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(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
384 dprintk("RPC: %s: rdma_create_id() failed %i\n",
389 ia->ri_async_rc = -ETIMEDOUT;
390 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
392 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
396 wait_for_completion_interruptible_timeout(&ia->ri_done,
397 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
400 * Until xprtrdma supports DEVICE_REMOVAL, the provider must
401 * be pinned while there are active NFS/RDMA mounts to prevent
402 * hangs and crashes at umount time.
404 if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
405 dprintk("RPC: %s: Failed to get device module\n",
407 ia->ri_async_rc = -ENODEV;
409 rc = ia->ri_async_rc;
413 ia->ri_async_rc = -ETIMEDOUT;
414 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
416 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
420 wait_for_completion_interruptible_timeout(&ia->ri_done,
421 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
422 rc = ia->ri_async_rc;
428 module_put(id->device->owner);
435 * Drain any cq, prior to teardown.
438 rpcrdma_clean_cq(struct ib_cq *cq)
443 while (1 == ib_poll_cq(cq, 1, &wc))
447 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
448 __func__, count, wc.opcode);
452 * Exported functions.
456 * Open and initialize an Interface Adapter.
457 * o initializes fields of struct rpcrdma_ia, including
458 * interface and provider attributes and protection zone.
461 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
463 struct rpcrdma_ia *ia = &xprt->rx_ia;
464 struct ib_device_attr *devattr = &ia->ri_devattr;
467 ia->ri_dma_mr = NULL;
469 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
470 if (IS_ERR(ia->ri_id)) {
471 rc = PTR_ERR(ia->ri_id);
474 ia->ri_device = ia->ri_id->device;
476 ia->ri_pd = ib_alloc_pd(ia->ri_device);
477 if (IS_ERR(ia->ri_pd)) {
478 rc = PTR_ERR(ia->ri_pd);
479 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
484 rc = ib_query_device(ia->ri_device, devattr);
486 dprintk("RPC: %s: ib_query_device failed %d\n",
491 if (memreg == RPCRDMA_FRMR) {
492 if (!(devattr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) ||
493 (devattr->max_fast_reg_page_list_len == 0)) {
494 dprintk("RPC: %s: FRMR registration "
495 "not supported by HCA\n", __func__);
496 memreg = RPCRDMA_MTHCAFMR;
499 if (memreg == RPCRDMA_MTHCAFMR) {
500 if (!ia->ri_device->alloc_fmr) {
501 dprintk("RPC: %s: MTHCAFMR registration "
502 "not supported by HCA\n", __func__);
510 ia->ri_ops = &rpcrdma_frwr_memreg_ops;
512 case RPCRDMA_ALLPHYSICAL:
513 ia->ri_ops = &rpcrdma_physical_memreg_ops;
515 case RPCRDMA_MTHCAFMR:
516 ia->ri_ops = &rpcrdma_fmr_memreg_ops;
519 printk(KERN_ERR "RPC: Unsupported memory "
520 "registration mode: %d\n", memreg);
524 dprintk("RPC: %s: memory registration strategy is '%s'\n",
525 __func__, ia->ri_ops->ro_displayname);
527 rwlock_init(&ia->ri_qplock);
531 ib_dealloc_pd(ia->ri_pd);
534 rpcrdma_destroy_id(ia->ri_id);
541 * Clean up/close an IA.
542 * o if event handles and PD have been initialized, free them.
546 rpcrdma_ia_close(struct rpcrdma_ia *ia)
548 dprintk("RPC: %s: entering\n", __func__);
549 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
551 rdma_destroy_qp(ia->ri_id);
552 rpcrdma_destroy_id(ia->ri_id);
556 /* If the pd is still busy, xprtrdma missed freeing a resource */
557 if (ia->ri_pd && !IS_ERR(ia->ri_pd))
558 ib_dealloc_pd(ia->ri_pd);
562 * Create unconnected endpoint.
