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;
372 /* last octet in a range */
374 last_byte_offset(u64 start, u64 len)
380 return end > start ? end - 1 : NFS4_MAX_UINT64;
384 * is l2 fully contained in l1?
386 * [----------------------------------)
391 lo_seg_contained(struct pnfs_layout_range *l1,
392 struct pnfs_layout_range *l2)
394 u64 start1 = l1->offset;
395 u64 end1 = end_offset(start1, l1->length);
396 u64 start2 = l2->offset;
397 u64 end2 = end_offset(start2, l2->length);
399 return (start1 <= start2) && (end1 >= end2);
403 * is l1 and l2 intersecting?
405 * [----------------------------------)
410 lo_seg_intersecting(struct pnfs_layout_range *l1,
411 struct pnfs_layout_range *l2)
413 u64 start1 = l1->offset;
414 u64 end1 = end_offset(start1, l1->length);
415 u64 start2 = l2->offset;
416 u64 end2 = end_offset(start2, l2->length);
418 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
419 (end2 == NFS4_MAX_UINT64 || end2 > start1);
423 should_free_lseg(struct pnfs_layout_range *lseg_range,
424 struct pnfs_layout_range *recall_range)
426 return (recall_range->iomode == IOMODE_ANY ||
427 lseg_range->iomode == recall_range->iomode) &&
428 lo_seg_intersecting(lseg_range, recall_range);
431 /* Returns 1 if lseg is removed from list, 0 otherwise */
432 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
433 struct list_head *tmp_list)
437 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
438 /* Remove the reference keeping the lseg in the
439 * list. It will now be removed when all
440 * outstanding io is finished.
442 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
443 atomic_read(&lseg->pls_refcount));
444 if (atomic_dec_and_test(&lseg->pls_refcount)) {
445 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
446 list_add(&lseg->pls_list, tmp_list);
453 /* Returns count of number of matching invalid lsegs remaining in list
457 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
458 struct list_head *tmp_list,
459 struct pnfs_layout_range *recall_range)
461 struct pnfs_layout_segment *lseg, *next;
462 int invalid = 0, removed = 0;
464 dprintk("%s:Begin lo %p\n", __func__, lo);
466 if (list_empty(&lo->plh_segs))
468 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
470 should_free_lseg(&lseg->pls_range, recall_range)) {
471 dprintk("%s: freeing lseg %p iomode %d "
472 "offset %llu length %llu\n", __func__,
473 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
474 lseg->pls_range.length);
476 removed += mark_lseg_invalid(lseg, tmp_list);
478 dprintk("%s:Return %i\n", __func__, invalid - removed);
479 return invalid - removed;
482 /* note free_me must contain lsegs from a single layout_hdr */
484 pnfs_free_lseg_list(struct list_head *free_me)
486 struct pnfs_layout_segment *lseg, *tmp;
488 if (list_empty(free_me))
491 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
492 list_del(&lseg->pls_list);
493 pnfs_free_lseg(lseg);
498 pnfs_destroy_layout(struct nfs_inode *nfsi)
500 struct pnfs_layout_hdr *lo;
503 spin_lock(&nfsi->vfs_inode.i_lock);
506 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
507 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
508 pnfs_get_layout_hdr(lo);
509 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
510 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
511 spin_unlock(&nfsi->vfs_inode.i_lock);
512 pnfs_free_lseg_list(&tmp_list);
513 pnfs_put_layout_hdr(lo);
515 spin_unlock(&nfsi->vfs_inode.i_lock);
517 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
520 * Called by the state manger to remove all layouts established under an
524 pnfs_destroy_all_layouts(struct nfs_client *clp)
526 struct nfs_server *server;
527 struct pnfs_layout_hdr *lo;
530 nfs4_deviceid_mark_client_invalid(clp);
531 nfs4_deviceid_purge_client(clp);
533 spin_lock(&clp->cl_lock);
535 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
536 if (!list_empty(&server->layouts))
537 list_splice_init(&server->layouts, &tmp_list);
540 spin_unlock(&clp->cl_lock);
542 while (!list_empty(&tmp_list)) {
543 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
545 dprintk("%s freeing layout for inode %lu\n", __func__,
546 lo->plh_inode->i_ino);
547 list_del_init(&lo->plh_layouts);
548 pnfs_destroy_layout(NFS_I(lo->plh_inode));
552 /* update lo->plh_stateid with new if is more recent */
554 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
559 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
560 newseq = be32_to_cpu(new->seqid);
561 if ((int)(newseq - oldseq) > 0) {
562 nfs4_stateid_copy(&lo->plh_stateid, new);
563 if (update_barrier) {
564 u32 new_barrier = be32_to_cpu(new->seqid);
566 if ((int)(new_barrier - lo->plh_barrier))
567 lo->plh_barrier = new_barrier;
569 /* Because of wraparound, we want to keep the barrier
570 * "close" to the current seqids. It needs to be
571 * within 2**31 to count as "behind", so if it
572 * gets too near that limit, give us a litle leeway
573 * and bring it to within 2**30.
