xref: /openbmc/linux/fs/nfs/pnfs.c (revision 9aab6601)
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 static bool
1230 pnfs_layout_segments_returnable(struct pnfs_layout_hdr *lo,
1231 				enum pnfs_iomode iomode,
1232 				u32 seq)
1233 {
1234 	struct pnfs_layout_range recall_range = {
1235 		.length = NFS4_MAX_UINT64,
1236 		.iomode = iomode,
1237 	};
1238 	return pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs,
1239 					       &recall_range, seq) != -EBUSY;
1240 }
1241 
1242 /* Return true if layoutreturn is needed */
1243 static bool
1244 pnfs_layout_need_return(struct pnfs_layout_hdr *lo)
1245 {
1246 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1247 		return false;
1248 	return pnfs_layout_segments_returnable(lo, lo->plh_return_iomode,
1249 					       lo->plh_return_seq);
1250 }
1251 
1252 static void pnfs_layoutreturn_before_put_layout_hdr(struct pnfs_layout_hdr *lo)
1253 {
1254 	struct inode *inode= lo->plh_inode;
1255 
1256 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1257 		return;
1258 	spin_lock(&inode->i_lock);
1259 	if (pnfs_layout_need_return(lo)) {
1260 		const struct cred *cred;
1261 		nfs4_stateid stateid;
1262 		enum pnfs_iomode iomode;
1263 		bool send;
1264 
1265 		send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
1266 		spin_unlock(&inode->i_lock);
1267 		if (send) {
1268 			/* Send an async layoutreturn so we dont deadlock */
1269 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
1270 		}
1271 	} else
1272 		spin_unlock(&inode->i_lock);
1273 }
1274 
1275 /*
1276  * Initiates a LAYOUTRETURN(FILE), and removes the pnfs_layout_hdr
1277  * when the layout segment list is empty.
1278  *
1279  * Note that a pnfs_layout_hdr can exist with an empty layout segment
1280  * list when LAYOUTGET has failed, or when LAYOUTGET succeeded, but the
1281  * deviceid is marked invalid.
1282  */
1283 int
1284 _pnfs_return_layout(struct inode *ino)
1285 {
1286 	struct pnfs_layout_hdr *lo = NULL;
1287 	struct nfs_inode *nfsi = NFS_I(ino);
1288 	LIST_HEAD(tmp_list);
1289 	const struct cred *cred;
1290 	nfs4_stateid stateid;
1291 	int status = 0;
1292 	bool send, valid_layout;
1293 
1294 	dprintk("NFS: %s for inode %lu\n", __func__, ino->i_ino);
1295 
1296 	spin_lock(&ino->i_lock);
1297 	lo = nfsi->layout;
1298 	if (!lo) {
1299 		spin_unlock(&ino->i_lock);
1300 		dprintk("NFS: %s no layout to return\n", __func__);
1301 		goto out;
1302 	}
1303 	/* Reference matched in nfs4_layoutreturn_release */
1304 	pnfs_get_layout_hdr(lo);
1305 	/* Is there an outstanding layoutreturn ? */
1306 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1307 		spin_unlock(&ino->i_lock);
1308 		if (wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1309 					TASK_UNINTERRUPTIBLE))
1310 			goto out_put_layout_hdr;
1311 		spin_lock(&ino->i_lock);
1312 	}
1313 	valid_layout = pnfs_layout_is_valid(lo);
1314 	pnfs_clear_layoutcommit(ino, &tmp_list);
1315 	pnfs_mark_matching_lsegs_invalid(lo, &tmp_list, NULL, 0);
1316 
1317 	if (NFS_SERVER(ino)->pnfs_curr_ld->return_range) {
1318 		struct pnfs_layout_range range = {
1319 			.iomode		= IOMODE_ANY,
1320 			.offset		= 0,
1321 			.length		= NFS4_MAX_UINT64,
1322 		};
1323 		NFS_SERVER(ino)->pnfs_curr_ld->return_range(lo, &range);
1324 	}
1325 
1326 	/* Don't send a LAYOUTRETURN if list was initially empty */
1327 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags) ||
1328 			!valid_layout) {
1329 		spin_unlock(&ino->i_lock);
1330 		dprintk("NFS: %s no layout segments to return\n", __func__);
1331 		goto out_wait_layoutreturn;
1332 	}
1333 
1334 	send = pnfs_prepare_layoutreturn(lo, &stateid, &cred, NULL);
1335 	spin_unlock(&ino->i_lock);
1336 	if (send)
1337 		status = pnfs_send_layoutreturn(lo, &stateid, &cred, IOMODE_ANY, true);
1338 out_wait_layoutreturn:
1339 	wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN, TASK_UNINTERRUPTIBLE);
1340 out_put_layout_hdr:
1341 	pnfs_free_lseg_list(&tmp_list);
1342 	pnfs_put_layout_hdr(lo);
1343 out:
1344 	dprintk("<-- %s status: %d\n", __func__, status);
1345 	return status;
1346 }
1347 
1348 int
1349 pnfs_commit_and_return_layout(struct inode *inode)
1350 {
1351 	struct pnfs_layout_hdr *lo;
1352 	int ret;
1353 
1354 	spin_lock(&inode->i_lock);
1355 	lo = NFS_I(inode)->layout;
1356 	if (lo == NULL) {
1357 		spin_unlock(&inode->i_lock);
1358 		return 0;
1359 	}
1360 	pnfs_get_layout_hdr(lo);
1361 	/* Block new layoutgets and read/write to ds */
1362 	lo->plh_block_lgets++;
1363 	spin_unlock(&inode->i_lock);
1364 	filemap_fdatawait(inode->i_mapping);
1365 	ret = pnfs_layoutcommit_inode(inode, true);
1366 	if (ret == 0)
1367 		ret = _pnfs_return_layout(inode);
1368 	spin_lock(&inode->i_lock);
1369 	lo->plh_block_lgets--;
1370 	spin_unlock(&inode->i_lock);
1371 	pnfs_put_layout_hdr(lo);
1372 	return ret;
1373 }
1374 
1375 bool pnfs_roc(struct inode *ino,
1376 		struct nfs4_layoutreturn_args *args,
1377 		struct nfs4_layoutreturn_res *res,
1378 		const struct cred *cred)
1379 {
1380 	struct nfs_inode *nfsi = NFS_I(ino);
1381 	struct nfs_open_context *ctx;
1382 	struct nfs4_state *state;
1383 	struct pnfs_layout_hdr *lo;
1384 	struct pnfs_layout_segment *lseg, *next;
1385 	const struct cred *lc_cred;
1386 	nfs4_stateid stateid;
1387 	enum pnfs_iomode iomode = 0;
1388 	bool layoutreturn = false, roc = false;
1389 	bool skip_read = false;
1390 
1391 	if (!nfs_have_layout(ino))
1392 		return false;
1393 retry:
1394 	rcu_read_lock();
1395 	spin_lock(&ino->i_lock);
1396 	lo = nfsi->layout;
1397 	if (!lo || !pnfs_layout_is_valid(lo) ||
1398 	    test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1399 		lo = NULL;
