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 inode
*inode
= hdr
->inode
;
89 struct rpc_task
*task
= &data
->task
;
91 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
92 dprintk("%s Reset task %5u for i/o through MDS "
93 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__
,
96 (long long)NFS_FILEID(inode
),
98 (unsigned long long)data
->args
.offset
);
100 task
->tk_status
= pnfs_write_done_resend_to_mds(hdr
->inode
,
102 hdr
->completion_ops
);
104 /* balance nfs_get_client in filelayout_write_pagelist */
105 nfs_put_client(data
->ds_clp
);
109 static void filelayout_reset_read(struct nfs_read_data
*data
)
111 struct nfs_pgio_header
*hdr
= data
->header
;
112 struct inode
*inode
= hdr
->inode
;
113 struct rpc_task
*task
= &data
->task
;
115 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
116 dprintk("%s Reset task %5u for i/o through MDS "
117 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__
,
120 (long long)NFS_FILEID(inode
),
122 (unsigned long long)data
->args
.offset
);
124 task
->tk_status
= pnfs_read_done_resend_to_mds(hdr
->inode
,
126 hdr
->completion_ops
);
128 /* balance nfs_get_client in filelayout_read_pagelist */
129 nfs_put_client(data
->ds_clp
);
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 inode
*inode
= lseg
->pls_layout
->plh_inode
;
139 struct nfs_server
*mds_server
= NFS_SERVER(inode
);
140 struct nfs4_deviceid_node
*devid
= FILELAYOUT_DEVID_NODE(lseg
);
141 struct nfs_client
*mds_client
= mds_server
->nfs_client
;
142 struct nfs4_slot_table
*tbl
= &clp
->cl_session
->fc_slot_table
;
144 if (task
->tk_status
>= 0)
147 switch (task
->tk_status
) {
148 /* MDS state errors */
149 case -NFS4ERR_DELEG_REVOKED
:
150 case -NFS4ERR_ADMIN_REVOKED
:
151 case -NFS4ERR_BAD_STATEID
:
154 nfs_remove_bad_delegation(state
->inode
);
155 case -NFS4ERR_OPENMODE
:
158 nfs4_schedule_stateid_recovery(mds_server
, state
);
159 goto wait_on_recovery
;
160 case -NFS4ERR_EXPIRED
:
162 nfs4_schedule_stateid_recovery(mds_server
, state
);
163 nfs4_schedule_lease_recovery(mds_client
);
164 goto wait_on_recovery
;
165 /* DS session errors */
166 case -NFS4ERR_BADSESSION
:
167 case -NFS4ERR_BADSLOT
:
168 case -NFS4ERR_BAD_HIGH_SLOT
:
169 case -NFS4ERR_DEADSESSION
:
170 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION
:
171 case -NFS4ERR_SEQ_FALSE_RETRY
:
172 case -NFS4ERR_SEQ_MISORDERED
:
173 dprintk("%s ERROR %d, Reset session. Exchangeid "
174 "flags 0x%x\n", __func__
, task
->tk_status
,
175 clp
->cl_exchange_flags
);
176 nfs4_schedule_session_recovery(clp
->cl_session
);
181 rpc_delay(task
, FILELAYOUT_POLL_RETRY_MAX
);
183 case -NFS4ERR_RETRY_UNCACHED_REP
:
185 /* Invalidate Layout errors */
186 case -NFS4ERR_PNFS_NO_LAYOUT
:
187 case -ESTALE
: /* mapped NFS4ERR_STALE */
188 case -EBADHANDLE
: /* mapped NFS4ERR_BADHANDLE */
189 case -EISDIR
: /* mapped NFS4ERR_ISDIR */
190 case -NFS4ERR_FHEXPIRED
:
191 case -NFS4ERR_WRONG_TYPE
:
192 dprintk("%s Invalid layout error %d\n", __func__
,
195 * Destroy layout so new i/o will get a new layout.
196 * Layout will not be destroyed until all current lseg
197 * references are put. Mark layout as invalid to resend failed
198 * i/o and all i/o waiting on the slot table to the MDS until
199 * layout is destroyed and a new valid layout is obtained.
201 set_bit(NFS_LAYOUT_INVALID
,
202 &NFS_I(state
->inode
)->layout
->plh_flags
);
203 pnfs_destroy_layout(NFS_I(state
->inode
));
204 rpc_wake_up(&tbl
->slot_tbl_waitq
);
206 /* RPC connection errors */
214 dprintk("%s DS connection error %d\n", __func__
,
216 if (!filelayout_test_devid_invalid(devid
))
217 _pnfs_return_layout(state
->inode
);
218 filelayout_mark_devid_invalid(devid
);
219 rpc_wake_up(&tbl
->slot_tbl_waitq
);
220 nfs4_ds_disconnect(clp
);
224 dprintk("%s Retry through MDS. Error %d\n", __func__
,
226 return -NFS4ERR_RESET_TO_MDS
;
232 rpc_sleep_on(&mds_client
->cl_rpcwaitq
, task
, NULL
);
233 if (test_bit(NFS4CLNT_MANAGER_RUNNING
, &mds_client
->cl_state
) == 0)
234 rpc_wake_up_queued_task(&mds_client
->cl_rpcwaitq
, task
);
238 /* NFS_PROTO call done callback routines */
240 static int filelayout_read_done_cb(struct rpc_task
*task
,
241 struct nfs_read_data
*data
)
243 struct nfs_pgio_header
*hdr
= data
->header
;
246 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
247 data
->ds_clp
, hdr
->lseg
);
250 case -NFS4ERR_RESET_TO_MDS
:
251 filelayout_reset_read(data
);
252 return task
->tk_status
;
254 rpc_restart_call_prepare(task
);
262 * We reference the rpc_cred of the first WRITE that triggers the need for
263 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
264 * rfc5661 is not clear about which credential should be used.
