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:
148 case -NFS4ERR_OPENMODE:
151 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
152 goto out_bad_stateid;
153 goto wait_on_recovery;
154 case -NFS4ERR_EXPIRED:
156 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
157 goto out_bad_stateid;
159 nfs4_schedule_lease_recovery(mds_client);
160 goto wait_on_recovery;
161 /* DS session errors */
162 case -NFS4ERR_BADSESSION:
163 case -NFS4ERR_BADSLOT:
164 case -NFS4ERR_BAD_HIGH_SLOT:
165 case -NFS4ERR_DEADSESSION:
166 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
167 case -NFS4ERR_SEQ_FALSE_RETRY:
168 case -NFS4ERR_SEQ_MISORDERED:
169 dprintk("%s ERROR %d, Reset session. Exchangeid "
170 "flags 0x%x\n", __func__, task->tk_status,
171 clp->cl_exchange_flags);
172 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
176 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
178 case -NFS4ERR_RETRY_UNCACHED_REP:
180 /* Invalidate Layout errors */
181 case -NFS4ERR_PNFS_NO_LAYOUT:
182 case -ESTALE: /* mapped NFS4ERR_STALE */
183 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
184 case -EISDIR: /* mapped NFS4ERR_ISDIR */
185 case -NFS4ERR_FHEXPIRED:
186 case -NFS4ERR_WRONG_TYPE:
187 dprintk("%s Invalid layout error %d\n", __func__,
190 * Destroy layout so new i/o will get a new layout.
191 * Layout will not be destroyed until all current lseg
192 * references are put. Mark layout as invalid to resend failed
193 * i/o and all i/o waiting on the slot table to the MDS until
194 * layout is destroyed and a new valid layout is obtained.
196 pnfs_destroy_layout(NFS_I(inode));
197 rpc_wake_up(&tbl->slot_tbl_waitq);
199 /* RPC connection errors */
207 dprintk("%s DS connection error %d\n", __func__,
209 nfs4_mark_deviceid_unavailable(devid);
210 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
211 rpc_wake_up(&tbl->slot_tbl_waitq);
215 dprintk("%s Retry through MDS. Error %d\n", __func__,
217 return -NFS4ERR_RESET_TO_MDS;
223 task->tk_status = -EIO;
226 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
227 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
228 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
232 /* NFS_PROTO call done callback routines */
234 static int filelayout_read_done_cb(struct rpc_task *task,
235 struct nfs_pgio_header *hdr)
239 trace_nfs4_pnfs_read(hdr, task->tk_status);
240 err = filelayout_async_handle_error(task, hdr->args.context->state,
241 hdr->ds_clp, hdr->lseg);
244 case -NFS4ERR_RESET_TO_MDS:
245 filelayout_reset_read(hdr);
246 return task->tk_status;
248 rpc_restart_call_prepare(task);
256 * We reference the rpc_cred of the first WRITE that triggers the need for
257 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
258 * rfc5661 is not clear about which credential should be used.
261 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
264 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
265 hdr->res.verf->committed != NFS_DATA_SYNC)
268 pnfs_set_layoutcommit(hdr);
269 dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
270 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
274 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
276 return filelayout_test_devid_invalid(node) ||
277 nfs4_test_deviceid_unavailable(node);
281 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
283 struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
285 return filelayout_test_devid_unavailable(node);
289 * Call ops for the async read/write cases
290 * In the case of dense layouts, the offset needs to be reset to its
293 static void filelayout_read_prepare(struct rpc_task *task, void *data)
295 struct nfs_pgio_header *hdr = data;
297 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
298 rpc_exit(task, -EIO);
301 if (filelayout_reset_to_mds(hdr->lseg)) {
302 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
303 filelayout_reset_read(hdr);
307 hdr->pgio_done_cb = filelayout_read_done_cb;
309 if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
314 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
315 hdr->args.lock_context, FMODE_READ) == -EIO)
316 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
319 static void filelayout_read_call_done(struct rpc_task *task, void *data)
321 struct nfs_pgio_header *hdr = data;
323 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
325 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
326 task->tk_status == 0) {
327 nfs41_sequence_done(task, &hdr->res.seq_res);
331 /* Note this may cause RPC to be resent */
332 hdr->mds_ops->rpc_call_done(task, data);
