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"
40 #include "delegation.h"
41 #include "nfs4filelayout.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_write_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/%lld, %u bytes @ offset %llu)\n", __func__,
96 hdr->inode->i_sb->s_id,
97 (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_read_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/%lld, %u bytes @ offset %llu)\n", __func__,
117 hdr->inode->i_sb->s_id,
118 (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_read_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_write_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 ionde %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_read_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->read_done_cb = filelayout_read_done_cb;
322 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
323 &rdata->args.seq_args,
327 nfs4_set_rw_stateid(&rdata->args.stateid, rdata->args.context,
328 rdata->args.lock_context, FMODE_READ);
331 static void filelayout_read_call_done(struct rpc_task *task, void *data)
333 struct nfs_read_data *rdata = data;
335 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
337 if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
338 task->tk_status == 0)
341 /* Note this may cause RPC to be resent */
342 rdata->header->mds_ops->rpc_call_done(task, data);
345 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
347 struct nfs_read_data *rdata = data;
349 rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
352 static void filelayout_read_release(void *data)
354 struct nfs_read_data *rdata = data;
355 struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
357 filelayout_fenceme(lo->plh_inode, lo);
358 nfs_put_client(rdata->ds_clp);
359 rdata->header->mds_ops->rpc_release(data);
362 static int filelayout_write_done_cb(struct rpc_task *task,
363 struct nfs_write_data *data)
365 struct nfs_pgio_header *hdr = data->header;
368 trace_nfs4_pnfs_write(data, task->tk_status);
369 err = filelayout_async_handle_error(task, data->args.context->state,
370 data->ds_clp, hdr->lseg);
373 case -NFS4ERR_RESET_TO_MDS:
374 filelayout_reset_write(data);
375 return task->tk_status;
377 rpc_restart_call_prepare(task);
381 filelayout_set_layoutcommit(data);
385 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
386 static void prepare_to_resend_writes(struct nfs_commit_data *data)
388 struct nfs_page *first = nfs_list_entry(data->pages.next);
390 data->task.tk_status = 0;
391 memcpy(&data->verf.verifier, &first->wb_verf,
392 sizeof(data->verf.verifier));
393 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
396 static int filelayout_commit_done_cb(struct rpc_task *task,
397 struct nfs_commit_data *data)
401 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
402 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
406 case -NFS4ERR_RESET_TO_MDS:
407 prepare_to_resend_writes(data);
410 rpc_restart_call_prepare(task);
417 static void filelayout_write_prepare(struct rpc_task *task, void *data)
419 struct nfs_write_data *wdata = data;
421 if (unlikely(test_bit(NFS_CONTEXT_BAD, &wdata->args.context->flags))) {
422 rpc_exit(task, -EIO);
425 if (filelayout_reset_to_mds(wdata->header->lseg)) {
426 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
427 filelayout_reset_write(wdata);
431 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
432 &wdata->args.seq_args,
436 nfs4_set_rw_stateid(&wdata->args.stateid, wdata->args.context,
437 wdata->args.lock_context, FMODE_WRITE);
440 static void filelayout_write_call_done(struct rpc_task *task, void *data)
442 struct nfs_write_data *wdata = data;
444 if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
445 task->tk_status == 0)
448 /* Note this may cause RPC to be resent */
449 wdata->header->mds_ops->rpc_call_done(task, data);
452 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
454 struct nfs_write_data *wdata = data;
456 rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
459 static void filelayout_write_release(void *data)
461 struct nfs_write_data *wdata = data;
462 struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
