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>
36 #include "nfs4trace.h"
38 #define NFSDBG_FACILITY NFSDBG_PNFS
39 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
44 * protects pnfs_modules_tbl.
46 static DEFINE_SPINLOCK(pnfs_spinlock
);
49 * pnfs_modules_tbl holds all pnfs modules
51 static LIST_HEAD(pnfs_modules_tbl
);
53 /* Return the registered pnfs layout driver module matching given id */
54 static struct pnfs_layoutdriver_type
*
55 find_pnfs_driver_locked(u32 id
)
57 struct pnfs_layoutdriver_type
*local
;
59 list_for_each_entry(local
, &pnfs_modules_tbl
, pnfs_tblid
)
64 dprintk("%s: Searching for id %u, found %p\n", __func__
, id
, local
);
68 static struct pnfs_layoutdriver_type
*
69 find_pnfs_driver(u32 id
)
71 struct pnfs_layoutdriver_type
*local
;
73 spin_lock(&pnfs_spinlock
);
74 local
= find_pnfs_driver_locked(id
);
75 if (local
!= NULL
&& !try_module_get(local
->owner
)) {
76 dprintk("%s: Could not grab reference on module\n", __func__
);
79 spin_unlock(&pnfs_spinlock
);
84 unset_pnfs_layoutdriver(struct nfs_server
*nfss
)
86 if (nfss
->pnfs_curr_ld
) {
87 if (nfss
->pnfs_curr_ld
->clear_layoutdriver
)
88 nfss
->pnfs_curr_ld
->clear_layoutdriver(nfss
);
89 /* Decrement the MDS count. Purge the deviceid cache if zero */
90 if (atomic_dec_and_test(&nfss
->nfs_client
->cl_mds_count
))
91 nfs4_deviceid_purge_client(nfss
->nfs_client
);
92 module_put(nfss
->pnfs_curr_ld
->owner
);
94 nfss
->pnfs_curr_ld
= NULL
;
98 * Try to set the server's pnfs module to the pnfs layout type specified by id.
99 * Currently only one pNFS layout driver per filesystem is supported.
101 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
104 set_pnfs_layoutdriver(struct nfs_server
*server
, const struct nfs_fh
*mntfh
,
107 struct pnfs_layoutdriver_type
*ld_type
= NULL
;
111 if (!(server
->nfs_client
->cl_exchange_flags
&
112 (EXCHGID4_FLAG_USE_NON_PNFS
| EXCHGID4_FLAG_USE_PNFS_MDS
))) {
113 printk(KERN_ERR
"NFS: %s: id %u cl_exchange_flags 0x%x\n",
114 __func__
, id
, server
->nfs_client
->cl_exchange_flags
);
117 ld_type
= find_pnfs_driver(id
);
119 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX
, id
);
120 ld_type
= find_pnfs_driver(id
);
122 dprintk("%s: No pNFS module found for %u.\n",
127 server
->pnfs_curr_ld
= ld_type
;
128 if (ld_type
->set_layoutdriver
129 && ld_type
->set_layoutdriver(server
, mntfh
)) {
130 printk(KERN_ERR
"NFS: %s: Error initializing pNFS layout "
131 "driver %u.\n", __func__
, id
);
132 module_put(ld_type
->owner
);
135 /* Bump the MDS count */
136 atomic_inc(&server
->nfs_client
->cl_mds_count
);
138 dprintk("%s: pNFS module for %u set\n", __func__
, id
);
142 dprintk("%s: Using NFSv4 I/O\n", __func__
);
143 server
->pnfs_curr_ld
= NULL
;
147 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
149 int status
= -EINVAL
;
150 struct pnfs_layoutdriver_type
*tmp
;
152 if (ld_type
->id
== 0) {
153 printk(KERN_ERR
"NFS: %s id 0 is reserved\n", __func__
);
156 if (!ld_type
->alloc_lseg
|| !ld_type
->free_lseg
) {
157 printk(KERN_ERR
"NFS: %s Layout driver must provide "
158 "alloc_lseg and free_lseg.\n", __func__
);
162 spin_lock(&pnfs_spinlock
);
163 tmp
= find_pnfs_driver_locked(ld_type
->id
);
165 list_add(&ld_type
->pnfs_tblid
, &pnfs_modules_tbl
);
167 dprintk("%s Registering id:%u name:%s\n", __func__
, ld_type
->id
,
170 printk(KERN_ERR
"NFS: %s Module with id %d already loaded!\n",
171 __func__
, ld_type
->id
);
173 spin_unlock(&pnfs_spinlock
);
177 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver
);
180 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type
*ld_type
)
182 dprintk("%s Deregistering id:%u\n", __func__
, ld_type
->id
);
183 spin_lock(&pnfs_spinlock
);
184 list_del(&ld_type
->pnfs_tblid
);
185 spin_unlock(&pnfs_spinlock
);
187 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver
);
190 * pNFS client layout cache
193 /* Need to hold i_lock if caller does not already hold reference */
195 pnfs_get_layout_hdr(struct pnfs_layout_hdr
*lo
)
197 atomic_inc(&lo
->plh_refcount
);
200 static struct pnfs_layout_hdr
*
201 pnfs_alloc_layout_hdr(struct inode
*ino
, gfp_t gfp_flags
)
203 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(ino
)->pnfs_curr_ld
;
204 return ld
->alloc_layout_hdr(ino
, 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
->free_layout_hdr(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 clear_bit(NFS_LAYOUT_BULK_RECALL
, &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
);
364 static void pnfs_free_lseg_async_work(struct work_struct
*work
)
366 struct pnfs_layout_segment
*lseg
;
367 struct pnfs_layout_hdr
*lo
;
369 lseg
= container_of(work
, struct pnfs_layout_segment
, pls_work
);
370 lo
= lseg
->pls_layout
;
372 pnfs_free_lseg(lseg
);
373 pnfs_put_layout_hdr(lo
);
376 static void pnfs_free_lseg_async(struct pnfs_layout_segment
*lseg
)
378 INIT_WORK(&lseg
->pls_work
, pnfs_free_lseg_async_work
);
379 schedule_work(&lseg
->pls_work
);
383 pnfs_put_lseg_locked(struct pnfs_layout_segment
*lseg
)
388 assert_spin_locked(&lseg
->pls_layout
->plh_inode
->i_lock
);
390 dprintk("%s: lseg %p ref %d valid %d\n", __func__
, lseg
,
391 atomic_read(&lseg
->pls_refcount
),
392 test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
));
393 if (atomic_dec_and_test(&lseg
->pls_refcount
)) {
394 struct pnfs_layout_hdr
*lo
= lseg
->pls_layout
;
395 pnfs_get_layout_hdr(lo
);
396 pnfs_layout_remove_lseg(lo
, lseg
);
397 pnfs_free_lseg_async(lseg
);
400 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked
);
403 end_offset(u64 start
, u64 len
)
408 return end
>= start
? end
: NFS4_MAX_UINT64
;
412 * is l2 fully contained in l1?
414 * [----------------------------------)
419 pnfs_lseg_range_contained(const struct pnfs_layout_range
*l1
,
420 const struct pnfs_layout_range
*l2
)
422 u64 start1
= l1
->offset
;
423 u64 end1
= end_offset(start1
, l1
->length
);
424 u64 start2
= l2
->offset
;
425 u64 end2
= end_offset(start2
, l2
->length
);
427 return (start1
<= start2
) && (end1
>= end2
);
431 * is l1 and l2 intersecting?
