SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/sunrpc/xprtrdma/svc_rdma_transport.c

   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 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
  88 {
  89         int target;
  90         int at_least_one = 0;
  91         struct svc_rdma_op_ctxt *ctxt;
  92
  93         target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
  94                      xprt->sc_ctxt_max);
  95
  96         spin_lock_bh(&xprt->sc_ctxt_lock);
  97         while (xprt->sc_ctxt_cnt < target) {
  98                 xprt->sc_ctxt_cnt++;
  99                 spin_unlock_bh(&xprt->sc_ctxt_lock);
 100
 101                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 102
 103                 spin_lock_bh(&xprt->sc_ctxt_lock);
 104                 if (ctxt) {
 105                         at_least_one = 1;
 106                         ctxt->next = xprt->sc_ctxt_head;
 107                         xprt->sc_ctxt_head = ctxt;
 108                 } else {
 109                         /* kmalloc failed...give up for now */
 110                         xprt->sc_ctxt_cnt--;
 111                         break;
 112                 }
 113         }
 114         spin_unlock_bh(&xprt->sc_ctxt_lock);
 115         dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
 116                 xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
 117         return at_least_one;
 118 }
 119
 120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
 121 {
 122         struct svc_rdma_op_ctxt *ctxt;
 123
 124         while (1) {
 125                 spin_lock_bh(&xprt->sc_ctxt_lock);
 126                 if (unlikely(xprt->sc_ctxt_head == NULL)) {
 127                         /* Try to bump my cache. */
 128                         spin_unlock_bh(&xprt->sc_ctxt_lock);
 129
 130                         if (rdma_bump_context_cache(xprt))
 131                                 continue;
 132
 133                         printk(KERN_INFO "svcrdma: sleeping waiting for "
 134                                "context memory on xprt=%p\n",
 135                                xprt);
 136                         schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 137                         continue;
 138                 }
 139                 ctxt = xprt->sc_ctxt_head;
 140                 xprt->sc_ctxt_head = ctxt->next;
 141                 spin_unlock_bh(&xprt->sc_ctxt_lock);
 142                 ctxt->xprt = xprt;
 143                 INIT_LIST_HEAD(&ctxt->dto_q);
 144                 ctxt->count = 0;
 145                 break;
 146         }
 147         return ctxt;
 148 }
 149
 150 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
 151 {
 152         struct svcxprt_rdma *xprt;
 153         int i;
 154
 155         BUG_ON(!ctxt);
 156         xprt = ctxt->xprt;
 157         if (free_pages)
 158                 for (i = 0; i < ctxt->count; i++)
 159                         put_page(ctxt->pages[i]);
 160
 161         for (i = 0; i < ctxt->count; i++)
 162                 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
 163                                  ctxt->sge[i].addr,
 164                                  ctxt->sge[i].length,
 165                                  ctxt->direction);
 166         spin_lock_bh(&xprt->sc_ctxt_lock);
 167         ctxt->next = xprt->sc_ctxt_head;
 168         xprt->sc_ctxt_head = ctxt;
 169         spin_unlock_bh(&xprt->sc_ctxt_lock);
 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         /*
 252          * Set the bit regardless of whether or not it's on the list
 253          * because it may be on the list already due to an SQ
 254          * completion.
 255         */
 256         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 257
 258         /*
 259          * If this transport is not already on the DTO transport queue,
 260          * add it
 261          */
 262         spin_lock_irqsave(&dto_lock, flags);
 263         if (list_empty(&xprt->sc_dto_q)) {
 264                 svc_xprt_get(&xprt->sc_xprt);
 265                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 266         }
 267         spin_unlock_irqrestore(&dto_lock, flags);
 268
 269         /* Tasklet does all the work to avoid irqsave locks. */
 270         tasklet_schedule(&dto_tasklet);
 271 }
 272
 273 /*
 274  * rq_cq_reap - Process the RQ CQ.
 275  *
 276  * Take all completing WC off the CQE and enqueue the associated DTO
 277  * context on the dto_q for the transport.
