc_qp = newxprt->sc_cm_id->qp;
858 /* Register all of physical memory */
859 newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
860 IB_ACCESS_LOCAL_WRITE |
861 IB_ACCESS_REMOTE_WRITE);
862 if (IS_ERR(newxprt->sc_phys_mr)) {
863 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
867 /* Post receive buffers */
868 for (i = 0; i < newxprt->sc_max_requests; i++) {
869 ret = svc_rdma_post_recv(newxprt);
871 dprintk("svcrdma: failure posting receive buffers\n");
876 /* Swap out the handler */
877 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
879 /* Accept Connection */
880 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
881 memset(&conn_param, 0, sizeof conn_param);
882 conn_param.responder_resources = 0;
883 conn_param.initiator_depth = newxprt->sc_ord;
884 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
886 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
891 dprintk("svcrdma: new connection %p accepted with the following "
893 " local_ip : %d.%d.%d.%d\n"
895 " remote_ip : %d.%d.%d.%d\n"
896 " remote_port : %d\n"
899 " max_requests : %d\n"
902 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
903 route.addr.src_addr)->sin_addr.s_addr),
904 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
905 route.addr.src_addr)->sin_port),
906 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
907 route.addr.dst_addr)->sin_addr.s_addr),
908 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
909 route.addr.dst_addr)->sin_port),
911 newxprt->sc_sq_depth,
912 newxprt->sc_max_requests,
915 /* Set the local and remote addresses in the transport */
916 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
917 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
918 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
919 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
921 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
922 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
923 return &newxprt->sc_xprt;
926 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
927 /* Take a reference in case the DTO handler runs */
928 svc_xprt_get(&newxprt->sc_xprt);
929 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
930 ib_destroy_qp(newxprt->sc_qp);
931 svc_xprt_put(&newxprt->sc_xprt);
933 rdma_destroy_id(newxprt->sc_cm_id);
934 /* This call to put will destroy the transport */
935 svc_xprt_put(&newxprt->sc_xprt);
939 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
944 * When connected, an svc_xprt has at least three references:
946 * - A reference held by the QP. We still hold that here because this
947 * code deletes the QP and puts the reference.
949 * - A reference held by the cm_id between the ESTABLISHED and
950 * DISCONNECTED events. If the remote peer disconnected first, this
951 * reference could be gone.
953 * - A reference held by the svc_recv code that called this function
954 * as part of close processing.
956 * At a minimum two references should still be held.
958 static void svc_rdma_detach(struct svc_xprt *xprt)
960 struct svcxprt_rdma *rdma =
961 container_of(xprt, struct svcxprt_rdma, sc_xprt);
962 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
964 /* Disconnect and flush posted WQE */
965 rdma_disconnect(rdma->sc_cm_id);
967 /* Destroy the QP if present (not a listener) */
968 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
969 ib_destroy_qp(rdma->sc_qp);
973 /* Destroy the CM ID */
974 rdma_destroy_id(rdma->sc_cm_id);
977 static void __svc_rdma_free(struct work_struct *work)
979 struct svcxprt_rdma *rdma =
980 container_of(work, struct svcxprt_rdma, sc_work);
981 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
983 /* We should only be called from kref_put */
984 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
986 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
987 ib_destroy_cq(rdma->sc_sq_cq);
989 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
990 ib_destroy_cq(rdma->sc_rq_cq);
992 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
993 ib_dereg_mr(rdma->sc_phys_mr);
995 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
996 ib_dealloc_pd(rdma->sc_pd);
998 destroy_context_cache(rdma);
1002 static void svc_rdma_free(struct svc_xprt *xprt)
1004 struct svcxprt_rdma *rdma =
1005 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1006 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1007 schedule_work(&rdma->sc_work);
1010 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1012 struct svcxprt_rdma *rdma =
1013 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1016 * If there are fewer SQ WR available than required to send a
1017 * simple response, return false.
1019 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1023 * ...or there are already waiters on the SQ,
1026 if (waitqueue_active(&rdma->sc_send_wait))
1029 /* Otherwise return true. */
1033 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1035 struct ib_send_wr *bad_wr;
1038 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1041 BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1042 BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1044 /* If the SQ is full, wait until an SQ entry is available */
1046 spin_lock_bh(&xprt->sc_lock);
1047 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1048 spin_unlock_bh(&xprt->sc_lock);
1049 atomic_inc(&rdma_stat_sq_starve);
1051 /* See if we can opportunistically reap SQ WR to make room */
1054 /* Wait until SQ WR available if SQ still full */
1055 wait_event(xprt->sc_send_wait,
1056 atomic_read(&xprt->sc_sq_count) <
1058 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1062 /* Bumped used SQ WR count and post */
1063 svc_xprt_get(&xprt->sc_xprt);
1064 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1066 atomic_inc(&xprt->sc_sq_count);
1068 svc_xprt_put(&xprt->sc_xprt);
1069 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1070 "sc_sq_count=%d, sc_sq_depth=%d\n",
1071 ret, atomic_read(&xprt->sc_sq_count),
1074 spin_unlock_bh(&xprt->sc_lock);
1080 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1081 enum rpcrdma_errcode err)
1083 struct ib_send_wr err_wr;
1086 struct svc_rdma_op_ctxt *ctxt;
1091 p = svc_rdma_get_page();
1092 va = page_address(p);
1094 /* XDR encode error */
1095 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1097 /* Prepare SGE for local address */
1098 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1099 p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1100 sge.lkey = xprt->sc_phys_mr->lkey;
1101 sge.length = length;
1103 ctxt = svc_rdma_get_context(xprt);
1107 /* Prepare SEND WR */
1108 memset(&err_wr, 0, sizeof err_wr);
1109 ctxt->wr_op = IB_WR_SEND;
1110 err_wr.wr_id = (unsigned long)ctxt;
1111 err_wr.sg_list = &sge;
1113 err_wr.opcode = IB_WR_SEND;
1114 err_wr.send_flags = IB_SEND_SIGNALED;
1117 ret = svc_rdma_send(xprt, &err_wr);
1119 dprintk("svcrdma: Error posting send = %d\n", ret);
1120 svc_rdma_put_context(ctxt, 1);