565 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
566 struct rpcrdma_create_data_internal *cdata)
568 struct ib_device_attr *devattr = &ia->ri_devattr;
569 struct ib_cq *sendcq, *recvcq;
570 struct ib_cq_init_attr cq_attr = {};
571 unsigned int max_qp_wr;
574 if (devattr->max_sge < RPCRDMA_MAX_IOVS) {
575 dprintk("RPC: %s: insufficient sge's available\n",
580 if (devattr->max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
581 dprintk("RPC: %s: insufficient wqe's available\n",
585 max_qp_wr = devattr->max_qp_wr - RPCRDMA_BACKWARD_WRS;
587 /* check provider's send/recv wr limits */
588 if (cdata->max_requests > max_qp_wr)
589 cdata->max_requests = max_qp_wr;
591 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
592 ep->rep_attr.qp_context = ep;
593 ep->rep_attr.srq = NULL;
594 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
595 ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
596 rc = ia->ri_ops->ro_open(ia, ep, cdata);
599 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
600 ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
601 ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
602 ep->rep_attr.cap.max_recv_sge = 1;
603 ep->rep_attr.cap.max_inline_data = 0;
604 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
605 ep->rep_attr.qp_type = IB_QPT_RC;
606 ep->rep_attr.port_num = ~0;
608 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
609 "iovs: send %d recv %d\n",
611 ep->rep_attr.cap.max_send_wr,
612 ep->rep_attr.cap.max_recv_wr,
613 ep->rep_attr.cap.max_send_sge,
614 ep->rep_attr.cap.max_recv_sge);
616 /* set trigger for requesting send completion */
617 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
618 if (ep->rep_cqinit > RPCRDMA_MAX_UNSIGNALED_SENDS)
619 ep->rep_cqinit = RPCRDMA_MAX_UNSIGNALED_SENDS;
620 else if (ep->rep_cqinit <= 2)
623 init_waitqueue_head(&ep->rep_connect_wait);
624 INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
626 cq_attr.cqe = ep->rep_attr.cap.max_send_wr + 1;
627 sendcq = ib_create_cq(ia->ri_device, rpcrdma_sendcq_upcall,
628 rpcrdma_cq_async_error_upcall, NULL, &cq_attr);
629 if (IS_ERR(sendcq)) {
630 rc = PTR_ERR(sendcq);
631 dprintk("RPC: %s: failed to create send CQ: %i\n",
636 rc = ib_req_notify_cq(sendcq, IB_CQ_NEXT_COMP);
638 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
643 cq_attr.cqe = ep->rep_attr.cap.max_recv_wr + 1;
644 recvcq = ib_create_cq(ia->ri_device, rpcrdma_recvcq_upcall,
645 rpcrdma_cq_async_error_upcall, NULL, &cq_attr);
646 if (IS_ERR(recvcq)) {
647 rc = PTR_ERR(recvcq);
648 dprintk("RPC: %s: failed to create recv CQ: %i\n",
653 rc = ib_req_notify_cq(recvcq, IB_CQ_NEXT_COMP);
655 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
657 ib_destroy_cq(recvcq);
661 ep->rep_attr.send_cq = sendcq;
662 ep->rep_attr.recv_cq = recvcq;
664 /* Initialize cma parameters */
666 /* RPC/RDMA does not use private data */
667 ep->rep_remote_cma.private_data = NULL;
668 ep->rep_remote_cma.private_data_len = 0;
670 /* Client offers RDMA Read but does not initiate */
671 ep->rep_remote_cma.initiator_depth = 0;
672 if (devattr->max_qp_rd_atom > 32) /* arbitrary but <= 255 */
673 ep->rep_remote_cma.responder_resources = 32;
675 ep->rep_remote_cma.responder_resources =
676 devattr->max_qp_rd_atom;
678 ep->rep_remote_cma.retry_count = 7;
679 ep->rep_remote_cma.flow_control = 0;
680 ep->rep_remote_cma.rnr_retry_count = 0;
685 err = ib_destroy_cq(sendcq);
687 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
691 ib_dereg_mr(ia->ri_dma_mr);
698 * Disconnect and destroy endpoint. After this, the only
699 * valid operations on the ep are to free it (if dynamically
700 * allocated) or re-create it.
