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"
43 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
45 MODULE_LICENSE("GPL");
46 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
47 MODULE_DESCRIPTION("The NFSv4 file layout driver");
49 #define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
52 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
55 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
59 offset -= flseg->pattern_offset;
60 stripe_no = div_u64(offset, stripe_width);
61 div_u64_rem(offset, flseg->stripe_unit, &rem);
63 return stripe_no * flseg->stripe_unit + rem;
66 /* This function is used by the layout driver to calculate the
67 * offset of the file on the dserver based on whether the
68 * layout type is STRIPE_DENSE or STRIPE_SPARSE
71 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
73 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
75 switch (flseg->stripe_type) {
80 return filelayout_get_dense_offset(flseg, offset);
86 static void filelayout_reset_write(struct nfs_write_data *data)
88 struct nfs_pgio_header *hdr = data->header;
89 struct rpc_task *task = &data->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/%lld, %u bytes @ offset %llu)\n", __func__,
95 hdr->inode->i_sb->s_id,
96 (long long)NFS_FILEID(hdr->inode),
98 (unsigned long long)data->args.offset);
100 task->tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
102 hdr->completion_ops);
106 static void filelayout_reset_read(struct nfs_read_data *data)
108 struct nfs_pgio_header *hdr = data->header;
109 struct rpc_task *task = &data->task;
111 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
112 dprintk("%s Reset task %5u for i/o through MDS "
113 "(req %s/%lld, %u bytes @ offset %llu)\n", __func__,
115 hdr->inode->i_sb->s_id,
116 (long long)NFS_FILEID(hdr->inode),
118 (unsigned long long)data->args.offset);
120 task->tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
122 hdr->completion_ops);
126 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
128 if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
130 clear_bit(NFS_INO_LAYOUTCOMMIT, &NFS_I(inode)->flags);
131 pnfs_return_layout(inode);
134 static int filelayout_async_handle_error(struct rpc_task *task,
135 struct nfs4_state *state,
136 struct nfs_client *clp,
137 struct pnfs_layout_segment *lseg)
139 struct pnfs_layout_hdr *lo = lseg->pls_layout;
140 struct inode *inode = lo->plh_inode;
141 struct nfs_server *mds_server = NFS_SERVER(inode);
142 struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
143 struct nfs_client *mds_client = mds_server->nfs_client;
144 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
146 if (task->tk_status >= 0)
149 switch (task->tk_status) {
150 /* MDS state errors */
151 case -NFS4ERR_DELEG_REVOKED:
152 case -NFS4ERR_ADMIN_REVOKED:
153 case -NFS4ERR_BAD_STATEID:
156 nfs_remove_bad_delegation(state->inode);
157 case -NFS4ERR_OPENMODE:
160 nfs4_schedule_stateid_recovery(mds_server, state);
161 goto wait_on_recovery;
162 case -NFS4ERR_EXPIRED:
164 nfs4_schedule_stateid_recovery(mds_server, state);
165 nfs4_schedule_lease_recovery(mds_client);
166 goto wait_on_recovery;
167 /* DS session errors */
168 case -NFS4ERR_BADSESSION:
169 case -NFS4ERR_BADSLOT:
170 case -NFS4ERR_BAD_HIGH_SLOT:
171 case -NFS4ERR_DEADSESSION:
172 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
173 case -NFS4ERR_SEQ_FALSE_RETRY:
174 case -NFS4ERR_SEQ_MISORDERED:
175 dprintk("%s ERROR %d, Reset session. Exchangeid "
176 "flags 0x%x\n", __func__, task->tk_status,
177 clp->cl_exchange_flags);
178 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
183 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
185 case -NFS4ERR_RETRY_UNCACHED_REP:
187 /* Invalidate Layout errors */
188 case -NFS4ERR_PNFS_NO_LAYOUT:
189 case -ESTALE: /* mapped NFS4ERR_STALE */
190 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
191 case -EISDIR: /* mapped NFS4ERR_ISDIR */
192 case -NFS4ERR_FHEXPIRED:
193 case -NFS4ERR_WRONG_TYPE:
194 dprintk("%s Invalid layout error %d\n", __func__,
197 * Destroy layout so new i/o will get a new layout.
