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:
288 complete(&ia->ri_done);
290 case RDMA_CM_EVENT_ADDR_ERROR:
291 ia->ri_async_rc = -EHOSTUNREACH;
292 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
294 complete(&ia->ri_done);
296 case RDMA_CM_EVENT_ROUTE_ERROR:
297 ia->ri_async_rc = -ENETUNREACH;
298 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
300 complete(&ia->ri_done);
302 case RDMA_CM_EVENT_ESTABLISHED:
304 ib_query_qp(ia->ri_id->qp, &attr,
305 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
307 dprintk("RPC: %s: %d responder resources"
309 __func__, attr.max_dest_rd_atomic, attr.max_rd_atomic);
311 case RDMA_CM_EVENT_CONNECT_ERROR:
312 connstate = -ENOTCONN;
314 case RDMA_CM_EVENT_UNREACHABLE:
315 connstate = -ENETDOWN;
317 case RDMA_CM_EVENT_REJECTED:
318 connstate = -ECONNREFUSED;
320 case RDMA_CM_EVENT_DISCONNECTED:
321 connstate = -ECONNABORTED;
323 case RDMA_CM_EVENT_DEVICE_REMOVAL:
326 dprintk("RPC: %s: %s: %u.%u.%u.%u:%u"
327 " (ep 0x%p event 0x%x)\n",
329 (event->event <= 11) ? conn[event->event] :
330 "unknown connection error",
331 NIPQUAD(addr->sin_addr.s_addr),
332 ntohs(addr->sin_port),
334 atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
335 dprintk("RPC: %s: %sconnected\n",
336 __func__, connstate > 0 ? "" : "dis");
337 ep->rep_connected = connstate;
339 wake_up_all(&ep->rep_connect_wait);
342 dprintk("RPC: %s: unexpected CM event %d\n",
343 __func__, event->event);
350 static struct rdma_cm_id *
351 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
352 struct rpcrdma_ia *ia, struct sockaddr *addr)
354 struct rdma_cm_id *id;
357 init_completion(&ia->ri_done);
359 id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP);
362 dprintk("RPC: %s: rdma_create_id() failed %i\n",
367 ia->ri_async_rc = -ETIMEDOUT;
368 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
370 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
374 wait_for_completion_interruptible_timeout(&ia->ri_done,
375 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
376 rc = ia->ri_async_rc;
380 ia->ri_async_rc = -ETIMEDOUT;
381 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
383 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
387 wait_for_completion_interruptible_timeout(&ia->ri_done,
388 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
389 rc = ia->ri_async_rc;
401 * Drain any cq, prior to teardown.
404 rpcrdma_clean_cq(struct ib_cq *cq)
409 while (1 == ib_poll_cq(cq, 1, &wc))
413 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
414 __func__, count, wc.opcode);
418 * Exported functions.
422 * Open and initialize an Interface Adapter.
423 * o initializes fields of struct rpcrdma_ia, including
424 * interface and provider attributes and protection zone.
427 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
430 struct ib_device_attr devattr;
431 struct rpcrdma_ia *ia = &xprt->rx_ia;
433 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
434 if (IS_ERR(ia->ri_id)) {
435 rc = PTR_ERR(ia->ri_id);
439 ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
440 if (IS_ERR(ia->ri_pd)) {
441 rc = PTR_ERR(ia->ri_pd);
442 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
448 * Query the device to determine if the requested memory
449 * registration strategy is supported. If it isn't, set the
450 * strategy to a globally supported model.
452 rc = ib_query_device(ia->ri_id->device, &devattr);
454 dprintk("RPC: %s: ib_query_device failed %d\n",
459 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
460 ia->ri_have_dma_lkey = 1;
461 ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
465 case RPCRDMA_MEMWINDOWS:
466 case RPCRDMA_MEMWINDOWS_ASYNC:
467 if (!(devattr.device_cap_flags & IB_DEVICE_MEM_WINDOW)) {
468 dprintk("RPC: %s: MEMWINDOWS registration "
469 "specified but not supported by adapter, "
470 "using slower RPCRDMA_REGISTER\n",
472 memreg = RPCRDMA_REGISTER;
475 case RPCRDMA_MTHCAFMR:
476 if (!ia->ri_id->device->alloc_fmr) {
477 #if RPCRDMA_PERSISTENT_REGISTRATION
478 dprintk("RPC: %s: MTHCAFMR registration "
479 "specified but not supported by adapter, "
480 "using riskier RPCRDMA_ALLPHYSICAL\n",
482 memreg = RPCRDMA_ALLPHYSICAL;
484 dprintk("RPC: %s: MTHCAFMR registration "
485 "specified but not supported by adapter, "
486 "using slower RPCRDMA_REGISTER\n",
488 memreg = RPCRDMA_REGISTER;
493 /* Requires both frmr reg and local dma lkey */
494 if ((devattr.device_cap_flags &
495 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
496 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) {
497 #if RPCRDMA_PERSISTENT_REGISTRATION
498 dprintk("RPC: %s: FRMR registration "
499 "specified but not supported by adapter, "
500 "using riskier RPCRDMA_ALLPHYSICAL\n",
502 memreg = RPCRDMA_ALLPHYSICAL;
504 dprintk("RPC: %s: FRMR registration "
505 "specified but not supported by adapter, "
506 "using slower RPCRDMA_REGISTER\n",
508 memreg = RPCRDMA_REGISTER;
515 * Optionally obtain an underlying physical identity mapping in
516 * order to do a memory window-based bind. This base registration
517 * is protected from remote access - that is enabled only by binding
518 * for the specific bytes targeted during each RPC operation, and
519 * revoked after the corresponding completion similar to a storage
523 case RPCRDMA_BOUNCEBUFFERS:
524 case RPCRDMA_REGISTER:
527 #if RPCRDMA_PERSISTENT_REGISTRATION
528 case RPCRDMA_ALLPHYSICAL:
529 mem_priv = IB_ACCESS_LOCAL_WRITE |
530 IB_ACCESS_REMOTE_WRITE |
531 IB_ACCESS_REMOTE_READ;
534 case RPCRDMA_MEMWINDOWS_ASYNC:
535 case RPCRDMA_MEMWINDOWS:
536 mem_priv = IB_ACCESS_LOCAL_WRITE |
539 case RPCRDMA_MTHCAFMR:
540 if (ia->ri_have_dma_lkey)
542 mem_priv = IB_ACCESS_LOCAL_WRITE;
544 ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
545 if (IS_ERR(ia->ri_bind_mem)) {
546 printk(KERN_ALERT "%s: ib_get_dma_mr for "
547 "phys register failed with %lX\n\t"
548 "Will continue with degraded performance\n",
549 __func__, PTR_ERR(ia->ri_bind_mem));
550 memreg = RPCRDMA_REGISTER;
551 ia->ri_bind_mem = NULL;
555 printk(KERN_ERR "%s: invalid memory registration mode %d\n",
560 dprintk("RPC: %s: memory registration strategy is %d\n",
563 /* Else will do memory reg/dereg for each chunk */
564 ia->ri_memreg_strategy = memreg;
568 rdma_destroy_id(ia->ri_id);
575 * Clean up/close an IA.
