drivers/infiniband/hw/qib/qib_keys.c

   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
  49 int 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;
  99 success:
 100         spin_unlock_irqrestore(&rkt->lock, flags);
 101 out:
 102         return ret;
 103 bail:
 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  */
 113 void 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;
 132 out:
 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  */
 151 int 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->iova || off + sge->length > mr->length ||
 190                      (mr->access_flags & acc) == 0))
 191                 goto bail;
 192         if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 193                 goto bail;
 194         rcu_read_unlock();
 195
 196         off += mr->offset;
 197         if (mr->page_shift) {
 198                 /*
 199                 page sizes are uniform power of 2 so no loop is necessary
 200                 entries_spanned_by_off is the number of times the loop below
 201                 would have executed.
 202                 */
 203                 size_t entries_spanned_by_off;
 204
 205                 entries_spanned_by_off = off >> mr->page_shift;
 206                 off -= (entries_spanned_by_off << mr->page_shift);
 207                 m = entries_spanned_by_off/QIB_SEGSZ;
 208                 n = entries_spanned_by_off%QIB_SEGSZ;
 209         } else {
 210                 m = 0;
 211                 n = 0;
 212                 while (off >= mr->map[m]->segs[n].length) {
 213                         off -= mr->map[m]->segs[n].length;
 214                         n++;
 215                         if (n >= QIB_SEGSZ) {
 216                                 m++;
 217                                 n = 0;
 218                         }
 219                 }
 220         }
 221         isge->mr = mr;
 222         isge->vaddr = mr->map[m]->segs[n].vaddr + off;
 223         isge->length = mr->map[m]->segs[n].length - off;
 224         isge->sge_length = sge->length;
 225         isge->m = m;
 226         isge->n = n;
 227 ok:
 228         return 1;
 229 bail:
 230         rcu_read_unlock();
 231         return 0;
 232 }
 233
 234 /**
 235  * qib_rkey_ok - check the IB virtual address, length, and RKEY
 236  * @qp: qp for validation
 237  * @sge: SGE state
 238  * @len: length of data
 239  * @vaddr: virtual address to place data
 240  * @rkey: rkey to check
 241  * @acc: access flags
 242  *
 243  * Return 1 if successful, otherwise 0.
 244  *
 245  * increments the reference count upon success
 246  */
 247 int qib_rkey_ok(struct qib_qp *qp, struct qib_sge *sge,
 248                 u32 len, u64 vaddr, u32 rkey, int acc)
 249 {
 250         struct qib_lkey_table *rkt = &to_idev(qp->ibqp.device)->lk_table;
 251         struct qib_mregion *mr;
 252         unsigned n, m;
 253         size_t off;
 254
 255         /*
 256          * We use RKEY == zero for kernel virtual addresses
 257          * (see qib_get_dma_mr and qib_dma.c).
 258          */
 259         rcu_read_lock();
 260         if (rkey == 0) {
 261                 struct qib_pd *pd = to_ipd(qp->ibqp.pd);
 262                 struct qib_ibdev *dev = to_idev(pd->ibpd.device);
 263
 264                 if (pd->user)
 265                         goto bail;
 266                 mr = rcu_dereference(dev->dma_mr);
 267                 if (!mr)
 268                         goto bail;
 269                 if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 270                         goto bail;
 271                 rcu_read_unlock();
 272
 273                 sge->mr = mr;
 274                 sge->vaddr = (void *) vaddr;
 275                 sge->length = len;
 276                 sge->sge_length = len;
 277                 sge->m = 0;
 278                 sge->n = 0;
 279                 goto ok;
 280         }
 281
 282         mr = rcu_dereference(
 283                 rkt->table[(rkey >> (32 - ib_qib_lkey_table_size))]);
 284         if (unlikely(!mr || mr->lkey != rkey || qp->ibqp.pd != mr->pd))
 285                 goto bail;
 286
 287         off = vaddr - mr->iova;
 288         if (unlikely(vaddr < mr->iova || off + len > mr->length ||
 289                      (mr->access_flags & acc) == 0))
 290                 goto bail;
 291         if (unlikely(!atomic_inc_not_zero(&mr->refcount)))
 292                 goto bail;
 293         rcu_read_unlock();
 294
 295         off += mr->offset;
 296         if (mr->page_shift) {
 297                 /*
 298                 page sizes are uniform power of 2 so no loop is necessary
 299                 entries_spanned_by_off is the number of times the loop below
 300                 would have executed.