267 filelayout_set_layoutcommit(struct nfs_write_data
*wdata
)
269 struct nfs_pgio_header
*hdr
= wdata
->header
;
271 if (FILELAYOUT_LSEG(hdr
->lseg
)->commit_through_mds
||
272 wdata
->res
.verf
->committed
== NFS_FILE_SYNC
)
275 pnfs_set_layoutcommit(wdata
);
276 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__
, hdr
->inode
->i_ino
,
277 (unsigned long) NFS_I(hdr
->inode
)->layout
->plh_lwb
);
281 * Call ops for the async read/write cases
282 * In the case of dense layouts, the offset needs to be reset to its
285 static void filelayout_read_prepare(struct rpc_task
*task
, void *data
)
287 struct nfs_read_data
*rdata
= data
;
289 if (filelayout_reset_to_mds(rdata
->header
->lseg
)) {
290 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
291 filelayout_reset_read(rdata
);
295 rdata
->read_done_cb
= filelayout_read_done_cb
;
297 if (nfs41_setup_sequence(rdata
->ds_clp
->cl_session
,
298 &rdata
->args
.seq_args
, &rdata
->res
.seq_res
,
302 rpc_call_start(task
);
305 static void filelayout_read_call_done(struct rpc_task
*task
, void *data
)
307 struct nfs_read_data
*rdata
= data
;
309 dprintk("--> %s task->tk_status %d\n", __func__
, task
->tk_status
);
311 if (test_bit(NFS_IOHDR_REDO
, &rdata
->header
->flags
))
314 /* Note this may cause RPC to be resent */
315 rdata
->header
->mds_ops
->rpc_call_done(task
, data
);
318 static void filelayout_read_count_stats(struct rpc_task
*task
, void *data
)
320 struct nfs_read_data
*rdata
= data
;
322 rpc_count_iostats(task
, NFS_SERVER(rdata
->header
->inode
)->client
->cl_metrics
);
325 static void filelayout_read_release(void *data
)
327 struct nfs_read_data
*rdata
= data
;
329 if (!test_bit(NFS_IOHDR_REDO
, &rdata
->header
->flags
))
330 nfs_put_client(rdata
->ds_clp
);
331 rdata
->header
->mds_ops
->rpc_release(data
);
334 static int filelayout_write_done_cb(struct rpc_task
*task
,
335 struct nfs_write_data
*data
)
337 struct nfs_pgio_header
*hdr
= data
->header
;
340 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
341 data
->ds_clp
, hdr
->lseg
);
344 case -NFS4ERR_RESET_TO_MDS
:
345 filelayout_reset_write(data
);
346 return task
->tk_status
;
348 rpc_restart_call_prepare(task
);
352 filelayout_set_layoutcommit(data
);
356 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
357 static void prepare_to_resend_writes(struct nfs_commit_data
*data
)
359 struct nfs_page
*first
= nfs_list_entry(data
->pages
.next
);
361 data
->task
.tk_status
= 0;
362 memcpy(data
->verf
.verifier
, first
->wb_verf
.verifier
,
363 sizeof(first
->wb_verf
.verifier
));
364 data
->verf
.verifier
[0]++; /* ensure verifier mismatch */
367 static int filelayout_commit_done_cb(struct rpc_task
*task
,
368 struct nfs_commit_data
*data
)
372 err
= filelayout_async_handle_error(task
, NULL
, data
->ds_clp
,
376 case -NFS4ERR_RESET_TO_MDS
:
377 prepare_to_resend_writes(data
);
380 rpc_restart_call_prepare(task
);
387 static void filelayout_write_prepare(struct rpc_task
*task
, void *data
)
389 struct nfs_write_data
*wdata
= data
;
391 if (filelayout_reset_to_mds(wdata
->header
->lseg
)) {
392 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
393 filelayout_reset_write(wdata
);
397 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
398 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
402 rpc_call_start(task
);
405 static void filelayout_write_call_done(struct rpc_task
*task
, void *data
)
407 struct nfs_write_data
*wdata
= data
;
409 if (test_bit(NFS_IOHDR_REDO
, &wdata
->header
->flags
))
412 /* Note this may cause RPC to be resent */
413 wdata
->header
->mds_ops
->rpc_call_done(task
, data
);
416 static void filelayout_write_count_stats(struct rpc_task
*task
, void *data
)
418 struct nfs_write_data
*wdata
= data
;
420 rpc_count_iostats(task
, NFS_SERVER(wdata
->header
->inode
)->client
->cl_metrics
);
423 static void filelayout_write_release(void *data
)
425 struct nfs_write_data
*wdata
= data
;
427 if (!test_bit(NFS_IOHDR_REDO
, &wdata
->header
->flags
))
428 nfs_put_client(wdata
->ds_clp
);
429 wdata
->header
->mds_ops
->rpc_release(data
);
