2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved.
3 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the BSD-type
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
15 * Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
18 * Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials provided
21 * with the distribution.
23 * Neither the name of the Network Appliance, Inc. nor the names of
24 * its contributors may be used to endorse or promote products
25 * derived from this software without specific prior written
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 * Author: Tom Tucker <tom@opengridcomputing.com>
43 #include <linux/sunrpc/svc_xprt.h>
44 #include <linux/sunrpc/debug.h>
45 #include <linux/sunrpc/rpc_rdma.h>
46 #include <linux/interrupt.h>
47 #include <linux/sched.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/workqueue.h>
51 #include <rdma/ib_verbs.h>
52 #include <rdma/rdma_cm.h>
53 #include <linux/sunrpc/svc_rdma.h>
54 #include <linux/export.h>
55 #include "xprt_rdma.h"
57 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
59 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
61 struct sockaddr *sa, int salen,
63 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
64 static void svc_rdma_release_rqst(struct svc_rqst *);
65 static void dto_tasklet_func(unsigned long data);
66 static void svc_rdma_detach(struct svc_xprt *xprt);
67 static void svc_rdma_free(struct svc_xprt *xprt);
68 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
69 static int svc_rdma_secure_port(struct svc_rqst *);
70 static void rq_cq_reap(struct svcxprt_rdma *xprt);
71 static void sq_cq_reap(struct svcxprt_rdma *xprt);
73 static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
74 static DEFINE_SPINLOCK(dto_lock);
75 static LIST_HEAD(dto_xprt_q);
77 static struct svc_xprt_ops svc_rdma_ops = {
78 .xpo_create = svc_rdma_create,
79 .xpo_recvfrom = svc_rdma_recvfrom,
80 .xpo_sendto = svc_rdma_sendto,
81 .xpo_release_rqst = svc_rdma_release_rqst,
82 .xpo_detach = svc_rdma_detach,
83 .xpo_free = svc_rdma_free,
84 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
85 .xpo_has_wspace = svc_rdma_has_wspace,
86 .xpo_accept = svc_rdma_accept,
87 .xpo_secure_port = svc_rdma_secure_port,
90 struct svc_xprt_class svc_rdma_class = {
92 .xcl_owner = THIS_MODULE,
93 .xcl_ops = &svc_rdma_ops,
94 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA,
95 .xcl_ident = XPRT_TRANSPORT_RDMA,
98 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
100 struct svc_rdma_op_ctxt *ctxt;
102 ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep,
103 GFP_KERNEL | __GFP_NOFAIL);
105 INIT_LIST_HEAD(&ctxt->dto_q);
108 atomic_inc(&xprt->sc_ctxt_used);
112 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
114 struct svcxprt_rdma *xprt = ctxt->xprt;
116 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
118 * Unmap the DMA addr in the SGE if the lkey matches
119 * the sc_dma_lkey, otherwise, ignore it since it is
120 * an FRMR lkey and will be unmapped later when the
121 * last WR that uses it completes.
123 if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) {
124 atomic_dec(&xprt->sc_dma_used);
125 ib_dma_unmap_page(xprt->sc_cm_id->device,
133 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
135 struct svcxprt_rdma *xprt;
140 for (i = 0; i < ctxt->count; i++)
141 put_page(ctxt->pages[i]);
143 kmem_cache_free(svc_rdma_ctxt_cachep, ctxt);
144 atomic_dec(&xprt->sc_ctxt_used);
148 * Temporary NFS req mappings are shared across all transport
149 * instances. These are short lived and should be bounded by the number
150 * of concurrent server threads * depth of the SQ.