703 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
707 dprintk("RPC: %s: entering, connected is %d\n",
708 __func__, ep->rep_connected);
710 cancel_delayed_work_sync(&ep->rep_connect_worker);
713 rpcrdma_ep_disconnect(ep, ia);
715 rpcrdma_clean_cq(ep->rep_attr.recv_cq);
716 rpcrdma_clean_cq(ep->rep_attr.send_cq);
719 rdma_destroy_qp(ia->ri_id);
720 ia->ri_id->qp = NULL;
723 rc = ib_destroy_cq(ep->rep_attr.recv_cq);
725 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
728 rc = ib_destroy_cq(ep->rep_attr.send_cq);
730 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
734 rc = ib_dereg_mr(ia->ri_dma_mr);
735 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
741 * Connect unconnected endpoint.
744 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
746 struct rdma_cm_id *id, *old;
750 if (ep->rep_connected != 0) {
751 struct rpcrdma_xprt *xprt;
753 dprintk("RPC: %s: reconnecting...\n", __func__);
755 rpcrdma_ep_disconnect(ep, ia);
756 rpcrdma_flush_cqs(ep);
758 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
759 id = rpcrdma_create_id(xprt, ia,
760 (struct sockaddr *)&xprt->rx_data.addr);
765 /* TEMP TEMP TEMP - fail if new device:
766 * Deregister/remarshal *all* requests!
767 * Close and recreate adapter, pd, etc!
768 * Re-determine all attributes still sane!
769 * More stuff I haven't thought of!
772 if (ia->ri_device != id->device) {
773 printk("RPC: %s: can't reconnect on "
774 "different device!\n", __func__);
775 rpcrdma_destroy_id(id);
780 rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
782 dprintk("RPC: %s: rdma_create_qp failed %i\n",
784 rpcrdma_destroy_id(id);
789 write_lock(&ia->ri_qplock);
792 write_unlock(&ia->ri_qplock);
794 rdma_destroy_qp(old);
795 rpcrdma_destroy_id(old);
797 dprintk("RPC: %s: connecting...\n", __func__);
798 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
800 dprintk("RPC: %s: rdma_create_qp failed %i\n",
802 /* do not update ep->rep_connected */
807 ep->rep_connected = 0;
809 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
811 dprintk("RPC: %s: rdma_connect() failed with %i\n",
816 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
819 * Check state. A non-peer reject indicates no listener
820 * (ECONNREFUSED), which may be a transient state. All
821 * others indicate a transport condition which has already
822 * undergone a best-effort.
824 if (ep->rep_connected == -ECONNREFUSED &&
825 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
826 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
829 if (ep->rep_connected <= 0) {
830 /* Sometimes, the only way to reliably connect to remote
831 * CMs is to use same nonzero values for ORD and IRD. */
832 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
833 (ep->rep_remote_cma.responder_resources == 0 ||
834 ep->rep_remote_cma.initiator_depth !=
835 ep->rep_remote_cma.responder_resources)) {
836 if (ep->rep_remote_cma.responder_resources == 0)
837 ep->rep_remote_cma.responder_resources = 1;
838 ep->rep_remote_cma.initiator_depth =
839 ep->rep_remote_cma.responder_resources;
842 rc = ep->rep_connected;
844 struct rpcrdma_xprt *r_xprt;
847 dprintk("RPC: %s: connected\n", __func__);
849 r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
850 extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
853 rc = rpcrdma_ep_post_extra_recv(r_xprt, extras);
855 pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
863 ep->rep_connected = rc;
868 * rpcrdma_ep_disconnect
870 * This is separate from destroy to facilitate the ability
871 * to reconnect without recreating the endpoint.
