4448ff829cbb936cf51eceadbd641cc37e2c7f12
[releases.git] / pnfs.c
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  */
50 static DEFINE_SPINLOCK(pnfs_spinlock);
51
52 /*
53  * pnfs_modules_tbl holds all pnfs modules
54  */
55 static LIST_HEAD(pnfs_modules_tbl);
56
57 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
58 static 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);
62 static 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 */
66 static struct pnfs_layoutdriver_type *
67 find_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;
75 out:
76         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
77         return local;
78 }
79
80 static struct pnfs_layoutdriver_type *
81 find_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
95 const struct pnfs_layoutdriver_type *pnfs_find_layoutdriver(u32 id)
96 {
97         return find_pnfs_driver(id);
98 }
99
100 void pnfs_put_layoutdriver(const struct pnfs_layoutdriver_type *ld)
101 {
102         if (ld)
103                 module_put(ld->owner);
104 }
105
106 void
107 unset_pnfs_layoutdriver(struct nfs_server *nfss)
108 {
109         if (nfss->pnfs_curr_ld) {
110                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
111                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
112                 /* Decrement the MDS count. Purge the deviceid cache if zero */
113                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
114                         nfs4_deviceid_purge_client(nfss->nfs_client);
115                 module_put(nfss->pnfs_curr_ld->owner);
116         }
117         nfss->pnfs_curr_ld = NULL;
118 }
119
120 /*
121  * When the server sends a list of layout types, we choose one in the order
122  * given in the list below.
123  *
124  * FIXME: should this list be configurable in some fashion? module param?
125  *        mount option? something else?
126  */
127 static const u32 ld_prefs[] = {
128         LAYOUT_SCSI,
129         LAYOUT_BLOCK_VOLUME,
130         LAYOUT_OSD2_OBJECTS,
131         LAYOUT_FLEX_FILES,
132         LAYOUT_NFSV4_1_FILES,
133         0
134 };
135
136 static int
137 ld_cmp(const void *e1, const void *e2)
138 {
139         u32 ld1 = *((u32 *)e1);
140         u32 ld2 = *((u32 *)e2);
141         int i;
142
143         for (i = 0; ld_prefs[i] != 0; i++) {
144                 if (ld1 == ld_prefs[i])
145                         return -1;
146
147                 if (ld2 == ld_prefs[i])
148                         return 1;
149         }
150         return 0;
151 }
152
153 /*
154  * Try to set the server's pnfs module to the pnfs layout type specified by id.
155  * Currently only one pNFS layout driver per filesystem is supported.
156  *
157  * @ids array of layout types supported by MDS.
158  */
159 void
160 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
161                       struct nfs_fsinfo *fsinfo)
162 {
163         struct pnfs_layoutdriver_type *ld_type = NULL;
164         u32 id;
165         int i;
166
167         if (fsinfo->nlayouttypes == 0)
168                 goto out_no_driver;
169         if (!(server->nfs_client->cl_exchange_flags &
170                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
171                 printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
172                         __func__, server->nfs_client->cl_exchange_flags);
173                 goto out_no_driver;
174         }
175
176         sort(fsinfo->layouttype, fsinfo->nlayouttypes,
177                 sizeof(*fsinfo->layouttype), ld_cmp, NULL);
178
179         for (i = 0; i < fsinfo->nlayouttypes; i++) {
180                 id = fsinfo->layouttype[i];
181                 ld_type = find_pnfs_driver(id);
182                 if (!ld_type) {
183                         request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
184                                         id);
185                         ld_type = find_pnfs_driver(id);
186                 }
187                 if (ld_type)
188                         break;
189         }
190
191         if (!ld_type) {
192                 dprintk("%s: No pNFS module found!\n", __func__);
193                 goto out_no_driver;
194         }
195
196         server->pnfs_curr_ld = ld_type;
197         if (ld_type->set_layoutdriver
198             && ld_type->set_layoutdriver(server, mntfh)) {
199                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
200                         "driver %u.\n", __func__, id);
201                 module_put(ld_type->owner);
202                 goto out_no_driver;
203         }
204         /* Bump the MDS count */
205         atomic_inc(&server->nfs_client->cl_mds_count);
206
207         dprintk("%s: pNFS module for %u set\n", __func__, id);
208         return;
209
210 out_no_driver:
211         dprintk("%s: Using NFSv4 I/O\n", __func__);
212         server->pnfs_curr_ld = NULL;
213 }
214
215 int
216 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
217 {
218         int status = -EINVAL;
219         struct pnfs_layoutdriver_type *tmp;
220
221         if (ld_type->id == 0) {
222                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
223                 return status;
224         }
225         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
226                 printk(KERN_ERR "NFS: %s Layout driver must provide "
227                        "alloc_lseg and free_lseg.\n", __func__);
228                 return status;
229         }
230
231         spin_lock(&pnfs_spinlock);
232         tmp = find_pnfs_driver_locked(ld_type->id);
233         if (!tmp) {
234                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
235                 status = 0;
236                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
237                         ld_type->name);
238         } else {
239                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
240                         __func__, ld_type->id);
241         }
242         spin_unlock(&pnfs_spinlock);
243
244         return status;
245 }
246 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
247
248 void
249 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
250 {
251         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
252         spin_lock(&pnfs_spinlock);
253         list_del(&ld_type->pnfs_tblid);
254         spin_unlock(&pnfs_spinlock);
255 }
256 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
257
258 /*
259  * pNFS client layout cache
260  */
261
262 /* Need to hold i_lock if caller does not already hold reference */
263 void
264 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
265 {
266         refcount_inc(&lo->plh_refcount);
267 }
268
269 static struct pnfs_layout_hdr *
270 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
271 {
272         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
273         return ld->alloc_layout_hdr(ino, gfp_flags);
274 }
275
276 static void
277 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
278 {
279         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
280         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
281
282         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
283                 struct nfs_client *clp = server->nfs_client;
284
285                 spin_lock(&clp->cl_lock);
286                 list_del_rcu(&lo->plh_layouts);
287                 spin_unlock(&clp->cl_lock);
288         }
289         put_cred(lo->plh_lc_cred);
290         return ld->free_layout_hdr(lo);
291 }
292
293 static void
294 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
295 {
296         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
297         dprintk("%s: freeing layout cache %p\n", __func__, lo);
298         nfsi->layout = NULL;
299         /* Reset MDS Threshold I/O counters */
300         nfsi->write_io = 0;
301         nfsi->read_io = 0;
302 }
303
304 void
305 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
306 {
307         struct inode *inode;
308         unsigned long i_state;
309
310         if (!lo)
311                 return;
312         inode = lo->plh_inode;
313         pnfs_layoutreturn_before_put_layout_hdr(lo);
314
315         if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
316                 if (!list_empty(&lo->plh_segs))
317                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
318                 pnfs_detach_layout_hdr(lo);
319                 i_state = inode->i_state;
320                 spin_unlock(&inode->i_lock);
321                 pnfs_free_layout_hdr(lo);
322                 /* Notify pnfs_destroy_layout_final() that we're done */
323                 if (i_state & (I_FREEING | I_CLEAR))
324                         wake_up_var(lo);
325         }
326 }
327
328 static struct inode *
329 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
330 {
331         struct inode *inode = igrab(lo->plh_inode);
332         if (inode)
333                 return inode;
334         set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
335         return NULL;
336 }
337
338 /*
339  * Compare 2 layout stateid sequence ids, to see which is newer,
340  * taking into account wraparound issues.
341  */
342 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
343 {
344         return (s32)(s1 - s2) > 0;
345 }
346
347 static void pnfs_barrier_update(struct pnfs_layout_hdr *lo, u32 newseq)
348 {
349         if (pnfs_seqid_is_newer(newseq, lo->plh_barrier) || !lo->plh_barrier)
350                 lo->plh_barrier = newseq;
351 }
352
353 static void
354 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
355                          u32 seq)
356 {
357         if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
358                 iomode = IOMODE_ANY;
359         lo->plh_return_iomode = iomode;
360         set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
361         /*
362          * We must set lo->plh_return_seq to avoid livelocks with
363          * pnfs_layout_need_return()
364          */
365         if (seq == 0)
366                 seq = be32_to_cpu(lo->plh_stateid.seqid);
367         if (!lo->plh_return_seq || pnfs_seqid_is_newer(seq, lo->plh_return_seq))
368                 lo->plh_return_seq = seq;
369         pnfs_barrier_update(lo, seq);
370 }
371
372 static void
373 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
374 {
375         struct pnfs_layout_segment *lseg;
376         lo->plh_return_iomode = 0;
377         lo->plh_return_seq = 0;
378         clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
379         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
380                 if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
381                         continue;
382                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
383         }
384 }
385
386 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
387 {
388         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
389         clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
390         smp_mb__after_atomic();
391         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
392         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
393 }
394
395 static void
396 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
397                 struct list_head *free_me)
398 {
399         clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
400         clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
401         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
402                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
403         if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
404                 pnfs_lseg_dec_and_remove_zero(lseg, free_me);
405 }
406
407 /*
408  * Update the seqid of a layout stateid after receiving
409  * NFS4ERR_OLD_STATEID
410  */
411 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
412                 struct pnfs_layout_range *dst_range,
413                 struct inode *inode)
414 {
415         struct pnfs_layout_hdr *lo;
416         struct pnfs_layout_range range = {
417                 .iomode = IOMODE_ANY,
418                 .offset = 0,
419                 .length = NFS4_MAX_UINT64,
420         };
421         bool ret = false;
422         LIST_HEAD(head);
423         int err;
424
425         spin_lock(&inode->i_lock);
426         lo = NFS_I(inode)->layout;
427         if (lo &&  pnfs_layout_is_valid(lo) &&
428             nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
429                 /* Is our call using the most recent seqid? If so, bump it */
430                 if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
431                         nfs4_stateid_seqid_inc(dst);
432                         ret = true;
433                         goto out;
434                 }
435                 /* Try to update the seqid to the most recent */
436                 err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
437                 if (err != -EBUSY) {
438                         dst->seqid = lo->plh_stateid.seqid;
439                         *dst_range = range;
440                         ret = true;
441                 }
442         }
443 out:
444         spin_unlock(&inode->i_lock);
445         pnfs_free_lseg_list(&head);
446         return ret;
447 }
448
449 /*
450  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
451  *
452  * In order to continue using the pnfs_layout_hdr, a full recovery
453  * is required.
454  * Note that caller must hold inode->i_lock.
455  */
456 int
457 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
458                 struct list_head *lseg_list)
459 {
460         struct pnfs_layout_range range = {
461                 .iomode = IOMODE_ANY,
462                 .offset = 0,
463                 .length = NFS4_MAX_UINT64,
464         };
465         struct pnfs_layout_segment *lseg, *next;
466
467         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
468         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
469                 pnfs_clear_lseg_state(lseg, lseg_list);
470         pnfs_clear_layoutreturn_info(lo);
471         pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
472         set_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags);
473         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
474             !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
475                 pnfs_clear_layoutreturn_waitbit(lo);
476         return !list_empty(&lo->plh_segs);
477 }
478
479 static int
480 pnfs_iomode_to_fail_bit(u32 iomode)
481 {
482         return iomode == IOMODE_RW ?
