]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - fs/nfs/pnfs.c
Merge branch 'core-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[mirror_ubuntu-bionic-kernel.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 void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
56
57 /* Return the registered pnfs layout driver module matching given id */
58 static struct pnfs_layoutdriver_type *
59 find_pnfs_driver_locked(u32 id)
60 {
61 struct pnfs_layoutdriver_type *local;
62
63 list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
64 if (local->id == id)
65 goto out;
66 local = NULL;
67 out:
68 dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
69 return local;
70 }
71
72 static struct pnfs_layoutdriver_type *
73 find_pnfs_driver(u32 id)
74 {
75 struct pnfs_layoutdriver_type *local;
76
77 spin_lock(&pnfs_spinlock);
78 local = find_pnfs_driver_locked(id);
79 if (local != NULL && !try_module_get(local->owner)) {
80 dprintk("%s: Could not grab reference on module\n", __func__);
81 local = NULL;
82 }
83 spin_unlock(&pnfs_spinlock);
84 return local;
85 }
86
87 void
88 unset_pnfs_layoutdriver(struct nfs_server *nfss)
89 {
90 if (nfss->pnfs_curr_ld) {
91 if (nfss->pnfs_curr_ld->clear_layoutdriver)
92 nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
93 /* Decrement the MDS count. Purge the deviceid cache if zero */
94 if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
95 nfs4_deviceid_purge_client(nfss->nfs_client);
96 module_put(nfss->pnfs_curr_ld->owner);
97 }
98 nfss->pnfs_curr_ld = NULL;
99 }
100
101 /*
102 * Try to set the server's pnfs module to the pnfs layout type specified by id.
103 * Currently only one pNFS layout driver per filesystem is supported.
104 *
105 * @id layout type. Zero (illegal layout type) indicates pNFS not in use.
106 */
107 void
108 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
109 u32 id)
110 {
111 struct pnfs_layoutdriver_type *ld_type = NULL;
112
113 if (id == 0)
114 goto out_no_driver;
115 if (!(server->nfs_client->cl_exchange_flags &
116 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
117 printk(KERN_ERR "NFS: %s: id %u cl_exchange_flags 0x%x\n",
118 __func__, id, server->nfs_client->cl_exchange_flags);
119 goto out_no_driver;
120 }
121 ld_type = find_pnfs_driver(id);
122 if (!ld_type) {
123 request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX, id);
124 ld_type = find_pnfs_driver(id);
125 if (!ld_type) {
126 dprintk("%s: No pNFS module found for %u.\n",
127 __func__, id);
128 goto out_no_driver;
129 }
130 }
131 server->pnfs_curr_ld = ld_type;
132 if (ld_type->set_layoutdriver
133 && ld_type->set_layoutdriver(server, mntfh)) {
134 printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
135 "driver %u.\n", __func__, id);
136 module_put(ld_type->owner);
137 goto out_no_driver;
138 }
139 /* Bump the MDS count */
140 atomic_inc(&server->nfs_client->cl_mds_count);
141
142 dprintk("%s: pNFS module for %u set\n", __func__, id);
143 return;
144
145 out_no_driver:
146 dprintk("%s: Using NFSv4 I/O\n", __func__);
147 server->pnfs_curr_ld = NULL;
148 }
149
150 int
151 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
152 {
153 int status = -EINVAL;
154 struct pnfs_layoutdriver_type *tmp;
155
156 if (ld_type->id == 0) {
157 printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
158 return status;
159 }
160 if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
161 printk(KERN_ERR "NFS: %s Layout driver must provide "
162 "alloc_lseg and free_lseg.\n", __func__);
163 return status;
164 }
165
166 spin_lock(&pnfs_spinlock);
167 tmp = find_pnfs_driver_locked(ld_type->id);
168 if (!tmp) {
169 list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
170 status = 0;
171 dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
172 ld_type->name);
173 } else {
174 printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
175 __func__, ld_type->id);
176 }
177 spin_unlock(&pnfs_spinlock);
178
179 return status;
180 }
181 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
182
183 void
184 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
185 {
186 dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
187 spin_lock(&pnfs_spinlock);
188 list_del(&ld_type->pnfs_tblid);
189 spin_unlock(&pnfs_spinlock);
190 }
191 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
192
193 /*
194 * pNFS client layout cache
195 */
196
197 /* Need to hold i_lock if caller does not already hold reference */
198 void
199 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
200 {
201 atomic_inc(&lo->plh_refcount);
202 }
203
204 static struct pnfs_layout_hdr *
205 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
206 {
207 struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
208 return ld->alloc_layout_hdr(ino, gfp_flags);
209 }
210
211 static void
212 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
213 {
214 struct nfs_server *server = NFS_SERVER(lo->plh_inode);
215 struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
216
217 if (!list_empty(&lo->plh_layouts)) {
218 struct nfs_client *clp = server->nfs_client;
219
220 spin_lock(&clp->cl_lock);
221 list_del_init(&lo->plh_layouts);
222 spin_unlock(&clp->cl_lock);
223 }
224 put_rpccred(lo->plh_lc_cred);
225 return ld->free_layout_hdr(lo);
226 }
227
228 static void
229 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
230 {
231 struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
232 dprintk("%s: freeing layout cache %p\n", __func__, lo);
233 nfsi->layout = NULL;
234 /* Reset MDS Threshold I/O counters */
235 nfsi->write_io = 0;
236 nfsi->read_io = 0;
237 }
238
239 void
240 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
241 {
242 struct inode *inode = lo->plh_inode;
243
244 pnfs_layoutreturn_before_put_layout_hdr(lo);
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 /*
256 * Mark a pnfs_layout_hdr and all associated layout segments as invalid
257 *
258 * In order to continue using the pnfs_layout_hdr, a full recovery
259 * is required.
260 * Note that caller must hold inode->i_lock.
261 */
262 static int
263 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
264 struct list_head *lseg_list)
265 {
266 struct pnfs_layout_range range = {
267 .iomode = IOMODE_ANY,
268 .offset = 0,
269 .length = NFS4_MAX_UINT64,
270 };
271
272 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
273 return pnfs_mark_matching_lsegs_invalid(lo, lseg_list, &range);
274 }
275
276 static int
277 pnfs_iomode_to_fail_bit(u32 iomode)
278 {
279 return iomode == IOMODE_RW ?
280 NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
281 }
282
283 static void
284 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
285 {
286 lo->plh_retry_timestamp = jiffies;
287 if (!test_and_set_bit(fail_bit, &lo->plh_flags))
288 atomic_inc(&lo->plh_refcount);
289 }
290
291 static void
292 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
293 {
294 if (test_and_clear_bit(fail_bit, &lo->plh_flags))
295 atomic_dec(&lo->plh_refcount);
296 }
297
298 static void
299 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
300 {
301 struct inode *inode = lo->plh_inode;
302 struct pnfs_layout_range range = {
303 .iomode = iomode,
304 .offset = 0,
305 .length = NFS4_MAX_UINT64,
306 };
307 LIST_HEAD(head);
308
309 spin_lock(&inode->i_lock);
310 pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
311 pnfs_mark_matching_lsegs_invalid(lo, &head, &range);
312 spin_unlock(&inode->i_lock);
313 pnfs_free_lseg_list(&head);
314 dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
315 iomode == IOMODE_RW ? "RW" : "READ");
316 }
317
318 static bool
319 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
320 {
321 unsigned long start, end;
322 int fail_bit = pnfs_iomode_to_fail_bit(iomode);
323
324 if (test_bit(fail_bit, &lo->plh_flags) == 0)
325 return false;
326 end = jiffies;
327 start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
328 if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
329 /* It is time to retry the failed layoutgets */
330 pnfs_layout_clear_fail_bit(lo, fail_bit);
331 return false;
332 }
333 return true;
334 }
335
336 static void
337 init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg)
338 {
339 INIT_LIST_HEAD(&lseg->pls_list);
340 INIT_LIST_HEAD(&lseg->pls_lc_list);
341 atomic_set(&lseg->pls_refcount, 1);
342 smp_mb();
343 set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
344 lseg->pls_layout = lo;
345 }
346
347 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
348 {
349 struct inode *ino = lseg->pls_layout->plh_inode;
350
351 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
352 }
353
354 static void
355 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
356 struct pnfs_layout_segment *lseg)
357 {
358 struct inode *inode = lo->plh_inode;
359
360 WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
361 list_del_init(&lseg->pls_list);
362 /* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
363 atomic_dec(&lo->plh_refcount);
364 if (list_empty(&lo->plh_segs))
365 clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
366 rpc_wake_up(&NFS_SERVER(inode)->roc_rpcwaitq);
367 }
368
369 void
370 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
371 {
372 struct pnfs_layout_hdr *lo;
373 struct inode *inode;
374
375 if (!lseg)
376 return;
377
378 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
379 atomic_read(&lseg->pls_refcount),
380 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
381
382 lo = lseg->pls_layout;
383 inode = lo->plh_inode;
384
385 if (atomic_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
386 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
387 spin_unlock(&inode->i_lock);
388 return;
389 }
390 pnfs_get_layout_hdr(lo);
391 pnfs_layout_remove_lseg(lo, lseg);
392 spin_unlock(&inode->i_lock);
393 pnfs_free_lseg(lseg);
394 pnfs_put_layout_hdr(lo);
395 }
396 }
397 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
398
399 static void pnfs_free_lseg_async_work(struct work_struct *work)
400 {
401 struct pnfs_layout_segment *lseg;
402 struct pnfs_layout_hdr *lo;
403
404 lseg = container_of(work, struct pnfs_layout_segment, pls_work);
405 lo = lseg->pls_layout;
406
407 pnfs_free_lseg(lseg);
408 pnfs_put_layout_hdr(lo);
409 }
410
411 static void pnfs_free_lseg_async(struct pnfs_layout_segment *lseg)
412 {
413 INIT_WORK(&lseg->pls_work, pnfs_free_lseg_async_work);
414 schedule_work(&lseg->pls_work);
415 }
416
417 void
418 pnfs_put_lseg_locked(struct pnfs_layout_segment *lseg)
419 {
420 if (!lseg)
421 return;
422
423 assert_spin_locked(&lseg->pls_layout->plh_inode->i_lock);
424
425 dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
426 atomic_read(&lseg->pls_refcount),
427 test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
428 if (atomic_dec_and_test(&lseg->pls_refcount)) {
429 struct pnfs_layout_hdr *lo = lseg->pls_layout;
430 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
431 return;
432 pnfs_get_layout_hdr(lo);
433 pnfs_layout_remove_lseg(lo, lseg);
434 pnfs_free_lseg_async(lseg);
435 }
436 }
437 EXPORT_SYMBOL_GPL(pnfs_put_lseg_locked);
438
439 static u64
440 end_offset(u64 start, u64 len)
441 {
442 u64 end;
443
444 end = start + len;
445 return end >= start ? end : NFS4_MAX_UINT64;
446 }
447
448 /*
449 * is l2 fully contained in l1?
