fs/exofs/ore.c

   1 /*
   2  * Copyright (C) 2005, 2006
   3  * Avishay Traeger (avishay@gmail.com)
   4  * Copyright (C) 2008, 2009
   5  * Boaz Harrosh <ooo@electrozaur.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 <linux/module.h>
  27 #include <asm/div64.h>
  28 #include <linux/lcm.h>
  29
  30 #include "ore_raid.h"
  31
  32 MODULE_AUTHOR("Boaz Harrosh <ooo@electrozaur.com>");
  33 MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
  34 MODULE_LICENSE("GPL");
  35
  36 /* ore_verify_layout does a couple of things:
  37  * 1. Given a minimum number of needed parameters fixes up the rest of the
  38  *    members to be operatonals for the ore. The needed parameters are those
  39  *    that are defined by the pnfs-objects layout STD.
  40  * 2. Check to see if the current ore code actually supports these parameters
  41  *    for example stripe_unit must be a multple of the system PAGE_SIZE,
  42  *    and etc...
  43  * 3. Cache some havily used calculations that will be needed by users.
  44  */
  45
  46 enum { BIO_MAX_PAGES_KMALLOC =
  47                 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),};
  48
  49 int ore_verify_layout(unsigned total_comps, struct ore_layout *layout)
  50 {
  51         u64 stripe_length;
  52
  53         switch (layout->raid_algorithm) {
  54         case PNFS_OSD_RAID_0:
  55                 layout->parity = 0;
  56                 break;
  57         case PNFS_OSD_RAID_5:
  58                 layout->parity = 1;
  59                 break;
  60         case PNFS_OSD_RAID_PQ:
  61                 layout->parity = 2;
  62                 break;
  63         case PNFS_OSD_RAID_4:
  64         default:
  65                 ORE_ERR("Only RAID_0/5/6 for now received-enum=%d\n",
  66                         layout->raid_algorithm);
  67                 return -EINVAL;
  68         }
  69         if (0 != (layout->stripe_unit & ~PAGE_MASK)) {
  70                 ORE_ERR("Stripe Unit(0x%llx)"
  71                           " must be Multples of PAGE_SIZE(0x%lx)\n",
  72                           _LLU(layout->stripe_unit), PAGE_SIZE);
  73                 return -EINVAL;
  74         }
  75         if (layout->group_width) {
  76                 if (!layout->group_depth) {
  77                         ORE_ERR("group_depth == 0 && group_width != 0\n");
  78                         return -EINVAL;
  79                 }
  80                 if (total_comps < (layout->group_width * layout->mirrors_p1)) {
  81                         ORE_ERR("Data Map wrong, "
  82                                 "numdevs=%d < group_width=%d * mirrors=%d\n",
  83                                 total_comps, layout->group_width,
  84                                 layout->mirrors_p1);
  85                         return -EINVAL;
  86                 }
  87                 layout->group_count = total_comps / layout->mirrors_p1 /
  88                                                 layout->group_width;
  89         } else {
  90                 if (layout->group_depth) {
  91                         printk(KERN_NOTICE "Warning: group_depth ignored "
  92                                 "group_width == 0 && group_depth == %lld\n",
  93                                 _LLU(layout->group_depth));
  94                 }
  95                 layout->group_width = total_comps / layout->mirrors_p1;
  96                 layout->group_depth = -1;
  97                 layout->group_count = 1;
  98         }
  99
 100         stripe_length = (u64)layout->group_width * layout->stripe_unit;
 101         if (stripe_length >= (1ULL << 32)) {
 102                 ORE_ERR("Stripe_length(0x%llx) >= 32bit is not supported\n",
 103                         _LLU(stripe_length));
 104                 return -EINVAL;
 105         }
 106
 107         layout->max_io_length =
 108                 (BIO_MAX_PAGES_KMALLOC * PAGE_SIZE - layout->stripe_unit) *
 109                                         (layout->group_width - layout->parity);
 110         if (layout->parity) {
 111                 unsigned stripe_length =
 112                                 (layout->group_width - layout->parity) *
 113                                 layout->stripe_unit;
 114
 115                 layout->max_io_length /= stripe_length;
 116                 layout->max_io_length *= stripe_length;
 117         }
 118         ORE_DBGMSG("max_io_length=0x%lx\n", layout->max_io_length);
 119
 120         return 0;
 121 }
 122 EXPORT_SYMBOL(ore_verify_layout);
 123
 124 static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
 125 {
 126         return ios->oc->comps[index & ios->oc->single_comp].cred;
 127 }
 128
 129 static struct osd_obj_id *_ios_obj(struct ore_io_state *ios, unsigned index)
