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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lov/lov_pack.c
  33  *
  34  * (Un)packing of OST/MDS requests
  35  *
  36  * Author: Andreas Dilger <adilger@clusterfs.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_LOV
  40
  41 #include <lustre/lustre_idl.h>
  42 #include <lustre/lustre_user.h>
  43
  44 #include <lustre_net.h>
  45 #include <lustre_swab.h>
  46 #include <obd.h>
  47 #include <obd_class.h>
  48 #include <obd_support.h>
  49
  50 #include "lov_cl_internal.h"
  51 #include "lov_internal.h"
  52
  53 void lov_dump_lmm_common(int level, void *lmmp)
  54 {
  55         struct lov_mds_md *lmm = lmmp;
  56         struct ost_id   oi;
  57
  58         lmm_oi_le_to_cpu(&oi, &lmm->lmm_oi);
  59         CDEBUG(level, "objid "DOSTID", magic 0x%08x, pattern %#x\n",
  60                POSTID(&oi), le32_to_cpu(lmm->lmm_magic),
  61                le32_to_cpu(lmm->lmm_pattern));
  62         CDEBUG(level, "stripe_size %u, stripe_count %u, layout_gen %u\n",
  63                le32_to_cpu(lmm->lmm_stripe_size),
  64                le16_to_cpu(lmm->lmm_stripe_count),
  65                le16_to_cpu(lmm->lmm_layout_gen));
  66 }
  67
  68 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
  69                                  int stripe_count)
  70 {
  71         int i;
  72
  73         if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
  74                 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
  75                        stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
  76                 return;
  77         }
  78
  79         for (i = 0; i < stripe_count; ++i, ++lod) {
  80                 struct ost_id   oi;
  81
  82                 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
  83                 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
  84                        le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
  85         }
  86 }
  87
  88 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
  89 {
  90         lov_dump_lmm_common(level, lmm);
  91         lov_dump_lmm_objects(level, lmm->lmm_objects,
  92                              le16_to_cpu(lmm->lmm_stripe_count));
  93 }
  94
  95 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
  96 {
  97         lov_dump_lmm_common(level, lmm);
  98         CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
  99         lov_dump_lmm_objects(level, lmm->lmm_objects,
 100                              le16_to_cpu(lmm->lmm_stripe_count));
 101 }
 102
 103 void lov_dump_lmm(int level, void *lmm)
 104 {
 105         int magic;
 106
 107         magic = le32_to_cpu(((struct lov_mds_md *)lmm)->lmm_magic);
 108         switch (magic) {
 109         case LOV_MAGIC_V1:
 110                 lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)lmm);
 111                 break;
 112         case LOV_MAGIC_V3:
 113                 lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)lmm);
 114                 break;
 115         default:
 116                 CDEBUG(level, "unrecognized lmm_magic %x, assuming %x\n",
 117                        magic, LOV_MAGIC_V1);
 118                 lov_dump_lmm_common(level, lmm);
 119                 break;
 120         }
 121 }
 122
 123 /**
 124  * Pack LOV striping metadata for disk storage format (in little
 125  * endian byte order).
 126  *
 127  * This follows the getxattr() conventions. If \a buf_size is zero
 128  * then return the size needed. If \a buf_size is too small then
 129  * return -ERANGE. Otherwise return the size of the result.