450 * start1 end1
451 * [----------------------------------)
452 * start2 end2
453 * [----------------)
454 */
455 static bool
456 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
457 const struct pnfs_layout_range *l2)
458 {
459 u64 start1 = l1->offset;
460 u64 end1 = end_offset(start1, l1->length);
461 u64 start2 = l2->offset;
462 u64 end2 = end_offset(start2, l2->length);
463
464 return (start1 <= start2) && (end1 >= end2);
465 }
466
467 /*
468 * is l1 and l2 intersecting?
469 * start1 end1
470 * [----------------------------------)
471 * start2 end2
472 * [----------------)
473 */
474 static bool
475 pnfs_lseg_range_intersecting(const struct pnfs_layout_range *l1,
476 const struct pnfs_layout_range *l2)
477 {
478 u64 start1 = l1->offset;
479 u64 end1 = end_offset(start1, l1->length);
480 u64 start2 = l2->offset;
481 u64 end2 = end_offset(start2, l2->length);
482
483 return (end1 == NFS4_MAX_UINT64 || end1 > start2) &&
484 (end2 == NFS4_MAX_UINT64 || end2 > start1);
485 }
486
487 static bool
488 should_free_lseg(const struct pnfs_layout_range *lseg_range,
489 const struct pnfs_layout_range *recall_range)
490 {
491 return (recall_range->iomode == IOMODE_ANY ||
492 lseg_range->iomode == recall_range->iomode) &&
493 pnfs_lseg_range_intersecting(lseg_range, recall_range);
494 }
495
496 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
497 struct list_head *tmp_list)
498 {
499 if (!atomic_dec_and_test(&lseg->pls_refcount))
500 return false;
501 pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
502 list_add(&lseg->pls_list, tmp_list);
503 return true;
504 }
505
506 /* Returns 1 if lseg is removed from list, 0 otherwise */
507 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
508 struct list_head *tmp_list)
509 {
510 int rv = 0;
511
512 if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
513 /* Remove the reference keeping the lseg in the
514 * list. It will now be removed when all
515 * outstanding io is finished.
516 */
517 dprintk("%s: lseg %p ref %d\n", __func__, lseg,
518 atomic_read(&lseg->pls_refcount));
519 if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
520 rv = 1;
521 }
522 return rv;
523 }
524
525 /* Returns count of number of matching invalid lsegs remaining in list
526 * after call.
527 */
528 int
529 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
530 struct list_head *tmp_list,
531 const struct pnfs_layout_range *recall_range)
532 {
533 struct pnfs_layout_segment *lseg, *next;
534 int remaining = 0;
535
536 dprintk("%s:Begin lo %p\n", __func__, lo);
537
538 if (list_empty(&lo->plh_segs))
539 return 0;
540 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
541 if (!recall_range ||
542 should_free_lseg(&lseg->pls_range, recall_range)) {
543 dprintk("%s: freeing lseg %p iomode %d "
544 "offset %llu length %llu\n", __func__,
545 lseg, lseg->pls_range.iomode, lseg->pls_range.offset,
546 lseg->pls_range.length);
547 if (!mark_lseg_invalid(lseg, tmp_list))
548 remaining++;
549 }
550 dprintk("%s:Return %i\n", __func__, remaining);
551 return remaining;
552 }
553
554 /* note free_me must contain lsegs from a single layout_hdr */
555 void
556 pnfs_free_lseg_list(struct list_head *free_me)
557 {
558 struct pnfs_layout_segment *lseg, *tmp;
559
560 if (list_empty(free_me))
561 return;
562
563 list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
564 list_del(&lseg->pls_list);
565 pnfs_free_lseg(lseg);
566 }
567 }
568
569 void
570 pnfs_destroy_layout(struct nfs_inode *nfsi)
571 {
572 struct pnfs_layout_hdr *lo;
573 LIST_HEAD(tmp_list);
574
575 spin_lock(&nfsi->vfs_inode.i_lock);
576 lo = nfsi->layout;
577 if (lo) {
578 pnfs_get_layout_hdr(lo);
579 pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
580 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
581 pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
582 spin_unlock(&nfsi->vfs_inode.i_lock);
583 pnfs_free_lseg_list(&tmp_list);
584 pnfs_put_layout_hdr(lo);
585 } else
586 spin_unlock(&nfsi->vfs_inode.i_lock);
587 }
588 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
589
590 static bool
591 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
592 struct list_head *layout_list)
593 {
594 struct pnfs_layout_hdr *lo;
595 bool ret = false;
596
597 spin_lock(&inode->i_lock);
598 lo = NFS_I(inode)->layout;
599 if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
600 pnfs_get_layout_hdr(lo);
601 list_add(&lo->plh_bulk_destroy, layout_list);
602 ret = true;
603 }
604 spin_unlock(&inode->i_lock);
605 return ret;
606 }
607
608 /* Caller must hold rcu_read_lock and clp->cl_lock */
609 static int
610 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
611 struct nfs_server *server,
612 struct list_head *layout_list)
613 {
614 struct pnfs_layout_hdr *lo, *next;
615 struct inode *inode;
616
617 list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
618 inode = igrab(lo->plh_inode);
619 if (inode == NULL)
620 continue;
621 list_del_init(&lo->plh_layouts);
622 if (pnfs_layout_add_bulk_destroy_list(inode, layout_list))
623 continue;
624 rcu_read_unlock();
625 spin_unlock(&clp->cl_lock);
626 iput(inode);
627 spin_lock(&clp->cl_lock);
628 rcu_read_lock();
629 return -EAGAIN;
630 }
631 return 0;
632 }
633
634 static int
635 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
636 bool is_bulk_recall)
637 {
638 struct pnfs_layout_hdr *lo;
639 struct inode *inode;
640 LIST_HEAD(lseg_list);
641 int ret = 0;
642
643 while (!list_empty(layout_list)) {
644 lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
645 plh_bulk_destroy);
646 dprintk("%s freeing layout for inode %lu\n", __func__,
647 lo->plh_inode->i_ino);
648 inode = lo->plh_inode;
649
650 pnfs_layoutcommit_inode(inode, false);
651
652 spin_lock(&inode->i_lock);
653 list_del_init(&lo->plh_bulk_destroy);
654 if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
655 if (is_bulk_recall)
656 set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
657 ret = -EAGAIN;
658 }
659 spin_unlock(&inode->i_lock);
660 pnfs_free_lseg_list(&lseg_list);
661 /* Free all lsegs that are attached to commit buckets */
662 nfs_commit_inode(inode, 0);
663 pnfs_put_layout_hdr(lo);
664 iput(inode);
665 }
666 return ret;
667 }
668
669 int
670 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
671 struct nfs_fsid *fsid,
672 bool is_recall)
673 {
674 struct nfs_server *server;
675 LIST_HEAD(layout_list);
676
677 spin_lock(&clp->cl_lock);
678 rcu_read_lock();
679 restart:
680 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
681 if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
682 continue;
683 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
684 server,
685 &layout_list) != 0)
686 goto restart;
687 }
688 rcu_read_unlock();
689 spin_unlock(&clp->cl_lock);
690
691 if (list_empty(&layout_list))
692 return 0;
693 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
694 }
695
696 int
697 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
698 bool is_recall)
699 {
700 struct nfs_server *server;
701 LIST_HEAD(layout_list);
702
703 spin_lock(&clp->cl_lock);
704 rcu_read_lock();
705 restart:
706 list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
707 if (pnfs_layout_bulk_destroy_byserver_locked(clp,
708 server,
709 &layout_list) != 0)
710 goto restart;
711 }
712 rcu_read_unlock();
713 spin_unlock(&clp->cl_lock);
714
715 if (list_empty(&layout_list))
716 return 0;
717 return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
718 }
719
720 /*
721 * Called by the state manger to remove all layouts established under an
722 * expired lease.
