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>
38 #include "../nfs4session.h"
39 #include "../internal.h"
40 #include "../delegation.h"
41 #include "filelayout.h"
42 #include "../nfs4trace.h"
44 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
46 MODULE_LICENSE("GPL");
47 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
48 MODULE_DESCRIPTION("The NFSv4 file layout driver");
50 #define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
53 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
56 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
60 offset -= flseg->pattern_offset;
61 stripe_no = div_u64(offset, stripe_width);
62 div_u64_rem(offset, flseg->stripe_unit, &rem);
64 return stripe_no * flseg->stripe_unit + rem;
67 /* This function is used by the layout driver to calculate the
68 * offset of the file on the dserver based on whether the
69 * layout type is STRIPE_DENSE or STRIPE_SPARSE
72 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
74 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
76 switch (flseg->stripe_type) {
81 return filelayout_get_dense_offset(flseg, offset);
87 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
89 struct rpc_task *task = &hdr->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/%llu, %u bytes @ offset %llu)\n", __func__,
95 hdr->inode->i_sb->s_id,
96 (unsigned long long)NFS_FILEID(hdr->inode),
98 (unsigned long long)hdr->args.offset);
100 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
104 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
106 struct rpc_task *task = &hdr->task;
108 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
109 dprintk("%s Reset task %5u for i/o through MDS "
110 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
112 hdr->inode->i_sb->s_id,
113 (unsigned long long)NFS_FILEID(hdr->inode),
115 (unsigned long long)hdr->args.offset);
117 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
121 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
123 if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
125 pnfs_return_layout(inode);
128 static int filelayout_async_handle_error(struct rpc_task *task,
129 struct nfs4_state *state,
130 struct nfs_client *clp,
131 struct pnfs_layout_segment *lseg)
133 struct pnfs_layout_hdr *lo = lseg->pls_layout;
134 struct inode *inode = lo->plh_inode;
135 struct nfs_server *mds_server = NFS_SERVER(inode);
136 struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
137 struct nfs_client *mds_client = mds_server->nfs_client;
138 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
140 if (task->tk_status >= 0)
143 switch (task->tk_status) {
144 /* MDS state errors */
145 case -NFS4ERR_DELEG_REVOKED:
146 case -NFS4ERR_ADMIN_REVOKED:
147 case -NFS4ERR_BAD_STATEID:
150 nfs_remove_bad_delegation(state->inode);
151 case -NFS4ERR_OPENMODE:
154 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
155 goto out_bad_stateid;
156 goto wait_on_recovery;
157 case -NFS4ERR_EXPIRED:
159 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
160 goto out_bad_stateid;
162 nfs4_schedule_lease_recovery(mds_client);
163 goto wait_on_recovery;
164 /* DS session errors */
165 case -NFS4ERR_BADSESSION:
166 case -NFS4ERR_BADSLOT:
167 case -NFS4ERR_BAD_HIGH_SLOT:
168 case -NFS4ERR_DEADSESSION:
169 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
170 case -NFS4ERR_SEQ_FALSE_RETRY:
171 case -NFS4ERR_SEQ_MISORDERED:
172 dprintk("%s ERROR %d, Reset session. Exchangeid "
173 "flags 0x%x\n", __func__, task->tk_status,
174 clp->cl_exchange_flags);
175 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
179 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
181 case -NFS4ERR_RETRY_UNCACHED_REP:
183 /* Invalidate Layout errors */
184 case -NFS4ERR_PNFS_NO_LAYOUT:
185 case -ESTALE: /* mapped NFS4ERR_STALE */
186 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
187 case -EISDIR: /* mapped NFS4ERR_ISDIR */
188 case -NFS4ERR_FHEXPIRED:
189 case -NFS4ERR_WRONG_TYPE:
190 dprintk("%s Invalid layout error %d\n", __func__,
193 * Destroy layout so new i/o will get a new layout.
194 * Layout will not be destroyed until all current lseg
195 * references are put. Mark layout as invalid to resend failed
196 * i/o and all i/o waiting on the slot table to the MDS until
197 * layout is destroyed and a new valid layout is obtained.
