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         struct ib_fast_reg_page_list *page_list;
  51
  52         struct rds_iw_mapping   mapping;
  53         unsigned char           remap_count;
  54 };
  55
  56 /*
  57  * Our own little MR pool
  58  */
  59 struct rds_iw_mr_pool {
  60         struct rds_iw_device    *device;                /* back ptr to the device that owns us */
  61
  62         struct mutex            flush_lock;             /* serialize fmr invalidate */
  63         struct work_struct      flush_worker;           /* flush worker */
  64
  65         spinlock_t              list_lock;              /* protect variables below */
  66         atomic_t                item_count;             /* total # of MRs */
  67         atomic_t                dirty_count;            /* # dirty of MRs */
  68         struct list_head        dirty_list;             /* dirty mappings */
  69         struct list_head        clean_list;             /* unused & unamapped MRs */
  70         atomic_t                free_pinned;            /* memory pinned by free MRs */
  71         unsigned long           max_message_size;       /* in pages */
  72         unsigned long           max_items;
  73         unsigned long           max_items_soft;
  74         unsigned long           max_free_pinned;
  75         int                     max_pages;
  76 };
  77
  78 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all);
  79 static void rds_iw_mr_pool_flush_worker(struct work_struct *work);
  80 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  81 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
  82                           struct rds_iw_mr *ibmr,
  83                           struct scatterlist *sg, unsigned int nents);
  84 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  85 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
  86                         struct list_head *unmap_list,
  87                         struct list_head *kill_list);
  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 rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
  91 {
  92         struct rds_iw_device *iwdev;
  93         struct rds_iw_cm_id *i_cm_id;
  94
  95         *rds_iwdev = NULL;
  96         *cm_id = NULL;
  97
  98         list_for_each_entry(iwdev, &rds_iw_devices, list) {
  99                 spin_lock_irq(&iwdev->spinlock);
 100                 list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
 101                         struct sockaddr_in *src_addr, *dst_addr;
 102
 103                         src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
 104                         dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
 105
 106                         rdsdebug("local ipaddr = %x port %d, "
 107                                  "remote ipaddr = %x port %d"
 108                                  "..looking for %x port %d, "
 109                                  "remote ipaddr = %x port %d\n",
 110                                 src_addr->sin_addr.s_addr,
 111                                 src_addr->sin_port,
 112                                 dst_addr->sin_addr.s_addr,
 113                                 dst_addr->sin_port,
 114                                 rs->rs_bound_addr,
 115                                 rs->rs_bound_port,
 116                                 rs->rs_conn_addr,
 117                                 rs->rs_conn_port);
 118 #ifdef WORKING_TUPLE_DETECTION
 119                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
 120                             src_addr->sin_port == rs->rs_bound_port &&
 121                             dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
 122                             dst_addr->sin_port == rs->rs_conn_port) {
 123 #else
 124                         /* FIXME - needs to compare the local and remote
 125                          * ipaddr/port tuple, but the ipaddr is the only
 126                          * available information in the rds_sock (as the rest are
 127                          * zero'ed.  It doesn't appear to be properly populated
 128                          * during connection setup...