433 * [----------------------------------)
438 pnfs_lseg_range_intersecting(const struct pnfs_layout_range
*l1
,
439 const struct pnfs_layout_range
*l2
)
441 u64 start1
= l1
->offset
;
442 u64 end1
= end_offset(start1
, l1
->length
);
443 u64 start2
= l2
->offset
;
444 u64 end2
= end_offset(start2
, l2
->length
);
446 return (end1
== NFS4_MAX_UINT64
|| end1
> start2
) &&
447 (end2
== NFS4_MAX_UINT64
|| end2
> start1
);
451 should_free_lseg(const struct pnfs_layout_range
*lseg_range
,
452 const struct pnfs_layout_range
*recall_range
)
454 return (recall_range
->iomode
== IOMODE_ANY
||
455 lseg_range
->iomode
== recall_range
->iomode
) &&
456 pnfs_lseg_range_intersecting(lseg_range
, recall_range
);
459 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment
*lseg
,
460 struct list_head
*tmp_list
)
462 if (!atomic_dec_and_test(&lseg
->pls_refcount
))
464 pnfs_layout_remove_lseg(lseg
->pls_layout
, lseg
);
465 list_add(&lseg
->pls_list
, tmp_list
);
469 /* Returns 1 if lseg is removed from list, 0 otherwise */
470 static int mark_lseg_invalid(struct pnfs_layout_segment
*lseg
,
471 struct list_head
*tmp_list
)
475 if (test_and_clear_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
)) {
476 /* Remove the reference keeping the lseg in the
477 * list. It will now be removed when all
478 * outstanding io is finished.
480 dprintk("%s: lseg %p ref %d\n", __func__
, lseg
,
481 atomic_read(&lseg
->pls_refcount
));
482 if (pnfs_lseg_dec_and_remove_zero(lseg
, tmp_list
))
488 /* Returns count of number of matching invalid lsegs remaining in list
492 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr
*lo
,
493 struct list_head
*tmp_list
,
494 struct pnfs_layout_range
*recall_range
)
496 struct pnfs_layout_segment
*lseg
, *next
;
497 int invalid
= 0, removed
= 0;
499 dprintk("%s:Begin lo %p\n", __func__
, lo
);
501 if (list_empty(&lo
->plh_segs
))
503 list_for_each_entry_safe(lseg
, next
, &lo
->plh_segs
, pls_list
)
505 should_free_lseg(&lseg
->pls_range
, recall_range
)) {
506 dprintk("%s: freeing lseg %p iomode %d "
507 "offset %llu length %llu\n", __func__
,
508 lseg
, lseg
->pls_range
.iomode
, lseg
->pls_range
.offset
,
509 lseg
->pls_range
.length
);
511 removed
+= mark_lseg_invalid(lseg
, tmp_list
);
513 dprintk("%s:Return %i\n", __func__
, invalid
- removed
);
514 return invalid
- removed
;
517 /* note free_me must contain lsegs from a single layout_hdr */
519 pnfs_free_lseg_list(struct list_head
*free_me
)
521 struct pnfs_layout_segment
*lseg
, *tmp
;
523 if (list_empty(free_me
))
526 list_for_each_entry_safe(lseg
, tmp
, free_me
, pls_list
) {
527 list_del(&lseg
->pls_list
);
528 pnfs_free_lseg(lseg
);
533 pnfs_destroy_layout(struct nfs_inode
*nfsi
)
535 struct pnfs_layout_hdr
*lo
;
538 spin_lock(&nfsi
->vfs_inode
.i_lock
);
541 lo
->plh_block_lgets
++; /* permanently block new LAYOUTGETs */
542 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
543 pnfs_get_layout_hdr(lo
);
544 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RO_FAILED
);
545 pnfs_layout_clear_fail_bit(lo
, NFS_LAYOUT_RW_FAILED
);
546 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
547 pnfs_free_lseg_list(&tmp_list
);
548 pnfs_put_layout_hdr(lo
);
550 spin_unlock(&nfsi
->vfs_inode
.i_lock
);
552 EXPORT_SYMBOL_GPL(pnfs_destroy_layout
);
555 pnfs_layout_add_bulk_destroy_list(struct inode
*inode
,
556 struct list_head
*layout_list
)
558 struct pnfs_layout_hdr
*lo
;
561 spin_lock(&inode
->i_lock
);
562 lo
= NFS_I(inode
)->layout
;
563 if (lo
!= NULL
&& list_empty(&lo
->plh_bulk_destroy
)) {
564 pnfs_get_layout_hdr(lo
);
565 list_add(&lo
->plh_bulk_destroy
, layout_list
);
568 spin_unlock(&inode
->i_lock
);
572 /* Caller must hold rcu_read_lock and clp->cl_lock */
574 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client
*clp
,
575 struct nfs_server
*server
,
576 struct list_head
*layout_list
)
578 struct pnfs_layout_hdr
*lo
, *next
;
581 list_for_each_entry_safe(lo
, next
, &server
->layouts
, plh_layouts
) {
582 inode
= igrab(lo
->plh_inode
);
585 list_del_init(&lo
->plh_layouts
);
586 if (pnfs_layout_add_bulk_destroy_list(inode
, layout_list
))
589 spin_unlock(&clp
->cl_lock
);
591 spin_lock(&clp
->cl_lock
);
599 pnfs_layout_free_bulk_destroy_list(struct list_head
*layout_list
,
602 struct pnfs_layout_hdr
*lo
;
604 struct pnfs_layout_range range
= {
605 .iomode
= IOMODE_ANY
,
607 .length
= NFS4_MAX_UINT64
,
609 LIST_HEAD(lseg_list
);
612 while (!list_empty(layout_list
)) {
613 lo
= list_entry(layout_list
->next
, struct pnfs_layout_hdr
,
615 dprintk("%s freeing layout for inode %lu\n", __func__
,
616 lo
->plh_inode
->i_ino
);
617 inode
= lo
->plh_inode
;
619 pnfs_layoutcommit_inode(inode
, false);
621 spin_lock(&inode
->i_lock
);
622 list_del_init(&lo
->plh_bulk_destroy
);
623 lo
->plh_block_lgets
++; /* permanently block new LAYOUTGETs */
625 set_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
);
626 if (pnfs_mark_matching_lsegs_invalid(lo
, &lseg_list
, &range
))
628 spin_unlock(&inode
->i_lock
);
629 pnfs_free_lseg_list(&lseg_list
);
630 pnfs_put_layout_hdr(lo
);
637 pnfs_destroy_layouts_byfsid(struct nfs_client
*clp
,
638 struct nfs_fsid
*fsid
,
641 struct nfs_server
*server
;
642 LIST_HEAD(layout_list
);
644 spin_lock(&clp
->cl_lock
);
647 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
648 if (memcmp(&server
->fsid
, fsid
, sizeof(*fsid
)) != 0)
650 if (pnfs_layout_bulk_destroy_byserver_locked(clp
,
656 spin_unlock(&clp
->cl_lock
);
658 if (list_empty(&layout_list
))
660 return pnfs_layout_free_bulk_destroy_list(&layout_list
, is_recall
);
664 pnfs_destroy_layouts_byclid(struct nfs_client
*clp
,
667 struct nfs_server
*server
;
668 LIST_HEAD(layout_list
);
670 spin_lock(&clp
->cl_lock
);
673 list_for_each_entry_rcu(server
, &clp
->cl_superblocks
, client_link
) {
674 if (pnfs_layout_bulk_destroy_byserver_locked(clp
,
680 spin_unlock(&clp
->cl_lock
);
682 if (list_empty(&layout_list
))
684 return pnfs_layout_free_bulk_destroy_list(&layout_list
, is_recall
);
688 * Called by the state manger to remove all layouts established under an
692 pnfs_destroy_all_layouts(struct nfs_client
*clp
)
694 nfs4_deviceid_mark_client_invalid(clp
);
695 nfs4_deviceid_purge_client(clp
);
697 pnfs_destroy_layouts_byclid(clp
, false);
701 * Compare 2 layout stateid sequence ids, to see which is newer,
702 * taking into account wraparound issues.
