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);
180 static struct pnfs_layout_hdr *
181 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
183 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
184 return ld->alloc_layout_hdr ? ld->alloc_layout_hdr(ino, gfp_flags) :
185 kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
189 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
191 struct pnfs_layoutdriver_type *ld = NFS_SERVER(lo->plh_inode)->pnfs_curr_ld;
192 return ld->alloc_layout_hdr ? ld->free_layout_hdr(lo) : kfree(lo);
196 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
198 dprintk("%s: freeing layout cache %p\n", __func__, lo);
199 BUG_ON(!list_empty(&lo->plh_layouts));
200 NFS_I(lo->plh_inode)->layout = NULL;
201 pnfs_free_layout_hdr(lo);
205 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
207 if (atomic_dec_and_test(&lo->plh_refcount))
208 destroy_layout_hdr(lo);
212 put_layout_hdr(struct pnfs_layout_hdr *lo)
214 struct inode *inode = lo->plh_inode;
216 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
217 destroy_layout_hdr(lo);
218 spin_unlock(&inode->i_lock);
223 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
225 INIT_LIST_HEAD(&lseg->pls_list);
226 atomic_set(&lseg->pls_refcount, 1);
228 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
229 lseg->pls_layout = lo;
232 static void free_lseg(struct pnfs_layout_segment *lseg)
234 struct inode *ino = lseg->pls_layout->plh_inode;
236 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
237 /* Matched by get_layout_hdr in pnfs_insert_layout */
238 put_layout_hdr(NFS_I(ino)->layout);
242 put_lseg_common(struct pnfs_layout_segment *lseg)
244 struct inode *inode = lseg->pls_layout->plh_inode;
246 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
247 list_del_init(&lseg->pls_list);
248 if (list_empty(&lseg->pls_layout->plh_segs)) {
249 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
250 /* Matched by initial refcount set in alloc_init_layout_hdr */
251 put_layout_hdr_locked(lseg->pls_layout);
253 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
257 put_lseg(struct pnfs_layout_segment *lseg)
264 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
265 atomic_read(&lseg->pls_refcount),
266 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
267 inode = lseg->pls_layout->plh_inode;
268 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
271 put_lseg_common(lseg);
272 list_add(&lseg->pls_list, &free_me);
273 spin_unlock(&inode->i_lock);
274 pnfs_free_lseg_list(&free_me);
277 EXPORT_SYMBOL_GPL(put_lseg);
280 end_offset(u64 start, u64 len)
285 return end >= start ? end : NFS4_MAX_UINT64;
288 /* last octet in a range */
290 last_byte_offset(u64 start, u64 len)
296 return end > start ? end - 1 : NFS4_MAX_UINT64;
300 * is l2 fully contained in l1?
302 * [----------------------------------)
307 lo_seg_contained(struct pnfs_layout_range *l1,
308 struct pnfs_layout_range *l2)
310 u64 start1 = l1->offset;
311 u64 end1 = end_offset(start1, l1->length);
312 u64 start2 = l2->offset;
313 u64 end2 = end_offset(start2, l2->length);
315 return (start1 <= start2) && (end1 >= end2);
319 * is l1 and l2 intersecting?
321 * [----------------------------------)
326 lo_seg_intersecting(struct pnfs_layout_range *l1,
327 struct pnfs_layout_range *l2)
329 u64 start1 = l1->offset;
330 u64 end1 = end_offset(start1, l1->length);
331 u64 start2 = l2->offset;
332 u64 end2 = end_offset(start2, l2->length);
334 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
335 (end2 == NFS4_MAX_UINT64 || end2 > start1);
339 should_free_lseg(struct pnfs_layout_range *lseg_range,
340 struct pnfs_layout_range *recall_range)
342 return (recall_range->iomode == IOMODE_ANY ||
343 lseg_range->iomode == recall_range->iomode) &&
344 lo_seg_intersecting(lseg_range, recall_range);
347 /* Returns 1 if lseg is removed from list, 0 otherwise */
348 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
349 struct list_head *tmp_list)
353 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
354 /* Remove the reference keeping the lseg in the
355 * list. It will now be removed when all
356 * outstanding io is finished.
