Merge branch 'for-next' of git://git.samba.org/sfrench/cifs-2.6
[firefly-linux-kernel-4.4.55.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 #include "nfs4trace.h"
37 #include "delegation.h"
38 #include "nfs42.h"
39
40 #define NFSDBG_FACILITY         NFSDBG_PNFS
41 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
42
43 /* Locking:
44  *
45  * pnfs_spinlock:
46  *      protects pnfs_modules_tbl.
47  */
48 static DEFINE_SPINLOCK(pnfs_spinlock);
49
50 /*
51  * pnfs_modules_tbl holds all pnfs modules
52  */
53 static LIST_HEAD(pnfs_modules_tbl);
54
55 static int
56 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
57                        enum pnfs_iomode iomode, bool sync);
58
59 /* Return the registered pnfs layout driver module matching given id */
60 static struct pnfs_layoutdriver_type *
61 find_pnfs_driver_locked(u32 id)
62 {
63         struct pnfs_layoutdriver_type *local;
64
65         list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
66                 if (local->id == id)
67                         goto out;
68         local = NULL;
69 out:
70         dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
71         return local;
72 }
73
74 static struct pnfs_layoutdriver_type *
75 find_pnfs_driver(u32 id)
76 {
77         struct pnfs_layoutdriver_type *local;
78
79         spin_lock(&pnfs_spinlock);
80         local = find_pnfs_driver_locked(id);
81         if (local != NULL && !try_module_get(local->owner)) {
82                 dprintk("%s: Could not grab reference on module\n", __func__);
83                 local = NULL;
84         }
85         spin_unlock(&pnfs_spinlock);
86         return local;
87 }
88
89 void
90 unset_pnfs_layoutdriver(struct nfs_server *nfss)
91 {
92         if (nfss->pnfs_curr_ld) {
93                 if (nfss->pnfs_curr_ld->clear_layoutdriver)
94                         nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
95                 /* Decrement the MDS count. Purge the deviceid cache if zero */
96                 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
97                         nfs4_deviceid_purge_client(nfss->nfs_client);
98                 module_put(nfss->pnfs_curr_ld->owner);
99         }
100         nfss->pnfs_curr_ld = NULL;
101 }
102
103 /*
104  * Try to set the server's pnfs module to the pnfs layout type specified by id.
105  * Currently only one pNFS layout driver per filesystem is supported.
106  *
107  * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
108  */
109 void
110 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
111                       u32 id)
112 {
113         struct pnfs_layoutdriver_type *ld_type = NULL;
114
115         if (id == 0)
116                 goto out_no_driver;
117         if (!(server->nfs_client->cl_exchange_flags &
118                  (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
119                 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
120                         __func__, id, server->nfs_client->cl_exchange_flags);
121                 goto out_no_driver;
122         }
123         ld_type = find_pnfs_driver(id);
124         if (!ld_type) {
125                 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
126                 ld_type = find_pnfs_driver(id);
127                 if (!ld_type) {
128                         dprintk("%s: No pNFS module found for %u.\n",
129                                 __func__, id);
130                         goto out_no_driver;
131                 }
132         }
133         server->pnfs_curr_ld = ld_type;
134         if (ld_type->set_layoutdriver
135             && ld_type->set_layoutdriver(server, mntfh)) {
136                 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
137                         "driver %u.\n", __func__, id);
138                 module_put(ld_type->owner);
139                 goto out_no_driver;
140         }
141         /* Bump the MDS count */
142         atomic_inc(&server->nfs_client->cl_mds_count);
143
144         dprintk("%s: pNFS module for %u set\n", __func__, id);
145         return;
146
147 out_no_driver:
148         dprintk("%s: Using NFSv4 I/O\n", __func__);
149         server->pnfs_curr_ld = NULL;
150 }
151
152 int
153 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
154 {
155         int status = -EINVAL;
156         struct pnfs_layoutdriver_type *tmp;
157
158         if (ld_type->id == 0) {
159                 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
160                 return status;
161         }
162         if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
163                 printk(KERN_ERR "NFS: %s Layout driver must provide "
164                        "alloc_lseg and free_lseg.\n", __func__);
165                 return status;
166         }
167
168         spin_lock(&pnfs_spinlock);
169         tmp = find_pnfs_driver_locked(ld_type->id);
170         if (!tmp) {
171                 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
172                 status = 0;
173                 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
174                         ld_type->name);
175         } else {
176                 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
177                         __func__, ld_type->id);
178         }
179         spin_unlock(&pnfs_spinlock);
180
181         return status;
182 }
183 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
184
185 void
186 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
187 {
188         dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
189         spin_lock(&pnfs_spinlock);
190         list_del(&ld_type->pnfs_tblid);
191         spin_unlock(&pnfs_spinlock);
192 }
193 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
194
195 /*
196  * pNFS client layout cache
197  */
198
199 /* Need to hold i_lock if caller does not already hold reference */
200 void
201 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
202 {
203         atomic_inc(&lo->plh_refcount);
204 }
205
206 static struct pnfs_layout_hdr *
207 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
208 {
209         struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
210         return ld->alloc_layout_hdr(ino, gfp_flags);
211 }
212
213 static void
214 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
215 {
216         struct nfs_server *server = NFS_SERVER(lo->plh_inode);
217         struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
218
219         if (!list_empty(&lo->plh_layouts)) {
220                 struct nfs_client *clp = server->nfs_client;
221
222                 spin_lock(&clp->cl_lock);
223                 list_del_init(&lo->plh_layouts);
224                 spin_unlock(&clp->cl_lock);
225         }
226         put_rpccred(lo->plh_lc_cred);
227         return ld->free_layout_hdr(lo);
228 }
229
230 static void
231 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
232 {
233         struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
234         dprintk("%s: freeing layout cache %p\n", __func__, lo);
235         nfsi->layout = NULL;
236         /* Reset MDS Threshold I/O counters */
237         nfsi->write_io = 0;
238         nfsi->read_io = 0;
239 }
240
241 void
242 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
243 {
244         struct inode *inode = lo->plh_inode;
245
246         if (atomic_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
247                 if (!list_empty(&lo->plh_segs))
248                         WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
249                 pnfs_detach_layout_hdr(lo);
250                 spin_unlock(&inode->i_lock);
251                 pnfs_free_layout_hdr(lo);
252         }
253 }
254
255 static int
256 pnfs_iomode_to_fail_bit(u32 iomode)
257 {
258         return iomode == IOMODE_RW ?
