pnfs-obj: osd raid engine read/write implementation
[pandora-kernel.git] / fs / nfs / objlayout / objio_osd.c
1 /*
2  *  pNFS Objects layout implementation over open-osd initiator library
3  *
4  *  Copyright (C) 2009 Panasas Inc. [year of first publication]
5  *  All rights reserved.
6  *
7  *  Benny Halevy <bhalevy@panasas.com>
8  *  Boaz Harrosh <bharrosh@panasas.com>
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2
12  *  See the file COPYING included with this distribution for more details.
13  *
14  *  Redistribution and use in source and binary forms, with or without
15  *  modification, are permitted provided that the following conditions
16  *  are met:
17  *
18  *  1. Redistributions of source code must retain the above copyright
19  *     notice, this list of conditions and the following disclaimer.
20  *  2. Redistributions in binary form must reproduce the above copyright
21  *     notice, this list of conditions and the following disclaimer in the
22  *     documentation and/or other materials provided with the distribution.
23  *  3. Neither the name of the Panasas company nor the names of its
24  *     contributors may be used to endorse or promote products derived
25  *     from this software without specific prior written permission.
26  *
27  *  THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
28  *  WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
29  *  MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
30  *  DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31  *  FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32  *  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33  *  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
34  *  BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
35  *  LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
36  *  NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
37  *  SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38  */
39
40 #include <linux/module.h>
41 #include <scsi/osd_initiator.h>
42
43 #include "objlayout.h"
44
45 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
46
47 #define _LLU(x) ((unsigned long long)x)
48
49 enum { BIO_MAX_PAGES_KMALLOC =
50                 (PAGE_SIZE - sizeof(struct bio)) / sizeof(struct bio_vec),
51 };
52
53 struct objio_dev_ent {
54         struct nfs4_deviceid_node id_node;
55         struct osd_dev *od;
56 };
57
58 static void
59 objio_free_deviceid_node(struct nfs4_deviceid_node *d)
60 {
61         struct objio_dev_ent *de = container_of(d, struct objio_dev_ent, id_node);
62
63         dprintk("%s: free od=%p\n", __func__, de->od);
64         osduld_put_device(de->od);
65         kfree(de);
66 }
67
68 static struct objio_dev_ent *_dev_list_find(const struct nfs_server *nfss,
69         const struct nfs4_deviceid *d_id)
70 {
71         struct nfs4_deviceid_node *d;
72         struct objio_dev_ent *de;
73
74         d = nfs4_find_get_deviceid(nfss->pnfs_curr_ld, nfss->nfs_client, d_id);
75         if (!d)
76                 return NULL;
77
78         de = container_of(d, struct objio_dev_ent, id_node);
79         return de;
80 }
81
82 static struct objio_dev_ent *
83 _dev_list_add(const struct nfs_server *nfss,
84         const struct nfs4_deviceid *d_id, struct osd_dev *od,
85         gfp_t gfp_flags)
86 {
87         struct nfs4_deviceid_node *d;
88         struct objio_dev_ent *de = kzalloc(sizeof(*de), gfp_flags);
89         struct objio_dev_ent *n;
90
91         if (!de) {
92                 dprintk("%s: -ENOMEM od=%p\n", __func__, od);
93                 return NULL;
94         }
95
96         dprintk("%s: Adding od=%p\n", __func__, od);
97         nfs4_init_deviceid_node(&de->id_node,
98                                 nfss->pnfs_curr_ld,
99                                 nfss->nfs_client,
100                                 d_id);
101         de->od = od;
102
103         d = nfs4_insert_deviceid_node(&de->id_node);
104         n = container_of(d, struct objio_dev_ent, id_node);
105         if (n != de) {
106                 dprintk("%s: Race with other n->od=%p\n", __func__, n->od);
107                 objio_free_deviceid_node(&de->id_node);
108                 de = n;
109         }
110
111         atomic_inc(&de->id_node.ref);
112         return de;
113 }
114
115 struct caps_buffers {
116         u8 caps_key[OSD_CRYPTO_KEYID_SIZE];
117         u8 creds[OSD_CAP_LEN];
118 };