199 pnfs_destroy_layout(NFS_I(inode));
200 rpc_wake_up(&tbl->slot_tbl_waitq);
202 /* RPC connection errors */
210 dprintk("%s DS connection error %d\n", __func__,
212 nfs4_mark_deviceid_unavailable(devid);
213 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
214 rpc_wake_up(&tbl->slot_tbl_waitq);
218 dprintk("%s Retry through MDS. Error %d\n", __func__,
220 return -NFS4ERR_RESET_TO_MDS;
226 task->tk_status = -EIO;
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_pgio_header *hdr)
242 trace_nfs4_pnfs_read(hdr, task->tk_status);
243 err = filelayout_async_handle_error(task, hdr->args.context->state,
244 hdr->ds_clp, hdr->lseg);
247 case -NFS4ERR_RESET_TO_MDS:
248 filelayout_reset_read(hdr);
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_pgio_header *hdr)
267 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
268 hdr->res.verf->committed != NFS_DATA_SYNC)
271 pnfs_set_layoutcommit(hdr);
272 dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
273 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
277 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
279 return filelayout_test_devid_invalid(node) ||
280 nfs4_test_deviceid_unavailable(node);
284 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
286 struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
288 return filelayout_test_devid_unavailable(node);
292 * Call ops for the async read/write cases
293 * In the case of dense layouts, the offset needs to be reset to its
296 static void filelayout_read_prepare(struct rpc_task *task, void *data)
298 struct nfs_pgio_header *hdr = data;
300 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
301 rpc_exit(task, -EIO);
304 if (filelayout_reset_to_mds(hdr->lseg)) {
305 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
306 filelayout_reset_read(hdr);
310 hdr->pgio_done_cb = filelayout_read_done_cb;
312 if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
317 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
318 hdr->args.lock_context, FMODE_READ) == -EIO)
319 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
322 static void filelayout_read_call_done(struct rpc_task *task, void *data)
324 struct nfs_pgio_header *hdr = data;
326 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
328 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
329 task->tk_status == 0) {
330 nfs41_sequence_done(task, &hdr->res.seq_res);
334 /* Note this may cause RPC to be resent */
335 hdr->mds_ops->rpc_call_done(task, data);
338 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
340 struct nfs_pgio_header *hdr = data;
342 rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
345 static void filelayout_read_release(void *data)
347 struct nfs_pgio_header *hdr = data;
348 struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
350 filelayout_fenceme(lo->plh_inode, lo);
351 nfs_put_client(hdr->ds_clp);
352 hdr->mds_ops->rpc_release(data);
355 static int filelayout_write_done_cb(struct rpc_task *task,
356 struct nfs_pgio_header *hdr)
360 trace_nfs4_pnfs_write(hdr, task->tk_status);
361 err = filelayout_async_handle_error(task, hdr->args.context->state,
362 hdr->ds_clp, hdr->lseg);
365 case -NFS4ERR_RESET_TO_MDS:
366 filelayout_reset_write(hdr);
367 return task->tk_status;
369 rpc_restart_call_prepare(task);
373 filelayout_set_layoutcommit(hdr);
377 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
378 static void prepare_to_resend_writes(struct nfs_commit_data *data)
380 struct nfs_page *first = nfs_list_entry(data->pages.next);
382 data->task.tk_status = 0;
383 memcpy(&data->verf.verifier, &first->wb_verf,
384 sizeof(data->verf.verifier));
385 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
388 static int filelayout_commit_done_cb(struct rpc_task *task,
389 struct nfs_commit_data *data)
393 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
394 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
398 case -NFS4ERR_RESET_TO_MDS:
399 prepare_to_resend_writes(data);
402 rpc_restart_call_prepare(task);
406 if (data->verf.committed == NFS_UNSTABLE)
407 pnfs_commit_set_layoutcommit(data);
412 static void filelayout_write_prepare(struct rpc_task *task, void *data)
414 struct nfs_pgio_header *hdr = data;
416 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
417 rpc_exit(task, -EIO);
420 if (filelayout_reset_to_mds(hdr->lseg)) {
421 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
422 filelayout_reset_write(hdr);
426 if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
431 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
432 hdr->args.lock_context, FMODE_WRITE) == -EIO)
433 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
436 static void filelayout_write_call_done(struct rpc_task *task, void *data)
