fs/exofs/ore.c

   1 /*
   2  * Copyright (C) 2005, 2006
   3  * Avishay Traeger (avishay@gmail.com)
   4  * Copyright (C) 2008, 2009
   5  * Boaz Harrosh <bharrosh@panasas.com>
   6  *
   7  * This file is part of exofs.
   8  *
   9  * exofs is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation.  Since it is based on ext2, and the only
  12  * valid version of GPL for the Linux kernel is version 2, the only valid
  13  * version of GPL for exofs is version 2.
  14  *
  15  * exofs is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with exofs; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  23  */
  24
  25 #include <linux/slab.h>
  26 #include <asm/div64.h>
  27 #include <linux/lcm.h>
  28
  29 #include "ore_raid.h"
  30
  31 MODULE_AUTHOR("Boaz Harrosh <bharrosh@panasas.com>");
  32 MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
  33 MODULE_LICENSE("GPL");
  34
  35 /* ore_verify_layout does a couple of things:
  36  * 1. Given a minimum number of needed parameters fixes up the rest of the
  37  *    members to be operatonals for the ore. The needed parameters are those
  38  *    that are defined by the pnfs-objects layout STD.
  39  * 2. Check to see if the current ore code actually supports these parameters
  40  *    for example stripe_unit must be a multple of the system PAGE_SIZE,
  41  *    and etc...
  42  * 3. Cache some havily used calculations that will be needed by users.
  43  */
  44
  45 enum { BIO_MAX_PAGES_KMALLOC =
  46                 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),};
  47
  48 int ore_verify_layout(unsigned total_comps, struct ore_layout *layout)
  49 {
  50         u64 stripe_length;
  51
  52         switch (layout->raid_algorithm) {
  53         case PNFS_OSD_RAID_0:
  54                 layout->parity = 0;
  55                 break;
  56         case PNFS_OSD_RAID_5:
  57                 layout->parity = 1;
  58                 break;
  59         case PNFS_OSD_RAID_PQ:
  60         case PNFS_OSD_RAID_4:
  61         default:
  62                 ORE_ERR("Only RAID_0/5 for now\n");
  63                 return -EINVAL;
  64         }
  65         if (0 != (layout->stripe_unit & ~PAGE_MASK)) {
  66                 ORE_ERR("Stripe Unit(0x%llx)"
  67                           " must be Multples of PAGE_SIZE(0x%lx)\n",
  68                           _LLU(layout->stripe_unit), PAGE_SIZE);
  69                 return -EINVAL;
  70         }
  71         if (layout->group_width) {
  72                 if (!layout->group_depth) {
  73                         ORE_ERR("group_depth == 0 && group_width != 0\n");
  74                         return -EINVAL;
  75                 }
  76                 if (total_comps < (layout->group_width * layout->mirrors_p1)) {
  77                         ORE_ERR("Data Map wrong, "
  78                                 "numdevs=%d < group_width=%d * mirrors=%d\n",
  79                                 total_comps, layout->group_width,
  80                                 layout->mirrors_p1);
  81                         return -EINVAL;
  82                 }
  83                 layout->group_count = total_comps / layout->mirrors_p1 /
  84                                                 layout->group_width;
  85         } else {
  86                 if (layout->group_depth) {
  87                         printk(KERN_NOTICE "Warning: group_depth ignored "
  88                                 "group_width == 0 && group_depth == %lld\n",
  89                                 _LLU(layout->group_depth));
  90                 }
  91                 layout->group_width = total_comps / layout->mirrors_p1;
  92                 layout->group_depth = -1;
  93                 layout->group_count = 1;
  94         }
  95
  96         stripe_length = (u64)layout->group_width * layout->stripe_unit;
  97         if (stripe_length >= (1ULL << 32)) {
  98                 ORE_ERR("Stripe_length(0x%llx) >= 32bit is not supported\n",
  99                         _LLU(stripe_length));
 100                 return -EINVAL;
 101         }
 102
 103         layout->max_io_length =
 104                 (BIO_MAX_PAGES_KMALLOC * PAGE_SIZE - layout->stripe_unit) *
 105                                                         layout->group_width;
 106         if (layout->parity) {
 107                 unsigned stripe_length =
 108                                 (layout->group_width - layout->parity) *
 109                                 layout->stripe_unit;
 110
 111                 layout->max_io_length /= stripe_length;
 112                 layout->max_io_length *= stripe_length;
 113         }
 114         return 0;
 115 }
 116 EXPORT_SYMBOL(ore_verify_layout);
 117
 118 static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
 119 {
 120         return ios->oc->comps[index & ios->oc->single_comp].cred;
