lustre/lov/lov_pack.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Andreas Dilger <adilger@clusterfs.com>
   6  *
   7  *   This file is part of Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  * (Un)packing of OST/MDS requests
  23  *
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_LOV
  27 #ifndef __KERNEL__
  28 #include <liblustre.h>
  29 #endif
  30
  31 #include <lustre_net.h>
  32 #include <obd.h>
  33 #include <obd_lov.h>
  34 #include <obd_class.h>
  35 #include <obd_support.h>
  36 #include <lustre/lustre_user.h>
  37
  38 #include "lov_internal.h"
  39
  40 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  41 {
  42         struct lov_ost_data_v1 *lod;
  43         int i;
  44         CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
  45                le64_to_cpu(lmm->lmm_object_id), le32_to_cpu(lmm->lmm_magic),
  46                le32_to_cpu(lmm->lmm_pattern));
  47         CDEBUG(level,"stripe_size %u, stripe_count %u\n",
  48                le32_to_cpu(lmm->lmm_stripe_size),
  49                le32_to_cpu(lmm->lmm_stripe_count));
  50         for (i = 0, lod = lmm->lmm_objects;
  51              i < le32_to_cpu(lmm->lmm_stripe_count); i++, lod++)
  52                 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n",
  53                        i, le32_to_cpu(lod->l_ost_idx),
  54                        le64_to_cpu(lod->l_object_gr),
  55                        le64_to_cpu(lod->l_object_id));
  56 }
  57
  58 void lov_dump_lmm_join(int level, struct lov_mds_md_join *lmmj)
  59 {
  60
  61         CDEBUG(level, "objid "LPX64", magic 0x%08X, pattern %#X\n",
  62                le64_to_cpu(lmmj->lmmj_md.lmm_object_id),
  63                le32_to_cpu(lmmj->lmmj_md.lmm_magic),
  64                le32_to_cpu(lmmj->lmmj_md.lmm_pattern));
  65         CDEBUG(level,"stripe_size %u, stripe_count %u extent_count %u \n",
  66                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_size),
  67                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_count),
  68                le32_to_cpu(lmmj->lmmj_extent_count));
  69 }
  70
  71 #define LMM_ASSERT(test)                                                \
  72 do {                                                                    \
  73         if (!(test)) lov_dump_lmm(D_ERROR, lmm);                        \
  74         LASSERT(test); /* so we know what assertion failed */           \
  75 } while(0)
  76
  77 /* Pack LOV object metadata for disk storage.  It is packed in LE byte
  78  * order and is opaque to the networking layer.
  79  *
  80  * XXX In the future, this will be enhanced to get the EA size from the
  81  *     underlying OSC device(s) to get their EA sizes so we can stack
  82  *     LOVs properly.  For now lov_mds_md_size() just assumes one obd_id
  83  *     per stripe.
  84  */
  85 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  86                struct lov_stripe_md *lsm)
  87 {
  88         struct obd_device *obd = class_exp2obd(exp);
  89         struct lov_obd *lov = &obd->u.lov;
  90         struct lov_oinfo *loi;
  91         struct lov_mds_md *lmm;
  92         int stripe_count = lov->desc.ld_tgt_count;
  93         int lmm_size;
  94         int i;
  95         ENTRY;
  96
  97         if (lsm) {
  98                 if (lsm->lsm_magic != LOV_MAGIC) {
  99                         CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X\n",
 100                                lsm->lsm_magic, LOV_MAGIC);
 101                         RETURN(-EINVAL);
 102                 }
 103                 stripe_count = lsm->lsm_stripe_count;
 104         }
 105
 106         /* XXX LOV STACKING call into osc for sizes */
 107         lmm_size = lov_mds_md_size(stripe_count);
 108
 109         if (!lmmp)
 110                 RETURN(lmm_size);
 111
 112         if (*lmmp && !lsm) {
 113                 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
 114                 OBD_FREE(*lmmp, lov_mds_md_size(stripe_count));
 115                 *lmmp = NULL;
 116                 RETURN(0);
 117         }
 118
 119         if (!*lmmp) {
 120                 OBD_ALLOC(*lmmp, lmm_size);
 121                 if (!*lmmp)
 122                         RETURN(-ENOMEM);
 123         }
 124
 125         lmm = *lmmp;
 126         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC); /* only write new format */
 127
 128         if (!lsm)
 129                 RETURN(lmm_size);
 130
 131         lmm->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 132         lmm->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
 133         lmm->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 134         lmm->lmm_stripe_count = cpu_to_le32(stripe_count);
 135         lmm->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 136
 137         for (i = 0, loi = lsm->lsm_oinfo; i < stripe_count; i++, loi++) {
 138                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 139                 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
 140                          lmm->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
 141                 lmm->lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
 142                 lmm->lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
 143                 lmm->lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 144                 lmm->lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 145         }
 146
 147         RETURN(lmm_size);
 148 }
 149
 150 /* Find the max stripecount we should use */
 151 int lov_get_stripecnt(struct lov_obd *lov, int stripe_count)
 152 {
 153         if (!stripe_count)
 154                 stripe_count = lov->desc.ld_default_stripe_count;
 155         if (stripe_count > lov->desc.ld_active_tgt_count)
