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>
35 #define NFSDBG_FACILITY NFSDBG_PNFS
40 * protects pnfs_modules_tbl.
42 static DEFINE_SPINLOCK(pnfs_spinlock);
45 * pnfs_modules_tbl holds all pnfs modules
47 static LIST_HEAD(pnfs_modules_tbl);
49 /* Return the registered pnfs layout driver module matching given id */
50 static struct pnfs_layoutdriver_type *
51 find_pnfs_driver_locked(u32 id)
53 struct pnfs_layoutdriver_type *local;
55 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
60 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
64 static struct pnfs_layoutdriver_type *
65 find_pnfs_driver(u32 id)
67 struct pnfs_layoutdriver_type *local;
69 spin_lock(&pnfs_spinlock);
70 local = find_pnfs_driver_locked(id);
71 spin_unlock(&pnfs_spinlock);
76 unset_pnfs_layoutdriver(struct nfs_server *nfss)
78 if (nfss->pnfs_curr_ld)
79 module_put(nfss->pnfs_curr_ld->owner);
80 nfss->pnfs_curr_ld = NULL;
84 * Try to set the server's pnfs module to the pnfs layout type specified by id.
85 * Currently only one pNFS layout driver per filesystem is supported.
87 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
90 set_pnfs_layoutdriver(struct nfs_server *server, u32 id)
92 struct pnfs_layoutdriver_type *ld_type = NULL;
96 if (!(server->nfs_client->cl_exchange_flags &
97 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
98 printk(KERN_ERR "%s: id %u cl_exchange_flags 0x%x\n", __func__,
99 id, server->nfs_client->cl_exchange_flags);
102 ld_type = find_pnfs_driver(id);
104 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
105 ld_type = find_pnfs_driver(id);
107 dprintk("%s: No pNFS module found for %u.\n",
112 if (!try_module_get(ld_type->owner)) {
113 dprintk("%s: Could not grab reference on module\n", __func__);
116 server->pnfs_curr_ld = ld_type;
118 dprintk("%s: pNFS module for %u set\n", __func__, id);
122 dprintk("%s: Using NFSv4 I/O\n", __func__);
123 server->pnfs_curr_ld = NULL;
127 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
129 int status = -EINVAL;
130 struct pnfs_layoutdriver_type *tmp;
132 if (ld_type->id == 0) {
133 printk(KERN_ERR "%s id 0 is reserved\n", __func__);
136 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
137 printk(KERN_ERR "%s Layout driver must provide "
138 "alloc_lseg and free_lseg.\n", __func__);
142 spin_lock(&pnfs_spinlock);
143 tmp = find_pnfs_driver_locked(ld_type->id);
145 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
147 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
150 printk(KERN_ERR "%s Module with id %d already loaded!\n",
151 __func__, ld_type->id);
153 spin_unlock(&pnfs_spinlock);
157 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
160 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
162 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
163 spin_lock(&pnfs_spinlock);
164 list_del(&ld_type->pnfs_tblid);
165 spin_unlock(&pnfs_spinlock);
167 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
170 * pNFS client layout cache
173 /* Need to hold i_lock if caller does not already hold reference */
175 get_layout_hdr(struct pnfs_layout_hdr *lo)
177 atomic_inc(&lo->plh_refcount);
181 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
183 dprintk("%s: freeing layout cache %p\n", __func__, lo);
184 BUG_ON(!list_empty(&lo->plh_layouts));
185 NFS_I(lo->plh_inode)->layout = NULL;
190 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
192 if (atomic_dec_and_test(&lo->plh_refcount))
193 destroy_layout_hdr(lo);
197 put_layout_hdr(struct pnfs_layout_hdr *lo)
199 struct inode *inode = lo->plh_inode;
201 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
202 destroy_layout_hdr(lo);
203 spin_unlock(&inode->i_lock);
208 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
210 INIT_LIST_HEAD(&lseg->pls_list);
211 atomic_set(&lseg->pls_refcount, 1);
213 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
214 lseg->pls_layout = lo;
217 static void free_lseg(struct pnfs_layout_segment *lseg)
219 struct inode *ino = lseg->pls_layout->plh_inode;
221 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
222 /* Matched by get_layout_hdr in pnfs_insert_layout */
223 put_layout_hdr(NFS_I(ino)->layout);
227 put_lseg_common(struct pnfs_layout_segment *lseg)
229 struct inode *inode = lseg->pls_layout->plh_inode;
231 BUG_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
232 list_del_init(&lseg->pls_list);
233 if (list_empty(&lseg->pls_layout->plh_segs)) {
234 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
235 /* Matched by initial refcount set in alloc_init_layout_hdr */
236 put_layout_hdr_locked(lseg->pls_layout);
238 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
242 put_lseg(struct pnfs_layout_segment *lseg)
249 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
250 atomic_read(&lseg->pls_refcount),
251 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
252 inode = lseg->pls_layout->plh_inode;
253 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
256 put_lseg_common(lseg);
257 list_add(&lseg->pls_list, &free_me);
258 spin_unlock(&inode->i_lock);
259 pnfs_free_lseg_list(&free_me);
262 EXPORT_SYMBOL_GPL(put_lseg);
265 should_free_lseg(u32 lseg_iomode, u32 recall_iomode)
267 return (recall_iomode == IOMODE_ANY ||
268 lseg_iomode == recall_iomode);
271 /* Returns 1 if lseg is removed from list, 0 otherwise */
272 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
273 struct list_head *tmp_list)
277 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
278 /* Remove the reference keeping the lseg in the
279 * list. It will now be removed when all
280 * outstanding io is finished.
