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