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 pnfs_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
);
326 pnfs_layout_remove_lseg(struct pnfs_layout_hdr
*lo
,
327 struct pnfs_layout_segment
*lseg
)
329 struct inode
*inode
= lo
->plh_inode
;
331 WARN_ON(test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
332 list_del_init(&lseg
->pls_list
);
333 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
334 atomic_dec(&lo
->plh_refcount
);
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
)) {
355 pnfs_get_layout_hdr(lo
);
356 pnfs_layout_remove_lseg(lo
, lseg
);
357 spin_unlock(&inode
->i_lock
);
358 pnfs_free_lseg(lseg
);
359 pnfs_put_layout_hdr(lo
);
362 EXPORT_SYMBOL_GPL(pnfs_put_lseg
);
365 end_offset(u64 start
, u64 len
)
370 return end
>= start
? end
: NFS4_MAX_UINT64
;
373 /* last octet in a range */
375 last_byte_offset(u64 start
, u64 len
)
381 return end
> start
? end
- 1 : NFS4_MAX_UINT64
;
385 * is l2 fully contained in l1?
387 * [----------------------------------)
392 lo_seg_contained(struct pnfs_layout_range
*l1
,
393 struct pnfs_layout_range
*l2
)
395 u64 start1
= l1
->offset
;
396 u64 end1
= end_offset(start1
, l1
->length
);
397 u64 start2
= l2
->offset
;
398 u64 end2
= end_offset(start2
, l2
->length
);
400 return (start1
<= start2
) && (end1
>= end2
);
404 * is l1 and l2 intersecting?
406 * [----------------------------------)
411 lo_seg_intersecting(struct pnfs_layout_range
*l1
,
412 struct pnfs_layout_range
*l2
)
414 u64 start1
= l1
->offset
;
415 u64 end1
= end_offset(start1
, l1
->length
);
416 u64 start2
= l2
->offset
;
417 u64 end2
= end_offset(start2
, l2
->length
);
419 return (end1
== NFS4_MAX_UINT64
|| end1
> start2
) &&
420 (end2
== NFS4_MAX_UINT64
|| end2
> start1
);
424 should_free_lseg(struct pnfs_layout_range
*lseg_range
,
425 struct pnfs_layout_range
*recall_range
)
427 return (recall_range
->iomode
== IOMODE_ANY
||
428 lseg_range
->iomode
== recall_range
->iomode
) &&
429 lo_seg_intersecting(lseg_range
, recall_range
);
432 /* Returns 1 if lseg is removed from list, 0 otherwise */
433 static int mark_lseg_invalid(struct pnfs_layout_segment
*lseg
,
434 struct list_head
*tmp_list
)
438 if (test_and_clear_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
)) {
439 /* Remove the reference keeping the lseg in the
440 * list. It will now be removed when all
441 * outstanding io is finished.
443 dprintk("%s: lseg %p ref %d\n", __func__
, lseg
,
444 atomic_read(&lseg
->pls_refcount
));
445 if (atomic_dec_and_test(&lseg
->pls_refcount
)) {
446 pnfs_layout_remove_lseg(lseg
->pls_layout
, lseg
);
447 list_add(&lseg
->pls_list
, tmp_list
);
454 /* Returns count of number of matching invalid lsegs remaining in list
458 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr
*lo
,
459 struct list_head
*tmp_list
,
460 struct pnfs_layout_range
*recall_range
)
462 struct pnfs_layout_segment
*lseg
, *next
;
463 int invalid
= 0, removed
= 0;
465 dprintk("%s:Begin lo %p\n", __func__
, lo
);
467 if (list_empty(&lo
->plh_segs
)) {
468 set_bit(NFS_LAYOUT_DESTROYED
, &lo
->plh_flags
);
471 list_for_each_entry_safe(lseg
, next
, &lo
->plh_segs
, pls_list
)
473 should_free_lseg(&lseg
->pls_range
, recall_range
)) {
474 dprintk("%s: freeing lseg %p iomode %d "
475 "offset %llu length %llu\n", __func__
,
476 lseg
, lseg
->pls_range
.iomode
, lseg
->pls_range
.offset
,
477 lseg
->pls_range
.length
);
479 removed
+= mark_lseg_invalid(lseg
, tmp_list
);
481 dprintk("%s:Return %i\n", __func__
, invalid
- removed
);
482 return invalid
- removed
;
485 /* note free_me must contain lsegs from a single layout_hdr */
487 pnfs_free_lseg_list(struct list_head
*free_me
)
489 struct pnfs_layout_segment
*lseg
, *tmp
;
491 if (list_empty(free_me
))
494 list_for_each_entry_safe(lseg
, tmp
, free_me
, pls_list
) {
495 list_del(&lseg
->pls_list
);
496 pnfs_free_lseg(lseg
);
501 pnfs_destroy_layout(struct nfs_inode
*nfsi
)
503 struct pnfs_layout_hdr
*lo
;
506 spin_lock(&nfsi
->vfs_inode
.i_lock
);
509 lo
->plh_block_lgets
++; /* permanently block new LAYOUTGETs */
510 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
511 pnfs_get_layout_hdr(lo
);
512 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RO_FAILED
);
513 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RW_FAILED
);
514 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
515 pnfs_free_lseg_list(&tmp_list
);
516 pnfs_put_layout_hdr(lo
);
518 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
520 EXPORT_SYMBOL_GPL(pnfs_destroy_layout
);
523 * Called by the state manger to remove all layouts established under an
527 pnfs_destroy_all_layouts(struct nfs_client
*clp
)
529 struct nfs_server
*server
;
530 struct pnfs_layout_hdr
*lo
;
533 nfs4_deviceid_mark_client_invalid(clp
);
534 nfs4_deviceid_purge_client(clp
);
536 spin_lock(&clp
->cl_lock
);
538 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
539 if (!list_empty(&server
->layouts
))
540 list_splice_init(&server
->layouts
, &tmp_list
);
543 spin_unlock(&clp
->cl_lock
);
545 while (!list_empty(&tmp_list
)) {
546 lo
= list_entry(tmp_list
.next
, struct pnfs_layout_hdr
,
548 dprintk("%s freeing layout for inode %lu\n", __func__
,
549 lo
->plh_inode
->i_ino
);
550 list_del_init(&lo
->plh_layouts
);
551 pnfs_destroy_layout(NFS_I(lo
->plh_inode
));
555 /* update lo->plh_stateid with new if is more recent */
557 pnfs_set_layout_stateid(struct pnfs_layout_hdr
*lo
, const nfs4_stateid
*new,
562 oldseq
= be32_to_cpu(lo
->plh_stateid
.seqid
);
563 newseq
= be32_to_cpu(new->seqid
);
564 if ((int)(newseq
- oldseq
) > 0) {
565 nfs4_stateid_copy(&lo
->plh_stateid
, new);
566 if (update_barrier
) {
567 u32 new_barrier
= be32_to_cpu(new->seqid
);
569 if ((int)(new_barrier
- lo
->plh_barrier
))
570 lo
->plh_barrier
= new_barrier
;
572 /* Because of wraparound, we want to keep the barrier
573 * "close" to the current seqids. It needs to be
574 * within 2**31 to count as "behind", so if it
575 * gets too near that limit, give us a litle leeway
576 * and bring it to within 2**30.