576 * o if event handles and PD have been initialized, free them.
580 rpcrdma_ia_close(struct rpcrdma_ia *ia)
584 dprintk("RPC: %s: entering\n", __func__);
585 if (ia->ri_bind_mem != NULL) {
586 rc = ib_dereg_mr(ia->ri_bind_mem);
587 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
590 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
592 rdma_destroy_qp(ia->ri_id);
593 rdma_destroy_id(ia->ri_id);
596 if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
597 rc = ib_dealloc_pd(ia->ri_pd);
598 dprintk("RPC: %s: ib_dealloc_pd returned %i\n",
604 * Create unconnected endpoint.
607 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
608 struct rpcrdma_create_data_internal *cdata)
610 struct ib_device_attr devattr;
613 rc = ib_query_device(ia->ri_id->device, &devattr);
615 dprintk("RPC: %s: ib_query_device failed %d\n",
620 /* check provider's send/recv wr limits */
621 if (cdata->max_requests > devattr.max_qp_wr)
622 cdata->max_requests = devattr.max_qp_wr;
624 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
625 ep->rep_attr.qp_context = ep;
626 /* send_cq and recv_cq initialized below */
627 ep->rep_attr.srq = NULL;
628 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
629 switch (ia->ri_memreg_strategy) {
631 /* Add room for frmr register and invalidate WRs */
632 ep->rep_attr.cap.max_send_wr *= 3;
633 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
636 case RPCRDMA_MEMWINDOWS_ASYNC:
637 case RPCRDMA_MEMWINDOWS:
638 /* Add room for mw_binds+unbinds - overkill! */
639 ep->rep_attr.cap.max_send_wr++;
640 ep->rep_attr.cap.max_send_wr *= (2 * RPCRDMA_MAX_SEGS);
641 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
647 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
648 ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
649 ep->rep_attr.cap.max_recv_sge = 1;
650 ep->rep_attr.cap.max_inline_data = 0;
651 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
652 ep->rep_attr.qp_type = IB_QPT_RC;
653 ep->rep_attr.port_num = ~0;
655 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
656 "iovs: send %d recv %d\n",
658 ep->rep_attr.cap.max_send_wr,
659 ep->rep_attr.cap.max_recv_wr,
660 ep->rep_attr.cap.max_send_sge,
661 ep->rep_attr.cap.max_recv_sge);
663 /* set trigger for requesting send completion */
664 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 /* - 1*/;
665 switch (ia->ri_memreg_strategy) {
666 case RPCRDMA_MEMWINDOWS_ASYNC:
667 case RPCRDMA_MEMWINDOWS:
668 ep->rep_cqinit -= RPCRDMA_MAX_SEGS;
673 if (ep->rep_cqinit <= 2)
677 init_waitqueue_head(&ep->rep_connect_wait);
680 * Create a single cq for receive dto and mw_bind (only ever
681 * care about unbind, really). Send completions are suppressed.
682 * Use single threaded tasklet upcalls to maintain ordering.
684 ep->rep_cq = ib_create_cq(ia->ri_id->device, rpcrdma_cq_event_upcall,
685 rpcrdma_cq_async_error_upcall, NULL,
686 ep->rep_attr.cap.max_recv_wr +
687 ep->rep_attr.cap.max_send_wr + 1, 0);
688 if (IS_ERR(ep->rep_cq)) {
689 rc = PTR_ERR(ep->rep_cq);
690 dprintk("RPC: %s: ib_create_cq failed: %i\n",
695 rc = ib_req_notify_cq(ep->rep_cq, IB_CQ_NEXT_COMP);
697 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
702 ep->rep_attr.send_cq = ep->rep_cq;
703 ep->rep_attr.recv_cq = ep->rep_cq;
705 /* Initialize cma parameters */
707 /* RPC/RDMA does not use private data */
708 ep->rep_remote_cma.private_data = NULL;
709 ep->rep_remote_cma.private_data_len = 0;
711 /* Client offers RDMA Read but does not initiate */
712 ep->rep_remote_cma.initiator_depth = 0;
713 if (ia->ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS)
714 ep->rep_remote_cma.responder_resources = 0;
715 else if (devattr.max_qp_rd_atom > 32) /* arbitrary but <= 255 */
716 ep->rep_remote_cma.responder_resources = 32;
718 ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;
720 ep->rep_remote_cma.retry_count = 7;
721 ep->rep_remote_cma.flow_control = 0;
722 ep->rep_remote_cma.rnr_retry_count = 0;
727 err = ib_destroy_cq(ep->rep_cq);
729 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
738 * Disconnect and destroy endpoint. After this, the only
739 * valid operations on the ep are to free it (if dynamically
740 * allocated) or re-create it.