152 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
154 struct svc_rdma_req_map *map;
155 map = kmem_cache_alloc(svc_rdma_map_cachep,
156 GFP_KERNEL | __GFP_NOFAIL);
161 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
163 kmem_cache_free(svc_rdma_map_cachep, map);
166 /* ib_cq event handler */
167 static void cq_event_handler(struct ib_event *event, void *context)
169 struct svc_xprt *xprt = context;
170 dprintk("svcrdma: received CQ event %s (%d), context=%p\n",
171 ib_event_msg(event->event), event->event, context);
172 set_bit(XPT_CLOSE, &xprt->xpt_flags);
175 /* QP event handler */
176 static void qp_event_handler(struct ib_event *event, void *context)
178 struct svc_xprt *xprt = context;
180 switch (event->event) {
181 /* These are considered benign events */
182 case IB_EVENT_PATH_MIG:
183 case IB_EVENT_COMM_EST:
184 case IB_EVENT_SQ_DRAINED:
185 case IB_EVENT_QP_LAST_WQE_REACHED:
186 dprintk("svcrdma: QP event %s (%d) received for QP=%p\n",
187 ib_event_msg(event->event), event->event,
190 /* These are considered fatal events */
191 case IB_EVENT_PATH_MIG_ERR:
192 case IB_EVENT_QP_FATAL:
193 case IB_EVENT_QP_REQ_ERR:
194 case IB_EVENT_QP_ACCESS_ERR:
195 case IB_EVENT_DEVICE_FATAL:
197 dprintk("svcrdma: QP ERROR event %s (%d) received for QP=%p, "
198 "closing transport\n",
199 ib_event_msg(event->event), event->event,
201 set_bit(XPT_CLOSE, &xprt->xpt_flags);
207 * Data Transfer Operation Tasklet
209 * Walks a list of transports with I/O pending, removing entries as
210 * they are added to the server's I/O pending list. Two bits indicate
211 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
212 * spinlock that serializes access to the transport list with the RQ
213 * and SQ interrupt handlers.
215 static void dto_tasklet_func(unsigned long data)
217 struct svcxprt_rdma *xprt;
220 spin_lock_irqsave(&dto_lock, flags);
221 while (!list_empty(&dto_xprt_q)) {
222 xprt = list_entry(dto_xprt_q.next,
223 struct svcxprt_rdma, sc_dto_q);
224 list_del_init(&xprt->sc_dto_q);
225 spin_unlock_irqrestore(&dto_lock, flags);
230 svc_xprt_put(&xprt->sc_xprt);
231 spin_lock_irqsave(&dto_lock, flags);
233 spin_unlock_irqrestore(&dto_lock, flags);
237 * Receive Queue Completion Handler
239 * Since an RQ completion handler is called on interrupt context, we
240 * need to defer the handling of the I/O to a tasklet
242 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
244 struct svcxprt_rdma *xprt = cq_context;
247 /* Guard against unconditional flush call for destroyed QP */
248 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
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
256 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
259 * If this transport is not already on the DTO transport queue,
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);
267 spin_unlock_irqrestore(&dto_lock, flags);
269 /* Tasklet does all the work to avoid irqsave locks. */
270 tasklet_schedule(&dto_tasklet);
274 * rq_cq_reap - Process the RQ CQ.
276 * Take all completing WC off the CQE and enqueue the associated DTO
277 * context on the dto_q for the transport.
279 * Note that caller must hold a transport reference.
281 static void rq_cq_reap(struct svcxprt_rdma *xprt)
285 struct svc_rdma_op_ctxt *ctxt = NULL;
287 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
290 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
291 atomic_inc(&rdma_stat_rq_poll);
293 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
294 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
295 ctxt->wc_status = wc.status;
296 ctxt->byte_len = wc.byte_len;
297 svc_rdma_unmap_dma(ctxt);
298 if (wc.status != IB_WC_SUCCESS) {
299 /* Close the transport */
300 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
301 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
302 svc_rdma_put_context(ctxt, 1);
303 svc_xprt_put(&xprt->sc_xprt);
306 spin_lock_bh(&xprt->sc_rq_dto_lock);
307 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
308 spin_unlock_bh(&xprt->sc_rq_dto_lock);
309 svc_xprt_put(&xprt->sc_xprt);
313 atomic_inc(&rdma_stat_rq_prod);
315 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
317 * If data arrived before established event,
318 * don't enqueue. This defers RPC I/O until the
319 * RDMA connection is complete.
