2 * pNFS functions to call and manage layout drivers.
4 * Copyright (c) 2002 [year of first publication]
5 * The Regents of the University of Michigan
8 * Dean Hildebrand <dhildebz@umich.edu>
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.
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
30 #include <linux/nfs_fs.h>
31 #include <linux/nfs_page.h>
32 #include <linux/module.h>
33 #include <linux/sort.h>
37 #include "nfs4trace.h"
38 #include "delegation.h"
41 #define NFSDBG_FACILITY NFSDBG_PNFS
42 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
47 * protects pnfs_modules_tbl.
49 static DEFINE_SPINLOCK(pnfs_spinlock);
52 * pnfs_modules_tbl holds all pnfs modules
54 static LIST_HEAD(pnfs_modules_tbl);
56 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
57 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
58 struct list_head *free_me,
59 const struct pnfs_layout_range *range,
61 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
62 struct list_head *tmp_list);
64 /* Return the registered pnfs layout driver module matching given id */
65 static struct pnfs_layoutdriver_type *
66 find_pnfs_driver_locked(u32 id)
68 struct pnfs_layoutdriver_type *local;
70 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
75 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
79 static struct pnfs_layoutdriver_type *
80 find_pnfs_driver(u32 id)
82 struct pnfs_layoutdriver_type *local;
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__);
90 spin_unlock(&pnfs_spinlock);
95 unset_pnfs_layoutdriver(struct nfs_server *nfss)
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);
105 nfss->pnfs_curr_ld = NULL;
109 * When the server sends a list of layout types, we choose one in the order
110 * given in the list below.
112 * FIXME: should this list be configurable in some fashion? module param?
113 * mount option? something else?
115 static const u32 ld_prefs[] = {
120 LAYOUT_NFSV4_1_FILES,
125 ld_cmp(const void *e1, const void *e2)
127 u32 ld1 = *((u32 *)e1);
128 u32 ld2 = *((u32 *)e2);
131 for (i = 0; ld_prefs[i] != 0; i++) {
132 if (ld1 == ld_prefs[i])
135 if (ld2 == ld_prefs[i])
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.
145 * @ids array of layout types supported by MDS.
148 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
149 struct nfs_fsinfo *fsinfo)
151 struct pnfs_layoutdriver_type *ld_type = NULL;
155 if (fsinfo->nlayouttypes == 0)
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);
164 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
165 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
167 for (i = 0; i < fsinfo->nlayouttypes; i++) {
168 id = fsinfo->layouttype[i];
169 ld_type = find_pnfs_driver(id);
171 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
173 ld_type = find_pnfs_driver(id);
180 dprintk("%s: No pNFS module found!\n", __func__);
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);
192 /* Bump the MDS count */
193 atomic_inc(&server->nfs_client->cl_mds_count);
195 dprintk("%s: pNFS module for %u set\n", __func__, id);
199 dprintk("%s: Using NFSv4 I/O\n", __func__);
200 server->pnfs_curr_ld = NULL;
204 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
206 int status = -EINVAL;
207 struct pnfs_layoutdriver_type *tmp;
209 if (ld_type->id == 0) {
210 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
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__);
219 spin_lock(&pnfs_spinlock);
220 tmp = find_pnfs_driver_locked(ld_type->id);
222 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
224 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
227 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
228 __func__, ld_type->id);
230 spin_unlock(&pnfs_spinlock);
234 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
237 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
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);
244 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
247 * pNFS client layout cache
250 /* Need to hold i_lock if caller does not already hold reference */
252 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
254 atomic_inc(&lo->plh_refcount);
257 static struct pnfs_layout_hdr *
258 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
260 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
261 return ld->alloc_layout_hdr(ino, gfp_flags);
265 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
267 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
268 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
270 if (!list_empty(&lo->plh_layouts)) {
271 struct nfs_client *clp = server->nfs_client;
273 spin_lock(&clp->cl_lock);
274 list_del_init(&lo->plh_layouts);
275 spin_unlock(&clp->cl_lock);
277 put_rpccred(lo->plh_lc_cred);
278 return ld->free_layout_hdr(lo);
282 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
284 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
285 dprintk("%s: freeing layout cache %p\n", __func__, lo);
287 /* Reset MDS Threshold I/O counters */
293 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
295 struct inode *inode = lo->plh_inode;
297 pnfs_layoutreturn_before_put_layout_hdr(lo);
299 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
300 if (!list_empty(&lo->plh_segs))
301 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
302 pnfs_detach_layout_hdr(lo);
303 spin_unlock(&inode->i_lock);
304 pnfs_free_layout_hdr(lo);
309 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
312 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
314 lo->plh_return_iomode = iomode;
315 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
317 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
318 lo->plh_return_seq = seq;
323 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
325 lo->plh_return_iomode = 0;
326 lo->plh_return_seq = 0;
327 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
330 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
332 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
333 clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
334 smp_mb__after_atomic();
335 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
336 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
340 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
341 struct list_head *free_me)
343 clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
344 clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
345 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
346 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
347 if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
348 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
352 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
354 * In order to continue using the pnfs_layout_hdr, a full recovery
356 * Note that caller must hold inode->i_lock.
359 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
360 struct list_head *lseg_list)
362 struct pnfs_layout_range range = {
363 .iomode = IOMODE_ANY,
365 .length = NFS4_MAX_UINT64,
367 struct pnfs_layout_segment *lseg, *next;
369 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
370 pnfs_clear_layoutreturn_info(lo);
371 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
372 pnfs_clear_lseg_state(lseg, lseg_list);
373 pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
374 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
375 !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
376 pnfs_clear_layoutreturn_waitbit(lo);
377 return !list_empty(&lo->plh_segs);
381 pnfs_iomode_to_fail_bit(u32 iomode)
383 return iomode == IOMODE_RW ?
