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/interrupt.h> 46 #include <linux/sched.h> 47 #include <linux/slab.h> 48 #include <linux/spinlock.h> 49 #include <linux/workqueue.h> 50 #include <rdma/ib_verbs.h> 51 #include <rdma/rdma_cm.h> 52 #include <linux/sunrpc/svc_rdma.h> 53 #include <linux/export.h> 54 #include "xprt_rdma.h" 55 56 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 57 58 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 59 struct net *net, 60 struct sockaddr *sa, int salen, 61 int flags); 62 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt); 63 static void svc_rdma_release_rqst(struct svc_rqst *); 64 static void dto_tasklet_func(unsigned long data); 65 static void svc_rdma_detach(struct svc_xprt *xprt); 66 static void svc_rdma_free(struct svc_xprt *xprt); 67 static int svc_rdma_has_wspace(struct svc_xprt *xprt); 68 static void rq_cq_reap(struct svcxprt_rdma *xprt); 69 static void sq_cq_reap(struct svcxprt_rdma *xprt); 70 71 static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL); 72 static DEFINE_SPINLOCK(dto_lock); 73 static LIST_HEAD(dto_xprt_q); 74 75 static struct svc_xprt_ops svc_rdma_ops = { 76 .xpo_create = svc_rdma_create, 77 .xpo_recvfrom = svc_rdma_recvfrom, 78 .xpo_sendto = svc_rdma_sendto, 79 .xpo_release_rqst = svc_rdma_release_rqst, 80 .xpo_detach = svc_rdma_detach, 81 .xpo_free = svc_rdma_free, 82 .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr, 83 .xpo_has_wspace = svc_rdma_has_wspace, 84 .xpo_accept = svc_rdma_accept, 85 }; 86 87 struct svc_xprt_class svc_rdma_class = { 88 .xcl_name = "rdma", 89 .xcl_owner = THIS_MODULE, 90 .xcl_ops = &svc_rdma_ops, 91 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP, 92 }; 93 94 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt) 95 { 96 struct svc_rdma_op_ctxt *ctxt; 97 98 while (1) { 99 ctxt = kmem_cache_alloc(svc_rdma_ctxt_cachep, GFP_KERNEL); 100 if (ctxt) 101 break; 102 schedule_timeout_uninterruptible(msecs_to_jiffies(500)); 103 } 104 ctxt->xprt = xprt; 105 INIT_LIST_HEAD(&ctxt->dto_q); 106 ctxt->count = 0; 107 ctxt->frmr = NULL; 108 atomic_inc(&xprt->sc_ctxt_used); 109 return ctxt; 110 } 111 112 void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt) 113 { 114 struct svcxprt_rdma *xprt = ctxt->xprt; 115 int i; 116 for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) { 117 /* 118 * Unmap the DMA addr in the SGE if the lkey matches 119 * the sc_dma_lkey, otherwise, ignore it since it is 120 * an FRMR lkey and will be unmapped later when the 121 * last WR that uses it completes. 122 */ 123 if (ctxt->sge[i].lkey == xprt->sc_dma_lkey) { 124 atomic_dec(&xprt->sc_dma_used); 125 ib_dma_unmap_page(xprt->sc_cm_id->device, 126 ctxt->sge[i].addr, 127 ctxt->sge[i].length, 128 ctxt->direction); 129 } 130 } 131 } 132 133 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages) 134 { 135 struct svcxprt_rdma *xprt; 136 int i; 137 138 BUG_ON(!ctxt); 139 xprt = ctxt->xprt; 140 if (free_pages) 141 for (i = 0; i < ctxt->count; i++) 142 put_page(ctxt->pages[i]); 143 144 kmem_cache_free(svc_rdma_ctxt_cachep, ctxt); 145 atomic_dec(&xprt->sc_ctxt_used); 146 } 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 * Process 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(&init_net, &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 1s\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_page(xprt->sc_cm_id->device, 504 page, 0, 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 ctxt->count = sge_no + 1; 513 buflen += PAGE_SIZE; 514 } 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_unmap_dma(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 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 581 svc_xprt_enqueue(&listen_xprt->sc_xprt); 582 } 583 584 /* 585 * Handles events generated on the listening endpoint. These events will be 586 * either be incoming connect requests or adapter removal events. 