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
);
106 static void filelayout_reset_read(struct nfs_read_data
*data
)
108 struct nfs_pgio_header
*hdr
= data
->header
;
109 struct inode
*inode
= hdr
->inode
;
110 struct rpc_task
*task
= &data
->task
;
112 if (!test_and_set_bit(NFS_IOHDR_REDO
, &hdr
->flags
)) {
113 dprintk("%s Reset task %5u for i/o through MDS "
114 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__
,
117 (long long)NFS_FILEID(inode
),
119 (unsigned long long)data
->args
.offset
);
121 task
->tk_status
= pnfs_read_done_resend_to_mds(hdr
->inode
,
123 hdr
->completion_ops
);
127 static int filelayout_async_handle_error(struct rpc_task
*task
,
128 struct nfs4_state
*state
,
129 struct nfs_client
*clp
,
130 struct pnfs_layout_segment
*lseg
)
132 struct inode
*inode
= lseg
->pls_layout
->plh_inode
;
133 struct nfs_server
*mds_server
= NFS_SERVER(inode
);
134 struct nfs4_deviceid_node
*devid
= FILELAYOUT_DEVID_NODE(lseg
);
135 struct nfs_client
*mds_client
= mds_server
->nfs_client
;
137 if (task
->tk_status
>= 0)
140 switch (task
->tk_status
) {
141 /* MDS state errors */
142 case -NFS4ERR_DELEG_REVOKED
:
143 case -NFS4ERR_ADMIN_REVOKED
:
144 case -NFS4ERR_BAD_STATEID
:
147 nfs_remove_bad_delegation(state
->inode
);
148 case -NFS4ERR_OPENMODE
:
151 nfs4_schedule_stateid_recovery(mds_server
, state
);
152 goto wait_on_recovery
;
153 case -NFS4ERR_EXPIRED
:
155 nfs4_schedule_stateid_recovery(mds_server
, state
);
156 nfs4_schedule_lease_recovery(mds_client
);
157 goto wait_on_recovery
;
158 /* DS session errors */
159 case -NFS4ERR_BADSESSION
:
160 case -NFS4ERR_BADSLOT
:
161 case -NFS4ERR_BAD_HIGH_SLOT
:
162 case -NFS4ERR_DEADSESSION
:
163 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION
:
164 case -NFS4ERR_SEQ_FALSE_RETRY
:
165 case -NFS4ERR_SEQ_MISORDERED
:
166 dprintk("%s ERROR %d, Reset session. Exchangeid "
167 "flags 0x%x\n", __func__
, task
->tk_status
,
168 clp
->cl_exchange_flags
);
169 nfs4_schedule_session_recovery(clp
->cl_session
);
174 rpc_delay(task
, FILELAYOUT_POLL_RETRY_MAX
);
176 case -NFS4ERR_RETRY_UNCACHED_REP
:
178 /* RPC connection errors */
186 dprintk("%s DS connection error %d\n", __func__
,
188 filelayout_mark_devid_invalid(devid
);
191 dprintk("%s Retry through MDS. Error %d\n", __func__
,
193 return -NFS4ERR_RESET_TO_MDS
;
199 rpc_sleep_on(&mds_client
->cl_rpcwaitq
, task
, NULL
);
200 if (test_bit(NFS4CLNT_MANAGER_RUNNING
, &mds_client
->cl_state
) == 0)
201 rpc_wake_up_queued_task(&mds_client
->cl_rpcwaitq
, task
);
205 /* NFS_PROTO call done callback routines */
207 static int filelayout_read_done_cb(struct rpc_task
*task
,
208 struct nfs_read_data
*data
)
210 struct nfs_pgio_header
*hdr
= data
->header
;
213 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
214 data
->ds_clp
, hdr
->lseg
);
217 case -NFS4ERR_RESET_TO_MDS
:
218 filelayout_reset_read(data
);
219 return task
->tk_status
;
221 rpc_restart_call_prepare(task
);
229 * We reference the rpc_cred of the first WRITE that triggers the need for
230 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
231 * rfc5661 is not clear about which credential should be used.
