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>
32 #include <linux/module.h>
37 #define NFSDBG_FACILITY NFSDBG_PNFS
38 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
43 * protects pnfs_modules_tbl.
45 static DEFINE_SPINLOCK(pnfs_spinlock
);
48 * pnfs_modules_tbl holds all pnfs modules
50 static LIST_HEAD(pnfs_modules_tbl
);
52 /* Return the registered pnfs layout driver module matching given id */
53 static struct pnfs_layoutdriver_type
*
54 find_pnfs_driver_locked(u32 id
)
56 struct pnfs_layoutdriver_type
*local
;
58 list_for_each_entry(local
, &pnfs_modules_tbl
, pnfs_tblid
)
63 dprintk("%s: Searching for id %u, found %p\n", __func__
, id
, local
);
67 static struct pnfs_layoutdriver_type
*
68 find_pnfs_driver(u32 id
)
70 struct pnfs_layoutdriver_type
*local
;
72 spin_lock(&pnfs_spinlock
);
73 local
= find_pnfs_driver_locked(id
);
74 if (local
!= NULL
&& !try_module_get(local
->owner
)) {
75 dprintk("%s: Could not grab reference on module\n", __func__
);
78 spin_unlock(&pnfs_spinlock
);
83 unset_pnfs_layoutdriver(struct nfs_server
*nfss
)
85 if (nfss
->pnfs_curr_ld
) {
86 if (nfss
->pnfs_curr_ld
->clear_layoutdriver
)
87 nfss
->pnfs_curr_ld
->clear_layoutdriver(nfss
);
88 /* Decrement the MDS count. Purge the deviceid cache if zero */
89 if (atomic_dec_and_test(&nfss
->nfs_client
->cl_mds_count
))
90 nfs4_deviceid_purge_client(nfss
->nfs_client
);
91 module_put(nfss
->pnfs_curr_ld
->owner
);
93 nfss
->pnfs_curr_ld
= NULL
;
97 * Try to set the server's pnfs module to the pnfs layout type specified by id.
98 * Currently only one pNFS layout driver per filesystem is supported.
100 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
103 set_pnfs_layoutdriver(struct nfs_server
*server
, const struct nfs_fh
*mntfh
,
106 struct pnfs_layoutdriver_type
*ld_type
= NULL
;
110 if (!(server
->nfs_client
->cl_exchange_flags
&
111 (EXCHGID4_FLAG_USE_NON_PNFS
| EXCHGID4_FLAG_USE_PNFS_MDS
))) {
112 printk(KERN_ERR
"NFS: %s: id %u cl_exchange_flags 0x%x\n",
113 __func__
, id
, server
->nfs_client
->cl_exchange_flags
);
116 ld_type
= find_pnfs_driver(id
);
118 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX
, id
);
119 ld_type
= find_pnfs_driver(id
);
121 dprintk("%s: No pNFS module found for %u.\n",
126 server
->pnfs_curr_ld
= ld_type
;
127 if (ld_type
->set_layoutdriver
128 && ld_type
->set_layoutdriver(server
, mntfh
)) {
129 printk(KERN_ERR
"NFS: %s: Error initializing pNFS layout "
130 "driver %u.\n", __func__
, id
);
131 module_put(ld_type
->owner
);
134 /* Bump the MDS count */
135 atomic_inc(&server
->nfs_client
->cl_mds_count
);
137 dprintk("%s: pNFS module for %u set\n", __func__
, id
);
141 dprintk("%s: Using NFSv4 I/O\n", __func__
);
142 server
->pnfs_curr_ld
= NULL
;
146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
148 int status
= -EINVAL
;
149 struct pnfs_layoutdriver_type
*tmp
;
151 if (ld_type
->id
== 0) {
152 printk(KERN_ERR
"NFS: %s id 0 is reserved\n", __func__
);
155 if (!ld_type
->alloc_lseg
|| !ld_type
->free_lseg
) {
156 printk(KERN_ERR
"NFS: %s Layout driver must provide "
157 "alloc_lseg and free_lseg.\n", __func__
);
161 spin_lock(&pnfs_spinlock
);
162 tmp
= find_pnfs_driver_locked(ld_type
->id
);
164 list_add(&ld_type
->pnfs_tblid
, &pnfs_modules_tbl
);
166 dprintk("%s Registering id:%u name:%s\n", __func__
, ld_type
->id
,
169 printk(KERN_ERR
"NFS: %s Module with id %d already loaded!\n",
170 __func__
, ld_type
->id
);
172 spin_unlock(&pnfs_spinlock
);
176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver
);
179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
181 dprintk("%s Deregistering id:%u\n", __func__
, ld_type
->id
);
182 spin_lock(&pnfs_spinlock
);
183 list_del(&ld_type
->pnfs_tblid
);
184 spin_unlock(&pnfs_spinlock
);
186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver
);
189 * pNFS client layout cache
192 /* Need to hold i_lock if caller does not already hold reference */
194 pnfs_get_layout_hdr(struct pnfs_layout_hdr
*lo
)
196 atomic_inc(&lo
->plh_refcount
);
199 static struct pnfs_layout_hdr
*
200 pnfs_alloc_layout_hdr(struct inode
*ino
, gfp_t gfp_flags
)
202 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(ino
)->pnfs_curr_ld
;
203 return ld
->alloc_layout_hdr
? ld
->alloc_layout_hdr(ino
, gfp_flags
) :
204 kzalloc(sizeof(struct pnfs_layout_hdr
), gfp_flags
);
208 pnfs_free_layout_hdr(struct pnfs_layout_hdr
*lo
)
210 struct nfs_server
*server
= NFS_SERVER(lo
->plh_inode
);
211 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
213 if (!list_empty(&lo
->plh_layouts
)) {
214 struct nfs_client
*clp
= server
->nfs_client
;
216 spin_lock(&clp
->cl_lock
);
217 list_del_init(&lo
->plh_layouts
);
218 spin_unlock(&clp
->cl_lock
);
220 put_rpccred(lo
->plh_lc_cred
);
221 return ld
->alloc_layout_hdr
? ld
->free_layout_hdr(lo
) : kfree(lo
);
225 pnfs_detach_layout_hdr(struct pnfs_layout_hdr
*lo
)
227 struct nfs_inode
*nfsi
= NFS_I(lo
->plh_inode
);
228 dprintk("%s: freeing layout cache %p\n", __func__
, lo
);
230 /* Reset MDS Threshold I/O counters */
236 pnfs_put_layout_hdr(struct pnfs_layout_hdr
*lo
)
238 struct inode
*inode
= lo
->plh_inode
;
240 if (atomic_dec_and_lock(&lo
->plh_refcount
, &inode
->i_lock
)) {
241 pnfs_detach_layout_hdr(lo
);
242 spin_unlock(&inode
->i_lock
);
243 pnfs_free_layout_hdr(lo
);
248 pnfs_iomode_to_fail_bit(u32 iomode
)
250 return iomode
== IOMODE_RW
?