 121 }
 122
 123 static struct osd_obj_id *_ios_obj(struct ore_io_state *ios, unsigned index)
 124 {
 125         return &ios->oc->comps[index & ios->oc->single_comp].obj;
 126 }
 127
 128 static struct osd_dev *_ios_od(struct ore_io_state *ios, unsigned index)
 129 {
 130         ORE_DBGMSG2("oc->first_dev=%d oc->numdevs=%d i=%d oc->ods=%p\n",
 131                     ios->oc->first_dev, ios->oc->numdevs, index,
 132                     ios->oc->ods);
 133
 134         return ore_comp_dev(ios->oc, index);
 135 }
 136
 137 int  _ore_get_io_state(struct ore_layout *layout,
 138                         struct ore_components *oc, unsigned numdevs,
 139                         unsigned sgs_per_dev, unsigned num_par_pages,
 140                         struct ore_io_state **pios)
 141 {
 142         struct ore_io_state *ios;
 143         struct page **pages;
 144         struct osd_sg_entry *sgilist;
 145         struct __alloc_all_io_state {
 146                 struct ore_io_state ios;
 147                 struct ore_per_dev_state per_dev[numdevs];
 148                 union {
 149                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
 150                         struct page *pages[num_par_pages];
 151                 };
 152         } *_aios;
 153
 154         if (likely(sizeof(*_aios) <= PAGE_SIZE)) {
 155                 _aios = kzalloc(sizeof(*_aios), GFP_KERNEL);
 156                 if (unlikely(!_aios)) {
 157                         ORE_DBGMSG("Failed kzalloc bytes=%zd\n",
 158                                    sizeof(*_aios));
 159                         *pios = NULL;
 160                         return -ENOMEM;
 161                 }
 162                 pages = num_par_pages ? _aios->pages : NULL;
 163                 sgilist = sgs_per_dev ? _aios->sglist : NULL;
 164                 ios = &_aios->ios;
 165         } else {
 166                 struct __alloc_small_io_state {
 167                         struct ore_io_state ios;
 168                         struct ore_per_dev_state per_dev[numdevs];
 169                 } *_aio_small;
 170                 union __extra_part {
 171                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
 172                         struct page *pages[num_par_pages];
 173                 } *extra_part;
 174
 175                 _aio_small = kzalloc(sizeof(*_aio_small), GFP_KERNEL);
 176                 if (unlikely(!_aio_small)) {
 177                         ORE_DBGMSG("Failed alloc first part bytes=%zd\n",
 178                                    sizeof(*_aio_small));
 179                         *pios = NULL;
 180                         return -ENOMEM;
 181                 }
 182                 extra_part = kzalloc(sizeof(*extra_part), GFP_KERNEL);
 183                 if (unlikely(!extra_part)) {
 184                         ORE_DBGMSG("Failed alloc second part bytes=%zd\n",
 185                                    sizeof(*extra_part));
 186                         kfree(_aio_small);
 187                         *pios = NULL;
 188                         return -ENOMEM;
 189                 }
 190
 191                 pages = num_par_pages ? extra_part->pages : NULL;
 192                 sgilist = sgs_per_dev ? extra_part->sglist : NULL;
 193                 /* In this case the per_dev[0].sgilist holds the pointer to
 194                  * be freed
 195                  */
 196                 ios = &_aio_small->ios;
 197                 ios->extra_part_alloc = true;
 198         }
 199
 200         if (pages) {
 201                 ios->parity_pages = pages;
 202                 ios->max_par_pages = num_par_pages;
 203         }
 204         if (sgilist) {
 205                 unsigned d;
 206
 207                 for (d = 0; d < numdevs; ++d) {
 208                         ios->per_dev[d].sglist = sgilist;
 209                         sgilist += sgs_per_dev;
 210                 }
 211                 ios->sgs_per_dev = sgs_per_dev;
 212         }
 213
 214         ios->layout = layout;
 215         ios->oc = oc;
 216         *pios = ios;
 217         return 0;
 218 }
 219
 220 /* Allocate an io_state for only a single group of devices
 221  *
 222  * If a user needs to call ore_read/write() this version must be used becase it
 223  * allocates extra stuff for striping and raid.
 224  * The ore might decide to only IO less then @length bytes do to alignmets
 225  * and constrains as follows:
 226  * - The IO cannot cross group boundary.
 227  * - In raid5/6 The end of the IO must align at end of a stripe eg.
 228  *   (@offset + @length) % strip_size == 0. Or the complete range is within a
 229  *   single stripe.
 230  * - Memory condition only permitted a shorter IO. (A user can use @length=~0
 231  *   And check the returned ios->length for max_io_size.)
