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