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_TCP, 95 }; 96 97 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt) 98 { 99 struct svc_rdma_op_ctxt *ctxt; 100 101 while (1) { 102 ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL); 103 if (ctxt) 104 break; 105 schedule_timeout_uninterruptible(msecs_to_jiffies(500)); 106 } 107 ctxt->xprt = xprt; 108 INIT_LIST_HEAD(&ctxt->dto_q); 109 ctxt->count = 0; 110 ctxt->frmr = NULL; 111 atomic_inc(&xprt->sc_ctxt_used); 112 return ctxt; 113 } 114 115 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt) 116 { 117 struct svcxprt_rdma *xprt = ctxt->xprt; 118 int i; 119 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) { 120 /* 121 * Unmap the DMA addr in the SGE if the lkey matches 122 * the sc_dma_lkey, otherwise, ignore it since it is 123 * an FRMR lkey and will be unmapped later when the 124 * last WR that uses it completes. 125 */ 126 if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) { 127 atomic_dec(&xprt->sc_dma_used); 128 ib_dma_unmap_page(xprt->sc_cm_id->device, 129 ctxt->sge[i].addr, 130 ctxt->sge[i].length, 131 ctxt->direction); 132 } 133 } 134 } 135 136 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages) 137 { 138 struct svcxprt_rdma *xprt; 139 int i; 140 141 BUG_ON(!ctxt); 142 xprt = ctxt->xprt; 143 if (free_pages) 144 for (i = 0; i < ctxt->count; i++) 145 put_page(ctxt->pages[i]); 146 147 kmem_cache_free(svc_rdma_ctxt_cachep, ctxt); 148 atomic_dec(&xprt->sc_ctxt_used); 149 } 150 151 /* 152 * Temporary NFS req mappings are shared across all transport 153 * instances. These are short lived and should be bounded by the number 154 * of concurrent server threads * depth of the SQ. 155 */ 156 struct svc_rdma_req_map *svc_rdma_get_req_map(void) 157 { 158 struct svc_rdma_req_map *map; 159 while (1) { 160 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL); 161 if (map) 162 break; 163 schedule_timeout_uninterruptible(msecs_to_jiffies(500)); 164 } 165 map->count = 0; 166 return map; 167 } 168 169 void svc_rdma_put_req_map(struct svc_rdma_req_map *map) 170 { 171 kmem_cache_free(svc_rdma_map_cachep, map); 172 } 173 174 /* ib_cq event handler */ 175 static void cq_event_handler(struct ib_event *event, void *context) 176 { 177 struct svc_xprt *xprt = context; 178 dprintk("svcrdma: received CQ event id=%d, context=%p\n", 179 event->event, context); 180 set_bit(XPT_CLOSE, &xprt->xpt_flags); 181 } 182 183 /* QP event handler */ 184 static void qp_event_handler(struct ib_event *event, void *context) 185 { 186 struct svc_xprt *xprt = context; 187 188 switch (event->event) { 189 /* These are considered benign events */ 190 case IB_EVENT_PATH_MIG: 191 case IB_EVENT_COMM_EST: 192 case IB_EVENT_SQ_DRAINED: 193 case IB_EVENT_QP_LAST_WQE_REACHED: 194 dprintk("svcrdma: QP event %d received for QP=%p\n", 195 event->event, event->element.qp); 196 break; 197 /* These are considered fatal events */ 198 case IB_EVENT_PATH_MIG_ERR: 199 case IB_EVENT_QP_FATAL: 200 case IB_EVENT_QP_REQ_ERR: 201 case IB_EVENT_QP_ACCESS_ERR: 202 case IB_EVENT_DEVICE_FATAL: 203 default: 204 dprintk("svcrdma: QP ERROR event %d received for QP=%p, " 205 "closing transport\n", 206 event->event, event->element.qp); 207 set_bit(XPT_CLOSE, &xprt->xpt_flags); 208 break; 209 } 210 } 211 212 /* 213 * Data Transfer Operation Tasklet 214 * 215 * Walks a list of transports with I/O pending, removing entries as 216 * they are added to the server's I/O pending list. Two bits indicate 217 * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave 218 * spinlock that serializes access to the transport list with the RQ 219 * and SQ interrupt handlers. 220 */ 221 static void dto_tasklet_func(unsigned long data) 222 { 223 struct svcxprt_rdma *xprt; 224 unsigned long flags; 225 226 spin_lock_irqsave(&dto_lock, flags); 227 while (!list_empty(&dto_xprt_q)) { 228 xprt = list_entry(dto_xprt_q.next, 229 struct svcxprt_rdma, sc_dto_q); 230 list_del_init(&xprt->sc_dto_q); 231 spin_unlock_irqrestore(&dto_lock, flags); 232 233 rq_cq_reap(xprt); 234 sq_cq_reap(xprt); 235 236 svc_xprt_put(&xprt->sc_xprt); 237 spin_lock_irqsave(&dto_lock, flags); 238 } 239 spin_unlock_irqrestore(&dto_lock, flags); 240 } 241 242 /* 243 * Receive Queue Completion Handler 244 * 245 * Since an RQ completion handler is called on interrupt context, we 246 * need to defer the handling of the I/O to a tasklet 247 */ 248 static void rq_comp_handler(struct ib_cq *cq, void *cq_context) 249 { 250 struct svcxprt_rdma *xprt = cq_context; 251 unsigned long flags; 252 253 /* Guard against unconditional flush call for destroyed QP */ 254 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) 255 return; 256 257 /* 258 * Set the bit regardless of whether or not it's on the list 259 * because it may be on the list already due to an SQ 260 * completion. 