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_FILE_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);
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)->pnfs_curr_ld,
650 NFS_SERVER(lo->plh_inode)->nfs_client, id);
652 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
653 lo->plh_lc_cred, gfp_flags);
657 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
658 /* Found deviceid is unavailable */
659 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
664 if (fl->first_stripe_index >= dsaddr->stripe_count) {
665 dprintk("%s Bad first_stripe_index %u\n",
666 __func__, fl->first_stripe_index);
670 if ((fl->stripe_type == STRIPE_SPARSE &&
671 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
672 (fl->stripe_type == STRIPE_DENSE &&
673 fl->num_fh != dsaddr->stripe_count)) {
674 dprintk("%s num_fh %u not valid for given packing\n",
675 __func__, fl->num_fh);
681 dprintk("--> %s returns %d\n", __func__, status);
684 nfs4_fl_put_deviceid(dsaddr);
688 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
692 for (i = 0; i < fl->num_fh; i++) {
693 if (!fl->fh_array[i])
695 kfree(fl->fh_array[i]);
702 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
704 filelayout_free_fh_array(fl);
709 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
710 struct nfs4_filelayout_segment *fl,
711 struct nfs4_layoutget_res *lgr,
712 struct nfs4_deviceid *id,
715 struct xdr_stream stream;
717 struct page *scratch;
722 dprintk("%s: set_layout_map Begin\n", __func__);
724 scratch = alloc_page(gfp_flags);
728 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
729 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
731 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
733 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
737 memcpy(id, p, sizeof(*id));
738 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
739 nfs4_print_deviceid(id);
741 nfl_util = be32_to_cpup(p++);
742 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
743 fl->commit_through_mds = 1;
744 if (nfl_util & NFL4_UFLG_DENSE)
745 fl->stripe_type = STRIPE_DENSE;
747 fl->stripe_type = STRIPE_SPARSE;
748 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
750 fl->first_stripe_index = be32_to_cpup(p++);
751 p = xdr_decode_hyper(p, &fl->pattern_offset);
752 fl->num_fh = be32_to_cpup(p++);
754 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
755 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
758 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
759 * Futher checking is done in filelayout_check_layout */
761 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
764 if (fl->num_fh > 0) {
765 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
771 for (i = 0; i < fl->num_fh; i++) {
772 /* Do we want to use a mempool here? */
773 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
774 if (!fl->fh_array[i])
777 p = xdr_inline_decode(&stream, 4);
780 fl->fh_array[i]->size = be32_to_cpup(p++);
781 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
782 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
783 i, fl->fh_array[i]->size);
787 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
790 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
791 dprintk("DEBUG: %s: fh len %d\n", __func__,
792 fl->fh_array[i]->size);
795 __free_page(scratch);
799 filelayout_free_fh_array(fl);
801 __free_page(scratch);
806 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
808 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
810 dprintk("--> %s\n", __func__);
811 nfs4_fl_put_deviceid(fl->dsaddr);
812 /* This assumes a single RW lseg */
813 if (lseg->pls_range.iomode == IOMODE_RW) {
814 struct nfs4_filelayout *flo;
816 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
817 flo->commit_info.nbuckets = 0;
818 kfree(flo->commit_info.buckets);
819 flo->commit_info.buckets = NULL;
821 _filelayout_free_lseg(fl);
825 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
826 struct nfs_commit_info *cinfo,
829 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
830 struct pnfs_commit_bucket *buckets;
833 if (fl->commit_through_mds)
836 size = (fl->stripe_type == STRIPE_SPARSE) ?
837 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
839 if (cinfo->ds->nbuckets >= size) {
840 /* This assumes there is only one IOMODE_RW lseg. What
841 * we really want to do is have a layout_hdr level
842 * dictionary of <multipath_list4, fh> keys, each
843 * associated with a struct list_head, populated by calls
844 * to filelayout_write_pagelist().