 130  */
 131 ssize_t lov_lsm_pack_v1v3(const struct lov_stripe_md *lsm, void *buf,
 132                           size_t buf_size)
 133 {
 134         struct lov_mds_md_v1 *lmmv1 = buf;
 135         struct lov_mds_md_v3 *lmmv3 = buf;
 136         struct lov_ost_data_v1 *lmm_objects;
 137         size_t lmm_size;
 138         unsigned int i;
 139         ENTRY;
 140
 141         lmm_size = lov_mds_md_size(lsm->lsm_entries[0]->lsme_stripe_count,
 142                                    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(
 155                                 lsm->lsm_entries[0]->lsme_stripe_size);
 156         lmmv1->lmm_stripe_count = cpu_to_le16(
 157                                 lsm->lsm_entries[0]->lsme_stripe_count);
 158         lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_entries[0]->lsme_pattern);
 159         lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
 160
 161         if (lsm->lsm_magic == LOV_MAGIC_V3) {
 162                 CLASSERT(sizeof(lsm->lsm_entries[0]->lsme_pool_name) ==
 163                          sizeof(lmmv3->lmm_pool_name));
 164                 strlcpy(lmmv3->lmm_pool_name,
 165                         lsm->lsm_entries[0]->lsme_pool_name,
 166                         sizeof(lmmv3->lmm_pool_name));
 167                 lmm_objects = lmmv3->lmm_objects;
 168         } else {
 169                 lmm_objects = lmmv1->lmm_objects;
 170         }
 171
 172         if (lsm->lsm_is_released)
 173                 RETURN(lmm_size);
 174
 175         for (i = 0; i < lsm->lsm_entries[0]->lsme_stripe_count; i++) {
 176                 struct lov_oinfo *loi = lsm->lsm_entries[0]->lsme_oinfo[i];
 177
 178                 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
 179                 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
 180                 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
 181         }
 182
 183         RETURN(lmm_size);
 184 }
 185
 186 ssize_t lov_lsm_pack(const struct lov_stripe_md *lsm, void *buf,
 187                      size_t buf_size)
 188 {
 189         struct lov_comp_md_v1 *lcmv1 = buf;
 190         struct lov_comp_md_entry_v1 *lcme;
 191         struct lov_ost_data_v1 *lmm_objects;
 192         size_t lmm_size;
 193         unsigned int entry;
 194         unsigned int offset;
 195         unsigned int size;
 196         unsigned int i;
 197         ENTRY;
 198
 199         if (lsm->lsm_magic == LOV_MAGIC_V1 || lsm->lsm_magic == LOV_MAGIC_V3)
 200                 return lov_lsm_pack_v1v3(lsm, buf, buf_size);
 201
 202         lmm_size = lov_comp_md_size(lsm);
 203         if (buf_size == 0)
 204                 RETURN(lmm_size);
 205
 206         if (buf_size < lmm_size)
 207                 RETURN(-ERANGE);
 208
 209         lcmv1->lcm_magic = cpu_to_le32(lsm->lsm_magic);
 210         lcmv1->lcm_size = cpu_to_le32(lmm_size);
 211         lcmv1->lcm_layout_gen = cpu_to_le32(lsm->lsm_layout_gen);
 212         lcmv1->lcm_entry_count = cpu_to_le16(lsm->lsm_entry_count);
 213
 214         offset = sizeof(*lcmv1) + sizeof(*lcme) * lsm->lsm_entry_count;
 215
 216         for (entry = 0; entry < lsm->lsm_entry_count; entry++) {
 217                 struct lov_stripe_md_entry *lsme;
 218                 struct lov_mds_md *lmm;
 219                 __u16 stripecnt;
 220
 221                 lsme = lsm->lsm_entries[entry];
 222                 lcme = &lcmv1->lcm_entries[entry];
 223
 224                 lcme->lcme_id = cpu_to_le32(lsme->lsme_id);
 225                 lcme->lcme_flags = cpu_to_le32(lsme->lsme_flags);
 226                 lcme->lcme_extent.e_start =
 227                         cpu_to_le64(lsme->lsme_extent.e_start);
 228                 lcme->lcme_extent.e_end =
 229                         cpu_to_le64(lsme->lsme_extent.e_end);
 230                 lcme->lcme_offset = cpu_to_le32(offset);
 231
 232                 lmm = (struct lov_mds_md *)((char *)lcmv1 + offset);
 233                 lmm->lmm_magic = cpu_to_le32(lsme->lsme_magic);
 234                 /* lmm->lmm_oi not set */
 235                 lmm->lmm_pattern = cpu_to_le32(lsme->lsme_pattern);
 236                 lmm->lmm_stripe_size = cpu_to_le32(lsme->lsme_stripe_size);
 237                 lmm->lmm_stripe_count = cpu_to_le16(lsme->lsme_stripe_count);