577 * NOTE - and yes, this is all unsigned arithmetic.
579 if (unlikely((newseq
- lo
->plh_barrier
) > (3 << 29)))
580 lo
->plh_barrier
= newseq
- (1 << 30);
585 /* lget is set to 1 if called from inside send_layoutget call chain */
587 pnfs_layoutgets_blocked(struct pnfs_layout_hdr
*lo
, nfs4_stateid
*stateid
,
591 (int)(lo
->plh_barrier
- be32_to_cpu(stateid
->seqid
)) >= 0)
593 return lo
->plh_block_lgets
||
594 test_bit(NFS_LAYOUT_DESTROYED
, &lo
->plh_flags
) ||
595 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
) ||
596 (list_empty(&lo
->plh_segs
) &&
597 (atomic_read(&lo
->plh_outstanding
) > lget
));
601 pnfs_choose_layoutget_stateid(nfs4_stateid
*dst
, struct pnfs_layout_hdr
*lo
,
602 struct nfs4_state
*open_state
)
606 dprintk("--> %s\n", __func__
);
607 spin_lock(&lo
->plh_inode
->i_lock
);
608 if (pnfs_layoutgets_blocked(lo
, NULL
, 1)) {
610 } else if (list_empty(&lo
->plh_segs
)) {
614 seq
= read_seqbegin(&open_state
->seqlock
);
615 nfs4_stateid_copy(dst
, &open_state
->stateid
);
616 } while (read_seqretry(&open_state
->seqlock
, seq
));
618 nfs4_stateid_copy(dst
, &lo
->plh_stateid
);
619 spin_unlock(&lo
->plh_inode
->i_lock
);
620 dprintk("<-- %s\n", __func__
);
625 * Get layout from server.
626 * for now, assume that whole file layouts are requested.
628 * arg->length: all ones
630 static struct pnfs_layout_segment
*
631 send_layoutget(struct pnfs_layout_hdr
*lo
,
632 struct nfs_open_context
*ctx
,
633 struct pnfs_layout_range
*range
,
636 struct inode
*ino
= lo
->plh_inode
;
637 struct nfs_server
*server
= NFS_SERVER(ino
);
638 struct nfs4_layoutget
*lgp
;
639 struct pnfs_layout_segment
*lseg
;
641 dprintk("--> %s\n", __func__
);
644 lgp
= kzalloc(sizeof(*lgp
), gfp_flags
);
648 lgp
->args
.minlength
= PAGE_CACHE_SIZE
;
649 if (lgp
->args
.minlength
> range
->length
)
650 lgp
->args
.minlength
= range
->length
;
651 lgp
->args
.maxcount
= PNFS_LAYOUT_MAXSIZE
;
652 lgp
->args
.range
= *range
;
653 lgp
->args
.type
= server
->pnfs_curr_ld
->id
;
654 lgp
->args
.inode
= ino
;
655 lgp
->args
.ctx
= get_nfs_open_context(ctx
);
656 lgp
->gfp_flags
= gfp_flags
;
658 /* Synchronously retrieve layout information from server and
661 lseg
= nfs4_proc_layoutget(lgp
, gfp_flags
);
663 switch (PTR_ERR(lseg
)) {
668 /* remember that LAYOUTGET failed and suspend trying */
669 pnfs_layout_io_set_failed(lo
, range
->iomode
);
678 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
679 * when the layout segment list is empty.
681 * Note that a pnfs_layout_hdr can exist with an empty layout segment
682 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
683 * deviceid is marked invalid.
