drivers/staging/lustre/lnet/lnet/lib-md.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/lnet/lib-md.c
  37  *
  38  * Memory Descriptor management routines
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_LNET
  42
  43 #include "../../include/linux/lnet/lib-lnet.h"
  44
  45 /* must be called with lnet_res_lock held */
  46 void
  47 lnet_md_unlink(lnet_libmd_t *md)
  48 {
  49         if (!(md->md_flags & LNET_MD_FLAG_ZOMBIE)) {
  50                 /* first unlink attempt... */
  51                 lnet_me_t *me = md->md_me;
  52
  53                 md->md_flags |= LNET_MD_FLAG_ZOMBIE;
  54
  55                 /*
  56                  * Disassociate from ME (if any),
  57                  * and unlink it if it was created
  58                  * with LNET_UNLINK
  59                  */
  60                 if (me) {
  61                         /* detach MD from portal */
  62                         lnet_ptl_detach_md(me, md);
  63                         if (me->me_unlink == LNET_UNLINK)
  64                                 lnet_me_unlink(me);
  65                 }
  66
  67                 /* ensure all future handle lookups fail */
  68                 lnet_res_lh_invalidate(&md->md_lh);
  69         }
  70
  71         if (md->md_refcount) {
  72                 CDEBUG(D_NET, "Queueing unlink of md %p\n", md);
  73                 return;
  74         }
  75
  76         CDEBUG(D_NET, "Unlinking md %p\n", md);
  77
  78         if (md->md_eq) {
  79                 int cpt = lnet_cpt_of_cookie(md->md_lh.lh_cookie);
  80
  81                 LASSERT(*md->md_eq->eq_refs[cpt] > 0);
  82                 (*md->md_eq->eq_refs[cpt])--;
  83         }
  84
  85         LASSERT(!list_empty(&md->md_list));
  86         list_del_init(&md->md_list);
  87         lnet_md_free(md);
  88 }
  89
  90 static int
  91 lnet_md_build(lnet_libmd_t *lmd, lnet_md_t *umd, int unlink)
  92 {
  93         int i;
  94         unsigned int niov;
  95         int total_length = 0;
  96
  97         lmd->md_me = NULL;
  98         lmd->md_start = umd->start;
  99         lmd->md_offset = 0;
 100         lmd->md_max_size = umd->max_size;
 101         lmd->md_options = umd->options;
 102         lmd->md_user_ptr = umd->user_ptr;
 103         lmd->md_eq = NULL;
 104         lmd->md_threshold = umd->threshold;
 105         lmd->md_refcount = 0;
 106         lmd->md_flags = (unlink == LNET_UNLINK) ? LNET_MD_FLAG_AUTO_UNLINK : 0;
 107
 108         if (umd->options & LNET_MD_IOVEC) {
 109                 if (umd->options & LNET_MD_KIOV) /* Can't specify both */
 110                         return -EINVAL;
 111
 112                 niov = umd->length;
 113                 lmd->md_niov = umd->length;
 114                 memcpy(lmd->md_iov.iov, umd->start,
 115                        niov * sizeof(lmd->md_iov.iov[0]));
 116
 117                 for (i = 0; i < (int)niov; i++) {
 118                         /* We take the base address on trust */
 119                         /* invalid length */
 120                         if (lmd->md_iov.iov[i].iov_len <= 0)
 121                                 return -EINVAL;
 122
 123                         total_length += lmd->md_iov.iov[i].iov_len;
 124                 }
 125
 126                 lmd->md_length = total_length;
 127
 128                 if ((umd->options & LNET_MD_MAX_SIZE) && /* use max size */
 129                     (umd->max_size < 0 ||
 130                      umd->max_size > total_length)) /* illegal max_size */
 131                         return -EINVAL;
 132
 133         } else if (umd->options & LNET_MD_KIOV) {
 134                 niov = umd->length;
 135                 lmd->md_niov = umd->length;
 136                 memcpy(lmd->md_iov.kiov, umd->start,
 137                        niov * sizeof(lmd->md_iov.kiov[0]));
 138
 139                 for (i = 0; i < (int)niov; i++) {
 140                         /* We take the page pointer on trust */
 141                         if (lmd->md_iov.kiov[i].kiov_offset +
 142                             lmd->md_iov.kiov[i].kiov_len > PAGE_SIZE)
