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>
36 #include "nfs4trace.h"
38 #define NFSDBG_FACILITY NFSDBG_PNFS
39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
44 * protects pnfs_modules_tbl.
46 static DEFINE_SPINLOCK(pnfs_spinlock);
49 * pnfs_modules_tbl holds all pnfs modules
51 static LIST_HEAD(pnfs_modules_tbl);
53 /* Return the registered pnfs layout driver module matching given id */
54 static struct pnfs_layoutdriver_type *
55 find_pnfs_driver_locked(u32 id)
57 struct pnfs_layoutdriver_type *local;
59 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
64 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
68 static struct pnfs_layoutdriver_type *
69 find_pnfs_driver(u32 id)
71 struct pnfs_layoutdriver_type *local;
73 spin_lock(&pnfs_spinlock);
74 local = find_pnfs_driver_locked(id);
75 if (local != NULL && !try_module_get(local->owner)) {
76 dprintk("%s: Could not grab reference on module\n", __func__);
79 spin_unlock(&pnfs_spinlock);
84 unset_pnfs_layoutdriver(struct nfs_server *nfss)
86 if (nfss->pnfs_curr_ld) {
87 if (nfss->pnfs_curr_ld->clear_layoutdriver)
88 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
89 /* Decrement the MDS count. Purge the deviceid cache if zero */
90 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
91 nfs4_deviceid_purge_client(nfss->nfs_client);
92 module_put(nfss->pnfs_curr_ld->owner);
94 nfss->pnfs_curr_ld = NULL;
98 * Try to set the server's pnfs module to the pnfs layout type specified by id.
99 * Currently only one pNFS layout driver per filesystem is supported.
101 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
104 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
107 struct pnfs_layoutdriver_type *ld_type = NULL;
111 if (!(server->nfs_client->cl_exchange_flags &
112 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
113 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
114 __func__, id, server->nfs_client->cl_exchange_flags);
117 ld_type = find_pnfs_driver(id);
119 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
120 ld_type = find_pnfs_driver(id);
122 dprintk("%s: No pNFS module found for %u.\n",
127 server->pnfs_curr_ld = ld_type;
128 if (ld_type->set_layoutdriver
129 && ld_type->set_layoutdriver(server, mntfh)) {
130 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
131 "driver %u.\n", __func__, id);
132 module_put(ld_type->owner);
135 /* Bump the MDS count */
136 atomic_inc(&server->nfs_client->cl_mds_count);
138 dprintk("%s: pNFS module for %u set\n", __func__, id);
142 dprintk("%s: Using NFSv4 I/O\n", __func__);
143 server->pnfs_curr_ld = NULL;
147 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
149 int status = -EINVAL;
150 struct pnfs_layoutdriver_type *tmp;
152 if (ld_type->id == 0) {
153 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
156 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
157 printk(KERN_ERR "NFS: %s Layout driver must provide "
158 "alloc_lseg and free_lseg.\n", __func__);
162 spin_lock(&pnfs_spinlock);
163 tmp = find_pnfs_driver_locked(ld_type->id);
165 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
167 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
170 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
171 __func__, ld_type->id);
173 spin_unlock(&pnfs_spinlock);
177 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
180 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
182 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
183 spin_lock(&pnfs_spinlock);
184 list_del(&ld_type->pnfs_tblid);
185 spin_unlock(&pnfs_spinlock);
187 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
190 * pNFS client layout cache
193 /* Need to hold i_lock if caller does not already hold reference */
195 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
197 atomic_inc(&lo->plh_refcount);
200 static struct pnfs_layout_hdr *
201 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
203 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
204 return ld->alloc_layout_hdr(ino, gfp_flags);
208 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
210 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
211 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
213 if (!list_empty(&lo->plh_layouts)) {
214 struct nfs_client *clp = server->nfs_client;
216 spin_lock(&clp->cl_lock);
217 list_del_init(&lo->plh_layouts);
218 spin_unlock(&clp->cl_lock);
220 put_rpccred(lo->plh_lc_cred);
221 return ld->free_layout_hdr(lo);
225 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
227 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
228 dprintk("%s: freeing layout cache %p\n", __func__, lo);
230 /* Reset MDS Threshold I/O counters */
236 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
238 struct inode *inode = lo->plh_inode;
240 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
241 pnfs_detach_layout_hdr(lo);
242 spin_unlock(&inode->i_lock);
243 pnfs_free_layout_hdr(lo);
248 pnfs_iomode_to_fail_bit(u32 iomode)
250 return iomode == IOMODE_RW ?
