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
  45 #include <lustre_net.h>
  46 #include <obd.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49 #include <lustre/lustre_user.h>
  50
  51 #include "lov_internal.h"
  52
  53 void lov_dump_lmm_common(int level, void *lmmp)
  54 {
  55         struct lov_mds_md *lmm = lmmp;
  56         struct ost_id   oi;
  57
  58         lmm_oi_le_to_cpu(&oi, &lmm->lmm_oi);
  59         CDEBUG(level, "objid "DOSTID", magic 0x%08x, pattern %#x\n",
  60                POSTID(&oi), le32_to_cpu(lmm->lmm_magic),
  61                le32_to_cpu(lmm->lmm_pattern));
  62         CDEBUG(level, "stripe_size %u, stripe_count %u, layout_gen %u\n",
  63                le32_to_cpu(lmm->lmm_stripe_size),
  64                le16_to_cpu(lmm->lmm_stripe_count),
  65                le16_to_cpu(lmm->lmm_layout_gen));
  66 }
  67
  68 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
  69                                  int stripe_count)
  70 {
  71         int i;
  72
  73         if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
  74                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  75                        stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
  76                 return;
  77         }
  78
  79         for (i = 0; i < stripe_count; ++i, ++lod) {
  80                 struct ost_id   oi;
  81
  82                 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
  83                 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
  84                        le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
  85         }
  86 }
  87
  88 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  89 {
  90         lov_dump_lmm_common(level, lmm);
  91         lov_dump_lmm_objects(level, lmm->lmm_objects,
  92                              le16_to_cpu(lmm->lmm_stripe_count));
  93 }
  94
  95 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
  96 {
  97         lov_dump_lmm_common(level, lmm);
  98         CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
  99         lov_dump_lmm_objects(level, lmm->lmm_objects,
 100                              le16_to_cpu(lmm->lmm_stripe_count));
 101 }
 102
 103 void lov_dump_lmm(int level, void *lmm)
 104 {
 105         int magic;
 106
 107         magic = le32_to_cpu(((struct lov_mds_md *)lmm)->lmm_magic);
 108         switch (magic) {
 109         case LOV_MAGIC_V1:
 110                 lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)lmm);
 111                 break;
 112         case LOV_MAGIC_V3:
 113                 lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)lmm);
 114                 break;
 115         default:
 116                 CDEBUG(level, "unrecognized lmm_magic %x, assuming %x\n",
 117                        magic, LOV_MAGIC_V1);
 118                 lov_dump_lmm_common(level, lmm);
 119                 break;
 120         }
 121 }
 122
 123 /* Pack LOV object metadata for disk storage.  It is packed in LE byte
 124  * order and is opaque to the networking layer.
 125  *
 126  * XXX In the future, this will be enhanced to get the EA size from the
 127  *     underlying OSC device(s) to get their EA sizes so we can stack
 128  *     LOVs properly.  For now lov_mds_md_size() just assumes one obd_id
 129  *     per stripe.
 130  */
 131 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
 132                struct lov_stripe_md *lsm)
 133 {
 134         struct obd_device *obd = class_exp2obd(exp);
 135         struct lov_obd *lov = &obd->u.lov;
 136         struct lov_mds_md_v1 *lmmv1;
 137         struct lov_mds_md_v3 *lmmv3;
 138         __u16 stripe_count;
 139         struct lov_ost_data_v1 *lmm_objects;
 140         int lmm_size, lmm_magic;
 141         int i;
 142         int cplen = 0;
 143         ENTRY;
 144
 145         if (lsm) {
 146                 lmm_magic = lsm->lsm_magic;
 147         } else {
 148                 if (lmmp && *lmmp)
 149                         lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
 150                 else
 151                         /* lsm == NULL and lmmp == NULL */
 152                         lmm_magic = LOV_MAGIC;
 153         }
 154
 155         if ((lmm_magic != LOV_MAGIC_V1) &&
 156             (lmm_magic != LOV_MAGIC_V3)) {
 157                 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
 158                         lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
 159                 RETURN(-EINVAL);
 160
 161         }
 162
 163         if (lsm) {
 164                 /* If we are just sizing the EA, limit the stripe count
 165                  * to the actual number of OSTs in this filesystem. */
 166                 if (!lmmp) {
 167                         stripe_count = lov_get_stripecnt(lov, lmm_magic,
 168                                                         lsm->lsm_stripe_count);
 169                         lsm->lsm_stripe_count = stripe_count;
 170                 } else if (!lsm_is_released(lsm)) {
 171                         stripe_count = lsm->lsm_stripe_count;
 172                 } else {
 173                         stripe_count = 0;
 174                 }
 175         } else {
 176                 /* No need to allocate more than maximum supported stripes.