321 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
322 svc_xprt_enqueue(&xprt->sc_xprt);
326 * Process a completion context
328 static void process_context(struct svcxprt_rdma *xprt,
329 struct svc_rdma_op_ctxt *ctxt)
331 svc_rdma_unmap_dma(ctxt);
333 switch (ctxt->wr_op) {
336 pr_err("svcrdma: SEND: ctxt->frmr != NULL\n");
337 svc_rdma_put_context(ctxt, 1);
340 case IB_WR_RDMA_WRITE:
342 pr_err("svcrdma: WRITE: ctxt->frmr != NULL\n");
343 svc_rdma_put_context(ctxt, 0);
346 case IB_WR_RDMA_READ:
347 case IB_WR_RDMA_READ_WITH_INV:
348 svc_rdma_put_frmr(xprt, ctxt->frmr);
349 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
350 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
352 spin_lock_bh(&xprt->sc_rq_dto_lock);
353 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
354 list_add_tail(&read_hdr->dto_q,
355 &xprt->sc_read_complete_q);
356 spin_unlock_bh(&xprt->sc_rq_dto_lock);
358 pr_err("svcrdma: ctxt->read_hdr == NULL\n");
360 svc_xprt_enqueue(&xprt->sc_xprt);
362 svc_rdma_put_context(ctxt, 0);
366 printk(KERN_ERR "svcrdma: unexpected completion type, "
374 * Send Queue Completion Handler - potentially called on interrupt context.
376 * Note that caller must hold a transport reference.
378 static void sq_cq_reap(struct svcxprt_rdma *xprt)
380 struct svc_rdma_op_ctxt *ctxt = NULL;
381 struct ib_wc wc_a[6];
383 struct ib_cq *cq = xprt->sc_sq_cq;
386 memset(wc_a, 0, sizeof(wc_a));
388 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
391 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
392 atomic_inc(&rdma_stat_sq_poll);
393 while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) {
396 for (i = 0; i < ret; i++) {
398 if (wc->status != IB_WC_SUCCESS) {
399 dprintk("svcrdma: sq wc err status %s (%d)\n",
400 ib_wc_status_msg(wc->status),
403 /* Close the transport */
404 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
407 /* Decrement used SQ WR count */
408 atomic_dec(&xprt->sc_sq_count);
409 wake_up(&xprt->sc_send_wait);
411 ctxt = (struct svc_rdma_op_ctxt *)
412 (unsigned long)wc->wr_id;
414 process_context(xprt, ctxt);
416 svc_xprt_put(&xprt->sc_xprt);
421 atomic_inc(&rdma_stat_sq_prod);
424 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
426 struct svcxprt_rdma *xprt = cq_context;
429 /* Guard against unconditional flush call for destroyed QP */
430 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
434 * Set the bit regardless of whether or not it's on the list
435 * because it may be on the list already due to an RQ
438 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
441 * If this transport is not already on the DTO transport queue,
444 spin_lock_irqsave(&dto_lock, flags);
445 if (list_empty(&xprt->sc_dto_q)) {
446 svc_xprt_get(&xprt->sc_xprt);
447 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
449 spin_unlock_irqrestore(&dto_lock, flags);
451 /* Tasklet does all the work to avoid irqsave locks. */
452 tasklet_schedule(&dto_tasklet);
455 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
458 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
462 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv);
463 INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
464 INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
465 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
466 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
467 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q);
468 init_waitqueue_head(&cma_xprt->sc_send_wait);
470 spin_lock_init(&cma_xprt->sc_lock);
471 spin_lock_init(&cma_xprt->sc_rq_dto_lock);
472 spin_lock_init(&cma_xprt->sc_frmr_q_lock);
474 cma_xprt->sc_ord = svcrdma_ord;
476 cma_xprt->sc_max_req_size = svcrdma_max_req_size;
477 cma_xprt->sc_max_requests = svcrdma_max_requests;
478 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
479 atomic_set(&cma_xprt->sc_sq_count, 0);
480 atomic_set(&cma_xprt->sc_ctxt_used, 0);
483 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
488 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
490 struct ib_recv_wr recv_wr, *bad_recv_wr;
491 struct svc_rdma_op_ctxt *ctxt;
498 ctxt = svc_rdma_get_context(xprt);
500 ctxt->direction = DMA_FROM_DEVICE;
501 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
502 if (sge_no >= xprt->sc_max_sge) {
503 pr_err("svcrdma: Too many sges (%d)\n", sge_no);
506 page = alloc_page(GFP_KERNEL | __GFP_NOFAIL);
507 ctxt->pages[sge_no] = page;
508 pa = ib_dma_map_page(xprt->sc_cm_id->device,
511 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa))
513 atomic_inc(&xprt->sc_dma_used);
514 ctxt->sge[sge_no].addr = pa;
515 ctxt->sge[sge_no].length = PAGE_SIZE;
516 ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey;
517 ctxt->count = sge_no + 1;
521 recv_wr.sg_list = &ctxt->sge[0];
522 recv_wr.num_sge = ctxt->count;
523 recv_wr.wr_id = (u64)(unsigned long)ctxt;
525 svc_xprt_get(&xprt->sc_xprt);
526 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
528 svc_rdma_unmap_dma(ctxt);
529 svc_rdma_put_context(ctxt, 1);
530 svc_xprt_put(&xprt->sc_xprt);
535 svc_rdma_unmap_dma(ctxt);
536 svc_rdma_put_context(ctxt, 1);
541 * This function handles the CONNECT_REQUEST event on a listening
542 * endpoint. It is passed the cma_id for the _new_ connection. The context in
543 * this cma_id is inherited from the listening cma_id and is the svc_xprt
544 * structure for the listening endpoint.
