2 * Device operations for the pnfs nfs4 file 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/vmalloc.h>
11 #include <linux/module.h>
12 #include <linux/sunrpc/addr.h>
14 #include "../internal.h"
15 #include "../nfs4session.h"
16 #include "flexfilelayout.h"
18 #define NFSDBG_FACILITY NFSDBG_PNFS_LD
20 static unsigned int dataserver_timeo = NFS_DEF_TCP_RETRANS;
21 static unsigned int dataserver_retrans;
23 static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg);
25 void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
27 if (!IS_ERR_OR_NULL(mirror_ds))
28 nfs4_put_deviceid_node(&mirror_ds->id_node);
31 void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
33 nfs4_print_deviceid(&mirror_ds->id_node.deviceid);
34 nfs4_pnfs_ds_put(mirror_ds->ds);
35 kfree(mirror_ds->ds_versions);
36 kfree_rcu(mirror_ds, id_node.rcu);
39 /* Decode opaque device data and construct new_ds using it */
40 struct nfs4_ff_layout_ds *
41 nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
44 struct xdr_stream stream;
47 struct list_head dsaddrs;
48 struct nfs4_pnfs_ds_addr *da;
49 struct nfs4_ff_layout_ds *new_ds = NULL;
50 struct nfs4_ff_ds_version *ds_versions = NULL;
56 /* set up xdr stream */
57 scratch = alloc_page(gfp_flags);
61 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
65 nfs4_init_deviceid_node(&new_ds->id_node,
68 INIT_LIST_HEAD(&dsaddrs);
70 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
71 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
74 p = xdr_inline_decode(&stream, 4);
76 goto out_err_drain_dsaddrs;
77 mp_count = be32_to_cpup(p);
78 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
80 for (i = 0; i < mp_count; i++) {
82 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
85 list_add_tail(&da->da_node, &dsaddrs);
87 if (list_empty(&dsaddrs)) {
88 dprintk("%s: no suitable DS addresses found\n",
91 goto out_err_drain_dsaddrs;
95 p = xdr_inline_decode(&stream, 4);
97 goto out_err_drain_dsaddrs;
98 version_count = be32_to_cpup(p);
99 dprintk("%s: version count %d\n", __func__, version_count);
101 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
106 for (i = 0; i < version_count; i++) {
107 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
108 * tightly_coupled(4) */
109 p = xdr_inline_decode(&stream, 20);
111 goto out_err_drain_dsaddrs;
112 ds_versions[i].version = be32_to_cpup(p++);
113 ds_versions[i].minor_version = be32_to_cpup(p++);
114 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
115 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
116 ds_versions[i].tightly_coupled = be32_to_cpup(p);
118 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
119 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
120 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
121 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
124 * check for valid major/minor combination.
125 * currently we support dataserver which talk:
126 * v3, v4.0, v4.1, v4.2
128 if (!((ds_versions[i].version == 3 && ds_versions[i].minor_version == 0) ||
129 (ds_versions[i].version == 4 && ds_versions[i].minor_version < 3))) {
130 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
131 i, ds_versions[i].version,
132 ds_versions[i].minor_version);
133 ret = -EPROTONOSUPPORT;
134 goto out_err_drain_dsaddrs;
137 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
138 __func__, i, ds_versions[i].version,
139 ds_versions[i].minor_version,
140 ds_versions[i].rsize,
141 ds_versions[i].wsize,
142 ds_versions[i].tightly_coupled);
145 new_ds->ds_versions = ds_versions;
146 new_ds->ds_versions_cnt = version_count;
148 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
150 goto out_err_drain_dsaddrs;
152 /* If DS was already in cache, free ds addrs */
153 while (!list_empty(&dsaddrs)) {
154 da = list_first_entry(&dsaddrs,
155 struct nfs4_pnfs_ds_addr,
157 list_del_init(&da->da_node);
158 kfree(da->da_remotestr);
162 __free_page(scratch);
165 out_err_drain_dsaddrs:
166 while (!list_empty(&dsaddrs)) {
167 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
169 list_del_init(&da->da_node);
170 kfree(da->da_remotestr);
176 __free_page(scratch);
180 dprintk("%s ERROR: returning %d\n", __func__, ret);
184 static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
185 struct nfs4_deviceid_node *devid)
187 nfs4_delete_deviceid(devid->ld, devid->nfs_client, &devid->deviceid);
188 if (!ff_layout_has_available_ds(lseg))
189 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
193 static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
194 struct nfs4_ff_layout_mirror *mirror,
197 if (mirror == NULL || IS_ERR(mirror->mirror_ds))
199 if (mirror->mirror_ds == NULL) {
201 struct nfs4_deviceid_node *node;
202 struct pnfs_layout_hdr *lh = lseg->pls_layout;
203 struct nfs4_ff_layout_ds *mirror_ds = ERR_PTR(-ENODEV);
205 node = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
206 &mirror->devid, lh->plh_lc_cred,
209 mirror_ds = FF_LAYOUT_MIRROR_DS(node);
211 /* check for race with another call to this function */
212 if (cmpxchg(&mirror->mirror_ds, NULL, mirror_ds) &&
213 mirror_ds != ERR_PTR(-ENODEV))
214 nfs4_put_deviceid_node(node);
219 if (IS_ERR(mirror->mirror_ds))
222 if (mirror->mirror_ds->ds == NULL) {
223 struct nfs4_deviceid_node *devid;
224 devid = &mirror->mirror_ds->id_node;
225 ff_layout_mark_devid_invalid(lseg, devid);
230 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
