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 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_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 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_xprt_put(&xprt->sc_xprt); 524 svc_rdma_put_context(ctxt, 1); 525 } 526 return ret; 527 528 err_put_ctxt: 529 svc_rdma_put_context(ctxt, 1); 530 return -ENOMEM; 531 } 532 533 /* 534 * This function handles the CONNECT_REQUEST event on a listening 535 * endpoint. It is passed the cma_id for the _new_ connection. The context in 536 * this cma_id is inherited from the listening cma_id and is the svc_xprt 537 * structure for the listening endpoint. 538 * 539 * This function creates a new xprt for the new connection and enqueues it on 540 * the accept queue for the listent xprt. When the listen thread is kicked, it 541 * will call the recvfrom method on the listen xprt which will accept the new 542 * connection. 543 */ 544 static void handle_connect_req(struct rdma_cm_id *new_cma_id, size_t client_ird) 545 { 546 struct svcxprt_rdma *listen_xprt = new_cma_id->context; 547 struct svcxprt_rdma *newxprt; 548 struct sockaddr *sa; 549 550 /* Create a new transport */ 551 newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0); 552 if (!newxprt) { 553 dprintk("svcrdma: failed to create new transport\n"); 554 return; 555 } 556 newxprt->sc_cm_id = new_cma_id; 557 new_cma_id->context = newxprt; 558 dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n", 559 newxprt, newxprt->sc_cm_id, listen_xprt); 560 561 /* Save client advertised inbound read limit for use later in accept. */ 562 newxprt->sc_ord = client_ird; 563 564 /* Set the local and remote addresses in the transport */ 565 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr; 566 svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 567 sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr; 568 svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa)); 569 570 /* 571 * Enqueue the new transport on the accept queue of the listening 572 * transport 573 */ 574 spin_lock_bh(&listen_xprt->sc_lock); 575 list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q); 576 spin_unlock_bh(&listen_xprt->sc_lock); 577 578 /* 579 * Can't use svc_xprt_received here because we are not on a 580 * rqstp thread 581 */ 582 set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags); 583 svc_xprt_enqueue(&listen_xprt->sc_xprt); 584 } 585 586 /* 587 * Handles events generated on the listening endpoint. These events will be 588 * either be incoming connect requests or adapter removal events. 589 */ 590 static int rdma_listen_handler(struct rdma_cm_id *cma_id, 591 struct rdma_cm_event *event) 592 { 593 struct svcxprt_rdma *xprt = cma_id->context; 594 int ret = 0; 595 596 switch (event->event) { 597 case RDMA_CM_EVENT_CONNECT_REQUEST: 598 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, " 599 "event=%d\n", cma_id, cma_id->context, event->event); 600 handle_connect_req(cma_id, 601 event->param.conn.initiator_depth); 602 break; 603 604 case RDMA_CM_EVENT_ESTABLISHED: 605 /* Accept complete */ 606 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, " 607 "cm_id=%p\n", xprt, cma_id); 608 break; 609 610 case RDMA_CM_EVENT_DEVICE_REMOVAL: 611 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n", 612 xprt, cma_id); 613 if (xprt) 614 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 615 break; 616 617 default: 618 dprintk("svcrdma: Unexpected event on listening endpoint %p, " 619 "event=%d\n", cma_id, event->event); 620 break; 621 } 622 623 return ret; 624 } 625 626 static int rdma_cma_handler(struct rdma_cm_id *cma_id, 627 struct rdma_cm_event *event) 628 { 629 struct svc_xprt *xprt = cma_id->context; 630 struct svcxprt_rdma *rdma = 631 container_of(xprt, struct svcxprt_rdma, sc_xprt); 632 switch (event->event) { 633 case RDMA_CM_EVENT_ESTABLISHED: 634 /* Accept complete */ 635 svc_xprt_get(xprt); 636 dprintk("svcrdma: Connection completed on DTO xprt=%p, " 637 "cm_id=%p\n", xprt, cma_id); 638 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags); 639 svc_xprt_enqueue(xprt); 640 break; 641 case RDMA_CM_EVENT_DISCONNECTED: 642 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n", 643 xprt, cma_id); 644 if (xprt) { 645 set_bit(XPT_CLOSE, &xprt->xpt_flags); 646 svc_xprt_enqueue(xprt); 647 svc_xprt_put(xprt); 648 } 649 break; 650 case RDMA_CM_EVENT_DEVICE_REMOVAL: 651 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, " 652 "event=%d\n", cma_id, xprt, event->event); 653 if (xprt) { 654 set_bit(XPT_CLOSE, &xprt->xpt_flags); 655 svc_xprt_enqueue(xprt); 656 } 657 break; 658 default: 659 dprintk("svcrdma: Unexpected event on DTO endpoint %p, " 660 "event=%d\n", cma_id, event->event); 661 break; 662 } 663 return 0; 664 } 665 666 /* 667 * Create a listening RDMA service endpoint. 668 */ 669 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv, 670 struct sockaddr *sa, int salen, 671 int flags) 672 { 673 struct rdma_cm_id *listen_id; 674 struct svcxprt_rdma *cma_xprt; 675 struct svc_xprt *xprt; 676 int ret; 677 678 dprintk("svcrdma: Creating RDMA socket\n"); 679 680 cma_xprt = rdma_create_xprt(serv, 1); 681 if (!cma_xprt) 682 return ERR_PTR(-ENOMEM); 683 xprt = &cma_xprt->sc_xprt; 684 685 listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP); 686 if (IS_ERR(listen_id)) { 687 ret = PTR_ERR(listen_id); 688 dprintk("svcrdma: rdma_create_id failed = %d\n", ret); 689 goto err0; 690 } 691 692 ret = rdma_bind_addr(listen_id, sa); 693 if (ret) { 694 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret); 695 goto err1; 696 } 697 cma_xprt->sc_cm_id = listen_id; 698 699 ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG); 700 if (ret) { 701 dprintk("svcrdma: rdma_listen failed = %d\n", ret); 702 goto err1; 703 } 704 705 /* 706 * We need to use the address from the cm_id in case the 707 * caller specified 0 for the port number. 708 */ 709 sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr; 710 svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen); 711 712 return &cma_xprt->sc_xprt; 713 714 err1: 715 rdma_destroy_id(listen_id); 716 err0: 717 kfree(cma_xprt); 718 return ERR_PTR(ret); 719 } 720 721 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt) 722 { 723 struct ib_mr *mr; 724 struct ib_fast_reg_page_list *pl; 725 struct svc_rdma_fastreg_mr *frmr; 726 727 frmr = kmalloc(sizeof(*frmr), GFP_KERNEL); 728 if (!frmr) 729 goto err; 730 731 mr = ib_alloc_fast_reg_mr(xprt->sc_pd, RPCSVC_MAXPAGES); 732 if (!mr) 733 goto err_free_frmr; 734 735 pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device, 736 RPCSVC_MAXPAGES); 737 if (!pl) 738 goto err_free_mr; 739 740 frmr->mr = mr; 741 frmr->page_list = pl; 742 INIT_LIST_HEAD(&frmr->frmr_list); 743 return frmr; 744 745 err_free_mr: 746 ib_dereg_mr(mr); 747 err_free_frmr: 748 kfree(frmr); 749 err: 750 return ERR_PTR(-ENOMEM); 751 } 752 753 static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt) 754 { 755 struct svc_rdma_fastreg_mr *frmr; 756 757 while (!list_empty(&xprt->sc_frmr_q)) { 758 frmr = list_entry(xprt->sc_frmr_q.next, 759 struct svc_rdma_fastreg_mr, frmr_list); 760 list_del_init(&frmr->frmr_list); 761 ib_dereg_mr(frmr->mr); 762 ib_free_fast_reg_page_list(frmr->page_list); 763 kfree(frmr); 764 } 765 } 766 767 struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma) 768 { 769 struct svc_rdma_fastreg_mr *frmr = NULL; 770 771 spin_lock_bh(&rdma->sc_frmr_q_lock); 