251 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
255 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
257 lo->plh_retry_timestamp = jiffies;
258 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
259 atomic_inc(&lo->plh_refcount);
263 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
265 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
266 atomic_dec(&lo->plh_refcount);
270 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
272 struct inode *inode = lo->plh_inode;
273 struct pnfs_layout_range range = {
276 .length = NFS4_MAX_UINT64,
280 spin_lock(&inode->i_lock);
281 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
282 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
283 spin_unlock(&inode->i_lock);
284 pnfs_free_lseg_list(&head);
285 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
286 iomode == IOMODE_RW ? "RW" : "READ");
290 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
292 unsigned long start, end;
293 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
295 if (test_bit(fail_bit, &lo->plh_flags) == 0)
298 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
299 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
300 /* It is time to retry the failed layoutgets */
301 pnfs_layout_clear_fail_bit(lo, fail_bit);
308 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
310 INIT_LIST_HEAD(&lseg->pls_list);
311 INIT_LIST_HEAD(&lseg->pls_lc_list);
312 atomic_set(&lseg->pls_refcount, 1);
314 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
315 lseg->pls_layout = lo;
318 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
320 struct inode *ino = lseg->pls_layout->plh_inode;
322 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
326 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
327 struct pnfs_layout_segment *lseg)
329 struct inode *inode = lo->plh_inode;
331 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
332 list_del_init(&lseg->pls_list);
333 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
334 atomic_dec(&lo->plh_refcount);
335 if (list_empty(&lo->plh_segs))
336 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
337 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
341 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
343 struct pnfs_layout_hdr *lo;
349 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
350 atomic_read(&lseg->pls_refcount),
351 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
352 lo = lseg->pls_layout;
353 inode = lo->plh_inode;
354 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
355 pnfs_get_layout_hdr(lo);
356 pnfs_layout_remove_lseg(lo, lseg);
357 spin_unlock(&inode->i_lock);
358 pnfs_free_lseg(lseg);
359 pnfs_put_layout_hdr(lo);
362 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
364 static void pnfs_put_lseg_async_work(struct work_struct *work)
366 struct pnfs_layout_segment *lseg;
368 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
374 pnfs_put_lseg_async(struct pnfs_layout_segment *lseg)
376 INIT_WORK(&lseg->pls_work, pnfs_put_lseg_async_work);
377 schedule_work(&lseg->pls_work);
379 EXPORT_SYMBOL_GPL(pnfs_put_lseg_async);
382 end_offset(u64 start, u64 len)
387 return end >= start ? end : NFS4_MAX_UINT64;
391 * is l2 fully contained in l1?
393 * [----------------------------------)
398 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
399 const struct pnfs_layout_range *l2)
401 u64 start1 = l1->offset;
402 u64 end1 = end_offset(start1, l1->length);
403 u64 start2 = l2->offset;
404 u64 end2 = end_offset(start2, l2->length);
406 return (start1 <= start2) && (end1 >= end2);
410 * is l1 and l2 intersecting?
412 * [----------------------------------)
417 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
418 const struct pnfs_layout_range *l2)
420 u64 start1 = l1->offset;
421 u64 end1 = end_offset(start1, l1->length);
422 u64 start2 = l2->offset;
423 u64 end2 = end_offset(start2, l2->length);
425 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
426 (end2 == NFS4_MAX_UINT64 || end2 > start1);
430 should_free_lseg(const struct pnfs_layout_range *lseg_range,
431 const struct pnfs_layout_range *recall_range)
433 return (recall_range->iomode == IOMODE_ANY ||
434 lseg_range->iomode == recall_range->iomode) &&
435 pnfs_lseg_range_intersecting(lseg_range, recall_range);
438 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
439 struct list_head *tmp_list)
441 if (!atomic_dec_and_test(&lseg->pls_refcount))
443 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
444 list_add(&lseg->pls_list, tmp_list);
448 /* Returns 1 if lseg is removed from list, 0 otherwise */
449 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
450 struct list_head *tmp_list)
454 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
455 /* Remove the reference keeping the lseg in the
456 * list. It will now be removed when all
457 * outstanding io is finished.
459 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
460 atomic_read(&lseg->pls_refcount));
461 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
467 /* Returns count of number of matching invalid lsegs remaining in list
471 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
472 struct list_head *tmp_list,
473 struct pnfs_layout_range *recall_range)
475 struct pnfs_layout_segment *lseg, *next;
476 int invalid = 0, removed = 0;
478 dprintk("%s:Begin lo %p\n", __func__, lo);
480 if (list_empty(&lo->plh_segs))
482 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
484 should_free_lseg(&lseg->pls_range, recall_range)) {
485 dprintk("%s: freeing lseg %p iomode %d "
486 "offset %llu length %llu\n", __func__,
487 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
488 lseg->pls_range.length);
490 removed += mark_lseg_invalid(lseg, tmp_list);
492 dprintk("%s:Return %i\n", __func__, invalid - removed);
493 return invalid - removed;
496 /* note free_me must contain lsegs from a single layout_hdr */
498 pnfs_free_lseg_list(struct list_head *free_me)
500 struct pnfs_layout_segment *lseg, *tmp;
502 if (list_empty(free_me))
505 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
506 list_del(&lseg->pls_list);
507 pnfs_free_lseg(lseg);
512 pnfs_destroy_layout(struct nfs_inode *nfsi)
514 struct pnfs_layout_hdr *lo;
517 spin_lock(&nfsi->vfs_inode.i_lock);
520 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
521 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
522 pnfs_get_layout_hdr(lo);
523 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
524 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
525 spin_unlock(&nfsi->vfs_inode.i_lock);
526 pnfs_free_lseg_list(&tmp_list);
