2 * Module for the pnfs nfs4 file layout driver.
3 * Defines all I/O and Policy interface operations, plus code
4 * to register itself with the pNFS client.
7 * The Regents of the University of Michigan
10 * Dean Hildebrand <dhildebz@umich.edu>
12 * Permission is granted to use, copy, create derivative works, and
13 * redistribute this software and such derivative works for any purpose,
14 * so long as the name of the University of Michigan is not used in
15 * any advertising or publicity pertaining to the use or distribution
16 * of this software without specific, written prior authorization. If
17 * the above copyright notice or any other identification of the
18 * University of Michigan is included in any copy of any portion of
19 * this software, then the disclaimer below must also be included.
21 * This software is provided as is, without representation or warranty
22 * of any kind either express or implied, including without limitation
23 * the implied warranties of merchantability, fitness for a particular
24 * purpose, or noninfringement. The Regents of the University of
25 * Michigan shall not be liable for any damages, including special,
26 * indirect, incidental, or consequential damages, with respect to any
27 * claim arising out of or in connection with the use of the software,
28 * even if it has been or is hereafter advised of the possibility of
32 #include <linux/nfs_fs.h>
33 #include <linux/nfs_page.h>
34 #include <linux/module.h>
36 #include <linux/sunrpc/metrics.h>
39 #include "delegation.h"
40 #include "nfs4filelayout.h"
42 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
44 MODULE_LICENSE("GPL");
45 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
46 MODULE_DESCRIPTION("The NFSv4 file layout driver");
48 #define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
51 filelayout_get_dense_offset(struct nfs4_filelayout_segment
*flseg
,
54 u32 stripe_width
= flseg
->stripe_unit
* flseg
->dsaddr
->stripe_count
;
58 offset
-= flseg
->pattern_offset
;
59 stripe_no
= div_u64(offset
, stripe_width
);
60 div_u64_rem(offset
, flseg
->stripe_unit
, &rem
);
62 return stripe_no
* flseg
->stripe_unit
+ rem
;
65 /* This function is used by the layout driver to calculate the
66 * offset of the file on the dserver based on whether the
67 * layout type is STRIPE_DENSE or STRIPE_SPARSE
70 filelayout_get_dserver_offset(struct pnfs_layout_segment
*lseg
, loff_t offset
)
72 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
74 switch (flseg
->stripe_type
) {
79 return filelayout_get_dense_offset(flseg
, offset
);
85 static void filelayout_reset_write(struct nfs_write_data
*data
)
87 struct nfs_pgio_header
*hdr
= data
->header
;
88 struct rpc_task
*task
= &data
->task
;
90 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
91 dprintk("%s Reset task %5u for i/o through MDS "
92 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__
,
94 hdr
->inode
->i_sb
->s_id
,
95 (long long)NFS_FILEID(hdr
->inode
),
97 (unsigned long long)data
->args
.offset
);
99 task
->tk_status
= pnfs_write_done_resend_to_mds(hdr
->inode
,
101 hdr
->completion_ops
);
105 static void filelayout_reset_read(struct nfs_read_data
*data
)
107 struct nfs_pgio_header
*hdr
= data
->header
;
108 struct rpc_task
*task
= &data
->task
;
110 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
111 dprintk("%s Reset task %5u for i/o through MDS "
112 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__
,
114 hdr
->inode
->i_sb
->s_id
,
115 (long long)NFS_FILEID(hdr
->inode
),
117 (unsigned long long)data
->args
.offset
);
119 task
->tk_status
= pnfs_read_done_resend_to_mds(hdr
->inode
,
121 hdr
->completion_ops
);
125 static void filelayout_fenceme(struct inode
*inode
, struct pnfs_layout_hdr
*lo
)
127 if (!test_and_clear_bit(NFS_LAYOUT_RETURN
, &lo
->plh_flags
))
129 clear_bit(NFS_INO_LAYOUTCOMMIT
, &NFS_I(inode
)->flags
);
130 pnfs_return_layout(inode
);
133 static int filelayout_async_handle_error(struct rpc_task
*task
,
134 struct nfs4_state
*state
,
135 struct nfs_client
*clp
,
136 struct pnfs_layout_segment
*lseg
)
138 struct pnfs_layout_hdr
*lo
= lseg
->pls_layout
;
139 struct inode
*inode
= lo
->plh_inode
;
140 struct nfs_server
*mds_server
= NFS_SERVER(inode
);
141 struct nfs4_deviceid_node
*devid
= FILELAYOUT_DEVID_NODE(lseg
);
142 struct nfs_client
*mds_client
= mds_server
->nfs_client
;
143 struct nfs4_slot_table
*tbl
= &clp
->cl_session
->fc_slot_table
;
145 if (task
->tk_status
>= 0)
148 switch (task
->tk_status
) {
149 /* MDS state errors */
150 case -NFS4ERR_DELEG_REVOKED
:
151 case -NFS4ERR_ADMIN_REVOKED
:
152 case -NFS4ERR_BAD_STATEID
:
155 nfs_remove_bad_delegation(state
->inode
);
156 case -NFS4ERR_OPENMODE
:
159 nfs4_schedule_stateid_recovery(mds_server
, state
);
160 goto wait_on_recovery
;
161 case -NFS4ERR_EXPIRED
:
163 nfs4_schedule_stateid_recovery(mds_server
, state
);
164 nfs4_schedule_lease_recovery(mds_client
);
165 goto wait_on_recovery
;
166 /* DS session errors */
167 case -NFS4ERR_BADSESSION
:
168 case -NFS4ERR_BADSLOT
:
169 case -NFS4ERR_BAD_HIGH_SLOT
:
170 case -NFS4ERR_DEADSESSION
:
171 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION
:
172 case -NFS4ERR_SEQ_FALSE_RETRY
:
173 case -NFS4ERR_SEQ_MISORDERED
:
174 dprintk("%s ERROR %d, Reset session. Exchangeid "
175 "flags 0x%x\n", __func__
, task
->tk_status
,
176 clp
->cl_exchange_flags
);
177 nfs4_schedule_session_recovery(clp
->cl_session
, task
->tk_status
);
182 rpc_delay(task
, FILELAYOUT_POLL_RETRY_MAX
);
184 case -NFS4ERR_RETRY_UNCACHED_REP
:
186 /* Invalidate Layout errors */
187 case -NFS4ERR_PNFS_NO_LAYOUT
:
188 case -ESTALE
: /* mapped NFS4ERR_STALE */
189 case -EBADHANDLE
: /* mapped NFS4ERR_BADHANDLE */
190 case -EISDIR
: /* mapped NFS4ERR_ISDIR */
191 case -NFS4ERR_FHEXPIRED
:
192 case -NFS4ERR_WRONG_TYPE
:
193 dprintk("%s Invalid layout error %d\n", __func__
,
196 * Destroy layout so new i/o will get a new layout.