 232  *
 233  * The caller must check returned ios->length (and/or ios->nr_pages) and
 234  * re-issue these pages that fall outside of ios->length
 235  */
 236 int  ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
 237                       bool is_reading, u64 offset, u64 length,
 238                       struct ore_io_state **pios)
 239 {
 240         struct ore_io_state *ios;
 241         unsigned numdevs = layout->group_width * layout->mirrors_p1;
 242         unsigned sgs_per_dev = 0, max_par_pages = 0;
 243         int ret;
 244
 245         if (layout->parity && length) {
 246                 unsigned data_devs = layout->group_width - layout->parity;
 247                 unsigned stripe_size = layout->stripe_unit * data_devs;
 248                 unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
 249                 u32 remainder;
 250                 u64 num_stripes;
 251                 u64 num_raid_units;
 252
 253                 num_stripes = div_u64_rem(length, stripe_size, &remainder);
 254                 if (remainder)
 255                         ++num_stripes;
 256
 257                 num_raid_units =  num_stripes * layout->parity;
 258
 259                 if (is_reading) {
 260                         /* For reads add per_dev sglist array */
 261                         /* TODO: Raid 6 we need twice more. Actually:
 262                         *         num_stripes / LCMdP(W,P);
 263                         *         if (W%P != 0) num_stripes *= parity;
 264                         */
 265
 266                         /* first/last seg is split */
 267                         num_raid_units += layout->group_width;
 268                         sgs_per_dev = div_u64(num_raid_units, data_devs);
 269                 } else {
 270                         /* For Writes add parity pages array. */
 271                         max_par_pages = num_raid_units * pages_in_unit *
 272                                                 sizeof(struct page *);
 273                 }
 274         }
 275
 276         ret = _ore_get_io_state(layout, oc, numdevs, sgs_per_dev, max_par_pages,
 277                                 pios);
 278         if (unlikely(ret))
 279                 return ret;
 280
 281         ios = *pios;
 282         ios->reading = is_reading;
 283         ios->offset = offset;
 284
 285         if (length) {
 286                 ore_calc_stripe_info(layout, offset, length, &ios->si);
 287                 ios->length = ios->si.length;
 288                 ios->nr_pages = (ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
 289                 if (layout->parity)
 290                         _ore_post_alloc_raid_stuff(ios);
 291         }
 292
 293         return 0;
 294 }
 295 EXPORT_SYMBOL(ore_get_rw_state);
 296
 297 /* Allocate an io_state for all the devices in the comps array
 298  *
 299  * This version of io_state allocation is used mostly by create/remove
 300  * and trunc where we currently need all the devices. The only wastful
 301  * bit is the read/write_attributes with no IO. Those sites should
 302  * be converted to use ore_get_rw_state() with length=0
 303  */
 304 int  ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
 305                       struct ore_io_state **pios)
 306 {
 307         return _ore_get_io_state(layout, oc, oc->numdevs, 0, 0, pios);
 308 }
 309 EXPORT_SYMBOL(ore_get_io_state);
 310
 311 void ore_put_io_state(struct ore_io_state *ios)
 312 {
 313         if (ios) {
 314                 unsigned i;
 315
 316                 for (i = 0; i < ios->numdevs; i++) {
 317                         struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 318
 319                         if (per_dev->or)
 320                                 osd_end_request(per_dev->or);
 321                         if (per_dev->bio)
 322                                 bio_put(per_dev->bio);
 323                 }
 324
 325                 _ore_free_raid_stuff(ios);
 326                 kfree(ios);
 327         }
 328 }
 329 EXPORT_SYMBOL(ore_put_io_state);
 330
 331 static void _sync_done(struct ore_io_state *ios, void *p)
 332 {
 333         struct completion *waiting = p;
 334
 335         complete(waiting);
 336 }
 337
 338 static void _last_io(struct kref *kref)
 339 {
 340         struct ore_io_state *ios = container_of(
 341                                         kref, struct ore_io_state, kref);
 342
 343         ios->done(ios, ios->private);
 344 }
 345
 346 static void _done_io(struct osd_request *or, void *p)
 347 {
 348         struct ore_io_state *ios = p;
 349
 350         kref_put(&ios->kref, _last_io);
 351 }
 352
 353 int ore_io_execute(struct ore_io_state *ios)
 354 {
 355         DECLARE_COMPLETION_ONSTACK(wait);
 356         bool sync = (ios->done == NULL);
 357         int i, ret;
 358
 359         if (sync) {
 360                 ios->done = _sync_done;
 361                 ios->private = &wait;
 362         }
 363
 364         for (i = 0; i < ios->numdevs; i++) {
 365                 struct osd_request *or = ios->per_dev[i].or;
 366                 if (unlikely(!or))
 367                         continue;
 368
 369                 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
 370                 if (unlikely(ret)) {
 371                         ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
 372                                      ret);
 373                         return ret;
 374                 }
 375         }
 376
 377         kref_init(&ios->kref);
 378
 379         for (i = 0; i < ios->numdevs; i++) {
 380                 struct osd_request *or = ios->per_dev[i].or;
 381                 if (unlikely(!or))
 382                         continue;
 383
 384                 kref_get(&ios->kref);
 385                 osd_execute_request_async(or, _done_io, ios);
 386         }
 387
 388         kref_put(&ios->kref, _last_io);
 389         ret = 0;
 390
 391         if (sync) {
 392                 wait_for_completion(&wait);
 393                 ret = ore_check_io(ios, NULL);
 394         }
 395         return ret;
 396 }
 397
 398 static void _clear_bio(struct bio *bio)
 399 {
 400         struct bio_vec *bv;
 401         unsigned i;
 402
 403         __bio_for_each_segment(bv, bio, i, 0) {
 404                 unsigned this_count = bv->bv_len;
 405
 406                 if (likely(PAGE_SIZE == this_count))
 407                         clear_highpage(bv->bv_page);
 408                 else
 409                         zero_user(bv->bv_page, bv->bv_offset, this_count);
 410         }
 411 }
 412
 413 int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
 414 {
 415         enum osd_err_priority acumulated_osd_err = 0;
 416         int acumulated_lin_err = 0;
 417         int i;
 418
 419         for (i = 0; i < ios->numdevs; i++) {
 420                 struct osd_sense_info osi;
 421                 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 422                 struct osd_request *or = per_dev->or;
 423                 int ret;
 424
 425                 if (unlikely(!or))
 426                         continue;
 427
 428                 ret = osd_req_decode_sense(or, &osi);
 429                 if (likely(!ret))
 430                         continue;
 431
 432                 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
 433                         /* start read offset passed endof file */
 434                         _clear_bio(per_dev->bio);
 435                         ORE_DBGMSG("start read offset passed end of file "
 436                                 "offset=0x%llx, length=0x%llx\n",
 437                                 _LLU(per_dev->offset),
 438                                 _LLU(per_dev->length));
 439
 440                         continue; /* we recovered */
 441                 }
 442
 443                 if (on_dev_error) {
 444                         u64 residual = ios->reading ?