742 * The caller's error handling must be sure to not leak the endpoint
743 * if this function fails.
746 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
750 dprintk("RPC: %s: entering, connected is %d\n",
751 __func__, ep->rep_connected);
754 rc = rpcrdma_ep_disconnect(ep, ia);
756 dprintk("RPC: %s: rpcrdma_ep_disconnect"
757 " returned %i\n", __func__, rc);
758 rdma_destroy_qp(ia->ri_id);
759 ia->ri_id->qp = NULL;
762 /* padding - could be done in rpcrdma_buffer_destroy... */
763 if (ep->rep_pad_mr) {
764 rpcrdma_deregister_internal(ia, ep->rep_pad_mr, &ep->rep_pad);
765 ep->rep_pad_mr = NULL;
768 rpcrdma_clean_cq(ep->rep_cq);
769 rc = ib_destroy_cq(ep->rep_cq);
771 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
778 * Connect unconnected endpoint.
781 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
783 struct rdma_cm_id *id;
786 int reconnect = (ep->rep_connected != 0);
789 struct rpcrdma_xprt *xprt;
791 rc = rpcrdma_ep_disconnect(ep, ia);
792 if (rc && rc != -ENOTCONN)
793 dprintk("RPC: %s: rpcrdma_ep_disconnect"
794 " status %i\n", __func__, rc);
795 rpcrdma_clean_cq(ep->rep_cq);
797 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
798 id = rpcrdma_create_id(xprt, ia,
799 (struct sockaddr *)&xprt->rx_data.addr);
804 /* TEMP TEMP TEMP - fail if new device:
805 * Deregister/remarshal *all* requests!
806 * Close and recreate adapter, pd, etc!
807 * Re-determine all attributes still sane!
808 * More stuff I haven't thought of!
811 if (ia->ri_id->device != id->device) {
812 printk("RPC: %s: can't reconnect on "
813 "different device!\n", __func__);
819 rdma_destroy_qp(ia->ri_id);
820 rdma_destroy_id(ia->ri_id);
824 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
826 dprintk("RPC: %s: rdma_create_qp failed %i\n",
831 /* XXX Tavor device performs badly with 2K MTU! */
832 if (strnicmp(ia->ri_id->device->dma_device->bus->name, "pci", 3) == 0) {
833 struct pci_dev *pcid = to_pci_dev(ia->ri_id->device->dma_device);
834 if (pcid->device == PCI_DEVICE_ID_MELLANOX_TAVOR &&
835 (pcid->vendor == PCI_VENDOR_ID_MELLANOX ||
836 pcid->vendor == PCI_VENDOR_ID_TOPSPIN)) {
837 struct ib_qp_attr attr = {
838 .path_mtu = IB_MTU_1024
840 rc = ib_modify_qp(ia->ri_id->qp, &attr, IB_QP_PATH_MTU);
844 ep->rep_connected = 0;
846 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
848 dprintk("RPC: %s: rdma_connect() failed with %i\n",
856 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
859 * Check state. A non-peer reject indicates no listener
860 * (ECONNREFUSED), which may be a transient state. All
861 * others indicate a transport condition which has already
862 * undergone a best-effort.
864 if (ep->rep_connected == -ECONNREFUSED
865 && ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
866 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
869 if (ep->rep_connected <= 0) {
870 /* Sometimes, the only way to reliably connect to remote
871 * CMs is to use same nonzero values for ORD and IRD. */
872 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
873 (ep->rep_remote_cma.responder_resources == 0 ||
874 ep->rep_remote_cma.initiator_depth !=
875 ep->rep_remote_cma.responder_resources)) {
876 if (ep->rep_remote_cma.responder_resources == 0)
877 ep->rep_remote_cma.responder_resources = 1;
878 ep->rep_remote_cma.initiator_depth =
879 ep->rep_remote_cma.responder_resources;
882 rc = ep->rep_connected;
884 dprintk("RPC: %s: connected\n", __func__);
889 ep->rep_connected = rc;
894 * rpcrdma_ep_disconnect
896 * This is separate from destroy to facilitate the ability
897 * to reconnect without recreating the endpoint.
899 * This call is not reentrant, and must not be made in parallel
900 * on the same endpoint.
