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/pci.h> /* for Tavor hack below */
52 #include <linux/slab.h>
54 #include "xprt_rdma.h"
61 # define RPCDBG_FACILITY RPCDBG_TRANS
69 * handle replies in tasklet context, using a single, global list
70 * rdma tasklet function -- just turn around and call the func
71 * for all replies on the list
74 static DEFINE_SPINLOCK(rpcrdma_tk_lock_g);
75 static LIST_HEAD(rpcrdma_tasklets_g);
78 rpcrdma_run_tasklet(unsigned long data)
80 struct rpcrdma_rep *rep;
81 void (*func)(struct rpcrdma_rep *);
85 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
86 while (!list_empty(&rpcrdma_tasklets_g)) {
87 rep = list_entry(rpcrdma_tasklets_g.next,
88 struct rpcrdma_rep, rr_list);
89 list_del(&rep->rr_list);
92 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
97 rpcrdma_recv_buffer_put(rep);
99 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
101 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
104 static DECLARE_TASKLET(rpcrdma_tasklet_g, rpcrdma_run_tasklet, 0UL);
107 rpcrdma_schedule_tasklet(struct rpcrdma_rep *rep)
111 spin_lock_irqsave(&rpcrdma_tk_lock_g, flags);
112 list_add_tail(&rep->rr_list, &rpcrdma_tasklets_g);
113 spin_unlock_irqrestore(&rpcrdma_tk_lock_g, flags);
114 tasklet_schedule(&rpcrdma_tasklet_g);
118 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
120 struct rpcrdma_ep *ep = context;
122 dprintk("RPC: %s: QP error %X on device %s ep %p\n",
123 __func__, event->event, event->device->name, context);
124 if (ep->rep_connected == 1) {
125 ep->rep_connected = -EIO;
127 wake_up_all(&ep->rep_connect_wait);
132 rpcrdma_cq_async_error_upcall(struct ib_event *event, void *context)
134 struct rpcrdma_ep *ep = context;
136 dprintk("RPC: %s: CQ error %X on device %s ep %p\n",
137 __func__, event->event, event->device->name, context);
138 if (ep->rep_connected == 1) {
139 ep->rep_connected = -EIO;
141 wake_up_all(&ep->rep_connect_wait);
146 void rpcrdma_event_process(struct ib_wc *wc)
148 struct rpcrdma_mw *frmr;
149 struct rpcrdma_rep *rep =
150 (struct rpcrdma_rep *)(unsigned long) wc->wr_id;
152 dprintk("RPC: %s: event rep %p status %X opcode %X length %u\n",
153 __func__, rep, wc->status, wc->opcode, wc->byte_len);
155 if (!rep) /* send or bind completion that we don't care about */
158 if (IB_WC_SUCCESS != wc->status) {
159 dprintk("RPC: %s: WC opcode %d status %X, connection lost\n",
160 __func__, wc->opcode, wc->status);
162 if (wc->opcode != IB_WC_FAST_REG_MR && wc->opcode != IB_WC_LOCAL_INV)
163 rpcrdma_schedule_tasklet(rep);
167 switch (wc->opcode) {
168 case IB_WC_FAST_REG_MR:
169 frmr = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
170 frmr->r.frmr.state = FRMR_IS_VALID;
172 case IB_WC_LOCAL_INV:
173 frmr = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
174 frmr->r.frmr.state = FRMR_IS_INVALID;
177 rep->rr_len = wc->byte_len;
178 ib_dma_sync_single_for_cpu(
179 rdmab_to_ia(rep->rr_buffer)->ri_id->device,
180 rep->rr_iov.addr, rep->rr_len, DMA_FROM_DEVICE);
181 /* Keep (only) the most recent credits, after check validity */
182 if (rep->rr_len >= 16) {
183 struct rpcrdma_msg *p =
184 (struct rpcrdma_msg *) rep->rr_base;
185 unsigned int credits = ntohl(p->rm_credit);
187 dprintk("RPC: %s: server"
188 " dropped credits to 0!\n", __func__);
191 } else if (credits > rep->rr_buffer->rb_max_requests) {
192 dprintk("RPC: %s: server"
193 " over-crediting: %d (%d)\n",
195 rep->rr_buffer->rb_max_requests);
196 credits = rep->rr_buffer->rb_max_requests;
198 atomic_set(&rep->rr_buffer->rb_credits, credits);
202 rpcrdma_schedule_tasklet(rep);
205 dprintk("RPC: %s: unexpected WC event %X\n",
206 __func__, wc->opcode);
212 rpcrdma_cq_poll(struct ib_cq *cq)
218 rc = ib_poll_cq(cq, 1, &wc);
220 dprintk("RPC: %s: ib_poll_cq failed %i\n",
227 rpcrdma_event_process(&wc);
234 * rpcrdma_cq_event_upcall
236 * This upcall handles recv, send, bind and unbind events.
237 * It is reentrant but processes single events in order to maintain
238 * ordering of receives to keep server credits.
240 * It is the responsibility of the scheduled tasklet to return
241 * recv buffers to the pool. NOTE: this affects synchronization of
242 * connection shutdown. That is, the structures required for
243 * the completion of the reply handler must remain intact until
244 * all memory has been reclaimed.
246 * Note that send events are suppressed and do not result in an upcall.