587 */ 588 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 589 struct rdma_cm_event *event) 590 { 591 struct svcxprt_rdma *xprt = cma_id->context; 592 int ret = 0; 593 594 switch (event->event) { 595 case RDMA_CM_EVENT_CONNECT_REQUEST: 596 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 597 "event=%d\n", cma_id, cma_id->context, event->event); 598 handle_connect_req(cma_id, 599 event->param.conn.initiator_depth); 600 break; 601 602 case RDMA_CM_EVENT_ESTABLISHED: 603 /* Accept complete */ 604 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " 605 "cm_id=%p\n", xprt, cma_id); 606 break; 607 608 case RDMA_CM_EVENT_DEVICE_REMOVAL: 609 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", 610 xprt, cma_id); 611 if (xprt) 612 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 613 break; 614 615 default: 616 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 617 "event=%d\n", cma_id, event->event); 618 break; 619 } 620 621 return ret; 622 } 623 624 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 625 struct rdma_cm_event *event) 626 { 627 struct svc_xprt *xprt = cma_id->context; 628 struct svcxprt_rdma *rdma = 629 container_of(xprt, struct svcxprt_rdma, sc_xprt); 630 switch (event->event) { 631 case RDMA_CM_EVENT_ESTABLISHED: 632 /* Accept complete */ 633 svc_xprt_get(xprt); 634 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 635 "cm_id=%p\n", xprt, cma_id); 636 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 637 svc_xprt_enqueue(xprt); 638 break; 639 case RDMA_CM_EVENT_DISCONNECTED: 640 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 641 xprt, cma_id); 642 if (xprt) { 643 set_bit(XPT_CLOSE, &xprt->xpt_flags); 644 svc_xprt_enqueue(xprt); 645 svc_xprt_put(xprt); 646 } 647 break; 648 case RDMA_CM_EVENT_DEVICE_REMOVAL: 649 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 650 "event=%d\n", cma_id, xprt, event->event); 651 if (xprt) { 652 set_bit(XPT_CLOSE, &xprt->xpt_flags); 653 svc_xprt_enqueue(xprt); 654 } 655 break; 656 default: 657 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 658 "event=%d\n", cma_id, event->event); 659 break; 660 } 661 return 0; 662 } 663 664 /* 665 * Create a listening RDMA service endpoint. 666 */ 667 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 668 struct net *net, 669 struct sockaddr *sa, int salen, 670 int flags) 671 { 672 struct rdma_cm_id *listen_id; 673 struct svcxprt_rdma *cma_xprt; 674 struct svc_xprt *xprt; 675 int ret; 676 677 dprintk("svcrdma: Creating RDMA socket\n"); 678 if (sa->sa_family != AF_INET) { 679 dprintk("svcrdma: Address family %d is not supported.\n", sa->sa_family); 680 return ERR_PTR(-EAFNOSUPPORT); 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 IB_QPT_RC); 689 if (IS_ERR(listen_id)) { 690 ret = PTR_ERR(listen_id); 691 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 692 goto err0; 693 } 694 695 ret = rdma_bind_addr(listen_id, sa); 696 if (ret) { 697 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 698 goto err1; 699 } 700 cma_xprt->sc_cm_id = listen_id; 701 702 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 703 if (ret) { 704 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 705 goto err1; 706 } 707 708 /* 709 * We need to use the address from the cm_id in case the 710 * caller specified 0 for the port number. 