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