1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Module for pnfs flexfile layout driver.
4 *
5 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
6 *
7 * Tao Peng <bergwolf@primarydata.com>
8 */
9
10 #include <linux/nfs_fs.h>
11 #include <linux/nfs_mount.h>
12 #include <linux/nfs_page.h>
13 #include <linux/module.h>
14 #include <linux/file.h>
15 #include <linux/sched/mm.h>
16
17 #include <linux/sunrpc/metrics.h>
18
19 #include "flexfilelayout.h"
20 #include "../nfs4session.h"
21 #include "../nfs4idmap.h"
22 #include "../internal.h"
23 #include "../delegation.h"
24 #include "../nfs4trace.h"
25 #include "../iostat.h"
26 #include "../nfs.h"
27 #include "../nfs42.h"
28
29 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
30
31 #define FF_LAYOUT_POLL_RETRY_MAX (15*HZ)
32 #define FF_LAYOUTRETURN_MAXERR 20
33
34 enum nfs4_ff_op_type {
35 NFS4_FF_OP_LAYOUTSTATS,
36 NFS4_FF_OP_LAYOUTRETURN,
37 };
38
39 static unsigned short io_maxretrans;
40
41 static const struct pnfs_commit_ops ff_layout_commit_ops;
42 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
43 struct nfs_pgio_header *hdr);
44 static int
45 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo,
46 struct nfs42_layoutstat_devinfo *devinfo,
47 int dev_limit, enum nfs4_ff_op_type type);
48 static void ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr,
49 const struct nfs42_layoutstat_devinfo *devinfo,
50 struct nfs4_ff_layout_ds_stripe *dss_info);
51
52 static struct pnfs_layout_hdr *
ff_layout_alloc_layout_hdr(struct inode * inode,gfp_t gfp_flags)53 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
54 {
55 struct nfs4_flexfile_layout *ffl;
56
57 ffl = kzalloc_obj(*ffl, gfp_flags);
58 if (ffl) {
59 pnfs_init_ds_commit_info(&ffl->commit_info);
60 INIT_LIST_HEAD(&ffl->error_list);
61 INIT_LIST_HEAD(&ffl->mirrors);
62 ffl->last_report_time = ktime_get();
63 ffl->commit_info.ops = &ff_layout_commit_ops;
64 return &ffl->generic_hdr;
65 } else
66 return NULL;
67 }
68
69 static void
ff_layout_free_layout_hdr(struct pnfs_layout_hdr * lo)70 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
71 {
72 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(lo);
73 struct nfs4_ff_layout_ds_err *err, *n;
74
75 list_for_each_entry_safe(err, n, &ffl->error_list, list) {
76 list_del(&err->list);
77 kfree(err);
78 }
79 kfree_rcu(ffl, generic_hdr.plh_rcu);
80 }
81
decode_pnfs_stateid(struct xdr_stream * xdr,nfs4_stateid * stateid)82 static int decode_pnfs_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
83 {
84 __be32 *p;
85
86 p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
87 if (unlikely(p == NULL))
88 return -ENOBUFS;
89 stateid->type = NFS4_PNFS_DS_STATEID_TYPE;
90 memcpy(stateid->data, p, NFS4_STATEID_SIZE);
91 dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
92 p[0], p[1], p[2], p[3]);
93 return 0;
94 }
95
decode_deviceid(struct xdr_stream * xdr,struct nfs4_deviceid * devid)96 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
97 {
98 __be32 *p;
99
100 p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
101 if (unlikely(!p))
102 return -ENOBUFS;
103 memcpy(devid, p, NFS4_DEVICEID4_SIZE);
104 nfs4_print_deviceid(devid);
105 return 0;
106 }
107
decode_nfs_fh(struct xdr_stream * xdr,struct nfs_fh * fh)108 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
109 {
110 __be32 *p;
111
112 p = xdr_inline_decode(xdr, 4);
113 if (unlikely(!p))
114 return -ENOBUFS;
115 fh->size = be32_to_cpup(p++);
116 if (fh->size > NFS_MAXFHSIZE) {
117 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
118 fh->size);
119 return -EOVERFLOW;
120 }
121 /* fh.data */
122 p = xdr_inline_decode(xdr, fh->size);
123 if (unlikely(!p))
124 return -ENOBUFS;
125 memcpy(&fh->data, p, fh->size);
126 dprintk("%s: fh len %d\n", __func__, fh->size);
127
128 return 0;
129 }
130
131 /*
132 * Currently only stringified uids and gids are accepted.
133 * I.e., kerberos is not supported to the DSes, so no pricipals.
134 *
135 * That means that one common function will suffice, but when
136 * principals are added, this should be split to accomodate
137 * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
138 */
139 static int
decode_name(struct xdr_stream * xdr,u32 * id)140 decode_name(struct xdr_stream *xdr, u32 *id)
141 {
142 __be32 *p;
143 int len;
144
145 /* opaque_length(4)*/
146 p = xdr_inline_decode(xdr, 4);
147 if (unlikely(!p))
148 return -ENOBUFS;
149 len = be32_to_cpup(p++);
150 if (len < 0)
151 return -EINVAL;
152
153 dprintk("%s: len %u\n", __func__, len);
154
155 /* opaque body */
156 p = xdr_inline_decode(xdr, len);
157 if (unlikely(!p))
158 return -ENOBUFS;
159
160 if (!nfs_map_string_to_numeric((char *)p, len, id))
161 return -EINVAL;
162
163 return 0;
164 }
165
166 static struct nfsd_file *
ff_local_open_fh(struct pnfs_layout_segment * lseg,u32 ds_idx,u32 dss_id,struct nfs_client * clp,const struct cred * cred,struct nfs_fh * fh,fmode_t mode)167 ff_local_open_fh(struct pnfs_layout_segment *lseg, u32 ds_idx, u32 dss_id,
168 struct nfs_client *clp, const struct cred *cred,
169 struct nfs_fh *fh, fmode_t mode)
170 {
171 #if IS_ENABLED(CONFIG_NFS_LOCALIO)
172 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
173
174 return nfs_local_open_fh(clp, cred, fh, &mirror->dss[dss_id].nfl, mode);
175 #else
176 return NULL;
177 #endif
178 }
179
ff_dss_match_fh(const struct nfs4_ff_layout_ds_stripe * dss1,const struct nfs4_ff_layout_ds_stripe * dss2)180 static bool ff_dss_match_fh(const struct nfs4_ff_layout_ds_stripe *dss1,
181 const struct nfs4_ff_layout_ds_stripe *dss2)
182 {
183 int i, j;
184
185 if (dss1->fh_versions_cnt != dss2->fh_versions_cnt)
186 return false;
187
188 for (i = 0; i < dss1->fh_versions_cnt; i++) {
189 bool found_fh = false;
190 for (j = 0; j < dss2->fh_versions_cnt; j++) {
191 if (nfs_compare_fh(&dss1->fh_versions[i],
192 &dss2->fh_versions[j]) == 0) {
193 found_fh = true;
194 break;
195 }
196 }
197 if (!found_fh)
198 return false;
199 }
200 return true;
201 }
202
ff_mirror_match_fh(const struct nfs4_ff_layout_mirror * m1,const struct nfs4_ff_layout_mirror * m2)203 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1,
204 const struct nfs4_ff_layout_mirror *m2)
205 {
206 u32 dss_id;
207
208 if (m1->dss_count != m2->dss_count)
209 return false;
210
211 for (dss_id = 0; dss_id < m1->dss_count; dss_id++)
212 if (!ff_dss_match_fh(&m1->dss[dss_id], &m2->dss[dss_id]))
213 return false;
214
215 return true;
216 }
217
ff_mirror_match_devid(const struct nfs4_ff_layout_mirror * m1,const struct nfs4_ff_layout_mirror * m2)218 static bool ff_mirror_match_devid(const struct nfs4_ff_layout_mirror *m1,
219 const struct nfs4_ff_layout_mirror *m2)
220 {
221 u32 dss_id;
222
223 if (m1->dss_count != m2->dss_count)
224 return false;
225
226 for (dss_id = 0; dss_id < m1->dss_count; dss_id++)
227 if (memcmp(&m1->dss[dss_id].devid,
228 &m2->dss[dss_id].devid,
229 sizeof(m1->dss[dss_id].devid)) != 0)
230 return false;
231
232 return true;
233 }
234
235 static struct nfs4_ff_layout_mirror *
ff_layout_add_mirror(struct pnfs_layout_hdr * lo,struct nfs4_ff_layout_mirror * mirror)236 ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
237 struct nfs4_ff_layout_mirror *mirror)
238 {
239 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
240 struct nfs4_ff_layout_mirror *pos;
241 struct inode *inode = lo->plh_inode;
242
243 spin_lock(&inode->i_lock);
244 list_for_each_entry(pos, &ff_layout->mirrors, mirrors) {
245 if (!ff_mirror_match_devid(mirror, pos))
246 continue;
247 if (!ff_mirror_match_fh(mirror, pos))
248 continue;
249 if (refcount_inc_not_zero(&pos->ref)) {
250 spin_unlock(&inode->i_lock);
251 return pos;
252 }
253 }
254 list_add(&mirror->mirrors, &ff_layout->mirrors);
255 mirror->layout = lo;
256 spin_unlock(&inode->i_lock);
257 return mirror;
258 }
259
260 static void
ff_layout_remove_mirror(struct nfs4_ff_layout_mirror * mirror)261 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror)
262 {
263 struct inode *inode;
264 if (mirror->layout == NULL)
265 return;
266 inode = mirror->layout->plh_inode;
267 spin_lock(&inode->i_lock);
268 list_del(&mirror->mirrors);
269 spin_unlock(&inode->i_lock);
270 mirror->layout = NULL;
271 }
272
ff_layout_alloc_mirror(u32 dss_count,gfp_t gfp_flags)273 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(u32 dss_count,
274 gfp_t gfp_flags)
275 {
276 struct nfs4_ff_layout_mirror *mirror;
277
278 mirror = kzalloc_obj(*mirror, gfp_flags);
279 if (mirror == NULL)
280 return NULL;
281
282 spin_lock_init(&mirror->lock);
283 refcount_set(&mirror->ref, 1);
284 INIT_LIST_HEAD(&mirror->mirrors);
285
286 mirror->dss_count = dss_count;
287 mirror->dss =
288 kzalloc_objs(struct nfs4_ff_layout_ds_stripe, dss_count,
289 gfp_flags);
290 if (mirror->dss == NULL) {
291 kfree(mirror);
292 return NULL;
293 }
294
295 for (u32 dss_id = 0; dss_id < mirror->dss_count; dss_id++)
296 nfs_localio_file_init(&mirror->dss[dss_id].nfl);
297
298 return mirror;
299 }
300
ff_layout_free_mirror(struct nfs4_ff_layout_mirror * mirror)301 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror)
302 {
303 const struct cred *cred;
304 u32 dss_id;
305
306 ff_layout_remove_mirror(mirror);
307
308 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++) {
309 kfree(mirror->dss[dss_id].fh_versions);
310 cred = rcu_access_pointer(mirror->dss[dss_id].ro_cred);
311 put_cred(cred);
312 cred = rcu_access_pointer(mirror->dss[dss_id].rw_cred);
313 put_cred(cred);
314 nfs_close_local_fh(&mirror->dss[dss_id].nfl);
315 nfs4_ff_layout_put_deviceid(mirror->dss[dss_id].mirror_ds);
316 }
317
318 kfree(mirror->dss);
319 kfree(mirror);
320 }
321
ff_layout_put_mirror(struct nfs4_ff_layout_mirror * mirror)322 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror)
323 {
324 if (mirror != NULL && refcount_dec_and_test(&mirror->ref))
325 ff_layout_free_mirror(mirror);
326 }
327
ff_layout_free_mirror_array(struct nfs4_ff_layout_segment * fls)328 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
329 {
330 u32 i;
331
332 for (i = 0; i < fls->mirror_array_cnt; i++)
333 ff_layout_put_mirror(fls->mirror_array[i]);
334 }
335
_ff_layout_free_lseg(struct nfs4_ff_layout_segment * fls)336 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
337 {
338 if (fls) {
339 ff_layout_free_mirror_array(fls);
340 kfree(fls);
341 }
342 }
343
344 static bool
ff_lseg_match_mirrors(struct pnfs_layout_segment * l1,struct pnfs_layout_segment * l2)345 ff_lseg_match_mirrors(struct pnfs_layout_segment *l1,
346 struct pnfs_layout_segment *l2)
347 {
348 const struct nfs4_ff_layout_segment *fl1 = FF_LAYOUT_LSEG(l1);
349 const struct nfs4_ff_layout_segment *fl2 = FF_LAYOUT_LSEG(l2);
350 u32 i;
351
352 if (fl1->mirror_array_cnt != fl2->mirror_array_cnt)
353 return false;
354 for (i = 0; i < fl1->mirror_array_cnt; i++) {
355 if (fl1->mirror_array[i] != fl2->mirror_array[i])
356 return false;
357 }
358 return true;
359 }
360
361 static bool
ff_lseg_range_is_after(const struct pnfs_layout_range * l1,const struct pnfs_layout_range * l2)362 ff_lseg_range_is_after(const struct pnfs_layout_range *l1,
363 const struct pnfs_layout_range *l2)
364 {
365 u64 end1, end2;
366
367 if (l1->iomode != l2->iomode)
368 return l1->iomode != IOMODE_READ;
369 end1 = pnfs_calc_offset_end(l1->offset, l1->length);
370 end2 = pnfs_calc_offset_end(l2->offset, l2->length);
371 if (end1 < l2->offset)
372 return false;
373 if (end2 < l1->offset)
374 return true;
375 return l2->offset <= l1->offset;
376 }
377
378 static bool
ff_lseg_merge(struct pnfs_layout_segment * new,struct pnfs_layout_segment * old)379 ff_lseg_merge(struct pnfs_layout_segment *new,
380 struct pnfs_layout_segment *old)
381 {
382 u64 new_end, old_end;
383
384 if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
385 return false;
386 if (new->pls_range.iomode != old->pls_range.iomode)
387 return false;
388 old_end = pnfs_calc_offset_end(old->pls_range.offset,
389 old->pls_range.length);
390 if (old_end < new->pls_range.offset)
391 return false;
392 new_end = pnfs_calc_offset_end(new->pls_range.offset,
393 new->pls_range.length);
394 if (new_end < old->pls_range.offset)
395 return false;
396 if (!ff_lseg_match_mirrors(new, old))
397 return false;
398
399 /* Mergeable: copy info from 'old' to 'new' */
400 if (new_end < old_end)
401 new_end = old_end;
402 if (new->pls_range.offset < old->pls_range.offset)
403 new->pls_range.offset = old->pls_range.offset;
404 new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
405 new_end);
406 if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
407 set_bit(NFS_LSEG_ROC, &new->pls_flags);
408 return true;
409 }
410
411 static void
ff_layout_add_lseg(struct pnfs_layout_hdr * lo,struct pnfs_layout_segment * lseg,struct list_head * free_me)412 ff_layout_add_lseg(struct pnfs_layout_hdr *lo,
413 struct pnfs_layout_segment *lseg,
414 struct list_head *free_me)
415 {
416 pnfs_generic_layout_insert_lseg(lo, lseg,
417 ff_lseg_range_is_after,
418 ff_lseg_merge,
419 free_me);
420 }
421
ff_mirror_efficiency_sum(const struct nfs4_ff_layout_mirror * mirror)422 static u32 ff_mirror_efficiency_sum(const struct nfs4_ff_layout_mirror *mirror)
423 {
424 u32 dss_id, sum = 0;
425
426 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++)
427 sum += mirror->dss[dss_id].efficiency;
428
429 return sum;
430 }
431
ff_layout_sort_mirrors(struct nfs4_ff_layout_segment * fls)432 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
433 {
434 int i, j;
435
436 for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
437 for (j = i + 1; j < fls->mirror_array_cnt; j++)
438 if (ff_mirror_efficiency_sum(fls->mirror_array[i]) <
439 ff_mirror_efficiency_sum(fls->mirror_array[j]))
440 swap(fls->mirror_array[i],
441 fls->mirror_array[j]);
442 }
443 }
444
445 static struct pnfs_layout_segment *
ff_layout_alloc_lseg(struct pnfs_layout_hdr * lh,struct nfs4_layoutget_res * lgr,gfp_t gfp_flags)446 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
447 struct nfs4_layoutget_res *lgr,
448 gfp_t gfp_flags)
449 {
450 struct pnfs_layout_segment *ret;
451 struct nfs4_ff_layout_segment *fls = NULL;
452 struct xdr_stream stream;
453 struct xdr_buf buf;
454 struct folio *scratch;
455 u64 stripe_unit;
456 u32 mirror_array_cnt;
457 __be32 *p;
458 int i, rc;
459 struct nfs4_ff_layout_ds_stripe *dss_info;
460
461 dprintk("--> %s\n", __func__);
462 scratch = folio_alloc(gfp_flags, 0);
463 if (!scratch)
464 return ERR_PTR(-ENOMEM);
465
466 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
467 lgr->layoutp->len);
468 xdr_set_scratch_folio(&stream, scratch);
469
470 /* stripe unit and mirror_array_cnt */
471 rc = -EIO;
472 p = xdr_inline_decode(&stream, 8 + 4);
473 if (!p)
474 goto out_err_free;
475
476 p = xdr_decode_hyper(p, &stripe_unit);
477 mirror_array_cnt = be32_to_cpup(p++);
478 dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
479 stripe_unit, mirror_array_cnt);
480
481 if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
482 mirror_array_cnt == 0)
483 goto out_err_free;
484
485 rc = -ENOMEM;
486 fls = kzalloc_flex(*fls, mirror_array, mirror_array_cnt, gfp_flags);
487 if (!fls)
488 goto out_err_free;
489
490 fls->mirror_array_cnt = mirror_array_cnt;
491 fls->stripe_unit = stripe_unit;
492
493 u32 dss_count = 0;
494 for (i = 0; i < fls->mirror_array_cnt; i++) {
495 struct nfs4_ff_layout_mirror *mirror;
496 struct cred *kcred;
497 const struct cred __rcu *cred;
498 kuid_t uid;
499 kgid_t gid;
500 u32 fh_count, id;
501 int j, dss_id;
502
503 rc = -EIO;
504 p = xdr_inline_decode(&stream, 4);
505 if (!p)
506 goto out_err_free;
507
508 // Ensure all mirrors have same stripe count.