464 filelayout_fenceme(lo->plh_inode, lo);
465 nfs_put_client(wdata->ds_clp);
466 wdata->header->mds_ops->rpc_release(data);
469 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
471 struct nfs_commit_data *wdata = data;
473 nfs41_setup_sequence(wdata->ds_clp->cl_session,
474 &wdata->args.seq_args,
479 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
481 struct nfs_commit_data *wdata = data;
483 /* Note this may cause RPC to be resent */
484 wdata->mds_ops->rpc_call_done(task, data);
487 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
489 struct nfs_commit_data *cdata = data;
491 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
494 static void filelayout_commit_release(void *calldata)
496 struct nfs_commit_data *data = calldata;
498 data->completion_ops->completion(data);
499 pnfs_put_lseg(data->lseg);
500 nfs_put_client(data->ds_clp);
501 nfs_commitdata_release(data);
504 static const struct rpc_call_ops filelayout_read_call_ops = {
505 .rpc_call_prepare = filelayout_read_prepare,
506 .rpc_call_done = filelayout_read_call_done,
507 .rpc_count_stats = filelayout_read_count_stats,
508 .rpc_release = filelayout_read_release,
511 static const struct rpc_call_ops filelayout_write_call_ops = {
512 .rpc_call_prepare = filelayout_write_prepare,
513 .rpc_call_done = filelayout_write_call_done,
514 .rpc_count_stats = filelayout_write_count_stats,
515 .rpc_release = filelayout_write_release,
518 static const struct rpc_call_ops filelayout_commit_call_ops = {
519 .rpc_call_prepare = filelayout_commit_prepare,
520 .rpc_call_done = filelayout_write_commit_done,
521 .rpc_count_stats = filelayout_commit_count_stats,
522 .rpc_release = filelayout_commit_release,
525 static enum pnfs_try_status
526 filelayout_read_pagelist(struct nfs_read_data *data)
528 struct nfs_pgio_header *hdr = data->header;
529 struct pnfs_layout_segment *lseg = hdr->lseg;
530 struct nfs4_pnfs_ds *ds;
531 loff_t offset = data->args.offset;
535 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
536 __func__, hdr->inode->i_ino,
537 data->args.pgbase, (size_t)data->args.count, offset);
539 /* Retrieve the correct rpc_client for the byte range */
540 j = nfs4_fl_calc_j_index(lseg, offset);
541 idx = nfs4_fl_calc_ds_index(lseg, j);
542 ds = nfs4_fl_prepare_ds(lseg, idx);
544 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 data->ds_clp = ds->ds_clp;
551 fh = nfs4_fl_select_ds_fh(lseg, j);
555 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
556 data->mds_offset = offset;
558 /* Perform an asynchronous read to ds */
559 nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
560 &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
561 return PNFS_ATTEMPTED;
564 /* Perform async writes. */
565 static enum pnfs_try_status
566 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
568 struct nfs_pgio_header *hdr = data->header;
569 struct pnfs_layout_segment *lseg = hdr->lseg;
570 struct nfs4_pnfs_ds *ds;
571 loff_t offset = data->args.offset;
575 /* Retrieve the correct rpc_client for the byte range */
576 j = nfs4_fl_calc_j_index(lseg, offset);
577 idx = nfs4_fl_calc_ds_index(lseg, j);
578 ds = nfs4_fl_prepare_ds(lseg, idx);
580 return PNFS_NOT_ATTEMPTED;
581 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
582 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
583 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
585 data->write_done_cb = filelayout_write_done_cb;
586 atomic_inc(&ds->ds_clp->cl_count);
587 data->ds_clp = ds->ds_clp;
588 fh = nfs4_fl_select_ds_fh(lseg, j);
592 * Get the file offset on the dserver. Set the write offset to
593 * this offset and save the original offset.
595 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
597 /* Perform an asynchronous write */
598 nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
599 &filelayout_write_call_ops, sync,
601 return PNFS_ATTEMPTED;
605 * filelayout_check_layout()
607 * Make sure layout segment parameters are sane WRT the device.
608 * At this point no generic layer initialization of the lseg has occurred,
609 * and nothing has been added to the layout_hdr cache.
