2 * Copyright (C) 2005, 2006
3 * Avishay Traeger (avishay@gmail.com)
4 * Copyright (C) 2008, 2009
5 * Boaz Harrosh <bharrosh@panasas.com>
7 * This file is part of exofs.
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.
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.
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
25 #include <linux/slab.h>
26 #include <asm/div64.h>
28 #include <scsi/osd_ore.h>
30 #define ORE_ERR(fmt, a...) printk(KERN_ERR "ore: " fmt, ##a)
32 #ifdef CONFIG_EXOFS_DEBUG
33 #define ORE_DBGMSG(fmt, a...) \
34 printk(KERN_NOTICE "ore @%s:%d: " fmt, __func__, __LINE__, ##a)
36 #define ORE_DBGMSG(fmt, a...) \
37 do { if (0) printk(fmt, ##a); } while (0)
40 /* u64 has problems with printk this will cast it to unsigned long long */
41 #define _LLU(x) (unsigned long long)(x)
43 #define ORE_DBGMSG2(M...) do {} while (0)
44 /* #define ORE_DBGMSG2 ORE_DBGMSG */
46 MODULE_AUTHOR("Boaz Harrosh <bharrosh@panasas.com>");
47 MODULE_DESCRIPTION("Objects Raid Engine ore.ko");
48 MODULE_LICENSE("GPL");
50 static void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
51 struct ore_striping_info *si);
53 static u8 *_ios_cred(struct ore_io_state *ios, unsigned index)
55 return ios->oc->comps[index & ios->oc->single_comp].cred;
58 static struct osd_obj_id *_ios_obj(struct ore_io_state *ios, unsigned index)
60 return &ios->oc->comps[index & ios->oc->single_comp].obj;
63 static struct osd_dev *_ios_od(struct ore_io_state *ios, unsigned index)
65 return ore_comp_dev(ios->oc, index);
68 static int _get_io_state(struct ore_layout *layout,
69 struct ore_components *oc, unsigned numdevs,
70 struct ore_io_state **pios)
72 struct ore_io_state *ios;
74 /*TODO: Maybe use kmem_cach per sbi of size
75 * exofs_io_state_size(layout->s_numdevs)
77 ios = kzalloc(ore_io_state_size(numdevs), GFP_KERNEL);
79 ORE_DBGMSG("Failed kzalloc bytes=%d\n",
80 ore_io_state_size(numdevs));
91 /* Allocate an io_state for only a single group of devices
93 * If a user needs to call ore_read/write() this version must be used becase it
94 * allocates extra stuff for striping and raid.
95 * The ore might decide to only IO less then @length bytes do to alignmets
96 * and constrains as follows:
97 * - The IO cannot cross group boundary.
98 * - In raid5/6 The end of the IO must align at end of a stripe eg.
99 * (@offset + @length) % strip_size == 0. Or the complete range is within a
101 * - Memory condition only permitted a shorter IO. (A user can use @length=~0
102 * And check the returned ios->length for max_io_size.)
104 * The caller must check returned ios->length (and/or ios->nr_pages) and
105 * re-issue these pages that fall outside of ios->length
107 int ore_get_rw_state(struct ore_layout *layout, struct ore_components *oc,
108 bool is_reading, u64 offset, u64 length,
109 struct ore_io_state **pios)
111 struct ore_io_state *ios;
112 unsigned numdevs = layout->group_width * layout->mirrors_p1;
115 ret = _get_io_state(layout, oc, numdevs, pios);
120 ios->reading = is_reading;
121 ios->offset = offset;
124 ore_calc_stripe_info(layout, offset, &ios->si);
125 ios->length = (length <= ios->si.group_length) ? length :
126 ios->si.group_length;
127 ios->nr_pages = (ios->length + PAGE_SIZE - 1) / PAGE_SIZE;
132 EXPORT_SYMBOL(ore_get_rw_state);
134 /* Allocate an io_state for all the devices in the comps array
136 * This version of io_state allocation is used mostly by create/remove
137 * and trunc where we currently need all the devices. The only wastful
138 * bit is the read/write_attributes with no IO. Those sites should
139 * be converted to use ore_get_rw_state() with length=0