1400 		goto out_noroc;
1401 	}
1402 	pnfs_get_layout_hdr(lo);
1403 	if (test_bit(NFS_LAYOUT_RETURN_LOCK, &lo->plh_flags)) {
1404 		spin_unlock(&ino->i_lock);
1405 		rcu_read_unlock();
1406 		wait_on_bit(&lo->plh_flags, NFS_LAYOUT_RETURN,
1407 				TASK_UNINTERRUPTIBLE);
1408 		pnfs_put_layout_hdr(lo);
1409 		goto retry;
1410 	}
1411 
1412 	/* no roc if we hold a delegation */
1413 	if (nfs4_check_delegation(ino, FMODE_READ)) {
1414 		if (nfs4_check_delegation(ino, FMODE_WRITE))
1415 			goto out_noroc;
1416 		skip_read = true;
1417 	}
1418 
1419 	list_for_each_entry_rcu(ctx, &nfsi->open_files, list) {
1420 		state = ctx->state;
1421 		if (state == NULL)
1422 			continue;
1423 		/* Don't return layout if there is open file state */
1424 		if (state->state & FMODE_WRITE)
1425 			goto out_noroc;
1426 		if (state->state & FMODE_READ)
1427 			skip_read = true;
1428 	}
1429 
1430 
1431 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list) {
1432 		if (skip_read && lseg->pls_range.iomode == IOMODE_READ)
1433 			continue;
1434 		/* If we are sending layoutreturn, invalidate all valid lsegs */
1435 		if (!test_and_clear_bit(NFS_LSEG_ROC, &lseg->pls_flags))
1436 			continue;
1437 		/*
1438 		 * Note: mark lseg for return so pnfs_layout_remove_lseg
1439 		 * doesn't invalidate the layout for us.
1440 		 */
1441 		set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
1442 		if (!mark_lseg_invalid(lseg, &lo->plh_return_segs))
1443 			continue;
1444 		pnfs_set_plh_return_info(lo, lseg->pls_range.iomode, 0);
1445 	}
1446 
1447 	if (!test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
1448 		goto out_noroc;
1449 
1450 	/* ROC in two conditions:
1451 	 * 1. there are ROC lsegs
1452 	 * 2. we don't send layoutreturn
1453 	 */
1454 	/* lo ref dropped in pnfs_roc_release() */
1455 	layoutreturn = pnfs_prepare_layoutreturn(lo, &stateid, &lc_cred, &iomode);
1456 	/* If the creds don't match, we can't compound the layoutreturn */
1457 	if (!layoutreturn || cred_fscmp(cred, lc_cred) != 0)
1458 		goto out_noroc;
1459 
1460 	roc = layoutreturn;
1461 	pnfs_init_layoutreturn_args(args, lo, &stateid, iomode);
1462 	res->lrs_present = 0;
1463 	layoutreturn = false;
1464 	put_cred(lc_cred);
1465 
1466 out_noroc:
1467 	spin_unlock(&ino->i_lock);
1468 	rcu_read_unlock();
1469 	pnfs_layoutcommit_inode(ino, true);
1470 	if (roc) {
1471 		struct pnfs_layoutdriver_type *ld = NFS_SERVER(ino)->pnfs_curr_ld;
1472 		if (ld->prepare_layoutreturn)
1473 			ld->prepare_layoutreturn(args);
1474 		pnfs_put_layout_hdr(lo);
1475 		return true;
1476 	}
1477 	if (layoutreturn)
1478 		pnfs_send_layoutreturn(lo, &stateid, &lc_cred, iomode, true);
1479 	pnfs_put_layout_hdr(lo);
1480 	return false;
1481 }
1482 
1483 int pnfs_roc_done(struct rpc_task *task, struct inode *inode,
1484 		struct nfs4_layoutreturn_args **argpp,
1485 		struct nfs4_layoutreturn_res **respp,
1486 		int *ret)
1487 {
1488 	struct nfs4_layoutreturn_args *arg = *argpp;
1489 	int retval = -EAGAIN;
1490 
1491 	if (!arg)
1492 		return 0;
1493 	/* Handle Layoutreturn errors */
1494 	switch (*ret) {
1495 	case 0:
1496 		retval = 0;
1497 		break;
1498 	case -NFS4ERR_NOMATCHING_LAYOUT:
1499 		/* Was there an RPC level error? If not, retry */
1500 		if (task->tk_rpc_status == 0)
1501 			break;
1502 		/* If the call was not sent, let caller handle it */
1503 		if (!RPC_WAS_SENT(task))
1504 			return 0;
1505 		/*
1506 		 * Otherwise, assume the call succeeded and
1507 		 * that we need to release the layout
1508 		 */
1509 		*ret = 0;
1510 		(*respp)->lrs_present = 0;
1511 		retval = 0;
1512 		break;
1513 	case -NFS4ERR_DELAY:
1514 		/* Let the caller handle the retry */
1515 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1516 		return 0;
1517 	case -NFS4ERR_OLD_STATEID:
1518 		if (!nfs4_layout_refresh_old_stateid(&arg->stateid,
1519 					&arg->range, inode))
1520 			break;
1521 		*ret = -NFS4ERR_NOMATCHING_LAYOUT;
1522 		return -EAGAIN;
1523 	}
1524 	*argpp = NULL;
1525 	*respp = NULL;
1526 	return retval;
1527 }
1528 
1529 void pnfs_roc_release(struct nfs4_layoutreturn_args *args,
1530 		struct nfs4_layoutreturn_res *res,
1531 		int ret)
1532 {
1533 	struct pnfs_layout_hdr *lo = args->layout;
1534 	const nfs4_stateid *arg_stateid = NULL;
1535 	const nfs4_stateid *res_stateid = NULL;
1536 	struct nfs4_xdr_opaque_data *ld_private = args->ld_private;
1537 
1538 	switch (ret) {
1539 	case -NFS4ERR_NOMATCHING_LAYOUT:
1540 		break;
1541 	case 0:
1542 		if (res->lrs_present)
1543 			res_stateid = &res->stateid;
1544 		fallthrough;
1545 	default:
1546 		arg_stateid = &args->stateid;
1547 	}
1548 	trace_nfs4_layoutreturn_on_close(args->inode, &args->stateid, ret);
1549 	pnfs_layoutreturn_free_lsegs(lo, arg_stateid, &args->range,
1550 			res_stateid);
1551 	if (ld_private && ld_private->ops && ld_private->ops->free)
1552 		ld_private->ops->free(ld_private);
1553 	pnfs_put_layout_hdr(lo);
1554 }
1555 
1556 bool pnfs_wait_on_layoutreturn(struct inode *ino, struct rpc_task *task)
1557 {
1558 	struct nfs_inode *nfsi = NFS_I(ino);
1559         struct pnfs_layout_hdr *lo;
1560         bool sleep = false;
1561 
1562 	/* we might not have grabbed lo reference. so need to check under
1563 	 * i_lock */
1564         spin_lock(&ino->i_lock);
1565         lo = nfsi->layout;
1566         if (lo && test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
1567                 rpc_sleep_on(&NFS_SERVER(ino)->roc_rpcwaitq, task, NULL);
1568                 sleep = true;
1569 	}
1570         spin_unlock(&ino->i_lock);
1571         return sleep;
1572 }
1573 
1574 /*
1575  * Compare two layout segments for sorting into layout cache.
