2 * Copyright (c) 2003-2007 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.
43 * Encapsulates the major functions managing:
50 #include <linux/interrupt.h>
51 #include <linux/slab.h>
52 #include <linux/prefetch.h>
53 #include <linux/sunrpc/addr.h>
54 #include <asm/bitops.h>
55 #include <linux/module.h> /* try_module_get()/module_put() */
57 #include "xprt_rdma.h"
63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
64 # define RPCDBG_FACILITY RPCDBG_TRANS
71 static struct workqueue_struct *rpcrdma_receive_wq;
74 rpcrdma_alloc_wq(void)
76 struct workqueue_struct *recv_wq;
78 recv_wq = alloc_workqueue("xprtrdma_receive",
79 WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
84 rpcrdma_receive_wq = recv_wq;
89 rpcrdma_destroy_wq(void)
91 struct workqueue_struct *wq;
93 if (rpcrdma_receive_wq) {
94 wq = rpcrdma_receive_wq;
95 rpcrdma_receive_wq = NULL;
96 destroy_workqueue(wq);
101 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
103 struct rpcrdma_ep *ep = context;
105 pr_err("RPC: %s: %s on device %s ep %p\n",
106 __func__, ib_event_msg(event->event),
107 event->device->name, context);
108 if (ep->rep_connected == 1) {
109 ep->rep_connected = -EIO;
110 rpcrdma_conn_func(ep);
111 wake_up_all(&ep->rep_connect_wait);
116 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
118 struct rpcrdma_ep *ep = context;
120 pr_err("RPC: %s: %s on device %s ep %p\n",
121 __func__, ib_event_msg(event->event),
122 event->device->name, context);
123 if (ep->rep_connected == 1) {
124 ep->rep_connected = -EIO;
125 rpcrdma_conn_func(ep);
126 wake_up_all(&ep->rep_connect_wait);
131 rpcrdma_sendcq_process_wc(struct ib_wc *wc)
133 /* WARNING: Only wr_id and status are reliable at this point */
134 if (wc->wr_id == RPCRDMA_IGNORE_COMPLETION) {
135 if (wc->status != IB_WC_SUCCESS &&
136 wc->status != IB_WC_WR_FLUSH_ERR)
137 pr_err("RPC: %s: SEND: %s\n",
138 __func__, ib_wc_status_msg(wc->status));
140 struct rpcrdma_mw *r;
142 r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
143 r->mw_sendcompletion(wc);
147 /* The common case is a single send completion is waiting. By
148 * passing two WC entries to ib_poll_cq, a return code of 1
149 * means there is exactly one WC waiting and no more. We don't
150 * have to invoke ib_poll_cq again to know that the CQ has been
154 rpcrdma_sendcq_poll(struct ib_cq *cq)
156 struct ib_wc *pos, wcs[2];
162 rc = ib_poll_cq(cq, ARRAY_SIZE(wcs), pos);
168 rpcrdma_sendcq_process_wc(pos++);
169 } while (rc == ARRAY_SIZE(wcs));
173 /* Handle provider send completion upcalls.
