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
  35 #include "rds.h"
  36 #include "rdma.h"
  37 #include "iw.h"
  38
  39
  40 /*
  41  * This is stored as mr->r_trans_private.
  42  */
  43 struct rds_iw_mr {
  44         struct rds_iw_device    *device;
  45         struct rds_iw_mr_pool   *pool;
  46         struct rdma_cm_id       *cm_id;
  47
  48         struct ib_mr    *mr;
  49         struct ib_fast_reg_page_list *page_list;
  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 int 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_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  80 static int rds_iw_map_fastreg(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 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool, struct rds_iw_mr *ibmr);
  88
  89 static int rds_iw_get_device(struct rds_sock *rs, struct rds_iw_device **rds_iwdev, struct rdma_cm_id **cm_id)
  90 {
  91         struct rds_iw_device *iwdev;
  92         struct rds_iw_cm_id *i_cm_id;
  93
  94         *rds_iwdev = NULL;
  95         *cm_id = NULL;
  96
  97         list_for_each_entry(iwdev, &rds_iw_devices, list) {
  98                 spin_lock_irq(&iwdev->spinlock);
  99                 list_for_each_entry(i_cm_id, &iwdev->cm_id_list, list) {
 100                         struct sockaddr_in *src_addr, *dst_addr;
 101
 102                         src_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.src_addr;
 103                         dst_addr = (struct sockaddr_in *)&i_cm_id->cm_id->route.addr.dst_addr;
 104
 105                         rdsdebug("local ipaddr = %x port %d, "
 106                                  "remote ipaddr = %x port %d"
 107                                  "..looking for %x port %d, "
 108                                  "remote ipaddr = %x port %d\n",
 109                                 src_addr->sin_addr.s_addr,
 110                                 src_addr->sin_port,
 111                                 dst_addr->sin_addr.s_addr,
 112                                 dst_addr->sin_port,
 113                                 rs->rs_bound_addr,
 114                                 rs->rs_bound_port,
 115                                 rs->rs_conn_addr,
 116                                 rs->rs_conn_port);
 117 #ifdef WORKING_TUPLE_DETECTION
 118                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr &&
 119                             src_addr->sin_port == rs->rs_bound_port &&
 120                             dst_addr->sin_addr.s_addr == rs->rs_conn_addr &&
 121                             dst_addr->sin_port == rs->rs_conn_port) {
 122 #else
 123                         /* FIXME - needs to compare the local and remote
 124                          * ipaddr/port tuple, but the ipaddr is the only
 125                          * available infomation in the rds_sock (as the rest are
 126                          * zero'ed.  It doesn't appear to be properly populated
 127                          * during connection setup...
 128                          */
 129                         if (src_addr->sin_addr.s_addr == rs->rs_bound_addr) {
 130 #endif
 131                                 spin_unlock_irq(&iwdev->spinlock);
 132                                 *rds_iwdev = iwdev;
 133                                 *cm_id = i_cm_id->cm_id;
 134                                 return 0;
 135                         }
 136                 }
 137                 spin_unlock_irq(&iwdev->spinlock);
 138         }
 139
 140         return 1;
 141 }
 142
 143 static int rds_iw_add_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 144 {
 145         struct rds_iw_cm_id *i_cm_id;
 146
 147         i_cm_id = kmalloc(sizeof *i_cm_id, GFP_KERNEL);
 148         if (!i_cm_id)
 149                 return -ENOMEM;
 150
 151         i_cm_id->cm_id = cm_id;
