Loading...
Note: File does not exist in v6.8.
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
32MODULE_AUTHOR("Boaz Harrosh <bharrosh@panasas.com>");
33MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
34MODULE_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
46enum { BIO_MAX_PAGES_KMALLOC =
47 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),};
48
49int ore_verify_layout(unsigned total_comps, struct ore_layout *layout)
50{
51 u64 stripe_length;
52
53 switch (layout->raid_algorithm) {
54 case PNFS_OSD_RAID_0:
55 layout->parity = 0;
56 break;
57 case PNFS_OSD_RAID_5:
58 layout->parity = 1;
59 break;
60 case PNFS_OSD_RAID_PQ:
61 case PNFS_OSD_RAID_4:
62 default:
63 ORE_ERR("Only RAID_0/5 for now\n");
64 return -EINVAL;
65 }
66 if (0 != (layout->stripe_unit & ~PAGE_MASK)) {
67 ORE_ERR("Stripe Unit(0x%llx)"
68 " must be Multples of PAGE_SIZE(0x%lx)\n",
69 _LLU(layout->stripe_unit), PAGE_SIZE);
70 return -EINVAL;
71 }
72 if (layout->group_width) {
73 if (!layout->group_depth) {
74 ORE_ERR("group_depth == 0 && group_width != 0\n");
75 return -EINVAL;
76 }
77 if (total_comps < (layout->group_width * layout->mirrors_p1)) {
78 ORE_ERR("Data Map wrong, "
79 "numdevs=%d < group_width=%d * mirrors=%d\n",
80 total_comps, layout->group_width,
81 layout->mirrors_p1);
82 return -EINVAL;
83 }
84 layout->group_count = total_comps / layout->mirrors_p1 /
85 layout->group_width;
86 } else {
87 if (layout->group_depth) {
88 printk(KERN_NOTICE "Warning: group_depth ignored "
89 "group_width == 0 && group_depth == %lld\n",
90 _LLU(layout->group_depth));
91 }
92 layout->group_width = total_comps / layout->mirrors_p1;
93 layout->group_depth = -1;
94 layout->group_count = 1;
95 }
96
97 stripe_length = (u64)layout->group_width * layout->stripe_unit;
98 if (stripe_length >= (1ULL << 32)) {
99 ORE_ERR("Stripe_length(0x%llx) >= 32bit is not supported\n",
100 _LLU(stripe_length));
101 return -EINVAL;
102 }
103
104 layout->max_io_length =
105 (BIO_MAX_PAGES_KMALLOC * PAGE_SIZE - layout->stripe_unit) *
106 (layout->group_width - layout->parity);
107 if (layout->parity) {
108 unsigned stripe_length =
109 (layout->group_width - layout->parity) *
110 layout->stripe_unit;
111
112 layout->max_io_length /= stripe_length;
113 layout->max_io_length *= stripe_length;
114 }
115 return 0;
116}
117EXPORT_SYMBOL(ore_verify_layout);
118
119static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
120{
121 return ios->oc->comps[index & ios->oc->single_comp].cred;
122}
123
124static 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
129static 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
138int _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 */
237int ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
238 bool is_reading, u64 offset, u64 length,
239 struct ore_io_state **pios)
240{
241 struct ore_io_state *ios;
242 unsigned numdevs = layout->group_width * layout->mirrors_p1;
243 unsigned sgs_per_dev = 0, max_par_pages = 0;
244 int ret;
245
246 if (layout->parity && length) {
247 unsigned data_devs = layout->group_width - layout->parity;
248 unsigned stripe_size = layout->stripe_unit * data_devs;
249 unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
250 u32 remainder;
251 u64 num_stripes;
252 u64 num_raid_units;
253
254 num_stripes = div_u64_rem(length, stripe_size, &remainder);
255 if (remainder)
256 ++num_stripes;
257
258 num_raid_units = num_stripes * layout->parity;
259
260 if (is_reading) {
261 /* For reads add per_dev sglist array */
262 /* TODO: Raid 6 we need twice more. Actually:
263 * num_stripes / LCMdP(W,P);
264 * if (W%P != 0) num_stripes *= parity;
265 */
266
267 /* first/last seg is split */
268 num_raid_units += layout->group_width;
269 sgs_per_dev = div_u64(num_raid_units, data_devs) + 2;
270 } else {
271 /* For Writes add parity pages array. */
272 max_par_pages = num_raid_units * pages_in_unit *
273 sizeof(struct page *);