261 */ 262 set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags); 263 264 /* 265 * If this transport is not already on the DTO transport queue, 266 * add it 267 */ 268 spin_lock_irqsave(&dto_lock, flags); 269 if (list_empty(&xprt->sc_dto_q)) { 270 svc_xprt_get(&xprt->sc_xprt); 271 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); 272 } 273 spin_unlock_irqrestore(&dto_lock, flags); 274 275 /* Tasklet does all the work to avoid irqsave locks. */ 276 tasklet_schedule(&dto_tasklet); 277 } 278 279 /* 280 * rq_cq_reap - Process the RQ CQ. 281 * 282 * Take all completing WC off the CQE and enqueue the associated DTO 283 * context on the dto_q for the transport. 284 * 285 * Note that caller must hold a transport reference. 286 */ 287 static void rq_cq_reap(struct svcxprt_rdma *xprt) 288 { 289 int ret; 290 struct ib_wc wc; 291 struct svc_rdma_op_ctxt *ctxt = NULL; 292 293 if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) 294 return; 295 296 ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP); 297 atomic_inc(&rdma_stat_rq_poll); 298 299 while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) { 300 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id; 301 ctxt->wc_status = wc.status; 302 ctxt->byte_len = wc.byte_len; 303 svc_rdma_unmap_dma(ctxt); 304 if (wc.status != IB_WC_SUCCESS) { 305 /* Close the transport */ 306 dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt); 307 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 308 svc_rdma_put_context(ctxt, 1); 309 svc_xprt_put(&xprt->sc_xprt); 310 continue; 311 } 312 spin_lock_bh(&xprt->sc_rq_dto_lock); 313 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q); 314 spin_unlock_bh(&xprt->sc_rq_dto_lock); 315 svc_xprt_put(&xprt->sc_xprt); 316 } 317 318 if (ctxt) 319 atomic_inc(&rdma_stat_rq_prod); 320 321 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 322 /* 323 * If data arrived before established event, 324 * don't enqueue. This defers RPC I/O until the 325 * RDMA connection is complete. 326 */ 327 if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags)) 328 svc_xprt_enqueue(&xprt->sc_xprt); 329 } 330 331 /* 332 * Process a completion context 333 */ 334 static void process_context(struct svcxprt_rdma *xprt, 335 struct svc_rdma_op_ctxt *ctxt) 336 { 337 svc_rdma_unmap_dma(ctxt); 338 339 switch (ctxt->wr_op) { 340 case IB_WR_SEND: 341 BUG_ON(ctxt->frmr); 342 svc_rdma_put_context(ctxt, 1); 343 break; 344 345 case IB_WR_RDMA_WRITE: 346 BUG_ON(ctxt->frmr); 347 svc_rdma_put_context(ctxt, 0); 348 break; 349 350 case IB_WR_RDMA_READ: 351 case IB_WR_RDMA_READ_WITH_INV: 352 svc_rdma_put_frmr(xprt, ctxt->frmr); 353 if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) { 354 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr; 355 BUG_ON(!read_hdr); 356 spin_lock_bh(&xprt->sc_rq_dto_lock); 357 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags); 358 list_add_tail(&read_hdr->dto_q, 359 &xprt->sc_read_complete_q); 360 spin_unlock_bh(&xprt->sc_rq_dto_lock); 361 svc_xprt_enqueue(&xprt->sc_xprt); 362 } 363 svc_rdma_put_context(ctxt, 0); 364 break; 365 366 default: 367 BUG_ON(1); 368 printk(KERN_ERR "svcrdma: unexpected completion type, " 369 "opcode=%d\n", 370 ctxt->wr_op); 371 break; 372 } 373 } 374 375 /* 376 * Send Queue Completion Handler - potentially called on interrupt context. 377 * 378 * Note that caller must hold a transport reference. 379 */ 380 static void sq_cq_reap(struct svcxprt_rdma *xprt) 381 { 382 struct svc_rdma_op_ctxt *ctxt = NULL; 383 struct ib_wc wc_a[6]; 384 struct ib_wc *wc; 385 struct ib_cq *cq = xprt->sc_sq_cq; 386 int ret; 387 388 memset(wc_a, 0, sizeof(wc_a)); 389 390 if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) 391 return; 392 393 ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP); 394 atomic_inc(&rdma_stat_sq_poll); 395 while ((ret = ib_poll_cq(cq, ARRAY_SIZE(wc_a), wc_a)) > 0) { 396 int i; 397 398 for (i = 0; i < ret; i++) { 399 wc = &wc_a[i]; 400 if (wc->status != IB_WC_SUCCESS) { 401 dprintk("svcrdma: sq wc err status %d\n", 402 wc->status); 403 404 /* Close the transport */ 405 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 406 } 407 408 /* Decrement used SQ WR count */ 409 atomic_dec(&xprt->sc_sq_count); 410 wake_up(&xprt->sc_send_wait); 411 412 ctxt = (struct svc_rdma_op_ctxt *) 413 (unsigned long)wc->wr_id; 414 if (ctxt) 415 process_context(xprt, ctxt); 416 417 svc_xprt_put(&xprt->sc_xprt); 418 } 419 } 420 421 if (ctxt) 422 atomic_inc(&rdma_stat_sq_prod); 423 } 424 425 static void sq_comp_handler(struct ib_cq *cq, void *cq_context) 426 { 427 struct svcxprt_rdma *xprt = cq_context; 428 unsigned long flags; 429 430 /* Guard against unconditional flush call for destroyed QP */ 431 if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0) 432 return; 433 434 /* 435 * Set the bit regardless of whether or not it's on the list 436 * because it may be on the list already due to an RQ 437 * completion. 