 152
 153         spin_lock_irq(&rds_iwdev->spinlock);
 154         list_add_tail(&i_cm_id->list, &rds_iwdev->cm_id_list);
 155         spin_unlock_irq(&rds_iwdev->spinlock);
 156
 157         return 0;
 158 }
 159
 160 void rds_iw_remove_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 161 {
 162         struct rds_iw_cm_id *i_cm_id;
 163
 164         spin_lock_irq(&rds_iwdev->spinlock);
 165         list_for_each_entry(i_cm_id, &rds_iwdev->cm_id_list, list) {
 166                 if (i_cm_id->cm_id == cm_id) {
 167                         list_del(&i_cm_id->list);
 168                         kfree(i_cm_id);
 169                         break;
 170                 }
 171         }
 172         spin_unlock_irq(&rds_iwdev->spinlock);
 173 }
 174
 175
 176 int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id)
 177 {
 178         struct sockaddr_in *src_addr, *dst_addr;
 179         struct rds_iw_device *rds_iwdev_old;
 180         struct rds_sock rs;
 181         struct rdma_cm_id *pcm_id;
 182         int rc;
 183
 184         src_addr = (struct sockaddr_in *)&cm_id->route.addr.src_addr;
 185         dst_addr = (struct sockaddr_in *)&cm_id->route.addr.dst_addr;
 186
 187         rs.rs_bound_addr = src_addr->sin_addr.s_addr;
 188         rs.rs_bound_port = src_addr->sin_port;
 189         rs.rs_conn_addr = dst_addr->sin_addr.s_addr;
 190         rs.rs_conn_port = dst_addr->sin_port;
 191
 192         rc = rds_iw_get_device(&rs, &rds_iwdev_old, &pcm_id);
 193         if (rc)
 194                 rds_iw_remove_cm_id(rds_iwdev, cm_id);
 195
 196         return rds_iw_add_cm_id(rds_iwdev, cm_id);
 197 }
 198
 199 void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 200 {
 201         struct rds_iw_connection *ic = conn->c_transport_data;
 202
 203         /* conn was previously on the nodev_conns_list */
 204         spin_lock_irq(&iw_nodev_conns_lock);
 205         BUG_ON(list_empty(&iw_nodev_conns));
 206         BUG_ON(list_empty(&ic->iw_node));
 207         list_del(&ic->iw_node);
 208
 209         spin_lock_irq(&rds_iwdev->spinlock);
 210         list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
 211         spin_unlock_irq(&rds_iwdev->spinlock);
 212         spin_unlock_irq(&iw_nodev_conns_lock);
 213
 214         ic->rds_iwdev = rds_iwdev;
 215 }
 216
 217 void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
 218 {
 219         struct rds_iw_connection *ic = conn->c_transport_data;
 220
 221         /* place conn on nodev_conns_list */
 222         spin_lock(&iw_nodev_conns_lock);
 223
 224         spin_lock_irq(&rds_iwdev->spinlock);
 225         BUG_ON(list_empty(&ic->iw_node));
 226         list_del(&ic->iw_node);
 227         spin_unlock_irq(&rds_iwdev->spinlock);
 228
 229         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 230
 231         spin_unlock(&iw_nodev_conns_lock);
 232
 233         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
 234         ic->rds_iwdev = NULL;
 235 }
 236
 237 void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
 238 {
 239         struct rds_iw_connection *ic, *_ic;
 240         LIST_HEAD(tmp_list);
 241
 242         /* avoid calling conn_destroy with irqs off */
 243         spin_lock_irq(list_lock);
 244         list_splice(list, &tmp_list);
 245         INIT_LIST_HEAD(list);
 246         spin_unlock_irq(list_lock);
 247
 248         list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node)
 249                 rds_conn_destroy(ic->conn);
 250 }
 251
 252 static void rds_iw_set_scatterlist(struct rds_iw_scatterlist *sg,
 253                 struct scatterlist *list, unsigned int sg_len)
 254 {
 255         sg->list = list;
 256         sg->len = sg_len;
 257         sg->dma_len = 0;
 258         sg->dma_npages = 0;
 259         sg->bytes = 0;
 260 }
 261
