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_TIMEO;
21 static unsigned int dataserver_retrans;
23 void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
26 nfs4_put_deviceid_node(&mirror_ds->id_node);
29 void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
31 nfs4_print_deviceid(&mirror_ds->id_node.deviceid);
32 nfs4_pnfs_ds_put(mirror_ds->ds);
33 kfree(mirror_ds->ds_versions);
34 kfree_rcu(mirror_ds, id_node.rcu);
37 /* Decode opaque device data and construct new_ds using it */
38 struct nfs4_ff_layout_ds *
39 nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
42 struct xdr_stream stream;
45 struct list_head dsaddrs;
46 struct nfs4_pnfs_ds_addr *da;
47 struct nfs4_ff_layout_ds *new_ds = NULL;
48 struct nfs4_ff_ds_version *ds_versions = NULL;
54 /* set up xdr stream */
55 scratch = alloc_page(gfp_flags);
59 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
63 nfs4_init_deviceid_node(&new_ds->id_node,
66 INIT_LIST_HEAD(&dsaddrs);
68 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
69 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
72 p = xdr_inline_decode(&stream, 4);
74 goto out_err_drain_dsaddrs;
75 mp_count = be32_to_cpup(p);
76 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
78 for (i = 0; i < mp_count; i++) {
80 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
83 list_add_tail(&da->da_node, &dsaddrs);
85 if (list_empty(&dsaddrs)) {
86 dprintk("%s: no suitable DS addresses found\n",
89 goto out_err_drain_dsaddrs;
93 p = xdr_inline_decode(&stream, 4);
95 goto out_err_drain_dsaddrs;
96 version_count = be32_to_cpup(p);
97 dprintk("%s: version count %d\n", __func__, version_count);
99 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
104 for (i = 0; i < version_count; i++) {
105 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
106 * tightly_coupled(4) */
107 p = xdr_inline_decode(&stream, 20);
109 goto out_err_drain_dsaddrs;
110 ds_versions[i].version = be32_to_cpup(p++);
111 ds_versions[i].minor_version = be32_to_cpup(p++);
112 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
113 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
114 ds_versions[i].tightly_coupled = be32_to_cpup(p);
116 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
117 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
118 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
119 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
121 if (ds_versions[i].version != 3 || ds_versions[i].minor_version != 0) {
122 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
123 i, ds_versions[i].version,
124 ds_versions[i].minor_version);
125 ret = -EPROTONOSUPPORT;
126 goto out_err_drain_dsaddrs;
129 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
130 __func__, i, ds_versions[i].version,
131 ds_versions[i].minor_version,
132 ds_versions[i].rsize,
133 ds_versions[i].wsize,
134 ds_versions[i].tightly_coupled);
137 new_ds->ds_versions = ds_versions;
138 new_ds->ds_versions_cnt = version_count;
140 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
142 goto out_err_drain_dsaddrs;
144 /* If DS was already in cache, free ds addrs */
145 while (!list_empty(&dsaddrs)) {
146 da = list_first_entry(&dsaddrs,
147 struct nfs4_pnfs_ds_addr,
149 list_del_init(&da->da_node);
150 kfree(da->da_remotestr);
154 __free_page(scratch);
157 out_err_drain_dsaddrs:
158 while (!list_empty(&dsaddrs)) {
159 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
161 list_del_init(&da->da_node);
162 kfree(da->da_remotestr);
168 __free_page(scratch);
172 dprintk("%s ERROR: returning %d\n", __func__, ret);
176 static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
177 struct nfs4_deviceid_node *devid)
179 nfs4_mark_deviceid_unavailable(devid);
180 if (!ff_layout_has_available_ds(lseg))
181 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
185 static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
186 struct nfs4_ff_layout_mirror *mirror)
188 if (mirror == NULL || mirror->mirror_ds == NULL) {
189 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
193 if (mirror->mirror_ds->ds == NULL) {
194 struct nfs4_deviceid_node *devid;
195 devid = &mirror->mirror_ds->id_node;
196 ff_layout_mark_devid_invalid(lseg, devid);
203 end_offset(u64 start, u64 len)
208 return end >= start ? end : NFS4_MAX_UINT64;
211 static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
212 u64 offset, u64 length)
216 end = max_t(u64, end_offset(err->offset, err->length),
217 end_offset(offset, length));
218 err->offset = min_t(u64, err->offset, offset);
219 err->length = end - err->offset;
223 ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
224 const struct nfs4_ff_layout_ds_err *e2)
228 if (e1->opnum != e2->opnum)
229 return e1->opnum < e2->opnum ? -1 : 1;
230 if (e1->status != e2->status)
231 return e1->status < e2->status ? -1 : 1;
232 ret = memcmp(e1->stateid.data, e2->stateid.data,
