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);
365 end_offset(u64 start, u64 len)
370 return end >= start ? end : NFS4_MAX_UINT64;
374 * is l2 fully contained in l1?
376 * [----------------------------------)
381 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
382 const struct pnfs_layout_range *l2)
384 u64 start1 = l1->offset;
385 u64 end1 = end_offset(start1, l1->length);
386 u64 start2 = l2->offset;
387 u64 end2 = end_offset(start2, l2->length);
389 return (start1 <= start2) && (end1 >= end2);
393 * is l1 and l2 intersecting?
395 * [----------------------------------)
400 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
401 const struct pnfs_layout_range *l2)
403 u64 start1 = l1->offset;
404 u64 end1 = end_offset(start1, l1->length);
405 u64 start2 = l2->offset;
406 u64 end2 = end_offset(start2, l2->length);
408 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
409 (end2 == NFS4_MAX_UINT64 || end2 > start1);
413 should_free_lseg(const struct pnfs_layout_range *lseg_range,
414 const struct pnfs_layout_range *recall_range)
416 return (recall_range->iomode == IOMODE_ANY ||
417 lseg_range->iomode == recall_range->iomode) &&
418 pnfs_lseg_range_intersecting(lseg_range, recall_range);
421 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
422 struct list_head *tmp_list)
424 if (!atomic_dec_and_test(&lseg->pls_refcount))
426 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
427 list_add(&lseg->pls_list, tmp_list);
431 /* Returns 1 if lseg is removed from list, 0 otherwise */
432 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
433 struct list_head *tmp_list)
437 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
438 /* Remove the reference keeping the lseg in the
439 * list. It will now be removed when all
440 * outstanding io is finished.
442 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
443 atomic_read(&lseg->pls_refcount));
444 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
450 /* Returns count of number of matching invalid lsegs remaining in list
454 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
455 struct list_head *tmp_list,
456 struct pnfs_layout_range *recall_range)
458 struct pnfs_layout_segment *lseg, *next;
459 int invalid = 0, removed = 0;
461 dprintk("%s:Begin lo %p\n", __func__, lo);
463 if (list_empty(&lo->plh_segs))
465 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
467 should_free_lseg(&lseg->pls_range, recall_range)) {
468 dprintk("%s: freeing lseg %p iomode %d "
469 "offset %llu length %llu\n", __func__,
470 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
471 lseg->pls_range.length);
473 removed += mark_lseg_invalid(lseg, tmp_list);
475 dprintk("%s:Return %i\n", __func__, invalid - removed);
476 return invalid - removed;
479 /* note free_me must contain lsegs from a single layout_hdr */
481 pnfs_free_lseg_list(struct list_head *free_me)
483 struct pnfs_layout_segment *lseg, *tmp;
485 if (list_empty(free_me))
488 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
489 list_del(&lseg->pls_list);
490 pnfs_free_lseg(lseg);
495 pnfs_destroy_layout(struct nfs_inode *nfsi)
497 struct pnfs_layout_hdr *lo;
500 spin_lock(&nfsi->vfs_inode.i_lock);
503 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
504 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
505 pnfs_get_layout_hdr(lo);
506 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
507 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
508 spin_unlock(&nfsi->vfs_inode.i_lock);
509 pnfs_free_lseg_list(&tmp_list);
510 pnfs_put_layout_hdr(lo);
512 spin_unlock(&nfsi->vfs_inode.i_lock);
514 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
517 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
518 struct list_head *layout_list)
520 struct pnfs_layout_hdr *lo;
523 spin_lock(&inode->i_lock);
524 lo = NFS_I(inode)->layout;
525 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
526 pnfs_get_layout_hdr(lo);
527 list_add(&lo->plh_bulk_destroy, layout_list);
530 spin_unlock(&inode->i_lock);
534 /* Caller must hold rcu_read_lock and clp->cl_lock */
536 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
537 struct nfs_server *server,
538 struct list_head *layout_list)
540 struct pnfs_layout_hdr *lo, *next;
543 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
544 inode = igrab(lo->plh_inode);
547 list_del_init(&lo->plh_layouts);
548 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
551 spin_unlock(&clp->cl_lock);
553 spin_lock(&clp->cl_lock);
561 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
564 struct pnfs_layout_hdr *lo;
566 struct pnfs_layout_range range = {
567 .iomode = IOMODE_ANY,
569 .length = NFS4_MAX_UINT64,
571 LIST_HEAD(lseg_list);
574 while (!list_empty(layout_list)) {
575 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
577 dprintk("%s freeing layout for inode %lu\n", __func__,
578 lo->plh_inode->i_ino);
579 inode = lo->plh_inode;
580 spin_lock(&inode->i_lock);
581 list_del_init(&lo->plh_bulk_destroy);
582 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
584 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
585 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
587 spin_unlock(&inode->i_lock);
588 pnfs_free_lseg_list(&lseg_list);
589 pnfs_put_layout_hdr(lo);
596 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
597 struct nfs_fsid *fsid,
600 struct nfs_server *server;
601 LIST_HEAD(layout_list);
603 spin_lock(&clp->cl_lock);
606 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
607 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
609 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
615 spin_unlock(&clp->cl_lock);
617 if (list_empty(&layout_list))
619 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
623 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
626 struct nfs_server *server;
627 LIST_HEAD(layout_list);
629 spin_lock(&clp->cl_lock);
632 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
633 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
639 spin_unlock(&clp->cl_lock);
641 if (list_empty(&layout_list))
643 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
647 * Called by the state manger to remove all layouts established under an
651 pnfs_destroy_all_layouts(struct nfs_client *clp)
653 nfs4_deviceid_mark_client_invalid(clp);
654 nfs4_deviceid_purge_client(clp);
656 pnfs_destroy_layouts_byclid(clp, false);
660 * Compare 2 layout stateid sequence ids, to see which is newer,
661 * taking into account wraparound issues.
