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;
 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->length + PAGE_SIZE - 1) / PAGE_SIZE;
 290                 if (layout->parity)
 291                         _ore_post_alloc_raid_stuff(ios);
 292         }
 293
 294         return 0;
 295 }
 296 EXPORT_SYMBOL(ore_get_rw_state);
 297
 298 /* Allocate an io_state for all the devices in the comps array
 299  *
 300  * This version of io_state allocation is used mostly by create/remove
 301  * and trunc where we currently need all the devices. The only wastful
 302  * bit is the read/write_attributes with no IO. Those sites should
 303  * be converted to use ore_get_rw_state() with length=0
 304  */
 305 int  ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
 306                       struct ore_io_state **pios)
 307 {
 308         return _ore_get_io_state(layout, oc, oc->numdevs, 0, 0, pios);
 309 }
 310 EXPORT_SYMBOL(ore_get_io_state);
 311
 312 void ore_put_io_state(struct ore_io_state *ios)
 313 {
 314         if (ios) {
 315                 unsigned i;
 316
 317                 for (i = 0; i < ios->numdevs; i++) {
 318                         struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 319
 320                         if (per_dev->or)
 321                                 osd_end_request(per_dev->or);
 322                         if (per_dev->bio)
 323                                 bio_put(per_dev->bio);
 324                 }
 325
 326                 _ore_free_raid_stuff(ios);
 327                 kfree(ios);
 328         }
 329 }
 330 EXPORT_SYMBOL(ore_put_io_state);
 331
 332 static void _sync_done(struct ore_io_state *ios, void *p)
 333 {
 334         struct completion *waiting = p;
 335
 336         complete(waiting);
 337 }
 338
 339 static void _last_io(struct kref *kref)
 340 {
 341         struct ore_io_state *ios = container_of(
 342                                         kref, struct ore_io_state, kref);
 343
 344         ios->done(ios, ios->private);
 345 }
 346
 347 static void _done_io(struct osd_request *or, void *p)
 348 {
 349         struct ore_io_state *ios = p;
 350
 351         kref_put(&ios->kref, _last_io);
 352 }
 353
 354 int ore_io_execute(struct ore_io_state *ios)
 355 {
 356         DECLARE_COMPLETION_ONSTACK(wait);
 357         bool sync = (ios->done == NULL);
 358         int i, ret;
 359
 360         if (sync) {
 361                 ios->done = _sync_done;
 362                 ios->private = &wait;
 363         }
 364
 365         for (i = 0; i < ios->numdevs; i++) {
 366                 struct osd_request *or = ios->per_dev[i].or;
 367                 if (unlikely(!or))
 368                         continue;
 369
 370                 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
 371                 if (unlikely(ret)) {
 372                         ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
 373                                      ret);
 374                         return ret;
 375                 }
 376         }
 377
 378         kref_init(&ios->kref);
 379
 380         for (i = 0; i < ios->numdevs; i++) {
 381                 struct osd_request *or = ios->per_dev[i].or;
 382                 if (unlikely(!or))
 383                         continue;
 384
 385                 kref_get(&ios->kref);
 386                 osd_execute_request_async(or, _done_io, ios);
 387         }
 388
 389         kref_put(&ios->kref, _last_io);
 390         ret = 0;
 391
 392         if (sync) {
 393                 wait_for_completion(&wait);
 394                 ret = ore_check_io(ios, NULL);
 395         }
 396         return ret;
 397 }
 398
 399 static void _clear_bio(struct bio *bio)
 400 {
 401         struct bio_vec *bv;
 402         unsigned i;
 403
 404         bio_for_each_segment_all(bv, bio, i) {
 405                 unsigned this_count = bv->bv_len;
 406
 407                 if (likely(PAGE_SIZE == this_count))
 408                         clear_highpage(bv->bv_page);
 409                 else
 410                         zero_user(bv->bv_page, bv->bv_offset, this_count);
 411         }
 412 }
 413
 414 int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
 415 {
 416         enum osd_err_priority acumulated_osd_err = 0;
 417         int acumulated_lin_err = 0;
 418         int i;
 419
 420         for (i = 0; i < ios->numdevs; i++) {
 421                 struct osd_sense_info osi;
 422                 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
 423                 struct osd_request *or = per_dev->or;
 424                 int ret;
 425
 426                 if (unlikely(!or))
 427                         continue;
 428
 429                 ret = osd_req_decode_sense(or, &osi);
 430                 if (likely(!ret))
 431                         continue;
 432
 433                 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
 434                         /* start read offset passed endof file */
 435                         _clear_bio(per_dev->bio);
 436                         ORE_DBGMSG("start read offset passed end of file "
 437                                 "offset=0x%llx, length=0x%llx\n",
 438                                 _LLU(per_dev->offset),
 439                                 _LLU(per_dev->length));
 440
 441                         continue; /* we recovered */
 442                 }
 443
 444                 if (on_dev_error) {
 445                         u64 residual = ios->reading ?
