Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 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> |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 17 | #include <linux/async_tx.h> |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 18 | |
| 19 | #include "ore_raid.h" |
| 20 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 21 | #undef ORE_DBGMSG2 |
| 22 | #define ORE_DBGMSG2 ORE_DBGMSG |
| 23 | |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 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 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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; |
Boaz Harrosh | 537632e | 2012-07-11 15:27:13 +0300 | [diff] [blame] | 147 | int p, c; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 148 | |
| 149 | if (!sp2d->needed) |
| 150 | return; |
| 151 | |
Boaz Harrosh | 537632e | 2012-07-11 15:27:13 +0300 | [diff] [blame] | 152 | for (c = data_devs - 1; c >= 0; --c) |
| 153 | for (p = sp2d->pages_in_unit - 1; p >= 0; --p) { |
| 154 | struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p]; |
| 155 | |
| 156 | if (_1ps->page_is_read[c]) { |
| 157 | struct page *page = _1ps->pages[c]; |
| 158 | |
| 159 | r4w->put_page(priv, page); |
| 160 | _1ps->page_is_read[c] = false; |
| 161 | } |
| 162 | } |
| 163 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 164 | for (p = 0; p < sp2d->pages_in_unit; p++) { |
| 165 | struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p]; |
| 166 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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 | { |
Dan Carpenter | 74b217d | 2012-10-02 11:28:45 +0300 | [diff] [blame] | 206 | int p; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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 | |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 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 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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 | */ |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 331 | static int _add_to_r4w(struct ore_io_state *ios, struct ore_striping_info *si, |
| 332 | struct page *page, unsigned pg_len) |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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)); |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 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)) { |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 372 | ORE_DBGMSG("Failed to bio_add_pc_page bi_vcnt=%d\n", |
| 373 | per_dev->bio->bi_vcnt); |
| 374 | return -ENOMEM; |
| 375 | } |
| 376 | |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 377 | per_dev->length += pg_len; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 378 | return 0; |
| 379 | } |
| 380 | |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 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 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 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 | */ |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 464 | static int _read_4_write_first_stripe(struct ore_io_state *ios) |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 465 | { |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 466 | struct ore_striping_info read_si; |
| 467 | struct __stripe_pages_2d *sp2d = ios->sp2d; |
| 468 | u64 offset = ios->si.first_stripe_start; |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 469 | unsigned c, p, min_p = sp2d->pages_in_unit, max_p = -1; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 470 | |
| 471 | if (offset == ios->offset) /* Go to start collect $200 */ |
| 472 | goto read_last_stripe; |
| 473 | |
| 474 | min_p = _sp2d_min_pg(sp2d); |
| 475 | max_p = _sp2d_max_pg(sp2d); |
| 476 | |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 477 | ORE_DBGMSG("stripe_start=0x%llx ios->offset=0x%llx min_p=%d max_p=%d\n", |
| 478 | offset, ios->offset, min_p, max_p); |
| 479 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 480 | for (c = 0; ; c++) { |
| 481 | ore_calc_stripe_info(ios->layout, offset, 0, &read_si); |
| 482 | read_si.obj_offset += min_p * PAGE_SIZE; |
| 483 | offset += min_p * PAGE_SIZE; |
| 484 | for (p = min_p; p <= max_p; p++) { |
| 485 | struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p]; |
| 486 | struct page **pp = &_1ps->pages[c]; |
| 487 | bool uptodate; |
| 488 | |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 489 | if (*pp) { |
| 490 | if (ios->offset % PAGE_SIZE) |
| 491 | /* Read the remainder of the page */ |
| 492 | _add_to_r4w_first_page(ios, *pp); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 493 | /* to-be-written pages start here */ |
| 494 | goto read_last_stripe; |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 495 | } |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 496 | |