234 static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
235 u64 offset, u64 length)
239 end = max_t(u64, pnfs_end_offset(err->offset, err->length),
240 pnfs_end_offset(offset, length));
241 err->offset = min_t(u64, err->offset, offset);
242 err->length = end - err->offset;
246 ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
247 const struct nfs4_ff_layout_ds_err *e2)
251 if (e1->opnum != e2->opnum)
252 return e1->opnum < e2->opnum ? -1 : 1;
253 if (e1->status != e2->status)
254 return e1->status < e2->status ? -1 : 1;
255 ret = memcmp(e1->stateid.data, e2->stateid.data,
256 sizeof(e1->stateid.data));
259 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
262 if (pnfs_end_offset(e1->offset, e1->length) < e2->offset)
264 if (e1->offset > pnfs_end_offset(e2->offset, e2->length))
266 /* If ranges overlap or are contiguous, they are the same */
271 ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
272 struct nfs4_ff_layout_ds_err *dserr)
274 struct nfs4_ff_layout_ds_err *err, *tmp;
275 struct list_head *head = &flo->error_list;
278 /* Do insertion sort w/ merges */
279 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
280 match = ff_ds_error_match(err, dserr);
284 /* Add entry "dserr" _before_ entry "err" */
288 /* Entries match, so merge "err" into "dserr" */
289 extend_ds_error(dserr, err->offset, err->length);
290 list_replace(&err->list, &dserr->list);
295 list_add_tail(&dserr->list, head);
298 int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
299 struct nfs4_ff_layout_mirror *mirror, u64 offset,
300 u64 length, int status, enum nfs_opnum4 opnum,
303 struct nfs4_ff_layout_ds_err *dserr;
308 if (mirror->mirror_ds == NULL)
311 dserr = kmalloc(sizeof(*dserr), gfp_flags);
315 INIT_LIST_HEAD(&dserr->list);
316 dserr->offset = offset;
317 dserr->length = length;
318 dserr->status = status;
319 dserr->opnum = opnum;
320 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
321 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
322 NFS4_DEVICEID4_SIZE);
324 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
325 ff_layout_add_ds_error_locked(flo, dserr);
326 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
331 static struct rpc_cred *
332 ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
334 struct rpc_cred *cred, __rcu **pcred;
336 if (iomode == IOMODE_READ)
337 pcred = &mirror->ro_cred;
339 pcred = &mirror->rw_cred;
343 cred = rcu_dereference(*pcred);
347 cred = get_rpccred_rcu(cred);
354 nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
356 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
357 struct nfs_fh *fh = NULL;
359 if (!ff_layout_mirror_valid(lseg, mirror, false)) {
360 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
361 __func__, mirror_idx);
365 /* FIXME: For now assume there is only 1 version available for the DS */
366 fh = &mirror->fh_versions[0];
372 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
373 * @lseg: the layout segment we're operating on
374 * @ds_idx: index of the DS to use
375 * @fail_return: return layout on connect failure?
377 * Try to prepare a DS connection to accept an RPC call. This involves
378 * selecting a mirror to use and connecting the client to it if it's not
381 * Since we only need a single functioning mirror to satisfy a read, we don't
382 * want to return the layout if there is one. For writes though, any down
383 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
384 * between the two cases.
386 * Returns a pointer to a connected DS object on success or NULL on failure.
388 struct nfs4_pnfs_ds *
389 nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
392 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
393 struct nfs4_pnfs_ds *ds = NULL;
394 struct nfs4_deviceid_node *devid;
395 struct inode *ino = lseg->pls_layout->plh_inode;
396 struct nfs_server *s = NFS_SERVER(ino);
397 unsigned int max_payload;
400 if (!ff_layout_mirror_valid(lseg, mirror, true)) {
401 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
406 devid = &mirror->mirror_ds->id_node;
407 if (ff_layout_test_devid_unavailable(devid))
410 ds = mirror->mirror_ds->ds;
411 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
416 /* FIXME: For now we assume the server sent only one version of NFS
419 status = nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
421 mirror->mirror_ds->ds_versions[0].version,
422 mirror->mirror_ds->ds_versions[0].minor_version);
424 /* connect success, check rsize/wsize limit */
427 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
429 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
430 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
431 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
432 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
436 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
437 mirror, lseg->pls_range.offset,
438 lseg->pls_range.length, NFS4ERR_NXIO,
439 OP_ILLEGAL, GFP_NOIO);
440 if (fail_return || !ff_layout_has_available_ds(lseg))
441 pnfs_error_mark_layout_for_return(ino, lseg);
448 ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
449 struct rpc_cred *mdscred)
451 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
452 struct rpc_cred *cred;
455 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
457 cred = get_rpccred(mdscred);
459 cred = get_rpccred(mdscred);
465 * Find or create a DS rpc client with th MDS server rpc client auth flavor
466 * in the nfs_client cl_ds_clients list.