233 sizeof(e1->stateid.data));
236 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
239 if (end_offset(e1->offset, e1->length) < e2->offset)
241 if (e1->offset > end_offset(e2->offset, e2->length))
243 /* If ranges overlap or are contiguous, they are the same */
248 ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
249 struct nfs4_ff_layout_ds_err *dserr)
251 struct nfs4_ff_layout_ds_err *err, *tmp;
252 struct list_head *head = &flo->error_list;
255 /* Do insertion sort w/ merges */
256 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
257 match = ff_ds_error_match(err, dserr);
261 /* Add entry "dserr" _before_ entry "err" */
265 /* Entries match, so merge "err" into "dserr" */
266 extend_ds_error(dserr, err->offset, err->length);
267 list_del(&err->list);
271 list_add_tail(&dserr->list, head);
274 int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
275 struct nfs4_ff_layout_mirror *mirror, u64 offset,
276 u64 length, int status, enum nfs_opnum4 opnum,
279 struct nfs4_ff_layout_ds_err *dserr;
284 if (mirror->mirror_ds == NULL)
287 dserr = kmalloc(sizeof(*dserr), gfp_flags);
291 INIT_LIST_HEAD(&dserr->list);
292 dserr->offset = offset;
293 dserr->length = length;
294 dserr->status = status;
295 dserr->opnum = opnum;
296 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
297 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
298 NFS4_DEVICEID4_SIZE);
300 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
301 ff_layout_add_ds_error_locked(flo, dserr);
302 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
307 static struct rpc_cred *
308 ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
310 struct rpc_cred *cred, __rcu **pcred;
312 if (iomode == IOMODE_READ)
313 pcred = &mirror->ro_cred;
315 pcred = &mirror->rw_cred;
319 cred = rcu_dereference(*pcred);
323 cred = get_rpccred_rcu(cred);
330 nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
332 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
333 struct nfs_fh *fh = NULL;
335 if (!ff_layout_mirror_valid(lseg, mirror)) {
336 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
337 __func__, mirror_idx);
341 /* FIXME: For now assume there is only 1 version available for the DS */
342 fh = &mirror->fh_versions[0];
348 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
349 * @lseg: the layout segment we're operating on
350 * @ds_idx: index of the DS to use
351 * @fail_return: return layout on connect failure?
353 * Try to prepare a DS connection to accept an RPC call. This involves
354 * selecting a mirror to use and connecting the client to it if it's not
357 * Since we only need a single functioning mirror to satisfy a read, we don't
358 * want to return the layout if there is one. For writes though, any down
359 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
360 * between the two cases.
362 * Returns a pointer to a connected DS object on success or NULL on failure.
364 struct nfs4_pnfs_ds *
365 nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
368 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
369 struct nfs4_pnfs_ds *ds = NULL;
370 struct nfs4_deviceid_node *devid;
371 struct inode *ino = lseg->pls_layout->plh_inode;
372 struct nfs_server *s = NFS_SERVER(ino);
373 unsigned int max_payload;
375 if (!ff_layout_mirror_valid(lseg, mirror)) {
376 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
381 devid = &mirror->mirror_ds->id_node;
382 if (ff_layout_test_devid_unavailable(devid))
385 ds = mirror->mirror_ds->ds;
386 if (READ_ONCE(ds->ds_clp))
388 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
391 /* FIXME: For now we assume the server sent only one version of NFS
394 nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
396 mirror->mirror_ds->ds_versions[0].version,
397 mirror->mirror_ds->ds_versions[0].minor_version,
400 /* connect success, check rsize/wsize limit */
403 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
405 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
406 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
407 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
408 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
411 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
412 mirror, lseg->pls_range.offset,
413 lseg->pls_range.length, NFS4ERR_NXIO,
414 OP_ILLEGAL, GFP_NOIO);
416 if (fail_return || !ff_layout_has_available_ds(lseg))
417 pnfs_error_mark_layout_for_return(ino, lseg);
424 ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
425 struct rpc_cred *mdscred)
427 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
428 struct rpc_cred *cred;
431 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
433 cred = get_rpccred(mdscred);
435 cred = get_rpccred(mdscred);
441 * Find or create a DS rpc client with th MDS server rpc client auth flavor
442 * in the nfs_client cl_ds_clients list.
