net/rds/ib_frmr.c

   1 /*
   2  * Copyright (c) 2016 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 "ib_mr.h"
  34
  35 static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev,
  36                                            int npages)
  37 {
  38         struct rds_ib_mr_pool *pool;
  39         struct rds_ib_mr *ibmr = NULL;
  40         struct rds_ib_frmr *frmr;
  41         int err = 0;
  42
  43         if (npages <= RDS_MR_8K_MSG_SIZE)
  44                 pool = rds_ibdev->mr_8k_pool;
  45         else
  46                 pool = rds_ibdev->mr_1m_pool;
  47
  48         ibmr = rds_ib_try_reuse_ibmr(pool);
  49         if (ibmr)
  50                 return ibmr;
  51
  52         ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL,
  53                             rdsibdev_to_node(rds_ibdev));
  54         if (!ibmr) {
  55                 err = -ENOMEM;
  56                 goto out_no_cigar;
  57         }
  58
  59         frmr = &ibmr->u.frmr;
  60         frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG,
  61                          pool->fmr_attr.max_pages);
  62         if (IS_ERR(frmr->mr)) {
  63                 pr_warn("RDS/IB: %s failed to allocate MR", __func__);
  64                 err = PTR_ERR(frmr->mr);
  65                 goto out_no_cigar;
  66         }
  67
  68         ibmr->pool = pool;
  69         if (pool->pool_type == RDS_IB_MR_8K_POOL)
  70                 rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
  71         else
  72                 rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
  73
  74         if (atomic_read(&pool->item_count) > pool->max_items_soft)
  75                 pool->max_items_soft = pool->max_items;
  76
  77         frmr->fr_state = FRMR_IS_FREE;
  78         return ibmr;
  79
  80 out_no_cigar:
  81         kfree(ibmr);
  82         atomic_dec(&pool->item_count);
  83         return ERR_PTR(err);
  84 }
  85
  86 static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop)
  87 {
  88         struct rds_ib_mr_pool *pool = ibmr->pool;
  89
  90         if (drop)
  91                 llist_add(&ibmr->llnode, &pool->drop_list);
  92         else
  93                 llist_add(&ibmr->llnode, &pool->free_list);
  94         atomic_add(ibmr->sg_len, &pool->free_pinned);
  95         atomic_inc(&pool->dirty_count);
  96
  97         /* If we've pinned too many pages, request a flush */
  98         if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
  99             atomic_read(&pool->dirty_count) >= pool->max_items / 5)
 100                 queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10);
 101 }
 102
 103 static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr)
 104 {
 105         struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 106         struct ib_send_wr *failed_wr;
 107         struct ib_reg_wr reg_wr;
 108         int ret, off = 0;
 109
 110         while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) {
 111                 atomic_inc(&ibmr->ic->i_fastreg_wrs);
 112                 cpu_relax();
 113         }
 114
 115         ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_dma_len,
 116                                 &off, PAGE_SIZE);
 117         if (unlikely(ret != ibmr->sg_dma_len))
 118                 return ret < 0 ? ret : -EINVAL;
 119
 120         /* Perform a WR for the fast_reg_mr. Each individual page
 121          * in the sg list is added to the fast reg page list and placed
 122          * inside the fast_reg_mr WR.  The key used is a rolling 8bit
 123          * counter, which should guarantee uniqueness.
