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