546 * This function creates a new xprt for the new connection and enqueues it on
547 * the accept queue for the listent xprt. When the listen thread is kicked, it
548 * will call the recvfrom method on the listen xprt which will accept the new
551 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird)
553 struct svcxprt_rdma *listen_xprt = new_cma_id->context;
554 struct svcxprt_rdma *newxprt;
557 /* Create a new transport */
558 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
560 dprintk("svcrdma: failed to create new transport\n");
563 newxprt->sc_cm_id = new_cma_id;
564 new_cma_id->context = newxprt;
565 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
566 newxprt, newxprt->sc_cm_id, listen_xprt);
568 /* Save client advertised inbound read limit for use later in accept. */
569 newxprt->sc_ord = client_ird;
571 /* Set the local and remote addresses in the transport */
572 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
573 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
574 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
575 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
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);
585 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
586 svc_xprt_enqueue(&listen_xprt->sc_xprt);
590 * Handles events generated on the listening endpoint. These events will be
591 * either be incoming connect requests or adapter removal events.
593 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
594 struct rdma_cm_event *event)
596 struct svcxprt_rdma *xprt = cma_id->context;
599 switch (event->event) {
600 case RDMA_CM_EVENT_CONNECT_REQUEST:
601 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
602 "event = %s (%d)\n", cma_id, cma_id->context,
603 rdma_event_msg(event->event), event->event);
604 handle_connect_req(cma_id,
605 event->param.conn.initiator_depth);
608 case RDMA_CM_EVENT_ESTABLISHED:
609 /* Accept complete */
610 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
611 "cm_id=%p\n", xprt, cma_id);
614 case RDMA_CM_EVENT_DEVICE_REMOVAL:
615 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
618 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
622 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
623 "event = %s (%d)\n", cma_id,
624 rdma_event_msg(event->event), event->event);
631 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
632 struct rdma_cm_event *event)
634 struct svc_xprt *xprt = cma_id->context;
635 struct svcxprt_rdma *rdma =
636 container_of(xprt, struct svcxprt_rdma, sc_xprt);
637 switch (event->event) {
638 case RDMA_CM_EVENT_ESTABLISHED:
639 /* Accept complete */
641 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
642 "cm_id=%p\n", xprt, cma_id);
643 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
644 svc_xprt_enqueue(xprt);
646 case RDMA_CM_EVENT_DISCONNECTED:
647 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
650 set_bit(XPT_CLOSE, &xprt->xpt_flags);
651 svc_xprt_enqueue(xprt);
655 case RDMA_CM_EVENT_DEVICE_REMOVAL:
656 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
657 "event = %s (%d)\n", cma_id, xprt,
658 rdma_event_msg(event->event), event->event);
660 set_bit(XPT_CLOSE, &xprt->xpt_flags);
661 svc_xprt_enqueue(xprt);
666 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
667 "event = %s (%d)\n", cma_id,
668 rdma_event_msg(event->event), event->event);
675 * Create a listening RDMA service endpoint.