384 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
388 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
390 lo->plh_retry_timestamp = jiffies;
391 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
392 atomic_inc(&lo->plh_refcount);
396 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
398 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
399 atomic_dec(&lo->plh_refcount);
403 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
405 struct inode *inode = lo->plh_inode;
406 struct pnfs_layout_range range = {
409 .length = NFS4_MAX_UINT64,
413 spin_lock(&inode->i_lock);
414 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
415 pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
416 spin_unlock(&inode->i_lock);
417 pnfs_free_lseg_list(&head);
418 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
419 iomode == IOMODE_RW ? "RW" : "READ");
423 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
425 unsigned long start, end;
426 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
428 if (test_bit(fail_bit, &lo->plh_flags) == 0)
431 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
432 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
433 /* It is time to retry the failed layoutgets */
434 pnfs_layout_clear_fail_bit(lo, fail_bit);
441 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
442 const struct pnfs_layout_range *range,
443 const nfs4_stateid *stateid)
445 INIT_LIST_HEAD(&lseg->pls_list);
446 INIT_LIST_HEAD(&lseg->pls_lc_list);
447 atomic_set(&lseg->pls_refcount, 1);
448 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
449 lseg->pls_layout = lo;
450 lseg->pls_range = *range;
451 lseg->pls_seq = be32_to_cpu(stateid->seqid);
454 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
457 struct inode *inode = lseg->pls_layout->plh_inode;
458 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
463 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
464 struct pnfs_layout_segment *lseg)
466 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
467 list_del_init(&lseg->pls_list);
468 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
469 atomic_dec(&lo->plh_refcount);
470 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
472 if (list_empty(&lo->plh_segs) &&
473 !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
474 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
475 if (atomic_read(&lo->plh_outstanding) == 0)
476 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
477 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
482 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
483 struct pnfs_layout_segment *lseg)
485 if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
486 pnfs_layout_is_valid(lo)) {
487 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
488 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
495 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
497 struct pnfs_layout_hdr *lo;
503 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
504 atomic_read(&lseg->pls_refcount),
505 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
507 lo = lseg->pls_layout;
508 inode = lo->plh_inode;
510 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
511 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
512 spin_unlock(&inode->i_lock);
515 pnfs_get_layout_hdr(lo);
516 pnfs_layout_remove_lseg(lo, lseg);
517 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
519 spin_unlock(&inode->i_lock);
520 pnfs_free_lseg(lseg);
521 pnfs_put_layout_hdr(lo);
524 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
526 static void pnfs_free_lseg_async_work(struct work_struct *work)
528 struct pnfs_layout_segment *lseg;
529 struct pnfs_layout_hdr *lo;
531 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
532 lo = lseg->pls_layout;
534 pnfs_free_lseg(lseg);
535 pnfs_put_layout_hdr(lo);
538 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
540 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
541 schedule_work(&lseg->pls_work);
545 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
550 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
552 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
553 atomic_read(&lseg->pls_refcount),
554 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
555 if (atomic_dec_and_test(&lseg->pls_refcount)) {
556 struct pnfs_layout_hdr *lo = lseg->pls_layout;
557 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
559 pnfs_layout_remove_lseg(lo, lseg);
560 if (!pnfs_cache_lseg_for_layoutreturn(lo, lseg)) {
561 pnfs_get_layout_hdr(lo);
562 pnfs_free_lseg_async(lseg);
566 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
569 * is l2 fully contained in l1?
571 * [----------------------------------)
576 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
577 const struct pnfs_layout_range *l2)
579 u64 start1 = l1->offset;
580 u64 end1 = pnfs_end_offset(start1, l1->length);
581 u64 start2 = l2->offset;
582 u64 end2 = pnfs_end_offset(start2, l2->length);
584 return (start1 <= start2) && (end1 >= end2);
587 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
588 struct list_head *tmp_list)
590 if (!atomic_dec_and_test(&lseg->pls_refcount))
592 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
593 list_add(&lseg->pls_list, tmp_list);
597 /* Returns 1 if lseg is removed from list, 0 otherwise */
598 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
599 struct list_head *tmp_list)
603 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
604 /* Remove the reference keeping the lseg in the
605 * list. It will now be removed when all
606 * outstanding io is finished.
608 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
609 atomic_read(&lseg->pls_refcount));
610 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
617 * Compare 2 layout stateid sequence ids, to see which is newer,
618 * taking into account wraparound issues.
620 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
622 return (s32)(s1 - s2) > 0;
626 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
627 const struct pnfs_layout_range *recall_range)
629 return (recall_range->iomode == IOMODE_ANY ||
630 lseg_range->iomode == recall_range->iomode) &&
631 pnfs_lseg_range_intersecting(lseg_range, recall_range);
635 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
636 const struct pnfs_layout_range *recall_range,
639 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
641 if (recall_range == NULL)
643 return pnfs_should_free_range(&lseg->pls_range, recall_range);
647 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
648 * @lo: layout header containing the lsegs
649 * @tmp_list: list head where doomed lsegs should go
650 * @recall_range: optional recall range argument to match (may be NULL)
651 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
653 * Walk the list of lsegs in the layout header, and tear down any that should
654 * be destroyed. If "recall_range" is specified then the segment must match
655 * that range. If "seq" is non-zero, then only match segments that were handed
656 * out at or before that sequence.
658 * Returns number of matching invalid lsegs remaining in list after scanning
659 * it and purging them.
662 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
663 struct list_head *tmp_list,
664 const struct pnfs_layout_range *recall_range,
667 struct pnfs_layout_segment *lseg, *next;
670 dprintk("%s:Begin lo %p\n", __func__, lo);
672 if (list_empty(&lo->plh_segs))
674 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
675 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
676 dprintk("%s: freeing lseg %p iomode %d seq %u"
677 "offset %llu length %llu\n", __func__,
678 lseg, lseg->pls_range.iomode, lseg->pls_seq,
679 lseg->pls_range.offset, lseg->pls_range.length);
680 if (!mark_lseg_invalid(lseg, tmp_list))
683 dprintk("%s:Return %i\n", __func__, remaining);
688 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
689 struct list_head *free_me,
690 const struct pnfs_layout_range *range,
693 struct pnfs_layout_segment *lseg, *next;
695 list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
696 if (pnfs_match_lseg_recall(lseg, range, seq))
697 list_move_tail(&lseg->pls_list, free_me);
701 /* note free_me must contain lsegs from a single layout_hdr */
703 pnfs_free_lseg_list(struct list_head *free_me)
705 struct pnfs_layout_segment *lseg, *tmp;
707 if (list_empty(free_me))
710 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
711 list_del(&lseg->pls_list);
712 pnfs_free_lseg(lseg);
717 pnfs_destroy_layout(struct nfs_inode *nfsi)
719 struct pnfs_layout_hdr *lo;
722 spin_lock(&nfsi->vfs_inode.i_lock);
725 pnfs_get_layout_hdr(lo);
726 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
727 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
728 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
729 spin_unlock(&nfsi->vfs_inode.i_lock);
730 pnfs_free_lseg_list(&tmp_list);
731 pnfs_put_layout_hdr(lo);
733 spin_unlock(&nfsi->vfs_inode.i_lock);
735 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
738 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
739 struct list_head *layout_list)
741 struct pnfs_layout_hdr *lo;
744 spin_lock(&inode->i_lock);
745 lo = NFS_I(inode)->layout;
746 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
747 pnfs_get_layout_hdr(lo);
748 list_add(&lo->plh_bulk_destroy, layout_list);
751 spin_unlock(&inode->i_lock);
755 /* Caller must hold rcu_read_lock and clp->cl_lock */
757 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
758 struct nfs_server *server,
759 struct list_head *layout_list)
761 struct pnfs_layout_hdr *lo, *next;
764 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
765 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
767 inode = igrab(lo->plh_inode);
770 list_del_init(&lo->plh_layouts);
771 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
774 spin_unlock(&clp->cl_lock);
776 spin_lock(&clp->cl_lock);
784 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
787 struct pnfs_layout_hdr *lo;
789 LIST_HEAD(lseg_list);
792 while (!list_empty(layout_list)) {
793 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
795 dprintk("%s freeing layout for inode %lu\n", __func__,
796 lo->plh_inode->i_ino);
797 inode = lo->plh_inode;
799 pnfs_layoutcommit_inode(inode, false);
801 spin_lock(&inode->i_lock);
802 list_del_init(&lo->plh_bulk_destroy);
803 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
805 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
808 spin_unlock(&inode->i_lock);
809 pnfs_free_lseg_list(&lseg_list);
810 /* Free all lsegs that are attached to commit buckets */
811 nfs_commit_inode(inode, 0);
812 pnfs_put_layout_hdr(lo);
819 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
820 struct nfs_fsid *fsid,
823 struct nfs_server *server;
824 LIST_HEAD(layout_list);
826 spin_lock(&clp->cl_lock);
829 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
830 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
832 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
838 spin_unlock(&clp->cl_lock);
840 if (list_empty(&layout_list))
842 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
846 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
849 struct nfs_server *server;
850 LIST_HEAD(layout_list);
852 spin_lock(&clp->cl_lock);
855 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
856 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
862 spin_unlock(&clp->cl_lock);
864 if (list_empty(&layout_list))
866 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
870 * Called by the state manger to remove all layouts established under an
874 pnfs_destroy_all_layouts(struct nfs_client *clp)
876 nfs4_deviceid_mark_client_invalid(clp);
877 nfs4_deviceid_purge_client(clp);
879 pnfs_destroy_layouts_byclid(clp, false);
882 /* update lo->plh_stateid with new if is more recent */
884 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
887 u32 oldseq, newseq, new_barrier = 0;
889 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
890 newseq = be32_to_cpu(new->seqid);
892 if (!pnfs_layout_is_valid(lo)) {
893 nfs4_stateid_copy(&lo->plh_stateid, new);
894 lo->plh_barrier = newseq;
895 pnfs_clear_layoutreturn_info(lo);
896 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
899 if (pnfs_seqid_is_newer(newseq, oldseq)) {
900 nfs4_stateid_copy(&lo->plh_stateid, new);
902 * Because of wraparound, we want to keep the barrier
903 * "close" to the current seqids.