574 * NOTE - and yes, this is all unsigned arithmetic.
576 if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
577 lo->plh_barrier = newseq - (1 << 30);
582 /* lget is set to 1 if called from inside send_layoutget call chain */
584 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
588 (int)(lo->plh_barrier - be32_to_cpu(stateid->seqid)) >= 0)
590 return lo->plh_block_lgets ||
591 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
592 (list_empty(&lo->plh_segs) &&
593 (atomic_read(&lo->plh_outstanding) > lget));
597 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
598 struct nfs4_state *open_state)
602 dprintk("--> %s\n", __func__);
603 spin_lock(&lo->plh_inode->i_lock);
604 if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
606 } else if (list_empty(&lo->plh_segs)) {
610 seq = read_seqbegin(&open_state->seqlock);
611 nfs4_stateid_copy(dst, &open_state->stateid);
612 } while (read_seqretry(&open_state->seqlock, seq));
614 nfs4_stateid_copy(dst, &lo->plh_stateid);
615 spin_unlock(&lo->plh_inode->i_lock);
616 dprintk("<-- %s\n", __func__);
621 * Get layout from server.
622 * for now, assume that whole file layouts are requested.
624 * arg->length: all ones
626 static struct pnfs_layout_segment *
627 send_layoutget(struct pnfs_layout_hdr *lo,
628 struct nfs_open_context *ctx,
629 struct pnfs_layout_range *range,
632 struct inode *ino = lo->plh_inode;
633 struct nfs_server *server = NFS_SERVER(ino);
634 struct nfs4_layoutget *lgp;
635 struct pnfs_layout_segment *lseg;
637 dprintk("--> %s\n", __func__);
640 lgp = kzalloc(sizeof(*lgp), gfp_flags);
644 lgp->args.minlength = PAGE_CACHE_SIZE;
645 if (lgp->args.minlength > range->length)
646 lgp->args.minlength = range->length;
647 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
648 lgp->args.range = *range;
649 lgp->args.type = server->pnfs_curr_ld->id;
650 lgp->args.inode = ino;
651 lgp->args.ctx = get_nfs_open_context(ctx);
652 lgp->gfp_flags = gfp_flags;
654 /* Synchronously retrieve layout information from server and
657 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
659 switch (PTR_ERR(lseg)) {
664 /* remember that LAYOUTGET failed and suspend trying */
665 pnfs_layout_io_set_failed(lo, range->iomode);
674 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
675 * when the layout segment list is empty.
677 * Note that a pnfs_layout_hdr can exist with an empty layout segment
678 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
679 * deviceid is marked invalid.