251 NFS_LAYOUT_RW_FAILED
: NFS_LAYOUT_RO_FAILED
;
255 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr
*lo
, int fail_bit
)
257 lo
->plh_retry_timestamp
= jiffies
;
258 if (test_and_set_bit(fail_bit
, &lo
->plh_flags
))
259 atomic_inc(&lo
->plh_refcount
);
263 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr
*lo
, int fail_bit
)
265 if (test_and_clear_bit(fail_bit
, &lo
->plh_flags
))
266 atomic_dec(&lo
->plh_refcount
);
270 pnfs_layout_io_set_failed(struct pnfs_layout_hdr
*lo
, u32 iomode
)
272 struct inode
*inode
= lo
->plh_inode
;
273 struct pnfs_layout_range range
= {
276 .length
= NFS4_MAX_UINT64
,
280 spin_lock(&inode
->i_lock
);
281 pnfs_layout_set_fail_bit(lo
, pnfs_iomode_to_fail_bit(iomode
));
282 pnfs_mark_matching_lsegs_invalid(lo
, &head
, &range
);
283 spin_unlock(&inode
->i_lock
);
284 pnfs_free_lseg_list(&head
);
285 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__
,
286 iomode
== IOMODE_RW
? "RW" : "READ");
290 pnfs_layout_io_test_failed(struct pnfs_layout_hdr
*lo
, u32 iomode
)
292 unsigned long start
, end
;
293 int fail_bit
= pnfs_iomode_to_fail_bit(iomode
);
295 if (test_bit(fail_bit
, &lo
->plh_flags
) == 0)
298 start
= end
- PNFS_LAYOUTGET_RETRY_TIMEOUT
;
299 if (!time_in_range(lo
->plh_retry_timestamp
, start
, end
)) {
300 /* It is time to retry the failed layoutgets */
301 pnfs_layout_clear_fail_bit(lo
, fail_bit
);
308 init_lseg(struct pnfs_layout_hdr
*lo
, struct pnfs_layout_segment
*lseg
)
310 INIT_LIST_HEAD(&lseg
->pls_list
);
311 INIT_LIST_HEAD(&lseg
->pls_lc_list
);
312 atomic_set(&lseg
->pls_refcount
, 1);
314 set_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
);
315 lseg
->pls_layout
= lo
;
318 static void free_lseg(struct pnfs_layout_segment
*lseg
)
320 struct inode
*ino
= lseg
->pls_layout
->plh_inode
;
322 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
323 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
324 pnfs_put_layout_hdr(NFS_I(ino
)->layout
);
328 pnfs_layout_remove_lseg(struct pnfs_layout_hdr
*lo
,
329 struct pnfs_layout_segment
*lseg
)
331 struct inode
*inode
= lo
->plh_inode
;
333 WARN_ON(test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
334 list_del_init(&lseg
->pls_list
);
335 if (list_empty(&lo
->plh_segs
))
336 set_bit(NFS_LAYOUT_DESTROYED
, &lo
->plh_flags
);
337 rpc_wake_up(&NFS_SERVER(inode
)->roc_rpcwaitq
);
341 pnfs_put_lseg(struct pnfs_layout_segment
*lseg
)
343 struct pnfs_layout_hdr
*lo
;
349 dprintk("%s: lseg %p ref %d valid %d\n", __func__
, lseg
,
350 atomic_read(&lseg
->pls_refcount
),
351 test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
352 lo
= lseg
->pls_layout
;
353 inode
= lo
->plh_inode
;
354 if (atomic_dec_and_lock(&lseg
->pls_refcount
, &inode
->i_lock
)) {
357 pnfs_layout_remove_lseg(lo
, lseg
);
358 spin_unlock(&inode
->i_lock
);
359 list_add(&lseg
->pls_list
, &free_me
);
360 pnfs_free_lseg_list(&free_me
);
363 EXPORT_SYMBOL_GPL(pnfs_put_lseg
);
366 end_offset(u64 start
, u64 len
)
371 return end
>= start
? end
: NFS4_MAX_UINT64
;
374 /* last octet in a range */
376 last_byte_offset(u64 start
, u64 len
)
382 return end
> start
? end
- 1 : NFS4_MAX_UINT64
;
386 * is l2 fully contained in l1?
388 * [----------------------------------)
393 lo_seg_contained(struct pnfs_layout_range
*l1
,
394 struct pnfs_layout_range
*l2
)
396 u64 start1
= l1
->offset
;
397 u64 end1
= end_offset(start1
, l1
->length
);
398 u64 start2
= l2
->offset
;
399 u64 end2
= end_offset(start2
, l2
->length
);
401 return (start1
<= start2
) && (end1
>= end2
);
405 * is l1 and l2 intersecting?
407 * [----------------------------------)
412 lo_seg_intersecting(struct pnfs_layout_range
*l1
,
413 struct pnfs_layout_range
*l2
)
415 u64 start1
= l1
->offset
;
416 u64 end1
= end_offset(start1
, l1
->length
);
417 u64 start2
= l2
->offset
;
418 u64 end2
= end_offset(start2
, l2
->length
);
420 return (end1
== NFS4_MAX_UINT64
|| end1
> start2
) &&
421 (end2
== NFS4_MAX_UINT64
|| end2
> start1
);
425 should_free_lseg(struct pnfs_layout_range
*lseg_range
,
426 struct pnfs_layout_range
*recall_range
)
428 return (recall_range
->iomode
== IOMODE_ANY
||
429 lseg_range
->iomode
== recall_range
->iomode
) &&
430 lo_seg_intersecting(lseg_range
, recall_range
);
433 /* Returns 1 if lseg is removed from list, 0 otherwise */
434 static int mark_lseg_invalid(struct pnfs_layout_segment
*lseg
,
435 struct list_head
*tmp_list
)
439 if (test_and_clear_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
)) {
440 /* Remove the reference keeping the lseg in the
441 * list. It will now be removed when all
442 * outstanding io is finished.