527 pnfs_put_layout_hdr(lo);
529 spin_unlock(&nfsi->vfs_inode.i_lock);
531 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
534 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
535 struct list_head *layout_list)
537 struct pnfs_layout_hdr *lo;
540 spin_lock(&inode->i_lock);
541 lo = NFS_I(inode)->layout;
542 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
543 pnfs_get_layout_hdr(lo);
544 list_add(&lo->plh_bulk_destroy, layout_list);
547 spin_unlock(&inode->i_lock);
551 /* Caller must hold rcu_read_lock and clp->cl_lock */
553 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
554 struct nfs_server *server,
555 struct list_head *layout_list)
557 struct pnfs_layout_hdr *lo, *next;
560 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
561 inode = igrab(lo->plh_inode);
564 list_del_init(&lo->plh_layouts);
565 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
568 spin_unlock(&clp->cl_lock);
570 spin_lock(&clp->cl_lock);
578 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
581 struct pnfs_layout_hdr *lo;
583 struct pnfs_layout_range range = {
584 .iomode = IOMODE_ANY,
586 .length = NFS4_MAX_UINT64,
588 LIST_HEAD(lseg_list);
591 while (!list_empty(layout_list)) {
592 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
594 dprintk("%s freeing layout for inode %lu\n", __func__,
595 lo->plh_inode->i_ino);
596 inode = lo->plh_inode;
597 spin_lock(&inode->i_lock);
598 list_del_init(&lo->plh_bulk_destroy);
599 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
601 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
602 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
604 spin_unlock(&inode->i_lock);
605 pnfs_free_lseg_list(&lseg_list);
606 pnfs_put_layout_hdr(lo);
613 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
614 struct nfs_fsid *fsid,
617 struct nfs_server *server;
618 LIST_HEAD(layout_list);
620 spin_lock(&clp->cl_lock);
623 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
624 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
626 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
632 spin_unlock(&clp->cl_lock);
634 if (list_empty(&layout_list))
636 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
640 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
643 struct nfs_server *server;
644 LIST_HEAD(layout_list);
646 spin_lock(&clp->cl_lock);
649 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
650 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
656 spin_unlock(&clp->cl_lock);
658 if (list_empty(&layout_list))
660 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
664 * Called by the state manger to remove all layouts established under an
668 pnfs_destroy_all_layouts(struct nfs_client *clp)
670 nfs4_deviceid_mark_client_invalid(clp);
671 nfs4_deviceid_purge_client(clp);
673 pnfs_destroy_layouts_byclid(clp, false);
677 * Compare 2 layout stateid sequence ids, to see which is newer,
678 * taking into account wraparound issues.
680 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
682 return (s32)(s1 - s2) > 0;
685 /* update lo->plh_stateid with new if is more recent */
687 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
690 u32 oldseq, newseq, new_barrier;
691 int empty = list_empty(&lo->plh_segs);
693 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
694 newseq = be32_to_cpu(new->seqid);
695 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
696 nfs4_stateid_copy(&lo->plh_stateid, new);
697 if (update_barrier) {
698 new_barrier = be32_to_cpu(new->seqid);
700 /* Because of wraparound, we want to keep the barrier
701 * "close" to the current seqids.
703 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
705 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
706 lo->plh_barrier = new_barrier;
711 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
712 const nfs4_stateid *stateid)
714 u32 seqid = be32_to_cpu(stateid->seqid);
716 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
719 /* lget is set to 1 if called from inside send_layoutget call chain */
721 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
723 return lo->plh_block_lgets ||
724 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
725 (list_empty(&lo->plh_segs) &&
726 (atomic_read(&lo->plh_outstanding) > lget));
730 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
731 struct nfs4_state *open_state)
735 dprintk("--> %s\n", __func__);
736 spin_lock(&lo->plh_inode->i_lock);
737 if (pnfs_layoutgets_blocked(lo, 1)) {
739 } else if (!nfs4_valid_open_stateid(open_state)) {
741 } else if (list_empty(&lo->plh_segs) ||
742 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
746 seq = read_seqbegin(&open_state->seqlock);
747 nfs4_stateid_copy(dst, &open_state->stateid);
748 } while (read_seqretry(&open_state->seqlock, seq));
750 nfs4_stateid_copy(dst, &lo->plh_stateid);
751 spin_unlock(&lo->plh_inode->i_lock);
752 dprintk("<-- %s\n", __func__);
757 * Get layout from server.
758 * for now, assume that whole file layouts are requested.
760 * arg->length: all ones
762 static struct pnfs_layout_segment *
763 send_layoutget(struct pnfs_layout_hdr *lo,
764 struct nfs_open_context *ctx,
765 struct pnfs_layout_range *range,
768 struct inode *ino = lo->plh_inode;
769 struct nfs_server *server = NFS_SERVER(ino);
770 struct nfs4_layoutget *lgp;
771 struct pnfs_layout_segment *lseg;
773 dprintk("--> %s\n", __func__);
775 lgp = kzalloc(sizeof(*lgp), gfp_flags);
779 lgp->args.minlength = PAGE_CACHE_SIZE;
780 if (lgp->args.minlength > range->length)
781 lgp->args.minlength = range->length;
782 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
783 lgp->args.range = *range;
784 lgp->args.type = server->pnfs_curr_ld->id;
785 lgp->args.inode = ino;
786 lgp->args.ctx = get_nfs_open_context(ctx);
787 lgp->gfp_flags = gfp_flags;
788 lgp->cred = lo->plh_lc_cred;
790 /* Synchronously retrieve layout information from server and
793 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
795 switch (PTR_ERR(lseg)) {
800 /* remember that LAYOUTGET failed and suspend trying */
801 pnfs_layout_io_set_failed(lo, range->iomode);
809 static void pnfs_clear_layoutcommit(struct inode *inode,
810 struct list_head *head)
812 struct nfs_inode *nfsi = NFS_I(inode);
813 struct pnfs_layout_segment *lseg, *tmp;
815 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
817 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
818 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
820 pnfs_lseg_dec_and_remove_zero(lseg, head);
825 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
826 * when the layout segment list is empty.