682 _pnfs_return_layout(struct inode *ino)
684 struct pnfs_layout_hdr *lo = NULL;
685 struct nfs_inode *nfsi = NFS_I(ino);
687 struct nfs4_layoutreturn *lrp;
688 nfs4_stateid stateid;
689 int status = 0, empty;
691 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
693 spin_lock(&ino->i_lock);
695 if (!lo || pnfs_test_layout_returned(lo)) {
696 spin_unlock(&ino->i_lock);
697 dprintk("NFS: %s no layout to return\n", __func__);
700 stateid = nfsi->layout->plh_stateid;
701 /* Reference matched in nfs4_layoutreturn_release */
702 pnfs_get_layout_hdr(lo);
703 empty = list_empty(&lo->plh_segs);
704 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
705 /* Don't send a LAYOUTRETURN if list was initially empty */
707 spin_unlock(&ino->i_lock);
708 pnfs_put_layout_hdr(lo);
709 dprintk("NFS: %s no layout segments to return\n", __func__);
712 lo->plh_block_lgets++;
713 pnfs_mark_layout_returned(lo);
714 spin_unlock(&ino->i_lock);
715 pnfs_free_lseg_list(&tmp_list);
717 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
719 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
720 if (unlikely(lrp == NULL)) {
722 pnfs_layout_io_set_failed(lo, IOMODE_RW);
723 pnfs_layout_io_set_failed(lo, IOMODE_READ);
724 pnfs_clear_layout_returned(lo);
725 pnfs_put_layout_hdr(lo);
729 lrp->args.stateid = stateid;
730 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
731 lrp->args.inode = ino;
732 lrp->args.layout = lo;
733 lrp->clp = NFS_SERVER(ino)->nfs_client;
735 status = nfs4_proc_layoutreturn(lrp);
737 dprintk("<-- %s status: %d\n", __func__, status);
740 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
742 bool pnfs_roc(struct inode *ino)
744 struct pnfs_layout_hdr *lo;
745 struct pnfs_layout_segment *lseg, *tmp;
749 spin_lock(&ino->i_lock);
750 lo = NFS_I(ino)->layout;
751 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
752 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
754 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
755 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
756 mark_lseg_invalid(lseg, &tmp_list);
761 lo->plh_block_lgets++;
762 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
763 spin_unlock(&ino->i_lock);
764 pnfs_free_lseg_list(&tmp_list);
768 spin_unlock(&ino->i_lock);
772 void pnfs_roc_release(struct inode *ino)
774 struct pnfs_layout_hdr *lo;
776 spin_lock(&ino->i_lock);
777 lo = NFS_I(ino)->layout;
778 lo->plh_block_lgets--;
779 if (atomic_dec_and_test(&lo->plh_refcount)) {
780 pnfs_detach_layout_hdr(lo);
781 spin_unlock(&ino->i_lock);
782 pnfs_free_layout_hdr(lo);
784 spin_unlock(&ino->i_lock);
787 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
789 struct pnfs_layout_hdr *lo;
791 spin_lock(&ino->i_lock);
792 lo = NFS_I(ino)->layout;
793 if ((int)(barrier - lo->plh_barrier) > 0)
794 lo->plh_barrier = barrier;
795 spin_unlock(&ino->i_lock);
798 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
800 struct nfs_inode *nfsi = NFS_I(ino);
801 struct pnfs_layout_hdr *lo;
802 struct pnfs_layout_segment *lseg;
806 spin_lock(&ino->i_lock);
807 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
808 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
809 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
814 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
816 /* Since close does not return a layout stateid for use as
817 * a barrier, we choose the worst-case barrier.
819 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
821 spin_unlock(&ino->i_lock);
826 * Compare two layout segments for sorting into layout cache.
