Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Device operations for the pnfs nfs4 file layout driver.
4 *
5 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
6 *
7 * Tao Peng <bergwolf@primarydata.com>
8 */
9
10#include <linux/nfs_fs.h>
11#include <linux/vmalloc.h>
12#include <linux/module.h>
13#include <linux/sunrpc/addr.h>
14
15#include "../internal.h"
16#include "../nfs4session.h"
17#include "flexfilelayout.h"
18
19#define NFSDBG_FACILITY NFSDBG_PNFS_LD
20
21static unsigned int dataserver_timeo = NFS_DEF_TCP_RETRANS;
22static unsigned int dataserver_retrans;
23
24static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg);
25
26void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
27{
28 if (!IS_ERR_OR_NULL(mirror_ds))
29 nfs4_put_deviceid_node(&mirror_ds->id_node);
30}
31
32void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
33{
34 nfs4_print_deviceid(&mirror_ds->id_node.deviceid);
35 nfs4_pnfs_ds_put(mirror_ds->ds);
36 kfree(mirror_ds->ds_versions);
37 kfree_rcu(mirror_ds, id_node.rcu);
38}
39
40/* Decode opaque device data and construct new_ds using it */
41struct nfs4_ff_layout_ds *
42nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
43 gfp_t gfp_flags)
44{
45 struct xdr_stream stream;
46 struct xdr_buf buf;
47 struct page *scratch;
48 struct list_head dsaddrs;
49 struct nfs4_pnfs_ds_addr *da;
50 struct nfs4_ff_layout_ds *new_ds = NULL;
51 struct nfs4_ff_ds_version *ds_versions = NULL;
52 u32 mp_count;
53 u32 version_count;
54 __be32 *p;
55 int i, ret = -ENOMEM;
56
57 /* set up xdr stream */
58 scratch = alloc_page(gfp_flags);
59 if (!scratch)
60 goto out_err;
61
62 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
63 if (!new_ds)
64 goto out_scratch;
65
66 nfs4_init_deviceid_node(&new_ds->id_node,
67 server,
68 &pdev->dev_id);
69 INIT_LIST_HEAD(&dsaddrs);
70
71 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
72 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
73
74 /* multipath count */
75 p = xdr_inline_decode(&stream, 4);
76 if (unlikely(!p))
77 goto out_err_drain_dsaddrs;
78 mp_count = be32_to_cpup(p);
79 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
80
81 for (i = 0; i < mp_count; i++) {
82 /* multipath ds */
83 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
84 &stream, gfp_flags);
85 if (da)
86 list_add_tail(&da->da_node, &dsaddrs);
87 }
88 if (list_empty(&dsaddrs)) {
89 dprintk("%s: no suitable DS addresses found\n",
90 __func__);
91 ret = -ENOMEDIUM;
92 goto out_err_drain_dsaddrs;
93 }
94
95 /* version count */
96 p = xdr_inline_decode(&stream, 4);
97 if (unlikely(!p))
98 goto out_err_drain_dsaddrs;
99 version_count = be32_to_cpup(p);
100 dprintk("%s: version count %d\n", __func__, version_count);
101
102 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
103 gfp_flags);
104 if (!ds_versions)
105 goto out_scratch;
106
107 for (i = 0; i < version_count; i++) {
108 /* 20 = version(4) + minor_version(4) + rsize(4) + wsize(4) +
109 * tightly_coupled(4) */
110 p = xdr_inline_decode(&stream, 20);
111 if (unlikely(!p))
112 goto out_err_drain_dsaddrs;
113 ds_versions[i].version = be32_to_cpup(p++);
114 ds_versions[i].minor_version = be32_to_cpup(p++);
115 ds_versions[i].rsize = nfs_block_size(be32_to_cpup(p++), NULL);
116 ds_versions[i].wsize = nfs_block_size(be32_to_cpup(p++), NULL);
117 ds_versions[i].tightly_coupled = be32_to_cpup(p);
118
119 if (ds_versions[i].rsize > NFS_MAX_FILE_IO_SIZE)
120 ds_versions[i].rsize = NFS_MAX_FILE_IO_SIZE;
121 if (ds_versions[i].wsize > NFS_MAX_FILE_IO_SIZE)
122 ds_versions[i].wsize = NFS_MAX_FILE_IO_SIZE;
123
124 /*
125 * check for valid major/minor combination.