 446                                         or->in.residual : or->out.residual;
 447                         u64 offset = (ios->offset + ios->length) - residual;
 448                         unsigned dev = per_dev->dev - ios->oc->first_dev;
 449                         struct ore_dev *od = ios->oc->ods[dev];
 450
 451                         on_dev_error(ios, od, dev, osi.osd_err_pri,
 452                                      offset, residual);
 453                 }
 454                 if (osi.osd_err_pri >= acumulated_osd_err) {
 455                         acumulated_osd_err = osi.osd_err_pri;
 456                         acumulated_lin_err = ret;
 457                 }
 458         }
 459
 460         return acumulated_lin_err;
 461 }
 462 EXPORT_SYMBOL(ore_check_io);
 463
 464 /*
 465  * L - logical offset into the file
 466  *
 467  * D - number of Data devices
 468  *      D = group_width - parity
 469  *
 470  * U - The number of bytes in a stripe within a group
 471  *      U =  stripe_unit * D
 472  *
 473  * T - The number of bytes striped within a group of component objects
 474  *     (before advancing to the next group)
 475  *      T = U * group_depth
 476  *
 477  * S - The number of bytes striped across all component objects
 478  *     before the pattern repeats
 479  *      S = T * group_count
 480  *
 481  * M - The "major" (i.e., across all components) cycle number
 482  *      M = L / S
 483  *
 484  * G - Counts the groups from the beginning of the major cycle
 485  *      G = (L - (M * S)) / T   [or (L % S) / T]
 486  *
 487  * H - The byte offset within the group
 488  *      H = (L - (M * S)) % T   [or (L % S) % T]
 489  *
 490  * N - The "minor" (i.e., across the group) stripe number
 491  *      N = H / U
 492  *
 493  * C - The component index coresponding to L
 494  *
 495  *      C = (H - (N * U)) / stripe_unit + G * D
 496  *      [or (L % U) / stripe_unit + G * D]
 497  *
 498  * O - The component offset coresponding to L
 499  *      O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
 500  *
 501  * LCMdP – Parity cycle: Lowest Common Multiple of group_width, parity
 502  *          divide by parity
 503  *      LCMdP = lcm(group_width, parity) / parity
 504  *
 505  * R - The parity Rotation stripe
 506  *     (Note parity cycle always starts at a group's boundary)
 507  *      R = N % LCMdP
 508  *
 509  * I = the first parity device index
 510  *      I = (group_width + group_width - R*parity - parity) % group_width
 511  *
 512  * Craid - The component index Rotated
 513  *      Craid = (group_width + C - R*parity) % group_width
 514  *      (We add the group_width to avoid negative numbers modulo math)
 515  */
 516 void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
 517                           u64 length, struct ore_striping_info *si)
 518 {
 519         u32     stripe_unit = layout->stripe_unit;
 520         u32     group_width = layout->group_width;
 521         u64     group_depth = layout->group_depth;
 522         u32     parity      = layout->parity;
 523
 524         u32     D = group_width - parity;
 525         u32     U = D * stripe_unit;
 526         u64     T = U * group_depth;
 527         u64     S = T * layout->group_count;
 528         u64     M = div64_u64(file_offset, S);
 529
 530         /*
 531         G = (L - (M * S)) / T
 532         H = (L - (M * S)) % T
 533         */
 534         u64     LmodS = file_offset - M * S;
 535         u32     G = div64_u64(LmodS, T);
 536         u64     H = LmodS - G * T;
 537
 538         u32     N = div_u64(H, U);
 539
 540         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
 541         u32     C = (u32)(H - (N * U)) / stripe_unit + G * group_width;
 542
 543         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
 544
