Loading...
1/*
2 * pNFS functions to call and manage layout drivers.
3 *
4 * Copyright (c) 2002 [year of first publication]
5 * The Regents of the University of Michigan
6 * All Rights Reserved
7 *
8 * Dean Hildebrand <dhildebz@umich.edu>
9 *
10 * Permission is granted to use, copy, create derivative works, and
11 * redistribute this software and such derivative works for any purpose,
12 * so long as the name of the University of Michigan is not used in
13 * any advertising or publicity pertaining to the use or distribution
14 * of this software without specific, written prior authorization. If
15 * the above copyright notice or any other identification of the
16 * University of Michigan is included in any copy of any portion of
17 * this software, then the disclaimer below must also be included.
18 *
19 * This software is provided as is, without representation or warranty
20 * of any kind either express or implied, including without limitation
21 * the implied warranties of merchantability, fitness for a particular
22 * purpose, or noninfringement. The Regents of the University of
23 * Michigan shall not be liable for any damages, including special,
24 * indirect, incidental, or consequential damages, with respect to any
25 * claim arising out of or in connection with the use of the software,
26 * even if it has been or is hereafter advised of the possibility of
27 * such damages.
28 */
29
30#include <linux/nfs_fs.h>
31#include <linux/nfs_page.h>
32#include <linux/module.h>
33#include <linux/sort.h>
34#include "internal.h"
35#include "pnfs.h"
36#include "iostat.h"
37#include "nfs4trace.h"
38#include "delegation.h"
39#include "nfs42.h"
40
41#define NFSDBG_FACILITY NFSDBG_PNFS
42#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
43
44/* Locking:
45 *
46 * pnfs_spinlock:
47 * protects pnfs_modules_tbl.
48 */
49static DEFINE_SPINLOCK(pnfs_spinlock);
50
51/*
52 * pnfs_modules_tbl holds all pnfs modules
53 */
54static LIST_HEAD(pnfs_modules_tbl);
55
56static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
57static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
58 struct list_head *free_me,
59 const struct pnfs_layout_range *range,
60 u32 seq);
61static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
62 struct list_head *tmp_list);
63
64/* Return the registered pnfs layout driver module matching given id */
65static struct pnfs_layoutdriver_type *
66find_pnfs_driver_locked(u32 id)
67{
68 struct pnfs_layoutdriver_type *local;
69
70 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
71 if (local->id == id)
72 goto out;
73 local = NULL;
74out:
75 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
76 return local;
77}
78
79static struct pnfs_layoutdriver_type *
80find_pnfs_driver(u32 id)
81{
82 struct pnfs_layoutdriver_type *local;
83
84 spin_lock(&pnfs_spinlock);
85 local = find_pnfs_driver_locked(id);
86 if (local != NULL && !try_module_get(local->owner)) {
87 dprintk("%s: Could not grab reference on module\n", __func__);
88 local = NULL;
89 }
90 spin_unlock(&pnfs_spinlock);
91 return local;
92}
93
94void
95unset_pnfs_layoutdriver(struct nfs_server *nfss)
96{
97 if (nfss->pnfs_curr_ld) {
98 if (nfss->pnfs_curr_ld->clear_layoutdriver)
99 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
100 /* Decrement the MDS count. Purge the deviceid cache if zero */
101 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
102 nfs4_deviceid_purge_client(nfss->nfs_client);
103 module_put(nfss->pnfs_curr_ld->owner);
104 }
105 nfss->pnfs_curr_ld = NULL;
106}
107
108/*
109 * When the server sends a list of layout types, we choose one in the order
110 * given in the list below.
111 *
112 * FIXME: should this list be configurable in some fashion? module param?
113 * mount option? something else?
114 */
115static const u32 ld_prefs[] = {
116 LAYOUT_SCSI,
117 LAYOUT_BLOCK_VOLUME,
118 LAYOUT_OSD2_OBJECTS,
119 LAYOUT_FLEX_FILES,
120 LAYOUT_NFSV4_1_FILES,
121 0
122};
123
124static int
125ld_cmp(const void *e1, const void *e2)
126{
127 u32 ld1 = *((u32 *)e1);
128 u32 ld2 = *((u32 *)e2);
129 int i;
130
131 for (i = 0; ld_prefs[i] != 0; i++) {
132 if (ld1 == ld_prefs[i])
133 return -1;
134
135 if (ld2 == ld_prefs[i])
136 return 1;
137 }
138 return 0;
139}
140
141/*
142 * Try to set the server's pnfs module to the pnfs layout type specified by id.
143 * Currently only one pNFS layout driver per filesystem is supported.
144 *
145 * @ids array of layout types supported by MDS.
146 */
147void
148set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
149 struct nfs_fsinfo *fsinfo)
150{
151 struct pnfs_layoutdriver_type *ld_type = NULL;
152 u32 id;
153 int i;
154
155 if (fsinfo->nlayouttypes == 0)
156 goto out_no_driver;
157 if (!(server->nfs_client->cl_exchange_flags &
158 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
159 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
160 __func__, server->nfs_client->cl_exchange_flags);
161 goto out_no_driver;
162 }
163
164 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
165 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
166
167 for (i = 0; i < fsinfo->nlayouttypes; i++) {
168 id = fsinfo->layouttype[i];
169 ld_type = find_pnfs_driver(id);
170 if (!ld_type) {
171 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
172 id);
173 ld_type = find_pnfs_driver(id);
174 }
175 if (ld_type)
176 break;
177 }
178
179 if (!ld_type) {
180 dprintk("%s: No pNFS module found!\n", __func__);
181 goto out_no_driver;
182 }
183
184 server->pnfs_curr_ld = ld_type;
185 if (ld_type->set_layoutdriver
186 && ld_type->set_layoutdriver(server, mntfh)) {
187 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
188 "driver %u.\n", __func__, id);
189 module_put(ld_type->owner);
190 goto out_no_driver;
191 }
192 /* Bump the MDS count */
193 atomic_inc(&server->nfs_client->cl_mds_count);
194
195 dprintk("%s: pNFS module for %u set\n", __func__, id);
196 return;
197
198out_no_driver:
199 dprintk("%s: Using NFSv4 I/O\n", __func__);
200 server->pnfs_curr_ld = NULL;
201}
202
203int
204pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
205{
206 int status = -EINVAL;
207 struct pnfs_layoutdriver_type *tmp;
208
209 if (ld_type->id == 0) {
210 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
211 return status;
212 }
213 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
214 printk(KERN_ERR "NFS: %s Layout driver must provide "
215 "alloc_lseg and free_lseg.\n", __func__);
216 return status;
217 }
218
219 spin_lock(&pnfs_spinlock);
220 tmp = find_pnfs_driver_locked(ld_type->id);
221 if (!tmp) {
222 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
223 status = 0;
224 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
225 ld_type->name);
226 } else {
227 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
228 __func__, ld_type->id);
229 }
230 spin_unlock(&pnfs_spinlock);
231
232 return status;
233}
234EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
235
236void
237pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
238{
239 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
240 spin_lock(&pnfs_spinlock);
241 list_del(&ld_type->pnfs_tblid);
242 spin_unlock(&pnfs_spinlock);
243}
244EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
245
246/*
247 * pNFS client layout cache
248 */
249
250/* Need to hold i_lock if caller does not already hold reference */
251void
252pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
253{
254 refcount_inc(&lo->plh_refcount);
255}
256
257static struct pnfs_layout_hdr *
258pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
259{
260 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
261 return ld->alloc_layout_hdr(ino, gfp_flags);
262}
263
264static void
265pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
266{
267 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
268 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
269
270 if (!list_empty(&lo->plh_layouts)) {
271 struct nfs_client *clp = server->nfs_client;
272
273 spin_lock(&clp->cl_lock);
274 list_del_init(&lo->plh_layouts);
275 spin_unlock(&clp->cl_lock);
276 }
277 put_rpccred(lo->plh_lc_cred);
278 return ld->free_layout_hdr(lo);
279}
280
281static void
282pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
283{
284 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
285 dprintk("%s: freeing layout cache %p\n", __func__, lo);
286 nfsi->layout = NULL;
287 /* Reset MDS Threshold I/O counters */
288 nfsi->write_io = 0;
289 nfsi->read_io = 0;
290}
291
292void
293pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
294{
295 struct inode *inode;
296
297 if (!lo)
298 return;
299 inode = lo->plh_inode;
300 pnfs_layoutreturn_before_put_layout_hdr(lo);
301
302 if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
303 if (!list_empty(&lo->plh_segs))
304 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
305 pnfs_detach_layout_hdr(lo);
306 spin_unlock(&inode->i_lock);
307 pnfs_free_layout_hdr(lo);
308 }
309}
310
311static void
312pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
313 u32 seq)
314{
315 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
316 iomode = IOMODE_ANY;
317 lo->plh_return_iomode = iomode;
318 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
319 if (seq != 0) {
320 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
321 lo->plh_return_seq = seq;
322 }
323}
324
325static void
326pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
327{
328 struct pnfs_layout_segment *lseg;
329 lo->plh_return_iomode = 0;
330 lo->plh_return_seq = 0;
331 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
332 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
333 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
334 continue;
335 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
336 }
337}
338
339static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
340{
341 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
342 clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
343 smp_mb__after_atomic();
344 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
345 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
346}
347
348static void
349pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
350 struct list_head *free_me)
351{
352 clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
353 clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
354 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
355 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
356 if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
357 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
358}
359
360/*
361 * Update the seqid of a layout stateid
362 */
363bool nfs4_refresh_layout_stateid(nfs4_stateid *dst, struct inode *inode)
364{
365 struct pnfs_layout_hdr *lo;
366 bool ret = false;
367
368 spin_lock(&inode->i_lock);
369 lo = NFS_I(inode)->layout;
370 if (lo && nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
371 dst->seqid = lo->plh_stateid.seqid;
372 ret = true;
373 }
374 spin_unlock(&inode->i_lock);
375 return ret;
376}
377
378/*
379 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
380 *
381 * In order to continue using the pnfs_layout_hdr, a full recovery
382 * is required.
383 * Note that caller must hold inode->i_lock.
384 */
385int
386pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
387 struct list_head *lseg_list)
388{
389 struct pnfs_layout_range range = {
390 .iomode = IOMODE_ANY,
391 .offset = 0,
392 .length = NFS4_MAX_UINT64,
393 };
394 struct pnfs_layout_segment *lseg, *next;
395
396 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
397 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
398 pnfs_clear_lseg_state(lseg, lseg_list);
399 pnfs_clear_layoutreturn_info(lo);
400 pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
401 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
402 !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
403 pnfs_clear_layoutreturn_waitbit(lo);
404 return !list_empty(&lo->plh_segs);
405}
406
407static int
408pnfs_iomode_to_fail_bit(u32 iomode)
409{
410 return iomode == IOMODE_RW ?
411 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
412}
413
414static void
415pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
416{
417 lo->plh_retry_timestamp = jiffies;
418 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
419 refcount_inc(&lo->plh_refcount);
420}
421
422static void
423pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
424{
425 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
426 refcount_dec(&lo->plh_refcount);
427}
428
429static void
430pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
431{
432 struct inode *inode = lo->plh_inode;
433 struct pnfs_layout_range range = {
434 .iomode = iomode,
435 .offset = 0,
436 .length = NFS4_MAX_UINT64,
437 };
438 LIST_HEAD(head);
439
440 spin_lock(&inode->i_lock);
441 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
442 pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
443 spin_unlock(&inode->i_lock);
444 pnfs_free_lseg_list(&head);
445 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
446 iomode == IOMODE_RW ? "RW" : "READ");
447}
448
449static bool
450pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
451{
452 unsigned long start, end;
453 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
454
455 if (test_bit(fail_bit, &lo->plh_flags) == 0)
456 return false;
457 end = jiffies;
458 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
459 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
460 /* It is time to retry the failed layoutgets */
461 pnfs_layout_clear_fail_bit(lo, fail_bit);
462 return false;
463 }
464 return true;
465}
466
467static void
468pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
469 const struct pnfs_layout_range *range,
470 const nfs4_stateid *stateid)
471{
472 INIT_LIST_HEAD(&lseg->pls_list);
473 INIT_LIST_HEAD(&lseg->pls_lc_list);
474 refcount_set(&lseg->pls_refcount, 1);
475 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
476 lseg->pls_layout = lo;
477 lseg->pls_range = *range;
478 lseg->pls_seq = be32_to_cpu(stateid->seqid);
479}
480
481static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
482{
483 if (lseg != NULL) {
484 struct inode *inode = lseg->pls_layout->plh_inode;
485 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
486 }
487}
488
489static void
490pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
491 struct pnfs_layout_segment *lseg)
492{
493 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
494 list_del_init(&lseg->pls_list);
495 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
496 refcount_dec(&lo->plh_refcount);
497 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
498 return;
499 if (list_empty(&lo->plh_segs) &&
500 !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
501 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
502 if (atomic_read(&lo->plh_outstanding) == 0)
503 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
504 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
505 }
506}
507
508static bool
509pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
510 struct pnfs_layout_segment *lseg)
511{
512 if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
513 pnfs_layout_is_valid(lo)) {
514 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
515 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
516 return true;
517 }
518 return false;
519}
520
521void
522pnfs_put_lseg(struct pnfs_layout_segment *lseg)
523{
524 struct pnfs_layout_hdr *lo;
525 struct inode *inode;
526
527 if (!lseg)
528 return;
529
530 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
531 refcount_read(&lseg->pls_refcount),
532 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
533
534 lo = lseg->pls_layout;
535 inode = lo->plh_inode;
536
537 if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
538 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
539 spin_unlock(&inode->i_lock);
540 return;
541 }
542 pnfs_get_layout_hdr(lo);
543 pnfs_layout_remove_lseg(lo, lseg);
544 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
545 lseg = NULL;
546 spin_unlock(&inode->i_lock);
547 pnfs_free_lseg(lseg);
548 pnfs_put_layout_hdr(lo);
549 }
550}
551EXPORT_SYMBOL_GPL(pnfs_put_lseg);
552
553/*
554 * is l2 fully contained in l1?
555 * start1 end1
556 * [----------------------------------)
557 * start2 end2
558 * [----------------)
559 */
560static bool
561pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
562 const struct pnfs_layout_range *l2)
563{
564 u64 start1 = l1->offset;
565 u64 end1 = pnfs_end_offset(start1, l1->length);
566 u64 start2 = l2->offset;
567 u64 end2 = pnfs_end_offset(start2, l2->length);
568
569 return (start1 <= start2) && (end1 >= end2);
570}
571
572static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
573 struct list_head *tmp_list)
574{
575 if (!refcount_dec_and_test(&lseg->pls_refcount))
576 return false;
577 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
578 list_add(&lseg->pls_list, tmp_list);
579 return true;
580}
581
582/* Returns 1 if lseg is removed from list, 0 otherwise */
583static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
584 struct list_head *tmp_list)
585{
586 int rv = 0;
587
588 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
589 /* Remove the reference keeping the lseg in the
590 * list. It will now be removed when all
591 * outstanding io is finished.
592 */
593 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
594 refcount_read(&lseg->pls_refcount));
595 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
596 rv = 1;
597 }
598 return rv;
599}
600
601/*
602 * Compare 2 layout stateid sequence ids, to see which is newer,
603 * taking into account wraparound issues.
604 */
605static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
606{
607 return (s32)(s1 - s2) > 0;
608}
609
610static bool
611pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
612 const struct pnfs_layout_range *recall_range)
613{
614 return (recall_range->iomode == IOMODE_ANY ||
615 lseg_range->iomode == recall_range->iomode) &&
616 pnfs_lseg_range_intersecting(lseg_range, recall_range);
617}
618
619static bool
620pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
621 const struct pnfs_layout_range *recall_range,
622 u32 seq)
623{
624 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
625 return false;
626 if (recall_range == NULL)
627 return true;
628 return pnfs_should_free_range(&lseg->pls_range, recall_range);
629}
630
631/**
632 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
633 * @lo: layout header containing the lsegs
634 * @tmp_list: list head where doomed lsegs should go
635 * @recall_range: optional recall range argument to match (may be NULL)
636 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
637 *
638 * Walk the list of lsegs in the layout header, and tear down any that should
639 * be destroyed. If "recall_range" is specified then the segment must match
640 * that range. If "seq" is non-zero, then only match segments that were handed
641 * out at or before that sequence.
642 *
643 * Returns number of matching invalid lsegs remaining in list after scanning
644 * it and purging them.
645 */
646int
647pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
648 struct list_head *tmp_list,
649 const struct pnfs_layout_range *recall_range,
650 u32 seq)
651{
652 struct pnfs_layout_segment *lseg, *next;
653 int remaining = 0;
654
655 dprintk("%s:Begin lo %p\n", __func__, lo);
656
657 if (list_empty(&lo->plh_segs))
658 return 0;
659 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
660 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
661 dprintk("%s: freeing lseg %p iomode %d seq %u "
662 "offset %llu length %llu\n", __func__,
663 lseg, lseg->pls_range.iomode, lseg->pls_seq,
664 lseg->pls_range.offset, lseg->pls_range.length);
665 if (!mark_lseg_invalid(lseg, tmp_list))
666 remaining++;
667 }
668 dprintk("%s:Return %i\n", __func__, remaining);
669 return remaining;
670}
671
672static void
673pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
674 struct list_head *free_me,
675 const struct pnfs_layout_range *range,
676 u32 seq)
677{
678 struct pnfs_layout_segment *lseg, *next;
679
680 list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
681 if (pnfs_match_lseg_recall(lseg, range, seq))
682 list_move_tail(&lseg->pls_list, free_me);
683 }
684}
685
686/* note free_me must contain lsegs from a single layout_hdr */
687void
688pnfs_free_lseg_list(struct list_head *free_me)
689{
690 struct pnfs_layout_segment *lseg, *tmp;
691
692 if (list_empty(free_me))
693 return;
694
695 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
696 list_del(&lseg->pls_list);
697 pnfs_free_lseg(lseg);
698 }
699}
700
701void
702pnfs_destroy_layout(struct nfs_inode *nfsi)
703{
704 struct pnfs_layout_hdr *lo;
705 LIST_HEAD(tmp_list);
706
707 spin_lock(&nfsi->vfs_inode.i_lock);
708 lo = nfsi->layout;
709 if (lo) {
710 pnfs_get_layout_hdr(lo);
711 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
712 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
713 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
714 spin_unlock(&nfsi->vfs_inode.i_lock);
715 pnfs_free_lseg_list(&tmp_list);
716 nfs_commit_inode(&nfsi->vfs_inode, 0);
717 pnfs_put_layout_hdr(lo);
718 } else
719 spin_unlock(&nfsi->vfs_inode.i_lock);
720}
721EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
722
723static bool
724pnfs_layout_add_bulk_destroy_list(struct inode *inode,
725 struct list_head *layout_list)
726{
727 struct pnfs_layout_hdr *lo;
728 bool ret = false;
729
730 spin_lock(&inode->i_lock);
731 lo = NFS_I(inode)->layout;
732 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
733 pnfs_get_layout_hdr(lo);
734 list_add(&lo->plh_bulk_destroy, layout_list);
735 ret = true;
736 }
737 spin_unlock(&inode->i_lock);
738 return ret;
739}
740
741/* Caller must hold rcu_read_lock and clp->cl_lock */
742static int
743pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
744 struct nfs_server *server,
745 struct list_head *layout_list)
746{
747 struct pnfs_layout_hdr *lo, *next;
748 struct inode *inode;
749
750 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
751 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
752 continue;
753 inode = igrab(lo->plh_inode);
754 if (inode == NULL)
755 continue;
756 list_del_init(&lo->plh_layouts);
757 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
758 continue;
759 rcu_read_unlock();
760 spin_unlock(&clp->cl_lock);
761 iput(inode);
762 spin_lock(&clp->cl_lock);
763 rcu_read_lock();
764 return -EAGAIN;
765 }
766 return 0;
767}
768
769static int
770pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
771 bool is_bulk_recall)
772{
773 struct pnfs_layout_hdr *lo;
774 struct inode *inode;
775 LIST_HEAD(lseg_list);
776 int ret = 0;
777
778 while (!list_empty(layout_list)) {
779 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
780 plh_bulk_destroy);
781 dprintk("%s freeing layout for inode %lu\n", __func__,
782 lo->plh_inode->i_ino);
783 inode = lo->plh_inode;
784
785 pnfs_layoutcommit_inode(inode, false);
786
787 spin_lock(&inode->i_lock);
788 list_del_init(&lo->plh_bulk_destroy);
789 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
790 if (is_bulk_recall)
791 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
792 ret = -EAGAIN;
793 }
794 spin_unlock(&inode->i_lock);
795 pnfs_free_lseg_list(&lseg_list);
796 /* Free all lsegs that are attached to commit buckets */
797 nfs_commit_inode(inode, 0);
798 pnfs_put_layout_hdr(lo);
799 iput(inode);
800 }
801 return ret;
802}
803
804int
805pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
806 struct nfs_fsid *fsid,
807 bool is_recall)
808{
809 struct nfs_server *server;
810 LIST_HEAD(layout_list);
811
812 spin_lock(&clp->cl_lock);
813 rcu_read_lock();
814restart:
815 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
816 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
817 continue;
818 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
819 server,
820 &layout_list) != 0)
821 goto restart;
822 }
823 rcu_read_unlock();
824 spin_unlock(&clp->cl_lock);
825
826 if (list_empty(&layout_list))
827 return 0;
828 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
829}
830
831int
832pnfs_destroy_layouts_byclid(struct nfs_client *clp,
833 bool is_recall)
834{
835 struct nfs_server *server;
836 LIST_HEAD(layout_list);
837
838 spin_lock(&clp->cl_lock);
839 rcu_read_lock();
840restart:
841 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
842 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
843 server,
844 &layout_list) != 0)
845 goto restart;
846 }
847 rcu_read_unlock();
848 spin_unlock(&clp->cl_lock);
849
850 if (list_empty(&layout_list))
851 return 0;
852 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
853}
854
855/*
856 * Called by the state manger to remove all layouts established under an
857 * expired lease.
