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);
58 /* Return the registered pnfs layout driver module matching given id */
59 static struct pnfs_layoutdriver_type *
60 find_pnfs_driver_locked(u32 id)
62 struct pnfs_layoutdriver_type *local;
64 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
69 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
73 static struct pnfs_layoutdriver_type *
74 find_pnfs_driver(u32 id)
76 struct pnfs_layoutdriver_type *local;
78 spin_lock(&pnfs_spinlock);
79 local = find_pnfs_driver_locked(id);
80 if (local != NULL && !try_module_get(local->owner)) {
81 dprintk("%s: Could not grab reference on module\n", __func__);
84 spin_unlock(&pnfs_spinlock);
89 unset_pnfs_layoutdriver(struct nfs_server *nfss)
91 if (nfss->pnfs_curr_ld) {
92 if (nfss->pnfs_curr_ld->clear_layoutdriver)
93 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
94 /* Decrement the MDS count. Purge the deviceid cache if zero */
95 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
96 nfs4_deviceid_purge_client(nfss->nfs_client);
97 module_put(nfss->pnfs_curr_ld->owner);
99 nfss->pnfs_curr_ld = NULL;
103 * When the server sends a list of layout types, we choose one in the order
104 * given in the list below.
106 * FIXME: should this list be configurable in some fashion? module param?
107 * mount option? something else?
109 static const u32 ld_prefs[] = {
114 LAYOUT_NFSV4_1_FILES,
119 ld_cmp(const void *e1, const void *e2)
121 u32 ld1 = *((u32 *)e1);
122 u32 ld2 = *((u32 *)e2);
125 for (i = 0; ld_prefs[i] != 0; i++) {
126 if (ld1 == ld_prefs[i])
129 if (ld2 == ld_prefs[i])
136 * Try to set the server's pnfs module to the pnfs layout type specified by id.
137 * Currently only one pNFS layout driver per filesystem is supported.
139 * @ids array of layout types supported by MDS.
142 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
143 struct nfs_fsinfo *fsinfo)
145 struct pnfs_layoutdriver_type *ld_type = NULL;
149 if (!(server->nfs_client->cl_exchange_flags &
150 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
151 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
152 __func__, server->nfs_client->cl_exchange_flags);
156 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
157 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
159 for (i = 0; i < fsinfo->nlayouttypes; i++) {
160 id = fsinfo->layouttype[i];
161 ld_type = find_pnfs_driver(id);
163 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
165 ld_type = find_pnfs_driver(id);
172 dprintk("%s: No pNFS module found!\n", __func__);
176 server->pnfs_curr_ld = ld_type;
177 if (ld_type->set_layoutdriver
178 && ld_type->set_layoutdriver(server, mntfh)) {
179 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
180 "driver %u.\n", __func__, id);
181 module_put(ld_type->owner);
184 /* Bump the MDS count */
185 atomic_inc(&server->nfs_client->cl_mds_count);
187 dprintk("%s: pNFS module for %u set\n", __func__, id);
191 dprintk("%s: Using NFSv4 I/O\n", __func__);
192 server->pnfs_curr_ld = NULL;
196 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
198 int status = -EINVAL;
199 struct pnfs_layoutdriver_type *tmp;
201 if (ld_type->id == 0) {
202 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
205 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
206 printk(KERN_ERR "NFS: %s Layout driver must provide "
207 "alloc_lseg and free_lseg.\n", __func__);
211 spin_lock(&pnfs_spinlock);
212 tmp = find_pnfs_driver_locked(ld_type->id);
214 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
216 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
219 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
220 __func__, ld_type->id);
222 spin_unlock(&pnfs_spinlock);
226 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
229 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
231 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
232 spin_lock(&pnfs_spinlock);
233 list_del(&ld_type->pnfs_tblid);
234 spin_unlock(&pnfs_spinlock);
236 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
239 * pNFS client layout cache
242 /* Need to hold i_lock if caller does not already hold reference */
244 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
246 atomic_inc(&lo->plh_refcount);
249 static struct pnfs_layout_hdr *
250 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
252 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
253 return ld->alloc_layout_hdr(ino, gfp_flags);
257 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
259 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
260 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
262 if (!list_empty(&lo->plh_layouts)) {
263 struct nfs_client *clp = server->nfs_client;
265 spin_lock(&clp->cl_lock);
266 list_del_init(&lo->plh_layouts);
267 spin_unlock(&clp->cl_lock);
269 put_rpccred(lo->plh_lc_cred);
270 return ld->free_layout_hdr(lo);
274 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
276 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
277 dprintk("%s: freeing layout cache %p\n", __func__, lo);
279 /* Reset MDS Threshold I/O counters */
285 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
287 struct inode *inode = lo->plh_inode;
289 pnfs_layoutreturn_before_put_layout_hdr(lo);
291 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
292 if (!list_empty(&lo->plh_segs))
293 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
294 pnfs_detach_layout_hdr(lo);
295 spin_unlock(&inode->i_lock);
296 pnfs_free_layout_hdr(lo);
301 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
303 * In order to continue using the pnfs_layout_hdr, a full recovery
305 * Note that caller must hold inode->i_lock.
308 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
309 struct list_head *lseg_list)
311 struct pnfs_layout_range range = {
312 .iomode = IOMODE_ANY,
314 .length = NFS4_MAX_UINT64,
317 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
318 return pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range, 0);
322 pnfs_iomode_to_fail_bit(u32 iomode)
324 return iomode == IOMODE_RW ?
