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>
39 #include "delegation.h"
40 #include "nfs4filelayout.h"
42 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
44 MODULE_LICENSE("GPL");
45 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
46 MODULE_DESCRIPTION("The NFSv4 file layout driver");
48 #define FILELAYOUT_POLL_RETRY_MAX (15*HZ)
51 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
54 u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
58 offset -= flseg->pattern_offset;
59 stripe_no = div_u64(offset, stripe_width);
60 div_u64_rem(offset, flseg->stripe_unit, &rem);
62 return stripe_no * flseg->stripe_unit + rem;
65 /* This function is used by the layout driver to calculate the
66 * offset of the file on the dserver based on whether the
67 * layout type is STRIPE_DENSE or STRIPE_SPARSE
70 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
72 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
74 switch (flseg->stripe_type) {
79 return filelayout_get_dense_offset(flseg, offset);
85 static int filelayout_async_handle_error(struct rpc_task *task,
86 struct nfs4_state *state,
87 struct nfs_client *clp,
90 struct nfs_server *mds_server = NFS_SERVER(state->inode);
91 struct nfs_client *mds_client = mds_server->nfs_client;
93 if (task->tk_status >= 0)
97 switch (task->tk_status) {
98 /* MDS state errors */
99 case -NFS4ERR_DELEG_REVOKED:
100 case -NFS4ERR_ADMIN_REVOKED:
101 case -NFS4ERR_BAD_STATEID:
102 nfs_remove_bad_delegation(state->inode);
103 case -NFS4ERR_OPENMODE:
104 nfs4_schedule_stateid_recovery(mds_server, state);
105 goto wait_on_recovery;
106 case -NFS4ERR_EXPIRED:
107 nfs4_schedule_stateid_recovery(mds_server, state);
108 nfs4_schedule_lease_recovery(mds_client);
109 goto wait_on_recovery;
110 /* DS session errors */
111 case -NFS4ERR_BADSESSION:
112 case -NFS4ERR_BADSLOT:
113 case -NFS4ERR_BAD_HIGH_SLOT:
114 case -NFS4ERR_DEADSESSION:
115 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
116 case -NFS4ERR_SEQ_FALSE_RETRY:
117 case -NFS4ERR_SEQ_MISORDERED:
118 dprintk("%s ERROR %d, Reset session. Exchangeid "
119 "flags 0x%x\n", __func__, task->tk_status,
120 clp->cl_exchange_flags);
121 nfs4_schedule_session_recovery(clp->cl_session);
126 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
128 case -NFS4ERR_RETRY_UNCACHED_REP:
131 dprintk("%s DS error. Retry through MDS %d\n", __func__,
140 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
141 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
142 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
146 /* NFS_PROTO call done callback routines */
148 static int filelayout_read_done_cb(struct rpc_task *task,
149 struct nfs_read_data *data)
153 dprintk("%s DS read\n", __func__);
155 if (filelayout_async_handle_error(task, data->args.context->state,
156 data->ds_clp, &reset) == -EAGAIN) {
157 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
158 __func__, data->ds_clp, data->ds_clp->cl_session);
160 nfs4_reset_read(task, data);
161 rpc_restart_call_prepare(task);
169 * We reference the rpc_cred of the first WRITE that triggers the need for
170 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
171 * rfc5661 is not clear about which credential should be used.