873 * This call is not reentrant, and must not be made in parallel
874 * on the same endpoint.
877 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
881 rpcrdma_flush_cqs(ep);
882 rc = rdma_disconnect(ia->ri_id);
884 /* returns without wait if not connected */
885 wait_event_interruptible(ep->rep_connect_wait,
886 ep->rep_connected != 1);
887 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
888 (ep->rep_connected == 1) ? "still " : "dis");
890 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
891 ep->rep_connected = rc;
896 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
898 struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
899 struct rpcrdma_req *req;
901 req = kzalloc(sizeof(*req), GFP_KERNEL);
903 return ERR_PTR(-ENOMEM);
905 INIT_LIST_HEAD(&req->rl_free);
906 spin_lock(&buffer->rb_reqslock);
907 list_add(&req->rl_all, &buffer->rb_allreqs);
908 spin_unlock(&buffer->rb_reqslock);
909 req->rl_buffer = &r_xprt->rx_buf;
914 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
916 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
917 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
918 struct rpcrdma_rep *rep;
922 rep = kzalloc(sizeof(*rep), GFP_KERNEL);
926 rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
928 if (IS_ERR(rep->rr_rdmabuf)) {
929 rc = PTR_ERR(rep->rr_rdmabuf);
933 rep->rr_device = ia->ri_device;
934 rep->rr_rxprt = r_xprt;
935 INIT_WORK(&rep->rr_work, rpcrdma_receive_worker);
945 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
947 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
948 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
951 buf->rb_max_requests = r_xprt->rx_data.max_requests;
952 buf->rb_bc_srv_max_requests = 0;
953 spin_lock_init(&buf->rb_lock);
955 rc = ia->ri_ops->ro_init(r_xprt);
959 INIT_LIST_HEAD(&buf->rb_send_bufs);
960 INIT_LIST_HEAD(&buf->rb_allreqs);
961 spin_lock_init(&buf->rb_reqslock);
962 for (i = 0; i < buf->rb_max_requests; i++) {
963 struct rpcrdma_req *req;
965 req = rpcrdma_create_req(r_xprt);
967 dprintk("RPC: %s: request buffer %d alloc"
968 " failed\n", __func__, i);
972 req->rl_backchannel = false;
973 list_add(&req->rl_free, &buf->rb_send_bufs);
976 INIT_LIST_HEAD(&buf->rb_recv_bufs);
977 for (i = 0; i < buf->rb_max_requests + 2; i++) {
978 struct rpcrdma_rep *rep;
980 rep = rpcrdma_create_rep(r_xprt);
982 dprintk("RPC: %s: reply buffer %d alloc failed\n",
987 list_add(&rep->rr_list, &buf->rb_recv_bufs);
992 rpcrdma_buffer_destroy(buf);
996 static struct rpcrdma_req *
997 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
999 struct rpcrdma_req *req;
1001 req = list_first_entry(&buf->rb_send_bufs,
1002 struct rpcrdma_req, rl_free);
1003 list_del(&req->rl_free);
1007 static struct rpcrdma_rep *
1008 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1010 struct rpcrdma_rep *rep;
1012 rep = list_first_entry(&buf->rb_recv_bufs,
1013 struct rpcrdma_rep, rr_list);
1014 list_del(&rep->rr_list);
1019 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
1021 rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
1026 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
1028 rpcrdma_free_regbuf(ia, req->rl_sendbuf);
1029 rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
1034 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1036 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1038 while (!list_empty(&buf->rb_recv_bufs)) {
1039 struct rpcrdma_rep *rep;
1041 rep = rpcrdma_buffer_get_rep_locked(buf);
1042 rpcrdma_destroy_rep(ia, rep);
1045 spin_lock(&buf->rb_reqslock);
1046 while (!list_empty(&buf->rb_allreqs)) {
1047 struct rpcrdma_req *req;
1049 req = list_first_entry(&buf->rb_allreqs,
1050 struct rpcrdma_req, rl_all);
1051 list_del(&req->rl_all);
1053 spin_unlock(&buf->rb_reqslock);
1054 rpcrdma_destroy_req(ia, req);
1055 spin_lock(&buf->rb_reqslock);
1057 spin_unlock(&buf->rb_reqslock);
1059 ia->ri_ops->ro_destroy(buf);
1063 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1065 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1066 struct rpcrdma_mw *mw = NULL;
1068 spin_lock(&buf->rb_mwlock);
1069 if (!list_empty(&buf->rb_mws)) {
1070 mw = list_first_entry(&buf->rb_mws,
1071 struct rpcrdma_mw, mw_list);
1072 list_del_init(&mw->mw_list);
1074 spin_unlock(&buf->rb_mwlock);
1077 pr_err("RPC: %s: no MWs available\n", __func__);
1082 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1084 struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1086 spin_lock(&buf->rb_mwlock);
1087 list_add_tail(&mw->mw_list, &buf->rb_mws);
1088 spin_unlock(&buf->rb_mwlock);
1092 * Get a set of request/reply buffers.