 545         si->obj_offset = si->unit_off + (N * stripe_unit) +
 546                                   (M * group_depth * stripe_unit);
 547
 548         if (parity) {
 549                 u32 LCMdP = lcm(group_width, parity) / parity;
 550                 /* R     = N % LCMdP; */
 551                 u32 RxP   = (N % LCMdP) * parity;
 552                 u32 first_dev = C - C % group_width;
 553
 554                 si->par_dev = (group_width + group_width - parity - RxP) %
 555                               group_width + first_dev;
 556                 si->dev = (group_width + C - RxP) % group_width + first_dev;
 557                 si->bytes_in_stripe = U;
 558                 si->first_stripe_start = M * S + G * T + N * U;
 559         } else {
 560                 /* Make the math correct see _prepare_one_group */
 561                 si->par_dev = group_width;
 562                 si->dev = C;
 563         }
 564
 565         si->dev *= layout->mirrors_p1;
 566         si->par_dev *= layout->mirrors_p1;
 567         si->offset = file_offset;
 568         si->length = T - H;
 569         if (si->length > length)
 570                 si->length = length;
 571         si->M = M;
 572 }
 573 EXPORT_SYMBOL(ore_calc_stripe_info);
 574
 575 int _ore_add_stripe_unit(struct ore_io_state *ios,  unsigned *cur_pg,
 576                          unsigned pgbase, struct page **pages,
 577                          struct ore_per_dev_state *per_dev, int cur_len)
 578 {
 579         unsigned pg = *cur_pg;
 580         struct request_queue *q =
 581                         osd_request_queue(_ios_od(ios, per_dev->dev));
 582         unsigned len = cur_len;
 583         int ret;
 584
 585         if (per_dev->bio == NULL) {
 586                 unsigned pages_in_stripe = ios->layout->group_width *
 587                                         (ios->layout->stripe_unit / PAGE_SIZE);
 588                 unsigned nr_pages = ios->nr_pages * ios->layout->group_width /
 589                                         (ios->layout->group_width -
 590                                          ios->layout->parity);
 591                 unsigned bio_size = (nr_pages + pages_in_stripe) /
 592                                         ios->layout->group_width;
 593
 594                 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
 595                 if (unlikely(!per_dev->bio)) {
 596                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
 597                                      bio_size);
 598                         ret = -ENOMEM;
 599                         goto out;
 600                 }
 601         }
 602
 603         while (cur_len > 0) {
 604                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
 605                 unsigned added_len;
 606
 607                 cur_len -= pglen;
 608
 609                 added_len = bio_add_pc_page(q, per_dev->bio, pages[pg],
 610                                             pglen, pgbase);
 611                 if (unlikely(pglen != added_len)) {
 612                         ORE_DBGMSG("Failed bio_add_pc_page bi_vcnt=%u\n",
 613                                    per_dev->bio->bi_vcnt);
 614                         ret = -ENOMEM;
 615                         goto out;
 616                 }
 617                 _add_stripe_page(ios->sp2d, &ios->si, pages[pg]);
 618
 619                 pgbase = 0;
 620                 ++pg;
 621         }
 622         BUG_ON(cur_len);
 623
 624         per_dev->length += len;
 625         *cur_pg = pg;
 626         ret = 0;
 627 out:    /* we fail the complete unit on an error eg don't advance
 628          * per_dev->length and cur_pg. This means that we might have a bigger
 629          * bio than the CDB requested length (per_dev->length). That's fine
 630          * only the oposite is fatal.