 129                          */
 130                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
 131 #endif
 132                                 spin_unlock_irq(&iwdev->spinlock);
 133                                 *rds_iwdev = iwdev;
 134                                 *cm_id = i_cm_id->cm_id;
 135                                 return 0;
 136                         }
 137                 }
 138                 spin_unlock_irq(&iwdev->spinlock);
 139         }
 140
 141         return 1;
 142 }
 143
 144 static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 145 {
 146         struct rds_iw_cm_id *i_cm_id;
 147
 148         i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
 149         if (!i_cm_id)
 150                 return -ENOMEM;
 151
 152         i_cm_id->cm_id = cm_id;
 153
 154         spin_lock_irq(&rds_iwdev->spinlock);
 155         list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
 156         spin_unlock_irq(&rds_iwdev->spinlock);
 157
 158         return 0;
 159 }
 160
 161 static void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev,
 162                                 struct rdma_cm_id *cm_id)
 163 {
 164         struct rds_iw_cm_id *i_cm_id;
 165
 166         spin_lock_irq(&rds_iwdev->spinlock);
 167         list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
 168                 if (i_cm_id->cm_id == cm_id) {
 169                         list_del(&i_cm_id->list);
 170                         kfree(i_cm_id);
 171                         break;
 172                 }
 173         }
 174         spin_unlock_irq(&rds_iwdev->spinlock);
 175 }
 176
 177
 178 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 179 {
 180         struct sockaddr_in *src_addr, *dst_addr;
 181         struct rds_iw_device *rds_iwdev_old;
 182         struct rds_sock rs;
 183         struct rdma_cm_id *pcm_id;
 184         int rc;
 185
 186         src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
 187         dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
 188
 189         rs.rs_bound_addr = src_addr->sin_addr.s_addr;
 190         rs.rs_bound_port = src_addr->sin_port;
 191         rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
 192         rs.rs_conn_port = dst_addr->sin_port;
 193
 194         rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
 195         if (rc)
 196                 rds_iw_remove_cm_id(rds_iwdev, cm_id);
 197
 198         return rds_iw_add_cm_id(rds_iwdev, cm_id);
 199 }
 200
 201 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 202 {
 203         struct rds_iw_connection *ic = conn->c_transport_data;
 204
 205         /* conn was previously on the nodev_conns_list */
 206         spin_lock_irq(&iw_nodev_conns_lock);
 207         BUG_ON(list_empty(&iw_nodev_conns));
 208         BUG_ON(list_empty(&ic->iw_node));
 209         list_del(&ic->iw_node);
 210
 211         spin_lock(&rds_iwdev->spinlock);
 212         list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
 213         spin_unlock(&rds_iwdev->spinlock);
 214         spin_unlock_irq(&iw_nodev_conns_lock);
 215
 216         ic->rds_iwdev = rds_iwdev;
 217 }
 218
 219 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 220 {
 221         struct rds_iw_connection *ic = conn->c_transport_data;
 222
 223         /* place conn on nodev_conns_list */
 224         spin_lock(&iw_nodev_conns_lock);
 225
 226         spin_lock_irq(&rds_iwdev->spinlock);
 227         BUG_ON(list_empty(&ic->iw_node));
 228         list_del(&ic->iw_node);
 229         spin_unlock_irq(&rds_iwdev->spinlock);
 230
 231         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 232
 233         spin_unlock(&iw_nodev_conns_lock);
 234
 235         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
 236         ic->rds_iwdev = NULL;
 237 }
 238
 239 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
 240 {
 241         struct rds_iw_connection *ic, *_ic;
 242         LIST_HEAD(tmp_list);
 243
 244         /* avoid calling conn_destroy with irqs off */
 245         spin_lock_irq(list_lock);
 246         list_splice(list, &tmp_list);
 247         INIT_LIST_HEAD(list);
 248         spin_unlock_irq(list_lock);
 249
 250         list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
 251                 rds_conn_destroy(ic->conn);
 252 }
 253
 254 static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
 255                 struct scatterlist *list, unsigned int sg_len)
 256 {
 257         sg->list = list;
 258         sg->len = sg_len;
 259         sg->dma_len = 0;
 260         sg->dma_npages = 0;
 261         sg->bytes = 0;
 262 }
 263
 264 static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
 265                         struct rds_iw_scatterlist *sg)
 266 {
 267         struct ib_device *dev = rds_iwdev->dev;
 268         u64 *dma_pages = NULL;
 269         int i, j, ret;
 270
 271         WARN_ON(sg->dma_len);
 272
 273         sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 274         if (unlikely(!sg->dma_len)) {
 275                 printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
