net/rds/iw_rdma.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  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
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <linux/ratelimit.h>
  36
  37 #include "rds.h"
  38 #include "iw.h"
  39
  40
  41 /*
  42  * This is stored as mr->r_trans_private.
  43  */
  44 struct rds_iw_mr {
  45         struct rds_iw_device    *device;
  46         struct rds_iw_mr_pool   *pool;
  47         struct rdma_cm_id       *cm_id;
  48
  49         struct ib_mr    *mr;
  50
  51         struct rds_iw_mapping   mapping;
  52         unsigned char           remap_count;
  53 };
  54
  55 /*
  56  * Our own little MR pool
  57  */
  58 struct rds_iw_mr_pool {
  59         struct rds_iw_device    *device;                /* back ptr to the device that owns us */
  60
  61         struct mutex            flush_lock;             /* serialize fmr invalidate */
  62         struct work_struct      flush_worker;           /* flush worker */
  63
  64         spinlock_t              list_lock;              /* protect variables below */
  65         atomic_t                item_count;             /* total # of MRs */
  66         atomic_t                dirty_count;            /* # dirty of MRs */
  67         struct list_head        dirty_list;             /* dirty mappings */
  68         struct list_head        clean_list;             /* unused & unamapped MRs */
  69         atomic_t                free_pinned;            /* memory pinned by free MRs */
  70         unsigned long           max_message_size;       /* in pages */
  71         unsigned long           max_items;
  72         unsigned long           max_items_soft;
  73         unsigned long           max_free_pinned;
  74         int                     max_pages;
  75 };
  76
  77 static void rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
  78 static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
  79 static int rds_iw_init_reg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  80 static int rds_iw_map_reg(struct rds_iw_mr_pool *pool,
  81                           struct rds_iw_mr *ibmr,
  82                           struct scatterlist *sg, unsigned int nents);
  83 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  84 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
  85                         struct list_head *unmap_list,
  86                         struct list_head *kill_list,
  87                         int *unpinned);
  88 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  89
  90 static int rds_iw_get_device(struct sockaddr_in *src, struct sockaddr_in *dst,
  91                              struct rds_iw_device **rds_iwdev,
  92                              struct rdma_cm_id **cm_id)
  93 {
  94         struct rds_iw_device *iwdev;
  95         struct rds_iw_cm_id *i_cm_id;
  96
  97         *rds_iwdev = NULL;
  98         *cm_id = NULL;
  99
 100         list_for_each_entry(iwdev, &rds_iw_devices, list) {
 101                 spin_lock_irq(&iwdev->spinlock);
 102                 list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
 103                         struct sockaddr_in *src_addr, *dst_addr;
 104
 105                         src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
 106                         dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
 107
 108                         rdsdebug("local ipaddr = %x port %d, "
 109                                  "remote ipaddr = %x port %d"
 110                                  "..looking for %x port %d, "
 111                                  "remote ipaddr = %x port %d\n",
 112                                 src_addr->sin_addr.s_addr,
 113                                 src_addr->sin_port,
 114                                 dst_addr->sin_addr.s_addr,
 115                                 dst_addr->sin_port,
 116                                 src->sin_addr.s_addr,
 117                                 src->sin_port,
 118                                 dst->sin_addr.s_addr,
 119                                 dst->sin_port);
 120 #ifdef WORKING_TUPLE_DETECTION
 121                         if (src_addr->sin_addr.s_addr == src->sin_addr.s_addr &&
 122                             src_addr->sin_port == src->sin_port &&
 123                             dst_addr->sin_addr.s_addr == dst->sin_addr.s_addr &&
 124                             dst_addr->sin_port == dst->sin_port) {
 125 #else
 126                         /* FIXME - needs to compare the local and remote
 127                          * ipaddr/port tuple, but the ipaddr is the only
 128                          * available information in the rds_sock (as the rest are
 129                          * zero'ed.  It doesn't appear to be properly populated
 130                          * during connection setup...