249 rpcrdma_cq_event_upcall(struct ib_cq *cq, void *context)
253 rc = rpcrdma_cq_poll(cq);
257 rc = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
259 dprintk("RPC: %s: ib_req_notify_cq failed %i\n",
268 static const char * const conn[] = {
285 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
287 struct rpcrdma_xprt *xprt = id->context;
288 struct rpcrdma_ia *ia = &xprt->rx_ia;
289 struct rpcrdma_ep *ep = &xprt->rx_ep;
291 struct sockaddr_in *addr = (struct sockaddr_in *) &ep->rep_remote_addr;
293 struct ib_qp_attr attr;
294 struct ib_qp_init_attr iattr;
297 switch (event->event) {
298 case RDMA_CM_EVENT_ADDR_RESOLVED:
299 case RDMA_CM_EVENT_ROUTE_RESOLVED:
301 complete(&ia->ri_done);
303 case RDMA_CM_EVENT_ADDR_ERROR:
304 ia->ri_async_rc = -EHOSTUNREACH;
305 dprintk("RPC: %s: CM address resolution error, ep 0x%p\n",
307 complete(&ia->ri_done);
309 case RDMA_CM_EVENT_ROUTE_ERROR:
310 ia->ri_async_rc = -ENETUNREACH;
311 dprintk("RPC: %s: CM route resolution error, ep 0x%p\n",
313 complete(&ia->ri_done);
315 case RDMA_CM_EVENT_ESTABLISHED:
317 ib_query_qp(ia->ri_id->qp, &attr,
318 IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
320 dprintk("RPC: %s: %d responder resources"
322 __func__, attr.max_dest_rd_atomic, attr.max_rd_atomic);
324 case RDMA_CM_EVENT_CONNECT_ERROR:
325 connstate = -ENOTCONN;
327 case RDMA_CM_EVENT_UNREACHABLE:
328 connstate = -ENETDOWN;
330 case RDMA_CM_EVENT_REJECTED:
331 connstate = -ECONNREFUSED;
333 case RDMA_CM_EVENT_DISCONNECTED:
334 connstate = -ECONNABORTED;
336 case RDMA_CM_EVENT_DEVICE_REMOVAL:
339 dprintk("RPC: %s: %s: %pI4:%u (ep 0x%p event 0x%x)\n",
341 (event->event <= 11) ? conn[event->event] :
342 "unknown connection error",
343 &addr->sin_addr.s_addr,
344 ntohs(addr->sin_port),
346 atomic_set(&rpcx_to_rdmax(ep->rep_xprt)->rx_buf.rb_credits, 1);
347 dprintk("RPC: %s: %sconnected\n",
348 __func__, connstate > 0 ? "" : "dis");
349 ep->rep_connected = connstate;
351 wake_up_all(&ep->rep_connect_wait);
354 dprintk("RPC: %s: unexpected CM event %d\n",
355 __func__, event->event);
360 if (connstate == 1) {
361 int ird = attr.max_dest_rd_atomic;
362 int tird = ep->rep_remote_cma.responder_resources;
363 printk(KERN_INFO "rpcrdma: connection to %pI4:%u "
364 "on %s, memreg %d slots %d ird %d%s\n",
365 &addr->sin_addr.s_addr,
366 ntohs(addr->sin_port),
367 ia->ri_id->device->name,
368 ia->ri_memreg_strategy,
369 xprt->rx_buf.rb_max_requests,
370 ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
371 } else if (connstate < 0) {
372 printk(KERN_INFO "rpcrdma: connection to %pI4:%u closed (%d)\n",
373 &addr->sin_addr.s_addr,
374 ntohs(addr->sin_port),
382 static struct rdma_cm_id *
383 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
384 struct rpcrdma_ia *ia, struct sockaddr *addr)
386 struct rdma_cm_id *id;
389 init_completion(&ia->ri_done);
391 id = rdma_create_id(rpcrdma_conn_upcall, xprt, RDMA_PS_TCP, IB_QPT_RC);
394 dprintk("RPC: %s: rdma_create_id() failed %i\n",
399 ia->ri_async_rc = -ETIMEDOUT;
400 rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
402 dprintk("RPC: %s: rdma_resolve_addr() failed %i\n",
406 wait_for_completion_interruptible_timeout(&ia->ri_done,
407 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
408 rc = ia->ri_async_rc;
412 ia->ri_async_rc = -ETIMEDOUT;
413 rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
415 dprintk("RPC: %s: rdma_resolve_route() failed %i\n",
419 wait_for_completion_interruptible_timeout(&ia->ri_done,
420 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
421 rc = ia->ri_async_rc;
433 * Drain any cq, prior to teardown.
436 rpcrdma_clean_cq(struct ib_cq *cq)
441 while (1 == ib_poll_cq(cq, 1, &wc))
445 dprintk("RPC: %s: flushed %d events (last 0x%x)\n",
446 __func__, count, wc.opcode);
450 * Exported functions.
454 * Open and initialize an Interface Adapter.
455 * o initializes fields of struct rpcrdma_ia, including
456 * interface and provider attributes and protection zone.
459 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
462 struct ib_device_attr devattr;
463 struct rpcrdma_ia *ia = &xprt->rx_ia;
465 ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
466 if (IS_ERR(ia->ri_id)) {
467 rc = PTR_ERR(ia->ri_id);
471 ia->ri_pd = ib_alloc_pd(ia->ri_id->device);
472 if (IS_ERR(ia->ri_pd)) {
473 rc = PTR_ERR(ia->ri_pd);
474 dprintk("RPC: %s: ib_alloc_pd() failed %i\n",
480 * Query the device to determine if the requested memory
481 * registration strategy is supported. If it isn't, set the
482 * strategy to a globally supported model.
484 rc = ib_query_device(ia->ri_id->device, &devattr);
486 dprintk("RPC: %s: ib_query_device failed %d\n",
491 if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY) {
492 ia->ri_have_dma_lkey = 1;
493 ia->ri_dma_lkey = ia->ri_id->device->local_dma_lkey;
497 case RPCRDMA_MEMWINDOWS:
498 case RPCRDMA_MEMWINDOWS_ASYNC:
499 if (!(devattr.device_cap_flags & IB_DEVICE_MEM_WINDOW)) {
500 dprintk("RPC: %s: MEMWINDOWS registration "
501 "specified but not supported by adapter, "
502 "using slower RPCRDMA_REGISTER\n",
504 memreg = RPCRDMA_REGISTER;
507 case RPCRDMA_MTHCAFMR:
508 if (!ia->ri_id->device->alloc_fmr) {
509 #if RPCRDMA_PERSISTENT_REGISTRATION
510 dprintk("RPC: %s: MTHCAFMR registration "
511 "specified but not supported by adapter, "
512 "using riskier RPCRDMA_ALLPHYSICAL\n",
514 memreg = RPCRDMA_ALLPHYSICAL;
516 dprintk("RPC: %s: MTHCAFMR registration "
517 "specified but not supported by adapter, "
518 "using slower RPCRDMA_REGISTER\n",
520 memreg = RPCRDMA_REGISTER;
525 /* Requires both frmr reg and local dma lkey */
526 if ((devattr.device_cap_flags &
527 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) !=
528 (IB_DEVICE_MEM_MGT_EXTENSIONS|IB_DEVICE_LOCAL_DMA_LKEY)) {
529 #if RPCRDMA_PERSISTENT_REGISTRATION
530 dprintk("RPC: %s: FRMR registration "
531 "specified but not supported by adapter, "
532 "using riskier RPCRDMA_ALLPHYSICAL\n",
534 memreg = RPCRDMA_ALLPHYSICAL;
536 dprintk("RPC: %s: FRMR registration "
537 "specified but not supported by adapter, "
538 "using slower RPCRDMA_REGISTER\n",
540 memreg = RPCRDMA_REGISTER;
547 * Optionally obtain an underlying physical identity mapping in
548 * order to do a memory window-based bind. This base registration
549 * is protected from remote access - that is enabled only by binding
550 * for the specific bytes targeted during each RPC operation, and
551 * revoked after the corresponding completion similar to a storage
555 case RPCRDMA_BOUNCEBUFFERS:
556 case RPCRDMA_REGISTER:
559 #if RPCRDMA_PERSISTENT_REGISTRATION
560 case RPCRDMA_ALLPHYSICAL:
561 mem_priv = IB_ACCESS_LOCAL_WRITE |
562 IB_ACCESS_REMOTE_WRITE |
563 IB_ACCESS_REMOTE_READ;
566 case RPCRDMA_MEMWINDOWS_ASYNC:
567 case RPCRDMA_MEMWINDOWS:
568 mem_priv = IB_ACCESS_LOCAL_WRITE |
571 case RPCRDMA_MTHCAFMR:
572 if (ia->ri_have_dma_lkey)
574 mem_priv = IB_ACCESS_LOCAL_WRITE;
576 ia->ri_bind_mem = ib_get_dma_mr(ia->ri_pd, mem_priv);
577 if (IS_ERR(ia->ri_bind_mem)) {
578 printk(KERN_ALERT "%s: ib_get_dma_mr for "
579 "phys register failed with %lX\n\t"
580 "Will continue with degraded performance\n",
581 __func__, PTR_ERR(ia->ri_bind_mem));
582 memreg = RPCRDMA_REGISTER;
583 ia->ri_bind_mem = NULL;
587 printk(KERN_ERR "%s: invalid memory registration mode %d\n",
592 dprintk("RPC: %s: memory registration strategy is %d\n",
595 /* Else will do memory reg/dereg for each chunk */
596 ia->ri_memreg_strategy = memreg;
601 ib_dealloc_pd(ia->ri_pd);
604 rdma_destroy_id(ia->ri_id);
611 * Clean up/close an IA.