197 * Layout will not be destroyed until all current lseg
198 * references are put. Mark layout as invalid to resend failed
199 * i/o and all i/o waiting on the slot table to the MDS until
200 * layout is destroyed and a new valid layout is obtained.
202 pnfs_destroy_layout(NFS_I(inode
));
203 rpc_wake_up(&tbl
->slot_tbl_waitq
);
205 /* RPC connection errors */
213 dprintk("%s DS connection error %d\n", __func__
,
215 nfs4_mark_deviceid_unavailable(devid
);
216 set_bit(NFS_LAYOUT_RETURN
, &lo
->plh_flags
);
217 rpc_wake_up(&tbl
->slot_tbl_waitq
);
221 dprintk("%s Retry through MDS. Error %d\n", __func__
,
223 return -NFS4ERR_RESET_TO_MDS
;
229 rpc_sleep_on(&mds_client
->cl_rpcwaitq
, task
, NULL
);
230 if (test_bit(NFS4CLNT_MANAGER_RUNNING
, &mds_client
->cl_state
) == 0)
231 rpc_wake_up_queued_task(&mds_client
->cl_rpcwaitq
, task
);
235 /* NFS_PROTO call done callback routines */
237 static int filelayout_read_done_cb(struct rpc_task
*task
,
238 struct nfs_read_data
*data
)
240 struct nfs_pgio_header
*hdr
= data
->header
;
243 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
244 data
->ds_clp
, hdr
->lseg
);
247 case -NFS4ERR_RESET_TO_MDS
:
248 filelayout_reset_read(data
);
249 return task
->tk_status
;
251 rpc_restart_call_prepare(task
);
259 * We reference the rpc_cred of the first WRITE that triggers the need for
260 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
261 * rfc5661 is not clear about which credential should be used.
264 filelayout_set_layoutcommit(struct nfs_write_data
*wdata
)
266 struct nfs_pgio_header
*hdr
= wdata
->header
;
268 if (FILELAYOUT_LSEG(hdr
->lseg
)->commit_through_mds
||
269 wdata
->res
.verf
->committed
== NFS_FILE_SYNC
)
272 pnfs_set_layoutcommit(wdata
);
273 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__
, hdr
->inode
->i_ino
,
274 (unsigned long) NFS_I(hdr
->inode
)->layout
->plh_lwb
);
278 filelayout_test_devid_unavailable(struct nfs4_deviceid_node
*node
)
280 return filelayout_test_devid_invalid(node
) ||
281 nfs4_test_deviceid_unavailable(node
);
285 filelayout_reset_to_mds(struct pnfs_layout_segment
*lseg
)
287 struct nfs4_deviceid_node
*node
= FILELAYOUT_DEVID_NODE(lseg
);
289 return filelayout_test_devid_unavailable(node
);
293 * Call ops for the async read/write cases
294 * In the case of dense layouts, the offset needs to be reset to its
297 static void filelayout_read_prepare(struct rpc_task
*task
, void *data
)
299 struct nfs_read_data
*rdata
= data
;
301 if (filelayout_reset_to_mds(rdata
->header
->lseg
)) {
302 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
303 filelayout_reset_read(rdata
);
307 rdata
->read_done_cb
= filelayout_read_done_cb
;
309 if (nfs41_setup_sequence(rdata
->ds_clp
->cl_session
,
310 &rdata
->args
.seq_args
, &rdata
->res
.seq_res
,
314 rpc_call_start(task
);
317 static void filelayout_read_call_done(struct rpc_task
*task
, void *data
)
319 struct nfs_read_data
*rdata
= data
;
321 dprintk("--> %s task->tk_status %d\n", __func__
, task
->tk_status
);
323 if (test_bit(NFS_IOHDR_REDO
, &rdata
->header
->flags
) &&
324 task
->tk_status
== 0)
327 /* Note this may cause RPC to be resent */
328 rdata
->header
->mds_ops
->rpc_call_done(task
, data
);
331 static void filelayout_read_count_stats(struct rpc_task
*task
, void *data
)
333 struct nfs_read_data
*rdata
= data
;
335 rpc_count_iostats(task
, NFS_SERVER(rdata
->header
->inode
)->client
->cl_metrics
);
338 static void filelayout_read_release(void *data
)
340 struct nfs_read_data
*rdata
= data
;
341 struct pnfs_layout_hdr
*lo
= rdata
->header
->lseg
->pls_layout
;
343 filelayout_fenceme(lo
->plh_inode
, lo
);
344 nfs_put_client(rdata
->ds_clp
);
345 rdata
->header
->mds_ops
->rpc_release(data
);
348 static int filelayout_write_done_cb(struct rpc_task
*task
,
349 struct nfs_write_data
*data
)
351 struct nfs_pgio_header
*hdr
= data
->header
;
354 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
355 data
->ds_clp
, hdr
->lseg
);
358 case -NFS4ERR_RESET_TO_MDS
:
359 filelayout_reset_write(data
);
360 return task
->tk_status
;