483                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
484 }
485
486 static void
487 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
488 {
489         lo->plh_retry_timestamp = jiffies;
490         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
491                 refcount_inc(&lo->plh_refcount);
492 }
493
494 static void
495 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
496 {
497         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
498                 refcount_dec(&lo->plh_refcount);
499 }
500
501 static void
502 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
503 {
504         struct inode *inode = lo->plh_inode;
505         struct pnfs_layout_range range = {
506                 .iomode = iomode,
507                 .offset = 0,
508                 .length = NFS4_MAX_UINT64,
509         };
510         LIST_HEAD(head);
511
512         spin_lock(&inode->i_lock);
513         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
514         pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
515         spin_unlock(&inode->i_lock);
516         pnfs_free_lseg_list(&head);
517         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
518                         iomode == IOMODE_RW ?  "RW" : "READ");
519 }
520
521 static bool
522 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
523 {
524         unsigned long start, end;
525         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
526
527         if (test_bit(fail_bit, &lo->plh_flags) == 0)
528                 return false;
529         end = jiffies;
530         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
531         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
532                 /* It is time to retry the failed layoutgets */
533                 pnfs_layout_clear_fail_bit(lo, fail_bit);
534                 return false;
535         }
536         return true;
537 }
538
539 static void
540 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
541                 const struct pnfs_layout_range *range,
542                 const nfs4_stateid *stateid)
543 {
544         INIT_LIST_HEAD(&lseg->pls_list);
545         INIT_LIST_HEAD(&lseg->pls_lc_list);
546         INIT_LIST_HEAD(&lseg->pls_commits);
547         refcount_set(&lseg->pls_refcount, 1);
548         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
549         lseg->pls_layout = lo;
550         lseg->pls_range = *range;
551         lseg->pls_seq = be32_to_cpu(stateid->seqid);
552 }
553
554 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
555 {
556         if (lseg != NULL) {
557                 struct inode *inode = lseg->pls_layout->plh_inode;
558                 NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
559         }
560 }
561
562 static void
563 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
564                 struct pnfs_layout_segment *lseg)
565 {
566         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
567         list_del_init(&lseg->pls_list);
568         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
569         refcount_dec(&lo->plh_refcount);
570         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
571                 return;
572         if (list_empty(&lo->plh_segs) &&
573             !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
574             !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
575                 if (atomic_read(&lo->plh_outstanding) == 0)
576                         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
577                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
578         }
579 }
580
581 static bool
582 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
583                 struct pnfs_layout_segment *lseg)
584 {
585         if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
586             pnfs_layout_is_valid(lo)) {
587                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
588                 list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
589                 return true;
590         }
591         return false;
592 }
593
594 void
595 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
596 {
597         struct pnfs_layout_hdr *lo;
598         struct inode *inode;
599
600         if (!lseg)
601                 return;
602
603         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
604                 refcount_read(&lseg->pls_refcount),
605                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
606
607         lo = lseg->pls_layout;
608         inode = lo->plh_inode;
609
610         if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
611                 pnfs_get_layout_hdr(lo);
612                 pnfs_layout_remove_lseg(lo, lseg);
613                 if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
614                         lseg = NULL;
615                 spin_unlock(&inode->i_lock);
616                 pnfs_free_lseg(lseg);
617                 pnfs_put_layout_hdr(lo);
618         }
619 }
620 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
621
622 /*
623  * is l2 fully contained in l1?
624  *   start1                             end1
625  *   [----------------------------------)
626  *           start2           end2
627  *           [----------------)
628  */
629 static bool
630 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
631                  const struct pnfs_layout_range *l2)
632 {
633         u64 start1 = l1->offset;
634         u64 end1 = pnfs_end_offset(start1, l1->length);
635         u64 start2 = l2->offset;
636         u64 end2 = pnfs_end_offset(start2, l2->length);
637
638         return (start1 <= start2) && (end1 >= end2);
639 }
640
641 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
642                 struct list_head *tmp_list)
643 {
644         if (!refcount_dec_and_test(&lseg->pls_refcount))
645                 return false;
646         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
647         list_add(&lseg->pls_list, tmp_list);
648         return true;
649 }
650
651 /* Returns 1 if lseg is removed from list, 0 otherwise */
652 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
653                              struct list_head *tmp_list)
654 {
655         int rv = 0;
656
657         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
658                 /* Remove the reference keeping the lseg in the
659                  * list.  It will now be removed when all
660                  * outstanding io is finished.
661                  */
662                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
663                         refcount_read(&lseg->pls_refcount));
664                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
665                         rv = 1;
666         }
667         return rv;
668 }
669
670 static bool
671 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
672                  const struct pnfs_layout_range *recall_range)
673 {
674         return (recall_range->iomode == IOMODE_ANY ||
675                 lseg_range->iomode == recall_range->iomode) &&
676                pnfs_lseg_range_intersecting(lseg_range, recall_range);
677 }
678
679 static bool
680 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
681                 const struct pnfs_layout_range *recall_range,
682                 u32 seq)
683 {
684         if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
685                 return false;
686         if (recall_range == NULL)
687                 return true;
688         return pnfs_should_free_range(&lseg->pls_range, recall_range);
689 }
690
691 /**
692  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
693  * @lo: layout header containing the lsegs
694  * @tmp_list: list head where doomed lsegs should go
695  * @recall_range: optional recall range argument to match (may be NULL)
696  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
697  *
698  * Walk the list of lsegs in the layout header, and tear down any that should
699  * be destroyed. If "recall_range" is specified then the segment must match
700  * that range. If "seq" is non-zero, then only match segments that were handed
701  * out at or before that sequence.
702  *
703  * Returns number of matching invalid lsegs remaining in list after scanning
704  * it and purging them.
705  */
706 int
707 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
708                             struct list_head *tmp_list,
709                             const struct pnfs_layout_range *recall_range,
710                             u32 seq)
711 {
712         struct pnfs_layout_segment *lseg, *next;
713         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
714         int remaining = 0;
715
716         dprintk("%s:Begin lo %p\n", __func__, lo);
717
718         if (list_empty(&lo->plh_segs))
719                 return 0;
720         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
721                 if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
722                         dprintk("%s: freeing lseg %p iomode %d seq %u "
723                                 "offset %llu length %llu\n", __func__,
724                                 lseg, lseg->pls_range.iomode, lseg->pls_seq,
725                                 lseg->pls_range.offset, lseg->pls_range.length);
726                         if (mark_lseg_invalid(lseg, tmp_list))
727                                 continue;
728                         remaining++;
729                         pnfs_lseg_cancel_io(server, lseg);
730                 }
731         dprintk("%s:Return %i\n", __func__, remaining);
732         return remaining;
733 }
734
735 static void
736 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
737                 struct list_head *free_me,
738                 const struct pnfs_layout_range *range,
739                 u32 seq)
740 {
741         struct pnfs_layout_segment *lseg, *next;
742
743         list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
744                 if (pnfs_match_lseg_recall(lseg, range, seq))
745                         list_move_tail(&lseg->pls_list, free_me);
746         }
747 }
748
749 /* note free_me must contain lsegs from a single layout_hdr */
750 void
751 pnfs_free_lseg_list(struct list_head *free_me)
752 {
753         struct pnfs_layout_segment *lseg, *tmp;
754
755         if (list_empty(free_me))
756                 return;
757
758         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
759                 list_del(&lseg->pls_list);
760                 pnfs_free_lseg(lseg);
761         }
762 }
763
764 static struct pnfs_layout_hdr *__pnfs_destroy_layout(struct nfs_inode *nfsi)
765 {
766         struct pnfs_layout_hdr *lo;
767         LIST_HEAD(tmp_list);
768
769         spin_lock(&nfsi->vfs_inode.i_lock);
770         lo = nfsi->layout;
771         if (lo) {
772                 pnfs_get_layout_hdr(lo);
773                 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
774                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
775                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
776                 spin_unlock(&nfsi->vfs_inode.i_lock);
777                 pnfs_free_lseg_list(&tmp_list);
778                 nfs_commit_inode(&nfsi->vfs_inode, 0);
779                 pnfs_put_layout_hdr(lo);
780         } else
781                 spin_unlock(&nfsi->vfs_inode.i_lock);
782         return lo;
783 }
784
785 void pnfs_destroy_layout(struct nfs_inode *nfsi)
786 {
787         __pnfs_destroy_layout(nfsi);
788 }
789 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
790
791 static bool pnfs_layout_removed(struct nfs_inode *nfsi,
792                                 struct pnfs_layout_hdr *lo)
793 {
794         bool ret;
795
796         spin_lock(&nfsi->vfs_inode.i_lock);
797         ret = nfsi->layout != lo;
798         spin_unlock(&nfsi->vfs_inode.i_lock);
799         return ret;
800 }
801
802 void pnfs_destroy_layout_final(struct nfs_inode *nfsi)
803 {
804         struct pnfs_layout_hdr *lo = __pnfs_destroy_layout(nfsi);
805
806         if (lo)
807                 wait_var_event(lo, pnfs_layout_removed(nfsi, lo));
808 }
809
810 static bool
811 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
812                 struct list_head *layout_list)
813 {
814         struct pnfs_layout_hdr *lo;
815         bool ret = false;
816
817         spin_lock(&inode->i_lock);
818         lo = NFS_I(inode)->layout;
819         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
820                 pnfs_get_layout_hdr(lo);
821                 list_add(&lo->plh_bulk_destroy, layout_list);
822                 ret = true;
823         }
824         spin_unlock(&inode->i_lock);
825         return ret;
826 }
827
828 /* Caller must hold rcu_read_lock and clp->cl_lock */
829 static int
830 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
831                 struct nfs_server *server,
832                 struct list_head *layout_list)
833         __must_hold(&clp->cl_lock)
834         __must_hold(RCU)
835 {
836         struct pnfs_layout_hdr *lo, *next;
837         struct inode *inode;
838
839         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
840                 if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
841                     test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
842                     !list_empty(&lo->plh_bulk_destroy))
843                         continue;
844                 /* If the sb is being destroyed, just bail */
845                 if (!nfs_sb_active(server->super))
846                         break;
847                 inode = pnfs_grab_inode_layout_hdr(lo);
848                 if (inode != NULL) {
849                         if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
850                                 list_del_rcu(&lo->plh_layouts);
851                         if (pnfs_layout_add_bulk_destroy_list(inode,
852                                                 layout_list))
853                                 continue;
854                         rcu_read_unlock();
855                         spin_unlock(&clp->cl_lock);
856                         iput(inode);
857                 } else {
858                         rcu_read_unlock();
859                         spin_unlock(&clp->cl_lock);
860                 }
861                 nfs_sb_deactive(server->super);
862                 spin_lock(&clp->cl_lock);
863                 rcu_read_lock();
864                 return -EAGAIN;
865         }
866         return 0;
867 }
868
869 static int
870 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
871                 bool is_bulk_recall)
872 {
873         struct pnfs_layout_hdr *lo;
874         struct inode *inode;
875         LIST_HEAD(lseg_list);
876         int ret = 0;
877
878         while (!list_empty(layout_list)) {
879                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
880                                 plh_bulk_destroy);
881                 dprintk("%s freeing layout for inode %lu\n", __func__,
882                         lo->plh_inode->i_ino);
883                 inode = lo->plh_inode;
884
885                 pnfs_layoutcommit_inode(inode, false);
886
887                 spin_lock(&inode->i_lock);
888                 list_del_init(&lo->plh_bulk_destroy);
889                 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
890                         if (is_bulk_recall)
891                                 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
892                         ret = -EAGAIN;
893                 }
894                 spin_unlock(&inode->i_lock);
895                 pnfs_free_lseg_list(&lseg_list);
896                 /* Free all lsegs that are attached to commit buckets */
897                 nfs_commit_inode(inode, 0);
898                 pnfs_put_layout_hdr(lo);
899                 nfs_iput_and_deactive(inode);
900         }
901         return ret;
902 }
903
904 int
905 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
906                 struct nfs_fsid *fsid,
907                 bool is_recall)
908 {
909         struct nfs_server *server;
910         LIST_HEAD(layout_list);
911
912         spin_lock(&clp->cl_lock);
913         rcu_read_lock();
914 restart:
915         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
916                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
917                         continue;
918                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
919                                 server,
920                                 &layout_list) != 0)
921                         goto restart;
922         }
923         rcu_read_unlock();
924         spin_unlock(&clp->cl_lock);
925
926         if (list_empty(&layout_list))
927                 return 0;
928         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
929 }
930
931 int
932 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
933                 bool is_recall)
934 {
935         struct nfs_server *server;
936         LIST_HEAD(layout_list);
937
938         spin_lock(&clp->cl_lock);
939         rcu_read_lock();
940 restart:
941         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
942                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
943                                         server,
944                                         &layout_list) != 0)
945                         goto restart;
946         }
947         rcu_read_unlock();
948         spin_unlock(&clp->cl_lock);
949
950         if (list_empty(&layout_list))
951                 return 0;
952         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
953 }
954
955 /*
956  * Called by the state manager to remove all layouts established under an
957  * expired lease.