198 * Layout will not be destroyed until all current lseg
199 * references are put. Mark layout as invalid to resend failed
200 * i/o and all i/o waiting on the slot table to the MDS until
201 * layout is destroyed and a new valid layout is obtained.
203 pnfs_destroy_layout(NFS_I(inode));
204 rpc_wake_up(&tbl->slot_tbl_waitq);
206 /* RPC connection errors */
214 dprintk("%s DS connection error %d\n", __func__,
216 nfs4_mark_deviceid_unavailable(devid);
217 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
218 rpc_wake_up(&tbl->slot_tbl_waitq);
222 dprintk("%s Retry through MDS. Error %d\n", __func__,
224 return -NFS4ERR_RESET_TO_MDS;
230 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
231 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
232 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
236 /* NFS_PROTO call done callback routines */
238 static int filelayout_read_done_cb(struct rpc_task *task,
239 struct nfs_read_data *data)
241 struct nfs_pgio_header *hdr = data->header;
244 err = filelayout_async_handle_error(task, data->args.context->state,
245 data->ds_clp, hdr->lseg);
248 case -NFS4ERR_RESET_TO_MDS:
249 filelayout_reset_read(data);
250 return task->tk_status;
252 rpc_restart_call_prepare(task);
260 * We reference the rpc_cred of the first WRITE that triggers the need for
261 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
262 * rfc5661 is not clear about which credential should be used.
265 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
267 struct nfs_pgio_header *hdr = wdata->header;
269 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
270 wdata->res.verf->committed == NFS_FILE_SYNC)
273 pnfs_set_layoutcommit(wdata);
274 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
275 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
279 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
281 return filelayout_test_devid_invalid(node) ||
282 nfs4_test_deviceid_unavailable(node);
286 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
288 struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
290 return filelayout_test_devid_unavailable(node);
294 * Call ops for the async read/write cases
295 * In the case of dense layouts, the offset needs to be reset to its
298 static void filelayout_read_prepare(struct rpc_task *task, void *data)
300 struct nfs_read_data *rdata = data;
302 if (filelayout_reset_to_mds(rdata->header->lseg)) {
303 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
304 filelayout_reset_read(rdata);
308 rdata->read_done_cb = filelayout_read_done_cb;
310 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
311 &rdata->args.seq_args, &rdata->res.seq_res,
315 rpc_call_start(task);
318 static void filelayout_read_call_done(struct rpc_task *task, void *data)
320 struct nfs_read_data *rdata = data;
322 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
324 if (test_bit(NFS_IOHDR_REDO, &rdata->header->flags) &&
325 task->tk_status == 0)
328 /* Note this may cause RPC to be resent */
329 rdata->header->mds_ops->rpc_call_done(task, data);
332 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
334 struct nfs_read_data *rdata = data;
336 rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
339 static void filelayout_read_release(void *data)
341 struct nfs_read_data *rdata = data;
342 struct pnfs_layout_hdr *lo = rdata->header->lseg->pls_layout;
344 filelayout_fenceme(lo->plh_inode, lo);
345 nfs_put_client(rdata->ds_clp);
346 rdata->header->mds_ops->rpc_release(data);
349 static int filelayout_write_done_cb(struct rpc_task *task,
350 struct nfs_write_data *data)
352 struct nfs_pgio_header *hdr = data->header;
355 err = filelayout_async_handle_error(task, data->args.context->state,
356 data->ds_clp, hdr->lseg);
359 case -NFS4ERR_RESET_TO_MDS:
360 filelayout_reset_write(data);