126 * currently we support dataserver which talk:
127 * v3, v4.0, v4.1, v4.2
128 */
129 if (!((ds_versions[i].version == 3 && ds_versions[i].minor_version == 0) ||
130 (ds_versions[i].version == 4 && ds_versions[i].minor_version < 3))) {
131 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
132 i, ds_versions[i].version,
133 ds_versions[i].minor_version);
134 ret = -EPROTONOSUPPORT;
135 goto out_err_drain_dsaddrs;
136 }
137
138 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
139 __func__, i, ds_versions[i].version,
140 ds_versions[i].minor_version,
141 ds_versions[i].rsize,
142 ds_versions[i].wsize,
143 ds_versions[i].tightly_coupled);
144 }
145
146 new_ds->ds_versions = ds_versions;
147 new_ds->ds_versions_cnt = version_count;
148
149 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
150 if (!new_ds->ds)
151 goto out_err_drain_dsaddrs;
152
153 /* If DS was already in cache, free ds addrs */
154 while (!list_empty(&dsaddrs)) {
155 da = list_first_entry(&dsaddrs,
156 struct nfs4_pnfs_ds_addr,
157 da_node);
158 list_del_init(&da->da_node);
159 kfree(da->da_remotestr);
160 kfree(da);
161 }
162
163 __free_page(scratch);
164 return new_ds;
165
166out_err_drain_dsaddrs:
167 while (!list_empty(&dsaddrs)) {
168 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
169 da_node);
170 list_del_init(&da->da_node);
171 kfree(da->da_remotestr);
172 kfree(da);
173 }
174
175 kfree(ds_versions);
176out_scratch:
177 __free_page(scratch);
178out_err:
179 kfree(new_ds);
180
181 dprintk("%s ERROR: returning %d\n", __func__, ret);
182 return NULL;
183}
184
185static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
186 struct nfs4_deviceid_node *devid)
187{
188 nfs4_delete_deviceid(devid->ld, devid->nfs_client, &devid->deviceid);
189 if (!ff_layout_has_available_ds(lseg))
190 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
191 lseg);
192}
193
194static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
195 struct nfs4_ff_layout_mirror *mirror,
196 bool create)
197{
198 if (mirror == NULL || IS_ERR(mirror->mirror_ds))
199 goto outerr;
200 if (mirror->mirror_ds == NULL) {
201 if (create) {
202 struct nfs4_deviceid_node *node;
203 struct pnfs_layout_hdr *lh = lseg->pls_layout;
204 struct nfs4_ff_layout_ds *mirror_ds = ERR_PTR(-ENODEV);
205
206 node = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
207 &mirror->devid, lh->plh_lc_cred,
208 GFP_KERNEL);
209 if (node)
210 mirror_ds = FF_LAYOUT_MIRROR_DS(node);
211
212 /* check for race with another call to this function */
213 if (cmpxchg(&mirror->mirror_ds, NULL, mirror_ds) &&
214 mirror_ds != ERR_PTR(-ENODEV))
215 nfs4_put_deviceid_node(node);
216 } else
217 goto outerr;
218 }
219
220 if (IS_ERR(mirror->mirror_ds))
221 goto outerr;
222
223 if (mirror->mirror_ds->ds == NULL) {
224 struct nfs4_deviceid_node *devid;
225 devid = &mirror->mirror_ds->id_node;
226 ff_layout_mark_devid_invalid(lseg, devid);
227 return false;
228 }
229 return true;
230outerr:
231 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
232 return false;
233}
234
235static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
236 u64 offset, u64 length)
237{
238 u64 end;
239
240 end = max_t(u64, pnfs_end_offset(err->offset, err->length),
241 pnfs_end_offset(offset, length));
242 err->offset = min_t(u64, err->offset, offset);
243 err->length = end - err->offset;
244}
245
246static int
247ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
248 const struct nfs4_ff_layout_ds_err *e2)
249{
250 int ret;
251
252 if (e1->opnum != e2->opnum)
253 return e1->opnum < e2->opnum ? -1 : 1;
254 if (e1->status != e2->status)
255 return e1->status < e2->status ? -1 : 1;
256 ret = memcmp(e1->stateid.data, e2->stateid.data,
257 sizeof(e1->stateid.data));
258 if (ret != 0)
259 return ret;
260 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
261 if (ret != 0)
262 return ret;
263 if (pnfs_end_offset(e1->offset, e1->length) < e2->offset)
264 return -1;
265 if (e1->offset > pnfs_end_offset(e2->offset, e2->length))
266 return 1;
267 /* If ranges overlap or are contiguous, they are the same */
268 return 0;
269}
270
271static void
272ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
273 struct nfs4_ff_layout_ds_err *dserr)
274{
275 struct nfs4_ff_layout_ds_err *err, *tmp;
276 struct list_head *head = &flo->error_list;
277 int match;
278
279 /* Do insertion sort w/ merges */
280 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
281 match = ff_ds_error_match(err, dserr);
282 if (match < 0)
283 continue;
284 if (match > 0) {
285 /* Add entry "dserr" _before_ entry "err" */
286 head = &err->list;
287 break;
288 }
289 /* Entries match, so merge "err" into "dserr" */
290 extend_ds_error(dserr, err->offset, err->length);
291 list_replace(&err->list, &dserr->list);
292 kfree(err);
293 return;
294 }
295
296 list_add_tail(&dserr->list, head);
297}
298
299int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
300 struct nfs4_ff_layout_mirror *mirror, u64 offset,
301 u64 length, int status, enum nfs_opnum4 opnum,
302 gfp_t gfp_flags)
303{
304 struct nfs4_ff_layout_ds_err *dserr;
305
306 if (status == 0)
307 return 0;
308
309 if (mirror->mirror_ds == NULL)
310 return -EINVAL;
311
312 dserr = kmalloc(sizeof(*dserr), gfp_flags);
313 if (!dserr)
314 return -ENOMEM;
315
316 INIT_LIST_HEAD(&dserr->list);
317 dserr->offset = offset;
318 dserr->length = length;
319 dserr->status = status;
320 dserr->opnum = opnum;
321 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
322 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
323 NFS4_DEVICEID4_SIZE);
324
325 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
326 ff_layout_add_ds_error_locked(flo, dserr);
327 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
328
329 return 0;
330}
331
332static struct rpc_cred *
333ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
334{
335 struct rpc_cred *cred, __rcu **pcred;
336
337 if (iomode == IOMODE_READ)
338 pcred = &mirror->ro_cred;
339 else
340 pcred = &mirror->rw_cred;
341
342 rcu_read_lock();
343 do {
344 cred = rcu_dereference(*pcred);
345 if (!cred)
346 break;
347
348 cred = get_rpccred_rcu(cred);
349 } while(!cred);
350 rcu_read_unlock();
351 return cred;
352}
353
354struct nfs_fh *
355nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
356{
357 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
358 struct nfs_fh *fh = NULL;
359
360 if (!ff_layout_mirror_valid(lseg, mirror, false)) {
361 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
362 __func__, mirror_idx);
363 goto out;
364 }
365
366 /* FIXME: For now assume there is only 1 version available for the DS */
367 fh = &mirror->fh_versions[0];
368out:
369 return fh;
370}
371
372/**
373 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
374 * @lseg: the layout segment we're operating on
375 * @ds_idx: index of the DS to use
376 * @fail_return: return layout on connect failure?