335 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
337 struct nfs_pgio_header *hdr = data;
339 rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
342 static void filelayout_read_release(void *data)
344 struct nfs_pgio_header *hdr = data;
345 struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
347 filelayout_fenceme(lo->plh_inode, lo);
348 nfs_put_client(hdr->ds_clp);
349 hdr->mds_ops->rpc_release(data);
352 static int filelayout_write_done_cb(struct rpc_task *task,
353 struct nfs_pgio_header *hdr)
357 trace_nfs4_pnfs_write(hdr, task->tk_status);
358 err = filelayout_async_handle_error(task, hdr->args.context->state,
359 hdr->ds_clp, hdr->lseg);
362 case -NFS4ERR_RESET_TO_MDS:
363 filelayout_reset_write(hdr);
364 return task->tk_status;
366 rpc_restart_call_prepare(task);
370 filelayout_set_layoutcommit(hdr);
374 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
375 static void prepare_to_resend_writes(struct nfs_commit_data *data)
377 struct nfs_page *first = nfs_list_entry(data->pages.next);
379 data->task.tk_status = 0;
380 memcpy(&data->verf.verifier, &first->wb_verf,
381 sizeof(data->verf.verifier));
382 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
385 static int filelayout_commit_done_cb(struct rpc_task *task,
386 struct nfs_commit_data *data)
390 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
391 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
395 case -NFS4ERR_RESET_TO_MDS:
396 prepare_to_resend_writes(data);
399 rpc_restart_call_prepare(task);
403 if (data->verf.committed == NFS_UNSTABLE)
404 pnfs_commit_set_layoutcommit(data);
409 static void filelayout_write_prepare(struct rpc_task *task, void *data)
411 struct nfs_pgio_header *hdr = data;
413 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
414 rpc_exit(task, -EIO);
417 if (filelayout_reset_to_mds(hdr->lseg)) {
418 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
419 filelayout_reset_write(hdr);
423 if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
428 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
429 hdr->args.lock_context, FMODE_WRITE) == -EIO)
430 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
433 static void filelayout_write_call_done(struct rpc_task *task, void *data)
435 struct nfs_pgio_header *hdr = data;
437 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
438 task->tk_status == 0) {
439 nfs41_sequence_done(task, &hdr->res.seq_res);
443 /* Note this may cause RPC to be resent */
444 hdr->mds_ops->rpc_call_done(task, data);
447 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
449 struct nfs_pgio_header *hdr = data;
451 rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
454 static void filelayout_write_release(void *data)
456 struct nfs_pgio_header *hdr = data;
457 struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
459 filelayout_fenceme(lo->plh_inode, lo);
460 nfs_put_client(hdr->ds_clp);
461 hdr->mds_ops->rpc_release(data);
464 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
466 struct nfs_commit_data *wdata = data;
468 nfs41_setup_sequence(wdata->ds_clp->cl_session,
469 &wdata->args.seq_args,
474 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
476 struct nfs_commit_data *wdata = data;
478 /* Note this may cause RPC to be resent */
479 wdata->mds_ops->rpc_call_done(task, data);
482 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
484 struct nfs_commit_data *cdata = data;
486 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
489 static void filelayout_commit_release(void *calldata)
491 struct nfs_commit_data *data = calldata;
493 data->completion_ops->completion(data);
494 pnfs_put_lseg(data->lseg);
495 nfs_put_client(data->ds_clp);
496 nfs_commitdata_release(data);
499 static const struct rpc_call_ops filelayout_read_call_ops = {
500 .rpc_call_prepare = filelayout_read_prepare,
501 .rpc_call_done = filelayout_read_call_done,
502 .rpc_count_stats = filelayout_read_count_stats,
503 .rpc_release = filelayout_read_release,
506 static const struct rpc_call_ops filelayout_write_call_ops = {
507 .rpc_call_prepare = filelayout_write_prepare,
508 .rpc_call_done = filelayout_write_call_done,
509 .rpc_count_stats = filelayout_write_count_stats,
510 .rpc_release = filelayout_write_release,
513 static const struct rpc_call_ops filelayout_commit_call_ops = {
514 .rpc_call_prepare = filelayout_commit_prepare,
515 .rpc_call_done = filelayout_write_commit_done,
516 .rpc_count_stats = filelayout_commit_count_stats,
517 .rpc_release = filelayout_commit_release,
520 static enum pnfs_try_status
