xref: /openbmc/linux/fs/nfs/pnfs.c (revision e95770af)
1 /*
2  *  pNFS functions to call and manage layout drivers.
3  *
4  *  Copyright (c) 2002 [year of first publication]
5  *  The Regents of the University of Michigan
6  *  All Rights Reserved
7  *
8  *  Dean Hildebrand <dhildebz@umich.edu>
9  *
10  *  Permission is granted to use, copy, create derivative works, and
11  *  redistribute this software and such derivative works for any purpose,
12  *  so long as the name of the University of Michigan is not used in
13  *  any advertising or publicity pertaining to the use or distribution
14  *  of this software without specific, written prior authorization. If
15  *  the above copyright notice or any other identification of the
16  *  University of Michigan is included in any copy of any portion of
17  *  this software, then the disclaimer below must also be included.
18  *
19  *  This software is provided as is, without representation or warranty
20  *  of any kind either express or implied, including without limitation
21  *  the implied warranties of merchantability, fitness for a particular
22  *  purpose, or noninfringement.  The Regents of the University of
23  *  Michigan shall not be liable for any damages, including special,
24  *  indirect, incidental, or consequential damages, with respect to any
25  *  claim arising out of or in connection with the use of the software,
26  *  even if it has been or is hereafter advised of the possibility of
27  *  such damages.
28  */
29 
30 #include <linux/nfs_fs.h>
31 #include <linux/nfs_page.h>
32 #include <linux/module.h>
33 #include <linux/sort.h>
34 #include "internal.h"
35 #include "pnfs.h"
36 #include "iostat.h"
37 #include "nfs4trace.h"
38 #include "delegation.h"
39 #include "nfs42.h"
40 #include "nfs4_fs.h"
41 
42 #define NFSDBG_FACILITY		NFSDBG_PNFS
43 #define PNFS_LAYOUTGET_RETRY_TIMEOUT (120*HZ)
44 
45 /* Locking:
46  *
47  * pnfs_spinlock:
48  *      protects pnfs_modules_tbl.
49  */
50 static DEFINE_SPINLOCK(pnfs_spinlock);
51 
52 /*
53  * pnfs_modules_tbl holds all pnfs modules
54  */
55 static LIST_HEAD(pnfs_modules_tbl);
56 
57 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo);
58 static void pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
59 		struct list_head *free_me,
60 		const struct pnfs_layout_range *range,
61 		u32 seq);
62 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
63 		                struct list_head *tmp_list);
64 
65 /* Return the registered pnfs layout driver module matching given id */
66 static struct pnfs_layoutdriver_type *
67 find_pnfs_driver_locked(u32 id)
68 {
69 	struct pnfs_layoutdriver_type *local;
70 
71 	list_for_each_entry(local, &pnfs_modules_tbl, pnfs_tblid)
72 		if (local->id == id)
73 			goto out;
74 	local = NULL;
75 out:
76 	dprintk("%s: Searching for id %u, found %p\n", __func__, id, local);
77 	return local;
78 }
79 
80 static struct pnfs_layoutdriver_type *
81 find_pnfs_driver(u32 id)
82 {
83 	struct pnfs_layoutdriver_type *local;
84 
85 	spin_lock(&pnfs_spinlock);
86 	local = find_pnfs_driver_locked(id);
87 	if (local != NULL && !try_module_get(local->owner)) {
88 		dprintk("%s: Could not grab reference on module\n", __func__);
89 		local = NULL;
90 	}
91 	spin_unlock(&pnfs_spinlock);
92 	return local;
93 }
94 
95 void
96 unset_pnfs_layoutdriver(struct nfs_server *nfss)
97 {
98 	if (nfss->pnfs_curr_ld) {
99 		if (nfss->pnfs_curr_ld->clear_layoutdriver)
100 			nfss->pnfs_curr_ld->clear_layoutdriver(nfss);
101 		/* Decrement the MDS count. Purge the deviceid cache if zero */
102 		if (atomic_dec_and_test(&nfss->nfs_client->cl_mds_count))
103 			nfs4_deviceid_purge_client(nfss->nfs_client);
104 		module_put(nfss->pnfs_curr_ld->owner);
105 	}
106 	nfss->pnfs_curr_ld = NULL;
107 }
108 
109 /*
110  * When the server sends a list of layout types, we choose one in the order
111  * given in the list below.
112  *
113  * FIXME: should this list be configurable in some fashion? module param?
114  * 	  mount option? something else?
115  */
116 static const u32 ld_prefs[] = {
117 	LAYOUT_SCSI,
118 	LAYOUT_BLOCK_VOLUME,
119 	LAYOUT_OSD2_OBJECTS,
120 	LAYOUT_FLEX_FILES,
121 	LAYOUT_NFSV4_1_FILES,
122 	0
123 };
124 
125 static int
126 ld_cmp(const void *e1, const void *e2)
127 {
128 	u32 ld1 = *((u32 *)e1);
129 	u32 ld2 = *((u32 *)e2);
130 	int i;
131 
132 	for (i = 0; ld_prefs[i] != 0; i++) {
133 		if (ld1 == ld_prefs[i])
134 			return -1;
135 
136 		if (ld2 == ld_prefs[i])
137 			return 1;
138 	}
139 	return 0;
140 }
141 
142 /*
143  * Try to set the server's pnfs module to the pnfs layout type specified by id.
144  * Currently only one pNFS layout driver per filesystem is supported.
145  *
146  * @ids array of layout types supported by MDS.
147  */
148 void
149 set_pnfs_layoutdriver(struct nfs_server *server, const struct nfs_fh *mntfh,
150 		      struct nfs_fsinfo *fsinfo)
151 {
152 	struct pnfs_layoutdriver_type *ld_type = NULL;
153 	u32 id;
154 	int i;
155 
156 	if (fsinfo->nlayouttypes == 0)
157 		goto out_no_driver;
158 	if (!(server->nfs_client->cl_exchange_flags &
159 		 (EXCHGID4_FLAG_USE_NON_PNFS | EXCHGID4_FLAG_USE_PNFS_MDS))) {
160 		printk(KERN_ERR "NFS: %s: cl_exchange_flags 0x%x\n",
161 			__func__, server->nfs_client->cl_exchange_flags);
162 		goto out_no_driver;
163 	}
164 
165 	sort(fsinfo->layouttype, fsinfo->nlayouttypes,
166 		sizeof(*fsinfo->layouttype), ld_cmp, NULL);
167 
168 	for (i = 0; i < fsinfo->nlayouttypes; i++) {
169 		id = fsinfo->layouttype[i];
170 		ld_type = find_pnfs_driver(id);
171 		if (!ld_type) {
172 			request_module("%s-%u", LAYOUT_NFSV4_1_MODULE_PREFIX,
173 					id);
174 			ld_type = find_pnfs_driver(id);
175 		}
176 		if (ld_type)
177 			break;
178 	}
179 
180 	if (!ld_type) {
181 		dprintk("%s: No pNFS module found!\n", __func__);
182 		goto out_no_driver;
183 	}
184 
185 	server->pnfs_curr_ld = ld_type;
186 	if (ld_type->set_layoutdriver
187 	    && ld_type->set_layoutdriver(server, mntfh)) {
188 		printk(KERN_ERR "NFS: %s: Error initializing pNFS layout "
189 			"driver %u.\n", __func__, id);
190 		module_put(ld_type->owner);
191 		goto out_no_driver;
192 	}
193 	/* Bump the MDS count */
194 	atomic_inc(&server->nfs_client->cl_mds_count);
195 
196 	dprintk("%s: pNFS module for %u set\n", __func__, id);
197 	return;
198 
199 out_no_driver:
200 	dprintk("%s: Using NFSv4 I/O\n", __func__);
201 	server->pnfs_curr_ld = NULL;
202 }
203 
204 int
205 pnfs_register_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
206 {
207 	int status = -EINVAL;
208 	struct pnfs_layoutdriver_type *tmp;
209 
210 	if (ld_type->id == 0) {
211 		printk(KERN_ERR "NFS: %s id 0 is reserved\n", __func__);
212 		return status;
213 	}
214 	if (!ld_type->alloc_lseg || !ld_type->free_lseg) {
215 		printk(KERN_ERR "NFS: %s Layout driver must provide "
216 		       "alloc_lseg and free_lseg.\n", __func__);
217 		return status;
218 	}
219 
220 	spin_lock(&pnfs_spinlock);
221 	tmp = find_pnfs_driver_locked(ld_type->id);
222 	if (!tmp) {
223 		list_add(&ld_type->pnfs_tblid, &pnfs_modules_tbl);
224 		status = 0;
225 		dprintk("%s Registering id:%u name:%s\n", __func__, ld_type->id,
226 			ld_type->name);
227 	} else {
228 		printk(KERN_ERR "NFS: %s Module with id %d already loaded!\n",
229 			__func__, ld_type->id);
230 	}
231 	spin_unlock(&pnfs_spinlock);
232 
233 	return status;
234 }
235 EXPORT_SYMBOL_GPL(pnfs_register_layoutdriver);
236 
237 void
238 pnfs_unregister_layoutdriver(struct pnfs_layoutdriver_type *ld_type)
239 {
240 	dprintk("%s Deregistering id:%u\n", __func__, ld_type->id);
241 	spin_lock(&pnfs_spinlock);
242 	list_del(&ld_type->pnfs_tblid);
243 	spin_unlock(&pnfs_spinlock);
244 }
245 EXPORT_SYMBOL_GPL(pnfs_unregister_layoutdriver);
246 
247 /*
248  * pNFS client layout cache
249  */
250 
251 /* Need to hold i_lock if caller does not already hold reference */
252 void
253 pnfs_get_layout_hdr(struct pnfs_layout_hdr *lo)
254 {
255 	refcount_inc(&lo->plh_refcount);
256 }
257 
258 static struct pnfs_layout_hdr *
259 pnfs_alloc_layout_hdr(struct inode *ino, gfp_t gfp_flags)
260 {
261 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
262 	return ld->alloc_layout_hdr(ino, gfp_flags);
263 }
264 
265 static void
266 pnfs_free_layout_hdr(struct pnfs_layout_hdr *lo)
267 {
268 	struct nfs_server *server = NFS_SERVER(lo->plh_inode);
269 	struct pnfs_layoutdriver_type *ld = server->pnfs_curr_ld;
270 
271 	if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
272 		struct nfs_client *clp = server->nfs_client;
273 
274 		spin_lock(&clp->cl_lock);
275 		list_del_rcu(&lo->plh_layouts);
276 		spin_unlock(&clp->cl_lock);
277 	}
278 	put_cred(lo->plh_lc_cred);
279 	return ld->free_layout_hdr(lo);
280 }
281 
282 static void
283 pnfs_detach_layout_hdr(struct pnfs_layout_hdr *lo)
284 {
285 	struct nfs_inode *nfsi = NFS_I(lo->plh_inode);
286 	dprintk("%s: freeing layout cache %p\n", __func__, lo);
287 	nfsi->layout = NULL;
288 	/* Reset MDS Threshold I/O counters */
289 	nfsi->write_io = 0;
290 	nfsi->read_io = 0;
291 }
292 
293 void
294 pnfs_put_layout_hdr(struct pnfs_layout_hdr *lo)
295 {
296 	struct inode *inode;
297 
298 	if (!lo)
299 		return;
300 	inode = lo->plh_inode;
301 	pnfs_layoutreturn_before_put_layout_hdr(lo);
302 
303 	if (refcount_dec_and_lock(&lo->plh_refcount, &inode->i_lock)) {
304 		if (!list_empty(&lo->plh_segs))
305 			WARN_ONCE(1, "NFS: BUG unfreed layout segments.\n");
306 		pnfs_detach_layout_hdr(lo);
307 		spin_unlock(&inode->i_lock);
308 		pnfs_free_layout_hdr(lo);
309 	}
310 }
311 
312 static struct inode *
313 pnfs_grab_inode_layout_hdr(struct pnfs_layout_hdr *lo)
314 {
315 	struct inode *inode = igrab(lo->plh_inode);
316 	if (inode)
317 		return inode;
318 	set_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags);
319 	return NULL;
320 }
321 
322 static void
323 pnfs_set_plh_return_info(struct pnfs_layout_hdr *lo, enum pnfs_iomode iomode,
324 			 u32 seq)
325 {
326 	if (lo->plh_return_iomode != 0 && lo->plh_return_iomode != iomode)
327 		iomode = IOMODE_ANY;
328 	lo->plh_return_iomode = iomode;
329 	set_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
330 	if (seq != 0) {
331 		WARN_ON_ONCE(lo->plh_return_seq != 0 && lo->plh_return_seq != seq);
332 		lo->plh_return_seq = seq;
333 	}
334 }
335 
336 static void
337 pnfs_clear_layoutreturn_info(struct pnfs_layout_hdr *lo)
338 {
339 	struct pnfs_layout_segment *lseg;
340 	lo->plh_return_iomode = 0;
341 	lo->plh_return_seq = 0;
342 	clear_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags);
343 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
344 		if (!test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
345 			continue;
346 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
347 	}
348 }
349 
350 static void pnfs_clear_layoutreturn_waitbit(struct pnfs_layout_hdr *lo)
351 {
352 	clear_bit_unlock(NFS_LAYOUT_RETURN, &lo->plh_flags);
353 	clear_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags);
354 	smp_mb__after_atomic();
355 	wake_up_bit(&lo->plh_flags, NFS_LAYOUT_RETURN);
356 	rpc_wake_up(&NFS_SERVER(lo->plh_inode)->roc_rpcwaitq);
357 }
358 
359 static void
360 pnfs_clear_lseg_state(struct pnfs_layout_segment *lseg,
361 		struct list_head *free_me)
362 {
363 	clear_bit(NFS_LSEG_ROC, &lseg->pls_flags);
364 	clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
365 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags))
366 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
367 	if (test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
368 		pnfs_lseg_dec_and_remove_zero(lseg, free_me);
369 }
370 
371 /*
372  * Update the seqid of a layout stateid after receiving
373  * NFS4ERR_OLD_STATEID
374  */
375 bool nfs4_layout_refresh_old_stateid(nfs4_stateid *dst,
376 		struct pnfs_layout_range *dst_range,
377 		struct inode *inode)
378 {
379 	struct pnfs_layout_hdr *lo;
380 	struct pnfs_layout_range range = {
381 		.iomode = IOMODE_ANY,
382 		.offset = 0,
383 		.length = NFS4_MAX_UINT64,
384 	};
385 	bool ret = false;
386 	LIST_HEAD(head);
387 	int err;
388 
389 	spin_lock(&inode->i_lock);
390 	lo = NFS_I(inode)->layout;
391 	if (lo &&  pnfs_layout_is_valid(lo) &&
392 	    nfs4_stateid_match_other(dst, &lo->plh_stateid)) {
393 		/* Is our call using the most recent seqid? If so, bump it */
394 		if (!nfs4_stateid_is_newer(&lo->plh_stateid, dst)) {
395 			nfs4_stateid_seqid_inc(dst);
396 			ret = true;
397 			goto out;
398 		}
399 		/* Try to update the seqid to the most recent */
400 		err = pnfs_mark_matching_lsegs_return(lo, &head, &range, 0);
401 		if (err != -EBUSY) {
402 			dst->seqid = lo->plh_stateid.seqid;
403 			*dst_range = range;
404 			ret = true;
405 		}
406 	}
407 out:
408 	spin_unlock(&inode->i_lock);
409 	pnfs_free_lseg_list(&head);
410 	return ret;
411 }
412 
413 /*
414  * Mark a pnfs_layout_hdr and all associated layout segments as invalid
415  *
416  * In order to continue using the pnfs_layout_hdr, a full recovery
417  * is required.
