Loading...
1/*
2 * pNFS functions to call and manage layout drivers.
3 *
4 * Copyright (c) 2002 [year of first publication]
5 * The Regents of the University of Michigan
6 * All Rights Reserved
7 *
8 * Dean Hildebrand <dhildebz@umich.edu>
9 *
10 * Permission is granted to use, copy, create derivative works, and
11 * redistribute this software and such derivative works for any purpose,
12 * so long as the name of the University of Michigan is not used in
13 * any advertising or publicity pertaining to the use or distribution
14 * of this software without specific, written prior authorization. If
15 * the above copyright notice or any other identification of the
16 * University of Michigan is included in any copy of any portion of
17 * this software, then the disclaimer below must also be included.
18 *
19 * This software is provided as is, without representation or warranty
20 * of any kind either express or implied, including without limitation
21 * the implied warranties of merchantability, fitness for a particular
22 * purpose, or noninfringement. The Regents of the University of
23 * Michigan shall not be liable for any damages, including special,
24 * indirect, incidental, or consequential damages, with respect to any
25 * claim arising out of or in connection with the use of the software,
26 * even if it has been or is hereafter advised of the possibility of
27 * such damages.
28 */
29
30#include <linux/nfs_fs.h>
31#include <linux/nfs_page.h>
32#include <linux/module.h>
33#include <linux/sort.h>
34#include "internal.h"
35#include "pnfs.h"
36#include "iostat.h"
37#include "nfs4trace.h"
38#include "delegation.h"
39#include "nfs42.h"
40#include "nfs4_fs.h"
41
42#define NFSDBG_FACILITY NFSDBG_PNFS
43#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
44
45/* Locking:
46 *
47 * pnfs_spinlock:
48 * protects pnfs_modules_tbl.
49 */
50static DEFINE_SPINLOCK(pnfs_spinlock);
51
52/*
53 * pnfs_modules_tbl holds all pnfs modules
54 */
55static LIST_HEAD(pnfs_modules_tbl);
56
57static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
58static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
59 struct list_head *free_me,
60 const struct pnfs_layout_range *range,
61 u32 seq);
62static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
63 struct list_head *tmp_list);
64
65/* Return the registered pnfs layout driver module matching given id */
66static struct pnfs_layoutdriver_type *
67find_pnfs_driver_locked(u32 id)
68{
69 struct pnfs_layoutdriver_type *local;
70
71 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
72 if (local->id == id)
73 goto out;
74 local = NULL;
75out:
76 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
77 return local;
78}
79
80static struct pnfs_layoutdriver_type *
81find_pnfs_driver(u32 id)
82{
83 struct pnfs_layoutdriver_type *local;
84
85 spin_lock(&pnfs_spinlock);
86 local = find_pnfs_driver_locked(id);
87 if (local != NULL && !try_module_get(local->owner)) {
88 dprintk("%s: Could not grab reference on module\n", __func__);
89 local = NULL;
90 }
91 spin_unlock(&pnfs_spinlock);
92 return local;
93}
94
95const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
96{
97 return find_pnfs_driver(id);
98}
99
100void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
101{
102 if (ld)
103 module_put(ld->owner);
104}
105
106void
107unset_pnfs_layoutdriver(struct nfs_server *nfss)
108{
109 if (nfss->pnfs_curr_ld) {
110 if (nfss->pnfs_curr_ld->clear_layoutdriver)
111 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
112 /* Decrement the MDS count. Purge the deviceid cache if zero */
113 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
114 nfs4_deviceid_purge_client(nfss->nfs_client);
115 module_put(nfss->pnfs_curr_ld->owner);
116 }
117 nfss->pnfs_curr_ld = NULL;
118}
119
120/*
121 * When the server sends a list of layout types, we choose one in the order
122 * given in the list below.
123 *
124 * FIXME: should this list be configurable in some fashion? module param?
125 * mount option? something else?
126 */
127static const u32 ld_prefs[] = {
128 LAYOUT_SCSI,
129 LAYOUT_BLOCK_VOLUME,
130 LAYOUT_OSD2_OBJECTS,
131 LAYOUT_FLEX_FILES,
132 LAYOUT_NFSV4_1_FILES,
133 0
134};
135
136static int
137ld_cmp(const void *e1, const void *e2)
138{
139 u32 ld1 = *((u32 *)e1);
140 u32 ld2 = *((u32 *)e2);
141 int i;
142
143 for (i = 0; ld_prefs[i] != 0; i++) {
144 if (ld1 == ld_prefs[i])
145 return -1;
146
147 if (ld2 == ld_prefs[i])
148 return 1;
149 }
150 return 0;
151}
152
153/*
154 * Try to set the server's pnfs module to the pnfs layout type specified by id.
155 * Currently only one pNFS layout driver per filesystem is supported.
156 *
157 * @ids array of layout types supported by MDS.
158 */
159void
160set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
161 struct nfs_fsinfo *fsinfo)
162{
163 struct pnfs_layoutdriver_type *ld_type = NULL;
164 u32 id;
165 int i;
166
167 if (fsinfo->nlayouttypes == 0)
168 goto out_no_driver;
169 if (!(server->nfs_client->cl_exchange_flags &
170 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
171 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
172 __func__, server->nfs_client->cl_exchange_flags);
173 goto out_no_driver;
174 }
175
176 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
177 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
178
179 for (i = 0; i < fsinfo->nlayouttypes; i++) {
180 id = fsinfo->layouttype[i];
181 ld_type = find_pnfs_driver(id);
182 if (!ld_type) {
183 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
184 id);
185 ld_type = find_pnfs_driver(id);
186 }
187 if (ld_type)
188 break;
189 }
190
191 if (!ld_type) {
192 dprintk("%s: No pNFS module found!\n", __func__);
193 goto out_no_driver;
194 }
195
196 server->pnfs_curr_ld = ld_type;
197 if (ld_type->set_layoutdriver
198 && ld_type->set_layoutdriver(server, mntfh)) {
199 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
200 "driver %u.\n", __func__, id);
201 module_put(ld_type->owner);
202 goto out_no_driver;
203 }
204 /* Bump the MDS count */
205 atomic_inc(&server->nfs_client->cl_mds_count);
206
207 dprintk("%s: pNFS module for %u set\n", __func__, id);
208 return;
209
210out_no_driver:
211 dprintk("%s: Using NFSv4 I/O\n", __func__);
212 server->pnfs_curr_ld = NULL;
213}
214
215int
216pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
217{
218 int status = -EINVAL;
219 struct pnfs_layoutdriver_type *tmp;
220
221 if (ld_type->id == 0) {
222 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
223 return status;
224 }
225 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
226 printk(KERN_ERR "NFS: %s Layout driver must provide "
227 "alloc_lseg and free_lseg.\n", __func__);
228 return status;
229 }
230
231 spin_lock(&pnfs_spinlock);
232 tmp = find_pnfs_driver_locked(ld_type->id);
233 if (!tmp) {
234 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
235 status = 0;
236 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
237 ld_type->name);
238 } else {
239 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
240 __func__, ld_type->id);
241 }
242 spin_unlock(&pnfs_spinlock);
243
244 return status;
245}
246EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
247
248void
249pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
250{
251 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
252 spin_lock(&pnfs_spinlock);
253 list_del(&ld_type->pnfs_tblid);
254 spin_unlock(&pnfs_spinlock);
255}
256EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
257
258/*
259 * pNFS client layout cache
260 */
261
262/* Need to hold i_lock if caller does not already hold reference */
263void
264pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
265{
266 refcount_inc(&lo->plh_refcount);
267}
268
269static struct pnfs_layout_hdr *
270pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
271{
272 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
273 return ld->alloc_layout_hdr(ino, gfp_flags);
274}
275
276static void
277pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
278{
279 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
280 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
281
282 if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
283 struct nfs_client *clp = server->nfs_client;
284
285 spin_lock(&clp->cl_lock);
286 list_del_rcu(&lo->plh_layouts);
287 spin_unlock(&clp->cl_lock);
288 }
289 put_cred(lo->plh_lc_cred);
290 return ld->free_layout_hdr(lo);
291}
292
293static void
294pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
295{
296 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
297 dprintk("%s: freeing layout cache %p\n", __func__, lo);
298 nfsi->layout = NULL;
299 /* Reset MDS Threshold I/O counters */
300 nfsi->write_io = 0;
301 nfsi->read_io = 0;
302}
303
304void
305pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
306{
307 struct inode *inode;
308 unsigned long i_state;
309
310 if (!lo)
311 return;
312 inode = lo->plh_inode;
313 pnfs_layoutreturn_before_put_layout_hdr(lo);
314
315 if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
316 if (!list_empty(&lo->plh_segs))
317 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
318 pnfs_detach_layout_hdr(lo);
319 i_state = inode->i_state;
320 spin_unlock(&inode->i_lock);
321 pnfs_free_layout_hdr(lo);
322 /* Notify pnfs_destroy_layout_final() that we're done */
323 if (i_state & (I_FREEING | I_CLEAR))
324 wake_up_var(lo);
325 }
326}
327
328static struct inode *
329pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
330{
331 struct inode *inode = igrab(lo->plh_inode);
332 if (inode)
333 return inode;
334 set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
335 return NULL;
336}
337
338/*
339 * Compare 2 layout stateid sequence ids, to see which is newer,
340 * taking into account wraparound issues.
341 */
342static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
343{
344 return (s32)(s1 - s2) > 0;
345}
346
347static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
348{
349 if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
350 lo->plh_barrier = newseq;
351}
352
353static void
354pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
355 u32 seq)
356{
357 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
358 iomode = IOMODE_ANY;
359 lo->plh_return_iomode = iomode;
360 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
361 /*
362 * We must set lo->plh_return_seq to avoid livelocks with
363 * pnfs_layout_need_return()
364 */
365 if (seq == 0)
366 seq = be32_to_cpu(lo->plh_stateid.seqid);
367 if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
368 lo->plh_return_seq = seq;
369 pnfs_barrier_update(lo, seq);
370}
371
372static void
373pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
374{
375 struct pnfs_layout_segment *lseg;
376 lo->plh_return_iomode = 0;
377 lo->plh_return_seq = 0;
378 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
379 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
380 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
381 continue;
382 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
383 }
384}
385
386static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
387{
388 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
389 clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
390 smp_mb__after_atomic();
391 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
392 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
393}
394
395static void
396pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
397 struct list_head *free_me)
398{
399 clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
400 clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
401 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
402 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
403 if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
404 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
405}
406
407/*
408 * Update the seqid of a layout stateid after receiving
409 * NFS4ERR_OLD_STATEID
410 */
411bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
412 struct pnfs_layout_range *dst_range,
413 struct inode *inode)
414{
415 struct pnfs_layout_hdr *lo;
416 struct pnfs_layout_range range = {
417 .iomode = IOMODE_ANY,
418 .offset = 0,
419 .length = NFS4_MAX_UINT64,
420 };
421 bool ret = false;
422 LIST_HEAD(head);
423 int err;
424
425 spin_lock(&inode->i_lock);
426 lo = NFS_I(inode)->layout;
427 if (lo && pnfs_layout_is_valid(lo) &&
428 nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
429 /* Is our call using the most recent seqid? If so, bump it */
430 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
431 nfs4_stateid_seqid_inc(dst);
432 ret = true;
433 goto out;
434 }
435 /* Try to update the seqid to the most recent */
436 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
437 if (err != -EBUSY) {
438 dst->seqid = lo->plh_stateid.seqid;
439 *dst_range = range;
440 ret = true;
441 }
442 }
443out:
444 spin_unlock(&inode->i_lock);
445 pnfs_free_lseg_list(&head);
446 return ret;
447}
448
449/*
450 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
451 *
452 * In order to continue using the pnfs_layout_hdr, a full recovery
453 * is required.
454 * Note that caller must hold inode->i_lock.
455 */
456int
457pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
458 struct list_head *lseg_list)
459{
460 struct pnfs_layout_range range = {
461 .iomode = IOMODE_ANY,
462 .offset = 0,
463 .length = NFS4_MAX_UINT64,
464 };
465 struct pnfs_layout_segment *lseg, *next;
466
467 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
468 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
469 pnfs_clear_lseg_state(lseg, lseg_list);
470 pnfs_clear_layoutreturn_info(lo);
471 pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
472 set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
473 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
474 !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
475 pnfs_clear_layoutreturn_waitbit(lo);
476 return !list_empty(&lo->plh_segs);
477}
478
479static int
480pnfs_iomode_to_fail_bit(u32 iomode)
481{
482 return iomode == IOMODE_RW ?
483 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
484}
485
486static void
487pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
488{
489 lo->plh_retry_timestamp = jiffies;
490 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
491 refcount_inc(&lo->plh_refcount);
492}
493
494static void
495pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
496{
497 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
498 refcount_dec(&lo->plh_refcount);
499}
500
501static void
502pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
503{
504 struct inode *inode = lo->plh_inode;
505 struct pnfs_layout_range range = {
506 .iomode = iomode,
507 .offset = 0,
508 .length = NFS4_MAX_UINT64,
509 };
510 LIST_HEAD(head);
511
512 spin_lock(&inode->i_lock);
513 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
514 pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
515 spin_unlock(&inode->i_lock);
516 pnfs_free_lseg_list(&head);
517 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
518 iomode == IOMODE_RW ? "RW" : "READ");
519}
520
521static bool
522pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
523{
524 unsigned long start, end;
525 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
526
527 if (test_bit(fail_bit, &lo->plh_flags) == 0)
528 return false;
529 end = jiffies;
530 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
531 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
532 /* It is time to retry the failed layoutgets */
533 pnfs_layout_clear_fail_bit(lo, fail_bit);
534 return false;
535 }
536 return true;
537}
538
539static void
540pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
541 const struct pnfs_layout_range *range,
542 const nfs4_stateid *stateid)
543{
544 INIT_LIST_HEAD(&lseg->pls_list);
545 INIT_LIST_HEAD(&lseg->pls_lc_list);
546 INIT_LIST_HEAD(&lseg->pls_commits);
547 refcount_set(&lseg->pls_refcount, 1);
548 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
549 lseg->pls_layout = lo;
550 lseg->pls_range = *range;
551 lseg->pls_seq = be32_to_cpu(stateid->seqid);
552}
553
554static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
555{
556 if (lseg != NULL) {
557 struct inode *inode = lseg->pls_layout->plh_inode;
558 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
559 }
560}
561
562static void
563pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
564 struct pnfs_layout_segment *lseg)
565{
566 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
567 list_del_init(&lseg->pls_list);
568 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
569 refcount_dec(&lo->plh_refcount);
570 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
571 return;
572 if (list_empty(&lo->plh_segs) &&
573 !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
574 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
575 if (atomic_read(&lo->plh_outstanding) == 0)
576 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
577 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
578 }
579}
580
581static bool
582pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
583 struct pnfs_layout_segment *lseg)
584{
585 if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
586 pnfs_layout_is_valid(lo)) {
587 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
588 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
589 return true;
590 }
591 return false;
592}
593
594void
595pnfs_put_lseg(struct pnfs_layout_segment *lseg)
596{
597 struct pnfs_layout_hdr *lo;
598 struct inode *inode;
599
600 if (!lseg)
601 return;
602
603 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
604 refcount_read(&lseg->pls_refcount),
605 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
606
607 lo = lseg->pls_layout;
608 inode = lo->plh_inode;
609
610 if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
611 pnfs_get_layout_hdr(lo);
612 pnfs_layout_remove_lseg(lo, lseg);
613 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
614 lseg = NULL;
615 spin_unlock(&inode->i_lock);
616 pnfs_free_lseg(lseg);
617 pnfs_put_layout_hdr(lo);
618 }
619}
620EXPORT_SYMBOL_GPL(pnfs_put_lseg);
621
622/*
623 * is l2 fully contained in l1?
624 * start1 end1
625 * [----------------------------------)
626 * start2 end2
627 * [----------------)
628 */
629static bool
630pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
631 const struct pnfs_layout_range *l2)
632{
633 u64 start1 = l1->offset;
634 u64 end1 = pnfs_end_offset(start1, l1->length);
635 u64 start2 = l2->offset;
636 u64 end2 = pnfs_end_offset(start2, l2->length);
637
638 return (start1 <= start2) && (end1 >= end2);
639}
640
641static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
642 struct list_head *tmp_list)
643{
644 if (!refcount_dec_and_test(&lseg->pls_refcount))
645 return false;
646 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
647 list_add(&lseg->pls_list, tmp_list);
648 return true;
649}
650
651/* Returns 1 if lseg is removed from list, 0 otherwise */
652static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
653 struct list_head *tmp_list)
654{
655 int rv = 0;
656
657 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
658 /* Remove the reference keeping the lseg in the
659 * list. It will now be removed when all
660 * outstanding io is finished.
661 */
662 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
663 refcount_read(&lseg->pls_refcount));
664 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
665 rv = 1;
666 }
667 return rv;
668}
669
670static bool
671pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
672 const struct pnfs_layout_range *recall_range)
673{
674 return (recall_range->iomode == IOMODE_ANY ||
675 lseg_range->iomode == recall_range->iomode) &&
676 pnfs_lseg_range_intersecting(lseg_range, recall_range);
677}
678
679static bool
680pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
681 const struct pnfs_layout_range *recall_range,
682 u32 seq)
683{
684 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
685 return false;
686 if (recall_range == NULL)
687 return true;
688 return pnfs_should_free_range(&lseg->pls_range, recall_range);
689}
690
691/**
692 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
693 * @lo: layout header containing the lsegs
694 * @tmp_list: list head where doomed lsegs should go
695 * @recall_range: optional recall range argument to match (may be NULL)
696 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
697 *
698 * Walk the list of lsegs in the layout header, and tear down any that should
699 * be destroyed. If "recall_range" is specified then the segment must match
700 * that range. If "seq" is non-zero, then only match segments that were handed
701 * out at or before that sequence.
702 *
703 * Returns number of matching invalid lsegs remaining in list after scanning
704 * it and purging them.