 130 {
 131         return &ios->oc->comps[index & ios->oc->single_comp].obj;
 132 }
 133
 134 static struct osd_dev *_ios_od(struct ore_io_state *ios, unsigned index)
 135 {
 136         ORE_DBGMSG2("oc->first_dev=%d oc->numdevs=%d i=%d oc->ods=%p\n",
 137                     ios->oc->first_dev, ios->oc->numdevs, index,
 138                     ios->oc->ods);
 139
 140         return ore_comp_dev(ios->oc, index);
 141 }
 142
 143 int  _ore_get_io_state(struct ore_layout *layout,
 144                         struct ore_components *oc, unsigned numdevs,
 145                         unsigned sgs_per_dev, unsigned num_par_pages,
 146                         struct ore_io_state **pios)
 147 {
 148         struct ore_io_state *ios;
 149         struct page **pages;
 150         struct osd_sg_entry *sgilist;
 151         struct __alloc_all_io_state {
 152                 struct ore_io_state ios;
 153                 struct ore_per_dev_state per_dev[numdevs];
 154                 union {
 155                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
 156                         struct page *pages[num_par_pages];
 157                 };
 158         } *_aios;
 159
 160         if (likely(sizeof(*_aios) <= PAGE_SIZE)) {
 161                 _aios = kzalloc(sizeof(*_aios), GFP_KERNEL);
 162                 if (unlikely(!_aios)) {
 163                         ORE_DBGMSG("Failed kzalloc bytes=%zd\n",
 164                                    sizeof(*_aios));
 165                         *pios = NULL;
 166                         return -ENOMEM;
 167                 }
 168                 pages = num_par_pages ? _aios->pages : NULL;
 169                 sgilist = sgs_per_dev ? _aios->sglist : NULL;
 170                 ios = &_aios->ios;
 171         } else {
 172                 struct __alloc_small_io_state {
 173                         struct ore_io_state ios;
 174                         struct ore_per_dev_state per_dev[numdevs];
 175                 } *_aio_small;
 176                 union __extra_part {
 177                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
 178                         struct page *pages[num_par_pages];
 179                 } *extra_part;
 180
 181                 _aio_small = kzalloc(sizeof(*_aio_small), GFP_KERNEL);
 182                 if (unlikely(!_aio_small)) {
 183                         ORE_DBGMSG("Failed alloc first part bytes=%zd\n",
 184                                    sizeof(*_aio_small));
 185                         *pios = NULL;
 186                         return -ENOMEM;
 187                 }
 188                 extra_part = kzalloc(sizeof(*extra_part), GFP_KERNEL);
 189                 if (unlikely(!extra_part)) {
 190                         ORE_DBGMSG("Failed alloc second part bytes=%zd\n",
 191                                    sizeof(*extra_part));
 192                         kfree(_aio_small);
 193                         *pios = NULL;
 194                         return -ENOMEM;
 195                 }
 196
 197                 pages = num_par_pages ? extra_part->pages : NULL;
 198                 sgilist = sgs_per_dev ? extra_part->sglist : NULL;
 199                 /* In this case the per_dev[0].sgilist holds the pointer to
 200                  * be freed
 201                  */
 202                 ios = &_aio_small->ios;
 203                 ios->extra_part_alloc = true;
 204         }
 205
 206         if (pages) {
 207                 ios->parity_pages = pages;
 208                 ios->max_par_pages = num_par_pages;
 209         }
 210         if (sgilist) {
 211                 unsigned d;
 212
 213                 for (d = 0; d < numdevs; ++d) {
 214                         ios->per_dev[d].sglist = sgilist;
 215                         sgilist += sgs_per_dev;
 216                 }
 217                 ios->sgs_per_dev = sgs_per_dev;
 218         }
 219
 220         ios->layout = layout;
 221         ios->oc = oc;
 222         *pios = ios;
 223         return 0;
 224 }
 225
 226 /* Allocate an io_state for only a single group of devices
 227  *
 228  * If a user needs to call ore_read/write() this version must be used becase it
 229  * allocates extra stuff for striping and raid.
 230  * The ore might decide to only IO less then @length bytes do to alignmets
 231  * and constrains as follows:
 232  * - The IO cannot cross group boundary.
 233  * - In raid5/6 The end of the IO must align at end of a stripe eg.
 234  *   (@offset + @length) % strip_size == 0. Or the complete range is within a
 235  *   single stripe.
 236  * - Memory condition only permitted a shorter IO. (A user can use @length=~0
 237  *   And check the returned ios->length for max_io_size.)