723 */
724 void
725 pnfs_destroy_all_layouts(struct nfs_client *clp)
726 {
727 nfs4_deviceid_mark_client_invalid(clp);
728 nfs4_deviceid_purge_client(clp);
729
730 pnfs_destroy_layouts_byclid(clp, false);
731 }
732
733 /*
734 * Compare 2 layout stateid sequence ids, to see which is newer,
735 * taking into account wraparound issues.
736 */
737 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
738 {
739 return (s32)(s1 - s2) > 0;
740 }
741
742 /* update lo->plh_stateid with new if is more recent */
743 void
744 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
745 bool update_barrier)
746 {
747 u32 oldseq, newseq, new_barrier;
748 int empty = list_empty(&lo->plh_segs);
749
750 oldseq = be32_to_cpu(lo->plh_stateid.seqid);
751 newseq = be32_to_cpu(new->seqid);
752 if (empty || pnfs_seqid_is_newer(newseq, oldseq)) {
753 nfs4_stateid_copy(&lo->plh_stateid, new);
754 if (update_barrier) {
755 new_barrier = be32_to_cpu(new->seqid);
756 } else {
757 /* Because of wraparound, we want to keep the barrier
758 * "close" to the current seqids.
759 */
760 new_barrier = newseq - atomic_read(&lo->plh_outstanding);
761 }
762 if (empty || pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
763 lo->plh_barrier = new_barrier;
764 }
765 }
766
767 static bool
768 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
769 const nfs4_stateid *stateid)
770 {
771 u32 seqid = be32_to_cpu(stateid->seqid);
772
773 return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
774 }
775
776 /* lget is set to 1 if called from inside send_layoutget call chain */
777 static bool
778 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
779 {
780 return lo->plh_block_lgets ||
781 test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
782 }
783
784 int
785 pnfs_choose_layoutget_stateid(nfs4_stateid *dst, struct pnfs_layout_hdr *lo,
786 const struct pnfs_layout_range *range,
787 struct nfs4_state *open_state)
788 {
789 int status = 0;
790
791 dprintk("--> %s\n", __func__);
792 spin_lock(&lo->plh_inode->i_lock);
793 if (pnfs_layoutgets_blocked(lo)) {
794 status = -EAGAIN;
795 } else if (!nfs4_valid_open_stateid(open_state)) {
796 status = -EBADF;
797 } else if (list_empty(&lo->plh_segs) ||
798 test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
799 int seq;
800
801 do {
802 seq = read_seqbegin(&open_state->seqlock);
803 nfs4_stateid_copy(dst, &open_state->stateid);
804 } while (read_seqretry(&open_state->seqlock, seq));
805 } else
806 nfs4_stateid_copy(dst, &lo->plh_stateid);
807 spin_unlock(&lo->plh_inode->i_lock);
808 dprintk("<-- %s\n", __func__);
809 return status;
810 }
811
812 /*
813 * Get layout from server.
814 * for now, assume that whole file layouts are requested.
815 * arg->offset: 0
816 * arg->length: all ones
817 */
818 static struct pnfs_layout_segment *
819 send_layoutget(struct pnfs_layout_hdr *lo,
820 struct nfs_open_context *ctx,
821 const struct pnfs_layout_range *range,
822 gfp_t gfp_flags)
823 {
824 struct inode *ino = lo->plh_inode;
825 struct nfs_server *server = NFS_SERVER(ino);
826 struct nfs4_layoutget *lgp;
827 struct pnfs_layout_segment *lseg;
828 loff_t i_size;
829
830 dprintk("--> %s\n", __func__);
831
832 /*
833 * Synchronously retrieve layout information from server and
834 * store in lseg. If we race with a concurrent seqid morphing
835 * op, then re-send the LAYOUTGET.
836 */
837 do {
838 lgp = kzalloc(sizeof(*lgp), gfp_flags);
839 if (lgp == NULL)
840 return NULL;
841
842 i_size = i_size_read(ino);
843
844 lgp->args.minlength = PAGE_SIZE;
845 if (lgp->args.minlength > range->length)
846 lgp->args.minlength = range->length;
847 if (range->iomode == IOMODE_READ) {
848 if (range->offset >= i_size)
849 lgp->args.minlength = 0;
850 else if (i_size - range->offset < lgp->args.minlength)
851 lgp->args.minlength = i_size - range->offset;
852 }
853 lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
854 pnfs_copy_range(&lgp->args.range, range);
855 lgp->args.type = server->pnfs_curr_ld->id;
856 lgp->args.inode = ino;
857 lgp->args.ctx = get_nfs_open_context(ctx);
858 lgp->gfp_flags = gfp_flags;
859 lgp->cred = lo->plh_lc_cred;
860
861 lseg = nfs4_proc_layoutget(lgp, gfp_flags);
862 } while (lseg == ERR_PTR(-EAGAIN));
863
864 if (IS_ERR(lseg) && !nfs_error_is_fatal(PTR_ERR(lseg)))
865 lseg = NULL;
866 else
867 pnfs_layout_clear_fail_bit(lo,
868 pnfs_iomode_to_fail_bit(range->iomode));
869
870 return lseg;
871 }
872
873 static void pnfs_clear_layoutcommit(struct inode *inode,
874 struct list_head *head)
875 {
876 struct nfs_inode *nfsi = NFS_I(inode);
877 struct pnfs_layout_segment *lseg, *tmp;
878
879 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
880 return;
881 list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
882 if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
883 continue;
884 pnfs_lseg_dec_and_remove_zero(lseg, head);
885 }
886 }
887
888 void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
889 {
890 clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
891 smp_mb__after_atomic();
892 wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
893 rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
894 }
895
896 static bool
897 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
898 {
899 if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
900 return false;
901 lo->plh_return_iomode = 0;
902 pnfs_get_layout_hdr(lo);
903 clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
904 return true;
905 }
906
907 static int
908 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo, const nfs4_stateid *stateid,
909 enum pnfs_iomode iomode, bool sync)
910 {
911 struct inode *ino = lo->plh_inode;
912 struct nfs4_layoutreturn *lrp;
913 int status = 0;
914
915 lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
916 if (unlikely(lrp == NULL)) {
917 status = -ENOMEM;
918 spin_lock(&ino->i_lock);
919 pnfs_clear_layoutreturn_waitbit(lo);
920 spin_unlock(&ino->i_lock);
921 pnfs_put_layout_hdr(lo);
922 goto out;
923 }
924
925 nfs4_stateid_copy(&lrp->args.stateid, stateid);
926 lrp->args.layout_type = NFS_SERVER(ino)->pnfs_curr_ld->id;
927 lrp->args.inode = ino;
928 lrp->args.range.iomode = iomode;
929 lrp->args.range.offset = 0;
930 lrp->args.range.length = NFS4_MAX_UINT64;
931 lrp->args.layout = lo;
932 lrp->clp = NFS_SERVER(ino)->nfs_client;
933 lrp->cred = lo->plh_lc_cred;
934
935 status = nfs4_proc_layoutreturn(lrp, sync);
936 out:
937 dprintk("<-- %s status: %d\n", __func__, status);
938 return status;
939 }
940
941 /* Return true if layoutreturn is needed */
942 static bool
943 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
944 {
945 struct pnfs_layout_segment *s;
946
947 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
948 return false;
949
950 /* Defer layoutreturn until all lsegs are done */
951 list_for_each_entry(s, &lo->plh_segs, pls_list) {
952 if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
953 return false;
954 }
955
956 return true;
957 }
958
959 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
960 {
961 struct inode *inode= lo->plh_inode;
962
963 if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
964 return;
965 spin_lock(&inode->i_lock);
966 if (pnfs_layout_need_return(lo)) {
967 nfs4_stateid stateid;
968 enum pnfs_iomode iomode;
969 bool send;
970
971 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
972 iomode = lo->plh_return_iomode;
973 send = pnfs_prepare_layoutreturn(lo);
974 spin_unlock(&inode->i_lock);
975 if (send) {
976 /* Send an async layoutreturn so we dont deadlock */
977 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
978 }
979 } else
980 spin_unlock(&inode->i_lock);
981 }
982
983 /*
984 * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
985 * when the layout segment list is empty.
986 *
987 * Note that a pnfs_layout_hdr can exist with an empty layout segment
988 * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
989 * deviceid is marked invalid.