 301                 */
 302                 size_t entries_spanned_by_off;
 303
 304                 entries_spanned_by_off = off >> mr->page_shift;
 305                 off -= (entries_spanned_by_off << mr->page_shift);
 306                 m = entries_spanned_by_off/QIB_SEGSZ;
 307                 n = entries_spanned_by_off%QIB_SEGSZ;
 308         } else {
 309                 m = 0;
 310                 n = 0;
 311                 while (off >= mr->map[m]->segs[n].length) {
 312                         off -= mr->map[m]->segs[n].length;
 313                         n++;
 314                         if (n >= QIB_SEGSZ) {
 315                                 m++;
 316                                 n = 0;
 317                         }
 318                 }
 319         }
 320         sge->mr = mr;
 321         sge->vaddr = mr->map[m]->segs[n].vaddr + off;
 322         sge->length = mr->map[m]->segs[n].length - off;
 323         sge->sge_length = len;
 324         sge->m = m;
 325         sge->n = n;
 326 ok:
 327         return 1;
 328 bail:
 329         rcu_read_unlock();
 330         return 0;
 331 }
 332
 333 /*
 334  * Initialize the memory region specified by the work reqeust.
 335  */
 336 int qib_fast_reg_mr(struct qib_qp *qp, struct ib_send_wr *wr)
 337 {
 338         struct qib_lkey_table *rkt = &to_idev(qp->ibqp.device)->lk_table;
 339         struct qib_pd *pd = to_ipd(qp->ibqp.pd);
 340         struct qib_mregion *mr;
 341         u32 rkey = wr->wr.fast_reg.rkey;
 342         unsigned i, n, m;
 343         int ret = -EINVAL;
 344         unsigned long flags;
 345         u64 *page_list;
 346         size_t ps;
 347
 348         spin_lock_irqsave(&rkt->lock, flags);
 349         if (pd->user || rkey == 0)
 350                 goto bail;
 351
 352         mr = rcu_dereference_protected(
 353                 rkt->table[(rkey >> (32 - ib_qib_lkey_table_size))],
 354                 lockdep_is_held(&rkt->lock));
 355         if (unlikely(mr == NULL || qp->ibqp.pd != mr->pd))
 356                 goto bail;
 357
 358         if (wr->wr.fast_reg.page_list_len > mr->max_segs)
 359                 goto bail;
 360
 361         ps = 1UL << wr->wr.fast_reg.page_shift;
 362         if (wr->wr.fast_reg.length > ps * wr->wr.fast_reg.page_list_len)
 363                 goto bail;
 364
 365         mr->user_base = wr->wr.fast_reg.iova_start;
 366         mr->iova = wr->wr.fast_reg.iova_start;
 367         mr->lkey = rkey;
 368         mr->length = wr->wr.fast_reg.length;
 369         mr->access_flags = wr->wr.fast_reg.access_flags;
 370         page_list = wr->wr.fast_reg.page_list->page_list;
 371         m = 0;
 372         n = 0;
 373         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 374                 mr->map[m]->segs[n].vaddr = (void *) page_list[i];
 375                 mr->map[m]->segs[n].length = ps;
 376                 if (++n == QIB_SEGSZ) {
 377                         m++;
 378                         n = 0;
 379                 }
 380         }
 381
 382         ret = 0;
 383 bail:
 384         spin_unlock_irqrestore(&rkt->lock, flags);
 385         return ret;
 386 }