 278  */
 279 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 280 {
 281         int ret;
 282         struct ib_wc wc;
 283         struct svc_rdma_op_ctxt *ctxt = NULL;
 284
 285         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
 286                 return;
 287
 288         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 289         atomic_inc(&rdma_stat_rq_poll);
 290
 291         spin_lock_bh(&xprt->sc_rq_dto_lock);
 292         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 293                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 294                 ctxt->wc_status = wc.status;
 295                 ctxt->byte_len = wc.byte_len;
 296                 if (wc.status != IB_WC_SUCCESS) {
 297                         /* Close the transport */
 298                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 299                         svc_rdma_put_context(ctxt, 1);
 300                         continue;
 301                 }
 302                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 303         }
 304         spin_unlock_bh(&xprt->sc_rq_dto_lock);
 305
 306         if (ctxt)
 307                 atomic_inc(&rdma_stat_rq_prod);
 308
 309         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 310         /*
 311          * If data arrived before established event,
 312          * don't enqueue. This defers RPC I/O until the
 313          * RDMA connection is complete.
 314          */
 315         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 316                 svc_xprt_enqueue(&xprt->sc_xprt);
 317 }
 318
 319 /*
 320  * Send Queue Completion Handler - potentially called on interrupt context.
 321  */
 322 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 323 {
 324         struct svc_rdma_op_ctxt *ctxt = NULL;
 325         struct ib_wc wc;
 326         struct ib_cq *cq = xprt->sc_sq_cq;
 327         int ret;
 328
 329
 330         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 331                 return;
 332
 333         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 334         atomic_inc(&rdma_stat_sq_poll);
 335         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 336                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 337                 xprt = ctxt->xprt;
 338
 339                 if (wc.status != IB_WC_SUCCESS)
 340                         /* Close the transport */
 341                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 342
 343                 /* Decrement used SQ WR count */
 344                 atomic_dec(&xprt->sc_sq_count);
 345                 wake_up(&xprt->sc_send_wait);
 346
 347                 switch (ctxt->wr_op) {
 348                 case IB_WR_SEND:
 349                 case IB_WR_RDMA_WRITE:
 350                         svc_rdma_put_context(ctxt, 1);
 351                         break;
 352
 353                 case IB_WR_RDMA_READ:
 354                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 355                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 356                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 357                                 list_add_tail(&ctxt->dto_q,
 358                                               &xprt->sc_read_complete_q);
 359                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 360                                 svc_xprt_enqueue(&xprt->sc_xprt);
 361                         }
 362                         break;
 363
 364                 default:
 365                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 366                                "opcode=%d, status=%d\n",
 367                                wc.opcode, wc.status);
 368                         break;
 369                 }
 370         }
 371
 372         if (ctxt)
 373                 atomic_inc(&rdma_stat_sq_prod);
 374 }
 375
 376 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 377 {
 378         struct svcxprt_rdma *xprt = cq_context;
 379         unsigned long flags;
 380
 381         /*
 382          * Set the bit regardless of whether or not it's on the list
 383          * because it may be on the list already due to an RQ
 384          * completion.
