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>
34 #define NFSDBG_FACILITY NFSDBG_PNFS
39 * protects pnfs_modules_tbl.
41 static DEFINE_SPINLOCK(pnfs_spinlock);
44 * pnfs_modules_tbl holds all pnfs modules
46 static LIST_HEAD(pnfs_modules_tbl);
48 /* Return the registered pnfs layout driver module matching given id */
49 static struct pnfs_layoutdriver_type *
50 find_pnfs_driver_locked(u32 id)
52 struct pnfs_layoutdriver_type *local;
54 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
59 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
63 static struct pnfs_layoutdriver_type *
64 find_pnfs_driver(u32 id)
66 struct pnfs_layoutdriver_type *local;
68 spin_lock(&pnfs_spinlock);
69 local = find_pnfs_driver_locked(id);
70 spin_unlock(&pnfs_spinlock);
75 unset_pnfs_layoutdriver(struct nfs_server *nfss)
77 if (nfss->pnfs_curr_ld) {
78 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
79 module_put(nfss->pnfs_curr_ld->owner);
81 nfss->pnfs_curr_ld = NULL;
85 * Try to set the server's pnfs module to the pnfs layout type specified by id.
86 * Currently only one pNFS layout driver per filesystem is supported.
88 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
91 set_pnfs_layoutdriver(struct nfs_server *server, u32 id)
93 struct pnfs_layoutdriver_type *ld_type = NULL;
97 if (!(server->nfs_client->cl_exchange_flags &
98 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
99 printk(KERN_ERR "%s: id %u cl_exchange_flags 0x%x\n", __func__,
100 id, server->nfs_client->cl_exchange_flags);
103 ld_type = find_pnfs_driver(id);
105 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
106 ld_type = find_pnfs_driver(id);
108 dprintk("%s: No pNFS module found for %u.\n",
113 if (!try_module_get(ld_type->owner)) {
114 dprintk("%s: Could not grab reference on module\n", __func__);
117 server->pnfs_curr_ld = ld_type;
118 if (ld_type->set_layoutdriver(server)) {
120 "%s: Error initializing mount point for layout driver %u.\n",
122 module_put(ld_type->owner);
125 dprintk("%s: pNFS module for %u set\n", __func__, id);
129 dprintk("%s: Using NFSv4 I/O\n", __func__);
130 server->pnfs_curr_ld = NULL;
134 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
136 int status = -EINVAL;
137 struct pnfs_layoutdriver_type *tmp;
139 if (ld_type->id == 0) {
140 printk(KERN_ERR "%s id 0 is reserved\n", __func__);
143 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
144 printk(KERN_ERR "%s Layout driver must provide "
145 "alloc_lseg and free_lseg.\n", __func__);
149 spin_lock(&pnfs_spinlock);
150 tmp = find_pnfs_driver_locked(ld_type->id);
152 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
154 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
157 printk(KERN_ERR "%s Module with id %d already loaded!\n",
158 __func__, ld_type->id);
160 spin_unlock(&pnfs_spinlock);
164 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
167 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
169 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
170 spin_lock(&pnfs_spinlock);
171 list_del(&ld_type->pnfs_tblid);
172 spin_unlock(&pnfs_spinlock);
174 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
177 * pNFS client layout cache
181 get_layout_hdr_locked(struct pnfs_layout_hdr *lo)
183 assert_spin_locked(&lo->plh_inode->i_lock);
188 put_layout_hdr_locked(struct pnfs_layout_hdr *lo)
190 assert_spin_locked(&lo->plh_inode->i_lock);
191 BUG_ON(lo->plh_refcount == 0);
194 if (!lo->plh_refcount) {
195 dprintk("%s: freeing layout cache %p\n", __func__, lo);
196 BUG_ON(!list_empty(&lo->plh_layouts));
197 NFS_I(lo->plh_inode)->layout = NULL;
203 put_layout_hdr(struct inode *inode)
205 spin_lock(&inode->i_lock);
206 put_layout_hdr_locked(NFS_I(inode)->layout);
207 spin_unlock(&inode->i_lock);
211 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
213 INIT_LIST_HEAD(&lseg->pls_list);
214 atomic_set(&lseg->pls_refcount, 1);
216 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
217 lseg->pls_layout = lo;
220 static void free_lseg(struct pnfs_layout_segment *lseg)
222 struct inode *ino = lseg->pls_layout->plh_inode;
224 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
225 /* Matched by get_layout_hdr in pnfs_insert_layout */
229 /* The use of tmp_list is necessary because pnfs_curr_ld->free_lseg
230 * could sleep, so must be called outside of the lock.
