net/rds/ib_rdma.c

0001 /*
0002  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/slab.h>
0035 #include <linux/rculist.h>
0036 #include <linux/llist.h>
0037
0038 #include "rds_single_path.h"
0039 #include "ib_mr.h"
0040 #include "rds.h"
0041
0042 struct workqueue_struct *rds_ib_mr_wq;
0043 struct rds_ib_dereg_odp_mr {
0044     struct work_struct work;
0045     struct ib_mr *mr;
0046 };
0047
0048 static void rds_ib_odp_mr_worker(struct work_struct *work);
0049
0050 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
0051 {
0052     struct rds_ib_device *rds_ibdev;
0053     struct rds_ib_ipaddr *i_ipaddr;
0054
0055     rcu_read_lock();
0056     list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
0057         list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
0058             if (i_ipaddr->ipaddr == ipaddr) {
0059                 refcount_inc(&rds_ibdev->refcount);
0060                 rcu_read_unlock();
0061                 return rds_ibdev;
0062             }
0063         }
0064     }
0065     rcu_read_unlock();
0066
0067     return NULL;
0068 }
0069
0070 static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
0071 {
0072     struct rds_ib_ipaddr *i_ipaddr;
0073
0074     i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
0075     if (!i_ipaddr)
0076         return -ENOMEM;
0077
0078     i_ipaddr->ipaddr = ipaddr;
0079
0080     spin_lock_irq(&rds_ibdev->spinlock);
0081     list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
0082     spin_unlock_irq(&rds_ibdev->spinlock);
0083
0084     return 0;
0085 }
0086
0087 static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
0088 {
0089     struct rds_ib_ipaddr *i_ipaddr;
0090     struct rds_ib_ipaddr *to_free = NULL;
0091
0092
0093     spin_lock_irq(&rds_ibdev->spinlock);
0094     list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
0095         if (i_ipaddr->ipaddr == ipaddr) {
0096             list_del_rcu(&i_ipaddr->list);
0097             to_free = i_ipaddr;
0098             break;
0099         }
0100     }
0101     spin_unlock_irq(&rds_ibdev->spinlock);
0102
0103     if (to_free)
0104         kfree_rcu(to_free, rcu);
0105 }
0106
0107 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
0108              struct in6_addr *ipaddr)
0109 {
0110     struct rds_ib_device *rds_ibdev_old;
0111
0112     rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]);
0113     if (!rds_ibdev_old)
0114         return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
0115
0116     if (rds_ibdev_old != rds_ibdev) {
0117         rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]);
0118         rds_ib_dev_put(rds_ibdev_old);
0119         return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]);
0120     }
0121     rds_ib_dev_put(rds_ibdev_old);
0122
0123     return 0;
0124 }
0125
0126 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
0127 {
0128     struct rds_ib_connection *ic = conn->c_transport_data;
0129
0130     /* conn was previously on the nodev_conns_list */
0131     spin_lock_irq(&ib_nodev_conns_lock);
0132     BUG_ON(list_empty(&ib_nodev_conns));
0133     BUG_ON(list_empty(&ic->ib_node));
0134     list_del(&ic->ib_node);
0135
0136     spin_lock(&rds_ibdev->spinlock);
0137     list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
0138     spin_unlock(&rds_ibdev->spinlock);
0139     spin_unlock_irq(&ib_nodev_conns_lock);
0140
0141     ic->rds_ibdev = rds_ibdev;
0142     refcount_inc(&rds_ibdev->refcount);
0143 }
0144
0145 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
0146 {
0147     struct rds_ib_connection *ic = conn->c_transport_data;
0148
0149     /* place conn on nodev_conns_list */
0150     spin_lock(&ib_nodev_conns_lock);
0151
0152     spin_lock_irq(&rds_ibdev->spinlock);
0153     BUG_ON(list_empty(&ic->ib_node));
0154     list_del(&ic->ib_node);
0155     spin_unlock_irq(&rds_ibdev->spinlock);
0156
0157     list_add_tail(&ic->ib_node, &ib_nodev_conns);
0158
0159     spin_unlock(&ib_nodev_conns_lock);
0160
0161     ic->rds_ibdev = NULL;
0162     rds_ib_dev_put(rds_ibdev);
0163 }
0164
0165 void rds_ib_destroy_nodev_conns(void)
0166 {
0167     struct rds_ib_connection *ic, *_ic;
0168     LIST_HEAD(tmp_list);
0169
0170     /* avoid calling conn_destroy with irqs off */
0171     spin_lock_irq(&ib_nodev_conns_lock);
0172     list_splice(&ib_nodev_conns, &tmp_list);
0173     spin_unlock_irq(&ib_nodev_conns_lock);
0174
