]> git.ipfire.org Git - people/ms/linux.git/blame - fs/nfs/flexfilelayout/flexfilelayout.c
projects / people / ms / linux.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
pNFS/flexfiles: track when layout is first used
[people/ms/linux.git] / fs / nfs / flexfilelayout / flexfilelayout.c
CommitLineData
d67ae825
TH		 1/*
2 * Module for pnfs flexfile layout driver.
3 *
4 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
5 *
6 * Tao Peng <bergwolf@primarydata.com>
7 */
8
9#include <linux/nfs_fs.h>
10#include <linux/nfs_page.h>
11#include <linux/module.h>
12
13#include <linux/sunrpc/metrics.h>
d67ae825
TH		 14
15#include "flexfilelayout.h"
16#include "../nfs4session.h"
40c64c26 17#include "../nfs4idmap.h"
d67ae825
TH		 18#include "../internal.h"
19#include "../delegation.h"
20#include "../nfs4trace.h"
21#include "../iostat.h"
22#include "../nfs.h"
23
24#define NFSDBG_FACILITY NFSDBG_PNFS_LD
25
26#define FF_LAYOUT_POLL_RETRY_MAX (15*HZ)
27
28static struct pnfs_layout_hdr *
29ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
30{
31 struct nfs4_flexfile_layout *ffl;
32
33 ffl = kzalloc(sizeof(*ffl), gfp_flags);
34 if (ffl) {
35 INIT_LIST_HEAD(&ffl->error_list);
36 return &ffl->generic_hdr;
37 } else
38 return NULL;
39}
40
41static void
42ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
43{
44 struct nfs4_ff_layout_ds_err *err, *n;
45
46 list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list,
47 list) {
48 list_del(&err->list);
49 kfree(err);
50 }
51 kfree(FF_LAYOUT_FROM_HDR(lo));
52}
53
54static int decode_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
55{
56 __be32 *p;
57
58 p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
59 if (unlikely(p == NULL))
60 return -ENOBUFS;
61 memcpy(stateid, p, NFS4_STATEID_SIZE);
62 dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
63 p[0], p[1], p[2], p[3]);
64 return 0;
65}
66
67static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
68{
69 __be32 *p;
70
71 p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
72 if (unlikely(!p))
73 return -ENOBUFS;
74 memcpy(devid, p, NFS4_DEVICEID4_SIZE);
75 nfs4_print_deviceid(devid);
76 return 0;
77}
78
79static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
80{
81 __be32 *p;
82
83 p = xdr_inline_decode(xdr, 4);
84 if (unlikely(!p))
85 return -ENOBUFS;
86 fh->size = be32_to_cpup(p++);
87 if (fh->size > sizeof(struct nfs_fh)) {
88 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
89 fh->size);
90 return -EOVERFLOW;
91 }
92 /* fh.data */
93 p = xdr_inline_decode(xdr, fh->size);
94 if (unlikely(!p))
95 return -ENOBUFS;
96 memcpy(&fh->data, p, fh->size);
97 dprintk("%s: fh len %d\n", __func__, fh->size);
98
99 return 0;
100}
101
102/*
103 * Currently only stringified uids and gids are accepted.
104 * I.e., kerberos is not supported to the DSes, so no pricipals.
105 *
106 * That means that one common function will suffice, but when
107 * principals are added, this should be split to accomodate
108 * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
109 */
110static int
111decode_name(struct xdr_stream *xdr, u32 *id)
112{
113 __be32 *p;
114 int len;
115
116 /* opaque_length(4)*/
117 p = xdr_inline_decode(xdr, 4);
118 if (unlikely(!p))
119 return -ENOBUFS;
120 len = be32_to_cpup(p++);
121 if (len < 0)
122 return -EINVAL;
123
124 dprintk("%s: len %u\n", __func__, len);
125
126 /* opaque body */
127 p = xdr_inline_decode(xdr, len);
128 if (unlikely(!p))
129 return -ENOBUFS;
130
131 if (!nfs_map_string_to_numeric((char *)p, len, id))
132 return -EINVAL;
133
134 return 0;
135}
136
137static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
138{
139 int i;
140
141 if (fls->mirror_array) {
142 for (i = 0; i < fls->mirror_array_cnt; i++) {
143 /* normally mirror_ds is freed in
144 * .free_deviceid_node but we still do it here
145 * for .alloc_lseg error path */
146 if (fls->mirror_array[i]) {
147 kfree(fls->mirror_array[i]->fh_versions);
148 nfs4_ff_layout_put_deviceid(fls->mirror_array[i]->mirror_ds);
149 kfree(fls->mirror_array[i]);
150 }
151 }
152 kfree(fls->mirror_array);
153 fls->mirror_array = NULL;
154 }
155}
156
157static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr)
158{
159 int ret = 0;
160
161 dprintk("--> %s\n", __func__);
162
163 /* FIXME: remove this check when layout segment support is added */
164 if (lgr->range.offset != 0 ||
165 lgr->range.length != NFS4_MAX_UINT64) {
166 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
167 __func__);
168 ret = -EINVAL;
169 }
170
171 dprintk("--> %s returns %d\n", __func__, ret);
172 return ret;
173}
174
175static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
176{
177 if (fls) {
178 ff_layout_free_mirror_array(fls);
179 kfree(fls);
180 }
181}
182
183static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
184{
d67ae825
TH		 185 int i, j;
186
187 for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
188 for (j = i + 1; j < fls->mirror_array_cnt; j++)
189 if (fls->mirror_array[i]->efficiency <
455b6ee6
FF		 190 fls->mirror_array[j]->efficiency)
191 swap(fls->mirror_array[i],
192 fls->mirror_array[j]);
d67ae825
TH		 193 }
194}
195
196static struct pnfs_layout_segment *
197ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
198 struct nfs4_layoutget_res *lgr,
199 gfp_t gfp_flags)
200{
201 struct pnfs_layout_segment *ret;
202 struct nfs4_ff_layout_segment *fls = NULL;
203 struct xdr_stream stream;
204 struct xdr_buf buf;
205 struct page *scratch;
206 u64 stripe_unit;
207 u32 mirror_array_cnt;
208 __be32 *p;
209 int i, rc;
210
211 dprintk("--> %s\n", __func__);
212 scratch = alloc_page(gfp_flags);
213 if (!scratch)