903 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
907 rpcrdma_clean_cq(ep->rep_cq);
908 rc = rdma_disconnect(ia->ri_id);
910 /* returns without wait if not connected */
911 wait_event_interruptible(ep->rep_connect_wait,
912 ep->rep_connected != 1);
913 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
914 (ep->rep_connected == 1) ? "still " : "dis");
916 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
917 ep->rep_connected = rc;
923 * Initialize buffer memory
926 rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
927 struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
932 struct rpcrdma_mw *r;
934 buf->rb_max_requests = cdata->max_requests;
935 spin_lock_init(&buf->rb_lock);
936 atomic_set(&buf->rb_credits, 1);
939 * 1. arrays for send and recv pointers
940 * 2. arrays of struct rpcrdma_req to fill in pointers
941 * 3. array of struct rpcrdma_rep for replies
943 * 5. mw's, fmr's or frmr's, if any
944 * Send/recv buffers in req/rep need to be registered
947 len = buf->rb_max_requests *
948 (sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
949 len += cdata->padding;
950 switch (ia->ri_memreg_strategy) {
952 len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
953 sizeof(struct rpcrdma_mw);
955 case RPCRDMA_MTHCAFMR:
956 /* TBD we are perhaps overallocating here */
957 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
958 sizeof(struct rpcrdma_mw);
960 case RPCRDMA_MEMWINDOWS_ASYNC:
961 case RPCRDMA_MEMWINDOWS:
962 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
963 sizeof(struct rpcrdma_mw);
969 /* allocate 1, 4 and 5 in one shot */
970 p = kzalloc(len, GFP_KERNEL);
972 dprintk("RPC: %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
977 buf->rb_pool = p; /* for freeing it later */
979 buf->rb_send_bufs = (struct rpcrdma_req **) p;
980 p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
981 buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
982 p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
985 * Register the zeroed pad buffer, if any.
987 if (cdata->padding) {
988 rc = rpcrdma_register_internal(ia, p, cdata->padding,
989 &ep->rep_pad_mr, &ep->rep_pad);
996 * Allocate the fmr's, or mw's for mw_bind chunk registration.
997 * We "cycle" the mw's in order to minimize rkey reuse,
998 * and also reduce unbind-to-bind collision.
1000 INIT_LIST_HEAD(&buf->rb_mws);
1001 r = (struct rpcrdma_mw *)p;
1002 switch (ia->ri_memreg_strategy) {
1004 for (i = buf->rb_max_requests * RPCRDMA_MAX_SEGS; i; i--) {
1005 r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
1007 if (IS_ERR(r->r.frmr.fr_mr)) {
1008 rc = PTR_ERR(r->r.frmr.fr_mr);
1009 dprintk("RPC: %s: ib_alloc_fast_reg_mr"
1010 " failed %i\n", __func__, rc);
1014 ib_alloc_fast_reg_page_list(ia->ri_id->device,
1016 if (IS_ERR(r->r.frmr.fr_pgl)) {
1017 rc = PTR_ERR(r->r.frmr.fr_pgl);
1019 "ib_alloc_fast_reg_page_list "
1020 "failed %i\n", __func__, rc);
1023 list_add(&r->mw_list, &buf->rb_mws);
1027 case RPCRDMA_MTHCAFMR:
1028 /* TBD we are perhaps overallocating here */
1029 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1030 static struct ib_fmr_attr fa =
1031 { RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT };
1032 r->r.fmr = ib_alloc_fmr(ia->ri_pd,
1033 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
1035 if (IS_ERR(r->r.fmr)) {
1036 rc = PTR_ERR(r->r.fmr);
1037 dprintk("RPC: %s: ib_alloc_fmr"
1038 " failed %i\n", __func__, rc);
1041 list_add(&r->mw_list, &buf->rb_mws);
1045 case RPCRDMA_MEMWINDOWS_ASYNC:
1046 case RPCRDMA_MEMWINDOWS:
1047 /* Allocate one extra request's worth, for full cycling */
1048 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1049 r->r.mw = ib_alloc_mw(ia->ri_pd);
1050 if (IS_ERR(r->r.mw)) {
1051 rc = PTR_ERR(r->r.mw);
1052 dprintk("RPC: %s: ib_alloc_mw"
1053 " failed %i\n", __func__, rc);
1056 list_add(&r->mw_list, &buf->rb_mws);
1065 * Allocate/init the request/reply buffers. Doing this
1066 * using kmalloc for now -- one for each buf.
1068 for (i = 0; i < buf->rb_max_requests; i++) {
1069 struct rpcrdma_req *req;
1070 struct rpcrdma_rep *rep;
1072 len = cdata->inline_wsize + sizeof(struct rpcrdma_req);
1073 /* RPC layer requests *double* size + 1K RPC_SLACK_SPACE! */
1074 /* Typical ~2400b, so rounding up saves work later */
1077 req = kmalloc(len, GFP_KERNEL);
1079 dprintk("RPC: %s: request buffer %d alloc"
1080 " failed\n", __func__, i);
1084 memset(req, 0, sizeof(struct rpcrdma_req));
1085 buf->rb_send_bufs[i] = req;
1086 buf->rb_send_bufs[i]->rl_buffer = buf;
1088 rc = rpcrdma_register_internal(ia, req->rl_base,
1089 len - offsetof(struct rpcrdma_req, rl_base),
1090 &buf->rb_send_bufs[i]->rl_handle,
1091 &buf->rb_send_bufs[i]->rl_iov);
1095 buf->rb_send_bufs[i]->rl_size = len-sizeof(struct rpcrdma_req);
1097 len = cdata->inline_rsize + sizeof(struct rpcrdma_rep);
1098 rep = kmalloc(len, GFP_KERNEL);
1100 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1105 memset(rep, 0, sizeof(struct rpcrdma_rep));
1106 buf->rb_recv_bufs[i] = rep;
1107 buf->rb_recv_bufs[i]->rr_buffer = buf;
1108 init_waitqueue_head(&rep->rr_unbind);
1110 rc = rpcrdma_register_internal(ia, rep->rr_base,
1111 len - offsetof(struct rpcrdma_rep, rr_base),
1112 &buf->rb_recv_bufs[i]->rr_handle,
1113 &buf->rb_recv_bufs[i]->rr_iov);
1118 dprintk("RPC: %s: max_requests %d\n",
1119 __func__, buf->rb_max_requests);
1123 rpcrdma_buffer_destroy(buf);
1128 * Unregister and destroy buffer memory. Need to deal with
1129 * partial initialization, so it's callable from failed create.