119
120 struct objio_segment {
121         struct pnfs_layout_segment lseg;
122
123         struct pnfs_osd_object_cred *comps;
124
125         unsigned mirrors_p1;
126         unsigned stripe_unit;
127         unsigned group_width;   /* Data stripe_units without integrity comps */
128         u64 group_depth;
129         unsigned group_count;
130
131         unsigned comps_index;
132         unsigned num_comps;
133         /* variable length */
134         struct objio_dev_ent *ods[];
135 };
136
137 static inline struct objio_segment *
138 OBJIO_LSEG(struct pnfs_layout_segment *lseg)
139 {
140         return container_of(lseg, struct objio_segment, lseg);
141 }
142
143 struct objio_state;
144 typedef ssize_t (*objio_done_fn)(struct objio_state *ios);
145
146 struct objio_state {
147         /* Generic layer */
148         struct objlayout_io_state ol_state;
149
150         struct objio_segment *layout;
151
152         struct kref kref;
153         objio_done_fn done;
154         void *private;
155
156         unsigned long length;
157         unsigned numdevs; /* Actually used devs in this IO */
158         /* A per-device variable array of size numdevs */
159         struct _objio_per_comp {
160                 struct bio *bio;
161                 struct osd_request *or;
162                 unsigned long length;
163                 u64 offset;
164                 unsigned dev;
165         } per_dev[];
166 };
167
168 /* Send and wait for a get_device_info of devices in the layout,
169    then look them up with the osd_initiator library */
170 static struct objio_dev_ent *_device_lookup(struct pnfs_layout_hdr *pnfslay,
171                                 struct objio_segment *objio_seg, unsigned comp,
172                                 gfp_t gfp_flags)
173 {
174         struct pnfs_osd_deviceaddr *deviceaddr;
175         struct nfs4_deviceid *d_id;
176         struct objio_dev_ent *ode;
177         struct osd_dev *od;
178         struct osd_dev_info odi;
179         int err;
180
181         d_id = &objio_seg->comps[comp].oc_object_id.oid_device_id;
182
183         ode = _dev_list_find(NFS_SERVER(pnfslay->plh_inode), d_id);
184         if (ode)
185                 return ode;
186
187         err = objlayout_get_deviceinfo(pnfslay, d_id, &deviceaddr, gfp_flags);
188         if (unlikely(err)) {
189                 dprintk("%s: objlayout_get_deviceinfo dev(%llx:%llx) =>%d\n",
190                         __func__, _DEVID_LO(d_id), _DEVID_HI(d_id), err);
191                 return ERR_PTR(err);
192         }
193
194         odi.systemid_len = deviceaddr->oda_systemid.len;
195         if (odi.systemid_len > sizeof(odi.systemid)) {
196                 err = -EINVAL;
197                 goto out;
198         } else if (odi.systemid_len)
199                 memcpy(odi.systemid, deviceaddr->oda_systemid.data,
200                        odi.systemid_len);
201         odi.osdname_len  = deviceaddr->oda_osdname.len;
202         odi.osdname      = (u8 *)deviceaddr->oda_osdname.data;
203
204         if (!odi.osdname_len && !odi.systemid_len) {
205                 dprintk("%s: !odi.osdname_len && !odi.systemid_len\n",
206                         __func__);
207                 err = -ENODEV;
208                 goto out;
209         }
210
211         od = osduld_info_lookup(&odi);
212         if (unlikely(IS_ERR(od))) {
213                 err = PTR_ERR(od);
214                 dprintk("%s: osduld_info_lookup => %d\n", __func__, err);
215                 goto out;
216         }
217
218         ode = _dev_list_add(NFS_SERVER(pnfslay->plh_inode), d_id, od,
219                             gfp_flags);
220
221 out:
222         dprintk("%s: return=%d\n", __func__, err);
223         objlayout_put_deviceinfo(deviceaddr);
224         return err ? ERR_PTR(err) : ode;
225 }
226
227 static int objio_devices_lookup(struct pnfs_layout_hdr *pnfslay,
228         struct objio_segment *objio_seg,
229         gfp_t gfp_flags)
230 {
231         unsigned i;
232         int err;
233
234         /* lookup all devices */
235         for (i = 0; i < objio_seg->num_comps; i++) {
236                 struct objio_dev_ent *ode;
237
238                 ode = _device_lookup(pnfslay, objio_seg, i, gfp_flags);
239                 if (unlikely(IS_ERR(ode))) {
240                         err = PTR_ERR(ode);
241                         goto out;
242                 }
