1 /* 2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved. 3 * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the BSD-type 9 * license below: 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 15 * Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 18 * Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials provided 21 * with the distribution. 22 * 23 * Neither the name of the Network Appliance, Inc. nor the names of 24 * its contributors may be used to endorse or promote products 25 * derived from this software without specific prior written 26 * permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Author: Tom Tucker <tom@opengridcomputing.com> 41 */ 42 43 #include <linux/sunrpc/svc_xprt.h> 44 #include <linux/sunrpc/debug.h> 45 #include <linux/sunrpc/rpc_rdma.h> 46 #include <linux/interrupt.h> 47 #include <linux/sched.h> 48 #include <linux/slab.h> 49 #include <linux/spinlock.h> 50 #include <linux/workqueue.h> 51 #include <rdma/ib_verbs.h> 52 #include <rdma/rdma_cm.h> 53 #include <linux/sunrpc/svc_rdma.h> 54 #include <linux/export.h> 55 #include "xprt_rdma.h" 56 57 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 58 59 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 60 struct net *net, 61 struct sockaddr *sa, int salen, 62 int flags); 63 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt); 64 static void svc_rdma_release_rqst(struct svc_rqst *); 65 static void dto_tasklet_func(unsigned long data); 66 static void svc_rdma_detach(struct svc_xprt *xprt); 67 static void svc_rdma_free(struct svc_xprt *xprt); 68 static int svc_rdma_has_wspace(struct svc_xprt *xprt); 69 static int svc_rdma_secure_port(struct svc_rqst *); 70 static void rq_cq_reap(struct svcxprt_rdma *xprt); 71 static void sq_cq_reap(struct svcxprt_rdma *xprt); 72 73 static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL); 74 static DEFINE_SPINLOCK(dto_lock); 75 static LIST_HEAD(dto_xprt_q); 76 77 static struct svc_xprt_ops svc_rdma_ops = { 78 .xpo_create = svc_rdma_create, 79 .xpo_recvfrom = svc_rdma_recvfrom, 80 .xpo_sendto = svc_rdma_sendto, 81 .xpo_release_rqst = svc_rdma_release_rqst, 82 .xpo_detach = svc_rdma_detach, 83 .xpo_free = svc_rdma_free, 84 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, 85 .xpo_has_wspace = svc_rdma_has_wspace, 86 .xpo_accept = svc_rdma_accept, 87 .xpo_secure_port = svc_rdma_secure_port, 88 }; 89 90 struct svc_xprt_class svc_rdma_class = { 91 .xcl_name = "rdma", 92 .xcl_owner = THIS_MODULE, 93 .xcl_ops = &svc_rdma_ops, 94 .xcl_max_payload = RPCSVC_MAXPAYLOAD_RDMA, 95 .xcl_ident = XPRT_TRANSPORT_RDMA, 96 }; 97 98 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt) 99 { 100 struct svc_rdma_op_ctxt *ctxt; 101 102 while (1) { 103 ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL); 104 if (ctxt) 105 break; 106 schedule_timeout_uninterruptible(msecs_to_jiffies(500)); 107 } 108 ctxt->xprt = xprt; 109 INIT_LIST_HEAD(&ctxt->dto_q); 110 ctxt->count = 0; 111 ctxt->frmr = NULL; 112 atomic_inc(&xprt->sc_ctxt_used); 113 return ctxt; 114 } 115 116 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt) 117 { 118 struct svcxprt_rdma *xprt = ctxt->xprt; 119 int i; 120 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) { 121 /* 122 * Unmap the DMA addr in the SGE if the lkey matches 123 * the sc_dma_lkey, otherwise, ignore it since it is 124 * an FRMR lkey and will be unmapped later when the 125 * last WR that uses it completes. 126 */ 127 if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) { 128 atomic_dec(&xprt->sc_dma_used); 129 ib_dma_unmap_page(xprt->sc_cm_id->device, 130 ctxt->sge[i].addr, 131 ctxt->sge[i].length, 132 ctxt->direction); 133 } 134 } 135 } 136 137 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages) 138 { 139 struct svcxprt_rdma *xprt; 140 int i; 141 142 BUG_ON(!ctxt); 143 xprt = ctxt->xprt; 144 if (free_pages) 145 for (i = 0; i < ctxt->count; i++) 146 put_page(ctxt->pages[i]); 147 148 kmem_cache_free(svc_rdma_ctxt_cachep, ctxt); 149 atomic_dec(&xprt->sc_ctxt_used); 150 } 151 152 /* 153 * Temporary NFS req mappings are shared across all transport 154 * instances. These are short lived and should be bounded by the number 155 * of concurrent server threads * depth of the SQ. 156 */ 157 struct svc_rdma_req_map *svc_rdma_get_req_map(void) 158 { 159 struct svc_rdma_req_map *map; 160 while (1) { 161 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL); 162 if (map) 163 break; 164 schedule_timeout_uninterruptible(msecs_to_jiffies(500)); 165 } 166 map->count = 0; 167 return map; 168 } 169 170 void svc_rdma_put_req_map(struct svc_rdma_req_map *map) 171 { 172 kmem_cache_free(svc_rdma_map_cachep, map); 173 } 174 175 /* ib_cq event handler */ 176 static void cq_event_handler(struct ib_event *event, void *context) 177 { 178 struct svc_xprt *xprt = context; 179 dprintk("svcrdma: received CQ event id=%d, context=%p\n", 180 event->event, context); 181 set_bit(XPT_CLOSE, &xprt->xpt_flags); 182 } 183 184 /* QP event handler */ 185 static void qp_event_handler(struct ib_event *event, void *context) 186 { 187 struct svc_xprt *xprt = context; 188 189 switch (event->event) { 190 /* These are considered benign events */ 191 case IB_EVENT_PATH_MIG: 192 case IB_EVENT_COMM_EST: 193 case IB_EVENT_SQ_DRAINED: 194 case IB_EVENT_QP_LAST_WQE_REACHED: 195 dprintk("svcrdma: QP event %d received for QP=%p\n", 196 event->event, event->element.qp); 197 break; 198 /* These are considered fatal events */ 199 case IB_EVENT_PATH_MIG_ERR: 200 case IB_EVENT_QP_FATAL: 201 case IB_EVENT_QP_REQ_ERR: 202 case IB_EVENT_QP_ACCESS_ERR: 203 case IB_EVENT_DEVICE_FATAL: 204 default: 205 dprintk("svcrdma: QP ERROR event %d received for QP=%p, " 206 "closing transport\n", 207 event->event, event->element.qp); 208 set_bit(XPT_CLOSE, &xprt->xpt_flags); 209 break; 210 } 211 } 212 213 /* 214 * Data Transfer Operation Tasklet 215 * 216 * Walks a list of transports with I/O pending, removing entries as 217 * they are added to the server's I/O pending list. Two bits indicate 218 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave 219 * spinlock that serializes access to the transport list with the RQ 220 * and SQ interrupt handlers. 221 */ 222 static void dto_tasklet_func(unsigned long data) 223 { 224 struct svcxprt_rdma *xprt; 225 unsigned long flags; 226 227 spin_lock_irqsave(&dto_lock, flags); 228 while (!list_empty(&dto_xprt_q)) { 229 xprt = list_entry(dto_xprt_q.next, 230 struct svcxprt_rdma, sc_dto_q); 231 list_del_init(&xprt->sc_dto_q); 232 spin_unlock_irqrestore(&dto_lock, flags); 233 234 rq_cq_reap(xprt); 235 sq_cq_reap(xprt); 236 237 svc_xprt_put(&xprt->sc_xprt); 238 spin_lock_irqsave(&dto_lock, flags); 239 } 240 spin_unlock_irqrestore(&dto_lock, flags); 241 } 242 243 /* 244 * Receive Queue Completion Handler 245 * 246 * Since an RQ completion handler is called on interrupt context, we 247 * need to defer the handling of the I/O to a tasklet 248 */ 249 static void rq_comp_handler(struct ib_cq *cq, void *cq_context) 250 { 251 struct svcxprt_rdma *xprt = cq_context; 252 unsigned long flags; 253 254 /* Guard against unconditional flush call for destroyed QP */ 255 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) 256 return; 257 258 /* 259 * Set the bit regardless of whether or not it's on the list 260 * because it may be on the list already due to an SQ 261 * completion. 262 */ 263 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags); 264 265 /* 266 * If this transport is not already on the DTO transport queue, 267 * add it 268 */ 269 spin_lock_irqsave(&dto_lock, flags); 270 if (list_empty(&xprt->sc_dto_q)) { 271 svc_xprt_get(&xprt->sc_xprt); 272 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); 273 } 274 spin_unlock_irqrestore(&dto_lock, flags); 275 276 /* Tasklet does all the work to avoid irqsave locks. */ 277 tasklet_schedule(&dto_tasklet); 278 } 279 280 /* 281 * rq_cq_reap - Process the RQ CQ. 282 * 283 * Take all completing WC off the CQE and enqueue the associated DTO 284 * context on the dto_q for the transport. 285 * 286 * Note that caller must hold a transport reference. 287 */ 288 static void rq_cq_reap(struct svcxprt_rdma *xprt) 289 { 290 int ret; 291 struct ib_wc wc; 292 struct svc_rdma_op_ctxt *ctxt = NULL; 293 294 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) 295 return; 296 297 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP); 298 atomic_inc(&rdma_stat_rq_poll); 299 300 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) { 301 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id; 302 ctxt->wc_status = wc.status; 303 ctxt->byte_len = wc.byte_len; 304 svc_rdma_unmap_dma(ctxt); 305 if (wc.status != IB_WC_SUCCESS) { 306 /* Close the transport */ 307 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt); 308 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 309 svc_rdma_put_context(ctxt, 1); 310 svc_xprt_put(&xprt->sc_xprt); 311 continue; 312 } 313 spin_lock_bh(&xprt->sc_rq_dto_lock); 314 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q); 315 spin_unlock_bh(&xprt->sc_rq_dto_lock); 316 svc_xprt_put(&xprt->sc_xprt); 317 } 318 319 if (ctxt) 320 atomic_inc(&rdma_stat_rq_prod); 321 322 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 323 /* 324 * If data arrived before established event, 325 * don't enqueue. This defers RPC I/O until the 326 * RDMA connection is complete. 327 */ 328 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags)) 329 svc_xprt_enqueue(&xprt->sc_xprt); 330 } 331 332 /* 333 * Process a completion context 334 */ 335 static void process_context(struct svcxprt_rdma *xprt, 336 struct svc_rdma_op_ctxt *ctxt) 337 { 338 svc_rdma_unmap_dma(ctxt); 339 340 switch (ctxt->wr_op) { 341 case IB_WR_SEND: 342 BUG_ON(ctxt->frmr); 343 svc_rdma_put_context(ctxt, 1); 344 break; 345 346 case IB_WR_RDMA_WRITE: 347 BUG_ON(ctxt->frmr); 348 svc_rdma_put_context(ctxt, 0); 349 break; 350 351 case IB_WR_RDMA_READ: 352 case IB_WR_RDMA_READ_WITH_INV: 353 svc_rdma_put_frmr(xprt, ctxt->frmr); 354 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) { 355 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr; 356 BUG_ON(!read_hdr); 357 spin_lock_bh(&xprt->sc_rq_dto_lock); 358 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 359 list_add_tail(&read_hdr->dto_q, 360 &xprt->sc_read_complete_q); 361 spin_unlock_bh(&xprt->sc_rq_dto_lock); 362 svc_xprt_enqueue(&xprt->sc_xprt); 363 } 364 svc_rdma_put_context(ctxt, 0); 365 break; 366 367 default: 368 BUG_ON(1); 369 printk(KERN_ERR "svcrdma: unexpected completion type, " 370 "opcode=%d\n", 371 ctxt->wr_op); 372 break; 373 } 374 } 375 376 /* 377 * Send Queue Completion Handler - potentially called on interrupt context. 