438 */ 439 set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags); 440 441 /* 442 * If this transport is not already on the DTO transport queue, 443 * add it 444 */ 445 spin_lock_irqsave(&dto_lock, flags); 446 if (list_empty(&xprt->sc_dto_q)) { 447 svc_xprt_get(&xprt->sc_xprt); 448 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q); 449 } 450 spin_unlock_irqrestore(&dto_lock, flags); 451 452 /* Tasklet does all the work to avoid irqsave locks. */ 453 tasklet_schedule(&dto_tasklet); 454 } 455 456 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv, 457 int listener) 458 { 459 struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL); 460 461 if (!cma_xprt) 462 return NULL; 463 svc_xprt_init(&init_net, &svc_rdma_class, &cma_xprt->sc_xprt, serv); 464 INIT_LIST_HEAD(&cma_xprt->sc_accept_q); 465 INIT_LIST_HEAD(&cma_xprt->sc_dto_q); 466 INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q); 467 INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q); 468 INIT_LIST_HEAD(&cma_xprt->sc_frmr_q); 469 init_waitqueue_head(&cma_xprt->sc_send_wait); 470 471 spin_lock_init(&cma_xprt->sc_lock); 472 spin_lock_init(&cma_xprt->sc_rq_dto_lock); 473 spin_lock_init(&cma_xprt->sc_frmr_q_lock); 474 475 cma_xprt->sc_ord = svcrdma_ord; 476 477 cma_xprt->sc_max_req_size = svcrdma_max_req_size; 478 cma_xprt->sc_max_requests = svcrdma_max_requests; 479 cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT; 480 atomic_set(&cma_xprt->sc_sq_count, 0); 481 atomic_set(&cma_xprt->sc_ctxt_used, 0); 482 483 if (listener) 484 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags); 485 486 return cma_xprt; 487 } 488 489 struct page *svc_rdma_get_page(void) 490 { 491 struct page *page; 492 493 while ((page = alloc_page(GFP_KERNEL)) == NULL) { 494 /* If we can't get memory, wait a bit and try again */ 495 printk(KERN_INFO "svcrdma: out of memory...retrying in 1s\n"); 496 schedule_timeout_uninterruptible(msecs_to_jiffies(1000)); 497 } 498 return page; 499 } 500 501 int svc_rdma_post_recv(struct svcxprt_rdma *xprt) 502 { 503 struct ib_recv_wr recv_wr, *bad_recv_wr; 504 struct svc_rdma_op_ctxt *ctxt; 505 struct page *page; 506 dma_addr_t pa; 507 int sge_no; 508 int buflen; 509 int ret; 510 511 ctxt = svc_rdma_get_context(xprt); 512 buflen = 0; 513 ctxt->direction = DMA_FROM_DEVICE; 514 for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) { 515 BUG_ON(sge_no >= xprt->sc_max_sge); 516 page = svc_rdma_get_page(); 517 ctxt->pages[sge_no] = page; 518 pa = ib_dma_map_page(xprt->sc_cm_id->device, 519 page, 0, PAGE_SIZE, 520 DMA_FROM_DEVICE); 521 if (ib_dma_mapping_error(xprt->sc_cm_id->device, pa)) 522 goto err_put_ctxt; 523 atomic_inc(&xprt->sc_dma_used); 524 ctxt->sge[sge_no].addr = pa; 525 ctxt->sge[sge_no].length = PAGE_SIZE; 526 ctxt->sge[sge_no].lkey = xprt->sc_dma_lkey; 527 ctxt->count = sge_no + 1; 528 buflen += PAGE_SIZE; 529 } 530 recv_wr.next = NULL; 531 recv_wr.sg_list = &ctxt->sge[0]; 532 recv_wr.num_sge = ctxt->count; 533 recv_wr.wr_id = (u64)(unsigned long)ctxt; 534 535 svc_xprt_get(&xprt->sc_xprt); 536 ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr); 537 if (ret) { 538 svc_rdma_unmap_dma(ctxt); 539 svc_rdma_put_context(ctxt, 1); 540 svc_xprt_put(&xprt->sc_xprt); 541 } 542 return ret; 543 544 err_put_ctxt: 545 svc_rdma_unmap_dma(ctxt); 546 svc_rdma_put_context(ctxt, 1); 547 return -ENOMEM; 548 } 549 550 /* 551 * This function handles the CONNECT_REQUEST event on a listening 552 * endpoint. It is passed the cma_id for the _new_ connection. The context in 553 * this cma_id is inherited from the listening cma_id and is the svc_xprt 554 * structure for the listening endpoint. 555 * 556 * This function creates a new xprt for the new connection and enqueues it on 557 * the accept queue for the listent xprt. When the listen thread is kicked, it 558 * will call the recvfrom method on the listen xprt which will accept the new 559 * connection. 560 */ 561 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird) 562 { 563 struct svcxprt_rdma *listen_xprt = new_cma_id->context; 564 struct svcxprt_rdma *newxprt; 565 struct sockaddr *sa; 566 567 /* Create a new transport */ 568 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); 569 if (!newxprt) { 570 dprintk("svcrdma: failed to create new transport\n"); 571 return; 572 } 573 newxprt->sc_cm_id = new_cma_id; 574 new_cma_id->context = newxprt; 575 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", 576 newxprt, newxprt->sc_cm_id, listen_xprt); 577 578 /* Save client advertised inbound read limit for use later in accept. */ 579 newxprt->sc_ord = client_ird; 580 581 /* Set the local and remote addresses in the transport */ 582 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 583 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 584 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 585 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 586 587 /* 588 * Enqueue the new transport on the accept queue of the listening 589 * transport 590 */ 591 spin_lock_bh(&listen_xprt->sc_lock); 592 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); 593 spin_unlock_bh(&listen_xprt->sc_lock); 594 595 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 596 svc_xprt_enqueue(&listen_xprt->sc_xprt); 597 } 598 599 /* 600 * Handles events generated on the listening endpoint. These events will be 601 * either be incoming connect requests or adapter removal events. 