827 * We want to preferentially return RW over RO layouts, so ensure those
831 cmp_layout(struct pnfs_layout_range *l1,
832 struct pnfs_layout_range *l2)
836 /* high offset > low offset */
837 d = l1->offset - l2->offset;
841 /* short length > long length */
842 d = l2->length - l1->length;
846 /* read > read/write */
847 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
851 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
852 struct pnfs_layout_segment *lseg)
854 struct pnfs_layout_segment *lp;
856 dprintk("%s:Begin\n", __func__);
858 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
859 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
861 list_add_tail(&lseg->pls_list, &lp->pls_list);
862 dprintk("%s: inserted lseg %p "
863 "iomode %d offset %llu length %llu before "
864 "lp %p iomode %d offset %llu length %llu\n",
865 __func__, lseg, lseg->pls_range.iomode,
866 lseg->pls_range.offset, lseg->pls_range.length,
867 lp, lp->pls_range.iomode, lp->pls_range.offset,
868 lp->pls_range.length);
871 list_add_tail(&lseg->pls_list, &lo->plh_segs);
872 dprintk("%s: inserted lseg %p "
873 "iomode %d offset %llu length %llu at tail\n",
874 __func__, lseg, lseg->pls_range.iomode,
875 lseg->pls_range.offset, lseg->pls_range.length);
877 pnfs_get_layout_hdr(lo);
879 dprintk("%s:Return\n", __func__);
882 static struct pnfs_layout_hdr *
883 alloc_init_layout_hdr(struct inode *ino,
884 struct nfs_open_context *ctx,
887 struct pnfs_layout_hdr *lo;
889 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
892 atomic_set(&lo->plh_refcount, 1);
893 INIT_LIST_HEAD(&lo->plh_layouts);
894 INIT_LIST_HEAD(&lo->plh_segs);
895 INIT_LIST_HEAD(&lo->plh_bulk_recall);
897 lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
901 static struct pnfs_layout_hdr *
902 pnfs_find_alloc_layout(struct inode *ino,
903 struct nfs_open_context *ctx,
906 struct nfs_inode *nfsi = NFS_I(ino);
907 struct pnfs_layout_hdr *new = NULL;
909 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
912 pnfs_get_layout_hdr(nfsi->layout);
915 spin_unlock(&ino->i_lock);
916 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
917 spin_lock(&ino->i_lock);
919 if (likely(nfsi->layout == NULL)) /* Won the race? */
922 pnfs_free_layout_hdr(new);
927 * iomode matching rules:
938 is_matching_lseg(struct pnfs_layout_range *ls_range,
939 struct pnfs_layout_range *range)
941 struct pnfs_layout_range range1;
943 if ((range->iomode == IOMODE_RW &&
944 ls_range->iomode != IOMODE_RW) ||
945 !lo_seg_intersecting(ls_range, range))
948 /* range1 covers only the first byte in the range */
951 return lo_seg_contained(ls_range, &range1);
955 * lookup range in layout
957 static struct pnfs_layout_segment *
958 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
959 struct pnfs_layout_range *range)
961 struct pnfs_layout_segment *lseg, *ret = NULL;
963 dprintk("%s:Begin\n", __func__);
965 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
966 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
967 is_matching_lseg(&lseg->pls_range, range)) {
968 ret = pnfs_get_lseg(lseg);
971 if (lseg->pls_range.offset > range->offset)
975 dprintk("%s:Return lseg %p ref %d\n",
976 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
981 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
982 * to the MDS or over pNFS
984 * The nfs_inode read_io and write_io fields are cumulative counters reset
985 * when there are no layout segments. Note that in pnfs_update_layout iomode
986 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
989 * A return of true means use MDS I/O.
992 * If a file's size is smaller than the file size threshold, data accesses
993 * SHOULD be sent to the metadata server. If an I/O request has a length that
994 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
995 * server. If both file size and I/O size are provided, the client SHOULD
996 * reach or exceed both thresholds before sending its read or write
997 * requests to the data server.
999 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1000 struct inode *ino, int iomode)
1002 struct nfs4_threshold *t = ctx->mdsthreshold;
1003 struct nfs_inode *nfsi = NFS_I(ino);
1004 loff_t fsize = i_size_read(ino);
1005 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1010 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1011 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1015 if (t->bm & THRESHOLD_RD) {
1016 dprintk("%s fsize %llu\n", __func__, fsize);
1018 if (fsize < t->rd_sz)
1021 if (t->bm & THRESHOLD_RD_IO) {
1022 dprintk("%s nfsi->read_io %llu\n", __func__,
1025 if (nfsi->read_io < t->rd_io_sz)
1030 if (t->bm & THRESHOLD_WR) {
1031 dprintk("%s fsize %llu\n", __func__, fsize);
1033 if (fsize < t->wr_sz)
1036 if (t->bm & THRESHOLD_WR_IO) {
1037 dprintk("%s nfsi->write_io %llu\n", __func__,
1040 if (nfsi->write_io < t->wr_io_sz)
1045 if (size_set && io_set) {
1048 } else if (size || io)
1051 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1056 * Layout segment is retreived from the server if not cached.