378 * 379 * Note that caller must hold a transport reference. 380 */ 381 static void sq_cq_reap(struct svcxprt_rdma *xprt) 382 { 383 struct svc_rdma_op_ctxt *ctxt = NULL; 384 struct ib_wc wc_a[6]; 385 struct ib_wc *wc; 386 struct ib_cq *cq = xprt->sc_sq_cq; 387 int ret; 388 389 memset(wc_a, 0, sizeof(wc_a)); 390 391 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) 392 return; 393 394 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP); 395 atomic_inc(&rdma_stat_sq_poll); 396 while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) { 397 int i; 398 399 for (i = 0; i < ret; i++) { 400 wc = &wc_a[i]; 401 if (wc->status != IB_WC_SUCCESS) { 402 dprintk("svcrdma: sq wc err status %d\n", 403 wc->status); 404 405 /* Close the transport */ 406 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 407 } 408 409 /* Decrement used SQ WR count */ 410 atomic_dec(&xprt->sc_sq_count); 411 wake_up(&xprt->sc_send_wait); 412 413 ctxt = (struct svc_rdma_op_ctxt *) 414 (unsigned long)wc->wr_id; 415 if (ctxt) 416 process_context(xprt, ctxt); 417 418 svc_xprt_put(&xprt->sc_xprt); 419 } 420 } 421 422 if (ctxt) 423 atomic_inc(&rdma_stat_sq_prod); 424 } 425 426 static void sq_comp_handler(struct ib_cq *cq, void *cq_context) 427 { 428 struct svcxprt_rdma *xprt = cq_context; 429 unsigned long flags; 430 431 /* Guard against unconditional flush call for destroyed QP */ 432 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) 433 return; 434 435 /* 436 * Set the bit regardless of whether or not it's on the list 437 * because it may be on the list already due to an RQ 438 * completion. 439 */ 440 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags); 441 442 /* 443 * If this transport is not already on the DTO transport queue, 444 * add it 445 */ 446 spin_lock_irqsave(&dto_lock, flags); 447 if (list_empty(&xprt->sc_dto_q)) { 448 svc_xprt_get(&xprt->sc_xprt); 449 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); 450 } 451 spin_unlock_irqrestore(&dto_lock, flags); 452 453 /* Tasklet does all the work to avoid irqsave locks. */ 454 tasklet_schedule(&dto_tasklet); 455 } 456 457 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv, 458 int listener) 459 { 460 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); 461 462 if (!cma_xprt) 463 return NULL; 464 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv); 465 INIT_LIST_HEAD(&cma_xprt->sc_accept_q); 466 INIT_LIST_HEAD(&cma_xprt->sc_dto_q); 467 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); 468 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); 469 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q); 470 init_waitqueue_head(&cma_xprt->sc_send_wait); 471 472 spin_lock_init(&cma_xprt->sc_lock); 473 spin_lock_init(&cma_xprt->sc_rq_dto_lock); 474 spin_lock_init(&cma_xprt->sc_frmr_q_lock); 475 476 cma_xprt->sc_ord = svcrdma_ord; 477 478 cma_xprt->sc_max_req_size = svcrdma_max_req_size; 479 cma_xprt->sc_max_requests = svcrdma_max_requests; 480 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT; 481 atomic_set(&cma_xprt->sc_sq_count, 0); 482 atomic_set(&cma_xprt->sc_ctxt_used, 0); 483 484 if (listener) 485 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); 486 487 return cma_xprt; 488 } 489 490 struct page *svc_rdma_get_page(void) 491 { 492 struct page *page; 493 494 while ((page = alloc_page(GFP_KERNEL)) == NULL) { 495 /* If we can't get memory, wait a bit and try again */ 496 printk(KERN_INFO "svcrdma: out of memory...retrying in 1s\n"); 497 schedule_timeout_uninterruptible(msecs_to_jiffies(1000)); 498 } 499 return page; 500 } 501 502 int svc_rdma_post_recv(struct svcxprt_rdma *xprt) 503 { 504 struct ib_recv_wr recv_wr, *bad_recv_wr; 505 struct svc_rdma_op_ctxt *ctxt; 506 struct page *page; 507 dma_addr_t pa; 508 int sge_no; 509 int buflen; 510 int ret; 511 512 ctxt = svc_rdma_get_context(xprt); 513 buflen = 0; 514 ctxt->direction = DMA_FROM_DEVICE; 515 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) { 516 BUG_ON(sge_no >= xprt->sc_max_sge); 517 page = svc_rdma_get_page(); 518 ctxt->pages[sge_no] = page; 519 pa = ib_dma_map_page(xprt->sc_cm_id->device, 520 page, 0, PAGE_SIZE, 521 DMA_FROM_DEVICE); 522 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa)) 523 goto err_put_ctxt; 524 atomic_inc(&xprt->sc_dma_used); 525 ctxt->sge[sge_no].addr = pa; 526 ctxt->sge[sge_no].length = PAGE_SIZE; 527 ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey; 528 ctxt->count = sge_no + 1; 529 buflen += PAGE_SIZE; 530 } 531 recv_wr.next = NULL; 532 recv_wr.sg_list = &ctxt->sge[0]; 533 recv_wr.num_sge = ctxt->count; 534 recv_wr.wr_id = (u64)(unsigned long)ctxt; 535 536 svc_xprt_get(&xprt->sc_xprt); 537 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr); 538 if (ret) { 539 svc_rdma_unmap_dma(ctxt); 540 svc_rdma_put_context(ctxt, 1); 541 svc_xprt_put(&xprt->sc_xprt); 542 } 543 return ret; 544 545 err_put_ctxt: 546 svc_rdma_unmap_dma(ctxt); 547 svc_rdma_put_context(ctxt, 1); 548 return -ENOMEM; 549 } 550 551 /* 552 * This function handles the CONNECT_REQUEST event on a listening 553 * endpoint. It is passed the cma_id for the _new_ connection. The context in 554 * this cma_id is inherited from the listening cma_id and is the svc_xprt 555 * structure for the listening endpoint. 