772 if (!list_empty(&rdma->sc_frmr_q)) { 773 frmr = list_entry(rdma->sc_frmr_q.next, 774 struct svc_rdma_fastreg_mr, frmr_list); 775 list_del_init(&frmr->frmr_list); 776 frmr->map_len = 0; 777 frmr->page_list_len = 0; 778 } 779 spin_unlock_bh(&rdma->sc_frmr_q_lock); 780 if (frmr) 781 return frmr; 782 783 return rdma_alloc_frmr(rdma); 784 } 785 786 static void frmr_unmap_dma(struct svcxprt_rdma *xprt, 787 struct svc_rdma_fastreg_mr *frmr) 788 { 789 int page_no; 790 for (page_no = 0; page_no < frmr->page_list_len; page_no++) { 791 dma_addr_t addr = frmr->page_list->page_list[page_no]; 792 if (ib_dma_mapping_error(frmr->mr->device, addr)) 793 continue; 794 atomic_dec(&xprt->sc_dma_used); 795 ib_dma_unmap_single(frmr->mr->device, addr, PAGE_SIZE, 796 frmr->direction); 797 } 798 } 799 800 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma, 801 struct svc_rdma_fastreg_mr *frmr) 802 { 803 if (frmr) { 804 frmr_unmap_dma(rdma, frmr); 805 spin_lock_bh(&rdma->sc_frmr_q_lock); 806 BUG_ON(!list_empty(&frmr->frmr_list)); 807 list_add(&frmr->frmr_list, &rdma->sc_frmr_q); 808 spin_unlock_bh(&rdma->sc_frmr_q_lock); 809 } 810 } 811 812 /* 813 * This is the xpo_recvfrom function for listening endpoints. Its 814 * purpose is to accept incoming connections. The CMA callback handler 815 * has already created a new transport and attached it to the new CMA 816 * ID. 817 * 818 * There is a queue of pending connections hung on the listening 819 * transport. This queue contains the new svc_xprt structure. This 820 * function takes svc_xprt structures off the accept_q and completes 821 * the connection. 822 */ 823 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) 824 { 825 struct svcxprt_rdma *listen_rdma; 826 struct svcxprt_rdma *newxprt = NULL; 827 struct rdma_conn_param conn_param; 828 struct ib_qp_init_attr qp_attr; 829 struct ib_device_attr devattr; 830 int dma_mr_acc; 831 int need_dma_mr; 832 int ret; 833 int i; 834 835 listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt); 836 clear_bit(XPT_CONN, &xprt->xpt_flags); 837 /* Get the next entry off the accept list */ 838 spin_lock_bh(&listen_rdma->sc_lock); 839 if (!list_empty(&listen_rdma->sc_accept_q)) { 840 newxprt = list_entry(listen_rdma->sc_accept_q.next, 841 struct svcxprt_rdma, sc_accept_q); 842 list_del_init(&newxprt->sc_accept_q); 843 } 844 if (!list_empty(&listen_rdma->sc_accept_q)) 845 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags); 846 spin_unlock_bh(&listen_rdma->sc_lock); 847 if (!newxprt) 848 return NULL; 849 850 dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n", 851 newxprt, newxprt->sc_cm_id); 852 853 ret = ib_query_device(newxprt->sc_cm_id->device, &devattr); 854 if (ret) { 855 dprintk("svcrdma: could not query device attributes on " 856 "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret); 857 goto errout; 858 } 859 860 /* Qualify the transport resource defaults with the 861 * capabilities of this particular device */ 862 newxprt->sc_max_sge = min((size_t)devattr.max_sge, 863 (size_t)RPCSVC_MAXPAGES); 864 newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr, 865 (size_t)svcrdma_max_requests); 866 newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests; 867 868 /* 869 * Limit ORD based on client limit, local device limit, and 870 * configured svcrdma limit. 