 238  *
 239  * The caller must check returned ios->length (and/or ios->nr_pages) and
 240  * re-issue these pages that fall outside of ios->length
 241  */
 242 int  ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
 243                       bool is_reading, u64 offset, u64 length,
 244                       struct ore_io_state **pios)
 245 {
 246         struct ore_io_state *ios;
 247         unsigned numdevs = layout->group_width * layout->mirrors_p1;
 248         unsigned sgs_per_dev = 0, max_par_pages = 0;
 249         int ret;
 250
 251         if (layout->parity && length) {
 252                 unsigned data_devs = layout->group_width - layout->parity;
 253                 unsigned stripe_size = layout->stripe_unit * data_devs;
 254                 unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
 255                 u32 remainder;
 256                 u64 num_stripes;
 257                 u64 num_raid_units;
 258
 259                 num_stripes = div_u64_rem(length, stripe_size, &remainder);
 260                 if (remainder)
 261                         ++num_stripes;
 262
 263                 num_raid_units =  num_stripes * layout->parity;
 264
 265                 if (is_reading) {
 266                         /* For reads add per_dev sglist array */
 267                         /* TODO: Raid 6 we need twice more. Actually:
 268                         *         num_stripes / LCMdP(W,P);
 269                         *         if (W%P != 0) num_stripes *= parity;
 270                         */
 271
 272                         /* first/last seg is split */
 273                         num_raid_units += layout->group_width;
 274                         sgs_per_dev = div_u64(num_raid_units, data_devs) + 2;
 275                 } else {
 276                         /* For Writes add parity pages array. */
 277                         max_par_pages = num_raid_units * pages_in_unit *
 278                                                 sizeof(struct page *);
 279                 }
 280         }
 281
 282         ret = _ore_get_io_state(layout, oc, numdevs, sgs_per_dev, max_par_pages,
 283                                 pios);
 284         if (unlikely(ret))
 285                 return ret;
 286
 287         ios = *pios;
 288         ios->reading = is_reading;
 289         ios->offset = offset;
 290
 291         if (length) {
 292                 ore_calc_stripe_info(layout, offset, length, &ios->si);
 293                 ios->length = ios->si.length;
 294                 ios->nr_pages = ((ios->offset & (PAGE_SIZE - 1)) +
 295                                  ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
 296                 if (layout->parity)
 297                         _ore_post_alloc_raid_stuff(ios);
 298         }
 299
 300         return 0;
 301 }
 302 EXPORT_SYMBOL(ore_get_rw_state);
 303
 304 /* Allocate an io_state for all the devices in the comps array
 305  *
 306  * This version of io_state allocation is used mostly by create/remove
 307  * and trunc where we currently need all the devices. The only wastful
 308  * bit is the read/write_attributes with no IO. Those sites should
 309  * be converted to use ore_get_rw_state() with length=0
 310  */
 311 int  ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
 312                       struct ore_io_state **pios)
 313 {
 314         return _ore_get_io_state(layout, oc, oc->numdevs, 0, 0, pios);
 315 }
 316 EXPORT_SYMBOL(ore_get_io_state);
 317
 318 void ore_put_io_state(struct ore_io_state *ios)
 319 {
 320         if (ios) {
 321                 unsigned i;
 322
 323                 for (i = 0; i < ios->numdevs; i++) {
 324                         struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 325
 326                         if (per_dev->or)
 327                                 osd_end_request(per_dev->or);
 328                         if (per_dev->bio)
 329                                 bio_put(per_dev->bio);
 330                 }
 331
 332                 _ore_free_raid_stuff(ios);
 333                 kfree(ios);
 334         }
 335 }
 336 EXPORT_SYMBOL(ore_put_io_state);
 337
 338 static void _sync_done(struct ore_io_state *ios, void *p)
 339 {
 340         struct completion *waiting = p;
 341
 342         complete(waiting);
 343 }
 344
 345 static void _last_io(struct kref *kref)
 346 {
 347         struct ore_io_state *ios = container_of(
 348                                         kref, struct ore_io_state, kref);
 349
 350         ios->done(ios, ios->private);
 351 }
 352
 353 static void _done_io(struct osd_request *or, void *p)
 354 {
 355         struct ore_io_state *ios = p;
 356
 357         kref_put(&ios->kref, _last_io);
 358 }
 359
 360 int ore_io_execute(struct ore_io_state *ios)
 361 {
 362         DECLARE_COMPLETION_ONSTACK(wait);
 363         bool sync = (ios->done == NULL);
 364         int i, ret;
 365
 366         if (sync) {
 367                 ios->done = _sync_done;
 368                 ios->private = &wait;
 369         }
 370
 371         for (i = 0; i < ios->numdevs; i++) {
 372                 struct osd_request *or = ios->per_dev[i].or;
 373                 if (unlikely(!or))
 374                         continue;
 375
 376                 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
 377                 if (unlikely(ret)) {
 378                         ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
 379                                      ret);
 380                         return ret;
 381                 }
 382         }
 383
 384         kref_init(&ios->kref);
 385
 386         for (i = 0; i < ios->numdevs; i++) {
 387                 struct osd_request *or = ios->per_dev[i].or;
 388                 if (unlikely(!or))
 389                         continue;
 390
 391                 kref_get(&ios->kref);
 392                 osd_execute_request_async(or, _done_io, ios);
 393         }
 394
 395         kref_put(&ios->kref, _last_io);
 396         ret = 0;
 397
 398         if (sync) {
 399                 wait_for_completion(&wait);
 400                 ret = ore_check_io(ios, NULL);
 401         }
 402         return ret;
 403 }
 404
 405 static void _clear_bio(struct bio *bio)
 406 {
 407         struct bio_vec *bv;
 408         unsigned i;
 409
 410         bio_for_each_segment_all(bv, bio, i) {
 411                 unsigned this_count = bv->bv_len;
 412
 413                 if (likely(PAGE_SIZE == this_count))
 414                         clear_highpage(bv->bv_page);
 415                 else
 416                         zero_user(bv->bv_page, bv->bv_offset, this_count);
 417         }
 418 }
 419
 420 int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
 421 {
 422         enum osd_err_priority acumulated_osd_err = 0;
 423         int acumulated_lin_err = 0;
 424         int i;
 425
 426         for (i = 0; i < ios->numdevs; i++) {
 427                 struct osd_sense_info osi;
 428                 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 429                 struct osd_request *or = per_dev->or;
 430                 int ret;
 431
 432                 if (unlikely(!or))
 433                         continue;
 434
 435                 ret = osd_req_decode_sense(or, &osi);
 436                 if (likely(!ret))
 437                         continue;
 438
 439                 if ((OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) &&
 440                     per_dev->bio) {
 441                         /* start read offset passed endof file.
