2 * Module for pnfs flexfile layout driver.
4 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
6 * Tao Peng <bergwolf@primarydata.com>
9 #include <linux/nfs_fs.h>
10 #include <linux/nfs_page.h>
11 #include <linux/module.h>
13 #include <linux/sunrpc/metrics.h>
15 #include "flexfilelayout.h"
16 #include "../nfs4session.h"
17 #include "../nfs4idmap.h"
18 #include "../internal.h"
19 #include "../delegation.h"
20 #include "../nfs4trace.h"
21 #include "../iostat.h"
25 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
27 #define FF_LAYOUT_POLL_RETRY_MAX (15*HZ)
29 static struct pnfs_layout_hdr *
30 ff_layout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
32 struct nfs4_flexfile_layout *ffl;
34 ffl = kzalloc(sizeof(*ffl), gfp_flags);
36 INIT_LIST_HEAD(&ffl->error_list);
37 INIT_LIST_HEAD(&ffl->mirrors);
38 return &ffl->generic_hdr;
44 ff_layout_free_layout_hdr(struct pnfs_layout_hdr *lo)
46 struct nfs4_ff_layout_ds_err *err, *n;
48 list_for_each_entry_safe(err, n, &FF_LAYOUT_FROM_HDR(lo)->error_list,
53 kfree(FF_LAYOUT_FROM_HDR(lo));
56 static int decode_stateid(struct xdr_stream *xdr, nfs4_stateid *stateid)
60 p = xdr_inline_decode(xdr, NFS4_STATEID_SIZE);
61 if (unlikely(p == NULL))
63 memcpy(stateid, p, NFS4_STATEID_SIZE);
64 dprintk("%s: stateid id= [%x%x%x%x]\n", __func__,
65 p[0], p[1], p[2], p[3]);
69 static int decode_deviceid(struct xdr_stream *xdr, struct nfs4_deviceid *devid)
73 p = xdr_inline_decode(xdr, NFS4_DEVICEID4_SIZE);
76 memcpy(devid, p, NFS4_DEVICEID4_SIZE);
77 nfs4_print_deviceid(devid);
81 static int decode_nfs_fh(struct xdr_stream *xdr, struct nfs_fh *fh)
85 p = xdr_inline_decode(xdr, 4);
88 fh->size = be32_to_cpup(p++);
89 if (fh->size > sizeof(struct nfs_fh)) {
90 printk(KERN_ERR "NFS flexfiles: Too big fh received %d\n",
95 p = xdr_inline_decode(xdr, fh->size);
98 memcpy(&fh->data, p, fh->size);
99 dprintk("%s: fh len %d\n", __func__, fh->size);
105 * Currently only stringified uids and gids are accepted.
106 * I.e., kerberos is not supported to the DSes, so no pricipals.
108 * That means that one common function will suffice, but when
109 * principals are added, this should be split to accomodate
110 * calls to both nfs_map_name_to_uid() and nfs_map_group_to_gid().
113 decode_name(struct xdr_stream *xdr, u32 *id)
118 /* opaque_length(4)*/
119 p = xdr_inline_decode(xdr, 4);
122 len = be32_to_cpup(p++);
126 dprintk("%s: len %u\n", __func__, len);
129 p = xdr_inline_decode(xdr, len);
133 if (!nfs_map_string_to_numeric((char *)p, len, id))
139 static bool ff_mirror_match_fh(const struct nfs4_ff_layout_mirror *m1,
140 const struct nfs4_ff_layout_mirror *m2)
144 if (m1->fh_versions_cnt != m2->fh_versions_cnt)
146 for (i = 0; i < m1->fh_versions_cnt; i++) {
147 bool found_fh = false;
148 for (j = 0; j < m2->fh_versions_cnt; i++) {
149 if (nfs_compare_fh(&m1->fh_versions[i],
150 &m2->fh_versions[j]) == 0) {
161 static struct nfs4_ff_layout_mirror *
162 ff_layout_add_mirror(struct pnfs_layout_hdr *lo,
163 struct nfs4_ff_layout_mirror *mirror)
165 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
166 struct nfs4_ff_layout_mirror *pos;
167 struct inode *inode = lo->plh_inode;
169 spin_lock(&inode->i_lock);
170 list_for_each_entry(pos, &ff_layout->mirrors, mirrors) {
171 if (mirror->mirror_ds != pos->mirror_ds)
173 if (!ff_mirror_match_fh(mirror, pos))
175 if (atomic_inc_not_zero(&pos->ref)) {
176 spin_unlock(&inode->i_lock);
180 list_add(&mirror->mirrors, &ff_layout->mirrors);
182 spin_unlock(&inode->i_lock);
187 ff_layout_remove_mirror(struct nfs4_ff_layout_mirror *mirror)
190 if (mirror->layout == NULL)
192 inode = mirror->layout->plh_inode;
193 spin_lock(&inode->i_lock);
194 list_del(&mirror->mirrors);
195 spin_unlock(&inode->i_lock);
196 mirror->layout = NULL;
199 static struct nfs4_ff_layout_mirror *ff_layout_alloc_mirror(gfp_t gfp_flags)
201 struct nfs4_ff_layout_mirror *mirror;
203 mirror = kzalloc(sizeof(*mirror), gfp_flags);
204 if (mirror != NULL) {
205 spin_lock_init(&mirror->lock);
206 atomic_set(&mirror->ref, 1);
207 INIT_LIST_HEAD(&mirror->mirrors);
212 static void ff_layout_free_mirror(struct nfs4_ff_layout_mirror *mirror)
214 ff_layout_remove_mirror(mirror);
215 kfree(mirror->fh_versions);
217 put_rpccred(mirror->cred);
218 nfs4_ff_layout_put_deviceid(mirror->mirror_ds);
222 static void ff_layout_put_mirror(struct nfs4_ff_layout_mirror *mirror)
224 if (mirror != NULL && atomic_dec_and_test(&mirror->ref))
225 ff_layout_free_mirror(mirror);
228 static void ff_layout_free_mirror_array(struct nfs4_ff_layout_segment *fls)
232 if (fls->mirror_array) {
233 for (i = 0; i < fls->mirror_array_cnt; i++) {
234 /* normally mirror_ds is freed in
235 * .free_deviceid_node but we still do it here
236 * for .alloc_lseg error path */
237 ff_layout_put_mirror(fls->mirror_array[i]);
239 kfree(fls->mirror_array);
240 fls->mirror_array = NULL;
244 static int ff_layout_check_layout(struct nfs4_layoutget_res *lgr)
248 dprintk("--> %s\n", __func__);
250 /* FIXME: remove this check when layout segment support is added */
251 if (lgr->range.offset != 0 ||
252 lgr->range.length != NFS4_MAX_UINT64) {
253 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
258 dprintk("--> %s returns %d\n", __func__, ret);
262 static void _ff_layout_free_lseg(struct nfs4_ff_layout_segment *fls)
265 ff_layout_free_mirror_array(fls);
271 ff_lseg_range_is_after(const struct pnfs_layout_range *l1,
272 const struct pnfs_layout_range *l2)
276 if (l1->iomode != l2->iomode)
277 return l1->iomode != IOMODE_READ;
278 end1 = pnfs_calc_offset_end(l1->offset, l1->length);
279 end2 = pnfs_calc_offset_end(l2->offset, l2->length);
280 if (end1 < l2->offset)
282 if (end2 < l1->offset)
284 return l2->offset <= l1->offset;
288 ff_lseg_merge(struct pnfs_layout_segment *new,
289 struct pnfs_layout_segment *old)
291 u64 new_end, old_end;
293 if (new->pls_range.iomode != old->pls_range.iomode)
295 old_end = pnfs_calc_offset_end(old->pls_range.offset,
296 old->pls_range.length);
297 if (old_end < new->pls_range.offset)
299 new_end = pnfs_calc_offset_end(new->pls_range.offset,
300 new->pls_range.length);
301 if (new_end < old->pls_range.offset)
304 /* Mergeable: copy info from 'old' to 'new' */
305 if (new_end < old_end)
307 if (new->pls_range.offset < old->pls_range.offset)
308 new->pls_range.offset = old->pls_range.offset;
309 new->pls_range.length = pnfs_calc_offset_length(new->pls_range.offset,
311 if (test_bit(NFS_LSEG_ROC, &old->pls_flags))
312 set_bit(NFS_LSEG_ROC, &new->pls_flags);
313 if (test_bit(NFS_LSEG_LAYOUTRETURN, &old->pls_flags))
314 set_bit(NFS_LSEG_LAYOUTRETURN, &new->pls_flags);
