NFSv4.1: Always clear the NFS_INO_LAYOUTCOMMIT in layoutreturn
[linux-3.10.git] / fs / nfs / 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 "internal.h"
34 #include "pnfs.h"
35 #include "iostat.h"
36
37 #define NFSDBG_FACILITY         NFSDBG_PNFS
38 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
39
40 /* Locking:
41  *
42  * pnfs_spinlock:
43  *      protects pnfs_modules_tbl.
44  */
45 static DEFINE_SPINLOCK(pnfs_spinlock);
46
47 /*
48  * pnfs_modules_tbl holds all pnfs modules
49  */
50 static LIST_HEAD(pnfs_modules_tbl);
51
52 /* Return the registered pnfs layout driver module matching given id */
53 static struct pnfs_layoutdriver_type *
54 find_pnfs_driver_locked(u32 id)
55 {
56         struct pnfs_layoutdriver_type *local;
57
58         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
59                 if (local->id == id)
60                         goto out;
61         local = NULL;
62 out:
63         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
64         return local;
65 }
66
67 static struct pnfs_layoutdriver_type *
68 find_pnfs_driver(u32 id)
69 {
70         struct pnfs_layoutdriver_type *local;
71
72         spin_lock(&pnfs_spinlock);
73         local = find_pnfs_driver_locked(id);
74         if (local != NULL && !try_module_get(local->owner)) {
75                 dprintk("%s: Could not grab reference on module\n", __func__);
76                 local = NULL;
77         }
78         spin_unlock(&pnfs_spinlock);
79         return local;
80 }
81
82 void
83 unset_pnfs_layoutdriver(struct nfs_server *nfss)
84 {
85         if (nfss->pnfs_curr_ld) {
86                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
87                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
88                 /* Decrement the MDS count. Purge the deviceid cache if zero */
89                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
90                         nfs4_deviceid_purge_client(nfss->nfs_client);
91                 module_put(nfss->pnfs_curr_ld->owner);
92         }
93         nfss->pnfs_curr_ld = NULL;
94 }
95
96 /*
97  * Try to set the server's pnfs module to the pnfs layout type specified by id.
98  * Currently only one pNFS layout driver per filesystem is supported.
99  *
100  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
101  */
102 void
103 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
104                       u32 id)
105 {
106         struct pnfs_layoutdriver_type *ld_type = NULL;
107
108         if (id == 0)
109                 goto out_no_driver;
110         if (!(server->nfs_client->cl_exchange_flags &
111                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
112                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
113                         __func__, id, server->nfs_client->cl_exchange_flags);
114                 goto out_no_driver;
115         }
116         ld_type = find_pnfs_driver(id);
117         if (!ld_type) {
118                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
119                 ld_type = find_pnfs_driver(id);
120                 if (!ld_type) {
121                         dprintk("%s: No pNFS module found for %u.\n",
122                                 __func__, id);
123                         goto out_no_driver;
124                 }
125         }
126         server->pnfs_curr_ld = ld_type;
127         if (ld_type->set_layoutdriver
128             && ld_type->set_layoutdriver(server, mntfh)) {
129                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
130                         "driver %u.\n", __func__, id);
131                 module_put(ld_type->owner);
132                 goto out_no_driver;
133         }
134         /* Bump the MDS count */
135         atomic_inc(&server->nfs_client->cl_mds_count);
136
137         dprintk("%s: pNFS module for %u set\n", __func__, id);
138         return;
139
140 out_no_driver:
141         dprintk("%s: Using NFSv4 I/O\n", __func__);
142         server->pnfs_curr_ld = NULL;
143 }
144
145 int
146 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
147 {
148         int status = -EINVAL;
149         struct pnfs_layoutdriver_type *tmp;
150
151         if (ld_type->id == 0) {
152                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
153                 return status;
154         }
155         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
156                 printk(KERN_ERR "NFS: %s Layout driver must provide "
157                        "alloc_lseg and free_lseg.\n", __func__);
158                 return status;
159         }
160
161         spin_lock(&pnfs_spinlock);
162         tmp = find_pnfs_driver_locked(ld_type->id);
163         if (!tmp) {
164                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
165                 status = 0;
166                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
167                         ld_type->name);
168         } else {
169                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
170                         __func__, ld_type->id);
171         }
172         spin_unlock(&pnfs_spinlock);
173
174         return status;
175 }
176 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
177
178 void
179 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
180 {
181         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
182         spin_lock(&pnfs_spinlock);
183         list_del(&ld_type->pnfs_tblid);
184         spin_unlock(&pnfs_spinlock);
185 }
186 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
187
188 /*
189  * pNFS client layout cache
190  */
191
192 /* Need to hold i_lock if caller does not already hold reference */
193 void
194 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
195 {
196         atomic_inc(&lo->plh_refcount);
197 }
198
199 static struct pnfs_layout_hdr *
200 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
201 {
202         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
203         return ld->alloc_layout_hdr(ino, gfp_flags);
204 }
205
206 static void
207 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
208 {
209         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
210         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
211
212         if (!list_empty(&lo->plh_layouts)) {
213                 struct nfs_client *clp = server->nfs_client;
214
215                 spin_lock(&clp->cl_lock);
216                 list_del_init(&lo->plh_layouts);
217                 spin_unlock(&clp->cl_lock);
218         }
219         put_rpccred(lo->plh_lc_cred);
220         return ld->free_layout_hdr(lo);
221 }
222
223 static void
224 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
225 {
226         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
227         dprintk("%s: freeing layout cache %p\n", __func__, lo);
228         nfsi->layout = NULL;
229         /* Reset MDS Threshold I/O counters */
230         nfsi->write_io = 0;
231         nfsi->read_io = 0;
232 }
233
234 void
235 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
236 {
237         struct inode *inode = lo->plh_inode;
238
239         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
240                 pnfs_detach_layout_hdr(lo);
241                 spin_unlock(&inode->i_lock);
242                 pnfs_free_layout_hdr(lo);
243         }
244 }
245
246 static int
247 pnfs_iomode_to_fail_bit(u32 iomode)
248 {
249         return iomode == IOMODE_RW ?