521 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
523 struct pnfs_layout_segment *lseg = hdr->lseg;
524 struct nfs4_pnfs_ds *ds;
525 struct rpc_clnt *ds_clnt;
526 loff_t offset = hdr->args.offset;
530 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
531 __func__, hdr->inode->i_ino,
532 hdr->args.pgbase, (size_t)hdr->args.count, offset);
534 /* Retrieve the correct rpc_client for the byte range */
535 j = nfs4_fl_calc_j_index(lseg, offset);
536 idx = nfs4_fl_calc_ds_index(lseg, j);
537 ds = nfs4_fl_prepare_ds(lseg, idx);
539 return PNFS_NOT_ATTEMPTED;
541 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
543 return PNFS_NOT_ATTEMPTED;
545 dprintk("%s USE DS: %s cl_count %d\n", __func__,
546 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
548 /* No multipath support. Use first DS */
549 atomic_inc(&ds->ds_clp->cl_count);
550 hdr->ds_clp = ds->ds_clp;
552 fh = nfs4_fl_select_ds_fh(lseg, j);
556 hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
557 hdr->mds_offset = offset;
559 /* Perform an asynchronous read to ds */
560 nfs_initiate_pgio(ds_clnt, hdr,
561 &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
562 return PNFS_ATTEMPTED;
565 /* Perform async writes. */
566 static enum pnfs_try_status
567 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
569 struct pnfs_layout_segment *lseg = hdr->lseg;
570 struct nfs4_pnfs_ds *ds;
571 struct rpc_clnt *ds_clnt;
572 loff_t offset = hdr->args.offset;
576 /* Retrieve the correct rpc_client for the byte range */
577 j = nfs4_fl_calc_j_index(lseg, offset);
578 idx = nfs4_fl_calc_ds_index(lseg, j);
579 ds = nfs4_fl_prepare_ds(lseg, idx);
581 return PNFS_NOT_ATTEMPTED;
583 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
585 return PNFS_NOT_ATTEMPTED;
587 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
588 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
589 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
591 hdr->pgio_done_cb = filelayout_write_done_cb;
592 atomic_inc(&ds->ds_clp->cl_count);
593 hdr->ds_clp = ds->ds_clp;
595 fh = nfs4_fl_select_ds_fh(lseg, j);
598 hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
600 /* Perform an asynchronous write */
601 nfs_initiate_pgio(ds_clnt, hdr,
602 &filelayout_write_call_ops, sync,
604 return PNFS_ATTEMPTED;
608 * filelayout_check_layout()
610 * Make sure layout segment parameters are sane WRT the device.
611 * At this point no generic layer initialization of the lseg has occurred,
612 * and nothing has been added to the layout_hdr cache.
616 filelayout_check_layout(struct pnfs_layout_hdr *lo,
617 struct nfs4_filelayout_segment *fl,
618 struct nfs4_layoutget_res *lgr,
619 struct nfs4_deviceid *id,
622 struct nfs4_deviceid_node *d;
623 struct nfs4_file_layout_dsaddr *dsaddr;
624 int status = -EINVAL;
626 dprintk("--> %s\n", __func__);
628 /* FIXME: remove this check when layout segment support is added */
629 if (lgr->range.offset != 0 ||
630 lgr->range.length != NFS4_MAX_UINT64) {
631 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
636 if (fl->pattern_offset > lgr->range.offset) {
637 dprintk("%s pattern_offset %lld too large\n",
638 __func__, fl->pattern_offset);
642 if (!fl->stripe_unit) {
643 dprintk("%s Invalid stripe unit (%u)\n",
644 __func__, fl->stripe_unit);
648 /* find and reference the deviceid */
649 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
650 lo->plh_lc_cred, gfp_flags);
654 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
655 /* Found deviceid is unavailable */
656 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
661 if (fl->first_stripe_index >= dsaddr->stripe_count) {
662 dprintk("%s Bad first_stripe_index %u\n",
663 __func__, fl->first_stripe_index);
667 if ((fl->stripe_type == STRIPE_SPARSE &&
668 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
669 (fl->stripe_type == STRIPE_DENSE &&
670 fl->num_fh != dsaddr->stripe_count)) {
671 dprintk("%s num_fh %u not valid for given packing\n",
672 __func__, fl->num_fh);
678 dprintk("--> %s returns %d\n", __func__, status);
681 nfs4_fl_put_deviceid(dsaddr);
685 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
689 for (i = 0; i < fl->num_fh; i++) {
690 if (!fl->fh_array[i])
692 kfree(fl->fh_array[i]);
699 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
701 filelayout_free_fh_array(fl);
706 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
707 struct nfs4_filelayout_segment *fl,
708 struct nfs4_layoutget_res *lgr,
709 struct nfs4_deviceid *id,
712 struct xdr_stream stream;
714 struct page *scratch;