 238                 lmm->lmm_layout_gen = cpu_to_le16(lsme->lsme_layout_gen);
 239
 240                 if (lsme->lsme_magic == LOV_MAGIC_V3) {
 241                         struct lov_mds_md_v3 *lmmv3 =
 242                                                 (struct lov_mds_md_v3 *)lmm;
 243
 244                         strlcpy(lmmv3->lmm_pool_name, lsme->lsme_pool_name,
 245                                 sizeof(lmmv3->lmm_pool_name));
 246                         lmm_objects = lmmv3->lmm_objects;
 247                 } else {
 248                         lmm_objects =
 249                                 ((struct lov_mds_md_v1 *)lmm)->lmm_objects;
 250                 }
 251
 252                 if (lsme_inited(lsme) &&
 253                     !(lsme->lsme_pattern & LOV_PATTERN_F_RELEASED))
 254                         stripecnt = lsme->lsme_stripe_count;
 255                 else
 256                         stripecnt = 0;
 257
 258                 for (i = 0; i < stripecnt; i++) {
 259                         struct lov_oinfo *loi = lsme->lsme_oinfo[i];
 260
 261                         ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
 262                         lmm_objects[i].l_ost_gen =
 263                                         cpu_to_le32(loi->loi_ost_gen);
 264                         lmm_objects[i].l_ost_idx =
 265                                         cpu_to_le32(loi->loi_ost_idx);
 266                 }
 267
 268                 size = lov_mds_md_size(stripecnt, lsme->lsme_magic);
 269                 lcme->lcme_size = cpu_to_le32(size);
 270                 offset += size;
 271         } /* for each layout component */
 272
 273         RETURN(lmm_size);
 274 }
 275
 276 /* Find the max stripecount we should use */
 277 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
 278 {
 279         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
 280
 281         if (!stripe_count)
 282                 stripe_count = lov->desc.ld_default_stripe_count;
 283         if (stripe_count > lov->desc.ld_active_tgt_count)
 284                 stripe_count = lov->desc.ld_active_tgt_count;
 285         if (!stripe_count)
 286                 stripe_count = 1;
 287
 288         /* stripe count is based on whether ldiskfs can handle
 289          * larger EA sizes */
 290         if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
 291             lov->lov_ocd.ocd_max_easize)
 292                 max_stripes = lov_mds_md_max_stripe_count(
 293                         lov->lov_ocd.ocd_max_easize, magic);
 294
 295         if (stripe_count > max_stripes)
 296                 stripe_count = max_stripes;
 297
 298         return stripe_count;
 299 }
 300
 301 int lov_free_memmd(struct lov_stripe_md **lsmp)
 302 {
 303         struct lov_stripe_md *lsm = *lsmp;
 304         int refc;
 305
 306         *lsmp = NULL;
 307         refc = atomic_dec_return(&lsm->lsm_refc);
 308         LASSERT(refc >= 0);
 309         if (refc == 0)
 310                 lsm_free(lsm);
 311
 312         return refc;
 313 }
 314
 315 /* Unpack LOV object metadata from disk storage.  It is packed in LE byte
 316  * order and is opaque to the networking layer.
 317  */
 318 struct lov_stripe_md *lov_unpackmd(struct lov_obd *lov, void *buf,
 319                                    size_t buf_size)
 320 {
 321         const struct lsm_operations *op;
 322         struct lov_stripe_md *lsm;
 323         u32 magic;
 324         ENTRY;
 325
 326         if (buf_size < sizeof(magic))
 327                 RETURN(ERR_PTR(-EINVAL));
 328
 329         magic = le32_to_cpu(*(u32 *)buf);
 330         op = lsm_op_find(magic);
 331         if (op == NULL)
 332                 RETURN(ERR_PTR(-EINVAL));
 333
 334         lsm = op->lsm_unpackmd(lov, buf, buf_size);
 335
 336         RETURN(lsm);
 337 }
 338
 339 /* Retrieve object striping information.
 340  *
 341  * @lump is a pointer to an in-core struct with lmm_ost_count indicating
 342  * the maximum number of OST indices which will fit in the user buffer.
 343  * lmm_magic must be LOV_USER_MAGIC.
 344  *
 345  * If @size > 0, User specified limited buffer size, usually the buffer is from
 346  * ll_lov_setstripe(), and the buffer can only hold basic layout template info.
