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_rcu(mirror_ds, id_node.rcu);
38 /* Decode opaque device data and construct new_ds using it */
39 struct nfs4_ff_layout_ds *
40 nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
43 struct xdr_stream stream;
46 struct list_head dsaddrs;
47 struct nfs4_pnfs_ds_addr *da;
48 struct nfs4_ff_layout_ds *new_ds = NULL;
49 struct nfs4_ff_ds_version *ds_versions = NULL;
55 /* set up xdr stream */
56 scratch = alloc_page(gfp_flags);
60 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
64 nfs4_init_deviceid_node(&new_ds->id_node,
67 INIT_LIST_HEAD(&dsaddrs);
69 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
70 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
73 p = xdr_inline_decode(&stream, 4);
75 goto out_err_drain_dsaddrs;
76 mp_count = be32_to_cpup(p);
77 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
79 for (i = 0; i < mp_count; i++) {
81 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
84 list_add_tail(&da->da_node, &dsaddrs);
86 if (list_empty(&dsaddrs)) {
87 dprintk("%s: no suitable DS addresses found\n",
90 goto out_err_drain_dsaddrs;
94 p = xdr_inline_decode(&stream, 4);
96 goto out_err_drain_dsaddrs;
97 version_count = be32_to_cpup(p);
98 dprintk("%s: version count %d\n", __func__, version_count);
100 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
105 for (i = 0; i < version_count; i++) {
106 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
107 * tightly_coupled(4) */
108 p = xdr_inline_decode(&stream, 20);
110 goto out_err_drain_dsaddrs;
111 ds_versions[i].version = be32_to_cpup(p++);
112 ds_versions[i].minor_version = be32_to_cpup(p++);
113 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
114 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
115 ds_versions[i].tightly_coupled = be32_to_cpup(p);
117 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
118 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
119 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
120 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
123 * check for valid major/minor combination.
124 * currently we support dataserver which talk:
125 * v3, v4.0, v4.1, v4.2
127 if (!((ds_versions[i].version == 3 && ds_versions[i].minor_version == 0) ||
128 (ds_versions[i].version == 4 && ds_versions[i].minor_version < 3))) {
129 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
130 i, ds_versions[i].version,
131 ds_versions[i].minor_version);
132 ret = -EPROTONOSUPPORT;
133 goto out_err_drain_dsaddrs;
136 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
137 __func__, i, ds_versions[i].version,
138 ds_versions[i].minor_version,
139 ds_versions[i].rsize,
140 ds_versions[i].wsize,
141 ds_versions[i].tightly_coupled);
144 new_ds->ds_versions = ds_versions;
145 new_ds->ds_versions_cnt = version_count;
147 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
149 goto out_err_drain_dsaddrs;
151 /* If DS was already in cache, free ds addrs */
152 while (!list_empty(&dsaddrs)) {
153 da = list_first_entry(&dsaddrs,
154 struct nfs4_pnfs_ds_addr,
156 list_del_init(&da->da_node);
157 kfree(da->da_remotestr);
161 __free_page(scratch);
164 out_err_drain_dsaddrs:
165 while (!list_empty(&dsaddrs)) {
166 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
168 list_del_init(&da->da_node);
169 kfree(da->da_remotestr);
175 __free_page(scratch);
179 dprintk("%s ERROR: returning %d\n", __func__, ret);
183 static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
184 struct nfs4_deviceid_node *devid)
186 nfs4_delete_deviceid(devid->ld, devid->nfs_client, &devid->deviceid);
187 if (!ff_layout_has_available_ds(lseg))
188 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
192 static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
193 struct nfs4_ff_layout_mirror *mirror,
196 if (mirror == NULL || IS_ERR(mirror->mirror_ds))
198 if (mirror->mirror_ds == NULL) {
200 struct nfs4_deviceid_node *node;
201 struct pnfs_layout_hdr *lh = lseg->pls_layout;
202 struct nfs4_ff_layout_ds *mirror_ds = ERR_PTR(-ENODEV);
204 node = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
205 &mirror->devid, lh->plh_lc_cred,
208 mirror_ds = FF_LAYOUT_MIRROR_DS(node);
210 /* check for race with another call to this function */
211 if (cmpxchg(&mirror->mirror_ds, NULL, mirror_ds) &&
212 mirror_ds != ERR_PTR(-ENODEV))
213 nfs4_put_deviceid_node(node);
218 if (IS_ERR(mirror->mirror_ds))
221 if (mirror->mirror_ds->ds == NULL) {
222 struct nfs4_deviceid_node *devid;
223 devid = &mirror->mirror_ds->id_node;
224 ff_layout_mark_devid_invalid(lseg, devid);
229 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
233 static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
234 u64 offset, u64 length)
238 end = max_t(u64, pnfs_end_offset(err->offset, err->length),
239 pnfs_end_offset(offset, length));
240 err->offset = min_t(u64, err->offset, offset);
241 err->length = end - err->offset;
245 ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
246 const struct nfs4_ff_layout_ds_err *e2)
250 if (e1->opnum != e2->opnum)
