fs/exofs/ore.c

   1 /*
   2  * Copyright (C) 2005, 2006
   3  * Avishay Traeger (avishay@gmail.com)
   4  * Copyright (C) 2008, 2009
   5  * Boaz Harrosh <bharrosh@panasas.com>
   6  *
   7  * This file is part of exofs.
   8  *
   9  * exofs is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation.  Since it is based on ext2, and the only
  12  * valid version of GPL for the Linux kernel is version 2, the only valid
  13  * version of GPL for exofs is version 2.
  14  *
  15  * exofs is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with exofs; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  23  */
  24
  25 #include <linux/slab.h>
  26 #include <asm/div64.h>
  27
  28 #include <scsi/osd_ore.h>
  29
  30 #define ORE_ERR(fmt, a...) printk(KERN_ERR "ore: " fmt, ##a)
  31
  32 #ifdef CONFIG_EXOFS_DEBUG
  33 #define ORE_DBGMSG(fmt, a...) \
  34         printk(KERN_NOTICE "ore @%s:%d: " fmt, __func__, __LINE__, ##a)
  35 #else
  36 #define ORE_DBGMSG(fmt, a...) \
  37         do { if (0) printk(fmt, ##a); } while (0)
  38 #endif
  39
  40 /* u64 has problems with printk this will cast it to unsigned long long */
  41 #define _LLU(x) (unsigned long long)(x)
  42
  43 #define ORE_DBGMSG2(M...) do {} while (0)
  44 /* #define ORE_DBGMSG2 ORE_DBGMSG */
  45
  46 MODULE_AUTHOR("Boaz Harrosh <bharrosh@panasas.com>");
  47 MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
  48 MODULE_LICENSE("GPL");
  49
  50 /* ore_verify_layout does a couple of things:
  51  * 1. Given a minimum number of needed parameters fixes up the rest of the
  52  *    members to be operatonals for the ore. The needed parameters are those
  53  *    that are defined by the pnfs-objects layout STD.
  54  * 2. Check to see if the current ore code actually supports these parameters
  55  *    for example stripe_unit must be a multple of the system PAGE_SIZE,
  56  *    and etc...
  57  * 3. Cache some havily used calculations that will be needed by users.
  58  */
  59
  60 enum { BIO_MAX_PAGES_KMALLOC =
  61                 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),};
  62
  63 int ore_verify_layout(unsigned total_comps, struct ore_layout *layout)
  64 {
  65         u64 stripe_length;
  66
  67 /* FIXME: Only raid0 is supported for now. */
  68         if (layout->raid_algorithm != PNFS_OSD_RAID_0) {
  69                 ORE_ERR("Only RAID_0 for now\n");
  70                 return -EINVAL;
  71         }
  72         if (0 != (layout->stripe_unit & ~PAGE_MASK)) {
  73                 ORE_ERR("Stripe Unit(0x%llx)"
  74                           " must be Multples of PAGE_SIZE(0x%lx)\n",
  75                           _LLU(layout->stripe_unit), PAGE_SIZE);
  76                 return -EINVAL;
  77         }
  78         if (layout->group_width) {
  79                 if (!layout->group_depth) {
  80                         ORE_ERR("group_depth == 0 && group_width != 0\n");
  81                         return -EINVAL;
  82                 }
  83                 if (total_comps < (layout->group_width * layout->mirrors_p1)) {
  84                         ORE_ERR("Data Map wrong, "
  85                                 "numdevs=%d < group_width=%d * mirrors=%d\n",
  86                                 total_comps, layout->group_width,
  87                                 layout->mirrors_p1);
  88                         return -EINVAL;
  89                 }