705 */
706int
707pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
708 struct list_head *tmp_list,
709 const struct pnfs_layout_range *recall_range,
710 u32 seq)
711{
712 struct pnfs_layout_segment *lseg, *next;
713 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
714 int remaining = 0;
715
716 dprintk("%s:Begin lo %p\n", __func__, lo);
717
718 if (list_empty(&lo->plh_segs))
719 return 0;
720 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
721 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
722 dprintk("%s: freeing lseg %p iomode %d seq %u "
723 "offset %llu length %llu\n", __func__,
724 lseg, lseg->pls_range.iomode, lseg->pls_seq,
725 lseg->pls_range.offset, lseg->pls_range.length);
726 if (mark_lseg_invalid(lseg, tmp_list))
727 continue;
728 remaining++;
729 pnfs_lseg_cancel_io(server, lseg);
730 }
731 dprintk("%s:Return %i\n", __func__, remaining);
732 return remaining;
733}
734
735static void
736pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
737 struct list_head *free_me,
738 const struct pnfs_layout_range *range,
739 u32 seq)
740{
741 struct pnfs_layout_segment *lseg, *next;
742
743 list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
744 if (pnfs_match_lseg_recall(lseg, range, seq))
745 list_move_tail(&lseg->pls_list, free_me);
746 }
747}
748
749/* note free_me must contain lsegs from a single layout_hdr */
750void
751pnfs_free_lseg_list(struct list_head *free_me)
752{
753 struct pnfs_layout_segment *lseg, *tmp;
754
755 if (list_empty(free_me))
756 return;
757
758 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
759 list_del(&lseg->pls_list);
760 pnfs_free_lseg(lseg);
761 }
762}
763
764static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
765{
766 struct pnfs_layout_hdr *lo;
767 LIST_HEAD(tmp_list);
768
769 spin_lock(&nfsi->vfs_inode.i_lock);
770 lo = nfsi->layout;
771 if (lo) {
772 pnfs_get_layout_hdr(lo);
773 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
774 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
775 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
776 spin_unlock(&nfsi->vfs_inode.i_lock);
777 pnfs_free_lseg_list(&tmp_list);
778 nfs_commit_inode(&nfsi->vfs_inode, 0);
779 pnfs_put_layout_hdr(lo);
780 } else
781 spin_unlock(&nfsi->vfs_inode.i_lock);
782 return lo;
783}
784
785void pnfs_destroy_layout(struct nfs_inode *nfsi)
786{
787 __pnfs_destroy_layout(nfsi);
788}
789EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
790
791static bool pnfs_layout_removed(struct nfs_inode *nfsi,
792 struct pnfs_layout_hdr *lo)
793{
794 bool ret;
795
796 spin_lock(&nfsi->vfs_inode.i_lock);
797 ret = nfsi->layout != lo;
798 spin_unlock(&nfsi->vfs_inode.i_lock);
799 return ret;
800}
801
802void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
803{
804 struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
805
806 if (lo)
807 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
808}
809
810static bool
811pnfs_layout_add_bulk_destroy_list(struct inode *inode,
812 struct list_head *layout_list)
813{
814 struct pnfs_layout_hdr *lo;
815 bool ret = false;
816
817 spin_lock(&inode->i_lock);
818 lo = NFS_I(inode)->layout;
819 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
820 pnfs_get_layout_hdr(lo);
821 list_add(&lo->plh_bulk_destroy, layout_list);
822 ret = true;
823 }
824 spin_unlock(&inode->i_lock);
825 return ret;
826}
827
828/* Caller must hold rcu_read_lock and clp->cl_lock */
829static int
830pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
831 struct nfs_server *server,
832 struct list_head *layout_list)
833 __must_hold(&clp->cl_lock)
834 __must_hold(RCU)
835{
836 struct pnfs_layout_hdr *lo, *next;
837 struct inode *inode;
838
839 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
840 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
841 test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
842 !list_empty(&lo->plh_bulk_destroy))
843 continue;
844 /* If the sb is being destroyed, just bail */
845 if (!nfs_sb_active(server->super))
846 break;
847 inode = pnfs_grab_inode_layout_hdr(lo);
848 if (inode != NULL) {
849 if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
850 list_del_rcu(&lo->plh_layouts);
851 if (pnfs_layout_add_bulk_destroy_list(inode,
852 layout_list))
853 continue;
854 rcu_read_unlock();
855 spin_unlock(&clp->cl_lock);
856 iput(inode);
857 } else {
858 rcu_read_unlock();
859 spin_unlock(&clp->cl_lock);
860 }
861 nfs_sb_deactive(server->super);
862 spin_lock(&clp->cl_lock);
863 rcu_read_lock();
864 return -EAGAIN;
865 }
866 return 0;
867}
868
869static int
870pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
871 bool is_bulk_recall)
872{
873 struct pnfs_layout_hdr *lo;
874 struct inode *inode;
875 LIST_HEAD(lseg_list);
876 int ret = 0;
877
878 while (!list_empty(layout_list)) {
879 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
880 plh_bulk_destroy);
881 dprintk("%s freeing layout for inode %lu\n", __func__,
882 lo->plh_inode->i_ino);
883 inode = lo->plh_inode;
884
885 pnfs_layoutcommit_inode(inode, false);
886
887 spin_lock(&inode->i_lock);
888 list_del_init(&lo->plh_bulk_destroy);
889 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
890 if (is_bulk_recall)
891 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
892 ret = -EAGAIN;
893 }
894 spin_unlock(&inode->i_lock);
895 pnfs_free_lseg_list(&lseg_list);
896 /* Free all lsegs that are attached to commit buckets */
897 nfs_commit_inode(inode, 0);
898 pnfs_put_layout_hdr(lo);
899 nfs_iput_and_deactive(inode);
900 }
901 return ret;
902}
903
904int
905pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
906 struct nfs_fsid *fsid,
907 bool is_recall)
908{
909 struct nfs_server *server;
910 LIST_HEAD(layout_list);
911
912 spin_lock(&clp->cl_lock);
913 rcu_read_lock();
914restart:
915 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
916 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
917 continue;
918 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
919 server,
920 &layout_list) != 0)
921 goto restart;
922 }
923 rcu_read_unlock();
924 spin_unlock(&clp->cl_lock);
925
926 if (list_empty(&layout_list))
927 return 0;
928 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
929}
930
931int
932pnfs_destroy_layouts_byclid(struct nfs_client *clp,
933 bool is_recall)
934{
935 struct nfs_server *server;
936 LIST_HEAD(layout_list);
937
938 spin_lock(&clp->cl_lock);
939 rcu_read_lock();
940restart:
941 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
942 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
943 server,
944 &layout_list) != 0)
945 goto restart;
946 }
947 rcu_read_unlock();
948 spin_unlock(&clp->cl_lock);
949
950 if (list_empty(&layout_list))
951 return 0;
952 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
953}
954
955/*
956 * Called by the state manager to remove all layouts established under an
957 * expired lease.
958 */
959void
960pnfs_destroy_all_layouts(struct nfs_client *clp)
961{
962 nfs4_deviceid_mark_client_invalid(clp);
963 nfs4_deviceid_purge_client(clp);
964
965 pnfs_destroy_layouts_byclid(clp, false);
966}
967
968static void
969pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
970{
971 const struct cred *old;
972
973 if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
974 old = xchg(&lo->plh_lc_cred, get_cred(cred));
975 put_cred(old);
976 }
977}
978
979/* update lo->plh_stateid with new if is more recent */
980void
981pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
982 const struct cred *cred, bool update_barrier)
983{
984 u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
985 u32 newseq = be32_to_cpu(new->seqid);
986
987 if (!pnfs_layout_is_valid(lo)) {
988 pnfs_set_layout_cred(lo, cred);
989 nfs4_stateid_copy(&lo->plh_stateid, new);
990 lo->plh_barrier = newseq;
991 pnfs_clear_layoutreturn_info(lo);
992 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
993 return;
994 }
995
996 if (pnfs_seqid_is_newer(newseq, oldseq))
997 nfs4_stateid_copy(&lo->plh_stateid, new);
998
999 if (update_barrier) {
1000 pnfs_barrier_update(lo, newseq);
1001 return;
1002 }
1003 /*
1004 * Because of wraparound, we want to keep the barrier
1005 * "close" to the current seqids. We really only want to
1006 * get here from a layoutget call.
1007 */
1008 if (atomic_read(&lo->plh_outstanding) == 1)
1009 pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1010}
1011
1012static bool
1013pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1014 const nfs4_stateid *stateid)
1015{
1016 u32 seqid = be32_to_cpu(stateid->seqid);
1017
1018 return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1019}
1020
1021/* lget is set to 1 if called from inside send_layoutget call chain */
1022static bool
1023pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1024{
1025 return lo->plh_block_lgets ||
1026 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1027}
1028
1029static struct nfs_server *
1030pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1031{
1032 struct nfs_server *server;
1033
1034 if (inode) {
1035 server = NFS_SERVER(inode);
1036 } else {
1037 struct dentry *parent_dir = dget_parent(ctx->dentry);
1038 server = NFS_SERVER(parent_dir->d_inode);
1039 dput(parent_dir);
1040 }
1041 return server;
1042}
1043
1044static void nfs4_free_pages(struct page **pages, size_t size)
1045{
1046 int i;
1047
1048 if (!pages)
1049 return;
1050
1051 for (i = 0; i < size; i++) {
1052 if (!pages[i])
1053 break;
1054 __free_page(pages[i]);
1055 }
1056 kfree(pages);
1057}
1058
1059static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1060{
1061 struct page **pages;
1062 int i;
1063
1064 pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1065 if (!pages) {
1066 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1067 return NULL;
1068 }
1069
1070 for (i = 0; i < size; i++) {
1071 pages[i] = alloc_page(gfp_flags);
1072 if (!pages[i]) {
1073 dprintk("%s: failed to allocate page\n", __func__);
1074 nfs4_free_pages(pages, i);
1075 return NULL;
1076 }
1077 }
1078
1079 return pages;
1080}
1081
1082static struct nfs4_layoutget *
1083pnfs_alloc_init_layoutget_args(struct inode *ino,
1084 struct nfs_open_context *ctx,
1085 const nfs4_stateid *stateid,
1086 const struct pnfs_layout_range *range,
1087 gfp_t gfp_flags)
1088{
1089 struct nfs_server *server = pnfs_find_server(ino, ctx);
1090 size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1091 size_t max_pages = max_response_pages(server);
1092 struct nfs4_layoutget *lgp;
1093
1094 dprintk("--> %s\n", __func__);
1095
1096 lgp = kzalloc(sizeof(*lgp), gfp_flags);
1097 if (lgp == NULL)
1098 return NULL;
1099
1100 if (max_reply_sz) {
1101 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1102 if (npages < max_pages)
1103 max_pages = npages;
1104 }
1105
1106 lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1107 if (!lgp->args.layout.pages) {
1108 kfree(lgp);
1109 return NULL;
1110 }
1111 lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1112 lgp->res.layoutp = &lgp->args.layout;
1113
1114 /* Don't confuse uninitialised result and success */
1115 lgp->res.status = -NFS4ERR_DELAY;
1116
1117 lgp->args.minlength = PAGE_SIZE;
1118 if (lgp->args.minlength > range->length)
1119 lgp->args.minlength = range->length;
1120 if (ino) {
1121 loff_t i_size = i_size_read(ino);
1122
1123 if (range->iomode == IOMODE_READ) {
1124 if (range->offset >= i_size)
1125 lgp->args.minlength = 0;
1126 else if (i_size - range->offset < lgp->args.minlength)
1127 lgp->args.minlength = i_size - range->offset;
1128 }
1129 }
1130 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1131 pnfs_copy_range(&lgp->args.range, range);
1132 lgp->args.type = server->pnfs_curr_ld->id;
1133 lgp->args.inode = ino;
1134 lgp->args.ctx = get_nfs_open_context(ctx);
1135 nfs4_stateid_copy(&lgp->args.stateid, stateid);
1136 lgp->gfp_flags = gfp_flags;
1137 lgp->cred = ctx->cred;
1138 return lgp;
1139}
1140
1141void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1142{
1143 size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1144
1145 nfs4_free_pages(lgp->args.layout.pages, max_pages);
1146 pnfs_put_layout_hdr(lgp->lo);
1147 put_nfs_open_context(lgp->args.ctx);
1148 kfree(lgp);
1149}
1150
1151static void pnfs_clear_layoutcommit(struct inode *inode,
1152 struct list_head *head)
1153{
1154 struct nfs_inode *nfsi = NFS_I(inode);
1155 struct pnfs_layout_segment *lseg, *tmp;
1156
1157 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1158 return;
1159 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1160 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1161 continue;
1162 pnfs_lseg_dec_and_remove_zero(lseg, head);
1163 }
1164}
1165
1166void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1167 const nfs4_stateid *arg_stateid,
1168 const struct pnfs_layout_range *range,
1169 const nfs4_stateid *stateid)
1170{
1171 struct inode *inode = lo->plh_inode;
1172 LIST_HEAD(freeme);
1173
1174 spin_lock(&inode->i_lock);
1175 if (!pnfs_layout_is_valid(lo) ||
1176 !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1177 goto out_unlock;
1178 if (stateid) {
1179 u32 seq = be32_to_cpu(arg_stateid->seqid);
1180
1181 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1182 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1183 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1184 } else
1185 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1186out_unlock:
1187 pnfs_clear_layoutreturn_waitbit(lo);
1188 spin_unlock(&inode->i_lock);
1189 pnfs_free_lseg_list(&freeme);
1190
1191}
1192
1193static bool
1194pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1195 nfs4_stateid *stateid,
1196 const struct cred **cred,
1197 enum pnfs_iomode *iomode)
1198{
1199 /* Serialise LAYOUTGET/LAYOUTRETURN */
1200 if (atomic_read(&lo->plh_outstanding) != 0)
1201 return false;
1202 if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1203 return false;
1204 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1205 pnfs_get_layout_hdr(lo);
1206 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1207 *cred = get_cred(lo->plh_lc_cred);
1208 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1209 if (lo->plh_return_seq != 0)
1210 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1211 if (iomode != NULL)
1212 *iomode = lo->plh_return_iomode;
1213 pnfs_clear_layoutreturn_info(lo);
1214 } else if (iomode != NULL)
1215 *iomode = IOMODE_ANY;
1216 pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1217 return true;
1218}
1219
1220static void
1221pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1222 struct pnfs_layout_hdr *lo,
1223 const nfs4_stateid *stateid,
1224 enum pnfs_iomode iomode)
1225{
1226 struct inode *inode = lo->plh_inode;
1227
1228 args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1229 args->inode = inode;
1230 args->range.iomode = iomode;
1231 args->range.offset = 0;
1232 args->range.length = NFS4_MAX_UINT64;
1233 args->layout = lo;
1234 nfs4_stateid_copy(&args->stateid, stateid);
1235}
1236
1237static int
1238pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1239 const nfs4_stateid *stateid,
1240 const struct cred **pcred,
1241 enum pnfs_iomode iomode,
1242 bool sync)
1243{
1244 struct inode *ino = lo->plh_inode;
1245 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1246 struct nfs4_layoutreturn *lrp;
1247 const struct cred *cred = *pcred;
1248 int status = 0;
1249
1250 *pcred = NULL;
1251 lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1252 if (unlikely(lrp == NULL)) {
1253 status = -ENOMEM;
1254 spin_lock(&ino->i_lock);
1255 pnfs_clear_layoutreturn_waitbit(lo);
1256 spin_unlock(&ino->i_lock);
1257 put_cred(cred);
1258 pnfs_put_layout_hdr(lo);
1259 goto out;
1260 }
1261
1262 pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1263 lrp->args.ld_private = &lrp->ld_private;
1264 lrp->clp = NFS_SERVER(ino)->nfs_client;
1265 lrp->cred = cred;
1266 if (ld->prepare_layoutreturn)
1267 ld->prepare_layoutreturn(&lrp->args);
1268
1269 status = nfs4_proc_layoutreturn(lrp, sync);
1270out:
1271 dprintk("<-- %s status: %d\n", __func__, status);
1272 return status;
1273}
1274
1275static bool
1276pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1277 enum pnfs_iomode iomode,
1278 u32 seq)
1279{
1280 struct pnfs_layout_range recall_range = {
1281 .length = NFS4_MAX_UINT64,
1282 .iomode = iomode,
1283 };
1284 return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1285 &recall_range, seq) != -EBUSY;
1286}
1287
1288/* Return true if layoutreturn is needed */
1289static bool
1290pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1291{
1292 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1293 return false;
1294 return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1295 lo->plh_return_seq);
1296}
1297
1298static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1299{
1300 struct inode *inode= lo->plh_inode;
1301
1302 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1303 return;
1304 spin_lock(&inode->i_lock);
1305 if (pnfs_layout_need_return(lo)) {
1306 const struct cred *cred;
1307 nfs4_stateid stateid;
1308 enum pnfs_iomode iomode;
1309 bool send;
1310
1311 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1312 spin_unlock(&inode->i_lock);
1313 if (send) {
1314 /* Send an async layoutreturn so we dont deadlock */
1315 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1316 }
1317 } else
1318 spin_unlock(&inode->i_lock);
1319}
1320
1321/*
1322 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1323 * when the layout segment list is empty.
1324 *
1325 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1326 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1327 * deviceid is marked invalid.
1328 */
1329int
1330_pnfs_return_layout(struct inode *ino)
1331{
1332 struct pnfs_layout_hdr *lo = NULL;
1333 struct nfs_inode *nfsi = NFS_I(ino);
1334 struct pnfs_layout_range range = {
1335 .iomode = IOMODE_ANY,
1336 .offset = 0,
1337 .length = NFS4_MAX_UINT64,
1338 };
1339 LIST_HEAD(tmp_list);
1340 const struct cred *cred;
1341 nfs4_stateid stateid;
1342 int status = 0;
1343 bool send, valid_layout;
1344
1345 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1346
1347 spin_lock(&ino->i_lock);
1348 lo = nfsi->layout;
1349 if (!lo) {
1350 spin_unlock(&ino->i_lock);
1351 dprintk("NFS: %s no layout to return\n", __func__);
1352 goto out;
1353 }
1354 /* Reference matched in nfs4_layoutreturn_release */
1355 pnfs_get_layout_hdr(lo);
1356 /* Is there an outstanding layoutreturn ? */
1357 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1358 spin_unlock(&ino->i_lock);
1359 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1360 TASK_UNINTERRUPTIBLE))
1361 goto out_put_layout_hdr;
1362 spin_lock(&ino->i_lock);
1363 }
1364 valid_layout = pnfs_layout_is_valid(lo);
1365 pnfs_clear_layoutcommit(ino, &tmp_list);
1366 pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1367
1368 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1369 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1370
1371 /* Don't send a LAYOUTRETURN if list was initially empty */
1372 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1373 !valid_layout) {
1374 spin_unlock(&ino->i_lock);
1375 dprintk("NFS: %s no layout segments to return\n", __func__);
1376 goto out_wait_layoutreturn;
1377 }
1378
1379 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1380 spin_unlock(&ino->i_lock);
1381 if (send)
1382 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1383out_wait_layoutreturn:
1384 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1385out_put_layout_hdr:
1386 pnfs_free_lseg_list(&tmp_list);
1387 pnfs_put_layout_hdr(lo);
1388out:
1389 dprintk("<-- %s status: %d\n", __func__, status);
1390 return status;
1391}
1392
1393int
1394pnfs_commit_and_return_layout(struct inode *inode)
1395{
1396 struct pnfs_layout_hdr *lo;
1397 int ret;
1398
1399 spin_lock(&inode->i_lock);
1400 lo = NFS_I(inode)->layout;
1401 if (lo == NULL) {
1402 spin_unlock(&inode->i_lock);
1403 return 0;
1404 }
1405 pnfs_get_layout_hdr(lo);
1406 /* Block new layoutgets and read/write to ds */
1407 lo->plh_block_lgets++;
1408 spin_unlock(&inode->i_lock);
1409 filemap_fdatawait(inode->i_mapping);
1410 ret = pnfs_layoutcommit_inode(inode, true);
1411 if (ret == 0)
1412 ret = _pnfs_return_layout(inode);
1413 spin_lock(&inode->i_lock);
1414 lo->plh_block_lgets--;
1415 spin_unlock(&inode->i_lock);
1416 pnfs_put_layout_hdr(lo);
1417 return ret;
1418}
1419
1420bool pnfs_roc(struct inode *ino,
1421 struct nfs4_layoutreturn_args *args,
1422 struct nfs4_layoutreturn_res *res,
1423 const struct cred *cred)
1424{
1425 struct nfs_inode *nfsi = NFS_I(ino);
1426 struct nfs_open_context *ctx;
1427 struct nfs4_state *state;
1428 struct pnfs_layout_hdr *lo;
1429 struct pnfs_layout_segment *lseg, *next;
1430 const struct cred *lc_cred;
1431 nfs4_stateid stateid;
1432 enum pnfs_iomode iomode = 0;
1433 bool layoutreturn = false, roc = false;
1434 bool skip_read = false;
1435
1436 if (!nfs_have_layout(ino))
1437 return false;
1438retry:
1439 rcu_read_lock();
1440 spin_lock(&ino->i_lock);
1441 lo = nfsi->layout;
1442 if (!lo || !pnfs_layout_is_valid(lo) ||
1443 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1444 lo = NULL;
1445 goto out_noroc;
1446 }
1447 pnfs_get_layout_hdr(lo);
1448 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1449 spin_unlock(&ino->i_lock);
1450 rcu_read_unlock();
1451 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1452 TASK_UNINTERRUPTIBLE);
1453 pnfs_put_layout_hdr(lo);
1454 goto retry;
1455 }
1456
1457 /* no roc if we hold a delegation */
1458 if (nfs4_check_delegation(ino, FMODE_READ)) {
1459 if (nfs4_check_delegation(ino, FMODE_WRITE))
1460 goto out_noroc;
1461 skip_read = true;
1462 }
1463
1464 list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1465 state = ctx->state;
1466 if (state == NULL)
1467 continue;
1468 /* Don't return layout if there is open file state */
1469 if (state->state & FMODE_WRITE)
1470 goto out_noroc;
1471 if (state->state & FMODE_READ)
1472 skip_read = true;
1473 }
1474
1475
1476 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1477 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1478 continue;
1479 /* If we are sending layoutreturn, invalidate all valid lsegs */
1480 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1481 continue;
1482 /*
1483 * Note: mark lseg for return so pnfs_layout_remove_lseg
1484 * doesn't invalidate the layout for us.