282 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
283 atomic_read(&lseg->pls_refcount));
284 if (atomic_dec_and_test(&lseg->pls_refcount)) {
285 put_lseg_common(lseg);
286 list_add(&lseg->pls_list, tmp_list);
293 /* Returns count of number of matching invalid lsegs remaining in list
297 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
298 struct list_head *tmp_list,
301 struct pnfs_layout_segment *lseg, *next;
302 int invalid = 0, removed = 0;
304 dprintk("%s:Begin lo %p\n", __func__, lo);
306 if (list_empty(&lo->plh_segs)) {
307 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
308 put_layout_hdr_locked(lo);
311 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
312 if (should_free_lseg(lseg->pls_range.iomode, iomode)) {
313 dprintk("%s: freeing lseg %p iomode %d "
314 "offset %llu length %llu\n", __func__,
315 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
316 lseg->pls_range.length);
318 removed += mark_lseg_invalid(lseg, tmp_list);
320 dprintk("%s:Return %i\n", __func__, invalid - removed);
321 return invalid - removed;
324 /* note free_me must contain lsegs from a single layout_hdr */
326 pnfs_free_lseg_list(struct list_head *free_me)
328 struct pnfs_layout_segment *lseg, *tmp;
329 struct pnfs_layout_hdr *lo;
331 if (list_empty(free_me))
334 lo = list_first_entry(free_me, struct pnfs_layout_segment,
335 pls_list)->pls_layout;
337 if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
338 struct nfs_client *clp;
340 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
341 spin_lock(&clp->cl_lock);
342 list_del_init(&lo->plh_layouts);
343 spin_unlock(&clp->cl_lock);
345 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
346 list_del(&lseg->pls_list);
352 pnfs_destroy_layout(struct nfs_inode *nfsi)
354 struct pnfs_layout_hdr *lo;
357 spin_lock(&nfsi->vfs_inode.i_lock);
360 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
361 mark_matching_lsegs_invalid(lo, &tmp_list, IOMODE_ANY);
363 spin_unlock(&nfsi->vfs_inode.i_lock);
364 pnfs_free_lseg_list(&tmp_list);
368 * Called by the state manger to remove all layouts established under an
372 pnfs_destroy_all_layouts(struct nfs_client *clp)
374 struct pnfs_layout_hdr *lo;
377 spin_lock(&clp->cl_lock);
378 list_splice_init(&clp->cl_layouts, &tmp_list);
379 spin_unlock(&clp->cl_lock);
381 while (!list_empty(&tmp_list)) {
382 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
384 dprintk("%s freeing layout for inode %lu\n", __func__,
385 lo->plh_inode->i_ino);
386 list_del_init(&lo->plh_layouts);
387 pnfs_destroy_layout(NFS_I(lo->plh_inode));
391 /* update lo->plh_stateid with new if is more recent */
393 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
398 oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
399 newseq = be32_to_cpu(new->stateid.seqid);
400 if ((int)(newseq - oldseq) > 0) {
401 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
402 if (update_barrier) {
403 u32 new_barrier = be32_to_cpu(new->stateid.seqid);
405 if ((int)(new_barrier - lo->plh_barrier))
406 lo->plh_barrier = new_barrier;
408 /* Because of wraparound, we want to keep the barrier
409 * "close" to the current seqids. It needs to be
410 * within 2**31 to count as "behind", so if it
411 * gets too near that limit, give us a litle leeway
412 * and bring it to within 2**30.