174 filelayout_set_layoutcommit(struct nfs_write_data *wdata)
176 struct nfs_pgio_header *hdr = wdata->header;
178 if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
179 wdata->res.verf->committed == NFS_FILE_SYNC)
182 pnfs_set_layoutcommit(wdata);
183 dprintk("%s ionde %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
184 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
188 * Call ops for the async read/write cases
189 * In the case of dense layouts, the offset needs to be reset to its
192 static void filelayout_read_prepare(struct rpc_task *task, void *data)
194 struct nfs_read_data *rdata = data;
196 rdata->read_done_cb = filelayout_read_done_cb;
198 if (nfs41_setup_sequence(rdata->ds_clp->cl_session,
199 &rdata->args.seq_args, &rdata->res.seq_res,
203 rpc_call_start(task);
206 static void filelayout_read_call_done(struct rpc_task *task, void *data)
208 struct nfs_read_data *rdata = data;
210 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
212 /* Note this may cause RPC to be resent */
213 rdata->header->mds_ops->rpc_call_done(task, data);
216 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
218 struct nfs_read_data *rdata = data;
220 rpc_count_iostats(task, NFS_SERVER(rdata->header->inode)->client->cl_metrics);
223 static void filelayout_read_release(void *data)
225 struct nfs_read_data *rdata = data;
227 rdata->header->mds_ops->rpc_release(data);
230 static int filelayout_write_done_cb(struct rpc_task *task,
231 struct nfs_write_data *data)
235 if (filelayout_async_handle_error(task, data->args.context->state,
236 data->ds_clp, &reset) == -EAGAIN) {
237 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
238 __func__, data->ds_clp, data->ds_clp->cl_session);
240 nfs4_reset_write(task, data);
241 rpc_restart_call_prepare(task);
245 filelayout_set_layoutcommit(data);
249 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
250 static void prepare_to_resend_writes(struct nfs_commit_data *data)
252 struct nfs_page *first = nfs_list_entry(data->pages.next);
254 data->task.tk_status = 0;
255 memcpy(data->verf.verifier, first->wb_verf.verifier,
256 sizeof(first->wb_verf.verifier));
257 data->verf.verifier[0]++; /* ensure verifier mismatch */
260 static int filelayout_commit_done_cb(struct rpc_task *task,
261 struct nfs_commit_data *data)
265 if (filelayout_async_handle_error(task, data->context->state,
266 data->ds_clp, &reset) == -EAGAIN) {
267 dprintk("%s calling restart ds_clp %p ds_clp->cl_session %p\n",
268 __func__, data->ds_clp, data->ds_clp->cl_session);
270 prepare_to_resend_writes(data);
272 rpc_restart_call_prepare(task);
279 static void filelayout_write_prepare(struct rpc_task *task, void *data)
281 struct nfs_write_data *wdata = data;
283 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
284 &wdata->args.seq_args, &wdata->res.seq_res,
288 rpc_call_start(task);
291 static void filelayout_write_call_done(struct rpc_task *task, void *data)
293 struct nfs_write_data *wdata = data;
295 /* Note this may cause RPC to be resent */
296 wdata->header->mds_ops->rpc_call_done(task, data);
299 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
301 struct nfs_write_data *wdata = data;
303 rpc_count_iostats(task, NFS_SERVER(wdata->header->inode)->client->cl_metrics);
306 static void filelayout_write_release(void *data)
308 struct nfs_write_data *wdata = data;
310 wdata->header->mds_ops->rpc_release(data);
313 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
315 struct nfs_commit_data *wdata = data;
317 if (nfs41_setup_sequence(wdata->ds_clp->cl_session,
318 &wdata->args.seq_args, &wdata->res.seq_res,
322 rpc_call_start(task);
325 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
327 struct nfs_commit_data *wdata = data;