231 * Returns 1 if object was removed, otherwise return 0.
234 put_lseg_locked(struct pnfs_layout_segment *lseg,
235 struct list_head *tmp_list)
237 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
238 atomic_read(&lseg->pls_refcount),
239 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
240 if (atomic_dec_and_test(&lseg->pls_refcount)) {
241 struct inode *ino = lseg->pls_layout->plh_inode;
243 BUG_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
244 list_del(&lseg->pls_list);
245 if (list_empty(&lseg->pls_layout->plh_segs)) {
246 struct nfs_client *clp;
248 clp = NFS_SERVER(ino)->nfs_client;
249 spin_lock(&clp->cl_lock);
250 /* List does not take a reference, so no need for put here */
251 list_del_init(&lseg->pls_layout->plh_layouts);
252 spin_unlock(&clp->cl_lock);
254 list_add(&lseg->pls_list, tmp_list);
261 should_free_lseg(u32 lseg_iomode, u32 recall_iomode)
263 return (recall_iomode == IOMODE_ANY ||
264 lseg_iomode == recall_iomode);
267 /* Returns 1 if lseg is removed from list, 0 otherwise */
268 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
269 struct list_head *tmp_list)
273 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
274 /* Remove the reference keeping the lseg in the
275 * list. It will now be removed when all
276 * outstanding io is finished.
278 rv = put_lseg_locked(lseg, tmp_list);
283 /* Returns count of number of matching invalid lsegs remaining in list
287 mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
288 struct list_head *tmp_list,
291 struct pnfs_layout_segment *lseg, *next;
292 int invalid = 0, removed = 0;
294 dprintk("%s:Begin lo %p\n", __func__, lo);
296 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
297 if (should_free_lseg(lseg->pls_range.iomode, iomode)) {
298 dprintk("%s: freeing lseg %p iomode %d "
299 "offset %llu length %llu\n", __func__,
300 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
301 lseg->pls_range.length);
303 removed += mark_lseg_invalid(lseg, tmp_list);
305 dprintk("%s:Return %i\n", __func__, invalid - removed);
306 return invalid - removed;
310 pnfs_free_lseg_list(struct list_head *free_me)
312 struct pnfs_layout_segment *lseg, *tmp;
314 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
315 list_del(&lseg->pls_list);
321 pnfs_destroy_layout(struct nfs_inode *nfsi)
323 struct pnfs_layout_hdr *lo;
326 spin_lock(&nfsi->vfs_inode.i_lock);
329 set_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags);
330 mark_matching_lsegs_invalid(lo, &tmp_list, IOMODE_ANY);
331 /* Matched by refcount set to 1 in alloc_init_layout_hdr */
332 put_layout_hdr_locked(lo);
334 spin_unlock(&nfsi->vfs_inode.i_lock);
335 pnfs_free_lseg_list(&tmp_list);
339 * Called by the state manger to remove all layouts established under an
343 pnfs_destroy_all_layouts(struct nfs_client *clp)
345 struct pnfs_layout_hdr *lo;
348 spin_lock(&clp->cl_lock);
349 list_splice_init(&clp->cl_layouts, &tmp_list);
350 spin_unlock(&clp->cl_lock);
352 while (!list_empty(&tmp_list)) {
353 lo = list_entry(tmp_list.next, struct pnfs_layout_hdr,
355 dprintk("%s freeing layout for inode %lu\n", __func__,
356 lo->plh_inode->i_ino);
357 pnfs_destroy_layout(NFS_I(lo->plh_inode));
361 /* update lo->plh_stateid with new if is more recent */
363 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo,
364 const nfs4_stateid *new)
368 oldseq = be32_to_cpu(lo->plh_stateid.stateid.seqid);
369 newseq = be32_to_cpu(new->stateid.seqid);
370 if ((int)(newseq - oldseq) > 0)
371 memcpy(&lo->plh_stateid, &new->stateid, sizeof(new->stateid));
375 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
376 struct nfs4_state *open_state)
380 dprintk("--> %s\n", __func__);
381 spin_lock(&lo->plh_inode->i_lock);
382 if (list_empty(&lo->plh_segs)) {
386 seq = read_seqbegin(&open_state->seqlock);
387 memcpy(dst->data, open_state->stateid.data,
388 sizeof(open_state->stateid.data));
389 } while (read_seqretry(&open_state->seqlock, seq));
391 memcpy(dst->data, lo->plh_stateid.data, sizeof(lo->plh_stateid.data));
392 spin_unlock(&lo->plh_inode->i_lock);
393 dprintk("<-- %s\n", __func__);
398 * Get layout from server.