413 * NOTE - and yes, this is all unsigned arithmetic.
415 if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
416 lo->plh_barrier = newseq - (1 << 30);
421 /* lget is set to 1 if called from inside send_layoutget call chain */
423 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
427 (int)(lo->plh_barrier - be32_to_cpu(stateid->stateid.seqid)) >= 0)
429 return lo->plh_block_lgets ||
430 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
431 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
432 (list_empty(&lo->plh_segs) &&
433 (atomic_read(&lo->plh_outstanding) > lget));
437 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
438 struct nfs4_state *open_state)
442 dprintk("--> %s\n", __func__);
443 spin_lock(&lo->plh_inode->i_lock);
444 if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
446 } else if (list_empty(&lo->plh_segs)) {
450 seq = read_seqbegin(&open_state->seqlock);
451 memcpy(dst->data, open_state->stateid.data,
452 sizeof(open_state->stateid.data));
453 } while (read_seqretry(&open_state->seqlock, seq));
455 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
456 spin_unlock(&lo->plh_inode->i_lock);
457 dprintk("<-- %s\n", __func__);
462 * Get layout from server.
463 * for now, assume that whole file layouts are requested.
465 * arg->length: all ones
467 static struct pnfs_layout_segment *
468 send_layoutget(struct pnfs_layout_hdr *lo,
469 struct nfs_open_context *ctx,
472 struct inode *ino = lo->plh_inode;
473 struct nfs_server *server = NFS_SERVER(ino);
474 struct nfs4_layoutget *lgp;
475 struct pnfs_layout_segment *lseg = NULL;
476 struct page **pages = NULL;
478 u32 max_resp_sz, max_pages;
480 dprintk("--> %s\n", __func__);
483 lgp = kzalloc(sizeof(*lgp), GFP_KERNEL);
487 /* allocate pages for xdr post processing */
488 max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
489 max_pages = max_resp_sz >> PAGE_SHIFT;
491 pages = kzalloc(max_pages * sizeof(struct page *), GFP_KERNEL);
495 for (i = 0; i < max_pages; i++) {
496 pages[i] = alloc_page(GFP_KERNEL);
501 lgp->args.minlength = NFS4_MAX_UINT64;
502 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
503 lgp->args.range.iomode = iomode;
504 lgp->args.range.offset = 0;
505 lgp->args.range.length = NFS4_MAX_UINT64;
506 lgp->args.type = server->pnfs_curr_ld->id;
507 lgp->args.inode = ino;
508 lgp->args.ctx = get_nfs_open_context(ctx);
509 lgp->args.layout.pages = pages;
510 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
513 /* Synchronously retrieve layout information from server and
516 nfs4_proc_layoutget(lgp);
518 /* remember that LAYOUTGET failed and suspend trying */
519 set_bit(lo_fail_bit(iomode), &lo->plh_flags);
523 for (i = 0; i < max_pages; i++)
524 __free_page(pages[i]);
530 /* free any allocated xdr pages, lgp as it's not used */
532 for (i = 0; i < max_pages; i++) {
535 __free_page(pages[i]);
543 bool pnfs_roc(struct inode *ino)
545 struct pnfs_layout_hdr *lo;
546 struct pnfs_layout_segment *lseg, *tmp;
550 spin_lock(&ino->i_lock);
551 lo = NFS_I(ino)->layout;
552 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
553 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
555 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
556 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
557 mark_lseg_invalid(lseg, &tmp_list);
562 lo->plh_block_lgets++;
563 get_layout_hdr(lo); /* matched in pnfs_roc_release */
564 spin_unlock(&ino->i_lock);
565 pnfs_free_lseg_list(&tmp_list);
569 spin_unlock(&ino->i_lock);
573 void pnfs_roc_release(struct inode *ino)
575 struct pnfs_layout_hdr *lo;
577 spin_lock(&ino->i_lock);
578 lo = NFS_I(ino)->layout;
579 lo->plh_block_lgets--;
580 put_layout_hdr_locked(lo);
581 spin_unlock(&ino->i_lock);
584 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
586 struct pnfs_layout_hdr *lo;
588 spin_lock(&ino->i_lock);
589 lo = NFS_I(ino)->layout;
590 if ((int)(barrier - lo->plh_barrier) > 0)
591 lo->plh_barrier = barrier;
592 spin_unlock(&ino->i_lock);
595 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
597 struct nfs_inode *nfsi = NFS_I(ino);
598 struct pnfs_layout_segment *lseg;
601 spin_lock(&ino->i_lock);
602 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
603 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
608 struct pnfs_layout_hdr *lo = nfsi->layout;
609 u32 current_seqid = be32_to_cpu(lo->plh_stateid.stateid.seqid);
611 /* Since close does not return a layout stateid for use as
612 * a barrier, we choose the worst-case barrier.