828 * Note that a pnfs_layout_hdr can exist with an empty layout segment
829 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
830 * deviceid is marked invalid.
833 _pnfs_return_layout(struct inode *ino)
835 struct pnfs_layout_hdr *lo = NULL;
836 struct nfs_inode *nfsi = NFS_I(ino);
838 struct nfs4_layoutreturn *lrp;
839 nfs4_stateid stateid;
840 int status = 0, empty;
842 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
844 spin_lock(&ino->i_lock);
847 spin_unlock(&ino->i_lock);
848 dprintk("NFS: %s no layout to return\n", __func__);
851 stateid = nfsi->layout->plh_stateid;
852 /* Reference matched in nfs4_layoutreturn_release */
853 pnfs_get_layout_hdr(lo);
854 empty = list_empty(&lo->plh_segs);
855 pnfs_clear_layoutcommit(ino, &tmp_list);
856 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
857 /* Don't send a LAYOUTRETURN if list was initially empty */
859 spin_unlock(&ino->i_lock);
860 pnfs_put_layout_hdr(lo);
861 dprintk("NFS: %s no layout segments to return\n", __func__);
865 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
866 lo->plh_block_lgets++;
867 spin_unlock(&ino->i_lock);
868 pnfs_free_lseg_list(&tmp_list);
870 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
871 if (unlikely(lrp == NULL)) {
873 spin_lock(&ino->i_lock);
874 lo->plh_block_lgets--;
875 spin_unlock(&ino->i_lock);
876 pnfs_put_layout_hdr(lo);
880 lrp->args.stateid = stateid;
881 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
882 lrp->args.inode = ino;
883 lrp->args.layout = lo;
884 lrp->clp = NFS_SERVER(ino)->nfs_client;
885 lrp->cred = lo->plh_lc_cred;
887 status = nfs4_proc_layoutreturn(lrp);
889 dprintk("<-- %s status: %d\n", __func__, status);
892 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
895 pnfs_commit_and_return_layout(struct inode *inode)
897 struct pnfs_layout_hdr *lo;
900 spin_lock(&inode->i_lock);
901 lo = NFS_I(inode)->layout;
903 spin_unlock(&inode->i_lock);
906 pnfs_get_layout_hdr(lo);
907 /* Block new layoutgets and read/write to ds */
908 lo->plh_block_lgets++;
909 spin_unlock(&inode->i_lock);
910 filemap_fdatawait(inode->i_mapping);
911 ret = pnfs_layoutcommit_inode(inode, true);
913 ret = _pnfs_return_layout(inode);
914 spin_lock(&inode->i_lock);
915 lo->plh_block_lgets--;
916 spin_unlock(&inode->i_lock);
917 pnfs_put_layout_hdr(lo);
921 bool pnfs_roc(struct inode *ino)
923 struct pnfs_layout_hdr *lo;
924 struct pnfs_layout_segment *lseg, *tmp;
928 spin_lock(&ino->i_lock);
929 lo = NFS_I(ino)->layout;
930 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
931 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
933 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
934 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
935 mark_lseg_invalid(lseg, &tmp_list);
940 lo->plh_block_lgets++;
941 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
942 spin_unlock(&ino->i_lock);
943 pnfs_free_lseg_list(&tmp_list);
947 spin_unlock(&ino->i_lock);
951 void pnfs_roc_release(struct inode *ino)
953 struct pnfs_layout_hdr *lo;
955 spin_lock(&ino->i_lock);
956 lo = NFS_I(ino)->layout;
957 lo->plh_block_lgets--;
958 if (atomic_dec_and_test(&lo->plh_refcount)) {
959 pnfs_detach_layout_hdr(lo);
960 spin_unlock(&ino->i_lock);
961 pnfs_free_layout_hdr(lo);
963 spin_unlock(&ino->i_lock);
966 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
968 struct pnfs_layout_hdr *lo;
970 spin_lock(&ino->i_lock);
971 lo = NFS_I(ino)->layout;
972 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
973 lo->plh_barrier = barrier;
974 spin_unlock(&ino->i_lock);
977 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
979 struct nfs_inode *nfsi = NFS_I(ino);
980 struct pnfs_layout_hdr *lo;
981 struct pnfs_layout_segment *lseg;
985 spin_lock(&ino->i_lock);
986 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
987 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
988 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
993 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
995 /* Since close does not return a layout stateid for use as
996 * a barrier, we choose the worst-case barrier.