858 */
859void
860pnfs_destroy_all_layouts(struct nfs_client *clp)
861{
862 nfs4_deviceid_mark_client_invalid(clp);
863 nfs4_deviceid_purge_client(clp);
864
865 pnfs_destroy_layouts_byclid(clp, false);
866}
867
868/* update lo->plh_stateid with new if is more recent */
869void
870pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
871 bool update_barrier)
872{
873 u32 oldseq, newseq, new_barrier = 0;
874
875 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
876 newseq = be32_to_cpu(new->seqid);
877
878 if (!pnfs_layout_is_valid(lo)) {
879 nfs4_stateid_copy(&lo->plh_stateid, new);
880 lo->plh_barrier = newseq;
881 pnfs_clear_layoutreturn_info(lo);
882 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
883 return;
884 }
885 if (pnfs_seqid_is_newer(newseq, oldseq)) {
886 nfs4_stateid_copy(&lo->plh_stateid, new);
887 /*
888 * Because of wraparound, we want to keep the barrier
889 * "close" to the current seqids.
890 */
891 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
892 }
893 if (update_barrier)
894 new_barrier = be32_to_cpu(new->seqid);
895 else if (new_barrier == 0)
896 return;
897 if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
898 lo->plh_barrier = new_barrier;
899}
900
901static bool
902pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
903 const nfs4_stateid *stateid)
904{
905 u32 seqid = be32_to_cpu(stateid->seqid);
906
907 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
908}
909
910/* lget is set to 1 if called from inside send_layoutget call chain */
911static bool
912pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
913{
914 return lo->plh_block_lgets ||
915 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
916}
917
918/*
919 * Get layout from server.
920 * for now, assume that whole file layouts are requested.
921 * arg->offset: 0
922 * arg->length: all ones
923 */
924static struct pnfs_layout_segment *
925send_layoutget(struct pnfs_layout_hdr *lo,
926 struct nfs_open_context *ctx,
927 nfs4_stateid *stateid,
928 const struct pnfs_layout_range *range,
929 long *timeout, gfp_t gfp_flags)
930{
931 struct inode *ino = lo->plh_inode;
932 struct nfs_server *server = NFS_SERVER(ino);
933 struct nfs4_layoutget *lgp;
934 loff_t i_size;
935
936 dprintk("--> %s\n", __func__);
937
938 /*
939 * Synchronously retrieve layout information from server and
940 * store in lseg. If we race with a concurrent seqid morphing
941 * op, then re-send the LAYOUTGET.
942 */
943 lgp = kzalloc(sizeof(*lgp), gfp_flags);
944 if (lgp == NULL)
945 return ERR_PTR(-ENOMEM);
946
947 i_size = i_size_read(ino);
948
949 lgp->args.minlength = PAGE_SIZE;
950 if (lgp->args.minlength > range->length)
951 lgp->args.minlength = range->length;
952 if (range->iomode == IOMODE_READ) {
953 if (range->offset >= i_size)
954 lgp->args.minlength = 0;
955 else if (i_size - range->offset < lgp->args.minlength)
956 lgp->args.minlength = i_size - range->offset;
957 }
958 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
959 pnfs_copy_range(&lgp->args.range, range);
960 lgp->args.type = server->pnfs_curr_ld->id;
961 lgp->args.inode = ino;
962 lgp->args.ctx = get_nfs_open_context(ctx);
963 nfs4_stateid_copy(&lgp->args.stateid, stateid);
964 lgp->gfp_flags = gfp_flags;
965 lgp->cred = lo->plh_lc_cred;
966
967 return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
968}
969
970static void pnfs_clear_layoutcommit(struct inode *inode,
971 struct list_head *head)
972{
973 struct nfs_inode *nfsi = NFS_I(inode);
974 struct pnfs_layout_segment *lseg, *tmp;
975
976 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
977 return;
978 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
979 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
980 continue;
981 pnfs_lseg_dec_and_remove_zero(lseg, head);
982 }
983}
984
985void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
986 const nfs4_stateid *arg_stateid,
987 const struct pnfs_layout_range *range,
988 const nfs4_stateid *stateid)
989{
990 struct inode *inode = lo->plh_inode;
991 LIST_HEAD(freeme);
992
993 spin_lock(&inode->i_lock);
994 if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
995 !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
996 goto out_unlock;
997 if (stateid) {
998 u32 seq = be32_to_cpu(arg_stateid->seqid);
999
1000 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1001 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1002 pnfs_set_layout_stateid(lo, stateid, true);
1003 } else
1004 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1005out_unlock:
1006 pnfs_clear_layoutreturn_waitbit(lo);
1007 spin_unlock(&inode->i_lock);
1008 pnfs_free_lseg_list(&freeme);
1009
1010}
1011
1012static bool
1013pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1014 nfs4_stateid *stateid,
1015 enum pnfs_iomode *iomode)
1016{
1017 /* Serialise LAYOUTGET/LAYOUTRETURN */
1018 if (atomic_read(&lo->plh_outstanding) != 0)
1019 return false;
1020 if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1021 return false;
1022 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1023 pnfs_get_layout_hdr(lo);
1024 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1025 if (stateid != NULL) {
1026 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1027 if (lo->plh_return_seq != 0)
1028 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1029 }
1030 if (iomode != NULL)
1031 *iomode = lo->plh_return_iomode;
1032 pnfs_clear_layoutreturn_info(lo);
1033 return true;
1034 }
1035 if (stateid != NULL)
1036 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1037 if (iomode != NULL)
1038 *iomode = IOMODE_ANY;
1039 return true;
1040}
1041
1042static void
1043pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1044 struct pnfs_layout_hdr *lo,
1045 const nfs4_stateid *stateid,
1046 enum pnfs_iomode iomode)
1047{
1048 struct inode *inode = lo->plh_inode;
1049
1050 args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1051 args->inode = inode;
1052 args->range.iomode = iomode;
1053 args->range.offset = 0;
1054 args->range.length = NFS4_MAX_UINT64;
1055 args->layout = lo;
1056 nfs4_stateid_copy(&args->stateid, stateid);
1057}
1058
1059static int
1060pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1061 enum pnfs_iomode iomode, bool sync)
1062{
1063 struct inode *ino = lo->plh_inode;
1064 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1065 struct nfs4_layoutreturn *lrp;
1066 int status = 0;
1067
1068 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1069 if (unlikely(lrp == NULL)) {
1070 status = -ENOMEM;
1071 spin_lock(&ino->i_lock);
1072 pnfs_clear_layoutreturn_waitbit(lo);
1073 spin_unlock(&ino->i_lock);
1074 pnfs_put_layout_hdr(lo);
1075 goto out;
1076 }
1077
1078 pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1079 lrp->args.ld_private = &lrp->ld_private;
1080 lrp->clp = NFS_SERVER(ino)->nfs_client;
1081 lrp->cred = lo->plh_lc_cred;
1082 if (ld->prepare_layoutreturn)
1083 ld->prepare_layoutreturn(&lrp->args);
1084
1085 status = nfs4_proc_layoutreturn(lrp, sync);
1086out:
1087 dprintk("<-- %s status: %d\n", __func__, status);
1088 return status;
1089}
1090
1091/* Return true if layoutreturn is needed */
1092static bool
1093pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1094{
1095 struct pnfs_layout_segment *s;
1096
1097 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1098 return false;
1099
1100 /* Defer layoutreturn until all lsegs are done */
1101 list_for_each_entry(s, &lo->plh_segs, pls_list) {
1102 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1103 return false;
1104 }
1105
1106 return true;
1107}
1108
1109static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1110{
1111 struct inode *inode= lo->plh_inode;
1112
1113 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1114 return;
1115 spin_lock(&inode->i_lock);
1116 if (pnfs_layout_need_return(lo)) {
1117 nfs4_stateid stateid;
1118 enum pnfs_iomode iomode;
1119 bool send;
1120
1121 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1122 spin_unlock(&inode->i_lock);
1123 if (send) {
1124 /* Send an async layoutreturn so we dont deadlock */
1125 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1126 }
1127 } else
1128 spin_unlock(&inode->i_lock);
1129}
1130
1131/*
1132 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1133 * when the layout segment list is empty.
1134 *
1135 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1136 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1137 * deviceid is marked invalid.
1138 */
1139int
1140_pnfs_return_layout(struct inode *ino)
1141{
1142 struct pnfs_layout_hdr *lo = NULL;
1143 struct nfs_inode *nfsi = NFS_I(ino);
1144 LIST_HEAD(tmp_list);
1145 nfs4_stateid stateid;
1146 int status = 0;
1147 bool send;
1148
1149 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1150
1151 spin_lock(&ino->i_lock);
1152 lo = nfsi->layout;
1153 if (!lo) {
1154 spin_unlock(&ino->i_lock);
1155 dprintk("NFS: %s no layout to return\n", __func__);
1156 goto out;
1157 }
1158 /* Reference matched in nfs4_layoutreturn_release */
1159 pnfs_get_layout_hdr(lo);
1160 /* Is there an outstanding layoutreturn ? */
1161 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1162 spin_unlock(&ino->i_lock);
1163 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1164 TASK_UNINTERRUPTIBLE))
1165 goto out_put_layout_hdr;
1166 spin_lock(&ino->i_lock);
1167 }
1168 pnfs_clear_layoutcommit(ino, &tmp_list);
1169 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1170
1171 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1172 struct pnfs_layout_range range = {
1173 .iomode = IOMODE_ANY,
1174 .offset = 0,
1175 .length = NFS4_MAX_UINT64,
1176 };
1177 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1178 }
1179
1180 /* Don't send a LAYOUTRETURN if list was initially empty */
1181 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1182 spin_unlock(&ino->i_lock);
1183 dprintk("NFS: %s no layout segments to return\n", __func__);
1184 goto out_put_layout_hdr;
1185 }
1186
1187 send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1188 spin_unlock(&ino->i_lock);
1189 if (send)
1190 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1191out_put_layout_hdr:
1192 pnfs_free_lseg_list(&tmp_list);
1193 pnfs_put_layout_hdr(lo);
1194out:
1195 dprintk("<-- %s status: %d\n", __func__, status);
1196 return status;
1197}
1198
1199int
1200pnfs_commit_and_return_layout(struct inode *inode)
1201{
1202 struct pnfs_layout_hdr *lo;
1203 int ret;
1204
1205 spin_lock(&inode->i_lock);
1206 lo = NFS_I(inode)->layout;
1207 if (lo == NULL) {
1208 spin_unlock(&inode->i_lock);
1209 return 0;
1210 }
1211 pnfs_get_layout_hdr(lo);
1212 /* Block new layoutgets and read/write to ds */
1213 lo->plh_block_lgets++;
1214 spin_unlock(&inode->i_lock);
1215 filemap_fdatawait(inode->i_mapping);
1216 ret = pnfs_layoutcommit_inode(inode, true);
1217 if (ret == 0)
1218 ret = _pnfs_return_layout(inode);
1219 spin_lock(&inode->i_lock);
1220 lo->plh_block_lgets--;
1221 spin_unlock(&inode->i_lock);
1222 pnfs_put_layout_hdr(lo);
1223 return ret;
1224}
1225
1226bool pnfs_roc(struct inode *ino,
1227 struct nfs4_layoutreturn_args *args,
1228 struct nfs4_layoutreturn_res *res,
1229 const struct rpc_cred *cred)
1230{
1231 struct nfs_inode *nfsi = NFS_I(ino);
1232 struct nfs_open_context *ctx;
1233 struct nfs4_state *state;
1234 struct pnfs_layout_hdr *lo;
1235 struct pnfs_layout_segment *lseg, *next;
1236 nfs4_stateid stateid;
1237 enum pnfs_iomode iomode = 0;
1238 bool layoutreturn = false, roc = false;
1239 bool skip_read = false;
1240
1241 if (!nfs_have_layout(ino))
1242 return false;
1243retry:
1244 spin_lock(&ino->i_lock);
1245 lo = nfsi->layout;
1246 if (!lo || !pnfs_layout_is_valid(lo) ||
1247 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1248 lo = NULL;
1249 goto out_noroc;
1250 }
1251 pnfs_get_layout_hdr(lo);
1252 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1253 spin_unlock(&ino->i_lock);
1254 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1255 TASK_UNINTERRUPTIBLE);
1256 pnfs_put_layout_hdr(lo);
1257 goto retry;
1258 }
1259
1260 /* no roc if we hold a delegation */
1261 if (nfs4_check_delegation(ino, FMODE_READ)) {
1262 if (nfs4_check_delegation(ino, FMODE_WRITE))
1263 goto out_noroc;
1264 skip_read = true;
1265 }
1266
1267 list_for_each_entry(ctx, &nfsi->open_files, list) {
1268 state = ctx->state;
1269 if (state == NULL)
1270 continue;
1271 /* Don't return layout if there is open file state */
1272 if (state->state & FMODE_WRITE)
1273 goto out_noroc;
1274 if (state->state & FMODE_READ)
1275 skip_read = true;
1276 }
1277
1278
1279 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1280 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1281 continue;
1282 /* If we are sending layoutreturn, invalidate all valid lsegs */
1283 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1284 continue;
1285 /*
1286 * Note: mark lseg for return so pnfs_layout_remove_lseg
1287 * doesn't invalidate the layout for us.
1288 */
1289 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1290 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1291 continue;
1292 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1293 }
1294
1295 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1296 goto out_noroc;
1297
1298 /* ROC in two conditions:
1299 * 1. there are ROC lsegs
1300 * 2. we don't send layoutreturn
1301 */
1302 /* lo ref dropped in pnfs_roc_release() */
1303 layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1304 /* If the creds don't match, we can't compound the layoutreturn */
1305 if (!layoutreturn || cred != lo->plh_lc_cred)
1306 goto out_noroc;
1307
1308 roc = layoutreturn;
1309 pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1310 res->lrs_present = 0;
1311 layoutreturn = false;
1312
1313out_noroc:
1314 spin_unlock(&ino->i_lock);
1315 pnfs_layoutcommit_inode(ino, true);
1316 if (roc) {
1317 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1318 if (ld->prepare_layoutreturn)
1319 ld->prepare_layoutreturn(args);
1320 pnfs_put_layout_hdr(lo);
1321 return true;
1322 }
1323 if (layoutreturn)
1324 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1325 pnfs_put_layout_hdr(lo);
1326 return false;
1327}
1328
1329void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1330 struct nfs4_layoutreturn_res *res,
1331 int ret)
1332{
1333 struct pnfs_layout_hdr *lo = args->layout;
1334 const nfs4_stateid *arg_stateid = NULL;
1335 const nfs4_stateid *res_stateid = NULL;
1336 struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1337
1338 if (ret == 0) {
1339 arg_stateid = &args->stateid;
1340 if (res->lrs_present)
1341 res_stateid = &res->stateid;
1342 }
1343 pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1344 res_stateid);
1345 if (ld_private && ld_private->ops && ld_private->ops->free)
1346 ld_private->ops->free(ld_private);
1347 pnfs_put_layout_hdr(lo);
1348 trace_nfs4_layoutreturn_on_close(args->inode, 0);
1349}
1350
1351bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1352{
1353 struct nfs_inode *nfsi = NFS_I(ino);
1354 struct pnfs_layout_hdr *lo;
1355 bool sleep = false;
1356
1357 /* we might not have grabbed lo reference. so need to check under
1358 * i_lock */
1359 spin_lock(&ino->i_lock);
1360 lo = nfsi->layout;
1361 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1362 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1363 sleep = true;
1364 }
1365 spin_unlock(&ino->i_lock);
1366 return sleep;
1367}
1368
1369/*
1370 * Compare two layout segments for sorting into layout cache.
1371 * We want to preferentially return RW over RO layouts, so ensure those
1372 * are seen first.
1373 */
1374static s64
1375pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1376 const struct pnfs_layout_range *l2)
1377{
1378 s64 d;
1379
1380 /* high offset > low offset */
1381 d = l1->offset - l2->offset;
1382 if (d)
1383 return d;
1384
1385 /* short length > long length */
1386 d = l2->length - l1->length;
1387 if (d)
1388 return d;
1389
1390 /* read > read/write */
1391 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1392}
1393
1394static bool
1395pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1396 const struct pnfs_layout_range *l2)
1397{
1398 return pnfs_lseg_range_cmp(l1, l2) > 0;
1399}
1400
1401static bool
1402pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1403 struct pnfs_layout_segment *old)
1404{
1405 return false;
1406}
1407
1408void
1409pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1410 struct pnfs_layout_segment *lseg,
1411 bool (*is_after)(const struct pnfs_layout_range *,
1412 const struct pnfs_layout_range *),
1413 bool (*do_merge)(struct pnfs_layout_segment *,
1414 struct pnfs_layout_segment *),
1415 struct list_head *free_me)
1416{
1417 struct pnfs_layout_segment *lp, *tmp;
1418
1419 dprintk("%s:Begin\n", __func__);
1420
1421 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1422 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1423 continue;
1424 if (do_merge(lseg, lp)) {
1425 mark_lseg_invalid(lp, free_me);
1426 continue;
1427 }
1428 if (is_after(&lseg->pls_range, &lp->pls_range))
1429 continue;
1430 list_add_tail(&lseg->pls_list, &lp->pls_list);
1431 dprintk("%s: inserted lseg %p "
1432 "iomode %d offset %llu length %llu before "
1433 "lp %p iomode %d offset %llu length %llu\n",
1434 __func__, lseg, lseg->pls_range.iomode,
1435 lseg->pls_range.offset, lseg->pls_range.length,
1436 lp, lp->pls_range.iomode, lp->pls_range.offset,
1437 lp->pls_range.length);
1438 goto out;
1439 }
1440 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1441 dprintk("%s: inserted lseg %p "
1442 "iomode %d offset %llu length %llu at tail\n",
1443 __func__, lseg, lseg->pls_range.iomode,
1444 lseg->pls_range.offset, lseg->pls_range.length);
1445out:
1446 pnfs_get_layout_hdr(lo);
1447
1448 dprintk("%s:Return\n", __func__);
1449}
1450EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1451
1452static void
1453pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1454 struct pnfs_layout_segment *lseg,
1455 struct list_head *free_me)
1456{
1457 struct inode *inode = lo->plh_inode;
1458 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1459
1460 if (ld->add_lseg != NULL)
1461 ld->add_lseg(lo, lseg, free_me);
1462 else
1463 pnfs_generic_layout_insert_lseg(lo, lseg,
1464 pnfs_lseg_range_is_after,
1465 pnfs_lseg_no_merge,
1466 free_me);
1467}
1468
1469static struct pnfs_layout_hdr *
1470alloc_init_layout_hdr(struct inode *ino,
1471 struct nfs_open_context *ctx,
1472 gfp_t gfp_flags)
1473{
1474 struct pnfs_layout_hdr *lo;
1475
1476 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1477 if (!lo)
1478 return NULL;
1479 refcount_set(&lo->plh_refcount, 1);
1480 INIT_LIST_HEAD(&lo->plh_layouts);
1481 INIT_LIST_HEAD(&lo->plh_segs);
1482 INIT_LIST_HEAD(&lo->plh_return_segs);
1483 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1484 lo->plh_inode = ino;
1485 lo->plh_lc_cred = get_rpccred(ctx->cred);
1486 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1487 return lo;
1488}
1489
1490static struct pnfs_layout_hdr *
1491pnfs_find_alloc_layout(struct inode *ino,
1492 struct nfs_open_context *ctx,
1493 gfp_t gfp_flags)
1494 __releases(&ino->i_lock)
1495 __acquires(&ino->i_lock)
1496{
1497 struct nfs_inode *nfsi = NFS_I(ino);
1498 struct pnfs_layout_hdr *new = NULL;
1499
1500 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1501
1502 if (nfsi->layout != NULL)
1503 goto out_existing;
1504 spin_unlock(&ino->i_lock);
1505 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1506 spin_lock(&ino->i_lock);
1507
1508 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1509 nfsi->layout = new;
1510 return new;
1511 } else if (new != NULL)
1512 pnfs_free_layout_hdr(new);
1513out_existing:
1514 pnfs_get_layout_hdr(nfsi->layout);
1515 return nfsi->layout;
1516}
1517
1518/*
1519 * iomode matching rules:
1520 * iomode lseg strict match
1521 * iomode
1522 * ----- ----- ------ -----
1523 * ANY READ N/A true
1524 * ANY RW N/A true
1525 * RW READ N/A false
1526 * RW RW N/A true
1527 * READ READ N/A true
1528 * READ RW true false
1529 * READ RW false true
1530 */
1531static bool
1532pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1533 const struct pnfs_layout_range *range,
1534 bool strict_iomode)
1535{
1536 struct pnfs_layout_range range1;
1537
1538 if ((range->iomode == IOMODE_RW &&
1539 ls_range->iomode != IOMODE_RW) ||
1540 (range->iomode != ls_range->iomode &&
1541 strict_iomode) ||
1542 !pnfs_lseg_range_intersecting(ls_range, range))
1543 return 0;
1544
1545 /* range1 covers only the first byte in the range */
1546 range1 = *range;
1547 range1.length = 1;
1548 return pnfs_lseg_range_contained(ls_range, &range1);
1549}
1550
1551/*
1552 * lookup range in layout
1553 */
1554static struct pnfs_layout_segment *
1555pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1556 struct pnfs_layout_range *range,
1557 bool strict_iomode)
1558{
1559 struct pnfs_layout_segment *lseg, *ret = NULL;
1560
1561 dprintk("%s:Begin\n", __func__);
1562
1563 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1564 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1565 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1566 pnfs_lseg_range_match(&lseg->pls_range, range,
1567 strict_iomode)) {
1568 ret = pnfs_get_lseg(lseg);
1569 break;
1570 }
1571 }
1572
1573 dprintk("%s:Return lseg %p ref %d\n",
1574 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1575 return ret;
1576}
1577
1578/*
1579 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1580 * to the MDS or over pNFS
1581 *
1582 * The nfs_inode read_io and write_io fields are cumulative counters reset
1583 * when there are no layout segments. Note that in pnfs_update_layout iomode
1584 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1585 * WRITE request.