704 static bool pnfs_seqid_is_newer(u32 s1
, u32 s2
)
706 return (s32
)(s1
- s2
) > 0;
709 /* update lo->plh_stateid with new if is more recent */
711 pnfs_set_layout_stateid(struct pnfs_layout_hdr
*lo
, const nfs4_stateid
*new,
714 u32 oldseq
, newseq
, new_barrier
;
715 int empty
= list_empty(&lo
->plh_segs
);
717 oldseq
= be32_to_cpu(lo
->plh_stateid
.seqid
);
718 newseq
= be32_to_cpu(new->seqid
);
719 if (empty
|| pnfs_seqid_is_newer(newseq
, oldseq
)) {
720 nfs4_stateid_copy(&lo
->plh_stateid
, new);
721 if (update_barrier
) {
722 new_barrier
= be32_to_cpu(new->seqid
);
724 /* Because of wraparound, we want to keep the barrier
725 * "close" to the current seqids.
727 new_barrier
= newseq
- atomic_read(&lo
->plh_outstanding
);
729 if (empty
|| pnfs_seqid_is_newer(new_barrier
, lo
->plh_barrier
))
730 lo
->plh_barrier
= new_barrier
;
735 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr
*lo
,
736 const nfs4_stateid
*stateid
)
738 u32 seqid
= be32_to_cpu(stateid
->seqid
);
740 return !pnfs_seqid_is_newer(seqid
, lo
->plh_barrier
);
743 /* lget is set to 1 if called from inside send_layoutget call chain */
745 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr
*lo
, int lget
)
747 return lo
->plh_block_lgets
||
748 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
) ||
749 (list_empty(&lo
->plh_segs
) &&
750 (atomic_read(&lo
->plh_outstanding
) > lget
));
754 pnfs_choose_layoutget_stateid(nfs4_stateid
*dst
, struct pnfs_layout_hdr
*lo
,
755 struct nfs4_state
*open_state
)
759 dprintk("--> %s\n", __func__
);
760 spin_lock(&lo
->plh_inode
->i_lock
);
761 if (pnfs_layoutgets_blocked(lo
, 1)) {
763 } else if (!nfs4_valid_open_stateid(open_state
)) {
765 } else if (list_empty(&lo
->plh_segs
) ||
766 test_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
)) {
770 seq
= read_seqbegin(&open_state
->seqlock
);
771 nfs4_stateid_copy(dst
, &open_state
->stateid
);
772 } while (read_seqretry(&open_state
->seqlock
, seq
));
774 nfs4_stateid_copy(dst
, &lo
->plh_stateid
);
775 spin_unlock(&lo
->plh_inode
->i_lock
);
776 dprintk("<-- %s\n", __func__
);
781 * Get layout from server.
782 * for now, assume that whole file layouts are requested.
784 * arg->length: all ones
786 static struct pnfs_layout_segment
*
787 send_layoutget(struct pnfs_layout_hdr
*lo
,
788 struct nfs_open_context
*ctx
,
789 struct pnfs_layout_range
*range
,
792 struct inode
*ino
= lo
->plh_inode
;
793 struct nfs_server
*server
= NFS_SERVER(ino
);
794 struct nfs4_layoutget
*lgp
;
795 struct pnfs_layout_segment
*lseg
;
797 dprintk("--> %s\n", __func__
);
799 lgp
= kzalloc(sizeof(*lgp
), gfp_flags
);
803 lgp
->args
.minlength
= PAGE_CACHE_SIZE
;
804 if (lgp
->args
.minlength
> range
->length
)
805 lgp
->args
.minlength
= range
->length
;
806 lgp
->args
.maxcount
= PNFS_LAYOUT_MAXSIZE
;
807 lgp
->args
.range
= *range
;
808 lgp
->args
.type
= server
->pnfs_curr_ld
->id
;
809 lgp
->args
.inode
= ino
;
810 lgp
->args
.ctx
= get_nfs_open_context(ctx
);
811 lgp
->gfp_flags
= gfp_flags
;
812 lgp
->cred
= lo
->plh_lc_cred
;
814 /* Synchronously retrieve layout information from server and
817 lseg
= nfs4_proc_layoutget(lgp
, gfp_flags
);
819 switch (PTR_ERR(lseg
)) {
824 /* remember that LAYOUTGET failed and suspend trying */
825 pnfs_layout_io_set_failed(lo
, range
->iomode
);
833 static void pnfs_clear_layoutcommit(struct inode
*inode
,
834 struct list_head
*head
)
836 struct nfs_inode
*nfsi
= NFS_I(inode
);
837 struct pnfs_layout_segment
*lseg
, *tmp
;
839 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
841 list_for_each_entry_safe(lseg
, tmp
, &nfsi
->layout
->plh_segs
, pls_list
) {
842 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
844 pnfs_lseg_dec_and_remove_zero(lseg
, head
);
849 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
850 * when the layout segment list is empty.
852 * Note that a pnfs_layout_hdr can exist with an empty layout segment
853 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
854 * deviceid is marked invalid.