 385         */
 386         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 387
 388         /*
 389          * If this transport is not already on the DTO transport queue,
 390          * add it
 391          */
 392         spin_lock_irqsave(&dto_lock, flags);
 393         if (list_empty(&xprt->sc_dto_q)) {
 394                 svc_xprt_get(&xprt->sc_xprt);
 395                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 396         }
 397         spin_unlock_irqrestore(&dto_lock, flags);
 398
 399         /* Tasklet does all the work to avoid irqsave locks. */
 400         tasklet_schedule(&dto_tasklet);
 401 }
 402
 403 static void create_context_cache(struct svcxprt_rdma *xprt,
 404                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 405 {
 406         struct svc_rdma_op_ctxt *ctxt;
 407         int i;
 408
 409         xprt->sc_ctxt_max = ctxt_max;
 410         xprt->sc_ctxt_bump = ctxt_bump;
 411         xprt->sc_ctxt_cnt = 0;
 412         xprt->sc_ctxt_head = NULL;
 413         for (i = 0; i < ctxt_count; i++) {
 414                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 415                 if (ctxt) {
 416                         ctxt->next = xprt->sc_ctxt_head;
 417                         xprt->sc_ctxt_head = ctxt;
 418                         xprt->sc_ctxt_cnt++;
 419                 }
 420         }
 421 }
 422
 423 static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
 424 {
 425         struct svc_rdma_op_ctxt *next;
 426         if (!ctxt)
 427                 return;
 428
 429         do {
 430                 next = ctxt->next;
 431                 kfree(ctxt);
 432                 ctxt = next;
 433         } while (next);
 434 }
 435
 436 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 437                                              int listener)
 438 {
 439         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 440
 441         if (!cma_xprt)
 442                 return NULL;
 443         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 444         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 445         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 446         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 447         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 448         init_waitqueue_head(&cma_xprt->sc_send_wait);
 449
 450         spin_lock_init(&cma_xprt->sc_lock);
 451         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 452         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 453         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 454
 455         cma_xprt->sc_ord = svcrdma_ord;
 456
 457         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 458         cma_xprt->sc_max_requests = svcrdma_max_requests;
 459         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 460         atomic_set(&cma_xprt->sc_sq_count, 0);
 461
 462         if (!listener) {
 463                 int reqs = cma_xprt->sc_max_requests;
 464                 create_context_cache(cma_xprt,
 465                                      reqs << 1, /* starting size */
 466                                      reqs,      /* bump amount */
 467                                      reqs +
 468                                      cma_xprt->sc_sq_depth +
 469                                      RPCRDMA_MAX_THREADS + 1); /* max */
 470                 if (!cma_xprt->sc_ctxt_head) {
 471                         kfree(cma_xprt);
 472                         return NULL;
 473                 }
 474                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 475         } else
 476                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 477
 478         return cma_xprt;
 479 }
 480
 481 struct page *svc_rdma_get_page(void)
 482 {
 483         struct page *page;
 484
 485         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 486                 /* If we can't get memory, wait a bit and try again */
 487                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 488                        "jiffies.\n");
 489                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 490         }
 491         return page;
 492 }
 493
 494 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 495 {
 496         struct ib_recv_wr recv_wr, *bad_recv_wr;
 497         struct svc_rdma_op_ctxt *ctxt;
 498         struct page *page;
 499         unsigned long pa;
 500         int sge_no;
 501         int buflen;
 502         int ret;
 503
 504         ctxt = svc_rdma_get_context(xprt);
 505         buflen = 0;
 506         ctxt->direction = DMA_FROM_DEVICE;
 507         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 508                 BUG_ON(sge_no >= xprt->sc_max_sge);
 509                 page = svc_rdma_get_page();
 510                 ctxt->pages[sge_no] = page;
 511                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 512                                      page, 0, PAGE_SIZE,
 513                                      DMA_FROM_DEVICE);
 514                 ctxt->sge[sge_no].addr = pa;
 515                 ctxt->sge[sge_no].length = PAGE_SIZE;
 516                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 517                 buflen += PAGE_SIZE;
 518         }
 519         ctxt->count = sge_no;
 520         recv_wr.next = NULL;
 521         recv_wr.sg_list = &ctxt->sge[0];
 522         recv_wr.num_sge = ctxt->count;
 523         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 524
 525         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 526         if (ret)
 527                 svc_rdma_put_context(ctxt, 1);
 528         return ret;
 529 }
 530
 531 /*
 532  * This function handles the CONNECT_REQUEST event on a listening
 533  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 534  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 535  * structure for the listening endpoint.
 536  *
 537  * This function creates a new xprt for the new connection and enqueues it on
 538  * the accept queue for the listent xprt. When the listen thread is kicked, it
 539  * will call the recvfrom method on the listen xprt which will accept the new
 540  * connection.
