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>
36 #define NFSDBG_FACILITY NFSDBG_PNFS
41 * protects pnfs_modules_tbl.
43 static DEFINE_SPINLOCK(pnfs_spinlock);
46 * pnfs_modules_tbl holds all pnfs modules
48 static LIST_HEAD(pnfs_modules_tbl);
50 /* Return the registered pnfs layout driver module matching given id */
51 static struct pnfs_layoutdriver_type *
52 find_pnfs_driver_locked(u32 id)
54 struct pnfs_layoutdriver_type *local;
56 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
61 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
65 static struct pnfs_layoutdriver_type *
66 find_pnfs_driver(u32 id)
68 struct pnfs_layoutdriver_type *local;
70 spin_lock(&pnfs_spinlock);
71 local = find_pnfs_driver_locked(id);
72 spin_unlock(&pnfs_spinlock);
77 unset_pnfs_layoutdriver(struct nfs_server *nfss)
79 if (nfss->pnfs_curr_ld) {
80 if (nfss->pnfs_curr_ld->clear_layoutdriver)
81 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
82 module_put(nfss->pnfs_curr_ld->owner);
84 nfss->pnfs_curr_ld = NULL;
88 * Try to set the server's pnfs module to the pnfs layout type specified by id.
89 * Currently only one pNFS layout driver per filesystem is supported.
91 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
94 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
97 struct pnfs_layoutdriver_type *ld_type = NULL;
101 if (!(server->nfs_client->cl_exchange_flags &
102 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
103 printk(KERN_ERR "%s: id %u cl_exchange_flags 0x%x\n", __func__,
104 id, server->nfs_client->cl_exchange_flags);
107 ld_type = find_pnfs_driver(id);
109 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
110 ld_type = find_pnfs_driver(id);
112 dprintk("%s: No pNFS module found for %u.\n",
117 if (!try_module_get(ld_type->owner)) {
118 dprintk("%s: Could not grab reference on module\n", __func__);
121 server->pnfs_curr_ld = ld_type;
122 if (ld_type->set_layoutdriver
123 && ld_type->set_layoutdriver(server, mntfh)) {
124 printk(KERN_ERR "%s: Error initializing pNFS layout driver %u.\n",
126 module_put(ld_type->owner);
130 dprintk("%s: pNFS module for %u set\n", __func__, id);
134 dprintk("%s: Using NFSv4 I/O\n", __func__);
135 server->pnfs_curr_ld = NULL;
139 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
141 int status = -EINVAL;
142 struct pnfs_layoutdriver_type *tmp;
144 if (ld_type->id == 0) {
145 printk(KERN_ERR "%s id 0 is reserved\n", __func__);
148 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
149 printk(KERN_ERR "%s Layout driver must provide "
150 "alloc_lseg and free_lseg.\n", __func__);
154 spin_lock(&pnfs_spinlock);
155 tmp = find_pnfs_driver_locked(ld_type->id);
157 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
159 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
162 printk(KERN_ERR "%s Module with id %d already loaded!\n",
163 __func__, ld_type->id);
165 spin_unlock(&pnfs_spinlock);
169 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
172 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
174 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
175 spin_lock(&pnfs_spinlock);
176 list_del(&ld_type->pnfs_tblid);
177 spin_unlock(&pnfs_spinlock);
179 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
182 * pNFS client layout cache
185 /* Need to hold i_lock if caller does not already hold reference */
187 get_layout_hdr(struct pnfs_layout_hdr *lo)
189 atomic_inc(&lo->plh_refcount);
192 static struct pnfs_layout_hdr *
193 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
195 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
196 return ld->alloc_layout_hdr ? ld->alloc_layout_hdr(ino, gfp_flags) :
197 kzalloc(sizeof(struct pnfs_layout_hdr), gfp_flags);
201 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