1485 */
1486 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1487 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1488 continue;
1489 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1490 }
1491
1492 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1493 goto out_noroc;
1494
1495 /* ROC in two conditions:
1496 * 1. there are ROC lsegs
1497 * 2. we don't send layoutreturn
1498 */
1499 /* lo ref dropped in pnfs_roc_release() */
1500 layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1501 /* If the creds don't match, we can't compound the layoutreturn */
1502 if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1503 goto out_noroc;
1504
1505 roc = layoutreturn;
1506 pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1507 res->lrs_present = 0;
1508 layoutreturn = false;
1509 put_cred(lc_cred);
1510
1511out_noroc:
1512 spin_unlock(&ino->i_lock);
1513 rcu_read_unlock();
1514 pnfs_layoutcommit_inode(ino, true);
1515 if (roc) {
1516 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1517 if (ld->prepare_layoutreturn)
1518 ld->prepare_layoutreturn(args);
1519 pnfs_put_layout_hdr(lo);
1520 return true;
1521 }
1522 if (layoutreturn)
1523 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1524 pnfs_put_layout_hdr(lo);
1525 return false;
1526}
1527
1528int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1529 struct nfs4_layoutreturn_res **respp, int *ret)
1530{
1531 struct nfs4_layoutreturn_args *arg = *argpp;
1532 int retval = -EAGAIN;
1533
1534 if (!arg)
1535 return 0;
1536 /* Handle Layoutreturn errors */
1537 switch (*ret) {
1538 case 0:
1539 retval = 0;
1540 break;
1541 case -NFS4ERR_NOMATCHING_LAYOUT:
1542 /* Was there an RPC level error? If not, retry */
1543 if (task->tk_rpc_status == 0)
1544 break;
1545 /* If the call was not sent, let caller handle it */
1546 if (!RPC_WAS_SENT(task))
1547 return 0;
1548 /*
1549 * Otherwise, assume the call succeeded and
1550 * that we need to release the layout
1551 */
1552 *ret = 0;
1553 (*respp)->lrs_present = 0;
1554 retval = 0;
1555 break;
1556 case -NFS4ERR_DELAY:
1557 /* Let the caller handle the retry */
1558 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1559 return 0;
1560 case -NFS4ERR_OLD_STATEID:
1561 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1562 &arg->range, arg->inode))
1563 break;
1564 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1565 return -EAGAIN;
1566 }
1567 *argpp = NULL;
1568 *respp = NULL;
1569 return retval;
1570}
1571
1572void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1573 struct nfs4_layoutreturn_res *res,
1574 int ret)
1575{
1576 struct pnfs_layout_hdr *lo = args->layout;
1577 struct inode *inode = args->inode;
1578 const nfs4_stateid *res_stateid = NULL;
1579 struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1580
1581 switch (ret) {
1582 case -NFS4ERR_NOMATCHING_LAYOUT:
1583 spin_lock(&inode->i_lock);
1584 if (pnfs_layout_is_valid(lo) &&
1585 nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1586 pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1587 pnfs_clear_layoutreturn_waitbit(lo);
1588 spin_unlock(&inode->i_lock);
1589 break;
1590 case 0:
1591 if (res->lrs_present)
1592 res_stateid = &res->stateid;
1593 fallthrough;
1594 default:
1595 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1596 res_stateid);
1597 }
1598 trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1599 if (ld_private && ld_private->ops && ld_private->ops->free)
1600 ld_private->ops->free(ld_private);
1601 pnfs_put_layout_hdr(lo);
1602}
1603
1604bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1605{
1606 struct nfs_inode *nfsi = NFS_I(ino);
1607 struct pnfs_layout_hdr *lo;
1608 bool sleep = false;
1609
1610 /* we might not have grabbed lo reference. so need to check under
1611 * i_lock */
1612 spin_lock(&ino->i_lock);
1613 lo = nfsi->layout;
1614 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1615 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1616 sleep = true;
1617 }
1618 spin_unlock(&ino->i_lock);
1619 return sleep;
1620}
1621
1622/*
1623 * Compare two layout segments for sorting into layout cache.
1624 * We want to preferentially return RW over RO layouts, so ensure those
1625 * are seen first.
1626 */
1627static s64
1628pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1629 const struct pnfs_layout_range *l2)
1630{
1631 s64 d;
1632
1633 /* high offset > low offset */
1634 d = l1->offset - l2->offset;
1635 if (d)
1636 return d;
1637
1638 /* short length > long length */
1639 d = l2->length - l1->length;
1640 if (d)
1641 return d;
1642
1643 /* read > read/write */
1644 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1645}
1646
1647static bool
1648pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1649 const struct pnfs_layout_range *l2)
1650{
1651 return pnfs_lseg_range_cmp(l1, l2) > 0;
1652}
1653
1654static bool
1655pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1656 struct pnfs_layout_segment *old)
1657{
1658 return false;
1659}
1660
1661void
1662pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1663 struct pnfs_layout_segment *lseg,
1664 bool (*is_after)(const struct pnfs_layout_range *,
1665 const struct pnfs_layout_range *),
1666 bool (*do_merge)(struct pnfs_layout_segment *,
1667 struct pnfs_layout_segment *),
1668 struct list_head *free_me)
1669{
1670 struct pnfs_layout_segment *lp, *tmp;
1671
1672 dprintk("%s:Begin\n", __func__);
1673
1674 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1675 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1676 continue;
1677 if (do_merge(lseg, lp)) {
1678 mark_lseg_invalid(lp, free_me);
1679 continue;
1680 }
1681 if (is_after(&lseg->pls_range, &lp->pls_range))
1682 continue;
1683 list_add_tail(&lseg->pls_list, &lp->pls_list);
1684 dprintk("%s: inserted lseg %p "
1685 "iomode %d offset %llu length %llu before "
1686 "lp %p iomode %d offset %llu length %llu\n",
1687 __func__, lseg, lseg->pls_range.iomode,
1688 lseg->pls_range.offset, lseg->pls_range.length,
1689 lp, lp->pls_range.iomode, lp->pls_range.offset,
1690 lp->pls_range.length);
1691 goto out;
1692 }
1693 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1694 dprintk("%s: inserted lseg %p "
1695 "iomode %d offset %llu length %llu at tail\n",
1696 __func__, lseg, lseg->pls_range.iomode,
1697 lseg->pls_range.offset, lseg->pls_range.length);
1698out:
1699 pnfs_get_layout_hdr(lo);
1700
1701 dprintk("%s:Return\n", __func__);
1702}
1703EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1704
1705static void
1706pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1707 struct pnfs_layout_segment *lseg,
1708 struct list_head *free_me)
1709{
1710 struct inode *inode = lo->plh_inode;
1711 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1712
1713 if (ld->add_lseg != NULL)
1714 ld->add_lseg(lo, lseg, free_me);
1715 else
1716 pnfs_generic_layout_insert_lseg(lo, lseg,
1717 pnfs_lseg_range_is_after,
1718 pnfs_lseg_no_merge,
1719 free_me);
1720}
1721
1722static struct pnfs_layout_hdr *
1723alloc_init_layout_hdr(struct inode *ino,
1724 struct nfs_open_context *ctx,
1725 gfp_t gfp_flags)
1726{
1727 struct pnfs_layout_hdr *lo;
1728
1729 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1730 if (!lo)
1731 return NULL;
1732 refcount_set(&lo->plh_refcount, 1);
1733 INIT_LIST_HEAD(&lo->plh_layouts);
1734 INIT_LIST_HEAD(&lo->plh_segs);
1735 INIT_LIST_HEAD(&lo->plh_return_segs);
1736 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1737 lo->plh_inode = ino;
1738 lo->plh_lc_cred = get_cred(ctx->cred);
1739 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1740 return lo;
1741}
1742
1743static struct pnfs_layout_hdr *
1744pnfs_find_alloc_layout(struct inode *ino,
1745 struct nfs_open_context *ctx,
1746 gfp_t gfp_flags)
1747 __releases(&ino->i_lock)
1748 __acquires(&ino->i_lock)
1749{
1750 struct nfs_inode *nfsi = NFS_I(ino);
1751 struct pnfs_layout_hdr *new = NULL;
1752
1753 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1754
1755 if (nfsi->layout != NULL)
1756 goto out_existing;
1757 spin_unlock(&ino->i_lock);
1758 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1759 spin_lock(&ino->i_lock);
1760
1761 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1762 nfsi->layout = new;
1763 return new;
1764 } else if (new != NULL)
1765 pnfs_free_layout_hdr(new);
1766out_existing:
1767 pnfs_get_layout_hdr(nfsi->layout);
1768 return nfsi->layout;
1769}
1770
1771/*
1772 * iomode matching rules:
1773 * iomode lseg strict match
1774 * iomode
1775 * ----- ----- ------ -----
1776 * ANY READ N/A true
1777 * ANY RW N/A true
1778 * RW READ N/A false
1779 * RW RW N/A true
1780 * READ READ N/A true
1781 * READ RW true false
1782 * READ RW false true
1783 */
1784static bool
1785pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1786 const struct pnfs_layout_range *range,
1787 bool strict_iomode)
1788{
1789 struct pnfs_layout_range range1;
1790
1791 if ((range->iomode == IOMODE_RW &&
1792 ls_range->iomode != IOMODE_RW) ||
1793 (range->iomode != ls_range->iomode &&
1794 strict_iomode) ||
1795 !pnfs_lseg_range_intersecting(ls_range, range))
1796 return false;
1797
1798 /* range1 covers only the first byte in the range */
1799 range1 = *range;
1800 range1.length = 1;
1801 return pnfs_lseg_range_contained(ls_range, &range1);
1802}
1803
1804/*
1805 * lookup range in layout
1806 */
1807static struct pnfs_layout_segment *
1808pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1809 struct pnfs_layout_range *range,
1810 bool strict_iomode)
1811{
1812 struct pnfs_layout_segment *lseg, *ret = NULL;
1813
1814 dprintk("%s:Begin\n", __func__);
1815
1816 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1817 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1818 pnfs_lseg_range_match(&lseg->pls_range, range,
1819 strict_iomode)) {
1820 ret = pnfs_get_lseg(lseg);
1821 break;
1822 }
1823 }
1824
1825 dprintk("%s:Return lseg %p ref %d\n",
1826 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1827 return ret;
1828}
1829
1830/*
1831 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1832 * to the MDS or over pNFS
1833 *
1834 * The nfs_inode read_io and write_io fields are cumulative counters reset
1835 * when there are no layout segments. Note that in pnfs_update_layout iomode
1836 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1837 * WRITE request.
1838 *
1839 * A return of true means use MDS I/O.
1840 *
1841 * From rfc 5661:
1842 * If a file's size is smaller than the file size threshold, data accesses
1843 * SHOULD be sent to the metadata server. If an I/O request has a length that
1844 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1845 * server. If both file size and I/O size are provided, the client SHOULD
1846 * reach or exceed both thresholds before sending its read or write
1847 * requests to the data server.
1848 */
1849static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1850 struct inode *ino, int iomode)
1851{
1852 struct nfs4_threshold *t = ctx->mdsthreshold;
1853 struct nfs_inode *nfsi = NFS_I(ino);
1854 loff_t fsize = i_size_read(ino);
1855 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1856
1857 if (t == NULL)
1858 return ret;
1859
1860 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1861 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1862
1863 switch (iomode) {
1864 case IOMODE_READ:
1865 if (t->bm & THRESHOLD_RD) {
1866 dprintk("%s fsize %llu\n", __func__, fsize);
1867 size_set = true;
1868 if (fsize < t->rd_sz)
1869 size = true;
1870 }
1871 if (t->bm & THRESHOLD_RD_IO) {
1872 dprintk("%s nfsi->read_io %llu\n", __func__,
1873 nfsi->read_io);
1874 io_set = true;
1875 if (nfsi->read_io < t->rd_io_sz)
1876 io = true;
1877 }
1878 break;
1879 case IOMODE_RW:
1880 if (t->bm & THRESHOLD_WR) {
1881 dprintk("%s fsize %llu\n", __func__, fsize);
1882 size_set = true;
1883 if (fsize < t->wr_sz)
1884 size = true;
1885 }
1886 if (t->bm & THRESHOLD_WR_IO) {
1887 dprintk("%s nfsi->write_io %llu\n", __func__,
1888 nfsi->write_io);
1889 io_set = true;
1890 if (nfsi->write_io < t->wr_io_sz)
1891 io = true;
1892 }
1893 break;
1894 }
1895 if (size_set && io_set) {
1896 if (size && io)
1897 ret = true;
1898 } else if (size || io)
1899 ret = true;
1900
1901 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1902 return ret;
1903}
1904
1905static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1906{
1907 /*
1908 * send layoutcommit as it can hold up layoutreturn due to lseg
1909 * reference
1910 */
1911 pnfs_layoutcommit_inode(lo->plh_inode, false);
1912 return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1913 nfs_wait_bit_killable,
1914 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
1915}
1916
1917static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1918{
1919 atomic_inc(&lo->plh_outstanding);
1920}
1921
1922static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1923{
1924 if (atomic_dec_and_test(&lo->plh_outstanding) &&
1925 test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1926 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1927}
1928
1929static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1930{
1931 return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1932}
1933
1934static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1935{
1936 unsigned long *bitlock = &lo->plh_flags;
1937
1938 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1939 smp_mb__after_atomic();
1940 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1941}
1942
1943static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1944 struct nfs_server *server)
1945{
1946 if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1947 struct nfs_client *clp = server->nfs_client;
1948
1949 /* The lo must be on the clp list if there is any
1950 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1951 */
1952 spin_lock(&clp->cl_lock);
1953 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1954 spin_unlock(&clp->cl_lock);
1955 }
1956}
1957
1958/*
1959 * Layout segment is retreived from the server if not cached.
1960 * The appropriate layout segment is referenced and returned to the caller.
1961 */
1962struct pnfs_layout_segment *
1963pnfs_update_layout(struct inode *ino,
1964 struct nfs_open_context *ctx,
1965 loff_t pos,
1966 u64 count,
1967 enum pnfs_iomode iomode,
1968 bool strict_iomode,
1969 gfp_t gfp_flags)
1970{
1971 struct pnfs_layout_range arg = {
1972 .iomode = iomode,
1973 .offset = pos,
1974 .length = count,
1975 };
1976 unsigned pg_offset;
1977 struct nfs_server *server = NFS_SERVER(ino);
1978 struct nfs_client *clp = server->nfs_client;
1979 struct pnfs_layout_hdr *lo = NULL;
1980 struct pnfs_layout_segment *lseg = NULL;
1981 struct nfs4_layoutget *lgp;
1982 nfs4_stateid stateid;
1983 struct nfs4_exception exception = {
1984 .inode = ino,
1985 };
1986 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1987 bool first;
1988
1989 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1990 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1991 PNFS_UPDATE_LAYOUT_NO_PNFS);
1992 goto out;
1993 }
1994
1995 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1996 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1997 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1998 goto out;
1999 }
2000
2001lookup_again:
2002 lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2003 if (IS_ERR(lseg))
2004 goto out;
2005 first = false;
2006 spin_lock(&ino->i_lock);
2007 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2008 if (lo == NULL) {
2009 spin_unlock(&ino->i_lock);
2010 lseg = ERR_PTR(-ENOMEM);
2011 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2012 PNFS_UPDATE_LAYOUT_NOMEM);
2013 goto out;
2014 }
2015
2016 /* Do we even need to bother with this? */
2017 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2018 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2019 PNFS_UPDATE_LAYOUT_BULK_RECALL);
2020 dprintk("%s matches recall, use MDS\n", __func__);
2021 goto out_unlock;
2022 }
2023
2024 /* if LAYOUTGET already failed once we don't try again */
2025 if (pnfs_layout_io_test_failed(lo, iomode)) {
2026 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2027 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2028 goto out_unlock;
2029 }
2030
2031 /*
2032 * If the layout segment list is empty, but there are outstanding
2033 * layoutget calls, then they might be subject to a layoutrecall.
2034 */
2035 if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2036 atomic_read(&lo->plh_outstanding) != 0) {
2037 spin_unlock(&ino->i_lock);
2038 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2039 TASK_KILLABLE));
2040 if (IS_ERR(lseg))
2041 goto out_put_layout_hdr;
2042 pnfs_put_layout_hdr(lo);
2043 goto lookup_again;
2044 }
2045
2046 /*
2047 * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2048 * for LAYOUTRETURN.
2049 */
2050 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2051 spin_unlock(&ino->i_lock);
2052 dprintk("%s wait for layoutreturn\n", __func__);
2053 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2054 if (!IS_ERR(lseg)) {
2055 pnfs_put_layout_hdr(lo);
2056 dprintk("%s retrying\n", __func__);
2057 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2058 lseg,
2059 PNFS_UPDATE_LAYOUT_RETRY);
2060 goto lookup_again;
2061 }
2062 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2063 PNFS_UPDATE_LAYOUT_RETURN);
2064 goto out_put_layout_hdr;
2065 }
2066
2067 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2068 if (lseg) {
2069 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2070 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2071 goto out_unlock;
2072 }
2073
2074 /*
2075 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2076 * stateid, or it has been invalidated, then we must use the open
2077 * stateid.
2078 */
2079 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2080 int status;
2081
2082 /*
2083 * The first layoutget for the file. Need to serialize per
2084 * RFC 5661 Errata 3208.
2085 */
2086 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2087 &lo->plh_flags)) {
2088 spin_unlock(&ino->i_lock);
2089 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2090 NFS_LAYOUT_FIRST_LAYOUTGET,
2091 TASK_KILLABLE));
2092 if (IS_ERR(lseg))
2093 goto out_put_layout_hdr;
2094 pnfs_put_layout_hdr(lo);
2095 dprintk("%s retrying\n", __func__);
2096 goto lookup_again;
2097 }
2098
2099 spin_unlock(&ino->i_lock);
2100 first = true;
2101 status = nfs4_select_rw_stateid(ctx->state,
2102 iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2103 NULL, &stateid, NULL);
2104 if (status != 0) {
2105 lseg = ERR_PTR(status);
2106 trace_pnfs_update_layout(ino, pos, count,
2107 iomode, lo, lseg,
2108 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2109 nfs4_schedule_stateid_recovery(server, ctx->state);
2110 pnfs_clear_first_layoutget(lo);
2111 pnfs_put_layout_hdr(lo);
2112 goto lookup_again;
2113 }
2114 spin_lock(&ino->i_lock);
2115 } else {
2116 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2117 }
2118
2119 if (pnfs_layoutgets_blocked(lo)) {
2120 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2121 PNFS_UPDATE_LAYOUT_BLOCKED);
2122 goto out_unlock;
2123 }
2124 nfs_layoutget_begin(lo);
2125 spin_unlock(&ino->i_lock);
2126
2127 _add_to_server_list(lo, server);
2128
2129 pg_offset = arg.offset & ~PAGE_MASK;
2130 if (pg_offset) {
2131 arg.offset -= pg_offset;
2132 arg.length += pg_offset;
2133 }
2134 if (arg.length != NFS4_MAX_UINT64)
2135 arg.length = PAGE_ALIGN(arg.length);
2136
2137 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2138 if (!lgp) {
2139 lseg = ERR_PTR(-ENOMEM);
2140 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2141 PNFS_UPDATE_LAYOUT_NOMEM);
2142 nfs_layoutget_end(lo);
2143 goto out_put_layout_hdr;
2144 }
2145
2146 lgp->lo = lo;
2147 pnfs_get_layout_hdr(lo);
2148
2149 lseg = nfs4_proc_layoutget(lgp, &exception);
2150 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2151 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2152 nfs_layoutget_end(lo);
2153 if (IS_ERR(lseg)) {
2154 switch(PTR_ERR(lseg)) {
2155 case -EBUSY:
2156 if (time_after(jiffies, giveup))
2157 lseg = NULL;
2158 break;
2159 case -ERECALLCONFLICT:
2160 case -EAGAIN:
2161 break;
2162 case -ENODATA:
2163 /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2164 pnfs_layout_set_fail_bit(
2165 lo, pnfs_iomode_to_fail_bit(iomode));
2166 lseg = NULL;
2167 goto out_put_layout_hdr;
2168 default:
2169 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2170 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2171 lseg = NULL;
2172 }
2173 goto out_put_layout_hdr;
2174 }
2175 if (lseg) {
2176 if (!exception.retry)
2177 goto out_put_layout_hdr;
2178 if (first)
2179 pnfs_clear_first_layoutget(lo);
2180 trace_pnfs_update_layout(ino, pos, count,
2181 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2182 pnfs_put_layout_hdr(lo);
2183 goto lookup_again;
2184 }
2185 } else {
2186 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2187 }
2188
2189out_put_layout_hdr:
2190 if (first)
2191 pnfs_clear_first_layoutget(lo);
2192 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2193 PNFS_UPDATE_LAYOUT_EXIT);
2194 pnfs_put_layout_hdr(lo);
2195out:
2196 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2197 "(%s, offset: %llu, length: %llu)\n",
2198 __func__, ino->i_sb->s_id,
2199 (unsigned long long)NFS_FILEID(ino),
2200 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2201 iomode==IOMODE_RW ? "read/write" : "read-only",
2202 (unsigned long long)pos,
2203 (unsigned long long)count);
2204 return lseg;
2205out_unlock:
2206 spin_unlock(&ino->i_lock);
2207 goto out_put_layout_hdr;
2208}
2209EXPORT_SYMBOL_GPL(pnfs_update_layout);
2210
2211static bool
2212pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2213{
2214 switch (range->iomode) {
2215 case IOMODE_READ:
2216 case IOMODE_RW:
2217 break;
2218 default:
2219 return false;
2220 }
2221 if (range->offset == NFS4_MAX_UINT64)
2222 return false;
2223 if (range->length == 0)
2224 return false;
2225 if (range->length != NFS4_MAX_UINT64 &&
2226 range->length > NFS4_MAX_UINT64 - range->offset)
2227 return false;
2228 return true;
2229}
2230
2231static struct pnfs_layout_hdr *
2232_pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2233{
2234 struct pnfs_layout_hdr *lo;
2235
2236 spin_lock(&ino->i_lock);
2237 lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2238 if (!lo)
2239 goto out_unlock;
2240 if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2241 goto out_unlock;
2242 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2243 goto out_unlock;
2244 if (pnfs_layoutgets_blocked(lo))
2245 goto out_unlock;
2246 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2247 goto out_unlock;
2248 nfs_layoutget_begin(lo);
2249 spin_unlock(&ino->i_lock);
2250 _add_to_server_list(lo, NFS_SERVER(ino));
2251 return lo;
2252
2253out_unlock:
2254 spin_unlock(&ino->i_lock);
2255 pnfs_put_layout_hdr(lo);
2256 return NULL;
2257}
2258
2259static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2260 struct nfs_open_context *ctx)
2261{
2262 struct inode *ino = data->dentry->d_inode;
2263 struct pnfs_layout_range rng = {
2264 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2265 IOMODE_RW: IOMODE_READ,
2266 .offset = 0,
2267 .length = NFS4_MAX_UINT64,
2268 };
2269 struct nfs4_layoutget *lgp;
2270 struct pnfs_layout_hdr *lo;
2271
2272 /* Heuristic: don't send layoutget if we have cached data */
2273 if (rng.iomode == IOMODE_READ &&
2274 (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2275 return;
2276
2277 lo = _pnfs_grab_empty_layout(ino, ctx);
2278 if (!lo)
2279 return;
2280 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, ¤t_stateid, &rng,
2281 nfs_io_gfp_mask());
2282 if (!lgp) {
2283 pnfs_clear_first_layoutget(lo);
2284 nfs_layoutget_end(lo);
2285 pnfs_put_layout_hdr(lo);
2286 return;
2287 }
2288 lgp->lo = lo;
2289 data->lgp = lgp;
2290 data->o_arg.lg_args = &lgp->args;
2291 data->o_res.lg_res = &lgp->res;
2292}
2293
2294static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2295 struct nfs_open_context *ctx)
2296{
2297 struct inode *ino = data->dentry->d_inode;
2298 struct pnfs_layout_range rng = {
2299 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2300 IOMODE_RW: IOMODE_READ,
2301 .offset = 0,
2302 .length = NFS4_MAX_UINT64,
2303 };
2304 struct nfs4_layoutget *lgp;
2305
2306 lgp = pnfs_alloc_init_layoutget_args(ino, ctx, ¤t_stateid, &rng,
2307 nfs_io_gfp_mask());
2308 if (!lgp)
2309 return;
2310 data->lgp = lgp;
2311 data->o_arg.lg_args = &lgp->args;
2312 data->o_res.lg_res = &lgp->res;
2313}
2314
2315void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2316 struct nfs_open_context *ctx)
2317{
2318 struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2319
2320 if (!(pnfs_enabled_sb(server) &&
2321 server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2322 return;
2323 /* Could check on max_ops, but currently hardcoded high enough */
2324 if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2325 return;
2326 if (data->lgp)
2327 return;
2328 if (data->state)
2329 _lgopen_prepare_attached(data, ctx);
2330 else
2331 _lgopen_prepare_floating(data, ctx);
2332}
2333
2334void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2335 struct nfs_open_context *ctx)
2336{
2337 struct pnfs_layout_hdr *lo;
2338 struct pnfs_layout_segment *lseg;
2339 struct nfs_server *srv = NFS_SERVER(ino);
2340 u32 iomode;
2341
2342 if (!lgp)
2343 return;
2344 dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2345 if (lgp->res.status) {
2346 switch (lgp->res.status) {
2347 default:
2348 break;
2349 /*
2350 * Halt lgopen attempts if the server doesn't recognise
2351 * the "current stateid" value, the layout type, or the
2352 * layoutget operation as being valid.