 143                                 return -EINVAL; /* invalid length */
 144
 145                         total_length += lmd->md_iov.kiov[i].kiov_len;
 146                 }
 147
 148                 lmd->md_length = total_length;
 149
 150                 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
 151                     (umd->max_size < 0 ||
 152                      umd->max_size > total_length)) /* illegal max_size */
 153                         return -EINVAL;
 154         } else {   /* contiguous */
 155                 lmd->md_length = umd->length;
 156                 niov = 1;
 157                 lmd->md_niov = 1;
 158                 lmd->md_iov.iov[0].iov_base = umd->start;
 159                 lmd->md_iov.iov[0].iov_len = umd->length;
 160
 161                 if ((umd->options & LNET_MD_MAX_SIZE) && /* max size used */
 162                     (umd->max_size < 0 ||
 163                      umd->max_size > (int)umd->length)) /* illegal max_size */
 164                         return -EINVAL;
 165         }
 166
 167         return 0;
 168 }
 169
 170 /* must be called with resource lock held */
 171 static int
 172 lnet_md_link(lnet_libmd_t *md, lnet_handle_eq_t eq_handle, int cpt)
 173 {
 174         struct lnet_res_container *container = the_lnet.ln_md_containers[cpt];
 175
 176         /*
 177          * NB we are passed an allocated, but inactive md.
 178          * if we return success, caller may lnet_md_unlink() it.
 179          * otherwise caller may only lnet_md_free() it.
 180          */
 181         /*
 182          * This implementation doesn't know how to create START events or
 183          * disable END events.  Best to LASSERT our caller is compliant so
 184          * we find out quickly...
 185          */
 186         /*
 187          * TODO - reevaluate what should be here in light of
 188          * the removal of the start and end events
 189          * maybe there we shouldn't even allow LNET_EQ_NONE!)
 190          * LASSERT(!eq);
 191          */
 192         if (!LNetHandleIsInvalid(eq_handle)) {
 193                 md->md_eq = lnet_handle2eq(&eq_handle);
 194
 195                 if (!md->md_eq)
 196                         return -ENOENT;
 197
 198                 (*md->md_eq->eq_refs[cpt])++;
 199         }
 200
 201         lnet_res_lh_initialize(container, &md->md_lh);
 202
 203         LASSERT(list_empty(&md->md_list));
 204         list_add(&md->md_list, &container->rec_active);
 205
 206         return 0;
 207 }
 208
 209 /* must be called with lnet_res_lock held */
 210 void
 211 lnet_md_deconstruct(lnet_libmd_t *lmd, lnet_md_t *umd)
 212 {
 213         /* NB this doesn't copy out all the iov entries so when a
 214          * discontiguous MD is copied out, the target gets to know the
 215          * original iov pointer (in start) and the number of entries it had
 216          * and that's all.
 217          */
 218         umd->start = lmd->md_start;
 219         umd->length = !(lmd->md_options &
 220                       (LNET_MD_IOVEC | LNET_MD_KIOV)) ?
 221                       lmd->md_length : lmd->md_niov;
 222         umd->threshold = lmd->md_threshold;
 223         umd->max_size = lmd->md_max_size;
 224         umd->options = lmd->md_options;
 225         umd->user_ptr = lmd->md_user_ptr;
 226         lnet_eq2handle(&umd->eq_handle, lmd->md_eq);
 227 }
 228
 229 static int
 230 lnet_md_validate(lnet_md_t *umd)
 231 {
 232         if (!umd->start && umd->length) {
 233                 CERROR("MD start pointer can not be NULL with length %u\n",
 234                        umd->length);
 235                 return -EINVAL;
 236         }
 237
 238         if ((umd->options & (LNET_MD_KIOV | LNET_MD_IOVEC)) &&
 239             umd->length > LNET_MAX_IOV) {
 240                 CERROR("Invalid option: too many fragments %u, %d max\n",
 241                        umd->length, LNET_MAX_IOV);
 242                 return -EINVAL;
 243         }
 244
 245         return 0;
 246 }
 247
 248 /**
 249  * Create a memory descriptor and attach it to a ME
 250  *
 251  * \param meh A handle for a ME to associate the new MD with.