1576  * We want to preferentially return RW over RO layouts, so ensure those
1577  * are seen first.
1578  */
1579 static s64
1580 pnfs_lseg_range_cmp(const struct pnfs_layout_range *l1,
1581 	   const struct pnfs_layout_range *l2)
1582 {
1583 	s64 d;
1584 
1585 	/* high offset > low offset */
1586 	d = l1->offset - l2->offset;
1587 	if (d)
1588 		return d;
1589 
1590 	/* short length > long length */
1591 	d = l2->length - l1->length;
1592 	if (d)
1593 		return d;
1594 
1595 	/* read > read/write */
1596 	return (int)(l1->iomode == IOMODE_READ) - (int)(l2->iomode == IOMODE_READ);
1597 }
1598 
1599 static bool
1600 pnfs_lseg_range_is_after(const struct pnfs_layout_range *l1,
1601 		const struct pnfs_layout_range *l2)
1602 {
1603 	return pnfs_lseg_range_cmp(l1, l2) > 0;
1604 }
1605 
1606 static bool
1607 pnfs_lseg_no_merge(struct pnfs_layout_segment *lseg,
1608 		struct pnfs_layout_segment *old)
1609 {
1610 	return false;
1611 }
1612 
1613 void
1614 pnfs_generic_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1615 		   struct pnfs_layout_segment *lseg,
1616 		   bool (*is_after)(const struct pnfs_layout_range *,
1617 			   const struct pnfs_layout_range *),
1618 		   bool (*do_merge)(struct pnfs_layout_segment *,
1619 			   struct pnfs_layout_segment *),
1620 		   struct list_head *free_me)
1621 {
1622 	struct pnfs_layout_segment *lp, *tmp;
1623 
1624 	dprintk("%s:Begin\n", __func__);
1625 
1626 	list_for_each_entry_safe(lp, tmp, &lo->plh_segs, pls_list) {
1627 		if (test_bit(NFS_LSEG_VALID, &lp->pls_flags) == 0)
1628 			continue;
1629 		if (do_merge(lseg, lp)) {
1630 			mark_lseg_invalid(lp, free_me);
1631 			continue;
1632 		}
1633 		if (is_after(&lseg->pls_range, &lp->pls_range))
1634 			continue;
1635 		list_add_tail(&lseg->pls_list, &lp->pls_list);
1636 		dprintk("%s: inserted lseg %p "
1637 			"iomode %d offset %llu length %llu before "
1638 			"lp %p iomode %d offset %llu length %llu\n",
1639 			__func__, lseg, lseg->pls_range.iomode,
1640 			lseg->pls_range.offset, lseg->pls_range.length,
1641 			lp, lp->pls_range.iomode, lp->pls_range.offset,
1642 			lp->pls_range.length);
1643 		goto out;
1644 	}
1645 	list_add_tail(&lseg->pls_list, &lo->plh_segs);
1646 	dprintk("%s: inserted lseg %p "
1647 		"iomode %d offset %llu length %llu at tail\n",
1648 		__func__, lseg, lseg->pls_range.iomode,
1649 		lseg->pls_range.offset, lseg->pls_range.length);
1650 out:
1651 	pnfs_get_layout_hdr(lo);
1652 
1653 	dprintk("%s:Return\n", __func__);
1654 }
1655 EXPORT_SYMBOL_GPL(pnfs_generic_layout_insert_lseg);
1656 
1657 static void
1658 pnfs_layout_insert_lseg(struct pnfs_layout_hdr *lo,
1659 		   struct pnfs_layout_segment *lseg,
1660 		   struct list_head *free_me)
1661 {
1662 	struct inode *inode = lo->plh_inode;
1663 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
1664 
1665 	if (ld->add_lseg != NULL)
1666 		ld->add_lseg(lo, lseg, free_me);
1667 	else
1668 		pnfs_generic_layout_insert_lseg(lo, lseg,
1669 				pnfs_lseg_range_is_after,
1670 				pnfs_lseg_no_merge,
1671 				free_me);
1672 }
1673 
1674 static struct pnfs_layout_hdr *
1675 alloc_init_layout_hdr(struct inode *ino,
1676 		      struct nfs_open_context *ctx,
1677 		      gfp_t gfp_flags)
1678 {
1679 	struct pnfs_layout_hdr *lo;
1680 
1681 	lo = pnfs_alloc_layout_hdr(ino, gfp_flags);
1682 	if (!lo)
1683 		return NULL;
1684 	refcount_set(&lo->plh_refcount, 1);
1685 	INIT_LIST_HEAD(&lo->plh_layouts);
1686 	INIT_LIST_HEAD(&lo->plh_segs);
1687 	INIT_LIST_HEAD(&lo->plh_return_segs);
1688 	INIT_LIST_HEAD(&lo->plh_bulk_destroy);
1689 	lo->plh_inode = ino;
1690 	lo->plh_lc_cred = get_cred(ctx->cred);
1691 	lo->plh_flags |= 1 << NFS_LAYOUT_INVALID_STID;
1692 	return lo;
1693 }
1694 
1695 static struct pnfs_layout_hdr *
1696 pnfs_find_alloc_layout(struct inode *ino,
1697 		       struct nfs_open_context *ctx,
1698 		       gfp_t gfp_flags)
1699 	__releases(&ino->i_lock)
1700 	__acquires(&ino->i_lock)
1701 {
1702 	struct nfs_inode *nfsi = NFS_I(ino);
1703 	struct pnfs_layout_hdr *new = NULL;
1704 
1705 	dprintk("%s Begin ino=%p layout=%p\n", __func__, ino, nfsi->layout);
1706 
1707 	if (nfsi->layout != NULL)
1708 		goto out_existing;
1709 	spin_unlock(&ino->i_lock);
1710 	new = alloc_init_layout_hdr(ino, ctx, gfp_flags);
1711 	spin_lock(&ino->i_lock);
1712 
1713 	if (likely(nfsi->layout == NULL)) {	/* Won the race? */
1714 		nfsi->layout = new;
1715 		return new;
1716 	} else if (new != NULL)
1717 		pnfs_free_layout_hdr(new);
1718 out_existing:
1719 	pnfs_get_layout_hdr(nfsi->layout);
1720 	return nfsi->layout;
1721 }
1722 
1723 /*