319 ff_layout_add_lseg(struct pnfs_layout_hdr *lo,
320 struct pnfs_layout_segment *lseg,
321 struct list_head *free_me)
323 pnfs_generic_layout_insert_lseg(lo, lseg,
324 ff_lseg_range_is_after,
329 static void ff_layout_sort_mirrors(struct nfs4_ff_layout_segment *fls)
333 for (i = 0; i < fls->mirror_array_cnt - 1; i++) {
334 for (j = i + 1; j < fls->mirror_array_cnt; j++)
335 if (fls->mirror_array[i]->efficiency <
336 fls->mirror_array[j]->efficiency)
337 swap(fls->mirror_array[i],
338 fls->mirror_array[j]);
342 static void ff_layout_mark_devices_valid(struct nfs4_ff_layout_segment *fls)
344 struct nfs4_deviceid_node *node;
347 if (!(fls->flags & FF_FLAGS_NO_IO_THRU_MDS))
349 for (i = 0; i < fls->mirror_array_cnt; i++) {
350 node = &fls->mirror_array[i]->mirror_ds->id_node;
351 clear_bit(NFS_DEVICEID_UNAVAILABLE, &node->flags);
355 static struct pnfs_layout_segment *
356 ff_layout_alloc_lseg(struct pnfs_layout_hdr *lh,
357 struct nfs4_layoutget_res *lgr,
360 struct pnfs_layout_segment *ret;
361 struct nfs4_ff_layout_segment *fls = NULL;
362 struct xdr_stream stream;
364 struct page *scratch;
366 u32 mirror_array_cnt;
370 dprintk("--> %s\n", __func__);
371 scratch = alloc_page(gfp_flags);
373 return ERR_PTR(-ENOMEM);
375 xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages,
377 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
379 /* stripe unit and mirror_array_cnt */
381 p = xdr_inline_decode(&stream, 8 + 4);
385 p = xdr_decode_hyper(p, &stripe_unit);
386 mirror_array_cnt = be32_to_cpup(p++);
387 dprintk("%s: stripe_unit=%llu mirror_array_cnt=%u\n", __func__,
388 stripe_unit, mirror_array_cnt);
390 if (mirror_array_cnt > NFS4_FLEXFILE_LAYOUT_MAX_MIRROR_CNT ||
391 mirror_array_cnt == 0)
395 fls = kzalloc(sizeof(*fls), gfp_flags);
399 fls->mirror_array_cnt = mirror_array_cnt;
400 fls->stripe_unit = stripe_unit;
401 fls->mirror_array = kcalloc(fls->mirror_array_cnt,
402 sizeof(fls->mirror_array[0]), gfp_flags);
403 if (fls->mirror_array == NULL)
406 for (i = 0; i < fls->mirror_array_cnt; i++) {
407 struct nfs4_ff_layout_mirror *mirror;
408 struct nfs4_deviceid devid;
409 struct nfs4_deviceid_node *idnode;
415 p = xdr_inline_decode(&stream, 4);
418 ds_count = be32_to_cpup(p);
420 /* FIXME: allow for striping? */
424 fls->mirror_array[i] = ff_layout_alloc_mirror(gfp_flags);
425 if (fls->mirror_array[i] == NULL) {
430 fls->mirror_array[i]->ds_count = ds_count;
433 rc = decode_deviceid(&stream, &devid);
437 idnode = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
438 &devid, lh->plh_lc_cred,
441 * upon success, mirror_ds is allocated by previous
442 * getdeviceinfo, or newly by .alloc_deviceid_node
443 * nfs4_find_get_deviceid failure is indeed getdeviceinfo falure
446 fls->mirror_array[i]->mirror_ds =
447 FF_LAYOUT_MIRROR_DS(idnode);
453 p = xdr_inline_decode(&stream, 4);
456 fls->mirror_array[i]->efficiency = be32_to_cpup(p);
459 rc = decode_stateid(&stream, &fls->mirror_array[i]->stateid);
464 p = xdr_inline_decode(&stream, 4);
467 fh_count = be32_to_cpup(p);
469 fls->mirror_array[i]->fh_versions =
470 kzalloc(fh_count * sizeof(struct nfs_fh),
472 if (fls->mirror_array[i]->fh_versions == NULL) {
477 for (j = 0; j < fh_count; j++) {
478 rc = decode_nfs_fh(&stream,
479 &fls->mirror_array[i]->fh_versions[j]);
484 fls->mirror_array[i]->fh_versions_cnt = fh_count;
487 rc = decode_name(&stream, &fls->mirror_array[i]->uid);
492 rc = decode_name(&stream, &fls->mirror_array[i]->gid);
496 mirror = ff_layout_add_mirror(lh, fls->mirror_array[i]);
497 if (mirror != fls->mirror_array[i]) {
498 ff_layout_free_mirror(fls->mirror_array[i]);
499 fls->mirror_array[i] = mirror;
502 dprintk("%s: uid %d gid %d\n", __func__,
503 fls->mirror_array[i]->uid,
504 fls->mirror_array[i]->gid);
507 p = xdr_inline_decode(&stream, 4);
509 fls->flags = be32_to_cpup(p);
511 ff_layout_sort_mirrors(fls);
512 rc = ff_layout_check_layout(lgr);
515 ff_layout_mark_devices_valid(fls);
517 ret = &fls->generic_hdr;
518 dprintk("<-- %s (success)\n", __func__);
520 __free_page(scratch);
523 _ff_layout_free_lseg(fls);
525 dprintk("<-- %s (%d)\n", __func__, rc);
529 static bool ff_layout_has_rw_segments(struct pnfs_layout_hdr *layout)
531 struct pnfs_layout_segment *lseg;
533 list_for_each_entry(lseg, &layout->plh_segs, pls_list)
534 if (lseg->pls_range.iomode == IOMODE_RW)
541 ff_layout_free_lseg(struct pnfs_layout_segment *lseg)
543 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
545 dprintk("--> %s\n", __func__);
547 if (lseg->pls_range.iomode == IOMODE_RW) {
548 struct nfs4_flexfile_layout *ffl;
551 ffl = FF_LAYOUT_FROM_HDR(lseg->pls_layout);
552 inode = ffl->generic_hdr.plh_inode;
553 spin_lock(&inode->i_lock);
554 if (!ff_layout_has_rw_segments(lseg->pls_layout)) {
555 ffl->commit_info.nbuckets = 0;
556 kfree(ffl->commit_info.buckets);
557 ffl->commit_info.buckets = NULL;
559 spin_unlock(&inode->i_lock);
561 _ff_layout_free_lseg(fls);
564 /* Return 1 until we have multiple lsegs support */
566 ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
572 nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
574 /* first IO request? */
575 if (atomic_inc_return(&timer->n_ops) == 1) {
576 timer->start_time = now;
581 nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer, ktime_t now)
585 if (atomic_dec_return(&timer->n_ops) < 0)
588 start = timer->start_time;
589 timer->start_time = now;
590 return ktime_sub(now, start);
594 nfs4_ff_layoutstat_start_io(struct nfs4_ff_layout_mirror *mirror,
595 struct nfs4_ff_layoutstat *layoutstat,
598 static const ktime_t notime = {0};
599 s64 report_interval = FF_LAYOUTSTATS_REPORT_INTERVAL;
601 nfs4_ff_start_busy_timer(&layoutstat->busy_timer, now);
602 if (ktime_equal(mirror->start_time, notime))
603 mirror->start_time = now;
604 if (ktime_equal(mirror->last_report_time, notime))
605 mirror->last_report_time = now;
606 if (layoutstats_timer != 0)
607 report_interval = (s64)layoutstats_timer * 1000LL;
608 if (ktime_to_ms(ktime_sub(now, mirror->last_report_time)) >=
610 mirror->last_report_time = now;
618 nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
621 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
623 iostat->ops_requested++;
624 iostat->bytes_requested += requested;
628 nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
631 ktime_t time_completed,
632 ktime_t time_started)
634 struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
635 ktime_t completion_time = ktime_sub(time_completed, time_started);
638 iostat->ops_completed++;
639 iostat->bytes_completed += completed;
640 iostat->bytes_not_delivered += requested - completed;
