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_data *data)
89 struct nfs_pgio_header *hdr = data->header;
90 struct rpc_task *task = &data->task;
92 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
93 dprintk("%s Reset task %5u for i/o through MDS "
94 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
96 hdr->inode->i_sb->s_id,
97 (unsigned long long)NFS_FILEID(hdr->inode),
99 (unsigned long long)data->args.offset);
101 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
108 static void filelayout_reset_read(struct nfs_pgio_data *data)
110 struct nfs_pgio_header *hdr = data->header;
111 struct rpc_task *task = &data->task;
113 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
114 dprintk("%s Reset task %5u for i/o through MDS "
115 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
117 hdr->inode->i_sb->s_id,
118 (unsigned long long)NFS_FILEID(hdr->inode),
120 (unsigned long long)data->args.offset);
122 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
129 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
131 if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
133 pnfs_return_layout(inode);
136 static int filelayout_async_handle_error(struct rpc_task *task,
137 struct nfs4_state *state,
138 struct nfs_client *clp,
139 struct pnfs_layout_segment *lseg)
141 struct pnfs_layout_hdr *lo = lseg->pls_layout;
142 struct inode *inode = lo->plh_inode;
143 struct nfs_server *mds_server = NFS_SERVER(inode);
144 struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
145 struct nfs_client *mds_client = mds_server->nfs_client;
146 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
148 if (task->tk_status >= 0)
151 switch (task->tk_status) {
152 /* MDS state errors */
153 case -NFS4ERR_DELEG_REVOKED:
154 case -NFS4ERR_ADMIN_REVOKED:
155 case -NFS4ERR_BAD_STATEID:
158 nfs_remove_bad_delegation(state->inode);
159 case -NFS4ERR_OPENMODE:
162 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
163 goto out_bad_stateid;
164 goto wait_on_recovery;
165 case -NFS4ERR_EXPIRED:
167 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
168 goto out_bad_stateid;
170 nfs4_schedule_lease_recovery(mds_client);
171 goto wait_on_recovery;
172 /* DS session errors */
173 case -NFS4ERR_BADSESSION:
174 case -NFS4ERR_BADSLOT:
175 case -NFS4ERR_BAD_HIGH_SLOT:
176 case -NFS4ERR_DEADSESSION:
177 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
178 case -NFS4ERR_SEQ_FALSE_RETRY:
179 case -NFS4ERR_SEQ_MISORDERED:
180 dprintk("%s ERROR %d, Reset session. Exchangeid "
181 "flags 0x%x\n", __func__, task->tk_status,
182 clp->cl_exchange_flags);
183 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
187 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
189 case -NFS4ERR_RETRY_UNCACHED_REP:
191 /* Invalidate Layout errors */
192 case -NFS4ERR_PNFS_NO_LAYOUT:
193 case -ESTALE: /* mapped NFS4ERR_STALE */
194 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
195 case -EISDIR: /* mapped NFS4ERR_ISDIR */
196 case -NFS4ERR_FHEXPIRED:
197 case -NFS4ERR_WRONG_TYPE:
198 dprintk("%s Invalid layout error %d\n", __func__,
201 * Destroy layout so new i/o will get a new layout.
202 * Layout will not be destroyed until all current lseg
203 * references are put. Mark layout as invalid to resend failed
204 * i/o and all i/o waiting on the slot table to the MDS until
205 * layout is destroyed and a new valid layout is obtained.