1724  * iomode matching rules:
1725  * iomode	lseg	strict match
1726  *                      iomode
1727  * -----	-----	------ -----
1728  * ANY		READ	N/A    true
1729  * ANY		RW	N/A    true
1730  * RW		READ	N/A    false
1731  * RW		RW	N/A    true
1732  * READ		READ	N/A    true
1733  * READ		RW	true   false
1734  * READ		RW	false  true
1735  */
1736 static bool
1737 pnfs_lseg_range_match(const struct pnfs_layout_range *ls_range,
1738 		 const struct pnfs_layout_range *range,
1739 		 bool strict_iomode)
1740 {
1741 	struct pnfs_layout_range range1;
1742 
1743 	if ((range->iomode == IOMODE_RW &&
1744 	     ls_range->iomode != IOMODE_RW) ||
1745 	    (range->iomode != ls_range->iomode &&
1746 	     strict_iomode) ||
1747 	    !pnfs_lseg_range_intersecting(ls_range, range))
1748 		return false;
1749 
1750 	/* range1 covers only the first byte in the range */
1751 	range1 = *range;
1752 	range1.length = 1;
1753 	return pnfs_lseg_range_contained(ls_range, &range1);
1754 }
1755 
1756 /*
1757  * lookup range in layout
1758  */
1759 static struct pnfs_layout_segment *
1760 pnfs_find_lseg(struct pnfs_layout_hdr *lo,
1761 		struct pnfs_layout_range *range,
1762 		bool strict_iomode)
1763 {
1764 	struct pnfs_layout_segment *lseg, *ret = NULL;
1765 
1766 	dprintk("%s:Begin\n", __func__);
1767 
1768 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
1769 		if (test_bit(NFS_LSEG_VALID, &lseg->pls_flags) &&
1770 		    pnfs_lseg_range_match(&lseg->pls_range, range,
1771 					  strict_iomode)) {
1772 			ret = pnfs_get_lseg(lseg);
1773 			break;
1774 		}
1775 	}
1776 
1777 	dprintk("%s:Return lseg %p ref %d\n",
1778 		__func__, ret, ret ? refcount_read(&ret->pls_refcount) : 0);
1779 	return ret;
1780 }
1781 
1782 /*
1783  * Use mdsthreshold hints set at each OPEN to determine if I/O should go
1784  * to the MDS or over pNFS
1785  *
1786  * The nfs_inode read_io and write_io fields are cumulative counters reset
1787  * when there are no layout segments. Note that in pnfs_update_layout iomode
1788  * is set to IOMODE_READ for a READ request, and set to IOMODE_RW for a
1789  * WRITE request.
1790  *
1791  * A return of true means use MDS I/O.
1792  *
1793  * From rfc 5661:
1794  * If a file's size is smaller than the file size threshold, data accesses
1795  * SHOULD be sent to the metadata server.  If an I/O request has a length that
1796  * is below the I/O size threshold, the I/O SHOULD be sent to the metadata
1797  * server.  If both file size and I/O size are provided, the client SHOULD
1798  * reach or exceed  both thresholds before sending its read or write
1799  * requests to the data server.
1800  */
1801 static bool pnfs_within_mdsthreshold(struct nfs_open_context *ctx,
1802 				     struct inode *ino, int iomode)
1803 {
1804 	struct nfs4_threshold *t = ctx->mdsthreshold;
1805 	struct nfs_inode *nfsi = NFS_I(ino);
1806 	loff_t fsize = i_size_read(ino);
1807 	bool size = false, size_set = false, io = false, io_set = false, ret = false;
1808 
1809 	if (t == NULL)
1810 		return ret;
1811 
1812 	dprintk("%s bm=0x%x rd_sz=%llu wr_sz=%llu rd_io=%llu wr_io=%llu\n",
1813 		__func__, t->bm, t->rd_sz, t->wr_sz, t->rd_io_sz, t->wr_io_sz);
1814 
1815 	switch (iomode) {
1816 	case IOMODE_READ:
1817 		if (t->bm & THRESHOLD_RD) {
1818 			dprintk("%s fsize %llu\n", __func__, fsize);
1819 			size_set = true;
1820 			if (fsize < t->rd_sz)
1821 				size = true;
1822 		}
1823 		if (t->bm & THRESHOLD_RD_IO) {
1824 			dprintk("%s nfsi->read_io %llu\n", __func__,
1825 				nfsi->read_io);
1826 			io_set = true;
1827 			if (nfsi->read_io < t->rd_io_sz)
1828 				io = true;
1829 		}
1830 		break;
1831 	case IOMODE_RW:
1832 		if (t->bm & THRESHOLD_WR) {
1833 			dprintk("%s fsize %llu\n", __func__, fsize);
1834 			size_set = true;
1835 			if (fsize < t->wr_sz)
1836 				size = true;
1837 		}
1838 		if (t->bm & THRESHOLD_WR_IO) {
1839 			dprintk("%s nfsi->write_io %llu\n", __func__,
1840 				nfsi->write_io);
1841 			io_set = true;
1842 			if (nfsi->write_io < t->wr_io_sz)
1843 				io = true;
1844 		}
1845 		break;
1846 	}
1847 	if (size_set && io_set) {
1848 		if (size && io)
1849 			ret = true;
1850 	} else if (size || io)
1851 		ret = true;
1852 
1853 	dprintk("<-- %s size %d io %d ret %d\n", __func__, size, io, ret);
1854 	return ret;
1855 }
1856 
1857 static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr *lo)
1858 {
1859 	/*
1860 	 * send layoutcommit as it can hold up layoutreturn due to lseg
1861 	 * reference
1862 	 */
1863 	pnfs_layoutcommit_inode(lo->plh_inode, false);
1864 	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
