xref: /openbmc/linux/fs/nfs/pnfs.c (revision 0936cdfb)
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_wait_layoutreturn;
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_wait_layoutreturn:
1343 	wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1344 out_put_layout_hdr:
1345 	pnfs_free_lseg_list(&tmp_list);
1346 	pnfs_put_layout_hdr(lo);
1347 out:
1348 	dprintk("<-- %s status: %d\n", __func__, status);
1349 	return status;
1350 }
1351 
1352 int
1353 pnfs_commit_and_return_layout(struct inode *inode)
1354 {
1355 	struct pnfs_layout_hdr *lo;
1356 	int ret;
1357 
1358 	spin_lock(&inode->i_lock);
1359 	lo = NFS_I(inode)->layout;
1360 	if (lo == NULL) {
1361 		spin_unlock(&inode->i_lock);
1362 		return 0;
1363 	}
1364 	pnfs_get_layout_hdr(lo);
1365 	/* Block new layoutgets and read/write to ds */
1366 	lo->plh_block_lgets++;
1367 	spin_unlock(&inode->i_lock);
1368 	filemap_fdatawait(inode->i_mapping);
1369 	ret = pnfs_layoutcommit_inode(inode, true);
1370 	if (ret == 0)
1371 		ret = _pnfs_return_layout(inode);
1372 	spin_lock(&inode->i_lock);
1373 	lo->plh_block_lgets--;
1374 	spin_unlock(&inode->i_lock);
1375 	pnfs_put_layout_hdr(lo);
1376 	return ret;
1377 }
1378 
1379 bool pnfs_roc(struct inode *ino,
1380 		struct nfs4_layoutreturn_args *args,
1381 		struct nfs4_layoutreturn_res *res,
1382 		const struct cred *cred)
1383 {
1384 	struct nfs_inode *nfsi = NFS_I(ino);
1385 	struct nfs_open_context *ctx;
1386 	struct nfs4_state *state;
1387 	struct pnfs_layout_hdr *lo;
1388 	struct pnfs_layout_segment *lseg, *next;
1389 	const struct cred *lc_cred;
1390 	nfs4_stateid stateid;
1391 	enum pnfs_iomode iomode = 0;
1392 	bool layoutreturn = false, roc = false;
1393 	bool skip_read = false;
1394 
1395 	if (!nfs_have_layout(ino))
1396 		return false;
1397 retry:
1398 	rcu_read_lock();
1399 	spin_lock(&ino->i_lock);
1400 	lo = nfsi->layout;
1401 	if (!lo || !pnfs_layout_is_valid(lo) ||
1402 	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1403 		lo = NULL;
1404 		goto out_noroc;
1405 	}
1406 	pnfs_get_layout_hdr(lo);
1407 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1408 		spin_unlock(&ino->i_lock);
1409 		rcu_read_unlock();
1410 		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1411 				TASK_UNINTERRUPTIBLE);
1412 		pnfs_put_layout_hdr(lo);
1413 		goto retry;
1414 	}
1415 
1416 	/* no roc if we hold a delegation */
1417 	if (nfs4_check_delegation(ino, FMODE_READ)) {
1418 		if (nfs4_check_delegation(ino, FMODE_WRITE))
1419 			goto out_noroc;
1420 		skip_read = true;
1421 	}
1422 
1423 	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1424 		state = ctx->state;
1425 		if (state == NULL)
1426 			continue;
1427 		/* Don't return layout if there is open file state */
1428 		if (state->state & FMODE_WRITE)
1429 			goto out_noroc;
1430 		if (state->state & FMODE_READ)
1431 			skip_read = true;
1432 	}
1433 
1434 
1435 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1436 		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1437 			continue;
1438 		/* If we are sending layoutreturn, invalidate all valid lsegs */
1439 		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1440 			continue;
1441 		/*
1442 		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1443 		 * doesn't invalidate the layout for us.
1444 		 */
1445 		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1446 		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1447 			continue;
1448 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1449 	}
1450 
1451 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1452 		goto out_noroc;
1453 
1454 	/* ROC in two conditions:
1455 	 * 1. there are ROC lsegs
1456 	 * 2. we don't send layoutreturn
1457 	 */
1458 	/* lo ref dropped in pnfs_roc_release() */
1459 	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1460 	/* If the creds don't match, we can't compound the layoutreturn */
1461 	if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1462 		goto out_noroc;
1463 
1464 	roc = layoutreturn;
1465 	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1466 	res->lrs_present = 0;
1467 	layoutreturn = false;
1468 	put_cred(lc_cred);
1469 
1470 out_noroc:
1471 	spin_unlock(&ino->i_lock);
1472 	rcu_read_unlock();
1473 	pnfs_layoutcommit_inode(ino, true);
1474 	if (roc) {
1475 		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1476 		if (ld->prepare_layoutreturn)
1477 			ld->prepare_layoutreturn(args);
1478 		pnfs_put_layout_hdr(lo);
1479 		return true;
1480 	}
1481 	if (layoutreturn)
1482 		pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1483 	pnfs_put_layout_hdr(lo);
1484 	return false;
1485 }
1486 
1487 int pnfs_roc_done(struct rpc_task *task, struct inode *inode,
1488 		struct nfs4_layoutreturn_args **argpp,
1489 		struct nfs4_layoutreturn_res **respp,
1490 		int *ret)
1491 {
1492 	struct nfs4_layoutreturn_args *arg = *argpp;
1493 	int retval = -EAGAIN;
1494 
1495 	if (!arg)
1496 		return 0;
1497 	/* Handle Layoutreturn errors */
1498 	switch (*ret) {
1499 	case 0:
1500 		retval = 0;
1501 		break;
1502 	case -NFS4ERR_NOMATCHING_LAYOUT:
1503 		/* Was there an RPC level error? If not, retry */
1504 		if (task->tk_rpc_status == 0)
1505 			break;
1506 		/* If the call was not sent, let caller handle it */
1507 		if (!RPC_WAS_SENT(task))
1508 			return 0;
1509 		/*
1510 		 * Otherwise, assume the call succeeded and
1511 		 * that we need to release the layout
1512 		 */
1513 		*ret = 0;
1514 		(*respp)->lrs_present = 0;
1515 		retval = 0;
1516 		break;
1517 	case -NFS4ERR_DELAY:
1518 		/* Let the caller handle the retry */
1519 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1520 		return 0;
1521 	case -NFS4ERR_OLD_STATEID:
1522 		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1523 					&arg->range, inode))
1524 			break;
1525 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1526 		return -EAGAIN;
1527 	}
1528 	*argpp = NULL;
1529 	*respp = NULL;
1530 	return retval;
1531 }
1532 
1533 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1534 		struct nfs4_layoutreturn_res *res,
1535 		int ret)
1536 {
1537 	struct pnfs_layout_hdr *lo = args->layout;
1538 	const nfs4_stateid *arg_stateid = NULL;
1539 	const nfs4_stateid *res_stateid = NULL;
1540 	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1541 
1542 	switch (ret) {
1543 	case -NFS4ERR_NOMATCHING_LAYOUT:
1544 		break;
1545 	case 0:
1546 		if (res->lrs_present)
1547 			res_stateid = &res->stateid;
1548 		/* Fallthrough */
1549 	default:
1550 		arg_stateid = &args->stateid;
1551 	}
1552 	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1553 			res_stateid);
1554 	if (ld_private && ld_private->ops && ld_private->ops->free)
1555 		ld_private->ops->free(ld_private);
1556 	pnfs_put_layout_hdr(lo);
1557 	trace_nfs4_layoutreturn_on_close(args->inode, 0);
1558 }
1559 
1560 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1561 {
1562 	struct nfs_inode *nfsi = NFS_I(ino);
1563         struct pnfs_layout_hdr *lo;
1564         bool sleep = false;
1565 
1566 	/* we might not have grabbed lo reference. so need to check under
1567 	 * i_lock */
1568         spin_lock(&ino->i_lock);
1569         lo = nfsi->layout;
1570         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1571                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1572                 sleep = true;
1573 	}
1574         spin_unlock(&ino->i_lock);
1575         return sleep;
1576 }
1577 
1578 /*
1579  * Compare two layout segments for sorting into layout cache.
