lustre/lov/lov_pack.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  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
  45 #include <lustre_net.h>
  46 #include <obd.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49 #include <lustre/lustre_user.h>
  50
  51 #include "lov_internal.h"
  52 #include "lov_cl_internal.h"
  53
  54 void lov_dump_lmm_common(int level, void *lmmp)
  55 {
  56         struct lov_mds_md *lmm = lmmp;
  57         struct ost_id   oi;
  58
  59         lmm_oi_le_to_cpu(&oi, &lmm->lmm_oi);
  60         CDEBUG(level, "objid "DOSTID", magic 0x%08x, pattern %#x\n",
  61                POSTID(&oi), le32_to_cpu(lmm->lmm_magic),
  62                le32_to_cpu(lmm->lmm_pattern));
  63         CDEBUG(level, "stripe_size %u, stripe_count %u, layout_gen %u\n",
  64                le32_to_cpu(lmm->lmm_stripe_size),
  65                le16_to_cpu(lmm->lmm_stripe_count),
  66                le16_to_cpu(lmm->lmm_layout_gen));
  67 }
  68
  69 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
  70                                  int stripe_count)
  71 {
  72         int i;
  73
  74         if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
  75                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  76                        stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
  77                 return;
  78         }
  79
  80         for (i = 0; i < stripe_count; ++i, ++lod) {
  81                 struct ost_id   oi;
  82
  83                 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
  84                 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
  85                        le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
  86         }
  87 }
  88
  89 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  90 {
  91         lov_dump_lmm_common(level, lmm);
  92         lov_dump_lmm_objects(level, lmm->lmm_objects,
  93                              le16_to_cpu(lmm->lmm_stripe_count));
  94 }
  95
  96 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
  97 {
  98         lov_dump_lmm_common(level, lmm);
  99         CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
 100         lov_dump_lmm_objects(level, lmm->lmm_objects,
 101                              le16_to_cpu(lmm->lmm_stripe_count));
 102 }
 103
 104 void lov_dump_lmm(int level, void *lmm)
 105 {
 106         int magic;
 107
 108         magic = le32_to_cpu(((struct lov_mds_md *)lmm)->lmm_magic);
 109         switch (magic) {
 110         case LOV_MAGIC_V1:
 111                 lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)lmm);
 112                 break;
 113         case LOV_MAGIC_V3:
 114                 lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)lmm);
 115                 break;
 116         default:
 117                 CDEBUG(level, "unrecognized lmm_magic %x, assuming %x\n",
 118                        magic, LOV_MAGIC_V1);
 119                 lov_dump_lmm_common(level, lmm);
 120                 break;
 121         }
 122 }
 123
 124 /**
 125  * Pack LOV striping metadata for disk storage format (in little
 126  * endian byte order).
 127  *
 128  * This follows the getxattr() conventions. If \a buf_size is zero
 129  * then return the size needed. If \a buf_size is too small then
 130  * return -ERANGE. Otherwise return the size of the result.
 131  */
 132 ssize_t lov_lsm_pack(const struct lov_stripe_md *lsm, void *buf,
 133                      size_t buf_size)
 134 {
 135         struct lov_mds_md_v1 *lmmv1 = buf;
 136         struct lov_mds_md_v3 *lmmv3 = buf;
 137         struct lov_ost_data_v1 *lmm_objects;
 138         size_t lmm_size;
 139         unsigned int i;
 140         ENTRY;
 141
 142         lmm_size = lov_mds_md_size(lsm->lsm_stripe_count, lsm->lsm_magic);
 143         if (buf_size == 0)
 144                 RETURN(lmm_size);
 145
 146         if (buf_size < lmm_size)
 147                 RETURN(-ERANGE);
 148
 149         /* lmmv1 and lmmv3 point to the same struct and have the
 150          * same first fields
 151          */
 152         lmmv1->lmm_magic = cpu_to_le32(lsm->lsm_magic);
 153         lmm_oi_cpu_to_le(&lmmv1->lmm_oi, &lsm->lsm_oi);
 154         lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
 155         lmmv1->lmm_stripe_count = cpu_to_le16(lsm->lsm_stripe_count);
 156         lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
 157         lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
 158
 159         if (lsm->lsm_magic == LOV_MAGIC_V3) {
 160                 CLASSERT(sizeof(lsm->lsm_pool_name) ==
 161                          sizeof(lmmv3->lmm_pool_name));
 162                 strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
 163                         sizeof(lmmv3->lmm_pool_name));
 164                 lmm_objects = lmmv3->lmm_objects;
 165         } else {
 166                 lmm_objects = lmmv1->lmm_objects;
 167         }
 168
 169         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 170                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
 171
 172                 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