438 struct nfs_pgio_header *hdr = data;
440 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
441 task->tk_status == 0) {
442 nfs41_sequence_done(task, &hdr->res.seq_res);
446 /* Note this may cause RPC to be resent */
447 hdr->mds_ops->rpc_call_done(task, data);
450 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
452 struct nfs_pgio_header *hdr = data;
454 rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
457 static void filelayout_write_release(void *data)
459 struct nfs_pgio_header *hdr = data;
460 struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
462 filelayout_fenceme(lo->plh_inode, lo);
463 nfs_put_client(hdr->ds_clp);
464 hdr->mds_ops->rpc_release(data);
467 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
469 struct nfs_commit_data *wdata = data;
471 nfs41_setup_sequence(wdata->ds_clp->cl_session,
472 &wdata->args.seq_args,
477 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
479 struct nfs_commit_data *wdata = data;
481 /* Note this may cause RPC to be resent */
482 wdata->mds_ops->rpc_call_done(task, data);
485 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
487 struct nfs_commit_data *cdata = data;
489 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
492 static void filelayout_commit_release(void *calldata)
494 struct nfs_commit_data *data = calldata;
496 data->completion_ops->completion(data);
497 pnfs_put_lseg(data->lseg);
498 nfs_put_client(data->ds_clp);
499 nfs_commitdata_release(data);
502 static const struct rpc_call_ops filelayout_read_call_ops = {
503 .rpc_call_prepare = filelayout_read_prepare,
504 .rpc_call_done = filelayout_read_call_done,
505 .rpc_count_stats = filelayout_read_count_stats,
506 .rpc_release = filelayout_read_release,
509 static const struct rpc_call_ops filelayout_write_call_ops = {
510 .rpc_call_prepare = filelayout_write_prepare,
511 .rpc_call_done = filelayout_write_call_done,
512 .rpc_count_stats = filelayout_write_count_stats,
513 .rpc_release = filelayout_write_release,
516 static const struct rpc_call_ops filelayout_commit_call_ops = {
517 .rpc_call_prepare = filelayout_commit_prepare,
518 .rpc_call_done = filelayout_write_commit_done,
519 .rpc_count_stats = filelayout_commit_count_stats,
520 .rpc_release = filelayout_commit_release,
523 static enum pnfs_try_status
524 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
526 struct pnfs_layout_segment *lseg = hdr->lseg;
527 struct nfs4_pnfs_ds *ds;
528 struct rpc_clnt *ds_clnt;
529 loff_t offset = hdr->args.offset;
533 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
534 __func__, hdr->inode->i_ino,
535 hdr->args.pgbase, (size_t)hdr->args.count, offset);
537 /* Retrieve the correct rpc_client for the byte range */
538 j = nfs4_fl_calc_j_index(lseg, offset);
539 idx = nfs4_fl_calc_ds_index(lseg, j);
540 ds = nfs4_fl_prepare_ds(lseg, idx);
542 return PNFS_NOT_ATTEMPTED;
544 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
546 return PNFS_NOT_ATTEMPTED;
548 dprintk("%s USE DS: %s cl_count %d\n", __func__,
549 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
551 /* No multipath support. Use first DS */
552 atomic_inc(&ds->ds_clp->cl_count);
553 hdr->ds_clp = ds->ds_clp;
555 fh = nfs4_fl_select_ds_fh(lseg, j);
559 hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
560 hdr->mds_offset = offset;
562 /* Perform an asynchronous read to ds */
563 nfs_initiate_pgio(ds_clnt, hdr,
564 &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
565 return PNFS_ATTEMPTED;
568 /* Perform async writes. */
569 static enum pnfs_try_status
570 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
572 struct pnfs_layout_segment *lseg = hdr->lseg;
573 struct nfs4_pnfs_ds *ds;
574 struct rpc_clnt *ds_clnt;
575 loff_t offset = hdr->args.offset;
579 /* Retrieve the correct rpc_client for the byte range */
580 j = nfs4_fl_calc_j_index(lseg, offset);
581 idx = nfs4_fl_calc_ds_index(lseg, j);
582 ds = nfs4_fl_prepare_ds(lseg, idx);
584 return PNFS_NOT_ATTEMPTED;
586 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
588 return PNFS_NOT_ATTEMPTED;
590 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
591 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
592 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
594 hdr->pgio_done_cb = filelayout_write_done_cb;
595 atomic_inc(&ds->ds_clp->cl_count);
596 hdr->ds_clp = ds->ds_clp;
598 fh = nfs4_fl_select_ds_fh(lseg, j);
601 hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
603 /* Perform an asynchronous write */
604 nfs_initiate_pgio(ds_clnt, hdr,
605 &filelayout_write_call_ops, sync,
607 return PNFS_ATTEMPTED;
611 * filelayout_check_layout()
613 * Make sure layout segment parameters are sane WRT the device.