207 pnfs_destroy_layout(NFS_I(inode));
208 rpc_wake_up(&tbl->slot_tbl_waitq);
210 /* RPC connection errors */
218 dprintk("%s DS connection error %d\n", __func__,
220 nfs4_mark_deviceid_unavailable(devid);
221 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
222 rpc_wake_up(&tbl->slot_tbl_waitq);
226 dprintk("%s Retry through MDS. Error %d\n", __func__,
228 return -NFS4ERR_RESET_TO_MDS;
234 task->tk_status = -EIO;
237 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
238 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
239 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
243 /* NFS_PROTO call done callback routines */
245 static int filelayout_read_done_cb(struct rpc_task *task,
246 struct nfs_pgio_data *data)
248 struct nfs_pgio_header *hdr = data->header;
251 trace_nfs4_pnfs_read(data, task->tk_status);
252 err = filelayout_async_handle_error(task, data->args.context->state,
253 data->ds_clp, hdr->lseg);
256 case -NFS4ERR_RESET_TO_MDS:
257 filelayout_reset_read(data);
258 return task->tk_status;
260 rpc_restart_call_prepare(task);
268 * We reference the rpc_cred of the first WRITE that triggers the need for
269 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
270 * rfc5661 is not clear about which credential should be used.
273 filelayout_set_layoutcommit(struct nfs_pgio_data *wdata)
275 struct nfs_pgio_header *hdr = wdata->header;
277 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
278 wdata->res.verf->committed == NFS_FILE_SYNC)
281 pnfs_set_layoutcommit(wdata);
282 dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
283 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
287 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
289 return filelayout_test_devid_invalid(node) ||
290 nfs4_test_deviceid_unavailable(node);
294 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
296 struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
298 return filelayout_test_devid_unavailable(node);
302 * Call ops for the async read/write cases
303 * In the case of dense layouts, the offset needs to be reset to its
306 static void filelayout_read_prepare(struct rpc_task *task, void *data)
308 struct nfs_pgio_data *rdata = data;
310 if (unlikely(test_bit(NFS_CONTEXT_BAD, &rdata->args.context->flags))) {
311 rpc_exit(task, -EIO);
314 if (filelayout_reset_to_mds(rdata->header->lseg)) {
315 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
316 filelayout_reset_read(rdata);
320 rdata->pgio_done_cb = filelayout_read_done_cb;
322 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
323 &rdata->args.seq_args,
327 if (nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
328 rdata->args.lock_context, FMODE_READ) == -EIO)
329 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
332 static void filelayout_read_call_done(struct rpc_task *task, void *data)
334 struct nfs_pgio_data *rdata = data;
336 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
338 if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
339 task->tk_status == 0) {
340 nfs41_sequence_done(task, &rdata->res.seq_res);
344 /* Note this may cause RPC to be resent */
345 rdata->header->mds_ops->rpc_call_done(task, data);
348 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
350 struct nfs_pgio_data *rdata = data;
352 rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
355 static void filelayout_read_release(void *data)
357 struct nfs_pgio_data *rdata = data;
358 struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
360 filelayout_fenceme(lo->plh_inode, lo);
361 nfs_put_client(rdata->ds_clp);
362 rdata->header->mds_ops->rpc_release(data);
365 static int filelayout_write_done_cb(struct rpc_task *task,
366 struct nfs_pgio_data *data)
368 struct nfs_pgio_header *hdr = data->header;
371 trace_nfs4_pnfs_write(data, task->tk_status);
372 err = filelayout_async_handle_error(task, data->args.context->state,
373 data->ds_clp, hdr->lseg);
376 case -NFS4ERR_RESET_TO_MDS:
377 filelayout_reset_write(data);
378 return task->tk_status;
380 rpc_restart_call_prepare(task);
384 filelayout_set_layoutcommit(data);
388 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
389 static void prepare_to_resend_writes(struct nfs_commit_data *data)
391 struct nfs_page *first = nfs_list_entry(data->pages.next);