905 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
908 new_barrier = be32_to_cpu(new->seqid);
909 else if (new_barrier == 0)
911 if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
912 lo->plh_barrier = new_barrier;
916 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
917 const nfs4_stateid *stateid)
919 u32 seqid = be32_to_cpu(stateid->seqid);
921 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
924 /* lget is set to 1 if called from inside send_layoutget call chain */
926 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
928 return lo->plh_block_lgets ||
929 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
933 * Get layout from server.
934 * for now, assume that whole file layouts are requested.
936 * arg->length: all ones
938 static struct pnfs_layout_segment *
939 send_layoutget(struct pnfs_layout_hdr *lo,
940 struct nfs_open_context *ctx,
941 nfs4_stateid *stateid,
942 const struct pnfs_layout_range *range,
943 long *timeout, gfp_t gfp_flags)
945 struct inode *ino = lo->plh_inode;
946 struct nfs_server *server = NFS_SERVER(ino);
947 struct nfs4_layoutget *lgp;
950 dprintk("--> %s\n", __func__);
953 * Synchronously retrieve layout information from server and
954 * store in lseg. If we race with a concurrent seqid morphing
955 * op, then re-send the LAYOUTGET.
957 lgp = kzalloc(sizeof(*lgp), gfp_flags);
959 return ERR_PTR(-ENOMEM);
961 i_size = i_size_read(ino);
963 lgp->args.minlength = PAGE_SIZE;
964 if (lgp->args.minlength > range->length)
965 lgp->args.minlength = range->length;
966 if (range->iomode == IOMODE_READ) {
967 if (range->offset >= i_size)
968 lgp->args.minlength = 0;
969 else if (i_size - range->offset < lgp->args.minlength)
970 lgp->args.minlength = i_size - range->offset;
972 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
973 pnfs_copy_range(&lgp->args.range, range);
974 lgp->args.type = server->pnfs_curr_ld->id;
975 lgp->args.inode = ino;
976 lgp->args.ctx = get_nfs_open_context(ctx);
977 nfs4_stateid_copy(&lgp->args.stateid, stateid);
978 lgp->gfp_flags = gfp_flags;
979 lgp->cred = lo->plh_lc_cred;
981 return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
984 static void pnfs_clear_layoutcommit(struct inode *inode,
985 struct list_head *head)
987 struct nfs_inode *nfsi = NFS_I(inode);
988 struct pnfs_layout_segment *lseg, *tmp;
990 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
992 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
993 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
995 pnfs_lseg_dec_and_remove_zero(lseg, head);
999 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1000 const nfs4_stateid *arg_stateid,
1001 const struct pnfs_layout_range *range,
1002 const nfs4_stateid *stateid)
1004 struct inode *inode = lo->plh_inode;
1007 spin_lock(&inode->i_lock);
1008 if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1009 !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1012 u32 seq = be32_to_cpu(arg_stateid->seqid);
1014 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1015 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1016 pnfs_set_layout_stateid(lo, stateid, true);
1018 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1020 pnfs_clear_layoutreturn_waitbit(lo);
1021 spin_unlock(&inode->i_lock);
1022 pnfs_free_lseg_list(&freeme);
1027 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1028 nfs4_stateid *stateid,
1029 enum pnfs_iomode *iomode)
1031 /* Serialise LAYOUTGET/LAYOUTRETURN */
1032 if (atomic_read(&lo->plh_outstanding) != 0)
1034 if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1036 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1037 pnfs_get_layout_hdr(lo);
1038 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1039 if (stateid != NULL) {
1040 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1041 if (lo->plh_return_seq != 0)
1042 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1045 *iomode = lo->plh_return_iomode;
1046 pnfs_clear_layoutreturn_info(lo);
1049 if (stateid != NULL)
1050 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1052 *iomode = IOMODE_ANY;
1057 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1058 struct pnfs_layout_hdr *lo,
1059 const nfs4_stateid *stateid,
1060 enum pnfs_iomode iomode)
1062 struct inode *inode = lo->plh_inode;
1064 args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1065 args->inode = inode;
1066 args->range.iomode = iomode;
1067 args->range.offset = 0;
1068 args->range.length = NFS4_MAX_UINT64;
1070 nfs4_stateid_copy(&args->stateid, stateid);
1074 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1075 enum pnfs_iomode iomode, bool sync)
1077 struct inode *ino = lo->plh_inode;
1078 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1079 struct nfs4_layoutreturn *lrp;
1082 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1083 if (unlikely(lrp == NULL)) {
1085 spin_lock(&ino->i_lock);
1086 pnfs_clear_layoutreturn_waitbit(lo);
1087 spin_unlock(&ino->i_lock);
1088 pnfs_put_layout_hdr(lo);
1092 pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1093 lrp->args.ld_private = &lrp->ld_private;
1094 lrp->clp = NFS_SERVER(ino)->nfs_client;
1095 lrp->cred = lo->plh_lc_cred;
1096 if (ld->prepare_layoutreturn)
1097 ld->prepare_layoutreturn(&lrp->args);
1099 status = nfs4_proc_layoutreturn(lrp, sync);
1101 dprintk("<-- %s status: %d\n", __func__, status);
1105 /* Return true if layoutreturn is needed */
1107 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1109 struct pnfs_layout_segment *s;
1111 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1114 /* Defer layoutreturn until all lsegs are done */
1115 list_for_each_entry(s, &lo->plh_segs, pls_list) {
1116 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1123 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1125 struct inode *inode= lo->plh_inode;
1127 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1129 spin_lock(&inode->i_lock);
1130 if (pnfs_layout_need_return(lo)) {
1131 nfs4_stateid stateid;
1132 enum pnfs_iomode iomode;
1135 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1136 spin_unlock(&inode->i_lock);
1138 /* Send an async layoutreturn so we dont deadlock */
1139 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1142 spin_unlock(&inode->i_lock);
1146 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1147 * when the layout segment list is empty.