0175     list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
0176         rds_conn_destroy(ic->conn);
0177 }
0178
0179 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
0180 {
0181     struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
0182
0183     iinfo->rdma_mr_max = pool_1m->max_items;
0184     iinfo->rdma_mr_size = pool_1m->max_pages;
0185 }
0186
0187 #if IS_ENABLED(CONFIG_IPV6)
0188 void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
0189              struct rds6_info_rdma_connection *iinfo6)
0190 {
0191     struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
0192
0193     iinfo6->rdma_mr_max = pool_1m->max_items;
0194     iinfo6->rdma_mr_size = pool_1m->max_pages;
0195 }
0196 #endif
0197
0198 struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool)
0199 {
0200     struct rds_ib_mr *ibmr = NULL;
0201     struct llist_node *ret;
0202     unsigned long flags;
0203
0204     spin_lock_irqsave(&pool->clean_lock, flags);
0205     ret = llist_del_first(&pool->clean_list);
0206     spin_unlock_irqrestore(&pool->clean_lock, flags);
0207     if (ret) {
0208         ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
0209         if (pool->pool_type == RDS_IB_MR_8K_POOL)
0210             rds_ib_stats_inc(s_ib_rdma_mr_8k_reused);
0211         else
0212             rds_ib_stats_inc(s_ib_rdma_mr_1m_reused);
0213     }
0214
0215     return ibmr;
0216 }
0217
0218 void rds_ib_sync_mr(void *trans_private, int direction)
0219 {
0220     struct rds_ib_mr *ibmr = trans_private;
0221     struct rds_ib_device *rds_ibdev = ibmr->device;
0222
0223     if (ibmr->odp)
0224         return;
0225
0226     switch (direction) {
0227     case DMA_FROM_DEVICE:
0228         ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
0229             ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
0230         break;
0231     case DMA_TO_DEVICE:
0232         ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
0233             ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
0234         break;
0235     }
0236 }
0237
0238 void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
0239 {
0240     struct rds_ib_device *rds_ibdev = ibmr->device;
0241
0242     if (ibmr->sg_dma_len) {
0243         ib_dma_unmap_sg(rds_ibdev->dev,
0244                 ibmr->sg, ibmr->sg_len,
0245                 DMA_BIDIRECTIONAL);
0246         ibmr->sg_dma_len = 0;
0247     }
0248
0249     /* Release the s/g list */
0250     if (ibmr->sg_len) {
0251         unsigned int i;
0252
0253         for (i = 0; i < ibmr->sg_len; ++i) {
0254             struct page *page = sg_page(&ibmr->sg[i]);
0255
0256             /* FIXME we need a way to tell a r/w MR
0257              * from a r/o MR */
0258             WARN_ON(!page->mapping && irqs_disabled());
0259             set_page_dirty(page);
0260             put_page(page);
0261         }
0262         kfree(ibmr->sg);
0263
0264         ibmr->sg = NULL;
0265         ibmr->sg_len = 0;
0266     }
0267 }
0268
0269 void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
0270 {
0271     unsigned int pinned = ibmr->sg_len;
0272
0273     __rds_ib_teardown_mr(ibmr);
0274     if (pinned) {
0275         struct rds_ib_mr_pool *pool = ibmr->pool;
0276
0277         atomic_sub(pinned, &pool->free_pinned);
0278     }
0279 }
0280
0281 static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
0282 {
0283     unsigned int item_count;
0284
0285     item_count = atomic_read(&pool->item_count);
0286     if (free_all)
0287         return item_count;
0288
0289     return 0;
0290 }
0291
0292 /*
0293  * given an llist of mrs, put them all into the list_head for more processing
0294  */
0295 static unsigned int llist_append_to_list(struct llist_head *llist,
0296                      struct list_head *list)
0297 {
0298     struct rds_ib_mr *ibmr;
0299     struct llist_node *node;
0300     struct llist_node *next;
0301     unsigned int count = 0;
0302
0303     node = llist_del_all(llist);
0304     while (node) {
0305         next = node->next;
0306         ibmr = llist_entry(node, struct rds_ib_mr, llnode);
0307         list_add_tail(&ibmr->unmap_list, list);
0308         node = next;
0309         count++;
0310     }
0311     return count;
0312 }
0313
0314 /*
0315  * this takes a list head of mrs and turns it into linked llist nodes
0316  * of clusters.  Each cluster has linked llist nodes of
0317  * MR_CLUSTER_SIZE mrs that are ready for reuse.