325 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
329 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
331 lo->plh_retry_timestamp = jiffies;
332 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
333 atomic_inc(&lo->plh_refcount);
337 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
339 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
340 atomic_dec(&lo->plh_refcount);
344 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
346 struct inode *inode = lo->plh_inode;
347 struct pnfs_layout_range range = {
350 .length = NFS4_MAX_UINT64,
354 spin_lock(&inode->i_lock);
355 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
356 pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
357 spin_unlock(&inode->i_lock);
358 pnfs_free_lseg_list(&head);
359 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
360 iomode == IOMODE_RW ? "RW" : "READ");
364 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
366 unsigned long start, end;
367 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
369 if (test_bit(fail_bit, &lo->plh_flags) == 0)
372 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
373 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
374 /* It is time to retry the failed layoutgets */
375 pnfs_layout_clear_fail_bit(lo, fail_bit);
382 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
383 const struct pnfs_layout_range *range,
384 const nfs4_stateid *stateid)
386 INIT_LIST_HEAD(&lseg->pls_list);
387 INIT_LIST_HEAD(&lseg->pls_lc_list);
388 atomic_set(&lseg->pls_refcount, 1);
389 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
390 lseg->pls_layout = lo;
391 lseg->pls_range = *range;
392 lseg->pls_seq = be32_to_cpu(stateid->seqid);
395 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
397 struct inode *ino = lseg->pls_layout->plh_inode;
399 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
403 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
404 struct pnfs_layout_segment *lseg)
406 struct inode *inode = lo->plh_inode;
408 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
409 list_del_init(&lseg->pls_list);
410 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
411 atomic_dec(&lo->plh_refcount);
412 if (list_empty(&lo->plh_segs)) {
413 if (atomic_read(&lo->plh_outstanding) == 0)
414 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
415 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
417 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
421 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
423 struct pnfs_layout_hdr *lo;
429 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
430 atomic_read(&lseg->pls_refcount),
431 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
433 lo = lseg->pls_layout;
434 inode = lo->plh_inode;
436 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
437 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
438 spin_unlock(&inode->i_lock);
441 pnfs_get_layout_hdr(lo);
442 pnfs_layout_remove_lseg(lo, lseg);
443 spin_unlock(&inode->i_lock);
444 pnfs_free_lseg(lseg);
445 pnfs_put_layout_hdr(lo);
448 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
450 static void pnfs_free_lseg_async_work(struct work_struct *work)
452 struct pnfs_layout_segment *lseg;
453 struct pnfs_layout_hdr *lo;
455 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
456 lo = lseg->pls_layout;
458 pnfs_free_lseg(lseg);
459 pnfs_put_layout_hdr(lo);
462 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
464 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
465 schedule_work(&lseg->pls_work);
469 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
474 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
476 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
477 atomic_read(&lseg->pls_refcount),
478 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
479 if (atomic_dec_and_test(&lseg->pls_refcount)) {
480 struct pnfs_layout_hdr *lo = lseg->pls_layout;
481 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
483 pnfs_get_layout_hdr(lo);
484 pnfs_layout_remove_lseg(lo, lseg);
485 pnfs_free_lseg_async(lseg);
488 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
491 end_offset(u64 start, u64 len)
496 return end >= start ? end : NFS4_MAX_UINT64;
500 * is l2 fully contained in l1?
502 * [----------------------------------)
507 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
508 const struct pnfs_layout_range *l2)
510 u64 start1 = l1->offset;
511 u64 end1 = end_offset(start1, l1->length);
512 u64 start2 = l2->offset;
513 u64 end2 = end_offset(start2, l2->length);
515 return (start1 <= start2) && (end1 >= end2);
519 * is l1 and l2 intersecting?
521 * [----------------------------------)
526 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
527 const struct pnfs_layout_range *l2)
529 u64 start1 = l1->offset;
530 u64 end1 = end_offset(start1, l1->length);
531 u64 start2 = l2->offset;
532 u64 end2 = end_offset(start2, l2->length);
534 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
535 (end2 == NFS4_MAX_UINT64 || end2 > start1);
538 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
539 struct list_head *tmp_list)
541 if (!atomic_dec_and_test(&lseg->pls_refcount))
543 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
544 list_add(&lseg->pls_list, tmp_list);
548 /* Returns 1 if lseg is removed from list, 0 otherwise */
549 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
550 struct list_head *tmp_list)
554 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
555 /* Remove the reference keeping the lseg in the
556 * list. It will now be removed when all
557 * outstanding io is finished.
559 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
560 atomic_read(&lseg->pls_refcount));
561 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
568 * Compare 2 layout stateid sequence ids, to see which is newer,
569 * taking into account wraparound issues.
571 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
573 return (s32)(s1 - s2) > 0;
577 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
578 const struct pnfs_layout_range *recall_range)
580 return (recall_range->iomode == IOMODE_ANY ||
581 lseg_range->iomode == recall_range->iomode) &&
582 pnfs_lseg_range_intersecting(lseg_range, recall_range);
586 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
587 const struct pnfs_layout_range *recall_range,
590 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
592 if (recall_range == NULL)
594 return pnfs_should_free_range(&lseg->pls_range, recall_range);
598 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
599 * @lo: layout header containing the lsegs
600 * @tmp_list: list head where doomed lsegs should go
601 * @recall_range: optional recall range argument to match (may be NULL)
602 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
604 * Walk the list of lsegs in the layout header, and tear down any that should
605 * be destroyed. If "recall_range" is specified then the segment must match
606 * that range. If "seq" is non-zero, then only match segments that were handed
607 * out at or before that sequence.
609 * Returns number of matching invalid lsegs remaining in list after scanning
610 * it and purging them.
613 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
614 struct list_head *tmp_list,
615 const struct pnfs_layout_range *recall_range,
618 struct pnfs_layout_segment *lseg, *next;
621 dprintk("%s:Begin lo %p\n", __func__, lo);
623 if (list_empty(&lo->plh_segs))
625 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
626 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
627 dprintk("%s: freeing lseg %p iomode %d seq %u"
628 "offset %llu length %llu\n", __func__,
629 lseg, lseg->pls_range.iomode, lseg->pls_seq,
630 lseg->pls_range.offset, lseg->pls_range.length);
631 if (!mark_lseg_invalid(lseg, tmp_list))
634 dprintk("%s:Return %i\n", __func__, remaining);
638 /* note free_me must contain lsegs from a single layout_hdr */
640 pnfs_free_lseg_list(struct list_head *free_me)
642 struct pnfs_layout_segment *lseg, *tmp;
644 if (list_empty(free_me))
647 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
648 list_del(&lseg->pls_list);
649 pnfs_free_lseg(lseg);
654 pnfs_destroy_layout(struct nfs_inode *nfsi)
656 struct pnfs_layout_hdr *lo;
659 spin_lock(&nfsi->vfs_inode.i_lock);
662 pnfs_get_layout_hdr(lo);
663 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
664 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
665 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
666 spin_unlock(&nfsi->vfs_inode.i_lock);
667 pnfs_free_lseg_list(&tmp_list);
668 pnfs_put_layout_hdr(lo);
670 spin_unlock(&nfsi->vfs_inode.i_lock);
672 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
675 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
676 struct list_head *layout_list)