  90                 layout->group_count = total_comps / layout->mirrors_p1 /
  91                                                 layout->group_width;
  92         } else {
  93                 if (layout->group_depth) {
  94                         printk(KERN_NOTICE "Warning: group_depth ignored "
  95                                 "group_width == 0 && group_depth == %lld\n",
  96                                 _LLU(layout->group_depth));
  97                 }
  98                 layout->group_width = total_comps / layout->mirrors_p1;
  99                 layout->group_depth = -1;
 100                 layout->group_count = 1;
 101         }
 102
 103         stripe_length = (u64)layout->group_width * layout->stripe_unit;
 104         if (stripe_length >= (1ULL << 32)) {
 105                 ORE_ERR("Stripe_length(0x%llx) >= 32bit is not supported\n",
 106                         _LLU(stripe_length));
 107                 return -EINVAL;
 108         }
 109
 110         layout->max_io_length =
 111                 (BIO_MAX_PAGES_KMALLOC * PAGE_SIZE - layout->stripe_unit) *
 112                                                         layout->group_width;
 113         return 0;
 114 }
 115 EXPORT_SYMBOL(ore_verify_layout);
 116
 117 static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
 118 {
 119         return ios->oc->comps[index & ios->oc->single_comp].cred;
 120 }
 121
 122 static struct osd_obj_id *_ios_obj(struct ore_io_state *ios, unsigned index)
 123 {
 124         return &ios->oc->comps[index & ios->oc->single_comp].obj;
 125 }
 126
 127 static struct osd_dev *_ios_od(struct ore_io_state *ios, unsigned index)
 128 {
 129         ORE_DBGMSG2("oc->first_dev=%d oc->numdevs=%d i=%d oc->ods=%p\n",
 130                     ios->oc->first_dev, ios->oc->numdevs, index,
 131                     ios->oc->ods);
 132
 133         return ore_comp_dev(ios->oc, index);
 134 }
 135
 136 static int  _get_io_state(struct ore_layout *layout,
 137                           struct ore_components *oc, unsigned numdevs,
 138                           struct ore_io_state **pios)
 139 {
 140         struct ore_io_state *ios;
 141
 142         /*TODO: Maybe use kmem_cach per sbi of size
 143          * exofs_io_state_size(layout->s_numdevs)
 144          */
 145         ios = kzalloc(ore_io_state_size(numdevs), GFP_KERNEL);
 146         if (unlikely(!ios)) {
 147                 ORE_DBGMSG("Failed kzalloc bytes=%d\n",
 148                            ore_io_state_size(numdevs));
 149                 *pios = NULL;
 150                 return -ENOMEM;
 151         }
 152
 153         ios->layout = layout;
 154         ios->oc = oc;
 155         *pios = ios;
 156         return 0;
 157 }
 158
 159 /* Allocate an io_state for only a single group of devices
 160  *
 161  * If a user needs to call ore_read/write() this version must be used becase it
 162  * allocates extra stuff for striping and raid.
 163  * The ore might decide to only IO less then @length bytes do to alignmets
 164  * and constrains as follows:
 165  * - The IO cannot cross group boundary.
 166  * - In raid5/6 The end of the IO must align at end of a stripe eg.
 167  *   (@offset + @length) % strip_size == 0. Or the complete range is within a
 168  *   single stripe.
 169  * - Memory condition only permitted a shorter IO. (A user can use @length=~0
 170  *   And check the returned ios->length for max_io_size.)