393 data->task.tk_status = 0;
394 memcpy(&data->verf.verifier, &first->wb_verf,
395 sizeof(data->verf.verifier));
396 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
399 static int filelayout_commit_done_cb(struct rpc_task *task,
400 struct nfs_commit_data *data)
404 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
405 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
409 case -NFS4ERR_RESET_TO_MDS:
410 prepare_to_resend_writes(data);
413 rpc_restart_call_prepare(task);
420 static void filelayout_write_prepare(struct rpc_task *task, void *data)
422 struct nfs_pgio_data *wdata = data;
424 if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
425 rpc_exit(task, -EIO);
428 if (filelayout_reset_to_mds(wdata->header->lseg)) {
429 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
430 filelayout_reset_write(wdata);
434 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
435 &wdata->args.seq_args,
439 if (nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
440 wdata->args.lock_context, FMODE_WRITE) == -EIO)
441 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
444 static void filelayout_write_call_done(struct rpc_task *task, void *data)
446 struct nfs_pgio_data *wdata = data;
448 if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
449 task->tk_status == 0) {
450 nfs41_sequence_done(task, &wdata->res.seq_res);
454 /* Note this may cause RPC to be resent */
455 wdata->header->mds_ops->rpc_call_done(task, data);
458 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
460 struct nfs_pgio_data *wdata = data;
462 rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
465 static void filelayout_write_release(void *data)
467 struct nfs_pgio_data *wdata = data;
468 struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
470 filelayout_fenceme(lo->plh_inode, lo);
471 nfs_put_client(wdata->ds_clp);
472 wdata->header->mds_ops->rpc_release(data);
475 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
477 struct nfs_commit_data *wdata = data;
479 nfs41_setup_sequence(wdata->ds_clp->cl_session,
480 &wdata->args.seq_args,
485 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
487 struct nfs_commit_data *wdata = data;
489 /* Note this may cause RPC to be resent */
490 wdata->mds_ops->rpc_call_done(task, data);
493 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
495 struct nfs_commit_data *cdata = data;
497 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
500 static void filelayout_commit_release(void *calldata)
502 struct nfs_commit_data *data = calldata;
504 data->completion_ops->completion(data);
505 pnfs_put_lseg(data->lseg);
506 nfs_put_client(data->ds_clp);
507 nfs_commitdata_release(data);
510 static const struct rpc_call_ops filelayout_read_call_ops = {
511 .rpc_call_prepare = filelayout_read_prepare,
512 .rpc_call_done = filelayout_read_call_done,
513 .rpc_count_stats = filelayout_read_count_stats,
514 .rpc_release = filelayout_read_release,
517 static const struct rpc_call_ops filelayout_write_call_ops = {
518 .rpc_call_prepare = filelayout_write_prepare,
519 .rpc_call_done = filelayout_write_call_done,
520 .rpc_count_stats = filelayout_write_count_stats,
521 .rpc_release = filelayout_write_release,
524 static const struct rpc_call_ops filelayout_commit_call_ops = {
525 .rpc_call_prepare = filelayout_commit_prepare,
526 .rpc_call_done = filelayout_write_commit_done,
527 .rpc_count_stats = filelayout_commit_count_stats,
528 .rpc_release = filelayout_commit_release,
531 static enum pnfs_try_status
532 filelayout_read_pagelist(struct nfs_pgio_data *data)
534 struct nfs_pgio_header *hdr = data->header;
535 struct pnfs_layout_segment *lseg = hdr->lseg;
536 struct nfs4_pnfs_ds *ds;
537 struct rpc_clnt *ds_clnt;
538 loff_t offset = data->args.offset;
542 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
543 __func__, hdr->inode->i_ino,
544 data->args.pgbase, (size_t)data->args.count, offset);
546 /* Retrieve the correct rpc_client for the byte range */
547 j = nfs4_fl_calc_j_index(lseg, offset);
548 idx = nfs4_fl_calc_ds_index(lseg, j);
549 ds = nfs4_fl_prepare_ds(lseg, idx);
551 return PNFS_NOT_ATTEMPTED;
553 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
555 return PNFS_NOT_ATTEMPTED;
557 dprintk("%s USE DS: %s cl_count %d\n", __func__,