998 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1000 spin_unlock(&ino->i_lock);
1005 * Compare two layout segments for sorting into layout cache.
1006 * We want to preferentially return RW over RO layouts, so ensure those
1010 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1011 const struct pnfs_layout_range *l2)
1015 /* high offset > low offset */
1016 d = l1->offset - l2->offset;
1020 /* short length > long length */
1021 d = l2->length - l1->length;
1025 /* read > read/write */
1026 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1030 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1031 struct pnfs_layout_segment *lseg)
1033 struct pnfs_layout_segment *lp;
1035 dprintk("%s:Begin\n", __func__);
1037 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1038 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1040 list_add_tail(&lseg->pls_list, &lp->pls_list);
1041 dprintk("%s: inserted lseg %p "
1042 "iomode %d offset %llu length %llu before "
1043 "lp %p iomode %d offset %llu length %llu\n",
1044 __func__, lseg, lseg->pls_range.iomode,
1045 lseg->pls_range.offset, lseg->pls_range.length,
1046 lp, lp->pls_range.iomode, lp->pls_range.offset,
1047 lp->pls_range.length);
1050 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1051 dprintk("%s: inserted lseg %p "
1052 "iomode %d offset %llu length %llu at tail\n",
1053 __func__, lseg, lseg->pls_range.iomode,
1054 lseg->pls_range.offset, lseg->pls_range.length);
1056 pnfs_get_layout_hdr(lo);
1058 dprintk("%s:Return\n", __func__);
1061 static struct pnfs_layout_hdr *
1062 alloc_init_layout_hdr(struct inode *ino,
1063 struct nfs_open_context *ctx,
1066 struct pnfs_layout_hdr *lo;
1068 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1071 atomic_set(&lo->plh_refcount, 1);
1072 INIT_LIST_HEAD(&lo->plh_layouts);
1073 INIT_LIST_HEAD(&lo->plh_segs);
1074 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1075 lo->plh_inode = ino;
1076 lo->plh_lc_cred = get_rpccred(ctx->cred);
1080 static struct pnfs_layout_hdr *
1081 pnfs_find_alloc_layout(struct inode *ino,
1082 struct nfs_open_context *ctx,
1085 struct nfs_inode *nfsi = NFS_I(ino);
1086 struct pnfs_layout_hdr *new = NULL;
1088 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1090 if (nfsi->layout != NULL)
1092 spin_unlock(&ino->i_lock);
1093 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1094 spin_lock(&ino->i_lock);
1096 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1099 } else if (new != NULL)
1100 pnfs_free_layout_hdr(new);
1102 pnfs_get_layout_hdr(nfsi->layout);
1103 return nfsi->layout;
1107 * iomode matching rules:
1118 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1119 const struct pnfs_layout_range *range)
1121 struct pnfs_layout_range range1;
1123 if ((range->iomode == IOMODE_RW &&
1124 ls_range->iomode != IOMODE_RW) ||
1125 !pnfs_lseg_range_intersecting(ls_range, range))
1128 /* range1 covers only the first byte in the range */
1131 return pnfs_lseg_range_contained(ls_range, &range1);
1135 * lookup range in layout
1137 static struct pnfs_layout_segment *
1138 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1139 struct pnfs_layout_range *range)
1141 struct pnfs_layout_segment *lseg, *ret = NULL;
1143 dprintk("%s:Begin\n", __func__);
1145 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1146 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1147 pnfs_lseg_range_match(&lseg->pls_range, range)) {
1148 ret = pnfs_get_lseg(lseg);
1151 if (lseg->pls_range.offset > range->offset)
1155 dprintk("%s:Return lseg %p ref %d\n",
1156 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1161 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1162 * to the MDS or over pNFS
1164 * The nfs_inode read_io and write_io fields are cumulative counters reset
1165 * when there are no layout segments. Note that in pnfs_update_layout iomode
1166 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1169 * A return of true means use MDS I/O.
1172 * If a file's size is smaller than the file size threshold, data accesses
1173 * SHOULD be sent to the metadata server. If an I/O request has a length that
1174 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1175 * server. If both file size and I/O size are provided, the client SHOULD
1176 * reach or exceed both thresholds before sending its read or write
1177 * requests to the data server.
1179 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1180 struct inode *ino, int iomode)
1182 struct nfs4_threshold *t = ctx->mdsthreshold;
1183 struct nfs_inode *nfsi = NFS_I(ino);
1184 loff_t fsize = i_size_read(ino);
1185 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1190 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1191 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1195 if (t->bm & THRESHOLD_RD) {
1196 dprintk("%s fsize %llu\n", __func__, fsize);
1198 if (fsize < t->rd_sz)
1201 if (t->bm & THRESHOLD_RD_IO) {
1202 dprintk("%s nfsi->read_io %llu\n", __func__,
1205 if (nfsi->read_io < t->rd_io_sz)
1210 if (t->bm & THRESHOLD_WR) {
1211 dprintk("%s fsize %llu\n", __func__, fsize);
1213 if (fsize < t->wr_sz)
1216 if (t->bm & THRESHOLD_WR_IO) {
1217 dprintk("%s nfsi->write_io %llu\n", __func__,
1220 if (nfsi->write_io < t->wr_io_sz)
1225 if (size_set && io_set) {
1228 } else if (size || io)
1231 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1236 * Layout segment is retreived from the server if not cached.