234 filelayout_set_layoutcommit(struct nfs_write_data
*wdata
)
236 struct nfs_pgio_header
*hdr
= wdata
->header
;
238 if (FILELAYOUT_LSEG(hdr
->lseg
)->commit_through_mds
||
239 wdata
->res
.verf
->committed
== NFS_FILE_SYNC
)
242 pnfs_set_layoutcommit(wdata
);
243 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__
, hdr
->inode
->i_ino
,
244 (unsigned long) NFS_I(hdr
->inode
)->layout
->plh_lwb
);
248 * Call ops for the async read/write cases
249 * In the case of dense layouts, the offset needs to be reset to its
252 static void filelayout_read_prepare(struct rpc_task
*task
, void *data
)
254 struct nfs_read_data
*rdata
= data
;
255 struct pnfs_layout_segment
*lseg
= rdata
->header
->lseg
;
257 if (filelayout_test_devid_invalid(FILELAYOUT_DEVID_NODE(lseg
))) {
258 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
259 filelayout_reset_read(rdata
);
263 rdata
->read_done_cb
= filelayout_read_done_cb
;
265 if (nfs41_setup_sequence(rdata
->ds_clp
->cl_session
,
266 &rdata
->args
.seq_args
, &rdata
->res
.seq_res
,
270 rpc_call_start(task
);
273 static void filelayout_read_call_done(struct rpc_task
*task
, void *data
)
275 struct nfs_read_data
*rdata
= data
;
277 dprintk("--> %s task->tk_status %d\n", __func__
, task
->tk_status
);
279 if (test_bit(NFS_IOHDR_REDO
, &rdata
->header
->flags
))
282 /* Note this may cause RPC to be resent */
283 rdata
->header
->mds_ops
->rpc_call_done(task
, data
);
286 static void filelayout_read_count_stats(struct rpc_task
*task
, void *data
)
288 struct nfs_read_data
*rdata
= data
;
290 rpc_count_iostats(task
, NFS_SERVER(rdata
->header
->inode
)->client
->cl_metrics
);
293 static void filelayout_read_release(void *data
)
295 struct nfs_read_data
*rdata
= data
;
297 rdata
->header
->mds_ops
->rpc_release(data
);
300 static int filelayout_write_done_cb(struct rpc_task
*task
,
301 struct nfs_write_data
*data
)
303 struct nfs_pgio_header
*hdr
= data
->header
;
306 err
= filelayout_async_handle_error(task
, data
->args
.context
->state
,
307 data
->ds_clp
, hdr
->lseg
);
310 case -NFS4ERR_RESET_TO_MDS
:
311 filelayout_reset_write(data
);
312 return task
->tk_status
;
314 rpc_restart_call_prepare(task
);
318 filelayout_set_layoutcommit(data
);
322 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
323 static void prepare_to_resend_writes(struct nfs_commit_data
*data
)
325 struct nfs_page
*first
= nfs_list_entry(data
->pages
.next
);
327 data
->task
.tk_status
= 0;
328 memcpy(data
->verf
.verifier
, first
->wb_verf
.verifier
,
329 sizeof(first
->wb_verf
.verifier
));
330 data
->verf
.verifier
[0]++; /* ensure verifier mismatch */
333 static int filelayout_commit_done_cb(struct rpc_task
*task
,
334 struct nfs_commit_data
*data
)
338 err
= filelayout_async_handle_error(task
, NULL
, data
->ds_clp
,
342 case -NFS4ERR_RESET_TO_MDS
:
343 prepare_to_resend_writes(data
);
346 rpc_restart_call_prepare(task
);
353 static void filelayout_write_prepare(struct rpc_task
*task
, void *data
)
355 struct nfs_write_data
*wdata
= data
;
356 struct pnfs_layout_segment
*lseg
= wdata
->header
->lseg
;
358 if (filelayout_test_devid_invalid(FILELAYOUT_DEVID_NODE(lseg
))) {
359 dprintk("%s task %u reset io to MDS\n", __func__
, task
->tk_pid
);
360 filelayout_reset_write(wdata
);
364 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
365 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
369 rpc_call_start(task
);
372 static void filelayout_write_call_done(struct rpc_task
*task
, void *data
)
374 struct nfs_write_data
*wdata
= data
;
376 if (test_bit(NFS_IOHDR_REDO
, &wdata
->header
->flags
))
379 /* Note this may cause RPC to be resent */
380 wdata
->header
->mds_ops
->rpc_call_done(task
, data
);
383 static void filelayout_write_count_stats(struct rpc_task
*task
, void *data
)
385 struct nfs_write_data
*wdata
= data
;
387 rpc_count_iostats(task
, NFS_SERVER(wdata
->header
->inode
)->client
->cl_metrics
);
390 static void filelayout_write_release(void *data
)
392 struct nfs_write_data
*wdata
= data
;
394 wdata
->header
->mds_ops
->rpc_release(data
);