250                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
251 }
252
253 static void
254 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
255 {
256         lo->plh_retry_timestamp = jiffies;
257         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
258                 atomic_inc(&lo->plh_refcount);
259 }
260
261 static void
262 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
263 {
264         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
265                 atomic_dec(&lo->plh_refcount);
266 }
267
268 static void
269 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
270 {
271         struct inode *inode = lo->plh_inode;
272         struct pnfs_layout_range range = {
273                 .iomode = iomode,
274                 .offset = 0,
275                 .length = NFS4_MAX_UINT64,
276         };
277         LIST_HEAD(head);
278
279         spin_lock(&inode->i_lock);
280         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
281         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
282         spin_unlock(&inode->i_lock);
283         pnfs_free_lseg_list(&head);
284         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
285                         iomode == IOMODE_RW ?  "RW" : "READ");
286 }
287
288 static bool
289 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
290 {
291         unsigned long start, end;
292         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
293
294         if (test_bit(fail_bit, &lo->plh_flags) == 0)
295                 return false;
296         end = jiffies;
297         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
298         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
299                 /* It is time to retry the failed layoutgets */
300                 pnfs_layout_clear_fail_bit(lo, fail_bit);
301                 return false;
302         }
303         return true;
304 }
305
306 static void
307 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
308 {
309         INIT_LIST_HEAD(&lseg->pls_list);
310         INIT_LIST_HEAD(&lseg->pls_lc_list);
311         atomic_set(&lseg->pls_refcount, 1);
312         smp_mb();
313         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
314         lseg->pls_layout = lo;
315 }
316
317 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
318 {
319         struct inode *ino = lseg->pls_layout->plh_inode;
320
321         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
322 }
323
324 static void
325 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
326                 struct pnfs_layout_segment *lseg)
327 {
328         struct inode *inode = lo->plh_inode;
329
330         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
331         list_del_init(&lseg->pls_list);
332         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
333         atomic_dec(&lo->plh_refcount);
334         if (list_empty(&lo->plh_segs))
335                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
336         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
337 }
338
339 void
340 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
341 {
342         struct pnfs_layout_hdr *lo;
343         struct inode *inode;
344
345         if (!lseg)
346                 return;
347
348         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
349                 atomic_read(&lseg->pls_refcount),
350                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
351         lo = lseg->pls_layout;
352         inode = lo->plh_inode;
353         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
354                 pnfs_get_layout_hdr(lo);
355                 pnfs_layout_remove_lseg(lo, lseg);
356                 spin_unlock(&inode->i_lock);
357                 pnfs_free_lseg(lseg);
358                 pnfs_put_layout_hdr(lo);
359         }
360 }
361 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
362
363 static inline u64
364 end_offset(u64 start, u64 len)
365 {
366         u64 end;
367
368         end = start + len;
369         return end >= start ? end : NFS4_MAX_UINT64;
370 }
371
372 /*
373  * is l2 fully contained in l1?
374  *   start1                             end1
375  *   [----------------------------------)
376  *           start2           end2
377  *           [----------------)
378  */
379 static inline int
380 lo_seg_contained(struct pnfs_layout_range *l1,
381                  struct pnfs_layout_range *l2)
382 {
383         u64 start1 = l1->offset;
384         u64 end1 = end_offset(start1, l1->length);
385         u64 start2 = l2->offset;
386         u64 end2 = end_offset(start2, l2->length);
387
388         return (start1 <= start2) && (end1 >= end2);
389 }
390
391 /*
392  * is l1 and l2 intersecting?
393  *   start1                             end1
394  *   [----------------------------------)
395  *                              start2           end2
396  *                              [----------------)
397  */
398 static inline int
399 lo_seg_intersecting(struct pnfs_layout_range *l1,
400                     struct pnfs_layout_range *l2)
401 {
402         u64 start1 = l1->offset;
403         u64 end1 = end_offset(start1, l1->length);
404         u64 start2 = l2->offset;
405         u64 end2 = end_offset(start2, l2->length);
406
407         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
408                (end2 == NFS4_MAX_UINT64 || end2 > start1);
409 }
410
411 static bool
412 should_free_lseg(struct pnfs_layout_range *lseg_range,
413                  struct pnfs_layout_range *recall_range)
414 {
415         return (recall_range->iomode == IOMODE_ANY ||
416                 lseg_range->iomode == recall_range->iomode) &&
417                lo_seg_intersecting(lseg_range, recall_range);
418 }
419
420 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
421                 struct list_head *tmp_list)
422 {
423         if (!atomic_dec_and_test(&lseg->pls_refcount))
424                 return false;
425         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
426         list_add(&lseg->pls_list, tmp_list);
427         return true;
428 }
429
430 /* Returns 1 if lseg is removed from list, 0 otherwise */
431 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
432                              struct list_head *tmp_list)
433 {
434         int rv = 0;
435
436         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
437                 /* Remove the reference keeping the lseg in the
438                  * list.  It will now be removed when all
439                  * outstanding io is finished.
440                  */
441                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
442                         atomic_read(&lseg->pls_refcount));
443                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
444                         rv = 1;
445         }
446         return rv;
447 }
448
449 /* Returns count of number of matching invalid lsegs remaining in list
450  * after call.
451  */
452 int
453 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
454                             struct list_head *tmp_list,
455                             struct pnfs_layout_range *recall_range)
456 {
457         struct pnfs_layout_segment *lseg, *next;
458         int invalid = 0, removed = 0;
459
460         dprintk("%s:Begin lo %p\n", __func__, lo);
461
462         if (list_empty(&lo->plh_segs))
463                 return 0;
464         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
465                 if (!recall_range ||
466                     should_free_lseg(&lseg->pls_range, recall_range)) {
467                         dprintk("%s: freeing lseg %p iomode %d "
468                                 "offset %llu length %llu\n", __func__,
469                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
470                                 lseg->pls_range.length);
471                         invalid++;
472                         removed += mark_lseg_invalid(lseg, tmp_list);
473                 }
474         dprintk("%s:Return %i\n", __func__, invalid - removed);
475         return invalid - removed;
476 }
477
478 /* note free_me must contain lsegs from a single layout_hdr */
479 void
480 pnfs_free_lseg_list(struct list_head *free_me)
481 {
482         struct pnfs_layout_segment *lseg, *tmp;
483
484         if (list_empty(free_me))
485                 return;
486
487         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
488                 list_del(&lseg->pls_list);
489                 pnfs_free_lseg(lseg);
490         }
491 }
492
493 void
494 pnfs_destroy_layout(struct nfs_inode *nfsi)
495 {
496         struct pnfs_layout_hdr *lo;
497         LIST_HEAD(tmp_list);
498
499         spin_lock(&nfsi->vfs_inode.i_lock);
500         lo = nfsi->layout;
501         if (lo) {
502                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
503                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
504                 pnfs_get_layout_hdr(lo);
505                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
506                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
507                 spin_unlock(&nfsi->vfs_inode.i_lock);
508                 pnfs_free_lseg_list(&tmp_list);
509                 pnfs_put_layout_hdr(lo);
510         } else
511                 spin_unlock(&nfsi->vfs_inode.i_lock);
512 }
513 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
514
515 static bool
516 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
517                 struct list_head *layout_list)
518 {
519         struct pnfs_layout_hdr *lo;
520         bool ret = false;
521
522         spin_lock(&inode->i_lock);
523         lo = NFS_I(inode)->layout;
524         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
525                 pnfs_get_layout_hdr(lo);
526                 list_add(&lo->plh_bulk_destroy, layout_list);
527                 ret = true;
528         }
529         spin_unlock(&inode->i_lock);
530         return ret;
531 }
532
533 /* Caller must hold rcu_read_lock and clp->cl_lock */
534 static int
535 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
536                 struct nfs_server *server,
537                 struct list_head *layout_list)
538 {
539         struct pnfs_layout_hdr *lo, *next;
540         struct inode *inode;
541
542         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
543                 inode = igrab(lo->plh_inode);
544                 if (inode == NULL)
545                         continue;
546                 list_del_init(&lo->plh_layouts);
547                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
548                         continue;
549                 rcu_read_unlock();
550                 spin_unlock(&clp->cl_lock);
551                 iput(inode);
552                 spin_lock(&clp->cl_lock);
553                 rcu_read_lock();
554                 return -EAGAIN;
555         }
556         return 0;
557 }
558
559 static int
560 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
561                 bool is_bulk_recall)
562 {
563         struct pnfs_layout_hdr *lo;
564         struct inode *inode;
565         struct pnfs_layout_range range = {
566                 .iomode = IOMODE_ANY,
567                 .offset = 0,
568                 .length = NFS4_MAX_UINT64,
569         };
570         LIST_HEAD(lseg_list);
571         int ret = 0;
572
573         while (!list_empty(layout_list)) {
574                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
575                                 plh_bulk_destroy);
576                 dprintk("%s freeing layout for inode %lu\n", __func__,
577                         lo->plh_inode->i_ino);
578                 inode = lo->plh_inode;
579                 spin_lock(&inode->i_lock);
580                 list_del_init(&lo->plh_bulk_destroy);
581                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
582                 if (is_bulk_recall)
583                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
584                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
585                         ret = -EAGAIN;
586                 spin_unlock(&inode->i_lock);
587                 pnfs_free_lseg_list(&lseg_list);
588                 pnfs_put_layout_hdr(lo);
589                 iput(inode);
590         }
591         return ret;
592 }
593
594 int
595 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
596                 struct nfs_fsid *fsid,
597                 bool is_recall)
598 {
599         struct nfs_server *server;
600         LIST_HEAD(layout_list);
601
602         spin_lock(&clp->cl_lock);
603         rcu_read_lock();
604 restart:
605         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
606                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
607                         continue;
608                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
609                                 server,
610                                 &layout_list) != 0)
611                         goto restart;
612         }
613         rcu_read_unlock();
614         spin_unlock(&clp->cl_lock);
615
616         if (list_empty(&layout_list))
617                 return 0;
618         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
619 }
620
621 int
622 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
623                 bool is_recall)
624 {
625         struct nfs_server *server;
626         LIST_HEAD(layout_list);
627
628         spin_lock(&clp->cl_lock);
629         rcu_read_lock();
630 restart:
631         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
632                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
633                                         server,
634                                         &layout_list) != 0)
635                         goto restart;
636         }
637         rcu_read_unlock();
638         spin_unlock(&clp->cl_lock);
639
640         if (list_empty(&layout_list))
641                 return 0;
642         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
643 }
644
645 /*
646  * Called by the state manger to remove all layouts established under an
647  * expired lease.