2353 * Also if it complains about too many ops in the compound
2354 * or of the request/reply being too big.
2355 */
2356 case -NFS4ERR_BAD_STATEID:
2357 case -NFS4ERR_NOTSUPP:
2358 case -NFS4ERR_REP_TOO_BIG:
2359 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2360 case -NFS4ERR_REQ_TOO_BIG:
2361 case -NFS4ERR_TOO_MANY_OPS:
2362 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2363 srv->caps &= ~NFS_CAP_LGOPEN;
2364 }
2365 return;
2366 }
2367 if (!lgp->lo) {
2368 lo = _pnfs_grab_empty_layout(ino, ctx);
2369 if (!lo)
2370 return;
2371 lgp->lo = lo;
2372 } else
2373 lo = lgp->lo;
2374
2375 lseg = pnfs_layout_process(lgp);
2376 if (!IS_ERR(lseg)) {
2377 iomode = lgp->args.range.iomode;
2378 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2379 pnfs_put_lseg(lseg);
2380 }
2381}
2382
2383void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2384{
2385 if (lgp != NULL) {
2386 if (lgp->lo) {
2387 pnfs_clear_first_layoutget(lgp->lo);
2388 nfs_layoutget_end(lgp->lo);
2389 }
2390 pnfs_layoutget_free(lgp);
2391 }
2392}
2393
2394struct pnfs_layout_segment *
2395pnfs_layout_process(struct nfs4_layoutget *lgp)
2396{
2397 struct pnfs_layout_hdr *lo = lgp->lo;
2398 struct nfs4_layoutget_res *res = &lgp->res;
2399 struct pnfs_layout_segment *lseg;
2400 struct inode *ino = lo->plh_inode;
2401 LIST_HEAD(free_me);
2402
2403 if (!pnfs_sanity_check_layout_range(&res->range))
2404 return ERR_PTR(-EINVAL);
2405
2406 /* Inject layout blob into I/O device driver */
2407 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2408 if (IS_ERR_OR_NULL(lseg)) {
2409 if (!lseg)
2410 lseg = ERR_PTR(-ENOMEM);
2411
2412 dprintk("%s: Could not allocate layout: error %ld\n",
2413 __func__, PTR_ERR(lseg));
2414 return lseg;
2415 }
2416
2417 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2418
2419 spin_lock(&ino->i_lock);
2420 if (pnfs_layoutgets_blocked(lo)) {
2421 dprintk("%s forget reply due to state\n", __func__);
2422 goto out_forget;
2423 }
2424
2425 if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2426 !pnfs_is_first_layoutget(lo))
2427 goto out_forget;
2428
2429 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2430 /* existing state ID, make sure the sequence number matches. */
2431 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2432 if (!pnfs_layout_is_valid(lo))
2433 lo->plh_barrier = 0;
2434 dprintk("%s forget reply due to sequence\n", __func__);
2435 goto out_forget;
2436 }
2437 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2438 } else if (pnfs_layout_is_valid(lo)) {
2439 /*
2440 * We got an entirely new state ID. Mark all segments for the
2441 * inode invalid, and retry the layoutget
2442 */
2443 struct pnfs_layout_range range = {
2444 .iomode = IOMODE_ANY,
2445 .length = NFS4_MAX_UINT64,
2446 };
2447 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2448 goto out_forget;
2449 } else {
2450 /* We have a completely new layout */
2451 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2452 }
2453
2454 pnfs_get_lseg(lseg);
2455 pnfs_layout_insert_lseg(lo, lseg, &free_me);
2456
2457
2458 if (res->return_on_close)
2459 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2460
2461 spin_unlock(&ino->i_lock);
2462 pnfs_free_lseg_list(&free_me);
2463 return lseg;
2464
2465out_forget:
2466 spin_unlock(&ino->i_lock);
2467 lseg->pls_layout = lo;
2468 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2469 return ERR_PTR(-EAGAIN);
2470}
2471
2472/**
2473 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2474 * @lo: pointer to layout header
2475 * @tmp_list: list header to be used with pnfs_free_lseg_list()
2476 * @return_range: describe layout segment ranges to be returned
2477 * @seq: stateid seqid to match
2478 *
2479 * This function is mainly intended for use by layoutrecall. It attempts
2480 * to free the layout segment immediately, or else to mark it for return
2481 * as soon as its reference count drops to zero.
2482 *
2483 * Returns
2484 * - 0: a layoutreturn needs to be scheduled.
2485 * - EBUSY: there are layout segment that are still in use.
2486 * - ENOENT: there are no layout segments that need to be returned.
2487 */
2488int
2489pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2490 struct list_head *tmp_list,
2491 const struct pnfs_layout_range *return_range,
2492 u32 seq)
2493{
2494 struct pnfs_layout_segment *lseg, *next;
2495 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2496 int remaining = 0;
2497
2498 dprintk("%s:Begin lo %p\n", __func__, lo);
2499
2500 assert_spin_locked(&lo->plh_inode->i_lock);
2501
2502 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2503 tmp_list = &lo->plh_return_segs;
2504
2505 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2506 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2507 dprintk("%s: marking lseg %p iomode %d "
2508 "offset %llu length %llu\n", __func__,
2509 lseg, lseg->pls_range.iomode,
2510 lseg->pls_range.offset,
2511 lseg->pls_range.length);
2512 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2513 tmp_list = &lo->plh_return_segs;
2514 if (mark_lseg_invalid(lseg, tmp_list))
2515 continue;
2516 remaining++;
2517 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2518 pnfs_lseg_cancel_io(server, lseg);
2519 }
2520
2521 if (remaining) {
2522 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2523 return -EBUSY;
2524 }
2525
2526 if (!list_empty(&lo->plh_return_segs)) {
2527 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2528 return 0;
2529 }
2530
2531 return -ENOENT;
2532}
2533
2534static void
2535pnfs_mark_layout_for_return(struct inode *inode,
2536 const struct pnfs_layout_range *range)
2537{
2538 struct pnfs_layout_hdr *lo;
2539 bool return_now = false;
2540
2541 spin_lock(&inode->i_lock);
2542 lo = NFS_I(inode)->layout;
2543 if (!pnfs_layout_is_valid(lo)) {
2544 spin_unlock(&inode->i_lock);
2545 return;
2546 }
2547 pnfs_set_plh_return_info(lo, range->iomode, 0);
2548 /*
2549 * mark all matching lsegs so that we are sure to have no live
2550 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2551 * for how it works.
2552 */
2553 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2554 const struct cred *cred;
2555 nfs4_stateid stateid;
2556 enum pnfs_iomode iomode;
2557
2558 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2559 spin_unlock(&inode->i_lock);
2560 if (return_now)
2561 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2562 } else {
2563 spin_unlock(&inode->i_lock);
2564 nfs_commit_inode(inode, 0);
2565 }
2566}
2567
2568void pnfs_error_mark_layout_for_return(struct inode *inode,
2569 struct pnfs_layout_segment *lseg)
2570{
2571 struct pnfs_layout_range range = {
2572 .iomode = lseg->pls_range.iomode,
2573 .offset = 0,
2574 .length = NFS4_MAX_UINT64,
2575 };
2576
2577 pnfs_mark_layout_for_return(inode, &range);
2578}
2579EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2580
2581static bool
2582pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2583{
2584 return pnfs_layout_is_valid(lo) &&
2585 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2586 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2587}
2588
2589static struct pnfs_layout_segment *
2590pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2591 const struct pnfs_layout_range *range,
2592 enum pnfs_iomode iomode)
2593{
2594 struct pnfs_layout_segment *lseg;
2595
2596 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2597 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2598 continue;
2599 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2600 continue;
2601 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2602 continue;
2603 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2604 return lseg;
2605 }
2606 return NULL;
2607}
2608
2609/* Find open file states whose mode matches that of the range */
2610static bool
2611pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2612 const struct pnfs_layout_range *range)
2613{
2614 struct list_head *head;
2615 struct nfs_open_context *ctx;
2616 fmode_t mode = 0;
2617
2618 if (!pnfs_layout_can_be_returned(lo) ||
2619 !pnfs_find_first_lseg(lo, range, range->iomode))
2620 return false;
2621
2622 head = &NFS_I(lo->plh_inode)->open_files;
2623 list_for_each_entry_rcu(ctx, head, list) {
2624 if (ctx->state)
2625 mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2626 }
2627
2628 switch (range->iomode) {
2629 default:
2630 break;
2631 case IOMODE_READ:
2632 mode &= ~FMODE_WRITE;
2633 break;
2634 case IOMODE_RW:
2635 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2636 mode &= ~FMODE_READ;
2637 }
2638 return mode == 0;
2639}
2640
2641static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2642 void *data)
2643{
2644 const struct pnfs_layout_range *range = data;
2645 const struct cred *cred;
2646 struct pnfs_layout_hdr *lo;
2647 struct inode *inode;
2648 nfs4_stateid stateid;
2649 enum pnfs_iomode iomode;
2650
2651restart:
2652 rcu_read_lock();
2653 list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2654 inode = lo->plh_inode;
2655 if (!inode || !pnfs_layout_can_be_returned(lo) ||
2656 test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2657 continue;
2658 spin_lock(&inode->i_lock);
2659 if (!lo->plh_inode ||
2660 !pnfs_should_return_unused_layout(lo, range)) {
2661 spin_unlock(&inode->i_lock);
2662 continue;
2663 }
2664 pnfs_get_layout_hdr(lo);
2665 pnfs_set_plh_return_info(lo, range->iomode, 0);
2666 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2667 range, 0) != 0 ||
2668 !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2669 spin_unlock(&inode->i_lock);
2670 rcu_read_unlock();
2671 pnfs_put_layout_hdr(lo);
2672 cond_resched();
2673 goto restart;
2674 }
2675 spin_unlock(&inode->i_lock);
2676 rcu_read_unlock();
2677 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2678 pnfs_put_layout_hdr(lo);
2679 cond_resched();
2680 goto restart;
2681 }
2682 rcu_read_unlock();
2683 return 0;
2684}
2685
2686void
2687pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2688 enum pnfs_iomode iomode)
2689{
2690 struct pnfs_layout_range range = {
2691 .iomode = iomode,
2692 .offset = 0,
2693 .length = NFS4_MAX_UINT64,
2694 };
2695
2696 nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2697 &range);
2698}
2699
2700void
2701pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2702{
2703 if (pgio->pg_lseg == NULL ||
2704 test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2705 return;
2706 pnfs_put_lseg(pgio->pg_lseg);
2707 pgio->pg_lseg = NULL;
2708}
2709EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2710
2711/*
2712 * Check for any intersection between the request and the pgio->pg_lseg,
2713 * and if none, put this pgio->pg_lseg away.
2714 */
2715void
2716pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2717{
2718 if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2719 pnfs_put_lseg(pgio->pg_lseg);
2720 pgio->pg_lseg = NULL;
2721 }
2722}
2723EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2724
2725void
2726pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2727{
2728 u64 rd_size;
2729
2730 pnfs_generic_pg_check_layout(pgio);
2731 pnfs_generic_pg_check_range(pgio, req);
2732 if (pgio->pg_lseg == NULL) {
2733 if (pgio->pg_dreq == NULL)
2734 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2735 else
2736 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq,
2737 req_offset(req));
2738
2739 pgio->pg_lseg =
2740 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2741 req_offset(req), rd_size,
2742 IOMODE_READ, false,
2743 nfs_io_gfp_mask());
2744 if (IS_ERR(pgio->pg_lseg)) {
2745 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2746 pgio->pg_lseg = NULL;
2747 return;
2748 }
2749 }
2750 /* If no lseg, fall back to read through mds */
2751 if (pgio->pg_lseg == NULL)
2752 nfs_pageio_reset_read_mds(pgio);
2753
2754}
2755EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2756
2757void
2758pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2759 struct nfs_page *req, u64 wb_size)
2760{
2761 pnfs_generic_pg_check_layout(pgio);
2762 pnfs_generic_pg_check_range(pgio, req);
2763 if (pgio->pg_lseg == NULL) {
2764 pgio->pg_lseg =
2765 pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2766 req_offset(req), wb_size, IOMODE_RW,
2767 false, nfs_io_gfp_mask());
2768 if (IS_ERR(pgio->pg_lseg)) {
2769 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2770 pgio->pg_lseg = NULL;
2771 return;
2772 }
2773 }
2774 /* If no lseg, fall back to write through mds */
2775 if (pgio->pg_lseg == NULL)
2776 nfs_pageio_reset_write_mds(pgio);
2777}
2778EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2779
2780void
2781pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2782{
2783 if (desc->pg_lseg) {
2784 pnfs_put_lseg(desc->pg_lseg);
2785 desc->pg_lseg = NULL;
2786 }
2787}
2788EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2789
2790/*
2791 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2792 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2793 */
2794size_t
2795pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2796 struct nfs_page *prev, struct nfs_page *req)
2797{
2798 unsigned int size;
2799 u64 seg_end, req_start, seg_left;
2800
2801 size = nfs_generic_pg_test(pgio, prev, req);
2802 if (!size)
2803 return 0;
2804
2805 /*
2806 * 'size' contains the number of bytes left in the current page (up
2807 * to the original size asked for in @req->wb_bytes).
2808 *
2809 * Calculate how many bytes are left in the layout segment
2810 * and if there are less bytes than 'size', return that instead.
2811 *
2812 * Please also note that 'end_offset' is actually the offset of the
2813 * first byte that lies outside the pnfs_layout_range. FIXME?
2814 *
2815 */
2816 if (pgio->pg_lseg) {
2817 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2818 pgio->pg_lseg->pls_range.length);
2819 req_start = req_offset(req);
2820
2821 /* start of request is past the last byte of this segment */
2822 if (req_start >= seg_end)
2823 return 0;
2824
2825 /* adjust 'size' iff there are fewer bytes left in the
2826 * segment than what nfs_generic_pg_test returned */
2827 seg_left = seg_end - req_start;
2828 if (seg_left < size)
2829 size = (unsigned int)seg_left;
2830 }
2831
2832 return size;
2833}
2834EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2835
2836int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2837{
2838 struct nfs_pageio_descriptor pgio;
2839
2840 /* Resend all requests through the MDS */
2841 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2842 hdr->completion_ops);
2843 return nfs_pageio_resend(&pgio, hdr);
2844}
2845EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2846
2847static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2848{
2849
2850 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2851 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2852 PNFS_LAYOUTRET_ON_ERROR) {
2853 pnfs_return_layout(hdr->inode);
2854 }
2855 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2856 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2857}
2858
2859/*
2860 * Called by non rpc-based layout drivers
2861 */
2862void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2863{
2864 if (likely(!hdr->pnfs_error)) {
2865 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2866 hdr->mds_offset + hdr->res.count);
2867 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2868 }
2869 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2870 if (unlikely(hdr->pnfs_error))
2871 pnfs_ld_handle_write_error(hdr);
2872 hdr->mds_ops->rpc_release(hdr);
2873}
2874EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2875
2876static void
2877pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2878 struct nfs_pgio_header *hdr)
2879{
2880 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2881
2882 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2883 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2884 nfs_pageio_reset_write_mds(desc);
2885 mirror->pg_recoalesce = 1;
2886 }
2887 hdr->completion_ops->completion(hdr);
2888}
2889
2890static enum pnfs_try_status
2891pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2892 const struct rpc_call_ops *call_ops,
2893 struct pnfs_layout_segment *lseg,
2894 int how)
2895{
2896 struct inode *inode = hdr->inode;
2897 enum pnfs_try_status trypnfs;
2898 struct nfs_server *nfss = NFS_SERVER(inode);
2899
2900 hdr->mds_ops = call_ops;
2901
2902 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2903 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2904 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2905 if (trypnfs != PNFS_NOT_ATTEMPTED)
2906 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2907 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2908 return trypnfs;
2909}
2910
2911static void
2912pnfs_do_write(struct nfs_pageio_descriptor *desc,
2913 struct nfs_pgio_header *hdr, int how)
2914{
2915 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2916 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2917 enum pnfs_try_status trypnfs;
2918
2919 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2920 switch (trypnfs) {
2921 case PNFS_NOT_ATTEMPTED:
2922 pnfs_write_through_mds(desc, hdr);
2923 break;
2924 case PNFS_ATTEMPTED:
2925 break;
2926 case PNFS_TRY_AGAIN:
2927 /* cleanup hdr and prepare to redo pnfs */
2928 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2929 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2930 list_splice_init(&hdr->pages, &mirror->pg_list);
2931 mirror->pg_recoalesce = 1;
2932 }
2933 hdr->mds_ops->rpc_release(hdr);
2934 }
2935}
2936
2937static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2938{
2939 pnfs_put_lseg(hdr->lseg);
2940 nfs_pgio_header_free(hdr);
2941}
2942
2943int
2944pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2945{
2946 struct nfs_pgio_header *hdr;
2947 int ret;
2948
2949 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2950 if (!hdr) {
2951 desc->pg_error = -ENOMEM;
2952 return desc->pg_error;
2953 }
2954 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2955
2956 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2957 ret = nfs_generic_pgio(desc, hdr);
2958 if (!ret)
2959 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2960
2961 return ret;
2962}
2963EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2964
2965int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2966{
2967 struct nfs_pageio_descriptor pgio;
2968
2969 /* Resend all requests through the MDS */
2970 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2971 return nfs_pageio_resend(&pgio, hdr);
2972}
2973EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2974
2975static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2976{
2977 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2978 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2979 PNFS_LAYOUTRET_ON_ERROR) {
2980 pnfs_return_layout(hdr->inode);
2981 }
2982 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2983 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2984}
2985
2986/*
2987 * Called by non rpc-based layout drivers
2988 */
2989void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2990{
2991 if (likely(!hdr->pnfs_error))
2992 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2993 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2994 if (unlikely(hdr->pnfs_error))
2995 pnfs_ld_handle_read_error(hdr);
2996 hdr->mds_ops->rpc_release(hdr);
2997}
2998EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2999
3000static void
3001pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
3002 struct nfs_pgio_header *hdr)
3003{
3004 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3005
3006 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3007 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3008 nfs_pageio_reset_read_mds(desc);
3009 mirror->pg_recoalesce = 1;
3010 }
3011 hdr->completion_ops->completion(hdr);
3012}
3013
3014/*
3015 * Call the appropriate parallel I/O subsystem read function.
