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>
37 #define NFSDBG_FACILITY NFSDBG_PNFS
38 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
43 * protects pnfs_modules_tbl.
45 static DEFINE_SPINLOCK(pnfs_spinlock);
48 * pnfs_modules_tbl holds all pnfs modules
50 static LIST_HEAD(pnfs_modules_tbl);
52 /* Return the registered pnfs layout driver module matching given id */
53 static struct pnfs_layoutdriver_type *
54 find_pnfs_driver_locked(u32 id)
56 struct pnfs_layoutdriver_type *local;
58 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
63 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
67 static struct pnfs_layoutdriver_type *
68 find_pnfs_driver(u32 id)
70 struct pnfs_layoutdriver_type *local;
72 spin_lock(&pnfs_spinlock);
73 local = find_pnfs_driver_locked(id);
74 if (local != NULL && !try_module_get(local->owner)) {
75 dprintk("%s: Could not grab reference on module\n", __func__);
78 spin_unlock(&pnfs_spinlock);
83 unset_pnfs_layoutdriver(struct nfs_server *nfss)
85 if (nfss->pnfs_curr_ld) {
86 if (nfss->pnfs_curr_ld->clear_layoutdriver)
87 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
88 /* Decrement the MDS count. Purge the deviceid cache if zero */
89 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
90 nfs4_deviceid_purge_client(nfss->nfs_client);
91 module_put(nfss->pnfs_curr_ld->owner);
93 nfss->pnfs_curr_ld = NULL;
97 * Try to set the server's pnfs module to the pnfs layout type specified by id.
98 * Currently only one pNFS layout driver per filesystem is supported.
100 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
103 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
106 struct pnfs_layoutdriver_type *ld_type = NULL;
110 if (!(server->nfs_client->cl_exchange_flags &
111 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
112 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
113 __func__, id, server->nfs_client->cl_exchange_flags);
116 ld_type = find_pnfs_driver(id);
118 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
119 ld_type = find_pnfs_driver(id);
121 dprintk("%s: No pNFS module found for %u.\n",
126 server->pnfs_curr_ld = ld_type;
127 if (ld_type->set_layoutdriver
128 && ld_type->set_layoutdriver(server, mntfh)) {
129 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
130 "driver %u.\n", __func__, id);
131 module_put(ld_type->owner);
134 /* Bump the MDS count */
135 atomic_inc(&server->nfs_client->cl_mds_count);
137 dprintk("%s: pNFS module for %u set\n", __func__, id);
141 dprintk("%s: Using NFSv4 I/O\n", __func__);
142 server->pnfs_curr_ld = NULL;
146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
148 int status = -EINVAL;
149 struct pnfs_layoutdriver_type *tmp;
151 if (ld_type->id == 0) {
152 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
155 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
156 printk(KERN_ERR "NFS: %s Layout driver must provide "
157 "alloc_lseg and free_lseg.\n", __func__);
161 spin_lock(&pnfs_spinlock);
162 tmp = find_pnfs_driver_locked(ld_type->id);
164 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
166 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
169 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
170 __func__, ld_type->id);
172 spin_unlock(&pnfs_spinlock);
176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
181 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
182 spin_lock(&pnfs_spinlock);
183 list_del(&ld_type->pnfs_tblid);
184 spin_unlock(&pnfs_spinlock);
186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
189 * pNFS client layout cache
192 /* Need to hold i_lock if caller does not already hold reference */
194 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
196 atomic_inc(&lo->plh_refcount);
199 static struct pnfs_layout_hdr *
200 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
202 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
203 return ld->alloc_layout_hdr(ino, gfp_flags);
207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
209 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
210 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
212 if (!list_empty(&lo->plh_layouts)) {
213 struct nfs_client *clp = server->nfs_client;
215 spin_lock(&clp->cl_lock);
216 list_del_init(&lo->plh_layouts);
217 spin_unlock(&clp->cl_lock);
219 put_rpccred(lo->plh_lc_cred);
220 return ld->free_layout_hdr(lo);
224 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
226 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
227 dprintk("%s: freeing layout cache %p\n", __func__, lo);
229 /* Reset MDS Threshold I/O counters */
235 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
237 struct inode *inode = lo->plh_inode;
239 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
240 pnfs_detach_layout_hdr(lo);
241 spin_unlock(&inode->i_lock);
242 pnfs_free_layout_hdr(lo);
247 pnfs_iomode_to_fail_bit(u32 iomode)
249 return iomode == IOMODE_RW ?