 445                                         or->in.residual : or->out.residual;
 446                         u64 offset = (ios->offset + ios->length) - residual;
 447                         struct ore_dev *od = ios->oc->ods[
 448                                         per_dev->dev - ios->oc->first_dev];
 449
 450                         on_dev_error(ios, od, per_dev->dev, osi.osd_err_pri,
 451                                      offset, residual);
 452                 }
 453                 if (osi.osd_err_pri >= acumulated_osd_err) {
 454                         acumulated_osd_err = osi.osd_err_pri;
 455                         acumulated_lin_err = ret;
 456                 }
 457         }
 458
 459         return acumulated_lin_err;
 460 }
 461 EXPORT_SYMBOL(ore_check_io);
 462
 463 /*
 464  * L - logical offset into the file
 465  *
 466  * D - number of Data devices
 467  *      D = group_width - parity
 468  *
 469  * U - The number of bytes in a stripe within a group
 470  *      U =  stripe_unit * D
 471  *
 472  * T - The number of bytes striped within a group of component objects
 473  *     (before advancing to the next group)
 474  *      T = U * group_depth
 475  *
 476  * S - The number of bytes striped across all component objects
 477  *     before the pattern repeats
 478  *      S = T * group_count
 479  *
 480  * M - The "major" (i.e., across all components) cycle number
 481  *      M = L / S
 482  *
 483  * G - Counts the groups from the beginning of the major cycle
 484  *      G = (L - (M * S)) / T   [or (L % S) / T]
 485  *
 486  * H - The byte offset within the group
 487  *      H = (L - (M * S)) % T   [or (L % S) % T]
 488  *
 489  * N - The "minor" (i.e., across the group) stripe number
 490  *      N = H / U
 491  *
 492  * C - The component index coresponding to L
 493  *
 494  *      C = (H - (N * U)) / stripe_unit + G * D
 495  *      [or (L % U) / stripe_unit + G * D]
 496  *
 497  * O - The component offset coresponding to L
 498  *      O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
 499  *
 500  * LCMdP – Parity cycle: Lowest Common Multiple of group_width, parity
 501  *          divide by parity
 502  *      LCMdP = lcm(group_width, parity) / parity
 503  *
 504  * R - The parity Rotation stripe
 505  *     (Note parity cycle always starts at a group's boundary)
 506  *      R = N % LCMdP
 507  *
 508  * I = the first parity device index
 509  *      I = (group_width + group_width - R*parity - parity) % group_width
 510  *
 511  * Craid - The component index Rotated
 512  *      Craid = (group_width + C - R*parity) % group_width
 513  *      (We add the group_width to avoid negative numbers modulo math)
 514  */
 515 void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
 516                           u64 length, struct ore_striping_info *si)
 517 {
 518         u32     stripe_unit = layout->stripe_unit;
 519         u32     group_width = layout->group_width;
 520         u64     group_depth = layout->group_depth;
 521         u32     parity      = layout->parity;
 522
 523         u32     D = group_width - parity;
 524         u32     U = D * stripe_unit;
 525         u64     T = U * group_depth;
 526         u64     S = T * layout->group_count;
 527         u64     M = div64_u64(file_offset, S);
 528
 529         /*
 530         G = (L - (M * S)) / T
 531         H = (L - (M * S)) % T
 532         */
 533         u64     LmodS = file_offset - M * S;
 534         u32     G = div64_u64(LmodS, T);
 535         u64     H = LmodS - G * T;
 536
 537         u32     N = div_u64(H, U);
 538
 539         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
 540         u32     C = (u32)(H - (N * U)) / stripe_unit + G * group_width;
 541
 542         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
 543
 544         si->obj_offset = si->unit_off + (N * stripe_unit) +
 545                                   (M * group_depth * stripe_unit);
 546
 547         if (parity) {
 548                 u32 LCMdP = lcm(group_width, parity) / parity;
 549                 /* R     = N % LCMdP; */
 550                 u32 RxP   = (N % LCMdP) * parity;
 551                 u32 first_dev = C - C % group_width;
 552
 553                 si->par_dev = (group_width + group_width - parity - RxP) %
 554                               group_width + first_dev;
 555                 si->dev = (group_width + C - RxP) % group_width + first_dev;
 556                 si->bytes_in_stripe = U;
 557                 si->first_stripe_start = M * S + G * T + N * U;
 558         } else {
 559                 /* Make the math correct see _prepare_one_group */
 560                 si->par_dev = group_width;
 561                 si->dev = C;
 562         }
 563
 564         si->dev *= layout->mirrors_p1;
 565         si->par_dev *= layout->mirrors_p1;
 566         si->offset = file_offset;
 567         si->length = T - H;
 568         if (si->length > length)
 569                 si->length = length;
 570         si->M = M;
 571 }
 572 EXPORT_SYMBOL(ore_calc_stripe_info);
 573
 574 int _ore_add_stripe_unit(struct ore_io_state *ios,  unsigned *cur_pg,
 575                          unsigned pgbase, struct page **pages,
 576                          struct ore_per_dev_state *per_dev, int cur_len)
 577 {
 578         unsigned pg = *cur_pg;
 579         struct request_queue *q =
 580                         osd_request_queue(_ios_od(ios, per_dev->dev));
 581         unsigned len = cur_len;
 582         int ret;
 583
