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/pci.h> /* for Tavor hack below */
52 #include "xprt_rdma.h"
59 # define RPCDBG_FACILITY RPCDBG_TRANS
67 * handle replies in tasklet context, using a single, global list
68 * rdma tasklet function -- just turn around and call the func
69 * for all replies on the list
72 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
73 static LIST_HEAD(rpcrdma_tasklets_g);
76 rpcrdma_run_tasklet(unsigned long data)
78 struct rpcrdma_rep *rep;
79 void (*func)(struct rpcrdma_rep *);
83 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
84 while (!list_empty(&rpcrdma_tasklets_g)) {
85 rep = list_entry(rpcrdma_tasklets_g.next,
86 struct rpcrdma_rep, rr_list);
87 list_del(&rep->rr_list);
90 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
95 rpcrdma_recv_buffer_put(rep);
97 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
99 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
102 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
105 rpcrdma_schedule_tasklet(struct rpcrdma_rep *rep)
109 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
110 list_add_tail(&rep->rr_list, &rpcrdma_tasklets_g);
111 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
112 tasklet_schedule(&rpcrdma_tasklet_g);
116 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
118 struct rpcrdma_ep *ep = context;
120 dprintk("RPC: %s: QP error %X on device %s ep %p\n",
121 __func__, event->event, event->device->name, context);
122 if (ep->rep_connected == 1) {
123 ep->rep_connected = -EIO;
125 wake_up_all(&ep->rep_connect_wait);
130 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
132 struct rpcrdma_ep *ep = context;
134 dprintk("RPC: %s: CQ error %X on device %s ep %p\n",
135 __func__, event->event, event->device->name, context);
136 if (ep->rep_connected == 1) {
137 ep->rep_connected = -EIO;
139 wake_up_all(&ep->rep_connect_wait);
144 void rpcrdma_event_process(struct ib_wc *wc)
146 struct rpcrdma_rep *rep =
147 (struct rpcrdma_rep *)(unsigned long) wc->wr_id;
149 dprintk("RPC: %s: event rep %p status %X opcode %X length %u\n",
150 __func__, rep, wc->status, wc->opcode, wc->byte_len);
152 if (!rep) /* send or bind completion that we don't care about */
155 if (IB_WC_SUCCESS != wc->status) {
156 dprintk("RPC: %s: %s WC status %X, connection lost\n",
157 __func__, (wc->opcode & IB_WC_RECV) ? "recv" : "send",
160 rpcrdma_schedule_tasklet(rep);
164 switch (wc->opcode) {
166 rep->rr_len = wc->byte_len;
167 ib_dma_sync_single_for_cpu(
168 rdmab_to_ia(rep->rr_buffer)->ri_id->device,
169 rep->rr_iov.addr, rep->rr_len, DMA_FROM_DEVICE);
170 /* Keep (only) the most recent credits, after check validity */
171 if (rep->rr_len >= 16) {
172 struct rpcrdma_msg *p =
173 (struct rpcrdma_msg *) rep->rr_base;
174 unsigned int credits = ntohl(p->rm_credit);
176 dprintk("RPC: %s: server"
177 " dropped credits to 0!\n", __func__);
180 } else if (credits > rep->rr_buffer->rb_max_requests) {
181 dprintk("RPC: %s: server"
182 " over-crediting: %d (%d)\n",
184 rep->rr_buffer->rb_max_requests);
185 credits = rep->rr_buffer->rb_max_requests;
187 atomic_set(&rep->rr_buffer->rb_credits, credits);
191 rpcrdma_schedule_tasklet(rep);
194 dprintk("RPC: %s: unexpected WC event %X\n",
195 __func__, wc->opcode);
201 rpcrdma_cq_poll(struct ib_cq *cq)
207 rc = ib_poll_cq(cq, 1, &wc);
209 dprintk("RPC: %s: ib_poll_cq failed %i\n",
216 rpcrdma_event_process(&wc);
223 * rpcrdma_cq_event_upcall
225 * This upcall handles recv, send, bind and unbind events.
226 * It is reentrant but processes single events in order to maintain
227 * ordering of receives to keep server credits.
229 * It is the responsibility of the scheduled tasklet to return
230 * recv buffers to the pool. NOTE: this affects synchronization of
231 * connection shutdown. That is, the structures required for
232 * the completion of the reply handler must remain intact until
233 * all memory has been reclaimed.
235 * Note that send events are suppressed and do not result in an upcall.
238 rpcrdma_cq_event_upcall(struct ib_cq *cq, void *context)
242 rc = rpcrdma_cq_poll(cq);
246 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
248 dprintk("RPC: %s: ib_req_notify_cq failed %i\n",
257 static const char * const conn[] = {
274 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
276 struct rpcrdma_xprt *xprt = id->context;
277 struct rpcrdma_ia *ia = &xprt->rx_ia;
278 struct rpcrdma_ep *ep = &xprt->rx_ep;
279 struct sockaddr_in *addr = (struct sockaddr_in *) &ep->rep_remote_addr;
280 struct ib_qp_attr attr;
281 struct ib_qp_init_attr iattr;
284 switch (event->event) {
285 case RDMA_CM_EVENT_ADDR_RESOLVED:
286 case RDMA_CM_EVENT_ROUTE_RESOLVED:
287 complete(&ia->ri_done);
289 case RDMA_CM_EVENT_ADDR_ERROR:
290 ia->ri_async_rc = -EHOSTUNREACH;
291 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
293 complete(&ia->ri_done);
295 case RDMA_CM_EVENT_ROUTE_ERROR:
296 ia->ri_async_rc = -ENETUNREACH;
297 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
299 complete(&ia->ri_done);
301 case RDMA_CM_EVENT_ESTABLISHED:
303 ib_query_qp(ia->ri_id->qp, &attr,
304 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
306 dprintk("RPC: %s: %d responder resources"
308 __func__, attr.max_dest_rd_atomic, attr.max_rd_atomic);
310 case RDMA_CM_EVENT_CONNECT_ERROR:
311 connstate = -ENOTCONN;
313 case RDMA_CM_EVENT_UNREACHABLE:
314 connstate = -ENETDOWN;
316 case RDMA_CM_EVENT_REJECTED:
317 connstate = -ECONNREFUSED;
319 case RDMA_CM_EVENT_DISCONNECTED:
320 connstate = -ECONNABORTED;
322 case RDMA_CM_EVENT_DEVICE_REMOVAL:
325 dprintk("RPC: %s: %s: %u.%u.%u.%u:%u"
326 " (ep 0x%p event 0x%x)\n",
328 (event->event <= 11) ? conn[event->event] :
329 "unknown connection error",
330 NIPQUAD(addr->sin_addr.s_addr),
331 ntohs(addr->sin_port),
333 atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
334 dprintk("RPC: %s: %sconnected\n",
335 __func__, connstate > 0 ? "" : "dis");
336 ep->rep_connected = connstate;
338 wake_up_all(&ep->rep_connect_wait);
341 ia->ri_async_rc = -EINVAL;
342 dprintk("RPC: %s: unexpected CM event %X\n",
343 __func__, event->event);
344 complete(&ia->ri_done);
351 static struct rdma_cm_id *
352 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
353 struct rpcrdma_ia *ia, struct sockaddr *addr)
355 struct rdma_cm_id *id;
358 id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP);
361 dprintk("RPC: %s: rdma_create_id() failed %i\n",
367 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
369 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
373 wait_for_completion(&ia->ri_done);
374 rc = ia->ri_async_rc;
379 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
381 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
385 wait_for_completion(&ia->ri_done);
386 rc = ia->ri_async_rc;
398 * Drain any cq, prior to teardown.
