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 "internal.h"
34#include "pnfs.h"
35#include "iostat.h"
36#include "nfs4trace.h"
37
38#define NFSDBG_FACILITY NFSDBG_PNFS
39#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
40
41/* Locking:
42 *
43 * pnfs_spinlock:
44 * protects pnfs_modules_tbl.
45 */
46static DEFINE_SPINLOCK(pnfs_spinlock);
47
48/*
49 * pnfs_modules_tbl holds all pnfs modules
50 */
51static LIST_HEAD(pnfs_modules_tbl);
52
53/* Return the registered pnfs layout driver module matching given id */
54static struct pnfs_layoutdriver_type *
55find_pnfs_driver_locked(u32 id)
56{
57 struct pnfs_layoutdriver_type *local;
58
59 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
60 if (local->id == id)
61 goto out;
62 local = NULL;
63out:
64 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
65 return local;
66}
67
68static struct pnfs_layoutdriver_type *
69find_pnfs_driver(u32 id)
70{
71 struct pnfs_layoutdriver_type *local;
72
73 spin_lock(&pnfs_spinlock);
74 local = find_pnfs_driver_locked(id);
75 if (local != NULL && !try_module_get(local->owner)) {
76 dprintk("%s: Could not grab reference on module\n", __func__);
77 local = NULL;
78 }
79 spin_unlock(&pnfs_spinlock);
80 return local;
81}
82
83void
84unset_pnfs_layoutdriver(struct nfs_server *nfss)
85{
86 if (nfss->pnfs_curr_ld) {
87 if (nfss->pnfs_curr_ld->clear_layoutdriver)
88 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
89 /* Decrement the MDS count. Purge the deviceid cache if zero */
90 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
91 nfs4_deviceid_purge_client(nfss->nfs_client);
92 module_put(nfss->pnfs_curr_ld->owner);
93 }
94 nfss->pnfs_curr_ld = NULL;
95}
96
97/*
98 * Try to set the server's pnfs module to the pnfs layout type specified by id.
99 * Currently only one pNFS layout driver per filesystem is supported.
100 *
101 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
102 */
103void
104set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
105 u32 id)
106{
107 struct pnfs_layoutdriver_type *ld_type = NULL;
108
109 if (id == 0)
110 goto out_no_driver;
111 if (!(server->nfs_client->cl_exchange_flags &
112 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
113 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
114 __func__, id, server->nfs_client->cl_exchange_flags);
115 goto out_no_driver;
116 }
117 ld_type = find_pnfs_driver(id);
118 if (!ld_type) {
119 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
120 ld_type = find_pnfs_driver(id);
121 if (!ld_type) {
122 dprintk("%s: No pNFS module found for %u.\n",
123 __func__, id);
124 goto out_no_driver;
125 }
126 }
127 server->pnfs_curr_ld = ld_type;
128 if (ld_type->set_layoutdriver
129 && ld_type->set_layoutdriver(server, mntfh)) {
130 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
131 "driver %u.\n", __func__, id);
132 module_put(ld_type->owner);
133 goto out_no_driver;
134 }
135 /* Bump the MDS count */
136 atomic_inc(&server->nfs_client->cl_mds_count);
137
138 dprintk("%s: pNFS module for %u set\n", __func__, id);
139 return;
140
141out_no_driver:
142 dprintk("%s: Using NFSv4 I/O\n", __func__);
143 server->pnfs_curr_ld = NULL;
144}
145
146int
147pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
148{
149 int status = -EINVAL;
150 struct pnfs_layoutdriver_type *tmp;
151
152 if (ld_type->id == 0) {
153 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
154 return status;
155 }
156 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
157 printk(KERN_ERR "NFS: %s Layout driver must provide "
158 "alloc_lseg and free_lseg.\n", __func__);
159 return status;
160 }
161
162 spin_lock(&pnfs_spinlock);
163 tmp = find_pnfs_driver_locked(ld_type->id);
164 if (!tmp) {
165 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
166 status = 0;
167 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
168 ld_type->name);
169 } else {
170 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
171 __func__, ld_type->id);
172 }
173 spin_unlock(&pnfs_spinlock);
174
175 return status;
176}
177EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
178
179void
180pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
181{
182 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
183 spin_lock(&pnfs_spinlock);
184 list_del(&ld_type->pnfs_tblid);
185 spin_unlock(&pnfs_spinlock);
186}
187EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
188
189/*
190 * pNFS client layout cache
191 */
192
193/* Need to hold i_lock if caller does not already hold reference */
194void
195pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
196{
197 atomic_inc(&lo->plh_refcount);
198}
199
200static struct pnfs_layout_hdr *
201pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
202{
203 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
204 return ld->alloc_layout_hdr(ino, gfp_flags);
205}
206
207static void
208pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
209{
210 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
211 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
212
213 if (!list_empty(&lo->plh_layouts)) {
214 struct nfs_client *clp = server->nfs_client;
215
216 spin_lock(&clp->cl_lock);
217 list_del_init(&lo->plh_layouts);
218 spin_unlock(&clp->cl_lock);
219 }
220 put_rpccred(lo->plh_lc_cred);
221 return ld->free_layout_hdr(lo);
222}
223
224static void
225pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
226{
227 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
228 dprintk("%s: freeing layout cache %p\n", __func__, lo);
229 nfsi->layout = NULL;
230 /* Reset MDS Threshold I/O counters */
231 nfsi->write_io = 0;
232 nfsi->read_io = 0;
233}
234
235void
236pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
237{
238 struct inode *inode = lo->plh_inode;
239
240 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
241 pnfs_detach_layout_hdr(lo);
242 spin_unlock(&inode->i_lock);
243 pnfs_free_layout_hdr(lo);
244 }
245}
246
247static int
248pnfs_iomode_to_fail_bit(u32 iomode)
249{
250 return iomode == IOMODE_RW ?