397 static void filelayout_commit_prepare(struct rpc_task
*task
, void *data
)
399 struct nfs_commit_data
*wdata
= data
;
401 if (nfs41_setup_sequence(wdata
->ds_clp
->cl_session
,
402 &wdata
->args
.seq_args
, &wdata
->res
.seq_res
,
406 rpc_call_start(task
);
409 static void filelayout_write_commit_done(struct rpc_task
*task
, void *data
)
411 struct nfs_commit_data
*wdata
= data
;
413 /* Note this may cause RPC to be resent */
414 wdata
->mds_ops
->rpc_call_done(task
, data
);
417 static void filelayout_commit_count_stats(struct rpc_task
*task
, void *data
)
419 struct nfs_commit_data
*cdata
= data
;
421 rpc_count_iostats(task
, NFS_SERVER(cdata
->inode
)->client
->cl_metrics
);
424 static void filelayout_commit_release(void *calldata
)
426 struct nfs_commit_data
*data
= calldata
;
428 data
->completion_ops
->completion(data
);
429 put_lseg(data
->lseg
);
430 nfs_commitdata_release(data
);
433 static const struct rpc_call_ops filelayout_read_call_ops
= {
434 .rpc_call_prepare
= filelayout_read_prepare
,
435 .rpc_call_done
= filelayout_read_call_done
,
436 .rpc_count_stats
= filelayout_read_count_stats
,
437 .rpc_release
= filelayout_read_release
,
440 static const struct rpc_call_ops filelayout_write_call_ops
= {
441 .rpc_call_prepare
= filelayout_write_prepare
,
442 .rpc_call_done
= filelayout_write_call_done
,
443 .rpc_count_stats
= filelayout_write_count_stats
,
444 .rpc_release
= filelayout_write_release
,
447 static const struct rpc_call_ops filelayout_commit_call_ops
= {
448 .rpc_call_prepare
= filelayout_commit_prepare
,
449 .rpc_call_done
= filelayout_write_commit_done
,
450 .rpc_count_stats
= filelayout_commit_count_stats
,
451 .rpc_release
= filelayout_commit_release
,
454 static enum pnfs_try_status
455 filelayout_read_pagelist(struct nfs_read_data
*data
)
457 struct nfs_pgio_header
*hdr
= data
->header
;
458 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
459 struct nfs4_pnfs_ds
*ds
;
460 loff_t offset
= data
->args
.offset
;
465 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
466 __func__
, hdr
->inode
->i_ino
,
467 data
->args
.pgbase
, (size_t)data
->args
.count
, offset
);
469 /* Retrieve the correct rpc_client for the byte range */
470 j
= nfs4_fl_calc_j_index(lseg
, offset
);
471 idx
= nfs4_fl_calc_ds_index(lseg
, j
);
472 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
474 return PNFS_NOT_ATTEMPTED
;
475 dprintk("%s USE DS: %s\n", __func__
, ds
->ds_remotestr
);
477 /* No multipath support. Use first DS */
478 data
->ds_clp
= ds
->ds_clp
;
479 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
483 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
484 data
->mds_offset
= offset
;
486 /* Perform an asynchronous read to ds */
487 status
= nfs_initiate_read(ds
->ds_clp
->cl_rpcclient
, data
,
488 &filelayout_read_call_ops
, RPC_TASK_SOFTCONN
);
490 return PNFS_ATTEMPTED
;
493 /* Perform async writes. */
494 static enum pnfs_try_status
495 filelayout_write_pagelist(struct nfs_write_data
*data
, int sync
)
497 struct nfs_pgio_header
*hdr
= data
->header
;
498 struct pnfs_layout_segment
*lseg
= hdr
->lseg
;
499 struct nfs4_pnfs_ds
*ds
;
500 loff_t offset
= data
->args
.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 ino %lu sync %d req %Zu@%llu DS: %s\n", __func__
,
512 hdr
->inode
->i_ino
, sync
, (size_t) data
->args
.count
, offset
,
515 data
->write_done_cb
= filelayout_write_done_cb
;
516 data
->ds_clp
= ds
->ds_clp
;
517 fh
= nfs4_fl_select_ds_fh(lseg
, j
);
521 * Get the file offset on the dserver. Set the write offset to
522 * this offset and save the original offset.
524 data
->args
.offset
= filelayout_get_dserver_offset(lseg
, offset
);
526 /* Perform an asynchronous write */
527 status
= nfs_initiate_write(ds
->ds_clp
->cl_rpcclient
, data
,
528 &filelayout_write_call_ops
, sync
,
531 return PNFS_ATTEMPTED
;
535 * filelayout_check_layout()
537 * Make sure layout segment parameters are sane WRT the device.
538 * At this point no generic layer initialization of the lseg has occurred,
539 * and nothing has been added to the layout_hdr cache.