176 rpcrdma_sendcq_upcall(struct ib_cq *cq, void *cq_context)
179 rpcrdma_sendcq_poll(cq);
180 } while (ib_req_notify_cq(cq, IB_CQ_NEXT_COMP |
181 IB_CQ_REPORT_MISSED_EVENTS) > 0);
185 rpcrdma_receive_worker(struct work_struct *work)
187 struct rpcrdma_rep *rep =
188 container_of(work, struct rpcrdma_rep, rr_work);
190 rpcrdma_reply_handler(rep);
194 rpcrdma_recvcq_process_wc(struct ib_wc *wc)
196 struct rpcrdma_rep *rep =
197 (struct rpcrdma_rep *)(unsigned long)wc->wr_id;
199 /* WARNING: Only wr_id and status are reliable at this point */
200 if (wc->status != IB_WC_SUCCESS)
203 /* status == SUCCESS means all fields in wc are trustworthy */
204 if (wc->opcode != IB_WC_RECV)
207 dprintk("RPC: %s: rep %p opcode 'recv', length %u: success\n",
208 __func__, rep, wc->byte_len);
210 rep->rr_len = wc->byte_len;
211 ib_dma_sync_single_for_cpu(rep->rr_device,
212 rdmab_addr(rep->rr_rdmabuf),
213 rep->rr_len, DMA_FROM_DEVICE);
214 prefetch(rdmab_to_msg(rep->rr_rdmabuf));
217 queue_work(rpcrdma_receive_wq, &rep->rr_work);
221 if (wc->status != IB_WC_WR_FLUSH_ERR)
222 pr_err("RPC: %s: rep %p: %s\n",
223 __func__, rep, ib_wc_status_msg(wc->status));
224 rep->rr_len = RPCRDMA_BAD_LEN;
228 /* The wc array is on stack: automatic memory is always CPU-local.
230 * struct ib_wc is 64 bytes, making the poll array potentially
231 * large. But this is at the bottom of the call chain. Further
232 * substantial work is done in another thread.
235 rpcrdma_recvcq_poll(struct ib_cq *cq)
237 struct ib_wc *pos, wcs[4];
243 rc = ib_poll_cq(cq, ARRAY_SIZE(wcs), pos);
249 rpcrdma_recvcq_process_wc(pos++);
250 } while (rc == ARRAY_SIZE(wcs));
253 /* Handle provider receive completion upcalls.
256 rpcrdma_recvcq_upcall(struct ib_cq *cq, void *cq_context)
259 rpcrdma_recvcq_poll(cq);
260 } while (ib_req_notify_cq(cq, IB_CQ_NEXT_COMP |
261 IB_CQ_REPORT_MISSED_EVENTS) > 0);
265 rpcrdma_flush_cqs(struct rpcrdma_ep *ep)
269 while (ib_poll_cq(ep->rep_attr.recv_cq, 1, &wc) > 0)
270 rpcrdma_recvcq_process_wc(&wc);
271 while (ib_poll_cq(ep->rep_attr.send_cq, 1, &wc) > 0)
272 rpcrdma_sendcq_process_wc(&wc);
276 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
278 struct rpcrdma_xprt *xprt = id->context;
279 struct rpcrdma_ia *ia = &xprt->rx_ia;
280 struct rpcrdma_ep *ep = &xprt->rx_ep;
281 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
282 struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
284 struct ib_qp_attr *attr = &ia->ri_qp_attr;
285 struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
288 switch (event->event) {
289 case RDMA_CM_EVENT_ADDR_RESOLVED:
290 case RDMA_CM_EVENT_ROUTE_RESOLVED:
292 complete(&ia->ri_done);
294 case RDMA_CM_EVENT_ADDR_ERROR:
295 ia->ri_async_rc = -EHOSTUNREACH;
296 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
298 complete(&ia->ri_done);
300 case RDMA_CM_EVENT_ROUTE_ERROR:
301 ia->ri_async_rc = -ENETUNREACH;
302 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
304 complete(&ia->ri_done);
306 case RDMA_CM_EVENT_ESTABLISHED:
308 ib_query_qp(ia->ri_id->qp, attr,
309 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
311 dprintk("RPC: %s: %d responder resources"
313 __func__, attr->max_dest_rd_atomic,
314 attr->max_rd_atomic);
316 case RDMA_CM_EVENT_CONNECT_ERROR:
317 connstate = -ENOTCONN;
319 case RDMA_CM_EVENT_UNREACHABLE:
320 connstate = -ENETDOWN;
322 case RDMA_CM_EVENT_REJECTED:
323 connstate = -ECONNREFUSED;
325 case RDMA_CM_EVENT_DISCONNECTED:
326 connstate = -ECONNABORTED;
328 case RDMA_CM_EVENT_DEVICE_REMOVAL:
331 dprintk("RPC: %s: %sconnected\n",
332 __func__, connstate > 0 ? "" : "dis");
333 ep->rep_connected = connstate;
334 rpcrdma_conn_func(ep);
335 wake_up_all(&ep->rep_connect_wait);
338 dprintk("RPC: %s: %pIS:%u (ep 0x%p): %s\n",
339 __func__, sap, rpc_get_port(sap), ep,
340 rdma_event_msg(event->event));