251 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
252}
253
254static void
255pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
256{
257 lo->plh_retry_timestamp = jiffies;
258 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
259 atomic_inc(&lo->plh_refcount);
260}
261
262static void
263pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
264{
265 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
266 atomic_dec(&lo->plh_refcount);
267}
268
269static void
270pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
271{
272 struct inode *inode = lo->plh_inode;
273 struct pnfs_layout_range range = {
274 .iomode = iomode,
275 .offset = 0,
276 .length = NFS4_MAX_UINT64,
277 };
278 LIST_HEAD(head);
279
280 spin_lock(&inode->i_lock);
281 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
282 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
283 spin_unlock(&inode->i_lock);
284 pnfs_free_lseg_list(&head);
285 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
286 iomode == IOMODE_RW ? "RW" : "READ");
287}
288
289static bool
290pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
291{
292 unsigned long start, end;
293 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
294
295 if (test_bit(fail_bit, &lo->plh_flags) == 0)
296 return false;
297 end = jiffies;
298 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
299 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
300 /* It is time to retry the failed layoutgets */
301 pnfs_layout_clear_fail_bit(lo, fail_bit);
302 return false;
303 }
304 return true;
305}
306
307static void
308init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
309{
310 INIT_LIST_HEAD(&lseg->pls_list);
311 INIT_LIST_HEAD(&lseg->pls_lc_list);
312 atomic_set(&lseg->pls_refcount, 1);
313 smp_mb();
314 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
315 lseg->pls_layout = lo;
316}
317
318static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
319{
320 struct inode *ino = lseg->pls_layout->plh_inode;
321
322 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
323}
324
325static void
326pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
327 struct pnfs_layout_segment *lseg)
328{
329 struct inode *inode = lo->plh_inode;
330
331 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
332 list_del_init(&lseg->pls_list);
333 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
334 atomic_dec(&lo->plh_refcount);
335 if (list_empty(&lo->plh_segs))
336 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
337 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
338}
339
340void
341pnfs_put_lseg(struct pnfs_layout_segment *lseg)
342{
343 struct pnfs_layout_hdr *lo;
344 struct inode *inode;
345
346 if (!lseg)
347 return;
348
349 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
350 atomic_read(&lseg->pls_refcount),
351 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
352 lo = lseg->pls_layout;
353 inode = lo->plh_inode;
354 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
355 pnfs_get_layout_hdr(lo);
356 pnfs_layout_remove_lseg(lo, lseg);
357 spin_unlock(&inode->i_lock);
358 pnfs_free_lseg(lseg);
359 pnfs_put_layout_hdr(lo);
360 }
361}
362EXPORT_SYMBOL_GPL(pnfs_put_lseg);
363
364static u64
365end_offset(u64 start, u64 len)
366{
367 u64 end;
368
369 end = start + len;
370 return end >= start ? end : NFS4_MAX_UINT64;
371}
372
373/*
374 * is l2 fully contained in l1?
375 * start1 end1
376 * [----------------------------------)
377 * start2 end2
378 * [----------------)
379 */
380static bool
381pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
382 const struct pnfs_layout_range *l2)
383{
384 u64 start1 = l1->offset;
385 u64 end1 = end_offset(start1, l1->length);
386 u64 start2 = l2->offset;
387 u64 end2 = end_offset(start2, l2->length);
388
389 return (start1 <= start2) && (end1 >= end2);
390}
391
392/*
393 * is l1 and l2 intersecting?
394 * start1 end1
395 * [----------------------------------)
396 * start2 end2
397 * [----------------)
398 */
399static bool
400pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
401 const struct pnfs_layout_range *l2)
402{
403 u64 start1 = l1->offset;
404 u64 end1 = end_offset(start1, l1->length);
405 u64 start2 = l2->offset;
406 u64 end2 = end_offset(start2, l2->length);
407
408 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
409 (end2 == NFS4_MAX_UINT64 || end2 > start1);
410}
411
412static bool
413should_free_lseg(const struct pnfs_layout_range *lseg_range,
414 const struct pnfs_layout_range *recall_range)
415{
416 return (recall_range->iomode == IOMODE_ANY ||
417 lseg_range->iomode == recall_range->iomode) &&
418 pnfs_lseg_range_intersecting(lseg_range, recall_range);
419}
420
421static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
422 struct list_head *tmp_list)
423{
424 if (!atomic_dec_and_test(&lseg->pls_refcount))
425 return false;
426 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
427 list_add(&lseg->pls_list, tmp_list);
428 return true;
429}
430
431/* Returns 1 if lseg is removed from list, 0 otherwise */
432static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
433 struct list_head *tmp_list)
434{
435 int rv = 0;
436
437 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
438 /* Remove the reference keeping the lseg in the
439 * list. It will now be removed when all
440 * outstanding io is finished.
441 */
442 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
443 atomic_read(&lseg->pls_refcount));
444 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
445 rv = 1;
446 }
447 return rv;
448}
449
450/* Returns count of number of matching invalid lsegs remaining in list
451 * after call.
452 */
453int
454pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
455 struct list_head *tmp_list,
456 struct pnfs_layout_range *recall_range)
457{
458 struct pnfs_layout_segment *lseg, *next;
459 int invalid = 0, removed = 0;
460
461 dprintk("%s:Begin lo %p\n", __func__, lo);
462
463 if (list_empty(&lo->plh_segs))
464 return 0;
465 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
466 if (!recall_range ||
467 should_free_lseg(&lseg->pls_range, recall_range)) {
468 dprintk("%s: freeing lseg %p iomode %d "
469 "offset %llu length %llu\n", __func__,
470 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
471 lseg->pls_range.length);
472 invalid++;
473 removed += mark_lseg_invalid(lseg, tmp_list);
474 }
475 dprintk("%s:Return %i\n", __func__, invalid - removed);
476 return invalid - removed;
477}
478
479/* note free_me must contain lsegs from a single layout_hdr */
480void
481pnfs_free_lseg_list(struct list_head *free_me)
482{
483 struct pnfs_layout_segment *lseg, *tmp;
484
485 if (list_empty(free_me))
486 return;
487
488 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
489 list_del(&lseg->pls_list);
490 pnfs_free_lseg(lseg);
491 }
492}
493
494void
495pnfs_destroy_layout(struct nfs_inode *nfsi)
496{
497 struct pnfs_layout_hdr *lo;
498 LIST_HEAD(tmp_list);
499
500 spin_lock(&nfsi->vfs_inode.i_lock);
501 lo = nfsi->layout;
502 if (lo) {
503 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
504 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
505 pnfs_get_layout_hdr(lo);
506 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
507 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
508 spin_unlock(&nfsi->vfs_inode.i_lock);
509 pnfs_free_lseg_list(&tmp_list);
510 pnfs_put_layout_hdr(lo);
511 } else
512 spin_unlock(&nfsi->vfs_inode.i_lock);
513}
514EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
515
516static bool
517pnfs_layout_add_bulk_destroy_list(struct inode *inode,
518 struct list_head *layout_list)
519{
520 struct pnfs_layout_hdr *lo;
521 bool ret = false;
522
523 spin_lock(&inode->i_lock);
524 lo = NFS_I(inode)->layout;
525 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
526 pnfs_get_layout_hdr(lo);
527 list_add(&lo->plh_bulk_destroy, layout_list);
528 ret = true;
529 }
530 spin_unlock(&inode->i_lock);
