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>
33 #include <linux/sort.h>
37 #include "nfs4trace.h"
38 #include "delegation.h"
41 #define NFSDBG_FACILITY NFSDBG_PNFS
42 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
47 * protects pnfs_modules_tbl.
49 static DEFINE_SPINLOCK(pnfs_spinlock
);
52 * pnfs_modules_tbl holds all pnfs modules
54 static LIST_HEAD(pnfs_modules_tbl
);
56 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr
*lo
);
58 /* Return the registered pnfs layout driver module matching given id */
59 static struct pnfs_layoutdriver_type
*
60 find_pnfs_driver_locked(u32 id
)
62 struct pnfs_layoutdriver_type
*local
;
64 list_for_each_entry(local
, &pnfs_modules_tbl
, pnfs_tblid
)
69 dprintk("%s: Searching for id %u, found %p\n", __func__
, id
, local
);
73 static struct pnfs_layoutdriver_type
*
74 find_pnfs_driver(u32 id
)
76 struct pnfs_layoutdriver_type
*local
;
78 spin_lock(&pnfs_spinlock
);
79 local
= find_pnfs_driver_locked(id
);
80 if (local
!= NULL
&& !try_module_get(local
->owner
)) {
81 dprintk("%s: Could not grab reference on module\n", __func__
);
84 spin_unlock(&pnfs_spinlock
);
89 unset_pnfs_layoutdriver(struct nfs_server
*nfss
)
91 if (nfss
->pnfs_curr_ld
) {
92 if (nfss
->pnfs_curr_ld
->clear_layoutdriver
)
93 nfss
->pnfs_curr_ld
->clear_layoutdriver(nfss
);
94 /* Decrement the MDS count. Purge the deviceid cache if zero */
95 if (atomic_dec_and_test(&nfss
->nfs_client
->cl_mds_count
))
96 nfs4_deviceid_purge_client(nfss
->nfs_client
);
97 module_put(nfss
->pnfs_curr_ld
->owner
);
99 nfss
->pnfs_curr_ld
= NULL
;
103 * When the server sends a list of layout types, we choose one in the order
104 * given in the list below.
106 * FIXME: should this list be configurable in some fashion? module param?
107 * mount option? something else?
109 static const u32 ld_prefs
[] = {
114 LAYOUT_NFSV4_1_FILES
,
119 ld_cmp(const void *e1
, const void *e2
)
121 u32 ld1
= *((u32
*)e1
);
122 u32 ld2
= *((u32
*)e2
);
125 for (i
= 0; ld_prefs
[i
] != 0; i
++) {
126 if (ld1
== ld_prefs
[i
])
129 if (ld2
== ld_prefs
[i
])
136 * Try to set the server's pnfs module to the pnfs layout type specified by id.
137 * Currently only one pNFS layout driver per filesystem is supported.
139 * @ids array of layout types supported by MDS.
142 set_pnfs_layoutdriver(struct nfs_server
*server
, const struct nfs_fh
*mntfh
,
143 struct nfs_fsinfo
*fsinfo
)
145 struct pnfs_layoutdriver_type
*ld_type
= NULL
;
149 if (fsinfo
->nlayouttypes
== 0)
151 if (!(server
->nfs_client
->cl_exchange_flags
&
152 (EXCHGID4_FLAG_USE_NON_PNFS
| EXCHGID4_FLAG_USE_PNFS_MDS
))) {
153 printk(KERN_ERR
"NFS: %s: cl_exchange_flags 0x%x\n",
154 __func__
, server
->nfs_client
->cl_exchange_flags
);
158 sort(fsinfo
->layouttype
, fsinfo
->nlayouttypes
,
159 sizeof(*fsinfo
->layouttype
), ld_cmp
, NULL
);
161 for (i
= 0; i
< fsinfo
->nlayouttypes
; i
++) {
162 id
= fsinfo
->layouttype
[i
];
163 ld_type
= find_pnfs_driver(id
);
165 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX
,
167 ld_type
= find_pnfs_driver(id
);
174 dprintk("%s: No pNFS module found!\n", __func__
);
178 server
->pnfs_curr_ld
= ld_type
;
179 if (ld_type
->set_layoutdriver
180 && ld_type
->set_layoutdriver(server
, mntfh
)) {
181 printk(KERN_ERR
"NFS: %s: Error initializing pNFS layout "
182 "driver %u.\n", __func__
, id
);
183 module_put(ld_type
->owner
);
186 /* Bump the MDS count */
187 atomic_inc(&server
->nfs_client
->cl_mds_count
);
189 dprintk("%s: pNFS module for %u set\n", __func__
, id
);
193 dprintk("%s: Using NFSv4 I/O\n", __func__
);
194 server
->pnfs_curr_ld
= NULL
;
198 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
200 int status
= -EINVAL
;
201 struct pnfs_layoutdriver_type
*tmp
;
203 if (ld_type
->id
== 0) {
204 printk(KERN_ERR
"NFS: %s id 0 is reserved\n", __func__
);
207 if (!ld_type
->alloc_lseg
|| !ld_type
->free_lseg
) {
208 printk(KERN_ERR
"NFS: %s Layout driver must provide "
209 "alloc_lseg and free_lseg.\n", __func__
);
213 spin_lock(&pnfs_spinlock
);
214 tmp
= find_pnfs_driver_locked(ld_type
->id
);
216 list_add(&ld_type
->pnfs_tblid
, &pnfs_modules_tbl
);
218 dprintk("%s Registering id:%u name:%s\n", __func__
, ld_type
->id
,
221 printk(KERN_ERR
"NFS: %s Module with id %d already loaded!\n",
222 __func__
, ld_type
->id
);
224 spin_unlock(&pnfs_spinlock
);
228 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver
);
231 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
233 dprintk("%s Deregistering id:%u\n", __func__
, ld_type
->id
);
234 spin_lock(&pnfs_spinlock
);
235 list_del(&ld_type
->pnfs_tblid
);
236 spin_unlock(&pnfs_spinlock
);
238 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver
);
241 * pNFS client layout cache
244 /* Need to hold i_lock if caller does not already hold reference */
246 pnfs_get_layout_hdr(struct pnfs_layout_hdr
*lo
)
248 atomic_inc(&lo
->plh_refcount
);
251 static struct pnfs_layout_hdr
*
252 pnfs_alloc_layout_hdr(struct inode
*ino
, gfp_t gfp_flags
)
254 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(ino
)->pnfs_curr_ld
;
255 return ld
->alloc_layout_hdr(ino
, gfp_flags
);
259 pnfs_free_layout_hdr(struct pnfs_layout_hdr
*lo
)
261 struct nfs_server
*server
= NFS_SERVER(lo
->plh_inode
);
262 struct pnfs_layoutdriver_type
*ld
= server
->pnfs_curr_ld
;
264 if (!list_empty(&lo
->plh_layouts
)) {
265 struct nfs_client
*clp
= server
->nfs_client
;
267 spin_lock(&clp
->cl_lock
);
268 list_del_init(&lo
->plh_layouts
);
269 spin_unlock(&clp
->cl_lock
);
271 put_rpccred(lo
->plh_lc_cred
);
272 return ld
->free_layout_hdr(lo
);
276 pnfs_detach_layout_hdr(struct pnfs_layout_hdr
*lo
)
278 struct nfs_inode
*nfsi
= NFS_I(lo
->plh_inode
);
279 dprintk("%s: freeing layout cache %p\n", __func__
, lo
);
281 /* Reset MDS Threshold I/O counters */
287 pnfs_put_layout_hdr(struct pnfs_layout_hdr
*lo
)
289 struct inode
*inode
= lo
->plh_inode
;
291 pnfs_layoutreturn_before_put_layout_hdr(lo
);
293 if (atomic_dec_and_lock(&lo
->plh_refcount
, &inode
->i_lock
)) {
294 if (!list_empty(&lo
->plh_segs
))
295 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
296 pnfs_detach_layout_hdr(lo
);
297 spin_unlock(&inode
->i_lock
);
298 pnfs_free_layout_hdr(lo
);
303 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
305 * In order to continue using the pnfs_layout_hdr, a full recovery
307 * Note that caller must hold inode->i_lock.
310 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr
*lo
,
311 struct list_head
*lseg_list
)
313 struct pnfs_layout_range range
= {
314 .iomode
= IOMODE_ANY
,
316 .length
= NFS4_MAX_UINT64
,
319 set_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
);
320 return pnfs_mark_matching_lsegs_invalid(lo
, lseg_list
, &range
, 0);
324 pnfs_iomode_to_fail_bit(u32 iomode
)
326 return iomode
== IOMODE_RW
?