1094 * Reply buffer (if available) is attached to send buffer upon return.
1096 struct rpcrdma_req *
1097 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1099 struct rpcrdma_req *req;
1101 spin_lock(&buffers->rb_lock);
1102 if (list_empty(&buffers->rb_send_bufs))
1104 req = rpcrdma_buffer_get_req_locked(buffers);
1105 if (list_empty(&buffers->rb_recv_bufs))
1107 req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers);
1108 spin_unlock(&buffers->rb_lock);
1112 spin_unlock(&buffers->rb_lock);
1113 pr_warn("RPC: %s: out of request buffers\n", __func__);
1116 spin_unlock(&buffers->rb_lock);
1117 pr_warn("RPC: %s: out of reply buffers\n", __func__);
1118 req->rl_reply = NULL;
1123 * Put request/reply buffers back into pool.
1124 * Pre-decrement counter/array index.
1127 rpcrdma_buffer_put(struct rpcrdma_req *req)
1129 struct rpcrdma_buffer *buffers = req->rl_buffer;
1130 struct rpcrdma_rep *rep = req->rl_reply;
1133 req->rl_reply = NULL;
1135 spin_lock(&buffers->rb_lock);
1136 list_add_tail(&req->rl_free, &buffers->rb_send_bufs);
1138 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1139 spin_unlock(&buffers->rb_lock);
1143 * Recover reply buffers from pool.
1144 * This happens when recovering from disconnect.
1147 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1149 struct rpcrdma_buffer *buffers = req->rl_buffer;
1151 spin_lock(&buffers->rb_lock);
1152 if (!list_empty(&buffers->rb_recv_bufs))
1153 req->rl_reply = rpcrdma_buffer_get_rep_locked(buffers);
1154 spin_unlock(&buffers->rb_lock);
1158 * Put reply buffers back into pool when not attached to
1159 * request. This happens in error conditions.
1162 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1164 struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1166 spin_lock(&buffers->rb_lock);
1167 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1168 spin_unlock(&buffers->rb_lock);
1172 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1176 rpcrdma_mapping_error(struct rpcrdma_mr_seg *seg)
1178 dprintk("RPC: map_one: offset %p iova %llx len %zu\n",
1180 (unsigned long long)seg->mr_dma, seg->mr_dmalen);
1184 * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1185 * @ia: controlling rpcrdma_ia
1186 * @size: size of buffer to be allocated, in bytes
1189 * Returns pointer to private header of an area of internally
1190 * registered memory, or an ERR_PTR. The registered buffer follows
1191 * the end of the private header.