259                 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
260 }
261
262 static void
263 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
264 {
265         lo->plh_retry_timestamp = jiffies;
266         if (!test_and_set_bit(fail_bit, &lo->plh_flags))
267                 atomic_inc(&lo->plh_refcount);
268 }
269
270 static void
271 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
272 {
273         if (test_and_clear_bit(fail_bit, &lo->plh_flags))
274                 atomic_dec(&lo->plh_refcount);
275 }
276
277 static void
278 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
279 {
280         struct inode *inode = lo->plh_inode;
281         struct pnfs_layout_range range = {
282                 .iomode = iomode,
283                 .offset = 0,
284                 .length = NFS4_MAX_UINT64,
285         };
286         LIST_HEAD(head);
287
288         spin_lock(&inode->i_lock);
289         pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
290         pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
291         spin_unlock(&inode->i_lock);
292         pnfs_free_lseg_list(&head);
293         dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
294                         iomode == IOMODE_RW ?  "RW" : "READ");
295 }
296
297 static bool
298 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
299 {
300         unsigned long start, end;
301         int fail_bit = pnfs_iomode_to_fail_bit(iomode);
302
303         if (test_bit(fail_bit, &lo->plh_flags) == 0)
304                 return false;
305         end = jiffies;
306         start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
307         if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
308                 /* It is time to retry the failed layoutgets */
309                 pnfs_layout_clear_fail_bit(lo, fail_bit);
310                 return false;
311         }
312         return true;
313 }
314
315 static void
316 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
317 {
318         INIT_LIST_HEAD(&lseg->pls_list);
319         INIT_LIST_HEAD(&lseg->pls_lc_list);
320         atomic_set(&lseg->pls_refcount, 1);
321         smp_mb();
322         set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
323         lseg->pls_layout = lo;
324 }
325
326 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
327 {
328         struct inode *ino = lseg->pls_layout->plh_inode;
329
330         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
331 }
332
333 static void
334 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
335                 struct pnfs_layout_segment *lseg)
336 {
337         struct inode *inode = lo->plh_inode;
338
339         WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
340         list_del_init(&lseg->pls_list);
341         /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
342         atomic_dec(&lo->plh_refcount);
343         if (list_empty(&lo->plh_segs))
344                 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
345         rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
346 }
347
348 /* Return true if layoutreturn is needed */
349 static bool
350 pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
351                         struct pnfs_layout_segment *lseg)
352 {
353         struct pnfs_layout_segment *s;
354
355         if (!test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
356                 return false;
357
358         list_for_each_entry(s, &lo->plh_segs, pls_list)
359                 if (s != lseg && test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
360                         return false;
361
362         return true;
363 }
364
365 static bool
366 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
367 {
368         if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
369                 return false;
370         lo->plh_return_iomode = 0;
371         pnfs_get_layout_hdr(lo);
372         clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE, &lo->plh_flags);
373         return true;
374 }
375
376 static void pnfs_layoutreturn_before_put_lseg(struct pnfs_layout_segment *lseg,
377                 struct pnfs_layout_hdr *lo, struct inode *inode)
378 {
379         lo = lseg->pls_layout;
380         inode = lo->plh_inode;
381
382         spin_lock(&inode->i_lock);
383         if (pnfs_layout_need_return(lo, lseg)) {
384                 nfs4_stateid stateid;
385                 enum pnfs_iomode iomode;
386                 bool send;
387
388                 stateid = lo->plh_stateid;
389                 iomode = lo->plh_return_iomode;
390                 send = pnfs_prepare_layoutreturn(lo);
391                 spin_unlock(&inode->i_lock);
392                 if (send) {
393                         /* Send an async layoutreturn so we dont deadlock */
394                         pnfs_send_layoutreturn(lo, stateid, iomode, false);
395                 }
396         } else
397                 spin_unlock(&inode->i_lock);
398 }
399
400 void
401 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
402 {
403         struct pnfs_layout_hdr *lo;
404         struct inode *inode;
405
406         if (!lseg)
407                 return;
408
409         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
410                 atomic_read(&lseg->pls_refcount),
411                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
412
413         /* Handle the case where refcount != 1 */
414         if (atomic_add_unless(&lseg->pls_refcount, -1, 1))
415                 return;
416
417         lo = lseg->pls_layout;
418         inode = lo->plh_inode;
419         /* Do we need a layoutreturn? */
420         if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
421                 pnfs_layoutreturn_before_put_lseg(lseg, lo, inode);
422
423         if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
424                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
425                         spin_unlock(&inode->i_lock);
426                         return;
427                 }
428                 pnfs_get_layout_hdr(lo);
429                 pnfs_layout_remove_lseg(lo, lseg);
430                 spin_unlock(&inode->i_lock);
431                 pnfs_free_lseg(lseg);
432                 pnfs_put_layout_hdr(lo);
433         }
434 }
435 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
436
437 static void pnfs_free_lseg_async_work(struct work_struct *work)
438 {
439         struct pnfs_layout_segment *lseg;
440         struct pnfs_layout_hdr *lo;
441
442         lseg = container_of(work, struct pnfs_layout_segment, pls_work);
443         lo = lseg->pls_layout;
444
445         pnfs_free_lseg(lseg);
446         pnfs_put_layout_hdr(lo);
447 }
448
449 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
450 {
451         INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
452         schedule_work(&lseg->pls_work);
453 }
454
455 void
456 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
457 {
458         if (!lseg)
459                 return;
460
461         assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
462
463         dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
464                 atomic_read(&lseg->pls_refcount),
465                 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
466         if (atomic_dec_and_test(&lseg->pls_refcount)) {
467                 struct pnfs_layout_hdr *lo = lseg->pls_layout;
468                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
469                         return;
470                 pnfs_get_layout_hdr(lo);
471                 pnfs_layout_remove_lseg(lo, lseg);
472                 pnfs_free_lseg_async(lseg);
473         }
474 }
475 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
476
477 static u64
478 end_offset(u64 start, u64 len)
479 {
480         u64 end;
481
482         end = start + len;
483         return end >= start ? end : NFS4_MAX_UINT64;
484 }
485
486 /*
487  * is l2 fully contained in l1?
488  *   start1                             end1
489  *   [----------------------------------)
490  *           start2           end2
491  *           [----------------)
492  */
493 static bool
494 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
495                  const struct pnfs_layout_range *l2)
496 {
497         u64 start1 = l1->offset;
498         u64 end1 = end_offset(start1, l1->length);
499         u64 start2 = l2->offset;
500         u64 end2 = end_offset(start2, l2->length);
501
502         return (start1 <= start2) && (end1 >= end2);
503 }
504
505 /*
506  * is l1 and l2 intersecting?
507  *   start1                             end1
508  *   [----------------------------------)
509  *                              start2           end2
510  *                              [----------------)
511  */
512 static bool
513 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
514                     const struct pnfs_layout_range *l2)
515 {
516         u64 start1 = l1->offset;
517         u64 end1 = end_offset(start1, l1->length);
518         u64 start2 = l2->offset;
519         u64 end2 = end_offset(start2, l2->length);
520
521         return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
522                (end2 == NFS4_MAX_UINT64 || end2 > start1);
523 }
524
525 static bool
526 should_free_lseg(const struct pnfs_layout_range *lseg_range,
527                  const struct pnfs_layout_range *recall_range)
528 {
529         return (recall_range->iomode == IOMODE_ANY ||
530                 lseg_range->iomode == recall_range->iomode) &&
531                pnfs_lseg_range_intersecting(lseg_range, recall_range);
532 }
533
534 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
535                 struct list_head *tmp_list)
536 {
537         if (!atomic_dec_and_test(&lseg->pls_refcount))
538                 return false;
539         pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
540         list_add(&lseg->pls_list, tmp_list);
541         return true;
542 }
543
544 /* Returns 1 if lseg is removed from list, 0 otherwise */
545 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
546                              struct list_head *tmp_list)
547 {
548         int rv = 0;
549
550         if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
551                 /* Remove the reference keeping the lseg in the
552                  * list.  It will now be removed when all
553                  * outstanding io is finished.
554                  */
555                 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
556                         atomic_read(&lseg->pls_refcount));
557                 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
558                         rv = 1;
559         }
560         return rv;
561 }
562
563 /* Returns count of number of matching invalid lsegs remaining in list
564  * after call.
565  */
566 int
567 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
568                             struct list_head *tmp_list,
569                             struct pnfs_layout_range *recall_range)
570 {
571         struct pnfs_layout_segment *lseg, *next;
572         int invalid = 0, removed = 0;
573
574         dprintk("%s:Begin lo %p\n", __func__, lo);
575
576         if (list_empty(&lo->plh_segs))
577                 return 0;
578         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
579                 if (!recall_range ||
580                     should_free_lseg(&lseg->pls_range, recall_range)) {
581                         dprintk("%s: freeing lseg %p iomode %d "
582                                 "offset %llu length %llu\n", __func__,
583                                 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
584                                 lseg->pls_range.length);
585                         invalid++;
586                         removed += mark_lseg_invalid(lseg, tmp_list);
587                 }
588         dprintk("%s:Return %i\n", __func__, invalid - removed);
589         return invalid - removed;
590 }
591
592 /* note free_me must contain lsegs from a single layout_hdr */
593 void
594 pnfs_free_lseg_list(struct list_head *free_me)
595 {
596         struct pnfs_layout_segment *lseg, *tmp;
597
598         if (list_empty(free_me))
599                 return;
600
601         list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
602                 list_del(&lseg->pls_list);
603                 pnfs_free_lseg(lseg);
604         }
605 }
606
607 void
608 pnfs_destroy_layout(struct nfs_inode *nfsi)
609 {
610         struct pnfs_layout_hdr *lo;
611         LIST_HEAD(tmp_list);
612
613         spin_lock(&nfsi->vfs_inode.i_lock);
614         lo = nfsi->layout;
615         if (lo) {
616                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
617                 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
618                 pnfs_get_layout_hdr(lo);
619                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
620                 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
621                 pnfs_clear_retry_layoutget(lo);
622                 spin_unlock(&nfsi->vfs_inode.i_lock);
623                 pnfs_free_lseg_list(&tmp_list);
624                 pnfs_put_layout_hdr(lo);
625         } else
626                 spin_unlock(&nfsi->vfs_inode.i_lock);
627 }
628 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
629
630 static bool
631 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
632                 struct list_head *layout_list)
633 {
634         struct pnfs_layout_hdr *lo;
635         bool ret = false;
636
637         spin_lock(&inode->i_lock);
638         lo = NFS_I(inode)->layout;
639         if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
640                 pnfs_get_layout_hdr(lo);
641                 list_add(&lo->plh_bulk_destroy, layout_list);
642                 ret = true;
643         }
644         spin_unlock(&inode->i_lock);
645         return ret;
646 }
647
648 /* Caller must hold rcu_read_lock and clp->cl_lock */
649 static int
650 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
651                 struct nfs_server *server,
652                 struct list_head *layout_list)
653 {
654         struct pnfs_layout_hdr *lo, *next;
655         struct inode *inode;
656
657         list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
658                 inode = igrab(lo->plh_inode);
659                 if (inode == NULL)
660                         continue;
661                 list_del_init(&lo->plh_layouts);
662                 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
663                         continue;
664                 rcu_read_unlock();
665                 spin_unlock(&clp->cl_lock);
666                 iput(inode);
667                 spin_lock(&clp->cl_lock);
668                 rcu_read_lock();
669                 return -EAGAIN;
670         }
671         return 0;
672 }
673
674 static int
675 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
676                 bool is_bulk_recall)
677 {
678         struct pnfs_layout_hdr *lo;
679         struct inode *inode;
680         struct pnfs_layout_range range = {
681                 .iomode = IOMODE_ANY,
682                 .offset = 0,
683                 .length = NFS4_MAX_UINT64,
684         };
685         LIST_HEAD(lseg_list);
686         int ret = 0;
687
688         while (!list_empty(layout_list)) {
689                 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
690                                 plh_bulk_destroy);
691                 dprintk("%s freeing layout for inode %lu\n", __func__,
692                         lo->plh_inode->i_ino);
693                 inode = lo->plh_inode;
694
695                 pnfs_layoutcommit_inode(inode, false);
696
697                 spin_lock(&inode->i_lock);
698                 list_del_init(&lo->plh_bulk_destroy);
699                 lo->plh_block_lgets++; /* permanently block new LAYOUTGETs */
700                 if (is_bulk_recall)
701                         set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
702                 if (pnfs_mark_matching_lsegs_invalid(lo, &lseg_list, &range))
703                         ret = -EAGAIN;
704                 spin_unlock(&inode->i_lock);
705                 pnfs_free_lseg_list(&lseg_list);
706                 pnfs_put_layout_hdr(lo);
707                 iput(inode);
708         }
709         return ret;
710 }
711
712 int
713 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
714                 struct nfs_fsid *fsid,
715                 bool is_recall)
716 {
717         struct nfs_server *server;
718         LIST_HEAD(layout_list);
719
720         spin_lock(&clp->cl_lock);
721         rcu_read_lock();
722 restart:
723         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
724                 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
725                         continue;
726                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
727                                 server,
728                                 &layout_list) != 0)
729                         goto restart;
730         }
731         rcu_read_unlock();
732         spin_unlock(&clp->cl_lock);
733
734         if (list_empty(&layout_list))
735                 return 0;
736         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
737 }
738
739 int
740 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
741                 bool is_recall)
742 {
743         struct nfs_server *server;
744         LIST_HEAD(layout_list);
745
746         spin_lock(&clp->cl_lock);
747         rcu_read_lock();
748 restart:
749         list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
750                 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
751                                         server,
752                                         &layout_list) != 0)
753                         goto restart;
754         }
755         rcu_read_unlock();
756         spin_unlock(&clp->cl_lock);
757
758         if (list_empty(&layout_list))
759                 return 0;
760         return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
761 }
762
763 /*
764  * Called by the state manger to remove all layouts established under an
765  * expired lease.