 171  *
 172  * The caller must check returned ios->length (and/or ios->nr_pages) and
 173  * re-issue these pages that fall outside of ios->length
 174  */
 175 int  ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
 176                       bool is_reading, u64 offset, u64 length,
 177                       struct ore_io_state **pios)
 178 {
 179         struct ore_io_state *ios;
 180         unsigned numdevs = layout->group_width * layout->mirrors_p1;
 181         int ret;
 182
 183         ret = _get_io_state(layout, oc, numdevs, pios);
 184         if (unlikely(ret))
 185                 return ret;
 186
 187         ios = *pios;
 188         ios->reading = is_reading;
 189         ios->offset = offset;
 190
 191         if (length) {
 192                 ore_calc_stripe_info(layout, offset, &ios->si);
 193                 ios->length = (length <= ios->si.group_length) ? length :
 194                                                         ios->si.group_length;
 195                 ios->nr_pages = (ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
 196         }
 197
 198         return 0;
 199 }
 200 EXPORT_SYMBOL(ore_get_rw_state);
 201
 202 /* Allocate an io_state for all the devices in the comps array
 203  *
 204  * This version of io_state allocation is used mostly by create/remove
 205  * and trunc where we currently need all the devices. The only wastful
 206  * bit is the read/write_attributes with no IO. Those sites should
 207  * be converted to use ore_get_rw_state() with length=0
 208  */
 209 int  ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
 210                       struct ore_io_state **pios)
 211 {
 212         return _get_io_state(layout, oc, oc->numdevs, pios);
 213 }
 214 EXPORT_SYMBOL(ore_get_io_state);
 215
 216 void ore_put_io_state(struct ore_io_state *ios)
 217 {
 218         if (ios) {
 219                 unsigned i;
 220
 221                 for (i = 0; i < ios->numdevs; i++) {
 222                         struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 223
 224                         if (per_dev->or)
 225                                 osd_end_request(per_dev->or);
 226                         if (per_dev->bio)
 227                                 bio_put(per_dev->bio);
 228                 }
 229
 230                 kfree(ios);
 231         }
 232 }
 233 EXPORT_SYMBOL(ore_put_io_state);
 234
 235 static void _sync_done(struct ore_io_state *ios, void *p)
 236 {
 237         struct completion *waiting = p;
 238
 239         complete(waiting);
 240 }
 241
 242 static void _last_io(struct kref *kref)
 243 {
 244         struct ore_io_state *ios = container_of(
 245                                         kref, struct ore_io_state, kref);
 246
 247         ios->done(ios, ios->private);
 248 }
 249
 250 static void _done_io(struct osd_request *or, void *p)
 251 {
 252         struct ore_io_state *ios = p;
 253
 254         kref_put(&ios->kref, _last_io);
 255 }
 256
 257 static int ore_io_execute(struct ore_io_state *ios)
 258 {
 259         DECLARE_COMPLETION_ONSTACK(wait);
 260         bool sync = (ios->done == NULL);
 261         int i, ret;
 262
 263         if (sync) {
 264                 ios->done = _sync_done;
 265                 ios->private = &wait;
 266         }
 267
 268         for (i = 0; i < ios->numdevs; i++) {
 269                 struct osd_request *or = ios->per_dev[i].or;
 270                 if (unlikely(!or))
 271                         continue;
 272
 273                 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
 274                 if (unlikely(ret)) {
 275                         ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
 276                                      ret);
 277                         return ret;
 278                 }
 279         }
 280
 281         kref_init(&ios->kref);
 282
 283         for (i = 0; i < ios->numdevs; i++) {
 284                 struct osd_request *or = ios->per_dev[i].or;
 285                 if (unlikely(!or))
 286                         continue;
 287
 288                 kref_get(&ios->kref);
 289                 osd_execute_request_async(or, _done_io, ios);
 290         }
 291
 292         kref_put(&ios->kref, _last_io);
 293         ret = 0;
 294
 295         if (sync) {
 296                 wait_for_completion(&wait);
 297                 ret = ore_check_io(ios, NULL);
 298         }
 299         return ret;
 300 }
 301
 302 static void _clear_bio(struct bio *bio)
 303 {
 304         struct bio_vec *bv;
 305         unsigned i;
 306
 307         __bio_for_each_segment(bv, bio, i, 0) {
 308                 unsigned this_count = bv->bv_len;
 309
 310                 if (likely(PAGE_SIZE == this_count))
 311                         clear_highpage(bv->bv_page);
 312                 else
 313                         zero_user(bv->bv_page, bv->bv_offset, this_count);
 314         }
 315 }
 316
 317 int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
 318 {
 319         enum osd_err_priority acumulated_osd_err = 0;
 320         int acumulated_lin_err = 0;
 321         int i;
 322
 323         for (i = 0; i < ios->numdevs; i++) {
 324                 struct osd_sense_info osi;
 325                 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 326                 struct osd_request *or = per_dev->or;
 327                 int ret;
 328
 329                 if (unlikely(!or))
 330                         continue;
 331
 332                 ret = osd_req_decode_sense(or, &osi);
 333                 if (likely(!ret))
 334                         continue;
 335
 336                 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
 337                         /* start read offset passed endof file */
 338                         _clear_bio(per_dev->bio);
 339                         ORE_DBGMSG("start read offset passed end of file "
 340                                 "offset=0x%llx, length=0x%llx\n",
 341                                 _LLU(per_dev->offset),
 342                                 _LLU(per_dev->length));
 343
 344                         continue; /* we recovered */
 345                 }
 346
 347                 if (on_dev_error) {
 348                         u64 residual = ios->reading ?