558 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
560 /* No multipath support. Use first DS */
561 atomic_inc(&ds->ds_clp->cl_count);
562 data->ds_clp = ds->ds_clp;
564 fh = nfs4_fl_select_ds_fh(lseg, j);
568 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
569 data->mds_offset = offset;
571 /* Perform an asynchronous read to ds */
572 nfs_initiate_pgio(ds_clnt, data,
573 &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
574 return PNFS_ATTEMPTED;
577 /* Perform async writes. */
578 static enum pnfs_try_status
579 filelayout_write_pagelist(struct nfs_pgio_data *data, int sync)
581 struct nfs_pgio_header *hdr = data->header;
582 struct pnfs_layout_segment *lseg = hdr->lseg;
583 struct nfs4_pnfs_ds *ds;
584 struct rpc_clnt *ds_clnt;
585 loff_t offset = data->args.offset;
589 /* Retrieve the correct rpc_client for the byte range */
590 j = nfs4_fl_calc_j_index(lseg, offset);
591 idx = nfs4_fl_calc_ds_index(lseg, j);
592 ds = nfs4_fl_prepare_ds(lseg, idx);
594 return PNFS_NOT_ATTEMPTED;
596 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
598 return PNFS_NOT_ATTEMPTED;
600 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
601 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
602 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
604 data->pgio_done_cb = filelayout_write_done_cb;
605 atomic_inc(&ds->ds_clp->cl_count);
606 data->ds_clp = ds->ds_clp;
608 fh = nfs4_fl_select_ds_fh(lseg, j);
612 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
614 /* Perform an asynchronous write */
615 nfs_initiate_pgio(ds_clnt, data,
616 &filelayout_write_call_ops, sync,
618 return PNFS_ATTEMPTED;
622 * filelayout_check_layout()
624 * Make sure layout segment parameters are sane WRT the device.
625 * At this point no generic layer initialization of the lseg has occurred,
626 * and nothing has been added to the layout_hdr cache.
630 filelayout_check_layout(struct pnfs_layout_hdr *lo,
631 struct nfs4_filelayout_segment *fl,
632 struct nfs4_layoutget_res *lgr,
633 struct nfs4_deviceid *id,
636 struct nfs4_deviceid_node *d;
637 struct nfs4_file_layout_dsaddr *dsaddr;
638 int status = -EINVAL;
640 dprintk("--> %s\n", __func__);
642 /* FIXME: remove this check when layout segment support is added */
643 if (lgr->range.offset != 0 ||
644 lgr->range.length != NFS4_MAX_UINT64) {
645 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
650 if (fl->pattern_offset > lgr->range.offset) {
651 dprintk("%s pattern_offset %lld too large\n",
652 __func__, fl->pattern_offset);
656 if (!fl->stripe_unit) {
657 dprintk("%s Invalid stripe unit (%u)\n",
658 __func__, fl->stripe_unit);
662 /* find and reference the deviceid */
663 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
664 NFS_SERVER(lo->plh_inode)->nfs_client, id);
666 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
667 lo->plh_lc_cred, gfp_flags);
671 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
672 /* Found deviceid is unavailable */
673 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
678 if (fl->first_stripe_index >= dsaddr->stripe_count) {
679 dprintk("%s Bad first_stripe_index %u\n",
680 __func__, fl->first_stripe_index);
684 if ((fl->stripe_type == STRIPE_SPARSE &&
685 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
686 (fl->stripe_type == STRIPE_DENSE &&
687 fl->num_fh != dsaddr->stripe_count)) {
688 dprintk("%s num_fh %u not valid for given packing\n",
689 __func__, fl->num_fh);
695 dprintk("--> %s returns %d\n", __func__, status);
698 nfs4_fl_put_deviceid(dsaddr);
702 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
706 for (i = 0; i < fl->num_fh; i++) {
707 if (!fl->fh_array[i])
709 kfree(fl->fh_array[i]);
716 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
718 filelayout_free_fh_array(fl);
723 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
724 struct nfs4_filelayout_segment *fl,
725 struct nfs4_layoutget_res *lgr,
726 struct nfs4_deviceid *id,
729 struct xdr_stream stream;
731 struct page *scratch;
736 dprintk("%s: set_layout_map Begin\n", __func__);
738 scratch = alloc_page(gfp_flags);
742 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
743 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
745 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