1130 * Must be called before destroying endpoint, as registrations
1134 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1137 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1138 struct rpcrdma_mw *r;
1140 /* clean up in reverse order from create
1141 * 1. recv mr memory (mr free, then kfree)
1142 * 1a. bind mw memory
1143 * 2. send mr memory (mr free, then kfree)
1144 * 3. padding (if any) [moved to rpcrdma_ep_destroy]
1147 dprintk("RPC: %s: entering\n", __func__);
1149 for (i = 0; i < buf->rb_max_requests; i++) {
1150 if (buf->rb_recv_bufs && buf->rb_recv_bufs[i]) {
1151 rpcrdma_deregister_internal(ia,
1152 buf->rb_recv_bufs[i]->rr_handle,
1153 &buf->rb_recv_bufs[i]->rr_iov);
1154 kfree(buf->rb_recv_bufs[i]);
1156 if (buf->rb_send_bufs && buf->rb_send_bufs[i]) {
1157 while (!list_empty(&buf->rb_mws)) {
1158 r = list_entry(buf->rb_mws.next,
1159 struct rpcrdma_mw, mw_list);
1160 list_del(&r->mw_list);
1161 switch (ia->ri_memreg_strategy) {
1163 rc = ib_dereg_mr(r->r.frmr.fr_mr);
1169 ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
1171 case RPCRDMA_MTHCAFMR:
1172 rc = ib_dealloc_fmr(r->r.fmr);
1179 case RPCRDMA_MEMWINDOWS_ASYNC:
1180 case RPCRDMA_MEMWINDOWS:
1181 rc = ib_dealloc_mw(r->r.mw);
1192 rpcrdma_deregister_internal(ia,
1193 buf->rb_send_bufs[i]->rl_handle,
1194 &buf->rb_send_bufs[i]->rl_iov);
1195 kfree(buf->rb_send_bufs[i]);
1199 kfree(buf->rb_pool);
1203 * Get a set of request/reply buffers.
1205 * Reply buffer (if needed) is attached to send buffer upon return.
1207 * rb_send_index and rb_recv_index MUST always be pointing to the
1208 * *next* available buffer (non-NULL). They are incremented after
1209 * removing buffers, and decremented *before* returning them.
1211 struct rpcrdma_req *
1212 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1214 struct rpcrdma_req *req;
1215 unsigned long flags;
1217 struct rpcrdma_mw *r;
1219 spin_lock_irqsave(&buffers->rb_lock, flags);
1220 if (buffers->rb_send_index == buffers->rb_max_requests) {
1221 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1222 dprintk("RPC: %s: out of request buffers\n", __func__);
1223 return ((struct rpcrdma_req *)NULL);
1226 req = buffers->rb_send_bufs[buffers->rb_send_index];
1227 if (buffers->rb_send_index < buffers->rb_recv_index) {
1228 dprintk("RPC: %s: %d extra receives outstanding (ok)\n",
1230 buffers->rb_recv_index - buffers->rb_send_index);
1231 req->rl_reply = NULL;
1233 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1234 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1236 buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1237 if (!list_empty(&buffers->rb_mws)) {
1238 i = RPCRDMA_MAX_SEGS - 1;
1240 r = list_entry(buffers->rb_mws.next,
1241 struct rpcrdma_mw, mw_list);
1242 list_del(&r->mw_list);
1243 req->rl_segments[i].mr_chunk.rl_mw = r;
1246 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1251 * Put request/reply buffers back into pool.
1252 * Pre-decrement counter/array index.
1255 rpcrdma_buffer_put(struct rpcrdma_req *req)
1257 struct rpcrdma_buffer *buffers = req->rl_buffer;
1258 struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1260 unsigned long flags;
1262 BUG_ON(req->rl_nchunks != 0);
1263 spin_lock_irqsave(&buffers->rb_lock, flags);
1264 buffers->rb_send_bufs[--buffers->rb_send_index] = req;
1266 if (req->rl_reply) {
1267 buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
1268 init_waitqueue_head(&req->rl_reply->rr_unbind);
1269 req->rl_reply->rr_func = NULL;
1270 req->rl_reply = NULL;
1272 switch (ia->ri_memreg_strategy) {
1274 case RPCRDMA_MTHCAFMR:
1275 case RPCRDMA_MEMWINDOWS_ASYNC:
1276 case RPCRDMA_MEMWINDOWS:
1278 * Cycle mw's back in reverse order, and "spin" them.
1279 * This delays and scrambles reuse as much as possible.
1283 struct rpcrdma_mw **mw;
1284 mw = &req->rl_segments[i].mr_chunk.rl_mw;
1285 list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
1287 } while (++i < RPCRDMA_MAX_SEGS);
1288 list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
1290 req->rl_segments[0].mr_chunk.rl_mw = NULL;
1295 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1299 * Recover reply buffers from pool.
1300 * This happens when recovering from error conditions.
1301 * Post-increment counter/array index.