766  */
767 void
768 pnfs_destroy_all_layouts(struct nfs_client *clp)
769 {
770         nfs4_deviceid_mark_client_invalid(clp);
771         nfs4_deviceid_purge_client(clp);
772
773         pnfs_destroy_layouts_byclid(clp, false);
774 }
775
776 /*
777  * Compare 2 layout stateid sequence ids, to see which is newer,
778  * taking into account wraparound issues.
779  */
780 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
781 {
782         return (s32)(s1 - s2) > 0;
783 }
784
785 /* update lo->plh_stateid with new if is more recent */
786 void
787 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
788                         bool update_barrier)
789 {
790         u32 oldseq, newseq, new_barrier;
791         int empty = list_empty(&lo->plh_segs);
792
793         oldseq = be32_to_cpu(lo->plh_stateid.seqid);
794         newseq = be32_to_cpu(new->seqid);
795         if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
796                 nfs4_stateid_copy(&lo->plh_stateid, new);
797                 if (update_barrier) {
798                         new_barrier = be32_to_cpu(new->seqid);
799                 } else {
800                         /* Because of wraparound, we want to keep the barrier
801                          * "close" to the current seqids.
802                          */
803                         new_barrier = newseq - atomic_read(&lo->plh_outstanding);
804                 }
805                 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
806                         lo->plh_barrier = new_barrier;
807         }
808 }
809
810 static bool
811 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
812                 const nfs4_stateid *stateid)
813 {
814         u32 seqid = be32_to_cpu(stateid->seqid);
815
816         return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
817 }
818
819 /* lget is set to 1 if called from inside send_layoutget call chain */
820 static bool
821 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
822 {
823         return lo->plh_block_lgets ||
824                 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
825 }
826
827 int
828 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
829                               struct pnfs_layout_range *range,
830                               struct nfs4_state *open_state)
831 {
832         int status = 0;
833
834         dprintk("--> %s\n", __func__);
835         spin_lock(&lo->plh_inode->i_lock);
836         if (pnfs_layoutgets_blocked(lo)) {
837                 status = -EAGAIN;
838         } else if (!nfs4_valid_open_stateid(open_state)) {
839                 status = -EBADF;
840         } else if (list_empty(&lo->plh_segs) ||
841                    test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
842                 int seq;
843
844                 do {
845                         seq = read_seqbegin(&open_state->seqlock);
846                         nfs4_stateid_copy(dst, &open_state->stateid);
847                 } while (read_seqretry(&open_state->seqlock, seq));
848         } else
849                 nfs4_stateid_copy(dst, &lo->plh_stateid);
850         spin_unlock(&lo->plh_inode->i_lock);
851         dprintk("<-- %s\n", __func__);
852         return status;
853 }
854
855 /*
856 * Get layout from server.
857 *    for now, assume that whole file layouts are requested.
858 *    arg->offset: 0
859 *    arg->length: all ones
860 */
861 static struct pnfs_layout_segment *
862 send_layoutget(struct pnfs_layout_hdr *lo,
863            struct nfs_open_context *ctx,
864            struct pnfs_layout_range *range,
865            gfp_t gfp_flags)
866 {
867         struct inode *ino = lo->plh_inode;
868         struct nfs_server *server = NFS_SERVER(ino);
869         struct nfs4_layoutget *lgp;
870         struct pnfs_layout_segment *lseg;
871         loff_t i_size;
872
873         dprintk("--> %s\n", __func__);
874
875         lgp = kzalloc(sizeof(*lgp), gfp_flags);
876         if (lgp == NULL)
877                 return NULL;
878
879         i_size = i_size_read(ino);
880
881         lgp->args.minlength = PAGE_CACHE_SIZE;
882         if (lgp->args.minlength > range->length)
883                 lgp->args.minlength = range->length;
884         if (range->iomode == IOMODE_READ) {
885                 if (range->offset >= i_size)
886                         lgp->args.minlength = 0;
887                 else if (i_size - range->offset < lgp->args.minlength)
888                         lgp->args.minlength = i_size - range->offset;
889         }
890         lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
891         lgp->args.range = *range;
892         lgp->args.type = server->pnfs_curr_ld->id;
893         lgp->args.inode = ino;
894         lgp->args.ctx = get_nfs_open_context(ctx);
895         lgp->gfp_flags = gfp_flags;
896         lgp->cred = lo->plh_lc_cred;
897
898         /* Synchronously retrieve layout information from server and
899          * store in lseg.
900          */
901         lseg = nfs4_proc_layoutget(lgp, gfp_flags);
902         if (IS_ERR(lseg)) {
903                 switch (PTR_ERR(lseg)) {
904                 case -ENOMEM:
905                 case -ERESTARTSYS:
906                         break;
907                 default:
908                         /* remember that LAYOUTGET failed and suspend trying */
909                         pnfs_layout_io_set_failed(lo, range->iomode);
910                 }
911                 return NULL;
912         } else
913                 pnfs_layout_clear_fail_bit(lo,
914                                 pnfs_iomode_to_fail_bit(range->iomode));
915
916         return lseg;
917 }
918
919 static void pnfs_clear_layoutcommit(struct inode *inode,
920                 struct list_head *head)
921 {
922         struct nfs_inode *nfsi = NFS_I(inode);
923         struct pnfs_layout_segment *lseg, *tmp;
924
925         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
926                 return;
927         list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
928                 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
929                         continue;
930                 pnfs_lseg_dec_and_remove_zero(lseg, head);
931         }
932 }
933
934 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
935 {
936         clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
937         smp_mb__after_atomic();
938         wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
939         rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
940 }
941
942 static int
943 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, nfs4_stateid stateid,
944                        enum pnfs_iomode iomode, bool sync)
945 {
946         struct inode *ino = lo->plh_inode;
947         struct nfs4_layoutreturn *lrp;
948         int status = 0;
949
950         lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
951         if (unlikely(lrp == NULL)) {
952                 status = -ENOMEM;
953                 spin_lock(&ino->i_lock);
954                 pnfs_clear_layoutreturn_waitbit(lo);
955                 spin_unlock(&ino->i_lock);
956                 pnfs_put_layout_hdr(lo);
957                 goto out;
958         }
959
960         lrp->args.stateid = stateid;
961         lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
962         lrp->args.inode = ino;
963         lrp->args.range.iomode = iomode;
964         lrp->args.range.offset = 0;
965         lrp->args.range.length = NFS4_MAX_UINT64;
966         lrp->args.layout = lo;
967         lrp->clp = NFS_SERVER(ino)->nfs_client;
968         lrp->cred = lo->plh_lc_cred;
969
970         status = nfs4_proc_layoutreturn(lrp, sync);
971 out:
972         dprintk("<-- %s status: %d\n", __func__, status);
973         return status;
974 }
975
976 /*
977  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
978  * when the layout segment list is empty.