469 nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
470 struct nfs_client *ds_clp, struct inode *inode)
472 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
474 switch (mirror->mirror_ds->ds_versions[0].version) {
476 /* For NFSv3 DS, flavor is set when creating DS connections */
477 return ds_clp->cl_rpcclient;
479 return nfs4_find_or_create_ds_client(ds_clp, inode);
485 void ff_layout_free_ds_ioerr(struct list_head *head)
487 struct nfs4_ff_layout_ds_err *err;
489 while (!list_empty(head)) {
490 err = list_first_entry(head,
491 struct nfs4_ff_layout_ds_err,
493 list_del(&err->list);
498 /* called with inode i_lock held */
499 int ff_layout_encode_ds_ioerr(struct xdr_stream *xdr, const struct list_head *head)
501 struct nfs4_ff_layout_ds_err *err;
504 list_for_each_entry(err, head, list) {
505 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
506 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
507 * + status(4) + opnum(4)
509 p = xdr_reserve_space(xdr,
510 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
513 p = xdr_encode_hyper(p, err->offset);
514 p = xdr_encode_hyper(p, err->length);
515 p = xdr_encode_opaque_fixed(p, &err->stateid,
518 *p++ = cpu_to_be32(1);
519 p = xdr_encode_opaque_fixed(p, &err->deviceid,
520 NFS4_DEVICEID4_SIZE);
521 *p++ = cpu_to_be32(err->status);
522 *p++ = cpu_to_be32(err->opnum);
523 dprintk("%s: offset %llu length %llu status %d op %d\n",
524 __func__, err->offset, err->length, err->status,
532 unsigned int do_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
533 const struct pnfs_layout_range *range,
534 struct list_head *head,
537 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
538 struct inode *inode = lo->plh_inode;
539 struct nfs4_ff_layout_ds_err *err, *n;
540 unsigned int ret = 0;
542 spin_lock(&inode->i_lock);
543 list_for_each_entry_safe(err, n, &flo->error_list, list) {
544 if (!pnfs_is_range_intersecting(err->offset,
545 pnfs_end_offset(err->offset, err->length),
547 pnfs_end_offset(range->offset, range->length)))
551 list_move(&err->list, head);
555 spin_unlock(&inode->i_lock);
559 unsigned int ff_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
560 const struct pnfs_layout_range *range,
561 struct list_head *head,
566 ret = do_layout_fetch_ds_ioerr(lo, range, head, maxnum);
567 /* If we're over the max, discard all remaining entries */
570 do_layout_fetch_ds_ioerr(lo, range, &discard, -1);
571 ff_layout_free_ds_ioerr(&discard);
576 static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
578 struct nfs4_ff_layout_mirror *mirror;
579 struct nfs4_deviceid_node *devid;
582 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
583 mirror = FF_LAYOUT_COMP(lseg, idx);
585 if (!mirror->mirror_ds)
587 if (IS_ERR(mirror->mirror_ds))
589 devid = &mirror->mirror_ds->id_node;
590 if (!ff_layout_test_devid_unavailable(devid))
598 static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
600 struct nfs4_ff_layout_mirror *mirror;
601 struct nfs4_deviceid_node *devid;
604 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
605 mirror = FF_LAYOUT_COMP(lseg, idx);
606 if (!mirror || IS_ERR(mirror->mirror_ds))
608 if (!mirror->mirror_ds)
610 devid = &mirror->mirror_ds->id_node;
611 if (ff_layout_test_devid_unavailable(devid))
615 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
618 static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
620 if (lseg->pls_range.iomode == IOMODE_READ)
621 return ff_read_layout_has_available_ds(lseg);
622 /* Note: RW layout needs all mirrors available */
623 return ff_rw_layout_has_available_ds(lseg);
626 bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
628 return ff_layout_no_fallback_to_mds(lseg) ||
629 ff_layout_has_available_ds(lseg);
632 bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
634 return lseg->pls_range.iomode == IOMODE_RW &&
635 ff_layout_no_read_on_rw(lseg);
638 module_param(dataserver_retrans, uint, 0644);
639 MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
640 "retries a request before it attempts further "
641 " recovery action.");
642 module_param(dataserver_timeo, uint, 0644);
643 MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
644 "NFSv4.1 client waits for a response from a "
645 " data server before it retries an NFS request.");