602 */ 603 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 604 struct rdma_cm_event *event) 605 { 606 struct svcxprt_rdma *xprt = cma_id->context; 607 int ret = 0; 608 609 switch (event->event) { 610 case RDMA_CM_EVENT_CONNECT_REQUEST: 611 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 612 "event=%d\n", cma_id, cma_id->context, event->event); 613 handle_connect_req(cma_id, 614 event->param.conn.initiator_depth); 615 break; 616 617 case RDMA_CM_EVENT_ESTABLISHED: 618 /* Accept complete */ 619 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " 620 "cm_id=%p\n", xprt, cma_id); 621 break; 622 623 case RDMA_CM_EVENT_DEVICE_REMOVAL: 624 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", 625 xprt, cma_id); 626 if (xprt) 627 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 628 break; 629 630 default: 631 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 632 "event=%d\n", cma_id, event->event); 633 break; 634 } 635 636 return ret; 637 } 638 639 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 640 struct rdma_cm_event *event) 641 { 642 struct svc_xprt *xprt = cma_id->context; 643 struct svcxprt_rdma *rdma = 644 container_of(xprt, struct svcxprt_rdma, sc_xprt); 645 switch (event->event) { 646 case RDMA_CM_EVENT_ESTABLISHED: 647 /* Accept complete */ 648 svc_xprt_get(xprt); 649 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 650 "cm_id=%p\n", xprt, cma_id); 651 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 652 svc_xprt_enqueue(xprt); 653 break; 654 case RDMA_CM_EVENT_DISCONNECTED: 655 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 656 xprt, cma_id); 657 if (xprt) { 658 set_bit(XPT_CLOSE, &xprt->xpt_flags); 659 svc_xprt_enqueue(xprt); 660 svc_xprt_put(xprt); 661 } 662 break; 663 case RDMA_CM_EVENT_DEVICE_REMOVAL: 664 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 665 "event=%d\n", cma_id, xprt, event->event); 666 if (xprt) { 667 set_bit(XPT_CLOSE, &xprt->xpt_flags); 668 svc_xprt_enqueue(xprt); 669 } 670 break; 671 default: 672 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 673 "event=%d\n", cma_id, event->event); 674 break; 675 } 676 return 0; 677 } 678 679 /* 680 * Create a listening RDMA service endpoint. 681 */ 682 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 683 struct net *net, 684 struct sockaddr *sa, int salen, 685 int flags) 686 { 687 struct rdma_cm_id *listen_id; 688 struct svcxprt_rdma *cma_xprt; 689 struct svc_xprt *xprt; 690 int ret; 691 692 dprintk("svcrdma: Creating RDMA socket\n"); 693 if (sa->sa_family != AF_INET) { 694 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); 695 return ERR_PTR(-EAFNOSUPPORT); 696 } 697 cma_xprt = rdma_create_xprt(serv, 1); 698 if (!cma_xprt) 699 return ERR_PTR(-ENOMEM); 700 xprt = &cma_xprt->sc_xprt; 701 702 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP, 703 IB_QPT_RC); 704 if (IS_ERR(listen_id)) { 705 ret = PTR_ERR(listen_id); 706 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 707 goto err0; 708 } 709 710 ret = rdma_bind_addr(listen_id, sa); 711 if (ret) { 712 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 713 goto err1; 714 } 715 cma_xprt->sc_cm_id = listen_id; 716 717 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 718 if (ret) { 719 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 720 goto err1; 721 } 722 723 /* 724 * We need to use the address from the cm_id in case the 725 * caller specified 0 for the port number. 