614 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
616 spin_unlock(&ino->i_lock);
621 * Compare two layout segments for sorting into layout cache.
622 * We want to preferentially return RW over RO layouts, so ensure those
626 cmp_layout(u32 iomode1, u32 iomode2)
628 /* read > read/write */
629 return (int)(iomode2 == IOMODE_READ) - (int)(iomode1 == IOMODE_READ);
633 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
634 struct pnfs_layout_segment *lseg)
636 struct pnfs_layout_segment *lp;
639 dprintk("%s:Begin\n", __func__);
641 assert_spin_locked(&lo->plh_inode->i_lock);
642 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
643 if (cmp_layout(lp->pls_range.iomode, lseg->pls_range.iomode) > 0)
645 list_add_tail(&lseg->pls_list, &lp->pls_list);
646 dprintk("%s: inserted lseg %p "
647 "iomode %d offset %llu length %llu before "
648 "lp %p iomode %d offset %llu length %llu\n",
649 __func__, lseg, lseg->pls_range.iomode,
650 lseg->pls_range.offset, lseg->pls_range.length,
651 lp, lp->pls_range.iomode, lp->pls_range.offset,
652 lp->pls_range.length);
657 list_add_tail(&lseg->pls_list, &lo->plh_segs);
658 dprintk("%s: inserted lseg %p "
659 "iomode %d offset %llu length %llu at tail\n",
660 __func__, lseg, lseg->pls_range.iomode,
661 lseg->pls_range.offset, lseg->pls_range.length);
665 dprintk("%s:Return\n", __func__);
668 static struct pnfs_layout_hdr *
669 alloc_init_layout_hdr(struct inode *ino)
671 struct pnfs_layout_hdr *lo;
673 lo = kzalloc(sizeof(struct pnfs_layout_hdr), GFP_KERNEL);
676 atomic_set(&lo->plh_refcount, 1);
677 INIT_LIST_HEAD(&lo->plh_layouts);
678 INIT_LIST_HEAD(&lo->plh_segs);
679 INIT_LIST_HEAD(&lo->plh_bulk_recall);
684 static struct pnfs_layout_hdr *
685 pnfs_find_alloc_layout(struct inode *ino)
687 struct nfs_inode *nfsi = NFS_I(ino);
688 struct pnfs_layout_hdr *new = NULL;
690 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
692 assert_spin_locked(&ino->i_lock);
694 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
699 spin_unlock(&ino->i_lock);
700 new = alloc_init_layout_hdr(ino);
701 spin_lock(&ino->i_lock);
703 if (likely(nfsi->layout == NULL)) /* Won the race? */
711 * iomode matching rules:
722 is_matching_lseg(struct pnfs_layout_segment *lseg, u32 iomode)
724 return (iomode != IOMODE_RW || lseg->pls_range.iomode == IOMODE_RW);
728 * lookup range in layout
730 static struct pnfs_layout_segment *
731 pnfs_find_lseg(struct pnfs_layout_hdr *lo, u32 iomode)
733 struct pnfs_layout_segment *lseg, *ret = NULL;
735 dprintk("%s:Begin\n", __func__);
737 assert_spin_locked(&lo->plh_inode->i_lock);
738 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
739 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
740 is_matching_lseg(lseg, iomode)) {
741 ret = get_lseg(lseg);
744 if (cmp_layout(iomode, lseg->pls_range.iomode) > 0)
748 dprintk("%s:Return lseg %p ref %d\n",
749 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
754 * Layout segment is retreived from the server if not cached.
755 * The appropriate layout segment is referenced and returned to the caller.