203 struct pnfs_layoutdriver_type *ld = NFS_SERVER(lo->plh_inode)->pnfs_curr_ld;
204 put_rpccred(lo->plh_lc_cred);
205 return ld->alloc_layout_hdr ? ld->free_layout_hdr(lo) : kfree(lo);
209 destroy_layout_hdr(struct pnfs_layout_hdr *lo)
211 dprintk("%s: freeing layout cache %p\n", __func__, lo);
212 BUG_ON(!list_empty(&lo->plh_layouts));
213 NFS_I(lo->plh_inode)->layout = NULL;
214 pnfs_free_layout_hdr(lo);
218 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
220 if (atomic_dec_and_test(&lo->plh_refcount))
221 destroy_layout_hdr(lo);
225 put_layout_hdr(struct pnfs_layout_hdr *lo)
227 struct inode *inode = lo->plh_inode;
229 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
230 destroy_layout_hdr(lo);
231 spin_unlock(&inode->i_lock);
236 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
238 INIT_LIST_HEAD(&lseg->pls_list);
239 INIT_LIST_HEAD(&lseg->pls_lc_list);
240 atomic_set(&lseg->pls_refcount, 1);
242 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
243 lseg->pls_layout = lo;
246 static void free_lseg(struct pnfs_layout_segment *lseg)
248 struct inode *ino = lseg->pls_layout->plh_inode;
250 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
251 /* Matched by get_layout_hdr in pnfs_insert_layout */
252 put_layout_hdr(NFS_I(ino)->layout);
256 put_lseg_common(struct pnfs_layout_segment *lseg)
258 struct inode *inode = lseg->pls_layout->plh_inode;
260 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
261 list_del_init(&lseg->pls_list);
262 if (list_empty(&lseg->pls_layout->plh_segs)) {
263 set_bit(NFS_LAYOUT_DESTROYED, &lseg->pls_layout->plh_flags);
264 /* Matched by initial refcount set in alloc_init_layout_hdr */
265 put_layout_hdr_locked(lseg->pls_layout);
267 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
271 put_lseg(struct pnfs_layout_segment *lseg)
278 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
279 atomic_read(&lseg->pls_refcount),
280 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
281 inode = lseg->pls_layout->plh_inode;
282 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
285 put_lseg_common(lseg);
286 list_add(&lseg->pls_list, &free_me);
287 spin_unlock(&inode->i_lock);
288 pnfs_free_lseg_list(&free_me);
291 EXPORT_SYMBOL_GPL(put_lseg);
294 end_offset(u64 start, u64 len)
299 return end >= start ? end : NFS4_MAX_UINT64;
302 /* last octet in a range */
304 last_byte_offset(u64 start, u64 len)
310 return end > start ? end - 1 : NFS4_MAX_UINT64;
314 * is l2 fully contained in l1?
316 * [----------------------------------)
321 lo_seg_contained(struct pnfs_layout_range *l1,
322 struct pnfs_layout_range *l2)
324 u64 start1 = l1->offset;
325 u64 end1 = end_offset(start1, l1->length);
326 u64 start2 = l2->offset;
327 u64 end2 = end_offset(start2, l2->length);
329 return (start1 <= start2) && (end1 >= end2);
333 * is l1 and l2 intersecting?
335 * [----------------------------------)
340 lo_seg_intersecting(struct pnfs_layout_range *l1,
341 struct pnfs_layout_range *l2)
343 u64 start1 = l1->offset;
344 u64 end1 = end_offset(start1, l1->length);
345 u64 start2 = l2->offset;
346 u64 end2 = end_offset(start2, l2->length);
348 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
349 (end2 == NFS4_MAX_UINT64 || end2 > start1);
353 should_free_lseg(struct pnfs_layout_range *lseg_range,
354 struct pnfs_layout_range *recall_range)
356 return (recall_range->iomode == IOMODE_ANY ||
357 lseg_range->iomode == recall_range->iomode) &&
358 lo_seg_intersecting(lseg_range, recall_range);
361 /* Returns 1 if lseg is removed from list, 0 otherwise */
362 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
363 struct list_head *tmp_list)
367 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
368 /* Remove the reference keeping the lseg in the
369 * list. It will now be removed when all
370 * outstanding io is finished.