 124          */
 125         ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++);
 126         frmr->fr_state = FRMR_IS_INUSE;
 127
 128         memset(&reg_wr, 0, sizeof(reg_wr));
 129         reg_wr.wr.wr_id = (unsigned long)(void *)ibmr;
 130         reg_wr.wr.opcode = IB_WR_REG_MR;
 131         reg_wr.wr.num_sge = 0;
 132         reg_wr.mr = frmr->mr;
 133         reg_wr.key = frmr->mr->rkey;
 134         reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 135                         IB_ACCESS_REMOTE_READ |
 136                         IB_ACCESS_REMOTE_WRITE;
 137         reg_wr.wr.send_flags = IB_SEND_SIGNALED;
 138
 139         failed_wr = &reg_wr.wr;
 140         ret = ib_post_send(ibmr->ic->i_cm_id->qp, &reg_wr.wr, &failed_wr);
 141         WARN_ON(failed_wr != &reg_wr.wr);
 142         if (unlikely(ret)) {
 143                 /* Failure here can be because of -ENOMEM as well */
 144                 frmr->fr_state = FRMR_IS_STALE;
 145                 atomic_inc(&ibmr->ic->i_fastreg_wrs);
 146                 if (printk_ratelimit())
 147                         pr_warn("RDS/IB: %s returned error(%d)\n",
 148                                 __func__, ret);
 149         }
 150         return ret;
 151 }
 152
 153 static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev,
 154                            struct rds_ib_mr_pool *pool,
 155                            struct rds_ib_mr *ibmr,
 156                            struct scatterlist *sg, unsigned int sg_len)
 157 {
 158         struct ib_device *dev = rds_ibdev->dev;
 159         struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 160         int i;
 161         u32 len;
 162         int ret = 0;
 163
 164         /* We want to teardown old ibmr values here and fill it up with
 165          * new sg values
 166          */
 167         rds_ib_teardown_mr(ibmr);
 168
 169         ibmr->sg = sg;
 170         ibmr->sg_len = sg_len;
 171         ibmr->sg_dma_len = 0;
 172         frmr->sg_byte_len = 0;
 173         WARN_ON(ibmr->sg_dma_len);
 174         ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len,
 175                                          DMA_BIDIRECTIONAL);
 176         if (unlikely(!ibmr->sg_dma_len)) {
 177                 pr_warn("RDS/IB: %s failed!\n", __func__);
 178                 return -EBUSY;
 179         }
 180
 181         frmr->sg_byte_len = 0;
 182         frmr->dma_npages = 0;
 183         len = 0;
 184
 185         ret = -EINVAL;
 186         for (i = 0; i < ibmr->sg_dma_len; ++i) {
 187                 unsigned int dma_len = ib_sg_dma_len(dev, &ibmr->sg[i]);
 188                 u64 dma_addr = ib_sg_dma_address(dev, &ibmr->sg[i]);
 189
 190                 frmr->sg_byte_len += dma_len;
 191                 if (dma_addr & ~PAGE_MASK) {
 192                         if (i > 0)
 193                                 goto out_unmap;
 194                         else
 195                                 ++frmr->dma_npages;
 196                 }
 197
 198                 if ((dma_addr + dma_len) & ~PAGE_MASK) {
 199                         if (i < ibmr->sg_dma_len - 1)
 200                                 goto out_unmap;
 201                         else
 202                                 ++frmr->dma_npages;
 203                 }
 204
 205                 len += dma_len;
 206         }
 207         frmr->dma_npages += len >> PAGE_SHIFT;
 208
 209         if (frmr->dma_npages > ibmr->pool->fmr_attr.max_pages) {
 210                 ret = -EMSGSIZE;
 211                 goto out_unmap;
 212         }
 213
 214         ret = rds_ib_post_reg_frmr(ibmr);
 215         if (ret)
 216                 goto out_unmap;
 217
 218         if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
 219                 rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
 220         else
 221                 rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
 222
 223         return ret;
 224
 225 out_unmap:
 226         ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len,
 227                         DMA_BIDIRECTIONAL);
 228         ibmr->sg_dma_len = 0;
 229         return ret;
 230 }
 231
 232 static int rds_ib_post_inv(struct rds_ib_mr *ibmr)
 233 {
 234         struct ib_send_wr *s_wr, *failed_wr;
 235         struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 236         struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id;
 237         int ret = -EINVAL;
 238
 239         if (!i_cm_id || !i_cm_id->qp || !frmr->mr)
 240                 goto out;
 241
 242         if (frmr->fr_state != FRMR_IS_INUSE)
 243                 goto out;
 244
 245         while (atomic_dec_return(&ibmr->ic->i_fastunreg_wrs) <= 0) {
 246                 atomic_inc(&ibmr->ic->i_fastunreg_wrs);
 247                 cpu_relax();
 248         }
 249
 250         frmr->fr_inv = true;
 251         s_wr = &frmr->fr_wr;
 252
 253         memset(s_wr, 0, sizeof(*s_wr));
 254         s_wr->wr_id = (unsigned long)(void *)ibmr;
 255         s_wr->opcode = IB_WR_LOCAL_INV;
 256         s_wr->ex.invalidate_rkey = frmr->mr->rkey;