 252  * \param umd Provides initial values for the user-visible parts of a MD.
 253  * Other than its use for initialization, there is no linkage between this
 254  * structure and the MD maintained by the LNet.
 255  * \param unlink A flag to indicate whether the MD is automatically unlinked
 256  * when it becomes inactive, either because the operation threshold drops to
 257  * zero or because the available memory becomes less than \a umd.max_size.
 258  * (Note that the check for unlinking a MD only occurs after the completion
 259  * of a successful operation on the MD.) The value LNET_UNLINK enables auto
 260  * unlinking; the value LNET_RETAIN disables it.
 261  * \param handle On successful returns, a handle to the newly created MD is
 262  * saved here. This handle can be used later in LNetMDUnlink().
 263  *
 264  * \retval 0       On success.
 265  * \retval -EINVAL If \a umd is not valid.
 266  * \retval -ENOMEM If new MD cannot be allocated.
 267  * \retval -ENOENT Either \a meh or \a umd.eq_handle does not point to a
 268  * valid object. Note that it's OK to supply a NULL \a umd.eq_handle by
 269  * calling LNetInvalidateHandle() on it.
 270  * \retval -EBUSY  If the ME pointed to by \a meh is already associated with
 271  * a MD.
 272  */
 273 int
 274 LNetMDAttach(lnet_handle_me_t meh, lnet_md_t umd,
 275              lnet_unlink_t unlink, lnet_handle_md_t *handle)
 276 {
 277         LIST_HEAD(matches);
 278         LIST_HEAD(drops);
 279         struct lnet_me *me;
 280         struct lnet_libmd *md;
 281         int cpt;
 282         int rc;
 283
 284         LASSERT(the_lnet.ln_refcount > 0);
 285
 286         if (lnet_md_validate(&umd))
 287                 return -EINVAL;
 288
 289         if (!(umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
 290                 CERROR("Invalid option: no MD_OP set\n");
 291                 return -EINVAL;
 292         }
 293
 294         md = lnet_md_alloc(&umd);
 295         if (!md)
 296                 return -ENOMEM;
 297
 298         rc = lnet_md_build(md, &umd, unlink);
 299         cpt = lnet_cpt_of_cookie(meh.cookie);
 300
 301         lnet_res_lock(cpt);
 302         if (rc)
 303                 goto failed;
 304
 305         me = lnet_handle2me(&meh);
 306         if (!me)
 307                 rc = -ENOENT;
 308         else if (me->me_md)
 309                 rc = -EBUSY;
 310         else
 311                 rc = lnet_md_link(md, umd.eq_handle, cpt);
 312
 313         if (rc)
 314                 goto failed;
 315
 316         /*
 317          * attach this MD to portal of ME and check if it matches any
 318          * blocked msgs on this portal
 319          */
 320         lnet_ptl_attach_md(me, md, &matches, &drops);
 321
 322         lnet_md2handle(handle, md);
 323
 324         lnet_res_unlock(cpt);
 325
 326         lnet_drop_delayed_msg_list(&drops, "Bad match");
 327         lnet_recv_delayed_msg_list(&matches);
 328
 329         return 0;
 330
 331  failed:
 332         lnet_md_free(md);
 333
 334         lnet_res_unlock(cpt);
 335         return rc;
 336 }
 337 EXPORT_SYMBOL(LNetMDAttach);
 338
 339 /**
 340  * Create a "free floating" memory descriptor - a MD that is not associated
 341  * with a ME. Such MDs are usually used in LNetPut() and LNetGet() operations.
 342  *
 343  * \param umd,unlink See the discussion for LNetMDAttach().
 344  * \param handle On successful returns, a handle to the newly created MD is
 345  * saved here. This handle can be used later in LNetMDUnlink(), LNetPut(),
 346  * and LNetGet() operations.
 347  *
 348  * \retval 0       On success.
 349  * \retval -EINVAL If \a umd is not valid.
 350  * \retval -ENOMEM If new MD cannot be allocated.
 351  * \retval -ENOENT \a umd.eq_handle does not point to a valid EQ. Note that
 352  * it's OK to supply a NULL \a umd.eq_handle by calling
 353  * LNetInvalidateHandle() on it.