274 }
275 }
276
277 ret = _ore_get_io_state(layout, oc, numdevs, sgs_per_dev, max_par_pages,
278 pios);
279 if (unlikely(ret))
280 return ret;
281
282 ios = *pios;
283 ios->reading = is_reading;
284 ios->offset = offset;
285
286 if (length) {
287 ore_calc_stripe_info(layout, offset, length, &ios->si);
288 ios->length = ios->si.length;
289 ios->nr_pages = ((ios->offset & (PAGE_SIZE - 1)) +
290 ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
291 if (layout->parity)
292 _ore_post_alloc_raid_stuff(ios);
293 }
294
295 return 0;
296}
297EXPORT_SYMBOL(ore_get_rw_state);
298
299/* Allocate an io_state for all the devices in the comps array
300 *
301 * This version of io_state allocation is used mostly by create/remove
302 * and trunc where we currently need all the devices. The only wastful
303 * bit is the read/write_attributes with no IO. Those sites should
304 * be converted to use ore_get_rw_state() with length=0
305 */
306int ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
307 struct ore_io_state **pios)
308{
309 return _ore_get_io_state(layout, oc, oc->numdevs, 0, 0, pios);
310}
311EXPORT_SYMBOL(ore_get_io_state);
312
313void ore_put_io_state(struct ore_io_state *ios)
314{
315 if (ios) {
316 unsigned i;
317
318 for (i = 0; i < ios->numdevs; i++) {
319 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
320
321 if (per_dev->or)
322 osd_end_request(per_dev->or);
323 if (per_dev->bio)
324 bio_put(per_dev->bio);
325 }
326
327 _ore_free_raid_stuff(ios);
328 kfree(ios);
329 }
330}
331EXPORT_SYMBOL(ore_put_io_state);
332
333static void _sync_done(struct ore_io_state *ios, void *p)
334{
335 struct completion *waiting = p;
336
337 complete(waiting);
338}
339
340static void _last_io(struct kref *kref)
341{
342 struct ore_io_state *ios = container_of(
343 kref, struct ore_io_state, kref);
344
345 ios->done(ios, ios->private);
346}
347
348static void _done_io(struct osd_request *or, void *p)
349{
350 struct ore_io_state *ios = p;
351
352 kref_put(&ios->kref, _last_io);
353}
354
355int ore_io_execute(struct ore_io_state *ios)
356{
357 DECLARE_COMPLETION_ONSTACK(wait);
358 bool sync = (ios->done == NULL);
359 int i, ret;
360
361 if (sync) {
362 ios->done = _sync_done;
363 ios->private = &wait;
364 }
365
366 for (i = 0; i < ios->numdevs; i++) {
367 struct osd_request *or = ios->per_dev[i].or;
368 if (unlikely(!or))
369 continue;
370
371 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
372 if (unlikely(ret)) {
373 ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
374 ret);
375 return ret;
376 }
377 }
378
379 kref_init(&ios->kref);
380
381 for (i = 0; i < ios->numdevs; i++) {
382 struct osd_request *or = ios->per_dev[i].or;
383 if (unlikely(!or))
384 continue;
385
386 kref_get(&ios->kref);
387 osd_execute_request_async(or, _done_io, ios);
388 }
389
390 kref_put(&ios->kref, _last_io);
391 ret = 0;
392
393 if (sync) {
394 wait_for_completion(&wait);
395 ret = ore_check_io(ios, NULL);
396 }
397 return ret;
398}
399
400static void _clear_bio(struct bio *bio)
401{
402 struct bio_vec *bv;
403 unsigned i;
404
405 bio_for_each_segment_all(bv, bio, i) {
406 unsigned this_count = bv->bv_len;
407
408 if (likely(PAGE_SIZE == this_count))
409 clear_highpage(bv->bv_page);
410 else
411 zero_user(bv->bv_page, bv->bv_offset, this_count);
412 }
413}
414
415int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
416{
417 enum osd_err_priority acumulated_osd_err = 0;
418 int acumulated_lin_err = 0;
419 int i;
420
421 for (i = 0; i < ios->numdevs; i++) {
422 struct osd_sense_info osi;
423 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
424 struct osd_request *or = per_dev->or;
425 int ret;
426
427 if (unlikely(!or))
428 continue;
429
430 ret = osd_req_decode_sense(or, &osi);
431 if (likely(!ret))
432 continue;
433
434 if ((OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) &&
435 per_dev->bio) {
436 /* start read offset passed endof file.