979  *
980  * Note that a pnfs_layout_hdr can exist with an empty layout segment
981  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
982  * deviceid is marked invalid.
983  */
984 int
985 _pnfs_return_layout(struct inode *ino)
986 {
987         struct pnfs_layout_hdr *lo = NULL;
988         struct nfs_inode *nfsi = NFS_I(ino);
989         LIST_HEAD(tmp_list);
990         nfs4_stateid stateid;
991         int status = 0, empty;
992         bool send;
993
994         dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
995
996         spin_lock(&ino->i_lock);
997         lo = nfsi->layout;
998         if (!lo) {
999                 spin_unlock(&ino->i_lock);
1000                 dprintk("NFS: %s no layout to return\n", __func__);
1001                 goto out;
1002         }
1003         stateid = nfsi->layout->plh_stateid;
1004         /* Reference matched in nfs4_layoutreturn_release */
1005         pnfs_get_layout_hdr(lo);
1006         empty = list_empty(&lo->plh_segs);
1007         pnfs_clear_layoutcommit(ino, &tmp_list);
1008         pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1009
1010         if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1011                 struct pnfs_layout_range range = {
1012                         .iomode         = IOMODE_ANY,
1013                         .offset         = 0,
1014                         .length         = NFS4_MAX_UINT64,
1015                 };
1016                 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1017         }
1018
1019         /* Don't send a LAYOUTRETURN if list was initially empty */
1020         if (empty) {
1021                 spin_unlock(&ino->i_lock);
1022                 dprintk("NFS: %s no layout segments to return\n", __func__);
1023                 goto out_put_layout_hdr;
1024         }
1025
1026         set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1027         send = pnfs_prepare_layoutreturn(lo);
1028         spin_unlock(&ino->i_lock);
1029         pnfs_free_lseg_list(&tmp_list);
1030         if (send)
1031                 status = pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1032 out_put_layout_hdr:
1033         pnfs_put_layout_hdr(lo);
1034 out:
1035         dprintk("<-- %s status: %d\n", __func__, status);
1036         return status;
1037 }
1038 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1039
1040 int
1041 pnfs_commit_and_return_layout(struct inode *inode)
1042 {
1043         struct pnfs_layout_hdr *lo;
1044         int ret;
1045
1046         spin_lock(&inode->i_lock);
1047         lo = NFS_I(inode)->layout;
1048         if (lo == NULL) {
1049                 spin_unlock(&inode->i_lock);
1050                 return 0;
1051         }
1052         pnfs_get_layout_hdr(lo);
1053         /* Block new layoutgets and read/write to ds */
1054         lo->plh_block_lgets++;
1055         spin_unlock(&inode->i_lock);
1056         filemap_fdatawait(inode->i_mapping);
1057         ret = pnfs_layoutcommit_inode(inode, true);
1058         if (ret == 0)
1059                 ret = _pnfs_return_layout(inode);
1060         spin_lock(&inode->i_lock);
1061         lo->plh_block_lgets--;
1062         spin_unlock(&inode->i_lock);
1063         pnfs_put_layout_hdr(lo);
1064         return ret;
1065 }
1066
1067 bool pnfs_roc(struct inode *ino)
1068 {
1069         struct nfs_inode *nfsi = NFS_I(ino);
1070         struct nfs_open_context *ctx;
1071         struct nfs4_state *state;
1072         struct pnfs_layout_hdr *lo;
1073         struct pnfs_layout_segment *lseg, *tmp;
1074         nfs4_stateid stateid;
1075         LIST_HEAD(tmp_list);
1076         bool found = false, layoutreturn = false, roc = false;
1077
1078         spin_lock(&ino->i_lock);
1079         lo = nfsi->layout;
1080         if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1081                 goto out_noroc;
1082
1083         /* no roc if we hold a delegation */
1084         if (nfs4_check_delegation(ino, FMODE_READ))
1085                 goto out_noroc;
1086
1087         list_for_each_entry(ctx, &nfsi->open_files, list) {
1088                 state = ctx->state;
1089                 /* Don't return layout if there is open file state */
1090                 if (state != NULL && state->state != 0)
1091                         goto out_noroc;
1092         }
1093
1094         stateid = lo->plh_stateid;
1095         /* always send layoutreturn if being marked so */
1096         if (test_and_clear_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1097                                    &lo->plh_flags))
1098                 layoutreturn = pnfs_prepare_layoutreturn(lo);
1099
1100         pnfs_clear_retry_layoutget(lo);
1101         list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1102                 /* If we are sending layoutreturn, invalidate all valid lsegs */
1103                 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1104                         mark_lseg_invalid(lseg, &tmp_list);
1105                         found = true;
1106                 }
1107         /* ROC in two conditions:
1108          * 1. there are ROC lsegs
1109          * 2. we don't send layoutreturn
1110          */
1111         if (found && !layoutreturn) {
1112                 /* lo ref dropped in pnfs_roc_release() */
1113                 pnfs_get_layout_hdr(lo);
1114                 roc = true;
1115         }
1116
1117 out_noroc:
1118         spin_unlock(&ino->i_lock);
1119         pnfs_free_lseg_list(&tmp_list);
1120         pnfs_layoutcommit_inode(ino, true);
1121         if (layoutreturn)
1122                 pnfs_send_layoutreturn(lo, stateid, IOMODE_ANY, true);
1123         return roc;
1124 }
1125
1126 void pnfs_roc_release(struct inode *ino)
1127 {
1128         struct pnfs_layout_hdr *lo;
1129
1130         spin_lock(&ino->i_lock);
1131         lo = NFS_I(ino)->layout;
1132         pnfs_clear_layoutreturn_waitbit(lo);
1133         if (atomic_dec_and_test(&lo->plh_refcount)) {
1134                 pnfs_detach_layout_hdr(lo);
1135                 spin_unlock(&ino->i_lock);
1136                 pnfs_free_layout_hdr(lo);
1137         } else
1138                 spin_unlock(&ino->i_lock);
1139 }
1140
1141 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1142 {
1143         struct pnfs_layout_hdr *lo;
1144
1145         spin_lock(&ino->i_lock);
1146         lo = NFS_I(ino)->layout;
1147         if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1148                 lo->plh_barrier = barrier;
1149         spin_unlock(&ino->i_lock);
1150         trace_nfs4_layoutreturn_on_close(ino, 0);
1151 }
1152
1153 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1154 {
1155         struct nfs_inode *nfsi = NFS_I(ino);
1156         struct pnfs_layout_hdr *lo;
1157         u32 current_seqid;
1158
1159         spin_lock(&ino->i_lock);
1160         lo = nfsi->layout;
1161         current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1162
1163         /* Since close does not return a layout stateid for use as
1164          * a barrier, we choose the worst-case barrier.
1165          */
1166         *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1167         spin_unlock(&ino->i_lock);
1168 }
1169
1170 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1171 {
1172         struct nfs_inode *nfsi = NFS_I(ino);
1173         struct pnfs_layout_hdr *lo;
1174         bool sleep = false;
1175
1176         /* we might not have grabbed lo reference. so need to check under
1177          * i_lock */
1178         spin_lock(&ino->i_lock);
1179         lo = nfsi->layout;
1180         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1181                 sleep = true;
1182         spin_unlock(&ino->i_lock);
1183
1184         if (sleep)
1185                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1186
1187         return sleep;
1188 }
1189
1190 /*
1191  * Compare two layout segments for sorting into layout cache.
1192  * We want to preferentially return RW over RO layouts, so ensure those
1193  * are seen first.