377 *
378 * Try to prepare a DS connection to accept an RPC call. This involves
379 * selecting a mirror to use and connecting the client to it if it's not
380 * already connected.
381 *
382 * Since we only need a single functioning mirror to satisfy a read, we don't
383 * want to return the layout if there is one. For writes though, any down
384 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
385 * between the two cases.
386 *
387 * Returns a pointer to a connected DS object on success or NULL on failure.
388 */
389struct nfs4_pnfs_ds *
390nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
391 bool fail_return)
392{
393 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
394 struct nfs4_pnfs_ds *ds = NULL;
395 struct nfs4_deviceid_node *devid;
396 struct inode *ino = lseg->pls_layout->plh_inode;
397 struct nfs_server *s = NFS_SERVER(ino);
398 unsigned int max_payload;
399 int status;
400
401 if (!ff_layout_mirror_valid(lseg, mirror, true)) {
402 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
403 __func__, ds_idx);
404 goto out;
405 }
406
407 devid = &mirror->mirror_ds->id_node;
408 if (ff_layout_test_devid_unavailable(devid))
409 goto out_fail;
410
411 ds = mirror->mirror_ds->ds;
412 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
413 smp_rmb();
414 if (ds->ds_clp)
415 goto out;
416
417 /* FIXME: For now we assume the server sent only one version of NFS
418 * to use for the DS.
419 */
420 status = nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
421 dataserver_retrans,
422 mirror->mirror_ds->ds_versions[0].version,
423 mirror->mirror_ds->ds_versions[0].minor_version);
424
425 /* connect success, check rsize/wsize limit */
426 if (!status) {
427 max_payload =
428 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
429 NULL);
430 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
431 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
432 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
433 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
434 goto out;
435 }
436out_fail:
437 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
438 mirror, lseg->pls_range.offset,
439 lseg->pls_range.length, NFS4ERR_NXIO,
440 OP_ILLEGAL, GFP_NOIO);
441 if (fail_return || !ff_layout_has_available_ds(lseg))
442 pnfs_error_mark_layout_for_return(ino, lseg);
443 ds = NULL;
444out:
445 return ds;
446}
447
448struct rpc_cred *
449ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
450 struct rpc_cred *mdscred)
451{
452 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
453 struct rpc_cred *cred;
454
455 if (mirror) {
456 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
457 if (!cred)
458 cred = get_rpccred(mdscred);
459 } else {
460 cred = get_rpccred(mdscred);
461 }
462 return cred;
463}
464
465/**
466* Find or create a DS rpc client with th MDS server rpc client auth flavor
467* in the nfs_client cl_ds_clients list.