344 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
345 if (connstate == 1) {
346 int ird = attr->max_dest_rd_atomic;
347 int tird = ep->rep_remote_cma.responder_resources;
349 pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
350 sap, rpc_get_port(sap),
352 ia->ri_ops->ro_displayname,
353 xprt->rx_buf.rb_max_requests,
354 ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
355 } else if (connstate < 0) {
356 pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
357 sap, rpc_get_port(sap), connstate);
364 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
367 module_put(id->device->owner);
372 static struct rdma_cm_id *
373 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
374 struct rpcrdma_ia *ia, struct sockaddr *addr)
376 struct rdma_cm_id *id;
379 init_completion(&ia->ri_done);
381 id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
385 dprintk("RPC: %s: rdma_create_id() failed %i\n",
390 ia->ri_async_rc = -ETIMEDOUT;
391 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
393 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
397 wait_for_completion_interruptible_timeout(&ia->ri_done,
398 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
401 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
402 * be pinned while there are active NFS/RDMA mounts to prevent
403 * hangs and crashes at umount time.
405 if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
406 dprintk("RPC: %s: Failed to get device module\n",
408 ia->ri_async_rc = -ENODEV;
410 rc = ia->ri_async_rc;
414 ia->ri_async_rc = -ETIMEDOUT;
415 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
417 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
421 wait_for_completion_interruptible_timeout(&ia->ri_done,
422 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
423 rc = ia->ri_async_rc;
429 module_put(id->device->owner);
436 * Drain any cq, prior to teardown.
439 rpcrdma_clean_cq(struct ib_cq *cq)
444 while (1 == ib_poll_cq(cq, 1, &wc))
448 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
449 __func__, count, wc.opcode);
453 * Exported functions.
457 * Open and initialize an Interface Adapter.
458 * o initializes fields of struct rpcrdma_ia, including
459 * interface and provider attributes and protection zone.
462 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
464 struct rpcrdma_ia *ia = &xprt->rx_ia;
465 struct ib_device_attr *devattr = &ia->ri_devattr;
468 ia->ri_dma_mr = NULL;
470 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
471 if (IS_ERR(ia->ri_id)) {
472 rc = PTR_ERR(ia->ri_id);
475 ia->ri_device = ia->ri_id->device;
477 ia->ri_pd = ib_alloc_pd(ia->ri_device);
478 if (IS_ERR(ia->ri_pd)) {
479 rc = PTR_ERR(ia->ri_pd);
480 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
485 rc = ib_query_device(ia->ri_device, devattr);
487 dprintk("RPC: %s: ib_query_device failed %d\n",
492 if (memreg == RPCRDMA_FRMR) {
493 if (!(devattr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) ||
494 (devattr->max_fast_reg_page_list_len == 0)) {
495 dprintk("RPC: %s: FRMR registration "
496 "not supported by HCA\n", __func__);
497 memreg = RPCRDMA_MTHCAFMR;
500 if (memreg == RPCRDMA_MTHCAFMR) {
501 if (!ia->ri_device->alloc_fmr) {
502 dprintk("RPC: %s: MTHCAFMR registration "
503 "not supported by HCA\n", __func__);
511 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
513 case RPCRDMA_ALLPHYSICAL:
514 ia->ri_ops = &rpcrdma_physical_memreg_ops;
516 case RPCRDMA_MTHCAFMR:
517 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
520 printk(KERN_ERR "RPC: Unsupported memory "
521 "registration mode: %d\n", memreg);
525 dprintk("RPC: %s: memory registration strategy is '%s'\n",
526 __func__, ia->ri_ops->ro_displayname);
528 rwlock_init(&ia->ri_qplock);
532 ib_dealloc_pd(ia->ri_pd);
535 rpcrdma_destroy_id(ia->ri_id);
542 * Clean up/close an IA.
