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
  45         CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
  46                le64_to_cpu(lmm->lmm_object_id), le32_to_cpu(lmm->lmm_magic),
  47                le32_to_cpu(lmm->lmm_pattern));
  48         CDEBUG(level,"stripe_size %u, stripe_count %u\n",
  49                le32_to_cpu(lmm->lmm_stripe_size),
  50                le32_to_cpu(lmm->lmm_stripe_count));
  51
  52         if (le32_to_cpu(lmm->lmm_stripe_count) <= LOV_V1_INSANE_STRIPE_COUNT) {
  53                 for (i = 0, lod = lmm->lmm_objects;
  54                      i < (int)le32_to_cpu(lmm->lmm_stripe_count); i++, lod++)
  55                          CDEBUG(level,
  56                                 "stripe %u idx %u subobj "LPX64"/"LPX64"\n",
  57                                 i, le32_to_cpu(lod->l_ost_idx),
  58                                 le64_to_cpu(lod->l_object_gr),
  59                                 le64_to_cpu(lod->l_object_id));
  60         } else {
  61                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  62                        le32_to_cpu(lmm->lmm_stripe_count),
  63                        LOV_V1_INSANE_STRIPE_COUNT);
  64         }
  65 }
  66
  67 void lov_dump_lmm_join(int level, struct lov_mds_md_join *lmmj)
  68 {
  69
  70         CDEBUG(level, "objid "LPX64", magic 0x%08X, pattern %#X\n",
  71                le64_to_cpu(lmmj->lmmj_md.lmm_object_id),
  72                le32_to_cpu(lmmj->lmmj_md.lmm_magic),
  73                le32_to_cpu(lmmj->lmmj_md.lmm_pattern));
  74         CDEBUG(level,"stripe_size %u, stripe_count %u extent_count %u \n",
  75                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_size),
  76                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_count),
  77                le32_to_cpu(lmmj->lmmj_extent_count));
  78 }
  79
  80 #define LMM_ASSERT(test)                                                \
  81 do {                                                                    \
  82         if (!(test)) lov_dump_lmm(D_ERROR, lmm);                        \
  83         LASSERT(test); /* so we know what assertion failed */           \
  84 } while(0)
  85
  86 /* Pack LOV object metadata for disk storage.  It is packed in LE byte
  87  * order and is opaque to the networking layer.
  88  *
  89  * XXX In the future, this will be enhanced to get the EA size from the
  90  *     underlying OSC device(s) to get their EA sizes so we can stack
  91  *     LOVs properly.  For now lov_mds_md_size() just assumes one obd_id
  92  *     per stripe.
  93  */
  94 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
  95                struct lov_stripe_md *lsm)
  96 {
  97         struct obd_device *obd = class_exp2obd(exp);
  98         struct lov_obd *lov = &obd->u.lov;
  99         struct lov_oinfo *loi;
 100         struct lov_mds_md *lmm;
 101         int stripe_count = lov->desc.ld_tgt_count;
 102         int lmm_size;
 103         int i;
 104         ENTRY;
 105
 106         if (lsm) {
 107                 if (lsm->lsm_magic != LOV_MAGIC) {
 108                         CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X\n",
 109                                lsm->lsm_magic, LOV_MAGIC);
 110                         RETURN(-EINVAL);
 111                 }
 112                 if (!lmmp) {
 113                         stripe_count = lov_get_stripecnt(lov, lsm->lsm_stripe_count);
 114                         lsm->lsm_stripe_count = stripe_count;
 115                 } else {
 116                         stripe_count = lsm->lsm_stripe_count;
 117                 }
 118         }
 119
 120         /* XXX LOV STACKING call into osc for sizes */
 121         lmm_size = lov_mds_md_size(stripe_count);
 122
 123         if (!lmmp)
 124                 RETURN(lmm_size);
 125
 126         if (*lmmp && !lsm) {
 127                 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
 128                 OBD_FREE(*lmmp, lov_mds_md_size(stripe_count));
 129                 *lmmp = NULL;
 130                 RETURN(0);
 131         }
 132
 133         if (!*lmmp) {
 134                 OBD_ALLOC(*lmmp, lmm_size);
 135                 if (!*lmmp)
 136                         RETURN(-ENOMEM);
 137         }
 138
 139         lmm = *lmmp;
 140         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC); /* only write new format */
 141
 142         if (!lsm)
 143                 RETURN(lmm_size);
 144
 145         lmm->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 146         lmm->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
 147         lmm->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 148         lmm->lmm_stripe_count = cpu_to_le32(stripe_count);
 149         lmm->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 150
 151         for (i = 0; i < stripe_count; i++) {
 152                 loi = lsm->lsm_oinfo[i];
 153                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 154                 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
 155                          lmm->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
 156                 lmm->lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
 157                 lmm->lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
 158                 lmm->lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 159                 lmm->lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 160         }
 161
 162         RETURN(lmm_size);
 163 }
 164
 165 /* Find the max stripecount we should use */
 166 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
 167 {
 168         if (!stripe_count)
 169                 stripe_count = lov->desc.ld_default_stripe_count;