686 _pnfs_return_layout(struct inode
*ino
)
688 struct pnfs_layout_hdr
*lo
= NULL
;
689 struct nfs_inode
*nfsi
= NFS_I(ino
);
691 struct nfs4_layoutreturn
*lrp
;
692 nfs4_stateid stateid
;
693 int status
= 0, empty
;
695 dprintk("NFS: %s for inode %lu\n", __func__
, ino
->i_ino
);
697 spin_lock(&ino
->i_lock
);
699 if (!lo
|| pnfs_test_layout_returned(lo
)) {
700 spin_unlock(&ino
->i_lock
);
701 dprintk("NFS: %s no layout to return\n", __func__
);
704 stateid
= nfsi
->layout
->plh_stateid
;
705 /* Reference matched in nfs4_layoutreturn_release */
706 pnfs_get_layout_hdr(lo
);
707 empty
= list_empty(&lo
->plh_segs
);
708 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
709 /* Don't send a LAYOUTRETURN if list was initially empty */
711 spin_unlock(&ino
->i_lock
);
712 pnfs_put_layout_hdr(lo
);
713 dprintk("NFS: %s no layout segments to return\n", __func__
);
716 lo
->plh_block_lgets
++;
717 pnfs_mark_layout_returned(lo
);
718 spin_unlock(&ino
->i_lock
);
719 pnfs_free_lseg_list(&tmp_list
);
721 WARN_ON(test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
));
723 lrp
= kzalloc(sizeof(*lrp
), GFP_KERNEL
);
724 if (unlikely(lrp
== NULL
)) {
726 pnfs_layout_io_set_failed(lo
, IOMODE_RW
);
727 pnfs_layout_io_set_failed(lo
, IOMODE_READ
);
728 pnfs_clear_layout_returned(lo
);
729 pnfs_put_layout_hdr(lo
);
733 lrp
->args
.stateid
= stateid
;
734 lrp
->args
.layout_type
= NFS_SERVER(ino
)->pnfs_curr_ld
->id
;
735 lrp
->args
.inode
= ino
;
736 lrp
->args
.layout
= lo
;
737 lrp
->clp
= NFS_SERVER(ino
)->nfs_client
;
739 status
= nfs4_proc_layoutreturn(lrp
);
741 dprintk("<-- %s status: %d\n", __func__
, status
);
744 EXPORT_SYMBOL_GPL(_pnfs_return_layout
);
746 bool pnfs_roc(struct inode
*ino
)
748 struct pnfs_layout_hdr
*lo
;
749 struct pnfs_layout_segment
*lseg
, *tmp
;
753 spin_lock(&ino
->i_lock
);
754 lo
= NFS_I(ino
)->layout
;
755 if (!lo
|| !test_and_clear_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
) ||
756 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
))
758 list_for_each_entry_safe(lseg
, tmp
, &lo
->plh_segs
, pls_list
)
759 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
760 mark_lseg_invalid(lseg
, &tmp_list
);
765 lo
->plh_block_lgets
++;
766 pnfs_get_layout_hdr(lo
); /* matched in pnfs_roc_release */
767 spin_unlock(&ino
->i_lock
);
768 pnfs_free_lseg_list(&tmp_list
);
772 spin_unlock(&ino
->i_lock
);
776 void pnfs_roc_release(struct inode
*ino
)
778 struct pnfs_layout_hdr
*lo
;
780 spin_lock(&ino
->i_lock
);
781 lo
= NFS_I(ino
)->layout
;
782 lo
->plh_block_lgets
--;
783 if (atomic_dec_and_test(&lo
->plh_refcount
)) {
784 pnfs_detach_layout_hdr(lo
);
785 spin_unlock(&ino
->i_lock
);
786 pnfs_free_layout_hdr(lo
);
788 spin_unlock(&ino
->i_lock
);
791 void pnfs_roc_set_barrier(struct inode
*ino
, u32 barrier
)
793 struct pnfs_layout_hdr
*lo
;
795 spin_lock(&ino
->i_lock
);
796 lo
= NFS_I(ino
)->layout
;
797 if ((int)(barrier
- lo
->plh_barrier
) > 0)
798 lo
->plh_barrier
= barrier
;
799 spin_unlock(&ino
->i_lock
);
802 bool pnfs_roc_drain(struct inode
*ino
, u32
*barrier
, struct rpc_task
*task
)
804 struct nfs_inode
*nfsi
= NFS_I(ino
);
805 struct pnfs_layout_hdr
*lo
;
806 struct pnfs_layout_segment
*lseg
;
810 spin_lock(&ino
->i_lock
);
811 list_for_each_entry(lseg
, &nfsi
->layout
->plh_segs
, pls_list
)
812 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
813 rpc_sleep_on(&NFS_SERVER(ino
)->roc_rpcwaitq
, task
, NULL
);
818 current_seqid
= be32_to_cpu(lo
->plh_stateid
.seqid
);
820 /* Since close does not return a layout stateid for use as
821 * a barrier, we choose the worst-case barrier.
823 *barrier
= current_seqid
+ atomic_read(&lo
->plh_outstanding
);
825 spin_unlock(&ino
->i_lock
);
830 * Compare two layout segments for sorting into layout cache.