327 NFS_LAYOUT_RW_FAILED
: NFS_LAYOUT_RO_FAILED
;
331 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr
*lo
, int fail_bit
)
333 lo
->plh_retry_timestamp
= jiffies
;
334 if (!test_and_set_bit(fail_bit
, &lo
->plh_flags
))
335 atomic_inc(&lo
->plh_refcount
);
339 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr
*lo
, int fail_bit
)
341 if (test_and_clear_bit(fail_bit
, &lo
->plh_flags
))
342 atomic_dec(&lo
->plh_refcount
);
346 pnfs_layout_io_set_failed(struct pnfs_layout_hdr
*lo
, u32 iomode
)
348 struct inode
*inode
= lo
->plh_inode
;
349 struct pnfs_layout_range range
= {
352 .length
= NFS4_MAX_UINT64
,
356 spin_lock(&inode
->i_lock
);
357 pnfs_layout_set_fail_bit(lo
, pnfs_iomode_to_fail_bit(iomode
));
358 pnfs_mark_matching_lsegs_invalid(lo
, &head
, &range
, 0);
359 spin_unlock(&inode
->i_lock
);
360 pnfs_free_lseg_list(&head
);
361 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__
,
362 iomode
== IOMODE_RW
? "RW" : "READ");
366 pnfs_layout_io_test_failed(struct pnfs_layout_hdr
*lo
, u32 iomode
)
368 unsigned long start
, end
;
369 int fail_bit
= pnfs_iomode_to_fail_bit(iomode
);
371 if (test_bit(fail_bit
, &lo
->plh_flags
) == 0)
374 start
= end
- PNFS_LAYOUTGET_RETRY_TIMEOUT
;
375 if (!time_in_range(lo
->plh_retry_timestamp
, start
, end
)) {
376 /* It is time to retry the failed layoutgets */
377 pnfs_layout_clear_fail_bit(lo
, fail_bit
);
384 pnfs_init_lseg(struct pnfs_layout_hdr
*lo
, struct pnfs_layout_segment
*lseg
,
385 const struct pnfs_layout_range
*range
,
386 const nfs4_stateid
*stateid
)
388 INIT_LIST_HEAD(&lseg
->pls_list
);
389 INIT_LIST_HEAD(&lseg
->pls_lc_list
);
390 atomic_set(&lseg
->pls_refcount
, 1);
391 set_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
);
392 lseg
->pls_layout
= lo
;
393 lseg
->pls_range
= *range
;
394 lseg
->pls_seq
= be32_to_cpu(stateid
->seqid
);
397 static void pnfs_free_lseg(struct pnfs_layout_segment
*lseg
)
399 struct inode
*ino
= lseg
->pls_layout
->plh_inode
;
401 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
405 pnfs_layout_remove_lseg(struct pnfs_layout_hdr
*lo
,
406 struct pnfs_layout_segment
*lseg
)
408 struct inode
*inode
= lo
->plh_inode
;
410 WARN_ON(test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
411 list_del_init(&lseg
->pls_list
);
412 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
413 atomic_dec(&lo
->plh_refcount
);
414 if (list_empty(&lo
->plh_segs
)) {
415 if (atomic_read(&lo
->plh_outstanding
) == 0)
416 set_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
);
417 clear_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
);
419 rpc_wake_up(&NFS_SERVER(inode
)->roc_rpcwaitq
);
423 pnfs_put_lseg(struct pnfs_layout_segment
*lseg
)
425 struct pnfs_layout_hdr
*lo
;
431 dprintk("%s: lseg %p ref %d valid %d\n", __func__
, lseg
,
432 atomic_read(&lseg
->pls_refcount
),
433 test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
435 lo
= lseg
->pls_layout
;
436 inode
= lo
->plh_inode
;
438 if (atomic_dec_and_lock(&lseg
->pls_refcount
, &inode
->i_lock
)) {
439 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
)) {
440 spin_unlock(&inode
->i_lock
);
443 pnfs_get_layout_hdr(lo
);
444 pnfs_layout_remove_lseg(lo
, lseg
);
445 spin_unlock(&inode
->i_lock
);
446 pnfs_free_lseg(lseg
);
447 pnfs_put_layout_hdr(lo
);
450 EXPORT_SYMBOL_GPL(pnfs_put_lseg
);
452 static void pnfs_free_lseg_async_work(struct work_struct
*work
)
454 struct pnfs_layout_segment
*lseg
;
455 struct pnfs_layout_hdr
*lo
;
457 lseg
= container_of(work
, struct pnfs_layout_segment
, pls_work
);
458 lo
= lseg
->pls_layout
;
460 pnfs_free_lseg(lseg
);
461 pnfs_put_layout_hdr(lo
);
464 static void pnfs_free_lseg_async(struct pnfs_layout_segment
*lseg
)
466 INIT_WORK(&lseg
->pls_work
, pnfs_free_lseg_async_work
);
467 schedule_work(&lseg
->pls_work
);
471 pnfs_put_lseg_locked(struct pnfs_layout_segment
*lseg
)
476 assert_spin_locked(&lseg
->pls_layout
->plh_inode
->i_lock
);
478 dprintk("%s: lseg %p ref %d valid %d\n", __func__
, lseg
,
479 atomic_read(&lseg
->pls_refcount
),
480 test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
481 if (atomic_dec_and_test(&lseg
->pls_refcount
)) {
482 struct pnfs_layout_hdr
*lo
= lseg
->pls_layout
;
483 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
))
485 pnfs_get_layout_hdr(lo
);
486 pnfs_layout_remove_lseg(lo
, lseg
);
487 pnfs_free_lseg_async(lseg
);
490 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked
);
493 end_offset(u64 start
, u64 len
)
498 return end
>= start
? end
: NFS4_MAX_UINT64
;
502 * is l2 fully contained in l1?
504 * [----------------------------------)
509 pnfs_lseg_range_contained(const struct pnfs_layout_range
*l1
,
510 const struct pnfs_layout_range
*l2
)
512 u64 start1
= l1
->offset
;
513 u64 end1
= end_offset(start1
, l1
->length
);
514 u64 start2
= l2
->offset
;
515 u64 end2
= end_offset(start2
, l2
->length
);
517 return (start1
<= start2
) && (end1
>= end2
);
521 * is l1 and l2 intersecting?
523 * [----------------------------------)
528 pnfs_lseg_range_intersecting(const struct pnfs_layout_range
*l1
,
529 const struct pnfs_layout_range
*l2
)
531 u64 start1
= l1
->offset
;
532 u64 end1
= end_offset(start1
, l1
->length
);
533 u64 start2
= l2
->offset
;
534 u64 end2
= end_offset(start2
, l2
->length
);
536 return (end1
== NFS4_MAX_UINT64
|| end1
> start2
) &&
537 (end2
== NFS4_MAX_UINT64
|| end2
> start1
);
540 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment
*lseg
,
541 struct list_head
*tmp_list
)
543 if (!atomic_dec_and_test(&lseg
->pls_refcount
))
545 pnfs_layout_remove_lseg(lseg
->pls_layout
, lseg
);
546 list_add(&lseg
->pls_list
, tmp_list
);
550 /* Returns 1 if lseg is removed from list, 0 otherwise */
551 static int mark_lseg_invalid(struct pnfs_layout_segment
*lseg
,
552 struct list_head
*tmp_list
)
556 if (test_and_clear_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
)) {
557 /* Remove the reference keeping the lseg in the
558 * list. It will now be removed when all
559 * outstanding io is finished.
561 dprintk("%s: lseg %p ref %d\n", __func__
, lseg
,
562 atomic_read(&lseg
->pls_refcount
));
563 if (pnfs_lseg_dec_and_remove_zero(lseg
, tmp_list
))
570 * Compare 2 layout stateid sequence ids, to see which is newer,
571 * taking into account wraparound issues.
573 static bool pnfs_seqid_is_newer(u32 s1
, u32 s2
)
575 return (s32
)(s1
- s2
) > 0;
579 pnfs_should_free_range(const struct pnfs_layout_range
*lseg_range
,
580 const struct pnfs_layout_range
*recall_range
)
582 return (recall_range
->iomode
== IOMODE_ANY
||
583 lseg_range
->iomode
== recall_range
->iomode
) &&
584 pnfs_lseg_range_intersecting(lseg_range
, recall_range
);
588 pnfs_match_lseg_recall(const struct pnfs_layout_segment
*lseg
,
589 const struct pnfs_layout_range
*recall_range
,
592 if (seq
!= 0 && pnfs_seqid_is_newer(lseg
->pls_seq
, seq
))
594 if (recall_range
== NULL
)
596 return pnfs_should_free_range(&lseg
->pls_range
, recall_range
);
600 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
601 * @lo: layout header containing the lsegs
602 * @tmp_list: list head where doomed lsegs should go
603 * @recall_range: optional recall range argument to match (may be NULL)
604 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
606 * Walk the list of lsegs in the layout header, and tear down any that should
607 * be destroyed. If "recall_range" is specified then the segment must match
608 * that range. If "seq" is non-zero, then only match segments that were handed
609 * out at or before that sequence.
611 * Returns number of matching invalid lsegs remaining in list after scanning
612 * it and purging them.