251 return e1->opnum < e2->opnum ? -1 : 1;
252 if (e1->status != e2->status)
253 return e1->status < e2->status ? -1 : 1;
254 ret = memcmp(e1->stateid.data, e2->stateid.data,
255 sizeof(e1->stateid.data));
258 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
261 if (pnfs_end_offset(e1->offset, e1->length) < e2->offset)
263 if (e1->offset > pnfs_end_offset(e2->offset, e2->length))
265 /* If ranges overlap or are contiguous, they are the same */
270 ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
271 struct nfs4_ff_layout_ds_err *dserr)
273 struct nfs4_ff_layout_ds_err *err, *tmp;
274 struct list_head *head = &flo->error_list;
277 /* Do insertion sort w/ merges */
278 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
279 match = ff_ds_error_match(err, dserr);
283 /* Add entry "dserr" _before_ entry "err" */
287 /* Entries match, so merge "err" into "dserr" */
288 extend_ds_error(dserr, err->offset, err->length);
289 list_replace(&err->list, &dserr->list);
294 list_add_tail(&dserr->list, head);
297 int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
298 struct nfs4_ff_layout_mirror *mirror, u64 offset,
299 u64 length, int status, enum nfs_opnum4 opnum,
302 struct nfs4_ff_layout_ds_err *dserr;
307 if (mirror->mirror_ds == NULL)
310 dserr = kmalloc(sizeof(*dserr), gfp_flags);
314 INIT_LIST_HEAD(&dserr->list);
315 dserr->offset = offset;
316 dserr->length = length;
317 dserr->status = status;
318 dserr->opnum = opnum;
319 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
320 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
321 NFS4_DEVICEID4_SIZE);
323 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
324 ff_layout_add_ds_error_locked(flo, dserr);
325 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
330 static struct rpc_cred *
331 ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
333 struct rpc_cred *cred, __rcu **pcred;
335 if (iomode == IOMODE_READ)
336 pcred = &mirror->ro_cred;
338 pcred = &mirror->rw_cred;
342 cred = rcu_dereference(*pcred);
346 cred = get_rpccred_rcu(cred);
353 nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
355 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
356 struct nfs_fh *fh = NULL;
358 if (!ff_layout_mirror_valid(lseg, mirror, false)) {
359 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
360 __func__, mirror_idx);
364 /* FIXME: For now assume there is only 1 version available for the DS */
365 fh = &mirror->fh_versions[0];
371 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
372 * @lseg: the layout segment we're operating on
373 * @ds_idx: index of the DS to use
374 * @fail_return: return layout on connect failure?
376 * Try to prepare a DS connection to accept an RPC call. This involves
377 * selecting a mirror to use and connecting the client to it if it's not
380 * Since we only need a single functioning mirror to satisfy a read, we don't
381 * want to return the layout if there is one. For writes though, any down
382 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
383 * between the two cases.
385 * Returns a pointer to a connected DS object on success or NULL on failure.
387 struct nfs4_pnfs_ds *
388 nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
391 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
392 struct nfs4_pnfs_ds *ds = NULL;
393 struct nfs4_deviceid_node *devid;
394 struct inode *ino = lseg->pls_layout->plh_inode;
395 struct nfs_server *s = NFS_SERVER(ino);
396 unsigned int max_payload;
399 if (!ff_layout_mirror_valid(lseg, mirror, true)) {
400 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
405 devid = &mirror->mirror_ds->id_node;
406 if (ff_layout_test_devid_unavailable(devid))
409 ds = mirror->mirror_ds->ds;
410 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
415 /* FIXME: For now we assume the server sent only one version of NFS
418 status = nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
420 mirror->mirror_ds->ds_versions[0].version,
421 mirror->mirror_ds->ds_versions[0].minor_version);
423 /* connect success, check rsize/wsize limit */
426 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
428 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
429 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
430 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
431 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
435 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
436 mirror, lseg->pls_range.offset,
437 lseg->pls_range.length, NFS4ERR_NXIO,
438 OP_ILLEGAL, GFP_NOIO);
439 if (fail_return || !ff_layout_has_available_ds(lseg))
440 pnfs_error_mark_layout_for_return(ino, lseg);
447 ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
448 struct rpc_cred *mdscred)
450 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
451 struct rpc_cred *cred;
454 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
456 cred = get_rpccred(mdscred);
458 cred = get_rpccred(mdscred);
464 * Find or create a DS rpc client with th MDS server rpc client auth flavor
465 * in the nfs_client cl_ds_clients list.