358 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
359 atomic_read(&lseg->pls_refcount));
360 if (atomic_dec_and_test(&lseg->pls_refcount)) {
361 put_lseg_common(lseg);
362 list_add(&lseg->pls_list, tmp_list);
369 /* Returns count of number of matching invalid lsegs remaining in list
373 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
374 struct list_head *tmp_list,
375 struct pnfs_layout_range *recall_range)
377 struct pnfs_layout_segment *lseg, *next;
378 int invalid = 0, removed = 0;
380 dprintk("%s:Begin lo %p\n", __func__, lo);
382 if (list_empty(&lo->plh_segs)) {
383 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
384 put_layout_hdr_locked(lo);
387 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
389 should_free_lseg(&lseg->pls_range, recall_range)) {
390 dprintk("%s: freeing lseg %p iomode %d "
391 "offset %llu length %llu\n", __func__,
392 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
393 lseg->pls_range.length);
395 removed += mark_lseg_invalid(lseg, tmp_list);
397 dprintk("%s:Return %i\n", __func__, invalid - removed);
398 return invalid - removed;
401 /* note free_me must contain lsegs from a single layout_hdr */
403 pnfs_free_lseg_list(struct list_head *free_me)
405 struct pnfs_layout_segment *lseg, *tmp;
406 struct pnfs_layout_hdr *lo;
408 if (list_empty(free_me))
411 lo = list_first_entry(free_me, struct pnfs_layout_segment,
412 pls_list)->pls_layout;
414 if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
415 struct nfs_client *clp;
417 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
418 spin_lock(&clp->cl_lock);
419 list_del_init(&lo->plh_layouts);
420 spin_unlock(&clp->cl_lock);
422 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
423 list_del(&lseg->pls_list);
429 pnfs_destroy_layout(struct nfs_inode *nfsi)
431 struct pnfs_layout_hdr *lo;
434 spin_lock(&nfsi->vfs_inode.i_lock);
437 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
438 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
440 spin_unlock(&nfsi->vfs_inode.i_lock);
441 pnfs_free_lseg_list(&tmp_list);
445 * Called by the state manger to remove all layouts established under an
449 pnfs_destroy_all_layouts(struct nfs_client *clp)
451 struct nfs_server *server;
452 struct pnfs_layout_hdr *lo;
455 spin_lock(&clp->cl_lock);
457 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
458 if (!list_empty(&server->layouts))
459 list_splice_init(&server->layouts, &tmp_list);
462 spin_unlock(&clp->cl_lock);
464 while (!list_empty(&tmp_list)) {
465 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
467 dprintk("%s freeing layout for inode %lu\n", __func__,
468 lo->plh_inode->i_ino);
469 list_del_init(&lo->plh_layouts);
470 pnfs_destroy_layout(NFS_I(lo->plh_inode));
474 /* update lo->plh_stateid with new if is more recent */
476 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
481 oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
482 newseq = be32_to_cpu(new->stateid.seqid);
483 if ((int)(newseq - oldseq) > 0) {
484 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
485 if (update_barrier) {
486 u32 new_barrier = be32_to_cpu(new->stateid.seqid);
488 if ((int)(new_barrier - lo->plh_barrier))
489 lo->plh_barrier = new_barrier;
491 /* Because of wraparound, we want to keep the barrier
492 * "close" to the current seqids. It needs to be
493 * within 2**31 to count as "behind", so if it
494 * gets too near that limit, give us a litle leeway
495 * and bring it to within 2**30.
496 * NOTE - and yes, this is all unsigned arithmetic.