418  * Note that caller must hold inode->i_lock.
419  */
420 int
421 pnfs_mark_layout_stateid_invalid(struct pnfs_layout_hdr *lo,
422 		struct list_head *lseg_list)
423 {
424 	struct pnfs_layout_range range = {
425 		.iomode = IOMODE_ANY,
426 		.offset = 0,
427 		.length = NFS4_MAX_UINT64,
428 	};
429 	struct pnfs_layout_segment *lseg, *next;
430 
431 	set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
432 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
433 		pnfs_clear_lseg_state(lseg, lseg_list);
434 	pnfs_clear_layoutreturn_info(lo);
435 	pnfs_free_returned_lsegs(lo, lseg_list, &range, 0);
436 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags) &&
437 	    !test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
438 		pnfs_clear_layoutreturn_waitbit(lo);
439 	return !list_empty(&lo->plh_segs);
440 }
441 
442 static int
443 pnfs_iomode_to_fail_bit(u32 iomode)
444 {
445 	return iomode == IOMODE_RW ?
446 		NFS_LAYOUT_RW_FAILED : NFS_LAYOUT_RO_FAILED;
447 }
448 
449 static void
450 pnfs_layout_set_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
451 {
452 	lo->plh_retry_timestamp = jiffies;
453 	if (!test_and_set_bit(fail_bit, &lo->plh_flags))
454 		refcount_inc(&lo->plh_refcount);
455 }
456 
457 static void
458 pnfs_layout_clear_fail_bit(struct pnfs_layout_hdr *lo, int fail_bit)
459 {
460 	if (test_and_clear_bit(fail_bit, &lo->plh_flags))
461 		refcount_dec(&lo->plh_refcount);
462 }
463 
464 static void
465 pnfs_layout_io_set_failed(struct pnfs_layout_hdr *lo, u32 iomode)
466 {
467 	struct inode *inode = lo->plh_inode;
468 	struct pnfs_layout_range range = {
469 		.iomode = iomode,
470 		.offset = 0,
471 		.length = NFS4_MAX_UINT64,
472 	};
473 	LIST_HEAD(head);
474 
475 	spin_lock(&inode->i_lock);
476 	pnfs_layout_set_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
477 	pnfs_mark_matching_lsegs_invalid(lo, &head, &range, 0);
478 	spin_unlock(&inode->i_lock);
479 	pnfs_free_lseg_list(&head);
480 	dprintk("%s Setting layout IOMODE_%s fail bit\n", __func__,
481 			iomode == IOMODE_RW ?  "RW" : "READ");
482 }
483 
484 static bool
485 pnfs_layout_io_test_failed(struct pnfs_layout_hdr *lo, u32 iomode)
486 {
487 	unsigned long start, end;
488 	int fail_bit = pnfs_iomode_to_fail_bit(iomode);
489 
490 	if (test_bit(fail_bit, &lo->plh_flags) == 0)
491 		return false;
492 	end = jiffies;
493 	start = end - PNFS_LAYOUTGET_RETRY_TIMEOUT;
494 	if (!time_in_range(lo->plh_retry_timestamp, start, end)) {
495 		/* It is time to retry the failed layoutgets */
496 		pnfs_layout_clear_fail_bit(lo, fail_bit);
497 		return false;
498 	}
499 	return true;
500 }
501 
502 static void
503 pnfs_init_lseg(struct pnfs_layout_hdr *lo, struct pnfs_layout_segment *lseg,
504 		const struct pnfs_layout_range *range,
505 		const nfs4_stateid *stateid)
506 {
507 	INIT_LIST_HEAD(&lseg->pls_list);
508 	INIT_LIST_HEAD(&lseg->pls_lc_list);
509 	INIT_LIST_HEAD(&lseg->pls_commits);
510 	refcount_set(&lseg->pls_refcount, 1);
511 	set_bit(NFS_LSEG_VALID, &lseg->pls_flags);
512 	lseg->pls_layout = lo;
513 	lseg->pls_range = *range;
514 	lseg->pls_seq = be32_to_cpu(stateid->seqid);
515 }
516 
517 static void pnfs_free_lseg(struct pnfs_layout_segment *lseg)
518 {
519 	if (lseg != NULL) {
520 		struct inode *inode = lseg->pls_layout->plh_inode;
521 		NFS_SERVER(inode)->pnfs_curr_ld->free_lseg(lseg);
522 	}
523 }
524 
525 static void
526 pnfs_layout_remove_lseg(struct pnfs_layout_hdr *lo,
527 		struct pnfs_layout_segment *lseg)
528 {
529 	WARN_ON(test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
530 	list_del_init(&lseg->pls_list);
531 	/* Matched by pnfs_get_layout_hdr in pnfs_layout_insert_lseg */
532 	refcount_dec(&lo->plh_refcount);
533 	if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
534 		return;
535 	if (list_empty(&lo->plh_segs) &&
536 	    !test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) &&
537 	    !test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
538 		if (atomic_read(&lo->plh_outstanding) == 0)
539 			set_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
540 		clear_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
541 	}
542 }
543 
544 static bool
545 pnfs_cache_lseg_for_layoutreturn(struct pnfs_layout_hdr *lo,
546 		struct pnfs_layout_segment *lseg)
547 {
548 	if (test_and_clear_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags) &&
549 	    pnfs_layout_is_valid(lo)) {
550 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
551 		list_move_tail(&lseg->pls_list, &lo->plh_return_segs);
552 		return true;
553 	}
554 	return false;
555 }
556 
557 void
558 pnfs_put_lseg(struct pnfs_layout_segment *lseg)
559 {
560 	struct pnfs_layout_hdr *lo;
561 	struct inode *inode;
562 
563 	if (!lseg)
564 		return;
565 
566 	dprintk("%s: lseg %p ref %d valid %d\n", __func__, lseg,
567 		refcount_read(&lseg->pls_refcount),
568 		test_bit(NFS_LSEG_VALID, &lseg->pls_flags));
569 
570 	lo = lseg->pls_layout;
571 	inode = lo->plh_inode;
572 
573 	if (refcount_dec_and_lock(&lseg->pls_refcount, &inode->i_lock)) {
574 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
575 			spin_unlock(&inode->i_lock);
576 			return;
577 		}
578 		pnfs_get_layout_hdr(lo);
579 		pnfs_layout_remove_lseg(lo, lseg);
580 		if (pnfs_cache_lseg_for_layoutreturn(lo, lseg))
581 			lseg = NULL;
582 		spin_unlock(&inode->i_lock);
583 		pnfs_free_lseg(lseg);
584 		pnfs_put_layout_hdr(lo);
585 	}
586 }
587 EXPORT_SYMBOL_GPL(pnfs_put_lseg);
588 
589 /*
590  * is l2 fully contained in l1?
591  *   start1                             end1
592  *   [----------------------------------)
593  *           start2           end2
594  *           [----------------)
595  */
596 static bool
597 pnfs_lseg_range_contained(const struct pnfs_layout_range *l1,
598 		 const struct pnfs_layout_range *l2)
599 {
600 	u64 start1 = l1->offset;
601 	u64 end1 = pnfs_end_offset(start1, l1->length);
602 	u64 start2 = l2->offset;
603 	u64 end2 = pnfs_end_offset(start2, l2->length);
604 
605 	return (start1 <= start2) && (end1 >= end2);
606 }
607 
608 static bool pnfs_lseg_dec_and_remove_zero(struct pnfs_layout_segment *lseg,
609 		struct list_head *tmp_list)
610 {
611 	if (!refcount_dec_and_test(&lseg->pls_refcount))
612 		return false;
613 	pnfs_layout_remove_lseg(lseg->pls_layout, lseg);
614 	list_add(&lseg->pls_list, tmp_list);
615 	return true;
616 }
617 
618 /* Returns 1 if lseg is removed from list, 0 otherwise */
619 static int mark_lseg_invalid(struct pnfs_layout_segment *lseg,
620 			     struct list_head *tmp_list)
621 {
622 	int rv = 0;
623 
624 	if (test_and_clear_bit(NFS_LSEG_VALID, &lseg->pls_flags)) {
625 		/* Remove the reference keeping the lseg in the
626 		 * list.  It will now be removed when all
627 		 * outstanding io is finished.
628 		 */
629 		dprintk("%s: lseg %p ref %d\n", __func__, lseg,
630 			refcount_read(&lseg->pls_refcount));
631 		if (pnfs_lseg_dec_and_remove_zero(lseg, tmp_list))
632 			rv = 1;
633 	}
634 	return rv;
635 }
636 
637 /*
638  * Compare 2 layout stateid sequence ids, to see which is newer,
639  * taking into account wraparound issues.
640  */
641 static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
642 {
643 	return (s32)(s1 - s2) > 0;
644 }
645 
646 static bool
647 pnfs_should_free_range(const struct pnfs_layout_range *lseg_range,
648 		 const struct pnfs_layout_range *recall_range)
649 {
650 	return (recall_range->iomode == IOMODE_ANY ||
651 		lseg_range->iomode == recall_range->iomode) &&
652 	       pnfs_lseg_range_intersecting(lseg_range, recall_range);
653 }
654 
655 static bool
656 pnfs_match_lseg_recall(const struct pnfs_layout_segment *lseg,
657 		const struct pnfs_layout_range *recall_range,
658 		u32 seq)
659 {
660 	if (seq != 0 && pnfs_seqid_is_newer(lseg->pls_seq, seq))
661 		return false;
662 	if (recall_range == NULL)
663 		return true;
664 	return pnfs_should_free_range(&lseg->pls_range, recall_range);
665 }
666 
667 /**
668  * pnfs_mark_matching_lsegs_invalid - tear down lsegs or mark them for later
669  * @lo: layout header containing the lsegs
670  * @tmp_list: list head where doomed lsegs should go
671  * @recall_range: optional recall range argument to match (may be NULL)
672  * @seq: only invalidate lsegs obtained prior to this sequence (may be 0)
673  *
674  * Walk the list of lsegs in the layout header, and tear down any that should
675  * be destroyed. If "recall_range" is specified then the segment must match
676  * that range. If "seq" is non-zero, then only match segments that were handed
677  * out at or before that sequence.
678  *
679  * Returns number of matching invalid lsegs remaining in list after scanning
680  * it and purging them.