543 filelayout_check_layout(struct pnfs_layout_hdr
*lo
,
544 struct nfs4_filelayout_segment
*fl
,
545 struct nfs4_layoutget_res
*lgr
,
546 struct nfs4_deviceid
*id
,
549 struct nfs4_deviceid_node
*d
;
550 struct nfs4_file_layout_dsaddr
*dsaddr
;
551 int status
= -EINVAL
;
552 struct nfs_server
*nfss
= NFS_SERVER(lo
->plh_inode
);
554 dprintk("--> %s\n", __func__
);
556 /* FIXME: remove this check when layout segment support is added */
557 if (lgr
->range
.offset
!= 0 ||
558 lgr
->range
.length
!= NFS4_MAX_UINT64
) {
559 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
564 if (fl
->pattern_offset
> lgr
->range
.offset
) {
565 dprintk("%s pattern_offset %lld too large\n",
566 __func__
, fl
->pattern_offset
);
570 if (!fl
->stripe_unit
|| fl
->stripe_unit
% PAGE_SIZE
) {
571 dprintk("%s Invalid stripe unit (%u)\n",
572 __func__
, fl
->stripe_unit
);
576 /* find and reference the deviceid */
577 d
= nfs4_find_get_deviceid(NFS_SERVER(lo
->plh_inode
)->pnfs_curr_ld
,
578 NFS_SERVER(lo
->plh_inode
)->nfs_client
, id
);
580 dsaddr
= get_device_info(lo
->plh_inode
, id
, gfp_flags
);
584 dsaddr
= container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
);
585 /* Found deviceid is being reaped */
586 if (test_bit(NFS_DEVICEID_INVALID
, &dsaddr
->id_node
.flags
))
591 if (fl
->first_stripe_index
>= dsaddr
->stripe_count
) {
592 dprintk("%s Bad first_stripe_index %u\n",
593 __func__
, fl
->first_stripe_index
);
597 if ((fl
->stripe_type
== STRIPE_SPARSE
&&
598 fl
->num_fh
> 1 && fl
->num_fh
!= dsaddr
->ds_num
) ||
599 (fl
->stripe_type
== STRIPE_DENSE
&&
600 fl
->num_fh
!= dsaddr
->stripe_count
)) {
601 dprintk("%s num_fh %u not valid for given packing\n",
602 __func__
, fl
->num_fh
);
606 if (fl
->stripe_unit
% nfss
->rsize
|| fl
->stripe_unit
% nfss
->wsize
) {
607 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
608 "wsize %u\n", __func__
, fl
->stripe_unit
, nfss
->rsize
,
614 dprintk("--> %s returns %d\n", __func__
, status
);
617 nfs4_fl_put_deviceid(dsaddr
);
621 static void filelayout_free_fh_array(struct nfs4_filelayout_segment
*fl
)
625 for (i
= 0; i
< fl
->num_fh
; i
++) {
626 if (!fl
->fh_array
[i
])
628 kfree(fl
->fh_array
[i
]);
635 _filelayout_free_lseg(struct nfs4_filelayout_segment
*fl
)
637 filelayout_free_fh_array(fl
);
642 filelayout_decode_layout(struct pnfs_layout_hdr
*flo
,
643 struct nfs4_filelayout_segment
*fl
,
644 struct nfs4_layoutget_res
*lgr
,
645 struct nfs4_deviceid
*id
,
648 struct xdr_stream stream
;
650 struct page
*scratch
;
655 dprintk("%s: set_layout_map Begin\n", __func__
);
657 scratch
= alloc_page(gfp_flags
);
661 xdr_init_decode_pages(&stream
, &buf
, lgr
->layoutp
->pages
, lgr
->layoutp
->len
);
662 xdr_set_scratch_buffer(&stream
, page_address(scratch
), PAGE_SIZE
);
664 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
666 p
= xdr_inline_decode(&stream
, NFS4_DEVICEID4_SIZE
+ 20);
670 memcpy(id
, p
, sizeof(*id
));
671 p
+= XDR_QUADLEN(NFS4_DEVICEID4_SIZE
);
672 nfs4_print_deviceid(id
);
674 nfl_util
= be32_to_cpup(p
++);
675 if (nfl_util
& NFL4_UFLG_COMMIT_THRU_MDS
)
676 fl
->commit_through_mds
= 1;
677 if (nfl_util
& NFL4_UFLG_DENSE
)
678 fl
->stripe_type
= STRIPE_DENSE
;
680 fl
->stripe_type
= STRIPE_SPARSE
;
681 fl
->stripe_unit
= nfl_util
& ~NFL4_UFLG_MASK
;
683 fl
->first_stripe_index
= be32_to_cpup(p
++);
684 p
= xdr_decode_hyper(p
, &fl
->pattern_offset
);
685 fl
->num_fh
= be32_to_cpup(p
++);
687 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
