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