871 */ 872 newxprt->sc_ord = min_t(size_t, devattr.max_qp_rd_atom, newxprt->sc_ord); 873 newxprt->sc_ord = min_t(size_t, svcrdma_ord, newxprt->sc_ord); 874 875 newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device); 876 if (IS_ERR(newxprt->sc_pd)) { 877 dprintk("svcrdma: error creating PD for connect request\n"); 878 goto errout; 879 } 880 newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device, 881 sq_comp_handler, 882 cq_event_handler, 883 newxprt, 884 newxprt->sc_sq_depth, 885 0); 886 if (IS_ERR(newxprt->sc_sq_cq)) { 887 dprintk("svcrdma: error creating SQ CQ for connect request\n"); 888 goto errout; 889 } 890 newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device, 891 rq_comp_handler, 892 cq_event_handler, 893 newxprt, 894 newxprt->sc_max_requests, 895 0); 896 if (IS_ERR(newxprt->sc_rq_cq)) { 897 dprintk("svcrdma: error creating RQ CQ for connect request\n"); 898 goto errout; 899 } 900 901 memset(&qp_attr, 0, sizeof qp_attr); 902 qp_attr.event_handler = qp_event_handler; 903 qp_attr.qp_context = &newxprt->sc_xprt; 904 qp_attr.cap.max_send_wr = newxprt->sc_sq_depth; 905 qp_attr.cap.max_recv_wr = newxprt->sc_max_requests; 906 qp_attr.cap.max_send_sge = newxprt->sc_max_sge; 907 qp_attr.cap.max_recv_sge = newxprt->sc_max_sge; 908 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR; 909 qp_attr.qp_type = IB_QPT_RC; 910 qp_attr.send_cq = newxprt->sc_sq_cq; 911 qp_attr.recv_cq = newxprt->sc_rq_cq; 912 dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n" 913 " cm_id->device=%p, sc_pd->device=%p\n" 914 " cap.max_send_wr = %d\n" 915 " cap.max_recv_wr = %d\n" 916 " cap.max_send_sge = %d\n" 917 " cap.max_recv_sge = %d\n", 918 newxprt->sc_cm_id, newxprt->sc_pd, 919 newxprt->sc_cm_id->device, newxprt->sc_pd->device, 920 qp_attr.cap.max_send_wr, 921 qp_attr.cap.max_recv_wr, 922 qp_attr.cap.max_send_sge, 923 qp_attr.cap.max_recv_sge); 924 925 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr); 926 if (ret) { 927 /* 928 * XXX: This is a hack. We need a xx_request_qp interface 929 * that will adjust the qp_attr's with a best-effort 930 * number 931 */ 932 qp_attr.cap.max_send_sge -= 2; 933 qp_attr.cap.max_recv_sge -= 2; 934 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, 935 &qp_attr); 936 if (ret) { 937 dprintk("svcrdma: failed to create QP, ret=%d\n", ret); 938 goto errout; 939 } 940 newxprt->sc_max_sge = qp_attr.cap.max_send_sge; 941 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge; 942 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr; 943 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr; 944 } 945 newxprt->sc_qp = newxprt->sc_cm_id->qp; 946 947 /* 948 * Use the most secure set of MR resources based on the 949 * transport type and available memory management features in 950 * the device. Here's the table implemented below: 951 * 952 * Fast Global DMA Remote WR 953 * Reg LKEY MR Access 954 * Sup'd Sup'd Needed Needed 955 * 956 * IWARP N N Y Y 957 * N Y Y Y 958 * Y N Y N 959 * Y Y N - 960 * 961 * IB N N Y N 962 * N Y N - 963 * Y N Y N 964 * Y Y N - 965 * 966 * NB: iWARP requires remote write access for the data sink 967 * of an RDMA_READ. IB does not. 968 */ 969 if (devattr.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 970 newxprt->sc_frmr_pg_list_len = 971 devattr.max_fast_reg_page_list_len; 972 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_FAST_REG; 973 } 974 975 /* 976 * Determine if a DMA MR is required and if so, what privs are required 977 */ 978 switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { 979 case RDMA_TRANSPORT_IWARP: 980 newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; 981 if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { 982 need_dma_mr = 1; 983 dma_mr_acc = 984 (IB_ACCESS_LOCAL_WRITE | 985 IB_ACCESS_REMOTE_WRITE); 986 } else if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 987 need_dma_mr = 1; 988 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 989 } else 990 need_dma_mr = 0; 991 break; 992 case RDMA_TRANSPORT_IB: 993 if (!(devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)) { 994 need_dma_mr = 1; 995 dma_mr_acc = IB_ACCESS_LOCAL_WRITE; 996 } else 997 need_dma_mr = 0; 998 break; 999 default: 1000 goto errout; 1001 } 1002 1003 /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ 1004 if (need_dma_mr) { 1005 /* Register all of physical memory */ 1006 newxprt->sc_phys_mr = 1007 ib_get_dma_mr(newxprt->sc_pd, dma_mr_acc); 1008 if (IS_ERR(newxprt->sc_phys_mr)) { 1009 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", 1010 ret); 1011 goto errout; 1012 } 1013 newxprt->sc_dma_lkey = newxprt->sc_phys_mr->lkey; 1014 } else 1015 newxprt->sc_dma_lkey = 1016 newxprt->sc_cm_id->device->local_dma_lkey; 1017 1018 /* Post receive buffers */ 1019 for (i = 0; i < newxprt->sc_max_requests; i++) { 1020 ret = svc_rdma_post_recv(newxprt); 1021 if (ret) { 1022 dprintk("svcrdma: failure posting receive buffers\n"); 1023 goto errout; 1024 } 1025 } 1026 1027 /* Swap out the handler */ 1028 newxprt->sc_cm_id->event_handler = rdma_cma_handler; 1029 1030 /* 1031 * Arm the CQs for the SQ and RQ before accepting so we can't 1032 * miss the first message 1033 */ 1034 ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP); 1035 ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP); 1036 1037 /* Accept Connection */ 1038 set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags); 1039 memset(&conn_param, 0, sizeof conn_param); 1040 conn_param.responder_resources = 0; 1041 conn_param.initiator_depth = newxprt->sc_ord; 1042 ret = rdma_accept(newxprt->sc_cm_id, &conn_param); 1043 if (ret) { 1044 dprintk("svcrdma: failed to accept new connection, ret=%d\n", 1045 ret); 1046 goto errout; 1047 } 1048 1049 dprintk("svcrdma: new connection %p accepted with the following " 1050 "attributes:\n" 1051 " local_ip : %d.%d.%d.%d\n" 1052 " local_port : %d\n" 1053 " remote_ip : %d.%d.%d.%d\n" 1054 " remote_port : %d\n" 1055 " max_sge : %d\n" 1056 " sq_depth : %d\n" 1057 " max_requests : %d\n" 1058 " ord : %d\n", 1059 newxprt, 1060 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1061 route.addr.src_addr)->sin_addr.s_addr), 1062 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1063 route.addr.src_addr)->sin_port), 1064 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1065 route.addr.dst_addr)->sin_addr.s_addr), 1066 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id-> 1067 route.addr.dst_addr)->sin_port), 1068 newxprt->sc_max_sge, 1069 newxprt->sc_sq_depth, 1070 newxprt->sc_max_requests, 1071 newxprt->sc_ord); 1072 1073 return &newxprt->sc_xprt; 1074 1075 errout: 1076 dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret); 1077 /* Take a reference in case the DTO handler runs */ 1078 svc_xprt_get(&newxprt->sc_xprt); 1079 if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) 1080 ib_destroy_qp(newxprt->sc_qp); 1081 rdma_destroy_id(newxprt->sc_cm_id); 1082 /* This call to put will destroy the transport */ 1083 svc_xprt_put(&newxprt->sc_xprt); 1084 return NULL; 1085 } 1086 1087 static void svc_rdma_release_rqst(struct svc_rqst *rqstp) 1088 { 1089 } 1090 1091 /* 1092 * When connected, an svc_xprt has at least two references: 1093 * 1094 * - A reference held by the cm_id between the ESTABLISHED and 1095 * DISCONNECTED events. If the remote peer disconnected first, this 1096 * reference could be gone. 1097 * 1098 * - A reference held by the svc_recv code that called this function 1099 * as part of close processing. 