509 if (dss_count == 0)
510 dss_count = be32_to_cpup(p);
511 else if (dss_count != be32_to_cpup(p))
512 goto out_err_free;
513
514 if (dss_count > NFS4_FLEXFILE_LAYOUT_MAX_STRIPE_CNT ||
515 dss_count == 0)
516 goto out_err_free;
517
518 if (dss_count > 1 && stripe_unit == 0)
519 goto out_err_free;
520
521 fls->mirror_array[i] = ff_layout_alloc_mirror(dss_count, gfp_flags);
522 if (fls->mirror_array[i] == NULL) {
523 rc = -ENOMEM;
524 goto out_err_free;
525 }
526
527 for (dss_id = 0; dss_id < dss_count; dss_id++) {
528 dss_info = &fls->mirror_array[i]->dss[dss_id];
529 dss_info->mirror = fls->mirror_array[i];
530
531 /* deviceid */
532 rc = decode_deviceid(&stream, &dss_info->devid);
533 if (rc)
534 goto out_err_free;
535
536 /* efficiency */
537 rc = -EIO;
538 p = xdr_inline_decode(&stream, 4);
539 if (!p)
540 goto out_err_free;
541 dss_info->efficiency = be32_to_cpup(p);
542
543 /* stateid */
544 rc = decode_pnfs_stateid(&stream, &dss_info->stateid);
545 if (rc)
546 goto out_err_free;
547
548 /* fh */
549 rc = -EIO;
550 p = xdr_inline_decode(&stream, 4);
551 if (!p)
552 goto out_err_free;
553 fh_count = be32_to_cpup(p);
554 if (fh_count == 0) {
555 rc = -EINVAL;
556 goto out_err_free;
557 }
558
559 dss_info->fh_versions =
560 kzalloc_objs(struct nfs_fh, fh_count, gfp_flags);
561 if (dss_info->fh_versions == NULL) {
562 rc = -ENOMEM;
563 goto out_err_free;
564 }
565
566 for (j = 0; j < fh_count; j++) {
567 rc = decode_nfs_fh(&stream,
568 &dss_info->fh_versions[j]);
569 if (rc)
570 goto out_err_free;
571 }
572
573 dss_info->fh_versions_cnt = fh_count;
574
575 /* user */
576 rc = decode_name(&stream, &id);
577 if (rc)
578 goto out_err_free;
579
580 uid = make_kuid(&init_user_ns, id);
581
582 /* group */
583 rc = decode_name(&stream, &id);
584 if (rc)
585 goto out_err_free;
586
587 gid = make_kgid(&init_user_ns, id);
588
589 if (gfp_flags & __GFP_FS)
590 kcred = prepare_kernel_cred(&init_task);
591 else {
592 unsigned int nofs_flags = memalloc_nofs_save();
593
594 kcred = prepare_kernel_cred(&init_task);
595 memalloc_nofs_restore(nofs_flags);
596 }
597 rc = -ENOMEM;
598 if (!kcred)
599 goto out_err_free;
600 kcred->fsuid = uid;
601 kcred->fsgid = gid;
602 cred = RCU_INITIALIZER(kcred);
603
604 if (lgr->range.iomode == IOMODE_READ)
605 rcu_assign_pointer(dss_info->ro_cred, cred);
606 else
607 rcu_assign_pointer(dss_info->rw_cred, cred);
608 }
609
610 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]);
611 if (mirror != fls->mirror_array[i]) {
612 for (dss_id = 0; dss_id < dss_count; dss_id++) {
613 dss_info = &fls->mirror_array[i]->dss[dss_id];
614 /* swap cred ptrs so free_mirror will clean up old */
615 if (lgr->range.iomode == IOMODE_READ) {
616 cred = xchg(&mirror->dss[dss_id].ro_cred,
617 dss_info->ro_cred);
618 rcu_assign_pointer(dss_info->ro_cred, cred);
619 } else {
620 cred = xchg(&mirror->dss[dss_id].rw_cred,
621 dss_info->rw_cred);
622 rcu_assign_pointer(dss_info->rw_cred, cred);
623 }
624 }
625 ff_layout_free_mirror(fls->mirror_array[i]);
626 fls->mirror_array[i] = mirror;
627 }
628
629 dprintk("%s: iomode %s uid %u gid %u\n", __func__,
630 lgr->range.iomode == IOMODE_READ ? "READ" : "RW",
631 from_kuid(&init_user_ns, uid),
632 from_kgid(&init_user_ns, gid));
633 }
634
635 p = xdr_inline_decode(&stream, 4);
636 if (!p)
637 goto out_sort_mirrors;
638 fls->flags = be32_to_cpup(p);
639 if (fls->flags & FF_FLAGS_NO_IO_THRU_MDS)
640 set_bit(NFS4_FF_HDR_NO_IO_THRU_MDS,
641 &FF_LAYOUT_FROM_HDR(lh)->flags);
642
643 p = xdr_inline_decode(&stream, 4);
644 if (!p)
645 goto out_sort_mirrors;
646 for (i=0; i < fls->mirror_array_cnt; i++)
647 fls->mirror_array[i]->report_interval = be32_to_cpup(p);
648
649 out_sort_mirrors:
650 ff_layout_sort_mirrors(fls);
651 ret = &fls->generic_hdr;
652 dprintk("<-- %s (success)\n", __func__);
653 out_free_page:
654 folio_put(scratch);
655 return ret;
656 out_err_free:
657 _ff_layout_free_lseg(fls);
658 ret = ERR_PTR(rc);
659 dprintk("<-- %s (%d)\n", __func__, rc);
660 goto out_free_page;
661 }
662
663 static void
ff_layout_free_lseg(struct pnfs_layout_segment * lseg)664 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
665 {
666 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
667
668 dprintk("--> %s\n", __func__);
669
670 if (lseg->pls_range.iomode == IOMODE_RW) {
671 struct nfs4_flexfile_layout *ffl;
672 struct inode *inode;
673
674 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
675 inode = ffl->generic_hdr.plh_inode;
676 spin_lock(&inode->i_lock);
677 pnfs_generic_ds_cinfo_release_lseg(&ffl->commit_info, lseg);
678 spin_unlock(&inode->i_lock);
679 }
680 _ff_layout_free_lseg(fls);
681 }
682
calc_commit_idx(struct pnfs_layout_segment * lseg,u32 mirror_idx,u32 dss_id)683 static u32 calc_commit_idx(struct pnfs_layout_segment *lseg,
684 u32 mirror_idx, u32 dss_id)
685 {
686 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
687
688 return (mirror_idx * flseg->mirror_array[0]->dss_count) + dss_id;
689 }
690
calc_mirror_idx_from_commit(struct pnfs_layout_segment * lseg,u32 commit_index)691 static u32 calc_mirror_idx_from_commit(struct pnfs_layout_segment *lseg,
692 u32 commit_index)
693 {
694 return commit_index / FF_LAYOUT_LSEG(lseg)->mirror_array[0]->dss_count;
695 }
696
calc_dss_id_from_commit(struct pnfs_layout_segment * lseg,u32 commit_index)697 static u32 calc_dss_id_from_commit(struct pnfs_layout_segment *lseg,
698 u32 commit_index)
699 {
700 return commit_index % FF_LAYOUT_LSEG(lseg)->mirror_array[0]->dss_count;
701 }
702
703 static void
nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer * timer,ktime_t now)704 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
705 {
706 /* first IO request? */
707 if (atomic_inc_return(&timer->n_ops) == 1) {
708 timer->start_time = now;
709 }
710 }
711
712 static ktime_t
nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer * timer,ktime_t now)713 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
714 {
715 ktime_t start;
716
717 if (atomic_dec_return(&timer->n_ops) < 0)
718 WARN_ON_ONCE(1);
719
720 start = timer->start_time;
721 timer->start_time = now;
722 return ktime_sub(now, start);
723 }
724
725 static bool
nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror * mirror,u32 dss_id,struct nfs4_ff_layoutstat * layoutstat,ktime_t now)726 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
727 u32 dss_id,
728 struct nfs4_ff_layoutstat *layoutstat,
729 ktime_t now)
730 {
731 s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
732 struct nfs4_flexfile_layout *ffl = FF_LAYOUT_FROM_HDR(mirror->layout);
733
734 nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
735 if (!mirror->dss[dss_id].start_time)
736 mirror->dss[dss_id].start_time = now;
737 if (mirror->report_interval != 0)
738 report_interval = (s64)mirror->report_interval * 1000LL;
739 else if (layoutstats_timer != 0)
740 report_interval = (s64)layoutstats_timer * 1000LL;
741 if (ktime_to_ms(ktime_sub(now, ffl->last_report_time)) >=
742 report_interval) {
743 ffl->last_report_time = now;
744 return true;
745 }
746
747 return false;
748 }
749
750 static void
nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat * layoutstat,__u64 requested)751 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
752 __u64 requested)
753 {
754 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
755
756 iostat->ops_requested++;
757 iostat->bytes_requested += requested;
758 }
759
760 static void
nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat * layoutstat,__u64 requested,__u64 completed,ktime_t time_completed,ktime_t time_started)761 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
762 __u64 requested,
763 __u64 completed,
764 ktime_t time_completed,
765 ktime_t time_started)
766 {
767 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
768 ktime_t completion_time = ktime_sub(time_completed, time_started);
769 ktime_t timer;
770
771 iostat->ops_completed++;
772 iostat->bytes_completed += completed;
773 iostat->bytes_not_delivered += requested - completed;
774
775 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
776 iostat->total_busy_time =
777 ktime_add(iostat->total_busy_time, timer);
778 iostat->aggregate_completion_time =
779 ktime_add(iostat->aggregate_completion_time,
780 completion_time);
781 }
782
783 static void
nfs4_ff_layout_stat_io_start_read(struct inode * inode,struct nfs4_ff_layout_mirror * mirror,u32 dss_id,__u64 requested,ktime_t now)784 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
785 struct nfs4_ff_layout_mirror *mirror,
786 u32 dss_id,
787 __u64 requested, ktime_t now)
788 {
789 bool report;
790
791 spin_lock(&mirror->lock);
792 report = nfs4_ff_layoutstat_start_io(
793 mirror, dss_id, &mirror->dss[dss_id].read_stat, now);
794 nfs4_ff_layout_stat_io_update_requested(
795 &mirror->dss[dss_id].read_stat, requested);
796 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags);
797 spin_unlock(&mirror->lock);
798
799 if (report)
800 pnfs_report_layoutstat(inode, nfs_io_gfp_mask());
801 }
802
803 static void
nfs4_ff_layout_stat_io_end_read(struct rpc_task * task,struct nfs4_ff_layout_mirror * mirror,u32 dss_id,__u64 requested,__u64 completed)804 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
805 struct nfs4_ff_layout_mirror *mirror,
806 u32 dss_id,
807 __u64 requested,
808 __u64 completed)
809 {
810 spin_lock(&mirror->lock);
811 nfs4_ff_layout_stat_io_update_completed(&mirror->dss[dss_id].read_stat,
812 requested, completed,
813 ktime_get(), task->tk_start);
814 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags);
815 spin_unlock(&mirror->lock);
816 }
817
818 static void
nfs4_ff_layout_stat_io_start_write(struct inode * inode,struct nfs4_ff_layout_mirror * mirror,u32 dss_id,__u64 requested,ktime_t now)819 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
820 struct nfs4_ff_layout_mirror *mirror,
821 u32 dss_id,
822 __u64 requested, ktime_t now)
823 {
824 bool report;
825
826 spin_lock(&mirror->lock);
827 report = nfs4_ff_layoutstat_start_io(
828 mirror,
829 dss_id,
830 &mirror->dss[dss_id].write_stat,
831 now);
832 nfs4_ff_layout_stat_io_update_requested(
833 &mirror->dss[dss_id].write_stat,
834 requested);
835 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags);
836 spin_unlock(&mirror->lock);
837
838 if (report)
839 pnfs_report_layoutstat(inode, nfs_io_gfp_mask());
840 }
841
842 static void
nfs4_ff_layout_stat_io_end_write(struct rpc_task * task,struct nfs4_ff_layout_mirror * mirror,u32 dss_id,__u64 requested,__u64 completed,enum nfs3_stable_how committed)843 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
844 struct nfs4_ff_layout_mirror *mirror,
845 u32 dss_id,
846 __u64 requested,
847 __u64 completed,
848 enum nfs3_stable_how committed)
849 {
850 if (committed == NFS_UNSTABLE)
851 requested = completed = 0;
852
853 spin_lock(&mirror->lock);
854 nfs4_ff_layout_stat_io_update_completed(&mirror->dss[dss_id].write_stat,
855 requested, completed, ktime_get(), task->tk_start);
856 set_bit(NFS4_FF_MIRROR_STAT_AVAIL, &mirror->flags);
857 spin_unlock(&mirror->lock);
858 }
859
860 static void
ff_layout_mark_ds_unreachable(struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id)861 ff_layout_mark_ds_unreachable(struct pnfs_layout_segment *lseg, u32 idx, u32 dss_id)
862 {
863 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id);
864
865 if (devid)