 262 static u64 *rds_iw_map_scatterlist(struct rds_iw_device *rds_iwdev,
 263                         struct rds_iw_scatterlist *sg)
 264 {
 265         struct ib_device *dev = rds_iwdev->dev;
 266         u64 *dma_pages = NULL;
 267         int i, j, ret;
 268
 269         WARN_ON(sg->dma_len);
 270
 271         sg->dma_len = ib_dma_map_sg(dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 272         if (unlikely(!sg->dma_len)) {
 273                 printk(KERN_WARNING "RDS/IW: dma_map_sg failed!\n");
 274                 return ERR_PTR(-EBUSY);
 275         }
 276
 277         sg->bytes = 0;
 278         sg->dma_npages = 0;
 279
 280         ret = -EINVAL;
 281         for (i = 0; i < sg->dma_len; ++i) {
 282                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
 283                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
 284                 u64 end_addr;
 285
 286                 sg->bytes += dma_len;
 287
 288                 end_addr = dma_addr + dma_len;
 289                 if (dma_addr & PAGE_MASK) {
 290                         if (i > 0)
 291                                 goto out_unmap;
 292                         dma_addr &= ~PAGE_MASK;
 293                 }
 294                 if (end_addr & PAGE_MASK) {
 295                         if (i < sg->dma_len - 1)
 296                                 goto out_unmap;
 297                         end_addr = (end_addr + PAGE_MASK) & ~PAGE_MASK;
 298                 }
 299
 300                 sg->dma_npages += (end_addr - dma_addr) >> PAGE_SHIFT;
 301         }
 302
 303         /* Now gather the dma addrs into one list */
 304         if (sg->dma_npages > fastreg_message_size)
 305                 goto out_unmap;
 306
 307         dma_pages = kmalloc(sizeof(u64) * sg->dma_npages, GFP_ATOMIC);
 308         if (!dma_pages) {
 309                 ret = -ENOMEM;
 310                 goto out_unmap;
 311         }
 312
 313         for (i = j = 0; i < sg->dma_len; ++i) {
 314                 unsigned int dma_len = ib_sg_dma_len(dev, &sg->list[i]);
 315                 u64 dma_addr = ib_sg_dma_address(dev, &sg->list[i]);
 316                 u64 end_addr;
 317
 318                 end_addr = dma_addr + dma_len;
 319                 dma_addr &= ~PAGE_MASK;
 320                 for (; dma_addr < end_addr; dma_addr += PAGE_SIZE)
 321                         dma_pages[j++] = dma_addr;
 322                 BUG_ON(j > sg->dma_npages);
 323         }
 324
 325         return dma_pages;
 326
 327 out_unmap:
 328         ib_dma_unmap_sg(rds_iwdev->dev, sg->list, sg->len, DMA_BIDIRECTIONAL);
 329         sg->dma_len = 0;
 330         kfree(dma_pages);
 331         return ERR_PTR(ret);
 332 }
 333
 334
 335 struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *rds_iwdev)
 336 {
 337         struct rds_iw_mr_pool *pool;
 338
 339         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 340         if (!pool) {
 341                 printk(KERN_WARNING "RDS/IW: rds_iw_create_mr_pool alloc error\n");
 342                 return ERR_PTR(-ENOMEM);
 343         }
 344
 345         pool->device = rds_iwdev;
 346         INIT_LIST_HEAD(&pool->dirty_list);
 347         INIT_LIST_HEAD(&pool->clean_list);
 348         mutex_init(&pool->flush_lock);
 349         spin_lock_init(&pool->list_lock);
 350         INIT_WORK(&pool->flush_worker, rds_iw_mr_pool_flush_worker);
 351
 352         pool->max_message_size = fastreg_message_size;
 353         pool->max_items = fastreg_pool_size;
 354         pool->max_free_pinned = pool->max_items * pool->max_message_size / 4;
 355         pool->max_pages = fastreg_message_size;
 356
 357         /* We never allow more than max_items MRs to be allocated.
 358          * When we exceed more than max_items_soft, we start freeing
 359          * items more aggressively.