372 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
373 atomic_read(&lseg->pls_refcount));
374 if (atomic_dec_and_test(&lseg->pls_refcount)) {
375 put_lseg_common(lseg);
376 list_add(&lseg->pls_list, tmp_list);
383 /* Returns count of number of matching invalid lsegs remaining in list
387 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
388 struct list_head *tmp_list,
389 struct pnfs_layout_range *recall_range)
391 struct pnfs_layout_segment *lseg, *next;
392 int invalid = 0, removed = 0;
394 dprintk("%s:Begin lo %p\n", __func__, lo);
396 if (list_empty(&lo->plh_segs)) {
397 if (!test_and_set_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags))
398 put_layout_hdr_locked(lo);
401 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
403 should_free_lseg(&lseg->pls_range, recall_range)) {
404 dprintk("%s: freeing lseg %p iomode %d "
405 "offset %llu length %llu\n", __func__,
406 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
407 lseg->pls_range.length);
409 removed += mark_lseg_invalid(lseg, tmp_list);
411 dprintk("%s:Return %i\n", __func__, invalid - removed);
412 return invalid - removed;
415 /* note free_me must contain lsegs from a single layout_hdr */
417 pnfs_free_lseg_list(struct list_head *free_me)
419 struct pnfs_layout_segment *lseg, *tmp;
420 struct pnfs_layout_hdr *lo;
422 if (list_empty(free_me))
425 lo = list_first_entry(free_me, struct pnfs_layout_segment,
426 pls_list)->pls_layout;
428 if (test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags)) {
429 struct nfs_client *clp;
431 clp = NFS_SERVER(lo->plh_inode)->nfs_client;
432 spin_lock(&clp->cl_lock);
433 list_del_init(&lo->plh_layouts);
434 spin_unlock(&clp->cl_lock);
436 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
437 list_del(&lseg->pls_list);
443 pnfs_destroy_layout(struct nfs_inode *nfsi)
445 struct pnfs_layout_hdr *lo;
448 spin_lock(&nfsi->vfs_inode.i_lock);
451 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
452 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
454 spin_unlock(&nfsi->vfs_inode.i_lock);
455 pnfs_free_lseg_list(&tmp_list);
459 * Called by the state manger to remove all layouts established under an
463 pnfs_destroy_all_layouts(struct nfs_client *clp)
465 struct nfs_server *server;
466 struct pnfs_layout_hdr *lo;
469 nfs4_deviceid_mark_client_invalid(clp);
470 nfs4_deviceid_purge_client(clp);
472 spin_lock(&clp->cl_lock);
474 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
475 if (!list_empty(&server->layouts))
476 list_splice_init(&server->layouts, &tmp_list);
479 spin_unlock(&clp->cl_lock);
481 while (!list_empty(&tmp_list)) {
482 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
484 dprintk("%s freeing layout for inode %lu\n", __func__,
485 lo->plh_inode->i_ino);
486 list_del_init(&lo->plh_layouts);
487 pnfs_destroy_layout(NFS_I(lo->plh_inode));
491 /* update lo->plh_stateid with new if is more recent */
493 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
498 oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
499 newseq = be32_to_cpu(new->stateid.seqid);
500 if ((int)(newseq - oldseq) > 0) {
501 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
502 if (update_barrier) {
503 u32 new_barrier = be32_to_cpu(new->stateid.seqid);
505 if ((int)(new_barrier - lo->plh_barrier))
506 lo->plh_barrier = new_barrier;
508 /* Because of wraparound, we want to keep the barrier
509 * "close" to the current seqids. It needs to be
510 * within 2**31 to count as "behind", so if it
511 * gets too near that limit, give us a litle leeway
512 * and bring it to within 2**30.
513 * NOTE - and yes, this is all unsigned arithmetic.
515 if (unlikely((newseq - lo->plh_barrier) > (3 << 29)))
516 lo->plh_barrier = newseq - (1 << 30);
521 /* lget is set to 1 if called from inside send_layoutget call chain */
523 pnfs_layoutgets_blocked(struct pnfs_layout_hdr *lo, nfs4_stateid *stateid,
527 (int)(lo->plh_barrier - be32_to_cpu(stateid->stateid.seqid)) >= 0)
529 return lo->plh_block_lgets ||
530 test_bit(NFS_LAYOUT_DESTROYED, &lo->plh_flags) ||
531 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
532 (list_empty(&lo->plh_segs) &&
533 (atomic_read(&lo->plh_outstanding) > lget));
537 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
538 struct nfs4_state *open_state)
542 dprintk("--> %s\n", __func__);
543 spin_lock(&lo->plh_inode->i_lock);
544 if (pnfs_layoutgets_blocked(lo, NULL, 1)) {
546 } else if (list_empty(&lo->plh_segs)) {
550 seq = read_seqbegin(&open_state->seqlock);
551 memcpy(dst->data, open_state->stateid.data,
552 sizeof(open_state->stateid.data));
553 } while (read_seqretry(&open_state->seqlock, seq));
555 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
556 spin_unlock(&lo->plh_inode->i_lock);
557 dprintk("<-- %s\n", __func__);
562 * Get layout from server.