1580  * We want to preferentially return RW over RO layouts, so ensure those
1581  * are seen first.
1582  */
1583 static s64
1584 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1585 	   const struct pnfs_layout_range *l2)
1586 {
1587 	s64 d;
1588 
1589 	/* high offset > low offset */
1590 	d = l1->offset - l2->offset;
1591 	if (d)
1592 		return d;
1593 
1594 	/* short length > long length */
1595 	d = l2->length - l1->length;
1596 	if (d)
1597 		return d;
1598 
1599 	/* read > read/write */
1600 	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1601 }
1602 
1603 static bool
1604 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1605 		const struct pnfs_layout_range *l2)
1606 {
1607 	return pnfs_lseg_range_cmp(l1, l2) > 0;
1608 }
1609 
1610 static bool
1611 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1612 		struct pnfs_layout_segment *old)
1613 {
1614 	return false;
1615 }
1616 
1617 void
1618 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1619 		   struct pnfs_layout_segment *lseg,
1620 		   bool (*is_after)(const struct pnfs_layout_range *,
1621 			   const struct pnfs_layout_range *),
1622 		   bool (*do_merge)(struct pnfs_layout_segment *,
1623 			   struct pnfs_layout_segment *),
1624 		   struct list_head *free_me)
1625 {
1626 	struct pnfs_layout_segment *lp, *tmp;
1627 
1628 	dprintk("%s:Begin\n", __func__);
1629 
1630 	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1631 		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1632 			continue;
1633 		if (do_merge(lseg, lp)) {
1634 			mark_lseg_invalid(lp, free_me);
1635 			continue;
1636 		}
1637 		if (is_after(&lseg->pls_range, &lp->pls_range))
1638 			continue;
1639 		list_add_tail(&lseg->pls_list, &lp->pls_list);
1640 		dprintk("%s: inserted lseg %p "
1641 			"iomode %d offset %llu length %llu before "
1642 			"lp %p iomode %d offset %llu length %llu\n",
1643 			__func__, lseg, lseg->pls_range.iomode,
1644 			lseg->pls_range.offset, lseg->pls_range.length,
1645 			lp, lp->pls_range.iomode, lp->pls_range.offset,
1646 			lp->pls_range.length);
1647 		goto out;
1648 	}
1649 	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1650 	dprintk("%s: inserted lseg %p "
1651 		"iomode %d offset %llu length %llu at tail\n",
1652 		__func__, lseg, lseg->pls_range.iomode,
1653 		lseg->pls_range.offset, lseg->pls_range.length);
1654 out:
1655 	pnfs_get_layout_hdr(lo);
1656 
1657 	dprintk("%s:Return\n", __func__);
1658 }
1659 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1660 
1661 static void
1662 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1663 		   struct pnfs_layout_segment *lseg,
1664 		   struct list_head *free_me)
1665 {
1666 	struct inode *inode = lo->plh_inode;
1667 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1668 
1669 	if (ld->add_lseg != NULL)
1670 		ld->add_lseg(lo, lseg, free_me);
1671 	else
1672 		pnfs_generic_layout_insert_lseg(lo, lseg,
1673 				pnfs_lseg_range_is_after,
1674 				pnfs_lseg_no_merge,
1675 				free_me);
1676 }
1677 
1678 static struct pnfs_layout_hdr *
1679 alloc_init_layout_hdr(struct inode *ino,
1680 		      struct nfs_open_context *ctx,
1681 		      gfp_t gfp_flags)
1682 {
1683 	struct pnfs_layout_hdr *lo;
1684 
1685 	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1686 	if (!lo)
1687 		return NULL;
1688 	refcount_set(&lo->plh_refcount, 1);
1689 	INIT_LIST_HEAD(&lo->plh_layouts);
1690 	INIT_LIST_HEAD(&lo->plh_segs);
1691 	INIT_LIST_HEAD(&lo->plh_return_segs);
1692 	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1693 	lo->plh_inode = ino;
1694 	lo->plh_lc_cred = get_cred(ctx->cred);
1695 	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1696 	return lo;
1697 }
1698 
1699 static struct pnfs_layout_hdr *
1700 pnfs_find_alloc_layout(struct inode *ino,
1701 		       struct nfs_open_context *ctx,
1702 		       gfp_t gfp_flags)
1703 	__releases(&ino->i_lock)
1704 	__acquires(&ino->i_lock)
1705 {
1706 	struct nfs_inode *nfsi = NFS_I(ino);
1707 	struct pnfs_layout_hdr *new = NULL;
1708 
1709 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1710 
1711 	if (nfsi->layout != NULL)
1712 		goto out_existing;
1713 	spin_unlock(&ino->i_lock);
1714 	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1715 	spin_lock(&ino->i_lock);
1716 
1717 	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1718 		nfsi->layout = new;
1719 		return new;
1720 	} else if (new != NULL)
1721 		pnfs_free_layout_hdr(new);
1722 out_existing:
1723 	pnfs_get_layout_hdr(nfsi->layout);
1724 	return nfsi->layout;
1725 }
1726 
1727 /*
1728  * iomode matching rules:
1729  * iomode	lseg	strict match
1730  *                      iomode
1731  * -----	-----	------ -----
1732  * ANY		READ	N/A    true
1733  * ANY		RW	N/A    true
1734  * RW		READ	N/A    false
1735  * RW		RW	N/A    true
1736  * READ		READ	N/A    true
1737  * READ		RW	true   false
1738  * READ		RW	false  true
1739  */
1740 static bool
1741 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1742 		 const struct pnfs_layout_range *range,
1743 		 bool strict_iomode)
1744 {
1745 	struct pnfs_layout_range range1;
1746 
1747 	if ((range->iomode == IOMODE_RW &&
1748 	     ls_range->iomode != IOMODE_RW) ||
1749 	    (range->iomode != ls_range->iomode &&
1750 	     strict_iomode) ||
1751 	    !pnfs_lseg_range_intersecting(ls_range, range))
1752 		return false;
1753 
1754 	/* range1 covers only the first byte in the range */
1755 	range1 = *range;
1756 	range1.length = 1;
1757 	return pnfs_lseg_range_contained(ls_range, &range1);
1758 }
1759 
1760 /*
1761  * lookup range in layout
1762  */
1763 static struct pnfs_layout_segment *
1764 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1765 		struct pnfs_layout_range *range,
1766 		bool strict_iomode)
1767 {
1768 	struct pnfs_layout_segment *lseg, *ret = NULL;
1769 
1770 	dprintk("%s:Begin\n", __func__);
1771 
1772 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1773 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1774 		    pnfs_lseg_range_match(&lseg->pls_range, range,
1775 					  strict_iomode)) {
1776 			ret = pnfs_get_lseg(lseg);
1777 			break;
1778 		}
1779 	}
1780 
1781 	dprintk("%s:Return lseg %p ref %d\n",
1782 		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1783 	return ret;
1784 }
1785 
1786 /*
1787  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1788  * to the MDS or over pNFS
1789  *
1790  * The nfs_inode read_io and write_io fields are cumulative counters reset
1791  * when there are no layout segments. Note that in pnfs_update_layout iomode
1792  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1793  * WRITE request.