3016 */
3017static enum pnfs_try_status
3018pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3019 const struct rpc_call_ops *call_ops,
3020 struct pnfs_layout_segment *lseg)
3021{
3022 struct inode *inode = hdr->inode;
3023 struct nfs_server *nfss = NFS_SERVER(inode);
3024 enum pnfs_try_status trypnfs;
3025
3026 hdr->mds_ops = call_ops;
3027
3028 dprintk("%s: Reading ino:%lu %u@%llu\n",
3029 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3030
3031 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3032 if (trypnfs != PNFS_NOT_ATTEMPTED)
3033 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3034 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3035 return trypnfs;
3036}
3037
3038/* Resend all requests through pnfs. */
3039void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3040 unsigned int mirror_idx)
3041{
3042 struct nfs_pageio_descriptor pgio;
3043
3044 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3045 /* Prevent deadlocks with layoutreturn! */
3046 pnfs_put_lseg(hdr->lseg);
3047 hdr->lseg = NULL;
3048
3049 nfs_pageio_init_read(&pgio, hdr->inode, false,
3050 hdr->completion_ops);
3051 pgio.pg_mirror_idx = mirror_idx;
3052 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3053 }
3054}
3055EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3056
3057static void
3058pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3059{
3060 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3061 struct pnfs_layout_segment *lseg = desc->pg_lseg;
3062 enum pnfs_try_status trypnfs;
3063
3064 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3065 switch (trypnfs) {
3066 case PNFS_NOT_ATTEMPTED:
3067 pnfs_read_through_mds(desc, hdr);
3068 break;
3069 case PNFS_ATTEMPTED:
3070 break;
3071 case PNFS_TRY_AGAIN:
3072 /* cleanup hdr and prepare to redo pnfs */
3073 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3074 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3075 list_splice_init(&hdr->pages, &mirror->pg_list);
3076 mirror->pg_recoalesce = 1;
3077 }
3078 hdr->mds_ops->rpc_release(hdr);
3079 }
3080}
3081
3082static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3083{
3084 pnfs_put_lseg(hdr->lseg);
3085 nfs_pgio_header_free(hdr);
3086}
3087
3088int
3089pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3090{
3091 struct nfs_pgio_header *hdr;
3092 int ret;
3093
3094 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3095 if (!hdr) {
3096 desc->pg_error = -ENOMEM;
3097 return desc->pg_error;
3098 }
3099 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3100 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3101 ret = nfs_generic_pgio(desc, hdr);
3102 if (!ret)
3103 pnfs_do_read(desc, hdr);
3104 return ret;
3105}
3106EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3107
3108static void pnfs_clear_layoutcommitting(struct inode *inode)
3109{
3110 unsigned long *bitlock = &NFS_I(inode)->flags;
3111
3112 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3113 smp_mb__after_atomic();
3114 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3115}
3116
3117/*
3118 * There can be multiple RW segments.
3119 */
3120static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3121{
3122 struct pnfs_layout_segment *lseg;
3123
3124 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3125 if (lseg->pls_range.iomode == IOMODE_RW &&
3126 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3127 list_add(&lseg->pls_lc_list, listp);
3128 }
3129}
3130
3131static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3132{
3133 struct pnfs_layout_segment *lseg, *tmp;
3134
3135 /* Matched by references in pnfs_set_layoutcommit */
3136 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3137 list_del_init(&lseg->pls_lc_list);
3138 pnfs_put_lseg(lseg);
3139 }
3140
3141 pnfs_clear_layoutcommitting(inode);
3142}
3143
3144void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3145{
3146 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3147}
3148EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3149
3150void
3151pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3152 loff_t end_pos)
3153{
3154 struct nfs_inode *nfsi = NFS_I(inode);
3155 bool mark_as_dirty = false;
3156
3157 spin_lock(&inode->i_lock);
3158 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3159 nfsi->layout->plh_lwb = end_pos;
3160 mark_as_dirty = true;
3161 dprintk("%s: Set layoutcommit for inode %lu ",
3162 __func__, inode->i_ino);
3163 } else if (end_pos > nfsi->layout->plh_lwb)
3164 nfsi->layout->plh_lwb = end_pos;
3165 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3166 /* references matched in nfs4_layoutcommit_release */
3167 pnfs_get_lseg(lseg);
3168 }
3169 spin_unlock(&inode->i_lock);
3170 dprintk("%s: lseg %p end_pos %llu\n",
3171 __func__, lseg, nfsi->layout->plh_lwb);
3172
3173 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3174 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3175 if (mark_as_dirty)
3176 mark_inode_dirty_sync(inode);
3177}
3178EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3179
3180void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3181{
3182 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3183
3184 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3185 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3186 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3187}
3188
3189/*
3190 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3191 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3192 * data to disk to allow the server to recover the data if it crashes.
3193 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3194 * is off, and a COMMIT is sent to a data server, or
3195 * if WRITEs to a data server return NFS_DATA_SYNC.
3196 */
3197int
3198pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3199{
3200 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3201 struct nfs4_layoutcommit_data *data;
3202 struct nfs_inode *nfsi = NFS_I(inode);
3203 loff_t end_pos;
3204 int status;
3205
3206 if (!pnfs_layoutcommit_outstanding(inode))
3207 return 0;
3208
3209 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3210
3211 status = -EAGAIN;
3212 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3213 if (!sync)
3214 goto out;
3215 status = wait_on_bit_lock_action(&nfsi->flags,
3216 NFS_INO_LAYOUTCOMMITTING,
3217 nfs_wait_bit_killable,
3218 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3219 if (status)
3220 goto out;
3221 }
3222
3223 status = -ENOMEM;
3224 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3225 data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3226 if (!data)
3227 goto clear_layoutcommitting;
3228
3229 status = 0;
3230 spin_lock(&inode->i_lock);
3231 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3232 goto out_unlock;
3233
3234 INIT_LIST_HEAD(&data->lseg_list);
3235 pnfs_list_write_lseg(inode, &data->lseg_list);
3236
3237 end_pos = nfsi->layout->plh_lwb;
3238
3239 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3240 data->cred = get_cred(nfsi->layout->plh_lc_cred);
3241 spin_unlock(&inode->i_lock);
3242
3243 data->args.inode = inode;
3244 nfs_fattr_init(&data->fattr);
3245 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3246 data->res.fattr = &data->fattr;
3247 if (end_pos != 0)
3248 data->args.lastbytewritten = end_pos - 1;
3249 else
3250 data->args.lastbytewritten = U64_MAX;
3251 data->res.server = NFS_SERVER(inode);
3252
3253 if (ld->prepare_layoutcommit) {
3254 status = ld->prepare_layoutcommit(&data->args);
3255 if (status) {
3256 put_cred(data->cred);
3257 spin_lock(&inode->i_lock);
3258 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3259 if (end_pos > nfsi->layout->plh_lwb)
3260 nfsi->layout->plh_lwb = end_pos;
3261 goto out_unlock;
3262 }
3263 }
3264
3265
3266 status = nfs4_proc_layoutcommit(data, sync);
3267out:
3268 if (status)
3269 mark_inode_dirty_sync(inode);
3270 dprintk("<-- %s status %d\n", __func__, status);
3271 return status;
3272out_unlock:
3273 spin_unlock(&inode->i_lock);
3274 kfree(data);
3275clear_layoutcommitting:
3276 pnfs_clear_layoutcommitting(inode);
3277 goto out;
3278}
3279EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3280
3281int
3282pnfs_generic_sync(struct inode *inode, bool datasync)
3283{
3284 return pnfs_layoutcommit_inode(inode, true);
3285}
3286EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3287
3288struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3289{
3290 struct nfs4_threshold *thp;
3291
3292 thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3293 if (!thp) {
3294 dprintk("%s mdsthreshold allocation failed\n", __func__);
3295 return NULL;
3296 }
3297 return thp;
3298}
3299
3300#if IS_ENABLED(CONFIG_NFS_V4_2)
3301int
3302pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3303{
3304 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3305 struct nfs_server *server = NFS_SERVER(inode);
3306 struct nfs_inode *nfsi = NFS_I(inode);
3307 struct nfs42_layoutstat_data *data;
3308 struct pnfs_layout_hdr *hdr;
3309 int status = 0;
3310
3311 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3312 goto out;
3313
3314 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3315 goto out;
3316
3317 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3318 goto out;
3319
3320 spin_lock(&inode->i_lock);
3321 if (!NFS_I(inode)->layout) {
3322 spin_unlock(&inode->i_lock);
3323 goto out_clear_layoutstats;
3324 }
3325 hdr = NFS_I(inode)->layout;
3326 pnfs_get_layout_hdr(hdr);
3327 spin_unlock(&inode->i_lock);
3328
3329 data = kzalloc(sizeof(*data), gfp_flags);
3330 if (!data) {
3331 status = -ENOMEM;
3332 goto out_put;
3333 }
3334
3335 data->args.fh = NFS_FH(inode);
3336 data->args.inode = inode;
3337 status = ld->prepare_layoutstats(&data->args);
3338 if (status)
3339 goto out_free;
3340
3341 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3342
3343out:
3344 dprintk("%s returns %d\n", __func__, status);
3345 return status;
3346
3347out_free:
3348 kfree(data);
3349out_put:
3350 pnfs_put_layout_hdr(hdr);
3351out_clear_layoutstats:
3352 smp_mb__before_atomic();
3353 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3354 smp_mb__after_atomic();
3355 goto out;
3356}
3357EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3358#endif
3359
3360unsigned int layoutstats_timer;
3361module_param(layoutstats_timer, uint, 0644);
3362EXPORT_SYMBOL_GPL(layoutstats_timer);
1/*
2 * pNFS functions to call and manage layout drivers.
3 *
4 * Copyright (c) 2002 [year of first publication]
5 * The Regents of the University of Michigan
6 * All Rights Reserved
7 *
8 * Dean Hildebrand <dhildebz@umich.edu>
9 *
10 * Permission is granted to use, copy, create derivative works, and
11 * redistribute this software and such derivative works for any purpose,
12 * so long as the name of the University of Michigan is not used in
13 * any advertising or publicity pertaining to the use or distribution
14 * of this software without specific, written prior authorization. If
15 * the above copyright notice or any other identification of the
16 * University of Michigan is included in any copy of any portion of
17 * this software, then the disclaimer below must also be included.
18 *
19 * This software is provided as is, without representation or warranty
20 * of any kind either express or implied, including without limitation
21 * the implied warranties of merchantability, fitness for a particular
22 * purpose, or noninfringement. The Regents of the University of
23 * Michigan shall not be liable for any damages, including special,
24 * indirect, incidental, or consequential damages, with respect to any
25 * claim arising out of or in connection with the use of the software,
26 * even if it has been or is hereafter advised of the possibility of
27 * such damages.
28 */
29
30#include <linux/nfs_fs.h>
31#include <linux/nfs_page.h>
32#include <linux/module.h>
33#include <linux/sort.h>
34#include "internal.h"
35#include "pnfs.h"
36#include "iostat.h"
37#include "nfs4trace.h"
38#include "delegation.h"
39#include "nfs42.h"
40
41#define NFSDBG_FACILITY NFSDBG_PNFS
42#define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
43
44/* Locking:
45 *
46 * pnfs_spinlock:
47 * protects pnfs_modules_tbl.
48 */
49static DEFINE_SPINLOCK(pnfs_spinlock);
50
51/*
52 * pnfs_modules_tbl holds all pnfs modules
53 */
54static LIST_HEAD(pnfs_modules_tbl);
55
56static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
57static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
58 struct list_head *free_me,
59 const struct pnfs_layout_range *range,
60 u32 seq);
61static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
62 struct list_head *tmp_list);
63
64/* Return the registered pnfs layout driver module matching given id */
65static struct pnfs_layoutdriver_type *
66find_pnfs_driver_locked(u32 id)
67{
68 struct pnfs_layoutdriver_type *local;
69
70 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
71 if (local->id == id)
72 goto out;
73 local = NULL;
74out:
75 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
76 return local;
77}
78
79static struct pnfs_layoutdriver_type *
80find_pnfs_driver(u32 id)
81{
82 struct pnfs_layoutdriver_type *local;
83
84 spin_lock(&pnfs_spinlock);
85 local = find_pnfs_driver_locked(id);
86 if (local != NULL && !try_module_get(local->owner)) {
87 dprintk("%s: Could not grab reference on module\n", __func__);
88 local = NULL;
89 }
90 spin_unlock(&pnfs_spinlock);
91 return local;
92}
93
94void
95unset_pnfs_layoutdriver(struct nfs_server *nfss)
96{
97 if (nfss->pnfs_curr_ld) {
98 if (nfss->pnfs_curr_ld->clear_layoutdriver)
99 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
100 /* Decrement the MDS count. Purge the deviceid cache if zero */
101 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
102 nfs4_deviceid_purge_client(nfss->nfs_client);
103 module_put(nfss->pnfs_curr_ld->owner);
104 }
105 nfss->pnfs_curr_ld = NULL;
106}
107
108/*
109 * When the server sends a list of layout types, we choose one in the order
110 * given in the list below.
111 *
112 * FIXME: should this list be configurable in some fashion? module param?
113 * mount option? something else?
114 */
115static const u32 ld_prefs[] = {
116 LAYOUT_SCSI,
117 LAYOUT_BLOCK_VOLUME,
118 LAYOUT_OSD2_OBJECTS,
119 LAYOUT_FLEX_FILES,
120 LAYOUT_NFSV4_1_FILES,
121 0
122};
123
124static int
125ld_cmp(const void *e1, const void *e2)
126{
127 u32 ld1 = *((u32 *)e1);
128 u32 ld2 = *((u32 *)e2);
129 int i;
130
131 for (i = 0; ld_prefs[i] != 0; i++) {
132 if (ld1 == ld_prefs[i])
133 return -1;
134
135 if (ld2 == ld_prefs[i])
136 return 1;
137 }
138 return 0;
139}
140
141/*
142 * Try to set the server's pnfs module to the pnfs layout type specified by id.
143 * Currently only one pNFS layout driver per filesystem is supported.
144 *
145 * @ids array of layout types supported by MDS.
146 */
147void
148set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
149 struct nfs_fsinfo *fsinfo)
150{
151 struct pnfs_layoutdriver_type *ld_type = NULL;
152 u32 id;
153 int i;
154
155 if (fsinfo->nlayouttypes == 0)
156 goto out_no_driver;
157 if (!(server->nfs_client->cl_exchange_flags &
158 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
159 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
160 __func__, server->nfs_client->cl_exchange_flags);
161 goto out_no_driver;
162 }
163
164 sort(fsinfo->layouttype, fsinfo->nlayouttypes,
165 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
166
167 for (i = 0; i < fsinfo->nlayouttypes; i++) {
168 id = fsinfo->layouttype[i];
169 ld_type = find_pnfs_driver(id);
170 if (!ld_type) {
171 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
172 id);
173 ld_type = find_pnfs_driver(id);
174 }
175 if (ld_type)
176 break;
177 }
178
179 if (!ld_type) {
180 dprintk("%s: No pNFS module found!\n", __func__);
181 goto out_no_driver;
182 }
183
184 server->pnfs_curr_ld = ld_type;
185 if (ld_type->set_layoutdriver
186 && ld_type->set_layoutdriver(server, mntfh)) {
187 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
188 "driver %u.\n", __func__, id);
189 module_put(ld_type->owner);
190 goto out_no_driver;
191 }
192 /* Bump the MDS count */
193 atomic_inc(&server->nfs_client->cl_mds_count);
194
195 dprintk("%s: pNFS module for %u set\n", __func__, id);
196 return;
197
198out_no_driver:
199 dprintk("%s: Using NFSv4 I/O\n", __func__);
200 server->pnfs_curr_ld = NULL;
201}
202
203int
204pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
205{
206 int status = -EINVAL;
207 struct pnfs_layoutdriver_type *tmp;
208
209 if (ld_type->id == 0) {
210 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
211 return status;
212 }
213 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
214 printk(KERN_ERR "NFS: %s Layout driver must provide "
215 "alloc_lseg and free_lseg.\n", __func__);
216 return status;
217 }
218
219 spin_lock(&pnfs_spinlock);
220 tmp = find_pnfs_driver_locked(ld_type->id);
221 if (!tmp) {
222 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
223 status = 0;
224 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
225 ld_type->name);
226 } else {
227 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
228 __func__, ld_type->id);
229 }
230 spin_unlock(&pnfs_spinlock);
231
232 return status;
233}
234EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
235
236void
237pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
238{
239 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
240 spin_lock(&pnfs_spinlock);
241 list_del(&ld_type->pnfs_tblid);
242 spin_unlock(&pnfs_spinlock);
243}
244EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
245
246/*
247 * pNFS client layout cache
248 */
249
250/* Need to hold i_lock if caller does not already hold reference */
251void
252pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
253{
254 atomic_inc(&lo->plh_refcount);
255}
256
257static struct pnfs_layout_hdr *
258pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
259{
260 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
261 return ld->alloc_layout_hdr(ino, gfp_flags);
262}
263
264static void
265pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
266{
267 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
268 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
269
270 if (!list_empty(&lo->plh_layouts)) {
271 struct nfs_client *clp = server->nfs_client;
272
273 spin_lock(&clp->cl_lock);
274 list_del_init(&lo->plh_layouts);
275 spin_unlock(&clp->cl_lock);
276 }
277 put_rpccred(lo->plh_lc_cred);
278 return ld->free_layout_hdr(lo);
279}
280
281static void
282pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
283{
284 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
285 dprintk("%s: freeing layout cache %p\n", __func__, lo);
286 nfsi->layout = NULL;
287 /* Reset MDS Threshold I/O counters */
288 nfsi->write_io = 0;
289 nfsi->read_io = 0;
290}
291
292void
293pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
294{
295 struct inode *inode = lo->plh_inode;
296
297 pnfs_layoutreturn_before_put_layout_hdr(lo);
298
299 if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
300 if (!list_empty(&lo->plh_segs))
301 WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
302 pnfs_detach_layout_hdr(lo);
303 spin_unlock(&inode->i_lock);
304 pnfs_free_layout_hdr(lo);
305 }
306}
307
308static void
309pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
310 u32 seq)
311{
312 if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
313 iomode = IOMODE_ANY;
314 lo->plh_return_iomode = iomode;
315 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
316 if (seq != 0) {
317 WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
318 lo->plh_return_seq = seq;
319 }
320}
321
322static void
323pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
324{
325 lo->plh_return_iomode = 0;
326 lo->plh_return_seq = 0;
327 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
328}
329
330static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
331{
332 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
333 clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
334 smp_mb__after_atomic();
335 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
336 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
337}
338
339static void
340pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
341 struct list_head *free_me)
342{
343 clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
344 clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
345 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
346 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
347 if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
348 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
349}
350
351/*
352 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
353 *
354 * In order to continue using the pnfs_layout_hdr, a full recovery
355 * is required.
356 * Note that caller must hold inode->i_lock.
357 */
358int
359pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
360 struct list_head *lseg_list)
361{
362 struct pnfs_layout_range range = {
363 .iomode = IOMODE_ANY,
364 .offset = 0,
365 .length = NFS4_MAX_UINT64,
366 };
367 struct pnfs_layout_segment *lseg, *next;
368
369 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
370 pnfs_clear_layoutreturn_info(lo);
371 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
372 pnfs_clear_lseg_state(lseg, lseg_list);
373 pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
374 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
375 !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
376 pnfs_clear_layoutreturn_waitbit(lo);
377 return !list_empty(&lo->plh_segs);
378}
379
380static int
381pnfs_iomode_to_fail_bit(u32 iomode)
382{
383 return iomode == IOMODE_RW ?
