lustre/lov/lov_pack.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, 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  * lustre/lov/lov_pack.c
  37  *
  38  * (Un)packing of OST/MDS requests
  39  *
  40  * Author: Andreas Dilger <adilger@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LOV
  44 #ifndef __KERNEL__
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <lustre_net.h>
  49 #include <obd.h>
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <lustre/lustre_user.h>
  53
  54 #include "lov_internal.h"
  55
  56 void lov_dump_lmm_common(int level, void *lmmp)
  57 {
  58         struct lov_mds_md *lmm = lmmp;
  59         struct ost_id   oi;
  60
  61         lmm_oi_le_to_cpu(&oi, &lmm->lmm_oi);
  62         CDEBUG(level, "objid "DOSTID", magic 0x%08x, pattern %#x\n",
  63                POSTID(&oi), le32_to_cpu(lmm->lmm_magic),
  64                le32_to_cpu(lmm->lmm_pattern));
  65         CDEBUG(level, "stripe_size %u, stripe_count %u, layout_gen %u\n",
  66                le32_to_cpu(lmm->lmm_stripe_size),
  67                le16_to_cpu(lmm->lmm_stripe_count),
  68                le16_to_cpu(lmm->lmm_layout_gen));
  69 }
  70
  71 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
  72                                  int stripe_count)
  73 {
  74         int i;
  75
  76         if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
  77                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  78                        stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
  79                 return;
  80         }
  81
  82         for (i = 0; i < stripe_count; ++i, ++lod) {
  83                 struct ost_id   oi;
  84
  85                 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
  86                 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
  87                        le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
  88         }
  89 }
  90
  91 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  92 {
  93         lov_dump_lmm_common(level, lmm);
  94         lov_dump_lmm_objects(level, lmm->lmm_objects,
  95                              le16_to_cpu(lmm->lmm_stripe_count));
  96 }
  97
  98 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
  99 {
 100         lov_dump_lmm_common(level, lmm);
 101         CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
 102         lov_dump_lmm_objects(level, lmm->lmm_objects,
 103                              le16_to_cpu(lmm->lmm_stripe_count));
 104 }
 105
 106 void lov_dump_lmm(int level, void *lmm)
 107 {
 108         int magic;
 109
 110         magic = le32_to_cpu(((struct lov_mds_md *)lmm)->lmm_magic);
 111         switch (magic) {
 112         case LOV_MAGIC_V1:
 113                 lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)lmm);
 114                 break;
 115         case LOV_MAGIC_V3:
 116                 lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)lmm);
 117                 break;
 118         default:
 119                 CDEBUG(level, "unrecognized lmm_magic %x, assuming %x\n",
 120                        magic, LOV_MAGIC_V1);
 121                 lov_dump_lmm_common(level, lmm);
 122                 break;
 123         }
 124 }
 125
 126 /* Pack LOV object metadata for disk storage.  It is packed in LE byte
 127  * order and is opaque to the networking layer.
 128  *
 129  * XXX In the future, this will be enhanced to get the EA size from the
 130  *     underlying OSC device(s) to get their EA sizes so we can stack
 131  *     LOVs properly.  For now lov_mds_md_size() just assumes one obd_id
 132  *     per stripe.
 133  */
 134 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
 135                struct lov_stripe_md *lsm)
 136 {
 137         struct obd_device *obd = class_exp2obd(exp);
 138         struct lov_obd *lov = &obd->u.lov;
 139         struct lov_mds_md_v1 *lmmv1;
 140         struct lov_mds_md_v3 *lmmv3;
 141         __u16 stripe_count;
 142         struct lov_ost_data_v1 *lmm_objects;
 143         int lmm_size, lmm_magic;
 144         int i;
 145         int cplen = 0;
 146         ENTRY;
 147
 148         if (lsm) {
 149                 lmm_magic = lsm->lsm_magic;
 150         } else {
 151                 if (lmmp && *lmmp)
 152                         lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
 153                 else
 154                         /* lsm == NULL and lmmp == NULL */
 155                         lmm_magic = LOV_MAGIC;
 156         }
 157
 158         if ((lmm_magic != LOV_MAGIC_V1) &&
 159             (lmm_magic != LOV_MAGIC_V3)) {
 160                 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
 161                         lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
 162                 RETURN(-EINVAL);
 163
 164         }
 165
 166         if (lsm) {
 167                 /* If we are just sizing the EA, limit the stripe count
 168                  * to the actual number of OSTs in this filesystem. */
 169                 if (!lmmp) {
 170                         stripe_count = lov_get_stripecnt(lov, lmm_magic,
 171                                                         lsm->lsm_stripe_count);
 172                         lsm->lsm_stripe_count = stripe_count;
 173                 } else if (!lsm_is_released(lsm)) {
 174                         stripe_count = lsm->lsm_stripe_count;
 175                 } else {
 176                         stripe_count = 0;
 177                 }
 178         } else {
 179                 /* No need to allocate more than maximum supported stripes.