 360          * Make sure that max_items > max_items_soft > max_items / 2
 361          */
 362         pool->max_items_soft = pool->max_items * 3 / 4;
 363
 364         return pool;
 365 }
 366
 367 void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo)
 368 {
 369         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 370
 371         iinfo->rdma_mr_max = pool->max_items;
 372         iinfo->rdma_mr_size = pool->max_pages;
 373 }
 374
 375 void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *pool)
 376 {
 377         flush_workqueue(rds_wq);
 378         rds_iw_flush_mr_pool(pool, 1);
 379         BUG_ON(atomic_read(&pool->item_count));
 380         BUG_ON(atomic_read(&pool->free_pinned));
 381         kfree(pool);
 382 }
 383
 384 static inline struct rds_iw_mr *rds_iw_reuse_fmr(struct rds_iw_mr_pool *pool)
 385 {
 386         struct rds_iw_mr *ibmr = NULL;
 387         unsigned long flags;
 388
 389         spin_lock_irqsave(&pool->list_lock, flags);
 390         if (!list_empty(&pool->clean_list)) {
 391                 ibmr = list_entry(pool->clean_list.next, struct rds_iw_mr, mapping.m_list);
 392                 list_del_init(&ibmr->mapping.m_list);
 393         }
 394         spin_unlock_irqrestore(&pool->list_lock, flags);
 395
 396         return ibmr;
 397 }
 398
 399 static struct rds_iw_mr *rds_iw_alloc_mr(struct rds_iw_device *rds_iwdev)
 400 {
 401         struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 402         struct rds_iw_mr *ibmr = NULL;
 403         int err = 0, iter = 0;
 404
 405         while (1) {
 406                 ibmr = rds_iw_reuse_fmr(pool);
 407                 if (ibmr)
 408                         return ibmr;
 409
 410                 /* No clean MRs - now we have the choice of either
 411                  * allocating a fresh MR up to the limit imposed by the
 412                  * driver, or flush any dirty unused MRs.
 413                  * We try to avoid stalling in the send path if possible,
 414                  * so we allocate as long as we're allowed to.
 415                  *
 416                  * We're fussy with enforcing the FMR limit, though. If the driver
 417                  * tells us we can't use more than N fmrs, we shouldn't start
 418                  * arguing with it */
 419                 if (atomic_inc_return(&pool->item_count) <= pool->max_items)
 420                         break;
 421
 422                 atomic_dec(&pool->item_count);
 423
 424                 if (++iter > 2) {
 425                         rds_iw_stats_inc(s_iw_rdma_mr_pool_depleted);
 426                         return ERR_PTR(-EAGAIN);
 427                 }
 428
 429                 /* We do have some empty MRs. Flush them out. */
 430                 rds_iw_stats_inc(s_iw_rdma_mr_pool_wait);
 431                 rds_iw_flush_mr_pool(pool, 0);
 432         }
 433
 434         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
 435         if (!ibmr) {
 436                 err = -ENOMEM;
 437                 goto out_no_cigar;
 438         }
 439
 440         spin_lock_init(&ibmr->mapping.m_lock);
 441         INIT_LIST_HEAD(&ibmr->mapping.m_list);
 442         ibmr->mapping.m_mr = ibmr;
 443
 444         err = rds_iw_init_fastreg(pool, ibmr);
 445         if (err)
 446                 goto out_no_cigar;
 447
 448         rds_iw_stats_inc(s_iw_rdma_mr_alloc);
 449         return ibmr;
 450
 451 out_no_cigar:
 452         if (ibmr) {
 453                 rds_iw_destroy_fastreg(pool, ibmr);
 454                 kfree(ibmr);
 455         }
 456         atomic_dec(&pool->item_count);
 457         return ERR_PTR(err);
 458 }
 459
 460 void rds_iw_sync_mr(void *trans_private, int direction)
 461 {
 462         struct rds_iw_mr *ibmr = trans_private;
 463         struct rds_iw_device *rds_iwdev = ibmr->device;
 464
 465         switch (direction) {
 466         case DMA_FROM_DEVICE:
 467                 ib_dma_sync_sg_for_cpu(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 468                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 469                 break;
 470         case DMA_TO_DEVICE:
 471                 ib_dma_sync_sg_for_device(rds_iwdev->dev, ibmr->mapping.m_sg.list,
 472                         ibmr->mapping.m_sg.dma_len, DMA_BIDIRECTIONAL);
 473                 break;
 474         }
 475 }
 476
 477 static inline unsigned int rds_iw_flush_goal(struct rds_iw_mr_pool *pool, int free_all)
 478 {
 479         unsigned int item_count;
 480
 481         item_count = atomic_read(&pool->item_count);
 482         if (free_all)
 483                 return item_count;
 484
 485         return 0;
 486 }
 487
 488 /*
 489  * Flush our pool of MRs.
 490  * At a minimum, all currently unused MRs are unmapped.