612 * o if event handles and PD have been initialized, free them.
616 rpcrdma_ia_close(struct rpcrdma_ia *ia)
620 dprintk("RPC: %s: entering\n", __func__);
621 if (ia->ri_bind_mem != NULL) {
622 rc = ib_dereg_mr(ia->ri_bind_mem);
623 dprintk("RPC: %s: ib_dereg_mr returned %i\n",
626 if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
628 rdma_destroy_qp(ia->ri_id);
629 rdma_destroy_id(ia->ri_id);
632 if (ia->ri_pd != NULL && !IS_ERR(ia->ri_pd)) {
633 rc = ib_dealloc_pd(ia->ri_pd);
634 dprintk("RPC: %s: ib_dealloc_pd returned %i\n",
640 * Create unconnected endpoint.
643 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
644 struct rpcrdma_create_data_internal *cdata)
646 struct ib_device_attr devattr;
649 rc = ib_query_device(ia->ri_id->device, &devattr);
651 dprintk("RPC: %s: ib_query_device failed %d\n",
656 /* check provider's send/recv wr limits */
657 if (cdata->max_requests > devattr.max_qp_wr)
658 cdata->max_requests = devattr.max_qp_wr;
660 ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
661 ep->rep_attr.qp_context = ep;
662 /* send_cq and recv_cq initialized below */
663 ep->rep_attr.srq = NULL;
664 ep->rep_attr.cap.max_send_wr = cdata->max_requests;
665 switch (ia->ri_memreg_strategy) {
667 /* Add room for frmr register and invalidate WRs.
668 * 1. FRMR reg WR for head
669 * 2. FRMR invalidate WR for head
670 * 3. FRMR reg WR for pagelist
671 * 4. FRMR invalidate WR for pagelist
672 * 5. FRMR reg WR for tail
673 * 6. FRMR invalidate WR for tail
674 * 7. The RDMA_SEND WR
676 ep->rep_attr.cap.max_send_wr *= 7;
677 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr) {
678 cdata->max_requests = devattr.max_qp_wr / 7;
679 if (!cdata->max_requests)
681 ep->rep_attr.cap.max_send_wr = cdata->max_requests * 7;
684 case RPCRDMA_MEMWINDOWS_ASYNC:
685 case RPCRDMA_MEMWINDOWS:
686 /* Add room for mw_binds+unbinds - overkill! */
687 ep->rep_attr.cap.max_send_wr++;
688 ep->rep_attr.cap.max_send_wr *= (2 * RPCRDMA_MAX_SEGS);
689 if (ep->rep_attr.cap.max_send_wr > devattr.max_qp_wr)
695 ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
696 ep->rep_attr.cap.max_send_sge = (cdata->padding ? 4 : 2);
697 ep->rep_attr.cap.max_recv_sge = 1;
698 ep->rep_attr.cap.max_inline_data = 0;
699 ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
700 ep->rep_attr.qp_type = IB_QPT_RC;
701 ep->rep_attr.port_num = ~0;
703 dprintk("RPC: %s: requested max: dtos: send %d recv %d; "
704 "iovs: send %d recv %d\n",
706 ep->rep_attr.cap.max_send_wr,
707 ep->rep_attr.cap.max_recv_wr,
708 ep->rep_attr.cap.max_send_sge,
709 ep->rep_attr.cap.max_recv_sge);
711 /* set trigger for requesting send completion */
712 ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 /* - 1*/;
713 switch (ia->ri_memreg_strategy) {
714 case RPCRDMA_MEMWINDOWS_ASYNC:
715 case RPCRDMA_MEMWINDOWS:
716 ep->rep_cqinit -= RPCRDMA_MAX_SEGS;
721 if (ep->rep_cqinit <= 2)
725 init_waitqueue_head(&ep->rep_connect_wait);
728 * Create a single cq for receive dto and mw_bind (only ever
729 * care about unbind, really). Send completions are suppressed.
730 * Use single threaded tasklet upcalls to maintain ordering.
732 ep->rep_cq = ib_create_cq(ia->ri_id->device, rpcrdma_cq_event_upcall,
733 rpcrdma_cq_async_error_upcall, NULL,
734 ep->rep_attr.cap.max_recv_wr +
735 ep->rep_attr.cap.max_send_wr + 1, 0);
736 if (IS_ERR(ep->rep_cq)) {
737 rc = PTR_ERR(ep->rep_cq);
738 dprintk("RPC: %s: ib_create_cq failed: %i\n",
743 rc = ib_req_notify_cq(ep->rep_cq, IB_CQ_NEXT_COMP);
745 dprintk("RPC: %s: ib_req_notify_cq failed: %i\n",
750 ep->rep_attr.send_cq = ep->rep_cq;
751 ep->rep_attr.recv_cq = ep->rep_cq;
753 /* Initialize cma parameters */
755 /* RPC/RDMA does not use private data */
756 ep->rep_remote_cma.private_data = NULL;
757 ep->rep_remote_cma.private_data_len = 0;
759 /* Client offers RDMA Read but does not initiate */
760 ep->rep_remote_cma.initiator_depth = 0;
761 if (ia->ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS)
762 ep->rep_remote_cma.responder_resources = 0;
763 else if (devattr.max_qp_rd_atom > 32) /* arbitrary but <= 255 */
764 ep->rep_remote_cma.responder_resources = 32;
766 ep->rep_remote_cma.responder_resources = devattr.max_qp_rd_atom;
768 ep->rep_remote_cma.retry_count = 7;
769 ep->rep_remote_cma.flow_control = 0;
770 ep->rep_remote_cma.rnr_retry_count = 0;
775 err = ib_destroy_cq(ep->rep_cq);
777 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
786 * Disconnect and destroy endpoint. After this, the only
787 * valid operations on the ep are to free it (if dynamically
788 * allocated) or re-create it.
790 * The caller's error handling must be sure to not leak the endpoint
791 * if this function fails.
794 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
798 dprintk("RPC: %s: entering, connected is %d\n",
799 __func__, ep->rep_connected);
802 rc = rpcrdma_ep_disconnect(ep, ia);
804 dprintk("RPC: %s: rpcrdma_ep_disconnect"
805 " returned %i\n", __func__, rc);
806 rdma_destroy_qp(ia->ri_id);
807 ia->ri_id->qp = NULL;
810 /* padding - could be done in rpcrdma_buffer_destroy... */
811 if (ep->rep_pad_mr) {
812 rpcrdma_deregister_internal(ia, ep->rep_pad_mr, &ep->rep_pad);
813 ep->rep_pad_mr = NULL;
816 rpcrdma_clean_cq(ep->rep_cq);
817 rc = ib_destroy_cq(ep->rep_cq);
819 dprintk("RPC: %s: ib_destroy_cq returned %i\n",
826 * Connect unconnected endpoint.