401 rpcrdma_clean_cq(struct ib_cq *cq)
406 while (1 == ib_poll_cq(cq, 1, &wc))
410 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
411 __func__, count, wc.opcode);
415 * Exported functions.
419 * Open and initialize an Interface Adapter.
420 * o initializes fields of struct rpcrdma_ia, including
421 * interface and provider attributes and protection zone.
424 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
427 struct ib_device_attr devattr;
428 struct rpcrdma_ia *ia = &xprt->rx_ia;
430 init_completion(&ia->ri_done);
432 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
433 if (IS_ERR(ia->ri_id)) {
434 rc = PTR_ERR(ia->ri_id);
438 ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
439 if (IS_ERR(ia->ri_pd)) {
440 rc = PTR_ERR(ia->ri_pd);
441 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
447 * Query the device to determine if the requested memory
448 * registration strategy is supported. If it isn't, set the
449 * strategy to a globally supported model.
451 rc = ib_query_device(ia->ri_id->device, &devattr);
453 dprintk("RPC: %s: ib_query_device failed %d\n",
458 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
459 ia->ri_have_dma_lkey = 1;
460 ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
464 case RPCRDMA_MEMWINDOWS:
465 case RPCRDMA_MEMWINDOWS_ASYNC:
466 if (!(devattr.device_cap_flags & IB_DEVICE_MEM_WINDOW)) {
467 dprintk("RPC: %s: MEMWINDOWS registration "
468 "specified but not supported by adapter, "
469 "using slower RPCRDMA_REGISTER\n",
471 memreg = RPCRDMA_REGISTER;
474 case RPCRDMA_MTHCAFMR:
475 if (!ia->ri_id->device->alloc_fmr) {
476 #if RPCRDMA_PERSISTENT_REGISTRATION
477 dprintk("RPC: %s: MTHCAFMR registration "
478 "specified but not supported by adapter, "
479 "using riskier RPCRDMA_ALLPHYSICAL\n",
481 memreg = RPCRDMA_ALLPHYSICAL;
483 dprintk("RPC: %s: MTHCAFMR registration "
484 "specified but not supported by adapter, "
485 "using slower RPCRDMA_REGISTER\n",
487 memreg = RPCRDMA_REGISTER;
492 /* Requires both frmr reg and local dma lkey */
493 if ((devattr.device_cap_flags &
494 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
495 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) {
496 #if RPCRDMA_PERSISTENT_REGISTRATION
497 dprintk("RPC: %s: FRMR registration "
498 "specified but not supported by adapter, "
499 "using riskier RPCRDMA_ALLPHYSICAL\n",
501 memreg = RPCRDMA_ALLPHYSICAL;
503 dprintk("RPC: %s: FRMR registration "
504 "specified but not supported by adapter, "
505 "using slower RPCRDMA_REGISTER\n",
507 memreg = RPCRDMA_REGISTER;
514 * Optionally obtain an underlying physical identity mapping in
515 * order to do a memory window-based bind. This base registration
516 * is protected from remote access - that is enabled only by binding
517 * for the specific bytes targeted during each RPC operation, and
518 * revoked after the corresponding completion similar to a storage
522 case RPCRDMA_BOUNCEBUFFERS:
523 case RPCRDMA_REGISTER:
526 #if RPCRDMA_PERSISTENT_REGISTRATION
527 case RPCRDMA_ALLPHYSICAL:
528 mem_priv = IB_ACCESS_LOCAL_WRITE |
529 IB_ACCESS_REMOTE_WRITE |
530 IB_ACCESS_REMOTE_READ;
533 case RPCRDMA_MEMWINDOWS_ASYNC:
534 case RPCRDMA_MEMWINDOWS:
535 mem_priv = IB_ACCESS_LOCAL_WRITE |
538 case RPCRDMA_MTHCAFMR:
539 if (ia->ri_have_dma_lkey)
541 mem_priv = IB_ACCESS_LOCAL_WRITE;
543 ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
544 if (IS_ERR(ia->ri_bind_mem)) {
545 printk(KERN_ALERT "%s: ib_get_dma_mr for "
546 "phys register failed with %lX\n\t"
547 "Will continue with degraded performance\n",
548 __func__, PTR_ERR(ia->ri_bind_mem));
549 memreg = RPCRDMA_REGISTER;
550 ia->ri_bind_mem = NULL;
554 printk(KERN_ERR "%s: invalid memory registration mode %d\n",
559 dprintk("RPC: %s: memory registration strategy is %d\n",
562 /* Else will do memory reg/dereg for each chunk */
563 ia->ri_memreg_strategy = memreg;
567 rdma_destroy_id(ia->ri_id);
574 * Clean up/close an IA.
575 * o if event handles and PD have been initialized, free them.