361 return task->tk_status;
363 rpc_restart_call_prepare(task);
367 filelayout_set_layoutcommit(data);
371 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
372 static void prepare_to_resend_writes(struct nfs_commit_data *data)
374 struct nfs_page *first = nfs_list_entry(data->pages.next);
376 data->task.tk_status = 0;
377 memcpy(&data->verf.verifier, &first->wb_verf,
378 sizeof(data->verf.verifier));
379 data->verf.verifier.data[0]++; /* ensure verifier mismatch */
382 static int filelayout_commit_done_cb(struct rpc_task *task,
383 struct nfs_commit_data *data)
387 err = filelayout_async_handle_error(task, NULL, data->ds_clp,
391 case -NFS4ERR_RESET_TO_MDS:
392 prepare_to_resend_writes(data);
395 rpc_restart_call_prepare(task);
402 static void filelayout_write_prepare(struct rpc_task *task, void *data)
404 struct nfs_write_data *wdata = data;
406 if (filelayout_reset_to_mds(wdata->header->lseg)) {
407 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
408 filelayout_reset_write(wdata);
412 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
413 &wdata->args.seq_args, &wdata->res.seq_res,
417 rpc_call_start(task);
420 static void filelayout_write_call_done(struct rpc_task *task, void *data)
422 struct nfs_write_data *wdata = data;
424 if (test_bit(NFS_IOHDR_REDO, &wdata->header->flags) &&
425 task->tk_status == 0)
428 /* Note this may cause RPC to be resent */
429 wdata->header->mds_ops->rpc_call_done(task, data);
432 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
434 struct nfs_write_data *wdata = data;
436 rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
439 static void filelayout_write_release(void *data)
441 struct nfs_write_data *wdata = data;
442 struct pnfs_layout_hdr *lo = wdata->header->lseg->pls_layout;
444 filelayout_fenceme(lo->plh_inode, lo);
445 nfs_put_client(wdata->ds_clp);
446 wdata->header->mds_ops->rpc_release(data);
449 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
451 struct nfs_commit_data *wdata = data;
453 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
454 &wdata->args.seq_args, &wdata->res.seq_res,
458 rpc_call_start(task);
461 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
463 struct nfs_commit_data *wdata = data;
465 /* Note this may cause RPC to be resent */
466 wdata->mds_ops->rpc_call_done(task, data);
469 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
471 struct nfs_commit_data *cdata = data;
473 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
476 static void filelayout_commit_release(void *calldata)
478 struct nfs_commit_data *data = calldata;
480 data->completion_ops->completion(data);
481 pnfs_put_lseg(data->lseg);
482 nfs_put_client(data->ds_clp);
483 nfs_commitdata_release(data);
486 static const struct rpc_call_ops filelayout_read_call_ops = {
487 .rpc_call_prepare = filelayout_read_prepare,
488 .rpc_call_done = filelayout_read_call_done,
489 .rpc_count_stats = filelayout_read_count_stats,
490 .rpc_release = filelayout_read_release,
493 static const struct rpc_call_ops filelayout_write_call_ops = {
494 .rpc_call_prepare = filelayout_write_prepare,
495 .rpc_call_done = filelayout_write_call_done,
496 .rpc_count_stats = filelayout_write_count_stats,
497 .rpc_release = filelayout_write_release,
500 static const struct rpc_call_ops filelayout_commit_call_ops = {
501 .rpc_call_prepare = filelayout_commit_prepare,
502 .rpc_call_done = filelayout_write_commit_done,
503 .rpc_count_stats = filelayout_commit_count_stats,
504 .rpc_release = filelayout_commit_release,
507 static enum pnfs_try_status
508 filelayout_read_pagelist(struct nfs_read_data *data)
510 struct nfs_pgio_header *hdr = data->header;
511 struct pnfs_layout_segment *lseg = hdr->lseg;
512 struct nfs4_pnfs_ds *ds;
513 loff_t offset = data->args.offset;
517 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
