NFSv4.1: layoutcommit
[linux-2.6.git] / fs / nfs / nfs4filelayout.c
1 /*
2  *  Module for the pnfs nfs4 file layout driver.
3  *  Defines all I/O and Policy interface operations, plus code
4  *  to register itself with the pNFS client.
5  *
6  *  Copyright (c) 2002
7  *  The Regents of the University of Michigan
8  *  All Rights Reserved
9  *
10  *  Dean Hildebrand <dhildebz@umich.edu>
11  *
12  *  Permission is granted to use, copy, create derivative works, and
13  *  redistribute this software and such derivative works for any purpose,
14  *  so long as the name of the University of Michigan is not used in
15  *  any advertising or publicity pertaining to the use or distribution
16  *  of this software without specific, written prior authorization. If
17  *  the above copyright notice or any other identification of the
18  *  University of Michigan is included in any copy of any portion of
19  *  this software, then the disclaimer below must also be included.
20  *
21  *  This software is provided as is, without representation or warranty
22  *  of any kind either express or implied, including without limitation
23  *  the implied warranties of merchantability, fitness for a particular
24  *  purpose, or noninfringement.  The Regents of the University of
25  *  Michigan shall not be liable for any damages, including special,
26  *  indirect, incidental, or consequential damages, with respect to any
27  *  claim arising out of or in connection with the use of the software,
28  *  even if it has been or is hereafter advised of the possibility of
29  *  such damages.
30  */
31
32 #include <linux/nfs_fs.h>
33
34 #include "internal.h"
35 #include "nfs4filelayout.h"
36
37 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
38
39 MODULE_LICENSE("GPL");
40 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
41 MODULE_DESCRIPTION("The NFSv4 file layout driver");
42
43 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
44
45 static loff_t
46 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
47                             loff_t offset)
48 {
49         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
50         u64 tmp;
51
52         offset -= flseg->pattern_offset;
53         tmp = offset;
54         do_div(tmp, stripe_width);
55
56         return tmp * flseg->stripe_unit + do_div(offset, flseg->stripe_unit);
57 }
58
59 /* This function is used by the layout driver to calculate the
60  * offset of the file on the dserver based on whether the
61  * layout type is STRIPE_DENSE or STRIPE_SPARSE
62  */
63 static loff_t
64 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
65 {
66         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
67
68         switch (flseg->stripe_type) {
69         case STRIPE_SPARSE:
70                 return offset;
71
72         case STRIPE_DENSE:
73                 return filelayout_get_dense_offset(flseg, offset);
74         }
75
76         BUG();
77 }
78
79 /* For data server errors we don't recover from */
80 static void
81 filelayout_set_lo_fail(struct pnfs_layout_segment *lseg)
82 {
83         if (lseg->pls_range.iomode == IOMODE_RW) {
84                 dprintk("%s Setting layout IOMODE_RW fail bit\n", __func__);
85                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
86         } else {
87                 dprintk("%s Setting layout IOMODE_READ fail bit\n", __func__);
88                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
89         }
90 }
91
92 static int filelayout_async_handle_error(struct rpc_task *task,
93                                          struct nfs4_state *state,
94                                          struct nfs_client *clp,
95                                          int *reset)
96 {
97         if (task->tk_status >= 0)
98                 return 0;
99
100         *reset = 0;
101
102         switch (task->tk_status) {
103         case -NFS4ERR_BADSESSION:
104         case -NFS4ERR_BADSLOT:
105         case -NFS4ERR_BAD_HIGH_SLOT:
106         case -NFS4ERR_DEADSESSION:
107         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
108         case -NFS4ERR_SEQ_FALSE_RETRY:
109         case -NFS4ERR_SEQ_MISORDERED:
110                 dprintk("%s ERROR %d, Reset session. Exchangeid "
111                         "flags 0x%x\n", __func__, task->tk_status,
112                         clp->cl_exchange_flags);
113                 nfs4_schedule_session_recovery(clp->cl_session);
114                 break;
115         case -NFS4ERR_DELAY:
116         case -NFS4ERR_GRACE:
117         case -EKEYEXPIRED:
118                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