 442                          * Note: if we do not have bio it means read-attributes
 443                          * In this case we should return error to caller.
 444                          */
 445                         _clear_bio(per_dev->bio);
 446                         ORE_DBGMSG("start read offset passed end of file "
 447                                 "offset=0x%llx, length=0x%llx\n",
 448                                 _LLU(per_dev->offset),
 449                                 _LLU(per_dev->length));
 450
 451                         continue; /* we recovered */
 452                 }
 453
 454                 if (on_dev_error) {
 455                         u64 residual = ios->reading ?
 456                                         or->in.residual : or->out.residual;
 457                         u64 offset = (ios->offset + ios->length) - residual;
 458                         unsigned dev = per_dev->dev - ios->oc->first_dev;
 459                         struct ore_dev *od = ios->oc->ods[dev];
 460
 461                         on_dev_error(ios, od, dev, osi.osd_err_pri,
 462                                      offset, residual);
 463                 }
 464                 if (osi.osd_err_pri >= acumulated_osd_err) {
 465                         acumulated_osd_err = osi.osd_err_pri;
 466                         acumulated_lin_err = ret;
 467                 }
 468         }
 469
 470         return acumulated_lin_err;
 471 }
 472 EXPORT_SYMBOL(ore_check_io);
 473
 474 /*
 475  * L - logical offset into the file
 476  *
 477  * D - number of Data devices
 478  *      D = group_width - parity
 479  *
 480  * U - The number of bytes in a stripe within a group
 481  *      U =  stripe_unit * D
 482  *
 483  * T - The number of bytes striped within a group of component objects
 484  *     (before advancing to the next group)
 485  *      T = U * group_depth
 486  *
 487  * S - The number of bytes striped across all component objects
 488  *     before the pattern repeats
 489  *      S = T * group_count
 490  *
 491  * M - The "major" (i.e., across all components) cycle number
 492  *      M = L / S
 493  *
 494  * G - Counts the groups from the beginning of the major cycle
 495  *      G = (L - (M * S)) / T   [or (L % S) / T]
 496  *
 497  * H - The byte offset within the group
 498  *      H = (L - (M * S)) % T   [or (L % S) % T]
 499  *
 500  * N - The "minor" (i.e., across the group) stripe number
 501  *      N = H / U
 502  *
 503  * C - The component index coresponding to L
 504  *
 505  *      C = (H - (N * U)) / stripe_unit + G * D
 506  *      [or (L % U) / stripe_unit + G * D]
 507  *
 508  * O - The component offset coresponding to L
 509  *      O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
 510  *
 511  * LCMdP – Parity cycle: Lowest Common Multiple of group_width, parity
 512  *          divide by parity
 513  *      LCMdP = lcm(group_width, parity) / parity
 514  *
 515  * R - The parity Rotation stripe
 516  *     (Note parity cycle always starts at a group's boundary)
 517  *      R = N % LCMdP
 518  *
 519  * I = the first parity device index
 520  *      I = (group_width + group_width - R*parity - parity) % group_width
 521  *
 522  * Craid - The component index Rotated
 523  *      Craid = (group_width + C - R*parity) % group_width
 524  *      (We add the group_width to avoid negative numbers modulo math)
 525  */
 526 void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
 527                           u64 length, struct ore_striping_info *si)
 528 {
 529         u32     stripe_unit = layout->stripe_unit;
 530         u32     group_width = layout->group_width;
 531         u64     group_depth = layout->group_depth;
 532         u32     parity      = layout->parity;
 533
 534         u32     D = group_width - parity;
 535         u32     U = D * stripe_unit;
 536         u64     T = U * group_depth;
 537         u64     S = T * layout->group_count;
 538         u64     M = div64_u64(file_offset, S);
 539
 540         /*
 541         G = (L - (M * S)) / T
 542         H = (L - (M * S)) % T
 543         */
 544         u64     LmodS = file_offset - M * S;
 545         u32     G = div64_u64(LmodS, T);
 546         u64     H = LmodS - G * T;
 547
 548         u32     N = div_u64(H, U);
 549         u32     Nlast;
 550
 551         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
 552         u32     C = (u32)(H - (N * U)) / stripe_unit + G * group_width;
 553         u32 first_dev = C - C % group_width;
 554
 555         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
 556
 557         si->obj_offset = si->unit_off + (N * stripe_unit) +
 558                                   (M * group_depth * stripe_unit);
 559         si->cur_comp = C - first_dev;
 560         si->cur_pg = si->unit_off / PAGE_SIZE;
 561
 562         if (parity) {
 563                 u32 LCMdP = lcm(group_width, parity) / parity;