 584         if (per_dev->bio == NULL) {
 585                 unsigned pages_in_stripe = ios->layout->group_width *
 586                                         (ios->layout->stripe_unit / PAGE_SIZE);
 587                 unsigned nr_pages = ios->nr_pages * ios->layout->group_width /
 588                                         (ios->layout->group_width -
 589                                          ios->layout->parity);
 590                 unsigned bio_size = (nr_pages + pages_in_stripe) /
 591                                         ios->layout->group_width;
 592
 593                 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
 594                 if (unlikely(!per_dev->bio)) {
 595                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
 596                                      bio_size);
 597                         ret = -ENOMEM;
 598                         goto out;
 599                 }
 600         }
 601
 602         while (cur_len > 0) {
 603                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
 604                 unsigned added_len;
 605
 606                 cur_len -= pglen;
 607
 608                 added_len = bio_add_pc_page(q, per_dev->bio, pages[pg],
 609                                             pglen, pgbase);
 610                 if (unlikely(pglen != added_len)) {
 611                         ORE_DBGMSG("Failed bio_add_pc_page bi_vcnt=%u\n",
 612                                    per_dev->bio->bi_vcnt);
 613                         ret = -ENOMEM;
 614                         goto out;
 615                 }
 616                 _add_stripe_page(ios->sp2d, &ios->si, pages[pg]);
 617
 618                 pgbase = 0;
 619                 ++pg;
 620         }
 621         BUG_ON(cur_len);
 622
 623         per_dev->length += len;
 624         *cur_pg = pg;
 625         ret = 0;
 626 out:    /* we fail the complete unit on an error eg don't advance
 627          * per_dev->length and cur_pg. This means that we might have a bigger
 628          * bio than the CDB requested length (per_dev->length). That's fine
 629          * only the oposite is fatal.
 630          */
 631         return ret;
 632 }
 633
 634 static int _prepare_for_striping(struct ore_io_state *ios)
 635 {
 636         struct ore_striping_info *si = &ios->si;
 637         unsigned stripe_unit = ios->layout->stripe_unit;
 638         unsigned mirrors_p1 = ios->layout->mirrors_p1;
 639         unsigned group_width = ios->layout->group_width;
 640         unsigned devs_in_group = group_width * mirrors_p1;
 641         unsigned dev = si->dev;
 642         unsigned first_dev = dev - (dev % devs_in_group);
 643         unsigned dev_order;
 644         unsigned cur_pg = ios->pages_consumed;
 645         u64 length = ios->length;
 646         int ret = 0;
 647
 648         if (!ios->pages) {
 649                 ios->numdevs = ios->layout->mirrors_p1;
 650                 return 0;
 651         }
 652
 653         BUG_ON(length > si->length);
 654
 655         dev_order = _dev_order(devs_in_group, mirrors_p1, si->par_dev, dev);
 656         si->cur_comp = dev_order;
 657         si->cur_pg = si->unit_off / PAGE_SIZE;
 658
 659         while (length) {
 660                 unsigned comp = dev - first_dev;
 661                 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
 662                 unsigned cur_len, page_off = 0;
 663
 664                 if (!per_dev->length) {
 665                         per_dev->dev = dev;
 666                         if (dev == si->dev) {
 667                                 WARN_ON(dev == si->par_dev);
 668                                 per_dev->offset = si->obj_offset;
 669                                 cur_len = stripe_unit - si->unit_off;
 670                                 page_off = si->unit_off & ~PAGE_MASK;
 671                                 BUG_ON(page_off && (page_off != ios->pgbase));
 672                         } else {
 673                                 if (si->cur_comp > dev_order)
 674                                         per_dev->offset =
 675                                                 si->obj_offset - si->unit_off;
 676                                 else /* si->cur_comp < dev_order */
 677                                         per_dev->offset =
 678                                                 si->obj_offset + stripe_unit -
 679                                                                    si->unit_off;
 680                                 cur_len = stripe_unit;
 681                         }
 682                 } else {
 683                         cur_len = stripe_unit;
 684                 }
 685                 if (cur_len >= length)
 686                         cur_len = length;
 687
 688                 ret = _ore_add_stripe_unit(ios, &cur_pg, page_off, ios->pages,
 689                                            per_dev, cur_len);
 690                 if (unlikely(ret))
 691                         goto out;
 692
 693                 dev += mirrors_p1;
 694                 dev = (dev % devs_in_group) + first_dev;
 695
 696                 length -= cur_len;
 697
 698                 si->cur_comp = (si->cur_comp + 1) % group_width;
 699                 if (unlikely((dev == si->par_dev) || (!length && ios->sp2d))) {
 700                         if (!length && ios->sp2d) {
 701                                 /* If we are writing and this is the very last
 702                                  * stripe. then operate on parity dev.
