ore: RAID5 Write
[linux-3.10.git] / fs / exofs / ore.c
1 /*
2  * Copyright (C) 2005, 2006
3  * Avishay Traeger (avishay@gmail.com)
4  * Copyright (C) 2008, 2009
5  * Boaz Harrosh <bharrosh@panasas.com>
6  *
7  * This file is part of exofs.
8  *
9  * exofs is free software; you can redistribute it and/or modify
10  * it under the terms of the GNU General Public License as published by
11  * the Free Software Foundation.  Since it is based on ext2, and the only
12  * valid version of GPL for the Linux kernel is version 2, the only valid
13  * version of GPL for exofs is version 2.
14  *
15  * exofs is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
18  * GNU General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public License
21  * along with exofs; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
23  */
24
25 #include <linux/slab.h>
26 #include <asm/div64.h>
27 #include <linux/lcm.h>
28
29 #include "ore_raid.h"
30
31 MODULE_AUTHOR("Boaz Harrosh <bharrosh@panasas.com>");
32 MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
33 MODULE_LICENSE("GPL");
34
35 /* ore_verify_layout does a couple of things:
36  * 1. Given a minimum number of needed parameters fixes up the rest of the
37  *    members to be operatonals for the ore. The needed parameters are those
38  *    that are defined by the pnfs-objects layout STD.
39  * 2. Check to see if the current ore code actually supports these parameters
40  *    for example stripe_unit must be a multple of the system PAGE_SIZE,
41  *    and etc...
42  * 3. Cache some havily used calculations that will be needed by users.
43  */
44
45 enum { BIO_MAX_PAGES_KMALLOC =
46                 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),};
47
48 int ore_verify_layout(unsigned total_comps, struct ore_layout *layout)
49 {
50         u64 stripe_length;
51
52 /* FIXME: Only raid0 is supported for now. */
53         if (layout->raid_algorithm != PNFS_OSD_RAID_0) {
54                 ORE_ERR("Only RAID_0 for now\n");
55                 return -EINVAL;
56         }
57         if (0 != (layout->stripe_unit & ~PAGE_MASK)) {
58                 ORE_ERR("Stripe Unit(0x%llx)"
59                           " must be Multples of PAGE_SIZE(0x%lx)\n",
60                           _LLU(layout->stripe_unit), PAGE_SIZE);
61                 return -EINVAL;
62         }
63         if (layout->group_width) {
64                 if (!layout->group_depth) {
65                         ORE_ERR("group_depth == 0 && group_width != 0\n");
66                         return -EINVAL;
67                 }
68                 if (total_comps < (layout->group_width * layout->mirrors_p1)) {
69                         ORE_ERR("Data Map wrong, "
70                                 "numdevs=%d < group_width=%d * mirrors=%d\n",
71                                 total_comps, layout->group_width,
72                                 layout->mirrors_p1);
73                         return -EINVAL;
74                 }
75                 layout->group_count = total_comps / layout->mirrors_p1 /
76                                                 layout->group_width;
77         } else {
78                 if (layout->group_depth) {
79                         printk(KERN_NOTICE "Warning: group_depth ignored "
80                                 "group_width == 0 && group_depth == %lld\n",
81                                 _LLU(layout->group_depth));
82                 }
83                 layout->group_width = total_comps / layout->mirrors_p1;
84                 layout->group_depth = -1;
85                 layout->group_count = 1;
86         }
87
88         stripe_length = (u64)layout->group_width * layout->stripe_unit;
89         if (stripe_length >= (1ULL << 32)) {
90                 ORE_ERR("Stripe_length(0x%llx) >= 32bit is not supported\n",
91                         _LLU(stripe_length));
92                 return -EINVAL;
93         }
94
95         layout->max_io_length =
96                 (BIO_MAX_PAGES_KMALLOC * PAGE_SIZE - layout->stripe_unit) *
97                                                         layout->group_width;
98         if (layout->parity) {
99                 unsigned stripe_length =
100                                 (layout->group_width - layout->parity) *
101                                 layout->stripe_unit;
102
103                 layout->max_io_length /= stripe_length;
104                 layout->max_io_length *= stripe_length;
105         }
106         return 0;
107 }
108 EXPORT_SYMBOL(ore_verify_layout);
109
110 static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
111 {
112         return ios->oc->comps[index & ios->oc->single_comp].cred;
113 }
114