663 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
665 return (s32)(s1 - s2) > 0;
669 pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
670 const nfs4_stateid *new,
671 struct list_head *free_me_list)
673 if (nfs4_stateid_match_other(&lo->plh_stateid, new))
675 /* Layout is new! Kill existing layout segments */
676 pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
679 /* update lo->plh_stateid with new if is more recent */
681 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
684 u32 oldseq, newseq, new_barrier;
685 int empty = list_empty(&lo->plh_segs);
687 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
688 newseq = be32_to_cpu(new->seqid);
689 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
690 nfs4_stateid_copy(&lo->plh_stateid, new);
691 if (update_barrier) {
692 new_barrier = be32_to_cpu(new->seqid);
694 /* Because of wraparound, we want to keep the barrier
695 * "close" to the current seqids.
697 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
699 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
700 lo->plh_barrier = new_barrier;
705 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
706 const nfs4_stateid *stateid)
708 u32 seqid = be32_to_cpu(stateid->seqid);
710 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
713 /* lget is set to 1 if called from inside send_layoutget call chain */
715 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
717 return lo->plh_block_lgets ||
718 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
719 (list_empty(&lo->plh_segs) &&
720 (atomic_read(&lo->plh_outstanding) > lget));
724 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
725 struct nfs4_state *open_state)
729 dprintk("--> %s\n", __func__);
730 spin_lock(&lo->plh_inode->i_lock);
731 if (pnfs_layoutgets_blocked(lo, 1)) {
733 } else if (!nfs4_valid_open_stateid(open_state)) {
735 } else if (list_empty(&lo->plh_segs)) {
739 seq = read_seqbegin(&open_state->seqlock);
740 nfs4_stateid_copy(dst, &open_state->stateid);
741 } while (read_seqretry(&open_state->seqlock, seq));
743 nfs4_stateid_copy(dst, &lo->plh_stateid);
744 spin_unlock(&lo->plh_inode->i_lock);
745 dprintk("<-- %s\n", __func__);
750 * Get layout from server.
751 * for now, assume that whole file layouts are requested.
753 * arg->length: all ones
755 static struct pnfs_layout_segment *
756 send_layoutget(struct pnfs_layout_hdr *lo,
757 struct nfs_open_context *ctx,
758 struct pnfs_layout_range *range,
761 struct inode *ino = lo->plh_inode;
762 struct nfs_server *server = NFS_SERVER(ino);
763 struct nfs4_layoutget *lgp;
764 struct pnfs_layout_segment *lseg;
766 dprintk("--> %s\n", __func__);
768 lgp = kzalloc(sizeof(*lgp), gfp_flags);
772 lgp->args.minlength = PAGE_CACHE_SIZE;
773 if (lgp->args.minlength > range->length)
774 lgp->args.minlength = range->length;
775 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
776 lgp->args.range = *range;
777 lgp->args.type = server->pnfs_curr_ld->id;
778 lgp->args.inode = ino;
779 lgp->args.ctx = get_nfs_open_context(ctx);
780 lgp->gfp_flags = gfp_flags;
781 lgp->cred = lo->plh_lc_cred;
783 /* Synchronously retrieve layout information from server and
786 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
788 switch (PTR_ERR(lseg)) {
793 /* remember that LAYOUTGET failed and suspend trying */
794 pnfs_layout_io_set_failed(lo, range->iomode);
802 static void pnfs_clear_layoutcommit(struct inode *inode,
803 struct list_head *head)
805 struct nfs_inode *nfsi = NFS_I(inode);
806 struct pnfs_layout_segment *lseg, *tmp;
808 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
810 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
811 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
813 pnfs_lseg_dec_and_remove_zero(lseg, head);
818 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
819 * when the layout segment list is empty.