 703                                  */
 704                                 dev = si->par_dev;
 705                         }
 706                         if (ios->sp2d)
 707                                 /* In writes cur_len just means if it's the
 708                                  * last one. See _ore_add_parity_unit.
 709                                  */
 710                                 cur_len = length;
 711                         per_dev = &ios->per_dev[dev - first_dev];
 712                         if (!per_dev->length) {
 713                                 /* Only/always the parity unit of the first
 714                                  * stripe will be empty. So this is a chance to
 715                                  * initialize the per_dev info.
 716                                  */
 717                                 per_dev->dev = dev;
 718                                 per_dev->offset = si->obj_offset - si->unit_off;
 719                         }
 720
 721                         ret = _ore_add_parity_unit(ios, si, per_dev, cur_len);
 722                         if (unlikely(ret))
 723                                         goto out;
 724
 725                         /* Rotate next par_dev backwards with wraping */
 726                         si->par_dev = (devs_in_group + si->par_dev -
 727                                        ios->layout->parity * mirrors_p1) %
 728                                       devs_in_group + first_dev;
 729                         /* Next stripe, start fresh */
 730                         si->cur_comp = 0;
 731                         si->cur_pg = 0;
 732                 }
 733         }
 734 out:
 735         ios->numdevs = devs_in_group;
 736         ios->pages_consumed = cur_pg;
 737         if (unlikely(ret)) {
 738                 if (length == ios->length)
 739                         return ret;
 740                 else
 741                         ios->length -= length;
 742         }
 743         return 0;
 744 }
 745
 746 int ore_create(struct ore_io_state *ios)
 747 {
 748         int i, ret;
 749
 750         for (i = 0; i < ios->oc->numdevs; i++) {
 751                 struct osd_request *or;
 752
 753                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 754                 if (unlikely(!or)) {
 755                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 756                         ret = -ENOMEM;
 757                         goto out;
 758                 }
 759                 ios->per_dev[i].or = or;
 760                 ios->numdevs++;
 761
 762                 osd_req_create_object(or, _ios_obj(ios, i));
 763         }
 764         ret = ore_io_execute(ios);
 765
 766 out:
 767         return ret;
 768 }
 769 EXPORT_SYMBOL(ore_create);
 770
 771 int ore_remove(struct ore_io_state *ios)
 772 {
 773         int i, ret;
 774
 775         for (i = 0; i < ios->oc->numdevs; i++) {
 776                 struct osd_request *or;
 777
 778                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 779                 if (unlikely(!or)) {
 780                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 781                         ret = -ENOMEM;
 782                         goto out;
 783                 }
 784                 ios->per_dev[i].or = or;
 785                 ios->numdevs++;
 786
 787                 osd_req_remove_object(or, _ios_obj(ios, i));
 788         }
 789         ret = ore_io_execute(ios);
 790
 791 out:
 792         return ret;
 793 }
 794 EXPORT_SYMBOL(ore_remove);
 795
 796 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
 797 {
 798         struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
 799         unsigned dev = ios->per_dev[cur_comp].dev;
 800         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
 801         int ret = 0;
 802
 803         if (ios->pages && !master_dev->length)
 804                 return 0; /* Just an empty slot */
 805
 806         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
 807                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 808                 struct osd_request *or;
 809
 810                 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
 811                 if (unlikely(!or)) {
 812                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 813                         ret = -ENOMEM;
 814                         goto out;
 815                 }
 816                 per_dev->or = or;
 817
 818                 if (ios->pages) {
 819                         struct bio *bio;
 820
 821                         if (per_dev != master_dev) {
 822                                 bio = bio_kmalloc(GFP_KERNEL,
 823                                                   master_dev->bio->bi_max_vecs);
 824                                 if (unlikely(!bio)) {
 825                                         ORE_DBGMSG(
 826                                               "Failed to allocate BIO size=%u\n",
 827                                               master_dev->bio->bi_max_vecs);
 828                                         ret = -ENOMEM;
 829                                         goto out;
 830                                 }
 831
 832                                 __bio_clone(bio, master_dev->bio);
 833                                 bio->bi_bdev = NULL;
 834                                 bio->bi_next = NULL;
 835                                 per_dev->offset = master_dev->offset;
 836                                 per_dev->length = master_dev->length;
 837                                 per_dev->bio =  bio;
 838                                 per_dev->dev = dev;
 839                         } else {
