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_mds_md *lmm;
 100         int stripe_count = lov->desc.ld_tgt_count;
 101         int lmm_size;
 102         int i;
 103         ENTRY;
 104
 105         if (lsm) {
 106                 if (lsm->lsm_magic != LOV_MAGIC) {
 107                         CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X\n",
 108                                lsm->lsm_magic, LOV_MAGIC);
 109                         RETURN(-EINVAL);
 110                 }
 111                 /* If we are just sizing the EA, limit the stripe count
 112                  * to the actual number of OSTs in this filesystem. */
 113                 if (!lmmp) {
 114                         stripe_count = lov_get_stripecnt(lov, lsm->lsm_stripe_count);
 115                         lsm->lsm_stripe_count = stripe_count;
 116                 } else {
 117                         stripe_count = lsm->lsm_stripe_count;
 118                 }
 119         }
 120
 121         /* XXX LOV STACKING call into osc for sizes */
 122         lmm_size = lov_mds_md_size(stripe_count);
 123
 124         if (!lmmp)
 125                 RETURN(lmm_size);
 126
 127         if (*lmmp && !lsm) {
 128                 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
 129                 OBD_FREE(*lmmp, lov_mds_md_size(stripe_count));
 130                 *lmmp = NULL;
 131                 RETURN(0);
 132         }
 133
 134         if (!*lmmp) {
 135                 OBD_ALLOC(*lmmp, lmm_size);
 136                 if (!*lmmp)
 137                         RETURN(-ENOMEM);
 138         }
 139
 140         lmm = *lmmp;
 141         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC); /* only write new format */
 142
 143         if (!lsm)
 144                 RETURN(lmm_size);
 145
 146         lmm->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 147         lmm->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
 148         lmm->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 149         lmm->lmm_stripe_count = cpu_to_le32(stripe_count);
 150         lmm->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 151
 152         for (i = 0; i < stripe_count; i++) {
 153                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 154
 155                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 156                 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
 157                          lmm->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
 158                 lmm->lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
 159                 lmm->lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
 160                 lmm->lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 161                 lmm->lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 162         }
 163
 164         RETURN(lmm_size);
 165 }
 166
 167 /* Find the max stripecount we should use */
 168 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
 169 {
 170         if (!stripe_count)
 171                 stripe_count = lov->desc.ld_default_stripe_count;
 172         if (stripe_count > lov->desc.ld_active_tgt_count)
 173                 stripe_count = lov->desc.ld_active_tgt_count;
 174         if (!stripe_count)
 175                 stripe_count = 1;
 176         /* for now, we limit the stripe count directly, when bug 4424 is
 177          * fixed this needs to be somewhat dynamic based on whether ext3
 178          * can handle larger EA sizes. */
 179         if (stripe_count > LOV_MAX_STRIPE_COUNT)
 180                 stripe_count = LOV_MAX_STRIPE_COUNT;
 181
 182         return stripe_count;
 183 }
 184
 185
 186 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
 187 {
 188         int rc;
 189
 190         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 191                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping V1 LMM:\n",
 192                        le32_to_cpu(*(__u32 *)lmm));
 193                 lov_dump_lmm_v1(D_WARNING, lmm);
 194                 return -EINVAL;
 195         }
 196         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 197                                      lmm_bytes, stripe_count);
 198         return rc;
 199 }
 200
 201 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
 202                       int pattern, int magic)
 203 {
 204         int i, lsm_size;
 205         ENTRY;
 206
 207         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 208
 209         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 210         if (!*lsmp) {
 211                 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
 212                 RETURN(-ENOMEM);
 213         }
 214
 215         spin_lock_init(&(*lsmp)->lsm_lock);
 216         (*lsmp)->lsm_magic = magic;
 217         (*lsmp)->lsm_stripe_count = stripe_count;
 218         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 219         (*lsmp)->lsm_pattern = pattern;
 220         (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 221
 222         for (i = 0; i < stripe_count; i++)
 223                 loi_init((*lsmp)->lsm_oinfo[i]);
 224
 225         RETURN(lsm_size);
 226 }
 227
 228 void lov_free_memmd(struct lov_stripe_md **lsmp)
 229 {
 230         struct lov_stripe_md *lsm = *lsmp;
 231
 232         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 233         lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 234
 235         *lsmp = NULL;
 236 }
 237
 238
 239 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 240  * order and is opaque to the networking layer.