141 int ore_get_io_state(struct ore_layout *layout, struct ore_components *oc,
142 struct ore_io_state **pios)
144 return _get_io_state(layout, oc, oc->numdevs, pios);
146 EXPORT_SYMBOL(ore_get_io_state);
148 void ore_put_io_state(struct ore_io_state *ios)
153 for (i = 0; i < ios->numdevs; i++) {
154 struct ore_per_dev_state *per_dev = &ios->per_dev[i];
157 osd_end_request(per_dev->or);
159 bio_put(per_dev->bio);
165 EXPORT_SYMBOL(ore_put_io_state);
167 static void _sync_done(struct ore_io_state *ios, void *p)
169 struct completion *waiting = p;
174 static void _last_io(struct kref *kref)
176 struct ore_io_state *ios = container_of(
177 kref, struct ore_io_state, kref);
179 ios->done(ios, ios->private);
182 static void _done_io(struct osd_request *or, void *p)
184 struct ore_io_state *ios = p;
186 kref_put(&ios->kref, _last_io);
189 static int ore_io_execute(struct ore_io_state *ios)
191 DECLARE_COMPLETION_ONSTACK(wait);
192 bool sync = (ios->done == NULL);
196 ios->done = _sync_done;
197 ios->private = &wait;
200 for (i = 0; i < ios->numdevs; i++) {
201 struct osd_request *or = ios->per_dev[i].or;
205 ret = osd_finalize_request(or, 0, _ios_cred(ios, i), NULL);
207 ORE_DBGMSG("Failed to osd_finalize_request() => %d\n",
213 kref_init(&ios->kref);
215 for (i = 0; i < ios->numdevs; i++) {
216 struct osd_request *or = ios->per_dev[i].or;
220 kref_get(&ios->kref);
221 osd_execute_request_async(or, _done_io, ios);
224 kref_put(&ios->kref, _last_io);
228 wait_for_completion(&wait);
229 ret = ore_check_io(ios, NULL);
234 static void _clear_bio(struct bio *bio)
239 __bio_for_each_segment(bv, bio, i, 0) {
240 unsigned this_count = bv->bv_len;
242 if (likely(PAGE_SIZE == this_count))
243 clear_highpage(bv->bv_page);
245 zero_user(bv->bv_page, bv->bv_offset, this_count);
249 int ore_check_io(struct ore_io_state *ios, u64 *resid)
251 enum osd_err_priority acumulated_osd_err = 0;
252 int acumulated_lin_err = 0;
255 for (i = 0; i < ios->numdevs; i++) {
256 struct osd_sense_info osi;
257 struct osd_request *or = ios->per_dev[i].or;
263 ret = osd_req_decode_sense(or, &osi);
267 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
268 /* start read offset passed endof file */
269 _clear_bio(ios->per_dev[i].bio);
270 ORE_DBGMSG("start read offset passed end of file "
271 "offset=0x%llx, length=0x%llx\n",
272 _LLU(ios->per_dev[i].offset),
273 _LLU(ios->per_dev[i].length));
275 continue; /* we recovered */
278 if (osi.osd_err_pri >= acumulated_osd_err) {
279 acumulated_osd_err = osi.osd_err_pri;
280 acumulated_lin_err = ret;
284 /* TODO: raid specific residual calculations */
286 if (likely(!acumulated_lin_err))
289 *resid = ios->length;
292 return acumulated_lin_err;
294 EXPORT_SYMBOL(ore_check_io);
297 * L - logical offset into the file
299 * U - The number of bytes in a stripe within a group
301 * U = stripe_unit * group_width
303 * T - The number of bytes striped within a group of component objects
304 * (before advancing to the next group)
306 * T = stripe_unit * group_width * group_depth
308 * S - The number of bytes striped across all component objects
309 * before the pattern repeats
311 * S = stripe_unit * group_width * group_depth * group_count
313 * M - The "major" (i.e., across all components) stripe number
317 * G - Counts the groups from the beginning of the major stripe
319 * G = (L - (M * S)) / T [or (L % S) / T]
321 * H - The byte offset within the group
323 * H = (L - (M * S)) % T [or (L % S) % T]
325 * N - The "minor" (i.e., across the group) stripe number
329 * C - The component index coresponding to L
331 * C = (H - (N * U)) / stripe_unit + G * group_width
332 * [or (L % U) / stripe_unit + G * group_width]