1304 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1306 struct rpcrdma_buffer *buffers = req->rl_buffer;
1307 unsigned long flags;
1309 if (req->rl_iov.length == 0) /* special case xprt_rdma_allocate() */
1310 buffers = ((struct rpcrdma_req *) buffers)->rl_buffer;
1311 spin_lock_irqsave(&buffers->rb_lock, flags);
1312 if (buffers->rb_recv_index < buffers->rb_max_requests) {
1313 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1314 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1316 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1320 * Put reply buffers back into pool when not attached to
1321 * request. This happens in error conditions, and when
1322 * aborting unbinds. Pre-decrement counter/array index.
1325 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1327 struct rpcrdma_buffer *buffers = rep->rr_buffer;
1328 unsigned long flags;
1330 rep->rr_func = NULL;
1331 spin_lock_irqsave(&buffers->rb_lock, flags);
1332 buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1333 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1337 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1341 rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
1342 struct ib_mr **mrp, struct ib_sge *iov)
1344 struct ib_phys_buf ipb;
1349 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
1351 iov->addr = ib_dma_map_single(ia->ri_id->device,
1352 va, len, DMA_BIDIRECTIONAL);
1355 if (ia->ri_have_dma_lkey) {
1357 iov->lkey = ia->ri_dma_lkey;
1359 } else if (ia->ri_bind_mem != NULL) {
1361 iov->lkey = ia->ri_bind_mem->lkey;
1365 ipb.addr = iov->addr;
1366 ipb.size = iov->length;
1367 mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
1368 IB_ACCESS_LOCAL_WRITE, &iov->addr);
1370 dprintk("RPC: %s: phys convert: 0x%llx "
1371 "registered 0x%llx length %d\n",
1372 __func__, (unsigned long long)ipb.addr,
1373 (unsigned long long)iov->addr, len);
1378 dprintk("RPC: %s: failed with %i\n", __func__, rc);
1381 iov->lkey = mr->lkey;
1389 rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
1390 struct ib_mr *mr, struct ib_sge *iov)
1394 ib_dma_unmap_single(ia->ri_id->device,
1395 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1400 rc = ib_dereg_mr(mr);
1402 dprintk("RPC: %s: ib_dereg_mr failed %i\n", __func__, rc);
1407 * Wrappers for chunk registration, shared by read/write chunk code.
1411 rpcrdma_map_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg, int writing)
1413 seg->mr_dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
1414 seg->mr_dmalen = seg->mr_len;
1416 seg->mr_dma = ib_dma_map_page(ia->ri_id->device,
1417 seg->mr_page, offset_in_page(seg->mr_offset),
1418 seg->mr_dmalen, seg->mr_dir);
1420 seg->mr_dma = ib_dma_map_single(ia->ri_id->device,
1422 seg->mr_dmalen, seg->mr_dir);
1426 rpcrdma_unmap_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg)
1429 ib_dma_unmap_page(ia->ri_id->device,
1430 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1432 ib_dma_unmap_single(ia->ri_id->device,
1433 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1437 rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
1438 int *nsegs, int writing, struct rpcrdma_ia *ia,
1439 struct rpcrdma_xprt *r_xprt)
1441 struct rpcrdma_mr_seg *seg1 = seg;
1442 struct ib_send_wr frmr_wr, *bad_wr;
1447 pageoff = offset_in_page(seg1->mr_offset);
1448 seg1->mr_offset -= pageoff; /* start of page */
1449 seg1->mr_len += pageoff;
1451 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1452 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1453 for (i = 0; i < *nsegs;) {
1454 rpcrdma_map_one(ia, seg, writing);
1455 seg1->mr_chunk.rl_mw->r.frmr.fr_pgl->page_list[i] = seg->mr_dma;
1459 /* Check for holes */
1460 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1461 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1464 dprintk("RPC: %s: Using frmr %p to map %d segments\n",
1465 __func__, seg1->mr_chunk.rl_mw, i);
1468 key = (u8)(seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey & 0x000000FF);
1469 ib_update_fast_reg_key(seg1->mr_chunk.rl_mw->r.frmr.fr_mr, ++key);
1471 /* Prepare FRMR WR */
1472 memset(&frmr_wr, 0, sizeof frmr_wr);
1473 frmr_wr.opcode = IB_WR_FAST_REG_MR;
1474 frmr_wr.send_flags = 0; /* unsignaled */
1475 frmr_wr.wr.fast_reg.iova_start = (unsigned long)seg1->mr_dma;
1476 frmr_wr.wr.fast_reg.page_list = seg1->mr_chunk.rl_mw->r.frmr.fr_pgl;
1477 frmr_wr.wr.fast_reg.page_list_len = i;
1478 frmr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1479 frmr_wr.wr.fast_reg.length = i << PAGE_SHIFT;
1480 frmr_wr.wr.fast_reg.access_flags = (writing ?