642 timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer, time_completed);
643 iostat->total_busy_time =
644 ktime_add(iostat->total_busy_time, timer);
645 iostat->aggregate_completion_time =
646 ktime_add(iostat->aggregate_completion_time,
651 nfs4_ff_layout_stat_io_start_read(struct inode *inode,
652 struct nfs4_ff_layout_mirror *mirror,
653 __u64 requested, ktime_t now)
657 spin_lock(&mirror->lock);
658 report = nfs4_ff_layoutstat_start_io(mirror, &mirror->read_stat, now);
659 nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
660 spin_unlock(&mirror->lock);
663 pnfs_report_layoutstat(inode, GFP_KERNEL);
667 nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
668 struct nfs4_ff_layout_mirror *mirror,
672 spin_lock(&mirror->lock);
673 nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
674 requested, completed,
675 ktime_get(), task->tk_start);
676 spin_unlock(&mirror->lock);
680 nfs4_ff_layout_stat_io_start_write(struct inode *inode,
681 struct nfs4_ff_layout_mirror *mirror,
682 __u64 requested, ktime_t now)
686 spin_lock(&mirror->lock);
687 report = nfs4_ff_layoutstat_start_io(mirror , &mirror->write_stat, now);
688 nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
689 spin_unlock(&mirror->lock);
692 pnfs_report_layoutstat(inode, GFP_NOIO);
696 nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
697 struct nfs4_ff_layout_mirror *mirror,
700 enum nfs3_stable_how committed)
702 if (committed == NFS_UNSTABLE)
703 requested = completed = 0;
705 spin_lock(&mirror->lock);
706 nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
707 requested, completed, ktime_get(), task->tk_start);
708 spin_unlock(&mirror->lock);
712 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
713 struct nfs_commit_info *cinfo,
716 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
717 struct pnfs_commit_bucket *buckets;
720 if (cinfo->ds->nbuckets != 0) {
721 /* This assumes there is only one RW lseg per file.
722 * To support multiple lseg per file, we need to
723 * change struct pnfs_commit_bucket to allow dynamic
724 * increasing nbuckets.
729 size = ff_layout_get_lseg_count(fls) * FF_LAYOUT_MIRROR_COUNT(lseg);
731 buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
738 spin_lock(cinfo->lock);
739 if (cinfo->ds->nbuckets != 0)
742 cinfo->ds->buckets = buckets;
743 cinfo->ds->nbuckets = size;
744 for (i = 0; i < size; i++) {
745 INIT_LIST_HEAD(&buckets[i].written);
746 INIT_LIST_HEAD(&buckets[i].committing);
747 /* mark direct verifier as unset */
748 buckets[i].direct_verf.committed =
749 NFS_INVALID_STABLE_HOW;
752 spin_unlock(cinfo->lock);
757 static struct nfs4_pnfs_ds *
758 ff_layout_choose_best_ds_for_read(struct pnfs_layout_segment *lseg,
762 struct nfs4_ff_layout_segment *fls = FF_LAYOUT_LSEG(lseg);
763 struct nfs4_pnfs_ds *ds;
766 /* mirrors are sorted by efficiency */
767 for (idx = start_idx; idx < fls->mirror_array_cnt; idx++) {
768 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
779 ff_layout_pg_init_read(struct nfs_pageio_descriptor *pgio,
780 struct nfs_page *req)
782 struct nfs_pgio_mirror *pgm;
783 struct nfs4_ff_layout_mirror *mirror;
784 struct nfs4_pnfs_ds *ds;
787 /* Use full layout for now */
789 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
795 /* If no lseg, fall back to read through mds */
796 if (pgio->pg_lseg == NULL)
799 ds = ff_layout_choose_best_ds_for_read(pgio->pg_lseg, 0, &ds_idx);
802 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, ds_idx);
804 pgio->pg_mirror_idx = ds_idx;
806 /* read always uses only one mirror - idx 0 for pgio layer */
807 pgm = &pgio->pg_mirrors[0];
808 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].rsize;
812 pnfs_put_lseg(pgio->pg_lseg);
813 pgio->pg_lseg = NULL;
814 nfs_pageio_reset_read_mds(pgio);
818 ff_layout_pg_init_write(struct nfs_pageio_descriptor *pgio,
819 struct nfs_page *req)
821 struct nfs4_ff_layout_mirror *mirror;
822 struct nfs_pgio_mirror *pgm;
823 struct nfs_commit_info cinfo;
824 struct nfs4_pnfs_ds *ds;
829 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
835 /* If no lseg, fall back to write through mds */
836 if (pgio->pg_lseg == NULL)
839 nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
840 status = ff_layout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
844 /* Use a direct mapping of ds_idx to pgio mirror_idx */
845 if (WARN_ON_ONCE(pgio->pg_mirror_count !=
846 FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg)))
849 for (i = 0; i < pgio->pg_mirror_count; i++) {
850 ds = nfs4_ff_layout_prepare_ds(pgio->pg_lseg, i, true);
853 pgm = &pgio->pg_mirrors[i];
854 mirror = FF_LAYOUT_COMP(pgio->pg_lseg, i);
855 pgm->pg_bsize = mirror->mirror_ds->ds_versions[0].wsize;
861 pnfs_put_lseg(pgio->pg_lseg);
862 pgio->pg_lseg = NULL;
863 nfs_pageio_reset_write_mds(pgio);
867 ff_layout_pg_get_mirror_count_write(struct nfs_pageio_descriptor *pgio,
868 struct nfs_page *req)
871 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
878 return FF_LAYOUT_MIRROR_COUNT(pgio->pg_lseg);
880 /* no lseg means that pnfs is not in use, so no mirroring here */
881 nfs_pageio_reset_write_mds(pgio);
885 static const struct nfs_pageio_ops ff_layout_pg_read_ops = {
886 .pg_init = ff_layout_pg_init_read,
887 .pg_test = pnfs_generic_pg_test,
888 .pg_doio = pnfs_generic_pg_readpages,
889 .pg_cleanup = pnfs_generic_pg_cleanup,
892 static const struct nfs_pageio_ops ff_layout_pg_write_ops = {
893 .pg_init = ff_layout_pg_init_write,
894 .pg_test = pnfs_generic_pg_test,
895 .pg_doio = pnfs_generic_pg_writepages,
896 .pg_get_mirror_count = ff_layout_pg_get_mirror_count_write,
897 .pg_cleanup = pnfs_generic_pg_cleanup,
900 static void ff_layout_reset_write(struct nfs_pgio_header *hdr, bool retry_pnfs)
902 struct rpc_task *task = &hdr->task;
904 pnfs_layoutcommit_inode(hdr->inode, false);
907 dprintk("%s Reset task %5u for i/o through pNFS "
908 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
910 hdr->inode->i_sb->s_id,
911 (unsigned long long)NFS_FILEID(hdr->inode),
913 (unsigned long long)hdr->args.offset);
916 struct nfs_open_context *ctx;
918 ctx = nfs_list_entry(hdr->pages.next)->wb_context;
919 set_bit(NFS_CONTEXT_RESEND_WRITES, &ctx->flags);
920 hdr->completion_ops->error_cleanup(&hdr->pages);
922 nfs_direct_set_resched_writes(hdr->dreq);
923 /* fake unstable write to let common nfs resend pages */
924 hdr->verf.committed = NFS_UNSTABLE;
925 hdr->good_bytes = hdr->args.count;
930 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
931 dprintk("%s Reset task %5u for i/o through MDS "
932 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
934 hdr->inode->i_sb->s_id,
