drivers/infiniband/hw/qib/qib_mr.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008, 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 <rdma/ib_umem.h>
  35 #include <rdma/ib_smi.h>
  36
  37 #include "qib.h"
  38
  39 /* Fast memory region */
  40 struct qib_fmr {
  41         struct ib_fmr ibfmr;
  42         struct qib_mregion mr;        /* must be last */
  43 };
  44
  45 static inline struct qib_fmr *to_ifmr(struct ib_fmr *ibfmr)
  46 {
  47         return container_of(ibfmr, struct qib_fmr, ibfmr);
  48 }
  49
  50 static int init_qib_mregion(struct qib_mregion *mr, struct ib_pd *pd,
  51         int count)
  52 {
  53         int m, i = 0;
  54         int rval = 0;
  55
  56         m = (count + QIB_SEGSZ - 1) / QIB_SEGSZ;
  57         for (; i < m; i++) {
  58                 mr->map[i] = kzalloc(sizeof(*mr->map[0]), GFP_KERNEL);
  59                 if (!mr->map[i])
  60                         goto bail;
  61         }
  62         mr->mapsz = m;
  63         init_completion(&mr->comp);
  64         /* count returning the ptr to user */
  65         atomic_set(&mr->refcount, 1);
  66         mr->pd = pd;
  67         mr->max_segs = count;
  68 out:
  69         return rval;
  70 bail:
  71         while (i)
  72                 kfree(mr->map[--i]);
  73         rval = -ENOMEM;
  74         goto out;
  75 }
  76
  77 static void deinit_qib_mregion(struct qib_mregion *mr)
  78 {
  79         int i = mr->mapsz;
  80
  81         mr->mapsz = 0;
  82         while (i)
  83                 kfree(mr->map[--i]);
  84 }
  85
  86
  87 /**
  88  * qib_get_dma_mr - get a DMA memory region
  89  * @pd: protection domain for this memory region
  90  * @acc: access flags
  91  *
  92  * Returns the memory region on success, otherwise returns an errno.
  93  * Note that all DMA addresses should be created via the
  94  * struct ib_dma_mapping_ops functions (see qib_dma.c).
  95  */
  96 struct ib_mr *qib_get_dma_mr(struct ib_pd *pd, int acc)
  97 {
  98         struct qib_mr *mr = NULL;
  99         struct ib_mr *ret;
 100         int rval;
 101
 102         if (to_ipd(pd)->user) {
 103                 ret = ERR_PTR(-EPERM);
 104                 goto bail;
 105         }
 106
 107         mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 108         if (!mr) {
 109                 ret = ERR_PTR(-ENOMEM);
 110                 goto bail;
 111         }
 112
 113         rval = init_qib_mregion(&mr->mr, pd, 0);
 114         if (rval) {
 115                 ret = ERR_PTR(rval);
 116                 goto bail;
 117         }
 118
 119
 120         rval = qib_alloc_lkey(&mr->mr, 1);
 121         if (rval) {
 122                 ret = ERR_PTR(rval);
 123                 goto bail_mregion;
 124         }
 125
 126         mr->mr.access_flags = acc;
 127         ret = &mr->ibmr;
 128 done:
 129         return ret;
 130
 131 bail_mregion:
 132         deinit_qib_mregion(&mr->mr);
 133 bail:
 134         kfree(mr);
 135         goto done;
 136 }
 137
 138 static struct qib_mr *alloc_mr(int count, struct ib_pd *pd)
 139 {
 140         struct qib_mr *mr;
 141         int rval = -ENOMEM;
 142         int m;
 143
 144         /* Allocate struct plus pointers to first level page tables. */
 145         m = (count + QIB_SEGSZ - 1) / QIB_SEGSZ;
 146         mr = kzalloc(sizeof(*mr) + m * sizeof(mr->mr.map[0]), GFP_KERNEL);
 147         if (!mr)
 148                 goto bail;
 149
 150         rval = init_qib_mregion(&mr->mr, pd, count);
 151         if (rval)
 152                 goto bail;
 153         /*
 154          * ib_reg_phys_mr() will initialize mr->ibmr except for
 155          * lkey and rkey.