468*/
469struct rpc_clnt *
470nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
471 struct nfs_client *ds_clp, struct inode *inode)
472{
473 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
474
475 switch (mirror->mirror_ds->ds_versions[0].version) {
476 case 3:
477 /* For NFSv3 DS, flavor is set when creating DS connections */
478 return ds_clp->cl_rpcclient;
479 case 4:
480 return nfs4_find_or_create_ds_client(ds_clp, inode);
481 default:
482 BUG();
483 }
484}
485
486void ff_layout_free_ds_ioerr(struct list_head *head)
487{
488 struct nfs4_ff_layout_ds_err *err;
489
490 while (!list_empty(head)) {
491 err = list_first_entry(head,
492 struct nfs4_ff_layout_ds_err,
493 list);
494 list_del(&err->list);
495 kfree(err);
496 }
497}
498
499/* called with inode i_lock held */
500int ff_layout_encode_ds_ioerr(struct xdr_stream *xdr, const struct list_head *head)
501{
502 struct nfs4_ff_layout_ds_err *err;
503 __be32 *p;
504
505 list_for_each_entry(err, head, list) {
506 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
507 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
508 * + status(4) + opnum(4)
509 */
510 p = xdr_reserve_space(xdr,
511 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
512 if (unlikely(!p))
513 return -ENOBUFS;
514 p = xdr_encode_hyper(p, err->offset);
515 p = xdr_encode_hyper(p, err->length);
516 p = xdr_encode_opaque_fixed(p, &err->stateid,
517 NFS4_STATEID_SIZE);
518 /* Encode 1 error */
519 *p++ = cpu_to_be32(1);
520 p = xdr_encode_opaque_fixed(p, &err->deviceid,
521 NFS4_DEVICEID4_SIZE);
522 *p++ = cpu_to_be32(err->status);
523 *p++ = cpu_to_be32(err->opnum);
524 dprintk("%s: offset %llu length %llu status %d op %d\n",
525 __func__, err->offset, err->length, err->status,
526 err->opnum);
527 }
528
529 return 0;
530}
531
532static
533unsigned int do_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
534 const struct pnfs_layout_range *range,
535 struct list_head *head,
536 unsigned int maxnum)
537{
538 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
539 struct inode *inode = lo->plh_inode;
540 struct nfs4_ff_layout_ds_err *err, *n;
541 unsigned int ret = 0;
542
543 spin_lock(&inode->i_lock);
544 list_for_each_entry_safe(err, n, &flo->error_list, list) {
545 if (!pnfs_is_range_intersecting(err->offset,
546 pnfs_end_offset(err->offset, err->length),
547 range->offset,
548 pnfs_end_offset(range->offset, range->length)))
549 continue;
550 if (!maxnum)
551 break;
552 list_move(&err->list, head);
553 maxnum--;
554 ret++;
555 }
556 spin_unlock(&inode->i_lock);
557 return ret;
558}
559
560unsigned int ff_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
561 const struct pnfs_layout_range *range,
562 struct list_head *head,
563 unsigned int maxnum)
564{
565 unsigned int ret;
566
567 ret = do_layout_fetch_ds_ioerr(lo, range, head, maxnum);
568 /* If we're over the max, discard all remaining entries */
569 if (ret == maxnum) {
570 LIST_HEAD(discard);
571 do_layout_fetch_ds_ioerr(lo, range, &discard, -1);
572 ff_layout_free_ds_ioerr(&discard);
573 }
574 return ret;
575}
576
577static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
578{
579 struct nfs4_ff_layout_mirror *mirror;
580 struct nfs4_deviceid_node *devid;
581 u32 idx;
582
583 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
584 mirror = FF_LAYOUT_COMP(lseg, idx);
585 if (mirror) {
586 if (!mirror->mirror_ds)
587 return true;
588 if (IS_ERR(mirror->mirror_ds))
589 continue;
590 devid = &mirror->mirror_ds->id_node;
591 if (!ff_layout_test_devid_unavailable(devid))
592 return true;
593 }
594 }
595
596 return false;
597}
598
599static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
600{
601 struct nfs4_ff_layout_mirror *mirror;
602 struct nfs4_deviceid_node *devid;
603 u32 idx;
604
605 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
606 mirror = FF_LAYOUT_COMP(lseg, idx);
607 if (!mirror || IS_ERR(mirror->mirror_ds))
608 return false;
609 if (!mirror->mirror_ds)
610 continue;
611 devid = &mirror->mirror_ds->id_node;
612 if (ff_layout_test_devid_unavailable(devid))
613 return false;
614 }
615
616 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
617}
618
619static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
620{
621 if (lseg->pls_range.iomode == IOMODE_READ)
622 return ff_read_layout_has_available_ds(lseg);
623 /* Note: RW layout needs all mirrors available */
624 return ff_rw_layout_has_available_ds(lseg);
625}
626
627bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
628{
629 return ff_layout_no_fallback_to_mds(lseg) ||
630 ff_layout_has_available_ds(lseg);
631}
632
633bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
634{
635 return lseg->pls_range.iomode == IOMODE_RW &&
636 ff_layout_no_read_on_rw(lseg);
637}
638
639module_param(dataserver_retrans, uint, 0644);
640MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
641 "retries a request before it attempts further "
642 " recovery action.");
643module_param(dataserver_timeo, uint, 0644);
644MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
645 "NFSv4.1 client waits for a response from a "
646 " data server before it retries an NFS request.");
1/*
2 * Device operations for the pnfs nfs4 file layout driver.
3 *
4 * Copyright (c) 2014, Primary Data, Inc. All rights reserved.