866 nfs4_mark_deviceid_unavailable(devid);
867 }
868
869 static void
ff_layout_mark_ds_reachable(struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id)870 ff_layout_mark_ds_reachable(struct pnfs_layout_segment *lseg, u32 idx, u32 dss_id)
871 {
872 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id);
873
874 if (devid)
875 nfs4_mark_deviceid_available(devid);
876 }
877
878 static struct nfs4_pnfs_ds *
ff_layout_choose_ds_for_read(struct pnfs_layout_segment * lseg,u32 start_idx,u32 * best_idx,u32 offset,u32 * dss_id,bool check_device)879 ff_layout_choose_ds_for_read(struct pnfs_layout_segment *lseg,
880 u32 start_idx, u32 *best_idx,
881 u32 offset, u32 *dss_id,
882 bool check_device)
883 {
884 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
885 struct nfs4_ff_layout_mirror *mirror;
886 struct nfs4_pnfs_ds *ds = ERR_PTR(-EAGAIN);
887 u32 idx;
888
889 /* mirrors are initially sorted by efficiency */
890 for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
891 mirror = FF_LAYOUT_COMP(lseg, idx);
892 *dss_id = nfs4_ff_layout_calc_dss_id(
893 fls->stripe_unit,
894 fls->mirror_array[idx]->dss_count,
895 offset);
896 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, *dss_id, false);
897 if (IS_ERR(ds))
898 continue;
899
900 if (check_device &&
901 nfs4_test_deviceid_unavailable(&mirror->dss[*dss_id].mirror_ds->id_node)) {
902 // reinitialize the error state in case if this is the last iteration
903 ds = ERR_PTR(-EINVAL);
904 continue;
905 }
906
907 *best_idx = idx;
908 break;
909 }
910
911 return ds;
912 }
913
914 static struct nfs4_pnfs_ds *
ff_layout_choose_any_ds_for_read(struct pnfs_layout_segment * lseg,u32 start_idx,u32 * best_idx,u32 offset,u32 * dss_id)915 ff_layout_choose_any_ds_for_read(struct pnfs_layout_segment *lseg,
916 u32 start_idx, u32 *best_idx,
917 u32 offset, u32 *dss_id)
918 {
919 return ff_layout_choose_ds_for_read(lseg, start_idx, best_idx,
920 offset, dss_id, false);
921 }
922
923 static struct nfs4_pnfs_ds *
ff_layout_choose_valid_ds_for_read(struct pnfs_layout_segment * lseg,u32 start_idx,u32 * best_idx,u32 offset,u32 * dss_id)924 ff_layout_choose_valid_ds_for_read(struct pnfs_layout_segment *lseg,
925 u32 start_idx, u32 *best_idx,
926 u32 offset, u32 *dss_id)
927 {
928 return ff_layout_choose_ds_for_read(lseg, start_idx, best_idx,
929 offset, dss_id, true);
930 }
931
932 static struct nfs4_pnfs_ds *
ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment * lseg,u32 start_idx,u32 * best_idx,u32 offset,u32 * dss_id)933 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
934 u32 start_idx, u32 *best_idx,
935 u32 offset, u32 *dss_id)
936 {
937 struct nfs4_pnfs_ds *ds;
938
939 ds = ff_layout_choose_valid_ds_for_read(lseg, start_idx, best_idx,
940 offset, dss_id);
941 if (!IS_ERR(ds))
942 return ds;
943 return ff_layout_choose_any_ds_for_read(lseg, start_idx, best_idx,
944 offset, dss_id);
945 }
946
947 static struct nfs4_pnfs_ds *
ff_layout_get_ds_for_read(struct nfs_pageio_descriptor * pgio,u32 * best_idx,u32 offset,u32 * dss_id)948 ff_layout_get_ds_for_read(struct nfs_pageio_descriptor *pgio,
949 u32 *best_idx,
950 u32 offset,
951 u32 *dss_id)
952 {
953 struct pnfs_layout_segment *lseg = pgio->pg_lseg;
954 struct nfs4_pnfs_ds *ds;
955
956 ds = ff_layout_choose_best_ds_for_read(lseg, pgio->pg_mirror_idx,
957 best_idx, offset, dss_id);
958 if (!IS_ERR(ds) || !pgio->pg_mirror_idx)
959 return ds;
960 return ff_layout_choose_best_ds_for_read(lseg, 0, best_idx,
961 offset, dss_id);
962 }
963
964 static void
ff_layout_pg_get_read(struct nfs_pageio_descriptor * pgio,struct nfs_page * req,bool strict_iomode)965 ff_layout_pg_get_read(struct nfs_pageio_descriptor *pgio,
966 struct nfs_page *req,
967 bool strict_iomode)
968 {
969 pnfs_put_lseg(pgio->pg_lseg);
970 pgio->pg_lseg =
971 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
972 req_offset(req), req->wb_bytes, IOMODE_READ,
973 strict_iomode, nfs_io_gfp_mask());
974 if (IS_ERR(pgio->pg_lseg)) {
975 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
976 pgio->pg_lseg = NULL;
977 }
978 }
979
980 static bool
ff_layout_lseg_is_striped(const struct nfs4_ff_layout_segment * fls)981 ff_layout_lseg_is_striped(const struct nfs4_ff_layout_segment *fls)
982 {
983 return fls->mirror_array[0]->dss_count > 1;
984 }
985
986 /*
987 * ff_layout_pg_test(). Called by nfs_can_coalesce_requests()
988 *
989 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
990 * of bytes (maximum @req->wb_bytes) that can be coalesced.
991 */
992 static size_t
ff_layout_pg_test(struct nfs_pageio_descriptor * pgio,struct nfs_page * prev,struct nfs_page * req)993 ff_layout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
994 struct nfs_page *req)
995 {
996 unsigned int size;
997 u64 p_stripe, r_stripe;
998 u32 stripe_offset;
999 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
1000 u32 stripe_unit = FF_LAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
1001
1002 /* calls nfs_generic_pg_test */
1003 size = pnfs_generic_pg_test(pgio, prev, req);
1004 if (!size)
1005 return 0;
1006 else if (!ff_layout_lseg_is_striped(FF_LAYOUT_LSEG(pgio->pg_lseg)))
1007 return size;
1008
1009 /* see if req and prev are in the same stripe */
1010 if (prev) {
1011 p_stripe = (u64)req_offset(prev) - segment_offset;
1012 r_stripe = (u64)req_offset(req) - segment_offset;
1013 do_div(p_stripe, stripe_unit);
1014 do_div(r_stripe, stripe_unit);
1015
1016 if (p_stripe != r_stripe)
1017 return 0;
1018 }
1019
1020 /* calculate remaining bytes in the current stripe */
1021 div_u64_rem((u64)req_offset(req) - segment_offset,
1022 stripe_unit,
1023 &stripe_offset);
1024 WARN_ON_ONCE(stripe_offset > stripe_unit);
1025 if (stripe_offset >= stripe_unit)
1026 return 0;
1027 return min(stripe_unit - (unsigned int)stripe_offset, size);
1028 }
1029
1030 static void
ff_layout_pg_init_read(struct nfs_pageio_descriptor * pgio,struct nfs_page * req)1031 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
1032 struct nfs_page *req)
1033 {
1034 struct nfs_pgio_mirror *pgm;
1035 struct nfs4_ff_layout_mirror *mirror;
1036 struct nfs4_pnfs_ds *ds;
1037 u32 ds_idx, dss_id;
1038
1039 if (NFS_SERVER(pgio->pg_inode)->flags &
1040 (NFS_MOUNT_SOFT|NFS_MOUNT_SOFTERR))
1041 pgio->pg_maxretrans = io_maxretrans;
1042 retry:
1043 pnfs_generic_pg_check_layout(pgio, req);
1044 /* Use full layout for now */
1045 if (!pgio->pg_lseg) {
1046 ff_layout_pg_get_read(pgio, req, false);
1047 if (!pgio->pg_lseg)
1048 goto out_nolseg;
1049 }
1050 if (ff_layout_avoid_read_on_rw(pgio->pg_lseg)) {
1051 ff_layout_pg_get_read(pgio, req, true);
1052 if (!pgio->pg_lseg)
1053 goto out_nolseg;
1054 }
1055 /* Reset wb_nio, since getting layout segment was successful */
1056 req->wb_nio = 0;
1057
1058 ds = ff_layout_get_ds_for_read(pgio, &ds_idx,
1059 req_offset(req), &dss_id);
1060 if (IS_ERR(ds)) {
1061 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
1062 goto out_mds;
1063 pnfs_generic_pg_cleanup(pgio);
1064 /* Sleep for 1 second before retrying */
1065 ssleep(1);
1066 goto retry;
1067 }
1068
1069 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
1070 pgm = &pgio->pg_mirrors[0];
1071 pgm->pg_bsize = mirror->dss[dss_id].mirror_ds->ds_versions[0].rsize;
1072
1073 pgio->pg_mirror_idx = ds_idx;
1074 return;
1075 out_nolseg:
1076 if (pgio->pg_error < 0) {
1077 if (pgio->pg_error != -EAGAIN)
1078 return;
1079 /* Retry getting layout segment if lower layer returned -EAGAIN */
1080 if (pgio->pg_maxretrans && req->wb_nio++ > pgio->pg_maxretrans) {
1081 if (NFS_SERVER(pgio->pg_inode)->flags & NFS_MOUNT_SOFTERR)
1082 pgio->pg_error = -ETIMEDOUT;
1083 else
1084 pgio->pg_error = -EIO;
1085 return;
1086 }
1087 pgio->pg_error = 0;
1088 /* Sleep for 1 second before retrying */
1089 ssleep(1);
1090 goto retry;
1091 }
1092 out_mds:
1093 trace_pnfs_mds_fallback_pg_init_read(pgio->pg_inode,
1094 0, NFS4_MAX_UINT64, IOMODE_READ,
1095 NFS_I(pgio->pg_inode)->layout,
1096 pgio->pg_lseg);
1097 pgio->pg_maxretrans = 0;
1098 nfs_pageio_reset_read_mds(pgio);
1099 }
1100
1101 static void
ff_layout_pg_init_write(struct nfs_pageio_descriptor * pgio,struct nfs_page * req)1102 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
1103 struct nfs_page *req)
1104 {
1105 struct nfs4_ff_layout_mirror *mirror;
1106 struct nfs_pgio_mirror *pgm;
1107 struct nfs4_pnfs_ds *ds;
1108 u32 i, dss_id;
1109
1110 retry:
1111 pnfs_generic_pg_check_layout(pgio, req);
1112 if (!pgio->pg_lseg) {
1113 pgio->pg_lseg =
1114 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
1115 req_offset(req), req->wb_bytes,
1116 IOMODE_RW, false, nfs_io_gfp_mask());
1117 if (IS_ERR(pgio->pg_lseg)) {
1118 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1119 pgio->pg_lseg = NULL;
1120 return;
1121 }
1122 }
1123 /* If no lseg, fall back to write through mds */
1124 if (pgio->pg_lseg == NULL)
1125 goto out_mds;
1126
1127 /* Use a direct mapping of ds_idx to pgio mirror_idx */
1128 if (pgio->pg_mirror_count != FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg))
1129 goto out_eagain;
1130
1131 for (i = 0; i < pgio->pg_mirror_count; i++) {
1132 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
1133 dss_id = nfs4_ff_layout_calc_dss_id(
1134 FF_LAYOUT_LSEG(pgio->pg_lseg)->stripe_unit,
1135 mirror->dss_count,
1136 req_offset(req));
1137 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, mirror,
1138 dss_id, true);
1139 if (IS_ERR(ds)) {
1140 if (!ff_layout_no_fallback_to_mds(pgio->pg_lseg))
1141 goto out_mds;
1142 pnfs_generic_pg_cleanup(pgio);
1143 /* Sleep for 1 second before retrying */
1144 ssleep(1);
1145 goto retry;
1146 }
1147 pgm = &pgio->pg_mirrors[i];
1148 pgm->pg_bsize = mirror->dss[dss_id].mirror_ds->ds_versions[0].wsize;
1149 }
1150
1151 if (NFS_SERVER(pgio->pg_inode)->flags &
1152 (NFS_MOUNT_SOFT|NFS_MOUNT_SOFTERR))
1153 pgio->pg_maxretrans = io_maxretrans;
1154 return;
1155 out_eagain:
1156 pnfs_generic_pg_cleanup(pgio);
1157 pgio->pg_error = -EAGAIN;
1158 return;
1159 out_mds:
1160 trace_pnfs_mds_fallback_pg_init_write(pgio->pg_inode,
1161 0, NFS4_MAX_UINT64, IOMODE_RW,
1162 NFS_I(pgio->pg_inode)->layout,
1163 pgio->pg_lseg);
1164 pgio->pg_maxretrans = 0;
1165 nfs_pageio_reset_write_mds(pgio);
1166 pgio->pg_error = -EAGAIN;
1167 }
1168
1169 static unsigned int
ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor * pgio,struct nfs_page * req)1170 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
1171 struct nfs_page *req)
1172 {
1173 if (!pgio->pg_lseg) {
1174 pgio->pg_lseg =
1175 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
1176 req_offset(req), req->wb_bytes,
1177 IOMODE_RW, false, nfs_io_gfp_mask());
1178 if (IS_ERR(pgio->pg_lseg)) {
1179 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1180 pgio->pg_lseg = NULL;
1181 goto out;
1182 }
1183 }
1184 if (pgio->pg_lseg)
1185 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
1186
1187 trace_pnfs_mds_fallback_pg_get_mirror_count(pgio->pg_inode,
1188 0, NFS4_MAX_UINT64, IOMODE_RW,
1189 NFS_I(pgio->pg_inode)->layout,
1190 pgio->pg_lseg);
1191 if (NFS_I(pgio->pg_inode)->layout &&
1192 ff_layout_hdr_no_fallback_to_mds(NFS_I(pgio->pg_inode)->layout)) {
1193 /*
1194 * FF_FLAGS_NO_IO_THRU_MDS: no current lseg but the server's
1195 * policy forbids MDS fallback. Surface -EAGAIN so writeback
1196 * retries rather than silently issuing the WRITE via MDS.