678 struct pnfs_layout_hdr *lo;
681 spin_lock(&inode->i_lock);
682 lo = NFS_I(inode)->layout;
683 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
684 pnfs_get_layout_hdr(lo);
685 list_add(&lo->plh_bulk_destroy, layout_list);
688 spin_unlock(&inode->i_lock);
692 /* Caller must hold rcu_read_lock and clp->cl_lock */
694 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
695 struct nfs_server *server,
696 struct list_head *layout_list)
698 struct pnfs_layout_hdr *lo, *next;
701 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
702 inode = igrab(lo->plh_inode);
705 list_del_init(&lo->plh_layouts);
706 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
709 spin_unlock(&clp->cl_lock);
711 spin_lock(&clp->cl_lock);
719 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
722 struct pnfs_layout_hdr *lo;
724 LIST_HEAD(lseg_list);
727 while (!list_empty(layout_list)) {
728 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
730 dprintk("%s freeing layout for inode %lu\n", __func__,
731 lo->plh_inode->i_ino);
732 inode = lo->plh_inode;
734 pnfs_layoutcommit_inode(inode, false);
736 spin_lock(&inode->i_lock);
737 list_del_init(&lo->plh_bulk_destroy);
738 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
740 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
743 spin_unlock(&inode->i_lock);
744 pnfs_free_lseg_list(&lseg_list);
745 /* Free all lsegs that are attached to commit buckets */
746 nfs_commit_inode(inode, 0);
747 pnfs_put_layout_hdr(lo);
754 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
755 struct nfs_fsid *fsid,
758 struct nfs_server *server;
759 LIST_HEAD(layout_list);
761 spin_lock(&clp->cl_lock);
764 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
765 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
767 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
773 spin_unlock(&clp->cl_lock);
775 if (list_empty(&layout_list))
777 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
781 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
784 struct nfs_server *server;
785 LIST_HEAD(layout_list);
787 spin_lock(&clp->cl_lock);
790 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
791 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
797 spin_unlock(&clp->cl_lock);
799 if (list_empty(&layout_list))
801 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
805 * Called by the state manger to remove all layouts established under an
809 pnfs_destroy_all_layouts(struct nfs_client *clp)
811 nfs4_deviceid_mark_client_invalid(clp);
812 nfs4_deviceid_purge_client(clp);
814 pnfs_destroy_layouts_byclid(clp, false);
818 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
820 lo->plh_return_iomode = 0;
821 lo->plh_return_seq = 0;
822 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
825 /* update lo->plh_stateid with new if is more recent */
827 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
830 u32 oldseq, newseq, new_barrier = 0;
832 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
833 newseq = be32_to_cpu(new->seqid);
835 if (!pnfs_layout_is_valid(lo)) {
836 nfs4_stateid_copy(&lo->plh_stateid, new);
837 lo->plh_barrier = newseq;
838 pnfs_clear_layoutreturn_info(lo);
839 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
842 if (pnfs_seqid_is_newer(newseq, oldseq)) {
843 nfs4_stateid_copy(&lo->plh_stateid, new);
845 * Because of wraparound, we want to keep the barrier
846 * "close" to the current seqids.
848 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
851 new_barrier = be32_to_cpu(new->seqid);
852 else if (new_barrier == 0)
854 if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
855 lo->plh_barrier = new_barrier;
859 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
860 const nfs4_stateid *stateid)
862 u32 seqid = be32_to_cpu(stateid->seqid);
864 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
867 /* lget is set to 1 if called from inside send_layoutget call chain */
869 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
871 return lo->plh_block_lgets ||
872 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
876 * Get layout from server.
877 * for now, assume that whole file layouts are requested.
879 * arg->length: all ones
881 static struct pnfs_layout_segment *
882 send_layoutget(struct pnfs_layout_hdr *lo,
883 struct nfs_open_context *ctx,
884 nfs4_stateid *stateid,
885 const struct pnfs_layout_range *range,
886 long *timeout, gfp_t gfp_flags)
888 struct inode *ino = lo->plh_inode;
889 struct nfs_server *server = NFS_SERVER(ino);
890 struct nfs4_layoutget *lgp;
893 dprintk("--> %s\n", __func__);
896 * Synchronously retrieve layout information from server and
897 * store in lseg. If we race with a concurrent seqid morphing
898 * op, then re-send the LAYOUTGET.
900 lgp = kzalloc(sizeof(*lgp), gfp_flags);
902 return ERR_PTR(-ENOMEM);
904 i_size = i_size_read(ino);
906 lgp->args.minlength = PAGE_SIZE;
907 if (lgp->args.minlength > range->length)
908 lgp->args.minlength = range->length;
909 if (range->iomode == IOMODE_READ) {
910 if (range->offset >= i_size)
911 lgp->args.minlength = 0;
912 else if (i_size - range->offset < lgp->args.minlength)
913 lgp->args.minlength = i_size - range->offset;
915 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
916 pnfs_copy_range(&lgp->args.range, range);
917 lgp->args.type = server->pnfs_curr_ld->id;
918 lgp->args.inode = ino;
919 lgp->args.ctx = get_nfs_open_context(ctx);
920 nfs4_stateid_copy(&lgp->args.stateid, stateid);
921 lgp->gfp_flags = gfp_flags;
922 lgp->cred = lo->plh_lc_cred;
924 return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
927 static void pnfs_clear_layoutcommit(struct inode *inode,
928 struct list_head *head)
930 struct nfs_inode *nfsi = NFS_I(inode);
931 struct pnfs_layout_segment *lseg, *tmp;
933 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
935 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
936 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
938 pnfs_lseg_dec_and_remove_zero(lseg, head);
942 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
944 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
945 smp_mb__after_atomic();
946 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
947 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
951 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
952 nfs4_stateid *stateid,
953 enum pnfs_iomode *iomode)
955 /* Serialise LAYOUTGET/LAYOUTRETURN */
956 if (atomic_read(&lo->plh_outstanding) != 0)
958 if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
960 pnfs_get_layout_hdr(lo);
961 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
962 if (stateid != NULL) {
963 nfs4_stateid_copy(stateid, &lo->plh_stateid);
964 if (lo->plh_return_seq != 0)
965 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
968 *iomode = lo->plh_return_iomode;
969 pnfs_clear_layoutreturn_info(lo);
973 nfs4_stateid_copy(stateid, &lo->plh_stateid);
975 *iomode = IOMODE_ANY;
980 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
981 enum pnfs_iomode iomode, bool sync)
983 struct inode *ino = lo->plh_inode;
984 struct nfs4_layoutreturn *lrp;
987 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
988 if (unlikely(lrp == NULL)) {
990 spin_lock(&ino->i_lock);
991 pnfs_clear_layoutreturn_waitbit(lo);
992 spin_unlock(&ino->i_lock);
993 pnfs_put_layout_hdr(lo);
997 nfs4_stateid_copy(&lrp->args.stateid, stateid);
998 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
999 lrp->args.inode = ino;
1000 lrp->args.range.iomode = iomode;
1001 lrp->args.range.offset = 0;
1002 lrp->args.range.length = NFS4_MAX_UINT64;
1003 lrp->args.layout = lo;
1004 lrp->clp = NFS_SERVER(ino)->nfs_client;
1005 lrp->cred = lo->plh_lc_cred;
1007 status = nfs4_proc_layoutreturn(lrp, sync);
1009 dprintk("<-- %s status: %d\n", __func__, status);
1013 /* Return true if layoutreturn is needed */
1015 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1017 struct pnfs_layout_segment *s;
1019 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1022 /* Defer layoutreturn until all lsegs are done */
1023 list_for_each_entry(s, &lo->plh_segs, pls_list) {
1024 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1031 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1033 struct inode *inode= lo->plh_inode;
1035 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1037 spin_lock(&inode->i_lock);
1038 if (pnfs_layout_need_return(lo)) {
1039 nfs4_stateid stateid;
1040 enum pnfs_iomode iomode;
1043 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1044 spin_unlock(&inode->i_lock);
1046 /* Send an async layoutreturn so we dont deadlock */
1047 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1050 spin_unlock(&inode->i_lock);
1054 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1055 * when the layout segment list is empty.