5 *
6 * Tao Peng <bergwolf@primarydata.com>
7 */
8
9#include <linux/nfs_fs.h>
10#include <linux/vmalloc.h>
11#include <linux/module.h>
12#include <linux/sunrpc/addr.h>
13
14#include "../internal.h"
15#include "../nfs4session.h"
16#include "flexfilelayout.h"
17
18#define NFSDBG_FACILITY NFSDBG_PNFS_LD
19
20static unsigned int dataserver_timeo = NFS_DEF_TCP_RETRANS;
21static unsigned int dataserver_retrans;
22
23static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg);
24
25void nfs4_ff_layout_put_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
26{
27 if (!IS_ERR_OR_NULL(mirror_ds))
28 nfs4_put_deviceid_node(&mirror_ds->id_node);
29}
30
31void nfs4_ff_layout_free_deviceid(struct nfs4_ff_layout_ds *mirror_ds)
32{
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);
36}
37
38/* Decode opaque device data and construct new_ds using it */
39struct nfs4_ff_layout_ds *
40nfs4_ff_alloc_deviceid_node(struct nfs_server *server, struct pnfs_device *pdev,
41 gfp_t gfp_flags)
42{
43 struct xdr_stream stream;
44 struct xdr_buf buf;
45 struct page *scratch;
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;
50 u32 mp_count;
51 u32 version_count;
52 __be32 *p;
53 int i, ret = -ENOMEM;
54
55 /* set up xdr stream */
56 scratch = alloc_page(gfp_flags);
57 if (!scratch)
58 goto out_err;
59
60 new_ds = kzalloc(sizeof(struct nfs4_ff_layout_ds), gfp_flags);
61 if (!new_ds)
62 goto out_scratch;
63
64 nfs4_init_deviceid_node(&new_ds->id_node,
65 server,
66 &pdev->dev_id);
67 INIT_LIST_HEAD(&dsaddrs);
68
69 xdr_init_decode_pages(&stream, &buf, pdev->pages, pdev->pglen);
70 xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
71
72 /* multipath count */
73 p = xdr_inline_decode(&stream, 4);
74 if (unlikely(!p))
75 goto out_err_drain_dsaddrs;
76 mp_count = be32_to_cpup(p);
77 dprintk("%s: multipath ds count %d\n", __func__, mp_count);
78
79 for (i = 0; i < mp_count; i++) {
80 /* multipath ds */
81 da = nfs4_decode_mp_ds_addr(server->nfs_client->cl_net,
82 &stream, gfp_flags);
83 if (da)
84 list_add_tail(&da->da_node, &dsaddrs);
85 }
86 if (list_empty(&dsaddrs)) {
87 dprintk("%s: no suitable DS addresses found\n",
88 __func__);
89 ret = -ENOMEDIUM;
90 goto out_err_drain_dsaddrs;
91 }
92
93 /* version count */
94 p = xdr_inline_decode(&stream, 4);
95 if (unlikely(!p))
96 goto out_err_drain_dsaddrs;
97 version_count = be32_to_cpup(p);
98 dprintk("%s: version count %d\n", __func__, version_count);
99
100 ds_versions = kzalloc(version_count * sizeof(struct nfs4_ff_ds_version),
101 gfp_flags);
102 if (!ds_versions)
103 goto out_scratch;
104
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);
109 if (unlikely(!p))
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);
116
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;
121
122 if (ds_versions[i].version != 3 || ds_versions[i].minor_version != 0) {
123 dprintk("%s: [%d] unsupported ds version %d-%d\n", __func__,
124 i, ds_versions[i].version,
125 ds_versions[i].minor_version);
126 ret = -EPROTONOSUPPORT;
127 goto out_err_drain_dsaddrs;
128 }
129
130 dprintk("%s: [%d] vers %u minor_ver %u rsize %u wsize %u coupled %d\n",
131 __func__, i, ds_versions[i].version,
132 ds_versions[i].minor_version,
133 ds_versions[i].rsize,
134 ds_versions[i].wsize,
135 ds_versions[i].tightly_coupled);
136 }
137
138 new_ds->ds_versions = ds_versions;
139 new_ds->ds_versions_cnt = version_count;
140
141 new_ds->ds = nfs4_pnfs_ds_add(&dsaddrs, gfp_flags);
142 if (!new_ds->ds)
143 goto out_err_drain_dsaddrs;
144
145 /* If DS was already in cache, free ds addrs */
146 while (!list_empty(&dsaddrs)) {
147 da = list_first_entry(&dsaddrs,
148 struct nfs4_pnfs_ds_addr,
149 da_node);
150 list_del_init(&da->da_node);
151 kfree(da->da_remotestr);
152 kfree(da);
153 }
154
155 __free_page(scratch);
156 return new_ds;
157
158out_err_drain_dsaddrs:
159 while (!list_empty(&dsaddrs)) {
160 da = list_first_entry(&dsaddrs, struct nfs4_pnfs_ds_addr,
161 da_node);
162 list_del_init(&da->da_node);
163 kfree(da->da_remotestr);
164 kfree(da);
165 }
166
167 kfree(ds_versions);
168out_scratch:
169 __free_page(scratch);
170out_err:
171 kfree(new_ds);
172
173 dprintk("%s ERROR: returning %d\n", __func__, ret);
174 return NULL;
175}
176
177static void ff_layout_mark_devid_invalid(struct pnfs_layout_segment *lseg,
178 struct nfs4_deviceid_node *devid)
179{
180 nfs4_delete_deviceid(devid->ld, devid->nfs_client, &devid->deviceid);
181 if (!ff_layout_has_available_ds(lseg))
182 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode,
183 lseg);
184}
185
186static bool ff_layout_mirror_valid(struct pnfs_layout_segment *lseg,
187 struct nfs4_ff_layout_mirror *mirror,
188 bool create)
189{
190 if (mirror == NULL || IS_ERR(mirror->mirror_ds))
191 goto outerr;
192 if (mirror->mirror_ds == NULL) {
193 if (create) {
194 struct nfs4_deviceid_node *node;