 170         if (stripe_count > lov->desc.ld_active_tgt_count)
 171                 stripe_count = lov->desc.ld_active_tgt_count;
 172         if (!stripe_count)
 173                 stripe_count = 1;
 174         /* for now, we limit the stripe count directly, when bug 4424 is
 175          * fixed this needs to be somewhat dynamic based on whether ext3
 176          * can handle larger EA sizes. */
 177         if (stripe_count > LOV_MAX_STRIPE_COUNT)
 178                 stripe_count = LOV_MAX_STRIPE_COUNT;
 179
 180         return stripe_count;
 181 }
 182
 183
 184 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
 185 {
 186         int rc;
 187
 188         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 189                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping V1 LMM:\n",
 190                        le32_to_cpu(*(__u32 *)lmm));
 191                 lov_dump_lmm_v1(D_WARNING, lmm);
 192                 return -EINVAL;
 193         }
 194         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 195                                      lmm_bytes, stripe_count);
 196         return rc;
 197 }
 198
 199 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
 200                       int pattern, int magic)
 201 {
 202         int i, lsm_size;
 203         ENTRY;
 204
 205         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 206
 207         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 208         if (!*lsmp) {
 209                 CERROR("can't allocate lsmp, stripe_count %d\n", stripe_count);
 210                 RETURN(-ENOMEM);
 211         }
 212
 213         spin_lock_init(&(*lsmp)->lsm_lock);
 214         (*lsmp)->lsm_magic = magic;
 215         (*lsmp)->lsm_stripe_count = stripe_count;
 216         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 217         (*lsmp)->lsm_pattern = pattern;
 218         (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 219
 220         for (i = 0; i < stripe_count; i++)
 221                 loi_init((*lsmp)->lsm_oinfo[i]);
 222
 223         RETURN(lsm_size);
 224 }
 225
 226 void lov_free_memmd(struct lov_stripe_md **lsmp)
 227 {
 228         struct lov_stripe_md *lsm = *lsmp;
 229
 230         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 231         lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 232
 233         *lsmp = NULL;
 234 }
 235
 236
 237 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 238  * order and is opaque to the networking layer.
 239  */
 240 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 241                  struct lov_mds_md *lmm, int lmm_bytes)
 242 {
 243         struct obd_device *obd = class_exp2obd(exp);
 244         struct lov_obd *lov = &obd->u.lov;
 245         int rc = 0, stripe_count, lsm_size;
 246         __u32 magic;
 247         ENTRY;
 248
 249         /* If passed an MDS struct use values from there, otherwise defaults */
 250         if (lmm) {
 251                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 252                 if (rc)
 253                         RETURN(rc);
 254                 magic = le32_to_cpu(lmm->lmm_magic);
 255         } else {
 256                 stripe_count = lov_get_stripecnt(lov, 0);
 257                 magic = LOV_MAGIC;
 258         }
 259
 260         /* If we aren't passed an lsmp struct, we just want the size */
 261         if (!lsmp) {
 262                 /* XXX LOV STACKING call into osc for sizes */
 263                 LBUG();
 264                 RETURN(lov_stripe_md_size(stripe_count));
 265         }
 266
 267         /* If we are passed an allocated struct but nothing to unpack, free */
 268         if (*lsmp && !lmm) {
 269                 lov_free_memmd(lsmp);
 270                 RETURN(0);
 271         }
 272
 273         lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
 274                                    magic);
 275         if (lsm_size < 0)
 276                 RETURN(lsm_size);
 277
 278         /* If we are passed a pointer but nothing to unpack, we only alloc */
 279         if (!lmm)
 280                 RETURN(lsm_size);
 281
 282         LASSERT(lsm_op_find(magic) != NULL);
 283         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 284         if (rc) {
 285                 lov_free_memmd(lsmp);
 286                 RETURN(rc);
 287         }
 288
 289         RETURN(lsm_size);
 290 }
 291
 292 /* Configure object striping information on a new file.
 293  *
 294  * @lmmu is a pointer to a user struct with one or more of the fields set to
 295  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 296  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 297  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 298  */
 299 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 300                   struct lov_user_md *lump)
 301 {
 302         struct obd_device *obd = class_exp2obd(exp);
 303         struct lov_obd *lov = &obd->u.lov;
 304         struct lov_user_md lum;
 305         int stripe_count;
 306         int rc;
 307         ENTRY;
 308
 309         rc = copy_from_user(&lum, lump, sizeof(lum));
 310         if (rc)
 311                 RETURN(-EFAULT);
 312
 313         if (lum.lmm_magic != LOV_USER_MAGIC) {
 314                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 315                         lustre_swab_lov_user_md(&lum);
 316                 } else {
 317                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 318                                " %#08x != %#08x\n",
 319                                lum.lmm_magic, LOV_USER_MAGIC);
 320                         RETURN(-EINVAL);
 321                 }
 322         }
 323
 324         if (lum.lmm_pattern == 0) {
 325                 lum.lmm_pattern = lov->desc.ld_pattern ?