614 * At this point no generic layer initialization of the lseg has occurred,
615 * and nothing has been added to the layout_hdr cache.
619 filelayout_check_layout(struct pnfs_layout_hdr *lo,
620 struct nfs4_filelayout_segment *fl,
621 struct nfs4_layoutget_res *lgr,
622 struct nfs4_deviceid *id,
625 struct nfs4_deviceid_node *d;
626 struct nfs4_file_layout_dsaddr *dsaddr;
627 int status = -EINVAL;
629 dprintk("--> %s\n", __func__);
631 /* FIXME: remove this check when layout segment support is added */
632 if (lgr->range.offset != 0 ||
633 lgr->range.length != NFS4_MAX_UINT64) {
634 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
639 if (fl->pattern_offset > lgr->range.offset) {
640 dprintk("%s pattern_offset %lld too large\n",
641 __func__, fl->pattern_offset);
645 if (!fl->stripe_unit) {
646 dprintk("%s Invalid stripe unit (%u)\n",
647 __func__, fl->stripe_unit);
651 /* find and reference the deviceid */
652 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
653 NFS_SERVER(lo->plh_inode)->nfs_client, id);
655 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
656 lo->plh_lc_cred, gfp_flags);
660 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
661 /* Found deviceid is unavailable */
662 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
667 if (fl->first_stripe_index >= dsaddr->stripe_count) {
668 dprintk("%s Bad first_stripe_index %u\n",
669 __func__, fl->first_stripe_index);
673 if ((fl->stripe_type == STRIPE_SPARSE &&
674 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
675 (fl->stripe_type == STRIPE_DENSE &&
676 fl->num_fh != dsaddr->stripe_count)) {
677 dprintk("%s num_fh %u not valid for given packing\n",
678 __func__, fl->num_fh);
684 dprintk("--> %s returns %d\n", __func__, status);
687 nfs4_fl_put_deviceid(dsaddr);
691 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
695 for (i = 0; i < fl->num_fh; i++) {
696 if (!fl->fh_array[i])
698 kfree(fl->fh_array[i]);
705 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
707 filelayout_free_fh_array(fl);
712 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
713 struct nfs4_filelayout_segment *fl,
714 struct nfs4_layoutget_res *lgr,
715 struct nfs4_deviceid *id,
718 struct xdr_stream stream;
720 struct page *scratch;
725 dprintk("%s: set_layout_map Begin\n", __func__);
727 scratch = alloc_page(gfp_flags);
731 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
732 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
734 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
736 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
740 memcpy(id, p, sizeof(*id));
741 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
742 nfs4_print_deviceid(id);
744 nfl_util = be32_to_cpup(p++);
745 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
746 fl->commit_through_mds = 1;
747 if (nfl_util & NFL4_UFLG_DENSE)
748 fl->stripe_type = STRIPE_DENSE;
750 fl->stripe_type = STRIPE_SPARSE;
751 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
753 fl->first_stripe_index = be32_to_cpup(p++);
754 p = xdr_decode_hyper(p, &fl->pattern_offset);
755 fl->num_fh = be32_to_cpup(p++);
757 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
758 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
761 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
762 * Futher checking is done in filelayout_check_layout */
764 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