195 struct pnfs_layout_hdr *lh = lseg->pls_layout;
196 struct nfs4_ff_layout_ds *mirror_ds = ERR_PTR(-ENODEV);
197
198 node = nfs4_find_get_deviceid(NFS_SERVER(lh->plh_inode),
199 &mirror->devid, lh->plh_lc_cred,
200 GFP_KERNEL);
201 if (node)
202 mirror_ds = FF_LAYOUT_MIRROR_DS(node);
203
204 /* check for race with another call to this function */
205 if (cmpxchg(&mirror->mirror_ds, NULL, mirror_ds) &&
206 mirror_ds != ERR_PTR(-ENODEV))
207 nfs4_put_deviceid_node(node);
208 } else
209 goto outerr;
210 }
211
212 if (IS_ERR(mirror->mirror_ds))
213 goto outerr;
214
215 if (mirror->mirror_ds->ds == NULL) {
216 struct nfs4_deviceid_node *devid;
217 devid = &mirror->mirror_ds->id_node;
218 ff_layout_mark_devid_invalid(lseg, devid);
219 return false;
220 }
221 return true;
222outerr:
223 pnfs_error_mark_layout_for_return(lseg->pls_layout->plh_inode, lseg);
224 return false;
225}
226
227static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
228 u64 offset, u64 length)
229{
230 u64 end;
231
232 end = max_t(u64, pnfs_end_offset(err->offset, err->length),
233 pnfs_end_offset(offset, length));
234 err->offset = min_t(u64, err->offset, offset);
235 err->length = end - err->offset;
236}
237
238static int
239ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
240 const struct nfs4_ff_layout_ds_err *e2)
241{
242 int ret;
243
244 if (e1->opnum != e2->opnum)
245 return e1->opnum < e2->opnum ? -1 : 1;
246 if (e1->status != e2->status)
247 return e1->status < e2->status ? -1 : 1;
248 ret = memcmp(e1->stateid.data, e2->stateid.data,
249 sizeof(e1->stateid.data));
250 if (ret != 0)
251 return ret;
252 ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
253 if (ret != 0)
254 return ret;
255 if (pnfs_end_offset(e1->offset, e1->length) < e2->offset)
256 return -1;
257 if (e1->offset > pnfs_end_offset(e2->offset, e2->length))
258 return 1;
259 /* If ranges overlap or are contiguous, they are the same */
260 return 0;
261}
262
263static void
264ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
265 struct nfs4_ff_layout_ds_err *dserr)
266{
267 struct nfs4_ff_layout_ds_err *err, *tmp;
268 struct list_head *head = &flo->error_list;
269 int match;
270
271 /* Do insertion sort w/ merges */
272 list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
273 match = ff_ds_error_match(err, dserr);
274 if (match < 0)
275 continue;
276 if (match > 0) {
277 /* Add entry "dserr" _before_ entry "err" */
278 head = &err->list;
279 break;
280 }
281 /* Entries match, so merge "err" into "dserr" */
282 extend_ds_error(dserr, err->offset, err->length);
283 list_replace(&err->list, &dserr->list);
284 kfree(err);
285 return;
286 }
287
288 list_add_tail(&dserr->list, head);
289}
290
291int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
292 struct nfs4_ff_layout_mirror *mirror, u64 offset,
293 u64 length, int status, enum nfs_opnum4 opnum,
294 gfp_t gfp_flags)
295{
296 struct nfs4_ff_layout_ds_err *dserr;
297
298 if (status == 0)
299 return 0;
300
301 if (mirror->mirror_ds == NULL)
302 return -EINVAL;
303
304 dserr = kmalloc(sizeof(*dserr), gfp_flags);
305 if (!dserr)
306 return -ENOMEM;
307
308 INIT_LIST_HEAD(&dserr->list);
309 dserr->offset = offset;
310 dserr->length = length;
311 dserr->status = status;
312 dserr->opnum = opnum;
313 nfs4_stateid_copy(&dserr->stateid, &mirror->stateid);
314 memcpy(&dserr->deviceid, &mirror->mirror_ds->id_node.deviceid,
315 NFS4_DEVICEID4_SIZE);
316
317 spin_lock(&flo->generic_hdr.plh_inode->i_lock);
318 ff_layout_add_ds_error_locked(flo, dserr);
319 spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
320
321 return 0;
322}
323
324static struct rpc_cred *
325ff_layout_get_mirror_cred(struct nfs4_ff_layout_mirror *mirror, u32 iomode)
326{
327 struct rpc_cred *cred, __rcu **pcred;
328
329 if (iomode == IOMODE_READ)
330 pcred = &mirror->ro_cred;
331 else
332 pcred = &mirror->rw_cred;
333
334 rcu_read_lock();
335 do {
336 cred = rcu_dereference(*pcred);
337 if (!cred)
338 break;
339
340 cred = get_rpccred_rcu(cred);
341 } while(!cred);
342 rcu_read_unlock();
343 return cred;
344}
345
346struct nfs_fh *
347nfs4_ff_layout_select_ds_fh(struct pnfs_layout_segment *lseg, u32 mirror_idx)
348{
349 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, mirror_idx);
350 struct nfs_fh *fh = NULL;
351
352 if (!ff_layout_mirror_valid(lseg, mirror, false)) {
353 pr_err_ratelimited("NFS: %s: No data server for mirror offset index %d\n",
354 __func__, mirror_idx);
355 goto out;
356 }
357
358 /* FIXME: For now assume there is only 1 version available for the DS */
359 fh = &mirror->fh_versions[0];
360out:
361 return fh;
362}
363
364/**
365 * nfs4_ff_layout_prepare_ds - prepare a DS connection for an RPC call
366 * @lseg: the layout segment we're operating on
367 * @ds_idx: index of the DS to use
368 * @fail_return: return layout on connect failure?