1057 * The appropriate layout segment is referenced and returned to the caller.
1059 struct pnfs_layout_segment *
1060 pnfs_update_layout(struct inode *ino,
1061 struct nfs_open_context *ctx,
1064 enum pnfs_iomode iomode,
1067 struct pnfs_layout_range arg = {
1073 struct nfs_server *server = NFS_SERVER(ino);
1074 struct nfs_client *clp = server->nfs_client;
1075 struct pnfs_layout_hdr *lo;
1076 struct pnfs_layout_segment *lseg = NULL;
1079 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1082 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1085 spin_lock(&ino->i_lock);
1086 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1088 spin_unlock(&ino->i_lock);
1092 /* Do we even need to bother with this? */
1093 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1094 dprintk("%s matches recall, use MDS\n", __func__);
1098 /* if LAYOUTGET already failed once we don't try again */
1099 if (pnfs_layout_io_test_failed(lo, iomode))
1102 /* Check to see if the layout for the given range already exists */
1103 lseg = pnfs_find_lseg(lo, &arg);
1107 if (pnfs_layoutgets_blocked(lo, NULL, 0))
1109 atomic_inc(&lo->plh_outstanding);
1111 if (list_empty(&lo->plh_segs))
1114 /* Enable LAYOUTRETURNs */
1115 pnfs_clear_layout_returned(lo);
1117 spin_unlock(&ino->i_lock);
1119 /* The lo must be on the clp list if there is any
1120 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1122 spin_lock(&clp->cl_lock);
1123 BUG_ON(!list_empty(&lo->plh_layouts));
1124 list_add_tail(&lo->plh_layouts, &server->layouts);
1125 spin_unlock(&clp->cl_lock);
1128 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1130 arg.offset -= pg_offset;
1131 arg.length += pg_offset;
1133 if (arg.length != NFS4_MAX_UINT64)
1134 arg.length = PAGE_CACHE_ALIGN(arg.length);
1136 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1137 atomic_dec(&lo->plh_outstanding);
1139 pnfs_put_layout_hdr(lo);
1141 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1142 "(%s, offset: %llu, length: %llu)\n",
1143 __func__, ino->i_sb->s_id,
1144 (unsigned long long)NFS_FILEID(ino),
1145 lseg == NULL ? "not found" : "found",
1146 iomode==IOMODE_RW ? "read/write" : "read-only",
1147 (unsigned long long)pos,
1148 (unsigned long long)count);
1151 spin_unlock(&ino->i_lock);
1152 goto out_put_layout_hdr;
1154 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1156 struct pnfs_layout_segment *
1157 pnfs_layout_process(struct nfs4_layoutget *lgp)
1159 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1160 struct nfs4_layoutget_res *res = &lgp->res;
1161 struct pnfs_layout_segment *lseg;
1162 struct inode *ino = lo->plh_inode;
1165 /* Inject layout blob into I/O device driver */
1166 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1167 if (!lseg || IS_ERR(lseg)) {
1171 status = PTR_ERR(lseg);
1172 dprintk("%s: Could not allocate layout: error %d\n",
1177 spin_lock(&ino->i_lock);
1178 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1179 dprintk("%s forget reply due to recall\n", __func__);
1180 goto out_forget_reply;
1183 if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1184 dprintk("%s forget reply due to state\n", __func__);
1185 goto out_forget_reply;
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 /* Done processing layoutget. Set the layout stateid */
1198 pnfs_set_layout_stateid(lo, &res->stateid, false);
1199 spin_unlock(&ino->i_lock);
1202 return ERR_PTR(status);
1205 spin_unlock(&ino->i_lock);
1206 lseg->pls_layout = lo;
1207 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1212 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1214 BUG_ON(pgio->pg_lseg != NULL);
1216 if (req->wb_offset != req->wb_pgbase) {
1217 nfs_pageio_reset_read_mds(pgio);
1220 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1226 /* If no lseg, fall back to read through mds */
1227 if (pgio->pg_lseg == NULL)
1228 nfs_pageio_reset_read_mds(pgio);
1231 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1234 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1236 BUG_ON(pgio->pg_lseg != NULL);
1238 if (req->wb_offset != req->wb_pgbase) {