688 __func__
, nfl_util
, fl
->num_fh
, fl
->first_stripe_index
,
691 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
692 * Futher checking is done in filelayout_check_layout */
694 max(NFS4_PNFS_MAX_STRIPE_CNT
, NFS4_PNFS_MAX_MULTI_CNT
))
697 if (fl
->num_fh
> 0) {
698 fl
->fh_array
= kzalloc(fl
->num_fh
* sizeof(struct nfs_fh
*),
704 for (i
= 0; i
< fl
->num_fh
; i
++) {
705 /* Do we want to use a mempool here? */
706 fl
->fh_array
[i
] = kmalloc(sizeof(struct nfs_fh
), gfp_flags
);
707 if (!fl
->fh_array
[i
])
710 p
= xdr_inline_decode(&stream
, 4);
713 fl
->fh_array
[i
]->size
= be32_to_cpup(p
++);
714 if (sizeof(struct nfs_fh
) < fl
->fh_array
[i
]->size
) {
715 printk(KERN_ERR
"NFS: Too big fh %d received %d\n",
716 i
, fl
->fh_array
[i
]->size
);
720 p
= xdr_inline_decode(&stream
, fl
->fh_array
[i
]->size
);
723 memcpy(fl
->fh_array
[i
]->data
, p
, fl
->fh_array
[i
]->size
);
724 dprintk("DEBUG: %s: fh len %d\n", __func__
,
725 fl
->fh_array
[i
]->size
);
728 __free_page(scratch
);
732 filelayout_free_fh_array(fl
);
734 __free_page(scratch
);
739 filelayout_free_lseg(struct pnfs_layout_segment
*lseg
)
741 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
743 dprintk("--> %s\n", __func__
);
744 nfs4_fl_put_deviceid(fl
->dsaddr
);
745 /* This assumes a single RW lseg */
746 if (lseg
->pls_range
.iomode
== IOMODE_RW
) {
747 struct nfs4_filelayout
*flo
;
749 flo
= FILELAYOUT_FROM_HDR(lseg
->pls_layout
);
750 flo
->commit_info
.nbuckets
= 0;
751 kfree(flo
->commit_info
.buckets
);
752 flo
->commit_info
.buckets
= NULL
;
754 _filelayout_free_lseg(fl
);
758 filelayout_alloc_commit_info(struct pnfs_layout_segment
*lseg
,
759 struct nfs_commit_info
*cinfo
,
762 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
763 struct pnfs_commit_bucket
*buckets
;
766 if (fl
->commit_through_mds
)
768 if (cinfo
->ds
->nbuckets
!= 0) {
769 /* This assumes there is only one IOMODE_RW lseg. What
770 * we really want to do is have a layout_hdr level
771 * dictionary of <multipath_list4, fh> keys, each
772 * associated with a struct list_head, populated by calls
773 * to filelayout_write_pagelist().
778 size
= (fl
->stripe_type
== STRIPE_SPARSE
) ?
779 fl
->dsaddr
->ds_num
: fl
->dsaddr
->stripe_count
;
781 buckets
= kcalloc(size
, sizeof(struct pnfs_commit_bucket
),
788 spin_lock(cinfo
->lock
);
789 if (cinfo
->ds
->nbuckets
!= 0)
792 cinfo
->ds
->buckets
= buckets
;
793 cinfo
->ds
->nbuckets
= size
;
794 for (i
= 0; i
< size
; i
++) {
795 INIT_LIST_HEAD(&buckets
[i
].written
);
796 INIT_LIST_HEAD(&buckets
[i
].committing
);
799 spin_unlock(cinfo
->lock
);
804 static struct pnfs_layout_segment
*
805 filelayout_alloc_lseg(struct pnfs_layout_hdr
*layoutid
,
806 struct nfs4_layoutget_res
*lgr
,
809 struct nfs4_filelayout_segment
*fl
;
811 struct nfs4_deviceid id
;
813 dprintk("--> %s\n", __func__
);
814 fl
= kzalloc(sizeof(*fl
), gfp_flags
);
818 rc
= filelayout_decode_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
);
819 if (rc
!= 0 || filelayout_check_layout(layoutid
, fl
, lgr
, &id
, gfp_flags
)) {
820 _filelayout_free_lseg(fl
);
823 return &fl
->generic_hdr
;
827 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
829 * return true : coalesce page
830 * return false : don't coalesce page
833 filelayout_pg_test(struct nfs_pageio_descriptor
*pgio
, struct nfs_page
*prev
,
834 struct nfs_page
*req
)
836 u64 p_stripe
, r_stripe
;
839 if (!pnfs_generic_pg_test(pgio
