fs/exofs/ore_raid.c

   1 /*
   2  * Copyright (C) 2011
   3  * Boaz Harrosh <bharrosh@panasas.com>
   4  *
   5  * This file is part of the objects raid engine (ore).
   6  *
   7  * It is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as published
   9  * by the Free Software Foundation.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * along with "ore". If not, write to the Free Software Foundation, Inc:
  13  *      "Free Software Foundation <info@fsf.org>"
  14  */
  15
  16 #include <linux/gfp.h>
  17 #include <linux/async_tx.h>
  18
  19 #include "ore_raid.h"
  20
  21 #undef ORE_DBGMSG2
  22 #define ORE_DBGMSG2 ORE_DBGMSG
  23
  24 struct page *_raid_page_alloc(void)
  25 {
  26         return alloc_page(GFP_KERNEL);
  27 }
  28
  29 void _raid_page_free(struct page *p)
  30 {
  31         __free_page(p);
  32 }
  33
  34 /* This struct is forward declare in ore_io_state, but is private to here.
  35  * It is put on ios->sp2d for RAID5/6 writes only. See _gen_xor_unit.
  36  *
  37  * __stripe_pages_2d is a 2d array of pages, and it is also a corner turn.
  38  * Ascending page index access is sp2d(p-minor, c-major). But storage is
  39  * sp2d[p-minor][c-major], so it can be properlly presented to the async-xor
  40  * API.
  41  */
  42 struct __stripe_pages_2d {
  43         /* Cache some hot path repeated calculations */
  44         unsigned parity;
  45         unsigned data_devs;
  46         unsigned pages_in_unit;
  47
  48         bool needed ;
  49
  50         /* Array size is pages_in_unit (layout->stripe_unit / PAGE_SIZE) */
  51         struct __1_page_stripe {
  52                 bool alloc;
  53                 unsigned write_count;
  54                 struct async_submit_ctl submit;
  55                 struct dma_async_tx_descriptor *tx;
  56
  57                 /* The size of this array is data_devs + parity */
  58                 struct page **pages;
  59                 struct page **scribble;
  60                 /* bool array, size of this array is data_devs */
  61                 char *page_is_read;
  62         } _1p_stripes[];
  63 };
  64
  65 /* This can get bigger then a page. So support multiple page allocations
  66  * _sp2d_free should be called even if _sp2d_alloc fails (by returning
  67  * none-zero).
  68  */
  69 static int _sp2d_alloc(unsigned pages_in_unit, unsigned group_width,
  70                        unsigned parity, struct __stripe_pages_2d **psp2d)
  71 {
  72         struct __stripe_pages_2d *sp2d;
  73         unsigned data_devs = group_width - parity;
  74         struct _alloc_all_bytes {
  75                 struct __alloc_stripe_pages_2d {
  76                         struct __stripe_pages_2d sp2d;
  77                         struct __1_page_stripe _1p_stripes[pages_in_unit];
  78                 } __asp2d;
  79                 struct __alloc_1p_arrays {
  80                         struct page *pages[group_width];
  81                         struct page *scribble[group_width];
  82                         char page_is_read[data_devs];
  83                 } __a1pa[pages_in_unit];
  84         } *_aab;
  85         struct __alloc_1p_arrays *__a1pa;
  86         struct __alloc_1p_arrays *__a1pa_end;
  87         const unsigned sizeof__a1pa = sizeof(_aab->__a1pa[0]);
  88         unsigned num_a1pa, alloc_size, i;
  89
  90         /* FIXME: check these numbers in ore_verify_layout */
  91         BUG_ON(sizeof(_aab->__asp2d) > PAGE_SIZE);
  92         BUG_ON(sizeof__a1pa > PAGE_SIZE);
  93
  94         if (sizeof(*_aab) > PAGE_SIZE) {
  95                 num_a1pa = (PAGE_SIZE - sizeof(_aab->__asp2d)) / sizeof__a1pa;
  96                 alloc_size = sizeof(_aab->__asp2d) + sizeof__a1pa * num_a1pa;
  97         } else {
  98                 num_a1pa = pages_in_unit;
  99                 alloc_size = sizeof(*_aab);
 100         }
 101
 102         _aab = kzalloc(alloc_size, GFP_KERNEL);
 103         if (unlikely(!_aab)) {
 104                 ORE_DBGMSG("!! Failed to alloc sp2d size=%d\n", alloc_size);
 105                 return -ENOMEM;
 106         }
 107
 108         sp2d = &_aab->__asp2d.sp2d;
 109         *psp2d = sp2d; /* From here Just call _sp2d_free */
 110
 111         __a1pa = _aab->__a1pa;
 112         __a1pa_end = __a1pa + num_a1pa;
 113
 114         for (i = 0; i < pages_in_unit; ++i) {
 115                 if (unlikely(__a1pa >= __a1pa_end)) {
 116                         num_a1pa = min_t(unsigned, PAGE_SIZE / sizeof__a1pa,
 117                                                         pages_in_unit - i);
 118
