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 end_offset(u64 start, u64 len)
270 return end >= start ? end : NFS4_MAX_UINT64;
273 /* last octet in a range */
275 last_byte_offset(u64 start, u64 len)
281 return end > start ? end - 1 : NFS4_MAX_UINT64;
285 * is l2 fully contained in l1?
287 * [----------------------------------)
292 lo_seg_contained(struct pnfs_layout_range *l1,
293 struct pnfs_layout_range *l2)
295 u64 start1 = l1->offset;
296 u64 end1 = end_offset(start1, l1->length);
297 u64 start2 = l2->offset;
298 u64 end2 = end_offset(start2, l2->length);
300 return (start1 <= start2) && (end1 >= end2);
304 * is l1 and l2 intersecting?
306 * [----------------------------------)
311 lo_seg_intersecting(struct pnfs_layout_range *l1,
312 struct pnfs_layout_range *l2)
314 u64 start1 = l1->offset;
315 u64 end1 = end_offset(start1, l1->length);
316 u64 start2 = l2->offset;
317 u64 end2 = end_offset(start2, l2->length);
319 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
320 (end2 == NFS4_MAX_UINT64 || end2 > start1);
324 should_free_lseg(u32 lseg_iomode, u32 recall_iomode)
326 return (recall_iomode == IOMODE_ANY ||
327 lseg_iomode == recall_iomode);
330 /* Returns 1 if lseg is removed from list, 0 otherwise */
331 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
332 struct list_head *tmp_list)
336 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
337 /* Remove the reference keeping the lseg in the
338 * list. It will now be removed when all
339 * outstanding io is finished.
341 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
342 atomic_read(&lseg->pls_refcount));
343 if (atomic_dec_and_test(&lseg->pls_refcount)) {
344 put_lseg_common(lseg);
345 list_add(&lseg->pls_list, tmp_list);
352 /* Returns count of number of matching invalid lsegs remaining in list
356 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
357 struct list_head *tmp_list,
360 struct pnfs_layout_segment *lseg, *next;
361 int invalid = 0, removed = 0;
363 dprintk("%s:Begin lo %p\n", __func__, lo);
365 if (list_empty(&lo->plh_segs)) {
366 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
367 put_layout_hdr_locked(lo);
370 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
371 if (should_free_lseg(lseg->pls_range.iomode, iomode)) {
372 dprintk("%s: freeing lseg %p iomode %d "
373 "offset %llu length %llu\n", __func__,
374 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
375 lseg->pls_range.length);
377 removed += mark_lseg_invalid(lseg, tmp_list);
379 dprintk("%s:Return %i\n", __func__, invalid - removed);
380 return invalid - removed;
383 /* note free_me must contain lsegs from a single layout_hdr */
385 pnfs_free_lseg_list(struct list_head *free_me)
387 struct pnfs_layout_segment *lseg, *tmp;
388 struct pnfs_layout_hdr *lo;
390 if (list_empty(free_me))
393 lo = list_first_entry(free_me, struct pnfs_layout_segment,
394 pls_list)->pls_layout;
396 if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
397 struct nfs_client *clp;
399 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
400 spin_lock(&clp->cl_lock);
401 list_del_init(&lo->plh_layouts);
402 spin_unlock(&clp->cl_lock);
404 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
405 list_del(&lseg->pls_list);
411 pnfs_destroy_layout(struct nfs_inode *nfsi)
413 struct pnfs_layout_hdr *lo;
416 spin_lock(&nfsi->vfs_inode.i_lock);
419 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
420 mark_matching_lsegs_invalid(lo, &tmp_list, IOMODE_ANY);
422 spin_unlock(&nfsi->vfs_inode.i_lock);
423 pnfs_free_lseg_list(&tmp_list);
427 * Called by the state manger to remove all layouts established under an
431 pnfs_destroy_all_layouts(struct nfs_client *clp)
433 struct pnfs_layout_hdr *lo;
436 spin_lock(&clp->cl_lock);
437 list_splice_init(&clp->cl_layouts, &tmp_list);
438 spin_unlock(&clp->cl_lock);
440 while (!list_empty(&tmp_list)) {
441 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
443 dprintk("%s freeing layout for inode %lu\n", __func__,
444 lo->plh_inode->i_ino);
445 list_del_init(&lo->plh_layouts);
446 pnfs_destroy_layout(NFS_I(lo->plh_inode));
450 /* update lo->plh_stateid with new if is more recent */
452 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
457 oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
458 newseq = be32_to_cpu(new->stateid.seqid);
459 if ((int)(newseq - oldseq) > 0) {
460 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
461 if (update_barrier) {
462 u32 new_barrier = be32_to_cpu(new->stateid.seqid);
464 if ((int)(new_barrier - lo->plh_barrier))
465 lo->plh_barrier = new_barrier;
467 /* Because of wraparound, we want to keep the barrier
468 * "close" to the current seqids. It needs to be
469 * within 2**31 to count as "behind", so if it
470 * gets too near that limit, give us a litle leeway
471 * and bring it to within 2**30.