1237 * The appropriate layout segment is referenced and returned to the caller.
1239 struct pnfs_layout_segment *
1240 pnfs_update_layout(struct inode *ino,
1241 struct nfs_open_context *ctx,
1244 enum pnfs_iomode iomode,
1247 struct pnfs_layout_range arg = {
1253 struct nfs_server *server = NFS_SERVER(ino);
1254 struct nfs_client *clp = server->nfs_client;
1255 struct pnfs_layout_hdr *lo;
1256 struct pnfs_layout_segment *lseg = NULL;
1259 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1262 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1265 spin_lock(&ino->i_lock);
1266 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1268 spin_unlock(&ino->i_lock);
1272 /* Do we even need to bother with this? */
1273 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1274 dprintk("%s matches recall, use MDS\n", __func__);
1278 /* if LAYOUTGET already failed once we don't try again */
1279 if (pnfs_layout_io_test_failed(lo, iomode))
1282 /* Check to see if the layout for the given range already exists */
1283 lseg = pnfs_find_lseg(lo, &arg);
1287 if (pnfs_layoutgets_blocked(lo, 0))
1289 atomic_inc(&lo->plh_outstanding);
1291 first = list_empty(&lo->plh_layouts) ? true : false;
1292 spin_unlock(&ino->i_lock);
1295 /* The lo must be on the clp list if there is any
1296 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1298 spin_lock(&clp->cl_lock);
1299 list_add_tail(&lo->plh_layouts, &server->layouts);
1300 spin_unlock(&clp->cl_lock);
1303 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1305 arg.offset -= pg_offset;
1306 arg.length += pg_offset;
1308 if (arg.length != NFS4_MAX_UINT64)
1309 arg.length = PAGE_CACHE_ALIGN(arg.length);
1311 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1312 atomic_dec(&lo->plh_outstanding);
1314 pnfs_put_layout_hdr(lo);
1316 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1317 "(%s, offset: %llu, length: %llu)\n",
1318 __func__, ino->i_sb->s_id,
1319 (unsigned long long)NFS_FILEID(ino),
1320 lseg == NULL ? "not found" : "found",
1321 iomode==IOMODE_RW ? "read/write" : "read-only",
1322 (unsigned long long)pos,
1323 (unsigned long long)count);
1326 spin_unlock(&ino->i_lock);
1327 goto out_put_layout_hdr;
1329 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1331 struct pnfs_layout_segment *
1332 pnfs_layout_process(struct nfs4_layoutget *lgp)
1334 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1335 struct nfs4_layoutget_res *res = &lgp->res;
1336 struct pnfs_layout_segment *lseg;
1337 struct inode *ino = lo->plh_inode;
1341 /* Inject layout blob into I/O device driver */
1342 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1343 if (!lseg || IS_ERR(lseg)) {
1347 status = PTR_ERR(lseg);
1348 dprintk("%s: Could not allocate layout: error %d\n",
1353 init_lseg(lo, lseg);
1354 lseg->pls_range = res->range;
1356 spin_lock(&ino->i_lock);
1357 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1358 dprintk("%s forget reply due to recall\n", __func__);
1359 goto out_forget_reply;
1362 if (pnfs_layoutgets_blocked(lo, 1)) {
1363 dprintk("%s forget reply due to state\n", __func__);
1364 goto out_forget_reply;
1367 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1368 /* existing state ID, make sure the sequence number matches. */
1369 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1370 dprintk("%s forget reply due to sequence\n", __func__);
1371 goto out_forget_reply;
1373 pnfs_set_layout_stateid(lo, &res->stateid, false);
1376 * We got an entirely new state ID. Mark all segments for the
1377 * inode invalid, and don't bother validating the stateid
1380 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1382 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1383 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1386 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1388 pnfs_get_lseg(lseg);
1389 pnfs_layout_insert_lseg(lo, lseg);
1391 if (res->return_on_close) {
1392 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1393 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1396 spin_unlock(&ino->i_lock);
1397 pnfs_free_lseg_list(&free_me);
1400 return ERR_PTR(status);
1403 spin_unlock(&ino->i_lock);
1404 lseg->pls_layout = lo;
1405 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1410 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1412 u64 rd_size = req->wb_bytes;
1414 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1416 if (pgio->pg_dreq == NULL)
1417 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1419 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1421 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1427 /* If no lseg, fall back to read through mds */
1428 if (pgio->pg_lseg == NULL)
1429 nfs_pageio_reset_read_mds(pgio);
1432 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1435 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1436 struct nfs_page *req, u64 wb_size)
1438 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1440 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1446 /* If no lseg, fall back to write through mds */
1447 if (pgio->pg_lseg == NULL)
1448 nfs_pageio_reset_write_mds(pgio);
1450 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1453 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1454 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1457 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1458 struct nfs_page *req)
1461 u64 seg_end, req_start, seg_left;
1463 size = nfs_generic_pg_test(pgio, prev, req);
1468 * 'size' contains the number of bytes left in the current page (up
1469 * to the original size asked for in @req->wb_bytes).