518 __func__, hdr->inode->i_ino,
519 data->args.pgbase, (size_t)data->args.count, offset);
521 /* Retrieve the correct rpc_client for the byte range */
522 j = nfs4_fl_calc_j_index(lseg, offset);
523 idx = nfs4_fl_calc_ds_index(lseg, j);
524 ds = nfs4_fl_prepare_ds(lseg, idx);
526 return PNFS_NOT_ATTEMPTED;
527 dprintk("%s USE DS: %s cl_count %d\n", __func__,
528 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
530 /* No multipath support. Use first DS */
531 atomic_inc(&ds->ds_clp->cl_count);
532 data->ds_clp = ds->ds_clp;
533 fh = nfs4_fl_select_ds_fh(lseg, j);
537 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
538 data->mds_offset = offset;
540 /* Perform an asynchronous read to ds */
541 nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
542 &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
543 return PNFS_ATTEMPTED;
546 /* Perform async writes. */
547 static enum pnfs_try_status
548 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
550 struct nfs_pgio_header *hdr = data->header;
551 struct pnfs_layout_segment *lseg = hdr->lseg;
552 struct nfs4_pnfs_ds *ds;
553 loff_t offset = data->args.offset;
557 /* Retrieve the correct rpc_client for the byte range */
558 j = nfs4_fl_calc_j_index(lseg, offset);
559 idx = nfs4_fl_calc_ds_index(lseg, j);
560 ds = nfs4_fl_prepare_ds(lseg, idx);
562 return PNFS_NOT_ATTEMPTED;
563 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
564 __func__, hdr->inode->i_ino, sync, (size_t) data->args.count,
565 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
567 data->write_done_cb = filelayout_write_done_cb;
568 atomic_inc(&ds->ds_clp->cl_count);
569 data->ds_clp = ds->ds_clp;
570 fh = nfs4_fl_select_ds_fh(lseg, j);
574 * Get the file offset on the dserver. Set the write offset to
575 * this offset and save the original offset.
577 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
579 /* Perform an asynchronous write */
580 nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
581 &filelayout_write_call_ops, sync,
583 return PNFS_ATTEMPTED;
587 * filelayout_check_layout()
589 * Make sure layout segment parameters are sane WRT the device.
590 * At this point no generic layer initialization of the lseg has occurred,
591 * and nothing has been added to the layout_hdr cache.
595 filelayout_check_layout(struct pnfs_layout_hdr *lo,
596 struct nfs4_filelayout_segment *fl,
597 struct nfs4_layoutget_res *lgr,
598 struct nfs4_deviceid *id,
601 struct nfs4_deviceid_node *d;
602 struct nfs4_file_layout_dsaddr *dsaddr;
603 int status = -EINVAL;
604 struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
606 dprintk("--> %s\n", __func__);
608 /* FIXME: remove this check when layout segment support is added */
609 if (lgr->range.offset != 0 ||
610 lgr->range.length != NFS4_MAX_UINT64) {
611 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
616 if (fl->pattern_offset > lgr->range.offset) {
617 dprintk("%s pattern_offset %lld too large\n",
618 __func__, fl->pattern_offset);
622 if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
623 dprintk("%s Invalid stripe unit (%u)\n",
624 __func__, fl->stripe_unit);
628 /* find and reference the deviceid */
629 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
630 NFS_SERVER(lo->plh_inode)->nfs_client, id);
632 dsaddr = filelayout_get_device_info(lo->plh_inode, id, gfp_flags);
636 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
637 /* Found deviceid is unavailable */
638 if (filelayout_test_devid_unavailable(&dsaddr->id_node))
643 if (fl->first_stripe_index >= dsaddr->stripe_count) {
644 dprintk("%s Bad first_stripe_index %u\n",
645 __func__, fl->first_stripe_index);
649 if ((fl->stripe_type == STRIPE_SPARSE &&
650 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
651 (fl->stripe_type == STRIPE_DENSE &&
652 fl->num_fh != dsaddr->stripe_count)) {