399 * for now, assume that whole file layouts are requested.
401 * arg->length: all ones
403 static struct pnfs_layout_segment *
404 send_layoutget(struct pnfs_layout_hdr *lo,
405 struct nfs_open_context *ctx,
408 struct inode *ino = lo->plh_inode;
409 struct nfs_server *server = NFS_SERVER(ino);
410 struct nfs4_layoutget *lgp;
411 struct pnfs_layout_segment *lseg = NULL;
413 dprintk("--> %s\n", __func__);
416 lgp = kzalloc(sizeof(*lgp), GFP_KERNEL);
418 put_layout_hdr(lo->plh_inode);
421 lgp->args.minlength = NFS4_MAX_UINT64;
422 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
423 lgp->args.range.iomode = iomode;
424 lgp->args.range.offset = 0;
425 lgp->args.range.length = NFS4_MAX_UINT64;
426 lgp->args.type = server->pnfs_curr_ld->id;
427 lgp->args.inode = ino;
428 lgp->args.ctx = get_nfs_open_context(ctx);
431 /* Synchronously retrieve layout information from server and
434 nfs4_proc_layoutget(lgp);
436 /* remember that LAYOUTGET failed and suspend trying */
437 set_bit(lo_fail_bit(iomode), &lo->plh_flags);
443 * Compare two layout segments for sorting into layout cache.
444 * We want to preferentially return RW over RO layouts, so ensure those
448 cmp_layout(u32 iomode1, u32 iomode2)
450 /* read > read/write */
451 return (int)(iomode2 == IOMODE_READ) - (int)(iomode1 == IOMODE_READ);
455 pnfs_insert_layout(struct pnfs_layout_hdr *lo,
456 struct pnfs_layout_segment *lseg)
458 struct pnfs_layout_segment *lp;
461 dprintk("%s:Begin\n", __func__);
463 assert_spin_locked(&lo->plh_inode->i_lock);
464 if (list_empty(&lo->plh_segs)) {
465 struct nfs_client *clp = NFS_SERVER(lo->plh_inode)->nfs_client;
467 spin_lock(&clp->cl_lock);
468 BUG_ON(!list_empty(&lo->plh_layouts));
469 list_add_tail(&lo->plh_layouts, &clp->cl_layouts);
470 spin_unlock(&clp->cl_lock);
472 list_for_each_entry(lp, &lo->plh_segs, pls_list) {
473 if (cmp_layout(lp->pls_range.iomode, lseg->pls_range.iomode) > 0)
475 list_add_tail(&lseg->pls_list, &lp->pls_list);
476 dprintk("%s: inserted lseg %p "
477 "iomode %d offset %llu length %llu before "
478 "lp %p iomode %d offset %llu length %llu\n",
479 __func__, lseg, lseg->pls_range.iomode,
480 lseg->pls_range.offset, lseg->pls_range.length,
481 lp, lp->pls_range.iomode, lp->pls_range.offset,
482 lp->pls_range.length);
487 list_add_tail(&lseg->pls_list, &lo->plh_segs);
488 dprintk("%s: inserted lseg %p "
489 "iomode %d offset %llu length %llu at tail\n",
490 __func__, lseg, lseg->pls_range.iomode,
491 lseg->pls_range.offset, lseg->pls_range.length);
493 get_layout_hdr_locked(lo);
495 dprintk("%s:Return\n", __func__);
498 static struct pnfs_layout_hdr *
499 alloc_init_layout_hdr(struct inode *ino)
501 struct pnfs_layout_hdr *lo;
503 lo = kzalloc(sizeof(struct pnfs_layout_hdr), GFP_KERNEL);
506 lo->plh_refcount = 1;
507 INIT_LIST_HEAD(&lo->plh_layouts);
508 INIT_LIST_HEAD(&lo->plh_segs);
513 static struct pnfs_layout_hdr *
514 pnfs_find_alloc_layout(struct inode *ino)
516 struct nfs_inode *nfsi = NFS_I(ino);
517 struct pnfs_layout_hdr *new = NULL;
519 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