115 static struct osd_obj_id *_ios_obj(struct ore_io_state *ios, unsigned index)
116 {
117         return &ios->oc->comps[index & ios->oc->single_comp].obj;
118 }
119
120 static struct osd_dev *_ios_od(struct ore_io_state *ios, unsigned index)
121 {
122         ORE_DBGMSG2("oc->first_dev=%d oc->numdevs=%d i=%d oc->ods=%p\n",
123                     ios->oc->first_dev, ios->oc->numdevs, index,
124                     ios->oc->ods);
125
126         return ore_comp_dev(ios->oc, index);
127 }
128
129 int  _ore_get_io_state(struct ore_layout *layout,
130                         struct ore_components *oc, unsigned numdevs,
131                         unsigned sgs_per_dev, unsigned num_par_pages,
132                         struct ore_io_state **pios)
133 {
134         struct ore_io_state *ios;
135         struct page **pages;
136         struct osd_sg_entry *sgilist;
137         struct __alloc_all_io_state {
138                 struct ore_io_state ios;
139                 struct ore_per_dev_state per_dev[numdevs];
140                 union {
141                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
142                         struct page *pages[num_par_pages];
143                 };
144         } *_aios;
145
146         if (likely(sizeof(*_aios) <= PAGE_SIZE)) {
147                 _aios = kzalloc(sizeof(*_aios), GFP_KERNEL);
148                 if (unlikely(!_aios)) {
149                         ORE_DBGMSG("Failed kzalloc bytes=%zd\n",
150                                    sizeof(*_aios));
151                         *pios = NULL;
152                         return -ENOMEM;
153                 }
154                 pages = num_par_pages ? _aios->pages : NULL;
155                 sgilist = sgs_per_dev ? _aios->sglist : NULL;
156                 ios = &_aios->ios;
157         } else {
158                 struct __alloc_small_io_state {
159                         struct ore_io_state ios;
160                         struct ore_per_dev_state per_dev[numdevs];
161                 } *_aio_small;
162                 union __extra_part {
163                         struct osd_sg_entry sglist[sgs_per_dev * numdevs];
164                         struct page *pages[num_par_pages];
165                 } *extra_part;
166
167                 _aio_small = kzalloc(sizeof(*_aio_small), GFP_KERNEL);
168                 if (unlikely(!_aio_small)) {
169                         ORE_DBGMSG("Failed alloc first part bytes=%zd\n",
170                                    sizeof(*_aio_small));
171                         *pios = NULL;
172                         return -ENOMEM;
173                 }
174                 extra_part = kzalloc(sizeof(*extra_part), GFP_KERNEL);
175                 if (unlikely(!extra_part)) {
176                         ORE_DBGMSG("Failed alloc second part bytes=%zd\n",
177                                    sizeof(*extra_part));
178                         kfree(_aio_small);
179                         *pios = NULL;
180                         return -ENOMEM;
181                 }
182
183                 pages = num_par_pages ? extra_part->pages : NULL;
184                 sgilist = sgs_per_dev ? extra_part->sglist : NULL;
185                 /* In this case the per_dev[0].sgilist holds the pointer to
186                  * be freed
187                  */
188                 ios = &_aio_small->ios;
189                 ios->extra_part_alloc = true;
190         }
191
192         if (pages) {
193                 ios->parity_pages = pages;
194                 ios->max_par_pages = num_par_pages;
195         }
196         if (sgilist) {
197                 unsigned d;
198
199                 for (d = 0; d < numdevs; ++d) {
200                         ios->per_dev[d].sglist = sgilist;
201                         sgilist += sgs_per_dev;
202                 }
203                 ios->sgs_per_dev = sgs_per_dev;
204         }
205
206         ios->layout = layout;
207         ios->oc = oc;
208         *pios = ios;
209         return 0;
210 }
211
212 /* Allocate an io_state for only a single group of devices
213  *
214  * If a user needs to call ore_read/write() this version must be used becase it
215  * allocates extra stuff for striping and raid.
216  * The ore might decide to only IO less then @length bytes do to alignmets
217  * and constrains as follows:
218  * - The IO cannot cross group boundary.
219  * - In raid5/6 The end of the IO must align at end of a stripe eg.
220  *   (@offset + @length) % strip_size == 0. Or the complete range is within a
221  *   single stripe.
222  * - Memory condition only permitted a shorter IO. (A user can use @length=~0
223  *   And check the returned ios->length for max_io_size.)