711 */ 712 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 713 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 714 715 return &cma_xprt->sc_xprt; 716 717 err1: 718 rdma_destroy_id(listen_id); 719 err0: 720 kfree(cma_xprt); 721 return ERR_PTR(ret); 722 } 723 724 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) 725 { 726 struct ib_mr *mr; 727 struct ib_fast_reg_page_list *pl; 728 struct svc_rdma_fastreg_mr *frmr; 729 730 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); 731 if (!frmr) 732 goto err; 733 734 mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); 735 if (IS_ERR(mr)) 736 goto err_free_frmr; 737 738 pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, 739 RPCSVC_MAXPAGES); 740 if (IS_ERR(pl)) 741 goto err_free_mr; 742 743 frmr->mr = mr; 744 frmr->page_list = pl; 745 INIT_LIST_HEAD(&frmr->frmr_list); 746 return frmr; 747 748 err_free_mr: 749 ib_dereg_mr(mr); 750 err_free_frmr: 751 kfree(frmr); 752 err: 753 return ERR_PTR(-ENOMEM); 754 } 755 756 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) 757 { 758 struct svc_rdma_fastreg_mr *frmr; 759 760 while (!list_empty(&xprt->sc_frmr_q)) { 761 frmr = list_entry(xprt->sc_frmr_q.next, 762 struct svc_rdma_fastreg_mr, frmr_list); 763 list_del_init(&frmr->frmr_list); 764 ib_dereg_mr(frmr->mr); 765 ib_free_fast_reg_page_list(frmr->page_list); 766 kfree(frmr); 767 } 768 } 769 770 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) 771 { 772 struct svc_rdma_fastreg_mr *frmr = NULL; 773 774 spin_lock_bh(&rdma->sc_frmr_q_lock); 775 if (!list_empty(&rdma->sc_frmr_q)) { 776 frmr = list_entry(rdma->sc_frmr_q.next, 777 struct svc_rdma_fastreg_mr, frmr_list); 778 list_del_init(&frmr->frmr_list); 779 frmr->map_len = 0; 780 frmr->page_list_len = 0; 781 } 782 spin_unlock_bh(&rdma->sc_frmr_q_lock); 783 if (frmr) 784 return frmr; 785 786 return rdma_alloc_frmr(rdma); 787 } 788 789 static void frmr_unmap_dma(struct svcxprt_rdma *xprt, 790 struct svc_rdma_fastreg_mr *frmr) 791 { 792 int page_no; 793 for (page_no = 0; page_no < frmr->page_list_len; page_no++) { 794 dma_addr_t addr = frmr->page_list->page_list[page_no]; 795 if (ib_dma_mapping_error(frmr->mr->device, addr)) 796 continue; 797 atomic_dec(&xprt->sc_dma_used); 798 ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE, 799 frmr->direction); 800 } 801 } 802 803 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, 804 struct svc_rdma_fastreg_mr *frmr) 805 { 806 if (frmr) { 807 frmr_unmap_dma(rdma, frmr); 808 spin_lock_bh(&rdma->sc_frmr_q_lock); 809 BUG_ON(!list_empty(&frmr->frmr_list)); 810 list_add(&frmr->frmr_list, &rdma->sc_frmr_q); 811 spin_unlock_bh(&rdma->sc_frmr_q_lock); 812 } 813 } 814 815 /* 816 * This is the xpo_recvfrom function for listening endpoints. Its 817 * purpose is to accept incoming connections. The CMA callback handler 818 * has already created a new transport and attached it to the new CMA 819 * ID. 820 * 821 * There is a queue of pending connections hung on the listening 822 * transport. This queue contains the new svc_xprt structure. This 823 * function takes svc_xprt structures off the accept_q and completes 824 * the connection. 825 */ 826 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 827 { 828 struct svcxprt_rdma *listen_rdma; 829 struct svcxprt_rdma *newxprt = NULL; 830 struct rdma_conn_param conn_param; 831 struct ib_qp_init_attr qp_attr; 832 struct ib_device_attr devattr; 833 int uninitialized_var(dma_mr_acc); 834 int need_dma_mr; 835 int ret; 836 int i; 837 838 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 839 clear_bit(XPT_CONN, &xprt->xpt_flags); 840 /* Get the next entry off the accept list */ 841 spin_lock_bh(&listen_rdma->sc_lock); 842 if (!list_empty(&listen_rdma->sc_accept_q)) { 843 newxprt = list_entry(listen_rdma->sc_accept_q.next, 844 struct svcxprt_rdma, sc_accept_q); 845 list_del_init(&newxprt->sc_accept_q); 846 } 847 if (!list_empty(&listen_rdma->sc_accept_q)) 848 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 849 spin_unlock_bh(&listen_rdma->sc_lock); 850 if (!newxprt) 851 return NULL; 852 853 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 854 newxprt, newxprt->sc_cm_id); 855 856 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); 857 if (ret) { 858 dprintk("svcrdma: could not query device attributes on " 859 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); 860 goto errout; 861 } 862 863 /* Qualify the transport resource defaults with the 864 * capabilities of this particular device */ 865 newxprt->sc_max_sge = min((size_t)devattr.max_sge, 866 (size_t)RPCSVC_MAXPAGES); 867 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, 868 (size_t)svcrdma_max_requests); 869 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; 870 871 /* 872 * Limit ORD based on client limit, local device limit, and 873 * configured svcrdma limit. 