 349                                         or->in.residual : or->out.residual;
 350                         u64 offset = (ios->offset + ios->length) - residual;
 351                         struct ore_dev *od = ios->oc->ods[
 352                                         per_dev->dev - ios->oc->first_dev];
 353
 354                         on_dev_error(ios, od, per_dev->dev, osi.osd_err_pri,
 355                                      offset, residual);
 356                 }
 357                 if (osi.osd_err_pri >= acumulated_osd_err) {
 358                         acumulated_osd_err = osi.osd_err_pri;
 359                         acumulated_lin_err = ret;
 360                 }
 361         }
 362
 363         return acumulated_lin_err;
 364 }
 365 EXPORT_SYMBOL(ore_check_io);
 366
 367 /*
 368  * L - logical offset into the file
 369  *
 370  * U - The number of bytes in a stripe within a group
 371  *
 372  *      U = stripe_unit * group_width
 373  *
 374  * T - The number of bytes striped within a group of component objects
 375  *     (before advancing to the next group)
 376  *
 377  *      T = stripe_unit * group_width * group_depth
 378  *
 379  * S - The number of bytes striped across all component objects
 380  *     before the pattern repeats
 381  *
 382  *      S = stripe_unit * group_width * group_depth * group_count
 383  *
 384  * M - The "major" (i.e., across all components) stripe number
 385  *
 386  *      M = L / S
 387  *
 388  * G - Counts the groups from the beginning of the major stripe
 389  *
 390  *      G = (L - (M * S)) / T   [or (L % S) / T]
 391  *
 392  * H - The byte offset within the group
 393  *
 394  *      H = (L - (M * S)) % T   [or (L % S) % T]
 395  *
 396  * N - The "minor" (i.e., across the group) stripe number
 397  *
 398  *      N = H / U
 399  *
 400  * C - The component index coresponding to L
 401  *
 402  *      C = (H - (N * U)) / stripe_unit + G * group_width
 403  *      [or (L % U) / stripe_unit + G * group_width]
 404  *
 405  * O - The component offset coresponding to L
 406  *
 407  *      O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
 408  */
 409 void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
 410                           struct ore_striping_info *si)
 411 {
 412         u32     stripe_unit = layout->stripe_unit;
 413         u32     group_width = layout->group_width;
 414         u64     group_depth = layout->group_depth;
 415
 416         u32     U = stripe_unit * group_width;
 417         u64     T = U * group_depth;
 418         u64     S = T * layout->group_count;
 419         u64     M = div64_u64(file_offset, S);
 420
 421         /*
 422         G = (L - (M * S)) / T
 423         H = (L - (M * S)) % T
 424         */
 425         u64     LmodS = file_offset - M * S;
 426         u32     G = div64_u64(LmodS, T);
 427         u64     H = LmodS - G * T;
 428
 429         u32     N = div_u64(H, U);
 430
 431         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
 432         si->dev = (u32)(H - (N * U)) / stripe_unit + G * group_width;
 433         si->dev *= layout->mirrors_p1;
 434
 435         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
 436
 437         si->obj_offset = si->unit_off + (N * stripe_unit) +
 438                                   (M * group_depth * stripe_unit);
 439
 440         si->group_length = T - H;
 441         si->M = M;
 442 }
 443 EXPORT_SYMBOL(ore_calc_stripe_info);
 444
 445 static int _add_stripe_unit(struct ore_io_state *ios,  unsigned *cur_pg,
 446                 unsigned pgbase, struct ore_per_dev_state *per_dev,
 447                 int cur_len)
 448 {
 449         unsigned pg = *cur_pg;
 450         struct request_queue *q =
 451                         osd_request_queue(_ios_od(ios, per_dev->dev));
 452         unsigned len = cur_len;
 453         int ret;
 454
 455         if (per_dev->bio == NULL) {
 456                 unsigned pages_in_stripe = ios->layout->group_width *
 457                                         (ios->layout->stripe_unit / PAGE_SIZE);
 458                 unsigned bio_size = (ios->nr_pages + pages_in_stripe) /
 459                                                 ios->layout->group_width;
 460
 461                 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
 462                 if (unlikely(!per_dev->bio)) {
 463                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
 464                                      bio_size);
 465                         ret = -ENOMEM;
 466                         goto out;
 467                 }