857 _pnfs_return_layout(struct inode
*ino
)
859 struct pnfs_layout_hdr
*lo
= NULL
;
860 struct nfs_inode
*nfsi
= NFS_I(ino
);
862 struct nfs4_layoutreturn
*lrp
;
863 nfs4_stateid stateid
;
864 int status
= 0, empty
;
866 dprintk("NFS: %s for inode %lu\n", __func__
, ino
->i_ino
);
868 spin_lock(&ino
->i_lock
);
871 spin_unlock(&ino
->i_lock
);
872 dprintk("NFS: %s no layout to return\n", __func__
);
875 stateid
= nfsi
->layout
->plh_stateid
;
876 /* Reference matched in nfs4_layoutreturn_release */
877 pnfs_get_layout_hdr(lo
);
878 empty
= list_empty(&lo
->plh_segs
);
879 pnfs_clear_layoutcommit(ino
, &tmp_list
);
880 pnfs_mark_matching_lsegs_invalid(lo
, &tmp_list
, NULL
);
882 if (NFS_SERVER(ino
)->pnfs_curr_ld
->return_range
) {
883 struct pnfs_layout_range range
= {
884 .iomode
= IOMODE_ANY
,
886 .length
= NFS4_MAX_UINT64
,
888 NFS_SERVER(ino
)->pnfs_curr_ld
->return_range(lo
, &range
);
891 /* Don't send a LAYOUTRETURN if list was initially empty */
893 spin_unlock(&ino
->i_lock
);
894 pnfs_put_layout_hdr(lo
);
895 dprintk("NFS: %s no layout segments to return\n", __func__
);
899 set_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
);
900 lo
->plh_block_lgets
++;
901 spin_unlock(&ino
->i_lock
);
902 pnfs_free_lseg_list(&tmp_list
);
904 lrp
= kzalloc(sizeof(*lrp
), GFP_KERNEL
);
905 if (unlikely(lrp
== NULL
)) {
907 spin_lock(&ino
->i_lock
);
908 lo
->plh_block_lgets
--;
909 spin_unlock(&ino
->i_lock
);
910 pnfs_put_layout_hdr(lo
);
914 lrp
->args
.stateid
= stateid
;
915 lrp
->args
.layout_type
= NFS_SERVER(ino
)->pnfs_curr_ld
->id
;
916 lrp
->args
.inode
= ino
;
917 lrp
->args
.layout
= lo
;
918 lrp
->clp
= NFS_SERVER(ino
)->nfs_client
;
919 lrp
->cred
= lo
->plh_lc_cred
;
921 status
= nfs4_proc_layoutreturn(lrp
);
923 dprintk("<-- %s status: %d\n", __func__
, status
);
926 EXPORT_SYMBOL_GPL(_pnfs_return_layout
);
929 pnfs_commit_and_return_layout(struct inode
*inode
)
931 struct pnfs_layout_hdr
*lo
;
934 spin_lock(&inode
->i_lock
);
935 lo
= NFS_I(inode
)->layout
;
937 spin_unlock(&inode
->i_lock
);
940 pnfs_get_layout_hdr(lo
);
941 /* Block new layoutgets and read/write to ds */
942 lo
->plh_block_lgets
++;
943 spin_unlock(&inode
->i_lock
);
944 filemap_fdatawait(inode
->i_mapping
);
945 ret
= pnfs_layoutcommit_inode(inode
, true);
947 ret
= _pnfs_return_layout(inode
);
948 spin_lock(&inode
->i_lock
);
949 lo
->plh_block_lgets
--;
950 spin_unlock(&inode
->i_lock
);
951 pnfs_put_layout_hdr(lo
);
955 bool pnfs_roc(struct inode
*ino
)
957 struct pnfs_layout_hdr
*lo
;
958 struct pnfs_layout_segment
*lseg
, *tmp
;
962 spin_lock(&ino
->i_lock
);
963 lo
= NFS_I(ino
)->layout
;
964 if (!lo
|| !test_and_clear_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
) ||
965 test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
))
967 list_for_each_entry_safe(lseg
, tmp
, &lo
->plh_segs
, pls_list
)
968 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
969 mark_lseg_invalid(lseg
, &tmp_list
);
974 lo
->plh_block_lgets
++;
975 pnfs_get_layout_hdr(lo
); /* matched in pnfs_roc_release */
976 spin_unlock(&ino
->i_lock
);
977 pnfs_free_lseg_list(&tmp_list
);
981 spin_unlock(&ino
->i_lock
);
985 void pnfs_roc_release(struct inode
*ino
)
987 struct pnfs_layout_hdr
*lo
;
989 spin_lock(&ino
->i_lock
);
990 lo
= NFS_I(ino
)->layout
;
991 lo
->plh_block_lgets
--;
992 if (atomic_dec_and_test(&lo
->plh_refcount
)) {
993 pnfs_detach_layout_hdr(lo
);
994 spin_unlock(&ino
->i_lock
);
995 pnfs_free_layout_hdr(lo
);
997 spin_unlock(&ino
->i_lock
);
1000 void pnfs_roc_set_barrier(struct inode
*ino
, u32 barrier
)
1002 struct pnfs_layout_hdr
*lo
;
1004 spin_lock(&ino
->i_lock
);
1005 lo
= NFS_I(ino
)->layout
;
1006 if (pnfs_seqid_is_newer(barrier
, lo
->plh_barrier
))
1007 lo
->plh_barrier
= barrier
;
1008 spin_unlock(&ino
->i_lock
);
1011 bool pnfs_roc_drain(struct inode
*ino
, u32
*barrier
, struct rpc_task
*task
)
1013 struct nfs_inode
*nfsi
= NFS_I(ino
);
1014 struct pnfs_layout_hdr
*lo
;
1015 struct pnfs_layout_segment
*lseg
;
1019 spin_lock(&ino
->i_lock
);
1020 list_for_each_entry(lseg
, &nfsi
->layout
->plh_segs
, pls_list
)
1021 if (test_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
)) {
1022 rpc_sleep_on(&NFS_SERVER(ino
)->roc_rpcwaitq
, task
, NULL
);
1027 current_seqid
= be32_to_cpu(lo
->plh_stateid
.seqid
);
1029 /* Since close does not return a layout stateid for use as
1030 * a barrier, we choose the worst-case barrier.
1032 *barrier
= current_seqid
+ atomic_read(&lo
->plh_outstanding
);
1034 spin_unlock(&ino
->i_lock
);
1039 * Compare two layout segments for sorting into layout cache.
1040 * We want to preferentially return RW over RO layouts, so ensure those
1044 pnfs_lseg_range_cmp(const struct pnfs_layout_range
*l1
,
1045 const struct pnfs_layout_range
*l2
)
1049 /* high offset > low offset */
1050 d
= l1
->offset
- l2
->offset
;
1054 /* short length > long length */
1055 d
= l2
->length
- l1
->length
;
1059 /* read > read/write */
1060 return (int)(l1
->iomode
== IOMODE_READ
) - (int)(l2
->iomode
== IOMODE_READ
);
1064 pnfs_layout_insert_lseg(struct pnfs_layout_hdr
*lo
,
1065 struct pnfs_layout_segment
*lseg
)
1067 struct pnfs_layout_segment
*lp
;
1069 dprintk("%s:Begin\n", __func__
);
1071 list_for_each_entry(lp
, &lo
->plh_segs
, pls_list
) {
1072 if (pnfs_lseg_range_cmp(&lseg
->pls_range
, &lp
->pls_range
) > 0)
1074 list_add_tail(&lseg
->pls_list
, &lp
->pls_list
);
1075 dprintk("%s: inserted lseg %p "
1076 "iomode %d offset %llu length %llu before "
1077 "lp %p iomode %d offset %llu length %llu\n",
1078 __func__
, lseg
, lseg
->pls_range
.iomode
,
1079 lseg
->pls_range
.offset
, lseg
->pls_range
.length
,
1080 lp
, lp
->pls_range
.iomode
, lp
->pls_range
.offset
,
1081 lp
->pls_range
.length
);
1084 list_add_tail(&lseg
->pls_list
, &lo
->plh_segs
);
1085 dprintk("%s: inserted lseg %p "
1086 "iomode %d offset %llu length %llu at tail\n",
1087 __func__
, lseg
, lseg
->pls_range
.iomode
,
1088 lseg
->pls_range
.offset
, lseg
->pls_range
.length
);