384 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
385}
386
387static void
388pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
389{
390 lo->plh_retry_timestamp = jiffies;
391 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
392 atomic_inc(&lo->plh_refcount);
393}
394
395static void
396pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
397{
398 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
399 atomic_dec(&lo->plh_refcount);
400}
401
402static void
403pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
404{
405 struct inode *inode = lo->plh_inode;
406 struct pnfs_layout_range range = {
407 .iomode = iomode,
408 .offset = 0,
409 .length = NFS4_MAX_UINT64,
410 };
411 LIST_HEAD(head);
412
413 spin_lock(&inode->i_lock);
414 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
415 pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
416 spin_unlock(&inode->i_lock);
417 pnfs_free_lseg_list(&head);
418 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
419 iomode == IOMODE_RW ? "RW" : "READ");
420}
421
422static bool
423pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
424{
425 unsigned long start, end;
426 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
427
428 if (test_bit(fail_bit, &lo->plh_flags) == 0)
429 return false;
430 end = jiffies;
431 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
432 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
433 /* It is time to retry the failed layoutgets */
434 pnfs_layout_clear_fail_bit(lo, fail_bit);
435 return false;
436 }
437 return true;
438}
439
440static void
441pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
442 const struct pnfs_layout_range *range,
443 const nfs4_stateid *stateid)
444{
445 INIT_LIST_HEAD(&lseg->pls_list);
446 INIT_LIST_HEAD(&lseg->pls_lc_list);
447 atomic_set(&lseg->pls_refcount, 1);
448 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
449 lseg->pls_layout = lo;
450 lseg->pls_range = *range;
451 lseg->pls_seq = be32_to_cpu(stateid->seqid);
452}
453
454static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
455{
456 if (lseg != NULL) {
457 struct inode *inode = lseg->pls_layout->plh_inode;
458 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
459 }
460}
461
462static void
463pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
464 struct pnfs_layout_segment *lseg)
465{
466 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
467 list_del_init(&lseg->pls_list);
468 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
469 atomic_dec(&lo->plh_refcount);
470 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
471 return;
472 if (list_empty(&lo->plh_segs) &&
473 !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
474 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
475 if (atomic_read(&lo->plh_outstanding) == 0)
476 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
477 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
478 }
479}
480
481static bool
482pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
483 struct pnfs_layout_segment *lseg)
484{
485 if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
486 pnfs_layout_is_valid(lo)) {
487 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
488 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
489 return true;
490 }
491 return false;
492}
493
494void
495pnfs_put_lseg(struct pnfs_layout_segment *lseg)
496{
497 struct pnfs_layout_hdr *lo;
498 struct inode *inode;
499
500 if (!lseg)
501 return;
502
503 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
504 atomic_read(&lseg->pls_refcount),
505 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
506
507 lo = lseg->pls_layout;
508 inode = lo->plh_inode;
509
510 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
511 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
512 spin_unlock(&inode->i_lock);
513 return;
514 }
515 pnfs_get_layout_hdr(lo);
516 pnfs_layout_remove_lseg(lo, lseg);
517 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
518 lseg = NULL;
519 spin_unlock(&inode->i_lock);
520 pnfs_free_lseg(lseg);
521 pnfs_put_layout_hdr(lo);
522 }
523}
524EXPORT_SYMBOL_GPL(pnfs_put_lseg);
525
526static void pnfs_free_lseg_async_work(struct work_struct *work)
527{
528 struct pnfs_layout_segment *lseg;
529 struct pnfs_layout_hdr *lo;
530
531 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
532 lo = lseg->pls_layout;
533
534 pnfs_free_lseg(lseg);
535 pnfs_put_layout_hdr(lo);
536}
537
538static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
539{
540 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
541 schedule_work(&lseg->pls_work);
542}
543
544void
545pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
546{
547 if (!lseg)
548 return;
549
550 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
551
552 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
553 atomic_read(&lseg->pls_refcount),
554 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
555 if (atomic_dec_and_test(&lseg->pls_refcount)) {
556 struct pnfs_layout_hdr *lo = lseg->pls_layout;
557 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
558 return;
559 pnfs_layout_remove_lseg(lo, lseg);
560 if (!pnfs_cache_lseg_for_layoutreturn(lo, lseg)) {
561 pnfs_get_layout_hdr(lo);
562 pnfs_free_lseg_async(lseg);
563 }
564 }
565}
566EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
567
568/*
569 * is l2 fully contained in l1?
570 * start1 end1
571 * [----------------------------------)
572 * start2 end2
573 * [----------------)
574 */
575static bool
576pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
577 const struct pnfs_layout_range *l2)
578{
579 u64 start1 = l1->offset;
580 u64 end1 = pnfs_end_offset(start1, l1->length);
581 u64 start2 = l2->offset;
582 u64 end2 = pnfs_end_offset(start2, l2->length);
583
584 return (start1 <= start2) && (end1 >= end2);
585}
586
587static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
588 struct list_head *tmp_list)
589{
590 if (!atomic_dec_and_test(&lseg->pls_refcount))
591 return false;
592 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
593 list_add(&lseg->pls_list, tmp_list);
594 return true;
595}
596
597/* Returns 1 if lseg is removed from list, 0 otherwise */
598static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
599 struct list_head *tmp_list)
600{
601 int rv = 0;
602
603 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
604 /* Remove the reference keeping the lseg in the
605 * list. It will now be removed when all
606 * outstanding io is finished.
607 */
608 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
609 atomic_read(&lseg->pls_refcount));
610 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
611 rv = 1;
612 }
613 return rv;
614}
615
616/*
617 * Compare 2 layout stateid sequence ids, to see which is newer,
618 * taking into account wraparound issues.
619 */
620static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
621{
622 return (s32)(s1 - s2) > 0;
623}
624
625static bool
626pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
627 const struct pnfs_layout_range *recall_range)
628{
629 return (recall_range->iomode == IOMODE_ANY ||
630 lseg_range->iomode == recall_range->iomode) &&
631 pnfs_lseg_range_intersecting(lseg_range, recall_range);
632}
633
634static bool
635pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
636 const struct pnfs_layout_range *recall_range,
637 u32 seq)
638{
639 if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
640 return false;
641 if (recall_range == NULL)
642 return true;
643 return pnfs_should_free_range(&lseg->pls_range, recall_range);
644}
645
646/**
647 * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
648 * @lo: layout header containing the lsegs
649 * @tmp_list: list head where doomed lsegs should go
650 * @recall_range: optional recall range argument to match (may be NULL)
651 * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
652 *
653 * Walk the list of lsegs in the layout header, and tear down any that should
654 * be destroyed. If "recall_range" is specified then the segment must match
655 * that range. If "seq" is non-zero, then only match segments that were handed
656 * out at or before that sequence.
657 *
658 * Returns number of matching invalid lsegs remaining in list after scanning
659 * it and purging them.
660 */
661int
662pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
663 struct list_head *tmp_list,
664 const struct pnfs_layout_range *recall_range,
665 u32 seq)
666{
667 struct pnfs_layout_segment *lseg, *next;
668 int remaining = 0;
669
670 dprintk("%s:Begin lo %p\n", __func__, lo);
671
672 if (list_empty(&lo->plh_segs))
673 return 0;
674 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
675 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
676 dprintk("%s: freeing lseg %p iomode %d seq %u"
677 "offset %llu length %llu\n", __func__,
678 lseg, lseg->pls_range.iomode, lseg->pls_seq,
679 lseg->pls_range.offset, lseg->pls_range.length);
680 if (!mark_lseg_invalid(lseg, tmp_list))
681 remaining++;
682 }
683 dprintk("%s:Return %i\n", __func__, remaining);
684 return remaining;
685}
686
687static void
688pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
689 struct list_head *free_me,
690 const struct pnfs_layout_range *range,
691 u32 seq)
692{
693 struct pnfs_layout_segment *lseg, *next;
694
695 list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
696 if (pnfs_match_lseg_recall(lseg, range, seq))
697 list_move_tail(&lseg->pls_list, free_me);
698 }
699}
700
701/* note free_me must contain lsegs from a single layout_hdr */
702void
703pnfs_free_lseg_list(struct list_head *free_me)
704{
705 struct pnfs_layout_segment *lseg, *tmp;
706
707 if (list_empty(free_me))
708 return;
709
710 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
711 list_del(&lseg->pls_list);
712 pnfs_free_lseg(lseg);
713 }
714}
715
716void
717pnfs_destroy_layout(struct nfs_inode *nfsi)
718{
719 struct pnfs_layout_hdr *lo;
720 LIST_HEAD(tmp_list);
721
722 spin_lock(&nfsi->vfs_inode.i_lock);
723 lo = nfsi->layout;
724 if (lo) {
725 pnfs_get_layout_hdr(lo);
726 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
727 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
728 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
729 spin_unlock(&nfsi->vfs_inode.i_lock);
730 pnfs_free_lseg_list(&tmp_list);
731 pnfs_put_layout_hdr(lo);
732 } else
733 spin_unlock(&nfsi->vfs_inode.i_lock);
734}
735EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
736
737static bool
738pnfs_layout_add_bulk_destroy_list(struct inode *inode,
739 struct list_head *layout_list)
740{
741 struct pnfs_layout_hdr *lo;
742 bool ret = false;
743
744 spin_lock(&inode->i_lock);
745 lo = NFS_I(inode)->layout;
746 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
747 pnfs_get_layout_hdr(lo);
748 list_add(&lo->plh_bulk_destroy, layout_list);
749 ret = true;
750 }
751 spin_unlock(&inode->i_lock);
752 return ret;
753}
754
755/* Caller must hold rcu_read_lock and clp->cl_lock */
756static int
757pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
758 struct nfs_server *server,
759 struct list_head *layout_list)
760{
761 struct pnfs_layout_hdr *lo, *next;
762 struct inode *inode;
763
764 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
765 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
766 continue;
767 inode = igrab(lo->plh_inode);
768 if (inode == NULL)
769 continue;
770 list_del_init(&lo->plh_layouts);
771 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
772 continue;
773 rcu_read_unlock();
774 spin_unlock(&clp->cl_lock);
775 iput(inode);
776 spin_lock(&clp->cl_lock);
777 rcu_read_lock();
778 return -EAGAIN;
779 }
780 return 0;
781}
782
783static int
784pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
785 bool is_bulk_recall)
786{
787 struct pnfs_layout_hdr *lo;
788 struct inode *inode;
789 LIST_HEAD(lseg_list);
790 int ret = 0;
791
792 while (!list_empty(layout_list)) {
793 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
794 plh_bulk_destroy);
795 dprintk("%s freeing layout for inode %lu\n", __func__,
796 lo->plh_inode->i_ino);
797 inode = lo->plh_inode;
798
799 pnfs_layoutcommit_inode(inode, false);
800
801 spin_lock(&inode->i_lock);
802 list_del_init(&lo->plh_bulk_destroy);
803 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
804 if (is_bulk_recall)
805 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
806 ret = -EAGAIN;
807 }
808 spin_unlock(&inode->i_lock);
809 pnfs_free_lseg_list(&lseg_list);
810 /* Free all lsegs that are attached to commit buckets */
811 nfs_commit_inode(inode, 0);
812 pnfs_put_layout_hdr(lo);
813 iput(inode);
814 }
815 return ret;
816}
817
818int
819pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
820 struct nfs_fsid *fsid,
821 bool is_recall)
822{
823 struct nfs_server *server;
824 LIST_HEAD(layout_list);
825
826 spin_lock(&clp->cl_lock);
827 rcu_read_lock();
828restart:
829 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
830 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
831 continue;
832 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
833 server,
834 &layout_list) != 0)
835 goto restart;
836 }
837 rcu_read_unlock();
838 spin_unlock(&clp->cl_lock);
839
840 if (list_empty(&layout_list))
841 return 0;
842 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
843}
844
845int
846pnfs_destroy_layouts_byclid(struct nfs_client *clp,
847 bool is_recall)
848{
849 struct nfs_server *server;
850 LIST_HEAD(layout_list);
851
852 spin_lock(&clp->cl_lock);
853 rcu_read_lock();
854restart:
855 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
856 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
857 server,
858 &layout_list) != 0)
859 goto restart;
860 }
861 rcu_read_unlock();
862 spin_unlock(&clp->cl_lock);
863
864 if (list_empty(&layout_list))
865 return 0;
866 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
867}
868
869/*
870 * Called by the state manger to remove all layouts established under an
871 * expired lease.
872 */
873void
874pnfs_destroy_all_layouts(struct nfs_client *clp)
875{
876 nfs4_deviceid_mark_client_invalid(clp);
877 nfs4_deviceid_purge_client(clp);
878
879 pnfs_destroy_layouts_byclid(clp, false);
880}
881
882/* update lo->plh_stateid with new if is more recent */
883void
884pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
885 bool update_barrier)
886{
887 u32 oldseq, newseq, new_barrier = 0;
888
889 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
890 newseq = be32_to_cpu(new->seqid);
891
892 if (!pnfs_layout_is_valid(lo)) {
893 nfs4_stateid_copy(&lo->plh_stateid, new);
894 lo->plh_barrier = newseq;
895 pnfs_clear_layoutreturn_info(lo);
896 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
897 return;
898 }
899 if (pnfs_seqid_is_newer(newseq, oldseq)) {
900 nfs4_stateid_copy(&lo->plh_stateid, new);
901 /*
902 * Because of wraparound, we want to keep the barrier
903 * "close" to the current seqids.
904 */
905 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
906 }
907 if (update_barrier)
908 new_barrier = be32_to_cpu(new->seqid);
909 else if (new_barrier == 0)
910 return;
911 if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
912 lo->plh_barrier = new_barrier;
913}
914
915static bool
916pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
917 const nfs4_stateid *stateid)
918{
919 u32 seqid = be32_to_cpu(stateid->seqid);
920
921 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
922}
923
924/* lget is set to 1 if called from inside send_layoutget call chain */
925static bool
926pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
927{
928 return lo->plh_block_lgets ||
929 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
930}
931
932/*
933 * Get layout from server.
934 * for now, assume that whole file layouts are requested.
935 * arg->offset: 0
936 * arg->length: all ones
937 */
938static struct pnfs_layout_segment *
939send_layoutget(struct pnfs_layout_hdr *lo,
940 struct nfs_open_context *ctx,
941 nfs4_stateid *stateid,
942 const struct pnfs_layout_range *range,
943 long *timeout, gfp_t gfp_flags)
944{
945 struct inode *ino = lo->plh_inode;
946 struct nfs_server *server = NFS_SERVER(ino);
947 struct nfs4_layoutget *lgp;
948 loff_t i_size;
949
950 dprintk("--> %s\n", __func__);
951
952 /*
953 * Synchronously retrieve layout information from server and
954 * store in lseg. If we race with a concurrent seqid morphing
955 * op, then re-send the LAYOUTGET.
956 */
957 lgp = kzalloc(sizeof(*lgp), gfp_flags);
958 if (lgp == NULL)
959 return ERR_PTR(-ENOMEM);
960
961 i_size = i_size_read(ino);
962
963 lgp->args.minlength = PAGE_SIZE;
964 if (lgp->args.minlength > range->length)
965 lgp->args.minlength = range->length;
966 if (range->iomode == IOMODE_READ) {
967 if (range->offset >= i_size)
968 lgp->args.minlength = 0;
969 else if (i_size - range->offset < lgp->args.minlength)
970 lgp->args.minlength = i_size - range->offset;
971 }
972 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
973 pnfs_copy_range(&lgp->args.range, range);
974 lgp->args.type = server->pnfs_curr_ld->id;
975 lgp->args.inode = ino;
976 lgp->args.ctx = get_nfs_open_context(ctx);
977 nfs4_stateid_copy(&lgp->args.stateid, stateid);
978 lgp->gfp_flags = gfp_flags;
979 lgp->cred = lo->plh_lc_cred;
980
981 return nfs4_proc_layoutget(lgp, timeout, gfp_flags);
982}
983
984static void pnfs_clear_layoutcommit(struct inode *inode,
985 struct list_head *head)
986{
987 struct nfs_inode *nfsi = NFS_I(inode);
988 struct pnfs_layout_segment *lseg, *tmp;
989
990 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
991 return;
992 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
993 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
994 continue;
995 pnfs_lseg_dec_and_remove_zero(lseg, head);
996 }
997}
998
999void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1000 const nfs4_stateid *arg_stateid,
1001 const struct pnfs_layout_range *range,
1002 const nfs4_stateid *stateid)
1003{
1004 struct inode *inode = lo->plh_inode;
1005 LIST_HEAD(freeme);
1006
1007 spin_lock(&inode->i_lock);
1008 if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1009 !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1010 goto out_unlock;
1011 if (stateid) {
1012 u32 seq = be32_to_cpu(arg_stateid->seqid);
1013
1014 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1015 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1016 pnfs_set_layout_stateid(lo, stateid, true);
1017 } else
1018 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1019out_unlock:
1020 pnfs_clear_layoutreturn_waitbit(lo);
1021 spin_unlock(&inode->i_lock);
1022 pnfs_free_lseg_list(&freeme);
1023
1024}
1025
1026static bool
1027pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1028 nfs4_stateid *stateid,
1029 enum pnfs_iomode *iomode)
1030{
1031 /* Serialise LAYOUTGET/LAYOUTRETURN */
1032 if (atomic_read(&lo->plh_outstanding) != 0)
1033 return false;
1034 if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1035 return false;
1036 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1037 pnfs_get_layout_hdr(lo);
1038 if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1039 if (stateid != NULL) {
1040 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1041 if (lo->plh_return_seq != 0)
1042 stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1043 }
1044 if (iomode != NULL)
1045 *iomode = lo->plh_return_iomode;
1046 pnfs_clear_layoutreturn_info(lo);
1047 return true;
1048 }
1049 if (stateid != NULL)
1050 nfs4_stateid_copy(stateid, &lo->plh_stateid);
1051 if (iomode != NULL)
1052 *iomode = IOMODE_ANY;
1053 return true;
1054}
1055
1056static void
1057pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1058 struct pnfs_layout_hdr *lo,
1059 const nfs4_stateid *stateid,
1060 enum pnfs_iomode iomode)
1061{
1062 struct inode *inode = lo->plh_inode;
1063
1064 args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1065 args->inode = inode;
1066 args->range.iomode = iomode;
1067 args->range.offset = 0;
1068 args->range.length = NFS4_MAX_UINT64;
1069 args->layout = lo;
1070 nfs4_stateid_copy(&args->stateid, stateid);
1071}
1072
1073static int
1074pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
1075 enum pnfs_iomode iomode, bool sync)
1076{
1077 struct inode *ino = lo->plh_inode;
1078 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1079 struct nfs4_layoutreturn *lrp;
1080 int status = 0;
1081
1082 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1083 if (unlikely(lrp == NULL)) {
1084 status = -ENOMEM;
1085 spin_lock(&ino->i_lock);
1086 pnfs_clear_layoutreturn_waitbit(lo);
1087 spin_unlock(&ino->i_lock);
1088 pnfs_put_layout_hdr(lo);
1089 goto out;
1090 }
1091
1092 pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1093 lrp->args.ld_private = &lrp->ld_private;
1094 lrp->clp = NFS_SERVER(ino)->nfs_client;
1095 lrp->cred = lo->plh_lc_cred;
1096 if (ld->prepare_layoutreturn)
1097 ld->prepare_layoutreturn(&lrp->args);
1098
1099 status = nfs4_proc_layoutreturn(lrp, sync);
1100out:
1101 dprintk("<-- %s status: %d\n", __func__, status);
1102 return status;
1103}
1104
1105/* Return true if layoutreturn is needed */
1106static bool
1107pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1108{
1109 struct pnfs_layout_segment *s;
1110
1111 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1112 return false;
1113
1114 /* Defer layoutreturn until all lsegs are done */
1115 list_for_each_entry(s, &lo->plh_segs, pls_list) {
1116 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1117 return false;
1118 }
1119
1120 return true;
1121}
1122
1123static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1124{
1125 struct inode *inode= lo->plh_inode;
1126
1127 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1128 return;
1129 spin_lock(&inode->i_lock);
1130 if (pnfs_layout_need_return(lo)) {
1131 nfs4_stateid stateid;
1132 enum pnfs_iomode iomode;
1133 bool send;
1134
1135 send = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1136 spin_unlock(&inode->i_lock);
1137 if (send) {
1138 /* Send an async layoutreturn so we dont deadlock */
1139 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1140 }
1141 } else
1142 spin_unlock(&inode->i_lock);
1143}
1144
1145/*
1146 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1147 * when the layout segment list is empty.
1148 *
1149 * Note that a pnfs_layout_hdr can exist with an empty layout segment
1150 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1151 * deviceid is marked invalid.
1152 */
1153int
1154_pnfs_return_layout(struct inode *ino)
1155{
1156 struct pnfs_layout_hdr *lo = NULL;
1157 struct nfs_inode *nfsi = NFS_I(ino);
1158 LIST_HEAD(tmp_list);
1159 nfs4_stateid stateid;
1160 int status = 0;
1161 bool send;
1162
1163 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1164
1165 spin_lock(&ino->i_lock);
1166 lo = nfsi->layout;
1167 if (!lo) {
1168 spin_unlock(&ino->i_lock);
1169 dprintk("NFS: %s no layout to return\n", __func__);
1170 goto out;
1171 }
1172 /* Reference matched in nfs4_layoutreturn_release */
1173 pnfs_get_layout_hdr(lo);
1174 /* Is there an outstanding layoutreturn ? */
1175 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1176 spin_unlock(&ino->i_lock);
1177 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1178 TASK_UNINTERRUPTIBLE))
1179 goto out_put_layout_hdr;
1180 spin_lock(&ino->i_lock);
1181 }
1182 pnfs_clear_layoutcommit(ino, &tmp_list);
1183 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1184
1185 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1186 struct pnfs_layout_range range = {
1187 .iomode = IOMODE_ANY,
1188 .offset = 0,
1189 .length = NFS4_MAX_UINT64,
1190 };
1191 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1192 }
1193
1194 /* Don't send a LAYOUTRETURN if list was initially empty */
1195 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1196 spin_unlock(&ino->i_lock);
1197 dprintk("NFS: %s no layout segments to return\n", __func__);
1198 goto out_put_layout_hdr;
1199 }
1200
1201 send = pnfs_prepare_layoutreturn(lo, &stateid, NULL);
1202 spin_unlock(&ino->i_lock);
1203 if (send)
1204 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1205out_put_layout_hdr:
1206 pnfs_free_lseg_list(&tmp_list);
1207 pnfs_put_layout_hdr(lo);
1208out:
1209 dprintk("<-- %s status: %d\n", __func__, status);
1210 return status;
1211}
1212EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1213
1214int
1215pnfs_commit_and_return_layout(struct inode *inode)
1216{
1217 struct pnfs_layout_hdr *lo;
1218 int ret;
1219
1220 spin_lock(&inode->i_lock);
1221 lo = NFS_I(inode)->layout;
1222 if (lo == NULL) {
1223 spin_unlock(&inode->i_lock);
1224 return 0;
1225 }
1226 pnfs_get_layout_hdr(lo);
1227 /* Block new layoutgets and read/write to ds */
1228 lo->plh_block_lgets++;
1229 spin_unlock(&inode->i_lock);
1230 filemap_fdatawait(inode->i_mapping);
1231 ret = pnfs_layoutcommit_inode(inode, true);
1232 if (ret == 0)
1233 ret = _pnfs_return_layout(inode);
1234 spin_lock(&inode->i_lock);
1235 lo->plh_block_lgets--;
1236 spin_unlock(&inode->i_lock);
1237 pnfs_put_layout_hdr(lo);
1238 return ret;
1239}
1240
1241bool pnfs_roc(struct inode *ino,
1242 struct nfs4_layoutreturn_args *args,
1243 struct nfs4_layoutreturn_res *res,
1244 const struct rpc_cred *cred)
1245{
1246 struct nfs_inode *nfsi = NFS_I(ino);
1247 struct nfs_open_context *ctx;
1248 struct nfs4_state *state;
1249 struct pnfs_layout_hdr *lo;
1250 struct pnfs_layout_segment *lseg, *next;
1251 nfs4_stateid stateid;
1252 enum pnfs_iomode iomode = 0;
1253 bool layoutreturn = false, roc = false;
1254 bool skip_read = false;
1255
1256 if (!nfs_have_layout(ino))
1257 return false;
1258retry:
1259 spin_lock(&ino->i_lock);
1260 lo = nfsi->layout;
1261 if (!lo || !pnfs_layout_is_valid(lo) ||
1262 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1263 goto out_noroc;
1264 if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1265 pnfs_get_layout_hdr(lo);
1266 spin_unlock(&ino->i_lock);
1267 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1268 TASK_UNINTERRUPTIBLE);
1269 pnfs_put_layout_hdr(lo);
1270 goto retry;
1271 }
1272
1273 /* no roc if we hold a delegation */
1274 if (nfs4_check_delegation(ino, FMODE_READ)) {
1275 if (nfs4_check_delegation(ino, FMODE_WRITE))
1276 goto out_noroc;
1277 skip_read = true;
1278 }
1279
1280 list_for_each_entry(ctx, &nfsi->open_files, list) {
1281 state = ctx->state;
1282 if (state == NULL)
1283 continue;
1284 /* Don't return layout if there is open file state */
1285 if (state->state & FMODE_WRITE)
1286 goto out_noroc;
1287 if (state->state & FMODE_READ)
1288 skip_read = true;
1289 }
1290
1291
1292 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1293 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1294 continue;
1295 /* If we are sending layoutreturn, invalidate all valid lsegs */
1296 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1297 continue;
1298 /*
1299 * Note: mark lseg for return so pnfs_layout_remove_lseg
1300 * doesn't invalidate the layout for us.