681  */
682 int
683 pnfs_mark_matching_lsegs_invalid(struct pnfs_layout_hdr *lo,
684 			    struct list_head *tmp_list,
685 			    const struct pnfs_layout_range *recall_range,
686 			    u32 seq)
687 {
688 	struct pnfs_layout_segment *lseg, *next;
689 	int remaining = 0;
690 
691 	dprintk("%s:Begin lo %p\n", __func__, lo);
692 
693 	if (list_empty(&lo->plh_segs))
694 		return 0;
695 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
696 		if (pnfs_match_lseg_recall(lseg, recall_range, seq)) {
697 			dprintk("%s: freeing lseg %p iomode %d seq %u "
698 				"offset %llu length %llu\n", __func__,
699 				lseg, lseg->pls_range.iomode, lseg->pls_seq,
700 				lseg->pls_range.offset, lseg->pls_range.length);
701 			if (!mark_lseg_invalid(lseg, tmp_list))
702 				remaining++;
703 		}
704 	dprintk("%s:Return %i\n", __func__, remaining);
705 	return remaining;
706 }
707 
708 static void
709 pnfs_free_returned_lsegs(struct pnfs_layout_hdr *lo,
710 		struct list_head *free_me,
711 		const struct pnfs_layout_range *range,
712 		u32 seq)
713 {
714 	struct pnfs_layout_segment *lseg, *next;
715 
716 	list_for_each_entry_safe(lseg, next, &lo->plh_return_segs, pls_list) {
717 		if (pnfs_match_lseg_recall(lseg, range, seq))
718 			list_move_tail(&lseg->pls_list, free_me);
719 	}
720 }
721 
722 /* note free_me must contain lsegs from a single layout_hdr */
723 void
724 pnfs_free_lseg_list(struct list_head *free_me)
725 {
726 	struct pnfs_layout_segment *lseg, *tmp;
727 
728 	if (list_empty(free_me))
729 		return;
730 
731 	list_for_each_entry_safe(lseg, tmp, free_me, pls_list) {
732 		list_del(&lseg->pls_list);
733 		pnfs_free_lseg(lseg);
734 	}
735 }
736 
737 void
738 pnfs_destroy_layout(struct nfs_inode *nfsi)
739 {
740 	struct pnfs_layout_hdr *lo;
741 	LIST_HEAD(tmp_list);
742 
743 	spin_lock(&nfsi->vfs_inode.i_lock);
744 	lo = nfsi->layout;
745 	if (lo) {
746 		pnfs_get_layout_hdr(lo);
747 		pnfs_mark_layout_stateid_invalid(lo, &tmp_list);
748 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RO_FAILED);
749 		pnfs_layout_clear_fail_bit(lo, NFS_LAYOUT_RW_FAILED);
750 		spin_unlock(&nfsi->vfs_inode.i_lock);
751 		pnfs_free_lseg_list(&tmp_list);
752 		nfs_commit_inode(&nfsi->vfs_inode, 0);
753 		pnfs_put_layout_hdr(lo);
754 	} else
755 		spin_unlock(&nfsi->vfs_inode.i_lock);
756 }
757 EXPORT_SYMBOL_GPL(pnfs_destroy_layout);
758 
759 static bool
760 pnfs_layout_add_bulk_destroy_list(struct inode *inode,
761 		struct list_head *layout_list)
762 {
763 	struct pnfs_layout_hdr *lo;
764 	bool ret = false;
765 
766 	spin_lock(&inode->i_lock);
767 	lo = NFS_I(inode)->layout;
768 	if (lo != NULL && list_empty(&lo->plh_bulk_destroy)) {
769 		pnfs_get_layout_hdr(lo);
770 		list_add(&lo->plh_bulk_destroy, layout_list);
771 		ret = true;
772 	}
773 	spin_unlock(&inode->i_lock);
774 	return ret;
775 }
776 
777 /* Caller must hold rcu_read_lock and clp->cl_lock */
778 static int
779 pnfs_layout_bulk_destroy_byserver_locked(struct nfs_client *clp,
780 		struct nfs_server *server,
781 		struct list_head *layout_list)
782 	__must_hold(&clp->cl_lock)
783 	__must_hold(RCU)
784 {
785 	struct pnfs_layout_hdr *lo, *next;
786 	struct inode *inode;
787 
788 	list_for_each_entry_safe(lo, next, &server->layouts, plh_layouts) {
789 		if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags) ||
790 		    test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) ||
791 		    !list_empty(&lo->plh_bulk_destroy))
792 			continue;
793 		/* If the sb is being destroyed, just bail */
794 		if (!nfs_sb_active(server->super))
795 			break;
796 		inode = pnfs_grab_inode_layout_hdr(lo);
797 		if (inode != NULL) {
798 			if (test_and_clear_bit(NFS_LAYOUT_HASHED, &lo->plh_flags))
799 				list_del_rcu(&lo->plh_layouts);
800 			if (pnfs_layout_add_bulk_destroy_list(inode,
801 						layout_list))
802 				continue;
803 			rcu_read_unlock();
804 			spin_unlock(&clp->cl_lock);
805 			iput(inode);
806 		} else {
807 			rcu_read_unlock();
808 			spin_unlock(&clp->cl_lock);
809 		}
810 		nfs_sb_deactive(server->super);
811 		spin_lock(&clp->cl_lock);
812 		rcu_read_lock();
813 		return -EAGAIN;
814 	}
815 	return 0;
816 }
817 
818 static int
819 pnfs_layout_free_bulk_destroy_list(struct list_head *layout_list,
820 		bool is_bulk_recall)
821 {
822 	struct pnfs_layout_hdr *lo;
823 	struct inode *inode;
824 	LIST_HEAD(lseg_list);
825 	int ret = 0;
826 
827 	while (!list_empty(layout_list)) {
828 		lo = list_entry(layout_list->next, struct pnfs_layout_hdr,
829 				plh_bulk_destroy);
830 		dprintk("%s freeing layout for inode %lu\n", __func__,
831 			lo->plh_inode->i_ino);
832 		inode = lo->plh_inode;
833 
834 		pnfs_layoutcommit_inode(inode, false);
835 
836 		spin_lock(&inode->i_lock);
837 		list_del_init(&lo->plh_bulk_destroy);
838 		if (pnfs_mark_layout_stateid_invalid(lo, &lseg_list)) {
839 			if (is_bulk_recall)
840 				set_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
841 			ret = -EAGAIN;
842 		}
843 		spin_unlock(&inode->i_lock);
844 		pnfs_free_lseg_list(&lseg_list);
845 		/* Free all lsegs that are attached to commit buckets */
846 		nfs_commit_inode(inode, 0);
847 		pnfs_put_layout_hdr(lo);
848 		nfs_iput_and_deactive(inode);
849 	}
850 	return ret;
851 }
852 
853 int
854 pnfs_destroy_layouts_byfsid(struct nfs_client *clp,
855 		struct nfs_fsid *fsid,
856 		bool is_recall)
857 {
858 	struct nfs_server *server;
859 	LIST_HEAD(layout_list);
860 
861 	spin_lock(&clp->cl_lock);
862 	rcu_read_lock();
863 restart:
864 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
865 		if (memcmp(&server->fsid, fsid, sizeof(*fsid)) != 0)
866 			continue;
867 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
868 				server,
869 				&layout_list) != 0)
870 			goto restart;
871 	}
872 	rcu_read_unlock();
873 	spin_unlock(&clp->cl_lock);
874 
875 	if (list_empty(&layout_list))
876 		return 0;
877 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
878 }
879 
880 int
881 pnfs_destroy_layouts_byclid(struct nfs_client *clp,
882 		bool is_recall)
883 {
884 	struct nfs_server *server;
885 	LIST_HEAD(layout_list);
886 
887 	spin_lock(&clp->cl_lock);
888 	rcu_read_lock();
889 restart:
890 	list_for_each_entry_rcu(server, &clp->cl_superblocks, client_link) {
891 		if (pnfs_layout_bulk_destroy_byserver_locked(clp,
892 					server,
893 					&layout_list) != 0)
894 			goto restart;
895 	}
896 	rcu_read_unlock();
897 	spin_unlock(&clp->cl_lock);
898 
899 	if (list_empty(&layout_list))
900 		return 0;
901 	return pnfs_layout_free_bulk_destroy_list(&layout_list, is_recall);
902 }
903 
904 /*
905  * Called by the state manger to remove all layouts established under an
906  * expired lease.
907  */
908 void
909 pnfs_destroy_all_layouts(struct nfs_client *clp)
910 {
911 	nfs4_deviceid_mark_client_invalid(clp);
912 	nfs4_deviceid_purge_client(clp);
913 
914 	pnfs_destroy_layouts_byclid(clp, false);
915 }
916 
917 static void
918 pnfs_set_layout_cred(struct pnfs_layout_hdr *lo, const struct cred *cred)
919 {
920 	const struct cred *old;
921 
922 	if (cred && cred_fscmp(lo->plh_lc_cred, cred) != 0) {
923 		old = xchg(&lo->plh_lc_cred, get_cred(cred));
924 		put_cred(old);
925 	}
926 }
927 
928 /* update lo->plh_stateid with new if is more recent */
929 void
930 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
931 			const struct cred *cred, bool update_barrier)
932 {
933 	u32 oldseq, newseq, new_barrier = 0;
934 
935 	oldseq = be32_to_cpu(lo->plh_stateid.seqid);
936 	newseq = be32_to_cpu(new->seqid);
937 
938 	if (!pnfs_layout_is_valid(lo)) {
939 		pnfs_set_layout_cred(lo, cred);
940 		nfs4_stateid_copy(&lo->plh_stateid, new);
941 		lo->plh_barrier = newseq;
942 		pnfs_clear_layoutreturn_info(lo);
943 		clear_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags);
944 		return;
945 	}
946 	if (pnfs_seqid_is_newer(newseq, oldseq)) {
947 		nfs4_stateid_copy(&lo->plh_stateid, new);
948 		/*
949 		 * Because of wraparound, we want to keep the barrier
950 		 * "close" to the current seqids.
951 		 */
952 		new_barrier = newseq - atomic_read(&lo->plh_outstanding);
953 	}
954 	if (update_barrier)
955 		new_barrier = be32_to_cpu(new->seqid);
956 	else if (new_barrier == 0)
957 		return;
958 	if (pnfs_seqid_is_newer(new_barrier, lo->plh_barrier))
959 		lo->plh_barrier = new_barrier;
960 }
961 
962 static bool
963 pnfs_layout_stateid_blocked(const struct pnfs_layout_hdr *lo,
964 		const nfs4_stateid *stateid)
965 {
966 	u32 seqid = be32_to_cpu(stateid->seqid);
967 
968 	return !pnfs_seqid_is_newer(seqid, lo->plh_barrier);
969 }
970 
971 /* lget is set to 1 if called from inside send_layoutget call chain */
972 static bool
973 pnfs_layoutgets_blocked(const struct pnfs_layout_hdr *lo)
974 {
975 	return lo->plh_block_lgets ||
976 		test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags);
977 }
978 
979 static struct nfs_server *
980 pnfs_find_server(struct inode *inode, struct nfs_open_context *ctx)
981 {
982 	struct nfs_server *server;
983 
984 	if (inode) {
985 		server = NFS_SERVER(inode);
986 	} else {
987 		struct dentry *parent_dir = dget_parent(ctx->dentry);
988 		server = NFS_SERVER(parent_dir->d_inode);
989 		dput(parent_dir);
990 	}
991 	return server;
992 }
993 
994 static void nfs4_free_pages(struct page **pages, size_t size)
995 {
996 	int i;
997 
998 	if (!pages)
999 		return;
1000 
1001 	for (i = 0; i < size; i++) {
1002 		if (!pages[i])
1003 			break;
1004 		__free_page(pages[i]);
1005 	}
1006 	kfree(pages);
1007 }
1008 
1009 static struct page **nfs4_alloc_pages(size_t size, gfp_t gfp_flags)
1010 {
1011 	struct page **pages;
1012 	int i;
1013 
1014 	pages = kmalloc_array(size, sizeof(struct page *), gfp_flags);
1015 	if (!pages) {
1016 		dprintk("%s: can't alloc array of %zu pages\n", __func__, size);
1017 		return NULL;
1018 	}
1019 
1020 	for (i = 0; i < size; i++) {
1021 		pages[i] = alloc_page(gfp_flags);
1022 		if (!pages[i]) {
1023 			dprintk("%s: failed to allocate page\n", __func__);
1024 			nfs4_free_pages(pages, i);
1025 			return NULL;
1026 		}
1027 	}
1028 
1029 	return pages;
1030 }
1031 
1032 static struct nfs4_layoutget *
1033 pnfs_alloc_init_layoutget_args(struct inode *ino,
1034 	   struct nfs_open_context *ctx,
1035 	   const nfs4_stateid *stateid,
1036 	   const struct pnfs_layout_range *range,
1037 	   gfp_t gfp_flags)
1038 {
1039 	struct nfs_server *server = pnfs_find_server(ino, ctx);
1040 	size_t max_reply_sz = server->pnfs_curr_ld->max_layoutget_response;
1041 	size_t max_pages = max_response_pages(server);
1042 	struct nfs4_layoutget *lgp;
1043 
1044 	dprintk("--> %s\n", __func__);
1045 
1046 	lgp = kzalloc(sizeof(*lgp), gfp_flags);
1047 	if (lgp == NULL)
1048 		return NULL;
1049 
1050 	if (max_reply_sz) {
1051 		size_t npages = (max_reply_sz + PAGE_SIZE - 1) >> PAGE_SHIFT;
1052 		if (npages < max_pages)
1053 			max_pages = npages;
1054 	}
1055 
1056 	lgp->args.layout.pages = nfs4_alloc_pages(max_pages, gfp_flags);
1057 	if (!lgp->args.layout.pages) {
1058 		kfree(lgp);
1059 		return NULL;
1060 	}
1061 	lgp->args.layout.pglen = max_pages * PAGE_SIZE;
1062 	lgp->res.layoutp = &lgp->args.layout;
1063 
1064 	/* Don't confuse uninitialised result and success */
1065 	lgp->res.status = -NFS4ERR_DELAY;
1066 
1067 	lgp->args.minlength = PAGE_SIZE;
1068 	if (lgp->args.minlength > range->length)
1069 		lgp->args.minlength = range->length;
1070 	if (ino) {
1071 		loff_t i_size = i_size_read(ino);
1072 
1073 		if (range->iomode == IOMODE_READ) {
1074 			if (range->offset >= i_size)
1075 				lgp->args.minlength = 0;
1076 			else if (i_size - range->offset < lgp->args.minlength)
1077 				lgp->args.minlength = i_size - range->offset;
1078 		}