 541  */
 542 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 543 {
 544         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 545         struct svcxprt_rdma *newxprt;
 546
 547         /* Create a new transport */
 548         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 549         if (!newxprt) {
 550                 dprintk("svcrdma: failed to create new transport\n");
 551                 return;
 552         }
 553         newxprt->sc_cm_id = new_cma_id;
 554         new_cma_id->context = newxprt;
 555         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 556                 newxprt, newxprt->sc_cm_id, listen_xprt);
 557
 558         /*
 559          * Enqueue the new transport on the accept queue of the listening
 560          * transport
 561          */
 562         spin_lock_bh(&listen_xprt->sc_lock);
 563         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 564         spin_unlock_bh(&listen_xprt->sc_lock);
 565
 566         /*
 567          * Can't use svc_xprt_received here because we are not on a
 568          * rqstp thread
 569         */
 570         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 571         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 572 }
 573
 574 /*
 575  * Handles events generated on the listening endpoint. These events will be
 576  * either be incoming connect requests or adapter removal  events.
 577  */
 578 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 579                                struct rdma_cm_event *event)
 580 {
 581         struct svcxprt_rdma *xprt = cma_id->context;
 582         int ret = 0;
 583
 584         switch (event->event) {
 585         case RDMA_CM_EVENT_CONNECT_REQUEST:
 586                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 587                         "event=%d\n", cma_id, cma_id->context, event->event);
 588                 handle_connect_req(cma_id);
 589                 break;
 590
 591         case RDMA_CM_EVENT_ESTABLISHED:
 592                 /* Accept complete */
 593                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 594                         "cm_id=%p\n", xprt, cma_id);
 595                 break;
 596
 597         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 598                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 599                         xprt, cma_id);
 600                 if (xprt)
 601                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 602                 break;
 603
 604         default:
 605                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 606                         "event=%d\n", cma_id, event->event);
 607                 break;
 608         }
 609
 610         return ret;
 611 }
 612
 613 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 614                             struct rdma_cm_event *event)
 615 {
 616         struct svc_xprt *xprt = cma_id->context;
 617         struct svcxprt_rdma *rdma =
 618                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 619         switch (event->event) {
 620         case RDMA_CM_EVENT_ESTABLISHED:
 621                 /* Accept complete */
 622                 svc_xprt_get(xprt);
 623                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 624                         "cm_id=%p\n", xprt, cma_id);
 625                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 626                 svc_xprt_enqueue(xprt);
 627                 break;
 628         case RDMA_CM_EVENT_DISCONNECTED:
 629                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 630                         xprt, cma_id);
 631                 if (xprt) {
 632                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 633                         svc_xprt_enqueue(xprt);
 634                         svc_xprt_put(xprt);
 635                 }
 636                 break;
 637         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 638                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 639                         "event=%d\n", cma_id, xprt, event->event);
 640                 if (xprt) {
 641                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 642                         svc_xprt_enqueue(xprt);
 643                 }
 644                 break;
 645         default:
 646                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 647                         "event=%d\n", cma_id, event->event);
 648                 break;
 649         }
 650         return 0;
 651 }
 652
 653 /*
 654  * Create a listening RDMA service endpoint.
 655  */
 656 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 657                                         struct sockaddr *sa, int salen,
 658                                         int flags)
 659 {
 660         struct rdma_cm_id *listen_id;
 661         struct svcxprt_rdma *cma_xprt;
 662         struct svc_xprt *xprt;
 663         int ret;
 664
 665         dprintk("svcrdma: Creating RDMA socket\n");
 666
 667         cma_xprt = rdma_create_xprt(serv, 1);
 668         if (!cma_xprt)
 669                 return ERR_PTR(-ENOMEM);
 670         xprt = &cma_xprt->sc_xprt;
 671
 672         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 673         if (IS_ERR(listen_id)) {
 674                 ret = PTR_ERR(listen_id);
 675                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
 676                 goto err0;
 677         }
 678
 679         ret = rdma_bind_addr(listen_id, sa);
 680         if (ret) {
 681                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 682                 goto err1;
 683         }
 684         cma_xprt->sc_cm_id = listen_id;
 685
 686         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 687         if (ret) {
 688                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 689                 goto err1;
 690         }
 691
 692         /*
 693          * We need to use the address from the cm_id in case the
 694          * caller specified 0 for the port number.
 695          */
 696         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 697         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 698
 699         return &cma_xprt->sc_xprt;
 700
 701  err1:
 702         rdma_destroy_id(listen_id);
 703  err0:
 704         kfree(cma_xprt);
 705         return ERR_PTR(ret);
 706 }
 707
 708 /*
 709  * This is the xpo_recvfrom function for listening endpoints. Its
 710  * purpose is to accept incoming connections. The CMA callback handler
 711  * has already created a new transport and attached it to the new CMA
 712  * ID.