719 dprintk("%s: set_layout_map Begin\n", __func__);
721 scratch = alloc_page(gfp_flags);
725 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
726 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
728 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
730 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
734 memcpy(id, p, sizeof(*id));
735 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
736 nfs4_print_deviceid(id);
738 nfl_util = be32_to_cpup(p++);
739 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
740 fl->commit_through_mds = 1;
741 if (nfl_util & NFL4_UFLG_DENSE)
742 fl->stripe_type = STRIPE_DENSE;
744 fl->stripe_type = STRIPE_SPARSE;
745 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
747 fl->first_stripe_index = be32_to_cpup(p++);
748 p = xdr_decode_hyper(p, &fl->pattern_offset);
749 fl->num_fh = be32_to_cpup(p++);
751 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
752 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
755 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
756 * Futher checking is done in filelayout_check_layout */
758 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
761 if (fl->num_fh > 0) {
762 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
768 for (i = 0; i < fl->num_fh; i++) {
769 /* Do we want to use a mempool here? */
770 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
771 if (!fl->fh_array[i])
774 p = xdr_inline_decode(&stream, 4);
777 fl->fh_array[i]->size = be32_to_cpup(p++);
778 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
779 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
780 i, fl->fh_array[i]->size);
784 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
787 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
788 dprintk("DEBUG: %s: fh len %d\n", __func__,
789 fl->fh_array[i]->size);
792 __free_page(scratch);
796 filelayout_free_fh_array(fl);
798 __free_page(scratch);
803 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
805 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
807 dprintk("--> %s\n", __func__);
808 nfs4_fl_put_deviceid(fl->dsaddr);
809 /* This assumes a single RW lseg */
810 if (lseg->pls_range.iomode == IOMODE_RW) {
811 struct nfs4_filelayout *flo;
813 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
814 flo->commit_info.nbuckets = 0;
815 kfree(flo->commit_info.buckets);
816 flo->commit_info.buckets = NULL;
818 _filelayout_free_lseg(fl);
822 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
823 struct nfs_commit_info *cinfo,
826 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
827 struct pnfs_commit_bucket *buckets;
830 if (fl->commit_through_mds)
833 size = (fl->stripe_type == STRIPE_SPARSE) ?
834 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
836 if (cinfo->ds->nbuckets >= size) {
837 /* This assumes there is only one IOMODE_RW lseg. What
838 * we really want to do is have a layout_hdr level
839 * dictionary of <multipath_list4, fh> keys, each
840 * associated with a struct list_head, populated by calls
841 * to filelayout_write_pagelist().
846 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
850 for (i = 0; i < size; i++) {
851 INIT_LIST_HEAD(&buckets[i].written);
852 INIT_LIST_HEAD(&buckets[i].committing);
853 /* mark direct verifier as unset */
854 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
857 spin_lock(cinfo->lock);
858 if (cinfo->ds->nbuckets >= size)
860 for (i = 0; i < cinfo->ds->nbuckets; i++) {
861 list_splice(&cinfo->ds->buckets[i].written,
862 &buckets[i].written);
863 list_splice(&cinfo->ds->buckets[i].committing,
864 &buckets[i].committing);
865 buckets[i].direct_verf.committed =
866 cinfo->ds->buckets[i].direct_verf.committed;
867 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
868 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
870 swap(cinfo->ds->buckets, buckets);
871 cinfo->ds->nbuckets = size;
873 spin_unlock(cinfo->lock);
878 static struct pnfs_layout_segment *
879 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
880 struct nfs4_layoutget_res *lgr,
883 struct nfs4_filelayout_segment *fl;
885 struct nfs4_deviceid id;
887 dprintk("--> %s\n", __func__);
888 fl = kzalloc(sizeof(*fl), gfp_flags);
892 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
893 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
894 _filelayout_free_lseg(fl);
897 return &fl->generic_hdr;
901 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
903 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
904 * of bytes (maximum @req->wb_bytes) that can be coalesced.