224  *
225  * The caller must check returned ios->length (and/or ios->nr_pages) and
226  * re-issue these pages that fall outside of ios->length
227  */
228 int  ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
229                       bool is_reading, u64 offset, u64 length,
230                       struct ore_io_state **pios)
231 {
232         struct ore_io_state *ios;
233         unsigned numdevs = layout->group_width * layout->mirrors_p1;
234         unsigned sgs_per_dev = 0, max_par_pages = 0;
235         int ret;
236
237         if (layout->parity && length) {
238                 unsigned data_devs = layout->group_width - layout->parity;
239                 unsigned stripe_size = layout->stripe_unit * data_devs;
240                 unsigned pages_in_unit = layout->stripe_unit / PAGE_SIZE;
241                 u32 remainder;
242                 u64 num_stripes;
243                 u64 num_raid_units;
244
245                 num_stripes = div_u64_rem(length, stripe_size, &remainder);
246                 if (remainder)
247                         ++num_stripes;
248
249                 num_raid_units =  num_stripes * layout->parity;
250
251                 if (is_reading) {
252                         /* For reads add per_dev sglist array */
253                         /* TODO: Raid 6 we need twice more. Actually:
254                         *         num_stripes / LCMdP(W,P);
255                         *         if (W%P != 0) num_stripes *= parity;
256                         */
257
258                         /* first/last seg is split */
259                         num_raid_units += layout->group_width;
260                         sgs_per_dev = div_u64(num_raid_units, data_devs);
261                 } else {
262                         /* For Writes add parity pages array. */
263                         max_par_pages = num_raid_units * pages_in_unit *
264                                                 sizeof(struct page *);
265                 }
266         }
267
268         ret = _ore_get_io_state(layout, oc, numdevs, sgs_per_dev, max_par_pages,
269                                 pios);
270         if (unlikely(ret))
271                 return ret;
272
273         ios = *pios;
274         ios->reading = is_reading;
275         ios->offset = offset;
276
277         if (length) {
278                 ore_calc_stripe_info(layout, offset, length, &ios->si);
279                 ios->length = ios->si.length;
280                 ios->nr_pages = (ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
281                 if (layout->parity)
282                         _ore_post_alloc_raid_stuff(ios);
283         }
284
285         return 0;
286 }
287 EXPORT_SYMBOL(ore_get_rw_state);
288
289 /* Allocate an io_state for all the devices in the comps array
290  *
291  * This version of io_state allocation is used mostly by create/remove
292  * and trunc where we currently need all the devices. The only wastful
293  * bit is the read/write_attributes with no IO. Those sites should
294  * be converted to use ore_get_rw_state() with length=0
295  */
296 int  ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
297                       struct ore_io_state **pios)
298 {
299         return _ore_get_io_state(layout, oc, oc->numdevs, 0, 0, pios);
300 }
301 EXPORT_SYMBOL(ore_get_io_state);
302
303 void ore_put_io_state(struct ore_io_state *ios)
304 {
305         if (ios) {
306                 unsigned i;
307
308                 for (i = 0; i < ios->numdevs; i++) {
309                         struct ore_per_dev_state *per_dev = &ios->per_dev[i];
310
311                         if (per_dev->or)
312                                 osd_end_request(per_dev->or);
313                         if (per_dev->bio)
314                                 bio_put(per_dev->bio);
315                 }
316
317                 _ore_free_raid_stuff(ios);
318                 kfree(ios);
319         }
320 }
321 EXPORT_SYMBOL(ore_put_io_state);
322
323 static void _sync_done(struct ore_io_state *ios, void *p)
324 {
325         struct completion *waiting = p;
326
327         complete(waiting);
328 }
329
330 static void _last_io(struct kref *kref)
331 {
332         struct ore_io_state *ios = container_of(
333                                         kref, struct ore_io_state, kref);
334
335         ios->done(ios, ios->private);
336 }
337
338 static void _done_io(struct osd_request *or, void *p)
339 {
340         struct ore_io_state *ios = p;
341
342         kref_put(&ios->kref, _last_io);
343 }
344
345 int ore_io_execute(struct ore_io_state *ios)
346 {
347         DECLARE_COMPLETION_ONSTACK(wait);
348         bool sync = (ios->done == NULL);
349         int i, ret;
350
351         if (sync) {
352                 ios->done = _sync_done;
353                 ios->private = &wait;
354         }
355
356         for (i = 0; i < ios->numdevs; i++) {
357                 struct osd_request *or = ios->per_dev[i].or;
358                 if (unlikely(!or))
359                         continue;
360
361                 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
362                 if (unlikely(ret)) {
363                         ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
364                                      ret);
365                         return ret;
366                 }
367         }
368
369         kref_init(&ios->kref);
370
371         for (i = 0; i < ios->numdevs; i++) {
372                 struct osd_request *or = ios->per_dev[i].or;
373                 if (unlikely(!or))
374                         continue;
375
376                 kref_get(&ios->kref);
377                 osd_execute_request_async(or, _done_io, ios);
378         }
379
380         kref_put(&ios->kref, _last_io);
381         ret = 0;
382
383         if (sync) {
384                 wait_for_completion(&wait);
385                 ret = ore_check_io(ios, NULL);
386         }
387         return ret;
388 }
389
390 static void _clear_bio(struct bio *bio)
391 {
392         struct bio_vec *bv;
393         unsigned i;
394
395         __bio_for_each_segment(bv, bio, i, 0) {
396                 unsigned this_count = bv->bv_len;
397
398                 if (likely(PAGE_SIZE == this_count))
399                         clear_highpage(bv->bv_page);
400                 else
401                         zero_user(bv->bv_page, bv->bv_offset, this_count);
402         }
403 }
404
405 int ore_check_io(struct ore_io_state *ios, ore_on_dev_error on_dev_error)
406 {
407         enum osd_err_priority acumulated_osd_err = 0;
408         int acumulated_lin_err = 0;
409         int i;
410
411         for (i = 0; i < ios->numdevs; i++) {
412                 struct osd_sense_info osi;
413                 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
414                 struct osd_request *or = per_dev->or;
415                 int ret;
416
417                 if (unlikely(!or))
418                         continue;
419
420                 ret = osd_req_decode_sense(or, &osi);
421                 if (likely(!ret))
422                         continue;
423
424                 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
425                         /* start read offset passed endof file */
426                         _clear_bio(per_dev->bio);
427                         ORE_DBGMSG("start read offset passed end of file "
428                                 "offset=0x%llx, length=0x%llx\n",
429                                 _LLU(per_dev->offset),
430                                 _LLU(per_dev->length));
431
432                         continue; /* we recovered */
433                 }
434
435                 if (on_dev_error) {
436                         u64 residual = ios->reading ?