444 dprintk("%s: lseg %p ref %d\n", __func__
, lseg
,
445 atomic_read(&lseg
->pls_refcount
));
446 if (atomic_dec_and_test(&lseg
->pls_refcount
)) {
447 pnfs_layout_remove_lseg(lseg
->pls_layout
, lseg
);
448 list_add(&lseg
->pls_list
, tmp_list
);
455 /* Returns count of number of matching invalid lsegs remaining in list
459 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr
*lo
,
460 struct list_head
*tmp_list
,
461 struct pnfs_layout_range
*recall_range
)
463 struct pnfs_layout_segment
*lseg
, *next
;
464 int invalid
= 0, removed
= 0;
466 dprintk("%s:Begin lo %p\n", __func__
, lo
);
468 if (list_empty(&lo
->plh_segs
)) {
469 set_bit(NFS_LAYOUT_DESTROYED
, &lo
->plh_flags
);
472 list_for_each_entry_safe(lseg
, next
, &lo
->plh_segs
, pls_list
)
474 should_free_lseg(&lseg
->pls_range
, recall_range
)) {
475 dprintk("%s: freeing lseg %p iomode %d "
476 "offset %llu length %llu\n", __func__
,
477 lseg
, lseg
->pls_range
.iomode
, lseg
->pls_range
.offset
,
478 lseg
->pls_range
.length
);
480 removed
+= mark_lseg_invalid(lseg
, tmp_list
);
482 dprintk("%s:Return %i\n", __func__
, invalid
- removed
);
483 return invalid
- removed
;
486 /* note free_me must contain lsegs from a single layout_hdr */
488 pnfs_free_lseg_list(struct list_head
*free_me
)
490 struct pnfs_layout_segment
*lseg
, *tmp
;
492 if (list_empty(free_me
))
495 list_for_each_entry_safe(lseg
, tmp
, free_me
, pls_list
) {
496 list_del(&lseg
->pls_list
);
502 pnfs_destroy_layout(struct nfs_inode
*nfsi
)
504 struct pnfs_layout_hdr
*lo
;
507 spin_lock(&nfsi
->vfs_inode
.i_lock
);
510 lo
->plh_block_lgets
++; /* permanently block new LAYOUTGETs */
511 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
512 pnfs_get_layout_hdr(lo
);
513 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RO_FAILED
);
514 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RW_FAILED
);
515 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
516 pnfs_free_lseg_list(&tmp_list
);
517 pnfs_put_layout_hdr(lo
);
519 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
521 EXPORT_SYMBOL_GPL(pnfs_destroy_layout
);
524 * Called by the state manger to remove all layouts established under an
528 pnfs_destroy_all_layouts(struct nfs_client
*clp
)
530 struct nfs_server
*server
;
531 struct pnfs_layout_hdr
*lo
;
534 nfs4_deviceid_mark_client_invalid(clp
);
535 nfs4_deviceid_purge_client(clp
);
537 spin_lock(&clp
->cl_lock
);
539 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
540 if (!list_empty(&server
->layouts
))
541 list_splice_init(&server
->layouts
, &tmp_list
);
544 spin_unlock(&clp
->cl_lock
);
546 while (!list_empty(&tmp_list
)) {
547 lo
= list_entry(tmp_list
.next
, struct pnfs_layout_hdr
,
549 dprintk("%s freeing layout for inode %lu\n", __func__
,
550 lo
->plh_inode
->i_ino
);
551 list_del_init(&lo
->plh_layouts
);
552 pnfs_destroy_layout(NFS_I(lo
->plh_inode
));
556 /* update lo->plh_stateid with new if is more recent */
558 pnfs_set_layout_stateid(struct pnfs_layout_hdr
*lo
, const nfs4_stateid
*new,
563 oldseq
= be32_to_cpu(lo
->plh_stateid
.seqid
);
564 newseq
= be32_to_cpu(new->seqid
);
565 if ((int)(newseq
- oldseq
) > 0) {
566 nfs4_stateid_copy(&lo
->plh_stateid
, new);
567 if (update_barrier
) {
568 u32 new_barrier
= be32_to_cpu(new->seqid
);
570 if ((int)(new_barrier
- lo
->plh_barrier
))
571 lo
->plh_barrier
= new_barrier
;
573 /* Because of wraparound, we want to keep the barrier
574 * "close" to the current seqids. It needs to be
575 * within 2**31 to count as "behind", so if it
576 * gets too near that limit, give us a litle leeway
577 * and bring it to within 2**30.
578 * NOTE - and yes, this is all unsigned arithmetic.
580 if (unlikely((newseq
- lo
->plh_barrier
) > (3 << 29)))
581 lo
->plh_barrier
= newseq
- (1 << 30);
586 /* lget is set to 1 if called from inside send_layoutget call chain */
588 pnfs_layoutgets_blocked(struct pnfs_layout_hdr
*lo
, nfs4_stateid
*stateid
,
592 (int)(lo
->plh_barrier
- be32_to_cpu(stateid
->seqid
)) >= 0)
594 return lo
->plh_block_lgets
||
595 test_bit(NFS_LAYOUT_DESTROYED
, &lo
->plh_flags
) ||
596 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
) ||
597 (list_empty(&lo
->plh_segs
) &&
598 (atomic_read(&lo
->plh_outstanding
) > lget
));
602 pnfs_choose_layoutget_stateid(nfs4_stateid
*dst
, struct pnfs_layout_hdr
*lo
,
603 struct nfs4_state
*open_state
)
607 dprintk("--> %s\n", __func__
);
608 spin_lock(&lo
->plh_inode
->i_lock
);
609 if (pnfs_layoutgets_blocked(lo
, NULL
, 1)) {
611 } else if (list_empty(&lo
->plh_segs
)) {
615 seq
= read_seqbegin(&open_state
->seqlock
);
616 nfs4_stateid_copy(dst
, &open_state
->stateid
);
617 } while (read_seqretry(&open_state
->seqlock
, seq
));
619 nfs4_stateid_copy(dst
, &lo
->plh_stateid
);
620 spin_unlock(&lo
->plh_inode
->i_lock
);
621 dprintk("<-- %s\n", __func__
);
626 * Get layout from server.
627 * for now, assume that whole file layouts are requested.
629 * arg->length: all ones
631 static struct pnfs_layout_segment
*
632 send_layoutget(struct pnfs_layout_hdr
*lo
,
633 struct nfs_open_context
*ctx
,
634 struct pnfs_layout_range
*range
,
637 struct inode
*ino
= lo
->plh_inode
;
638 struct nfs_server
*server
= NFS_SERVER(ino
);
639 struct nfs4_layoutget
*lgp
;
640 struct pnfs_layout_segment
*lseg
;
642 dprintk("--> %s\n", __func__
);
645 lgp
= kzalloc(sizeof(*lgp
), gfp_flags
);
649 lgp
->args
.minlength
= PAGE_CACHE_SIZE
;
650 if (lgp
->args
.minlength
> range
->length
)
651 lgp
->args
.minlength
= range
->length
;
652 lgp
->args
.maxcount
= PNFS_LAYOUT_MAXSIZE
;
653 lgp
->args
.range
= *range
;
654 lgp
->args
.type
= server
->pnfs_curr_ld
->id
;
655 lgp
->args
.inode
= ino
;
656 lgp
->args
.ctx
= get_nfs_open_context(ctx
);
657 lgp
->gfp_flags
= gfp_flags
;
659 /* Synchronously retrieve layout information from server and
662 lseg
= nfs4_proc_layoutget(lgp
, gfp_flags
);
664 switch (PTR_ERR(lseg
)) {
669 /* remember that LAYOUTGET failed and suspend trying */
670 pnfs_layout_io_set_failed(lo
, range
->iomode
);
679 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
680 * when the layout segment list is empty.
682 * Note that a pnfs_layout_hdr can exist with an empty layout segment
683 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
684 * deviceid is marked invalid.