990 */
991 int
992 _pnfs_return_layout(struct inode *ino)
993 {
994 struct pnfs_layout_hdr *lo = NULL;
995 struct nfs_inode *nfsi = NFS_I(ino);
996 LIST_HEAD(tmp_list);
997 nfs4_stateid stateid;
998 int status = 0, empty;
999 bool send;
1000
1001 dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1002
1003 spin_lock(&ino->i_lock);
1004 lo = nfsi->layout;
1005 if (!lo) {
1006 spin_unlock(&ino->i_lock);
1007 dprintk("NFS: %s no layout to return\n", __func__);
1008 goto out;
1009 }
1010 nfs4_stateid_copy(&stateid, &nfsi->layout->plh_stateid);
1011 /* Reference matched in nfs4_layoutreturn_release */
1012 pnfs_get_layout_hdr(lo);
1013 empty = list_empty(&lo->plh_segs);
1014 pnfs_clear_layoutcommit(ino, &tmp_list);
1015 pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL);
1016
1017 if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1018 struct pnfs_layout_range range = {
1019 .iomode = IOMODE_ANY,
1020 .offset = 0,
1021 .length = NFS4_MAX_UINT64,
1022 };
1023 NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1024 }
1025
1026 /* Don't send a LAYOUTRETURN if list was initially empty */
1027 if (empty) {
1028 spin_unlock(&ino->i_lock);
1029 dprintk("NFS: %s no layout segments to return\n", __func__);
1030 goto out_put_layout_hdr;
1031 }
1032
1033 set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1034 send = pnfs_prepare_layoutreturn(lo);
1035 spin_unlock(&ino->i_lock);
1036 pnfs_free_lseg_list(&tmp_list);
1037 if (send)
1038 status = pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1039 out_put_layout_hdr:
1040 pnfs_put_layout_hdr(lo);
1041 out:
1042 dprintk("<-- %s status: %d\n", __func__, status);
1043 return status;
1044 }
1045 EXPORT_SYMBOL_GPL(_pnfs_return_layout);
1046
1047 int
1048 pnfs_commit_and_return_layout(struct inode *inode)
1049 {
1050 struct pnfs_layout_hdr *lo;
1051 int ret;
1052
1053 spin_lock(&inode->i_lock);
1054 lo = NFS_I(inode)->layout;
1055 if (lo == NULL) {
1056 spin_unlock(&inode->i_lock);
1057 return 0;
1058 }
1059 pnfs_get_layout_hdr(lo);
1060 /* Block new layoutgets and read/write to ds */
1061 lo->plh_block_lgets++;
1062 spin_unlock(&inode->i_lock);
1063 filemap_fdatawait(inode->i_mapping);
1064 ret = pnfs_layoutcommit_inode(inode, true);
1065 if (ret == 0)
1066 ret = _pnfs_return_layout(inode);
1067 spin_lock(&inode->i_lock);
1068 lo->plh_block_lgets--;
1069 spin_unlock(&inode->i_lock);
1070 pnfs_put_layout_hdr(lo);
1071 return ret;
1072 }
1073
1074 bool pnfs_roc(struct inode *ino)
1075 {
1076 struct nfs_inode *nfsi = NFS_I(ino);
1077 struct nfs_open_context *ctx;
1078 struct nfs4_state *state;
1079 struct pnfs_layout_hdr *lo;
1080 struct pnfs_layout_segment *lseg, *tmp;
1081 nfs4_stateid stateid;
1082 LIST_HEAD(tmp_list);
1083 bool found = false, layoutreturn = false, roc = false;
1084
1085 spin_lock(&ino->i_lock);
1086 lo = nfsi->layout;
1087 if (!lo || test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags))
1088 goto out_noroc;
1089
1090 /* no roc if we hold a delegation */
1091 if (nfs4_check_delegation(ino, FMODE_READ))
1092 goto out_noroc;
1093
1094 list_for_each_entry(ctx, &nfsi->open_files, list) {
1095 state = ctx->state;
1096 /* Don't return layout if there is open file state */
1097 if (state != NULL && state->state != 0)
1098 goto out_noroc;
1099 }
1100
1101 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1102 /* always send layoutreturn if being marked so */
1103 if (test_and_clear_bit(NFS_LAYOUT_RETURN_REQUESTED,
1104 &lo->plh_flags))
1105 layoutreturn = pnfs_prepare_layoutreturn(lo);
1106
1107 list_for_each_entry_safe(lseg, tmp, &lo->plh_segs, pls_list)
1108 /* If we are sending layoutreturn, invalidate all valid lsegs */
1109 if (layoutreturn || test_bit(NFS_LSEG_ROC, &lseg->pls_flags)) {
1110 mark_lseg_invalid(lseg, &tmp_list);
1111 found = true;
1112 }
1113 /* ROC in two conditions:
1114 * 1. there are ROC lsegs
1115 * 2. we don't send layoutreturn
1116 */
1117 if (found && !layoutreturn) {
1118 /* lo ref dropped in pnfs_roc_release() */
1119 pnfs_get_layout_hdr(lo);
1120 roc = true;
1121 }
1122
1123 out_noroc:
1124 spin_unlock(&ino->i_lock);
1125 pnfs_free_lseg_list(&tmp_list);
1126 pnfs_layoutcommit_inode(ino, true);
1127 if (layoutreturn)
1128 pnfs_send_layoutreturn(lo, &stateid, IOMODE_ANY, true);
1129 return roc;
1130 }
1131
1132 void pnfs_roc_release(struct inode *ino)
1133 {
1134 struct pnfs_layout_hdr *lo;
1135
1136 spin_lock(&ino->i_lock);
1137 lo = NFS_I(ino)->layout;
1138 pnfs_clear_layoutreturn_waitbit(lo);
1139 if (atomic_dec_and_test(&lo->plh_refcount)) {
1140 pnfs_detach_layout_hdr(lo);
1141 spin_unlock(&ino->i_lock);
1142 pnfs_free_layout_hdr(lo);
1143 } else
1144 spin_unlock(&ino->i_lock);
1145 }
1146
1147 void pnfs_roc_set_barrier(struct inode *ino, u32 barrier)
1148 {
1149 struct pnfs_layout_hdr *lo;
1150
1151 spin_lock(&ino->i_lock);
1152 lo = NFS_I(ino)->layout;
1153 pnfs_mark_layout_returned_if_empty(lo);
1154 if (pnfs_seqid_is_newer(barrier, lo->plh_barrier))
1155 lo->plh_barrier = barrier;
1156 spin_unlock(&ino->i_lock);
1157 trace_nfs4_layoutreturn_on_close(ino, 0);
1158 }
1159
1160 void pnfs_roc_get_barrier(struct inode *ino, u32 *barrier)
1161 {
1162 struct nfs_inode *nfsi = NFS_I(ino);
1163 struct pnfs_layout_hdr *lo;
1164 u32 current_seqid;
1165
1166 spin_lock(&ino->i_lock);
1167 lo = nfsi->layout;
1168 current_seqid = be32_to_cpu(lo->plh_stateid.seqid);
1169
1170 /* Since close does not return a layout stateid for use as
1171 * a barrier, we choose the worst-case barrier.
1172 */
1173 *barrier = current_seqid + atomic_read(&lo->plh_outstanding);
1174 spin_unlock(&ino->i_lock);
1175 }
1176
1177 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1178 {
1179 struct nfs_inode *nfsi = NFS_I(ino);
1180 struct pnfs_layout_hdr *lo;
1181 bool sleep = false;
1182
1183 /* we might not have grabbed lo reference. so need to check under
1184 * i_lock */
1185 spin_lock(&ino->i_lock);
1186 lo = nfsi->layout;
1187 if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
1188 sleep = true;
1189 spin_unlock(&ino->i_lock);
1190
1191 if (sleep)
1192 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1193
1194 return sleep;
1195 }
1196
1197 /*
1198 * Compare two layout segments for sorting into layout cache.
1199 * We want to preferentially return RW over RO layouts, so ensure those
1200 * are seen first.