243                 objio_seg->ods[i] = ode;
244         }
245         err = 0;
246
247 out:
248         dprintk("%s: return=%d\n", __func__, err);
249         return err;
250 }
251
252 static int _verify_data_map(struct pnfs_osd_layout *layout)
253 {
254         struct pnfs_osd_data_map *data_map = &layout->olo_map;
255         u64 stripe_length;
256         u32 group_width;
257
258 /* FIXME: Only raid0 for now. if not go through MDS */
259         if (data_map->odm_raid_algorithm != PNFS_OSD_RAID_0) {
260                 printk(KERN_ERR "Only RAID_0 for now\n");
261                 return -ENOTSUPP;
262         }
263         if (0 != (data_map->odm_num_comps % (data_map->odm_mirror_cnt + 1))) {
264                 printk(KERN_ERR "Data Map wrong, num_comps=%u mirrors=%u\n",
265                           data_map->odm_num_comps, data_map->odm_mirror_cnt);
266                 return -EINVAL;
267         }
268
269         if (data_map->odm_group_width)
270                 group_width = data_map->odm_group_width;
271         else
272                 group_width = data_map->odm_num_comps /
273                                                 (data_map->odm_mirror_cnt + 1);
274
275         stripe_length = (u64)data_map->odm_stripe_unit * group_width;
276         if (stripe_length >= (1ULL << 32)) {
277                 printk(KERN_ERR "Total Stripe length(0x%llx)"
278                           " >= 32bit is not supported\n", _LLU(stripe_length));
279                 return -ENOTSUPP;
280         }
281
282         if (0 != (data_map->odm_stripe_unit & ~PAGE_MASK)) {
283                 printk(KERN_ERR "Stripe Unit(0x%llx)"
284                           " must be Multples of PAGE_SIZE(0x%lx)\n",
285                           _LLU(data_map->odm_stripe_unit), PAGE_SIZE);
286                 return -ENOTSUPP;
287         }
288
289         return 0;
290 }
291
292 static void copy_single_comp(struct pnfs_osd_object_cred *cur_comp,
293                              struct pnfs_osd_object_cred *src_comp,
294                              struct caps_buffers *caps_p)
295 {
296         WARN_ON(src_comp->oc_cap_key.cred_len > sizeof(caps_p->caps_key));
297         WARN_ON(src_comp->oc_cap.cred_len > sizeof(caps_p->creds));
298
299         *cur_comp = *src_comp;
300
301         memcpy(caps_p->caps_key, src_comp->oc_cap_key.cred,
302                sizeof(caps_p->caps_key));
303         cur_comp->oc_cap_key.cred = caps_p->caps_key;
304
305         memcpy(caps_p->creds, src_comp->oc_cap.cred,
306                sizeof(caps_p->creds));
307         cur_comp->oc_cap.cred = caps_p->creds;
308 }
309
310 int objio_alloc_lseg(struct pnfs_layout_segment **outp,
311         struct pnfs_layout_hdr *pnfslay,
312         struct pnfs_layout_range *range,
313         struct xdr_stream *xdr,
314         gfp_t gfp_flags)
315 {
316         struct objio_segment *objio_seg;
317         struct pnfs_osd_xdr_decode_layout_iter iter;
318         struct pnfs_osd_layout layout;
319         struct pnfs_osd_object_cred *cur_comp, src_comp;
320         struct caps_buffers *caps_p;
321         int err;
322
323         err = pnfs_osd_xdr_decode_layout_map(&layout, &iter, xdr);
324         if (unlikely(err))
325                 return err;
326
327         err = _verify_data_map(&layout);
328         if (unlikely(err))
329                 return err;
330
331         objio_seg = kzalloc(sizeof(*objio_seg) +
332                             sizeof(objio_seg->ods[0]) * layout.olo_num_comps +
333                             sizeof(*objio_seg->comps) * layout.olo_num_comps +
334                             sizeof(struct caps_buffers) * layout.olo_num_comps,
335                             gfp_flags);
336         if (!objio_seg)
337                 return -ENOMEM;
338
339         objio_seg->comps = (void *)(objio_seg->ods + layout.olo_num_comps);
340         cur_comp = objio_seg->comps;
341         caps_p = (void *)(cur_comp + layout.olo_num_comps);
342         while (pnfs_osd_xdr_decode_layout_comp(&src_comp, &iter, xdr, &err))
343                 copy_single_comp(cur_comp++, &src_comp, caps_p++);
344         if (unlikely(err))
345                 goto err;
346
347         objio_seg->num_comps = layout.olo_num_comps;
348         objio_seg->comps_index = layout.olo_comps_index;
349         err = objio_devices_lookup(pnfslay, objio_seg, gfp_flags);
350         if (err)
351                 goto err;
352