468 nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
469 struct nfs_client *ds_clp, struct inode *inode)
471 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
473 switch (mirror->mirror_ds->ds_versions[0].version) {
475 /* For NFSv3 DS, flavor is set when creating DS connections */
476 return ds_clp->cl_rpcclient;
478 return nfs4_find_or_create_ds_client(ds_clp, inode);
484 void ff_layout_free_ds_ioerr(struct list_head *head)
486 struct nfs4_ff_layout_ds_err *err;
488 while (!list_empty(head)) {
489 err = list_first_entry(head,
490 struct nfs4_ff_layout_ds_err,
492 list_del(&err->list);
497 /* called with inode i_lock held */
498 int ff_layout_encode_ds_ioerr(struct xdr_stream *xdr, const struct list_head *head)
500 struct nfs4_ff_layout_ds_err *err;
503 list_for_each_entry(err, head, list) {
504 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
505 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
506 * + status(4) + opnum(4)
508 p = xdr_reserve_space(xdr,
509 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
512 p = xdr_encode_hyper(p, err->offset);
513 p = xdr_encode_hyper(p, err->length);
514 p = xdr_encode_opaque_fixed(p, &err->stateid,
517 *p++ = cpu_to_be32(1);
518 p = xdr_encode_opaque_fixed(p, &err->deviceid,
519 NFS4_DEVICEID4_SIZE);
520 *p++ = cpu_to_be32(err->status);
521 *p++ = cpu_to_be32(err->opnum);
522 dprintk("%s: offset %llu length %llu status %d op %d\n",
523 __func__, err->offset, err->length, err->status,
531 unsigned int do_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
532 const struct pnfs_layout_range *range,
533 struct list_head *head,
536 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
537 struct inode *inode = lo->plh_inode;
538 struct nfs4_ff_layout_ds_err *err, *n;
539 unsigned int ret = 0;
541 spin_lock(&inode->i_lock);
542 list_for_each_entry_safe(err, n, &flo->error_list, list) {
543 if (!pnfs_is_range_intersecting(err->offset,
544 pnfs_end_offset(err->offset, err->length),
546 pnfs_end_offset(range->offset, range->length)))
550 list_move(&err->list, head);
554 spin_unlock(&inode->i_lock);
558 unsigned int ff_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
559 const struct pnfs_layout_range *range,
560 struct list_head *head,
565 ret = do_layout_fetch_ds_ioerr(lo, range, head, maxnum);
566 /* If we're over the max, discard all remaining entries */
569 do_layout_fetch_ds_ioerr(lo, range, &discard, -1);
570 ff_layout_free_ds_ioerr(&discard);
575 static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
577 struct nfs4_ff_layout_mirror *mirror;
578 struct nfs4_deviceid_node *devid;
581 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
582 mirror = FF_LAYOUT_COMP(lseg, idx);
584 if (!mirror->mirror_ds)
586 if (IS_ERR(mirror->mirror_ds))
588 devid = &mirror->mirror_ds->id_node;
589 if (!ff_layout_test_devid_unavailable(devid))
597 static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
599 struct nfs4_ff_layout_mirror *mirror;
600 struct nfs4_deviceid_node *devid;
603 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
604 mirror = FF_LAYOUT_COMP(lseg, idx);
605 if (!mirror || IS_ERR(mirror->mirror_ds))
607 if (!mirror->mirror_ds)
609 devid = &mirror->mirror_ds->id_node;
610 if (ff_layout_test_devid_unavailable(devid))
614 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
617 static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
619 if (lseg->pls_range.iomode == IOMODE_READ)
620 return ff_read_layout_has_available_ds(lseg);
621 /* Note: RW layout needs all mirrors available */
622 return ff_rw_layout_has_available_ds(lseg);
625 bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
627 return ff_layout_no_fallback_to_mds(lseg) ||
628 ff_layout_has_available_ds(lseg);
631 bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
633 return lseg->pls_range.iomode == IOMODE_RW &&
634 ff_layout_no_read_on_rw(lseg);
637 module_param(dataserver_retrans, uint, 0644);
638 MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
639 "retries a request before it attempts further "
640 " recovery action.");
641 module_param(dataserver_timeo, uint, 0644);
642 MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
643 "NFSv4.1 client waits for a response from a "
644 " data server before it retries an NFS request.");