, prev
, req
) ||
840 !nfs_generic_pg_test(pgio
, prev
, req
))
843 p_stripe
= (u64
)req_offset(prev
);
844 r_stripe
= (u64
)req_offset(req
);
845 stripe_unit
= FILELAYOUT_LSEG(pgio
->pg_lseg
)->stripe_unit
;
847 do_div(p_stripe
, stripe_unit
);
848 do_div(r_stripe
, stripe_unit
);
850 return (p_stripe
== r_stripe
);
854 filelayout_pg_init_read(struct nfs_pageio_descriptor
*pgio
,
855 struct nfs_page
*req
)
857 BUG_ON(pgio
->pg_lseg
!= NULL
);
859 if (req
->wb_offset
!= req
->wb_pgbase
) {
861 * Handling unaligned pages is difficult, because have to
862 * somehow split a req in two in certain cases in the
863 * pg.test code. Avoid this by just not using pnfs
866 nfs_pageio_reset_read_mds(pgio
);
869 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
875 /* If no lseg, fall back to read through mds */
876 if (pgio
->pg_lseg
== NULL
)
877 nfs_pageio_reset_read_mds(pgio
);
881 filelayout_pg_init_write(struct nfs_pageio_descriptor
*pgio
,
882 struct nfs_page
*req
)
884 struct nfs_commit_info cinfo
;
887 BUG_ON(pgio
->pg_lseg
!= NULL
);
889 if (req
->wb_offset
!= req
->wb_pgbase
)
891 pgio
->pg_lseg
= pnfs_update_layout(pgio
->pg_inode
,
897 /* If no lseg, fall back to write through mds */
898 if (pgio
->pg_lseg
== NULL
)
900 nfs_init_cinfo(&cinfo
, pgio
->pg_inode
, pgio
->pg_dreq
);
901 status
= filelayout_alloc_commit_info(pgio
->pg_lseg
, &cinfo
, GFP_NOFS
);
903 put_lseg(pgio
->pg_lseg
);
904 pgio
->pg_lseg
= NULL
;
909 nfs_pageio_reset_write_mds(pgio
);
912 static const struct nfs_pageio_ops filelayout_pg_read_ops
= {
913 .pg_init
= filelayout_pg_init_read
,
914 .pg_test
= filelayout_pg_test
,
915 .pg_doio
= pnfs_generic_pg_readpages
,
918 static const struct nfs_pageio_ops filelayout_pg_write_ops
= {
919 .pg_init
= filelayout_pg_init_write
,
920 .pg_test
= filelayout_pg_test
,
921 .pg_doio
= pnfs_generic_pg_writepages
,
924 static u32
select_bucket_index(struct nfs4_filelayout_segment
*fl
, u32 j
)
926 if (fl
->stripe_type
== STRIPE_SPARSE
)
927 return nfs4_fl_calc_ds_index(&fl
->generic_hdr
, j
);
932 /* The generic layer is about to remove the req from the commit list.
933 * If this will make the bucket empty, it will need to put the lseg reference.
936 filelayout_clear_request_commit(struct nfs_page
*req
,
937 struct nfs_commit_info
*cinfo
)
939 struct pnfs_layout_segment
*freeme
= NULL
;
941 spin_lock(cinfo
->lock
);
942 if (!test_and_clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
))
944 cinfo
->ds
->nwritten
--;
945 if (list_is_singular(&req
->wb_list
)) {
946 struct pnfs_commit_bucket
*bucket
;
948 bucket
= list_first_entry(&req
->wb_list
,
949 struct pnfs_commit_bucket
,
951 freeme
= bucket
->wlseg
;
952 bucket
->wlseg
= NULL
;
955 nfs_request_remove_commit_list(req
, cinfo
);
956 spin_unlock(cinfo
->lock
);
960 static struct list_head
*
961 filelayout_choose_commit_list(struct nfs_page
*req
,
962 struct pnfs_layout_segment
*lseg
,
963 struct nfs_commit_info
*cinfo
)
965 struct nfs4_filelayout_segment
*fl
= FILELAYOUT_LSEG(lseg
);
967 struct list_head
*list
;
968 struct pnfs_commit_bucket
*buckets
;
970 if (fl
->commit_through_mds
)
971 return &cinfo
->mds
->list
;
973 /* Note that we are calling nfs4_fl_calc_j_index on each page
974 * that ends up being committed to a data server. An attractive
975 * alternative is to add a field to nfs_write_data and nfs_page
976 * to store the value calculated in filelayout_write_pagelist
977 * and just use that here.