677 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
679 struct sockaddr *sa, int salen,
682 struct rdma_cm_id *listen_id;
683 struct svcxprt_rdma *cma_xprt;
686 dprintk("svcrdma: Creating RDMA socket\n");
687 if (sa->sa_family != AF_INET) {
688 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family);
689 return ERR_PTR(-EAFNOSUPPORT);
691 cma_xprt = rdma_create_xprt(serv, 1);
693 return ERR_PTR(-ENOMEM);
695 listen_id = rdma_create_id(&init_net, rdma_listen_handler, cma_xprt,
696 RDMA_PS_TCP, IB_QPT_RC);
697 if (IS_ERR(listen_id)) {
698 ret = PTR_ERR(listen_id);
699 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
703 ret = rdma_bind_addr(listen_id, sa);
705 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
708 cma_xprt->sc_cm_id = listen_id;
710 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
712 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
717 * We need to use the address from the cm_id in case the
718 * caller specified 0 for the port number.
720 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
721 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
723 return &cma_xprt->sc_xprt;
726 rdma_destroy_id(listen_id);
732 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
735 struct scatterlist *sg;
736 struct svc_rdma_fastreg_mr *frmr;
739 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL);
743 num_sg = min_t(u32, RPCSVC_MAXPAGES, xprt->sc_frmr_pg_list_len);
744 mr = ib_alloc_mr(xprt->sc_pd, IB_MR_TYPE_MEM_REG, num_sg);
748 sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
752 sg_init_table(sg, RPCSVC_MAXPAGES);
756 INIT_LIST_HEAD(&frmr->frmr_list);
764 return ERR_PTR(-ENOMEM);
767 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
769 struct svc_rdma_fastreg_mr *frmr;
771 while (!list_empty(&xprt->sc_frmr_q)) {
772 frmr = list_entry(xprt->sc_frmr_q.next,
773 struct svc_rdma_fastreg_mr, frmr_list);
774 list_del_init(&frmr->frmr_list);
776 ib_dereg_mr(frmr->mr);
781 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
783 struct svc_rdma_fastreg_mr *frmr = NULL;
785 spin_lock_bh(&rdma->sc_frmr_q_lock);
786 if (!list_empty(&rdma->sc_frmr_q)) {
787 frmr = list_entry(rdma->sc_frmr_q.next,
788 struct svc_rdma_fastreg_mr, frmr_list);
789 list_del_init(&frmr->frmr_list);
792 spin_unlock_bh(&rdma->sc_frmr_q_lock);
796 return rdma_alloc_frmr(rdma);
799 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
800 struct svc_rdma_fastreg_mr *frmr)
803 ib_dma_unmap_sg(rdma->sc_cm_id->device,
804 frmr->sg, frmr->sg_nents, frmr->direction);
805 atomic_dec(&rdma->sc_dma_used);
806 spin_lock_bh(&rdma->sc_frmr_q_lock);
807 WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
808 list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
809 spin_unlock_bh(&rdma->sc_frmr_q_lock);
814 * This is the xpo_recvfrom function for listening endpoints. Its
815 * purpose is to accept incoming connections. The CMA callback handler
816 * has already created a new transport and attached it to the new CMA
819 * There is a queue of pending connections hung on the listening
820 * transport. This queue contains the new svc_xprt structure. This
821 * function takes svc_xprt structures off the accept_q and completes
824 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
826 struct svcxprt_rdma *listen_rdma;
827 struct svcxprt_rdma *newxprt = NULL;
828 struct rdma_conn_param conn_param;
829 struct ib_cq_init_attr cq_attr = {};
830 struct ib_qp_init_attr qp_attr;
831 struct ib_device_attr devattr;
832 int uninitialized_var(dma_mr_acc);
837 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
838 clear_bit(XPT_CONN, &xprt->xpt_flags);
839 /* Get the next entry off the accept list */
840 spin_lock_bh(&listen_rdma->sc_lock);
841 if (!list_empty(&listen_rdma->sc_accept_q)) {
842 newxprt = list_entry(listen_rdma->sc_accept_q.next,
843 struct svcxprt_rdma, sc_accept_q);
844 list_del_init(&newxprt->sc_accept_q);
846 if (!list_empty(&listen_rdma->sc_accept_q))
847 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
848 spin_unlock_bh(&listen_rdma->sc_lock);
852 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
853 newxprt, newxprt->sc_cm_id);
855 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
857 dprintk("svcrdma: could not query device attributes on "
858 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
862 /* Qualify the transport resource defaults with the
863 * capabilities of this particular device */
864 newxprt->sc_max_sge = min((size_t)devattr.max_sge,
865 (size_t)RPCSVC_MAXPAGES);
866 newxprt->sc_max_sge_rd = min_t(size_t, devattr.max_sge_rd,
868 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
869 (size_t)svcrdma_max_requests);
870 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
873 * Limit ORD based on client limit, local device limit, and
874 * configured svcrdma limit.