648  */
649 void
650 pnfs_destroy_all_layouts(struct nfs_client *clp)
651 {
652         nfs4_deviceid_mark_client_invalid(clp);
653         nfs4_deviceid_purge_client(clp);
654
655         pnfs_destroy_layouts_byclid(clp, false);
656 }
657
658 /*
659  * Compare 2 layout stateid sequence ids, to see which is newer,
660  * taking into account wraparound issues.
661  */
662 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
663 {
664         return (s32)s1 - (s32)s2 > 0;
665 }
666
667 /* update lo->plh_stateid with new if is more recent */
668 void
669 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
670                         bool update_barrier)
671 {
672         u32 oldseq, newseq, new_barrier;
673         int empty = list_empty(&lo->plh_segs);
674
675         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
676         newseq = be32_to_cpu(new->seqid);
677         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
678                 nfs4_stateid_copy(&lo->plh_stateid, new);
679                 if (update_barrier) {
680                         new_barrier = be32_to_cpu(new->seqid);
681                 } else {
682                         /* Because of wraparound, we want to keep the barrier
683                          * "close" to the current seqids.
684                          */
685                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
686                 }
687                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
688                         lo->plh_barrier = new_barrier;
689         }
690 }
691
692 static bool
693 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
694                 const nfs4_stateid *stateid)
695 {
696         u32 seqid = be32_to_cpu(stateid->seqid);
697
698         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
699 }
700
701 /* lget is set to 1 if called from inside send_layoutget call chain */
702 static bool
703 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo, int lget)
704 {
705         return lo->plh_block_lgets ||
706                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags) ||
707                 (list_empty(&lo->plh_segs) &&
708                  (atomic_read(&lo->plh_outstanding) > lget));
709 }
710
711 int
712 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
713                               struct nfs4_state *open_state)
714 {
715         int status = 0;
716
717         dprintk("--> %s\n", __func__);
718         spin_lock(&lo->plh_inode->i_lock);
719         if (pnfs_layoutgets_blocked(lo, 1)) {
720                 status = -EAGAIN;
721         } else if (list_empty(&lo->plh_segs)) {
722                 int seq;
723
724                 do {
725                         seq = read_seqbegin(&open_state->seqlock);
726                         nfs4_stateid_copy(dst, &open_state->stateid);
727                 } while (read_seqretry(&open_state->seqlock, seq));
728         } else
729                 nfs4_stateid_copy(dst, &lo->plh_stateid);
730         spin_unlock(&lo->plh_inode->i_lock);
731         dprintk("<-- %s\n", __func__);
732         return status;
733 }
734
735 /*
736 * Get layout from server.
737 *    for now, assume that whole file layouts are requested.
738 *    arg->offset: 0
739 *    arg->length: all ones
740 */
741 static struct pnfs_layout_segment *
742 send_layoutget(struct pnfs_layout_hdr *lo,
743            struct nfs_open_context *ctx,
744            struct pnfs_layout_range *range,
745            gfp_t gfp_flags)
746 {
747         struct inode *ino = lo->plh_inode;
748         struct nfs_server *server = NFS_SERVER(ino);
749         struct nfs4_layoutget *lgp;
750         struct pnfs_layout_segment *lseg;
751
752         dprintk("--> %s\n", __func__);
753
754         lgp = kzalloc(sizeof(*lgp), gfp_flags);
755         if (lgp == NULL)
756                 return NULL;
757
758         lgp->args.minlength = PAGE_CACHE_SIZE;
759         if (lgp->args.minlength > range->length)
760                 lgp->args.minlength = range->length;
761         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
762         lgp->args.range = *range;
763         lgp->args.type = server->pnfs_curr_ld->id;
764         lgp->args.inode = ino;
765         lgp->args.ctx = get_nfs_open_context(ctx);
766         lgp->gfp_flags = gfp_flags;
767
768         /* Synchronously retrieve layout information from server and
769          * store in lseg.
770          */
771         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
772         if (IS_ERR(lseg)) {
773                 switch (PTR_ERR(lseg)) {
774                 case -ENOMEM:
775                 case -ERESTARTSYS:
776                         break;
777                 default:
778                         /* remember that LAYOUTGET failed and suspend trying */
779                         pnfs_layout_io_set_failed(lo, range->iomode);
780                 }
781                 return NULL;
782         }
783
784         return lseg;
785 }
786
787 static void pnfs_clear_layoutcommit(struct inode *inode,
788                 struct list_head *head)
789 {
790         struct nfs_inode *nfsi = NFS_I(inode);
791         struct pnfs_layout_segment *lseg, *tmp;
792
793         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
794                 return;
795         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
796                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
797                         continue;
798                 pnfs_lseg_dec_and_remove_zero(lseg, head);
799         }
800 }
801
802 /*
803  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
804  * when the layout segment list is empty.
805  *
806  * Note that a pnfs_layout_hdr can exist with an empty layout segment
807  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
808  * deviceid is marked invalid.