831 * We want to preferentially return RW over RO layouts, so ensure those
835 cmp_layout(struct pnfs_layout_range
*l1
,
836 struct pnfs_layout_range
*l2
)
840 /* high offset > low offset */
841 d
= l1
->offset
- l2
->offset
;
845 /* short length > long length */
846 d
= l2
->length
- l1
->length
;
850 /* read > read/write */
851 return (int)(l1
->iomode
== IOMODE_READ
) - (int)(l2
->iomode
== IOMODE_READ
);
855 pnfs_layout_insert_lseg(struct pnfs_layout_hdr
*lo
,
856 struct pnfs_layout_segment
*lseg
)
858 struct pnfs_layout_segment
*lp
;
860 dprintk("%s:Begin\n", __func__
);
862 assert_spin_locked(&lo
->plh_inode
->i_lock
);
863 list_for_each_entry(lp
, &lo
->plh_segs
, pls_list
) {
864 if (cmp_layout(&lseg
->pls_range
, &lp
->pls_range
) > 0)
866 list_add_tail(&lseg
->pls_list
, &lp
->pls_list
);
867 dprintk("%s: inserted lseg %p "
868 "iomode %d offset %llu length %llu before "
869 "lp %p iomode %d offset %llu length %llu\n",
870 __func__
, lseg
, lseg
->pls_range
.iomode
,
871 lseg
->pls_range
.offset
, lseg
->pls_range
.length
,
872 lp
, lp
->pls_range
.iomode
, lp
->pls_range
.offset
,
873 lp
->pls_range
.length
);
876 list_add_tail(&lseg
->pls_list
, &lo
->plh_segs
);
877 dprintk("%s: inserted lseg %p "
878 "iomode %d offset %llu length %llu at tail\n",
879 __func__
, lseg
, lseg
->pls_range
.iomode
,
880 lseg
->pls_range
.offset
, lseg
->pls_range
.length
);
882 pnfs_get_layout_hdr(lo
);
884 dprintk("%s:Return\n", __func__
);
887 static struct pnfs_layout_hdr
*
888 alloc_init_layout_hdr(struct inode
*ino
,
889 struct nfs_open_context
*ctx
,
892 struct pnfs_layout_hdr
*lo
;
894 lo
= pnfs_alloc_layout_hdr(ino
, gfp_flags
);
897 atomic_set(&lo
->plh_refcount
, 1);
898 INIT_LIST_HEAD(&lo
->plh_layouts
);
899 INIT_LIST_HEAD(&lo
->plh_segs
);
900 INIT_LIST_HEAD(&lo
->plh_bulk_recall
);
902 lo
->plh_lc_cred
= get_rpccred(ctx
->state
->owner
->so_cred
);
906 static struct pnfs_layout_hdr
*
907 pnfs_find_alloc_layout(struct inode
*ino
,
908 struct nfs_open_context
*ctx
,
911 struct nfs_inode
*nfsi
= NFS_I(ino
);
912 struct pnfs_layout_hdr
*new = NULL
;
914 dprintk("%s Begin ino=%p layout=%p\n", __func__
, ino
, nfsi
->layout
);
916 assert_spin_locked(&ino
->i_lock
);
918 if (test_bit(NFS_LAYOUT_DESTROYED
, &nfsi
->layout
->plh_flags
))
920 pnfs_get_layout_hdr(nfsi
->layout
);
923 spin_unlock(&ino
->i_lock
);
924 new = alloc_init_layout_hdr(ino
, ctx
, gfp_flags
);
925 spin_lock(&ino
->i_lock
);
927 if (likely(nfsi
->layout
== NULL
)) /* Won the race? */
930 pnfs_free_layout_hdr(new);
935 * iomode matching rules:
946 is_matching_lseg(struct pnfs_layout_range
*ls_range
,
947 struct pnfs_layout_range
*range
)
949 struct pnfs_layout_range range1
;
951 if ((range
->iomode
== IOMODE_RW
&&
952 ls_range
->iomode
!= IOMODE_RW
) ||
953 !lo_seg_intersecting(ls_range
, range
))
956 /* range1 covers only the first byte in the range */
959 return lo_seg_contained(ls_range
, &range1
);
963 * lookup range in layout
965 static struct pnfs_layout_segment
*
966 pnfs_find_lseg(struct pnfs_layout_hdr
*lo
,
967 struct pnfs_layout_range
*range
)
969 struct pnfs_layout_segment
*lseg
, *ret
= NULL
;
971 dprintk("%s:Begin\n", __func__
);
973 assert_spin_locked(&lo
->plh_inode
->i_lock
);
974 list_for_each_entry(lseg
, &lo
->plh_segs
, pls_list
) {
975 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
) &&
976 is_matching_lseg(&lseg
->pls_range
, range
)) {
977 ret
= pnfs_get_lseg(lseg
);
980 if (lseg
->pls_range
.offset
> range
->offset
)
984 dprintk("%s:Return lseg %p ref %d\n",
985 __func__
, ret
, ret
? atomic_read(&ret
->pls_refcount
) : 0);
990 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
991 * to the MDS or over pNFS
993 * The nfs_inode read_io and write_io fields are cumulative counters reset
994 * when there are no layout segments. Note that in pnfs_update_layout iomode
995 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
998 * A return of true means use MDS I/O.
1001 * If a file's size is smaller than the file size threshold, data accesses
1002 * SHOULD be sent to the metadata server. If an I/O request has a length that
1003 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1004 * server. If both file size and I/O size are provided, the client SHOULD
1005 * reach or exceed both thresholds before sending its read or write
1006 * requests to the data server.
1008 static bool pnfs_within_mdsthreshold(struct nfs_open_context
*ctx
,
1009 struct inode
*ino
, int iomode
)
1011 struct nfs4_threshold
*t
= ctx
->mdsthreshold
;
1012 struct nfs_inode
*nfsi
= NFS_I(ino
);
1013 loff_t fsize
= i_size_read(ino
);
1014 bool size
= false, size_set
= false, io
= false, io_set
= false, ret
= false;
1019 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1020 __func__
, t
->bm
, t
->rd_sz
, t
->wr_sz
, t
->rd_io_sz
, t
->wr_io_sz
);
1024 if (t
->bm
& THRESHOLD_RD
) {
1025 dprintk("%s fsize %llu\n", __func__
, fsize
);
1027 if (fsize
< t
->rd_sz
)
1030 if (t
->bm
& THRESHOLD_RD_IO
) {
1031 dprintk("%s nfsi->read_io %llu\n", __func__
,
1034 if (nfsi
->read_io
< t
->rd_io_sz
)
1039 if (t
->bm
& THRESHOLD_WR
) {
1040 dprintk("%s fsize %llu\n", __func__
, fsize
);
1042 if (fsize
< t
->wr_sz
)
1045 if (t
->bm
& THRESHOLD_WR_IO
) {
1046 dprintk("%s nfsi->write_io %llu\n", __func__
,
1049 if (nfsi
->write_io
< t
->wr_io_sz
)
1054 if (size_set
&& io_set
) {
1057 } else if (size
|| io
)
1060 dprintk("<-- %s size %d io %d ret %d\n", __func__
, size
, io
, ret
);
1065 * Layout segment is retreived from the server if not cached.
1066 * The appropriate layout segment is referenced and returned to the caller.