329 /* Note this may cause RPC to be resent */
330 wdata->mds_ops->rpc_call_done(task, data);
333 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
335 struct nfs_commit_data *cdata = data;
337 rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
340 static void filelayout_commit_release(void *calldata)
342 struct nfs_commit_data *data = calldata;
344 data->completion_ops->completion(data);
345 put_lseg(data->lseg);
346 nfs_commitdata_release(data);
349 static const struct rpc_call_ops filelayout_read_call_ops = {
350 .rpc_call_prepare = filelayout_read_prepare,
351 .rpc_call_done = filelayout_read_call_done,
352 .rpc_count_stats = filelayout_read_count_stats,
353 .rpc_release = filelayout_read_release,
356 static const struct rpc_call_ops filelayout_write_call_ops = {
357 .rpc_call_prepare = filelayout_write_prepare,
358 .rpc_call_done = filelayout_write_call_done,
359 .rpc_count_stats = filelayout_write_count_stats,
360 .rpc_release = filelayout_write_release,
363 static const struct rpc_call_ops filelayout_commit_call_ops = {
364 .rpc_call_prepare = filelayout_commit_prepare,
365 .rpc_call_done = filelayout_write_commit_done,
366 .rpc_count_stats = filelayout_commit_count_stats,
367 .rpc_release = filelayout_commit_release,
370 static enum pnfs_try_status
371 filelayout_read_pagelist(struct nfs_read_data *data)
373 struct nfs_pgio_header *hdr = data->header;
374 struct pnfs_layout_segment *lseg = hdr->lseg;
375 struct nfs4_pnfs_ds *ds;
376 loff_t offset = data->args.offset;
381 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
382 __func__, hdr->inode->i_ino,
383 data->args.pgbase, (size_t)data->args.count, offset);
385 /* Retrieve the correct rpc_client for the byte range */
386 j = nfs4_fl_calc_j_index(lseg, offset);
387 idx = nfs4_fl_calc_ds_index(lseg, j);
388 ds = nfs4_fl_prepare_ds(lseg, idx);
390 return PNFS_NOT_ATTEMPTED;
391 dprintk("%s USE DS: %s\n", __func__, ds->ds_remotestr);
393 /* No multipath support. Use first DS */
394 data->ds_clp = ds->ds_clp;
395 fh = nfs4_fl_select_ds_fh(lseg, j);
399 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
400 data->mds_offset = offset;
402 /* Perform an asynchronous read to ds */
403 status = nfs_initiate_read(ds->ds_clp->cl_rpcclient, data,
404 &filelayout_read_call_ops, RPC_TASK_SOFTCONN);
406 return PNFS_ATTEMPTED;
409 /* Perform async writes. */
410 static enum pnfs_try_status
411 filelayout_write_pagelist(struct nfs_write_data *data, int sync)
413 struct nfs_pgio_header *hdr = data->header;
414 struct pnfs_layout_segment *lseg = hdr->lseg;
415 struct nfs4_pnfs_ds *ds;
416 loff_t offset = data->args.offset;
421 /* Retrieve the correct rpc_client for the byte range */
422 j = nfs4_fl_calc_j_index(lseg, offset);
423 idx = nfs4_fl_calc_ds_index(lseg, j);
424 ds = nfs4_fl_prepare_ds(lseg, idx);
426 return PNFS_NOT_ATTEMPTED;
427 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s\n", __func__,
428 hdr->inode->i_ino, sync, (size_t) data->args.count, offset,
431 data->write_done_cb = filelayout_write_done_cb;
432 data->ds_clp = ds->ds_clp;
433 fh = nfs4_fl_select_ds_fh(lseg, j);
437 * Get the file offset on the dserver. Set the write offset to
438 * this offset and save the original offset.
440 data->args.offset = filelayout_get_dserver_offset(lseg, offset);
442 /* Perform an asynchronous write */
443 status = nfs_initiate_write(ds->ds_clp->cl_rpcclient, data,
444 &filelayout_write_call_ops, sync,
447 return PNFS_ATTEMPTED;
451 * filelayout_check_layout()
453 * Make sure layout segment parameters are sane WRT the device.
454 * At this point no generic layer initialization of the lseg has occurred,
455 * and nothing has been added to the layout_hdr cache.