 257         s_wr->send_flags = IB_SEND_SIGNALED;
 258
 259         failed_wr = s_wr;
 260         ret = ib_post_send(i_cm_id->qp, s_wr, &failed_wr);
 261         WARN_ON(failed_wr != s_wr);
 262         if (unlikely(ret)) {
 263                 frmr->fr_state = FRMR_IS_STALE;
 264                 frmr->fr_inv = false;
 265                 atomic_inc(&ibmr->ic->i_fastunreg_wrs);
 266                 pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret);
 267                 goto out;
 268         }
 269 out:
 270         return ret;
 271 }
 272
 273 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
 274 {
 275         struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id;
 276         struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 277
 278         if (wc->status != IB_WC_SUCCESS) {
 279                 frmr->fr_state = FRMR_IS_STALE;
 280                 if (rds_conn_up(ic->conn))
 281                         rds_ib_conn_error(ic->conn,
 282                                           "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n",
 283                                           &ic->conn->c_laddr,
 284                                           &ic->conn->c_faddr,
 285                                           wc->status,
 286                                           ib_wc_status_msg(wc->status),
 287                                           wc->vendor_err);
 288         }
 289
 290         if (frmr->fr_inv) {
 291                 frmr->fr_state = FRMR_IS_FREE;
 292                 frmr->fr_inv = false;
 293                 atomic_inc(&ic->i_fastreg_wrs);
 294         } else {
 295                 atomic_inc(&ic->i_fastunreg_wrs);
 296         }
 297 }
 298
 299 void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed,
 300                        unsigned long *unpinned, unsigned int goal)
 301 {
 302         struct rds_ib_mr *ibmr, *next;
 303         struct rds_ib_frmr *frmr;
 304         int ret = 0;
 305         unsigned int freed = *nfreed;
 306
 307         /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
 308         list_for_each_entry(ibmr, list, unmap_list) {
 309                 if (ibmr->sg_dma_len)
 310                         ret |= rds_ib_post_inv(ibmr);
 311         }
 312         if (ret)
 313                 pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret);
 314
 315         /* Now we can destroy the DMA mapping and unpin any pages */
 316         list_for_each_entry_safe(ibmr, next, list, unmap_list) {
 317                 *unpinned += ibmr->sg_len;
 318                 frmr = &ibmr->u.frmr;
 319                 __rds_ib_teardown_mr(ibmr);
 320                 if (freed < goal || frmr->fr_state == FRMR_IS_STALE) {
 321                         /* Don't de-allocate if the MR is not free yet */
 322                         if (frmr->fr_state == FRMR_IS_INUSE)
 323                                 continue;
 324
 325                         if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
 326                                 rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
 327                         else
 328                                 rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
 329                         list_del(&ibmr->unmap_list);
 330                         if (frmr->mr)
 331                                 ib_dereg_mr(frmr->mr);
 332                         kfree(ibmr);
 333                         freed++;
 334                 }
 335         }
 336         *nfreed = freed;
 337 }
 338
 339 struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev,
 340                                   struct rds_ib_connection *ic,
 341                                   struct scatterlist *sg,
 342                                   unsigned long nents, u32 *key)
 343 {
 344         struct rds_ib_mr *ibmr = NULL;
 345         struct rds_ib_frmr *frmr;
 346         int ret;
 347
 348         do {
 349                 if (ibmr)
 350                         rds_ib_free_frmr(ibmr, true);
 351                 ibmr = rds_ib_alloc_frmr(rds_ibdev, nents);
 352                 if (IS_ERR(ibmr))
 353                         return ibmr;
 354                 frmr = &ibmr->u.frmr;
 355         } while (frmr->fr_state != FRMR_IS_FREE);
 356
 357         ibmr->ic = ic;
 358         ibmr->device = rds_ibdev;
 359         ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents);
 360         if (ret == 0) {
 361                 *key = frmr->mr->rkey;
 362         } else {
 363                 rds_ib_free_frmr(ibmr, false);
 364                 ibmr = ERR_PTR(ret);
 365         }
 366
 367         return ibmr;
 368 }
 369
 370 void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr)
 371 {
 372         struct rds_ib_mr_pool *pool = ibmr->pool;
 373         struct rds_ib_frmr *frmr = &ibmr->u.frmr;
 374
 375         if (frmr->fr_state == FRMR_IS_STALE)
 376                 llist_add(&ibmr->llnode, &pool->drop_list);
 377         else
 378                 llist_add(&ibmr->llnode, &pool->free_list);
 379 }