935 (unsigned long long)NFS_FILEID(hdr->inode),
937 (unsigned long long)hdr->args.offset);
939 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
943 static void ff_layout_reset_read(struct nfs_pgio_header *hdr)
945 struct rpc_task *task = &hdr->task;
947 pnfs_layoutcommit_inode(hdr->inode, false);
949 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
950 dprintk("%s Reset task %5u for i/o through MDS "
951 "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
953 hdr->inode->i_sb->s_id,
954 (unsigned long long)NFS_FILEID(hdr->inode),
956 (unsigned long long)hdr->args.offset);
958 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
962 static int ff_layout_async_handle_error_v4(struct rpc_task *task,
963 struct nfs4_state *state,
964 struct nfs_client *clp,
965 struct pnfs_layout_segment *lseg,
968 struct pnfs_layout_hdr *lo = lseg->pls_layout;
969 struct inode *inode = lo->plh_inode;
970 struct nfs_server *mds_server = NFS_SERVER(inode);
972 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
973 struct nfs_client *mds_client = mds_server->nfs_client;
974 struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
976 if (task->tk_status >= 0)
979 switch (task->tk_status) {
980 /* MDS state errors */
981 case -NFS4ERR_DELEG_REVOKED:
982 case -NFS4ERR_ADMIN_REVOKED:
983 case -NFS4ERR_BAD_STATEID:
986 nfs_remove_bad_delegation(state->inode);
987 case -NFS4ERR_OPENMODE:
990 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
991 goto out_bad_stateid;
992 goto wait_on_recovery;
993 case -NFS4ERR_EXPIRED:
995 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
996 goto out_bad_stateid;
998 nfs4_schedule_lease_recovery(mds_client);
999 goto wait_on_recovery;
1000 /* DS session errors */
1001 case -NFS4ERR_BADSESSION:
1002 case -NFS4ERR_BADSLOT:
1003 case -NFS4ERR_BAD_HIGH_SLOT:
1004 case -NFS4ERR_DEADSESSION:
1005 case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
1006 case -NFS4ERR_SEQ_FALSE_RETRY:
1007 case -NFS4ERR_SEQ_MISORDERED:
1008 dprintk("%s ERROR %d, Reset session. Exchangeid "
1009 "flags 0x%x\n", __func__, task->tk_status,
1010 clp->cl_exchange_flags);
1011 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
1013 case -NFS4ERR_DELAY:
1014 case -NFS4ERR_GRACE:
1015 rpc_delay(task, FF_LAYOUT_POLL_RETRY_MAX);
1017 case -NFS4ERR_RETRY_UNCACHED_REP:
1019 /* Invalidate Layout errors */
1020 case -NFS4ERR_PNFS_NO_LAYOUT:
1021 case -ESTALE: /* mapped NFS4ERR_STALE */
1022 case -EBADHANDLE: /* mapped NFS4ERR_BADHANDLE */
1023 case -EISDIR: /* mapped NFS4ERR_ISDIR */
1024 case -NFS4ERR_FHEXPIRED:
1025 case -NFS4ERR_WRONG_TYPE:
1026 dprintk("%s Invalid layout error %d\n", __func__,
1029 * Destroy layout so new i/o will get a new layout.
1030 * Layout will not be destroyed until all current lseg
1031 * references are put. Mark layout as invalid to resend failed
1032 * i/o and all i/o waiting on the slot table to the MDS until
1033 * layout is destroyed and a new valid layout is obtained.
1035 pnfs_destroy_layout(NFS_I(inode));
1036 rpc_wake_up(&tbl->slot_tbl_waitq);
1038 /* RPC connection errors */
1046 dprintk("%s DS connection error %d\n", __func__,
1048 nfs4_mark_deviceid_unavailable(devid);
1049 rpc_wake_up(&tbl->slot_tbl_waitq);
1052 if (ff_layout_no_fallback_to_mds(lseg) ||
1053 ff_layout_has_available_ds(lseg))
1054 return -NFS4ERR_RESET_TO_PNFS;
1056 dprintk("%s Retry through MDS. Error %d\n", __func__,
1058 return -NFS4ERR_RESET_TO_MDS;
1061 task->tk_status = 0;
1064 task->tk_status = -EIO;
1067 rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
1068 if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
1069 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
1073 /* Retry all errors through either pNFS or MDS except for -EJUKEBOX */
1074 static int ff_layout_async_handle_error_v3(struct rpc_task *task,
1075 struct pnfs_layout_segment *lseg,
1078 struct nfs4_deviceid_node *devid = FF_LAYOUT_DEVID_NODE(lseg, idx);
1080 if (task->tk_status >= 0)
1083 switch (task->tk_status) {
1084 /* File access problems. Don't mark the device as unavailable */
1093 nfs_inc_stats(lseg->pls_layout->plh_inode, NFSIOS_DELAY);
1096 dprintk("%s DS connection error %d\n", __func__,
1098 nfs4_mark_deviceid_unavailable(devid);
1100 /* FIXME: Need to prevent infinite looping here. */
1101 return -NFS4ERR_RESET_TO_PNFS;
1103 task->tk_status = 0;
1104 rpc_restart_call(task);
1105 rpc_delay(task, NFS_JUKEBOX_RETRY_TIME);
1109 static int ff_layout_async_handle_error(struct rpc_task *task,
1110 struct nfs4_state *state,
1111 struct nfs_client *clp,
1112 struct pnfs_layout_segment *lseg,
1115 int vers = clp->cl_nfs_mod->rpc_vers->number;
1119 return ff_layout_async_handle_error_v3(task, lseg, idx);
1121 return ff_layout_async_handle_error_v4(task, state, clp,
1124 /* should never happen */
1130 static void ff_layout_io_track_ds_error(struct pnfs_layout_segment *lseg,
1131 int idx, u64 offset, u64 length,
1132 u32 status, int opnum, int error)
1134 struct nfs4_ff_layout_mirror *mirror;
1141 case -EPROTONOSUPPORT:
1152 status = NFS4ERR_NXIO;
1155 status = NFS4ERR_ACCESS;
1162 mirror = FF_LAYOUT_COMP(lseg, idx);
1163 err = ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
1164 mirror, offset, length, status, opnum,
1166 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
1167 dprintk("%s: err %d op %d status %u\n", __func__, err, opnum, status);
1170 /* NFS_PROTO call done callback routines */
1171 static int ff_layout_read_done_cb(struct rpc_task *task,
1172 struct nfs_pgio_header *hdr)
1176 trace_nfs4_pnfs_read(hdr, task->tk_status);
1177 if (task->tk_status < 0)
1178 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1179 hdr->args.offset, hdr->args.count,
1180 hdr->res.op_status, OP_READ,
1182 err = ff_layout_async_handle_error(task, hdr->args.context->state,
1183 hdr->ds_clp, hdr->lseg,
1184 hdr->pgio_mirror_idx);
1187 case -NFS4ERR_RESET_TO_PNFS:
1188 if (ff_layout_choose_best_ds_for_read(hdr->lseg,
1189 hdr->pgio_mirror_idx + 1,
1190 &hdr->pgio_mirror_idx))
1192 set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1193 &hdr->lseg->pls_layout->plh_flags);
1194 pnfs_read_resend_pnfs(hdr);
1195 return task->tk_status;
1196 case -NFS4ERR_RESET_TO_MDS:
1197 ff_layout_reset_read(hdr);
1198 return task->tk_status;
1205 rpc_restart_call_prepare(task);
1210 ff_layout_need_layoutcommit(struct pnfs_layout_segment *lseg)
1212 return !(FF_LAYOUT_LSEG(lseg)->flags & FF_FLAGS_NO_LAYOUTCOMMIT);
1216 * We reference the rpc_cred of the first WRITE that triggers the need for
1217 * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
1218 * rfc5661 is not clear about which credential should be used.