543 * o if event handles and PD have been initialized, free them.
547 rpcrdma_ia_close(struct rpcrdma_ia *ia)
549 dprintk("RPC: %s: entering\n", __func__);
550 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
552 rdma_destroy_qp(ia->ri_id);
553 rpcrdma_destroy_id(ia->ri_id);
557 /* If the pd is still busy, xprtrdma missed freeing a resource */
558 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
559 ib_dealloc_pd(ia->ri_pd);
563 * Create unconnected endpoint.
566 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
567 struct rpcrdma_create_data_internal *cdata)
569 struct ib_device_attr *devattr = &ia->ri_devattr;
570 struct ib_cq *sendcq, *recvcq;
571 struct ib_cq_init_attr cq_attr = {};
572 unsigned int max_qp_wr;
575 if (devattr->max_sge < RPCRDMA_MAX_IOVS) {
576 dprintk("RPC: %s: insufficient sge's available\n",
581 if (devattr->max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
582 dprintk("RPC: %s: insufficient wqe's available\n",
586 max_qp_wr = devattr->max_qp_wr - RPCRDMA_BACKWARD_WRS;
588 /* check provider's send/recv wr limits */
589 if (cdata->max_requests > max_qp_wr)
590 cdata->max_requests = max_qp_wr;
592 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
593 ep->rep_attr.qp_context = ep;
594 ep->rep_attr.srq = NULL;
595 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
596 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
597 rc = ia->ri_ops->ro_open(ia, ep, cdata);
600 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
601 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
602 ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
603 ep->rep_attr.cap.max_recv_sge = 1;
604 ep->rep_attr.cap.max_inline_data = 0;
605 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
606 ep->rep_attr.qp_type = IB_QPT_RC;
607 ep->rep_attr.port_num = ~0;
609 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
610 "iovs: send %d recv %d\n",
612 ep->rep_attr.cap.max_send_wr,
613 ep->rep_attr.cap.max_recv_wr,
614 ep->rep_attr.cap.max_send_sge,
615 ep->rep_attr.cap.max_recv_sge);
617 /* set trigger for requesting send completion */
618 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
619 if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS)
620 ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS;
621 else if (ep->rep_cqinit <= 2)
624 init_waitqueue_head(&ep->rep_connect_wait);
625 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
627 cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1;
628 sendcq = ib_create_cq(ia->ri_device, rpcrdma_sendcq_upcall,
629 rpcrdma_cq_async_error_upcall, NULL, &cq_attr);
630 if (IS_ERR(sendcq)) {
631 rc = PTR_ERR(sendcq);
632 dprintk("RPC: %s: failed to create send CQ: %i\n",
637 rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP);
639 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
644 cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1;
645 recvcq = ib_create_cq(ia->ri_device, rpcrdma_recvcq_upcall,
646 rpcrdma_cq_async_error_upcall, NULL, &cq_attr);
647 if (IS_ERR(recvcq)) {
648 rc = PTR_ERR(recvcq);
649 dprintk("RPC: %s: failed to create recv CQ: %i\n",
654 rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP);
656 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
658 ib_destroy_cq(recvcq);
662 ep->rep_attr.send_cq = sendcq;
663 ep->rep_attr.recv_cq = recvcq;
665 /* Initialize cma parameters */
667 /* RPC/RDMA does not use private data */
668 ep->rep_remote_cma.private_data = NULL;
669 ep->rep_remote_cma.private_data_len = 0;
671 /* Client offers RDMA Read but does not initiate */
672 ep->rep_remote_cma.initiator_depth = 0;
673 if (devattr->max_qp_rd_atom > 32) /* arbitrary but <= 255 */
674 ep->rep_remote_cma.responder_resources = 32;
676 ep->rep_remote_cma.responder_resources =
677 devattr->max_qp_rd_atom;
679 ep->rep_remote_cma.retry_count = 7;
680 ep->rep_remote_cma.flow_control = 0;
681 ep->rep_remote_cma.rnr_retry_count = 0;
686 err = ib_destroy_cq(sendcq);
688 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
692 ib_dereg_mr(ia->ri_dma_mr);
699 * Disconnect and destroy endpoint. After this, the only
700 * valid operations on the ep are to free it (if dynamically
701 * allocated) or re-create it.