459 filelayout_check_layout(struct pnfs_layout_hdr *lo,
460 struct nfs4_filelayout_segment *fl,
461 struct nfs4_layoutget_res *lgr,
462 struct nfs4_deviceid *id,
465 struct nfs4_deviceid_node *d;
466 struct nfs4_file_layout_dsaddr *dsaddr;
467 int status = -EINVAL;
468 struct nfs_server *nfss = NFS_SERVER(lo->plh_inode);
470 dprintk("--> %s\n", __func__);
472 /* FIXME: remove this check when layout segment support is added */
473 if (lgr->range.offset != 0 ||
474 lgr->range.length != NFS4_MAX_UINT64) {
475 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
480 if (fl->pattern_offset > lgr->range.offset) {
481 dprintk("%s pattern_offset %lld too large\n",
482 __func__, fl->pattern_offset);
486 if (!fl->stripe_unit || fl->stripe_unit % PAGE_SIZE) {
487 dprintk("%s Invalid stripe unit (%u)\n",
488 __func__, fl->stripe_unit);
492 /* find and reference the deviceid */
493 d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode)->pnfs_curr_ld,
494 NFS_SERVER(lo->plh_inode)->nfs_client, id);
496 dsaddr = get_device_info(lo->plh_inode, id, gfp_flags);
500 dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
501 /* Found deviceid is being reaped */
502 if (test_bit(NFS_DEVICEID_INVALID, &dsaddr->id_node.flags))
507 if (fl->first_stripe_index >= dsaddr->stripe_count) {
508 dprintk("%s Bad first_stripe_index %u\n",
509 __func__, fl->first_stripe_index);
513 if ((fl->stripe_type == STRIPE_SPARSE &&
514 fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
515 (fl->stripe_type == STRIPE_DENSE &&
516 fl->num_fh != dsaddr->stripe_count)) {
517 dprintk("%s num_fh %u not valid for given packing\n",
518 __func__, fl->num_fh);
522 if (fl->stripe_unit % nfss->rsize || fl->stripe_unit % nfss->wsize) {
523 dprintk("%s Stripe unit (%u) not aligned with rsize %u "
524 "wsize %u\n", __func__, fl->stripe_unit, nfss->rsize,
530 dprintk("--> %s returns %d\n", __func__, status);
533 nfs4_fl_put_deviceid(dsaddr);
537 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
541 for (i = 0; i < fl->num_fh; i++) {
542 if (!fl->fh_array[i])
544 kfree(fl->fh_array[i]);
551 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
553 filelayout_free_fh_array(fl);
558 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
559 struct nfs4_filelayout_segment *fl,
560 struct nfs4_layoutget_res *lgr,
561 struct nfs4_deviceid *id,
564 struct xdr_stream stream;
566 struct page *scratch;
571 dprintk("%s: set_layout_map Begin\n", __func__);
573 scratch = alloc_page(gfp_flags);
577 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
578 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
580 /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
582 p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
586 memcpy(id, p, sizeof(*id));
587 p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
588 nfs4_print_deviceid(id);
590 nfl_util = be32_to_cpup(p++);
591 if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
592 fl->commit_through_mds = 1;
593 if (nfl_util & NFL4_UFLG_DENSE)
594 fl->stripe_type = STRIPE_DENSE;
596 fl->stripe_type = STRIPE_SPARSE;
597 fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
599 fl->first_stripe_index = be32_to_cpup(p++);
600 p = xdr_decode_hyper(p, &fl->pattern_offset);
601 fl->num_fh = be32_to_cpup(p++);
603 dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
604 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
607 /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
608 * Futher checking is done in filelayout_check_layout */
610 max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
613 if (fl->num_fh > 0) {
614 fl->fh_array = kzalloc(fl->num_fh * sizeof(struct nfs_fh *),
620 for (i = 0; i < fl->num_fh; i++) {
621 /* Do we want to use a mempool here? */
622 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
623 if (!fl->fh_array[i])
626 p = xdr_inline_decode(&stream, 4);
629 fl->fh_array[i]->size = be32_to_cpup(p++);
630 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
631 printk(KERN_ERR "NFS: Too big fh %d received %d\n",
632 i, fl->fh_array[i]->size);
636 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
639 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
640 dprintk("DEBUG: %s: fh len %d\n", __func__,
641 fl->fh_array[i]->size);
644 __free_page(scratch);
648 filelayout_free_fh_array(fl);
650 __free_page(scratch);
655 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
657 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
659 dprintk("--> %s\n", __func__);
660 nfs4_fl_put_deviceid(fl->dsaddr);
661 /* This assumes a single RW lseg */
662 if (lseg->pls_range.iomode == IOMODE_RW) {
663 struct nfs4_filelayout *flo;
665 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
666 flo->commit_info.nbuckets = 0;
667 kfree(flo->commit_info.buckets);
668 flo->commit_info.buckets = NULL;
670 _filelayout_free_lseg(fl);
674 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
675 struct nfs_commit_info *cinfo,
678 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
679 struct pnfs_commit_bucket *buckets;
682 if (fl->commit_through_mds)
684 if (cinfo->ds->nbuckets != 0) {
685 /* This assumes there is only one IOMODE_RW lseg. What
686 * we really want to do is have a layout_hdr level
687 * dictionary of <multipath_list4, fh> keys, each
688 * associated with a struct list_head, populated by calls
689 * to filelayout_write_pagelist().