556 * 557 * This function creates a new xprt for the new connection and enqueues it on 558 * the accept queue for the listent xprt. When the listen thread is kicked, it 559 * will call the recvfrom method on the listen xprt which will accept the new 560 * connection. 561 */ 562 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird) 563 { 564 struct svcxprt_rdma *listen_xprt = new_cma_id->context; 565 struct svcxprt_rdma *newxprt; 566 struct sockaddr *sa; 567 568 /* Create a new transport */ 569 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); 570 if (!newxprt) { 571 dprintk("svcrdma: failed to create new transport\n"); 572 return; 573 } 574 newxprt->sc_cm_id = new_cma_id; 575 new_cma_id->context = newxprt; 576 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", 577 newxprt, newxprt->sc_cm_id, listen_xprt); 578 579 /* Save client advertised inbound read limit for use later in accept. */ 580 newxprt->sc_ord = client_ird; 581 582 /* Set the local and remote addresses in the transport */ 583 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 584 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 585 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 586 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 587 588 /* 589 * Enqueue the new transport on the accept queue of the listening 590 * transport 591 */ 592 spin_lock_bh(&listen_xprt->sc_lock); 593 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); 594 spin_unlock_bh(&listen_xprt->sc_lock); 595 596 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 597 svc_xprt_enqueue(&listen_xprt->sc_xprt); 598 } 599 600 /* 601 * Handles events generated on the listening endpoint. These events will be 602 * either be incoming connect requests or adapter removal events. 603 */ 604 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 605 struct rdma_cm_event *event) 606 { 607 struct svcxprt_rdma *xprt = cma_id->context; 608 int ret = 0; 609 610 switch (event->event) { 611 case RDMA_CM_EVENT_CONNECT_REQUEST: 612 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 613 "event=%d\n", cma_id, cma_id->context, event->event); 614 handle_connect_req(cma_id, 615 event->param.conn.initiator_depth); 616 break; 617 618 case RDMA_CM_EVENT_ESTABLISHED: 619 /* Accept complete */ 620 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " 621 "cm_id=%p\n", xprt, cma_id); 622 break; 623 624 case RDMA_CM_EVENT_DEVICE_REMOVAL: 625 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", 626 xprt, cma_id); 627 if (xprt) 628 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 629 break; 630 631 default: 632 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 633 "event=%d\n", cma_id, event->event); 634 break; 635 } 636 637 return ret; 638 } 639 640 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 641 struct rdma_cm_event *event) 642 { 643 struct svc_xprt *xprt = cma_id->context; 644 struct svcxprt_rdma *rdma = 645 container_of(xprt, struct svcxprt_rdma, sc_xprt); 646 switch (event->event) { 647 case RDMA_CM_EVENT_ESTABLISHED: 648 /* Accept complete */ 649 svc_xprt_get(xprt); 650 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 651 "cm_id=%p\n", xprt, cma_id); 652 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 653 svc_xprt_enqueue(xprt); 654 break; 655 case RDMA_CM_EVENT_DISCONNECTED: 656 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 657 xprt, cma_id); 658 if (xprt) { 659 set_bit(XPT_CLOSE, &xprt->xpt_flags); 660 svc_xprt_enqueue(xprt); 661 svc_xprt_put(xprt); 662 } 663 break; 664 case RDMA_CM_EVENT_DEVICE_REMOVAL: 665 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 666 "event=%d\n", cma_id, xprt, event->event); 667 if (xprt) { 668 set_bit(XPT_CLOSE, &xprt->xpt_flags); 669 svc_xprt_enqueue(xprt); 670 } 671 break; 672 default: 673 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 674 "event=%d\n", cma_id, event->event); 675 break; 676 } 677 return 0; 678 } 679 680 /* 681 * Create a listening RDMA service endpoint. 