1068 struct pnfs_layout_segment
*
1069 pnfs_update_layout(struct inode
*ino
,
1070 struct nfs_open_context
*ctx
,
1073 enum pnfs_iomode iomode
,
1076 struct pnfs_layout_range arg
= {
1082 struct nfs_server
*server
= NFS_SERVER(ino
);
1083 struct nfs_client
*clp
= server
->nfs_client
;
1084 struct pnfs_layout_hdr
*lo
;
1085 struct pnfs_layout_segment
*lseg
= NULL
;
1088 if (!pnfs_enabled_sb(NFS_SERVER(ino
)))
1091 if (pnfs_within_mdsthreshold(ctx
, ino
, iomode
))
1094 spin_lock(&ino
->i_lock
);
1095 lo
= pnfs_find_alloc_layout(ino
, ctx
, gfp_flags
);
1097 spin_unlock(&ino
->i_lock
);
1101 /* Do we even need to bother with this? */
1102 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1103 dprintk("%s matches recall, use MDS\n", __func__
);
1107 /* if LAYOUTGET already failed once we don't try again */
1108 if (pnfs_layout_io_test_failed(lo
, iomode
))
1111 /* Check to see if the layout for the given range already exists */
1112 lseg
= pnfs_find_lseg(lo
, &arg
);
1116 if (pnfs_layoutgets_blocked(lo
, NULL
, 0))
1118 atomic_inc(&lo
->plh_outstanding
);
1120 if (list_empty(&lo
->plh_segs
))
1123 /* Enable LAYOUTRETURNs */
1124 pnfs_clear_layout_returned(lo
);
1126 spin_unlock(&ino
->i_lock
);
1128 /* The lo must be on the clp list if there is any
1129 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1131 spin_lock(&clp
->cl_lock
);
1132 BUG_ON(!list_empty(&lo
->plh_layouts
));
1133 list_add_tail(&lo
->plh_layouts
, &server
->layouts
);
1134 spin_unlock(&clp
->cl_lock
);
1137 pg_offset
= arg
.offset
& ~PAGE_CACHE_MASK
;
1139 arg
.offset
-= pg_offset
;
1140 arg
.length
+= pg_offset
;
1142 if (arg
.length
!= NFS4_MAX_UINT64
)
1143 arg
.length
= PAGE_CACHE_ALIGN(arg
.length
);
1145 lseg
= send_layoutget(lo
, ctx
, &arg
, gfp_flags
);
1146 atomic_dec(&lo
->plh_outstanding
);
1148 pnfs_put_layout_hdr(lo
);
1150 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1151 "(%s, offset: %llu, length: %llu)\n",
1152 __func__
, ino
->i_sb
->s_id
,
1153 (unsigned long long)NFS_FILEID(ino
),
1154 lseg
== NULL
? "not found" : "found",
1155 iomode
==IOMODE_RW
? "read/write" : "read-only",
1156 (unsigned long long)pos
,
1157 (unsigned long long)count
);
1160 spin_unlock(&ino
->i_lock
);
1161 goto out_put_layout_hdr
;
1163 EXPORT_SYMBOL_GPL(pnfs_update_layout
);
1165 struct pnfs_layout_segment
*
1166 pnfs_layout_process(struct nfs4_layoutget
*lgp
)
1168 struct pnfs_layout_hdr
*lo
= NFS_I(lgp
->args
.inode
)->layout
;
1169 struct nfs4_layoutget_res
*res
= &lgp
->res
;
1170 struct pnfs_layout_segment
*lseg
;
1171 struct inode
*ino
= lo
->plh_inode
;
1174 /* Inject layout blob into I/O device driver */
1175 lseg
= NFS_SERVER(ino
)->pnfs_curr_ld
->alloc_lseg(lo
, res
, lgp
->gfp_flags
);
1176 if (!lseg
|| IS_ERR(lseg
)) {
1180 status
= PTR_ERR(lseg
);
1181 dprintk("%s: Could not allocate layout: error %d\n",
1186 spin_lock(&ino
->i_lock
);
1187 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1188 dprintk("%s forget reply due to recall\n", __func__
);
1189 goto out_forget_reply
;
1192 if (pnfs_layoutgets_blocked(lo
, &res
->stateid
, 1)) {
1193 dprintk("%s forget reply due to state\n", __func__
);
1194 goto out_forget_reply
;
1196 init_lseg(lo
, lseg
);
1197 lseg
->pls_range
= res
->range
;
1198 pnfs_get_lseg(lseg
);
1199 pnfs_layout_insert_lseg(lo
, lseg
);
1201 if (res
->return_on_close
) {
1202 set_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
);
1203 set_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
);
1206 /* Done processing layoutget. Set the layout stateid */
1207 pnfs_set_layout_stateid(lo
, &res
->stateid
, false);
1208 spin_unlock(&ino
->i_lock
);
1211 return ERR_PTR(status
);
1214 spin_unlock(&ino
->i_lock
);
1215 lseg
->pls_layout
= lo
;
1216 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
1221 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1223 BUG_ON(pgio
->pg_lseg
!= NULL
);
1225 if (req
->wb_offset
!= req
->wb_pgbase
) {
1226 nfs_pageio_reset_read_mds(pgio
);
1229 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1235 /* If no lseg, fall back to read through mds */
1236 if (pgio
->pg_lseg
== NULL
)
1237 nfs_pageio_reset_read_mds(pgio
);
1240 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read
);
1243 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1245 BUG_ON(pgio
->pg_lseg
!= NULL
);
1247 if (req
->wb_offset
!= req
->wb_pgbase
) {
1248 nfs_pageio_reset_write_mds(pgio
);
1251 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1257 /* If no lseg, fall back to write through mds */
1258 if (pgio
->pg_lseg
== NULL
)
1259 nfs_pageio_reset_write_mds(pgio
);
1261 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write
);
1264 pnfs_pageio_init_read(struct nfs_pageio_descriptor
*pgio
, struct inode
*inode
,
1265 const struct nfs_pgio_completion_ops
*compl_ops
)
1267 struct nfs_server
*server
= NFS_SERVER(inode
);
1268 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
1271 nfs_pageio_init_read(pgio
, inode
, compl_ops
);
1273 nfs_pageio_init(pgio
, inode
, ld
->pg_read_ops
, compl_ops
, server
->rsize
, 0);
1277 pnfs_pageio_init_write(struct nfs_pageio_descriptor
*pgio
, struct inode
*inode
,
1279 const struct nfs_pgio_completion_ops
*compl_ops
)
1281 struct nfs_server
*server
= NFS_SERVER(inode
);
1282 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
1285 nfs_pageio_init_write(pgio
, inode
, ioflags
, compl_ops
);
1287 nfs_pageio_init(pgio
, inode
, ld
->pg_write_ops
, compl_ops
, server
->wsize
, ioflags
);
1291 pnfs_generic_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
1292 struct nfs_page
*req
)
1294 if (pgio
->pg_lseg
== NULL
)
1295 return nfs_generic_pg_test(pgio
, prev
, req
);
1298 * Test if a nfs_page is fully contained in the pnfs_layout_range.
