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