1471 * Calculate how many bytes are left in the layout segment
1472 * and if there are less bytes than 'size', return that instead.
1474 * Please also note that 'end_offset' is actually the offset of the
1475 * first byte that lies outside the pnfs_layout_range. FIXME?
1478 if (pgio->pg_lseg) {
1479 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1480 pgio->pg_lseg->pls_range.length);
1481 req_start = req_offset(req);
1482 WARN_ON_ONCE(req_start > seg_end);
1483 /* start of request is past the last byte of this segment */
1484 if (req_start >= seg_end)
1487 /* adjust 'size' iff there are fewer bytes left in the
1488 * segment than what nfs_generic_pg_test returned */
1489 seg_left = seg_end - req_start;
1490 if (seg_left < size)
1491 size = (unsigned int)seg_left;
1496 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1498 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1500 struct nfs_pageio_descriptor pgio;
1502 /* Resend all requests through the MDS */
1503 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1504 hdr->completion_ops);
1505 return nfs_pageio_resend(&pgio, hdr);
1507 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1509 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1512 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1513 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1514 PNFS_LAYOUTRET_ON_ERROR) {
1515 pnfs_return_layout(hdr->inode);
1517 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1518 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1522 * Called by non rpc-based layout drivers
1524 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1526 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1527 if (!hdr->pnfs_error) {
1528 pnfs_set_layoutcommit(hdr);
1529 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1531 pnfs_ld_handle_write_error(hdr);
1532 hdr->mds_ops->rpc_release(hdr);
1534 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1537 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1538 struct nfs_pgio_header *hdr)
1540 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1541 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1542 nfs_pageio_reset_write_mds(desc);
1543 desc->pg_recoalesce = 1;
1545 nfs_pgio_data_destroy(hdr);
1548 static enum pnfs_try_status
1549 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1550 const struct rpc_call_ops *call_ops,
1551 struct pnfs_layout_segment *lseg,
1554 struct inode *inode = hdr->inode;
1555 enum pnfs_try_status trypnfs;
1556 struct nfs_server *nfss = NFS_SERVER(inode);
1558 hdr->mds_ops = call_ops;
1560 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1561 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1562 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1563 if (trypnfs != PNFS_NOT_ATTEMPTED)
1564 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1565 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1570 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1571 struct nfs_pgio_header *hdr, int how)
1573 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1574 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1575 enum pnfs_try_status trypnfs;
1577 desc->pg_lseg = NULL;
1578 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1579 if (trypnfs == PNFS_NOT_ATTEMPTED)
1580 pnfs_write_through_mds(desc, hdr);
1581 pnfs_put_lseg(lseg);
1584 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1586 pnfs_put_lseg(hdr->lseg);
1587 nfs_pgio_header_free(hdr);
1589 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1592 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1594 struct nfs_pgio_header *hdr;
1597 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1599 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1600 pnfs_put_lseg(desc->pg_lseg);
1601 desc->pg_lseg = NULL;
1604 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1605 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1606 ret = nfs_generic_pgio(desc, hdr);
1608 pnfs_put_lseg(desc->pg_lseg);
1609 desc->pg_lseg = NULL;
1611 pnfs_do_write(desc, hdr, desc->pg_ioflags);
1614 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1616 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
1618 struct nfs_pageio_descriptor pgio;
1620 /* Resend all requests through the MDS */
1621 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
1622 return nfs_pageio_resend(&pgio, hdr);
1624 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1626 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
1628 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1629 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1630 PNFS_LAYOUTRET_ON_ERROR) {
1631 pnfs_return_layout(hdr->inode);
1633 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1634 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
1638 * Called by non rpc-based layout drivers
1640 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
1642 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
1643 if (likely(!hdr->pnfs_error)) {
1644 __nfs4_read_done_cb(hdr);
1645 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1647 pnfs_ld_handle_read_error(hdr);
1648 hdr->mds_ops->rpc_release(hdr);
1650 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1653 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1654 struct nfs_pgio_header *hdr)
1656 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1657 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1658 nfs_pageio_reset_read_mds(desc);
1659 desc->pg_recoalesce = 1;
1661 nfs_pgio_data_destroy(hdr);
1665 * Call the appropriate parallel I/O subsystem read function.
1667 static enum pnfs_try_status
1668 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
1669 const struct rpc_call_ops *call_ops,
1670 struct pnfs_layout_segment *lseg)
1672 struct inode *inode = hdr->inode;
1673 struct nfs_server *nfss = NFS_SERVER(inode);
1674 enum pnfs_try_status trypnfs;
1676 hdr->mds_ops = call_ops;
1678 dprintk("%s: Reading ino:%lu %u@%llu\n",
1679 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
1681 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
1682 if (trypnfs != PNFS_NOT_ATTEMPTED)
1683 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1684 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1689 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
1691 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1692 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1693 enum pnfs_try_status trypnfs;
1695 desc->pg_lseg = NULL;
1696 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
1697 if (trypnfs == PNFS_NOT_ATTEMPTED)
1698 pnfs_read_through_mds(desc, hdr);
1699 pnfs_put_lseg(lseg);
1702 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1704 pnfs_put_lseg(hdr->lseg);
1705 nfs_pgio_header_free(hdr);
1707 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1710 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1712 struct nfs_pgio_header *hdr;
1715 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1717 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1719 pnfs_put_lseg(desc->pg_lseg);
1720 desc->pg_lseg = NULL;
1723 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1724 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1725 ret = nfs_generic_pgio(desc, hdr);
1727 pnfs_put_lseg(desc->pg_lseg);
1728 desc->pg_lseg = NULL;
1730 pnfs_do_read(desc, hdr);
1733 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1735 static void pnfs_clear_layoutcommitting(struct inode *inode)
1737 unsigned long *bitlock = &NFS_I(inode)->flags;
1739 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1740 smp_mb__after_atomic();
1741 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1745 * There can be multiple RW segments.