1220 * Flexlayout client should treat DS replied FILE_SYNC as DATA_SYNC, so
1221 * to follow http://www.rfc-editor.org/errata_search.php?rfc=5661&eid=2751
1222 * we always send layoutcommit after DS writes.
1225 ff_layout_set_layoutcommit(struct nfs_pgio_header *hdr)
1227 if (!ff_layout_need_layoutcommit(hdr->lseg))
1230 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1231 hdr->mds_offset + hdr->res.count);
1232 dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
1233 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
1237 ff_layout_reset_to_mds(struct pnfs_layout_segment *lseg, int idx)
1239 /* No mirroring for now */
1240 struct nfs4_deviceid_node *node = FF_LAYOUT_DEVID_NODE(lseg, idx);
1242 return ff_layout_test_devid_unavailable(node);
1245 static int ff_layout_read_prepare_common(struct rpc_task *task,
1246 struct nfs_pgio_header *hdr)
1248 nfs4_ff_layout_stat_io_start_read(hdr->inode,
1249 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1253 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1254 rpc_exit(task, -EIO);
1257 if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1258 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
1259 if (ff_layout_has_available_ds(hdr->lseg))
1260 pnfs_read_resend_pnfs(hdr);
1262 ff_layout_reset_read(hdr);
1266 hdr->pgio_done_cb = ff_layout_read_done_cb;
1272 * Call ops for the async read/write cases
1273 * In the case of dense layouts, the offset needs to be reset to its
1276 static void ff_layout_read_prepare_v3(struct rpc_task *task, void *data)
1278 struct nfs_pgio_header *hdr = data;
1280 if (ff_layout_read_prepare_common(task, hdr))
1283 rpc_call_start(task);
1286 static int ff_layout_setup_sequence(struct nfs_client *ds_clp,
1287 struct nfs4_sequence_args *args,
1288 struct nfs4_sequence_res *res,
1289 struct rpc_task *task)
1291 if (ds_clp->cl_session)
1292 return nfs41_setup_sequence(ds_clp->cl_session,
1296 return nfs40_setup_sequence(ds_clp->cl_slot_tbl,
1302 static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
1304 struct nfs_pgio_header *hdr = data;
1306 if (ff_layout_setup_sequence(hdr->ds_clp,
1307 &hdr->args.seq_args,
1312 if (ff_layout_read_prepare_common(task, hdr))
1315 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1316 hdr->args.lock_context, FMODE_READ) == -EIO)
1317 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1320 static void ff_layout_read_call_done(struct rpc_task *task, void *data)
1322 struct nfs_pgio_header *hdr = data;
1324 dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
1326 nfs4_ff_layout_stat_io_end_read(task,
1327 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1328 hdr->args.count, hdr->res.count);
1330 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1331 task->tk_status == 0) {
1332 nfs4_sequence_done(task, &hdr->res.seq_res);
1336 /* Note this may cause RPC to be resent */
1337 hdr->mds_ops->rpc_call_done(task, hdr);
1340 static void ff_layout_read_count_stats(struct rpc_task *task, void *data)
1342 struct nfs_pgio_header *hdr = data;
1344 rpc_count_iostats_metrics(task,
1345 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_READ]);
1348 static int ff_layout_write_done_cb(struct rpc_task *task,
1349 struct nfs_pgio_header *hdr)
1353 trace_nfs4_pnfs_write(hdr, task->tk_status);
1354 if (task->tk_status < 0)
1355 ff_layout_io_track_ds_error(hdr->lseg, hdr->pgio_mirror_idx,
1356 hdr->args.offset, hdr->args.count,
1357 hdr->res.op_status, OP_WRITE,
1359 err = ff_layout_async_handle_error(task, hdr->args.context->state,
1360 hdr->ds_clp, hdr->lseg,
1361 hdr->pgio_mirror_idx);
1364 case -NFS4ERR_RESET_TO_PNFS:
1365 pnfs_set_retry_layoutget(hdr->lseg->pls_layout);
1366 ff_layout_reset_write(hdr, true);
1367 return task->tk_status;
1368 case -NFS4ERR_RESET_TO_MDS:
1369 pnfs_clear_retry_layoutget(hdr->lseg->pls_layout);
1370 ff_layout_reset_write(hdr, false);
1371 return task->tk_status;
1373 rpc_restart_call_prepare(task);
1377 if (hdr->res.verf->committed == NFS_FILE_SYNC ||
1378 hdr->res.verf->committed == NFS_DATA_SYNC)
1379 ff_layout_set_layoutcommit(hdr);
1381 /* zero out fattr since we don't care DS attr at all */
1382 hdr->fattr.valid = 0;
1383 if (task->tk_status >= 0)
1384 nfs_writeback_update_inode(hdr);
1389 static int ff_layout_commit_done_cb(struct rpc_task *task,
1390 struct nfs_commit_data *data)
1394 trace_nfs4_pnfs_commit_ds(data, task->tk_status);
1395 if (task->tk_status < 0)
1396 ff_layout_io_track_ds_error(data->lseg, data->ds_commit_index,
1397 data->args.offset, data->args.count,
1398 data->res.op_status, OP_COMMIT,
1400 err = ff_layout_async_handle_error(task, NULL, data->ds_clp,
1401 data->lseg, data->ds_commit_index);
1404 case -NFS4ERR_RESET_TO_PNFS:
1405 pnfs_set_retry_layoutget(data->lseg->pls_layout);
1406 pnfs_generic_prepare_to_resend_writes(data);
1408 case -NFS4ERR_RESET_TO_MDS:
1409 pnfs_clear_retry_layoutget(data->lseg->pls_layout);
1410 pnfs_generic_prepare_to_resend_writes(data);
1413 rpc_restart_call_prepare(task);
1417 if (data->verf.committed == NFS_UNSTABLE
1418 && ff_layout_need_layoutcommit(data->lseg))
1419 pnfs_set_layoutcommit(data->inode, data->lseg, data->lwb);
1424 static int ff_layout_write_prepare_common(struct rpc_task *task,
1425 struct nfs_pgio_header *hdr)
1427 nfs4_ff_layout_stat_io_start_write(hdr->inode,
1428 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1432 if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
1433 rpc_exit(task, -EIO);
1437 if (ff_layout_reset_to_mds(hdr->lseg, hdr->pgio_mirror_idx)) {
1440 retry_pnfs = ff_layout_has_available_ds(hdr->lseg);
1441 dprintk("%s task %u reset io to %s\n", __func__,
1442 task->tk_pid, retry_pnfs ? "pNFS" : "MDS");
1443 ff_layout_reset_write(hdr, retry_pnfs);
1451 static void ff_layout_write_prepare_v3(struct rpc_task *task, void *data)
1453 struct nfs_pgio_header *hdr = data;
1455 if (ff_layout_write_prepare_common(task, hdr))
1458 rpc_call_start(task);
1461 static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
1463 struct nfs_pgio_header *hdr = data;
1465 if (ff_layout_setup_sequence(hdr->ds_clp,
1466 &hdr->args.seq_args,
1471 if (ff_layout_write_prepare_common(task, hdr))
1474 if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
1475 hdr->args.lock_context, FMODE_WRITE) == -EIO)
1476 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