432 static void filelayout_commit_prepare(struct rpc_task
*task
, void *data
)
434 struct nfs_commit_data
*wdata
= data
;
436 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
437 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
441 rpc_call_start(task
);
444 static void filelayout_write_commit_done(struct rpc_task
*task
, void *data
)
446 struct nfs_commit_data
*wdata
= data
;
448 /* Note this may cause RPC to be resent */
449 wdata
->mds_ops
->rpc_call_done(task
, data
);
452 static void filelayout_commit_count_stats(struct rpc_task
*task
, void *data
)
454 struct nfs_commit_data
*cdata
= data
;
456 rpc_count_iostats(task
, NFS_SERVER(cdata
->inode
)->client
->cl_metrics
);
459 static void filelayout_commit_release(void *calldata
)
461 struct nfs_commit_data
*data
= calldata
;
463 data
->completion_ops
->completion(data
);
464 put_lseg(data
->lseg
);
465 nfs_put_client(data
->ds_clp
);
466 nfs_commitdata_release(data
);
469 static const struct rpc_call_ops filelayout_read_call_ops
= {
470 .rpc_call_prepare
= filelayout_read_prepare
,
471 .rpc_call_done
= filelayout_read_call_done
,
472 .rpc_count_stats
= filelayout_read_count_stats
,
473 .rpc_release
= filelayout_read_release
,
476 static const struct rpc_call_ops filelayout_write_call_ops
= {
477 .rpc_call_prepare
= filelayout_write_prepare
,
478 .rpc_call_done
= filelayout_write_call_done
,
479 .rpc_count_stats
= filelayout_write_count_stats
,
480 .rpc_release
= filelayout_write_release
,
483 static const struct rpc_call_ops filelayout_commit_call_ops
= {
484 .rpc_call_prepare
= filelayout_commit_prepare
,
485 .rpc_call_done
= filelayout_write_commit_done
,
486 .rpc_count_stats
= filelayout_commit_count_stats
,
487 .rpc_release
= filelayout_commit_release
,
490 static enum pnfs_try_status
491 filelayout_read_pagelist(struct nfs_read_data
*data
)
493 struct nfs_pgio_header
*hdr
= data
->header
;
494 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
495 struct nfs4_pnfs_ds
*ds
;
496 loff_t offset
= data
->args
.offset
;
501 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
502 __func__
, hdr
->inode
->i_ino
,
503 data
->args
.pgbase
, (size_t)data
->args
.count
, offset
);
505 /* Retrieve the correct rpc_client for the byte range */
506 j
= nfs4_fl_calc_j_index(lseg
, offset
);
507 idx
= nfs4_fl_calc_ds_index(lseg
, j
);
508 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
510 return PNFS_NOT_ATTEMPTED
;
511 dprintk("%s USE DS: %s cl_count %d\n", __func__
,
512 ds
->ds_remotestr
, atomic_read(&ds
->ds_clp
->cl_count
));
514 /* No multipath support. Use first DS */
515 atomic_inc(&ds
->ds_clp
->cl_count
);
516 data
->ds_clp
= ds
->ds_clp
;
517 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
521 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
522 data
->mds_offset
= offset
;
524 /* Perform an asynchronous read to ds */
525 status
= nfs_initiate_read(ds
->ds_clp
->cl_rpcclient
, data
,
526 &filelayout_read_call_ops
, RPC_TASK_SOFTCONN
);
528 return PNFS_ATTEMPTED
;
531 /* Perform async writes. */
532 static enum pnfs_try_status
533 filelayout_write_pagelist(struct nfs_write_data
*data
, int sync
)
535 struct nfs_pgio_header
*hdr
= data
->header
;
536 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
537 struct nfs4_pnfs_ds
*ds
;
538 loff_t offset
= data
->args
.offset
;
543 /* Retrieve the correct rpc_client for the byte range */
544 j
= nfs4_fl_calc_j_index(lseg
, offset
);
545 idx
= nfs4_fl_calc_ds_index(lseg
, j
);
546 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
548 return PNFS_NOT_ATTEMPTED
;
549 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
550 __func__
, hdr
->inode
->i_ino
, sync
, (size_t) data
->args
.count
,
551 offset
, ds
->ds_remotestr
, atomic_read(&ds
->ds_clp
->cl_count
));
553 data
->write_done_cb
= filelayout_write_done_cb
;
554 atomic_inc(&ds
->ds_clp
->cl_count
);
555 data
->ds_clp
= ds
->ds_clp
;
556 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
560 * Get the file offset on the dserver. Set the write offset to
561 * this offset and save the original offset.