1090 pnfs_get_layout_hdr(lo
);
1092 dprintk("%s:Return\n", __func__
);
1095 static struct pnfs_layout_hdr
*
1096 alloc_init_layout_hdr(struct inode
*ino
,
1097 struct nfs_open_context
*ctx
,
1100 struct pnfs_layout_hdr
*lo
;
1102 lo
= pnfs_alloc_layout_hdr(ino
, gfp_flags
);
1105 atomic_set(&lo
->plh_refcount
, 1);
1106 INIT_LIST_HEAD(&lo
->plh_layouts
);
1107 INIT_LIST_HEAD(&lo
->plh_segs
);
1108 INIT_LIST_HEAD(&lo
->plh_bulk_destroy
);
1109 lo
->plh_inode
= ino
;
1110 lo
->plh_lc_cred
= get_rpccred(ctx
->cred
);
1114 static struct pnfs_layout_hdr
*
1115 pnfs_find_alloc_layout(struct inode
*ino
,
1116 struct nfs_open_context
*ctx
,
1119 struct nfs_inode
*nfsi
= NFS_I(ino
);
1120 struct pnfs_layout_hdr
*new = NULL
;
1122 dprintk("%s Begin ino=%p layout=%p\n", __func__
, ino
, nfsi
->layout
);
1124 if (nfsi
->layout
!= NULL
)
1126 spin_unlock(&ino
->i_lock
);
1127 new = alloc_init_layout_hdr(ino
, ctx
, gfp_flags
);
1128 spin_lock(&ino
->i_lock
);
1130 if (likely(nfsi
->layout
== NULL
)) { /* Won the race? */
1133 } else if (new != NULL
)
1134 pnfs_free_layout_hdr(new);
1136 pnfs_get_layout_hdr(nfsi
->layout
);
1137 return nfsi
->layout
;
1141 * iomode matching rules:
1152 pnfs_lseg_range_match(const struct pnfs_layout_range
*ls_range
,
1153 const struct pnfs_layout_range
*range
)
1155 struct pnfs_layout_range range1
;
1157 if ((range
->iomode
== IOMODE_RW
&&
1158 ls_range
->iomode
!= IOMODE_RW
) ||
1159 !pnfs_lseg_range_intersecting(ls_range
, range
))
1162 /* range1 covers only the first byte in the range */
1165 return pnfs_lseg_range_contained(ls_range
, &range1
);
1169 * lookup range in layout
1171 static struct pnfs_layout_segment
*
1172 pnfs_find_lseg(struct pnfs_layout_hdr
*lo
,
1173 struct pnfs_layout_range
*range
)
1175 struct pnfs_layout_segment
*lseg
, *ret
= NULL
;
1177 dprintk("%s:Begin\n", __func__
);
1179 list_for_each_entry(lseg
, &lo
->plh_segs
, pls_list
) {
1180 if (test_bit(NFS_LSEG_VALID
, &lseg
->pls_flags
) &&
1181 pnfs_lseg_range_match(&lseg
->pls_range
, range
)) {
1182 ret
= pnfs_get_lseg(lseg
);
1185 if (lseg
->pls_range
.offset
> range
->offset
)
1189 dprintk("%s:Return lseg %p ref %d\n",
1190 __func__
, ret
, ret
? atomic_read(&ret
->pls_refcount
) : 0);
1195 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1196 * to the MDS or over pNFS
1198 * The nfs_inode read_io and write_io fields are cumulative counters reset
1199 * when there are no layout segments. Note that in pnfs_update_layout iomode
1200 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1203 * A return of true means use MDS I/O.
1206 * If a file's size is smaller than the file size threshold, data accesses
1207 * SHOULD be sent to the metadata server. If an I/O request has a length that
1208 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1209 * server. If both file size and I/O size are provided, the client SHOULD
1210 * reach or exceed both thresholds before sending its read or write
1211 * requests to the data server.
1213 static bool pnfs_within_mdsthreshold(struct nfs_open_context
*ctx
,
1214 struct inode
*ino
, int iomode
)
1216 struct nfs4_threshold
*t
= ctx
->mdsthreshold
;
1217 struct nfs_inode
*nfsi
= NFS_I(ino
);
1218 loff_t fsize
= i_size_read(ino
);
1219 bool size
= false, size_set
= false, io
= false, io_set
= false, ret
= false;
1224 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1225 __func__
, t
->bm
, t
->rd_sz
, t
->wr_sz
, t
->rd_io_sz
, t
->wr_io_sz
);
1229 if (t
->bm
& THRESHOLD_RD
) {
1230 dprintk("%s fsize %llu\n", __func__
, fsize
);
1232 if (fsize
< t
->rd_sz
)
1235 if (t
->bm
& THRESHOLD_RD_IO
) {
1236 dprintk("%s nfsi->read_io %llu\n", __func__
,
1239 if (nfsi
->read_io
< t
->rd_io_sz
)
1244 if (t
->bm
& THRESHOLD_WR
) {
1245 dprintk("%s fsize %llu\n", __func__
, fsize
);
1247 if (fsize
< t
->wr_sz
)
1250 if (t
->bm
& THRESHOLD_WR_IO
) {
1251 dprintk("%s nfsi->write_io %llu\n", __func__
,
1254 if (nfsi
->write_io
< t
->wr_io_sz
)
1259 if (size_set
&& io_set
) {
1262 } else if (size
|| io
)
1265 dprintk("<-- %s size %d io %d ret %d\n", __func__
, size
, io
, ret
);
1270 * Layout segment is retreived from the server if not cached.
1271 * The appropriate layout segment is referenced and returned to the caller.
1273 struct pnfs_layout_segment
*
1274 pnfs_update_layout(struct inode
*ino
,
1275 struct nfs_open_context
*ctx
,
1278 enum pnfs_iomode iomode
,
1281 struct pnfs_layout_range arg
= {
1287 struct nfs_server
*server
= NFS_SERVER(ino
);
1288 struct nfs_client
*clp
= server
->nfs_client
;
1289 struct pnfs_layout_hdr
*lo
;
1290 struct pnfs_layout_segment
*lseg
= NULL
;
1293 if (!pnfs_enabled_sb(NFS_SERVER(ino
)))
1296 if (pnfs_within_mdsthreshold(ctx
, ino
, iomode
))
1299 spin_lock(&ino
->i_lock
);
1300 lo
= pnfs_find_alloc_layout(ino
, ctx
, gfp_flags
);
1302 spin_unlock(&ino
->i_lock
);
1306 /* Do we even need to bother with this? */
1307 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1308 dprintk("%s matches recall, use MDS\n", __func__
);
1312 /* if LAYOUTGET already failed once we don't try again */
1313 if (pnfs_layout_io_test_failed(lo
, iomode
))
1316 /* Check to see if the layout for the given range already exists */
1317 lseg
= pnfs_find_lseg(lo
, &arg
);
1321 if (pnfs_layoutgets_blocked(lo
, 0))
1323 atomic_inc(&lo
->plh_outstanding
);
1325 first
= list_empty(&lo
->plh_layouts
) ? true : false;
1326 spin_unlock(&ino
->i_lock
);
1329 /* The lo must be on the clp list if there is any
1330 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1332 spin_lock(&clp
->cl_lock
);
1333 list_add_tail(&lo
->plh_layouts
, &server
->layouts
);
1334 spin_unlock(&clp
->cl_lock
);
1337 pg_offset
= arg
.offset
& ~PAGE_CACHE_MASK
;
1339 arg
.offset
-= pg_offset
;
1340 arg
.length
+= pg_offset
;
1342 if (arg
.length
!= NFS4_MAX_UINT64
)
1343 arg
.length
= PAGE_CACHE_ALIGN(arg
.length
);
1345 lseg
= send_layoutget(lo
, ctx
, &arg
, gfp_flags
);
1346 atomic_dec(&lo
->plh_outstanding
);
1348 pnfs_put_layout_hdr(lo
);