 241  */
 242 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 243                  struct lov_mds_md *lmm, int lmm_bytes)
 244 {
 245         struct obd_device *obd = class_exp2obd(exp);
 246         struct lov_obd *lov = &obd->u.lov;
 247         int rc = 0, stripe_count, lsm_size;
 248         __u32 magic;
 249         ENTRY;
 250
 251         /* If passed an MDS struct use values from there, otherwise defaults */
 252         if (lmm) {
 253                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 254                 if (rc)
 255                         RETURN(rc);
 256                 magic = le32_to_cpu(lmm->lmm_magic);
 257         } else {
 258                 stripe_count = lov_get_stripecnt(lov, 0);
 259                 magic = LOV_MAGIC;
 260         }
 261
 262         /* If we aren't passed an lsmp struct, we just want the size */
 263         if (!lsmp) {
 264                 /* XXX LOV STACKING call into osc for sizes */
 265                 LBUG();
 266                 RETURN(lov_stripe_md_size(stripe_count));
 267         }
 268         /* If we are passed an allocated struct but nothing to unpack, free */
 269         if (*lsmp && !lmm) {
 270                 lov_free_memmd(lsmp);
 271                 RETURN(0);
 272         }
 273
 274         lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
 275                                    magic);
 276         if (lsm_size < 0)
 277                 RETURN(lsm_size);
 278
 279         /* If we are passed a pointer but nothing to unpack, we only alloc */
 280         if (!lmm)
 281                 RETURN(lsm_size);
 282
 283         LASSERT(lsm_op_find(magic) != NULL);
 284         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 285         if (rc) {
 286                 lov_free_memmd(lsmp);
 287                 RETURN(rc);
 288         }
 289
 290         RETURN(lsm_size);
 291 }
 292
 293 static int __lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 294                            struct lov_user_md *lump)
 295 {
 296         struct obd_device *obd = class_exp2obd(exp);
 297         struct lov_obd *lov = &obd->u.lov;
 298         struct lov_user_md lum;
 299         int stripe_count;
 300         int rc;
 301         ENTRY;
 302
 303         rc = copy_from_user(&lum, lump, sizeof(lum));
 304         if (rc)
 305                 RETURN(-EFAULT);
 306
 307         if (lum.lmm_magic != LOV_USER_MAGIC) {
 308                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 309                         lustre_swab_lov_user_md(&lum);
 310                 } else {
 311                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 312                                " %#08x != %#08x\n",
 313                                lum.lmm_magic, LOV_USER_MAGIC);
 314                         RETURN(-EINVAL);
 315                 }
 316         }
 317
 318         if (lum.lmm_pattern == 0) {
 319                 lum.lmm_pattern = lov->desc.ld_pattern ?
 320                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 321         }
 322
 323         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 324                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 325                        lum.lmm_pattern);
 326                 RETURN(-EINVAL);
 327         }
 328
 329         /* 64kB is the largest common page size we see (ia64), and matches the
 330          * check in lfs */
 331         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 332                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 333                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 334                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 335         }
 336
 337         if ((lum.lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
 338             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 339                 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
 340                        lum.lmm_stripe_offset, lov->desc.ld_tgt_count);
 341                 RETURN(-EINVAL);
 342         }
 343         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 344
 345         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0UL) {
 346                 CDEBUG(D_IOCTL, "stripe width %ux%i exeeds %lu bytes\n",
 347                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0UL);
 348                 RETURN(-EINVAL);
 349         }
 350
 351         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 352
 353         if (rc >= 0) {
 354                 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lum.lmm_stripe_offset;
 355                 (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 356                 rc = 0;
 357         }
 358
 359         RETURN(0);
 360 }
 361
 362 /* Configure object striping information on a new file.