 491  * If the number of MRs allocated exceeds the limit, we also try
 492  * to free as many MRs as needed to get back to this limit.
 493  */
 494 static int rds_iw_flush_mr_pool(struct rds_iw_mr_pool *pool, int free_all)
 495 {
 496         struct rds_iw_mr *ibmr, *next;
 497         LIST_HEAD(unmap_list);
 498         LIST_HEAD(kill_list);
 499         unsigned long flags;
 500         unsigned int nfreed = 0, ncleaned = 0, free_goal;
 501         int ret = 0;
 502
 503         rds_iw_stats_inc(s_iw_rdma_mr_pool_flush);
 504
 505         mutex_lock(&pool->flush_lock);
 506
 507         spin_lock_irqsave(&pool->list_lock, flags);
 508         /* Get the list of all mappings to be destroyed */
 509         list_splice_init(&pool->dirty_list, &unmap_list);
 510         if (free_all)
 511                 list_splice_init(&pool->clean_list, &kill_list);
 512         spin_unlock_irqrestore(&pool->list_lock, flags);
 513
 514         free_goal = rds_iw_flush_goal(pool, free_all);
 515
 516         /* Batched invalidate of dirty MRs.
 517          * For FMR based MRs, the mappings on the unmap list are
 518          * actually members of an ibmr (ibmr->mapping). They either
 519          * migrate to the kill_list, or have been cleaned and should be
 520          * moved to the clean_list.
 521          * For fastregs, they will be dynamically allocated, and
 522          * will be destroyed by the unmap function.
 523          */
 524         if (!list_empty(&unmap_list)) {
 525                 ncleaned = rds_iw_unmap_fastreg_list(pool, &unmap_list, &kill_list);
 526                 /* If we've been asked to destroy all MRs, move those
 527                  * that were simply cleaned to the kill list */
 528                 if (free_all)
 529                         list_splice_init(&unmap_list, &kill_list);
 530         }
 531
 532         /* Destroy any MRs that are past their best before date */
 533         list_for_each_entry_safe(ibmr, next, &kill_list, mapping.m_list) {
 534                 rds_iw_stats_inc(s_iw_rdma_mr_free);
 535                 list_del(&ibmr->mapping.m_list);
 536                 rds_iw_destroy_fastreg(pool, ibmr);
 537                 kfree(ibmr);
 538                 nfreed++;
 539         }
 540
 541         /* Anything that remains are laundered ibmrs, which we can add
 542          * back to the clean list. */
 543         if (!list_empty(&unmap_list)) {
 544                 spin_lock_irqsave(&pool->list_lock, flags);
 545                 list_splice(&unmap_list, &pool->clean_list);
 546                 spin_unlock_irqrestore(&pool->list_lock, flags);
 547         }
 548
 549         atomic_sub(ncleaned, &pool->dirty_count);
 550         atomic_sub(nfreed, &pool->item_count);
 551
 552         mutex_unlock(&pool->flush_lock);
 553         return ret;
 554 }
 555
 556 static void rds_iw_mr_pool_flush_worker(struct work_struct *work)
 557 {
 558         struct rds_iw_mr_pool *pool = container_of(work, struct rds_iw_mr_pool, flush_worker);
 559
 560         rds_iw_flush_mr_pool(pool, 0);
 561 }
 562
 563 void rds_iw_free_mr(void *trans_private, int invalidate)
 564 {
 565         struct rds_iw_mr *ibmr = trans_private;
 566         struct rds_iw_mr_pool *pool = ibmr->device->mr_pool;
 567
 568         rdsdebug("RDS/IW: free_mr nents %u\n", ibmr->mapping.m_sg.len);
 569         if (!pool)
 570                 return;
 571
 572         /* Return it to the pool's free list */
 573         rds_iw_free_fastreg(pool, ibmr);
 574
 575         /* If we've pinned too many pages, request a flush */
 576         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
 577             atomic_read(&pool->dirty_count) >= pool->max_items / 10)
 578                 queue_work(rds_wq, &pool->flush_worker);
 579
 580         if (invalidate) {
 581                 if (likely(!in_interrupt())) {
 582                         rds_iw_flush_mr_pool(pool, 0);