1794  *
1795  * A return of true means use MDS I/O.
1796  *
1797  * From rfc 5661:
1798  * If a file's size is smaller than the file size threshold, data accesses
1799  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1800  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1801  * server.  If both file size and I/O size are provided, the client SHOULD
1802  * reach or exceed  both thresholds before sending its read or write
1803  * requests to the data server.
1804  */
1805 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1806 				     struct inode *ino, int iomode)
1807 {
1808 	struct nfs4_threshold *t = ctx->mdsthreshold;
1809 	struct nfs_inode *nfsi = NFS_I(ino);
1810 	loff_t fsize = i_size_read(ino);
1811 	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1812 
1813 	if (t == NULL)
1814 		return ret;
1815 
1816 	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1817 		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1818 
1819 	switch (iomode) {
1820 	case IOMODE_READ:
1821 		if (t->bm & THRESHOLD_RD) {
1822 			dprintk("%s fsize %llu\n", __func__, fsize);
1823 			size_set = true;
1824 			if (fsize < t->rd_sz)
1825 				size = true;
1826 		}
1827 		if (t->bm & THRESHOLD_RD_IO) {
1828 			dprintk("%s nfsi->read_io %llu\n", __func__,
1829 				nfsi->read_io);
1830 			io_set = true;
1831 			if (nfsi->read_io < t->rd_io_sz)
1832 				io = true;
1833 		}
1834 		break;
1835 	case IOMODE_RW:
1836 		if (t->bm & THRESHOLD_WR) {
1837 			dprintk("%s fsize %llu\n", __func__, fsize);
1838 			size_set = true;
1839 			if (fsize < t->wr_sz)
1840 				size = true;
1841 		}
1842 		if (t->bm & THRESHOLD_WR_IO) {
1843 			dprintk("%s nfsi->write_io %llu\n", __func__,
1844 				nfsi->write_io);
1845 			io_set = true;
1846 			if (nfsi->write_io < t->wr_io_sz)
1847 				io = true;
1848 		}
1849 		break;
1850 	}
1851 	if (size_set && io_set) {
1852 		if (size && io)
1853 			ret = true;
1854 	} else if (size || io)
1855 		ret = true;
1856 
1857 	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1858 	return ret;
1859 }
1860 
1861 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1862 {
1863 	/*
1864 	 * send layoutcommit as it can hold up layoutreturn due to lseg
1865 	 * reference
1866 	 */
1867 	pnfs_layoutcommit_inode(lo->plh_inode, false);
1868 	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1869 				   nfs_wait_bit_killable,
1870 				   TASK_KILLABLE);
1871 }
1872 
1873 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1874 {
1875 	atomic_inc(&lo->plh_outstanding);
1876 }
1877 
1878 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1879 {
1880 	if (atomic_dec_and_test(&lo->plh_outstanding))
1881 		wake_up_var(&lo->plh_outstanding);
1882 }
1883 
1884 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1885 {
1886 	unsigned long *bitlock = &lo->plh_flags;
1887 
1888 	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1889 	smp_mb__after_atomic();
1890 	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1891 }
1892 
1893 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1894 				struct nfs_server *server)
1895 {
1896 	if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1897 		struct nfs_client *clp = server->nfs_client;
1898 
1899 		/* The lo must be on the clp list if there is any
1900 		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1901 		 */
1902 		spin_lock(&clp->cl_lock);
1903 		list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1904 		spin_unlock(&clp->cl_lock);
1905 	}
1906 }
1907 
1908 /*
1909  * Layout segment is retreived from the server if not cached.
1910  * The appropriate layout segment is referenced and returned to the caller.
1911  */
1912 struct pnfs_layout_segment *
1913 pnfs_update_layout(struct inode *ino,
1914 		   struct nfs_open_context *ctx,
1915 		   loff_t pos,
1916 		   u64 count,
1917 		   enum pnfs_iomode iomode,
1918 		   bool strict_iomode,
1919 		   gfp_t gfp_flags)
1920 {
1921 	struct pnfs_layout_range arg = {
1922 		.iomode = iomode,
1923 		.offset = pos,
1924 		.length = count,
1925 	};
1926 	unsigned pg_offset;
1927 	struct nfs_server *server = NFS_SERVER(ino);
1928 	struct nfs_client *clp = server->nfs_client;
1929 	struct pnfs_layout_hdr *lo = NULL;
1930 	struct pnfs_layout_segment *lseg = NULL;
1931 	struct nfs4_layoutget *lgp;
1932 	nfs4_stateid stateid;
1933 	long timeout = 0;
1934 	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1935 	bool first;
1936 
1937 	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1938 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1939 				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1940 		goto out;
1941 	}
1942 
1943 	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1944 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1945 				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1946 		goto out;
1947 	}
1948 
1949 lookup_again:
1950 	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1951 	if (IS_ERR(lseg))
1952 		goto out;
1953 	first = false;
1954 	spin_lock(&ino->i_lock);
1955 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1956 	if (lo == NULL) {
1957 		spin_unlock(&ino->i_lock);
1958 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1959 				 PNFS_UPDATE_LAYOUT_NOMEM);
1960 		goto out;
1961 	}
1962 
1963 	/* Do we even need to bother with this? */
1964 	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1965 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1966 				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1967 		dprintk("%s matches recall, use MDS\n", __func__);
1968 		goto out_unlock;
1969 	}
1970 
1971 	/* if LAYOUTGET already failed once we don't try again */
1972 	if (pnfs_layout_io_test_failed(lo, iomode)) {
1973 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1974 				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1975 		goto out_unlock;
1976 	}
1977 
1978 	/*
1979 	 * If the layout segment list is empty, but there are outstanding
1980 	 * layoutget calls, then they might be subject to a layoutrecall.
1981 	 */
1982 	if (list_empty(&lo->plh_segs) &&
1983 	    atomic_read(&lo->plh_outstanding) != 0) {
1984 		spin_unlock(&ino->i_lock);
1985 		lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
1986 					!atomic_read(&lo->plh_outstanding)));
1987 		if (IS_ERR(lseg))
1988 			goto out_put_layout_hdr;
1989 		pnfs_put_layout_hdr(lo);
1990 		goto lookup_again;
1991 	}
1992 
1993 	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1994 	if (lseg) {
1995 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1996 				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1997 		goto out_unlock;
1998 	}
1999 
2000 	/*
2001 	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
2002 	 * stateid, or it has been invalidated, then we must use the open
2003 	 * stateid.
2004 	 */
2005 	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2006 		int status;
2007 
2008 		/*
2009 		 * The first layoutget for the file. Need to serialize per
2010 		 * RFC 5661 Errata 3208.
2011 		 */
2012 		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2013 				     &lo->plh_flags)) {
2014 			spin_unlock(&ino->i_lock);
2015 			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2016 						NFS_LAYOUT_FIRST_LAYOUTGET,
2017 						TASK_KILLABLE));
2018 			if (IS_ERR(lseg))
2019 				goto out_put_layout_hdr;
2020 			pnfs_put_layout_hdr(lo);
2021 			dprintk("%s retrying\n", __func__);
2022 			goto lookup_again;
2023 		}
2024 
2025 		spin_unlock(&ino->i_lock);
2026 		first = true;
2027 		status = nfs4_select_rw_stateid(ctx->state,
2028 					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2029 					NULL, &stateid, NULL);
2030 		if (status != 0) {
2031 			lseg = ERR_PTR(status);
2032 			trace_pnfs_update_layout(ino, pos, count,
2033 					iomode, lo, lseg,
2034 					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2035 			nfs4_schedule_stateid_recovery(server, ctx->state);
2036 			pnfs_clear_first_layoutget(lo);
2037 			pnfs_put_layout_hdr(lo);
2038 			goto lookup_again;
2039 		}
2040 		spin_lock(&ino->i_lock);
2041 	} else {
2042 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2043 	}
2044 
2045 	/*
2046 	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
2047 	 * for LAYOUTRETURN even if first is true.