563 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
565 /* Perform an asynchronous write */
566 status
= nfs_initiate_write(ds
->ds_clp
->cl_rpcclient
, data
,
567 &filelayout_write_call_ops
, sync
,
570 return PNFS_ATTEMPTED
;
574 * filelayout_check_layout()
576 * Make sure layout segment parameters are sane WRT the device.
577 * At this point no generic layer initialization of the lseg has occurred,
578 * and nothing has been added to the layout_hdr cache.
582 filelayout_check_layout(struct pnfs_layout_hdr
*lo
,
583 struct nfs4_filelayout_segment
*fl
,
584 struct nfs4_layoutget_res
*lgr
,
585 struct nfs4_deviceid
*id
,
588 struct nfs4_deviceid_node
*d
;
589 struct nfs4_file_layout_dsaddr
*dsaddr
;
590 int status
= -EINVAL
;
591 struct nfs_server
*nfss
= NFS_SERVER(lo
->plh_inode
);
593 dprintk("--> %s\n", __func__
);
595 /* FIXME: remove this check when layout segment support is added */
596 if (lgr
->range
.offset
!= 0 ||
597 lgr
->range
.length
!= NFS4_MAX_UINT64
) {
598 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
603 if (fl
->pattern_offset
> lgr
->range
.offset
) {
604 dprintk("%s pattern_offset %lld too large\n",
605 __func__
, fl
->pattern_offset
);
609 if (!fl
->stripe_unit
|| fl
->stripe_unit
% PAGE_SIZE
) {
610 dprintk("%s Invalid stripe unit (%u)\n",
611 __func__
, fl
->stripe_unit
);
615 /* find and reference the deviceid */
616 d
= nfs4_find_get_deviceid(NFS_SERVER(lo
->plh_inode
)->pnfs_curr_ld
,
617 NFS_SERVER(lo
->plh_inode
)->nfs_client
, id
);
619 dsaddr
= get_device_info(lo
->plh_inode
, id
, gfp_flags
);
623 dsaddr
= container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
);
624 /* Found deviceid is being reaped */
625 if (test_bit(NFS_DEVICEID_INVALID
, &dsaddr
->id_node
.flags
))
630 if (fl
->first_stripe_index
>= dsaddr
->stripe_count
) {
631 dprintk("%s Bad first_stripe_index %u\n",
632 __func__
, fl
->first_stripe_index
);
636 if ((fl
->stripe_type
== STRIPE_SPARSE
&&
637 fl
->num_fh
> 1 && fl
->num_fh
!= dsaddr
->ds_num
) ||
638 (fl
->stripe_type
== STRIPE_DENSE
&&
639 fl
->num_fh
!= dsaddr
->stripe_count
)) {
640 dprintk("%s num_fh %u not valid for given packing\n",
641 __func__
, fl
->num_fh
);
645 if (fl
->stripe_unit
% nfss
->rsize
|| fl
->stripe_unit
% nfss
->wsize
) {
646 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
647 "wsize %u\n", __func__
, fl
->stripe_unit
, nfss
->rsize
,
653 dprintk("--> %s returns %d\n", __func__
, status
);
656 nfs4_fl_put_deviceid(dsaddr
);
660 static void filelayout_free_fh_array(struct nfs4_filelayout_segment
*fl
)
664 for (i
= 0; i
< fl
->num_fh
; i
++) {
665 if (!fl
->fh_array
[i
])
667 kfree(fl
->fh_array
[i
]);
674 _filelayout_free_lseg(struct nfs4_filelayout_segment
*fl
)
676 filelayout_free_fh_array(fl
);
681 filelayout_decode_layout(struct pnfs_layout_hdr
*flo
,
682 struct nfs4_filelayout_segment
*fl
,
683 struct nfs4_layoutget_res
*lgr
,
684 struct nfs4_deviceid
*id
,
687 struct xdr_stream stream
;
689 struct page
*scratch
;
694 dprintk("%s: set_layout_map Begin\n", __func__
);
696 scratch
= alloc_page(gfp_flags
);
700 xdr_init_decode_pages(&stream
, &buf
, lgr
->layoutp
->pages
, lgr
->layoutp
->len
);
701 xdr_set_scratch_buffer(&stream
, page_address(scratch
), PAGE_SIZE
);
703 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
705 p
= xdr_inline_decode(&stream
, NFS4_DEVICEID4_SIZE
+ 20);
709 memcpy(id
, p
, sizeof(*id
));
710 p
+= XDR_QUADLEN(NFS4_DEVICEID4_SIZE
);
711 nfs4_print_deviceid(id
);
713 nfl_util
= be32_to_cpup(p
++);
714 if (nfl_util
& NFL4_UFLG_COMMIT_THRU_MDS
)
715 fl
->commit_through_mds
= 1;
716 if (nfl_util