694 size = (fl->stripe_type == STRIPE_SPARSE) ?
695 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
697 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
704 spin_lock(cinfo->lock);
705 if (cinfo->ds->nbuckets != 0)
708 cinfo->ds->buckets = buckets;
709 cinfo->ds->nbuckets = size;
710 for (i = 0; i < size; i++) {
711 INIT_LIST_HEAD(&buckets[i].written);
712 INIT_LIST_HEAD(&buckets[i].committing);
715 spin_unlock(cinfo->lock);
720 static struct pnfs_layout_segment *
721 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
722 struct nfs4_layoutget_res *lgr,
725 struct nfs4_filelayout_segment *fl;
727 struct nfs4_deviceid id;
729 dprintk("--> %s\n", __func__);
730 fl = kzalloc(sizeof(*fl), gfp_flags);
734 rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
735 if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
736 _filelayout_free_lseg(fl);
739 return &fl->generic_hdr;
743 * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
745 * return true : coalesce page
746 * return false : don't coalesce page
749 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
750 struct nfs_page *req)
752 u64 p_stripe, r_stripe;
755 if (!pnfs_generic_pg_test(pgio, prev, req) ||
756 !nfs_generic_pg_test(pgio, prev, req))
759 p_stripe = (u64)req_offset(prev);
760 r_stripe = (u64)req_offset(req);
761 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
763 do_div(p_stripe, stripe_unit);
764 do_div(r_stripe, stripe_unit);
766 return (p_stripe == r_stripe);
770 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
771 struct nfs_page *req)
773 BUG_ON(pgio->pg_lseg != NULL);
775 if (req->wb_offset != req->wb_pgbase) {
777 * Handling unaligned pages is difficult, because have to
778 * somehow split a req in two in certain cases in the
779 * pg.test code. Avoid this by just not using pnfs
782 nfs_pageio_reset_read_mds(pgio);
785 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
791 /* If no lseg, fall back to read through mds */
792 if (pgio->pg_lseg == NULL)
793 nfs_pageio_reset_read_mds(pgio);
797 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
798 struct nfs_page *req)
800 struct nfs_commit_info cinfo;
803 BUG_ON(pgio->pg_lseg != NULL);
805 if (req->wb_offset != req->wb_pgbase)
807 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
813 /* If no lseg, fall back to write through mds */
814 if (pgio->pg_lseg == NULL)
816 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
817 status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
819 put_lseg(pgio->pg_lseg);
820 pgio->pg_lseg = NULL;
825 nfs_pageio_reset_write_mds(pgio);
828 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
829 .pg_init = filelayout_pg_init_read,
830 .pg_test = filelayout_pg_test,
831 .pg_doio = pnfs_generic_pg_readpages,
834 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
835 .pg_init = filelayout_pg_init_write,
836 .pg_test = filelayout_pg_test,
837 .pg_doio = pnfs_generic_pg_writepages,
840 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
842 if (fl->stripe_type == STRIPE_SPARSE)
843 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
848 /* The generic layer is about to remove the req from the commit list.
849 * If this will make the bucket empty, it will need to put the lseg reference.
852 filelayout_clear_request_commit(struct nfs_page *req,
853 struct nfs_commit_info *cinfo)
855 struct pnfs_layout_segment *freeme = NULL;
857 spin_lock(cinfo->lock);
858 if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
860 cinfo->ds->nwritten--;
861 if (list_is_singular(&req->wb_list)) {
862 struct pnfs_commit_bucket *bucket;
864 bucket = list_first_entry(&req->wb_list,
865 struct pnfs_commit_bucket,
867 freeme = bucket->wlseg;
868 bucket->wlseg = NULL;
871 nfs_request_remove_commit_list(req, cinfo);
872 spin_unlock(cinfo->lock);
876 static struct list_head *
877 filelayout_choose_commit_list(struct nfs_page *req,
878 struct pnfs_layout_segment *lseg,
879 struct nfs_commit_info *cinfo)
881 struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
883 struct list_head *list;
884 struct pnfs_commit_bucket *buckets;
886 if (fl->commit_through_mds)
887 return &cinfo->mds->list;
889 /* Note that we are calling nfs4_fl_calc_j_index on each page
890 * that ends up being committed to a data server. An attractive
891 * alternative is to add a field to nfs_write_data and nfs_page
892 * to store the value calculated in filelayout_write_pagelist
893 * and just use that here.