1865 				   nfs_wait_bit_killable,
1866 				   TASK_KILLABLE);
1867 }
1868 
1869 static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
1870 {
1871 	atomic_inc(&lo->plh_outstanding);
1872 }
1873 
1874 static void nfs_layoutget_end(struct pnfs_layout_hdr *lo)
1875 {
1876 	if (atomic_dec_and_test(&lo->plh_outstanding))
1877 		wake_up_var(&lo->plh_outstanding);
1878 }
1879 
1880 static void pnfs_clear_first_layoutget(struct pnfs_layout_hdr *lo)
1881 {
1882 	unsigned long *bitlock = &lo->plh_flags;
1883 
1884 	clear_bit_unlock(NFS_LAYOUT_FIRST_LAYOUTGET, bitlock);
1885 	smp_mb__after_atomic();
1886 	wake_up_bit(bitlock, NFS_LAYOUT_FIRST_LAYOUTGET);
1887 }
1888 
1889 static void _add_to_server_list(struct pnfs_layout_hdr *lo,
1890 				struct nfs_server *server)
1891 {
1892 	if (!test_and_set_bit(NFS_LAYOUT_HASHED, &lo->plh_flags)) {
1893 		struct nfs_client *clp = server->nfs_client;
1894 
1895 		/* The lo must be on the clp list if there is any
1896 		 * chance of a CB_LAYOUTRECALL(FILE) coming in.
1897 		 */
1898 		spin_lock(&clp->cl_lock);
1899 		list_add_tail_rcu(&lo->plh_layouts, &server->layouts);
1900 		spin_unlock(&clp->cl_lock);
1901 	}
1902 }
1903 
1904 /*
1905  * Layout segment is retreived from the server if not cached.
1906  * The appropriate layout segment is referenced and returned to the caller.
1907  */
1908 struct pnfs_layout_segment *
1909 pnfs_update_layout(struct inode *ino,
1910 		   struct nfs_open_context *ctx,
1911 		   loff_t pos,
1912 		   u64 count,
1913 		   enum pnfs_iomode iomode,
1914 		   bool strict_iomode,
1915 		   gfp_t gfp_flags)
1916 {
1917 	struct pnfs_layout_range arg = {
1918 		.iomode = iomode,
1919 		.offset = pos,
1920 		.length = count,
1921 	};
1922 	unsigned pg_offset;
1923 	struct nfs_server *server = NFS_SERVER(ino);
1924 	struct nfs_client *clp = server->nfs_client;
1925 	struct pnfs_layout_hdr *lo = NULL;
1926 	struct pnfs_layout_segment *lseg = NULL;
1927 	struct nfs4_layoutget *lgp;
1928 	nfs4_stateid stateid;
1929 	long timeout = 0;
1930 	unsigned long giveup = jiffies + (clp->cl_lease_time << 1);
1931 	bool first;
1932 
1933 	if (!pnfs_enabled_sb(NFS_SERVER(ino))) {
1934 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1935 				 PNFS_UPDATE_LAYOUT_NO_PNFS);
1936 		goto out;
1937 	}
1938 
1939 	if (pnfs_within_mdsthreshold(ctx, ino, iomode)) {
1940 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1941 				 PNFS_UPDATE_LAYOUT_MDSTHRESH);
1942 		goto out;
1943 	}
1944 
1945 lookup_again:
1946 	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
1947 	if (IS_ERR(lseg))
1948 		goto out;
1949 	first = false;
1950 	spin_lock(&ino->i_lock);
1951 	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
1952 	if (lo == NULL) {
1953 		spin_unlock(&ino->i_lock);
1954 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1955 				 PNFS_UPDATE_LAYOUT_NOMEM);
1956 		goto out;
1957 	}
1958 
1959 	/* Do we even need to bother with this? */
1960 	if (test_bit(NFS_LAYOUT_BULK_RECALL, &lo->plh_flags)) {
1961 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1962 				 PNFS_UPDATE_LAYOUT_BULK_RECALL);
1963 		dprintk("%s matches recall, use MDS\n", __func__);
1964 		goto out_unlock;
1965 	}
1966 
1967 	/* if LAYOUTGET already failed once we don't try again */
1968 	if (pnfs_layout_io_test_failed(lo, iomode)) {
1969 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1970 				 PNFS_UPDATE_LAYOUT_IO_TEST_FAIL);
1971 		goto out_unlock;
1972 	}
1973 
1974 	/*
1975 	 * If the layout segment list is empty, but there are outstanding
1976 	 * layoutget calls, then they might be subject to a layoutrecall.
1977 	 */
1978 	if (list_empty(&lo->plh_segs) &&
1979 	    atomic_read(&lo->plh_outstanding) != 0) {
1980 		spin_unlock(&ino->i_lock);
1981 		lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
1982 					!atomic_read(&lo->plh_outstanding)));
1983 		if (IS_ERR(lseg))
1984 			goto out_put_layout_hdr;
1985 		pnfs_put_layout_hdr(lo);
1986 		goto lookup_again;
1987 	}
1988 
1989 	lseg = pnfs_find_lseg(lo, &arg, strict_iomode);
1990 	if (lseg) {
1991 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
1992 				PNFS_UPDATE_LAYOUT_FOUND_CACHED);
1993 		goto out_unlock;
1994 	}
1995 
1996 	/*
1997 	 * Choose a stateid for the LAYOUTGET. If we don't have a layout
1998 	 * stateid, or it has been invalidated, then we must use the open
1999 	 * stateid.
2000 	 */
2001 	if (test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags)) {
2002 		int status;
2003 
2004 		/*
2005 		 * The first layoutget for the file. Need to serialize per
2006 		 * RFC 5661 Errata 3208.