250 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
254 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
256 lo->plh_retry_timestamp = jiffies;
257 if (test_and_set_bit(fail_bit, &lo->plh_flags))
258 atomic_inc(&lo->plh_refcount);
262 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
264 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
265 atomic_dec(&lo->plh_refcount);
269 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
271 struct inode *inode = lo->plh_inode;
272 struct pnfs_layout_range range = {
275 .length = NFS4_MAX_UINT64,
279 spin_lock(&inode->i_lock);
280 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
281 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
282 spin_unlock(&inode->i_lock);
283 pnfs_free_lseg_list(&head);
284 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
285 iomode == IOMODE_RW ? "RW" : "READ");
289 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
291 unsigned long start, end;
292 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
294 if (test_bit(fail_bit, &lo->plh_flags) == 0)
297 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
298 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
299 /* It is time to retry the failed layoutgets */
300 pnfs_layout_clear_fail_bit(lo, fail_bit);
307 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
309 INIT_LIST_HEAD(&lseg->pls_list);
310 INIT_LIST_HEAD(&lseg->pls_lc_list);
311 atomic_set(&lseg->pls_refcount, 1);
313 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
314 lseg->pls_layout = lo;
317 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
319 struct inode *ino = lseg->pls_layout->plh_inode;
321 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
325 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
326 struct pnfs_layout_segment *lseg)
328 struct inode *inode = lo->plh_inode;
330 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
331 list_del_init(&lseg->pls_list);
332 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
333 atomic_dec(&lo->plh_refcount);
334 if (list_empty(&lo->plh_segs))
335 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
336 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
340 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
342 struct pnfs_layout_hdr *lo;
348 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
349 atomic_read(&lseg->pls_refcount),
350 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
351 lo = lseg->pls_layout;
352 inode = lo->plh_inode;
353 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
354 pnfs_get_layout_hdr(lo);
355 pnfs_layout_remove_lseg(lo, lseg);
356 spin_unlock(&inode->i_lock);
357 pnfs_free_lseg(lseg);
358 pnfs_put_layout_hdr(lo);
361 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
364 end_offset(u64 start, u64 len)
369 return end >= start ? end : NFS4_MAX_UINT64;
373 * is l2 fully contained in l1?
375 * [----------------------------------)
380 lo_seg_contained(struct pnfs_layout_range *l1,
381 struct pnfs_layout_range *l2)
383 u64 start1 = l1->offset;
384 u64 end1 = end_offset(start1, l1->length);
385 u64 start2 = l2->offset;
386 u64 end2 = end_offset(start2, l2->length);
388 return (start1 <= start2) && (end1 >= end2);
392 * is l1 and l2 intersecting?
394 * [----------------------------------)
399 lo_seg_intersecting(struct pnfs_layout_range *l1,
400 struct pnfs_layout_range *l2)
402 u64 start1 = l1->offset;
403 u64 end1 = end_offset(start1, l1->length);
404 u64 start2 = l2->offset;
405 u64 end2 = end_offset(start2, l2->length);
407 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
408 (end2 == NFS4_MAX_UINT64 || end2 > start1);
412 should_free_lseg(struct pnfs_layout_range *lseg_range,
413 struct pnfs_layout_range *recall_range)
415 return (recall_range->iomode == IOMODE_ANY ||
416 lseg_range->iomode == recall_range->iomode) &&
417 lo_seg_intersecting(lseg_range, recall_range);
420 /* Returns 1 if lseg is removed from list, 0 otherwise */
421 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
422 struct list_head *tmp_list)
426 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
427 /* Remove the reference keeping the lseg in the
428 * list. It will now be removed when all
429 * outstanding io is finished.
431 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
432 atomic_read(&lseg->pls_refcount));
433 if (atomic_dec_and_test(&lseg->pls_refcount)) {
434 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
435 list_add(&lseg->pls_list, tmp_list);
442 /* Returns count of number of matching invalid lsegs remaining in list
446 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
447 struct list_head *tmp_list,
448 struct pnfs_layout_range *recall_range)
450 struct pnfs_layout_segment *lseg, *next;
451 int invalid = 0, removed = 0;
453 dprintk("%s:Begin lo %p\n", __func__, lo);
455 if (list_empty(&lo->plh_segs))
457 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
459 should_free_lseg(&lseg->pls_range, recall_range)) {
460 dprintk("%s: freeing lseg %p iomode %d "
461 "offset %llu length %llu\n", __func__,
462 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
463 lseg->pls_range.length);
465 removed += mark_lseg_invalid(lseg, tmp_list);
467 dprintk("%s:Return %i\n", __func__, invalid - removed);
468 return invalid - removed;
471 /* note free_me must contain lsegs from a single layout_hdr */
473 pnfs_free_lseg_list(struct list_head *free_me)
475 struct pnfs_layout_segment *lseg, *tmp;
477 if (list_empty(free_me))
480 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
481 list_del(&lseg->pls_list);
482 pnfs_free_lseg(lseg);
487 pnfs_destroy_layout(struct nfs_inode *nfsi)
489 struct pnfs_layout_hdr *lo;
492 spin_lock(&nfsi->vfs_inode.i_lock);
495 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
496 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
497 pnfs_get_layout_hdr(lo);
498 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
499 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
500 spin_unlock(&nfsi->vfs_inode.i_lock);
501 pnfs_free_lseg_list(&tmp_list);
502 pnfs_put_layout_hdr(lo);
504 spin_unlock(&nfsi->vfs_inode.i_lock);
506 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
509 * Called by the state manger to remove all layouts established under an
513 pnfs_destroy_all_layouts(struct nfs_client *clp)
515 struct nfs_server *server;
516 struct pnfs_layout_hdr *lo;
519 nfs4_deviceid_mark_client_invalid(clp);
520 nfs4_deviceid_purge_client(clp);
522 spin_lock(&clp->cl_lock);
524 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
525 if (!list_empty(&server->layouts))
526 list_splice_init(&server->layouts, &tmp_list);
529 spin_unlock(&clp->cl_lock);
531 while (!list_empty(&tmp_list)) {
532 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
534 dprintk("%s freeing layout for inode %lu\n", __func__,
535 lo->plh_inode->i_ino);
536 list_del_init(&lo->plh_layouts);
537 pnfs_destroy_layout(NFS_I(lo->plh_inode));
542 * Compare 2 layout stateid sequence ids, to see which is newer,
543 * taking into account wraparound issues.