895 j = nfs4_fl_calc_j_index(lseg, req_offset(req));
896 i = select_bucket_index(fl, j);
897 buckets = cinfo->ds->buckets;
898 list = &buckets[i].written;
899 if (list_empty(list)) {
900 /* Non-empty buckets hold a reference on the lseg. That ref
901 * is normally transferred to the COMMIT call and released
902 * there. It could also be released if the last req is pulled
903 * off due to a rewrite, in which case it will be done in
904 * filelayout_clear_request_commit
906 buckets[i].wlseg = get_lseg(lseg);
908 set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
909 cinfo->ds->nwritten++;
914 filelayout_mark_request_commit(struct nfs_page *req,
915 struct pnfs_layout_segment *lseg,
916 struct nfs_commit_info *cinfo)
918 struct list_head *list;
920 list = filelayout_choose_commit_list(req, lseg, cinfo);
921 nfs_request_add_commit_list(req, list, cinfo);
924 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
926 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
928 if (flseg->stripe_type == STRIPE_SPARSE)
931 return nfs4_fl_calc_ds_index(lseg, i);
934 static struct nfs_fh *
935 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
937 struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
939 if (flseg->stripe_type == STRIPE_SPARSE) {
940 if (flseg->num_fh == 1)
942 else if (flseg->num_fh == 0)
943 /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
946 return flseg->fh_array[i];
949 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
951 struct pnfs_layout_segment *lseg = data->lseg;
952 struct nfs4_pnfs_ds *ds;
956 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
957 ds = nfs4_fl_prepare_ds(lseg, idx);
959 prepare_to_resend_writes(data);
960 filelayout_commit_release(data);
963 dprintk("%s ino %lu, how %d\n", __func__, data->inode->i_ino, how);
964 data->commit_done_cb = filelayout_commit_done_cb;
965 data->ds_clp = ds->ds_clp;
966 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
969 return nfs_initiate_commit(ds->ds_clp->cl_rpcclient, data,
970 &filelayout_commit_call_ops, how,
975 transfer_commit_list(struct list_head *src, struct list_head *dst,
976 struct nfs_commit_info *cinfo, int max)
978 struct nfs_page *req, *tmp;
981 list_for_each_entry_safe(req, tmp, src, wb_list) {
982 if (!nfs_lock_request(req))
984 if (cond_resched_lock(cinfo->lock))
985 list_safe_reset_next(req, tmp, wb_list);
986 nfs_request_remove_commit_list(req, cinfo);
987 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
988 nfs_list_add_request(req, dst);
990 if ((ret == max) && !cinfo->dreq)
997 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
998 struct nfs_commit_info *cinfo,
1001 struct list_head *src = &bucket->written;
1002 struct list_head *dst = &bucket->committing;
1005 ret = transfer_commit_list(src, dst, cinfo, max);
1007 cinfo->ds->nwritten -= ret;
1008 cinfo->ds->ncommitting += ret;
1009 bucket->clseg = bucket->wlseg;
1010 if (list_empty(src))
1011 bucket->wlseg = NULL;
1013 get_lseg(bucket->clseg);
1018 /* Move reqs from written to committing lists, returning count of number moved.
1019 * Note called with cinfo->lock held.
1021 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1026 for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1027 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1035 /* Pull everything off the committing lists and dump into @dst */
1036 static void filelayout_recover_commit_reqs(struct list_head *dst,
1037 struct nfs_commit_info *cinfo)
1039 struct pnfs_commit_bucket *b;
1042 /* NOTE cinfo->lock is NOT held, relying on fact that this is
1043 * only called on single thread per dreq.