437                                         or->in.residual : or->out.residual;
438                         u64 offset = (ios->offset + ios->length) - residual;
439                         struct ore_dev *od = ios->oc->ods[
440                                         per_dev->dev - ios->oc->first_dev];
441
442                         on_dev_error(ios, od, per_dev->dev, osi.osd_err_pri,
443                                      offset, residual);
444                 }
445                 if (osi.osd_err_pri >= acumulated_osd_err) {
446                         acumulated_osd_err = osi.osd_err_pri;
447                         acumulated_lin_err = ret;
448                 }
449         }
450
451         return acumulated_lin_err;
452 }
453 EXPORT_SYMBOL(ore_check_io);
454
455 /*
456  * L - logical offset into the file
457  *
458  * D - number of Data devices
459  *      D = group_width - parity
460  *
461  * U - The number of bytes in a stripe within a group
462  *      U =  stripe_unit * D
463  *
464  * T - The number of bytes striped within a group of component objects
465  *     (before advancing to the next group)
466  *      T = U * group_depth
467  *
468  * S - The number of bytes striped across all component objects
469  *     before the pattern repeats
470  *      S = T * group_count
471  *
472  * M - The "major" (i.e., across all components) cycle number
473  *      M = L / S
474  *
475  * G - Counts the groups from the beginning of the major cycle
476  *      G = (L - (M * S)) / T   [or (L % S) / T]
477  *
478  * H - The byte offset within the group
479  *      H = (L - (M * S)) % T   [or (L % S) % T]
480  *
481  * N - The "minor" (i.e., across the group) stripe number
482  *      N = H / U
483  *
484  * C - The component index coresponding to L
485  *
486  *      C = (H - (N * U)) / stripe_unit + G * D
487  *      [or (L % U) / stripe_unit + G * D]
488  *
489  * O - The component offset coresponding to L
490  *      O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
491  *
492  * LCMdP – Parity cycle: Lowest Common Multiple of group_width, parity
493  *          divide by parity
494  *      LCMdP = lcm(group_width, parity) / parity
495  *
496  * R - The parity Rotation stripe
497  *     (Note parity cycle always starts at a group's boundary)
498  *      R = N % LCMdP
499  *
500  * I = the first parity device index
501  *      I = (group_width + group_width - R*parity - parity) % group_width
502  *
503  * Craid - The component index Rotated
504  *      Craid = (group_width + C - R*parity) % group_width
505  *      (We add the group_width to avoid negative numbers modulo math)
506  */
507 void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
508                           u64 length, struct ore_striping_info *si)
509 {
510         u32     stripe_unit = layout->stripe_unit;
511         u32     group_width = layout->group_width;
512         u64     group_depth = layout->group_depth;
513         u32     parity      = layout->parity;
514
515         u32     D = group_width - parity;
516         u32     U = D * stripe_unit;
517         u64     T = U * group_depth;
518         u64     S = T * layout->group_count;
519         u64     M = div64_u64(file_offset, S);
520
521         /*
522         G = (L - (M * S)) / T
523         H = (L - (M * S)) % T
524         */
525         u64     LmodS = file_offset - M * S;
526         u32     G = div64_u64(LmodS, T);
527         u64     H = LmodS - G * T;
528
529         u32     N = div_u64(H, U);
530
531         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
532         u32     C = (u32)(H - (N * U)) / stripe_unit + G * group_width;
533
534         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
535
536         si->obj_offset = si->unit_off + (N * stripe_unit) +
537                                   (M * group_depth * stripe_unit);
538
539         if (parity) {
540                 u32 LCMdP = lcm(group_width, parity) / parity;
541                 /* R     = N % LCMdP; */
542                 u32 RxP   = (N % LCMdP) * parity;
543                 u32 first_dev = C - C % group_width;
544
545                 si->par_dev = (group_width + group_width - parity - RxP) %
546                               group_width + first_dev;
547                 si->dev = (group_width + C - RxP) % group_width + first_dev;
548                 si->bytes_in_stripe = U;
549                 si->first_stripe_start = M * S + G * T + N * U;
550         } else {
551                 /* Make the math correct see _prepare_one_group */
552                 si->par_dev = group_width;
553                 si->dev = C;
554         }
555
556         si->dev *= layout->mirrors_p1;
557         si->par_dev *= layout->mirrors_p1;
558         si->offset = file_offset;
559         si->length = T - H;
560         if (si->length > length)
561                 si->length = length;
562         si->M = M;
563 }
564 EXPORT_SYMBOL(ore_calc_stripe_info);
565
566 int _ore_add_stripe_unit(struct ore_io_state *ios,  unsigned *cur_pg,
567                          unsigned pgbase, struct page **pages,
568                          struct ore_per_dev_state *per_dev, int cur_len)
569 {
570         unsigned pg = *cur_pg;
571         struct request_queue *q =
572                         osd_request_queue(_ios_od(ios, per_dev->dev));
573         unsigned len = cur_len;
574         int ret;
575
576         if (per_dev->bio == NULL) {