 583                 } else {
 584                         /* We get here if the user created a MR marked
 585                          * as use_once and invalidate at the same time. */
 586                         queue_work(rds_wq, &pool->flush_worker);
 587                 }
 588         }
 589 }
 590
 591 void rds_iw_flush_mrs(void)
 592 {
 593         struct rds_iw_device *rds_iwdev;
 594
 595         list_for_each_entry(rds_iwdev, &rds_iw_devices, list) {
 596                 struct rds_iw_mr_pool *pool = rds_iwdev->mr_pool;
 597
 598                 if (pool)
 599                         rds_iw_flush_mr_pool(pool, 0);
 600         }
 601 }
 602
 603 void *rds_iw_get_mr(struct scatterlist *sg, unsigned long nents,
 604                     struct rds_sock *rs, u32 *key_ret)
 605 {
 606         struct rds_iw_device *rds_iwdev;
 607         struct rds_iw_mr *ibmr = NULL;
 608         struct rdma_cm_id *cm_id;
 609         int ret;
 610
 611         ret = rds_iw_get_device(rs, &rds_iwdev, &cm_id);
 612         if (ret || !cm_id) {
 613                 ret = -ENODEV;
 614                 goto out;
 615         }
 616
 617         if (!rds_iwdev->mr_pool) {
 618                 ret = -ENODEV;
 619                 goto out;
 620         }
 621
 622         ibmr = rds_iw_alloc_mr(rds_iwdev);
 623         if (IS_ERR(ibmr))
 624                 return ibmr;
 625
 626         ibmr->cm_id = cm_id;
 627         ibmr->device = rds_iwdev;
 628
 629         ret = rds_iw_map_fastreg(rds_iwdev->mr_pool, ibmr, sg, nents);
 630         if (ret == 0)
 631                 *key_ret = ibmr->mr->rkey;
 632         else
 633                 printk(KERN_WARNING "RDS/IW: failed to map mr (errno=%d)\n", ret);
 634
 635 out:
 636         if (ret) {
 637                 if (ibmr)
 638                         rds_iw_free_mr(ibmr, 0);
 639                 ibmr = ERR_PTR(ret);
 640         }
 641         return ibmr;
 642 }
 643
 644 /*
 645  * iWARP fastreg handling
 646  *
 647  * The life cycle of a fastreg registration is a bit different from
 648  * FMRs.
 649  * The idea behind fastreg is to have one MR, to which we bind different
 650  * mappings over time. To avoid stalling on the expensive map and invalidate
 651  * operations, these operations are pipelined on the same send queue on
 652  * which we want to send the message containing the r_key.
 653  *
 654  * This creates a bit of a problem for us, as we do not have the destination
 655  * IP in GET_MR, so the connection must be setup prior to the GET_MR call for
 656  * RDMA to be correctly setup.  If a fastreg request is present, rds_iw_xmit
 657  * will try to queue a LOCAL_INV (if needed) and a FAST_REG_MR work request
 658  * before queuing the SEND. When completions for these arrive, they are
 659  * dispatched to the MR has a bit set showing that RDMa can be performed.
 660  *
 661  * There is another interesting aspect that's related to invalidation.
 662  * The application can request that a mapping is invalidated in FREE_MR.
 663  * The expectation there is that this invalidation step includes ALL
 664  * PREVIOUSLY FREED MRs.
 665  */
 666 static int rds_iw_init_fastreg(struct rds_iw_mr_pool *pool,
 667                                 struct rds_iw_mr *ibmr)
 668 {
 669         struct rds_iw_device *rds_iwdev = pool->device;
 670         struct ib_fast_reg_page_list *page_list = NULL;
 671         struct ib_mr *mr;
 672         int err;
 673
 674         mr = ib_alloc_fast_reg_mr(rds_iwdev->pd, pool->max_message_size);
 675         if (IS_ERR(mr)) {
 676                 err = PTR_ERR(mr);
 677
 678                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed (err=%d)\n", err);
 679                 return err;
 680         }
 681
 682         /* FIXME - this is overkill, but mapping->m_sg.dma_len/mapping->m_sg.dma_npages
 683          * is not filled in.