2048 	 */
2049 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2050 		spin_unlock(&ino->i_lock);
2051 		dprintk("%s wait for layoutreturn\n", __func__);
2052 		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2053 		if (!IS_ERR(lseg)) {
2054 			if (first)
2055 				pnfs_clear_first_layoutget(lo);
2056 			pnfs_put_layout_hdr(lo);
2057 			dprintk("%s retrying\n", __func__);
2058 			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2059 					lseg, PNFS_UPDATE_LAYOUT_RETRY);
2060 			goto lookup_again;
2061 		}
2062 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2063 				PNFS_UPDATE_LAYOUT_RETURN);
2064 		goto out_put_layout_hdr;
2065 	}
2066 
2067 	if (pnfs_layoutgets_blocked(lo)) {
2068 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2069 				PNFS_UPDATE_LAYOUT_BLOCKED);
2070 		goto out_unlock;
2071 	}
2072 	nfs_layoutget_begin(lo);
2073 	spin_unlock(&ino->i_lock);
2074 
2075 	_add_to_server_list(lo, server);
2076 
2077 	pg_offset = arg.offset & ~PAGE_MASK;
2078 	if (pg_offset) {
2079 		arg.offset -= pg_offset;
2080 		arg.length += pg_offset;
2081 	}
2082 	if (arg.length != NFS4_MAX_UINT64)
2083 		arg.length = PAGE_ALIGN(arg.length);
2084 
2085 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2086 	if (!lgp) {
2087 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2088 					 PNFS_UPDATE_LAYOUT_NOMEM);
2089 		nfs_layoutget_end(lo);
2090 		goto out_put_layout_hdr;
2091 	}
2092 
2093 	lseg = nfs4_proc_layoutget(lgp, &timeout);
2094 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2095 				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2096 	nfs_layoutget_end(lo);
2097 	if (IS_ERR(lseg)) {
2098 		switch(PTR_ERR(lseg)) {
2099 		case -EBUSY:
2100 			if (time_after(jiffies, giveup))
2101 				lseg = NULL;
2102 			break;
2103 		case -ERECALLCONFLICT:
2104 		case -EAGAIN:
2105 			break;
2106 		default:
2107 			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2108 				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2109 				lseg = NULL;
2110 			}
2111 			goto out_put_layout_hdr;
2112 		}
2113 		if (lseg) {
2114 			if (first)
2115 				pnfs_clear_first_layoutget(lo);
2116 			trace_pnfs_update_layout(ino, pos, count,
2117 				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2118 			pnfs_put_layout_hdr(lo);
2119 			goto lookup_again;
2120 		}
2121 	} else {
2122 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2123 	}
2124 
2125 out_put_layout_hdr:
2126 	if (first)
2127 		pnfs_clear_first_layoutget(lo);
2128 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2129 				 PNFS_UPDATE_LAYOUT_EXIT);
2130 	pnfs_put_layout_hdr(lo);
2131 out:
2132 	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2133 			"(%s, offset: %llu, length: %llu)\n",
2134 			__func__, ino->i_sb->s_id,
2135 			(unsigned long long)NFS_FILEID(ino),
2136 			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2137 			iomode==IOMODE_RW ?  "read/write" : "read-only",
2138 			(unsigned long long)pos,
2139 			(unsigned long long)count);
2140 	return lseg;
2141 out_unlock:
2142 	spin_unlock(&ino->i_lock);
2143 	goto out_put_layout_hdr;
2144 }
2145 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2146 
2147 static bool
2148 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2149 {
2150 	switch (range->iomode) {
2151 	case IOMODE_READ:
2152 	case IOMODE_RW:
2153 		break;
2154 	default:
2155 		return false;
2156 	}
2157 	if (range->offset == NFS4_MAX_UINT64)
2158 		return false;
2159 	if (range->length == 0)
2160 		return false;
2161 	if (range->length != NFS4_MAX_UINT64 &&
2162 	    range->length > NFS4_MAX_UINT64 - range->offset)
2163 		return false;
2164 	return true;
2165 }
2166 
2167 static struct pnfs_layout_hdr *
2168 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2169 {
2170 	struct pnfs_layout_hdr *lo;
2171 
2172 	spin_lock(&ino->i_lock);
2173 	lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2174 	if (!lo)
2175 		goto out_unlock;
2176 	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2177 		goto out_unlock;
2178 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2179 		goto out_unlock;
2180 	if (pnfs_layoutgets_blocked(lo))
2181 		goto out_unlock;
2182 	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2183 		goto out_unlock;
2184 	nfs_layoutget_begin(lo);
2185 	spin_unlock(&ino->i_lock);
2186 	_add_to_server_list(lo, NFS_SERVER(ino));
2187 	return lo;
2188 
2189 out_unlock:
2190 	spin_unlock(&ino->i_lock);
2191 	pnfs_put_layout_hdr(lo);
2192 	return NULL;
2193 }
2194 
2195 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2196 				     struct nfs_open_context *ctx)
2197 {
2198 	struct inode *ino = data->dentry->d_inode;
2199 	struct pnfs_layout_range rng = {
2200 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2201 			  IOMODE_RW: IOMODE_READ,
2202 		.offset = 0,
2203 		.length = NFS4_MAX_UINT64,
2204 	};
2205 	struct nfs4_layoutget *lgp;
2206 	struct pnfs_layout_hdr *lo;
2207 
2208 	/* Heuristic: don't send layoutget if we have cached data */
2209 	if (rng.iomode == IOMODE_READ &&
2210 	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2211 		return;
2212 
2213 	lo = _pnfs_grab_empty_layout(ino, ctx);
2214 	if (!lo)
2215 		return;
2216 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2217 					     &rng, GFP_KERNEL);
2218 	if (!lgp) {
2219 		pnfs_clear_first_layoutget(lo);
2220 		pnfs_put_layout_hdr(lo);
2221 		return;
2222 	}
2223 	data->lgp = lgp;
2224 	data->o_arg.lg_args = &lgp->args;
2225 	data->o_res.lg_res = &lgp->res;
2226 }
2227 
2228 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2229 				     struct nfs_open_context *ctx)
2230 {
2231 	struct pnfs_layout_range rng = {
2232 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2233 			  IOMODE_RW: IOMODE_READ,
2234 		.offset = 0,
2235 		.length = NFS4_MAX_UINT64,
2236 	};
2237 	struct nfs4_layoutget *lgp;
2238 
2239 	lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2240 					     &rng, GFP_KERNEL);
2241 	if (!lgp)
2242 		return;
2243 	data->lgp = lgp;
2244 	data->o_arg.lg_args = &lgp->args;
2245 	data->o_res.lg_res = &lgp->res;
2246 }
2247 
2248 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2249 			 struct nfs_open_context *ctx)
2250 {
2251 	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2252 
2253 	if (!(pnfs_enabled_sb(server) &&
2254 	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2255 		return;
2256 	/* Could check on max_ops, but currently hardcoded high enough */
2257 	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2258 		return;
2259 	if (data->state)
2260 		_lgopen_prepare_attached(data, ctx);
2261 	else
2262 		_lgopen_prepare_floating(data, ctx);
2263 }
2264 
2265 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2266 		       struct nfs_open_context *ctx)
2267 {
2268 	struct pnfs_layout_hdr *lo;
2269 	struct pnfs_layout_segment *lseg;
2270 	struct nfs_server *srv = NFS_SERVER(ino);
2271 	u32 iomode;
2272 
2273 	if (!lgp)
2274 		return;
2275 	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2276 	if (lgp->res.status) {
2277 		switch (lgp->res.status) {
2278 		default:
2279 			break;
2280 		/*
2281 		 * Halt lgopen attempts if the server doesn't recognise
2282 		 * the "current stateid" value, the layout type, or the
2283 		 * layoutget operation as being valid.