958  */
959 void
960 pnfs_destroy_all_layouts(struct nfs_client *clp)
961 {
962         nfs4_deviceid_mark_client_invalid(clp);
963         nfs4_deviceid_purge_client(clp);
964
965         pnfs_destroy_layouts_byclid(clp, false);
966 }
967
968 static void
969 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
970 {
971         const struct cred *old;
972
973         if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
974                 old = xchg(&lo->plh_lc_cred, get_cred(cred));
975                 put_cred(old);
976         }
977 }
978
979 /* update lo->plh_stateid with new if is more recent */
980 void
981 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
982                         const struct cred *cred, bool update_barrier)
983 {
984         u32 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
985         u32 newseq = be32_to_cpu(new->seqid);
986
987         if (!pnfs_layout_is_valid(lo)) {
988                 pnfs_set_layout_cred(lo, cred);
989                 nfs4_stateid_copy(&lo->plh_stateid, new);
990                 lo->plh_barrier = newseq;
991                 pnfs_clear_layoutreturn_info(lo);
992                 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
993                 return;
994         }
995
996         if (pnfs_seqid_is_newer(newseq, oldseq))
997                 nfs4_stateid_copy(&lo->plh_stateid, new);
998
999         if (update_barrier) {
1000                 pnfs_barrier_update(lo, newseq);
1001                 return;
1002         }
1003         /*
1004          * Because of wraparound, we want to keep the barrier
1005          * "close" to the current seqids. We really only want to
1006          * get here from a layoutget call.
1007          */
1008         if (atomic_read(&lo->plh_outstanding) == 1)
1009                  pnfs_barrier_update(lo, be32_to_cpu(lo->plh_stateid.seqid));
1010 }
1011
1012 static bool
1013 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
1014                 const nfs4_stateid *stateid)
1015 {
1016         u32 seqid = be32_to_cpu(stateid->seqid);
1017
1018         return lo->plh_barrier && pnfs_seqid_is_newer(lo->plh_barrier, seqid);
1019 }
1020
1021 /* lget is set to 1 if called from inside send_layoutget call chain */
1022 static bool
1023 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
1024 {
1025         return lo->plh_block_lgets ||
1026                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
1027 }
1028
1029 static struct nfs_server *
1030 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
1031 {
1032         struct nfs_server *server;
1033
1034         if (inode) {
1035                 server = NFS_SERVER(inode);
1036         } else {
1037                 struct dentry *parent_dir = dget_parent(ctx->dentry);
1038                 server = NFS_SERVER(parent_dir->d_inode);
1039                 dput(parent_dir);
1040         }
1041         return server;
1042 }
1043
1044 static void nfs4_free_pages(struct page **pages, size_t size)
1045 {
1046         int i;
1047
1048         if (!pages)
1049                 return;
1050
1051         for (i = 0; i < size; i++) {
1052                 if (!pages[i])
1053                         break;
1054                 __free_page(pages[i]);
1055         }
1056         kfree(pages);
1057 }
1058
1059 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1060 {
1061         struct page **pages;
1062         int i;
1063
1064         pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1065         if (!pages) {
1066                 dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1067                 return NULL;
1068         }
1069
1070         for (i = 0; i < size; i++) {
1071                 pages[i] = alloc_page(gfp_flags);
1072                 if (!pages[i]) {
1073                         dprintk("%s: failed to allocate page\n", __func__);
1074                         nfs4_free_pages(pages, i);
1075                         return NULL;
1076                 }
1077         }
1078
1079         return pages;
1080 }
1081
1082 static struct nfs4_layoutget *
1083 pnfs_alloc_init_layoutget_args(struct inode *ino,
1084            struct nfs_open_context *ctx,
1085            const nfs4_stateid *stateid,
1086            const struct pnfs_layout_range *range,
1087            gfp_t gfp_flags)
1088 {
1089         struct nfs_server *server = pnfs_find_server(ino, ctx);
1090         size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1091         size_t max_pages = max_response_pages(server);
1092         struct nfs4_layoutget *lgp;
1093
1094         dprintk("--> %s\n", __func__);
1095
1096         lgp = kzalloc(sizeof(*lgp), gfp_flags);
1097         if (lgp == NULL)
1098                 return NULL;
1099
1100         if (max_reply_sz) {
1101                 size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1102                 if (npages < max_pages)
1103                         max_pages = npages;
1104         }
1105
1106         lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1107         if (!lgp->args.layout.pages) {
1108                 kfree(lgp);
1109                 return NULL;
1110         }
1111         lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1112         lgp->res.layoutp = &lgp->args.layout;
1113
1114         /* Don't confuse uninitialised result and success */
1115         lgp->res.status = -NFS4ERR_DELAY;
1116
1117         lgp->args.minlength = PAGE_SIZE;
1118         if (lgp->args.minlength > range->length)
1119                 lgp->args.minlength = range->length;
1120         if (ino) {
1121                 loff_t i_size = i_size_read(ino);
1122
1123                 if (range->iomode == IOMODE_READ) {
1124                         if (range->offset >= i_size)
1125                                 lgp->args.minlength = 0;
1126                         else if (i_size - range->offset < lgp->args.minlength)
1127                                 lgp->args.minlength = i_size - range->offset;
1128                 }
1129         }
1130         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1131         pnfs_copy_range(&lgp->args.range, range);
1132         lgp->args.type = server->pnfs_curr_ld->id;
1133         lgp->args.inode = ino;
1134         lgp->args.ctx = get_nfs_open_context(ctx);
1135         nfs4_stateid_copy(&lgp->args.stateid, stateid);
1136         lgp->gfp_flags = gfp_flags;
1137         lgp->cred = ctx->cred;
1138         return lgp;
1139 }
1140
1141 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1142 {
1143         size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1144
1145         nfs4_free_pages(lgp->args.layout.pages, max_pages);
1146         pnfs_put_layout_hdr(lgp->lo);
1147         put_nfs_open_context(lgp->args.ctx);
1148         kfree(lgp);
1149 }
1150
1151 static void pnfs_clear_layoutcommit(struct inode *inode,
1152                 struct list_head *head)
1153 {
1154         struct nfs_inode *nfsi = NFS_I(inode);
1155         struct pnfs_layout_segment *lseg, *tmp;
1156
1157         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1158                 return;
1159         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1160                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1161                         continue;
1162                 pnfs_lseg_dec_and_remove_zero(lseg, head);
1163         }
1164 }
1165
1166 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1167                 const nfs4_stateid *arg_stateid,
1168                 const struct pnfs_layout_range *range,
1169                 const nfs4_stateid *stateid)
1170 {
1171         struct inode *inode = lo->plh_inode;
1172         LIST_HEAD(freeme);
1173
1174         spin_lock(&inode->i_lock);
1175         if (!pnfs_layout_is_valid(lo) ||
1176             !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1177                 goto out_unlock;
1178         if (stateid) {
1179                 u32 seq = be32_to_cpu(arg_stateid->seqid);
1180
1181                 pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1182                 pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1183                 pnfs_set_layout_stateid(lo, stateid, NULL, true);
1184         } else
1185                 pnfs_mark_layout_stateid_invalid(lo, &freeme);
1186 out_unlock:
1187         pnfs_clear_layoutreturn_waitbit(lo);
1188         spin_unlock(&inode->i_lock);
1189         pnfs_free_lseg_list(&freeme);
1190
1191 }
1192
1193 static bool
1194 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1195                 nfs4_stateid *stateid,
1196                 const struct cred **cred,
1197                 enum pnfs_iomode *iomode)
1198 {
1199         /* Serialise LAYOUTGET/LAYOUTRETURN */
1200         if (atomic_read(&lo->plh_outstanding) != 0)
1201                 return false;
1202         if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1203                 return false;
1204         set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1205         pnfs_get_layout_hdr(lo);
1206         nfs4_stateid_copy(stateid, &lo->plh_stateid);
1207         *cred = get_cred(lo->plh_lc_cred);
1208         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1209                 if (lo->plh_return_seq != 0)
1210                         stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1211                 if (iomode != NULL)
1212                         *iomode = lo->plh_return_iomode;
1213                 pnfs_clear_layoutreturn_info(lo);
1214         } else if (iomode != NULL)
1215                 *iomode = IOMODE_ANY;
1216         pnfs_barrier_update(lo, be32_to_cpu(stateid->seqid));
1217         return true;
1218 }
1219
1220 static void
1221 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1222                 struct pnfs_layout_hdr *lo,
1223                 const nfs4_stateid *stateid,
1224                 enum pnfs_iomode iomode)
1225 {
1226         struct inode *inode = lo->plh_inode;
1227
1228         args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1229         args->inode = inode;
1230         args->range.iomode = iomode;
1231         args->range.offset = 0;
1232         args->range.length = NFS4_MAX_UINT64;
1233         args->layout = lo;
1234         nfs4_stateid_copy(&args->stateid, stateid);
1235 }
1236
1237 static int
1238 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1239                        const nfs4_stateid *stateid,
1240                        const struct cred **pcred,
1241                        enum pnfs_iomode iomode,
1242                        bool sync)
1243 {
1244         struct inode *ino = lo->plh_inode;
1245         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1246         struct nfs4_layoutreturn *lrp;
1247         const struct cred *cred = *pcred;
1248         int status = 0;
1249
1250         *pcred = NULL;
1251         lrp = kzalloc(sizeof(*lrp), nfs_io_gfp_mask());
1252         if (unlikely(lrp == NULL)) {
1253                 status = -ENOMEM;
1254                 spin_lock(&ino->i_lock);
1255                 pnfs_clear_layoutreturn_waitbit(lo);
1256                 spin_unlock(&ino->i_lock);
1257                 put_cred(cred);
1258                 pnfs_put_layout_hdr(lo);
1259                 goto out;
1260         }
1261
1262         pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1263         lrp->args.ld_private = &lrp->ld_private;
1264         lrp->clp = NFS_SERVER(ino)->nfs_client;
1265         lrp->cred = cred;
1266         if (ld->prepare_layoutreturn)
1267                 ld->prepare_layoutreturn(&lrp->args);
1268
1269         status = nfs4_proc_layoutreturn(lrp, sync);
1270 out:
1271         dprintk("<-- %s status: %d\n", __func__, status);
1272         return status;
1273 }
1274
1275 static bool
1276 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1277                                 enum pnfs_iomode iomode,
1278                                 u32 seq)
1279 {
1280         struct pnfs_layout_range recall_range = {
1281                 .length = NFS4_MAX_UINT64,
1282                 .iomode = iomode,
1283         };
1284         return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1285                                                &recall_range, seq) != -EBUSY;
1286 }
1287
1288 /* Return true if layoutreturn is needed */
1289 static bool
1290 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1291 {
1292         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1293                 return false;
1294         return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1295                                                lo->plh_return_seq);
1296 }
1297
1298 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1299 {
1300         struct inode *inode= lo->plh_inode;
1301
1302         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1303                 return;
1304         spin_lock(&inode->i_lock);
1305         if (pnfs_layout_need_return(lo)) {
1306                 const struct cred *cred;
1307                 nfs4_stateid stateid;
1308                 enum pnfs_iomode iomode;
1309                 bool send;
1310
1311                 send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1312                 spin_unlock(&inode->i_lock);
1313                 if (send) {
1314                         /* Send an async layoutreturn so we dont deadlock */
1315                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1316                 }
1317         } else
1318                 spin_unlock(&inode->i_lock);
1319 }
1320
1321 /*
1322  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1323  * when the layout segment list is empty.
1324  *
1325  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1326  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1327  * deviceid is marked invalid.