1586 *
1587 * A return of true means use MDS I/O.
1588 *
1589 * From rfc 5661:
1590 * If a file's size is smaller than the file size threshold, data accesses
1591 * SHOULD be sent to the metadata server. If an I/O request has a length that
1592 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1593 * server. If both file size and I/O size are provided, the client SHOULD
1594 * reach or exceed both thresholds before sending its read or write
1595 * requests to the data server.
1596 */
1597static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1598 struct inode *ino, int iomode)
1599{
1600 struct nfs4_threshold *t = ctx->mdsthreshold;
1601 struct nfs_inode *nfsi = NFS_I(ino);
1602 loff_t fsize = i_size_read(ino);
1603 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1604
1605 if (t == NULL)
1606 return ret;
1607
1608 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1609 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1610
1611 switch (iomode) {
1612 case IOMODE_READ:
1613 if (t->bm & THRESHOLD_RD) {
1614 dprintk("%s fsize %llu\n", __func__, fsize);
1615 size_set = true;
1616 if (fsize < t->rd_sz)
1617 size = true;
1618 }
1619 if (t->bm & THRESHOLD_RD_IO) {
1620 dprintk("%s nfsi->read_io %llu\n", __func__,
1621 nfsi->read_io);
1622 io_set = true;
1623 if (nfsi->read_io < t->rd_io_sz)
1624 io = true;
1625 }
1626 break;
1627 case IOMODE_RW:
1628 if (t->bm & THRESHOLD_WR) {
1629 dprintk("%s fsize %llu\n", __func__, fsize);
1630 size_set = true;
1631 if (fsize < t->wr_sz)
1632 size = true;
1633 }
1634 if (t->bm & THRESHOLD_WR_IO) {
1635 dprintk("%s nfsi->write_io %llu\n", __func__,
1636 nfsi->write_io);
1637 io_set = true;
1638 if (nfsi->write_io < t->wr_io_sz)
1639 io = true;
1640 }
1641 break;
1642 }
1643 if (size_set && io_set) {
1644 if (size && io)
1645 ret = true;
1646 } else if (size || io)
1647 ret = true;
1648
1649 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1650 return ret;
1651}
1652
1653static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1654{
1655 /*
1656 * send layoutcommit as it can hold up layoutreturn due to lseg
1657 * reference
1658 */
1659 pnfs_layoutcommit_inode(lo->plh_inode, false);
1660 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1661 nfs_wait_bit_killable,
1662 TASK_UNINTERRUPTIBLE);
1663}
1664
1665static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1666{
1667 unsigned long *bitlock = &lo->plh_flags;
1668
1669 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1670 smp_mb__after_atomic();
1671 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1672}
1673
1674/*
1675 * Layout segment is retreived from the server if not cached.
1676 * The appropriate layout segment is referenced and returned to the caller.
1677 */
1678struct pnfs_layout_segment *
1679pnfs_update_layout(struct inode *ino,
1680 struct nfs_open_context *ctx,
1681 loff_t pos,
1682 u64 count,
1683 enum pnfs_iomode iomode,
1684 bool strict_iomode,
1685 gfp_t gfp_flags)
1686{
1687 struct pnfs_layout_range arg = {
1688 .iomode = iomode,
1689 .offset = pos,
1690 .length = count,
1691 };
1692 unsigned pg_offset;
1693 struct nfs_server *server = NFS_SERVER(ino);
1694 struct nfs_client *clp = server->nfs_client;
1695 struct pnfs_layout_hdr *lo = NULL;
1696 struct pnfs_layout_segment *lseg = NULL;
1697 nfs4_stateid stateid;
1698 long timeout = 0;
1699 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1700 bool first;
1701
1702 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1703 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1704 PNFS_UPDATE_LAYOUT_NO_PNFS);
1705 goto out;
1706 }
1707
1708 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1709 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1710 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1711 goto out;
1712 }
1713
1714 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1715 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1716 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1717 goto out;
1718 }
1719
1720lookup_again:
1721 nfs4_client_recover_expired_lease(clp);
1722 first = false;
1723 spin_lock(&ino->i_lock);
1724 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1725 if (lo == NULL) {
1726 spin_unlock(&ino->i_lock);
1727 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1728 PNFS_UPDATE_LAYOUT_NOMEM);
1729 goto out;
1730 }
1731
1732 /* Do we even need to bother with this? */
1733 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1734 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1735 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1736 dprintk("%s matches recall, use MDS\n", __func__);
1737 goto out_unlock;
1738 }
1739
1740 /* if LAYOUTGET already failed once we don't try again */
1741 if (pnfs_layout_io_test_failed(lo, iomode)) {
1742 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1743 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1744 goto out_unlock;
1745 }
1746
1747 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1748 if (lseg) {
1749 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1750 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1751 goto out_unlock;
1752 }
1753
1754 if (!nfs4_valid_open_stateid(ctx->state)) {
1755 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1756 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1757 goto out_unlock;
1758 }
1759
1760 /*
1761 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1762 * stateid, or it has been invalidated, then we must use the open
1763 * stateid.
1764 */
1765 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1766
1767 /*
1768 * The first layoutget for the file. Need to serialize per
1769 * RFC 5661 Errata 3208.
1770 */
1771 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1772 &lo->plh_flags)) {
1773 spin_unlock(&ino->i_lock);
1774 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1775 TASK_UNINTERRUPTIBLE);
1776 pnfs_put_layout_hdr(lo);
1777 dprintk("%s retrying\n", __func__);
1778 goto lookup_again;
1779 }
1780
1781 first = true;
1782 if (nfs4_select_rw_stateid(ctx->state,
1783 iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
1784 NULL, &stateid, NULL) != 0) {
1785 trace_pnfs_update_layout(ino, pos, count,
1786 iomode, lo, lseg,
1787 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1788 goto out_unlock;
1789 }
1790 } else {
1791 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1792 }
1793
1794 /*
1795 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1796 * for LAYOUTRETURN even if first is true.
1797 */
1798 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1799 spin_unlock(&ino->i_lock);
1800 dprintk("%s wait for layoutreturn\n", __func__);
1801 if (pnfs_prepare_to_retry_layoutget(lo)) {
1802 if (first)
1803 pnfs_clear_first_layoutget(lo);
1804 pnfs_put_layout_hdr(lo);
1805 dprintk("%s retrying\n", __func__);
1806 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1807 lseg, PNFS_UPDATE_LAYOUT_RETRY);
1808 goto lookup_again;
1809 }
1810 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1811 PNFS_UPDATE_LAYOUT_RETURN);
1812 goto out_put_layout_hdr;
1813 }
1814
1815 if (pnfs_layoutgets_blocked(lo)) {
1816 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1817 PNFS_UPDATE_LAYOUT_BLOCKED);
1818 goto out_unlock;
1819 }
1820 atomic_inc(&lo->plh_outstanding);
1821 spin_unlock(&ino->i_lock);
1822
1823 if (list_empty(&lo->plh_layouts)) {
1824 /* The lo must be on the clp list if there is any
1825 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1826 */
1827 spin_lock(&clp->cl_lock);
1828 if (list_empty(&lo->plh_layouts))
1829 list_add_tail(&lo->plh_layouts, &server->layouts);
1830 spin_unlock(&clp->cl_lock);
1831 }
1832
1833 pg_offset = arg.offset & ~PAGE_MASK;
1834 if (pg_offset) {
1835 arg.offset -= pg_offset;
1836 arg.length += pg_offset;
1837 }
1838 if (arg.length != NFS4_MAX_UINT64)
1839 arg.length = PAGE_ALIGN(arg.length);
1840
1841 lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1842 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1843 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1844 atomic_dec(&lo->plh_outstanding);
1845 if (IS_ERR(lseg)) {
1846 switch(PTR_ERR(lseg)) {
1847 case -EBUSY:
1848 if (time_after(jiffies, giveup))
1849 lseg = NULL;
1850 break;
1851 case -ERECALLCONFLICT:
1852 /* Huh? We hold no layouts, how is there a recall? */
1853 if (first) {
1854 lseg = NULL;
1855 break;
1856 }
1857 /* Destroy the existing layout and start over */
1858 if (time_after(jiffies, giveup))
1859 pnfs_destroy_layout(NFS_I(ino));
1860 /* Fallthrough */
1861 case -EAGAIN:
1862 break;
1863 default:
1864 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1865 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1866 lseg = NULL;
1867 }
1868 goto out_put_layout_hdr;
1869 }
1870 if (lseg) {
1871 if (first)
1872 pnfs_clear_first_layoutget(lo);
1873 trace_pnfs_update_layout(ino, pos, count,
1874 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1875 pnfs_put_layout_hdr(lo);
1876 goto lookup_again;
1877 }
1878 } else {
1879 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1880 }
1881
1882out_put_layout_hdr:
1883 if (first)
1884 pnfs_clear_first_layoutget(lo);
1885 pnfs_put_layout_hdr(lo);
1886out:
1887 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1888 "(%s, offset: %llu, length: %llu)\n",
1889 __func__, ino->i_sb->s_id,
1890 (unsigned long long)NFS_FILEID(ino),
1891 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1892 iomode==IOMODE_RW ? "read/write" : "read-only",
1893 (unsigned long long)pos,
1894 (unsigned long long)count);
1895 return lseg;
1896out_unlock:
1897 spin_unlock(&ino->i_lock);
1898 goto out_put_layout_hdr;
1899}
1900EXPORT_SYMBOL_GPL(pnfs_update_layout);
1901
1902static bool
1903pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1904{
1905 switch (range->iomode) {
1906 case IOMODE_READ:
1907 case IOMODE_RW:
1908 break;
1909 default:
1910 return false;
1911 }
1912 if (range->offset == NFS4_MAX_UINT64)
1913 return false;
1914 if (range->length == 0)
1915 return false;
1916 if (range->length != NFS4_MAX_UINT64 &&
1917 range->length > NFS4_MAX_UINT64 - range->offset)
1918 return false;
1919 return true;
1920}
1921
1922struct pnfs_layout_segment *
1923pnfs_layout_process(struct nfs4_layoutget *lgp)
1924{
1925 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1926 struct nfs4_layoutget_res *res = &lgp->res;
1927 struct pnfs_layout_segment *lseg;
1928 struct inode *ino = lo->plh_inode;
1929 LIST_HEAD(free_me);
1930
1931 if (!pnfs_sanity_check_layout_range(&res->range))
1932 return ERR_PTR(-EINVAL);
1933
1934 /* Inject layout blob into I/O device driver */
1935 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1936 if (IS_ERR_OR_NULL(lseg)) {
1937 if (!lseg)
1938 lseg = ERR_PTR(-ENOMEM);
1939
1940 dprintk("%s: Could not allocate layout: error %ld\n",
1941 __func__, PTR_ERR(lseg));
1942 return lseg;
1943 }
1944
1945 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1946
1947 spin_lock(&ino->i_lock);
1948 if (pnfs_layoutgets_blocked(lo)) {
1949 dprintk("%s forget reply due to state\n", __func__);
1950 goto out_forget;
1951 }
1952
1953 if (!pnfs_layout_is_valid(lo)) {
1954 /* We have a completely new layout */
1955 pnfs_set_layout_stateid(lo, &res->stateid, true);
1956 } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1957 /* existing state ID, make sure the sequence number matches. */
1958 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1959 dprintk("%s forget reply due to sequence\n", __func__);
1960 goto out_forget;
1961 }
1962 pnfs_set_layout_stateid(lo, &res->stateid, false);
1963 } else {
1964 /*
1965 * We got an entirely new state ID. Mark all segments for the
1966 * inode invalid, and retry the layoutget
1967 */
1968 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1969 goto out_forget;
1970 }
1971
1972 pnfs_get_lseg(lseg);
1973 pnfs_layout_insert_lseg(lo, lseg, &free_me);
1974
1975
1976 if (res->return_on_close)
1977 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1978
1979 spin_unlock(&ino->i_lock);
1980 pnfs_free_lseg_list(&free_me);
1981 return lseg;
1982
1983out_forget:
1984 spin_unlock(&ino->i_lock);
1985 lseg->pls_layout = lo;
1986 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1987 if (!pnfs_layout_is_valid(lo))
1988 nfs_commit_inode(ino, 0);
1989 return ERR_PTR(-EAGAIN);
1990}
1991
1992/**
1993 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1994 * @lo: pointer to layout header
1995 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1996 * @return_range: describe layout segment ranges to be returned
1997 *
1998 * This function is mainly intended for use by layoutrecall. It attempts
1999 * to free the layout segment immediately, or else to mark it for return
2000 * as soon as its reference count drops to zero.
2001 */
2002int
2003pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2004 struct list_head *tmp_list,
2005 const struct pnfs_layout_range *return_range,
2006 u32 seq)
2007{
2008 struct pnfs_layout_segment *lseg, *next;
2009 int remaining = 0;
2010
2011 dprintk("%s:Begin lo %p\n", __func__, lo);
2012
2013 if (list_empty(&lo->plh_segs))
2014 return 0;
2015
2016 assert_spin_locked(&lo->plh_inode->i_lock);
2017
2018 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2019 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2020 dprintk("%s: marking lseg %p iomode %d "
2021 "offset %llu length %llu\n", __func__,
2022 lseg, lseg->pls_range.iomode,
2023 lseg->pls_range.offset,
2024 lseg->pls_range.length);
2025 if (mark_lseg_invalid(lseg, tmp_list))
2026 continue;
2027 remaining++;
2028 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2029 }
2030
2031 if (remaining)
2032 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2033
2034 return remaining;
2035}
2036
2037void pnfs_error_mark_layout_for_return(struct inode *inode,
2038 struct pnfs_layout_segment *lseg)
2039{
2040 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2041 struct pnfs_layout_range range = {
2042 .iomode = lseg->pls_range.iomode,
2043 .offset = 0,
2044 .length = NFS4_MAX_UINT64,
2045 };
2046 bool return_now = false;
2047
2048 spin_lock(&inode->i_lock);
2049 if (!pnfs_layout_is_valid(lo)) {
2050 spin_unlock(&inode->i_lock);
2051 return;
2052 }
2053 pnfs_set_plh_return_info(lo, range.iomode, 0);
2054 /*
2055 * mark all matching lsegs so that we are sure to have no live
2056 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2057 * for how it works.
2058 */
2059 if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2060 nfs4_stateid stateid;
2061 enum pnfs_iomode iomode;
2062
2063 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2064 spin_unlock(&inode->i_lock);
2065 if (return_now)
2066 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2067 } else {
2068 spin_unlock(&inode->i_lock);
2069 nfs_commit_inode(inode, 0);
2070 }
2071}
2072EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2073
2074void
2075pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2076{
2077 if (pgio->pg_lseg == NULL ||
2078 test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2079 return;
2080 pnfs_put_lseg(pgio->pg_lseg);
2081 pgio->pg_lseg = NULL;
2082}
2083EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2084
2085/*
2086 * Check for any intersection between the request and the pgio->pg_lseg,
2087 * and if none, put this pgio->pg_lseg away.
2088 */
2089static void
2090pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2091{
2092 if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2093 pnfs_put_lseg(pgio->pg_lseg);
2094 pgio->pg_lseg = NULL;
2095 }
2096}
2097
2098void
2099pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2100{
2101 u64 rd_size = req->wb_bytes;
2102
2103 pnfs_generic_pg_check_layout(pgio);
2104 pnfs_generic_pg_check_range(pgio, req);
2105 if (pgio->pg_lseg == NULL) {
2106 if (pgio->pg_dreq == NULL)
2107 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2108 else
2109 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2110
2111 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2112 req->wb_context,
2113 req_offset(req),
2114 rd_size,
2115 IOMODE_READ,
2116 false,
2117 GFP_KERNEL);
2118 if (IS_ERR(pgio->pg_lseg)) {
2119 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2120 pgio->pg_lseg = NULL;
2121 return;
2122 }
2123 }
2124 /* If no lseg, fall back to read through mds */
2125 if (pgio->pg_lseg == NULL)
2126 nfs_pageio_reset_read_mds(pgio);
2127
2128}
2129EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2130
2131void
2132pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2133 struct nfs_page *req, u64 wb_size)
2134{
2135 pnfs_generic_pg_check_layout(pgio);
2136 pnfs_generic_pg_check_range(pgio, req);
2137 if (pgio->pg_lseg == NULL) {
2138 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2139 req->wb_context,
2140 req_offset(req),
2141 wb_size,
2142 IOMODE_RW,
2143 false,
2144 GFP_NOFS);
2145 if (IS_ERR(pgio->pg_lseg)) {
2146 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2147 pgio->pg_lseg = NULL;
2148 return;
2149 }
2150 }
2151 /* If no lseg, fall back to write through mds */
2152 if (pgio->pg_lseg == NULL)
2153 nfs_pageio_reset_write_mds(pgio);
2154}
2155EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2156
2157void
2158pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2159{
2160 if (desc->pg_lseg) {
2161 pnfs_put_lseg(desc->pg_lseg);
2162 desc->pg_lseg = NULL;
2163 }
2164}
2165EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2166
2167/*
2168 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2169 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2170 */
2171size_t
2172pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2173 struct nfs_page *prev, struct nfs_page *req)
2174{
2175 unsigned int size;
2176 u64 seg_end, req_start, seg_left;
2177
2178 size = nfs_generic_pg_test(pgio, prev, req);
2179 if (!size)
2180 return 0;
2181
2182 /*
2183 * 'size' contains the number of bytes left in the current page (up
2184 * to the original size asked for in @req->wb_bytes).
2185 *
2186 * Calculate how many bytes are left in the layout segment
2187 * and if there are less bytes than 'size', return that instead.
2188 *
2189 * Please also note that 'end_offset' is actually the offset of the
2190 * first byte that lies outside the pnfs_layout_range. FIXME?