472 * NOTE - and yes, this is all unsigned arithmetic.
474 if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
475 lo->plh_barrier = newseq - (1 << 30);
480 /* lget is set to 1 if called from inside send_layoutget call chain */
482 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
486 (int)(lo->plh_barrier - be32_to_cpu(stateid->stateid.seqid)) >= 0)
488 return lo->plh_block_lgets ||
489 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
490 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
491 (list_empty(&lo->plh_segs) &&
492 (atomic_read(&lo->plh_outstanding) > lget));
496 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
497 struct nfs4_state *open_state)
501 dprintk("--> %s\n", __func__);
502 spin_lock(&lo->plh_inode->i_lock);
503 if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
505 } else if (list_empty(&lo->plh_segs)) {
509 seq = read_seqbegin(&open_state->seqlock);
510 memcpy(dst->data, open_state->stateid.data,
511 sizeof(open_state->stateid.data));
512 } while (read_seqretry(&open_state->seqlock, seq));
514 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
515 spin_unlock(&lo->plh_inode->i_lock);
516 dprintk("<-- %s\n", __func__);
521 * Get layout from server.
522 * for now, assume that whole file layouts are requested.
524 * arg->length: all ones
526 static struct pnfs_layout_segment *
527 send_layoutget(struct pnfs_layout_hdr *lo,
528 struct nfs_open_context *ctx,
529 struct pnfs_layout_range *range,
532 struct inode *ino = lo->plh_inode;
533 struct nfs_server *server = NFS_SERVER(ino);
534 struct nfs4_layoutget *lgp;
535 struct pnfs_layout_segment *lseg = NULL;
536 struct page **pages = NULL;
538 u32 max_resp_sz, max_pages;
540 dprintk("--> %s\n", __func__);
543 lgp = kzalloc(sizeof(*lgp), gfp_flags);
547 /* allocate pages for xdr post processing */
548 max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
549 max_pages = max_resp_sz >> PAGE_SHIFT;
551 pages = kzalloc(max_pages * sizeof(struct page *), gfp_flags);
555 for (i = 0; i < max_pages; i++) {
556 pages[i] = alloc_page(gfp_flags);
561 lgp->args.minlength = PAGE_CACHE_SIZE;
562 if (lgp->args.minlength > range->length)
563 lgp->args.minlength = range->length;
564 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
565 lgp->args.range = *range;
566 lgp->args.type = server->pnfs_curr_ld->id;
567 lgp->args.inode = ino;
568 lgp->args.ctx = get_nfs_open_context(ctx);
569 lgp->args.layout.pages = pages;
570 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
572 lgp->gfp_flags = gfp_flags;
574 /* Synchronously retrieve layout information from server and
577 nfs4_proc_layoutget(lgp);
579 /* remember that LAYOUTGET failed and suspend trying */
580 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
584 for (i = 0; i < max_pages; i++)
585 __free_page(pages[i]);
591 /* free any allocated xdr pages, lgp as it's not used */
593 for (i = 0; i < max_pages; i++) {
596 __free_page(pages[i]);
604 bool pnfs_roc(struct inode *ino)
606 struct pnfs_layout_hdr *lo;
607 struct pnfs_layout_segment *lseg, *tmp;
611 spin_lock(&ino->i_lock);
612 lo = NFS_I(ino)->layout;
613 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
614 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
616 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
617 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
618 mark_lseg_invalid(lseg, &tmp_list);
623 lo->plh_block_lgets++;
624 get_layout_hdr(lo); /* matched in pnfs_roc_release */
625 spin_unlock(&ino->i_lock);
626 pnfs_free_lseg_list(&tmp_list);
630 spin_unlock(&ino->i_lock);
634 void pnfs_roc_release(struct inode *ino)
636 struct pnfs_layout_hdr *lo;
638 spin_lock(&ino->i_lock);
639 lo = NFS_I(ino)->layout;
640 lo->plh_block_lgets--;
641 put_layout_hdr_locked(lo);
642 spin_unlock(&ino->i_lock);
645 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
647 struct pnfs_layout_hdr *lo;
649 spin_lock(&ino->i_lock);
650 lo = NFS_I(ino)->layout;
651 if ((int)(barrier - lo->plh_barrier) > 0)
652 lo->plh_barrier = barrier;
653 spin_unlock(&ino->i_lock);
656 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
658 struct nfs_inode *nfsi = NFS_I(ino);
659 struct pnfs_layout_segment *lseg;
662 spin_lock(&ino->i_lock);
663 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
664 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
669 struct pnfs_layout_hdr *lo = nfsi->layout;
670 u32 current_seqid = be32_to_cpu(lo->plh_stateid.stateid.seqid);
672 /* Since close does not return a layout stateid for use as
673 * a barrier, we choose the worst-case barrier.