829 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
831 struct rdma_cm_id *id;
835 if (ep->rep_connected != 0) {
836 struct rpcrdma_xprt *xprt;
838 rc = rpcrdma_ep_disconnect(ep, ia);
839 if (rc && rc != -ENOTCONN)
840 dprintk("RPC: %s: rpcrdma_ep_disconnect"
841 " status %i\n", __func__, rc);
842 rpcrdma_clean_cq(ep->rep_cq);
844 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
845 id = rpcrdma_create_id(xprt, ia,
846 (struct sockaddr *)&xprt->rx_data.addr);
851 /* TEMP TEMP TEMP - fail if new device:
852 * Deregister/remarshal *all* requests!
853 * Close and recreate adapter, pd, etc!
854 * Re-determine all attributes still sane!
855 * More stuff I haven't thought of!
858 if (ia->ri_id->device != id->device) {
859 printk("RPC: %s: can't reconnect on "
860 "different device!\n", __func__);
866 rdma_destroy_qp(ia->ri_id);
867 rdma_destroy_id(ia->ri_id);
871 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
873 dprintk("RPC: %s: rdma_create_qp failed %i\n",
878 /* XXX Tavor device performs badly with 2K MTU! */
879 if (strnicmp(ia->ri_id->device->dma_device->bus->name, "pci", 3) == 0) {
880 struct pci_dev *pcid = to_pci_dev(ia->ri_id->device->dma_device);
881 if (pcid->device == PCI_DEVICE_ID_MELLANOX_TAVOR &&
882 (pcid->vendor == PCI_VENDOR_ID_MELLANOX ||
883 pcid->vendor == PCI_VENDOR_ID_TOPSPIN)) {
884 struct ib_qp_attr attr = {
885 .path_mtu = IB_MTU_1024
887 rc = ib_modify_qp(ia->ri_id->qp, &attr, IB_QP_PATH_MTU);
891 ep->rep_connected = 0;
893 rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
895 dprintk("RPC: %s: rdma_connect() failed with %i\n",
900 wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
903 * Check state. A non-peer reject indicates no listener
904 * (ECONNREFUSED), which may be a transient state. All
905 * others indicate a transport condition which has already
906 * undergone a best-effort.
908 if (ep->rep_connected == -ECONNREFUSED &&
909 ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
910 dprintk("RPC: %s: non-peer_reject, retry\n", __func__);
913 if (ep->rep_connected <= 0) {
914 /* Sometimes, the only way to reliably connect to remote
915 * CMs is to use same nonzero values for ORD and IRD. */
916 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
917 (ep->rep_remote_cma.responder_resources == 0 ||
918 ep->rep_remote_cma.initiator_depth !=
919 ep->rep_remote_cma.responder_resources)) {
920 if (ep->rep_remote_cma.responder_resources == 0)
921 ep->rep_remote_cma.responder_resources = 1;
922 ep->rep_remote_cma.initiator_depth =
923 ep->rep_remote_cma.responder_resources;
926 rc = ep->rep_connected;
928 dprintk("RPC: %s: connected\n", __func__);
933 ep->rep_connected = rc;
938 * rpcrdma_ep_disconnect
940 * This is separate from destroy to facilitate the ability
941 * to reconnect without recreating the endpoint.
943 * This call is not reentrant, and must not be made in parallel
944 * on the same endpoint.
947 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
951 rpcrdma_clean_cq(ep->rep_cq);
952 rc = rdma_disconnect(ia->ri_id);
954 /* returns without wait if not connected */
955 wait_event_interruptible(ep->rep_connect_wait,
956 ep->rep_connected != 1);
957 dprintk("RPC: %s: after wait, %sconnected\n", __func__,
958 (ep->rep_connected == 1) ? "still " : "dis");
960 dprintk("RPC: %s: rdma_disconnect %i\n", __func__, rc);
961 ep->rep_connected = rc;
967 * Initialize buffer memory
970 rpcrdma_buffer_create(struct rpcrdma_buffer *buf, struct rpcrdma_ep *ep,
971 struct rpcrdma_ia *ia, struct rpcrdma_create_data_internal *cdata)
976 struct rpcrdma_mw *r;
978 buf->rb_max_requests = cdata->max_requests;
979 spin_lock_init(&buf->rb_lock);
980 atomic_set(&buf->rb_credits, 1);
983 * 1. arrays for send and recv pointers
984 * 2. arrays of struct rpcrdma_req to fill in pointers
985 * 3. array of struct rpcrdma_rep for replies
987 * 5. mw's, fmr's or frmr's, if any
988 * Send/recv buffers in req/rep need to be registered
991 len = buf->rb_max_requests *
992 (sizeof(struct rpcrdma_req *) + sizeof(struct rpcrdma_rep *));
993 len += cdata->padding;
994 switch (ia->ri_memreg_strategy) {
996 len += buf->rb_max_requests * RPCRDMA_MAX_SEGS *
997 sizeof(struct rpcrdma_mw);
999 case RPCRDMA_MTHCAFMR:
1000 /* TBD we are perhaps overallocating here */
1001 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
1002 sizeof(struct rpcrdma_mw);
1004 case RPCRDMA_MEMWINDOWS_ASYNC:
1005 case RPCRDMA_MEMWINDOWS:
1006 len += (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS *
1007 sizeof(struct rpcrdma_mw);
1013 /* allocate 1, 4 and 5 in one shot */
1014 p = kzalloc(len, GFP_KERNEL);
1016 dprintk("RPC: %s: req_t/rep_t/pad kzalloc(%zd) failed\n",
1021 buf->rb_pool = p; /* for freeing it later */
1023 buf->rb_send_bufs = (struct rpcrdma_req **) p;
1024 p = (char *) &buf->rb_send_bufs[buf->rb_max_requests];
1025 buf->rb_recv_bufs = (struct rpcrdma_rep **) p;
1026 p = (char *) &buf->rb_recv_bufs[buf->rb_max_requests];
1029 * Register the zeroed pad buffer, if any.
1031 if (cdata->padding) {
1032 rc = rpcrdma_register_internal(ia, p, cdata->padding,
1033 &ep->rep_pad_mr, &ep->rep_pad);
1037 p += cdata->padding;
1040 * Allocate the fmr's, or mw's for mw_bind chunk registration.
1041 * We "cycle" the mw's in order to minimize rkey reuse,
1042 * and also reduce unbind-to-bind collision.