498 if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
499 lo->plh_barrier = newseq - (1 << 30);
504 /* lget is set to 1 if called from inside send_layoutget call chain */
506 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
510 (int)(lo->plh_barrier - be32_to_cpu(stateid->stateid.seqid)) >= 0)
512 return lo->plh_block_lgets ||
513 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
514 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
515 (list_empty(&lo->plh_segs) &&
516 (atomic_read(&lo->plh_outstanding) > lget));
520 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
521 struct nfs4_state *open_state)
525 dprintk("--> %s\n", __func__);
526 spin_lock(&lo->plh_inode->i_lock);
527 if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
529 } else if (list_empty(&lo->plh_segs)) {
533 seq = read_seqbegin(&open_state->seqlock);
534 memcpy(dst->data, open_state->stateid.data,
535 sizeof(open_state->stateid.data));
536 } while (read_seqretry(&open_state->seqlock, seq));
538 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
539 spin_unlock(&lo->plh_inode->i_lock);
540 dprintk("<-- %s\n", __func__);
545 * Get layout from server.
546 * for now, assume that whole file layouts are requested.
548 * arg->length: all ones
550 static struct pnfs_layout_segment *
551 send_layoutget(struct pnfs_layout_hdr *lo,
552 struct nfs_open_context *ctx,
553 struct pnfs_layout_range *range,
556 struct inode *ino = lo->plh_inode;
557 struct nfs_server *server = NFS_SERVER(ino);
558 struct nfs4_layoutget *lgp;
559 struct pnfs_layout_segment *lseg = NULL;
560 struct page **pages = NULL;
562 u32 max_resp_sz, max_pages;
564 dprintk("--> %s\n", __func__);
567 lgp = kzalloc(sizeof(*lgp), gfp_flags);
571 /* allocate pages for xdr post processing */
572 max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
573 max_pages = max_resp_sz >> PAGE_SHIFT;
575 pages = kzalloc(max_pages * sizeof(struct page *), gfp_flags);
579 for (i = 0; i < max_pages; i++) {
580 pages[i] = alloc_page(gfp_flags);
585 lgp->args.minlength = PAGE_CACHE_SIZE;
586 if (lgp->args.minlength > range->length)
587 lgp->args.minlength = range->length;
588 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
589 lgp->args.range = *range;
590 lgp->args.type = server->pnfs_curr_ld->id;
591 lgp->args.inode = ino;
592 lgp->args.ctx = get_nfs_open_context(ctx);
593 lgp->args.layout.pages = pages;
594 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
596 lgp->gfp_flags = gfp_flags;
598 /* Synchronously retrieve layout information from server and
601 nfs4_proc_layoutget(lgp);
603 /* remember that LAYOUTGET failed and suspend trying */
604 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
608 for (i = 0; i < max_pages; i++)
609 __free_page(pages[i]);
615 /* free any allocated xdr pages, lgp as it's not used */
617 for (i = 0; i < max_pages; i++) {
620 __free_page(pages[i]);
628 /* Initiates a LAYOUTRETURN(FILE) */
630 _pnfs_return_layout(struct inode *ino)
632 struct pnfs_layout_hdr *lo = NULL;
633 struct nfs_inode *nfsi = NFS_I(ino);
635 struct nfs4_layoutreturn *lrp;
636 nfs4_stateid stateid;
639 dprintk("--> %s\n", __func__);
641 spin_lock(&ino->i_lock);
644 spin_unlock(&ino->i_lock);
645 dprintk("%s: no layout to return\n", __func__);
648 stateid = nfsi->layout->plh_stateid;
649 /* Reference matched in nfs4_layoutreturn_release */
651 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
652 lo->plh_block_lgets++;
653 spin_unlock(&ino->i_lock);
654 pnfs_free_lseg_list(&tmp_list);