& NFL4_UFLG_DENSE
)
717 fl
->stripe_type
= STRIPE_DENSE
;
719 fl
->stripe_type
= STRIPE_SPARSE
;
720 fl
->stripe_unit
= nfl_util
& ~NFL4_UFLG_MASK
;
722 fl
->first_stripe_index
= be32_to_cpup(p
++);
723 p
= xdr_decode_hyper(p
, &fl
->pattern_offset
);
724 fl
->num_fh
= be32_to_cpup(p
++);
726 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
727 __func__
, nfl_util
, fl
->num_fh
, fl
->first_stripe_index
,
730 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
731 * Futher checking is done in filelayout_check_layout */
733 max(NFS4_PNFS_MAX_STRIPE_CNT
, NFS4_PNFS_MAX_MULTI_CNT
))
736 if (fl
->num_fh
> 0) {
737 fl
->fh_array
= kzalloc(fl
->num_fh
* sizeof(struct nfs_fh
*),
743 for (i
= 0; i
< fl
->num_fh
; i
++) {
744 /* Do we want to use a mempool here? */
745 fl
->fh_array
[i
] = kmalloc(sizeof(struct nfs_fh
), gfp_flags
);
746 if (!fl
->fh_array
[i
])
749 p
= xdr_inline_decode(&stream
, 4);
752 fl
->fh_array
[i
]->size
= be32_to_cpup(p
++);
753 if (sizeof(struct nfs_fh
) < fl
->fh_array
[i
]->size
) {
754 printk(KERN_ERR
"NFS: Too big fh %d received %d\n",
755 i
, fl
->fh_array
[i
]->size
);
759 p
= xdr_inline_decode(&stream
, fl
->fh_array
[i
]->size
);
762 memcpy(fl
->fh_array
[i
]->data
, p
, fl
->fh_array
[i
]->size
);
763 dprintk("DEBUG: %s: fh len %d\n", __func__
,
764 fl
->fh_array
[i
]->size
);
767 __free_page(scratch
);
771 filelayout_free_fh_array(fl
);
773 __free_page(scratch
);
778 filelayout_free_lseg(struct pnfs_layout_segment
*lseg
)
780 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
782 dprintk("--> %s\n", __func__
);
783 nfs4_fl_put_deviceid(fl
->dsaddr
);
784 /* This assumes a single RW lseg */
785 if (lseg
->pls_range
.iomode
== IOMODE_RW
) {
786 struct nfs4_filelayout
*flo
;
788 flo
= FILELAYOUT_FROM_HDR(lseg
->pls_layout
);
789 flo
->commit_info
.nbuckets
= 0;
790 kfree(flo
->commit_info
.buckets
);
791 flo
->commit_info
.buckets
= NULL
;
793 _filelayout_free_lseg(fl
);
797 filelayout_alloc_commit_info(struct pnfs_layout_segment
*lseg
,
798 struct nfs_commit_info
*cinfo
,
801 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
802 struct pnfs_commit_bucket
*buckets
;
805 if (fl
->commit_through_mds
)
807 if (cinfo
->ds
->nbuckets
!= 0) {
808 /* This assumes there is only one IOMODE_RW lseg. What
809 * we really want to do is have a layout_hdr level
810 * dictionary of <multipath_list4, fh> keys, each
811 * associated with a struct list_head, populated by calls
812 * to filelayout_write_pagelist().
817 size
= (fl
->stripe_type
== STRIPE_SPARSE
) ?
818 fl
->dsaddr
->ds_num
: fl
->dsaddr
->stripe_count
;
820 buckets
= kcalloc(size
, sizeof(struct pnfs_commit_bucket
),
827 spin_lock(cinfo
->lock
);
828 if (cinfo
->ds
->nbuckets
!= 0)
831 cinfo
->ds
->buckets
= buckets
;
832 cinfo
->ds
->nbuckets
= size
;
833 for (i
= 0; i
< size
; i
++) {
834 INIT_LIST_HEAD(&buckets
[i
].written
);
835 INIT_LIST_HEAD(&buckets
[i
].committing
);
838 spin_unlock(cinfo
->lock
);
843 static struct pnfs_layout_segment
*
844 filelayout_alloc_lseg(struct pnfs_layout_hdr
*layoutid
,
845 struct nfs4_layoutget_res
*lgr
,
848 struct nfs4_filelayout_segment
*fl
;
850 struct nfs4_deviceid id
;
852 dprintk("--> %s\n", __func__
);
853 fl
= kzalloc(sizeof(*fl
), gfp_flags
);
857 rc
= filelayout_decode_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
);
858 if (rc
!= 0 || filelayout_check_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
)) {
859 _filelayout_free_lseg(fl
);
862 return &fl
->generic_hdr
;
866 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
868 * return true : coalesce page
869 * return false : don't coalesce page