 119                         __a1pa = kzalloc(num_a1pa * sizeof__a1pa, GFP_KERNEL);
 120                         if (unlikely(!__a1pa)) {
 121                                 ORE_DBGMSG("!! Failed to _alloc_1p_arrays=%d\n",
 122                                            num_a1pa);
 123                                 return -ENOMEM;
 124                         }
 125                         __a1pa_end = __a1pa + num_a1pa;
 126                         /* First *pages is marked for kfree of the buffer */
 127                         sp2d->_1p_stripes[i].alloc = true;
 128                 }
 129
 130                 sp2d->_1p_stripes[i].pages = __a1pa->pages;
 131                 sp2d->_1p_stripes[i].scribble = __a1pa->scribble ;
 132                 sp2d->_1p_stripes[i].page_is_read = __a1pa->page_is_read;
 133                 ++__a1pa;
 134         }
 135
 136         sp2d->parity = parity;
 137         sp2d->data_devs = data_devs;
 138         sp2d->pages_in_unit = pages_in_unit;
 139         return 0;
 140 }
 141
 142 static void _sp2d_reset(struct __stripe_pages_2d *sp2d,
 143                         const struct _ore_r4w_op *r4w, void *priv)
 144 {
 145         unsigned data_devs = sp2d->data_devs;
 146         unsigned group_width = data_devs + sp2d->parity;
 147         unsigned p;
 148
 149         if (!sp2d->needed)
 150                 return;
 151
 152         for (p = 0; p < sp2d->pages_in_unit; p++) {
 153                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 154
 155                 if (_1ps->write_count < group_width) {
 156                         unsigned c;
 157
 158                         for (c = 0; c < data_devs; c++)
 159                                 if (_1ps->page_is_read[c]) {
 160                                         struct page *page = _1ps->pages[c];
 161
 162                                         r4w->put_page(priv, page);
 163                                         _1ps->page_is_read[c] = false;
 164                                 }
 165                 }
 166
 167                 memset(_1ps->pages, 0, group_width * sizeof(*_1ps->pages));
 168                 _1ps->write_count = 0;
 169                 _1ps->tx = NULL;
 170         }
 171
 172         sp2d->needed = false;
 173 }
 174
 175 static void _sp2d_free(struct __stripe_pages_2d *sp2d)
 176 {
 177         unsigned i;
 178
 179         if (!sp2d)
 180                 return;
 181
 182         for (i = 0; i < sp2d->pages_in_unit; ++i) {
 183                 if (sp2d->_1p_stripes[i].alloc)
 184                         kfree(sp2d->_1p_stripes[i].pages);
 185         }
 186
 187         kfree(sp2d);
 188 }
 189
 190 static unsigned _sp2d_min_pg(struct __stripe_pages_2d *sp2d)
 191 {
 192         unsigned p;
 193
 194         for (p = 0; p < sp2d->pages_in_unit; p++) {
 195                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 196
 197                 if (_1ps->write_count)
 198                         return p;
 199         }
 200
 201         return ~0;
 202 }
 203
 204 static unsigned _sp2d_max_pg(struct __stripe_pages_2d *sp2d)
 205 {
 206         unsigned p;
 207
 208         for (p = sp2d->pages_in_unit - 1; p >= 0; --p) {
 209                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 210
 211                 if (_1ps->write_count)
 212                         return p;
 213         }
 214
 215         return ~0;
 216 }
 217
 218 static void _gen_xor_unit(struct __stripe_pages_2d *sp2d)
 219 {
 220         unsigned p;
 221         for (p = 0; p < sp2d->pages_in_unit; p++) {
 222                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 223
 224                 if (!_1ps->write_count)
 225                         continue;
 226
 227                 init_async_submit(&_1ps->submit,
 228                         ASYNC_TX_XOR_ZERO_DST | ASYNC_TX_ACK,
 229                         NULL,
 230                         NULL, NULL,
 231                         (addr_conv_t *)_1ps->scribble);
 232
 233                 /* TODO: raid6 */
 234                 _1ps->tx = async_xor(_1ps->pages[sp2d->data_devs], _1ps->pages,
 235                                      0, sp2d->data_devs, PAGE_SIZE,
 236                                      &_1ps->submit);
 237         }
 238
 239         for (p = 0; p < sp2d->pages_in_unit; p++) {
 240                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 241                 /* NOTE: We wait for HW synchronously (I don't have such HW
 242                  * to test with.) Is parallelism needed with today's multi
 243                  * cores?