1057 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1058 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1059 * deviceid is marked invalid.
1062 _pnfs_return_layout(struct inode *ino)
1064 struct pnfs_layout_hdr *lo = NULL;
1065 struct nfs_inode *nfsi = NFS_I(ino);
1066 LIST_HEAD(tmp_list);
1067 nfs4_stateid stateid;
1068 int status = 0, empty;
1071 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1073 spin_lock(&ino->i_lock);
1076 spin_unlock(&ino->i_lock);
1077 dprintk("NFS: %s no layout to return\n", __func__);
1080 /* Reference matched in nfs4_layoutreturn_release */
1081 pnfs_get_layout_hdr(lo);
1082 empty = list_empty(&lo->plh_segs);
1083 pnfs_clear_layoutcommit(ino, &tmp_list);
1084 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1086 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1087 struct pnfs_layout_range range = {
1088 .iomode = IOMODE_ANY,
1090 .length = NFS4_MAX_UINT64,
1092 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1095 /* Don't send a LAYOUTRETURN if list was initially empty */
1097 spin_unlock(&ino->i_lock);
1098 dprintk("NFS: %s no layout segments to return\n", __func__);
1099 goto out_put_layout_hdr;
1102 send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1103 spin_unlock(&ino->i_lock);
1104 pnfs_free_lseg_list(&tmp_list);
1106 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1108 pnfs_put_layout_hdr(lo);
1110 dprintk("<-- %s status: %d\n", __func__, status);
1113 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1116 pnfs_commit_and_return_layout(struct inode *inode)
1118 struct pnfs_layout_hdr *lo;
1121 spin_lock(&inode->i_lock);
1122 lo = NFS_I(inode)->layout;
1124 spin_unlock(&inode->i_lock);
1127 pnfs_get_layout_hdr(lo);
1128 /* Block new layoutgets and read/write to ds */
1129 lo->plh_block_lgets++;
1130 spin_unlock(&inode->i_lock);
1131 filemap_fdatawait(inode->i_mapping);
1132 ret = pnfs_layoutcommit_inode(inode, true);
1134 ret = _pnfs_return_layout(inode);
1135 spin_lock(&inode->i_lock);
1136 lo->plh_block_lgets--;
1137 spin_unlock(&inode->i_lock);
1138 pnfs_put_layout_hdr(lo);
1142 bool pnfs_roc(struct inode *ino)
1144 struct nfs_inode *nfsi = NFS_I(ino);
1145 struct nfs_open_context *ctx;
1146 struct nfs4_state *state;
1147 struct pnfs_layout_hdr *lo;
1148 struct pnfs_layout_segment *lseg, *tmp;
1149 nfs4_stateid stateid;
1150 LIST_HEAD(tmp_list);
1151 bool found = false, layoutreturn = false, roc = false;
1153 spin_lock(&ino->i_lock);
1155 if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1158 /* no roc if we hold a delegation */
1159 if (nfs4_check_delegation(ino, FMODE_READ))
1162 list_for_each_entry(ctx, &nfsi->open_files, list) {
1164 /* Don't return layout if there is open file state */
1165 if (state != NULL && state->state != 0)
1169 /* always send layoutreturn if being marked so */
1170 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1171 layoutreturn = pnfs_prepare_layoutreturn(lo,
1174 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1175 /* If we are sending layoutreturn, invalidate all valid lsegs */
1176 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1177 mark_lseg_invalid(lseg, &tmp_list);
1180 /* ROC in two conditions:
1181 * 1. there are ROC lsegs
1182 * 2. we don't send layoutreturn
1184 if (found && !layoutreturn) {
1185 /* lo ref dropped in pnfs_roc_release() */
1186 pnfs_get_layout_hdr(lo);
1191 spin_unlock(&ino->i_lock);
1192 pnfs_free_lseg_list(&tmp_list);
1193 pnfs_layoutcommit_inode(ino, true);
1195 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1199 void pnfs_roc_release(struct inode *ino)
1201 struct pnfs_layout_hdr *lo;
1203 spin_lock(&ino->i_lock);
1204 lo = NFS_I(ino)->layout;
1205 pnfs_clear_layoutreturn_waitbit(lo);
1206 if (atomic_dec_and_test(&lo->plh_refcount)) {
1207 pnfs_detach_layout_hdr(lo);
1208 spin_unlock(&ino->i_lock);
1209 pnfs_free_layout_hdr(lo);
1211 spin_unlock(&ino->i_lock);
1214 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1216 struct pnfs_layout_hdr *lo;
1218 spin_lock(&ino->i_lock);
1219 lo = NFS_I(ino)->layout;
1220 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1221 lo->plh_barrier = barrier;
1222 spin_unlock(&ino->i_lock);
1223 trace_nfs4_layoutreturn_on_close(ino, 0);
1226 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1228 struct nfs_inode *nfsi = NFS_I(ino);
1229 struct pnfs_layout_hdr *lo;
1232 spin_lock(&ino->i_lock);
1234 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1236 /* Since close does not return a layout stateid for use as
1237 * a barrier, we choose the worst-case barrier.
1239 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1240 spin_unlock(&ino->i_lock);
1243 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1245 struct nfs_inode *nfsi = NFS_I(ino);
1246 struct pnfs_layout_hdr *lo;
1249 /* we might not have grabbed lo reference. so need to check under
1251 spin_lock(&ino->i_lock);
1253 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1255 spin_unlock(&ino->i_lock);
1258 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1264 * Compare two layout segments for sorting into layout cache.