809  */
810 int
811 _pnfs_return_layout(struct inode *ino)
812 {
813         struct pnfs_layout_hdr *lo = NULL;
814         struct nfs_inode *nfsi = NFS_I(ino);
815         LIST_HEAD(tmp_list);
816         struct nfs4_layoutreturn *lrp;
817         nfs4_stateid stateid;
818         int status = 0, empty;
819
820         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
821
822         spin_lock(&ino->i_lock);
823         lo = nfsi->layout;
824         if (!lo) {
825                 spin_unlock(&ino->i_lock);
826                 dprintk("NFS: %s no layout to return\n", __func__);
827                 goto out;
828         }
829         stateid = nfsi->layout->plh_stateid;
830         /* Reference matched in nfs4_layoutreturn_release */
831         pnfs_get_layout_hdr(lo);
832         empty = list_empty(&lo->plh_segs);
833         pnfs_clear_layoutcommit(ino, &tmp_list);
834         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
835         /* Don't send a LAYOUTRETURN if list was initially empty */
836         if (empty) {
837                 spin_unlock(&ino->i_lock);
838                 pnfs_put_layout_hdr(lo);
839                 dprintk("NFS: %s no layout segments to return\n", __func__);
840                 goto out;
841         }
842         lo->plh_block_lgets++;
843         spin_unlock(&ino->i_lock);
844         pnfs_free_lseg_list(&tmp_list);
845
846         lrp = kzalloc(sizeof(*lrp), GFP_KERNEL);
847         if (unlikely(lrp == NULL)) {
848                 status = -ENOMEM;
849                 spin_lock(&ino->i_lock);
850                 lo->plh_block_lgets--;
851                 spin_unlock(&ino->i_lock);
852                 pnfs_put_layout_hdr(lo);
853                 goto out;
854         }
855
856         lrp->args.stateid = stateid;
857         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
858         lrp->args.inode = ino;
859         lrp->args.layout = lo;
860         lrp->clp = NFS_SERVER(ino)->nfs_client;
861
862         status = nfs4_proc_layoutreturn(lrp);
863 out:
864         dprintk("<-- %s status: %d\n", __func__, status);
865         return status;
866 }
867 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
868
869 bool pnfs_roc(struct inode *ino)
870 {
871         struct pnfs_layout_hdr *lo;
872         struct pnfs_layout_segment *lseg, *tmp;
873         LIST_HEAD(tmp_list);
874         bool found = false;
875
876         spin_lock(&ino->i_lock);
877         lo = NFS_I(ino)->layout;
878         if (!lo || !test_and_clear_bit(NFS_LAYOUT_ROC, &lo->plh_flags) ||
879             test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
880                 goto out_nolayout;
881         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
882                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
883                         mark_lseg_invalid(lseg, &tmp_list);
884                         found = true;
885                 }
886         if (!found)
887                 goto out_nolayout;
888         lo->plh_block_lgets++;
889         pnfs_get_layout_hdr(lo); /* matched in pnfs_roc_release */
890         spin_unlock(&ino->i_lock);
891         pnfs_free_lseg_list(&tmp_list);
892         return true;
893
894 out_nolayout:
895         spin_unlock(&ino->i_lock);
896         return false;
897 }
898
899 void pnfs_roc_release(struct inode *ino)
900 {
901         struct pnfs_layout_hdr *lo;
902
903         spin_lock(&ino->i_lock);
904         lo = NFS_I(ino)->layout;
905         lo->plh_block_lgets--;
906         if (atomic_dec_and_test(&lo->plh_refcount)) {
907                 pnfs_detach_layout_hdr(lo);
908                 spin_unlock(&ino->i_lock);
909                 pnfs_free_layout_hdr(lo);
910         } else
911                 spin_unlock(&ino->i_lock);
912 }
913
914 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
915 {
916         struct pnfs_layout_hdr *lo;
917
918         spin_lock(&ino->i_lock);
919         lo = NFS_I(ino)->layout;
920         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
921                 lo->plh_barrier = barrier;
922         spin_unlock(&ino->i_lock);
923 }
924
925 bool pnfs_roc_drain(struct inode *ino, u32 *barrier, struct rpc_task *task)
926 {
927         struct nfs_inode *nfsi = NFS_I(ino);
928         struct pnfs_layout_hdr *lo;
929         struct pnfs_layout_segment *lseg;
930         u32 current_seqid;
931         bool found = false;
932
933         spin_lock(&ino->i_lock);
934         list_for_each_entry(lseg, &nfsi->layout->plh_segs, pls_list)
935                 if (test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
936                         rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
937                         found = true;
938                         goto out;
939                 }
940         lo = nfsi->layout;
941         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
942
943         /* Since close does not return a layout stateid for use as
944          * a barrier, we choose the worst-case barrier.
945          */
946         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
947 out:
948         spin_unlock(&ino->i_lock);
949         return found;
950 }
951
952 /*
953  * Compare two layout segments for sorting into layout cache.
954  * We want to preferentially return RW over RO layouts, so ensure those
955  * are seen first.
956  */
957 static s64
958 cmp_layout(struct pnfs_layout_range *l1,
959            struct pnfs_layout_range *l2)
960 {
961         s64 d;
962
963         /* high offset > low offset */
964         d = l1->offset - l2->offset;
965         if (d)
966                 return d;
967
968         /* short length > long length */
969         d = l2->length - l1->length;
970         if (d)
971                 return d;
972
973         /* read > read/write */
974         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
975 }
976
977 static void
978 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
979                    struct pnfs_layout_segment *lseg)
980 {
981         struct pnfs_layout_segment *lp;
982
983         dprintk("%s:Begin\n", __func__);
984
985         list_for_each_entry(lp, &lo->plh_segs, pls_list) {
986                 if (cmp_layout(&lseg->pls_range, &lp->pls_range) > 0)
987                         continue;
988                 list_add_tail(&lseg->pls_list, &lp->pls_list);
989                 dprintk("%s: inserted lseg %p "
990                         "iomode %d offset %llu length %llu before "
991                         "lp %p iomode %d offset %llu length %llu\n",
992                         __func__, lseg, lseg->pls_range.iomode,
993                         lseg->pls_range.offset, lseg->pls_range.length,
994                         lp, lp->pls_range.iomode, lp->pls_range.offset,
995                         lp->pls_range.length);
996                 goto out;
997         }
998         list_add_tail(&lseg->pls_list, &lo->plh_segs);
999         dprintk("%s: inserted lseg %p "
1000                 "iomode %d offset %llu length %llu at tail\n",
1001                 __func__, lseg, lseg->pls_range.iomode,
1002                 lseg->pls_range.offset, lseg->pls_range.length);
1003 out:
1004         pnfs_get_layout_hdr(lo);
1005
1006         dprintk("%s:Return\n", __func__);
1007 }
1008
1009 static struct pnfs_layout_hdr *
1010 alloc_init_layout_hdr(struct inode *ino,
1011                       struct nfs_open_context *ctx,
1012                       gfp_t gfp_flags)
1013 {
1014         struct pnfs_layout_hdr *lo;
1015
1016         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1017         if (!lo)
1018                 return NULL;
1019         atomic_set(&lo->plh_refcount, 1);
1020         INIT_LIST_HEAD(&lo->plh_layouts);
1021         INIT_LIST_HEAD(&lo->plh_segs);
1022         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1023         lo->plh_inode = ino;
1024         lo->plh_lc_cred = get_rpccred(ctx->state->owner->so_cred);
1025         return lo;
1026 }
1027
1028 static struct pnfs_layout_hdr *
1029 pnfs_find_alloc_layout(struct inode *ino,
1030                        struct nfs_open_context *ctx,
1031                        gfp_t gfp_flags)
1032 {
1033         struct nfs_inode *nfsi = NFS_I(ino);
1034         struct pnfs_layout_hdr *new = NULL;
1035
1036         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1037
1038         if (nfsi->layout != NULL)
1039                 goto out_existing;
1040         spin_unlock(&ino->i_lock);