1328  */
1329 int
1330 _pnfs_return_layout(struct inode *ino)
1331 {
1332         struct pnfs_layout_hdr *lo = NULL;
1333         struct nfs_inode *nfsi = NFS_I(ino);
1334         struct pnfs_layout_range range = {
1335                 .iomode         = IOMODE_ANY,
1336                 .offset         = 0,
1337                 .length         = NFS4_MAX_UINT64,
1338         };
1339         LIST_HEAD(tmp_list);
1340         const struct cred *cred;
1341         nfs4_stateid stateid;
1342         int status = 0;
1343         bool send, valid_layout;
1344
1345         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1346
1347         spin_lock(&ino->i_lock);
1348         lo = nfsi->layout;
1349         if (!lo) {
1350                 spin_unlock(&ino->i_lock);
1351                 dprintk("NFS: %s no layout to return\n", __func__);
1352                 goto out;
1353         }
1354         /* Reference matched in nfs4_layoutreturn_release */
1355         pnfs_get_layout_hdr(lo);
1356         /* Is there an outstanding layoutreturn ? */
1357         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1358                 spin_unlock(&ino->i_lock);
1359                 if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1360                                         TASK_UNINTERRUPTIBLE))
1361                         goto out_put_layout_hdr;
1362                 spin_lock(&ino->i_lock);
1363         }
1364         valid_layout = pnfs_layout_is_valid(lo);
1365         pnfs_clear_layoutcommit(ino, &tmp_list);
1366         pnfs_mark_matching_lsegs_return(lo, &tmp_list, &range, 0);
1367
1368         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range)
1369                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1370
1371         /* Don't send a LAYOUTRETURN if list was initially empty */
1372         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1373                         !valid_layout) {
1374                 spin_unlock(&ino->i_lock);
1375                 dprintk("NFS: %s no layout segments to return\n", __func__);
1376                 goto out_wait_layoutreturn;
1377         }
1378
1379         send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1380         spin_unlock(&ino->i_lock);
1381         if (send)
1382                 status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1383 out_wait_layoutreturn:
1384         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1385 out_put_layout_hdr:
1386         pnfs_free_lseg_list(&tmp_list);
1387         pnfs_put_layout_hdr(lo);
1388 out:
1389         dprintk("<-- %s status: %d\n", __func__, status);
1390         return status;
1391 }
1392
1393 int
1394 pnfs_commit_and_return_layout(struct inode *inode)
1395 {
1396         struct pnfs_layout_hdr *lo;
1397         int ret;
1398
1399         spin_lock(&inode->i_lock);
1400         lo = NFS_I(inode)->layout;
1401         if (lo == NULL) {
1402                 spin_unlock(&inode->i_lock);
1403                 return 0;
1404         }
1405         pnfs_get_layout_hdr(lo);
1406         /* Block new layoutgets and read/write to ds */
1407         lo->plh_block_lgets++;
1408         spin_unlock(&inode->i_lock);
1409         filemap_fdatawait(inode->i_mapping);
1410         ret = pnfs_layoutcommit_inode(inode, true);
1411         if (ret == 0)
1412                 ret = _pnfs_return_layout(inode);
1413         spin_lock(&inode->i_lock);
1414         lo->plh_block_lgets--;
1415         spin_unlock(&inode->i_lock);
1416         pnfs_put_layout_hdr(lo);
1417         return ret;
1418 }
1419
1420 bool pnfs_roc(struct inode *ino,
1421                 struct nfs4_layoutreturn_args *args,
1422                 struct nfs4_layoutreturn_res *res,
1423                 const struct cred *cred)
1424 {
1425         struct nfs_inode *nfsi = NFS_I(ino);
1426         struct nfs_open_context *ctx;
1427         struct nfs4_state *state;
1428         struct pnfs_layout_hdr *lo;
1429         struct pnfs_layout_segment *lseg, *next;
1430         const struct cred *lc_cred;
1431         nfs4_stateid stateid;
1432         enum pnfs_iomode iomode = 0;
1433         bool layoutreturn = false, roc = false;
1434         bool skip_read = false;
1435
1436         if (!nfs_have_layout(ino))
1437                 return false;
1438 retry:
1439         rcu_read_lock();
1440         spin_lock(&ino->i_lock);
1441         lo = nfsi->layout;
1442         if (!lo || !pnfs_layout_is_valid(lo) ||
1443             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1444                 lo = NULL;
1445                 goto out_noroc;
1446         }
1447         pnfs_get_layout_hdr(lo);
1448         if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1449                 spin_unlock(&ino->i_lock);
1450                 rcu_read_unlock();
1451                 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1452                                 TASK_UNINTERRUPTIBLE);
1453                 pnfs_put_layout_hdr(lo);
1454                 goto retry;
1455         }
1456
1457         /* no roc if we hold a delegation */
1458         if (nfs4_check_delegation(ino, FMODE_READ)) {
1459                 if (nfs4_check_delegation(ino, FMODE_WRITE))
1460                         goto out_noroc;
1461                 skip_read = true;
1462         }
1463
1464         list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1465                 state = ctx->state;
1466                 if (state == NULL)
1467                         continue;
1468                 /* Don't return layout if there is open file state */
1469                 if (state->state & FMODE_WRITE)
1470                         goto out_noroc;
1471                 if (state->state & FMODE_READ)
1472                         skip_read = true;
1473         }
1474
1475
1476         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1477                 if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1478                         continue;
1479                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1480                 if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1481                         continue;
1482                 /*
1483                  * Note: mark lseg for return so pnfs_layout_remove_lseg
1484                  * doesn't invalidate the layout for us.
1485                  */
1486                 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1487                 if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1488                         continue;
1489                 pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1490         }
1491
1492         if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1493                 goto out_noroc;
1494
1495         /* ROC in two conditions:
1496          * 1. there are ROC lsegs
1497          * 2. we don't send layoutreturn
1498          */
1499         /* lo ref dropped in pnfs_roc_release() */
1500         layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1501         /* If the creds don't match, we can't compound the layoutreturn */
1502         if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1503                 goto out_noroc;
1504
1505         roc = layoutreturn;
1506         pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1507         res->lrs_present = 0;
1508         layoutreturn = false;
1509         put_cred(lc_cred);
1510
1511 out_noroc:
1512         spin_unlock(&ino->i_lock);
1513         rcu_read_unlock();
1514         pnfs_layoutcommit_inode(ino, true);
1515         if (roc) {
1516                 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1517                 if (ld->prepare_layoutreturn)
1518                         ld->prepare_layoutreturn(args);
1519                 pnfs_put_layout_hdr(lo);
1520                 return true;
1521         }
1522         if (layoutreturn)
1523                 pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1524         pnfs_put_layout_hdr(lo);
1525         return false;
1526 }
1527
1528 int pnfs_roc_done(struct rpc_task *task, struct nfs4_layoutreturn_args **argpp,
1529                   struct nfs4_layoutreturn_res **respp, int *ret)
1530 {
1531         struct nfs4_layoutreturn_args *arg = *argpp;
1532         int retval = -EAGAIN;
1533
1534         if (!arg)
1535                 return 0;
1536         /* Handle Layoutreturn errors */
1537         switch (*ret) {
1538         case 0:
1539                 retval = 0;
1540                 break;
1541         case -NFS4ERR_NOMATCHING_LAYOUT:
1542                 /* Was there an RPC level error? If not, retry */
1543                 if (task->tk_rpc_status == 0)
1544                         break;
1545                 /* If the call was not sent, let caller handle it */
1546                 if (!RPC_WAS_SENT(task))
1547                         return 0;
1548                 /*
1549                  * Otherwise, assume the call succeeded and
1550                  * that we need to release the layout
1551                  */
1552                 *ret = 0;
1553                 (*respp)->lrs_present = 0;
1554                 retval = 0;
1555                 break;
1556         case -NFS4ERR_DELAY:
1557                 /* Let the caller handle the retry */
1558                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1559                 return 0;
1560         case -NFS4ERR_OLD_STATEID:
1561                 if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1562                                                      &arg->range, arg->inode))
1563                         break;
1564                 *ret = -NFS4ERR_NOMATCHING_LAYOUT;
1565                 return -EAGAIN;
1566         }
1567         *argpp = NULL;
1568         *respp = NULL;
1569         return retval;
1570 }
1571
1572 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1573                 struct nfs4_layoutreturn_res *res,
1574                 int ret)
1575 {
1576         struct pnfs_layout_hdr *lo = args->layout;
1577         struct inode *inode = args->inode;
1578         const nfs4_stateid *res_stateid = NULL;
1579         struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1580
1581         switch (ret) {
1582         case -NFS4ERR_NOMATCHING_LAYOUT:
1583                 spin_lock(&inode->i_lock);
1584                 if (pnfs_layout_is_valid(lo) &&
1585                     nfs4_stateid_match_other(&args->stateid, &lo->plh_stateid))
1586                         pnfs_set_plh_return_info(lo, args->range.iomode, 0);
1587                 pnfs_clear_layoutreturn_waitbit(lo);
1588                 spin_unlock(&inode->i_lock);
1589                 break;
1590         case 0:
1591                 if (res->lrs_present)
1592                         res_stateid = &res->stateid;
1593                 fallthrough;
1594         default:
1595                 pnfs_layoutreturn_free_lsegs(lo, &args->stateid, &args->range,
1596                                              res_stateid);
1597         }
1598         trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1599         if (ld_private && ld_private->ops && ld_private->ops->free)
1600                 ld_private->ops->free(ld_private);
1601         pnfs_put_layout_hdr(lo);
1602 }
1603
1604 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1605 {
1606         struct nfs_inode *nfsi = NFS_I(ino);
1607         struct pnfs_layout_hdr *lo;
1608         bool sleep = false;
1609
1610         /* we might not have grabbed lo reference. so need to check under
1611          * i_lock */
1612         spin_lock(&ino->i_lock);
1613         lo = nfsi->layout;
1614         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1615                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1616                 sleep = true;
1617         }
1618         spin_unlock(&ino->i_lock);
1619         return sleep;
1620 }
1621
1622 /*
1623  * Compare two layout segments for sorting into layout cache.
1624  * We want to preferentially return RW over RO layouts, so ensure those
1625  * are seen first.
1626  */
1627 static s64
1628 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1629            const struct pnfs_layout_range *l2)
1630 {
1631         s64 d;
1632
1633         /* high offset > low offset */
1634         d = l1->offset - l2->offset;
1635         if (d)
1636                 return d;
1637
1638         /* short length > long length */
1639         d = l2->length - l1->length;
1640         if (d)
1641                 return d;
1642
1643         /* read > read/write */
1644         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1645 }
1646
1647 static bool
1648 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1649                 const struct pnfs_layout_range *l2)
1650 {
1651         return pnfs_lseg_range_cmp(l1, l2) > 0;
1652 }
1653
1654 static bool
1655 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1656                 struct pnfs_layout_segment *old)
1657 {
1658         return false;
1659 }
1660
1661 void
1662 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1663                    struct pnfs_layout_segment *lseg,
1664                    bool (*is_after)(const struct pnfs_layout_range *,
1665                            const struct pnfs_layout_range *),
1666                    bool (*do_merge)(struct pnfs_layout_segment *,
1667                            struct pnfs_layout_segment *),
1668                    struct list_head *free_me)
1669 {
1670         struct pnfs_layout_segment *lp, *tmp;
1671
1672         dprintk("%s:Begin\n", __func__);
1673
1674         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1675                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1676                         continue;
1677                 if (do_merge(lseg, lp)) {
1678                         mark_lseg_invalid(lp, free_me);
1679                         continue;
1680                 }
1681                 if (is_after(&lseg->pls_range, &lp->pls_range))
1682                         continue;
1683                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1684                 dprintk("%s: inserted lseg %p "
1685                         "iomode %d offset %llu length %llu before "
1686                         "lp %p iomode %d offset %llu length %llu\n",
1687                         __func__, lseg, lseg->pls_range.iomode,
1688                         lseg->pls_range.offset, lseg->pls_range.length,
1689                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1690                         lp->pls_range.length);
1691                 goto out;
1692         }
1693         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1694         dprintk("%s: inserted lseg %p "
1695                 "iomode %d offset %llu length %llu at tail\n",
1696                 __func__, lseg, lseg->pls_range.iomode,
1697                 lseg->pls_range.offset, lseg->pls_range.length);
1698 out:
1699         pnfs_get_layout_hdr(lo);
1700
1701         dprintk("%s:Return\n", __func__);
1702 }
1703 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1704
1705 static void
1706 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1707                    struct pnfs_layout_segment *lseg,
1708                    struct list_head *free_me)
1709 {
1710         struct inode *inode = lo->plh_inode;
1711         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1712
1713         if (ld->add_lseg != NULL)
1714                 ld->add_lseg(lo, lseg, free_me);
1715         else
1716                 pnfs_generic_layout_insert_lseg(lo, lseg,
1717                                 pnfs_lseg_range_is_after,
1718                                 pnfs_lseg_no_merge,
1719                                 free_me);
1720 }
1721
1722 static struct pnfs_layout_hdr *
1723 alloc_init_layout_hdr(struct inode *ino,
1724                       struct nfs_open_context *ctx,
1725                       gfp_t gfp_flags)
1726 {
1727         struct pnfs_layout_hdr *lo;
1728
1729         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1730         if (!lo)
1731                 return NULL;
1732         refcount_set(&lo->plh_refcount, 1);
1733         INIT_LIST_HEAD(&lo->plh_layouts);
1734         INIT_LIST_HEAD(&lo->plh_segs);
1735         INIT_LIST_HEAD(&lo->plh_return_segs);
1736         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1737         lo->plh_inode = ino;
1738         lo->plh_lc_cred = get_cred(ctx->cred);
1739         lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1740         return lo;
1741 }
1742
1743 static struct pnfs_layout_hdr *
1744 pnfs_find_alloc_layout(struct inode *ino,
1745                        struct nfs_open_context *ctx,
1746                        gfp_t gfp_flags)
1747         __releases(&ino->i_lock)
1748         __acquires(&ino->i_lock)
1749 {
1750         struct nfs_inode *nfsi = NFS_I(ino);
1751         struct pnfs_layout_hdr *new = NULL;
1752
1753         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1754
1755         if (nfsi->layout != NULL)
1756                 goto out_existing;
1757         spin_unlock(&ino->i_lock);
1758         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1759         spin_lock(&ino->i_lock);
1760
1761         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1762                 nfsi->layout = new;
1763                 return new;
1764         } else if (new != NULL)
1765                 pnfs_free_layout_hdr(new);
1766 out_existing:
1767         pnfs_get_layout_hdr(nfsi->layout);
1768         return nfsi->layout;
1769 }
1770
1771 /*
1772  * iomode matching rules:
1773  * iomode       lseg    strict match
1774  *                      iomode
1775  * -----        -----   ------ -----
1776  * ANY          READ    N/A    true
1777  * ANY          RW      N/A    true
1778  * RW           READ    N/A    false
1779  * RW           RW      N/A    true
1780  * READ         READ    N/A    true
1781  * READ         RW      true   false
1782  * READ         RW      false  true
1783  */
1784 static bool
1785 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1786                  const struct pnfs_layout_range *range,
1787                  bool strict_iomode)
1788 {
1789         struct pnfs_layout_range range1;
1790
1791         if ((range->iomode == IOMODE_RW &&
1792              ls_range->iomode != IOMODE_RW) ||
1793             (range->iomode != ls_range->iomode &&
1794              strict_iomode) ||
1795             !pnfs_lseg_range_intersecting(ls_range, range))
1796                 return false;
1797
1798         /* range1 covers only the first byte in the range */
1799         range1 = *range;
1800         range1.length = 1;
1801         return pnfs_lseg_range_contained(ls_range, &range1);
1802 }
1803
1804 /*
1805  * lookup range in layout
1806  */
1807 static struct pnfs_layout_segment *
1808 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1809                 struct pnfs_layout_range *range,
1810                 bool strict_iomode)
1811 {
1812         struct pnfs_layout_segment *lseg, *ret = NULL;
1813
1814         dprintk("%s:Begin\n", __func__);
1815
1816         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1817                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1818                     pnfs_lseg_range_match(&lseg->pls_range, range,
1819                                           strict_iomode)) {
1820                         ret = pnfs_get_lseg(lseg);
1821                         break;
1822                 }
1823         }
1824
1825         dprintk("%s:Return lseg %p ref %d\n",
1826                 __func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1827         return ret;
1828 }
1829
1830 /*
1831  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1832  * to the MDS or over pNFS
1833  *
1834  * The nfs_inode read_io and write_io fields are cumulative counters reset
1835  * when there are no layout segments. Note that in pnfs_update_layout iomode
1836  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1837  * WRITE request.