579 rpcrdma_ia_close(struct rpcrdma_ia *ia)
583 dprintk("RPC: %s: entering\n", __func__);
584 if (ia->ri_bind_mem != NULL) {
585 rc = ib_dereg_mr(ia->ri_bind_mem);
586 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
589 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
591 rdma_destroy_qp(ia->ri_id);
592 rdma_destroy_id(ia->ri_id);
595 if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
596 rc = ib_dealloc_pd(ia->ri_pd);
597 dprintk("RPC: %s: ib_dealloc_pd returned %i\n",
603 * Create unconnected endpoint.
606 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
607 struct rpcrdma_create_data_internal *cdata)
609 struct ib_device_attr devattr;
612 rc = ib_query_device(ia->ri_id->device, &devattr);
614 dprintk("RPC: %s: ib_query_device failed %d\n",
619 /* check provider's send/recv wr limits */
620 if (cdata->max_requests > devattr.max_qp_wr)
621 cdata->max_requests = devattr.max_qp_wr;
623 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
624 ep->rep_attr.qp_context = ep;
625 /* send_cq and recv_cq initialized below */
626 ep->rep_attr.srq = NULL;
627 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
628 switch (ia->ri_memreg_strategy) {
630 /* Add room for frmr register and invalidate WRs */
631 ep->rep_attr.cap.max_send_wr *= 3;
632 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
635 case RPCRDMA_MEMWINDOWS_ASYNC:
636 case RPCRDMA_MEMWINDOWS:
637 /* Add room for mw_binds+unbinds - overkill! */
638 ep->rep_attr.cap.max_send_wr++;
639 ep->rep_attr.cap.max_send_wr *= (2 * RPCRDMA_MAX_SEGS);
640 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
646 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
647 ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
648 ep->rep_attr.cap.max_recv_sge = 1;
649 ep->rep_attr.cap.max_inline_data = 0;
650 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
651 ep->rep_attr.qp_type = IB_QPT_RC;
652 ep->rep_attr.port_num = ~0;
654 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
655 "iovs: send %d recv %d\n",
657 ep->rep_attr.cap.max_send_wr,
658 ep->rep_attr.cap.max_recv_wr,
659 ep->rep_attr.cap.max_send_sge,
660 ep->rep_attr.cap.max_recv_sge);
662 /* set trigger for requesting send completion */
663 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 /* - 1*/;
664 switch (ia->ri_memreg_strategy) {
665 case RPCRDMA_MEMWINDOWS_ASYNC:
666 case RPCRDMA_MEMWINDOWS:
667 ep->rep_cqinit -= RPCRDMA_MAX_SEGS;
672 if (ep->rep_cqinit <= 2)
676 init_waitqueue_head(&ep->rep_connect_wait);
679 * Create a single cq for receive dto and mw_bind (only ever
680 * care about unbind, really). Send completions are suppressed.
681 * Use single threaded tasklet upcalls to maintain ordering.
683 ep->rep_cq = ib_create_cq(ia->ri_id->device, rpcrdma_cq_event_upcall,
684 rpcrdma_cq_async_error_upcall, NULL,
685 ep->rep_attr.cap.max_recv_wr +
686 ep->rep_attr.cap.max_send_wr + 1, 0);
687 if (IS_ERR(ep->rep_cq)) {
688 rc = PTR_ERR(ep->rep_cq);
689 dprintk("RPC: %s: ib_create_cq failed: %i\n",
694 rc = ib_req_notify_cq(ep->rep_cq, IB_CQ_NEXT_COMP);
696 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
701 ep->rep_attr.send_cq = ep->rep_cq;
702 ep->rep_attr.recv_cq = ep->rep_cq;
704 /* Initialize cma parameters */
706 /* RPC/RDMA does not use private data */
707 ep->rep_remote_cma.private_data = NULL;
708 ep->rep_remote_cma.private_data_len = 0;
710 /* Client offers RDMA Read but does not initiate */
711 ep->rep_remote_cma.initiator_depth = 0;
712 if (ia->ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS)
713 ep->rep_remote_cma.responder_resources = 0;
714 else if (devattr.max_qp_rd_atom > 32) /* arbitrary but <= 255 */
715 ep->rep_remote_cma.responder_resources = 32;
717 ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;
719 ep->rep_remote_cma.retry_count = 7;
720 ep->rep_remote_cma.flow_control = 0;
721 ep->rep_remote_cma.rnr_retry_count = 0;
726 err = ib_destroy_cq(ep->rep_cq);
728 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
737 * Disconnect and destroy endpoint. After this, the only
738 * valid operations on the ep are to free it (if dynamically
739 * allocated) or re-create it.
741 * The caller's error handling must be sure to not leak the endpoint
742 * if this function fails.
745 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
749 dprintk("RPC: %s: entering, connected is %d\n",
750 __func__, ep->rep_connected);
753 rc = rpcrdma_ep_disconnect(ep, ia);
755 dprintk("RPC: %s: rpcrdma_ep_disconnect"
756 " returned %i\n", __func__, rc);
757 rdma_destroy_qp(ia->ri_id);
758 ia->ri_id->qp = NULL;
761 /* padding - could be done in rpcrdma_buffer_destroy... */
762 if (ep->rep_pad_mr) {
763 rpcrdma_deregister_internal(ia, ep->rep_pad_mr, &ep->rep_pad);
764 ep->rep_pad_mr = NULL;
767 rpcrdma_clean_cq(ep->rep_cq);
768 rc = ib_destroy_cq(ep->rep_cq);
770 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
777 * Connect unconnected endpoint.
780 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
782 struct rdma_cm_id *id;
785 int reconnect = (ep->rep_connected != 0);
788 struct rpcrdma_xprt *xprt;
790 rc = rpcrdma_ep_disconnect(ep, ia);
791 if (rc && rc != -ENOTCONN)
792 dprintk("RPC: %s: rpcrdma_ep_disconnect"
793 " status %i\n", __func__, rc);
794 rpcrdma_clean_cq(ep->rep_cq);
796 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
797 id = rpcrdma_create_id(xprt, ia,
798 (struct sockaddr *)&xprt->rx_data.addr);
803 /* TEMP TEMP TEMP - fail if new device:
804 * Deregister/remarshal *all* requests!