 564                 /* R     = N % LCMdP; */
 565                 u32 RxP   = (N % LCMdP) * parity;
 566
 567                 si->par_dev = (group_width + group_width - parity - RxP) %
 568                               group_width + first_dev;
 569                 si->dev = (group_width + group_width + C - RxP) %
 570                           group_width + first_dev;
 571                 si->bytes_in_stripe = U;
 572                 si->first_stripe_start = M * S + G * T + N * U;
 573         } else {
 574                 /* Make the math correct see _prepare_one_group */
 575                 si->par_dev = group_width;
 576                 si->dev = C;
 577         }
 578
 579         si->dev *= layout->mirrors_p1;
 580         si->par_dev *= layout->mirrors_p1;
 581         si->offset = file_offset;
 582         si->length = T - H;
 583         if (si->length > length)
 584                 si->length = length;
 585
 586         Nlast = div_u64(H + si->length + U - 1, U);
 587         si->maxdevUnits = Nlast - N;
 588
 589         si->M = M;
 590 }
 591 EXPORT_SYMBOL(ore_calc_stripe_info);
 592
 593 int _ore_add_stripe_unit(struct ore_io_state *ios,  unsigned *cur_pg,
 594                          unsigned pgbase, struct page **pages,
 595                          struct ore_per_dev_state *per_dev, int cur_len)
 596 {
 597         unsigned pg = *cur_pg;
 598         struct request_queue *q =
 599                         osd_request_queue(_ios_od(ios, per_dev->dev));
 600         unsigned len = cur_len;
 601         int ret;
 602
 603         if (per_dev->bio == NULL) {
 604                 unsigned bio_size;
 605
 606                 if (!ios->reading) {
 607                         bio_size = ios->si.maxdevUnits;
 608                 } else {
 609                         bio_size = (ios->si.maxdevUnits + 1) *
 610                              (ios->layout->group_width - ios->layout->parity) /
 611                              ios->layout->group_width;
 612                 }
 613                 bio_size *= (ios->layout->stripe_unit / PAGE_SIZE);
 614
 615                 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
 616                 if (unlikely(!per_dev->bio)) {
 617                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
 618                                      bio_size);
 619                         ret = -ENOMEM;
 620                         goto out;
 621                 }
 622         }
 623
 624         while (cur_len > 0) {
 625                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
 626                 unsigned added_len;
 627
 628                 cur_len -= pglen;
 629
 630                 added_len = bio_add_pc_page(q, per_dev->bio, pages[pg],
 631                                             pglen, pgbase);
 632                 if (unlikely(pglen != added_len)) {
 633                         /* If bi_vcnt == bi_max then this is a SW BUG */
 634                         ORE_DBGMSG("Failed bio_add_pc_page bi_vcnt=0x%x "
 635                                    "bi_max=0x%x BIO_MAX=0x%x cur_len=0x%x\n",
 636                                    per_dev->bio->bi_vcnt,
 637                                    per_dev->bio->bi_max_vecs,
 638                                    BIO_MAX_PAGES_KMALLOC, cur_len);
 639                         ret = -ENOMEM;
 640                         goto out;
 641                 }
 642                 _add_stripe_page(ios->sp2d, &ios->si, pages[pg]);
 643
 644                 pgbase = 0;
 645                 ++pg;
 646         }
 647         BUG_ON(cur_len);
 648
 649         per_dev->length += len;
 650         *cur_pg = pg;
 651         ret = 0;
 652 out:    /* we fail the complete unit on an error eg don't advance
 653          * per_dev->length and cur_pg. This means that we might have a bigger
 654          * bio than the CDB requested length (per_dev->length). That's fine
 655          * only the oposite is fatal.
 656          */
 657         return ret;
 658 }
 659
 660 static int _add_parity_units(struct ore_io_state *ios,
 661                              struct ore_striping_info *si,
 662                              unsigned dev, unsigned first_dev,
 663                              unsigned mirrors_p1, unsigned devs_in_group,
 664                              unsigned cur_len)
 665 {
 666         unsigned do_parity;
 667         int ret = 0;
 668
 669         for (do_parity = ios->layout->parity; do_parity; --do_parity) {
 670                 struct ore_per_dev_state *per_dev;
 671
 672                 per_dev = &ios->per_dev[dev - first_dev];
 673                 if (!per_dev->length && !per_dev->offset) {
 674                         /* Only/always the parity unit of the first
 675                          * stripe will be empty. So this is a chance to
 676                          * initialize the per_dev info.