353         objio_seg->mirrors_p1 = layout.olo_map.odm_mirror_cnt + 1;
354         objio_seg->stripe_unit = layout.olo_map.odm_stripe_unit;
355         if (layout.olo_map.odm_group_width) {
356                 objio_seg->group_width = layout.olo_map.odm_group_width;
357                 objio_seg->group_depth = layout.olo_map.odm_group_depth;
358                 objio_seg->group_count = layout.olo_map.odm_num_comps /
359                                                 objio_seg->mirrors_p1 /
360                                                 objio_seg->group_width;
361         } else {
362                 objio_seg->group_width = layout.olo_map.odm_num_comps /
363                                                 objio_seg->mirrors_p1;
364                 objio_seg->group_depth = -1;
365                 objio_seg->group_count = 1;
366         }
367
368         *outp = &objio_seg->lseg;
369         return 0;
370
371 err:
372         kfree(objio_seg);
373         dprintk("%s: Error: return %d\n", __func__, err);
374         *outp = NULL;
375         return err;
376 }
377
378 void objio_free_lseg(struct pnfs_layout_segment *lseg)
379 {
380         int i;
381         struct objio_segment *objio_seg = OBJIO_LSEG(lseg);
382
383         for (i = 0; i < objio_seg->num_comps; i++) {
384                 if (!objio_seg->ods[i])
385                         break;
386                 nfs4_put_deviceid_node(&objio_seg->ods[i]->id_node);
387         }
388         kfree(objio_seg);
389 }
390
391 int objio_alloc_io_state(struct pnfs_layout_segment *lseg,
392                          struct objlayout_io_state **outp,
393                          gfp_t gfp_flags)
394 {
395         struct objio_segment *objio_seg = OBJIO_LSEG(lseg);
396         struct objio_state *ios;
397         const unsigned first_size = sizeof(*ios) +
398                                 objio_seg->num_comps * sizeof(ios->per_dev[0]);
399
400         ios = kzalloc(first_size, gfp_flags);
401         if (unlikely(!ios))
402                 return -ENOMEM;
403
404         ios->layout = objio_seg;
405
406         *outp = &ios->ol_state;
407         return 0;
408 }
409
410 void objio_free_io_state(struct objlayout_io_state *ol_state)
411 {
412         struct objio_state *ios = container_of(ol_state, struct objio_state,
413                                                ol_state);
414
415         kfree(ios);
416 }
417
418 static void _clear_bio(struct bio *bio)
419 {
420         struct bio_vec *bv;
421         unsigned i;
422
423         __bio_for_each_segment(bv, bio, i, 0) {
424                 unsigned this_count = bv->bv_len;
425
426                 if (likely(PAGE_SIZE == this_count))
427                         clear_highpage(bv->bv_page);
428                 else
429                         zero_user(bv->bv_page, bv->bv_offset, this_count);
430         }
431 }
432
433 static int _io_check(struct objio_state *ios, bool is_write)
434 {
435         enum osd_err_priority oep = OSD_ERR_PRI_NO_ERROR;
436         int lin_ret = 0;
437         int i;
438
439         for (i = 0; i <  ios->numdevs; i++) {
440                 struct osd_sense_info osi;
441                 struct osd_request *or = ios->per_dev[i].or;
442                 unsigned dev;
443                 int ret;
444
445                 if (!or)
446                         continue;
447
448                 ret = osd_req_decode_sense(or, &osi);
449                 if (likely(!ret))
450                         continue;
451
452                 if (OSD_ERR_PRI_CLEAR_PAGES == osi.osd_err_pri) {
453                         /* start read offset passed endof file */
454                         BUG_ON(is_write);
455                         _clear_bio(ios->per_dev[i].bio);
456                         dprintk("%s: start read offset passed end of file "
457                                 "offset=0x%llx, length=0x%lx\n", __func__,
458                                 _LLU(ios->per_dev[i].offset),
459                                 ios->per_dev[i].length);
460
461                         continue; /* we recovered */
462                 }
463                 dev = ios->per_dev[i].dev;
464
465                 if (osi.osd_err_pri >= oep) {
466                         oep = osi.osd_err_pri;
467                         lin_ret = ret;
468                 }
469         }
470
471         return lin_ret;
472 }
473
474 /*
475  * Common IO state helpers.