615 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr
*lo
,
616 struct list_head
*tmp_list
,
617 const struct pnfs_layout_range
*recall_range
,
620 struct pnfs_layout_segment
*lseg
, *next
;
623 dprintk("%s:Begin lo %p\n", __func__
, lo
);
625 if (list_empty(&lo
->plh_segs
))
627 list_for_each_entry_safe(lseg
, next
, &lo
->plh_segs
, pls_list
)
628 if (pnfs_match_lseg_recall(lseg
, recall_range
, seq
)) {
629 dprintk("%s: freeing lseg %p iomode %d seq %u"
630 "offset %llu length %llu\n", __func__
,
631 lseg
, lseg
->pls_range
.iomode
, lseg
->pls_seq
,
632 lseg
->pls_range
.offset
, lseg
->pls_range
.length
);
633 if (!mark_lseg_invalid(lseg
, tmp_list
))
636 dprintk("%s:Return %i\n", __func__
, remaining
);
640 /* note free_me must contain lsegs from a single layout_hdr */
642 pnfs_free_lseg_list(struct list_head
*free_me
)
644 struct pnfs_layout_segment
*lseg
, *tmp
;
646 if (list_empty(free_me
))
649 list_for_each_entry_safe(lseg
, tmp
, free_me
, pls_list
) {
650 list_del(&lseg
->pls_list
);
651 pnfs_free_lseg(lseg
);
656 pnfs_destroy_layout(struct nfs_inode
*nfsi
)
658 struct pnfs_layout_hdr
*lo
;
661 spin_lock(&nfsi
->vfs_inode
.i_lock
);
664 pnfs_get_layout_hdr(lo
);
665 pnfs_mark_layout_stateid_invalid(lo
, &tmp_list
);
666 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RO_FAILED
);
667 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RW_FAILED
);
668 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
669 pnfs_free_lseg_list(&tmp_list
);
670 pnfs_put_layout_hdr(lo
);
672 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
674 EXPORT_SYMBOL_GPL(pnfs_destroy_layout
);
677 pnfs_layout_add_bulk_destroy_list(struct inode
*inode
,
678 struct list_head
*layout_list
)
680 struct pnfs_layout_hdr
*lo
;
683 spin_lock(&inode
->i_lock
);
684 lo
= NFS_I(inode
)->layout
;
685 if (lo
!= NULL
&& list_empty(&lo
->plh_bulk_destroy
)) {
686 pnfs_get_layout_hdr(lo
);
687 list_add(&lo
->plh_bulk_destroy
, layout_list
);
690 spin_unlock(&inode
->i_lock
);
694 /* Caller must hold rcu_read_lock and clp->cl_lock */
696 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client
*clp
,
697 struct nfs_server
*server
,
698 struct list_head
*layout_list
)
700 struct pnfs_layout_hdr
*lo
, *next
;
703 list_for_each_entry_safe(lo
, next
, &server
->layouts
, plh_layouts
) {
704 inode
= igrab(lo
->plh_inode
);
707 list_del_init(&lo
->plh_layouts
);
708 if (pnfs_layout_add_bulk_destroy_list(inode
, layout_list
))
711 spin_unlock(&clp
->cl_lock
);
713 spin_lock(&clp
->cl_lock
);
721 pnfs_layout_free_bulk_destroy_list(struct list_head
*layout_list
,
724 struct pnfs_layout_hdr
*lo
;
726 LIST_HEAD(lseg_list
);
729 while (!list_empty(layout_list
)) {
730 lo
= list_entry(layout_list
->next
, struct pnfs_layout_hdr
,
732 dprintk("%s freeing layout for inode %lu\n", __func__
,
733 lo
->plh_inode
->i_ino
);
734 inode
= lo
->plh_inode
;
736 pnfs_layoutcommit_inode(inode
, false);
738 spin_lock(&inode
->i_lock
);
739 list_del_init(&lo
->plh_bulk_destroy
);
740 if (pnfs_mark_layout_stateid_invalid(lo
, &lseg_list
)) {
742 set_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
);
745 spin_unlock(&inode
->i_lock
);
746 pnfs_free_lseg_list(&lseg_list
);
747 /* Free all lsegs that are attached to commit buckets */
748 nfs_commit_inode(inode
, 0);
749 pnfs_put_layout_hdr(lo
);
756 pnfs_destroy_layouts_byfsid(struct nfs_client
*clp
,
757 struct nfs_fsid
*fsid
,
760 struct nfs_server
*server
;
761 LIST_HEAD(layout_list
);
763 spin_lock(&clp
->cl_lock
);
766 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
767 if (memcmp(&server
->fsid
, fsid
, sizeof(*fsid
)) != 0)
769 if (pnfs_layout_bulk_destroy_byserver_locked(clp
,
775 spin_unlock(&clp
->cl_lock
);
777 if (list_empty(&layout_list
))
779 return pnfs_layout_free_bulk_destroy_list(&layout_list
, is_recall
);
783 pnfs_destroy_layouts_byclid(struct nfs_client
*clp
,
786 struct nfs_server
*server
;
787 LIST_HEAD(layout_list
);
789 spin_lock(&clp
->cl_lock
);
792 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
793 if (pnfs_layout_bulk_destroy_byserver_locked(clp
,
799 spin_unlock(&clp
->cl_lock
);
801 if (list_empty(&layout_list
))
803 return pnfs_layout_free_bulk_destroy_list(&layout_list
, is_recall
);
807 * Called by the state manger to remove all layouts established under an
811 pnfs_destroy_all_layouts(struct nfs_client
*clp
)
813 nfs4_deviceid_mark_client_invalid(clp
);
814 nfs4_deviceid_purge_client(clp
);
816 pnfs_destroy_layouts_byclid(clp
, false);
820 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr
*lo
)
822 lo
->plh_return_iomode
= 0;
823 lo
->plh_return_seq
= 0;
824 clear_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
);
827 /* update lo->plh_stateid with new if is more recent */
829 pnfs_set_layout_stateid(struct pnfs_layout_hdr
*lo
, const nfs4_stateid
*new,
832 u32 oldseq
, newseq
, new_barrier
= 0;
834 oldseq
= be32_to_cpu(lo
->plh_stateid
.seqid
);
835 newseq
= be32_to_cpu(new->seqid
);
837 if (!pnfs_layout_is_valid(lo
)) {
838 nfs4_stateid_copy(&lo
->plh_stateid
, new);
839 lo
->plh_barrier
= newseq
;
840 pnfs_clear_layoutreturn_info(lo
);
841 clear_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
);
844 if (pnfs_seqid_is_newer(newseq
, oldseq
)) {
845 nfs4_stateid_copy(&lo
->plh_stateid
, new);
847 * Because of wraparound, we want to keep the barrier
848 * "close" to the current seqids.
850 new_barrier
= newseq
- atomic_read(&lo
->plh_outstanding
);
853 new_barrier
= be32_to_cpu(new->seqid
);
854 else if (new_barrier
== 0)
856 if (pnfs_seqid_is_newer(new_barrier
, lo
->plh_barrier
))
857 lo
->plh_barrier
= new_barrier
;
861 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr
*lo
,
862 const nfs4_stateid
*stateid
)
864 u32 seqid
= be32_to_cpu(stateid
->seqid
);
866 return !pnfs_seqid_is_newer(seqid
, lo
->plh_barrier
);
869 /* lget is set to 1 if called from inside send_layoutget call chain */
871 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr
*lo
)
873 return lo
->plh_block_lgets
||
874 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
);
878 * Get layout from server.
879 * for now, assume that whole file layouts are requested.
881 * arg->length: all ones
883 static struct pnfs_layout_segment
*
884 send_layoutget(struct pnfs_layout_hdr
*lo
,
885 struct nfs_open_context
*ctx
,
886 nfs4_stateid
*stateid
,
887 const struct pnfs_layout_range
*range
,
888 long *timeout
, gfp_t gfp_flags
)
890 struct inode
*ino
= lo
->plh_inode
;
891 struct nfs_server
*server
= NFS_SERVER(ino
);
892 struct nfs4_layoutget
*lgp
;
895 dprintk("--> %s\n", __func__
);
898 * Synchronously retrieve layout information from server and
899 * store in lseg. If we race with a concurrent seqid morphing
900 * op, then re-send the LAYOUTGET.
902 lgp
= kzalloc(sizeof(*lgp
), gfp_flags
);
904 return ERR_PTR(-ENOMEM
);
906 i_size
= i_size_read(ino
);
908 lgp
->args
.minlength
= PAGE_SIZE
;
909 if (lgp
->args
.minlength
> range
->length
)
910 lgp
->args
.minlength
= range
->length
;
911 if (range
->iomode
== IOMODE_READ
) {
912 if (range
->offset
>= i_size
)
913 lgp
->args
.minlength
= 0;
914 else if (i_size
- range
->offset
< lgp
->args
.minlength
)
915 lgp
->args
.minlength
= i_size
- range
->offset
;
917 lgp
->args
.maxcount
= PNFS_LAYOUT_MAXSIZE
;
918 pnfs_copy_range(&lgp
->args
.range
, range
);
919 lgp
->args
.type
= server
->pnfs_curr_ld
->id
;
920 lgp
->args
.inode
= ino
;
921 lgp
->args
.ctx
= get_nfs_open_context(ctx
);
922 nfs4_stateid_copy(&lgp
->args
.stateid
, stateid
);
923 lgp
->gfp_flags
= gfp_flags
;
924 lgp
->cred
= lo
->plh_lc_cred
;
926 return nfs4_proc_layoutget(lgp
, timeout
, gfp_flags
);
929 static void pnfs_clear_layoutcommit(struct inode
*inode
,
930 struct list_head
*head
)
932 struct nfs_inode
*nfsi
= NFS_I(inode
);
933 struct pnfs_layout_segment
*lseg
, *tmp
;
935 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
937 list_for_each_entry_safe(lseg
, tmp
, &nfsi
->layout
->plh_segs
, pls_list
) {
938 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
940 pnfs_lseg_dec_and_remove_zero(lseg
, head
);
944 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr
*lo
)
946 clear_bit_unlock(NFS_LAYOUT_RETURN
, &lo
->plh_flags
);
947 smp_mb__after_atomic();
948 wake_up_bit(&lo
->plh_flags
, NFS_LAYOUT_RETURN
);
949 rpc_wake_up(&NFS_SERVER(lo
->plh_inode
)->roc_rpcwaitq
);
953 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr
*lo
,
954 nfs4_stateid
*stateid
,
955 enum pnfs_iomode
*iomode
)
957 /* Serialise LAYOUTGET/LAYOUTRETURN */
958 if (atomic_read(&lo
->plh_outstanding
) != 0)
960 if (test_and_set_bit(NFS_LAYOUT_RETURN
, &lo
->plh_flags
))
962 pnfs_get_layout_hdr(lo
);
963 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
)) {
964 if (stateid
!= NULL
) {
965 nfs4_stateid_copy(stateid
, &lo
->plh_stateid
);
966 if (lo
->plh_return_seq
!= 0)
967 stateid
->seqid
= cpu_to_be32(lo
->plh_return_seq
);
970 *iomode
= lo
->plh_return_iomode
;
971 pnfs_clear_layoutreturn_info(lo
);
975 nfs4_stateid_copy(stateid
, &lo
->plh_stateid
);
977 *iomode
= IOMODE_ANY
;
982 pnfs_send_layoutreturn(struct pnfs_layout_hdr
*lo
, const nfs4_stateid
*stateid
,
983 enum pnfs_iomode iomode
, bool sync
)
985 struct inode
*ino
= lo
->plh_inode
;
986 struct nfs4_layoutreturn
*lrp
;
989 lrp
= kzalloc(sizeof(*lrp
), GFP_NOFS
);
990 if (unlikely(lrp
== NULL
)) {
992 spin_lock(&ino
->i_lock
);
993 pnfs_clear_layoutreturn_waitbit(lo
);
994 spin_unlock(&ino
->i_lock
);
995 pnfs_put_layout_hdr(lo
);
999 nfs4_stateid_copy(&lrp
->args
.stateid
, stateid
);
1000 lrp
->args
.layout_type
= NFS_SERVER(ino
)->pnfs_curr_ld
->id
;
1001 lrp
->args
.inode
= ino
;
1002 lrp
->args
.range
.iomode
= iomode
;
1003 lrp
->args
.range
.offset
= 0;
1004 lrp
->args
.range
.length
= NFS4_MAX_UINT64
;
1005 lrp
->args
.layout
= lo
;
1006 lrp
->clp
= NFS_SERVER(ino
)->nfs_client
;
1007 lrp
->cred
= lo
->plh_lc_cred
;
1009 status
= nfs4_proc_layoutreturn(lrp
, sync
);
1011 dprintk("<-- %s status: %d\n", __func__
, status
);
1015 /* Return true if layoutreturn is needed */
1017 pnfs_layout_need_return(struct pnfs_layout_hdr
*lo
)
1019 struct pnfs_layout_segment
*s
;
1021 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
))
1024 /* Defer layoutreturn until all lsegs are done */
1025 list_for_each_entry(s
, &lo
->plh_segs
, pls_list
) {
1026 if (test_bit(NFS_LSEG_LAYOUTRETURN
, &s
->pls_flags
))
1033 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr
*lo
)
1035 struct inode
*inode
= lo
->plh_inode
;
1037 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
))
1039 spin_lock(&inode
->i_lock
);
1040 if (pnfs_layout_need_return(lo
)) {
1041 nfs4_stateid stateid
;
1042 enum pnfs_iomode iomode
;
1045 send
= pnfs_prepare_layoutreturn(lo
, &stateid
, &iomode
);
1046 spin_unlock(&inode
->i_lock
);
1048 /* Send an async layoutreturn so we dont deadlock */
1049 pnfs_send_layoutreturn(lo
, &stateid
, iomode
, false);
1052 spin_unlock(&inode
->i_lock
);
1056 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1057 * when the layout segment list is empty.