 173                 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 174                 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 175         }
 176
 177         RETURN(lmm_size);
 178 }
 179
 180 /* Find the max stripecount we should use */
 181 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
 182 {
 183         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
 184
 185         if (!stripe_count)
 186                 stripe_count = lov->desc.ld_default_stripe_count;
 187         if (stripe_count > lov->desc.ld_active_tgt_count)
 188                 stripe_count = lov->desc.ld_active_tgt_count;
 189         if (!stripe_count)
 190                 stripe_count = 1;
 191
 192         /* stripe count is based on whether ldiskfs can handle
 193          * larger EA sizes */
 194         if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
 195             lov->lov_ocd.ocd_max_easize)
 196                 max_stripes = lov_mds_md_max_stripe_count(
 197                         lov->lov_ocd.ocd_max_easize, magic);
 198
 199         if (stripe_count > max_stripes)
 200                 stripe_count = max_stripes;
 201
 202         return stripe_count;
 203 }
 204
 205 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
 206 {
 207         int rc;
 208
 209         if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
 210                 char *buffer;
 211                 int sz;
 212
 213                 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
 214                        le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
 215                 sz = lmm_bytes * 2 + 1;
 216                 OBD_ALLOC_LARGE(buffer, sz);
 217                 if (buffer != NULL) {
 218                         int i;
 219
 220                         for (i = 0; i < lmm_bytes; i++)
 221                                 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
 222                         buffer[sz - 1] = '\0';
 223                         CERROR("%s\n", buffer);
 224                         OBD_FREE_LARGE(buffer, sz);
 225                 }
 226                 return -EINVAL;
 227         }
 228         rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
 229                                      lmm_bytes, stripe_count);
 230         return rc;
 231 }
 232
 233 struct lov_stripe_md *lov_lsm_alloc(u16 stripe_count, u32 pattern, u32 magic)
 234 {
 235         struct lov_stripe_md *lsm;
 236         unsigned int i;
 237         ENTRY;
 238
 239         CDEBUG(D_INFO, "alloc lsm, stripe_count %u\n",
 240                (unsigned int)stripe_count);
 241
 242         lsm = lsm_alloc_plain(stripe_count);
 243         if (lsm == NULL) {
 244                 CERROR("cannot allocate LSM stripe_count %u\n",
 245                        (unsigned int)stripe_count);
 246                 RETURN(ERR_PTR(-ENOMEM));
 247         }
 248
 249         atomic_set(&lsm->lsm_refc, 1);
 250         spin_lock_init(&lsm->lsm_lock);
 251         lsm->lsm_magic = magic;
 252         lsm->lsm_stripe_count = stripe_count;
 253         lsm->lsm_maxbytes = LUSTRE_EXT3_STRIPE_MAXBYTES * stripe_count;
 254         lsm->lsm_pattern = pattern;
 255         lsm->lsm_pool_name[0] = '\0';
 256         lsm->lsm_layout_gen = 0;
 257         if (stripe_count > 0)
 258                 lsm->lsm_oinfo[0]->loi_ost_idx = ~0;
 259
 260         for (i = 0; i < stripe_count; i++)
 261                 loi_init(lsm->lsm_oinfo[i]);
 262
 263         RETURN(lsm);
 264 }
 265
 266 int lov_free_memmd(struct lov_stripe_md **lsmp)
 267 {
 268         struct lov_stripe_md *lsm = *lsmp;
 269         int refc;
 270
 271         *lsmp = NULL;
 272         refc = atomic_dec_return(&lsm->lsm_refc);
 273         LASSERT(refc >= 0);
 274         if (refc == 0) {
 275                 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
 276                 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
 277         }
 278         return refc;
 279 }
 280
 281 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 282  * order and is opaque to the networking layer.
 283  */
 284 struct lov_stripe_md *lov_unpackmd(struct lov_obd *lov, struct lov_mds_md *lmm,
 285                                    size_t lmm_size)
 286 {
 287         struct lov_stripe_md *lsm;
 288         u16 stripe_count;
 289         u32 magic;
 290         u32 pattern;
 291         int rc;
 292         ENTRY;
 293
 294         rc = lov_verify_lmm(lmm, lmm_size, &stripe_count);
 295         if (rc != 0)
 296                 RETURN(ERR_PTR(rc));
 297
 298         magic = le32_to_cpu(lmm->lmm_magic);
 299         pattern = le32_to_cpu(lmm->lmm_pattern);
 300
 301         lsm = lov_lsm_alloc(stripe_count, pattern, magic);
 302         if (IS_ERR(lsm))
 303                 RETURN(lsm);
 304
 305         LASSERT(lsm_op_find(magic) != NULL);
 306         rc = lsm_op_find(magic)->lsm_unpackmd(lov, lsm, lmm);
 307         if (rc != 0) {
 308                 lov_free_memmd(&lsm);
 309                 RETURN(ERR_PTR(rc));
 310         }
 311
 312         RETURN(lsm);
 313 }
 314
 315 /* Retrieve object striping information.