1079 	}
1080 	lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
1081 	pnfs_copy_range(&lgp->args.range, range);
1082 	lgp->args.type = server->pnfs_curr_ld->id;
1083 	lgp->args.inode = ino;
1084 	lgp->args.ctx = get_nfs_open_context(ctx);
1085 	nfs4_stateid_copy(&lgp->args.stateid, stateid);
1086 	lgp->gfp_flags = gfp_flags;
1087 	lgp->cred = ctx->cred;
1088 	return lgp;
1089 }
1090 
1091 void pnfs_layoutget_free(struct nfs4_layoutget *lgp)
1092 {
1093 	size_t max_pages = lgp->args.layout.pglen / PAGE_SIZE;
1094 
1095 	nfs4_free_pages(lgp->args.layout.pages, max_pages);
1096 	if (lgp->args.inode)
1097 		pnfs_put_layout_hdr(NFS_I(lgp->args.inode)->layout);
1098 	put_nfs_open_context(lgp->args.ctx);
1099 	kfree(lgp);
1100 }
1101 
1102 static void pnfs_clear_layoutcommit(struct inode *inode,
1103 		struct list_head *head)
1104 {
1105 	struct nfs_inode *nfsi = NFS_I(inode);
1106 	struct pnfs_layout_segment *lseg, *tmp;
1107 
1108 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
1109 		return;
1110 	list_for_each_entry_safe(lseg, tmp, &nfsi->layout->plh_segs, pls_list) {
1111 		if (!test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
1112 			continue;
1113 		pnfs_lseg_dec_and_remove_zero(lseg, head);
1114 	}
1115 }
1116 
1117 void pnfs_layoutreturn_free_lsegs(struct pnfs_layout_hdr *lo,
1118 		const nfs4_stateid *arg_stateid,
1119 		const struct pnfs_layout_range *range,
1120 		const nfs4_stateid *stateid)
1121 {
1122 	struct inode *inode = lo->plh_inode;
1123 	LIST_HEAD(freeme);
1124 
1125 	spin_lock(&inode->i_lock);
1126 	if (!pnfs_layout_is_valid(lo) || !arg_stateid ||
1127 	    !nfs4_stateid_match_other(&lo->plh_stateid, arg_stateid))
1128 		goto out_unlock;
1129 	if (stateid) {
1130 		u32 seq = be32_to_cpu(arg_stateid->seqid);
1131 
1132 		pnfs_mark_matching_lsegs_invalid(lo, &freeme, range, seq);
1133 		pnfs_free_returned_lsegs(lo, &freeme, range, seq);
1134 		pnfs_set_layout_stateid(lo, stateid, NULL, true);
1135 	} else
1136 		pnfs_mark_layout_stateid_invalid(lo, &freeme);
1137 out_unlock:
1138 	pnfs_clear_layoutreturn_waitbit(lo);
1139 	spin_unlock(&inode->i_lock);
1140 	pnfs_free_lseg_list(&freeme);
1141 
1142 }
1143 
1144 static bool
1145 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo,
1146 		nfs4_stateid *stateid,
1147 		const struct cred **cred,
1148 		enum pnfs_iomode *iomode)
1149 {
1150 	/* Serialise LAYOUTGET/LAYOUTRETURN */
1151 	if (atomic_read(&lo->plh_outstanding) != 0)
1152 		return false;
1153 	if (test_and_set_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags))
1154 		return false;
1155 	set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
1156 	pnfs_get_layout_hdr(lo);
1157 	if (test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags)) {
1158 		nfs4_stateid_copy(stateid, &lo->plh_stateid);
1159 		*cred = get_cred(lo->plh_lc_cred);
1160 		if (lo->plh_return_seq != 0)
1161 			stateid->seqid = cpu_to_be32(lo->plh_return_seq);
1162 		if (iomode != NULL)
1163 			*iomode = lo->plh_return_iomode;
1164 		pnfs_clear_layoutreturn_info(lo);
1165 		return true;
1166 	}
1167 	nfs4_stateid_copy(stateid, &lo->plh_stateid);
1168 	*cred = get_cred(lo->plh_lc_cred);
1169 	if (iomode != NULL)
1170 		*iomode = IOMODE_ANY;
1171 	return true;
1172 }
1173 
1174 static void
1175 pnfs_init_layoutreturn_args(struct nfs4_layoutreturn_args *args,
1176 		struct pnfs_layout_hdr *lo,
1177 		const nfs4_stateid *stateid,
1178 		enum pnfs_iomode iomode)
1179 {
1180 	struct inode *inode = lo->plh_inode;
1181 
1182 	args->layout_type = NFS_SERVER(inode)->pnfs_curr_ld->id;
1183 	args->inode = inode;
1184 	args->range.iomode = iomode;
1185 	args->range.offset = 0;
1186 	args->range.length = NFS4_MAX_UINT64;
1187 	args->layout = lo;
1188 	nfs4_stateid_copy(&args->stateid, stateid);
1189 }
1190 
1191 static int
1192 pnfs_send_layoutreturn(struct pnfs_layout_hdr *lo,
1193 		       const nfs4_stateid *stateid,
1194 		       const struct cred **pcred,
1195 		       enum pnfs_iomode iomode,
1196 		       bool sync)
1197 {
1198 	struct inode *ino = lo->plh_inode;
1199 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1200 	struct nfs4_layoutreturn *lrp;
1201 	const struct cred *cred = *pcred;
1202 	int status = 0;
1203 
1204 	*pcred = NULL;
1205 	lrp = kzalloc(sizeof(*lrp), GFP_NOFS);
1206 	if (unlikely(lrp == NULL)) {
1207 		status = -ENOMEM;
1208 		spin_lock(&ino->i_lock);
1209 		pnfs_clear_layoutreturn_waitbit(lo);
1210 		spin_unlock(&ino->i_lock);
1211 		put_cred(cred);
1212 		pnfs_put_layout_hdr(lo);
1213 		goto out;
1214 	}
1215 
1216 	pnfs_init_layoutreturn_args(&lrp->args, lo, stateid, iomode);
1217 	lrp->args.ld_private = &lrp->ld_private;
1218 	lrp->clp = NFS_SERVER(ino)->nfs_client;
1219 	lrp->cred = cred;
1220 	if (ld->prepare_layoutreturn)
1221 		ld->prepare_layoutreturn(&lrp->args);
1222 
1223 	status = nfs4_proc_layoutreturn(lrp, sync);
1224 out:
1225 	dprintk("<-- %s status: %d\n", __func__, status);
1226 	return status;
1227 }
1228 
1229 /* Return true if layoutreturn is needed */
1230 static bool
1231 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1232 {
1233 	struct pnfs_layout_segment *s;
1234 	enum pnfs_iomode iomode;
1235 	u32 seq;
1236 
1237 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1238 		return false;
1239 
1240 	seq = lo->plh_return_seq;
1241 	iomode = lo->plh_return_iomode;
1242 
1243 	/* Defer layoutreturn until all recalled lsegs are done */
1244 	list_for_each_entry(s, &lo->plh_segs, pls_list) {
1245 		if (seq && pnfs_seqid_is_newer(s->pls_seq, seq))
1246 			continue;
1247 		if (iomode != IOMODE_ANY && s->pls_range.iomode != iomode)
1248 			continue;
1249 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &s->pls_flags))
1250 			return false;
1251 	}
1252 
1253 	return true;
1254 }
1255 
1256 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1257 {
1258 	struct inode *inode= lo->plh_inode;
1259 
1260 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1261 		return;
1262 	spin_lock(&inode->i_lock);
1263 	if (pnfs_layout_need_return(lo)) {
1264 		const struct cred *cred;
1265 		nfs4_stateid stateid;
1266 		enum pnfs_iomode iomode;
1267 		bool send;
1268 
1269 		send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1270 		spin_unlock(&inode->i_lock);
1271 		if (send) {
1272 			/* Send an async layoutreturn so we dont deadlock */
1273 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1274 		}
1275 	} else
1276 		spin_unlock(&inode->i_lock);
1277 }
1278 
1279 /*
1280  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1281  * when the layout segment list is empty.
1282  *
1283  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1284  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1285  * deviceid is marked invalid.
1286  */
1287 int
1288 _pnfs_return_layout(struct inode *ino)
1289 {
1290 	struct pnfs_layout_hdr *lo = NULL;
1291 	struct nfs_inode *nfsi = NFS_I(ino);
1292 	LIST_HEAD(tmp_list);
1293 	const struct cred *cred;
1294 	nfs4_stateid stateid;
1295 	int status = 0;
1296 	bool send, valid_layout;
1297 
1298 	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1299 
1300 	spin_lock(&ino->i_lock);
1301 	lo = nfsi->layout;
1302 	if (!lo) {
1303 		spin_unlock(&ino->i_lock);
1304 		dprintk("NFS: %s no layout to return\n", __func__);
1305 		goto out;
1306 	}
1307 	/* Reference matched in nfs4_layoutreturn_release */
1308 	pnfs_get_layout_hdr(lo);
1309 	/* Is there an outstanding layoutreturn ? */
1310 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1311 		spin_unlock(&ino->i_lock);
1312 		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1313 					TASK_UNINTERRUPTIBLE))
1314 			goto out_put_layout_hdr;
1315 		spin_lock(&ino->i_lock);
1316 	}
1317 	valid_layout = pnfs_layout_is_valid(lo);
1318 	pnfs_clear_layoutcommit(ino, &tmp_list);
1319 	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1320 
1321 	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1322 		struct pnfs_layout_range range = {
1323 			.iomode		= IOMODE_ANY,
1324 			.offset		= 0,
1325 			.length		= NFS4_MAX_UINT64,
1326 		};
1327 		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1328 	}
1329 
1330 	/* Don't send a LAYOUTRETURN if list was initially empty */
1331 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1332 			!valid_layout) {
1333 		spin_unlock(&ino->i_lock);
1334 		dprintk("NFS: %s no layout segments to return\n", __func__);
1335 		goto out_put_layout_hdr;
1336 	}
1337 
1338 	send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1339 	spin_unlock(&ino->i_lock);
1340 	if (send)
1341 		status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1342 out_put_layout_hdr:
1343 	pnfs_free_lseg_list(&tmp_list);
1344 	pnfs_put_layout_hdr(lo);
1345 out:
1346 	dprintk("<-- %s status: %d\n", __func__, status);
1347 	return status;
1348 }
1349 
1350 int
1351 pnfs_commit_and_return_layout(struct inode *inode)
1352 {
1353 	struct pnfs_layout_hdr *lo;
1354 	int ret;
1355 
1356 	spin_lock(&inode->i_lock);
1357 	lo = NFS_I(inode)->layout;
1358 	if (lo == NULL) {
1359 		spin_unlock(&inode->i_lock);
1360 		return 0;
1361 	}
1362 	pnfs_get_layout_hdr(lo);
1363 	/* Block new layoutgets and read/write to ds */
1364 	lo->plh_block_lgets++;
1365 	spin_unlock(&inode->i_lock);
1366 	filemap_fdatawait(inode->i_mapping);
1367 	ret = pnfs_layoutcommit_inode(inode, true);
1368 	if (ret == 0)
1369 		ret = _pnfs_return_layout(inode);
1370 	spin_lock(&inode->i_lock);
1371 	lo->plh_block_lgets--;
1372 	spin_unlock(&inode->i_lock);
1373 	pnfs_put_layout_hdr(lo);
1374 	return ret;
1375 }
1376 
1377 bool pnfs_roc(struct inode *ino,
1378 		struct nfs4_layoutreturn_args *args,
1379 		struct nfs4_layoutreturn_res *res,
1380 		const struct cred *cred)
1381 {
1382 	struct nfs_inode *nfsi = NFS_I(ino);
1383 	struct nfs_open_context *ctx;
1384 	struct nfs4_state *state;
1385 	struct pnfs_layout_hdr *lo;
1386 	struct pnfs_layout_segment *lseg, *next;
1387 	const struct cred *lc_cred;
1388 	nfs4_stateid stateid;
1389 	enum pnfs_iomode iomode = 0;
1390 	bool layoutreturn = false, roc = false;
1391 	bool skip_read = false;
1392 
1393 	if (!nfs_have_layout(ino))
1394 		return false;
1395 retry:
1396 	rcu_read_lock();
1397 	spin_lock(&ino->i_lock);
1398 	lo = nfsi->layout;
1399 	if (!lo || !pnfs_layout_is_valid(lo) ||
1400 	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1401 		lo = NULL;
1402 		goto out_noroc;
1403 	}
1404 	pnfs_get_layout_hdr(lo);
1405 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1406 		spin_unlock(&ino->i_lock);
1407 		rcu_read_unlock();
1408 		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1409 				TASK_UNINTERRUPTIBLE);
1410 		pnfs_put_layout_hdr(lo);
1411 		goto retry;
1412 	}
1413 
1414 	/* no roc if we hold a delegation */
1415 	if (nfs4_check_delegation(ino, FMODE_READ)) {
1416 		if (nfs4_check_delegation(ino, FMODE_WRITE))
1417 			goto out_noroc;
1418 		skip_read = true;
1419 	}
1420 
1421 	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1422 		state = ctx->state;
1423 		if (state == NULL)
1424 			continue;
1425 		/* Don't return layout if there is open file state */
1426 		if (state->state & FMODE_WRITE)
1427 			goto out_noroc;
1428 		if (state->state & FMODE_READ)
1429 			skip_read = true;
1430 	}
1431 
1432 
1433 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1434 		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1435 			continue;
1436 		/* If we are sending layoutreturn, invalidate all valid lsegs */
1437 		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1438 			continue;
1439 		/*
1440 		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1441 		 * doesn't invalidate the layout for us.