1059 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1060 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1061 * deviceid is marked invalid.
1064 _pnfs_return_layout(struct inode
*ino
)
1066 struct pnfs_layout_hdr
*lo
= NULL
;
1067 struct nfs_inode
*nfsi
= NFS_I(ino
);
1068 LIST_HEAD(tmp_list
);
1069 nfs4_stateid stateid
;
1070 int status
= 0, empty
;
1073 dprintk("NFS: %s for inode %lu\n", __func__
, ino
->i_ino
);
1075 spin_lock(&ino
->i_lock
);
1078 spin_unlock(&ino
->i_lock
);
1079 dprintk("NFS: %s no layout to return\n", __func__
);
1082 /* Reference matched in nfs4_layoutreturn_release */
1083 pnfs_get_layout_hdr(lo
);
1084 empty
= list_empty(&lo
->plh_segs
);
1085 pnfs_clear_layoutcommit(ino
, &tmp_list
);
1086 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
, 0);
1088 if (NFS_SERVER(ino
)->pnfs_curr_ld
->return_range
) {
1089 struct pnfs_layout_range range
= {
1090 .iomode
= IOMODE_ANY
,
1092 .length
= NFS4_MAX_UINT64
,
1094 NFS_SERVER(ino
)->pnfs_curr_ld
->return_range(lo
, &range
);
1097 /* Don't send a LAYOUTRETURN if list was initially empty */
1099 spin_unlock(&ino
->i_lock
);
1100 dprintk("NFS: %s no layout segments to return\n", __func__
);
1101 goto out_put_layout_hdr
;
1104 send
= pnfs_prepare_layoutreturn(lo
, &stateid
, NULL
);
1105 spin_unlock(&ino
->i_lock
);
1106 pnfs_free_lseg_list(&tmp_list
);
1108 status
= pnfs_send_layoutreturn(lo
, &stateid
, IOMODE_ANY
, true);
1110 pnfs_put_layout_hdr(lo
);
1112 dprintk("<-- %s status: %d\n", __func__
, status
);
1115 EXPORT_SYMBOL_GPL(_pnfs_return_layout
);
1118 pnfs_commit_and_return_layout(struct inode
*inode
)
1120 struct pnfs_layout_hdr
*lo
;
1123 spin_lock(&inode
->i_lock
);
1124 lo
= NFS_I(inode
)->layout
;
1126 spin_unlock(&inode
->i_lock
);
1129 pnfs_get_layout_hdr(lo
);
1130 /* Block new layoutgets and read/write to ds */
1131 lo
->plh_block_lgets
++;
1132 spin_unlock(&inode
->i_lock
);
1133 filemap_fdatawait(inode
->i_mapping
);
1134 ret
= pnfs_layoutcommit_inode(inode
, true);
1136 ret
= _pnfs_return_layout(inode
);
1137 spin_lock(&inode
->i_lock
);
1138 lo
->plh_block_lgets
--;
1139 spin_unlock(&inode
->i_lock
);
1140 pnfs_put_layout_hdr(lo
);
1144 bool pnfs_roc(struct inode
*ino
)
1146 struct nfs_inode
*nfsi
= NFS_I(ino
);
1147 struct nfs_open_context
*ctx
;
1148 struct nfs4_state
*state
;
1149 struct pnfs_layout_hdr
*lo
;
1150 struct pnfs_layout_segment
*lseg
, *tmp
;
1151 nfs4_stateid stateid
;
1152 LIST_HEAD(tmp_list
);
1153 bool found
= false, layoutreturn
= false, roc
= false;
1155 spin_lock(&ino
->i_lock
);
1157 if (!lo
|| test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
))
1160 /* no roc if we hold a delegation */
1161 if (nfs4_check_delegation(ino
, FMODE_READ
))
1164 list_for_each_entry(ctx
, &nfsi
->open_files
, list
) {
1166 /* Don't return layout if there is open file state */
1167 if (state
!= NULL
&& state
->state
!= 0)
1171 /* always send layoutreturn if being marked so */
1172 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
))
1173 layoutreturn
= pnfs_prepare_layoutreturn(lo
,
1176 list_for_each_entry_safe(lseg
, tmp
, &lo
->plh_segs
, pls_list
)
1177 /* If we are sending layoutreturn, invalidate all valid lsegs */
1178 if (layoutreturn
|| test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
1179 mark_lseg_invalid(lseg
, &tmp_list
);
1182 /* ROC in two conditions:
1183 * 1. there are ROC lsegs
1184 * 2. we don't send layoutreturn
1186 if (found
&& !layoutreturn
) {
1187 /* lo ref dropped in pnfs_roc_release() */
1188 pnfs_get_layout_hdr(lo
);
1193 spin_unlock(&ino
->i_lock
);
1194 pnfs_free_lseg_list(&tmp_list
);
1195 pnfs_layoutcommit_inode(ino
, true);
1197 pnfs_send_layoutreturn(lo
, &stateid
, IOMODE_ANY
, true);
1201 void pnfs_roc_release(struct inode
*ino
)
1203 struct pnfs_layout_hdr
*lo
;
1205 spin_lock(&ino
->i_lock
);
1206 lo
= NFS_I(ino
)->layout
;
1207 pnfs_clear_layoutreturn_waitbit(lo
);
1208 if (atomic_dec_and_test(&lo
->plh_refcount
)) {
1209 pnfs_detach_layout_hdr(lo
);
1210 spin_unlock(&ino
->i_lock
);
1211 pnfs_free_layout_hdr(lo
);
1213 spin_unlock(&ino
->i_lock
);
1216 void pnfs_roc_set_barrier(struct inode
*ino
, u32 barrier
)
1218 struct pnfs_layout_hdr
*lo
;
1220 spin_lock(&ino
->i_lock
);
1221 lo
= NFS_I(ino
)->layout
;
1222 if (pnfs_seqid_is_newer(barrier
, lo
->plh_barrier
))
1223 lo
->plh_barrier
= barrier
;
1224 spin_unlock(&ino
->i_lock
);
1225 trace_nfs4_layoutreturn_on_close(ino
, 0);
1228 void pnfs_roc_get_barrier(struct inode
*ino
, u32
*barrier
)
1230 struct nfs_inode
*nfsi
= NFS_I(ino
);
1231 struct pnfs_layout_hdr
*lo
;
1234 spin_lock(&ino
->i_lock
);
1236 current_seqid
= be32_to_cpu(lo
->plh_stateid
.seqid
);
1238 /* Since close does not return a layout stateid for use as
1239 * a barrier, we choose the worst-case barrier.
1241 *barrier
= current_seqid
+ atomic_read(&lo
->plh_outstanding
);
1242 spin_unlock(&ino
->i_lock
);
1245 bool pnfs_wait_on_layoutreturn(struct inode
*ino
, struct rpc_task
*task
)
1247 struct nfs_inode
*nfsi
= NFS_I(ino
);
1248 struct pnfs_layout_hdr
*lo
;
1251 /* we might not have grabbed lo reference. so need to check under
1253 spin_lock(&ino
->i_lock
);
1255 if (lo
&& test_bit(NFS_LAYOUT_RETURN
, &lo
->plh_flags
))
1257 spin_unlock(&ino
->i_lock
);
1260 rpc_sleep_on(&NFS_SERVER(ino
)->roc_rpcwaitq
, task
, NULL
);
1266 * Compare two layout segments for sorting into layout cache.