 244                  */
 245                 async_tx_issue_pending(_1ps->tx);
 246         }
 247 }
 248
 249 void _ore_add_stripe_page(struct __stripe_pages_2d *sp2d,
 250                        struct ore_striping_info *si, struct page *page)
 251 {
 252         struct __1_page_stripe *_1ps;
 253
 254         sp2d->needed = true;
 255
 256         _1ps = &sp2d->_1p_stripes[si->cur_pg];
 257         _1ps->pages[si->cur_comp] = page;
 258         ++_1ps->write_count;
 259
 260         si->cur_pg = (si->cur_pg + 1) % sp2d->pages_in_unit;
 261         /* si->cur_comp is advanced outside at main loop */
 262 }
 263
 264 void _ore_add_sg_seg(struct ore_per_dev_state *per_dev, unsigned cur_len,
 265                      bool not_last)
 266 {
 267         struct osd_sg_entry *sge;
 268
 269         ORE_DBGMSG("dev=%d cur_len=0x%x not_last=%d cur_sg=%d "
 270                      "offset=0x%llx length=0x%x last_sgs_total=0x%x\n",
 271                      per_dev->dev, cur_len, not_last, per_dev->cur_sg,
 272                      _LLU(per_dev->offset), per_dev->length,
 273                      per_dev->last_sgs_total);
 274
 275         if (!per_dev->cur_sg) {
 276                 sge = per_dev->sglist;
 277
 278                 /* First time we prepare two entries */
 279                 if (per_dev->length) {
 280                         ++per_dev->cur_sg;
 281                         sge->offset = per_dev->offset;
 282                         sge->len = per_dev->length;
 283                 } else {
 284                         /* Here the parity is the first unit of this object.
 285                          * This happens every time we reach a parity device on
 286                          * the same stripe as the per_dev->offset. We need to
 287                          * just skip this unit.
 288                          */
 289                         per_dev->offset += cur_len;
 290                         return;
 291                 }
 292         } else {
 293                 /* finalize the last one */
 294                 sge = &per_dev->sglist[per_dev->cur_sg - 1];
 295                 sge->len = per_dev->length - per_dev->last_sgs_total;
 296         }
 297
 298         if (not_last) {
 299                 /* Partly prepare the next one */
 300                 struct osd_sg_entry *next_sge = sge + 1;
 301
 302                 ++per_dev->cur_sg;
 303                 next_sge->offset = sge->offset + sge->len + cur_len;
 304                 /* Save cur len so we know how mutch was added next time */
 305                 per_dev->last_sgs_total = per_dev->length;
 306                 next_sge->len = 0;
 307         } else if (!sge->len) {
 308                 /* Optimize for when the last unit is a parity */
 309                 --per_dev->cur_sg;
 310         }
 311 }
 312
 313 static int _alloc_read_4_write(struct ore_io_state *ios)
 314 {
 315         struct ore_layout *layout = ios->layout;
 316         int ret;
 317         /* We want to only read those pages not in cache so worst case
 318          * is a stripe populated with every other page
 319          */
 320         unsigned sgs_per_dev = ios->sp2d->pages_in_unit + 2;
 321
 322         ret = _ore_get_io_state(layout, ios->oc,
 323                                 layout->group_width * layout->mirrors_p1,
 324                                 sgs_per_dev, 0, &ios->ios_read_4_write);
 325         return ret;
 326 }
 327
 328 /* @si contains info of the to-be-inserted page. Update of @si should be
 329  * maintained by caller. Specificaly si->dev, si->obj_offset, ...