 156          */
 157         rval = qib_alloc_lkey(&mr->mr, 0);
 158         if (rval)
 159                 goto bail_mregion;
 160         mr->ibmr.lkey = mr->mr.lkey;
 161         mr->ibmr.rkey = mr->mr.lkey;
 162 done:
 163         return mr;
 164
 165 bail_mregion:
 166         deinit_qib_mregion(&mr->mr);
 167 bail:
 168         kfree(mr);
 169         mr = ERR_PTR(rval);
 170         goto done;
 171 }
 172
 173 /**
 174  * qib_reg_phys_mr - register a physical memory region
 175  * @pd: protection domain for this memory region
 176  * @buffer_list: pointer to the list of physical buffers to register
 177  * @num_phys_buf: the number of physical buffers to register
 178  * @iova_start: the starting address passed over IB which maps to this MR
 179  *
 180  * Returns the memory region on success, otherwise returns an errno.
 181  */
 182 struct ib_mr *qib_reg_phys_mr(struct ib_pd *pd,
 183                               struct ib_phys_buf *buffer_list,
 184                               int num_phys_buf, int acc, u64 *iova_start)
 185 {
 186         struct qib_mr *mr;
 187         int n, m, i;
 188         struct ib_mr *ret;
 189
 190         mr = alloc_mr(num_phys_buf, pd);
 191         if (IS_ERR(mr)) {
 192                 ret = (struct ib_mr *)mr;
 193                 goto bail;
 194         }
 195
 196         mr->mr.user_base = *iova_start;
 197         mr->mr.iova = *iova_start;
 198         mr->mr.access_flags = acc;
 199
 200         m = 0;
 201         n = 0;
 202         for (i = 0; i < num_phys_buf; i++) {
 203                 mr->mr.map[m]->segs[n].vaddr = (void *) buffer_list[i].addr;
 204                 mr->mr.map[m]->segs[n].length = buffer_list[i].size;
 205                 mr->mr.length += buffer_list[i].size;
 206                 n++;
 207                 if (n == QIB_SEGSZ) {
 208                         m++;
 209                         n = 0;
 210                 }
 211         }
 212
 213         ret = &mr->ibmr;
 214
 215 bail:
 216         return ret;
 217 }
 218
 219 /**
 220  * qib_reg_user_mr - register a userspace memory region
 221  * @pd: protection domain for this memory region
 222  * @start: starting userspace address
 223  * @length: length of region to register
 224  * @mr_access_flags: access flags for this memory region
 225  * @udata: unused by the QLogic_IB driver
 226  *
 227  * Returns the memory region on success, otherwise returns an errno.
 228  */
 229 struct ib_mr *qib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 230                               u64 virt_addr, int mr_access_flags,
 231                               struct ib_udata *udata)
 232 {
 233         struct qib_mr *mr;
 234         struct ib_umem *umem;
 235         struct scatterlist *sg;
 236         int n, m, entry;
 237         struct ib_mr *ret;
 238
 239         if (length == 0) {
 240                 ret = ERR_PTR(-EINVAL);
 241                 goto bail;
 242         }
 243
 244         umem = ib_umem_get(pd->uobject->context, start, length,
 245                            mr_access_flags, 0);
 246         if (IS_ERR(umem))
 247                 return (void *) umem;
 248
 249         n = umem->nmap;
 250
 251         mr = alloc_mr(n, pd);
 252         if (IS_ERR(mr)) {
 253                 ret = (struct ib_mr *)mr;
 254                 ib_umem_release(umem);
 255                 goto bail;
 256         }
 257
 258         mr->mr.user_base = start;
 259         mr->mr.iova = virt_addr;
 260         mr->mr.length = length;
 261         mr->mr.offset = ib_umem_offset(umem);
 262         mr->mr.access_flags = mr_access_flags;
 263         mr->umem = umem;
 264
 265         if (is_power_of_2(umem->page_size))
 266                 mr->mr.page_shift = ilog2(umem->page_size);
 267         m = 0;
 268         n = 0;
 269         for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
 270                         void *vaddr;
 271
 272                         vaddr = page_address(sg_page(sg));
 273                         if (!vaddr) {
 274                                 ret = ERR_PTR(-EINVAL);
 275                                 goto bail;
 276                         }
 277                         mr->mr.map[m]->segs[n].vaddr = vaddr;
 278                         mr->mr.map[m]->segs[n].length = umem->page_size;
 279                         n++;
 280                         if (n == QIB_SEGSZ) {
 281                                 m++;
 282                                 n = 0;
 283                         }
 284         }
 285         ret = &mr->ibmr;
 286
 287 bail:
 288         return ret;
 289 }
 290
 291 /**
 292  * qib_dereg_mr - unregister and free a memory region
 293  * @ibmr: the memory region to free
 294  *
 295  * Returns 0 on success.