531 return ret;
532}
533
534/* Caller must hold rcu_read_lock and clp->cl_lock */
535static int
536pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
537 struct nfs_server *server,
538 struct list_head *layout_list)
539{
540 struct pnfs_layout_hdr *lo, *next;
541 struct inode *inode;
542
543 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
544 inode = igrab(lo->plh_inode);
545 if (inode == NULL)
546 continue;
547 list_del_init(&lo->plh_layouts);
548 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
549 continue;
550 rcu_read_unlock();
551 spin_unlock(&clp->cl_lock);
552 iput(inode);
553 spin_lock(&clp->cl_lock);
554 rcu_read_lock();
555 return -EAGAIN;
556 }
557 return 0;
558}
559
560static int
561pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
562 bool is_bulk_recall)
563{
564 struct pnfs_layout_hdr *lo;
565 struct inode *inode;
566 struct pnfs_layout_range range = {
567 .iomode = IOMODE_ANY,
568 .offset = 0,
569 .length = NFS4_MAX_UINT64,
570 };
571 LIST_HEAD(lseg_list);
572 int ret = 0;
573
574 while (!list_empty(layout_list)) {
575 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
576 plh_bulk_destroy);
577 dprintk("%s freeing layout for inode %lu\n", __func__,
578 lo->plh_inode->i_ino);
579 inode = lo->plh_inode;
580 spin_lock(&inode->i_lock);
581 list_del_init(&lo->plh_bulk_destroy);
582 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
583 if (is_bulk_recall)
584 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
585 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
586 ret = -EAGAIN;
587 spin_unlock(&inode->i_lock);
588 pnfs_free_lseg_list(&lseg_list);
589 pnfs_put_layout_hdr(lo);
590 iput(inode);
591 }
592 return ret;
593}
594
595int
596pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
597 struct nfs_fsid *fsid,
598 bool is_recall)
599{
600 struct nfs_server *server;
601 LIST_HEAD(layout_list);
602
603 spin_lock(&clp->cl_lock);
604 rcu_read_lock();
605restart:
606 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
607 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
608 continue;
609 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
610 server,
611 &layout_list) != 0)
612 goto restart;
613 }
614 rcu_read_unlock();
615 spin_unlock(&clp->cl_lock);
616
617 if (list_empty(&layout_list))
618 return 0;
619 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
620}
621
622int
623pnfs_destroy_layouts_byclid(struct nfs_client *clp,
624 bool is_recall)
625{
626 struct nfs_server *server;
627 LIST_HEAD(layout_list);
628
629 spin_lock(&clp->cl_lock);
630 rcu_read_lock();
631restart:
632 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
633 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
634 server,
635 &layout_list) != 0)
636 goto restart;
637 }
638 rcu_read_unlock();
639 spin_unlock(&clp->cl_lock);
640
641 if (list_empty(&layout_list))
642 return 0;
643 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
644}
645
646/*
647 * Called by the state manger to remove all layouts established under an
648 * expired lease.
649 */
650void
651pnfs_destroy_all_layouts(struct nfs_client *clp)
652{
653 nfs4_deviceid_mark_client_invalid(clp);
654 nfs4_deviceid_purge_client(clp);
655
656 pnfs_destroy_layouts_byclid(clp, false);
657}
658
659/*
660 * Compare 2 layout stateid sequence ids, to see which is newer,
661 * taking into account wraparound issues.
662 */
663static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
664{
665 return (s32)(s1 - s2) > 0;
666}
667
668static void
669pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
670 const nfs4_stateid *new,
671 struct list_head *free_me_list)
672{
673 if (nfs4_stateid_match_other(&lo->plh_stateid, new))
674 return;
675 /* Layout is new! Kill existing layout segments */
676 pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
677}
678
679/* update lo->plh_stateid with new if is more recent */
680void
681pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
682 bool update_barrier)
683{
684 u32 oldseq, newseq, new_barrier;
685 int empty = list_empty(&lo->plh_segs);
686
687 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
688 newseq = be32_to_cpu(new->seqid);
689 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
690 nfs4_stateid_copy(&lo->plh_stateid, new);
691 if (update_barrier) {
692 new_barrier = be32_to_cpu(new->seqid);
693 } else {
694 /* Because of wraparound, we want to keep the barrier
695 * "close" to the current seqids.
696 */
697 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
698 }
699 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
700 lo->plh_barrier = new_barrier;
701 }
702}
703
704static bool
705pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
706 const nfs4_stateid *stateid)
707{
708 u32 seqid = be32_to_cpu(stateid->seqid);
709
710 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
711}
712
713/* lget is set to 1 if called from inside send_layoutget call chain */
714static bool
715pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
716{
717 return lo->plh_block_lgets ||
718 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
719 (list_empty(&lo->plh_segs) &&
720 (atomic_read(&lo->plh_outstanding) > lget));
721}
722
723int
724pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
725 struct nfs4_state *open_state)
726{
727 int status = 0;
728
729 dprintk("--> %s\n", __func__);
730 spin_lock(&lo->plh_inode->i_lock);
731 if (pnfs_layoutgets_blocked(lo, 1)) {
732 status = -EAGAIN;
733 } else if (!nfs4_valid_open_stateid(open_state)) {
734 status = -EBADF;
735 } else if (list_empty(&lo->plh_segs)) {
736 int seq;
737
738 do {
739 seq = read_seqbegin(&open_state->seqlock);
740 nfs4_stateid_copy(dst, &open_state->stateid);
741 } while (read_seqretry(&open_state->seqlock, seq));
742 } else
743 nfs4_stateid_copy(dst, &lo->plh_stateid);
744 spin_unlock(&lo->plh_inode->i_lock);
745 dprintk("<-- %s\n", __func__);
746 return status;
747}
748
749/*
750* Get layout from server.
751* for now, assume that whole file layouts are requested.
752* arg->offset: 0
753* arg->length: all ones
754*/
755static struct pnfs_layout_segment *
756send_layoutget(struct pnfs_layout_hdr *lo,
757 struct nfs_open_context *ctx,
758 struct pnfs_layout_range *range,
759 gfp_t gfp_flags)
760{
761 struct inode *ino = lo->plh_inode;
762 struct nfs_server *server = NFS_SERVER(ino);
763 struct nfs4_layoutget *lgp;
764 struct pnfs_layout_segment *lseg;
765
766 dprintk("--> %s\n", __func__);
767
768 lgp = kzalloc(sizeof(*lgp), gfp_flags);
769 if (lgp == NULL)
770 return NULL;
771
772 lgp->args.minlength = PAGE_CACHE_SIZE;
773 if (lgp->args.minlength > range->length)
774 lgp->args.minlength = range->length;
775 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
776 lgp->args.range = *range;
777 lgp->args.type = server->pnfs_curr_ld->id;
778 lgp->args.inode = ino;
779 lgp->args.ctx = get_nfs_open_context(ctx);
780 lgp->gfp_flags = gfp_flags;
781 lgp->cred = lo->plh_lc_cred;
782
783 /* Synchronously retrieve layout information from server and
784 * store in lseg.
785 */
786 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
787 if (IS_ERR(lseg)) {
788 switch (PTR_ERR(lseg)) {
789 case -ENOMEM:
790 case -ERESTARTSYS:
791 break;
792 default:
793 /* remember that LAYOUTGET failed and suspend trying */
794 pnfs_layout_io_set_failed(lo, range->iomode);
795 }
796 return NULL;
797 }
798
799 return lseg;
800}
801
802static void pnfs_clear_layoutcommit(struct inode *inode,
803 struct list_head *head)
804{
805 struct nfs_inode *nfsi = NFS_I(inode);
806 struct pnfs_layout_segment *lseg, *tmp;
807
808 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
809 return;
810 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
811 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
812 continue;
813 pnfs_lseg_dec_and_remove_zero(lseg, head);
814 }
815}
816
817/*
818 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
819 * when the layout segment list is empty.