821 * Note that a pnfs_layout_hdr can exist with an empty layout segment
822 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
823 * deviceid is marked invalid.
826 _pnfs_return_layout(struct inode *ino)
828 struct pnfs_layout_hdr *lo = NULL;
829 struct nfs_inode *nfsi = NFS_I(ino);
831 struct nfs4_layoutreturn *lrp;
832 nfs4_stateid stateid;
833 int status = 0, empty;
835 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
837 spin_lock(&ino->i_lock);
840 spin_unlock(&ino->i_lock);
841 dprintk("NFS: %s no layout to return\n", __func__);
844 stateid = nfsi->layout->plh_stateid;
845 /* Reference matched in nfs4_layoutreturn_release */
846 pnfs_get_layout_hdr(lo);
847 empty = list_empty(&lo->plh_segs);
848 pnfs_clear_layoutcommit(ino, &tmp_list);
849 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
850 /* Don't send a LAYOUTRETURN if list was initially empty */
852 spin_unlock(&ino->i_lock);
853 pnfs_put_layout_hdr(lo);
854 dprintk("NFS: %s no layout segments to return\n", __func__);
857 lo->plh_block_lgets++;
858 spin_unlock(&ino->i_lock);
859 pnfs_free_lseg_list(&tmp_list);
861 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
862 if (unlikely(lrp == NULL)) {
864 spin_lock(&ino->i_lock);
865 lo->plh_block_lgets--;
866 spin_unlock(&ino->i_lock);
867 pnfs_put_layout_hdr(lo);
871 lrp->args.stateid = stateid;
872 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
873 lrp->args.inode = ino;
874 lrp->args.layout = lo;
875 lrp->clp = NFS_SERVER(ino)->nfs_client;
876 lrp->cred = lo->plh_lc_cred;
878 status = nfs4_proc_layoutreturn(lrp);
880 dprintk("<-- %s status: %d\n", __func__, status);
883 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
886 pnfs_commit_and_return_layout(struct inode *inode)
888 struct pnfs_layout_hdr *lo;
891 spin_lock(&inode->i_lock);
892 lo = NFS_I(inode)->layout;
894 spin_unlock(&inode->i_lock);
897 pnfs_get_layout_hdr(lo);
898 /* Block new layoutgets and read/write to ds */
899 lo->plh_block_lgets++;
900 spin_unlock(&inode->i_lock);
901 filemap_fdatawait(inode->i_mapping);
902 ret = pnfs_layoutcommit_inode(inode, true);
904 ret = _pnfs_return_layout(inode);
905 spin_lock(&inode->i_lock);
906 lo->plh_block_lgets--;
907 spin_unlock(&inode->i_lock);
908 pnfs_put_layout_hdr(lo);
912 bool pnfs_roc(struct inode *ino)
914 struct pnfs_layout_hdr *lo;
915 struct pnfs_layout_segment *lseg, *tmp;
919 spin_lock(&ino->i_lock);
920 lo = NFS_I(ino)->layout;
921 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
922 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
924 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
925 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
926 mark_lseg_invalid(lseg, &tmp_list);
931 lo->plh_block_lgets++;
932 pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
933 spin_unlock(&ino->i_lock);
934 pnfs_free_lseg_list(&tmp_list);
938 spin_unlock(&ino->i_lock);
942 void pnfs_roc_release(struct inode *ino)
944 struct pnfs_layout_hdr *lo;
946 spin_lock(&ino->i_lock);
947 lo = NFS_I(ino)->layout;
948 lo->plh_block_lgets--;
949 if (atomic_dec_and_test(&lo->plh_refcount)) {
950 pnfs_detach_layout_hdr(lo);
951 spin_unlock(&ino->i_lock);
952 pnfs_free_layout_hdr(lo);
954 spin_unlock(&ino->i_lock);
957 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
959 struct pnfs_layout_hdr *lo;
961 spin_lock(&ino->i_lock);
962 lo = NFS_I(ino)->layout;
963 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
964 lo->plh_barrier = barrier;
965 spin_unlock(&ino->i_lock);
968 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
970 struct nfs_inode *nfsi = NFS_I(ino);
971 struct pnfs_layout_hdr *lo;
972 struct pnfs_layout_segment *lseg;
976 spin_lock(&ino->i_lock);
977 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
978 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
979 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
984 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
986 /* Since close does not return a layout stateid for use as
987 * a barrier, we choose the worst-case barrier.