656 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
658 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
659 if (unlikely(lrp == NULL)) {
661 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
662 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
667 lrp->args.stateid = stateid;
668 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
669 lrp->args.inode = ino;
670 lrp->clp = NFS_SERVER(ino)->nfs_client;
672 status = nfs4_proc_layoutreturn(lrp);
674 dprintk("<-- %s status: %d\n", __func__, status);
678 bool pnfs_roc(struct inode *ino)
680 struct pnfs_layout_hdr *lo;
681 struct pnfs_layout_segment *lseg, *tmp;
685 spin_lock(&ino->i_lock);
686 lo = NFS_I(ino)->layout;
687 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
688 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
690 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
691 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
692 mark_lseg_invalid(lseg, &tmp_list);
697 lo->plh_block_lgets++;
698 get_layout_hdr(lo); /* matched in pnfs_roc_release */
699 spin_unlock(&ino->i_lock);
700 pnfs_free_lseg_list(&tmp_list);
704 spin_unlock(&ino->i_lock);
708 void pnfs_roc_release(struct inode *ino)
710 struct pnfs_layout_hdr *lo;
712 spin_lock(&ino->i_lock);
713 lo = NFS_I(ino)->layout;
714 lo->plh_block_lgets--;
715 put_layout_hdr_locked(lo);
716 spin_unlock(&ino->i_lock);
719 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
721 struct pnfs_layout_hdr *lo;
723 spin_lock(&ino->i_lock);
724 lo = NFS_I(ino)->layout;
725 if ((int)(barrier - lo->plh_barrier) > 0)
726 lo->plh_barrier = barrier;
727 spin_unlock(&ino->i_lock);
730 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
732 struct nfs_inode *nfsi = NFS_I(ino);
733 struct pnfs_layout_segment *lseg;
736 spin_lock(&ino->i_lock);
737 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
738 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
743 struct pnfs_layout_hdr *lo = nfsi->layout;
744 u32 current_seqid = be32_to_cpu(lo->plh_stateid.stateid.seqid);
746 /* Since close does not return a layout stateid for use as
747 * a barrier, we choose the worst-case barrier.
749 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
751 spin_unlock(&ino->i_lock);
756 * Compare two layout segments for sorting into layout cache.
757 * We want to preferentially return RW over RO layouts, so ensure those
761 cmp_layout(struct pnfs_layout_range *l1,
762 struct pnfs_layout_range *l2)
766 /* high offset > low offset */
767 d = l1->offset - l2->offset;
771 /* short length > long length */
772 d = l2->length - l1->length;
776 /* read > read/write */
777 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
781 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
782 struct pnfs_layout_segment *lseg)
784 struct pnfs_layout_segment *lp;
786 dprintk("%s:Begin\n", __func__);
788 assert_spin_locked(&lo->plh_inode->i_lock);
789 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
790 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
792 list_add_tail(&lseg->pls_list, &lp->pls_list);
793 dprintk("%s: inserted lseg %p "
794 "iomode %d offset %llu length %llu before "
795 "lp %p iomode %d offset %llu length %llu\n",
796 __func__, lseg, lseg->pls_range.iomode,
797 lseg->pls_range.offset, lseg->pls_range.length,
798 lp, lp->pls_range.iomode, lp->pls_range.offset,
799 lp->pls_range.length);
802 list_add_tail(&lseg->pls_list, &lo->plh_segs);
803 dprintk("%s: inserted lseg %p "
804 "iomode %d offset %llu length %llu at tail\n",