1149 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1150 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1151 * deviceid is marked invalid.
1154 _pnfs_return_layout(struct inode *ino)
1156 struct pnfs_layout_hdr *lo = NULL;
1157 struct nfs_inode *nfsi = NFS_I(ino);
1158 LIST_HEAD(tmp_list);
1159 nfs4_stateid stateid;
1163 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1165 spin_lock(&ino->i_lock);
1168 spin_unlock(&ino->i_lock);
1169 dprintk("NFS: %s no layout to return\n", __func__);
1172 /* Reference matched in nfs4_layoutreturn_release */
1173 pnfs_get_layout_hdr(lo);
1174 /* Is there an outstanding layoutreturn ? */
1175 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1176 spin_unlock(&ino->i_lock);
1177 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1178 TASK_UNINTERRUPTIBLE))
1179 goto out_put_layout_hdr;
1180 spin_lock(&ino->i_lock);
1182 pnfs_clear_layoutcommit(ino, &tmp_list);
1183 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1185 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1186 struct pnfs_layout_range range = {
1187 .iomode = IOMODE_ANY,
1189 .length = NFS4_MAX_UINT64,
1191 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1194 /* Don't send a LAYOUTRETURN if list was initially empty */
1195 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1196 spin_unlock(&ino->i_lock);
1197 dprintk("NFS: %s no layout segments to return\n", __func__);
1198 goto out_put_layout_hdr;
1201 send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1202 spin_unlock(&ino->i_lock);
1203 pnfs_free_lseg_list(&tmp_list);
1205 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1207 pnfs_put_layout_hdr(lo);
1209 dprintk("<-- %s status: %d\n", __func__, status);
1212 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1215 pnfs_commit_and_return_layout(struct inode *inode)
1217 struct pnfs_layout_hdr *lo;
1220 spin_lock(&inode->i_lock);
1221 lo = NFS_I(inode)->layout;
1223 spin_unlock(&inode->i_lock);
1226 pnfs_get_layout_hdr(lo);
1227 /* Block new layoutgets and read/write to ds */
1228 lo->plh_block_lgets++;
1229 spin_unlock(&inode->i_lock);
1230 filemap_fdatawait(inode->i_mapping);
1231 ret = pnfs_layoutcommit_inode(inode, true);
1233 ret = _pnfs_return_layout(inode);
1234 spin_lock(&inode->i_lock);
1235 lo->plh_block_lgets--;
1236 spin_unlock(&inode->i_lock);
1237 pnfs_put_layout_hdr(lo);
1241 bool pnfs_roc(struct inode *ino,
1242 struct nfs4_layoutreturn_args *args,
1243 struct nfs4_layoutreturn_res *res,
1244 const struct rpc_cred *cred)
1246 struct nfs_inode *nfsi = NFS_I(ino);
1247 struct nfs_open_context *ctx;
1248 struct nfs4_state *state;
1249 struct pnfs_layout_hdr *lo;
1250 struct pnfs_layout_segment *lseg, *next;
1251 nfs4_stateid stateid;
1252 enum pnfs_iomode iomode = 0;
1253 bool layoutreturn = false, roc = false;
1255 if (!nfs_have_layout(ino))
1258 spin_lock(&ino->i_lock);
1260 if (!lo || !pnfs_layout_is_valid(lo) ||
1261 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1263 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1264 pnfs_get_layout_hdr(lo);
1265 spin_unlock(&ino->i_lock);
1266 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1267 TASK_UNINTERRUPTIBLE);
1268 pnfs_put_layout_hdr(lo);
1272 /* no roc if we hold a delegation */
1273 if (nfs4_check_delegation(ino, FMODE_READ))
1276 list_for_each_entry(ctx, &nfsi->open_files, list) {
1278 /* Don't return layout if there is open file state */
1279 if (state != NULL && state->state != 0)
1284 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1285 /* If we are sending layoutreturn, invalidate all valid lsegs */
1286 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1289 * Note: mark lseg for return so pnfs_layout_remove_lseg
1290 * doesn't invalidate the layout for us.
1292 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1293 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1295 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1298 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1301 /* ROC in two conditions:
1302 * 1. there are ROC lsegs
1303 * 2. we don't send layoutreturn
1305 /* lo ref dropped in pnfs_roc_release() */
1306 layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1307 /* If the creds don't match, we can't compound the layoutreturn */
1308 if (!layoutreturn || cred != lo->plh_lc_cred)
1312 pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1313 res->lrs_present = 0;
1314 layoutreturn = false;
1317 spin_unlock(&ino->i_lock);
1318 pnfs_layoutcommit_inode(ino, true);
1320 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1321 if (ld->prepare_layoutreturn)
1322 ld->prepare_layoutreturn(args);
1326 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1330 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1331 struct nfs4_layoutreturn_res *res,
1334 struct pnfs_layout_hdr *lo = args->layout;
1335 const nfs4_stateid *arg_stateid = NULL;
1336 const nfs4_stateid *res_stateid = NULL;
1337 struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1340 arg_stateid = &args->stateid;
1341 if (res->lrs_present)
1342 res_stateid = &res->stateid;
1344 pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1346 if (ld_private && ld_private->ops && ld_private->ops->free)
1347 ld_private->ops->free(ld_private);
1348 pnfs_put_layout_hdr(lo);
1349 trace_nfs4_layoutreturn_on_close(args->inode, 0);
1352 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1354 struct nfs_inode *nfsi = NFS_I(ino);
1355 struct pnfs_layout_hdr *lo;
1358 /* we might not have grabbed lo reference. so need to check under
1360 spin_lock(&ino->i_lock);
1362 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1363 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1366 spin_unlock(&ino->i_lock);
1371 * Compare two layout segments for sorting into layout cache.