989 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
991 spin_unlock(&ino->i_lock);
996 * Compare two layout segments for sorting into layout cache.
997 * We want to preferentially return RW over RO layouts, so ensure those
1001 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1002 const struct pnfs_layout_range *l2)
1006 /* high offset > low offset */
1007 d = l1->offset - l2->offset;
1011 /* short length > long length */
1012 d = l2->length - l1->length;
1016 /* read > read/write */
1017 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1021 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1022 struct pnfs_layout_segment *lseg)
1024 struct pnfs_layout_segment *lp;
1026 dprintk("%s:Begin\n", __func__);
1028 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
1029 if (pnfs_lseg_range_cmp(&lseg->pls_range, &lp->pls_range) > 0)
1031 list_add_tail(&lseg->pls_list, &lp->pls_list);
1032 dprintk("%s: inserted lseg %p "
1033 "iomode %d offset %llu length %llu before "
1034 "lp %p iomode %d offset %llu length %llu\n",
1035 __func__, lseg, lseg->pls_range.iomode,
1036 lseg->pls_range.offset, lseg->pls_range.length,
1037 lp, lp->pls_range.iomode, lp->pls_range.offset,
1038 lp->pls_range.length);
1041 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1042 dprintk("%s: inserted lseg %p "
1043 "iomode %d offset %llu length %llu at tail\n",
1044 __func__, lseg, lseg->pls_range.iomode,
1045 lseg->pls_range.offset, lseg->pls_range.length);
1047 pnfs_get_layout_hdr(lo);
1049 dprintk("%s:Return\n", __func__);
1052 static struct pnfs_layout_hdr *
1053 alloc_init_layout_hdr(struct inode *ino,
1054 struct nfs_open_context *ctx,
1057 struct pnfs_layout_hdr *lo;
1059 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1062 atomic_set(&lo->plh_refcount, 1);
1063 INIT_LIST_HEAD(&lo->plh_layouts);
1064 INIT_LIST_HEAD(&lo->plh_segs);
1065 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1066 lo->plh_inode = ino;
1067 lo->plh_lc_cred = get_rpccred(ctx->cred);
1071 static struct pnfs_layout_hdr *
1072 pnfs_find_alloc_layout(struct inode *ino,
1073 struct nfs_open_context *ctx,
1076 struct nfs_inode *nfsi = NFS_I(ino);
1077 struct pnfs_layout_hdr *new = NULL;
1079 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1081 if (nfsi->layout != NULL)
1083 spin_unlock(&ino->i_lock);
1084 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1085 spin_lock(&ino->i_lock);
1087 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1090 } else if (new != NULL)
1091 pnfs_free_layout_hdr(new);
1093 pnfs_get_layout_hdr(nfsi->layout);
1094 return nfsi->layout;
1098 * iomode matching rules:
1109 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1110 const struct pnfs_layout_range *range)
1112 struct pnfs_layout_range range1;
1114 if ((range->iomode == IOMODE_RW &&
1115 ls_range->iomode != IOMODE_RW) ||
1116 !pnfs_lseg_range_intersecting(ls_range, range))
1119 /* range1 covers only the first byte in the range */
1122 return pnfs_lseg_range_contained(ls_range, &range1);
1126 * lookup range in layout
1128 static struct pnfs_layout_segment *
1129 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1130 struct pnfs_layout_range *range)
1132 struct pnfs_layout_segment *lseg, *ret = NULL;
1134 dprintk("%s:Begin\n", __func__);
1136 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1137 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1138 pnfs_lseg_range_match(&lseg->pls_range, range)) {
1139 ret = pnfs_get_lseg(lseg);
1142 if (lseg->pls_range.offset > range->offset)
1146 dprintk("%s:Return lseg %p ref %d\n",
1147 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1152 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1153 * to the MDS or over pNFS
1155 * The nfs_inode read_io and write_io fields are cumulative counters reset
1156 * when there are no layout segments. Note that in pnfs_update_layout iomode
1157 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1160 * A return of true means use MDS I/O.
1163 * If a file's size is smaller than the file size threshold, data accesses
1164 * SHOULD be sent to the metadata server. If an I/O request has a length that
1165 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1166 * server. If both file size and I/O size are provided, the client SHOULD
1167 * reach or exceed both thresholds before sending its read or write
1168 * requests to the data server.