653 dprintk("%s num_fh %u not valid for given packing\n",
654 __func__, fl->num_fh);
658 if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
659 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
660 "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
666 dprintk("--> %s returns %d\n", __func__, status);
669 nfs4_fl_put_deviceid(dsaddr);
673 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
677 for (i = 0; i < fl->num_fh; i++) {
678 if (!fl->fh_array[i])
680 kfree(fl->fh_array[i]);
687 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
689 filelayout_free_fh_array(fl);
694 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
695 struct nfs4_filelayout_segment *fl,
696 struct nfs4_layoutget_res *lgr,
697 struct nfs4_deviceid *id,
700 struct xdr_stream stream;
702 struct page *scratch;
707 dprintk("%s: set_layout_map Begin\n", __func__);
709 scratch = alloc_page(gfp_flags);
713 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
714 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
716 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
718 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
722 memcpy(id, p, sizeof(*id));
723 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
724 nfs4_print_deviceid(id);
726 nfl_util = be32_to_cpup(p++);
727 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
728 fl->commit_through_mds = 1;
729 if (nfl_util & NFL4_UFLG_DENSE)
730 fl->stripe_type = STRIPE_DENSE;
732 fl->stripe_type = STRIPE_SPARSE;
733 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
735 fl->first_stripe_index = be32_to_cpup(p++);
736 p = xdr_decode_hyper(p, &fl->pattern_offset);
737 fl->num_fh = be32_to_cpup(p++);
739 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
740 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
743 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
744 * Futher checking is done in filelayout_check_layout */
746 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
749 if (fl->num_fh > 0) {
750 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
756 for (i = 0; i < fl->num_fh; i++) {
757 /* Do we want to use a mempool here? */
758 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
759 if (!fl->fh_array[i])
762 p = xdr_inline_decode(&stream, 4);
765 fl->fh_array[i]->size = be32_to_cpup(p++);
766 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
767 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
768 i, fl->fh_array[i]->size);
772 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
775 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
776 dprintk("DEBUG: %s: fh len %d\n", __func__,
777 fl->fh_array[i]->size);
780 __free_page(scratch);
784 filelayout_free_fh_array(fl);
786 __free_page(scratch);
791 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
793 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
795 dprintk("--> %s\n", __func__);
796 nfs4_fl_put_deviceid(fl->dsaddr);
797 /* This assumes a single RW lseg */
798 if (lseg->pls_range.iomode == IOMODE_RW) {
799 struct nfs4_filelayout *flo;
801 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
802 flo->commit_info.nbuckets = 0;
803 kfree(flo->commit_info.buckets);
804 flo->commit_info.buckets = NULL;
806 _filelayout_free_lseg(fl);
810 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
811 struct nfs_commit_info *cinfo,
814 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
815 struct pnfs_commit_bucket *buckets;
818 if (fl->commit_through_mds)
820 if (cinfo->ds->nbuckets != 0) {
821 /* This assumes there is only one IOMODE_RW lseg. What
822 * we really want to do is have a layout_hdr level
823 * dictionary of <multipath_list4, fh> keys, each
824 * associated with a struct list_head, populated by calls
825 * to filelayout_write_pagelist().
830 size = (fl->stripe_type == STRIPE_SPARSE) ?