1265 * We want to preferentially return RW over RO layouts, so ensure those
1269 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1270 const struct pnfs_layout_range *l2)
1274 /* high offset > low offset */
1275 d = l1->offset - l2->offset;
1279 /* short length > long length */
1280 d = l2->length - l1->length;
1284 /* read > read/write */
1285 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1289 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1290 const struct pnfs_layout_range *l2)
1292 return pnfs_lseg_range_cmp(l1, l2) > 0;
1296 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1297 struct pnfs_layout_segment *old)
1303 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1304 struct pnfs_layout_segment *lseg,
1305 bool (*is_after)(const struct pnfs_layout_range *,
1306 const struct pnfs_layout_range *),
1307 bool (*do_merge)(struct pnfs_layout_segment *,
1308 struct pnfs_layout_segment *),
1309 struct list_head *free_me)
1311 struct pnfs_layout_segment *lp, *tmp;
1313 dprintk("%s:Begin\n", __func__);
1315 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1316 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1318 if (do_merge(lseg, lp)) {
1319 mark_lseg_invalid(lp, free_me);
1322 if (is_after(&lseg->pls_range, &lp->pls_range))
1324 list_add_tail(&lseg->pls_list, &lp->pls_list);
1325 dprintk("%s: inserted lseg %p "
1326 "iomode %d offset %llu length %llu before "
1327 "lp %p iomode %d offset %llu length %llu\n",
1328 __func__, lseg, lseg->pls_range.iomode,
1329 lseg->pls_range.offset, lseg->pls_range.length,
1330 lp, lp->pls_range.iomode, lp->pls_range.offset,
1331 lp->pls_range.length);
1334 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1335 dprintk("%s: inserted lseg %p "
1336 "iomode %d offset %llu length %llu at tail\n",
1337 __func__, lseg, lseg->pls_range.iomode,
1338 lseg->pls_range.offset, lseg->pls_range.length);
1340 pnfs_get_layout_hdr(lo);
1342 dprintk("%s:Return\n", __func__);
1344 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1347 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1348 struct pnfs_layout_segment *lseg,
1349 struct list_head *free_me)
1351 struct inode *inode = lo->plh_inode;
1352 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1354 if (ld->add_lseg != NULL)
1355 ld->add_lseg(lo, lseg, free_me);
1357 pnfs_generic_layout_insert_lseg(lo, lseg,
1358 pnfs_lseg_range_is_after,
1363 static struct pnfs_layout_hdr *
1364 alloc_init_layout_hdr(struct inode *ino,
1365 struct nfs_open_context *ctx,
1368 struct pnfs_layout_hdr *lo;
1370 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1373 atomic_set(&lo->plh_refcount, 1);
1374 INIT_LIST_HEAD(&lo->plh_layouts);
1375 INIT_LIST_HEAD(&lo->plh_segs);
1376 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1377 lo->plh_inode = ino;
1378 lo->plh_lc_cred = get_rpccred(ctx->cred);
1379 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1383 static struct pnfs_layout_hdr *
1384 pnfs_find_alloc_layout(struct inode *ino,
1385 struct nfs_open_context *ctx,
1387 __releases(&ino->i_lock)
1388 __acquires(&ino->i_lock)
1390 struct nfs_inode *nfsi = NFS_I(ino);
1391 struct pnfs_layout_hdr *new = NULL;
1393 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1395 if (nfsi->layout != NULL)
1397 spin_unlock(&ino->i_lock);
1398 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1399 spin_lock(&ino->i_lock);
1401 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1404 } else if (new != NULL)
1405 pnfs_free_layout_hdr(new);
1407 pnfs_get_layout_hdr(nfsi->layout);
1408 return nfsi->layout;
1412 * iomode matching rules:
1413 * iomode lseg strict match
1415 * ----- ----- ------ -----
1420 * READ READ N/A true
1421 * READ RW true false
1422 * READ RW false true
1425 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1426 const struct pnfs_layout_range *range,
1429 struct pnfs_layout_range range1;
1431 if ((range->iomode == IOMODE_RW &&
1432 ls_range->iomode != IOMODE_RW) ||
1433 (range->iomode != ls_range->iomode &&
1434 strict_iomode == true) ||
1435 !pnfs_lseg_range_intersecting(ls_range, range))
1438 /* range1 covers only the first byte in the range */
1441 return pnfs_lseg_range_contained(ls_range, &range1);
1445 * lookup range in layout
1447 static struct pnfs_layout_segment *
1448 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1449 struct pnfs_layout_range *range,
1452 struct pnfs_layout_segment *lseg, *ret = NULL;
1454 dprintk("%s:Begin\n", __func__);
1456 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1457 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1458 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1459 pnfs_lseg_range_match(&lseg->pls_range, range,
1461 ret = pnfs_get_lseg(lseg);
1466 dprintk("%s:Return lseg %p ref %d\n",
1467 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1472 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1473 * to the MDS or over pNFS
1475 * The nfs_inode read_io and write_io fields are cumulative counters reset
1476 * when there are no layout segments. Note that in pnfs_update_layout iomode
1477 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1480 * A return of true means use MDS I/O.
1483 * If a file's size is smaller than the file size threshold, data accesses
1484 * SHOULD be sent to the metadata server. If an I/O request has a length that
1485 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1486 * server. If both file size and I/O size are provided, the client SHOULD
1487 * reach or exceed both thresholds before sending its read or write
1488 * requests to the data server.
1490 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1491 struct inode *ino, int iomode)
1493 struct nfs4_threshold *t = ctx->mdsthreshold;
1494 struct nfs_inode *nfsi = NFS_I(ino);
1495 loff_t fsize = i_size_read(ino);
1496 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1501 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1502 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1506 if (t->bm & THRESHOLD_RD) {
1507 dprintk("%s fsize %llu\n", __func__, fsize);
1509 if (fsize < t->rd_sz)
1512 if (t->bm & THRESHOLD_RD_IO) {
1513 dprintk("%s nfsi->read_io %llu\n", __func__,
1516 if (nfsi->read_io < t->rd_io_sz)
1521 if (t->bm & THRESHOLD_WR) {
1522 dprintk("%s fsize %llu\n", __func__, fsize);
1524 if (fsize < t->wr_sz)
1527 if (t->bm & THRESHOLD_WR_IO) {
1528 dprintk("%s nfsi->write_io %llu\n", __func__,
1531 if (nfsi->write_io < t->wr_io_sz)
1536 if (size_set && io_set) {
1539 } else if (size || io)
1542 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1546 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1549 * send layoutcommit as it can hold up layoutreturn due to lseg
1552 pnfs_layoutcommit_inode(lo->plh_inode, false);
1553 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1554 nfs_wait_bit_killable,
1555 TASK_UNINTERRUPTIBLE);
1558 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1560 unsigned long *bitlock = &lo->plh_flags;
1562 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1563 smp_mb__after_atomic();
1564 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1568 * Layout segment is retreived from the server if not cached.
1569 * The appropriate layout segment is referenced and returned to the caller.
1571 struct pnfs_layout_segment *
1572 pnfs_update_layout(struct inode *ino,
1573 struct nfs_open_context *ctx,
1576 enum pnfs_iomode iomode,
1580 struct pnfs_layout_range arg = {
1585 unsigned pg_offset, seq;
1586 struct nfs_server *server = NFS_SERVER(ino);
1587 struct nfs_client *clp = server->nfs_client;
1588 struct pnfs_layout_hdr *lo = NULL;
1589 struct pnfs_layout_segment *lseg = NULL;
1590 nfs4_stateid stateid;
1592 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1595 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1596 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1597 PNFS_UPDATE_LAYOUT_NO_PNFS);
1601 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1602 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1603 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1607 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1608 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1609 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1614 nfs4_client_recover_expired_lease(clp);
1616 spin_lock(&ino->i_lock);
1617 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1619 spin_unlock(&ino->i_lock);
1620 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1621 PNFS_UPDATE_LAYOUT_NOMEM);
1625 /* Do we even need to bother with this? */
1626 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1627 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1628 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1629 dprintk("%s matches recall, use MDS\n", __func__);
1633 /* if LAYOUTGET already failed once we don't try again */
1634 if (pnfs_layout_io_test_failed(lo, iomode)) {
1635 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1636 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1640 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1642 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1643 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1647 if (!nfs4_valid_open_stateid(ctx->state)) {
1648 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1649 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1654 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1655 * stateid, or it has been invalidated, then we must use the open
1658 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1661 * The first layoutget for the file. Need to serialize per
1662 * RFC 5661 Errata 3208.