1479 static void ff_layout_write_call_done(struct rpc_task *task, void *data)
1481 struct nfs_pgio_header *hdr = data;
1483 nfs4_ff_layout_stat_io_end_write(task,
1484 FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
1485 hdr->args.count, hdr->res.count,
1486 hdr->res.verf->committed);
1488 if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
1489 task->tk_status == 0) {
1490 nfs4_sequence_done(task, &hdr->res.seq_res);
1494 /* Note this may cause RPC to be resent */
1495 hdr->mds_ops->rpc_call_done(task, hdr);
1498 static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
1500 struct nfs_pgio_header *hdr = data;
1502 rpc_count_iostats_metrics(task,
1503 &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
1506 static void ff_layout_commit_prepare_common(struct rpc_task *task,
1507 struct nfs_commit_data *cdata)
1509 nfs4_ff_layout_stat_io_start_write(cdata->inode,
1510 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1514 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
1516 ff_layout_commit_prepare_common(task, data);
1517 rpc_call_start(task);
1520 static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
1522 struct nfs_commit_data *wdata = data;
1524 if (ff_layout_setup_sequence(wdata->ds_clp,
1525 &wdata->args.seq_args,
1526 &wdata->res.seq_res,
1529 ff_layout_commit_prepare_common(task, data);
1532 static void ff_layout_commit_done(struct rpc_task *task, void *data)
1534 struct nfs_commit_data *cdata = data;
1535 struct nfs_page *req;
1538 if (task->tk_status == 0) {
1539 list_for_each_entry(req, &cdata->pages, wb_list)
1540 count += req->wb_bytes;
1543 nfs4_ff_layout_stat_io_end_write(task,
1544 FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
1545 count, count, NFS_FILE_SYNC);
1547 pnfs_generic_write_commit_done(task, data);
1550 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
1552 struct nfs_commit_data *cdata = data;
1554 rpc_count_iostats_metrics(task,
1555 &NFS_CLIENT(cdata->inode)->cl_metrics[NFSPROC4_CLNT_COMMIT]);
1558 static const struct rpc_call_ops ff_layout_read_call_ops_v3 = {
1559 .rpc_call_prepare = ff_layout_read_prepare_v3,
1560 .rpc_call_done = ff_layout_read_call_done,
1561 .rpc_count_stats = ff_layout_read_count_stats,
1562 .rpc_release = pnfs_generic_rw_release,
1565 static const struct rpc_call_ops ff_layout_read_call_ops_v4 = {
1566 .rpc_call_prepare = ff_layout_read_prepare_v4,
1567 .rpc_call_done = ff_layout_read_call_done,
1568 .rpc_count_stats = ff_layout_read_count_stats,
1569 .rpc_release = pnfs_generic_rw_release,
1572 static const struct rpc_call_ops ff_layout_write_call_ops_v3 = {
1573 .rpc_call_prepare = ff_layout_write_prepare_v3,
1574 .rpc_call_done = ff_layout_write_call_done,
1575 .rpc_count_stats = ff_layout_write_count_stats,
1576 .rpc_release = pnfs_generic_rw_release,
1579 static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
1580 .rpc_call_prepare = ff_layout_write_prepare_v4,
1581 .rpc_call_done = ff_layout_write_call_done,
1582 .rpc_count_stats = ff_layout_write_count_stats,
1583 .rpc_release = pnfs_generic_rw_release,
1586 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
1587 .rpc_call_prepare = ff_layout_commit_prepare_v3,
1588 .rpc_call_done = ff_layout_commit_done,
1589 .rpc_count_stats = ff_layout_commit_count_stats,
1590 .rpc_release = pnfs_generic_commit_release,
1593 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
1594 .rpc_call_prepare = ff_layout_commit_prepare_v4,
1595 .rpc_call_done = ff_layout_commit_done,
1596 .rpc_count_stats = ff_layout_commit_count_stats,
1597 .rpc_release = pnfs_generic_commit_release,
1600 static enum pnfs_try_status
1601 ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
1603 struct pnfs_layout_segment *lseg = hdr->lseg;
1604 struct nfs4_pnfs_ds *ds;
1605 struct rpc_clnt *ds_clnt;
1606 struct rpc_cred *ds_cred;
1607 loff_t offset = hdr->args.offset;
1608 u32 idx = hdr->pgio_mirror_idx;
1612 dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
1613 __func__, hdr->inode->i_ino,
1614 hdr->args.pgbase, (size_t)hdr->args.count, offset);
1616 ds = nfs4_ff_layout_prepare_ds(lseg, idx, false);
1620 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1622 if (IS_ERR(ds_clnt))
1625 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1626 if (IS_ERR(ds_cred))
1629 vers = nfs4_ff_layout_ds_version(lseg, idx);
1631 dprintk("%s USE DS: %s cl_count %d vers %d\n", __func__,
1632 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count), vers);
1634 atomic_inc(&ds->ds_clp->cl_count);
1635 hdr->ds_clp = ds->ds_clp;
1636 fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1640 * Note that if we ever decide to split across DSes,
1641 * then we may need to handle dense-like offsets.
1643 hdr->args.offset = offset;
1644 hdr->mds_offset = offset;
1646 /* Perform an asynchronous read to ds */
1647 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1648 vers == 3 ? &ff_layout_read_call_ops_v3 :
1649 &ff_layout_read_call_ops_v4,
1650 0, RPC_TASK_SOFTCONN);
1652 return PNFS_ATTEMPTED;
1655 if (ff_layout_has_available_ds(lseg))
1656 return PNFS_TRY_AGAIN;
1657 return PNFS_NOT_ATTEMPTED;
1660 /* Perform async writes. */
1661 static enum pnfs_try_status
1662 ff_layout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
1664 struct pnfs_layout_segment *lseg = hdr->lseg;
1665 struct nfs4_pnfs_ds *ds;
1666 struct rpc_clnt *ds_clnt;
1667 struct rpc_cred *ds_cred;
1668 loff_t offset = hdr->args.offset;
1671 int idx = hdr->pgio_mirror_idx;
1673 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1675 return PNFS_NOT_ATTEMPTED;
1677 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1679 if (IS_ERR(ds_clnt))
1680 return PNFS_NOT_ATTEMPTED;
1682 ds_cred = ff_layout_get_ds_cred(lseg, idx, hdr->cred);
1683 if (IS_ERR(ds_cred))
1684 return PNFS_NOT_ATTEMPTED;
1686 vers = nfs4_ff_layout_ds_version(lseg, idx);
1688 dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d vers %d\n",
1689 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
1690 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count),
1693 hdr->pgio_done_cb = ff_layout_write_done_cb;
1694 atomic_inc(&ds->ds_clp->cl_count);
1695 hdr->ds_clp = ds->ds_clp;
1696 hdr->ds_commit_idx = idx;
1697 fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
1702 * Note that if we ever decide to split across DSes,
1703 * then we may need to handle dense-like offsets.