820 *
821 * Note that a pnfs_layout_hdr can exist with an empty layout segment
822 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
823 * deviceid is marked invalid.
824 */
825int
826_pnfs_return_layout(struct inode *ino)
827{
828 struct pnfs_layout_hdr *lo = NULL;
829 struct nfs_inode *nfsi = NFS_I(ino);
830 LIST_HEAD(tmp_list);
831 struct nfs4_layoutreturn *lrp;
832 nfs4_stateid stateid;
833 int status = 0, empty;
834
835 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
836
837 spin_lock(&ino->i_lock);
838 lo = nfsi->layout;
839 if (!lo) {
840 spin_unlock(&ino->i_lock);
841 dprintk("NFS: %s no layout to return\n", __func__);
842 goto out;
843 }
844 stateid = nfsi->layout->plh_stateid;
845 /* Reference matched in nfs4_layoutreturn_release */
846 pnfs_get_layout_hdr(lo);
847 empty = list_empty(&lo->plh_segs);
848 pnfs_clear_layoutcommit(ino, &tmp_list);
849 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
850 /* Don't send a LAYOUTRETURN if list was initially empty */
851 if (empty) {
852 spin_unlock(&ino->i_lock);
853 pnfs_put_layout_hdr(lo);
854 dprintk("NFS: %s no layout segments to return\n", __func__);
855 goto out;
856 }
857 lo->plh_block_lgets++;
858 spin_unlock(&ino->i_lock);
859 pnfs_free_lseg_list(&tmp_list);
860
861 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
862 if (unlikely(lrp == NULL)) {
863 status = -ENOMEM;
864 spin_lock(&ino->i_lock);
865 lo->plh_block_lgets--;
866 spin_unlock(&ino->i_lock);
867 pnfs_put_layout_hdr(lo);
868 goto out;
869 }
870
871 lrp->args.stateid = stateid;
872 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
873 lrp->args.inode = ino;
874 lrp->args.layout = lo;
875 lrp->clp = NFS_SERVER(ino)->nfs_client;
876 lrp->cred = lo->plh_lc_cred;
877
878 status = nfs4_proc_layoutreturn(lrp);
879out:
880 dprintk("<-- %s status: %d\n", __func__, status);
881 return status;
882}
883EXPORT_SYMBOL_GPL(_pnfs_return_layout);
884
885int
886pnfs_commit_and_return_layout(struct inode *inode)
887{
888 struct pnfs_layout_hdr *lo;
889 int ret;
890
891 spin_lock(&inode->i_lock);
892 lo = NFS_I(inode)->layout;
893 if (lo == NULL) {
894 spin_unlock(&inode->i_lock);
895 return 0;
896 }
897 pnfs_get_layout_hdr(lo);
898 /* Block new layoutgets and read/write to ds */
899 lo->plh_block_lgets++;
900 spin_unlock(&inode->i_lock);
901 filemap_fdatawait(inode->i_mapping);
902 ret = pnfs_layoutcommit_inode(inode, true);
903 if (ret == 0)
904 ret = _pnfs_return_layout(inode);
905 spin_lock(&inode->i_lock);
906 lo->plh_block_lgets--;
907 spin_unlock(&inode->i_lock);
908 pnfs_put_layout_hdr(lo);
909 return ret;
910}
911
912bool pnfs_roc(struct inode *ino)
913{
914 struct pnfs_layout_hdr *lo;
915 struct pnfs_layout_segment *lseg, *tmp;
916 LIST_HEAD(tmp_list);
917 bool found = false;
918
919 spin_lock(&ino->i_lock);
920 lo = NFS_I(ino)->layout;
921 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
922 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
923 goto out_nolayout;
924 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
925 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
926 mark_lseg_invalid(lseg, &tmp_list);
927 found = true;
928 }
929 if (!found)
930 goto out_nolayout;
931 lo->plh_block_lgets++;
932 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
933 spin_unlock(&ino->i_lock);
934 pnfs_free_lseg_list(&tmp_list);
935 return true;
936
937out_nolayout:
938 spin_unlock(&ino->i_lock);
939 return false;
940}
941
942void pnfs_roc_release(struct inode *ino)
943{
944 struct pnfs_layout_hdr *lo;
945
946 spin_lock(&ino->i_lock);
947 lo = NFS_I(ino)->layout;
948 lo->plh_block_lgets--;
949 if (atomic_dec_and_test(&lo->plh_refcount)) {
950 pnfs_detach_layout_hdr(lo);
951 spin_unlock(&ino->i_lock);
952 pnfs_free_layout_hdr(lo);
953 } else
954 spin_unlock(&ino->i_lock);
955}
956
957void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
958{
959 struct pnfs_layout_hdr *lo;
960
961 spin_lock(&ino->i_lock);
962 lo = NFS_I(ino)->layout;
963 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
964 lo->plh_barrier = barrier;
965 spin_unlock(&ino->i_lock);
966}
967
968bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
969{
970 struct nfs_inode *nfsi = NFS_I(ino);
971 struct pnfs_layout_hdr *lo;
972 struct pnfs_layout_segment *lseg;
973 u32 current_seqid;
974 bool found = false;
975
976 spin_lock(&ino->i_lock);
977 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
978 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
979 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
980 found = true;
981 goto out;
982 }
983 lo = nfsi->layout;
984 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
985
986 /* Since close does not return a layout stateid for use as
987 * a barrier, we choose the worst-case barrier.
988 */
989 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
990out:
991 spin_unlock(&ino->i_lock);
992 return found;
993}
994
995/*
996 * Compare two layout segments for sorting into layout cache.
997 * We want to preferentially return RW over RO layouts, so ensure those
998 * are seen first.