767 if (fl->num_fh > 0) {
768 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
774 for (i = 0; i < fl->num_fh; i++) {
775 /* Do we want to use a mempool here? */
776 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
777 if (!fl->fh_array[i])
780 p = xdr_inline_decode(&stream, 4);
783 fl->fh_array[i]->size = be32_to_cpup(p++);
784 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
785 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
786 i, fl->fh_array[i]->size);
790 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
793 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
794 dprintk("DEBUG: %s: fh len %d\n", __func__,
795 fl->fh_array[i]->size);
798 __free_page(scratch);
802 filelayout_free_fh_array(fl);
804 __free_page(scratch);
809 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
811 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
813 dprintk("--> %s\n", __func__);
814 nfs4_fl_put_deviceid(fl->dsaddr);
815 /* This assumes a single RW lseg */
816 if (lseg->pls_range.iomode == IOMODE_RW) {
817 struct nfs4_filelayout *flo;
819 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
820 flo->commit_info.nbuckets = 0;
821 kfree(flo->commit_info.buckets);
822 flo->commit_info.buckets = NULL;
824 _filelayout_free_lseg(fl);
828 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
829 struct nfs_commit_info *cinfo,
832 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
833 struct pnfs_commit_bucket *buckets;
836 if (fl->commit_through_mds)
839 size = (fl->stripe_type == STRIPE_SPARSE) ?
840 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
842 if (cinfo->ds->nbuckets >= size) {
843 /* This assumes there is only one IOMODE_RW lseg. What
844 * we really want to do is have a layout_hdr level
845 * dictionary of <multipath_list4, fh> keys, each
846 * associated with a struct list_head, populated by calls
847 * to filelayout_write_pagelist().
852 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
856 for (i = 0; i < size; i++) {
857 INIT_LIST_HEAD(&buckets[i].written);
858 INIT_LIST_HEAD(&buckets[i].committing);
859 /* mark direct verifier as unset */
860 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
863 spin_lock(cinfo->lock);
864 if (cinfo->ds->nbuckets >= size)
866 for (i = 0; i < cinfo->ds->nbuckets; i++) {
867 list_splice(&cinfo->ds->buckets[i].written,
868 &buckets[i].written);
869 list_splice(&cinfo->ds->buckets[i].committing,
870 &buckets[i].committing);
871 buckets[i].direct_verf.committed =
872 cinfo->ds->buckets[i].direct_verf.committed;
873 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
874 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
876 swap(cinfo->ds->buckets, buckets);
877 cinfo->ds->nbuckets = size;
879 spin_unlock(cinfo->lock);
884 static struct pnfs_layout_segment *
885 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
886 struct nfs4_layoutget_res *lgr,
889 struct nfs4_filelayout_segment *fl;
891 struct nfs4_deviceid id;
893 dprintk("--> %s\n", __func__);
894 fl = kzalloc(sizeof(*fl), gfp_flags);
898 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
899 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
900 _filelayout_free_lseg(fl);
903 return &fl->generic_hdr;
907 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
909 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
910 * of bytes (maximum @req->wb_bytes) that can be coalesced.