 330  */
 331 static int _add_to_r4w(struct ore_io_state *ios, struct ore_striping_info *si,
 332                        struct page *page, unsigned pg_len)
 333 {
 334         struct request_queue *q;
 335         struct ore_per_dev_state *per_dev;
 336         struct ore_io_state *read_ios;
 337         unsigned first_dev = si->dev - (si->dev %
 338                           (ios->layout->group_width * ios->layout->mirrors_p1));
 339         unsigned comp = si->dev - first_dev;
 340         unsigned added_len;
 341
 342         if (!ios->ios_read_4_write) {
 343                 int ret = _alloc_read_4_write(ios);
 344
 345                 if (unlikely(ret))
 346                         return ret;
 347         }
 348
 349         read_ios = ios->ios_read_4_write;
 350         read_ios->numdevs = ios->layout->group_width * ios->layout->mirrors_p1;
 351
 352         per_dev = &read_ios->per_dev[comp];
 353         if (!per_dev->length) {
 354                 per_dev->bio = bio_kmalloc(GFP_KERNEL,
 355                                            ios->sp2d->pages_in_unit);
 356                 if (unlikely(!per_dev->bio)) {
 357                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
 358                                      ios->sp2d->pages_in_unit);
 359                         return -ENOMEM;
 360                 }
 361                 per_dev->offset = si->obj_offset;
 362                 per_dev->dev = si->dev;
 363         } else if (si->obj_offset != (per_dev->offset + per_dev->length)) {
 364                 u64 gap = si->obj_offset - (per_dev->offset + per_dev->length);
 365
 366                 _ore_add_sg_seg(per_dev, gap, true);
 367         }
 368         q = osd_request_queue(ore_comp_dev(read_ios->oc, per_dev->dev));
 369         added_len = bio_add_pc_page(q, per_dev->bio, page, pg_len,
 370                                     si->obj_offset % PAGE_SIZE);
 371         if (unlikely(added_len != pg_len)) {
 372                 ORE_DBGMSG("Failed to bio_add_pc_page bi_vcnt=%d\n",
 373                               per_dev->bio->bi_vcnt);
 374                 return -ENOMEM;
 375         }
 376
 377         per_dev->length += pg_len;
 378         return 0;
 379 }
 380
 381 /* read the beginning of an unaligned first page */
 382 static int _add_to_r4w_first_page(struct ore_io_state *ios, struct page *page)
 383 {
 384         struct ore_striping_info si;
 385         unsigned pg_len;
 386
 387         ore_calc_stripe_info(ios->layout, ios->offset, 0, &si);
 388
 389         pg_len = si.obj_offset % PAGE_SIZE;
 390         si.obj_offset -= pg_len;
 391
 392         ORE_DBGMSG("offset=0x%llx len=0x%x index=0x%lx dev=%x\n",
 393                    _LLU(si.obj_offset), pg_len, page->index, si.dev);
 394
 395         return _add_to_r4w(ios, &si, page, pg_len);
 396 }
 397
 398 /* read the end of an incomplete last page */
 399 static int _add_to_r4w_last_page(struct ore_io_state *ios, u64 *offset)
 400 {
 401         struct ore_striping_info si;
 402         struct page *page;
 403         unsigned pg_len, p, c;
 404
 405         ore_calc_stripe_info(ios->layout, *offset, 0, &si);
 406
 407         p = si.unit_off / PAGE_SIZE;
 408         c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1,
 409                        ios->layout->mirrors_p1, si.par_dev, si.dev);
 410         page = ios->sp2d->_1p_stripes[p].pages[c];
 411
 412         pg_len = PAGE_SIZE - (si.unit_off % PAGE_SIZE);
 413         *offset += pg_len;
 414
 415         ORE_DBGMSG("p=%d, c=%d next-offset=0x%llx len=0x%x dev=%x par_dev=%d\n",
 416                    p, c, _LLU(*offset), pg_len, si.dev, si.par_dev);
 417
 418         BUG_ON(!page);
 419
 420         return _add_to_r4w(ios, &si, page, pg_len);
 421 }
 422
 423 static void _mark_read4write_pages_uptodate(struct ore_io_state *ios, int ret)
 424 {
 425         struct bio_vec *bv;
 426         unsigned i, d;
 427
 428         /* loop on all devices all pages */
 429         for (d = 0; d < ios->numdevs; d++) {
 430                 struct bio *bio = ios->per_dev[d].bio;
 431
 432                 if (!bio)
 433                         continue;
 434
 435                 __bio_for_each_segment(bv, bio, i, 0) {
 436                         struct page *page = bv->bv_page;
 437
 438                         SetPageUptodate(page);
 439                         if (PageError(page))
 440                                 ClearPageError(page);
 441                 }
 442         }
 443 }
 444
 445 /* read_4_write is hacked to read the start of the first stripe and/or
 446  * the end of the last stripe. If needed, with an sg-gap at each device/page.