704 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
708 dprintk("RPC: %s: entering, connected is %d\n",
709 __func__, ep->rep_connected);
711 cancel_delayed_work_sync(&ep->rep_connect_worker);
714 rpcrdma_ep_disconnect(ep, ia);
716 rpcrdma_clean_cq(ep->rep_attr.recv_cq);
717 rpcrdma_clean_cq(ep->rep_attr.send_cq);
720 rdma_destroy_qp(ia->ri_id);
721 ia->ri_id->qp = NULL;
724 rc = ib_destroy_cq(ep->rep_attr.recv_cq);
726 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
729 rc = ib_destroy_cq(ep->rep_attr.send_cq);
731 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
735 rc = ib_dereg_mr(ia->ri_dma_mr);
736 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
742 * Connect unconnected endpoint.
745 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
747 struct rdma_cm_id *id, *old;
751 if (ep->rep_connected != 0) {
752 struct rpcrdma_xprt *xprt;
754 dprintk("RPC: %s: reconnecting...\n", __func__);
756 rpcrdma_ep_disconnect(ep, ia);
757 rpcrdma_flush_cqs(ep);
759 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
760 id = rpcrdma_create_id(xprt, ia,
761 (struct sockaddr *)&xprt->rx_data.addr);
766 /* TEMP TEMP TEMP - fail if new device:
767 * Deregister/remarshal *all* requests!
768 * Close and recreate adapter, pd, etc!
769 * Re-determine all attributes still sane!
770 * More stuff I haven't thought of!
773 if (ia->ri_device != id->device) {
774 printk("RPC: %s: can't reconnect on "
775 "different device!\n", __func__);
776 rpcrdma_destroy_id(id);
781 rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
783 dprintk("RPC: %s: rdma_create_qp failed %i\n",
785 rpcrdma_destroy_id(id);
790 write_lock(&ia->ri_qplock);
793 write_unlock(&ia->ri_qplock);
795 rdma_destroy_qp(old);
796 rpcrdma_destroy_id(old);
798 dprintk("RPC: %s: connecting...\n", __func__);
799 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
801 dprintk("RPC: %s: rdma_create_qp failed %i\n",
803 /* do not update ep->rep_connected */
808 ep->rep_connected = 0;
810 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
812 dprintk("RPC: %s: rdma_connect() failed with %i\n",
817 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
820 * Check state. A non-peer reject indicates no listener
821 * (ECONNREFUSED), which may be a transient state. All
822 * others indicate a transport condition which has already
823 * undergone a best-effort.
825 if (ep->rep_connected == -ECONNREFUSED &&
826 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
827 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
830 if (ep->rep_connected <= 0) {
831 /* Sometimes, the only way to reliably connect to remote
832 * CMs is to use same nonzero values for ORD and IRD. */
833 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
834 (ep->rep_remote_cma.responder_resources == 0 ||
835 ep->rep_remote_cma.initiator_depth !=
836 ep->rep_remote_cma.responder_resources)) {
837 if (ep->rep_remote_cma.responder_resources == 0)
838 ep->rep_remote_cma.responder_resources = 1;
839 ep->rep_remote_cma.initiator_depth =
840 ep->rep_remote_cma.responder_resources;
843 rc = ep->rep_connected;
845 struct rpcrdma_xprt *r_xprt;
848 dprintk("RPC: %s: connected\n", __func__);
850 r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
851 extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
854 rc = rpcrdma_ep_post_extra_recv(r_xprt, extras);
856 pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
864 ep->rep_connected = rc;
869 * rpcrdma_ep_disconnect
871 * This is separate from destroy to facilitate the ability
872 * to reconnect without recreating the endpoint.