362 rpc_restart_call_prepare(task
);
366 filelayout_set_layoutcommit(data
);
370 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
371 static void prepare_to_resend_writes(struct nfs_commit_data
*data
)
373 struct nfs_page
*first
= nfs_list_entry(data
->pages
.next
);
375 data
->task
.tk_status
= 0;
376 memcpy(&data
->verf
.verifier
, &first
->wb_verf
,
377 sizeof(data
->verf
.verifier
));
378 data
->verf
.verifier
.data
[0]++; /* ensure verifier mismatch */
381 static int filelayout_commit_done_cb(struct rpc_task
*task
,
382 struct nfs_commit_data
*data
)
386 err
= filelayout_async_handle_error(task
, NULL
, data
->ds_clp
,
390 case -NFS4ERR_RESET_TO_MDS
:
391 prepare_to_resend_writes(data
);
394 rpc_restart_call_prepare(task
);
401 static void filelayout_write_prepare(struct rpc_task
*task
, void *data
)
403 struct nfs_write_data
*wdata
= data
;
405 if (filelayout_reset_to_mds(wdata
->header
->lseg
)) {
406 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
407 filelayout_reset_write(wdata
);
411 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
412 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
416 rpc_call_start(task
);
419 static void filelayout_write_call_done(struct rpc_task
*task
, void *data
)
421 struct nfs_write_data
*wdata
= data
;
423 if (test_bit(NFS_IOHDR_REDO
, &wdata
->header
->flags
) &&
424 task
->tk_status
== 0)
427 /* Note this may cause RPC to be resent */
428 wdata
->header
->mds_ops
->rpc_call_done(task
, data
);
431 static void filelayout_write_count_stats(struct rpc_task
*task
, void *data
)
433 struct nfs_write_data
*wdata
= data
;
435 rpc_count_iostats(task
, NFS_SERVER(wdata
->header
->inode
)->client
->cl_metrics
);
438 static void filelayout_write_release(void *data
)
440 struct nfs_write_data
*wdata
= data
;
441 struct pnfs_layout_hdr
*lo
= wdata
->header
->lseg
->pls_layout
;
443 filelayout_fenceme(lo
->plh_inode
, lo
);
444 nfs_put_client(wdata
->ds_clp
);
445 wdata
->header
->mds_ops
->rpc_release(data
);
448 static void filelayout_commit_prepare(struct rpc_task
*task
, void *data
)
450 struct nfs_commit_data
*wdata
= data
;
452 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
453 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
457 rpc_call_start(task
);
460 static void filelayout_write_commit_done(struct rpc_task
*task
, void *data
)
462 struct nfs_commit_data
*wdata
= data
;
464 /* Note this may cause RPC to be resent */
465 wdata
->mds_ops
->rpc_call_done(task
, data
);
468 static void filelayout_commit_count_stats(struct rpc_task
*task
, void *data
)
470 struct nfs_commit_data
*cdata
= data
;
472 rpc_count_iostats(task
, NFS_SERVER(cdata
->inode
)->client
->cl_metrics
);
475 static void filelayout_commit_release(void *calldata
)
477 struct nfs_commit_data
*data
= calldata
;
479 data
->completion_ops
->completion(data
);
480 pnfs_put_lseg(data
->lseg
);
481 nfs_put_client(data
->ds_clp
);
482 nfs_commitdata_release(data
);
485 static const struct rpc_call_ops filelayout_read_call_ops
= {
486 .rpc_call_prepare
= filelayout_read_prepare
,
487 .rpc_call_done
= filelayout_read_call_done
,
488 .rpc_count_stats
= filelayout_read_count_stats
,
489 .rpc_release
= filelayout_read_release
,
492 static const struct rpc_call_ops filelayout_write_call_ops
= {
493 .rpc_call_prepare
= filelayout_write_prepare
,
494 .rpc_call_done
= filelayout_write_call_done
,
495 .rpc_count_stats
= filelayout_write_count_stats
,
496 .rpc_release
= filelayout_write_release
,
499 static const struct rpc_call_ops filelayout_commit_call_ops
= {
500 .rpc_call_prepare
= filelayout_commit_prepare
,
501 .rpc_call_done
= filelayout_write_commit_done
,
502 .rpc_count_stats
= filelayout_commit_count_stats
,
503 .rpc_release
= filelayout_commit_release
,
506 static enum pnfs_try_status
507 filelayout_read_pagelist(struct nfs_read_data
*data
)
509 struct nfs_pgio_header
*hdr
= data
->header
;
510 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
511 struct nfs4_pnfs_ds
*ds
;
512 loff_t offset
= data
->args
.offset
;
517 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