913 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
914 struct nfs_page *req)
917 u64 p_stripe, r_stripe;
919 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
920 u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
922 /* calls nfs_generic_pg_test */
923 size = pnfs_generic_pg_test(pgio, prev, req);
927 /* see if req and prev are in the same stripe */
929 p_stripe = (u64)req_offset(prev) - segment_offset;
930 r_stripe = (u64)req_offset(req) - segment_offset;
931 do_div(p_stripe, stripe_unit);
932 do_div(r_stripe, stripe_unit);
934 if (p_stripe != r_stripe)
938 /* calculate remaining bytes in the current stripe */
939 div_u64_rem((u64)req_offset(req) - segment_offset,
942 WARN_ON_ONCE(stripe_offset > stripe_unit);
943 if (stripe_offset >= stripe_unit)
945 return min(stripe_unit - (unsigned int)stripe_offset, size);
949 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
950 struct nfs_page *req)
953 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
959 /* If no lseg, fall back to read through mds */
960 if (pgio->pg_lseg == NULL)
961 nfs_pageio_reset_read_mds(pgio);
965 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
966 struct nfs_page *req)
968 struct nfs_commit_info cinfo;
972 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
978 /* If no lseg, fall back to write through mds */
979 if (pgio->pg_lseg == NULL)
981 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
982 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
984 pnfs_put_lseg(pgio->pg_lseg);
985 pgio->pg_lseg = NULL;
990 nfs_pageio_reset_write_mds(pgio);
993 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
994 .pg_init = filelayout_pg_init_read,
995 .pg_test = filelayout_pg_test,
996 .pg_doio = pnfs_generic_pg_readpages,
999 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
1000 .pg_init = filelayout_pg_init_write,
1001 .pg_test = filelayout_pg_test,
1002 .pg_doio = pnfs_generic_pg_writepages,
1005 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1007 if (fl->stripe_type == STRIPE_SPARSE)
1008 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1013 /* The generic layer is about to remove the req from the commit list.
1014 * If this will make the bucket empty, it will need to put the lseg reference.
1015 * Note this is must be called holding the inode (/cinfo) lock
1018 filelayout_clear_request_commit(struct nfs_page *req,
1019 struct nfs_commit_info *cinfo)
1021 struct pnfs_layout_segment *freeme = NULL;
1023 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1025 cinfo->ds->nwritten--;
1026 if (list_is_singular(&req->wb_list)) {
1027 struct pnfs_commit_bucket *bucket;
1029 bucket = list_first_entry(&req->wb_list,
1030 struct pnfs_commit_bucket,
1032 freeme = bucket->wlseg;
1033 bucket->wlseg = NULL;
1036 nfs_request_remove_commit_list(req, cinfo);
1037 pnfs_put_lseg_async(freeme);
1041 filelayout_mark_request_commit(struct nfs_page *req,
1042 struct pnfs_layout_segment *lseg,
1043 struct nfs_commit_info *cinfo)
1046 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1048 struct list_head *list;
1049 struct pnfs_commit_bucket *buckets;
1051 if (fl->commit_through_mds) {
1052 list = &cinfo->mds->list;
1053 spin_lock(cinfo->lock);
1057 /* Note that we are calling nfs4_fl_calc_j_index on each page
1058 * that ends up being committed to a data server. An attractive
1059 * alternative is to add a field to nfs_write_data and nfs_page
1060 * to store the value calculated in filelayout_write_pagelist
1061 * and just use that here.
1063 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1064 i = select_bucket_index(fl, j);
1065 spin_lock(cinfo->lock);
1066 buckets = cinfo->ds->buckets;
1067 list = &buckets[i].written;
1068 if (list_empty(list)) {
1069 /* Non-empty buckets hold a reference on the lseg. That ref
1070 * is normally transferred to the COMMIT call and released
1071 * there. It could also be released if the last req is pulled
1072 * off due to a rewrite, in which case it will be done in
1073 * filelayout_clear_request_commit
1075 buckets[i].wlseg = pnfs_get_lseg(lseg);
1077 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1078 cinfo->ds->nwritten++;
1081 /* nfs_request_add_commit_list(). We need to add req to list without
1082 * dropping cinfo lock.