| 497 | *pp = ios->r4w->get_page(ios->private, offset, |
| 498 | &uptodate); |
| 499 | if (unlikely(!*pp)) |
| 500 | return -ENOMEM; |
| 501 | |
| 502 | if (!uptodate) |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 503 | _add_to_r4w(ios, &read_si, *pp, PAGE_SIZE); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 504 | |
| 505 | /* Mark read-pages to be cache_released */ |
| 506 | _1ps->page_is_read[c] = true; |
| 507 | read_si.obj_offset += PAGE_SIZE; |
| 508 | offset += PAGE_SIZE; |
| 509 | } |
| 510 | offset += (sp2d->pages_in_unit - p) * PAGE_SIZE; |
| 511 | } |
| 512 | |
| 513 | read_last_stripe: |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 514 | return 0; |
| 515 | } |
| 516 | |
| 517 | static int _read_4_write_last_stripe(struct ore_io_state *ios) |
| 518 | { |
| 519 | struct ore_striping_info read_si; |
| 520 | struct __stripe_pages_2d *sp2d = ios->sp2d; |
| 521 | u64 offset; |
| 522 | u64 last_stripe_end; |
| 523 | unsigned bytes_in_stripe = ios->si.bytes_in_stripe; |
| 524 | unsigned c, p, min_p = sp2d->pages_in_unit, max_p = -1; |
| 525 | |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 526 | offset = ios->offset + ios->length; |
| 527 | if (offset % PAGE_SIZE) |
| 528 | _add_to_r4w_last_page(ios, &offset); |
| 529 | /* offset will be aligned to next page */ |
| 530 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 531 | last_stripe_end = div_u64(offset + bytes_in_stripe - 1, bytes_in_stripe) |
| 532 | * bytes_in_stripe; |
| 533 | if (offset == last_stripe_end) /* Optimize for the aligned case */ |
| 534 | goto read_it; |
| 535 | |
| 536 | ore_calc_stripe_info(ios->layout, offset, 0, &read_si); |
| 537 | p = read_si.unit_off / PAGE_SIZE; |
| 538 | c = _dev_order(ios->layout->group_width * ios->layout->mirrors_p1, |
| 539 | ios->layout->mirrors_p1, read_si.par_dev, read_si.dev); |
| 540 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 541 | if (min_p == sp2d->pages_in_unit) { |
| 542 | /* Didn't do it yet */ |
| 543 | min_p = _sp2d_min_pg(sp2d); |
| 544 | max_p = _sp2d_max_pg(sp2d); |
| 545 | } |
| 546 | |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 547 | ORE_DBGMSG("offset=0x%llx stripe_end=0x%llx min_p=%d max_p=%d\n", |
| 548 | offset, last_stripe_end, min_p, max_p); |
| 549 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 550 | while (offset < last_stripe_end) { |
| 551 | struct __1_page_stripe *_1ps = &sp2d->_1p_stripes[p]; |
| 552 | |
| 553 | if ((min_p <= p) && (p <= max_p)) { |
| 554 | struct page *page; |
| 555 | bool uptodate; |
| 556 | |
| 557 | BUG_ON(_1ps->pages[c]); |
| 558 | page = ios->r4w->get_page(ios->private, offset, |
| 559 | &uptodate); |
| 560 | if (unlikely(!page)) |
| 561 | return -ENOMEM; |
| 562 | |
| 563 | _1ps->pages[c] = page; |
| 564 | /* Mark read-pages to be cache_released */ |
| 565 | _1ps->page_is_read[c] = true; |
| 566 | if (!uptodate) |
Boaz Harrosh | 724577c | 2011-12-28 19:21:45 +0200 | [diff] [blame] | 567 | _add_to_r4w(ios, &read_si, page, PAGE_SIZE); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 568 | } |
| 569 | |
| 570 | offset += PAGE_SIZE; |
| 571 | if (p == (sp2d->pages_in_unit - 1)) { |
| 572 | ++c; |
| 573 | p = 0; |
| 574 | ore_calc_stripe_info(ios->layout, offset, 0, &read_si); |
| 575 | } else { |
| 576 | read_si.obj_offset += PAGE_SIZE; |
| 577 | ++p; |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | read_it: |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 582 | return 0; |
| 583 | } |
| 584 | |
| 585 | static int _read_4_write_execute(struct ore_io_state *ios) |
| 586 | { |
| 587 | struct ore_io_state *ios_read; |
| 588 | unsigned i; |
| 589 | int ret; |
| 590 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 591 | ios_read = ios->ios_read_4_write; |
| 592 | if (!ios_read) |
| 593 | return 0; |
| 594 | |
| 595 | /* FIXME: Ugly to signal _sbi_read_mirror that we have bio(s). Change |
| 596 | * to check for per_dev->bio |
| 597 | */ |
| 598 | ios_read->pages = ios->pages; |
| 599 | |
| 600 | /* Now read these devices */ |
| 601 | for (i = 0; i < ios_read->numdevs; i += ios_read->layout->mirrors_p1) { |
| 602 | ret = _ore_read_mirror(ios_read, i); |
| 603 | if (unlikely(ret)) |
| 604 | return ret; |
| 605 | } |
| 606 | |
| 607 | ret = ore_io_execute(ios_read); /* Synchronus execution */ |
| 608 | if (unlikely(ret)) { |
| 609 | ORE_DBGMSG("!! ore_io_execute => %d\n", ret); |
| 610 | return ret; |
| 611 | } |
| 612 | |
| 613 | _mark_read4write_pages_uptodate(ios_read, ret); |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 614 | ore_put_io_state(ios_read); |
| 615 | ios->ios_read_4_write = NULL; /* Might need a reuse at last stripe */ |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 616 | return 0; |
| 617 | } |
| 618 | |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 619 | /* In writes @cur_len means length left. .i.e cur_len==0 is the last parity U */ |