 713  *
 714  * There is a queue of pending connections hung on the listening
 715  * transport. This queue contains the new svc_xprt structure. This
 716  * function takes svc_xprt structures off the accept_q and completes
 717  * the connection.
 718  */
 719 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 720 {
 721         struct svcxprt_rdma *listen_rdma;
 722         struct svcxprt_rdma *newxprt = NULL;
 723         struct rdma_conn_param conn_param;
 724         struct ib_qp_init_attr qp_attr;
 725         struct ib_device_attr devattr;
 726         struct sockaddr *sa;
 727         int ret;
 728         int i;
 729
 730         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 731         clear_bit(XPT_CONN, &xprt->xpt_flags);
 732         /* Get the next entry off the accept list */
 733         spin_lock_bh(&listen_rdma->sc_lock);
 734         if (!list_empty(&listen_rdma->sc_accept_q)) {
 735                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 736                                      struct svcxprt_rdma, sc_accept_q);
 737                 list_del_init(&newxprt->sc_accept_q);
 738         }
 739         if (!list_empty(&listen_rdma->sc_accept_q))
 740                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 741         spin_unlock_bh(&listen_rdma->sc_lock);
 742         if (!newxprt)
 743                 return NULL;
 744
 745         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 746                 newxprt, newxprt->sc_cm_id);
 747
 748         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 749         if (ret) {
 750                 dprintk("svcrdma: could not query device attributes on "
 751                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 752                 goto errout;
 753         }
 754
 755         /* Qualify the transport resource defaults with the
 756          * capabilities of this particular device */
 757         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 758                                   (size_t)RPCSVC_MAXPAGES);
 759         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 760                                    (size_t)svcrdma_max_requests);
 761         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 762
 763         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 764                                (size_t)svcrdma_ord);
 765
 766         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 767         if (IS_ERR(newxprt->sc_pd)) {
 768                 dprintk("svcrdma: error creating PD for connect request\n");
 769                 goto errout;
 770         }
 771         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 772                                          sq_comp_handler,
 773                                          cq_event_handler,
 774                                          newxprt,
 775                                          newxprt->sc_sq_depth,
 776                                          0);
 777         if (IS_ERR(newxprt->sc_sq_cq)) {
 778                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 779                 goto errout;
 780         }
 781         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 782                                          rq_comp_handler,
 783                                          cq_event_handler,
 784                                          newxprt,
 785                                          newxprt->sc_max_requests,
 786                                          0);
 787         if (IS_ERR(newxprt->sc_rq_cq)) {
 788                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 789                 goto errout;
 790         }
 791
 792         memset(&qp_attr, 0, sizeof qp_attr);
 793         qp_attr.event_handler = qp_event_handler;
 794         qp_attr.qp_context = &newxprt->sc_xprt;
 795         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 796         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 797         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 798         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 799         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 800         qp_attr.qp_type = IB_QPT_RC;
 801         qp_attr.send_cq = newxprt->sc_sq_cq;
 802         qp_attr.recv_cq = newxprt->sc_rq_cq;
 803         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 804                 "    cm_id->device=%p, sc_pd->device=%p\n"
 805                 "    cap.max_send_wr = %d\n"
 806                 "    cap.max_recv_wr = %d\n"
 807                 "    cap.max_send_sge = %d\n"
 808                 "    cap.max_recv_sge = %d\n",
 809                 newxprt->sc_cm_id, newxprt->sc_pd,
 810                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 811                 qp_attr.cap.max_send_wr,
 812                 qp_attr.cap.max_recv_wr,
 813                 qp_attr.cap.max_send_sge,
 814                 qp_attr.cap.max_recv_sge);
 815
 816         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 817         if (ret) {
 818                 /*
 819                  * XXX: This is a hack. We need a xx_request_qp interface
 820                  * that will adjust the qp_attr's with a best-effort
 821                  * number