1664 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1666 spin_unlock(&ino->i_lock);
1667 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1668 TASK_UNINTERRUPTIBLE);
1669 pnfs_put_layout_hdr(lo);
1670 dprintk("%s retrying\n", __func__);
1676 seq = read_seqbegin(&ctx->state->seqlock);
1677 nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1678 } while (read_seqretry(&ctx->state->seqlock, seq));
1680 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1684 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1685 * for LAYOUTRETURN even if first is true.
1687 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1688 spin_unlock(&ino->i_lock);
1689 dprintk("%s wait for layoutreturn\n", __func__);
1690 if (pnfs_prepare_to_retry_layoutget(lo)) {
1692 pnfs_clear_first_layoutget(lo);
1693 pnfs_put_layout_hdr(lo);
1694 dprintk("%s retrying\n", __func__);
1695 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1696 lseg, PNFS_UPDATE_LAYOUT_RETRY);
1699 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1700 PNFS_UPDATE_LAYOUT_RETURN);
1701 goto out_put_layout_hdr;
1704 if (pnfs_layoutgets_blocked(lo)) {
1705 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1706 PNFS_UPDATE_LAYOUT_BLOCKED);
1709 atomic_inc(&lo->plh_outstanding);
1710 spin_unlock(&ino->i_lock);
1712 if (list_empty(&lo->plh_layouts)) {
1713 /* The lo must be on the clp list if there is any
1714 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1716 spin_lock(&clp->cl_lock);
1717 if (list_empty(&lo->plh_layouts))
1718 list_add_tail(&lo->plh_layouts, &server->layouts);
1719 spin_unlock(&clp->cl_lock);
1722 pg_offset = arg.offset & ~PAGE_MASK;
1724 arg.offset -= pg_offset;
1725 arg.length += pg_offset;
1727 if (arg.length != NFS4_MAX_UINT64)
1728 arg.length = PAGE_ALIGN(arg.length);
1730 lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1731 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1732 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1733 atomic_dec(&lo->plh_outstanding);
1735 switch(PTR_ERR(lseg)) {
1737 if (time_after(jiffies, giveup))
1740 case -ERECALLCONFLICT:
1741 /* Huh? We hold no layouts, how is there a recall? */
1746 /* Destroy the existing layout and start over */
1747 if (time_after(jiffies, giveup))
1748 pnfs_destroy_layout(NFS_I(ino));
1753 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1754 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1757 goto out_put_layout_hdr;
1761 pnfs_clear_first_layoutget(lo);
1762 trace_pnfs_update_layout(ino, pos, count,
1763 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1764 pnfs_put_layout_hdr(lo);
1768 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1773 pnfs_clear_first_layoutget(lo);
1774 pnfs_put_layout_hdr(lo);
1776 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1777 "(%s, offset: %llu, length: %llu)\n",
1778 __func__, ino->i_sb->s_id,
1779 (unsigned long long)NFS_FILEID(ino),
1780 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1781 iomode==IOMODE_RW ? "read/write" : "read-only",
1782 (unsigned long long)pos,
1783 (unsigned long long)count);
1786 spin_unlock(&ino->i_lock);
1787 goto out_put_layout_hdr;
1789 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1792 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1794 switch (range->iomode) {
1801 if (range->offset == NFS4_MAX_UINT64)
1803 if (range->length == 0)
1805 if (range->length != NFS4_MAX_UINT64 &&
1806 range->length > NFS4_MAX_UINT64 - range->offset)
1811 struct pnfs_layout_segment *
1812 pnfs_layout_process(struct nfs4_layoutget *lgp)
1814 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1815 struct nfs4_layoutget_res *res = &lgp->res;
1816 struct pnfs_layout_segment *lseg;
1817 struct inode *ino = lo->plh_inode;
1820 if (!pnfs_sanity_check_layout_range(&res->range))
1821 return ERR_PTR(-EINVAL);
1823 /* Inject layout blob into I/O device driver */
1824 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1825 if (IS_ERR_OR_NULL(lseg)) {
1827 lseg = ERR_PTR(-ENOMEM);
1829 dprintk("%s: Could not allocate layout: error %ld\n",
1830 __func__, PTR_ERR(lseg));
1834 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1836 spin_lock(&ino->i_lock);
1837 if (pnfs_layoutgets_blocked(lo)) {
1838 dprintk("%s forget reply due to state\n", __func__);
1842 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1843 /* existing state ID, make sure the sequence number matches. */
1844 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1845 dprintk("%s forget reply due to sequence\n", __func__);
1848 pnfs_set_layout_stateid(lo, &res->stateid, false);
1851 * We got an entirely new state ID. Mark all segments for the
1852 * inode invalid, and don't bother validating the stateid
1855 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1857 pnfs_set_layout_stateid(lo, &res->stateid, true);
1860 pnfs_get_lseg(lseg);
1861 pnfs_layout_insert_lseg(lo, lseg, &free_me);
1864 if (res->return_on_close)
1865 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1867 spin_unlock(&ino->i_lock);
1868 pnfs_free_lseg_list(&free_me);
1872 spin_unlock(&ino->i_lock);
1873 lseg->pls_layout = lo;
1874 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1875 return ERR_PTR(-EAGAIN);
1879 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
1882 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
1883 iomode = IOMODE_ANY;
1884 lo->plh_return_iomode = iomode;
1885 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
1887 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
1888 lo->plh_return_seq = seq;
1893 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1894 * @lo: pointer to layout header
1895 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1896 * @return_range: describe layout segment ranges to be returned
1898 * This function is mainly intended for use by layoutrecall. It attempts
1899 * to free the layout segment immediately, or else to mark it for return
1900 * as soon as its reference count drops to zero.