119                 break;
120         default:
121                 dprintk("%s DS error. Retry through MDS %d\n", __func__,
122                         task->tk_status);
123                 *reset = 1;
124                 break;
125         }
126         task->tk_status = 0;
127         return -EAGAIN;
128 }
129
130 /* NFS_PROTO call done callback routines */
131
132 static int filelayout_read_done_cb(struct rpc_task *task,
133                                 struct nfs_read_data *data)
134 {
135         struct nfs_client *clp = data->ds_clp;
136         int reset = 0;
137
138         dprintk("%s DS read\n", __func__);
139
140         if (filelayout_async_handle_error(task, data->args.context->state,
141                                           data->ds_clp, &reset) == -EAGAIN) {
142                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
143                         __func__, data->ds_clp, data->ds_clp->cl_session);
144                 if (reset) {
145                         filelayout_set_lo_fail(data->lseg);
146                         nfs4_reset_read(task, data);
147                         clp = NFS_SERVER(data->inode)->nfs_client;
148                 }
149                 nfs_restart_rpc(task, clp);
150                 return -EAGAIN;
151         }
152
153         return 0;
154 }
155
156 /*
157  * We reference the rpc_cred of the first WRITE that triggers the need for
158  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
159  * rfc5661 is not clear about which credential should be used.
160  */
161 static void
162 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
163 {
164         if (FILELAYOUT_LSEG(wdata->lseg)->commit_through_mds ||
165             wdata->res.verf->committed == NFS_FILE_SYNC)
166                 return;
167
168         pnfs_set_layoutcommit(wdata);
169         dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, wdata->inode->i_ino,
170                 (unsigned long) wdata->lseg->pls_end_pos);
171 }
172
173 /*
174  * Call ops for the async read/write cases
175  * In the case of dense layouts, the offset needs to be reset to its
176  * original value.
177  */
178 static void filelayout_read_prepare(struct rpc_task *task, void *data)
179 {
180         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
181
182         rdata->read_done_cb = filelayout_read_done_cb;
183
184         if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
185                                 &rdata->args.seq_args, &rdata->res.seq_res,
186                                 0, task))
187                 return;
188
189         rpc_call_start(task);
190 }
191
192 static void filelayout_read_call_done(struct rpc_task *task, void *data)
193 {
194         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
195
196         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
197
198         /* Note this may cause RPC to be resent */
199         rdata->mds_ops->rpc_call_done(task, data);
200 }
201
202 static void filelayout_read_release(void *data)
203 {
204         struct nfs_read_data *rdata = (struct nfs_read_data *)data;
205
206         rdata->mds_ops->rpc_release(data);
207 }
208
209 static int filelayout_write_done_cb(struct rpc_task *task,
210                                 struct nfs_write_data *data)
211 {
212         int reset = 0;
213
214         if (filelayout_async_handle_error(task, data->args.context->state,
215                                           data->ds_clp, &reset) == -EAGAIN) {
216                 struct nfs_client *clp;
217
218                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
219                         __func__, data->ds_clp, data->ds_clp->cl_session);
220                 if (reset) {
221                         filelayout_set_lo_fail(data->lseg);
222                         nfs4_reset_write(task, data);
223                         clp = NFS_SERVER(data->inode)->nfs_client;
224                 } else
225                         clp = data->ds_clp;
226                 nfs_restart_rpc(task, clp);
227                 return -EAGAIN;
228         }
229
230         filelayout_set_layoutcommit(data);
231         return 0;
232 }
233
234 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
235 static void prepare_to_resend_writes(struct nfs_write_data *data)
236 {
237         struct nfs_page *first = nfs_list_entry(data->pages.next);
238
239         data->task.tk_status = 0;
240         memcpy(data->verf.verifier, first->wb_verf.verifier,
241                sizeof(first->wb_verf.verifier));
242         data->verf.verifier[0]++; /* ensure verifier mismatch */