563 * for now, assume that whole file layouts are requested.
565 * arg->length: all ones
567 static struct pnfs_layout_segment *
568 send_layoutget(struct pnfs_layout_hdr *lo,
569 struct nfs_open_context *ctx,
570 struct pnfs_layout_range *range,
573 struct inode *ino = lo->plh_inode;
574 struct nfs_server *server = NFS_SERVER(ino);
575 struct nfs4_layoutget *lgp;
576 struct pnfs_layout_segment *lseg = NULL;
577 struct page **pages = NULL;
579 u32 max_resp_sz, max_pages;
581 dprintk("--> %s\n", __func__);
584 lgp = kzalloc(sizeof(*lgp), gfp_flags);
588 /* allocate pages for xdr post processing */
589 max_resp_sz = server->nfs_client->cl_session->fc_attrs.max_resp_sz;
590 max_pages = max_resp_sz >> PAGE_SHIFT;
592 pages = kzalloc(max_pages * sizeof(struct page *), gfp_flags);
596 for (i = 0; i < max_pages; i++) {
597 pages[i] = alloc_page(gfp_flags);
602 lgp->args.minlength = PAGE_CACHE_SIZE;
603 if (lgp->args.minlength > range->length)
604 lgp->args.minlength = range->length;
605 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
606 lgp->args.range = *range;
607 lgp->args.type = server->pnfs_curr_ld->id;
608 lgp->args.inode = ino;
609 lgp->args.ctx = get_nfs_open_context(ctx);
610 lgp->args.layout.pages = pages;
611 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
613 lgp->gfp_flags = gfp_flags;
615 /* Synchronously retrieve layout information from server and
618 nfs4_proc_layoutget(lgp);
620 /* remember that LAYOUTGET failed and suspend trying */
621 set_bit(lo_fail_bit(range->iomode), &lo->plh_flags);
625 for (i = 0; i < max_pages; i++)
626 __free_page(pages[i]);
632 /* free any allocated xdr pages, lgp as it's not used */
634 for (i = 0; i < max_pages; i++) {
637 __free_page(pages[i]);
645 /* Initiates a LAYOUTRETURN(FILE) */
647 _pnfs_return_layout(struct inode *ino)
649 struct pnfs_layout_hdr *lo = NULL;
650 struct nfs_inode *nfsi = NFS_I(ino);
652 struct nfs4_layoutreturn *lrp;
653 nfs4_stateid stateid;
656 dprintk("--> %s\n", __func__);
658 spin_lock(&ino->i_lock);
661 spin_unlock(&ino->i_lock);
662 dprintk("%s: no layout to return\n", __func__);
665 stateid = nfsi->layout->plh_stateid;
666 /* Reference matched in nfs4_layoutreturn_release */
668 mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
669 lo->plh_block_lgets++;
670 spin_unlock(&ino->i_lock);
671 pnfs_free_lseg_list(&tmp_list);
673 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags));
675 lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
676 if (unlikely(lrp == NULL)) {
678 set_bit(NFS_LAYOUT_RW_FAILED, &lo->plh_flags);
679 set_bit(NFS_LAYOUT_RO_FAILED, &lo->plh_flags);
684 lrp->args.stateid = stateid;
685 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
686 lrp->args.inode = ino;
687 lrp->args.layout = lo;
688 lrp->clp = NFS_SERVER(ino)->nfs_client;
690 status = nfs4_proc_layoutreturn(lrp);
692 dprintk("<-- %s status: %d\n", __func__, status);
696 bool pnfs_roc(struct inode *ino)
698 struct pnfs_layout_hdr *lo;
699 struct pnfs_layout_segment *lseg, *tmp;
703 spin_lock(&ino->i_lock);
704 lo = NFS_I(ino)->layout;
705 if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
706 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
708 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
709 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
710 mark_lseg_invalid(lseg, &tmp_list);
715 lo->plh_block_lgets++;
716 get_layout_hdr(lo); /* matched in pnfs_roc_release */
717 spin_unlock(&ino->i_lock);
718 pnfs_free_lseg_list(&tmp_list);
722 spin_unlock(&ino->i_lock);
726 void pnfs_roc_release(struct inode *ino)
728 struct pnfs_layout_hdr *lo;
730 spin_lock(&ino->i_lock);
731 lo = NFS_I(ino)->layout;
732 lo->plh_block_lgets--;
733 put_layout_hdr_locked(lo);
734 spin_unlock(&ino->i_lock);
737 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
739 struct pnfs_layout_hdr *lo;
741 spin_lock(&ino->i_lock);
742 lo = NFS_I(ino)->layout;
743 if ((int)(barrier - lo->plh_barrier) > 0)
744 lo->plh_barrier = barrier;
745 spin_unlock(&ino->i_lock);
748 bool pnfs_roc_drain(struct inode *ino, u32 *barrier)
750 struct nfs_inode *nfsi = NFS_I(ino);
751 struct pnfs_layout_segment *lseg;
754 spin_lock(&ino->i_lock);
755 list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
756 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
761 struct pnfs_layout_hdr *lo = nfsi->layout;
762 u32 current_seqid = be32_to_cpu(lo->plh_stateid.stateid.seqid);
764 /* Since close does not return a layout stateid for use as
765 * a barrier, we choose the worst-case barrier.