1299 * Note that this test makes several assumptions:
1300 * - that the previous nfs_page in the struct nfs_pageio_descriptor
1301 * is known to lie within the range.
1302 * - that the nfs_page being tested is known to be contiguous with the
1303 * previous nfs_page.
1304 * - Layout ranges are page aligned, so we only have to test the
1305 * start offset of the request.
1307 * Please also note that 'end_offset' is actually the offset of the
1308 * first byte that lies outside the pnfs_layout_range. FIXME?
1311 return req_offset(req
) < end_offset(pgio
->pg_lseg
->pls_range
.offset
,
1312 pgio
->pg_lseg
->pls_range
.length
);
1314 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test
);
1316 int pnfs_write_done_resend_to_mds(struct inode
*inode
,
1317 struct list_head
*head
,
1318 const struct nfs_pgio_completion_ops
*compl_ops
)
1320 struct nfs_pageio_descriptor pgio
;
1323 /* Resend all requests through the MDS */
1324 nfs_pageio_init_write(&pgio
, inode
, FLUSH_STABLE
, compl_ops
);
1325 while (!list_empty(head
)) {
1326 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1328 nfs_list_remove_request(req
);
1329 if (!nfs_pageio_add_request(&pgio
, req
))
1330 nfs_list_add_request(req
, &failed
);
1332 nfs_pageio_complete(&pgio
);
1334 if (!list_empty(&failed
)) {
1335 /* For some reason our attempt to resend pages. Mark the
1336 * overall send request as having failed, and let
1337 * nfs_writeback_release_full deal with the error.
1339 list_move(&failed
, head
);
1344 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds
);
1346 static void pnfs_ld_handle_write_error(struct nfs_write_data
*data
)
1348 struct nfs_pgio_header
*hdr
= data
->header
;
1350 dprintk("pnfs write error = %d\n", hdr
->pnfs_error
);
1351 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1352 PNFS_LAYOUTRET_ON_ERROR
) {
1353 clear_bit(NFS_INO_LAYOUTCOMMIT
, &NFS_I(hdr
->inode
)->flags
);
1354 pnfs_return_layout(hdr
->inode
);
1356 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1357 data
->task
.tk_status
= pnfs_write_done_resend_to_mds(hdr
->inode
,
1359 hdr
->completion_ops
);
1363 * Called by non rpc-based layout drivers
1365 void pnfs_ld_write_done(struct nfs_write_data
*data
)
1367 struct nfs_pgio_header
*hdr
= data
->header
;
1369 if (!hdr
->pnfs_error
) {
1370 pnfs_set_layoutcommit(data
);
1371 hdr
->mds_ops
->rpc_call_done(&data
->task
, data
);
1373 pnfs_ld_handle_write_error(data
);
1374 hdr
->mds_ops
->rpc_release(data
);
1376 EXPORT_SYMBOL_GPL(pnfs_ld_write_done
);
1379 pnfs_write_through_mds(struct nfs_pageio_descriptor
*desc
,
1380 struct nfs_write_data
*data
)
1382 struct nfs_pgio_header
*hdr
= data
->header
;
1384 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1385 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1386 nfs_pageio_reset_write_mds(desc
);
1387 desc
->pg_recoalesce
= 1;
1389 nfs_writedata_release(data
);
1392 static enum pnfs_try_status
1393 pnfs_try_to_write_data(struct nfs_write_data
*wdata
,
1394 const struct rpc_call_ops
*call_ops
,
1395 struct pnfs_layout_segment
*lseg
,
1398 struct nfs_pgio_header
*hdr
= wdata
->header
;
1399 struct inode
*inode
= hdr
->inode
;
1400 enum pnfs_try_status trypnfs
;
1401 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1403 hdr
->mds_ops
= call_ops
;
1405 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__
,
1406 inode
->i_ino
, wdata
->args
.count
, wdata
->args
.offset
, how
);
1407 trypnfs
= nfss
->pnfs_curr_ld
->write_pagelist(wdata
, how
);
1408 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1409 nfs_inc_stats(inode
, NFSIOS_PNFS_WRITE
);
1410 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1415 pnfs_do_multiple_writes(struct nfs_pageio_descriptor
*desc
, struct list_head
*head
, int how
)
1417 struct nfs_write_data
*data
;
1418 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1419 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1421 desc
->pg_lseg
= NULL
;
1422 while (!list_empty(head
)) {
1423 enum pnfs_try_status trypnfs
;
1425 data
= list_first_entry(head
, struct nfs_write_data
, list
);
1426 list_del_init(&data
->list
);
1428 trypnfs
= pnfs_try_to_write_data(data
, call_ops
, lseg
, how
);
1429 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1430 pnfs_write_through_mds(desc
, data
);
1432 pnfs_put_lseg(lseg
);
1435 static void pnfs_writehdr_free(struct nfs_pgio_header
*hdr
)
1437 pnfs_put_lseg(hdr
->lseg
);
1438 nfs_writehdr_free(hdr
);
1440 EXPORT_SYMBOL_GPL(pnfs_writehdr_free
);
1443 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor
*desc
)
1445 struct nfs_write_header
*whdr
;
1446 struct nfs_pgio_header
*hdr
;
1449 whdr
= nfs_writehdr_alloc();
1451 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1452 pnfs_put_lseg(desc
->pg_lseg
);
1453 desc
->pg_lseg
= NULL
;
1456 hdr
= &whdr
->header
;
1457 nfs_pgheader_init(desc
, hdr
, pnfs_writehdr_free
);