577                 unsigned pages_in_stripe = ios->layout->group_width *
578                                         (ios->layout->stripe_unit / PAGE_SIZE);
579                 unsigned nr_pages = ios->nr_pages * ios->layout->group_width /
580                                         (ios->layout->group_width -
581                                          ios->layout->parity);
582                 unsigned bio_size = (nr_pages + pages_in_stripe) /
583                                         ios->layout->group_width;
584
585                 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
586                 if (unlikely(!per_dev->bio)) {
587                         ORE_DBGMSG("Failed to allocate BIO size=%u\n",
588                                      bio_size);
589                         ret = -ENOMEM;
590                         goto out;
591                 }
592         }
593
594         while (cur_len > 0) {
595                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
596                 unsigned added_len;
597
598                 cur_len -= pglen;
599
600                 added_len = bio_add_pc_page(q, per_dev->bio, pages[pg],
601                                             pglen, pgbase);
602                 if (unlikely(pglen != added_len)) {
603                         ORE_DBGMSG("Failed bio_add_pc_page bi_vcnt=%u\n",
604                                    per_dev->bio->bi_vcnt);
605                         ret = -ENOMEM;
606                         goto out;
607                 }
608                 _add_stripe_page(ios->sp2d, &ios->si, pages[pg]);
609
610                 pgbase = 0;
611                 ++pg;
612         }
613         BUG_ON(cur_len);
614
615         per_dev->length += len;
616         *cur_pg = pg;
617         ret = 0;
618 out:    /* we fail the complete unit on an error eg don't advance
619          * per_dev->length and cur_pg. This means that we might have a bigger
620          * bio than the CDB requested length (per_dev->length). That's fine
621          * only the oposite is fatal.
622          */
623         return ret;
624 }
625
626 static int _prepare_for_striping(struct ore_io_state *ios)
627 {
628         struct ore_striping_info *si = &ios->si;
629         unsigned stripe_unit = ios->layout->stripe_unit;
630         unsigned mirrors_p1 = ios->layout->mirrors_p1;
631         unsigned group_width = ios->layout->group_width;
632         unsigned devs_in_group = group_width * mirrors_p1;
633         unsigned dev = si->dev;
634         unsigned first_dev = dev - (dev % devs_in_group);
635         unsigned dev_order;
636         unsigned cur_pg = ios->pages_consumed;
637         u64 length = ios->length;
638         int ret = 0;
639
640         if (!ios->pages) {
641                 ios->numdevs = ios->layout->mirrors_p1;
642                 return 0;
643         }
644
645         BUG_ON(length > si->length);
646
647         dev_order = _dev_order(devs_in_group, mirrors_p1, si->par_dev, dev);
648         si->cur_comp = dev_order;
649         si->cur_pg = si->unit_off / PAGE_SIZE;
650
651         while (length) {
652                 unsigned comp = dev - first_dev;
653                 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
654                 unsigned cur_len, page_off = 0;
655
656                 if (!per_dev->length) {
657                         per_dev->dev = dev;
658                         if (dev == si->dev) {
659                                 WARN_ON(dev == si->par_dev);
660                                 per_dev->offset = si->obj_offset;
661                                 cur_len = stripe_unit - si->unit_off;
662                                 page_off = si->unit_off & ~PAGE_MASK;
663                                 BUG_ON(page_off && (page_off != ios->pgbase));
664                         } else {
665                                 if (si->cur_comp > dev_order)
666                                         per_dev->offset =
667                                                 si->obj_offset - si->unit_off;
668                                 else /* si->cur_comp < dev_order */
669                                         per_dev->offset =
670                                                 si->obj_offset + stripe_unit -
671                                                                    si->unit_off;
672                                 cur_len = stripe_unit;
673                         }
674                 } else {
675                         cur_len = stripe_unit;
676                 }
677                 if (cur_len >= length)
678                         cur_len = length;
679
680                 ret = _ore_add_stripe_unit(ios, &cur_pg, page_off, ios->pages,
681                                            per_dev, cur_len);
682                 if (unlikely(ret))
683                         goto out;
684
685                 dev += mirrors_p1;
686                 dev = (dev % devs_in_group) + first_dev;
687
688                 length -= cur_len;
689
690                 si->cur_comp = (si->cur_comp + 1) % group_width;
691                 if (unlikely((dev == si->par_dev) || (!length && ios->sp2d))) {
692                         if (!length && ios->sp2d) {
693                                 /* If we are writing and this is the very last
694                                  * stripe. then operate on parity dev.