437 * Note: if we do not have bio it means read-attributes
438 * In this case we should return error to caller.
439 */
440 _clear_bio(per_dev->bio);
441 ORE_DBGMSG("start read offset passed end of file "
442 "offset=0x%llx, length=0x%llx\n",
443 _LLU(per_dev->offset),
444 _LLU(per_dev->length));
445
446 continue; /* we recovered */
447 }
448
449 if (on_dev_error) {
450 u64 residual = ios->reading ?
451 or->in.residual : or->out.residual;
452 u64 offset = (ios->offset + ios->length) - residual;
453 unsigned dev = per_dev->dev - ios->oc->first_dev;
454 struct ore_dev *od = ios->oc->ods[dev];
455
456 on_dev_error(ios, od, dev, osi.osd_err_pri,
457 offset, residual);
458 }
459 if (osi.osd_err_pri >= acumulated_osd_err) {
460 acumulated_osd_err = osi.osd_err_pri;
461 acumulated_lin_err = ret;
462 }
463 }
464
465 return acumulated_lin_err;
466}
467EXPORT_SYMBOL(ore_check_io);
468
469/*
470 * L - logical offset into the file
471 *
472 * D - number of Data devices
473 * D = group_width - parity
474 *
475 * U - The number of bytes in a stripe within a group
476 * U = stripe_unit * D
477 *
478 * T - The number of bytes striped within a group of component objects
479 * (before advancing to the next group)
480 * T = U * group_depth
481 *
482 * S - The number of bytes striped across all component objects
483 * before the pattern repeats
484 * S = T * group_count
485 *
486 * M - The "major" (i.e., across all components) cycle number
487 * M = L / S
488 *
489 * G - Counts the groups from the beginning of the major cycle
490 * G = (L - (M * S)) / T [or (L % S) / T]
491 *
492 * H - The byte offset within the group
493 * H = (L - (M * S)) % T [or (L % S) % T]
494 *
495 * N - The "minor" (i.e., across the group) stripe number
496 * N = H / U
497 *
498 * C - The component index coresponding to L
499 *
500 * C = (H - (N * U)) / stripe_unit + G * D
501 * [or (L % U) / stripe_unit + G * D]
502 *
503 * O - The component offset coresponding to L
504 * O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
505 *
506 * LCMdP – Parity cycle: Lowest Common Multiple of group_width, parity
507 * divide by parity
508 * LCMdP = lcm(group_width, parity) / parity
509 *
510 * R - The parity Rotation stripe
511 * (Note parity cycle always starts at a group's boundary)
512 * R = N % LCMdP
513 *
514 * I = the first parity device index
515 * I = (group_width + group_width - R*parity - parity) % group_width
516 *
517 * Craid - The component index Rotated
518 * Craid = (group_width + C - R*parity) % group_width
519 * (We add the group_width to avoid negative numbers modulo math)
520 */
521void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
522 u64 length, struct ore_striping_info *si)
523{
524 u32 stripe_unit = layout->stripe_unit;
525 u32 group_width = layout->group_width;
526 u64 group_depth = layout->group_depth;
527 u32 parity = layout->parity;
528
529 u32 D = group_width - parity;
530 u32 U = D * stripe_unit;
531 u64 T = U * group_depth;
532 u64 S = T * layout->group_count;
533 u64 M = div64_u64(file_offset, S);
534
535 /*
536 G = (L - (M * S)) / T
537 H = (L - (M * S)) % T
538 */