1201 */
1202 static s64
1203 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1204 const struct pnfs_layout_range *l2)
1205 {
1206 s64 d;
1207
1208 /* high offset > low offset */
1209 d = l1->offset - l2->offset;
1210 if (d)
1211 return d;
1212
1213 /* short length > long length */
1214 d = l2->length - l1->length;
1215 if (d)
1216 return d;
1217
1218 /* read > read/write */
1219 return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1220 }
1221
1222 static bool
1223 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1224 const struct pnfs_layout_range *l2)
1225 {
1226 return pnfs_lseg_range_cmp(l1, l2) > 0;
1227 }
1228
1229 static bool
1230 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1231 struct pnfs_layout_segment *old)
1232 {
1233 return false;
1234 }
1235
1236 void
1237 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1238 struct pnfs_layout_segment *lseg,
1239 bool (*is_after)(const struct pnfs_layout_range *,
1240 const struct pnfs_layout_range *),
1241 bool (*do_merge)(struct pnfs_layout_segment *,
1242 struct pnfs_layout_segment *),
1243 struct list_head *free_me)
1244 {
1245 struct pnfs_layout_segment *lp, *tmp;
1246
1247 dprintk("%s:Begin\n", __func__);
1248
1249 list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1250 if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1251 continue;
1252 if (do_merge(lseg, lp)) {
1253 mark_lseg_invalid(lp, free_me);
1254 continue;
1255 }
1256 if (is_after(&lseg->pls_range, &lp->pls_range))
1257 continue;
1258 list_add_tail(&lseg->pls_list, &lp->pls_list);
1259 dprintk("%s: inserted lseg %p "
1260 "iomode %d offset %llu length %llu before "
1261 "lp %p iomode %d offset %llu length %llu\n",
1262 __func__, lseg, lseg->pls_range.iomode,
1263 lseg->pls_range.offset, lseg->pls_range.length,
1264 lp, lp->pls_range.iomode, lp->pls_range.offset,
1265 lp->pls_range.length);
1266 goto out;
1267 }
1268 list_add_tail(&lseg->pls_list, &lo->plh_segs);
1269 dprintk("%s: inserted lseg %p "
1270 "iomode %d offset %llu length %llu at tail\n",
1271 __func__, lseg, lseg->pls_range.iomode,
1272 lseg->pls_range.offset, lseg->pls_range.length);
1273 out:
1274 pnfs_get_layout_hdr(lo);
1275
1276 dprintk("%s:Return\n", __func__);
1277 }
1278 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1279
1280 static void
1281 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1282 struct pnfs_layout_segment *lseg,
1283 struct list_head *free_me)
1284 {
1285 struct inode *inode = lo->plh_inode;
1286 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1287
1288 if (ld->add_lseg != NULL)
1289 ld->add_lseg(lo, lseg, free_me);
1290 else
1291 pnfs_generic_layout_insert_lseg(lo, lseg,
1292 pnfs_lseg_range_is_after,
1293 pnfs_lseg_no_merge,
1294 free_me);
1295 }
1296
1297 static struct pnfs_layout_hdr *
1298 alloc_init_layout_hdr(struct inode *ino,
1299 struct nfs_open_context *ctx,
1300 gfp_t gfp_flags)
1301 {
1302 struct pnfs_layout_hdr *lo;
1303
1304 lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1305 if (!lo)
1306 return NULL;
1307 atomic_set(&lo->plh_refcount, 1);
1308 INIT_LIST_HEAD(&lo->plh_layouts);
1309 INIT_LIST_HEAD(&lo->plh_segs);
1310 INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1311 lo->plh_inode = ino;
1312 lo->plh_lc_cred = get_rpccred(ctx->cred);
1313 return lo;
1314 }
1315
1316 static struct pnfs_layout_hdr *
1317 pnfs_find_alloc_layout(struct inode *ino,
1318 struct nfs_open_context *ctx,
1319 gfp_t gfp_flags)
1320 {
1321 struct nfs_inode *nfsi = NFS_I(ino);
1322 struct pnfs_layout_hdr *new = NULL;
1323
1324 dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1325
1326 if (nfsi->layout != NULL)
1327 goto out_existing;
1328 spin_unlock(&ino->i_lock);
1329 new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1330 spin_lock(&ino->i_lock);
1331
1332 if (likely(nfsi->layout == NULL)) { /* Won the race? */
1333 nfsi->layout = new;
1334 return new;
1335 } else if (new != NULL)
1336 pnfs_free_layout_hdr(new);
1337 out_existing:
1338 pnfs_get_layout_hdr(nfsi->layout);
1339 return nfsi->layout;
1340 }
1341
1342 /*
1343 * iomode matching rules:
1344 * iomode lseg match
1345 * ----- ----- -----
1346 * ANY READ true
1347 * ANY RW true
1348 * RW READ false
1349 * RW RW true
1350 * READ READ true
1351 * READ RW true
1352 */
1353 static bool
1354 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1355 const struct pnfs_layout_range *range)
1356 {
1357 struct pnfs_layout_range range1;
1358
1359 if ((range->iomode == IOMODE_RW &&
1360 ls_range->iomode != IOMODE_RW) ||
1361 !pnfs_lseg_range_intersecting(ls_range, range))
1362 return 0;
1363
1364 /* range1 covers only the first byte in the range */
1365 range1 = *range;
1366 range1.length = 1;
1367 return pnfs_lseg_range_contained(ls_range, &range1);
1368 }
1369
1370 /*
1371 * lookup range in layout
1372 */
1373 static struct pnfs_layout_segment *
1374 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1375 struct pnfs_layout_range *range)
1376 {
1377 struct pnfs_layout_segment *lseg, *ret = NULL;
1378
1379 dprintk("%s:Begin\n", __func__);
1380
1381 list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1382 if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1383 !test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
1384 pnfs_lseg_range_match(&lseg->pls_range, range)) {
1385 ret = pnfs_get_lseg(lseg);
1386 break;
1387 }
1388 }
1389
1390 dprintk("%s:Return lseg %p ref %d\n",
1391 __func__, ret, ret ? atomic_read(&ret->pls_refcount) : 0);
1392 return ret;
1393 }
1394
1395 /*
1396 * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1397 * to the MDS or over pNFS
1398 *
1399 * The nfs_inode read_io and write_io fields are cumulative counters reset
1400 * when there are no layout segments. Note that in pnfs_update_layout iomode
1401 * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1402 * WRITE request.
1403 *
1404 * A return of true means use MDS I/O.
1405 *
1406 * From rfc 5661:
1407 * If a file's size is smaller than the file size threshold, data accesses
1408 * SHOULD be sent to the metadata server. If an I/O request has a length that
1409 * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1410 * server. If both file size and I/O size are provided, the client SHOULD
1411 * reach or exceed both thresholds before sending its read or write
1412 * requests to the data server.
1413 */
1414 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1415 struct inode *ino, int iomode)
1416 {
1417 struct nfs4_threshold *t = ctx->mdsthreshold;
1418 struct nfs_inode *nfsi = NFS_I(ino);
1419 loff_t fsize = i_size_read(ino);
1420 bool size = false, size_set = false, io = false, io_set = false, ret = false;
1421
1422 if (t == NULL)
1423 return ret;
1424
1425 dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1426 __func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1427
1428 switch (iomode) {
1429 case IOMODE_READ:
1430 if (t->bm & THRESHOLD_RD) {
1431 dprintk("%s fsize %llu\n", __func__, fsize);
1432 size_set = true;
1433 if (fsize < t->rd_sz)
1434 size = true;
1435 }
1436 if (t->bm & THRESHOLD_RD_IO) {
1437 dprintk("%s nfsi->read_io %llu\n", __func__,
1438 nfsi->read_io);
1439 io_set = true;
1440 if (nfsi->read_io < t->rd_io_sz)
1441 io = true;
1442 }
1443 break;
1444 case IOMODE_RW:
1445 if (t->bm & THRESHOLD_WR) {
1446 dprintk("%s fsize %llu\n", __func__, fsize);
1447 size_set = true;
1448 if (fsize < t->wr_sz)
1449 size = true;
1450 }
1451 if (t->bm & THRESHOLD_WR_IO) {
1452 dprintk("%s nfsi->write_io %llu\n", __func__,
1453 nfsi->write_io);
1454 io_set = true;
1455 if (nfsi->write_io < t->wr_io_sz)
1456 io = true;
1457 }
1458 break;
1459 }
1460 if (size_set && io_set) {
1461 if (size && io)
1462 ret = true;
1463 } else if (size || io)
1464 ret = true;
1465
1466 dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1467 return ret;
1468 }
1469
1470 static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1471 {
1472 /*
1473 * send layoutcommit as it can hold up layoutreturn due to lseg
1474 * reference
1475 */
1476 pnfs_layoutcommit_inode(lo->plh_inode, false);
1477 return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1478 nfs_wait_bit_killable,
1479 TASK_UNINTERRUPTIBLE);
1480 }
1481
1482 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1483 {
1484 unsigned long *bitlock = &lo->plh_flags;
1485
1486 clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1487 smp_mb__after_atomic();
1488 wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1489 }
1490
1491 /*
1492 * Layout segment is retreived from the server if not cached.
1493 * The appropriate layout segment is referenced and returned to the caller.
1494 */
1495 struct pnfs_layout_segment *
1496 pnfs_update_layout(struct inode *ino,
1497 struct nfs_open_context *ctx,
1498 loff_t pos,
1499 u64 count,
1500 enum pnfs_iomode iomode,
1501 gfp_t gfp_flags)
1502 {
1503 struct pnfs_layout_range arg = {
1504 .iomode = iomode,
1505 .offset = pos,
1506 .length = count,
1507 };
1508 unsigned pg_offset;
1509 struct nfs_server *server = NFS_SERVER(ino);
1510 struct nfs_client *clp = server->nfs_client;
1511 struct pnfs_layout_hdr *lo;
1512 struct pnfs_layout_segment *lseg = NULL;
1513 bool first;
1514
1515 if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1516 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1517 PNFS_UPDATE_LAYOUT_NO_PNFS);
1518 goto out;
1519 }
1520
1521 if (iomode == IOMODE_READ && i_size_read(ino) == 0) {
1522 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1523 PNFS_UPDATE_LAYOUT_RD_ZEROLEN);
1524 goto out;
1525 }
1526
1527 if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1528 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1529 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1530 goto out;
1531 }
1532
1533 lookup_again:
1534 first = false;
1535 spin_lock(&ino->i_lock);
1536 lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1537 if (lo == NULL) {
1538 spin_unlock(&ino->i_lock);
1539 trace_pnfs_update_layout(ino, pos, count, iomode, NULL,
1540 PNFS_UPDATE_LAYOUT_NOMEM);
1541 goto out;
1542 }
1543
1544 /* Do we even need to bother with this? */
1545 if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1546 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1547 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1548 dprintk("%s matches recall, use MDS\n", __func__);
1549 goto out_unlock;
1550 }
1551
1552 /* if LAYOUTGET already failed once we don't try again */
1553 if (pnfs_layout_io_test_failed(lo, iomode)) {
1554 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1555 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1556 goto out_unlock;
1557 }
1558
1559 first = list_empty(&lo->plh_segs);
1560 if (first) {
1561 /* The first layoutget for the file. Need to serialize per
1562 * RFC 5661 Errata 3208.