675 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
677 spin_unlock(&ino->i_lock);
682 * Compare two layout segments for sorting into layout cache.
683 * We want to preferentially return RW over RO layouts, so ensure those
687 cmp_layout(struct pnfs_layout_range *l1,
688 struct pnfs_layout_range *l2)
692 /* high offset > low offset */
693 d = l1->offset - l2->offset;
697 /* short length > long length */
698 d = l2->length - l1->length;
702 /* read > read/write */
703 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
707 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
708 struct pnfs_layout_segment *lseg)
710 struct pnfs_layout_segment *lp;
712 dprintk("%s:Begin\n", __func__);
714 assert_spin_locked(&lo->plh_inode->i_lock);
715 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
716 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
718 list_add_tail(&lseg->pls_list, &lp->pls_list);
719 dprintk("%s: inserted lseg %p "
720 "iomode %d offset %llu length %llu before "
721 "lp %p iomode %d offset %llu length %llu\n",
722 __func__, lseg, lseg->pls_range.iomode,
723 lseg->pls_range.offset, lseg->pls_range.length,
724 lp, lp->pls_range.iomode, lp->pls_range.offset,
725 lp->pls_range.length);
728 list_add_tail(&lseg->pls_list, &lo->plh_segs);
729 dprintk("%s: inserted lseg %p "
730 "iomode %d offset %llu length %llu at tail\n",
731 __func__, lseg, lseg->pls_range.iomode,
732 lseg->pls_range.offset, lseg->pls_range.length);
736 dprintk("%s:Return\n", __func__);
739 static struct pnfs_layout_hdr *
740 alloc_init_layout_hdr(struct inode *ino, gfp_t gfp_flags)
742 struct pnfs_layout_hdr *lo;
744 lo = kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
747 atomic_set(&lo->plh_refcount, 1);
748 INIT_LIST_HEAD(&lo->plh_layouts);
749 INIT_LIST_HEAD(&lo->plh_segs);
750 INIT_LIST_HEAD(&lo->plh_bulk_recall);
755 static struct pnfs_layout_hdr *
756 pnfs_find_alloc_layout(struct inode *ino, gfp_t gfp_flags)
758 struct nfs_inode *nfsi = NFS_I(ino);
759 struct pnfs_layout_hdr *new = NULL;
761 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
763 assert_spin_locked(&ino->i_lock);
765 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
770 spin_unlock(&ino->i_lock);
771 new = alloc_init_layout_hdr(ino, gfp_flags);
772 spin_lock(&ino->i_lock);
774 if (likely(nfsi->layout == NULL)) /* Won the race? */
782 * iomode matching rules:
793 is_matching_lseg(struct pnfs_layout_range *ls_range,
794 struct pnfs_layout_range *range)
796 struct pnfs_layout_range range1;
798 if ((range->iomode == IOMODE_RW &&
799 ls_range->iomode != IOMODE_RW) ||
800 !lo_seg_intersecting(ls_range, range))
803 /* range1 covers only the first byte in the range */
806 return lo_seg_contained(ls_range, &range1);
810 * lookup range in layout
812 static struct pnfs_layout_segment *
813 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
814 struct pnfs_layout_range *range)
816 struct pnfs_layout_segment *lseg, *ret = NULL;
818 dprintk("%s:Begin\n", __func__);
820 assert_spin_locked(&lo->plh_inode->i_lock);
821 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
822 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
823 is_matching_lseg(&lseg->pls_range, range)) {
824 ret = get_lseg(lseg);
827 if (cmp_layout(range, &lseg->pls_range) > 0)
831 dprintk("%s:Return lseg %p ref %d\n",
832 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
837 * Layout segment is retreived from the server if not cached.