539 u64 LmodS = file_offset - M * S;
540 u32 G = div64_u64(LmodS, T);
541 u64 H = LmodS - G * T;
542
543 u32 N = div_u64(H, U);
544 u32 Nlast;
545
546 /* "H - (N * U)" is just "H % U" so it's bound to u32 */
547 u32 C = (u32)(H - (N * U)) / stripe_unit + G * group_width;
548
549 div_u64_rem(file_offset, stripe_unit, &si->unit_off);
550
551 si->obj_offset = si->unit_off + (N * stripe_unit) +
552 (M * group_depth * stripe_unit);
553
554 if (parity) {
555 u32 LCMdP = lcm(group_width, parity) / parity;
556 /* R = N % LCMdP; */
557 u32 RxP = (N % LCMdP) * parity;
558 u32 first_dev = C - C % group_width;
559
560 si->par_dev = (group_width + group_width - parity - RxP) %
561 group_width + first_dev;
562 si->dev = (group_width + C - RxP) % group_width + first_dev;
563 si->bytes_in_stripe = U;
564 si->first_stripe_start = M * S + G * T + N * U;
565 } else {
566 /* Make the math correct see _prepare_one_group */
567 si->par_dev = group_width;
568 si->dev = C;
569 }
570
571 si->dev *= layout->mirrors_p1;
572 si->par_dev *= layout->mirrors_p1;
573 si->offset = file_offset;
574 si->length = T - H;
575 if (si->length > length)
576 si->length = length;
577
578 Nlast = div_u64(H + si->length + U - 1, U);
579 si->maxdevUnits = Nlast - N;
580
581 si->M = M;
582}
583EXPORT_SYMBOL(ore_calc_stripe_info);
584
585int _ore_add_stripe_unit(struct ore_io_state *ios, unsigned *cur_pg,
586 unsigned pgbase, struct page **pages,
587 struct ore_per_dev_state *per_dev, int cur_len)
588{
589 unsigned pg = *cur_pg;
590 struct request_queue *q =
591 osd_request_queue(_ios_od(ios, per_dev->dev));
592 unsigned len = cur_len;
593 int ret;
594
595 if (per_dev->bio == NULL) {
596 unsigned bio_size;
597
598 if (!ios->reading) {
599 bio_size = ios->si.maxdevUnits;
600 } else {
601 bio_size = (ios->si.maxdevUnits + 1) *
602 (ios->layout->group_width - ios->layout->parity) /
603 ios->layout->group_width;
604 }
605 bio_size *= (ios->layout->stripe_unit / PAGE_SIZE);
606
607 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
608 if (unlikely(!per_dev->bio)) {
609 ORE_DBGMSG("Failed to allocate BIO size=%u\n",
610 bio_size);
611 ret = -ENOMEM;
612 goto out;
613 }
614 }
615
616 while (cur_len > 0) {
617 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
618 unsigned added_len;
619
620 cur_len -= pglen;
621
622 added_len = bio_add_pc_page(q, per_dev->bio, pages[pg],
623 pglen, pgbase);
624 if (unlikely(pglen != added_len)) {
625 /* If bi_vcnt == bi_max then this is a SW BUG */
626 ORE_DBGMSG("Failed bio_add_pc_page bi_vcnt=0x%x "
627 "bi_max=0x%x BIO_MAX=0x%x cur_len=0x%x\n",
628 per_dev->bio->bi_vcnt,
629 per_dev->bio->bi_max_vecs,
630 BIO_MAX_PAGES_KMALLOC, cur_len);
631 ret = -ENOMEM;
632 goto out;
633 }
634 _add_stripe_page(ios->sp2d, &ios->si, pages[pg]);
635
636 pgbase = 0;
637 ++pg;
638 }
639 BUG_ON(cur_len);
640
641 per_dev->length += len;
642 *cur_pg = pg;
643 ret = 0;
644out: /* we fail the complete unit on an error eg don't advance
645 * per_dev->length and cur_pg. This means that we might have a bigger
646 * bio than the CDB requested length (per_dev->length). That's fine
647 * only the oposite is fatal.