 296  *
 297  * Note that this is called to free MRs created by qib_get_dma_mr()
 298  * or qib_reg_user_mr().
 299  */
 300 int qib_dereg_mr(struct ib_mr *ibmr)
 301 {
 302         struct qib_mr *mr = to_imr(ibmr);
 303         int ret = 0;
 304         unsigned long timeout;
 305
 306         qib_free_lkey(&mr->mr);
 307
 308         qib_put_mr(&mr->mr); /* will set completion if last */
 309         timeout = wait_for_completion_timeout(&mr->mr.comp,
 310                 5 * HZ);
 311         if (!timeout) {
 312                 qib_get_mr(&mr->mr);
 313                 ret = -EBUSY;
 314                 goto out;
 315         }
 316         deinit_qib_mregion(&mr->mr);
 317         if (mr->umem)
 318                 ib_umem_release(mr->umem);
 319         kfree(mr);
 320 out:
 321         return ret;
 322 }
 323
 324 /*
 325  * Allocate a memory region usable with the
 326  * IB_WR_FAST_REG_MR send work request.
 327  *
 328  * Return the memory region on success, otherwise return an errno.
 329  */
 330 struct ib_mr *qib_alloc_mr(struct ib_pd *pd,
 331                            enum ib_mr_type mr_type,
 332                            u32 max_num_sg)
 333 {
 334         struct qib_mr *mr;
 335
 336         if (mr_type != IB_MR_TYPE_MEM_REG)
 337                 return ERR_PTR(-EINVAL);
 338
 339         mr = alloc_mr(max_num_sg, pd);
 340         if (IS_ERR(mr))
 341                 return (struct ib_mr *)mr;
 342
 343         return &mr->ibmr;
 344 }
 345
 346 struct ib_fast_reg_page_list *
 347 qib_alloc_fast_reg_page_list(struct ib_device *ibdev, int page_list_len)
 348 {
 349         unsigned size = page_list_len * sizeof(u64);
 350         struct ib_fast_reg_page_list *pl;
 351
 352         if (size > PAGE_SIZE)
 353                 return ERR_PTR(-EINVAL);
 354
 355         pl = kzalloc(sizeof(*pl), GFP_KERNEL);
 356         if (!pl)
 357                 return ERR_PTR(-ENOMEM);
 358
 359         pl->page_list = kzalloc(size, GFP_KERNEL);
 360         if (!pl->page_list)
 361                 goto err_free;
 362
 363         return pl;
 364
 365 err_free:
 366         kfree(pl);
 367         return ERR_PTR(-ENOMEM);
 368 }
 369
 370 void qib_free_fast_reg_page_list(struct ib_fast_reg_page_list *pl)
 371 {
 372         kfree(pl->page_list);
 373         kfree(pl);
 374 }
 375
 376 /**
 377  * qib_alloc_fmr - allocate a fast memory region
 378  * @pd: the protection domain for this memory region
 379  * @mr_access_flags: access flags for this memory region
 380  * @fmr_attr: fast memory region attributes
 381  *
 382  * Returns the memory region on success, otherwise returns an errno.
 383  */
 384 struct ib_fmr *qib_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
 385                              struct ib_fmr_attr *fmr_attr)
 386 {
 387         struct qib_fmr *fmr;
 388         int m;
 389         struct ib_fmr *ret;
 390         int rval = -ENOMEM;
 391
 392         /* Allocate struct plus pointers to first level page tables. */
 393         m = (fmr_attr->max_pages + QIB_SEGSZ - 1) / QIB_SEGSZ;
 394         fmr = kzalloc(sizeof(*fmr) + m * sizeof(fmr->mr.map[0]), GFP_KERNEL);
 395         if (!fmr)
 396                 goto bail;
 397
 398         rval = init_qib_mregion(&fmr->mr, pd, fmr_attr->max_pages);
 399         if (rval)
 400                 goto bail;
 401
 402         /*
 403          * ib_alloc_fmr() will initialize fmr->ibfmr except for lkey &
 404          * rkey.
 405          */
 406         rval = qib_alloc_lkey(&fmr->mr, 0);
 407         if (rval)
 408                 goto bail_mregion;
 409         fmr->ibfmr.rkey = fmr->mr.lkey;
 410         fmr->ibfmr.lkey = fmr->mr.lkey;
 411         /*
 412          * Resources are allocated but no valid mapping (RKEY can't be
 413          * used).