767 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
769 spin_unlock(&ino->i_lock);
774 * Compare two layout segments for sorting into layout cache.
775 * We want to preferentially return RW over RO layouts, so ensure those
779 cmp_layout(struct pnfs_layout_range *l1,
780 struct pnfs_layout_range *l2)
784 /* high offset > low offset */
785 d = l1->offset - l2->offset;
789 /* short length > long length */
790 d = l2->length - l1->length;
794 /* read > read/write */
795 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
799 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
800 struct pnfs_layout_segment *lseg)
802 struct pnfs_layout_segment *lp;
804 dprintk("%s:Begin\n", __func__);
806 assert_spin_locked(&lo->plh_inode->i_lock);
807 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
808 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
810 list_add_tail(&lseg->pls_list, &lp->pls_list);
811 dprintk("%s: inserted lseg %p "
812 "iomode %d offset %llu length %llu before "
813 "lp %p iomode %d offset %llu length %llu\n",
814 __func__, lseg, lseg->pls_range.iomode,
815 lseg->pls_range.offset, lseg->pls_range.length,
816 lp, lp->pls_range.iomode, lp->pls_range.offset,
817 lp->pls_range.length);
820 list_add_tail(&lseg->pls_list, &lo->plh_segs);
821 dprintk("%s: inserted lseg %p "
822 "iomode %d offset %llu length %llu at tail\n",
823 __func__, lseg, lseg->pls_range.iomode,
824 lseg->pls_range.offset, lseg->pls_range.length);
828 dprintk("%s:Return\n", __func__);
831 static struct pnfs_layout_hdr *
832 alloc_init_layout_hdr(struct inode *ino,
833 struct nfs_open_context *ctx,
836 struct pnfs_layout_hdr *lo;
838 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
841 atomic_set(&lo->plh_refcount, 1);
842 INIT_LIST_HEAD(&lo->plh_layouts);
843 INIT_LIST_HEAD(&lo->plh_segs);
844 INIT_LIST_HEAD(&lo->plh_bulk_recall);
846 lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
850 static struct pnfs_layout_hdr *
851 pnfs_find_alloc_layout(struct inode *ino,
852 struct nfs_open_context *ctx,
855 struct nfs_inode *nfsi = NFS_I(ino);
856 struct pnfs_layout_hdr *new = NULL;
858 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
860 assert_spin_locked(&ino->i_lock);
862 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
867 spin_unlock(&ino->i_lock);
868 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
869 spin_lock(&ino->i_lock);
871 if (likely(nfsi->layout == NULL)) /* Won the race? */
874 pnfs_free_layout_hdr(new);
879 * iomode matching rules:
890 is_matching_lseg(struct pnfs_layout_range *ls_range,
891 struct pnfs_layout_range *range)
893 struct pnfs_layout_range range1;
895 if ((range->iomode == IOMODE_RW &&
896 ls_range->iomode != IOMODE_RW) ||
897 !lo_seg_intersecting(ls_range, range))
900 /* range1 covers only the first byte in the range */
903 return lo_seg_contained(ls_range, &range1);
907 * lookup range in layout
909 static struct pnfs_layout_segment *
910 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
911 struct pnfs_layout_range *range)
913 struct pnfs_layout_segment *lseg, *ret = NULL;
915 dprintk("%s:Begin\n", __func__);
917 assert_spin_locked(&lo->plh_inode->i_lock);
918 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
919 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
920 is_matching_lseg(&lseg->pls_range, range)) {
921 ret = get_lseg(lseg);
924 if (lseg->pls_range.offset > range->offset)
928 dprintk("%s:Return lseg %p ref %d\n",
929 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
934 * Layout segment is retreived from the server if not cached.