682 */ 683 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 684 struct net *net, 685 struct sockaddr *sa, int salen, 686 int flags) 687 { 688 struct rdma_cm_id *listen_id; 689 struct svcxprt_rdma *cma_xprt; 690 struct svc_xprt *xprt; 691 int ret; 692 693 dprintk("svcrdma: Creating RDMA socket\n"); 694 if (sa->sa_family != AF_INET) { 695 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); 696 return ERR_PTR(-EAFNOSUPPORT); 697 } 698 cma_xprt = rdma_create_xprt(serv, 1); 699 if (!cma_xprt) 700 return ERR_PTR(-ENOMEM); 701 xprt = &cma_xprt->sc_xprt; 702 703 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP, 704 IB_QPT_RC); 705 if (IS_ERR(listen_id)) { 706 ret = PTR_ERR(listen_id); 707 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 708 goto err0; 709 } 710 711 ret = rdma_bind_addr(listen_id, sa); 712 if (ret) { 713 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 714 goto err1; 715 } 716 cma_xprt->sc_cm_id = listen_id; 717 718 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 719 if (ret) { 720 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 721 goto err1; 722 } 723 724 /* 725 * We need to use the address from the cm_id in case the 726 * caller specified 0 for the port number. 727 */ 728 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 729 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 730 731 return &cma_xprt->sc_xprt; 732 733 err1: 734 rdma_destroy_id(listen_id); 735 err0: 736 kfree(cma_xprt); 737 return ERR_PTR(ret); 738 } 739 740 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) 741 { 742 struct ib_mr *mr; 743 struct ib_fast_reg_page_list *pl; 744 struct svc_rdma_fastreg_mr *frmr; 745 746 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); 747 if (!frmr) 748 goto err; 749 750 mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); 751 if (IS_ERR(mr)) 752 goto err_free_frmr; 753 754 pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, 755 RPCSVC_MAXPAGES); 756 if (IS_ERR(pl)) 757 goto err_free_mr; 758 759 frmr->mr = mr; 760 frmr->page_list = pl; 761 INIT_LIST_HEAD(&frmr->frmr_list); 762 return frmr; 763 764 err_free_mr: 765 ib_dereg_mr(mr); 766 err_free_frmr: 767 kfree(frmr); 768 err: 769 return ERR_PTR(-ENOMEM); 770 } 771 772 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) 773 { 774 struct svc_rdma_fastreg_mr *frmr; 775 776 while (!list_empty(&xprt->sc_frmr_q)) { 777 frmr = list_entry(xprt->sc_frmr_q.next, 778 struct svc_rdma_fastreg_mr, frmr_list); 779 list_del_init(&frmr->frmr_list); 780 ib_dereg_mr(frmr->mr); 781 ib_free_fast_reg_page_list(frmr->page_list); 782 kfree(frmr); 783 } 784 } 785 786 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) 787 { 788 struct svc_rdma_fastreg_mr *frmr = NULL; 789 790 spin_lock_bh(&rdma->sc_frmr_q_lock); 791 if (!list_empty(&rdma->sc_frmr_q)) { 792 frmr = list_entry(rdma->sc_frmr_q.next, 793 struct svc_rdma_fastreg_mr, frmr_list); 794 list_del_init(&frmr->frmr_list); 795 frmr->map_len = 0; 796 frmr->page_list_len = 0; 797 } 798 spin_unlock_bh(&rdma->sc_frmr_q_lock); 799 if (frmr) 800 return frmr; 801 802 return rdma_alloc_frmr(rdma); 803 } 804 805 static void frmr_unmap_dma(struct svcxprt_rdma *xprt, 806 struct svc_rdma_fastreg_mr *frmr) 807 { 808 int page_no; 809 for (page_no = 0; page_no < frmr->page_list_len; page_no++) { 810 dma_addr_t addr = frmr->page_list->page_list[page_no]; 811 if (ib_dma_mapping_error(frmr->mr->device, addr)) 812 continue; 813 atomic_dec(&xprt->sc_dma_used); 814 ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE, 815 frmr->direction); 816 } 817 } 818 819 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, 820 struct svc_rdma_fastreg_mr *frmr) 821 { 822 if (frmr) { 823 frmr_unmap_dma(rdma, frmr); 824 spin_lock_bh(&rdma->sc_frmr_q_lock); 825 BUG_ON(!list_empty(&frmr->frmr_list)); 826 list_add(&frmr->frmr_list, &rdma->sc_frmr_q); 827 spin_unlock_bh(&rdma->sc_frmr_q_lock); 828 } 829 } 830 831 /* 832 * This is the xpo_recvfrom function for listening endpoints. Its 833 * purpose is to accept incoming connections. The CMA callback handler 834 * has already created a new transport and attached it to the new CMA 835 * ID. 836 * 837 * There is a queue of pending connections hung on the listening 838 * transport. This queue contains the new svc_xprt structure. This 839 * function takes svc_xprt structures off the accept_q and completes 840 * the connection. 841 */ 842 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 843 { 844 struct svcxprt_rdma *listen_rdma; 845 struct svcxprt_rdma *newxprt = NULL; 846 struct rdma_conn_param conn_param; 847 struct ib_qp_init_attr qp_attr; 848 struct ib_device_attr devattr; 849 int uninitialized_var(dma_mr_acc); 850 int need_dma_mr; 851 int ret; 852 int i; 853 854 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 855 clear_bit(XPT_CONN, &xprt->xpt_flags); 856 /* Get the next entry off the accept list */ 857 spin_lock_bh(&listen_rdma->sc_lock); 858 if (!list_empty(&listen_rdma->sc_accept_q)) { 859 newxprt = list_entry(listen_rdma->sc_accept_q.next, 860 struct svcxprt_rdma, sc_accept_q); 861 list_del_init(&newxprt->sc_accept_q); 862 } 863 if (!list_empty(&listen_rdma->sc_accept_q)) 864 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 865 spin_unlock_bh(&listen_rdma->sc_lock); 866 if (!newxprt) 867 return NULL; 868 869 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 870 newxprt, newxprt->sc_cm_id); 871 872 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); 873 if (ret) { 874 dprintk("svcrdma: could not query device attributes on " 875 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); 876 goto errout; 877 } 878 879 /* Qualify the transport resource defaults with the 880 * capabilities of this particular device */ 881 newxprt->sc_max_sge = min((size_t)devattr.max_sge, 882 (size_t)RPCSVC_MAXPAGES); 883 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, 884 (size_t)svcrdma_max_requests); 885 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; 886 887 /* 888 * Limit ORD based on client limit, local device limit, and 889 * configured svcrdma limit. 