 276                 return ERR_PTR(-EBUSY);
 277         }
 278
 279         sg->bytes = 0;
 280         sg->dma_npages = 0;
 281
 282         ret = -EINVAL;
 283         for (i = 0; i < sg->dma_len; ++i) {
 284                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
 285                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
 286                 u64 end_addr;
 287
 288                 sg->bytes += dma_len;
 289
 290                 end_addr = dma_addr + dma_len;
 291                 if (dma_addr & PAGE_MASK) {
 292                         if (i > 0)
 293                                 goto out_unmap;
 294                         dma_addr &= ~PAGE_MASK;
 295                 }
 296                 if (end_addr & PAGE_MASK) {
 297                         if (i < sg->dma_len - 1)
 298                                 goto out_unmap;
 299                         end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
 300                 }
 301
 302                 sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
 303         }
 304
 305         /* Now gather the dma addrs into one list */
 306         if (sg->dma_npages > fastreg_message_size)
 307                 goto out_unmap;
 308
 309         dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
 310         if (!dma_pages) {
 311                 ret = -ENOMEM;
 312                 goto out_unmap;
 313         }
 314
 315         for (i = j = 0; i < sg->dma_len; ++i) {
 316                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
 317                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
 318                 u64 end_addr;
 319
 320                 end_addr = dma_addr + dma_len;
 321                 dma_addr &= ~PAGE_MASK;
 322                 for (; dma_addr < end_addr; dma_addr += PAGE_SIZE)
 323                         dma_pages[j++] = dma_addr;
 324                 BUG_ON(j > sg->dma_npages);
 325         }
 326
 327         return dma_pages;
 328
 329 out_unmap:
 330         ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 331         sg->dma_len = 0;
 332         kfree(dma_pages);
 333         return ERR_PTR(ret);
 334 }
 335
 336
 337 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
 338 {
 339         struct rds_iw_mr_pool *pool;
 340
 341         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 342         if (!pool) {
 343                 printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
 344                 return ERR_PTR(-ENOMEM);
 345         }
 346
 347         pool->device = rds_iwdev;
 348         INIT_LIST_HEAD(&pool->dirty_list);
 349         INIT_LIST_HEAD(&pool->clean_list);
 350         mutex_init(&pool->flush_lock);
 351         spin_lock_init(&pool->list_lock);
 352         INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
 353
 354         pool->max_message_size = fastreg_message_size;
 355         pool->max_items = fastreg_pool_size;
 356         pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
 357         pool->max_pages = fastreg_message_size;
 358
 359         /* We never allow more than max_items MRs to be allocated.
 360          * When we exceed more than max_items_soft, we start freeing
 361          * items more aggressively.
 362          * Make sure that max_items > max_items_soft > max_items / 2
 363          */
 364         pool->max_items_soft = pool->max_items * 3 / 4;
 365
 366         return pool;
 367 }
 368
 369 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
 370 {
 371         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 372
 373         iinfo->rdma_mr_max = pool->max_items;
 374         iinfo->rdma_mr_size = pool->max_pages;
 375 }
 376
 377 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
 378 {
 379         flush_workqueue(rds_wq);
 380         rds_iw_flush_mr_pool(pool, 1);
 381         BUG_ON(atomic_read(&pool->item_count));
 382         BUG_ON(atomic_read(&pool->free_pinned));
 383         kfree(pool);
 384 }
 385
 386 static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
 387 {
 388         struct rds_iw_mr *ibmr = NULL;
 389         unsigned long flags;
 390
 391         spin_lock_irqsave(&pool->list_lock, flags);
 392         if (!list_empty(&pool->clean_list)) {
 393                 ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
 394                 list_del_init(&ibmr->mapping.m_list);
 395         }
 396         spin_unlock_irqrestore(&pool->list_lock, flags);
 397
 398         return ibmr;
 399 }
 400
 401 static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
 402 {
 403         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 404         struct rds_iw_mr *ibmr = NULL;
 405         int err = 0, iter = 0;
 406
 407         while (1) {
 408                 ibmr = rds_iw_reuse_fmr(pool);
 409                 if (ibmr)
 410                         return ibmr;
 411
 412                 /* No clean MRs - now we have the choice of either
 413                  * allocating a fresh MR up to the limit imposed by the
 414                  * driver, or flush any dirty unused MRs.