 684          */
 685         page_list = ib_alloc_fast_reg_page_list(rds_iwdev->dev, pool->max_message_size);
 686         if (IS_ERR(page_list)) {
 687                 err = PTR_ERR(page_list);
 688
 689                 printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed (err=%d)\n", err);
 690                 ib_dereg_mr(mr);
 691                 return err;
 692         }
 693
 694         ibmr->page_list = page_list;
 695         ibmr->mr = mr;
 696         return 0;
 697 }
 698
 699 static int rds_iw_rdma_build_fastreg(struct rds_iw_mapping *mapping)
 700 {
 701         struct rds_iw_mr *ibmr = mapping->m_mr;
 702         struct ib_send_wr f_wr, *failed_wr;
 703         int ret;
 704
 705         /*
 706          * Perform a WR for the fast_reg_mr. Each individual page
 707          * in the sg list is added to the fast reg page list and placed
 708          * inside the fast_reg_mr WR.  The key used is a rolling 8bit
 709          * counter, which should guarantee uniqueness.
 710          */
 711         ib_update_fast_reg_key(ibmr->mr, ibmr->remap_count++);
 712         mapping->m_rkey = ibmr->mr->rkey;
 713
 714         memset(&f_wr, 0, sizeof(f_wr));
 715         f_wr.wr_id = RDS_IW_FAST_REG_WR_ID;
 716         f_wr.opcode = IB_WR_FAST_REG_MR;
 717         f_wr.wr.fast_reg.length = mapping->m_sg.bytes;
 718         f_wr.wr.fast_reg.rkey = mapping->m_rkey;
 719         f_wr.wr.fast_reg.page_list = ibmr->page_list;
 720         f_wr.wr.fast_reg.page_list_len = mapping->m_sg.dma_len;
 721         f_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
 722         f_wr.wr.fast_reg.access_flags = IB_ACCESS_LOCAL_WRITE |
 723                                 IB_ACCESS_REMOTE_READ |
 724                                 IB_ACCESS_REMOTE_WRITE;
 725         f_wr.wr.fast_reg.iova_start = 0;
 726         f_wr.send_flags = IB_SEND_SIGNALED;
 727
 728         failed_wr = &f_wr;
 729         ret = ib_post_send(ibmr->cm_id->qp, &f_wr, &failed_wr);
 730         BUG_ON(failed_wr != &f_wr);
 731         if (ret && printk_ratelimit())
 732                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 733                         __func__, __LINE__, ret);
 734         return ret;
 735 }
 736
 737 static int rds_iw_rdma_fastreg_inv(struct rds_iw_mr *ibmr)
 738 {
 739         struct ib_send_wr s_wr, *failed_wr;
 740         int ret = 0;
 741
 742         if (!ibmr->cm_id->qp || !ibmr->mr)
 743                 goto out;
 744
 745         memset(&s_wr, 0, sizeof(s_wr));
 746         s_wr.wr_id = RDS_IW_LOCAL_INV_WR_ID;
 747         s_wr.opcode = IB_WR_LOCAL_INV;
 748         s_wr.ex.invalidate_rkey = ibmr->mr->rkey;
 749         s_wr.send_flags = IB_SEND_SIGNALED;
 750
 751         failed_wr = &s_wr;
 752         ret = ib_post_send(ibmr->cm_id->qp, &s_wr, &failed_wr);
 753         if (ret && printk_ratelimit()) {
 754                 printk(KERN_WARNING "RDS/IW: %s:%d ib_post_send returned %d\n",
 755                         __func__, __LINE__, ret);
 756                 goto out;
 757         }
 758 out:
 759         return ret;
 760 }
 761
 762 static int rds_iw_map_fastreg(struct rds_iw_mr_pool *pool,
 763                         struct rds_iw_mr *ibmr,
 764                         struct scatterlist *sg,
 765                         unsigned int sg_len)
 766 {
 767         struct rds_iw_device *rds_iwdev = pool->device;
 768         struct rds_iw_mapping *mapping = &ibmr->mapping;
 769         u64 *dma_pages;
 770         int i, ret = 0;
 771
 772         rds_iw_set_scatterlist(&mapping->m_sg, sg, sg_len);
 773
 774         dma_pages = rds_iw_map_scatterlist(rds_iwdev, &mapping->m_sg);
 775         if (IS_ERR(dma_pages)) {
 776                 ret = PTR_ERR(dma_pages);
 777                 dma_pages = NULL;
 778                 goto out;
 779         }
 780
 781         if (mapping->m_sg.dma_len > pool->max_message_size) {
 782                 ret = -EMSGSIZE;
 783                 goto out;
 784         }
 785
 786         for (i = 0; i < mapping->m_sg.dma_npages; ++i)
 787                 ibmr->page_list->page_list[i] = dma_pages[i];
 788
 789         ret = rds_iw_rdma_build_fastreg(mapping);
 790         if (ret)
 791                 goto out;
 792
 793         rds_iw_stats_inc(s_iw_rdma_mr_used);
 794
 795 out:
 796         kfree(dma_pages);
 797
 798         return ret;
 799 }
 800
 801 /*
 802  * "Free" a fastreg MR.
