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  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see [sun.com URL with a
  20  * copy of GPLv2].
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lov/lov_pack.c
  37  *
  38  * (Un)packing of OST/MDS requests
  39  *
  40  * Author: Andreas Dilger <adilger@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_LOV
  44 #ifndef __KERNEL__
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <lustre_net.h>
  49 #include <obd.h>
  50 #include <obd_lov.h>
  51 #include <obd_class.h>
  52 #include <obd_support.h>
  53 #include <lustre/lustre_user.h>
  54
  55 #include "lov_internal.h"
  56
  57 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  58 {
  59         struct lov_ost_data_v1 *lod;
  60         int i;
  61
  62         CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
  63                le64_to_cpu(lmm->lmm_object_id), le32_to_cpu(lmm->lmm_magic),
  64                le32_to_cpu(lmm->lmm_pattern));
  65         CDEBUG(level,"stripe_size %u, stripe_count %u\n",
  66                le32_to_cpu(lmm->lmm_stripe_size),
  67                le32_to_cpu(lmm->lmm_stripe_count));
  68
  69         if (le32_to_cpu(lmm->lmm_stripe_count) <= LOV_V1_INSANE_STRIPE_COUNT) {
  70                 for (i = 0, lod = lmm->lmm_objects;
  71                      i < (int)le32_to_cpu(lmm->lmm_stripe_count); i++, lod++)
  72                          CDEBUG(level,
  73                                 "stripe %u idx %u subobj "LPX64"/"LPX64"\n",
  74                                 i, le32_to_cpu(lod->l_ost_idx),
  75                                 le64_to_cpu(lod->l_object_gr),
  76                                 le64_to_cpu(lod->l_object_id));
  77         } else {
  78                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  79                        le32_to_cpu(lmm->lmm_stripe_count),
  80                        LOV_V1_INSANE_STRIPE_COUNT);
  81         }
  82 }
  83
  84 void lov_dump_lmm_join(int level, struct lov_mds_md_join *lmmj)
  85 {
  86
  87         CDEBUG(level, "objid "LPX64", magic 0x%08X, pattern %#X\n",
  88                le64_to_cpu(lmmj->lmmj_md.lmm_object_id),
  89                le32_to_cpu(lmmj->lmmj_md.lmm_magic),
  90                le32_to_cpu(lmmj->lmmj_md.lmm_pattern));
  91         CDEBUG(level,"stripe_size %u, stripe_count %u extent_count %u \n",
  92                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_size),
  93                le32_to_cpu(lmmj->lmmj_md.lmm_stripe_count),
  94                le32_to_cpu(lmmj->lmmj_extent_count));
  95 }
  96
  97 #define LMM_ASSERT(test)                                                \
  98 do {                                                                    \
  99         if (!(test)) lov_dump_lmm(D_ERROR, lmm);                        \
 100         LASSERT(test); /* so we know what assertion failed */           \
 101 } while(0)
 102
 103 /* Pack LOV object metadata for disk storage.  It is packed in LE byte
 104  * order and is opaque to the networking layer.
 105  *
 106  * XXX In the future, this will be enhanced to get the EA size from the
 107  *     underlying OSC device(s) to get their EA sizes so we can stack
 108  *     LOVs properly.  For now lov_mds_md_size() just assumes one obd_id
 109  *     per stripe.