747 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
751 memcpy(id, p, sizeof(*id));
752 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
753 nfs4_print_deviceid(id);
755 nfl_util = be32_to_cpup(p++);
756 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
757 fl->commit_through_mds = 1;
758 if (nfl_util & NFL4_UFLG_DENSE)
759 fl->stripe_type = STRIPE_DENSE;
761 fl->stripe_type = STRIPE_SPARSE;
762 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
764 fl->first_stripe_index = be32_to_cpup(p++);
765 p = xdr_decode_hyper(p, &fl->pattern_offset);
766 fl->num_fh = be32_to_cpup(p++);
768 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
769 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
772 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
773 * Futher checking is done in filelayout_check_layout */
775 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
778 if (fl->num_fh > 0) {
779 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
785 for (i = 0; i < fl->num_fh; i++) {
786 /* Do we want to use a mempool here? */
787 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
788 if (!fl->fh_array[i])
791 p = xdr_inline_decode(&stream, 4);
794 fl->fh_array[i]->size = be32_to_cpup(p++);
795 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
796 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
797 i, fl->fh_array[i]->size);
801 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
804 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
805 dprintk("DEBUG: %s: fh len %d\n", __func__,
806 fl->fh_array[i]->size);
809 __free_page(scratch);
813 filelayout_free_fh_array(fl);
815 __free_page(scratch);
820 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
822 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
824 dprintk("--> %s\n", __func__);
825 nfs4_fl_put_deviceid(fl->dsaddr);
826 /* This assumes a single RW lseg */
827 if (lseg->pls_range.iomode == IOMODE_RW) {
828 struct nfs4_filelayout *flo;
830 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
831 flo->commit_info.nbuckets = 0;
832 kfree(flo->commit_info.buckets);
833 flo->commit_info.buckets = NULL;
835 _filelayout_free_lseg(fl);
839 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
840 struct nfs_commit_info *cinfo,
843 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
844 struct pnfs_commit_bucket *buckets;
847 if (fl->commit_through_mds)
850 size = (fl->stripe_type == STRIPE_SPARSE) ?
851 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
853 if (cinfo->ds->nbuckets >= size) {
854 /* This assumes there is only one IOMODE_RW lseg. What
855 * we really want to do is have a layout_hdr level
856 * dictionary of <multipath_list4, fh> keys, each
857 * associated with a struct list_head, populated by calls
858 * to filelayout_write_pagelist().
863 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
867 for (i = 0; i < size; i++) {
868 INIT_LIST_HEAD(&buckets[i].written);
869 INIT_LIST_HEAD(&buckets[i].committing);
870 /* mark direct verifier as unset */
871 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
874 spin_lock(cinfo->lock);
875 if (cinfo->ds->nbuckets >= size)
877 for (i = 0; i < cinfo->ds->nbuckets; i++) {
878 list_splice(&cinfo->ds->buckets[i].written,
879 &buckets[i].written);
880 list_splice(&cinfo->ds->buckets[i].committing,
881 &buckets[i].committing);
882 buckets[i].direct_verf.committed =
883 cinfo->ds->buckets[i].direct_verf.committed;
884 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
885 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
887 swap(cinfo->ds->buckets, buckets);
888 cinfo->ds->nbuckets = size;
890 spin_unlock(cinfo->lock);
895 static struct pnfs_layout_segment *
896 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
897 struct nfs4_layoutget_res *lgr,
900 struct nfs4_filelayout_segment *fl;
902 struct nfs4_deviceid id;
904 dprintk("--> %s\n", __func__);
905 fl = kzalloc(sizeof(*fl), gfp_flags);
909 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
910 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
911 _filelayout_free_lseg(fl);
914 return &fl->generic_hdr;
918 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
920 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
921 * of bytes (maximum @req->wb_bytes) that can be coalesced.