1041         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1042         spin_lock(&ino->i_lock);
1043
1044         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1045                 nfsi->layout = new;
1046                 return new;
1047         } else if (new != NULL)
1048                 pnfs_free_layout_hdr(new);
1049 out_existing:
1050         pnfs_get_layout_hdr(nfsi->layout);
1051         return nfsi->layout;
1052 }
1053
1054 /*
1055  * iomode matching rules:
1056  * iomode       lseg    match
1057  * -----        -----   -----
1058  * ANY          READ    true
1059  * ANY          RW      true
1060  * RW           READ    false
1061  * RW           RW      true
1062  * READ         READ    true
1063  * READ         RW      true
1064  */
1065 static int
1066 is_matching_lseg(struct pnfs_layout_range *ls_range,
1067                  struct pnfs_layout_range *range)
1068 {
1069         struct pnfs_layout_range range1;
1070
1071         if ((range->iomode == IOMODE_RW &&
1072              ls_range->iomode != IOMODE_RW) ||
1073             !lo_seg_intersecting(ls_range, range))
1074                 return 0;
1075
1076         /* range1 covers only the first byte in the range */
1077         range1 = *range;
1078         range1.length = 1;
1079         return lo_seg_contained(ls_range, &range1);
1080 }
1081
1082 /*
1083  * lookup range in layout
1084  */
1085 static struct pnfs_layout_segment *
1086 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1087                 struct pnfs_layout_range *range)
1088 {
1089         struct pnfs_layout_segment *lseg, *ret = NULL;
1090
1091         dprintk("%s:Begin\n", __func__);
1092
1093         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1094                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1095                     is_matching_lseg(&lseg->pls_range, range)) {
1096                         ret = pnfs_get_lseg(lseg);
1097                         break;
1098                 }
1099                 if (lseg->pls_range.offset > range->offset)
1100                         break;
1101         }
1102
1103         dprintk("%s:Return lseg %p ref %d\n",
1104                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1105         return ret;
1106 }
1107
1108 /*
1109  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1110  * to the MDS or over pNFS
1111  *
1112  * The nfs_inode read_io and write_io fields are cumulative counters reset
1113  * when there are no layout segments. Note that in pnfs_update_layout iomode
1114  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1115  * WRITE request.
1116  *
1117  * A return of true means use MDS I/O.
1118  *
1119  * From rfc 5661:
1120  * If a file's size is smaller than the file size threshold, data accesses
1121  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1122  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1123  * server.  If both file size and I/O size are provided, the client SHOULD
1124  * reach or exceed  both thresholds before sending its read or write
1125  * requests to the data server.
1126  */
1127 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1128                                      struct inode *ino, int iomode)
1129 {
1130         struct nfs4_threshold *t = ctx->mdsthreshold;
1131         struct nfs_inode *nfsi = NFS_I(ino);
1132         loff_t fsize = i_size_read(ino);
1133         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1134
1135         if (t == NULL)
1136                 return ret;
1137
1138         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1139                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1140
1141         switch (iomode) {
1142         case IOMODE_READ:
1143                 if (t->bm & THRESHOLD_RD) {
1144                         dprintk("%s fsize %llu\n", __func__, fsize);
1145                         size_set = true;
1146                         if (fsize < t->rd_sz)
1147                                 size = true;
1148                 }
1149                 if (t->bm & THRESHOLD_RD_IO) {
1150                         dprintk("%s nfsi->read_io %llu\n", __func__,
1151                                 nfsi->read_io);
1152                         io_set = true;
1153                         if (nfsi->read_io < t->rd_io_sz)
1154                                 io = true;
1155                 }
1156                 break;
1157         case IOMODE_RW:
1158                 if (t->bm & THRESHOLD_WR) {
1159                         dprintk("%s fsize %llu\n", __func__, fsize);
1160                         size_set = true;
1161                         if (fsize < t->wr_sz)
1162                                 size = true;
1163                 }
1164                 if (t->bm & THRESHOLD_WR_IO) {
1165                         dprintk("%s nfsi->write_io %llu\n", __func__,
1166                                 nfsi->write_io);
1167                         io_set = true;
1168                         if (nfsi->write_io < t->wr_io_sz)
1169                                 io = true;
1170                 }
1171                 break;
1172         }
1173         if (size_set && io_set) {
1174                 if (size && io)
1175                         ret = true;
1176         } else if (size || io)
1177                 ret = true;
1178
1179         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1180         return ret;
1181 }
1182
1183 /*
1184  * Layout segment is retreived from the server if not cached.
1185  * The appropriate layout segment is referenced and returned to the caller.
1186  */
1187 struct pnfs_layout_segment *
1188 pnfs_update_layout(struct inode *ino,
1189                    struct nfs_open_context *ctx,
1190                    loff_t pos,
1191                    u64 count,
1192                    enum pnfs_iomode iomode,
1193                    gfp_t gfp_flags)
1194 {
1195         struct pnfs_layout_range arg = {
1196                 .iomode = iomode,
1197                 .offset = pos,
1198                 .length = count,
1199         };
1200         unsigned pg_offset;
1201         struct nfs_server *server = NFS_SERVER(ino);
1202         struct nfs_client *clp = server->nfs_client;
1203         struct pnfs_layout_hdr *lo;
1204         struct pnfs_layout_segment *lseg = NULL;
1205         bool first;
1206
1207         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1208                 goto out;
1209
1210         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1211                 goto out;
1212
1213         spin_lock(&ino->i_lock);
1214         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1215         if (lo == NULL) {
1216                 spin_unlock(&ino->i_lock);
1217                 goto out;
1218         }
1219
1220         /* Do we even need to bother with this? */
1221         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1222                 dprintk("%s matches recall, use MDS\n", __func__);
1223                 goto out_unlock;
1224         }
1225
1226         /* if LAYOUTGET already failed once we don't try again */
1227         if (pnfs_layout_io_test_failed(lo, iomode))
1228                 goto out_unlock;
1229
1230         /* Check to see if the layout for the given range already exists */
1231         lseg = pnfs_find_lseg(lo, &arg);
1232         if (lseg)
1233                 goto out_unlock;
1234
1235         if (pnfs_layoutgets_blocked(lo, 0))
1236                 goto out_unlock;
1237         atomic_inc(&lo->plh_outstanding);
1238
1239         first = list_empty(&lo->plh_layouts) ? true : false;
1240         spin_unlock(&ino->i_lock);
1241
1242         if (first) {
1243                 /* The lo must be on the clp list if there is any
1244                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1245                  */
1246                 spin_lock(&clp->cl_lock);
1247                 list_add_tail(&lo->plh_layouts, &server->layouts);
1248                 spin_unlock(&clp->cl_lock);
1249         }
1250
1251         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1252         if (pg_offset) {
1253                 arg.offset -= pg_offset;
1254                 arg.length += pg_offset;
1255         }
1256         if (arg.length != NFS4_MAX_UINT64)
1257                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1258
1259         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1260         atomic_dec(&lo->plh_outstanding);
1261 out_put_layout_hdr:
1262         pnfs_put_layout_hdr(lo);
1263 out:
1264         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1265                         "(%s, offset: %llu, length: %llu)\n",
1266                         __func__, ino->i_sb->s_id,
1267                         (unsigned long long)NFS_FILEID(ino),
1268                         lseg == NULL ? "not found" : "found",
1269                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1270                         (unsigned long long)pos,
1271                         (unsigned long long)count);
1272         return lseg;
1273 out_unlock:
1274         spin_unlock(&ino->i_lock);
1275         goto out_put_layout_hdr;
1276 }
1277 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1278
1279 struct pnfs_layout_segment *
1280 pnfs_layout_process(struct nfs4_layoutget *lgp)
1281 {
1282         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1283         struct nfs4_layoutget_res *res = &lgp->res;
1284         struct pnfs_layout_segment *lseg;
1285         struct inode *ino = lo->plh_inode;
1286         int status = 0;
1287
1288         /* Inject layout blob into I/O device driver */
1289         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1290         if (!lseg || IS_ERR(lseg)) {
1291                 if (!lseg)
1292                         status = -ENOMEM;
1293                 else
1294                         status = PTR_ERR(lseg);
1295                 dprintk("%s: Could not allocate layout: error %d\n",
1296                        __func__, status);
1297                 goto out;
1298         }
1299
1300         spin_lock(&ino->i_lock);
1301         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1302                 dprintk("%s forget reply due to recall\n", __func__);
1303                 goto out_forget_reply;
1304         }
1305
1306         if (pnfs_layoutgets_blocked(lo, 1) ||
1307             pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1308                 dprintk("%s forget reply due to state\n", __func__);
1309                 goto out_forget_reply;
1310         }
1311
1312         /* Done processing layoutget. Set the layout stateid */
1313         pnfs_set_layout_stateid(lo, &res->stateid, false);
1314
1315         init_lseg(lo, lseg);
1316         lseg->pls_range = res->range;
1317         pnfs_get_lseg(lseg);
1318         pnfs_layout_insert_lseg(lo, lseg);
1319
1320         if (res->return_on_close) {
1321                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1322                 set_bit(NFS_LAYOUT_ROC, &lo->plh_flags);
1323         }
1324
1325         spin_unlock(&ino->i_lock);
1326         return lseg;
1327 out:
1328         return ERR_PTR(status);
1329
1330 out_forget_reply:
1331         spin_unlock(&ino->i_lock);
1332         lseg->pls_layout = lo;
1333         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1334         goto out;
1335 }
1336
1337 void
1338 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1339 {
1340         u64 rd_size = req->wb_bytes;
1341
1342         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1343
1344         if (req->wb_offset != req->wb_pgbase) {
1345                 nfs_pageio_reset_read_mds(pgio);
1346                 return;
1347         }
1348
1349         if (pgio->pg_dreq == NULL)
1350                 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1351         else
1352                 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1353
1354         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1355                                            req->wb_context,
1356                                            req_offset(req),
1357                                            rd_size,
1358                                            IOMODE_READ,
1359                                            GFP_KERNEL);
1360         /* If no lseg, fall back to read through mds */
1361         if (pgio->pg_lseg == NULL)
1362                 nfs_pageio_reset_read_mds(pgio);
1363
1364 }
1365 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1366
1367 void
1368 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1369                            struct nfs_page *req, u64 wb_size)
1370 {
1371         WARN_ON_ONCE(pgio->pg_lseg != NULL);
1372
1373         if (req->wb_offset != req->wb_pgbase) {
1374                 nfs_pageio_reset_write_mds(pgio);
1375                 return;
1376         }
1377
1378         pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1379                                            req->wb_context,
1380                                            req_offset(req),
1381                                            wb_size,
1382                                            IOMODE_RW,
1383                                            GFP_NOFS);
1384         /* If no lseg, fall back to write through mds */
1385         if (pgio->pg_lseg == NULL)
1386                 nfs_pageio_reset_write_mds(pgio);
1387 }
1388 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1389
1390 void
1391 pnfs_pageio_init_read(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1392                       const struct nfs_pgio_completion_ops *compl_ops)
1393 {
1394         struct nfs_server *server = NFS_SERVER(inode);
1395         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1396
1397         if (ld == NULL)
1398                 nfs_pageio_init_read(pgio, inode, compl_ops);
1399         else
1400                 nfs_pageio_init(pgio, inode, ld->pg_read_ops, compl_ops, server->rsize, 0);
1401 }
1402
1403 void
1404 pnfs_pageio_init_write(struct nfs_pageio_descriptor *pgio, struct inode *inode,
1405                        int ioflags,
1406                        const struct nfs_pgio_completion_ops *compl_ops)
1407 {
1408         struct nfs_server *server = NFS_SERVER(inode);
1409         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
1410
1411         if (ld == NULL)
1412                 nfs_pageio_init_write(pgio, inode, ioflags, compl_ops);
1413         else
1414                 nfs_pageio_init(pgio, inode, ld->pg_write_ops, compl_ops, server->wsize, ioflags);
1415 }
1416
1417 bool
1418 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
1419                      struct nfs_page *req)
1420 {
1421         if (pgio->pg_lseg == NULL)
1422                 return nfs_generic_pg_test(pgio, prev, req);
1423
1424         /*
1425          * Test if a nfs_page is fully contained in the pnfs_layout_range.
1426          * Note that this test makes several assumptions:
1427          * - that the previous nfs_page in the struct nfs_pageio_descriptor
1428          *   is known to lie within the range.
1429          *   - that the nfs_page being tested is known to be contiguous with the
1430          *   previous nfs_page.
1431          *   - Layout ranges are page aligned, so we only have to test the
1432          *   start offset of the request.
1433          *
1434          * Please also note that 'end_offset' is actually the offset of the
1435          * first byte that lies outside the pnfs_layout_range. FIXME?
1436          *
1437          */
1438         return req_offset(req) < end_offset(pgio->pg_lseg->pls_range.offset,
1439                                          pgio->pg_lseg->pls_range.length);
1440 }
1441 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1442
1443 int pnfs_write_done_resend_to_mds(struct inode *inode,
1444                                 struct list_head *head,
1445                                 const struct nfs_pgio_completion_ops *compl_ops,
1446                                 struct nfs_direct_req *dreq)
1447 {
1448         struct nfs_pageio_descriptor pgio;
1449         LIST_HEAD(failed);
1450
1451         /* Resend all requests through the MDS */
1452         nfs_pageio_init_write(&pgio, inode, FLUSH_STABLE, compl_ops);
1453         pgio.pg_dreq = dreq;
1454         while (!list_empty(head)) {
1455                 struct nfs_page *req = nfs_list_entry(head->next);
1456
1457                 nfs_list_remove_request(req);
1458                 if (!nfs_pageio_add_request(&pgio, req))
1459                         nfs_list_add_request(req, &failed);
1460         }
1461         nfs_pageio_complete(&pgio);
1462
1463         if (!list_empty(&failed)) {
1464                 /* For some reason our attempt to resend pages. Mark the
1465                  * overall send request as having failed, and let
1466                  * nfs_writeback_release_full deal with the error.