 131                          */
 132                         if (src_addr->sin_addr.s_addr == src->sin_addr.s_addr) {
 133 #endif
 134                                 spin_unlock_irq(&iwdev->spinlock);
 135                                 *rds_iwdev = iwdev;
 136                                 *cm_id = i_cm_id->cm_id;
 137                                 return 0;
 138                         }
 139                 }
 140                 spin_unlock_irq(&iwdev->spinlock);
 141         }
 142
 143         return 1;
 144 }
 145
 146 static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 147 {
 148         struct rds_iw_cm_id *i_cm_id;
 149
 150         i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
 151         if (!i_cm_id)
 152                 return -ENOMEM;
 153
 154         i_cm_id->cm_id = cm_id;
 155
 156         spin_lock_irq(&rds_iwdev->spinlock);
 157         list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
 158         spin_unlock_irq(&rds_iwdev->spinlock);
 159
 160         return 0;
 161 }
 162
 163 static void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev,
 164                                 struct rdma_cm_id *cm_id)
 165 {
 166         struct rds_iw_cm_id *i_cm_id;
 167
 168         spin_lock_irq(&rds_iwdev->spinlock);
 169         list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
 170                 if (i_cm_id->cm_id == cm_id) {
 171                         list_del(&i_cm_id->list);
 172                         kfree(i_cm_id);
 173                         break;
 174                 }
 175         }
 176         spin_unlock_irq(&rds_iwdev->spinlock);
 177 }
 178
 179
 180 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 181 {
 182         struct sockaddr_in *src_addr, *dst_addr;
 183         struct rds_iw_device *rds_iwdev_old;
 184         struct rdma_cm_id *pcm_id;
 185         int rc;
 186
 187         src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
 188         dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
 189
 190         rc = rds_iw_get_device(src_addr, dst_addr, &rds_iwdev_old, &pcm_id);
 191         if (rc)
 192                 rds_iw_remove_cm_id(rds_iwdev, cm_id);
 193
 194         return rds_iw_add_cm_id(rds_iwdev, cm_id);
 195 }
 196
 197 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 198 {
 199         struct rds_iw_connection *ic = conn->c_transport_data;
 200
 201         /* conn was previously on the nodev_conns_list */
 202         spin_lock_irq(&iw_nodev_conns_lock);
 203         BUG_ON(list_empty(&iw_nodev_conns));
 204         BUG_ON(list_empty(&ic->iw_node));
 205         list_del(&ic->iw_node);
 206
 207         spin_lock(&rds_iwdev->spinlock);
 208         list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
 209         spin_unlock(&rds_iwdev->spinlock);
 210         spin_unlock_irq(&iw_nodev_conns_lock);
 211
 212         ic->rds_iwdev = rds_iwdev;
 213 }
 214
 215 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 216 {
 217         struct rds_iw_connection *ic = conn->c_transport_data;
 218
 219         /* place conn on nodev_conns_list */
 220         spin_lock(&iw_nodev_conns_lock);
 221
 222         spin_lock_irq(&rds_iwdev->spinlock);
 223         BUG_ON(list_empty(&ic->iw_node));
 224         list_del(&ic->iw_node);
 225         spin_unlock_irq(&rds_iwdev->spinlock);
 226
 227         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 228
 229         spin_unlock(&iw_nodev_conns_lock);
 230
 231         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
 232         ic->rds_iwdev = NULL;
 233 }
 234
 235 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
 236 {
 237         struct rds_iw_connection *ic, *_ic;
 238         LIST_HEAD(tmp_list);
 239
 240         /* avoid calling conn_destroy with irqs off */
 241         spin_lock_irq(list_lock);
 242         list_splice(list, &tmp_list);
 243         INIT_LIST_HEAD(list);
 244         spin_unlock_irq(list_lock);
 245
 246         list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
 247                 rds_conn_destroy(ic->conn);
 248 }
 249
 250 static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
 251                 struct scatterlist *list, unsigned int sg_len)
 252 {
 253         sg->list = list;
 254         sg->len = sg_len;
 255         sg->dma_len = 0;
 256         sg->dma_npages = 0;
 257         sg->bytes = 0;
 258 }
 259
 260 static int rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
 261                                   struct rds_iw_scatterlist *sg)
 262 {
 263         struct ib_device *dev = rds_iwdev->dev;