369 *
370 * Try to prepare a DS connection to accept an RPC call. This involves
371 * selecting a mirror to use and connecting the client to it if it's not
372 * already connected.
373 *
374 * Since we only need a single functioning mirror to satisfy a read, we don't
375 * want to return the layout if there is one. For writes though, any down
376 * mirror should result in a LAYOUTRETURN. @fail_return is how we distinguish
377 * between the two cases.
378 *
379 * Returns a pointer to a connected DS object on success or NULL on failure.
380 */
381struct nfs4_pnfs_ds *
382nfs4_ff_layout_prepare_ds(struct pnfs_layout_segment *lseg, u32 ds_idx,
383 bool fail_return)
384{
385 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
386 struct nfs4_pnfs_ds *ds = NULL;
387 struct nfs4_deviceid_node *devid;
388 struct inode *ino = lseg->pls_layout->plh_inode;
389 struct nfs_server *s = NFS_SERVER(ino);
390 unsigned int max_payload;
391
392 if (!ff_layout_mirror_valid(lseg, mirror, true)) {
393 pr_err_ratelimited("NFS: %s: No data server for offset index %d\n",
394 __func__, ds_idx);
395 goto out;
396 }
397
398 devid = &mirror->mirror_ds->id_node;
399 if (ff_layout_test_devid_unavailable(devid))
400 goto out_fail;
401
402 ds = mirror->mirror_ds->ds;
403 /* matching smp_wmb() in _nfs4_pnfs_v3/4_ds_connect */
404 smp_rmb();
405 if (ds->ds_clp)
406 goto out;
407
408 /* FIXME: For now we assume the server sent only one version of NFS
409 * to use for the DS.
410 */
411 nfs4_pnfs_ds_connect(s, ds, devid, dataserver_timeo,
412 dataserver_retrans,
413 mirror->mirror_ds->ds_versions[0].version,
414 mirror->mirror_ds->ds_versions[0].minor_version);
415
416 /* connect success, check rsize/wsize limit */
417 if (ds->ds_clp) {
418 max_payload =
419 nfs_block_size(rpc_max_payload(ds->ds_clp->cl_rpcclient),
420 NULL);
421 if (mirror->mirror_ds->ds_versions[0].rsize > max_payload)
422 mirror->mirror_ds->ds_versions[0].rsize = max_payload;
423 if (mirror->mirror_ds->ds_versions[0].wsize > max_payload)
424 mirror->mirror_ds->ds_versions[0].wsize = max_payload;
425 goto out;
426 }
427 ff_layout_track_ds_error(FF_LAYOUT_FROM_HDR(lseg->pls_layout),
428 mirror, lseg->pls_range.offset,
429 lseg->pls_range.length, NFS4ERR_NXIO,
430 OP_ILLEGAL, GFP_NOIO);
431out_fail:
432 if (fail_return || !ff_layout_has_available_ds(lseg))
433 pnfs_error_mark_layout_for_return(ino, lseg);
434 ds = NULL;
435out:
436 return ds;
437}
438
439struct rpc_cred *
440ff_layout_get_ds_cred(struct pnfs_layout_segment *lseg, u32 ds_idx,
441 struct rpc_cred *mdscred)
442{
443 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
444 struct rpc_cred *cred;
445
446 if (mirror) {
447 cred = ff_layout_get_mirror_cred(mirror, lseg->pls_range.iomode);
448 if (!cred)
449 cred = get_rpccred(mdscred);
450 } else {
451 cred = get_rpccred(mdscred);
452 }
453 return cred;
454}
455
456/**
457* Find or create a DS rpc client with th MDS server rpc client auth flavor
458* in the nfs_client cl_ds_clients list.