687 _pnfs_return_layout(struct inode
*ino
)
689 struct pnfs_layout_hdr
*lo
= NULL
;
690 struct nfs_inode
*nfsi
= NFS_I(ino
);
692 struct nfs4_layoutreturn
*lrp
;
693 nfs4_stateid stateid
;
694 int status
= 0, empty
;
696 dprintk("NFS: %s for inode %lu\n", __func__
, ino
->i_ino
);
698 spin_lock(&ino
->i_lock
);
700 if (!lo
|| pnfs_test_layout_returned(lo
)) {
701 spin_unlock(&ino
->i_lock
);
702 dprintk("NFS: %s no layout to return\n", __func__
);
705 stateid
= nfsi
->layout
->plh_stateid
;
706 /* Reference matched in nfs4_layoutreturn_release */
707 pnfs_get_layout_hdr(lo
);
708 empty
= list_empty(&lo
->plh_segs
);
709 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
710 /* Don't send a LAYOUTRETURN if list was initially empty */
712 spin_unlock(&ino
->i_lock
);
713 pnfs_put_layout_hdr(lo
);
714 dprintk("NFS: %s no layout segments to return\n", __func__
);
717 lo
->plh_block_lgets
++;
718 pnfs_mark_layout_returned(lo
);
719 spin_unlock(&ino
->i_lock
);
720 pnfs_free_lseg_list(&tmp_list
);
722 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
));
724 lrp
= kzalloc(sizeof(*lrp
), GFP_KERNEL
);
725 if (unlikely(lrp
== NULL
)) {
727 pnfs_layout_io_set_failed(lo
, IOMODE_RW
);
728 pnfs_layout_io_set_failed(lo
, IOMODE_READ
);
729 pnfs_clear_layout_returned(lo
);
730 pnfs_put_layout_hdr(lo
);
734 lrp
->args
.stateid
= stateid
;
735 lrp
->args
.layout_type
= NFS_SERVER(ino
)->pnfs_curr_ld
->id
;
736 lrp
->args
.inode
= ino
;
737 lrp
->args
.layout
= lo
;
738 lrp
->clp
= NFS_SERVER(ino
)->nfs_client
;
740 status
= nfs4_proc_layoutreturn(lrp
);
742 dprintk("<-- %s status: %d\n", __func__
, status
);
745 EXPORT_SYMBOL_GPL(_pnfs_return_layout
);
747 bool pnfs_roc(struct inode
*ino
)
749 struct pnfs_layout_hdr
*lo
;
750 struct pnfs_layout_segment
*lseg
, *tmp
;
754 spin_lock(&ino
->i_lock
);
755 lo
= NFS_I(ino
)->layout
;
756 if (!lo
|| !test_and_clear_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
) ||
757 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
))
759 list_for_each_entry_safe(lseg
, tmp
, &lo
->plh_segs
, pls_list
)
760 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
761 mark_lseg_invalid(lseg
, &tmp_list
);
766 lo
->plh_block_lgets
++;
767 pnfs_get_layout_hdr(lo
); /* matched in pnfs_roc_release */
768 spin_unlock(&ino
->i_lock
);
769 pnfs_free_lseg_list(&tmp_list
);
773 spin_unlock(&ino
->i_lock
);
777 void pnfs_roc_release(struct inode
*ino
)
779 struct pnfs_layout_hdr
*lo
;
781 spin_lock(&ino
->i_lock
);
782 lo
= NFS_I(ino
)->layout
;
783 lo
->plh_block_lgets
--;
784 if (atomic_dec_and_test(&lo
->plh_refcount
)) {
785 pnfs_detach_layout_hdr(lo
);
786 spin_unlock(&ino
->i_lock
);
787 pnfs_free_layout_hdr(lo
);
789 spin_unlock(&ino
->i_lock
);
792 void pnfs_roc_set_barrier(struct inode
*ino
, u32 barrier
)
794 struct pnfs_layout_hdr
*lo
;
796 spin_lock(&ino
->i_lock
);
797 lo
= NFS_I(ino
)->layout
;
798 if ((int)(barrier
- lo
->plh_barrier
) > 0)
799 lo
->plh_barrier
= barrier
;
800 spin_unlock(&ino
->i_lock
);
803 bool pnfs_roc_drain(struct inode
*ino
, u32
*barrier
, struct rpc_task
*task
)
805 struct nfs_inode
*nfsi
= NFS_I(ino
);
806 struct pnfs_layout_hdr
*lo
;
807 struct pnfs_layout_segment
*lseg
;
811 spin_lock(&ino
->i_lock
);
812 list_for_each_entry(lseg
, &nfsi
->layout
->plh_segs
, pls_list
)
813 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
814 rpc_sleep_on(&NFS_SERVER(ino
)->roc_rpcwaitq
, task
, NULL
);
819 current_seqid
= be32_to_cpu(lo
->plh_stateid
.seqid
);
821 /* Since close does not return a layout stateid for use as
822 * a barrier, we choose the worst-case barrier.
824 *barrier
= current_seqid
+ atomic_read(&lo
->plh_outstanding
);
826 spin_unlock(&ino
->i_lock
);
831 * Compare two layout segments for sorting into layout cache.
832 * We want to preferentially return RW over RO layouts, so ensure those
836 cmp_layout(struct pnfs_layout_range
*l1
,
837 struct pnfs_layout_range
*l2
)
841 /* high offset > low offset */
842 d
= l1
->offset
- l2
->offset
;
846 /* short length > long length */
847 d
= l2
->length
- l1
->length
;
851 /* read > read/write */
852 return (int)(l1
->iomode
== IOMODE_READ
) - (int)(l2
->iomode
== IOMODE_READ
);
856 pnfs_layout_insert_lseg(struct pnfs_layout_hdr
*lo
,
857 struct pnfs_layout_segment
*lseg
)
859 struct pnfs_layout_segment
*lp
;
861 dprintk("%s:Begin\n", __func__
);
863 assert_spin_locked(&lo
->plh_inode
->i_lock
);
864 list_for_each_entry(lp
, &lo
->plh_segs
, pls_list
) {
865 if (cmp_layout(&lseg
->pls_range
, &lp
->pls_range
) > 0)
867 list_add_tail(&lseg
->pls_list
, &lp
->pls_list
);
868 dprintk("%s: inserted lseg %p "
869 "iomode %d offset %llu length %llu before "