518 __func__
, hdr
->inode
->i_ino
,
519 data
->args
.pgbase
, (size_t)data
->args
.count
, offset
);
521 /* Retrieve the correct rpc_client for the byte range */
522 j
= nfs4_fl_calc_j_index(lseg
, offset
);
523 idx
= nfs4_fl_calc_ds_index(lseg
, j
);
524 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
526 return PNFS_NOT_ATTEMPTED
;
527 dprintk("%s USE DS: %s cl_count %d\n", __func__
,
528 ds
->ds_remotestr
, atomic_read(&ds
->ds_clp
->cl_count
));
530 /* No multipath support. Use first DS */
531 atomic_inc(&ds
->ds_clp
->cl_count
);
532 data
->ds_clp
= ds
->ds_clp
;
533 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
537 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
538 data
->mds_offset
= offset
;
540 /* Perform an asynchronous read to ds */
541 status
= nfs_initiate_read(ds
->ds_clp
->cl_rpcclient
, data
,
542 &filelayout_read_call_ops
, RPC_TASK_SOFTCONN
);
544 return PNFS_ATTEMPTED
;
547 /* Perform async writes. */
548 static enum pnfs_try_status
549 filelayout_write_pagelist(struct nfs_write_data
*data
, int sync
)
551 struct nfs_pgio_header
*hdr
= data
->header
;
552 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
553 struct nfs4_pnfs_ds
*ds
;
554 loff_t offset
= data
->args
.offset
;
559 /* Retrieve the correct rpc_client for the byte range */
560 j
= nfs4_fl_calc_j_index(lseg
, offset
);
561 idx
= nfs4_fl_calc_ds_index(lseg
, j
);
562 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
564 return PNFS_NOT_ATTEMPTED
;
565 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
566 __func__
, hdr
->inode
->i_ino
, sync
, (size_t) data
->args
.count
,
567 offset
, ds
->ds_remotestr
, atomic_read(&ds
->ds_clp
->cl_count
));
569 data
->write_done_cb
= filelayout_write_done_cb
;
570 atomic_inc(&ds
->ds_clp
->cl_count
);
571 data
->ds_clp
= ds
->ds_clp
;
572 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
576 * Get the file offset on the dserver. Set the write offset to
577 * this offset and save the original offset.
579 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
581 /* Perform an asynchronous write */
582 status
= nfs_initiate_write(ds
->ds_clp
->cl_rpcclient
, data
,
583 &filelayout_write_call_ops
, sync
,
586 return PNFS_ATTEMPTED
;
590 * filelayout_check_layout()
592 * Make sure layout segment parameters are sane WRT the device.
593 * At this point no generic layer initialization of the lseg has occurred,
594 * and nothing has been added to the layout_hdr cache.
598 filelayout_check_layout(struct pnfs_layout_hdr
*lo
,
599 struct nfs4_filelayout_segment
*fl
,
600 struct nfs4_layoutget_res
*lgr
,
601 struct nfs4_deviceid
*id
,
604 struct nfs4_deviceid_node
*d
;
605 struct nfs4_file_layout_dsaddr
*dsaddr
;
606 int status
= -EINVAL
;
607 struct nfs_server
*nfss
= NFS_SERVER(lo
->plh_inode
);
609 dprintk("--> %s\n", __func__
);
611 /* FIXME: remove this check when layout segment support is added */
612 if (lgr
->range
.offset
!= 0 ||
613 lgr
->range
.length
!= NFS4_MAX_UINT64
) {
614 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
619 if (fl
->pattern_offset
> lgr
->range
.offset
) {
620 dprintk("%s pattern_offset %lld too large\n",
621 __func__
, fl
->pattern_offset
);
625 if (!fl
->stripe_unit
|| fl
->stripe_unit
% PAGE_SIZE
) {
626 dprintk("%s Invalid stripe unit (%u)\n",
627 __func__
, fl
->stripe_unit
);
631 /* find and reference the deviceid */
632 d
= nfs4_find_get_deviceid(NFS_SERVER(lo
->plh_inode
)->pnfs_curr_ld
,
633 NFS_SERVER(lo
->plh_inode
)->nfs_client
, id
);
635 dsaddr
= filelayout_get_device_info(lo
->plh_inode
, id
, gfp_flags
);
639 dsaddr
= container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
);
640 /* Found deviceid is unavailable */
641 if (filelayout_test_devid_unavailable(&dsaddr
->id_node
))
646 if (fl
->first_stripe_index
>= dsaddr
->stripe_count
) {
647 dprintk("%s Bad first_stripe_index %u\n",
648 __func__
, fl
->first_stripe_index
);
652 if ((fl
->stripe_type
== STRIPE_SPARSE
&&
653 fl
->num_fh
> 1 && fl
->num_fh
!= dsaddr
->ds_num
) ||
654 (fl
->stripe_type
== STRIPE_DENSE
&&
655 fl
->num_fh