545 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
547 return (s32)s1 - (s32)s2 > 0;
550 /* update lo->plh_stateid with new if is more recent */
552 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
555 u32 oldseq, newseq, new_barrier;
556 int empty = list_empty(&lo->plh_segs);
558 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
559 newseq = be32_to_cpu(new->seqid);
560 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
561 nfs4_stateid_copy(&lo->plh_stateid, new);
562 if (update_barrier) {
563 new_barrier = be32_to_cpu(new->seqid);
565 /* Because of wraparound, we want to keep the barrier
566 * "close" to the current seqids.
568 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
570 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
571 lo->plh_barrier = new_barrier;
576 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
577 const nfs4_stateid *stateid)
579 u32 seqid = be32_to_cpu(stateid->seqid);
581 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
584 /* lget is set to 1 if called from inside send_layoutget call chain */
586 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
588 return lo->plh_block_lgets ||
589 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
590 (list_empty(&lo->plh_segs) &&
591 (atomic_read(&lo->plh_outstanding) > lget));
595 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
596 struct nfs4_state *open_state)
600 dprintk("--> %s\n", __func__);
601 spin_lock(&lo->plh_inode->i_lock);
602 if (pnfs_layoutgets_blocked(lo, 1)) {
604 } else if (list_empty(&lo->plh_segs)) {
608 seq = read_seqbegin(&open_state->seqlock);
609 nfs4_stateid_copy(dst, &open_state->stateid);
610 } while (read_seqretry(&open_state->seqlock, seq));
612 nfs4_stateid_copy(dst, &lo->plh_stateid);
613 spin_unlock(&lo->plh_inode->i_lock);
614 dprintk("<-- %s\n", __func__);
619 * Get layout from server.
620 * for now, assume that whole file layouts are requested.
622 * arg->length: all ones
624 static struct pnfs_layout_segment *
625 send_layoutget(struct pnfs_layout_hdr *lo,
626 struct nfs_open_context *ctx,
627 struct pnfs_layout_range *range,
630 struct inode *ino = lo->plh_inode;
631 struct nfs_server *server = NFS_SERVER(ino);
632 struct nfs4_layoutget *lgp;
633 struct pnfs_layout_segment *lseg;
635 dprintk("--> %s\n", __func__);
638 lgp = kzalloc(sizeof(*lgp), gfp_flags);
642 lgp->args.minlength = PAGE_CACHE_SIZE;
643 if (lgp->args.minlength > range->length)
644 lgp->args.minlength = range->length;
645 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
646 lgp->args.range = *range;
647 lgp->args.type = server->pnfs_curr_ld->id;
648 lgp->args.inode = ino;
649 lgp->args.ctx = get_nfs_open_context(ctx);
650 lgp->gfp_flags = gfp_flags;
652 /* Synchronously retrieve layout information from server and
655 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
657 switch (PTR_ERR(lseg)) {
662 /* remember that LAYOUTGET failed and suspend trying */
663 pnfs_layout_io_set_failed(lo, range->iomode);
672 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
673 * when the layout segment list is empty.
675 * Note that a pnfs_layout_hdr can exist with an empty layout segment
676 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
677 * deviceid is marked invalid.
