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