1467                  */
1468                 list_move(&failed, head);
1469                 return -EIO;
1470         }
1471         return 0;
1472 }
1473 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1474
1475 static void pnfs_ld_handle_write_error(struct nfs_write_data *data)
1476 {
1477         struct nfs_pgio_header *hdr = data->header;
1478
1479         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1480         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1481             PNFS_LAYOUTRET_ON_ERROR) {
1482                 pnfs_return_layout(hdr->inode);
1483         }
1484         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1485                 data->task.tk_status = pnfs_write_done_resend_to_mds(hdr->inode,
1486                                                         &hdr->pages,
1487                                                         hdr->completion_ops,
1488                                                         hdr->dreq);
1489 }
1490
1491 /*
1492  * Called by non rpc-based layout drivers
1493  */
1494 void pnfs_ld_write_done(struct nfs_write_data *data)
1495 {
1496         struct nfs_pgio_header *hdr = data->header;
1497
1498         if (!hdr->pnfs_error) {
1499                 pnfs_set_layoutcommit(data);
1500                 hdr->mds_ops->rpc_call_done(&data->task, data);
1501         } else
1502                 pnfs_ld_handle_write_error(data);
1503         hdr->mds_ops->rpc_release(data);
1504 }
1505 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1506
1507 static void
1508 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1509                 struct nfs_write_data *data)
1510 {
1511         struct nfs_pgio_header *hdr = data->header;
1512
1513         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1514                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1515                 nfs_pageio_reset_write_mds(desc);
1516                 desc->pg_recoalesce = 1;
1517         }
1518         nfs_writedata_release(data);
1519 }
1520
1521 static enum pnfs_try_status
1522 pnfs_try_to_write_data(struct nfs_write_data *wdata,
1523                         const struct rpc_call_ops *call_ops,
1524                         struct pnfs_layout_segment *lseg,
1525                         int how)
1526 {
1527         struct nfs_pgio_header *hdr = wdata->header;
1528         struct inode *inode = hdr->inode;
1529         enum pnfs_try_status trypnfs;
1530         struct nfs_server *nfss = NFS_SERVER(inode);
1531
1532         hdr->mds_ops = call_ops;
1533
1534         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1535                 inode->i_ino, wdata->args.count, wdata->args.offset, how);
1536         trypnfs = nfss->pnfs_curr_ld->write_pagelist(wdata, how);
1537         if (trypnfs != PNFS_NOT_ATTEMPTED)
1538                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1539         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1540         return trypnfs;
1541 }
1542
1543 static void
1544 pnfs_do_multiple_writes(struct nfs_pageio_descriptor *desc, struct list_head *head, int how)
1545 {
1546         struct nfs_write_data *data;
1547         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1548         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1549
1550         desc->pg_lseg = NULL;
1551         while (!list_empty(head)) {
1552                 enum pnfs_try_status trypnfs;
1553
1554                 data = list_first_entry(head, struct nfs_write_data, list);
1555                 list_del_init(&data->list);
1556
1557                 trypnfs = pnfs_try_to_write_data(data, call_ops, lseg, how);
1558                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1559                         pnfs_write_through_mds(desc, data);
1560         }
1561         pnfs_put_lseg(lseg);
1562 }
1563
1564 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1565 {
1566         pnfs_put_lseg(hdr->lseg);
1567         nfs_writehdr_free(hdr);
1568 }
1569 EXPORT_SYMBOL_GPL(pnfs_writehdr_free);
1570
1571 int
1572 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1573 {
1574         struct nfs_write_header *whdr;
1575         struct nfs_pgio_header *hdr;
1576         int ret;
1577
1578         whdr = nfs_writehdr_alloc();
1579         if (!whdr) {
1580                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1581                 pnfs_put_lseg(desc->pg_lseg);
1582                 desc->pg_lseg = NULL;
1583                 return -ENOMEM;
1584         }
1585         hdr = &whdr->header;
1586         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1587         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1588         atomic_inc(&hdr->refcnt);
1589         ret = nfs_generic_flush(desc, hdr);
1590         if (ret != 0) {
1591                 pnfs_put_lseg(desc->pg_lseg);
1592                 desc->pg_lseg = NULL;
1593         } else
1594                 pnfs_do_multiple_writes(desc, &hdr->rpc_list, desc->pg_ioflags);
1595         if (atomic_dec_and_test(&hdr->refcnt))
1596                 hdr->completion_ops->completion(hdr);
1597         return ret;
1598 }
1599 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
1600
1601 int pnfs_read_done_resend_to_mds(struct inode *inode,
1602                                 struct list_head *head,
1603                                 const struct nfs_pgio_completion_ops *compl_ops,
1604                                 struct nfs_direct_req *dreq)
1605 {
1606         struct nfs_pageio_descriptor pgio;
1607         LIST_HEAD(failed);
1608
1609         /* Resend all requests through the MDS */
1610         nfs_pageio_init_read(&pgio, inode, compl_ops);
1611         pgio.pg_dreq = dreq;
1612         while (!list_empty(head)) {
1613                 struct nfs_page *req = nfs_list_entry(head->next);
1614
1615                 nfs_list_remove_request(req);
1616                 if (!nfs_pageio_add_request(&pgio, req))
1617                         nfs_list_add_request(req, &failed);
1618         }
1619         nfs_pageio_complete(&pgio);
1620
1621         if (!list_empty(&failed)) {
1622                 list_move(&failed, head);
1623                 return -EIO;
1624         }
1625         return 0;
1626 }
1627 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
1628
1629 static void pnfs_ld_handle_read_error(struct nfs_read_data *data)
1630 {
1631         struct nfs_pgio_header *hdr = data->header;
1632
1633         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
1634         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1635             PNFS_LAYOUTRET_ON_ERROR) {
1636                 pnfs_return_layout(hdr->inode);
1637         }
1638         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1639                 data->task.tk_status = pnfs_read_done_resend_to_mds(hdr->inode,
1640                                                         &hdr->pages,
1641                                                         hdr->completion_ops,
1642                                                         hdr->dreq);
1643 }
1644
1645 /*
1646  * Called by non rpc-based layout drivers
1647  */
1648 void pnfs_ld_read_done(struct nfs_read_data *data)
1649 {
1650         struct nfs_pgio_header *hdr = data->header;
1651
1652         if (likely(!hdr->pnfs_error)) {
1653                 __nfs4_read_done_cb(data);
1654                 hdr->mds_ops->rpc_call_done(&data->task, data);
1655         } else
1656                 pnfs_ld_handle_read_error(data);
1657         hdr->mds_ops->rpc_release(data);
1658 }
1659 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
1660
1661 static void
1662 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
1663                 struct nfs_read_data *data)
1664 {
1665         struct nfs_pgio_header *hdr = data->header;
1666
1667         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1668                 list_splice_tail_init(&hdr->pages, &desc->pg_list);
1669                 nfs_pageio_reset_read_mds(desc);
1670                 desc->pg_recoalesce = 1;
1671         }
1672         nfs_readdata_release(data);
1673 }
1674
1675 /*
1676  * Call the appropriate parallel I/O subsystem read function.