1170 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1171 struct inode *ino, int iomode)
1173 struct nfs4_threshold *t = ctx->mdsthreshold;
1174 struct nfs_inode *nfsi = NFS_I(ino);
1175 loff_t fsize = i_size_read(ino);
1176 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1181 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1182 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1186 if (t->bm & THRESHOLD_RD) {
1187 dprintk("%s fsize %llu\n", __func__, fsize);
1189 if (fsize < t->rd_sz)
1192 if (t->bm & THRESHOLD_RD_IO) {
1193 dprintk("%s nfsi->read_io %llu\n", __func__,
1196 if (nfsi->read_io < t->rd_io_sz)
1201 if (t->bm & THRESHOLD_WR) {
1202 dprintk("%s fsize %llu\n", __func__, fsize);
1204 if (fsize < t->wr_sz)
1207 if (t->bm & THRESHOLD_WR_IO) {
1208 dprintk("%s nfsi->write_io %llu\n", __func__,
1211 if (nfsi->write_io < t->wr_io_sz)
1216 if (size_set && io_set) {
1219 } else if (size || io)
1222 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1227 * Layout segment is retreived from the server if not cached.
1228 * The appropriate layout segment is referenced and returned to the caller.
1230 struct pnfs_layout_segment *
1231 pnfs_update_layout(struct inode *ino,
1232 struct nfs_open_context *ctx,
1235 enum pnfs_iomode iomode,
1238 struct pnfs_layout_range arg = {
1244 struct nfs_server *server = NFS_SERVER(ino);
1245 struct nfs_client *clp = server->nfs_client;
1246 struct pnfs_layout_hdr *lo;
1247 struct pnfs_layout_segment *lseg = NULL;
1250 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1253 if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1256 spin_lock(&ino->i_lock);
1257 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1259 spin_unlock(&ino->i_lock);
1263 /* Do we even need to bother with this? */
1264 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1265 dprintk("%s matches recall, use MDS\n", __func__);
1269 /* if LAYOUTGET already failed once we don't try again */
1270 if (pnfs_layout_io_test_failed(lo, iomode))
1273 /* Check to see if the layout for the given range already exists */
1274 lseg = pnfs_find_lseg(lo, &arg);
1278 if (pnfs_layoutgets_blocked(lo, 0))
1280 atomic_inc(&lo->plh_outstanding);
1282 first = list_empty(&lo->plh_layouts) ? true : false;
1283 spin_unlock(&ino->i_lock);
1286 /* The lo must be on the clp list if there is any
1287 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1289 spin_lock(&clp->cl_lock);
1290 list_add_tail(&lo->plh_layouts, &server->layouts);
1291 spin_unlock(&clp->cl_lock);
1294 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1296 arg.offset -= pg_offset;
1297 arg.length += pg_offset;
1299 if (arg.length != NFS4_MAX_UINT64)
1300 arg.length = PAGE_CACHE_ALIGN(arg.length);
1302 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1303 atomic_dec(&lo->plh_outstanding);
1305 pnfs_put_layout_hdr(lo);
1307 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1308 "(%s, offset: %llu, length: %llu)\n",
1309 __func__, ino->i_sb->s_id,
1310 (unsigned long long)NFS_FILEID(ino),
1311 lseg == NULL ? "not found" : "found",
1312 iomode==IOMODE_RW ? "read/write" : "read-only",
1313 (unsigned long long)pos,
1314 (unsigned long long)count);
1317 spin_unlock(&ino->i_lock);
1318 goto out_put_layout_hdr;
1320 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1322 struct pnfs_layout_segment *
1323 pnfs_layout_process(struct nfs4_layoutget *lgp)
1325 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1326 struct nfs4_layoutget_res *res = &lgp->res;
1327 struct pnfs_layout_segment *lseg;
1328 struct inode *ino = lo->plh_inode;
1332 /* Inject layout blob into I/O device driver */
1333 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1334 if (!lseg || IS_ERR(lseg)) {
1338 status = PTR_ERR(lseg);
1339 dprintk("%s: Could not allocate layout: error %d\n",
1344 spin_lock(&ino->i_lock);
1345 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1346 dprintk("%s forget reply due to recall\n", __func__);
1347 goto out_forget_reply;
1350 if (pnfs_layoutgets_blocked(lo, 1) ||
1351 pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1352 dprintk("%s forget reply due to state\n", __func__);
1353 goto out_forget_reply;
1356 /* Check that the new stateid matches the old stateid */
1357 pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
1358 /* Done processing layoutget. Set the layout stateid */
1359 pnfs_set_layout_stateid(lo, &res->stateid, false);
1361 init_lseg(lo, lseg);
1362 lseg->pls_range = res->range;
1363 pnfs_get_lseg(lseg);
1364 pnfs_layout_insert_lseg(lo, lseg);
1366 if (res->return_on_close) {
1367 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1368 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1371 spin_unlock(&ino->i_lock);
1372 pnfs_free_lseg_list(&free_me);
1375 return ERR_PTR(status);
1378 spin_unlock(&ino->i_lock);
1379 lseg->pls_layout = lo;
1380 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1385 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1387 u64 rd_size = req->wb_bytes;
1389 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1391 if (req->wb_offset != req->wb_pgbase) {
1392 nfs_pageio_reset_read_mds(pgio);
1396 if (pgio->pg_dreq == NULL)