 677                          */
 678                         per_dev->dev = dev;
 679                         per_dev->offset = si->obj_offset - si->unit_off;
 680                 }
 681
 682                 ret = _ore_add_parity_unit(ios, si, per_dev, cur_len,
 683                                            do_parity == 1);
 684                 if (unlikely(ret))
 685                                 break;
 686
 687                 if (do_parity != 1) {
 688                         dev = ((dev + mirrors_p1) % devs_in_group) + first_dev;
 689                         si->cur_comp = (si->cur_comp + 1) %
 690                                                        ios->layout->group_width;
 691                 }
 692         }
 693
 694         return ret;
 695 }
 696
 697 static int _prepare_for_striping(struct ore_io_state *ios)
 698 {
 699         struct ore_striping_info *si = &ios->si;
 700         unsigned stripe_unit = ios->layout->stripe_unit;
 701         unsigned mirrors_p1 = ios->layout->mirrors_p1;
 702         unsigned group_width = ios->layout->group_width;
 703         unsigned devs_in_group = group_width * mirrors_p1;
 704         unsigned dev = si->dev;
 705         unsigned first_dev = dev - (dev % devs_in_group);
 706         unsigned cur_pg = ios->pages_consumed;
 707         u64 length = ios->length;
 708         int ret = 0;
 709
 710         if (!ios->pages) {
 711                 ios->numdevs = ios->layout->mirrors_p1;
 712                 return 0;
 713         }
 714
 715         BUG_ON(length > si->length);
 716
 717         while (length) {
 718                 struct ore_per_dev_state *per_dev =
 719                                                 &ios->per_dev[dev - first_dev];
 720                 unsigned cur_len, page_off = 0;
 721
 722                 if (!per_dev->length && !per_dev->offset) {
 723                         /* First time initialize the per_dev info. */
 724                         per_dev->dev = dev;
 725                         if (dev == si->dev) {
 726                                 WARN_ON(dev == si->par_dev);
 727                                 per_dev->offset = si->obj_offset;
 728                                 cur_len = stripe_unit - si->unit_off;
 729                                 page_off = si->unit_off & ~PAGE_MASK;
 730                                 BUG_ON(page_off && (page_off != ios->pgbase));
 731                         } else {
 732                                 per_dev->offset = si->obj_offset - si->unit_off;
 733                                 cur_len = stripe_unit;
 734                         }
 735                 } else {
 736                         cur_len = stripe_unit;
 737                 }
 738                 if (cur_len >= length)
 739                         cur_len = length;
 740
 741                 ret = _ore_add_stripe_unit(ios, &cur_pg, page_off, ios->pages,
 742                                            per_dev, cur_len);
 743                 if (unlikely(ret))
 744                         goto out;
 745
 746                 length -= cur_len;
 747
 748                 dev = ((dev + mirrors_p1) % devs_in_group) + first_dev;
 749                 si->cur_comp = (si->cur_comp + 1) % group_width;
 750                 if (unlikely((dev == si->par_dev) || (!length && ios->sp2d))) {
 751                         if (!length && ios->sp2d) {
 752                                 /* If we are writing and this is the very last
 753                                  * stripe. then operate on parity dev.
 754                                  */
 755                                 dev = si->par_dev;
 756                                 /* If last stripe operate on parity comp */
 757                                 si->cur_comp = group_width - ios->layout->parity;
 758                         }
 759
 760                         /* In writes cur_len just means if it's the
 761                          * last one. See _ore_add_parity_unit.
 762                          */
 763                         ret = _add_parity_units(ios, si, dev, first_dev,
 764                                                 mirrors_p1, devs_in_group,
 765                                                 ios->sp2d ? length : cur_len);
 766                         if (unlikely(ret))
 767                                         goto out;
 768
 769                         /* Rotate next par_dev backwards with wraping */
 770                         si->par_dev = (devs_in_group + si->par_dev -
 771                                        ios->layout->parity * mirrors_p1) %
 772                                       devs_in_group + first_dev;
 773                         /* Next stripe, start fresh */
 774                         si->cur_comp = 0;
 775                         si->cur_pg = 0;
 776                         si->obj_offset += cur_len;
 777                         si->unit_off = 0;
 778                 }
 779         }
 780 out:
 781         ios->numdevs = devs_in_group;
 782         ios->pages_consumed = cur_pg;
 783         return ret;
 784 }
 785
 786 int ore_create(struct ore_io_state *ios)
 787 {
 788         int i, ret;
 789
 790         for (i = 0; i < ios->oc->numdevs; i++) {
 791                 struct osd_request *or;
 792
 793                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 794                 if (unlikely(!or)) {
 795                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 796                         ret = -ENOMEM;
 797                         goto out;
 798                 }
 799                 ios->per_dev[i].or = or;
 800                 ios->numdevs++;
 801
 802                 osd_req_create_object(or, _ios_obj(ios, i));
 803         }
 804         ret = ore_io_execute(ios);
 805
 806 out:
 807         return ret;
 808 }
 809 EXPORT_SYMBOL(ore_create);
 810
 811 int ore_remove(struct ore_io_state *ios)
 812 {
 813         int i, ret;
 814
 815         for (i = 0; i < ios->oc->numdevs; i++) {
 816                 struct osd_request *or;
 817
 818                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 819                 if (unlikely(!or)) {
 820                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 821                         ret = -ENOMEM;
 822                         goto out;
 823                 }
 824                 ios->per_dev[i].or = or;
 825                 ios->numdevs++;
 826
 827                 osd_req_remove_object(or, _ios_obj(ios, i));
 828         }
 829         ret = ore_io_execute(ios);
 830
 831 out:
 832         return ret;
 833 }
 834 EXPORT_SYMBOL(ore_remove);
 835
 836 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
 837 {
 838         struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