999 */
1000static s64
1001pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1002 const struct pnfs_layout_range *l2)
1003{
1004 s64 d;
1005
1006 /* high offset > low offset */
1007 d = l1->offset - l2->offset;
1008 if (d)
1009 return d;
1010
1011 /* short length > long length */
1012 d = l2->length - l1->length;
1013 if (d)
1014 return d;
1015
1016 /* read > read/write */
1017 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1018}
1019
1020static void
1021pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1022 struct pnfs_layout_segment *lseg)
1023{
1024 struct pnfs_layout_segment *lp;
1025
1026 dprintk("%s:Begin\n", __func__);
1027
1028 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1029 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1030 continue;
1031 list_add_tail(&lseg->pls_list, &lp->pls_list);
1032 dprintk("%s: inserted lseg %p "
1033 "iomode %d offset %llu length %llu before "
1034 "lp %p iomode %d offset %llu length %llu\n",
1035 __func__, lseg, lseg->pls_range.iomode,
1036 lseg->pls_range.offset, lseg->pls_range.length,
1037 lp, lp->pls_range.iomode, lp->pls_range.offset,
1038 lp->pls_range.length);
1039 goto out;
1040 }
1041 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1042 dprintk("%s: inserted lseg %p "
1043 "iomode %d offset %llu length %llu at tail\n",
1044 __func__, lseg, lseg->pls_range.iomode,
1045 lseg->pls_range.offset, lseg->pls_range.length);
1046out:
1047 pnfs_get_layout_hdr(lo);
1048
1049 dprintk("%s:Return\n", __func__);
1050}
1051
1052static struct pnfs_layout_hdr *
1053alloc_init_layout_hdr(struct inode *ino,
1054 struct nfs_open_context *ctx,
1055 gfp_t gfp_flags)
1056{
1057 struct pnfs_layout_hdr *lo;
1058
1059 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1060 if (!lo)
1061 return NULL;
1062 atomic_set(&lo->plh_refcount, 1);
1063 INIT_LIST_HEAD(&lo->plh_layouts);
1064 INIT_LIST_HEAD(&lo->plh_segs);
1065 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1066 lo->plh_inode = ino;
1067 lo->plh_lc_cred = get_rpccred(ctx->cred);
1068 return lo;
1069}
1070
1071static struct pnfs_layout_hdr *
1072pnfs_find_alloc_layout(struct inode *ino,
1073 struct nfs_open_context *ctx,
1074 gfp_t gfp_flags)
1075{
1076 struct nfs_inode *nfsi = NFS_I(ino);
1077 struct pnfs_layout_hdr *new = NULL;
1078
1079 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1080
1081 if (nfsi->layout != NULL)
1082 goto out_existing;
1083 spin_unlock(&ino->i_lock);
1084 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1085 spin_lock(&ino->i_lock);
1086
1087 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1088 nfsi->layout = new;
1089 return new;
1090 } else if (new != NULL)
1091 pnfs_free_layout_hdr(new);
1092out_existing:
1093 pnfs_get_layout_hdr(nfsi->layout);
1094 return nfsi->layout;
1095}
1096
1097/*
1098 * iomode matching rules:
1099 * iomode lseg match
1100 * ----- ----- -----
1101 * ANY READ true
1102 * ANY RW true
1103 * RW READ false
1104 * RW RW true
1105 * READ READ true
1106 * READ RW true
1107 */
1108static bool
1109pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1110 const struct pnfs_layout_range *range)
1111{
1112 struct pnfs_layout_range range1;
1113
1114 if ((range->iomode == IOMODE_RW &&
1115 ls_range->iomode != IOMODE_RW) ||
1116 !pnfs_lseg_range_intersecting(ls_range, range))
1117 return 0;
1118
1119 /* range1 covers only the first byte in the range */
1120 range1 = *range;
1121 range1.length = 1;
1122 return pnfs_lseg_range_contained(ls_range, &range1);
1123}
1124
1125/*
1126 * lookup range in layout
1127 */
1128static struct pnfs_layout_segment *
1129pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1130 struct pnfs_layout_range *range)
1131{
1132 struct pnfs_layout_segment *lseg, *ret = NULL;
1133
1134 dprintk("%s:Begin\n", __func__);
1135
1136 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1137 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1138 pnfs_lseg_range_match(&lseg->pls_range, range)) {
1139 ret = pnfs_get_lseg(lseg);
1140 break;
1141 }
1142 if (lseg->pls_range.offset > range->offset)
1143 break;
1144 }
1145
1146 dprintk("%s:Return lseg %p ref %d\n",
1147 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1148 return ret;
1149}
1150
1151/*
1152 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1153 * to the MDS or over pNFS
1154 *
1155 * The nfs_inode read_io and write_io fields are cumulative counters reset
1156 * when there are no layout segments. Note that in pnfs_update_layout iomode
1157 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1158 * WRITE request.
1159 *
1160 * A return of true means use MDS I/O.
1161 *
1162 * From rfc 5661:
1163 * If a file's size is smaller than the file size threshold, data accesses
1164 * SHOULD be sent to the metadata server. If an I/O request has a length that
1165 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1166 * server. If both file size and I/O size are provided, the client SHOULD
1167 * reach or exceed both thresholds before sending its read or write
1168 * requests to the data server.
1169 */
1170static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1171 struct inode *ino, int iomode)
1172{
1173 struct nfs4_threshold *t = ctx->mdsthreshold;
1174 struct nfs_inode *nfsi = NFS_I(ino);
1175 loff_t fsize = i_size_read(ino);
1176 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1177
1178 if (t == NULL)
1179 return ret;
1180
1181 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1182 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1183
1184 switch (iomode) {
1185 case IOMODE_READ:
1186 if (t->bm & THRESHOLD_RD) {
1187 dprintk("%s fsize %llu\n", __func__, fsize);
1188 size_set = true;
1189 if (fsize < t->rd_sz)
1190 size = true;
1191 }
1192 if (t->bm & THRESHOLD_RD_IO) {
1193 dprintk("%s nfsi->read_io %llu\n", __func__,
1194 nfsi->read_io);
1195 io_set = true;
1196 if (nfsi->read_io < t->rd_io_sz)
1197 io = true;
1198 }
1199 break;
1200 case IOMODE_RW:
1201 if (t->bm & THRESHOLD_WR) {
1202 dprintk("%s fsize %llu\n", __func__, fsize);
1203 size_set = true;
1204 if (fsize < t->wr_sz)
1205 size = true;
1206 }
1207 if (t->bm & THRESHOLD_WR_IO) {
1208 dprintk("%s nfsi->write_io %llu\n", __func__,
1209 nfsi->write_io);
1210 io_set = true;
1211 if (nfsi->write_io < t->wr_io_sz)
1212 io = true;
1213 }
1214 break;
1215 }
1216 if (size_set && io_set) {
1217 if (size && io)
1218 ret = true;
1219 } else if (size || io)
1220 ret = true;
1221
1222 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1223 return ret;
1224}
1225
1226/*
1227 * Layout segment is retreived from the server if not cached.
1228 * The appropriate layout segment is referenced and returned to the caller.