1838  *
1839  * A return of true means use MDS I/O.
1840  *
1841  * From rfc 5661:
1842  * If a file's size is smaller than the file size threshold, data accesses
1843  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1844  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1845  * server.  If both file size and I/O size are provided, the client SHOULD
1846  * reach or exceed  both thresholds before sending its read or write
1847  * requests to the data server.
1848  */
1849 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1850                                      struct inode *ino, int iomode)
1851 {
1852         struct nfs4_threshold *t = ctx->mdsthreshold;
1853         struct nfs_inode *nfsi = NFS_I(ino);
1854         loff_t fsize = i_size_read(ino);
1855         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1856
1857         if (t == NULL)
1858                 return ret;
1859
1860         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1861                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1862
1863         switch (iomode) {
1864         case IOMODE_READ:
1865                 if (t->bm & THRESHOLD_RD) {
1866                         dprintk("%s fsize %llu\n", __func__, fsize);
1867                         size_set = true;
1868                         if (fsize < t->rd_sz)
1869                                 size = true;
1870                 }
1871                 if (t->bm & THRESHOLD_RD_IO) {
1872                         dprintk("%s nfsi->read_io %llu\n", __func__,
1873                                 nfsi->read_io);
1874                         io_set = true;
1875                         if (nfsi->read_io < t->rd_io_sz)
1876                                 io = true;
1877                 }
1878                 break;
1879         case IOMODE_RW:
1880                 if (t->bm & THRESHOLD_WR) {
1881                         dprintk("%s fsize %llu\n", __func__, fsize);
1882                         size_set = true;
1883                         if (fsize < t->wr_sz)
1884                                 size = true;
1885                 }
1886                 if (t->bm & THRESHOLD_WR_IO) {
1887                         dprintk("%s nfsi->write_io %llu\n", __func__,
1888                                 nfsi->write_io);
1889                         io_set = true;
1890                         if (nfsi->write_io < t->wr_io_sz)
1891                                 io = true;
1892                 }
1893                 break;
1894         }
1895         if (size_set && io_set) {
1896                 if (size && io)
1897                         ret = true;
1898         } else if (size || io)
1899                 ret = true;
1900
1901         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1902         return ret;
1903 }
1904
1905 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1906 {
1907         /*
1908          * send layoutcommit as it can hold up layoutreturn due to lseg
1909          * reference
1910          */
1911         pnfs_layoutcommit_inode(lo->plh_inode, false);
1912         return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1913                                    nfs_wait_bit_killable,
1914                                    TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
1915 }
1916
1917 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1918 {
1919         atomic_inc(&lo->plh_outstanding);
1920 }
1921
1922 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1923 {
1924         if (atomic_dec_and_test(&lo->plh_outstanding) &&
1925             test_and_clear_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags))
1926                 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN);
1927 }
1928
1929 static bool pnfs_is_first_layoutget(struct pnfs_layout_hdr *lo)
1930 {
1931         return test_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags);
1932 }
1933
1934 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1935 {
1936         unsigned long *bitlock = &lo->plh_flags;
1937
1938         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1939         smp_mb__after_atomic();
1940         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1941 }
1942
1943 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1944                                 struct nfs_server *server)
1945 {
1946         if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1947                 struct nfs_client *clp = server->nfs_client;
1948
1949                 /* The lo must be on the clp list if there is any
1950                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1951                  */
1952                 spin_lock(&clp->cl_lock);
1953                 list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1954                 spin_unlock(&clp->cl_lock);
1955         }
1956 }
1957
1958 /*
1959  * Layout segment is retreived from the server if not cached.
1960  * The appropriate layout segment is referenced and returned to the caller.
1961  */
1962 struct pnfs_layout_segment *
1963 pnfs_update_layout(struct inode *ino,
1964                    struct nfs_open_context *ctx,
1965                    loff_t pos,
1966                    u64 count,
1967                    enum pnfs_iomode iomode,
1968                    bool strict_iomode,
1969                    gfp_t gfp_flags)
1970 {
1971         struct pnfs_layout_range arg = {
1972                 .iomode = iomode,
1973                 .offset = pos,
1974                 .length = count,
1975         };
1976         unsigned pg_offset;
1977         struct nfs_server *server = NFS_SERVER(ino);
1978         struct nfs_client *clp = server->nfs_client;
1979         struct pnfs_layout_hdr *lo = NULL;
1980         struct pnfs_layout_segment *lseg = NULL;
1981         struct nfs4_layoutget *lgp;
1982         nfs4_stateid stateid;
1983         long timeout = 0;
1984         unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1985         bool first;
1986
1987         if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1988                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1989                                  PNFS_UPDATE_LAYOUT_NO_PNFS);
1990                 goto out;
1991         }
1992
1993         if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1994                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1995                                  PNFS_UPDATE_LAYOUT_MDSTHRESH);
1996                 goto out;
1997         }
1998
1999 lookup_again:
2000         lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
2001         if (IS_ERR(lseg))
2002                 goto out;
2003         first = false;
2004         spin_lock(&ino->i_lock);
2005         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
2006         if (lo == NULL) {
2007                 spin_unlock(&ino->i_lock);
2008                 lseg = ERR_PTR(-ENOMEM);
2009                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2010                                  PNFS_UPDATE_LAYOUT_NOMEM);
2011                 goto out;
2012         }
2013
2014         /* Do we even need to bother with this? */
2015         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
2016                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2017                                  PNFS_UPDATE_LAYOUT_BULK_RECALL);
2018                 dprintk("%s matches recall, use MDS\n", __func__);
2019                 goto out_unlock;
2020         }
2021
2022         /* if LAYOUTGET already failed once we don't try again */
2023         if (pnfs_layout_io_test_failed(lo, iomode)) {
2024                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2025                                  PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
2026                 goto out_unlock;
2027         }
2028
2029         /*
2030          * If the layout segment list is empty, but there are outstanding
2031          * layoutget calls, then they might be subject to a layoutrecall.
2032          */
2033         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2034             atomic_read(&lo->plh_outstanding) != 0) {
2035                 spin_unlock(&ino->i_lock);
2036                 lseg = ERR_PTR(wait_on_bit(&lo->plh_flags, NFS_LAYOUT_DRAIN,
2037                                            TASK_KILLABLE));
2038                 if (IS_ERR(lseg))
2039                         goto out_put_layout_hdr;
2040                 pnfs_put_layout_hdr(lo);
2041                 goto lookup_again;
2042         }
2043
2044         /*
2045          * Because we free lsegs when sending LAYOUTRETURN, we need to wait
2046          * for LAYOUTRETURN.
2047          */
2048         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2049                 spin_unlock(&ino->i_lock);
2050                 dprintk("%s wait for layoutreturn\n", __func__);
2051                 lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2052                 if (!IS_ERR(lseg)) {
2053                         pnfs_put_layout_hdr(lo);
2054                         dprintk("%s retrying\n", __func__);
2055                         trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2056                                                  lseg,
2057                                                  PNFS_UPDATE_LAYOUT_RETRY);
2058                         goto lookup_again;
2059                 }
2060                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2061                                          PNFS_UPDATE_LAYOUT_RETURN);
2062                 goto out_put_layout_hdr;
2063         }
2064
2065         lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
2066         if (lseg) {
2067                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2068                                 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
2069                 goto out_unlock;
2070         }
2071
2072         /*
2073          * Choose a stateid for the LAYOUTGET. If we don't have a layout
2074          * stateid, or it has been invalidated, then we must use the open
2075          * stateid.
2076          */
2077         if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2078                 int status;
2079
2080                 /*
2081                  * The first layoutget for the file. Need to serialize per
2082                  * RFC 5661 Errata 3208.