1563 */
1564 if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
1565 &lo->plh_flags)) {
1566 spin_unlock(&ino->i_lock);
1567 wait_on_bit(&lo->plh_flags, NFS_LAYOUT_FIRST_LAYOUTGET,
1568 TASK_UNINTERRUPTIBLE);
1569 pnfs_put_layout_hdr(lo);
1570 goto lookup_again;
1571 }
1572 } else {
1573 /* Check to see if the layout for the given range
1574 * already exists
1575 */
1576 lseg = pnfs_find_lseg(lo, &arg);
1577 if (lseg) {
1578 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1579 PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1580 goto out_unlock;
1581 }
1582 }
1583
1584 /*
1585 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
1586 * for LAYOUTRETURN even if first is true.
1587 */
1588 if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1589 spin_unlock(&ino->i_lock);
1590 dprintk("%s wait for layoutreturn\n", __func__);
1591 if (pnfs_prepare_to_retry_layoutget(lo)) {
1592 if (first)
1593 pnfs_clear_first_layoutget(lo);
1594 pnfs_put_layout_hdr(lo);
1595 dprintk("%s retrying\n", __func__);
1596 goto lookup_again;
1597 }
1598 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1599 PNFS_UPDATE_LAYOUT_RETURN);
1600 goto out_put_layout_hdr;
1601 }
1602
1603 if (pnfs_layoutgets_blocked(lo)) {
1604 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1605 PNFS_UPDATE_LAYOUT_BLOCKED);
1606 goto out_unlock;
1607 }
1608 atomic_inc(&lo->plh_outstanding);
1609 spin_unlock(&ino->i_lock);
1610
1611 if (list_empty(&lo->plh_layouts)) {
1612 /* The lo must be on the clp list if there is any
1613 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1614 */
1615 spin_lock(&clp->cl_lock);
1616 if (list_empty(&lo->plh_layouts))
1617 list_add_tail(&lo->plh_layouts, &server->layouts);
1618 spin_unlock(&clp->cl_lock);
1619 }
1620
1621 pg_offset = arg.offset & ~PAGE_MASK;
1622 if (pg_offset) {
1623 arg.offset -= pg_offset;
1624 arg.length += pg_offset;
1625 }
1626 if (arg.length != NFS4_MAX_UINT64)
1627 arg.length = PAGE_ALIGN(arg.length);
1628
1629 lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
1630 atomic_dec(&lo->plh_outstanding);
1631 trace_pnfs_update_layout(ino, pos, count, iomode, lo,
1632 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
1633 out_put_layout_hdr:
1634 if (first)
1635 pnfs_clear_first_layoutget(lo);
1636 pnfs_put_layout_hdr(lo);
1637 out:
1638 dprintk("%s: inode %s/%llu pNFS layout segment %s for "
1639 "(%s, offset: %llu, length: %llu)\n",
1640 __func__, ino->i_sb->s_id,
1641 (unsigned long long)NFS_FILEID(ino),
1642 IS_ERR_OR_NULL(lseg) ? "not found" : "found",
1643 iomode==IOMODE_RW ? "read/write" : "read-only",
1644 (unsigned long long)pos,
1645 (unsigned long long)count);
1646 return lseg;
1647 out_unlock:
1648 spin_unlock(&ino->i_lock);
1649 goto out_put_layout_hdr;
1650 }
1651 EXPORT_SYMBOL_GPL(pnfs_update_layout);
1652
1653 static bool
1654 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
1655 {
1656 switch (range->iomode) {
1657 case IOMODE_READ:
1658 case IOMODE_RW:
1659 break;
1660 default:
1661 return false;
1662 }
1663 if (range->offset == NFS4_MAX_UINT64)
1664 return false;
1665 if (range->length == 0)
1666 return false;
1667 if (range->length != NFS4_MAX_UINT64 &&
1668 range->length > NFS4_MAX_UINT64 - range->offset)
1669 return false;
1670 return true;
1671 }
1672
1673 struct pnfs_layout_segment *
1674 pnfs_layout_process(struct nfs4_layoutget *lgp)
1675 {
1676 struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
1677 struct nfs4_layoutget_res *res = &lgp->res;
1678 struct pnfs_layout_segment *lseg;
1679 struct inode *ino = lo->plh_inode;
1680 LIST_HEAD(free_me);
1681 int status = -EINVAL;
1682
1683 if (!pnfs_sanity_check_layout_range(&res->range))
1684 goto out;
1685
1686 /* Inject layout blob into I/O device driver */
1687 lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
1688 if (!lseg || IS_ERR(lseg)) {
1689 if (!lseg)
1690 status = -ENOMEM;
1691 else
1692 status = PTR_ERR(lseg);
1693 dprintk("%s: Could not allocate layout: error %d\n",
1694 __func__, status);
1695 goto out;
1696 }
1697
1698 init_lseg(lo, lseg);
1699 lseg->pls_range = res->range;
1700
1701 spin_lock(&ino->i_lock);
1702 if (pnfs_layoutgets_blocked(lo)) {
1703 dprintk("%s forget reply due to state\n", __func__);
1704 goto out_forget_reply;
1705 }
1706
1707 if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
1708 /* existing state ID, make sure the sequence number matches. */
1709 if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
1710 dprintk("%s forget reply due to sequence\n", __func__);
1711 status = -EAGAIN;
1712 goto out_forget_reply;
1713 }
1714 pnfs_set_layout_stateid(lo, &res->stateid, false);
1715 } else {
1716 /*
1717 * We got an entirely new state ID. Mark all segments for the
1718 * inode invalid, and don't bother validating the stateid
1719 * sequence number.
1720 */
1721 pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
1722
1723 nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
1724 lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
1725 }
1726
1727 clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
1728
1729 pnfs_get_lseg(lseg);
1730 pnfs_layout_insert_lseg(lo, lseg, &free_me);
1731
1732 if (res->return_on_close)
1733 set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
1734
1735 spin_unlock(&ino->i_lock);
1736 pnfs_free_lseg_list(&free_me);
1737 return lseg;
1738 out:
1739 return ERR_PTR(status);
1740
1741 out_forget_reply:
1742 spin_unlock(&ino->i_lock);
1743 lseg->pls_layout = lo;
1744 NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
1745 goto out;
1746 }
1747
1748 static void
1749 pnfs_set_plh_return_iomode(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode)
1750 {
1751 if (lo->plh_return_iomode == iomode)
1752 return;
1753 if (lo->plh_return_iomode != 0)
1754 iomode = IOMODE_ANY;
1755 lo->plh_return_iomode = iomode;
1756 set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
1757 }
1758
1759 /**
1760 * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
1761 * @lo: pointer to layout header
1762 * @tmp_list: list header to be used with pnfs_free_lseg_list()
1763 * @return_range: describe layout segment ranges to be returned
1764 *
1765 * This function is mainly intended for use by layoutrecall. It attempts
1766 * to free the layout segment immediately, or else to mark it for return
1767 * as soon as its reference count drops to zero.
1768 */
1769 int
1770 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
1771 struct list_head *tmp_list,
1772 const struct pnfs_layout_range *return_range)
1773 {
1774 struct pnfs_layout_segment *lseg, *next;
1775 int remaining = 0;
1776
1777 dprintk("%s:Begin lo %p\n", __func__, lo);
1778
1779 if (list_empty(&lo->plh_segs))
1780 return 0;
1781
1782 assert_spin_locked(&lo->plh_inode->i_lock);
1783
1784 list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
1785 if (should_free_lseg(&lseg->pls_range, return_range)) {
1786 dprintk("%s: marking lseg %p iomode %d "
1787 "offset %llu length %llu\n", __func__,
1788 lseg, lseg->pls_range.iomode,
1789 lseg->pls_range.offset,
1790 lseg->pls_range.length);
1791 if (mark_lseg_invalid(lseg, tmp_list))
1792 continue;
1793 remaining++;
1794 set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1795 pnfs_set_plh_return_iomode(lo, return_range->iomode);
1796 }
1797 return remaining;
1798 }
1799
1800 void pnfs_error_mark_layout_for_return(struct inode *inode,
1801 struct pnfs_layout_segment *lseg)
1802 {
1803 struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
1804 struct pnfs_layout_range range = {
1805 .iomode = lseg->pls_range.iomode,
1806 .offset = 0,
1807 .length = NFS4_MAX_UINT64,
1808 };
1809 LIST_HEAD(free_me);
1810 bool return_now = false;
1811
1812 spin_lock(&inode->i_lock);
1813 pnfs_set_plh_return_iomode(lo, range.iomode);
1814 /*
1815 * mark all matching lsegs so that we are sure to have no live
1816 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
1817 * for how it works.