1301 */
1302 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1303 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1304 continue;
1305 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1306 }
1307
1308 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1309 goto out_noroc;
1310
1311 /* ROC in two conditions:
1312 * 1. there are ROC lsegs
1313 * 2. we don't send layoutreturn
1314 */
1315 /* lo ref dropped in pnfs_roc_release() */
1316 layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
1317 /* If the creds don't match, we can't compound the layoutreturn */
1318 if (!layoutreturn || cred != lo->plh_lc_cred)
1319 goto out_noroc;
1320
1321 roc = layoutreturn;
1322 pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1323 res->lrs_present = 0;
1324 layoutreturn = false;
1325
1326out_noroc:
1327 spin_unlock(&ino->i_lock);
1328 pnfs_layoutcommit_inode(ino, true);
1329 if (roc) {
1330 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1331 if (ld->prepare_layoutreturn)
1332 ld->prepare_layoutreturn(args);
1333 return true;
1334 }
1335 if (layoutreturn)
1336 pnfs_send_layoutreturn(lo, &stateid, iomode, true);
1337 return false;
1338}
1339
1340void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1341 struct nfs4_layoutreturn_res *res,
1342 int ret)
1343{
1344 struct pnfs_layout_hdr *lo = args->layout;
1345 const nfs4_stateid *arg_stateid = NULL;
1346 const nfs4_stateid *res_stateid = NULL;
1347 struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1348
1349 if (ret == 0) {
1350 arg_stateid = &args->stateid;
1351 if (res->lrs_present)
1352 res_stateid = &res->stateid;
1353 }
1354 pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1355 res_stateid);
1356 if (ld_private && ld_private->ops && ld_private->ops->free)
1357 ld_private->ops->free(ld_private);
1358 pnfs_put_layout_hdr(lo);
1359 trace_nfs4_layoutreturn_on_close(args->inode, 0);
1360}
1361
1362bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1363{
1364 struct nfs_inode *nfsi = NFS_I(ino);
1365 struct pnfs_layout_hdr *lo;
1366 bool sleep = false;
1367
1368 /* we might not have grabbed lo reference. so need to check under
1369 * i_lock */
1370 spin_lock(&ino->i_lock);
1371 lo = nfsi->layout;
1372 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1373 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1374 sleep = true;
1375 }
1376 spin_unlock(&ino->i_lock);
1377 return sleep;
1378}
1379
1380/*
1381 * Compare two layout segments for sorting into layout cache.
1382 * We want to preferentially return RW over RO layouts, so ensure those
1383 * are seen first.
1384 */
1385static s64
1386pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1387 const struct pnfs_layout_range *l2)
1388{
1389 s64 d;
1390
1391 /* high offset > low offset */
1392 d = l1->offset - l2->offset;
1393 if (d)
1394 return d;
1395
1396 /* short length > long length */
1397 d = l2->length - l1->length;
1398 if (d)
1399 return d;
1400
1401 /* read > read/write */
1402 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1403}
1404
1405static bool
1406pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1407 const struct pnfs_layout_range *l2)
1408{
1409 return pnfs_lseg_range_cmp(l1, l2) > 0;
1410}
1411
1412static bool
1413pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1414 struct pnfs_layout_segment *old)
1415{
1416 return false;
1417}
1418
1419void
1420pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1421 struct pnfs_layout_segment *lseg,
1422 bool (*is_after)(const struct pnfs_layout_range *,
1423 const struct pnfs_layout_range *),
1424 bool (*do_merge)(struct pnfs_layout_segment *,
1425 struct pnfs_layout_segment *),
1426 struct list_head *free_me)
1427{
1428 struct pnfs_layout_segment *lp, *tmp;
1429
1430 dprintk("%s:Begin\n", __func__);
1431
1432 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1433 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1434 continue;
1435 if (do_merge(lseg, lp)) {
1436 mark_lseg_invalid(lp, free_me);
1437 continue;
1438 }
1439 if (is_after(&lseg->pls_range, &lp->pls_range))
1440 continue;
1441 list_add_tail(&lseg->pls_list, &lp->pls_list);
1442 dprintk("%s: inserted lseg %p "
1443 "iomode %d offset %llu length %llu before "
1444 "lp %p iomode %d offset %llu length %llu\n",
1445 __func__, lseg, lseg->pls_range.iomode,
1446 lseg->pls_range.offset, lseg->pls_range.length,
1447 lp, lp->pls_range.iomode, lp->pls_range.offset,
1448 lp->pls_range.length);
1449 goto out;
1450 }
1451 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1452 dprintk("%s: inserted lseg %p "
1453 "iomode %d offset %llu length %llu at tail\n",
1454 __func__, lseg, lseg->pls_range.iomode,
1455 lseg->pls_range.offset, lseg->pls_range.length);
1456out:
1457 pnfs_get_layout_hdr(lo);
1458
1459 dprintk("%s:Return\n", __func__);
1460}
1461EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1462
1463static void
1464pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1465 struct pnfs_layout_segment *lseg,
1466 struct list_head *free_me)
1467{
1468 struct inode *inode = lo->plh_inode;
1469 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1470
1471 if (ld->add_lseg != NULL)
1472 ld->add_lseg(lo, lseg, free_me);
1473 else
1474 pnfs_generic_layout_insert_lseg(lo, lseg,
1475 pnfs_lseg_range_is_after,
1476 pnfs_lseg_no_merge,
1477 free_me);
1478}
1479
1480static struct pnfs_layout_hdr *
1481alloc_init_layout_hdr(struct inode *ino,
1482 struct nfs_open_context *ctx,
1483 gfp_t gfp_flags)
1484{
1485 struct pnfs_layout_hdr *lo;
1486
1487 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1488 if (!lo)
1489 return NULL;
1490 atomic_set(&lo->plh_refcount, 1);
1491 INIT_LIST_HEAD(&lo->plh_layouts);
1492 INIT_LIST_HEAD(&lo->plh_segs);
1493 INIT_LIST_HEAD(&lo->plh_return_segs);
1494 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1495 lo->plh_inode = ino;
1496 lo->plh_lc_cred = get_rpccred(ctx->cred);
1497 lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1498 return lo;
1499}
1500
1501static struct pnfs_layout_hdr *
1502pnfs_find_alloc_layout(struct inode *ino,
1503 struct nfs_open_context *ctx,
1504 gfp_t gfp_flags)
1505 __releases(&ino->i_lock)
1506 __acquires(&ino->i_lock)
1507{
1508 struct nfs_inode *nfsi = NFS_I(ino);
1509 struct pnfs_layout_hdr *new = NULL;
1510
1511 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1512
1513 if (nfsi->layout != NULL)
1514 goto out_existing;
1515 spin_unlock(&ino->i_lock);
1516 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1517 spin_lock(&ino->i_lock);
1518
1519 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1520 nfsi->layout = new;
1521 return new;
1522 } else if (new != NULL)
1523 pnfs_free_layout_hdr(new);
1524out_existing:
1525 pnfs_get_layout_hdr(nfsi->layout);
1526 return nfsi->layout;
1527}
1528
1529/*
1530 * iomode matching rules:
1531 * iomode lseg strict match
1532 * iomode
1533 * ----- ----- ------ -----
1534 * ANY READ N/A true
1535 * ANY RW N/A true
1536 * RW READ N/A false
1537 * RW RW N/A true
1538 * READ READ N/A true
1539 * READ RW true false
1540 * READ RW false true
1541 */
1542static bool
1543pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1544 const struct pnfs_layout_range *range,
1545 bool strict_iomode)
1546{
1547 struct pnfs_layout_range range1;
1548
1549 if ((range->iomode == IOMODE_RW &&
1550 ls_range->iomode != IOMODE_RW) ||
1551 (range->iomode != ls_range->iomode &&
1552 strict_iomode == true) ||
1553 !pnfs_lseg_range_intersecting(ls_range, range))
1554 return 0;
1555
1556 /* range1 covers only the first byte in the range */
1557 range1 = *range;
1558 range1.length = 1;
1559 return pnfs_lseg_range_contained(ls_range, &range1);
1560}
1561
1562/*
1563 * lookup range in layout
1564 */
1565static struct pnfs_layout_segment *
1566pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1567 struct pnfs_layout_range *range,
1568 bool strict_iomode)
1569{
1570 struct pnfs_layout_segment *lseg, *ret = NULL;
1571
1572 dprintk("%s:Begin\n", __func__);
1573
1574 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1575 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1576 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1577 pnfs_lseg_range_match(&lseg->pls_range, range,
1578 strict_iomode)) {
1579 ret = pnfs_get_lseg(lseg);
1580 break;
1581 }
1582 }
1583
1584 dprintk("%s:Return lseg %p ref %d\n",
1585 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1586 return ret;
1587}
1588
1589/*
1590 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1591 * to the MDS or over pNFS
1592 *
1593 * The nfs_inode read_io and write_io fields are cumulative counters reset
1594 * when there are no layout segments. Note that in pnfs_update_layout iomode
1595 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1596 * WRITE request.
1597 *
1598 * A return of true means use MDS I/O.
1599 *
1600 * From rfc 5661:
1601 * If a file's size is smaller than the file size threshold, data accesses
1602 * SHOULD be sent to the metadata server. If an I/O request has a length that
1603 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1604 * server. If both file size and I/O size are provided, the client SHOULD
1605 * reach or exceed both thresholds before sending its read or write
1606 * requests to the data server.
1607 */
1608static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1609 struct inode *ino, int iomode)
1610{
1611 struct nfs4_threshold *t = ctx->mdsthreshold;
1612 struct nfs_inode *nfsi = NFS_I(ino);
1613 loff_t fsize = i_size_read(ino);
1614 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1615
1616 if (t == NULL)
1617 return ret;
1618
1619 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1620 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1621
1622 switch (iomode) {
1623 case IOMODE_READ:
1624 if (t->bm & THRESHOLD_RD) {
1625 dprintk("%s fsize %llu\n", __func__, fsize);
1626 size_set = true;
1627 if (fsize < t->rd_sz)
1628 size = true;
1629 }
1630 if (t->bm & THRESHOLD_RD_IO) {
1631 dprintk("%s nfsi->read_io %llu\n", __func__,
1632 nfsi->read_io);
1633 io_set = true;
1634 if (nfsi->read_io < t->rd_io_sz)
1635 io = true;
1636 }
1637 break;
1638 case IOMODE_RW:
1639 if (t->bm & THRESHOLD_WR) {
1640 dprintk("%s fsize %llu\n", __func__, fsize);
1641 size_set = true;
1642 if (fsize < t->wr_sz)
1643 size = true;
1644 }
1645 if (t->bm & THRESHOLD_WR_IO) {
1646 dprintk("%s nfsi->write_io %llu\n", __func__,
1647 nfsi->write_io);
1648 io_set = true;
1649 if (nfsi->write_io < t->wr_io_sz)
1650 io = true;
1651 }
1652 break;
1653 }
1654 if (size_set && io_set) {
1655 if (size && io)
1656 ret = true;
1657 } else if (size || io)
1658 ret = true;
1659
1660 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1661 return ret;
1662}
1663
1664static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1665{
1666 /*
1667 * send layoutcommit as it can hold up layoutreturn due to lseg
1668 * reference
1669 */
1670 pnfs_layoutcommit_inode(lo->plh_inode, false);
1671 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1672 nfs_wait_bit_killable,
1673 TASK_UNINTERRUPTIBLE);
1674}
1675
1676static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1677{
1678 unsigned long *bitlock = &lo->plh_flags;
1679
1680 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1681 smp_mb__after_atomic();
1682 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1683}
1684
1685/*
1686 * Layout segment is retreived from the server if not cached.
1687 * The appropriate layout segment is referenced and returned to the caller.
1688 */
1689struct pnfs_layout_segment *
1690pnfs_update_layout(struct inode *ino,
1691 struct nfs_open_context *ctx,
1692 loff_t pos,
1693 u64 count,
1694 enum pnfs_iomode iomode,
1695 bool strict_iomode,
1696 gfp_t gfp_flags)
1697{
1698 struct pnfs_layout_range arg = {
1699 .iomode = iomode,
1700 .offset = pos,
1701 .length = count,
1702 };
1703 unsigned pg_offset, seq;
1704 struct nfs_server *server = NFS_SERVER(ino);
1705 struct nfs_client *clp = server->nfs_client;
1706 struct pnfs_layout_hdr *lo = NULL;
1707 struct pnfs_layout_segment *lseg = NULL;
1708 nfs4_stateid stateid;
1709 long timeout = 0;
1710 unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1711 bool first;
1712
1713 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1714 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1715 PNFS_UPDATE_LAYOUT_NO_PNFS);
1716 goto out;
1717 }
1718
1719 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1720 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1721 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1722 goto out;
1723 }
1724
1725 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1726 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1727 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1728 goto out;
1729 }
1730
1731lookup_again:
1732 nfs4_client_recover_expired_lease(clp);
1733 first = false;
1734 spin_lock(&ino->i_lock);
1735 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1736 if (lo == NULL) {
1737 spin_unlock(&ino->i_lock);
1738 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1739 PNFS_UPDATE_LAYOUT_NOMEM);
1740 goto out;
1741 }
1742
1743 /* Do we even need to bother with this? */
1744 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1745 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1746 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1747 dprintk("%s matches recall, use MDS\n", __func__);
1748 goto out_unlock;
1749 }
1750
1751 /* if LAYOUTGET already failed once we don't try again */
1752 if (pnfs_layout_io_test_failed(lo, iomode)) {
1753 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1754 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1755 goto out_unlock;
1756 }
1757
1758 lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1759 if (lseg) {
1760 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1761 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1762 goto out_unlock;
1763 }
1764
1765 if (!nfs4_valid_open_stateid(ctx->state)) {
1766 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1767 PNFS_UPDATE_LAYOUT_INVALID_OPEN);
1768 goto out_unlock;
1769 }
1770
1771 /*
1772 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1773 * stateid, or it has been invalidated, then we must use the open
1774 * stateid.
1775 */
1776 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
1777
1778 /*
1779 * The first layoutget for the file. Need to serialize per
1780 * RFC 5661 Errata 3208.
1781 */
1782 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1783 &lo->plh_flags)) {
1784 spin_unlock(&ino->i_lock);
1785 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1786 TASK_UNINTERRUPTIBLE);
1787 pnfs_put_layout_hdr(lo);
1788 dprintk("%s retrying\n", __func__);
1789 goto lookup_again;
1790 }
1791
1792 first = true;
1793 do {
1794 seq = read_seqbegin(&ctx->state->seqlock);
1795 nfs4_stateid_copy(&stateid, &ctx->state->stateid);
1796 } while (read_seqretry(&ctx->state->seqlock, seq));
1797 } else {
1798 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1799 }
1800
1801 /*
1802 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1803 * for LAYOUTRETURN even if first is true.
1804 */
1805 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1806 spin_unlock(&ino->i_lock);
1807 dprintk("%s wait for layoutreturn\n", __func__);
1808 if (pnfs_prepare_to_retry_layoutget(lo)) {
1809 if (first)
1810 pnfs_clear_first_layoutget(lo);
1811 pnfs_put_layout_hdr(lo);
1812 dprintk("%s retrying\n", __func__);
1813 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1814 lseg, PNFS_UPDATE_LAYOUT_RETRY);
1815 goto lookup_again;
1816 }
1817 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1818 PNFS_UPDATE_LAYOUT_RETURN);
1819 goto out_put_layout_hdr;
1820 }
1821
1822 if (pnfs_layoutgets_blocked(lo)) {
1823 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1824 PNFS_UPDATE_LAYOUT_BLOCKED);
1825 goto out_unlock;
1826 }
1827 atomic_inc(&lo->plh_outstanding);
1828 spin_unlock(&ino->i_lock);
1829
1830 if (list_empty(&lo->plh_layouts)) {
1831 /* The lo must be on the clp list if there is any
1832 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1833 */
1834 spin_lock(&clp->cl_lock);
1835 if (list_empty(&lo->plh_layouts))
1836 list_add_tail(&lo->plh_layouts, &server->layouts);
1837 spin_unlock(&clp->cl_lock);
1838 }
1839
1840 pg_offset = arg.offset & ~PAGE_MASK;
1841 if (pg_offset) {
1842 arg.offset -= pg_offset;
1843 arg.length += pg_offset;
1844 }
1845 if (arg.length != NFS4_MAX_UINT64)
1846 arg.length = PAGE_ALIGN(arg.length);
1847
1848 lseg = send_layoutget(lo, ctx, &stateid, &arg, &timeout, gfp_flags);
1849 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1850 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1851 atomic_dec(&lo->plh_outstanding);
1852 if (IS_ERR(lseg)) {
1853 switch(PTR_ERR(lseg)) {
1854 case -EBUSY:
1855 if (time_after(jiffies, giveup))
1856 lseg = NULL;
1857 break;
1858 case -ERECALLCONFLICT:
1859 /* Huh? We hold no layouts, how is there a recall? */
1860 if (first) {
1861 lseg = NULL;
1862 break;
1863 }
1864 /* Destroy the existing layout and start over */
1865 if (time_after(jiffies, giveup))
1866 pnfs_destroy_layout(NFS_I(ino));
1867 /* Fallthrough */
1868 case -EAGAIN:
1869 break;
1870 default:
1871 if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
1872 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1873 lseg = NULL;
1874 }
1875 goto out_put_layout_hdr;
1876 }
1877 if (lseg) {
1878 if (first)
1879 pnfs_clear_first_layoutget(lo);
1880 trace_pnfs_update_layout(ino, pos, count,
1881 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
1882 pnfs_put_layout_hdr(lo);
1883 goto lookup_again;
1884 }
1885 } else {
1886 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
1887 }
1888
1889out_put_layout_hdr:
1890 if (first)
1891 pnfs_clear_first_layoutget(lo);
1892 pnfs_put_layout_hdr(lo);
1893out:
1894 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1895 "(%s, offset: %llu, length: %llu)\n",
1896 __func__, ino->i_sb->s_id,
1897 (unsigned long long)NFS_FILEID(ino),
1898 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1899 iomode==IOMODE_RW ? "read/write" : "read-only",
1900 (unsigned long long)pos,
1901 (unsigned long long)count);
1902 return lseg;
1903out_unlock:
1904 spin_unlock(&ino->i_lock);
1905 goto out_put_layout_hdr;
1906}
1907EXPORT_SYMBOL_GPL(pnfs_update_layout);
1908
1909static bool
1910pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1911{
1912 switch (range->iomode) {
1913 case IOMODE_READ:
1914 case IOMODE_RW:
1915 break;
1916 default:
1917 return false;
1918 }
1919 if (range->offset == NFS4_MAX_UINT64)
1920 return false;
1921 if (range->length == 0)
1922 return false;
1923 if (range->length != NFS4_MAX_UINT64 &&
1924 range->length > NFS4_MAX_UINT64 - range->offset)
1925 return false;
1926 return true;
1927}
1928
1929struct pnfs_layout_segment *
1930pnfs_layout_process(struct nfs4_layoutget *lgp)
1931{
1932 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1933 struct nfs4_layoutget_res *res = &lgp->res;
1934 struct pnfs_layout_segment *lseg;
1935 struct inode *ino = lo->plh_inode;
1936 LIST_HEAD(free_me);
1937
1938 if (!pnfs_sanity_check_layout_range(&res->range))
1939 return ERR_PTR(-EINVAL);
1940
1941 /* Inject layout blob into I/O device driver */
1942 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1943 if (IS_ERR_OR_NULL(lseg)) {
1944 if (!lseg)
1945 lseg = ERR_PTR(-ENOMEM);
1946
1947 dprintk("%s: Could not allocate layout: error %ld\n",
1948 __func__, PTR_ERR(lseg));
1949 return lseg;
1950 }
1951
1952 pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
1953
1954 spin_lock(&ino->i_lock);
1955 if (pnfs_layoutgets_blocked(lo)) {
1956 dprintk("%s forget reply due to state\n", __func__);
1957 goto out_forget;
1958 }
1959
1960 if (!pnfs_layout_is_valid(lo)) {
1961 /* We have a completely new layout */
1962 pnfs_set_layout_stateid(lo, &res->stateid, true);
1963 } else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1964 /* existing state ID, make sure the sequence number matches. */
1965 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1966 dprintk("%s forget reply due to sequence\n", __func__);
1967 goto out_forget;
1968 }
1969 pnfs_set_layout_stateid(lo, &res->stateid, false);
1970 } else {
1971 /*
1972 * We got an entirely new state ID. Mark all segments for the
1973 * inode invalid, and retry the layoutget
1974 */
1975 pnfs_mark_layout_stateid_invalid(lo, &free_me);
1976 goto out_forget;
1977 }
1978
1979 pnfs_get_lseg(lseg);
1980 pnfs_layout_insert_lseg(lo, lseg, &free_me);
1981
1982
1983 if (res->return_on_close)
1984 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1985
1986 spin_unlock(&ino->i_lock);
1987 pnfs_free_lseg_list(&free_me);
1988 return lseg;
1989
1990out_forget:
1991 spin_unlock(&ino->i_lock);
1992 lseg->pls_layout = lo;
1993 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1994 return ERR_PTR(-EAGAIN);
1995}
1996
1997/**
1998 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1999 * @lo: pointer to layout header
2000 * @tmp_list: list header to be used with pnfs_free_lseg_list()
2001 * @return_range: describe layout segment ranges to be returned
2002 *
2003 * This function is mainly intended for use by layoutrecall. It attempts
2004 * to free the layout segment immediately, or else to mark it for return
2005 * as soon as its reference count drops to zero.