214 return ERR_PTR(-ENOMEM);
215
216 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
217 lgr->layoutp->len);
218 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
219
220 /* stripe unit and mirror_array_cnt */
221 rc = -EIO;
222 p = xdr_inline_decode(&stream, 8 + 4);
223 if (!p)
224 goto out_err_free;
225
226 p = xdr_decode_hyper(p, &stripe_unit);
227 mirror_array_cnt = be32_to_cpup(p++);
228 dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
229 stripe_unit, mirror_array_cnt);
230
231 if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
232 mirror_array_cnt == 0)
233 goto out_err_free;
234
235 rc = -ENOMEM;
236 fls = kzalloc(sizeof(*fls), gfp_flags);
237 if (!fls)
238 goto out_err_free;
239
240 fls->mirror_array_cnt = mirror_array_cnt;
241 fls->stripe_unit = stripe_unit;
242 fls->mirror_array = kcalloc(fls->mirror_array_cnt,
243 sizeof(fls->mirror_array[0]), gfp_flags);
244 if (fls->mirror_array == NULL)
245 goto out_err_free;
246
247 for (i = 0; i < fls->mirror_array_cnt; i++) {
248 struct nfs4_deviceid devid;
249 struct nfs4_deviceid_node *idnode;
250 u32 ds_count;
251 u32 fh_count;
252 int j;
253
254 rc = -EIO;
255 p = xdr_inline_decode(&stream, 4);
256 if (!p)
257 goto out_err_free;
258 ds_count = be32_to_cpup(p);
259
260 /* FIXME: allow for striping? */
261 if (ds_count != 1)
262 goto out_err_free;
263
264 fls->mirror_array[i] =
265 kzalloc(sizeof(struct nfs4_ff_layout_mirror),
266 gfp_flags);
267 if (fls->mirror_array[i] == NULL) {
268 rc = -ENOMEM;
269 goto out_err_free;
270 }
271
272 spin_lock_init(&fls->mirror_array[i]->lock);
273 fls->mirror_array[i]->ds_count = ds_count;
274
275 /* deviceid */
276 rc = decode_deviceid(&stream, &devid);
277 if (rc)
278 goto out_err_free;
279
280 idnode = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
281 &devid, lh->plh_lc_cred,
282 gfp_flags);
283 /*
284 * upon success, mirror_ds is allocated by previous
285 * getdeviceinfo, or newly by .alloc_deviceid_node
286 * nfs4_find_get_deviceid failure is indeed getdeviceinfo falure
287 */
288 if (idnode)
289 fls->mirror_array[i]->mirror_ds =
290 FF_LAYOUT_MIRROR_DS(idnode);
291 else
292 goto out_err_free;
293
294 /* efficiency */
295 rc = -EIO;
296 p = xdr_inline_decode(&stream, 4);
297 if (!p)
298 goto out_err_free;
299 fls->mirror_array[i]->efficiency = be32_to_cpup(p);
300
301 /* stateid */
302 rc = decode_stateid(&stream, &fls->mirror_array[i]->stateid);
303 if (rc)
304 goto out_err_free;
305
306 /* fh */
307 p = xdr_inline_decode(&stream, 4);
308 if (!p)
309 goto out_err_free;
310 fh_count = be32_to_cpup(p);
311
312 fls->mirror_array[i]->fh_versions =
313 kzalloc(fh_count * sizeof(struct nfs_fh),
314 gfp_flags);
315 if (fls->mirror_array[i]->fh_versions == NULL) {
316 rc = -ENOMEM;
317 goto out_err_free;
318 }
319
320 for (j = 0; j < fh_count; j++) {
321 rc = decode_nfs_fh(&stream,
322 &fls->mirror_array[i]->fh_versions[j]);
323 if (rc)
324 goto out_err_free;
325 }
326
327 fls->mirror_array[i]->fh_versions_cnt = fh_count;
328
329 /* user */
330 rc = decode_name(&stream, &fls->mirror_array[i]->uid);
331 if (rc)
332 goto out_err_free;
333
334 /* group */
335 rc = decode_name(&stream, &fls->mirror_array[i]->gid);
336 if (rc)
337 goto out_err_free;
338
339 dprintk("%s: uid %d gid %d\n", __func__,
340 fls->mirror_array[i]->uid,
341 fls->mirror_array[i]->gid);
342 }
343
344 ff_layout_sort_mirrors(fls);
345 rc = ff_layout_check_layout(lgr);
346 if (rc)
347 goto out_err_free;
348
349 ret = &fls->generic_hdr;
350 dprintk("<-- %s (success)\n", __func__);
351out_free_page:
352 __free_page(scratch);
353 return ret;
354out_err_free:
355 _ff_layout_free_lseg(fls);
356 ret = ERR_PTR(rc);
357 dprintk("<-- %s (%d)\n", __func__, rc);
358 goto out_free_page;
359}
360
361static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
362{
363 struct pnfs_layout_segment *lseg;
364
365 list_for_each_entry(lseg, &layout->plh_segs, pls_list)
366 if (lseg->pls_range.iomode == IOMODE_RW)
367 return true;
368
369 return false;
370}
371
372static void
373ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
374{
375 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
376 int i;
377
378 dprintk("--> %s\n", __func__);
379
380 for (i = 0; i < fls->mirror_array_cnt; i++) {
381 if (fls->mirror_array[i]) {
382 nfs4_ff_layout_put_deviceid(fls->mirror_array[i]->mirror_ds);
383 fls->mirror_array[i]->mirror_ds = NULL;
384 if (fls->mirror_array[i]->cred) {
385 put_rpccred(fls->mirror_array[i]->cred);
386 fls->mirror_array[i]->cred = NULL;
387 }
388 }
389 }
390
391 if (lseg->pls_range.iomode == IOMODE_RW) {
392 struct nfs4_flexfile_layout *ffl;
393 struct inode *inode;
394
395 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
396 inode = ffl->generic_hdr.plh_inode;
397 spin_lock(&inode->i_lock);
398 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
399 ffl->commit_info.nbuckets = 0;
400 kfree(ffl->commit_info.buckets);
401 ffl->commit_info.buckets = NULL;
402 }
403 spin_unlock(&inode->i_lock);
404 }
405 _ff_layout_free_lseg(fls);
406}
407
408/* Return 1 until we have multiple lsegs support */
409static int
410ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
411{
412 return 1;
413}
414
abcb7bfc
TM		 415static void
416nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer)
417{
418 ktime_t old, new;
419
420 /*
421 * Note: careful here!
422 * If the counter is zero, then we must not increment it until after
423 * we've set the start_time.
424 * If we were instead to use atomic_inc_return(), then another
425 * request might come in, bump, and then call end_busy_timer()
426 * before we've set the timer->start_time.