476  */
477 static void _io_free(struct objio_state *ios)
478 {
479         unsigned i;
480
481         for (i = 0; i < ios->numdevs; i++) {
482                 struct _objio_per_comp *per_dev = &ios->per_dev[i];
483
484                 if (per_dev->or) {
485                         osd_end_request(per_dev->or);
486                         per_dev->or = NULL;
487                 }
488
489                 if (per_dev->bio) {
490                         bio_put(per_dev->bio);
491                         per_dev->bio = NULL;
492                 }
493         }
494 }
495
496 struct osd_dev *_io_od(struct objio_state *ios, unsigned dev)
497 {
498         unsigned min_dev = ios->layout->comps_index;
499         unsigned max_dev = min_dev + ios->layout->num_comps;
500
501         BUG_ON(dev < min_dev || max_dev <= dev);
502         return ios->layout->ods[dev - min_dev]->od;
503 }
504
505 struct _striping_info {
506         u64 obj_offset;
507         u64 group_length;
508         unsigned dev;
509         unsigned unit_off;
510 };
511
512 static void _calc_stripe_info(struct objio_state *ios, u64 file_offset,
513                               struct _striping_info *si)
514 {
515         u32     stripe_unit = ios->layout->stripe_unit;
516         u32     group_width = ios->layout->group_width;
517         u64     group_depth = ios->layout->group_depth;
518         u32     U = stripe_unit * group_width;
519
520         u64     T = U * group_depth;
521         u64     S = T * ios->layout->group_count;
522         u64     M = div64_u64(file_offset, S);
523
524         /*
525         G = (L - (M * S)) / T
526         H = (L - (M * S)) % T
527         */
528         u64     LmodU = file_offset - M * S;
529         u32     G = div64_u64(LmodU, T);
530         u64     H = LmodU - G * T;
531
532         u32     N = div_u64(H, U);
533
534         div_u64_rem(file_offset, stripe_unit, &si->unit_off);
535         si->obj_offset = si->unit_off + (N * stripe_unit) +
536                                   (M * group_depth * stripe_unit);
537
538         /* "H - (N * U)" is just "H % U" so it's bound to u32 */
539         si->dev = (u32)(H - (N * U)) / stripe_unit + G * group_width;
540         si->dev *= ios->layout->mirrors_p1;
541
542         si->group_length = T - H;
543 }
544
545 static int _add_stripe_unit(struct objio_state *ios,  unsigned *cur_pg,
546                 unsigned pgbase, struct _objio_per_comp *per_dev, int cur_len,
547                 gfp_t gfp_flags)
548 {
549         unsigned pg = *cur_pg;
550         struct request_queue *q =
551                         osd_request_queue(_io_od(ios, per_dev->dev));
552
553         per_dev->length += cur_len;
554
555         if (per_dev->bio == NULL) {
556                 unsigned stripes = ios->layout->num_comps /
557                                                      ios->layout->mirrors_p1;
558                 unsigned pages_in_stripe = stripes *
559                                       (ios->layout->stripe_unit / PAGE_SIZE);
560                 unsigned bio_size = (ios->ol_state.nr_pages + pages_in_stripe) /
561                                     stripes;
562
563                 if (BIO_MAX_PAGES_KMALLOC < bio_size)
564                         bio_size = BIO_MAX_PAGES_KMALLOC;
565
566                 per_dev->bio = bio_kmalloc(gfp_flags, bio_size);
567                 if (unlikely(!per_dev->bio)) {
568                         dprintk("Faild to allocate BIO size=%u\n", bio_size);
569                         return -ENOMEM;
570                 }
571         }
572
573         while (cur_len > 0) {
574                 unsigned pglen = min_t(unsigned, PAGE_SIZE - pgbase, cur_len);
575                 unsigned added_len;
576
577                 BUG_ON(ios->ol_state.nr_pages <= pg);
578                 cur_len -= pglen;
579
580                 added_len = bio_add_pc_page(q, per_dev->bio,
581                                         ios->ol_state.pages[pg], pglen, pgbase);
582                 if (unlikely(pglen != added_len))
583                         return -ENOMEM;
584                 pgbase = 0;
585                 ++pg;
586         }
587         BUG_ON(cur_len);
588
589         *cur_pg = pg;
590         return 0;
591 }
592
593 static int _prepare_one_group(struct objio_state *ios, u64 length,
594                               struct _striping_info *si, unsigned *last_pg,
595                               gfp_t gfp_flags)
596 {
597         unsigned stripe_unit = ios->layout->stripe_unit;
598         unsigned mirrors_p1 = ios->layout->mirrors_p1;
599         unsigned devs_in_group = ios->layout->group_width * mirrors_p1;