 631          */
 632         return ret;
 633 }
 634
 635 static int _prepare_for_striping(struct ore_io_state *ios)
 636 {
 637         struct ore_striping_info *si = &ios->si;
 638         unsigned stripe_unit = ios->layout->stripe_unit;
 639         unsigned mirrors_p1 = ios->layout->mirrors_p1;
 640         unsigned group_width = ios->layout->group_width;
 641         unsigned devs_in_group = group_width * mirrors_p1;
 642         unsigned dev = si->dev;
 643         unsigned first_dev = dev - (dev % devs_in_group);
 644         unsigned dev_order;
 645         unsigned cur_pg = ios->pages_consumed;
 646         u64 length = ios->length;
 647         int ret = 0;
 648
 649         if (!ios->pages) {
 650                 ios->numdevs = ios->layout->mirrors_p1;
 651                 return 0;
 652         }
 653
 654         BUG_ON(length > si->length);
 655
 656         dev_order = _dev_order(devs_in_group, mirrors_p1, si->par_dev, dev);
 657         si->cur_comp = dev_order;
 658         si->cur_pg = si->unit_off / PAGE_SIZE;
 659
 660         while (length) {
 661                 unsigned comp = dev - first_dev;
 662                 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
 663                 unsigned cur_len, page_off = 0;
 664
 665                 if (!per_dev->length) {
 666                         per_dev->dev = dev;
 667                         if (dev == si->dev) {
 668                                 WARN_ON(dev == si->par_dev);
 669                                 per_dev->offset = si->obj_offset;
 670                                 cur_len = stripe_unit - si->unit_off;
 671                                 page_off = si->unit_off & ~PAGE_MASK;
 672                                 BUG_ON(page_off && (page_off != ios->pgbase));
 673                         } else {
 674                                 if (si->cur_comp > dev_order)
 675                                         per_dev->offset =
 676                                                 si->obj_offset - si->unit_off;
 677                                 else /* si->cur_comp < dev_order */
 678                                         per_dev->offset =
 679                                                 si->obj_offset + stripe_unit -
 680                                                                    si->unit_off;
 681                                 cur_len = stripe_unit;
 682                         }
 683                 } else {
 684                         cur_len = stripe_unit;
 685                 }
 686                 if (cur_len >= length)
 687                         cur_len = length;
 688
 689                 ret = _ore_add_stripe_unit(ios, &cur_pg, page_off, ios->pages,
 690                                            per_dev, cur_len);
 691                 if (unlikely(ret))
 692                         goto out;
 693
 694                 dev += mirrors_p1;
 695                 dev = (dev % devs_in_group) + first_dev;
 696
 697                 length -= cur_len;
 698
 699                 si->cur_comp = (si->cur_comp + 1) % group_width;
 700                 if (unlikely((dev == si->par_dev) || (!length && ios->sp2d))) {
 701                         if (!length && ios->sp2d) {
 702                                 /* If we are writing and this is the very last
 703                                  * stripe. then operate on parity dev.
 704                                  */
 705                                 dev = si->par_dev;
 706                         }
 707                         if (ios->sp2d)
 708                                 /* In writes cur_len just means if it's the
 709                                  * last one. See _ore_add_parity_unit.
 710                                  */
 711                                 cur_len = length;
 712                         per_dev = &ios->per_dev[dev - first_dev];
 713                         if (!per_dev->length) {
 714                                 /* Only/always the parity unit of the first
 715                                  * stripe will be empty. So this is a chance to
 716                                  * initialize the per_dev info.