1239 nfs_pageio_reset_write_mds(pgio);
1242 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1248 /* If no lseg, fall back to write through mds */
1249 if (pgio->pg_lseg == NULL)
1250 nfs_pageio_reset_write_mds(pgio);
1252 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1255 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1256 const struct nfs_pgio_completion_ops *compl_ops)
1258 struct nfs_server *server = NFS_SERVER(inode);
1259 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1262 nfs_pageio_init_read(pgio, inode, compl_ops);
1264 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1268 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1270 const struct nfs_pgio_completion_ops *compl_ops)
1272 struct nfs_server *server = NFS_SERVER(inode);
1273 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1276 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1278 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1282 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1283 struct nfs_page *req)
1285 if (pgio->pg_lseg == NULL)
1286 return nfs_generic_pg_test(pgio, prev, req);
1289 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1290 * Note that this test makes several assumptions:
1291 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1292 * is known to lie within the range.
1293 * - that the nfs_page being tested is known to be contiguous with the
1294 * previous nfs_page.
1295 * - Layout ranges are page aligned, so we only have to test the
1296 * start offset of the request.
1298 * Please also note that 'end_offset' is actually the offset of the
1299 * first byte that lies outside the pnfs_layout_range. FIXME?
1302 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1303 pgio->pg_lseg->pls_range.length);
1305 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1307 int pnfs_write_done_resend_to_mds(struct inode *inode,
1308 struct list_head *head,
1309 const struct nfs_pgio_completion_ops *compl_ops)
1311 struct nfs_pageio_descriptor pgio;
1314 /* Resend all requests through the MDS */
1315 nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1316 while (!list_empty(head)) {
1317 struct nfs_page *req = nfs_list_entry(head->next);
1319 nfs_list_remove_request(req);
1320 if (!nfs_pageio_add_request(&pgio, req))
1321 nfs_list_add_request(req, &failed);
1323 nfs_pageio_complete(&pgio);
1325 if (!list_empty(&failed)) {
1326 /* For some reason our attempt to resend pages. Mark the
1327 * overall send request as having failed, and let
1328 * nfs_writeback_release_full deal with the error.
1330 list_move(&failed, head);
1335 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1337 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1339 struct nfs_pgio_header *hdr = data->header;
1341 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1342 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1343 PNFS_LAYOUTRET_ON_ERROR) {
1344 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1345 pnfs_return_layout(hdr->inode);
1347 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1348 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1350 hdr->completion_ops);
1354 * Called by non rpc-based layout drivers
1356 void pnfs_ld_write_done(struct nfs_write_data *data)
1358 struct nfs_pgio_header *hdr = data->header;
1360 if (!hdr->pnfs_error) {
1361 pnfs_set_layoutcommit(data);
1362 hdr->mds_ops->rpc_call_done(&data->task, data);
1364 pnfs_ld_handle_write_error(data);
1365 hdr->mds_ops->rpc_release(data);
1367 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1370 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1371 struct nfs_write_data *data)
1373 struct nfs_pgio_header *hdr = data->header;
1375 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1376 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1377 nfs_pageio_reset_write_mds(desc);
1378 desc->pg_recoalesce = 1;
1380 nfs_writedata_release(data);
1383 static enum pnfs_try_status
1384 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1385 const struct rpc_call_ops *call_ops,