2007 		 */
2008 		if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET,
2009 				     &lo->plh_flags)) {
2010 			spin_unlock(&ino->i_lock);
2011 			lseg = ERR_PTR(wait_on_bit(&lo->plh_flags,
2012 						NFS_LAYOUT_FIRST_LAYOUTGET,
2013 						TASK_KILLABLE));
2014 			if (IS_ERR(lseg))
2015 				goto out_put_layout_hdr;
2016 			pnfs_put_layout_hdr(lo);
2017 			dprintk("%s retrying\n", __func__);
2018 			goto lookup_again;
2019 		}
2020 
2021 		spin_unlock(&ino->i_lock);
2022 		first = true;
2023 		status = nfs4_select_rw_stateid(ctx->state,
2024 					iomode == IOMODE_RW ? FMODE_WRITE : FMODE_READ,
2025 					NULL, &stateid, NULL);
2026 		if (status != 0) {
2027 			lseg = ERR_PTR(status);
2028 			trace_pnfs_update_layout(ino, pos, count,
2029 					iomode, lo, lseg,
2030 					PNFS_UPDATE_LAYOUT_INVALID_OPEN);
2031 			nfs4_schedule_stateid_recovery(server, ctx->state);
2032 			pnfs_clear_first_layoutget(lo);
2033 			pnfs_put_layout_hdr(lo);
2034 			goto lookup_again;
2035 		}
2036 		spin_lock(&ino->i_lock);
2037 	} else {
2038 		nfs4_stateid_copy(&stateid, &lo->plh_stateid);
2039 	}
2040 
2041 	/*
2042 	 * Because we free lsegs before sending LAYOUTRETURN, we need to wait
2043 	 * for LAYOUTRETURN even if first is true.
2044 	 */
2045 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags)) {
2046 		spin_unlock(&ino->i_lock);
2047 		dprintk("%s wait for layoutreturn\n", __func__);
2048 		lseg = ERR_PTR(pnfs_prepare_to_retry_layoutget(lo));
2049 		if (!IS_ERR(lseg)) {
2050 			if (first)
2051 				pnfs_clear_first_layoutget(lo);
2052 			pnfs_put_layout_hdr(lo);
2053 			dprintk("%s retrying\n", __func__);
2054 			trace_pnfs_update_layout(ino, pos, count, iomode, lo,
2055 					lseg, PNFS_UPDATE_LAYOUT_RETRY);
2056 			goto lookup_again;
2057 		}
2058 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2059 				PNFS_UPDATE_LAYOUT_RETURN);
2060 		goto out_put_layout_hdr;
2061 	}
2062 
2063 	if (pnfs_layoutgets_blocked(lo)) {
2064 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2065 				PNFS_UPDATE_LAYOUT_BLOCKED);
2066 		goto out_unlock;
2067 	}
2068 	nfs_layoutget_begin(lo);
2069 	spin_unlock(&ino->i_lock);
2070 
2071 	_add_to_server_list(lo, server);
2072 
2073 	pg_offset = arg.offset & ~PAGE_MASK;
2074 	if (pg_offset) {
2075 		arg.offset -= pg_offset;
2076 		arg.length += pg_offset;
2077 	}
2078 	if (arg.length != NFS4_MAX_UINT64)
2079 		arg.length = PAGE_ALIGN(arg.length);
2080 
2081 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &stateid, &arg, gfp_flags);
2082 	if (!lgp) {
2083 		trace_pnfs_update_layout(ino, pos, count, iomode, lo, NULL,
2084 					 PNFS_UPDATE_LAYOUT_NOMEM);
2085 		nfs_layoutget_end(lo);
2086 		goto out_put_layout_hdr;
2087 	}
2088 
2089 	lseg = nfs4_proc_layoutget(lgp, &timeout);
2090 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2091 				 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);
2092 	nfs_layoutget_end(lo);
2093 	if (IS_ERR(lseg)) {
2094 		switch(PTR_ERR(lseg)) {
2095 		case -EBUSY:
2096 			if (time_after(jiffies, giveup))
2097 				lseg = NULL;
2098 			break;
2099 		case -ERECALLCONFLICT:
2100 		case -EAGAIN:
2101 			break;
2102 		default:
2103 			if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
2104 				pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2105 				lseg = NULL;
2106 			}
2107 			goto out_put_layout_hdr;
2108 		}
2109 		if (lseg) {
2110 			if (first)
2111 				pnfs_clear_first_layoutget(lo);
2112 			trace_pnfs_update_layout(ino, pos, count,
2113 				iomode, lo, lseg, PNFS_UPDATE_LAYOUT_RETRY);
2114 			pnfs_put_layout_hdr(lo);
2115 			goto lookup_again;
2116 		}
2117 	} else {
2118 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2119 	}
2120 
2121 out_put_layout_hdr:
2122 	if (first)
2123 		pnfs_clear_first_layoutget(lo);
2124 	trace_pnfs_update_layout(ino, pos, count, iomode, lo, lseg,
2125 				 PNFS_UPDATE_LAYOUT_EXIT);
2126 	pnfs_put_layout_hdr(lo);
2127 out:
2128 	dprintk("%s: inode %s/%llu pNFS layout segment %s for "
2129 			"(%s, offset: %llu, length: %llu)\n",
2130 			__func__, ino->i_sb->s_id,
2131 			(unsigned long long)NFS_FILEID(ino),
2132 			IS_ERR_OR_NULL(lseg) ? "not found" : "found",
2133 			iomode==IOMODE_RW ?  "read/write" : "read-only",
2134 			(unsigned long long)pos,
2135 			(unsigned long long)count);
2136 	return lseg;
2137 out_unlock:
2138 	spin_unlock(&ino->i_lock);
2139 	goto out_put_layout_hdr;
2140 }