 717                                  */
 718                                 per_dev->dev = dev;
 719                                 per_dev->offset = si->obj_offset - si->unit_off;
 720                         }
 721
 722                         ret = _ore_add_parity_unit(ios, si, per_dev, cur_len);
 723                         if (unlikely(ret))
 724                                         goto out;
 725
 726                         /* Rotate next par_dev backwards with wraping */
 727                         si->par_dev = (devs_in_group + si->par_dev -
 728                                        ios->layout->parity * mirrors_p1) %
 729                                       devs_in_group + first_dev;
 730                         /* Next stripe, start fresh */
 731                         si->cur_comp = 0;
 732                         si->cur_pg = 0;
 733                 }
 734         }
 735 out:
 736         ios->numdevs = devs_in_group;
 737         ios->pages_consumed = cur_pg;
 738         return ret;
 739 }
 740
 741 int ore_create(struct ore_io_state *ios)
 742 {
 743         int i, ret;
 744
 745         for (i = 0; i < ios->oc->numdevs; i++) {
 746                 struct osd_request *or;
 747
 748                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 749                 if (unlikely(!or)) {
 750                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 751                         ret = -ENOMEM;
 752                         goto out;
 753                 }
 754                 ios->per_dev[i].or = or;
 755                 ios->numdevs++;
 756
 757                 osd_req_create_object(or, _ios_obj(ios, i));
 758         }
 759         ret = ore_io_execute(ios);
 760
 761 out:
 762         return ret;
 763 }
 764 EXPORT_SYMBOL(ore_create);
 765
 766 int ore_remove(struct ore_io_state *ios)
 767 {
 768         int i, ret;
 769
 770         for (i = 0; i < ios->oc->numdevs; i++) {
 771                 struct osd_request *or;
 772
 773                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
 774                 if (unlikely(!or)) {
 775                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 776                         ret = -ENOMEM;
 777                         goto out;
 778                 }
 779                 ios->per_dev[i].or = or;
 780                 ios->numdevs++;
 781
 782                 osd_req_remove_object(or, _ios_obj(ios, i));
 783         }
 784         ret = ore_io_execute(ios);
 785
 786 out:
 787         return ret;
 788 }
 789 EXPORT_SYMBOL(ore_remove);
 790
 791 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
 792 {
 793         struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
 794         unsigned dev = ios->per_dev[cur_comp].dev;
 795         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
 796         int ret = 0;
 797
 798         if (ios->pages && !master_dev->length)
 799                 return 0; /* Just an empty slot */
 800
 801         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
 802                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 803                 struct osd_request *or;
 804
 805                 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
 806                 if (unlikely(!or)) {
 807                         ORE_ERR("%s: osd_start_request failed\n", __func__);
 808                         ret = -ENOMEM;
 809                         goto out;
 810                 }
 811                 per_dev->or = or;
 812
 813                 if (ios->pages) {
 814                         struct bio *bio;
 815
 816                         if (per_dev != master_dev) {
 817                                 bio = bio_clone_kmalloc(master_dev->bio,
 818                                                         GFP_KERNEL);
 819                                 if (unlikely(!bio)) {
 820                                         ORE_DBGMSG(
 821                                               "Failed to allocate BIO size=%u\n",
 822                                               master_dev->bio->bi_max_vecs);
 823                                         ret = -ENOMEM;
 824                                         goto out;
 825                                 }
 826
 827                                 bio->bi_bdev = NULL;
 828                                 bio->bi_next = NULL;
 829                                 per_dev->offset = master_dev->offset;
 830                                 per_dev->length = master_dev->length;
 831                                 per_dev->bio =  bio;
 832                                 per_dev->dev = dev;
 833                         } else {
 834                                 bio = master_dev->bio;
 835                                 /* FIXME: bio_set_dir() */
 836                                 bio->bi_rw |= REQ_WRITE;
 837                         }
 838
 839                         osd_req_write(or, _ios_obj(ios, cur_comp),
 840                                       per_dev->offset, bio, per_dev->length);
 841                         ORE_DBGMSG("write(0x%llx) offset=0x%llx "
 842                                       "length=0x%llx dev=%d\n",
 843                                      _LLU(_ios_obj(ios, cur_comp)->id),
 844                                      _LLU(per_dev->offset),
 845                                      _LLU(per_dev->length), dev);
 846                 } else if (ios->kern_buff) {
 847                         per_dev->offset = ios->si.obj_offset;
 848                         per_dev->dev = ios->si.dev + dev;
 849
 850                         /* no cross device without page array */
 851                         BUG_ON((ios->layout->group_width > 1) &&
 852                                (ios->si.unit_off + ios->length >
 853                                 ios->layout->stripe_unit));
 854
 855                         ret = osd_req_write_kern(or, _ios_obj(ios, cur_comp),
 856                                                  per_dev->offset,
 857                                                  ios->kern_buff, ios->length);
 858                         if (unlikely(ret))
 859                                 goto out;
 860                         ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
 861                                       "length=0x%llx dev=%d\n",
 862                                      _LLU(_ios_obj(ios, cur_comp)->id),
 863                                      _LLU(per_dev->offset),
 864                                      _LLU(ios->length), per_dev->dev);
 865                 } else {
 866                         osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
 867                         ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
 868                                      _LLU(_ios_obj(ios, cur_comp)->id),
 869                                      ios->out_attr_len, dev);
 870                 }
 871
 872                 if (ios->out_attr)
 873                         osd_req_add_set_attr_list(or, ios->out_attr,
 874                                                   ios->out_attr_len);
 875
 876                 if (ios->in_attr)
 877                         osd_req_add_get_attr_list(or, ios->in_attr,
 878                                                   ios->in_attr_len);
 879         }
 880
 881 out:
 882         return ret;
 883 }
 884
 885 int ore_write(struct ore_io_state *ios)
 886 {
 887         int i;
 888         int ret;
 889
 890         if (unlikely(ios->sp2d && !ios->r4w)) {
 891                 /* A library is attempting a RAID-write without providing
 892                  * a pages lock interface.