695                                  */
696                                 dev = si->par_dev;
697                         }
698                         if (ios->sp2d)
699                                 /* In writes cur_len just means if it's the
700                                  * last one. See _ore_add_parity_unit.
701                                  */
702                                 cur_len = length;
703                         per_dev = &ios->per_dev[dev - first_dev];
704                         if (!per_dev->length) {
705                                 /* Only/always the parity unit of the first
706                                  * stripe will be empty. So this is a chance to
707                                  * initialize the per_dev info.
708                                  */
709                                 per_dev->dev = dev;
710                                 per_dev->offset = si->obj_offset - si->unit_off;
711                         }
712
713                         ret = _ore_add_parity_unit(ios, si, per_dev, cur_len);
714                         if (unlikely(ret))
715                                         goto out;
716
717                         /* Rotate next par_dev backwards with wraping */
718                         si->par_dev = (devs_in_group + si->par_dev -
719                                        ios->layout->parity * mirrors_p1) %
720                                       devs_in_group + first_dev;
721                         /* Next stripe, start fresh */
722                         si->cur_comp = 0;
723                         si->cur_pg = 0;
724                 }
725         }
726 out:
727         ios->numdevs = devs_in_group;
728         ios->pages_consumed = cur_pg;
729         if (unlikely(ret)) {
730                 if (length == ios->length)
731                         return ret;
732                 else
733                         ios->length -= length;
734         }
735         return 0;
736 }
737
738 int ore_create(struct ore_io_state *ios)
739 {
740         int i, ret;
741
742         for (i = 0; i < ios->oc->numdevs; i++) {
743                 struct osd_request *or;
744
745                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
746                 if (unlikely(!or)) {
747                         ORE_ERR("%s: osd_start_request failed\n", __func__);
748                         ret = -ENOMEM;
749                         goto out;
750                 }
751                 ios->per_dev[i].or = or;
752                 ios->numdevs++;
753
754                 osd_req_create_object(or, _ios_obj(ios, i));
755         }
756         ret = ore_io_execute(ios);
757
758 out:
759         return ret;
760 }
761 EXPORT_SYMBOL(ore_create);
762
763 int ore_remove(struct ore_io_state *ios)
764 {
765         int i, ret;
766
767         for (i = 0; i < ios->oc->numdevs; i++) {
768                 struct osd_request *or;
769
770                 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
771                 if (unlikely(!or)) {
772                         ORE_ERR("%s: osd_start_request failed\n", __func__);
773                         ret = -ENOMEM;
774                         goto out;
775                 }
776                 ios->per_dev[i].or = or;
777                 ios->numdevs++;
778
779                 osd_req_remove_object(or, _ios_obj(ios, i));
780         }
781         ret = ore_io_execute(ios);
782
783 out:
784         return ret;
785 }
786 EXPORT_SYMBOL(ore_remove);
787
788 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
789 {
790         struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
791         unsigned dev = ios->per_dev[cur_comp].dev;
792         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
793         int ret = 0;
794
795         if (ios->pages && !master_dev->length)
796                 return 0; /* Just an empty slot */
797
798         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
799                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
800                 struct osd_request *or;
801
802                 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
803                 if (unlikely(!or)) {
804                         ORE_ERR("%s: osd_start_request failed\n", __func__);
805                         ret = -ENOMEM;
806                         goto out;
807                 }
808                 per_dev->or = or;
809
810                 if (ios->pages) {
811                         struct bio *bio;
812
813                         if (per_dev != master_dev) {
814                                 bio = bio_kmalloc(GFP_KERNEL,
815                                                   master_dev->bio->bi_max_vecs);
816                                 if (unlikely(!bio)) {
817                                         ORE_DBGMSG(
818                                               "Failed to allocate BIO size=%u\n",
819                                               master_dev->bio->bi_max_vecs);
820                                         ret = -ENOMEM;
821                                         goto out;
822                                 }
823
824                                 __bio_clone(bio, master_dev->bio);
825                                 bio->bi_bdev = NULL;
826                                 bio->bi_next = NULL;
827                                 per_dev->offset = master_dev->offset;
828                                 per_dev->length = master_dev->length;
829                                 per_dev->bio =  bio;
830                                 per_dev->dev = dev;
831                         } else {
832                                 bio = master_dev->bio;