1705 hdr->args.offset = offset;
1707 /* Perform an asynchronous write */
1708 nfs_initiate_pgio(ds_clnt, hdr, ds_cred, ds->ds_clp->rpc_ops,
1709 vers == 3 ? &ff_layout_write_call_ops_v3 :
1710 &ff_layout_write_call_ops_v4,
1711 sync, RPC_TASK_SOFTCONN);
1712 return PNFS_ATTEMPTED;
1715 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1720 static struct nfs_fh *
1721 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1723 struct nfs4_ff_layout_segment *flseg = FF_LAYOUT_LSEG(lseg);
1725 /* FIXME: Assume that there is only one NFS version available
1728 return &flseg->mirror_array[i]->fh_versions[0];
1731 static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
1733 struct pnfs_layout_segment *lseg = data->lseg;
1734 struct nfs4_pnfs_ds *ds;
1735 struct rpc_clnt *ds_clnt;
1736 struct rpc_cred *ds_cred;
1741 idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1742 ds = nfs4_ff_layout_prepare_ds(lseg, idx, true);
1746 ds_clnt = nfs4_ff_find_or_create_ds_client(lseg, idx, ds->ds_clp,
1748 if (IS_ERR(ds_clnt))
1751 ds_cred = ff_layout_get_ds_cred(lseg, idx, data->cred);
1752 if (IS_ERR(ds_cred))
1755 vers = nfs4_ff_layout_ds_version(lseg, idx);
1757 dprintk("%s ino %lu, how %d cl_count %d vers %d\n", __func__,
1758 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count),
1760 data->commit_done_cb = ff_layout_commit_done_cb;
1761 data->cred = ds_cred;
1762 atomic_inc(&ds->ds_clp->cl_count);
1763 data->ds_clp = ds->ds_clp;
1764 fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1768 return nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
1769 vers == 3 ? &ff_layout_commit_call_ops_v3 :
1770 &ff_layout_commit_call_ops_v4,
1771 how, RPC_TASK_SOFTCONN);
1773 pnfs_generic_prepare_to_resend_writes(data);
1774 pnfs_generic_commit_release(data);
1779 ff_layout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1780 int how, struct nfs_commit_info *cinfo)
1782 return pnfs_generic_commit_pagelist(inode, mds_pages, how, cinfo,
1783 ff_layout_initiate_commit);
1786 static struct pnfs_ds_commit_info *
1787 ff_layout_get_ds_info(struct inode *inode)
1789 struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1794 return &FF_LAYOUT_FROM_HDR(layout)->commit_info;
1798 ff_layout_free_deviceid_node(struct nfs4_deviceid_node *d)
1800 nfs4_ff_layout_free_deviceid(container_of(d, struct nfs4_ff_layout_ds,
1804 static int ff_layout_encode_ioerr(struct nfs4_flexfile_layout *flo,
1805 struct xdr_stream *xdr,
1806 const struct nfs4_layoutreturn_args *args)
1808 struct pnfs_layout_hdr *hdr = &flo->generic_hdr;
1810 int count = 0, ret = 0;
1812 start = xdr_reserve_space(xdr, 4);
1813 if (unlikely(!start))
1816 /* This assume we always return _ALL_ layouts */
1817 spin_lock(&hdr->plh_inode->i_lock);
1818 ret = ff_layout_encode_ds_ioerr(flo, xdr, &count, &args->range);
1819 spin_unlock(&hdr->plh_inode->i_lock);
1821 *start = cpu_to_be32(count);
1826 /* report nothing for now */
1827 static void ff_layout_encode_iostats(struct nfs4_flexfile_layout *flo,
1828 struct xdr_stream *xdr,
1829 const struct nfs4_layoutreturn_args *args)
1833 p = xdr_reserve_space(xdr, 4);
1835 *p = cpu_to_be32(0);
1838 static struct nfs4_deviceid_node *
1839 ff_layout_alloc_deviceid_node(struct nfs_server *server,
1840 struct pnfs_device *pdev, gfp_t gfp_flags)
1842 struct nfs4_ff_layout_ds *dsaddr;
1844 dsaddr = nfs4_ff_alloc_deviceid_node(server, pdev, gfp_flags);
1847 return &dsaddr->id_node;
1851 ff_layout_encode_layoutreturn(struct pnfs_layout_hdr *lo,
1852 struct xdr_stream *xdr,
1853 const struct nfs4_layoutreturn_args *args)
1855 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
1858 dprintk("%s: Begin\n", __func__);
1859 start = xdr_reserve_space(xdr, 4);
1862 if (ff_layout_encode_ioerr(flo, xdr, args))
1865 ff_layout_encode_iostats(flo, xdr, args);
1867 *start = cpu_to_be32((xdr->p - start - 1) * 4);
1868 dprintk("%s: Return\n", __func__);
1872 ff_layout_ntop4(const struct sockaddr *sap, char *buf, const size_t buflen)
1874 const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
1876 return snprintf(buf, buflen, "%pI4", &sin->sin_addr);
1880 ff_layout_ntop6_noscopeid(const struct sockaddr *sap, char *buf,
1883 const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
1884 const struct in6_addr *addr = &sin6->sin6_addr;
1887 * RFC 4291, Section 2.2.2
1889 * Shorthanded ANY address
1891 if (ipv6_addr_any(addr))
1892 return snprintf(buf, buflen, "::");
1895 * RFC 4291, Section 2.2.2
1897 * Shorthanded loopback address
1899 if (ipv6_addr_loopback(addr))
1900 return snprintf(buf, buflen, "::1");
1903 * RFC 4291, Section 2.2.3
1905 * Special presentation address format for mapped v4
1908 if (ipv6_addr_v4mapped(addr))
1909 return snprintf(buf, buflen, "::ffff:%pI4",
1910 &addr->s6_addr32[3]);
1913 * RFC 4291, Section 2.2.1
1915 return snprintf(buf, buflen, "%pI6c", addr);
1918 /* Derived from rpc_sockaddr2uaddr */
1920 ff_layout_encode_netaddr(struct xdr_stream *xdr, struct nfs4_pnfs_ds_addr *da)
1922 struct sockaddr *sap = (struct sockaddr *)&da->da_addr;
1923 char portbuf[RPCBIND_MAXUADDRPLEN];
1924 char addrbuf[RPCBIND_MAXUADDRLEN];
1926 unsigned short port;
1930 switch (sap->sa_family) {
1932 if (ff_layout_ntop4(sap, addrbuf, sizeof(addrbuf)) == 0)
1934 port = ntohs(((struct sockaddr_in *)sap)->sin_port);
1939 if (ff_layout_ntop6_noscopeid(sap, addrbuf, sizeof(addrbuf)) == 0)
1941 port = ntohs(((struct sockaddr_in6 *)sap)->sin6_port);
1946 /* we only support tcp and tcp6 */
1951 snprintf(portbuf, sizeof(portbuf), ".%u.%u", port >> 8, port & 0xff);
1952 len = strlcat(addrbuf, portbuf, sizeof(addrbuf));
1954 p = xdr_reserve_space(xdr, 4 + netid_len);
1955 xdr_encode_opaque(p, netid, netid_len);
1957 p = xdr_reserve_space(xdr, 4 + len);
1958 xdr_encode_opaque(p, addrbuf, len);
1962 ff_layout_encode_nfstime(struct xdr_stream *xdr,
1965 struct timespec64 ts;
1968 p = xdr_reserve_space(xdr, 12);
1969 ts = ktime_to_timespec64(t);
1970 p = xdr_encode_hyper(p, ts.tv_sec);