874 */ 875 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); 876 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); 877 878 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); 879 if (IS_ERR(newxprt->sc_pd)) { 880 dprintk("svcrdma: error creating PD for connect request\n"); 881 goto errout; 882 } 883 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, 884 sq_comp_handler, 885 cq_event_handler, 886 newxprt, 887 newxprt->sc_sq_depth, 888 0); 889 if (IS_ERR(newxprt->sc_sq_cq)) { 890 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 891 goto errout; 892 } 893 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, 894 rq_comp_handler, 895 cq_event_handler, 896 newxprt, 897 newxprt->sc_max_requests, 898 0); 899 if (IS_ERR(newxprt->sc_rq_cq)) { 900 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 901 goto errout; 902 } 903 904 memset(&qp_attr, 0, sizeof qp_attr); 905 qp_attr.event_handler = qp_event_handler; 906 qp_attr.qp_context = &newxprt->sc_xprt; 907 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; 908 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; 909 qp_attr.cap.max_send_sge = newxprt->sc_max_sge; 910 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; 911 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 912 qp_attr.qp_type = IB_QPT_RC; 913 qp_attr.send_cq = newxprt->sc_sq_cq; 914 qp_attr.recv_cq = newxprt->sc_rq_cq; 915 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" 916 " cm_id->device=%p, sc_pd->device=%p\n" 917 " cap.max_send_wr = %d\n" 918 " cap.max_recv_wr = %d\n" 919 " cap.max_send_sge = %d\n" 920 " cap.max_recv_sge = %d\n", 921 newxprt->sc_cm_id, newxprt->sc_pd, 922 newxprt->sc_cm_id->device, newxprt->sc_pd->device, 923 qp_attr.cap.max_send_wr, 924 qp_attr.cap.max_recv_wr, 925 qp_attr.cap.max_send_sge, 926 qp_attr.cap.max_recv_sge); 927 928 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 929 if (ret) { 930 /* 931 * XXX: This is a hack. We need a xx_request_qp interface 932 * that will adjust the qp_attr's with a best-effort 933 * number 934 */ 935 qp_attr.cap.max_send_sge -= 2; 936 qp_attr.cap.max_recv_sge -= 2; 937 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, 938 &qp_attr); 939 if (ret) { 940 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 941 goto errout; 942 } 943 newxprt->sc_max_sge = qp_attr.cap.max_send_sge; 944 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge; 945 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr; 946 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr; 947 } 948 newxprt->sc_qp = newxprt->sc_cm_id->qp; 949 950 /* 951 * Use the most secure set of MR resources based on the 952 * transport type and available memory management features in 953 * the device. Here's the table implemented below: 954 * 955 * Fast Global DMA Remote WR 956 * Reg LKEY MR Access 957 * Sup'd Sup'd Needed Needed 958 * 959 * IWARP N N Y Y 960 * N Y Y Y 961 * Y N Y N 962 * Y Y N - 963 * 964 * IB N N Y N 965 * N Y N - 966 * Y N Y N 967 * Y Y N - 968 * 969 * NB: iWARP requires remote write access for the data sink 970 * of an RDMA_READ. IB does not. 971 */ 972 if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 973 newxprt->sc_frmr_pg_list_len = 974 devattr.max_fast_reg_page_list_len; 975 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; 976 } 977 978 /* 979 * Determine if a DMA MR is required and if