2083                  */
2084                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2085                                      &lo->plh_flags)) {
2086                         spin_unlock(&ino->i_lock);
2087                         lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2088                                                 NFS_LAYOUT_FIRST_LAYOUTGET,
2089                                                 TASK_KILLABLE));
2090                         if (IS_ERR(lseg))
2091                                 goto out_put_layout_hdr;
2092                         pnfs_put_layout_hdr(lo);
2093                         dprintk("%s retrying\n", __func__);
2094                         goto lookup_again;
2095                 }
2096
2097                 spin_unlock(&ino->i_lock);
2098                 first = true;
2099                 status = nfs4_select_rw_stateid(ctx->state,
2100                                         iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2101                                         NULL, &stateid, NULL);
2102                 if (status != 0) {
2103                         lseg = ERR_PTR(status);
2104                         trace_pnfs_update_layout(ino, pos, count,
2105                                         iomode, lo, lseg,
2106                                         PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2107                         nfs4_schedule_stateid_recovery(server, ctx->state);
2108                         pnfs_clear_first_layoutget(lo);
2109                         pnfs_put_layout_hdr(lo);
2110                         goto lookup_again;
2111                 }
2112                 spin_lock(&ino->i_lock);
2113         } else {
2114                 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2115         }
2116
2117         if (pnfs_layoutgets_blocked(lo)) {
2118                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2119                                 PNFS_UPDATE_LAYOUT_BLOCKED);
2120                 goto out_unlock;
2121         }
2122         nfs_layoutget_begin(lo);
2123         spin_unlock(&ino->i_lock);
2124
2125         _add_to_server_list(lo, server);
2126
2127         pg_offset = arg.offset & ~PAGE_MASK;
2128         if (pg_offset) {
2129                 arg.offset -= pg_offset;
2130                 arg.length += pg_offset;
2131         }
2132         if (arg.length != NFS4_MAX_UINT64)
2133                 arg.length = PAGE_ALIGN(arg.length);
2134
2135         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2136         if (!lgp) {
2137                 lseg = ERR_PTR(-ENOMEM);
2138                 trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2139                                          PNFS_UPDATE_LAYOUT_NOMEM);
2140                 nfs_layoutget_end(lo);
2141                 goto out_put_layout_hdr;
2142         }
2143
2144         lgp->lo = lo;
2145         pnfs_get_layout_hdr(lo);
2146
2147         lseg = nfs4_proc_layoutget(lgp, &timeout);
2148         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2149                                  PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2150         nfs_layoutget_end(lo);
2151         if (IS_ERR(lseg)) {
2152                 switch(PTR_ERR(lseg)) {
2153                 case -EBUSY:
2154                         if (time_after(jiffies, giveup))
2155                                 lseg = NULL;
2156                         break;
2157                 case -ERECALLCONFLICT:
2158                 case -EAGAIN:
2159                         break;
2160                 case -ENODATA:
2161                         /* The server returned NFS4ERR_LAYOUTUNAVAILABLE */
2162                         pnfs_layout_set_fail_bit(
2163                                 lo, pnfs_iomode_to_fail_bit(iomode));
2164                         lseg = NULL;
2165                         goto out_put_layout_hdr;
2166                 default:
2167                         if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2168                                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2169                                 lseg = NULL;
2170                         }
2171                         goto out_put_layout_hdr;
2172                 }
2173                 if (lseg) {
2174                         if (first)
2175                                 pnfs_clear_first_layoutget(lo);
2176                         trace_pnfs_update_layout(ino, pos, count,
2177                                 iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2178                         pnfs_put_layout_hdr(lo);
2179                         goto lookup_again;
2180                 }
2181         } else {
2182                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2183         }
2184
2185 out_put_layout_hdr:
2186         if (first)
2187                 pnfs_clear_first_layoutget(lo);
2188         trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2189                                  PNFS_UPDATE_LAYOUT_EXIT);
2190         pnfs_put_layout_hdr(lo);
2191 out:
2192         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2193                         "(%s, offset: %llu, length: %llu)\n",
2194                         __func__, ino->i_sb->s_id,
2195                         (unsigned long long)NFS_FILEID(ino),
2196                         IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2197                         iomode==IOMODE_RW ?  "read/write" : "read-only",
2198                         (unsigned long long)pos,
2199                         (unsigned long long)count);
2200         return lseg;
2201 out_unlock:
2202         spin_unlock(&ino->i_lock);
2203         goto out_put_layout_hdr;
2204 }
2205 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2206
2207 static bool
2208 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2209 {
2210         switch (range->iomode) {
2211         case IOMODE_READ:
2212         case IOMODE_RW:
2213                 break;
2214         default:
2215                 return false;
2216         }
2217         if (range->offset == NFS4_MAX_UINT64)
2218                 return false;
2219         if (range->length == 0)
2220                 return false;
2221         if (range->length != NFS4_MAX_UINT64 &&
2222             range->length > NFS4_MAX_UINT64 - range->offset)
2223                 return false;
2224         return true;
2225 }
2226
2227 static struct pnfs_layout_hdr *
2228 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2229 {
2230         struct pnfs_layout_hdr *lo;
2231
2232         spin_lock(&ino->i_lock);
2233         lo = pnfs_find_alloc_layout(ino, ctx, nfs_io_gfp_mask());
2234         if (!lo)
2235                 goto out_unlock;
2236         if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2237                 goto out_unlock;
2238         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2239                 goto out_unlock;
2240         if (pnfs_layoutgets_blocked(lo))
2241                 goto out_unlock;
2242         if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2243                 goto out_unlock;
2244         nfs_layoutget_begin(lo);
2245         spin_unlock(&ino->i_lock);
2246         _add_to_server_list(lo, NFS_SERVER(ino));
2247         return lo;
2248
2249 out_unlock:
2250         spin_unlock(&ino->i_lock);
2251         pnfs_put_layout_hdr(lo);
2252         return NULL;
2253 }
2254
2255 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2256                                      struct nfs_open_context *ctx)
2257 {
2258         struct inode *ino = data->dentry->d_inode;
2259         struct pnfs_layout_range rng = {
2260                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2261                           IOMODE_RW: IOMODE_READ,
2262                 .offset = 0,
2263                 .length = NFS4_MAX_UINT64,
2264         };
2265         struct nfs4_layoutget *lgp;
2266         struct pnfs_layout_hdr *lo;
2267
2268         /* Heuristic: don't send layoutget if we have cached data */
2269         if (rng.iomode == IOMODE_READ &&
2270            (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2271                 return;
2272
2273         lo = _pnfs_grab_empty_layout(ino, ctx);
2274         if (!lo)
2275                 return;
2276         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2277                                              nfs_io_gfp_mask());
2278         if (!lgp) {
2279                 pnfs_clear_first_layoutget(lo);
2280                 nfs_layoutget_end(lo);
2281                 pnfs_put_layout_hdr(lo);
2282                 return;
2283         }
2284         lgp->lo = lo;
2285         data->lgp = lgp;
2286         data->o_arg.lg_args = &lgp->args;
2287         data->o_res.lg_res = &lgp->res;
2288 }
2289
2290 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2291                                      struct nfs_open_context *ctx)
2292 {
2293         struct inode *ino = data->dentry->d_inode;
2294         struct pnfs_layout_range rng = {
2295                 .iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2296                           IOMODE_RW: IOMODE_READ,
2297                 .offset = 0,
2298                 .length = NFS4_MAX_UINT64,
2299         };
2300         struct nfs4_layoutget *lgp;
2301
2302         lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid, &rng,
2303                                              nfs_io_gfp_mask());
2304         if (!lgp)
2305                 return;
2306         data->lgp = lgp;
2307         data->o_arg.lg_args = &lgp->args;
2308         data->o_res.lg_res = &lgp->res;
2309 }
2310
2311 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2312                          struct nfs_open_context *ctx)
2313 {
2314         struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2315
2316         if (!(pnfs_enabled_sb(server) &&
2317               server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2318                 return;
2319         /* Could check on max_ops, but currently hardcoded high enough */
2320         if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2321                 return;
2322         if (data->lgp)
2323                 return;
2324         if (data->state)
2325                 _lgopen_prepare_attached(data, ctx);
2326         else
2327                 _lgopen_prepare_floating(data, ctx);
2328 }
2329
2330 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2331                        struct nfs_open_context *ctx)
2332 {
2333         struct pnfs_layout_hdr *lo;
2334         struct pnfs_layout_segment *lseg;
2335         struct nfs_server *srv = NFS_SERVER(ino);
2336         u32 iomode;
2337
2338         if (!lgp)
2339                 return;
2340         dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2341         if (lgp->res.status) {
2342                 switch (lgp->res.status) {
2343                 default:
2344                         break;
2345                 /*
2346                  * Halt lgopen attempts if the server doesn't recognise
2347                  * the "current stateid" value, the layout type, or the
2348                  * layoutget operation as being valid.
2349                  * Also if it complains about too many ops in the compound
2350                  * or of the request/reply being too big.
2351                  */
2352                 case -NFS4ERR_BAD_STATEID:
2353                 case -NFS4ERR_NOTSUPP:
2354                 case -NFS4ERR_REP_TOO_BIG:
2355                 case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2356                 case -NFS4ERR_REQ_TOO_BIG:
2357                 case -NFS4ERR_TOO_MANY_OPS:
2358                 case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2359                         srv->caps &= ~NFS_CAP_LGOPEN;
2360                 }
2361                 return;
2362         }
2363         if (!lgp->lo) {
2364                 lo = _pnfs_grab_empty_layout(ino, ctx);
2365                 if (!lo)
2366                         return;
2367                 lgp->lo = lo;
2368         } else
2369                 lo = lgp->lo;
2370
2371         lseg = pnfs_layout_process(lgp);
2372         if (!IS_ERR(lseg)) {
2373                 iomode = lgp->args.range.iomode;
2374                 pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2375                 pnfs_put_lseg(lseg);
2376         }
2377 }
2378
2379 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2380 {
2381         if (lgp != NULL) {
2382                 if (lgp->lo) {
2383                         pnfs_clear_first_layoutget(lgp->lo);
2384                         nfs_layoutget_end(lgp->lo);
2385                 }
2386                 pnfs_layoutget_free(lgp);
2387         }
2388 }
2389
2390 struct pnfs_layout_segment *
2391 pnfs_layout_process(struct nfs4_layoutget *lgp)
2392 {
2393         struct pnfs_layout_hdr *lo = lgp->lo;
2394         struct nfs4_layoutget_res *res = &lgp->res;
2395         struct pnfs_layout_segment *lseg;
2396         struct inode *ino = lo->plh_inode;
2397         LIST_HEAD(free_me);
2398
2399         if (!pnfs_sanity_check_layout_range(&res->range))
2400                 return ERR_PTR(-EINVAL);
2401
2402         /* Inject layout blob into I/O device driver */
2403         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2404         if (IS_ERR_OR_NULL(lseg)) {
2405                 if (!lseg)
2406                         lseg = ERR_PTR(-ENOMEM);
2407
2408                 dprintk("%s: Could not allocate layout: error %ld\n",
2409                        __func__, PTR_ERR(lseg));
2410                 return lseg;
2411         }
2412
2413         pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2414
2415         spin_lock(&ino->i_lock);
2416         if (pnfs_layoutgets_blocked(lo)) {
2417                 dprintk("%s forget reply due to state\n", __func__);
2418                 goto out_forget;
2419         }
2420
2421         if (test_bit(NFS_LAYOUT_DRAIN, &lo->plh_flags) &&
2422             !pnfs_is_first_layoutget(lo))
2423                 goto out_forget;
2424
2425         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2426                 /* existing state ID, make sure the sequence number matches. */
2427                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2428                         if (!pnfs_layout_is_valid(lo))
2429                                 lo->plh_barrier = 0;
2430                         dprintk("%s forget reply due to sequence\n", __func__);
2431                         goto out_forget;
2432                 }
2433                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2434         } else if (pnfs_layout_is_valid(lo)) {
2435                 /*
2436                  * We got an entirely new state ID.  Mark all segments for the
2437                  * inode invalid, and retry the layoutget
2438                  */
2439                 struct pnfs_layout_range range = {
2440                         .iomode = IOMODE_ANY,
2441                         .length = NFS4_MAX_UINT64,
2442                 };
2443                 pnfs_mark_matching_lsegs_return(lo, &free_me, &range, 0);
2444                 goto out_forget;
2445         } else {
2446                 /* We have a completely new layout */
2447                 pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2448         }
2449
2450         pnfs_get_lseg(lseg);
2451         pnfs_layout_insert_lseg(lo, lseg, &free_me);
2452
2453
2454         if (res->return_on_close)
2455                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2456
2457         spin_unlock(&ino->i_lock);
2458         pnfs_free_lseg_list(&free_me);
2459         return lseg;
2460
2461 out_forget:
2462         spin_unlock(&ino->i_lock);
2463         lseg->pls_layout = lo;
2464         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2465         return ERR_PTR(-EAGAIN);
2466 }
2467
2468 /**
2469  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2470  * @lo: pointer to layout header
2471  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2472  * @return_range: describe layout segment ranges to be returned
2473  * @seq: stateid seqid to match
2474  *
2475  * This function is mainly intended for use by layoutrecall. It attempts
2476  * to free the layout segment immediately, or else to mark it for return
2477  * as soon as its reference count drops to zero.
2478  *
2479  * Returns
2480  * - 0: a layoutreturn needs to be scheduled.
2481  * - EBUSY: there are layout segment that are still in use.
2482  * - ENOENT: there are no layout segments that need to be returned.