0318  */
0319 static void list_to_llist_nodes(struct list_head *list,
0320                 struct llist_node **nodes_head,
0321                 struct llist_node **nodes_tail)
0322 {
0323     struct rds_ib_mr *ibmr;
0324     struct llist_node *cur = NULL;
0325     struct llist_node **next = nodes_head;
0326
0327     list_for_each_entry(ibmr, list, unmap_list) {
0328         cur = &ibmr->llnode;
0329         *next = cur;
0330         next = &cur->next;
0331     }
0332     *next = NULL;
0333     *nodes_tail = cur;
0334 }
0335
0336 /*
0337  * Flush our pool of MRs.
0338  * At a minimum, all currently unused MRs are unmapped.
0339  * If the number of MRs allocated exceeds the limit, we also try
0340  * to free as many MRs as needed to get back to this limit.
0341  */
0342 int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
0343              int free_all, struct rds_ib_mr **ibmr_ret)
0344 {
0345     struct rds_ib_mr *ibmr;
0346     struct llist_node *clean_nodes;
0347     struct llist_node *clean_tail;
0348     LIST_HEAD(unmap_list);
0349     unsigned long unpinned = 0;
0350     unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
0351
0352     if (pool->pool_type == RDS_IB_MR_8K_POOL)
0353         rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
0354     else
0355         rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
0356
0357     if (ibmr_ret) {
0358         DEFINE_WAIT(wait);
0359         while (!mutex_trylock(&pool->flush_lock)) {
0360             ibmr = rds_ib_reuse_mr(pool);
0361             if (ibmr) {
0362                 *ibmr_ret = ibmr;
0363                 finish_wait(&pool->flush_wait, &wait);
0364                 goto out_nolock;
0365             }
0366
0367             prepare_to_wait(&pool->flush_wait, &wait,
0368                     TASK_UNINTERRUPTIBLE);
0369             if (llist_empty(&pool->clean_list))
0370                 schedule();
0371
0372             ibmr = rds_ib_reuse_mr(pool);
0373             if (ibmr) {
0374                 *ibmr_ret = ibmr;
0375                 finish_wait(&pool->flush_wait, &wait);
0376                 goto out_nolock;
0377             }
0378         }
0379         finish_wait(&pool->flush_wait, &wait);
0380     } else
0381         mutex_lock(&pool->flush_lock);
0382
0383     if (ibmr_ret) {
0384         ibmr = rds_ib_reuse_mr(pool);
0385         if (ibmr) {
0386             *ibmr_ret = ibmr;
0387             goto out;
0388         }
0389     }
0390
0391     /* Get the list of all MRs to be dropped. Ordering matters -
0392      * we want to put drop_list ahead of free_list.