427 */
428 old = timer->start_time;
429 if (atomic_inc_not_zero(&timer->n_ops) == 0) {
430 new = ktime_get();
431 cmpxchg(&timer->start_time.tv64, old.tv64, new.tv64);
432 atomic_inc(&timer->n_ops);
433 }
434}
435
436static ktime_t
437nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer)
438{
439 ktime_t start, now;
440
441 now = ktime_get();
442 start.tv64 = xchg(&timer->start_time.tv64, now.tv64);
443 atomic_dec(&timer->n_ops);
444 return ktime_sub(now, start);
445}
446
447static ktime_t
448nfs4_ff_layout_calc_completion_time(struct rpc_task *task)
449{
450 return ktime_sub(ktime_get(), task->tk_start);
451}
452
453static void
d983803d
PT		 454nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
455 struct nfs4_ff_layoutstat *layoutstat)
abcb7bfc 456{
d983803d
PT		 457 static const ktime_t notime = {0};
458
abcb7bfc 459 nfs4_ff_start_busy_timer(&layoutstat->busy_timer);
d983803d 460 cmpxchg(&mirror->start_time, notime, ktime_get());
abcb7bfc
TM		 461}
462
463static void
464nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
465 __u64 requested)
466{
467 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
468
469 iostat->ops_requested++;
470 iostat->bytes_requested += requested;
471}
472
473static void
474nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
475 __u64 requested,
476 __u64 completed,
477 ktime_t time_completed)
478{
479 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
480 ktime_t timer;
481
482 iostat->ops_completed++;
483 iostat->bytes_completed += completed;
484 iostat->bytes_not_delivered += requested - completed;
485
486 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer);
487 iostat->total_busy_time =
488 ktime_add(iostat->total_busy_time, timer);
489 iostat->aggregate_completion_time =
490 ktime_add(iostat->aggregate_completion_time, time_completed);
491}
492
493static void
494nfs4_ff_layout_stat_io_start_read(struct nfs4_ff_layout_mirror *mirror,
495 __u64 requested)
496{
497 spin_lock(&mirror->lock);
d983803d 498 nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat);
abcb7bfc
TM		 499 nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
500 spin_unlock(&mirror->lock);
501}
502
503static void
504nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
505 struct nfs4_ff_layout_mirror *mirror,
506 __u64 requested,
507 __u64 completed)
508{
509 spin_lock(&mirror->lock);
510 nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
511 requested, completed,
512 nfs4_ff_layout_calc_completion_time(task));
513 spin_unlock(&mirror->lock);
514}
515
516static void
517nfs4_ff_layout_stat_io_start_write(struct nfs4_ff_layout_mirror *mirror,
518 __u64 requested)
519{
520 spin_lock(&mirror->lock);
d983803d 521 nfs4_ff_layoutstat_start_io(mirror, &mirror->write_stat);
abcb7bfc
TM		 522 nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
523 spin_unlock(&mirror->lock);
524}
525
526static void
527nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
528 struct nfs4_ff_layout_mirror *mirror,
529 __u64 requested,
530 __u64 completed,
531 enum nfs3_stable_how committed)
532{
533 if (committed == NFS_UNSTABLE)
534 requested = completed = 0;
535
536 spin_lock(&mirror->lock);
537 nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
538 requested, completed,
539 nfs4_ff_layout_calc_completion_time(task));
540 spin_unlock(&mirror->lock);
541}
542
d67ae825
TH		 543static int
544ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
545 struct nfs_commit_info *cinfo,
546 gfp_t gfp_flags)
547{
548 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
549 struct pnfs_commit_bucket *buckets;
550 int size;
551
552 if (cinfo->ds->nbuckets != 0) {
553 /* This assumes there is only one RW lseg per file.
554 * To support multiple lseg per file, we need to
555 * change struct pnfs_commit_bucket to allow dynamic
556 * increasing nbuckets.
557 */
558 return 0;
559 }
560
561 size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
562
563 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
564 gfp_flags);
565 if (!buckets)
566 return -ENOMEM;
567 else {
568 int i;
569
570 spin_lock(cinfo->lock);
571 if (cinfo->ds->nbuckets != 0)
572 kfree(buckets);
573 else {
574 cinfo->ds->buckets = buckets;
575 cinfo->ds->nbuckets = size;
576 for (i = 0; i < size; i++) {
577 INIT_LIST_HEAD(&buckets[i].written);
578 INIT_LIST_HEAD(&buckets[i].committing);
579 /* mark direct verifier as unset */
580 buckets[i].direct_verf.committed =
581 NFS_INVALID_STABLE_HOW;
582 }
583 }
584 spin_unlock(cinfo->lock);
585 return 0;
586 }
587}
588
589static struct nfs4_pnfs_ds *
590ff_layout_choose_best_ds_for_read(struct nfs_pageio_descriptor *pgio,
591 int *best_idx)
592{
593 struct nfs4_ff_layout_segment *fls;
594 struct nfs4_pnfs_ds *ds;
595 int idx;
596
597 fls = FF_LAYOUT_LSEG(pgio->pg_lseg);
598 /* mirrors are sorted by efficiency */
599 for (idx = 0; idx < fls->mirror_array_cnt; idx++) {
600 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, idx, false);
601 if (ds) {
602 *best_idx = idx;
603 return ds;
604 }
605 }
606
607 return NULL;
608}
609
610static void
611ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
612 struct nfs_page *req)
613{
614 struct nfs_pgio_mirror *pgm;
615 struct nfs4_ff_layout_mirror *mirror;
616 struct nfs4_pnfs_ds *ds;
617 int ds_idx;
618
619 /* Use full layout for now */
620 if (!pgio->pg_lseg)
621 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
622 req->wb_context,
623 0,
624 NFS4_MAX_UINT64,
625 IOMODE_READ,
626 GFP_KERNEL);
627 /* If no lseg, fall back to read through mds */
628 if (pgio->pg_lseg == NULL)
629 goto out_mds;
630
631 ds = ff_layout_choose_best_ds_for_read(pgio, &ds_idx);
632 if (!ds)
633 goto out_mds;
634 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
635
636 pgio->pg_mirror_idx = ds_idx;
637
638 /* read always uses only one mirror - idx 0 for pgio layer */
639 pgm = &pgio->pg_mirrors[0];
640 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
641
642 return;
643out_mds:
644 pnfs_put_lseg(pgio->pg_lseg);
645 pgio->pg_lseg = NULL;
646 nfs_pageio_reset_read_mds(pgio);
647}
648
649static void
650ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
651 struct nfs_page *req)
652{
653 struct nfs4_ff_layout_mirror *mirror;
654 struct nfs_pgio_mirror *pgm;
655 struct nfs_commit_info cinfo;
656 struct nfs4_pnfs_ds *ds;
657 int i;
658 int status;
659
660 if (!pgio->pg_lseg)
661 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