872 filelayout_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
873 struct nfs_page
*req
)
875 u64 p_stripe
, r_stripe
;
878 if (!pnfs_generic_pg_test(pgio
, prev
, req
) ||
879 !nfs_generic_pg_test(pgio
, prev
, req
))
882 p_stripe
= (u64
)req_offset(prev
);
883 r_stripe
= (u64
)req_offset(req
);
884 stripe_unit
= FILELAYOUT_LSEG(pgio
->pg_lseg
)->stripe_unit
;
886 do_div(p_stripe
, stripe_unit
);
887 do_div(r_stripe
, stripe_unit
);
889 return (p_stripe
== r_stripe
);
893 filelayout_pg_init_read(struct nfs_pageio_descriptor
*pgio
,
894 struct nfs_page
*req
)
896 BUG_ON(pgio
->pg_lseg
!= NULL
);
898 if (req
->wb_offset
!= req
->wb_pgbase
) {
900 * Handling unaligned pages is difficult, because have to
901 * somehow split a req in two in certain cases in the
902 * pg.test code. Avoid this by just not using pnfs
905 nfs_pageio_reset_read_mds(pgio
);
908 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
914 /* If no lseg, fall back to read through mds */
915 if (pgio
->pg_lseg
== NULL
)
916 nfs_pageio_reset_read_mds(pgio
);
920 filelayout_pg_init_write(struct nfs_pageio_descriptor
*pgio
,
921 struct nfs_page
*req
)
923 struct nfs_commit_info cinfo
;
926 BUG_ON(pgio
->pg_lseg
!= NULL
);
928 if (req
->wb_offset
!= req
->wb_pgbase
)
930 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
936 /* If no lseg, fall back to write through mds */
937 if (pgio
->pg_lseg
== NULL
)
939 nfs_init_cinfo(&cinfo
, pgio
->pg_inode
, pgio
->pg_dreq
);
940 status
= filelayout_alloc_commit_info(pgio
->pg_lseg
, &cinfo
, GFP_NOFS
);
942 put_lseg(pgio
->pg_lseg
);
943 pgio
->pg_lseg
= NULL
;
948 nfs_pageio_reset_write_mds(pgio
);
951 static const struct nfs_pageio_ops filelayout_pg_read_ops
= {
952 .pg_init
= filelayout_pg_init_read
,
953 .pg_test
= filelayout_pg_test
,
954 .pg_doio
= pnfs_generic_pg_readpages
,
957 static const struct nfs_pageio_ops filelayout_pg_write_ops
= {
958 .pg_init
= filelayout_pg_init_write
,
959 .pg_test
= filelayout_pg_test
,
960 .pg_doio
= pnfs_generic_pg_writepages
,
963 static u32
select_bucket_index(struct nfs4_filelayout_segment
*fl
, u32 j
)
965 if (fl
->stripe_type
== STRIPE_SPARSE
)
966 return nfs4_fl_calc_ds_index(&fl
->generic_hdr
, j
);
971 /* The generic layer is about to remove the req from the commit list.
972 * If this will make the bucket empty, it will need to put the lseg reference.
975 filelayout_clear_request_commit(struct nfs_page
*req
,
976 struct nfs_commit_info
*cinfo
)
978 struct pnfs_layout_segment
*freeme
= NULL
;
980 spin_lock(cinfo
->lock
);
981 if (!test_and_clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
))
983 cinfo
->ds
->nwritten
--;
984 if (list_is_singular(&req
->wb_list
)) {
985 struct pnfs_commit_bucket
*bucket
;
987 bucket
= list_first_entry(&req
->wb_list
,
988 struct pnfs_commit_bucket
,
990 freeme
= bucket
->wlseg
;
991 bucket
->wlseg
= NULL
;
994 nfs_request_remove_commit_list(req
, cinfo
);
995 spin_unlock(cinfo
->lock
);
999 static struct list_head
*
1000 filelayout_choose_commit_list(struct nfs_page
*req
,
1001 struct pnfs_layout_segment
*lseg
,
1002 struct nfs_commit_info
*cinfo
)
1004 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
1006 struct list_head
*list
;
1007 struct pnfs_commit_bucket
*buckets
;
1009 if (fl
->commit_through_mds
)
1010 return &cinfo
->mds
->list
;
1012 /* Note that we are calling nfs4_fl_calc_j_index on each page
1013 * that ends up being committed to a data server. An attractive
1014 * alternative is to add a field to nfs_write_data and nfs_page
1015 * to store the value calculated in filelayout_write_pagelist
1016 * and just use that here.