2141 EXPORT_SYMBOL_GPL(pnfs_update_layout);
2142 
2143 static bool
2144 pnfs_sanity_check_layout_range(struct pnfs_layout_range *range)
2145 {
2146 	switch (range->iomode) {
2147 	case IOMODE_READ:
2148 	case IOMODE_RW:
2149 		break;
2150 	default:
2151 		return false;
2152 	}
2153 	if (range->offset == NFS4_MAX_UINT64)
2154 		return false;
2155 	if (range->length == 0)
2156 		return false;
2157 	if (range->length != NFS4_MAX_UINT64 &&
2158 	    range->length > NFS4_MAX_UINT64 - range->offset)
2159 		return false;
2160 	return true;
2161 }
2162 
2163 static struct pnfs_layout_hdr *
2164 _pnfs_grab_empty_layout(struct inode *ino, struct nfs_open_context *ctx)
2165 {
2166 	struct pnfs_layout_hdr *lo;
2167 
2168 	spin_lock(&ino->i_lock);
2169 	lo = pnfs_find_alloc_layout(ino, ctx, GFP_KERNEL);
2170 	if (!lo)
2171 		goto out_unlock;
2172 	if (!test_bit(NFS_LAYOUT_INVALID_STID, &lo->plh_flags))
2173 		goto out_unlock;
2174 	if (test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
2175 		goto out_unlock;
2176 	if (pnfs_layoutgets_blocked(lo))
2177 		goto out_unlock;
2178 	if (test_and_set_bit(NFS_LAYOUT_FIRST_LAYOUTGET, &lo->plh_flags))
2179 		goto out_unlock;
2180 	nfs_layoutget_begin(lo);
2181 	spin_unlock(&ino->i_lock);
2182 	_add_to_server_list(lo, NFS_SERVER(ino));
2183 	return lo;
2184 
2185 out_unlock:
2186 	spin_unlock(&ino->i_lock);
2187 	pnfs_put_layout_hdr(lo);
2188 	return NULL;
2189 }
2190 
2191 static void _lgopen_prepare_attached(struct nfs4_opendata *data,
2192 				     struct nfs_open_context *ctx)
2193 {
2194 	struct inode *ino = data->dentry->d_inode;
2195 	struct pnfs_layout_range rng = {
2196 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2197 			  IOMODE_RW: IOMODE_READ,
2198 		.offset = 0,
2199 		.length = NFS4_MAX_UINT64,
2200 	};
2201 	struct nfs4_layoutget *lgp;
2202 	struct pnfs_layout_hdr *lo;
2203 
2204 	/* Heuristic: don't send layoutget if we have cached data */
2205 	if (rng.iomode == IOMODE_READ &&
2206 	   (i_size_read(ino) == 0 || ino->i_mapping->nrpages != 0))
2207 		return;
2208 
2209 	lo = _pnfs_grab_empty_layout(ino, ctx);
2210 	if (!lo)
2211 		return;
2212 	lgp = pnfs_alloc_init_layoutget_args(ino, ctx, &current_stateid,
2213 					     &rng, GFP_KERNEL);
2214 	if (!lgp) {
2215 		pnfs_clear_first_layoutget(lo);
2216 		pnfs_put_layout_hdr(lo);
2217 		return;
2218 	}
2219 	data->lgp = lgp;
2220 	data->o_arg.lg_args = &lgp->args;
2221 	data->o_res.lg_res = &lgp->res;
2222 }
2223 
2224 static void _lgopen_prepare_floating(struct nfs4_opendata *data,
2225 				     struct nfs_open_context *ctx)
2226 {
2227 	struct pnfs_layout_range rng = {
2228 		.iomode = (data->o_arg.fmode & FMODE_WRITE) ?
2229 			  IOMODE_RW: IOMODE_READ,
2230 		.offset = 0,
2231 		.length = NFS4_MAX_UINT64,
2232 	};
2233 	struct nfs4_layoutget *lgp;
2234 
2235 	lgp = pnfs_alloc_init_layoutget_args(NULL, ctx, &current_stateid,
2236 					     &rng, GFP_KERNEL);
2237 	if (!lgp)
2238 		return;
2239 	data->lgp = lgp;
2240 	data->o_arg.lg_args = &lgp->args;
2241 	data->o_res.lg_res = &lgp->res;
2242 }
2243 
2244 void pnfs_lgopen_prepare(struct nfs4_opendata *data,
2245 			 struct nfs_open_context *ctx)
2246 {
2247 	struct nfs_server *server = NFS_SERVER(data->dir->d_inode);
2248 
2249 	if (!(pnfs_enabled_sb(server) &&
2250 	      server->pnfs_curr_ld->flags & PNFS_LAYOUTGET_ON_OPEN))
2251 		return;
2252 	/* Could check on max_ops, but currently hardcoded high enough */
2253 	if (!nfs_server_capable(data->dir->d_inode, NFS_CAP_LGOPEN))
2254 		return;
2255 	if (data->state)
2256 		_lgopen_prepare_attached(data, ctx);
2257 	else
2258 		_lgopen_prepare_floating(data, ctx);
2259 }
2260 
2261 void pnfs_parse_lgopen(struct inode *ino, struct nfs4_layoutget *lgp,
2262 		       struct nfs_open_context *ctx)
2263 {
2264 	struct pnfs_layout_hdr *lo;
2265 	struct pnfs_layout_segment *lseg;
2266 	struct nfs_server *srv = NFS_SERVER(ino);
2267 	u32 iomode;
2268 
2269 	if (!lgp)
2270 		return;
2271 	dprintk("%s: entered with status %i\n", __func__, lgp->res.status);
2272 	if (lgp->res.status) {
2273 		switch (lgp->res.status) {
2274 		default:
2275 			break;
2276 		/*
2277 		 * Halt lgopen attempts if the server doesn't recognise
2278 		 * the "current stateid" value, the layout type, or the
2279 		 * layoutget operation as being valid.
2280 		 * Also if it complains about too many ops in the compound
2281 		 * or of the request/reply being too big.