1397 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1399 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1401 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1407 /* If no lseg, fall back to read through mds */
1408 if (pgio->pg_lseg == NULL)
1409 nfs_pageio_reset_read_mds(pgio);
1412 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1415 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1416 struct nfs_page *req, u64 wb_size)
1418 WARN_ON_ONCE(pgio->pg_lseg != NULL);
1420 if (req->wb_offset != req->wb_pgbase) {
1421 nfs_pageio_reset_write_mds(pgio);
1425 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1431 /* If no lseg, fall back to write through mds */
1432 if (pgio->pg_lseg == NULL)
1433 nfs_pageio_reset_write_mds(pgio);
1435 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1438 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1439 struct nfs_page *req)
1441 if (pgio->pg_lseg == NULL)
1442 return nfs_generic_pg_test(pgio, prev, req);
1445 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1446 * Note that this test makes several assumptions:
1447 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1448 * is known to lie within the range.
1449 * - that the nfs_page being tested is known to be contiguous with the
1450 * previous nfs_page.
1451 * - Layout ranges are page aligned, so we only have to test the
1452 * start offset of the request.
1454 * Please also note that 'end_offset' is actually the offset of the
1455 * first byte that lies outside the pnfs_layout_range. FIXME?
1458 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1459 pgio->pg_lseg->pls_range.length);
1461 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1463 int pnfs_write_done_resend_to_mds(struct inode *inode,
1464 struct list_head *head,
1465 const struct nfs_pgio_completion_ops *compl_ops,
1466 struct nfs_direct_req *dreq)
1468 struct nfs_pageio_descriptor pgio;
1471 /* Resend all requests through the MDS */
1472 nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, true, compl_ops);
1473 pgio.pg_dreq = dreq;
1474 while (!list_empty(head)) {
1475 struct nfs_page *req = nfs_list_entry(head->next);
1477 nfs_list_remove_request(req);
1478 if (!nfs_pageio_add_request(&pgio, req))
1479 nfs_list_add_request(req, &failed);
1481 nfs_pageio_complete(&pgio);
1483 if (!list_empty(&failed)) {
1484 /* For some reason our attempt to resend pages. Mark the
1485 * overall send request as having failed, and let
1486 * nfs_writeback_release_full deal with the error.
1488 list_move(&failed, head);
1493 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1495 static void pnfs_ld_handle_write_error(struct nfs_pgio_data *data)
1497 struct nfs_pgio_header *hdr = data->header;
1499 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1500 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1501 PNFS_LAYOUTRET_ON_ERROR) {
1502 pnfs_return_layout(hdr->inode);
1504 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1505 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1507 hdr->completion_ops,
1512 * Called by non rpc-based layout drivers
1514 void pnfs_ld_write_done(struct nfs_pgio_data *data)
1516 struct nfs_pgio_header *hdr = data->header;
1518 trace_nfs4_pnfs_write(data, hdr->pnfs_error);
1519 if (!hdr->pnfs_error) {
1520 pnfs_set_layoutcommit(data);
1521 hdr->mds_ops->rpc_call_done(&data->task, data);
1523 pnfs_ld_handle_write_error(data);
1524 hdr->mds_ops->rpc_release(data);
1526 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1529 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1530 struct nfs_pgio_data *data)
1532 struct nfs_pgio_header *hdr = data->header;
1534 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1535 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1536 nfs_pageio_reset_write_mds(desc);
1537 desc->pg_recoalesce = 1;
1539 nfs_pgio_data_release(data);
1542 static enum pnfs_try_status
1543 pnfs_try_to_write_data(struct nfs_pgio_data *wdata,
1544 const struct rpc_call_ops *call_ops,
1545 struct pnfs_layout_segment *lseg,
1548 struct nfs_pgio_header *hdr = wdata->header;
1549 struct inode *inode = hdr->inode;
1550 enum pnfs_try_status trypnfs;
1551 struct nfs_server *nfss = NFS_SERVER(inode);
1553 hdr->mds_ops = call_ops;
1555 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1556 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1557 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1558 if (trypnfs != PNFS_NOT_ATTEMPTED)
1559 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1560 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1565 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1567 struct nfs_pgio_data *data;
1568 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1569 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1571 desc->pg_lseg = NULL;
1572 while (!list_empty(head)) {
1573 enum pnfs_try_status trypnfs;
1575 data = list_first_entry(head, struct nfs_pgio_data, list);
1576 list_del_init(&data->list);
1578 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1579 if (trypnfs == PNFS_NOT_ATTEMPTED)
1580 pnfs_write_through_mds(desc, data);
1582 pnfs_put_lseg(lseg);
1585 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1587 pnfs_put_lseg(hdr->lseg);
1588 nfs_rw_header_free(hdr);
1590 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1593 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1595 struct nfs_rw_header *whdr;