1372 * We want to preferentially return RW over RO layouts, so ensure those
1376 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1377 const struct pnfs_layout_range *l2)
1381 /* high offset > low offset */
1382 d = l1->offset - l2->offset;
1386 /* short length > long length */
1387 d = l2->length - l1->length;
1391 /* read > read/write */
1392 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1396 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1397 const struct pnfs_layout_range *l2)
1399 return pnfs_lseg_range_cmp(l1, l2) > 0;
1403 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1404 struct pnfs_layout_segment *old)
1410 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1411 struct pnfs_layout_segment *lseg,
1412 bool (*is_after)(const struct pnfs_layout_range *,
1413 const struct pnfs_layout_range *),
1414 bool (*do_merge)(struct pnfs_layout_segment *,
1415 struct pnfs_layout_segment *),
1416 struct list_head *free_me)
1418 struct pnfs_layout_segment *lp, *tmp;
1420 dprintk("%s:Begin\n", __func__);
1422 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1423 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1425 if (do_merge(lseg, lp)) {
1426 mark_lseg_invalid(lp, free_me);
1429 if (is_after(&lseg->pls_range, &lp->pls_range))
1431 list_add_tail(&lseg->pls_list, &lp->pls_list);
1432 dprintk("%s: inserted lseg %p "
1433 "iomode %d offset %llu length %llu before "
1434 "lp %p iomode %d offset %llu length %llu\n",
1435 __func__, lseg, lseg->pls_range.iomode,
1436 lseg->pls_range.offset, lseg->pls_range.length,
1437 lp, lp->pls_range.iomode, lp->pls_range.offset,
1438 lp->pls_range.length);
1441 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1442 dprintk("%s: inserted lseg %p "
1443 "iomode %d offset %llu length %llu at tail\n",
1444 __func__, lseg, lseg->pls_range.iomode,
1445 lseg->pls_range.offset, lseg->pls_range.length);
1447 pnfs_get_layout_hdr(lo);
1449 dprintk("%s:Return\n", __func__);
1451 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1454 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1455 struct pnfs_layout_segment *lseg,
1456 struct list_head *free_me)
1458 struct inode *inode = lo->plh_inode;
1459 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1461 if (ld->add_lseg != NULL)
1462 ld->add_lseg(lo, lseg, free_me);
1464 pnfs_generic_layout_insert_lseg(lo, lseg,
1465 pnfs_lseg_range_is_after,
1470 static struct pnfs_layout_hdr *
1471 alloc_init_layout_hdr(struct inode *ino,
1472 struct nfs_open_context *ctx,
1475 struct pnfs_layout_hdr *lo;
1477 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1480 atomic_set(&lo->plh_refcount, 1);
1481 INIT_LIST_HEAD(&lo->plh_layouts);
1482 INIT_LIST_HEAD(&lo->plh_segs);
1483 INIT_LIST_HEAD(&lo->plh_return_segs);
1484 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1485 lo->plh_inode = ino;
1486 lo->plh_lc_cred = get_rpccred(ctx->cred);
1487 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1491 static struct pnfs_layout_hdr *
1492 pnfs_find_alloc_layout(struct inode *ino,
1493 struct nfs_open_context *ctx,
1495 __releases(&ino->i_lock)
1496 __acquires(&ino->i_lock)
1498 struct nfs_inode *nfsi = NFS_I(ino);
1499 struct pnfs_layout_hdr *new = NULL;
1501 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1503 if (nfsi->layout != NULL)
1505 spin_unlock(&ino->i_lock);
1506 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1507 spin_lock(&ino->i_lock);
1509 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1512 } else if (new != NULL)
1513 pnfs_free_layout_hdr(new);
1515 pnfs_get_layout_hdr(nfsi->layout);
1516 return nfsi->layout;
1520 * iomode matching rules:
1521 * iomode lseg strict match
1523 * ----- ----- ------ -----
1528 * READ READ N/A true
1529 * READ RW true false
1530 * READ RW false true
1533 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1534 const struct pnfs_layout_range *range,
1537 struct pnfs_layout_range range1;
1539 if ((range->iomode == IOMODE_RW &&
1540 ls_range->iomode != IOMODE_RW) ||
1541 (range->iomode != ls_range->iomode &&
1542 strict_iomode == true) ||
1543 !pnfs_lseg_range_intersecting(ls_range, range))
1546 /* range1 covers only the first byte in the range */
1549 return pnfs_lseg_range_contained(ls_range, &range1);
1553 * lookup range in layout
1555 static struct pnfs_layout_segment *
1556 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1557 struct pnfs_layout_range *range,
1560 struct pnfs_layout_segment *lseg, *ret = NULL;
1562 dprintk("%s:Begin\n", __func__);
1564 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1565 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1566 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1567 pnfs_lseg_range_match(&lseg->pls_range, range,
1569 ret = pnfs_get_lseg(lseg);
1574 dprintk("%s:Return lseg %p ref %d\n",
1575 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1580 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1581 * to the MDS or over pNFS
1583 * The nfs_inode read_io and write_io fields are cumulative counters reset
1584 * when there are no layout segments. Note that in pnfs_update_layout iomode
1585 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1588 * A return of true means use MDS I/O.
1591 * If a file's size is smaller than the file size threshold, data accesses
1592 * SHOULD be sent to the metadata server. If an I/O request has a length that
1593 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1594 * server. If both file size and I/O size are provided, the client SHOULD
1595 * reach or exceed both thresholds before sending its read or write
1596 * requests to the data server.
1598 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1599 struct inode *ino, int iomode)
1601 struct nfs4_threshold *t = ctx->mdsthreshold;
1602 struct nfs_inode *nfsi = NFS_I(ino);
1603 loff_t fsize = i_size_read(ino);
1604 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1609 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1610 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1614 if (t->bm & THRESHOLD_RD) {
1615 dprintk("%s fsize %llu\n", __func__, fsize);
1617 if (fsize < t->rd_sz)
1620 if (t->bm & THRESHOLD_RD_IO) {
1621 dprintk("%s nfsi->read_io %llu\n", __func__,
1624 if (nfsi->read_io < t->rd_io_sz)
1629 if (t->bm & THRESHOLD_WR) {
1630 dprintk("%s fsize %llu\n", __func__, fsize);
1632 if (fsize < t->wr_sz)
1635 if (t->bm & THRESHOLD_WR_IO) {
1636 dprintk("%s nfsi->write_io %llu\n", __func__,
1639 if (nfsi->write_io < t->wr_io_sz)
1644 if (size_set && io_set) {
1647 } else if (size || io)
1650 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1654 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1657 * send layoutcommit as it can hold up layoutreturn due to lseg
1660 pnfs_layoutcommit_inode(lo->plh_inode, false);
1661 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1662 nfs_wait_bit_killable,
1663 TASK_UNINTERRUPTIBLE);
1666 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1668 unsigned long *bitlock = &lo->plh_flags;
1670 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1671 smp_mb__after_atomic();
1672 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1676 * Layout segment is retreived from the server if not cached.
1677 * The appropriate layout segment is referenced and returned to the caller.