680 _pnfs_return_layout(struct inode *ino)
682 struct pnfs_layout_hdr *lo = NULL;
683 struct nfs_inode *nfsi = NFS_I(ino);
685 struct nfs4_layoutreturn *lrp;
686 nfs4_stateid stateid;
687 int status = 0, empty;
689 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
691 spin_lock(&ino->i_lock);
694 spin_unlock(&ino->i_lock);
695 dprintk("NFS: %s no layout to return\n", __func__);
698 stateid = nfsi->layout->plh_stateid;
699 /* Reference matched in nfs4_layoutreturn_release */
700 pnfs_get_layout_hdr(lo);
701 empty = list_empty(&lo->plh_segs);
702 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
703 /* Don't send a LAYOUTRETURN if list was initially empty */
705 spin_unlock(&ino->i_lock);
706 pnfs_put_layout_hdr(lo);
707 dprintk("NFS: %s no layout segments to return\n", __func__);
710 lo->plh_block_lgets++;
711 spin_unlock(&ino->i_lock);
712 pnfs_free_lseg_list(&tmp_list);
714 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
716 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
717 if (unlikely(lrp == NULL)) {
719 spin_lock(&ino->i_lock);
720 lo->plh_block_lgets--;
721 spin_unlock(&ino->i_lock);
722 pnfs_put_layout_hdr(lo);
726 lrp->args.stateid = stateid;
727 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
728 lrp->args.inode = ino;
729 lrp->args.layout = lo;
730 lrp->clp = NFS_SERVER(ino)->nfs_client;
732 status = nfs4_proc_layoutreturn(lrp);
734 dprintk("<-- %s status: %d\n", __func__, status);
737 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
739 bool pnfs_roc(struct inode *ino)
741 struct pnfs_layout_hdr *lo;
742 struct pnfs_layout_segment *lseg, *tmp;
746 spin_lock(&ino->i_lock);
747 lo = NFS_I(ino)->layout;
748 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
749 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
751 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
752 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
753 mark_lseg_invalid(lseg, &tmp_list);
758 lo->plh_block_lgets++;
759 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
760 spin_unlock(&ino->i_lock);
761 pnfs_free_lseg_list(&tmp_list);
765 spin_unlock(&ino->i_lock);
769 void pnfs_roc_release(struct inode *ino)
771 struct pnfs_layout_hdr *lo;
773 spin_lock(&ino->i_lock);
774 lo = NFS_I(ino)->layout;
775 lo->plh_block_lgets--;
776 if (atomic_dec_and_test(&lo->plh_refcount)) {
777 pnfs_detach_layout_hdr(lo);
778 spin_unlock(&ino->i_lock);
779 pnfs_free_layout_hdr(lo);
781 spin_unlock(&ino->i_lock);
784 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
786 struct pnfs_layout_hdr *lo;
788 spin_lock(&ino->i_lock);
789 lo = NFS_I(ino)->layout;
790 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
791 lo->plh_barrier = barrier;
792 spin_unlock(&ino->i_lock);
795 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
797 struct nfs_inode *nfsi = NFS_I(ino);
798 struct pnfs_layout_hdr *lo;
799 struct pnfs_layout_segment *lseg;
803 spin_lock(&ino->i_lock);
804 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
805 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
806 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
811 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
813 /* Since close does not return a layout stateid for use as
814 * a barrier, we choose the worst-case barrier.
816 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
818 spin_unlock(&ino->i_lock);
823 * Compare two layout segments for sorting into layout cache.
824 * We want to preferentially return RW over RO layouts, so ensure those
828 cmp_layout(struct pnfs_layout_range *l1,
829 struct pnfs_layout_range *l2)
833 /* high offset > low offset */
834 d = l1->offset - l2->offset;
838 /* short length > long length */
839 d = l2->length - l1->length;
843 /* read > read/write */
844 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
848 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
849 struct pnfs_layout_segment *lseg)
851 struct pnfs_layout_segment *lp;
853 dprintk("%s:Begin\n", __func__);
855 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
856 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
858 list_add_tail(&lseg->pls_list, &lp->pls_list);
859 dprintk("%s: inserted lseg %p "
860 "iomode %d offset %llu length %llu before "
861 "lp %p iomode %d offset %llu length %llu\n",
862 __func__, lseg, lseg->pls_range.iomode,
863 lseg->pls_range.offset, lseg->pls_range.length,
864 lp, lp->pls_range.iomode, lp->pls_range.offset,
865 lp->pls_range.length);
868 list_add_tail(&lseg->pls_list, &lo->plh_segs);
869 dprintk("%s: inserted lseg %p "
870 "iomode %d offset %llu length %llu at tail\n",
871 __func__, lseg, lseg->pls_range.iomode,
872 lseg->pls_range.offset, lseg->pls_range.length);
874 pnfs_get_layout_hdr(lo);
876 dprintk("%s:Return\n", __func__);
879 static struct pnfs_layout_hdr *
880 alloc_init_layout_hdr(struct inode *ino,
881 struct nfs_open_context *ctx,
884 struct pnfs_layout_hdr *lo;
886 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
889 atomic_set(&lo->plh_refcount, 1);
890 INIT_LIST_HEAD(&lo->plh_layouts);
891 INIT_LIST_HEAD(&lo->plh_segs);
892 INIT_LIST_HEAD(&lo->plh_bulk_recall);
894 lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