726 */ 727 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 728 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 729 730 return &cma_xprt->sc_xprt; 731 732 err1: 733 rdma_destroy_id(listen_id); 734 err0: 735 kfree(cma_xprt); 736 return ERR_PTR(ret); 737 } 738 739 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) 740 { 741 struct ib_mr *mr; 742 struct ib_fast_reg_page_list *pl; 743 struct svc_rdma_fastreg_mr *frmr; 744 745 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); 746 if (!frmr) 747 goto err; 748 749 mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); 750 if (IS_ERR(mr)) 751 goto err_free_frmr; 752 753 pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, 754 RPCSVC_MAXPAGES); 755 if (IS_ERR(pl)) 756 goto err_free_mr; 757 758 frmr->mr = mr; 759 frmr->page_list = pl; 760 INIT_LIST_HEAD(&frmr->frmr_list); 761 return frmr; 762 763 err_free_mr: 764 ib_dereg_mr(mr); 765 err_free_frmr: 766 kfree(frmr); 767 err: 768 return ERR_PTR(-ENOMEM); 769 } 770 771 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) 772 { 773 struct svc_rdma_fastreg_mr *frmr; 774 775 while (!list_empty(&xprt->sc_frmr_q)) { 776 frmr = list_entry(xprt->sc_frmr_q.next, 777 struct svc_rdma_fastreg_mr, frmr_list); 778 list_del_init(&frmr->frmr_list); 779 ib_dereg_mr(frmr->mr); 780 ib_free_fast_reg_page_list(frmr->page_list); 781 kfree(frmr); 782 } 783 } 784 785 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) 786 { 787 struct svc_rdma_fastreg_mr *frmr = NULL; 788 789 spin_lock_bh(&rdma->sc_frmr_q_lock); 790 if (!list_empty(&rdma->sc_frmr_q)) { 791 frmr = list_entry(rdma->sc_frmr_q.next, 792 struct svc_rdma_fastreg_mr, frmr_list); 793 list_del_init(&frmr->frmr_list); 794 frmr->map_len = 0; 795 frmr->page_list_len = 0; 796 } 797 spin_unlock_bh(&rdma->sc_frmr_q_lock); 798 if (frmr) 799 return frmr; 800 801 return rdma_alloc_frmr(rdma); 802 } 803 804 static void frmr_unmap_dma(struct svcxprt_rdma *xprt, 805 struct svc_rdma_fastreg_mr *frmr) 806 { 807 int page_no; 808 for (page_no = 0; page_no < frmr->page_list_len; page_no++) { 809 dma_addr_t addr = frmr->page_list->page_list[page_no]; 810 if (ib_dma_mapping_error(frmr->mr->device, addr)) 811 continue; 812 atomic_dec(&xprt->sc_dma_used); 813 ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE, 814 frmr->direction); 815 } 816 } 817 818 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, 819 struct svc_rdma_fastreg_mr *frmr) 820 { 821 if (frmr) { 822 frmr_unmap_dma(rdma, frmr); 823 spin_lock_bh(&rdma->sc_frmr_q_lock); 824 BUG_ON(!list_empty(&frmr->frmr_list)); 825 list_add(&frmr->frmr_list, &rdma->sc_frmr_q); 826 spin_unlock_bh(&rdma->sc_frmr_q_lock); 827 } 828 } 829 830 /* 831 * This is the xpo_recvfrom function for listening endpoints. Its 832 * purpose is to accept incoming connections. The CMA callback handler 833 * has already created a new transport and attached it to the new CMA 834 * ID. 835 * 836 * There is a queue of pending connections hung on the listening 837 * transport. This queue contains the new svc_xprt structure. This 838 * function takes svc_xprt structures off the accept_q and completes 839 * the connection. 840 */ 841 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 842 { 843 struct svcxprt_rdma *listen_rdma; 844 struct svcxprt_rdma *newxprt = NULL; 845 struct rdma_conn_param conn_param; 846 struct ib_qp_init_attr qp_attr; 847 struct ib_device_attr devattr; 848 int uninitialized_var(dma_mr_acc); 849 int need_dma_mr; 850 int ret; 851 int i; 852 853 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 854 clear_bit(XPT_CONN, &xprt->xpt_flags); 855 /* Get the next entry off the accept list */ 856 spin_lock_bh(&listen_rdma->sc_lock); 857 if (!list_empty(&listen_rdma->sc_accept_q)) { 858 newxprt = list_entry(listen_rdma->sc_accept_q.next, 859 struct svcxprt_rdma, sc_accept_q); 860 list_del_init(&newxprt->sc_accept_q); 861 } 862 if (!list_empty(&listen_rdma->sc_accept_q)) 863 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 864 spin_unlock_bh(&listen_rdma->sc_lock); 865 if (!newxprt) 866 return NULL; 867 868 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 869 newxprt, newxprt->sc_cm_id); 870 871 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); 872 if (ret) { 873 dprintk("svcrdma: could not query device attributes on " 874 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); 875 goto errout; 876 } 877 878 /* Qualify the transport resource defaults with the 879 * capabilities of this particular device */ 880 newxprt->sc_max_sge = min((size_t)devattr.max_sge, 881 (size_t)RPCSVC_MAXPAGES); 882 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, 883 (size_t)svcrdma_max_requests); 884 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; 885 886 /* 887 * Limit ORD based on client limit, local device limit, and 888 * configured svcrdma limit. 