613 filelayout_check_layout(struct pnfs_layout_hdr *lo,
614 struct nfs4_filelayout_segment *fl,
615 struct nfs4_layoutget_res *lgr,
616 struct nfs4_deviceid *id,
619 struct nfs4_deviceid_node *d;
620 struct nfs4_file_layout_dsaddr *dsaddr;
621 int status = -EINVAL;
622 struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
624 dprintk("--> %s\n", __func__);
626 /* FIXME: remove this check when layout segment support is added */
627 if (lgr->range.offset != 0 ||
628 lgr->range.length != NFS4_MAX_UINT64) {
629 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
634 if (fl->pattern_offset > lgr->range.offset) {
635 dprintk("%s pattern_offset %lld too large\n",
636 __func__, fl->pattern_offset);
640 if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
641 dprintk("%s Invalid stripe unit (%u)\n",
642 __func__, fl->stripe_unit);
646 /* find and reference the deviceid */
647 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
648 NFS_SERVER(lo->plh_inode)->nfs_client, id);
650 dsaddr = filelayout_get_device_info(lo->plh_inode, id,
651 lo->plh_lc_cred, gfp_flags);
655 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
656 /* Found deviceid is unavailable */
657 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
662 if (fl->first_stripe_index >= dsaddr->stripe_count) {
663 dprintk("%s Bad first_stripe_index %u\n",
664 __func__, fl->first_stripe_index);
668 if ((fl->stripe_type == STRIPE_SPARSE &&
669 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
670 (fl->stripe_type == STRIPE_DENSE &&
671 fl->num_fh != dsaddr->stripe_count)) {
672 dprintk("%s num_fh %u not valid for given packing\n",
673 __func__, fl->num_fh);
677 if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
678 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
679 "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
685 dprintk("--> %s returns %d\n", __func__, status);
688 nfs4_fl_put_deviceid(dsaddr);
692 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
696 for (i = 0; i < fl->num_fh; i++) {
697 if (!fl->fh_array[i])
699 kfree(fl->fh_array[i]);
706 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
708 filelayout_free_fh_array(fl);
713 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
714 struct nfs4_filelayout_segment *fl,
715 struct nfs4_layoutget_res *lgr,
716 struct nfs4_deviceid *id,
719 struct xdr_stream stream;
721 struct page *scratch;
726 dprintk("%s: set_layout_map Begin\n", __func__);
728 scratch = alloc_page(gfp_flags);
732 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
733 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
735 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
737 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
741 memcpy(id, p, sizeof(*id));
742 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
743 nfs4_print_deviceid(id);
745 nfl_util = be32_to_cpup(p++);
746 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
747 fl->commit_through_mds = 1;
748 if (nfl_util & NFL4_UFLG_DENSE)
749 fl->stripe_type = STRIPE_DENSE;
751 fl->stripe_type = STRIPE_SPARSE;
752 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
754 fl->first_stripe_index = be32_to_cpup(p++);
755 p = xdr_decode_hyper(p, &fl->pattern_offset);
756 fl->num_fh = be32_to_cpup(p++);
758 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
759 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
762 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
763 * Futher checking is done in filelayout_check_layout */
765 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
768 if (fl->num_fh > 0) {
769 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
775 for (i = 0; i < fl->num_fh; i++) {
776 /* Do we want to use a mempool here? */
777 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
778 if (!fl->fh_array[i])
781 p = xdr_inline_decode(&stream, 4);
784 fl->fh_array[i]->size = be32_to_cpup(p++);
785 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
786 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
787 i, fl->fh_array[i]->size);
791 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
794 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
795 dprintk("DEBUG: %s: fh len %d\n", __func__,
796 fl->fh_array[i]->size);
799 __free_page(scratch);
803 filelayout_free_fh_array(fl);
805 __free_page(scratch);
810 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
812 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
814 dprintk("--> %s\n", __func__);
815 nfs4_fl_put_deviceid(fl->dsaddr);
816 /* This assumes a single RW lseg */
817 if (lseg->pls_range.iomode == IOMODE_RW) {
818 struct nfs4_filelayout *flo;
820 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
821 flo->commit_info.nbuckets = 0;
822 kfree(flo->commit_info.buckets);
823 flo->commit_info.buckets = NULL;
825 _filelayout_free_lseg(fl);
829 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
830 struct nfs_commit_info *cinfo,
833 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
834 struct pnfs_commit_bucket *buckets;
837 if (fl->commit_through_mds)
839 if (cinfo->ds->nbuckets != 0) {
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 size = (fl->stripe_type == STRIPE_SPARSE) ?