1442 		 */
1443 		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1444 		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1445 			continue;
1446 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1447 	}
1448 
1449 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1450 		goto out_noroc;
1451 
1452 	/* ROC in two conditions:
1453 	 * 1. there are ROC lsegs
1454 	 * 2. we don't send layoutreturn
1455 	 */
1456 	/* lo ref dropped in pnfs_roc_release() */
1457 	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1458 	/* If the creds don't match, we can't compound the layoutreturn */
1459 	if (!layoutreturn)
1460 		goto out_noroc;
1461 	if (cred_fscmp(cred, lc_cred) != 0)
1462 		goto out_noroc_put_cred;
1463 
1464 	roc = layoutreturn;
1465 	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1466 	res->lrs_present = 0;
1467 	layoutreturn = false;
1468 
1469 out_noroc_put_cred:
1470 	put_cred(lc_cred);
1471 out_noroc:
1472 	spin_unlock(&ino->i_lock);
1473 	rcu_read_unlock();
1474 	pnfs_layoutcommit_inode(ino, true);
1475 	if (roc) {
1476 		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1477 		if (ld->prepare_layoutreturn)
1478 			ld->prepare_layoutreturn(args);
1479 		pnfs_put_layout_hdr(lo);
1480 		return true;
1481 	}
1482 	if (layoutreturn)
1483 		pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1484 	pnfs_put_layout_hdr(lo);
1485 	return false;
1486 }
1487 
1488 int pnfs_roc_done(struct rpc_task *task, struct inode *inode,
1489 		struct nfs4_layoutreturn_args **argpp,
1490 		struct nfs4_layoutreturn_res **respp,
1491 		int *ret)
1492 {
1493 	struct nfs4_layoutreturn_args *arg = *argpp;
1494 	int retval = -EAGAIN;
1495 
1496 	if (!arg)
1497 		return 0;
1498 	/* Handle Layoutreturn errors */
1499 	switch (*ret) {
1500 	case 0:
1501 		retval = 0;
1502 		break;
1503 	case -NFS4ERR_NOMATCHING_LAYOUT:
1504 		/* Was there an RPC level error? If not, retry */
1505 		if (task->tk_rpc_status == 0)
1506 			break;
1507 		/* If the call was not sent, let caller handle it */
1508 		if (!RPC_WAS_SENT(task))
1509 			return 0;
1510 		/*
1511 		 * Otherwise, assume the call succeeded and
1512 		 * that we need to release the layout
1513 		 */
1514 		*ret = 0;
1515 		(*respp)->lrs_present = 0;
1516 		retval = 0;
1517 		break;
1518 	case -NFS4ERR_DELAY:
1519 		/* Let the caller handle the retry */
1520 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1521 		return 0;
1522 	case -NFS4ERR_OLD_STATEID:
1523 		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1524 					&arg->range, inode))
1525 			break;
1526 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1527 		return -EAGAIN;
1528 	}
1529 	*argpp = NULL;
1530 	*respp = NULL;
1531 	return retval;
1532 }
1533 
1534 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1535 		struct nfs4_layoutreturn_res *res,
1536 		int ret)
1537 {
1538 	struct pnfs_layout_hdr *lo = args->layout;
1539 	const nfs4_stateid *arg_stateid = NULL;
1540 	const nfs4_stateid *res_stateid = NULL;
1541 	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1542 
1543 	switch (ret) {
1544 	case -NFS4ERR_NOMATCHING_LAYOUT:
1545 		break;
1546 	case 0:
1547 		if (res->lrs_present)
1548 			res_stateid = &res->stateid;
1549 		/* Fallthrough */
1550 	default:
1551 		arg_stateid = &args->stateid;
1552 	}
1553 	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1554 			res_stateid);
1555 	if (ld_private && ld_private->ops && ld_private->ops->free)
1556 		ld_private->ops->free(ld_private);
1557 	pnfs_put_layout_hdr(lo);
1558 	trace_nfs4_layoutreturn_on_close(args->inode, 0);
1559 }
1560 
1561 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1562 {
1563 	struct nfs_inode *nfsi = NFS_I(ino);
1564         struct pnfs_layout_hdr *lo;
1565         bool sleep = false;
1566 
1567 	/* we might not have grabbed lo reference. so need to check under
1568 	 * i_lock */
1569         spin_lock(&ino->i_lock);
1570         lo = nfsi->layout;
1571         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1572                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1573                 sleep = true;
1574 	}
1575         spin_unlock(&ino->i_lock);
1576         return sleep;
1577 }
1578 
1579 /*
1580  * Compare two layout segments for sorting into layout cache.
1581  * We want to preferentially return RW over RO layouts, so ensure those
1582  * are seen first.
1583  */
1584 static s64
1585 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1586 	   const struct pnfs_layout_range *l2)
1587 {
1588 	s64 d;
1589 
1590 	/* high offset > low offset */
1591 	d = l1->offset - l2->offset;
1592 	if (d)
1593 		return d;
1594 
1595 	/* short length > long length */
1596 	d = l2->length - l1->length;
1597 	if (d)
1598 		return d;
1599 
1600 	/* read > read/write */
1601 	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1602 }
1603 
1604 static bool
1605 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1606 		const struct pnfs_layout_range *l2)
1607 {
1608 	return pnfs_lseg_range_cmp(l1, l2) > 0;
1609 }
1610 
1611 static bool
1612 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1613 		struct pnfs_layout_segment *old)
1614 {
1615 	return false;
1616 }
1617 
1618 void
1619 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1620 		   struct pnfs_layout_segment *lseg,
1621 		   bool (*is_after)(const struct pnfs_layout_range *,
1622 			   const struct pnfs_layout_range *),
1623 		   bool (*do_merge)(struct pnfs_layout_segment *,
1624 			   struct pnfs_layout_segment *),
1625 		   struct list_head *free_me)
1626 {
1627 	struct pnfs_layout_segment *lp, *tmp;
1628 
1629 	dprintk("%s:Begin\n", __func__);
1630 
1631 	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1632 		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1633 			continue;
1634 		if (do_merge(lseg, lp)) {
1635 			mark_lseg_invalid(lp, free_me);
1636 			continue;
1637 		}
1638 		if (is_after(&lseg->pls_range, &lp->pls_range))
1639 			continue;
1640 		list_add_tail(&lseg->pls_list, &lp->pls_list);
1641 		dprintk("%s: inserted lseg %p "
1642 			"iomode %d offset %llu length %llu before "
1643 			"lp %p iomode %d offset %llu length %llu\n",
1644 			__func__, lseg, lseg->pls_range.iomode,
1645 			lseg->pls_range.offset, lseg->pls_range.length,
1646 			lp, lp->pls_range.iomode, lp->pls_range.offset,
1647 			lp->pls_range.length);
1648 		goto out;
1649 	}
1650 	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1651 	dprintk("%s: inserted lseg %p "
1652 		"iomode %d offset %llu length %llu at tail\n",
1653 		__func__, lseg, lseg->pls_range.iomode,
1654 		lseg->pls_range.offset, lseg->pls_range.length);
1655 out:
1656 	pnfs_get_layout_hdr(lo);
1657 
1658 	dprintk("%s:Return\n", __func__);
1659 }
1660 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1661 
1662 static void
1663 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1664 		   struct pnfs_layout_segment *lseg,
1665 		   struct list_head *free_me)
1666 {
1667 	struct inode *inode = lo->plh_inode;
1668 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1669 
1670 	if (ld->add_lseg != NULL)
1671 		ld->add_lseg(lo, lseg, free_me);
1672 	else
1673 		pnfs_generic_layout_insert_lseg(lo, lseg,
1674 				pnfs_lseg_range_is_after,
1675 				pnfs_lseg_no_merge,
1676 				free_me);
1677 }
1678 
1679 static struct pnfs_layout_hdr *
1680 alloc_init_layout_hdr(struct inode *ino,
1681 		      struct nfs_open_context *ctx,
1682 		      gfp_t gfp_flags)
1683 {
1684 	struct pnfs_layout_hdr *lo;
1685 
1686 	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1687 	if (!lo)
1688 		return NULL;
1689 	refcount_set(&lo->plh_refcount, 1);
1690 	INIT_LIST_HEAD(&lo->plh_layouts);
1691 	INIT_LIST_HEAD(&lo->plh_segs);
1692 	INIT_LIST_HEAD(&lo->plh_return_segs);
1693 	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1694 	lo->plh_inode = ino;
1695 	lo->plh_lc_cred = get_cred(ctx->cred);
1696 	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1697 	return lo;
1698 }
1699 
1700 static struct pnfs_layout_hdr *
1701 pnfs_find_alloc_layout(struct inode *ino,
1702 		       struct nfs_open_context *ctx,
1703 		       gfp_t gfp_flags)
1704 	__releases(&ino->i_lock)
1705 	__acquires(&ino->i_lock)
1706 {
1707 	struct nfs_inode *nfsi = NFS_I(ino);
1708 	struct pnfs_layout_hdr *new = NULL;
1709 
1710 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1711 
1712 	if (nfsi->layout != NULL)
1713 		goto out_existing;
1714 	spin_unlock(&ino->i_lock);
1715 	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1716 	spin_lock(&ino->i_lock);
1717 
1718 	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1719 		nfsi->layout = new;
1720 		return new;
1721 	} else if (new != NULL)
1722 		pnfs_free_layout_hdr(new);
1723 out_existing:
1724 	pnfs_get_layout_hdr(nfsi->layout);
1725 	return nfsi->layout;
1726 }
1727 
1728 /*
1729  * iomode matching rules:
1730  * iomode	lseg	strict match
1731  *                      iomode
1732  * -----	-----	------ -----
1733  * ANY		READ	N/A    true
1734  * ANY		RW	N/A    true
1735  * RW		READ	N/A    false
1736  * RW		RW	N/A    true
1737  * READ		READ	N/A    true
1738  * READ		RW	true   false
1739  * READ		RW	false  true
1740  */
1741 static bool
1742 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1743 		 const struct pnfs_layout_range *range,
1744 		 bool strict_iomode)
1745 {
1746 	struct pnfs_layout_range range1;
1747 
1748 	if ((range->iomode == IOMODE_RW &&
1749 	     ls_range->iomode != IOMODE_RW) ||
1750 	    (range->iomode != ls_range->iomode &&
1751 	     strict_iomode) ||
1752 	    !pnfs_lseg_range_intersecting(ls_range, range))
1753 		return false;
1754 
1755 	/* range1 covers only the first byte in the range */
1756 	range1 = *range;
1757 	range1.length = 1;
1758 	return pnfs_lseg_range_contained(ls_range, &range1);
1759 }
1760 
1761 /*
1762  * lookup range in layout
1763  */
1764 static struct pnfs_layout_segment *
1765 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1766 		struct pnfs_layout_range *range,
1767 		bool strict_iomode)
1768 {
1769 	struct pnfs_layout_segment *lseg, *ret = NULL;
1770 
1771 	dprintk("%s:Begin\n", __func__);
1772 
1773 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1774 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1775 		    pnfs_lseg_range_match(&lseg->pls_range, range,
1776 					  strict_iomode)) {
1777 			ret = pnfs_get_lseg(lseg);
1778 			break;
1779 		}
1780 	}
1781 
1782 	dprintk("%s:Return lseg %p ref %d\n",
1783 		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1784 	return ret;
1785 }
1786 
1787 /*
1788  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1789  * to the MDS or over pNFS
1790  *
1791  * The nfs_inode read_io and write_io fields are cumulative counters reset
1792  * when there are no layout segments. Note that in pnfs_update_layout iomode
1793  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1794  * WRITE request.
1795  *
1796  * A return of true means use MDS I/O.
1797  *
1798  * From rfc 5661:
1799  * If a file's size is smaller than the file size threshold, data accesses
1800  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1801  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1802  * server.  If both file size and I/O size are provided, the client SHOULD
1803  * reach or exceed  both thresholds before sending its read or write
1804  * requests to the data server.
1805  */
1806 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1807 				     struct inode *ino, int iomode)
1808 {
1809 	struct nfs4_threshold *t = ctx->mdsthreshold;
1810 	struct nfs_inode *nfsi = NFS_I(ino);
1811 	loff_t fsize = i_size_read(ino);
1812 	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1813 
1814 	if (t == NULL)
1815 		return ret;
1816 
1817 	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1818 		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1819 
1820 	switch (iomode) {
1821 	case IOMODE_READ:
1822 		if (t->bm & THRESHOLD_RD) {
1823 			dprintk("%s fsize %llu\n", __func__, fsize);
1824 			size_set = true;
1825 			if (fsize < t->rd_sz)
1826 				size = true;
1827 		}
1828 		if (t->bm & THRESHOLD_RD_IO) {
1829 			dprintk("%s nfsi->read_io %llu\n", __func__,
1830 				nfsi->read_io);
1831 			io_set = true;
1832 			if (nfsi->read_io < t->rd_io_sz)
1833 				io = true;
1834 		}
1835 		break;
1836 	case IOMODE_RW:
1837 		if (t->bm & THRESHOLD_WR) {
1838 			dprintk("%s fsize %llu\n", __func__, fsize);
1839 			size_set = true;
1840 			if (fsize < t->wr_sz)
1841 				size = true;
1842 		}
1843 		if (t->bm & THRESHOLD_WR_IO) {
1844 			dprintk("%s nfsi->write_io %llu\n", __func__,
1845 				nfsi->write_io);
1846 			io_set = true;
1847 			if (nfsi->write_io < t->wr_io_sz)
1848 				io = true;
1849 		}
1850 		break;
1851 	}
1852 	if (size_set && io_set) {
1853 		if (size && io)
1854 			ret = true;
1855 	} else if (size || io)
1856 		ret = true;
1857 
1858 	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1859 	return ret;
1860 }
1861 
1862 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1863 {
1864 	/*
1865 	 * send layoutcommit as it can hold up layoutreturn due to lseg
1866 	 * reference
1867 	 */
1868 	pnfs_layoutcommit_inode(lo->plh_inode, false);
1869 	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1870 				   nfs_wait_bit_killable,
1871 				   TASK_KILLABLE);
1872 }
1873 
1874 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1875 {
1876 	atomic_inc(&lo->plh_outstanding);
1877 }
1878 
1879 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1880 {
1881 	if (atomic_dec_and_test(&lo->plh_outstanding))
1882 		wake_up_var(&lo->plh_outstanding);
1883 }
1884 
1885 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1886 {
1887 	unsigned long *bitlock = &lo->plh_flags;
1888 
1889 	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1890 	smp_mb__after_atomic();
1891 	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1892 }
1893 
1894 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1895 				struct nfs_server *server)
1896 {
1897 	if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1898 		struct nfs_client *clp = server->nfs_client;
1899 
1900 		/* The lo must be on the clp list if there is any
1901 		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1902 		 */
1903 		spin_lock(&clp->cl_lock);
1904 		list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1905 		spin_unlock(&clp->cl_lock);
1906 	}
1907 }
1908 
1909 /*
1910  * Layout segment is retreived from the server if not cached.