889 */ 890 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); 891 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); 892 893 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); 894 if (IS_ERR(newxprt->sc_pd)) { 895 dprintk("svcrdma: error creating PD for connect request\n"); 896 goto errout; 897 } 898 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, 899 sq_comp_handler, 900 cq_event_handler, 901 newxprt, 902 newxprt->sc_sq_depth, 903 0); 904 if (IS_ERR(newxprt->sc_sq_cq)) { 905 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 906 goto errout; 907 } 908 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, 909 rq_comp_handler, 910 cq_event_handler, 911 newxprt, 912 newxprt->sc_max_requests, 913 0); 914 if (IS_ERR(newxprt->sc_rq_cq)) { 915 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 916 goto errout; 917 } 918 919 memset(&qp_attr, 0, sizeof qp_attr); 920 qp_attr.event_handler = qp_event_handler; 921 qp_attr.qp_context = &newxprt->sc_xprt; 922 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; 923 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; 924 qp_attr.cap.max_send_sge = newxprt->sc_max_sge; 925 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; 926 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 927 qp_attr.qp_type = IB_QPT_RC; 928 qp_attr.send_cq = newxprt->sc_sq_cq; 929 qp_attr.recv_cq = newxprt->sc_rq_cq; 930 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" 931 " cm_id->device=%p, sc_pd->device=%p\n" 932 " cap.max_send_wr = %d\n" 933 " cap.max_recv_wr = %d\n" 934 " cap.max_send_sge = %d\n" 935 " cap.max_recv_sge = %d\n", 936 newxprt->sc_cm_id, newxprt->sc_pd, 937 newxprt->sc_cm_id->device, newxprt->sc_pd->device, 938 qp_attr.cap.max_send_wr, 939 qp_attr.cap.max_recv_wr, 940 qp_attr.cap.max_send_sge, 941 qp_attr.cap.max_recv_sge); 942 943 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 944 if (ret) { 945 /* 946 * XXX: This is a hack. We need a xx_request_qp interface 947 * that will adjust the qp_attr's with a best-effort 948 * number 949 */ 950 qp_attr.cap.max_send_sge -= 2; 951 qp_attr.cap.max_recv_sge -= 2; 952 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, 953 &qp_attr); 954 if (ret) { 955 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 956 goto errout; 957 } 958 newxprt->sc_max_sge = qp_attr.cap.max_send_sge; 959 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge; 960 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr; 961 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr; 962 } 963 newxprt->sc_qp = newxprt->sc_cm_id->qp; 964 965 /* 966 * Use the most secure set of MR resources based on the 967 * transport type and available memory management features in 968 * the device. Here's the table implemented below: 969 * 970 * Fast Global DMA Remote WR 971 * Reg LKEY MR Access 972 * Sup'd Sup'd Needed Needed 973 * 974 * IWARP N N Y Y 975 * N Y Y Y 976 * Y N Y N 977 * Y Y N - 978 * 979 * IB N N Y N 980 * N Y N - 981 * Y N Y N 982 * Y Y N - 983 * 984 * NB: iWARP requires remote write access for the data sink 985 * of an RDMA_READ. IB does not. 986 */ 987 if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 988 newxprt->sc_frmr_pg_list_len = 989 devattr.max_fast_reg_page_list_len; 990 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; 991 } 992 993 /* 994 * Determine if a DMA MR is required and if so, what privs are required 995 */ 996 switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { 997 case RDMA_TRANSPORT_IWARP: 998 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; 999 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 1000 need_dma_mr = 1; 1001 dma_mr_acc = 1002 (IB_ACCESS_LOCAL_WRITE | 1003 IB_ACCESS_REMOTE_WRITE); 1004 } else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 1005 need_dma_mr = 1; 1006 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 1007 } else 1008 need_dma_mr = 0; 1009 break; 1010 case RDMA_TRANSPORT_IB: 1011 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 1012 need_dma_mr = 1; 1013 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 1014 } else if (!(devattr.device_cap_flags & 1015 IB_DEVICE_LOCAL_DMA_LKEY)) { 1016 need_dma_mr = 1; 1017 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 1018 } else 1019 need_dma_mr = 0; 1020 break; 1021 default: 1022 goto errout; 1023 } 1024 1025 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ 1026 if (need_dma_mr) { 1027 /* Register all of physical memory */ 1028 newxprt->sc_phys_mr = 1029 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); 1030 if (IS_ERR(newxprt->sc_phys_mr)) { 1031 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", 1032 ret); 1033 goto errout; 1034 } 1035 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; 1036 } else 1037 newxprt->sc_dma_lkey = 1038 newxprt->sc_cm_id->device->local_dma_lkey; 1039 1040 /* Post receive buffers */ 1041 for (i = 0; i < newxprt->sc_max_requests; i++) { 1042 ret = svc_rdma_post_recv(newxprt); 1043 if (ret) { 1044 dprintk("svcrdma: failure posting receive buffers\n"); 1045 goto errout; 1046 } 1047 } 1048 1049 /* Swap out the handler */ 1050 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 1051 1052 /* 1053 * Arm the CQs for the SQ and RQ before accepting so we can't 1054 * miss the first message 1055 */ 1056 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); 1057 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); 1058 1059 /* Accept Connection */ 1060 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 1061 memset(&conn_param, 0, sizeof conn_param); 1062 