 264         int i, ret;
 265
 266         WARN_ON(sg->dma_len);
 267
 268         sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 269         if (unlikely(!sg->dma_len)) {
 270                 printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
 271                 return -EBUSY;
 272         }
 273
 274         sg->bytes = 0;
 275         sg->dma_npages = 0;
 276
 277         ret = -EINVAL;
 278         for (i = 0; i < sg->dma_len; ++i) {
 279                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
 280                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
 281                 u64 end_addr;
 282
 283                 sg->bytes += dma_len;
 284
 285                 end_addr = dma_addr + dma_len;
 286                 if (dma_addr & PAGE_MASK) {
 287                         if (i > 0)
 288                                 goto out_unmap;
 289                         dma_addr &= ~PAGE_MASK;
 290                 }
 291                 if (end_addr & PAGE_MASK) {
 292                         if (i < sg->dma_len - 1)
 293                                 goto out_unmap;
 294                         end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
 295                 }
 296
 297                 sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
 298         }
 299
 300         /* Now gather the dma addrs into one list */
 301         if (sg->dma_npages > fastreg_message_size)
 302                 goto out_unmap;
 303
 304
 305
 306         return 0;
 307
 308 out_unmap:
 309         ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 310         sg->dma_len = 0;
 311         return ret;
 312 }
 313
 314
 315 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
 316 {
 317         struct rds_iw_mr_pool *pool;
 318
 319         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 320         if (!pool) {
 321                 printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
 322                 return ERR_PTR(-ENOMEM);
 323         }
 324
 325         pool->device = rds_iwdev;
 326         INIT_LIST_HEAD(&pool->dirty_list);
 327         INIT_LIST_HEAD(&pool->clean_list);
 328         mutex_init(&pool->flush_lock);
 329         spin_lock_init(&pool->list_lock);
 330         INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
 331
 332         pool->max_message_size = fastreg_message_size;
 333         pool->max_items = fastreg_pool_size;
 334         pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
 335         pool->max_pages = fastreg_message_size;
 336
 337         /* We never allow more than max_items MRs to be allocated.
 338          * When we exceed more than max_items_soft, we start freeing
 339          * items more aggressively.
 340          * Make sure that max_items > max_items_soft > max_items / 2
 341          */
 342         pool->max_items_soft = pool->max_items * 3 / 4;
 343
 344         return pool;
 345 }
 346
 347 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
 348 {
 349         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 350
 351         iinfo->rdma_mr_max = pool->max_items;
 352         iinfo->rdma_mr_size = pool->max_pages;
 353 }
 354
 355 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
 356 {
 357         flush_workqueue(rds_wq);
 358         rds_iw_flush_mr_pool(pool, 1);
 359         BUG_ON(atomic_read(&pool->item_count));
 360         BUG_ON(atomic_read(&pool->free_pinned));
 361         kfree(pool);
 362 }
 363
 364 static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
 365 {
 366         struct rds_iw_mr *ibmr = NULL;
 367         unsigned long flags;
 368
 369         spin_lock_irqsave(&pool->list_lock, flags);
 370         if (!list_empty(&pool->clean_list)) {
 371                 ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
 372                 list_del_init(&ibmr->mapping.m_list);
 373         }
 374         spin_unlock_irqrestore(&pool->list_lock, flags);
 375
 376         return ibmr;
 377 }
 378
 379 static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
 380 {
 381         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 382         struct rds_iw_mr *ibmr = NULL;
 383         int err = 0, iter = 0;
 384
 385         while (1) {
 386                 ibmr = rds_iw_reuse_fmr(pool);
 387                 if (ibmr)
 388                         return ibmr;
 389
 390                 /* No clean MRs - now we have the choice of either
 391                  * allocating a fresh MR up to the limit imposed by the
 392                  * driver, or flush any dirty unused MRs.