 447  * It is assumed to be called after the to_be_written pages of the first stripe
 448  * are populating ios->sp2d[][]
 449  *
 450  * NOTE: We call ios->r4w->lock_fn for all pages needed for parity calculations
 451  * These pages are held at sp2d[p].pages[c] but with
 452  * sp2d[p].page_is_read[c] = true. At _sp2d_reset these pages are
 453  * ios->r4w->lock_fn(). The ios->r4w->lock_fn might signal that the page is
 454  * @uptodate=true, so we don't need to read it, only unlock, after IO.
 455  *
 456  * TODO: The read_4_write should calc a need_to_read_pages_count, if bigger then
 457  * to-be-written count, we should consider the xor-in-place mode.
 458  * need_to_read_pages_count is the actual number of pages not present in cache.
 459  * maybe "devs_in_group - ios->sp2d[p].write_count" is a good enough
 460  * approximation? In this mode the read pages are put in the empty places of
 461  * ios->sp2d[p][*], xor is calculated the same way. These pages are
 462  * allocated/freed and don't go through cache
 463  */
 464 static int _read_4_write(struct ore_io_state *ios)
 465 {
 466         struct ore_io_state *ios_read;
 467         struct ore_striping_info read_si;
 468         struct __stripe_pages_2d *sp2d = ios->sp2d;
 469         u64 offset = ios->si.first_stripe_start;
 470         u64 last_stripe_end;
 471         unsigned bytes_in_stripe = ios->si.bytes_in_stripe;
 472         unsigned i, c, p, min_p = sp2d->pages_in_unit, max_p = -1;
 473         int ret;
 474
 475         if (offset == ios->offset) /* Go to start collect $200 */
 476                 goto read_last_stripe;
 477
 478         min_p = _sp2d_min_pg(sp2d);
 479         max_p = _sp2d_max_pg(sp2d);
 480
 481         for (c = 0; ; c++) {
 482                 ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
 483                 read_si.obj_offset += min_p * PAGE_SIZE;
 484                 offset += min_p * PAGE_SIZE;
 485                 for (p = min_p; p <= max_p; p++) {
 486                         struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 487                         struct page **pp = &_1ps->pages[c];
 488                         bool uptodate;
 489
 490                         if (*pp) {
 491                                 if (ios->offset % PAGE_SIZE)
 492                                         /* Read the remainder of the page */
 493                                         _add_to_r4w_first_page(ios, *pp);
 494                                 /* to-be-written pages start here */
 495                                 goto read_last_stripe;
 496                         }
 497
 498                         *pp = ios->r4w->get_page(ios->private, offset,
 499                                                  &uptodate);
 500                         if (unlikely(!*pp))
 501                                 return -ENOMEM;
 502
 503                         if (!uptodate)
 504                                 _add_to_r4w(ios, &read_si, *pp, PAGE_SIZE);
 505
 506                         /* Mark read-pages to be cache_released */
 507                         _1ps->page_is_read[c] = true;
 508                         read_si.obj_offset += PAGE_SIZE;
 509                         offset += PAGE_SIZE;
 510                 }
 511                 offset += (sp2d->pages_in_unit - p) * PAGE_SIZE;
 512         }
 513
 514 read_last_stripe:
 515         offset = ios->offset + ios->length;
 516         if (offset % PAGE_SIZE)
 517                 _add_to_r4w_last_page(ios, &offset);
 518                 /* offset will be aligned to next page */
 519
 520         last_stripe_end = div_u64(offset + bytes_in_stripe - 1, bytes_in_stripe)
 521                                  * bytes_in_stripe;
 522         if (offset == last_stripe_end) /* Optimize for the aligned case */
 523                 goto read_it;
 524
 525         ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
 526         p = read_si.unit_off / PAGE_SIZE;
 527         c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1,
 528                        ios->layout->mirrors_p1, read_si.par_dev, read_si.dev);
 529
 530         BUG_ON(ios->si.first_stripe_start + bytes_in_stripe != last_stripe_end);