2284 		 * Also if it complains about too many ops in the compound
2285 		 * or of the request/reply being too big.
2286 		 */
2287 		case -NFS4ERR_BAD_STATEID:
2288 		case -NFS4ERR_NOTSUPP:
2289 		case -NFS4ERR_REP_TOO_BIG:
2290 		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2291 		case -NFS4ERR_REQ_TOO_BIG:
2292 		case -NFS4ERR_TOO_MANY_OPS:
2293 		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2294 			srv->caps &= ~NFS_CAP_LGOPEN;
2295 		}
2296 		return;
2297 	}
2298 	if (!lgp->args.inode) {
2299 		lo = _pnfs_grab_empty_layout(ino, ctx);
2300 		if (!lo)
2301 			return;
2302 		lgp->args.inode = ino;
2303 	} else
2304 		lo = NFS_I(lgp->args.inode)->layout;
2305 
2306 	lseg = pnfs_layout_process(lgp);
2307 	if (!IS_ERR(lseg)) {
2308 		iomode = lgp->args.range.iomode;
2309 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2310 		pnfs_put_lseg(lseg);
2311 	}
2312 }
2313 
2314 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2315 {
2316 	if (lgp != NULL) {
2317 		struct inode *inode = lgp->args.inode;
2318 		if (inode) {
2319 			struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2320 			pnfs_clear_first_layoutget(lo);
2321 			nfs_layoutget_end(lo);
2322 		}
2323 		pnfs_layoutget_free(lgp);
2324 	}
2325 }
2326 
2327 struct pnfs_layout_segment *
2328 pnfs_layout_process(struct nfs4_layoutget *lgp)
2329 {
2330 	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2331 	struct nfs4_layoutget_res *res = &lgp->res;
2332 	struct pnfs_layout_segment *lseg;
2333 	struct inode *ino = lo->plh_inode;
2334 	LIST_HEAD(free_me);
2335 
2336 	if (!pnfs_sanity_check_layout_range(&res->range))
2337 		return ERR_PTR(-EINVAL);
2338 
2339 	/* Inject layout blob into I/O device driver */
2340 	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2341 	if (IS_ERR_OR_NULL(lseg)) {
2342 		if (!lseg)
2343 			lseg = ERR_PTR(-ENOMEM);
2344 
2345 		dprintk("%s: Could not allocate layout: error %ld\n",
2346 		       __func__, PTR_ERR(lseg));
2347 		return lseg;
2348 	}
2349 
2350 	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2351 
2352 	spin_lock(&ino->i_lock);
2353 	if (pnfs_layoutgets_blocked(lo)) {
2354 		dprintk("%s forget reply due to state\n", __func__);
2355 		goto out_forget;
2356 	}
2357 
2358 	if (!pnfs_layout_is_valid(lo)) {
2359 		/* We have a completely new layout */
2360 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2361 	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2362 		/* existing state ID, make sure the sequence number matches. */
2363 		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2364 			dprintk("%s forget reply due to sequence\n", __func__);
2365 			goto out_forget;
2366 		}
2367 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2368 	} else {
2369 		/*
2370 		 * We got an entirely new state ID.  Mark all segments for the
2371 		 * inode invalid, and retry the layoutget
2372 		 */
2373 		pnfs_mark_layout_stateid_invalid(lo, &free_me);
2374 		goto out_forget;
2375 	}
2376 
2377 	pnfs_get_lseg(lseg);
2378 	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2379 
2380 
2381 	if (res->return_on_close)
2382 		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2383 
2384 	spin_unlock(&ino->i_lock);
2385 	pnfs_free_lseg_list(&free_me);
2386 	return lseg;
2387 
2388 out_forget:
2389 	spin_unlock(&ino->i_lock);
2390 	lseg->pls_layout = lo;
2391 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2392 	return ERR_PTR(-EAGAIN);
2393 }
2394 
2395 static int
2396 mark_lseg_invalid_or_return(struct pnfs_layout_segment *lseg,
2397 		struct list_head *tmp_list)
2398 {
2399 	if (!mark_lseg_invalid(lseg, tmp_list))
2400 		return 0;
2401 	pnfs_cache_lseg_for_layoutreturn(lseg->pls_layout, lseg);
2402 	return 1;
2403 }
2404 
2405 /**
2406  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2407  * @lo: pointer to layout header
2408  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2409  * @return_range: describe layout segment ranges to be returned
2410  * @seq: stateid seqid to match
2411  *
2412  * This function is mainly intended for use by layoutrecall. It attempts
2413  * to free the layout segment immediately, or else to mark it for return
2414  * as soon as its reference count drops to zero.
2415  *
2416  * Returns
2417  * - 0: a layoutreturn needs to be scheduled.
2418  * - EBUSY: there are layout segment that are still in use.
2419  * - ENOENT: there are no layout segments that need to be returned.
2420  */
2421 int
2422 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2423 				struct list_head *tmp_list,
2424 				const struct pnfs_layout_range *return_range,
2425 				u32 seq)
2426 {
2427 	struct pnfs_layout_segment *lseg, *next;
2428 	int remaining = 0;
2429 
2430 	dprintk("%s:Begin lo %p\n", __func__, lo);
2431 
2432 	assert_spin_locked(&lo->plh_inode->i_lock);
2433 
2434 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2435 		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2436 			dprintk("%s: marking lseg %p iomode %d "
2437 				"offset %llu length %llu\n", __func__,
2438 				lseg, lseg->pls_range.iomode,
2439 				lseg->pls_range.offset,
2440 				lseg->pls_range.length);
2441 			if (mark_lseg_invalid_or_return(lseg, tmp_list))
2442 				continue;
2443 			remaining++;
2444 			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2445 		}
2446 
2447 	if (remaining) {
2448 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2449 		return -EBUSY;
2450 	}
2451 
2452 	if (!list_empty(&lo->plh_return_segs)) {
2453 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2454 		return 0;
2455 	}
2456 
2457 	return -ENOENT;
2458 }
2459 
2460 static void
2461 pnfs_mark_layout_for_return(struct inode *inode,
2462 			    const struct pnfs_layout_range *range)
2463 {
2464 	struct pnfs_layout_hdr *lo;
2465 	bool return_now = false;
2466 
2467 	spin_lock(&inode->i_lock);
2468 	lo = NFS_I(inode)->layout;
2469 	if (!pnfs_layout_is_valid(lo)) {
2470 		spin_unlock(&inode->i_lock);
2471 		return;
2472 	}
2473 	pnfs_set_plh_return_info(lo, range->iomode, 0);
2474 	/*
2475 	 * mark all matching lsegs so that we are sure to have no live
2476 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2477 	 * for how it works.