 840                                 bio = master_dev->bio;
 841                                 /* FIXME: bio_set_dir() */
 842                                 bio->bi_rw |= REQ_WRITE;
 843                         }
 844
 845                         osd_req_write(or, _ios_obj(ios, dev), per_dev->offset,
 846                                       bio, per_dev->length);
 847                         ORE_DBGMSG("write(0x%llx) offset=0x%llx "
 848                                       "length=0x%llx dev=%d\n",
 849                                      _LLU(_ios_obj(ios, dev)->id),
 850                                      _LLU(per_dev->offset),
 851                                      _LLU(per_dev->length), dev);
 852                 } else if (ios->kern_buff) {
 853                         per_dev->offset = ios->si.obj_offset;
 854                         per_dev->dev = ios->si.dev + dev;
 855
 856                         /* no cross device without page array */
 857                         BUG_ON((ios->layout->group_width > 1) &&
 858                                (ios->si.unit_off + ios->length >
 859                                 ios->layout->stripe_unit));
 860
 861                         ret = osd_req_write_kern(or, _ios_obj(ios, per_dev->dev),
 862                                                  per_dev->offset,
 863                                                  ios->kern_buff, ios->length);
 864                         if (unlikely(ret))
 865                                 goto out;
 866                         ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
 867                                       "length=0x%llx dev=%d\n",
 868                                      _LLU(_ios_obj(ios, dev)->id),
 869                                      _LLU(per_dev->offset),
 870                                      _LLU(ios->length), per_dev->dev);
 871                 } else {
 872                         osd_req_set_attributes(or, _ios_obj(ios, dev));
 873                         ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
 874                                      _LLU(_ios_obj(ios, dev)->id),
 875                                      ios->out_attr_len, dev);
 876                 }
 877
 878                 if (ios->out_attr)
 879                         osd_req_add_set_attr_list(or, ios->out_attr,
 880                                                   ios->out_attr_len);
 881
 882                 if (ios->in_attr)
 883                         osd_req_add_get_attr_list(or, ios->in_attr,
 884                                                   ios->in_attr_len);
 885         }
 886
 887 out:
 888         return ret;
 889 }
 890
 891 int ore_write(struct ore_io_state *ios)
 892 {
 893         int i;
 894         int ret;
 895
 896         if (unlikely(ios->sp2d && !ios->r4w)) {
 897                 /* A library is attempting a RAID-write without providing
 898                  * a pages lock interface.
 899                  */
 900                 WARN_ON_ONCE(1);
 901                 return -ENOTSUPP;
 902         }
 903
 904         ret = _prepare_for_striping(ios);
 905         if (unlikely(ret))
 906                 return ret;
 907
 908         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 909                 ret = _write_mirror(ios, i);
 910                 if (unlikely(ret))
 911                         return ret;
 912         }
 913
 914         ret = ore_io_execute(ios);
 915         return ret;
 916 }
 917 EXPORT_SYMBOL(ore_write);
 918
 919 int _ore_read_mirror(struct ore_io_state *ios, unsigned cur_comp)
 920 {
 921         struct osd_request *or;
 922         struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 923         struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
 924         unsigned first_dev = (unsigned)obj->id;
 925
 926         if (ios->pages && !per_dev->length)
 927                 return 0; /* Just an empty slot */
 928
 929         first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
 930         or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
 931         if (unlikely(!or)) {
 932                 ORE_ERR("%s: osd_start_request failed\n", __func__);
 933                 return -ENOMEM;
 934         }
 935         per_dev->or = or;
 936
 937         if (ios->pages) {
 938                 if (per_dev->cur_sg) {
 939                         /* finalize the last sg_entry */
 940                         _ore_add_sg_seg(per_dev, 0, false);
 941                         if (unlikely(!per_dev->cur_sg))
 942                                 return 0; /* Skip parity only device */
 943
 944                         osd_req_read_sg(or, obj, per_dev->bio,
 945                                         per_dev->sglist, per_dev->cur_sg);
 946                 } else {
 947                         /* The no raid case */
 948                         osd_req_read(or, obj, per_dev->offset,
 949                                      per_dev->bio, per_dev->length);
 950                 }
 951
 952                 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
 953                              " dev=%d sg_len=%d\n", _LLU(obj->id),
 954                              _LLU(per_dev->offset), _LLU(per_dev->length),
 955                              first_dev, per_dev->cur_sg);
 956         } else {
 957                 BUG_ON(ios->kern_buff);
 958
 959                 osd_req_get_attributes(or, obj);
 960                 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
 961                               _LLU(obj->id),
 962                               ios->in_attr_len, first_dev);
 963         }
 964         if (ios->out_attr)
 965                 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
 966
 967         if (ios->in_attr)