 177                  * Anyway, this is pretty inaccurate since ld_tgt_count now
 178                  * represents max index and we should rely on the actual number
 179                  * of OSTs instead */
 180                 stripe_count = lov_mds_md_max_stripe_count(
 181                         lov->lov_ocd.ocd_max_easize, lmm_magic);
 182
 183                 if (stripe_count > lov->desc.ld_tgt_count)
 184                         stripe_count = lov->desc.ld_tgt_count;
 185         }
 186
 187         /* XXX LOV STACKING call into osc for sizes */
 188         lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
 189
 190         if (!lmmp)
 191                 RETURN(lmm_size);
 192
 193         if (*lmmp && !lsm) {
 194                 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
 195                 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
 196                 OBD_FREE_LARGE(*lmmp, lmm_size);
 197                 *lmmp = NULL;
 198                 RETURN(0);
 199         }
 200
 201         if (!*lmmp) {
 202                 OBD_ALLOC_LARGE(*lmmp, lmm_size);
 203                 if (!*lmmp)
 204                         RETURN(-ENOMEM);
 205         }
 206
 207         CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
 208                lmm_magic, lmm_size);
 209
 210         lmmv1 = *lmmp;
 211         lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
 212         if (lmm_magic == LOV_MAGIC_V3)
 213                 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
 214         else
 215                 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
 216
 217         if (!lsm)
 218                 RETURN(lmm_size);
 219
 220         /* lmmv1 and lmmv3 point to the same struct and have the
 221          * same first fields
 222          */
 223         lmm_oi_cpu_to_le(&lmmv1->lmm_oi, &lsm->lsm_oi);
 224         lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 225         lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
 226         lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 227         lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
 228         if (lsm->lsm_magic == LOV_MAGIC_V3) {
 229                 cplen = strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
 230                                 sizeof(lmmv3->lmm_pool_name));
 231                 if (cplen >= sizeof(lmmv3->lmm_pool_name))
 232                         RETURN(-E2BIG);
 233                 lmm_objects = lmmv3->lmm_objects;
 234         } else {
 235                 lmm_objects = lmmv1->lmm_objects;
 236         }
 237
 238         for (i = 0; i < stripe_count; i++) {
 239                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 240                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 241                 LASSERTF(ostid_id(&loi->loi_oi) != 0, "lmm_oi "DOSTID
 242                          " stripe %u/%u idx %u\n", POSTID(&lmmv1->lmm_oi),
 243                          i, stripe_count, loi->loi_ost_idx);
 244                 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
 245                 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 246                 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 247         }
 248
 249         RETURN(lmm_size);
 250 }
 251
 252 /* Find the max stripecount we should use */
 253 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
 254 {
 255         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
 256
 257         if (!stripe_count)
 258                 stripe_count = lov->desc.ld_default_stripe_count;
 259         if (stripe_count > lov->desc.ld_active_tgt_count)
 260                 stripe_count = lov->desc.ld_active_tgt_count;
 261         if (!stripe_count)
 262                 stripe_count = 1;
 263
 264         /* stripe count is based on whether ldiskfs can handle
 265          * larger EA sizes */
 266         if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
 267             lov->lov_ocd.ocd_max_easize)
 268                 max_stripes = lov_mds_md_max_stripe_count(
 269                         lov->lov_ocd.ocd_max_easize, magic);
 270
 271         if (stripe_count > max_stripes)
 272                 stripe_count = max_stripes;
 273
 274         return stripe_count;
 275 }
 276
 277
 278 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
 279 {
 280         int rc;
 281
 282         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 283                 char *buffer;
 284                 int sz;
 285
 286                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