1229 */
1230struct pnfs_layout_segment *
1231pnfs_update_layout(struct inode *ino,
1232 struct nfs_open_context *ctx,
1233 loff_t pos,
1234 u64 count,
1235 enum pnfs_iomode iomode,
1236 gfp_t gfp_flags)
1237{
1238 struct pnfs_layout_range arg = {
1239 .iomode = iomode,
1240 .offset = pos,
1241 .length = count,
1242 };
1243 unsigned pg_offset;
1244 struct nfs_server *server = NFS_SERVER(ino);
1245 struct nfs_client *clp = server->nfs_client;
1246 struct pnfs_layout_hdr *lo;
1247 struct pnfs_layout_segment *lseg = NULL;
1248 bool first;
1249
1250 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1251 goto out;
1252
1253 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1254 goto out;
1255
1256 spin_lock(&ino->i_lock);
1257 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1258 if (lo == NULL) {
1259 spin_unlock(&ino->i_lock);
1260 goto out;
1261 }
1262
1263 /* Do we even need to bother with this? */
1264 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1265 dprintk("%s matches recall, use MDS\n", __func__);
1266 goto out_unlock;
1267 }
1268
1269 /* if LAYOUTGET already failed once we don't try again */
1270 if (pnfs_layout_io_test_failed(lo, iomode))
1271 goto out_unlock;
1272
1273 /* Check to see if the layout for the given range already exists */
1274 lseg = pnfs_find_lseg(lo, &arg);
1275 if (lseg)
1276 goto out_unlock;
1277
1278 if (pnfs_layoutgets_blocked(lo, 0))
1279 goto out_unlock;
1280 atomic_inc(&lo->plh_outstanding);
1281
1282 first = list_empty(&lo->plh_layouts) ? true : false;
1283 spin_unlock(&ino->i_lock);
1284
1285 if (first) {
1286 /* The lo must be on the clp list if there is any
1287 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1288 */
1289 spin_lock(&clp->cl_lock);
1290 list_add_tail(&lo->plh_layouts, &server->layouts);
1291 spin_unlock(&clp->cl_lock);
1292 }
1293
1294 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1295 if (pg_offset) {
1296 arg.offset -= pg_offset;
1297 arg.length += pg_offset;
1298 }
1299 if (arg.length != NFS4_MAX_UINT64)
1300 arg.length = PAGE_CACHE_ALIGN(arg.length);
1301
1302 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1303 atomic_dec(&lo->plh_outstanding);
1304out_put_layout_hdr:
1305 pnfs_put_layout_hdr(lo);
1306out:
1307 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1308 "(%s, offset: %llu, length: %llu)\n",
1309 __func__, ino->i_sb->s_id,
1310 (unsigned long long)NFS_FILEID(ino),
1311 lseg == NULL ? "not found" : "found",
1312 iomode==IOMODE_RW ? "read/write" : "read-only",
1313 (unsigned long long)pos,
1314 (unsigned long long)count);
1315 return lseg;
1316out_unlock:
1317 spin_unlock(&ino->i_lock);
1318 goto out_put_layout_hdr;
1319}
1320EXPORT_SYMBOL_GPL(pnfs_update_layout);
1321
1322struct pnfs_layout_segment *
1323pnfs_layout_process(struct nfs4_layoutget *lgp)
1324{
1325 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1326 struct nfs4_layoutget_res *res = &lgp->res;
1327 struct pnfs_layout_segment *lseg;
1328 struct inode *ino = lo->plh_inode;
1329 LIST_HEAD(free_me);
1330 int status = 0;
1331
1332 /* Inject layout blob into I/O device driver */
1333 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1334 if (!lseg || IS_ERR(lseg)) {
1335 if (!lseg)
1336 status = -ENOMEM;
1337 else
1338 status = PTR_ERR(lseg);
1339 dprintk("%s: Could not allocate layout: error %d\n",
1340 __func__, status);
1341 goto out;
1342 }
1343
1344 spin_lock(&ino->i_lock);
1345 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1346 dprintk("%s forget reply due to recall\n", __func__);
1347 goto out_forget_reply;
1348 }
1349
1350 if (pnfs_layoutgets_blocked(lo, 1) ||
1351 pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1352 dprintk("%s forget reply due to state\n", __func__);
1353 goto out_forget_reply;
1354 }
1355
1356 /* Check that the new stateid matches the old stateid */
1357 pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
1358 /* Done processing layoutget. Set the layout stateid */
1359 pnfs_set_layout_stateid(lo, &res->stateid, false);
1360
1361 init_lseg(lo, lseg);
1362 lseg->pls_range = res->range;
1363 pnfs_get_lseg(lseg);
1364 pnfs_layout_insert_lseg(lo, lseg);
1365
1366 if (res->return_on_close) {
1367 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1368 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1369 }
1370
1371 spin_unlock(&ino->i_lock);
1372 pnfs_free_lseg_list(&free_me);
1373 return lseg;
1374out:
1375 return ERR_PTR(status);
1376
1377out_forget_reply:
1378 spin_unlock(&ino->i_lock);
1379 lseg->pls_layout = lo;
1380 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1381 goto out;
1382}
1383
1384void
1385pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1386{
1387 u64 rd_size = req->wb_bytes;
1388
1389 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1390
1391 if (req->wb_offset != req->wb_pgbase) {
1392 nfs_pageio_reset_read_mds(pgio);
1393 return;
1394 }
1395
1396 if (pgio->pg_dreq == NULL)
1397 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1398 else
1399 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1400
1401 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1402 req->wb_context,
1403 req_offset(req),
1404 rd_size,
1405 IOMODE_READ,
1406 GFP_KERNEL);
1407 /* If no lseg, fall back to read through mds */
1408 if (pgio->pg_lseg == NULL)
1409 nfs_pageio_reset_read_mds(pgio);
1410
1411}
1412EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1413
1414void
1415pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1416 struct nfs_page *req, u64 wb_size)
1417{
1418 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1419
1420 if (req->wb_offset != req->wb_pgbase) {
1421 nfs_pageio_reset_write_mds(pgio);
1422 return;
1423 }
1424
1425 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1426 req->wb_context,
1427 req_offset(req),
1428 wb_size,
1429 IOMODE_RW,
1430 GFP_NOFS);
1431 /* If no lseg, fall back to write through mds */
1432 if (pgio->pg_lseg == NULL)
1433 nfs_pageio_reset_write_mds(pgio);
1434}
1435EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1436
1437void
1438pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1439 const struct nfs_pgio_completion_ops *compl_ops)
1440{
1441 struct nfs_server *server = NFS_SERVER(inode);
1442 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1443
1444 if (ld == NULL)
1445 nfs_pageio_init_read(pgio, inode, compl_ops);
1446 else
1447 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1448}
1449
1450void
1451pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1452 int ioflags,
1453 const struct nfs_pgio_completion_ops *compl_ops)
1454{
1455 struct nfs_server *server = NFS_SERVER(inode);
1456 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1457
1458 if (ld == NULL)
1459 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1460 else
1461 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1462}
1463
1464bool
1465pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1466 struct nfs_page *req)
1467{
1468 if (pgio->pg_lseg == NULL)
1469 return nfs_generic_pg_test(pgio, prev, req);
1470
1471 /*
1472 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1473 * Note that this test makes several assumptions:
1474 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1475 * is known to lie within the range.
1476 * - that the nfs_page being tested is known to be contiguous with the
1477 * previous nfs_page.
1478 * - Layout ranges are page aligned, so we only have to test the
1479 * start offset of the request.
1480 *
1481 * Please also note that 'end_offset' is actually the offset of the
1482 * first byte that lies outside the pnfs_layout_range. FIXME?