1818 */
1819 if (!pnfs_mark_matching_lsegs_return(lo, &free_me, &range)) {
1820 nfs4_stateid stateid;
1821 enum pnfs_iomode iomode = lo->plh_return_iomode;
1822
1823 nfs4_stateid_copy(&stateid, &lo->plh_stateid);
1824 return_now = pnfs_prepare_layoutreturn(lo);
1825 spin_unlock(&inode->i_lock);
1826 if (return_now)
1827 pnfs_send_layoutreturn(lo, &stateid, iomode, false);
1828 } else {
1829 spin_unlock(&inode->i_lock);
1830 nfs_commit_inode(inode, 0);
1831 }
1832 pnfs_free_lseg_list(&free_me);
1833 }
1834 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
1835
1836 void
1837 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
1838 {
1839 u64 rd_size = req->wb_bytes;
1840
1841 if (pgio->pg_lseg == NULL) {
1842 if (pgio->pg_dreq == NULL)
1843 rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
1844 else
1845 rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
1846
1847 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1848 req->wb_context,
1849 req_offset(req),
1850 rd_size,
1851 IOMODE_READ,
1852 GFP_KERNEL);
1853 if (IS_ERR(pgio->pg_lseg)) {
1854 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1855 pgio->pg_lseg = NULL;
1856 return;
1857 }
1858 }
1859 /* If no lseg, fall back to read through mds */
1860 if (pgio->pg_lseg == NULL)
1861 nfs_pageio_reset_read_mds(pgio);
1862
1863 }
1864 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
1865
1866 void
1867 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
1868 struct nfs_page *req, u64 wb_size)
1869 {
1870 if (pgio->pg_lseg == NULL) {
1871 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
1872 req->wb_context,
1873 req_offset(req),
1874 wb_size,
1875 IOMODE_RW,
1876 GFP_NOFS);
1877 if (IS_ERR(pgio->pg_lseg)) {
1878 pgio->pg_error = PTR_ERR(pgio->pg_lseg);
1879 pgio->pg_lseg = NULL;
1880 return;
1881 }
1882 }
1883 /* If no lseg, fall back to write through mds */
1884 if (pgio->pg_lseg == NULL)
1885 nfs_pageio_reset_write_mds(pgio);
1886 }
1887 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
1888
1889 void
1890 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
1891 {
1892 if (desc->pg_lseg) {
1893 pnfs_put_lseg(desc->pg_lseg);
1894 desc->pg_lseg = NULL;
1895 }
1896 }
1897 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
1898
1899 /*
1900 * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
1901 * of bytes (maximum @req->wb_bytes) that can be coalesced.
1902 */
1903 size_t
1904 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
1905 struct nfs_page *prev, struct nfs_page *req)
1906 {
1907 unsigned int size;
1908 u64 seg_end, req_start, seg_left;
1909
1910 size = nfs_generic_pg_test(pgio, prev, req);
1911 if (!size)
1912 return 0;
1913
1914 /*
1915 * 'size' contains the number of bytes left in the current page (up
1916 * to the original size asked for in @req->wb_bytes).
1917 *
1918 * Calculate how many bytes are left in the layout segment
1919 * and if there are less bytes than 'size', return that instead.
1920 *
1921 * Please also note that 'end_offset' is actually the offset of the
1922 * first byte that lies outside the pnfs_layout_range. FIXME?
1923 *
1924 */
1925 if (pgio->pg_lseg) {
1926 seg_end = end_offset(pgio->pg_lseg->pls_range.offset,
1927 pgio->pg_lseg->pls_range.length);
1928 req_start = req_offset(req);
1929 WARN_ON_ONCE(req_start >= seg_end);
1930 /* start of request is past the last byte of this segment */
1931 if (req_start >= seg_end) {
1932 /* reference the new lseg */
1933 if (pgio->pg_ops->pg_cleanup)
1934 pgio->pg_ops->pg_cleanup(pgio);
1935 if (pgio->pg_ops->pg_init)
1936 pgio->pg_ops->pg_init(pgio, req);
1937 return 0;
1938 }
1939
1940 /* adjust 'size' iff there are fewer bytes left in the
1941 * segment than what nfs_generic_pg_test returned */
1942 seg_left = seg_end - req_start;
1943 if (seg_left < size)
1944 size = (unsigned int)seg_left;
1945 }
1946
1947 return size;
1948 }
1949 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
1950
1951 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
1952 {
1953 struct nfs_pageio_descriptor pgio;
1954
1955 /* Resend all requests through the MDS */
1956 nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
1957 hdr->completion_ops);
1958 set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
1959 return nfs_pageio_resend(&pgio, hdr);
1960 }
1961 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
1962
1963 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
1964 {
1965
1966 dprintk("pnfs write error = %d\n", hdr->pnfs_error);
1967 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
1968 PNFS_LAYOUTRET_ON_ERROR) {
1969 pnfs_return_layout(hdr->inode);
1970 }
1971 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
1972 hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
1973 }
1974
1975 /*
1976 * Called by non rpc-based layout drivers
1977 */
1978 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
1979 {
1980 if (likely(!hdr->pnfs_error)) {
1981 pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
1982 hdr->mds_offset + hdr->res.count);
1983 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
1984 }
1985 trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
1986 if (unlikely(hdr->pnfs_error))
1987 pnfs_ld_handle_write_error(hdr);
1988 hdr->mds_ops->rpc_release(hdr);
1989 }
1990 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
1991
1992 static void
1993 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
1994 struct nfs_pgio_header *hdr)
1995 {
1996 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1997
1998 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
1999 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2000 nfs_pageio_reset_write_mds(desc);
2001 mirror->pg_recoalesce = 1;
2002 }
2003 nfs_pgio_data_destroy(hdr);
2004 hdr->release(hdr);
2005 }
2006
2007 static enum pnfs_try_status
2008 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2009 const struct rpc_call_ops *call_ops,
2010 struct pnfs_layout_segment *lseg,
2011 int how)
2012 {
2013 struct inode *inode = hdr->inode;
2014 enum pnfs_try_status trypnfs;
2015 struct nfs_server *nfss = NFS_SERVER(inode);
2016
2017 hdr->mds_ops = call_ops;
2018
2019 dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2020 inode->i_ino, hdr->args.count, hdr->args.offset, how);
2021 trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2022 if (trypnfs != PNFS_NOT_ATTEMPTED)
2023 nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2024 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2025 return trypnfs;
2026 }
2027
2028 static void
2029 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2030 struct nfs_pgio_header *hdr, int how)
2031 {
2032 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2033 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2034 enum pnfs_try_status trypnfs;
2035
2036 trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2037 if (trypnfs == PNFS_NOT_ATTEMPTED)
2038 pnfs_write_through_mds(desc, hdr);
2039 }
2040
2041 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2042 {
2043 pnfs_put_lseg(hdr->lseg);
2044 nfs_pgio_header_free(hdr);
2045 }
2046
2047 int
2048 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2049 {
2050 struct nfs_pgio_header *hdr;
2051 int ret;
2052
2053 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2054 if (!hdr) {
2055 desc->pg_error = -ENOMEM;
2056 return desc->pg_error;
2057 }
2058 nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2059
2060 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2061 ret = nfs_generic_pgio(desc, hdr);
2062 if (!ret)
2063 pnfs_do_write(desc, hdr, desc->pg_ioflags);
2064
2065 return ret;
2066 }
2067 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2068
2069 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2070 {
2071 struct nfs_pageio_descriptor pgio;
2072
2073 /* Resend all requests through the MDS */
2074 nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2075 return nfs_pageio_resend(&pgio, hdr);
2076 }
2077 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2078
2079 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2080 {
2081 dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2082 if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2083 PNFS_LAYOUTRET_ON_ERROR) {
2084 pnfs_return_layout(hdr->inode);
2085 }
2086 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2087 hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2088 }
2089
2090 /*
2091 * Called by non rpc-based layout drivers
2092 */
2093 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2094 {
2095 if (likely(!hdr->pnfs_error)) {
2096 __nfs4_read_done_cb(hdr);
2097 hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2098 }
2099 trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2100 if (unlikely(hdr->pnfs_error))
2101 pnfs_ld_handle_read_error(hdr);
2102 hdr->mds_ops->rpc_release(hdr);
2103 }
2104 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2105
2106 static void
2107 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2108 struct nfs_pgio_header *hdr)
2109 {
2110 struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2111
2112 if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2113 list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2114 nfs_pageio_reset_read_mds(desc);
2115 mirror->pg_recoalesce = 1;
2116 }
2117 nfs_pgio_data_destroy(hdr);
2118 hdr->release(hdr);
2119 }
2120
2121 /*
2122 * Call the appropriate parallel I/O subsystem read function.