 468         }
 469
 470         while (cur_len > 0) {
 471                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
 472                 unsigned added_len;
 473
 474                 BUG_ON(ios->nr_pages <= pg);
 475                 cur_len -= pglen;
 476
 477                 added_len = bio_add_pc_page(q, per_dev->bio, ios->pages[pg],
 478                                             pglen, pgbase);
 479                 if (unlikely(pglen != added_len)) {
 480                         ret = -ENOMEM;
 481                         goto out;
 482                 }
 483                 pgbase = 0;
 484                 ++pg;
 485         }
 486         BUG_ON(cur_len);
 487
 488         per_dev->length += len;
 489         *cur_pg = pg;
 490         ret = 0;
 491 out:    /* we fail the complete unit on an error eg don't advance
 492          * per_dev->length and cur_pg. This means that we might have a bigger
 493          * bio than the CDB requested length (per_dev->length). That's fine
 494          * only the oposite is fatal.
 495          */
 496         return ret;
 497 }
 498
 499 static int _prepare_for_striping(struct ore_io_state *ios)
 500 {
 501         struct ore_striping_info *si = &ios->si;
 502         unsigned stripe_unit = ios->layout->stripe_unit;
 503         unsigned mirrors_p1 = ios->layout->mirrors_p1;
 504         unsigned devs_in_group = ios->layout->group_width * mirrors_p1;
 505         unsigned dev = si->dev;
 506         unsigned first_dev = dev - (dev % devs_in_group);
 507         unsigned cur_pg = ios->pages_consumed;
 508         u64 length = ios->length;
 509         int ret = 0;
 510
 511         if (!ios->pages) {
 512                 ios->numdevs = ios->layout->mirrors_p1;
 513                 return 0;
 514         }
 515
 516         BUG_ON(length > si->group_length);
 517
 518         while (length) {
 519                 unsigned comp = dev - first_dev;
 520                 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
 521                 unsigned cur_len, page_off = 0;
 522
 523                 if (!per_dev->length) {
 524                         per_dev->dev = dev;
 525                         if (dev < si->dev) {
 526                                 per_dev->offset = si->obj_offset + stripe_unit -
 527                                                                    si->unit_off;
 528                                 cur_len = stripe_unit;
 529                         } else if (dev == si->dev) {
 530                                 per_dev->offset = si->obj_offset;
 531                                 cur_len = stripe_unit - si->unit_off;
 532                                 page_off = si->unit_off & ~PAGE_MASK;
 533                                 BUG_ON(page_off && (page_off != ios->pgbase));
 534                         } else { /* dev > si->dev */
 535                                 per_dev->offset = si->obj_offset - si->unit_off;
 536                                 cur_len = stripe_unit;
 537                         }
 538                 } else {
 539                         cur_len = stripe_unit;
 540                 }
 541                 if (cur_len >= length)
 542                         cur_len = length;
 543
 544                 ret = _add_stripe_unit(ios, &cur_pg, page_off , per_dev,
 545                                        cur_len);
 546                 if (unlikely(ret))
 547                         goto out;
 548
 549                 dev += mirrors_p1;
 550                 dev = (dev % devs_in_group) + first_dev;
 551
 552                 length -= cur_len;
 553         }
 554 out:
 555         ios->numdevs = devs_in_group;
 556         ios->pages_consumed = cur_pg;
 557         if (unlikely(ret)) {
 558                 if (length == ios->length)
 559                         return ret;
 560                 else
 561                         ios->length -= length;
 562         }
 563         return 0;
 564 }
 565
 566 int ore_create(struct ore_io_state *ios)
 567 {
 568         int i, ret;
 569
 570         for (i = 0; i < ios->oc->numdevs; i++) {
 571                 struct osd_request *or;
 572
 573                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 574                 if (unlikely(!or)) {
 575                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 576                         ret = -ENOMEM;
 577                         goto out;
 578                 }
 579                 ios->per_dev[i].or = or;
 580                 ios->numdevs++;
 581
 582                 osd_req_create_object(or, _ios_obj(ios, i));