1044 INIT_LIST_HEAD(&buf->rb_mws);
1045 r = (struct rpcrdma_mw *)p;
1046 switch (ia->ri_memreg_strategy) {
1048 for (i = buf->rb_max_requests * RPCRDMA_MAX_SEGS; i; i--) {
1049 r->r.frmr.fr_mr = ib_alloc_fast_reg_mr(ia->ri_pd,
1051 if (IS_ERR(r->r.frmr.fr_mr)) {
1052 rc = PTR_ERR(r->r.frmr.fr_mr);
1053 dprintk("RPC: %s: ib_alloc_fast_reg_mr"
1054 " failed %i\n", __func__, rc);
1058 ib_alloc_fast_reg_page_list(ia->ri_id->device,
1060 if (IS_ERR(r->r.frmr.fr_pgl)) {
1061 rc = PTR_ERR(r->r.frmr.fr_pgl);
1063 "ib_alloc_fast_reg_page_list "
1064 "failed %i\n", __func__, rc);
1067 list_add(&r->mw_list, &buf->rb_mws);
1071 case RPCRDMA_MTHCAFMR:
1072 /* TBD we are perhaps overallocating here */
1073 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1074 static struct ib_fmr_attr fa =
1075 { RPCRDMA_MAX_DATA_SEGS, 1, PAGE_SHIFT };
1076 r->r.fmr = ib_alloc_fmr(ia->ri_pd,
1077 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ,
1079 if (IS_ERR(r->r.fmr)) {
1080 rc = PTR_ERR(r->r.fmr);
1081 dprintk("RPC: %s: ib_alloc_fmr"
1082 " failed %i\n", __func__, rc);
1085 list_add(&r->mw_list, &buf->rb_mws);
1089 case RPCRDMA_MEMWINDOWS_ASYNC:
1090 case RPCRDMA_MEMWINDOWS:
1091 /* Allocate one extra request's worth, for full cycling */
1092 for (i = (buf->rb_max_requests+1) * RPCRDMA_MAX_SEGS; i; i--) {
1093 r->r.mw = ib_alloc_mw(ia->ri_pd);
1094 if (IS_ERR(r->r.mw)) {
1095 rc = PTR_ERR(r->r.mw);
1096 dprintk("RPC: %s: ib_alloc_mw"
1097 " failed %i\n", __func__, rc);
1100 list_add(&r->mw_list, &buf->rb_mws);
1109 * Allocate/init the request/reply buffers. Doing this
1110 * using kmalloc for now -- one for each buf.
1112 for (i = 0; i < buf->rb_max_requests; i++) {
1113 struct rpcrdma_req *req;
1114 struct rpcrdma_rep *rep;
1116 len = cdata->inline_wsize + sizeof(struct rpcrdma_req);
1117 /* RPC layer requests *double* size + 1K RPC_SLACK_SPACE! */
1118 /* Typical ~2400b, so rounding up saves work later */
1121 req = kmalloc(len, GFP_KERNEL);
1123 dprintk("RPC: %s: request buffer %d alloc"
1124 " failed\n", __func__, i);
1128 memset(req, 0, sizeof(struct rpcrdma_req));
1129 buf->rb_send_bufs[i] = req;
1130 buf->rb_send_bufs[i]->rl_buffer = buf;
1132 rc = rpcrdma_register_internal(ia, req->rl_base,
1133 len - offsetof(struct rpcrdma_req, rl_base),
1134 &buf->rb_send_bufs[i]->rl_handle,
1135 &buf->rb_send_bufs[i]->rl_iov);
1139 buf->rb_send_bufs[i]->rl_size = len-sizeof(struct rpcrdma_req);
1141 len = cdata->inline_rsize + sizeof(struct rpcrdma_rep);
1142 rep = kmalloc(len, GFP_KERNEL);
1144 dprintk("RPC: %s: reply buffer %d alloc failed\n",
1149 memset(rep, 0, sizeof(struct rpcrdma_rep));
1150 buf->rb_recv_bufs[i] = rep;
1151 buf->rb_recv_bufs[i]->rr_buffer = buf;
1152 init_waitqueue_head(&rep->rr_unbind);
1154 rc = rpcrdma_register_internal(ia, rep->rr_base,
1155 len - offsetof(struct rpcrdma_rep, rr_base),
1156 &buf->rb_recv_bufs[i]->rr_handle,
1157 &buf->rb_recv_bufs[i]->rr_iov);
1162 dprintk("RPC: %s: max_requests %d\n",
1163 __func__, buf->rb_max_requests);
1167 rpcrdma_buffer_destroy(buf);
1172 * Unregister and destroy buffer memory. Need to deal with
1173 * partial initialization, so it's callable from failed create.
1174 * Must be called before destroying endpoint, as registrations
1178 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1181 struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1182 struct rpcrdma_mw *r;
1184 /* clean up in reverse order from create
1185 * 1. recv mr memory (mr free, then kfree)
1186 * 1a. bind mw memory
1187 * 2. send mr memory (mr free, then kfree)
1188 * 3. padding (if any) [moved to rpcrdma_ep_destroy]
1191 dprintk("RPC: %s: entering\n", __func__);
1193 for (i = 0; i < buf->rb_max_requests; i++) {
1194 if (buf->rb_recv_bufs && buf->rb_recv_bufs[i]) {
1195 rpcrdma_deregister_internal(ia,
1196 buf->rb_recv_bufs[i]->rr_handle,
1197 &buf->rb_recv_bufs[i]->rr_iov);
1198 kfree(buf->rb_recv_bufs[i]);
1200 if (buf->rb_send_bufs && buf->rb_send_bufs[i]) {
1201 while (!list_empty(&buf->rb_mws)) {
1202 r = list_entry(buf->rb_mws.next,
1203 struct rpcrdma_mw, mw_list);
1204 list_del(&r->mw_list);
1205 switch (ia->ri_memreg_strategy) {
1207 rc = ib_dereg_mr(r->r.frmr.fr_mr);
1213 ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
1215 case RPCRDMA_MTHCAFMR:
1216 rc = ib_dealloc_fmr(r->r.fmr);
1223 case RPCRDMA_MEMWINDOWS_ASYNC:
1224 case RPCRDMA_MEMWINDOWS:
1225 rc = ib_dealloc_mw(r->r.mw);
1236 rpcrdma_deregister_internal(ia,
1237 buf->rb_send_bufs[i]->rl_handle,
1238 &buf->rb_send_bufs[i]->rl_iov);
1239 kfree(buf->rb_send_bufs[i]);
1243 kfree(buf->rb_pool);
1247 * Get a set of request/reply buffers.
1249 * Reply buffer (if needed) is attached to send buffer upon return.
1251 * rb_send_index and rb_recv_index MUST always be pointing to the
1252 * *next* available buffer (non-NULL). They are incremented after
1253 * removing buffers, and decremented *before* returning them.
1255 struct rpcrdma_req *
1256 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1258 struct rpcrdma_req *req;
1259 unsigned long flags;
1261 struct rpcrdma_mw *r;
1263 spin_lock_irqsave(&buffers->rb_lock, flags);
1264 if (buffers->rb_send_index == buffers->rb_max_requests) {
1265 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1266 dprintk("RPC: %s: out of request buffers\n", __func__);
1267 return ((struct rpcrdma_req *)NULL);
1270 req = buffers->rb_send_bufs[buffers->rb_send_index];
1271 if (buffers->rb_send_index < buffers->rb_recv_index) {
1272 dprintk("RPC: %s: %d extra receives outstanding (ok)\n",
1274 buffers->rb_recv_index - buffers->rb_send_index);
1275 req->rl_reply = NULL;
1277 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1278 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1280 buffers->rb_send_bufs[buffers->rb_send_index++] = NULL;
1281 if (!list_empty(&buffers->rb_mws)) {
1282 i = RPCRDMA_MAX_SEGS - 1;
1284 r = list_entry(buffers->rb_mws.next,
1285 struct rpcrdma_mw, mw_list);
1286 list_del(&r->mw_list);
1287 req->rl_segments[i].mr_chunk.rl_mw = r;
1290 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1295 * Put request/reply buffers back into pool.