1197 */
1198 pgio->pg_error = -EAGAIN;
1199 goto out;
1200 }
1201 /* no lseg means that pnfs is not in use, so no mirroring here */
1202 nfs_pageio_reset_write_mds(pgio);
1203 out:
1204 return 1;
1205 }
1206
1207 static u32
ff_layout_pg_set_mirror_write(struct nfs_pageio_descriptor * desc,u32 idx)1208 ff_layout_pg_set_mirror_write(struct nfs_pageio_descriptor *desc, u32 idx)
1209 {
1210 u32 old = desc->pg_mirror_idx;
1211
1212 desc->pg_mirror_idx = idx;
1213 return old;
1214 }
1215
1216 static struct nfs_pgio_mirror *
ff_layout_pg_get_mirror_write(struct nfs_pageio_descriptor * desc,u32 idx)1217 ff_layout_pg_get_mirror_write(struct nfs_pageio_descriptor *desc, u32 idx)
1218 {
1219 return &desc->pg_mirrors[idx];
1220 }
1221
1222 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
1223 .pg_init = ff_layout_pg_init_read,
1224 .pg_test = ff_layout_pg_test,
1225 .pg_doio = pnfs_generic_pg_readpages,
1226 .pg_cleanup = pnfs_generic_pg_cleanup,
1227 };
1228
1229 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
1230 .pg_init = ff_layout_pg_init_write,
1231 .pg_test = ff_layout_pg_test,
1232 .pg_doio = pnfs_generic_pg_writepages,
1233 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
1234 .pg_cleanup = pnfs_generic_pg_cleanup,
1235 .pg_get_mirror = ff_layout_pg_get_mirror_write,
1236 .pg_set_mirror = ff_layout_pg_set_mirror_write,
1237 };
1238
ff_layout_reset_write(struct nfs_pgio_header * hdr,bool retry_pnfs)1239 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
1240 {
1241 struct rpc_task *task = &hdr->task;
1242
1243 pnfs_layoutcommit_inode(hdr->inode, false);
1244
1245 if (retry_pnfs) {
1246 dprintk("%s Reset task %5u for i/o through pNFS "
1247 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1248 hdr->task.tk_pid,
1249 hdr->inode->i_sb->s_id,
1250 (unsigned long long)hdr->inode->i_ino,
1251 hdr->args.count,
1252 (unsigned long long)hdr->args.offset);
1253
1254 hdr->completion_ops->reschedule_io(hdr);
1255 return;
1256 }
1257
1258 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1259 dprintk("%s Reset task %5u for i/o through MDS "
1260 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1261 hdr->task.tk_pid,
1262 hdr->inode->i_sb->s_id,
1263 (unsigned long long)hdr->inode->i_ino,
1264 hdr->args.count,
1265 (unsigned long long)hdr->args.offset);
1266
1267 trace_pnfs_mds_fallback_write_done(hdr->inode,
1268 hdr->args.offset, hdr->args.count,
1269 IOMODE_RW, NFS_I(hdr->inode)->layout,
1270 hdr->lseg);
1271 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
1272 }
1273 }
1274
ff_layout_resend_pnfs_read(struct nfs_pgio_header * hdr)1275 static void ff_layout_resend_pnfs_read(struct nfs_pgio_header *hdr)
1276 {
1277 u32 idx = hdr->pgio_mirror_idx + 1;
1278 u32 new_idx = 0;
1279 u32 dss_id = 0;
1280 struct nfs4_pnfs_ds *ds;
1281
1282 ds = ff_layout_choose_any_ds_for_read(hdr->lseg, idx, &new_idx,
1283 hdr->args.offset, &dss_id);
1284 if (IS_ERR(ds))
1285 pnfs_error_mark_layout_for_return(hdr->inode, hdr->lseg);
1286 else
1287 ff_layout_send_layouterror(hdr->lseg);
1288 pnfs_read_resend_pnfs(hdr, new_idx);
1289 }
1290
ff_layout_reset_read(struct nfs_pgio_header * hdr)1291 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
1292 {
1293 struct rpc_task *task = &hdr->task;
1294
1295 pnfs_layoutcommit_inode(hdr->inode, false);
1296 pnfs_error_mark_layout_for_return(hdr->inode, hdr->lseg);
1297
1298 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1299 dprintk("%s Reset task %5u for i/o through MDS "
1300 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
1301 hdr->task.tk_pid,
1302 hdr->inode->i_sb->s_id,
1303 (unsigned long long)hdr->inode->i_ino,
1304 hdr->args.count,
1305 (unsigned long long)hdr->args.offset);
1306
1307 trace_pnfs_mds_fallback_read_done(hdr->inode,
1308 hdr->args.offset, hdr->args.count,
1309 IOMODE_READ, NFS_I(hdr->inode)->layout,
1310 hdr->lseg);
1311 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
1312 }
1313 }
1314
ff_layout_async_handle_error_v4(struct rpc_task * task,u32 op_status,struct nfs4_state * state,struct nfs_client * clp,struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id)1315 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
1316 u32 op_status,
1317 struct nfs4_state *state,
1318 struct nfs_client *clp,
1319 struct pnfs_layout_segment *lseg,
1320 u32 idx, u32 dss_id)
1321 {
1322 struct pnfs_layout_hdr *lo = lseg->pls_layout;
1323 struct inode *inode = lo->plh_inode;
1324 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id);
1325 struct nfs4_slot_table *tbl = nfs4_has_session(clp) ?
1326 &clp->cl_session->fc_slot_table : clp->cl_slot_tbl;
1327
1328 switch (op_status) {
1329 case NFS4_OK:
1330 case NFS4ERR_NXIO:
1331 break;
1332 case NFSERR_PERM:
1333 if (!task->tk_xprt)
1334 break;
1335 xprt_force_disconnect(task->tk_xprt);
1336 goto out_retry;
1337 case NFS4ERR_BADSESSION:
1338 case NFS4ERR_BADSLOT:
1339 case NFS4ERR_BAD_HIGH_SLOT:
1340 case NFS4ERR_DEADSESSION:
1341 case NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1342 case NFS4ERR_SEQ_FALSE_RETRY:
1343 case NFS4ERR_SEQ_MISORDERED:
1344 dprintk("%s ERROR %d, Reset session. Exchangeid "
1345 "flags 0x%x\n", __func__, task->tk_status,
1346 clp->cl_exchange_flags);
1347 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1348 goto out_retry;
1349 case NFS4ERR_DELAY:
1350 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1351 fallthrough;
1352 case NFS4ERR_GRACE:
1353 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1354 goto out_retry;
1355 case NFS4ERR_RETRY_UNCACHED_REP:
1356 goto out_retry;
1357 /* Invalidate Layout errors */
1358 case NFS4ERR_PNFS_NO_LAYOUT:
1359 case NFS4ERR_STALE:
1360 case NFS4ERR_BADHANDLE:
1361 case NFS4ERR_ISDIR:
1362 case NFS4ERR_FHEXPIRED:
1363 case NFS4ERR_WRONG_TYPE:
1364 dprintk("%s Invalid layout error %d\n", __func__,
1365 task->tk_status);
1366 /*
1367 * Destroy layout so new i/o will get a new layout.
1368 * Layout will not be destroyed until all current lseg
1369 * references are put. Mark layout as invalid to resend failed
1370 * i/o and all i/o waiting on the slot table to the MDS until
1371 * layout is destroyed and a new valid layout is obtained.
1372 */
1373 pnfs_destroy_layout(NFS_I(inode));
1374 rpc_wake_up(&tbl->slot_tbl_waitq);
1375 goto reset;
1376 default:
1377 break;
1378 }
1379
1380 switch (task->tk_status) {
1381 /* RPC connection errors */
1382 case -ENETDOWN:
1383 case -ENETUNREACH:
1384 if (test_bit(NFS_CS_NETUNREACH_FATAL, &clp->cl_flags))
1385 return -NFS4ERR_FATAL_IOERROR;
1386 fallthrough;
1387 case -ECONNREFUSED:
1388 case -EHOSTDOWN:
1389 case -EHOSTUNREACH:
1390 case -EIO:
1391 case -ETIMEDOUT:
1392 case -EPIPE:
1393 case -EPROTO:
1394 case -ENODEV:
1395 dprintk("%s DS connection error %d\n", __func__,
1396 task->tk_status);
1397 nfs4_delete_deviceid(devid->ld, devid->nfs_client,
1398 &devid->deviceid);
1399 rpc_wake_up(&tbl->slot_tbl_waitq);
1400 break;
1401 default:
1402 break;
1403 }
1404
1405 if (ff_layout_avoid_mds_available_ds(lseg))
1406 return -NFS4ERR_RESET_TO_PNFS;
1407 reset:
1408 dprintk("%s Retry through MDS. Error %d\n", __func__,
1409 task->tk_status);
1410 return -NFS4ERR_RESET_TO_MDS;
1411
1412 out_retry:
1413 task->tk_status = 0;
1414 return -EAGAIN;
1415 }
1416
1417 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
ff_layout_async_handle_error_v3(struct rpc_task * task,u32 op_status,struct nfs_client * clp,struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id)1418 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1419 u32 op_status,
1420 struct nfs_client *clp,
1421 struct pnfs_layout_segment *lseg,
1422 u32 idx, u32 dss_id)
1423 {
1424 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx, dss_id);
1425
1426 switch (op_status) {
1427 case NFS_OK:
1428 case NFSERR_NXIO:
1429 break;
1430 case NFSERR_PERM:
1431 if (!task->tk_xprt)
1432 break;
1433 xprt_force_disconnect(task->tk_xprt);
1434 goto out_retry;
1435 case NFSERR_ACCES:
1436 case NFSERR_BADHANDLE:
1437 case NFSERR_FBIG:
1438 case NFSERR_IO:
1439 case NFSERR_NOSPC:
1440 case NFSERR_ROFS:
1441 case NFSERR_STALE:
1442 goto out_reset_to_pnfs;
1443 case NFSERR_JUKEBOX:
1444 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1445 goto out_retry;
1446 default:
1447 break;
1448 }
1449
1450 switch (task->tk_status) {
1451 /* File access problems. Don't mark the device as unavailable */
1452 case -EACCES:
1453 case -ESTALE:
1454 case -EISDIR:
1455 case -EBADHANDLE:
1456 case -ELOOP:
1457 case -ENOSPC:
1458 break;
1459 case -EJUKEBOX:
1460 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1461 goto out_retry;
1462 case -ENETDOWN:
1463 case -ENETUNREACH:
1464 if (test_bit(NFS_CS_NETUNREACH_FATAL, &clp->cl_flags))
1465 return -NFS4ERR_FATAL_IOERROR;
1466 fallthrough;
1467 default:
1468 dprintk("%s DS connection error %d\n", __func__,
1469 task->tk_status);
1470 nfs4_delete_deviceid(devid->ld, devid->nfs_client,
1471 &devid->deviceid);
1472 }
1473 out_reset_to_pnfs:
1474 /* FIXME: Need to prevent infinite looping here. */
1475 return -NFS4ERR_RESET_TO_PNFS;
1476 out_retry:
1477 task->tk_status = 0;
1478 rpc_restart_call_prepare(task);
1479 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1480 return -EAGAIN;
1481 }
1482
ff_layout_async_handle_error(struct rpc_task * task,u32 op_status,struct nfs4_state * state,struct nfs_client * clp,struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id)1483 static int ff_layout_async_handle_error(struct rpc_task *task,
1484 u32 op_status,
1485 struct nfs4_state *state,
1486 struct nfs_client *clp,
1487 struct pnfs_layout_segment *lseg,
1488 u32 idx, u32 dss_id)
1489 {
1490 int vers = clp->cl_nfs_mod->rpc_vers->number;
1491
1492 if (task->tk_status >= 0) {
1493 ff_layout_mark_ds_reachable(lseg, idx, dss_id);
1494 return 0;
1495 }
1496
1497 /* Handle the case of an invalid layout segment */
1498 if (!pnfs_is_valid_lseg(lseg))
1499 return -NFS4ERR_RESET_TO_PNFS;
1500
1501 switch (vers) {
1502 case 3:
1503 return ff_layout_async_handle_error_v3(task, op_status, clp,
1504 lseg, idx, dss_id);
1505 case 4:
1506 return ff_layout_async_handle_error_v4(task, op_status, state,
1507 clp, lseg, idx, dss_id);
1508 default:
1509 /* should never happen */
1510 WARN_ON_ONCE(1);
1511 return 0;
1512 }
1513 }
1514
ff_layout_io_track_ds_error(struct pnfs_layout_segment * lseg,u32 idx,u32 dss_id,u64 offset,u64 length,u32 * op_status,int opnum,int error)1515 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1516 u32 idx, u32 dss_id, u64 offset, u64 length,
1517 u32 *op_status, int opnum, int error)
1518 {
1519 struct nfs4_ff_layout_mirror *mirror;
1520 u32 status = *op_status;
1521 int err;
1522
1523 if (status == 0) {
1524 switch (error) {
1525 case -ETIMEDOUT:
1526 case -EPFNOSUPPORT:
1527 case -EPROTONOSUPPORT:
1528 case -EOPNOTSUPP:
1529 case -EINVAL:
1530 case -ECONNREFUSED:
1531 case -ECONNRESET:
1532 case -EHOSTDOWN:
1533 case -EHOSTUNREACH:
1534 case -ENETDOWN:
1535 case -ENETUNREACH:
1536 case -EADDRINUSE:
1537 case -ENOBUFS:
1538 case -EPIPE:
1539 case -EPERM:
1540 case -EPROTO:
1541 case -ENODEV:
1542 *op_status = status = NFS4ERR_NXIO;
1543 break;
1544 case -EACCES:
1545 *op_status = status = NFS4ERR_ACCESS;
1546 break;
1547 case -ECANCELED:
1548 /*
1549 * In-flight I/O we cancelled to return a recalled or
1550 * revoked layout. Report it as a failure to reach the
1551 * device (NFS4ERR_NXIO), like the transport errors
1552 * above, so the server can reconcile the affected mirror
1553 * instance. We aborted the I/O ourselves rather than
1554 * observe the device fail, so don't condemn it below.
1555 */
1556 *op_status = status = NFS4ERR_NXIO;
1557 break;
1558 default:
1559 return;
1560 }
1561 }
1562
1563 mirror = FF_LAYOUT_COMP(lseg, idx);
1564 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1565 mirror, dss_id, offset, length, status, opnum,
1566 nfs_io_gfp_mask());
1567
1568 /*
1569 * I/O we cancelled ourselves to return a recalled or revoked layout
1570 * is reported above so the server can reconcile the mirror, but we
1571 * have no evidence the device is at fault: don't mark it unreachable
1572 * or force a return.