1350 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1351 "(%s, offset: %llu, length: %llu)\n",
1352 __func__
, ino
->i_sb
->s_id
,
1353 (unsigned long long)NFS_FILEID(ino
),
1354 lseg
== NULL
? "not found" : "found",
1355 iomode
==IOMODE_RW
? "read/write" : "read-only",
1356 (unsigned long long)pos
,
1357 (unsigned long long)count
);
1360 spin_unlock(&ino
->i_lock
);
1361 goto out_put_layout_hdr
;
1363 EXPORT_SYMBOL_GPL(pnfs_update_layout
);
1365 struct pnfs_layout_segment
*
1366 pnfs_layout_process(struct nfs4_layoutget
*lgp
)
1368 struct pnfs_layout_hdr
*lo
= NFS_I(lgp
->args
.inode
)->layout
;
1369 struct nfs4_layoutget_res
*res
= &lgp
->res
;
1370 struct pnfs_layout_segment
*lseg
;
1371 struct inode
*ino
= lo
->plh_inode
;
1375 /* Inject layout blob into I/O device driver */
1376 lseg
= NFS_SERVER(ino
)->pnfs_curr_ld
->alloc_lseg(lo
, res
, lgp
->gfp_flags
);
1377 if (!lseg
|| IS_ERR(lseg
)) {
1381 status
= PTR_ERR(lseg
);
1382 dprintk("%s: Could not allocate layout: error %d\n",
1387 init_lseg(lo
, lseg
);
1388 lseg
->pls_range
= res
->range
;
1390 spin_lock(&ino
->i_lock
);
1391 if (test_bit(NFS_LAYOUT_BULK_RECALL
, &lo
->plh_flags
)) {
1392 dprintk("%s forget reply due to recall\n", __func__
);
1393 goto out_forget_reply
;
1396 if (pnfs_layoutgets_blocked(lo
, 1)) {
1397 dprintk("%s forget reply due to state\n", __func__
);
1398 goto out_forget_reply
;
1401 if (nfs4_stateid_match_other(&lo
->plh_stateid
, &res
->stateid
)) {
1402 /* existing state ID, make sure the sequence number matches. */
1403 if (pnfs_layout_stateid_blocked(lo
, &res
->stateid
)) {
1404 dprintk("%s forget reply due to sequence\n", __func__
);
1405 goto out_forget_reply
;
1407 pnfs_set_layout_stateid(lo
, &res
->stateid
, false);
1410 * We got an entirely new state ID. Mark all segments for the
1411 * inode invalid, and don't bother validating the stateid
1414 pnfs_mark_matching_lsegs_invalid(lo
, &free_me
, NULL
);
1416 nfs4_stateid_copy(&lo
->plh_stateid
, &res
->stateid
);
1417 lo
->plh_barrier
= be32_to_cpu(res
->stateid
.seqid
);
1420 clear_bit(NFS_LAYOUT_INVALID_STID
, &lo
->plh_flags
);
1422 pnfs_get_lseg(lseg
);
1423 pnfs_layout_insert_lseg(lo
, lseg
);
1425 if (res
->return_on_close
) {
1426 set_bit(NFS_LSEG_ROC
, &lseg
->pls_flags
);
1427 set_bit(NFS_LAYOUT_ROC
, &lo
->plh_flags
);
1430 spin_unlock(&ino
->i_lock
);
1431 pnfs_free_lseg_list(&free_me
);
1434 return ERR_PTR(status
);
1437 spin_unlock(&ino
->i_lock
);
1438 lseg
->pls_layout
= lo
;
1439 NFS_SERVER(ino
)->pnfs_curr_ld
->free_lseg(lseg
);
1444 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*req
)
1446 u64 rd_size
= req
->wb_bytes
;
1448 WARN_ON_ONCE(pgio
->pg_lseg
!= NULL
);
1450 if (pgio
->pg_dreq
== NULL
)
1451 rd_size
= i_size_read(pgio
->pg_inode
) - req_offset(req
);
1453 rd_size
= nfs_dreq_bytes_left(pgio
->pg_dreq
);
1455 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1461 /* If no lseg, fall back to read through mds */
1462 if (pgio
->pg_lseg
== NULL
)
1463 nfs_pageio_reset_read_mds(pgio
);
1466 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read
);
1469 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor
*pgio
,
1470 struct nfs_page
*req
, u64 wb_size
)
1472 WARN_ON_ONCE(pgio
->pg_lseg
!= NULL
);
1474 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
1480 /* If no lseg, fall back to write through mds */
1481 if (pgio
->pg_lseg
== NULL
)
1482 nfs_pageio_reset_write_mds(pgio
);
1484 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write
);
1487 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1488 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1491 pnfs_generic_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
1492 struct nfs_page
*req
)
1495 u64 seg_end
, req_start
, seg_left
;
1497 size
= nfs_generic_pg_test(pgio
, prev
, req
);
1502 * 'size' contains the number of bytes left in the current page (up
1503 * to the original size asked for in @req->wb_bytes).
1505 * Calculate how many bytes are left in the layout segment
1506 * and if there are less bytes than 'size', return that instead.
1508 * Please also note that 'end_offset' is actually the offset of the
1509 * first byte that lies outside the pnfs_layout_range. FIXME?
1512 if (pgio
->pg_lseg
) {
1513 seg_end
= end_offset(pgio
->pg_lseg
->pls_range
.offset
,
1514 pgio
->pg_lseg
->pls_range
.length
);
1515 req_start
= req_offset(req
);
1516 WARN_ON_ONCE(req_start
> seg_end
);
1517 /* start of request is past the last byte of this segment */
1518 if (req_start
>= seg_end
)
1521 /* adjust 'size' iff there are fewer bytes left in the
1522 * segment than what nfs_generic_pg_test returned */
1523 seg_left
= seg_end
- req_start
;
1524 if (seg_left
< size
)
1525 size
= (unsigned int)seg_left
;
1530 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test
);
1532 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header
*hdr
)
1534 struct nfs_pageio_descriptor pgio
;
1536 /* Resend all requests through the MDS */
1537 nfs_pageio_init_write(&pgio
, hdr
->inode
, FLUSH_STABLE
, true,
1538 hdr
->completion_ops
);
1539 return nfs_pageio_resend(&pgio
, hdr
);
1541 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds
);
1543 static void pnfs_ld_handle_write_error(struct nfs_pgio_header
*hdr
)
1546 dprintk("pnfs write error = %d\n", hdr
->pnfs_error
);
1547 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1548 PNFS_LAYOUTRET_ON_ERROR
) {
1549 pnfs_return_layout(hdr
->inode
);
1551 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1552 hdr
->task
.tk_status
= pnfs_write_done_resend_to_mds(hdr
);
1556 * Called by non rpc-based layout drivers
1558 void pnfs_ld_write_done(struct nfs_pgio_header
*hdr
)
1560 trace_nfs4_pnfs_write(hdr
, hdr
->pnfs_error
);
1561 if (!hdr
->pnfs_error
) {
1562 pnfs_set_layoutcommit(hdr
);
1563 hdr
->mds_ops
->rpc_call_done(&hdr
->task
, hdr
);
1565 pnfs_ld_handle_write_error(hdr
);
1566 hdr
->mds_ops
->rpc_release(hdr
);
1568 EXPORT_SYMBOL_GPL(pnfs_ld_write_done
);
1571 pnfs_write_through_mds(struct nfs_pageio_descriptor