334 * O - The component offset coresponding to L
336 * O = L % stripe_unit + N * stripe_unit + M * group_depth * stripe_unit
338 static void ore_calc_stripe_info(struct ore_layout *layout, u64 file_offset,
339 struct ore_striping_info *si)
341 u32 stripe_unit = layout->stripe_unit;
342 u32 group_width = layout->group_width;
343 u64 group_depth = layout->group_depth;
345 u32 U = stripe_unit * group_width;
346 u64 T = U * group_depth;
347 u64 S = T * layout->group_count;
348 u64 M = div64_u64(file_offset, S);
351 G = (L - (M * S)) / T
352 H = (L - (M * S)) % T
354 u64 LmodS = file_offset - M * S;
355 u32 G = div64_u64(LmodS, T);
356 u64 H = LmodS - G * T;
358 u32 N = div_u64(H, U);
360 /* "H - (N * U)" is just "H % U" so it's bound to u32 */
361 si->dev = (u32)(H - (N * U)) / stripe_unit + G * group_width;
362 si->dev *= layout->mirrors_p1;
364 div_u64_rem(file_offset, stripe_unit, &si->unit_off);
366 si->obj_offset = si->unit_off + (N * stripe_unit) +
367 (M * group_depth * stripe_unit);
369 si->group_length = T - H;
373 static int _add_stripe_unit(struct ore_io_state *ios, unsigned *cur_pg,
374 unsigned pgbase, struct ore_per_dev_state *per_dev,
377 unsigned pg = *cur_pg;
378 struct request_queue *q =
379 osd_request_queue(_ios_od(ios, per_dev->dev));
380 unsigned len = cur_len;
383 if (per_dev->bio == NULL) {
384 unsigned pages_in_stripe = ios->layout->group_width *
385 (ios->layout->stripe_unit / PAGE_SIZE);
386 unsigned bio_size = (ios->nr_pages + pages_in_stripe) /
387 ios->layout->group_width;
389 per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
390 if (unlikely(!per_dev->bio)) {
391 ORE_DBGMSG("Failed to allocate BIO size=%u\n",
398 while (cur_len > 0) {
399 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
402 BUG_ON(ios->nr_pages <= pg);
405 added_len = bio_add_pc_page(q, per_dev->bio, ios->pages[pg],
407 if (unlikely(pglen != added_len)) {
416 per_dev->length += len;
419 out: /* we fail the complete unit on an error eg don't advance
420 * per_dev->length and cur_pg. This means that we might have a bigger
421 * bio than the CDB requested length (per_dev->length). That's fine
422 * only the oposite is fatal.
427 static int _prepare_for_striping(struct ore_io_state *ios)
429 struct ore_striping_info *si = &ios->si;
430 unsigned stripe_unit = ios->layout->stripe_unit;
431 unsigned mirrors_p1 = ios->layout->mirrors_p1;
432 unsigned devs_in_group = ios->layout->group_width * mirrors_p1;
433 unsigned dev = si->dev;
434 unsigned first_dev = dev - (dev % devs_in_group);
435 unsigned cur_pg = ios->pages_consumed;
436 u64 length = ios->length;
440 ios->numdevs = ios->layout->mirrors_p1;
444 BUG_ON(length > si->group_length);
447 unsigned comp = dev - first_dev;
448 struct ore_per_dev_state *per_dev = &ios->per_dev[comp];
449 unsigned cur_len, page_off = 0;
451 if (!per_dev->length) {
454 per_dev->offset = si->obj_offset + stripe_unit -
456 cur_len = stripe_unit;
457 } else if (dev == si->dev) {
458 per_dev->offset = si->obj_offset;
459 cur_len = stripe_unit - si->unit_off;
460 page_off = si->unit_off & ~PAGE_MASK;
461 BUG_ON(page_off && (page_off != ios->pgbase));
462 } else { /* dev > si->dev */
463 per_dev->offset = si->obj_offset - si->unit_off;
464 cur_len = stripe_unit;
467 cur_len = stripe_unit;
469 if (cur_len >= length)
472 ret = _add_stripe_unit(ios, &cur_pg, page_off , per_dev,
478 dev = (dev % devs_in_group) + first_dev;
483 ios->numdevs = devs_in_group;
484 ios->pages_consumed = cur_pg;
486 if (length == ios->length)
489 ios->length -= length;
494 int ore_create(struct ore_io_state *ios)
498 for (i = 0; i < ios->oc->numdevs; i++) {
499 struct osd_request *or;
