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_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; i < stripe_count; i++) {
 138                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 139
 140                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 141                 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
 142                          lmm->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
 143                 lmm->lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
 144                 lmm->lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
 145                 lmm->lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 146                 lmm->lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 147         }
 148
 149         RETURN(lmm_size);
 150 }
 151
 152 /* Find the max stripecount we should use */
 153 int lov_get_stripecnt(struct lov_obd *lov, int stripe_count)
 154 {
 155         if (!stripe_count)
 156                 stripe_count = lov->desc.ld_default_stripe_count;
 157         if (stripe_count > lov->desc.ld_active_tgt_count)
 158                 stripe_count = lov->desc.ld_active_tgt_count;
 159         if (!stripe_count)
 160                 stripe_count = 1;
 161         /* for now, we limit the stripe count directly, when bug 4424 is
 162          * fixed this needs to be somewhat dynamic based on whether ext3
 163          * can handle larger EA sizes. */
 164         if (stripe_count > LOV_MAX_STRIPE_COUNT)
 165                 stripe_count = LOV_MAX_STRIPE_COUNT;
 166
 167         return stripe_count;
 168 }
 169
 170
 171 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
 172 {
 173         int rc;
 174
 175         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 176                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping V1 LMM:\n",
 177                        le32_to_cpu(*(__u32 *)lmm));
 178                 lov_dump_lmm_v1(D_WARNING, lmm);
 179                 return -EINVAL;
 180         }
 181         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 182                                      lmm_bytes, stripe_count);
 183         return rc;
 184 }
 185
 186 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
 187                       int pattern, int magic)
 188 {
 189         int i, lsm_size;
 190         ENTRY;
 191
 192         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 193
 194         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 195         if (!*lsmp) {
 196                 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
 197                 RETURN(-ENOMEM);
 198         }
 199
 200         spin_lock_init(&(*lsmp)->lsm_lock);
 201         (*lsmp)->lsm_magic = magic;
 202         (*lsmp)->lsm_stripe_count = stripe_count;
 203         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 204         (*lsmp)->lsm_pattern = pattern;
 205         (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 206
 207         for (i = 0; i < stripe_count; i++)
 208                 loi_init((*lsmp)->lsm_oinfo[i]);
 209
 210         RETURN(lsm_size);
 211 }
 212
 213 void lov_free_memmd(struct lov_stripe_md **lsmp)
 214 {
 215         struct lov_stripe_md *lsm = *lsmp;
 216
 217         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 218         lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 219
 220         *lsmp = NULL;
 221 }
 222
 223
 224 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 225  * order and is opaque to the networking layer.
 226  */
 227 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 228                  struct lov_mds_md *lmm, int lmm_bytes)
 229 {
 230         struct obd_device *obd = class_exp2obd(exp);
 231         struct lov_obd *lov = &obd->u.lov;
 232         int rc = 0, stripe_count, lsm_size;
 233         __u32 magic;
 234         ENTRY;
 235
 236         /* If passed an MDS struct use values from there, otherwise defaults */
 237         if (lmm) {
 238                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 239                 if (rc)
 240                         RETURN(rc);
 241                 magic = le32_to_cpu(lmm->lmm_magic);
 242         } else {
 243                 stripe_count = lov_get_stripecnt(lov, 0);
 244                 magic = LOV_MAGIC;
 245         }
 246
 247         /* If we aren't passed an lsmp struct, we just want the size */
 248         if (!lsmp) {
 249                 /* XXX LOV STACKING call into osc for sizes */
 250                 LBUG();
 251                 RETURN(lov_stripe_md_size(stripe_count));
 252         }
 253         /* If we are passed an allocated struct but nothing to unpack, free */
 254         if (*lsmp && !lmm) {
 255                 lov_free_memmd(lsmp);
 256                 RETURN(0);
 257         }
 258
 259         lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
 260                                    magic);
 261         if (lsm_size < 0)
 262                 RETURN(lsm_size);
 263
 264         /* If we are passed a pointer but nothing to unpack, we only alloc */
 265         if (!lmm)
 266                 RETURN(lsm_size);
 267
 268         LASSERT(lsm_op_find(magic) != NULL);
 269         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 270         if (rc) {
 271                 lov_free_memmd(lsmp);
 272                 RETURN(rc);
 273         }
 274
 275         RETURN(lsm_size);
 276 }
 277
 278 static int __lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 279                            struct lov_user_md *lump)
 280 {
 281         struct obd_device *obd = class_exp2obd(exp);
 282         struct lov_obd *lov = &obd->u.lov;
 283         struct lov_user_md lum;
 284         int stripe_count;
 285         int rc;
 286         ENTRY;
 287
 288         rc = copy_from_user(&lum, lump, sizeof(lum));
 289         if (rc)
 290                 RETURN(-EFAULT);
 291
 292         if (lum.lmm_magic != LOV_USER_MAGIC) {
 293                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 294                         lustre_swab_lov_user_md(&lum);
 295                 } else {
 296                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 297                                " %#08x != %#08x\n",
 298                                lum.lmm_magic, LOV_USER_MAGIC);
 299                         RETURN(-EINVAL);
 300                 }
 301         }
 302
 303         if (lum.lmm_pattern == 0) {
 304                 lum.lmm_pattern = lov->desc.ld_pattern ?