2483  */
2484 int
2485 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2486                                 struct list_head *tmp_list,
2487                                 const struct pnfs_layout_range *return_range,
2488                                 u32 seq)
2489 {
2490         struct pnfs_layout_segment *lseg, *next;
2491         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
2492         int remaining = 0;
2493
2494         dprintk("%s:Begin lo %p\n", __func__, lo);
2495
2496         assert_spin_locked(&lo->plh_inode->i_lock);
2497
2498         if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2499                 tmp_list = &lo->plh_return_segs;
2500
2501         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2502                 if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2503                         dprintk("%s: marking lseg %p iomode %d "
2504                                 "offset %llu length %llu\n", __func__,
2505                                 lseg, lseg->pls_range.iomode,
2506                                 lseg->pls_range.offset,
2507                                 lseg->pls_range.length);
2508                         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2509                                 tmp_list = &lo->plh_return_segs;
2510                         if (mark_lseg_invalid(lseg, tmp_list))
2511                                 continue;
2512                         remaining++;
2513                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2514                         pnfs_lseg_cancel_io(server, lseg);
2515                 }
2516
2517         if (remaining) {
2518                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2519                 return -EBUSY;
2520         }
2521
2522         if (!list_empty(&lo->plh_return_segs)) {
2523                 pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2524                 return 0;
2525         }
2526
2527         return -ENOENT;
2528 }
2529
2530 static void
2531 pnfs_mark_layout_for_return(struct inode *inode,
2532                             const struct pnfs_layout_range *range)
2533 {
2534         struct pnfs_layout_hdr *lo;
2535         bool return_now = false;
2536
2537         spin_lock(&inode->i_lock);
2538         lo = NFS_I(inode)->layout;
2539         if (!pnfs_layout_is_valid(lo)) {
2540                 spin_unlock(&inode->i_lock);
2541                 return;
2542         }
2543         pnfs_set_plh_return_info(lo, range->iomode, 0);
2544         /*
2545          * mark all matching lsegs so that we are sure to have no live
2546          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2547          * for how it works.
2548          */
2549         if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2550                 const struct cred *cred;
2551                 nfs4_stateid stateid;
2552                 enum pnfs_iomode iomode;
2553
2554                 return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2555                 spin_unlock(&inode->i_lock);
2556                 if (return_now)
2557                         pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2558         } else {
2559                 spin_unlock(&inode->i_lock);
2560                 nfs_commit_inode(inode, 0);
2561         }
2562 }
2563
2564 void pnfs_error_mark_layout_for_return(struct inode *inode,
2565                                        struct pnfs_layout_segment *lseg)
2566 {
2567         struct pnfs_layout_range range = {
2568                 .iomode = lseg->pls_range.iomode,
2569                 .offset = 0,
2570                 .length = NFS4_MAX_UINT64,
2571         };
2572
2573         pnfs_mark_layout_for_return(inode, &range);
2574 }
2575 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2576
2577 static bool
2578 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2579 {
2580         return pnfs_layout_is_valid(lo) &&
2581                 !test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2582                 !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2583 }
2584
2585 static struct pnfs_layout_segment *
2586 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2587                      const struct pnfs_layout_range *range,
2588                      enum pnfs_iomode iomode)
2589 {
2590         struct pnfs_layout_segment *lseg;
2591
2592         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2593                 if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2594                         continue;
2595                 if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2596                         continue;
2597                 if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2598                         continue;
2599                 if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2600                         return lseg;
2601         }
2602         return NULL;
2603 }
2604
2605 /* Find open file states whose mode matches that of the range */
2606 static bool
2607 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2608                                  const struct pnfs_layout_range *range)
2609 {
2610         struct list_head *head;
2611         struct nfs_open_context *ctx;
2612         fmode_t mode = 0;
2613
2614         if (!pnfs_layout_can_be_returned(lo) ||
2615             !pnfs_find_first_lseg(lo, range, range->iomode))
2616                 return false;
2617
2618         head = &NFS_I(lo->plh_inode)->open_files;
2619         list_for_each_entry_rcu(ctx, head, list) {
2620                 if (ctx->state)
2621                         mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2622         }
2623
2624         switch (range->iomode) {
2625         default:
2626                 break;
2627         case IOMODE_READ:
2628                 mode &= ~FMODE_WRITE;
2629                 break;
2630         case IOMODE_RW:
2631                 if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2632                         mode &= ~FMODE_READ;
2633         }
2634         return mode == 0;
2635 }
2636
2637 static int pnfs_layout_return_unused_byserver(struct nfs_server *server,
2638                                               void *data)
2639 {
2640         const struct pnfs_layout_range *range = data;
2641         const struct cred *cred;
2642         struct pnfs_layout_hdr *lo;
2643         struct inode *inode;
2644         nfs4_stateid stateid;
2645         enum pnfs_iomode iomode;
2646
2647 restart:
2648         rcu_read_lock();
2649         list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2650                 inode = lo->plh_inode;
2651                 if (!inode || !pnfs_layout_can_be_returned(lo) ||
2652                     test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2653                         continue;
2654                 spin_lock(&inode->i_lock);
2655                 if (!lo->plh_inode ||
2656                     !pnfs_should_return_unused_layout(lo, range)) {
2657                         spin_unlock(&inode->i_lock);
2658                         continue;
2659                 }
2660                 pnfs_get_layout_hdr(lo);
2661                 pnfs_set_plh_return_info(lo, range->iomode, 0);
2662                 if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
2663                                                     range, 0) != 0 ||
2664                     !pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode)) {
2665                         spin_unlock(&inode->i_lock);
2666                         rcu_read_unlock();
2667                         pnfs_put_layout_hdr(lo);
2668                         cond_resched();
2669                         goto restart;
2670                 }
2671                 spin_unlock(&inode->i_lock);
2672                 rcu_read_unlock();
2673                 pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2674                 pnfs_put_layout_hdr(lo);
2675                 cond_resched();
2676                 goto restart;
2677         }
2678         rcu_read_unlock();
2679         return 0;
2680 }
2681
2682 void
2683 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2684                                  enum pnfs_iomode iomode)
2685 {
2686         struct pnfs_layout_range range = {
2687                 .iomode = iomode,
2688                 .offset = 0,
2689                 .length = NFS4_MAX_UINT64,
2690         };
2691
2692         nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2693                         &range);
2694 }
2695
2696 void
2697 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2698 {
2699         if (pgio->pg_lseg == NULL ||
2700             test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2701                 return;
2702         pnfs_put_lseg(pgio->pg_lseg);
2703         pgio->pg_lseg = NULL;
2704 }
2705 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2706
2707 /*
2708  * Check for any intersection between the request and the pgio->pg_lseg,
2709  * and if none, put this pgio->pg_lseg away.
2710  */
2711 void
2712 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2713 {
2714         if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2715                 pnfs_put_lseg(pgio->pg_lseg);
2716                 pgio->pg_lseg = NULL;
2717         }
2718 }
2719 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2720
2721 void
2722 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2723 {
2724         u64 rd_size;
2725
2726         pnfs_generic_pg_check_layout(pgio);
2727         pnfs_generic_pg_check_range(pgio, req);
2728         if (pgio->pg_lseg == NULL) {
2729                 if (pgio->pg_dreq == NULL)
2730                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2731                 else
2732                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq,
2733                                                       req_offset(req));
2734
2735                 pgio->pg_lseg =
2736                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2737                                            req_offset(req), rd_size,
2738                                            IOMODE_READ, false,
2739                                            nfs_io_gfp_mask());
2740                 if (IS_ERR(pgio->pg_lseg)) {
2741                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2742                         pgio->pg_lseg = NULL;
2743                         return;
2744                 }
2745         }
2746         /* If no lseg, fall back to read through mds */
2747         if (pgio->pg_lseg == NULL)
2748                 nfs_pageio_reset_read_mds(pgio);
2749
2750 }
2751 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2752
2753 void
2754 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2755                            struct nfs_page *req, u64 wb_size)
2756 {
2757         pnfs_generic_pg_check_layout(pgio);
2758         pnfs_generic_pg_check_range(pgio, req);
2759         if (pgio->pg_lseg == NULL) {
2760                 pgio->pg_lseg =
2761                         pnfs_update_layout(pgio->pg_inode, nfs_req_openctx(req),
2762                                            req_offset(req), wb_size, IOMODE_RW,
2763                                            false, nfs_io_gfp_mask());
2764                 if (IS_ERR(pgio->pg_lseg)) {
2765                         pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2766                         pgio->pg_lseg = NULL;
2767                         return;
2768                 }
2769         }
2770         /* If no lseg, fall back to write through mds */
2771         if (pgio->pg_lseg == NULL)
2772                 nfs_pageio_reset_write_mds(pgio);
2773 }
2774 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2775
2776 void
2777 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2778 {
2779         if (desc->pg_lseg) {
2780                 pnfs_put_lseg(desc->pg_lseg);
2781                 desc->pg_lseg = NULL;
2782         }
2783 }
2784 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2785
2786 /*
2787  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2788  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2789  */
2790 size_t
2791 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2792                      struct nfs_page *prev, struct nfs_page *req)
2793 {
2794         unsigned int size;
2795         u64 seg_end, req_start, seg_left;
2796
2797         size = nfs_generic_pg_test(pgio, prev, req);
2798         if (!size)
2799                 return 0;
2800
2801         /*
2802          * 'size' contains the number of bytes left in the current page (up
2803          * to the original size asked for in @req->wb_bytes).
2804          *
2805          * Calculate how many bytes are left in the layout segment
2806          * and if there are less bytes than 'size', return that instead.
2807          *
2808          * Please also note that 'end_offset' is actually the offset of the
2809          * first byte that lies outside the pnfs_layout_range. FIXME?
2810          *
2811          */
2812         if (pgio->pg_lseg) {
2813                 seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2814                                      pgio->pg_lseg->pls_range.length);
2815                 req_start = req_offset(req);
2816
2817                 /* start of request is past the last byte of this segment */
2818                 if (req_start >= seg_end)
2819                         return 0;
2820
2821                 /* adjust 'size' iff there are fewer bytes left in the
2822                  * segment than what nfs_generic_pg_test returned */
2823                 seg_left = seg_end - req_start;
2824                 if (seg_left < size)
2825                         size = (unsigned int)seg_left;
2826         }
2827
2828         return size;
2829 }
2830 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2831
2832 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2833 {
2834         struct nfs_pageio_descriptor pgio;
2835
2836         /* Resend all requests through the MDS */
2837         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2838                               hdr->completion_ops);
2839         return nfs_pageio_resend(&pgio, hdr);
2840 }
2841 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2842
2843 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2844 {
2845
2846         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2847         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2848             PNFS_LAYOUTRET_ON_ERROR) {
2849                 pnfs_return_layout(hdr->inode);
2850         }
2851         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2852                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2853 }
2854
2855 /*
2856  * Called by non rpc-based layout drivers
2857  */
2858 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2859 {
2860         if (likely(!hdr->pnfs_error)) {
2861                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2862                                 hdr->mds_offset + hdr->res.count);
2863                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2864         }
2865         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2866         if (unlikely(hdr->pnfs_error))
2867                 pnfs_ld_handle_write_error(hdr);
2868         hdr->mds_ops->rpc_release(hdr);
2869 }
2870 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2871
2872 static void
2873 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2874                 struct nfs_pgio_header *hdr)
2875 {
2876         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2877
2878         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2879                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2880                 nfs_pageio_reset_write_mds(desc);
2881                 mirror->pg_recoalesce = 1;
2882         }
2883         hdr->completion_ops->completion(hdr);
2884 }
2885
2886 static enum pnfs_try_status
2887 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2888                         const struct rpc_call_ops *call_ops,
2889                         struct pnfs_layout_segment *lseg,
2890                         int how)
2891 {
2892         struct inode *inode = hdr->inode;
2893         enum pnfs_try_status trypnfs;
2894         struct nfs_server *nfss = NFS_SERVER(inode);
2895
2896         hdr->mds_ops = call_ops;
2897
2898         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2899                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2900         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2901         if (trypnfs != PNFS_NOT_ATTEMPTED)
2902                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2903         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2904         return trypnfs;
2905 }
2906
2907 static void
2908 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2909               struct nfs_pgio_header *hdr, int how)
2910 {
2911         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2912         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2913         enum pnfs_try_status trypnfs;
2914
2915         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2916         switch (trypnfs) {
2917         case PNFS_NOT_ATTEMPTED:
2918                 pnfs_write_through_mds(desc, hdr);
2919                 break;
2920         case PNFS_ATTEMPTED:
2921                 break;
2922         case PNFS_TRY_AGAIN:
2923                 /* cleanup hdr and prepare to redo pnfs */
2924                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2925                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2926                         list_splice_init(&hdr->pages, &mirror->pg_list);
2927                         mirror->pg_recoalesce = 1;
2928                 }
2929                 hdr->mds_ops->rpc_release(hdr);
2930         }
2931 }
2932
2933 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2934 {
2935         pnfs_put_lseg(hdr->lseg);
2936         nfs_pgio_header_free(hdr);
2937 }
2938
2939 int
2940 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2941 {
2942         struct nfs_pgio_header *hdr;
2943         int ret;
2944
2945         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2946         if (!hdr) {
2947                 desc->pg_error = -ENOMEM;
2948                 return desc->pg_error;
2949         }
2950         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2951
2952         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2953         ret = nfs_generic_pgio(desc, hdr);
2954         if (!ret)
2955                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2956
2957         return ret;
2958 }
2959 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2960
2961 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2962 {
2963         struct nfs_pageio_descriptor pgio;
2964
2965         /* Resend all requests through the MDS */
2966         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2967         return nfs_pageio_resend(&pgio, hdr);
2968 }
2969 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2970
2971 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2972 {
2973         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2974         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2975             PNFS_LAYOUTRET_ON_ERROR) {
2976                 pnfs_return_layout(hdr->inode);
2977         }
2978         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2979                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2980 }
2981
2982 /*
2983  * Called by non rpc-based layout drivers
2984  */
2985 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2986 {
2987         if (likely(!hdr->pnfs_error))
2988                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2989         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2990         if (unlikely(hdr->pnfs_error))
2991                 pnfs_ld_handle_read_error(hdr);
2992         hdr->mds_ops->rpc_release(hdr);
2993 }
2994 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2995
2996 static void
2997 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2998                 struct nfs_pgio_header *hdr)
2999 {
3000         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3001
3002         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3003                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
3004                 nfs_pageio_reset_read_mds(desc);
3005                 mirror->pg_recoalesce = 1;
3006         }
3007         hdr->completion_ops->completion(hdr);
3008 }
3009
3010 /*
3011  * Call the appropriate parallel I/O subsystem read function.