600         unsigned dev = si->dev;
601         unsigned first_dev = dev - (dev % devs_in_group);
602         unsigned max_comp = ios->numdevs ? ios->numdevs - mirrors_p1 : 0;
603         unsigned cur_pg = *last_pg;
604         int ret = 0;
605
606         while (length) {
607                 struct _objio_per_comp *per_dev = &ios->per_dev[dev];
608                 unsigned cur_len, page_off = 0;
609
610                 if (!per_dev->length) {
611                         per_dev->dev = dev;
612                         if (dev < si->dev) {
613                                 per_dev->offset = si->obj_offset + stripe_unit -
614                                                                    si->unit_off;
615                                 cur_len = stripe_unit;
616                         } else if (dev == si->dev) {
617                                 per_dev->offset = si->obj_offset;
618                                 cur_len = stripe_unit - si->unit_off;
619                                 page_off = si->unit_off & ~PAGE_MASK;
620                                 BUG_ON(page_off &&
621                                       (page_off != ios->ol_state.pgbase));
622                         } else { /* dev > si->dev */
623                                 per_dev->offset = si->obj_offset - si->unit_off;
624                                 cur_len = stripe_unit;
625                         }
626
627                         if (max_comp < dev)
628                                 max_comp = dev;
629                 } else {
630                         cur_len = stripe_unit;
631                 }
632                 if (cur_len >= length)
633                         cur_len = length;
634
635                 ret = _add_stripe_unit(ios, &cur_pg, page_off , per_dev,
636                                        cur_len, gfp_flags);
637                 if (unlikely(ret))
638                         goto out;
639
640                 dev += mirrors_p1;
641                 dev = (dev % devs_in_group) + first_dev;
642
643                 length -= cur_len;
644                 ios->length += cur_len;
645         }
646 out:
647         ios->numdevs = max_comp + mirrors_p1;
648         *last_pg = cur_pg;
649         return ret;
650 }
651
652 static int _io_rw_pagelist(struct objio_state *ios, gfp_t gfp_flags)
653 {
654         u64 length = ios->ol_state.count;
655         u64 offset = ios->ol_state.offset;
656         struct _striping_info si;
657         unsigned last_pg = 0;
658         int ret = 0;
659
660         while (length) {
661                 _calc_stripe_info(ios, offset, &si);
662
663                 if (length < si.group_length)
664                         si.group_length = length;
665
666                 ret = _prepare_one_group(ios, si.group_length, &si, &last_pg, gfp_flags);
667                 if (unlikely(ret))
668                         goto out;
669
670                 offset += si.group_length;
671                 length -= si.group_length;
672         }
673
674 out:
675         if (!ios->length)
676                 return ret;
677
678         return 0;
679 }
680
681 static ssize_t _sync_done(struct objio_state *ios)
682 {
683         struct completion *waiting = ios->private;
684
685         complete(waiting);
686         return 0;
687 }
688
689 static void _last_io(struct kref *kref)
690 {
691         struct objio_state *ios = container_of(kref, struct objio_state, kref);
692
693         ios->done(ios);
694 }
695
696 static void _done_io(struct osd_request *or, void *p)
697 {
698         struct objio_state *ios = p;
699
700         kref_put(&ios->kref, _last_io);
701 }
702
703 static ssize_t _io_exec(struct objio_state *ios)
704 {
705         DECLARE_COMPLETION_ONSTACK(wait);
706         ssize_t status = 0; /* sync status */
707         unsigned i;
708         objio_done_fn saved_done_fn = ios->done;
709         bool sync = ios->ol_state.sync;
710
711         if (sync) {
712                 ios->done = _sync_done;
713                 ios->private = &wait;
714         }
715
716         kref_init(&ios->kref);
717
718         for (i = 0; i < ios->numdevs; i++) {
719                 struct osd_request *or = ios->per_dev[i].or;
720
721                 if (!or)
722                         continue;
723
724                 kref_get(&ios->kref);
725                 osd_execute_request_async(or, _done_io, ios);
726         }
727
728         kref_put(&ios->kref, _last_io);
729
730         if (sync) {
731                 wait_for_completion(&wait);
732                 status = saved_done_fn(ios);
733         }
734
735         return status;
736 }
737
738 /*
739  * read
740  */