874 * This call is not reentrant, and must not be made in parallel
875 * on the same endpoint.
878 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
882 rpcrdma_flush_cqs(ep);
883 rc = rdma_disconnect(ia->ri_id);
885 /* returns without wait if not connected */
886 wait_event_interruptible(ep->rep_connect_wait,
887 ep->rep_connected != 1);
888 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
889 (ep->rep_connected == 1) ? "still " : "dis");
891 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
892 ep->rep_connected = rc;
897 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
899 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
900 struct rpcrdma_req *req;
902 req = kzalloc(sizeof(*req), GFP_KERNEL);
904 return ERR_PTR(-ENOMEM);
906 INIT_LIST_HEAD(&req->rl_free);
907 spin_lock(&buffer->rb_reqslock);
908 list_add(&req->rl_all, &buffer->rb_allreqs);
909 spin_unlock(&buffer->rb_reqslock);
910 req->rl_buffer = &r_xprt->rx_buf;
915 * rpcrdma_create_rep - Allocate an rpcrdma_rep object
916 * @r_xprt: controlling transport
918 * Returns 0 on success or a negative errno on failure.
921 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
923 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
924 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
925 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
926 struct rpcrdma_rep *rep;
930 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
934 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
936 if (IS_ERR(rep->rr_rdmabuf)) {
937 rc = PTR_ERR(rep->rr_rdmabuf);
941 rep->rr_device = ia->ri_device;
942 rep->rr_rxprt = r_xprt;
943 INIT_WORK(&rep->rr_work, rpcrdma_receive_worker);
945 spin_lock(&buf->rb_lock);
946 list_add(&rep->rr_list, &buf->rb_recv_bufs);
947 spin_unlock(&buf->rb_lock);
953 dprintk("RPC: %s: reply buffer %d alloc failed\n",
959 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
961 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
962 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
965 buf->rb_max_requests = r_xprt->rx_data.max_requests;
966 buf->rb_bc_srv_max_requests = 0;
967 spin_lock_init(&buf->rb_lock);
969 rc = ia->ri_ops->ro_init(r_xprt);
973 INIT_LIST_HEAD(&buf->rb_send_bufs);
974 INIT_LIST_HEAD(&buf->rb_allreqs);
975 spin_lock_init(&buf->rb_reqslock);
976 for (i = 0; i < buf->rb_max_requests; i++) {
977 struct rpcrdma_req *req;
979 req = rpcrdma_create_req(r_xprt);
981 dprintk("RPC: %s: request buffer %d alloc"
982 " failed\n", __func__, i);
986 req->rl_backchannel = false;
987 list_add(&req->rl_free, &buf->rb_send_bufs);
990 INIT_LIST_HEAD(&buf->rb_recv_bufs);
991 for (i = 0; i <= buf->rb_max_requests; i++) {
992 rc = rpcrdma_create_rep(r_xprt);
999 rpcrdma_buffer_destroy(buf);
1003 static struct rpcrdma_req *
1004 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1006 struct rpcrdma_req *req;
1008 req = list_first_entry(&buf->rb_send_bufs,
1009 struct rpcrdma_req, rl_free);
1010 list_del(&req->rl_free);
1014 static struct rpcrdma_rep *
1015 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1017 struct rpcrdma_rep *rep;
1019 rep = list_first_entry(&buf->rb_recv_bufs,
1020 struct rpcrdma_rep, rr_list);
1021 list_del(&rep->rr_list);
1026 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
1028 rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
1033 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
1035 rpcrdma_free_regbuf(ia, req->rl_sendbuf);
1036 rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
1041 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1043 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1045 while (!list_empty(&buf->rb_recv_bufs)) {