870 "lp %p iomode %d offset %llu length %llu\n",
871 __func__
, lseg
, lseg
->pls_range
.iomode
,
872 lseg
->pls_range
.offset
, lseg
->pls_range
.length
,
873 lp
, lp
->pls_range
.iomode
, lp
->pls_range
.offset
,
874 lp
->pls_range
.length
);
877 list_add_tail(&lseg
->pls_list
, &lo
->plh_segs
);
878 dprintk("%s: inserted lseg %p "
879 "iomode %d offset %llu length %llu at tail\n",
880 __func__
, lseg
, lseg
->pls_range
.iomode
,
881 lseg
->pls_range
.offset
, lseg
->pls_range
.length
);
883 pnfs_get_layout_hdr(lo
);
885 dprintk("%s:Return\n", __func__
);
888 static struct pnfs_layout_hdr
*
889 alloc_init_layout_hdr(struct inode
*ino
,
890 struct nfs_open_context
*ctx
,
893 struct pnfs_layout_hdr
*lo
;
895 lo
= pnfs_alloc_layout_hdr(ino
, gfp_flags
);
898 atomic_set(&lo
->plh_refcount
, 1);
899 INIT_LIST_HEAD(&lo
->plh_layouts
);
900 INIT_LIST_HEAD(&lo
->plh_segs
);
901 INIT_LIST_HEAD(&lo
->plh_bulk_recall
);
903 lo
->plh_lc_cred
= get_rpccred(ctx
->state
->owner
->so_cred
);
907 static struct pnfs_layout_hdr
*
908 pnfs_find_alloc_layout(struct inode
*ino
,
909 struct nfs_open_context
*ctx
,
912 struct nfs_inode
*nfsi
= NFS_I(ino
);
913 struct pnfs_layout_hdr
*new = NULL
;
915 dprintk("%s Begin ino=%p layout=%p\n", __func__
, ino
, nfsi
->layout
);
917 assert_spin_locked(&ino
->i_lock
);
919 if (test_bit(NFS_LAYOUT_DESTROYED
, &nfsi
->layout
->plh_flags
))
921 pnfs_get_layout_hdr(nfsi
->layout
);
924 spin_unlock(&ino
->i_lock
);
925 new = alloc_init_layout_hdr(ino
, ctx
, gfp_flags
);
926 spin_lock(&ino
->i_lock
);
928 if (likely(nfsi
->layout
== NULL
)) /* Won the race? */
931 pnfs_free_layout_hdr(new);
936 * iomode matching rules:
947 is_matching_lseg(struct pnfs_layout_range
*ls_range
,
948 struct pnfs_layout_range
*range
)
950 struct pnfs_layout_range range1
;
952 if ((range
->iomode
== IOMODE_RW
&&
953 ls_range
->iomode
!= IOMODE_RW
) ||
954 !lo_seg_intersecting(ls_range
, range
))
957 /* range1 covers only the first byte in the range */
960 return lo_seg_contained(ls_range
, &range1
);
964 * lookup range in layout
966 static struct pnfs_layout_segment
*
967 pnfs_find_lseg(struct pnfs_layout_hdr
*lo
,
968 struct pnfs_layout_range
*range
)
970 struct pnfs_layout_segment
*lseg
, *ret
= NULL
;
972 dprintk("%s:Begin\n", __func__
);
974 assert_spin_locked(&lo
->plh_inode
->i_lock
);
975 list_for_each_entry(lseg
, &lo
->plh_segs
, pls_list
) {
976 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
) &&
977 is_matching_lseg(&lseg
->pls_range
, range
)) {
978 ret
= pnfs_get_lseg(lseg
);
981 if (lseg
->pls_range
.offset
> range
->offset
)
985 dprintk("%s:Return lseg %p ref %d\n",
986 __func__
, ret
, ret
? atomic_read(&ret
->pls_refcount
) : 0);
991 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
992 * to the MDS or over pNFS
994 * The nfs_inode read_io and write_io fields are cumulative counters reset
995 * when there are no layout segments. Note that in pnfs_update_layout iomode
996 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
999 * A return of true means use MDS I/O.
1002 * If a file's size is smaller than the file size threshold, data accesses
1003 * SHOULD be sent to the metadata server. If an I/O request has a length that
1004 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1005 * server. If both file size and I/O size are provided, the client SHOULD
1006 * reach or exceed both thresholds before sending its read or write
1007 * requests to the data server.
1009 static bool pnfs_within_mdsthreshold(struct nfs_open_context
*ctx
,
1010 struct inode
*ino
, int iomode
)
1012 struct nfs4_threshold
*t
= ctx
->mdsthreshold
;
1013 struct nfs_inode
*nfsi
= NFS_I(ino
);
1014 loff_t fsize
= i_size_read(ino
);
1015 bool size
= false, size_set
= false, io
= false, io_set
= false, ret
= false;
1020 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1021 __func__
, t
->bm
, t
->rd_sz
, t
->wr_sz
, t
->rd_io_sz
, t
->wr_io_sz
);
1025 if (t
->bm
& THRESHOLD_RD
) {
1026 dprintk("%s fsize %llu\n", __func__
, fsize
);
1028 if (fsize
< t
->rd_sz
)
1031 if (t
->bm
& THRESHOLD_RD_IO
) {
1032 dprintk("%s nfsi->read_io %llu\n", __func__
,
1035 if (nfsi
->read_io
< t
->rd_io_sz
)
1040 if (t
->bm
& THRESHOLD_WR
) {
1041 dprintk("%s fsize %llu\n", __func__
, fsize
);
1043 if (fsize
< t
->wr_sz
)
1046 if (t
->bm
& THRESHOLD_WR_IO
) {
1047 dprintk("%s nfsi->write_io %llu\n", __func__
,
1050 if (nfsi
->write_io
< t
->wr_io_sz
)
1055 if (size_set
&& io_set
) {
1058 } else if (size
|| io
)
1061 dprintk("<-- %s size %d io %d ret %d\n", __func__
, size
, io
, ret
);
1066 * Layout segment is retreived from the server if not cached.
1067 * The appropriate layout segment is referenced and returned to the caller.