 180                  * Anyway, this is pretty inaccurate since ld_tgt_count now
 181                  * represents max index and we should rely on the actual number
 182                  * of OSTs instead */
 183                 stripe_count = lov_mds_md_max_stripe_count(
 184                         lov->lov_ocd.ocd_max_easize, lmm_magic);
 185
 186                 if (stripe_count > lov->desc.ld_tgt_count)
 187                         stripe_count = lov->desc.ld_tgt_count;
 188         }
 189
 190         /* XXX LOV STACKING call into osc for sizes */
 191         lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
 192
 193         if (!lmmp)
 194                 RETURN(lmm_size);
 195
 196         if (*lmmp && !lsm) {
 197                 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
 198                 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
 199                 OBD_FREE_LARGE(*lmmp, lmm_size);
 200                 *lmmp = NULL;
 201                 RETURN(0);
 202         }
 203
 204         if (!*lmmp) {
 205                 OBD_ALLOC_LARGE(*lmmp, lmm_size);
 206                 if (!*lmmp)
 207                         RETURN(-ENOMEM);
 208         }
 209
 210         CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
 211                lmm_magic, lmm_size);
 212
 213         lmmv1 = *lmmp;
 214         lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
 215         if (lmm_magic == LOV_MAGIC_V3)
 216                 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
 217         else
 218                 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
 219
 220         if (!lsm)
 221                 RETURN(lmm_size);
 222
 223         /* lmmv1 and lmmv3 point to the same struct and have the
 224          * same first fields
 225          */
 226         lmm_oi_cpu_to_le(&lmmv1->lmm_oi, &lsm->lsm_oi);
 227         lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 228         lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
 229         lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 230         lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
 231         if (lsm->lsm_magic == LOV_MAGIC_V3) {
 232                 cplen = strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
 233                                 sizeof(lmmv3->lmm_pool_name));
 234                 if (cplen >= sizeof(lmmv3->lmm_pool_name))
 235                         RETURN(-E2BIG);
 236                 lmm_objects = lmmv3->lmm_objects;
 237         } else {
 238                 lmm_objects = lmmv1->lmm_objects;
 239         }
 240
 241         for (i = 0; i < stripe_count; i++) {
 242                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 243                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 244                 LASSERTF(ostid_id(&loi->loi_oi) != 0, "lmm_oi "DOSTID
 245                          " stripe %u/%u idx %u\n", POSTID(&lmmv1->lmm_oi),
 246                          i, stripe_count, loi->loi_ost_idx);
 247                 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
 248                 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 249                 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 250         }
 251
 252         RETURN(lmm_size);
 253 }
 254
 255 /* Find the max stripecount we should use */
 256 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
 257 {
 258         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
 259
 260         if (!stripe_count)
 261                 stripe_count = lov->desc.ld_default_stripe_count;
 262         if (stripe_count > lov->desc.ld_active_tgt_count)
 263                 stripe_count = lov->desc.ld_active_tgt_count;
 264         if (!stripe_count)
 265                 stripe_count = 1;
 266
 267         /* stripe count is based on whether ldiskfs can handle
 268          * larger EA sizes */
 269         if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
 270             lov->lov_ocd.ocd_max_easize)
 271                 max_stripes = lov_mds_md_max_stripe_count(
 272                         lov->lov_ocd.ocd_max_easize, magic);
 273
 274         if (stripe_count > max_stripes)
 275                 stripe_count = max_stripes;
 276
 277         return stripe_count;
 278 }
 279
 280
 281 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
 282 {
 283         int rc;
 284
 285         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 286                 char *buffer;
 287                 int sz;
 288
 289                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