1481 IB_ACCESS_REMOTE_WRITE : IB_ACCESS_REMOTE_READ);
1482 frmr_wr.wr.fast_reg.rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1483 DECR_CQCOUNT(&r_xprt->rx_ep);
1485 rc = ib_post_send(ia->ri_id->qp, &frmr_wr, &bad_wr);
1488 dprintk("RPC: %s: failed ib_post_send for register,"
1489 " status %i\n", __func__, rc);
1491 rpcrdma_unmap_one(ia, --seg);
1493 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1494 seg1->mr_base = seg1->mr_dma + pageoff;
1503 rpcrdma_deregister_frmr_external(struct rpcrdma_mr_seg *seg,
1504 struct rpcrdma_ia *ia, struct rpcrdma_xprt *r_xprt)
1506 struct rpcrdma_mr_seg *seg1 = seg;
1507 struct ib_send_wr invalidate_wr, *bad_wr;
1510 while (seg1->mr_nsegs--)
1511 rpcrdma_unmap_one(ia, seg++);
1513 memset(&invalidate_wr, 0, sizeof invalidate_wr);
1514 invalidate_wr.opcode = IB_WR_LOCAL_INV;
1515 invalidate_wr.send_flags = 0; /* unsignaled */
1516 invalidate_wr.ex.invalidate_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1517 DECR_CQCOUNT(&r_xprt->rx_ep);
1519 rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
1521 dprintk("RPC: %s: failed ib_post_send for invalidate,"
1522 " status %i\n", __func__, rc);
1527 rpcrdma_register_fmr_external(struct rpcrdma_mr_seg *seg,
1528 int *nsegs, int writing, struct rpcrdma_ia *ia)
1530 struct rpcrdma_mr_seg *seg1 = seg;
1531 u64 physaddrs[RPCRDMA_MAX_DATA_SEGS];
1532 int len, pageoff, i, rc;
1534 pageoff = offset_in_page(seg1->mr_offset);
1535 seg1->mr_offset -= pageoff; /* start of page */
1536 seg1->mr_len += pageoff;
1538 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1539 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1540 for (i = 0; i < *nsegs;) {
1541 rpcrdma_map_one(ia, seg, writing);
1542 physaddrs[i] = seg->mr_dma;
1546 /* Check for holes */
1547 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1548 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1551 rc = ib_map_phys_fmr(seg1->mr_chunk.rl_mw->r.fmr,
1552 physaddrs, i, seg1->mr_dma);
1554 dprintk("RPC: %s: failed ib_map_phys_fmr "
1555 "%u@0x%llx+%i (%d)... status %i\n", __func__,
1556 len, (unsigned long long)seg1->mr_dma,
1559 rpcrdma_unmap_one(ia, --seg);
1561 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.fmr->rkey;
1562 seg1->mr_base = seg1->mr_dma + pageoff;
1571 rpcrdma_deregister_fmr_external(struct rpcrdma_mr_seg *seg,
1572 struct rpcrdma_ia *ia)
1574 struct rpcrdma_mr_seg *seg1 = seg;
1578 list_add(&seg1->mr_chunk.rl_mw->r.fmr->list, &l);
1579 rc = ib_unmap_fmr(&l);
1580 while (seg1->mr_nsegs--)
1581 rpcrdma_unmap_one(ia, seg++);
1583 dprintk("RPC: %s: failed ib_unmap_fmr,"
1584 " status %i\n", __func__, rc);
1589 rpcrdma_register_memwin_external(struct rpcrdma_mr_seg *seg,
1590 int *nsegs, int writing, struct rpcrdma_ia *ia,
1591 struct rpcrdma_xprt *r_xprt)
1593 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1594 IB_ACCESS_REMOTE_READ);
1595 struct ib_mw_bind param;
1599 rpcrdma_map_one(ia, seg, writing);
1600 param.mr = ia->ri_bind_mem;
1601 param.wr_id = 0ULL; /* no send cookie */
1602 param.addr = seg->mr_dma;
1603 param.length = seg->mr_len;
1604 param.send_flags = 0;
1605 param.mw_access_flags = mem_priv;
1607 DECR_CQCOUNT(&r_xprt->rx_ep);
1608 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1610 dprintk("RPC: %s: failed ib_bind_mw "
1611 "%u@0x%llx status %i\n",
1612 __func__, seg->mr_len,
1613 (unsigned long long)seg->mr_dma, rc);
1614 rpcrdma_unmap_one(ia, seg);
1616 seg->mr_rkey = seg->mr_chunk.rl_mw->r.mw->rkey;
1617 seg->mr_base = param.addr;
1624 rpcrdma_deregister_memwin_external(struct rpcrdma_mr_seg *seg,
1625 struct rpcrdma_ia *ia,
1626 struct rpcrdma_xprt *r_xprt, void **r)
1628 struct ib_mw_bind param;
1632 BUG_ON(seg->mr_nsegs != 1);
1633 param.mr = ia->ri_bind_mem;
1634 param.addr = 0ULL; /* unbind */
1636 param.mw_access_flags = 0;
1638 param.wr_id = (u64) (unsigned long) *r;
1639 param.send_flags = IB_SEND_SIGNALED;
1640 INIT_CQCOUNT(&r_xprt->rx_ep);
1643 param.send_flags = 0;
1644 DECR_CQCOUNT(&r_xprt->rx_ep);
1646 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1647 rpcrdma_unmap_one(ia, seg);
1649 dprintk("RPC: %s: failed ib_(un)bind_mw,"
1650 " status %i\n", __func__, rc);
1652 *r = NULL; /* will upcall on completion */
1657 rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
1658 int *nsegs, int writing, struct rpcrdma_ia *ia)
1660 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1661 IB_ACCESS_REMOTE_READ);
1662 struct rpcrdma_mr_seg *seg1 = seg;
1663 struct ib_phys_buf ipb[RPCRDMA_MAX_DATA_SEGS];
1666 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1667 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1668 for (len = 0, i = 0; i < *nsegs;) {
1669 rpcrdma_map_one(ia, seg, writing);