1069 struct pnfs_layout_segment
*
1070 pnfs_update_layout(struct inode
*ino
,
1071 struct nfs_open_context
*ctx
,
1074 enum pnfs_iomode iomode
,
1077 struct pnfs_layout_range arg
= {
1083 struct nfs_server
*server
= NFS_SERVER(ino
);
1084 struct nfs_client
*clp
= server
->nfs_client
;
1085 struct pnfs_layout_hdr
*lo
;
1086 struct pnfs_layout_segment
*lseg
= NULL
;
1089 if (!pnfs_enabled_sb(NFS_SERVER(ino
)))
1092 if (pnfs_within_mdsthreshold(ctx
, ino
, iomode
))
1095 spin_lock(&ino
->i_lock
);
1096 lo
= pnfs_find_alloc_layout(ino
, ctx
, gfp_flags
);
1098 spin_unlock(&ino
->i_lock
);
1102 /* Do we even need to bother with this? */
1103 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1104 dprintk("%s matches recall, use MDS\n", __func__
);
1108 /* if LAYOUTGET already failed once we don't try again */
1109 if (pnfs_layout_io_test_failed(lo
, iomode
))
1112 /* Check to see if the layout for the given range already exists */
1113 lseg
= pnfs_find_lseg(lo
, &arg
);
1117 if (pnfs_layoutgets_blocked(lo
, NULL
, 0))
1119 atomic_inc(&lo
->plh_outstanding
);
1121 if (list_empty(&lo
->plh_segs
))
1124 /* Enable LAYOUTRETURNs */
1125 pnfs_clear_layout_returned(lo
);
1127 spin_unlock(&ino
->i_lock
);
1129 /* The lo must be on the clp list if there is any
1130 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1132 spin_lock(&clp
->cl_lock
);
1133 BUG_ON(!list_empty(&lo
->plh_layouts
));
1134 list_add_tail(&lo
->plh_layouts
, &server
->layouts
);
1135 spin_unlock(&clp
->cl_lock
);
1138 pg_offset
= arg
.offset
& ~PAGE_CACHE_MASK
;
1140 arg
.offset
-= pg_offset
;
1141 arg
.length
+= pg_offset
;
1143 if (arg
.length
!= NFS4_MAX_UINT64
)
1144 arg
.length
= PAGE_CACHE_ALIGN(arg
.length
);
1146 lseg
= send_layoutget(lo
, ctx
, &arg
, gfp_flags
);
1147 atomic_dec(&lo
->plh_outstanding
);
1149 pnfs_put_layout_hdr(lo
);
1151 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1152 "(%s, offset: %llu, length: %llu)\n",
1153 __func__
, ino
->i_sb
->s_id
,
1154 (unsigned long long)NFS_FILEID(ino
),
1155 lseg
== NULL
? "not found" : "found",
1156 iomode
==IOMODE_RW
? "read/write" : "read-only",
1157 (unsigned long long)pos
,
1158 (unsigned long long)count
);
1161 spin_unlock(&ino
->i_lock
);
1162 goto out_put_layout_hdr
;
1164 EXPORT_SYMBOL_GPL(pnfs_update_layout
);
1166 struct pnfs_layout_segment
*
1167 pnfs_layout_process(struct nfs4_layoutget
*lgp
)
1169 struct pnfs_layout_hdr
*lo
= NFS_I(lgp
->args
.inode
)->layout
;
1170 struct nfs4_layoutget_res
*res
= &lgp
->res
;
1171 struct pnfs_layout_segment
*lseg
;
1172 struct inode
*ino
= lo
->plh_inode
;
1175 /* Inject layout blob into I/O device driver */
1176 lseg
= NFS_SERVER(ino
)->pnfs_curr_ld
->alloc_lseg(lo
, res
, lgp
->gfp_flags
);
1177 if (!lseg
|| IS_ERR(lseg
)) {
1181 status
= PTR_ERR(lseg
);
1182 dprintk("%s: Could not allocate layout: error %d\n",
1187 spin_lock(&ino
->i_lock
);
1188 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1189 dprintk("%s forget reply due to recall\n", __func__
);
1190 goto out_forget_reply
;
1193 if (pnfs_layoutgets_blocked(lo
, &res
->stateid
, 1)) {
1194 dprintk("%s forget reply due to state\n", __func__
);
1195 goto out_forget_reply
;
1197 init_lseg(lo
, lseg
);
1198 lseg
->pls_range
= res
->range
;
1199 pnfs_get_lseg(lseg
);
1200 pnfs_layout_insert_lseg(lo
, lseg
);
1202 if (res
->return_on_close
) {
1203 set_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
);
1204 set_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
);
1207 /* Done processing layoutget. Set the layout stateid */
1208 pnfs_set_layout_stateid(lo
, &res
->stateid
, false);
1209 spin_unlock(&ino
->i_lock
);
1212 return ERR_PTR(status
);
1215 spin_unlock(&ino
->i_lock
);
1216 lseg
->pls_layout
= lo
;
1217 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
1222 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1224 BUG_ON(pgio
->pg_lseg
!= NULL
);
1226 if (req
->wb_offset
!= req
->wb_pgbase
) {
1227 nfs_pageio_reset_read_mds(pgio
);
1230 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1236 /* If no lseg, fall back to read through mds */
1237 if (pgio
->pg_lseg
== NULL
)
1238 nfs_pageio_reset_read_mds(pgio
);
1241 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read
);
1244 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1246 BUG_ON(pgio
->pg_lseg
!= NULL
);
1248 if (req
->wb_offset
!= req
->wb_pgbase
) {
1249 nfs_pageio_reset_write_mds(pgio
);
1252 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1258 /* If no lseg, fall back to write through mds */
1259 if (pgio
->pg_lseg
== NULL
)
1260 nfs_pageio_reset_write_mds(pgio
);
1262 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write
);
1265 pnfs_pageio_init_read(struct nfs_pageio_descriptor
*pgio
, struct inode
*inode
,
1266 const struct nfs_pgio_completion_ops
*compl_ops
)
1268 struct nfs_server
*server
= NFS_SERVER(inode
);
1269 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
1272 nfs_pageio_init_read(pgio
, inode
, compl_ops
);
1274 nfs_pageio_init(pgio
, inode
, ld
->pg_read_ops
, compl_ops
, server
->rsize
, 0);
1278 pnfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
, struct inode
*inode
,
1280 const struct nfs_pgio_completion_ops
*compl_ops
)
1282 struct nfs_server
*server
= NFS_SERVER(inode
);
1283 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
1286 nfs_pageio_init_write(pgio
, inode
, ioflags
, compl_ops
);
1288 nfs_pageio_init(pgio
, inode
, ld
->pg_write_ops
, compl_ops
, server
->wsize
, ioflags
);
1292 pnfs_generic_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
1293 struct nfs_page
*req
)
1295 if (pgio
->pg_lseg
== NULL
)
1296 return nfs_generic_pg_test(pgio
, prev
, req
);
1299 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1300 * Note that this test makes several assumptions:
1301 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1302 * is known to lie within the range.
1303 * - that the nfs_page being tested is known to be contiguous with the
1304 * previous nfs_page.
1305 * - Layout ranges are page aligned, so we only have to test the
1306 * start offset of the request.
1308 * Please also note that 'end_offset' is actually the offset of the
1309 * first byte that lies outside the pnfs_layout_range. FIXME?
1312 return req_offset(req
) < end_offset(pgio
->pg_lseg
->pls_range
.offset
,
1313 pgio
->pg_lseg
->pls_range
.length
);
1315 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test
);
1317 int pnfs_write_done_resend_to_mds(struct inode
*inode
,
1318 struct list_head
*head
,
1319 const struct nfs_pgio_completion_ops
*compl_ops
)
1321 struct nfs_pageio_descriptor pgio
;
1324 /* Resend all requests through the MDS */
1325 nfs_pageio_init_write(&pgio
, inode
, FLUSH_STABLE
, compl_ops
);
1326 while (!list_empty(head
)) {
1327 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1329 nfs_list_remove_request(req
);
1330 if (!nfs_pageio_add_request(&pgio
, req
))
1331 nfs_list_add_request(req
, &failed
);
1333 nfs_pageio_complete(&pgio
);
1335 if (!list_empty(&failed
)) {
1336 /* For some reason our attempt to resend pages. Mark the
1337 * overall send request as having failed, and let
1338 * nfs_writeback_release_full deal with the error.