so, what privs are required 980 */ 981 switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { 982 case RDMA_TRANSPORT_IWARP: 983 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; 984 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 985 need_dma_mr = 1; 986 dma_mr_acc = 987 (IB_ACCESS_LOCAL_WRITE | 988 IB_ACCESS_REMOTE_WRITE); 989 } else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 990 need_dma_mr = 1; 991 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 992 } else 993 need_dma_mr = 0; 994 break; 995 case RDMA_TRANSPORT_IB: 996 if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 997 need_dma_mr = 1; 998 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 999 } else 1000 need_dma_mr = 0; 1001 break; 1002 default: 1003 goto errout; 1004 } 1005 1006 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ 1007 if (need_dma_mr) { 1008 /* Register all of physical memory */ 1009 newxprt->sc_phys_mr = 1010 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); 1011 if (IS_ERR(newxprt->sc_phys_mr)) { 1012 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", 1013 ret); 1014 goto errout; 1015 } 1016 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; 1017 } else 1018 newxprt->sc_dma_lkey = 1019 newxprt->sc_cm_id->device->local_dma_lkey; 1020 1021 /* Post receive buffers */ 1022 for (i = 0; i < newxprt->sc_max_requests; i++) { 1023 ret = svc_rdma_post_recv(newxprt); 1024 if (ret) { 1025 dprintk("svcrdma: failure posting receive buffers\n"); 1026 goto errout; 1027 } 1028 } 1029 1030 /* Swap out the handler */ 1031 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 1032 1033 /* 1034 * Arm the CQs for the SQ and RQ before accepting so we can't 1035 * miss the first message 1036 */ 1037 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); 1038 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); 1039 1040 /* Accept Connection */ 1041 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 1042 memset(&conn_param, 0, sizeof conn_param); 1043 conn_param.responder_resources = 0; 1044 conn_param.initiator_depth = newxprt->sc_ord; 1045 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 1046 if (ret) { 1047 dprintk("svcrdma: failed to accept new connection, ret=%d\n", 1048 ret); 1049 goto errout; 1050 } 1051 1052 dprintk("svcrdma: new connection %p accepted with the following " 1053 "attributes:\n" 1054 " local_ip : %pI4\n" 1055 " local_port : %d\n" 1056 " remote_ip : %pI4\n" 1057 " remote_port : %d\n" 1058 " max_sge : %d\n" 1059 " sq_depth : %d\n" 1060 " max_requests : %d\n" 1061 " ord : %d\n", 1062 newxprt, 1063 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1064 route.addr.src_addr)->sin_addr.s_addr, 1065 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1066 route.addr.src_addr)->sin_port), 1067 &((struct sockaddr_in *)&newxprt->sc_cm_id-> 1068 route.addr.dst_addr)->sin_addr.s_addr, 1069 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1070 route.addr.dst_addr)->sin_port), 1071 newxprt->sc_max_sge, 1072 newxprt->sc_sq_depth, 1073 newxprt->sc_max_requests, 1074 newxprt->sc_ord); 1075 1076 return &newxprt->sc_xprt; 1077 1078 errout: 1079 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 1080 /* Take a reference in case the DTO handler runs */ 1081 svc_xprt_get(&newxprt->sc_xprt); 1082 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 1083 ib_destroy_qp(newxprt->sc_qp); 1084 rdma_destroy_id(newxprt->sc_cm_id); 1085 /* This call to put will destroy the transport */ 1086 svc_xprt_put(&newxprt->sc_xprt); 1087 return NULL; 1088 } 1089 1090 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 1091 { 1092 } 1093 1094 /* 1095 * When connected, an svc_xprt has at least two references: 1096 * 1097 * - A reference held by the cm_id between the ESTABLISHED and 1098 * DISCONNECTED events. If the remote peer disconnected first, this 1099 * reference could be gone. 1100 * 1101 * - A reference held by the svc_recv code that called this function 1102 * as part of close processing. 