 968                 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
 969
 970         return 0;
 971 }
 972
 973 int ore_read(struct ore_io_state *ios)
 974 {
 975         int i;
 976         int ret;
 977
 978         ret = _prepare_for_striping(ios);
 979         if (unlikely(ret))
 980                 return ret;
 981
 982         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 983                 ret = _ore_read_mirror(ios, i);
 984                 if (unlikely(ret))
 985                         return ret;
 986         }
 987
 988         ret = ore_io_execute(ios);
 989         return ret;
 990 }
 991 EXPORT_SYMBOL(ore_read);
 992
 993 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
 994 {
 995         struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
 996         void *iter = NULL;
 997         int nelem;
 998
 999         do {
1000                 nelem = 1;
1001                 osd_req_decode_get_attr_list(ios->per_dev[0].or,
1002                                              &cur_attr, &nelem, &iter);
1003                 if ((cur_attr.attr_page == attr->attr_page) &&
1004                     (cur_attr.attr_id == attr->attr_id)) {
1005                         attr->len = cur_attr.len;
1006                         attr->val_ptr = cur_attr.val_ptr;
1007                         return 0;
1008                 }
1009         } while (iter);
1010
1011         return -EIO;
1012 }
1013 EXPORT_SYMBOL(extract_attr_from_ios);
1014
1015 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
1016                              struct osd_attr *attr)
1017 {
1018         int last_comp = cur_comp + ios->layout->mirrors_p1;
1019
1020         for (; cur_comp < last_comp; ++cur_comp) {
1021                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
1022                 struct osd_request *or;
1023
1024                 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
1025                 if (unlikely(!or)) {
1026                         ORE_ERR("%s: osd_start_request failed\n", __func__);
1027                         return -ENOMEM;
1028                 }
1029                 per_dev->or = or;
1030
1031                 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
1032                 osd_req_add_set_attr_list(or, attr, 1);
1033         }
1034
1035         return 0;
1036 }
1037
1038 struct _trunc_info {
1039         struct ore_striping_info si;
1040         u64 prev_group_obj_off;
1041         u64 next_group_obj_off;
1042
1043         unsigned first_group_dev;
1044         unsigned nex_group_dev;
1045 };
1046
1047 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
1048                              struct _trunc_info *ti)
1049 {
1050         unsigned stripe_unit = layout->stripe_unit;
1051
1052         ore_calc_stripe_info(layout, file_offset, 0, &ti->si);
1053
1054         ti->prev_group_obj_off = ti->si.M * stripe_unit;
1055         ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
1056
1057         ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
1058         ti->nex_group_dev = ti->first_group_dev + layout->group_width;
1059 }
1060
1061 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
1062                    u64 size)
1063 {
1064         struct ore_io_state *ios;
1065         struct exofs_trunc_attr {
1066                 struct osd_attr attr;
1067                 __be64 newsize;
1068         } *size_attrs;
1069         struct _trunc_info ti;
1070         int i, ret;
1071
1072         ret = ore_get_io_state(layout, oc, &ios);
1073         if (unlikely(ret))
1074                 return ret;
1075
1076         _calc_trunk_info(ios->layout, size, &ti);
1077
1078         size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
1079                              GFP_KERNEL);
1080         if (unlikely(!size_attrs)) {
1081                 ret = -ENOMEM;
1082                 goto out;
1083         }
1084
1085         ios->numdevs = ios->oc->numdevs;
1086
1087         for (i = 0; i < ios->numdevs; ++i) {
1088                 struct exofs_trunc_attr *size_attr = &size_attrs[i];
1089                 u64 obj_size;
1090
1091                 if (i < ti.first_group_dev)
1092                         obj_size = ti.prev_group_obj_off;
1093                 else if (i >= ti.nex_group_dev)
1094                         obj_size = ti.next_group_obj_off;
1095                 else if (i < ti.si.dev) /* dev within this group */
1096                         obj_size = ti.si.obj_offset +
1097                                       ios->layout->stripe_unit - ti.si.unit_off;
1098                 else if (i == ti.si.dev)
1099                         obj_size = ti.si.obj_offset;
1100                 else /* i > ti.dev */
1101                         obj_size = ti.si.obj_offset - ti.si.unit_off;
1102
1103                 size_attr->newsize = cpu_to_be64(obj_size);
1104                 size_attr->attr = g_attr_logical_length;
1105                 size_attr->attr.val_ptr = &size_attr->newsize;
1106
1107                 ORE_DBGMSG("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
1108                              _LLU(oc->comps->obj.id), _LLU(obj_size), i);
1109                 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
1110                                         &size_attr->attr);
1111                 if (unlikely(ret))
1112                         goto out;
1113         }
1114         ret = ore_io_execute(ios);
1115
1116 out:
1117         kfree(size_attrs);
1118         ore_put_io_state(ios);
1119         return ret;
1120 }
1121 EXPORT_SYMBOL(ore_truncate);
1122
1123 const struct osd_attr g_attr_logical_length = ATTR_DEF(
1124         OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
1125 EXPORT_SYMBOL(g_attr_logical_length);