 583         }
 584         ret = ore_io_execute(ios);
 585
 586 out:
 587         return ret;
 588 }
 589 EXPORT_SYMBOL(ore_create);
 590
 591 int ore_remove(struct ore_io_state *ios)
 592 {
 593         int i, ret;
 594
 595         for (i = 0; i < ios->oc->numdevs; i++) {
 596                 struct osd_request *or;
 597
 598                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 599                 if (unlikely(!or)) {
 600                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 601                         ret = -ENOMEM;
 602                         goto out;
 603                 }
 604                 ios->per_dev[i].or = or;
 605                 ios->numdevs++;
 606
 607                 osd_req_remove_object(or, _ios_obj(ios, i));
 608         }
 609         ret = ore_io_execute(ios);
 610
 611 out:
 612         return ret;
 613 }
 614 EXPORT_SYMBOL(ore_remove);
 615
 616 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
 617 {
 618         struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
 619         unsigned dev = ios->per_dev[cur_comp].dev;
 620         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
 621         int ret = 0;
 622
 623         if (ios->pages && !master_dev->length)
 624                 return 0; /* Just an empty slot */
 625
 626         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
 627                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 628                 struct osd_request *or;
 629
 630                 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
 631                 if (unlikely(!or)) {
 632                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 633                         ret = -ENOMEM;
 634                         goto out;
 635                 }
 636                 per_dev->or = or;
 637
 638                 if (ios->pages) {
 639                         struct bio *bio;
 640
 641                         if (per_dev != master_dev) {
 642                                 bio = bio_kmalloc(GFP_KERNEL,
 643                                                   master_dev->bio->bi_max_vecs);
 644                                 if (unlikely(!bio)) {
 645                                         ORE_DBGMSG(
 646                                               "Failed to allocate BIO size=%u\n",
 647                                               master_dev->bio->bi_max_vecs);
 648                                         ret = -ENOMEM;
 649                                         goto out;
 650                                 }
 651
 652                                 __bio_clone(bio, master_dev->bio);
 653                                 bio->bi_bdev = NULL;
 654                                 bio->bi_next = NULL;
 655                                 per_dev->offset = master_dev->offset;
 656                                 per_dev->length = master_dev->length;
 657                                 per_dev->bio =  bio;
 658                                 per_dev->dev = dev;
 659                         } else {
 660                                 bio = master_dev->bio;
 661                                 /* FIXME: bio_set_dir() */
 662                                 bio->bi_rw |= REQ_WRITE;
 663                         }
 664
 665                         osd_req_write(or, _ios_obj(ios, dev), per_dev->offset,
 666                                       bio, per_dev->length);
 667                         ORE_DBGMSG("write(0x%llx) offset=0x%llx "
 668                                       "length=0x%llx dev=%d\n",
 669                                      _LLU(_ios_obj(ios, dev)->id),
 670                                      _LLU(per_dev->offset),
 671                                      _LLU(per_dev->length), dev);
 672                 } else if (ios->kern_buff) {
 673                         per_dev->offset = ios->si.obj_offset;
 674                         per_dev->dev = ios->si.dev + dev;
 675
 676                         /* no cross device without page array */
 677                         BUG_ON((ios->layout->group_width > 1) &&
 678                                (ios->si.unit_off + ios->length >
 679                                 ios->layout->stripe_unit));
 680
 681                         ret = osd_req_write_kern(or, _ios_obj(ios, per_dev->dev),
 682                                                  per_dev->offset,
 683                                                  ios->kern_buff, ios->length);
 684                         if (unlikely(ret))