833                                 /* FIXME: bio_set_dir() */
834                                 bio->bi_rw |= REQ_WRITE;
835                         }
836
837                         osd_req_write(or, _ios_obj(ios, dev), per_dev->offset,
838                                       bio, per_dev->length);
839                         ORE_DBGMSG("write(0x%llx) offset=0x%llx "
840                                       "length=0x%llx dev=%d\n",
841                                      _LLU(_ios_obj(ios, dev)->id),
842                                      _LLU(per_dev->offset),
843                                      _LLU(per_dev->length), dev);
844                 } else if (ios->kern_buff) {
845                         per_dev->offset = ios->si.obj_offset;
846                         per_dev->dev = ios->si.dev + dev;
847
848                         /* no cross device without page array */
849                         BUG_ON((ios->layout->group_width > 1) &&
850                                (ios->si.unit_off + ios->length >
851                                 ios->layout->stripe_unit));
852
853                         ret = osd_req_write_kern(or, _ios_obj(ios, per_dev->dev),
854                                                  per_dev->offset,
855                                                  ios->kern_buff, ios->length);
856                         if (unlikely(ret))
857                                 goto out;
858                         ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
859                                       "length=0x%llx dev=%d\n",
860                                      _LLU(_ios_obj(ios, dev)->id),
861                                      _LLU(per_dev->offset),
862                                      _LLU(ios->length), per_dev->dev);
863                 } else {
864                         osd_req_set_attributes(or, _ios_obj(ios, dev));
865                         ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
866                                      _LLU(_ios_obj(ios, dev)->id),
867                                      ios->out_attr_len, dev);
868                 }
869
870                 if (ios->out_attr)
871                         osd_req_add_set_attr_list(or, ios->out_attr,
872                                                   ios->out_attr_len);
873
874                 if (ios->in_attr)
875                         osd_req_add_get_attr_list(or, ios->in_attr,
876                                                   ios->in_attr_len);
877         }
878
879 out:
880         return ret;
881 }
882
883 int ore_write(struct ore_io_state *ios)
884 {
885         int i;
886         int ret;
887
888         if (unlikely(ios->sp2d && !ios->r4w)) {
889                 /* A library is attempting a RAID-write without providing
890                  * a pages lock interface.
891                  */
892                 WARN_ON_ONCE(1);
893                 return -ENOTSUPP;
894         }
895
896         ret = _prepare_for_striping(ios);
897         if (unlikely(ret))
898                 return ret;
899
900         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
901                 ret = _write_mirror(ios, i);
902                 if (unlikely(ret))
903                         return ret;
904         }
905
906         ret = ore_io_execute(ios);
907         return ret;
908 }
909 EXPORT_SYMBOL(ore_write);
910
911 int _ore_read_mirror(struct ore_io_state *ios, unsigned cur_comp)
912 {
913         struct osd_request *or;
914         struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
915         struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
916         unsigned first_dev = (unsigned)obj->id;
917
918         if (ios->pages && !per_dev->length)
919                 return 0; /* Just an empty slot */
920
921         first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
922         or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
923         if (unlikely(!or)) {
924                 ORE_ERR("%s: osd_start_request failed\n", __func__);
925                 return -ENOMEM;
926         }
927         per_dev->or = or;
928
929         if (ios->pages) {
930                 if (per_dev->cur_sg) {
931                         /* finalize the last sg_entry */
932                         _ore_add_sg_seg(per_dev, 0, false);
933                         if (unlikely(!per_dev->cur_sg))
934                                 return 0; /* Skip parity only device */
935
936                         osd_req_read_sg(or, obj, per_dev->bio,
937                                         per_dev->sglist, per_dev->cur_sg);
938                 } else {
939                         /* The no raid case */
940                         osd_req_read(or, obj, per_dev->offset,
941                                      per_dev->bio, per_dev->length);
942                 }
943
944                 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
945                              " dev=%d sg_len=%d\n", _LLU(obj->id),
946                              _LLU(per_dev->offset), _LLU(per_dev->length),
947                              first_dev, per_dev->cur_sg);
948         } else {
949                 BUG_ON(ios->kern_buff);
950
951                 osd_req_get_attributes(or, obj);
952                 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
953                               _LLU(obj->id),
954                               ios->in_attr_len, first_dev);
955         }
956         if (ios->out_attr)
957                 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
958
959         if (ios->in_attr)
960                 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