1573 */
1574 if (error == -ECANCELED)
1575 goto out;
1576
1577 switch (status) {
1578 case NFS4ERR_DELAY:
1579 case NFS4ERR_GRACE:
1580 case NFS4ERR_PERM:
1581 break;
1582 case NFS4ERR_NXIO:
1583 ff_layout_mark_ds_unreachable(lseg, idx, dss_id);
1584 /*
1585 * Don't return the layout if this is a read and we still
1586 * have layouts to try
1587 */
1588 if (opnum == OP_READ)
1589 break;
1590 fallthrough;
1591 default:
1592 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
1593 lseg);
1594 }
1595
1596 out:
1597 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1598 }
1599
1600 /* NFS_PROTO call done callback routines */
ff_layout_read_done_cb(struct rpc_task * task,struct nfs_pgio_header * hdr)1601 static int ff_layout_read_done_cb(struct rpc_task *task,
1602 struct nfs_pgio_header *hdr)
1603 {
1604 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(hdr->lseg);
1605 u32 dss_id = nfs4_ff_layout_calc_dss_id(
1606 flseg->stripe_unit,
1607 flseg->mirror_array[hdr->pgio_mirror_idx]->dss_count,
1608 hdr->args.offset);
1609 int err;
1610
1611 if (task->tk_status < 0) {
1612 ff_layout_io_track_ds_error(hdr->lseg,
1613 hdr->pgio_mirror_idx, dss_id,
1614 hdr->args.offset, hdr->args.count,
1615 &hdr->res.op_status, OP_READ,
1616 task->tk_status);
1617 trace_ff_layout_read_error(hdr, task->tk_status);
1618 }
1619
1620 err = ff_layout_async_handle_error(task, hdr->res.op_status,
1621 hdr->args.context->state,
1622 hdr->ds_clp, hdr->lseg,
1623 hdr->pgio_mirror_idx,
1624 dss_id);
1625
1626 trace_nfs4_pnfs_read(hdr, err);
1627 clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
1628 clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
1629 switch (err) {
1630 case -NFS4ERR_RESET_TO_PNFS:
1631 set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
1632 return task->tk_status;
1633 case -NFS4ERR_RESET_TO_MDS:
1634 set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
1635 return task->tk_status;
1636 case -EAGAIN:
1637 goto out_eagain;
1638 case -NFS4ERR_FATAL_IOERROR:
1639 task->tk_status = -EIO;
1640 return 0;
1641 }
1642
1643 return 0;
1644 out_eagain:
1645 rpc_restart_call_prepare(task);
1646 return -EAGAIN;
1647 }
1648
1649 static bool
ff_layout_need_layoutcommit(struct pnfs_layout_segment * lseg)1650 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1651 {
1652 return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1653 }
1654
1655 /*
1656 * We reference the rpc_cred of the first WRITE that triggers the need for
1657 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1658 * rfc5661 is not clear about which credential should be used.
1659 *
1660 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1661 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1662 * we always send layoutcommit after DS writes.
1663 */
1664 static void
ff_layout_set_layoutcommit(struct inode * inode,struct pnfs_layout_segment * lseg,loff_t end_offset)1665 ff_layout_set_layoutcommit(struct inode *inode,
1666 struct pnfs_layout_segment *lseg,
1667 loff_t end_offset)
1668 {
1669 if (!ff_layout_need_layoutcommit(lseg))
1670 return;
1671
1672 pnfs_set_layoutcommit(inode, lseg, end_offset);
1673 dprintk("%s inode %llu pls_end_pos %llu\n", __func__, inode->i_ino,
1674 (unsigned long long) NFS_I(inode)->layout->plh_lwb);
1675 }
1676
ff_layout_read_record_layoutstats_start(struct rpc_task * task,struct nfs_pgio_header * hdr)1677 static void ff_layout_read_record_layoutstats_start(struct rpc_task *task,
1678 struct nfs_pgio_header *hdr)
1679 {
1680 struct nfs4_ff_layout_mirror *mirror;
1681 u32 dss_id;
1682
1683 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1684 return;
1685
1686 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx);
1687 dss_id = nfs4_ff_layout_calc_dss_id(
1688 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit,
1689 mirror->dss_count,
1690 hdr->args.offset);
1691
1692 nfs4_ff_layout_stat_io_start_read(
1693 hdr->inode,
1694 mirror,
1695 dss_id,
1696 hdr->args.count,
1697 task->tk_start);
1698 }
1699
ff_layout_read_record_layoutstats_done(struct rpc_task * task,struct nfs_pgio_header * hdr)1700 static void ff_layout_read_record_layoutstats_done(struct rpc_task *task,
1701 struct nfs_pgio_header *hdr)
1702 {
1703 struct nfs4_ff_layout_mirror *mirror;
1704 u32 dss_id;
1705
1706 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1707 return;
1708
1709 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx);
1710 dss_id = nfs4_ff_layout_calc_dss_id(
1711 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit,
1712 mirror->dss_count,
1713 hdr->args.offset);
1714
1715 nfs4_ff_layout_stat_io_end_read(
1716 task,
1717 mirror,
1718 dss_id,
1719 hdr->args.count,
1720 hdr->res.count);
1721 set_bit(NFS_LSEG_LAYOUTRETURN, &hdr->lseg->pls_flags);
1722 }
1723
ff_layout_read_prepare_common(struct rpc_task * task,struct nfs_pgio_header * hdr)1724 static int ff_layout_read_prepare_common(struct rpc_task *task,
1725 struct nfs_pgio_header *hdr)
1726 {
1727 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1728 rpc_exit(task, -EIO);
1729 return -EIO;
1730 }
1731
1732 if (!pnfs_is_valid_lseg(hdr->lseg)) {
1733 rpc_exit(task, -EAGAIN);
1734 return -EAGAIN;
1735 }
1736
1737 ff_layout_read_record_layoutstats_start(task, hdr);
1738 return 0;
1739 }
1740
1741 /*
1742 * Call ops for the async read/write cases
1743 * In the case of dense layouts, the offset needs to be reset to its
1744 * original value.
1745 */
ff_layout_read_prepare_v3(struct rpc_task * task,void * data)1746 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1747 {
1748 struct nfs_pgio_header *hdr = data;
1749
1750 if (ff_layout_read_prepare_common(task, hdr))
1751 return;
1752
1753 rpc_call_start(task);
1754 }
1755
ff_layout_read_prepare_v4(struct rpc_task * task,void * data)1756 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1757 {
1758 struct nfs_pgio_header *hdr = data;
1759
1760 if (nfs4_setup_sequence(hdr->ds_clp,
1761 &hdr->args.seq_args,
1762 &hdr->res.seq_res,
1763 task))
1764 return;
1765
1766 ff_layout_read_prepare_common(task, hdr);
1767 }
1768
ff_layout_read_call_done(struct rpc_task * task,void * data)1769 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1770 {
1771 struct nfs_pgio_header *hdr = data;
1772
1773 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1774 task->tk_status == 0) {
1775 nfs4_sequence_done(task, &hdr->res.seq_res);
1776 return;
1777 }
1778
1779 /* Note this may cause RPC to be resent */
1780 hdr->mds_ops->rpc_call_done(task, hdr);
1781 }
1782
ff_layout_read_count_stats(struct rpc_task * task,void * data)1783 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1784 {
1785 struct nfs_pgio_header *hdr = data;
1786
1787 ff_layout_read_record_layoutstats_done(task, hdr);
1788 rpc_count_iostats_metrics(task,
1789 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1790 }
1791
ff_layout_read_release(void * data)1792 static void ff_layout_read_release(void *data)
1793 {
1794 struct nfs_pgio_header *hdr = data;
1795
1796 ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
1797 if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags))
1798 ff_layout_resend_pnfs_read(hdr);
1799 else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags))
1800 ff_layout_reset_read(hdr);
1801 pnfs_generic_rw_release(data);
1802 }
1803
1804
ff_layout_write_done_cb(struct rpc_task * task,struct nfs_pgio_header * hdr)1805 static int ff_layout_write_done_cb(struct rpc_task *task,
1806 struct nfs_pgio_header *hdr)
1807 {
1808 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(hdr->lseg);
1809 u32 dss_id = nfs4_ff_layout_calc_dss_id(
1810 flseg->stripe_unit,
1811 flseg->mirror_array[hdr->pgio_mirror_idx]->dss_count,
1812 hdr->args.offset);
1813 loff_t end_offs = 0;
1814 int err;
1815
1816 if (task->tk_status < 0) {
1817 ff_layout_io_track_ds_error(hdr->lseg,
1818 hdr->pgio_mirror_idx, dss_id,
1819 hdr->args.offset, hdr->args.count,
1820 &hdr->res.op_status, OP_WRITE,
1821 task->tk_status);
1822 trace_ff_layout_write_error(hdr, task->tk_status);
1823 }
1824
1825 err = ff_layout_async_handle_error(task, hdr->res.op_status,
1826 hdr->args.context->state,
1827 hdr->ds_clp, hdr->lseg,
1828 hdr->pgio_mirror_idx,
1829 dss_id);
1830
1831 trace_nfs4_pnfs_write(hdr, err);
1832 clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
1833 clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
1834 switch (err) {
1835 case -NFS4ERR_RESET_TO_PNFS:
1836 set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
1837 return task->tk_status;
1838 case -NFS4ERR_RESET_TO_MDS:
1839 set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
1840 return task->tk_status;
1841 case -EAGAIN:
1842 return -EAGAIN;
1843 case -NFS4ERR_FATAL_IOERROR:
1844 task->tk_status = -EIO;
1845 return 0;
1846 }
1847
1848 if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1849 hdr->res.verf->committed == NFS_DATA_SYNC)
1850 end_offs = hdr->mds_offset + (loff_t)hdr->res.count;
1851
1852 /* Note: if the write is unstable, don't set end_offs until commit */
1853 ff_layout_set_layoutcommit(hdr->inode, hdr->lseg, end_offs);
1854
1855 /* zero out fattr since we don't care DS attr at all */
1856 hdr->fattr.valid = 0;
1857 if (task->tk_status >= 0)
1858 nfs_writeback_update_inode(hdr);
1859
1860 return 0;
1861 }
1862
ff_layout_commit_done_cb(struct rpc_task * task,struct nfs_commit_data * data)1863 static int ff_layout_commit_done_cb(struct rpc_task *task,
1864 struct nfs_commit_data *data)
1865 {
1866 int err;
1867 u32 idx = calc_mirror_idx_from_commit(data->lseg, data->ds_commit_index);
1868 u32 dss_id = calc_dss_id_from_commit(data->lseg, data->ds_commit_index);
1869
1870 if (task->tk_status < 0) {
1871 ff_layout_io_track_ds_error(data->lseg, idx, dss_id,
1872 data->args.offset, data->args.count,
1873 &data->res.op_status, OP_COMMIT,
1874 task->tk_status);
1875 trace_ff_layout_commit_error(data, task->tk_status);
1876 }
1877
1878 err = ff_layout_async_handle_error(task, data->res.op_status,
1879 NULL, data->ds_clp, data->lseg, idx,
1880 dss_id);
1881
1882 trace_nfs4_pnfs_commit_ds(data, err);
1883 switch (err) {
1884 case -NFS4ERR_RESET_TO_PNFS:
1885 pnfs_generic_prepare_to_resend_writes(data);
1886 return -EAGAIN;
1887 case -NFS4ERR_RESET_TO_MDS:
1888 pnfs_generic_prepare_to_resend_writes(data);
1889 return -EAGAIN;
1890 case -EAGAIN:
1891 rpc_restart_call_prepare(task);
1892 return -EAGAIN;
1893 case -NFS4ERR_FATAL_IOERROR:
1894 task->tk_status = -EIO;
1895 return 0;
1896 }
1897
1898 ff_layout_set_layoutcommit(data->inode, data->lseg, data->lwb);
1899 return 0;
1900 }
1901
ff_layout_write_record_layoutstats_start(struct rpc_task * task,struct nfs_pgio_header * hdr)1902 static void ff_layout_write_record_layoutstats_start(struct rpc_task *task,
1903 struct nfs_pgio_header *hdr)
1904 {
1905 struct nfs4_ff_layout_mirror *mirror;
1906 u32 dss_id;
1907
1908 if (test_and_set_bit(NFS_IOHDR_STAT, &hdr->flags))
1909 return;
1910
1911 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx);
1912 dss_id = nfs4_ff_layout_calc_dss_id(
1913 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit,
1914 mirror->dss_count,
1915 hdr->args.offset);
1916
1917 nfs4_ff_layout_stat_io_start_write(
1918 hdr->inode,
1919 mirror,
1920 dss_id,
1921 hdr->args.count,
1922 task->tk_start);
1923 }
1924
ff_layout_write_record_layoutstats_done(struct rpc_task * task,struct nfs_pgio_header * hdr)1925 static void ff_layout_write_record_layoutstats_done(struct rpc_task *task,
1926 struct nfs_pgio_header *hdr)
1927 {
1928 struct nfs4_ff_layout_mirror *mirror;
1929 u32 dss_id;
1930
1931 if (!test_and_clear_bit(NFS_IOHDR_STAT, &hdr->flags))
1932 return;
1933
1934 mirror = FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx);
1935 dss_id = nfs4_ff_layout_calc_dss_id(
1936 FF_LAYOUT_LSEG(hdr->lseg)->stripe_unit,
1937 mirror->dss_count,
1938 hdr->args.offset);
1939
1940 nfs4_ff_layout_stat_io_end_write(
1941 task,
1942 mirror,
1943 dss_id,
1944 hdr->args.count,
1945 hdr->res.count,
1946 hdr->res.verf->committed);
1947 set_bit(NFS_LSEG_LAYOUTRETURN, &hdr->lseg->pls_flags);
1948 }
1949
ff_layout_write_prepare_common(struct rpc_task * task,struct nfs_pgio_header * hdr)1950 static int ff_layout_write_prepare_common(struct rpc_task *task,
1951 struct nfs_pgio_header *hdr)
1952 {
1953 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1954 rpc_exit(task, -EIO);
1955 return -EIO;
1956 }
1957
1958 if (!pnfs_is_valid_lseg(hdr->lseg)) {
1959 rpc_exit(task, -EAGAIN);
1960 return -EAGAIN;
1961 }
1962
1963 ff_layout_write_record_layoutstats_start(task, hdr);
1964 return 0;
1965 }
1966
ff_layout_write_prepare_v3(struct rpc_task * task,void * data)1967 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1968 {
1969 struct nfs_pgio_header *hdr = data;
1970
1971 if (ff_layout_write_prepare_common(task, hdr))
1972 return;
1973
1974 rpc_call_start(task);
1975 }
1976
ff_layout_write_prepare_v4(struct rpc_task * task,void * data)1977 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1978 {