 685                                 goto out;
 686                         ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
 687                                       "length=0x%llx dev=%d\n",
 688                                      _LLU(_ios_obj(ios, dev)->id),
 689                                      _LLU(per_dev->offset),
 690                                      _LLU(ios->length), per_dev->dev);
 691                 } else {
 692                         osd_req_set_attributes(or, _ios_obj(ios, dev));
 693                         ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
 694                                      _LLU(_ios_obj(ios, dev)->id),
 695                                      ios->out_attr_len, dev);
 696                 }
 697
 698                 if (ios->out_attr)
 699                         osd_req_add_set_attr_list(or, ios->out_attr,
 700                                                   ios->out_attr_len);
 701
 702                 if (ios->in_attr)
 703                         osd_req_add_get_attr_list(or, ios->in_attr,
 704                                                   ios->in_attr_len);
 705         }
 706
 707 out:
 708         return ret;
 709 }
 710
 711 int ore_write(struct ore_io_state *ios)
 712 {
 713         int i;
 714         int ret;
 715
 716         ret = _prepare_for_striping(ios);
 717         if (unlikely(ret))
 718                 return ret;
 719
 720         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 721                 ret = _write_mirror(ios, i);
 722                 if (unlikely(ret))
 723                         return ret;
 724         }
 725
 726         ret = ore_io_execute(ios);
 727         return ret;
 728 }
 729 EXPORT_SYMBOL(ore_write);
 730
 731 static int _read_mirror(struct ore_io_state *ios, unsigned cur_comp)
 732 {
 733         struct osd_request *or;
 734         struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 735         struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
 736         unsigned first_dev = (unsigned)obj->id;
 737
 738         if (ios->pages && !per_dev->length)
 739                 return 0; /* Just an empty slot */
 740
 741         first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
 742         or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
 743         if (unlikely(!or)) {
 744                 ORE_ERR("%s: osd_start_request failed\n", __func__);
 745                 return -ENOMEM;
 746         }
 747         per_dev->or = or;
 748
 749         if (ios->pages) {
 750                 osd_req_read(or, obj, per_dev->offset,
 751                                 per_dev->bio, per_dev->length);
 752                 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
 753                              " dev=%d\n", _LLU(obj->id),
 754                              _LLU(per_dev->offset), _LLU(per_dev->length),
 755                              first_dev);
 756         } else {
 757                 BUG_ON(ios->kern_buff);
 758
 759                 osd_req_get_attributes(or, obj);
 760                 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
 761                               _LLU(obj->id),
 762                               ios->in_attr_len, first_dev);
 763         }
 764         if (ios->out_attr)
 765                 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
 766
 767         if (ios->in_attr)
 768                 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
 769
 770         return 0;
 771 }
 772
 773 int ore_read(struct ore_io_state *ios)
 774 {
 775         int i;
 776         int ret;
 777
 778         ret = _prepare_for_striping(ios);
 779         if (unlikely(ret))
 780                 return ret;
 781
 782         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 783                 ret = _read_mirror(ios, i);
 784                 if (unlikely(ret))
 785                         return ret;
 786         }
 787
 788         ret = ore_io_execute(ios);
 789         return ret;
 790 }
 791 EXPORT_SYMBOL(ore_read);
 792
 793 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
 794 {
 795         struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
 796         void *iter = NULL;
 797         int nelem;
 798
 799         do {
 800                 nelem = 1;
 801                 osd_req_decode_get_attr_list(ios->per_dev[0].or,
 802                                              &cur_attr, &nelem, &iter);
 803                 if ((cur_attr.attr_page == attr->attr_page) &&
 804                     (cur_attr.attr_id == attr->attr_id)) {