 305                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 306         }
 307
 308         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 309                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 310                        lum.lmm_pattern);
 311                 RETURN(-EINVAL);
 312         }
 313
 314         /* 64kB is the largest common page size we see (ia64), and matches the
 315          * check in lfs */
 316         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 317                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 318                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 319                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 320         }
 321
 322         if ((lum.lmm_stripe_offset >= lov->desc.ld_active_tgt_count) &&
 323             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 324                 CDEBUG(D_IOCTL, "stripe offset %u > number of active OSTs %u\n",
 325                        lum.lmm_stripe_offset, lov->desc.ld_active_tgt_count);
 326                 RETURN(-EINVAL);
 327         }
 328         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 329
 330         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0UL) {
 331                 CDEBUG(D_IOCTL, "stripe width %ux%i exeeds %lu bytes\n",
 332                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0UL);
 333                 RETURN(-EINVAL);
 334         }
 335
 336         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 337
 338         if (rc >= 0) {
 339                 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lum.lmm_stripe_offset;
 340                 (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 341                 rc = 0;
 342         }
 343
 344         RETURN(0);
 345 }
 346
 347 /* Configure object striping information on a new file.
 348  *
 349  * @lmmu is a pointer to a user struct with one or more of the fields set to
 350  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 351  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 352  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 353  */
 354 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 355                   struct lov_user_md *lump)
 356 {
 357         int rc;
 358         mm_segment_t seg;
 359
 360         seg = get_fs();
 361         set_fs(KERNEL_DS);
 362
 363         rc = __lov_setstripe(exp, lsmp, lump);
 364         set_fs(seg);
 365         RETURN(rc);
 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         /*
 418          * XXX huge struct allocated on stack.
 419          */
 420         struct lov_user_md lum;
 421         struct lov_mds_md *lmmk = NULL;
 422         int rc, lmm_size;
 423         mm_segment_t seg;
 424         ENTRY;
 425
 426         if (!lsm)
 427                 RETURN(-ENODATA);
 428
 429         /*
 430          * "Switch to kernel segment" to allow copying from kernel space by
 431          * copy_{to,from}_user().
 432          */
 433         seg = get_fs();
 434         set_fs(KERNEL_DS);
 435         rc = copy_from_user(&lum, lump, sizeof(lum));
 436         if (rc)
 437                 rc = -EFAULT;
 438         else if (lum.lmm_magic != LOV_USER_MAGIC)
 439                 rc = -EINVAL;
 440         else {
 441                 rc = lov_packmd(exp, &lmmk, lsm);
 442                 if (rc < 0)
 443                         RETURN(rc);
 444                 lmm_size = rc;
 445                 rc = 0;
 446
 447                 /* FIXME: Bug 1185 - copy fields properly when structs change */
 448                 CLASSERT(sizeof lum == sizeof *lmmk);
 449                 CLASSERT(sizeof lum.lmm_objects[0] ==
 450                          sizeof lmmk->lmm_objects[0]);
 451
 452                 /* User wasn't expecting this many OST entries */
 453                 if (lum.lmm_stripe_count == 0) {
 454                         if (copy_to_user(lump, lmmk, sizeof lum))
 455                                 rc = -EFAULT;
 456                 } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 457                         rc = -EOVERFLOW;
 458                 } else if (copy_to_user(lump, lmmk, sizeof lum))
 459                         rc = -EFAULT;
 460
 461                 obd_free_diskmd(exp, &lmmk);
 462         }
 463         set_fs(seg);
 464         RETURN(rc);
 465 }