1677  */
1678 static enum pnfs_try_status
1679 pnfs_try_to_read_data(struct nfs_read_data *rdata,
1680                        const struct rpc_call_ops *call_ops,
1681                        struct pnfs_layout_segment *lseg)
1682 {
1683         struct nfs_pgio_header *hdr = rdata->header;
1684         struct inode *inode = hdr->inode;
1685         struct nfs_server *nfss = NFS_SERVER(inode);
1686         enum pnfs_try_status trypnfs;
1687
1688         hdr->mds_ops = call_ops;
1689
1690         dprintk("%s: Reading ino:%lu %u@%llu\n",
1691                 __func__, inode->i_ino, rdata->args.count, rdata->args.offset);
1692
1693         trypnfs = nfss->pnfs_curr_ld->read_pagelist(rdata);
1694         if (trypnfs != PNFS_NOT_ATTEMPTED)
1695                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
1696         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1697         return trypnfs;
1698 }
1699
1700 static void
1701 pnfs_do_multiple_reads(struct nfs_pageio_descriptor *desc, struct list_head *head)
1702 {
1703         struct nfs_read_data *data;
1704         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1705         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1706
1707         desc->pg_lseg = NULL;
1708         while (!list_empty(head)) {
1709                 enum pnfs_try_status trypnfs;
1710
1711                 data = list_first_entry(head, struct nfs_read_data, list);
1712                 list_del_init(&data->list);
1713
1714                 trypnfs = pnfs_try_to_read_data(data, call_ops, lseg);
1715                 if (trypnfs == PNFS_NOT_ATTEMPTED)
1716                         pnfs_read_through_mds(desc, data);
1717         }
1718         pnfs_put_lseg(lseg);
1719 }
1720
1721 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
1722 {
1723         pnfs_put_lseg(hdr->lseg);
1724         nfs_readhdr_free(hdr);
1725 }
1726 EXPORT_SYMBOL_GPL(pnfs_readhdr_free);
1727
1728 int
1729 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
1730 {
1731         struct nfs_read_header *rhdr;
1732         struct nfs_pgio_header *hdr;
1733         int ret;
1734
1735         rhdr = nfs_readhdr_alloc();
1736         if (!rhdr) {
1737                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
1738                 ret = -ENOMEM;
1739                 pnfs_put_lseg(desc->pg_lseg);
1740                 desc->pg_lseg = NULL;
1741                 return ret;
1742         }
1743         hdr = &rhdr->header;
1744         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
1745         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1746         atomic_inc(&hdr->refcnt);
1747         ret = nfs_generic_pagein(desc, hdr);
1748         if (ret != 0) {
1749                 pnfs_put_lseg(desc->pg_lseg);
1750                 desc->pg_lseg = NULL;
1751         } else
1752                 pnfs_do_multiple_reads(desc, &hdr->rpc_list);
1753         if (atomic_dec_and_test(&hdr->refcnt))
1754                 hdr->completion_ops->completion(hdr);
1755         return ret;
1756 }
1757 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
1758
1759 /*
1760  * There can be multiple RW segments.
1761  */
1762 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
1763 {
1764         struct pnfs_layout_segment *lseg;
1765
1766         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
1767                 if (lseg->pls_range.iomode == IOMODE_RW &&
1768                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1769                         list_add(&lseg->pls_lc_list, listp);
1770         }
1771 }
1772
1773 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
1774 {
1775         struct pnfs_layout_segment *lseg, *tmp;
1776         unsigned long *bitlock = &NFS_I(inode)->flags;
1777
1778         /* Matched by references in pnfs_set_layoutcommit */
1779         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
1780                 list_del_init(&lseg->pls_lc_list);
1781                 pnfs_put_lseg(lseg);
1782         }
1783
1784         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
1785         smp_mb__after_clear_bit();
1786         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
1787 }
1788
1789 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
1790 {
1791         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
1792 }
1793 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
1794
1795 void
1796 pnfs_set_layoutcommit(struct nfs_write_data *wdata)
1797 {
1798         struct nfs_pgio_header *hdr = wdata->header;
1799         struct inode *inode = hdr->inode;
1800         struct nfs_inode *nfsi = NFS_I(inode);
1801         loff_t end_pos = wdata->mds_offset + wdata->res.count;
1802         bool mark_as_dirty = false;
1803
1804         spin_lock(&inode->i_lock);
1805         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1806                 mark_as_dirty = true;
1807                 dprintk("%s: Set layoutcommit for inode %lu ",
1808                         __func__, inode->i_ino);
1809         }
1810         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &hdr->lseg->pls_flags)) {
1811                 /* references matched in nfs4_layoutcommit_release */
1812                 pnfs_get_lseg(hdr->lseg);
1813         }
1814         if (end_pos > nfsi->layout->plh_lwb)
1815                 nfsi->layout->plh_lwb = end_pos;
1816         spin_unlock(&inode->i_lock);
1817         dprintk("%s: lseg %p end_pos %llu\n",
1818                 __func__, hdr->lseg, nfsi->layout->plh_lwb);
1819
1820         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
1821          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
1822         if (mark_as_dirty)
1823                 mark_inode_dirty_sync(inode);
1824 }
1825 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
1826
1827 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
1828 {
1829         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
1830
1831         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
1832                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
1833         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
1834 }
1835
1836 /*
1837  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
1838  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
1839  * data to disk to allow the server to recover the data if it crashes.
1840  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
1841  * is off, and a COMMIT is sent to a data server, or
1842  * if WRITEs to a data server return NFS_DATA_SYNC.
1843  */
1844 int
1845 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
1846 {
1847         struct nfs4_layoutcommit_data *data;
1848         struct nfs_inode *nfsi = NFS_I(inode);
1849         loff_t end_pos;
1850         int status = 0;
1851
1852         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
1853
1854         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1855                 return 0;
1856
1857         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
1858         data = kzalloc(sizeof(*data), GFP_NOFS);
1859         if (!data) {
1860                 status = -ENOMEM;
1861                 goto out;
1862         }
1863
1864         if (!test_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1865                 goto out_free;
1866
1867         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
1868                 if (!sync) {
1869                         status = -EAGAIN;
1870                         goto out_free;
1871                 }
1872                 status = wait_on_bit_lock(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING,
1873                                         nfs_wait_bit_killable, TASK_KILLABLE);
1874                 if (status)
1875                         goto out_free;
1876         }
1877
1878         INIT_LIST_HEAD(&data->lseg_list);
1879         spin_lock(&inode->i_lock);
1880         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
1881                 clear_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags);
1882                 spin_unlock(&inode->i_lock);
1883                 wake_up_bit(&nfsi->flags, NFS_INO_LAYOUTCOMMITTING);
1884                 goto out_free;
1885         }
1886
1887         pnfs_list_write_lseg(inode, &data->lseg_list);
1888
1889         end_pos = nfsi->layout->plh_lwb;
1890         nfsi->layout->plh_lwb = 0;
1891
1892         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
1893         spin_unlock(&inode->i_lock);
1894
1895         data->args.inode = inode;
1896         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
1897         nfs_fattr_init(&data->fattr);
1898         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
1899         data->res.fattr = &data->fattr;
1900         data->args.lastbytewritten = end_pos - 1;
1901         data->res.server = NFS_SERVER(inode);
1902
1903         status = nfs4_proc_layoutcommit(data, sync);
1904 out:
1905         if (status)
1906                 mark_inode_dirty_sync(inode);
1907         dprintk("<-- %s status %d\n", __func__, status);
1908         return status;
1909 out_free:
1910         kfree(data);
1911         goto out;
1912 }
1913
1914 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
1915 {
1916         struct nfs4_threshold *thp;
1917
1918         thp = kzalloc(sizeof(*thp), GFP_NOFS);
1919         if (!thp) {
1920                 dprintk("%s mdsthreshold allocation failed\n", __func__);
1921                 return NULL;
1922         }
1923         return thp;
1924 }