1103 * 1104 * At a minimum one references should still be held. 1105 */ 1106 static void svc_rdma_detach(struct svc_xprt *xprt) 1107 { 1108 struct svcxprt_rdma *rdma = 1109 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1110 dprintk("svc: svc_rdma_detach(%p)\n", xprt); 1111 1112 /* Disconnect and flush posted WQE */ 1113 rdma_disconnect(rdma->sc_cm_id); 1114 } 1115 1116 static void __svc_rdma_free(struct work_struct *work) 1117 { 1118 struct svcxprt_rdma *rdma = 1119 container_of(work, struct svcxprt_rdma, sc_work); 1120 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); 1121 1122 /* We should only be called from kref_put */ 1123 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0); 1124 1125 /* 1126 * Destroy queued, but not processed read completions. Note 1127 * that this cleanup has to be done before destroying the 1128 * cm_id because the device ptr is needed to unmap the dma in 1129 * svc_rdma_put_context. 1130 */ 1131 while (!list_empty(&rdma->sc_read_complete_q)) { 1132 struct svc_rdma_op_ctxt *ctxt; 1133 ctxt = list_entry(rdma->sc_read_complete_q.next, 1134 struct svc_rdma_op_ctxt, 1135 dto_q); 1136 list_del_init(&ctxt->dto_q); 1137 svc_rdma_put_context(ctxt, 1); 1138 } 1139 1140 /* Destroy queued, but not processed recv completions */ 1141 while (!list_empty(&rdma->sc_rq_dto_q)) { 1142 struct svc_rdma_op_ctxt *ctxt; 1143 ctxt = list_entry(rdma->sc_rq_dto_q.next, 1144 struct svc_rdma_op_ctxt, 1145 dto_q); 1146 list_del_init(&ctxt->dto_q); 1147 svc_rdma_put_context(ctxt, 1); 1148 } 1149 1150 /* Warn if we leaked a resource or under-referenced */ 1151 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0); 1152 WARN_ON(atomic_read(&rdma->sc_dma_used) != 0); 1153 1154 /* De-allocate fastreg mr */ 1155 rdma_dealloc_frmr_q(rdma); 1156 1157 /* Destroy the QP if present (not a listener) */ 1158 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1159 ib_destroy_qp(rdma->sc_qp); 1160 1161 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 1162 ib_destroy_cq(rdma->sc_sq_cq); 1163 1164 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 1165 ib_destroy_cq(rdma->sc_rq_cq); 1166 1167 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) 1168 ib_dereg_mr(rdma->sc_phys_mr); 1169 1170 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 1171 ib_dealloc_pd(rdma->sc_pd); 1172 1173 /* Destroy the CM ID */ 1174 rdma_destroy_id(rdma->sc_cm_id); 1175 1176 kfree(rdma); 1177 } 1178 1179 static void svc_rdma_free(struct svc_xprt *xprt) 1180 { 1181 struct svcxprt_rdma *rdma = 1182 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1183 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 1184 queue_work(svc_rdma_wq, &rdma->sc_work); 1185 } 1186 1187 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 1188 { 1189 struct svcxprt_rdma *rdma = 1190 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1191 1192 /* 1193 * If there are fewer SQ WR available than required to send a 1194 * simple response, return false. 1195 */ 1196 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3)) 1197 return 0; 1198 1199 /* 1200 * ...or there are already waiters on the SQ, 1201 * return false. 1202 */ 1203 if (waitqueue_active(&rdma->sc_send_wait)) 1204 return 0; 1205 1206 /* Otherwise return true. */ 1207 return 1; 1208 } 1209 1210 /* 1211 * Attempt to register the kvec representing the RPC memory with the 1212 * device. 1213 * 1214 * Returns: 1215 * NULL : The device does not support fastreg or there were no more 1216 * fastreg mr. 1217 * frmr : The kvec register request was successfully posted. 