1194  */
1195 static s64
1196 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1197            const struct pnfs_layout_range *l2)
1198 {
1199         s64 d;
1200
1201         /* high offset > low offset */
1202         d = l1->offset - l2->offset;
1203         if (d)
1204                 return d;
1205
1206         /* short length > long length */
1207         d = l2->length - l1->length;
1208         if (d)
1209                 return d;
1210
1211         /* read > read/write */
1212         return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1213 }
1214
1215 static bool
1216 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1217                 const struct pnfs_layout_range *l2)
1218 {
1219         return pnfs_lseg_range_cmp(l1, l2) > 0;
1220 }
1221
1222 static bool
1223 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1224                 struct pnfs_layout_segment *old)
1225 {
1226         return false;
1227 }
1228
1229 void
1230 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1231                    struct pnfs_layout_segment *lseg,
1232                    bool (*is_after)(const struct pnfs_layout_range *,
1233                            const struct pnfs_layout_range *),
1234                    bool (*do_merge)(struct pnfs_layout_segment *,
1235                            struct pnfs_layout_segment *),
1236                    struct list_head *free_me)
1237 {
1238         struct pnfs_layout_segment *lp, *tmp;
1239
1240         dprintk("%s:Begin\n", __func__);
1241
1242         list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1243                 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1244                         continue;
1245                 if (do_merge(lseg, lp)) {
1246                         mark_lseg_invalid(lp, free_me);
1247                         continue;
1248                 }
1249                 if (is_after(&lseg->pls_range, &lp->pls_range))
1250                         continue;
1251                 list_add_tail(&lseg->pls_list, &lp->pls_list);
1252                 dprintk("%s: inserted lseg %p "
1253                         "iomode %d offset %llu length %llu before "
1254                         "lp %p iomode %d offset %llu length %llu\n",
1255                         __func__, lseg, lseg->pls_range.iomode,
1256                         lseg->pls_range.offset, lseg->pls_range.length,
1257                         lp, lp->pls_range.iomode, lp->pls_range.offset,
1258                         lp->pls_range.length);
1259                 goto out;
1260         }
1261         list_add_tail(&lseg->pls_list, &lo->plh_segs);
1262         dprintk("%s: inserted lseg %p "
1263                 "iomode %d offset %llu length %llu at tail\n",
1264                 __func__, lseg, lseg->pls_range.iomode,
1265                 lseg->pls_range.offset, lseg->pls_range.length);
1266 out:
1267         pnfs_get_layout_hdr(lo);
1268
1269         dprintk("%s:Return\n", __func__);
1270 }
1271 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1272
1273 static void
1274 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1275                    struct pnfs_layout_segment *lseg,
1276                    struct list_head *free_me)
1277 {
1278         struct inode *inode = lo->plh_inode;
1279         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1280
1281         if (ld->add_lseg != NULL)
1282                 ld->add_lseg(lo, lseg, free_me);
1283         else
1284                 pnfs_generic_layout_insert_lseg(lo, lseg,
1285                                 pnfs_lseg_range_is_after,
1286                                 pnfs_lseg_no_merge,
1287                                 free_me);
1288 }
1289
1290 static struct pnfs_layout_hdr *
1291 alloc_init_layout_hdr(struct inode *ino,
1292                       struct nfs_open_context *ctx,
1293                       gfp_t gfp_flags)
1294 {
1295         struct pnfs_layout_hdr *lo;
1296
1297         lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1298         if (!lo)
1299                 return NULL;
1300         atomic_set(&lo->plh_refcount, 1);
1301         INIT_LIST_HEAD(&lo->plh_layouts);
1302         INIT_LIST_HEAD(&lo->plh_segs);
1303         INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1304         lo->plh_inode = ino;
1305         lo->plh_lc_cred = get_rpccred(ctx->cred);
1306         return lo;
1307 }
1308
1309 static struct pnfs_layout_hdr *
1310 pnfs_find_alloc_layout(struct inode *ino,
1311                        struct nfs_open_context *ctx,
1312                        gfp_t gfp_flags)
1313 {
1314         struct nfs_inode *nfsi = NFS_I(ino);
1315         struct pnfs_layout_hdr *new = NULL;
1316
1317         dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1318
1319         if (nfsi->layout != NULL)
1320                 goto out_existing;
1321         spin_unlock(&ino->i_lock);
1322         new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1323         spin_lock(&ino->i_lock);
1324
1325         if (likely(nfsi->layout == NULL)) {     /* Won the race? */
1326                 nfsi->layout = new;
1327                 return new;
1328         } else if (new != NULL)
1329                 pnfs_free_layout_hdr(new);
1330 out_existing:
1331         pnfs_get_layout_hdr(nfsi->layout);
1332         return nfsi->layout;
1333 }
1334
1335 /*
1336  * iomode matching rules:
1337  * iomode       lseg    match
1338  * -----        -----   -----
1339  * ANY          READ    true
1340  * ANY          RW      true
1341  * RW           READ    false
1342  * RW           RW      true
1343  * READ         READ    true
1344  * READ         RW      true
1345  */
1346 static bool
1347 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1348                  const struct pnfs_layout_range *range)
1349 {
1350         struct pnfs_layout_range range1;
1351
1352         if ((range->iomode == IOMODE_RW &&
1353              ls_range->iomode != IOMODE_RW) ||
1354             !pnfs_lseg_range_intersecting(ls_range, range))
1355                 return 0;
1356
1357         /* range1 covers only the first byte in the range */
1358         range1 = *range;
1359         range1.length = 1;
1360         return pnfs_lseg_range_contained(ls_range, &range1);
1361 }
1362
1363 /*
1364  * lookup range in layout
1365  */
1366 static struct pnfs_layout_segment *
1367 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1368                 struct pnfs_layout_range *range)
1369 {
1370         struct pnfs_layout_segment *lseg, *ret = NULL;
1371
1372         dprintk("%s:Begin\n", __func__);
1373
1374         list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1375                 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1376                     !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1377                     pnfs_lseg_range_match(&lseg->pls_range, range)) {
1378                         ret = pnfs_get_lseg(lseg);
1379                         break;
1380                 }
1381         }
1382
1383         dprintk("%s:Return lseg %p ref %d\n",
1384                 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1385         return ret;
1386 }
1387
1388 /*
1389  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1390  * to the MDS or over pNFS
1391  *
1392  * The nfs_inode read_io and write_io fields are cumulative counters reset
1393  * when there are no layout segments. Note that in pnfs_update_layout iomode
1394  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1395  * WRITE request.
1396  *
1397  * A return of true means use MDS I/O.
1398  *
1399  * From rfc 5661:
1400  * If a file's size is smaller than the file size threshold, data accesses
1401  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1402  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1403  * server.  If both file size and I/O size are provided, the client SHOULD
1404  * reach or exceed  both thresholds before sending its read or write
1405  * requests to the data server.
1406  */
1407 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1408                                      struct inode *ino, int iomode)
1409 {
1410         struct nfs4_threshold *t = ctx->mdsthreshold;
1411         struct nfs_inode *nfsi = NFS_I(ino);
1412         loff_t fsize = i_size_read(ino);
1413         bool size = false, size_set = false, io = false, io_set = false, ret = false;
1414
1415         if (t == NULL)
1416                 return ret;
1417
1418         dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1419                 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1420
1421         switch (iomode) {
1422         case IOMODE_READ:
1423                 if (t->bm & THRESHOLD_RD) {
1424                         dprintk("%s fsize %llu\n", __func__, fsize);
1425                         size_set = true;
1426                         if (fsize < t->rd_sz)
1427                                 size = true;
1428                 }
1429                 if (t->bm & THRESHOLD_RD_IO) {
1430                         dprintk("%s nfsi->read_io %llu\n", __func__,
1431                                 nfsi->read_io);
1432                         io_set = true;
1433                         if (nfsi->read_io < t->rd_io_sz)
1434                                 io = true;
1435                 }
1436                 break;
1437         case IOMODE_RW:
1438                 if (t->bm & THRESHOLD_WR) {
1439                         dprintk("%s fsize %llu\n", __func__, fsize);
1440                         size_set = true;
1441                         if (fsize < t->wr_sz)
1442                                 size = true;
1443                 }
1444                 if (t->bm & THRESHOLD_WR_IO) {
1445                         dprintk("%s nfsi->write_io %llu\n", __func__,
1446                                 nfsi->write_io);
1447                         io_set = true;
1448                         if (nfsi->write_io < t->wr_io_sz)
1449                                 io = true;
1450                 }
1451                 break;
1452         }
1453         if (size_set && io_set) {
1454                 if (size && io)
1455                         ret = true;
1456         } else if (size || io)
1457                 ret = true;
1458
1459         dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1460         return ret;
1461 }
1462
1463 /* stop waiting if someone clears NFS_LAYOUT_RETRY_LAYOUTGET bit. */
1464 static int pnfs_layoutget_retry_bit_wait(struct wait_bit_key *key)
1465 {
1466         if (!test_bit(NFS_LAYOUT_RETRY_LAYOUTGET, key->flags))
1467                 return 1;
1468         return nfs_wait_bit_killable(key);
1469 }
1470
1471 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1472 {
1473         if (!pnfs_should_retry_layoutget(lo))
1474                 return false;
1475         /*
1476          * send layoutcommit as it can hold up layoutreturn due to lseg
1477          * reference
1478          */
1479         pnfs_layoutcommit_inode(lo->plh_inode, false);
1480         return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1481                                    pnfs_layoutget_retry_bit_wait,
1482                                    TASK_UNINTERRUPTIBLE);
1483 }
1484
1485 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1486 {
1487         unsigned long *bitlock = &lo->plh_flags;
1488
1489         clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1490         smp_mb__after_atomic();
1491         wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1492 }
1493
1494 /*
1495  * Layout segment is retreived from the server if not cached.
1496  * The appropriate layout segment is referenced and returned to the caller.