2191 *
2192 */
2193 if (pgio->pg_lseg) {
2194 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2195 pgio->pg_lseg->pls_range.length);
2196 req_start = req_offset(req);
2197
2198 /* start of request is past the last byte of this segment */
2199 if (req_start >= seg_end)
2200 return 0;
2201
2202 /* adjust 'size' iff there are fewer bytes left in the
2203 * segment than what nfs_generic_pg_test returned */
2204 seg_left = seg_end - req_start;
2205 if (seg_left < size)
2206 size = (unsigned int)seg_left;
2207 }
2208
2209 return size;
2210}
2211EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2212
2213int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2214{
2215 struct nfs_pageio_descriptor pgio;
2216
2217 /* Resend all requests through the MDS */
2218 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2219 hdr->completion_ops);
2220 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2221 return nfs_pageio_resend(&pgio, hdr);
2222}
2223EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2224
2225static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2226{
2227
2228 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2229 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2230 PNFS_LAYOUTRET_ON_ERROR) {
2231 pnfs_return_layout(hdr->inode);
2232 }
2233 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2234 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2235}
2236
2237/*
2238 * Called by non rpc-based layout drivers
2239 */
2240void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2241{
2242 if (likely(!hdr->pnfs_error)) {
2243 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2244 hdr->mds_offset + hdr->res.count);
2245 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2246 }
2247 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2248 if (unlikely(hdr->pnfs_error))
2249 pnfs_ld_handle_write_error(hdr);
2250 hdr->mds_ops->rpc_release(hdr);
2251}
2252EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2253
2254static void
2255pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2256 struct nfs_pgio_header *hdr)
2257{
2258 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2259
2260 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2261 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2262 nfs_pageio_reset_write_mds(desc);
2263 mirror->pg_recoalesce = 1;
2264 }
2265 hdr->completion_ops->completion(hdr);
2266}
2267
2268static enum pnfs_try_status
2269pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2270 const struct rpc_call_ops *call_ops,
2271 struct pnfs_layout_segment *lseg,
2272 int how)
2273{
2274 struct inode *inode = hdr->inode;
2275 enum pnfs_try_status trypnfs;
2276 struct nfs_server *nfss = NFS_SERVER(inode);
2277
2278 hdr->mds_ops = call_ops;
2279
2280 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2281 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2282 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2283 if (trypnfs != PNFS_NOT_ATTEMPTED)
2284 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2285 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2286 return trypnfs;
2287}
2288
2289static void
2290pnfs_do_write(struct nfs_pageio_descriptor *desc,
2291 struct nfs_pgio_header *hdr, int how)
2292{
2293 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2294 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2295 enum pnfs_try_status trypnfs;
2296
2297 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2298 switch (trypnfs) {
2299 case PNFS_NOT_ATTEMPTED:
2300 pnfs_write_through_mds(desc, hdr);
2301 case PNFS_ATTEMPTED:
2302 break;
2303 case PNFS_TRY_AGAIN:
2304 /* cleanup hdr and prepare to redo pnfs */
2305 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2306 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2307 list_splice_init(&hdr->pages, &mirror->pg_list);
2308 mirror->pg_recoalesce = 1;
2309 }
2310 hdr->mds_ops->rpc_release(hdr);
2311 }
2312}
2313
2314static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2315{
2316 pnfs_put_lseg(hdr->lseg);
2317 nfs_pgio_header_free(hdr);
2318}
2319
2320int
2321pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2322{
2323 struct nfs_pgio_header *hdr;
2324 int ret;
2325
2326 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2327 if (!hdr) {
2328 desc->pg_error = -ENOMEM;
2329 return desc->pg_error;
2330 }
2331 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2332
2333 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2334 ret = nfs_generic_pgio(desc, hdr);
2335 if (!ret)
2336 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2337
2338 return ret;
2339}
2340EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2341
2342int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2343{
2344 struct nfs_pageio_descriptor pgio;
2345
2346 /* Resend all requests through the MDS */
2347 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2348 return nfs_pageio_resend(&pgio, hdr);
2349}
2350EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2351
2352static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2353{
2354 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2355 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2356 PNFS_LAYOUTRET_ON_ERROR) {
2357 pnfs_return_layout(hdr->inode);
2358 }
2359 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2360 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2361}
2362
2363/*
2364 * Called by non rpc-based layout drivers
2365 */
2366void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2367{
2368 if (likely(!hdr->pnfs_error))
2369 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2370 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2371 if (unlikely(hdr->pnfs_error))
2372 pnfs_ld_handle_read_error(hdr);
2373 hdr->mds_ops->rpc_release(hdr);
2374}
2375EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2376
2377static void
2378pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2379 struct nfs_pgio_header *hdr)
2380{
2381 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2382
2383 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2384 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2385 nfs_pageio_reset_read_mds(desc);
2386 mirror->pg_recoalesce = 1;
2387 }
2388 hdr->completion_ops->completion(hdr);
2389}
2390
2391/*
2392 * Call the appropriate parallel I/O subsystem read function.
2393 */
2394static enum pnfs_try_status
2395pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2396 const struct rpc_call_ops *call_ops,
2397 struct pnfs_layout_segment *lseg)
2398{
2399 struct inode *inode = hdr->inode;
2400 struct nfs_server *nfss = NFS_SERVER(inode);
2401 enum pnfs_try_status trypnfs;
2402
2403 hdr->mds_ops = call_ops;
2404
2405 dprintk("%s: Reading ino:%lu %u@%llu\n",
2406 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2407
2408 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2409 if (trypnfs != PNFS_NOT_ATTEMPTED)
2410 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2411 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2412 return trypnfs;
2413}
2414
2415/* Resend all requests through pnfs. */
2416void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2417{
2418 struct nfs_pageio_descriptor pgio;
2419
2420 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2421 /* Prevent deadlocks with layoutreturn! */
2422 pnfs_put_lseg(hdr->lseg);
2423 hdr->lseg = NULL;
2424
2425 nfs_pageio_init_read(&pgio, hdr->inode, false,
2426 hdr->completion_ops);
2427 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2428 }
2429}
2430EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2431
2432static void
2433pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2434{
2435 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2436 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2437 enum pnfs_try_status trypnfs;
2438
2439 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2440 switch (trypnfs) {
2441 case PNFS_NOT_ATTEMPTED:
2442 pnfs_read_through_mds(desc, hdr);
2443 case PNFS_ATTEMPTED:
2444 break;
2445 case PNFS_TRY_AGAIN:
2446 /* cleanup hdr and prepare to redo pnfs */
2447 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2448 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2449 list_splice_init(&hdr->pages, &mirror->pg_list);
2450 mirror->pg_recoalesce = 1;
2451 }
2452 hdr->mds_ops->rpc_release(hdr);
2453 }
2454}
2455
2456static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2457{
2458 pnfs_put_lseg(hdr->lseg);
2459 nfs_pgio_header_free(hdr);
2460}
2461
2462int
2463pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2464{
2465 struct nfs_pgio_header *hdr;
2466 int ret;
2467
2468 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2469 if (!hdr) {
2470 desc->pg_error = -ENOMEM;
2471 return desc->pg_error;
2472 }
2473 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2474 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2475 ret = nfs_generic_pgio(desc, hdr);
2476 if (!ret)
2477 pnfs_do_read(desc, hdr);
2478 return ret;
2479}
2480EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2481
2482static void pnfs_clear_layoutcommitting(struct inode *inode)
2483{
2484 unsigned long *bitlock = &NFS_I(inode)->flags;
2485
2486 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2487 smp_mb__after_atomic();
2488 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2489}
2490
2491/*
2492 * There can be multiple RW segments.
2493 */
2494static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2495{
2496 struct pnfs_layout_segment *lseg;
2497
2498 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2499 if (lseg->pls_range.iomode == IOMODE_RW &&
2500 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2501 list_add(&lseg->pls_lc_list, listp);
2502 }
2503}
2504
2505static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2506{
2507 struct pnfs_layout_segment *lseg, *tmp;
2508
2509 /* Matched by references in pnfs_set_layoutcommit */
2510 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2511 list_del_init(&lseg->pls_lc_list);
2512 pnfs_put_lseg(lseg);
2513 }
2514
2515 pnfs_clear_layoutcommitting(inode);
2516}
2517
2518void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2519{
2520 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2521}
2522EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2523
2524void
2525pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2526 loff_t end_pos)
2527{
2528 struct nfs_inode *nfsi = NFS_I(inode);
2529 bool mark_as_dirty = false;
2530
2531 spin_lock(&inode->i_lock);
2532 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2533 nfsi->layout->plh_lwb = end_pos;
2534 mark_as_dirty = true;
2535 dprintk("%s: Set layoutcommit for inode %lu ",
2536 __func__, inode->i_ino);
2537 } else if (end_pos > nfsi->layout->plh_lwb)
2538 nfsi->layout->plh_lwb = end_pos;
2539 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2540 /* references matched in nfs4_layoutcommit_release */
2541 pnfs_get_lseg(lseg);
2542 }
2543 spin_unlock(&inode->i_lock);
2544 dprintk("%s: lseg %p end_pos %llu\n",
2545 __func__, lseg, nfsi->layout->plh_lwb);
2546
2547 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2548 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2549 if (mark_as_dirty)
2550 mark_inode_dirty_sync(inode);
2551}
2552EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2553
2554void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2555{
2556 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2557
2558 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2559 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2560 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2561}
2562
2563/*
2564 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2565 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2566 * data to disk to allow the server to recover the data if it crashes.
2567 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2568 * is off, and a COMMIT is sent to a data server, or
2569 * if WRITEs to a data server return NFS_DATA_SYNC.
2570 */
2571int
2572pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2573{
2574 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2575 struct nfs4_layoutcommit_data *data;
2576 struct nfs_inode *nfsi = NFS_I(inode);
2577 loff_t end_pos;
2578 int status;
2579
2580 if (!pnfs_layoutcommit_outstanding(inode))
2581 return 0;
2582
2583 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2584
2585 status = -EAGAIN;
2586 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2587 if (!sync)
2588 goto out;
2589 status = wait_on_bit_lock_action(&nfsi->flags,
2590 NFS_INO_LAYOUTCOMMITTING,
2591 nfs_wait_bit_killable,
2592 TASK_KILLABLE);
2593 if (status)
2594 goto out;
2595 }
2596
2597 status = -ENOMEM;
2598 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2599 data = kzalloc(sizeof(*data), GFP_NOFS);
2600 if (!data)
2601 goto clear_layoutcommitting;
2602
2603 status = 0;
2604 spin_lock(&inode->i_lock);
2605 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2606 goto out_unlock;
2607
2608 INIT_LIST_HEAD(&data->lseg_list);
2609 pnfs_list_write_lseg(inode, &data->lseg_list);
2610
2611 end_pos = nfsi->layout->plh_lwb;
2612
2613 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2614 spin_unlock(&inode->i_lock);
2615
2616 data->args.inode = inode;
2617 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2618 nfs_fattr_init(&data->fattr);
2619 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2620 data->res.fattr = &data->fattr;
2621 if (end_pos != 0)
2622 data->args.lastbytewritten = end_pos - 1;
2623 else
2624 data->args.lastbytewritten = U64_MAX;
2625 data->res.server = NFS_SERVER(inode);
2626
2627 if (ld->prepare_layoutcommit) {
2628 status = ld->prepare_layoutcommit(&data->args);
2629 if (status) {
2630 put_rpccred(data->cred);
2631 spin_lock(&inode->i_lock);
2632 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2633 if (end_pos > nfsi->layout->plh_lwb)
2634 nfsi->layout->plh_lwb = end_pos;
2635 goto out_unlock;
2636 }
2637 }
2638
2639
2640 status = nfs4_proc_layoutcommit(data, sync);
2641out:
2642 if (status)
2643 mark_inode_dirty_sync(inode);
2644 dprintk("<-- %s status %d\n", __func__, status);
2645 return status;
2646out_unlock:
2647 spin_unlock(&inode->i_lock);
2648 kfree(data);
2649clear_layoutcommitting:
2650 pnfs_clear_layoutcommitting(inode);
2651 goto out;
2652}
2653EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2654
2655int
2656pnfs_generic_sync(struct inode *inode, bool datasync)
2657{
2658 return pnfs_layoutcommit_inode(inode, true);
2659}
2660EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2661
2662struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2663{
2664 struct nfs4_threshold *thp;
2665
2666 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2667 if (!thp) {
2668 dprintk("%s mdsthreshold allocation failed\n", __func__);
2669 return NULL;
2670 }
2671 return thp;
2672}
2673
2674#if IS_ENABLED(CONFIG_NFS_V4_2)
2675int
2676pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2677{
2678 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2679 struct nfs_server *server = NFS_SERVER(inode);
2680 struct nfs_inode *nfsi = NFS_I(inode);
2681 struct nfs42_layoutstat_data *data;
2682 struct pnfs_layout_hdr *hdr;
2683 int status = 0;
2684
2685 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2686 goto out;
2687
2688 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2689 goto out;
2690
2691 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2692 goto out;
2693
2694 spin_lock(&inode->i_lock);
2695 if (!NFS_I(inode)->layout) {
2696 spin_unlock(&inode->i_lock);
2697 goto out_clear_layoutstats;
2698 }
2699 hdr = NFS_I(inode)->layout;
2700 pnfs_get_layout_hdr(hdr);
2701 spin_unlock(&inode->i_lock);
2702
2703 data = kzalloc(sizeof(*data), gfp_flags);
2704 if (!data) {
2705 status = -ENOMEM;
2706 goto out_put;
2707 }
2708
2709 data->args.fh = NFS_FH(inode);
2710 data->args.inode = inode;
2711 status = ld->prepare_layoutstats(&data->args);
2712 if (status)
2713 goto out_free;
2714
2715 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2716
2717out:
2718 dprintk("%s returns %d\n", __func__, status);
2719 return status;
2720
2721out_free:
2722 kfree(data);
2723out_put:
2724 pnfs_put_layout_hdr(hdr);
2725out_clear_layoutstats:
2726 smp_mb__before_atomic();
2727 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2728 smp_mb__after_atomic();
2729 goto out;
2730}
2731EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2732#endif
2733
2734unsigned int layoutstats_timer;
2735module_param(layoutstats_timer, uint, 0644);
2736EXPORT_SYMBOL_GPL(layoutstats_timer);
1/*
2 * pNFS functions to call and manage layout drivers.
3 *
4 * Copyright (c) 2002 [year of first publication]
5 * The Regents of the University of Michigan
6 * All Rights Reserved
7 *
8 * Dean Hildebrand <dhildebz@umich.edu>
9 *
10 * Permission is granted to use, copy, create derivative works, and
11 * redistribute this software and such derivative works for any purpose,
12 * so long as the name of the University of Michigan is not used in
13 * any advertising or publicity pertaining to the use or distribution
14 * of this software without specific, written prior authorization. If
15 * the above copyright notice or any other identification of the
16 * University of Michigan is included in any copy of any portion of
17 * this software, then the disclaimer below must also be included.
18 *
19 * This software is provided as is, without representation or warranty
20 * of any kind either express or implied, including without limitation
21 * the implied warranties of merchantability, fitness for a particular
22 * purpose, or noninfringement. The Regents of the University of
23 * Michigan shall not be liable for any damages, including special,
24 * indirect, incidental, or consequential damages, with respect to any
25 * claim arising out of or in connection with the use of the software,
26 * even if it has been or is hereafter advised of the possibility of
27 * such damages.
28 */
29
30#include <linux/nfs_fs.h>
31#include <linux/nfs_page.h>
32#include <linux/module.h>
33#include <linux/sort.h>
34#include "internal.h"
35#include "pnfs.h"
36#include "iostat.h"
37#include "nfs4trace.h"
38#include "delegation.h"
39#include "nfs42.h"
40#include "nfs4_fs.h"
41
42#define NFSDBG_FACILITY NFSDBG_PNFS
43#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
44
45/* Locking:
46 *
47 * pnfs_spinlock:
48 * protects pnfs_modules_tbl.
49 */
50static DEFINE_SPINLOCK(pnfs_spinlock);
51
52/*
53 * pnfs_modules_tbl holds all pnfs modules
54 */
55static LIST_HEAD(pnfs_modules_tbl);
56
57static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
58static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
59 struct list_head *free_me,
60 const struct pnfs_layout_range *range,
61 u32 seq);
62static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
63 struct list_head *tmp_list);
64
65/* Return the registered pnfs layout driver module matching given id */
66static struct pnfs_layoutdriver_type *
67find_pnfs_driver_locked(u32 id)
68{
69 struct pnfs_layoutdriver_type *local;
70
71 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
72 if (local->id == id)
73 goto out;
74 local = NULL;
75out:
76 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
77 return local;
78}
79
80static struct pnfs_layoutdriver_type *
81find_pnfs_driver(u32 id)
82{
83 struct pnfs_layoutdriver_type *local;
84
85 spin_lock(&pnfs_spinlock);
86 local = find_pnfs_driver_locked(id);
87 if (local != NULL && !try_module_get(local->owner)) {
88 dprintk("%s: Could not grab reference on module\n", __func__);
89 local = NULL;
90 }
91 spin_unlock(&pnfs_spinlock);
92 return local;
93}
94
95const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
96{
97 return find_pnfs_driver(id);
98}
99
100void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
101{
102 if (ld)
103 module_put(ld->owner);
104}
105
106void
107unset_pnfs_layoutdriver(struct nfs_server *nfss)
108{
109 if (nfss->pnfs_curr_ld) {
110 if (nfss->pnfs_curr_ld->clear_layoutdriver)
111 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
112 /* Decrement the MDS count. Purge the deviceid cache if zero */
113 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
114 nfs4_deviceid_purge_client(nfss->nfs_client);
115 module_put(nfss->pnfs_curr_ld->owner);
116 }
117 nfss->pnfs_curr_ld = NULL;
118}
119
120/*
121 * When the server sends a list of layout types, we choose one in the order
122 * given in the list below.
123 *
124 * FIXME: should this list be configurable in some fashion? module param?
125 * mount option? something else?
126 */
127static const u32 ld_prefs[] = {
128 LAYOUT_SCSI,
129 LAYOUT_BLOCK_VOLUME,
130 LAYOUT_OSD2_OBJECTS,
131 LAYOUT_FLEX_FILES,
132 LAYOUT_NFSV4_1_FILES,
133 0
134};
135
136static int
137ld_cmp(const void *e1, const void *e2)
138{
139 u32 ld1 = *((u32 *)e1);
140 u32 ld2 = *((u32 *)e2);
141 int i;
142
143 for (i = 0; ld_prefs[i] != 0; i++) {
144 if (ld1 == ld_prefs[i])
145 return -1;
146
147 if (ld2 == ld_prefs[i])
148 return 1;
149 }
150 return 0;
151}
152
153/*
154 * Try to set the server's pnfs module to the pnfs layout type specified by id.
155 * Currently only one pNFS layout driver per filesystem is supported.
156 *
157 * @ids array of layout types supported by MDS.
158 */
159void
160set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
161 struct nfs_fsinfo *fsinfo)
162{
163 struct pnfs_layoutdriver_type *ld_type = NULL;
164 u32 id;
165 int i;
166
167 if (fsinfo->nlayouttypes == 0)
168 goto out_no_driver;
169 if (!(server->nfs_client->cl_exchange_flags &
170 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
171 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
172 __func__, server->nfs_client->cl_exchange_flags);
173 goto out_no_driver;
174 }
175
176 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
177 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
178
179 for (i = 0; i < fsinfo->nlayouttypes; i++) {
180 id = fsinfo->layouttype[i];
181 ld_type = find_pnfs_driver(id);
182 if (!ld_type) {
183 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
184 id);
185 ld_type = find_pnfs_driver(id);
186 }
187 if (ld_type)
188 break;
189 }
190
191 if (!ld_type) {
192 dprintk("%s: No pNFS module found!\n", __func__);
193 goto out_no_driver;
194 }
195
196 server->pnfs_curr_ld = ld_type;
197 if (ld_type->set_layoutdriver
198 && ld_type->set_layoutdriver(server, mntfh)) {
199 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
200 "driver %u.\n", __func__, id);
201 module_put(ld_type->owner);
202 goto out_no_driver;
203 }
204 /* Bump the MDS count */
205 atomic_inc(&server->nfs_client->cl_mds_count);
206
207 dprintk("%s: pNFS module for %u set\n", __func__, id);
208 return;
209
210out_no_driver:
211 dprintk("%s: Using NFSv4 I/O\n", __func__);
212 server->pnfs_curr_ld = NULL;
213}
214
215int
216pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
217{
218 int status = -EINVAL;
219 struct pnfs_layoutdriver_type *tmp;
220
221 if (ld_type->id == 0) {
222 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
223 return status;
224 }
225 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
226 printk(KERN_ERR "NFS: %s Layout driver must provide "
227 "alloc_lseg and free_lseg.\n", __func__);
228 return status;
229 }
230
231 spin_lock(&pnfs_spinlock);
232 tmp = find_pnfs_driver_locked(ld_type->id);
233 if (!tmp) {
234 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
235 status = 0;
236 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
237 ld_type->name);
238 } else {
239 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
240 __func__, ld_type->id);
241 }
242 spin_unlock(&pnfs_spinlock);
243
244 return status;
245}
246EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
247
248void
249pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
250{
251 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
252 spin_lock(&pnfs_spinlock);
253 list_del(&ld_type->pnfs_tblid);
254 spin_unlock(&pnfs_spinlock);
255}
256EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
257
258/*
259 * pNFS client layout cache
260 */
261
262/* Need to hold i_lock if caller does not already hold reference */
263void
264pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
265{
266 refcount_inc(&lo->plh_refcount);
267}
268
269static struct pnfs_layout_hdr *
270pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
271{
272 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
273 return ld->alloc_layout_hdr(ino, gfp_flags);
274}
275
276static void
277pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
278{
279 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
280 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
281
282 if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
283 struct nfs_client *clp = server->nfs_client;
284
285 spin_lock(&clp->cl_lock);
286 list_del_rcu(&lo->plh_layouts);
287 spin_unlock(&clp->cl_lock);
288 }
289 put_cred(lo->plh_lc_cred);
290 return ld->free_layout_hdr(lo);
291}
292
293static void
294pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
295{
296 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
297 dprintk("%s: freeing layout cache %p\n", __func__, lo);
298 nfsi->layout = NULL;
299 /* Reset MDS Threshold I/O counters */
300 nfsi->write_io = 0;
301 nfsi->read_io = 0;
302}
303
304void
305pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
306{
307 struct inode *inode;
308 unsigned long i_state;
309
310 if (!lo)
311 return;
312 inode = lo->plh_inode;
313 pnfs_layoutreturn_before_put_layout_hdr(lo);
314
315 if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
316 if (!list_empty(&lo->plh_segs))
317 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
318 pnfs_detach_layout_hdr(lo);
319 i_state = inode->i_state;
320 spin_unlock(&inode->i_lock);
321 pnfs_free_layout_hdr(lo);
322 /* Notify pnfs_destroy_layout_final() that we're done */
323 if (i_state & (I_FREEING | I_CLEAR))
324 wake_up_var(lo);
325 }
326}
327
328static struct inode *
329pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
330{
331 struct inode *inode = igrab(lo->plh_inode);
332 if (inode)
333 return inode;
334 set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
335 return NULL;
336}
337
338/*
339 * Compare 2 layout stateid sequence ids, to see which is newer,
340 * taking into account wraparound issues.