1971 *p++ = cpu_to_be32(ts.tv_nsec);
1975 ff_layout_encode_io_latency(struct xdr_stream *xdr,
1976 struct nfs4_ff_io_stat *stat)
1980 p = xdr_reserve_space(xdr, 5 * 8);
1981 p = xdr_encode_hyper(p, stat->ops_requested);
1982 p = xdr_encode_hyper(p, stat->bytes_requested);
1983 p = xdr_encode_hyper(p, stat->ops_completed);
1984 p = xdr_encode_hyper(p, stat->bytes_completed);
1985 p = xdr_encode_hyper(p, stat->bytes_not_delivered);
1986 ff_layout_encode_nfstime(xdr, stat->total_busy_time);
1987 ff_layout_encode_nfstime(xdr, stat->aggregate_completion_time);
1991 ff_layout_encode_layoutstats(struct xdr_stream *xdr,
1992 struct nfs42_layoutstat_args *args,
1993 struct nfs42_layoutstat_devinfo *devinfo)
1995 struct nfs4_ff_layout_mirror *mirror = devinfo->layout_private;
1996 struct nfs4_pnfs_ds_addr *da;
1997 struct nfs4_pnfs_ds *ds = mirror->mirror_ds->ds;
1998 struct nfs_fh *fh = &mirror->fh_versions[0];
2001 da = list_first_entry(&ds->ds_addrs, struct nfs4_pnfs_ds_addr, da_node);
2002 dprintk("%s: DS %s: encoding address %s\n",
2003 __func__, ds->ds_remotestr, da->da_remotestr);
2004 /* layoutupdate length */
2005 start = xdr_reserve_space(xdr, 4);
2007 ff_layout_encode_netaddr(xdr, da);
2009 p = xdr_reserve_space(xdr, 4 + fh->size);
2010 xdr_encode_opaque(p, fh->data, fh->size);
2011 /* ff_io_latency4 read */
2012 spin_lock(&mirror->lock);
2013 ff_layout_encode_io_latency(xdr, &mirror->read_stat.io_stat);
2014 /* ff_io_latency4 write */
2015 ff_layout_encode_io_latency(xdr, &mirror->write_stat.io_stat);
2016 spin_unlock(&mirror->lock);
2018 ff_layout_encode_nfstime(xdr, ktime_sub(ktime_get(), mirror->start_time));
2020 p = xdr_reserve_space(xdr, 4);
2021 *p = cpu_to_be32(false);
2023 *start = cpu_to_be32((xdr->p - start - 1) * 4);
2027 ff_layout_mirror_prepare_stats(struct nfs42_layoutstat_args *args,
2028 struct pnfs_layout_hdr *lo,
2031 struct nfs4_flexfile_layout *ff_layout = FF_LAYOUT_FROM_HDR(lo);
2032 struct nfs4_ff_layout_mirror *mirror;
2033 struct nfs4_deviceid_node *dev;
2034 struct nfs42_layoutstat_devinfo *devinfo;
2037 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2040 if (!mirror->mirror_ds)
2042 /* mirror refcount put in cleanup_layoutstats */
2043 if (!atomic_inc_not_zero(&mirror->ref))
2045 dev = &mirror->mirror_ds->id_node;
2046 devinfo = &args->devinfo[i];
2047 memcpy(&devinfo->dev_id, &dev->deviceid, NFS4_DEVICEID4_SIZE);
2048 devinfo->offset = 0;
2049 devinfo->length = NFS4_MAX_UINT64;
2050 devinfo->read_count = mirror->read_stat.io_stat.ops_completed;
2051 devinfo->read_bytes = mirror->read_stat.io_stat.bytes_completed;
2052 devinfo->write_count = mirror->write_stat.io_stat.ops_completed;
2053 devinfo->write_bytes = mirror->write_stat.io_stat.bytes_completed;
2054 devinfo->layout_type = LAYOUT_FLEX_FILES;
2055 devinfo->layoutstats_encode = ff_layout_encode_layoutstats;
2056 devinfo->layout_private = mirror;
2064 ff_layout_prepare_layoutstats(struct nfs42_layoutstat_args *args)
2066 struct nfs4_flexfile_layout *ff_layout;
2067 struct nfs4_ff_layout_mirror *mirror;
2070 spin_lock(&args->inode->i_lock);
2071 ff_layout = FF_LAYOUT_FROM_HDR(NFS_I(args->inode)->layout);
2072 list_for_each_entry(mirror, &ff_layout->mirrors, mirrors) {
2073 if (atomic_read(&mirror->ref) != 0)
2076 spin_unlock(&args->inode->i_lock);
2077 /* For now, send at most PNFS_LAYOUTSTATS_MAXDEV statistics */
2078 if (dev_count > PNFS_LAYOUTSTATS_MAXDEV) {
2079 dprintk("%s: truncating devinfo to limit (%d:%d)\n",
2080 __func__, dev_count, PNFS_LAYOUTSTATS_MAXDEV);
2081 dev_count = PNFS_LAYOUTSTATS_MAXDEV;
2083 args->devinfo = kmalloc_array(dev_count, sizeof(*args->devinfo), GFP_NOIO);
2087 spin_lock(&args->inode->i_lock);
2088 args->num_dev = ff_layout_mirror_prepare_stats(args,
2089 &ff_layout->generic_hdr, dev_count);
2090 spin_unlock(&args->inode->i_lock);
2096 ff_layout_cleanup_layoutstats(struct nfs42_layoutstat_data *data)
2098 struct nfs4_ff_layout_mirror *mirror;
2101 for (i = 0; i < data->args.num_dev; i++) {
2102 mirror = data->args.devinfo[i].layout_private;
2103 data->args.devinfo[i].layout_private = NULL;
2104 ff_layout_put_mirror(mirror);
2108 static struct pnfs_layoutdriver_type flexfilelayout_type = {
2109 .id = LAYOUT_FLEX_FILES,
2110 .name = "LAYOUT_FLEX_FILES",
2111 .owner = THIS_MODULE,
2112 .alloc_layout_hdr = ff_layout_alloc_layout_hdr,
2113 .free_layout_hdr = ff_layout_free_layout_hdr,
2114 .alloc_lseg = ff_layout_alloc_lseg,
2115 .free_lseg = ff_layout_free_lseg,
2116 .add_lseg = ff_layout_add_lseg,
2117 .pg_read_ops = &ff_layout_pg_read_ops,
2118 .pg_write_ops = &ff_layout_pg_write_ops,
2119 .get_ds_info = ff_layout_get_ds_info,
2120 .free_deviceid_node = ff_layout_free_deviceid_node,
2121 .mark_request_commit = pnfs_layout_mark_request_commit,
2122 .clear_request_commit = pnfs_generic_clear_request_commit,
2123 .scan_commit_lists = pnfs_generic_scan_commit_lists,
2124 .recover_commit_reqs = pnfs_generic_recover_commit_reqs,
2125 .commit_pagelist = ff_layout_commit_pagelist,
2126 .read_pagelist = ff_layout_read_pagelist,
2127 .write_pagelist = ff_layout_write_pagelist,
2128 .alloc_deviceid_node = ff_layout_alloc_deviceid_node,
2129 .encode_layoutreturn = ff_layout_encode_layoutreturn,
2130 .sync = pnfs_nfs_generic_sync,
2131 .prepare_layoutstats = ff_layout_prepare_layoutstats,
2132 .cleanup_layoutstats = ff_layout_cleanup_layoutstats,
2135 static int __init nfs4flexfilelayout_init(void)
2137 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Registering...\n",
2139 return pnfs_register_layoutdriver(&flexfilelayout_type);
2142 static void __exit nfs4flexfilelayout_exit(void)
2144 printk(KERN_INFO "%s: NFSv4 Flexfile Layout Driver Unregistering...\n",
2146 pnfs_unregister_layoutdriver(&flexfilelayout_type);
2149 MODULE_ALIAS("nfs-layouttype4-4");
2151 MODULE_LICENSE("GPL");
2152 MODULE_DESCRIPTION("The NFSv4 flexfile layout driver");
2154 module_init(nfs4flexfilelayout_init);
2155 module_exit(nfs4flexfilelayout_exit);