850 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
852 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
859 spin_lock(cinfo->lock);
860 if (cinfo->ds->nbuckets != 0)
863 cinfo->ds->buckets = buckets;
864 cinfo->ds->nbuckets = size;
865 for (i = 0; i < size; i++) {
866 INIT_LIST_HEAD(&buckets[i].written);
867 INIT_LIST_HEAD(&buckets[i].committing);
870 spin_unlock(cinfo->lock);
875 static struct pnfs_layout_segment *
876 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
877 struct nfs4_layoutget_res *lgr,
880 struct nfs4_filelayout_segment *fl;
882 struct nfs4_deviceid id;
884 dprintk("--> %s\n", __func__);
885 fl = kzalloc(sizeof(*fl), gfp_flags);
889 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
890 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
891 _filelayout_free_lseg(fl);
894 return &fl->generic_hdr;
898 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
900 * return true : coalesce page
901 * return false : don't coalesce page
904 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
905 struct nfs_page *req)
907 u64 p_stripe, r_stripe;
910 if (!pnfs_generic_pg_test(pgio, prev, req) ||
911 !nfs_generic_pg_test(pgio, prev, req))
914 p_stripe = (u64)req_offset(prev);
915 r_stripe = (u64)req_offset(req);
916 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
918 do_div(p_stripe, stripe_unit);
919 do_div(r_stripe, stripe_unit);
921 return (p_stripe == r_stripe);
925 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
926 struct nfs_page *req)
928 WARN_ON_ONCE(pgio->pg_lseg != NULL);
930 if (req->wb_offset != req->wb_pgbase) {
932 * Handling unaligned pages is difficult, because have to
933 * somehow split a req in two in certain cases in the
934 * pg.test code. Avoid this by just not using pnfs
937 nfs_pageio_reset_read_mds(pgio);
940 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
946 /* If no lseg, fall back to read through mds */
947 if (pgio->pg_lseg == NULL)
948 nfs_pageio_reset_read_mds(pgio);
952 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
953 struct nfs_page *req)
955 struct nfs_commit_info cinfo;
958 WARN_ON_ONCE(pgio->pg_lseg != NULL);
960 if (req->wb_offset != req->wb_pgbase)
962 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
968 /* If no lseg, fall back to write through mds */
969 if (pgio->pg_lseg == NULL)
971 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
972 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
974 pnfs_put_lseg(pgio->pg_lseg);
975 pgio->pg_lseg = NULL;
980 nfs_pageio_reset_write_mds(pgio);
983 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
984 .pg_init = filelayout_pg_init_read,
985 .pg_test = filelayout_pg_test,
986 .pg_doio = pnfs_generic_pg_readpages,
989 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
990 .pg_init = filelayout_pg_init_write,
991 .pg_test = filelayout_pg_test,
992 .pg_doio = pnfs_generic_pg_writepages,
995 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
997 if (fl->stripe_type == STRIPE_SPARSE)
998 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1003 /* The generic layer is about to remove the req from the commit list.
1004 * If this will make the bucket empty, it will need to put the lseg reference.
1007 filelayout_clear_request_commit(struct nfs_page *req,
1008 struct nfs_commit_info *cinfo)
1010 struct pnfs_layout_segment *freeme = NULL;
1012 spin_lock(cinfo->lock);
1013 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1015 cinfo->ds->nwritten--;
1016 if (list_is_singular(&req->wb_list)) {
1017 struct pnfs_commit_bucket *bucket;
1019 bucket = list_first_entry(&req->wb_list,
1020 struct pnfs_commit_bucket,
1022 freeme = bucket->wlseg;
1023 bucket->wlseg = NULL;
1026 nfs_request_remove_commit_list(req, cinfo);
1027 spin_unlock(cinfo->lock);
1028 pnfs_put_lseg(freeme);
1031 static struct list_head *
1032 filelayout_choose_commit_list(struct nfs_page *req,
1033 struct pnfs_layout_segment *lseg,
1034 struct nfs_commit_info *cinfo)
1036 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1038 struct list_head *list;
1039 struct pnfs_commit_bucket *buckets;
1041 if (fl->commit_through_mds)
1042 return &cinfo->mds->list;
1044 /* Note that we are calling nfs4_fl_calc_j_index on each page
1045 * that ends up being committed to a data server. An attractive
1046 * alternative is to add a field to nfs_write_data and nfs_page
1047 * to store the value calculated in filelayout_write_pagelist
1048 * and just use that here.