 803  */
 804 static void rds_iw_free_fastreg(struct rds_iw_mr_pool *pool,
 805                 struct rds_iw_mr *ibmr)
 806 {
 807         unsigned long flags;
 808         int ret;
 809
 810         if (!ibmr->mapping.m_sg.dma_len)
 811                 return;
 812
 813         ret = rds_iw_rdma_fastreg_inv(ibmr);
 814         if (ret)
 815                 return;
 816
 817         /* Try to post the LOCAL_INV WR to the queue. */
 818         spin_lock_irqsave(&pool->list_lock, flags);
 819
 820         list_add_tail(&ibmr->mapping.m_list, &pool->dirty_list);
 821         atomic_add(ibmr->mapping.m_sg.len, &pool->free_pinned);
 822         atomic_inc(&pool->dirty_count);
 823
 824         spin_unlock_irqrestore(&pool->list_lock, flags);
 825 }
 826
 827 static unsigned int rds_iw_unmap_fastreg_list(struct rds_iw_mr_pool *pool,
 828                                 struct list_head *unmap_list,
 829                                 struct list_head *kill_list)
 830 {
 831         struct rds_iw_mapping *mapping, *next;
 832         unsigned int ncleaned = 0;
 833         LIST_HEAD(laundered);
 834
 835         /* Batched invalidation of fastreg MRs.
 836          * Why do we do it this way, even though we could pipeline unmap
 837          * and remap? The reason is the application semantics - when the
 838          * application requests an invalidation of MRs, it expects all
 839          * previously released R_Keys to become invalid.
 840          *
 841          * If we implement MR reuse naively, we risk memory corruption
 842          * (this has actually been observed). So the default behavior
 843          * requires that a MR goes through an explicit unmap operation before
 844          * we can reuse it again.
 845          *
 846          * We could probably improve on this a little, by allowing immediate
 847          * reuse of a MR on the same socket (eg you could add small
 848          * cache of unused MRs to strct rds_socket - GET_MR could grab one
 849          * of these without requiring an explicit invalidate).
 850          */
 851         while (!list_empty(unmap_list)) {
 852                 unsigned long flags;
 853
 854                 spin_lock_irqsave(&pool->list_lock, flags);
 855                 list_for_each_entry_safe(mapping, next, unmap_list, m_list) {
 856                         list_move(&mapping->m_list, &laundered);
 857                         ncleaned++;
 858                 }
 859                 spin_unlock_irqrestore(&pool->list_lock, flags);
 860         }
 861
 862         /* Move all laundered mappings back to the unmap list.
 863          * We do not kill any WRs right now - it doesn't seem the
 864          * fastreg API has a max_remap limit. */
 865         list_splice_init(&laundered, unmap_list);
 866
 867         return ncleaned;
 868 }
 869
 870 static void rds_iw_destroy_fastreg(struct rds_iw_mr_pool *pool,
 871                 struct rds_iw_mr *ibmr)
 872 {
 873         if (ibmr->page_list)
 874                 ib_free_fast_reg_page_list(ibmr->page_list);
 875         if (ibmr->mr)
 876                 ib_dereg_mr(ibmr->mr);
 877 }