1267 * We want to preferentially return RW over RO layouts, so ensure those
1271 pnfs_lseg_range_cmp(const struct pnfs_layout_range
*l1
,
1272 const struct pnfs_layout_range
*l2
)
1276 /* high offset > low offset */
1277 d
= l1
->offset
- l2
->offset
;
1281 /* short length > long length */
1282 d
= l2
->length
- l1
->length
;
1286 /* read > read/write */
1287 return (int)(l1
->iomode
== IOMODE_READ
) - (int)(l2
->iomode
== IOMODE_READ
);
1291 pnfs_lseg_range_is_after(const struct pnfs_layout_range
*l1
,
1292 const struct pnfs_layout_range
*l2
)
1294 return pnfs_lseg_range_cmp(l1
, l2
) > 0;
1298 pnfs_lseg_no_merge(struct pnfs_layout_segment
*lseg
,
1299 struct pnfs_layout_segment
*old
)
1305 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr
*lo
,
1306 struct pnfs_layout_segment
*lseg
,
1307 bool (*is_after
)(const struct pnfs_layout_range
*,
1308 const struct pnfs_layout_range
*),
1309 bool (*do_merge
)(struct pnfs_layout_segment
*,
1310 struct pnfs_layout_segment
*),
1311 struct list_head
*free_me
)
1313 struct pnfs_layout_segment
*lp
, *tmp
;
1315 dprintk("%s:Begin\n", __func__
);
1317 list_for_each_entry_safe(lp
, tmp
, &lo
->plh_segs
, pls_list
) {
1318 if (test_bit(NFS_LSEG_VALID
, &lp
->pls_flags
) == 0)
1320 if (do_merge(lseg
, lp
)) {
1321 mark_lseg_invalid(lp
, free_me
);
1324 if (is_after(&lseg
->pls_range
, &lp
->pls_range
))
1326 list_add_tail(&lseg
->pls_list
, &lp
->pls_list
);
1327 dprintk("%s: inserted lseg %p "
1328 "iomode %d offset %llu length %llu before "
1329 "lp %p iomode %d offset %llu length %llu\n",
1330 __func__
, lseg
, lseg
->pls_range
.iomode
,
1331 lseg
->pls_range
.offset
, lseg
->pls_range
.length
,
1332 lp
, lp
->pls_range
.iomode
, lp
->pls_range
.offset
,
1333 lp
->pls_range
.length
);
1336 list_add_tail(&lseg
->pls_list
, &lo
->plh_segs
);
1337 dprintk("%s: inserted lseg %p "
1338 "iomode %d offset %llu length %llu at tail\n",
1339 __func__
, lseg
, lseg
->pls_range
.iomode
,
1340 lseg
->pls_range
.offset
, lseg
->pls_range
.length
);
1342 pnfs_get_layout_hdr(lo
);
1344 dprintk("%s:Return\n", __func__
);
1346 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg
);
1349 pnfs_layout_insert_lseg(struct pnfs_layout_hdr
*lo
,
1350 struct pnfs_layout_segment
*lseg
,
1351 struct list_head
*free_me
)
1353 struct inode
*inode
= lo
->plh_inode
;
1354 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(inode
)->pnfs_curr_ld
;
1356 if (ld
->add_lseg
!= NULL
)
1357 ld
->add_lseg(lo
, lseg
, free_me
);
1359 pnfs_generic_layout_insert_lseg(lo
, lseg
,
1360 pnfs_lseg_range_is_after
,
1365 static struct pnfs_layout_hdr
*
1366 alloc_init_layout_hdr(struct inode
*ino
,
1367 struct nfs_open_context
*ctx
,
1370 struct pnfs_layout_hdr
*lo
;
1372 lo
= pnfs_alloc_layout_hdr(ino
, gfp_flags
);
1375 atomic_set(&lo
->plh_refcount
, 1);
1376 INIT_LIST_HEAD(&lo
->plh_layouts
);
1377 INIT_LIST_HEAD(&lo
->plh_segs
);
1378 INIT_LIST_HEAD(&lo
->plh_bulk_destroy
);
1379 lo
->plh_inode
= ino
;
1380 lo
->plh_lc_cred
= get_rpccred(ctx
->cred
);
1381 lo
->plh_flags
|= 1 << NFS_LAYOUT_INVALID_STID
;
1385 static struct pnfs_layout_hdr
*
1386 pnfs_find_alloc_layout(struct inode
*ino
,
1387 struct nfs_open_context
*ctx
,
1389 __releases(&ino
->i_lock
)
1390 __acquires(&ino
->i_lock
)
1392 struct nfs_inode
*nfsi
= NFS_I(ino
);
1393 struct pnfs_layout_hdr
*new = NULL
;
1395 dprintk("%s Begin ino=%p layout=%p\n", __func__
, ino
, nfsi
->layout
);
1397 if (nfsi
->layout
!= NULL
)
1399 spin_unlock(&ino
->i_lock
);
1400 new = alloc_init_layout_hdr(ino
, ctx
, gfp_flags
);
1401 spin_lock(&ino
->i_lock
);
1403 if (likely(nfsi
->layout
== NULL
)) { /* Won the race? */
1406 } else if (new != NULL
)
1407 pnfs_free_layout_hdr(new);
1409 pnfs_get_layout_hdr(nfsi
->layout
);
1410 return nfsi
->layout
;
1414 * iomode matching rules:
1415 * iomode lseg strict match
1417 * ----- ----- ------ -----
1422 * READ READ N/A true
1423 * READ RW true false
1424 * READ RW false true
1427 pnfs_lseg_range_match(const struct pnfs_layout_range
*ls_range
,
1428 const struct pnfs_layout_range
*range
,
1431 struct pnfs_layout_range range1
;
1433 if ((range
->iomode
== IOMODE_RW
&&
1434 ls_range
->iomode
!= IOMODE_RW
) ||
1435 (range
->iomode
!= ls_range
->iomode
&&
1436 strict_iomode
== true) ||
1437 !pnfs_lseg_range_intersecting(ls_range
, range
))
1440 /* range1 covers only the first byte in the range */
1443 return pnfs_lseg_range_contained(ls_range
, &range1
);
1447 * lookup range in layout
1449 static struct pnfs_layout_segment
*
1450 pnfs_find_lseg(struct pnfs_layout_hdr
*lo
,
1451 struct pnfs_layout_range
*range
,
1454 struct pnfs_layout_segment
*lseg
, *ret
= NULL
;
1456 dprintk("%s:Begin\n", __func__
);
1458 list_for_each_entry(lseg
, &lo
->plh_segs
, pls_list
) {
1459 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
) &&
1460 !test_bit(NFS_LSEG_LAYOUTRETURN
, &lseg
->pls_flags
) &&
1461 pnfs_lseg_range_match(&lseg
->pls_range
, range
,
1463 ret
= pnfs_get_lseg(lseg
);
1468 dprintk("%s:Return lseg %p ref %d\n",
1469 __func__
, ret
, ret
? atomic_read(&ret
->pls_refcount
) : 0);
1474 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1475 * to the MDS or over pNFS
1477 * The nfs_inode read_io and write_io fields are cumulative counters reset
1478 * when there are no layout segments. Note that in pnfs_update_layout iomode
1479 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1482 * A return of true means use MDS I/O.
1485 * If a file's size is smaller than the file size threshold, data accesses
1486 * SHOULD be sent to the metadata server. If an I/O request has a length that
1487 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1488 * server. If both file size and I/O size are provided, the client SHOULD
1489 * reach or exceed both thresholds before sending its read or write
1490 * requests to the data server.
1492 static bool pnfs_within_mdsthreshold(struct nfs_open_context
*ctx
,
1493 struct inode
*ino
, int iomode
)
1495 struct nfs4_threshold
*t
= ctx
->mdsthreshold
;
1496 struct nfs_inode
*nfsi
= NFS_I(ino
);
1497 loff_t fsize
= i_size_read(ino
);
1498 bool size
= false, size_set
= false, io
= false, io_set
= false, ret
= false;
1503 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1504 __func__
, t
->bm
, t
->rd_sz
, t
->wr_sz
, t
->rd_io_sz
, t
->wr_io_sz
);
1508 if (t
->bm
& THRESHOLD_RD
) {
1509 dprintk("%s fsize %llu\n", __func__
, fsize
);
1511 if (fsize
< t
->rd_sz
)
1514 if (t
->bm
& THRESHOLD_RD_IO
) {
1515 dprintk("%s nfsi->read_io %llu\n", __func__
,
1518 if (nfsi
->read_io
< t
->rd_io_sz
)
1523 if (t
->bm
& THRESHOLD_WR
) {
1524 dprintk("%s fsize %llu\n", __func__
, fsize
);
1526 if (fsize
< t
->wr_sz
)
1529 if (t
->bm
& THRESHOLD_WR_IO
) {
1530 dprintk("%s nfsi->write_io %llu\n", __func__
,
1533 if (nfsi
->write_io
< t
->wr_io_sz
)
1538 if (size_set
&& io_set
) {
1541 } else if (size
|| io
)
1544 dprintk("<-- %s size %d io %d ret %d\n", __func__
, size
, io
, ret
);
1548 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr
*lo
)
1551 * send layoutcommit as it can hold up layoutreturn due to lseg
1554 pnfs_layoutcommit_inode(lo
->plh_inode
, false);
1555 return !wait_on_bit_action(&lo
->plh_flags
, NFS_LAYOUT_RETURN
,
1556 nfs_wait_bit_killable
,
1557 TASK_UNINTERRUPTIBLE
);
1560 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr
*lo
)
1562 unsigned long *bitlock
= &lo
->plh_flags
;
1564 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET
, bitlock
);
1565 smp_mb__after_atomic();
1566 wake_up_bit(bitlock
, NFS_LAYOUT_FIRST_LAYOUTGET
);
1570 * Layout segment is retreived from the server if not cached.
1571 * The appropriate layout segment is referenced and returned to the caller.
1573 struct pnfs_layout_segment
*
1574 pnfs_update_layout(struct inode
*ino
,
1575 struct nfs_open_context
*ctx
,
1578 enum pnfs_iomode iomode
,
1582 struct pnfs_layout_range arg
= {
1587 unsigned pg_offset
, seq
;
1588 struct nfs_server
*server
= NFS_SERVER(ino
);
1589 struct nfs_client
*clp
= server
->nfs_client
;
1590 struct pnfs_layout_hdr
*lo
= NULL
;
1591 struct pnfs_layout_segment
*lseg
= NULL
;
1592 nfs4_stateid stateid
;
1594 unsigned long giveup
= jiffies
+ (clp
->cl_lease_time
<< 1);
1597 if (!pnfs_enabled_sb(NFS_SERVER(ino
))) {
1598 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1599 PNFS_UPDATE_LAYOUT_NO_PNFS
);
1603 if (iomode
== IOMODE_READ
&& i_size_read(ino
) == 0) {
1604 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1605 PNFS_UPDATE_LAYOUT_RD_ZEROLEN
);
1609 if (pnfs_within_mdsthreshold(ctx
, ino
, iomode
)) {
1610 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1611 PNFS_UPDATE_LAYOUT_MDSTHRESH
);
1616 nfs4_client_recover_expired_lease(clp
);
1618 spin_lock(&ino
->i_lock
);
1619 lo
= pnfs_find_alloc_layout(ino
, ctx
, gfp_flags
);
1621 spin_unlock(&ino
->i_lock
);
1622 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1623 PNFS_UPDATE_LAYOUT_NOMEM
);
1627 /* Do we even need to bother with this? */
1628 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1629 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1630 PNFS_UPDATE_LAYOUT_BULK_RECALL
);
1631 dprintk("%s matches recall, use MDS\n", __func__
);
1635 /* if LAYOUTGET already failed once we don't try again */
1636 if (pnfs_layout_io_test_failed(lo
, iomode
)) {
1637 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1638 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL
);
1642 lseg
= pnfs_find_lseg(lo
, &arg
, strict_iomode
);
1644 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1645 PNFS_UPDATE_LAYOUT_FOUND_CACHED
);
1649 if (!nfs4_valid_open_stateid(ctx
->state
)) {
1650 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1651 PNFS_UPDATE_LAYOUT_INVALID_OPEN
);
1656 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1657 * stateid, or it has been invalidated, then we must use the open
1660 if (test_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
)) {
1663 * The first layoutget for the file. Need to serialize per
1664 * RFC 5661 Errata 3208.