445 nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
446 struct nfs_client *ds_clp, struct inode *inode)
448 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
450 switch (mirror->mirror_ds->ds_versions[0].version) {
452 /* For NFSv3 DS, flavor is set when creating DS connections */
453 return ds_clp->cl_rpcclient;
455 return nfs4_find_or_create_ds_client(ds_clp, inode);
461 static bool is_range_intersecting(u64 offset1, u64 length1,
462 u64 offset2, u64 length2)
464 u64 end1 = end_offset(offset1, length1);
465 u64 end2 = end_offset(offset2, length2);
467 return (end1 == NFS4_MAX_UINT64 || end1 > offset2) &&
468 (end2 == NFS4_MAX_UINT64 || end2 > offset1);
471 /* called with inode i_lock held */
472 int ff_layout_encode_ds_ioerr(struct nfs4_flexfile_layout *flo,
473 struct xdr_stream *xdr, int *count,
474 const struct pnfs_layout_range *range)
476 struct nfs4_ff_layout_ds_err *err, *n;
479 list_for_each_entry_safe(err, n, &flo->error_list, list) {
480 if (!is_range_intersecting(err->offset, err->length,
481 range->offset, range->length))
483 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
484 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
485 * + status(4) + opnum(4)
487 p = xdr_reserve_space(xdr,
488 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
491 p = xdr_encode_hyper(p, err->offset);
492 p = xdr_encode_hyper(p, err->length);
493 p = xdr_encode_opaque_fixed(p, &err->stateid,
496 *p++ = cpu_to_be32(1);
497 p = xdr_encode_opaque_fixed(p, &err->deviceid,
498 NFS4_DEVICEID4_SIZE);
499 *p++ = cpu_to_be32(err->status);
500 *p++ = cpu_to_be32(err->opnum);
502 list_del(&err->list);
503 dprintk("%s: offset %llu length %llu status %d op %d count %d\n",
504 __func__, err->offset, err->length, err->status,
512 static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
514 struct nfs4_ff_layout_mirror *mirror;
515 struct nfs4_deviceid_node *devid;
518 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
519 mirror = FF_LAYOUT_COMP(lseg, idx);
520 if (mirror && mirror->mirror_ds) {
521 devid = &mirror->mirror_ds->id_node;
522 if (!ff_layout_test_devid_unavailable(devid))
530 static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
532 struct nfs4_ff_layout_mirror *mirror;
533 struct nfs4_deviceid_node *devid;
536 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
537 mirror = FF_LAYOUT_COMP(lseg, idx);
538 if (!mirror || !mirror->mirror_ds)
540 devid = &mirror->mirror_ds->id_node;
541 if (ff_layout_test_devid_unavailable(devid))
545 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
548 bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
550 if (lseg->pls_range.iomode == IOMODE_READ)
551 return ff_read_layout_has_available_ds(lseg);
552 /* Note: RW layout needs all mirrors available */
553 return ff_rw_layout_has_available_ds(lseg);
556 bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
558 return ff_layout_no_fallback_to_mds(lseg) ||
559 ff_layout_has_available_ds(lseg);
562 bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
564 return lseg->pls_range.iomode == IOMODE_RW &&
565 ff_layout_no_read_on_rw(lseg);
568 module_param(dataserver_retrans, uint, 0644);
569 MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
570 "retries a request before it attempts further "
571 " recovery action.");
572 module_param(dataserver_timeo, uint, 0644);
573 MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
574 "NFSv4.1 client waits for a response from a "
575 " data server before it retries an NFS request.");