0393      */
0394     dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
0395     dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
0396     if (free_all) {
0397         unsigned long flags;
0398
0399         spin_lock_irqsave(&pool->clean_lock, flags);
0400         llist_append_to_list(&pool->clean_list, &unmap_list);
0401         spin_unlock_irqrestore(&pool->clean_lock, flags);
0402     }
0403
0404     free_goal = rds_ib_flush_goal(pool, free_all);
0405
0406     if (list_empty(&unmap_list))
0407         goto out;
0408
0409     rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal);
0410
0411     if (!list_empty(&unmap_list)) {
0412         unsigned long flags;
0413
0414         list_to_llist_nodes(&unmap_list, &clean_nodes, &clean_tail);
0415         if (ibmr_ret) {
0416             *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
0417             clean_nodes = clean_nodes->next;
0418         }
0419         /* more than one entry in llist nodes */
0420         if (clean_nodes) {
0421             spin_lock_irqsave(&pool->clean_lock, flags);
0422             llist_add_batch(clean_nodes, clean_tail,
0423                     &pool->clean_list);
0424             spin_unlock_irqrestore(&pool->clean_lock, flags);
0425         }
0426     }
0427
0428     atomic_sub(unpinned, &pool->free_pinned);
0429     atomic_sub(dirty_to_clean, &pool->dirty_count);
0430     atomic_sub(nfreed, &pool->item_count);
0431
0432 out:
0433     mutex_unlock(&pool->flush_lock);
0434     if (waitqueue_active(&pool->flush_wait))
0435         wake_up(&pool->flush_wait);
0436 out_nolock:
0437     return 0;
0438 }
0439
0440 struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool)
0441 {
0442     struct rds_ib_mr *ibmr = NULL;
0443     int iter = 0;
0444
0445     while (1) {
0446         ibmr = rds_ib_reuse_mr(pool);
0447         if (ibmr)
0448             return ibmr;
0449
0450         if (atomic_inc_return(&pool->item_count) <= pool->max_items)
0451             break;
0452
0453         atomic_dec(&pool->item_count);
0454
0455         if (++iter > 2) {
0456             if (pool->pool_type == RDS_IB_MR_8K_POOL)
0457                 rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
0458             else
0459                 rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
0460             break;
0461         }
0462
0463         /* We do have some empty MRs. Flush them out. */
0464         if (pool->pool_type == RDS_IB_MR_8K_POOL)
0465             rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
0466         else
0467             rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
0468
0469         rds_ib_flush_mr_pool(pool, 0, &ibmr);
0470         if (ibmr)
0471             return ibmr;
0472     }
0473
0474     return NULL;
0475 }
0476
0477 static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
0478 {
0479     struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
0480
0481     rds_ib_flush_mr_pool(pool, 0, NULL);
0482 }
0483
0484 void rds_ib_free_mr(void *trans_private, int invalidate)
0485 {
0486     struct rds_ib_mr *ibmr = trans_private;
0487     struct rds_ib_mr_pool *pool = ibmr->pool;
0488     struct rds_ib_device *rds_ibdev = ibmr->device;
0489
0490     rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
0491
0492     if (ibmr->odp) {
0493         /* A MR created and marked as use_once. We use delayed work,
0494          * because there is a change that we are in interrupt and can't
0495          * call to ib_dereg_mr() directly.
0496          */
0497         INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
0498         queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
0499         return;
0500     }
0501
0502     /* Return it to the pool's free list */
0503     rds_ib_free_frmr_list(ibmr);
0504
0505     atomic_add(ibmr->sg_len, &pool->free_pinned);
0506     atomic_inc(&pool->dirty_count);
0507
0508     /* If we've pinned too many pages, request a flush */
0509     if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
0510         atomic_read(&pool->dirty_count) >= pool->max_items / 5)
0511         queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
0512
0513     if (invalidate) {
0514         if (likely(!in_interrupt())) {
0515             rds_ib_flush_mr_pool(pool, 0, NULL);
0516         } else {
0517             /* We get here if the user created a MR marked
0518              * as use_once and invalidate at the same time.
0519              */
0520             queue_delayed_work(rds_ib_mr_wq,
0521                        &pool->flush_worker, 10);
0522         }
0523     }
0524
0525     rds_ib_dev_put(rds_ibdev);
0526 }
0527
0528 void rds_ib_flush_mrs(void)
0529 {
0530     struct rds_ib_device *rds_ibdev;
0531
0532     down_read(&rds_ib_devices_lock);
0533     list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
0534         if (rds_ibdev->mr_8k_pool)
0535             rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
0536
0537         if (rds_ibdev->mr_1m_pool)
0538             rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
0539     }
0540     up_read(&rds_ib_devices_lock);
0541 }