876 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord);
877 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord);
879 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
880 if (IS_ERR(newxprt->sc_pd)) {
881 dprintk("svcrdma: error creating PD for connect request\n");
884 cq_attr.cqe = newxprt->sc_sq_depth;
885 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
890 if (IS_ERR(newxprt->sc_sq_cq)) {
891 dprintk("svcrdma: error creating SQ CQ for connect request\n");
894 cq_attr.cqe = newxprt->sc_max_requests;
895 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
900 if (IS_ERR(newxprt->sc_rq_cq)) {
901 dprintk("svcrdma: error creating RQ CQ for connect request\n");
905 memset(&qp_attr, 0, sizeof qp_attr);
906 qp_attr.event_handler = qp_event_handler;
907 qp_attr.qp_context = &newxprt->sc_xprt;
908 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
909 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
910 qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
911 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
912 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
913 qp_attr.qp_type = IB_QPT_RC;
914 qp_attr.send_cq = newxprt->sc_sq_cq;
915 qp_attr.recv_cq = newxprt->sc_rq_cq;
916 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
917 " cm_id->device=%p, sc_pd->device=%p\n"
918 " cap.max_send_wr = %d\n"
919 " cap.max_recv_wr = %d\n"
920 " cap.max_send_sge = %d\n"
921 " cap.max_recv_sge = %d\n",
922 newxprt->sc_cm_id, newxprt->sc_pd,
923 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
924 qp_attr.cap.max_send_wr,
925 qp_attr.cap.max_recv_wr,
926 qp_attr.cap.max_send_sge,
927 qp_attr.cap.max_recv_sge);
929 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
931 dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
934 newxprt->sc_qp = newxprt->sc_cm_id->qp;
937 * Use the most secure set of MR resources based on the
938 * transport type and available memory management features in
939 * the device. Here's the table implemented below:
941 * Fast Global DMA Remote WR
943 * Sup'd Sup'd Needed Needed
955 * NB: iWARP requires remote write access for the data sink
956 * of an RDMA_READ. IB does not.
958 newxprt->sc_reader = rdma_read_chunk_lcl;
959 if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
960 newxprt->sc_frmr_pg_list_len =
961 devattr.max_fast_reg_page_list_len;
962 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG;
963 newxprt->sc_reader = rdma_read_chunk_frmr;
967 * Determine if a DMA MR is required and if so, what privs are required
969 if (!rdma_protocol_iwarp(newxprt->sc_cm_id->device,
970 newxprt->sc_cm_id->port_num) &&
971 !rdma_ib_or_roce(newxprt->sc_cm_id->device,
972 newxprt->sc_cm_id->port_num))
975 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG) ||
976 !(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) {
978 dma_mr_acc = IB_ACCESS_LOCAL_WRITE;
979 if (rdma_protocol_iwarp(newxprt->sc_cm_id->device,
980 newxprt->sc_cm_id->port_num) &&
981 !(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG))
982 dma_mr_acc |= IB_ACCESS_REMOTE_WRITE;
985 if (rdma_protocol_iwarp(newxprt->sc_cm_id->device,
986 newxprt->sc_cm_id->port_num))
987 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV;
989 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */
991 /* Register all of physical memory */
992 newxprt->sc_phys_mr =
993 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc);
994 if (IS_ERR(newxprt->sc_phys_mr)) {
995 dprintk("svcrdma: Failed to create DMA MR ret=%d\n",