 531         /* unaligned IO must be within a single stripe */
 532
 533         if (min_p == sp2d->pages_in_unit) {
 534                 /* Didn't do it yet */
 535                 min_p = _sp2d_min_pg(sp2d);
 536                 max_p = _sp2d_max_pg(sp2d);
 537         }
 538
 539         while (offset < last_stripe_end) {
 540                 struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p];
 541
 542                 if ((min_p <= p) && (p <= max_p)) {
 543                         struct page *page;
 544                         bool uptodate;
 545
 546                         BUG_ON(_1ps->pages[c]);
 547                         page = ios->r4w->get_page(ios->private, offset,
 548                                                   &uptodate);
 549                         if (unlikely(!page))
 550                                 return -ENOMEM;
 551
 552                         _1ps->pages[c] = page;
 553                         /* Mark read-pages to be cache_released */
 554                         _1ps->page_is_read[c] = true;
 555                         if (!uptodate)
 556                                 _add_to_r4w(ios, &read_si, page, PAGE_SIZE);
 557                 }
 558
 559                 offset += PAGE_SIZE;
 560                 if (p == (sp2d->pages_in_unit - 1)) {
 561                         ++c;
 562                         p = 0;
 563                         ore_calc_stripe_info(ios->layout, offset, 0, &read_si);
 564                 } else {
 565                         read_si.obj_offset += PAGE_SIZE;
 566                         ++p;
 567                 }
 568         }
 569
 570 read_it:
 571         ios_read = ios->ios_read_4_write;
 572         if (!ios_read)
 573                 return 0;
 574
 575         /* FIXME: Ugly to signal _sbi_read_mirror that we have bio(s). Change
 576          * to check for per_dev->bio
 577          */
 578         ios_read->pages = ios->pages;
 579
 580         /* Now read these devices */
 581         for (i = 0; i < ios_read->numdevs; i += ios_read->layout->mirrors_p1) {
 582                 ret = _ore_read_mirror(ios_read, i);
 583                 if (unlikely(ret))
 584                         return ret;
 585         }
 586
 587         ret = ore_io_execute(ios_read); /* Synchronus execution */
 588         if (unlikely(ret)) {
 589                 ORE_DBGMSG("!! ore_io_execute => %d\n", ret);
 590                 return ret;
 591         }
 592
 593         _mark_read4write_pages_uptodate(ios_read, ret);
 594         return 0;
 595 }
 596
 597 /* In writes @cur_len means length left. .i.e cur_len==0 is the last parity U */
 598 int _ore_add_parity_unit(struct ore_io_state *ios,
 599                             struct ore_striping_info *si,
 600                             struct ore_per_dev_state *per_dev,
 601                             unsigned cur_len)
 602 {
 603         if (ios->reading) {
 604                 if (per_dev->cur_sg >= ios->sgs_per_dev) {
 605                         ORE_DBGMSG("cur_sg(%d) >= sgs_per_dev(%d)\n" ,
 606                                 per_dev->cur_sg, ios->sgs_per_dev);
 607                         return -ENOMEM;
 608                 }
 609                 _ore_add_sg_seg(per_dev, cur_len, true);
 610         } else {
 611                 struct __stripe_pages_2d *sp2d = ios->sp2d;
 612                 struct page **pages = ios->parity_pages + ios->cur_par_page;
 613                 unsigned num_pages;
 614                 unsigned array_start = 0;
 615                 unsigned i;
 616                 int ret;
 617
 618                 si->cur_pg = _sp2d_min_pg(sp2d);
 619                 num_pages  = _sp2d_max_pg(sp2d) + 1 - si->cur_pg;
 620
 621                 if (!cur_len) /* If last stripe operate on parity comp */
 622                         si->cur_comp = sp2d->data_devs;
 623
 624                 if (!per_dev->length) {
 625                         per_dev->offset += si->cur_pg * PAGE_SIZE;
 626                         /* If first stripe, Read in all read4write pages
 627                          * (if needed) before we calculate the first parity.