935 * The appropriate layout segment is referenced and returned to the caller.
937 struct pnfs_layout_segment *
938 pnfs_update_layout(struct inode *ino,
939 struct nfs_open_context *ctx,
942 enum pnfs_iomode iomode,
945 struct pnfs_layout_range arg = {
951 struct nfs_inode *nfsi = NFS_I(ino);
952 struct nfs_server *server = NFS_SERVER(ino);
953 struct nfs_client *clp = server->nfs_client;
954 struct pnfs_layout_hdr *lo;
955 struct pnfs_layout_segment *lseg = NULL;
958 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
960 spin_lock(&ino->i_lock);
961 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
963 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
967 /* Do we even need to bother with this? */
968 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
969 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
970 dprintk("%s matches recall, use MDS\n", __func__);
974 /* if LAYOUTGET already failed once we don't try again */
975 if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
978 /* Check to see if the layout for the given range already exists */
979 lseg = pnfs_find_lseg(lo, &arg);
983 if (pnfs_layoutgets_blocked(lo, NULL, 0))
985 atomic_inc(&lo->plh_outstanding);
988 if (list_empty(&lo->plh_segs))
990 spin_unlock(&ino->i_lock);
992 /* The lo must be on the clp list if there is any
993 * chance of a CB_LAYOUTRECALL(FILE) coming in.
995 spin_lock(&clp->cl_lock);
996 BUG_ON(!list_empty(&lo->plh_layouts));
997 list_add_tail(&lo->plh_layouts, &server->layouts);
998 spin_unlock(&clp->cl_lock);
1001 pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1003 arg.offset -= pg_offset;
1004 arg.length += pg_offset;
1006 if (arg.length != NFS4_MAX_UINT64)
1007 arg.length = PAGE_CACHE_ALIGN(arg.length);
1009 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1010 if (!lseg && first) {
1011 spin_lock(&clp->cl_lock);
1012 list_del_init(&lo->plh_layouts);
1013 spin_unlock(&clp->cl_lock);
1015 atomic_dec(&lo->plh_outstanding);
1018 dprintk("%s end, state 0x%lx lseg %p\n", __func__,
1019 nfsi->layout ? nfsi->layout->plh_flags : -1, lseg);
1022 spin_unlock(&ino->i_lock);
1025 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1028 pnfs_layout_process(struct nfs4_layoutget *lgp)
1030 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1031 struct nfs4_layoutget_res *res = &lgp->res;
1032 struct pnfs_layout_segment *lseg;
1033 struct inode *ino = lo->plh_inode;
1034 struct nfs_client *clp = NFS_SERVER(ino)->nfs_client;
1037 /* Inject layout blob into I/O device driver */
1038 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1039 if (!lseg || IS_ERR(lseg)) {
1043 status = PTR_ERR(lseg);
1044 dprintk("%s: Could not allocate layout: error %d\n",
1049 spin_lock(&ino->i_lock);
1050 if (test_bit(NFS4CLNT_LAYOUTRECALL, &clp->cl_state) ||
1051 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1052 dprintk("%s forget reply due to recall\n", __func__);
1053 goto out_forget_reply;
1056 if (pnfs_layoutgets_blocked(lo, &res->stateid, 1)) {
1057 dprintk("%s forget reply due to state\n", __func__);
1058 goto out_forget_reply;
1060 init_lseg(lo, lseg);
1061 lseg->pls_range = res->range;
1062 *lgp->lsegpp = get_lseg(lseg);
1063 pnfs_insert_layout(lo, lseg);
1065 if (res->return_on_close) {
1066 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1067 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1070 /* Done processing layoutget. Set the layout stateid */
1071 pnfs_set_layout_stateid(lo, &res->stateid, false);
1072 spin_unlock(&ino->i_lock);
1077 spin_unlock(&ino->i_lock);
1078 lseg->pls_layout = lo;
1079 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1084 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1086 BUG_ON(pgio->pg_lseg != NULL);
1088 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1094 /* If no lseg, fall back to read through mds */
1095 if (pgio->pg_lseg == NULL)
1096 nfs_pageio_reset_read_mds(pgio);
1099 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1102 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1104 BUG_ON(pgio->pg_lseg != NULL);
1106 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1112 /* If no lseg, fall back to write through mds */
1113 if (pgio->pg_lseg == NULL)
1114 nfs_pageio_reset_write_mds(pgio);
1116 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1119 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode)
1121 struct nfs_server *server = NFS_SERVER(inode);
1122 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1126 nfs_pageio_init(pgio, inode, ld->pg_read_ops, server->rsize, 0);
1131 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode, int ioflags)
1133 struct nfs_server *server = NFS_SERVER(inode);
1134 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1138 nfs_pageio_init(pgio, inode, ld->pg_write_ops, server->wsize, ioflags);
1143 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1144 struct nfs_page *req)
1146 if (pgio->pg_lseg == NULL)
1147 return nfs_generic_pg_test(pgio, prev, req);
1150 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1151 * Note that this test makes several assumptions:
1152 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1153 * is known to lie within the range.