!= dsaddr
->stripe_count
)) {
656 dprintk("%s num_fh %u not valid for given packing\n",
657 __func__
, fl
->num_fh
);
661 if (fl
->stripe_unit
% nfss
->rsize
|| fl
->stripe_unit
% nfss
->wsize
) {
662 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
663 "wsize %u\n", __func__
, fl
->stripe_unit
, nfss
->rsize
,
669 dprintk("--> %s returns %d\n", __func__
, status
);
672 nfs4_fl_put_deviceid(dsaddr
);
676 static void filelayout_free_fh_array(struct nfs4_filelayout_segment
*fl
)
680 for (i
= 0; i
< fl
->num_fh
; i
++) {
681 if (!fl
->fh_array
[i
])
683 kfree(fl
->fh_array
[i
]);
690 _filelayout_free_lseg(struct nfs4_filelayout_segment
*fl
)
692 filelayout_free_fh_array(fl
);
697 filelayout_decode_layout(struct pnfs_layout_hdr
*flo
,
698 struct nfs4_filelayout_segment
*fl
,
699 struct nfs4_layoutget_res
*lgr
,
700 struct nfs4_deviceid
*id
,
703 struct xdr_stream stream
;
705 struct page
*scratch
;
710 dprintk("%s: set_layout_map Begin\n", __func__
);
712 scratch
= alloc_page(gfp_flags
);
716 xdr_init_decode_pages(&stream
, &buf
, lgr
->layoutp
->pages
, lgr
->layoutp
->len
);
717 xdr_set_scratch_buffer(&stream
, page_address(scratch
), PAGE_SIZE
);
719 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
721 p
= xdr_inline_decode(&stream
, NFS4_DEVICEID4_SIZE
+ 20);
725 memcpy(id
, p
, sizeof(*id
));
726 p
+= XDR_QUADLEN(NFS4_DEVICEID4_SIZE
);
727 nfs4_print_deviceid(id
);
729 nfl_util
= be32_to_cpup(p
++);
730 if (nfl_util
& NFL4_UFLG_COMMIT_THRU_MDS
)
731 fl
->commit_through_mds
= 1;
732 if (nfl_util
& NFL4_UFLG_DENSE
)
733 fl
->stripe_type
= STRIPE_DENSE
;
735 fl
->stripe_type
= STRIPE_SPARSE
;
736 fl
->stripe_unit
= nfl_util
& ~NFL4_UFLG_MASK
;
738 fl
->first_stripe_index
= be32_to_cpup(p
++);
739 p
= xdr_decode_hyper(p
, &fl
->pattern_offset
);
740 fl
->num_fh
= be32_to_cpup(p
++);
742 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
743 __func__
, nfl_util
, fl
->num_fh
, fl
->first_stripe_index
,
746 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
747 * Futher checking is done in filelayout_check_layout */
749 max(NFS4_PNFS_MAX_STRIPE_CNT
, NFS4_PNFS_MAX_MULTI_CNT
))
752 if (fl
->num_fh
> 0) {
753 fl
->fh_array
= kcalloc(fl
->num_fh
, sizeof(fl
->fh_array
[0]),
759 for (i
= 0; i
< fl
->num_fh
; i
++) {
760 /* Do we want to use a mempool here? */
761 fl
->fh_array
[i
] = kmalloc(sizeof(struct nfs_fh
), gfp_flags
);
762 if (!fl
->fh_array
[i
])
765 p
= xdr_inline_decode(&stream
, 4);
768 fl
->fh_array
[i
]->size
= be32_to_cpup(p
++);
769 if (sizeof(struct nfs_fh
) < fl
->fh_array
[i
]->size
) {
770 printk(KERN_ERR
"NFS: Too big fh %d received %d\n",
771 i
, fl
->fh_array
[i
]->size
);
775 p
= xdr_inline_decode(&stream
, fl
->fh_array
[i
]->size
);
778 memcpy(fl
->fh_array
[i
]->data
, p
, fl
->fh_array
[i
]->size
);
779 dprintk("DEBUG: %s: fh len %d\n", __func__
,
780 fl
->fh_array
[i
]->size
);
783 __free_page(scratch
);
787 filelayout_free_fh_array(fl
);
789 __free_page(scratch
);
794 filelayout_free_lseg(struct pnfs_layout_segment
*lseg
)
796 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
798 dprintk("--> %s\n", __func__
);
799 nfs4_fl_put_deviceid(fl
->dsaddr
);
800 /* This assumes a single RW lseg */
801 if (lseg
->pls_range
.iomode
== IOMODE_RW
) {
802 struct nfs4_filelayout
*flo
;
804 flo
= FILELAYOUT_FROM_HDR(lseg
->pls_layout
);
805 flo
->commit_info
.nbuckets
= 0;
806 kfree(flo
->commit_info
.buckets
);
807 flo
->commit_info
.buckets
= NULL
;
809 _filelayout_free_lseg(fl
);
813 filelayout_alloc_commit_info(struct pnfs_layout_segment
*lseg
,
814 struct nfs_commit_info
*cinfo
,
817 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
818 struct pnfs_commit_bucket
*buckets
;
821 if (fl
->commit_through_mds
)
823 if (cinfo
->ds
->nbuckets
!= 0) {
824 /* This assumes there is only one IOMODE_RW lseg. What
825 * we really want to do is have a layout_hdr level
826 * dictionary of <multipath_list4, fh> keys, each
827 * associated with a struct list_head, populated by calls
828 * to filelayout_write_pagelist().
833 size
= (fl
->stripe_type
== STRIPE_SPARSE
) ?