 415                  * We try to avoid stalling in the send path if possible,
 416                  * so we allocate as long as we're allowed to.
 417                  *
 418                  * We're fussy with enforcing the FMR limit, though. If the driver
 419                  * tells us we can't use more than N fmrs, we shouldn't start
 420                  * arguing with it */
 421                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 422                         break;
 423
 424                 atomic_dec(&pool->item_count);
 425
 426                 if (++iter > 2) {
 427                         rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
 428                         return ERR_PTR(-EAGAIN);
 429                 }
 430
 431                 /* We do have some empty MRs. Flush them out. */
 432                 rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
 433                 rds_iw_flush_mr_pool(pool, 0);
 434         }
 435
 436         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
 437         if (!ibmr) {
 438                 err = -ENOMEM;
 439                 goto out_no_cigar;
 440         }
 441
 442         spin_lock_init(&ibmr->mapping.m_lock);
 443         INIT_LIST_HEAD(&ibmr->mapping.m_list);
 444         ibmr->mapping.m_mr = ibmr;
 445
 446         err = rds_iw_init_fastreg(pool, ibmr);
 447         if (err)
 448                 goto out_no_cigar;
 449
 450         rds_iw_stats_inc(s_iw_rdma_mr_alloc);
 451         return ibmr;
 452
 453 out_no_cigar:
 454         if (ibmr) {
 455                 rds_iw_destroy_fastreg(pool, ibmr);
 456                 kfree(ibmr);
 457         }
 458         atomic_dec(&pool->item_count);
 459         return ERR_PTR(err);
 460 }
 461
 462 void rds_iw_sync_mr(void *trans_private, int direction)
 463 {
 464         struct rds_iw_mr *ibmr = trans_private;
 465         struct rds_iw_device *rds_iwdev = ibmr->device;
 466
 467         switch (direction) {
 468         case DMA_FROM_DEVICE:
 469                 ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 470                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 471                 break;
 472         case DMA_TO_DEVICE:
 473                 ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 474                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 475                 break;
 476         }
 477 }
 478
 479 static inline unsigned int rds_iw_flush_goal(struct rds_iw_mr_pool *pool, int free_all)
 480 {
 481         unsigned int item_count;
 482
 483         item_count = atomic_read(&pool->item_count);
 484         if (free_all)
 485                 return item_count;
 486
 487         return 0;
 488 }
 489
 490 /*
 491  * Flush our pool of MRs.
 492  * At a minimum, all currently unused MRs are unmapped.
 493  * If the number of MRs allocated exceeds the limit, we also try
 494  * to free as many MRs as needed to get back to this limit.
 495  */
 496 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
 497 {
 498         struct rds_iw_mr *ibmr, *next;
 499         LIST_HEAD(unmap_list);
 500         LIST_HEAD(kill_list);
 501         unsigned long flags;
 502         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 503         int ret = 0;
 504
 505         rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
 506
 507         mutex_lock(&pool->flush_lock);
 508
 509         spin_lock_irqsave(&pool->list_lock, flags);
 510         /* Get the list of all mappings to be destroyed */
 511         list_splice_init(&pool->dirty_list, &unmap_list);
 512         if (free_all)
 513                 list_splice_init(&pool->clean_list, &kill_list);
 514         spin_unlock_irqrestore(&pool->list_lock, flags);
 515
 516         free_goal = rds_iw_flush_goal(pool, free_all);
 517
 518         /* Batched invalidate of dirty MRs.
 519          * For FMR based MRs, the mappings on the unmap list are
 520          * actually members of an ibmr (ibmr->mapping). They either
 521          * migrate to the kill_list, or have been cleaned and should be
 522          * moved to the clean_list.