243 }
244
245 static int filelayout_commit_done_cb(struct rpc_task *task,
246                                      struct nfs_write_data *data)
247 {
248         int reset = 0;
249
250         if (filelayout_async_handle_error(task, data->args.context->state,
251                                           data->ds_clp, &reset) == -EAGAIN) {
252                 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
253                         __func__, data->ds_clp, data->ds_clp->cl_session);
254                 if (reset) {
255                         prepare_to_resend_writes(data);
256                         filelayout_set_lo_fail(data->lseg);
257                 } else
258                         nfs_restart_rpc(task, data->ds_clp);
259                 return -EAGAIN;
260         }
261
262         return 0;
263 }
264
265 static void filelayout_write_prepare(struct rpc_task *task, void *data)
266 {
267         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
268
269         if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
270                                 &wdata->args.seq_args, &wdata->res.seq_res,
271                                 0, task))
272                 return;
273
274         rpc_call_start(task);
275 }
276
277 static void filelayout_write_call_done(struct rpc_task *task, void *data)
278 {
279         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
280
281         /* Note this may cause RPC to be resent */
282         wdata->mds_ops->rpc_call_done(task, data);
283 }
284
285 static void filelayout_write_release(void *data)
286 {
287         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
288
289         wdata->mds_ops->rpc_release(data);
290 }
291
292 static void filelayout_commit_release(void *data)
293 {
294         struct nfs_write_data *wdata = (struct nfs_write_data *)data;
295
296         nfs_commit_release_pages(wdata);
297         if (atomic_dec_and_test(&NFS_I(wdata->inode)->commits_outstanding))
298                 nfs_commit_clear_lock(NFS_I(wdata->inode));
299         nfs_commitdata_release(wdata);
300 }
301
302 struct rpc_call_ops filelayout_read_call_ops = {
303         .rpc_call_prepare = filelayout_read_prepare,
304         .rpc_call_done = filelayout_read_call_done,
305         .rpc_release = filelayout_read_release,
306 };
307
308 struct rpc_call_ops filelayout_write_call_ops = {
309         .rpc_call_prepare = filelayout_write_prepare,
310         .rpc_call_done = filelayout_write_call_done,
311         .rpc_release = filelayout_write_release,
312 };
313
314 struct rpc_call_ops filelayout_commit_call_ops = {
315         .rpc_call_prepare = filelayout_write_prepare,
316         .rpc_call_done = filelayout_write_call_done,
317         .rpc_release = filelayout_commit_release,
318 };
319
320 static enum pnfs_try_status
321 filelayout_read_pagelist(struct nfs_read_data *data)
322 {
323         struct pnfs_layout_segment *lseg = data->lseg;
324         struct nfs4_pnfs_ds *ds;
325         loff_t offset = data->args.offset;
326         u32 j, idx;
327         struct nfs_fh *fh;
328         int status;
329
330         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
331                 __func__, data->inode->i_ino,
332                 data->args.pgbase, (size_t)data->args.count, offset);
333
334         /* Retrieve the correct rpc_client for the byte range */
335         j = nfs4_fl_calc_j_index(lseg, offset);
336         idx = nfs4_fl_calc_ds_index(lseg, j);
337         ds = nfs4_fl_prepare_ds(lseg, idx);
338         if (!ds) {
339                 /* Either layout fh index faulty, or ds connect failed */
340                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
341                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
342                 return PNFS_NOT_ATTEMPTED;
343         }
344         dprintk("%s USE DS:ip %x %hu\n", __func__,
345                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
346
347         /* No multipath support. Use first DS */
348         data->ds_clp = ds->ds_clp;
349         fh = nfs4_fl_select_ds_fh(lseg, j);
350         if (fh)
351                 data->args.fh = fh;
352
353         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
354         data->mds_offset = offset;
355
356         /* Perform an asynchronous read to ds */
357         status = nfs_initiate_read(data, ds->ds_clp->cl_rpcclient,
358                                    &filelayout_read_call_ops);
359         BUG_ON(status != 0);
360         return PNFS_ATTEMPTED;
361 }
362
363 /* Perform async writes. */
364 static enum pnfs_try_status