1296 * Pre-decrement counter/array index.
1299 rpcrdma_buffer_put(struct rpcrdma_req *req)
1301 struct rpcrdma_buffer *buffers = req->rl_buffer;
1302 struct rpcrdma_ia *ia = rdmab_to_ia(buffers);
1304 unsigned long flags;
1306 BUG_ON(req->rl_nchunks != 0);
1307 spin_lock_irqsave(&buffers->rb_lock, flags);
1308 buffers->rb_send_bufs[--buffers->rb_send_index] = req;
1310 if (req->rl_reply) {
1311 buffers->rb_recv_bufs[--buffers->rb_recv_index] = req->rl_reply;
1312 init_waitqueue_head(&req->rl_reply->rr_unbind);
1313 req->rl_reply->rr_func = NULL;
1314 req->rl_reply = NULL;
1316 switch (ia->ri_memreg_strategy) {
1318 case RPCRDMA_MTHCAFMR:
1319 case RPCRDMA_MEMWINDOWS_ASYNC:
1320 case RPCRDMA_MEMWINDOWS:
1322 * Cycle mw's back in reverse order, and "spin" them.
1323 * This delays and scrambles reuse as much as possible.
1327 struct rpcrdma_mw **mw;
1328 mw = &req->rl_segments[i].mr_chunk.rl_mw;
1329 list_add_tail(&(*mw)->mw_list, &buffers->rb_mws);
1331 } while (++i < RPCRDMA_MAX_SEGS);
1332 list_add_tail(&req->rl_segments[0].mr_chunk.rl_mw->mw_list,
1334 req->rl_segments[0].mr_chunk.rl_mw = NULL;
1339 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1343 * Recover reply buffers from pool.
1344 * This happens when recovering from error conditions.
1345 * Post-increment counter/array index.
1348 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1350 struct rpcrdma_buffer *buffers = req->rl_buffer;
1351 unsigned long flags;
1353 if (req->rl_iov.length == 0) /* special case xprt_rdma_allocate() */
1354 buffers = ((struct rpcrdma_req *) buffers)->rl_buffer;
1355 spin_lock_irqsave(&buffers->rb_lock, flags);
1356 if (buffers->rb_recv_index < buffers->rb_max_requests) {
1357 req->rl_reply = buffers->rb_recv_bufs[buffers->rb_recv_index];
1358 buffers->rb_recv_bufs[buffers->rb_recv_index++] = NULL;
1360 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1364 * Put reply buffers back into pool when not attached to
1365 * request. This happens in error conditions, and when
1366 * aborting unbinds. Pre-decrement counter/array index.
1369 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1371 struct rpcrdma_buffer *buffers = rep->rr_buffer;
1372 unsigned long flags;
1374 rep->rr_func = NULL;
1375 spin_lock_irqsave(&buffers->rb_lock, flags);
1376 buffers->rb_recv_bufs[--buffers->rb_recv_index] = rep;
1377 spin_unlock_irqrestore(&buffers->rb_lock, flags);
1381 * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1385 rpcrdma_register_internal(struct rpcrdma_ia *ia, void *va, int len,
1386 struct ib_mr **mrp, struct ib_sge *iov)
1388 struct ib_phys_buf ipb;
1393 * All memory passed here was kmalloc'ed, therefore phys-contiguous.
1395 iov->addr = ib_dma_map_single(ia->ri_id->device,
1396 va, len, DMA_BIDIRECTIONAL);
1399 if (ia->ri_have_dma_lkey) {
1401 iov->lkey = ia->ri_dma_lkey;
1403 } else if (ia->ri_bind_mem != NULL) {
1405 iov->lkey = ia->ri_bind_mem->lkey;
1409 ipb.addr = iov->addr;
1410 ipb.size = iov->length;
1411 mr = ib_reg_phys_mr(ia->ri_pd, &ipb, 1,
1412 IB_ACCESS_LOCAL_WRITE, &iov->addr);
1414 dprintk("RPC: %s: phys convert: 0x%llx "
1415 "registered 0x%llx length %d\n",
1416 __func__, (unsigned long long)ipb.addr,
1417 (unsigned long long)iov->addr, len);
1422 dprintk("RPC: %s: failed with %i\n", __func__, rc);
1425 iov->lkey = mr->lkey;
1433 rpcrdma_deregister_internal(struct rpcrdma_ia *ia,
1434 struct ib_mr *mr, struct ib_sge *iov)
1438 ib_dma_unmap_single(ia->ri_id->device,
1439 iov->addr, iov->length, DMA_BIDIRECTIONAL);
1444 rc = ib_dereg_mr(mr);
1446 dprintk("RPC: %s: ib_dereg_mr failed %i\n", __func__, rc);
1451 * Wrappers for chunk registration, shared by read/write chunk code.
1455 rpcrdma_map_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg, int writing)
1457 seg->mr_dir = writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
1458 seg->mr_dmalen = seg->mr_len;
1460 seg->mr_dma = ib_dma_map_page(ia->ri_id->device,
1461 seg->mr_page, offset_in_page(seg->mr_offset),
1462 seg->mr_dmalen, seg->mr_dir);
1464 seg->mr_dma = ib_dma_map_single(ia->ri_id->device,
1466 seg->mr_dmalen, seg->mr_dir);
1467 if (ib_dma_mapping_error(ia->ri_id->device, seg->mr_dma)) {
1468 dprintk("RPC: %s: mr_dma %llx mr_offset %p mr_dma_len %zu\n",
1470 (unsigned long long)seg->mr_dma,
1471 seg->mr_offset, seg->mr_dmalen);
1476 rpcrdma_unmap_one(struct rpcrdma_ia *ia, struct rpcrdma_mr_seg *seg)
1479 ib_dma_unmap_page(ia->ri_id->device,
1480 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1482 ib_dma_unmap_single(ia->ri_id->device,
1483 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
1487 rpcrdma_register_frmr_external(struct rpcrdma_mr_seg *seg,
1488 int *nsegs, int writing, struct rpcrdma_ia *ia,
1489 struct rpcrdma_xprt *r_xprt)
1491 struct rpcrdma_mr_seg *seg1 = seg;
1492 struct ib_send_wr invalidate_wr, frmr_wr, *bad_wr, *post_wr;
1498 pageoff = offset_in_page(seg1->mr_offset);
1499 seg1->mr_offset -= pageoff; /* start of page */
1500 seg1->mr_len += pageoff;
1502 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1503 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1504 for (i = 0; i < *nsegs;) {
1505 rpcrdma_map_one(ia, seg, writing);
1506 seg1->mr_chunk.rl_mw->r.frmr.fr_pgl->page_list[i] = seg->mr_dma;
1508 BUG_ON(seg->mr_len > PAGE_SIZE);
1511 /* Check for holes */
1512 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1513 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1516 dprintk("RPC: %s: Using frmr %p to map %d segments\n",
1517 __func__, seg1->mr_chunk.rl_mw, i);
1519 if (unlikely(seg1->mr_chunk.rl_mw->r.frmr.state == FRMR_IS_VALID)) {
1520 dprintk("RPC: %s: frmr %x left valid, posting invalidate.\n",
1522 seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey);
1523 /* Invalidate before using. */
1524 memset(&invalidate_wr, 0, sizeof invalidate_wr);
1525 invalidate_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
1526 invalidate_wr.next = &frmr_wr;
1527 invalidate_wr.opcode = IB_WR_LOCAL_INV;
1528 invalidate_wr.send_flags = IB_SEND_SIGNALED;
1529 invalidate_wr.ex.invalidate_rkey =
1530 seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1531 DECR_CQCOUNT(&r_xprt->rx_ep);
1532 post_wr = &invalidate_wr;
1537 key = (u8)(seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey & 0x000000FF);
1538 ib_update_fast_reg_key(seg1->mr_chunk.rl_mw->r.frmr.fr_mr, ++key);
1540 /* Prepare FRMR WR */
1541 memset(&frmr_wr, 0, sizeof frmr_wr);
1542 frmr_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
1543 frmr_wr.opcode = IB_WR_FAST_REG_MR;
1544 frmr_wr.send_flags = IB_SEND_SIGNALED;
1545 frmr_wr.wr.fast_reg.iova_start = seg1->mr_dma;
1546 frmr_wr.wr.fast_reg.page_list = seg1->mr_chunk.rl_mw->r.frmr.fr_pgl;
1547 frmr_wr.wr.fast_reg.page_list_len = i;
1548 frmr_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
1549 frmr_wr.wr.fast_reg.length = i << PAGE_SHIFT;
1550 BUG_ON(frmr_wr.wr.fast_reg.length < len);
1551 frmr_wr.wr.fast_reg.access_flags = (writing ?