 839         unsigned dev = ios->per_dev[cur_comp].dev;
 840         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
 841         int ret = 0;
 842
 843         if (ios->pages && !master_dev->length)
 844                 return 0; /* Just an empty slot */
 845
 846         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
 847                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 848                 struct osd_request *or;
 849
 850                 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
 851                 if (unlikely(!or)) {
 852                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 853                         ret = -ENOMEM;
 854                         goto out;
 855                 }
 856                 per_dev->or = or;
 857
 858                 if (ios->pages) {
 859                         struct bio *bio;
 860
 861                         if (per_dev != master_dev) {
 862                                 bio = bio_clone_kmalloc(master_dev->bio,
 863                                                         GFP_KERNEL);
 864                                 if (unlikely(!bio)) {
 865                                         ORE_DBGMSG(
 866                                               "Failed to allocate BIO size=%u\n",
 867                                               master_dev->bio->bi_max_vecs);
 868                                         ret = -ENOMEM;
 869                                         goto out;
 870                                 }
 871
 872                                 bio->bi_bdev = NULL;
 873                                 bio->bi_next = NULL;
 874                                 per_dev->offset = master_dev->offset;
 875                                 per_dev->length = master_dev->length;
 876                                 per_dev->bio =  bio;
 877                                 per_dev->dev = dev;
 878                         } else {
 879                                 bio = master_dev->bio;
 880                                 /* FIXME: bio_set_dir() */
 881                                 bio->bi_rw |= REQ_WRITE;
 882                         }
 883
 884                         osd_req_write(or, _ios_obj(ios, cur_comp),
 885                                       per_dev->offset, bio, per_dev->length);
 886                         ORE_DBGMSG("write(0x%llx) offset=0x%llx "
 887                                       "length=0x%llx dev=%d\n",
 888                                      _LLU(_ios_obj(ios, cur_comp)->id),
 889                                      _LLU(per_dev->offset),
 890                                      _LLU(per_dev->length), dev);
 891                 } else if (ios->kern_buff) {
 892                         per_dev->offset = ios->si.obj_offset;
 893                         per_dev->dev = ios->si.dev + dev;
 894
 895                         /* no cross device without page array */
 896                         BUG_ON((ios->layout->group_width > 1) &&
 897                                (ios->si.unit_off + ios->length >
 898                                 ios->layout->stripe_unit));
 899
 900                         ret = osd_req_write_kern(or, _ios_obj(ios, cur_comp),
 901                                                  per_dev->offset,
 902                                                  ios->kern_buff, ios->length);
 903                         if (unlikely(ret))
 904                                 goto out;
 905                         ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
 906                                       "length=0x%llx dev=%d\n",
 907                                      _LLU(_ios_obj(ios, cur_comp)->id),
 908                                      _LLU(per_dev->offset),
 909                                      _LLU(ios->length), per_dev->dev);
 910                 } else {
 911                         osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
 912                         ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
 913                                      _LLU(_ios_obj(ios, cur_comp)->id),
 914                                      ios->out_attr_len, dev);
 915                 }
 916
 917                 if (ios->out_attr)
 918                         osd_req_add_set_attr_list(or, ios->out_attr,
 919                                                   ios->out_attr_len);
 920
 921                 if (ios->in_attr)
 922                         osd_req_add_get_attr_list(or, ios->in_attr,
 923                                                   ios->in_attr_len);
 924         }
 925
 926 out:
 927         return ret;
 928 }
 929
 930 int ore_write(struct ore_io_state *ios)
 931 {
 932         int i;
 933         int ret;
 934
 935         if (unlikely(ios->sp2d && !ios->r4w)) {
 936                 /* A library is attempting a RAID-write without providing
 937                  * a pages lock interface.
 938                  */
 939                 WARN_ON_ONCE(1);
 940                 return -ENOTSUPP;
 941         }
 942
 943         ret = _prepare_for_striping(ios);
 944         if (unlikely(ret))
 945                 return ret;
 946
 947         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 948                 ret = _write_mirror(ios, i);
 949                 if (unlikely(ret))
 950                         return ret;
 951         }
 952
 953         ret = ore_io_execute(ios);
 954         return ret;
 955 }
 956 EXPORT_SYMBOL(ore_write);
 957
 958 int _ore_read_mirror(struct ore_io_state *ios, unsigned cur_comp)
 959 {
 960         struct osd_request *or;
 961         struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 962         struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
 963         unsigned first_dev = (unsigned)obj->id;
 964
 965         if (ios->pages && !per_dev->length)
 966                 return 0; /* Just an empty slot */
 967
 968         first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
 969         or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
 970         if (unlikely(!or)) {
 971                 ORE_ERR("%s: osd_start_request failed\n", __func__);
 972                 return -ENOMEM;
 973         }
 974         per_dev->or = or;
 975
 976         if (ios->pages) {
 977                 if (per_dev->cur_sg) {
 978                         /* finalize the last sg_entry */
 979                         _ore_add_sg_seg(per_dev, 0, false);
 980                         if (unlikely(!per_dev->cur_sg))
 981                                 return 0; /* Skip parity only device */
 982
 983                         osd_req_read_sg(or, obj, per_dev->bio,
 984                                         per_dev->sglist, per_dev->cur_sg);