521 assert_spin_locked(&ino->i_lock);
523 if (test_bit(NFS_LAYOUT_DESTROYED, &nfsi->layout->plh_flags))
528 spin_unlock(&ino->i_lock);
529 new = alloc_init_layout_hdr(ino);
530 spin_lock(&ino->i_lock);
532 if (likely(nfsi->layout == NULL)) /* Won the race? */
540 * iomode matching rules:
551 is_matching_lseg(struct pnfs_layout_segment *lseg, u32 iomode)
553 return (iomode != IOMODE_RW || lseg->pls_range.iomode == IOMODE_RW);
557 * lookup range in layout
559 static struct pnfs_layout_segment *
560 pnfs_has_layout(struct pnfs_layout_hdr *lo, u32 iomode)
562 struct pnfs_layout_segment *lseg, *ret = NULL;
564 dprintk("%s:Begin\n", __func__);
566 assert_spin_locked(&lo->plh_inode->i_lock);
567 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
568 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
569 is_matching_lseg(lseg, iomode)) {
573 if (cmp_layout(iomode, lseg->pls_range.iomode) > 0)
577 dprintk("%s:Return lseg %p ref %d\n",
578 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
583 * Layout segment is retreived from the server if not cached.
584 * The appropriate layout segment is referenced and returned to the caller.
586 struct pnfs_layout_segment *
587 pnfs_update_layout(struct inode *ino,
588 struct nfs_open_context *ctx,
589 enum pnfs_iomode iomode)
591 struct nfs_inode *nfsi = NFS_I(ino);
592 struct pnfs_layout_hdr *lo;
593 struct pnfs_layout_segment *lseg = NULL;
595 if (!pnfs_enabled_sb(NFS_SERVER(ino)))
597 spin_lock(&ino->i_lock);
598 lo = pnfs_find_alloc_layout(ino);
600 dprintk("%s ERROR: can't get pnfs_layout_hdr\n", __func__);
604 /* Check to see if the layout for the given range already exists */
605 lseg = pnfs_has_layout(lo, iomode);
607 dprintk("%s: Using cached lseg %p for iomode %d)\n",
608 __func__, lseg, iomode);
612 /* if LAYOUTGET already failed once we don't try again */
613 if (test_bit(lo_fail_bit(iomode), &nfsi->layout->plh_flags))
616 get_layout_hdr_locked(lo); /* Matched in nfs4_layoutget_release */
617 spin_unlock(&ino->i_lock);
619 lseg = send_layoutget(lo, ctx, iomode);
621 dprintk("%s end, state 0x%lx lseg %p\n", __func__,
622 nfsi->layout->plh_flags, lseg);
625 spin_unlock(&ino->i_lock);
630 pnfs_layout_process(struct nfs4_layoutget *lgp)
632 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
633 struct nfs4_layoutget_res *res = &lgp->res;
634 struct pnfs_layout_segment *lseg;
635 struct inode *ino = lo->plh_inode;
638 /* Inject layout blob into I/O device driver */
639 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res);
640 if (!lseg || IS_ERR(lseg)) {
644 status = PTR_ERR(lseg);
645 dprintk("%s: Could not allocate layout: error %d\n",
650 spin_lock(&ino->i_lock);
652 lseg->pls_range = res->range;
654 pnfs_insert_layout(lo, lseg);
656 /* Done processing layoutget. Set the layout stateid */
657 pnfs_set_layout_stateid(lo, &res->stateid);
658 spin_unlock(&ino->i_lock);
664 * Device ID cache. Currently supports one layout type per struct nfs_client.
665 * Add layout type to the lookup key to expand to support multiple types.