849 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
853 for (i = 0; i < size; i++) {
854 INIT_LIST_HEAD(&buckets[i].written);
855 INIT_LIST_HEAD(&buckets[i].committing);
856 /* mark direct verifier as unset */
857 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
860 spin_lock(cinfo->lock);
861 if (cinfo->ds->nbuckets >= size)
863 for (i = 0; i < cinfo->ds->nbuckets; i++) {
864 list_splice(&cinfo->ds->buckets[i].written,
865 &buckets[i].written);
866 list_splice(&cinfo->ds->buckets[i].committing,
867 &buckets[i].committing);
868 buckets[i].direct_verf.committed =
869 cinfo->ds->buckets[i].direct_verf.committed;
870 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
871 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
873 swap(cinfo->ds->buckets, buckets);
874 cinfo->ds->nbuckets = size;
876 spin_unlock(cinfo->lock);
881 static struct pnfs_layout_segment *
882 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
883 struct nfs4_layoutget_res *lgr,
886 struct nfs4_filelayout_segment *fl;
888 struct nfs4_deviceid id;
890 dprintk("--> %s\n", __func__);
891 fl = kzalloc(sizeof(*fl), gfp_flags);
895 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
896 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
897 _filelayout_free_lseg(fl);
900 return &fl->generic_hdr;
904 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
906 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
907 * of bytes (maximum @req->wb_bytes) that can be coalesced.
910 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
911 struct nfs_page *req)
914 u64 p_stripe, r_stripe;
916 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
917 u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
919 /* calls nfs_generic_pg_test */
920 size = pnfs_generic_pg_test(pgio, prev, req);
924 /* see if req and prev are in the same stripe */
926 p_stripe = (u64)req_offset(prev) - segment_offset;
927 r_stripe = (u64)req_offset(req) - segment_offset;
928 do_div(p_stripe, stripe_unit);
929 do_div(r_stripe, stripe_unit);
931 if (p_stripe != r_stripe)
935 /* calculate remaining bytes in the current stripe */
936 div_u64_rem((u64)req_offset(req) - segment_offset,
939 WARN_ON_ONCE(stripe_offset > stripe_unit);
940 if (stripe_offset >= stripe_unit)
942 return min(stripe_unit - (unsigned int)stripe_offset, size);
946 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
947 struct nfs_page *req)
950 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
956 /* If no lseg, fall back to read through mds */
957 if (pgio->pg_lseg == NULL)
958 nfs_pageio_reset_read_mds(pgio);
962 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
963 struct nfs_page *req)
965 struct nfs_commit_info cinfo;
969 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
975 /* If no lseg, fall back to write through mds */
976 if (pgio->pg_lseg == NULL)
978 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
979 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
981 pnfs_put_lseg(pgio->pg_lseg);
982 pgio->pg_lseg = NULL;
987 nfs_pageio_reset_write_mds(pgio);
990 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
991 .pg_init = filelayout_pg_init_read,
992 .pg_test = filelayout_pg_test,
993 .pg_doio = pnfs_generic_pg_readpages,
996 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
997 .pg_init = filelayout_pg_init_write,
998 .pg_test = filelayout_pg_test,
999 .pg_doio = pnfs_generic_pg_writepages,
1002 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1004 if (fl->stripe_type == STRIPE_SPARSE)
1005 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1010 /* The generic layer is about to remove the req from the commit list.
1011 * If this will make the bucket empty, it will need to put the lseg reference.
1012 * Note this is must be called holding the inode (/cinfo) lock
1015 filelayout_clear_request_commit(struct nfs_page *req,
1016 struct nfs_commit_info *cinfo)
1018 struct pnfs_layout_segment *freeme = NULL;
1020 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1022 cinfo->ds->nwritten--;
1023 if (list_is_singular(&req->wb_list)) {
1024 struct pnfs_commit_bucket *bucket;
1026 bucket = list_first_entry(&req->wb_list,
1027 struct pnfs_commit_bucket,
1029 freeme = bucket->wlseg;
1030 bucket->wlseg = NULL;
1033 nfs_request_remove_commit_list(req, cinfo);
1034 pnfs_put_lseg_async(freeme);
1038 filelayout_mark_request_commit(struct nfs_page *req,
1039 struct pnfs_layout_segment *lseg,
1040 struct nfs_commit_info *cinfo)
1043 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1045 struct list_head *list;
1046 struct pnfs_commit_bucket *buckets;
1048 if (fl->commit_through_mds) {
1049 list = &cinfo->mds->list;
1050 spin_lock(cinfo->lock);
1054 /* Note that we are calling nfs4_fl_calc_j_index on each page
1055 * that ends up being committed to a data server. An attractive
1056 * alternative is to add a field to nfs_write_data and nfs_page
1057 * to store the value calculated in filelayout_write_pagelist
1058 * and just use that here.
1060 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1061 i = select_bucket_index(fl, j);
1062 spin_lock(cinfo->lock);
1063 buckets = cinfo->ds->buckets;
1064 list = &buckets[i].written;
1065 if (list_empty(list)) {
1066 /* Non-empty buckets hold a reference on the lseg. That ref
1067 * is normally transferred to the COMMIT call and released
1068 * there. It could also be released if the last req is pulled
1069 * off due to a rewrite, in which case it will be done in
1070 * filelayout_clear_request_commit
1072 buckets[i].wlseg = pnfs_get_lseg(lseg);
1074 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1075 cinfo->ds->nwritten++;
1078 /* nfs_request_add_commit_list(). We need to add req to list without
1079 * dropping cinfo lock.