834 fl
->dsaddr
->ds_num
: fl
->dsaddr
->stripe_count
;
836 buckets
= kcalloc(size
, sizeof(struct pnfs_commit_bucket
),
843 spin_lock(cinfo
->lock
);
844 if (cinfo
->ds
->nbuckets
!= 0)
847 cinfo
->ds
->buckets
= buckets
;
848 cinfo
->ds
->nbuckets
= size
;
849 for (i
= 0; i
< size
; i
++) {
850 INIT_LIST_HEAD(&buckets
[i
].written
);
851 INIT_LIST_HEAD(&buckets
[i
].committing
);
854 spin_unlock(cinfo
->lock
);
859 static struct pnfs_layout_segment
*
860 filelayout_alloc_lseg(struct pnfs_layout_hdr
*layoutid
,
861 struct nfs4_layoutget_res
*lgr
,
864 struct nfs4_filelayout_segment
*fl
;
866 struct nfs4_deviceid id
;
868 dprintk("--> %s\n", __func__
);
869 fl
= kzalloc(sizeof(*fl
), gfp_flags
);
873 rc
= filelayout_decode_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
);
874 if (rc
!= 0 || filelayout_check_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
)) {
875 _filelayout_free_lseg(fl
);
878 return &fl
->generic_hdr
;
882 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
884 * return true : coalesce page
885 * return false : don't coalesce page
888 filelayout_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
889 struct nfs_page
*req
)
891 u64 p_stripe
, r_stripe
;
894 if (!pnfs_generic_pg_test(pgio
, prev
, req
) ||
895 !nfs_generic_pg_test(pgio
, prev
, req
))
898 p_stripe
= (u64
)req_offset(prev
);
899 r_stripe
= (u64
)req_offset(req
);
900 stripe_unit
= FILELAYOUT_LSEG(pgio
->pg_lseg
)->stripe_unit
;
902 do_div(p_stripe
, stripe_unit
);
903 do_div(r_stripe
, stripe_unit
);
905 return (p_stripe
== r_stripe
);
909 filelayout_pg_init_read(struct nfs_pageio_descriptor
*pgio
,
910 struct nfs_page
*req
)
912 BUG_ON(pgio
->pg_lseg
!= NULL
);
914 if (req
->wb_offset
!= req
->wb_pgbase
) {
916 * Handling unaligned pages is difficult, because have to
917 * somehow split a req in two in certain cases in the
918 * pg.test code. Avoid this by just not using pnfs
921 nfs_pageio_reset_read_mds(pgio
);
924 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
930 /* If no lseg, fall back to read through mds */
931 if (pgio
->pg_lseg
== NULL
)
932 nfs_pageio_reset_read_mds(pgio
);
936 filelayout_pg_init_write(struct nfs_pageio_descriptor
*pgio
,
937 struct nfs_page
*req
)
939 struct nfs_commit_info cinfo
;
942 BUG_ON(pgio
->pg_lseg
!= NULL
);
944 if (req
->wb_offset
!= req
->wb_pgbase
)
946 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
952 /* If no lseg, fall back to write through mds */
953 if (pgio
->pg_lseg
== NULL
)
955 nfs_init_cinfo(&cinfo
, pgio
->pg_inode
, pgio
->pg_dreq
);
956 status
= filelayout_alloc_commit_info(pgio
->pg_lseg
, &cinfo
, GFP_NOFS
);
958 pnfs_put_lseg(pgio
->pg_lseg
);
959 pgio
->pg_lseg
= NULL
;
964 nfs_pageio_reset_write_mds(pgio
);
967 static const struct nfs_pageio_ops filelayout_pg_read_ops
= {
968 .pg_init
= filelayout_pg_init_read
,
969 .pg_test
= filelayout_pg_test
,
970 .pg_doio
= pnfs_generic_pg_readpages
,
973 static const struct nfs_pageio_ops filelayout_pg_write_ops
= {
974 .pg_init
= filelayout_pg_init_write
,
975 .pg_test
= filelayout_pg_test
,
976 .pg_doio
= pnfs_generic_pg_writepages
,
979 static u32
select_bucket_index(struct nfs4_filelayout_segment
*fl
, u32 j
)
981 if (fl
->stripe_type
== STRIPE_SPARSE
)
982 return nfs4_fl_calc_ds_index(&fl
->generic_hdr
, j
);
987 /* The generic layer is about to remove the req from the commit list.
988 * If this will make the bucket empty, it will need to put the lseg reference.
991 filelayout_clear_request_commit(struct nfs_page
*req
,
992 struct nfs_commit_info
*cinfo
)
994 struct pnfs_layout_segment
*freeme
= NULL
;
996 spin_lock(cinfo
->lock
);
997 if (!test_and_clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
))
999 cinfo
->ds
->nwritten
--;
1000 if (list_is_singular(&req
->wb_list
)) {
1001 struct pnfs_commit_bucket
*bucket
;
1003 bucket
= list_first_entry(&req
->wb_list
,
1004 struct pnfs_commit_bucket
,
1006 freeme
= bucket
->wlseg
;
1007 bucket
->wlseg
= NULL
;
1010 nfs_request_remove_commit_list(req
, cinfo
);
1011 spin_unlock(cinfo
->lock
);
1012 pnfs_put_lseg(freeme
);
1015 static struct list_head
*
1016 filelayout_choose_commit_list(struct nfs_page
*req
,
1017 struct pnfs_layout_segment
*lseg
,
1018 struct nfs_commit_info
*cinfo
)
1020 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
1022 struct list_head
*list
;
1023 struct pnfs_commit_bucket
*buckets
;
1025 if (fl
->commit_through_mds
)
1026 return &cinfo
->mds
->list
;
1028 /* Note that we are calling nfs4_fl_calc_j_index on each page
1029 * that ends up being committed to a data server. An attractive
1030 * alternative is to add a field to nfs_write_data and nfs_page
1031 * to store the value calculated in filelayout_write_pagelist
1032 * and just use that here.