2478 	 */
2479 	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2480 		const struct cred *cred;
2481 		nfs4_stateid stateid;
2482 		enum pnfs_iomode iomode;
2483 
2484 		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2485 		spin_unlock(&inode->i_lock);
2486 		if (return_now)
2487 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2488 	} else {
2489 		spin_unlock(&inode->i_lock);
2490 		nfs_commit_inode(inode, 0);
2491 	}
2492 }
2493 
2494 void pnfs_error_mark_layout_for_return(struct inode *inode,
2495 				       struct pnfs_layout_segment *lseg)
2496 {
2497 	struct pnfs_layout_range range = {
2498 		.iomode = lseg->pls_range.iomode,
2499 		.offset = 0,
2500 		.length = NFS4_MAX_UINT64,
2501 	};
2502 
2503 	pnfs_mark_layout_for_return(inode, &range);
2504 }
2505 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2506 
2507 static bool
2508 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2509 {
2510 	return pnfs_layout_is_valid(lo) &&
2511 		!test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2512 		!test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2513 }
2514 
2515 static struct pnfs_layout_segment *
2516 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2517 		     const struct pnfs_layout_range *range,
2518 		     enum pnfs_iomode iomode)
2519 {
2520 	struct pnfs_layout_segment *lseg;
2521 
2522 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2523 		if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2524 			continue;
2525 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2526 			continue;
2527 		if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2528 			continue;
2529 		if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2530 			return lseg;
2531 	}
2532 	return NULL;
2533 }
2534 
2535 /* Find open file states whose mode matches that of the range */
2536 static bool
2537 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2538 				 const struct pnfs_layout_range *range)
2539 {
2540 	struct list_head *head;
2541 	struct nfs_open_context *ctx;
2542 	fmode_t mode = 0;
2543 
2544 	if (!pnfs_layout_can_be_returned(lo) ||
2545 	    !pnfs_find_first_lseg(lo, range, range->iomode))
2546 		return false;
2547 
2548 	head = &NFS_I(lo->plh_inode)->open_files;
2549 	list_for_each_entry_rcu(ctx, head, list) {
2550 		if (ctx->state)
2551 			mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2552 	}
2553 
2554 	switch (range->iomode) {
2555 	default:
2556 		break;
2557 	case IOMODE_READ:
2558 		mode &= ~FMODE_WRITE;
2559 		break;
2560 	case IOMODE_RW:
2561 		if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2562 			mode &= ~FMODE_READ;
2563 	}
2564 	return mode == 0;
2565 }
2566 
2567 static int
2568 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2569 {
2570 	const struct pnfs_layout_range *range = data;
2571 	struct pnfs_layout_hdr *lo;
2572 	struct inode *inode;
2573 restart:
2574 	rcu_read_lock();
2575 	list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2576 		if (!pnfs_layout_can_be_returned(lo) ||
2577 		    test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2578 			continue;
2579 		inode = lo->plh_inode;
2580 		spin_lock(&inode->i_lock);
2581 		if (!pnfs_should_return_unused_layout(lo, range)) {
2582 			spin_unlock(&inode->i_lock);
2583 			continue;
2584 		}
2585 		spin_unlock(&inode->i_lock);
2586 		inode = pnfs_grab_inode_layout_hdr(lo);
2587 		if (!inode)
2588 			continue;
2589 		rcu_read_unlock();
2590 		pnfs_mark_layout_for_return(inode, range);
2591 		iput(inode);
2592 		cond_resched();
2593 		goto restart;
2594 	}
2595 	rcu_read_unlock();
2596 	return 0;
2597 }
2598 
2599 void
2600 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2601 				 enum pnfs_iomode iomode)
2602 {
2603 	struct pnfs_layout_range range = {
2604 		.iomode = iomode,
2605 		.offset = 0,
2606 		.length = NFS4_MAX_UINT64,
2607 	};
2608 
2609 	nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2610 			&range);
2611 }
2612 
2613 void
2614 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2615 {
2616 	if (pgio->pg_lseg == NULL ||
2617 	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2618 		return;
2619 	pnfs_put_lseg(pgio->pg_lseg);
2620 	pgio->pg_lseg = NULL;
2621 }
2622 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2623 
2624 /*
2625  * Check for any intersection between the request and the pgio->pg_lseg,
2626  * and if none, put this pgio->pg_lseg away.
2627  */
2628 void
2629 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2630 {
2631 	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2632 		pnfs_put_lseg(pgio->pg_lseg);
2633 		pgio->pg_lseg = NULL;
2634 	}
2635 }
2636 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2637 
2638 void
2639 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2640 {
2641 	u64 rd_size = req->wb_bytes;
2642 
2643 	pnfs_generic_pg_check_layout(pgio);
2644 	pnfs_generic_pg_check_range(pgio, req);
2645 	if (pgio->pg_lseg == NULL) {
2646 		if (pgio->pg_dreq == NULL)
2647 			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2648 		else
2649 			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2650 
2651 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2652 						   nfs_req_openctx(req),
2653 						   req_offset(req),
2654 						   rd_size,
2655 						   IOMODE_READ,
2656 						   false,
2657 						   GFP_KERNEL);
2658 		if (IS_ERR(pgio->pg_lseg)) {
2659 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2660 			pgio->pg_lseg = NULL;
2661 			return;
2662 		}
2663 	}
2664 	/* If no lseg, fall back to read through mds */
2665 	if (pgio->pg_lseg == NULL)
2666 		nfs_pageio_reset_read_mds(pgio);
2667 
2668 }
2669 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2670 
2671 void
2672 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2673 			   struct nfs_page *req, u64 wb_size)
2674 {
2675 	pnfs_generic_pg_check_layout(pgio);
2676 	pnfs_generic_pg_check_range(pgio, req);
2677 	if (pgio->pg_lseg == NULL) {
2678 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2679 						   nfs_req_openctx(req),
2680 						   req_offset(req),
2681 						   wb_size,
2682 						   IOMODE_RW,
2683 						   false,
2684 						   GFP_KERNEL);
2685 		if (IS_ERR(pgio->pg_lseg)) {
2686 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2687 			pgio->pg_lseg = NULL;
2688 			return;
2689 		}
2690 	}
2691 	/* If no lseg, fall back to write through mds */
2692 	if (pgio->pg_lseg == NULL)
2693 		nfs_pageio_reset_write_mds(pgio);
2694 }
2695 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2696 
2697 void
2698 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2699 {
2700 	if (desc->pg_lseg) {
2701 		pnfs_put_lseg(desc->pg_lseg);
2702 		desc->pg_lseg = NULL;
2703 	}
2704 }
2705 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2706 
2707 /*
2708  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2709  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2710  */
2711 size_t
2712 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2713 		     struct nfs_page *prev, struct nfs_page *req)
2714 {
2715 	unsigned int size;
2716 	u64 seg_end, req_start, seg_left;
2717 
2718 	size = nfs_generic_pg_test(pgio, prev, req);
2719 	if (!size)
2720 		return 0;
2721 
2722 	/*
2723 	 * 'size' contains the number of bytes left in the current page (up
2724 	 * to the original size asked for in @req->wb_bytes).
2725 	 *
2726 	 * Calculate how many bytes are left in the layout segment
2727 	 * and if there are less bytes than 'size', return that instead.
2728 	 *
2729 	 * Please also note that 'end_offset' is actually the offset of the
2730 	 * first byte that lies outside the pnfs_layout_range. FIXME?