501 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
503 ORE_ERR("%s: osd_start_request failed\n", __func__);
507 ios->per_dev[i].or = or;
510 osd_req_create_object(or, _ios_obj(ios, i));
512 ret = ore_io_execute(ios);
517 EXPORT_SYMBOL(ore_create);
519 int ore_remove(struct ore_io_state *ios)
523 for (i = 0; i < ios->oc->numdevs; i++) {
524 struct osd_request *or;
526 or = osd_start_request(_ios_od(ios, i), GFP_KERNEL);
528 ORE_ERR("%s: osd_start_request failed\n", __func__);
532 ios->per_dev[i].or = or;
535 osd_req_remove_object(or, _ios_obj(ios, i));
537 ret = ore_io_execute(ios);
542 EXPORT_SYMBOL(ore_remove);
544 static int _write_mirror(struct ore_io_state *ios, int cur_comp)
546 struct ore_per_dev_state *master_dev = &ios->per_dev[cur_comp];
547 unsigned dev = ios->per_dev[cur_comp].dev;
548 unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
551 if (ios->pages && !master_dev->length)
552 return 0; /* Just an empty slot */
554 for (; cur_comp < last_comp; ++cur_comp, ++dev) {
555 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
556 struct osd_request *or;
558 or = osd_start_request(_ios_od(ios, dev), GFP_KERNEL);
560 ORE_ERR("%s: osd_start_request failed\n", __func__);
569 if (per_dev != master_dev) {
570 bio = bio_kmalloc(GFP_KERNEL,
571 master_dev->bio->bi_max_vecs);
572 if (unlikely(!bio)) {
574 "Failed to allocate BIO size=%u\n",
575 master_dev->bio->bi_max_vecs);
580 __bio_clone(bio, master_dev->bio);
583 per_dev->offset = master_dev->offset;
584 per_dev->length = master_dev->length;
588 bio = master_dev->bio;
589 /* FIXME: bio_set_dir() */
590 bio->bi_rw |= REQ_WRITE;
593 osd_req_write(or, _ios_obj(ios, dev), per_dev->offset,
594 bio, per_dev->length);
595 ORE_DBGMSG("write(0x%llx) offset=0x%llx "
596 "length=0x%llx dev=%d\n",
597 _LLU(_ios_obj(ios, dev)->id),
598 _LLU(per_dev->offset),
599 _LLU(per_dev->length), dev);
600 } else if (ios->kern_buff) {
601 per_dev->offset = ios->si.obj_offset;
602 per_dev->dev = ios->si.dev + dev;
604 /* no cross device without page array */
605 BUG_ON((ios->layout->group_width > 1) &&
606 (ios->si.unit_off + ios->length >
607 ios->layout->stripe_unit));
609 ret = osd_req_write_kern(or, _ios_obj(ios, per_dev->dev),
611 ios->kern_buff, ios->length);
614 ORE_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
615 "length=0x%llx dev=%d\n",
616 _LLU(_ios_obj(ios, dev)->id),
617 _LLU(per_dev->offset),
618 _LLU(ios->length), per_dev->dev);
620 osd_req_set_attributes(or, _ios_obj(ios, dev));
621 ORE_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
622 _LLU(_ios_obj(ios, dev)->id),
623 ios->out_attr_len, dev);
627 osd_req_add_set_attr_list(or, ios->out_attr,
631 osd_req_add_get_attr_list(or, ios->in_attr,
639 int ore_write(struct ore_io_state *ios)
644 ret = _prepare_for_striping(ios);
648 for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
649 ret = _write_mirror(ios, i);
654 ret = ore_io_execute(ios);
657 EXPORT_SYMBOL(ore_write);
659 static int _read_mirror(struct ore_io_state *ios, unsigned cur_comp)
661 struct osd_request *or;
662 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
663 struct osd_obj_id *obj = _ios_obj(ios, cur_comp);
664 unsigned first_dev = (unsigned)obj->id;
666 if (ios->pages && !per_dev->length)
667 return 0; /* Just an empty slot */
669 first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
670 or = osd_start_request(_ios_od(ios, first_dev), GFP_KERNEL);
672 ORE_ERR("%s: osd_start_request failed\n", __func__);
678 osd_req_read(or, obj, per_dev->offset,
679 per_dev->bio, per_dev->length);
680 ORE_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
681 " dev=%d\n", _LLU(obj->id),