979 j
= nfs4_fl_calc_j_index(lseg
, req_offset(req
));
980 i
= select_bucket_index(fl
, j
);
981 buckets
= cinfo
->ds
->buckets
;
982 list
= &buckets
[i
].written
;
983 if (list_empty(list
)) {
984 /* Non-empty buckets hold a reference on the lseg. That ref
985 * is normally transferred to the COMMIT call and released
986 * there. It could also be released if the last req is pulled
987 * off due to a rewrite, in which case it will be done in
988 * filelayout_clear_request_commit
990 buckets
[i
].wlseg
= get_lseg(lseg
);
992 set_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
993 cinfo
->ds
->nwritten
++;
998 filelayout_mark_request_commit(struct nfs_page
*req
,
999 struct pnfs_layout_segment
*lseg
,
1000 struct nfs_commit_info
*cinfo
)
1002 struct list_head
*list
;
1004 list
= filelayout_choose_commit_list(req
, lseg
, cinfo
);
1005 nfs_request_add_commit_list(req
, list
, cinfo
);
1008 static u32
calc_ds_index_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1010 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1012 if (flseg
->stripe_type
== STRIPE_SPARSE
)
1015 return nfs4_fl_calc_ds_index(lseg
, i
);
1018 static struct nfs_fh
*
1019 select_ds_fh_from_commit(struct pnfs_layout_segment
*lseg
, u32 i
)
1021 struct nfs4_filelayout_segment
*flseg
= FILELAYOUT_LSEG(lseg
);
1023 if (flseg
->stripe_type
== STRIPE_SPARSE
) {
1024 if (flseg
->num_fh
== 1)
1026 else if (flseg
->num_fh
== 0)
1027 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1030 return flseg
->fh_array
[i
];
1033 static int filelayout_initiate_commit(struct nfs_commit_data
*data
, int how
)
1035 struct pnfs_layout_segment
*lseg
= data
->lseg
;
1036 struct nfs4_pnfs_ds
*ds
;
1040 idx
= calc_ds_index_from_commit(lseg
, data
->ds_commit_index
);
1041 ds
= nfs4_fl_prepare_ds(lseg
, idx
);
1043 prepare_to_resend_writes(data
);
1044 filelayout_commit_release(data
);
1047 dprintk("%s ino %lu, how %d\n", __func__
, data
->inode
->i_ino
, how
);
1048 data
->commit_done_cb
= filelayout_commit_done_cb
;
1049 data
->ds_clp
= ds
->ds_clp
;
1050 fh
= select_ds_fh_from_commit(lseg
, data
->ds_commit_index
);
1053 return nfs_initiate_commit(ds
->ds_clp
->cl_rpcclient
, data
,
1054 &filelayout_commit_call_ops
, how
,
1059 transfer_commit_list(struct list_head
*src
, struct list_head
*dst
,
1060 struct nfs_commit_info
*cinfo
, int max
)
1062 struct nfs_page
*req
, *tmp
;
1065 list_for_each_entry_safe(req
, tmp
, src
, wb_list
) {
1066 if (!nfs_lock_request(req
))
1068 if (cond_resched_lock(cinfo
->lock
))
1069 list_safe_reset_next(req
, tmp
, wb_list
);
1070 nfs_request_remove_commit_list(req
, cinfo
);
1071 clear_bit(PG_COMMIT_TO_DS
, &req
->wb_flags
);
1072 nfs_list_add_request(req
, dst
);
1074 if ((ret
== max
) && !cinfo
->dreq
)
1081 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket
*bucket
,
1082 struct nfs_commit_info
*cinfo
,
1085 struct list_head
*src
= &bucket
->written
;
1086 struct list_head
*dst
= &bucket
->committing
;
1089 ret
= transfer_commit_list(src
, dst
, cinfo
, max
);
1091 cinfo
->ds
->nwritten
-= ret
;
1092 cinfo
->ds
->ncommitting
+= ret
;
1093 bucket
->clseg
= bucket
->wlseg
;
1094 if (list_empty(src
))
1095 bucket
->wlseg
= NULL
;
1097 get_lseg(bucket
->clseg
);
1102 /* Move reqs from written to committing lists, returning count of number moved.
1103 * Note called with cinfo->lock held.
1105 static int filelayout_scan_commit_lists(struct nfs_commit_info
*cinfo
,
1110 for (i
= 0; i
< cinfo
->ds
->nbuckets
&& max
!= 0; i
++) {
1111 cnt
= filelayout_scan_ds_commit_list(&cinfo
->ds
->buckets
[i
],
1119 /* Pull everything off the committing lists and dump into @dst */
1120 static void filelayout_recover_commit_reqs(struct list_head
*dst
,
1121 struct nfs_commit_info
*cinfo
)
1123 struct pnfs_commit_bucket
*b
;
1126 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1127 * only called on single thread per dreq.