1679 struct pnfs_layout_segment *
1680 pnfs_update_layout(struct inode *ino,
1681 struct nfs_open_context *ctx,
1684 enum pnfs_iomode iomode,
1688 struct pnfs_layout_range arg = {
1693 unsigned pg_offset, seq;
1694 struct nfs_server *server = NFS_SERVER(ino);
1695 struct nfs_client *clp = server->nfs_client;
1696 struct pnfs_layout_hdr *lo = NULL;
1697 struct pnfs_layout_segment *lseg = NULL;
1698 nfs4_stateid stateid;
1700 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1703 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1704 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1705 PNFS_UPDATE_LAYOUT_NO_PNFS);
1709 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1710 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1711 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1715 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1716 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1717 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1722 nfs4_client_recover_expired_lease(clp);
1724 spin_lock(&ino->i_lock);
1725 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1727 spin_unlock(&ino->i_lock);
1728 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1729 PNFS_UPDATE_LAYOUT_NOMEM);
1733 /* Do we even need to bother with this? */
1734 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1735 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1736 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1737 dprintk("%s matches recall, use MDS\n", __func__);
1741 /* if LAYOUTGET already failed once we don't try again */
1742 if (pnfs_layout_io_test_failed(lo, iomode)) {
1743 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1744 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1748 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1750 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1751 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1755 if (!nfs4_valid_open_stateid(ctx->state)) {
1756 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1757 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1762 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1763 * stateid, or it has been invalidated, then we must use the open
1766 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1769 * The first layoutget for the file. Need to serialize per
1770 * RFC 5661 Errata 3208.
1772 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1774 spin_unlock(&ino->i_lock);
1775 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1776 TASK_UNINTERRUPTIBLE);
1777 pnfs_put_layout_hdr(lo);
1778 dprintk("%s retrying\n", __func__);
1784 seq = read_seqbegin(&ctx->state->seqlock);
1785 nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1786 } while (read_seqretry(&ctx->state->seqlock, seq));
1788 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1792 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1793 * for LAYOUTRETURN even if first is true.
1795 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1796 spin_unlock(&ino->i_lock);
1797 dprintk("%s wait for layoutreturn\n", __func__);
1798 if (pnfs_prepare_to_retry_layoutget(lo)) {
1800 pnfs_clear_first_layoutget(lo);
1801 pnfs_put_layout_hdr(lo);
1802 dprintk("%s retrying\n", __func__);
1803 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1804 lseg, PNFS_UPDATE_LAYOUT_RETRY);
1807 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1808 PNFS_UPDATE_LAYOUT_RETURN);
1809 goto out_put_layout_hdr;
1812 if (pnfs_layoutgets_blocked(lo)) {
1813 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1814 PNFS_UPDATE_LAYOUT_BLOCKED);
1817 atomic_inc(&lo->plh_outstanding);
1818 spin_unlock(&ino->i_lock);
1820 if (list_empty(&lo->plh_layouts)) {
1821 /* The lo must be on the clp list if there is any
1822 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1824 spin_lock(&clp->cl_lock);
1825 if (list_empty(&lo->plh_layouts))
1826 list_add_tail(&lo->plh_layouts, &server->layouts);
1827 spin_unlock(&clp->cl_lock);
1830 pg_offset = arg.offset & ~PAGE_MASK;
1832 arg.offset -= pg_offset;
1833 arg.length += pg_offset;
1835 if (arg.length != NFS4_MAX_UINT64)
1836 arg.length = PAGE_ALIGN(arg.length);
1838 lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1839 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1840 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1841 atomic_dec(&lo->plh_outstanding);
1843 switch(PTR_ERR(lseg)) {
1845 if (time_after(jiffies, giveup))
1848 case -ERECALLCONFLICT:
1849 /* Huh? We hold no layouts, how is there a recall? */
1854 /* Destroy the existing layout and start over */
1855 if (time_after(jiffies, giveup))
1856 pnfs_destroy_layout(NFS_I(ino));
1861 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1862 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1865 goto out_put_layout_hdr;
1869 pnfs_clear_first_layoutget(lo);
1870 trace_pnfs_update_layout(ino, pos, count,
1871 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1872 pnfs_put_layout_hdr(lo);
1876 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1881 pnfs_clear_first_layoutget(lo);
1882 pnfs_put_layout_hdr(lo);
1884 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1885 "(%s, offset: %llu, length: %llu)\n",
1886 __func__, ino->i_sb->s_id,
1887 (unsigned long long)NFS_FILEID(ino),
1888 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1889 iomode==IOMODE_RW ? "read/write" : "read-only",
1890 (unsigned long long)pos,
1891 (unsigned long long)count);
1894 spin_unlock(&ino->i_lock);
1895 goto out_put_layout_hdr;
1897 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1900 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1902 switch (range->iomode) {
1909 if (range->offset == NFS4_MAX_UINT64)
1911 if (range->length == 0)
1913 if (range->length != NFS4_MAX_UINT64 &&
1914 range->length > NFS4_MAX_UINT64 - range->offset)
1919 struct pnfs_layout_segment *
1920 pnfs_layout_process(struct nfs4_layoutget *lgp)
1922 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1923 struct nfs4_layoutget_res *res = &lgp->res;
1924 struct pnfs_layout_segment *lseg;
1925 struct inode *ino = lo->plh_inode;
1928 if (!pnfs_sanity_check_layout_range(&res->range))
1929 return ERR_PTR(-EINVAL);
1931 /* Inject layout blob into I/O device driver */
1932 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1933 if (IS_ERR_OR_NULL(lseg)) {
1935 lseg = ERR_PTR(-ENOMEM);
1937 dprintk("%s: Could not allocate layout: error %ld\n",
1938 __func__, PTR_ERR(lseg));
1942 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1944 spin_lock(&ino->i_lock);
1945 if (pnfs_layoutgets_blocked(lo)) {
1946 dprintk("%s forget reply due to state\n", __func__);
1950 if (!pnfs_layout_is_valid(lo)) {
1951 /* We have a completely new layout */
1952 pnfs_set_layout_stateid(lo, &res->stateid, true);
1953 } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1954 /* existing state ID, make sure the sequence number matches. */
1955 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1956 dprintk("%s forget reply due to sequence\n", __func__);
1959 pnfs_set_layout_stateid(lo, &res->stateid, false);
1962 * We got an entirely new state ID. Mark all segments for the
1963 * inode invalid, and retry the layoutget
1965 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1969 pnfs_get_lseg(lseg);
1970 pnfs_layout_insert_lseg(lo, lseg, &free_me);
1973 if (res->return_on_close)
1974 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1976 spin_unlock(&ino->i_lock);
1977 pnfs_free_lseg_list(&free_me);
1981 spin_unlock(&ino->i_lock);
1982 lseg->pls_layout = lo;
1983 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1984 return ERR_PTR(-EAGAIN);
1988 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1989 * @lo: pointer to layout header
1990 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1991 * @return_range: describe layout segment ranges to be returned
1993 * This function is mainly intended for use by layoutrecall. It attempts
1994 * to free the layout segment immediately, or else to mark it for return
1995 * as soon as its reference count drops to zero.