1100 * 1101 * At a minimum one references should still be held. 1102 */ 1103 static void svc_rdma_detach(struct svc_xprt *xprt) 1104 { 1105 struct svcxprt_rdma *rdma = 1106 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1107 dprintk("svc: svc_rdma_detach(%p)\n", xprt); 1108 1109 /* Disconnect and flush posted WQE */ 1110 rdma_disconnect(rdma->sc_cm_id); 1111 } 1112 1113 static void __svc_rdma_free(struct work_struct *work) 1114 { 1115 struct svcxprt_rdma *rdma = 1116 container_of(work, struct svcxprt_rdma, sc_work); 1117 dprintk("svcrdma: svc_rdma_free(%p)\n", rdma); 1118 1119 /* We should only be called from kref_put */ 1120 BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0); 1121 1122 /* 1123 * Destroy queued, but not processed read completions. Note 1124 * that this cleanup has to be done before destroying the 1125 * cm_id because the device ptr is needed to unmap the dma in 1126 * svc_rdma_put_context. 1127 */ 1128 while (!list_empty(&rdma->sc_read_complete_q)) { 1129 struct svc_rdma_op_ctxt *ctxt; 1130 ctxt = list_entry(rdma->sc_read_complete_q.next, 1131 struct svc_rdma_op_ctxt, 1132 dto_q); 1133 list_del_init(&ctxt->dto_q); 1134 svc_rdma_put_context(ctxt, 1); 1135 } 1136 1137 /* Destroy queued, but not processed recv completions */ 1138 while (!list_empty(&rdma->sc_rq_dto_q)) { 1139 struct svc_rdma_op_ctxt *ctxt; 1140 ctxt = list_entry(rdma->sc_rq_dto_q.next, 1141 struct svc_rdma_op_ctxt, 1142 dto_q); 1143 list_del_init(&ctxt->dto_q); 1144 svc_rdma_put_context(ctxt, 1); 1145 } 1146 1147 /* Warn if we leaked a resource or under-referenced */ 1148 WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0); 1149 WARN_ON(atomic_read(&rdma->sc_dma_used) != 0); 1150 1151 /* De-allocate fastreg mr */ 1152 rdma_dealloc_frmr_q(rdma); 1153 1154 /* Destroy the QP if present (not a listener) */ 1155 if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) 1156 ib_destroy_qp(rdma->sc_qp); 1157 1158 if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq)) 1159 ib_destroy_cq(rdma->sc_sq_cq); 1160 1161 if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq)) 1162 ib_destroy_cq(rdma->sc_rq_cq); 1163 1164 if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr)) 1165 ib_dereg_mr(rdma->sc_phys_mr); 1166 1167 if (rdma->sc_pd && !IS_ERR(rdma->sc_pd)) 1168 ib_dealloc_pd(rdma->sc_pd); 1169 1170 /* Destroy the CM ID */ 1171 rdma_destroy_id(rdma->sc_cm_id); 1172 1173 kfree(rdma); 1174 } 1175 1176 static void svc_rdma_free(struct svc_xprt *xprt) 1177 { 1178 struct svcxprt_rdma *rdma = 1179 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1180 INIT_WORK(&rdma->sc_work, __svc_rdma_free); 1181 schedule_work(&rdma->sc_work); 1182 } 1183 1184 static int svc_rdma_has_wspace(struct svc_xprt *xprt) 1185 { 1186 struct svcxprt_rdma *rdma = 1187 container_of(xprt, struct svcxprt_rdma, sc_xprt); 1188 1189 /* 1190 * If there are fewer SQ WR available than required to send a 1191 * simple response, return false. 1192 */ 1193 if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3)) 1194 return 0; 1195 1196 /* 1197 * ...or there are already waiters on the SQ, 1198 * return false. 1199 */ 1200 if (waitqueue_active(&rdma->sc_send_wait)) 1201 return 0; 1202 1203 /* Otherwise return true. */ 1204 return 1; 1205 } 1206 1207 /* 1208 * Attempt to register the kvec representing the RPC memory with the 1209 * device. 1210 * 1211 * Returns: 1212 * NULL : The device does not support fastreg or there were no more 1213 * fastreg mr. 1214 * frmr : The kvec register request was successfully posted. 