831 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
833 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
840 spin_lock(cinfo->lock);
841 if (cinfo->ds->nbuckets != 0)
844 cinfo->ds->buckets = buckets;
845 cinfo->ds->nbuckets = size;
846 for (i = 0; i < size; i++) {
847 INIT_LIST_HEAD(&buckets[i].written);
848 INIT_LIST_HEAD(&buckets[i].committing);
851 spin_unlock(cinfo->lock);
856 static struct pnfs_layout_segment *
857 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
858 struct nfs4_layoutget_res *lgr,
861 struct nfs4_filelayout_segment *fl;
863 struct nfs4_deviceid id;
865 dprintk("--> %s\n", __func__);
866 fl = kzalloc(sizeof(*fl), gfp_flags);
870 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
871 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
872 _filelayout_free_lseg(fl);
875 return &fl->generic_hdr;
879 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
881 * return true : coalesce page
882 * return false : don't coalesce page
885 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
886 struct nfs_page *req)
888 u64 p_stripe, r_stripe;
891 if (!pnfs_generic_pg_test(pgio, prev, req) ||
892 !nfs_generic_pg_test(pgio, prev, req))
895 p_stripe = (u64)req_offset(prev);
896 r_stripe = (u64)req_offset(req);
897 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
899 do_div(p_stripe, stripe_unit);
900 do_div(r_stripe, stripe_unit);
902 return (p_stripe == r_stripe);
906 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
907 struct nfs_page *req)
909 WARN_ON_ONCE(pgio->pg_lseg != NULL);
911 if (req->wb_offset != req->wb_pgbase) {
913 * Handling unaligned pages is difficult, because have to
914 * somehow split a req in two in certain cases in the
915 * pg.test code. Avoid this by just not using pnfs
918 nfs_pageio_reset_read_mds(pgio);
921 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
927 /* If no lseg, fall back to read through mds */
928 if (pgio->pg_lseg == NULL)
929 nfs_pageio_reset_read_mds(pgio);
933 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
934 struct nfs_page *req)
936 struct nfs_commit_info cinfo;
939 WARN_ON_ONCE(pgio->pg_lseg != NULL);
941 if (req->wb_offset != req->wb_pgbase)
943 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
949 /* If no lseg, fall back to write through mds */
950 if (pgio->pg_lseg == NULL)
952 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
953 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
955 pnfs_put_lseg(pgio->pg_lseg);
956 pgio->pg_lseg = NULL;
961 nfs_pageio_reset_write_mds(pgio);
964 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
965 .pg_init = filelayout_pg_init_read,
966 .pg_test = filelayout_pg_test,
967 .pg_doio = pnfs_generic_pg_readpages,
970 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
971 .pg_init = filelayout_pg_init_write,
972 .pg_test = filelayout_pg_test,
973 .pg_doio = pnfs_generic_pg_writepages,
976 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
978 if (fl->stripe_type == STRIPE_SPARSE)
979 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
984 /* The generic layer is about to remove the req from the commit list.
985 * If this will make the bucket empty, it will need to put the lseg reference.
988 filelayout_clear_request_commit(struct nfs_page *req,
989 struct nfs_commit_info *cinfo)
991 struct pnfs_layout_segment *freeme = NULL;
993 spin_lock(cinfo->lock);
994 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
996 cinfo->ds->nwritten--;
997 if (list_is_singular(&req->wb_list)) {
998 struct pnfs_commit_bucket *bucket;
1000 bucket = list_first_entry(&req->wb_list,
1001 struct pnfs_commit_bucket,
1003 freeme = bucket->wlseg;
1004 bucket->wlseg = NULL;
1007 nfs_request_remove_commit_list(req, cinfo);
1008 spin_unlock(cinfo->lock);
1009 pnfs_put_lseg(freeme);
1012 static struct list_head *
1013 filelayout_choose_commit_list(struct nfs_page *req,
1014 struct pnfs_layout_segment *lseg,
1015 struct nfs_commit_info *cinfo)
1017 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1019 struct list_head *list;
1020 struct pnfs_commit_bucket *buckets;
1022 if (fl->commit_through_mds)
1023 return &cinfo->mds->list;
1025 /* Note that we are calling nfs4_fl_calc_j_index on each page
1026 * that ends up being committed to a data server. An attractive
1027 * alternative is to add a field to nfs_write_data and nfs_page
1028 * to store the value calculated in filelayout_write_pagelist
1029 * and just use that here.