805 __func__, lseg, lseg->pls_range.iomode,
806 lseg->pls_range.offset, lseg->pls_range.length);
810 dprintk("%s:Return\n", __func__);
813 static struct pnfs_layout_hdr *
814 alloc_init_layout_hdr(struct inode *ino, gfp_t gfp_flags)
816 struct pnfs_layout_hdr *lo;
818 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
821 atomic_set(&lo->plh_refcount, 1);
822 INIT_LIST_HEAD(&lo->plh_layouts);
823 INIT_LIST_HEAD(&lo->plh_segs);
824 INIT_LIST_HEAD(&lo->plh_bulk_recall);
829 static struct pnfs_layout_hdr *
830 pnfs_find_alloc_layout(struct inode *ino, gfp_t gfp_flags)
832 struct nfs_inode *nfsi = NFS_I(ino);
833 struct pnfs_layout_hdr *new = NULL;
835 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
837 assert_spin_locked(&ino->i_lock);
839 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
844 spin_unlock(&ino->i_lock);
845 new = alloc_init_layout_hdr(ino, gfp_flags);
846 spin_lock(&ino->i_lock);
848 if (likely(nfsi->layout == NULL)) /* Won the race? */
851 pnfs_free_layout_hdr(new);
856 * iomode matching rules:
867 is_matching_lseg(struct pnfs_layout_range *ls_range,
868 struct pnfs_layout_range *range)
870 struct pnfs_layout_range range1;
872 if ((range->iomode == IOMODE_RW &&
873 ls_range->iomode != IOMODE_RW) ||
874 !lo_seg_intersecting(ls_range, range))
877 /* range1 covers only the first byte in the range */
880 return lo_seg_contained(ls_range, &range1);
884 * lookup range in layout
886 static struct pnfs_layout_segment *
887 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
888 struct pnfs_layout_range *range)
890 struct pnfs_layout_segment *lseg, *ret = NULL;
892 dprintk("%s:Begin\n", __func__);
894 assert_spin_locked(&lo->plh_inode->i_lock);
895 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
896 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
897 is_matching_lseg(&lseg->pls_range, range)) {
898 ret = get_lseg(lseg);
901 if (lseg->pls_range.offset > range->offset)
905 dprintk("%s:Return lseg %p ref %d\n",
906 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
911 * Layout segment is retreived from the server if not cached.
912 * The appropriate layout segment is referenced and returned to the caller.
914 static struct pnfs_layout_segment *
915 pnfs_update_layout(struct inode *ino,
916 struct nfs_open_context *ctx,
919 enum pnfs_iomode iomode,
922 struct pnfs_layout_range arg = {
928 struct nfs_inode *nfsi = NFS_I(ino);
929 struct nfs_server *server = NFS_SERVER(ino);
930 struct nfs_client *clp = server->nfs_client;
931 struct pnfs_layout_hdr *lo;
932 struct pnfs_layout_segment *lseg = NULL;
935 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
937 spin_lock(&ino->i_lock);
938 lo = pnfs_find_alloc_layout(ino, gfp_flags);
940 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
944 /* Do we even need to bother with this? */
945 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
946 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
947 dprintk("%s matches recall, use MDS\n", __func__);
951 /* if LAYOUTGET already failed once we don't try again */
952 if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
955 /* Check to see if the layout for the given range already exists */
956 lseg = pnfs_find_lseg(lo, &arg);
960 if (pnfs_layoutgets_blocked(lo, NULL, 0))
962 atomic_inc(&lo->plh_outstanding);
965 if (list_empty(&lo->plh_segs))
967 spin_unlock(&ino->i_lock);
969 /* The lo must be on the clp list if there is any
970 * chance of a CB_LAYOUTRECALL(FILE) coming in.