 805                         attr->len = cur_attr.len;
 806                         attr->val_ptr = cur_attr.val_ptr;
 807                         return 0;
 808                 }
 809         } while (iter);
 810
 811         return -EIO;
 812 }
 813 EXPORT_SYMBOL(extract_attr_from_ios);
 814
 815 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
 816                              struct osd_attr *attr)
 817 {
 818         int last_comp = cur_comp + ios->layout->mirrors_p1;
 819
 820         for (; cur_comp < last_comp; ++cur_comp) {
 821                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 822                 struct osd_request *or;
 823
 824                 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
 825                 if (unlikely(!or)) {
 826                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 827                         return -ENOMEM;
 828                 }
 829                 per_dev->or = or;
 830
 831                 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
 832                 osd_req_add_set_attr_list(or, attr, 1);
 833         }
 834
 835         return 0;
 836 }
 837
 838 struct _trunc_info {
 839         struct ore_striping_info si;
 840         u64 prev_group_obj_off;
 841         u64 next_group_obj_off;
 842
 843         unsigned first_group_dev;
 844         unsigned nex_group_dev;
 845 };
 846
 847 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
 848                              struct _trunc_info *ti)
 849 {
 850         unsigned stripe_unit = layout->stripe_unit;
 851
 852         ore_calc_stripe_info(layout, file_offset, &ti->si);
 853
 854         ti->prev_group_obj_off = ti->si.M * stripe_unit;
 855         ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
 856
 857         ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
 858         ti->nex_group_dev = ti->first_group_dev + layout->group_width;
 859 }
 860
 861 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
 862                    u64 size)
 863 {
 864         struct ore_io_state *ios;
 865         struct exofs_trunc_attr {
 866                 struct osd_attr attr;
 867                 __be64 newsize;
 868         } *size_attrs;
 869         struct _trunc_info ti;
 870         int i, ret;
 871
 872         ret = ore_get_io_state(layout, oc, &ios);
 873         if (unlikely(ret))
 874                 return ret;
 875
 876         _calc_trunk_info(ios->layout, size, &ti);
 877
 878         size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
 879                              GFP_KERNEL);
 880         if (unlikely(!size_attrs)) {
 881                 ret = -ENOMEM;
 882                 goto out;
 883         }
 884
 885         ios->numdevs = ios->oc->numdevs;
 886
 887         for (i = 0; i < ios->numdevs; ++i) {
 888                 struct exofs_trunc_attr *size_attr = &size_attrs[i];
 889                 u64 obj_size;
 890
 891                 if (i < ti.first_group_dev)
 892                         obj_size = ti.prev_group_obj_off;
 893                 else if (i >= ti.nex_group_dev)
 894                         obj_size = ti.next_group_obj_off;
 895                 else if (i < ti.si.dev) /* dev within this group */
 896                         obj_size = ti.si.obj_offset +
 897                                       ios->layout->stripe_unit - ti.si.unit_off;
 898                 else if (i == ti.si.dev)
 899                         obj_size = ti.si.obj_offset;
 900                 else /* i > ti.dev */
 901                         obj_size = ti.si.obj_offset - ti.si.unit_off;
 902
 903                 size_attr->newsize = cpu_to_be64(obj_size);
 904                 size_attr->attr = g_attr_logical_length;
 905                 size_attr->attr.val_ptr = &size_attr->newsize;
 906
 907                 ORE_DBGMSG("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
 908                              _LLU(oc->comps->obj.id), _LLU(obj_size), i);
 909                 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
 910                                         &size_attr->attr);
 911                 if (unlikely(ret))
 912                         goto out;
 913         }
 914         ret = ore_io_execute(ios);
 915
 916 out:
 917         kfree(size_attrs);
 918         ore_put_io_state(ios);
 919         return ret;
 920 }
 921 EXPORT_SYMBOL(ore_truncate);
 922
 923 const struct osd_attr g_attr_logical_length = ATTR_DEF(
 924         OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
 925 EXPORT_SYMBOL(g_attr_logical_length);