 393                  * We try to avoid stalling in the send path if possible,
 394                  * so we allocate as long as we're allowed to.
 395                  *
 396                  * We're fussy with enforcing the FMR limit, though. If the driver
 397                  * tells us we can't use more than N fmrs, we shouldn't start
 398                  * arguing with it */
 399                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 400                         break;
 401
 402                 atomic_dec(&pool->item_count);
 403
 404                 if (++iter > 2) {
 405                         rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
 406                         return ERR_PTR(-EAGAIN);
 407                 }
 408
 409                 /* We do have some empty MRs. Flush them out. */
 410                 rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
 411                 rds_iw_flush_mr_pool(pool, 0);
 412         }
 413
 414         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
 415         if (!ibmr) {
 416                 err = -ENOMEM;
 417                 goto out_no_cigar;
 418         }
 419
 420         spin_lock_init(&ibmr->mapping.m_lock);
 421         INIT_LIST_HEAD(&ibmr->mapping.m_list);
 422         ibmr->mapping.m_mr = ibmr;
 423
 424         err = rds_iw_init_reg(pool, ibmr);
 425         if (err)
 426                 goto out_no_cigar;
 427
 428         rds_iw_stats_inc(s_iw_rdma_mr_alloc);
 429         return ibmr;
 430
 431 out_no_cigar:
 432         if (ibmr) {
 433                 rds_iw_destroy_fastreg(pool, ibmr);
 434                 kfree(ibmr);
 435         }
 436         atomic_dec(&pool->item_count);
 437         return ERR_PTR(err);
 438 }
 439
 440 void rds_iw_sync_mr(void *trans_private, int direction)
 441 {
 442         struct rds_iw_mr *ibmr = trans_private;
 443         struct rds_iw_device *rds_iwdev = ibmr->device;
 444
 445         switch (direction) {
 446         case DMA_FROM_DEVICE:
 447                 ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 448                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 449                 break;
 450         case DMA_TO_DEVICE:
 451                 ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 452                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 453                 break;
 454         }
 455 }
 456
 457 /*
 458  * Flush our pool of MRs.
 459  * At a minimum, all currently unused MRs are unmapped.
 460  * If the number of MRs allocated exceeds the limit, we also try
 461  * to free as many MRs as needed to get back to this limit.
 462  */
 463 static void rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
 464 {
 465         struct rds_iw_mr *ibmr, *next;
 466         LIST_HEAD(unmap_list);
 467         LIST_HEAD(kill_list);
 468         unsigned long flags;
 469         unsigned int nfreed = 0, ncleaned = 0, unpinned = 0;
 470
 471         rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
 472
 473         mutex_lock(&pool->flush_lock);
 474
 475         spin_lock_irqsave(&pool->list_lock, flags);
 476         /* Get the list of all mappings to be destroyed */
 477         list_splice_init(&pool->dirty_list, &unmap_list);
 478         if (free_all)
 479                 list_splice_init(&pool->clean_list, &kill_list);
 480         spin_unlock_irqrestore(&pool->list_lock, flags);
 481
 482         /* Batched invalidate of dirty MRs.
 483          * For FMR based MRs, the mappings on the unmap list are
 484          * actually members of an ibmr (ibmr->mapping). They either
 485          * migrate to the kill_list, or have been cleaned and should be
 486          * moved to the clean_list.
 487          * For fastregs, they will be dynamically allocated, and
 488          * will be destroyed by the unmap function.