757 struct pnfs_layout_segment *
758 pnfs_update_layout(struct inode *ino,
759 struct nfs_open_context *ctx,
760 enum pnfs_iomode iomode)
762 struct nfs_inode *nfsi = NFS_I(ino);
763 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
764 struct pnfs_layout_hdr *lo;
765 struct pnfs_layout_segment *lseg = NULL;
768 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
770 spin_lock(&ino->i_lock);
771 lo = pnfs_find_alloc_layout(ino);
773 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
777 /* Do we even need to bother with this? */
778 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
779 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
780 dprintk("%s matches recall, use MDS\n", __func__);
784 /* if LAYOUTGET already failed once we don't try again */
785 if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
788 /* Check to see if the layout for the given range already exists */
789 lseg = pnfs_find_lseg(lo, iomode);
793 if (pnfs_layoutgets_blocked(lo, NULL, 0))
795 atomic_inc(&lo->plh_outstanding);
798 if (list_empty(&lo->plh_segs))
800 spin_unlock(&ino->i_lock);
802 /* The lo must be on the clp list if there is any
803 * chance of a CB_LAYOUTRECALL(FILE) coming in.
805 spin_lock(&clp->cl_lock);
806 BUG_ON(!list_empty(&lo->plh_layouts));
807 list_add_tail(&lo->plh_layouts, &clp->cl_layouts);
808 spin_unlock(&clp->cl_lock);
811 lseg = send_layoutget(lo, ctx, iomode);
812 if (!lseg && first) {
813 spin_lock(&clp->cl_lock);
814 list_del_init(&lo->plh_layouts);
815 spin_unlock(&clp->cl_lock);
817 atomic_dec(&lo->plh_outstanding);
820 dprintk("%s end, state 0x%lx lseg %p\n", __func__,
821 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
824 spin_unlock(&ino->i_lock);
829 pnfs_layout_process(struct nfs4_layoutget *lgp)
831 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
832 struct nfs4_layoutget_res *res = &lgp->res;
833 struct pnfs_layout_segment *lseg;
834 struct inode *ino = lo->plh_inode;
835 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
838 /* Verify we got what we asked for.
839 * Note that because the xdr parsing only accepts a single
840 * element array, this can fail even if the server is behaving
843 if (lgp->args.range.iomode > res->range.iomode ||
844 res->range.offset != 0 ||
845 res->range.length != NFS4_MAX_UINT64) {
849 /* Inject layout blob into I/O device driver */
850 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res);
851 if (!lseg || IS_ERR(lseg)) {
855 status = PTR_ERR(lseg);
856 dprintk("%s: Could not allocate layout: error %d\n",
861 spin_lock(&ino->i_lock);
862 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
863 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
864 dprintk("%s forget reply due to recall\n", __func__);
865 goto out_forget_reply;
868 if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
869 dprintk("%s forget reply due to state\n", __func__);
870 goto out_forget_reply;
873 lseg->pls_range = res->range;
874 *lgp->lsegpp = get_lseg(lseg);
875 pnfs_insert_layout(lo, lseg);
877 if (res->return_on_close) {
878 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
879 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
882 /* Done processing layoutget. Set the layout stateid */
883 pnfs_set_layout_stateid(lo, &res->stateid, false);
884 spin_unlock(&ino->i_lock);
889 spin_unlock(&ino->i_lock);
890 lseg->pls_layout = lo;
891 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
895 static int pnfs_read_pg_test(struct nfs_pageio_descriptor *pgio,
896 struct nfs_page *prev,
897 struct nfs_page *req)
899 if (pgio->pg_count == prev->wb_bytes) {
900 /* This is first coelesce call for a series of nfs_pages */
901 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
905 return NFS_SERVER(pgio->pg_inode)->pnfs_curr_ld->pg_test(pgio, prev, req);
909 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
911 struct pnfs_layoutdriver_type *ld;
913 ld = NFS_SERVER(inode)->pnfs_curr_ld;
914 pgio->pg_test = (ld && ld->pg_test) ? pnfs_read_pg_test : NULL;
917 static int pnfs_write_pg_test(struct nfs_pageio_descriptor *pgio,
918 struct nfs_page *prev,
919 struct nfs_page *req)
921 if (pgio->pg_count == prev->wb_bytes) {
922 /* This is first coelesce call for a series of nfs_pages */
923 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
927 return NFS_SERVER(pgio->pg_inode)->pnfs_curr_ld->pg_test(pgio, prev, req);
931 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode)
933 struct pnfs_layoutdriver_type *ld;
935 ld = NFS_SERVER(inode)->pnfs_curr_ld;
936 pgio->pg_test = (ld && ld->pg_test) ? pnfs_write_pg_test : NULL;
940 pnfs_try_to_write_data(struct nfs_write_data *wdata,
941 const struct rpc_call_ops *call_ops, int how)
943 struct inode *inode = wdata->inode;
944 enum pnfs_try_status trypnfs;
945 struct nfs_server *nfss = NFS_SERVER(inode);
947 wdata->mds_ops = call_ops;
949 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
950 inode->i_ino, wdata->args.count, wdata->args.offset, how);
952 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
953 if (trypnfs == PNFS_NOT_ATTEMPTED) {
954 put_lseg(wdata->lseg);
957 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
959 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
964 * Call the appropriate parallel I/O subsystem read function.