2731 	 *
2732 	 */
2733 	if (pgio->pg_lseg) {
2734 		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2735 				     pgio->pg_lseg->pls_range.length);
2736 		req_start = req_offset(req);
2737 
2738 		/* start of request is past the last byte of this segment */
2739 		if (req_start >= seg_end)
2740 			return 0;
2741 
2742 		/* adjust 'size' iff there are fewer bytes left in the
2743 		 * segment than what nfs_generic_pg_test returned */
2744 		seg_left = seg_end - req_start;
2745 		if (seg_left < size)
2746 			size = (unsigned int)seg_left;
2747 	}
2748 
2749 	return size;
2750 }
2751 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2752 
2753 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2754 {
2755 	struct nfs_pageio_descriptor pgio;
2756 
2757 	/* Resend all requests through the MDS */
2758 	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2759 			      hdr->completion_ops);
2760 	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2761 	return nfs_pageio_resend(&pgio, hdr);
2762 }
2763 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2764 
2765 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2766 {
2767 
2768 	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2769 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2770 	    PNFS_LAYOUTRET_ON_ERROR) {
2771 		pnfs_return_layout(hdr->inode);
2772 	}
2773 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2774 		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2775 }
2776 
2777 /*
2778  * Called by non rpc-based layout drivers
2779  */
2780 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2781 {
2782 	if (likely(!hdr->pnfs_error)) {
2783 		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2784 				hdr->mds_offset + hdr->res.count);
2785 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2786 	}
2787 	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2788 	if (unlikely(hdr->pnfs_error))
2789 		pnfs_ld_handle_write_error(hdr);
2790 	hdr->mds_ops->rpc_release(hdr);
2791 }
2792 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2793 
2794 static void
2795 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2796 		struct nfs_pgio_header *hdr)
2797 {
2798 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2799 
2800 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2801 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2802 		nfs_pageio_reset_write_mds(desc);
2803 		mirror->pg_recoalesce = 1;
2804 	}
2805 	hdr->completion_ops->completion(hdr);
2806 }
2807 
2808 static enum pnfs_try_status
2809 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2810 			const struct rpc_call_ops *call_ops,
2811 			struct pnfs_layout_segment *lseg,
2812 			int how)
2813 {
2814 	struct inode *inode = hdr->inode;
2815 	enum pnfs_try_status trypnfs;
2816 	struct nfs_server *nfss = NFS_SERVER(inode);
2817 
2818 	hdr->mds_ops = call_ops;
2819 
2820 	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2821 		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2822 	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2823 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2824 		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2825 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2826 	return trypnfs;
2827 }
2828 
2829 static void
2830 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2831 	      struct nfs_pgio_header *hdr, int how)
2832 {
2833 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2834 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2835 	enum pnfs_try_status trypnfs;
2836 
2837 	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2838 	switch (trypnfs) {
2839 	case PNFS_NOT_ATTEMPTED:
2840 		pnfs_write_through_mds(desc, hdr);
2841 	case PNFS_ATTEMPTED:
2842 		break;
2843 	case PNFS_TRY_AGAIN:
2844 		/* cleanup hdr and prepare to redo pnfs */
2845 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2846 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2847 			list_splice_init(&hdr->pages, &mirror->pg_list);
2848 			mirror->pg_recoalesce = 1;
2849 		}
2850 		hdr->mds_ops->rpc_release(hdr);
2851 	}
2852 }
2853 
2854 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2855 {
2856 	pnfs_put_lseg(hdr->lseg);
2857 	nfs_pgio_header_free(hdr);
2858 }
2859 
2860 int
2861 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2862 {
2863 	struct nfs_pgio_header *hdr;
2864 	int ret;
2865 
2866 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2867 	if (!hdr) {
2868 		desc->pg_error = -ENOMEM;
2869 		return desc->pg_error;
2870 	}
2871 	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2872 
2873 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2874 	ret = nfs_generic_pgio(desc, hdr);
2875 	if (!ret)
2876 		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2877 
2878 	return ret;
2879 }
2880 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2881 
2882 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2883 {
2884 	struct nfs_pageio_descriptor pgio;
2885 
2886 	/* Resend all requests through the MDS */
2887 	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2888 	return nfs_pageio_resend(&pgio, hdr);
2889 }
2890 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2891 
2892 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2893 {
2894 	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2895 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2896 	    PNFS_LAYOUTRET_ON_ERROR) {
2897 		pnfs_return_layout(hdr->inode);
2898 	}
2899 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2900 		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2901 }
2902 
2903 /*
2904  * Called by non rpc-based layout drivers
2905  */
2906 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2907 {
2908 	if (likely(!hdr->pnfs_error))
2909 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2910 	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2911 	if (unlikely(hdr->pnfs_error))
2912 		pnfs_ld_handle_read_error(hdr);
2913 	hdr->mds_ops->rpc_release(hdr);
2914 }
2915 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2916 
2917 static void
2918 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2919 		struct nfs_pgio_header *hdr)
2920 {
2921 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2922 
2923 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2924 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2925 		nfs_pageio_reset_read_mds(desc);
2926 		mirror->pg_recoalesce = 1;
2927 	}
2928 	hdr->completion_ops->completion(hdr);
2929 }
2930 
2931 /*
2932  * Call the appropriate parallel I/O subsystem read function.
2933  */
2934 static enum pnfs_try_status
2935 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2936 		       const struct rpc_call_ops *call_ops,
2937 		       struct pnfs_layout_segment *lseg)
2938 {
2939 	struct inode *inode = hdr->inode;
2940 	struct nfs_server *nfss = NFS_SERVER(inode);
2941 	enum pnfs_try_status trypnfs;
2942 
2943 	hdr->mds_ops = call_ops;
2944 
2945 	dprintk("%s: Reading ino:%lu %u@%llu\n",
2946 		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2947 
2948 	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2949 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2950 		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2951 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2952 	return trypnfs;
2953 }
2954 
2955 /* Resend all requests through pnfs. */
2956 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr)
2957 {
2958 	struct nfs_pageio_descriptor pgio;
2959 
2960 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2961 		/* Prevent deadlocks with layoutreturn! */
2962 		pnfs_put_lseg(hdr->lseg);
2963 		hdr->lseg = NULL;
2964 
2965 		nfs_pageio_init_read(&pgio, hdr->inode, false,
2966 					hdr->completion_ops);
2967 		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2968 	}
2969 }
2970 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2971 
2972 static void
2973 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2974 {
2975 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2976 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2977 	enum pnfs_try_status trypnfs;
2978 
2979 	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2980 	switch (trypnfs) {
2981 	case PNFS_NOT_ATTEMPTED:
2982 		pnfs_read_through_mds(desc, hdr);
2983 	case PNFS_ATTEMPTED:
2984 		break;
2985 	case PNFS_TRY_AGAIN:
2986 		/* cleanup hdr and prepare to redo pnfs */
2987 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2988 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2989 			list_splice_init(&hdr->pages, &mirror->pg_list);
2990 			mirror->pg_recoalesce = 1;
2991 		}
2992 		hdr->mds_ops->rpc_release(hdr);
2993 	}
2994 }
2995 
2996 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2997 {
2998 	pnfs_put_lseg(hdr->lseg);
2999 	nfs_pgio_header_free(hdr);
3000 }
3001 
3002 int
3003 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
3004 {
3005 	struct nfs_pgio_header *hdr;
3006 	int ret;
3007 
3008 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
3009 	if (!hdr) {
3010 		desc->pg_error = -ENOMEM;
3011 		return desc->pg_error;
3012 	}
3013 	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3014 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3015 	ret = nfs_generic_pgio(desc, hdr);
3016 	if (!ret)
3017 		pnfs_do_read(desc, hdr);
3018 	return ret;
3019 }
3020 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3021 
3022 static void pnfs_clear_layoutcommitting(struct inode *inode)
3023 {
3024 	unsigned long *bitlock = &NFS_I(inode)->flags;
3025 
3026 	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3027 	smp_mb__after_atomic();
3028 	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3029 }
3030 
3031 /*
3032  * There can be multiple RW segments.