 326                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 327         }
 328
 329         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 330                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 331                        lum.lmm_pattern);
 332                 RETURN(-EINVAL);
 333         }
 334
 335         /* 64kB is the largest common page size we see (ia64), and matches the
 336          * check in lfs */
 337         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 338                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 339                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 340                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 341         }
 342
 343         if ((lum.lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
 344             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 345                 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
 346                        lum.lmm_stripe_offset, lov->desc.ld_tgt_count);
 347                 RETURN(-EINVAL);
 348         }
 349         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 350
 351         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0U) {
 352                 CDEBUG(D_IOCTL, "stripe width %ux%u exceeds %u bytes\n",
 353                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0U);
 354                 RETURN(-EINVAL);
 355         }
 356
 357         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 358
 359         if (rc < 0)
 360                 RETURN(rc);
 361
 362         (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lum.lmm_stripe_offset;
 363         (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 364
 365         RETURN(0);
 366 }
 367
 368 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
 369               struct lov_user_md *lump)
 370 {
 371         int i;
 372         int rc;
 373         struct obd_export *oexp;
 374         struct lov_obd *lov = &exp->exp_obd->u.lov;
 375         obd_id last_id = 0;
 376
 377         ENTRY;
 378         for (i = 0; i < lump->lmm_stripe_count; i++) {
 379                 __u32 len = sizeof(last_id);
 380                 oexp = lov->lov_tgts[lump->lmm_objects[i].l_ost_idx]->ltd_exp;
 381                 rc = obd_get_info(oexp, strlen("last_id"), "last_id",
 382                                   &len, &last_id);
 383                 if (rc)
 384                         RETURN(rc);
 385                 if (lump->lmm_objects[i].l_object_id > last_id) {
 386                         CERROR("Setting EA for object > than last id on "
 387                                "ost idx %d "LPD64" > "LPD64" \n",
 388                                lump->lmm_objects[i].l_ost_idx,
 389                                lump->lmm_objects[i].l_object_id, last_id);
 390                         RETURN(-EINVAL);
 391                 }
 392         }
 393
 394         rc = lov_setstripe(exp, lsmp, lump);
 395         if (rc)
 396                 RETURN(rc);
 397
 398         for (i = 0; i < lump->lmm_stripe_count; i++) {
 399                 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
 400                         lump->lmm_objects[i].l_ost_idx;
 401                 (*lsmp)->lsm_oinfo[i]->loi_id = lump->lmm_objects[i].l_object_id;
 402                 (*lsmp)->lsm_oinfo[i]->loi_gr = lump->lmm_objects[i].l_object_gr;
 403         }
 404         RETURN(0);
 405 }
 406
 407
 408 /* Retrieve object striping information.
 409  *
 410  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 411  * the maximum number of OST indices which will fit in the user buffer.
 412  * lmm_magic must be LOV_USER_MAGIC.
 413  */
 414 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 415                   struct lov_user_md *lump)
 416 {
 417         struct lov_user_md lum;
 418         struct lov_mds_md *lmmk = NULL;
 419         int rc, lmm_size;
 420         ENTRY;
 421
 422         if (!lsm)
 423                 RETURN(-ENODATA);
 424
 425         rc = copy_from_user(&lum, lump, sizeof(lum));
 426         if (rc)
 427                 RETURN(-EFAULT);
 428
 429         if (lum.lmm_magic != LOV_USER_MAGIC)
 430                 RETURN(-EINVAL);
 431
 432         rc = lov_packmd(exp, &lmmk, lsm);
 433         if (rc < 0)
 434                 RETURN(rc);
 435         lmm_size = rc;
 436         rc = 0;
 437
 438         /* FIXME: Bug 1185 - copy fields properly when structs change */
 439         LASSERT(sizeof(lum) == sizeof(*lmmk));
 440         LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lmmk->lmm_objects[0]));
 441
 442         /* User wasn't expecting this many OST entries */
 443         if (lum.lmm_stripe_count == 0) {
 444                 if (copy_to_user(lump, lmmk, sizeof(lum)))
 445                         rc = -EFAULT;
 446         } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 447                 rc = -EOVERFLOW;
 448         } else if (copy_to_user(lump, lmmk, lmm_size)) {
 449                 rc = -EFAULT;
 450         }
 451
 452         obd_free_diskmd(exp, &lmmk);
 453
 454         RETURN(rc);
 455 }