805 * Close and recreate adapter, pd, etc!
806 * Re-determine all attributes still sane!
807 * More stuff I haven't thought of!
810 if (ia->ri_id->device != id->device) {
811 printk("RPC: %s: can't reconnect on "
812 "different device!\n", __func__);
818 rdma_destroy_id(ia->ri_id);
822 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
824 dprintk("RPC: %s: rdma_create_qp failed %i\n",
829 /* XXX Tavor device performs badly with 2K MTU! */
830 if (strnicmp(ia->ri_id->device->dma_device->bus->name, "pci", 3) == 0) {
831 struct pci_dev *pcid = to_pci_dev(ia->ri_id->device->dma_device);
832 if (pcid->device == PCI_DEVICE_ID_MELLANOX_TAVOR &&
833 (pcid->vendor == PCI_VENDOR_ID_MELLANOX ||
834 pcid->vendor == PCI_VENDOR_ID_TOPSPIN)) {
835 struct ib_qp_attr attr = {
836 .path_mtu = IB_MTU_1024
838 rc = ib_modify_qp(ia->ri_id->qp, &attr, IB_QP_PATH_MTU);
842 ep->rep_connected = 0;
844 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
846 dprintk("RPC: %s: rdma_connect() failed with %i\n",
854 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
857 * Check state. A non-peer reject indicates no listener
858 * (ECONNREFUSED), which may be a transient state. All
859 * others indicate a transport condition which has already
860 * undergone a best-effort.
862 if (ep->rep_connected == -ECONNREFUSED
863 && ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
864 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
867 if (ep->rep_connected <= 0) {
868 /* Sometimes, the only way to reliably connect to remote
869 * CMs is to use same nonzero values for ORD and IRD. */
870 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
871 (ep->rep_remote_cma.responder_resources == 0 ||
872 ep->rep_remote_cma.initiator_depth !=
873 ep->rep_remote_cma.responder_resources)) {
874 if (ep->rep_remote_cma.responder_resources == 0)
875 ep->rep_remote_cma.responder_resources = 1;
876 ep->rep_remote_cma.initiator_depth =
877 ep->rep_remote_cma.responder_resources;
880 rc = ep->rep_connected;
882 dprintk("RPC: %s: connected\n", __func__);
887 ep->rep_connected = rc;
892 * rpcrdma_ep_disconnect
894 * This is separate from destroy to facilitate the ability
895 * to reconnect without recreating the endpoint.
897 * This call is not reentrant, and must not be made in parallel
898 * on the same endpoint.
901 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
905 rpcrdma_clean_cq(ep->rep_cq);
906 rc = rdma_disconnect(ia->ri_id);
908 /* returns without wait if not connected */
909 wait_event_interruptible(ep->rep_connect_wait,
910 ep->rep_connected != 1);
911 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
912 (ep->rep_connected == 1) ? "still " : "dis");
914 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
915 ep->rep_connected = rc;
921 * Initialize buffer memory
924 rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
925 struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
930 struct rpcrdma_mw *r;
932 buf->rb_max_requests = cdata->max_requests;
933 spin_lock_init(&buf->rb_lock);
934 atomic_set(&buf->rb_credits, 1);
937 * 1. arrays for send and recv pointers
938 * 2. arrays of struct rpcrdma_req to fill in pointers
939 * 3. array of struct rpcrdma_rep for replies
941 * 5. mw's, fmr's or frmr's, if any
942 * Send/recv buffers in req/rep need to be registered
945 len = buf->rb_max_requests *
946 (sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
947 len += cdata->padding;
948 switch (ia->ri_memreg_strategy) {
950 len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
951 sizeof(struct rpcrdma_mw);
953 case RPCRDMA_MTHCAFMR:
954 /* TBD we are perhaps overallocating here */
955 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
956 sizeof(struct rpcrdma_mw);
958 case RPCRDMA_MEMWINDOWS_ASYNC:
959 case RPCRDMA_MEMWINDOWS:
960 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
961 sizeof(struct rpcrdma_mw);
967 /* allocate 1, 4 and 5 in one shot */
968 p = kzalloc(len, GFP_KERNEL);
970 dprintk("RPC: %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
975 buf->rb_pool = p; /* for freeing it later */
977 buf->rb_send_bufs = (struct rpcrdma_req **) p;
978 p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
979 buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
980 p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
983 * Register the zeroed pad buffer, if any.
985 if (cdata->padding) {
986 rc = rpcrdma_register_internal(ia, p, cdata->padding,
987 &ep->rep_pad_mr, &ep->rep_pad);
994 * Allocate the fmr's, or mw's for mw_bind chunk registration.
995 * We "cycle" the mw's in order to minimize rkey reuse,
996 * and also reduce unbind-to-bind collision.
998 INIT_LIST_HEAD(&buf->rb_mws);
999 r = (struct rpcrdma_mw *)p;
1000 switch (ia->ri_memreg_strategy) {
1002 for (i = buf->rb_max_requests * RPCRDMA_MAX_SEGS; i; i--) {
1003 r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
1005 if (IS_ERR(r->r.frmr.fr_mr)) {
1006 rc = PTR_ERR(r->r.frmr.fr_mr);
1007 dprintk("RPC: %s: ib_alloc_fast_reg_mr"
1008 " failed %i\n", __func__, rc);
1012 ib_alloc_fast_reg_page_list(ia->ri_id->device,
1014 if (IS_ERR(r->r.frmr.fr_pgl)) {
1015 rc = PTR_ERR(r->r.frmr.fr_pgl);
1017 "ib_alloc_fast_reg_page_list "
1018 "failed %i\n", __func__, rc);
1021 list_add(&r->mw_list, &buf->rb_mws);
1025 case RPCRDMA_MTHCAFMR:
1026 /* TBD we are perhaps overallocating here */
1027 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1028 static struct ib_fmr_attr fa =
1029 { RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT };
1030 r->r.fmr = ib_alloc_fmr(ia->ri_pd,
1031 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
1033 if (IS_ERR(r->r.fmr)) {
1034 rc = PTR_ERR(r->r.fmr);
1035 dprintk("RPC: %s: ib_alloc_fmr"
1036 " failed %i\n", __func__, rc);
1039 list_add(&r->mw_list, &buf->rb_mws);
1043 case RPCRDMA_MEMWINDOWS_ASYNC:
1044 case RPCRDMA_MEMWINDOWS:
1045 /* Allocate one extra request's worth, for full cycling */
1046 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1047 r->r.mw = ib_alloc_mw(ia->ri_pd);
1048 if (IS_ERR(r->r.mw)) {
1049 rc = PTR_ERR(r->r.mw);
1050 dprintk("RPC: %s: ib_alloc_mw"
1051 " failed %i\n", __func__, rc);
1054 list_add(&r->mw_list, &buf->rb_mws);
1063 * Allocate/init the request/reply buffers. Doing this
1064 * using kmalloc for now -- one for each buf.