1034 j
= nfs4_fl_calc_j_index(lseg
, req_offset(req
));
1035 i
= select_bucket_index(fl
, j
);
1036 buckets
= cinfo
->ds
->buckets
;
1037 list
= &buckets
[i
].written
;
1038 if (list_empty(list
)) {
1039 /* Non-empty buckets hold a reference on the lseg. That ref
1040 * is normally transferred to the COMMIT call and released
1041 * there. It could also be released if the last req is pulled
1042 * off due to a rewrite, in which case it will be done in
1043 * filelayout_clear_request_commit
1045 buckets
[i
].wlseg
= pnfs_get_lseg(lseg
);
1047 set_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1048 cinfo
->ds
->nwritten
++;
1053 filelayout_mark_request_commit(struct nfs_page
*req
,
1054 struct pnfs_layout_segment
*lseg
,
1055 struct nfs_commit_info
*cinfo
)
1057 struct list_head
*list
;
1059 list
= filelayout_choose_commit_list(req
, lseg
, cinfo
);
1060 nfs_request_add_commit_list(req
, list
, cinfo
);
1063 static u32
calc_ds_index_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1065 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1067 if (flseg
->stripe_type
== STRIPE_SPARSE
)
1070 return nfs4_fl_calc_ds_index(lseg
, i
);
1073 static struct nfs_fh
*
1074 select_ds_fh_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1076 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1078 if (flseg
->stripe_type
== STRIPE_SPARSE
) {
1079 if (flseg
->num_fh
== 1)
1081 else if (flseg
->num_fh
== 0)
1082 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1085 return flseg
->fh_array
[i
];
1088 static int filelayout_initiate_commit(struct nfs_commit_data
*data
, int how
)
1090 struct pnfs_layout_segment
*lseg
= data
->lseg
;
1091 struct nfs4_pnfs_ds
*ds
;
1095 idx
= calc_ds_index_from_commit(lseg
, data
->ds_commit_index
);
1096 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
1098 prepare_to_resend_writes(data
);
1099 filelayout_commit_release(data
);
1102 dprintk("%s ino %lu, how %d cl_count %d\n", __func__
,
1103 data
->inode
->i_ino
, how
, atomic_read(&ds
->ds_clp
->cl_count
));
1104 data
->commit_done_cb
= filelayout_commit_done_cb
;
1105 atomic_inc(&ds
->ds_clp
->cl_count
);
1106 data
->ds_clp
= ds
->ds_clp
;
1107 fh
= select_ds_fh_from_commit(lseg
, data
->ds_commit_index
);
1110 return nfs_initiate_commit(ds
->ds_clp
->cl_rpcclient
, data
,
1111 &filelayout_commit_call_ops
, how
,
1116 transfer_commit_list(struct list_head
*src
, struct list_head
*dst
,
1117 struct nfs_commit_info
*cinfo
, int max
)
1119 struct nfs_page
*req
, *tmp
;
1122 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
1123 if (!nfs_lock_request(req
))
1125 kref_get(&req
->wb_kref
);
1126 if (cond_resched_lock(cinfo
->lock
))
1127 list_safe_reset_next(req
, tmp
, wb_list
);
1128 nfs_request_remove_commit_list(req
, cinfo
);
1129 clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1130 nfs_list_add_request(req
, dst
);
1132 if ((ret
== max
) && !cinfo
->dreq
)
1139 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket
*bucket
,
1140 struct nfs_commit_info
*cinfo
,
1143 struct list_head
*src
= &bucket
->written
;
1144 struct list_head
*dst
= &bucket
->committing
;
1147 ret
= transfer_commit_list(src
, dst
, cinfo
, max
);
1149 cinfo
->ds
->nwritten
-= ret
;
1150 cinfo
->ds
->ncommitting
+= ret
;
1151 bucket
->clseg
= bucket
->wlseg
;
1152 if (list_empty(src
))
1153 bucket
->wlseg
= NULL
;
1155 pnfs_get_lseg(bucket
->clseg
);
1160 /* Move reqs from written to committing lists, returning count of number moved.
1161 * Note called with cinfo->lock held.
1163 static int filelayout_scan_commit_lists(struct nfs_commit_info
*cinfo
,
1168 for (i
= 0; i
< cinfo
->ds
->nbuckets
&& max
!= 0; i
++) {
1169 cnt
= filelayout_scan_ds_commit_list(&cinfo
->ds
->buckets
[i
],
1177 /* Pull everything off the committing lists and dump into @dst */
1178 static void filelayout_recover_commit_reqs(struct list_head
*dst
,
1179 struct nfs_commit_info
*cinfo
)
1181 struct pnfs_commit_bucket
*b
;
1184 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1185 * only called on single thread per dreq.