1046 struct rpcrdma_rep *rep;
1048 rep = rpcrdma_buffer_get_rep_locked(buf);
1049 rpcrdma_destroy_rep(ia, rep);
1052 spin_lock(&buf->rb_reqslock);
1053 while (!list_empty(&buf->rb_allreqs)) {
1054 struct rpcrdma_req *req;
1056 req = list_first_entry(&buf->rb_allreqs,
1057 struct rpcrdma_req, rl_all);
1058 list_del(&req->rl_all);
1060 spin_unlock(&buf->rb_reqslock);
1061 rpcrdma_destroy_req(ia, req);
1062 spin_lock(&buf->rb_reqslock);
1064 spin_unlock(&buf->rb_reqslock);
1066 ia->ri_ops->ro_destroy(buf);
1070 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1072 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1073 struct rpcrdma_mw *mw = NULL;
1075 spin_lock(&buf->rb_mwlock);
1076 if (!list_empty(&buf->rb_mws)) {
1077 mw = list_first_entry(&buf->rb_mws,
1078 struct rpcrdma_mw, mw_list);
1079 list_del_init(&mw->mw_list);
1081 spin_unlock(&buf->rb_mwlock);
1084 pr_err("RPC: %s: no MWs available\n", __func__);
1089 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1091 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1093 spin_lock(&buf->rb_mwlock);
1094 list_add_tail(&mw->mw_list, &buf->rb_mws);
1095 spin_unlock(&buf->rb_mwlock);
1099 * Get a set of request/reply buffers.
1101 * Reply buffer (if available) is attached to send buffer upon return.
1103 struct rpcrdma_req *
1104 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1106 struct rpcrdma_req *req;
1108 spin_lock(&buffers->rb_lock);
1109 if (list_empty(&buffers->rb_send_bufs))
1111 req = rpcrdma_buffer_get_req_locked(buffers);
1112 if (list_empty(&buffers->rb_recv_bufs))
1114 req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers);
1115 spin_unlock(&buffers->rb_lock);
1119 spin_unlock(&buffers->rb_lock);
1120 pr_warn("RPC: %s: out of request buffers\n", __func__);
1123 spin_unlock(&buffers->rb_lock);
1124 pr_warn("RPC: %s: out of reply buffers\n", __func__);
1125 req->rl_reply = NULL;
1130 * Put request/reply buffers back into pool.
1131 * Pre-decrement counter/array index.
1134 rpcrdma_buffer_put(struct rpcrdma_req *req)
1136 struct rpcrdma_buffer *buffers = req->rl_buffer;
1137 struct rpcrdma_rep *rep = req->rl_reply;
1140 req->rl_reply = NULL;
1142 spin_lock(&buffers->rb_lock);
1143 list_add_tail(&req->rl_free, &buffers->rb_send_bufs);
1145 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1146 spin_unlock(&buffers->rb_lock);
1150 * Recover reply buffers from pool.
1151 * This happens when recovering from disconnect.
1154 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1156 struct rpcrdma_buffer *buffers = req->rl_buffer;
1158 spin_lock(&buffers->rb_lock);
1159 if (!list_empty(&buffers->rb_recv_bufs))
1160 req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers);
1161 spin_unlock(&buffers->rb_lock);
1165 * Put reply buffers back into pool when not attached to
1166 * request. This happens in error conditions.
1169 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1171 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1173 spin_lock(&buffers->rb_lock);
1174 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1175 spin_unlock(&buffers->rb_lock);
1179 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1183 rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
1185 dprintk("RPC: map_one: offset %p iova %llx len %zu\n",
1187 (unsigned long long)seg->mr_dma, seg->mr_dmalen);
1191 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1192 * @ia: controlling rpcrdma_ia
1193 * @size: size of buffer to be allocated, in bytes
1196 * Returns pointer to private header of an area of internally
1197 * registered memory, or an ERR_PTR. The registered buffer follows
1198 * the end of the private header.