 290                        le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
 291                 sz = lmm_bytes * 2 + 1;
 292                 OBD_ALLOC_LARGE(buffer, sz);
 293                 if (buffer != NULL) {
 294                         int i;
 295
 296                         for (i = 0; i < lmm_bytes; i++)
 297                                 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
 298                         buffer[sz - 1] = '\0';
 299                         CERROR("%s\n", buffer);
 300                         OBD_FREE_LARGE(buffer, sz);
 301                 }
 302                 return -EINVAL;
 303         }
 304         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 305                                      lmm_bytes, stripe_count);
 306         return rc;
 307 }
 308
 309 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
 310                     int pattern, int magic)
 311 {
 312         int i, lsm_size;
 313         ENTRY;
 314
 315         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 316
 317         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 318         if (!*lsmp) {
 319                 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
 320                 RETURN(-ENOMEM);
 321         }
 322
 323         atomic_set(&(*lsmp)->lsm_refc, 1);
 324         spin_lock_init(&(*lsmp)->lsm_lock);
 325         (*lsmp)->lsm_magic = magic;
 326         (*lsmp)->lsm_stripe_count = stripe_count;
 327         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 328         (*lsmp)->lsm_pattern = pattern;
 329         (*lsmp)->lsm_pool_name[0] = '\0';
 330         (*lsmp)->lsm_layout_gen = 0;
 331         if (stripe_count > 0)
 332                 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 333
 334         for (i = 0; i < stripe_count; i++)
 335                 loi_init((*lsmp)->lsm_oinfo[i]);
 336
 337         RETURN(lsm_size);
 338 }
 339
 340 int lov_free_memmd(struct lov_stripe_md **lsmp)
 341 {
 342         struct lov_stripe_md *lsm = *lsmp;
 343         int refc;
 344
 345         *lsmp = NULL;
 346         refc = atomic_dec_return(&lsm->lsm_refc);
 347         LASSERT(refc >= 0);
 348         if (refc == 0) {
 349                 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 350                 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 351         }
 352         return refc;
 353 }
 354
 355
 356 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 357  * order and is opaque to the networking layer.
 358  */
 359 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 360                  struct lov_mds_md *lmm, int lmm_bytes)
 361 {
 362         struct obd_device *obd = class_exp2obd(exp);
 363         struct lov_obd *lov = &obd->u.lov;
 364         int rc = 0, lsm_size;
 365         __u16 stripe_count;
 366         __u32 magic;
 367         __u32 pattern;
 368         ENTRY;
 369
 370         /* If passed an MDS struct use values from there, otherwise defaults */
 371         if (lmm) {
 372                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 373                 if (rc)
 374                         RETURN(rc);
 375                 magic = le32_to_cpu(lmm->lmm_magic);
 376                 pattern = le32_to_cpu(lmm->lmm_pattern);
 377         } else {
 378                 magic = LOV_MAGIC;
 379                 stripe_count = lov_get_stripecnt(lov, magic, 0);
 380                 pattern = LOV_PATTERN_RAID0;
 381         }
 382
 383         /* If we aren't passed an lsmp struct, we just want the size */
 384         if (!lsmp) {
 385                 /* XXX LOV STACKING call into osc for sizes */
 386                 LBUG();
 387                 RETURN(lov_stripe_md_size(stripe_count));
 388         }
 389         /* If we are passed an allocated struct but nothing to unpack, free */
 390         if (*lsmp && !lmm) {
 391                 lov_free_memmd(lsmp);
 392                 RETURN(0);
 393         }
 394
 395         lsm_size = lov_alloc_memmd(lsmp, stripe_count, pattern, magic);
 396         if (lsm_size < 0)
 397                 RETURN(lsm_size);
 398
 399         /* If we are passed a pointer but nothing to unpack, we only alloc */
 400         if (!lmm)
 401                 RETURN(lsm_size);
 402
 403         LASSERT(lsm_op_find(magic) != NULL);
 404         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 405         if (rc) {
 406                 lov_free_memmd(lsmp);
 407                 RETURN(rc);
 408         }
 409
 410         RETURN(lsm_size);
 411 }
 412
 413 /* Retrieve object striping information.