*desc
,
1572 struct nfs_pgio_header
*hdr
)
1574 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1575 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1576 nfs_pageio_reset_write_mds(desc
);
1577 desc
->pg_recoalesce
= 1;
1579 nfs_pgio_data_destroy(hdr
);
1582 static enum pnfs_try_status
1583 pnfs_try_to_write_data(struct nfs_pgio_header
*hdr
,
1584 const struct rpc_call_ops
*call_ops
,
1585 struct pnfs_layout_segment
*lseg
,
1588 struct inode
*inode
= hdr
->inode
;
1589 enum pnfs_try_status trypnfs
;
1590 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1592 hdr
->mds_ops
= call_ops
;
1594 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__
,
1595 inode
->i_ino
, hdr
->args
.count
, hdr
->args
.offset
, how
);
1596 trypnfs
= nfss
->pnfs_curr_ld
->write_pagelist(hdr
, how
);
1597 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1598 nfs_inc_stats(inode
, NFSIOS_PNFS_WRITE
);
1599 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1604 pnfs_do_write(struct nfs_pageio_descriptor
*desc
,
1605 struct nfs_pgio_header
*hdr
, int how
)
1607 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1608 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1609 enum pnfs_try_status trypnfs
;
1611 desc
->pg_lseg
= NULL
;
1612 trypnfs
= pnfs_try_to_write_data(hdr
, call_ops
, lseg
, how
);
1613 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1614 pnfs_write_through_mds(desc
, hdr
);
1615 pnfs_put_lseg(lseg
);
1618 static void pnfs_writehdr_free(struct nfs_pgio_header
*hdr
)
1620 pnfs_put_lseg(hdr
->lseg
);
1621 nfs_pgio_header_free(hdr
);
1623 EXPORT_SYMBOL_GPL(pnfs_writehdr_free
);
1626 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor
*desc
)
1628 struct nfs_pgio_header
*hdr
;
1631 hdr
= nfs_pgio_header_alloc(desc
->pg_rw_ops
);
1633 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1634 pnfs_put_lseg(desc
->pg_lseg
);
1635 desc
->pg_lseg
= NULL
;
1638 nfs_pgheader_init(desc
, hdr
, pnfs_writehdr_free
);
1639 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1640 ret
= nfs_generic_pgio(desc
, hdr
);
1642 pnfs_put_lseg(desc
->pg_lseg
);
1643 desc
->pg_lseg
= NULL
;
1645 pnfs_do_write(desc
, hdr
, desc
->pg_ioflags
);
1648 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages
);
1650 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header
*hdr
)
1652 struct nfs_pageio_descriptor pgio
;
1654 /* Resend all requests through the MDS */
1655 nfs_pageio_init_read(&pgio
, hdr
->inode
, true, hdr
->completion_ops
);
1656 return nfs_pageio_resend(&pgio
, hdr
);
1658 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds
);
1660 static void pnfs_ld_handle_read_error(struct nfs_pgio_header
*hdr
)
1662 dprintk("pnfs read error = %d\n", hdr
->pnfs_error
);
1663 if (NFS_SERVER(hdr
->inode
)->pnfs_curr_ld
->flags
&
1664 PNFS_LAYOUTRET_ON_ERROR
) {
1665 pnfs_return_layout(hdr
->inode
);
1667 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
))
1668 hdr
->task
.tk_status
= pnfs_read_done_resend_to_mds(hdr
);
1672 * Called by non rpc-based layout drivers
1674 void pnfs_ld_read_done(struct nfs_pgio_header
*hdr
)
1676 trace_nfs4_pnfs_read(hdr
, hdr
->pnfs_error
);
1677 if (likely(!hdr
->pnfs_error
)) {
1678 __nfs4_read_done_cb(hdr
);
1679 hdr
->mds_ops
->rpc_call_done(&hdr
->task
, hdr
);
1681 pnfs_ld_handle_read_error(hdr
);
1682 hdr
->mds_ops
->rpc_release(hdr
);
1684 EXPORT_SYMBOL_GPL(pnfs_ld_read_done
);
1687 pnfs_read_through_mds(struct nfs_pageio_descriptor
*desc
,
1688 struct nfs_pgio_header
*hdr
)
1690 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
1691 list_splice_tail_init(&hdr
->pages
, &desc
->pg_list
);
1692 nfs_pageio_reset_read_mds(desc
);
1693 desc
->pg_recoalesce
= 1;
1695 nfs_pgio_data_destroy(hdr
);
1699 * Call the appropriate parallel I/O subsystem read function.
1701 static enum pnfs_try_status
1702 pnfs_try_to_read_data(struct nfs_pgio_header
*hdr
,
1703 const struct rpc_call_ops
*call_ops
,
1704 struct pnfs_layout_segment
*lseg
)
1706 struct inode
*inode
= hdr
->inode
;
1707 struct nfs_server
*nfss
= NFS_SERVER(inode
);
1708 enum pnfs_try_status trypnfs
;
1710 hdr
->mds_ops
= call_ops
;
1712 dprintk("%s: Reading ino:%lu %u@%llu\n",
1713 __func__
, inode
->i_ino
, hdr
->args
.count
, hdr
->args
.offset
);
1715 trypnfs
= nfss
->pnfs_curr_ld
->read_pagelist(hdr
);
1716 if (trypnfs
!= PNFS_NOT_ATTEMPTED
)
1717 nfs_inc_stats(inode
, NFSIOS_PNFS_READ
);
1718 dprintk("%s End (trypnfs:%d)\n", __func__
, trypnfs
);
1723 pnfs_do_read(struct nfs_pageio_descriptor
*desc
, struct nfs_pgio_header
*hdr
)
1725 const struct rpc_call_ops
*call_ops
= desc
->pg_rpc_callops
;
1726 struct pnfs_layout_segment
*lseg
= desc
->pg_lseg
;
1727 enum pnfs_try_status trypnfs
;
1729 desc
->pg_lseg
= NULL
;
1730 trypnfs
= pnfs_try_to_read_data(hdr
, call_ops
, lseg
);
1731 if (trypnfs
== PNFS_NOT_ATTEMPTED
)
1732 pnfs_read_through_mds(desc
, hdr
);
1733 pnfs_put_lseg(lseg
);
1736 static void pnfs_readhdr_free(struct nfs_pgio_header
*hdr
)
1738 pnfs_put_lseg(hdr
->lseg
);
1739 nfs_pgio_header_free(hdr
);
1741 EXPORT_SYMBOL_GPL(pnfs_readhdr_free
);
1744 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor
*desc
)
1746 struct nfs_pgio_header
*hdr
;
1749 hdr
= nfs_pgio_header_alloc(desc
->pg_rw_ops
);
1751 desc
->pg_completion_ops
->error_cleanup(&desc
->pg_list
);
1753 pnfs_put_lseg(desc
->pg_lseg
);
1754 desc
->pg_lseg
= NULL
;
1757 nfs_pgheader_init(desc
, hdr
, pnfs_readhdr_free
);
1758 hdr
->lseg
= pnfs_get_lseg(desc
->pg_lseg
);
1759 ret
= nfs_generic_pgio(desc
, hdr
);
1761 pnfs_put_lseg(desc
->pg_lseg
);
1762 desc
->pg_lseg
= NULL
;
1764 pnfs_do_read(desc
, hdr
);
1767 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages
);
1769 static void pnfs_clear_layoutcommitting(struct inode
*inode
)
1771 unsigned long *bitlock
= &NFS_I(inode
)->flags
;
1773 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING
, bitlock
);
1774 smp_mb__after_atomic();
1775 wake_up_bit(bitlock
, NFS_INO_LAYOUTCOMMITTING
);
1779 * There can be multiple RW segments.