1081 set_bit(PG_CLEAN, &(req)->wb_flags);
1082 nfs_list_add_request(req, list);
1083 cinfo->mds->ncommit++;
1084 spin_unlock(cinfo->lock);
1086 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1087 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1089 __mark_inode_dirty(req->wb_context->dentry->d_inode,
1094 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1096 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1098 if (flseg->stripe_type == STRIPE_SPARSE)
1101 return nfs4_fl_calc_ds_index(lseg, i);
1104 static struct nfs_fh *
1105 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1107 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1109 if (flseg->stripe_type == STRIPE_SPARSE) {
1110 if (flseg->num_fh == 1)
1112 else if (flseg->num_fh == 0)
1113 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1116 return flseg->fh_array[i];
1119 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1121 struct pnfs_layout_segment *lseg = data->lseg;
1122 struct nfs4_pnfs_ds *ds;
1123 struct rpc_clnt *ds_clnt;
1127 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1128 ds = nfs4_fl_prepare_ds(lseg, idx);
1132 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1133 if (IS_ERR(ds_clnt))
1136 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1137 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1138 data->commit_done_cb = filelayout_commit_done_cb;
1139 atomic_inc(&ds->ds_clp->cl_count);
1140 data->ds_clp = ds->ds_clp;
1141 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1144 return nfs_initiate_commit(ds_clnt, data,
1145 &filelayout_commit_call_ops, how,
1148 prepare_to_resend_writes(data);
1149 filelayout_commit_release(data);
1154 transfer_commit_list(struct list_head *src, struct list_head *dst,
1155 struct nfs_commit_info *cinfo, int max)
1157 struct nfs_page *req, *tmp;
1160 list_for_each_entry_safe(req, tmp, src, wb_list) {
1161 if (!nfs_lock_request(req))
1163 kref_get(&req->wb_kref);
1164 if (cond_resched_lock(cinfo->lock))
1165 list_safe_reset_next(req, tmp, wb_list);
1166 nfs_request_remove_commit_list(req, cinfo);
1167 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1168 nfs_list_add_request(req, dst);
1170 if ((ret == max) && !cinfo->dreq)
1176 /* Note called with cinfo->lock held. */
1178 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1179 struct nfs_commit_info *cinfo,
1182 struct list_head *src = &bucket->written;
1183 struct list_head *dst = &bucket->committing;
1186 ret = transfer_commit_list(src, dst, cinfo, max);
1188 cinfo->ds->nwritten -= ret;
1189 cinfo->ds->ncommitting += ret;
1190 bucket->clseg = bucket->wlseg;
1191 if (list_empty(src))
1192 bucket->wlseg = NULL;
1194 pnfs_get_lseg(bucket->clseg);
1199 /* Move reqs from written to committing lists, returning count of number moved.
1200 * Note called with cinfo->lock held.
1202 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1207 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1208 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1216 /* Pull everything off the committing lists and dump into @dst */
1217 static void filelayout_recover_commit_reqs(struct list_head *dst,
1218 struct nfs_commit_info *cinfo)
1220 struct pnfs_commit_bucket *b;
1221 struct pnfs_layout_segment *freeme;
1225 spin_lock(cinfo->lock);
1226 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1227 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1230 spin_unlock(cinfo->lock);
1231 pnfs_put_lseg(freeme);
1235 cinfo->ds->nwritten = 0;
1236 spin_unlock(cinfo->lock);
1239 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1241 * @cinfo - commit info for current inode
1242 * @page - page to search for matching head request
1244 * Returns a the head request if one is found, otherwise returns NULL.