 156                 stripe_count = lov->desc.ld_active_tgt_count;
 157         if (!stripe_count)
 158                 stripe_count = 1;
 159         /* for now, we limit the stripe count directly, when bug 4424 is
 160          * fixed this needs to be somewhat dynamic based on whether ext3
 161          * can handle larger EA sizes. */
 162         if (stripe_count > LOV_MAX_STRIPE_COUNT)
 163                 stripe_count = LOV_MAX_STRIPE_COUNT;
 164
 165         return stripe_count;
 166 }
 167
 168
 169 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
 170 {
 171         int rc;
 172
 173         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 174                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping V1 LMM:\n",
 175                        le32_to_cpu(*(__u32 *)lmm));
 176                 lov_dump_lmm_v1(D_WARNING, lmm);
 177                 return -EINVAL;
 178         }
 179         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 180                                      lmm_bytes, stripe_count);
 181         return rc;
 182 }
 183
 184 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
 185                       int pattern, int magic)
 186 {
 187         int lsm_size = lov_stripe_md_size(stripe_count);
 188         struct lov_oinfo *loi;
 189         int i;
 190         ENTRY;
 191
 192         CDEBUG(D_INFO, "alloc lsm, stripe_count %d, lsm_size %d\n",
 193                stripe_count, lsm_size);
 194
 195         OBD_ALLOC(*lsmp, lsm_size);
 196         if (!*lsmp) {
 197                 CERROR("can not allocate lsmp lsm_size %d stripe_count %d\n",
 198                         lsm_size, stripe_count);
 199                 RETURN(-ENOMEM);
 200         }
 201
 202         spin_lock_init(&(*lsmp)->lsm_lock);
 203         (*lsmp)->lsm_magic = magic;
 204         (*lsmp)->lsm_stripe_count = stripe_count;
 205         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 206         (*lsmp)->lsm_pattern = pattern;
 207         (*lsmp)->lsm_oinfo[0].loi_ost_idx = ~0;
 208
 209         for (i = 0, loi = (*lsmp)->lsm_oinfo; i < stripe_count; i++, loi++)
 210                 loi_init(loi);
 211
 212         RETURN(lsm_size);
 213 }
 214
 215 void lov_free_memmd(struct lov_stripe_md **lsmp)
 216 {
 217         struct lov_stripe_md *lsm = *lsmp;
 218
 219         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 220         lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 221
 222         *lsmp = NULL;
 223 }
 224
 225
 226 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 227  * order and is opaque to the networking layer.
 228  */
 229 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 230                  struct lov_mds_md *lmm, int lmm_bytes)
 231 {
 232         struct obd_device *obd = class_exp2obd(exp);
 233         struct lov_obd *lov = &obd->u.lov;
 234         int rc = 0, stripe_count, lsm_size;
 235         __u32 magic;
 236         ENTRY;
 237
 238         /* If passed an MDS struct use values from there, otherwise defaults */
 239         if (lmm) {
 240                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 241                 if (rc)
 242                         RETURN(rc);
 243                 magic = le32_to_cpu(lmm->lmm_magic);
 244         } else {
 245                 stripe_count = lov_get_stripecnt(lov, 0);
 246                 magic = LOV_MAGIC;
 247         }
 248
 249         /* If we aren't passed an lsmp struct, we just want the size */
 250         if (!lsmp)
 251                 /* XXX LOV STACKING call into osc for sizes */
 252                 RETURN(lov_stripe_md_size(stripe_count));
 253
 254         /* If we are passed an allocated struct but nothing to unpack, free */
 255         if (*lsmp && !lmm) {
 256                 lov_free_memmd(lsmp);
 257                 RETURN(0);
 258         }
 259
 260         lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
 261                                    magic);
 262         if (lsm_size < 0)
 263                 RETURN(lsm_size);
 264
 265         /* If we are passed a pointer but nothing to unpack, we only alloc */
 266         if (!lmm)
 267                 RETURN(lsm_size);
 268
 269         LASSERT(lsm_op_find(magic) != NULL);
 270         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 271         if (rc) {
 272                 lov_free_memmd(lsmp);
 273                 RETURN(rc);
 274         }
 275
 276         RETURN(lsm_size);
 277 }
 278
 279 /* Configure object striping information on a new file.
 280  *
 281  * @lmmu is a pointer to a user struct with one or more of the fields set to
 282  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 283  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 284  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 285  */
 286 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 287                   struct lov_user_md *lump)
 288 {
 289         struct obd_device *obd = class_exp2obd(exp);
 290         struct lov_obd *lov = &obd->u.lov;
 291         struct lov_user_md lum;
 292         int stripe_count;
 293         int rc;
 294         ENTRY;
 295
 296         rc = copy_from_user(&lum, lump, sizeof(lum));
 297         if (rc)
 298                 RETURN(-EFAULT);
 299
 300         if (lum.lmm_magic != LOV_USER_MAGIC) {
 301                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 302                         lustre_swab_lov_user_md(&lum);
 303                 } else {
 304                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 305                                " %#08x != %#08x\n",
 306                                lum.lmm_magic, LOV_USER_MAGIC);
 307                         RETURN(-EINVAL);
 308                 }
 309         }
 310
 311         if (lum.lmm_pattern == 0) {
 312                 lum.lmm_pattern = lov->desc.ld_pattern ?