1084 set_bit(PG_CLEAN, &(req)->wb_flags);
1085 nfs_list_add_request(req, list);
1086 cinfo->mds->ncommit++;
1087 spin_unlock(cinfo->lock);
1089 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1090 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1092 __mark_inode_dirty(req->wb_context->dentry->d_inode,
1097 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1099 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1101 if (flseg->stripe_type == STRIPE_SPARSE)
1104 return nfs4_fl_calc_ds_index(lseg, i);
1107 static struct nfs_fh *
1108 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1110 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1112 if (flseg->stripe_type == STRIPE_SPARSE) {
1113 if (flseg->num_fh == 1)
1115 else if (flseg->num_fh == 0)
1116 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1119 return flseg->fh_array[i];
1122 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1124 struct pnfs_layout_segment *lseg = data->lseg;
1125 struct nfs4_pnfs_ds *ds;
1126 struct rpc_clnt *ds_clnt;
1130 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1131 ds = nfs4_fl_prepare_ds(lseg, idx);
1135 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1136 if (IS_ERR(ds_clnt))
1139 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1140 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1141 data->commit_done_cb = filelayout_commit_done_cb;
1142 atomic_inc(&ds->ds_clp->cl_count);
1143 data->ds_clp = ds->ds_clp;
1144 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1147 return nfs_initiate_commit(ds_clnt, data,
1148 &filelayout_commit_call_ops, how,
1151 prepare_to_resend_writes(data);
1152 filelayout_commit_release(data);
1157 transfer_commit_list(struct list_head *src, struct list_head *dst,
1158 struct nfs_commit_info *cinfo, int max)
1160 struct nfs_page *req, *tmp;
1163 list_for_each_entry_safe(req, tmp, src, wb_list) {
1164 if (!nfs_lock_request(req))
1166 kref_get(&req->wb_kref);
1167 if (cond_resched_lock(cinfo->lock))
1168 list_safe_reset_next(req, tmp, wb_list);
1169 nfs_request_remove_commit_list(req, cinfo);
1170 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1171 nfs_list_add_request(req, dst);
1173 if ((ret == max) && !cinfo->dreq)
1179 /* Note called with cinfo->lock held. */
1181 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1182 struct nfs_commit_info *cinfo,
1185 struct list_head *src = &bucket->written;
1186 struct list_head *dst = &bucket->committing;
1189 ret = transfer_commit_list(src, dst, cinfo, max);
1191 cinfo->ds->nwritten -= ret;
1192 cinfo->ds->ncommitting += ret;
1193 bucket->clseg = bucket->wlseg;
1194 if (list_empty(src))
1195 bucket->wlseg = NULL;
1197 pnfs_get_lseg(bucket->clseg);
1202 /* Move reqs from written to committing lists, returning count of number moved.
1203 * Note called with cinfo->lock held.
1205 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1210 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1211 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1219 /* Pull everything off the committing lists and dump into @dst */
1220 static void filelayout_recover_commit_reqs(struct list_head *dst,
1221 struct nfs_commit_info *cinfo)
1223 struct pnfs_commit_bucket *b;
1224 struct pnfs_layout_segment *freeme;
1228 spin_lock(cinfo->lock);
1229 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1230 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1233 spin_unlock(cinfo->lock);
1234 pnfs_put_lseg(freeme);
1238 cinfo->ds->nwritten = 0;
1239 spin_unlock(cinfo->lock);
1242 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1244 * @cinfo - commit info for current inode
1245 * @page - page to search for matching head request
1247 * Returns a the head request if one is found, otherwise returns NULL.