1018 j
= nfs4_fl_calc_j_index(lseg
, req_offset(req
));
1019 i
= select_bucket_index(fl
, j
);
1020 buckets
= cinfo
->ds
->buckets
;
1021 list
= &buckets
[i
].written
;
1022 if (list_empty(list
)) {
1023 /* Non-empty buckets hold a reference on the lseg. That ref
1024 * is normally transferred to the COMMIT call and released
1025 * there. It could also be released if the last req is pulled
1026 * off due to a rewrite, in which case it will be done in
1027 * filelayout_clear_request_commit
1029 buckets
[i
].wlseg
= get_lseg(lseg
);
1031 set_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1032 cinfo
->ds
->nwritten
++;
1037 filelayout_mark_request_commit(struct nfs_page
*req
,
1038 struct pnfs_layout_segment
*lseg
,
1039 struct nfs_commit_info
*cinfo
)
1041 struct list_head
*list
;
1043 list
= filelayout_choose_commit_list(req
, lseg
, cinfo
);
1044 nfs_request_add_commit_list(req
, list
, cinfo
);
1047 static u32
calc_ds_index_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1049 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1051 if (flseg
->stripe_type
== STRIPE_SPARSE
)
1054 return nfs4_fl_calc_ds_index(lseg
, i
);
1057 static struct nfs_fh
*
1058 select_ds_fh_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1060 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1062 if (flseg
->stripe_type
== STRIPE_SPARSE
) {
1063 if (flseg
->num_fh
== 1)
1065 else if (flseg
->num_fh
== 0)
1066 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1069 return flseg
->fh_array
[i
];
1072 static int filelayout_initiate_commit(struct nfs_commit_data
*data
, int how
)
1074 struct pnfs_layout_segment
*lseg
= data
->lseg
;
1075 struct nfs4_pnfs_ds
*ds
;
1079 idx
= calc_ds_index_from_commit(lseg
, data
->ds_commit_index
);
1080 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
1082 prepare_to_resend_writes(data
);
1083 filelayout_commit_release(data
);
1086 dprintk("%s ino %lu, how %d cl_count %d\n", __func__
,
1087 data
->inode
->i_ino
, how
, atomic_read(&ds
->ds_clp
->cl_count
));
1088 data
->commit_done_cb
= filelayout_commit_done_cb
;
1089 atomic_inc(&ds
->ds_clp
->cl_count
);
1090 data
->ds_clp
= ds
->ds_clp
;
1091 fh
= select_ds_fh_from_commit(lseg
, data
->ds_commit_index
);
1094 return nfs_initiate_commit(ds
->ds_clp
->cl_rpcclient
, data
,
1095 &filelayout_commit_call_ops
, how
,
1100 transfer_commit_list(struct list_head
*src
, struct list_head
*dst
,
1101 struct nfs_commit_info
*cinfo
, int max
)
1103 struct nfs_page
*req
, *tmp
;
1106 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
1107 if (!nfs_lock_request(req
))
1109 kref_get(&req
->wb_kref
);
1110 if (cond_resched_lock(cinfo
->lock
))
1111 list_safe_reset_next(req
, tmp
, wb_list
);
1112 nfs_request_remove_commit_list(req
, cinfo
);
1113 clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1114 nfs_list_add_request(req
, dst
);
1116 if ((ret
== max
) && !cinfo
->dreq
)
1123 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket
*bucket
,
1124 struct nfs_commit_info
*cinfo
,
1127 struct list_head
*src
= &bucket
->written
;
1128 struct list_head
*dst
= &bucket
->committing
;
1131 ret
= transfer_commit_list(src
, dst
, cinfo
, max
);
1133 cinfo
->ds
->nwritten
-= ret
;
1134 cinfo
->ds
->ncommitting
+= ret
;
1135 bucket
->clseg
= bucket
->wlseg
;
1136 if (list_empty(src
))
1137 bucket
->wlseg
= NULL
;
1139 get_lseg(bucket
->clseg
);
1144 /* Move reqs from written to committing lists, returning count of number moved.
1145 * Note called with cinfo->lock held.
1147 static int filelayout_scan_commit_lists(struct nfs_commit_info
*cinfo
,
1152 for (i
= 0; i
< cinfo
->ds
->nbuckets
&& max
!= 0; i
++) {
1153 cnt
= filelayout_scan_ds_commit_list(&cinfo
->ds
->buckets
[i
],
1161 /* Pull everything off the committing lists and dump into @dst */
1162 static void filelayout_recover_commit_reqs(struct list_head
*dst
,
1163 struct nfs_commit_info
*cinfo
)
1165 struct pnfs_commit_bucket
*b
;
1168 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1169 * only called on single thread per dreq.