1031 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1032 i = select_bucket_index(fl, j);
1033 buckets = cinfo->ds->buckets;
1034 list = &buckets[i].written;
1035 if (list_empty(list)) {
1036 /* Non-empty buckets hold a reference on the lseg. That ref
1037 * is normally transferred to the COMMIT call and released
1038 * there. It could also be released if the last req is pulled
1039 * off due to a rewrite, in which case it will be done in
1040 * filelayout_clear_request_commit
1042 buckets[i].wlseg = pnfs_get_lseg(lseg);
1044 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1045 cinfo->ds->nwritten++;
1050 filelayout_mark_request_commit(struct nfs_page *req,
1051 struct pnfs_layout_segment *lseg,
1052 struct nfs_commit_info *cinfo)
1054 struct list_head *list;
1056 list = filelayout_choose_commit_list(req, lseg, cinfo);
1057 nfs_request_add_commit_list(req, list, cinfo);
1060 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1062 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1064 if (flseg->stripe_type == STRIPE_SPARSE)
1067 return nfs4_fl_calc_ds_index(lseg, i);
1070 static struct nfs_fh *
1071 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1073 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1075 if (flseg->stripe_type == STRIPE_SPARSE) {
1076 if (flseg->num_fh == 1)
1078 else if (flseg->num_fh == 0)
1079 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1082 return flseg->fh_array[i];
1085 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1087 struct pnfs_layout_segment *lseg = data->lseg;
1088 struct nfs4_pnfs_ds *ds;
1092 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1093 ds = nfs4_fl_prepare_ds(lseg, idx);
1095 prepare_to_resend_writes(data);
1096 filelayout_commit_release(data);
1099 dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1100 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1101 data->commit_done_cb = filelayout_commit_done_cb;
1102 atomic_inc(&ds->ds_clp->cl_count);
1103 data->ds_clp = ds->ds_clp;
1104 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1107 return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
1108 &filelayout_commit_call_ops, how,
1113 transfer_commit_list(struct list_head *src, struct list_head *dst,
1114 struct nfs_commit_info *cinfo, int max)
1116 struct nfs_page *req, *tmp;
1119 list_for_each_entry_safe(req, tmp, src, wb_list) {
1120 if (!nfs_lock_request(req))
1122 kref_get(&req->wb_kref);
1123 if (cond_resched_lock(cinfo->lock))
1124 list_safe_reset_next(req, tmp, wb_list);
1125 nfs_request_remove_commit_list(req, cinfo);
1126 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1127 nfs_list_add_request(req, dst);
1129 if ((ret == max) && !cinfo->dreq)
1136 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1137 struct nfs_commit_info *cinfo,
1140 struct list_head *src = &bucket->written;
1141 struct list_head *dst = &bucket->committing;
1144 ret = transfer_commit_list(src, dst, cinfo, max);
1146 cinfo->ds->nwritten -= ret;
1147 cinfo->ds->ncommitting += ret;
1148 bucket->clseg = bucket->wlseg;
1149 if (list_empty(src))
1150 bucket->wlseg = NULL;
1152 pnfs_get_lseg(bucket->clseg);
1157 /* Move reqs from written to committing lists, returning count of number moved.
1158 * Note called with cinfo->lock held.
1160 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1165 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1166 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1174 /* Pull everything off the committing lists and dump into @dst */
1175 static void filelayout_recover_commit_reqs(struct list_head *dst,
1176 struct nfs_commit_info *cinfo)
1178 struct pnfs_commit_bucket *b;
1181 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1182 * only called on single thread per dreq.