1903 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1904 struct list_head *tmp_list,
1905 const struct pnfs_layout_range *return_range,
1908 struct pnfs_layout_segment *lseg, *next;
1911 dprintk("%s:Begin lo %p\n", __func__, lo);
1913 if (list_empty(&lo->plh_segs))
1916 assert_spin_locked(&lo->plh_inode->i_lock);
1918 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1919 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
1920 dprintk("%s: marking lseg %p iomode %d "
1921 "offset %llu length %llu\n", __func__,
1922 lseg, lseg->pls_range.iomode,
1923 lseg->pls_range.offset,
1924 lseg->pls_range.length);
1925 if (mark_lseg_invalid(lseg, tmp_list))
1928 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1932 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
1937 void pnfs_error_mark_layout_for_return(struct inode *inode,
1938 struct pnfs_layout_segment *lseg)
1940 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1941 struct pnfs_layout_range range = {
1942 .iomode = lseg->pls_range.iomode,
1944 .length = NFS4_MAX_UINT64,
1947 bool return_now = false;
1949 spin_lock(&inode->i_lock);
1950 pnfs_set_plh_return_info(lo, range.iomode, 0);
1952 * mark all matching lsegs so that we are sure to have no live
1953 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1956 if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0)) {
1957 nfs4_stateid stateid;
1958 enum pnfs_iomode iomode;
1960 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1961 spin_unlock(&inode->i_lock);
1963 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1965 spin_unlock(&inode->i_lock);
1966 nfs_commit_inode(inode, 0);
1968 pnfs_free_lseg_list(&free_me);
1970 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1973 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1975 u64 rd_size = req->wb_bytes;
1977 if (pgio->pg_lseg == NULL) {
1978 if (pgio->pg_dreq == NULL)
1979 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1981 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1983 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1990 if (IS_ERR(pgio->pg_lseg)) {
1991 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1992 pgio->pg_lseg = NULL;
1996 /* If no lseg, fall back to read through mds */
1997 if (pgio->pg_lseg == NULL)
1998 nfs_pageio_reset_read_mds(pgio);
2001 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2004 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2005 struct nfs_page *req, u64 wb_size)
2007 if (pgio->pg_lseg == NULL) {
2008 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2015 if (IS_ERR(pgio->pg_lseg)) {
2016 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2017 pgio->pg_lseg = NULL;
2021 /* If no lseg, fall back to write through mds */
2022 if (pgio->pg_lseg == NULL)
2023 nfs_pageio_reset_write_mds(pgio);
2025 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2028 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2030 if (desc->pg_lseg) {
2031 pnfs_put_lseg(desc->pg_lseg);
2032 desc->pg_lseg = NULL;
2035 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2038 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2039 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2042 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2043 struct nfs_page *prev, struct nfs_page *req)
2046 u64 seg_end, req_start, seg_left;
2048 size = nfs_generic_pg_test(pgio, prev, req);
2053 * 'size' contains the number of bytes left in the current page (up
2054 * to the original size asked for in @req->wb_bytes).
2056 * Calculate how many bytes are left in the layout segment
2057 * and if there are less bytes than 'size', return that instead.
2059 * Please also note that 'end_offset' is actually the offset of the
2060 * first byte that lies outside the pnfs_layout_range. FIXME?
2063 if (pgio->pg_lseg) {
2064 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
2065 pgio->pg_lseg->pls_range.length);
2066 req_start = req_offset(req);
2067 WARN_ON_ONCE(req_start >= seg_end);
2068 /* start of request is past the last byte of this segment */
2069 if (req_start >= seg_end) {
2070 /* reference the new lseg */
2071 if (pgio->pg_ops->pg_cleanup)
2072 pgio->pg_ops->pg_cleanup(pgio);
2073 if (pgio->pg_ops->pg_init)
2074 pgio->pg_ops->pg_init(pgio, req);
2078 /* adjust 'size' iff there are fewer bytes left in the
2079 * segment than what nfs_generic_pg_test returned */
2080 seg_left = seg_end - req_start;
2081 if (seg_left < size)
2082 size = (unsigned int)seg_left;
2087 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2089 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2091 struct nfs_pageio_descriptor pgio;
2093 /* Resend all requests through the MDS */
2094 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2095 hdr->completion_ops);
2096 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2097 return nfs_pageio_resend(&pgio, hdr);
2099 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2101 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2104 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2105 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2106 PNFS_LAYOUTRET_ON_ERROR) {
2107 pnfs_return_layout(hdr->inode);
2109 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2110 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2114 * Called by non rpc-based layout drivers
2116 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2118 if (likely(!hdr->pnfs_error)) {
2119 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2120 hdr->mds_offset + hdr->res.count);
2121 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2123 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2124 if (unlikely(hdr->pnfs_error))
2125 pnfs_ld_handle_write_error(hdr);
2126 hdr->mds_ops->rpc_release(hdr);
2128 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2131 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2132 struct nfs_pgio_header *hdr)
2134 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2136 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2137 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2138 nfs_pageio_reset_write_mds(desc);
2139 mirror->pg_recoalesce = 1;
2141 nfs_pgio_data_destroy(hdr);
2145 static enum pnfs_try_status
2146 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2147 const struct rpc_call_ops *call_ops,
2148 struct pnfs_layout_segment *lseg,
2151 struct inode *inode = hdr->inode;
2152 enum pnfs_try_status trypnfs;
2153 struct nfs_server *nfss = NFS_SERVER(inode);
2155 hdr->mds_ops = call_ops;
2157 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2158 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2159 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2160 if (trypnfs != PNFS_NOT_ATTEMPTED)
2161 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2162 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2167 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2168 struct nfs_pgio_header *hdr, int how)
2170 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2171 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2172 enum pnfs_try_status trypnfs;
2174 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2175 if (trypnfs == PNFS_NOT_ATTEMPTED)
2176 pnfs_write_through_mds(desc, hdr);
2179 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2181 pnfs_put_lseg(hdr->lseg);
2182 nfs_pgio_header_free(hdr);
2186 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2188 struct nfs_pgio_header *hdr;
2191 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2193 desc->pg_error = -ENOMEM;
2194 return desc->pg_error;
2196 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2198 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2199 ret = nfs_generic_pgio(desc, hdr);
2201 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2205 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2207 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2209 struct nfs_pageio_descriptor pgio;
2211 /* Resend all requests through the MDS */
2212 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2213 return nfs_pageio_resend(&pgio, hdr);
2215 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2217 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2219 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2220 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2221 PNFS_LAYOUTRET_ON_ERROR) {
2222 pnfs_return_layout(hdr->inode);
2224 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2225 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2229 * Called by non rpc-based layout drivers
2231 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2233 if (likely(!hdr->pnfs_error))
2234 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2235 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2236 if (unlikely(hdr->pnfs_error))
2237 pnfs_ld_handle_read_error(hdr);
2238 hdr->mds_ops->rpc_release(hdr);
2240 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2243 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2244 struct nfs_pgio_header *hdr)
2246 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2248 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2249 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2250 nfs_pageio_reset_read_mds(desc);
2251 mirror->pg_recoalesce = 1;
2253 nfs_pgio_data_destroy(hdr);
2258 * Call the appropriate parallel I/O subsystem read function.