1066 for (i = 0; i < buf->rb_max_requests; i++) {
1067 struct rpcrdma_req *req;
1068 struct rpcrdma_rep *rep;
1070 len = cdata->inline_wsize + sizeof(struct rpcrdma_req);
1071 /* RPC layer requests *double* size + 1K RPC_SLACK_SPACE! */
1072 /* Typical ~2400b, so rounding up saves work later */
1075 req = kmalloc(len, GFP_KERNEL);
1077 dprintk("RPC: %s: request buffer %d alloc"
1078 " failed\n", __func__, i);
1082 memset(req, 0, sizeof(struct rpcrdma_req));
1083 buf->rb_send_bufs[i] = req;
1084 buf->rb_send_bufs[i]->rl_buffer = buf;
1086 rc = rpcrdma_register_internal(ia, req->rl_base,
1087 len - offsetof(struct rpcrdma_req, rl_base),
1088 &buf->rb_send_bufs[i]->rl_handle,
1089 &buf->rb_send_bufs[i]->rl_iov);
1093 buf->rb_send_bufs[i]->rl_size = len-sizeof(struct rpcrdma_req);
1095 len = cdata->inline_rsize + sizeof(struct rpcrdma_rep);
1096 rep = kmalloc(len, GFP_KERNEL);
1098 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1103 memset(rep, 0, sizeof(struct rpcrdma_rep));
1104 buf->rb_recv_bufs[i] = rep;
1105 buf->rb_recv_bufs[i]->rr_buffer = buf;
1106 init_waitqueue_head(&rep->rr_unbind);
1108 rc = rpcrdma_register_internal(ia, rep->rr_base,
1109 len - offsetof(struct rpcrdma_rep, rr_base),
1110 &buf->rb_recv_bufs[i]->rr_handle,
1111 &buf->rb_recv_bufs[i]->rr_iov);
1116 dprintk("RPC: %s: max_requests %d\n",
1117 __func__, buf->rb_max_requests);
1121 rpcrdma_buffer_destroy(buf);
1126 * Unregister and destroy buffer memory. Need to deal with
1127 * partial initialization, so it's callable from failed create.
1128 * Must be called before destroying endpoint, as registrations
1132 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1135 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1136 struct rpcrdma_mw *r;
1138 /* clean up in reverse order from create
1139 * 1. recv mr memory (mr free, then kfree)
1140 * 1a. bind mw memory
1141 * 2. send mr memory (mr free, then kfree)
1142 * 3. padding (if any) [moved to rpcrdma_ep_destroy]
1145 dprintk("RPC: %s: entering\n", __func__);
1147 for (i = 0; i < buf->rb_max_requests; i++) {
1148 if (buf->rb_recv_bufs && buf->rb_recv_bufs[i]) {
1149 rpcrdma_deregister_internal(ia,
1150 buf->rb_recv_bufs[i]->rr_handle,
1151 &buf->rb_recv_bufs[i]->rr_iov);
1152 kfree(buf->rb_recv_bufs[i]);
1154 if (buf->rb_send_bufs && buf->rb_send_bufs[i]) {
1155 while (!list_empty(&buf->rb_mws)) {
1156 r = list_entry(buf->rb_mws.next,
1157 struct rpcrdma_mw, mw_list);
1158 list_del(&r->mw_list);
1159 switch (ia->ri_memreg_strategy) {
1161 rc = ib_dereg_mr(r->r.frmr.fr_mr);
1167 ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
1169 case RPCRDMA_MTHCAFMR:
1170 rc = ib_dealloc_fmr(r->r.fmr);
1177 case RPCRDMA_MEMWINDOWS_ASYNC:
1178 case RPCRDMA_MEMWINDOWS:
1179 rc = ib_dealloc_mw(r->r.mw);
1190 rpcrdma_deregister_internal(ia,
1191 buf->rb_send_bufs[i]->rl_handle,
1192 &buf->rb_send_bufs[i]->rl_iov);
1193 kfree(buf->rb_send_bufs[i]);
1197 kfree(buf->rb_pool);
1201 * Get a set of request/reply buffers.
1203 * Reply buffer (if needed) is attached to send buffer upon return.
1205 * rb_send_index and rb_recv_index MUST always be pointing to the
1206 * *next* available buffer (non-NULL). They are incremented after
1207 * removing buffers, and decremented *before* returning them.
1209 struct rpcrdma_req *
1210 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1212 struct rpcrdma_req *req;
1213 unsigned long flags;
1215 struct rpcrdma_mw *r;
1217 spin_lock_irqsave(&buffers->rb_lock, flags);
1218 if (buffers->rb_send_index == buffers->rb_max_requests) {
1219 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1220 dprintk("RPC: %s: out of request buffers\n", __func__);
1221 return ((struct rpcrdma_req *)NULL);
1224 req = buffers->rb_send_bufs[buffers->rb_send_index];
1225 if (buffers->rb_send_index < buffers->rb_recv_index) {
1226 dprintk("RPC: %s: %d extra receives outstanding (ok)\n",
1228 buffers->rb_recv_index - buffers->rb_send_index);
1229 req->rl_reply = NULL;
1231 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1232 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1234 buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1235 if (!list_empty(&buffers->rb_mws)) {
1236 i = RPCRDMA_MAX_SEGS - 1;
1238 r = list_entry(buffers->rb_mws.next,
1239 struct rpcrdma_mw, mw_list);
1240 list_del(&r->mw_list);
1241 req->rl_segments[i].mr_chunk.rl_mw = r;
1244 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1249 * Put request/reply buffers back into pool.