1386 struct pnfs_layout_segment *lseg,
1389 struct nfs_pgio_header *hdr = wdata->header;
1390 struct inode *inode = hdr->inode;
1391 enum pnfs_try_status trypnfs;
1392 struct nfs_server *nfss = NFS_SERVER(inode);
1394 hdr->mds_ops = call_ops;
1396 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1397 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1398 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1399 if (trypnfs != PNFS_NOT_ATTEMPTED)
1400 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1401 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1406 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1408 struct nfs_write_data *data;
1409 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1410 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1412 desc->pg_lseg = NULL;
1413 while (!list_empty(head)) {
1414 enum pnfs_try_status trypnfs;
1416 data = list_first_entry(head, struct nfs_write_data, list);
1417 list_del_init(&data->list);
1419 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1420 if (trypnfs == PNFS_NOT_ATTEMPTED)
1421 pnfs_write_through_mds(desc, data);
1423 pnfs_put_lseg(lseg);
1426 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1428 pnfs_put_lseg(hdr->lseg);
1429 nfs_writehdr_free(hdr);
1431 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1434 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1436 struct nfs_write_header *whdr;
1437 struct nfs_pgio_header *hdr;
1440 whdr = nfs_writehdr_alloc();
1442 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1443 pnfs_put_lseg(desc->pg_lseg);
1444 desc->pg_lseg = NULL;
1447 hdr = &whdr->header;
1448 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1449 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1450 atomic_inc(&hdr->refcnt);
1451 ret = nfs_generic_flush(desc, hdr);
1453 pnfs_put_lseg(desc->pg_lseg);
1454 desc->pg_lseg = NULL;
1456 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1457 if (atomic_dec_and_test(&hdr->refcnt))
1458 hdr->completion_ops->completion(hdr);
1461 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1463 int pnfs_read_done_resend_to_mds(struct inode *inode,
1464 struct list_head *head,
1465 const struct nfs_pgio_completion_ops *compl_ops)
1467 struct nfs_pageio_descriptor pgio;
1470 /* Resend all requests through the MDS */
1471 nfs_pageio_init_read(&pgio, inode, compl_ops);
1472 while (!list_empty(head)) {
1473 struct nfs_page *req = nfs_list_entry(head->next);
1475 nfs_list_remove_request(req);
1476 if (!nfs_pageio_add_request(&pgio, req))
1477 nfs_list_add_request(req, &failed);
1479 nfs_pageio_complete(&pgio);
1481 if (!list_empty(&failed)) {
1482 list_move(&failed, head);
1487 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1489 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1491 struct nfs_pgio_header *hdr = data->header;
1493 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1494 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1495 PNFS_LAYOUTRET_ON_ERROR) {
1496 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(hdr->inode)->flags);
1497 pnfs_return_layout(hdr->inode);
1499 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1500 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1502 hdr->completion_ops);
1506 * Called by non rpc-based layout drivers
1508 void pnfs_ld_read_done(struct nfs_read_data *data)
1510 struct nfs_pgio_header *hdr = data->header;
1512 if (likely(!hdr->pnfs_error)) {
1513 __nfs4_read_done_cb(data);
1514 hdr->mds_ops->rpc_call_done(&data->task, data);
1516 pnfs_ld_handle_read_error(data);
1517 hdr->mds_ops->rpc_release(data);
1519 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1522 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1523 struct nfs_read_data *data)
1525 struct nfs_pgio_header *hdr = data->header;
1527 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1528 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1529 nfs_pageio_reset_read_mds(desc);
1530 desc->pg_recoalesce = 1;
1532 nfs_readdata_release(data);
1536 * Call the appropriate parallel I/O subsystem read function.