 489          */
 490         if (!list_empty(&unmap_list)) {
 491                 ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list,
 492                                                      &kill_list, &unpinned);
 493                 /* If we've been asked to destroy all MRs, move those
 494                  * that were simply cleaned to the kill list */
 495                 if (free_all)
 496                         list_splice_init(&unmap_list, &kill_list);
 497         }
 498
 499         /* Destroy any MRs that are past their best before date */
 500         list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
 501                 rds_iw_stats_inc(s_iw_rdma_mr_free);
 502                 list_del(&ibmr->mapping.m_list);
 503                 rds_iw_destroy_fastreg(pool, ibmr);
 504                 kfree(ibmr);
 505                 nfreed++;
 506         }
 507
 508         /* Anything that remains are laundered ibmrs, which we can add
 509          * back to the clean list. */
 510         if (!list_empty(&unmap_list)) {
 511                 spin_lock_irqsave(&pool->list_lock, flags);
 512                 list_splice(&unmap_list, &pool->clean_list);
 513                 spin_unlock_irqrestore(&pool->list_lock, flags);
 514         }
 515
 516         atomic_sub(unpinned, &pool->free_pinned);
 517         atomic_sub(ncleaned, &pool->dirty_count);
 518         atomic_sub(nfreed, &pool->item_count);
 519
 520         mutex_unlock(&pool->flush_lock);
 521 }
 522
 523 static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
 524 {
 525         struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
 526
 527         rds_iw_flush_mr_pool(pool, 0);
 528 }
 529
 530 void rds_iw_free_mr(void *trans_private, int invalidate)
 531 {
 532         struct rds_iw_mr *ibmr = trans_private;
 533         struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
 534
 535         rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
 536         if (!pool)
 537                 return;
 538
 539         /* Return it to the pool's free list */
 540         rds_iw_free_fastreg(pool, ibmr);
 541
 542         /* If we've pinned too many pages, request a flush */
 543         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 544             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 545                 queue_work(rds_wq, &pool->flush_worker);
 546
 547         if (invalidate) {
 548                 if (likely(!in_interrupt())) {
 549                         rds_iw_flush_mr_pool(pool, 0);
 550                 } else {
 551                         /* We get here if the user created a MR marked
 552                          * as use_once and invalidate at the same time. */
 553                         queue_work(rds_wq, &pool->flush_worker);
 554                 }
 555         }
 556 }
 557
 558 void rds_iw_flush_mrs(void)
 559 {
 560         struct rds_iw_device *rds_iwdev;
 561
 562         list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
 563                 struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 564
 565                 if (pool)
 566                         rds_iw_flush_mr_pool(pool, 0);
 567         }
 568 }
 569
 570 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
 571                     struct rds_sock *rs, u32 *key_ret)
 572 {
 573         struct rds_iw_device *rds_iwdev;
 574         struct rds_iw_mr *ibmr = NULL;
 575         struct rdma_cm_id *cm_id;
 576         struct sockaddr_in src = {
 577                 .sin_addr.s_addr = rs->rs_bound_addr,
 578                 .sin_port = rs->rs_bound_port,
 579         };
 580         struct sockaddr_in dst = {
 581                 .sin_addr.s_addr = rs->rs_conn_addr,
 582                 .sin_port = rs->rs_conn_port,
 583         };
 584         int ret;
 585
 586         ret = rds_iw_get_device(&src, &dst, &rds_iwdev, &cm_id);
 587         if (ret || !cm_id) {
 588                 ret = -ENODEV;
 589                 goto out;
 590         }
 591
 592         if (!rds_iwdev->mr_pool) {
 593                 ret = -ENODEV;
 594                 goto out;
 595         }
 596
 597         ibmr = rds_iw_alloc_mr(rds_iwdev);
 598         if (IS_ERR(ibmr))
 599                 return ibmr;
 600
 601         ibmr->cm_id = cm_id;
 602         ibmr->device = rds_iwdev;
 603
 604         ret = rds_iw_map_reg(rds_iwdev->mr_pool, ibmr, sg, nents);
 605         if (ret == 0)
 606                 *key_ret = ibmr->mr->rkey;
 607         else
 608                 printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
 609
 610 out:
 611         if (ret) {
 612                 if (ibmr)
 613                         rds_iw_free_mr(ibmr, 0);
 614                 ibmr = ERR_PTR(ret);
 615         }
 616         return ibmr;
 617 }
 618
 619 /*
 620  * iWARP reg handling
 621  *
 622  * The life cycle of a fastreg registration is a bit different from
 623  * FMRs.