648 */
649 return ret;
650}
651
652static int _prepare_for_striping(struct ore_io_state *ios)
653{
654 struct ore_striping_info *si = &ios->si;
655 unsigned stripe_unit = ios->layout->stripe_unit;
656 unsigned mirrors_p1 = ios->layout->mirrors_p1;
657 unsigned group_width = ios->layout->group_width;
658 unsigned devs_in_group = group_width * mirrors_p1;
659 unsigned dev = si->dev;
660 unsigned first_dev = dev - (dev % devs_in_group);
661 unsigned dev_order;
662 unsigned cur_pg = ios->pages_consumed;
663 u64 length = ios->length;
664 int ret = 0;
665
666 if (!ios->pages) {
667 ios->numdevs = ios->layout->mirrors_p1;
668 return 0;
669 }
670
671 BUG_ON(length > si->length);
672
673 dev_order = _dev_order(devs_in_group, mirrors_p1, si->par_dev, dev);
674 si->cur_comp = dev_order;
675 si->cur_pg = si->unit_off / PAGE_SIZE;
676
677 while (length) {
678 unsigned comp = dev - first_dev;
679 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
680 unsigned cur_len, page_off = 0;
681
682 if (!per_dev->length) {
683 per_dev->dev = dev;
684 if (dev == si->dev) {
685 WARN_ON(dev == si->par_dev);
686 per_dev->offset = si->obj_offset;
687 cur_len = stripe_unit - si->unit_off;
688 page_off = si->unit_off & ~PAGE_MASK;
689 BUG_ON(page_off && (page_off != ios->pgbase));
690 } else {
691 if (si->cur_comp > dev_order)
692 per_dev->offset =
693 si->obj_offset - si->unit_off;
694 else /* si->cur_comp < dev_order */
695 per_dev->offset =
696 si->obj_offset + stripe_unit -
697 si->unit_off;
698 cur_len = stripe_unit;
699 }
700 } else {
701 cur_len = stripe_unit;
702 }
703 if (cur_len >= length)
704 cur_len = length;
705
706 ret = _ore_add_stripe_unit(ios, &cur_pg, page_off, ios->pages,
707 per_dev, cur_len);
708 if (unlikely(ret))
709 goto out;
710
711 dev += mirrors_p1;
712 dev = (dev % devs_in_group) + first_dev;
713
714 length -= cur_len;
715
716 si->cur_comp = (si->cur_comp + 1) % group_width;
717 if (unlikely((dev == si->par_dev) || (!length && ios->sp2d))) {
718 if (!length && ios->sp2d) {
719 /* If we are writing and this is the very last
720 * stripe. then operate on parity dev.
721 */
722 dev = si->par_dev;
723 }
724 if (ios->sp2d)
725 /* In writes cur_len just means if it's the
726 * last one. See _ore_add_parity_unit.
727 */
728 cur_len = length;
729 per_dev = &ios->per_dev[dev - first_dev];
730 if (!per_dev->length) {
731 /* Only/always the parity unit of the first
732 * stripe will be empty. So this is a chance to
733 * initialize the per_dev info.
