LXR linux/drivers/infiniband/hw/qib/qib

   1/*
   2 * Copyright (c) 2006, 2007, 2009 QLogic Corporation. All rights reserved.
   3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        provided with the distribution.
  23 *
  24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34#include "qib.h"
  35
  36/**
  37 * qib_alloc_lkey - allocate an lkey
  38 * @mr: memory region that this lkey protects
  39 * @dma_region: 0->normal key, 1->restricted DMA key
  40 *
  41 * Returns 0 if successful, otherwise returns -errno.
  42 *
  43 * Increments mr reference count as required.
  44 *
  45 * Sets the lkey field mr for non-dma regions.
  46 *
  47 */
  48
  49int qib_alloc_lkey(struct qib_mregion *mr, int dma_region)
  50{
  51        unsigned long flags;
  52        u32 r;
  53        u32 n;
  54        int ret = 0;
  55        struct qib_ibdev *dev = to_idev(mr->pd->device);
  56        struct qib_lkey_table *rkt = &dev->lk_table;
  57
  58        spin_lock_irqsave(&rkt->lock, flags);
  59
  60        /* special case for dma_mr lkey == 0 */
  61        if (dma_region) {
  62                struct qib_mregion *tmr;
  63
  64                tmr = rcu_dereference(dev->dma_mr);
  65                if (!tmr) {
  66                        qib_get_mr(mr);
  67                        rcu_assign_pointer(dev->dma_mr, mr);
  68                        mr->lkey_published = 1;
  69                }
  70                goto success;
  71        }
  72
  73        /* Find the next available LKEY */
  74        r = rkt->next;
  75        n = r;
  76        for (;;) {
  77                if (rkt->table[r] == NULL)
  78                        break;
  79                r = (r + 1) & (rkt->max - 1);
  80                if (r == n)
  81                        goto bail;
  82        }
  83        rkt->next = (r + 1) & (rkt->max - 1);
  84        /*
  85         * Make sure lkey is never zero which is reserved to indicate an
  86         * unrestricted LKEY.
  87         */
  88        rkt->gen++;
  89        mr->lkey = (r << (32 - ib_qib_lkey_table_size)) |
  90                ((((1 << (24 - ib_qib_lkey_table_size)) - 1) & rkt->gen)
  91                 << 8);
  92        if (mr->lkey == 0) {
  93                mr->lkey |= 1 << 8;
  94                rkt->gen++;
  95        }
  96        qib_get_mr(mr);
  97        rcu_assign_pointer(rkt->table[r], mr);
  98        mr->lkey_published = 1;
  99success:
 100        spin_unlock_irqrestore(&rkt->lock, flags);
 101out:
 102        return ret;
 103bail:
 104        spin_unlock_irqrestore(&rkt->lock, flags);
 105        ret = -ENOMEM;
 106        goto out;
 107}
 108
 109/**
 110 * qib_free_lkey - free an lkey
 111 * @mr: mr to free from tables
 112 */
 113void qib_free_lkey(struct qib_mregion *mr)
 114{
 115        unsigned long flags;
 116        u32 lkey = mr->lkey;
 117        u32 r;
 118        struct qib_ibdev *dev = to_idev(mr->pd->device);
 119        struct qib_lkey_table *rkt = &dev->lk_table;
 120
 121        spin_lock_irqsave(&rkt->lock, flags);
 122        if (!mr->lkey_published)
 123                goto out;
 124        if (lkey == 0)
 125                rcu_assign_pointer(dev->dma_mr, NULL);
 126        else {
 127                r = lkey >> (32 - ib_qib_lkey_table_size);
 128                rcu_assign_pointer(rkt->table[r], NULL);
 129        }
 130        qib_put_mr(mr);
 131        mr->lkey_published = 0;
 132out:
 133        spin_unlock_irqrestore(&rkt->lock, flags);
 134}
 135
 136/**
 137 * qib_lkey_ok - check IB SGE for validity and initialize
 138 * @rkt: table containing lkey to check SGE against
 139 * @pd: protection domain
 140 * @isge: outgoing internal SGE
 141 * @sge: SGE to check
 142 * @acc: access flags
 143 *
 144 * Return 1 if valid and successful, otherwise returns 0.
 145 *
 146 * increments the reference count upon success
 147 *
 148 * Check the IB SGE for validity and initialize our internal version
 149 * of it.