1250 * Pre-decrement counter/array index.
1253 rpcrdma_buffer_put(struct rpcrdma_req *req)
1255 struct rpcrdma_buffer *buffers = req->rl_buffer;
1256 struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1258 unsigned long flags;
1260 BUG_ON(req->rl_nchunks != 0);
1261 spin_lock_irqsave(&buffers->rb_lock, flags);
1262 buffers->rb_send_bufs[--buffers->rb_send_index] = req;
1264 if (req->rl_reply) {
1265 buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
1266 init_waitqueue_head(&req->rl_reply->rr_unbind);
1267 req->rl_reply->rr_func = NULL;
1268 req->rl_reply = NULL;
1270 switch (ia->ri_memreg_strategy) {
1272 case RPCRDMA_MTHCAFMR:
1273 case RPCRDMA_MEMWINDOWS_ASYNC:
1274 case RPCRDMA_MEMWINDOWS:
1276 * Cycle mw's back in reverse order, and "spin" them.
1277 * This delays and scrambles reuse as much as possible.
1281 struct rpcrdma_mw **mw;
1282 mw = &req->rl_segments[i].mr_chunk.rl_mw;
1283 list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
1285 } while (++i < RPCRDMA_MAX_SEGS);
1286 list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
1288 req->rl_segments[0].mr_chunk.rl_mw = NULL;
1293 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1297 * Recover reply buffers from pool.
1298 * This happens when recovering from error conditions.
1299 * Post-increment counter/array index.
1302 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1304 struct rpcrdma_buffer *buffers = req->rl_buffer;
1305 unsigned long flags;
1307 if (req->rl_iov.length == 0) /* special case xprt_rdma_allocate() */
1308 buffers = ((struct rpcrdma_req *) buffers)->rl_buffer;
1309 spin_lock_irqsave(&buffers->rb_lock, flags);
1310 if (buffers->rb_recv_index < buffers->rb_max_requests) {
1311 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1312 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1314 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1318 * Put reply buffers back into pool when not attached to
1319 * request. This happens in error conditions, and when
1320 * aborting unbinds. Pre-decrement counter/array index.
1323 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1325 struct rpcrdma_buffer *buffers = rep->rr_buffer;
1326 unsigned long flags;
1328 rep->rr_func = NULL;
1329 spin_lock_irqsave(&buffers->rb_lock, flags);
1330 buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1331 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1335 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1339 rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
1340 struct ib_mr **mrp, struct ib_sge *iov)
1342 struct ib_phys_buf ipb;
1347 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
1349 iov->addr = ib_dma_map_single(ia->ri_id->device,
1350 va, len, DMA_BIDIRECTIONAL);
1353 if (ia->ri_have_dma_lkey) {
1355 iov->lkey = ia->ri_dma_lkey;
1357 } else if (ia->ri_bind_mem != NULL) {
1359 iov->lkey = ia->ri_bind_mem->lkey;
1363 ipb.addr = iov->addr;
1364 ipb.size = iov->length;
1365 mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
1366 IB_ACCESS_LOCAL_WRITE, &iov->addr);
1368 dprintk("RPC: %s: phys convert: 0x%llx "
1369 "registered 0x%llx length %d\n",
1370 __func__, (unsigned long long)ipb.addr,
1371 (unsigned long long)iov->addr, len);
1376 dprintk("RPC: %s: failed with %i\n", __func__, rc);
1379 iov->lkey = mr->lkey;
1387 rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
1388 struct ib_mr *mr, struct ib_sge *iov)
1392 ib_dma_unmap_single(ia->ri_id->device,
1393 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1398 rc = ib_dereg_mr(mr);
1400 dprintk("RPC: %s: ib_dereg_mr failed %i\n", __func__, rc);
1405 * Wrappers for chunk registration, shared by read/write chunk code.
1409 rpcrdma_map_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg, int writing)
1411 seg->mr_dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
1412 seg->mr_dmalen = seg->mr_len;
1414 seg->mr_dma = ib_dma_map_page(ia->ri_id->device,
1415 seg->mr_page, offset_in_page(seg->mr_offset),
1416 seg->mr_dmalen, seg->mr_dir);
1418 seg->mr_dma = ib_dma_map_single(ia->ri_id->device,
1420 seg->mr_dmalen, seg->mr_dir);
1424 rpcrdma_unmap_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg)
1427 ib_dma_unmap_page(ia->ri_id->device,
1428 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1430 ib_dma_unmap_single(ia->ri_id->device,
1431 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1435 rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
1436 int *nsegs, int writing, struct rpcrdma_ia *ia,
1437 struct rpcrdma_xprt *r_xprt)
1439 struct rpcrdma_mr_seg *seg1 = seg;
1440 struct ib_send_wr frmr_wr, *bad_wr;
1445 pageoff = offset_in_page(seg1->mr_offset);
1446 seg1->mr_offset -= pageoff; /* start of page */
1447 seg1->mr_len += pageoff;
1449 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1450 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1451 for (i = 0; i < *nsegs;) {
1452 rpcrdma_map_one(ia, seg, writing);
1453 seg1->mr_chunk.rl_mw->r.frmr.fr_pgl->page_list[i] = seg->mr_dma;
1457 /* Check for holes */
1458 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1459 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1462 dprintk("RPC: %s: Using frmr %p to map %d segments\n",
1463 __func__, seg1->mr_chunk.rl_mw, i);
1466 key = (u8)(seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey & 0x000000FF);
1467 ib_update_fast_reg_key(seg1->mr_chunk.rl_mw->r.frmr.fr_mr, ++key);
1469 /* Prepare FRMR WR */
1470 memset(&frmr_wr, 0, sizeof frmr_wr);
1471 frmr_wr.opcode = IB_WR_FAST_REG_MR;
1472 frmr_wr.send_flags = 0; /* unsignaled */
1473 frmr_wr.wr.fast_reg.iova_start = (unsigned long)seg1->mr_dma;
1474 frmr_wr.wr.fast_reg.page_list = seg1->mr_chunk.rl_mw->r.frmr.fr_pgl;
1475 frmr_wr.wr.fast_reg.page_list_len = i;
1476 frmr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1477 frmr_wr.wr.fast_reg.length = i << PAGE_SHIFT;
1478 frmr_wr.wr.fast_reg.access_flags = (writing ?