 822                  */
 823                 qp_attr.cap.max_send_sge -= 2;
 824                 qp_attr.cap.max_recv_sge -= 2;
 825                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 826                                      &qp_attr);
 827                 if (ret) {
 828                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 829                         goto errout;
 830                 }
 831                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 832                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 833                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 834                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 835         }
 836         svc_xprt_get(&newxprt->sc_xprt);
 837         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 838
 839         /* Register all of physical memory */
 840         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 841                                             IB_ACCESS_LOCAL_WRITE |
 842                                             IB_ACCESS_REMOTE_WRITE);
 843         if (IS_ERR(newxprt->sc_phys_mr)) {
 844                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 845                 goto errout;
 846         }
 847
 848         /* Post receive buffers */
 849         for (i = 0; i < newxprt->sc_max_requests; i++) {
 850                 ret = svc_rdma_post_recv(newxprt);
 851                 if (ret) {
 852                         dprintk("svcrdma: failure posting receive buffers\n");
 853                         goto errout;
 854                 }
 855         }
 856
 857         /* Swap out the handler */
 858         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 859
 860         /* Accept Connection */
 861         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 862         memset(&conn_param, 0, sizeof conn_param);
 863         conn_param.responder_resources = 0;
 864         conn_param.initiator_depth = newxprt->sc_ord;
 865         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 866         if (ret) {
 867                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 868                        ret);
 869                 goto errout;
 870         }
 871
 872         dprintk("svcrdma: new connection %p accepted with the following "
 873                 "attributes:\n"
 874                 "    local_ip        : %d.%d.%d.%d\n"
 875                 "    local_port      : %d\n"
 876                 "    remote_ip       : %d.%d.%d.%d\n"
 877                 "    remote_port     : %d\n"
 878                 "    max_sge         : %d\n"
 879                 "    sq_depth        : %d\n"
 880                 "    max_requests    : %d\n"
 881                 "    ord             : %d\n",
 882                 newxprt,
 883                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 884                          route.addr.src_addr)->sin_addr.s_addr),
 885                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 886                        route.addr.src_addr)->sin_port),
 887                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 888                          route.addr.dst_addr)->sin_addr.s_addr),
 889                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 890                        route.addr.dst_addr)->sin_port),
 891                 newxprt->sc_max_sge,
 892                 newxprt->sc_sq_depth,
 893                 newxprt->sc_max_requests,
 894                 newxprt->sc_ord);
 895
 896         /* Set the local and remote addresses in the transport */
 897         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 898         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 899         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 900         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 901
 902         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 903         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 904         return &newxprt->sc_xprt;
 905
 906  errout:
 907         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 908         /* Take a reference in case the DTO handler runs */
 909         svc_xprt_get(&newxprt->sc_xprt);
 910         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
 911                 ib_destroy_qp(newxprt->sc_qp);
 912                 svc_xprt_put(&newxprt->sc_xprt);
 913         }
 914         rdma_destroy_id(newxprt->sc_cm_id);
 915         /* This call to put will destroy the transport */
 916         svc_xprt_put(&newxprt->sc_xprt);
 917         return NULL;
 918 }
 919
 920 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 921 {
 922 }
 923
 924 /*
 925  * When connected, an svc_xprt has at least three references:
 926  *
 927  * - A reference held by the QP. We still hold that here because this
 928  *   code deletes the QP and puts the reference.
 929  *
 930  * - A reference held by the cm_id between the ESTABLISHED and
 931  *   DISCONNECTED events. If the remote peer disconnected first, this
 932  *   reference could be gone.
 933  *
 934  * - A reference held by the svc_recv code that called this function
 935  *   as part of close processing.
 936  *
 937  * At a minimum two references should still be held.