 523          * For fastregs, they will be dynamically allocated, and
 524          * will be destroyed by the unmap function.
 525          */
 526         if (!list_empty(&unmap_list)) {
 527                 ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list, &kill_list);
 528                 /* If we've been asked to destroy all MRs, move those
 529                  * that were simply cleaned to the kill list */
 530                 if (free_all)
 531                         list_splice_init(&unmap_list, &kill_list);
 532         }
 533
 534         /* Destroy any MRs that are past their best before date */
 535         list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
 536                 rds_iw_stats_inc(s_iw_rdma_mr_free);
 537                 list_del(&ibmr->mapping.m_list);
 538                 rds_iw_destroy_fastreg(pool, ibmr);
 539                 kfree(ibmr);
 540                 nfreed++;
 541         }
 542
 543         /* Anything that remains are laundered ibmrs, which we can add
 544          * back to the clean list. */
 545         if (!list_empty(&unmap_list)) {
 546                 spin_lock_irqsave(&pool->list_lock, flags);
 547                 list_splice(&unmap_list, &pool->clean_list);
 548                 spin_unlock_irqrestore(&pool->list_lock, flags);
 549         }
 550
 551         atomic_sub(ncleaned, &pool->dirty_count);
 552         atomic_sub(nfreed, &pool->item_count);
 553
 554         mutex_unlock(&pool->flush_lock);
 555         return ret;
 556 }
 557
 558 static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
 559 {
 560         struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
 561
 562         rds_iw_flush_mr_pool(pool, 0);
 563 }
 564
 565 void rds_iw_free_mr(void *trans_private, int invalidate)
 566 {
 567         struct rds_iw_mr *ibmr = trans_private;
 568         struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
 569
 570         rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
 571         if (!pool)
 572                 return;
 573
 574         /* Return it to the pool's free list */
 575         rds_iw_free_fastreg(pool, ibmr);
 576
 577         /* If we've pinned too many pages, request a flush */
 578         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 579             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 580                 queue_work(rds_wq, &pool->flush_worker);
 581
 582         if (invalidate) {
 583                 if (likely(!in_interrupt())) {
 584                         rds_iw_flush_mr_pool(pool, 0);
 585                 } else {
 586                         /* We get here if the user created a MR marked
 587                          * as use_once and invalidate at the same time. */
 588                         queue_work(rds_wq, &pool->flush_worker);
 589                 }
 590         }
 591 }
 592
 593 void rds_iw_flush_mrs(void)
 594 {
 595         struct rds_iw_device *rds_iwdev;
 596
 597         list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
 598                 struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 599
 600                 if (pool)
 601                         rds_iw_flush_mr_pool(pool, 0);
 602         }
 603 }
 604
 605 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
 606                     struct rds_sock *rs, u32 *key_ret)
 607 {
 608         struct rds_iw_device *rds_iwdev;
 609         struct rds_iw_mr *ibmr = NULL;
 610         struct rdma_cm_id *cm_id;
 611         int ret;
 612
 613         ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
 614         if (ret || !cm_id) {
 615                 ret = -ENODEV;
 616                 goto out;
 617         }
 618
 619         if (!rds_iwdev->mr_pool) {
 620                 ret = -ENODEV;
 621                 goto out;
 622         }
 623
 624         ibmr = rds_iw_alloc_mr(rds_iwdev);
 625         if (IS_ERR(ibmr))
 626                 return ibmr;
 627
 628         ibmr->cm_id = cm_id;
 629         ibmr->device = rds_iwdev;
 630
 631         ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
 632         if (ret == 0)
 633                 *key_ret = ibmr->mr->rkey;
 634         else
 635                 printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
 636
 637 out:
 638         if (ret) {
 639                 if (ibmr)
 640                         rds_iw_free_mr(ibmr, 0);
 641                 ibmr = ERR_PTR(ret);
 642         }
 643         return ibmr;
 644 }
 645
 646 /*
 647  * iWARP fastreg handling
 648  *
 649  * The life cycle of a fastreg registration is a bit different from
 650  * FMRs.