| 620 | int _ore_add_parity_unit(struct ore_io_state *ios, |
| 621 | struct ore_striping_info *si, |
| 622 | struct ore_per_dev_state *per_dev, |
| 623 | unsigned cur_len) |
| 624 | { |
| 625 | if (ios->reading) { |
Boaz Harrosh | 361aba5 | 2011-12-28 19:14:23 +0200 | [diff] [blame] | 626 | if (per_dev->cur_sg >= ios->sgs_per_dev) { |
| 627 | ORE_DBGMSG("cur_sg(%d) >= sgs_per_dev(%d)\n" , |
| 628 | per_dev->cur_sg, ios->sgs_per_dev); |
| 629 | return -ENOMEM; |
| 630 | } |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 631 | _ore_add_sg_seg(per_dev, cur_len, true); |
| 632 | } else { |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 633 | struct __stripe_pages_2d *sp2d = ios->sp2d; |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 634 | struct page **pages = ios->parity_pages + ios->cur_par_page; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 635 | unsigned num_pages; |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 636 | unsigned array_start = 0; |
| 637 | unsigned i; |
| 638 | int ret; |
| 639 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 640 | si->cur_pg = _sp2d_min_pg(sp2d); |
| 641 | num_pages = _sp2d_max_pg(sp2d) + 1 - si->cur_pg; |
| 642 | |
| 643 | if (!cur_len) /* If last stripe operate on parity comp */ |
| 644 | si->cur_comp = sp2d->data_devs; |
| 645 | |
| 646 | if (!per_dev->length) { |
| 647 | per_dev->offset += si->cur_pg * PAGE_SIZE; |
| 648 | /* If first stripe, Read in all read4write pages |
| 649 | * (if needed) before we calculate the first parity. |
| 650 | */ |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 651 | _read_4_write_first_stripe(ios); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 652 | } |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 653 | if (!cur_len) /* If last stripe r4w pages of last stripe */ |
| 654 | _read_4_write_last_stripe(ios); |
| 655 | _read_4_write_execute(ios); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 656 | |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 657 | for (i = 0; i < num_pages; i++) { |
| 658 | pages[i] = _raid_page_alloc(); |
| 659 | if (unlikely(!pages[i])) |
| 660 | return -ENOMEM; |
| 661 | |
| 662 | ++(ios->cur_par_page); |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 663 | } |
| 664 | |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 665 | BUG_ON(si->cur_comp != sp2d->data_devs); |
| 666 | BUG_ON(si->cur_pg + num_pages > sp2d->pages_in_unit); |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 667 | |
| 668 | ret = _ore_add_stripe_unit(ios, &array_start, 0, pages, |
| 669 | per_dev, num_pages * PAGE_SIZE); |
| 670 | if (unlikely(ret)) |
| 671 | return ret; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 672 | |
| 673 | /* TODO: raid6 if (last_parity_dev) */ |
| 674 | _gen_xor_unit(sp2d); |
| 675 | _sp2d_reset(sp2d, ios->r4w, ios->private); |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 676 | } |
| 677 | return 0; |
| 678 | } |
| 679 | |
| 680 | int _ore_post_alloc_raid_stuff(struct ore_io_state *ios) |
| 681 | { |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 682 | if (ios->parity_pages) { |
Boaz Harrosh | 9ff1930 | 2012-06-08 01:19:07 +0300 | [diff] [blame] | 683 | struct ore_layout *layout = ios->layout; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 684 | unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE; |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 685 | |
| 686 | if (_sp2d_alloc(pages_in_unit, layout->group_width, |
| 687 | layout->parity, &ios->sp2d)) { |
| 688 | return -ENOMEM; |
| 689 | } |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 690 | } |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 691 | return 0; |
| 692 | } |
| 693 | |
| 694 | void _ore_free_raid_stuff(struct ore_io_state *ios) |
| 695 | { |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 696 | if (ios->sp2d) { /* writing and raid */ |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 697 | unsigned i; |
| 698 | |
| 699 | for (i = 0; i < ios->cur_par_page; i++) { |
| 700 | struct page *page = ios->parity_pages[i]; |
| 701 | |
| 702 | if (page) |
| 703 | _raid_page_free(page); |
| 704 | } |
| 705 | if (ios->extra_part_alloc) |
| 706 | kfree(ios->parity_pages); |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 707 | /* If IO returned an error pages might need unlocking */ |
| 708 | _sp2d_reset(ios->sp2d, ios->r4w, ios->private); |
| 709 | _sp2d_free(ios->sp2d); |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 710 | } else { |
| 711 | /* Will only be set if raid reading && sglist is big */ |
| 712 | if (ios->extra_part_alloc) |
| 713 | kfree(ios->per_dev[0].sglist); |
| 714 | } |
Boaz Harrosh | 769ba8d | 2011-10-14 15:33:51 +0200 | [diff] [blame] | 715 | if (ios->ios_read_4_write) |
| 716 | ore_put_io_state(ios->ios_read_4_write); |
Boaz Harrosh | a1fec1d | 2011-10-12 18:42:22 +0200 | [diff] [blame] | 717 | } |