1479 IB_ACCESS_REMOTE_WRITE : IB_ACCESS_REMOTE_READ);
1480 frmr_wr.wr.fast_reg.rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1481 DECR_CQCOUNT(&r_xprt->rx_ep);
1483 rc = ib_post_send(ia->ri_id->qp, &frmr_wr, &bad_wr);
1486 dprintk("RPC: %s: failed ib_post_send for register,"
1487 " status %i\n", __func__, rc);
1489 rpcrdma_unmap_one(ia, --seg);
1491 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1492 seg1->mr_base = seg1->mr_dma + pageoff;
1501 rpcrdma_deregister_frmr_external(struct rpcrdma_mr_seg *seg,
1502 struct rpcrdma_ia *ia, struct rpcrdma_xprt *r_xprt)
1504 struct rpcrdma_mr_seg *seg1 = seg;
1505 struct ib_send_wr invalidate_wr, *bad_wr;
1508 while (seg1->mr_nsegs--)
1509 rpcrdma_unmap_one(ia, seg++);
1511 memset(&invalidate_wr, 0, sizeof invalidate_wr);
1512 invalidate_wr.opcode = IB_WR_LOCAL_INV;
1513 invalidate_wr.send_flags = 0; /* unsignaled */
1514 invalidate_wr.ex.invalidate_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1515 DECR_CQCOUNT(&r_xprt->rx_ep);
1517 rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
1519 dprintk("RPC: %s: failed ib_post_send for invalidate,"
1520 " status %i\n", __func__, rc);
1525 rpcrdma_register_fmr_external(struct rpcrdma_mr_seg *seg,
1526 int *nsegs, int writing, struct rpcrdma_ia *ia)
1528 struct rpcrdma_mr_seg *seg1 = seg;
1529 u64 physaddrs[RPCRDMA_MAX_DATA_SEGS];
1530 int len, pageoff, i, rc;
1532 pageoff = offset_in_page(seg1->mr_offset);
1533 seg1->mr_offset -= pageoff; /* start of page */
1534 seg1->mr_len += pageoff;
1536 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1537 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1538 for (i = 0; i < *nsegs;) {
1539 rpcrdma_map_one(ia, seg, writing);
1540 physaddrs[i] = seg->mr_dma;
1544 /* Check for holes */
1545 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1546 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1549 rc = ib_map_phys_fmr(seg1->mr_chunk.rl_mw->r.fmr,
1550 physaddrs, i, seg1->mr_dma);
1552 dprintk("RPC: %s: failed ib_map_phys_fmr "
1553 "%u@0x%llx+%i (%d)... status %i\n", __func__,
1554 len, (unsigned long long)seg1->mr_dma,
1557 rpcrdma_unmap_one(ia, --seg);
1559 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.fmr->rkey;
1560 seg1->mr_base = seg1->mr_dma + pageoff;
1569 rpcrdma_deregister_fmr_external(struct rpcrdma_mr_seg *seg,
1570 struct rpcrdma_ia *ia)
1572 struct rpcrdma_mr_seg *seg1 = seg;
1576 list_add(&seg1->mr_chunk.rl_mw->r.fmr->list, &l);
1577 rc = ib_unmap_fmr(&l);
1578 while (seg1->mr_nsegs--)
1579 rpcrdma_unmap_one(ia, seg++);
1581 dprintk("RPC: %s: failed ib_unmap_fmr,"
1582 " status %i\n", __func__, rc);
1587 rpcrdma_register_memwin_external(struct rpcrdma_mr_seg *seg,
1588 int *nsegs, int writing, struct rpcrdma_ia *ia,
1589 struct rpcrdma_xprt *r_xprt)
1591 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1592 IB_ACCESS_REMOTE_READ);
1593 struct ib_mw_bind param;
1597 rpcrdma_map_one(ia, seg, writing);
1598 param.mr = ia->ri_bind_mem;
1599 param.wr_id = 0ULL; /* no send cookie */
1600 param.addr = seg->mr_dma;
1601 param.length = seg->mr_len;
1602 param.send_flags = 0;
1603 param.mw_access_flags = mem_priv;
1605 DECR_CQCOUNT(&r_xprt->rx_ep);
1606 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1608 dprintk("RPC: %s: failed ib_bind_mw "
1609 "%u@0x%llx status %i\n",
1610 __func__, seg->mr_len,
1611 (unsigned long long)seg->mr_dma, rc);
1612 rpcrdma_unmap_one(ia, seg);
1614 seg->mr_rkey = seg->mr_chunk.rl_mw->r.mw->rkey;
1615 seg->mr_base = param.addr;
1622 rpcrdma_deregister_memwin_external(struct rpcrdma_mr_seg *seg,
1623 struct rpcrdma_ia *ia,
1624 struct rpcrdma_xprt *r_xprt, void **r)
1626 struct ib_mw_bind param;
1630 BUG_ON(seg->mr_nsegs != 1);
1631 param.mr = ia->ri_bind_mem;
1632 param.addr = 0ULL; /* unbind */
1634 param.mw_access_flags = 0;
1636 param.wr_id = (u64) (unsigned long) *r;
1637 param.send_flags = IB_SEND_SIGNALED;
1638 INIT_CQCOUNT(&r_xprt->rx_ep);
1641 param.send_flags = 0;
1642 DECR_CQCOUNT(&r_xprt->rx_ep);
1644 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1645 rpcrdma_unmap_one(ia, seg);
1647 dprintk("RPC: %s: failed ib_(un)bind_mw,"
1648 " status %i\n", __func__, rc);
1650 *r = NULL; /* will upcall on completion */
1655 rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
1656 int *nsegs, int writing, struct rpcrdma_ia *ia)
1658 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1659 IB_ACCESS_REMOTE_READ);
1660 struct rpcrdma_mr_seg *seg1 = seg;
1661 struct ib_phys_buf ipb[RPCRDMA_MAX_DATA_SEGS];
1664 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1665 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1666 for (len = 0, i = 0; i < *nsegs;) {
1667 rpcrdma_map_one(ia, seg, writing);