1154 * - that the nfs_page being tested is known to be contiguous with the
1155 * previous nfs_page.
1156 * - Layout ranges are page aligned, so we only have to test the
1157 * start offset of the request.
1159 * Please also note that 'end_offset' is actually the offset of the
1160 * first byte that lies outside the pnfs_layout_range. FIXME?
1163 return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1164 pgio->pg_lseg->pls_range.length);
1166 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1169 * Called by non rpc-based layout drivers
1172 pnfs_ld_write_done(struct nfs_write_data *data)
1176 if (!data->pnfs_error) {
1177 pnfs_set_layoutcommit(data);
1178 data->mds_ops->rpc_call_done(&data->task, data);
1179 data->mds_ops->rpc_release(data);
1183 dprintk("%s: pnfs_error=%d, retry via MDS\n", __func__,
1185 status = nfs_initiate_write(data, NFS_CLIENT(data->inode),
1186 data->mds_ops, NFS_FILE_SYNC);
1187 return status ? : -EAGAIN;
1189 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1192 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1193 struct nfs_write_data *data)
1195 list_splice_tail_init(&data->pages, &desc->pg_list);
1196 if (data->req && list_empty(&data->req->wb_list))
1197 nfs_list_add_request(data->req, &desc->pg_list);
1198 nfs_pageio_reset_write_mds(desc);
1199 desc->pg_recoalesce = 1;
1200 nfs_writedata_release(data);
1203 static enum pnfs_try_status
1204 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1205 const struct rpc_call_ops *call_ops,
1206 struct pnfs_layout_segment *lseg,
1209 struct inode *inode = wdata->inode;
1210 enum pnfs_try_status trypnfs;
1211 struct nfs_server *nfss = NFS_SERVER(inode);
1213 wdata->mds_ops = call_ops;
1214 wdata->lseg = get_lseg(lseg);
1216 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1217 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1219 trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1220 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1221 put_lseg(wdata->lseg);
1224 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1226 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1231 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1233 struct nfs_write_data *data;
1234 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1235 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1237 desc->pg_lseg = NULL;
1238 while (!list_empty(head)) {
1239 enum pnfs_try_status trypnfs;
1241 data = list_entry(head->next, struct nfs_write_data, list);
1242 list_del_init(&data->list);
1244 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1245 if (trypnfs == PNFS_NOT_ATTEMPTED)
1246 pnfs_write_through_mds(desc, data);
1252 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1257 ret = nfs_generic_flush(desc, &head);
1259 put_lseg(desc->pg_lseg);
1260 desc->pg_lseg = NULL;
1263 pnfs_do_multiple_writes(desc, &head, desc->pg_ioflags);
1266 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1269 * Called by non rpc-based layout drivers
1272 pnfs_ld_read_done(struct nfs_read_data *data)
1276 if (!data->pnfs_error) {
1277 __nfs4_read_done_cb(data);
1278 data->mds_ops->rpc_call_done(&data->task, data);
1279 data->mds_ops->rpc_release(data);
1283 dprintk("%s: pnfs_error=%d, retry via MDS\n", __func__,
1285 status = nfs_initiate_read(data, NFS_CLIENT(data->inode),
1287 return status ? : -EAGAIN;
1289 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1292 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1293 struct nfs_read_data *data)
1295 list_splice_tail_init(&data->pages, &desc->pg_list);
1296 if (data->req && list_empty(&data->req->wb_list))
1297 nfs_list_add_request(data->req, &desc->pg_list);
1298 nfs_pageio_reset_read_mds(desc);
1299 desc->pg_recoalesce = 1;
1300 nfs_readdata_release(data);
1304 * Call the appropriate parallel I/O subsystem read function.