734 */
735 per_dev->dev = dev;
736 per_dev->offset = si->obj_offset - si->unit_off;
737 }
738
739 ret = _ore_add_parity_unit(ios, si, per_dev, cur_len);
740 if (unlikely(ret))
741 goto out;
742
743 /* Rotate next par_dev backwards with wraping */
744 si->par_dev = (devs_in_group + si->par_dev -
745 ios->layout->parity * mirrors_p1) %
746 devs_in_group + first_dev;
747 /* Next stripe, start fresh */
748 si->cur_comp = 0;
749 si->cur_pg = 0;
750 }
751 }
752out:
753 ios->numdevs = devs_in_group;
754 ios->pages_consumed = cur_pg;
755 return ret;
756}
757
758int ore_create(struct ore_io_state *ios)
759{
760 int i, ret;
761
762 for (i = 0; i < ios->oc->numdevs; i++) {
763 struct osd_request *or;
764
765 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
766 if (unlikely(!or)) {
767 ORE_ERR("%s: osd_start_request failed\n", __func__);
768 ret = -ENOMEM;
769 goto out;
770 }
771 ios->per_dev[i].or = or;
772 ios->numdevs++;
773
774 osd_req_create_object(or, _ios_obj(ios, i));
775 }
776 ret = ore_io_execute(ios);
777
778out:
779 return ret;
780}
781EXPORT_SYMBOL(ore_create);
782
783int ore_remove(struct ore_io_state *ios)
784{
785 int i, ret;
786
787 for (i = 0; i < ios->oc->numdevs; i++) {
788 struct osd_request *or;
789
790 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
791 if (unlikely(!or)) {
792 ORE_ERR("%s: osd_start_request failed\n", __func__);
793 ret = -ENOMEM;
794 goto out;
795 }
796 ios->per_dev[i].or = or;
797 ios->numdevs++;
798
799 osd_req_remove_object(or, _ios_obj(ios, i));
800 }
801 ret = ore_io_execute(ios);
802
803out:
804 return ret;
805}
806EXPORT_SYMBOL(ore_remove);
807
808static int _write_mirror(struct ore_io_state *ios, int cur_comp)
809{
810 struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
811 unsigned dev = ios->per_dev[cur_comp].dev;
812 unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
813 int ret = 0;
814
815 if (ios->pages && !master_dev->length)
816 return 0; /* Just an empty slot */
817
818 for (; cur_comp < last_comp; ++cur_comp, ++dev) {
819 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
820 struct osd_request *or;
821
822 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
823 if (unlikely(!or)) {
824 ORE_ERR("%s: osd_start_request failed\n", __func__);
825 ret = -ENOMEM;
826 goto out;
827 }
828 per_dev->or = or;
829
830 if (ios->pages) {
831 struct bio *bio;
832
833 if (per_dev != master_dev) {
834 bio = bio_clone_kmalloc(master_dev->bio,
835 GFP_KERNEL);
836 if (unlikely(!bio)) {
837 ORE_DBGMSG(
838 "Failed to allocate BIO size=%u\n",
839 master_dev->bio->bi_max_vecs);
840 ret = -ENOMEM;
841 goto out;
842 }
843
844 bio->bi_bdev = NULL;
845 bio->bi_next = NULL;
846 per_dev->offset = master_dev->offset;
847 per_dev->length = master_dev->length;
848 per_dev->bio = bio;
849 per_dev->dev = dev;
850 } else {
851 bio = master_dev->bio;
852 /* FIXME: bio_set_dir() */
853 bio->bi_rw |= REQ_WRITE;
854 }
855
856 osd_req_write(or, _ios_obj(ios, cur_comp),
857 per_dev->offset, bio, per_dev->length);
858 ORE_DBGMSG("write(0x%llx) offset=0x%llx "
859 "length=0x%llx dev=%d\n",
860 _LLU(_ios_obj(ios, cur_comp)->id),
861 _LLU(per_dev->offset),
862 _LLU(per_dev->length), dev);
863 } else if (ios->kern_buff) {
864 per_dev->offset = ios->si.obj_offset;
865 per_dev->dev = ios->si.dev + dev;
866
867 /* no cross device without page array */
868 BUG_ON((ios->layout->group_width > 1) &&
869 (ios->si.unit_off + ios->length >
870 ios->layout->stripe_unit));
871
872 ret = osd_req_write_kern(or, _ios_obj(ios, cur_comp),
873 per_dev->offset,
874 ios->kern_buff, ios->length);
875 if (unlikely(ret))
876 goto out;
877 ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
878 "length=0x%llx dev=%d\n",
879 _LLU(_ios_obj(ios, cur_comp)->id),
880 _LLU(per_dev->offset),
881 _LLU(ios->length), per_dev->dev);
882 } else {
883 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
884 ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
885 _LLU(_ios_obj(ios, cur_comp)->id),
886 ios->out_attr_len, dev);
887 }
888
889 if (ios->out_attr)
890 osd_req_add_set_attr_list(or, ios->out_attr,
891 ios->out_attr_len);
892
893 if (ios->in_attr)
894 osd_req_add_get_attr_list(or, ios->in_attr,
895 ios->in_attr_len);
896 }
897
898out:
899 return ret;
900}
901
902int ore_write(struct ore_io_state *ios)
903{
904 int i;
905 int ret;
906
907 if (unlikely(ios->sp2d && !ios->r4w)) {
908 /* A library is attempting a RAID-write without providing
909 * a pages lock interface.