999 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey;
1001 newxprt->sc_dma_lkey =
1002 newxprt->sc_cm_id->device->local_dma_lkey;
1004 /* Post receive buffers */
1005 for (i = 0; i < newxprt->sc_max_requests; i++) {
1006 ret = svc_rdma_post_recv(newxprt);
1008 dprintk("svcrdma: failure posting receive buffers\n");
1013 /* Swap out the handler */
1014 newxprt->sc_cm_id->event_handler = rdma_cma_handler;
1017 * Arm the CQs for the SQ and RQ before accepting so we can't
1018 * miss the first message
1020 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
1021 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
1023 /* Accept Connection */
1024 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
1025 memset(&conn_param, 0, sizeof conn_param);
1026 conn_param.responder_resources = 0;
1027 conn_param.initiator_depth = newxprt->sc_ord;
1028 ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
1030 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
1035 dprintk("svcrdma: new connection %p accepted with the following "
1037 " local_ip : %pI4\n"
1038 " local_port : %d\n"
1039 " remote_ip : %pI4\n"
1040 " remote_port : %d\n"
1042 " max_sge_rd : %d\n"
1044 " max_requests : %d\n"
1047 &((struct sockaddr_in *)&newxprt->sc_cm_id->
1048 route.addr.src_addr)->sin_addr.s_addr,
1049 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1050 route.addr.src_addr)->sin_port),
1051 &((struct sockaddr_in *)&newxprt->sc_cm_id->
1052 route.addr.dst_addr)->sin_addr.s_addr,
1053 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
1054 route.addr.dst_addr)->sin_port),
1055 newxprt->sc_max_sge,
1056 newxprt->sc_max_sge_rd,
1057 newxprt->sc_sq_depth,
1058 newxprt->sc_max_requests,
1061 return &newxprt->sc_xprt;
1064 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
1065 /* Take a reference in case the DTO handler runs */
1066 svc_xprt_get(&newxprt->sc_xprt);
1067 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
1068 ib_destroy_qp(newxprt->sc_qp);
1069 rdma_destroy_id(newxprt->sc_cm_id);
1070 /* This call to put will destroy the transport */
1071 svc_xprt_put(&newxprt->sc_xprt);
1075 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
1080 * When connected, an svc_xprt has at least two references:
1082 * - A reference held by the cm_id between the ESTABLISHED and
1083 * DISCONNECTED events. If the remote peer disconnected first, this
1084 * reference could be gone.
1086 * - A reference held by the svc_recv code that called this function
1087 * as part of close processing.
1089 * At a minimum one references should still be held.
1091 static void svc_rdma_detach(struct svc_xprt *xprt)
1093 struct svcxprt_rdma *rdma =
1094 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1095 dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1097 /* Disconnect and flush posted WQE */
1098 rdma_disconnect(rdma->sc_cm_id);
1101 static void __svc_rdma_free(struct work_struct *work)
1103 struct svcxprt_rdma *rdma =
1104 container_of(work, struct svcxprt_rdma, sc_work);
1105 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1107 /* We should only be called from kref_put */
1108 if (atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0)
1109 pr_err("svcrdma: sc_xprt still in use? (%d)\n",
1110 atomic_read(&rdma->sc_xprt.xpt_ref.refcount));
1113 * Destroy queued, but not processed read completions. Note
1114 * that this cleanup has to be done before destroying the
1115 * cm_id because the device ptr is needed to unmap the dma in
1116 * svc_rdma_put_context.