conn_param.responder_resources = 0; 1063 conn_param.initiator_depth = newxprt->sc_ord; 1064 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 1065 if (ret) { 1066 dprintk("svcrdma: failed to accept new connection, ret=%d\n", 1067 ret); 1068 goto errout; 1069 } 1070 1071 dprintk("svcrdma: new connection %p accepted with the following " 1072 "attributes:\n" 1073 " local_ip : %pI4\n" 1074 " local_port : %d\n" 1075 " remote_ip : %pI4\n" 1076 " remote_port : %d\n" 1077 " max_sge : %d\n" 1078 " sq_depth : %d\n" 1079 " max_requests : %d\n" 1080 " ord : %d\n", 1081 newxprt, 1082 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1083 route.addr.src_addr)->sin_addr.s_addr, 1084 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1085 route.addr.src_addr)->sin_port), 1086 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1087 route.addr.dst_addr)->sin_addr.s_addr, 1088 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1089 route.addr.dst_addr)->sin_port), 1090 newxprt->sc_max_sge, 1091 newxprt->sc_sq_depth, 1092 newxprt->sc_max_requests, 1093 newxprt->sc_ord); 1094 1095 return &newxprt->sc_xprt; 1096 1097 errout: 1098 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 1099 /* Take a reference in case the DTO handler runs */ 1100 svc_xprt_get(&newxprt->sc_xprt); 1101 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 1102 ib_destroy_qp(newxprt->sc_qp); 1103 rdma_destroy_id(newxprt->sc_cm_id); 1104 /* This call to put will destroy the transport */ 1105 svc_xprt_put(&newxprt->sc_xprt); 1106 return NULL; 1107 } 1108 1109 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 1110 { 1111 } 1112 1113 /* 1114 * When connected, an svc_xprt has at least two references: 1115 * 1116 * - A reference held by the cm_id between the ESTABLISHED and 1117 * DISCONNECTED events. If the remote peer disconnected first, this 1118 * reference could be gone. 1119 * 1120 * - A reference held by the svc_recv code that called this function 1121 * as part of close processing. 1122 * 1123 * At a minimum one references should still be held. 1124 */ 1125 static void svc_rdma_detach(struct svc_xprt *xprt) 1126 { 1127 struct svcxprt_rdma *rdma = 1128 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1129 dprintk("svc: svc_rdma_detach(%p)\n", xprt); 1130 1131 /* Disconnect and flush posted WQE */ 1132 rdma_disconnect(rdma->sc_cm_id); 1133 } 1134 1135 static void __svc_rdma_free(struct work_struct *work) 1136 { 1137 struct svcxprt_rdma *rdma = 1138 container_of(work, struct svcxprt_rdma, sc_work); 1139 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); 1140 1141 /* We should only be called from kref_put */ 1142 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0); 1143 1144 /* 1145 * Destroy queued, but not processed read completions. Note 1146 * that this cleanup has to be done before destroying the 1147 * cm_id because the device ptr is needed to unmap the dma in 1148 * svc_rdma_put_context. 1149 */ 1150 while (!list_empty(&rdma->sc_read_complete_q)) { 1151 struct svc_rdma_op_ctxt *ctxt; 1152 ctxt = list_entry(rdma->sc_read_complete_q.next, 1153 struct svc_rdma_op_ctxt, 1154 dto_q); 1155 list_del_init(&ctxt->dto_q); 1156 svc_rdma_put_context(ctxt, 1); 1157 } 1158 1159 /* Destroy queued, but not processed recv completions */ 1160 while (!list_empty(&rdma->sc_rq_dto_q)) { 1161 struct svc_rdma_op_ctxt *ctxt; 1162 ctxt = list_entry(rdma->sc_rq_dto_q.next, 1163 struct svc_rdma_op_ctxt, 1164 dto_q); 1165 list_del_init(&ctxt->dto_q); 1166 svc_rdma_put_context(ctxt, 1); 1167 } 1168 1169 /* Warn if we leaked a resource or under-referenced */ 1170 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0); 1171 WARN_ON(atomic_read(&rdma->sc_dma_used) != 0); 1172 1173 /* De-allocate fastreg mr */ 1174 rdma_dealloc_frmr_q(rdma); 1175 1176 /* Destroy the QP if present (not a listener) */ 1177 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1178 ib_destroy_qp(rdma->sc_qp); 1179 1180 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 1181 ib_destroy_cq(rdma->sc_sq_cq); 1182 1183 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 1184 ib_destroy_cq(rdma->sc_rq_cq); 1185 1186 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) 1187 ib_dereg_mr(rdma->sc_phys_mr); 1188 1189 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 1190 ib_dealloc_pd(rdma->sc_pd); 1191 1192 /* Destroy the CM ID */ 1193 rdma_destroy_id(rdma->sc_cm_id); 1194 1195 kfree(rdma); 1196 } 1197 1198 static void svc_rdma_free(struct svc_xprt *xprt) 1199 { 1200 struct svcxprt_rdma *rdma = 1201 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1202 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 1203 queue_work(svc_rdma_wq, &rdma->sc_work); 1204 } 1205 1206 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 1207 { 1208 struct svcxprt_rdma *rdma = 1209 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1210 1211 /* 1212 * If there are already waiters on the SQ, 1213 * return false. 