 150 */
 151int qib_lkey_ok(struct qib_lkey_table *rkt, struct qib_pd *pd,
 152                struct qib_sge *isge, struct ib_sge *sge, int acc)
 153{
 154        struct qib_mregion *mr;
 155        unsigned n, m;
 156        size_t off;
 157
 158        /*
 159         * We use LKEY == zero for kernel virtual addresses
 160         * (see qib_get_dma_mr and qib_dma.c).
 161         */
 162        rcu_read_lock();
 163        if (sge->lkey == 0) {
 164                struct qib_ibdev *dev = to_idev(pd->ibpd.device);
 165
 166                if (pd->user)
 167                        goto bail;
 168                mr = rcu_dereference(dev->dma_mr);
 169                if (!mr)
 170                        goto bail;
 171                if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 172                        goto bail;
 173                rcu_read_unlock();
 174
 175                isge->mr = mr;
 176                isge->vaddr = (void *) sge->addr;
 177                isge->length = sge->length;
 178                isge->sge_length = sge->length;
 179                isge->m = 0;
 180                isge->n = 0;
 181                goto ok;
 182        }
 183        mr = rcu_dereference(
 184                rkt->table[(sge->lkey >> (32 - ib_qib_lkey_table_size))]);
 185        if (unlikely(!mr || mr->lkey != sge->lkey || mr->pd != &pd->ibpd))
 186                goto bail;
 187
 188        off = sge->addr - mr->user_base;
 189        if (unlikely(sge->addr < mr->user_base ||
 190                     off + sge->length > mr->length ||
 191                     (mr->access_flags & acc) != acc))
 192                goto bail;
 193        if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 194                goto bail;
 195        rcu_read_unlock();
 196
 197        off += mr->offset;
 198        if (mr->page_shift) {
 199                /*
 200                page sizes are uniform power of 2 so no loop is necessary
 201                entries_spanned_by_off is the number of times the loop below
 202                would have executed.
 203                */
 204                size_t entries_spanned_by_off;
 205
 206                entries_spanned_by_off = off >> mr->page_shift;
 207                off -= (entries_spanned_by_off << mr->page_shift);
 208                m = entries_spanned_by_off/QIB_SEGSZ;
 209                n = entries_spanned_by_off%QIB_SEGSZ;
 210        } else {
 211                m = 0;
 212                n = 0;
 213                while (off >= mr->map[m]->segs[n].length) {
 214                        off -= mr->map[m]->segs[n].length;
 215                        n++;
 216                        if (n >= QIB_SEGSZ) {
 217                                m++;
 218                                n = 0;
 219                        }
 220                }
 221        }
 222        isge->mr = mr;
 223        isge->vaddr = mr->map[m]->segs[n].vaddr + off;
 224        isge->length = mr->map[m]->segs[n].length - off;
 225        isge->sge_length = sge->length;
 226        isge->m = m;
 227        isge->n = n;
 228ok:
 229        return 1;
 230bail:
 231        rcu_read_unlock();
 232        return 0;
 233}
 234
 235/**
 236 * qib_rkey_ok - check the IB virtual address, length, and RKEY
 237 * @qp: qp for validation
 238 * @sge: SGE state
 239 * @len: length of data
 240 * @vaddr: virtual address to place data
 241 * @rkey: rkey to check
 242 * @acc: access flags
 243 *
 244 * Return 1 if successful, otherwise 0.
 245 *
 246 * increments the reference count upon success
 247 */
 248int qib_rkey_ok(struct qib_qp *qp, struct qib_sge *sge,
 249                u32 len, u64 vaddr, u32 rkey, int acc)
 250{
 251        struct qib_lkey_table *rkt = &to_idev(qp->ibqp.device)->lk_table;
 252        struct qib_mregion *mr;
 253        unsigned n, m;
 254        size_t off;
 255
 256        /*
 257         * We use RKEY == zero for kernel virtual addresses
 258         * (see qib_get_dma_mr and qib_dma.c).