 624  * The idea behind fastreg is to have one MR, to which we bind different
 625  * mappings over time. To avoid stalling on the expensive map and invalidate
 626  * operations, these operations are pipelined on the same send queue on
 627  * which we want to send the message containing the r_key.
 628  *
 629  * This creates a bit of a problem for us, as we do not have the destination
 630  * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
 631  * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
 632  * will try to queue a LOCAL_INV (if needed) and a REG_MR work request
 633  * before queuing the SEND. When completions for these arrive, they are
 634  * dispatched to the MR has a bit set showing that RDMa can be performed.
 635  *
 636  * There is another interesting aspect that's related to invalidation.
 637  * The application can request that a mapping is invalidated in FREE_MR.
 638  * The expectation there is that this invalidation step includes ALL
 639  * PREVIOUSLY FREED MRs.
 640  */
 641 static int rds_iw_init_reg(struct rds_iw_mr_pool *pool,
 642                            struct rds_iw_mr *ibmr)
 643 {
 644         struct rds_iw_device *rds_iwdev = pool->device;
 645         struct ib_mr *mr;
 646         int err;
 647
 648         mr = ib_alloc_mr(rds_iwdev->pd, IB_MR_TYPE_MEM_REG,
 649                          pool->max_message_size);
 650         if (IS_ERR(mr)) {
 651                 err = PTR_ERR(mr);
 652
 653                 printk(KERN_WARNING "RDS/IW: ib_alloc_mr failed (err=%d)\n", err);
 654                 return err;
 655         }
 656
 657         ibmr->mr = mr;
 658         return 0;
 659 }
 660
 661 static int rds_iw_rdma_reg_mr(struct rds_iw_mapping *mapping)
 662 {
 663         struct rds_iw_mr *ibmr = mapping->m_mr;
 664         struct rds_iw_scatterlist *m_sg = &mapping->m_sg;
 665         struct ib_reg_wr reg_wr;
 666         struct ib_send_wr *failed_wr;
 667         int ret, n;
 668
 669         n = ib_map_mr_sg_zbva(ibmr->mr, m_sg->list, m_sg->len, PAGE_SIZE);
 670         if (unlikely(n != m_sg->len))
 671                 return n < 0 ? n : -EINVAL;
 672
 673         reg_wr.wr.next = NULL;
 674         reg_wr.wr.opcode = IB_WR_REG_MR;
 675         reg_wr.wr.wr_id = RDS_IW_REG_WR_ID;
 676         reg_wr.wr.num_sge = 0;
 677         reg_wr.mr = ibmr->mr;
 678         reg_wr.key = mapping->m_rkey;
 679         reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 680                         IB_ACCESS_REMOTE_READ |
 681                         IB_ACCESS_REMOTE_WRITE;
 682
 683         /*
 684          * Perform a WR for the reg_mr. Each individual page
 685          * in the sg list is added to the fast reg page list and placed
 686          * inside the reg_mr WR.  The key used is a rolling 8bit
 687          * counter, which should guarantee uniqueness.