1538 static enum pnfs_try_status
1539 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1540 const struct rpc_call_ops *call_ops,
1541 struct pnfs_layout_segment *lseg)
1543 struct nfs_pgio_header *hdr = rdata->header;
1544 struct inode *inode = hdr->inode;
1545 struct nfs_server *nfss = NFS_SERVER(inode);
1546 enum pnfs_try_status trypnfs;
1548 hdr->mds_ops = call_ops;
1550 dprintk("%s: Reading ino:%lu %u@%llu\n",
1551 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1553 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1554 if (trypnfs != PNFS_NOT_ATTEMPTED)
1555 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1556 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1561 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1563 struct nfs_read_data *data;
1564 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1565 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1567 desc->pg_lseg = NULL;
1568 while (!list_empty(head)) {
1569 enum pnfs_try_status trypnfs;
1571 data = list_first_entry(head, struct nfs_read_data, list);
1572 list_del_init(&data->list);
1574 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1575 if (trypnfs == PNFS_NOT_ATTEMPTED)
1576 pnfs_read_through_mds(desc, data);
1578 pnfs_put_lseg(lseg);
1581 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1583 pnfs_put_lseg(hdr->lseg);
1584 nfs_readhdr_free(hdr);
1586 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1589 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1591 struct nfs_read_header *rhdr;
1592 struct nfs_pgio_header *hdr;
1595 rhdr = nfs_readhdr_alloc();
1597 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1599 pnfs_put_lseg(desc->pg_lseg);
1600 desc->pg_lseg = NULL;
1603 hdr = &rhdr->header;
1604 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1605 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1606 atomic_inc(&hdr->refcnt);
1607 ret = nfs_generic_pagein(desc, hdr);
1609 pnfs_put_lseg(desc->pg_lseg);
1610 desc->pg_lseg = NULL;
1612 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1613 if (atomic_dec_and_test(&hdr->refcnt))
1614 hdr->completion_ops->completion(hdr);
1617 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1620 * There can be multiple RW segments.
1622 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1624 struct pnfs_layout_segment *lseg;
1626 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1627 if (lseg->pls_range.iomode == IOMODE_RW &&
1628 test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1629 list_add(&lseg->pls_lc_list, listp);
1633 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1635 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1637 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1640 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1642 struct nfs_pgio_header *hdr = wdata->header;
1643 struct inode *inode = hdr->inode;
1644 struct nfs_inode *nfsi = NFS_I(inode);
1645 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1646 bool mark_as_dirty = false;
1648 spin_lock(&inode->i_lock);
1649 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1650 mark_as_dirty = true;
1651 dprintk("%s: Set layoutcommit for inode %lu ",
1652 __func__, inode->i_ino);
1654 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1655 /* references matched in nfs4_layoutcommit_release */
1656 pnfs_get_lseg(hdr->lseg);
1658 if (end_pos > nfsi->layout->plh_lwb)
1659 nfsi->layout->plh_lwb = end_pos;
1660 spin_unlock(&inode->i_lock);
1661 dprintk("%s: lseg %p end_pos %llu\n",
1662 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1664 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1665 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1667 mark_inode_dirty_sync(inode);
1669 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1671 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1673 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1675 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1676 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1680 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1681 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1682 * data to disk to allow the server to recover the data if it crashes.
1683 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1684 * is off, and a COMMIT is sent to a data server, or
1685 * if WRITEs to a data server return NFS_DATA_SYNC.
1688 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1690 struct nfs4_layoutcommit_data *data;
1691 struct nfs_inode *nfsi = NFS_I(inode);
1695 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1697 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1700 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1701 data = kzalloc(sizeof(*data), GFP_NOFS);
1707 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1710 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1715 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1716 nfs_wait_bit_killable, TASK_KILLABLE);
1721 INIT_LIST_HEAD(&data->lseg_list);
1722 spin_lock(&inode->i_lock);
1723 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1724 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1725 spin_unlock(&inode->i_lock);
1726 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1730 pnfs_list_write_lseg(inode, &data->lseg_list);
1732 end_pos = nfsi->layout->plh_lwb;
1733 nfsi->layout->plh_lwb = 0;
1735 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1736 spin_unlock(&inode->i_lock);
1738 data->args.inode = inode;
1739 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1740 nfs_fattr_init(&data->fattr);
1741 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1742 data->res.fattr = &data->fattr;
1743 data->args.lastbytewritten = end_pos - 1;
1744 data->res.server = NFS_SERVER(inode);
1746 status = nfs4_proc_layoutcommit(data, sync);
1749 mark_inode_dirty_sync(inode);
1750 dprintk("<-- %s status %d\n", __func__, status);
1757 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1759 struct nfs4_threshold *thp;
1761 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1763 dprintk("%s mdsthreshold allocation failed\n", __func__);