1552 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
1553 IB_ACCESS_REMOTE_READ);
1554 frmr_wr.wr.fast_reg.rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1555 DECR_CQCOUNT(&r_xprt->rx_ep);
1557 rc = ib_post_send(ia->ri_id->qp, post_wr, &bad_wr);
1560 dprintk("RPC: %s: failed ib_post_send for register,"
1561 " status %i\n", __func__, rc);
1563 rpcrdma_unmap_one(ia, --seg);
1565 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1566 seg1->mr_base = seg1->mr_dma + pageoff;
1575 rpcrdma_deregister_frmr_external(struct rpcrdma_mr_seg *seg,
1576 struct rpcrdma_ia *ia, struct rpcrdma_xprt *r_xprt)
1578 struct rpcrdma_mr_seg *seg1 = seg;
1579 struct ib_send_wr invalidate_wr, *bad_wr;
1582 while (seg1->mr_nsegs--)
1583 rpcrdma_unmap_one(ia, seg++);
1585 memset(&invalidate_wr, 0, sizeof invalidate_wr);
1586 invalidate_wr.wr_id = (unsigned long)(void *)seg1->mr_chunk.rl_mw;
1587 invalidate_wr.opcode = IB_WR_LOCAL_INV;
1588 invalidate_wr.send_flags = IB_SEND_SIGNALED;
1589 invalidate_wr.ex.invalidate_rkey = seg1->mr_chunk.rl_mw->r.frmr.fr_mr->rkey;
1590 DECR_CQCOUNT(&r_xprt->rx_ep);
1592 rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
1594 dprintk("RPC: %s: failed ib_post_send for invalidate,"
1595 " status %i\n", __func__, rc);
1600 rpcrdma_register_fmr_external(struct rpcrdma_mr_seg *seg,
1601 int *nsegs, int writing, struct rpcrdma_ia *ia)
1603 struct rpcrdma_mr_seg *seg1 = seg;
1604 u64 physaddrs[RPCRDMA_MAX_DATA_SEGS];
1605 int len, pageoff, i, rc;
1607 pageoff = offset_in_page(seg1->mr_offset);
1608 seg1->mr_offset -= pageoff; /* start of page */
1609 seg1->mr_len += pageoff;
1611 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1612 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1613 for (i = 0; i < *nsegs;) {
1614 rpcrdma_map_one(ia, seg, writing);
1615 physaddrs[i] = seg->mr_dma;
1619 /* Check for holes */
1620 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1621 offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
1624 rc = ib_map_phys_fmr(seg1->mr_chunk.rl_mw->r.fmr,
1625 physaddrs, i, seg1->mr_dma);
1627 dprintk("RPC: %s: failed ib_map_phys_fmr "
1628 "%u@0x%llx+%i (%d)... status %i\n", __func__,
1629 len, (unsigned long long)seg1->mr_dma,
1632 rpcrdma_unmap_one(ia, --seg);
1634 seg1->mr_rkey = seg1->mr_chunk.rl_mw->r.fmr->rkey;
1635 seg1->mr_base = seg1->mr_dma + pageoff;
1644 rpcrdma_deregister_fmr_external(struct rpcrdma_mr_seg *seg,
1645 struct rpcrdma_ia *ia)
1647 struct rpcrdma_mr_seg *seg1 = seg;
1651 list_add(&seg1->mr_chunk.rl_mw->r.fmr->list, &l);
1652 rc = ib_unmap_fmr(&l);
1653 while (seg1->mr_nsegs--)
1654 rpcrdma_unmap_one(ia, seg++);
1656 dprintk("RPC: %s: failed ib_unmap_fmr,"
1657 " status %i\n", __func__, rc);
1662 rpcrdma_register_memwin_external(struct rpcrdma_mr_seg *seg,
1663 int *nsegs, int writing, struct rpcrdma_ia *ia,
1664 struct rpcrdma_xprt *r_xprt)
1666 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1667 IB_ACCESS_REMOTE_READ);
1668 struct ib_mw_bind param;
1672 rpcrdma_map_one(ia, seg, writing);
1673 param.mr = ia->ri_bind_mem;
1674 param.wr_id = 0ULL; /* no send cookie */
1675 param.addr = seg->mr_dma;
1676 param.length = seg->mr_len;
1677 param.send_flags = 0;
1678 param.mw_access_flags = mem_priv;
1680 DECR_CQCOUNT(&r_xprt->rx_ep);
1681 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1683 dprintk("RPC: %s: failed ib_bind_mw "
1684 "%u@0x%llx status %i\n",
1685 __func__, seg->mr_len,
1686 (unsigned long long)seg->mr_dma, rc);
1687 rpcrdma_unmap_one(ia, seg);
1689 seg->mr_rkey = seg->mr_chunk.rl_mw->r.mw->rkey;
1690 seg->mr_base = param.addr;
1697 rpcrdma_deregister_memwin_external(struct rpcrdma_mr_seg *seg,
1698 struct rpcrdma_ia *ia,
1699 struct rpcrdma_xprt *r_xprt, void **r)
1701 struct ib_mw_bind param;
1705 BUG_ON(seg->mr_nsegs != 1);
1706 param.mr = ia->ri_bind_mem;
1707 param.addr = 0ULL; /* unbind */
1709 param.mw_access_flags = 0;
1711 param.wr_id = (u64) (unsigned long) *r;
1712 param.send_flags = IB_SEND_SIGNALED;
1713 INIT_CQCOUNT(&r_xprt->rx_ep);
1716 param.send_flags = 0;
1717 DECR_CQCOUNT(&r_xprt->rx_ep);
1719 rc = ib_bind_mw(ia->ri_id->qp, seg->mr_chunk.rl_mw->r.mw, ¶m);
1720 rpcrdma_unmap_one(ia, seg);
1722 dprintk("RPC: %s: failed ib_(un)bind_mw,"
1723 " status %i\n", __func__, rc);
1725 *r = NULL; /* will upcall on completion */
1730 rpcrdma_register_default_external(struct rpcrdma_mr_seg *seg,
1731 int *nsegs, int writing, struct rpcrdma_ia *ia)
1733 int mem_priv = (writing ? IB_ACCESS_REMOTE_WRITE :
1734 IB_ACCESS_REMOTE_READ);
1735 struct rpcrdma_mr_seg *seg1 = seg;
1736 struct ib_phys_buf ipb[RPCRDMA_MAX_DATA_SEGS];
1739 if (*nsegs > RPCRDMA_MAX_DATA_SEGS)
1740 *nsegs = RPCRDMA_MAX_DATA_SEGS;
1741 for (len = 0, i = 0; i < *nsegs;) {