1186 * Can't take the lock because need to do pnfs_put_lseg
1188 for (i
= 0, b
= cinfo
->ds
->buckets
; i
< cinfo
->ds
->nbuckets
; i
++, b
++) {
1189 if (transfer_commit_list(&b
->written
, dst
, cinfo
, 0)) {
1190 BUG_ON(!list_empty(&b
->written
));
1191 pnfs_put_lseg(b
->wlseg
);
1195 cinfo
->ds
->nwritten
= 0;
1199 alloc_ds_commits(struct nfs_commit_info
*cinfo
, struct list_head
*list
)
1201 struct pnfs_ds_commit_info
*fl_cinfo
;
1202 struct pnfs_commit_bucket
*bucket
;
1203 struct nfs_commit_data
*data
;
1205 unsigned int nreq
= 0;
1207 fl_cinfo
= cinfo
->ds
;
1208 bucket
= fl_cinfo
->buckets
;
1209 for (i
= 0; i
< fl_cinfo
->nbuckets
; i
++, bucket
++) {
1210 if (list_empty(&bucket
->committing
))
1212 data
= nfs_commitdata_alloc();
1215 data
->ds_commit_index
= i
;
1216 data
->lseg
= bucket
->clseg
;
1217 bucket
->clseg
= NULL
;
1218 list_add(&data
->pages
, list
);
1222 /* Clean up on error */
1223 for (j
= i
; j
< fl_cinfo
->nbuckets
; j
++, bucket
++) {
1224 if (list_empty(&bucket
->committing
))
1226 nfs_retry_commit(&bucket
->committing
, bucket
->clseg
, cinfo
);
1227 pnfs_put_lseg(bucket
->clseg
);
1228 bucket
->clseg
= NULL
;
1230 /* Caller will clean up entries put on list */
1234 /* This follows nfs_commit_list pretty closely */
1236 filelayout_commit_pagelist(struct inode
*inode
, struct list_head
*mds_pages
,
1237 int how
, struct nfs_commit_info
*cinfo
)
1239 struct nfs_commit_data
*data
, *tmp
;
1241 unsigned int nreq
= 0;
1243 if (!list_empty(mds_pages
)) {
1244 data
= nfs_commitdata_alloc();
1247 list_add(&data
->pages
, &list
);
1250 nfs_retry_commit(mds_pages
, NULL
, cinfo
);
1253 nreq
+= alloc_ds_commits(cinfo
, &list
);
1256 cinfo
->completion_ops
->error_cleanup(NFS_I(inode
));
1260 atomic_add(nreq
, &cinfo
->mds
->rpcs_out
);
1262 list_for_each_entry_safe(data
, tmp
, &list
, pages
) {
1263 list_del_init(&data
->pages
);
1265 nfs_init_commit(data
, mds_pages
, NULL
, cinfo
);
1266 nfs_initiate_commit(NFS_CLIENT(inode
), data
,
1267 data
->mds_ops
, how
, 0);
1269 struct pnfs_commit_bucket
*buckets
;
1271 buckets
= cinfo
->ds
->buckets
;
1272 nfs_init_commit(data
, &buckets
[data
->ds_commit_index
].committing
, data
->lseg
, cinfo
);
1273 filelayout_initiate_commit(data
, how
);
1277 cinfo
->ds
->ncommitting
= 0;
1278 return PNFS_ATTEMPTED
;
1282 filelayout_free_deveiceid_node(struct nfs4_deviceid_node
*d
)
1284 nfs4_fl_free_deviceid(container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
));
1287 static struct pnfs_layout_hdr
*
1288 filelayout_alloc_layout_hdr(struct inode
*inode
, gfp_t gfp_flags
)
1290 struct nfs4_filelayout
*flo
;
1292 flo
= kzalloc(sizeof(*flo
), gfp_flags
);
1293 return &flo
->generic_hdr
;
1297 filelayout_free_layout_hdr(struct pnfs_layout_hdr
*lo
)
1299 kfree(FILELAYOUT_FROM_HDR(lo
));
1302 static struct pnfs_ds_commit_info
*
1303 filelayout_get_ds_info(struct inode
*inode
)
1305 struct pnfs_layout_hdr
*layout
= NFS_I(inode
)->layout
;
1310 return &FILELAYOUT_FROM_HDR(layout
)->commit_info
;
1313 static struct pnfs_layoutdriver_type filelayout_type
= {
1314 .id
= LAYOUT_NFSV4_1_FILES
,
1315 .name
= "LAYOUT_NFSV4_1_FILES",
1316 .owner
= THIS_MODULE
,
1317 .alloc_layout_hdr
= filelayout_alloc_layout_hdr
,
1318 .free_layout_hdr
= filelayout_free_layout_hdr
,
1319 .alloc_lseg
= filelayout_alloc_lseg
,
1320 .free_lseg
= filelayout_free_lseg
,
1321 .pg_read_ops
= &filelayout_pg_read_ops
,
1322 .pg_write_ops
= &filelayout_pg_write_ops
,
1323 .get_ds_info
= &filelayout_get_ds_info
,
1324 .mark_request_commit
= filelayout_mark_request_commit
,
1325 .clear_request_commit
= filelayout_clear_request_commit
,
1326 .scan_commit_lists
= filelayout_scan_commit_lists
,
1327 .recover_commit_reqs
= filelayout_recover_commit_reqs
,
1328 .commit_pagelist
= filelayout_commit_pagelist
,
1329 .read_pagelist
= filelayout_read_pagelist
,
1330 .write_pagelist
= filelayout_write_pagelist
,
1331 .free_deviceid_node
= filelayout_free_deveiceid_node
,
1334 static int __init
nfs4filelayout_init(void)
1336 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Registering...\n",
1338 return pnfs_register_layoutdriver(&filelayout_type
);
1341 static void __exit
nfs4filelayout_exit(void)
1343 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Unregistering...\n",
1345 pnfs_unregister_layoutdriver(&filelayout_type
);
1348 MODULE_ALIAS("nfs-layouttype4-1");
1350 module_init(nfs4filelayout_init
);
1351 module_exit(nfs4filelayout_exit
);