1246 static struct nfs_page *
1247 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1249 struct nfs_page *freq, *t;
1250 struct pnfs_commit_bucket *b;
1253 /* Linearly search the commit lists for each bucket until a matching
1254 * request is found */
1255 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1256 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1257 if (freq->wb_page == page)
1258 return freq->wb_head;
1260 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1261 if (freq->wb_page == page)
1262 return freq->wb_head;
1269 static void filelayout_retry_commit(struct nfs_commit_info *cinfo, int idx)
1271 struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
1272 struct pnfs_commit_bucket *bucket = fl_cinfo->buckets;
1273 struct pnfs_layout_segment *freeme;
1276 for (i = idx; i < fl_cinfo->nbuckets; i++, bucket++) {
1277 if (list_empty(&bucket->committing))
1279 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1280 spin_lock(cinfo->lock);
1281 freeme = bucket->clseg;
1282 bucket->clseg = NULL;
1283 spin_unlock(cinfo->lock);
1284 pnfs_put_lseg(freeme);
1289 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1291 struct pnfs_ds_commit_info *fl_cinfo;
1292 struct pnfs_commit_bucket *bucket;
1293 struct nfs_commit_data *data;
1295 unsigned int nreq = 0;
1297 fl_cinfo = cinfo->ds;
1298 bucket = fl_cinfo->buckets;
1299 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1300 if (list_empty(&bucket->committing))
1302 data = nfs_commitdata_alloc();
1305 data->ds_commit_index = i;
1306 spin_lock(cinfo->lock);
1307 data->lseg = bucket->clseg;
1308 bucket->clseg = NULL;
1309 spin_unlock(cinfo->lock);
1310 list_add(&data->pages, list);
1314 /* Clean up on error */
1315 filelayout_retry_commit(cinfo, i);
1316 /* Caller will clean up entries put on list */
1320 /* This follows nfs_commit_list pretty closely */
1322 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1323 int how, struct nfs_commit_info *cinfo)
1325 struct nfs_commit_data *data, *tmp;
1327 unsigned int nreq = 0;
1329 if (!list_empty(mds_pages)) {
1330 data = nfs_commitdata_alloc();
1333 list_add(&data->pages, &list);
1336 nfs_retry_commit(mds_pages, NULL, cinfo);
1337 filelayout_retry_commit(cinfo, 0);
1338 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1343 nreq += alloc_ds_commits(cinfo, &list);
1346 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1350 atomic_add(nreq, &cinfo->mds->rpcs_out);
1352 list_for_each_entry_safe(data, tmp, &list, pages) {
1353 list_del_init(&data->pages);
1355 nfs_init_commit(data, mds_pages, NULL, cinfo);
1356 nfs_initiate_commit(NFS_CLIENT(inode), data,
1357 data->mds_ops, how, 0);
1359 struct pnfs_commit_bucket *buckets;
1361 buckets = cinfo->ds->buckets;
1362 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1363 filelayout_initiate_commit(data, how);
1367 cinfo->ds->ncommitting = 0;
1368 return PNFS_ATTEMPTED;
1372 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1374 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1377 static struct pnfs_layout_hdr *
1378 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1380 struct nfs4_filelayout *flo;
1382 flo = kzalloc(sizeof(*flo), gfp_flags);
1383 return flo != NULL ? &flo->generic_hdr : NULL;
1387 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1389 kfree(FILELAYOUT_FROM_HDR(lo));
1392 static struct pnfs_ds_commit_info *
1393 filelayout_get_ds_info(struct inode *inode)
1395 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1400 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1403 static struct pnfs_layoutdriver_type filelayout_type = {
1404 .id = LAYOUT_NFSV4_1_FILES,
1405 .name = "LAYOUT_NFSV4_1_FILES",
1406 .owner = THIS_MODULE,
1407 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1408 .free_layout_hdr = filelayout_free_layout_hdr,
1409 .alloc_lseg = filelayout_alloc_lseg,
1410 .free_lseg = filelayout_free_lseg,
1411 .pg_read_ops = &filelayout_pg_read_ops,
1412 .pg_write_ops = &filelayout_pg_write_ops,
1413 .get_ds_info = &filelayout_get_ds_info,
1414 .mark_request_commit = filelayout_mark_request_commit,
1415 .clear_request_commit = filelayout_clear_request_commit,
1416 .scan_commit_lists = filelayout_scan_commit_lists,
1417 .recover_commit_reqs = filelayout_recover_commit_reqs,
1418 .search_commit_reqs = filelayout_search_commit_reqs,
1419 .commit_pagelist = filelayout_commit_pagelist,
1420 .read_pagelist = filelayout_read_pagelist,
1421 .write_pagelist = filelayout_write_pagelist,
1422 .free_deviceid_node = filelayout_free_deveiceid_node,
1425 static int __init nfs4filelayout_init(void)
1427 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1429 return pnfs_register_layoutdriver(&filelayout_type);
1432 static void __exit nfs4filelayout_exit(void)
1434 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1436 pnfs_unregister_layoutdriver(&filelayout_type);
1439 MODULE_ALIAS("nfs-layouttype4-1");
1441 module_init(nfs4filelayout_init);
1442 module_exit(nfs4filelayout_exit);