1458 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1459 atomic_inc(&hdr
->refcnt
);
1460 ret
= nfs_generic_flush(desc
, hdr
);
1462 pnfs_put_lseg(desc
->pg_lseg
);
1463 desc
->pg_lseg
= NULL
;
1465 pnfs_do_multiple_writes(desc
, &hdr
->rpc_list
, desc
->pg_ioflags
);
1466 if (atomic_dec_and_test(&hdr
->refcnt
))
1467 hdr
->completion_ops
->completion(hdr
);
1470 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages
);
1472 int pnfs_read_done_resend_to_mds(struct inode
*inode
,
1473 struct list_head
*head
,
1474 const struct nfs_pgio_completion_ops
*compl_ops
)
1476 struct nfs_pageio_descriptor pgio
;
1479 /* Resend all requests through the MDS */
1480 nfs_pageio_init_read(&pgio
, inode
, compl_ops
);
1481 while (!list_empty(head
)) {
1482 struct nfs_page
*req
= nfs_list_entry(head
->next
);
1484 nfs_list_remove_request(req
);
1485 if (!nfs_pageio_add_request(&pgio
, req
))
1486 nfs_list_add_request(req
, &failed
);
1488 nfs_pageio_complete(&pgio
);
1490 if (!list_empty(&failed
)) {
1491 list_move(&failed
, head
);
1496 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds
);
1498 static void pnfs_ld_handle_read_error(struct nfs_read_data
*data
)
1500 struct nfs_pgio_header
*hdr
= data
->header
;
1502 dprintk("pnfs read error = %d\n", hdr
->pnfs_error
);
1503 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1504 PNFS_LAYOUTRET_ON_ERROR
) {
1505 clear_bit(NFS_INO_LAYOUTCOMMIT
, &NFS_I(hdr
->inode
)->flags
);
1506 pnfs_return_layout(hdr
->inode
);
1508 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1509 data
->task
.tk_status
= pnfs_read_done_resend_to_mds(hdr
->inode
,
1511 hdr
->completion_ops
);
1515 * Called by non rpc-based layout drivers
1517 void pnfs_ld_read_done(struct nfs_read_data
*data
)
1519 struct nfs_pgio_header
*hdr
= data
->header
;
1521 if (likely(!hdr
->pnfs_error
)) {
1522 __nfs4_read_done_cb(data
);
1523 hdr
->mds_ops
->rpc_call_done(&data
->task
, data
);
1525 pnfs_ld_handle_read_error(data
);
1526 hdr
->mds_ops
->rpc_release(data
);
1528 EXPORT_SYMBOL_GPL(pnfs_ld_read_done
);
1531 pnfs_read_through_mds(struct nfs_pageio_descriptor
*desc
,
1532 struct nfs_read_data
*data
)
1534 struct nfs_pgio_header
*hdr
= data
->header
;
1536 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1537 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1538 nfs_pageio_reset_read_mds(desc
);
1539 desc
->pg_recoalesce
= 1;
1541 nfs_readdata_release(data
);
1545 * Call the appropriate parallel I/O subsystem read function.
1547 static enum pnfs_try_status
1548 pnfs_try_to_read_data(struct nfs_read_data
*rdata
,
1549 const struct rpc_call_ops
*call_ops
,
1550 struct pnfs_layout_segment
*lseg
)
1552 struct nfs_pgio_header
*hdr
= rdata
->header
;
1553 struct inode
*inode
= hdr
->inode
;
1554 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1555 enum pnfs_try_status trypnfs
;
1557 hdr
->mds_ops
= call_ops
;
1559 dprintk("%s: Reading ino:%lu %u@%llu\n",
1560 __func__
, inode
->i_ino
, rdata
->args
.count
, rdata
->args
.offset
);
1562 trypnfs
= nfss
->pnfs_curr_ld
->read_pagelist(rdata
);
1563 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1564 nfs_inc_stats(inode
, NFSIOS_PNFS_READ
);
1565 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1570 pnfs_do_multiple_reads(struct nfs_pageio_descriptor
*desc
, struct list_head
*head
)
1572 struct nfs_read_data
*data
;
1573 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1574 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1576 desc
->pg_lseg
= NULL
;
1577 while (!list_empty(head
)) {
1578 enum pnfs_try_status trypnfs
;
1580 data
= list_first_entry(head
, struct nfs_read_data
, list
);
1581 list_del_init(&data
->list
);
1583 trypnfs
= pnfs_try_to_read_data(data
, call_ops
, lseg
);
1584 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1585 pnfs_read_through_mds(desc
, data
);
1587 pnfs_put_lseg(lseg
);
1590 static void pnfs_readhdr_free(struct nfs_pgio_header
*hdr
)
1592 pnfs_put_lseg(hdr
->lseg
);
1593 nfs_readhdr_free(hdr
);
1595 EXPORT_SYMBOL_GPL(pnfs_readhdr_free
);
1598 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor
*desc
)
1600 struct nfs_read_header
*rhdr
;
1601 struct nfs_pgio_header
*hdr
;
1604 rhdr
= nfs_readhdr_alloc();
1606 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1608 pnfs_put_lseg(desc
->pg_lseg
);
1609 desc
->pg_lseg
= NULL
;
1612 hdr
= &rhdr
->header
;
1613 nfs_pgheader_init(desc
, hdr
, pnfs_readhdr_free
);
1614 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1615 atomic_inc(&hdr
->refcnt
);
1616 ret
= nfs_generic_pagein(desc
, hdr
);
1618 pnfs_put_lseg(desc
->pg_lseg
);
1619 desc
->pg_lseg
= NULL
;
1621 pnfs_do_multiple_reads(desc
, &hdr
->rpc_list
);
1622 if (atomic_dec_and_test(&hdr
->refcnt
))
1623 hdr
->completion_ops
->completion(hdr
);
1626 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages
);
1629 * There can be multiple RW segments.