1911  * The appropriate layout segment is referenced and returned to the caller.
1912  */
1913 struct pnfs_layout_segment *
1914 pnfs_update_layout(struct inode *ino,
1915 		   struct nfs_open_context *ctx,
1916 		   loff_t pos,
1917 		   u64 count,
1918 		   enum pnfs_iomode iomode,
1919 		   bool strict_iomode,
1920 		   gfp_t gfp_flags)
1921 {
1922 	struct pnfs_layout_range arg = {
1923 		.iomode = iomode,
1924 		.offset = pos,
1925 		.length = count,
1926 	};
1927 	unsigned pg_offset;
1928 	struct nfs_server *server = NFS_SERVER(ino);
1929 	struct nfs_client *clp = server->nfs_client;
1930 	struct pnfs_layout_hdr *lo = NULL;
1931 	struct pnfs_layout_segment *lseg = NULL;
1932 	struct nfs4_layoutget *lgp;
1933 	nfs4_stateid stateid;
1934 	long timeout = 0;
1935 	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1936 	bool first;
1937 
1938 	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1939 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1940 				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1941 		goto out;
1942 	}
1943 
1944 	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1945 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1946 				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1947 		goto out;
1948 	}
1949 
1950 lookup_again:
1951 	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1952 	if (IS_ERR(lseg))
1953 		goto out;
1954 	first = false;
1955 	spin_lock(&ino->i_lock);
1956 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1957 	if (lo == NULL) {
1958 		spin_unlock(&ino->i_lock);
1959 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1960 				 PNFS_UPDATE_LAYOUT_NOMEM);
1961 		goto out;
1962 	}
1963 
1964 	/* Do we even need to bother with this? */
1965 	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1966 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1967 				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1968 		dprintk("%s matches recall, use MDS\n", __func__);
1969 		goto out_unlock;
1970 	}
1971 
1972 	/* if LAYOUTGET already failed once we don't try again */
1973 	if (pnfs_layout_io_test_failed(lo, iomode)) {
1974 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1975 				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1976 		goto out_unlock;
1977 	}
1978 
1979 	/*
1980 	 * If the layout segment list is empty, but there are outstanding
1981 	 * layoutget calls, then they might be subject to a layoutrecall.
1982 	 */
1983 	if (list_empty(&lo->plh_segs) &&
1984 	    atomic_read(&lo->plh_outstanding) != 0) {
1985 		spin_unlock(&ino->i_lock);
1986 		lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
1987 					!atomic_read(&lo->plh_outstanding)));
1988 		if (IS_ERR(lseg))
1989 			goto out_put_layout_hdr;
1990 		pnfs_put_layout_hdr(lo);
1991 		goto lookup_again;
1992 	}
1993 
1994 	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1995 	if (lseg) {
1996 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1997 				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1998 		goto out_unlock;
1999 	}
2000 
2001 	/*
2002 	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2003 	 * stateid, or it has been invalidated, then we must use the open
2004 	 * stateid.
2005 	 */
2006 	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2007 		int status;
2008 
2009 		/*
2010 		 * The first layoutget for the file. Need to serialize per
2011 		 * RFC 5661 Errata 3208.
2012 		 */
2013 		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2014 				     &lo->plh_flags)) {
2015 			spin_unlock(&ino->i_lock);
2016 			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2017 						NFS_LAYOUT_FIRST_LAYOUTGET,
2018 						TASK_KILLABLE));
2019 			if (IS_ERR(lseg))
2020 				goto out_put_layout_hdr;
2021 			pnfs_put_layout_hdr(lo);
2022 			dprintk("%s retrying\n", __func__);
2023 			goto lookup_again;
2024 		}
2025 
2026 		spin_unlock(&ino->i_lock);
2027 		first = true;
2028 		status = nfs4_select_rw_stateid(ctx->state,
2029 					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2030 					NULL, &stateid, NULL);
2031 		if (status != 0) {
2032 			lseg = ERR_PTR(status);
2033 			trace_pnfs_update_layout(ino, pos, count,
2034 					iomode, lo, lseg,
2035 					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2036 			nfs4_schedule_stateid_recovery(server, ctx->state);
2037 			pnfs_clear_first_layoutget(lo);
2038 			pnfs_put_layout_hdr(lo);
2039 			goto lookup_again;
2040 		}
2041 		spin_lock(&ino->i_lock);
2042 	} else {
2043 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2044 	}
2045 
2046 	/*
2047 	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
2048 	 * for LAYOUTRETURN even if first is true.
2049 	 */
2050 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2051 		spin_unlock(&ino->i_lock);
2052 		dprintk("%s wait for layoutreturn\n", __func__);
2053 		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2054 		if (!IS_ERR(lseg)) {
2055 			if (first)
2056 				pnfs_clear_first_layoutget(lo);
2057 			pnfs_put_layout_hdr(lo);
2058 			dprintk("%s retrying\n", __func__);
2059 			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2060 					lseg, PNFS_UPDATE_LAYOUT_RETRY);
2061 			goto lookup_again;
2062 		}
2063 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2064 				PNFS_UPDATE_LAYOUT_RETURN);
2065 		goto out_put_layout_hdr;
2066 	}
2067 
2068 	if (pnfs_layoutgets_blocked(lo)) {
2069 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2070 				PNFS_UPDATE_LAYOUT_BLOCKED);
2071 		goto out_unlock;
2072 	}
2073 	nfs_layoutget_begin(lo);
2074 	spin_unlock(&ino->i_lock);
2075 
2076 	_add_to_server_list(lo, server);
2077 
2078 	pg_offset = arg.offset & ~PAGE_MASK;
2079 	if (pg_offset) {
2080 		arg.offset -= pg_offset;
2081 		arg.length += pg_offset;
2082 	}
2083 	if (arg.length != NFS4_MAX_UINT64)
2084 		arg.length = PAGE_ALIGN(arg.length);
2085 
2086 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2087 	if (!lgp) {
2088 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2089 					 PNFS_UPDATE_LAYOUT_NOMEM);
2090 		nfs_layoutget_end(lo);
2091 		goto out_put_layout_hdr;
2092 	}
2093 
2094 	lseg = nfs4_proc_layoutget(lgp, &timeout);
2095 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2096 				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2097 	nfs_layoutget_end(lo);
2098 	if (IS_ERR(lseg)) {
2099 		switch(PTR_ERR(lseg)) {
2100 		case -EBUSY:
2101 			if (time_after(jiffies, giveup))
2102 				lseg = NULL;
2103 			break;
2104 		case -ERECALLCONFLICT:
2105 		case -EAGAIN:
2106 			break;
2107 		default:
2108 			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2109 				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2110 				lseg = NULL;
2111 			}
2112 			goto out_put_layout_hdr;
2113 		}
2114 		if (lseg) {
2115 			if (first)
2116 				pnfs_clear_first_layoutget(lo);
2117 			trace_pnfs_update_layout(ino, pos, count,
2118 				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2119 			pnfs_put_layout_hdr(lo);
2120 			goto lookup_again;
2121 		}
2122 	} else {
2123 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2124 	}
2125 
2126 out_put_layout_hdr:
2127 	if (first)
2128 		pnfs_clear_first_layoutget(lo);
2129 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2130 				 PNFS_UPDATE_LAYOUT_EXIT);
2131 	pnfs_put_layout_hdr(lo);
2132 out:
2133 	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2134 			"(%s, offset: %llu, length: %llu)\n",
2135 			__func__, ino->i_sb->s_id,
2136 			(unsigned long long)NFS_FILEID(ino),
2137 			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2138 			iomode==IOMODE_RW ?  "read/write" : "read-only",
2139 			(unsigned long long)pos,
2140 			(unsigned long long)count);
2141 	return lseg;
2142 out_unlock:
2143 	spin_unlock(&ino->i_lock);
2144 	goto out_put_layout_hdr;
2145 }
2146 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2147 
2148 static bool
2149 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2150 {
2151 	switch (range->iomode) {
2152 	case IOMODE_READ:
2153 	case IOMODE_RW:
2154 		break;
2155 	default:
2156 		return false;
2157 	}
2158 	if (range->offset == NFS4_MAX_UINT64)
2159 		return false;
2160 	if (range->length == 0)
2161 		return false;
2162 	if (range->length != NFS4_MAX_UINT64 &&
2163 	    range->length > NFS4_MAX_UINT64 - range->offset)
2164 		return false;
2165 	return true;
2166 }
2167 
2168 static struct pnfs_layout_hdr *
2169 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2170 {
2171 	struct pnfs_layout_hdr *lo;
2172 
2173 	spin_lock(&ino->i_lock);
2174 	lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2175 	if (!lo)
2176 		goto out_unlock;
2177 	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2178 		goto out_unlock;
2179 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2180 		goto out_unlock;
2181 	if (pnfs_layoutgets_blocked(lo))
2182 		goto out_unlock;
2183 	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2184 		goto out_unlock;
2185 	nfs_layoutget_begin(lo);
2186 	spin_unlock(&ino->i_lock);
2187 	_add_to_server_list(lo, NFS_SERVER(ino));
2188 	return lo;
2189 
2190 out_unlock:
2191 	spin_unlock(&ino->i_lock);
2192 	pnfs_put_layout_hdr(lo);
2193 	return NULL;
2194 }
2195 
2196 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2197 				     struct nfs_open_context *ctx)
2198 {
2199 	struct inode *ino = data->dentry->d_inode;
2200 	struct pnfs_layout_range rng = {
2201 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2202 			  IOMODE_RW: IOMODE_READ,
2203 		.offset = 0,
2204 		.length = NFS4_MAX_UINT64,
2205 	};
2206 	struct nfs4_layoutget *lgp;
2207 	struct pnfs_layout_hdr *lo;
2208 
2209 	/* Heuristic: don't send layoutget if we have cached data */
2210 	if (rng.iomode == IOMODE_READ &&
2211 	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2212 		return;
2213 
2214 	lo = _pnfs_grab_empty_layout(ino, ctx);
2215 	if (!lo)
2216 		return;
2217 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2218 					     &rng, GFP_KERNEL);
2219 	if (!lgp) {
2220 		pnfs_clear_first_layoutget(lo);
2221 		pnfs_put_layout_hdr(lo);
2222 		return;
2223 	}
2224 	data->lgp = lgp;
2225 	data->o_arg.lg_args = &lgp->args;
2226 	data->o_res.lg_res = &lgp->res;
2227 }
2228 
2229 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2230 				     struct nfs_open_context *ctx)
2231 {
2232 	struct pnfs_layout_range rng = {
2233 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2234 			  IOMODE_RW: IOMODE_READ,
2235 		.offset = 0,
2236 		.length = NFS4_MAX_UINT64,
2237 	};
2238 	struct nfs4_layoutget *lgp;
2239 
2240 	lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2241 					     &rng, GFP_KERNEL);
2242 	if (!lgp)
2243 		return;
2244 	data->lgp = lgp;
2245 	data->o_arg.lg_args = &lgp->args;
2246 	data->o_res.lg_res = &lgp->res;
2247 }
2248 
2249 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2250 			 struct nfs_open_context *ctx)
2251 {
2252 	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2253 
2254 	if (!(pnfs_enabled_sb(server) &&
2255 	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2256 		return;
2257 	/* Could check on max_ops, but currently hardcoded high enough */
2258 	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2259 		return;
2260 	if (data->state)
2261 		_lgopen_prepare_attached(data, ctx);
2262 	else
2263 		_lgopen_prepare_floating(data, ctx);
2264 }
2265 
2266 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2267 		       struct nfs_open_context *ctx)
2268 {
2269 	struct pnfs_layout_hdr *lo;
2270 	struct pnfs_layout_segment *lseg;
2271 	struct nfs_server *srv = NFS_SERVER(ino);
2272 	u32 iomode;
2273 
2274 	if (!lgp)
2275 		return;
2276 	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2277 	if (lgp->res.status) {
2278 		switch (lgp->res.status) {
2279 		default:
2280 			break;
2281 		/*
2282 		 * Halt lgopen attempts if the server doesn't recognise
2283 		 * the "current stateid" value, the layout type, or the
2284 		 * layoutget operation as being valid.
2285 		 * Also if it complains about too many ops in the compound
2286 		 * or of the request/reply being too big.