 313                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 314         }
 315
 316         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 317                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 318                        lum.lmm_pattern);
 319                 RETURN(-EINVAL);
 320         }
 321
 322         /* 64kB is the largest common page size we see (ia64), and matches the
 323          * check in lfs */
 324         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 325                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 326                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 327                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 328         }
 329
 330         if ((lum.lmm_stripe_offset >= lov->desc.ld_active_tgt_count) &&
 331             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 332                 CDEBUG(D_IOCTL, "stripe offset %u > number of active OSTs %u\n",
 333                        lum.lmm_stripe_offset, lov->desc.ld_active_tgt_count);
 334                 RETURN(-EINVAL);
 335         }
 336         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 337
 338         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0UL) {
 339                 CDEBUG(D_IOCTL, "stripe width %ux%u > %lu on 32-bit system\n",
 340                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0UL);
 341                 RETURN(-EINVAL);
 342         }
 343
 344         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 345
 346         if (rc < 0)
 347                 RETURN(rc);
 348
 349         (*lsmp)->lsm_oinfo[0].loi_ost_idx = lum.lmm_stripe_offset;
 350         (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 351
 352         RETURN(0);
 353 }
 354
 355 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
 356               struct lov_user_md *lump)
 357 {
 358         int i;
 359         int rc;
 360         struct obd_export *oexp;
 361         struct lov_obd *lov = &exp->exp_obd->u.lov;
 362         obd_id last_id = 0;
 363
 364         ENTRY;
 365         for (i = 0; i < lump->lmm_stripe_count; i++) {
 366                 __u32 len = sizeof(last_id);
 367                 oexp = lov->lov_tgts[lump->lmm_objects[i].l_ost_idx]->ltd_exp;
 368                 rc = obd_get_info(oexp, strlen("last_id"), "last_id",
 369                                   &len, &last_id);
 370                 if (rc)
 371                         RETURN(rc);
 372                 if (lump->lmm_objects[i].l_object_id > last_id) {
 373                         CERROR("Setting EA for object > than last id on "
 374                                "ost idx %d "LPD64" > "LPD64" \n",
 375                                lump->lmm_objects[i].l_ost_idx,
 376                                lump->lmm_objects[i].l_object_id, last_id);
 377                         RETURN(-EINVAL);
 378                 }
 379         }
 380
 381         rc = lov_setstripe(exp, lsmp, lump);
 382         if (rc)
 383                 RETURN(rc);
 384
 385         for (i = 0; i < lump->lmm_stripe_count; i++) {
 386                 (*lsmp)->lsm_oinfo[i].loi_ost_idx =
 387                         lump->lmm_objects[i].l_ost_idx;
 388                 (*lsmp)->lsm_oinfo[i].loi_id = lump->lmm_objects[i].l_object_id;
 389                 (*lsmp)->lsm_oinfo[i].loi_gr = lump->lmm_objects[i].l_object_gr;
 390         }
 391         RETURN(0);
 392 }
 393
 394
 395 /* Retrieve object striping information.
 396  *
 397  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 398  * the maximum number of OST indices which will fit in the user buffer.
 399  * lmm_magic must be LOV_USER_MAGIC.
 400  */
 401 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 402                   struct lov_user_md *lump)
 403 {
 404         struct lov_user_md lum;
 405         struct lov_mds_md *lmmk = NULL;
 406         int rc, lmm_size;
 407         ENTRY;
 408
 409         if (!lsm)
 410                 RETURN(-ENODATA);
 411
 412         rc = copy_from_user(&lum, lump, sizeof(lum));
 413         if (rc)
 414                 RETURN(-EFAULT);
 415
 416         if (lum.lmm_magic != LOV_USER_MAGIC)
 417                 RETURN(-EINVAL);
 418
 419         rc = lov_packmd(exp, &lmmk, lsm);
 420         if (rc < 0)
 421                 RETURN(rc);
 422         lmm_size = rc;
 423         rc = 0;
 424
 425         /* FIXME: Bug 1185 - copy fields properly when structs change */
 426         LASSERT(sizeof(lum) == sizeof(*lmmk));
 427         LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lmmk->lmm_objects[0]));
 428
 429         /* User wasn't expecting this many OST entries */
 430         if (lum.lmm_stripe_count == 0) {
 431                 if (copy_to_user(lump, lmmk, sizeof(lum)))
 432                         rc = -EFAULT;
 433         } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 434                 rc = -EOVERFLOW;
 435         } else if (copy_to_user(lump, lmmk, lmm_size)) {
 436                 rc = -EFAULT;
 437         }
 438
 439         obd_free_diskmd(exp, &lmmk);
 440
 441         RETURN(rc);
 442 }