 363  *
 364  * @lmmu is a pointer to a user struct with one or more of the fields set to
 365  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 366  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 367  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 368  */
 369 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 370                   struct lov_user_md *lump)
 371 {
 372         int rc;
 373         mm_segment_t seg;
 374
 375         seg = get_fs();
 376         set_fs(KERNEL_DS);
 377
 378         rc = __lov_setstripe(exp, lsmp, lump);
 379         set_fs(seg);
 380         RETURN(rc);
 381 }
 382
 383 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
 384               struct lov_user_md *lump)
 385 {
 386         int i;
 387         int rc;
 388         struct obd_export *oexp;
 389         struct lov_obd *lov = &exp->exp_obd->u.lov;
 390         obd_id last_id = 0;
 391
 392         ENTRY;
 393         for (i = 0; i < lump->lmm_stripe_count; i++) {
 394                 __u32 len = sizeof(last_id);
 395                 oexp = lov->lov_tgts[lump->lmm_objects[i].l_ost_idx]->ltd_exp;
 396                 rc = obd_get_info(oexp, strlen("last_id"), "last_id",
 397                                   &len, &last_id);
 398                 if (rc)
 399                         RETURN(rc);
 400                 if (lump->lmm_objects[i].l_object_id > last_id) {
 401                         CERROR("Setting EA for object > than last id on "
 402                                "ost idx %d "LPD64" > "LPD64" \n",
 403                                lump->lmm_objects[i].l_ost_idx,
 404                                lump->lmm_objects[i].l_object_id, last_id);
 405                         RETURN(-EINVAL);
 406                 }
 407         }
 408
 409         rc = lov_setstripe(exp, lsmp, lump);
 410         if (rc)
 411                 RETURN(rc);
 412
 413         for (i = 0; i < lump->lmm_stripe_count; i++) {
 414                 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
 415                         lump->lmm_objects[i].l_ost_idx;
 416                 (*lsmp)->lsm_oinfo[i]->loi_id = lump->lmm_objects[i].l_object_id;
 417                 (*lsmp)->lsm_oinfo[i]->loi_gr = lump->lmm_objects[i].l_object_gr;
 418         }
 419         RETURN(0);
 420 }
 421
 422
 423 /* Retrieve object striping information.
 424  *
 425  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 426  * the maximum number of OST indices which will fit in the user buffer.
 427  * lmm_magic must be LOV_USER_MAGIC.
 428  */
 429 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 430                   struct lov_user_md *lump)
 431 {
 432         /*
 433          * XXX huge struct allocated on stack.
 434          */
 435         struct lov_user_md lum;
 436         struct lov_mds_md *lmmk = NULL;
 437         int rc, lmm_size;
 438         mm_segment_t seg;
 439         ENTRY;
 440
 441         if (!lsm)
 442                 RETURN(-ENODATA);
 443
 444         /*
 445          * "Switch to kernel segment" to allow copying from kernel space by
 446          * copy_{to,from}_user().
 447          */
 448         seg = get_fs();
 449         set_fs(KERNEL_DS);
 450         rc = copy_from_user(&lum, lump, sizeof(lum));
 451         if (rc)
 452                 rc = -EFAULT;
 453         else if (lum.lmm_magic != LOV_USER_MAGIC)
 454                 rc = -EINVAL;
 455         else {
 456                 rc = lov_packmd(exp, &lmmk, lsm);
 457                 if (rc < 0)
 458                         RETURN(rc);
 459                 lmm_size = rc;
 460                 rc = 0;
 461
 462                 /* FIXME: Bug 1185 - copy fields properly when structs change */
 463                 CLASSERT(sizeof lum == sizeof *lmmk);
 464                 CLASSERT(sizeof lum.lmm_objects[0] ==
 465                          sizeof lmmk->lmm_objects[0]);
 466
 467                 /* User wasn't expecting this many OST entries */
 468                 if (lum.lmm_stripe_count == 0) {
 469                         if (copy_to_user(lump, lmmk, sizeof lum))
 470                                 rc = -EFAULT;
 471                 } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 472                         rc = -EOVERFLOW;
 473                 } else if (copy_to_user(lump, lmmk, sizeof lum))
 474                         rc = -EFAULT;
 475
 476                 obd_free_diskmd(exp, &lmmk);
 477         }
 478         set_fs(seg);
 479         RETURN(rc);
 480 }