 651  * The idea behind fastreg is to have one MR, to which we bind different
 652  * mappings over time. To avoid stalling on the expensive map and invalidate
 653  * operations, these operations are pipelined on the same send queue on
 654  * which we want to send the message containing the r_key.
 655  *
 656  * This creates a bit of a problem for us, as we do not have the destination
 657  * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
 658  * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
 659  * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
 660  * before queuing the SEND. When completions for these arrive, they are
 661  * dispatched to the MR has a bit set showing that RDMa can be performed.
 662  *
 663  * There is another interesting aspect that's related to invalidation.
 664  * The application can request that a mapping is invalidated in FREE_MR.
 665  * The expectation there is that this invalidation step includes ALL
 666  * PREVIOUSLY FREED MRs.
 667  */
 668 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
 669                                 struct rds_iw_mr *ibmr)
 670 {
 671         struct rds_iw_device *rds_iwdev = pool->device;
 672         struct ib_fast_reg_page_list *page_list = NULL;
 673         struct ib_mr *mr;
 674         int err;
 675
 676         mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
 677         if (IS_ERR(mr)) {
 678                 err = PTR_ERR(mr);
 679
 680                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
 681                 return err;
 682         }
 683
 684         /* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
 685          * is not filled in.
 686          */
 687         page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
 688         if (IS_ERR(page_list)) {
 689                 err = PTR_ERR(page_list);
 690
 691                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
 692                 ib_dereg_mr(mr);
 693                 return err;
 694         }
 695
 696         ibmr->page_list = page_list;
 697         ibmr->mr = mr;
 698         return 0;
 699 }
 700
 701 static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
 702 {
 703         struct rds_iw_mr *ibmr = mapping->m_mr;
 704         struct ib_send_wr f_wr, *failed_wr;
 705         int ret;
 706
 707         /*
 708          * Perform a WR for the fast_reg_mr. Each individual page
 709          * in the sg list is added to the fast reg page list and placed
 710          * inside the fast_reg_mr WR.  The key used is a rolling 8bit
 711          * counter, which should guarantee uniqueness.
 712          */
 713         ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
 714         mapping->m_rkey = ibmr->mr->rkey;
 715
 716         memset(&f_wr, 0, sizeof(f_wr));
 717         f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
 718         f_wr.opcode = IB_WR_FAST_REG_MR;
 719         f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
 720         f_wr.wr.fast_reg.rkey = mapping->m_rkey;
 721         f_wr.wr.fast_reg.page_list = ibmr->page_list;
 722         f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
 723         f_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
 724         f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
 725                                 IB_ACCESS_REMOTE_READ |
 726                                 IB_ACCESS_REMOTE_WRITE;
 727         f_wr.wr.fast_reg.iova_start = 0;
 728         f_wr.send_flags = IB_SEND_SIGNALED;
 729
 730         failed_wr = &f_wr;
 731         ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
 732         BUG_ON(failed_wr != &f_wr);
 733         if (ret)
 734                 printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 735                         __func__, __LINE__, ret);
 736         return ret;
 737 }
 738
 739 static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
 740 {
 741         struct ib_send_wr s_wr, *failed_wr;
 742         int ret = 0;
 743
 744         if (!ibmr->cm_id->qp || !ibmr->mr)
 745                 goto out;
 746
 747         memset(&s_wr, 0, sizeof(s_wr));
 748         s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
 749         s_wr.opcode = IB_WR_LOCAL_INV;
 750         s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
 751         s_wr.send_flags = IB_SEND_SIGNALED;
 752
 753         failed_wr = &s_wr;
 754         ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
 755         if (ret) {