1215 * <0 : An error was encountered attempting to register the kvec. 1216 */ 1217 int svc_rdma_fastreg(struct svcxprt_rdma *xprt, 1218 struct svc_rdma_fastreg_mr *frmr) 1219 { 1220 struct ib_send_wr fastreg_wr; 1221 u8 key; 1222 1223 /* Bump the key */ 1224 key = (u8)(frmr->mr->lkey & 0x000000FF); 1225 ib_update_fast_reg_key(frmr->mr, ++key); 1226 1227 /* Prepare FASTREG WR */ 1228 memset(&fastreg_wr, 0, sizeof fastreg_wr); 1229 fastreg_wr.opcode = IB_WR_FAST_REG_MR; 1230 fastreg_wr.send_flags = IB_SEND_SIGNALED; 1231 fastreg_wr.wr.fast_reg.iova_start = (unsigned long)frmr->kva; 1232 fastreg_wr.wr.fast_reg.page_list = frmr->page_list; 1233 fastreg_wr.wr.fast_reg.page_list_len = frmr->page_list_len; 1234 fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT; 1235 fastreg_wr.wr.fast_reg.length = frmr->map_len; 1236 fastreg_wr.wr.fast_reg.access_flags = frmr->access_flags; 1237 fastreg_wr.wr.fast_reg.rkey = frmr->mr->lkey; 1238 return svc_rdma_send(xprt, &fastreg_wr); 1239 } 1240 1241 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr) 1242 { 1243 struct ib_send_wr *bad_wr, *n_wr; 1244 int wr_count; 1245 int i; 1246 int ret; 1247 1248 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1249 return -ENOTCONN; 1250 1251 BUG_ON(wr->send_flags != IB_SEND_SIGNALED); 1252 wr_count = 1; 1253 for (n_wr = wr->next; n_wr; n_wr = n_wr->next) 1254 wr_count++; 1255 1256 /* If the SQ is full, wait until an SQ entry is available */ 1257 while (1) { 1258 spin_lock_bh(&xprt->sc_lock); 1259 if (xprt->sc_sq_depth < atomic_read(&xprt->sc_sq_count) + wr_count) { 1260 spin_unlock_bh(&xprt->sc_lock); 1261 atomic_inc(&rdma_stat_sq_starve); 1262 1263 /* See if we can opportunistically reap SQ WR to make room */ 1264 sq_cq_reap(xprt); 1265 1266 /* Wait until SQ WR available if SQ still full */ 1267 wait_event(xprt->sc_send_wait, 1268 atomic_read(&xprt->sc_sq_count) < 1269 xprt->sc_sq_depth); 1270 if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags)) 1271 return 0; 1272 continue; 1273 } 1274 /* Take a transport ref for each WR posted */ 1275 for (i = 0; i < wr_count; i++) 1276 svc_xprt_get(&xprt->sc_xprt); 1277 1278 /* Bump used SQ WR count and post */ 1279 atomic_add(wr_count, &xprt->sc_sq_count); 1280 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr); 1281 if (ret) { 1282 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 1283 atomic_sub(wr_count, &xprt->sc_sq_count); 1284 for (i = 0; i < wr_count; i ++) 1285 svc_xprt_put(&xprt->sc_xprt); 1286 dprintk("svcrdma: failed to post SQ WR rc=%d, " 1287 "sc_sq_count=%d, sc_sq_depth=%d\n", 1288 ret, atomic_read(&xprt->sc_sq_count), 1289 xprt->sc_sq_depth); 1290 } 1291 spin_unlock_bh(&xprt->sc_lock); 1292 if (ret) 1293 wake_up(&xprt->sc_send_wait); 1294 break; 1295 } 1296 return ret; 1297 } 1298 1299 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp, 1300 enum rpcrdma_errcode err) 1301 { 1302 struct ib_send_wr err_wr; 1303 struct ib_sge sge; 1304 struct page *p; 1305 struct svc_rdma_op_ctxt *ctxt; 1306 u32 *va; 1307 int length; 1308 int ret; 1309 1310 p = svc_rdma_get_page(); 1311 va = page_address(p); 1312 1313 /* XDR encode error */ 1314 length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va); 1315 1316 /* Prepare SGE for local address */ 1317 sge.addr = ib_dma_map_page(xprt->sc_cm_id->device, 1318 p, 0, PAGE_SIZE, DMA_FROM_DEVICE); 1319 if (ib_dma_mapping_error(xprt->sc_cm_id->device, sge.addr)) { 1320 put_page(p); 1321 return; 1322 } 1323 atomic_inc(&xprt->sc_dma_used); 1324 sge.lkey = xprt->sc_dma_lkey; 1325 sge.length = length; 1326 1327 ctxt = svc_rdma_get_context(xprt); 1328 ctxt->count = 1; 1329 ctxt->pages[0] = p; 1330 1331 /* Prepare SEND WR */ 1332 memset(&err_wr, 0, sizeof err_wr); 1333 ctxt->wr_op = IB_WR_SEND; 1334 err_wr.wr_id = (unsigned long)ctxt; 1335 err_wr.sg_list = &sge; 1336 err_wr.num_sge = 1; 1337 err_wr.opcode = IB_WR_SEND; 1338 err_wr.send_flags = IB_SEND_SIGNALED; 1339 1340 /* Post It */ 1341 ret = svc_rdma_send(xprt, &err_wr); 1342 if (ret) { 1343 dprintk("svcrdma: Error %d posting send for protocol error\n", 1344 ret); 1345 ib_dma_unmap_page(xprt->sc_cm_id->device, 1346 sge.addr, PAGE_SIZE, 1347 DMA_FROM_DEVICE); 1348 svc_rdma_put_context(ctxt, 1); 1349 } 1350 } 1351