459*/
460struct rpc_clnt *
461nfs4_ff_find_or_create_ds_client(struct pnfs_layout_segment *lseg, u32 ds_idx,
462 struct nfs_client *ds_clp, struct inode *inode)
463{
464 struct nfs4_ff_layout_mirror *mirror = FF_LAYOUT_COMP(lseg, ds_idx);
465
466 switch (mirror->mirror_ds->ds_versions[0].version) {
467 case 3:
468 /* For NFSv3 DS, flavor is set when creating DS connections */
469 return ds_clp->cl_rpcclient;
470 case 4:
471 return nfs4_find_or_create_ds_client(ds_clp, inode);
472 default:
473 BUG();
474 }
475}
476
477void ff_layout_free_ds_ioerr(struct list_head *head)
478{
479 struct nfs4_ff_layout_ds_err *err;
480
481 while (!list_empty(head)) {
482 err = list_first_entry(head,
483 struct nfs4_ff_layout_ds_err,
484 list);
485 list_del(&err->list);
486 kfree(err);
487 }
488}
489
490/* called with inode i_lock held */
491int ff_layout_encode_ds_ioerr(struct xdr_stream *xdr, const struct list_head *head)
492{
493 struct nfs4_ff_layout_ds_err *err;
494 __be32 *p;
495
496 list_for_each_entry(err, head, list) {
497 /* offset(8) + length(8) + stateid(NFS4_STATEID_SIZE)
498 * + array length + deviceid(NFS4_DEVICEID4_SIZE)
499 * + status(4) + opnum(4)
500 */
501 p = xdr_reserve_space(xdr,
502 28 + NFS4_STATEID_SIZE + NFS4_DEVICEID4_SIZE);
503 if (unlikely(!p))
504 return -ENOBUFS;
505 p = xdr_encode_hyper(p, err->offset);
506 p = xdr_encode_hyper(p, err->length);
507 p = xdr_encode_opaque_fixed(p, &err->stateid,
508 NFS4_STATEID_SIZE);
509 /* Encode 1 error */
510 *p++ = cpu_to_be32(1);
511 p = xdr_encode_opaque_fixed(p, &err->deviceid,
512 NFS4_DEVICEID4_SIZE);
513 *p++ = cpu_to_be32(err->status);
514 *p++ = cpu_to_be32(err->opnum);
515 dprintk("%s: offset %llu length %llu status %d op %d\n",
516 __func__, err->offset, err->length, err->status,
517 err->opnum);
518 }
519
520 return 0;
521}
522
523static
524unsigned int do_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
525 const struct pnfs_layout_range *range,
526 struct list_head *head,
527 unsigned int maxnum)
528{
529 struct nfs4_flexfile_layout *flo = FF_LAYOUT_FROM_HDR(lo);
530 struct inode *inode = lo->plh_inode;
531 struct nfs4_ff_layout_ds_err *err, *n;
532 unsigned int ret = 0;
533
534 spin_lock(&inode->i_lock);
535 list_for_each_entry_safe(err, n, &flo->error_list, list) {
536 if (!pnfs_is_range_intersecting(err->offset,
537 pnfs_end_offset(err->offset, err->length),
538 range->offset,
539 pnfs_end_offset(range->offset, range->length)))
540 continue;
541 if (!maxnum)
542 break;
543 list_move(&err->list, head);
544 maxnum--;
545 ret++;
546 }
547 spin_unlock(&inode->i_lock);
548 return ret;
549}
550
551unsigned int ff_layout_fetch_ds_ioerr(struct pnfs_layout_hdr *lo,
552 const struct pnfs_layout_range *range,
553 struct list_head *head,
554 unsigned int maxnum)
555{
556 unsigned int ret;
557
558 ret = do_layout_fetch_ds_ioerr(lo, range, head, maxnum);
559 /* If we're over the max, discard all remaining entries */
560 if (ret == maxnum) {
561 LIST_HEAD(discard);
562 do_layout_fetch_ds_ioerr(lo, range, &discard, -1);
563 ff_layout_free_ds_ioerr(&discard);
564 }
565 return ret;
566}
567
568static bool ff_read_layout_has_available_ds(struct pnfs_layout_segment *lseg)
569{
570 struct nfs4_ff_layout_mirror *mirror;
571 struct nfs4_deviceid_node *devid;
572 u32 idx;
573
574 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
575 mirror = FF_LAYOUT_COMP(lseg, idx);
576 if (mirror) {
577 if (!mirror->mirror_ds)
578 return true;
579 if (IS_ERR(mirror->mirror_ds))
580 continue;
581 devid = &mirror->mirror_ds->id_node;
582 if (!ff_layout_test_devid_unavailable(devid))
583 return true;
584 }
585 }
586
587 return false;
588}
589
590static bool ff_rw_layout_has_available_ds(struct pnfs_layout_segment *lseg)
591{
592 struct nfs4_ff_layout_mirror *mirror;
593 struct nfs4_deviceid_node *devid;
594 u32 idx;
595
596 for (idx = 0; idx < FF_LAYOUT_MIRROR_COUNT(lseg); idx++) {
597 mirror = FF_LAYOUT_COMP(lseg, idx);
598 if (!mirror || IS_ERR(mirror->mirror_ds))
599 return false;
600 if (!mirror->mirror_ds)
601 continue;
602 devid = &mirror->mirror_ds->id_node;
603 if (ff_layout_test_devid_unavailable(devid))
604 return false;
605 }
606
607 return FF_LAYOUT_MIRROR_COUNT(lseg) != 0;
608}
609
610static bool ff_layout_has_available_ds(struct pnfs_layout_segment *lseg)
611{
612 if (lseg->pls_range.iomode == IOMODE_READ)
613 return ff_read_layout_has_available_ds(lseg);
614 /* Note: RW layout needs all mirrors available */
615 return ff_rw_layout_has_available_ds(lseg);
616}
617
618bool ff_layout_avoid_mds_available_ds(struct pnfs_layout_segment *lseg)
619{
620 return ff_layout_no_fallback_to_mds(lseg) ||
621 ff_layout_has_available_ds(lseg);
622}
623
624bool ff_layout_avoid_read_on_rw(struct pnfs_layout_segment *lseg)
625{
626 return lseg->pls_range.iomode == IOMODE_RW &&
627 ff_layout_no_read_on_rw(lseg);
628}
629
630module_param(dataserver_retrans, uint, 0644);
631MODULE_PARM_DESC(dataserver_retrans, "The number of times the NFSv4.1 client "
632 "retries a request before it attempts further "
633 " recovery action.");
634module_param(dataserver_timeo, uint, 0644);
635MODULE_PARM_DESC(dataserver_timeo, "The time (in tenths of a second) the "
636 "NFSv4.1 client waits for a response from a "
637 " data server before it retries an NFS request.");