1666 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET
,
1668 spin_unlock(&ino
->i_lock
);
1669 wait_on_bit(&lo
->plh_flags
, NFS_LAYOUT_FIRST_LAYOUTGET
,
1670 TASK_UNINTERRUPTIBLE
);
1671 pnfs_put_layout_hdr(lo
);
1672 dprintk("%s retrying\n", __func__
);
1678 seq
= read_seqbegin(&ctx
->state
->seqlock
);
1679 nfs4_stateid_copy(&stateid
, &ctx
->state
->stateid
);
1680 } while (read_seqretry(&ctx
->state
->seqlock
, seq
));
1682 nfs4_stateid_copy(&stateid
, &lo
->plh_stateid
);
1686 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1687 * for LAYOUTRETURN even if first is true.
1689 if (test_bit(NFS_LAYOUT_RETURN
, &lo
->plh_flags
)) {
1690 spin_unlock(&ino
->i_lock
);
1691 dprintk("%s wait for layoutreturn\n", __func__
);
1692 if (pnfs_prepare_to_retry_layoutget(lo
)) {
1694 pnfs_clear_first_layoutget(lo
);
1695 pnfs_put_layout_hdr(lo
);
1696 dprintk("%s retrying\n", __func__
);
1697 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
,
1698 lseg
, PNFS_UPDATE_LAYOUT_RETRY
);
1701 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1702 PNFS_UPDATE_LAYOUT_RETURN
);
1703 goto out_put_layout_hdr
;
1706 if (pnfs_layoutgets_blocked(lo
)) {
1707 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1708 PNFS_UPDATE_LAYOUT_BLOCKED
);
1711 atomic_inc(&lo
->plh_outstanding
);
1712 spin_unlock(&ino
->i_lock
);
1714 if (list_empty(&lo
->plh_layouts
)) {
1715 /* The lo must be on the clp list if there is any
1716 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1718 spin_lock(&clp
->cl_lock
);
1719 if (list_empty(&lo
->plh_layouts
))
1720 list_add_tail(&lo
->plh_layouts
, &server
->layouts
);
1721 spin_unlock(&clp
->cl_lock
);
1724 pg_offset
= arg
.offset
& ~PAGE_MASK
;
1726 arg
.offset
-= pg_offset
;
1727 arg
.length
+= pg_offset
;
1729 if (arg
.length
!= NFS4_MAX_UINT64
)
1730 arg
.length
= PAGE_ALIGN(arg
.length
);
1732 lseg
= send_layoutget(lo
, ctx
, &stateid
, &arg
, &timeout
, gfp_flags
);
1733 trace_pnfs_update_layout(ino
, pos
, count
, iomode
, lo
, lseg
,
1734 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET
);
1735 atomic_dec(&lo
->plh_outstanding
);
1737 switch(PTR_ERR(lseg
)) {
1739 if (time_after(jiffies
, giveup
))
1742 case -ERECALLCONFLICT
:
1743 /* Huh? We hold no layouts, how is there a recall? */
1748 /* Destroy the existing layout and start over */
1749 if (time_after(jiffies
, giveup
))
1750 pnfs_destroy_layout(NFS_I(ino
));
1755 if (!nfs_error_is_fatal(PTR_ERR(lseg
))) {
1756 pnfs_layout_clear_fail_bit(lo
, pnfs_iomode_to_fail_bit(iomode
));
1759 goto out_put_layout_hdr
;
1763 pnfs_clear_first_layoutget(lo
);
1764 trace_pnfs_update_layout(ino
, pos
, count
,
1765 iomode
, lo
, lseg
, PNFS_UPDATE_LAYOUT_RETRY
);
1766 pnfs_put_layout_hdr(lo
);
1770 pnfs_layout_clear_fail_bit(lo
, pnfs_iomode_to_fail_bit(iomode
));
1775 pnfs_clear_first_layoutget(lo
);
1776 pnfs_put_layout_hdr(lo
);
1778 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1779 "(%s, offset: %llu, length: %llu)\n",
1780 __func__
, ino
->i_sb
->s_id
,
1781 (unsigned long long)NFS_FILEID(ino
),
1782 IS_ERR_OR_NULL(lseg
) ? "not found" : "found",
1783 iomode
==IOMODE_RW
? "read/write" : "read-only",
1784 (unsigned long long)pos
,
1785 (unsigned long long)count
);
1788 spin_unlock(&ino
->i_lock
);
1789 goto out_put_layout_hdr
;
1791 EXPORT_SYMBOL_GPL(pnfs_update_layout
);
1794 pnfs_sanity_check_layout_range(struct pnfs_layout_range
*range
)
1796 switch (range
->iomode
) {
1803 if (range
->offset
== NFS4_MAX_UINT64
)
1805 if (range
->length
== 0)
1807 if (range
->length
!= NFS4_MAX_UINT64
&&
1808 range
->length
> NFS4_MAX_UINT64
- range
->offset
)
1813 struct pnfs_layout_segment
*
1814 pnfs_layout_process(struct nfs4_layoutget
*lgp
)
1816 struct pnfs_layout_hdr
*lo
= NFS_I(lgp
->args
.inode
)->layout
;
1817 struct nfs4_layoutget_res
*res
= &lgp
->res
;
1818 struct pnfs_layout_segment
*lseg
;
1819 struct inode
*ino
= lo
->plh_inode
;
1822 if (!pnfs_sanity_check_layout_range(&res
->range
))
1823 return ERR_PTR(-EINVAL
);
1825 /* Inject layout blob into I/O device driver */
1826 lseg
= NFS_SERVER(ino
)->pnfs_curr_ld
->alloc_lseg(lo
, res
, lgp
->gfp_flags
);
1827 if (IS_ERR_OR_NULL(lseg
)) {
1829 lseg
= ERR_PTR(-ENOMEM
);
1831 dprintk("%s: Could not allocate layout: error %ld\n",
1832 __func__
, PTR_ERR(lseg
));
1836 pnfs_init_lseg(lo
, lseg
, &res
->range
, &res
->stateid
);
1838 spin_lock(&ino
->i_lock
);
1839 if (pnfs_layoutgets_blocked(lo
)) {
1840 dprintk("%s forget reply due to state\n", __func__
);
1844 if (nfs4_stateid_match_other(&lo
->plh_stateid
, &res
->stateid
)) {
1845 /* existing state ID, make sure the sequence number matches. */
1846 if (pnfs_layout_stateid_blocked(lo
, &res
->stateid
)) {
1847 dprintk("%s forget reply due to sequence\n", __func__
);
1850 pnfs_set_layout_stateid(lo
, &res
->stateid
, false);
1853 * We got an entirely new state ID. Mark all segments for the
1854 * inode invalid, and don't bother validating the stateid
1857 pnfs_mark_layout_stateid_invalid(lo
, &free_me
);
1859 pnfs_set_layout_stateid(lo
, &res
->stateid
, true);
1862 pnfs_get_lseg(lseg
);
1863 pnfs_layout_insert_lseg(lo
, lseg
, &free_me
);
1866 if (res
->return_on_close
)
1867 set_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
);
1869 spin_unlock(&ino
->i_lock
);
1870 pnfs_free_lseg_list(&free_me
);
1874 spin_unlock(&ino
->i_lock
);
1875 lseg
->pls_layout
= lo
;
1876 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
1877 return ERR_PTR(-EAGAIN
);
1881 pnfs_set_plh_return_info(struct pnfs_layout_hdr
*lo
, enum pnfs_iomode iomode
,
1884 if (lo
->plh_return_iomode
!= 0 && lo
->plh_return_iomode
!= iomode
)
1885 iomode
= IOMODE_ANY
;
1886 lo
->plh_return_iomode
= iomode
;
1887 set_bit(NFS_LAYOUT_RETURN_REQUESTED
, &lo
->plh_flags
);
1889 WARN_ON_ONCE(lo
->plh_return_seq
!= 0 && lo
->plh_return_seq
!= seq
);
1890 lo
->plh_return_seq
= seq
;
1895 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1896 * @lo: pointer to layout header
1897 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1898 * @return_range: describe layout segment ranges to be returned
1900 * This function is mainly intended for use by layoutrecall. It attempts
1901 * to free the layout segment immediately, or else to mark it for return
1902 * as soon as its reference count drops to zero.