1200 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1201 * receiving the payload of RDMA RECV operations. regbufs are not
1202 * used for RDMA READ/WRITE operations, thus are registered only for
1205 struct rpcrdma_regbuf *
1206 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1208 struct rpcrdma_regbuf *rb;
1211 rb = kmalloc(sizeof(*rb) + size, flags);
1216 iov->addr = ib_dma_map_single(ia->ri_device,
1217 (void *)rb->rg_base, size,
1219 if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1223 iov->lkey = ia->ri_pd->local_dma_lkey;
1225 rb->rg_owner = NULL;
1231 return ERR_PTR(-ENOMEM);
1235 * rpcrdma_free_regbuf - deregister and free registered buffer
1236 * @ia: controlling rpcrdma_ia
1237 * @rb: regbuf to be deregistered and freed
1240 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1248 ib_dma_unmap_single(ia->ri_device,
1249 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1254 * Prepost any receive buffer, then post send.
1256 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1259 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1260 struct rpcrdma_ep *ep,
1261 struct rpcrdma_req *req)
1263 struct ib_device *device = ia->ri_device;
1264 struct ib_send_wr send_wr, *send_wr_fail;
1265 struct rpcrdma_rep *rep = req->rl_reply;
1266 struct ib_sge *iov = req->rl_send_iov;
1270 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1273 req->rl_reply = NULL;
1276 send_wr.next = NULL;
1277 send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
1278 send_wr.sg_list = iov;
1279 send_wr.num_sge = req->rl_niovs;
1280 send_wr.opcode = IB_WR_SEND;
1282 for (i = 0; i < send_wr.num_sge; i++)
1283 ib_dma_sync_single_for_device(device, iov[i].addr,
1284 iov[i].length, DMA_TO_DEVICE);
1285 dprintk("RPC: %s: posting %d s/g entries\n",
1286 __func__, send_wr.num_sge);
1288 if (DECR_CQCOUNT(ep) > 0)
1289 send_wr.send_flags = 0;
1290 else { /* Provider must take a send completion every now and then */
1292 send_wr.send_flags = IB_SEND_SIGNALED;
1295 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1297 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1304 * (Re)post a receive buffer.
1307 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1308 struct rpcrdma_ep *ep,
1309 struct rpcrdma_rep *rep)
1311 struct ib_recv_wr recv_wr, *recv_wr_fail;
1314 recv_wr.next = NULL;
1315 recv_wr.wr_id = (u64) (unsigned long) rep;
1316 recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1317 recv_wr.num_sge = 1;
1319 ib_dma_sync_single_for_cpu(ia->ri_device,
1320 rdmab_addr(rep->rr_rdmabuf),
1321 rdmab_length(rep->rr_rdmabuf),
1324 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1327 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,
1333 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1334 * @r_xprt: transport associated with these backchannel resources
1335 * @min_reqs: minimum number of incoming requests expected
1337 * Returns zero if all requested buffers were posted, or a negative errno.
1340 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1342 struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1343 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1344 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1345 struct rpcrdma_rep *rep;
1349 spin_lock(&buffers->rb_lock);
1350 if (list_empty(&buffers->rb_recv_bufs))
1352 rep = rpcrdma_buffer_get_rep_locked(buffers);
1353 spin_unlock(&buffers->rb_lock);
1355 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1363 spin_unlock(&buffers->rb_lock);
1364 pr_warn("%s: no extra receive buffers\n", __func__);
1368 rpcrdma_recv_buffer_put(rep);
1372 /* How many chunk list items fit within our inline buffers?
1375 rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
1377 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1378 int bytes, segments;
1380 bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
1381 bytes -= RPCRDMA_HDRLEN_MIN;
1382 if (bytes < sizeof(struct rpcrdma_segment) * 2) {
1383 pr_warn("RPC: %s: inline threshold too small\n",
1388 segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
1389 dprintk("RPC: %s: max chunk list size = %d segments\n",
1390 __func__, segments);