838 * The appropriate layout segment is referenced and returned to the caller.
840 struct pnfs_layout_segment *
841 pnfs_update_layout(struct inode *ino,
842 struct nfs_open_context *ctx,
845 enum pnfs_iomode iomode,
848 struct pnfs_layout_range arg = {
853 struct nfs_inode *nfsi = NFS_I(ino);
854 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
855 struct pnfs_layout_hdr *lo;
856 struct pnfs_layout_segment *lseg = NULL;
859 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
861 spin_lock(&ino->i_lock);
862 lo = pnfs_find_alloc_layout(ino, gfp_flags);
864 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
868 /* Do we even need to bother with this? */
869 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
870 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
871 dprintk("%s matches recall, use MDS\n", __func__);
875 /* if LAYOUTGET already failed once we don't try again */
876 if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
879 /* Check to see if the layout for the given range already exists */
880 lseg = pnfs_find_lseg(lo, &arg);
884 if (pnfs_layoutgets_blocked(lo, NULL, 0))
886 atomic_inc(&lo->plh_outstanding);
889 if (list_empty(&lo->plh_segs))
891 spin_unlock(&ino->i_lock);
893 /* The lo must be on the clp list if there is any
894 * chance of a CB_LAYOUTRECALL(FILE) coming in.
896 spin_lock(&clp->cl_lock);
897 BUG_ON(!list_empty(&lo->plh_layouts));
898 list_add_tail(&lo->plh_layouts, &clp->cl_layouts);
899 spin_unlock(&clp->cl_lock);
902 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
903 if (!lseg && first) {
904 spin_lock(&clp->cl_lock);
905 list_del_init(&lo->plh_layouts);
906 spin_unlock(&clp->cl_lock);
908 atomic_dec(&lo->plh_outstanding);
911 dprintk("%s end, state 0x%lx lseg %p\n", __func__,
912 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
915 spin_unlock(&ino->i_lock);
920 pnfs_layout_process(struct nfs4_layoutget *lgp)
922 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
923 struct nfs4_layoutget_res *res = &lgp->res;
924 struct pnfs_layout_segment *lseg;
925 struct inode *ino = lo->plh_inode;
926 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
929 /* Inject layout blob into I/O device driver */
930 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
931 if (!lseg || IS_ERR(lseg)) {
935 status = PTR_ERR(lseg);
936 dprintk("%s: Could not allocate layout: error %d\n",
941 spin_lock(&ino->i_lock);
942 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
943 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
944 dprintk("%s forget reply due to recall\n", __func__);
945 goto out_forget_reply;
948 if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
949 dprintk("%s forget reply due to state\n", __func__);
950 goto out_forget_reply;
953 lseg->pls_range = res->range;
954 *lgp->lsegpp = get_lseg(lseg);
955 pnfs_insert_layout(lo, lseg);
957 if (res->return_on_close) {
958 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
959 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
962 /* Done processing layoutget. Set the layout stateid */
963 pnfs_set_layout_stateid(lo, &res->stateid, false);
964 spin_unlock(&ino->i_lock);
969 spin_unlock(&ino->i_lock);
970 lseg->pls_layout = lo;
971 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
975 static int pnfs_read_pg_test(struct nfs_pageio_descriptor *pgio,
976 struct nfs_page *prev,
977 struct nfs_page *req)
979 if (pgio->pg_count == prev->wb_bytes) {
980 /* This is first coelesce call for a series of nfs_pages */
981 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
987 } else if (pgio->pg_lseg &&
988 req_offset(req) > end_offset(pgio->pg_lseg->pls_range.offset,
989 pgio->pg_lseg->pls_range.length))
991 return NFS_SERVER(pgio->pg_inode)->pnfs_curr_ld->pg_test(pgio, prev, req);
995 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
997 struct pnfs_layoutdriver_type *ld;
999 ld = NFS_SERVER(inode)->pnfs_curr_ld;
1000 pgio->pg_test = (ld && ld->pg_test) ? pnfs_read_pg_test : NULL;
1003 static int pnfs_write_pg_test(struct nfs_pageio_descriptor *pgio,
1004 struct nfs_page *prev,
1005 struct nfs_page *req)
1007 if (pgio->pg_count == prev->wb_bytes) {
1008 /* This is first coelesce call for a series of nfs_pages */
1009 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1015 } else if (pgio->pg_lseg &&
1016 req_offset(req) > end_offset(pgio->pg_lseg->pls_range.offset,
1017 pgio->pg_lseg->pls_range.length))
1019 return NFS_SERVER(pgio->pg_inode)->pnfs_curr_ld->pg_test(pgio, prev, req);
1023 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode)
1025 struct pnfs_layoutdriver_type *ld;
1027 ld = NFS_SERVER(inode)->pnfs_curr_ld;
1028 pgio->pg_test = (ld && ld->pg_test) ? pnfs_write_pg_test : NULL;
1031 enum pnfs_try_status
1032 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1033 const struct rpc_call_ops *call_ops, int how)
1035 struct inode *inode = wdata->inode;
1036 enum pnfs_try_status trypnfs;
1037 struct nfs_server *nfss = NFS_SERVER(inode);
1039 wdata->mds_ops = call_ops;
1041 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1042 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1044 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1045 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1046 put_lseg(wdata->lseg);
1049 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1051 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1056 * Call the appropriate parallel I/O subsystem read function.