 259         */
 260        rcu_read_lock();
 261        if (rkey == 0) {
 262                struct qib_pd *pd = to_ipd(qp->ibqp.pd);
 263                struct qib_ibdev *dev = to_idev(pd->ibpd.device);
 264
 265                if (pd->user)
 266                        goto bail;
 267                mr = rcu_dereference(dev->dma_mr);
 268                if (!mr)
 269                        goto bail;
 270                if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 271                        goto bail;
 272                rcu_read_unlock();
 273
 274                sge->mr = mr;
 275                sge->vaddr = (void *) vaddr;
 276                sge->length = len;
 277                sge->sge_length = len;
 278                sge->m = 0;
 279                sge->n = 0;
 280                goto ok;
 281        }
 282
 283        mr = rcu_dereference(
 284                rkt->table[(rkey >> (32 - ib_qib_lkey_table_size))]);
 285        if (unlikely(!mr || mr->lkey != rkey || qp->ibqp.pd != mr->pd))
 286                goto bail;
 287
 288        off = vaddr - mr->iova;
 289        if (unlikely(vaddr < mr->iova || off + len > mr->length ||
 290                     (mr->access_flags & acc) == 0))
 291                goto bail;
 292        if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 293                goto bail;
 294        rcu_read_unlock();
 295
 296        off += mr->offset;
 297        if (mr->page_shift) {
 298                /*
 299                page sizes are uniform power of 2 so no loop is necessary
 300                entries_spanned_by_off is the number of times the loop below
 301                would have executed.
 302                */
 303                size_t entries_spanned_by_off;
 304
 305                entries_spanned_by_off = off >> mr->page_shift;
 306                off -= (entries_spanned_by_off << mr->page_shift);
 307                m = entries_spanned_by_off/QIB_SEGSZ;
 308                n = entries_spanned_by_off%QIB_SEGSZ;
 309        } else {
 310                m = 0;
 311                n = 0;
 312                while (off >= mr->map[m]->segs[n].length) {
 313                        off -= mr->map[m]->segs[n].length;
 314                        n++;
 315                        if (n >= QIB_SEGSZ) {
 316                                m++;
 317                                n = 0;
 318                        }
 319                }
 320        }
 321        sge->mr = mr;
 322        sge->vaddr = mr->map[m]->segs[n].vaddr + off;
 323        sge->length = mr->map[m]->segs[n].length - off;
 324        sge->sge_length = len;
 325        sge->m = m;
 326        sge->n = n;
 327ok:
 328        return 1;
 329bail:
 330        rcu_read_unlock();
 331        return 0;
 332}
 333
 334/*
 335 * Initialize the memory region specified by the work reqeust.
 336 */
 337int qib_fast_reg_mr(struct qib_qp *qp, struct ib_send_wr *wr)
 338{
 339        struct qib_lkey_table *rkt = &to_idev(qp->ibqp.device)->lk_table;
 340        struct qib_pd *pd = to_ipd(qp->ibqp.pd);
 341        struct qib_mregion *mr;
 342        u32 rkey = wr->wr.fast_reg.rkey;
 343        unsigned i, n, m;
 344        int ret = -EINVAL;
 345        unsigned long flags;
 346        u64 *page_list;
 347        size_t ps;
 348
 349        spin_lock_irqsave(&rkt->lock, flags);
 350        if (pd->user || rkey == 0)
 351                goto bail;
 352
 353        mr = rcu_dereference_protected(
 354                rkt->table[(rkey >> (32 - ib_qib_lkey_table_size))],
 355                lockdep_is_held(&rkt->lock));
 356        if (unlikely(mr == NULL || qp->ibqp.pd != mr->pd))
 357                goto bail;
 358
 359        if (wr->wr.fast_reg.page_list_len > mr->max_segs)
 360                goto bail;
 361
 362        ps = 1UL << wr->wr.fast_reg.page_shift;
 363        if (wr->wr.fast_reg.length > ps * wr->wr.fast_reg.page_list_len)
 364                goto bail;
 365
 366        mr->user_base = wr->wr.fast_reg.iova_start;
 367        mr->iova = wr->wr.fast_reg.iova_start;
 368        mr->lkey = rkey;
 369        mr->length = wr->wr.fast_reg.length;
 370        mr->access_flags = wr->wr.fast_reg.access_flags;
 371        page_list = wr->wr.fast_reg.page_list->page_list;
 372        m = 0;
 373        n = 0;
 374        for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 375                mr->map[m]->segs[n].vaddr = (void *) page_list[i];
 376                mr->map[m]->segs[n].length = ps;
 377                if (++n == QIB_SEGSZ) {
 378                        m++;
 379                        n = 0;
 380                }
 381        }
 382
 383        ret = 0;
 384bail:
 385        spin_unlock_irqrestore(&rkt->lock, flags);
 386        return ret;
 387}
 388