1050 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1051 i = select_bucket_index(fl, j);
1052 buckets = cinfo->ds->buckets;
1053 list = &buckets[i].written;
1054 if (list_empty(list)) {
1055 /* Non-empty buckets hold a reference on the lseg. That ref
1056 * is normally transferred to the COMMIT call and released
1057 * there. It could also be released if the last req is pulled
1058 * off due to a rewrite, in which case it will be done in
1059 * filelayout_clear_request_commit
1061 buckets[i].wlseg = pnfs_get_lseg(lseg);
1063 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1064 cinfo->ds->nwritten++;
1069 filelayout_mark_request_commit(struct nfs_page *req,
1070 struct pnfs_layout_segment *lseg,
1071 struct nfs_commit_info *cinfo)
1073 struct list_head *list;
1075 list = filelayout_choose_commit_list(req, lseg, cinfo);
1076 nfs_request_add_commit_list(req, list, cinfo);
1079 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1081 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1083 if (flseg->stripe_type == STRIPE_SPARSE)
1086 return nfs4_fl_calc_ds_index(lseg, i);
1089 static struct nfs_fh *
1090 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1092 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1094 if (flseg->stripe_type == STRIPE_SPARSE) {
1095 if (flseg->num_fh == 1)
1097 else if (flseg->num_fh == 0)
1098 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1101 return flseg->fh_array[i];
1104 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1106 struct pnfs_layout_segment *lseg = data->lseg;
1107 struct nfs4_pnfs_ds *ds;
1111 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1112 ds = nfs4_fl_prepare_ds(lseg, idx);
1114 prepare_to_resend_writes(data);
1115 filelayout_commit_release(data);
1118 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1119 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1120 data->commit_done_cb = filelayout_commit_done_cb;
1121 atomic_inc(&ds->ds_clp->cl_count);
1122 data->ds_clp = ds->ds_clp;
1123 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1126 return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1127 &filelayout_commit_call_ops, how,
1132 transfer_commit_list(struct list_head *src, struct list_head *dst,
1133 struct nfs_commit_info *cinfo, int max)
1135 struct nfs_page *req, *tmp;
1138 list_for_each_entry_safe(req, tmp, src, wb_list) {
1139 if (!nfs_lock_request(req))
1141 kref_get(&req->wb_kref);
1142 if (cond_resched_lock(cinfo->lock))
1143 list_safe_reset_next(req, tmp, wb_list);
1144 nfs_request_remove_commit_list(req, cinfo);
1145 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1146 nfs_list_add_request(req, dst);
1148 if ((ret == max) && !cinfo->dreq)
1155 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1156 struct nfs_commit_info *cinfo,
1159 struct list_head *src = &bucket->written;
1160 struct list_head *dst = &bucket->committing;
1163 ret = transfer_commit_list(src, dst, cinfo, max);
1165 cinfo->ds->nwritten -= ret;
1166 cinfo->ds->ncommitting += ret;
1167 bucket->clseg = bucket->wlseg;
1168 if (list_empty(src))
1169 bucket->wlseg = NULL;
1171 pnfs_get_lseg(bucket->clseg);
1176 /* Move reqs from written to committing lists, returning count of number moved.
1177 * Note called with cinfo->lock held.
1179 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1184 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1185 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1193 /* Pull everything off the committing lists and dump into @dst */
1194 static void filelayout_recover_commit_reqs(struct list_head *dst,
1195 struct nfs_commit_info *cinfo)
1197 struct pnfs_commit_bucket *b;
1200 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1201 * only called on single thread per dreq.