 893                  */
 894                 WARN_ON_ONCE(1);
 895                 return -ENOTSUPP;
 896         }
 897
 898         ret = _prepare_for_striping(ios);
 899         if (unlikely(ret))
 900                 return ret;
 901
 902         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 903                 ret = _write_mirror(ios, i);
 904                 if (unlikely(ret))
 905                         return ret;
 906         }
 907
 908         ret = ore_io_execute(ios);
 909         return ret;
 910 }
 911 EXPORT_SYMBOL(ore_write);
 912
 913 int _ore_read_mirror(struct ore_io_state *ios, unsigned cur_comp)
 914 {
 915         struct osd_request *or;
 916         struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
 917         struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
 918         unsigned first_dev = (unsigned)obj->id;
 919
 920         if (ios->pages && !per_dev->length)
 921                 return 0; /* Just an empty slot */
 922
 923         first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
 924         or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
 925         if (unlikely(!or)) {
 926                 ORE_ERR("%s: osd_start_request failed\n", __func__);
 927                 return -ENOMEM;
 928         }
 929         per_dev->or = or;
 930
 931         if (ios->pages) {
 932                 if (per_dev->cur_sg) {
 933                         /* finalize the last sg_entry */
 934                         _ore_add_sg_seg(per_dev, 0, false);
 935                         if (unlikely(!per_dev->cur_sg))
 936                                 return 0; /* Skip parity only device */
 937
 938                         osd_req_read_sg(or, obj, per_dev->bio,
 939                                         per_dev->sglist, per_dev->cur_sg);
 940                 } else {
 941                         /* The no raid case */
 942                         osd_req_read(or, obj, per_dev->offset,
 943                                      per_dev->bio, per_dev->length);
 944                 }
 945
 946                 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
 947                              " dev=%d sg_len=%d\n", _LLU(obj->id),
 948                              _LLU(per_dev->offset), _LLU(per_dev->length),
 949                              first_dev, per_dev->cur_sg);
 950         } else {
 951                 BUG_ON(ios->kern_buff);
 952
 953                 osd_req_get_attributes(or, obj);
 954                 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
 955                               _LLU(obj->id),
 956                               ios->in_attr_len, first_dev);
 957         }
 958         if (ios->out_attr)
 959                 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
 960
 961         if (ios->in_attr)
 962                 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
 963
 964         return 0;
 965 }
 966
 967 int ore_read(struct ore_io_state *ios)
 968 {
 969         int i;
 970         int ret;
 971
 972         ret = _prepare_for_striping(ios);
 973         if (unlikely(ret))
 974                 return ret;
 975
 976         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
 977                 ret = _ore_read_mirror(ios, i);
 978                 if (unlikely(ret))
 979                         return ret;
 980         }
 981
 982         ret = ore_io_execute(ios);
 983         return ret;
 984 }
 985 EXPORT_SYMBOL(ore_read);
 986
 987 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
 988 {
 989         struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
 990         void *iter = NULL;
 991         int nelem;
 992
 993         do {
 994                 nelem = 1;
 995                 osd_req_decode_get_attr_list(ios->per_dev[0].or,
 996                                              &cur_attr, &nelem, &iter);
 997                 if ((cur_attr.attr_page == attr->attr_page) &&
 998                     (cur_attr.attr_id == attr->attr_id)) {
 999                         attr->len = cur_attr.len;
1000                         attr->val_ptr = cur_attr.val_ptr;
1001                         return 0;
1002                 }
1003         } while (iter);
1004
1005         return -EIO;
1006 }
1007 EXPORT_SYMBOL(extract_attr_from_ios);