898 static struct pnfs_layout_hdr *
899 pnfs_find_alloc_layout(struct inode *ino,
900 struct nfs_open_context *ctx,
903 struct nfs_inode *nfsi = NFS_I(ino);
904 struct pnfs_layout_hdr *new = NULL;
906 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
908 if (nfsi->layout != NULL)
910 spin_unlock(&ino->i_lock);
911 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
912 spin_lock(&ino->i_lock);
914 if (likely(nfsi->layout == NULL)) { /* Won the race? */
917 } else if (new != NULL)
918 pnfs_free_layout_hdr(new);
920 pnfs_get_layout_hdr(nfsi->layout);
925 * iomode matching rules:
936 is_matching_lseg(struct pnfs_layout_range *ls_range,
937 struct pnfs_layout_range *range)
939 struct pnfs_layout_range range1;
941 if ((range->iomode == IOMODE_RW &&
942 ls_range->iomode != IOMODE_RW) ||
943 !lo_seg_intersecting(ls_range, range))
946 /* range1 covers only the first byte in the range */
949 return lo_seg_contained(ls_range, &range1);
953 * lookup range in layout
955 static struct pnfs_layout_segment *
956 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
957 struct pnfs_layout_range *range)
959 struct pnfs_layout_segment *lseg, *ret = NULL;
961 dprintk("%s:Begin\n", __func__);
963 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
964 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
965 is_matching_lseg(&lseg->pls_range, range)) {
966 ret = pnfs_get_lseg(lseg);
969 if (lseg->pls_range.offset > range->offset)
973 dprintk("%s:Return lseg %p ref %d\n",
974 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
979 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
980 * to the MDS or over pNFS
982 * The nfs_inode read_io and write_io fields are cumulative counters reset
983 * when there are no layout segments. Note that in pnfs_update_layout iomode
984 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
987 * A return of true means use MDS I/O.
990 * If a file's size is smaller than the file size threshold, data accesses
991 * SHOULD be sent to the metadata server. If an I/O request has a length that
992 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
993 * server. If both file size and I/O size are provided, the client SHOULD
994 * reach or exceed both thresholds before sending its read or write
995 * requests to the data server.
997 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
998 struct inode *ino, int iomode)
1000 struct nfs4_threshold *t = ctx->mdsthreshold;
1001 struct nfs_inode *nfsi = NFS_I(ino);
1002 loff_t fsize = i_size_read(ino);
1003 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1008 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1009 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1013 if (t->bm & THRESHOLD_RD) {
1014 dprintk("%s fsize %llu\n", __func__, fsize);
1016 if (fsize < t->rd_sz)
1019 if (t->bm & THRESHOLD_RD_IO) {
1020 dprintk("%s nfsi->read_io %llu\n", __func__,
1023 if (nfsi->read_io < t->rd_io_sz)
1028 if (t->bm & THRESHOLD_WR) {
1029 dprintk("%s fsize %llu\n", __func__, fsize);
1031 if (fsize < t->wr_sz)
1034 if (t->bm & THRESHOLD_WR_IO) {
1035 dprintk("%s nfsi->write_io %llu\n", __func__,
1038 if (nfsi->write_io < t->wr_io_sz)
1043 if (size_set && io_set) {
1046 } else if (size || io)
1049 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1054 * Layout segment is retreived from the server if not cached.
1055 * The appropriate layout segment is referenced and returned to the caller.
1057 struct pnfs_layout_segment *
1058 pnfs_update_layout(struct inode *ino,
1059 struct nfs_open_context *ctx,
1062 enum pnfs_iomode iomode,
1065 struct pnfs_layout_range arg = {
1071 struct nfs_server *server = NFS_SERVER(ino);
1072 struct nfs_client *clp = server->nfs_client;
1073 struct pnfs_layout_hdr *lo;
1074 struct pnfs_layout_segment *lseg = NULL;
1077 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1080 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1083 spin_lock(&ino->i_lock);
1084 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1086 spin_unlock(&ino->i_lock);
1090 /* Do we even need to bother with this? */
1091 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1092 dprintk("%s matches recall, use MDS\n", __func__);
1096 /* if LAYOUTGET already failed once we don't try again */
1097 if (pnfs_layout_io_test_failed(lo, iomode))
1100 /* Check to see if the layout for the given range already exists */
1101 lseg = pnfs_find_lseg(lo, &arg);
1105 if (pnfs_layoutgets_blocked(lo, 0))
1107 atomic_inc(&lo->plh_outstanding);
1109 if (list_empty(&lo->plh_segs))
1112 spin_unlock(&ino->i_lock);
1114 /* The lo must be on the clp list if there is any
1115 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1117 spin_lock(&clp->cl_lock);
1118 BUG_ON(!list_empty(&lo->plh_layouts));
1119 list_add_tail(&lo->plh_layouts, &server->layouts);
1120 spin_unlock(&clp->cl_lock);
1123 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1125 arg.offset -= pg_offset;
1126 arg.length += pg_offset;
1128 if (arg.length != NFS4_MAX_UINT64)
1129 arg.length = PAGE_CACHE_ALIGN(arg.length);
1131 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1132 atomic_dec(&lo->plh_outstanding);
1134 pnfs_put_layout_hdr(lo);