341 */
342static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
343{
344 return (s32)(s1 - s2) > 0;
345}
346
347static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
348{
349 if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
350 lo->plh_barrier = newseq;
351}
352
353static void
354pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
355 u32 seq)
356{
357 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
358 iomode = IOMODE_ANY;
359 lo->plh_return_iomode = iomode;
360 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
361 /*
362 * We must set lo->plh_return_seq to avoid livelocks with
363 * pnfs_layout_need_return()
364 */
365 if (seq == 0)
366 seq = be32_to_cpu(lo->plh_stateid.seqid);
367 if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
368 lo->plh_return_seq = seq;
369 pnfs_barrier_update(lo, seq);
370}
371
372static void
373pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
374{
375 struct pnfs_layout_segment *lseg;
376 lo->plh_return_iomode = 0;
377 lo->plh_return_seq = 0;
378 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
379 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
380 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
381 continue;
382 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
383 }
384}
385
386static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
387{
388 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
389 clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
390 smp_mb__after_atomic();
391 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
392 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
393}
394
395static void
396pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
397 struct list_head *free_me)
398{
399 clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
400 clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
401 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
402 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
403 if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
404 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
405}
406
407/*
408 * Update the seqid of a layout stateid after receiving
409 * NFS4ERR_OLD_STATEID
410 */
411bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
412 struct pnfs_layout_range *dst_range,
413 struct inode *inode)
414{
415 struct pnfs_layout_hdr *lo;
416 struct pnfs_layout_range range = {
417 .iomode = IOMODE_ANY,
418 .offset = 0,
419 .length = NFS4_MAX_UINT64,
420 };
421 bool ret = false;
422 LIST_HEAD(head);
423 int err;
424
425 spin_lock(&inode->i_lock);
426 lo = NFS_I(inode)->layout;
427 if (lo && pnfs_layout_is_valid(lo) &&
428 nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
429 /* Is our call using the most recent seqid? If so, bump it */
430 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
431 nfs4_stateid_seqid_inc(dst);
432 ret = true;
433 goto out;
434 }
435 /* Try to update the seqid to the most recent */
436 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
437 if (err != -EBUSY) {
438 dst->seqid = lo->plh_stateid.seqid;
439 *dst_range = range;
440 ret = true;
441 }
442 }
443out:
444 spin_unlock(&inode->i_lock);
445 pnfs_free_lseg_list(&head);
446 return ret;
447}
448
449/*
450 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
451 *
452 * In order to continue using the pnfs_layout_hdr, a full recovery
453 * is required.
454 * Note that caller must hold inode->i_lock.
455 */
456int
457pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
458 struct list_head *lseg_list)
459{
460 struct pnfs_layout_range range = {
461 .iomode = IOMODE_ANY,
462 .offset = 0,
463 .length = NFS4_MAX_UINT64,
464 };
465 struct pnfs_layout_segment *lseg, *next;
466
467 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
468 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
469 pnfs_clear_lseg_state(lseg, lseg_list);
470 pnfs_clear_layoutreturn_info(lo);
471 pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
472 set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
473 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
474 !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
475 pnfs_clear_layoutreturn_waitbit(lo);
476 return !list_empty(&lo->plh_segs);
477}
478
479static int
480pnfs_iomode_to_fail_bit(u32 iomode)
481{
482 return iomode == IOMODE_RW ?
483 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
484}
485
486static void
487pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
488{
489 lo->plh_retry_timestamp = jiffies;
490 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
491 refcount_inc(&lo->plh_refcount);
492}
493
494static void
495pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
496{
497 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
498 refcount_dec(&lo->plh_refcount);
499}
500
501static void
502pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
503{
504 struct inode *inode = lo->plh_inode;
505 struct pnfs_layout_range range = {
506 .iomode = iomode,
507 .offset = 0,
508 .length = NFS4_MAX_UINT64,
509 };
510 LIST_HEAD(head);
511
512 spin_lock(&inode->i_lock);
513 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
514 pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
515 spin_unlock(&inode->i_lock);
516 pnfs_free_lseg_list(&head);
517 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
518 iomode == IOMODE_RW ? "RW" : "READ");
519}
520
521static bool
522pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
523{
524 unsigned long start, end;
525 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
526
527 if (test_bit(fail_bit, &lo->plh_flags) == 0)
528 return false;
529 end = jiffies;
530 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
531 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
532 /* It is time to retry the failed layoutgets */
533 pnfs_layout_clear_fail_bit(lo, fail_bit);
534 return false;
535 }
536 return true;
537}
538
539static void
540pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
541 const struct pnfs_layout_range *range,
542 const nfs4_stateid *stateid)
543{
544 INIT_LIST_HEAD(&lseg->pls_list);
545 INIT_LIST_HEAD(&lseg->pls_lc_list);
546 INIT_LIST_HEAD(&lseg->pls_commits);
547 refcount_set(&lseg->pls_refcount, 1);
548 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
549 lseg->pls_layout = lo;
550 lseg->pls_range = *range;
551 lseg->pls_seq = be32_to_cpu(stateid->seqid);
552}
553
554static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
555{
556 if (lseg != NULL) {
557 struct inode *inode = lseg->pls_layout->plh_inode;
558 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
559 }
560}
561
562static void
563pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
564 struct pnfs_layout_segment *lseg)
565{
566 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
567 list_del_init(&lseg->pls_list);
568 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
569 refcount_dec(&lo->plh_refcount);
570 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
571 return;
572 if (list_empty(&lo->plh_segs) &&
573 !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
574 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
575 if (atomic_read(&lo->plh_outstanding) == 0)
576 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
577 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
578 }
579}
580
581static bool
582pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
583 struct pnfs_layout_segment *lseg)
584{
585 if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
586 pnfs_layout_is_valid(lo)) {
587 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
588 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
589 return true;
590 }
591 return false;
592}
593
594void
595pnfs_put_lseg(struct pnfs_layout_segment *lseg)
596{
597 struct pnfs_layout_hdr *lo;
598 struct inode *inode;
599
600 if (!lseg)
601 return;
602
603 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
604 refcount_read(&lseg->pls_refcount),
605 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
606
607 lo = lseg->pls_layout;
608 inode = lo->plh_inode;
609
610 if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
611 pnfs_get_layout_hdr(lo);
612 pnfs_layout_remove_lseg(lo, lseg);
613 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
614 lseg = NULL;
615 spin_unlock(&inode->i_lock);
616 pnfs_free_lseg(lseg);
617 pnfs_put_layout_hdr(lo);
618 }
619}
620EXPORT_SYMBOL_GPL(pnfs_put_lseg);
621
622/*
623 * is l2 fully contained in l1?
624 * start1 end1
625 * [----------------------------------)
626 * start2 end2
627 * [----------------)
628 */
629static bool
630pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
631 const struct pnfs_layout_range *l2)
632{
633 u64 start1 = l1->offset;
634 u64 end1 = pnfs_end_offset(start1, l1->length);
635 u64 start2 = l2->offset;
636 u64 end2 = pnfs_end_offset(start2, l2->length);
637
638 return (start1 <= start2) && (end1 >= end2);
639}
640
641static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
642 struct list_head *tmp_list)
643{
644 if (!refcount_dec_and_test(&lseg->pls_refcount))
645 return false;
646 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
647 list_add(&lseg->pls_list, tmp_list);
648 return true;
649}
650
651/* Returns 1 if lseg is removed from list, 0 otherwise */
652static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
653 struct list_head *tmp_list)
654{
655 int rv = 0;
656
657 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
658 /* Remove the reference keeping the lseg in the
659 * list. It will now be removed when all
660 * outstanding io is finished.
661 */
662 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
663 refcount_read(&lseg->pls_refcount));
664 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
665 rv = 1;
666 }
667 return rv;
668}
669
670static bool
671pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
672 const struct pnfs_layout_range *recall_range)
673{
674 return (recall_range->iomode == IOMODE_ANY ||
675 lseg_range->iomode == recall_range->iomode) &&
676 pnfs_lseg_range_intersecting(lseg_range, recall_range);
677}
678
679static bool
680pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
681 const struct pnfs_layout_range *recall_range,
682 u32 seq)
683{
684 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
685 return false;
686 if (recall_range == NULL)
687 return true;
688 return pnfs_should_free_range(&lseg->pls_range, recall_range);
689}
690
691/**
692 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
693 * @lo: layout header containing the lsegs
694 * @tmp_list: list head where doomed lsegs should go
695 * @recall_range: optional recall range argument to match (may be NULL)
696 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
697 *
698 * Walk the list of lsegs in the layout header, and tear down any that should
699 * be destroyed. If "recall_range" is specified then the segment must match
700 * that range. If "seq" is non-zero, then only match segments that were handed
701 * out at or before that sequence.
702 *
703 * Returns number of matching invalid lsegs remaining in list after scanning
704 * it and purging them.
705 */
706int
707pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
708 struct list_head *tmp_list,
709 const struct pnfs_layout_range *recall_range,
710 u32 seq)
711{
712 struct pnfs_layout_segment *lseg, *next;
713 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
714 int remaining = 0;
715
716 dprintk("%s:Begin lo %p\n", __func__, lo);
717
718 if (list_empty(&lo->plh_segs))
719 return 0;
720 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
721 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
722 dprintk("%s: freeing lseg %p iomode %d seq %u "
723 "offset %llu length %llu\n", __func__,
724 lseg, lseg->pls_range.iomode, lseg->pls_seq,
725 lseg->pls_range.offset, lseg->pls_range.length);
726 if (mark_lseg_invalid(lseg, tmp_list))
727 continue;
728 remaining++;
729 pnfs_lseg_cancel_io(server, lseg);
730 }
731 dprintk("%s:Return %i\n", __func__, remaining);
732 return remaining;
733}
734
735static void
736pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
737 struct list_head *free_me,
738 const struct pnfs_layout_range *range,
739 u32 seq)
740{
741 struct pnfs_layout_segment *lseg, *next;
742
743 list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
744 if (pnfs_match_lseg_recall(lseg, range, seq))
745 list_move_tail(&lseg->pls_list, free_me);
746 }
747}
748
749/* note free_me must contain lsegs from a single layout_hdr */
750void
751pnfs_free_lseg_list(struct list_head *free_me)
752{
753 struct pnfs_layout_segment *lseg, *tmp;
754
755 if (list_empty(free_me))
756 return;
757
758 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
759 list_del(&lseg->pls_list);
760 pnfs_free_lseg(lseg);
761 }
762}
763
764static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
765{
766 struct pnfs_layout_hdr *lo;
767 LIST_HEAD(tmp_list);
768
769 spin_lock(&nfsi->vfs_inode.i_lock);
770 lo = nfsi->layout;
771 if (lo) {
772 pnfs_get_layout_hdr(lo);
773 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
774 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
775 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
776 spin_unlock(&nfsi->vfs_inode.i_lock);
777 pnfs_free_lseg_list(&tmp_list);
778 nfs_commit_inode(&nfsi->vfs_inode, 0);
779 pnfs_put_layout_hdr(lo);
780 } else
781 spin_unlock(&nfsi->vfs_inode.i_lock);
782 return lo;
783}
784
785void pnfs_destroy_layout(struct nfs_inode *nfsi)
786{
787 __pnfs_destroy_layout(nfsi);
788}
789EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
790
791static bool pnfs_layout_removed(struct nfs_inode *nfsi,
792 struct pnfs_layout_hdr *lo)
793{
794 bool ret;
795
796 spin_lock(&nfsi->vfs_inode.i_lock);
797 ret = nfsi->layout != lo;
798 spin_unlock(&nfsi->vfs_inode.i_lock);
799 return ret;
800}
801
802void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
803{
804 struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
805
806 if (lo)
807 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
808}
809
810static bool
811pnfs_layout_add_bulk_destroy_list(struct inode *inode,
812 struct list_head *layout_list)
813{
814 struct pnfs_layout_hdr *lo;
815 bool ret = false;
816
817 spin_lock(&inode->i_lock);
818 lo = NFS_I(inode)->layout;
819 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
820 pnfs_get_layout_hdr(lo);
821 list_add(&lo->plh_bulk_destroy, layout_list);
822 ret = true;
823 }
824 spin_unlock(&inode->i_lock);
825 return ret;
826}
827
828/* Caller must hold rcu_read_lock and clp->cl_lock */
829static int
830pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
831 struct nfs_server *server,
832 struct list_head *layout_list)
833 __must_hold(&clp->cl_lock)
834 __must_hold(RCU)
835{
836 struct pnfs_layout_hdr *lo, *next;
837 struct inode *inode;
838
839 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
840 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
841 test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
842 !list_empty(&lo->plh_bulk_destroy))
843 continue;
844 /* If the sb is being destroyed, just bail */
845 if (!nfs_sb_active(server->super))
846 break;
847 inode = pnfs_grab_inode_layout_hdr(lo);
848 if (inode != NULL) {
849 if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
850 list_del_rcu(&lo->plh_layouts);
851 if (pnfs_layout_add_bulk_destroy_list(inode,
852 layout_list))
853 continue;
854 rcu_read_unlock();
855 spin_unlock(&clp->cl_lock);
856 iput(inode);
857 } else {
858 rcu_read_unlock();
859 spin_unlock(&clp->cl_lock);
860 }
861 nfs_sb_deactive(server->super);
862 spin_lock(&clp->cl_lock);
863 rcu_read_lock();
864 return -EAGAIN;
865 }
866 return 0;
867}
868
869static int
870pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
871 bool is_bulk_recall)
872{
873 struct pnfs_layout_hdr *lo;
874 struct inode *inode;
875 LIST_HEAD(lseg_list);
876 int ret = 0;
877
878 while (!list_empty(layout_list)) {
879 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
880 plh_bulk_destroy);
881 dprintk("%s freeing layout for inode %lu\n", __func__,
882 lo->plh_inode->i_ino);
883 inode = lo->plh_inode;
884
885 pnfs_layoutcommit_inode(inode, false);
886
887 spin_lock(&inode->i_lock);
888 list_del_init(&lo->plh_bulk_destroy);
889 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
890 if (is_bulk_recall)
891 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
892 ret = -EAGAIN;
893 }
894 spin_unlock(&inode->i_lock);
895 pnfs_free_lseg_list(&lseg_list);
896 /* Free all lsegs that are attached to commit buckets */
897 nfs_commit_inode(inode, 0);
898 pnfs_put_layout_hdr(lo);
899 nfs_iput_and_deactive(inode);
900 }
901 return ret;
902}
903
904int
905pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
906 struct nfs_fsid *fsid,
907 bool is_recall)
908{
909 struct nfs_server *server;
910 LIST_HEAD(layout_list);
911
912 spin_lock(&clp->cl_lock);
913 rcu_read_lock();
914restart:
915 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
916 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
917 continue;
918 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
919 server,
920 &layout_list) != 0)
921 goto restart;
922 }
923 rcu_read_unlock();
924 spin_unlock(&clp->cl_lock);
925
926 if (list_empty(&layout_list))
927 return 0;
928 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
929}
930
931int
932pnfs_destroy_layouts_byclid(struct nfs_client *clp,
933 bool is_recall)
934{
935 struct nfs_server *server;
936 LIST_HEAD(layout_list);
937
938 spin_lock(&clp->cl_lock);
939 rcu_read_lock();
940restart:
941 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
942 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
943 server,
944 &layout_list) != 0)
945 goto restart;
946 }
947 rcu_read_unlock();
948 spin_unlock(&clp->cl_lock);
949
950 if (list_empty(&layout_list))
951 return 0;
952 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
953}
954
955/*
956 * Called by the state manager to remove all layouts established under an
957 * expired lease.
958 */
959void
960pnfs_destroy_all_layouts(struct nfs_client *clp)
961{
962 nfs4_deviceid_mark_client_invalid(clp);
963 nfs4_deviceid_purge_client(clp);
964
965 pnfs_destroy_layouts_byclid(clp, false);
966}
967
968static void
969pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
970{
971 const struct cred *old;
972
973 if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
974 old = xchg(&lo->plh_lc_cred, get_cred(cred));
975 put_cred(old);
976 }
977}
978
979/* update lo->plh_stateid with new if is more recent */
980void
981pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
982 const struct cred *cred, bool update_barrier)
983{
984 u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
985 u32 newseq = be32_to_cpu(new->seqid);
986
987 if (!pnfs_layout_is_valid(lo)) {
988 pnfs_set_layout_cred(lo, cred);
989 nfs4_stateid_copy(&lo->plh_stateid, new);
990 lo->plh_barrier = newseq;
991 pnfs_clear_layoutreturn_info(lo);
992 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
993 return;
994 }
995
996 if (pnfs_seqid_is_newer(newseq, oldseq))
997 nfs4_stateid_copy(&lo->plh_stateid, new);
998
999 if (update_barrier) {
1000 pnfs_barrier_update(lo, newseq);
1001 return;
1002 }
1003 /*
1004 * Because of wraparound, we want to keep the barrier
1005 * "close" to the current seqids. We really only want to
1006 * get here from a layoutget call.