682 _LLU(per_dev->offset), _LLU(per_dev->length),
685 BUG_ON(ios->kern_buff);
687 osd_req_get_attributes(or, obj);
688 ORE_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
690 ios->in_attr_len, first_dev);
693 osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
696 osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
701 int ore_read(struct ore_io_state *ios)
706 ret = _prepare_for_striping(ios);
710 for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
711 ret = _read_mirror(ios, i);
716 ret = ore_io_execute(ios);
719 EXPORT_SYMBOL(ore_read);
721 int extract_attr_from_ios(struct ore_io_state *ios, struct osd_attr *attr)
723 struct osd_attr cur_attr = {.attr_page = 0}; /* start with zeros */
729 osd_req_decode_get_attr_list(ios->per_dev[0].or,
730 &cur_attr, &nelem, &iter);
731 if ((cur_attr.attr_page == attr->attr_page) &&
732 (cur_attr.attr_id == attr->attr_id)) {
733 attr->len = cur_attr.len;
734 attr->val_ptr = cur_attr.val_ptr;
741 EXPORT_SYMBOL(extract_attr_from_ios);
743 static int _truncate_mirrors(struct ore_io_state *ios, unsigned cur_comp,
744 struct osd_attr *attr)
746 int last_comp = cur_comp + ios->layout->mirrors_p1;
748 for (; cur_comp < last_comp; ++cur_comp) {
749 struct ore_per_dev_state *per_dev = &ios->per_dev[cur_comp];
750 struct osd_request *or;
752 or = osd_start_request(_ios_od(ios, cur_comp), GFP_KERNEL);
754 ORE_ERR("%s: osd_start_request failed\n", __func__);
759 osd_req_set_attributes(or, _ios_obj(ios, cur_comp));
760 osd_req_add_set_attr_list(or, attr, 1);
767 struct ore_striping_info si;
768 u64 prev_group_obj_off;
769 u64 next_group_obj_off;
771 unsigned first_group_dev;
772 unsigned nex_group_dev;
775 static void _calc_trunk_info(struct ore_layout *layout, u64 file_offset,
776 struct _trunc_info *ti)
778 unsigned stripe_unit = layout->stripe_unit;
780 ore_calc_stripe_info(layout, file_offset, &ti->si);
782 ti->prev_group_obj_off = ti->si.M * stripe_unit;
783 ti->next_group_obj_off = ti->si.M ? (ti->si.M - 1) * stripe_unit : 0;
785 ti->first_group_dev = ti->si.dev - (ti->si.dev % layout->group_width);
786 ti->nex_group_dev = ti->first_group_dev + layout->group_width;
789 int ore_truncate(struct ore_layout *layout, struct ore_components *oc,
792 struct ore_io_state *ios;
793 struct exofs_trunc_attr {
794 struct osd_attr attr;
797 struct _trunc_info ti;
800 ret = ore_get_io_state(layout, oc, &ios);
804 _calc_trunk_info(ios->layout, size, &ti);
806 size_attrs = kcalloc(ios->oc->numdevs, sizeof(*size_attrs),
808 if (unlikely(!size_attrs)) {
813 ios->numdevs = ios->oc->numdevs;
815 for (i = 0; i < ios->numdevs; ++i) {
816 struct exofs_trunc_attr *size_attr = &size_attrs[i];
819 if (i < ti.first_group_dev)
820 obj_size = ti.prev_group_obj_off;
821 else if (i >= ti.nex_group_dev)
822 obj_size = ti.next_group_obj_off;
823 else if (i < ti.si.dev) /* dev within this group */
824 obj_size = ti.si.obj_offset +
825 ios->layout->stripe_unit - ti.si.unit_off;
826 else if (i == ti.si.dev)
827 obj_size = ti.si.obj_offset;
828 else /* i > ti.dev */
829 obj_size = ti.si.obj_offset - ti.si.unit_off;
831 size_attr->newsize = cpu_to_be64(obj_size);
832 size_attr->attr = g_attr_logical_length;
833 size_attr->attr.val_ptr = &size_attr->newsize;
835 ORE_DBGMSG("trunc(0x%llx) obj_offset=0x%llx dev=%d\n",
836 _LLU(oc->comps->obj.id), _LLU(obj_size), i);
837 ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
842 ret = ore_io_execute(ios);
846 ore_put_io_state(ios);
849 EXPORT_SYMBOL(ore_truncate);
851 const struct osd_attr g_attr_logical_length = ATTR_DEF(
852 OSD_APAGE_OBJECT_INFORMATION, OSD_ATTR_OI_LOGICAL_LENGTH, 8);
853 EXPORT_SYMBOL(g_attr_logical_length);