1668 ipb[i].addr = seg->mr_dma;
1669 ipb[i].size = seg->mr_len;
1673 /* Check for holes */
1674 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1675 offset_in_page((seg-1)->mr_offset+(seg-1)->mr_len))
1678 seg1->mr_base = seg1->mr_dma;
1679 seg1->mr_chunk.rl_mr = ib_reg_phys_mr(ia->ri_pd,
1680 ipb, i, mem_priv, &seg1->mr_base);
1681 if (IS_ERR(seg1->mr_chunk.rl_mr)) {
1682 rc = PTR_ERR(seg1->mr_chunk.rl_mr);
1683 dprintk("RPC: %s: failed ib_reg_phys_mr "
1684 "%u@0x%llx (%d)... status %i\n",
1686 (unsigned long long)seg1->mr_dma, i, rc);
1688 rpcrdma_unmap_one(ia, --seg);
1690 seg1->mr_rkey = seg1->mr_chunk.rl_mr->rkey;
1699 rpcrdma_deregister_default_external(struct rpcrdma_mr_seg *seg,
1700 struct rpcrdma_ia *ia)
1702 struct rpcrdma_mr_seg *seg1 = seg;
1705 rc = ib_dereg_mr(seg1->mr_chunk.rl_mr);
1706 seg1->mr_chunk.rl_mr = NULL;
1707 while (seg1->mr_nsegs--)
1708 rpcrdma_unmap_one(ia, seg++);
1710 dprintk("RPC: %s: failed ib_dereg_mr,"
1711 " status %i\n", __func__, rc);
1716 rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
1717 int nsegs, int writing, struct rpcrdma_xprt *r_xprt)
1719 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1722 switch (ia->ri_memreg_strategy) {
1724 #if RPCRDMA_PERSISTENT_REGISTRATION
1725 case RPCRDMA_ALLPHYSICAL:
1726 rpcrdma_map_one(ia, seg, writing);
1727 seg->mr_rkey = ia->ri_bind_mem->rkey;
1728 seg->mr_base = seg->mr_dma;
1734 /* Registration using frmr registration */
1736 rc = rpcrdma_register_frmr_external(seg, &nsegs, writing, ia, r_xprt);
1739 /* Registration using fmr memory registration */
1740 case RPCRDMA_MTHCAFMR:
1741 rc = rpcrdma_register_fmr_external(seg, &nsegs, writing, ia);
1744 /* Registration using memory windows */
1745 case RPCRDMA_MEMWINDOWS_ASYNC:
1746 case RPCRDMA_MEMWINDOWS:
1747 rc = rpcrdma_register_memwin_external(seg, &nsegs, writing, ia, r_xprt);
1750 /* Default registration each time */
1752 rc = rpcrdma_register_default_external(seg, &nsegs, writing, ia);
1762 rpcrdma_deregister_external(struct rpcrdma_mr_seg *seg,
1763 struct rpcrdma_xprt *r_xprt, void *r)
1765 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1766 int nsegs = seg->mr_nsegs, rc;
1768 switch (ia->ri_memreg_strategy) {
1770 #if RPCRDMA_PERSISTENT_REGISTRATION
1771 case RPCRDMA_ALLPHYSICAL:
1773 rpcrdma_unmap_one(ia, seg);
1779 rc = rpcrdma_deregister_frmr_external(seg, ia, r_xprt);
1782 case RPCRDMA_MTHCAFMR:
1783 rc = rpcrdma_deregister_fmr_external(seg, ia);
1786 case RPCRDMA_MEMWINDOWS_ASYNC:
1787 case RPCRDMA_MEMWINDOWS:
1788 rc = rpcrdma_deregister_memwin_external(seg, ia, r_xprt, &r);
1792 rc = rpcrdma_deregister_default_external(seg, ia);
1796 struct rpcrdma_rep *rep = r;
1797 void (*func)(struct rpcrdma_rep *) = rep->rr_func;
1798 rep->rr_func = NULL;
1799 func(rep); /* dereg done, callback now */
1805 * Prepost any receive buffer, then post send.
1807 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1810 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1811 struct rpcrdma_ep *ep,
1812 struct rpcrdma_req *req)
1814 struct ib_send_wr send_wr, *send_wr_fail;
1815 struct rpcrdma_rep *rep = req->rl_reply;
1819 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1822 req->rl_reply = NULL;
1825 send_wr.next = NULL;
1826 send_wr.wr_id = 0ULL; /* no send cookie */
1827 send_wr.sg_list = req->rl_send_iov;
1828 send_wr.num_sge = req->rl_niovs;
1829 send_wr.opcode = IB_WR_SEND;
1830 if (send_wr.num_sge == 4) /* no need to sync any pad (constant) */
1831 ib_dma_sync_single_for_device(ia->ri_id->device,
1832 req->rl_send_iov[3].addr, req->rl_send_iov[3].length,
1834 ib_dma_sync_single_for_device(ia->ri_id->device,
1835 req->rl_send_iov[1].addr, req->rl_send_iov[1].length,
1837 ib_dma_sync_single_for_device(ia->ri_id->device,
1838 req->rl_send_iov[0].addr, req->rl_send_iov[0].length,
1841 if (DECR_CQCOUNT(ep) > 0)
1842 send_wr.send_flags = 0;
1843 else { /* Provider must take a send completion every now and then */
1845 send_wr.send_flags = IB_SEND_SIGNALED;
1848 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1850 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1857 * (Re)post a receive buffer.
1860 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1861 struct rpcrdma_ep *ep,
1862 struct rpcrdma_rep *rep)
1864 struct ib_recv_wr recv_wr, *recv_wr_fail;
1867 recv_wr.next = NULL;
1868 recv_wr.wr_id = (u64) (unsigned long) rep;
1869 recv_wr.sg_list = &rep->rr_iov;
1870 recv_wr.num_sge = 1;
1872 ib_dma_sync_single_for_cpu(ia->ri_id->device,
1873 rep->rr_iov.addr, rep->rr_iov.length, DMA_BIDIRECTIONAL);
1876 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1879 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,