 287                        le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
 288                 sz = lmm_bytes * 2 + 1;
 289                 OBD_ALLOC_LARGE(buffer, sz);
 290                 if (buffer != NULL) {
 291                         int i;
 292
 293                         for (i = 0; i < lmm_bytes; i++)
 294                                 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
 295                         buffer[sz - 1] = '\0';
 296                         CERROR("%s\n", buffer);
 297                         OBD_FREE_LARGE(buffer, sz);
 298                 }
 299                 return -EINVAL;
 300         }
 301         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 302                                      lmm_bytes, stripe_count);
 303         return rc;
 304 }
 305
 306 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
 307                     int pattern, int magic)
 308 {
 309         int i, lsm_size;
 310         ENTRY;
 311
 312         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 313
 314         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 315         if (!*lsmp) {
 316                 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
 317                 RETURN(-ENOMEM);
 318         }
 319
 320         atomic_set(&(*lsmp)->lsm_refc, 1);
 321         spin_lock_init(&(*lsmp)->lsm_lock);
 322         (*lsmp)->lsm_magic = magic;
 323         (*lsmp)->lsm_stripe_count = stripe_count;
 324         (*lsmp)->lsm_maxbytes = LUSTRE_EXT3_STRIPE_MAXBYTES * stripe_count;
 325         (*lsmp)->lsm_pattern = pattern;
 326         (*lsmp)->lsm_pool_name[0] = '\0';
 327         (*lsmp)->lsm_layout_gen = 0;
 328         if (stripe_count > 0)
 329                 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 330
 331         for (i = 0; i < stripe_count; i++)
 332                 loi_init((*lsmp)->lsm_oinfo[i]);
 333
 334         RETURN(lsm_size);
 335 }
 336
 337 int lov_free_memmd(struct lov_stripe_md **lsmp)
 338 {
 339         struct lov_stripe_md *lsm = *lsmp;
 340         int refc;
 341
 342         *lsmp = NULL;
 343         refc = atomic_dec_return(&lsm->lsm_refc);
 344         LASSERT(refc >= 0);
 345         if (refc == 0) {
 346                 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 347                 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 348         }
 349         return refc;
 350 }
 351
 352
 353 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 354  * order and is opaque to the networking layer.
 355  */
 356 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 357                  struct lov_mds_md *lmm, int lmm_bytes)
 358 {
 359         struct obd_device *obd = class_exp2obd(exp);
 360         struct lov_obd *lov = &obd->u.lov;
 361         int rc = 0, lsm_size;
 362         __u16 stripe_count;
 363         __u32 magic;
 364         __u32 pattern;
 365         ENTRY;
 366
 367         /* If passed an MDS struct use values from there, otherwise defaults */
 368         if (lmm) {
 369                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 370                 if (rc)
 371                         RETURN(rc);
 372                 magic = le32_to_cpu(lmm->lmm_magic);
 373                 pattern = le32_to_cpu(lmm->lmm_pattern);
 374         } else {
 375                 magic = LOV_MAGIC;
 376                 stripe_count = lov_get_stripecnt(lov, magic, 0);
 377                 pattern = LOV_PATTERN_RAID0;
 378         }
 379
 380         /* If we aren't passed an lsmp struct, we just want the size */
 381         if (!lsmp) {
 382                 /* XXX LOV STACKING call into osc for sizes */
 383                 LBUG();
 384                 RETURN(lov_stripe_md_size(stripe_count));
 385         }
 386         /* If we are passed an allocated struct but nothing to unpack, free */
 387         if (*lsmp && !lmm) {
 388                 lov_free_memmd(lsmp);
 389                 RETURN(0);
 390         }
 391
 392         lsm_size = lov_alloc_memmd(lsmp, stripe_count, pattern, magic);
 393         if (lsm_size < 0)
 394                 RETURN(lsm_size);
 395
 396         /* If we are passed a pointer but nothing to unpack, we only alloc */
 397         if (!lmm)
 398                 RETURN(lsm_size);
 399
 400         LASSERT(lsm_op_find(magic) != NULL);
 401         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 402         if (rc) {
 403                 lov_free_memmd(lsmp);
 404                 RETURN(rc);
 405         }
 406
 407         RETURN(lsm_size);
 408 }
 409
 410 /* Retrieve object striping information.