1781 static void pnfs_list_write_lseg(struct inode
*inode
, struct list_head
*listp
)
1783 struct pnfs_layout_segment
*lseg
;
1785 list_for_each_entry(lseg
, &NFS_I(inode
)->layout
->plh_segs
, pls_list
) {
1786 if (lseg
->pls_range
.iomode
== IOMODE_RW
&&
1787 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT
, &lseg
->pls_flags
))
1788 list_add(&lseg
->pls_lc_list
, listp
);
1792 static void pnfs_list_write_lseg_done(struct inode
*inode
, struct list_head
*listp
)
1794 struct pnfs_layout_segment
*lseg
, *tmp
;
1796 /* Matched by references in pnfs_set_layoutcommit */
1797 list_for_each_entry_safe(lseg
, tmp
, listp
, pls_lc_list
) {
1798 list_del_init(&lseg
->pls_lc_list
);
1799 pnfs_put_lseg(lseg
);
1802 pnfs_clear_layoutcommitting(inode
);
1805 void pnfs_set_lo_fail(struct pnfs_layout_segment
*lseg
)
1807 pnfs_layout_io_set_failed(lseg
->pls_layout
, lseg
->pls_range
.iomode
);
1809 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail
);
1812 pnfs_set_layoutcommit(struct nfs_pgio_header
*hdr
)
1814 struct inode
*inode
= hdr
->inode
;
1815 struct nfs_inode
*nfsi
= NFS_I(inode
);
1816 loff_t end_pos
= hdr
->mds_offset
+ hdr
->res
.count
;
1817 bool mark_as_dirty
= false;
1819 spin_lock(&inode
->i_lock
);
1820 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1821 mark_as_dirty
= true;
1822 dprintk("%s: Set layoutcommit for inode %lu ",
1823 __func__
, inode
->i_ino
);
1825 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT
, &hdr
->lseg
->pls_flags
)) {
1826 /* references matched in nfs4_layoutcommit_release */
1827 pnfs_get_lseg(hdr
->lseg
);
1829 if (end_pos
> nfsi
->layout
->plh_lwb
)
1830 nfsi
->layout
->plh_lwb
= end_pos
;
1831 spin_unlock(&inode
->i_lock
);
1832 dprintk("%s: lseg %p end_pos %llu\n",
1833 __func__
, hdr
->lseg
, nfsi
->layout
->plh_lwb
);
1835 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1836 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1838 mark_inode_dirty_sync(inode
);
1840 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit
);
1842 void pnfs_commit_set_layoutcommit(struct nfs_commit_data
*data
)
1844 struct inode
*inode
= data
->inode
;
1845 struct nfs_inode
*nfsi
= NFS_I(inode
);
1846 bool mark_as_dirty
= false;
1848 spin_lock(&inode
->i_lock
);
1849 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
)) {
1850 mark_as_dirty
= true;
1851 dprintk("%s: Set layoutcommit for inode %lu ",
1852 __func__
, inode
->i_ino
);
1854 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT
, &data
->lseg
->pls_flags
)) {
1855 /* references matched in nfs4_layoutcommit_release */
1856 pnfs_get_lseg(data
->lseg
);
1858 if (data
->lwb
> nfsi
->layout
->plh_lwb
)
1859 nfsi
->layout
->plh_lwb
= data
->lwb
;
1860 spin_unlock(&inode
->i_lock
);
1861 dprintk("%s: lseg %p end_pos %llu\n",
1862 __func__
, data
->lseg
, nfsi
->layout
->plh_lwb
);
1864 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1865 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1867 mark_inode_dirty_sync(inode
);
1869 EXPORT_SYMBOL_GPL(pnfs_commit_set_layoutcommit
);
1871 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data
*data
)
1873 struct nfs_server
*nfss
= NFS_SERVER(data
->args
.inode
);
1875 if (nfss
->pnfs_curr_ld
->cleanup_layoutcommit
)
1876 nfss
->pnfs_curr_ld
->cleanup_layoutcommit(data
);
1877 pnfs_list_write_lseg_done(data
->args
.inode
, &data
->lseg_list
);
1881 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1882 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1883 * data to disk to allow the server to recover the data if it crashes.
1884 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1885 * is off, and a COMMIT is sent to a data server, or
1886 * if WRITEs to a data server return NFS_DATA_SYNC.
1889 pnfs_layoutcommit_inode(struct inode
*inode
, bool sync
)
1891 struct pnfs_layoutdriver_type
*ld
= NFS_SERVER(inode
)->pnfs_curr_ld
;
1892 struct nfs4_layoutcommit_data
*data
;
1893 struct nfs_inode
*nfsi
= NFS_I(inode
);
1897 if (!pnfs_layoutcommit_outstanding(inode
))
1900 dprintk("--> %s inode %lu\n", __func__
, inode
->i_ino
);
1903 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING
, &nfsi
->flags
)) {
1906 status
= wait_on_bit_lock_action(&nfsi
->flags
,
1907 NFS_INO_LAYOUTCOMMITTING
,
1908 nfs_wait_bit_killable
,
1915 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1916 data
= kzalloc(sizeof(*data
), GFP_NOFS
);
1918 goto clear_layoutcommitting
;
1921 spin_lock(&inode
->i_lock
);
1922 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
))
1925 INIT_LIST_HEAD(&data
->lseg_list
);
1926 pnfs_list_write_lseg(inode
, &data
->lseg_list
);
1928 end_pos
= nfsi
->layout
->plh_lwb
;
1929 nfsi
->layout
->plh_lwb
= 0;
1931 nfs4_stateid_copy(&data
->args
.stateid
, &nfsi
->layout
->plh_stateid
);
1932 spin_unlock(&inode
->i_lock
);
1934 data
->args
.inode
= inode
;
1935 data
->cred
= get_rpccred(nfsi
->layout
->plh_lc_cred
);
1936 nfs_fattr_init(&data
->fattr
);
1937 data
->args
.bitmask
= NFS_SERVER(inode
)->cache_consistency_bitmask
;
1938 data
->res
.fattr
= &data
->fattr
;
1939 data
->args
.lastbytewritten
= end_pos
- 1;
1940 data
->res
.server
= NFS_SERVER(inode
);
1942 if (ld
->prepare_layoutcommit
) {
1943 status
= ld
->prepare_layoutcommit(&data
->args
);
1945 spin_lock(&inode
->i_lock
);
1946 if (end_pos
< nfsi
->layout
->plh_lwb
)
1947 nfsi
->layout
->plh_lwb
= end_pos
;
1948 spin_unlock(&inode
->i_lock
);
1949 put_rpccred(data
->cred
);
1950 set_bit(NFS_INO_LAYOUTCOMMIT
, &nfsi
->flags
);
1951 goto clear_layoutcommitting
;
1956 status
= nfs4_proc_layoutcommit(data
, sync
);
1959 mark_inode_dirty_sync(inode
);
1960 dprintk("<-- %s status %d\n", __func__
, status
);
1963 spin_unlock(&inode
->i_lock
);
1965 clear_layoutcommitting
:
1966 pnfs_clear_layoutcommitting(inode
);
1970 struct nfs4_threshold
*pnfs_mdsthreshold_alloc(void)
1972 struct nfs4_threshold
*thp
;
1974 thp
= kzalloc(sizeof(*thp
), GFP_NOFS
);
1976 dprintk("%s mdsthreshold allocation failed\n", __func__
);