1483 *
1484 */
1485 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1486 pgio->pg_lseg->pls_range.length);
1487}
1488EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1489
1490int pnfs_write_done_resend_to_mds(struct inode *inode,
1491 struct list_head *head,
1492 const struct nfs_pgio_completion_ops *compl_ops,
1493 struct nfs_direct_req *dreq)
1494{
1495 struct nfs_pageio_descriptor pgio;
1496 LIST_HEAD(failed);
1497
1498 /* Resend all requests through the MDS */
1499 nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1500 pgio.pg_dreq = dreq;
1501 while (!list_empty(head)) {
1502 struct nfs_page *req = nfs_list_entry(head->next);
1503
1504 nfs_list_remove_request(req);
1505 if (!nfs_pageio_add_request(&pgio, req))
1506 nfs_list_add_request(req, &failed);
1507 }
1508 nfs_pageio_complete(&pgio);
1509
1510 if (!list_empty(&failed)) {
1511 /* For some reason our attempt to resend pages. Mark the
1512 * overall send request as having failed, and let
1513 * nfs_writeback_release_full deal with the error.
1514 */
1515 list_move(&failed, head);
1516 return -EIO;
1517 }
1518 return 0;
1519}
1520EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1521
1522static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1523{
1524 struct nfs_pgio_header *hdr = data->header;
1525
1526 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1527 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1528 PNFS_LAYOUTRET_ON_ERROR) {
1529 pnfs_return_layout(hdr->inode);
1530 }
1531 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1532 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1533 &hdr->pages,
1534 hdr->completion_ops,
1535 hdr->dreq);
1536}
1537
1538/*
1539 * Called by non rpc-based layout drivers
1540 */
1541void pnfs_ld_write_done(struct nfs_write_data *data)
1542{
1543 struct nfs_pgio_header *hdr = data->header;
1544
1545 trace_nfs4_pnfs_write(data, hdr->pnfs_error);
1546 if (!hdr->pnfs_error) {
1547 pnfs_set_layoutcommit(data);
1548 hdr->mds_ops->rpc_call_done(&data->task, data);
1549 } else
1550 pnfs_ld_handle_write_error(data);
1551 hdr->mds_ops->rpc_release(data);
1552}
1553EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1554
1555static void
1556pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1557 struct nfs_write_data *data)
1558{
1559 struct nfs_pgio_header *hdr = data->header;
1560
1561 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1562 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1563 nfs_pageio_reset_write_mds(desc);
1564 desc->pg_recoalesce = 1;
1565 }
1566 nfs_writedata_release(data);
1567}
1568
1569static enum pnfs_try_status
1570pnfs_try_to_write_data(struct nfs_write_data *wdata,
1571 const struct rpc_call_ops *call_ops,
1572 struct pnfs_layout_segment *lseg,
1573 int how)
1574{
1575 struct nfs_pgio_header *hdr = wdata->header;
1576 struct inode *inode = hdr->inode;
1577 enum pnfs_try_status trypnfs;
1578 struct nfs_server *nfss = NFS_SERVER(inode);
1579
1580 hdr->mds_ops = call_ops;
1581
1582 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1583 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1584 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1585 if (trypnfs != PNFS_NOT_ATTEMPTED)
1586 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1587 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1588 return trypnfs;
1589}
1590
1591static void
1592pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1593{
1594 struct nfs_write_data *data;
1595 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1596 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1597
1598 desc->pg_lseg = NULL;
1599 while (!list_empty(head)) {
1600 enum pnfs_try_status trypnfs;
1601
1602 data = list_first_entry(head, struct nfs_write_data, list);
1603 list_del_init(&data->list);
1604
1605 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1606 if (trypnfs == PNFS_NOT_ATTEMPTED)
1607 pnfs_write_through_mds(desc, data);
1608 }
1609 pnfs_put_lseg(lseg);
1610}
1611
1612static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1613{
1614 pnfs_put_lseg(hdr->lseg);
1615 nfs_writehdr_free(hdr);
1616}
1617EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1618
1619int
1620pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1621{
1622 struct nfs_write_header *whdr;
1623 struct nfs_pgio_header *hdr;
1624 int ret;
1625
1626 whdr = nfs_writehdr_alloc();
1627 if (!whdr) {
1628 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1629 pnfs_put_lseg(desc->pg_lseg);
1630 desc->pg_lseg = NULL;
1631 return -ENOMEM;
1632 }
1633 hdr = &whdr->header;
1634 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1635 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1636 atomic_inc(&hdr->refcnt);
1637 ret = nfs_generic_flush(desc, hdr);
1638 if (ret != 0) {
1639 pnfs_put_lseg(desc->pg_lseg);
1640 desc->pg_lseg = NULL;
1641 } else
1642 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1643 if (atomic_dec_and_test(&hdr->refcnt))
1644 hdr->completion_ops->completion(hdr);
1645 return ret;
1646}
1647EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1648
1649int pnfs_read_done_resend_to_mds(struct inode *inode,
1650 struct list_head *head,
1651 const struct nfs_pgio_completion_ops *compl_ops,
1652 struct nfs_direct_req *dreq)
1653{
1654 struct nfs_pageio_descriptor pgio;
1655 LIST_HEAD(failed);
1656
1657 /* Resend all requests through the MDS */
1658 nfs_pageio_init_read(&pgio, inode, compl_ops);
1659 pgio.pg_dreq = dreq;
1660 while (!list_empty(head)) {
1661 struct nfs_page *req = nfs_list_entry(head->next);
1662
1663 nfs_list_remove_request(req);
1664 if (!nfs_pageio_add_request(&pgio, req))
1665 nfs_list_add_request(req, &failed);
1666 }
1667 nfs_pageio_complete(&pgio);
1668
1669 if (!list_empty(&failed)) {
1670 list_move(&failed, head);
1671 return -EIO;
1672 }
1673 return 0;
1674}
1675EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1676
1677static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1678{
1679 struct nfs_pgio_header *hdr = data->header;
1680
1681 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1682 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1683 PNFS_LAYOUTRET_ON_ERROR) {
1684 pnfs_return_layout(hdr->inode);
1685 }
1686 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1687 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1688 &hdr->pages,
1689 hdr->completion_ops,
1690 hdr->dreq);
1691}
1692
1693/*
1694 * Called by non rpc-based layout drivers
1695 */
1696void pnfs_ld_read_done(struct nfs_read_data *data)
1697{
1698 struct nfs_pgio_header *hdr = data->header;
1699
1700 trace_nfs4_pnfs_read(data, hdr->pnfs_error);
1701 if (likely(!hdr->pnfs_error)) {
1702 __nfs4_read_done_cb(data);
1703 hdr->mds_ops->rpc_call_done(&data->task, data);
1704 } else
1705 pnfs_ld_handle_read_error(data);
1706 hdr->mds_ops->rpc_release(data);
1707}
1708EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1709
1710static void
1711pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1712 struct nfs_read_data *data)
1713{
1714 struct nfs_pgio_header *hdr = data->header;
1715
1716 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1717 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1718 nfs_pageio_reset_read_mds(desc);
1719 desc->pg_recoalesce = 1;
1720 }
1721 nfs_readdata_release(data);
1722}
1723
1724/*
1725 * Call the appropriate parallel I/O subsystem read function.