1979 struct nfs_pgio_header *hdr = data;
1980
1981 if (nfs4_setup_sequence(hdr->ds_clp,
1982 &hdr->args.seq_args,
1983 &hdr->res.seq_res,
1984 task))
1985 return;
1986
1987 ff_layout_write_prepare_common(task, hdr);
1988 }
1989
ff_layout_write_call_done(struct rpc_task * task,void * data)1990 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1991 {
1992 struct nfs_pgio_header *hdr = data;
1993
1994 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1995 task->tk_status == 0) {
1996 nfs4_sequence_done(task, &hdr->res.seq_res);
1997 return;
1998 }
1999
2000 /* Note this may cause RPC to be resent */
2001 hdr->mds_ops->rpc_call_done(task, hdr);
2002 }
2003
ff_layout_write_count_stats(struct rpc_task * task,void * data)2004 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
2005 {
2006 struct nfs_pgio_header *hdr = data;
2007
2008 ff_layout_write_record_layoutstats_done(task, hdr);
2009 rpc_count_iostats_metrics(task,
2010 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
2011 }
2012
ff_layout_write_release(void * data)2013 static void ff_layout_write_release(void *data)
2014 {
2015 struct nfs_pgio_header *hdr = data;
2016
2017 ff_layout_write_record_layoutstats_done(&hdr->task, hdr);
2018 if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags)) {
2019 ff_layout_send_layouterror(hdr->lseg);
2020 ff_layout_reset_write(hdr, true);
2021 } else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags))
2022 ff_layout_reset_write(hdr, false);
2023 pnfs_generic_rw_release(data);
2024 }
2025
ff_layout_commit_record_layoutstats_start(struct rpc_task * task,struct nfs_commit_data * cdata)2026 static void ff_layout_commit_record_layoutstats_start(struct rpc_task *task,
2027 struct nfs_commit_data *cdata)
2028 {
2029 u32 idx, dss_id;
2030
2031 if (test_and_set_bit(NFS_IOHDR_STAT, &cdata->flags))
2032 return;
2033
2034 idx = calc_mirror_idx_from_commit(cdata->lseg, cdata->ds_commit_index);
2035 dss_id = calc_dss_id_from_commit(cdata->lseg, cdata->ds_commit_index);
2036 nfs4_ff_layout_stat_io_start_write(cdata->inode,
2037 FF_LAYOUT_COMP(cdata->lseg, idx),
2038 dss_id,
2039 0, task->tk_start);
2040 }
2041
ff_layout_commit_record_layoutstats_done(struct rpc_task * task,struct nfs_commit_data * cdata)2042 static void ff_layout_commit_record_layoutstats_done(struct rpc_task *task,
2043 struct nfs_commit_data *cdata)
2044 {
2045 struct nfs_page *req;
2046 __u64 count = 0;
2047 u32 idx, dss_id;
2048
2049 if (!test_and_clear_bit(NFS_IOHDR_STAT, &cdata->flags))
2050 return;
2051
2052 if (task->tk_status == 0) {
2053 list_for_each_entry(req, &cdata->pages, wb_list)
2054 count += req->wb_bytes;
2055 }
2056
2057 idx = calc_mirror_idx_from_commit(cdata->lseg, cdata->ds_commit_index);
2058 dss_id = calc_dss_id_from_commit(cdata->lseg, cdata->ds_commit_index);
2059 nfs4_ff_layout_stat_io_end_write(task,
2060 FF_LAYOUT_COMP(cdata->lseg, idx),
2061 dss_id,
2062 count, count, NFS_FILE_SYNC);
2063 set_bit(NFS_LSEG_LAYOUTRETURN, &cdata->lseg->pls_flags);
2064 }
2065
ff_layout_commit_prepare_common(struct rpc_task * task,struct nfs_commit_data * cdata)2066 static int ff_layout_commit_prepare_common(struct rpc_task *task,
2067 struct nfs_commit_data *cdata)
2068 {
2069 if (!pnfs_is_valid_lseg(cdata->lseg)) {
2070 rpc_exit(task, -EAGAIN);
2071 return -EAGAIN;
2072 }
2073
2074 ff_layout_commit_record_layoutstats_start(task, cdata);
2075 return 0;
2076 }
2077
ff_layout_commit_prepare_v3(struct rpc_task * task,void * data)2078 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
2079 {
2080 if (ff_layout_commit_prepare_common(task, data))
2081 return;
2082
2083 rpc_call_start(task);
2084 }
2085
ff_layout_commit_prepare_v4(struct rpc_task * task,void * data)2086 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
2087 {
2088 struct nfs_commit_data *wdata = data;
2089
2090 if (nfs4_setup_sequence(wdata->ds_clp,
2091 &wdata->args.seq_args,
2092 &wdata->res.seq_res,
2093 task))
2094 return;
2095 ff_layout_commit_prepare_common(task, data);
2096 }
2097
ff_layout_commit_done(struct rpc_task * task,void * data)2098 static void ff_layout_commit_done(struct rpc_task *task, void *data)
2099 {
2100 pnfs_generic_write_commit_done(task, data);
2101 }
2102
ff_layout_commit_count_stats(struct rpc_task * task,void * data)2103 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
2104 {
2105 struct nfs_commit_data *cdata = data;
2106
2107 ff_layout_commit_record_layoutstats_done(task, cdata);
2108 rpc_count_iostats_metrics(task,
2109 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
2110 }
2111
ff_layout_commit_release(void * data)2112 static void ff_layout_commit_release(void *data)
2113 {
2114 struct nfs_commit_data *cdata = data;
2115
2116 ff_layout_commit_record_layoutstats_done(&cdata->task, cdata);
2117 pnfs_generic_commit_release(data);
2118 }
2119
2120 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
2121 .rpc_call_prepare = ff_layout_read_prepare_v3,
2122 .rpc_call_done = ff_layout_read_call_done,
2123 .rpc_count_stats = ff_layout_read_count_stats,
2124 .rpc_release = ff_layout_read_release,
2125 };
2126
2127 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
2128 .rpc_call_prepare = ff_layout_read_prepare_v4,
2129 .rpc_call_done = ff_layout_read_call_done,
2130 .rpc_count_stats = ff_layout_read_count_stats,
2131 .rpc_release = ff_layout_read_release,
2132 };
2133
2134 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
2135 .rpc_call_prepare = ff_layout_write_prepare_v3,
2136 .rpc_call_done = ff_layout_write_call_done,
2137 .rpc_count_stats = ff_layout_write_count_stats,
2138 .rpc_release = ff_layout_write_release,
2139 };
2140
2141 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
2142 .rpc_call_prepare = ff_layout_write_prepare_v4,
2143 .rpc_call_done = ff_layout_write_call_done,
2144 .rpc_count_stats = ff_layout_write_count_stats,
2145 .rpc_release = ff_layout_write_release,
2146 };
2147
2148 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
2149 .rpc_call_prepare = ff_layout_commit_prepare_v3,
2150 .rpc_call_done = ff_layout_commit_done,
2151 .rpc_count_stats = ff_layout_commit_count_stats,
2152 .rpc_release = ff_layout_commit_release,
2153 };
2154
2155 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
2156 .rpc_call_prepare = ff_layout_commit_prepare_v4,
2157 .rpc_call_done = ff_layout_commit_done,
2158 .rpc_count_stats = ff_layout_commit_count_stats,
2159 .rpc_release = ff_layout_commit_release,
2160 };
2161
2162 static enum pnfs_try_status
ff_layout_read_pagelist(struct nfs_pgio_header * hdr)2163 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
2164 {
2165 struct pnfs_layout_segment *lseg = hdr->lseg;
2166 struct nfs4_pnfs_ds *ds;
2167 struct rpc_clnt *ds_clnt;
2168 struct nfsd_file *localio;
2169 struct nfs4_ff_layout_mirror *mirror;
2170 const struct cred *ds_cred;
2171 loff_t offset = hdr->args.offset;
2172 u32 idx = hdr->pgio_mirror_idx;
2173 int vers;
2174 struct nfs_fh *fh;
2175 u32 dss_id;
2176 bool ds_fatal_error = false;
2177
2178 dprintk("--> %s ino %llu pgbase %u req %zu@%llu\n",
2179 __func__, hdr->inode->i_ino,
2180 hdr->args.pgbase, (size_t)hdr->args.count, offset);
2181
2182 mirror = FF_LAYOUT_COMP(lseg, idx);
2183 dss_id = nfs4_ff_layout_calc_dss_id(
2184 FF_LAYOUT_LSEG(lseg)->stripe_unit,
2185 mirror->dss_count,
2186 offset);
2187 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, false);
2188 if (IS_ERR(ds)) {
2189 ds_fatal_error = nfs_error_is_fatal(PTR_ERR(ds));
2190 goto out_failed;
2191 }
2192
2193 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp,
2194 hdr->inode, dss_id);
2195 if (IS_ERR(ds_clnt))
2196 goto out_failed;
2197
2198 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, hdr->cred, dss_id);
2199 if (!ds_cred)
2200 goto out_failed;
2201
2202 vers = nfs4_ff_layout_ds_version(mirror, dss_id);
2203
2204 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
2205 ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count), vers);
2206
2207 hdr->pgio_done_cb = ff_layout_read_done_cb;
2208 refcount_inc(&ds->ds_clp->cl_count);
2209 hdr->ds_clp = ds->ds_clp;
2210 fh = nfs4_ff_layout_select_ds_fh(mirror, dss_id);
2211 if (fh)
2212 hdr->args.fh = fh;
2213
2214 nfs4_ff_layout_select_ds_stateid(mirror, dss_id, &hdr->args.stateid);
2215
2216 /*
2217 * Note that if we ever decide to split across DSes,
2218 * then we may need to handle dense-like offsets.
2219 */
2220 hdr->args.offset = offset;
2221 hdr->mds_offset = offset;
2222
2223 /* Start IO accounting for local read */
2224 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh,
2225 FMODE_READ);
2226 if (localio) {
2227 hdr->task.tk_start = ktime_get();
2228 ff_layout_read_record_layoutstats_start(&hdr->task, hdr);
2229 }
2230
2231 /* Perform an asynchronous read to ds */
2232 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
2233 vers == 3 ? &ff_layout_read_call_ops_v3 :
2234 &ff_layout_read_call_ops_v4,
2235 0, RPC_TASK_SOFTCONN, localio);
2236 put_cred(ds_cred);
2237 return PNFS_ATTEMPTED;
2238
2239 out_failed:
2240 if (ff_layout_avoid_mds_available_ds(lseg) && !ds_fatal_error)
2241 return PNFS_TRY_AGAIN;
2242 if (ff_layout_no_fallback_to_mds(lseg)) {
2243 /*
2244 * FF_FLAGS_NO_IO_THRU_MDS: force fresh LAYOUTGET,
2245 * never fall through to MDS I/O.
2246 */
2247 pnfs_error_mark_layout_for_return(hdr->inode, lseg);
2248 return PNFS_TRY_AGAIN;
2249 }
2250 trace_pnfs_mds_fallback_read_pagelist(hdr->inode,
2251 hdr->args.offset, hdr->args.count,
2252 IOMODE_READ, NFS_I(hdr->inode)->layout, lseg);
2253 return PNFS_NOT_ATTEMPTED;
2254 }
2255
2256 /* Perform async writes. */
2257 static enum pnfs_try_status
ff_layout_write_pagelist(struct nfs_pgio_header * hdr,int sync)2258 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
2259 {
2260 struct pnfs_layout_segment *lseg = hdr->lseg;
2261 struct nfs4_pnfs_ds *ds;
2262 struct rpc_clnt *ds_clnt;
2263 struct nfsd_file *localio;
2264 struct nfs4_ff_layout_mirror *mirror;
2265 const struct cred *ds_cred;
2266 loff_t offset = hdr->args.offset;
2267 int vers;
2268 struct nfs_fh *fh;
2269 u32 idx = hdr->pgio_mirror_idx;
2270 u32 dss_id;
2271 bool ds_fatal_error = false;
2272
2273 mirror = FF_LAYOUT_COMP(lseg, idx);
2274 dss_id = nfs4_ff_layout_calc_dss_id(
2275 FF_LAYOUT_LSEG(lseg)->stripe_unit,
2276 mirror->dss_count,
2277 offset);
2278 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, true);
2279 if (IS_ERR(ds)) {
2280 ds_fatal_error = nfs_error_is_fatal(PTR_ERR(ds));
2281 goto out_failed;
2282 }
2283
2284 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp,
2285 hdr->inode, dss_id);
2286 if (IS_ERR(ds_clnt))
2287 goto out_failed;
2288
2289 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, hdr->cred, dss_id);
2290 if (!ds_cred)
2291 goto out_failed;
2292
2293 vers = nfs4_ff_layout_ds_version(mirror, dss_id);
2294
2295 dprintk("%s ino %llu sync %d req %zu@%llu DS: %s cl_count %d vers %d\n",
2296 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
2297 offset, ds->ds_remotestr, refcount_read(&ds->ds_clp->cl_count),
2298 vers);
2299
2300 hdr->pgio_done_cb = ff_layout_write_done_cb;
2301 refcount_inc(&ds->ds_clp->cl_count);
2302 hdr->ds_clp = ds->ds_clp;
2303 hdr->ds_commit_idx = calc_commit_idx(lseg, idx, dss_id);
2304 fh = nfs4_ff_layout_select_ds_fh(mirror, dss_id);
2305 if (fh)
2306 hdr->args.fh = fh;
2307
2308 nfs4_ff_layout_select_ds_stateid(mirror, dss_id, &hdr->args.stateid);
2309
2310 /*
2311 * Note that if we ever decide to split across DSes,
2312 * then we may need to handle dense-like offsets.
2313 */
2314 hdr->args.offset = offset;
2315
2316 /* Start IO accounting for local write */
2317 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh,
2318 FMODE_READ|FMODE_WRITE);
2319 if (localio) {
2320 hdr->task.tk_start = ktime_get();
2321 ff_layout_write_record_layoutstats_start(&hdr->task, hdr);
2322 }
2323
2324 /* Perform an asynchronous write */
2325 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
2326 vers == 3 ? &ff_layout_write_call_ops_v3 :
2327 &ff_layout_write_call_ops_v4,
2328 sync, RPC_TASK_SOFTCONN, localio);
2329 put_cred(ds_cred);
2330 return PNFS_ATTEMPTED;
2331
2332 out_failed:
2333 if (ff_layout_avoid_mds_available_ds(lseg) && !ds_fatal_error)
2334 return PNFS_TRY_AGAIN;
2335 if (ff_layout_no_fallback_to_mds(lseg)) {
2336 /*
2337 * FF_FLAGS_NO_IO_THRU_MDS: force fresh LAYOUTGET,
2338 * never fall through to MDS I/O.
2339 */
2340 pnfs_error_mark_layout_for_return(hdr->inode, lseg);
2341 return PNFS_TRY_AGAIN;
2342 }
2343 trace_pnfs_mds_fallback_write_pagelist(hdr->inode,
2344 hdr->args.offset, hdr->args.count,
2345 IOMODE_RW, NFS_I(hdr->inode)->layout, lseg);
2346 return PNFS_NOT_ATTEMPTED;
2347 }
2348
2349 static struct nfs_fh *
select_ds_fh_from_commit(struct pnfs_layout_segment * lseg,u32 i,u32 dss_id)2350 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i, u32 dss_id)
2351 {
2352 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
2353
2354 /* FIXME: Assume that there is only one NFS version available
2355 * for the DS.