1007 */
1008 if (atomic_read(&lo->plh_outstanding) == 1)
1009 pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1010}
1011
1012static bool
1013pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1014 const nfs4_stateid *stateid)
1015{
1016 u32 seqid = be32_to_cpu(stateid->seqid);
1017
1018 return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1019}
1020
1021/* lget is set to 1 if called from inside send_layoutget call chain */
1022static bool
1023pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1024{
1025 return lo->plh_block_lgets ||
1026 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1027}
1028
1029static struct nfs_server *
1030pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1031{
1032 struct nfs_server *server;
1033
1034 if (inode) {
1035 server = NFS_SERVER(inode);
1036 } else {
1037 struct dentry *parent_dir = dget_parent(ctx->dentry);
1038 server = NFS_SERVER(parent_dir->d_inode);
1039 dput(parent_dir);
1040 }
1041 return server;
1042}
1043
1044static void nfs4_free_pages(struct page **pages, size_t size)
1045{
1046 int i;
1047
1048 if (!pages)
1049 return;
1050
1051 for (i = 0; i < size; i++) {
1052 if (!pages[i])
1053 break;
1054 __free_page(pages[i]);
1055 }
1056 kfree(pages);
1057}
1058
1059static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1060{
1061 struct page **pages;
1062 int i;
1063
1064 pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1065 if (!pages) {
1066 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1067 return NULL;
1068 }
1069
1070 for (i = 0; i < size; i++) {
1071 pages[i] = alloc_page(gfp_flags);
1072 if (!pages[i]) {
1073 dprintk("%s: failed to allocate page\n", __func__);
1074 nfs4_free_pages(pages, i);
1075 return NULL;
1076 }
1077 }
1078
1079 return pages;
1080}
1081
1082static struct nfs4_layoutget *
1083pnfs_alloc_init_layoutget_args(struct inode *ino,
1084 struct nfs_open_context *ctx,
1085 const nfs4_stateid *stateid,
1086 const struct pnfs_layout_range *range,
1087 gfp_t gfp_flags)
1088{
1089 struct nfs_server *server = pnfs_find_server(ino, ctx);
1090 size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1091 size_t max_pages = max_response_pages(server);
1092 struct nfs4_layoutget *lgp;
1093
1094 dprintk("--> %s\n", __func__);
1095
1096 lgp = kzalloc(sizeof(*lgp), gfp_flags);
1097 if (lgp == NULL)
1098 return NULL;
1099
1100 if (max_reply_sz) {
1101 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1102 if (npages < max_pages)
1103 max_pages = npages;
1104 }
1105
1106 lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1107 if (!lgp->args.layout.pages) {
1108 kfree(lgp);
1109 return NULL;
1110 }
1111 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1112 lgp->res.layoutp = &lgp->args.layout;
1113
1114 /* Don't confuse uninitialised result and success */
1115 lgp->res.status = -NFS4ERR_DELAY;
1116
1117 lgp->args.minlength = PAGE_SIZE;
1118 if (lgp->args.minlength > range->length)
1119 lgp->args.minlength = range->length;
1120 if (ino) {
1121 loff_t i_size = i_size_read(ino);
1122
1123 if (range->iomode == IOMODE_READ) {
1124 if (range->offset >= i_size)
1125 lgp->args.minlength = 0;
1126 else if (i_size - range->offset < lgp->args.minlength)
1127 lgp->args.minlength = i_size - range->offset;
1128 }
1129 }
1130 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1131 pnfs_copy_range(&lgp->args.range, range);
1132 lgp->args.type = server->pnfs_curr_ld->id;
1133 lgp->args.inode = ino;
1134 lgp->args.ctx = get_nfs_open_context(ctx);
1135 nfs4_stateid_copy(&lgp->args.stateid, stateid);
1136 lgp->gfp_flags = gfp_flags;
1137 lgp->cred = ctx->cred;
1138 return lgp;
1139}
1140
1141void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1142{
1143 size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1144
1145 nfs4_free_pages(lgp->args.layout.pages, max_pages);
1146 pnfs_put_layout_hdr(lgp->lo);
1147 put_nfs_open_context(lgp->args.ctx);
1148 kfree(lgp);
1149}
1150
1151static void pnfs_clear_layoutcommit(struct inode *inode,
1152 struct list_head *head)
1153{
1154 struct nfs_inode *nfsi = NFS_I(inode);
1155 struct pnfs_layout_segment *lseg, *tmp;
1156
1157 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1158 return;
1159 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1160 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1161 continue;
1162 pnfs_lseg_dec_and_remove_zero(lseg, head);
1163 }
1164}
1165
1166void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1167 const nfs4_stateid *arg_stateid,
1168 const struct pnfs_layout_range *range,
1169 const nfs4_stateid *stateid)
1170{
1171 struct inode *inode = lo->plh_inode;
1172 LIST_HEAD(freeme);
1173
1174 spin_lock(&inode->i_lock);
1175 if (!pnfs_layout_is_valid(lo) ||
1176 !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1177 goto out_unlock;
1178 if (stateid) {
1179 u32 seq = be32_to_cpu(arg_stateid->seqid);
1180
1181 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1182 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1183 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1184 } else
1185 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1186out_unlock:
1187 pnfs_clear_layoutreturn_waitbit(lo);
1188 spin_unlock(&inode->i_lock);
1189 pnfs_free_lseg_list(&freeme);
1190
1191}
1192
1193static bool
1194pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1195 nfs4_stateid *stateid,
1196 const struct cred **cred,
1197 enum pnfs_iomode *iomode)
1198{
1199 /* Serialise LAYOUTGET/LAYOUTRETURN */
1200 if (atomic_read(&lo->plh_outstanding) != 0)
1201 return false;
1202 if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1203 return false;
1204 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1205 pnfs_get_layout_hdr(lo);
1206 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1207 *cred = get_cred(lo->plh_lc_cred);
1208 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1209 if (lo->plh_return_seq != 0)
1210 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1211 if (iomode != NULL)
1212 *iomode = lo->plh_return_iomode;
1213 pnfs_clear_layoutreturn_info(lo);
1214 } else if (iomode != NULL)
1215 *iomode = IOMODE_ANY;
1216 pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1217 return true;
1218}
1219
1220static void
1221pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1222 struct pnfs_layout_hdr *lo,
1223 const nfs4_stateid *stateid,
1224 enum pnfs_iomode iomode)
1225{
1226 struct inode *inode = lo->plh_inode;
1227
1228 args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1229 args->inode = inode;
1230 args->range.iomode = iomode;
1231 args->range.offset = 0;
1232 args->range.length = NFS4_MAX_UINT64;
1233 args->layout = lo;
1234 nfs4_stateid_copy(&args->stateid, stateid);
1235}
1236
1237static int
1238pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1239 const nfs4_stateid *stateid,
1240 const struct cred **pcred,
1241 enum pnfs_iomode iomode,
1242 bool sync)
1243{
1244 struct inode *ino = lo->plh_inode;
1245 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1246 struct nfs4_layoutreturn *lrp;
1247 const struct cred *cred = *pcred;
1248 int status = 0;
1249
1250 *pcred = NULL;
1251 lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1252 if (unlikely(lrp == NULL)) {
1253 status = -ENOMEM;
1254 spin_lock(&ino->i_lock);
1255 pnfs_clear_layoutreturn_waitbit(lo);
1256 spin_unlock(&ino->i_lock);
1257 put_cred(cred);
1258 pnfs_put_layout_hdr(lo);
1259 goto out;
1260 }
1261
1262 pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1263 lrp->args.ld_private = &lrp->ld_private;
1264 lrp->clp = NFS_SERVER(ino)->nfs_client;
1265 lrp->cred = cred;
1266 if (ld->prepare_layoutreturn)
1267 ld->prepare_layoutreturn(&lrp->args);
1268
1269 status = nfs4_proc_layoutreturn(lrp, sync);
1270out:
1271 dprintk("<-- %s status: %d\n", __func__, status);
1272 return status;
1273}
1274
1275static bool
1276pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1277 enum pnfs_iomode iomode,
1278 u32 seq)
1279{
1280 struct pnfs_layout_range recall_range = {
1281 .length = NFS4_MAX_UINT64,
1282 .iomode = iomode,
1283 };
1284 return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1285 &recall_range, seq) != -EBUSY;
1286}
1287
1288/* Return true if layoutreturn is needed */
1289static bool
1290pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1291{
1292 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1293 return false;
1294 return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1295 lo->plh_return_seq);
1296}
1297
1298static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1299{
1300 struct inode *inode= lo->plh_inode;
1301
1302 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1303 return;
1304 spin_lock(&inode->i_lock);
1305 if (pnfs_layout_need_return(lo)) {
1306 const struct cred *cred;
1307 nfs4_stateid stateid;
1308 enum pnfs_iomode iomode;
1309 bool send;
1310
1311 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1312 spin_unlock(&inode->i_lock);
1313 if (send) {
1314 /* Send an async layoutreturn so we dont deadlock */
1315 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1316 }
1317 } else
1318 spin_unlock(&inode->i_lock);
1319}
1320
1321/*
1322 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1323 * when the layout segment list is empty.
1324 *
1325 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1326 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1327 * deviceid is marked invalid.
1328 */
1329int
1330_pnfs_return_layout(struct inode *ino)
1331{
1332 struct pnfs_layout_hdr *lo = NULL;
1333 struct nfs_inode *nfsi = NFS_I(ino);
1334 struct pnfs_layout_range range = {
1335 .iomode = IOMODE_ANY,
1336 .offset = 0,
1337 .length = NFS4_MAX_UINT64,
1338 };
1339 LIST_HEAD(tmp_list);
1340 const struct cred *cred;
1341 nfs4_stateid stateid;
1342 int status = 0;
1343 bool send, valid_layout;
1344
1345 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1346
1347 spin_lock(&ino->i_lock);
1348 lo = nfsi->layout;
1349 if (!lo) {
1350 spin_unlock(&ino->i_lock);
1351 dprintk("NFS: %s no layout to return\n", __func__);
1352 goto out;
1353 }
1354 /* Reference matched in nfs4_layoutreturn_release */
1355 pnfs_get_layout_hdr(lo);
1356 /* Is there an outstanding layoutreturn ? */
1357 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1358 spin_unlock(&ino->i_lock);
1359 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1360 TASK_UNINTERRUPTIBLE))
1361 goto out_put_layout_hdr;
1362 spin_lock(&ino->i_lock);
1363 }
1364 valid_layout = pnfs_layout_is_valid(lo);
1365 pnfs_clear_layoutcommit(ino, &tmp_list);
1366 pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1367
1368 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1369 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1370
1371 /* Don't send a LAYOUTRETURN if list was initially empty */
1372 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1373 !valid_layout) {
1374 spin_unlock(&ino->i_lock);
1375 dprintk("NFS: %s no layout segments to return\n", __func__);
1376 goto out_wait_layoutreturn;
1377 }
1378
1379 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1380 spin_unlock(&ino->i_lock);
1381 if (send)
1382 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1383out_wait_layoutreturn:
1384 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1385out_put_layout_hdr:
1386 pnfs_free_lseg_list(&tmp_list);
1387 pnfs_put_layout_hdr(lo);
1388out:
1389 dprintk("<-- %s status: %d\n", __func__, status);
1390 return status;
1391}
1392
1393int
1394pnfs_commit_and_return_layout(struct inode *inode)
1395{
1396 struct pnfs_layout_hdr *lo;
1397 int ret;
1398
1399 spin_lock(&inode->i_lock);
1400 lo = NFS_I(inode)->layout;
1401 if (lo == NULL) {
1402 spin_unlock(&inode->i_lock);
1403 return 0;
1404 }
1405 pnfs_get_layout_hdr(lo);
1406 /* Block new layoutgets and read/write to ds */
1407 lo->plh_block_lgets++;
1408 spin_unlock(&inode->i_lock);
1409 filemap_fdatawait(inode->i_mapping);
1410 ret = pnfs_layoutcommit_inode(inode, true);
1411 if (ret == 0)
1412 ret = _pnfs_return_layout(inode);
1413 spin_lock(&inode->i_lock);
1414 lo->plh_block_lgets--;
1415 spin_unlock(&inode->i_lock);
1416 pnfs_put_layout_hdr(lo);
1417 return ret;
1418}
1419
1420bool pnfs_roc(struct inode *ino,
1421 struct nfs4_layoutreturn_args *args,
1422 struct nfs4_layoutreturn_res *res,
1423 const struct cred *cred)
1424{
1425 struct nfs_inode *nfsi = NFS_I(ino);
1426 struct nfs_open_context *ctx;
1427 struct nfs4_state *state;
1428 struct pnfs_layout_hdr *lo;
1429 struct pnfs_layout_segment *lseg, *next;
1430 const struct cred *lc_cred;
1431 nfs4_stateid stateid;
1432 enum pnfs_iomode iomode = 0;
1433 bool layoutreturn = false, roc = false;
1434 bool skip_read = false;
1435
1436 if (!nfs_have_layout(ino))
1437 return false;
1438retry:
1439 rcu_read_lock();
1440 spin_lock(&ino->i_lock);
1441 lo = nfsi->layout;
1442 if (!lo || !pnfs_layout_is_valid(lo) ||
1443 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1444 lo = NULL;
1445 goto out_noroc;
1446 }
1447 pnfs_get_layout_hdr(lo);
1448 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1449 spin_unlock(&ino->i_lock);
1450 rcu_read_unlock();
1451 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1452 TASK_UNINTERRUPTIBLE);
1453 pnfs_put_layout_hdr(lo);
1454 goto retry;
1455 }
1456
1457 /* no roc if we hold a delegation */
1458 if (nfs4_check_delegation(ino, FMODE_READ)) {
1459 if (nfs4_check_delegation(ino, FMODE_WRITE))
1460 goto out_noroc;
1461 skip_read = true;
1462 }
1463
1464 list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1465 state = ctx->state;
1466 if (state == NULL)
1467 continue;
1468 /* Don't return layout if there is open file state */
1469 if (state->state & FMODE_WRITE)
1470 goto out_noroc;
1471 if (state->state & FMODE_READ)
1472 skip_read = true;
1473 }
1474
1475
1476 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1477 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1478 continue;
1479 /* If we are sending layoutreturn, invalidate all valid lsegs */
1480 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1481 continue;
1482 /*
1483 * Note: mark lseg for return so pnfs_layout_remove_lseg
1484 * doesn't invalidate the layout for us.
1485 */
1486 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1487 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1488 continue;
1489 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1490 }
1491
1492 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1493 goto out_noroc;
1494
1495 /* ROC in two conditions:
1496 * 1. there are ROC lsegs
1497 * 2. we don't send layoutreturn
1498 */
1499 /* lo ref dropped in pnfs_roc_release() */
1500 layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1501 /* If the creds don't match, we can't compound the layoutreturn */
1502 if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1503 goto out_noroc;
1504
1505 roc = layoutreturn;
1506 pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1507 res->lrs_present = 0;
1508 layoutreturn = false;
1509 put_cred(lc_cred);
1510
1511out_noroc:
1512 spin_unlock(&ino->i_lock);
1513 rcu_read_unlock();
1514 pnfs_layoutcommit_inode(ino, true);
1515 if (roc) {
1516 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1517 if (ld->prepare_layoutreturn)
1518 ld->prepare_layoutreturn(args);
1519 pnfs_put_layout_hdr(lo);
1520 return true;
1521 }
1522 if (layoutreturn)
1523 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1524 pnfs_put_layout_hdr(lo);
1525 return false;
1526}
1527
1528int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1529 struct nfs4_layoutreturn_res **respp, int *ret)
1530{
1531 struct nfs4_layoutreturn_args *arg = *argpp;
1532 int retval = -EAGAIN;
1533
1534 if (!arg)
1535 return 0;
1536 /* Handle Layoutreturn errors */
1537 switch (*ret) {
1538 case 0:
1539 retval = 0;
1540 break;
1541 case -NFS4ERR_NOMATCHING_LAYOUT:
1542 /* Was there an RPC level error? If not, retry */
1543 if (task->tk_rpc_status == 0)
1544 break;
1545 /* If the call was not sent, let caller handle it */
1546 if (!RPC_WAS_SENT(task))
1547 return 0;
1548 /*
1549 * Otherwise, assume the call succeeded and
1550 * that we need to release the layout
1551 */
1552 *ret = 0;
1553 (*respp)->lrs_present = 0;
1554 retval = 0;
1555 break;
1556 case -NFS4ERR_DELAY:
1557 /* Let the caller handle the retry */
1558 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1559 return 0;
1560 case -NFS4ERR_OLD_STATEID:
1561 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1562 &arg->range, arg->inode))
1563 break;
1564 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1565 return -EAGAIN;
1566 }
1567 *argpp = NULL;
1568 *respp = NULL;
1569 return retval;
1570}
1571
1572void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1573 struct nfs4_layoutreturn_res *res,
1574 int ret)
1575{
1576 struct pnfs_layout_hdr *lo = args->layout;
1577 struct inode *inode = args->inode;
1578 const nfs4_stateid *res_stateid = NULL;
1579 struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1580
1581 switch (ret) {
1582 case -NFS4ERR_NOMATCHING_LAYOUT:
1583 spin_lock(&inode->i_lock);
1584 if (pnfs_layout_is_valid(lo) &&
1585 nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1586 pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1587 pnfs_clear_layoutreturn_waitbit(lo);
1588 spin_unlock(&inode->i_lock);
1589 break;
1590 case 0:
1591 if (res->lrs_present)
1592 res_stateid = &res->stateid;
1593 fallthrough;
1594 default:
1595 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1596 res_stateid);
1597 }
1598 trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1599 if (ld_private && ld_private->ops && ld_private->ops->free)
1600 ld_private->ops->free(ld_private);
1601 pnfs_put_layout_hdr(lo);
1602}
1603
1604bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1605{
1606 struct nfs_inode *nfsi = NFS_I(ino);
1607 struct pnfs_layout_hdr *lo;
1608 bool sleep = false;
1609
1610 /* we might not have grabbed lo reference. so need to check under
1611 * i_lock */
1612 spin_lock(&ino->i_lock);
1613 lo = nfsi->layout;
1614 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1615 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1616 sleep = true;
1617 }
1618 spin_unlock(&ino->i_lock);
1619 return sleep;
1620}
1621
1622/*
1623 * Compare two layout segments for sorting into layout cache.
1624 * We want to preferentially return RW over RO layouts, so ensure those
1625 * are seen first.
1626 */
1627static s64
1628pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1629 const struct pnfs_layout_range *l2)
1630{
1631 s64 d;
1632
1633 /* high offset > low offset */
1634 d = l1->offset - l2->offset;
1635 if (d)
1636 return d;
1637
1638 /* short length > long length */
1639 d = l2->length - l1->length;
1640 if (d)
1641 return d;
1642
1643 /* read > read/write */
1644 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1645}
1646
1647static bool
1648pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1649 const struct pnfs_layout_range *l2)
1650{
1651 return pnfs_lseg_range_cmp(l1, l2) > 0;
1652}
1653
1654static bool
1655pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1656 struct pnfs_layout_segment *old)
1657{
1658 return false;
1659}
1660
1661void
1662pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1663 struct pnfs_layout_segment *lseg,
1664 bool (*is_after)(const struct pnfs_layout_range *,
1665 const struct pnfs_layout_range *),
1666 bool (*do_merge)(struct pnfs_layout_segment *,
1667 struct pnfs_layout_segment *),
1668 struct list_head *free_me)
1669{
1670 struct pnfs_layout_segment *lp, *tmp;
1671
1672 dprintk("%s:Begin\n", __func__);
1673
1674 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1675 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1676 continue;
1677 if (do_merge(lseg, lp)) {
1678 mark_lseg_invalid(lp, free_me);
1679 continue;
1680 }
1681 if (is_after(&lseg->pls_range, &lp->pls_range))
1682 continue;
1683 list_add_tail(&lseg->pls_list, &lp->pls_list);
1684 dprintk("%s: inserted lseg %p "
1685 "iomode %d offset %llu length %llu before "
1686 "lp %p iomode %d offset %llu length %llu\n",
1687 __func__, lseg, lseg->pls_range.iomode,
1688 lseg->pls_range.offset, lseg->pls_range.length,
1689 lp, lp->pls_range.iomode, lp->pls_range.offset,
1690 lp->pls_range.length);
1691 goto out;
1692 }
1693 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1694 dprintk("%s: inserted lseg %p "
1695 "iomode %d offset %llu length %llu at tail\n",
1696 __func__, lseg, lseg->pls_range.iomode,
1697 lseg->pls_range.offset, lseg->pls_range.length);
1698out:
1699 pnfs_get_layout_hdr(lo);
1700
1701 dprintk("%s:Return\n", __func__);
1702}
1703EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1704
1705static void
1706pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1707 struct pnfs_layout_segment *lseg,
1708 struct list_head *free_me)
1709{
1710 struct inode *inode = lo->plh_inode;
1711 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1712
1713 if (ld->add_lseg != NULL)
1714 ld->add_lseg(lo, lseg, free_me);
1715 else
1716 pnfs_generic_layout_insert_lseg(lo, lseg,
1717 pnfs_lseg_range_is_after,
1718 pnfs_lseg_no_merge,
1719 free_me);
1720}
1721
1722static struct pnfs_layout_hdr *
1723alloc_init_layout_hdr(struct inode *ino,
1724 struct nfs_open_context *ctx,
1725 gfp_t gfp_flags)
1726{
1727 struct pnfs_layout_hdr *lo;
1728
1729 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1730 if (!lo)
1731 return NULL;
1732 refcount_set(&lo->plh_refcount, 1);
1733 INIT_LIST_HEAD(&lo->plh_layouts);
1734 INIT_LIST_HEAD(&lo->plh_segs);
1735 INIT_LIST_HEAD(&lo->plh_return_segs);
1736 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1737 lo->plh_inode = ino;
1738 lo->plh_lc_cred = get_cred(ctx->cred);
1739 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1740 return lo;
1741}
1742
1743static struct pnfs_layout_hdr *
1744pnfs_find_alloc_layout(struct inode *ino,
1745 struct nfs_open_context *ctx,
1746 gfp_t gfp_flags)
1747 __releases(&ino->i_lock)
1748 __acquires(&ino->i_lock)
1749{
1750 struct nfs_inode *nfsi = NFS_I(ino);
1751 struct pnfs_layout_hdr *new = NULL;
1752
1753 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1754
1755 if (nfsi->layout != NULL)
1756 goto out_existing;
1757 spin_unlock(&ino->i_lock);
1758 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1759 spin_lock(&ino->i_lock);
1760
1761 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1762 nfsi->layout = new;
1763 return new;
1764 } else if (new != NULL)
1765 pnfs_free_layout_hdr(new);
1766out_existing:
1767 pnfs_get_layout_hdr(nfsi->layout);
1768 return nfsi->layout;
1769}
1770
1771/*
1772 * iomode matching rules:
1773 * iomode lseg strict match
1774 * iomode
1775 * ----- ----- ------ -----
1776 * ANY READ N/A true
1777 * ANY RW N/A true
1778 * RW READ N/A false
1779 * RW RW N/A true
1780 * READ READ N/A true
1781 * READ RW true false
1782 * READ RW false true
1783 */
1784static bool
1785pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1786 const struct pnfs_layout_range *range,
1787 bool strict_iomode)
1788{
1789 struct pnfs_layout_range range1;
1790
1791 if ((range->iomode == IOMODE_RW &&
1792 ls_range->iomode != IOMODE_RW) ||
1793 (range->iomode != ls_range->iomode &&
1794 strict_iomode) ||
1795 !pnfs_lseg_range_intersecting(ls_range, range))
1796 return false;
1797
1798 /* range1 covers only the first byte in the range */
1799 range1 = *range;
1800 range1.length = 1;
1801 return pnfs_lseg_range_contained(ls_range, &range1);
1802}
1803
1804/*
1805 * lookup range in layout
1806 */
1807static struct pnfs_layout_segment *
1808pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1809 struct pnfs_layout_range *range,
1810 bool strict_iomode)
1811{
1812 struct pnfs_layout_segment *lseg, *ret = NULL;
1813
1814 dprintk("%s:Begin\n", __func__);
1815
1816 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1817 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1818 pnfs_lseg_range_match(&lseg->pls_range, range,
1819 strict_iomode)) {
1820 ret = pnfs_get_lseg(lseg);
1821 break;
1822 }
1823 }
1824
1825 dprintk("%s:Return lseg %p ref %d\n",
1826 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1827 return ret;
1828}
1829
1830/*
1831 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1832 * to the MDS or over pNFS
1833 *
1834 * The nfs_inode read_io and write_io fields are cumulative counters reset
1835 * when there are no layout segments. Note that in pnfs_update_layout iomode
1836 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1837 * WRITE request.
1838 *
1839 * A return of true means use MDS I/O.