1596 struct nfs_pgio_header *hdr;
1599 whdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1601 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1602 pnfs_put_lseg(desc->pg_lseg);
1603 desc->pg_lseg = NULL;
1606 hdr = &whdr->header;
1607 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1608 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1609 atomic_inc(&hdr->refcnt);
1610 ret = nfs_generic_flush(desc, hdr);
1612 pnfs_put_lseg(desc->pg_lseg);
1613 desc->pg_lseg = NULL;
1615 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1616 if (atomic_dec_and_test(&hdr->refcnt))
1617 hdr->completion_ops->completion(hdr);
1620 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1622 int pnfs_read_done_resend_to_mds(struct inode *inode,
1623 struct list_head *head,
1624 const struct nfs_pgio_completion_ops *compl_ops,
1625 struct nfs_direct_req *dreq)
1627 struct nfs_pageio_descriptor pgio;
1630 /* Resend all requests through the MDS */
1631 nfs_pageio_init_read(&pgio, inode, true, compl_ops);
1632 pgio.pg_dreq = dreq;
1633 while (!list_empty(head)) {
1634 struct nfs_page *req = nfs_list_entry(head->next);
1636 nfs_list_remove_request(req);
1637 if (!nfs_pageio_add_request(&pgio, req))
1638 nfs_list_add_request(req, &failed);
1640 nfs_pageio_complete(&pgio);
1642 if (!list_empty(&failed)) {
1643 list_move(&failed, head);
1648 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1650 static void pnfs_ld_handle_read_error(struct nfs_pgio_data *data)
1652 struct nfs_pgio_header *hdr = data->header;
1654 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1655 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1656 PNFS_LAYOUTRET_ON_ERROR) {
1657 pnfs_return_layout(hdr->inode);
1659 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1660 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1662 hdr->completion_ops,
1667 * Called by non rpc-based layout drivers
1669 void pnfs_ld_read_done(struct nfs_pgio_data *data)
1671 struct nfs_pgio_header *hdr = data->header;
1673 trace_nfs4_pnfs_read(data, hdr->pnfs_error);
1674 if (likely(!hdr->pnfs_error)) {
1675 __nfs4_read_done_cb(data);
1676 hdr->mds_ops->rpc_call_done(&data->task, data);
1678 pnfs_ld_handle_read_error(data);
1679 hdr->mds_ops->rpc_release(data);
1681 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1684 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1685 struct nfs_pgio_data *data)
1687 struct nfs_pgio_header *hdr = data->header;
1689 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1690 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1691 nfs_pageio_reset_read_mds(desc);
1692 desc->pg_recoalesce = 1;
1694 nfs_pgio_data_release(data);
1698 * Call the appropriate parallel I/O subsystem read function.
1700 static enum pnfs_try_status
1701 pnfs_try_to_read_data(struct nfs_pgio_data *rdata,
1702 const struct rpc_call_ops *call_ops,
1703 struct pnfs_layout_segment *lseg)
1705 struct nfs_pgio_header *hdr = rdata->header;
1706 struct inode *inode = hdr->inode;
1707 struct nfs_server *nfss = NFS_SERVER(inode);
1708 enum pnfs_try_status trypnfs;
1710 hdr->mds_ops = call_ops;
1712 dprintk("%s: Reading ino:%lu %u@%llu\n",
1713 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1715 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1716 if (trypnfs != PNFS_NOT_ATTEMPTED)
1717 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1718 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1723 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1725 struct nfs_pgio_data *data;
1726 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1727 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1729 desc->pg_lseg = NULL;
1730 while (!list_empty(head)) {
1731 enum pnfs_try_status trypnfs;
1733 data = list_first_entry(head, struct nfs_pgio_data, list);
1734 list_del_init(&data->list);
1736 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1737 if (trypnfs == PNFS_NOT_ATTEMPTED)
1738 pnfs_read_through_mds(desc, data);
1740 pnfs_put_lseg(lseg);
1743 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1745 pnfs_put_lseg(hdr->lseg);
1746 nfs_rw_header_free(hdr);
1748 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1751 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1753 struct nfs_rw_header *rhdr;
1754 struct nfs_pgio_header *hdr;
1757 rhdr = nfs_rw_header_alloc(desc->pg_rw_ops);
1759 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1761 pnfs_put_lseg(desc->pg_lseg);
1762 desc->pg_lseg = NULL;
1765 hdr = &rhdr->header;
1766 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1767 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1768 atomic_inc(&hdr->refcnt);
1769 ret = nfs_generic_pagein(desc, hdr);
1771 pnfs_put_lseg(desc->pg_lseg);
1772 desc->pg_lseg = NULL;
1774 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1775 if (atomic_dec_and_test(&hdr->refcnt))
1776 hdr->completion_ops->completion(hdr);
1779 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1781 static void pnfs_clear_layoutcommitting(struct inode *inode)
1783 unsigned long *bitlock = &NFS_I(inode)->flags;
1785 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1786 smp_mb__after_clear_bit();
1787 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1791 * There can be multiple RW segments.