924 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
925 struct nfs_page *req)
928 u64 p_stripe, r_stripe;
930 u64 segment_offset = pgio->pg_lseg->pls_range.offset;
931 u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
933 /* calls nfs_generic_pg_test */
934 size = pnfs_generic_pg_test(pgio, prev, req);
938 /* see if req and prev are in the same stripe */
940 p_stripe = (u64)req_offset(prev) - segment_offset;
941 r_stripe = (u64)req_offset(req) - segment_offset;
942 do_div(p_stripe, stripe_unit);
943 do_div(r_stripe, stripe_unit);
945 if (p_stripe != r_stripe)
949 /* calculate remaining bytes in the current stripe */
950 div_u64_rem((u64)req_offset(req) - segment_offset,
953 WARN_ON_ONCE(stripe_offset > stripe_unit);
954 if (stripe_offset >= stripe_unit)
956 return min(stripe_unit - (unsigned int)stripe_offset, size);
960 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
961 struct nfs_page *req)
964 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
970 /* If no lseg, fall back to read through mds */
971 if (pgio->pg_lseg == NULL)
972 nfs_pageio_reset_read_mds(pgio);
976 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
977 struct nfs_page *req)
979 struct nfs_commit_info cinfo;
983 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
989 /* If no lseg, fall back to write through mds */
990 if (pgio->pg_lseg == NULL)
992 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
993 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
995 pnfs_put_lseg(pgio->pg_lseg);
996 pgio->pg_lseg = NULL;
1001 nfs_pageio_reset_write_mds(pgio);
1004 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
1005 .pg_init = filelayout_pg_init_read,
1006 .pg_test = filelayout_pg_test,
1007 .pg_doio = pnfs_generic_pg_readpages,
1010 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
1011 .pg_init = filelayout_pg_init_write,
1012 .pg_test = filelayout_pg_test,
1013 .pg_doio = pnfs_generic_pg_writepages,
1016 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1018 if (fl->stripe_type == STRIPE_SPARSE)
1019 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1024 /* The generic layer is about to remove the req from the commit list.
1025 * If this will make the bucket empty, it will need to put the lseg reference.
1028 filelayout_clear_request_commit(struct nfs_page *req,
1029 struct nfs_commit_info *cinfo)
1031 struct pnfs_layout_segment *freeme = NULL;
1033 spin_lock(cinfo->lock);
1034 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1036 cinfo->ds->nwritten--;
1037 if (list_is_singular(&req->wb_list)) {
1038 struct pnfs_commit_bucket *bucket;
1040 bucket = list_first_entry(&req->wb_list,
1041 struct pnfs_commit_bucket,
1043 freeme = bucket->wlseg;
1044 bucket->wlseg = NULL;
1047 nfs_request_remove_commit_list(req, cinfo);
1048 spin_unlock(cinfo->lock);
1049 pnfs_put_lseg(freeme);
1052 static struct list_head *
1053 filelayout_choose_commit_list(struct nfs_page *req,
1054 struct pnfs_layout_segment *lseg,
1055 struct nfs_commit_info *cinfo)
1057 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1059 struct list_head *list;
1060 struct pnfs_commit_bucket *buckets;
1062 if (fl->commit_through_mds)
1063 return &cinfo->mds->list;
1065 /* Note that we are calling nfs4_fl_calc_j_index on each page
1066 * that ends up being committed to a data server. An attractive
1067 * alternative is to add a field to nfs_write_data and nfs_page
1068 * to store the value calculated in filelayout_write_pagelist
1069 * and just use that here.