1998 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1999 struct list_head *tmp_list,
2000 const struct pnfs_layout_range *return_range,
2003 struct pnfs_layout_segment *lseg, *next;
2006 dprintk("%s:Begin lo %p\n", __func__, lo);
2008 if (list_empty(&lo->plh_segs))
2011 assert_spin_locked(&lo->plh_inode->i_lock);
2013 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2014 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2015 dprintk("%s: marking lseg %p iomode %d "
2016 "offset %llu length %llu\n", __func__,
2017 lseg, lseg->pls_range.iomode,
2018 lseg->pls_range.offset,
2019 lseg->pls_range.length);
2020 if (mark_lseg_invalid(lseg, tmp_list))
2023 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2027 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2032 void pnfs_error_mark_layout_for_return(struct inode *inode,
2033 struct pnfs_layout_segment *lseg)
2035 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2036 struct pnfs_layout_range range = {
2037 .iomode = lseg->pls_range.iomode,
2039 .length = NFS4_MAX_UINT64,
2041 bool return_now = false;
2043 spin_lock(&inode->i_lock);
2044 pnfs_set_plh_return_info(lo, range.iomode, 0);
2045 /* Block LAYOUTGET */
2046 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2048 * mark all matching lsegs so that we are sure to have no live
2049 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2052 if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2053 nfs4_stateid stateid;
2054 enum pnfs_iomode iomode;
2056 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2057 spin_unlock(&inode->i_lock);
2059 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2061 spin_unlock(&inode->i_lock);
2062 nfs_commit_inode(inode, 0);
2065 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2068 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2070 u64 rd_size = req->wb_bytes;
2072 if (pgio->pg_lseg == NULL) {
2073 if (pgio->pg_dreq == NULL)
2074 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2076 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2078 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2085 if (IS_ERR(pgio->pg_lseg)) {
2086 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2087 pgio->pg_lseg = NULL;
2091 /* If no lseg, fall back to read through mds */
2092 if (pgio->pg_lseg == NULL)
2093 nfs_pageio_reset_read_mds(pgio);
2096 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2099 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2100 struct nfs_page *req, u64 wb_size)
2102 if (pgio->pg_lseg == NULL) {
2103 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2110 if (IS_ERR(pgio->pg_lseg)) {
2111 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2112 pgio->pg_lseg = NULL;
2116 /* If no lseg, fall back to write through mds */
2117 if (pgio->pg_lseg == NULL)
2118 nfs_pageio_reset_write_mds(pgio);
2120 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2123 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2125 if (desc->pg_lseg) {
2126 pnfs_put_lseg(desc->pg_lseg);
2127 desc->pg_lseg = NULL;
2130 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2133 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2134 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2137 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2138 struct nfs_page *prev, struct nfs_page *req)
2141 u64 seg_end, req_start, seg_left;
2143 size = nfs_generic_pg_test(pgio, prev, req);
2148 * 'size' contains the number of bytes left in the current page (up
2149 * to the original size asked for in @req->wb_bytes).
2151 * Calculate how many bytes are left in the layout segment
2152 * and if there are less bytes than 'size', return that instead.
2154 * Please also note that 'end_offset' is actually the offset of the
2155 * first byte that lies outside the pnfs_layout_range. FIXME?
2158 if (pgio->pg_lseg) {
2159 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2160 pgio->pg_lseg->pls_range.length);
2161 req_start = req_offset(req);
2162 WARN_ON_ONCE(req_start >= seg_end);
2163 /* start of request is past the last byte of this segment */
2164 if (req_start >= seg_end) {
2165 /* reference the new lseg */
2166 if (pgio->pg_ops->pg_cleanup)
2167 pgio->pg_ops->pg_cleanup(pgio);
2168 if (pgio->pg_ops->pg_init)
2169 pgio->pg_ops->pg_init(pgio, req);
2173 /* adjust 'size' iff there are fewer bytes left in the
2174 * segment than what nfs_generic_pg_test returned */
2175 seg_left = seg_end - req_start;
2176 if (seg_left < size)
2177 size = (unsigned int)seg_left;
2182 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2184 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2186 struct nfs_pageio_descriptor pgio;
2188 /* Resend all requests through the MDS */
2189 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2190 hdr->completion_ops);
2191 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2192 return nfs_pageio_resend(&pgio, hdr);
2194 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2196 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2199 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2200 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2201 PNFS_LAYOUTRET_ON_ERROR) {
2202 pnfs_return_layout(hdr->inode);
2204 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2205 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2209 * Called by non rpc-based layout drivers
2211 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2213 if (likely(!hdr->pnfs_error)) {
2214 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2215 hdr->mds_offset + hdr->res.count);
2216 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2218 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2219 if (unlikely(hdr->pnfs_error))
2220 pnfs_ld_handle_write_error(hdr);
2221 hdr->mds_ops->rpc_release(hdr);
2223 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2226 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2227 struct nfs_pgio_header *hdr)
2229 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2231 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2232 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2233 nfs_pageio_reset_write_mds(desc);
2234 mirror->pg_recoalesce = 1;
2236 nfs_pgio_data_destroy(hdr);
2240 static enum pnfs_try_status
2241 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2242 const struct rpc_call_ops *call_ops,
2243 struct pnfs_layout_segment *lseg,
2246 struct inode *inode = hdr->inode;
2247 enum pnfs_try_status trypnfs;
2248 struct nfs_server *nfss = NFS_SERVER(inode);
2250 hdr->mds_ops = call_ops;
2252 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2253 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2254 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2255 if (trypnfs != PNFS_NOT_ATTEMPTED)
2256 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2257 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2262 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2263 struct nfs_pgio_header *hdr, int how)
2265 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2266 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2267 enum pnfs_try_status trypnfs;
2269 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2270 if (trypnfs == PNFS_NOT_ATTEMPTED)
2271 pnfs_write_through_mds(desc, hdr);
2274 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2276 pnfs_put_lseg(hdr->lseg);
2277 nfs_pgio_header_free(hdr);
2281 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2283 struct nfs_pgio_header *hdr;
2286 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2288 desc->pg_error = -ENOMEM;
2289 return desc->pg_error;
2291 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2293 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2294 ret = nfs_generic_pgio(desc, hdr);
2296 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2300 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2302 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2304 struct nfs_pageio_descriptor pgio;
2306 /* Resend all requests through the MDS */
2307 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2308 return nfs_pageio_resend(&pgio, hdr);
2310 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2312 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2314 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2315 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2316 PNFS_LAYOUTRET_ON_ERROR) {
2317 pnfs_return_layout(hdr->inode);
2319 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2320 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2324 * Called by non rpc-based layout drivers
2326 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2328 if (likely(!hdr->pnfs_error))
2329 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2330 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2331 if (unlikely(hdr->pnfs_error))
2332 pnfs_ld_handle_read_error(hdr);
2333 hdr->mds_ops->rpc_release(hdr);
2335 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2338 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2339 struct nfs_pgio_header *hdr)
2341 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2343 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2344 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2345 nfs_pageio_reset_read_mds(desc);
2346 mirror->pg_recoalesce = 1;
2348 nfs_pgio_data_destroy(hdr);
2353 * Call the appropriate parallel I/O subsystem read function.