1497  */
1498 struct pnfs_layout_segment *
1499 pnfs_update_layout(struct inode *ino,
1500                    struct nfs_open_context *ctx,
1501                    loff_t pos,
1502                    u64 count,
1503                    enum pnfs_iomode iomode,
1504                    gfp_t gfp_flags)
1505 {
1506         struct pnfs_layout_range arg = {
1507                 .iomode = iomode,
1508                 .offset = pos,
1509                 .length = count,
1510         };
1511         unsigned pg_offset;
1512         struct nfs_server *server = NFS_SERVER(ino);
1513         struct nfs_client *clp = server->nfs_client;
1514         struct pnfs_layout_hdr *lo;
1515         struct pnfs_layout_segment *lseg = NULL;
1516         bool first;
1517
1518         if (!pnfs_enabled_sb(NFS_SERVER(ino)))
1519                 goto out;
1520
1521         if (iomode == IOMODE_READ && i_size_read(ino) == 0)
1522                 goto out;
1523
1524         if (pnfs_within_mdsthreshold(ctx, ino, iomode))
1525                 goto out;
1526
1527 lookup_again:
1528         first = false;
1529         spin_lock(&ino->i_lock);
1530         lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1531         if (lo == NULL) {
1532                 spin_unlock(&ino->i_lock);
1533                 goto out;
1534         }
1535
1536         /* Do we even need to bother with this? */
1537         if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1538                 dprintk("%s matches recall, use MDS\n", __func__);
1539                 goto out_unlock;
1540         }
1541
1542         /* if LAYOUTGET already failed once we don't try again */
1543         if (pnfs_layout_io_test_failed(lo, iomode) &&
1544             !pnfs_should_retry_layoutget(lo))
1545                 goto out_unlock;
1546
1547         first = list_empty(&lo->plh_segs);
1548         if (first) {
1549                 /* The first layoutget for the file. Need to serialize per
1550                  * RFC 5661 Errata 3208.
1551                  */
1552                 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1553                                      &lo->plh_flags)) {
1554                         spin_unlock(&ino->i_lock);
1555                         wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1556                                     TASK_UNINTERRUPTIBLE);
1557                         pnfs_put_layout_hdr(lo);
1558                         goto lookup_again;
1559                 }
1560         } else {
1561                 /* Check to see if the layout for the given range
1562                  * already exists
1563                  */
1564                 lseg = pnfs_find_lseg(lo, &arg);
1565                 if (lseg)
1566                         goto out_unlock;
1567         }
1568
1569         /*
1570          * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1571          * for LAYOUTRETURN even if first is true.
1572          */
1573         if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1574                 spin_unlock(&ino->i_lock);
1575                 dprintk("%s wait for layoutreturn\n", __func__);
1576                 if (pnfs_prepare_to_retry_layoutget(lo)) {
1577                         if (first)
1578                                 pnfs_clear_first_layoutget(lo);
1579                         pnfs_put_layout_hdr(lo);
1580                         dprintk("%s retrying\n", __func__);
1581                         goto lookup_again;
1582                 }
1583                 goto out_put_layout_hdr;
1584         }
1585
1586         if (pnfs_layoutgets_blocked(lo))
1587                 goto out_unlock;
1588         atomic_inc(&lo->plh_outstanding);
1589         spin_unlock(&ino->i_lock);
1590
1591         if (list_empty(&lo->plh_layouts)) {
1592                 /* The lo must be on the clp list if there is any
1593                  * chance of a CB_LAYOUTRECALL(FILE) coming in.
1594                  */
1595                 spin_lock(&clp->cl_lock);
1596                 if (list_empty(&lo->plh_layouts))
1597                         list_add_tail(&lo->plh_layouts, &server->layouts);
1598                 spin_unlock(&clp->cl_lock);
1599         }
1600
1601         pg_offset = arg.offset & ~PAGE_CACHE_MASK;
1602         if (pg_offset) {
1603                 arg.offset -= pg_offset;
1604                 arg.length += pg_offset;
1605         }
1606         if (arg.length != NFS4_MAX_UINT64)
1607                 arg.length = PAGE_CACHE_ALIGN(arg.length);
1608
1609         lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1610         pnfs_clear_retry_layoutget(lo);
1611         atomic_dec(&lo->plh_outstanding);
1612 out_put_layout_hdr:
1613         if (first)
1614                 pnfs_clear_first_layoutget(lo);
1615         pnfs_put_layout_hdr(lo);
1616 out:
1617         dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1618                         "(%s, offset: %llu, length: %llu)\n",
1619                         __func__, ino->i_sb->s_id,
1620                         (unsigned long long)NFS_FILEID(ino),
1621                         lseg == NULL ? "not found" : "found",
1622                         iomode==IOMODE_RW ?  "read/write" : "read-only",
1623                         (unsigned long long)pos,
1624                         (unsigned long long)count);
1625         return lseg;
1626 out_unlock:
1627         spin_unlock(&ino->i_lock);
1628         goto out_put_layout_hdr;
1629 }
1630 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1631
1632 static bool
1633 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1634 {
1635         switch (range->iomode) {
1636         case IOMODE_READ:
1637         case IOMODE_RW:
1638                 break;
1639         default:
1640                 return false;
1641         }
1642         if (range->offset == NFS4_MAX_UINT64)
1643                 return false;
1644         if (range->length == 0)
1645                 return false;
1646         if (range->length != NFS4_MAX_UINT64 &&
1647             range->length > NFS4_MAX_UINT64 - range->offset)
1648                 return false;
1649         return true;
1650 }
1651
1652 struct pnfs_layout_segment *
1653 pnfs_layout_process(struct nfs4_layoutget *lgp)
1654 {
1655         struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1656         struct nfs4_layoutget_res *res = &lgp->res;
1657         struct pnfs_layout_segment *lseg;
1658         struct inode *ino = lo->plh_inode;
1659         LIST_HEAD(free_me);
1660         int status = -EINVAL;
1661
1662         if (!pnfs_sanity_check_layout_range(&res->range))
1663                 goto out;
1664
1665         /* Inject layout blob into I/O device driver */
1666         lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1667         if (!lseg || IS_ERR(lseg)) {
1668                 if (!lseg)
1669                         status = -ENOMEM;
1670                 else
1671                         status = PTR_ERR(lseg);
1672                 dprintk("%s: Could not allocate layout: error %d\n",
1673                        __func__, status);
1674                 goto out;
1675         }
1676
1677         init_lseg(lo, lseg);
1678         lseg->pls_range = res->range;
1679
1680         spin_lock(&ino->i_lock);
1681         if (pnfs_layoutgets_blocked(lo)) {
1682                 dprintk("%s forget reply due to state\n", __func__);
1683                 goto out_forget_reply;
1684         }
1685
1686         if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1687                 /* existing state ID, make sure the sequence number matches. */
1688                 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1689                         dprintk("%s forget reply due to sequence\n", __func__);
1690                         goto out_forget_reply;
1691                 }
1692                 pnfs_set_layout_stateid(lo, &res->stateid, false);
1693         } else {
1694                 /*
1695                  * We got an entirely new state ID.  Mark all segments for the
1696                  * inode invalid, and don't bother validating the stateid
1697                  * sequence number.
1698                  */
1699                 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1700
1701                 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1702                 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1703         }
1704
1705         clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1706
1707         pnfs_get_lseg(lseg);
1708         pnfs_layout_insert_lseg(lo, lseg, &free_me);
1709
1710         if (res->return_on_close)
1711                 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1712
1713         spin_unlock(&ino->i_lock);
1714         pnfs_free_lseg_list(&free_me);
1715         return lseg;
1716 out:
1717         return ERR_PTR(status);
1718
1719 out_forget_reply:
1720         spin_unlock(&ino->i_lock);
1721         lseg->pls_layout = lo;
1722         NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1723         goto out;
1724 }
1725
1726 static void
1727 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1728                                 struct list_head *tmp_list,
1729                                 struct pnfs_layout_range *return_range)
1730 {
1731         struct pnfs_layout_segment *lseg, *next;
1732
1733         dprintk("%s:Begin lo %p\n", __func__, lo);
1734
1735         if (list_empty(&lo->plh_segs))
1736                 return;
1737
1738         list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1739                 if (should_free_lseg(&lseg->pls_range, return_range)) {
1740                         dprintk("%s: marking lseg %p iomode %d "
1741                                 "offset %llu length %llu\n", __func__,
1742                                 lseg, lseg->pls_range.iomode,
1743                                 lseg->pls_range.offset,
1744                                 lseg->pls_range.length);
1745                         set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1746                         mark_lseg_invalid(lseg, tmp_list);
1747                         set_bit(NFS_LAYOUT_RETURN_BEFORE_CLOSE,
1748                                         &lo->plh_flags);
1749                 }
1750 }
1751
1752 void pnfs_error_mark_layout_for_return(struct inode *inode,
1753                                        struct pnfs_layout_segment *lseg)
1754 {
1755         struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1756         int iomode = pnfs_iomode_to_fail_bit(lseg->pls_range.iomode);
1757         struct pnfs_layout_range range = {
1758                 .iomode = lseg->pls_range.iomode,
1759                 .offset = 0,
1760                 .length = NFS4_MAX_UINT64,
1761         };
1762         LIST_HEAD(free_me);
1763
1764         spin_lock(&inode->i_lock);
1765         /* set failure bit so that pnfs path will be retried later */
1766         pnfs_layout_set_fail_bit(lo, iomode);
1767         if (lo->plh_return_iomode == 0)
1768                 lo->plh_return_iomode = range.iomode;
1769         else if (lo->plh_return_iomode != range.iomode)
1770                 lo->plh_return_iomode = IOMODE_ANY;
1771         /*
1772          * mark all matching lsegs so that we are sure to have no live
1773          * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1774          * for how it works.