890 */ 891 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); 892 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); 893 894 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); 895 if (IS_ERR(newxprt->sc_pd)) { 896 dprintk("svcrdma: error creating PD for connect request\n"); 897 goto errout; 898 } 899 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, 900 sq_comp_handler, 901 cq_event_handler, 902 newxprt, 903 newxprt->sc_sq_depth, 904 0); 905 if (IS_ERR(newxprt->sc_sq_cq)) { 906 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 907 goto errout; 908 } 909 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, 910 rq_comp_handler, 911 cq_event_handler, 912 newxprt, 913 newxprt->sc_max_requests, 914 0); 915 if (IS_ERR(newxprt->sc_rq_cq)) { 916 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 917 goto errout; 918 } 919 920 memset(&qp_attr, 0, sizeof qp_attr); 921 qp_attr.event_handler = qp_event_handler; 922 qp_attr.qp_context = &newxprt->sc_xprt; 923 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; 924 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; 925 qp_attr.cap.max_send_sge = newxprt->sc_max_sge; 926 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; 927 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 928 qp_attr.qp_type = IB_QPT_RC; 929 qp_attr.send_cq = newxprt->sc_sq_cq; 930 qp_attr.recv_cq = newxprt->sc_rq_cq; 931 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" 932 " cm_id->device=%p, sc_pd->device=%p\n" 933 " cap.max_send_wr = %d\n" 934 " cap.max_recv_wr = %d\n" 935 " cap.max_send_sge = %d\n" 936 " cap.max_recv_sge = %d\n", 937 newxprt->sc_cm_id, newxprt->sc_pd, 938 newxprt->sc_cm_id->device, newxprt->sc_pd->device, 939 qp_attr.cap.max_send_wr, 940 qp_attr.cap.max_recv_wr, 941 qp_attr.cap.max_send_sge, 942 qp_attr.cap.max_recv_sge); 943 944 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 945 if (ret) { 946 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 947 goto errout; 948 } 949 newxprt->sc_qp = newxprt->sc_cm_id->qp; 950 951 /* 952 * Use the most secure set of MR resources based on the 953 * transport type and available memory management features in 954 * the device. Here's the table implemented below: 955 * 956 * Fast Global DMA Remote WR 957 * Reg LKEY MR Access 958 * Sup'd Sup'd Needed Needed 959 * 960 * IWARP N N Y Y 961 * N Y Y Y 962 * Y N Y N 963 * Y Y N - 964 * 965 * IB N N Y N 966 * N Y N - 967 * Y N Y N 968 * Y Y N - 969 * 970 * NB: iWARP requires remote write access for the data sink 971 * of an RDMA_READ. IB does not. 972 */ 973 if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 974 newxprt->sc_frmr_pg_list_len = 975 devattr.max_fast_reg_page_list_len; 976 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; 977 } 978 979 /* 980 * Determine if a DMA MR is required and if so, what privs are required 981 */ 982 switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { 983 case RDMA_TRANSPORT_IWARP: 984 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; 985 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 986 need_dma_mr = 1; 987 dma_mr_acc = 988 (IB_ACCESS_LOCAL_WRITE | 989 IB_ACCESS_REMOTE_WRITE); 990 } else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 991 need_dma_mr = 1; 992 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 993 } else 994 need_dma_mr = 0; 995 break; 996 case RDMA_TRANSPORT_IB: 997 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 998 need_dma_mr = 1; 999 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 1000 } else if (!(devattr.device_cap_flags & 1001 IB_DEVICE_LOCAL_DMA_LKEY)) { 1002 need_dma_mr = 1; 1003 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 1004 } else 1005 need_dma_mr = 0; 1006 break; 1007 default: 1008 goto errout; 1009 } 1010 1011 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ 1012 if (need_dma_mr) { 1013 /* Register all of physical memory */ 1014 newxprt->sc_phys_mr = 1015 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); 1016 if (IS_ERR(newxprt->sc_phys_mr)) { 1017 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", 1018 ret); 1019 goto errout; 1020 } 1021 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; 1022 } else 1023 newxprt->sc_dma_lkey = 1024 newxprt->sc_cm_id->device->local_dma_lkey; 1025 1026 /* Post receive buffers */ 1027 for (i = 0; i < newxprt->sc_max_requests; i++) { 1028 ret = svc_rdma_post_recv(newxprt); 1029 if (ret) { 1030 dprintk("svcrdma: failure posting receive buffers\n"); 1031 goto errout; 1032 } 1033 } 1034 1035 /* Swap out the handler */ 1036 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 1037 1038 /* 1039 * Arm the CQs for the SQ and RQ before accepting so we can't 1040 * miss the first message 1041 */ 1042 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); 1043 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); 1044 1045 /* Accept Connection */ 1046 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 1047 memset(&conn_param, 0, sizeof conn_param); 1048 conn_param.responder_resources = 0; 1049 conn_param.initiator_depth = newxprt->sc_ord; 1050 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 1051 if (ret) { 1052 dprintk("svcrdma: failed to accept new connection, ret=%d\n", 1053 ret); 1054 goto errout; 1055 } 1056 1057 dprintk("svcrdma: new connection %p accepted with the following " 1058 "attributes:\n" 1059 " local_ip : %pI4\n" 1060 " local_port : %d\n" 1061 " remote_ip : %pI4\n" 1062 " remote_port : %d\n" 1063 " max_sge : %d\n" 1064 " sq_depth : %d\n" 1065 " max_requests : %d\n" 1066 " ord : %d\n", 1067 newxprt, 1068 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1069 route.addr.src_addr)->sin_addr.s_addr, 1070 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1071 route.addr.src_addr)->sin_port), 1072 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1073 route.addr.dst_addr)->sin_addr.s_addr, 1074 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1075 route.addr.dst_addr)->sin_port), 1076 newxprt->sc_max_sge, 1077 newxprt->sc_sq_depth, 1078 newxprt->sc_max_requests, 1079 newxprt->sc_ord); 1080 1081 return &newxprt->sc_xprt; 1082 1083 errout: 1084 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 1085 /* Take a reference in case the DTO handler runs */ 1086 svc_xprt_get(&newxprt->sc_xprt); 1087 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 1088 ib_destroy_qp(newxprt->sc_qp); 1089 rdma_destroy_id(newxprt->sc_cm_id); 1090 /* This call to put will destroy the transport */ 1091 svc_xprt_put(&newxprt->sc_xprt); 1092 return NULL; 1093 } 1094 1095 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 1096 { 1097 } 1098 1099 /* 1100 * When connected, an svc_xprt has at least two references: 1101 * 1102 * - A reference held by the cm_id between the ESTABLISHED and 1103 * DISCONNECTED events. If the remote peer disconnected first, this 1104 * reference could be gone. 1105 * 1106 * - A reference held by the svc_recv code that called this function 1107 * as part of close processing. 1108 * 1109 * At a minimum one references should still be held. 1110 */ 1111 static void svc_rdma_detach(struct svc_xprt *xprt) 1112 { 1113 struct svcxprt_rdma *rdma = 1114 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1115 dprintk("svc: svc_rdma_detach(%p)\n", xprt); 1116 1117 /* Disconnect and flush posted WQE */ 1118 rdma_disconnect(rdma->sc_cm_id); 1119 } 1120 1121 static void __svc_rdma_free(struct work_struct *work) 1122 { 1123 struct svcxprt_rdma *rdma = 1124 container_of(work, struct svcxprt_rdma, sc_work); 1125 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); 1126 1127 /* We should only be called from kref_put */ 1128 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0); 1129 1130 /* 1131 * Destroy queued, but not processed read completions. Note 1132 * that this cleanup has to be done before destroying the 1133 * cm_id because the device ptr is needed to unmap the dma in 1134 * svc_rdma_put_context. 1135 */ 1136 while (!list_empty(&rdma->sc_read_complete_q)) { 1137 struct svc_rdma_op_ctxt *ctxt; 1138 ctxt = list_entry(rdma->sc_read_complete_q.next, 1139 struct svc_rdma_op_ctxt, 1140 dto_q); 1141 list_del_init(&ctxt->dto_q); 1142 svc_rdma_put_context(ctxt, 1); 1143 } 1144 1145 /* Destroy queued, but not processed recv completions */ 1146 while (!list_empty(&rdma->sc_rq_dto_q)) { 1147 struct svc_rdma_op_ctxt *ctxt; 1148 ctxt = list_entry(rdma->sc_rq_dto_q.next, 1149 struct svc_rdma_op_ctxt, 1150 dto_q); 1151 list_del_init(&ctxt->dto_q); 1152 svc_rdma_put_context(ctxt, 1); 1153 } 1154 1155 /* Warn if we leaked a resource or under-referenced */ 1156 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0); 1157 WARN_ON(atomic_read(&rdma->sc_dma_used) != 0); 1158 1159 /* De-allocate fastreg mr */ 1160 rdma_dealloc_frmr_q(rdma); 1161 1162 /* Destroy the QP if present (not a listener) */ 1163 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1164 ib_destroy_qp(rdma->sc_qp); 1165 1166 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 1167 ib_destroy_cq(rdma->sc_sq_cq); 1168 1169 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 1170 ib_destroy_cq(rdma->sc_rq_cq); 1171 1172 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) 1173 ib_dereg_mr(rdma->sc_phys_mr); 1174 1175 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 1176 ib_dealloc_pd(rdma->sc_pd); 1177 1178 /* Destroy the CM ID */ 1179 rdma_destroy_id(rdma->sc_cm_id); 1180 1181 kfree(rdma); 1182 } 1183 1184 static void svc_rdma_free(struct svc_xprt *xprt) 1185 { 1186 struct svcxprt_rdma *rdma = 1187 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1188 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 1189 queue_work(svc_rdma_wq, &rdma->sc_work); 1190 } 1191 1192 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 1193 { 1194 struct svcxprt_rdma *rdma = 1195 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1196 1197 /* 1198 * If there are already waiters on the SQ, 1199 * return false. 