2006 */
2007int
2008pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2009 struct list_head *tmp_list,
2010 const struct pnfs_layout_range *return_range,
2011 u32 seq)
2012{
2013 struct pnfs_layout_segment *lseg, *next;
2014 int remaining = 0;
2015
2016 dprintk("%s:Begin lo %p\n", __func__, lo);
2017
2018 if (list_empty(&lo->plh_segs))
2019 return 0;
2020
2021 assert_spin_locked(&lo->plh_inode->i_lock);
2022
2023 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2024 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2025 dprintk("%s: marking lseg %p iomode %d "
2026 "offset %llu length %llu\n", __func__,
2027 lseg, lseg->pls_range.iomode,
2028 lseg->pls_range.offset,
2029 lseg->pls_range.length);
2030 if (mark_lseg_invalid(lseg, tmp_list))
2031 continue;
2032 remaining++;
2033 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2034 }
2035
2036 if (remaining)
2037 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2038
2039 return remaining;
2040}
2041
2042void pnfs_error_mark_layout_for_return(struct inode *inode,
2043 struct pnfs_layout_segment *lseg)
2044{
2045 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2046 struct pnfs_layout_range range = {
2047 .iomode = lseg->pls_range.iomode,
2048 .offset = 0,
2049 .length = NFS4_MAX_UINT64,
2050 };
2051 bool return_now = false;
2052
2053 spin_lock(&inode->i_lock);
2054 pnfs_set_plh_return_info(lo, range.iomode, 0);
2055 /* Block LAYOUTGET */
2056 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2057 /*
2058 * mark all matching lsegs so that we are sure to have no live
2059 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2060 * for how it works.
2061 */
2062 if (!pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, &range, 0)) {
2063 nfs4_stateid stateid;
2064 enum pnfs_iomode iomode;
2065
2066 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &iomode);
2067 spin_unlock(&inode->i_lock);
2068 if (return_now)
2069 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
2070 } else {
2071 spin_unlock(&inode->i_lock);
2072 nfs_commit_inode(inode, 0);
2073 }
2074}
2075EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2076
2077void
2078pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2079{
2080 u64 rd_size = req->wb_bytes;
2081
2082 if (pgio->pg_lseg == NULL) {
2083 if (pgio->pg_dreq == NULL)
2084 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2085 else
2086 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2087
2088 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2089 req->wb_context,
2090 req_offset(req),
2091 rd_size,
2092 IOMODE_READ,
2093 false,
2094 GFP_KERNEL);
2095 if (IS_ERR(pgio->pg_lseg)) {
2096 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2097 pgio->pg_lseg = NULL;
2098 return;
2099 }
2100 }
2101 /* If no lseg, fall back to read through mds */
2102 if (pgio->pg_lseg == NULL)
2103 nfs_pageio_reset_read_mds(pgio);
2104
2105}
2106EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2107
2108void
2109pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2110 struct nfs_page *req, u64 wb_size)
2111{
2112 if (pgio->pg_lseg == NULL) {
2113 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2114 req->wb_context,
2115 req_offset(req),
2116 wb_size,
2117 IOMODE_RW,
2118 false,
2119 GFP_NOFS);
2120 if (IS_ERR(pgio->pg_lseg)) {
2121 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2122 pgio->pg_lseg = NULL;
2123 return;
2124 }
2125 }
2126 /* If no lseg, fall back to write through mds */
2127 if (pgio->pg_lseg == NULL)
2128 nfs_pageio_reset_write_mds(pgio);
2129}
2130EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2131
2132void
2133pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2134{
2135 if (desc->pg_lseg) {
2136 pnfs_put_lseg(desc->pg_lseg);
2137 desc->pg_lseg = NULL;
2138 }
2139}
2140EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2141
2142/*
2143 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2144 * of bytes (maximum @req->wb_bytes) that can be coalesced.
2145 */
2146size_t
2147pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2148 struct nfs_page *prev, struct nfs_page *req)
2149{
2150 unsigned int size;
2151 u64 seg_end, req_start, seg_left;
2152
2153 size = nfs_generic_pg_test(pgio, prev, req);
2154 if (!size)
2155 return 0;
2156
2157 /*
2158 * 'size' contains the number of bytes left in the current page (up
2159 * to the original size asked for in @req->wb_bytes).
2160 *
2161 * Calculate how many bytes are left in the layout segment
2162 * and if there are less bytes than 'size', return that instead.
2163 *
2164 * Please also note that 'end_offset' is actually the offset of the
2165 * first byte that lies outside the pnfs_layout_range. FIXME?
2166 *
2167 */
2168 if (pgio->pg_lseg) {
2169 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2170 pgio->pg_lseg->pls_range.length);
2171 req_start = req_offset(req);
2172 WARN_ON_ONCE(req_start >= seg_end);
2173 /* start of request is past the last byte of this segment */
2174 if (req_start >= seg_end) {
2175 /* reference the new lseg */
2176 if (pgio->pg_ops->pg_cleanup)
2177 pgio->pg_ops->pg_cleanup(pgio);
2178 if (pgio->pg_ops->pg_init)
2179 pgio->pg_ops->pg_init(pgio, req);
2180 return 0;
2181 }
2182
2183 /* adjust 'size' iff there are fewer bytes left in the
2184 * segment than what nfs_generic_pg_test returned */
2185 seg_left = seg_end - req_start;
2186 if (seg_left < size)
2187 size = (unsigned int)seg_left;
2188 }
2189
2190 return size;
2191}
2192EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2193
2194int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2195{
2196 struct nfs_pageio_descriptor pgio;
2197
2198 /* Resend all requests through the MDS */
2199 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2200 hdr->completion_ops);
2201 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2202 return nfs_pageio_resend(&pgio, hdr);
2203}
2204EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2205
2206static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2207{
2208
2209 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2210 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2211 PNFS_LAYOUTRET_ON_ERROR) {
2212 pnfs_return_layout(hdr->inode);
2213 }
2214 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2215 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2216}
2217
2218/*
2219 * Called by non rpc-based layout drivers
2220 */
2221void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2222{
2223 if (likely(!hdr->pnfs_error)) {
2224 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2225 hdr->mds_offset + hdr->res.count);
2226 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2227 }
2228 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2229 if (unlikely(hdr->pnfs_error))
2230 pnfs_ld_handle_write_error(hdr);
2231 hdr->mds_ops->rpc_release(hdr);
2232}
2233EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2234
2235static void
2236pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2237 struct nfs_pgio_header *hdr)
2238{
2239 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2240
2241 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2242 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2243 nfs_pageio_reset_write_mds(desc);
2244 mirror->pg_recoalesce = 1;
2245 }
2246 nfs_pgio_data_destroy(hdr);
2247 hdr->release(hdr);
2248}
2249
2250static enum pnfs_try_status
2251pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2252 const struct rpc_call_ops *call_ops,
2253 struct pnfs_layout_segment *lseg,
2254 int how)
2255{
2256 struct inode *inode = hdr->inode;
2257 enum pnfs_try_status trypnfs;
2258 struct nfs_server *nfss = NFS_SERVER(inode);
2259
2260 hdr->mds_ops = call_ops;
2261
2262 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2263 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2264 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2265 if (trypnfs != PNFS_NOT_ATTEMPTED)
2266 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2267 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2268 return trypnfs;
2269}
2270
2271static void
2272pnfs_do_write(struct nfs_pageio_descriptor *desc,
2273 struct nfs_pgio_header *hdr, int how)
2274{
2275 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2276 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2277 enum pnfs_try_status trypnfs;
2278
2279 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2280 if (trypnfs == PNFS_NOT_ATTEMPTED)
2281 pnfs_write_through_mds(desc, hdr);
2282}
2283
2284static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2285{
2286 pnfs_put_lseg(hdr->lseg);
2287 nfs_pgio_header_free(hdr);
2288}
2289
2290int
2291pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2292{
2293 struct nfs_pgio_header *hdr;
2294 int ret;
2295
2296 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2297 if (!hdr) {
2298 desc->pg_error = -ENOMEM;
2299 return desc->pg_error;
2300 }
2301 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2302
2303 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2304 ret = nfs_generic_pgio(desc, hdr);
2305 if (!ret)
2306 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2307
2308 return ret;
2309}
2310EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2311
2312int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2313{
2314 struct nfs_pageio_descriptor pgio;
2315
2316 /* Resend all requests through the MDS */
2317 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2318 return nfs_pageio_resend(&pgio, hdr);
2319}
2320EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2321
2322static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2323{
2324 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2325 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2326 PNFS_LAYOUTRET_ON_ERROR) {
2327 pnfs_return_layout(hdr->inode);
2328 }
2329 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2330 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2331}
2332
2333/*
2334 * Called by non rpc-based layout drivers
2335 */
2336void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2337{
2338 if (likely(!hdr->pnfs_error))
2339 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2340 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2341 if (unlikely(hdr->pnfs_error))
2342 pnfs_ld_handle_read_error(hdr);
2343 hdr->mds_ops->rpc_release(hdr);
2344}
2345EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2346
2347static void
2348pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2349 struct nfs_pgio_header *hdr)
2350{
2351 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2352
2353 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2354 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2355 nfs_pageio_reset_read_mds(desc);
2356 mirror->pg_recoalesce = 1;
2357 }
2358 nfs_pgio_data_destroy(hdr);
2359 hdr->release(hdr);
2360}
2361
2362/*
2363 * Call the appropriate parallel I/O subsystem read function.
2364 */
2365static enum pnfs_try_status
2366pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2367 const struct rpc_call_ops *call_ops,
2368 struct pnfs_layout_segment *lseg)
2369{
2370 struct inode *inode = hdr->inode;
2371 struct nfs_server *nfss = NFS_SERVER(inode);
2372 enum pnfs_try_status trypnfs;
2373
2374 hdr->mds_ops = call_ops;
2375
2376 dprintk("%s: Reading ino:%lu %u@%llu\n",
2377 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2378
2379 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2380 if (trypnfs != PNFS_NOT_ATTEMPTED)
2381 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2382 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2383 return trypnfs;
2384}
2385
2386/* Resend all requests through pnfs. */
2387void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2388{
2389 struct nfs_pageio_descriptor pgio;
2390
2391 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2392 /* Prevent deadlocks with layoutreturn! */
2393 pnfs_put_lseg(hdr->lseg);
2394 hdr->lseg = NULL;
2395
2396 nfs_pageio_init_read(&pgio, hdr->inode, false,
2397 hdr->completion_ops);
2398 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2399 }
2400}
2401EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2402
2403static void
2404pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2405{
2406 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2407 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2408 enum pnfs_try_status trypnfs;
2409
2410 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2411 if (trypnfs == PNFS_TRY_AGAIN)
2412 pnfs_read_resend_pnfs(hdr);
2413 if (trypnfs == PNFS_NOT_ATTEMPTED || hdr->task.tk_status)
2414 pnfs_read_through_mds(desc, hdr);
2415}
2416
2417static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2418{
2419 pnfs_put_lseg(hdr->lseg);
2420 nfs_pgio_header_free(hdr);
2421}
2422
2423int
2424pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2425{
2426 struct nfs_pgio_header *hdr;
2427 int ret;
2428
2429 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2430 if (!hdr) {
2431 desc->pg_error = -ENOMEM;
2432 return desc->pg_error;
2433 }
2434 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2435 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2436 ret = nfs_generic_pgio(desc, hdr);
2437 if (!ret)
2438 pnfs_do_read(desc, hdr);
2439 return ret;
2440}
2441EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2442
2443static void pnfs_clear_layoutcommitting(struct inode *inode)
2444{
2445 unsigned long *bitlock = &NFS_I(inode)->flags;
2446
2447 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2448 smp_mb__after_atomic();
2449 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2450}
2451
2452/*
2453 * There can be multiple RW segments.
2454 */
2455static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2456{
2457 struct pnfs_layout_segment *lseg;
2458
2459 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2460 if (lseg->pls_range.iomode == IOMODE_RW &&
2461 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2462 list_add(&lseg->pls_lc_list, listp);
2463 }
2464}
2465
2466static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2467{
2468 struct pnfs_layout_segment *lseg, *tmp;
2469
2470 /* Matched by references in pnfs_set_layoutcommit */
2471 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2472 list_del_init(&lseg->pls_lc_list);
2473 pnfs_put_lseg(lseg);
2474 }
2475
2476 pnfs_clear_layoutcommitting(inode);
2477}
2478
2479void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2480{
2481 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2482}
2483EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2484
2485void
2486pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2487 loff_t end_pos)
2488{
2489 struct nfs_inode *nfsi = NFS_I(inode);
2490 bool mark_as_dirty = false;
2491
2492 spin_lock(&inode->i_lock);
2493 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2494 nfsi->layout->plh_lwb = end_pos;
2495 mark_as_dirty = true;
2496 dprintk("%s: Set layoutcommit for inode %lu ",
2497 __func__, inode->i_ino);
2498 } else if (end_pos > nfsi->layout->plh_lwb)
2499 nfsi->layout->plh_lwb = end_pos;
2500 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2501 /* references matched in nfs4_layoutcommit_release */
2502 pnfs_get_lseg(lseg);
2503 }
2504 spin_unlock(&inode->i_lock);
2505 dprintk("%s: lseg %p end_pos %llu\n",
2506 __func__, lseg, nfsi->layout->plh_lwb);
2507
2508 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2509 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2510 if (mark_as_dirty)
2511 mark_inode_dirty_sync(inode);
2512}
2513EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2514
2515void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2516{
2517 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2518
2519 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2520 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2521 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2522}
2523
2524/*
2525 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2526 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2527 * data to disk to allow the server to recover the data if it crashes.
2528 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2529 * is off, and a COMMIT is sent to a data server, or
2530 * if WRITEs to a data server return NFS_DATA_SYNC.
2531 */
2532int
2533pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2534{
2535 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2536 struct nfs4_layoutcommit_data *data;
2537 struct nfs_inode *nfsi = NFS_I(inode);
2538 loff_t end_pos;
2539 int status;
2540
2541 if (!pnfs_layoutcommit_outstanding(inode))
2542 return 0;
2543
2544 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2545
2546 status = -EAGAIN;
2547 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2548 if (!sync)
2549 goto out;
2550 status = wait_on_bit_lock_action(&nfsi->flags,
2551 NFS_INO_LAYOUTCOMMITTING,
2552 nfs_wait_bit_killable,
2553 TASK_KILLABLE);
2554 if (status)
2555 goto out;
2556 }
2557
2558 status = -ENOMEM;
2559 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2560 data = kzalloc(sizeof(*data), GFP_NOFS);
2561 if (!data)
2562 goto clear_layoutcommitting;
2563
2564 status = 0;
2565 spin_lock(&inode->i_lock);
2566 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2567 goto out_unlock;
2568
2569 INIT_LIST_HEAD(&data->lseg_list);
2570 pnfs_list_write_lseg(inode, &data->lseg_list);
2571
2572 end_pos = nfsi->layout->plh_lwb;
2573
2574 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2575 spin_unlock(&inode->i_lock);
2576
2577 data->args.inode = inode;
2578 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2579 nfs_fattr_init(&data->fattr);
2580 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2581 data->res.fattr = &data->fattr;
2582 if (end_pos != 0)
2583 data->args.lastbytewritten = end_pos - 1;
2584 else
2585 data->args.lastbytewritten = U64_MAX;
2586 data->res.server = NFS_SERVER(inode);
2587
2588 if (ld->prepare_layoutcommit) {
2589 status = ld->prepare_layoutcommit(&data->args);
2590 if (status) {
2591 put_rpccred(data->cred);
2592 spin_lock(&inode->i_lock);
2593 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2594 if (end_pos > nfsi->layout->plh_lwb)
2595 nfsi->layout->plh_lwb = end_pos;
2596 goto out_unlock;
2597 }
2598 }
2599
2600
2601 status = nfs4_proc_layoutcommit(data, sync);
2602out:
2603 if (status)
2604 mark_inode_dirty_sync(inode);
2605 dprintk("<-- %s status %d\n", __func__, status);
2606 return status;
2607out_unlock:
2608 spin_unlock(&inode->i_lock);
2609 kfree(data);
2610clear_layoutcommitting:
2611 pnfs_clear_layoutcommitting(inode);
2612 goto out;
2613}
2614EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2615
2616int
2617pnfs_generic_sync(struct inode *inode, bool datasync)
2618{
2619 return pnfs_layoutcommit_inode(inode, true);
2620}
2621EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2622
2623struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2624{
2625 struct nfs4_threshold *thp;
2626
2627 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2628 if (!thp) {
2629 dprintk("%s mdsthreshold allocation failed\n", __func__);
2630 return NULL;
2631 }
2632 return thp;
2633}
2634
2635#if IS_ENABLED(CONFIG_NFS_V4_2)
2636int
2637pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2638{
2639 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2640 struct nfs_server *server = NFS_SERVER(inode);
2641 struct nfs_inode *nfsi = NFS_I(inode);
2642 struct nfs42_layoutstat_data *data;
2643 struct pnfs_layout_hdr *hdr;
2644 int status = 0;
2645
2646 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2647 goto out;
2648
2649 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2650 goto out;
2651
2652 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2653 goto out;
2654
2655 spin_lock(&inode->i_lock);
2656 if (!NFS_I(inode)->layout) {
2657 spin_unlock(&inode->i_lock);
2658 goto out_clear_layoutstats;
2659 }
2660 hdr = NFS_I(inode)->layout;
2661 pnfs_get_layout_hdr(hdr);
2662 spin_unlock(&inode->i_lock);
2663
2664 data = kzalloc(sizeof(*data), gfp_flags);
2665 if (!data) {
2666 status = -ENOMEM;
2667 goto out_put;
2668 }
2669
2670 data->args.fh = NFS_FH(inode);
2671 data->args.inode = inode;
2672 status = ld->prepare_layoutstats(&data->args);
2673 if (status)
2674 goto out_free;
2675
2676 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2677
2678out:
2679 dprintk("%s returns %d\n", __func__, status);
2680 return status;
2681
2682out_free:
2683 kfree(data);
2684out_put:
2685 pnfs_put_layout_hdr(hdr);
2686out_clear_layoutstats:
2687 smp_mb__before_atomic();
2688 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2689 smp_mb__after_atomic();
2690 goto out;
2691}
2692EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2693#endif
2694
2695unsigned int layoutstats_timer;
2696module_param(layoutstats_timer, uint, 0644);
2697EXPORT_SYMBOL_GPL(layoutstats_timer);