2356 */
2357 return &flseg->mirror_array[i]->dss[dss_id].fh_versions[0];
2358 }
2359
ff_layout_initiate_commit(struct nfs_commit_data * data,int how)2360 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
2361 {
2362 struct pnfs_layout_segment *lseg = data->lseg;
2363 struct nfs4_pnfs_ds *ds;
2364 struct rpc_clnt *ds_clnt;
2365 struct nfsd_file *localio;
2366 struct nfs4_ff_layout_mirror *mirror;
2367 const struct cred *ds_cred;
2368 u32 idx, dss_id;
2369 int vers, ret;
2370 struct nfs_fh *fh;
2371
2372 if (!lseg || !(pnfs_is_valid_lseg(lseg) ||
2373 test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags)))
2374 goto out_err;
2375
2376 idx = calc_mirror_idx_from_commit(lseg, data->ds_commit_index);
2377 mirror = FF_LAYOUT_COMP(lseg, idx);
2378 dss_id = calc_dss_id_from_commit(lseg, data->ds_commit_index);
2379 ds = nfs4_ff_layout_prepare_ds(lseg, mirror, dss_id, true);
2380 if (IS_ERR(ds))
2381 goto out_err;
2382
2383 ds_clnt = nfs4_ff_find_or_create_ds_client(mirror, ds->ds_clp,
2384 data->inode, dss_id);
2385 if (IS_ERR(ds_clnt))
2386 goto out_err;
2387
2388 ds_cred = ff_layout_get_ds_cred(mirror, &lseg->pls_range, data->cred, dss_id);
2389 if (!ds_cred)
2390 goto out_err;
2391
2392 vers = nfs4_ff_layout_ds_version(mirror, dss_id);
2393
2394 dprintk("%s ino %llu, how %d cl_count %d vers %d\n", __func__,
2395 data->inode->i_ino, how, refcount_read(&ds->ds_clp->cl_count),
2396 vers);
2397 data->commit_done_cb = ff_layout_commit_done_cb;
2398 data->cred = ds_cred;
2399 refcount_inc(&ds->ds_clp->cl_count);
2400 data->ds_clp = ds->ds_clp;
2401 fh = select_ds_fh_from_commit(lseg, idx, dss_id);
2402 if (fh)
2403 data->args.fh = fh;
2404
2405 /* Start IO accounting for local commit */
2406 localio = ff_local_open_fh(lseg, idx, dss_id, ds->ds_clp, ds_cred, fh,
2407 FMODE_READ|FMODE_WRITE);
2408 if (localio) {
2409 data->task.tk_start = ktime_get();
2410 ff_layout_commit_record_layoutstats_start(&data->task, data);
2411 }
2412
2413 ret = nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
2414 vers == 3 ? &ff_layout_commit_call_ops_v3 :
2415 &ff_layout_commit_call_ops_v4,
2416 how, RPC_TASK_SOFTCONN, localio);
2417 put_cred(ds_cred);
2418 return ret;
2419 out_err:
2420 pnfs_generic_prepare_to_resend_writes(data);
2421 pnfs_generic_commit_release(data);
2422 return -EAGAIN;
2423 }
2424
2425 static int
ff_layout_commit_pagelist(struct inode * inode,struct list_head * mds_pages,int how,struct nfs_commit_info * cinfo)2426 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
2427 int how, struct nfs_commit_info *cinfo)
2428 {
2429 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
2430 ff_layout_initiate_commit);
2431 }
2432
ff_layout_match_rw(const struct rpc_task * task,const struct nfs_pgio_header * hdr,const struct pnfs_layout_segment * lseg)2433 static bool ff_layout_match_rw(const struct rpc_task *task,
2434 const struct nfs_pgio_header *hdr,
2435 const struct pnfs_layout_segment *lseg)
2436 {
2437 return hdr->lseg == lseg;
2438 }
2439
ff_layout_match_commit(const struct rpc_task * task,const struct nfs_commit_data * cdata,const struct pnfs_layout_segment * lseg)2440 static bool ff_layout_match_commit(const struct rpc_task *task,
2441 const struct nfs_commit_data *cdata,
2442 const struct pnfs_layout_segment *lseg)
2443 {
2444 return cdata->lseg == lseg;
2445 }
2446
ff_layout_match_io(const struct rpc_task * task,const void * data)2447 static bool ff_layout_match_io(const struct rpc_task *task, const void *data)
2448 {
2449 const struct rpc_call_ops *ops = task->tk_ops;
2450
2451 if (ops == &ff_layout_read_call_ops_v3 ||
2452 ops == &ff_layout_read_call_ops_v4 ||
2453 ops == &ff_layout_write_call_ops_v3 ||
2454 ops == &ff_layout_write_call_ops_v4)
2455 return ff_layout_match_rw(task, task->tk_calldata, data);
2456 if (ops == &ff_layout_commit_call_ops_v3 ||
2457 ops == &ff_layout_commit_call_ops_v4)
2458 return ff_layout_match_commit(task, task->tk_calldata, data);
2459 return false;
2460 }
2461
ff_layout_cancel_io(struct pnfs_layout_segment * lseg)2462 static void ff_layout_cancel_io(struct pnfs_layout_segment *lseg)
2463 {
2464 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
2465 struct nfs4_ff_layout_mirror *mirror;
2466 struct nfs4_ff_layout_ds *mirror_ds;
2467 struct nfs4_pnfs_ds *ds;
2468 struct nfs_client *ds_clp;
2469 struct rpc_clnt *clnt;
2470 u32 idx, dss_id;
2471
2472 for (idx = 0; idx < flseg->mirror_array_cnt; idx++) {
2473 mirror = flseg->mirror_array[idx];
2474 for (dss_id = 0; dss_id < mirror->dss_count; dss_id++) {
2475 mirror_ds = mirror->dss[dss_id].mirror_ds;
2476 if (IS_ERR_OR_NULL(mirror_ds))
2477 continue;
2478 ds = mirror->dss[dss_id].mirror_ds->ds;
2479 if (!ds)
2480 continue;
2481 ds_clp = ds->ds_clp;
2482 if (!ds_clp)
2483 continue;
2484 clnt = ds_clp->cl_rpcclient;
2485 if (!clnt)
2486 continue;
2487 if (!rpc_cancel_tasks(clnt, -ECANCELED,
2488 ff_layout_match_io, lseg))
2489 continue;
2490 rpc_clnt_disconnect(clnt);
2491 }
2492 }
2493 }
2494
2495 static struct pnfs_ds_commit_info *
ff_layout_get_ds_info(struct inode * inode)2496 ff_layout_get_ds_info(struct inode *inode)
2497 {
2498 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
2499
2500 if (layout == NULL)
2501 return NULL;
2502
2503 return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
2504 }
2505
2506 static void
ff_layout_setup_ds_info(struct pnfs_ds_commit_info * fl_cinfo,struct pnfs_layout_segment * lseg)2507 ff_layout_setup_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
2508 struct pnfs_layout_segment *lseg)
2509 {
2510 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
2511 struct inode *inode = lseg->pls_layout->plh_inode;
2512 struct pnfs_commit_array *array, *new;
2513 u32 size = flseg->mirror_array_cnt * flseg->mirror_array[0]->dss_count;
2514
2515 new = pnfs_alloc_commit_array(size,
2516 nfs_io_gfp_mask());
2517 if (new) {
2518 spin_lock(&inode->i_lock);
2519 array = pnfs_add_commit_array(fl_cinfo, new, lseg);
2520 spin_unlock(&inode->i_lock);
2521 if (array != new)
2522 pnfs_free_commit_array(new);
2523 }
2524 }
2525
2526 static void
ff_layout_release_ds_info(struct pnfs_ds_commit_info * fl_cinfo,struct inode * inode)2527 ff_layout_release_ds_info(struct pnfs_ds_commit_info *fl_cinfo,
2528 struct inode *inode)
2529 {
2530 spin_lock(&inode->i_lock);
2531 pnfs_generic_ds_cinfo_destroy(fl_cinfo);
2532 spin_unlock(&inode->i_lock);
2533 }
2534
2535 static void
ff_layout_free_deviceid_node(struct nfs4_deviceid_node * d)2536 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
2537 {
2538 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
2539 id_node));
2540 }
2541
ff_layout_encode_ioerr(struct xdr_stream * xdr,const struct nfs4_layoutreturn_args * args,const struct nfs4_flexfile_layoutreturn_args * ff_args)2542 static int ff_layout_encode_ioerr(struct xdr_stream *xdr,
2543 const struct nfs4_layoutreturn_args *args,
2544 const struct nfs4_flexfile_layoutreturn_args *ff_args)
2545 {
2546 __be32 *start;
2547
2548 start = xdr_reserve_space(xdr, 4);
2549 if (unlikely(!start))
2550 return -E2BIG;
2551
2552 *start = cpu_to_be32(ff_args->num_errors);
2553 /* This assume we always return _ALL_ layouts */
2554 return ff_layout_encode_ds_ioerr(xdr, &ff_args->errors);
2555 }
2556
2557 static void
ff_layout_encode_ff_iostat_head(struct xdr_stream * xdr,const nfs4_stateid * stateid,const struct nfs42_layoutstat_devinfo * devinfo)2558 ff_layout_encode_ff_iostat_head(struct xdr_stream *xdr,
2559 const nfs4_stateid *stateid,
2560 const struct nfs42_layoutstat_devinfo *devinfo)
2561 {
2562 __be32 *p;
2563
2564 p = xdr_reserve_space(xdr, 8 + 8);
2565 p = xdr_encode_hyper(p, devinfo->offset);
2566 p = xdr_encode_hyper(p, devinfo->length);
2567 encode_opaque_fixed(xdr, stateid->data, NFS4_STATEID_SIZE);
2568 p = xdr_reserve_space(xdr, 4*8);
2569 p = xdr_encode_hyper(p, devinfo->read_count);
2570 p = xdr_encode_hyper(p, devinfo->read_bytes);
2571 p = xdr_encode_hyper(p, devinfo->write_count);
2572 p = xdr_encode_hyper(p, devinfo->write_bytes);
2573 encode_opaque_fixed(xdr, devinfo->dev_id.data, NFS4_DEVICEID4_SIZE);
2574 }
2575
2576 static void
ff_layout_encode_ff_iostat(struct xdr_stream * xdr,const nfs4_stateid * stateid,const struct nfs42_layoutstat_devinfo * devinfo)2577 ff_layout_encode_ff_iostat(struct xdr_stream *xdr,
2578 const nfs4_stateid *stateid,
2579 const struct nfs42_layoutstat_devinfo *devinfo)
2580 {
2581 ff_layout_encode_ff_iostat_head(xdr, stateid, devinfo);
2582 ff_layout_encode_ff_layoutupdate(xdr, devinfo,
2583 devinfo->ld_private.data);
2584 }
2585
2586 /* report nothing for now */
ff_layout_encode_iostats_array(struct xdr_stream * xdr,const struct nfs4_layoutreturn_args * args,struct nfs4_flexfile_layoutreturn_args * ff_args)2587 static void ff_layout_encode_iostats_array(struct xdr_stream *xdr,
2588 const struct nfs4_layoutreturn_args *args,
2589 struct nfs4_flexfile_layoutreturn_args *ff_args)
2590 {
2591 __be32 *p;
2592 int i;
2593
2594 p = xdr_reserve_space(xdr, 4);
2595 *p = cpu_to_be32(ff_args->num_dev);
2596 for (i = 0; i < ff_args->num_dev; i++)
2597 ff_layout_encode_ff_iostat(xdr,
2598 &args->layout->plh_stateid,
2599 &ff_args->devinfo[i]);
2600 }
2601
2602 static void
ff_layout_free_iostats_array(struct nfs42_layoutstat_devinfo * devinfo,unsigned int num_entries)2603 ff_layout_free_iostats_array(struct nfs42_layoutstat_devinfo *devinfo,
2604 unsigned int num_entries)
2605 {
2606 unsigned int i;
2607
2608 for (i = 0; i < num_entries; i++) {
2609 if (!devinfo[i].ld_private.ops)
2610 continue;
2611 if (!devinfo[i].ld_private.ops->free)
2612 continue;
2613 devinfo[i].ld_private.ops->free(&devinfo[i].ld_private);
2614 }
2615 }
2616
2617 static struct nfs4_deviceid_node *
ff_layout_alloc_deviceid_node(struct nfs_server * server,struct pnfs_device * pdev,gfp_t gfp_flags)2618 ff_layout_alloc_deviceid_node(struct nfs_server *server,
2619 struct pnfs_device *pdev, gfp_t gfp_flags)
2620 {
2621 struct nfs4_ff_layout_ds *dsaddr;
2622
2623 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
2624 if (!dsaddr)
2625 return NULL;
2626 return &dsaddr->id_node;
2627 }
2628
2629 static void
ff_layout_encode_layoutreturn(struct xdr_stream * xdr,const void * voidargs,const struct nfs4_xdr_opaque_data * ff_opaque)2630 ff_layout_encode_layoutreturn(struct xdr_stream *xdr,
2631 const void *voidargs,
2632 const struct nfs4_xdr_opaque_data *ff_opaque)
2633 {
2634 const struct nfs4_layoutreturn_args *args = voidargs;
2635 struct nfs4_flexfile_layoutreturn_args *ff_args = ff_opaque->data;
2636 struct xdr_buf tmp_buf = {
2637 .head = {
2638 [0] = {
2639 .iov_base = page_address(ff_args->pages[0]),
2640 },
2641 },
2642 .buflen = PAGE_SIZE,
2643 };
2644 struct xdr_stream tmp_xdr;
2645 __be32 *start;
2646
2647 dprintk("%s: Begin\n", __func__);
2648
2649 xdr_init_encode(&tmp_xdr, &tmp_buf, NULL, NULL);
2650
2651 ff_layout_encode_ioerr(&tmp_xdr, args, ff_args);
2652 ff_layout_encode_iostats_array(&tmp_xdr, args, ff_args);
2653
2654 start = xdr_reserve_space(xdr, 4);
2655 *start = cpu_to_be32(tmp_buf.len);
2656 xdr_write_pages(xdr, ff_args->pages, 0, tmp_buf.len);
2657
2658 dprintk("%s: Return\n", __func__);
2659 }
2660
2661 static void
ff_layout_free_layoutreturn(struct nfs4_xdr_opaque_data * args)2662 ff_layout_free_layoutreturn(struct nfs4_xdr_opaque_data *args)
2663 {
2664 struct nfs4_flexfile_layoutreturn_args *ff_args;
2665
2666 if (!args->data)
2667 return;
2668 ff_args = args->data;
2669 args->data = NULL;
2670
2671 ff_layout_free_ds_ioerr(&ff_args->errors);
2672 ff_layout_free_iostats_array(ff_args->devinfo, ff_args->num_dev);
2673
2674 put_page(ff_args->pages[0]);
2675 kfree(ff_args);
2676 }
2677
2678 static const struct nfs4_xdr_opaque_ops layoutreturn_ops = {
2679 .encode = ff_layout_encode_layoutreturn,
2680 .free = ff_layout_free_layoutreturn,
2681 };
2682
2683 static int
ff_layout_prepare_layoutreturn(struct nfs4_layoutreturn_args * args)2684 ff_layout_prepare_layoutreturn(struct nfs4_layoutreturn_args *args)
2685 {
2686 struct nfs4_flexfile_layoutreturn_args *ff_args;
2687 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(args->layout);
2688
2689 ff_args = kmalloc_obj(*ff_args, nfs_io_gfp_mask());
2690 if (!ff_args)
2691 goto out_nomem;
2692 ff_args->pages[0] = alloc_page(nfs_io_gfp_mask());
2693 if (!ff_args->pages[0])
2694 goto out_nomem_free;
2695
2696 INIT_LIST_HEAD(&ff_args->errors);
2697 ff_args->num_errors = ff_layout_fetch_ds_ioerr(args->layout,
2698 &args->range, &ff_args->errors,
2699 FF_LAYOUTRETURN_MAXERR);
2700
2701 spin_lock(&args->inode->i_lock);
2702 ff_args->num_dev = ff_layout_mirror_prepare_stats(
2703 &ff_layout->generic_hdr, &ff_args->devinfo[0],
2704 ARRAY_SIZE(ff_args->devinfo), NFS4_FF_OP_LAYOUTRETURN);
2705 spin_unlock(&args->inode->i_lock);
2706
2707 args->ld_private->ops = &layoutreturn_ops;
2708 args->ld_private->data = ff_args;
2709 return 0;
2710 out_nomem_free:
2711 kfree(ff_args);
2712 out_nomem:
2713 return -ENOMEM;
2714 }
2715
2716 #ifdef CONFIG_NFS_V4_2
2717 void
ff_layout_send_layouterror(struct pnfs_layout_segment * lseg)2718 ff_layout_send_layouterror(struct pnfs_layout_segment *lseg)
2719 {
2720 struct pnfs_layout_hdr *lo = lseg->pls_layout;
2721 struct nfs42_layout_error *errors;
2722 LIST_HEAD(head);
2723
2724 if (!nfs_server_capable(lo->plh_inode, NFS_CAP_LAYOUTERROR))
2725 return;
2726 ff_layout_fetch_ds_ioerr(lo, &lseg->pls_range, &head, -1);
2727 if (list_empty(&head))
2728 return;
2729
2730 errors = kmalloc_objs(*errors, NFS42_LAYOUTERROR_MAX, nfs_io_gfp_mask());
2731 if (errors != NULL) {
2732 const struct nfs4_ff_layout_ds_err *pos;
2733 size_t n = 0;
2734
2735 list_for_each_entry(pos, &head, list) {
2736 errors[n].offset = pos->offset;
2737 errors[n].length = pos->length;
2738 nfs4_stateid_copy(&errors[n].stateid, &pos->stateid);
2739 errors[n].errors[0].dev_id = pos->deviceid;
2740 errors[n].errors[0].status = pos->status;
2741 errors[n].errors[0].opnum = pos->opnum;
2742 n++;
2743 if (!list_is_last(&pos->list, &head) &&
2744 n < NFS42_LAYOUTERROR_MAX)
2745 continue;
2746 if (nfs42_proc_layouterror(lseg, errors, n) < 0)
2747 break;
2748 n = 0;
2749 }
2750 kfree(errors);
2751 }
2752 ff_layout_free_ds_ioerr(&head);
2753 }
2754 #else
2755 void
ff_layout_send_layouterror(struct pnfs_layout_segment * lseg)2756 ff_layout_send_layouterror(struct pnfs_layout_segment *lseg)
2757 {
2758 }
2759 #endif
2760
2761 static int
ff_layout_ntop4(const struct sockaddr * sap,char * buf,const size_t buflen)2762 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
2763 {
2764 const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
2765
2766 return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
2767 }
2768
2769 static size_t
ff_layout_ntop6_noscopeid(const struct sockaddr * sap,char * buf,const int buflen)2770 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
2771 const int buflen)
2772 {
2773 const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
2774 const struct in6_addr *addr = &sin6->sin6_addr;
2775
2776 /*
2777 * RFC 4291, Section 2.2.2
2778 *
2779 * Shorthanded ANY address
2780 */
2781 if (ipv6_addr_any(addr))
2782 return snprintf(buf, buflen, "::");
2783
2784 /*
2785 * RFC 4291, Section 2.2.2
2786 *
2787 * Shorthanded loopback address
2788 */
2789 if (ipv6_addr_loopback(addr))
2790 return snprintf(buf, buflen, "::1");
2791
2792 /*
2793 * RFC 4291, Section 2.2.3
2794 *
2795 * Special presentation address format for mapped v4
2796 * addresses.