907 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
908 struct nfs_page *req)
911 u64 p_stripe, r_stripe;
913 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
914 u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
916 /* calls nfs_generic_pg_test */
917 size = pnfs_generic_pg_test(pgio, prev, req);
921 /* see if req and prev are in the same stripe */
923 p_stripe = (u64)req_offset(prev) - segment_offset;
924 r_stripe = (u64)req_offset(req) - segment_offset;
925 do_div(p_stripe, stripe_unit);
926 do_div(r_stripe, stripe_unit);
928 if (p_stripe != r_stripe)
932 /* calculate remaining bytes in the current stripe */
933 div_u64_rem((u64)req_offset(req) - segment_offset,
936 WARN_ON_ONCE(stripe_offset > stripe_unit);
937 if (stripe_offset >= stripe_unit)
939 return min(stripe_unit - (unsigned int)stripe_offset, size);
943 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
944 struct nfs_page *req)
947 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
953 /* If no lseg, fall back to read through mds */
954 if (pgio->pg_lseg == NULL)
955 nfs_pageio_reset_read_mds(pgio);
959 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
960 struct nfs_page *req)
962 struct nfs_commit_info cinfo;
966 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
972 /* If no lseg, fall back to write through mds */
973 if (pgio->pg_lseg == NULL)
975 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
976 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
978 pnfs_put_lseg(pgio->pg_lseg);
979 pgio->pg_lseg = NULL;
984 nfs_pageio_reset_write_mds(pgio);
987 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
988 .pg_init = filelayout_pg_init_read,
989 .pg_test = filelayout_pg_test,
990 .pg_doio = pnfs_generic_pg_readpages,
993 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
994 .pg_init = filelayout_pg_init_write,
995 .pg_test = filelayout_pg_test,
996 .pg_doio = pnfs_generic_pg_writepages,
999 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1001 if (fl->stripe_type == STRIPE_SPARSE)
1002 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1007 /* The generic layer is about to remove the req from the commit list.
1008 * If this will make the bucket empty, it will need to put the lseg reference.
1009 * Note this is must be called holding the inode (/cinfo) lock
1012 filelayout_clear_request_commit(struct nfs_page *req,
1013 struct nfs_commit_info *cinfo)
1015 struct pnfs_layout_segment *freeme = NULL;
1017 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1019 cinfo->ds->nwritten--;
1020 if (list_is_singular(&req->wb_list)) {
1021 struct pnfs_commit_bucket *bucket;
1023 bucket = list_first_entry(&req->wb_list,
1024 struct pnfs_commit_bucket,
1026 freeme = bucket->wlseg;
1027 bucket->wlseg = NULL;
1030 nfs_request_remove_commit_list(req, cinfo);
1031 pnfs_put_lseg_locked(freeme);
1035 filelayout_mark_request_commit(struct nfs_page *req,
1036 struct pnfs_layout_segment *lseg,
1037 struct nfs_commit_info *cinfo)
1040 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1042 struct list_head *list;
1043 struct pnfs_commit_bucket *buckets;
1045 if (fl->commit_through_mds) {
1046 list = &cinfo->mds->list;
1047 spin_lock(cinfo->lock);
1051 /* Note that we are calling nfs4_fl_calc_j_index on each page
1052 * that ends up being committed to a data server. An attractive
1053 * alternative is to add a field to nfs_write_data and nfs_page
1054 * to store the value calculated in filelayout_write_pagelist
1055 * and just use that here.
1057 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1058 i = select_bucket_index(fl, j);
1059 spin_lock(cinfo->lock);
1060 buckets = cinfo->ds->buckets;
1061 list = &buckets[i].written;
1062 if (list_empty(list)) {
1063 /* Non-empty buckets hold a reference on the lseg. That ref
1064 * is normally transferred to the COMMIT call and released
1065 * there. It could also be released if the last req is pulled
1066 * off due to a rewrite, in which case it will be done in
1067 * filelayout_clear_request_commit
1069 buckets[i].wlseg = pnfs_get_lseg(lseg);
1071 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1072 cinfo->ds->nwritten++;
1075 /* nfs_request_add_commit_list(). We need to add req to list without
1076 * dropping cinfo lock.