1775          */
1776         pnfs_mark_matching_lsegs_return(lo, &free_me, &range);
1777         spin_unlock(&inode->i_lock);
1778         pnfs_free_lseg_list(&free_me);
1779 }
1780 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1781
1782 void
1783 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1784 {
1785         u64 rd_size = req->wb_bytes;
1786
1787         if (pgio->pg_lseg == NULL) {
1788                 if (pgio->pg_dreq == NULL)
1789                         rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1790                 else
1791                         rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1792
1793                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1794                                                    req->wb_context,
1795                                                    req_offset(req),
1796                                                    rd_size,
1797                                                    IOMODE_READ,
1798                                                    GFP_KERNEL);
1799         }
1800         /* If no lseg, fall back to read through mds */
1801         if (pgio->pg_lseg == NULL)
1802                 nfs_pageio_reset_read_mds(pgio);
1803
1804 }
1805 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1806
1807 void
1808 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1809                            struct nfs_page *req, u64 wb_size)
1810 {
1811         if (pgio->pg_lseg == NULL)
1812                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1813                                                    req->wb_context,
1814                                                    req_offset(req),
1815                                                    wb_size,
1816                                                    IOMODE_RW,
1817                                                    GFP_NOFS);
1818         /* If no lseg, fall back to write through mds */
1819         if (pgio->pg_lseg == NULL)
1820                 nfs_pageio_reset_write_mds(pgio);
1821 }
1822 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1823
1824 void
1825 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1826 {
1827         if (desc->pg_lseg) {
1828                 pnfs_put_lseg(desc->pg_lseg);
1829                 desc->pg_lseg = NULL;
1830         }
1831 }
1832 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1833
1834 /*
1835  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1836  * of bytes (maximum @req->wb_bytes) that can be coalesced.
1837  */
1838 size_t
1839 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1840                      struct nfs_page *prev, struct nfs_page *req)
1841 {
1842         unsigned int size;
1843         u64 seg_end, req_start, seg_left;
1844
1845         size = nfs_generic_pg_test(pgio, prev, req);
1846         if (!size)
1847                 return 0;
1848
1849         /*
1850          * 'size' contains the number of bytes left in the current page (up
1851          * to the original size asked for in @req->wb_bytes).
1852          *
1853          * Calculate how many bytes are left in the layout segment
1854          * and if there are less bytes than 'size', return that instead.
1855          *
1856          * Please also note that 'end_offset' is actually the offset of the
1857          * first byte that lies outside the pnfs_layout_range. FIXME?
1858          *
1859          */
1860         if (pgio->pg_lseg) {
1861                 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1862                                      pgio->pg_lseg->pls_range.length);
1863                 req_start = req_offset(req);
1864                 WARN_ON_ONCE(req_start >= seg_end);
1865                 /* start of request is past the last byte of this segment */
1866                 if (req_start >= seg_end) {
1867                         /* reference the new lseg */
1868                         if (pgio->pg_ops->pg_cleanup)
1869                                 pgio->pg_ops->pg_cleanup(pgio);
1870                         if (pgio->pg_ops->pg_init)
1871                                 pgio->pg_ops->pg_init(pgio, req);
1872                         return 0;
1873                 }
1874
1875                 /* adjust 'size' iff there are fewer bytes left in the
1876                  * segment than what nfs_generic_pg_test returned */
1877                 seg_left = seg_end - req_start;
1878                 if (seg_left < size)
1879                         size = (unsigned int)seg_left;
1880         }
1881
1882         return size;
1883 }
1884 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1885
1886 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1887 {
1888         struct nfs_pageio_descriptor pgio;
1889
1890         /* Resend all requests through the MDS */
1891         nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1892                               hdr->completion_ops);
1893         set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1894         return nfs_pageio_resend(&pgio, hdr);
1895 }
1896 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1897
1898 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1899 {
1900
1901         dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1902         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1903             PNFS_LAYOUTRET_ON_ERROR) {
1904                 pnfs_return_layout(hdr->inode);
1905         }
1906         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1907                 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1908 }
1909
1910 /*
1911  * Called by non rpc-based layout drivers
1912  */
1913 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1914 {
1915         if (likely(!hdr->pnfs_error)) {
1916                 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1917                                 hdr->mds_offset + hdr->res.count);
1918                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1919         }
1920         trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1921         if (unlikely(hdr->pnfs_error))
1922                 pnfs_ld_handle_write_error(hdr);
1923         hdr->mds_ops->rpc_release(hdr);
1924 }
1925 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1926
1927 static void
1928 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1929                 struct nfs_pgio_header *hdr)
1930 {
1931         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1932
1933         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1934                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
1935                 nfs_pageio_reset_write_mds(desc);
1936                 mirror->pg_recoalesce = 1;
1937         }
1938         nfs_pgio_data_destroy(hdr);
1939         hdr->release(hdr);
1940 }
1941
1942 static enum pnfs_try_status
1943 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
1944                         const struct rpc_call_ops *call_ops,
1945                         struct pnfs_layout_segment *lseg,
1946                         int how)
1947 {
1948         struct inode *inode = hdr->inode;
1949         enum pnfs_try_status trypnfs;
1950         struct nfs_server *nfss = NFS_SERVER(inode);
1951
1952         hdr->mds_ops = call_ops;
1953
1954         dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
1955                 inode->i_ino, hdr->args.count, hdr->args.offset, how);
1956         trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
1957         if (trypnfs != PNFS_NOT_ATTEMPTED)
1958                 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
1959         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
1960         return trypnfs;
1961 }
1962
1963 static void
1964 pnfs_do_write(struct nfs_pageio_descriptor *desc,
1965               struct nfs_pgio_header *hdr, int how)
1966 {
1967         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
1968         struct pnfs_layout_segment *lseg = desc->pg_lseg;
1969         enum pnfs_try_status trypnfs;
1970
1971         trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
1972         if (trypnfs == PNFS_NOT_ATTEMPTED)
1973                 pnfs_write_through_mds(desc, hdr);
1974 }
1975
1976 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
1977 {
1978         pnfs_put_lseg(hdr->lseg);
1979         nfs_pgio_header_free(hdr);
1980 }
1981
1982 int
1983 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
1984 {
1985         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1986
1987         struct nfs_pgio_header *hdr;
1988         int ret;
1989
1990         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
1991         if (!hdr) {
1992                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
1993                 return -ENOMEM;
1994         }
1995         nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
1996
1997         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
1998         ret = nfs_generic_pgio(desc, hdr);
1999         if (!ret)
2000                 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2001
2002         return ret;
2003 }
2004 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2005
2006 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2007 {
2008         struct nfs_pageio_descriptor pgio;
2009
2010         /* Resend all requests through the MDS */
2011         nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2012         return nfs_pageio_resend(&pgio, hdr);
2013 }
2014 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2015
2016 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2017 {
2018         dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2019         if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2020             PNFS_LAYOUTRET_ON_ERROR) {
2021                 pnfs_return_layout(hdr->inode);
2022         }
2023         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2024                 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2025 }
2026
2027 /*
2028  * Called by non rpc-based layout drivers
2029  */
2030 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2031 {
2032         if (likely(!hdr->pnfs_error)) {
2033                 __nfs4_read_done_cb(hdr);
2034                 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2035         }
2036         trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2037         if (unlikely(hdr->pnfs_error))
2038                 pnfs_ld_handle_read_error(hdr);
2039         hdr->mds_ops->rpc_release(hdr);
2040 }
2041 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2042
2043 static void
2044 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2045                 struct nfs_pgio_header *hdr)
2046 {
2047         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2048
2049         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2050                 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2051                 nfs_pageio_reset_read_mds(desc);
2052                 mirror->pg_recoalesce = 1;
2053         }
2054         nfs_pgio_data_destroy(hdr);
2055         hdr->release(hdr);
2056 }
2057
2058 /*
2059  * Call the appropriate parallel I/O subsystem read function.