1118 while (!list_empty(&rdma->sc_read_complete_q)) {
1119 struct svc_rdma_op_ctxt *ctxt;
1120 ctxt = list_entry(rdma->sc_read_complete_q.next,
1121 struct svc_rdma_op_ctxt,
1123 list_del_init(&ctxt->dto_q);
1124 svc_rdma_put_context(ctxt, 1);
1127 /* Destroy queued, but not processed recv completions */
1128 while (!list_empty(&rdma->sc_rq_dto_q)) {
1129 struct svc_rdma_op_ctxt *ctxt;
1130 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1131 struct svc_rdma_op_ctxt,
1133 list_del_init(&ctxt->dto_q);
1134 svc_rdma_put_context(ctxt, 1);
1137 /* Warn if we leaked a resource or under-referenced */
1138 if (atomic_read(&rdma->sc_ctxt_used) != 0)
1139 pr_err("svcrdma: ctxt still in use? (%d)\n",
1140 atomic_read(&rdma->sc_ctxt_used));
1141 if (atomic_read(&rdma->sc_dma_used) != 0)
1142 pr_err("svcrdma: dma still in use? (%d)\n",
1143 atomic_read(&rdma->sc_dma_used));
1145 /* De-allocate fastreg mr */
1146 rdma_dealloc_frmr_q(rdma);
1148 /* Destroy the QP if present (not a listener) */
1149 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1150 ib_destroy_qp(rdma->sc_qp);
1152 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1153 ib_destroy_cq(rdma->sc_sq_cq);
1155 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1156 ib_destroy_cq(rdma->sc_rq_cq);
1158 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1159 ib_dereg_mr(rdma->sc_phys_mr);
1161 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1162 ib_dealloc_pd(rdma->sc_pd);
1164 /* Destroy the CM ID */
1165 rdma_destroy_id(rdma->sc_cm_id);
1170 static void svc_rdma_free(struct svc_xprt *xprt)
1172 struct svcxprt_rdma *rdma =
1173 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1174 INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1175 queue_work(svc_rdma_wq, &rdma->sc_work);
1178 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1180 struct svcxprt_rdma *rdma =
1181 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1184 * If there are already waiters on the SQ,
1187 if (waitqueue_active(&rdma->sc_send_wait))
1190 /* Otherwise return true. */
1194 static int svc_rdma_secure_port(struct svc_rqst *rqstp)
1199 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1201 struct ib_send_wr *bad_wr, *n_wr;
1206 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1210 for (n_wr = wr->next; n_wr; n_wr = n_wr->next)
1213 /* If the SQ is full, wait until an SQ entry is available */
1215 spin_lock_bh(&xprt->sc_lock);
1216 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) {
1217 spin_unlock_bh(&xprt->sc_lock);
1218 atomic_inc(&rdma_stat_sq_starve);
1220 /* See if we can opportunistically reap SQ WR to make room */
1223 /* Wait until SQ WR available if SQ still full */
1224 wait_event(xprt->sc_send_wait,
1225 atomic_read(&xprt->sc_sq_count) <
1227 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1231 /* Take a transport ref for each WR posted */
1232 for (i = 0; i < wr_count; i++)
1233 svc_xprt_get(&xprt->sc_xprt);
1235 /* Bump used SQ WR count and post */
1236 atomic_add(wr_count, &xprt->sc_sq_count);
1237 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1239 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
1240 atomic_sub(wr_count, &xprt->sc_sq_count);
1241 for (i = 0; i < wr_count; i ++)
1242 svc_xprt_put(&xprt->sc_xprt);
1243 dprintk("svcrdma: failed to post SQ WR rc=%d, "
1244 "sc_sq_count=%d, sc_sq_depth=%d\n",
1245 ret, atomic_read(&xprt->sc_sq_count),
1248 spin_unlock_bh(&xprt->sc_lock);
1250 wake_up(&xprt->sc_send_wait);
1256 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1257 enum rpcrdma_errcode err)
1259 struct ib_send_wr err_wr;
1261 struct svc_rdma_op_ctxt *ctxt;
1266 p = alloc_page(GFP_KERNEL | __GFP_NOFAIL);
1267 va = page_address(p);
1269 /* XDR encode error */
1270 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1272 ctxt = svc_rdma_get_context(xprt);
1273 ctxt->direction = DMA_FROM_DEVICE;
1277 /* Prepare SGE for local address */
1278 ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device,
1279 p, 0, length, DMA_FROM_DEVICE);
1280 if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) {
1282 svc_rdma_put_context(ctxt, 1);
1285 atomic_inc(&xprt->sc_dma_used);
1286 ctxt->sge[0].lkey = xprt->sc_dma_lkey;
1287 ctxt->sge[0].length = length;
1289 /* Prepare SEND WR */
1290 memset(&err_wr, 0, sizeof err_wr);
1291 ctxt->wr_op = IB_WR_SEND;
1292 err_wr.wr_id = (unsigned long)ctxt;
1293 err_wr.sg_list = ctxt->sge;
1295 err_wr.opcode = IB_WR_SEND;
1296 err_wr.send_flags = IB_SEND_SIGNALED;
1299 ret = svc_rdma_send(xprt, &err_wr);
1301 dprintk("svcrdma: Error %d posting send for protocol error\n",
1303 svc_rdma_unmap_dma(ctxt);
1304 svc_rdma_put_context(ctxt, 1);