2123 */
2124 static enum pnfs_try_status
2125 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2126 const struct rpc_call_ops *call_ops,
2127 struct pnfs_layout_segment *lseg)
2128 {
2129 struct inode *inode = hdr->inode;
2130 struct nfs_server *nfss = NFS_SERVER(inode);
2131 enum pnfs_try_status trypnfs;
2132
2133 hdr->mds_ops = call_ops;
2134
2135 dprintk("%s: Reading ino:%lu %u@%llu\n",
2136 __func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2137
2138 trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2139 if (trypnfs != PNFS_NOT_ATTEMPTED)
2140 nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2141 dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2142 return trypnfs;
2143 }
2144
2145 /* Resend all requests through pnfs. */
2146 int pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2147 {
2148 struct nfs_pageio_descriptor pgio;
2149
2150 nfs_pageio_init_read(&pgio, hdr->inode, false, hdr->completion_ops);
2151 return nfs_pageio_resend(&pgio, hdr);
2152 }
2153 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2154
2155 static void
2156 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2157 {
2158 const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2159 struct pnfs_layout_segment *lseg = desc->pg_lseg;
2160 enum pnfs_try_status trypnfs;
2161 int err = 0;
2162
2163 trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2164 if (trypnfs == PNFS_TRY_AGAIN)
2165 err = pnfs_read_resend_pnfs(hdr);
2166 if (trypnfs == PNFS_NOT_ATTEMPTED || err)
2167 pnfs_read_through_mds(desc, hdr);
2168 }
2169
2170 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2171 {
2172 pnfs_put_lseg(hdr->lseg);
2173 nfs_pgio_header_free(hdr);
2174 }
2175
2176 int
2177 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2178 {
2179 struct nfs_pgio_header *hdr;
2180 int ret;
2181
2182 hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2183 if (!hdr) {
2184 desc->pg_error = -ENOMEM;
2185 return desc->pg_error;
2186 }
2187 nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
2188 hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2189 ret = nfs_generic_pgio(desc, hdr);
2190 if (!ret)
2191 pnfs_do_read(desc, hdr);
2192 return ret;
2193 }
2194 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
2195
2196 static void pnfs_clear_layoutcommitting(struct inode *inode)
2197 {
2198 unsigned long *bitlock = &NFS_I(inode)->flags;
2199
2200 clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
2201 smp_mb__after_atomic();
2202 wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
2203 }
2204
2205 /*
2206 * There can be multiple RW segments.
2207 */
2208 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
2209 {
2210 struct pnfs_layout_segment *lseg;
2211
2212 list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
2213 if (lseg->pls_range.iomode == IOMODE_RW &&
2214 test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
2215 list_add(&lseg->pls_lc_list, listp);
2216 }
2217 }
2218
2219 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
2220 {
2221 struct pnfs_layout_segment *lseg, *tmp;
2222
2223 /* Matched by references in pnfs_set_layoutcommit */
2224 list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
2225 list_del_init(&lseg->pls_lc_list);
2226 pnfs_put_lseg(lseg);
2227 }
2228
2229 pnfs_clear_layoutcommitting(inode);
2230 }
2231
2232 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
2233 {
2234 pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
2235 }
2236 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
2237
2238 void
2239 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
2240 loff_t end_pos)
2241 {
2242 struct nfs_inode *nfsi = NFS_I(inode);
2243 bool mark_as_dirty = false;
2244
2245 spin_lock(&inode->i_lock);
2246 if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
2247 nfsi->layout->plh_lwb = end_pos;
2248 mark_as_dirty = true;
2249 dprintk("%s: Set layoutcommit for inode %lu ",
2250 __func__, inode->i_ino);
2251 } else if (end_pos > nfsi->layout->plh_lwb)
2252 nfsi->layout->plh_lwb = end_pos;
2253 if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
2254 /* references matched in nfs4_layoutcommit_release */
2255 pnfs_get_lseg(lseg);
2256 }
2257 spin_unlock(&inode->i_lock);
2258 dprintk("%s: lseg %p end_pos %llu\n",
2259 __func__, lseg, nfsi->layout->plh_lwb);
2260
2261 /* if pnfs_layoutcommit_inode() runs between inode locks, the next one
2262 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
2263 if (mark_as_dirty)
2264 mark_inode_dirty_sync(inode);
2265 }
2266 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
2267
2268 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
2269 {
2270 struct nfs_server *nfss = NFS_SERVER(data->args.inode);
2271
2272 if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
2273 nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
2274 pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
2275 }
2276
2277 /*
2278 * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
2279 * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
2280 * data to disk to allow the server to recover the data if it crashes.
2281 * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
2282 * is off, and a COMMIT is sent to a data server, or
2283 * if WRITEs to a data server return NFS_DATA_SYNC.
2284 */
2285 int
2286 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
2287 {
2288 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2289 struct nfs4_layoutcommit_data *data;
2290 struct nfs_inode *nfsi = NFS_I(inode);
2291 loff_t end_pos;
2292 int status;
2293
2294 if (!pnfs_layoutcommit_outstanding(inode))
2295 return 0;
2296
2297 dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
2298
2299 status = -EAGAIN;
2300 if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
2301 if (!sync)
2302 goto out;
2303 status = wait_on_bit_lock_action(&nfsi->flags,
2304 NFS_INO_LAYOUTCOMMITTING,
2305 nfs_wait_bit_killable,
2306 TASK_KILLABLE);
2307 if (status)
2308 goto out;
2309 }
2310
2311 status = -ENOMEM;
2312 /* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
2313 data = kzalloc(sizeof(*data), GFP_NOFS);
2314 if (!data)
2315 goto clear_layoutcommitting;
2316
2317 status = 0;
2318 spin_lock(&inode->i_lock);
2319 if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
2320 goto out_unlock;
2321
2322 INIT_LIST_HEAD(&data->lseg_list);
2323 pnfs_list_write_lseg(inode, &data->lseg_list);
2324
2325 end_pos = nfsi->layout->plh_lwb;
2326
2327 nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
2328 spin_unlock(&inode->i_lock);
2329
2330 data->args.inode = inode;
2331 data->cred = get_rpccred(nfsi->layout->plh_lc_cred);
2332 nfs_fattr_init(&data->fattr);
2333 data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
2334 data->res.fattr = &data->fattr;
2335 data->args.lastbytewritten = end_pos - 1;
2336 data->res.server = NFS_SERVER(inode);
2337
2338 if (ld->prepare_layoutcommit) {
2339 status = ld->prepare_layoutcommit(&data->args);
2340 if (status) {
2341 put_rpccred(data->cred);
2342 spin_lock(&inode->i_lock);
2343 set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
2344 if (end_pos > nfsi->layout->plh_lwb)
2345 nfsi->layout->plh_lwb = end_pos;
2346 goto out_unlock;
2347 }
2348 }
2349
2350
2351 status = nfs4_proc_layoutcommit(data, sync);
2352 out:
2353 if (status)
2354 mark_inode_dirty_sync(inode);
2355 dprintk("<-- %s status %d\n", __func__, status);
2356 return status;
2357 out_unlock:
2358 spin_unlock(&inode->i_lock);
2359 kfree(data);
2360 clear_layoutcommitting:
2361 pnfs_clear_layoutcommitting(inode);
2362 goto out;
2363 }
2364 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
2365
2366 int
2367 pnfs_generic_sync(struct inode *inode, bool datasync)
2368 {
2369 return pnfs_layoutcommit_inode(inode, true);
2370 }
2371 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
2372
2373 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
2374 {
2375 struct nfs4_threshold *thp;
2376
2377 thp = kzalloc(sizeof(*thp), GFP_NOFS);
2378 if (!thp) {
2379 dprintk("%s mdsthreshold allocation failed\n", __func__);
2380 return NULL;
2381 }
2382 return thp;
2383 }
2384
2385 #if IS_ENABLED(CONFIG_NFS_V4_2)
2386 int
2387 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
2388 {
2389 struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
2390 struct nfs_server *server = NFS_SERVER(inode);
2391 struct nfs_inode *nfsi = NFS_I(inode);
2392 struct nfs42_layoutstat_data *data;
2393 struct pnfs_layout_hdr *hdr;
2394 int status = 0;
2395
2396 if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
2397 goto out;
2398
2399 if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
2400 goto out;
2401
2402 if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
2403 goto out;
2404
2405 spin_lock(&inode->i_lock);
2406 if (!NFS_I(inode)->layout) {
2407 spin_unlock(&inode->i_lock);
2408 goto out;
2409 }
2410 hdr = NFS_I(inode)->layout;
2411 pnfs_get_layout_hdr(hdr);
2412 spin_unlock(&inode->i_lock);
2413
2414 data = kzalloc(sizeof(*data), gfp_flags);
2415 if (!data) {
2416 status = -ENOMEM;
2417 goto out_put;
2418 }
2419
2420 data->args.fh = NFS_FH(inode);
2421 data->args.inode = inode;
2422 nfs4_stateid_copy(&data->args.stateid, &hdr->plh_stateid);
2423 status = ld->prepare_layoutstats(&data->args);
2424 if (status)
2425 goto out_free;
2426
2427 status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
2428
2429 out:
2430 dprintk("%s returns %d\n", __func__, status);
2431 return status;
2432
2433 out_free:
2434 kfree(data);
2435 out_put:
2436 pnfs_put_layout_hdr(hdr);
2437 smp_mb__before_atomic();
2438 clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
2439 smp_mb__after_atomic();
2440 goto out;
2441 }
2442 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
2443 #endif
2444
2445 unsigned int layoutstats_timer;
2446 module_param(layoutstats_timer, uint, 0644);
2447 EXPORT_SYMBOL_GPL(layoutstats_timer);