 938  */
 939 static void svc_rdma_detach(struct svc_xprt *xprt)
 940 {
 941         struct svcxprt_rdma *rdma =
 942                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 943         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
 944
 945         /* Disconnect and flush posted WQE */
 946         rdma_disconnect(rdma->sc_cm_id);
 947
 948         /* Destroy the QP if present (not a listener) */
 949         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
 950                 ib_destroy_qp(rdma->sc_qp);
 951                 svc_xprt_put(xprt);
 952         }
 953
 954         /* Destroy the CM ID */
 955         rdma_destroy_id(rdma->sc_cm_id);
 956 }
 957
 958 static void svc_rdma_free(struct svc_xprt *xprt)
 959 {
 960         struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
 961         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
 962         /* We should only be called from kref_put */
 963         BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
 964         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
 965                 ib_destroy_cq(rdma->sc_sq_cq);
 966
 967         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
 968                 ib_destroy_cq(rdma->sc_rq_cq);
 969
 970         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
 971                 ib_dereg_mr(rdma->sc_phys_mr);
 972
 973         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
 974                 ib_dealloc_pd(rdma->sc_pd);
 975
 976         destroy_context_cache(rdma->sc_ctxt_head);
 977         kfree(rdma);
 978 }
 979
 980 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
 981 {
 982         struct svcxprt_rdma *rdma =
 983                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 984
 985         /*
 986          * If there are fewer SQ WR available than required to send a
 987          * simple response, return false.
 988          */
 989         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
 990                 return 0;
 991
 992         /*
 993          * ...or there are already waiters on the SQ,
 994          * return false.
 995          */
 996         if (waitqueue_active(&rdma->sc_send_wait))
 997                 return 0;
 998
 999         /* Otherwise return true. */
1000         return 1;
1001 }
1002
1003 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1004 {
1005         struct ib_send_wr *bad_wr;
1006         int ret;
1007
1008         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1009                 return -ENOTCONN;
1010
1011         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1012         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1013                 wr->opcode);
1014         /* If the SQ is full, wait until an SQ entry is available */
1015         while (1) {
1016                 spin_lock_bh(&xprt->sc_lock);
1017                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1018                         spin_unlock_bh(&xprt->sc_lock);
1019                         atomic_inc(&rdma_stat_sq_starve);
1020
1021                         /* See if we can opportunistically reap SQ WR to make room */
1022                         sq_cq_reap(xprt);
1023
1024                         /* Wait until SQ WR available if SQ still full */
1025                         wait_event(xprt->sc_send_wait,
1026                                    atomic_read(&xprt->sc_sq_count) <
1027                                    xprt->sc_sq_depth);
1028                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1029                                 return 0;
1030                         continue;
1031                 }
1032                 /* Bumped used SQ WR count and post */
1033                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1034                 if (!ret)
1035                         atomic_inc(&xprt->sc_sq_count);
1036                 else
1037                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1038                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1039                                ret, atomic_read(&xprt->sc_sq_count),
1040                                xprt->sc_sq_depth);
1041                 spin_unlock_bh(&xprt->sc_lock);
1042                 break;
1043         }
1044         return ret;
1045 }
1046
1047 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1048                         enum rpcrdma_errcode err)
1049 {
1050         struct ib_send_wr err_wr;
1051         struct ib_sge sge;
1052         struct page *p;
1053         struct svc_rdma_op_ctxt *ctxt;
1054         u32 *va;
1055         int length;
1056         int ret;
1057
1058         p = svc_rdma_get_page();
1059         va = page_address(p);
1060
1061         /* XDR encode error */
1062         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1063
1064         /* Prepare SGE for local address */
1065         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1066                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1067         sge.lkey = xprt->sc_phys_mr->lkey;
1068         sge.length = length;
1069
1070         ctxt = svc_rdma_get_context(xprt);
1071         ctxt->count = 1;
1072         ctxt->pages[0] = p;
1073
1074         /* Prepare SEND WR */
1075         memset(&err_wr, 0, sizeof err_wr);
1076         ctxt->wr_op = IB_WR_SEND;
1077         err_wr.wr_id = (unsigned long)ctxt;
1078         err_wr.sg_list = &sge;
1079         err_wr.num_sge = 1;
1080         err_wr.opcode = IB_WR_SEND;
1081         err_wr.send_flags = IB_SEND_SIGNALED;
1082
1083         /* Post It */
1084         ret = svc_rdma_send(xprt, &err_wr);
1085         if (ret) {
1086                 dprintk("svcrdma: Error posting send = %d\n", ret);
1087                 svc_rdma_put_context(ctxt, 1);
1088         }
1089
1090         return ret;
1091 }