741 static ssize_t _read_done(struct objio_state *ios)
742 {
743         ssize_t status;
744         int ret = _io_check(ios, false);
745
746         _io_free(ios);
747
748         if (likely(!ret))
749                 status = ios->length;
750         else
751                 status = ret;
752
753         objlayout_read_done(&ios->ol_state, status, ios->ol_state.sync);
754         return status;
755 }
756
757 static int _read_mirrors(struct objio_state *ios, unsigned cur_comp)
758 {
759         struct osd_request *or = NULL;
760         struct _objio_per_comp *per_dev = &ios->per_dev[cur_comp];
761         unsigned dev = per_dev->dev;
762         struct pnfs_osd_object_cred *cred =
763                         &ios->layout->comps[dev];
764         struct osd_obj_id obj = {
765                 .partition = cred->oc_object_id.oid_partition_id,
766                 .id = cred->oc_object_id.oid_object_id,
767         };
768         int ret;
769
770         or = osd_start_request(_io_od(ios, dev), GFP_KERNEL);
771         if (unlikely(!or)) {
772                 ret = -ENOMEM;
773                 goto err;
774         }
775         per_dev->or = or;
776
777         osd_req_read(or, &obj, per_dev->offset, per_dev->bio, per_dev->length);
778
779         ret = osd_finalize_request(or, 0, cred->oc_cap.cred, NULL);
780         if (ret) {
781                 dprintk("%s: Faild to osd_finalize_request() => %d\n",
782                         __func__, ret);
783                 goto err;
784         }
785
786         dprintk("%s:[%d] dev=%d obj=0x%llx start=0x%llx length=0x%lx\n",
787                 __func__, cur_comp, dev, obj.id, _LLU(per_dev->offset),
788                 per_dev->length);
789
790 err:
791         return ret;
792 }
793
794 static ssize_t _read_exec(struct objio_state *ios)
795 {
796         unsigned i;
797         int ret;
798
799         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
800                 if (!ios->per_dev[i].length)
801                         continue;
802                 ret = _read_mirrors(ios, i);
803                 if (unlikely(ret))
804                         goto err;
805         }
806
807         ios->done = _read_done;
808         return _io_exec(ios); /* In sync mode exec returns the io status */
809
810 err:
811         _io_free(ios);
812         return ret;
813 }
814
815 ssize_t objio_read_pagelist(struct objlayout_io_state *ol_state)
816 {
817         struct objio_state *ios = container_of(ol_state, struct objio_state,
818                                                ol_state);
819         int ret;
820
821         ret = _io_rw_pagelist(ios, GFP_KERNEL);
822         if (unlikely(ret))
823                 return ret;
824
825         return _read_exec(ios);
826 }
827
828 /*
829  * write
830  */
831 static ssize_t _write_done(struct objio_state *ios)
832 {
833         ssize_t status;
834         int ret = _io_check(ios, true);
835
836         _io_free(ios);
837
838         if (likely(!ret)) {
839                 /* FIXME: should be based on the OSD's persistence model
840                  * See OSD2r05 Section 4.13 Data persistence model */
841                 ios->ol_state.committed = NFS_FILE_SYNC;
842                 status = ios->length;
843         } else {
844                 status = ret;
845         }
846
847         objlayout_write_done(&ios->ol_state, status, ios->ol_state.sync);
848         return status;
849 }
850
851 static int _write_mirrors(struct objio_state *ios, unsigned cur_comp)
852 {
853         struct _objio_per_comp *master_dev = &ios->per_dev[cur_comp];
854         unsigned dev = ios->per_dev[cur_comp].dev;
855         unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
856         int ret;
857
858         for (; cur_comp < last_comp; ++cur_comp, ++dev) {
859                 struct osd_request *or = NULL;
860                 struct pnfs_osd_object_cred *cred =
861                                         &ios->layout->comps[dev];
862                 struct osd_obj_id obj = {
863                         .partition = cred->oc_object_id.oid_partition_id,
864                         .id = cred->oc_object_id.oid_object_id,
865                 };
866                 struct _objio_per_comp *per_dev = &ios->per_dev[cur_comp];
867                 struct bio *bio;
868
869                 or = osd_start_request(_io_od(ios, dev), GFP_NOFS);
870                 if (unlikely(!or)) {
871                         ret = -ENOMEM;
872                         goto err;
873                 }
874                 per_dev->or = or;
875
876                 if (per_dev != master_dev) {
877                         bio = bio_kmalloc(GFP_NOFS,