1128 * Can't take the lock because need to do put_lseg
1130 for (i
= 0, b
= cinfo
->ds
->buckets
; i
< cinfo
->ds
->nbuckets
; i
++, b
++) {
1131 if (transfer_commit_list(&b
->written
, dst
, cinfo
, 0)) {
1132 BUG_ON(!list_empty(&b
->written
));
1137 cinfo
->ds
->nwritten
= 0;
1141 alloc_ds_commits(struct nfs_commit_info
*cinfo
, struct list_head
*list
)
1143 struct pnfs_ds_commit_info
*fl_cinfo
;
1144 struct pnfs_commit_bucket
*bucket
;
1145 struct nfs_commit_data
*data
;
1147 unsigned int nreq
= 0;
1149 fl_cinfo
= cinfo
->ds
;
1150 bucket
= fl_cinfo
->buckets
;
1151 for (i
= 0; i
< fl_cinfo
->nbuckets
; i
++, bucket
++) {
1152 if (list_empty(&bucket
->committing
))
1154 data
= nfs_commitdata_alloc();
1157 data
->ds_commit_index
= i
;
1158 data
->lseg
= bucket
->clseg
;
1159 bucket
->clseg
= NULL
;
1160 list_add(&data
->pages
, list
);
1164 /* Clean up on error */
1165 for (j
= i
; j
< fl_cinfo
->nbuckets
; j
++, bucket
++) {
1166 if (list_empty(&bucket
->committing
))
1168 nfs_retry_commit(&bucket
->committing
, bucket
->clseg
, cinfo
);
1169 put_lseg(bucket
->clseg
);
1170 bucket
->clseg
= NULL
;
1172 /* Caller will clean up entries put on list */
1176 /* This follows nfs_commit_list pretty closely */
1178 filelayout_commit_pagelist(struct inode
*inode
, struct list_head
*mds_pages
,
1179 int how
, struct nfs_commit_info
*cinfo
)
1181 struct nfs_commit_data
*data
, *tmp
;
1183 unsigned int nreq
= 0;
1185 if (!list_empty(mds_pages
)) {
1186 data
= nfs_commitdata_alloc();
1189 list_add(&data
->pages
, &list
);
1192 nfs_retry_commit(mds_pages
, NULL
, cinfo
);
1195 nreq
+= alloc_ds_commits(cinfo
, &list
);
1198 cinfo
->completion_ops
->error_cleanup(NFS_I(inode
));
1202 atomic_add(nreq
, &cinfo
->mds
->rpcs_out
);
1204 list_for_each_entry_safe(data
, tmp
, &list
, pages
) {
1205 list_del_init(&data
->pages
);
1207 nfs_init_commit(data
, mds_pages
, NULL
, cinfo
);
1208 nfs_initiate_commit(NFS_CLIENT(inode
), data
,
1209 data
->mds_ops
, how
, 0);
1211 struct pnfs_commit_bucket
*buckets
;
1213 buckets
= cinfo
->ds
->buckets
;
1214 nfs_init_commit(data
, &buckets
[data
->ds_commit_index
].committing
, data
->lseg
, cinfo
);
1215 filelayout_initiate_commit(data
, how
);
1219 cinfo
->ds
->ncommitting
= 0;
1220 return PNFS_ATTEMPTED
;
1224 filelayout_free_deveiceid_node(struct nfs4_deviceid_node
*d
)
1226 nfs4_fl_free_deviceid(container_of(d
, struct nfs4_file_layout_dsaddr
, id_node
));
1229 static struct pnfs_layout_hdr
*
1230 filelayout_alloc_layout_hdr(struct inode
*inode
, gfp_t gfp_flags
)
1232 struct nfs4_filelayout
*flo
;
1234 flo
= kzalloc(sizeof(*flo
), gfp_flags
);
1235 return &flo
->generic_hdr
;
1239 filelayout_free_layout_hdr(struct pnfs_layout_hdr
*lo
)
1241 kfree(FILELAYOUT_FROM_HDR(lo
));
1244 static struct pnfs_ds_commit_info
*
1245 filelayout_get_ds_info(struct inode
*inode
)
1247 struct pnfs_layout_hdr
*layout
= NFS_I(inode
)->layout
;
1252 return &FILELAYOUT_FROM_HDR(layout
)->commit_info
;
1255 static struct pnfs_layoutdriver_type filelayout_type
= {
1256 .id
= LAYOUT_NFSV4_1_FILES
,
1257 .name
= "LAYOUT_NFSV4_1_FILES",
1258 .owner
= THIS_MODULE
,
1259 .alloc_layout_hdr
= filelayout_alloc_layout_hdr
,
1260 .free_layout_hdr
= filelayout_free_layout_hdr
,
1261 .alloc_lseg
= filelayout_alloc_lseg
,
1262 .free_lseg
= filelayout_free_lseg
,
1263 .pg_read_ops
= &filelayout_pg_read_ops
,
1264 .pg_write_ops
= &filelayout_pg_write_ops
,
1265 .get_ds_info
= &filelayout_get_ds_info
,
1266 .mark_request_commit
= filelayout_mark_request_commit
,
1267 .clear_request_commit
= filelayout_clear_request_commit
,
1268 .scan_commit_lists
= filelayout_scan_commit_lists
,
1269 .recover_commit_reqs
= filelayout_recover_commit_reqs
,
1270 .commit_pagelist
= filelayout_commit_pagelist
,
1271 .read_pagelist
= filelayout_read_pagelist
,
1272 .write_pagelist
= filelayout_write_pagelist
,
1273 .free_deviceid_node
= filelayout_free_deveiceid_node
,
1276 static int __init
nfs4filelayout_init(void)
1278 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Registering...\n",
1280 return pnfs_register_layoutdriver(&filelayout_type
);
1283 static void __exit
nfs4filelayout_exit(void)
1285 printk(KERN_INFO
"%s: NFSv4 File Layout Driver Unregistering...\n",
1287 pnfs_unregister_layoutdriver(&filelayout_type
);
1290 MODULE_ALIAS("nfs-layouttype4-1");
1292 module_init(nfs4filelayout_init
);
1293 module_exit(nfs4filelayout_exit
);