1793 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1795 struct pnfs_layout_segment *lseg;
1797 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1798 if (lseg->pls_range.iomode == IOMODE_RW &&
1799 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1800 list_add(&lseg->pls_lc_list, listp);
1804 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1806 struct pnfs_layout_segment *lseg, *tmp;
1808 /* Matched by references in pnfs_set_layoutcommit */
1809 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1810 list_del_init(&lseg->pls_lc_list);
1811 pnfs_put_lseg(lseg);
1814 pnfs_clear_layoutcommitting(inode);
1817 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1819 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1821 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1824 pnfs_set_layoutcommit(struct nfs_pgio_data *wdata)
1826 struct nfs_pgio_header *hdr = wdata->header;
1827 struct inode *inode = hdr->inode;
1828 struct nfs_inode *nfsi = NFS_I(inode);
1829 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1830 bool mark_as_dirty = false;
1832 spin_lock(&inode->i_lock);
1833 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1834 mark_as_dirty = true;
1835 dprintk("%s: Set layoutcommit for inode %lu ",
1836 __func__, inode->i_ino);
1838 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1839 /* references matched in nfs4_layoutcommit_release */
1840 pnfs_get_lseg(hdr->lseg);
1842 if (end_pos > nfsi->layout->plh_lwb)
1843 nfsi->layout->plh_lwb = end_pos;
1844 spin_unlock(&inode->i_lock);
1845 dprintk("%s: lseg %p end_pos %llu\n",
1846 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1848 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1849 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1851 mark_inode_dirty_sync(inode);
1853 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1855 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1857 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1859 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1860 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1861 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1865 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1866 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1867 * data to disk to allow the server to recover the data if it crashes.
1868 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1869 * is off, and a COMMIT is sent to a data server, or
1870 * if WRITEs to a data server return NFS_DATA_SYNC.
1873 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1875 struct nfs4_layoutcommit_data *data;
1876 struct nfs_inode *nfsi = NFS_I(inode);
1880 if (!pnfs_layoutcommit_outstanding(inode))
1883 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1886 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1889 status = wait_on_bit_lock(&nfsi->flags,
1890 NFS_INO_LAYOUTCOMMITTING,
1891 nfs_wait_bit_killable,
1898 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1899 data = kzalloc(sizeof(*data), GFP_NOFS);
1901 goto clear_layoutcommitting;
1904 spin_lock(&inode->i_lock);
1905 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1908 INIT_LIST_HEAD(&data->lseg_list);
1909 pnfs_list_write_lseg(inode, &data->lseg_list);
1911 end_pos = nfsi->layout->plh_lwb;
1912 nfsi->layout->plh_lwb = 0;
1914 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1915 spin_unlock(&inode->i_lock);
1917 data->args.inode = inode;
1918 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1919 nfs_fattr_init(&data->fattr);
1920 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1921 data->res.fattr = &data->fattr;
1922 data->args.lastbytewritten = end_pos - 1;
1923 data->res.server = NFS_SERVER(inode);
1925 status = nfs4_proc_layoutcommit(data, sync);
1928 mark_inode_dirty_sync(inode);
1929 dprintk("<-- %s status %d\n", __func__, status);
1932 spin_unlock(&inode->i_lock);
1934 clear_layoutcommitting:
1935 pnfs_clear_layoutcommitting(inode);
1939 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1941 struct nfs4_threshold *thp;
1943 thp = kzalloc(sizeof(*thp), GFP_NOFS);
1945 dprintk("%s mdsthreshold allocation failed\n", __func__);