 110  */
 111 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
 112                struct lov_stripe_md *lsm)
 113 {
 114         struct obd_device *obd = class_exp2obd(exp);
 115         struct lov_obd *lov = &obd->u.lov;
 116         struct lov_mds_md *lmm;
 117         int stripe_count = lov->desc.ld_tgt_count;
 118         int lmm_size;
 119         int i;
 120         ENTRY;
 121
 122         if (lsm) {
 123                 if (lsm->lsm_magic != LOV_MAGIC) {
 124                         CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X\n",
 125                                lsm->lsm_magic, LOV_MAGIC);
 126                         RETURN(-EINVAL);
 127                 }
 128                 /* If we are just sizing the EA, limit the stripe count
 129                  * to the actual number of OSTs in this filesystem. */
 130                 if (!lmmp) {
 131                         stripe_count = lov_get_stripecnt(lov, lsm->lsm_stripe_count);
 132                         lsm->lsm_stripe_count = stripe_count;
 133                 } else {
 134                         stripe_count = lsm->lsm_stripe_count;
 135                 }
 136         }
 137
 138         /* XXX LOV STACKING call into osc for sizes */
 139         lmm_size = lov_mds_md_size(stripe_count);
 140
 141         if (!lmmp)
 142                 RETURN(lmm_size);
 143
 144         if (*lmmp && !lsm) {
 145                 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
 146                 OBD_FREE(*lmmp, lov_mds_md_size(stripe_count));
 147                 *lmmp = NULL;
 148                 RETURN(0);
 149         }
 150
 151         if (!*lmmp) {
 152                 OBD_ALLOC(*lmmp, lmm_size);
 153                 if (!*lmmp)
 154                         RETURN(-ENOMEM);
 155         }
 156
 157         lmm = *lmmp;
 158         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC); /* only write new format */
 159
 160         if (!lsm)
 161                 RETURN(lmm_size);
 162
 163         lmm->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
 164         lmm->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
 165         lmm->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 166         lmm->lmm_stripe_count = cpu_to_le32(stripe_count);
 167         lmm->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 168
 169         for (i = 0; i < stripe_count; i++) {
 170                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 171
 172                 /* XXX LOV STACKING call down to osc_packmd() to do packing */
 173                 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
 174                          lmm->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
 175                 lmm->lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
 176                 lmm->lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
 177                 lmm->lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 178                 lmm->lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 179         }
 180
 181         RETURN(lmm_size);
 182 }
 183
 184 /* Find the max stripecount we should use */
 185 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
 186 {
 187         if (!stripe_count)
 188                 stripe_count = lov->desc.ld_default_stripe_count;
 189         if (stripe_count > lov->desc.ld_active_tgt_count)
 190                 stripe_count = lov->desc.ld_active_tgt_count;
 191         if (!stripe_count)
 192                 stripe_count = 1;
 193         /* for now, we limit the stripe count directly, when bug 4424 is
 194          * fixed this needs to be somewhat dynamic based on whether ext3
 195          * can handle larger EA sizes. */
 196         if (stripe_count > LOV_MAX_STRIPE_COUNT)
 197                 stripe_count = LOV_MAX_STRIPE_COUNT;
 198
 199         return stripe_count;
 200 }
 201
 202
 203 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
 204 {
 205         int rc;
 206
 207         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 208                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping V1 LMM:\n",
 209                        le32_to_cpu(*(__u32 *)lmm));
 210                 lov_dump_lmm_v1(D_WARNING, lmm);
 211                 return -EINVAL;
 212         }
 213         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 214                                      lmm_bytes, stripe_count);
 215         return rc;
 216 }
 217
 218 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
 219                       int pattern, int magic)
 220 {
 221         int i, lsm_size;
 222         ENTRY;
 223
 224         CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
 225
 226         *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
 227         if (!*lsmp) {
 228                 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
 229                 RETURN(-ENOMEM);
 230         }
 231
 232         spin_lock_init(&(*lsmp)->lsm_lock);
 233         (*lsmp)->lsm_magic = magic;
 234         (*lsmp)->lsm_stripe_count = stripe_count;
 235         (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
 236         (*lsmp)->lsm_pattern = pattern;
 237         (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
 238
 239         for (i = 0; i < stripe_count; i++)
 240                 loi_init((*lsmp)->lsm_oinfo[i]);
 241
 242         RETURN(lsm_size);
 243 }
 244
 245 void lov_free_memmd(struct lov_stripe_md **lsmp)
 246 {
 247         struct lov_stripe_md *lsm = *lsmp;
 248
 249         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 250         lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 251
 252         *lsmp = NULL;
 253 }
 254
 255
 256 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 257  * order and is opaque to the networking layer.