1214 */ 1215 if (waitqueue_active(&rdma->sc_send_wait)) 1216 return 0; 1217 1218 /* Otherwise return true. */ 1219 return 1; 1220 } 1221 1222 static int svc_rdma_secure_port(struct svc_rqst *rqstp) 1223 { 1224 return 1; 1225 } 1226 1227 /* 1228 * Attempt to register the kvec representing the RPC memory with the 1229 * device. 1230 * 1231 * Returns: 1232 * NULL : The device does not support fastreg or there were no more 1233 * fastreg mr. 1234 * frmr : The kvec register request was successfully posted. 1235 * <0 : An error was encountered attempting to register the kvec. 1236 */ 1237 int svc_rdma_fastreg(struct svcxprt_rdma *xprt, 1238 struct svc_rdma_fastreg_mr *frmr) 1239 { 1240 struct ib_send_wr fastreg_wr; 1241 u8 key; 1242 1243 /* Bump the key */ 1244 key = (u8)(frmr->mr->lkey & 0x000000FF); 1245 ib_update_fast_reg_key(frmr->mr, ++key); 1246 1247 /* Prepare FASTREG WR */ 1248 memset(&fastreg_wr, 0, sizeof fastreg_wr); 1249 fastreg_wr.opcode = IB_WR_FAST_REG_MR; 1250 fastreg_wr.send_flags = IB_SEND_SIGNALED; 1251 fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; 1252 fastreg_wr.wr.fast_reg.page_list = frmr->page_list; 1253 fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; 1254 fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; 1255 fastreg_wr.wr.fast_reg.length = frmr->map_len; 1256 fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; 1257 fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; 1258 return svc_rdma_send(xprt, &fastreg_wr); 1259 } 1260 1261 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) 1262 { 1263 struct ib_send_wr *bad_wr, *n_wr; 1264 int wr_count; 1265 int i; 1266 int ret; 1267 1268 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1269 return -ENOTCONN; 1270 1271 BUG_ON(wr->send_flags != IB_SEND_SIGNALED); 1272 wr_count = 1; 1273 for (n_wr = wr->next; n_wr; n_wr = n_wr->next) 1274 wr_count++; 1275 1276 /* If the SQ is full, wait until an SQ entry is available */ 1277 while (1) { 1278 spin_lock_bh(&xprt->sc_lock); 1279 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { 1280 spin_unlock_bh(&xprt->sc_lock); 1281 atomic_inc(&rdma_stat_sq_starve); 1282 1283 /* See if we can opportunistically reap SQ WR to make room */ 1284 sq_cq_reap(xprt); 1285 1286 /* Wait until SQ WR available if SQ still full */ 1287 wait_event(xprt->sc_send_wait, 1288 atomic_read(&xprt->sc_sq_count) < 1289 xprt->sc_sq_depth); 1290 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1291 return -ENOTCONN; 1292 continue; 1293 } 1294 /* Take a transport ref for each WR posted */ 1295 for (i = 0; i < wr_count; i++) 1296 svc_xprt_get(&xprt->sc_xprt); 1297 1298 /* Bump used SQ WR count and post */ 1299 atomic_add(wr_count, &xprt->sc_sq_count); 1300 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); 1301 if (ret) { 1302 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 1303 atomic_sub(wr_count, &xprt->sc_sq_count); 1304 for (i = 0; i < wr_count; i ++) 1305 svc_xprt_put(&xprt->sc_xprt); 1306 dprintk("svcrdma: failed to post SQ WR rc=%d, " 1307 "sc_sq_count=%d, sc_sq_depth=%d\n", 1308 ret, atomic_read(&xprt->sc_sq_count), 1309 xprt->sc_sq_depth); 1310 } 1311 spin_unlock_bh(&xprt->sc_lock); 1312 if (ret) 1313 wake_up(&xprt->sc_send_wait); 1314 break; 1315 } 1316 return ret; 1317 } 1318 1319 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, 1320 enum rpcrdma_errcode err) 1321 { 1322 struct ib_send_wr err_wr; 1323 struct page *p; 1324 struct svc_rdma_op_ctxt *ctxt; 1325 u32 *va; 1326 int length; 1327 int ret; 1328 1329 p = svc_rdma_get_page(); 1330 va = page_address(p); 1331 1332 /* XDR encode error */ 1333 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va); 1334 1335 ctxt = svc_rdma_get_context(xprt); 1336 ctxt->direction = DMA_FROM_DEVICE; 1337 ctxt->count = 1; 1338 ctxt->pages[0] = p; 1339 1340 /* Prepare SGE for local address */ 1341 ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device, 1342 p, 0, length, DMA_FROM_DEVICE); 1343 if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) { 1344 put_page(p); 1345 svc_rdma_put_context(ctxt, 1); 1346 return; 1347 } 1348 atomic_inc(&xprt->sc_dma_used); 1349 ctxt->sge[0].lkey = xprt->sc_dma_lkey; 1350 ctxt->sge[0].length = length; 1351 1352 /* Prepare SEND WR */ 1353 memset(&err_wr, 0, sizeof err_wr); 1354 ctxt->wr_op = IB_WR_SEND; 1355 err_wr.wr_id = (unsigned long)ctxt; 1356 err_wr.sg_list = ctxt->sge; 1357 err_wr.num_sge = 1; 1358 err_wr.opcode = IB_WR_SEND; 1359 err_wr.send_flags = IB_SEND_SIGNALED; 1360 1361 /* Post It */ 1362 ret = svc_rdma_send(xprt, &err_wr); 1363 if (ret) { 1364 dprintk("svcrdma: Error %d posting send for protocol error\n", 1365 ret); 1366 svc_rdma_unmap_dma(ctxt); 1367 svc_rdma_put_context(ctxt, 1); 1368 } 1369 } 1370