662 req->wb_context,
663 0,
664 NFS4_MAX_UINT64,
665 IOMODE_RW,
666 GFP_NOFS);
667 /* If no lseg, fall back to write through mds */
668 if (pgio->pg_lseg == NULL)
669 goto out_mds;
670
671 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
672 status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
673 if (status < 0)
674 goto out_mds;
675
676 /* Use a direct mapping of ds_idx to pgio mirror_idx */
677 if (WARN_ON_ONCE(pgio->pg_mirror_count !=
678 FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
679 goto out_mds;
680
681 for (i = 0; i < pgio->pg_mirror_count; i++) {
682 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
683 if (!ds)
684 goto out_mds;
685 pgm = &pgio->pg_mirrors[i];
686 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
687 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
688 }
689
690 return;
691
692out_mds:
693 pnfs_put_lseg(pgio->pg_lseg);
694 pgio->pg_lseg = NULL;
695 nfs_pageio_reset_write_mds(pgio);
696}
697
698static unsigned int
699ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
700 struct nfs_page *req)
701{
702 if (!pgio->pg_lseg)
703 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
704 req->wb_context,
705 0,
706 NFS4_MAX_UINT64,
707 IOMODE_RW,
708 GFP_NOFS);
709 if (pgio->pg_lseg)
710 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
711
712 /* no lseg means that pnfs is not in use, so no mirroring here */
713 pnfs_put_lseg(pgio->pg_lseg);
714 pgio->pg_lseg = NULL;
715 nfs_pageio_reset_write_mds(pgio);
716 return 1;
717}
718
719static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
720 .pg_init = ff_layout_pg_init_read,
721 .pg_test = pnfs_generic_pg_test,
722 .pg_doio = pnfs_generic_pg_readpages,
723 .pg_cleanup = pnfs_generic_pg_cleanup,
724};
725
726static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
727 .pg_init = ff_layout_pg_init_write,
728 .pg_test = pnfs_generic_pg_test,
729 .pg_doio = pnfs_generic_pg_writepages,
730 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
731 .pg_cleanup = pnfs_generic_pg_cleanup,
732};
733
734static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
735{
736 struct rpc_task *task = &hdr->task;
737
738 pnfs_layoutcommit_inode(hdr->inode, false);
739
740 if (retry_pnfs) {
741 dprintk("%s Reset task %5u for i/o through pNFS "
742 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
743 hdr->task.tk_pid,
744 hdr->inode->i_sb->s_id,
745 (unsigned long long)NFS_FILEID(hdr->inode),
746 hdr->args.count,
747 (unsigned long long)hdr->args.offset);
748
749 if (!hdr->dreq) {
750 struct nfs_open_context *ctx;
751
752 ctx = nfs_list_entry(hdr->pages.next)->wb_context;
753 set_bit(NFS_CONTEXT_RESEND_WRITES, &ctx->flags);
754 hdr->completion_ops->error_cleanup(&hdr->pages);
755 } else {
756 nfs_direct_set_resched_writes(hdr->dreq);
757 /* fake unstable write to let common nfs resend pages */
758 hdr->verf.committed = NFS_UNSTABLE;
759 hdr->good_bytes = 0;
760 }
761 return;
762 }
763
764 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
765 dprintk("%s Reset task %5u for i/o through MDS "
766 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
767 hdr->task.tk_pid,
768 hdr->inode->i_sb->s_id,
769 (unsigned long long)NFS_FILEID(hdr->inode),
770 hdr->args.count,
771 (unsigned long long)hdr->args.offset);
772
773 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
774 }
775}
776
777static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
778{
779 struct rpc_task *task = &hdr->task;
780
781 pnfs_layoutcommit_inode(hdr->inode, false);
782
783 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
784 dprintk("%s Reset task %5u for i/o through MDS "
785 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
786 hdr->task.tk_pid,
787 hdr->inode->i_sb->s_id,
788 (unsigned long long)NFS_FILEID(hdr->inode),
789 hdr->args.count,
790 (unsigned long long)hdr->args.offset);
791
792 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
793 }
794}
795
796static int ff_layout_async_handle_error_v4(struct rpc_task *task,
797 struct nfs4_state *state,
798 struct nfs_client *clp,
799 struct pnfs_layout_segment *lseg,
800 int idx)
801{
802 struct pnfs_layout_hdr *lo = lseg->pls_layout;
803 struct inode *inode = lo->plh_inode;
804 struct nfs_server *mds_server = NFS_SERVER(inode);
805
806 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
807 struct nfs_client *mds_client = mds_server->nfs_client;
808 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
809
810 if (task->tk_status >= 0)
811 return 0;
812
813 switch (task->tk_status) {
814 /* MDS state errors */
815 case -NFS4ERR_DELEG_REVOKED:
816 case -NFS4ERR_ADMIN_REVOKED:
817 case -NFS4ERR_BAD_STATEID:
818 if (state == NULL)
819 break;
820 nfs_remove_bad_delegation(state->inode);
821 case -NFS4ERR_OPENMODE:
822 if (state == NULL)
823 break;
824 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
825 goto out_bad_stateid;
826 goto wait_on_recovery;
827 case -NFS4ERR_EXPIRED:
828 if (state != NULL) {
829 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
830 goto out_bad_stateid;
831 }
832 nfs4_schedule_lease_recovery(mds_client);
833 goto wait_on_recovery;
834 /* DS session errors */
835 case -NFS4ERR_BADSESSION:
836 case -NFS4ERR_BADSLOT:
837 case -NFS4ERR_BAD_HIGH_SLOT:
838 case -NFS4ERR_DEADSESSION:
839 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
840 case -NFS4ERR_SEQ_FALSE_RETRY:
841 case -NFS4ERR_SEQ_MISORDERED:
842 dprintk("%s ERROR %d, Reset session. Exchangeid "
843 "flags 0x%x\n", __func__, task->tk_status,
844 clp->cl_exchange_flags);
845 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
846 break;
847 case -NFS4ERR_DELAY:
848 case -NFS4ERR_GRACE:
849 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
850 break;
851 case -NFS4ERR_RETRY_UNCACHED_REP:
852 break;
853 /* Invalidate Layout errors */
854 case -NFS4ERR_PNFS_NO_LAYOUT:
855 case -ESTALE: /* mapped NFS4ERR_STALE */
856 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
857 case -EISDIR: /* mapped NFS4ERR_ISDIR */
858 case -NFS4ERR_FHEXPIRED:
859 case -NFS4ERR_WRONG_TYPE:
860 dprintk("%s Invalid layout error %d\n", __func__,
861 task->tk_status);
862 /*
863 * Destroy layout so new i/o will get a new layout.
864 * Layout will not be destroyed until all current lseg
865 * references are put. Mark layout as invalid to resend failed
866 * i/o and all i/o waiting on the slot table to the MDS until
867 * layout is destroyed and a new valid layout is obtained.