1905 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr
*lo
,
1906 struct list_head
*tmp_list
,
1907 const struct pnfs_layout_range
*return_range
,
1910 struct pnfs_layout_segment
*lseg
, *next
;
1913 dprintk("%s:Begin lo %p\n", __func__
, lo
);
1915 if (list_empty(&lo
->plh_segs
))
1918 assert_spin_locked(&lo
->plh_inode
->i_lock
);
1920 list_for_each_entry_safe(lseg
, next
, &lo
->plh_segs
, pls_list
)
1921 if (pnfs_match_lseg_recall(lseg
, return_range
, seq
)) {
1922 dprintk("%s: marking lseg %p iomode %d "
1923 "offset %llu length %llu\n", __func__
,
1924 lseg
, lseg
->pls_range
.iomode
,
1925 lseg
->pls_range
.offset
,
1926 lseg
->pls_range
.length
);
1927 if (mark_lseg_invalid(lseg
, tmp_list
))
1930 set_bit(NFS_LSEG_LAYOUTRETURN
, &lseg
->pls_flags
);
1934 pnfs_set_plh_return_info(lo
, return_range
->iomode
, seq
);
1939 void pnfs_error_mark_layout_for_return(struct inode
*inode
,
1940 struct pnfs_layout_segment
*lseg
)
1942 struct pnfs_layout_hdr
*lo
= NFS_I(inode
)->layout
;
1943 struct pnfs_layout_range range
= {
1944 .iomode
= lseg
->pls_range
.iomode
,
1946 .length
= NFS4_MAX_UINT64
,
1949 bool return_now
= false;
1951 spin_lock(&inode
->i_lock
);
1952 pnfs_set_plh_return_info(lo
, range
.iomode
, 0);
1954 * mark all matching lsegs so that we are sure to have no live
1955 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1958 if (!pnfs_mark_matching_lsegs_return(lo
, &free_me
, &range
, 0)) {
1959 nfs4_stateid stateid
;
1960 enum pnfs_iomode iomode
;
1962 return_now
= pnfs_prepare_layoutreturn(lo
, &stateid
, &iomode
);
1963 spin_unlock(&inode
->i_lock
);
1965 pnfs_send_layoutreturn(lo
, &stateid
, iomode
, false);
1967 spin_unlock(&inode
->i_lock
);
1968 nfs_commit_inode(inode
, 0);
1970 pnfs_free_lseg_list(&free_me
);
1972 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return
);
1975 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1977 u64 rd_size
= req
->wb_bytes
;
1979 if (pgio
->pg_lseg
== NULL
) {
1980 if (pgio
->pg_dreq
== NULL
)
1981 rd_size
= i_size_read(pgio
->pg_inode
) - req_offset(req
);
1983 rd_size
= nfs_dreq_bytes_left(pgio
->pg_dreq
);
1985 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1992 if (IS_ERR(pgio
->pg_lseg
)) {
1993 pgio
->pg_error
= PTR_ERR(pgio
->pg_lseg
);
1994 pgio
->pg_lseg
= NULL
;
1998 /* If no lseg, fall back to read through mds */
1999 if (pgio
->pg_lseg
== NULL
)
2000 nfs_pageio_reset_read_mds(pgio
);
2003 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read
);
2006 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor
*pgio
,
2007 struct nfs_page
*req
, u64 wb_size
)
2009 if (pgio
->pg_lseg
== NULL
) {
2010 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
2017 if (IS_ERR(pgio
->pg_lseg
)) {
2018 pgio
->pg_error
= PTR_ERR(pgio
->pg_lseg
);
2019 pgio
->pg_lseg
= NULL
;
2023 /* If no lseg, fall back to write through mds */
2024 if (pgio
->pg_lseg
== NULL
)
2025 nfs_pageio_reset_write_mds(pgio
);
2027 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write
);
2030 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor
*desc
)
2032 if (desc
->pg_lseg
) {
2033 pnfs_put_lseg(desc
->pg_lseg
);
2034 desc
->pg_lseg
= NULL
;
2037 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup
);
2040 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2041 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2044 pnfs_generic_pg_test(struct nfs_pageio_descriptor
*pgio
,
2045 struct nfs_page
*prev
, struct nfs_page
*req
)
2048 u64 seg_end
, req_start
, seg_left
;
2050 size
= nfs_generic_pg_test(pgio
, prev
, req
);
2055 * 'size' contains the number of bytes left in the current page (up
2056 * to the original size asked for in @req->wb_bytes).
2058 * Calculate how many bytes are left in the layout segment
2059 * and if there are less bytes than 'size', return that instead.
2061 * Please also note that 'end_offset' is actually the offset of the
2062 * first byte that lies outside the pnfs_layout_range. FIXME?
2065 if (pgio
->pg_lseg
) {
2066 seg_end
= end_offset(pgio
->pg_lseg
->pls_range
.offset
,
2067 pgio
->pg_lseg
->pls_range
.length
);
2068 req_start
= req_offset(req
);
2069 WARN_ON_ONCE(req_start
>= seg_end
);
2070 /* start of request is past the last byte of this segment */
2071 if (req_start
>= seg_end
) {
2072 /* reference the new lseg */
2073 if (pgio
->pg_ops
->pg_cleanup
)
2074 pgio
->pg_ops
->pg_cleanup(pgio
);
2075 if (pgio
->pg_ops
->pg_init
)
2076 pgio
->pg_ops
->pg_init(pgio
, req
);
2080 /* adjust 'size' iff there are fewer bytes left in the
2081 * segment than what nfs_generic_pg_test returned */
2082 seg_left
= seg_end
- req_start
;
2083 if (seg_left
< size
)
2084 size
= (unsigned int)seg_left
;
2089 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test
);
2091 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header
*hdr
)
2093 struct nfs_pageio_descriptor pgio
;
2095 /* Resend all requests through the MDS */
2096 nfs_pageio_init_write(&pgio
, hdr
->inode
, FLUSH_STABLE
, true,
2097 hdr
->completion_ops
);
2098 set_bit(NFS_CONTEXT_RESEND_WRITES
, &hdr
->args
.context
->flags
);
2099 return nfs_pageio_resend(&pgio
, hdr
);
2101 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds
);
2103 static void pnfs_ld_handle_write_error(struct nfs_pgio_header
*hdr
)
2106 dprintk("pnfs write error = %d\n", hdr
->pnfs_error
);
2107 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
2108 PNFS_LAYOUTRET_ON_ERROR
) {
2109 pnfs_return_layout(hdr
->inode
);
2111 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
2112 hdr
->task
.tk_status
= pnfs_write_done_resend_to_mds(hdr
);
2116 * Called by non rpc-based layout drivers
2118 void pnfs_ld_write_done(struct nfs_pgio_header
*hdr
)
2120 if (likely(!hdr
->pnfs_error
)) {
2121 pnfs_set_layoutcommit(hdr
->inode
, hdr
->lseg
,
2122 hdr
->mds_offset
+ hdr
->res
.count
);
2123 hdr
->mds_ops
->rpc_call_done(&hdr
->task
, hdr
);
2125 trace_nfs4_pnfs_write(hdr
, hdr
->pnfs_error
);
2126 if (unlikely(hdr
->pnfs_error
))
2127 pnfs_ld_handle_write_error(hdr
);
2128 hdr
->mds_ops
->rpc_release(hdr
);
2130 EXPORT_SYMBOL_GPL(pnfs_ld_write_done
);
2133 pnfs_write_through_mds(struct nfs_pageio_descriptor
*desc
,
2134 struct nfs_pgio_header
*hdr
)
2136 struct nfs_pgio_mirror
*mirror
= nfs_pgio_current_mirror(desc
);
2138 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
2139 list_splice_tail_init(&hdr
->pages
, &mirror
->pg_list
);
2140 nfs_pageio_reset_write_mds(desc
);
2141 mirror
->pg_recoalesce
= 1;
2143 nfs_pgio_data_destroy(hdr
);
2147 static enum pnfs_try_status
2148 pnfs_try_to_write_data(struct nfs_pgio_header
*hdr
,
2149 const struct rpc_call_ops
*call_ops
,
2150 struct pnfs_layout_segment
*lseg
,
2153 struct inode
*inode
= hdr
->inode
;
2154 enum pnfs_try_status trypnfs
;
2155 struct nfs_server
*nfss
= NFS_SERVER(inode
);
2157 hdr
->mds_ops
= call_ops
;
2159 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__
,
2160 inode
->i_ino
, hdr
->args
.count
, hdr
->args
.offset
, how
);
2161 trypnfs
= nfss
->pnfs_curr_ld
->write_pagelist(hdr
, how
);
2162 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
2163 nfs_inc_stats(inode
, NFSIOS_PNFS_WRITE
);
2164 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
2169 pnfs_do_write(struct nfs_pageio_descriptor
*desc
,
2170 struct nfs_pgio_header
*hdr
, int how
)
2172 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
2173 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
2174 enum pnfs_try_status trypnfs
;
2176 trypnfs
= pnfs_try_to_write_data(hdr
, call_ops
, lseg
, how
);
2177 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
2178 pnfs_write_through_mds(desc
, hdr
);
2181 static void pnfs_writehdr_free(struct nfs_pgio_header
*hdr
)
2183 pnfs_put_lseg(hdr
->lseg
);
2184 nfs_pgio_header_free(hdr
);
2188 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor
*desc
)
2190 struct nfs_pgio_header
*hdr
;
2193 hdr
= nfs_pgio_header_alloc(desc
->pg_rw_ops
);
2195 desc
->pg_error
= -ENOMEM
;
2196 return desc
->pg_error
;
2198 nfs_pgheader_init(desc
, hdr
, pnfs_writehdr_free
);
2200 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
2201 ret
= nfs_generic_pgio(desc
, hdr
);
2203 pnfs_do_write(desc
, hdr
, desc
->pg_ioflags
);
2207 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages
);
2209 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header
*hdr
)
2211 struct nfs_pageio_descriptor pgio
;
2213 /* Resend all requests through the MDS */
2214 nfs_pageio_init_read(&pgio
, hdr
->inode
, true, hdr
->completion_ops
);
2215 return nfs_pageio_resend(&pgio
, hdr
);
2217 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds
);
2219 static void pnfs_ld_handle_read_error(struct nfs_pgio_header
*hdr
)
2221 dprintk("pnfs read error = %d\n", hdr
->pnfs_error
);
2222 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
2223 PNFS_LAYOUTRET_ON_ERROR
) {
2224 pnfs_return_layout(hdr
->inode
);
2226 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
2227 hdr
->task
.tk_status
= pnfs_read_done_resend_to_mds(hdr
);
2231 * Called by non rpc-based layout drivers
2233 void pnfs_ld_read_done(struct nfs_pgio_header
*hdr
)
2235 if (likely(!hdr
->pnfs_error
))
2236 hdr
->mds_ops
->rpc_call_done(&hdr
->task
, hdr
);
2237 trace_nfs4_pnfs_read(hdr
, hdr
->pnfs_error
);
2238 if (unlikely(hdr
->pnfs_error
))
2239 pnfs_ld_handle_read_error(hdr
);
2240 hdr
->mds_ops
->rpc_release(hdr
);
2242 EXPORT_SYMBOL_GPL(pnfs_ld_read_done
);
2245 pnfs_read_through_mds(struct nfs_pageio_descriptor
*desc
,
2246 struct nfs_pgio_header
*hdr
)
2248 struct nfs_pgio_mirror
*mirror
= nfs_pgio_current_mirror(desc
);
2250 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
2251 list_splice_tail_init(&hdr
->pages
, &mirror
->pg_list
);
2252 nfs_pageio_reset_read_mds(desc
);
2253 mirror
->pg_recoalesce
= 1;
2255 nfs_pgio_data_destroy(hdr
);
2260 * Call the appropriate parallel I/O subsystem read function.