972 spin_lock(&clp->cl_lock);
973 BUG_ON(!list_empty(&lo->plh_layouts));
974 list_add_tail(&lo->plh_layouts, &server->layouts);
975 spin_unlock(&clp->cl_lock);
978 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
980 arg.offset -= pg_offset;
981 arg.length += pg_offset;
983 arg.length = PAGE_CACHE_ALIGN(arg.length);
985 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
986 if (!lseg && first) {
987 spin_lock(&clp->cl_lock);
988 list_del_init(&lo->plh_layouts);
989 spin_unlock(&clp->cl_lock);
991 atomic_dec(&lo->plh_outstanding);
994 dprintk("%s end, state 0x%lx lseg %p\n", __func__,
995 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
998 spin_unlock(&ino->i_lock);
1003 pnfs_layout_process(struct nfs4_layoutget *lgp)
1005 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1006 struct nfs4_layoutget_res *res = &lgp->res;
1007 struct pnfs_layout_segment *lseg;
1008 struct inode *ino = lo->plh_inode;
1009 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
1012 /* Inject layout blob into I/O device driver */
1013 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1014 if (!lseg || IS_ERR(lseg)) {
1018 status = PTR_ERR(lseg);
1019 dprintk("%s: Could not allocate layout: error %d\n",
1024 spin_lock(&ino->i_lock);
1025 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
1026 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1027 dprintk("%s forget reply due to recall\n", __func__);
1028 goto out_forget_reply;
1031 if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1032 dprintk("%s forget reply due to state\n", __func__);
1033 goto out_forget_reply;
1035 init_lseg(lo, lseg);
1036 lseg->pls_range = res->range;
1037 *lgp->lsegpp = get_lseg(lseg);
1038 pnfs_insert_layout(lo, lseg);
1040 if (res->return_on_close) {
1041 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1042 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1045 /* Done processing layoutget. Set the layout stateid */
1046 pnfs_set_layout_stateid(lo, &res->stateid, false);
1047 spin_unlock(&ino->i_lock);
1052 spin_unlock(&ino->i_lock);
1053 lseg->pls_layout = lo;
1054 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1059 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1061 BUG_ON(pgio->pg_lseg != NULL);
1063 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1069 /* If no lseg, fall back to read through mds */
1070 if (pgio->pg_lseg == NULL)
1071 nfs_pageio_init_read_mds(pgio, pgio->pg_inode);
1074 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1077 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1079 BUG_ON(pgio->pg_lseg != NULL);
1081 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1087 /* If no lseg, fall back to write through mds */
1088 if (pgio->pg_lseg == NULL)
1089 nfs_pageio_init_write_mds(pgio, pgio->pg_inode, pgio->pg_ioflags);
1091 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1094 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
1096 struct nfs_server *server = NFS_SERVER(inode);
1097 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1101 nfs_pageio_init(pgio, inode, ld->pg_read_ops, server->rsize, 0);
1106 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode, int ioflags)
1108 struct nfs_server *server = NFS_SERVER(inode);
1109 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1113 nfs_pageio_init(pgio, inode, ld->pg_write_ops, server->wsize, ioflags);
1118 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1119 struct nfs_page *req)
1121 if (pgio->pg_lseg == NULL)
1122 return nfs_generic_pg_test(pgio, prev, req);
1125 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1126 * Note that this test makes several assumptions:
1127 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1128 * is known to lie within the range.
1129 * - that the nfs_page being tested is known to be contiguous with the
1130 * previous nfs_page.
1131 * - Layout ranges are page aligned, so we only have to test the
1132 * start offset of the request.
1134 * Please also note that 'end_offset' is actually the offset of the
1135 * first byte that lies outside the pnfs_layout_range. FIXME?
1138 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1139 pgio->pg_lseg->pls_range.length);
1141 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1144 * Called by non rpc-based layout drivers
1147 pnfs_ld_write_done(struct nfs_write_data *data)
1151 if (!data->pnfs_error) {
1152 pnfs_set_layoutcommit(data);
1153 data->mds_ops->rpc_call_done(&data->task, data);
1154 data->mds_ops->rpc_release(data);
1158 dprintk("%s: pnfs_error=%d, retry via MDS\n", __func__,
1160 status = nfs_initiate_write(data, NFS_CLIENT(data->inode),
1161 data->mds_ops, NFS_FILE_SYNC);
1162 return status ? : -EAGAIN;
1164 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1166 enum pnfs_try_status
1167 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1168 const struct rpc_call_ops *call_ops, int how)
1170 struct inode *inode = wdata->inode;
1171 enum pnfs_try_status trypnfs;
1172 struct nfs_server *nfss = NFS_SERVER(inode);
1174 wdata->mds_ops = call_ops;
1176 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1177 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1179 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1180 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1181 put_lseg(wdata->lseg);
1184 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1186 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1191 * Called by non rpc-based layout drivers
1194 pnfs_ld_read_done(struct nfs_read_data *data)
1198 if (!data->pnfs_error) {
1199 __nfs4_read_done_cb(data);
1200 data->mds_ops->rpc_call_done(&data->task, data);
1201 data->mds_ops->rpc_release(data);
1205 dprintk("%s: pnfs_error=%d, retry via MDS\n", __func__,
1207 status = nfs_initiate_read(data, NFS_CLIENT(data->inode),
1209 return status ? : -EAGAIN;
1211 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1214 * Call the appropriate parallel I/O subsystem read function.