1071 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1072 i = select_bucket_index(fl, j);
1073 spin_lock(cinfo->lock);
1074 buckets = cinfo->ds->buckets;
1075 list = &buckets[i].written;
1076 if (list_empty(list)) {
1077 /* Non-empty buckets hold a reference on the lseg. That ref
1078 * is normally transferred to the COMMIT call and released
1079 * there. It could also be released if the last req is pulled
1080 * off due to a rewrite, in which case it will be done in
1081 * filelayout_clear_request_commit
1083 buckets[i].wlseg = pnfs_get_lseg(lseg);
1085 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1086 cinfo->ds->nwritten++;
1087 spin_unlock(cinfo->lock);
1092 filelayout_mark_request_commit(struct nfs_page *req,
1093 struct pnfs_layout_segment *lseg,
1094 struct nfs_commit_info *cinfo)
1096 struct list_head *list;
1098 list = filelayout_choose_commit_list(req, lseg, cinfo);
1099 nfs_request_add_commit_list(req, list, cinfo);
1102 static u32 calc_ds_index_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)
1109 return nfs4_fl_calc_ds_index(lseg, i);
1112 static struct nfs_fh *
1113 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1115 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1117 if (flseg->stripe_type == STRIPE_SPARSE) {
1118 if (flseg->num_fh == 1)
1120 else if (flseg->num_fh == 0)
1121 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1124 return flseg->fh_array[i];
1127 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1129 struct pnfs_layout_segment *lseg = data->lseg;
1130 struct nfs4_pnfs_ds *ds;
1131 struct rpc_clnt *ds_clnt;
1135 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1136 ds = nfs4_fl_prepare_ds(lseg, idx);
1140 ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1141 if (IS_ERR(ds_clnt))
1144 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1145 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1146 data->commit_done_cb = filelayout_commit_done_cb;
1147 atomic_inc(&ds->ds_clp->cl_count);
1148 data->ds_clp = ds->ds_clp;
1149 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1152 return nfs_initiate_commit(ds_clnt, data,
1153 &filelayout_commit_call_ops, how,
1156 prepare_to_resend_writes(data);
1157 filelayout_commit_release(data);
1162 transfer_commit_list(struct list_head *src, struct list_head *dst,
1163 struct nfs_commit_info *cinfo, int max)
1165 struct nfs_page *req, *tmp;
1168 list_for_each_entry_safe(req, tmp, src, wb_list) {
1169 if (!nfs_lock_request(req))
1171 kref_get(&req->wb_kref);
1172 if (cond_resched_lock(cinfo->lock))
1173 list_safe_reset_next(req, tmp, wb_list);
1174 nfs_request_remove_commit_list(req, cinfo);
1175 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1176 nfs_list_add_request(req, dst);
1178 if ((ret == max) && !cinfo->dreq)
1184 /* Note called with cinfo->lock held. */
1186 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1187 struct nfs_commit_info *cinfo,
1190 struct list_head *src = &bucket->written;
1191 struct list_head *dst = &bucket->committing;
1194 ret = transfer_commit_list(src, dst, cinfo, max);
1196 cinfo->ds->nwritten -= ret;
1197 cinfo->ds->ncommitting += ret;
1198 bucket->clseg = bucket->wlseg;
1199 if (list_empty(src))
1200 bucket->wlseg = NULL;
1202 pnfs_get_lseg(bucket->clseg);
1207 /* Move reqs from written to committing lists, returning count of number moved.
1208 * Note called with cinfo->lock held.