1670 ipb[i].addr = seg->mr_dma;
1671 ipb[i].size = seg->mr_len;
1675 /* Check for holes */
1676 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1677 offset_in_page((seg-1)->mr_offset+(seg-1)->mr_len))
1680 seg1->mr_base = seg1->mr_dma;
1681 seg1->mr_chunk.rl_mr = ib_reg_phys_mr(ia->ri_pd,
1682 ipb, i, mem_priv, &seg1->mr_base);
1683 if (IS_ERR(seg1->mr_chunk.rl_mr)) {
1684 rc = PTR_ERR(seg1->mr_chunk.rl_mr);
1685 dprintk("RPC: %s: failed ib_reg_phys_mr "
1686 "%u@0x%llx (%d)... status %i\n",
1688 (unsigned long long)seg1->mr_dma, i, rc);
1690 rpcrdma_unmap_one(ia, --seg);
1692 seg1->mr_rkey = seg1->mr_chunk.rl_mr->rkey;
1701 rpcrdma_deregister_default_external(struct rpcrdma_mr_seg *seg,
1702 struct rpcrdma_ia *ia)
1704 struct rpcrdma_mr_seg *seg1 = seg;
1707 rc = ib_dereg_mr(seg1->mr_chunk.rl_mr);
1708 seg1->mr_chunk.rl_mr = NULL;
1709 while (seg1->mr_nsegs--)
1710 rpcrdma_unmap_one(ia, seg++);
1712 dprintk("RPC: %s: failed ib_dereg_mr,"
1713 " status %i\n", __func__, rc);
1718 rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
1719 int nsegs, int writing, struct rpcrdma_xprt *r_xprt)
1721 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1724 switch (ia->ri_memreg_strategy) {
1726 #if RPCRDMA_PERSISTENT_REGISTRATION
1727 case RPCRDMA_ALLPHYSICAL:
1728 rpcrdma_map_one(ia, seg, writing);
1729 seg->mr_rkey = ia->ri_bind_mem->rkey;
1730 seg->mr_base = seg->mr_dma;
1736 /* Registration using frmr registration */
1738 rc = rpcrdma_register_frmr_external(seg, &nsegs, writing, ia, r_xprt);
1741 /* Registration using fmr memory registration */
1742 case RPCRDMA_MTHCAFMR:
1743 rc = rpcrdma_register_fmr_external(seg, &nsegs, writing, ia);
1746 /* Registration using memory windows */
1747 case RPCRDMA_MEMWINDOWS_ASYNC:
1748 case RPCRDMA_MEMWINDOWS:
1749 rc = rpcrdma_register_memwin_external(seg, &nsegs, writing, ia, r_xprt);
1752 /* Default registration each time */
1754 rc = rpcrdma_register_default_external(seg, &nsegs, writing, ia);
1764 rpcrdma_deregister_external(struct rpcrdma_mr_seg *seg,
1765 struct rpcrdma_xprt *r_xprt, void *r)
1767 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1768 int nsegs = seg->mr_nsegs, rc;
1770 switch (ia->ri_memreg_strategy) {
1772 #if RPCRDMA_PERSISTENT_REGISTRATION
1773 case RPCRDMA_ALLPHYSICAL:
1775 rpcrdma_unmap_one(ia, seg);
1781 rc = rpcrdma_deregister_frmr_external(seg, ia, r_xprt);
1784 case RPCRDMA_MTHCAFMR:
1785 rc = rpcrdma_deregister_fmr_external(seg, ia);
1788 case RPCRDMA_MEMWINDOWS_ASYNC:
1789 case RPCRDMA_MEMWINDOWS:
1790 rc = rpcrdma_deregister_memwin_external(seg, ia, r_xprt, &r);
1794 rc = rpcrdma_deregister_default_external(seg, ia);
1798 struct rpcrdma_rep *rep = r;
1799 void (*func)(struct rpcrdma_rep *) = rep->rr_func;
1800 rep->rr_func = NULL;
1801 func(rep); /* dereg done, callback now */
1807 * Prepost any receive buffer, then post send.
1809 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1812 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1813 struct rpcrdma_ep *ep,
1814 struct rpcrdma_req *req)
1816 struct ib_send_wr send_wr, *send_wr_fail;
1817 struct rpcrdma_rep *rep = req->rl_reply;
1821 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1824 req->rl_reply = NULL;
1827 send_wr.next = NULL;
1828 send_wr.wr_id = 0ULL; /* no send cookie */
1829 send_wr.sg_list = req->rl_send_iov;
1830 send_wr.num_sge = req->rl_niovs;
1831 send_wr.opcode = IB_WR_SEND;
1832 if (send_wr.num_sge == 4) /* no need to sync any pad (constant) */
1833 ib_dma_sync_single_for_device(ia->ri_id->device,
1834 req->rl_send_iov[3].addr, req->rl_send_iov[3].length,
1836 ib_dma_sync_single_for_device(ia->ri_id->device,
1837 req->rl_send_iov[1].addr, req->rl_send_iov[1].length,
1839 ib_dma_sync_single_for_device(ia->ri_id->device,
1840 req->rl_send_iov[0].addr, req->rl_send_iov[0].length,
1843 if (DECR_CQCOUNT(ep) > 0)
1844 send_wr.send_flags = 0;
1845 else { /* Provider must take a send completion every now and then */
1847 send_wr.send_flags = IB_SEND_SIGNALED;
1850 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1852 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1859 * (Re)post a receive buffer.
1862 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1863 struct rpcrdma_ep *ep,
1864 struct rpcrdma_rep *rep)
1866 struct ib_recv_wr recv_wr, *recv_wr_fail;
1869 recv_wr.next = NULL;
1870 recv_wr.wr_id = (u64) (unsigned long) rep;
1871 recv_wr.sg_list = &rep->rr_iov;
1872 recv_wr.num_sge = 1;
1874 ib_dma_sync_single_for_cpu(ia->ri_id->device,
1875 rep->rr_iov.addr, rep->rr_iov.length, DMA_BIDIRECTIONAL);
1878 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1881 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,