1078 set_bit(PG_CLEAN, &(req)->wb_flags);
1079 nfs_list_add_request(req, list);
1080 cinfo->mds->ncommit++;
1081 spin_unlock(cinfo->lock);
1083 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1084 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1086 __mark_inode_dirty(req->wb_context->dentry->d_inode,
1091 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1093 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1095 if (flseg->stripe_type == STRIPE_SPARSE)
1098 return nfs4_fl_calc_ds_index(lseg, i);
1101 static struct nfs_fh *
1102 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1104 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1106 if (flseg->stripe_type == STRIPE_SPARSE) {
1107 if (flseg->num_fh == 1)
1109 else if (flseg->num_fh == 0)
1110 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1113 return flseg->fh_array[i];
1116 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1118 struct pnfs_layout_segment *lseg = data->lseg;
1119 struct nfs4_pnfs_ds *ds;
1120 struct rpc_clnt *ds_clnt;
1124 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1125 ds = nfs4_fl_prepare_ds(lseg, idx);
1129 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1130 if (IS_ERR(ds_clnt))
1133 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1134 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1135 data->commit_done_cb = filelayout_commit_done_cb;
1136 atomic_inc(&ds->ds_clp->cl_count);
1137 data->ds_clp = ds->ds_clp;
1138 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1141 return nfs_initiate_commit(ds_clnt, data,
1142 &filelayout_commit_call_ops, how,
1145 prepare_to_resend_writes(data);
1146 filelayout_commit_release(data);
1151 transfer_commit_list(struct list_head *src, struct list_head *dst,
1152 struct nfs_commit_info *cinfo, int max)
1154 struct nfs_page *req, *tmp;
1157 list_for_each_entry_safe(req, tmp, src, wb_list) {
1158 if (!nfs_lock_request(req))
1160 kref_get(&req->wb_kref);
1161 if (cond_resched_lock(cinfo->lock))
1162 list_safe_reset_next(req, tmp, wb_list);
1163 nfs_request_remove_commit_list(req, cinfo);
1164 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1165 nfs_list_add_request(req, dst);
1167 if ((ret == max) && !cinfo->dreq)
1173 /* Note called with cinfo->lock held. */
1175 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1176 struct nfs_commit_info *cinfo,
1179 struct list_head *src = &bucket->written;
1180 struct list_head *dst = &bucket->committing;
1183 ret = transfer_commit_list(src, dst, cinfo, max);
1185 cinfo->ds->nwritten -= ret;
1186 cinfo->ds->ncommitting += ret;
1187 bucket->clseg = bucket->wlseg;
1188 if (list_empty(src))
1189 bucket->wlseg = NULL;
1191 pnfs_get_lseg(bucket->clseg);
1196 /* Move reqs from written to committing lists, returning count of number moved.
1197 * Note called with cinfo->lock held.
1199 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1204 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1205 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1213 /* Pull everything off the committing lists and dump into @dst */
1214 static void filelayout_recover_commit_reqs(struct list_head *dst,
1215 struct nfs_commit_info *cinfo)
1217 struct pnfs_commit_bucket *b;
1218 struct pnfs_layout_segment *freeme;
1222 spin_lock(cinfo->lock);
1223 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1224 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1227 spin_unlock(cinfo->lock);
1228 pnfs_put_lseg(freeme);
1232 cinfo->ds->nwritten = 0;
1233 spin_unlock(cinfo->lock);
1236 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1238 * @cinfo - commit info for current inode
1239 * @page - page to search for matching head request
1241 * Returns a the head request if one is found, otherwise returns NULL.