868 */
869 pnfs_destroy_layout(NFS_I(inode));
870 rpc_wake_up(&tbl->slot_tbl_waitq);
871 goto reset;
872 /* RPC connection errors */
873 case -ECONNREFUSED:
874 case -EHOSTDOWN:
875 case -EHOSTUNREACH:
876 case -ENETUNREACH:
877 case -EIO:
878 case -ETIMEDOUT:
879 case -EPIPE:
880 dprintk("%s DS connection error %d\n", __func__,
881 task->tk_status);
882 nfs4_mark_deviceid_unavailable(devid);
883 rpc_wake_up(&tbl->slot_tbl_waitq);
884 /* fall through */
885 default:
886 if (ff_layout_has_available_ds(lseg))
887 return -NFS4ERR_RESET_TO_PNFS;
888reset:
889 dprintk("%s Retry through MDS. Error %d\n", __func__,
890 task->tk_status);
891 return -NFS4ERR_RESET_TO_MDS;
892 }
893out:
894 task->tk_status = 0;
895 return -EAGAIN;
896out_bad_stateid:
897 task->tk_status = -EIO;
898 return 0;
899wait_on_recovery:
900 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
901 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
902 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
903 goto out;
904}
905
906/* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
907static int ff_layout_async_handle_error_v3(struct rpc_task *task,
908 struct pnfs_layout_segment *lseg,
909 int idx)
910{
911 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
912
913 if (task->tk_status >= 0)
914 return 0;
915
916 if (task->tk_status != -EJUKEBOX) {
917 dprintk("%s DS connection error %d\n", __func__,
918 task->tk_status);
919 nfs4_mark_deviceid_unavailable(devid);
920 if (ff_layout_has_available_ds(lseg))
921 return -NFS4ERR_RESET_TO_PNFS;
922 else
923 return -NFS4ERR_RESET_TO_MDS;
924 }
925
926 if (task->tk_status == -EJUKEBOX)
927 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
928 task->tk_status = 0;
929 rpc_restart_call(task);
930 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
931 return -EAGAIN;
932}
933
934static int ff_layout_async_handle_error(struct rpc_task *task,
935 struct nfs4_state *state,
936 struct nfs_client *clp,
937 struct pnfs_layout_segment *lseg,
938 int idx)
939{
940 int vers = clp->cl_nfs_mod->rpc_vers->number;
941
942 switch (vers) {
943 case 3:
944 return ff_layout_async_handle_error_v3(task, lseg, idx);
945 case 4:
946 return ff_layout_async_handle_error_v4(task, state, clp,
947 lseg, idx);
948 default:
949 /* should never happen */
950 WARN_ON_ONCE(1);
951 return 0;
952 }
953}
954
955static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
956 int idx, u64 offset, u64 length,
957 u32 status, int opnum)
958{
959 struct nfs4_ff_layout_mirror *mirror;
960 int err;
961
962 mirror = FF_LAYOUT_COMP(lseg, idx);
963 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
964 mirror, offset, length, status, opnum,
965 GFP_NOIO);
966 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
967}
968
969/* NFS_PROTO call done callback routines */
970
971static int ff_layout_read_done_cb(struct rpc_task *task,
972 struct nfs_pgio_header *hdr)
973{
974 struct inode *inode;
975 int err;
976
977 trace_nfs4_pnfs_read(hdr, task->tk_status);
978 if (task->tk_status == -ETIMEDOUT && !hdr->res.op_status)
979 hdr->res.op_status = NFS4ERR_NXIO;
980 if (task->tk_status < 0 && hdr->res.op_status)
981 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
982 hdr->args.offset, hdr->args.count,
983 hdr->res.op_status, OP_READ);
984 err = ff_layout_async_handle_error(task, hdr->args.context->state,
985 hdr->ds_clp, hdr->lseg,
986 hdr->pgio_mirror_idx);
987
988 switch (err) {
989 case -NFS4ERR_RESET_TO_PNFS:
990 set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
991 &hdr->lseg->pls_layout->plh_flags);
992 pnfs_read_resend_pnfs(hdr);
993 return task->tk_status;
994 case -NFS4ERR_RESET_TO_MDS:
995 inode = hdr->lseg->pls_layout->plh_inode;
996 pnfs_error_mark_layout_for_return(inode, hdr->lseg);
997 ff_layout_reset_read(hdr);
998 return task->tk_status;
999 case -EAGAIN:
1000 rpc_restart_call_prepare(task);
1001 return -EAGAIN;
1002 }
1003
1004 return 0;
1005}
1006
1007/*
1008 * We reference the rpc_cred of the first WRITE that triggers the need for
1009 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1010 * rfc5661 is not clear about which credential should be used.
1011 *
1012 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1013 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1014 * we always send layoutcommit after DS writes.
1015 */
1016static void
1017ff_layout_set_layoutcommit(struct nfs_pgio_header *hdr)
1018{
67af7611
TM		 1019 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1020 hdr->mds_offset + hdr->res.count);
d67ae825
TH		 1021 dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
1022 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
1023}
1024
1025static bool
1026ff_layout_reset_to_mds(struct pnfs_layout_segment *lseg, int idx)
1027{
1028 /* No mirroring for now */
1029 struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1030
1031 return ff_layout_test_devid_unavailable(node);
1032}
1033
1034static int ff_layout_read_prepare_common(struct rpc_task *task,
1035 struct nfs_pgio_header *hdr)
1036{
abcb7bfc
TM		 1037 nfs4_ff_layout_stat_io_start_read(
1038 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1039 hdr->args.count);
1040
d67ae825
TH		 1041 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1042 rpc_exit(task, -EIO);
1043 return -EIO;
1044 }
1045 if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1046 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
1047 if (ff_layout_has_available_ds(hdr->lseg))
1048 pnfs_read_resend_pnfs(hdr);
1049 else
1050 ff_layout_reset_read(hdr);
1051 rpc_exit(task, 0);
1052 return -EAGAIN;
1053 }
1054 hdr->pgio_done_cb = ff_layout_read_done_cb;
1055
1056 return 0;
1057}
1058
1059/*
1060 * Call ops for the async read/write cases
1061 * In the case of dense layouts, the offset needs to be reset to its
1062 * original value.