2287 		 */
2288 		case -NFS4ERR_BAD_STATEID:
2289 		case -NFS4ERR_NOTSUPP:
2290 		case -NFS4ERR_REP_TOO_BIG:
2291 		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2292 		case -NFS4ERR_REQ_TOO_BIG:
2293 		case -NFS4ERR_TOO_MANY_OPS:
2294 		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2295 			srv->caps &= ~NFS_CAP_LGOPEN;
2296 		}
2297 		return;
2298 	}
2299 	if (!lgp->args.inode) {
2300 		lo = _pnfs_grab_empty_layout(ino, ctx);
2301 		if (!lo)
2302 			return;
2303 		lgp->args.inode = ino;
2304 	} else
2305 		lo = NFS_I(lgp->args.inode)->layout;
2306 
2307 	lseg = pnfs_layout_process(lgp);
2308 	if (!IS_ERR(lseg)) {
2309 		iomode = lgp->args.range.iomode;
2310 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2311 		pnfs_put_lseg(lseg);
2312 	}
2313 }
2314 
2315 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2316 {
2317 	if (lgp != NULL) {
2318 		struct inode *inode = lgp->args.inode;
2319 		if (inode) {
2320 			struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2321 			pnfs_clear_first_layoutget(lo);
2322 			nfs_layoutget_end(lo);
2323 		}
2324 		pnfs_layoutget_free(lgp);
2325 	}
2326 }
2327 
2328 struct pnfs_layout_segment *
2329 pnfs_layout_process(struct nfs4_layoutget *lgp)
2330 {
2331 	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2332 	struct nfs4_layoutget_res *res = &lgp->res;
2333 	struct pnfs_layout_segment *lseg;
2334 	struct inode *ino = lo->plh_inode;
2335 	LIST_HEAD(free_me);
2336 
2337 	if (!pnfs_sanity_check_layout_range(&res->range))
2338 		return ERR_PTR(-EINVAL);
2339 
2340 	/* Inject layout blob into I/O device driver */
2341 	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2342 	if (IS_ERR_OR_NULL(lseg)) {
2343 		if (!lseg)
2344 			lseg = ERR_PTR(-ENOMEM);
2345 
2346 		dprintk("%s: Could not allocate layout: error %ld\n",
2347 		       __func__, PTR_ERR(lseg));
2348 		return lseg;
2349 	}
2350 
2351 	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2352 
2353 	spin_lock(&ino->i_lock);
2354 	if (pnfs_layoutgets_blocked(lo)) {
2355 		dprintk("%s forget reply due to state\n", __func__);
2356 		goto out_forget;
2357 	}
2358 
2359 	if (!pnfs_layout_is_valid(lo)) {
2360 		/* We have a completely new layout */
2361 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2362 	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2363 		/* existing state ID, make sure the sequence number matches. */
2364 		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2365 			dprintk("%s forget reply due to sequence\n", __func__);
2366 			goto out_forget;
2367 		}
2368 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2369 	} else {
2370 		/*
2371 		 * We got an entirely new state ID.  Mark all segments for the
2372 		 * inode invalid, and retry the layoutget
2373 		 */
2374 		pnfs_mark_layout_stateid_invalid(lo, &free_me);
2375 		goto out_forget;
2376 	}
2377 
2378 	pnfs_get_lseg(lseg);
2379 	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2380 
2381 
2382 	if (res->return_on_close)
2383 		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2384 
2385 	spin_unlock(&ino->i_lock);
2386 	pnfs_free_lseg_list(&free_me);
2387 	return lseg;
2388 
2389 out_forget:
2390 	spin_unlock(&ino->i_lock);
2391 	lseg->pls_layout = lo;
2392 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2393 	return ERR_PTR(-EAGAIN);
2394 }
2395 
2396 static int
2397 mark_lseg_invalid_or_return(struct pnfs_layout_segment *lseg,
2398 		struct list_head *tmp_list)
2399 {
2400 	if (!mark_lseg_invalid(lseg, tmp_list))
2401 		return 0;
2402 	pnfs_cache_lseg_for_layoutreturn(lseg->pls_layout, lseg);
2403 	return 1;
2404 }
2405 
2406 /**
2407  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2408  * @lo: pointer to layout header
2409  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2410  * @return_range: describe layout segment ranges to be returned
2411  * @seq: stateid seqid to match
2412  *
2413  * This function is mainly intended for use by layoutrecall. It attempts
2414  * to free the layout segment immediately, or else to mark it for return
2415  * as soon as its reference count drops to zero.
2416  *
2417  * Returns
2418  * - 0: a layoutreturn needs to be scheduled.
2419  * - EBUSY: there are layout segment that are still in use.
2420  * - ENOENT: there are no layout segments that need to be returned.
2421  */
2422 int
2423 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2424 				struct list_head *tmp_list,
2425 				const struct pnfs_layout_range *return_range,
2426 				u32 seq)
2427 {
2428 	struct pnfs_layout_segment *lseg, *next;
2429 	int remaining = 0;
2430 
2431 	dprintk("%s:Begin lo %p\n", __func__, lo);
2432 
2433 	assert_spin_locked(&lo->plh_inode->i_lock);
2434 
2435 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2436 		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2437 			dprintk("%s: marking lseg %p iomode %d "
2438 				"offset %llu length %llu\n", __func__,
2439 				lseg, lseg->pls_range.iomode,
2440 				lseg->pls_range.offset,
2441 				lseg->pls_range.length);
2442 			if (mark_lseg_invalid_or_return(lseg, tmp_list))
2443 				continue;
2444 			remaining++;
2445 			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2446 		}
2447 
2448 	if (remaining) {
2449 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2450 		return -EBUSY;
2451 	}
2452 
2453 	if (!list_empty(&lo->plh_return_segs)) {
2454 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2455 		return 0;
2456 	}
2457 
2458 	return -ENOENT;
2459 }
2460 
2461 static void
2462 pnfs_mark_layout_for_return(struct inode *inode,
2463 			    const struct pnfs_layout_range *range)
2464 {
2465 	struct pnfs_layout_hdr *lo;
2466 	bool return_now = false;
2467 
2468 	spin_lock(&inode->i_lock);
2469 	lo = NFS_I(inode)->layout;
2470 	if (!pnfs_layout_is_valid(lo)) {
2471 		spin_unlock(&inode->i_lock);
2472 		return;
2473 	}
2474 	pnfs_set_plh_return_info(lo, range->iomode, 0);
2475 	/*
2476 	 * mark all matching lsegs so that we are sure to have no live
2477 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2478 	 * for how it works.
2479 	 */
2480 	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2481 		const struct cred *cred;
2482 		nfs4_stateid stateid;
2483 		enum pnfs_iomode iomode;
2484 
2485 		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2486 		spin_unlock(&inode->i_lock);
2487 		if (return_now)
2488 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2489 	} else {
2490 		spin_unlock(&inode->i_lock);
2491 		nfs_commit_inode(inode, 0);
2492 	}
2493 }
2494 
2495 void pnfs_error_mark_layout_for_return(struct inode *inode,
2496 				       struct pnfs_layout_segment *lseg)
2497 {
2498 	struct pnfs_layout_range range = {
2499 		.iomode = lseg->pls_range.iomode,
2500 		.offset = 0,
2501 		.length = NFS4_MAX_UINT64,
2502 	};
2503 
2504 	pnfs_mark_layout_for_return(inode, &range);
2505 }
2506 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2507 
2508 static bool
2509 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2510 {
2511 	return pnfs_layout_is_valid(lo) &&
2512 		!test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2513 		!test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2514 }
2515 
2516 static struct pnfs_layout_segment *
2517 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2518 		     const struct pnfs_layout_range *range,
2519 		     enum pnfs_iomode iomode)
2520 {
2521 	struct pnfs_layout_segment *lseg;
2522 
2523 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2524 		if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2525 			continue;
2526 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2527 			continue;
2528 		if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2529 			continue;
2530 		if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2531 			return lseg;
2532 	}
2533 	return NULL;
2534 }
2535 
2536 /* Find open file states whose mode matches that of the range */
2537 static bool
2538 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2539 				 const struct pnfs_layout_range *range)
2540 {
2541 	struct list_head *head;
2542 	struct nfs_open_context *ctx;
2543 	fmode_t mode = 0;
2544 
2545 	if (!pnfs_layout_can_be_returned(lo) ||
2546 	    !pnfs_find_first_lseg(lo, range, range->iomode))
2547 		return false;
2548 
2549 	head = &NFS_I(lo->plh_inode)->open_files;
2550 	list_for_each_entry_rcu(ctx, head, list) {
2551 		if (ctx->state)
2552 			mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2553 	}
2554 
2555 	switch (range->iomode) {
2556 	default:
2557 		break;
2558 	case IOMODE_READ:
2559 		mode &= ~FMODE_WRITE;
2560 		break;
2561 	case IOMODE_RW:
2562 		if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2563 			mode &= ~FMODE_READ;
2564 	}
2565 	return mode == 0;
2566 }
2567 
2568 static int
2569 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2570 {
2571 	const struct pnfs_layout_range *range = data;
2572 	struct pnfs_layout_hdr *lo;
2573 	struct inode *inode;
2574 restart:
2575 	rcu_read_lock();
2576 	list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2577 		if (!pnfs_layout_can_be_returned(lo) ||
2578 		    test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2579 			continue;
2580 		inode = lo->plh_inode;
2581 		spin_lock(&inode->i_lock);
2582 		if (!pnfs_should_return_unused_layout(lo, range)) {
2583 			spin_unlock(&inode->i_lock);
2584 			continue;
2585 		}
2586 		spin_unlock(&inode->i_lock);
2587 		inode = pnfs_grab_inode_layout_hdr(lo);
2588 		if (!inode)
2589 			continue;
2590 		rcu_read_unlock();
2591 		pnfs_mark_layout_for_return(inode, range);
2592 		iput(inode);
2593 		cond_resched();
2594 		goto restart;
2595 	}
2596 	rcu_read_unlock();
2597 	return 0;
2598 }
2599 
2600 void
2601 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2602 				 enum pnfs_iomode iomode)
2603 {
2604 	struct pnfs_layout_range range = {
2605 		.iomode = iomode,
2606 		.offset = 0,
2607 		.length = NFS4_MAX_UINT64,
2608 	};
2609 
2610 	nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2611 			&range);
2612 }
2613 
2614 void
2615 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2616 {
2617 	if (pgio->pg_lseg == NULL ||
2618 	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2619 		return;
2620 	pnfs_put_lseg(pgio->pg_lseg);
2621 	pgio->pg_lseg = NULL;
2622 }
2623 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2624 
2625 /*
2626  * Check for any intersection between the request and the pgio->pg_lseg,
2627  * and if none, put this pgio->pg_lseg away.
2628  */
2629 void
2630 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2631 {
2632 	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2633 		pnfs_put_lseg(pgio->pg_lseg);
2634 		pgio->pg_lseg = NULL;
2635 	}
2636 }
2637 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2638 
2639 void
2640 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2641 {
2642 	u64 rd_size = req->wb_bytes;
2643 
2644 	pnfs_generic_pg_check_layout(pgio);
2645 	pnfs_generic_pg_check_range(pgio, req);
2646 	if (pgio->pg_lseg == NULL) {
2647 		if (pgio->pg_dreq == NULL)
2648 			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2649 		else
2650 			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2651 
2652 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2653 						   nfs_req_openctx(req),
2654 						   req_offset(req),
2655 						   rd_size,
2656 						   IOMODE_READ,
2657 						   false,
2658 						   GFP_KERNEL);
2659 		if (IS_ERR(pgio->pg_lseg)) {
2660 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2661 			pgio->pg_lseg = NULL;
2662 			return;
2663 		}
2664 	}
2665 	/* If no lseg, fall back to read through mds */
2666 	if (pgio->pg_lseg == NULL)
2667 		nfs_pageio_reset_read_mds(pgio);
2668 
2669 }
2670 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2671 
2672 void
2673 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2674 			   struct nfs_page *req, u64 wb_size)
2675 {
2676 	pnfs_generic_pg_check_layout(pgio);
2677 	pnfs_generic_pg_check_range(pgio, req);
2678 	if (pgio->pg_lseg == NULL) {
2679 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2680 						   nfs_req_openctx(req),
2681 						   req_offset(req),
2682 						   wb_size,
2683 						   IOMODE_RW,
2684 						   false,
2685 						   GFP_KERNEL);
2686 		if (IS_ERR(pgio->pg_lseg)) {
2687 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2688 			pgio->pg_lseg = NULL;
2689 			return;
2690 		}
2691 	}
2692 	/* If no lseg, fall back to write through mds */
2693 	if (pgio->pg_lseg == NULL)
2694 		nfs_pageio_reset_write_mds(pgio);
2695 }
2696 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2697 
2698 void
2699 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2700 {
2701 	if (desc->pg_lseg) {
2702 		pnfs_put_lseg(desc->pg_lseg);
2703 		desc->pg_lseg = NULL;
2704 	}
2705 }
2706 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2707 
2708 /*
2709  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2710  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2711  */
2712 size_t
2713 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2714 		     struct nfs_page *prev, struct nfs_page *req)
2715 {
2716 	unsigned int size;
2717 	u64 seg_end, req_start, seg_left;
2718 
2719 	size = nfs_generic_pg_test(pgio, prev, req);
2720 	if (!size)
2721 		return 0;
2722 
2723 	/*
2724 	 * 'size' contains the number of bytes left in the current page (up
2725 	 * to the original size asked for in @req->wb_bytes).
2726 	 *
2727 	 * Calculate how many bytes are left in the layout segment
2728 	 * and if there are less bytes than 'size', return that instead.
2729 	 *
2730 	 * Please also note that 'end_offset' is actually the offset of the
2731 	 * first byte that lies outside the pnfs_layout_range. FIXME?