967 pnfs_try_to_read_data(struct nfs_read_data *rdata,
968 const struct rpc_call_ops *call_ops)
970 struct inode *inode = rdata->inode;
971 struct nfs_server *nfss = NFS_SERVER(inode);
972 enum pnfs_try_status trypnfs;
974 rdata->mds_ops = call_ops;
976 dprintk("%s: Reading ino:%lu %u@%llu\n",
977 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
979 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
980 if (trypnfs == PNFS_NOT_ATTEMPTED) {
981 put_lseg(rdata->lseg);
984 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
986 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
991 * Currently there is only one (whole file) write lseg.
993 static struct pnfs_layout_segment *pnfs_list_write_lseg(struct inode *inode)
995 struct pnfs_layout_segment *lseg, *rv = NULL;
997 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
998 if (lseg->pls_range.iomode == IOMODE_RW)
1004 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1006 struct nfs_inode *nfsi = NFS_I(wdata->inode);
1007 loff_t end_pos = wdata->args.offset + wdata->res.count;
1008 bool mark_as_dirty = false;
1010 spin_lock(&nfsi->vfs_inode.i_lock);
1011 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1012 /* references matched in nfs4_layoutcommit_release */
1013 get_lseg(wdata->lseg);
1014 wdata->lseg->pls_lc_cred =
1015 get_rpccred(wdata->args.context->state->owner->so_cred);
1016 mark_as_dirty = true;
1017 dprintk("%s: Set layoutcommit for inode %lu ",
1018 __func__, wdata->inode->i_ino);
1020 if (end_pos > wdata->lseg->pls_end_pos)
1021 wdata->lseg->pls_end_pos = end_pos;
1022 spin_unlock(&nfsi->vfs_inode.i_lock);
1024 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1025 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1027 mark_inode_dirty_sync(wdata->inode);
1029 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1032 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1033 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1034 * data to disk to allow the server to recover the data if it crashes.
1035 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1036 * is off, and a COMMIT is sent to a data server, or
1037 * if WRITEs to a data server return NFS_DATA_SYNC.
1040 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1042 struct nfs4_layoutcommit_data *data;
1043 struct nfs_inode *nfsi = NFS_I(inode);
1044 struct pnfs_layout_segment *lseg;
1045 struct rpc_cred *cred;
1049 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1051 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1054 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1055 data = kzalloc(sizeof(*data), GFP_NOFS);
1057 mark_inode_dirty_sync(inode);
1062 spin_lock(&inode->i_lock);
1063 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1064 spin_unlock(&inode->i_lock);
1069 * Currently only one (whole file) write lseg which is referenced
1070 * in pnfs_set_layoutcommit and will be found.
1072 lseg = pnfs_list_write_lseg(inode);
1074 end_pos = lseg->pls_end_pos;
1075 cred = lseg->pls_lc_cred;
1076 lseg->pls_end_pos = 0;
1077 lseg->pls_lc_cred = NULL;
1079 memcpy(&data->args.stateid.data, nfsi->layout->plh_stateid.data,
1080 sizeof(nfsi->layout->plh_stateid.data));
1081 spin_unlock(&inode->i_lock);
1083 data->args.inode = inode;
1086 nfs_fattr_init(&data->fattr);
1087 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1088 data->res.fattr = &data->fattr;
1089 data->args.lastbytewritten = end_pos - 1;
1090 data->res.server = NFS_SERVER(inode);
1092 status = nfs4_proc_layoutcommit(data, sync);
1094 dprintk("<-- %s status %d\n", __func__, status);