1200 */ 1201 if (waitqueue_active(&rdma->sc_send_wait)) 1202 return 0; 1203 1204 /* Otherwise return true. */ 1205 return 1; 1206 } 1207 1208 static int svc_rdma_secure_port(struct svc_rqst *rqstp) 1209 { 1210 return 1; 1211 } 1212 1213 /* 1214 * Attempt to register the kvec representing the RPC memory with the 1215 * device. 1216 * 1217 * Returns: 1218 * NULL : The device does not support fastreg or there were no more 1219 * fastreg mr. 1220 * frmr : The kvec register request was successfully posted. 1221 * <0 : An error was encountered attempting to register the kvec. 1222 */ 1223 int svc_rdma_fastreg(struct svcxprt_rdma *xprt, 1224 struct svc_rdma_fastreg_mr *frmr) 1225 { 1226 struct ib_send_wr fastreg_wr; 1227 u8 key; 1228 1229 /* Bump the key */ 1230 key = (u8)(frmr->mr->lkey & 0x000000FF); 1231 ib_update_fast_reg_key(frmr->mr, ++key); 1232 1233 /* Prepare FASTREG WR */ 1234 memset(&fastreg_wr, 0, sizeof fastreg_wr); 1235 fastreg_wr.opcode = IB_WR_FAST_REG_MR; 1236 fastreg_wr.send_flags = IB_SEND_SIGNALED; 1237 fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; 1238 fastreg_wr.wr.fast_reg.page_list = frmr->page_list; 1239 fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; 1240 fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; 1241 fastreg_wr.wr.fast_reg.length = frmr->map_len; 1242 fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; 1243 fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; 1244 return svc_rdma_send(xprt, &fastreg_wr); 1245 } 1246 1247 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) 1248 { 1249 struct ib_send_wr *bad_wr, *n_wr; 1250 int wr_count; 1251 int i; 1252 int ret; 1253 1254 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1255 return -ENOTCONN; 1256 1257 BUG_ON(wr->send_flags != IB_SEND_SIGNALED); 1258 wr_count = 1; 1259 for (n_wr = wr->next; n_wr; n_wr = n_wr->next) 1260 wr_count++; 1261 1262 /* If the SQ is full, wait until an SQ entry is available */ 1263 while (1) { 1264 spin_lock_bh(&xprt->sc_lock); 1265 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { 1266 spin_unlock_bh(&xprt->sc_lock); 1267 atomic_inc(&rdma_stat_sq_starve); 1268 1269 /* See if we can opportunistically reap SQ WR to make room */ 1270 sq_cq_reap(xprt); 1271 1272 /* Wait until SQ WR available if SQ still full */ 1273 wait_event(xprt->sc_send_wait, 1274 atomic_read(&xprt->sc_sq_count) < 1275 xprt->sc_sq_depth); 1276 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1277 return -ENOTCONN; 1278 continue; 1279 } 1280 /* Take a transport ref for each WR posted */ 1281 for (i = 0; i < wr_count; i++) 1282 svc_xprt_get(&xprt->sc_xprt); 1283 1284 /* Bump used SQ WR count and post */ 1285 atomic_add(wr_count, &xprt->sc_sq_count); 1286 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); 1287 if (ret) { 1288 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 1289 atomic_sub(wr_count, &xprt->sc_sq_count); 1290 for (i = 0; i < wr_count; i ++) 1291 svc_xprt_put(&xprt->sc_xprt); 1292 dprintk("svcrdma: failed to post SQ WR rc=%d, " 1293 "sc_sq_count=%d, sc_sq_depth=%d\n", 1294 ret, atomic_read(&xprt->sc_sq_count), 1295 xprt->sc_sq_depth); 1296 } 1297 spin_unlock_bh(&xprt->sc_lock); 1298 if (ret) 1299 wake_up(&xprt->sc_send_wait); 1300 break; 1301 } 1302 return ret; 1303 } 1304 1305 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, 1306 enum rpcrdma_errcode err) 1307 { 1308 struct ib_send_wr err_wr; 1309 struct page *p; 1310 struct svc_rdma_op_ctxt *ctxt; 1311 u32 *va; 1312 int length; 1313 int ret; 1314 1315 p = svc_rdma_get_page(); 1316 va = page_address(p); 1317 1318 /* XDR encode error */ 1319 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va); 1320 1321 ctxt = svc_rdma_get_context(xprt); 1322 ctxt->direction = DMA_FROM_DEVICE; 1323 ctxt->count = 1; 1324 ctxt->pages[0] = p; 1325 1326 /* Prepare SGE for local address */ 1327 ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device, 1328 p, 0, length, DMA_FROM_DEVICE); 1329 if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) { 1330 put_page(p); 1331 svc_rdma_put_context(ctxt, 1); 1332 return; 1333 } 1334 atomic_inc(&xprt->sc_dma_used); 1335 ctxt->sge[0].lkey = xprt->sc_dma_lkey; 1336 ctxt->sge[0].length = length; 1337 1338 /* Prepare SEND WR */ 1339 memset(&err_wr, 0, sizeof err_wr); 1340 ctxt->wr_op = IB_WR_SEND; 1341 err_wr.wr_id = (unsigned long)ctxt; 1342 err_wr.sg_list = ctxt->sge; 1343 err_wr.num_sge = 1; 1344 err_wr.opcode = IB_WR_SEND; 1345 err_wr.send_flags = IB_SEND_SIGNALED; 1346 1347 /* Post It */ 1348 ret = svc_rdma_send(xprt, &err_wr); 1349 if (ret) { 1350 dprintk("svcrdma: Error %d posting send for protocol error\n", 1351 ret); 1352 svc_rdma_unmap_dma(ctxt); 1353 svc_rdma_put_context(ctxt, 1); 1354 } 1355 } 1356