2732 	 *
2733 	 */
2734 	if (pgio->pg_lseg) {
2735 		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2736 				     pgio->pg_lseg->pls_range.length);
2737 		req_start = req_offset(req);
2738 
2739 		/* start of request is past the last byte of this segment */
2740 		if (req_start >= seg_end)
2741 			return 0;
2742 
2743 		/* adjust 'size' iff there are fewer bytes left in the
2744 		 * segment than what nfs_generic_pg_test returned */
2745 		seg_left = seg_end - req_start;
2746 		if (seg_left < size)
2747 			size = (unsigned int)seg_left;
2748 	}
2749 
2750 	return size;
2751 }
2752 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2753 
2754 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2755 {
2756 	struct nfs_pageio_descriptor pgio;
2757 
2758 	/* Resend all requests through the MDS */
2759 	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2760 			      hdr->completion_ops);
2761 	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2762 	return nfs_pageio_resend(&pgio, hdr);
2763 }
2764 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2765 
2766 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2767 {
2768 
2769 	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2770 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2771 	    PNFS_LAYOUTRET_ON_ERROR) {
2772 		pnfs_return_layout(hdr->inode);
2773 	}
2774 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2775 		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2776 }
2777 
2778 /*
2779  * Called by non rpc-based layout drivers
2780  */
2781 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2782 {
2783 	if (likely(!hdr->pnfs_error)) {
2784 		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2785 				hdr->mds_offset + hdr->res.count);
2786 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2787 	}
2788 	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2789 	if (unlikely(hdr->pnfs_error))
2790 		pnfs_ld_handle_write_error(hdr);
2791 	hdr->mds_ops->rpc_release(hdr);
2792 }
2793 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2794 
2795 static void
2796 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2797 		struct nfs_pgio_header *hdr)
2798 {
2799 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2800 
2801 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2802 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2803 		nfs_pageio_reset_write_mds(desc);
2804 		mirror->pg_recoalesce = 1;
2805 	}
2806 	hdr->completion_ops->completion(hdr);
2807 }
2808 
2809 static enum pnfs_try_status
2810 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2811 			const struct rpc_call_ops *call_ops,
2812 			struct pnfs_layout_segment *lseg,
2813 			int how)
2814 {
2815 	struct inode *inode = hdr->inode;
2816 	enum pnfs_try_status trypnfs;
2817 	struct nfs_server *nfss = NFS_SERVER(inode);
2818 
2819 	hdr->mds_ops = call_ops;
2820 
2821 	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2822 		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2823 	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2824 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2825 		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2826 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2827 	return trypnfs;
2828 }
2829 
2830 static void
2831 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2832 	      struct nfs_pgio_header *hdr, int how)
2833 {
2834 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2835 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2836 	enum pnfs_try_status trypnfs;
2837 
2838 	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2839 	switch (trypnfs) {
2840 	case PNFS_NOT_ATTEMPTED:
2841 		pnfs_write_through_mds(desc, hdr);
2842 	case PNFS_ATTEMPTED:
2843 		break;
2844 	case PNFS_TRY_AGAIN:
2845 		/* cleanup hdr and prepare to redo pnfs */
2846 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2847 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2848 			list_splice_init(&hdr->pages, &mirror->pg_list);
2849 			mirror->pg_recoalesce = 1;
2850 		}
2851 		hdr->mds_ops->rpc_release(hdr);
2852 	}
2853 }
2854 
2855 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2856 {
2857 	pnfs_put_lseg(hdr->lseg);
2858 	nfs_pgio_header_free(hdr);
2859 }
2860 
2861 int
2862 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2863 {
2864 	struct nfs_pgio_header *hdr;
2865 	int ret;
2866 
2867 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2868 	if (!hdr) {
2869 		desc->pg_error = -ENOMEM;
2870 		return desc->pg_error;
2871 	}
2872 	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2873 
2874 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2875 	ret = nfs_generic_pgio(desc, hdr);
2876 	if (!ret)
2877 		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2878 
2879 	return ret;
2880 }
2881 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2882 
2883 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2884 {
2885 	struct nfs_pageio_descriptor pgio;
2886 
2887 	/* Resend all requests through the MDS */
2888 	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2889 	return nfs_pageio_resend(&pgio, hdr);
2890 }
2891 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2892 
2893 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2894 {
2895 	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2896 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2897 	    PNFS_LAYOUTRET_ON_ERROR) {
2898 		pnfs_return_layout(hdr->inode);
2899 	}
2900 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2901 		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2902 }
2903 
2904 /*
2905  * Called by non rpc-based layout drivers
2906  */
2907 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2908 {
2909 	if (likely(!hdr->pnfs_error))
2910 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2911 	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2912 	if (unlikely(hdr->pnfs_error))
2913 		pnfs_ld_handle_read_error(hdr);
2914 	hdr->mds_ops->rpc_release(hdr);
2915 }
2916 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2917 
2918 static void
2919 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2920 		struct nfs_pgio_header *hdr)
2921 {
2922 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2923 
2924 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2925 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2926 		nfs_pageio_reset_read_mds(desc);
2927 		mirror->pg_recoalesce = 1;
2928 	}
2929 	hdr->completion_ops->completion(hdr);
2930 }
2931 
2932 /*
2933  * Call the appropriate parallel I/O subsystem read function.
2934  */
2935 static enum pnfs_try_status
2936 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2937 		       const struct rpc_call_ops *call_ops,
2938 		       struct pnfs_layout_segment *lseg)
2939 {
2940 	struct inode *inode = hdr->inode;
2941 	struct nfs_server *nfss = NFS_SERVER(inode);
2942 	enum pnfs_try_status trypnfs;
2943 
2944 	hdr->mds_ops = call_ops;
2945 
2946 	dprintk("%s: Reading ino:%lu %u@%llu\n",
2947 		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2948 
2949 	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2950 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2951 		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2952 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2953 	return trypnfs;
2954 }
2955 
2956 /* Resend all requests through pnfs. */
2957 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2958 {
2959 	struct nfs_pageio_descriptor pgio;
2960 
2961 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2962 		/* Prevent deadlocks with layoutreturn! */
2963 		pnfs_put_lseg(hdr->lseg);
2964 		hdr->lseg = NULL;
2965 
2966 		nfs_pageio_init_read(&pgio, hdr->inode, false,
2967 					hdr->completion_ops);
2968 		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2969 	}
2970 }
2971 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2972 
2973 static void
2974 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2975 {
2976 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2977 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2978 	enum pnfs_try_status trypnfs;
2979 
2980 	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2981 	switch (trypnfs) {
2982 	case PNFS_NOT_ATTEMPTED:
2983 		pnfs_read_through_mds(desc, hdr);
2984 	case PNFS_ATTEMPTED:
2985 		break;
2986 	case PNFS_TRY_AGAIN:
2987 		/* cleanup hdr and prepare to redo pnfs */
2988 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2989 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2990 			list_splice_init(&hdr->pages, &mirror->pg_list);
2991 			mirror->pg_recoalesce = 1;
2992 		}
2993 		hdr->mds_ops->rpc_release(hdr);
2994 	}
2995 }
2996 
2997 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2998 {
2999 	pnfs_put_lseg(hdr->lseg);
3000 	nfs_pgio_header_free(hdr);
3001 }
3002 
3003 int
3004 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3005 {
3006 	struct nfs_pgio_header *hdr;
3007 	int ret;
3008 
3009 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3010 	if (!hdr) {
3011 		desc->pg_error = -ENOMEM;
3012 		return desc->pg_error;
3013 	}
3014 	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3015 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3016 	ret = nfs_generic_pgio(desc, hdr);
3017 	if (!ret)
3018 		pnfs_do_read(desc, hdr);
3019 	return ret;
3020 }
3021 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3022 
3023 static void pnfs_clear_layoutcommitting(struct inode *inode)
3024 {
3025 	unsigned long *bitlock = &NFS_I(inode)->flags;
3026 
3027 	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3028 	smp_mb__after_atomic();
3029 	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3030 }
3031 
3032 /*
3033  * There can be multiple RW segments.
3034  */
3035 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3036 {
3037 	struct pnfs_layout_segment *lseg;
3038 
3039 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3040 		if (lseg->pls_range.iomode == IOMODE_RW &&
3041 		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3042 			list_add(&lseg->pls_lc_list, listp);
3043 	}
3044 }
3045 
3046 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3047 {
3048 	struct pnfs_layout_segment *lseg, *tmp;
3049 
3050 	/* Matched by references in pnfs_set_layoutcommit */
3051 	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3052 		list_del_init(&lseg->pls_lc_list);
3053 		pnfs_put_lseg(lseg);
3054 	}
3055 
3056 	pnfs_clear_layoutcommitting(inode);
3057 }
3058 
3059 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3060 {
3061 	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3062 }
3063 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3064 
3065 void
3066 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3067 		loff_t end_pos)
3068 {
3069 	struct nfs_inode *nfsi = NFS_I(inode);
3070 	bool mark_as_dirty = false;
3071 
3072 	spin_lock(&inode->i_lock);
3073 	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3074 		nfsi->layout->plh_lwb = end_pos;
3075 		mark_as_dirty = true;
3076 		dprintk("%s: Set layoutcommit for inode %lu ",
3077 			__func__, inode->i_ino);
3078 	} else if (end_pos > nfsi->layout->plh_lwb)
3079 		nfsi->layout->plh_lwb = end_pos;
3080 	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3081 		/* references matched in nfs4_layoutcommit_release */
3082 		pnfs_get_lseg(lseg);
3083 	}
3084 	spin_unlock(&inode->i_lock);
3085 	dprintk("%s: lseg %p end_pos %llu\n",
3086 		__func__, lseg, nfsi->layout->plh_lwb);
3087 
3088 	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3089 	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3090 	if (mark_as_dirty)
3091 		mark_inode_dirty_sync(inode);
3092 }
3093 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3094 
3095 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3096 {
3097 	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3098 
3099 	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3100 		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3101 	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3102 }
3103 
3104 /*
3105  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3106  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3107  * data to disk to allow the server to recover the data if it crashes.
3108  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3109  * is off, and a COMMIT is sent to a data server, or
3110  * if WRITEs to a data server return NFS_DATA_SYNC.
3111  */
3112 int
3113 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3114 {
3115 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3116 	struct nfs4_layoutcommit_data *data;
3117 	struct nfs_inode *nfsi = NFS_I(inode);
3118 	loff_t end_pos;
3119 	int status;
3120 
3121 	if (!pnfs_layoutcommit_outstanding(inode))
3122 		return 0;
3123 
3124 	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3125 
3126 	status = -EAGAIN;
3127 	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3128 		if (!sync)
3129 			goto out;
3130 		status = wait_on_bit_lock_action(&nfsi->flags,
3131 				NFS_INO_LAYOUTCOMMITTING,
3132 				nfs_wait_bit_killable,
3133 				TASK_KILLABLE);
3134 		if (status)
3135 			goto out;
3136 	}
3137 
3138 	status = -ENOMEM;
3139 	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3140 	data = kzalloc(sizeof(*data), GFP_NOFS);
3141 	if (!data)
3142 		goto clear_layoutcommitting;
3143 
3144 	status = 0;
3145 	spin_lock(&inode->i_lock);
3146 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3147 		goto out_unlock;
3148 
3149 	INIT_LIST_HEAD(&data->lseg_list);
3150 	pnfs_list_write_lseg(inode, &data->lseg_list);
3151 
3152 	end_pos = nfsi->layout->plh_lwb;
3153 
3154 	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3155 	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3156 	spin_unlock(&inode->i_lock);
3157 
3158 	data->args.inode = inode;
3159 	nfs_fattr_init(&data->fattr);
3160 	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3161 	data->res.fattr = &data->fattr;
3162 	if (end_pos != 0)
3163 		data->args.lastbytewritten = end_pos - 1;
3164 	else
3165 		data->args.lastbytewritten = U64_MAX;
3166 	data->res.server = NFS_SERVER(inode);
3167 
3168 	if (ld->prepare_layoutcommit) {
3169 		status = ld->prepare_layoutcommit(&data->args);
3170 		if (status) {
3171 			put_cred(data->cred);
3172 			spin_lock(&inode->i_lock);
3173 			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3174 			if (end_pos > nfsi->layout->plh_lwb)
3175 				nfsi->layout->plh_lwb = end_pos;
3176 			goto out_unlock;
3177 		}
3178 	}
3179 
3180 
3181 	status = nfs4_proc_layoutcommit(data, sync);
3182 out:
3183 	if (status)
3184 		mark_inode_dirty_sync(inode);
3185 	dprintk("<-- %s status %d\n", __func__, status);
3186 	return status;
3187 out_unlock:
3188 	spin_unlock(&inode->i_lock);
3189 	kfree(data);
3190 clear_layoutcommitting:
3191 	pnfs_clear_layoutcommitting(inode);
3192 	goto out;
3193 }
3194 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3195 
3196 int
3197 pnfs_generic_sync(struct inode *inode, bool datasync)
3198 {
3199 	return pnfs_layoutcommit_inode(inode, true);
3200 }
3201 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3202 
3203 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3204 {
3205 	struct nfs4_threshold *thp;
3206 
3207 	thp = kzalloc(sizeof(*thp), GFP_NOFS);
3208 	if (!thp) {
3209 		dprintk("%s mdsthreshold allocation failed\n", __func__);
3210 		return NULL;
3211 	}
3212 	return thp;
3213 }
3214 
3215 #if IS_ENABLED(CONFIG_NFS_V4_2)
3216 int
3217 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3218 {
3219 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3220 	struct nfs_server *server = NFS_SERVER(inode);
3221 	struct nfs_inode *nfsi = NFS_I(inode);
3222 	struct nfs42_layoutstat_data *data;
3223 	struct pnfs_layout_hdr *hdr;
3224 	int status = 0;
3225 
3226 	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3227 		goto out;
3228 
3229 	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3230 		goto out;
3231 
3232 	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3233 		goto out;
3234 
3235 	spin_lock(&inode->i_lock);
3236 	if (!NFS_I(inode)->layout) {
3237 		spin_unlock(&inode->i_lock);
3238 		goto out_clear_layoutstats;
3239 	}
3240 	hdr = NFS_I(inode)->layout;
3241 	pnfs_get_layout_hdr(hdr);
3242 	spin_unlock(&inode->i_lock);
3243 
3244 	data = kzalloc(sizeof(*data), gfp_flags);
3245 	if (!data) {
3246 		status = -ENOMEM;
3247 		goto out_put;
3248 	}
3249 
3250 	data->args.fh = NFS_FH(inode);
3251 	data->args.inode = inode;
3252 	status = ld->prepare_layoutstats(&data->args);
3253 	if (status)
3254 		goto out_free;
3255 
3256 	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3257 
3258 out:
3259 	dprintk("%s returns %d\n", __func__, status);
3260 	return status;
3261 
3262 out_free:
3263 	kfree(data);
3264 out_put:
3265 	pnfs_put_layout_hdr(hdr);
3266 out_clear_layoutstats:
3267 	smp_mb__before_atomic();
3268 	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3269 	smp_mb__after_atomic();
3270 	goto out;
3271 }
3272 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3273 #endif
3274 
3275 unsigned int layoutstats_timer;
3276 module_param(layoutstats_timer, uint, 0644);
3277 EXPORT_SYMBOL_GPL(layoutstats_timer);
3278