 258  */
 259 int lov_unpackmd(struct obd_export *exp,  struct lov_stripe_md **lsmp,
 260                  struct lov_mds_md *lmm, int lmm_bytes)
 261 {
 262         struct obd_device *obd = class_exp2obd(exp);
 263         struct lov_obd *lov = &obd->u.lov;
 264         int rc = 0, stripe_count, lsm_size;
 265         __u32 magic;
 266         ENTRY;
 267
 268         /* If passed an MDS struct use values from there, otherwise defaults */
 269         if (lmm) {
 270                 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
 271                 if (rc)
 272                         RETURN(rc);
 273                 magic = le32_to_cpu(lmm->lmm_magic);
 274         } else {
 275                 stripe_count = lov_get_stripecnt(lov, 0);
 276                 magic = LOV_MAGIC;
 277         }
 278
 279         /* If we aren't passed an lsmp struct, we just want the size */
 280         if (!lsmp) {
 281                 /* XXX LOV STACKING call into osc for sizes */
 282                 LBUG();
 283                 RETURN(lov_stripe_md_size(stripe_count));
 284         }
 285         /* If we are passed an allocated struct but nothing to unpack, free */
 286         if (*lsmp && !lmm) {
 287                 lov_free_memmd(lsmp);
 288                 RETURN(0);
 289         }
 290
 291         lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
 292                                    magic);
 293         if (lsm_size < 0)
 294                 RETURN(lsm_size);
 295
 296         /* If we are passed a pointer but nothing to unpack, we only alloc */
 297         if (!lmm)
 298                 RETURN(lsm_size);
 299
 300         LASSERT(lsm_op_find(magic) != NULL);
 301         rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
 302         if (rc) {
 303                 lov_free_memmd(lsmp);
 304                 RETURN(rc);
 305         }
 306
 307         RETURN(lsm_size);
 308 }
 309
 310 static int __lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 311                            struct lov_user_md *lump)
 312 {
 313         struct obd_device *obd = class_exp2obd(exp);
 314         struct lov_obd *lov = &obd->u.lov;
 315         struct lov_user_md lum;
 316         int stripe_count;
 317         int rc;
 318         ENTRY;
 319
 320         rc = copy_from_user(&lum, lump, sizeof(lum));
 321         if (rc)
 322                 RETURN(-EFAULT);
 323
 324         if (lum.lmm_magic != LOV_USER_MAGIC) {
 325                 if (lum.lmm_magic == __swab32(LOV_USER_MAGIC)) {
 326                         lustre_swab_lov_user_md(&lum);
 327                 } else {
 328                         CDEBUG(D_IOCTL, "bad userland LOV MAGIC:"
 329                                " %#08x != %#08x\n",
 330                                lum.lmm_magic, LOV_USER_MAGIC);
 331                         RETURN(-EINVAL);
 332                 }
 333         }
 334
 335         if (lum.lmm_pattern == 0) {
 336                 lum.lmm_pattern = lov->desc.ld_pattern ?
 337                         lov->desc.ld_pattern : LOV_PATTERN_RAID0;
 338         }
 339
 340         if (lum.lmm_pattern != LOV_PATTERN_RAID0) {
 341                 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
 342                        lum.lmm_pattern);
 343                 RETURN(-EINVAL);
 344         }
 345
 346         /* 64kB is the largest common page size we see (ia64), and matches the
 347          * check in lfs */
 348         if (lum.lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 349                 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
 350                        lum.lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
 351                 lum.lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
 352         }
 353
 354         if ((lum.lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
 355             (lum.lmm_stripe_offset != (typeof(lum.lmm_stripe_offset))(-1))) {
 356                 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
 357                        lum.lmm_stripe_offset, lov->desc.ld_tgt_count);
 358                 RETURN(-EINVAL);
 359         }
 360         stripe_count = lov_get_stripecnt(lov, lum.lmm_stripe_count);
 361
 362         if ((__u64)lum.lmm_stripe_size * stripe_count > ~0UL) {
 363                 CDEBUG(D_IOCTL, "stripe width %ux%i exeeds %lu bytes\n",
 364                        lum.lmm_stripe_size, (int)lum.lmm_stripe_count, ~0UL);
 365                 RETURN(-EINVAL);
 366         }
 367
 368         rc = lov_alloc_memmd(lsmp, stripe_count, lum.lmm_pattern, LOV_MAGIC);
 369
 370         if (rc >= 0) {
 371                 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lum.lmm_stripe_offset;
 372                 (*lsmp)->lsm_stripe_size = lum.lmm_stripe_size;
 373                 rc = 0;
 374         }
 375
 376         RETURN(0);
 377 }
 378
 379 /* Configure object striping information on a new file.
 380  *
 381  * @lmmu is a pointer to a user struct with one or more of the fields set to
 382  * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
 383  * lmm_stripe_offset, and lmm_stripe_pattern.  lmm_magic must be LOV_MAGIC.