3033  */
3034 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3035 {
3036 	struct pnfs_layout_segment *lseg;
3037 
3038 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3039 		if (lseg->pls_range.iomode == IOMODE_RW &&
3040 		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3041 			list_add(&lseg->pls_lc_list, listp);
3042 	}
3043 }
3044 
3045 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3046 {
3047 	struct pnfs_layout_segment *lseg, *tmp;
3048 
3049 	/* Matched by references in pnfs_set_layoutcommit */
3050 	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3051 		list_del_init(&lseg->pls_lc_list);
3052 		pnfs_put_lseg(lseg);
3053 	}
3054 
3055 	pnfs_clear_layoutcommitting(inode);
3056 }
3057 
3058 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3059 {
3060 	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3061 }
3062 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3063 
3064 void
3065 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3066 		loff_t end_pos)
3067 {
3068 	struct nfs_inode *nfsi = NFS_I(inode);
3069 	bool mark_as_dirty = false;
3070 
3071 	spin_lock(&inode->i_lock);
3072 	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3073 		nfsi->layout->plh_lwb = end_pos;
3074 		mark_as_dirty = true;
3075 		dprintk("%s: Set layoutcommit for inode %lu ",
3076 			__func__, inode->i_ino);
3077 	} else if (end_pos > nfsi->layout->plh_lwb)
3078 		nfsi->layout->plh_lwb = end_pos;
3079 	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3080 		/* references matched in nfs4_layoutcommit_release */
3081 		pnfs_get_lseg(lseg);
3082 	}
3083 	spin_unlock(&inode->i_lock);
3084 	dprintk("%s: lseg %p end_pos %llu\n",
3085 		__func__, lseg, nfsi->layout->plh_lwb);
3086 
3087 	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3088 	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3089 	if (mark_as_dirty)
3090 		mark_inode_dirty_sync(inode);
3091 }
3092 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3093 
3094 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3095 {
3096 	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3097 
3098 	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3099 		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3100 	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3101 }
3102 
3103 /*
3104  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3105  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3106  * data to disk to allow the server to recover the data if it crashes.
3107  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3108  * is off, and a COMMIT is sent to a data server, or
3109  * if WRITEs to a data server return NFS_DATA_SYNC.
3110  */
3111 int
3112 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3113 {
3114 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3115 	struct nfs4_layoutcommit_data *data;
3116 	struct nfs_inode *nfsi = NFS_I(inode);
3117 	loff_t end_pos;
3118 	int status;
3119 
3120 	if (!pnfs_layoutcommit_outstanding(inode))
3121 		return 0;
3122 
3123 	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3124 
3125 	status = -EAGAIN;
3126 	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3127 		if (!sync)
3128 			goto out;
3129 		status = wait_on_bit_lock_action(&nfsi->flags,
3130 				NFS_INO_LAYOUTCOMMITTING,
3131 				nfs_wait_bit_killable,
3132 				TASK_KILLABLE);
3133 		if (status)
3134 			goto out;
3135 	}
3136 
3137 	status = -ENOMEM;
3138 	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3139 	data = kzalloc(sizeof(*data), GFP_NOFS);
3140 	if (!data)
3141 		goto clear_layoutcommitting;
3142 
3143 	status = 0;
3144 	spin_lock(&inode->i_lock);
3145 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3146 		goto out_unlock;
3147 
3148 	INIT_LIST_HEAD(&data->lseg_list);
3149 	pnfs_list_write_lseg(inode, &data->lseg_list);
3150 
3151 	end_pos = nfsi->layout->plh_lwb;
3152 
3153 	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3154 	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3155 	spin_unlock(&inode->i_lock);
3156 
3157 	data->args.inode = inode;
3158 	nfs_fattr_init(&data->fattr);
3159 	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3160 	data->res.fattr = &data->fattr;
3161 	if (end_pos != 0)
3162 		data->args.lastbytewritten = end_pos - 1;
3163 	else
3164 		data->args.lastbytewritten = U64_MAX;
3165 	data->res.server = NFS_SERVER(inode);
3166 
3167 	if (ld->prepare_layoutcommit) {
3168 		status = ld->prepare_layoutcommit(&data->args);
3169 		if (status) {
3170 			put_cred(data->cred);
3171 			spin_lock(&inode->i_lock);
3172 			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3173 			if (end_pos > nfsi->layout->plh_lwb)
3174 				nfsi->layout->plh_lwb = end_pos;
3175 			goto out_unlock;
3176 		}
3177 	}
3178 
3179 
3180 	status = nfs4_proc_layoutcommit(data, sync);
3181 out:
3182 	if (status)
3183 		mark_inode_dirty_sync(inode);
3184 	dprintk("<-- %s status %d\n", __func__, status);
3185 	return status;
3186 out_unlock:
3187 	spin_unlock(&inode->i_lock);
3188 	kfree(data);
3189 clear_layoutcommitting:
3190 	pnfs_clear_layoutcommitting(inode);
3191 	goto out;
3192 }
3193 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3194 
3195 int
3196 pnfs_generic_sync(struct inode *inode, bool datasync)
3197 {
3198 	return pnfs_layoutcommit_inode(inode, true);
3199 }
3200 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3201 
3202 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3203 {
3204 	struct nfs4_threshold *thp;
3205 
3206 	thp = kzalloc(sizeof(*thp), GFP_NOFS);
3207 	if (!thp) {
3208 		dprintk("%s mdsthreshold allocation failed\n", __func__);
3209 		return NULL;
3210 	}
3211 	return thp;
3212 }
3213 
3214 #if IS_ENABLED(CONFIG_NFS_V4_2)
3215 int
3216 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3217 {
3218 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3219 	struct nfs_server *server = NFS_SERVER(inode);
3220 	struct nfs_inode *nfsi = NFS_I(inode);
3221 	struct nfs42_layoutstat_data *data;
3222 	struct pnfs_layout_hdr *hdr;
3223 	int status = 0;
3224 
3225 	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3226 		goto out;
3227 
3228 	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3229 		goto out;
3230 
3231 	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3232 		goto out;
3233 
3234 	spin_lock(&inode->i_lock);
3235 	if (!NFS_I(inode)->layout) {
3236 		spin_unlock(&inode->i_lock);
3237 		goto out_clear_layoutstats;
3238 	}
3239 	hdr = NFS_I(inode)->layout;
3240 	pnfs_get_layout_hdr(hdr);
3241 	spin_unlock(&inode->i_lock);
3242 
3243 	data = kzalloc(sizeof(*data), gfp_flags);
3244 	if (!data) {
3245 		status = -ENOMEM;
3246 		goto out_put;
3247 	}
3248 
3249 	data->args.fh = NFS_FH(inode);
3250 	data->args.inode = inode;
3251 	status = ld->prepare_layoutstats(&data->args);
3252 	if (status)
3253 		goto out_free;
3254 
3255 	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3256 
3257 out:
3258 	dprintk("%s returns %d\n", __func__, status);
3259 	return status;
3260 
3261 out_free:
3262 	kfree(data);
3263 out_put:
3264 	pnfs_put_layout_hdr(hdr);
3265 out_clear_layoutstats:
3266 	smp_mb__before_atomic();
3267 	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3268 	smp_mb__after_atomic();
3269 	goto out;
3270 }
3271 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3272 #endif
3273 
3274 unsigned int layoutstats_timer;
3275 module_param(layoutstats_timer, uint, 0644);
3276 EXPORT_SYMBOL_GPL(layoutstats_timer);
3277