1840 *
1841 * From rfc 5661:
1842 * If a file's size is smaller than the file size threshold, data accesses
1843 * SHOULD be sent to the metadata server. If an I/O request has a length that
1844 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1845 * server. If both file size and I/O size are provided, the client SHOULD
1846 * reach or exceed both thresholds before sending its read or write
1847 * requests to the data server.
1848 */
1849static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1850 struct inode *ino, int iomode)
1851{
1852 struct nfs4_threshold *t = ctx->mdsthreshold;
1853 struct nfs_inode *nfsi = NFS_I(ino);
1854 loff_t fsize = i_size_read(ino);
1855 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1856
1857 if (t == NULL)
1858 return ret;
1859
1860 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1861 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1862
1863 switch (iomode) {
1864 case IOMODE_READ:
1865 if (t->bm & THRESHOLD_RD) {
1866 dprintk("%s fsize %llu\n", __func__, fsize);
1867 size_set = true;
1868 if (fsize < t->rd_sz)
1869 size = true;
1870 }
1871 if (t->bm & THRESHOLD_RD_IO) {
1872 dprintk("%s nfsi->read_io %llu\n", __func__,
1873 nfsi->read_io);
1874 io_set = true;
1875 if (nfsi->read_io < t->rd_io_sz)
1876 io = true;
1877 }
1878 break;
1879 case IOMODE_RW:
1880 if (t->bm & THRESHOLD_WR) {
1881 dprintk("%s fsize %llu\n", __func__, fsize);
1882 size_set = true;
1883 if (fsize < t->wr_sz)
1884 size = true;
1885 }
1886 if (t->bm & THRESHOLD_WR_IO) {
1887 dprintk("%s nfsi->write_io %llu\n", __func__,
1888 nfsi->write_io);
1889 io_set = true;
1890 if (nfsi->write_io < t->wr_io_sz)
1891 io = true;
1892 }
1893 break;
1894 }
1895 if (size_set && io_set) {
1896 if (size && io)
1897 ret = true;
1898 } else if (size || io)
1899 ret = true;
1900
1901 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1902 return ret;
1903}
1904
1905static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1906{
1907 /*
1908 * send layoutcommit as it can hold up layoutreturn due to lseg
1909 * reference
1910 */
1911 pnfs_layoutcommit_inode(lo->plh_inode, false);
1912 return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1913 nfs_wait_bit_killable,
1914 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
1915}
1916
1917static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1918{
1919 atomic_inc(&lo->plh_outstanding);
1920}
1921
1922static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1923{
1924 if (atomic_dec_and_test(&lo->plh_outstanding) &&
1925 test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1926 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1927}
1928
1929static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1930{
1931 return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1932}
1933
1934static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1935{
1936 unsigned long *bitlock = &lo->plh_flags;
1937
1938 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1939 smp_mb__after_atomic();
1940 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1941}
1942
1943static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1944 struct nfs_server *server)
1945{
1946 if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1947 struct nfs_client *clp = server->nfs_client;
1948
1949 /* The lo must be on the clp list if there is any
1950 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1951 */
1952 spin_lock(&clp->cl_lock);
1953 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1954 spin_unlock(&clp->cl_lock);
1955 }
1956}
1957
1958/*
1959 * Layout segment is retreived from the server if not cached.
1960 * The appropriate layout segment is referenced and returned to the caller.
1961 */
1962struct pnfs_layout_segment *
1963pnfs_update_layout(struct inode *ino,
1964 struct nfs_open_context *ctx,
1965 loff_t pos,
1966 u64 count,
1967 enum pnfs_iomode iomode,
1968 bool strict_iomode,
1969 gfp_t gfp_flags)
1970{
1971 struct pnfs_layout_range arg = {
1972 .iomode = iomode,
1973 .offset = pos,
1974 .length = count,
1975 };
1976 unsigned pg_offset;
1977 struct nfs_server *server = NFS_SERVER(ino);
1978 struct nfs_client *clp = server->nfs_client;
1979 struct pnfs_layout_hdr *lo = NULL;
1980 struct pnfs_layout_segment *lseg = NULL;
1981 struct nfs4_layoutget *lgp;
1982 nfs4_stateid stateid;
1983 long timeout = 0;
1984 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1985 bool first;
1986
1987 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1988 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1989 PNFS_UPDATE_LAYOUT_NO_PNFS);
1990 goto out;
1991 }
1992
1993 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1994 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1995 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1996 goto out;
1997 }
1998
1999lookup_again:
2000 lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2001 if (IS_ERR(lseg))
2002 goto out;
2003 first = false;
2004 spin_lock(&ino->i_lock);
2005 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2006 if (lo == NULL) {
2007 spin_unlock(&ino->i_lock);
2008 lseg = ERR_PTR(-ENOMEM);
2009 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2010 PNFS_UPDATE_LAYOUT_NOMEM);
2011 goto out;
2012 }
2013
2014 /* Do we even need to bother with this? */
2015 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2016 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2017 PNFS_UPDATE_LAYOUT_BULK_RECALL);
2018 dprintk("%s matches recall, use MDS\n", __func__);
2019 goto out_unlock;
2020 }
2021
2022 /* if LAYOUTGET already failed once we don't try again */
2023 if (pnfs_layout_io_test_failed(lo, iomode)) {
2024 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2025 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2026 goto out_unlock;
2027 }
2028
2029 /*
2030 * If the layout segment list is empty, but there are outstanding
2031 * layoutget calls, then they might be subject to a layoutrecall.
2032 */
2033 if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2034 atomic_read(&lo->plh_outstanding) != 0) {
2035 spin_unlock(&ino->i_lock);
2036 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2037 TASK_KILLABLE));
2038 if (IS_ERR(lseg))
2039 goto out_put_layout_hdr;
2040 pnfs_put_layout_hdr(lo);
2041 goto lookup_again;
2042 }
2043
2044 /*
2045 * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2046 * for LAYOUTRETURN.
2047 */
2048 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2049 spin_unlock(&ino->i_lock);
2050 dprintk("%s wait for layoutreturn\n", __func__);
2051 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2052 if (!IS_ERR(lseg)) {
2053 pnfs_put_layout_hdr(lo);
2054 dprintk("%s retrying\n", __func__);
2055 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2056 lseg,
2057 PNFS_UPDATE_LAYOUT_RETRY);
2058 goto lookup_again;
2059 }
2060 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2061 PNFS_UPDATE_LAYOUT_RETURN);
2062 goto out_put_layout_hdr;
2063 }
2064
2065 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2066 if (lseg) {
2067 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2068 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2069 goto out_unlock;
2070 }
2071
2072 /*
2073 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2074 * stateid, or it has been invalidated, then we must use the open
2075 * stateid.
2076 */
2077 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2078 int status;
2079
2080 /*
2081 * The first layoutget for the file. Need to serialize per
2082 * RFC 5661 Errata 3208.
2083 */
2084 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2085 &lo->plh_flags)) {
2086 spin_unlock(&ino->i_lock);
2087 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2088 NFS_LAYOUT_FIRST_LAYOUTGET,
2089 TASK_KILLABLE));
2090 if (IS_ERR(lseg))
2091 goto out_put_layout_hdr;
2092 pnfs_put_layout_hdr(lo);
2093 dprintk("%s retrying\n", __func__);
2094 goto lookup_again;
2095 }
2096
2097 spin_unlock(&ino->i_lock);
2098 first = true;
2099 status = nfs4_select_rw_stateid(ctx->state,
2100 iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2101 NULL, &stateid, NULL);
2102 if (status != 0) {
2103 lseg = ERR_PTR(status);
2104 trace_pnfs_update_layout(ino, pos, count,
2105 iomode, lo, lseg,
2106 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2107 nfs4_schedule_stateid_recovery(server, ctx->state);
2108 pnfs_clear_first_layoutget(lo);
2109 pnfs_put_layout_hdr(lo);
2110 goto lookup_again;
2111 }
2112 spin_lock(&ino->i_lock);
2113 } else {
2114 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2115 }
2116
2117 if (pnfs_layoutgets_blocked(lo)) {
2118 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2119 PNFS_UPDATE_LAYOUT_BLOCKED);
2120 goto out_unlock;
2121 }
2122 nfs_layoutget_begin(lo);
2123 spin_unlock(&ino->i_lock);
2124
2125 _add_to_server_list(lo, server);
2126
2127 pg_offset = arg.offset & ~PAGE_MASK;
2128 if (pg_offset) {
2129 arg.offset -= pg_offset;
2130 arg.length += pg_offset;
2131 }
2132 if (arg.length != NFS4_MAX_UINT64)
2133 arg.length = PAGE_ALIGN(arg.length);
2134
2135 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2136 if (!lgp) {
2137 lseg = ERR_PTR(-ENOMEM);
2138 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2139 PNFS_UPDATE_LAYOUT_NOMEM);
2140 nfs_layoutget_end(lo);
2141 goto out_put_layout_hdr;
2142 }
2143
2144 lgp->lo = lo;
2145 pnfs_get_layout_hdr(lo);
2146
2147 lseg = nfs4_proc_layoutget(lgp, &timeout);
2148 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2149 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2150 nfs_layoutget_end(lo);
2151 if (IS_ERR(lseg)) {
2152 switch(PTR_ERR(lseg)) {
2153 case -EBUSY:
2154 if (time_after(jiffies, giveup))
2155 lseg = NULL;
2156 break;
2157 case -ERECALLCONFLICT:
2158 case -EAGAIN:
2159 break;
2160 case -ENODATA:
2161 /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2162 pnfs_layout_set_fail_bit(
2163 lo, pnfs_iomode_to_fail_bit(iomode));
2164 lseg = NULL;
2165 goto out_put_layout_hdr;
2166 default:
2167 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2168 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2169 lseg = NULL;
2170 }
2171 goto out_put_layout_hdr;
2172 }
2173 if (lseg) {
2174 if (first)
2175 pnfs_clear_first_layoutget(lo);
2176 trace_pnfs_update_layout(ino, pos, count,
2177 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2178 pnfs_put_layout_hdr(lo);
2179 goto lookup_again;
2180 }
2181 } else {
2182 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2183 }
2184
2185out_put_layout_hdr:
2186 if (first)
2187 pnfs_clear_first_layoutget(lo);
2188 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2189 PNFS_UPDATE_LAYOUT_EXIT);
2190 pnfs_put_layout_hdr(lo);
2191out:
2192 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2193 "(%s, offset: %llu, length: %llu)\n",
2194 __func__, ino->i_sb->s_id,
2195 (unsigned long long)NFS_FILEID(ino),
2196 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2197 iomode==IOMODE_RW ? "read/write" : "read-only",
2198 (unsigned long long)pos,
2199 (unsigned long long)count);
2200 return lseg;
2201out_unlock:
2202 spin_unlock(&ino->i_lock);
2203 goto out_put_layout_hdr;
2204}
2205EXPORT_SYMBOL_GPL(pnfs_update_layout);
2206
2207static bool
2208pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2209{
2210 switch (range->iomode) {
2211 case IOMODE_READ:
2212 case IOMODE_RW:
2213 break;
2214 default:
2215 return false;
2216 }
2217 if (range->offset == NFS4_MAX_UINT64)
2218 return false;
2219 if (range->length == 0)
2220 return false;
2221 if (range->length != NFS4_MAX_UINT64 &&
2222 range->length > NFS4_MAX_UINT64 - range->offset)
2223 return false;
2224 return true;
2225}
2226
2227static struct pnfs_layout_hdr *
2228_pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2229{
2230 struct pnfs_layout_hdr *lo;
2231
2232 spin_lock(&ino->i_lock);
2233 lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2234 if (!lo)
2235 goto out_unlock;
2236 if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2237 goto out_unlock;
2238 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2239 goto out_unlock;
2240 if (pnfs_layoutgets_blocked(lo))
2241 goto out_unlock;
2242 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2243 goto out_unlock;
2244 nfs_layoutget_begin(lo);
2245 spin_unlock(&ino->i_lock);
2246 _add_to_server_list(lo, NFS_SERVER(ino));
2247 return lo;
2248
2249out_unlock:
2250 spin_unlock(&ino->i_lock);
2251 pnfs_put_layout_hdr(lo);
2252 return NULL;
2253}
2254
2255static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2256 struct nfs_open_context *ctx)
2257{
2258 struct inode *ino = data->dentry->d_inode;
2259 struct pnfs_layout_range rng = {
2260 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2261 IOMODE_RW: IOMODE_READ,
2262 .offset = 0,
2263 .length = NFS4_MAX_UINT64,
2264 };
2265 struct nfs4_layoutget *lgp;
2266 struct pnfs_layout_hdr *lo;
2267
2268 /* Heuristic: don't send layoutget if we have cached data */
2269 if (rng.iomode == IOMODE_READ &&
2270 (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2271 return;
2272
2273 lo = _pnfs_grab_empty_layout(ino, ctx);
2274 if (!lo)
2275 return;
2276 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, ¤t_stateid, &rng,
2277 nfs_io_gfp_mask());
2278 if (!lgp) {
2279 pnfs_clear_first_layoutget(lo);
2280 nfs_layoutget_end(lo);
2281 pnfs_put_layout_hdr(lo);
2282 return;
2283 }
2284 lgp->lo = lo;
2285 data->lgp = lgp;
2286 data->o_arg.lg_args = &lgp->args;
2287 data->o_res.lg_res = &lgp->res;
2288}
2289
2290static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2291 struct nfs_open_context *ctx)
2292{
2293 struct inode *ino = data->dentry->d_inode;
2294 struct pnfs_layout_range rng = {
2295 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2296 IOMODE_RW: IOMODE_READ,
2297 .offset = 0,
2298 .length = NFS4_MAX_UINT64,
2299 };
2300 struct nfs4_layoutget *lgp;
2301
2302 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, ¤t_stateid, &rng,
2303 nfs_io_gfp_mask());
2304 if (!lgp)
2305 return;
2306 data->lgp = lgp;
2307 data->o_arg.lg_args = &lgp->args;
2308 data->o_res.lg_res = &lgp->res;
2309}
2310
2311void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2312 struct nfs_open_context *ctx)
2313{
2314 struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2315
2316 if (!(pnfs_enabled_sb(server) &&
2317 server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2318 return;
2319 /* Could check on max_ops, but currently hardcoded high enough */
2320 if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2321 return;
2322 if (data->lgp)
2323 return;
2324 if (data->state)
2325 _lgopen_prepare_attached(data, ctx);
2326 else
2327 _lgopen_prepare_floating(data, ctx);
2328}
2329
2330void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2331 struct nfs_open_context *ctx)
2332{
2333 struct pnfs_layout_hdr *lo;
2334 struct pnfs_layout_segment *lseg;
2335 struct nfs_server *srv = NFS_SERVER(ino);
2336 u32 iomode;
2337
2338 if (!lgp)
2339 return;
2340 dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2341 if (lgp->res.status) {
2342 switch (lgp->res.status) {
2343 default:
2344 break;
2345 /*
2346 * Halt lgopen attempts if the server doesn't recognise
2347 * the "current stateid" value, the layout type, or the
2348 * layoutget operation as being valid.
2349 * Also if it complains about too many ops in the compound
2350 * or of the request/reply being too big.
2351 */
2352 case -NFS4ERR_BAD_STATEID:
2353 case -NFS4ERR_NOTSUPP:
2354 case -NFS4ERR_REP_TOO_BIG:
2355 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2356 case -NFS4ERR_REQ_TOO_BIG:
2357 case -NFS4ERR_TOO_MANY_OPS:
2358 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2359 srv->caps &= ~NFS_CAP_LGOPEN;
2360 }
2361 return;
2362 }
2363 if (!lgp->lo) {
2364 lo = _pnfs_grab_empty_layout(ino, ctx);
2365 if (!lo)
2366 return;
2367 lgp->lo = lo;
2368 } else
2369 lo = lgp->lo;
2370
2371 lseg = pnfs_layout_process(lgp);
2372 if (!IS_ERR(lseg)) {
2373 iomode = lgp->args.range.iomode;
2374 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2375 pnfs_put_lseg(lseg);
2376 }
2377}
2378
2379void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2380{
2381 if (lgp != NULL) {
2382 if (lgp->lo) {
2383 pnfs_clear_first_layoutget(lgp->lo);
2384 nfs_layoutget_end(lgp->lo);
2385 }
2386 pnfs_layoutget_free(lgp);
2387 }
2388}
2389
2390struct pnfs_layout_segment *
2391pnfs_layout_process(struct nfs4_layoutget *lgp)
2392{
2393 struct pnfs_layout_hdr *lo = lgp->lo;
2394 struct nfs4_layoutget_res *res = &lgp->res;
2395 struct pnfs_layout_segment *lseg;
2396 struct inode *ino = lo->plh_inode;
2397 LIST_HEAD(free_me);
2398
2399 if (!pnfs_sanity_check_layout_range(&res->range))
2400 return ERR_PTR(-EINVAL);
2401
2402 /* Inject layout blob into I/O device driver */
2403 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2404 if (IS_ERR_OR_NULL(lseg)) {
2405 if (!lseg)
2406 lseg = ERR_PTR(-ENOMEM);
2407
2408 dprintk("%s: Could not allocate layout: error %ld\n",
2409 __func__, PTR_ERR(lseg));
2410 return lseg;
2411 }
2412
2413 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2414
2415 spin_lock(&ino->i_lock);
2416 if (pnfs_layoutgets_blocked(lo)) {
2417 dprintk("%s forget reply due to state\n", __func__);
2418 goto out_forget;
2419 }
2420
2421 if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2422 !pnfs_is_first_layoutget(lo))
2423 goto out_forget;
2424
2425 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2426 /* existing state ID, make sure the sequence number matches. */
2427 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2428 if (!pnfs_layout_is_valid(lo))
2429 lo->plh_barrier = 0;
2430 dprintk("%s forget reply due to sequence\n", __func__);
2431 goto out_forget;
2432 }
2433 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2434 } else if (pnfs_layout_is_valid(lo)) {
2435 /*
2436 * We got an entirely new state ID. Mark all segments for the
2437 * inode invalid, and retry the layoutget
2438 */
2439 struct pnfs_layout_range range = {
2440 .iomode = IOMODE_ANY,
2441 .length = NFS4_MAX_UINT64,
2442 };
2443 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2444 goto out_forget;
2445 } else {
2446 /* We have a completely new layout */
2447 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2448 }
2449
2450 pnfs_get_lseg(lseg);
2451 pnfs_layout_insert_lseg(lo, lseg, &free_me);
2452
2453
2454 if (res->return_on_close)
2455 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2456
2457 spin_unlock(&ino->i_lock);
2458 pnfs_free_lseg_list(&free_me);
2459 return lseg;
2460
2461out_forget:
2462 spin_unlock(&ino->i_lock);
2463 lseg->pls_layout = lo;
2464 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2465 return ERR_PTR(-EAGAIN);
2466}
2467
2468/**
2469 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2470 * @lo: pointer to layout header
2471 * @tmp_list: list header to be used with pnfs_free_lseg_list()
2472 * @return_range: describe layout segment ranges to be returned
2473 * @seq: stateid seqid to match
2474 *
2475 * This function is mainly intended for use by layoutrecall. It attempts
2476 * to free the layout segment immediately, or else to mark it for return
2477 * as soon as its reference count drops to zero.
2478 *
2479 * Returns
2480 * - 0: a layoutreturn needs to be scheduled.
2481 * - EBUSY: there are layout segment that are still in use.
2482 * - ENOENT: there are no layout segments that need to be returned.
2483 */
2484int
2485pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2486 struct list_head *tmp_list,
2487 const struct pnfs_layout_range *return_range,
2488 u32 seq)
2489{
2490 struct pnfs_layout_segment *lseg, *next;
2491 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2492 int remaining = 0;
2493
2494 dprintk("%s:Begin lo %p\n", __func__, lo);
2495
2496 assert_spin_locked(&lo->plh_inode->i_lock);
2497
2498 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2499 tmp_list = &lo->plh_return_segs;
2500
2501 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2502 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2503 dprintk("%s: marking lseg %p iomode %d "
2504 "offset %llu length %llu\n", __func__,
2505 lseg, lseg->pls_range.iomode,
2506 lseg->pls_range.offset,
2507 lseg->pls_range.length);
2508 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2509 tmp_list = &lo->plh_return_segs;
2510 if (mark_lseg_invalid(lseg, tmp_list))
2511 continue;
2512 remaining++;
2513 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2514 pnfs_lseg_cancel_io(server, lseg);
2515 }
2516
2517 if (remaining) {
2518 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2519 return -EBUSY;
2520 }
2521
2522 if (!list_empty(&lo->plh_return_segs)) {
2523 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2524 return 0;
2525 }
2526
2527 return -ENOENT;
2528}
2529
2530static void
2531pnfs_mark_layout_for_return(struct inode *inode,
2532 const struct pnfs_layout_range *range)
2533{
2534 struct pnfs_layout_hdr *lo;
2535 bool return_now = false;
2536
2537 spin_lock(&inode->i_lock);
2538 lo = NFS_I(inode)->layout;
2539 if (!pnfs_layout_is_valid(lo)) {
2540 spin_unlock(&inode->i_lock);
2541 return;
2542 }
2543 pnfs_set_plh_return_info(lo, range->iomode, 0);
2544 /*
2545 * mark all matching lsegs so that we are sure to have no live
2546 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2547 * for how it works.