 347  */
 348 int lov_getstripe(const struct lu_env *env, struct lov_object *obj,
 349                   struct lov_stripe_md *lsm, struct lov_user_md __user *lump,
 350                   size_t size)
 351 {
 352         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
 353         struct lov_mds_md *lmmk, *lmm;
 354         struct lov_user_md_v1 lum;
 355         size_t  lmmk_size;
 356         ssize_t lmm_size, lum_size = 0;
 357         static bool printed;
 358         int     rc = 0;
 359         ENTRY;
 360
 361         if (lsm->lsm_magic != LOV_MAGIC_V1 && lsm->lsm_magic != LOV_MAGIC_V3 &&
 362             lsm->lsm_magic != LOV_MAGIC_COMP_V1) {
 363                 CERROR("bad LSM MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
 364                        lsm->lsm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
 365                 GOTO(out, rc = -EIO);
 366         }
 367
 368         if (!printed) {
 369                 LCONSOLE_WARN("%s: using old ioctl(LL_IOC_LOV_GETSTRIPE) on "
 370                               DFID", use llapi_layout_get_by_path()\n",
 371                               current->comm,
 372                               PFID(&obj->lo_cl.co_lu.lo_header->loh_fid));
 373                 printed = true;
 374         }
 375
 376         lmmk_size = lov_comp_md_size(lsm);
 377
 378         OBD_ALLOC_LARGE(lmmk, lmmk_size);
 379         if (lmmk == NULL)
 380                 GOTO(out, rc = -ENOMEM);
 381
 382         lmm_size = lov_lsm_pack(lsm, lmmk, lmmk_size);
 383         if (lmm_size < 0)
 384                 GOTO(out_free, rc = lmm_size);
 385
 386         if (cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) {
 387                 if (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1) ||
 388                     lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
 389                         lustre_swab_lov_mds_md(lmmk);
 390                         lustre_swab_lov_user_md_objects(
 391                                 (struct lov_user_ost_data *)lmmk->lmm_objects,
 392                                 lmmk->lmm_stripe_count);
 393                 } else if (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1)) {
 394                         lustre_swab_lov_comp_md_v1(
 395                                         (struct lov_comp_md_v1 *)lmmk);
 396                 }
 397         }
 398
 399         /* Legacy appication passes limited buffer, we need to figure out
 400          * the user buffer size by the passed in lmm_stripe_count. */
 401         if (copy_from_user(&lum, lump, sizeof(struct lov_user_md_v1)))
 402                 GOTO(out_free, rc = -EFAULT);
 403
 404         if (lum.lmm_magic == LOV_USER_MAGIC_V1 ||
 405             lum.lmm_magic == LOV_USER_MAGIC_V3)
 406                 lum_size = lov_user_md_size(lum.lmm_stripe_count,
 407                                             lum.lmm_magic);
 408
 409         if (lum_size != 0) {
 410                 struct lov_mds_md *comp_md = lmmk;
 411
 412                 /* Legacy app (ADIO for instance) treats the layout as V1/V3
 413                  * blindly, we'd return a reasonable V1/V3 for them. */
 414                 if (lmmk->lmm_magic == LOV_MAGIC_COMP_V1) {
 415                         struct lov_comp_md_v1 *comp_v1;
 416                         struct cl_object *cl_obj;
 417                         struct cl_attr attr;
 418                         int i;
 419
 420                         attr.cat_size = 0;
 421                         cl_obj = cl_object_top(&obj->lo_cl);
 422                         cl_object_attr_get(env, cl_obj, &attr);
 423
 424                         /* return the last instantiated component if file size
 425                          * is non-zero, otherwise, return the last component.*/
 426                         comp_v1 = (struct lov_comp_md_v1 *)lmmk;
 427                         i = attr.cat_size == 0 ? comp_v1->lcm_entry_count : 0;
 428                         for (; i < comp_v1->lcm_entry_count; i++) {
 429                                 if (!(comp_v1->lcm_entries[i].lcme_flags &
 430                                                 LCME_FL_INIT))
 431                                         break;
 432                         }
 433                         if (i > 0)
 434                                 i--;
 435                         comp_md = (struct lov_mds_md *)((char *)comp_v1 +
 436                                         comp_v1->lcm_entries[i].lcme_offset);
 437                 }
 438
 439                 lmm = comp_md;
 440                 lmm_size = lum_size;
 441         } else {
 442                 lmm = lmmk;
 443                 lmm_size = lmmk_size;
 444         }
 445         /**
 446          * User specified limited buffer size, usually the buffer is
 447          * from ll_lov_setstripe(), and the buffer can only hold basic
 448          * layout template info.
 449          */
 450         if (size == 0 || size > lmm_size)
 451                 size = lmm_size;
 452         if (copy_to_user(lump, lmm, size))
 453                 GOTO(out_free, rc = -EFAULT);
 454
 455 out_free:
 456         OBD_FREE_LARGE(lmmk, lmmk_size);
 457 out:
 458         RETURN(rc);
 459 }