 414  *
 415  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 416  * the maximum number of OST indices which will fit in the user buffer.
 417  * lmm_magic must be LOV_USER_MAGIC.
 418  */
 419 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 420                   struct lov_user_md *lump)
 421 {
 422         /*
 423          * XXX huge struct allocated on stack.
 424          */
 425         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
 426         struct lov_user_md_v3 lum;
 427         struct lov_mds_md *lmmk = NULL;
 428         int rc, lmm_size;
 429         int lum_size;
 430         ENTRY;
 431
 432         if (!lsm)
 433                 RETURN(-ENODATA);
 434
 435         /* we only need the header part from user space to get lmm_magic and
 436          * lmm_stripe_count, (the header part is common to v1 and v3) */
 437         lum_size = sizeof(struct lov_user_md_v1);
 438         if (copy_from_user(&lum, lump, lum_size))
 439                 GOTO(out_set, rc = -EFAULT);
 440         else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
 441                  (lum.lmm_magic != LOV_USER_MAGIC_V3))
 442                 GOTO(out_set, rc = -EINVAL);
 443
 444         if (lum.lmm_stripe_count &&
 445             (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
 446                 /* Return right size of stripe to user */
 447                 lum.lmm_stripe_count = lsm->lsm_stripe_count;
 448                 rc = copy_to_user(lump, &lum, lum_size);
 449                 GOTO(out_set, rc = -EOVERFLOW);
 450         }
 451         rc = lov_packmd(exp, &lmmk, lsm);
 452         if (rc < 0)
 453                 GOTO(out_set, rc);
 454         lmm_size = rc;
 455         rc = 0;
 456
 457         /* FIXME: Bug 1185 - copy fields properly when structs change */
 458         /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
 459         CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
 460         CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
 461
 462         if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
 463             ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
 464             (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
 465                 lustre_swab_lov_mds_md(lmmk);
 466                 lustre_swab_lov_user_md_objects(
 467                                 (struct lov_user_ost_data*)lmmk->lmm_objects,
 468                                 lmmk->lmm_stripe_count);
 469         }
 470         if (lum.lmm_magic == LOV_USER_MAGIC) {
 471                 /* User request for v1, we need skip lmm_pool_name */
 472                 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
 473                         memmove(((struct lov_mds_md_v1 *)lmmk)->lmm_objects,
 474                                 ((struct lov_mds_md_v3 *)lmmk)->lmm_objects,
 475                                 lmmk->lmm_stripe_count *
 476                                 sizeof(struct lov_ost_data_v1));
 477                         lmm_size -= LOV_MAXPOOLNAME;
 478                 }
 479         } else {
 480                 /* if v3 we just have to update the lum_size */
 481                 lum_size = sizeof(struct lov_user_md_v3);
 482         }
 483
 484         /* User wasn't expecting this many OST entries */
 485         if (lum.lmm_stripe_count == 0)
 486                 lmm_size = lum_size;
 487         else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
 488                 GOTO(out_set, rc = -EOVERFLOW);
 489         /*
 490          * Have a difference between lov_mds_md & lov_user_md.
 491          * So we have to re-order the data before copy to user.
 492          */
 493         lum.lmm_stripe_count = lmmk->lmm_stripe_count;
 494         lum.lmm_layout_gen = lmmk->lmm_layout_gen;
 495         ((struct lov_user_md *)lmmk)->lmm_layout_gen = lum.lmm_layout_gen;
 496         ((struct lov_user_md *)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
 497         if (copy_to_user(lump, lmmk, lmm_size))
 498                 rc = -EFAULT;
 499
 500         obd_free_diskmd(exp, &lmmk);
 501 out_set:
 502         RETURN(rc);
 503 }