 688          */
 689         ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
 690         mapping->m_rkey = ibmr->mr->rkey;
 691
 692         failed_wr = &reg_wr.wr;
 693         ret = ib_post_send(ibmr->cm_id->qp, &reg_wr.wr, &failed_wr);
 694         BUG_ON(failed_wr != &reg_wr.wr);
 695         if (ret)
 696                 printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 697                         __func__, __LINE__, ret);
 698         return ret;
 699 }
 700
 701 static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
 702 {
 703         struct ib_send_wr s_wr, *failed_wr;
 704         int ret = 0;
 705
 706         if (!ibmr->cm_id->qp || !ibmr->mr)
 707                 goto out;
 708
 709         memset(&s_wr, 0, sizeof(s_wr));
 710         s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
 711         s_wr.opcode = IB_WR_LOCAL_INV;
 712         s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
 713         s_wr.send_flags = IB_SEND_SIGNALED;
 714
 715         failed_wr = &s_wr;
 716         ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
 717         if (ret) {
 718                 printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 719                         __func__, __LINE__, ret);
 720                 goto out;
 721         }
 722 out:
 723         return ret;
 724 }
 725
 726 static int rds_iw_map_reg(struct rds_iw_mr_pool *pool,
 727                           struct rds_iw_mr *ibmr,
 728                           struct scatterlist *sg,
 729                           unsigned int sg_len)
 730 {
 731         struct rds_iw_device *rds_iwdev = pool->device;
 732         struct rds_iw_mapping *mapping = &ibmr->mapping;
 733         u64 *dma_pages;
 734         int ret = 0;
 735
 736         rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
 737
 738         ret = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
 739         if (ret) {
 740                 dma_pages = NULL;
 741                 goto out;
 742         }
 743
 744         if (mapping->m_sg.dma_len > pool->max_message_size) {
 745                 ret = -EMSGSIZE;
 746                 goto out;
 747         }
 748
 749         ret = rds_iw_rdma_reg_mr(mapping);
 750         if (ret)
 751                 goto out;
 752
 753         rds_iw_stats_inc(s_iw_rdma_mr_used);
 754
 755 out:
 756         kfree(dma_pages);
 757
 758         return ret;
 759 }
 760
 761 /*
 762  * "Free" a fastreg MR.
 763  */
 764 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
 765                 struct rds_iw_mr *ibmr)
 766 {
 767         unsigned long flags;
 768         int ret;
 769
 770         if (!ibmr->mapping.m_sg.dma_len)
 771                 return;
 772
 773         ret = rds_iw_rdma_fastreg_inv(ibmr);
 774         if (ret)
 775                 return;
 776
 777         /* Try to post the LOCAL_INV WR to the queue. */
 778         spin_lock_irqsave(&pool->list_lock, flags);
 779
 780         list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
 781         atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
 782         atomic_inc(&pool->dirty_count);
 783
 784         spin_unlock_irqrestore(&pool->list_lock, flags);
 785 }
 786
 787 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
 788                                 struct list_head *unmap_list,
 789                                 struct list_head *kill_list,
 790                                 int *unpinned)
 791 {
 792         struct rds_iw_mapping *mapping, *next;
 793         unsigned int ncleaned = 0;
 794         LIST_HEAD(laundered);
 795
 796         /* Batched invalidation of fastreg MRs.
 797          * Why do we do it this way, even though we could pipeline unmap
 798          * and remap? The reason is the application semantics - when the
 799          * application requests an invalidation of MRs, it expects all
 800          * previously released R_Keys to become invalid.
 801          *
 802          * If we implement MR reuse naively, we risk memory corruption
 803          * (this has actually been observed). So the default behavior
 804          * requires that a MR goes through an explicit unmap operation before
 805          * we can reuse it again.
 806          *
 807          * We could probably improve on this a little, by allowing immediate
 808          * reuse of a MR on the same socket (eg you could add small
 809          * cache of unused MRs to strct rds_socket - GET_MR could grab one
 810          * of these without requiring an explicit invalidate).
 811          */
 812         while (!list_empty(unmap_list)) {
 813                 unsigned long flags;
 814
 815                 spin_lock_irqsave(&pool->list_lock, flags);
 816                 list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
 817                         *unpinned += mapping->m_sg.len;
 818                         list_move(&mapping->m_list, &laundered);
 819                         ncleaned++;
 820                 }
 821                 spin_unlock_irqrestore(&pool->list_lock, flags);
 822         }
 823
 824         /* Move all laundered mappings back to the unmap list.
 825          * We do not kill any WRs right now - it doesn't seem the
 826          * fastreg API has a max_remap limit. */
 827         list_splice_init(&laundered, unmap_list);
 828
 829         return ncleaned;
 830 }
 831
 832 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
 833                 struct rds_iw_mr *ibmr)
 834 {
 835         if (ibmr->mr)
 836                 ib_dereg_mr(ibmr->mr);
 837 }