1136 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1137 "(%s, offset: %llu, length: %llu)\n",
1138 __func__, ino->i_sb->s_id,
1139 (unsigned long long)NFS_FILEID(ino),
1140 lseg == NULL ? "not found" : "found",
1141 iomode==IOMODE_RW ? "read/write" : "read-only",
1142 (unsigned long long)pos,
1143 (unsigned long long)count);
1146 spin_unlock(&ino->i_lock);
1147 goto out_put_layout_hdr;
1149 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1151 struct pnfs_layout_segment *
1152 pnfs_layout_process(struct nfs4_layoutget *lgp)
1154 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1155 struct nfs4_layoutget_res *res = &lgp->res;
1156 struct pnfs_layout_segment *lseg;
1157 struct inode *ino = lo->plh_inode;
1160 /* Inject layout blob into I/O device driver */
1161 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1162 if (!lseg || IS_ERR(lseg)) {
1166 status = PTR_ERR(lseg);
1167 dprintk("%s: Could not allocate layout: error %d\n",
1172 spin_lock(&ino->i_lock);
1173 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1174 dprintk("%s forget reply due to recall\n", __func__);
1175 goto out_forget_reply;
1178 if (pnfs_layoutgets_blocked(lo, 1) ||
1179 pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1180 dprintk("%s forget reply due to state\n", __func__);
1181 goto out_forget_reply;
1184 /* Done processing layoutget. Set the layout stateid */
1185 pnfs_set_layout_stateid(lo, &res->stateid, false);
1187 init_lseg(lo, lseg);
1188 lseg->pls_range = res->range;
1189 pnfs_get_lseg(lseg);
1190 pnfs_layout_insert_lseg(lo, lseg);
1192 if (res->return_on_close) {
1193 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1194 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1197 spin_unlock(&ino->i_lock);
1200 return ERR_PTR(status);
1203 spin_unlock(&ino->i_lock);
1204 lseg->pls_layout = lo;
1205 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1210 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1212 u64 rd_size = req->wb_bytes;
1214 BUG_ON(pgio->pg_lseg != NULL);
1216 if (req->wb_offset != req->wb_pgbase) {
1217 nfs_pageio_reset_read_mds(pgio);
1221 if (pgio->pg_dreq == NULL)
1222 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1224 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1226 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1232 /* If no lseg, fall back to read through mds */
1233 if (pgio->pg_lseg == NULL)
1234 nfs_pageio_reset_read_mds(pgio);
1237 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1240 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1241 struct nfs_page *req, u64 wb_size)
1243 BUG_ON(pgio->pg_lseg != NULL);
1245 if (req->wb_offset != req->wb_pgbase) {
1246 nfs_pageio_reset_write_mds(pgio);
1250 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1256 /* If no lseg, fall back to write through mds */
1257 if (pgio->pg_lseg == NULL)
1258 nfs_pageio_reset_write_mds(pgio);
1260 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1263 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1264 const struct nfs_pgio_completion_ops *compl_ops)
1266 struct nfs_server *server = NFS_SERVER(inode);
1267 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1270 nfs_pageio_init_read(pgio, inode, compl_ops);
1272 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1276 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1278 const struct nfs_pgio_completion_ops *compl_ops)
1280 struct nfs_server *server = NFS_SERVER(inode);
1281 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1284 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1286 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1290 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1291 struct nfs_page *req)
1293 if (pgio->pg_lseg == NULL)
1294 return nfs_generic_pg_test(pgio, prev, req);
1297 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1298 * Note that this test makes several assumptions:
1299 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1300 * is known to lie within the range.
1301 * - that the nfs_page being tested is known to be contiguous with the
1302 * previous nfs_page.
1303 * - Layout ranges are page aligned, so we only have to test the
1304 * start offset of the request.
1306 * Please also note that 'end_offset' is actually the offset of the
1307 * first byte that lies outside the pnfs_layout_range. FIXME?
1310 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1311 pgio->pg_lseg->pls_range.length);
1313 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1315 int pnfs_write_done_resend_to_mds(struct inode *inode,
1316 struct list_head *head,
1317 const struct nfs_pgio_completion_ops *compl_ops)
1319 struct nfs_pageio_descriptor pgio;
1322 /* Resend all requests through the MDS */
1323 nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1324 while (!list_empty(head)) {
1325 struct nfs_page *req = nfs_list_entry(head->next);
1327 nfs_list_remove_request(req);
1328 if (!nfs_pageio_add_request(&pgio, req))
1329 nfs_list_add_request(req, &failed);
1331 nfs_pageio_complete(&pgio);
1333 if (!list_empty(&failed)) {
1334 /* For some reason our attempt to resend pages. Mark the
1335 * overall send request as having failed, and let
1336 * nfs_writeback_release_full deal with the error.