1747 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1749 struct pnfs_layout_segment *lseg;
1751 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1752 if (lseg->pls_range.iomode == IOMODE_RW &&
1753 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1754 list_add(&lseg->pls_lc_list, listp);
1758 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1760 struct pnfs_layout_segment *lseg, *tmp;
1762 /* Matched by references in pnfs_set_layoutcommit */
1763 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1764 list_del_init(&lseg->pls_lc_list);
1765 pnfs_put_lseg(lseg);
1768 pnfs_clear_layoutcommitting(inode);
1771 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1773 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1775 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1778 pnfs_set_layoutcommit(struct nfs_pgio_header *hdr)
1780 struct inode *inode = hdr->inode;
1781 struct nfs_inode *nfsi = NFS_I(inode);
1782 loff_t end_pos = hdr->mds_offset + hdr->res.count;
1783 bool mark_as_dirty = false;
1785 spin_lock(&inode->i_lock);
1786 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1787 mark_as_dirty = true;
1788 dprintk("%s: Set layoutcommit for inode %lu ",
1789 __func__, inode->i_ino);
1791 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1792 /* references matched in nfs4_layoutcommit_release */
1793 pnfs_get_lseg(hdr->lseg);
1795 if (end_pos > nfsi->layout->plh_lwb)
1796 nfsi->layout->plh_lwb = end_pos;
1797 spin_unlock(&inode->i_lock);
1798 dprintk("%s: lseg %p end_pos %llu\n",
1799 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1801 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1802 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1804 mark_inode_dirty_sync(inode);
1806 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1808 void pnfs_commit_set_layoutcommit(struct nfs_commit_data *data)
1810 struct inode *inode = data->inode;
1811 struct nfs_inode *nfsi = NFS_I(inode);
1812 bool mark_as_dirty = false;
1814 spin_lock(&inode->i_lock);
1815 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1816 mark_as_dirty = true;
1817 dprintk("%s: Set layoutcommit for inode %lu ",
1818 __func__, inode->i_ino);
1820 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &data->lseg->pls_flags)) {
1821 /* references matched in nfs4_layoutcommit_release */
1822 pnfs_get_lseg(data->lseg);
1824 if (data->lwb > nfsi->layout->plh_lwb)
1825 nfsi->layout->plh_lwb = data->lwb;
1826 spin_unlock(&inode->i_lock);
1827 dprintk("%s: lseg %p end_pos %llu\n",
1828 __func__, data->lseg, nfsi->layout->plh_lwb);
1830 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1831 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1833 mark_inode_dirty_sync(inode);
1835 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit);
1837 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1839 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1841 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1842 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1843 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1847 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1848 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1849 * data to disk to allow the server to recover the data if it crashes.
1850 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1851 * is off, and a COMMIT is sent to a data server, or
1852 * if WRITEs to a data server return NFS_DATA_SYNC.
1855 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1857 struct nfs4_layoutcommit_data *data;
1858 struct nfs_inode *nfsi = NFS_I(inode);
1862 if (!pnfs_layoutcommit_outstanding(inode))
1865 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1868 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1871 status = wait_on_bit_lock_action(&nfsi->flags,
1872 NFS_INO_LAYOUTCOMMITTING,
1873 nfs_wait_bit_killable,
1880 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1881 data = kzalloc(sizeof(*data), GFP_NOFS);
1883 goto clear_layoutcommitting;
1886 spin_lock(&inode->i_lock);
1887 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1890 INIT_LIST_HEAD(&data->lseg_list);
1891 pnfs_list_write_lseg(inode, &data->lseg_list);
1893 end_pos = nfsi->layout->plh_lwb;
1894 nfsi->layout->plh_lwb = 0;
1896 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1897 spin_unlock(&inode->i_lock);
1899 data->args.inode = inode;
1900 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1901 nfs_fattr_init(&data->fattr);
1902 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1903 data->res.fattr = &data->fattr;
1904 data->args.lastbytewritten = end_pos - 1;
1905 data->res.server = NFS_SERVER(inode);
1907 status = nfs4_proc_layoutcommit(data, sync);
1910 mark_inode_dirty_sync(inode);
1911 dprintk("<-- %s status %d\n", __func__, status);
1914 spin_unlock(&inode->i_lock);
1916 clear_layoutcommitting:
1917 pnfs_clear_layoutcommitting(inode);
1921 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1923 struct nfs4_threshold *thp;
1925 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1927 dprintk("%s mdsthreshold allocation failed\n", __func__);