1008
1009 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
1010                              struct osd_attr *attr)
1011 {
1012         int last_comp = cur_comp + ios->layout->mirrors_p1;
1013
1014         for (; cur_comp < last_comp; ++cur_comp) {
1015                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
1016                 struct osd_request *or;
1017
1018                 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
1019                 if (unlikely(!or)) {
1020                         ORE_ERR("%s: osd_start_request failed\n", __func__);
1021                         return -ENOMEM;
1022                 }
1023                 per_dev->or = or;
1024
1025                 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
1026                 osd_req_add_set_attr_list(or, attr, 1);
1027         }
1028
1029         return 0;
1030 }
1031
1032 struct _trunc_info {
1033         struct ore_striping_info si;
1034         u64 prev_group_obj_off;
1035         u64 next_group_obj_off;
1036
1037         unsigned first_group_dev;
1038         unsigned nex_group_dev;
1039 };
1040
1041 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
1042                              struct _trunc_info *ti)
1043 {
1044         unsigned stripe_unit = layout->stripe_unit;
1045
1046         ore_calc_stripe_info(layout, file_offset, 0, &ti->si);
1047
1048         ti->prev_group_obj_off = ti->si.M * stripe_unit;
1049         ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
1050
1051         ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
1052         ti->nex_group_dev = ti->first_group_dev + layout->group_width;
1053 }
1054
1055 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
1056                    u64 size)
1057 {
1058         struct ore_io_state *ios;
1059         struct exofs_trunc_attr {
1060                 struct osd_attr attr;
1061                 __be64 newsize;
1062         } *size_attrs;
1063         struct _trunc_info ti;
1064         int i, ret;
1065
1066         ret = ore_get_io_state(layout, oc, &ios);
1067         if (unlikely(ret))
1068                 return ret;
1069
1070         _calc_trunk_info(ios->layout, size, &ti);
1071
1072         size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
1073                              GFP_KERNEL);
1074         if (unlikely(!size_attrs)) {
1075                 ret = -ENOMEM;
1076                 goto out;
1077         }
1078
1079         ios->numdevs = ios->oc->numdevs;
1080
1081         for (i = 0; i < ios->numdevs; ++i) {
1082                 struct exofs_trunc_attr *size_attr = &size_attrs[i];
1083                 u64 obj_size;
1084
1085                 if (i < ti.first_group_dev)
1086                         obj_size = ti.prev_group_obj_off;
1087                 else if (i >= ti.nex_group_dev)
1088                         obj_size = ti.next_group_obj_off;
1089                 else if (i < ti.si.dev) /* dev within this group */
1090                         obj_size = ti.si.obj_offset +
1091                                       ios->layout->stripe_unit - ti.si.unit_off;
1092                 else if (i == ti.si.dev)
1093                         obj_size = ti.si.obj_offset;
1094                 else /* i > ti.dev */
1095                         obj_size = ti.si.obj_offset - ti.si.unit_off;
1096
1097                 size_attr->newsize = cpu_to_be64(obj_size);
1098                 size_attr->attr = g_attr_logical_length;
1099                 size_attr->attr.val_ptr = &size_attr->newsize;
1100
1101                 ORE_DBGMSG("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
1102                              _LLU(oc->comps->obj.id), _LLU(obj_size), i);
1103                 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
1104                                         &size_attr->attr);
1105                 if (unlikely(ret))
1106                         goto out;
1107         }
1108         ret = ore_io_execute(ios);
1109
1110 out:
1111         kfree(size_attrs);
1112         ore_put_io_state(ios);
1113         return ret;
1114 }
1115 EXPORT_SYMBOL(ore_truncate);
1116
1117 const struct osd_attr g_attr_logical_length = ATTR_DEF(
1118         OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
1119 EXPORT_SYMBOL(g_attr_logical_length);