878                                           master_dev->bio->bi_max_vecs);
879                         if (unlikely(!bio)) {
880                                 dprintk("Faild to allocate BIO size=%u\n",
881                                         master_dev->bio->bi_max_vecs);
882                                 ret = -ENOMEM;
883                                 goto err;
884                         }
885
886                         __bio_clone(bio, master_dev->bio);
887                         bio->bi_bdev = NULL;
888                         bio->bi_next = NULL;
889                         per_dev->bio = bio;
890                         per_dev->dev = dev;
891                         per_dev->length = master_dev->length;
892                         per_dev->offset =  master_dev->offset;
893                 } else {
894                         bio = master_dev->bio;
895                         bio->bi_rw |= REQ_WRITE;
896                 }
897
898                 osd_req_write(or, &obj, per_dev->offset, bio, per_dev->length);
899
900                 ret = osd_finalize_request(or, 0, cred->oc_cap.cred, NULL);
901                 if (ret) {
902                         dprintk("%s: Faild to osd_finalize_request() => %d\n",
903                                 __func__, ret);
904                         goto err;
905                 }
906
907                 dprintk("%s:[%d] dev=%d obj=0x%llx start=0x%llx length=0x%lx\n",
908                         __func__, cur_comp, dev, obj.id, _LLU(per_dev->offset),
909                         per_dev->length);
910         }
911
912 err:
913         return ret;
914 }
915
916 static ssize_t _write_exec(struct objio_state *ios)
917 {
918         unsigned i;
919         int ret;
920
921         for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
922                 if (!ios->per_dev[i].length)
923                         continue;
924                 ret = _write_mirrors(ios, i);
925                 if (unlikely(ret))
926                         goto err;
927         }
928
929         ios->done = _write_done;
930         return _io_exec(ios); /* In sync mode exec returns the io->status */
931
932 err:
933         _io_free(ios);
934         return ret;
935 }
936
937 ssize_t objio_write_pagelist(struct objlayout_io_state *ol_state, bool stable)
938 {
939         struct objio_state *ios = container_of(ol_state, struct objio_state,
940                                                ol_state);
941         int ret;
942
943         /* TODO: ios->stable = stable; */
944         ret = _io_rw_pagelist(ios, GFP_NOFS);
945         if (unlikely(ret))
946                 return ret;
947
948         return _write_exec(ios);
949 }
950
951 /*
952  * objlayout_pg_test(). Called by nfs_can_coalesce_requests()
953  *
954  * return 1 :  coalesce page
955  * return 0 :  don't coalesce page
956  */
957 int
958 objlayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
959                    struct nfs_page *req)
960 {
961         return 1;
962 }
963
964 static struct pnfs_layoutdriver_type objlayout_type = {
965         .id = LAYOUT_OSD2_OBJECTS,
966         .name = "LAYOUT_OSD2_OBJECTS",
967
968         .alloc_layout_hdr        = objlayout_alloc_layout_hdr,
969         .free_layout_hdr         = objlayout_free_layout_hdr,
970
971         .alloc_lseg              = objlayout_alloc_lseg,
972         .free_lseg               = objlayout_free_lseg,
973
974         .read_pagelist           = objlayout_read_pagelist,
975         .write_pagelist          = objlayout_write_pagelist,
976         .pg_test                 = objlayout_pg_test,
977
978         .free_deviceid_node      = objio_free_deviceid_node,
979 };
980
981 MODULE_DESCRIPTION("pNFS Layout Driver for OSD2 objects");
982 MODULE_AUTHOR("Benny Halevy <bhalevy@panasas.com>");
983 MODULE_LICENSE("GPL");
984
985 static int __init
986 objlayout_init(void)
987 {
988         int ret = pnfs_register_layoutdriver(&objlayout_type);
989
990         if (ret)
991                 printk(KERN_INFO
992                         "%s: Registering OSD pNFS Layout Driver failed: error=%d\n",
993                         __func__, ret);
994         else
995                 printk(KERN_INFO "%s: Registered OSD pNFS Layout Driver\n",
996                         __func__);
997         return ret;
998 }
999
1000 static void __exit
1001 objlayout_exit(void)
1002 {
1003         pnfs_unregister_layoutdriver(&objlayout_type);
1004         printk(KERN_INFO "%s: Unregistered OSD pNFS Layout Driver\n",
1005                __func__);
1006 }
1007
1008 module_init(objlayout_init);
1009 module_exit(objlayout_exit);