1742 rpcrdma_map_one(ia, seg, writing);
1743 ipb[i].addr = seg->mr_dma;
1744 ipb[i].size = seg->mr_len;
1748 /* Check for holes */
1749 if ((i < *nsegs && offset_in_page(seg->mr_offset)) ||
1750 offset_in_page((seg-1)->mr_offset+(seg-1)->mr_len))
1753 seg1->mr_base = seg1->mr_dma;
1754 seg1->mr_chunk.rl_mr = ib_reg_phys_mr(ia->ri_pd,
1755 ipb, i, mem_priv, &seg1->mr_base);
1756 if (IS_ERR(seg1->mr_chunk.rl_mr)) {
1757 rc = PTR_ERR(seg1->mr_chunk.rl_mr);
1758 dprintk("RPC: %s: failed ib_reg_phys_mr "
1759 "%u@0x%llx (%d)... status %i\n",
1761 (unsigned long long)seg1->mr_dma, i, rc);
1763 rpcrdma_unmap_one(ia, --seg);
1765 seg1->mr_rkey = seg1->mr_chunk.rl_mr->rkey;
1774 rpcrdma_deregister_default_external(struct rpcrdma_mr_seg *seg,
1775 struct rpcrdma_ia *ia)
1777 struct rpcrdma_mr_seg *seg1 = seg;
1780 rc = ib_dereg_mr(seg1->mr_chunk.rl_mr);
1781 seg1->mr_chunk.rl_mr = NULL;
1782 while (seg1->mr_nsegs--)
1783 rpcrdma_unmap_one(ia, seg++);
1785 dprintk("RPC: %s: failed ib_dereg_mr,"
1786 " status %i\n", __func__, rc);
1791 rpcrdma_register_external(struct rpcrdma_mr_seg *seg,
1792 int nsegs, int writing, struct rpcrdma_xprt *r_xprt)
1794 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1797 switch (ia->ri_memreg_strategy) {
1799 #if RPCRDMA_PERSISTENT_REGISTRATION
1800 case RPCRDMA_ALLPHYSICAL:
1801 rpcrdma_map_one(ia, seg, writing);
1802 seg->mr_rkey = ia->ri_bind_mem->rkey;
1803 seg->mr_base = seg->mr_dma;
1809 /* Registration using frmr registration */
1811 rc = rpcrdma_register_frmr_external(seg, &nsegs, writing, ia, r_xprt);
1814 /* Registration using fmr memory registration */
1815 case RPCRDMA_MTHCAFMR:
1816 rc = rpcrdma_register_fmr_external(seg, &nsegs, writing, ia);
1819 /* Registration using memory windows */
1820 case RPCRDMA_MEMWINDOWS_ASYNC:
1821 case RPCRDMA_MEMWINDOWS:
1822 rc = rpcrdma_register_memwin_external(seg, &nsegs, writing, ia, r_xprt);
1825 /* Default registration each time */
1827 rc = rpcrdma_register_default_external(seg, &nsegs, writing, ia);
1837 rpcrdma_deregister_external(struct rpcrdma_mr_seg *seg,
1838 struct rpcrdma_xprt *r_xprt, void *r)
1840 struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1841 int nsegs = seg->mr_nsegs, rc;
1843 switch (ia->ri_memreg_strategy) {
1845 #if RPCRDMA_PERSISTENT_REGISTRATION
1846 case RPCRDMA_ALLPHYSICAL:
1848 rpcrdma_unmap_one(ia, seg);
1854 rc = rpcrdma_deregister_frmr_external(seg, ia, r_xprt);
1857 case RPCRDMA_MTHCAFMR:
1858 rc = rpcrdma_deregister_fmr_external(seg, ia);
1861 case RPCRDMA_MEMWINDOWS_ASYNC:
1862 case RPCRDMA_MEMWINDOWS:
1863 rc = rpcrdma_deregister_memwin_external(seg, ia, r_xprt, &r);
1867 rc = rpcrdma_deregister_default_external(seg, ia);
1871 struct rpcrdma_rep *rep = r;
1872 void (*func)(struct rpcrdma_rep *) = rep->rr_func;
1873 rep->rr_func = NULL;
1874 func(rep); /* dereg done, callback now */
1880 * Prepost any receive buffer, then post send.
1882 * Receive buffer is donated to hardware, reclaimed upon recv completion.
1885 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1886 struct rpcrdma_ep *ep,
1887 struct rpcrdma_req *req)
1889 struct ib_send_wr send_wr, *send_wr_fail;
1890 struct rpcrdma_rep *rep = req->rl_reply;
1894 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1897 req->rl_reply = NULL;
1900 send_wr.next = NULL;
1901 send_wr.wr_id = 0ULL; /* no send cookie */
1902 send_wr.sg_list = req->rl_send_iov;
1903 send_wr.num_sge = req->rl_niovs;
1904 send_wr.opcode = IB_WR_SEND;
1905 if (send_wr.num_sge == 4) /* no need to sync any pad (constant) */
1906 ib_dma_sync_single_for_device(ia->ri_id->device,
1907 req->rl_send_iov[3].addr, req->rl_send_iov[3].length,
1909 ib_dma_sync_single_for_device(ia->ri_id->device,
1910 req->rl_send_iov[1].addr, req->rl_send_iov[1].length,
1912 ib_dma_sync_single_for_device(ia->ri_id->device,
1913 req->rl_send_iov[0].addr, req->rl_send_iov[0].length,
1916 if (DECR_CQCOUNT(ep) > 0)
1917 send_wr.send_flags = 0;
1918 else { /* Provider must take a send completion every now and then */
1920 send_wr.send_flags = IB_SEND_SIGNALED;
1923 rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1925 dprintk("RPC: %s: ib_post_send returned %i\n", __func__,
1932 * (Re)post a receive buffer.
1935 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1936 struct rpcrdma_ep *ep,
1937 struct rpcrdma_rep *rep)
1939 struct ib_recv_wr recv_wr, *recv_wr_fail;
1942 recv_wr.next = NULL;
1943 recv_wr.wr_id = (u64) (unsigned long) rep;
1944 recv_wr.sg_list = &rep->rr_iov;
1945 recv_wr.num_sge = 1;
1947 ib_dma_sync_single_for_cpu(ia->ri_id->device,
1948 rep->rr_iov.addr, rep->rr_iov.length, DMA_BIDIRECTIONAL);
1951 rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1954 dprintk("RPC: %s: ib_post_recv returned %i\n", __func__,