365 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
366 {
367         struct pnfs_layout_segment *lseg = data->lseg;
368         struct nfs4_pnfs_ds *ds;
369         loff_t offset = data->args.offset;
370         u32 j, idx;
371         struct nfs_fh *fh;
372         int status;
373
374         /* Retrieve the correct rpc_client for the byte range */
375         j = nfs4_fl_calc_j_index(lseg, offset);
376         idx = nfs4_fl_calc_ds_index(lseg, j);
377         ds = nfs4_fl_prepare_ds(lseg, idx);
378         if (!ds) {
379                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
380                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
381                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
382                 return PNFS_NOT_ATTEMPTED;
383         }
384         dprintk("%s ino %lu sync %d req %Zu@%llu DS:%x:%hu\n", __func__,
385                 data->inode->i_ino, sync, (size_t) data->args.count, offset,
386                 ntohl(ds->ds_ip_addr), ntohs(ds->ds_port));
387
388         /* We can't handle commit to ds yet */
389         if (!FILELAYOUT_LSEG(lseg)->commit_through_mds)
390                 data->args.stable = NFS_FILE_SYNC;
391
392         data->write_done_cb = filelayout_write_done_cb;
393         data->ds_clp = ds->ds_clp;
394         fh = nfs4_fl_select_ds_fh(lseg, j);
395         if (fh)
396                 data->args.fh = fh;
397         /*
398          * Get the file offset on the dserver. Set the write offset to
399          * this offset and save the original offset.
400          */
401         data->args.offset = filelayout_get_dserver_offset(lseg, offset);
402         data->mds_offset = offset;
403
404         /* Perform an asynchronous write */
405         status = nfs_initiate_write(data, ds->ds_clp->cl_rpcclient,
406                                     &filelayout_write_call_ops, sync);
407         BUG_ON(status != 0);
408         return PNFS_ATTEMPTED;
409 }
410
411 /*
412  * filelayout_check_layout()
413  *
414  * Make sure layout segment parameters are sane WRT the device.
415  * At this point no generic layer initialization of the lseg has occurred,
416  * and nothing has been added to the layout_hdr cache.
417  *
418  */
419 static int
420 filelayout_check_layout(struct pnfs_layout_hdr *lo,
421                         struct nfs4_filelayout_segment *fl,
422                         struct nfs4_layoutget_res *lgr,
423                         struct nfs4_deviceid *id)
424 {
425         struct nfs4_file_layout_dsaddr *dsaddr;
426         int status = -EINVAL;
427         struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
428
429         dprintk("--> %s\n", __func__);
430
431         if (fl->pattern_offset > lgr->range.offset) {
432                 dprintk("%s pattern_offset %lld to large\n",
433                                 __func__, fl->pattern_offset);
434                 goto out;
435         }
436
437         if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
438                 dprintk("%s Invalid stripe unit (%u)\n",
439                         __func__, fl->stripe_unit);
440                 goto out;
441         }
442
443         /* find and reference the deviceid */
444         dsaddr = nfs4_fl_find_get_deviceid(id);
445         if (dsaddr == NULL) {
446                 dsaddr = get_device_info(lo->plh_inode, id);
447                 if (dsaddr == NULL)
448                         goto out;
449         }
450         fl->dsaddr = dsaddr;
451
452         if (fl->first_stripe_index < 0 ||
453             fl->first_stripe_index >= dsaddr->stripe_count) {
454                 dprintk("%s Bad first_stripe_index %d\n",
455                                 __func__, fl->first_stripe_index);
456                 goto out_put;
457         }
458
459         if ((fl->stripe_type == STRIPE_SPARSE &&
460             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
461             (fl->stripe_type == STRIPE_DENSE &&
462             fl->num_fh != dsaddr->stripe_count)) {
463                 dprintk("%s num_fh %u not valid for given packing\n",
464                         __func__, fl->num_fh);
465                 goto out_put;
466         }
467
468         if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
469                 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
470                         "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
471                         nfss->wsize);
472         }
473
474         status = 0;
475 out:
476         dprintk("--> %s returns %d\n", __func__, status);
477         return status;
478 out_put:
479         nfs4_fl_put_deviceid(dsaddr);
480         goto out;
481 }
482
483 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
484 {
485         int i;
486
487         for (i = 0; i < fl->num_fh; i++) {
488                 if (!fl->fh_array[i])
489                         break;
490                 kfree(fl->fh_array[i]);
491         }
492         kfree(fl->fh_array);
493         fl->fh_array = NULL;
494 }
495
496 static void
497 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
498 {
499         filelayout_free_fh_array(fl);
500         kfree(fl);
501 }
502
503 static int
504 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
505                          struct nfs4_filelayout_segment *fl,
506                          struct nfs4_layoutget_res *lgr,
507                          struct nfs4_deviceid *id)
508 {
509         uint32_t *p = (uint32_t *)lgr->layout.buf;
510         uint32_t nfl_util;
511         int i;
512
513         dprintk("%s: set_layout_map Begin\n", __func__);
514
515         memcpy(id, p, sizeof(*id));
516         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
517         print_deviceid(id);
518
519         nfl_util = be32_to_cpup(p++);
520         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
521                 fl->commit_through_mds = 1;
522         if (nfl_util & NFL4_UFLG_DENSE)
523                 fl->stripe_type = STRIPE_DENSE;
524         else
525                 fl->stripe_type = STRIPE_SPARSE;
526         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
527
528         fl->first_stripe_index = be32_to_cpup(p++);
529         p = xdr_decode_hyper(p, &fl->pattern_offset);
530         fl->num_fh = be32_to_cpup(p++);
531
532         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
533                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
534                 fl->pattern_offset);
535
536         fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
537                                GFP_KERNEL);
538         if (!fl->fh_array)
539                 return -ENOMEM;
540
541         for (i = 0; i < fl->num_fh; i++) {
542                 /* Do we want to use a mempool here? */
543                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), GFP_KERNEL);
544                 if (!fl->fh_array[i]) {
545                         filelayout_free_fh_array(fl);
546                         return -ENOMEM;
547                 }
548                 fl->fh_array[i]->size = be32_to_cpup(p++);
549                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
550                         printk(KERN_ERR "Too big fh %d received %d\n",
551                                i, fl->fh_array[i]->size);
552                         filelayout_free_fh_array(fl);
553                         return -EIO;
554                 }
555                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
556                 p += XDR_QUADLEN(fl->fh_array[i]->size);
557                 dprintk("DEBUG: %s: fh len %d\n", __func__,
558                         fl->fh_array[i]->size);
559         }
560
561         return 0;
562 }
563
564 static void
565 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
566 {
567         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
568
569         dprintk("--> %s\n", __func__);
570         nfs4_fl_put_deviceid(fl->dsaddr);
571         kfree(fl->commit_buckets);
572         _filelayout_free_lseg(fl);
573 }
574
575 static struct pnfs_layout_segment *
576 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
577                       struct nfs4_layoutget_res *lgr)
578 {
579         struct nfs4_filelayout_segment *fl;
580         int rc;
581         struct nfs4_deviceid id;
582
583         dprintk("--> %s\n", __func__);
584         fl = kzalloc(sizeof(*fl), GFP_KERNEL);
585         if (!fl)
586                 return NULL;
587
588         rc = filelayout_decode_layout(layoutid, fl, lgr, &id);
589         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id)) {
590                 _filelayout_free_lseg(fl);
591                 return NULL;
592         }
593
594         /* This assumes there is only one IOMODE_RW lseg.  What
595          * we really want to do is have a layout_hdr level
596          * dictionary of <multipath_list4, fh> keys, each
597          * associated with a struct list_head, populated by calls
598          * to filelayout_write_pagelist().
599          * */
600         if ((!fl->commit_through_mds) && (lgr->range.iomode == IOMODE_RW)) {
601                 int i;
602                 int size = (fl->stripe_type == STRIPE_SPARSE) ?