2797 */
2798 if (ipv6_addr_v4mapped(addr))
2799 return snprintf(buf, buflen, "::ffff:%pI4",
2800 &addr->s6_addr32[3]);
2801
2802 /*
2803 * RFC 4291, Section 2.2.1
2804 */
2805 return snprintf(buf, buflen, "%pI6c", addr);
2806 }
2807
2808 /* Derived from rpc_sockaddr2uaddr */
2809 static void
ff_layout_encode_netaddr(struct xdr_stream * xdr,struct nfs4_pnfs_ds_addr * da)2810 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
2811 {
2812 struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
2813 char portbuf[RPCBIND_MAXUADDRPLEN];
2814 char addrbuf[RPCBIND_MAXUADDRLEN];
2815 unsigned short port;
2816 int len, netid_len;
2817 __be32 *p;
2818
2819 switch (sap->sa_family) {
2820 case AF_INET:
2821 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
2822 return;
2823 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
2824 break;
2825 case AF_INET6:
2826 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
2827 return;
2828 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
2829 break;
2830 default:
2831 WARN_ON_ONCE(1);
2832 return;
2833 }
2834
2835 snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
2836 len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
2837
2838 netid_len = strlen(da->da_netid);
2839 p = xdr_reserve_space(xdr, 4 + netid_len);
2840 xdr_encode_opaque(p, da->da_netid, netid_len);
2841
2842 p = xdr_reserve_space(xdr, 4 + len);
2843 xdr_encode_opaque(p, addrbuf, len);
2844 }
2845
2846 static void
ff_layout_encode_nfstime(struct xdr_stream * xdr,ktime_t t)2847 ff_layout_encode_nfstime(struct xdr_stream *xdr,
2848 ktime_t t)
2849 {
2850 struct timespec64 ts;
2851 __be32 *p;
2852
2853 p = xdr_reserve_space(xdr, 12);
2854 ts = ktime_to_timespec64(t);
2855 p = xdr_encode_hyper(p, ts.tv_sec);
2856 *p++ = cpu_to_be32(ts.tv_nsec);
2857 }
2858
2859 static void
ff_layout_encode_io_latency(struct xdr_stream * xdr,struct nfs4_ff_io_stat * stat)2860 ff_layout_encode_io_latency(struct xdr_stream *xdr,
2861 struct nfs4_ff_io_stat *stat)
2862 {
2863 __be32 *p;
2864
2865 p = xdr_reserve_space(xdr, 5 * 8);
2866 p = xdr_encode_hyper(p, stat->ops_requested);
2867 p = xdr_encode_hyper(p, stat->bytes_requested);
2868 p = xdr_encode_hyper(p, stat->ops_completed);
2869 p = xdr_encode_hyper(p, stat->bytes_completed);
2870 p = xdr_encode_hyper(p, stat->bytes_not_delivered);
2871 ff_layout_encode_nfstime(xdr, stat->total_busy_time);
2872 ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
2873 }
2874
2875 static void
ff_layout_encode_ff_layoutupdate(struct xdr_stream * xdr,const struct nfs42_layoutstat_devinfo * devinfo,struct nfs4_ff_layout_ds_stripe * dss_info)2876 ff_layout_encode_ff_layoutupdate(struct xdr_stream *xdr,
2877 const struct nfs42_layoutstat_devinfo *devinfo,
2878 struct nfs4_ff_layout_ds_stripe *dss_info)
2879 {
2880 struct nfs4_pnfs_ds_addr *da;
2881 struct nfs4_pnfs_ds *ds = dss_info->mirror_ds->ds;
2882 struct nfs_fh *fh = &dss_info->fh_versions[0];
2883 __be32 *p;
2884
2885 da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2886 dprintk("%s: DS %s: encoding address %s\n",
2887 __func__, ds->ds_remotestr, da->da_remotestr);
2888 /* netaddr4 */
2889 ff_layout_encode_netaddr(xdr, da);
2890 /* nfs_fh4 */
2891 p = xdr_reserve_space(xdr, 4 + fh->size);
2892 xdr_encode_opaque(p, fh->data, fh->size);
2893 /* ff_io_latency4 read */
2894 spin_lock(&dss_info->mirror->lock);
2895 ff_layout_encode_io_latency(xdr,
2896 &dss_info->read_stat.io_stat);
2897 /* ff_io_latency4 write */
2898 ff_layout_encode_io_latency(xdr,
2899 &dss_info->write_stat.io_stat);
2900 spin_unlock(&dss_info->mirror->lock);
2901 /* nfstime4 */
2902 ff_layout_encode_nfstime(xdr,
2903 ktime_sub(ktime_get(),
2904 dss_info->start_time));
2905 /* bool */
2906 p = xdr_reserve_space(xdr, 4);
2907 *p = cpu_to_be32(false);
2908 }
2909
2910 static void
ff_layout_encode_layoutstats(struct xdr_stream * xdr,const void * args,const struct nfs4_xdr_opaque_data * opaque)2911 ff_layout_encode_layoutstats(struct xdr_stream *xdr, const void *args,
2912 const struct nfs4_xdr_opaque_data *opaque)
2913 {
2914 struct nfs42_layoutstat_devinfo *devinfo = container_of(opaque,
2915 struct nfs42_layoutstat_devinfo, ld_private);
2916 __be32 *start;
2917
2918 /* layoutupdate length */
2919 start = xdr_reserve_space(xdr, 4);
2920 ff_layout_encode_ff_layoutupdate(xdr, devinfo, opaque->data);
2921
2922 *start = cpu_to_be32((xdr->p - start - 1) * 4);
2923 }
2924
2925 static void
ff_layout_free_layoutstats(struct nfs4_xdr_opaque_data * opaque)2926 ff_layout_free_layoutstats(struct nfs4_xdr_opaque_data *opaque)
2927 {
2928 struct nfs4_ff_layout_ds_stripe *dss_info = opaque->data;
2929 struct nfs4_ff_layout_mirror *mirror = dss_info->mirror;
2930
2931 ff_layout_put_mirror(mirror);
2932 }
2933
2934 static const struct nfs4_xdr_opaque_ops layoutstat_ops = {
2935 .encode = ff_layout_encode_layoutstats,
2936 .free = ff_layout_free_layoutstats,
2937 };
2938
2939 static int
ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr * lo,struct nfs42_layoutstat_devinfo * devinfo,int dev_limit,enum nfs4_ff_op_type type)2940 ff_layout_mirror_prepare_stats(struct pnfs_layout_hdr *lo,
2941 struct nfs42_layoutstat_devinfo *devinfo,
2942 int dev_limit, enum nfs4_ff_op_type type)
2943 {
2944 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2945 struct nfs4_ff_layout_mirror *mirror;
2946 struct nfs4_ff_layout_ds_stripe *dss_info;
2947 struct nfs4_deviceid_node *dev;
2948 int i = 0, dss_id;
2949
2950 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2951 for (dss_id = 0; dss_id < mirror->dss_count; ++dss_id) {
2952 dss_info = &mirror->dss[dss_id];
2953 if (i >= dev_limit)
2954 break;
2955 if (IS_ERR_OR_NULL(dss_info->mirror_ds))
2956 continue;
2957 if (!test_and_clear_bit(NFS4_FF_MIRROR_STAT_AVAIL,
2958 &mirror->flags) &&
2959 type != NFS4_FF_OP_LAYOUTRETURN)
2960 continue;
2961 /* mirror refcount put in cleanup_layoutstats */
2962 if (!refcount_inc_not_zero(&mirror->ref))
2963 continue;
2964 dev = &dss_info->mirror_ds->id_node;
2965 memcpy(&devinfo->dev_id,
2966 &dev->deviceid,
2967 NFS4_DEVICEID4_SIZE);
2968 devinfo->offset = 0;
2969 devinfo->length = NFS4_MAX_UINT64;
2970 spin_lock(&mirror->lock);
2971 devinfo->read_count =
2972 dss_info->read_stat.io_stat.ops_completed;
2973 devinfo->read_bytes =
2974 dss_info->read_stat.io_stat.bytes_completed;
2975 devinfo->write_count =
2976 dss_info->write_stat.io_stat.ops_completed;
2977 devinfo->write_bytes =
2978 dss_info->write_stat.io_stat.bytes_completed;
2979 spin_unlock(&mirror->lock);
2980 devinfo->layout_type = LAYOUT_FLEX_FILES;
2981 devinfo->ld_private.ops = &layoutstat_ops;
2982 devinfo->ld_private.data = &mirror->dss[dss_id];
2983
2984 devinfo++;
2985 i++;
2986 }
2987 }
2988 return i;
2989 }
2990
ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args * args)2991 static int ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2992 {
2993 struct pnfs_layout_hdr *lo;
2994 struct nfs4_flexfile_layout *ff_layout;
2995 const int dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2996
2997 /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2998 args->devinfo = kmalloc_objs(*args->devinfo, dev_count,
2999 nfs_io_gfp_mask());
3000 if (!args->devinfo)
3001 return -ENOMEM;
3002
3003 spin_lock(&args->inode->i_lock);
3004 lo = NFS_I(args->inode)->layout;
3005 if (lo && pnfs_layout_is_valid(lo)) {
3006 ff_layout = FF_LAYOUT_FROM_HDR(lo);
3007 args->num_dev = ff_layout_mirror_prepare_stats(
3008 &ff_layout->generic_hdr, &args->devinfo[0], dev_count,
3009 NFS4_FF_OP_LAYOUTSTATS);
3010 } else
3011 args->num_dev = 0;
3012 spin_unlock(&args->inode->i_lock);
3013 if (!args->num_dev) {
3014 kfree(args->devinfo);
3015 args->devinfo = NULL;
3016 return -ENOENT;
3017 }
3018
3019 return 0;
3020 }
3021
3022 static int
ff_layout_set_layoutdriver(struct nfs_server * server,const struct nfs_fh * dummy)3023 ff_layout_set_layoutdriver(struct nfs_server *server,
3024 const struct nfs_fh *dummy)
3025 {
3026 #if IS_ENABLED(CONFIG_NFS_V4_2)
3027 server->caps |= NFS_CAP_LAYOUTSTATS | NFS_CAP_REBOOT_LAYOUTRETURN;
3028 #endif
3029 return 0;
3030 }
3031
3032 static const struct pnfs_commit_ops ff_layout_commit_ops = {
3033 .setup_ds_info = ff_layout_setup_ds_info,
3034 .release_ds_info = ff_layout_release_ds_info,
3035 .mark_request_commit = pnfs_layout_mark_request_commit,
3036 .clear_request_commit = pnfs_generic_clear_request_commit,
3037 .scan_commit_lists = pnfs_generic_scan_commit_lists,
3038 .recover_commit_reqs = pnfs_generic_recover_commit_reqs,
3039 .commit_pagelist = ff_layout_commit_pagelist,
3040 };
3041
3042 static struct pnfs_layoutdriver_type flexfilelayout_type = {
3043 .id = LAYOUT_FLEX_FILES,
3044 .name = "LAYOUT_FLEX_FILES",
3045 .owner = THIS_MODULE,
3046 .flags = PNFS_LAYOUTGET_ON_OPEN,
3047 .max_layoutget_response = 4096, /* 1 page or so... */
3048 .set_layoutdriver = ff_layout_set_layoutdriver,
3049 .alloc_layout_hdr = ff_layout_alloc_layout_hdr,
3050 .free_layout_hdr = ff_layout_free_layout_hdr,
3051 .alloc_lseg = ff_layout_alloc_lseg,
3052 .free_lseg = ff_layout_free_lseg,
3053 .add_lseg = ff_layout_add_lseg,
3054 .pg_read_ops = &ff_layout_pg_read_ops,
3055 .pg_write_ops = &ff_layout_pg_write_ops,
3056 .get_ds_info = ff_layout_get_ds_info,
3057 .free_deviceid_node = ff_layout_free_deviceid_node,
3058 .read_pagelist = ff_layout_read_pagelist,
3059 .write_pagelist = ff_layout_write_pagelist,
3060 .alloc_deviceid_node = ff_layout_alloc_deviceid_node,
3061 .prepare_layoutreturn = ff_layout_prepare_layoutreturn,
3062 .sync = pnfs_nfs_generic_sync,
3063 .prepare_layoutstats = ff_layout_prepare_layoutstats,
3064 .cancel_io = ff_layout_cancel_io,
3065 };
3066
nfs4flexfilelayout_init(void)3067 static int __init nfs4flexfilelayout_init(void)
3068 {
3069 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
3070 __func__);
3071 return pnfs_register_layoutdriver(&flexfilelayout_type);
3072 }
3073
nfs4flexfilelayout_exit(void)3074 static void __exit nfs4flexfilelayout_exit(void)
3075 {
3076 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
3077 __func__);
3078 pnfs_unregister_layoutdriver(&flexfilelayout_type);
3079 }
3080
3081 MODULE_ALIAS("nfs-layouttype4-4");
3082
3083 MODULE_LICENSE("GPL");
3084 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
3085
3086 module_init(nfs4flexfilelayout_init);
3087 module_exit(nfs4flexfilelayout_exit);
3088
3089 module_param(io_maxretrans, ushort, 0644);
3090 MODULE_PARM_DESC(io_maxretrans, "The number of times the NFSv4.1 client "
3091 "retries an I/O request before returning an error. ");
3092