 384  * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
 385  */
 386 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
 387                   struct lov_user_md *lump)
 388 {
 389         int rc;
 390         mm_segment_t seg;
 391
 392         seg = get_fs();
 393         set_fs(KERNEL_DS);
 394
 395         rc = __lov_setstripe(exp, lsmp, lump);
 396         set_fs(seg);
 397         RETURN(rc);
 398 }
 399
 400 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
 401               struct lov_user_md *lump)
 402 {
 403         int i;
 404         int rc;
 405         struct obd_export *oexp;
 406         struct lov_obd *lov = &exp->exp_obd->u.lov;
 407         obd_id last_id = 0;
 408
 409         ENTRY;
 410         for (i = 0; i < lump->lmm_stripe_count; i++) {
 411                 __u32 len = sizeof(last_id);
 412                 oexp = lov->lov_tgts[lump->lmm_objects[i].l_ost_idx]->ltd_exp;
 413                 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
 414                                   &len, &last_id, NULL);
 415                 if (rc)
 416                         RETURN(rc);
 417                 if (lump->lmm_objects[i].l_object_id > last_id) {
 418                         CERROR("Setting EA for object > than last id on "
 419                                "ost idx %d "LPD64" > "LPD64" \n",
 420                                lump->lmm_objects[i].l_ost_idx,
 421                                lump->lmm_objects[i].l_object_id, last_id);
 422                         RETURN(-EINVAL);
 423                 }
 424         }
 425
 426         rc = lov_setstripe(exp, lsmp, lump);
 427         if (rc)
 428                 RETURN(rc);
 429
 430         for (i = 0; i < lump->lmm_stripe_count; i++) {
 431                 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
 432                         lump->lmm_objects[i].l_ost_idx;
 433                 (*lsmp)->lsm_oinfo[i]->loi_id = lump->lmm_objects[i].l_object_id;
 434                 (*lsmp)->lsm_oinfo[i]->loi_gr = lump->lmm_objects[i].l_object_gr;
 435         }
 436         RETURN(0);
 437 }
 438
 439
 440 /* Retrieve object striping information.
 441  *
 442  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 443  * the maximum number of OST indices which will fit in the user buffer.
 444  * lmm_magic must be LOV_USER_MAGIC.
 445  */
 446 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
 447                   struct lov_user_md *lump)
 448 {
 449         /*
 450          * XXX huge struct allocated on stack.
 451          */
 452         struct lov_user_md lum;
 453         struct lov_mds_md *lmmk = NULL;
 454         int rc, lmm_size;
 455         mm_segment_t seg;
 456         ENTRY;
 457
 458         if (!lsm)
 459                 RETURN(-ENODATA);
 460
 461         /*
 462          * "Switch to kernel segment" to allow copying from kernel space by
 463          * copy_{to,from}_user().
 464          */
 465         seg = get_fs();
 466         set_fs(KERNEL_DS);
 467         rc = copy_from_user(&lum, lump, sizeof(lum));
 468         if (rc)
 469                 rc = -EFAULT;
 470         else if (lum.lmm_magic != LOV_USER_MAGIC)
 471                 rc = -EINVAL;
 472         else {
 473                 rc = lov_packmd(exp, &lmmk, lsm);
 474                 if (rc < 0)
 475                         RETURN(rc);
 476                 lmm_size = rc;
 477                 rc = 0;
 478
 479                 /* FIXME: Bug 1185 - copy fields properly when structs change */
 480                 CLASSERT(sizeof lum == sizeof *lmmk);
 481                 CLASSERT(sizeof lum.lmm_objects[0] ==
 482                          sizeof lmmk->lmm_objects[0]);
 483
 484                 /* User wasn't expecting this many OST entries */
 485                 if (lum.lmm_stripe_count == 0) {
 486                         if (copy_to_user(lump, lmmk, sizeof lum))
 487                                 rc = -EFAULT;
 488                 } else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count) {
 489                         rc = -EOVERFLOW;
 490                 } else if (copy_to_user(lump, lmmk, sizeof lum))
 491                         rc = -EFAULT;
 492
 493                 obd_free_diskmd(exp, &lmmk);
 494         }
 495         set_fs(seg);
 496         RETURN(rc);
 497 }