1170 * Can't take the lock because need to do put_lseg
1172 for (i
= 0, b
= cinfo
->ds
->buckets
; i
< cinfo
->ds
->nbuckets
; i
++, b
++) {
1173 if (transfer_commit_list(&b
->written
, dst
, cinfo
, 0)) {
1174 BUG_ON(!list_empty(&b
->written
));
1179 cinfo
->ds
->nwritten
= 0;
1183 alloc_ds_commits(struct nfs_commit_info
*cinfo
, struct list_head
*list
)
1185 struct pnfs_ds_commit_info
*fl_cinfo
;
1186 struct pnfs_commit_bucket
*bucket
;
1187 struct nfs_commit_data
*data
;
1189 unsigned int nreq
= 0;
1191 fl_cinfo
= cinfo
->ds
;
1192 bucket
= fl_cinfo
->buckets
;
1193 for (i
= 0; i
< fl_cinfo
->nbuckets
; i
++, bucket
++) {
1194 if (list_empty(&bucket
->committing
))
1196 data
= nfs_commitdata_alloc();
1199 data
->ds_commit_index
= i
;
1200 data
->lseg
= bucket
->clseg
;
1201 bucket
->clseg
= NULL
;
1202 list_add(&data
->pages
, list
);
1206 /* Clean up on error */
1207 for (j
= i
; j
< fl_cinfo
->nbuckets
; j
++, bucket
++) {
1208 if (list_empty(&bucket
->committing
))
1210 nfs_retry_commit(&bucket
->committing
, bucket
->clseg
, cinfo
);
1211 put_lseg(bucket
->clseg
);
1212 bucket
->clseg
= NULL
;
1214 /* Caller will clean up entries put on list */
1218 /* This follows nfs_commit_list pretty closely */
1220 filelayout_commit_pagelist(struct inode
*inode
, struct list_head
*mds_pages
,
1221 int how
, struct nfs_commit_info
*cinfo
)
1223 struct nfs_commit_data
*data
, *tmp
;
1225 unsigned int nreq
= 0;
1227 if (!list_empty(mds_pages
)) {
1228 data
= nfs_commitdata_alloc();
1231 list_add(&data
->pages
, &list
);
1234 nfs_retry_commit(mds_pages
, NULL
, cinfo
);
1237 nreq
+= alloc_ds_commits(cinfo
, &list
);
1240 cinfo
->completion_ops
->error_cleanup(NFS_I(inode
));
1244 atomic_add(nreq
, &cinfo
->mds
->rpcs_out
);
1246 list_for_each_entry_safe(data
, tmp
, &list
, pages
) {
1247 list_del_init(&data
->pages
);
1249 nfs_init_commit(data
, mds_pages
, NULL
, cinfo
);
1250 nfs_initiate_commit(NFS_CLIENT(inode
), data
,
1251 data
->mds_ops
, how
, 0);
1253 struct pnfs_commit_bucket
*buckets
;
1255 buckets
= cinfo
->ds
->buckets
;
1256 nfs_init_commit(data
, &buckets
[data
->ds_commit_index
].committing
, data
->lseg
, cinfo
);
1257 filelayout_initiate_commit(data
, how
);
1261 cinfo
->ds
->ncommitting
= 0;
1262 return PNFS_ATTEMPTED
;
1266 filelayout_free_deveiceid_node(struct nfs4_deviceid_node
*d
)
1268 nfs4_fl_free_deviceid(container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
));
1271 static struct pnfs_layout_hdr
*
1272 filelayout_alloc_layout_hdr(struct inode
*inode
, gfp_t gfp_flags
)
1274 struct nfs4_filelayout
*flo
;
1276 flo
= kzalloc(sizeof(*flo
), gfp_flags
);
1277 return &flo
->generic_hdr
;
1281 filelayout_free_layout_hdr(struct pnfs_layout_hdr
*lo
)
1283 kfree(FILELAYOUT_FROM_HDR(lo
));
1286 static struct pnfs_ds_commit_info
*
1287 filelayout_get_ds_info(struct inode
*inode
)
1289 struct pnfs_layout_hdr
*layout
= NFS_I(inode
)->layout
;
1294 return &FILELAYOUT_FROM_HDR(layout
)->commit_info
;
1297 static struct pnfs_layoutdriver_type filelayout_type
= {
1298 .id
= LAYOUT_NFSV4_1_FILES
,
1299 .name
= "LAYOUT_NFSV4_1_FILES",
1300 .owner
= THIS_MODULE
,
1301 .alloc_layout_hdr
= filelayout_alloc_layout_hdr
,
1302 .free_layout_hdr
= filelayout_free_layout_hdr
,
1303 .alloc_lseg
= filelayout_alloc_lseg
,
1304 .free_lseg
= filelayout_free_lseg
,
1305 .pg_read_ops
= &filelayout_pg_read_ops
,
1306 .pg_write_ops
= &filelayout_pg_write_ops
,
1307 .get_ds_info
= &filelayout_get_ds_info
,
1308 .mark_request_commit
= filelayout_mark_request_commit
,
1309 .clear_request_commit
= filelayout_clear_request_commit
,
1310 .scan_commit_lists
= filelayout_scan_commit_lists
,
1311 .recover_commit_reqs
= filelayout_recover_commit_reqs
,
1312 .commit_pagelist
= filelayout_commit_pagelist
,
1313 .read_pagelist
= filelayout_read_pagelist
,
1314 .write_pagelist
= filelayout_write_pagelist
,
1315 .free_deviceid_node
= filelayout_free_deveiceid_node
,
1318 static int __init
nfs4filelayout_init(void)
1320 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Registering...\n",
1322 return pnfs_register_layoutdriver(&filelayout_type
);
1325 static void __exit
nfs4filelayout_exit(void)
1327 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Unregistering...\n",
1329 pnfs_unregister_layoutdriver(&filelayout_type
);
1332 MODULE_ALIAS("nfs-layouttype4-1");
1334 module_init(nfs4filelayout_init
);
1335 module_exit(nfs4filelayout_exit
);