668 pnfs_alloc_init_deviceid_cache(struct nfs_client *clp,
669 void (*free_callback)(struct pnfs_deviceid_node *))
671 struct pnfs_deviceid_cache *c;
673 c = kzalloc(sizeof(struct pnfs_deviceid_cache), GFP_KERNEL);
676 spin_lock(&clp->cl_lock);
677 if (clp->cl_devid_cache != NULL) {
678 atomic_inc(&clp->cl_devid_cache->dc_ref);
679 dprintk("%s [kref [%d]]\n", __func__,
680 atomic_read(&clp->cl_devid_cache->dc_ref));
683 /* kzalloc initializes hlists */
684 spin_lock_init(&c->dc_lock);
685 atomic_set(&c->dc_ref, 1);
686 c->dc_free_callback = free_callback;
687 clp->cl_devid_cache = c;
688 dprintk("%s [new]\n", __func__);
690 spin_unlock(&clp->cl_lock);
693 EXPORT_SYMBOL_GPL(pnfs_alloc_init_deviceid_cache);
696 * Called from pnfs_layoutdriver_type->free_lseg
697 * last layout segment reference frees deviceid
700 pnfs_put_deviceid(struct pnfs_deviceid_cache *c,
701 struct pnfs_deviceid_node *devid)
703 struct nfs4_deviceid *id = &devid->de_id;
704 struct pnfs_deviceid_node *d;
705 struct hlist_node *n;
706 long h = nfs4_deviceid_hash(id);
708 dprintk("%s [%d]\n", __func__, atomic_read(&devid->de_ref));
709 if (!atomic_dec_and_lock(&devid->de_ref, &c->dc_lock))
712 hlist_for_each_entry_rcu(d, n, &c->dc_deviceids[h], de_node)
713 if (!memcmp(&d->de_id, id, sizeof(*id))) {
714 hlist_del_rcu(&d->de_node);
715 spin_unlock(&c->dc_lock);
717 c->dc_free_callback(devid);
720 spin_unlock(&c->dc_lock);
721 /* Why wasn't it found in the list? */
724 EXPORT_SYMBOL_GPL(pnfs_put_deviceid);
726 /* Find and reference a deviceid */
727 struct pnfs_deviceid_node *
728 pnfs_find_get_deviceid(struct pnfs_deviceid_cache *c, struct nfs4_deviceid *id)
730 struct pnfs_deviceid_node *d;
731 struct hlist_node *n;
732 long hash = nfs4_deviceid_hash(id);
734 dprintk("--> %s hash %ld\n", __func__, hash);
736 hlist_for_each_entry_rcu(d, n, &c->dc_deviceids[hash], de_node) {
737 if (!memcmp(&d->de_id, id, sizeof(*id))) {
738 if (!atomic_inc_not_zero(&d->de_ref)) {
750 EXPORT_SYMBOL_GPL(pnfs_find_get_deviceid);
753 * Add a deviceid to the cache.
754 * GETDEVICEINFOs for same deviceid can race. If deviceid is found, discard new
756 struct pnfs_deviceid_node *
757 pnfs_add_deviceid(struct pnfs_deviceid_cache *c, struct pnfs_deviceid_node *new)
759 struct pnfs_deviceid_node *d;
760 long hash = nfs4_deviceid_hash(&new->de_id);
762 dprintk("--> %s hash %ld\n", __func__, hash);
763 spin_lock(&c->dc_lock);
764 d = pnfs_find_get_deviceid(c, &new->de_id);
766 spin_unlock(&c->dc_lock);
767 dprintk("%s [discard]\n", __func__);
768 c->dc_free_callback(new);
771 INIT_HLIST_NODE(&new->de_node);
772 atomic_set(&new->de_ref, 1);
773 hlist_add_head_rcu(&new->de_node, &c->dc_deviceids[hash]);
774 spin_unlock(&c->dc_lock);
775 dprintk("%s [new]\n", __func__);
778 EXPORT_SYMBOL_GPL(pnfs_add_deviceid);
781 pnfs_put_deviceid_cache(struct nfs_client *clp)
783 struct pnfs_deviceid_cache *local = clp->cl_devid_cache;
785 dprintk("--> %s cl_devid_cache %p\n", __func__, clp->cl_devid_cache);
786 if (atomic_dec_and_lock(&local->dc_ref, &clp->cl_lock)) {
788 /* Verify cache is empty */
789 for (i = 0; i < NFS4_DEVICE_ID_HASH_SIZE; i++)
790 BUG_ON(!hlist_empty(&local->dc_deviceids[i]));
791 clp->cl_devid_cache = NULL;
792 spin_unlock(&clp->cl_lock);
796 EXPORT_SYMBOL_GPL(pnfs_put_deviceid_cache);