2060  */
2061 static enum pnfs_try_status
2062 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2063                        const struct rpc_call_ops *call_ops,
2064                        struct pnfs_layout_segment *lseg)
2065 {
2066         struct inode *inode = hdr->inode;
2067         struct nfs_server *nfss = NFS_SERVER(inode);
2068         enum pnfs_try_status trypnfs;
2069
2070         hdr->mds_ops = call_ops;
2071
2072         dprintk("%s: Reading ino:%lu %u@%llu\n",
2073                 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2074
2075         trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2076         if (trypnfs != PNFS_NOT_ATTEMPTED)
2077                 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2078         dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2079         return trypnfs;
2080 }
2081
2082 /* Resend all requests through pnfs. */
2083 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2084 {
2085         struct nfs_pageio_descriptor pgio;
2086
2087         nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2088         return nfs_pageio_resend(&pgio, hdr);
2089 }
2090 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2091
2092 static void
2093 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2094 {
2095         const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2096         struct pnfs_layout_segment *lseg = desc->pg_lseg;
2097         enum pnfs_try_status trypnfs;
2098         int err = 0;
2099
2100         trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2101         if (trypnfs == PNFS_TRY_AGAIN)
2102                 err = pnfs_read_resend_pnfs(hdr);
2103         if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2104                 pnfs_read_through_mds(desc, hdr);
2105 }
2106
2107 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2108 {
2109         pnfs_put_lseg(hdr->lseg);
2110         nfs_pgio_header_free(hdr);
2111 }
2112
2113 int
2114 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2115 {
2116         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2117
2118         struct nfs_pgio_header *hdr;
2119         int ret;
2120
2121         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2122         if (!hdr) {
2123                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list);
2124                 return -ENOMEM;
2125         }
2126         nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2127         hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2128         ret = nfs_generic_pgio(desc, hdr);
2129         if (!ret)
2130                 pnfs_do_read(desc, hdr);
2131         return ret;
2132 }
2133 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2134
2135 static void pnfs_clear_layoutcommitting(struct inode *inode)
2136 {
2137         unsigned long *bitlock = &NFS_I(inode)->flags;
2138
2139         clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2140         smp_mb__after_atomic();
2141         wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2142 }
2143
2144 /*
2145  * There can be multiple RW segments.
2146  */
2147 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2148 {
2149         struct pnfs_layout_segment *lseg;
2150
2151         list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2152                 if (lseg->pls_range.iomode == IOMODE_RW &&
2153                     test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2154                         list_add(&lseg->pls_lc_list, listp);
2155         }
2156 }
2157
2158 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2159 {
2160         struct pnfs_layout_segment *lseg, *tmp;
2161
2162         /* Matched by references in pnfs_set_layoutcommit */
2163         list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2164                 list_del_init(&lseg->pls_lc_list);
2165                 pnfs_put_lseg(lseg);
2166         }
2167
2168         pnfs_clear_layoutcommitting(inode);
2169 }
2170
2171 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2172 {
2173         pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2174 }
2175 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2176
2177 void
2178 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2179                 loff_t end_pos)
2180 {
2181         struct nfs_inode *nfsi = NFS_I(inode);
2182         bool mark_as_dirty = false;
2183
2184         spin_lock(&inode->i_lock);
2185         if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2186                 nfsi->layout->plh_lwb = end_pos;
2187                 mark_as_dirty = true;
2188                 dprintk("%s: Set layoutcommit for inode %lu ",
2189                         __func__, inode->i_ino);
2190         } else if (end_pos > nfsi->layout->plh_lwb)
2191                 nfsi->layout->plh_lwb = end_pos;
2192         if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2193                 /* references matched in nfs4_layoutcommit_release */
2194                 pnfs_get_lseg(lseg);
2195         }
2196         spin_unlock(&inode->i_lock);
2197         dprintk("%s: lseg %p end_pos %llu\n",
2198                 __func__, lseg, nfsi->layout->plh_lwb);
2199
2200         /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2201          * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2202         if (mark_as_dirty)
2203                 mark_inode_dirty_sync(inode);
2204 }
2205 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2206
2207 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2208 {
2209         struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2210
2211         if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2212                 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2213         pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2214 }
2215
2216 /*
2217  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2218  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2219  * data to disk to allow the server to recover the data if it crashes.
2220  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2221  * is off, and a COMMIT is sent to a data server, or
2222  * if WRITEs to a data server return NFS_DATA_SYNC.
2223  */
2224 int
2225 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2226 {
2227         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2228         struct nfs4_layoutcommit_data *data;
2229         struct nfs_inode *nfsi = NFS_I(inode);
2230         loff_t end_pos;
2231         int status;
2232
2233         if (!pnfs_layoutcommit_outstanding(inode))
2234                 return 0;
2235
2236         dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2237
2238         status = -EAGAIN;
2239         if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2240                 if (!sync)
2241                         goto out;
2242                 status = wait_on_bit_lock_action(&nfsi->flags,
2243                                 NFS_INO_LAYOUTCOMMITTING,
2244                                 nfs_wait_bit_killable,
2245                                 TASK_KILLABLE);
2246                 if (status)
2247                         goto out;
2248         }
2249
2250         status = -ENOMEM;
2251         /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2252         data = kzalloc(sizeof(*data), GFP_NOFS);
2253         if (!data)
2254                 goto clear_layoutcommitting;
2255
2256         status = 0;
2257         spin_lock(&inode->i_lock);
2258         if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2259                 goto out_unlock;
2260
2261         INIT_LIST_HEAD(&data->lseg_list);
2262         pnfs_list_write_lseg(inode, &data->lseg_list);
2263
2264         end_pos = nfsi->layout->plh_lwb;
2265
2266         nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2267         spin_unlock(&inode->i_lock);
2268
2269         data->args.inode = inode;
2270         data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2271         nfs_fattr_init(&data->fattr);
2272         data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2273         data->res.fattr = &data->fattr;
2274         data->args.lastbytewritten = end_pos - 1;
2275         data->res.server = NFS_SERVER(inode);
2276
2277         if (ld->prepare_layoutcommit) {
2278                 status = ld->prepare_layoutcommit(&data->args);
2279                 if (status) {
2280                         put_rpccred(data->cred);
2281                         spin_lock(&inode->i_lock);
2282                         set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2283                         if (end_pos > nfsi->layout->plh_lwb)
2284                                 nfsi->layout->plh_lwb = end_pos;
2285                         goto out_unlock;
2286                 }
2287         }
2288
2289
2290         status = nfs4_proc_layoutcommit(data, sync);
2291 out:
2292         if (status)
2293                 mark_inode_dirty_sync(inode);
2294         dprintk("<-- %s status %d\n", __func__, status);
2295         return status;
2296 out_unlock:
2297         spin_unlock(&inode->i_lock);
2298         kfree(data);
2299 clear_layoutcommitting:
2300         pnfs_clear_layoutcommitting(inode);
2301         goto out;
2302 }
2303 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2304
2305 int
2306 pnfs_generic_sync(struct inode *inode, bool datasync)
2307 {
2308         return pnfs_layoutcommit_inode(inode, true);
2309 }
2310 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2311
2312 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2313 {
2314         struct nfs4_threshold *thp;
2315
2316         thp = kzalloc(sizeof(*thp), GFP_NOFS);
2317         if (!thp) {
2318                 dprintk("%s mdsthreshold allocation failed\n", __func__);
2319                 return NULL;
2320         }
2321         return thp;
2322 }
2323
2324 #if IS_ENABLED(CONFIG_NFS_V4_2)
2325 int
2326 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2327 {
2328         struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2329         struct nfs_server *server = NFS_SERVER(inode);
2330         struct nfs_inode *nfsi = NFS_I(inode);
2331         struct nfs42_layoutstat_data *data;
2332         struct pnfs_layout_hdr *hdr;
2333         int status = 0;
2334
2335         if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2336                 goto out;
2337
2338         if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2339                 goto out;
2340
2341         if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2342                 goto out;
2343
2344         spin_lock(&inode->i_lock);
2345         if (!NFS_I(inode)->layout) {
2346                 spin_unlock(&inode->i_lock);
2347                 goto out;
2348         }
2349         hdr = NFS_I(inode)->layout;
2350         pnfs_get_layout_hdr(hdr);
2351         spin_unlock(&inode->i_lock);
2352
2353         data = kzalloc(sizeof(*data), gfp_flags);
2354         if (!data) {
2355                 status = -ENOMEM;
2356                 goto out_put;
2357         }
2358
2359         data->args.fh = NFS_FH(inode);
2360         data->args.inode = inode;
2361         nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2362         status = ld->prepare_layoutstats(&data->args);
2363         if (status)
2364                 goto out_free;
2365
2366         status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2367
2368 out:
2369         dprintk("%s returns %d\n", __func__, status);
2370         return status;
2371
2372 out_free:
2373         kfree(data);
2374 out_put:
2375         pnfs_put_layout_hdr(hdr);
2376         smp_mb__before_atomic();
2377         clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2378         smp_mb__after_atomic();
2379         goto out;
2380 }
2381 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2382 #endif
2383
2384 unsigned int layoutstats_timer;
2385 module_param(layoutstats_timer, uint, 0644);
2386 EXPORT_SYMBOL_GPL(layoutstats_timer);