1193 * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1194 * receiving the payload of RDMA RECV operations. regbufs are not
1195 * used for RDMA READ/WRITE operations, thus are registered only for
1198 struct rpcrdma_regbuf *
1199 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1201 struct rpcrdma_regbuf *rb;
1204 rb = kmalloc(sizeof(*rb) + size, flags);
1209 iov->addr = ib_dma_map_single(ia->ri_device,
1210 (void *)rb->rg_base, size,
1212 if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1216 iov->lkey = ia->ri_pd->local_dma_lkey;
1218 rb->rg_owner = NULL;
1224 return ERR_PTR(-ENOMEM);
1228 * rpcrdma_free_regbuf - deregister and free registered buffer
1229 * @ia: controlling rpcrdma_ia
1230 * @rb: regbuf to be deregistered and freed
1233 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1241 ib_dma_unmap_single(ia->ri_device,
1242 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1247 * Prepost any receive buffer, then post send.
1249 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1252 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1253 struct rpcrdma_ep *ep,
1254 struct rpcrdma_req *req)
1256 struct ib_device *device = ia->ri_device;
1257 struct ib_send_wr send_wr, *send_wr_fail;
1258 struct rpcrdma_rep *rep = req->rl_reply;
1259 struct ib_sge *iov = req->rl_send_iov;
1263 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1266 req->rl_reply = NULL;
1269 send_wr.next = NULL;
1270 send_wr.wr_id = RPCRDMA_IGNORE_COMPLETION;
1271 send_wr.sg_list = iov;
1272 send_wr.num_sge = req->rl_niovs;
1273 send_wr.opcode = IB_WR_SEND;
1275 for (i = 0; i < send_wr.num_sge; i++)
1276 ib_dma_sync_single_for_device(device, iov[i].addr,
1277 iov[i].length, DMA_TO_DEVICE);
1278 dprintk("RPC: %s: posting %d s/g entries\n",
1279 __func__, send_wr.num_sge);
1281 if (DECR_CQCOUNT(ep) > 0)
1282 send_wr.send_flags = 0;
1283 else { /* Provider must take a send completion every now and then */
1285 send_wr.send_flags = IB_SEND_SIGNALED;
1288 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1290 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1297 * (Re)post a receive buffer.
1300 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1301 struct rpcrdma_ep *ep,
1302 struct rpcrdma_rep *rep)
1304 struct ib_recv_wr recv_wr, *recv_wr_fail;
1307 recv_wr.next = NULL;
1308 recv_wr.wr_id = (u64) (unsigned long) rep;
1309 recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1310 recv_wr.num_sge = 1;
1312 ib_dma_sync_single_for_cpu(ia->ri_device,
1313 rdmab_addr(rep->rr_rdmabuf),
1314 rdmab_length(rep->rr_rdmabuf),
1317 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1320 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,
1326 * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1327 * @r_xprt: transport associated with these backchannel resources
1328 * @min_reqs: minimum number of incoming requests expected
1330 * Returns zero if all requested buffers were posted, or a negative errno.
1333 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1335 struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1336 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1337 struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1338 struct rpcrdma_rep *rep;
1339 unsigned long flags;
1343 spin_lock_irqsave(&buffers->rb_lock, flags);
1344 if (list_empty(&buffers->rb_recv_bufs))
1346 rep = rpcrdma_buffer_get_rep_locked(buffers);
1347 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1349 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1357 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1358 pr_warn("%s: no extra receive buffers\n", __func__);
1362 rpcrdma_recv_buffer_put(rep);
1366 /* How many chunk list items fit within our inline buffers?
1369 rpcrdma_max_segments(struct rpcrdma_xprt *r_xprt)
1371 struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1372 int bytes, segments;
1374 bytes = min_t(unsigned int, cdata->inline_wsize, cdata->inline_rsize);
1375 bytes -= RPCRDMA_HDRLEN_MIN;
1376 if (bytes < sizeof(struct rpcrdma_segment) * 2) {
1377 pr_warn("RPC: %s: inline threshold too small\n",
1382 segments = 1 << (fls(bytes / sizeof(struct rpcrdma_segment)) - 1);
1383 dprintk("RPC: %s: max chunk list size = %d segments\n",
1384 __func__, segments);