2355 static enum pnfs_try_status
2356 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2357 const struct rpc_call_ops *call_ops,
2358 struct pnfs_layout_segment *lseg)
2360 struct inode *inode = hdr->inode;
2361 struct nfs_server *nfss = NFS_SERVER(inode);
2362 enum pnfs_try_status trypnfs;
2364 hdr->mds_ops = call_ops;
2366 dprintk("%s: Reading ino:%lu %u@%llu\n",
2367 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2369 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2370 if (trypnfs != PNFS_NOT_ATTEMPTED)
2371 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2372 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2376 /* Resend all requests through pnfs. */
2377 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2379 struct nfs_pageio_descriptor pgio;
2381 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2382 /* Prevent deadlocks with layoutreturn! */
2383 pnfs_put_lseg(hdr->lseg);
2386 nfs_pageio_init_read(&pgio, hdr->inode, false,
2387 hdr->completion_ops);
2388 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2391 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2394 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2396 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2397 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2398 enum pnfs_try_status trypnfs;
2400 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2401 if (trypnfs == PNFS_TRY_AGAIN)
2402 pnfs_read_resend_pnfs(hdr);
2403 if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status)
2404 pnfs_read_through_mds(desc, hdr);
2407 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2409 pnfs_put_lseg(hdr->lseg);
2410 nfs_pgio_header_free(hdr);
2414 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2416 struct nfs_pgio_header *hdr;
2419 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2421 desc->pg_error = -ENOMEM;
2422 return desc->pg_error;
2424 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2425 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2426 ret = nfs_generic_pgio(desc, hdr);
2428 pnfs_do_read(desc, hdr);
2431 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2433 static void pnfs_clear_layoutcommitting(struct inode *inode)
2435 unsigned long *bitlock = &NFS_I(inode)->flags;
2437 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2438 smp_mb__after_atomic();
2439 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2443 * There can be multiple RW segments.
2445 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2447 struct pnfs_layout_segment *lseg;
2449 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2450 if (lseg->pls_range.iomode == IOMODE_RW &&
2451 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2452 list_add(&lseg->pls_lc_list, listp);
2456 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2458 struct pnfs_layout_segment *lseg, *tmp;
2460 /* Matched by references in pnfs_set_layoutcommit */
2461 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2462 list_del_init(&lseg->pls_lc_list);
2463 pnfs_put_lseg(lseg);
2466 pnfs_clear_layoutcommitting(inode);
2469 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2471 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2473 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2476 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2479 struct nfs_inode *nfsi = NFS_I(inode);
2480 bool mark_as_dirty = false;
2482 spin_lock(&inode->i_lock);
2483 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2484 nfsi->layout->plh_lwb = end_pos;
2485 mark_as_dirty = true;
2486 dprintk("%s: Set layoutcommit for inode %lu ",
2487 __func__, inode->i_ino);
2488 } else if (end_pos > nfsi->layout->plh_lwb)
2489 nfsi->layout->plh_lwb = end_pos;
2490 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2491 /* references matched in nfs4_layoutcommit_release */
2492 pnfs_get_lseg(lseg);
2494 spin_unlock(&inode->i_lock);
2495 dprintk("%s: lseg %p end_pos %llu\n",
2496 __func__, lseg, nfsi->layout->plh_lwb);
2498 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2499 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2501 mark_inode_dirty_sync(inode);
2503 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2505 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2507 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2509 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2510 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2511 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2515 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2516 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2517 * data to disk to allow the server to recover the data if it crashes.
2518 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2519 * is off, and a COMMIT is sent to a data server, or
2520 * if WRITEs to a data server return NFS_DATA_SYNC.
2523 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2525 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2526 struct nfs4_layoutcommit_data *data;
2527 struct nfs_inode *nfsi = NFS_I(inode);
2531 if (!pnfs_layoutcommit_outstanding(inode))
2534 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2537 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2540 status = wait_on_bit_lock_action(&nfsi->flags,
2541 NFS_INO_LAYOUTCOMMITTING,
2542 nfs_wait_bit_killable,
2549 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2550 data = kzalloc(sizeof(*data), GFP_NOFS);
2552 goto clear_layoutcommitting;
2555 spin_lock(&inode->i_lock);
2556 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2559 INIT_LIST_HEAD(&data->lseg_list);
2560 pnfs_list_write_lseg(inode, &data->lseg_list);
2562 end_pos = nfsi->layout->plh_lwb;
2564 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2565 spin_unlock(&inode->i_lock);
2567 data->args.inode = inode;
2568 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2569 nfs_fattr_init(&data->fattr);
2570 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2571 data->res.fattr = &data->fattr;
2573 data->args.lastbytewritten = end_pos - 1;
2575 data->args.lastbytewritten = U64_MAX;
2576 data->res.server = NFS_SERVER(inode);
2578 if (ld->prepare_layoutcommit) {
2579 status = ld->prepare_layoutcommit(&data->args);
2581 put_rpccred(data->cred);
2582 spin_lock(&inode->i_lock);
2583 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2584 if (end_pos > nfsi->layout->plh_lwb)
2585 nfsi->layout->plh_lwb = end_pos;
2591 status = nfs4_proc_layoutcommit(data, sync);
2594 mark_inode_dirty_sync(inode);
2595 dprintk("<-- %s status %d\n", __func__, status);
2598 spin_unlock(&inode->i_lock);
2600 clear_layoutcommitting:
2601 pnfs_clear_layoutcommitting(inode);
2604 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2607 pnfs_generic_sync(struct inode *inode, bool datasync)
2609 return pnfs_layoutcommit_inode(inode, true);
2611 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2613 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2615 struct nfs4_threshold *thp;
2617 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2619 dprintk("%s mdsthreshold allocation failed\n", __func__);
2625 #if IS_ENABLED(CONFIG_NFS_V4_2)
2627 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2629 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2630 struct nfs_server *server = NFS_SERVER(inode);
2631 struct nfs_inode *nfsi = NFS_I(inode);
2632 struct nfs42_layoutstat_data *data;
2633 struct pnfs_layout_hdr *hdr;
2636 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2639 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2642 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2645 spin_lock(&inode->i_lock);
2646 if (!NFS_I(inode)->layout) {
2647 spin_unlock(&inode->i_lock);
2648 goto out_clear_layoutstats;
2650 hdr = NFS_I(inode)->layout;
2651 pnfs_get_layout_hdr(hdr);
2652 spin_unlock(&inode->i_lock);
2654 data = kzalloc(sizeof(*data), gfp_flags);
2660 data->args.fh = NFS_FH(inode);
2661 data->args.inode = inode;
2662 status = ld->prepare_layoutstats(&data->args);
2666 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2669 dprintk("%s returns %d\n", __func__, status);
2675 pnfs_put_layout_hdr(hdr);
2676 out_clear_layoutstats:
2677 smp_mb__before_atomic();
2678 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2679 smp_mb__after_atomic();
2682 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2685 unsigned int layoutstats_timer;
2686 module_param(layoutstats_timer, uint, 0644);
2687 EXPORT_SYMBOL_GPL(layoutstats_timer);