 414          */
 415         fmr->mr.access_flags = mr_access_flags;
 416         fmr->mr.max_segs = fmr_attr->max_pages;
 417         fmr->mr.page_shift = fmr_attr->page_shift;
 418
 419         ret = &fmr->ibfmr;
 420 done:
 421         return ret;
 422
 423 bail_mregion:
 424         deinit_qib_mregion(&fmr->mr);
 425 bail:
 426         kfree(fmr);
 427         ret = ERR_PTR(rval);
 428         goto done;
 429 }
 430
 431 /**
 432  * qib_map_phys_fmr - set up a fast memory region
 433  * @ibmfr: the fast memory region to set up
 434  * @page_list: the list of pages to associate with the fast memory region
 435  * @list_len: the number of pages to associate with the fast memory region
 436  * @iova: the virtual address of the start of the fast memory region
 437  *
 438  * This may be called from interrupt context.
 439  */
 440
 441 int qib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list,
 442                      int list_len, u64 iova)
 443 {
 444         struct qib_fmr *fmr = to_ifmr(ibfmr);
 445         struct qib_lkey_table *rkt;
 446         unsigned long flags;
 447         int m, n, i;
 448         u32 ps;
 449         int ret;
 450
 451         i = atomic_read(&fmr->mr.refcount);
 452         if (i > 2)
 453                 return -EBUSY;
 454
 455         if (list_len > fmr->mr.max_segs) {
 456                 ret = -EINVAL;
 457                 goto bail;
 458         }
 459         rkt = &to_idev(ibfmr->device)->lk_table;
 460         spin_lock_irqsave(&rkt->lock, flags);
 461         fmr->mr.user_base = iova;
 462         fmr->mr.iova = iova;
 463         ps = 1 << fmr->mr.page_shift;
 464         fmr->mr.length = list_len * ps;
 465         m = 0;
 466         n = 0;
 467         for (i = 0; i < list_len; i++) {
 468                 fmr->mr.map[m]->segs[n].vaddr = (void *) page_list[i];
 469                 fmr->mr.map[m]->segs[n].length = ps;
 470                 if (++n == QIB_SEGSZ) {
 471                         m++;
 472                         n = 0;
 473                 }
 474         }
 475         spin_unlock_irqrestore(&rkt->lock, flags);
 476         ret = 0;
 477
 478 bail:
 479         return ret;
 480 }
 481
 482 /**
 483  * qib_unmap_fmr - unmap fast memory regions
 484  * @fmr_list: the list of fast memory regions to unmap
 485  *
 486  * Returns 0 on success.
 487  */
 488 int qib_unmap_fmr(struct list_head *fmr_list)
 489 {
 490         struct qib_fmr *fmr;
 491         struct qib_lkey_table *rkt;
 492         unsigned long flags;
 493
 494         list_for_each_entry(fmr, fmr_list, ibfmr.list) {
 495                 rkt = &to_idev(fmr->ibfmr.device)->lk_table;
 496                 spin_lock_irqsave(&rkt->lock, flags);
 497                 fmr->mr.user_base = 0;
 498                 fmr->mr.iova = 0;
 499                 fmr->mr.length = 0;
 500                 spin_unlock_irqrestore(&rkt->lock, flags);
 501         }
 502         return 0;
 503 }
 504
 505 /**
 506  * qib_dealloc_fmr - deallocate a fast memory region
 507  * @ibfmr: the fast memory region to deallocate
 508  *
 509  * Returns 0 on success.
 510  */
 511 int qib_dealloc_fmr(struct ib_fmr *ibfmr)
 512 {
 513         struct qib_fmr *fmr = to_ifmr(ibfmr);
 514         int ret = 0;
 515         unsigned long timeout;
 516
 517         qib_free_lkey(&fmr->mr);
 518         qib_put_mr(&fmr->mr); /* will set completion if last */
 519         timeout = wait_for_completion_timeout(&fmr->mr.comp,
 520                 5 * HZ);
 521         if (!timeout) {
 522                 qib_get_mr(&fmr->mr);
 523                 ret = -EBUSY;
 524                 goto out;
 525         }
 526         deinit_qib_mregion(&fmr->mr);
 527         kfree(fmr);
 528 out:
 529         return ret;
 530 }
 531
 532 void mr_rcu_callback(struct rcu_head *list)
 533 {
 534         struct qib_mregion *mr = container_of(list, struct qib_mregion, list);
 535
 536         complete(&mr->comp);
 537 }