1631 static void pnfs_list_write_lseg(struct inode
*inode
, struct list_head
*listp
)
1633 struct pnfs_layout_segment
*lseg
;
1635 list_for_each_entry(lseg
, &NFS_I(inode
)->layout
->plh_segs
, pls_list
) {
1636 if (lseg
->pls_range
.iomode
== IOMODE_RW
&&
1637 test_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
1638 list_add(&lseg
->pls_lc_list
, listp
);
1642 void pnfs_set_lo_fail(struct pnfs_layout_segment
*lseg
)
1644 pnfs_layout_io_set_failed(lseg
->pls_layout
, lseg
->pls_range
.iomode
);
1646 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail
);
1649 pnfs_set_layoutcommit(struct nfs_write_data
*wdata
)
1651 struct nfs_pgio_header
*hdr
= wdata
->header
;
1652 struct inode
*inode
= hdr
->inode
;
1653 struct nfs_inode
*nfsi
= NFS_I(inode
);
1654 loff_t end_pos
= wdata
->mds_offset
+ wdata
->res
.count
;
1655 bool mark_as_dirty
= false;
1657 spin_lock(&inode
->i_lock
);
1658 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1659 mark_as_dirty
= true;
1660 dprintk("%s: Set layoutcommit for inode %lu ",
1661 __func__
, inode
->i_ino
);
1663 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT
, &hdr
->lseg
->pls_flags
)) {
1664 /* references matched in nfs4_layoutcommit_release */
1665 pnfs_get_lseg(hdr
->lseg
);
1667 if (end_pos
> nfsi
->layout
->plh_lwb
)
1668 nfsi
->layout
->plh_lwb
= end_pos
;
1669 spin_unlock(&inode
->i_lock
);
1670 dprintk("%s: lseg %p end_pos %llu\n",
1671 __func__
, hdr
->lseg
, nfsi
->layout
->plh_lwb
);
1673 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1674 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1676 mark_inode_dirty_sync(inode
);
1678 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit
);
1680 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data
*data
)
1682 struct nfs_server
*nfss
= NFS_SERVER(data
->args
.inode
);
1684 if (nfss
->pnfs_curr_ld
->cleanup_layoutcommit
)
1685 nfss
->pnfs_curr_ld
->cleanup_layoutcommit(data
);
1689 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1690 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1691 * data to disk to allow the server to recover the data if it crashes.
1692 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1693 * is off, and a COMMIT is sent to a data server, or
1694 * if WRITEs to a data server return NFS_DATA_SYNC.
1697 pnfs_layoutcommit_inode(struct inode
*inode
, bool sync
)
1699 struct nfs4_layoutcommit_data
*data
;
1700 struct nfs_inode
*nfsi
= NFS_I(inode
);
1704 dprintk("--> %s inode %lu\n", __func__
, inode
->i_ino
);
1706 if (!test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
1709 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1710 data
= kzalloc(sizeof(*data
), GFP_NOFS
);
1716 if (!test_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
1719 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
)) {
1724 status
= wait_on_bit_lock(&nfsi
->flags
, NFS_INO_LAYOUTCOMMITTING
,
1725 nfs_wait_bit_killable
, TASK_KILLABLE
);
1730 INIT_LIST_HEAD(&data
->lseg_list
);
1731 spin_lock(&inode
->i_lock
);
1732 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1733 clear_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
);
1734 spin_unlock(&inode
->i_lock
);
1735 wake_up_bit(&nfsi
->flags
, NFS_INO_LAYOUTCOMMITTING
);
1739 pnfs_list_write_lseg(inode
, &data
->lseg_list
);
1741 end_pos
= nfsi
->layout
->plh_lwb
;
1742 nfsi
->layout
->plh_lwb
= 0;
1744 nfs4_stateid_copy(&data
->args
.stateid
, &nfsi
->layout
->plh_stateid
);
1745 spin_unlock(&inode
->i_lock
);
1747 data
->args
.inode
= inode
;
1748 data
->cred
= get_rpccred(nfsi
->layout
->plh_lc_cred
);
1749 nfs_fattr_init(&data
->fattr
);
1750 data
->args
.bitmask
= NFS_SERVER(inode
)->cache_consistency_bitmask
;
1751 data
->res
.fattr
= &data
->fattr
;
1752 data
->args
.lastbytewritten
= end_pos
- 1;
1753 data
->res
.server
= NFS_SERVER(inode
);
1755 status
= nfs4_proc_layoutcommit(data
, sync
);
1758 mark_inode_dirty_sync(inode
);
1759 dprintk("<-- %s status %d\n", __func__
, status
);
1766 struct nfs4_threshold
*pnfs_mdsthreshold_alloc(void)
1768 struct nfs4_threshold
*thp
;
1770 thp
= kzalloc(sizeof(*thp
), GFP_NOFS
);
1772 dprintk("%s mdsthreshold allocation failed\n", __func__
);