1063 */
1064static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1065{
1066 struct nfs_pgio_header *hdr = data;
1067
1068 if (ff_layout_read_prepare_common(task, hdr))
1069 return;
1070
1071 rpc_call_start(task);
1072}
1073
1074static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1075 struct nfs4_sequence_args *args,
1076 struct nfs4_sequence_res *res,
1077 struct rpc_task *task)
1078{
1079 if (ds_clp->cl_session)
1080 return nfs41_setup_sequence(ds_clp->cl_session,
1081 args,
1082 res,
1083 task);
1084 return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1085 args,
1086 res,
1087 task);
1088}
1089
1090static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1091{
1092 struct nfs_pgio_header *hdr = data;
1093
d67ae825
TH		 1094 if (ff_layout_setup_sequence(hdr->ds_clp,
1095 &hdr->args.seq_args,
1096 &hdr->res.seq_res,
1097 task))
1098 return;
1099
abcb7bfc
TM		 1100 if (ff_layout_read_prepare_common(task, hdr))
1101 return;
1102
d67ae825
TH		 1103 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1104 hdr->args.lock_context, FMODE_READ) == -EIO)
1105 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1106}
1107
1108static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1109{
1110 struct nfs_pgio_header *hdr = data;
1111
1112 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1113
abcb7bfc
TM		 1114 nfs4_ff_layout_stat_io_end_read(task,
1115 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1116 hdr->args.count, hdr->res.count);
1117
d67ae825
TH		 1118 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1119 task->tk_status == 0) {
1120 nfs4_sequence_done(task, &hdr->res.seq_res);
1121 return;
1122 }
1123
1124 /* Note this may cause RPC to be resent */
1125 hdr->mds_ops->rpc_call_done(task, hdr);
1126}
1127
1128static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1129{
1130 struct nfs_pgio_header *hdr = data;
1131
1132 rpc_count_iostats_metrics(task,
1133 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1134}
1135
1136static int ff_layout_write_done_cb(struct rpc_task *task,
1137 struct nfs_pgio_header *hdr)
1138{
1139 struct inode *inode;
1140 int err;
1141
1142 trace_nfs4_pnfs_write(hdr, task->tk_status);
1143 if (task->tk_status == -ETIMEDOUT && !hdr->res.op_status)
1144 hdr->res.op_status = NFS4ERR_NXIO;
1145 if (task->tk_status < 0 && hdr->res.op_status)
1146 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1147 hdr->args.offset, hdr->args.count,
1148 hdr->res.op_status, OP_WRITE);
1149 err = ff_layout_async_handle_error(task, hdr->args.context->state,
1150 hdr->ds_clp, hdr->lseg,
1151 hdr->pgio_mirror_idx);
1152
1153 switch (err) {
1154 case -NFS4ERR_RESET_TO_PNFS:
1155 case -NFS4ERR_RESET_TO_MDS:
1156 inode = hdr->lseg->pls_layout->plh_inode;
1157 pnfs_error_mark_layout_for_return(inode, hdr->lseg);
1158 if (err == -NFS4ERR_RESET_TO_PNFS) {
1159 pnfs_set_retry_layoutget(hdr->lseg->pls_layout);
1160 ff_layout_reset_write(hdr, true);
1161 } else {
1162 pnfs_clear_retry_layoutget(hdr->lseg->pls_layout);
1163 ff_layout_reset_write(hdr, false);
1164 }
1165 return task->tk_status;
1166 case -EAGAIN:
1167 rpc_restart_call_prepare(task);
1168 return -EAGAIN;
1169 }
1170
1171 if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1172 hdr->res.verf->committed == NFS_DATA_SYNC)
1173 ff_layout_set_layoutcommit(hdr);
1174
1175 return 0;
1176}
1177
1178static int ff_layout_commit_done_cb(struct rpc_task *task,
1179 struct nfs_commit_data *data)
1180{
1181 struct inode *inode;
1182 int err;
1183
1184 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1185 if (task->tk_status == -ETIMEDOUT && !data->res.op_status)
1186 data->res.op_status = NFS4ERR_NXIO;
1187 if (task->tk_status < 0 && data->res.op_status)
1188 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1189 data->args.offset, data->args.count,
1190 data->res.op_status, OP_COMMIT);
1191 err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1192 data->lseg, data->ds_commit_index);
1193
1194 switch (err) {
1195 case -NFS4ERR_RESET_TO_PNFS:
1196 case -NFS4ERR_RESET_TO_MDS:
1197 inode = data->lseg->pls_layout->plh_inode;
1198 pnfs_error_mark_layout_for_return(inode, data->lseg);
1199 if (err == -NFS4ERR_RESET_TO_PNFS)
1200 pnfs_set_retry_layoutget(data->lseg->pls_layout);
1201 else
1202 pnfs_clear_retry_layoutget(data->lseg->pls_layout);
1203 pnfs_generic_prepare_to_resend_writes(data);
1204 return -EAGAIN;
1205 case -EAGAIN:
1206 rpc_restart_call_prepare(task);
1207 return -EAGAIN;
1208 }
1209
1210 if (data->verf.committed == NFS_UNSTABLE)
67af7611 1211 pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
d67ae825
TH		 1212
1213 return 0;
1214}
1215
1216static int ff_layout_write_prepare_common(struct rpc_task *task,
1217 struct nfs_pgio_header *hdr)
1218{
abcb7bfc
TM		 1219 nfs4_ff_layout_stat_io_start_write(
1220 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1221 hdr->args.count);
1222
d67ae825
TH		 1223 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1224 rpc_exit(task, -EIO);
1225 return -EIO;
1226 }
1227
1228 if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1229 bool retry_pnfs;
1230
1231 retry_pnfs = ff_layout_has_available_ds(hdr->lseg);
1232 dprintk("%s task %u reset io to %s\n", __func__,
1233 task->tk_pid, retry_pnfs ? "pNFS" : "MDS");
1234 ff_layout_reset_write(hdr, retry_pnfs);
1235 rpc_exit(task, 0);
1236 return -EAGAIN;
1237 }
1238
1239 return 0;
1240}
1241
1242static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1243{
1244 struct nfs_pgio_header *hdr = data;
1245
1246 if (ff_layout_write_prepare_common(task, hdr))
1247 return;
1248
1249 rpc_call_start(task);
1250}
1251
1252static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1253{
1254 struct nfs_pgio_header *hdr = data;
1255
d67ae825
TH		 1256 if (ff_layout_setup_sequence(hdr->ds_clp,
1257 &hdr->args.seq_args,
1258 &hdr->res.seq_res,
1259 task))
1260 return;
1261
abcb7bfc
TM		 1262 if (ff_layout_write_prepare_common(task, hdr))
1263 return;
1264
d67ae825
TH		 1265 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1266 hdr->args.lock_context, FMODE_WRITE) == -EIO)
1267 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1268}
1269
1270static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1271{
1272 struct nfs_pgio_header *hdr = data;
1273
abcb7bfc
TM		 1274 nfs4_ff_layout_stat_io_end_write(task,
1275 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1276 hdr->args.count, hdr->res.count,
1277 hdr->res.verf->committed);
1278
d67ae825
TH		 1279 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1280 task->tk_status == 0) {
1281 nfs4_sequence_done(task, &hdr->res.seq_res);
1282 return;
1283 }
1284
1285 /* Note this may cause RPC to be resent */