910 */
911 WARN_ON_ONCE(1);
912 return -ENOTSUPP;
913 }
914
915 ret = _prepare_for_striping(ios);
916 if (unlikely(ret))
917 return ret;
918
919 for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
920 ret = _write_mirror(ios, i);
921 if (unlikely(ret))
922 return ret;
923 }
924
925 ret = ore_io_execute(ios);
926 return ret;
927}
928EXPORT_SYMBOL(ore_write);
929
930int _ore_read_mirror(struct ore_io_state *ios, unsigned cur_comp)
931{
932 struct osd_request *or;
933 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
934 struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
935 unsigned first_dev = (unsigned)obj->id;
936
937 if (ios->pages && !per_dev->length)
938 return 0; /* Just an empty slot */
939
940 first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
941 or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
942 if (unlikely(!or)) {
943 ORE_ERR("%s: osd_start_request failed\n", __func__);
944 return -ENOMEM;
945 }
946 per_dev->or = or;
947
948 if (ios->pages) {
949 if (per_dev->cur_sg) {
950 /* finalize the last sg_entry */
951 _ore_add_sg_seg(per_dev, 0, false);
952 if (unlikely(!per_dev->cur_sg))
953 return 0; /* Skip parity only device */
954
955 osd_req_read_sg(or, obj, per_dev->bio,
956 per_dev->sglist, per_dev->cur_sg);
957 } else {
958 /* The no raid case */
959 osd_req_read(or, obj, per_dev->offset,
960 per_dev->bio, per_dev->length);
961 }
962
963 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
964 " dev=%d sg_len=%d\n", _LLU(obj->id),
965 _LLU(per_dev->offset), _LLU(per_dev->length),
966 first_dev, per_dev->cur_sg);
967 } else {
968 BUG_ON(ios->kern_buff);
969
970 osd_req_get_attributes(or, obj);
971 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
972 _LLU(obj->id),
973 ios->in_attr_len, first_dev);
974 }
975 if (ios->out_attr)
976 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
977
978 if (ios->in_attr)
979 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
980
981 return 0;
982}
983
984int ore_read(struct ore_io_state *ios)
985{
986 int i;
987 int ret;
988
989 ret = _prepare_for_striping(ios);
990 if (unlikely(ret))
991 return ret;
992
993 for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
994 ret = _ore_read_mirror(ios, i);
995 if (unlikely(ret))
996 return ret;
997 }
998
999 ret = ore_io_execute(ios);
1000 return ret;
1001}
1002EXPORT_SYMBOL(ore_read);
1003
1004int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
1005{
1006 struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
1007 void *iter = NULL;
1008 int nelem;
1009
1010 do {
1011 nelem = 1;
1012 osd_req_decode_get_attr_list(ios->per_dev[0].or,
1013 &cur_attr, &nelem, &iter);
1014 if ((cur_attr.attr_page == attr->attr_page) &&
1015 (cur_attr.attr_id == attr->attr_id)) {
1016 attr->len = cur_attr.len;
1017 attr->val_ptr = cur_attr.val_ptr;
1018 return 0;
1019 }
1020 } while (iter);
1021
1022 return -EIO;
1023}
1024EXPORT_SYMBOL(extract_attr_from_ios);
1025
1026static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
1027 struct osd_attr *attr)
1028{
1029 int last_comp = cur_comp + ios->layout->mirrors_p1;