 316  *
 317  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 318  * the maximum number of OST indices which will fit in the user buffer.
 319  * lmm_magic must be LOV_USER_MAGIC.
 320  */
 321 int lov_getstripe(struct lov_object *obj, struct lov_stripe_md *lsm,
 322                   struct lov_user_md __user *lump)
 323 {
 324         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
 325         struct lov_mds_md       *lmmk;
 326         struct lov_user_md_v3   lum;
 327         u32                     stripe_count;
 328         size_t                  lum_size;
 329         size_t                  lmmk_size;
 330         ssize_t                 lmm_size;
 331         int                     rc;
 332         ENTRY;
 333
 334         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3) {
 335                 CERROR("bad LSM MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
 336                        lsm->lsm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
 337                 GOTO(out, rc = -EIO);
 338         }
 339
 340         if (!lsm_is_released(lsm))
 341                 stripe_count = lsm->lsm_stripe_count;
 342         else
 343                 stripe_count = 0;
 344
 345         /* we only need the header part from user space to get lmm_magic and
 346          * lmm_stripe_count, (the header part is common to v1 and v3) */
 347         lum_size = sizeof(struct lov_user_md_v1);
 348         if (copy_from_user(&lum, lump, lum_size))
 349                 GOTO(out, rc = -EFAULT);
 350
 351         if (lum.lmm_magic != LOV_USER_MAGIC_V1 &&
 352             lum.lmm_magic != LOV_USER_MAGIC_V3 &&
 353             lum.lmm_magic != LOV_USER_MAGIC_SPECIFIC)
 354                 GOTO(out, rc = -EINVAL);
 355
 356         if (lum.lmm_stripe_count != 0 && lum.lmm_stripe_count < stripe_count) {
 357                 /* Return right size of stripe to user */
 358                 lum.lmm_stripe_count = stripe_count;
 359                 rc = copy_to_user(lump, &lum, lum_size);
 360                 GOTO(out, rc = -EOVERFLOW);
 361         }
 362
 363         lmmk_size = lov_mds_md_size(stripe_count, lsm->lsm_magic);
 364
 365         OBD_ALLOC_LARGE(lmmk, lmmk_size);
 366         if (lmmk == NULL)
 367                 GOTO(out, rc = -ENOMEM);
 368
 369         lmm_size = lov_lsm_pack(lsm, lmmk, lmmk_size);
 370         if (lmm_size < 0)
 371                 GOTO(out_free, rc = lmm_size);
 372
 373         /* FIXME: Bug 1185 - copy fields properly when structs change */
 374         /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
 375         CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
 376         CLASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lmmk->lmm_objects[0]));
 377
 378         if (cpu_to_le32(LOV_MAGIC) != LOV_MAGIC &&
 379             (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1) ||
 380              lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3))) {
 381                 lustre_swab_lov_mds_md(lmmk);
 382                 lustre_swab_lov_user_md_objects(
 383                                 (struct lov_user_ost_data *)lmmk->lmm_objects,
 384                                 lmmk->lmm_stripe_count);
 385         }
 386
 387         if (lum.lmm_magic == LOV_USER_MAGIC) {
 388                 /* User request for v1, we need skip lmm_pool_name */
 389                 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
 390                         memmove(((struct lov_mds_md_v1 *)lmmk)->lmm_objects,
 391                                 ((struct lov_mds_md_v3 *)lmmk)->lmm_objects,
 392                                 lmmk->lmm_stripe_count *
 393                                 sizeof(struct lov_ost_data_v1));
 394                         lmm_size -= LOV_MAXPOOLNAME;
 395                 }
 396         } else {
 397                 /* if v3 we just have to update the lum_size */
 398                 lum_size = sizeof(struct lov_user_md_v3);
 399         }
 400
 401         /* User wasn't expecting this many OST entries */
 402         if (lum.lmm_stripe_count == 0)
 403                 lmm_size = lum_size;
 404         else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
 405                 GOTO(out_free, rc = -EOVERFLOW);
 406         /*
 407          * Have a difference between lov_mds_md & lov_user_md.
 408          * So we have to re-order the data before copy to user.
 409          */
 410         lum.lmm_stripe_count = lmmk->lmm_stripe_count;
 411         lum.lmm_layout_gen = lmmk->lmm_layout_gen;
 412         ((struct lov_user_md *)lmmk)->lmm_layout_gen = lum.lmm_layout_gen;
 413         ((struct lov_user_md *)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
 414         if (copy_to_user(lump, lmmk, lmm_size))
 415                 GOTO(out_free, rc = -EFAULT);
 416
 417         GOTO(out_free, rc = 0);
 418 out_free:
 419         OBD_FREE_LARGE(lmmk, lmmk_size);
 420 out:
 421         return rc;
 422 }