1243 static struct nfs_page *
1244 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1246 struct nfs_page *freq, *t;
1247 struct pnfs_commit_bucket *b;
1250 /* Linearly search the commit lists for each bucket until a matching
1251 * request is found */
1252 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1253 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1254 if (freq->wb_page == page)
1255 return freq->wb_head;
1257 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1258 if (freq->wb_page == page)
1259 return freq->wb_head;
1266 static void filelayout_retry_commit(struct nfs_commit_info *cinfo, int idx)
1268 struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
1269 struct pnfs_commit_bucket *bucket;
1270 struct pnfs_layout_segment *freeme;
1273 for (i = idx; i < fl_cinfo->nbuckets; i++) {
1274 bucket = &fl_cinfo->buckets[i];
1275 if (list_empty(&bucket->committing))
1277 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1278 spin_lock(cinfo->lock);
1279 freeme = bucket->clseg;
1280 bucket->clseg = NULL;
1281 spin_unlock(cinfo->lock);
1282 pnfs_put_lseg(freeme);
1287 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1289 struct pnfs_ds_commit_info *fl_cinfo;
1290 struct pnfs_commit_bucket *bucket;
1291 struct nfs_commit_data *data;
1293 unsigned int nreq = 0;
1295 fl_cinfo = cinfo->ds;
1296 bucket = fl_cinfo->buckets;
1297 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1298 if (list_empty(&bucket->committing))
1300 data = nfs_commitdata_alloc();
1303 data->ds_commit_index = i;
1304 spin_lock(cinfo->lock);
1305 data->lseg = bucket->clseg;
1306 bucket->clseg = NULL;
1307 spin_unlock(cinfo->lock);
1308 list_add(&data->pages, list);
1312 /* Clean up on error */
1313 filelayout_retry_commit(cinfo, i);
1314 /* Caller will clean up entries put on list */
1318 /* This follows nfs_commit_list pretty closely */
1320 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1321 int how, struct nfs_commit_info *cinfo)
1323 struct nfs_commit_data *data, *tmp;
1325 unsigned int nreq = 0;
1327 if (!list_empty(mds_pages)) {
1328 data = nfs_commitdata_alloc();
1331 list_add(&data->pages, &list);
1334 nfs_retry_commit(mds_pages, NULL, cinfo);
1335 filelayout_retry_commit(cinfo, 0);
1336 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1341 nreq += alloc_ds_commits(cinfo, &list);
1344 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1348 atomic_add(nreq, &cinfo->mds->rpcs_out);
1350 list_for_each_entry_safe(data, tmp, &list, pages) {
1351 list_del_init(&data->pages);
1353 nfs_init_commit(data, mds_pages, NULL, cinfo);
1354 nfs_initiate_commit(NFS_CLIENT(inode), data,
1355 data->mds_ops, how, 0);
1357 struct pnfs_commit_bucket *buckets;
1359 buckets = cinfo->ds->buckets;
1360 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1361 filelayout_initiate_commit(data, how);
1365 cinfo->ds->ncommitting = 0;
1366 return PNFS_ATTEMPTED;
1368 static struct nfs4_deviceid_node *
1369 filelayout_alloc_deviceid_node(struct nfs_server *server,
1370 struct pnfs_device *pdev, gfp_t gfp_flags)
1372 struct nfs4_file_layout_dsaddr *dsaddr;
1374 dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1377 return &dsaddr->id_node;
1381 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1383 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1386 static struct pnfs_layout_hdr *
1387 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1389 struct nfs4_filelayout *flo;
1391 flo = kzalloc(sizeof(*flo), gfp_flags);
1392 return flo != NULL ? &flo->generic_hdr : NULL;
1396 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1398 kfree(FILELAYOUT_FROM_HDR(lo));
1401 static struct pnfs_ds_commit_info *
1402 filelayout_get_ds_info(struct inode *inode)
1404 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1409 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1412 static struct pnfs_layoutdriver_type filelayout_type = {
1413 .id = LAYOUT_NFSV4_1_FILES,
1414 .name = "LAYOUT_NFSV4_1_FILES",
1415 .owner = THIS_MODULE,
1416 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1417 .free_layout_hdr = filelayout_free_layout_hdr,
1418 .alloc_lseg = filelayout_alloc_lseg,
1419 .free_lseg = filelayout_free_lseg,
1420 .pg_read_ops = &filelayout_pg_read_ops,
1421 .pg_write_ops = &filelayout_pg_write_ops,
1422 .get_ds_info = &filelayout_get_ds_info,
1423 .mark_request_commit = filelayout_mark_request_commit,
1424 .clear_request_commit = filelayout_clear_request_commit,
1425 .scan_commit_lists = filelayout_scan_commit_lists,
1426 .recover_commit_reqs = filelayout_recover_commit_reqs,
1427 .search_commit_reqs = filelayout_search_commit_reqs,
1428 .commit_pagelist = filelayout_commit_pagelist,
1429 .read_pagelist = filelayout_read_pagelist,
1430 .write_pagelist = filelayout_write_pagelist,
1431 .alloc_deviceid_node = filelayout_alloc_deviceid_node,
1432 .free_deviceid_node = filelayout_free_deveiceid_node,
1435 static int __init nfs4filelayout_init(void)
1437 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1439 return pnfs_register_layoutdriver(&filelayout_type);
1442 static void __exit nfs4filelayout_exit(void)
1444 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1446 pnfs_unregister_layoutdriver(&filelayout_type);
1449 MODULE_ALIAS("nfs-layouttype4-1");
1451 module_init(nfs4filelayout_init);
1452 module_exit(nfs4filelayout_exit);