 354  */
 355 int
 356 LNetMDBind(lnet_md_t umd, lnet_unlink_t unlink, lnet_handle_md_t *handle)
 357 {
 358         lnet_libmd_t *md;
 359         int cpt;
 360         int rc;
 361
 362         LASSERT(the_lnet.ln_refcount > 0);
 363
 364         if (lnet_md_validate(&umd))
 365                 return -EINVAL;
 366
 367         if ((umd.options & (LNET_MD_OP_GET | LNET_MD_OP_PUT))) {
 368                 CERROR("Invalid option: GET|PUT illegal on active MDs\n");
 369                 return -EINVAL;
 370         }
 371
 372         md = lnet_md_alloc(&umd);
 373         if (!md)
 374                 return -ENOMEM;
 375
 376         rc = lnet_md_build(md, &umd, unlink);
 377
 378         cpt = lnet_res_lock_current();
 379         if (rc)
 380                 goto failed;
 381
 382         rc = lnet_md_link(md, umd.eq_handle, cpt);
 383         if (rc)
 384                 goto failed;
 385
 386         lnet_md2handle(handle, md);
 387
 388         lnet_res_unlock(cpt);
 389         return 0;
 390
 391  failed:
 392         lnet_md_free(md);
 393
 394         lnet_res_unlock(cpt);
 395         return rc;
 396 }
 397 EXPORT_SYMBOL(LNetMDBind);
 398
 399 /**
 400  * Unlink the memory descriptor from any ME it may be linked to and release
 401  * the internal resources associated with it. As a result, active messages
 402  * associated with the MD may get aborted.
 403  *
 404  * This function does not free the memory region associated with the MD;
 405  * i.e., the memory the user allocated for this MD. If the ME associated with
 406  * this MD is not NULL and was created with auto unlink enabled, the ME is
 407  * unlinked as well (see LNetMEAttach()).
 408  *
 409  * Explicitly unlinking a MD via this function call has the same behavior as
 410  * a MD that has been automatically unlinked, except that no LNET_EVENT_UNLINK
 411  * is generated in the latter case.
 412  *
 413  * An unlinked event can be reported in two ways:
 414  * - If there's no pending operations on the MD, it's unlinked immediately
 415  *   and an LNET_EVENT_UNLINK event is logged before this function returns.
 416  * - Otherwise, the MD is only marked for deletion when this function
 417  *   returns, and the unlinked event will be piggybacked on the event of
 418  *   the completion of the last operation by setting the unlinked field of
 419  *   the event. No dedicated LNET_EVENT_UNLINK event is generated.
 420  *
 421  * Note that in both cases the unlinked field of the event is always set; no
 422  * more event will happen on the MD after such an event is logged.
 423  *
 424  * \param mdh A handle for the MD to be unlinked.
 425  *
 426  * \retval 0       On success.
 427  * \retval -ENOENT If \a mdh does not point to a valid MD object.
 428  */
 429 int
 430 LNetMDUnlink(lnet_handle_md_t mdh)
 431 {
 432         lnet_event_t ev;
 433         lnet_libmd_t *md;
 434         int cpt;
 435
 436         LASSERT(the_lnet.ln_refcount > 0);
 437
 438         cpt = lnet_cpt_of_cookie(mdh.cookie);
 439         lnet_res_lock(cpt);
 440
 441         md = lnet_handle2md(&mdh);
 442         if (!md) {
 443                 lnet_res_unlock(cpt);
 444                 return -ENOENT;
 445         }
 446
 447         md->md_flags |= LNET_MD_FLAG_ABORTED;
 448         /*
 449          * If the MD is busy, lnet_md_unlink just marks it for deletion, and
 450          * when the LND is done, the completion event flags that the MD was
 451          * unlinked.  Otherwise, we enqueue an event now...
 452          */
 453         if (md->md_eq && !md->md_refcount) {
 454                 lnet_build_unlink_event(md, &ev);
 455                 lnet_eq_enqueue_event(md->md_eq, &ev);
 456         }
 457
 458         lnet_md_unlink(md);
 459
 460         lnet_res_unlock(cpt);
 461         return 0;
 462 }
 463 EXPORT_SYMBOL(LNetMDUnlink);