603                         fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
604
605                 fl->commit_buckets = kcalloc(size, sizeof(struct list_head), GFP_KERNEL);
606                 if (!fl->commit_buckets) {
607                         filelayout_free_lseg(&fl->generic_hdr);
608                         return NULL;
609                 }
610                 fl->number_of_buckets = size;
611                 for (i = 0; i < size; i++)
612                         INIT_LIST_HEAD(&fl->commit_buckets[i]);
613         }
614         return &fl->generic_hdr;
615 }
616
617 /*
618  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
619  *
620  * return 1 :  coalesce page
621  * return 0 :  don't coalesce page
622  */
623 int
624 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
625                    struct nfs_page *req)
626 {
627         u64 p_stripe, r_stripe;
628         u32 stripe_unit;
629
630         if (!pgio->pg_lseg)
631                 return 1;
632         p_stripe = (u64)prev->wb_index << PAGE_CACHE_SHIFT;
633         r_stripe = (u64)req->wb_index << PAGE_CACHE_SHIFT;
634         stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
635
636         do_div(p_stripe, stripe_unit);
637         do_div(r_stripe, stripe_unit);
638
639         return (p_stripe == r_stripe);
640 }
641
642 static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
643 {
644         return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
645 }
646
647 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
648 {
649         if (fl->stripe_type == STRIPE_SPARSE)
650                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
651         else
652                 return j;
653 }
654
655 struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
656 {
657         struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
658         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
659         u32 i, j;
660         struct list_head *list;
661
662         /* Note that we are calling nfs4_fl_calc_j_index on each page
663          * that ends up being committed to a data server.  An attractive
664          * alternative is to add a field to nfs_write_data and nfs_page
665          * to store the value calculated in filelayout_write_pagelist
666          * and just use that here.
667          */
668         j = nfs4_fl_calc_j_index(lseg,
669                                  (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
670         i = select_bucket_index(fl, j);
671         list = &fl->commit_buckets[i];
672         if (list_empty(list)) {
673                 /* Non-empty buckets hold a reference on the lseg */
674                 get_lseg(lseg);
675         }
676         return list;
677 }
678
679 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
680 {
681         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
682
683         if (flseg->stripe_type == STRIPE_SPARSE)
684                 return i;
685         else
686                 return nfs4_fl_calc_ds_index(lseg, i);
687 }
688
689 static struct nfs_fh *
690 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
691 {
692         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
693
694         if (flseg->stripe_type == STRIPE_SPARSE) {
695                 if (flseg->num_fh == 1)
696                         i = 0;
697                 else if (flseg->num_fh == 0)
698                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
699                         return NULL;
700         }
701         return flseg->fh_array[i];
702 }
703
704 static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
705 {
706         struct pnfs_layout_segment *lseg = data->lseg;
707         struct nfs4_pnfs_ds *ds;
708         u32 idx;
709         struct nfs_fh *fh;
710
711         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
712         ds = nfs4_fl_prepare_ds(lseg, idx);
713         if (!ds) {
714                 printk(KERN_ERR "%s: prepare_ds failed, use MDS\n", __func__);
715                 set_bit(lo_fail_bit(IOMODE_RW), &lseg->pls_layout->plh_flags);
716                 set_bit(lo_fail_bit(IOMODE_READ), &lseg->pls_layout->plh_flags);
717                 prepare_to_resend_writes(data);
718                 data->mds_ops->rpc_release(data);
719                 return -EAGAIN;
720         }
721         dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
722         data->write_done_cb = filelayout_commit_done_cb;
723         data->ds_clp = ds->ds_clp;
724         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
725         if (fh)
726                 data->args.fh = fh;
727         return nfs_initiate_commit(data, ds->ds_clp->cl_rpcclient,
728                                    &filelayout_commit_call_ops, how);
729 }
730
731 /*
732  * This is only useful while we are using whole file layouts.