1210 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1215 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1216 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1224 /* Pull everything off the committing lists and dump into @dst */
1225 static void filelayout_recover_commit_reqs(struct list_head *dst,
1226 struct nfs_commit_info *cinfo)
1228 struct pnfs_commit_bucket *b;
1229 struct pnfs_layout_segment *freeme;
1233 spin_lock(cinfo->lock);
1234 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1235 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1238 spin_unlock(cinfo->lock);
1239 pnfs_put_lseg(freeme);
1243 cinfo->ds->nwritten = 0;
1244 spin_unlock(cinfo->lock);
1248 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1250 struct pnfs_ds_commit_info *fl_cinfo;
1251 struct pnfs_commit_bucket *bucket;
1252 struct nfs_commit_data *data;
1254 unsigned int nreq = 0;
1255 struct pnfs_layout_segment *freeme;
1257 fl_cinfo = cinfo->ds;
1258 bucket = fl_cinfo->buckets;
1259 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1260 if (list_empty(&bucket->committing))
1262 data = nfs_commitdata_alloc();
1265 data->ds_commit_index = i;
1266 spin_lock(cinfo->lock);
1267 data->lseg = bucket->clseg;
1268 bucket->clseg = NULL;
1269 spin_unlock(cinfo->lock);
1270 list_add(&data->pages, list);
1274 /* Clean up on error */
1275 for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1276 if (list_empty(&bucket->committing))
1278 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1279 spin_lock(cinfo->lock);
1280 freeme = bucket->clseg;
1281 bucket->clseg = NULL;
1282 spin_unlock(cinfo->lock);
1283 pnfs_put_lseg(freeme);
1285 /* Caller will clean up entries put on list */
1289 /* This follows nfs_commit_list pretty closely */
1291 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1292 int how, struct nfs_commit_info *cinfo)
1294 struct nfs_commit_data *data, *tmp;
1296 unsigned int nreq = 0;
1298 if (!list_empty(mds_pages)) {
1299 data = nfs_commitdata_alloc();
1302 list_add(&data->pages, &list);
1305 nfs_retry_commit(mds_pages, NULL, cinfo);
1308 nreq += alloc_ds_commits(cinfo, &list);
1311 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1315 atomic_add(nreq, &cinfo->mds->rpcs_out);
1317 list_for_each_entry_safe(data, tmp, &list, pages) {
1318 list_del_init(&data->pages);
1320 nfs_init_commit(data, mds_pages, NULL, cinfo);
1321 nfs_initiate_commit(NFS_CLIENT(inode), data,
1322 data->mds_ops, how, 0);
1324 struct pnfs_commit_bucket *buckets;
1326 buckets = cinfo->ds->buckets;
1327 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1328 filelayout_initiate_commit(data, how);
1332 cinfo->ds->ncommitting = 0;
1333 return PNFS_ATTEMPTED;
1337 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1339 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1342 static struct pnfs_layout_hdr *
1343 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1345 struct nfs4_filelayout *flo;
1347 flo = kzalloc(sizeof(*flo), gfp_flags);
1348 return flo != NULL ? &flo->generic_hdr : NULL;
1352 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1354 kfree(FILELAYOUT_FROM_HDR(lo));
1357 static struct pnfs_ds_commit_info *
1358 filelayout_get_ds_info(struct inode *inode)
1360 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1365 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1368 static struct pnfs_layoutdriver_type filelayout_type = {
1369 .id = LAYOUT_NFSV4_1_FILES,
1370 .name = "LAYOUT_NFSV4_1_FILES",
1371 .owner = THIS_MODULE,
1372 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1373 .free_layout_hdr = filelayout_free_layout_hdr,
1374 .alloc_lseg = filelayout_alloc_lseg,
1375 .free_lseg = filelayout_free_lseg,
1376 .pg_read_ops = &filelayout_pg_read_ops,
1377 .pg_write_ops = &filelayout_pg_write_ops,
1378 .get_ds_info = &filelayout_get_ds_info,
1379 .mark_request_commit = filelayout_mark_request_commit,
1380 .clear_request_commit = filelayout_clear_request_commit,
1381 .scan_commit_lists = filelayout_scan_commit_lists,
1382 .recover_commit_reqs = filelayout_recover_commit_reqs,
1383 .commit_pagelist = filelayout_commit_pagelist,
1384 .read_pagelist = filelayout_read_pagelist,
1385 .write_pagelist = filelayout_write_pagelist,
1386 .free_deviceid_node = filelayout_free_deveiceid_node,
1389 static int __init nfs4filelayout_init(void)
1391 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1393 return pnfs_register_layoutdriver(&filelayout_type);
1396 static void __exit nfs4filelayout_exit(void)
1398 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1400 pnfs_unregister_layoutdriver(&filelayout_type);
1403 MODULE_ALIAS("nfs-layouttype4-1");
1405 module_init(nfs4filelayout_init);
1406 module_exit(nfs4filelayout_exit);