1058 enum pnfs_try_status
1059 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1060 const struct rpc_call_ops *call_ops)
1062 struct inode *inode = rdata->inode;
1063 struct nfs_server *nfss = NFS_SERVER(inode);
1064 enum pnfs_try_status trypnfs;
1066 rdata->mds_ops = call_ops;
1068 dprintk("%s: Reading ino:%lu %u@%llu\n",
1069 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1071 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1072 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1073 put_lseg(rdata->lseg);
1076 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1078 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1083 * Currently there is only one (whole file) write lseg.
1085 static struct pnfs_layout_segment *pnfs_list_write_lseg(struct inode *inode)
1087 struct pnfs_layout_segment *lseg, *rv = NULL;
1089 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
1090 if (lseg->pls_range.iomode == IOMODE_RW)
1096 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1098 struct nfs_inode *nfsi = NFS_I(wdata->inode);
1099 loff_t end_pos = wdata->args.offset + wdata->res.count;
1100 bool mark_as_dirty = false;
1102 spin_lock(&nfsi->vfs_inode.i_lock);
1103 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1104 /* references matched in nfs4_layoutcommit_release */
1105 get_lseg(wdata->lseg);
1106 wdata->lseg->pls_lc_cred =
1107 get_rpccred(wdata->args.context->state->owner->so_cred);
1108 mark_as_dirty = true;
1109 dprintk("%s: Set layoutcommit for inode %lu ",
1110 __func__, wdata->inode->i_ino);
1112 if (end_pos > wdata->lseg->pls_end_pos)
1113 wdata->lseg->pls_end_pos = end_pos;
1114 spin_unlock(&nfsi->vfs_inode.i_lock);
1116 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1117 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1119 mark_inode_dirty_sync(wdata->inode);
1121 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1124 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1125 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1126 * data to disk to allow the server to recover the data if it crashes.
1127 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1128 * is off, and a COMMIT is sent to a data server, or
1129 * if WRITEs to a data server return NFS_DATA_SYNC.
1132 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1134 struct nfs4_layoutcommit_data *data;
1135 struct nfs_inode *nfsi = NFS_I(inode);
1136 struct pnfs_layout_segment *lseg;
1137 struct rpc_cred *cred;
1141 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1143 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1146 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1147 data = kzalloc(sizeof(*data), GFP_NOFS);
1149 mark_inode_dirty_sync(inode);
1154 spin_lock(&inode->i_lock);
1155 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1156 spin_unlock(&inode->i_lock);
1161 * Currently only one (whole file) write lseg which is referenced
1162 * in pnfs_set_layoutcommit and will be found.
1164 lseg = pnfs_list_write_lseg(inode);
1166 end_pos = lseg->pls_end_pos;
1167 cred = lseg->pls_lc_cred;
1168 lseg->pls_end_pos = 0;
1169 lseg->pls_lc_cred = NULL;
1171 memcpy(&data->args.stateid.data, nfsi->layout->plh_stateid.data,
1172 sizeof(nfsi->layout->plh_stateid.data));
1173 spin_unlock(&inode->i_lock);
1175 data->args.inode = inode;
1178 nfs_fattr_init(&data->fattr);
1179 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1180 data->res.fattr = &data->fattr;
1181 data->args.lastbytewritten = end_pos - 1;
1182 data->res.server = NFS_SERVER(inode);
1184 status = nfs4_proc_layoutcommit(data, sync);
1186 dprintk("<-- %s status %d\n", __func__, status);