733  */
734 static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
735 {
736         struct pnfs_layout_segment *lseg, *rv = NULL;
737
738         spin_lock(&inode->i_lock);
739         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
740                 if (lseg->pls_range.iomode == IOMODE_RW)
741                         rv = get_lseg(lseg);
742         spin_unlock(&inode->i_lock);
743         return rv;
744 }
745
746 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
747 {
748         struct pnfs_layout_segment *lseg;
749         struct nfs4_filelayout_segment *fl;
750         struct nfs_write_data *data;
751         int i, j;
752
753         /* Won't need this when non-whole file layout segments are supported
754          * instead we will use a pnfs_layout_hdr structure */
755         lseg = find_only_write_lseg(inode);
756         if (!lseg)
757                 return 0;
758         fl = FILELAYOUT_LSEG(lseg);
759         for (i = 0; i < fl->number_of_buckets; i++) {
760                 if (list_empty(&fl->commit_buckets[i]))
761                         continue;
762                 data = nfs_commitdata_alloc();
763                 if (!data)
764                         goto out_bad;
765                 data->ds_commit_index = i;
766                 data->lseg = lseg;
767                 list_add(&data->pages, list);
768         }
769         put_lseg(lseg);
770         return 0;
771
772 out_bad:
773         for (j = i; j < fl->number_of_buckets; j++) {
774                 if (list_empty(&fl->commit_buckets[i]))
775                         continue;
776                 nfs_retry_commit(&fl->commit_buckets[i], lseg);
777                 put_lseg(lseg);  /* associated with emptying bucket */
778         }
779         put_lseg(lseg);
780         /* Caller will clean up entries put on list */
781         return -ENOMEM;
782 }
783
784 /* This follows nfs_commit_list pretty closely */
785 static int
786 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
787                            int how)
788 {
789         struct nfs_write_data   *data, *tmp;
790         LIST_HEAD(list);
791
792         if (!list_empty(mds_pages)) {
793                 data = nfs_commitdata_alloc();
794                 if (!data)
795                         goto out_bad;
796                 data->lseg = NULL;
797                 list_add(&data->pages, &list);
798         }
799
800         if (alloc_ds_commits(inode, &list))
801                 goto out_bad;
802
803         list_for_each_entry_safe(data, tmp, &list, pages) {
804                 list_del_init(&data->pages);
805                 atomic_inc(&NFS_I(inode)->commits_outstanding);
806                 if (!data->lseg) {
807                         nfs_init_commit(data, mds_pages, NULL);
808                         nfs_initiate_commit(data, NFS_CLIENT(inode),
809                                             data->mds_ops, how);
810                 } else {
811                         nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
812                         filelayout_initiate_commit(data, how);
813                 }
814         }
815         return 0;
816  out_bad:
817         list_for_each_entry_safe(data, tmp, &list, pages) {
818                 nfs_retry_commit(&data->pages, data->lseg);
819                 list_del_init(&data->pages);
820                 nfs_commit_free(data);
821         }
822         nfs_retry_commit(mds_pages, NULL);
823         nfs_commit_clear_lock(NFS_I(inode));
824         return -ENOMEM;
825 }
826
827 static struct pnfs_layoutdriver_type filelayout_type = {
828         .id                     = LAYOUT_NFSV4_1_FILES,
829         .name                   = "LAYOUT_NFSV4_1_FILES",
830         .owner                  = THIS_MODULE,
831         .alloc_lseg             = filelayout_alloc_lseg,
832         .free_lseg              = filelayout_free_lseg,
833         .pg_test                = filelayout_pg_test,
834         .mark_pnfs_commit       = filelayout_mark_pnfs_commit,
835         .choose_commit_list     = filelayout_choose_commit_list,
836         .commit_pagelist        = filelayout_commit_pagelist,
837         .read_pagelist          = filelayout_read_pagelist,
838         .write_pagelist         = filelayout_write_pagelist,
839 };
840
841 static int __init nfs4filelayout_init(void)
842 {
843         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
844                __func__);
845         return pnfs_register_layoutdriver(&filelayout_type);
846 }
847
848 static void __exit nfs4filelayout_exit(void)
849 {
850         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
851                __func__);
852         pnfs_unregister_layoutdriver(&filelayout_type);
853 }
854
855 module_init(nfs4filelayout_init);
856 module_exit(nfs4filelayout_exit);