1338 list_move(&failed, head);
1343 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1345 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1347 struct nfs_pgio_header *hdr = data->header;
1349 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1350 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1351 PNFS_LAYOUTRET_ON_ERROR) {
1352 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1353 pnfs_return_layout(hdr->inode);
1355 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1356 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1358 hdr->completion_ops);
1362 * Called by non rpc-based layout drivers
1364 void pnfs_ld_write_done(struct nfs_write_data *data)
1366 struct nfs_pgio_header *hdr = data->header;
1368 if (!hdr->pnfs_error) {
1369 pnfs_set_layoutcommit(data);
1370 hdr->mds_ops->rpc_call_done(&data->task, data);
1372 pnfs_ld_handle_write_error(data);
1373 hdr->mds_ops->rpc_release(data);
1375 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1378 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1379 struct nfs_write_data *data)
1381 struct nfs_pgio_header *hdr = data->header;
1383 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1384 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1385 nfs_pageio_reset_write_mds(desc);
1386 desc->pg_recoalesce = 1;
1388 nfs_writedata_release(data);
1391 static enum pnfs_try_status
1392 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1393 const struct rpc_call_ops *call_ops,
1394 struct pnfs_layout_segment *lseg,
1397 struct nfs_pgio_header *hdr = wdata->header;
1398 struct inode *inode = hdr->inode;
1399 enum pnfs_try_status trypnfs;
1400 struct nfs_server *nfss = NFS_SERVER(inode);
1402 hdr->mds_ops = call_ops;
1404 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1405 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1406 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1407 if (trypnfs != PNFS_NOT_ATTEMPTED)
1408 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1409 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1414 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1416 struct nfs_write_data *data;
1417 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1418 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1420 desc->pg_lseg = NULL;
1421 while (!list_empty(head)) {
1422 enum pnfs_try_status trypnfs;
1424 data = list_first_entry(head, struct nfs_write_data, list);
1425 list_del_init(&data->list);
1427 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1428 if (trypnfs == PNFS_NOT_ATTEMPTED)
1429 pnfs_write_through_mds(desc, data);
1431 pnfs_put_lseg(lseg);
1434 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1436 pnfs_put_lseg(hdr->lseg);
1437 nfs_writehdr_free(hdr);
1439 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1442 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1444 struct nfs_write_header *whdr;
1445 struct nfs_pgio_header *hdr;
1448 whdr = nfs_writehdr_alloc();
1450 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1451 pnfs_put_lseg(desc->pg_lseg);
1452 desc->pg_lseg = NULL;
1455 hdr = &whdr->header;
1456 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1457 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1458 atomic_inc(&hdr->refcnt);
1459 ret = nfs_generic_flush(desc, hdr);
1461 pnfs_put_lseg(desc->pg_lseg);
1462 desc->pg_lseg = NULL;
1464 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1465 if (atomic_dec_and_test(&hdr->refcnt))
1466 hdr->completion_ops->completion(hdr);
1469 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1471 int pnfs_read_done_resend_to_mds(struct inode *inode,
1472 struct list_head *head,
1473 const struct nfs_pgio_completion_ops *compl_ops)
1475 struct nfs_pageio_descriptor pgio;
1478 /* Resend all requests through the MDS */
1479 nfs_pageio_init_read(&pgio, inode, compl_ops);
1480 while (!list_empty(head)) {
1481 struct nfs_page *req = nfs_list_entry(head->next);
1483 nfs_list_remove_request(req);
1484 if (!nfs_pageio_add_request(&pgio, req))
1485 nfs_list_add_request(req, &failed);
1487 nfs_pageio_complete(&pgio);
1489 if (!list_empty(&failed)) {
1490 list_move(&failed, head);
1495 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1497 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1499 struct nfs_pgio_header *hdr = data->header;
1501 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1502 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1503 PNFS_LAYOUTRET_ON_ERROR) {
1504 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1505 pnfs_return_layout(hdr->inode);
1507 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1508 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1510 hdr->completion_ops);
1514 * Called by non rpc-based layout drivers
1516 void pnfs_ld_read_done(struct nfs_read_data *data)
1518 struct nfs_pgio_header *hdr = data->header;
1520 if (likely(!hdr->pnfs_error)) {
1521 __nfs4_read_done_cb(data);
1522 hdr->mds_ops->rpc_call_done(&data->task, data);
1524 pnfs_ld_handle_read_error(data);
1525 hdr->mds_ops->rpc_release(data);
1527 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1530 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1531 struct nfs_read_data *data)
1533 struct nfs_pgio_header *hdr = data->header;
1535 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1536 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1537 nfs_pageio_reset_read_mds(desc);
1538 desc->pg_recoalesce = 1;
1540 nfs_readdata_release(data);
1544 * Call the appropriate parallel I/O subsystem read function.