1249 static struct nfs_page *
1250 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1252 struct nfs_page *freq, *t;
1253 struct pnfs_commit_bucket *b;
1256 /* Linearly search the commit lists for each bucket until a matching
1257 * request is found */
1258 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1259 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1260 if (freq->wb_page == page)
1261 return freq->wb_head;
1263 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1264 if (freq->wb_page == page)
1265 return freq->wb_head;
1272 static void filelayout_retry_commit(struct nfs_commit_info *cinfo, int idx)
1274 struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
1275 struct pnfs_commit_bucket *bucket;
1276 struct pnfs_layout_segment *freeme;
1279 for (i = idx; i < fl_cinfo->nbuckets; i++) {
1280 bucket = &fl_cinfo->buckets[i];
1281 if (list_empty(&bucket->committing))
1283 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1284 spin_lock(cinfo->lock);
1285 freeme = bucket->clseg;
1286 bucket->clseg = NULL;
1287 spin_unlock(cinfo->lock);
1288 pnfs_put_lseg(freeme);
1293 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1295 struct pnfs_ds_commit_info *fl_cinfo;
1296 struct pnfs_commit_bucket *bucket;
1297 struct nfs_commit_data *data;
1299 unsigned int nreq = 0;
1301 fl_cinfo = cinfo->ds;
1302 bucket = fl_cinfo->buckets;
1303 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1304 if (list_empty(&bucket->committing))
1306 data = nfs_commitdata_alloc();
1309 data->ds_commit_index = i;
1310 spin_lock(cinfo->lock);
1311 data->lseg = bucket->clseg;
1312 bucket->clseg = NULL;
1313 spin_unlock(cinfo->lock);
1314 list_add(&data->pages, list);
1318 /* Clean up on error */
1319 filelayout_retry_commit(cinfo, i);
1320 /* Caller will clean up entries put on list */
1324 /* This follows nfs_commit_list pretty closely */
1326 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1327 int how, struct nfs_commit_info *cinfo)
1329 struct nfs_commit_data *data, *tmp;
1331 unsigned int nreq = 0;
1333 if (!list_empty(mds_pages)) {
1334 data = nfs_commitdata_alloc();
1337 list_add(&data->pages, &list);
1340 nfs_retry_commit(mds_pages, NULL, cinfo);
1341 filelayout_retry_commit(cinfo, 0);
1342 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1347 nreq += alloc_ds_commits(cinfo, &list);
1350 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1354 atomic_add(nreq, &cinfo->mds->rpcs_out);
1356 list_for_each_entry_safe(data, tmp, &list, pages) {
1357 list_del_init(&data->pages);
1359 nfs_init_commit(data, mds_pages, NULL, cinfo);
1360 nfs_initiate_commit(NFS_CLIENT(inode), data,
1361 data->mds_ops, how, 0);
1363 struct pnfs_commit_bucket *buckets;
1365 buckets = cinfo->ds->buckets;
1366 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1367 filelayout_initiate_commit(data, how);
1371 cinfo->ds->ncommitting = 0;
1372 return PNFS_ATTEMPTED;
1376 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1378 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1381 static struct pnfs_layout_hdr *
1382 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1384 struct nfs4_filelayout *flo;
1386 flo = kzalloc(sizeof(*flo), gfp_flags);
1387 return flo != NULL ? &flo->generic_hdr : NULL;
1391 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1393 kfree(FILELAYOUT_FROM_HDR(lo));
1396 static struct pnfs_ds_commit_info *
1397 filelayout_get_ds_info(struct inode *inode)
1399 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1404 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1407 static struct pnfs_layoutdriver_type filelayout_type = {
1408 .id = LAYOUT_NFSV4_1_FILES,
1409 .name = "LAYOUT_NFSV4_1_FILES",
1410 .owner = THIS_MODULE,
1411 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1412 .free_layout_hdr = filelayout_free_layout_hdr,
1413 .alloc_lseg = filelayout_alloc_lseg,
1414 .free_lseg = filelayout_free_lseg,
1415 .pg_read_ops = &filelayout_pg_read_ops,
1416 .pg_write_ops = &filelayout_pg_write_ops,
1417 .get_ds_info = &filelayout_get_ds_info,
1418 .mark_request_commit = filelayout_mark_request_commit,
1419 .clear_request_commit = filelayout_clear_request_commit,
1420 .scan_commit_lists = filelayout_scan_commit_lists,
1421 .recover_commit_reqs = filelayout_recover_commit_reqs,
1422 .search_commit_reqs = filelayout_search_commit_reqs,
1423 .commit_pagelist = filelayout_commit_pagelist,
1424 .read_pagelist = filelayout_read_pagelist,
1425 .write_pagelist = filelayout_write_pagelist,
1426 .free_deviceid_node = filelayout_free_deveiceid_node,
1429 static int __init nfs4filelayout_init(void)
1431 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1433 return pnfs_register_layoutdriver(&filelayout_type);
1436 static void __exit nfs4filelayout_exit(void)
1438 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1440 pnfs_unregister_layoutdriver(&filelayout_type);
1443 MODULE_ALIAS("nfs-layouttype4-1");
1445 module_init(nfs4filelayout_init);
1446 module_exit(nfs4filelayout_exit);