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