1216 enum pnfs_try_status
1217 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1218 const struct rpc_call_ops *call_ops)
1220 struct inode *inode = rdata->inode;
1221 struct nfs_server *nfss = NFS_SERVER(inode);
1222 enum pnfs_try_status trypnfs;
1224 rdata->mds_ops = call_ops;
1226 dprintk("%s: Reading ino:%lu %u@%llu\n",
1227 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1229 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1230 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1231 put_lseg(rdata->lseg);
1234 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1236 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1241 * Currently there is only one (whole file) write lseg.
1243 static struct pnfs_layout_segment *pnfs_list_write_lseg(struct inode *inode)
1245 struct pnfs_layout_segment *lseg, *rv = NULL;
1247 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
1248 if (lseg->pls_range.iomode == IOMODE_RW)
1254 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1256 struct nfs_inode *nfsi = NFS_I(wdata->inode);
1257 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1258 bool mark_as_dirty = false;
1260 spin_lock(&nfsi->vfs_inode.i_lock);
1261 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1262 /* references matched in nfs4_layoutcommit_release */
1263 get_lseg(wdata->lseg);
1264 wdata->lseg->pls_lc_cred =
1265 get_rpccred(wdata->args.context->state->owner->so_cred);
1266 mark_as_dirty = true;
1267 dprintk("%s: Set layoutcommit for inode %lu ",
1268 __func__, wdata->inode->i_ino);
1270 if (end_pos > wdata->lseg->pls_end_pos)
1271 wdata->lseg->pls_end_pos = end_pos;
1272 spin_unlock(&nfsi->vfs_inode.i_lock);
1274 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1275 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1277 mark_inode_dirty_sync(wdata->inode);
1279 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1282 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1283 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1284 * data to disk to allow the server to recover the data if it crashes.
1285 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1286 * is off, and a COMMIT is sent to a data server, or
1287 * if WRITEs to a data server return NFS_DATA_SYNC.
1290 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1292 struct nfs4_layoutcommit_data *data;
1293 struct nfs_inode *nfsi = NFS_I(inode);
1294 struct pnfs_layout_segment *lseg;
1295 struct rpc_cred *cred;
1299 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1301 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1304 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1305 data = kzalloc(sizeof(*data), GFP_NOFS);
1307 mark_inode_dirty_sync(inode);
1312 spin_lock(&inode->i_lock);
1313 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1314 spin_unlock(&inode->i_lock);
1319 * Currently only one (whole file) write lseg which is referenced
1320 * in pnfs_set_layoutcommit and will be found.
1322 lseg = pnfs_list_write_lseg(inode);
1324 end_pos = lseg->pls_end_pos;
1325 cred = lseg->pls_lc_cred;
1326 lseg->pls_end_pos = 0;
1327 lseg->pls_lc_cred = NULL;
1329 memcpy(&data->args.stateid.data, nfsi->layout->plh_stateid.data,
1330 sizeof(nfsi->layout->plh_stateid.data));
1331 spin_unlock(&inode->i_lock);
1333 data->args.inode = inode;
1336 nfs_fattr_init(&data->fattr);
1337 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1338 data->res.fattr = &data->fattr;
1339 data->args.lastbytewritten = end_pos - 1;
1340 data->res.server = NFS_SERVER(inode);
1342 status = nfs4_proc_layoutcommit(data, sync);
1344 dprintk("<-- %s status %d\n", __func__, status);