3012  */
3013 static enum pnfs_try_status
3014 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
3015                        const struct rpc_call_ops *call_ops,
3016                        struct pnfs_layout_segment *lseg)
3017 {
3018         struct inode *inode = hdr->inode;
3019         struct nfs_server *nfss = NFS_SERVER(inode);
3020         enum pnfs_try_status trypnfs;
3021
3022         hdr->mds_ops = call_ops;
3023
3024         dprintk("%s: Reading ino:%lu %u@%llu\n",
3025                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
3026
3027         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
3028         if (trypnfs != PNFS_NOT_ATTEMPTED)
3029                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
3030         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
3031         return trypnfs;
3032 }
3033
3034 /* Resend all requests through pnfs. */
3035 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
3036                            unsigned int mirror_idx)
3037 {
3038         struct nfs_pageio_descriptor pgio;
3039
3040         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3041                 /* Prevent deadlocks with layoutreturn! */
3042                 pnfs_put_lseg(hdr->lseg);
3043                 hdr->lseg = NULL;
3044
3045                 nfs_pageio_init_read(&pgio, hdr->inode, false,
3046                                         hdr->completion_ops);
3047                 pgio.pg_mirror_idx = mirror_idx;
3048                 hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
3049         }
3050 }
3051 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
3052
3053 static void
3054 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
3055 {
3056         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
3057         struct pnfs_layout_segment *lseg = desc->pg_lseg;
3058         enum pnfs_try_status trypnfs;
3059
3060         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
3061         switch (trypnfs) {
3062         case PNFS_NOT_ATTEMPTED:
3063                 pnfs_read_through_mds(desc, hdr);
3064                 break;
3065         case PNFS_ATTEMPTED:
3066                 break;
3067         case PNFS_TRY_AGAIN:
3068                 /* cleanup hdr and prepare to redo pnfs */
3069                 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
3070                         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
3071                         list_splice_init(&hdr->pages, &mirror->pg_list);
3072                         mirror->pg_recoalesce = 1;
3073                 }
3074                 hdr->mds_ops->rpc_release(hdr);
3075         }
3076 }
3077
3078 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
3079 {
3080         pnfs_put_lseg(hdr->lseg);
3081         nfs_pgio_header_free(hdr);
3082 }
3083
3084 int
3085 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3086 {
3087         struct nfs_pgio_header *hdr;
3088         int ret;
3089
3090         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3091         if (!hdr) {
3092                 desc->pg_error = -ENOMEM;
3093                 return desc->pg_error;
3094         }
3095         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3096         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3097         ret = nfs_generic_pgio(desc, hdr);
3098         if (!ret)
3099                 pnfs_do_read(desc, hdr);
3100         return ret;
3101 }
3102 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3103
3104 static void pnfs_clear_layoutcommitting(struct inode *inode)
3105 {
3106         unsigned long *bitlock = &NFS_I(inode)->flags;
3107
3108         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3109         smp_mb__after_atomic();
3110         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3111 }
3112
3113 /*
3114  * There can be multiple RW segments.
3115  */
3116 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3117 {
3118         struct pnfs_layout_segment *lseg;
3119
3120         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3121                 if (lseg->pls_range.iomode == IOMODE_RW &&
3122                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3123                         list_add(&lseg->pls_lc_list, listp);
3124         }
3125 }
3126
3127 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3128 {
3129         struct pnfs_layout_segment *lseg, *tmp;
3130
3131         /* Matched by references in pnfs_set_layoutcommit */
3132         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3133                 list_del_init(&lseg->pls_lc_list);
3134                 pnfs_put_lseg(lseg);
3135         }
3136
3137         pnfs_clear_layoutcommitting(inode);
3138 }
3139
3140 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3141 {
3142         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3143 }
3144 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3145
3146 void
3147 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3148                 loff_t end_pos)
3149 {
3150         struct nfs_inode *nfsi = NFS_I(inode);
3151         bool mark_as_dirty = false;
3152
3153         spin_lock(&inode->i_lock);
3154         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3155                 nfsi->layout->plh_lwb = end_pos;
3156                 mark_as_dirty = true;
3157                 dprintk("%s: Set layoutcommit for inode %lu ",
3158                         __func__, inode->i_ino);
3159         } else if (end_pos > nfsi->layout->plh_lwb)
3160                 nfsi->layout->plh_lwb = end_pos;
3161         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3162                 /* references matched in nfs4_layoutcommit_release */
3163                 pnfs_get_lseg(lseg);
3164         }
3165         spin_unlock(&inode->i_lock);
3166         dprintk("%s: lseg %p end_pos %llu\n",
3167                 __func__, lseg, nfsi->layout->plh_lwb);
3168
3169         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3170          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3171         if (mark_as_dirty)
3172                 mark_inode_dirty_sync(inode);
3173 }
3174 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3175
3176 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3177 {
3178         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3179
3180         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3181                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3182         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3183 }
3184
3185 /*
3186  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3187  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3188  * data to disk to allow the server to recover the data if it crashes.
3189  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3190  * is off, and a COMMIT is sent to a data server, or
3191  * if WRITEs to a data server return NFS_DATA_SYNC.
3192  */
3193 int
3194 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3195 {
3196         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3197         struct nfs4_layoutcommit_data *data;
3198         struct nfs_inode *nfsi = NFS_I(inode);
3199         loff_t end_pos;
3200         int status;
3201
3202         if (!pnfs_layoutcommit_outstanding(inode))
3203                 return 0;
3204
3205         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3206
3207         status = -EAGAIN;
3208         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3209                 if (!sync)
3210                         goto out;
3211                 status = wait_on_bit_lock_action(&nfsi->flags,
3212                                 NFS_INO_LAYOUTCOMMITTING,
3213                                 nfs_wait_bit_killable,
3214                                 TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
3215                 if (status)
3216                         goto out;
3217         }
3218
3219         status = -ENOMEM;
3220         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3221         data = kzalloc(sizeof(*data), nfs_io_gfp_mask());
3222         if (!data)
3223                 goto clear_layoutcommitting;
3224
3225         status = 0;
3226         spin_lock(&inode->i_lock);
3227         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3228                 goto out_unlock;
3229
3230         INIT_LIST_HEAD(&data->lseg_list);
3231         pnfs_list_write_lseg(inode, &data->lseg_list);
3232
3233         end_pos = nfsi->layout->plh_lwb;
3234
3235         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3236         data->cred = get_cred(nfsi->layout->plh_lc_cred);
3237         spin_unlock(&inode->i_lock);
3238
3239         data->args.inode = inode;
3240         nfs_fattr_init(&data->fattr);
3241         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3242         data->res.fattr = &data->fattr;
3243         if (end_pos != 0)
3244                 data->args.lastbytewritten = end_pos - 1;
3245         else
3246                 data->args.lastbytewritten = U64_MAX;
3247         data->res.server = NFS_SERVER(inode);
3248
3249         if (ld->prepare_layoutcommit) {
3250                 status = ld->prepare_layoutcommit(&data->args);
3251                 if (status) {
3252                         put_cred(data->cred);
3253                         spin_lock(&inode->i_lock);
3254                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3255                         if (end_pos > nfsi->layout->plh_lwb)
3256                                 nfsi->layout->plh_lwb = end_pos;
3257                         goto out_unlock;
3258                 }
3259         }
3260
3261
3262         status = nfs4_proc_layoutcommit(data, sync);
3263 out:
3264         if (status)
3265                 mark_inode_dirty_sync(inode);
3266         dprintk("<-- %s status %d\n", __func__, status);
3267         return status;
3268 out_unlock:
3269         spin_unlock(&inode->i_lock);
3270         kfree(data);
3271 clear_layoutcommitting:
3272         pnfs_clear_layoutcommitting(inode);
3273         goto out;
3274 }
3275 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3276
3277 int
3278 pnfs_generic_sync(struct inode *inode, bool datasync)
3279 {
3280         return pnfs_layoutcommit_inode(inode, true);
3281 }
3282 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3283
3284 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3285 {
3286         struct nfs4_threshold *thp;
3287
3288         thp = kzalloc(sizeof(*thp), nfs_io_gfp_mask());
3289         if (!thp) {
3290                 dprintk("%s mdsthreshold allocation failed\n", __func__);
3291                 return NULL;
3292         }
3293         return thp;
3294 }
3295
3296 #if IS_ENABLED(CONFIG_NFS_V4_2)
3297 int
3298 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3299 {
3300         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3301         struct nfs_server *server = NFS_SERVER(inode);
3302         struct nfs_inode *nfsi = NFS_I(inode);
3303         struct nfs42_layoutstat_data *data;
3304         struct pnfs_layout_hdr *hdr;
3305         int status = 0;
3306
3307         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3308                 goto out;
3309
3310         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3311                 goto out;
3312
3313         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3314                 goto out;
3315
3316         spin_lock(&inode->i_lock);
3317         if (!NFS_I(inode)->layout) {
3318                 spin_unlock(&inode->i_lock);
3319                 goto out_clear_layoutstats;
3320         }
3321         hdr = NFS_I(inode)->layout;
3322         pnfs_get_layout_hdr(hdr);
3323         spin_unlock(&inode->i_lock);
3324
3325         data = kzalloc(sizeof(*data), gfp_flags);
3326         if (!data) {
3327                 status = -ENOMEM;
3328                 goto out_put;
3329         }
3330
3331         data->args.fh = NFS_FH(inode);
3332         data->args.inode = inode;
3333         status = ld->prepare_layoutstats(&data->args);
3334         if (status)
3335                 goto out_free;
3336
3337         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3338
3339 out:
3340         dprintk("%s returns %d\n", __func__, status);
3341         return status;
3342
3343 out_free:
3344         kfree(data);
3345 out_put:
3346         pnfs_put_layout_hdr(hdr);
3347 out_clear_layoutstats:
3348         smp_mb__before_atomic();
3349         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3350         smp_mb__after_atomic();
3351         goto out;
3352 }
3353 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3354 #endif
3355
3356 unsigned int layoutstats_timer;
3357 module_param(layoutstats_timer, uint, 0644);
3358 EXPORT_SYMBOL_GPL(layoutstats_timer);