1306 static enum pnfs_try_status
1307 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1308 const struct rpc_call_ops *call_ops,
1309 struct pnfs_layout_segment *lseg)
1311 struct inode *inode = rdata->inode;
1312 struct nfs_server *nfss = NFS_SERVER(inode);
1313 enum pnfs_try_status trypnfs;
1315 rdata->mds_ops = call_ops;
1316 rdata->lseg = get_lseg(lseg);
1318 dprintk("%s: Reading ino:%lu %u@%llu\n",
1319 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1321 trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1322 if (trypnfs == PNFS_NOT_ATTEMPTED) {
1323 put_lseg(rdata->lseg);
1326 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1328 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1333 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1335 struct nfs_read_data *data;
1336 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1337 struct pnfs_layout_segment *lseg = desc->pg_lseg;
1339 desc->pg_lseg = NULL;
1340 while (!list_empty(head)) {
1341 enum pnfs_try_status trypnfs;
1343 data = list_entry(head->next, struct nfs_read_data, list);
1344 list_del_init(&data->list);
1346 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1347 if (trypnfs == PNFS_NOT_ATTEMPTED)
1348 pnfs_read_through_mds(desc, data);
1354 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1359 ret = nfs_generic_pagein(desc, &head);
1361 put_lseg(desc->pg_lseg);
1362 desc->pg_lseg = NULL;
1365 pnfs_do_multiple_reads(desc, &head);
1368 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1371 * There can be multiple RW segments.
1373 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1375 struct pnfs_layout_segment *lseg;
1377 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1378 if (lseg->pls_range.iomode == IOMODE_RW &&
1379 test_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1380 list_add(&lseg->pls_lc_list, listp);
1385 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1387 struct nfs_inode *nfsi = NFS_I(wdata->inode);
1388 loff_t end_pos = wdata->mds_offset + wdata->res.count;
1389 bool mark_as_dirty = false;
1391 spin_lock(&nfsi->vfs_inode.i_lock);
1392 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1393 mark_as_dirty = true;
1394 dprintk("%s: Set layoutcommit for inode %lu ",
1395 __func__, wdata->inode->i_ino);
1397 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &wdata->lseg->pls_flags)) {
1398 /* references matched in nfs4_layoutcommit_release */
1399 get_lseg(wdata->lseg);
1401 if (end_pos > nfsi->layout->plh_lwb)
1402 nfsi->layout->plh_lwb = end_pos;
1403 spin_unlock(&nfsi->vfs_inode.i_lock);
1404 dprintk("%s: lseg %p end_pos %llu\n",
1405 __func__, wdata->lseg, nfsi->layout->plh_lwb);
1407 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1408 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1410 mark_inode_dirty_sync(wdata->inode);
1412 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1414 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1416 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1418 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1419 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1423 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1424 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1425 * data to disk to allow the server to recover the data if it crashes.
1426 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1427 * is off, and a COMMIT is sent to a data server, or
1428 * if WRITEs to a data server return NFS_DATA_SYNC.
1431 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1433 struct nfs4_layoutcommit_data *data;
1434 struct nfs_inode *nfsi = NFS_I(inode);
1438 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1440 if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1443 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1444 data = kzalloc(sizeof(*data), GFP_NOFS);
1446 mark_inode_dirty_sync(inode);
1451 INIT_LIST_HEAD(&data->lseg_list);
1452 spin_lock(&inode->i_lock);
1453 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1454 spin_unlock(&inode->i_lock);
1459 pnfs_list_write_lseg(inode, &data->lseg_list);
1461 end_pos = nfsi->layout->plh_lwb;
1462 nfsi->layout->plh_lwb = 0;
1464 memcpy(&data->args.stateid.data, nfsi->layout->plh_stateid.data,
1465 sizeof(nfsi->layout->plh_stateid.data));
1466 spin_unlock(&inode->i_lock);
1468 data->args.inode = inode;
1469 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1470 nfs_fattr_init(&data->fattr);
1471 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1472 data->res.fattr = &data->fattr;
1473 data->args.lastbytewritten = end_pos - 1;
1474 data->res.server = NFS_SERVER(inode);
1476 status = nfs4_proc_layoutcommit(data, sync);
1478 dprintk("<-- %s status %d\n", __func__, status);