2260 static enum pnfs_try_status
2261 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2262 const struct rpc_call_ops *call_ops,
2263 struct pnfs_layout_segment *lseg)
2265 struct inode *inode = hdr->inode;
2266 struct nfs_server *nfss = NFS_SERVER(inode);
2267 enum pnfs_try_status trypnfs;
2269 hdr->mds_ops = call_ops;
2271 dprintk("%s: Reading ino:%lu %u@%llu\n",
2272 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2274 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2275 if (trypnfs != PNFS_NOT_ATTEMPTED)
2276 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2277 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2281 /* Resend all requests through pnfs. */
2282 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2284 struct nfs_pageio_descriptor pgio;
2286 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2287 nfs_pageio_init_read(&pgio, hdr->inode, false,
2288 hdr->completion_ops);
2289 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2292 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2295 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2297 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2298 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2299 enum pnfs_try_status trypnfs;
2301 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2302 if (trypnfs == PNFS_TRY_AGAIN)
2303 pnfs_read_resend_pnfs(hdr);
2304 if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status)
2305 pnfs_read_through_mds(desc, hdr);
2308 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2310 pnfs_put_lseg(hdr->lseg);
2311 nfs_pgio_header_free(hdr);
2315 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2317 struct nfs_pgio_header *hdr;
2320 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2322 desc->pg_error = -ENOMEM;
2323 return desc->pg_error;
2325 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2326 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2327 ret = nfs_generic_pgio(desc, hdr);
2329 pnfs_do_read(desc, hdr);
2332 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2334 static void pnfs_clear_layoutcommitting(struct inode *inode)
2336 unsigned long *bitlock = &NFS_I(inode)->flags;
2338 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2339 smp_mb__after_atomic();
2340 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2344 * There can be multiple RW segments.
2346 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2348 struct pnfs_layout_segment *lseg;
2350 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2351 if (lseg->pls_range.iomode == IOMODE_RW &&
2352 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2353 list_add(&lseg->pls_lc_list, listp);
2357 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2359 struct pnfs_layout_segment *lseg, *tmp;
2361 /* Matched by references in pnfs_set_layoutcommit */
2362 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2363 list_del_init(&lseg->pls_lc_list);
2364 pnfs_put_lseg(lseg);
2367 pnfs_clear_layoutcommitting(inode);
2370 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2372 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2374 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2377 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2380 struct nfs_inode *nfsi = NFS_I(inode);
2381 bool mark_as_dirty = false;
2383 spin_lock(&inode->i_lock);
2384 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2385 nfsi->layout->plh_lwb = end_pos;
2386 mark_as_dirty = true;
2387 dprintk("%s: Set layoutcommit for inode %lu ",
2388 __func__, inode->i_ino);
2389 } else if (end_pos > nfsi->layout->plh_lwb)
2390 nfsi->layout->plh_lwb = end_pos;
2391 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2392 /* references matched in nfs4_layoutcommit_release */
2393 pnfs_get_lseg(lseg);
2395 spin_unlock(&inode->i_lock);
2396 dprintk("%s: lseg %p end_pos %llu\n",
2397 __func__, lseg, nfsi->layout->plh_lwb);
2399 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2400 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2402 mark_inode_dirty_sync(inode);
2404 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2406 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2408 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2410 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2411 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2412 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2416 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2417 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2418 * data to disk to allow the server to recover the data if it crashes.
2419 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2420 * is off, and a COMMIT is sent to a data server, or
2421 * if WRITEs to a data server return NFS_DATA_SYNC.
2424 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2426 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2427 struct nfs4_layoutcommit_data *data;
2428 struct nfs_inode *nfsi = NFS_I(inode);
2432 if (!pnfs_layoutcommit_outstanding(inode))
2435 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2438 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2441 status = wait_on_bit_lock_action(&nfsi->flags,
2442 NFS_INO_LAYOUTCOMMITTING,
2443 nfs_wait_bit_killable,
2450 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2451 data = kzalloc(sizeof(*data), GFP_NOFS);
2453 goto clear_layoutcommitting;
2456 spin_lock(&inode->i_lock);
2457 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2460 INIT_LIST_HEAD(&data->lseg_list);
2461 pnfs_list_write_lseg(inode, &data->lseg_list);
2463 end_pos = nfsi->layout->plh_lwb;
2465 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2466 spin_unlock(&inode->i_lock);
2468 data->args.inode = inode;
2469 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2470 nfs_fattr_init(&data->fattr);
2471 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2472 data->res.fattr = &data->fattr;
2474 data->args.lastbytewritten = end_pos - 1;
2476 data->args.lastbytewritten = U64_MAX;
2477 data->res.server = NFS_SERVER(inode);
2479 if (ld->prepare_layoutcommit) {
2480 status = ld->prepare_layoutcommit(&data->args);
2482 put_rpccred(data->cred);
2483 spin_lock(&inode->i_lock);
2484 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2485 if (end_pos > nfsi->layout->plh_lwb)
2486 nfsi->layout->plh_lwb = end_pos;
2492 status = nfs4_proc_layoutcommit(data, sync);
2495 mark_inode_dirty_sync(inode);
2496 dprintk("<-- %s status %d\n", __func__, status);
2499 spin_unlock(&inode->i_lock);
2501 clear_layoutcommitting:
2502 pnfs_clear_layoutcommitting(inode);
2505 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2508 pnfs_generic_sync(struct inode *inode, bool datasync)
2510 return pnfs_layoutcommit_inode(inode, true);
2512 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2514 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2516 struct nfs4_threshold *thp;
2518 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2520 dprintk("%s mdsthreshold allocation failed\n", __func__);
2526 #if IS_ENABLED(CONFIG_NFS_V4_2)
2528 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2530 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2531 struct nfs_server *server = NFS_SERVER(inode);
2532 struct nfs_inode *nfsi = NFS_I(inode);
2533 struct nfs42_layoutstat_data *data;
2534 struct pnfs_layout_hdr *hdr;
2537 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2540 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2543 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2546 spin_lock(&inode->i_lock);
2547 if (!NFS_I(inode)->layout) {
2548 spin_unlock(&inode->i_lock);
2549 goto out_clear_layoutstats;
2551 hdr = NFS_I(inode)->layout;
2552 pnfs_get_layout_hdr(hdr);
2553 spin_unlock(&inode->i_lock);
2555 data = kzalloc(sizeof(*data), gfp_flags);
2561 data->args.fh = NFS_FH(inode);
2562 data->args.inode = inode;
2563 status = ld->prepare_layoutstats(&data->args);
2567 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2570 dprintk("%s returns %d\n", __func__, status);
2576 pnfs_put_layout_hdr(hdr);
2577 out_clear_layoutstats:
2578 smp_mb__before_atomic();
2579 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2580 smp_mb__after_atomic();
2583 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2586 unsigned int layoutstats_timer;
2587 module_param(layoutstats_timer, uint, 0644);
2588 EXPORT_SYMBOL_GPL(layoutstats_timer);