2262 static enum pnfs_try_status
2263 pnfs_try_to_read_data(struct nfs_pgio_header
*hdr
,
2264 const struct rpc_call_ops
*call_ops
,
2265 struct pnfs_layout_segment
*lseg
)
2267 struct inode
*inode
= hdr
->inode
;
2268 struct nfs_server
*nfss
= NFS_SERVER(inode
);
2269 enum pnfs_try_status trypnfs
;
2271 hdr
->mds_ops
= call_ops
;
2273 dprintk("%s: Reading ino:%lu %u@%llu\n",
2274 __func__
, inode
->i_ino
, hdr
->args
.count
, hdr
->args
.offset
);
2276 trypnfs
= nfss
->pnfs_curr_ld
->read_pagelist(hdr
);
2277 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
2278 nfs_inc_stats(inode
, NFSIOS_PNFS_READ
);
2279 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
2283 /* Resend all requests through pnfs. */
2284 void pnfs_read_resend_pnfs(struct nfs_pgio_header
*hdr
)
2286 struct nfs_pageio_descriptor pgio
;
2288 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
2289 nfs_pageio_init_read(&pgio
, hdr
->inode
, false,
2290 hdr
->completion_ops
);
2291 hdr
->task
.tk_status
= nfs_pageio_resend(&pgio
, hdr
);
2294 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs
);
2297 pnfs_do_read(struct nfs_pageio_descriptor
*desc
, struct nfs_pgio_header
*hdr
)
2299 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
2300 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
2301 enum pnfs_try_status trypnfs
;
2303 trypnfs
= pnfs_try_to_read_data(hdr
, call_ops
, lseg
);
2304 if (trypnfs
== PNFS_TRY_AGAIN
)
2305 pnfs_read_resend_pnfs(hdr
);
2306 if (trypnfs
== PNFS_NOT_ATTEMPTED
|| hdr
->task
.tk_status
)
2307 pnfs_read_through_mds(desc
, hdr
);
2310 static void pnfs_readhdr_free(struct nfs_pgio_header
*hdr
)
2312 pnfs_put_lseg(hdr
->lseg
);
2313 nfs_pgio_header_free(hdr
);
2317 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor
*desc
)
2319 struct nfs_pgio_header
*hdr
;
2322 hdr
= nfs_pgio_header_alloc(desc
->pg_rw_ops
);
2324 desc
->pg_error
= -ENOMEM
;
2325 return desc
->pg_error
;
2327 nfs_pgheader_init(desc
, hdr
, pnfs_readhdr_free
);
2328 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
2329 ret
= nfs_generic_pgio(desc
, hdr
);
2331 pnfs_do_read(desc
, hdr
);
2334 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages
);
2336 static void pnfs_clear_layoutcommitting(struct inode
*inode
)
2338 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
2340 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING
, bitlock
);
2341 smp_mb__after_atomic();
2342 wake_up_bit(bitlock
, NFS_INO_LAYOUTCOMMITTING
);
2346 * There can be multiple RW segments.
2348 static void pnfs_list_write_lseg(struct inode
*inode
, struct list_head
*listp
)
2350 struct pnfs_layout_segment
*lseg
;
2352 list_for_each_entry(lseg
, &NFS_I(inode
)->layout
->plh_segs
, pls_list
) {
2353 if (lseg
->pls_range
.iomode
== IOMODE_RW
&&
2354 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
2355 list_add(&lseg
->pls_lc_list
, listp
);
2359 static void pnfs_list_write_lseg_done(struct inode
*inode
, struct list_head
*listp
)
2361 struct pnfs_layout_segment
*lseg
, *tmp
;
2363 /* Matched by references in pnfs_set_layoutcommit */
2364 list_for_each_entry_safe(lseg
, tmp
, listp
, pls_lc_list
) {
2365 list_del_init(&lseg
->pls_lc_list
);
2366 pnfs_put_lseg(lseg
);
2369 pnfs_clear_layoutcommitting(inode
);
2372 void pnfs_set_lo_fail(struct pnfs_layout_segment
*lseg
)
2374 pnfs_layout_io_set_failed(lseg
->pls_layout
, lseg
->pls_range
.iomode
);
2376 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail
);
2379 pnfs_set_layoutcommit(struct inode
*inode
, struct pnfs_layout_segment
*lseg
,
2382 struct nfs_inode
*nfsi
= NFS_I(inode
);
2383 bool mark_as_dirty
= false;
2385 spin_lock(&inode
->i_lock
);
2386 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
2387 nfsi
->layout
->plh_lwb
= end_pos
;
2388 mark_as_dirty
= true;
2389 dprintk("%s: Set layoutcommit for inode %lu ",
2390 __func__
, inode
->i_ino
);
2391 } else if (end_pos
> nfsi
->layout
->plh_lwb
)
2392 nfsi
->layout
->plh_lwb
= end_pos
;
2393 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
)) {
2394 /* references matched in nfs4_layoutcommit_release */
2395 pnfs_get_lseg(lseg
);
2397 spin_unlock(&inode
->i_lock
);
2398 dprintk("%s: lseg %p end_pos %llu\n",
2399 __func__
, lseg
, nfsi
->layout
->plh_lwb
);
2401 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2402 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2404 mark_inode_dirty_sync(inode
);
2406 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit
);
2408 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data
*data
)
2410 struct nfs_server
*nfss
= NFS_SERVER(data
->args
.inode
);
2412 if (nfss
->pnfs_curr_ld
->cleanup_layoutcommit
)
2413 nfss
->pnfs_curr_ld
->cleanup_layoutcommit(data
);
2414 pnfs_list_write_lseg_done(data
->args
.inode
, &data
->lseg_list
);
2418 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2419 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2420 * data to disk to allow the server to recover the data if it crashes.
2421 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2422 * is off, and a COMMIT is sent to a data server, or
2423 * if WRITEs to a data server return NFS_DATA_SYNC.
2426 pnfs_layoutcommit_inode(struct inode
*inode
, bool sync
)
2428 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(inode
)->pnfs_curr_ld
;
2429 struct nfs4_layoutcommit_data
*data
;
2430 struct nfs_inode
*nfsi
= NFS_I(inode
);
2434 if (!pnfs_layoutcommit_outstanding(inode
))
2437 dprintk("--> %s inode %lu\n", __func__
, inode
->i_ino
);
2440 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
)) {
2443 status
= wait_on_bit_lock_action(&nfsi
->flags
,
2444 NFS_INO_LAYOUTCOMMITTING
,
2445 nfs_wait_bit_killable
,
2452 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2453 data
= kzalloc(sizeof(*data
), GFP_NOFS
);
2455 goto clear_layoutcommitting
;
2458 spin_lock(&inode
->i_lock
);
2459 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
2462 INIT_LIST_HEAD(&data
->lseg_list
);
2463 pnfs_list_write_lseg(inode
, &data
->lseg_list
);
2465 end_pos
= nfsi
->layout
->plh_lwb
;
2467 nfs4_stateid_copy(&data
->args
.stateid
, &nfsi
->layout
->plh_stateid
);
2468 spin_unlock(&inode
->i_lock
);
2470 data
->args
.inode
= inode
;
2471 data
->cred
= get_rpccred(nfsi
->layout
->plh_lc_cred
);
2472 nfs_fattr_init(&data
->fattr
);
2473 data
->args
.bitmask
= NFS_SERVER(inode
)->cache_consistency_bitmask
;
2474 data
->res
.fattr
= &data
->fattr
;
2476 data
->args
.lastbytewritten
= end_pos
- 1;
2478 data
->args
.lastbytewritten
= U64_MAX
;
2479 data
->res
.server
= NFS_SERVER(inode
);
2481 if (ld
->prepare_layoutcommit
) {
2482 status
= ld
->prepare_layoutcommit(&data
->args
);
2484 put_rpccred(data
->cred
);
2485 spin_lock(&inode
->i_lock
);
2486 set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
);
2487 if (end_pos
> nfsi
->layout
->plh_lwb
)
2488 nfsi
->layout
->plh_lwb
= end_pos
;
2494 status
= nfs4_proc_layoutcommit(data
, sync
);
2497 mark_inode_dirty_sync(inode
);
2498 dprintk("<-- %s status %d\n", __func__
, status
);
2501 spin_unlock(&inode
->i_lock
);
2503 clear_layoutcommitting
:
2504 pnfs_clear_layoutcommitting(inode
);
2507 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode
);
2510 pnfs_generic_sync(struct inode
*inode
, bool datasync
)
2512 return pnfs_layoutcommit_inode(inode
, true);
2514 EXPORT_SYMBOL_GPL(pnfs_generic_sync
);
2516 struct nfs4_threshold
*pnfs_mdsthreshold_alloc(void)
2518 struct nfs4_threshold
*thp
;
2520 thp
= kzalloc(sizeof(*thp
), GFP_NOFS
);
2522 dprintk("%s mdsthreshold allocation failed\n", __func__
);
2528 #if IS_ENABLED(CONFIG_NFS_V4_2)
2530 pnfs_report_layoutstat(struct inode
*inode
, gfp_t gfp_flags
)
2532 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(inode
)->pnfs_curr_ld
;
2533 struct nfs_server
*server
= NFS_SERVER(inode
);
2534 struct nfs_inode
*nfsi
= NFS_I(inode
);
2535 struct nfs42_layoutstat_data
*data
;
2536 struct pnfs_layout_hdr
*hdr
;
2539 if (!pnfs_enabled_sb(server
) || !ld
->prepare_layoutstats
)
2542 if (!nfs_server_capable(inode
, NFS_CAP_LAYOUTSTATS
))
2545 if (test_and_set_bit(NFS_INO_LAYOUTSTATS
, &nfsi
->flags
))
2548 spin_lock(&inode
->i_lock
);
2549 if (!NFS_I(inode
)->layout
) {
2550 spin_unlock(&inode
->i_lock
);
2551 goto out_clear_layoutstats
;
2553 hdr
= NFS_I(inode
)->layout
;
2554 pnfs_get_layout_hdr(hdr
);
2555 spin_unlock(&inode
->i_lock
);
2557 data
= kzalloc(sizeof(*data
), gfp_flags
);
2563 data
->args
.fh
= NFS_FH(inode
);
2564 data
->args
.inode
= inode
;
2565 status
= ld
->prepare_layoutstats(&data
->args
);
2569 status
= nfs42_proc_layoutstats_generic(NFS_SERVER(inode
), data
);
2572 dprintk("%s returns %d\n", __func__
, status
);
2578 pnfs_put_layout_hdr(hdr
);
2579 out_clear_layoutstats
:
2580 smp_mb__before_atomic();
2581 clear_bit(NFS_INO_LAYOUTSTATS
, &nfsi
->flags
);
2582 smp_mb__after_atomic();
2585 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat
);
2588 unsigned int layoutstats_timer
;
2589 module_param(layoutstats_timer
, uint
, 0644);
2590 EXPORT_SYMBOL_GPL(layoutstats_timer
);