1726 */
1727static enum pnfs_try_status
1728pnfs_try_to_read_data(struct nfs_read_data *rdata,
1729 const struct rpc_call_ops *call_ops,
1730 struct pnfs_layout_segment *lseg)
1731{
1732 struct nfs_pgio_header *hdr = rdata->header;
1733 struct inode *inode = hdr->inode;
1734 struct nfs_server *nfss = NFS_SERVER(inode);
1735 enum pnfs_try_status trypnfs;
1736
1737 hdr->mds_ops = call_ops;
1738
1739 dprintk("%s: Reading ino:%lu %u@%llu\n",
1740 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1741
1742 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1743 if (trypnfs != PNFS_NOT_ATTEMPTED)
1744 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1745 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1746 return trypnfs;
1747}
1748
1749static void
1750pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1751{
1752 struct nfs_read_data *data;
1753 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1754 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1755
1756 desc->pg_lseg = NULL;
1757 while (!list_empty(head)) {
1758 enum pnfs_try_status trypnfs;
1759
1760 data = list_first_entry(head, struct nfs_read_data, list);
1761 list_del_init(&data->list);
1762
1763 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1764 if (trypnfs == PNFS_NOT_ATTEMPTED)
1765 pnfs_read_through_mds(desc, data);
1766 }
1767 pnfs_put_lseg(lseg);
1768}
1769
1770static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1771{
1772 pnfs_put_lseg(hdr->lseg);
1773 nfs_readhdr_free(hdr);
1774}
1775EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1776
1777int
1778pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1779{
1780 struct nfs_read_header *rhdr;
1781 struct nfs_pgio_header *hdr;
1782 int ret;
1783
1784 rhdr = nfs_readhdr_alloc();
1785 if (!rhdr) {
1786 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1787 ret = -ENOMEM;
1788 pnfs_put_lseg(desc->pg_lseg);
1789 desc->pg_lseg = NULL;
1790 return ret;
1791 }
1792 hdr = &rhdr->header;
1793 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1794 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1795 atomic_inc(&hdr->refcnt);
1796 ret = nfs_generic_pagein(desc, hdr);
1797 if (ret != 0) {
1798 pnfs_put_lseg(desc->pg_lseg);
1799 desc->pg_lseg = NULL;
1800 } else
1801 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1802 if (atomic_dec_and_test(&hdr->refcnt))
1803 hdr->completion_ops->completion(hdr);
1804 return ret;
1805}
1806EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1807
1808static void pnfs_clear_layoutcommitting(struct inode *inode)
1809{
1810 unsigned long *bitlock = &NFS_I(inode)->flags;
1811
1812 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1813 smp_mb__after_clear_bit();
1814 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1815}
1816
1817/*
1818 * There can be multiple RW segments.
1819 */
1820static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1821{
1822 struct pnfs_layout_segment *lseg;
1823
1824 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1825 if (lseg->pls_range.iomode == IOMODE_RW &&
1826 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1827 list_add(&lseg->pls_lc_list, listp);
1828 }
1829}
1830
1831static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1832{
1833 struct pnfs_layout_segment *lseg, *tmp;
1834
1835 /* Matched by references in pnfs_set_layoutcommit */
1836 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1837 list_del_init(&lseg->pls_lc_list);
1838 pnfs_put_lseg(lseg);
1839 }
1840
1841 pnfs_clear_layoutcommitting(inode);
1842}
1843
1844void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1845{
1846 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1847}
1848EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1849
1850void
1851pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1852{
1853 struct nfs_pgio_header *hdr = wdata->header;
1854 struct inode *inode = hdr->inode;
1855 struct nfs_inode *nfsi = NFS_I(inode);
1856 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1857 bool mark_as_dirty = false;
1858
1859 spin_lock(&inode->i_lock);
1860 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1861 mark_as_dirty = true;
1862 dprintk("%s: Set layoutcommit for inode %lu ",
1863 __func__, inode->i_ino);
1864 }
1865 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1866 /* references matched in nfs4_layoutcommit_release */
1867 pnfs_get_lseg(hdr->lseg);
1868 }
1869 if (end_pos > nfsi->layout->plh_lwb)
1870 nfsi->layout->plh_lwb = end_pos;
1871 spin_unlock(&inode->i_lock);
1872 dprintk("%s: lseg %p end_pos %llu\n",
1873 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1874
1875 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1876 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1877 if (mark_as_dirty)
1878 mark_inode_dirty_sync(inode);
1879}
1880EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1881
1882void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1883{
1884 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1885
1886 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1887 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1888 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1889}
1890
1891/*
1892 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1893 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1894 * data to disk to allow the server to recover the data if it crashes.
1895 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1896 * is off, and a COMMIT is sent to a data server, or
1897 * if WRITEs to a data server return NFS_DATA_SYNC.
1898 */
1899int
1900pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1901{
1902 struct nfs4_layoutcommit_data *data;
1903 struct nfs_inode *nfsi = NFS_I(inode);
1904 loff_t end_pos;
1905 int status;
1906
1907 if (!pnfs_layoutcommit_outstanding(inode))
1908 return 0;
1909
1910 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1911
1912 status = -EAGAIN;
1913 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1914 if (!sync)
1915 goto out;
1916 status = wait_on_bit_lock(&nfsi->flags,
1917 NFS_INO_LAYOUTCOMMITTING,
1918 nfs_wait_bit_killable,
1919 TASK_KILLABLE);
1920 if (status)
1921 goto out;
1922 }
1923
1924 status = -ENOMEM;
1925 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1926 data = kzalloc(sizeof(*data), GFP_NOFS);
1927 if (!data)
1928 goto clear_layoutcommitting;
1929
1930 status = 0;
1931 spin_lock(&inode->i_lock);
1932 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1933 goto out_unlock;
1934
1935 INIT_LIST_HEAD(&data->lseg_list);
1936 pnfs_list_write_lseg(inode, &data->lseg_list);
1937
1938 end_pos = nfsi->layout->plh_lwb;
1939 nfsi->layout->plh_lwb = 0;
1940
1941 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1942 spin_unlock(&inode->i_lock);
1943
1944 data->args.inode = inode;
1945 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1946 nfs_fattr_init(&data->fattr);
1947 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1948 data->res.fattr = &data->fattr;
1949 data->args.lastbytewritten = end_pos - 1;
1950 data->res.server = NFS_SERVER(inode);
1951
1952 status = nfs4_proc_layoutcommit(data, sync);
1953out:
1954 if (status)
1955 mark_inode_dirty_sync(inode);
1956 dprintk("<-- %s status %d\n", __func__, status);
1957 return status;
1958out_unlock:
1959 spin_unlock(&inode->i_lock);
1960 kfree(data);
1961clear_layoutcommitting:
1962 pnfs_clear_layoutcommitting(inode);
1963 goto out;
1964}
1965
1966struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1967{
1968 struct nfs4_threshold *thp;
1969
1970 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1971 if (!thp) {
1972 dprintk("%s mdsthreshold allocation failed\n", __func__);
1973 return NULL;
1974 }
1975 return thp;
1976}