0542
0543 u32 rds_ib_get_lkey(void *trans_private)
0544 {
0545     struct rds_ib_mr *ibmr = trans_private;
0546
0547     return ibmr->u.mr->lkey;
0548 }
0549
0550 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
0551             struct rds_sock *rs, u32 *key_ret,
0552             struct rds_connection *conn,
0553             u64 start, u64 length, int need_odp)
0554 {
0555     struct rds_ib_device *rds_ibdev;
0556     struct rds_ib_mr *ibmr = NULL;
0557     struct rds_ib_connection *ic = NULL;
0558     int ret;
0559
0560     rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]);
0561     if (!rds_ibdev) {
0562         ret = -ENODEV;
0563         goto out;
0564     }
0565
0566     if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
0567         u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
0568         int access_flags =
0569             (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
0570              IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
0571              IB_ACCESS_ON_DEMAND);
0572         struct ib_sge sge = {};
0573         struct ib_mr *ib_mr;
0574
0575         if (!rds_ibdev->odp_capable) {
0576             ret = -EOPNOTSUPP;
0577             goto out;
0578         }
0579
0580         ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
0581                        access_flags);
0582
0583         if (IS_ERR(ib_mr)) {
0584             rdsdebug("rds_ib_get_user_mr returned %d\n",
0585                  IS_ERR(ib_mr));
0586             ret = PTR_ERR(ib_mr);
0587             goto out;
0588         }
0589         if (key_ret)
0590             *key_ret = ib_mr->rkey;
0591
0592         ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
0593         if (!ibmr) {
0594             ib_dereg_mr(ib_mr);
0595             ret = -ENOMEM;
0596             goto out;
0597         }
0598         ibmr->u.mr = ib_mr;
0599         ibmr->odp = 1;
0600
0601         sge.addr = virt_addr;
0602         sge.length = length;
0603         sge.lkey = ib_mr->lkey;
0604
0605         ib_advise_mr(rds_ibdev->pd,
0606                  IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
0607                  IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
0608         return ibmr;
0609     }
0610
0611     if (conn)
0612         ic = conn->c_transport_data;
0613
0614     if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
0615         ret = -ENODEV;
0616         goto out;
0617     }
0618
0619     ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret);
0620     if (IS_ERR(ibmr)) {
0621         ret = PTR_ERR(ibmr);
0622         pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret);
0623     } else {
0624         return ibmr;
0625     }
0626
0627  out:
0628     if (rds_ibdev)
0629         rds_ib_dev_put(rds_ibdev);
0630
0631     return ERR_PTR(ret);
0632 }
0633
0634 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
0635 {
0636     cancel_delayed_work_sync(&pool->flush_worker);
0637     rds_ib_flush_mr_pool(pool, 1, NULL);
0638     WARN_ON(atomic_read(&pool->item_count));
0639     WARN_ON(atomic_read(&pool->free_pinned));
0640     kfree(pool);
0641 }
0642
0643 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
0644                          int pool_type)
0645 {
0646     struct rds_ib_mr_pool *pool;
0647
0648     pool = kzalloc(sizeof(*pool), GFP_KERNEL);
0649     if (!pool)
0650         return ERR_PTR(-ENOMEM);
0651
0652     pool->pool_type = pool_type;
0653     init_llist_head(&pool->free_list);
0654     init_llist_head(&pool->drop_list);
0655     init_llist_head(&pool->clean_list);
0656     spin_lock_init(&pool->clean_lock);
0657     mutex_init(&pool->flush_lock);
0658     init_waitqueue_head(&pool->flush_wait);
0659     INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
0660
0661     if (pool_type == RDS_IB_MR_1M_POOL) {
0662         /* +1 allows for unaligned MRs */
0663         pool->max_pages = RDS_MR_1M_MSG_SIZE + 1;
0664         pool->max_items = rds_ibdev->max_1m_mrs;
0665     } else {
0666         /* pool_type == RDS_IB_MR_8K_POOL */
0667         pool->max_pages = RDS_MR_8K_MSG_SIZE + 1;
0668         pool->max_items = rds_ibdev->max_8k_mrs;
0669     }
0670
0671     pool->max_free_pinned = pool->max_items * pool->max_pages / 4;
0672     pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4;
0673
0674     return pool;
0675 }
0676
0677 int rds_ib_mr_init(void)
0678 {
0679     rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0);
0680     if (!rds_ib_mr_wq)
0681         return -ENOMEM;
0682     return 0;
0683 }
0684
0685 /* By the time this is called all the IB devices should have been torn down and
0686  * had their pools freed.  As each pool is freed its work struct is waited on,
0687  * so the pool flushing work queue should be idle by the time we get here.
0688  */
0689 void rds_ib_mr_exit(void)
0690 {
0691     destroy_workqueue(rds_ib_mr_wq);
0692 }
0693
0694 static void rds_ib_odp_mr_worker(struct work_struct  *work)
0695 {
0696     struct rds_ib_mr *ibmr;
0697
0698     ibmr = container_of(work, struct rds_ib_mr, work.work);
0699     ib_dereg_mr(ibmr->u.mr);
0700     kfree(ibmr);
0701 }