1183 * Can't take the lock because need to do pnfs_put_lseg
1185 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1186 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1187 pnfs_put_lseg(b->wlseg);
1191 cinfo->ds->nwritten = 0;
1195 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1197 struct pnfs_ds_commit_info *fl_cinfo;
1198 struct pnfs_commit_bucket *bucket;
1199 struct nfs_commit_data *data;
1201 unsigned int nreq = 0;
1203 fl_cinfo = cinfo->ds;
1204 bucket = fl_cinfo->buckets;
1205 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1206 if (list_empty(&bucket->committing))
1208 data = nfs_commitdata_alloc();
1211 data->ds_commit_index = i;
1212 data->lseg = bucket->clseg;
1213 bucket->clseg = NULL;
1214 list_add(&data->pages, list);
1218 /* Clean up on error */
1219 for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1220 if (list_empty(&bucket->committing))
1222 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1223 pnfs_put_lseg(bucket->clseg);
1224 bucket->clseg = NULL;
1226 /* Caller will clean up entries put on list */
1230 /* This follows nfs_commit_list pretty closely */
1232 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1233 int how, struct nfs_commit_info *cinfo)
1235 struct nfs_commit_data *data, *tmp;
1237 unsigned int nreq = 0;
1239 if (!list_empty(mds_pages)) {
1240 data = nfs_commitdata_alloc();
1243 list_add(&data->pages, &list);
1246 nfs_retry_commit(mds_pages, NULL, cinfo);
1249 nreq += alloc_ds_commits(cinfo, &list);
1252 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1256 atomic_add(nreq, &cinfo->mds->rpcs_out);
1258 list_for_each_entry_safe(data, tmp, &list, pages) {
1259 list_del_init(&data->pages);
1261 nfs_init_commit(data, mds_pages, NULL, cinfo);
1262 nfs_initiate_commit(NFS_CLIENT(inode), data,
1263 data->mds_ops, how, 0);
1265 struct pnfs_commit_bucket *buckets;
1267 buckets = cinfo->ds->buckets;
1268 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1269 filelayout_initiate_commit(data, how);
1273 cinfo->ds->ncommitting = 0;
1274 return PNFS_ATTEMPTED;
1278 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1280 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1283 static struct pnfs_layout_hdr *
1284 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1286 struct nfs4_filelayout *flo;
1288 flo = kzalloc(sizeof(*flo), gfp_flags);
1289 return &flo->generic_hdr;
1293 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1295 kfree(FILELAYOUT_FROM_HDR(lo));
1298 static struct pnfs_ds_commit_info *
1299 filelayout_get_ds_info(struct inode *inode)
1301 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1306 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1309 static struct pnfs_layoutdriver_type filelayout_type = {
1310 .id = LAYOUT_NFSV4_1_FILES,
1311 .name = "LAYOUT_NFSV4_1_FILES",
1312 .owner = THIS_MODULE,
1313 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1314 .free_layout_hdr = filelayout_free_layout_hdr,
1315 .alloc_lseg = filelayout_alloc_lseg,
1316 .free_lseg = filelayout_free_lseg,
1317 .pg_read_ops = &filelayout_pg_read_ops,
1318 .pg_write_ops = &filelayout_pg_write_ops,
1319 .get_ds_info = &filelayout_get_ds_info,
1320 .mark_request_commit = filelayout_mark_request_commit,
1321 .clear_request_commit = filelayout_clear_request_commit,
1322 .scan_commit_lists = filelayout_scan_commit_lists,
1323 .recover_commit_reqs = filelayout_recover_commit_reqs,
1324 .commit_pagelist = filelayout_commit_pagelist,
1325 .read_pagelist = filelayout_read_pagelist,
1326 .write_pagelist = filelayout_write_pagelist,
1327 .free_deviceid_node = filelayout_free_deveiceid_node,
1330 static int __init nfs4filelayout_init(void)
1332 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1334 return pnfs_register_layoutdriver(&filelayout_type);
1337 static void __exit nfs4filelayout_exit(void)
1339 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1341 pnfs_unregister_layoutdriver(&filelayout_type);
1344 MODULE_ALIAS("nfs-layouttype4-1");
1346 module_init(nfs4filelayout_init);
1347 module_exit(nfs4filelayout_exit);