1286 hdr->mds_ops->rpc_call_done(task, hdr);
1287}
1288
1289static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1290{
1291 struct nfs_pgio_header *hdr = data;
1292
1293 rpc_count_iostats_metrics(task,
1294 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1295}
1296
abcb7bfc
TM		 1297static void ff_layout_commit_prepare_common(struct rpc_task *task,
1298 struct nfs_commit_data *cdata)
1299{
1300 nfs4_ff_layout_stat_io_start_write(
1301 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1302 0);
1303}
1304
d67ae825
TH		 1305static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1306{
abcb7bfc 1307 ff_layout_commit_prepare_common(task, data);
d67ae825
TH		 1308 rpc_call_start(task);
1309}
1310
1311static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1312{
1313 struct nfs_commit_data *wdata = data;
1314
abcb7bfc 1315 if (ff_layout_setup_sequence(wdata->ds_clp,
d67ae825
TH		 1316 &wdata->args.seq_args,
1317 &wdata->res.seq_res,
abcb7bfc
TM		 1318 task))
1319 return;
1320 ff_layout_commit_prepare_common(task, data);
1321}
1322
1323static void ff_layout_commit_done(struct rpc_task *task, void *data)
1324{
1325 struct nfs_commit_data *cdata = data;
1326 struct nfs_page *req;
1327 __u64 count = 0;
1328
1329 if (task->tk_status == 0) {
1330 list_for_each_entry(req, &cdata->pages, wb_list)
1331 count += req->wb_bytes;
1332 }
1333
1334 nfs4_ff_layout_stat_io_end_write(task,
1335 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1336 count, count, NFS_FILE_SYNC);
1337
1338 pnfs_generic_write_commit_done(task, data);
d67ae825
TH		 1339}
1340
1341static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1342{
1343 struct nfs_commit_data *cdata = data;
1344
1345 rpc_count_iostats_metrics(task,
1346 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1347}
1348
1349static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1350 .rpc_call_prepare = ff_layout_read_prepare_v3,
1351 .rpc_call_done = ff_layout_read_call_done,
1352 .rpc_count_stats = ff_layout_read_count_stats,
1353 .rpc_release = pnfs_generic_rw_release,
1354};
1355
1356static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1357 .rpc_call_prepare = ff_layout_read_prepare_v4,
1358 .rpc_call_done = ff_layout_read_call_done,
1359 .rpc_count_stats = ff_layout_read_count_stats,
1360 .rpc_release = pnfs_generic_rw_release,
1361};
1362
1363static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1364 .rpc_call_prepare = ff_layout_write_prepare_v3,
1365 .rpc_call_done = ff_layout_write_call_done,
1366 .rpc_count_stats = ff_layout_write_count_stats,
1367 .rpc_release = pnfs_generic_rw_release,
1368};
1369
1370static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1371 .rpc_call_prepare = ff_layout_write_prepare_v4,
1372 .rpc_call_done = ff_layout_write_call_done,
1373 .rpc_count_stats = ff_layout_write_count_stats,
1374 .rpc_release = pnfs_generic_rw_release,
1375};
1376
1377static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1378 .rpc_call_prepare = ff_layout_commit_prepare_v3,
abcb7bfc 1379 .rpc_call_done = ff_layout_commit_done,
d67ae825
TH		 1380 .rpc_count_stats = ff_layout_commit_count_stats,
1381 .rpc_release = pnfs_generic_commit_release,
1382};
1383
1384static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1385 .rpc_call_prepare = ff_layout_commit_prepare_v4,
abcb7bfc 1386 .rpc_call_done = ff_layout_commit_done,
d67ae825
TH		 1387 .rpc_count_stats = ff_layout_commit_count_stats,
1388 .rpc_release = pnfs_generic_commit_release,
1389};
1390
1391static enum pnfs_try_status
1392ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1393{
1394 struct pnfs_layout_segment *lseg = hdr->lseg;
1395 struct nfs4_pnfs_ds *ds;
1396 struct rpc_clnt *ds_clnt;
1397 struct rpc_cred *ds_cred;
1398 loff_t offset = hdr->args.offset;
1399 u32 idx = hdr->pgio_mirror_idx;
1400 int vers;
1401 struct nfs_fh *fh;
1402
1403 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1404 __func__, hdr->inode->i_ino,
1405 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1406
1407 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1408 if (!ds)
1409 goto out_failed;
1410
1411 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1412 hdr->inode);
1413 if (IS_ERR(ds_clnt))
1414 goto out_failed;
1415
1416 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1417 if (IS_ERR(ds_cred))
1418 goto out_failed;
1419
1420 vers = nfs4_ff_layout_ds_version(lseg, idx);
1421
1422 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1423 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1424
1425 atomic_inc(&ds->ds_clp->cl_count);
1426 hdr->ds_clp = ds->ds_clp;
1427 fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1428 if (fh)
1429 hdr->args.fh = fh;
d67ae825
TH		 1430 /*
1431 * Note that if we ever decide to split across DSes,
1432 * then we may need to handle dense-like offsets.
1433 */
1434 hdr->args.offset = offset;
1435 hdr->mds_offset = offset;
1436
1437 /* Perform an asynchronous read to ds */
1438 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1439 vers == 3 ? &ff_layout_read_call_ops_v3 :
1440 &ff_layout_read_call_ops_v4,
1441 0, RPC_TASK_SOFTCONN);
1442
1443 return PNFS_ATTEMPTED;
1444
1445out_failed:
1446 if (ff_layout_has_available_ds(lseg))
1447 return PNFS_TRY_AGAIN;
1448 return PNFS_NOT_ATTEMPTED;
1449}
1450
1451/* Perform async writes. */
1452static enum pnfs_try_status
1453ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1454{
1455 struct pnfs_layout_segment *lseg = hdr->lseg;
1456 struct nfs4_pnfs_ds *ds;
1457 struct rpc_clnt *ds_clnt;
1458 struct rpc_cred *ds_cred;
1459 loff_t offset = hdr->args.offset;
1460 int vers;
1461 struct nfs_fh *fh;
1462 int idx = hdr->pgio_mirror_idx;
1463
1464 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1465 if (!ds)
1466 return PNFS_NOT_ATTEMPTED;
1467
1468 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1469 hdr->inode);
1470 if (IS_ERR(ds_clnt))
1471 return PNFS_NOT_ATTEMPTED;
1472
1473 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1474 if (IS_ERR(ds_cred))
1475 return PNFS_NOT_ATTEMPTED;
1476
1477 vers = nfs4_ff_layout_ds_version(lseg, idx);
1478
1479 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1480 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1481 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1482 vers);
1483
1484 hdr->pgio_done_cb = ff_layout_write_done_cb;
1485 atomic_inc(&ds->ds_clp->cl_count);
1486 hdr->ds_clp = ds->ds_clp;
1487 hdr->ds_commit_idx = idx;
1488 fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1489 if (fh)
1490 hdr->args.fh = fh;
1491
1492 /*
1493 * Note that if we ever decide to split across DSes,
1494 * then we may need to handle dense-like offsets.