1044 * Can't take the lock because need to do put_lseg
1046 for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1047 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1048 BUG_ON(!list_empty(&b->written));
1053 cinfo->ds->nwritten = 0;
1057 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1059 struct pnfs_ds_commit_info *fl_cinfo;
1060 struct pnfs_commit_bucket *bucket;
1061 struct nfs_commit_data *data;
1063 unsigned int nreq = 0;
1065 fl_cinfo = cinfo->ds;
1066 bucket = fl_cinfo->buckets;
1067 for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1068 if (list_empty(&bucket->committing))
1070 data = nfs_commitdata_alloc();
1073 data->ds_commit_index = i;
1074 data->lseg = bucket->clseg;
1075 bucket->clseg = NULL;
1076 list_add(&data->pages, list);
1080 /* Clean up on error */
1081 for (j = i; j < fl_cinfo->nbuckets; j++, bucket++) {
1082 if (list_empty(&bucket->committing))
1084 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1085 put_lseg(bucket->clseg);
1086 bucket->clseg = NULL;
1088 /* Caller will clean up entries put on list */
1092 /* This follows nfs_commit_list pretty closely */
1094 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1095 int how, struct nfs_commit_info *cinfo)
1097 struct nfs_commit_data *data, *tmp;
1099 unsigned int nreq = 0;
1101 if (!list_empty(mds_pages)) {
1102 data = nfs_commitdata_alloc();
1105 list_add(&data->pages, &list);
1108 nfs_retry_commit(mds_pages, NULL, cinfo);
1111 nreq += alloc_ds_commits(cinfo, &list);
1114 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1118 atomic_add(nreq, &cinfo->mds->rpcs_out);
1120 list_for_each_entry_safe(data, tmp, &list, pages) {
1121 list_del_init(&data->pages);
1123 nfs_init_commit(data, mds_pages, NULL, cinfo);
1124 nfs_initiate_commit(NFS_CLIENT(inode), data,
1125 data->mds_ops, how, 0);
1127 struct pnfs_commit_bucket *buckets;
1129 buckets = cinfo->ds->buckets;
1130 nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1131 filelayout_initiate_commit(data, how);
1135 cinfo->ds->ncommitting = 0;
1136 return PNFS_ATTEMPTED;
1140 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1142 nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1145 static struct pnfs_layout_hdr *
1146 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1148 struct nfs4_filelayout *flo;
1150 flo = kzalloc(sizeof(*flo), gfp_flags);
1151 return &flo->generic_hdr;
1155 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1157 kfree(FILELAYOUT_FROM_HDR(lo));
1160 static struct pnfs_ds_commit_info *
1161 filelayout_get_ds_info(struct inode *inode)
1163 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1168 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1171 static struct pnfs_layoutdriver_type filelayout_type = {
1172 .id = LAYOUT_NFSV4_1_FILES,
1173 .name = "LAYOUT_NFSV4_1_FILES",
1174 .owner = THIS_MODULE,
1175 .alloc_layout_hdr = filelayout_alloc_layout_hdr,
1176 .free_layout_hdr = filelayout_free_layout_hdr,
1177 .alloc_lseg = filelayout_alloc_lseg,
1178 .free_lseg = filelayout_free_lseg,
1179 .pg_read_ops = &filelayout_pg_read_ops,
1180 .pg_write_ops = &filelayout_pg_write_ops,
1181 .get_ds_info = &filelayout_get_ds_info,
1182 .mark_request_commit = filelayout_mark_request_commit,
1183 .clear_request_commit = filelayout_clear_request_commit,
1184 .scan_commit_lists = filelayout_scan_commit_lists,
1185 .recover_commit_reqs = filelayout_recover_commit_reqs,
1186 .commit_pagelist = filelayout_commit_pagelist,
1187 .read_pagelist = filelayout_read_pagelist,
1188 .write_pagelist = filelayout_write_pagelist,
1189 .free_deviceid_node = filelayout_free_deveiceid_node,
1192 static int __init nfs4filelayout_init(void)
1194 printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1196 return pnfs_register_layoutdriver(&filelayout_type);
1199 static void __exit nfs4filelayout_exit(void)
1201 printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1203 pnfs_unregister_layoutdriver(&filelayout_type);
1206 MODULE_ALIAS("nfs-layouttype4-1");
1208 module_init(nfs4filelayout_init);
1209 module_exit(nfs4filelayout_exit);