 628                          */
 629                         _read_4_write(ios);
 630                 }
 631
 632                 for (i = 0; i < num_pages; i++) {
 633                         pages[i] = _raid_page_alloc();
 634                         if (unlikely(!pages[i]))
 635                                 return -ENOMEM;
 636
 637                         ++(ios->cur_par_page);
 638                 }
 639
 640                 BUG_ON(si->cur_comp != sp2d->data_devs);
 641                 BUG_ON(si->cur_pg + num_pages > sp2d->pages_in_unit);
 642
 643                 ret = _ore_add_stripe_unit(ios,  &array_start, 0, pages,
 644                                            per_dev, num_pages * PAGE_SIZE);
 645                 if (unlikely(ret))
 646                         return ret;
 647
 648                 /* TODO: raid6 if (last_parity_dev) */
 649                 _gen_xor_unit(sp2d);
 650                 _sp2d_reset(sp2d, ios->r4w, ios->private);
 651         }
 652         return 0;
 653 }
 654
 655 int _ore_post_alloc_raid_stuff(struct ore_io_state *ios)
 656 {
 657         struct ore_layout *layout = ios->layout;
 658
 659         if (ios->parity_pages) {
 660                 unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
 661                 unsigned stripe_size = ios->si.bytes_in_stripe;
 662                 u64 last_stripe, first_stripe;
 663
 664                 if (_sp2d_alloc(pages_in_unit, layout->group_width,
 665                                 layout->parity, &ios->sp2d)) {
 666                         return -ENOMEM;
 667                 }
 668
 669                 /* Round io down to last full strip */
 670                 first_stripe = div_u64(ios->offset, stripe_size);
 671                 last_stripe = div_u64(ios->offset + ios->length, stripe_size);
 672
 673                 /* If an IO spans more then a single stripe it must end at
 674                  * a stripe boundary. The reminder at the end is pushed into the
 675                  * next IO.
 676                  */
 677                 if (last_stripe != first_stripe) {
 678                         ios->length = last_stripe * stripe_size - ios->offset;
 679
 680                         BUG_ON(!ios->length);
 681                         ios->nr_pages = (ios->length + PAGE_SIZE - 1) /
 682                                         PAGE_SIZE;
 683                         ios->si.length = ios->length; /*make it consistent */
 684                 }
 685         }
 686         return 0;
 687 }
 688
 689 void _ore_free_raid_stuff(struct ore_io_state *ios)
 690 {
 691         if (ios->sp2d) { /* writing and raid */
 692                 unsigned i;
 693
 694                 for (i = 0; i < ios->cur_par_page; i++) {
 695                         struct page *page = ios->parity_pages[i];
 696
 697                         if (page)
 698                                 _raid_page_free(page);
 699                 }
 700                 if (ios->extra_part_alloc)
 701                         kfree(ios->parity_pages);
 702                 /* If IO returned an error pages might need unlocking */
 703                 _sp2d_reset(ios->sp2d, ios->r4w, ios->private);
 704                 _sp2d_free(ios->sp2d);
 705         } else {
 706                 /* Will only be set if raid reading && sglist is big */
 707                 if (ios->extra_part_alloc)
 708                         kfree(ios->per_dev[0].sglist);
 709         }
 710         if (ios->ios_read_4_write)
 711                 ore_put_io_state(ios->ios_read_4_write);
 712 }