2548 */
2549 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2550 const struct cred *cred;
2551 nfs4_stateid stateid;
2552 enum pnfs_iomode iomode;
2553
2554 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2555 spin_unlock(&inode->i_lock);
2556 if (return_now)
2557 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2558 } else {
2559 spin_unlock(&inode->i_lock);
2560 nfs_commit_inode(inode, 0);
2561 }
2562}
2563
2564void pnfs_error_mark_layout_for_return(struct inode *inode,
2565 struct pnfs_layout_segment *lseg)
2566{
2567 struct pnfs_layout_range range = {
2568 .iomode = lseg->pls_range.iomode,
2569 .offset = 0,
2570 .length = NFS4_MAX_UINT64,
2571 };
2572
2573 pnfs_mark_layout_for_return(inode, &range);
2574}
2575EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2576
2577static bool
2578pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2579{
2580 return pnfs_layout_is_valid(lo) &&
2581 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2582 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2583}
2584
2585static struct pnfs_layout_segment *
2586pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2587 const struct pnfs_layout_range *range,
2588 enum pnfs_iomode iomode)
2589{
2590 struct pnfs_layout_segment *lseg;
2591
2592 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2593 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2594 continue;
2595 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2596 continue;
2597 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2598 continue;
2599 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2600 return lseg;
2601 }
2602 return NULL;
2603}
2604
2605/* Find open file states whose mode matches that of the range */
2606static bool
2607pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2608 const struct pnfs_layout_range *range)
2609{
2610 struct list_head *head;
2611 struct nfs_open_context *ctx;
2612 fmode_t mode = 0;
2613
2614 if (!pnfs_layout_can_be_returned(lo) ||
2615 !pnfs_find_first_lseg(lo, range, range->iomode))
2616 return false;
2617
2618 head = &NFS_I(lo->plh_inode)->open_files;
2619 list_for_each_entry_rcu(ctx, head, list) {
2620 if (ctx->state)
2621 mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2622 }
2623
2624 switch (range->iomode) {
2625 default:
2626 break;
2627 case IOMODE_READ:
2628 mode &= ~FMODE_WRITE;
2629 break;
2630 case IOMODE_RW:
2631 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2632 mode &= ~FMODE_READ;
2633 }
2634 return mode == 0;
2635}
2636
2637static int
2638pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2639{
2640 const struct pnfs_layout_range *range = data;
2641 struct pnfs_layout_hdr *lo;
2642 struct inode *inode;
2643restart:
2644 rcu_read_lock();
2645 list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2646 if (!pnfs_layout_can_be_returned(lo) ||
2647 test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2648 continue;
2649 inode = lo->plh_inode;
2650 spin_lock(&inode->i_lock);
2651 if (!pnfs_should_return_unused_layout(lo, range)) {
2652 spin_unlock(&inode->i_lock);
2653 continue;
2654 }
2655 spin_unlock(&inode->i_lock);
2656 inode = pnfs_grab_inode_layout_hdr(lo);
2657 if (!inode)
2658 continue;
2659 rcu_read_unlock();
2660 pnfs_mark_layout_for_return(inode, range);
2661 iput(inode);
2662 cond_resched();
2663 goto restart;
2664 }
2665 rcu_read_unlock();
2666 return 0;
2667}
2668
2669void
2670pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2671 enum pnfs_iomode iomode)
2672{
2673 struct pnfs_layout_range range = {
2674 .iomode = iomode,
2675 .offset = 0,
2676 .length = NFS4_MAX_UINT64,
2677 };
2678
2679 nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2680 &range);
2681}
2682
2683void
2684pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2685{
2686 if (pgio->pg_lseg == NULL ||
2687 test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2688 return;
2689 pnfs_put_lseg(pgio->pg_lseg);
2690 pgio->pg_lseg = NULL;
2691}
2692EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2693
2694/*
2695 * Check for any intersection between the request and the pgio->pg_lseg,
2696 * and if none, put this pgio->pg_lseg away.
2697 */
2698void
2699pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2700{
2701 if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2702 pnfs_put_lseg(pgio->pg_lseg);
2703 pgio->pg_lseg = NULL;
2704 }
2705}
2706EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2707
2708void
2709pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2710{
2711 u64 rd_size;
2712
2713 pnfs_generic_pg_check_layout(pgio);
2714 pnfs_generic_pg_check_range(pgio, req);
2715 if (pgio->pg_lseg == NULL) {
2716 if (pgio->pg_dreq == NULL)
2717 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2718 else
2719 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2720
2721 pgio->pg_lseg =
2722 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2723 req_offset(req), rd_size,
2724 IOMODE_READ, false,
2725 nfs_io_gfp_mask());
2726 if (IS_ERR(pgio->pg_lseg)) {
2727 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2728 pgio->pg_lseg = NULL;
2729 return;
2730 }
2731 }
2732 /* If no lseg, fall back to read through mds */
2733 if (pgio->pg_lseg == NULL)
2734 nfs_pageio_reset_read_mds(pgio);
2735
2736}
2737EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2738
2739void
2740pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2741 struct nfs_page *req, u64 wb_size)
2742{
2743 pnfs_generic_pg_check_layout(pgio);
2744 pnfs_generic_pg_check_range(pgio, req);
2745 if (pgio->pg_lseg == NULL) {
2746 pgio->pg_lseg =
2747 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2748 req_offset(req), wb_size, IOMODE_RW,
2749 false, nfs_io_gfp_mask());
2750 if (IS_ERR(pgio->pg_lseg)) {
2751 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2752 pgio->pg_lseg = NULL;
2753 return;
2754 }
2755 }
2756 /* If no lseg, fall back to write through mds */
2757 if (pgio->pg_lseg == NULL)
2758 nfs_pageio_reset_write_mds(pgio);
2759}
2760EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2761
2762void
2763pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2764{
2765 if (desc->pg_lseg) {
2766 pnfs_put_lseg(desc->pg_lseg);
2767 desc->pg_lseg = NULL;
2768 }
2769}
2770EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2771
2772/*
2773 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2774 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2775 */
2776size_t
2777pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2778 struct nfs_page *prev, struct nfs_page *req)
2779{
2780 unsigned int size;
2781 u64 seg_end, req_start, seg_left;
2782
2783 size = nfs_generic_pg_test(pgio, prev, req);
2784 if (!size)
2785 return 0;
2786
2787 /*
2788 * 'size' contains the number of bytes left in the current page (up
2789 * to the original size asked for in @req->wb_bytes).
2790 *
2791 * Calculate how many bytes are left in the layout segment
2792 * and if there are less bytes than 'size', return that instead.
2793 *
2794 * Please also note that 'end_offset' is actually the offset of the
2795 * first byte that lies outside the pnfs_layout_range. FIXME?
2796 *
2797 */
2798 if (pgio->pg_lseg) {
2799 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2800 pgio->pg_lseg->pls_range.length);
2801 req_start = req_offset(req);
2802
2803 /* start of request is past the last byte of this segment */
2804 if (req_start >= seg_end)
2805 return 0;
2806
2807 /* adjust 'size' iff there are fewer bytes left in the
2808 * segment than what nfs_generic_pg_test returned */
2809 seg_left = seg_end - req_start;
2810 if (seg_left < size)
2811 size = (unsigned int)seg_left;
2812 }
2813
2814 return size;
2815}
2816EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2817
2818int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2819{
2820 struct nfs_pageio_descriptor pgio;
2821
2822 /* Resend all requests through the MDS */
2823 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2824 hdr->completion_ops);
2825 return nfs_pageio_resend(&pgio, hdr);
2826}
2827EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2828
2829static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2830{
2831
2832 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2833 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2834 PNFS_LAYOUTRET_ON_ERROR) {
2835 pnfs_return_layout(hdr->inode);
2836 }
2837 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2838 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2839}
2840
2841/*
2842 * Called by non rpc-based layout drivers
2843 */
2844void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2845{
2846 if (likely(!hdr->pnfs_error)) {
2847 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2848 hdr->mds_offset + hdr->res.count);
2849 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2850 }
2851 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2852 if (unlikely(hdr->pnfs_error))
2853 pnfs_ld_handle_write_error(hdr);
2854 hdr->mds_ops->rpc_release(hdr);
2855}
2856EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2857
2858static void
2859pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2860 struct nfs_pgio_header *hdr)
2861{
2862 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2863
2864 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2865 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2866 nfs_pageio_reset_write_mds(desc);
2867 mirror->pg_recoalesce = 1;
2868 }
2869 hdr->completion_ops->completion(hdr);
2870}
2871
2872static enum pnfs_try_status
2873pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2874 const struct rpc_call_ops *call_ops,
2875 struct pnfs_layout_segment *lseg,
2876 int how)
2877{
2878 struct inode *inode = hdr->inode;
2879 enum pnfs_try_status trypnfs;
2880 struct nfs_server *nfss = NFS_SERVER(inode);
2881
2882 hdr->mds_ops = call_ops;
2883
2884 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2885 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2886 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2887 if (trypnfs != PNFS_NOT_ATTEMPTED)
2888 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2889 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2890 return trypnfs;
2891}
2892
2893static void
2894pnfs_do_write(struct nfs_pageio_descriptor *desc,
2895 struct nfs_pgio_header *hdr, int how)
2896{
2897 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2898 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2899 enum pnfs_try_status trypnfs;
2900
2901 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2902 switch (trypnfs) {
2903 case PNFS_NOT_ATTEMPTED:
2904 pnfs_write_through_mds(desc, hdr);
2905 break;
2906 case PNFS_ATTEMPTED:
2907 break;
2908 case PNFS_TRY_AGAIN:
2909 /* cleanup hdr and prepare to redo pnfs */
2910 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2911 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2912 list_splice_init(&hdr->pages, &mirror->pg_list);
2913 mirror->pg_recoalesce = 1;
2914 }
2915 hdr->mds_ops->rpc_release(hdr);
2916 }
2917}
2918
2919static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2920{
2921 pnfs_put_lseg(hdr->lseg);
2922 nfs_pgio_header_free(hdr);
2923}
2924
2925int
2926pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2927{
2928 struct nfs_pgio_header *hdr;
2929 int ret;
2930
2931 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2932 if (!hdr) {
2933 desc->pg_error = -ENOMEM;
2934 return desc->pg_error;
2935 }
2936 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2937
2938 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2939 ret = nfs_generic_pgio(desc, hdr);
2940 if (!ret)
2941 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2942
2943 return ret;
2944}
2945EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2946
2947int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2948{
2949 struct nfs_pageio_descriptor pgio;
2950
2951 /* Resend all requests through the MDS */
2952 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2953 return nfs_pageio_resend(&pgio, hdr);
2954}
2955EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2956
2957static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2958{
2959 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2960 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2961 PNFS_LAYOUTRET_ON_ERROR) {
2962 pnfs_return_layout(hdr->inode);
2963 }
2964 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2965 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2966}
2967
2968/*
2969 * Called by non rpc-based layout drivers
2970 */
2971void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2972{
2973 if (likely(!hdr->pnfs_error))
2974 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2975 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2976 if (unlikely(hdr->pnfs_error))
2977 pnfs_ld_handle_read_error(hdr);
2978 hdr->mds_ops->rpc_release(hdr);
2979}
2980EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2981
2982static void
2983pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2984 struct nfs_pgio_header *hdr)
2985{
2986 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2987
2988 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2989 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2990 nfs_pageio_reset_read_mds(desc);
2991 mirror->pg_recoalesce = 1;
2992 }
2993 hdr->completion_ops->completion(hdr);
2994}
2995
2996/*
2997 * Call the appropriate parallel I/O subsystem read function.
2998 */
2999static enum pnfs_try_status
3000pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3001 const struct rpc_call_ops *call_ops,
3002 struct pnfs_layout_segment *lseg)
3003{
3004 struct inode *inode = hdr->inode;
3005 struct nfs_server *nfss = NFS_SERVER(inode);
3006 enum pnfs_try_status trypnfs;
3007
3008 hdr->mds_ops = call_ops;
3009
3010 dprintk("%s: Reading ino:%lu %u@%llu\n",
3011 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3012
3013 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3014 if (trypnfs != PNFS_NOT_ATTEMPTED)
3015 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3016 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3017 return trypnfs;
3018}
3019
3020/* Resend all requests through pnfs. */
3021void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3022 unsigned int mirror_idx)
3023{
3024 struct nfs_pageio_descriptor pgio;
3025
3026 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3027 /* Prevent deadlocks with layoutreturn! */
3028 pnfs_put_lseg(hdr->lseg);
3029 hdr->lseg = NULL;
3030
3031 nfs_pageio_init_read(&pgio, hdr->inode, false,
3032 hdr->completion_ops);
3033 pgio.pg_mirror_idx = mirror_idx;
3034 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3035 }
3036}
3037EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3038
3039static void
3040pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3041{
3042 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3043 struct pnfs_layout_segment *lseg = desc->pg_lseg;
3044 enum pnfs_try_status trypnfs;
3045
3046 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3047 switch (trypnfs) {
3048 case PNFS_NOT_ATTEMPTED:
3049 pnfs_read_through_mds(desc, hdr);
3050 break;
3051 case PNFS_ATTEMPTED:
3052 break;
3053 case PNFS_TRY_AGAIN:
3054 /* cleanup hdr and prepare to redo pnfs */
3055 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3056 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3057 list_splice_init(&hdr->pages, &mirror->pg_list);
3058 mirror->pg_recoalesce = 1;
3059 }
3060 hdr->mds_ops->rpc_release(hdr);
3061 }
3062}
3063
3064static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3065{
3066 pnfs_put_lseg(hdr->lseg);
3067 nfs_pgio_header_free(hdr);
3068}
3069
3070int
3071pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3072{
3073 struct nfs_pgio_header *hdr;
3074 int ret;
3075
3076 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3077 if (!hdr) {
3078 desc->pg_error = -ENOMEM;
3079 return desc->pg_error;
3080 }
3081 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3082 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3083 ret = nfs_generic_pgio(desc, hdr);
3084 if (!ret)
3085 pnfs_do_read(desc, hdr);
3086 return ret;
3087}
3088EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3089
3090static void pnfs_clear_layoutcommitting(struct inode *inode)
3091{
3092 unsigned long *bitlock = &NFS_I(inode)->flags;
3093
3094 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3095 smp_mb__after_atomic();
3096 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3097}
3098
3099/*
3100 * There can be multiple RW segments.
3101 */
3102static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3103{
3104 struct pnfs_layout_segment *lseg;
3105
3106 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3107 if (lseg->pls_range.iomode == IOMODE_RW &&
3108 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3109 list_add(&lseg->pls_lc_list, listp);
3110 }
3111}
3112
3113static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3114{
3115 struct pnfs_layout_segment *lseg, *tmp;
3116
3117 /* Matched by references in pnfs_set_layoutcommit */
3118 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3119 list_del_init(&lseg->pls_lc_list);
3120 pnfs_put_lseg(lseg);
3121 }
3122
3123 pnfs_clear_layoutcommitting(inode);
3124}
3125
3126void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3127{
3128 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3129}
3130EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3131
3132void
3133pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3134 loff_t end_pos)
3135{
3136 struct nfs_inode *nfsi = NFS_I(inode);
3137 bool mark_as_dirty = false;
3138
3139 spin_lock(&inode->i_lock);
3140 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3141 nfsi->layout->plh_lwb = end_pos;
3142 mark_as_dirty = true;
3143 dprintk("%s: Set layoutcommit for inode %lu ",
3144 __func__, inode->i_ino);
3145 } else if (end_pos > nfsi->layout->plh_lwb)
3146 nfsi->layout->plh_lwb = end_pos;
3147 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3148 /* references matched in nfs4_layoutcommit_release */
3149 pnfs_get_lseg(lseg);
3150 }
3151 spin_unlock(&inode->i_lock);
3152 dprintk("%s: lseg %p end_pos %llu\n",
3153 __func__, lseg, nfsi->layout->plh_lwb);
3154
3155 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3156 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3157 if (mark_as_dirty)
3158 mark_inode_dirty_sync(inode);
3159}
3160EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3161
3162void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3163{
3164 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3165
3166 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3167 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3168 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3169}
3170
3171/*
3172 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3173 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3174 * data to disk to allow the server to recover the data if it crashes.
3175 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3176 * is off, and a COMMIT is sent to a data server, or
3177 * if WRITEs to a data server return NFS_DATA_SYNC.
3178 */
3179int
3180pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3181{
3182 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3183 struct nfs4_layoutcommit_data *data;
3184 struct nfs_inode *nfsi = NFS_I(inode);
3185 loff_t end_pos;
3186 int status;
3187
3188 if (!pnfs_layoutcommit_outstanding(inode))
3189 return 0;
3190
3191 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3192
3193 status = -EAGAIN;
3194 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3195 if (!sync)
3196 goto out;
3197 status = wait_on_bit_lock_action(&nfsi->flags,
3198 NFS_INO_LAYOUTCOMMITTING,
3199 nfs_wait_bit_killable,
3200 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3201 if (status)
3202 goto out;
3203 }
3204
3205 status = -ENOMEM;
3206 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3207 data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3208 if (!data)
3209 goto clear_layoutcommitting;
3210
3211 status = 0;
3212 spin_lock(&inode->i_lock);
3213 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3214 goto out_unlock;
3215
3216 INIT_LIST_HEAD(&data->lseg_list);
3217 pnfs_list_write_lseg(inode, &data->lseg_list);
3218
3219 end_pos = nfsi->layout->plh_lwb;
3220
3221 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3222 data->cred = get_cred(nfsi->layout->plh_lc_cred);
3223 spin_unlock(&inode->i_lock);
3224
3225 data->args.inode = inode;
3226 nfs_fattr_init(&data->fattr);
3227 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3228 data->res.fattr = &data->fattr;
3229 if (end_pos != 0)
3230 data->args.lastbytewritten = end_pos - 1;
3231 else
3232 data->args.lastbytewritten = U64_MAX;
3233 data->res.server = NFS_SERVER(inode);
3234
3235 if (ld->prepare_layoutcommit) {
3236 status = ld->prepare_layoutcommit(&data->args);
3237 if (status) {
3238 put_cred(data->cred);
3239 spin_lock(&inode->i_lock);
3240 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3241 if (end_pos > nfsi->layout->plh_lwb)
3242 nfsi->layout->plh_lwb = end_pos;
3243 goto out_unlock;
3244 }
3245 }
3246
3247
3248 status = nfs4_proc_layoutcommit(data, sync);
3249out:
3250 if (status)
3251 mark_inode_dirty_sync(inode);
3252 dprintk("<-- %s status %d\n", __func__, status);
3253 return status;
3254out_unlock:
3255 spin_unlock(&inode->i_lock);
3256 kfree(data);
3257clear_layoutcommitting:
3258 pnfs_clear_layoutcommitting(inode);
3259 goto out;
3260}
3261EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3262
3263int
3264pnfs_generic_sync(struct inode *inode, bool datasync)
3265{
3266 return pnfs_layoutcommit_inode(inode, true);
3267}
3268EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3269
3270struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3271{
3272 struct nfs4_threshold *thp;
3273
3274 thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3275 if (!thp) {
3276 dprintk("%s mdsthreshold allocation failed\n", __func__);
3277 return NULL;
3278 }
3279 return thp;
3280}
3281
3282#if IS_ENABLED(CONFIG_NFS_V4_2)
3283int
3284pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3285{
3286 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3287 struct nfs_server *server = NFS_SERVER(inode);
3288 struct nfs_inode *nfsi = NFS_I(inode);
3289 struct nfs42_layoutstat_data *data;
3290 struct pnfs_layout_hdr *hdr;
3291 int status = 0;
3292
3293 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3294 goto out;
3295
3296 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3297 goto out;
3298
3299 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3300 goto out;
3301
3302 spin_lock(&inode->i_lock);
3303 if (!NFS_I(inode)->layout) {
3304 spin_unlock(&inode->i_lock);
3305 goto out_clear_layoutstats;
3306 }
3307 hdr = NFS_I(inode)->layout;
3308 pnfs_get_layout_hdr(hdr);
3309 spin_unlock(&inode->i_lock);
3310
3311 data = kzalloc(sizeof(*data), gfp_flags);
3312 if (!data) {
3313 status = -ENOMEM;
3314 goto out_put;
3315 }
3316
3317 data->args.fh = NFS_FH(inode);
3318 data->args.inode = inode;
3319 status = ld->prepare_layoutstats(&data->args);
3320 if (status)
3321 goto out_free;
3322
3323 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3324
3325out:
3326 dprintk("%s returns %d\n", __func__, status);
3327 return status;
3328
3329out_free:
3330 kfree(data);
3331out_put:
3332 pnfs_put_layout_hdr(hdr);
3333out_clear_layoutstats:
3334 smp_mb__before_atomic();
3335 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3336 smp_mb__after_atomic();
3337 goto out;
3338}
3339EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3340#endif
3341
3342unsigned int layoutstats_timer;
3343module_param(layoutstats_timer, uint, 0644);
3344EXPORT_SYMBOL_GPL(layoutstats_timer);