1218 * <0 : An error was encountered attempting to register the kvec. 1219 */ 1220 int svc_rdma_fastreg(struct svcxprt_rdma *xprt, 1221 struct svc_rdma_fastreg_mr *frmr) 1222 { 1223 struct ib_send_wr fastreg_wr; 1224 u8 key; 1225 1226 /* Bump the key */ 1227 key = (u8)(frmr->mr->lkey & 0x000000FF); 1228 ib_update_fast_reg_key(frmr->mr, ++key); 1229 1230 /* Prepare FASTREG WR */ 1231 memset(&fastreg_wr, 0, sizeof fastreg_wr); 1232 fastreg_wr.opcode = IB_WR_FAST_REG_MR; 1233 fastreg_wr.send_flags = IB_SEND_SIGNALED; 1234 fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; 1235 fastreg_wr.wr.fast_reg.page_list = frmr->page_list; 1236 fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; 1237 fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; 1238 fastreg_wr.wr.fast_reg.length = frmr->map_len; 1239 fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; 1240 fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; 1241 return svc_rdma_send(xprt, &fastreg_wr); 1242 } 1243 1244 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) 1245 { 1246 struct ib_send_wr *bad_wr, *n_wr; 1247 int wr_count; 1248 int i; 1249 int ret; 1250 1251 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1252 return -ENOTCONN; 1253 1254 BUG_ON(wr->send_flags != IB_SEND_SIGNALED); 1255 wr_count = 1; 1256 for (n_wr = wr->next; n_wr; n_wr = n_wr->next) 1257 wr_count++; 1258 1259 /* If the SQ is full, wait until an SQ entry is available */ 1260 while (1) { 1261 spin_lock_bh(&xprt->sc_lock); 1262 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { 1263 spin_unlock_bh(&xprt->sc_lock); 1264 atomic_inc(&rdma_stat_sq_starve); 1265 1266 /* See if we can opportunistically reap SQ WR to make room */ 1267 sq_cq_reap(xprt); 1268 1269 /* Wait until SQ WR available if SQ still full */ 1270 wait_event(xprt->sc_send_wait, 1271 atomic_read(&xprt->sc_sq_count) < 1272 xprt->sc_sq_depth); 1273 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1274 return -ENOTCONN; 1275 continue; 1276 } 1277 /* Take a transport ref for each WR posted */ 1278 for (i = 0; i < wr_count; i++) 1279 svc_xprt_get(&xprt->sc_xprt); 1280 1281 /* Bump used SQ WR count and post */ 1282 atomic_add(wr_count, &xprt->sc_sq_count); 1283 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); 1284 if (ret) { 1285 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 1286 atomic_sub(wr_count, &xprt->sc_sq_count); 1287 for (i = 0; i < wr_count; i ++) 1288 svc_xprt_put(&xprt->sc_xprt); 1289 dprintk("svcrdma: failed to post SQ WR rc=%d, " 1290 "sc_sq_count=%d, sc_sq_depth=%d\n", 1291 ret, atomic_read(&xprt->sc_sq_count), 1292 xprt->sc_sq_depth); 1293 } 1294 spin_unlock_bh(&xprt->sc_lock); 1295 if (ret) 1296 wake_up(&xprt->sc_send_wait); 1297 break; 1298 } 1299 return ret; 1300 } 1301 1302 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, 1303 enum rpcrdma_errcode err) 1304 { 1305 struct ib_send_wr err_wr; 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 ctxt = svc_rdma_get_context(xprt); 1319 ctxt->direction = DMA_FROM_DEVICE; 1320 ctxt->count = 1; 1321 ctxt->pages[0] = p; 1322 1323 /* Prepare SGE for local address */ 1324 ctxt->sge[0].addr = ib_dma_map_page(xprt->sc_cm_id->device, 1325 p, 0, length, DMA_FROM_DEVICE); 1326 if (ib_dma_mapping_error(xprt->sc_cm_id->device, ctxt->sge[0].addr)) { 1327 put_page(p); 1328 svc_rdma_put_context(ctxt, 1); 1329 return; 1330 } 1331 atomic_inc(&xprt->sc_dma_used); 1332 ctxt->sge[0].lkey = xprt->sc_dma_lkey; 1333 ctxt->sge[0].length = length; 1334 1335 /* Prepare SEND WR */ 1336 memset(&err_wr, 0, sizeof err_wr); 1337 ctxt->wr_op = IB_WR_SEND; 1338 err_wr.wr_id = (unsigned long)ctxt; 1339 err_wr.sg_list = ctxt->sge; 1340 err_wr.num_sge = 1; 1341 err_wr.opcode = IB_WR_SEND; 1342 err_wr.send_flags = IB_SEND_SIGNALED; 1343 1344 /* Post It */ 1345 ret = svc_rdma_send(xprt, &err_wr); 1346 if (ret) { 1347 dprintk("svcrdma: Error %d posting send for protocol error\n", 1348 ret); 1349 svc_rdma_unmap_dma(ctxt); 1350 svc_rdma_put_context(ctxt, 1); 1351 } 1352 } 1353