 756                 printk_ratelimited(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 757                         __func__, __LINE__, ret);
 758                 goto out;
 759         }
 760 out:
 761         return ret;
 762 }
 763
 764 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
 765                         struct rds_iw_mr *ibmr,
 766                         struct scatterlist *sg,
 767                         unsigned int sg_len)
 768 {
 769         struct rds_iw_device *rds_iwdev = pool->device;
 770         struct rds_iw_mapping *mapping = &ibmr->mapping;
 771         u64 *dma_pages;
 772         int i, ret = 0;
 773
 774         rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
 775
 776         dma_pages = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
 777         if (IS_ERR(dma_pages)) {
 778                 ret = PTR_ERR(dma_pages);
 779                 dma_pages = NULL;
 780                 goto out;
 781         }
 782
 783         if (mapping->m_sg.dma_len > pool->max_message_size) {
 784                 ret = -EMSGSIZE;
 785                 goto out;
 786         }
 787
 788         for (i = 0; i < mapping->m_sg.dma_npages; ++i)
 789                 ibmr->page_list->page_list[i] = dma_pages[i];
 790
 791         ret = rds_iw_rdma_build_fastreg(mapping);
 792         if (ret)
 793                 goto out;
 794
 795         rds_iw_stats_inc(s_iw_rdma_mr_used);
 796
 797 out:
 798         kfree(dma_pages);
 799
 800         return ret;
 801 }
 802
 803 /*
 804  * "Free" a fastreg MR.
 805  */
 806 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
 807                 struct rds_iw_mr *ibmr)
 808 {
 809         unsigned long flags;
 810         int ret;
 811
 812         if (!ibmr->mapping.m_sg.dma_len)
 813                 return;
 814
 815         ret = rds_iw_rdma_fastreg_inv(ibmr);
 816         if (ret)
 817                 return;
 818
 819         /* Try to post the LOCAL_INV WR to the queue. */
 820         spin_lock_irqsave(&pool->list_lock, flags);
 821
 822         list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
 823         atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
 824         atomic_inc(&pool->dirty_count);
 825
 826         spin_unlock_irqrestore(&pool->list_lock, flags);
 827 }
 828
 829 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
 830                                 struct list_head *unmap_list,
 831                                 struct list_head *kill_list)
 832 {
 833         struct rds_iw_mapping *mapping, *next;
 834         unsigned int ncleaned = 0;
 835         LIST_HEAD(laundered);
 836
 837         /* Batched invalidation of fastreg MRs.
 838          * Why do we do it this way, even though we could pipeline unmap
 839          * and remap? The reason is the application semantics - when the
 840          * application requests an invalidation of MRs, it expects all
 841          * previously released R_Keys to become invalid.
 842          *
 843          * If we implement MR reuse naively, we risk memory corruption
 844          * (this has actually been observed). So the default behavior
 845          * requires that a MR goes through an explicit unmap operation before
 846          * we can reuse it again.
 847          *
 848          * We could probably improve on this a little, by allowing immediate
 849          * reuse of a MR on the same socket (eg you could add small
 850          * cache of unused MRs to strct rds_socket - GET_MR could grab one
 851          * of these without requiring an explicit invalidate).
 852          */
 853         while (!list_empty(unmap_list)) {
 854                 unsigned long flags;
 855
 856                 spin_lock_irqsave(&pool->list_lock, flags);
 857                 list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
 858                         list_move(&mapping->m_list, &laundered);
 859                         ncleaned++;
 860                 }
 861                 spin_unlock_irqrestore(&pool->list_lock, flags);
 862         }
 863
 864         /* Move all laundered mappings back to the unmap list.
 865          * We do not kill any WRs right now - it doesn't seem the
 866          * fastreg API has a max_remap limit. */
 867         list_splice_init(&laundered, unmap_list);
 868
 869         return ncleaned;
 870 }
 871
 872 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
 873                 struct rds_iw_mr *ibmr)
 874 {
 875         if (ibmr->page_list)
 876                 ib_free_fast_reg_page_list(ibmr->page_list);
 877         if (ibmr->mr)
 878                 ib_dereg_mr(ibmr->mr);
 879 }