961
962         return 0;
963 }
964
965 int ore_read(struct ore_io_state *ios)
966 {
967         int i;
968         int ret;
969
970         ret = _prepare_for_striping(ios);
971         if (unlikely(ret))
972                 return ret;
973
974         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
975                 ret = _ore_read_mirror(ios, i);
976                 if (unlikely(ret))
977                         return ret;
978         }
979
980         ret = ore_io_execute(ios);
981         return ret;
982 }
983 EXPORT_SYMBOL(ore_read);
984
985 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
986 {
987         struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
988         void *iter = NULL;
989         int nelem;
990
991         do {
992                 nelem = 1;
993                 osd_req_decode_get_attr_list(ios->per_dev[0].or,
994                                              &cur_attr, &nelem, &iter);
995                 if ((cur_attr.attr_page == attr->attr_page) &&
996                     (cur_attr.attr_id == attr->attr_id)) {
997                         attr->len = cur_attr.len;
998                         attr->val_ptr = cur_attr.val_ptr;
999                         return 0;
1000                 }
1001         } while (iter);
1002
1003         return -EIO;
1004 }
1005 EXPORT_SYMBOL(extract_attr_from_ios);
1006
1007 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
1008                              struct osd_attr *attr)
1009 {
1010         int last_comp = cur_comp + ios->layout->mirrors_p1;
1011
1012         for (; cur_comp < last_comp; ++cur_comp) {
1013                 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
1014                 struct osd_request *or;
1015
1016                 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
1017                 if (unlikely(!or)) {
1018                         ORE_ERR("%s: osd_start_request failed\n", __func__);
1019                         return -ENOMEM;
1020                 }
1021                 per_dev->or = or;
1022
1023                 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
1024                 osd_req_add_set_attr_list(or, attr, 1);
1025         }
1026
1027         return 0;
1028 }
1029
1030 struct _trunc_info {
1031         struct ore_striping_info si;
1032         u64 prev_group_obj_off;
1033         u64 next_group_obj_off;
1034
1035         unsigned first_group_dev;
1036         unsigned nex_group_dev;
1037 };
1038
1039 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
1040                              struct _trunc_info *ti)
1041 {
1042         unsigned stripe_unit = layout->stripe_unit;
1043
1044         ore_calc_stripe_info(layout, file_offset, 0, &ti->si);
1045
1046         ti->prev_group_obj_off = ti->si.M * stripe_unit;
1047         ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
1048
1049         ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
1050         ti->nex_group_dev = ti->first_group_dev + layout->group_width;
1051 }
1052
1053 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
1054                    u64 size)
1055 {
1056         struct ore_io_state *ios;
1057         struct exofs_trunc_attr {
1058                 struct osd_attr attr;
1059                 __be64 newsize;
1060         } *size_attrs;
1061         struct _trunc_info ti;
1062         int i, ret;
1063
1064         ret = ore_get_io_state(layout, oc, &ios);
1065         if (unlikely(ret))
1066                 return ret;
1067
1068         _calc_trunk_info(ios->layout, size, &ti);
1069
1070         size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
1071                              GFP_KERNEL);
1072         if (unlikely(!size_attrs)) {
1073                 ret = -ENOMEM;
1074                 goto out;
1075         }
1076
1077         ios->numdevs = ios->oc->numdevs;
1078
1079         for (i = 0; i < ios->numdevs; ++i) {
1080                 struct exofs_trunc_attr *size_attr = &size_attrs[i];
1081                 u64 obj_size;
1082
1083                 if (i < ti.first_group_dev)
1084                         obj_size = ti.prev_group_obj_off;
1085                 else if (i >= ti.nex_group_dev)
1086                         obj_size = ti.next_group_obj_off;
1087                 else if (i < ti.si.dev) /* dev within this group */
1088                         obj_size = ti.si.obj_offset +
1089                                       ios->layout->stripe_unit - ti.si.unit_off;
1090                 else if (i == ti.si.dev)
1091                         obj_size = ti.si.obj_offset;
1092                 else /* i > ti.dev */
1093                         obj_size = ti.si.obj_offset - ti.si.unit_off;
1094
1095                 size_attr->newsize = cpu_to_be64(obj_size);
1096                 size_attr->attr = g_attr_logical_length;
1097                 size_attr->attr.val_ptr = &size_attr->newsize;
1098
1099                 ORE_DBGMSG("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
1100                              _LLU(oc->comps->obj.id), _LLU(obj_size), i);
1101                 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
1102                                         &size_attr->attr);
1103                 if (unlikely(ret))
1104                         goto out;
1105         }
1106         ret = ore_io_execute(ios);
1107
1108 out:
1109         kfree(size_attrs);
1110         ore_put_io_state(ios);
1111         return ret;
1112 }
1113 EXPORT_SYMBOL(ore_truncate);
1114
1115 const struct osd_attr g_attr_logical_length = ATTR_DEF(
1116         OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
1117 EXPORT_SYMBOL(g_attr_logical_length);