1546 static enum pnfs_try_status
1547 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1548 const struct rpc_call_ops *call_ops,
1549 struct pnfs_layout_segment *lseg)
1551 struct nfs_pgio_header *hdr = rdata->header;
1552 struct inode *inode = hdr->inode;
1553 struct nfs_server *nfss = NFS_SERVER(inode);
1554 enum pnfs_try_status trypnfs;
1556 hdr->mds_ops = call_ops;
1558 dprintk("%s: Reading ino:%lu %u@%llu\n",
1559 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1561 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1562 if (trypnfs != PNFS_NOT_ATTEMPTED)
1563 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1564 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1569 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1571 struct nfs_read_data *data;
1572 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1573 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1575 desc->pg_lseg = NULL;
1576 while (!list_empty(head)) {
1577 enum pnfs_try_status trypnfs;
1579 data = list_first_entry(head, struct nfs_read_data, list);
1580 list_del_init(&data->list);
1582 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1583 if (trypnfs == PNFS_NOT_ATTEMPTED)
1584 pnfs_read_through_mds(desc, data);
1586 pnfs_put_lseg(lseg);
1589 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1591 pnfs_put_lseg(hdr->lseg);
1592 nfs_readhdr_free(hdr);
1594 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1597 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1599 struct nfs_read_header *rhdr;
1600 struct nfs_pgio_header *hdr;
1603 rhdr = nfs_readhdr_alloc();
1605 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1607 pnfs_put_lseg(desc->pg_lseg);
1608 desc->pg_lseg = NULL;
1611 hdr = &rhdr->header;
1612 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1613 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1614 atomic_inc(&hdr->refcnt);
1615 ret = nfs_generic_pagein(desc, hdr);
1617 pnfs_put_lseg(desc->pg_lseg);
1618 desc->pg_lseg = NULL;
1620 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1621 if (atomic_dec_and_test(&hdr->refcnt))
1622 hdr->completion_ops->completion(hdr);
1625 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1628 * There can be multiple RW segments.
1630 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1632 struct pnfs_layout_segment *lseg;
1634 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1635 if (lseg->pls_range.iomode == IOMODE_RW &&
1636 test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1637 list_add(&lseg->pls_lc_list, listp);
1641 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1643 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1645 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1648 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1650 struct nfs_pgio_header *hdr = wdata->header;
1651 struct inode *inode = hdr->inode;
1652 struct nfs_inode *nfsi = NFS_I(inode);
1653 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1654 bool mark_as_dirty = false;
1656 spin_lock(&inode->i_lock);
1657 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1658 mark_as_dirty = true;
1659 dprintk("%s: Set layoutcommit for inode %lu ",
1660 __func__, inode->i_ino);
1662 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1663 /* references matched in nfs4_layoutcommit_release */
1664 pnfs_get_lseg(hdr->lseg);
1666 if (end_pos > nfsi->layout->plh_lwb)
1667 nfsi->layout->plh_lwb = end_pos;
1668 spin_unlock(&inode->i_lock);
1669 dprintk("%s: lseg %p end_pos %llu\n",
1670 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1672 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1673 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1675 mark_inode_dirty_sync(inode);
1677 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1679 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1681 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1683 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1684 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1688 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1689 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1690 * data to disk to allow the server to recover the data if it crashes.
1691 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1692 * is off, and a COMMIT is sent to a data server, or
1693 * if WRITEs to a data server return NFS_DATA_SYNC.
1696 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1698 struct nfs4_layoutcommit_data *data;
1699 struct nfs_inode *nfsi = NFS_I(inode);
1703 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1705 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1708 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1709 data = kzalloc(sizeof(*data), GFP_NOFS);
1715 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1718 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1723 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1724 nfs_wait_bit_killable, TASK_KILLABLE);
1729 INIT_LIST_HEAD(&data->lseg_list);
1730 spin_lock(&inode->i_lock);
1731 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1732 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1733 spin_unlock(&inode->i_lock);
1734 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1738 pnfs_list_write_lseg(inode, &data->lseg_list);
1740 end_pos = nfsi->layout->plh_lwb;
1741 nfsi->layout->plh_lwb = 0;
1743 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1744 spin_unlock(&inode->i_lock);
1746 data->args.inode = inode;
1747 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1748 nfs_fattr_init(&data->fattr);
1749 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1750 data->res.fattr = &data->fattr;
1751 data->args.lastbytewritten = end_pos - 1;
1752 data->res.server = NFS_SERVER(inode);
1754 status = nfs4_proc_layoutcommit(data, sync);
1757 mark_inode_dirty_sync(inode);
1758 dprintk("<-- %s status %d\n", __func__, status);
1765 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1767 struct nfs4_threshold *thp;
1769 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1771 dprintk("%s mdsthreshold allocation failed\n", __func__);