1202 * Can't take the lock because need to do pnfs_put_lseg
1204 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1205 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1206 pnfs_put_lseg(b->wlseg);
1210 cinfo->ds->nwritten = 0;
1214 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1216 struct pnfs_ds_commit_info *fl_cinfo;
1217 struct pnfs_commit_bucket *bucket;
1218 struct nfs_commit_data *data;
1220 unsigned int nreq = 0;
1222 fl_cinfo = cinfo->ds;
1223 bucket = fl_cinfo->buckets;
1224 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1225 if (list_empty(&bucket->committing))
1227 data = nfs_commitdata_alloc();
1230 data->ds_commit_index = i;
1231 data->lseg = bucket->clseg;
1232 bucket->clseg = NULL;
1233 list_add(&data->pages, list);
1237 /* Clean up on error */
1238 for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1239 if (list_empty(&bucket->committing))
1241 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1242 pnfs_put_lseg(bucket->clseg);
1243 bucket->clseg = NULL;
1245 /* Caller will clean up entries put on list */
1249 /* This follows nfs_commit_list pretty closely */
1251 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1252 int how, struct nfs_commit_info *cinfo)
1254 struct nfs_commit_data *data, *tmp;
1256 unsigned int nreq = 0;
1258 if (!list_empty(mds_pages)) {
1259 data = nfs_commitdata_alloc();
1262 list_add(&data->pages, &list);
1265 nfs_retry_commit(mds_pages, NULL, cinfo);
1268 nreq += alloc_ds_commits(cinfo, &list);
1271 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1275 atomic_add(nreq, &cinfo->mds->rpcs_out);
1277 list_for_each_entry_safe(data, tmp, &list, pages) {
1278 list_del_init(&data->pages);
1280 nfs_init_commit(data, mds_pages, NULL, cinfo);
1281 nfs_initiate_commit(NFS_CLIENT(inode), data,
1282 data->mds_ops, how, 0);
1284 struct pnfs_commit_bucket *buckets;
1286 buckets = cinfo->ds->buckets;
1287 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1288 filelayout_initiate_commit(data, how);
1292 cinfo->ds->ncommitting = 0;
1293 return PNFS_ATTEMPTED;
1297 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1299 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1302 static struct pnfs_layout_hdr *
1303 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1305 struct nfs4_filelayout *flo;
1307 flo = kzalloc(sizeof(*flo), gfp_flags);
1308 return &flo->generic_hdr;
1312 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1314 kfree(FILELAYOUT_FROM_HDR(lo));
1317 static struct pnfs_ds_commit_info *
1318 filelayout_get_ds_info(struct inode *inode)
1320 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1325 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1328 static struct pnfs_layoutdriver_type filelayout_type = {
1329 .id = LAYOUT_NFSV4_1_FILES,
1330 .name = "LAYOUT_NFSV4_1_FILES",
1331 .owner = THIS_MODULE,
1332 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1333 .free_layout_hdr = filelayout_free_layout_hdr,
1334 .alloc_lseg = filelayout_alloc_lseg,
1335 .free_lseg = filelayout_free_lseg,
1336 .pg_read_ops = &filelayout_pg_read_ops,
1337 .pg_write_ops = &filelayout_pg_write_ops,
1338 .get_ds_info = &filelayout_get_ds_info,
1339 .mark_request_commit = filelayout_mark_request_commit,
1340 .clear_request_commit = filelayout_clear_request_commit,
1341 .scan_commit_lists = filelayout_scan_commit_lists,
1342 .recover_commit_reqs = filelayout_recover_commit_reqs,
1343 .commit_pagelist = filelayout_commit_pagelist,
1344 .read_pagelist = filelayout_read_pagelist,
1345 .write_pagelist = filelayout_write_pagelist,
1346 .free_deviceid_node = filelayout_free_deveiceid_node,
1349 static int __init nfs4filelayout_init(void)
1351 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1353 return pnfs_register_layoutdriver(&filelayout_type);
1356 static void __exit nfs4filelayout_exit(void)
1358 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1360 pnfs_unregister_layoutdriver(&filelayout_type);
1363 MODULE_ALIAS("nfs-layouttype4-1");
1365 module_init(nfs4filelayout_init);
1366 module_exit(nfs4filelayout_exit);