1340 list_move(&failed
, head
);
1345 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds
);
1347 static void pnfs_ld_handle_write_error(struct nfs_write_data
*data
)
1349 struct nfs_pgio_header
*hdr
= data
->header
;
1351 dprintk("pnfs write error = %d\n", hdr
->pnfs_error
);
1352 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1353 PNFS_LAYOUTRET_ON_ERROR
) {
1354 clear_bit(NFS_INO_LAYOUTCOMMIT
, &NFS_I(hdr
->inode
)->flags
);
1355 pnfs_return_layout(hdr
->inode
);
1357 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1358 data
->task
.tk_status
= pnfs_write_done_resend_to_mds(hdr
->inode
,
1360 hdr
->completion_ops
);
1364 * Called by non rpc-based layout drivers
1366 void pnfs_ld_write_done(struct nfs_write_data
*data
)
1368 struct nfs_pgio_header
*hdr
= data
->header
;
1370 if (!hdr
->pnfs_error
) {
1371 pnfs_set_layoutcommit(data
);
1372 hdr
->mds_ops
->rpc_call_done(&data
->task
, data
);
1374 pnfs_ld_handle_write_error(data
);
1375 hdr
->mds_ops
->rpc_release(data
);
1377 EXPORT_SYMBOL_GPL(pnfs_ld_write_done
);
1380 pnfs_write_through_mds(struct nfs_pageio_descriptor
*desc
,
1381 struct nfs_write_data
*data
)
1383 struct nfs_pgio_header
*hdr
= data
->header
;
1385 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1386 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1387 nfs_pageio_reset_write_mds(desc
);
1388 desc
->pg_recoalesce
= 1;
1390 nfs_writedata_release(data
);
1393 static enum pnfs_try_status
1394 pnfs_try_to_write_data(struct nfs_write_data
*wdata
,
1395 const struct rpc_call_ops
*call_ops
,
1396 struct pnfs_layout_segment
*lseg
,
1399 struct nfs_pgio_header
*hdr
= wdata
->header
;
1400 struct inode
*inode
= hdr
->inode
;
1401 enum pnfs_try_status trypnfs
;
1402 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1404 hdr
->mds_ops
= call_ops
;
1406 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__
,
1407 inode
->i_ino
, wdata
->args
.count
, wdata
->args
.offset
, how
);
1408 trypnfs
= nfss
->pnfs_curr_ld
->write_pagelist(wdata
, how
);
1409 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1410 nfs_inc_stats(inode
, NFSIOS_PNFS_WRITE
);
1411 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1416 pnfs_do_multiple_writes(struct nfs_pageio_descriptor
*desc
, struct list_head
*head
, int how
)
1418 struct nfs_write_data
*data
;
1419 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1420 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1422 desc
->pg_lseg
= NULL
;
1423 while (!list_empty(head
)) {
1424 enum pnfs_try_status trypnfs
;
1426 data
= list_first_entry(head
, struct nfs_write_data
, list
);
1427 list_del_init(&data
->list
);
1429 trypnfs
= pnfs_try_to_write_data(data
, call_ops
, lseg
, how
);
1430 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1431 pnfs_write_through_mds(desc
, data
);
1433 pnfs_put_lseg(lseg
);
1436 static void pnfs_writehdr_free(struct nfs_pgio_header
*hdr
)
1438 pnfs_put_lseg(hdr
->lseg
);
1439 nfs_writehdr_free(hdr
);
1441 EXPORT_SYMBOL_GPL(pnfs_writehdr_free
);
1444 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor
*desc
)
1446 struct nfs_write_header
*whdr
;
1447 struct nfs_pgio_header
*hdr
;
1450 whdr
= nfs_writehdr_alloc();
1452 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1453 pnfs_put_lseg(desc
->pg_lseg
);
1454 desc
->pg_lseg
= NULL
;
1457 hdr
= &whdr
->header
;
1458 nfs_pgheader_init(desc
, hdr
, pnfs_writehdr_free
);
1459 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1460 atomic_inc(&hdr
->refcnt
);
1461 ret
= nfs_generic_flush(desc
, hdr
);
1463 pnfs_put_lseg(desc
->pg_lseg
);
1464 desc
->pg_lseg
= NULL
;
1466 pnfs_do_multiple_writes(desc
, &hdr
->rpc_list
, desc
->pg_ioflags
);
1467 if (atomic_dec_and_test(&hdr
->refcnt
))
1468 hdr
->completion_ops
->completion(hdr
);
1471 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages
);
1473 int pnfs_read_done_resend_to_mds(struct inode
*inode
,
1474 struct list_head
*head
,
1475 const struct nfs_pgio_completion_ops
*compl_ops
)
1477 struct nfs_pageio_descriptor pgio
;
1480 /* Resend all requests through the MDS */
1481 nfs_pageio_init_read(&pgio
, inode
, compl_ops
);
1482 while (!list_empty(head
)) {
1483 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1485 nfs_list_remove_request(req
);
1486 if (!nfs_pageio_add_request(&pgio
, req
))
1487 nfs_list_add_request(req
, &failed
);
1489 nfs_pageio_complete(&pgio
);
1491 if (!list_empty(&failed
)) {
1492 list_move(&failed
, head
);
1497 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds
);
1499 static void pnfs_ld_handle_read_error(struct nfs_read_data
*data
)
1501 struct nfs_pgio_header
*hdr
= data
->header
;
1503 dprintk("pnfs read error = %d\n", hdr
->pnfs_error
);
1504 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1505 PNFS_LAYOUTRET_ON_ERROR
) {
1506 clear_bit(NFS_INO_LAYOUTCOMMIT
, &NFS_I(hdr
->inode
)->flags
);
1507 pnfs_return_layout(hdr
->inode
);
1509 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1510 data
->task
.tk_status
= pnfs_read_done_resend_to_mds(hdr
->inode
,
1512 hdr
->completion_ops
);
1516 * Called by non rpc-based layout drivers
1518 void pnfs_ld_read_done(struct nfs_read_data
*data
)
1520 struct nfs_pgio_header
*hdr
= data
->header
;
1522 if (likely(!hdr
->pnfs_error
)) {
1523 __nfs4_read_done_cb(data
);
1524 hdr
->mds_ops
->rpc_call_done(&data
->task
, data
);
1526 pnfs_ld_handle_read_error(data
);
1527 hdr
->mds_ops
->rpc_release(data
);
1529 EXPORT_SYMBOL_GPL(pnfs_ld_read_done
);
1532 pnfs_read_through_mds(struct nfs_pageio_descriptor
*desc
,
1533 struct nfs_read_data
*data
)
1535 struct nfs_pgio_header
*hdr
= data
->header
;
1537 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1538 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1539 nfs_pageio_reset_read_mds(desc
);
1540 desc
->pg_recoalesce
= 1;
1542 nfs_readdata_release(data
);
1546 * Call the appropriate parallel I/O subsystem read function.