1030
1031 for (; cur_comp < last_comp; ++cur_comp) {
1032 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
1033 struct osd_request *or;
1034
1035 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
1036 if (unlikely(!or)) {
1037 ORE_ERR("%s: osd_start_request failed\n", __func__);
1038 return -ENOMEM;
1039 }
1040 per_dev->or = or;
1041
1042 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
1043 osd_req_add_set_attr_list(or, attr, 1);
1044 }
1045
1046 return 0;
1047}
1048
1049struct _trunc_info {
1050 struct ore_striping_info si;
1051 u64 prev_group_obj_off;
1052 u64 next_group_obj_off;
1053
1054 unsigned first_group_dev;
1055 unsigned nex_group_dev;
1056};
1057
1058static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
1059 struct _trunc_info *ti)
1060{
1061 unsigned stripe_unit = layout->stripe_unit;
1062
1063 ore_calc_stripe_info(layout, file_offset, 0, &ti->si);
1064
1065 ti->prev_group_obj_off = ti->si.M * stripe_unit;
1066 ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
1067
1068 ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
1069 ti->nex_group_dev = ti->first_group_dev + layout->group_width;
1070}
1071
1072int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
1073 u64 size)
1074{
1075 struct ore_io_state *ios;
1076 struct exofs_trunc_attr {
1077 struct osd_attr attr;
1078 __be64 newsize;
1079 } *size_attrs;
1080 struct _trunc_info ti;
1081 int i, ret;
1082
1083 ret = ore_get_io_state(layout, oc, &ios);
1084 if (unlikely(ret))
1085 return ret;
1086
1087 _calc_trunk_info(ios->layout, size, &ti);
1088
1089 size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
1090 GFP_KERNEL);
1091 if (unlikely(!size_attrs)) {
1092 ret = -ENOMEM;
1093 goto out;
1094 }
1095
1096 ios->numdevs = ios->oc->numdevs;
1097
1098 for (i = 0; i < ios->numdevs; ++i) {
1099 struct exofs_trunc_attr *size_attr = &size_attrs[i];
1100 u64 obj_size;
1101
1102 if (i < ti.first_group_dev)
1103 obj_size = ti.prev_group_obj_off;
1104 else if (i >= ti.nex_group_dev)
1105 obj_size = ti.next_group_obj_off;
1106 else if (i < ti.si.dev) /* dev within this group */
1107 obj_size = ti.si.obj_offset +
1108 ios->layout->stripe_unit - ti.si.unit_off;
1109 else if (i == ti.si.dev)
1110 obj_size = ti.si.obj_offset;
1111 else /* i > ti.dev */
1112 obj_size = ti.si.obj_offset - ti.si.unit_off;
1113
1114 size_attr->newsize = cpu_to_be64(obj_size);
1115 size_attr->attr = g_attr_logical_length;
1116 size_attr->attr.val_ptr = &size_attr->newsize;
1117
1118 ORE_DBGMSG2("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
1119 _LLU(oc->comps->obj.id), _LLU(obj_size), i);
1120 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
1121 &size_attr->attr);
1122 if (unlikely(ret))
1123 goto out;
1124 }
1125 ret = ore_io_execute(ios);
1126
1127out:
1128 kfree(size_attrs);
1129 ore_put_io_state(ios);
1130 return ret;
1131}
1132EXPORT_SYMBOL(ore_truncate);
1133
1134const struct osd_attr g_attr_logical_length = ATTR_DEF(
1135 OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
1136EXPORT_SYMBOL(g_attr_logical_length);