 411  *
 412  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 413  * the maximum number of OST indices which will fit in the user buffer.
 414  * lmm_magic must be LOV_USER_MAGIC.
 415  */
 416 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 417                   struct lov_user_md __user *lump)
 418 {
 419         /*
 420          * XXX huge struct allocated on stack.
 421          */
 422         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
 423         struct lov_user_md_v3 lum;
 424         struct lov_mds_md *lmmk = NULL;
 425         int rc, lmm_size;
 426         int lum_size;
 427         ENTRY;
 428
 429         if (!lsm)
 430                 RETURN(-ENODATA);
 431
 432         /* we only need the header part from user space to get lmm_magic and
 433          * lmm_stripe_count, (the header part is common to v1 and v3) */
 434         lum_size = sizeof(struct lov_user_md_v1);
 435         if (copy_from_user(&lum, lump, lum_size))
 436                 GOTO(out_set, rc = -EFAULT);
 437         else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
 438                  (lum.lmm_magic != LOV_USER_MAGIC_V3))
 439                 GOTO(out_set, rc = -EINVAL);
 440
 441         if (lum.lmm_stripe_count &&
 442             (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
 443                 /* Return right size of stripe to user */
 444                 lum.lmm_stripe_count = lsm->lsm_stripe_count;
 445                 rc = copy_to_user(lump, &lum, lum_size);
 446                 GOTO(out_set, rc = -EOVERFLOW);
 447         }
 448         rc = lov_packmd(exp, &lmmk, lsm);
 449         if (rc < 0)
 450                 GOTO(out_set, rc);
 451         lmm_size = rc;
 452         rc = 0;
 453
 454         /* FIXME: Bug 1185 - copy fields properly when structs change */
 455         /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
 456         CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
 457         CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
 458
 459         if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
 460             ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
 461             (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
 462                 lustre_swab_lov_mds_md(lmmk);
 463                 lustre_swab_lov_user_md_objects(
 464                                 (struct lov_user_ost_data*)lmmk->lmm_objects,
 465                                 lmmk->lmm_stripe_count);
 466         }
 467         if (lum.lmm_magic == LOV_USER_MAGIC) {
 468                 /* User request for v1, we need skip lmm_pool_name */
 469                 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
 470                         memmove(((struct lov_mds_md_v1 *)lmmk)->lmm_objects,
 471                                 ((struct lov_mds_md_v3 *)lmmk)->lmm_objects,
 472                                 lmmk->lmm_stripe_count *
 473                                 sizeof(struct lov_ost_data_v1));
 474                         lmm_size -= LOV_MAXPOOLNAME;
 475                 }
 476         } else {
 477                 /* if v3 we just have to update the lum_size */
 478                 lum_size = sizeof(struct lov_user_md_v3);
 479         }
 480
 481         /* User wasn't expecting this many OST entries */
 482         if (lum.lmm_stripe_count == 0)
 483                 lmm_size = lum_size;
 484         else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
 485                 GOTO(out_set, rc = -EOVERFLOW);
 486         /*
 487          * Have a difference between lov_mds_md & lov_user_md.
 488          * So we have to re-order the data before copy to user.
 489          */
 490         lum.lmm_stripe_count = lmmk->lmm_stripe_count;
 491         lum.lmm_layout_gen = lmmk->lmm_layout_gen;
 492         ((struct lov_user_md *)lmmk)->lmm_layout_gen = lum.lmm_layout_gen;
 493         ((struct lov_user_md *)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
 494         if (copy_to_user(lump, lmmk, lmm_size))
 495                 rc = -EFAULT;
 496
 497         obd_free_diskmd(exp, &lmmk);
 498 out_set:
 499         RETURN(rc);
 500 }