1548 static enum pnfs_try_status
1549 pnfs_try_to_read_data(struct nfs_read_data
*rdata
,
1550 const struct rpc_call_ops
*call_ops
,
1551 struct pnfs_layout_segment
*lseg
)
1553 struct nfs_pgio_header
*hdr
= rdata
->header
;
1554 struct inode
*inode
= hdr
->inode
;
1555 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1556 enum pnfs_try_status trypnfs
;
1558 hdr
->mds_ops
= call_ops
;
1560 dprintk("%s: Reading ino:%lu %u@%llu\n",
1561 __func__
, inode
->i_ino
, rdata
->args
.count
, rdata
->args
.offset
);
1563 trypnfs
= nfss
->pnfs_curr_ld
->read_pagelist(rdata
);
1564 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1565 nfs_inc_stats(inode
, NFSIOS_PNFS_READ
);
1566 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1571 pnfs_do_multiple_reads(struct nfs_pageio_descriptor
*desc
, struct list_head
*head
)
1573 struct nfs_read_data
*data
;
1574 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1575 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1577 desc
->pg_lseg
= NULL
;
1578 while (!list_empty(head
)) {
1579 enum pnfs_try_status trypnfs
;
1581 data
= list_first_entry(head
, struct nfs_read_data
, list
);
1582 list_del_init(&data
->list
);
1584 trypnfs
= pnfs_try_to_read_data(data
, call_ops
, lseg
);
1585 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1586 pnfs_read_through_mds(desc
, data
);
1588 pnfs_put_lseg(lseg
);
1591 static void pnfs_readhdr_free(struct nfs_pgio_header
*hdr
)
1593 pnfs_put_lseg(hdr
->lseg
);
1594 nfs_readhdr_free(hdr
);
1596 EXPORT_SYMBOL_GPL(pnfs_readhdr_free
);
1599 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor
*desc
)
1601 struct nfs_read_header
*rhdr
;
1602 struct nfs_pgio_header
*hdr
;
1605 rhdr
= nfs_readhdr_alloc();
1607 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1609 pnfs_put_lseg(desc
->pg_lseg
);
1610 desc
->pg_lseg
= NULL
;
1613 hdr
= &rhdr
->header
;
1614 nfs_pgheader_init(desc
, hdr
, pnfs_readhdr_free
);
1615 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1616 atomic_inc(&hdr
->refcnt
);
1617 ret
= nfs_generic_pagein(desc
, hdr
);
1619 pnfs_put_lseg(desc
->pg_lseg
);
1620 desc
->pg_lseg
= NULL
;
1622 pnfs_do_multiple_reads(desc
, &hdr
->rpc_list
);
1623 if (atomic_dec_and_test(&hdr
->refcnt
))
1624 hdr
->completion_ops
->completion(hdr
);
1627 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages
);
1630 * There can be multiple RW segments.
1632 static void pnfs_list_write_lseg(struct inode
*inode
, struct list_head
*listp
)
1634 struct pnfs_layout_segment
*lseg
;
1636 list_for_each_entry(lseg
, &NFS_I(inode
)->layout
->plh_segs
, pls_list
) {
1637 if (lseg
->pls_range
.iomode
== IOMODE_RW
&&
1638 test_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
1639 list_add(&lseg
->pls_lc_list
, listp
);
1643 void pnfs_set_lo_fail(struct pnfs_layout_segment
*lseg
)
1645 pnfs_layout_io_set_failed(lseg
->pls_layout
, lseg
->pls_range
.iomode
);
1647 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail
);
1650 pnfs_set_layoutcommit(struct nfs_write_data
*wdata
)
1652 struct nfs_pgio_header
*hdr
= wdata
->header
;
1653 struct inode
*inode
= hdr
->inode
;
1654 struct nfs_inode
*nfsi
= NFS_I(inode
);
1655 loff_t end_pos
= wdata
->mds_offset
+ wdata
->res
.count
;
1656 bool mark_as_dirty
= false;
1658 spin_lock(&inode
->i_lock
);
1659 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1660 mark_as_dirty
= true;
1661 dprintk("%s: Set layoutcommit for inode %lu ",
1662 __func__
, inode
->i_ino
);
1664 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT
, &hdr
->lseg
->pls_flags
)) {
1665 /* references matched in nfs4_layoutcommit_release */
1666 pnfs_get_lseg(hdr
->lseg
);
1668 if (end_pos
> nfsi
->layout
->plh_lwb
)
1669 nfsi
->layout
->plh_lwb
= end_pos
;
1670 spin_unlock(&inode
->i_lock
);
1671 dprintk("%s: lseg %p end_pos %llu\n",
1672 __func__
, hdr
->lseg
, nfsi
->layout
->plh_lwb
);
1674 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1675 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1677 mark_inode_dirty_sync(inode
);
1679 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit
);
1681 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data
*data
)
1683 struct nfs_server
*nfss
= NFS_SERVER(data
->args
.inode
);
1685 if (nfss
->pnfs_curr_ld
->cleanup_layoutcommit
)
1686 nfss
->pnfs_curr_ld
->cleanup_layoutcommit(data
);
1690 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1691 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1692 * data to disk to allow the server to recover the data if it crashes.
1693 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1694 * is off, and a COMMIT is sent to a data server, or
1695 * if WRITEs to a data server return NFS_DATA_SYNC.
1698 pnfs_layoutcommit_inode(struct inode
*inode
, bool sync
)
1700 struct nfs4_layoutcommit_data
*data
;
1701 struct nfs_inode
*nfsi
= NFS_I(inode
);
1705 dprintk("--> %s inode %lu\n", __func__
, inode
->i_ino
);
1707 if (!test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
1710 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1711 data
= kzalloc(sizeof(*data
), GFP_NOFS
);
1717 if (!test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
1720 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
)) {
1725 status
= wait_on_bit_lock(&nfsi
->flags
, NFS_INO_LAYOUTCOMMITTING
,
1726 nfs_wait_bit_killable
, TASK_KILLABLE
);
1731 INIT_LIST_HEAD(&data
->lseg_list
);
1732 spin_lock(&inode
->i_lock
);
1733 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1734 clear_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
);
1735 spin_unlock(&inode
->i_lock
);
1736 wake_up_bit(&nfsi
->flags
, NFS_INO_LAYOUTCOMMITTING
);
1740 pnfs_list_write_lseg(inode
, &data
->lseg_list
);
1742 end_pos
= nfsi
->layout
->plh_lwb
;
1743 nfsi
->layout
->plh_lwb
= 0;
1745 nfs4_stateid_copy(&data
->args
.stateid
, &nfsi
->layout
->plh_stateid
);
1746 spin_unlock(&inode
->i_lock
);
1748 data
->args
.inode
= inode
;
1749 data
->cred
= get_rpccred(nfsi
->layout
->plh_lc_cred
);
1750 nfs_fattr_init(&data
->fattr
);
1751 data
->args
.bitmask
= NFS_SERVER(inode
)->cache_consistency_bitmask
;
1752 data
->res
.fattr
= &data
->fattr
;
1753 data
->args
.lastbytewritten
= end_pos
- 1;
1754 data
->res
.server
= NFS_SERVER(inode
);
1756 status
= nfs4_proc_layoutcommit(data
, sync
);
1759 mark_inode_dirty_sync(inode
);
1760 dprintk("<-- %s status %d\n", __func__
, status
);
1767 struct nfs4_threshold
*pnfs_mdsthreshold_alloc(void)
1769 struct nfs4_threshold
*thp
;
1771 thp
= kzalloc(sizeof(*thp
), GFP_NOFS
);
1773 dprintk("%s mdsthreshold allocation failed\n", __func__
);