 985                 } else {
 986                         /* The no raid case */
 987                         osd_req_read(or, obj, per_dev->offset,
 988                                      per_dev->bio, per_dev->length);
 989                 }
 990
 991                 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
 992                              " dev=%d sg_len=%d\n", _LLU(obj->id),
 993                              _LLU(per_dev->offset), _LLU(per_dev->length),
 994                              first_dev, per_dev->cur_sg);
 995         } else {
 996                 BUG_ON(ios->kern_buff);
 997
 998                 osd_req_get_attributes(or, obj);
 999                 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
1000                               _LLU(obj->id),
1001                               ios->in_attr_len, first_dev);
1002         }
1003         if (ios->out_attr)
1004                 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
1005
1006         if (ios->in_attr)
1007                 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
1008
1009         return 0;
1010 }
1011
1012 int ore_read(struct ore_io_state *ios)
1013 {
1014         int i;
1015         int ret;
1016
1017         ret = _prepare_for_striping(ios);
1018         if (unlikely(ret))
1019                 return ret;
1020
1021         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
1022                 ret = _ore_read_mirror(ios, i);
1023                 if (unlikely(ret))
1024                         return ret;
1025         }
1026
1027         ret = ore_io_execute(ios);
1028         return ret;
1029 }
1030 EXPORT_SYMBOL(ore_read);
1031
1032 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
1033 {
1034         struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
1035         void *iter = NULL;
1036         int nelem;
1037
1038         do {
1039                 nelem = 1;
1040                 osd_req_decode_get_attr_list(ios->per_dev[0].or,
1041                                              &cur_attr, &nelem, &iter);
1042                 if ((cur_attr.attr_page == attr->attr_page) &&
1043                     (cur_attr.attr_id == attr->attr_id)) {
1044                         attr->len = cur_attr.len;
1045                         attr->val_ptr = cur_attr.val_ptr;
1046                         return 0;
1047                 }
1048         } while (iter);
1049
1050         return -EIO;
1051 }
1052 EXPORT_SYMBOL(extract_attr_from_ios);
1053
1054 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
1055                              struct osd_attr *attr)
1056 {
1057         int last_comp = cur_comp + ios->layout->mirrors_p1;
1058
1059         for (; cur_comp < last_comp; ++cur_comp) {
1060                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
1061                 struct osd_request *or;
1062
1063                 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
1064                 if (unlikely(!or)) {
1065                         ORE_ERR("%s: osd_start_request failed\n", __func__);
1066                         return -ENOMEM;
1067                 }
1068                 per_dev->or = or;
1069
1070                 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
1071                 osd_req_add_set_attr_list(or, attr, 1);
1072         }
1073
1074         return 0;
1075 }
1076
1077 struct _trunc_info {
1078         struct ore_striping_info si;
1079         u64 prev_group_obj_off;
1080         u64 next_group_obj_off;
1081
1082         unsigned first_group_dev;
1083         unsigned nex_group_dev;
1084 };
1085
1086 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
1087                              struct _trunc_info *ti)
1088 {
1089         unsigned stripe_unit = layout->stripe_unit;
1090
1091         ore_calc_stripe_info(layout, file_offset, 0, &ti->si);
1092
1093         ti->prev_group_obj_off = ti->si.M * stripe_unit;
1094         ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
1095
1096         ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
1097         ti->nex_group_dev = ti->first_group_dev + layout->group_width;
1098 }
1099
1100 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
1101                    u64 size)
1102 {
1103         struct ore_io_state *ios;
1104         struct exofs_trunc_attr {
1105                 struct osd_attr attr;
1106                 __be64 newsize;
1107         } *size_attrs;
1108         struct _trunc_info ti;
1109         int i, ret;
1110
1111         ret = ore_get_io_state(layout, oc, &ios);
1112         if (unlikely(ret))
1113                 return ret;
1114
1115         _calc_trunk_info(ios->layout, size, &ti);
1116
1117         size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
1118                              GFP_KERNEL);
1119         if (unlikely(!size_attrs)) {
1120                 ret = -ENOMEM;
1121                 goto out;
1122         }
1123
1124         ios->numdevs = ios->oc->numdevs;
1125
1126         for (i = 0; i < ios->numdevs; ++i) {
1127                 struct exofs_trunc_attr *size_attr = &size_attrs[i];
1128                 u64 obj_size;
1129
1130                 if (i < ti.first_group_dev)
1131                         obj_size = ti.prev_group_obj_off;
1132                 else if (i >= ti.nex_group_dev)
1133                         obj_size = ti.next_group_obj_off;
1134                 else if (i < ti.si.dev) /* dev within this group */
1135                         obj_size = ti.si.obj_offset +
1136                                       ios->layout->stripe_unit - ti.si.unit_off;
1137                 else if (i == ti.si.dev)
1138                         obj_size = ti.si.obj_offset;
1139                 else /* i > ti.dev */
1140                         obj_size = ti.si.obj_offset - ti.si.unit_off;
1141
1142                 size_attr->newsize = cpu_to_be64(obj_size);
1143                 size_attr->attr = g_attr_logical_length;
1144                 size_attr->attr.val_ptr = &size_attr->newsize;
1145
1146                 ORE_DBGMSG2("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
1147                              _LLU(oc->comps->obj.id), _LLU(obj_size), i);
1148                 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
1149                                         &size_attr->attr);
1150                 if (unlikely(ret))
1151                         goto out;
1152         }
1153         ret = ore_io_execute(ios);
1154
1155 out:
1156         kfree(size_attrs);
1157         ore_put_io_state(ios);
1158         return ret;
1159 }
1160 EXPORT_SYMBOL(ore_truncate);
1161
1162 const struct osd_attr g_attr_logical_length = ATTR_DEF(
1163         OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
1164 EXPORT_SYMBOL(g_attr_logical_length);