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 static int __lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 280                            struct lov_user_md *lump)
 281 {
 282         struct obd_device *obd = class_exp2obd(exp);
 283         struct lov_obd *lov = &obd->u.lov;
 284         struct lov_user_md lum;
 285         int stripe_count;
 286         int rc;
 287         ENTRY;
 288
 289         rc = copy_from_user(&lum, lump, sizeof(lum));
 290         if (rc)
 291                 RETURN(-EFAULT);
 292
 293         if (lum.lmm_magic != LOV_USER_MAGIC) {
 294                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 295                         lustre_swab_lov_user_md(&lum);
 296                 } else {
 297                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 298                                " %#08x != %#08x\n",
 299                                lum.lmm_magic, LOV_USER_MAGIC);
 300                         RETURN(-EINVAL);
 301                 }
 302         }
 303
 304         if (lum.lmm_pattern == 0) {
 305                 lum.lmm_pattern = lov->desc.ld_pattern ?
 306                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 307         }
 308
 309         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 310                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 311                        lum.lmm_pattern);
 312                 RETURN(-EINVAL);
 313         }
 314
 315         /* 64kB is the largest common page size we see (ia64), and matches the
 316          * check in lfs */
 317         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 318                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 319                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 320                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 321         }
 322
 323         if ((lum.lmm_stripe_offset >= lov->desc.ld_active_tgt_count) &&
 324             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 325                 CDEBUG(D_IOCTL, "stripe offset %u > number of active OSTs %u\n",
 326                        lum.lmm_stripe_offset, lov->desc.ld_active_tgt_count);
 327                 RETURN(-EINVAL);
 328         }
 329         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 330
 331         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0UL) {
 332                 CDEBUG(D_IOCTL, "stripe width %ux%u > %lu on 32-bit system\n",
 333                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0UL);
 334                 RETURN(-EINVAL);
 335         }
 336
 337         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 338
 339         if (rc >= 0) {
 340                 (*lsmp)->lsm_oinfo[0].loi_ost_idx = lum.lmm_stripe_offset;
 341                 (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 342                 rc = 0;
 343         }
 344
 345         RETURN(0);
 346 }
 347
 348 /* Configure object striping information on a new file.
 349  *
 350  * @lmmu is a pointer to a user struct with one or more of the fields set to
 351  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 352  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 353  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 354  */
 355 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 356                   struct lov_user_md *lump)
 357 {
 358         int rc;
 359         mm_segment_t seg;
 360
 361         seg = get_fs();
 362         set_fs(KERNEL_DS);
 363
 364         rc = __lov_setstripe(exp, lsmp, lump);
 365         set_fs(seg);
 366         RETURN(rc);
 367 }
 368
 369 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
 370               struct lov_user_md *lump)
 371 {
 372         int i;
 373         int rc;
 374         struct obd_export *oexp;
 375         struct lov_obd *lov = &exp->exp_obd->u.lov;
 376         obd_id last_id = 0;
 377
 378         ENTRY;
 379         for (i = 0; i < lump->lmm_stripe_count; i++) {
 380                 __u32 len = sizeof(last_id);
 381                 oexp = lov->lov_tgts[lump->lmm_objects[i].l_ost_idx]->ltd_exp;
 382                 rc = obd_get_info(oexp, strlen("last_id"), "last_id",
 383                                   &len, &last_id);
 384                 if (rc)
 385                         RETURN(rc);
 386                 if (lump->lmm_objects[i].l_object_id > last_id) {
 387                         CERROR("Setting EA for object > than last id on "
 388                                "ost idx %d "LPD64" > "LPD64" \n",
 389                                lump->lmm_objects[i].l_ost_idx,
 390                                lump->lmm_objects[i].l_object_id, last_id);
 391                         RETURN(-EINVAL);
 392                 }
 393         }
 394
 395         rc = lov_setstripe(exp, lsmp, lump);
 396         if (rc)
 397                 RETURN(rc);
 398
 399         for (i = 0; i < lump->lmm_stripe_count; i++) {
 400                 (*lsmp)->lsm_oinfo[i].loi_ost_idx =
 401                         lump->lmm_objects[i].l_ost_idx;
 402                 (*lsmp)->lsm_oinfo[i].loi_id = lump->lmm_objects[i].l_object_id;
 403                 (*lsmp)->lsm_oinfo[i].loi_gr = lump->lmm_objects[i].l_object_gr;
 404         }
 405         RETURN(0);
 406 }
 407
 408
 409 /* Retrieve object striping information.
 410  *
 411  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 412  * the maximum number of OST indices which will fit in the user buffer.
 413  * lmm_magic must be LOV_USER_MAGIC.
 414  */
 415 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 416                   struct lov_user_md *lump)
 417 {
 418         /*
 419          * XXX huge struct allocated on stack.
 420          */
 421         struct lov_user_md lum;
 422         struct lov_mds_md *lmmk = NULL;
 423         int rc, lmm_size;
 424         mm_segment_t seg;
 425         ENTRY;
 426
 427         if (!lsm)
 428                 RETURN(-ENODATA);
 429
 430         /*
 431          * "Switch to kernel segment" to allow copying from kernel space by
 432          * copy_{to,from}_user().
 433          */
 434         seg = get_fs();
 435         set_fs(KERNEL_DS);
 436         rc = copy_from_user(&lum, lump, sizeof(lum));
 437         if (rc)
 438                 rc = -EFAULT;
 439         else if (lum.lmm_magic != LOV_USER_MAGIC)
 440                 rc = -EINVAL;
 441         else {
 442                 rc = lov_packmd(exp, &lmmk, lsm);
 443                 if (rc < 0)
 444                         RETURN(rc);
 445                 lmm_size = rc;
 446                 rc = 0;
 447
 448                 /* FIXME: Bug 1185 - copy fields properly when structs change */
 449                 CLASSERT(sizeof lum == sizeof *lmmk);
 450                 CLASSERT(sizeof lum.lmm_objects[0] ==
 451                          sizeof lmmk->lmm_objects[0]);
 452
 453                 /* User wasn't expecting this many OST entries */
 454                 if (lum.lmm_stripe_count == 0) {
 455                         if (copy_to_user(lump, lmmk, sizeof lum))
 456                                 rc = -EFAULT;
 457                 } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 458                         rc = -EOVERFLOW;
 459                 } else if (copy_to_user(lump, lmmk, sizeof lum))
 460                         rc = -EFAULT;
 461
 462                 obd_free_diskmd(exp, &lmmk);
 463         }
 464         set_fs(seg);
 465         RETURN(rc);
 466 }