1495 */
1496 hdr->args.offset = offset;
1497
1498 /* Perform an asynchronous write */
1499 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1500 vers == 3 ? &ff_layout_write_call_ops_v3 :
1501 &ff_layout_write_call_ops_v4,
1502 sync, RPC_TASK_SOFTCONN);
1503 return PNFS_ATTEMPTED;
1504}
1505
d67ae825
TH		 1506static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1507{
1508 return i;
1509}
1510
1511static struct nfs_fh *
1512select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1513{
1514 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1515
1516 /* FIXME: Assume that there is only one NFS version available
1517 * for the DS.
1518 */
1519 return &flseg->mirror_array[i]->fh_versions[0];
1520}
1521
1522static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1523{
1524 struct pnfs_layout_segment *lseg = data->lseg;
1525 struct nfs4_pnfs_ds *ds;
1526 struct rpc_clnt *ds_clnt;
1527 struct rpc_cred *ds_cred;
1528 u32 idx;
1529 int vers;
1530 struct nfs_fh *fh;
1531
1532 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1533 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1534 if (!ds)
1535 goto out_err;
1536
1537 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1538 data->inode);
1539 if (IS_ERR(ds_clnt))
1540 goto out_err;
1541
1542 ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1543 if (IS_ERR(ds_cred))
1544 goto out_err;
1545
1546 vers = nfs4_ff_layout_ds_version(lseg, idx);
1547
1548 dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1549 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1550 vers);
1551 data->commit_done_cb = ff_layout_commit_done_cb;
1552 data->cred = ds_cred;
1553 atomic_inc(&ds->ds_clp->cl_count);
1554 data->ds_clp = ds->ds_clp;
1555 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1556 if (fh)
1557 data->args.fh = fh;
abcb7bfc 1558
d67ae825
TH		 1559 return nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1560 vers == 3 ? &ff_layout_commit_call_ops_v3 :
1561 &ff_layout_commit_call_ops_v4,
1562 how, RPC_TASK_SOFTCONN);
1563out_err:
1564 pnfs_generic_prepare_to_resend_writes(data);
1565 pnfs_generic_commit_release(data);
1566 return -EAGAIN;
1567}
1568
1569static int
1570ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1571 int how, struct nfs_commit_info *cinfo)
1572{
1573 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1574 ff_layout_initiate_commit);
1575}
1576
1577static struct pnfs_ds_commit_info *
1578ff_layout_get_ds_info(struct inode *inode)
1579{
1580 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1581
1582 if (layout == NULL)
1583 return NULL;
1584
1585 return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1586}
1587
1588static void
fc87701b 1589ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
d67ae825
TH		 1590{
1591 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1592 id_node));
1593}
1594
1595static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1596 struct xdr_stream *xdr,
1597 const struct nfs4_layoutreturn_args *args)
1598{
1599 struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1600 __be32 *start;
1601 int count = 0, ret = 0;
1602
1603 start = xdr_reserve_space(xdr, 4);
1604 if (unlikely(!start))
1605 return -E2BIG;
1606
1607 /* This assume we always return _ALL_ layouts */
1608 spin_lock(&hdr->plh_inode->i_lock);
1609 ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1610 spin_unlock(&hdr->plh_inode->i_lock);
1611
1612 *start = cpu_to_be32(count);
1613
1614 return ret;
1615}
1616
1617/* report nothing for now */
1618static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1619 struct xdr_stream *xdr,
1620 const struct nfs4_layoutreturn_args *args)
1621{
1622 __be32 *p;
1623
1624 p = xdr_reserve_space(xdr, 4);
1625 if (likely(p))
1626 *p = cpu_to_be32(0);
1627}
1628
1629static struct nfs4_deviceid_node *
1630ff_layout_alloc_deviceid_node(struct nfs_server *server,
1631 struct pnfs_device *pdev, gfp_t gfp_flags)
1632{
1633 struct nfs4_ff_layout_ds *dsaddr;
1634
1635 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
1636 if (!dsaddr)
1637 return NULL;
1638 return &dsaddr->id_node;
1639}
1640
1641static void
1642ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
1643 struct xdr_stream *xdr,
1644 const struct nfs4_layoutreturn_args *args)
1645{
1646 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
1647 __be32 *start;
1648
1649 dprintk("%s: Begin\n", __func__);
1650 start = xdr_reserve_space(xdr, 4);
1651 BUG_ON(!start);
1652
1653 if (ff_layout_encode_ioerr(flo, xdr, args))
1654 goto out;
1655
1656 ff_layout_encode_iostats(flo, xdr, args);
1657out:
1658 *start = cpu_to_be32((xdr->p - start - 1) * 4);
1659 dprintk("%s: Return\n", __func__);
1660}
1661
1662static struct pnfs_layoutdriver_type flexfilelayout_type = {
1663 .id = LAYOUT_FLEX_FILES,
1664 .name = "LAYOUT_FLEX_FILES",
1665 .owner = THIS_MODULE,
1666 .alloc_layout_hdr = ff_layout_alloc_layout_hdr,
1667 .free_layout_hdr = ff_layout_free_layout_hdr,
1668 .alloc_lseg = ff_layout_alloc_lseg,
1669 .free_lseg = ff_layout_free_lseg,
1670 .pg_read_ops = &ff_layout_pg_read_ops,
1671 .pg_write_ops = &ff_layout_pg_write_ops,
1672 .get_ds_info = ff_layout_get_ds_info,
fc87701b 1673 .free_deviceid_node = ff_layout_free_deviceid_node,
338d00cf 1674 .mark_request_commit = pnfs_layout_mark_request_commit,
d67ae825
TH		 1675 .clear_request_commit = pnfs_generic_clear_request_commit,
1676 .scan_commit_lists = pnfs_generic_scan_commit_lists,
1677 .recover_commit_reqs = pnfs_generic_recover_commit_reqs,
1678 .commit_pagelist = ff_layout_commit_pagelist,
1679 .read_pagelist = ff_layout_read_pagelist,
1680 .write_pagelist = ff_layout_write_pagelist,
1681 .alloc_deviceid_node = ff_layout_alloc_deviceid_node,
1682 .encode_layoutreturn = ff_layout_encode_layoutreturn,
5bb89b47 1683 .sync = pnfs_nfs_generic_sync,
d67ae825
TH		 1684};
1685
1686static int __init nfs4flexfilelayout_init(void)
1687{
1688 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
1689 __func__);
1690 return pnfs_register_layoutdriver(&flexfilelayout_type);
1691}
1692
1693static void __exit nfs4flexfilelayout_exit(void)
1694{
1695 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
1696 __func__);
1697 pnfs_unregister_layoutdriver(&flexfilelayout_type);
1698}
1699
1700MODULE_ALIAS("nfs-layouttype4-4");
1701
1702MODULE_LICENSE("GPL");
1703MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
1704
1705module_init(nfs4flexfilelayout_init);
1706module_exit(nfs4flexfilelayout_exit);
Michael's Linux tree.