2282 		 */
2283 		case -NFS4ERR_BAD_STATEID:
2284 		case -NFS4ERR_NOTSUPP:
2285 		case -NFS4ERR_REP_TOO_BIG:
2286 		case -NFS4ERR_REP_TOO_BIG_TO_CACHE:
2287 		case -NFS4ERR_REQ_TOO_BIG:
2288 		case -NFS4ERR_TOO_MANY_OPS:
2289 		case -NFS4ERR_UNKNOWN_LAYOUTTYPE:
2290 			srv->caps &= ~NFS_CAP_LGOPEN;
2291 		}
2292 		return;
2293 	}
2294 	if (!lgp->args.inode) {
2295 		lo = _pnfs_grab_empty_layout(ino, ctx);
2296 		if (!lo)
2297 			return;
2298 		lgp->args.inode = ino;
2299 	} else
2300 		lo = NFS_I(lgp->args.inode)->layout;
2301 
2302 	lseg = pnfs_layout_process(lgp);
2303 	if (!IS_ERR(lseg)) {
2304 		iomode = lgp->args.range.iomode;
2305 		pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
2306 		pnfs_put_lseg(lseg);
2307 	}
2308 }
2309 
2310 void nfs4_lgopen_release(struct nfs4_layoutget *lgp)
2311 {
2312 	if (lgp != NULL) {
2313 		struct inode *inode = lgp->args.inode;
2314 		if (inode) {
2315 			struct pnfs_layout_hdr *lo = NFS_I(inode)->layout;
2316 			pnfs_clear_first_layoutget(lo);
2317 			nfs_layoutget_end(lo);
2318 		}
2319 		pnfs_layoutget_free(lgp);
2320 	}
2321 }
2322 
2323 struct pnfs_layout_segment *
2324 pnfs_layout_process(struct nfs4_layoutget *lgp)
2325 {
2326 	struct pnfs_layout_hdr *lo = NFS_I(lgp->args.inode)->layout;
2327 	struct nfs4_layoutget_res *res = &lgp->res;
2328 	struct pnfs_layout_segment *lseg;
2329 	struct inode *ino = lo->plh_inode;
2330 	LIST_HEAD(free_me);
2331 
2332 	if (!pnfs_sanity_check_layout_range(&res->range))
2333 		return ERR_PTR(-EINVAL);
2334 
2335 	/* Inject layout blob into I/O device driver */
2336 	lseg = NFS_SERVER(ino)->pnfs_curr_ld->alloc_lseg(lo, res, lgp->gfp_flags);
2337 	if (IS_ERR_OR_NULL(lseg)) {
2338 		if (!lseg)
2339 			lseg = ERR_PTR(-ENOMEM);
2340 
2341 		dprintk("%s: Could not allocate layout: error %ld\n",
2342 		       __func__, PTR_ERR(lseg));
2343 		return lseg;
2344 	}
2345 
2346 	pnfs_init_lseg(lo, lseg, &res->range, &res->stateid);
2347 
2348 	spin_lock(&ino->i_lock);
2349 	if (pnfs_layoutgets_blocked(lo)) {
2350 		dprintk("%s forget reply due to state\n", __func__);
2351 		goto out_forget;
2352 	}
2353 
2354 	if (!pnfs_layout_is_valid(lo)) {
2355 		/* We have a completely new layout */
2356 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, true);
2357 	} else if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
2358 		/* existing state ID, make sure the sequence number matches. */
2359 		if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
2360 			dprintk("%s forget reply due to sequence\n", __func__);
2361 			goto out_forget;
2362 		}
2363 		pnfs_set_layout_stateid(lo, &res->stateid, lgp->cred, false);
2364 	} else {
2365 		/*
2366 		 * We got an entirely new state ID.  Mark all segments for the
2367 		 * inode invalid, and retry the layoutget
2368 		 */
2369 		pnfs_mark_layout_stateid_invalid(lo, &free_me);
2370 		goto out_forget;
2371 	}
2372 
2373 	pnfs_get_lseg(lseg);
2374 	pnfs_layout_insert_lseg(lo, lseg, &free_me);
2375 
2376 
2377 	if (res->return_on_close)
2378 		set_bit(NFS_LSEG_ROC, &lseg->pls_flags);
2379 
2380 	spin_unlock(&ino->i_lock);
2381 	pnfs_free_lseg_list(&free_me);
2382 	return lseg;
2383 
2384 out_forget:
2385 	spin_unlock(&ino->i_lock);
2386 	lseg->pls_layout = lo;
2387 	NFS_SERVER(ino)->pnfs_curr_ld->free_lseg(lseg);
2388 	return ERR_PTR(-EAGAIN);
2389 }
2390 
2391 /**
2392  * pnfs_mark_matching_lsegs_return - Free or return matching layout segments
2393  * @lo: pointer to layout header
2394  * @tmp_list: list header to be used with pnfs_free_lseg_list()
2395  * @return_range: describe layout segment ranges to be returned
2396  * @seq: stateid seqid to match
2397  *
2398  * This function is mainly intended for use by layoutrecall. It attempts
2399  * to free the layout segment immediately, or else to mark it for return
2400  * as soon as its reference count drops to zero.
2401  *
2402  * Returns
2403  * - 0: a layoutreturn needs to be scheduled.
2404  * - EBUSY: there are layout segment that are still in use.
2405  * - ENOENT: there are no layout segments that need to be returned.
2406  */
2407 int
2408 pnfs_mark_matching_lsegs_return(struct pnfs_layout_hdr *lo,
2409 				struct list_head *tmp_list,
2410 				const struct pnfs_layout_range *return_range,
2411 				u32 seq)
2412 {
2413 	struct pnfs_layout_segment *lseg, *next;
2414 	int remaining = 0;
2415 
2416 	dprintk("%s:Begin lo %p\n", __func__, lo);
2417 
2418 	assert_spin_locked(&lo->plh_inode->i_lock);
2419 
2420 	list_for_each_entry_safe(lseg, next, &lo->plh_segs, pls_list)
2421 		if (pnfs_match_lseg_recall(lseg, return_range, seq)) {
2422 			dprintk("%s: marking lseg %p iomode %d "
2423 				"offset %llu length %llu\n", __func__,
2424 				lseg, lseg->pls_range.iomode,
2425 				lseg->pls_range.offset,
2426 				lseg->pls_range.length);
2427 			if (mark_lseg_invalid(lseg, tmp_list))
2428 				continue;
2429 			remaining++;
2430 			set_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags);
2431 		}
2432 
2433 	if (remaining) {
2434 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2435 		return -EBUSY;
2436 	}
2437 
2438 	if (!list_empty(&lo->plh_return_segs)) {
2439 		pnfs_set_plh_return_info(lo, return_range->iomode, seq);
2440 		return 0;
2441 	}
2442 
2443 	return -ENOENT;
2444 }
2445 
2446 static void
2447 pnfs_mark_layout_for_return(struct inode *inode,
2448 			    const struct pnfs_layout_range *range)
2449 {
2450 	struct pnfs_layout_hdr *lo;
2451 	bool return_now = false;
2452 
2453 	spin_lock(&inode->i_lock);
2454 	lo = NFS_I(inode)->layout;
2455 	if (!pnfs_layout_is_valid(lo)) {
2456 		spin_unlock(&inode->i_lock);
2457 		return;
2458 	}
2459 	pnfs_set_plh_return_info(lo, range->iomode, 0);
2460 	/*
2461 	 * mark all matching lsegs so that we are sure to have no live
2462 	 * segments at hand when sending layoutreturn. See pnfs_put_lseg()
2463 	 * for how it works.
2464 	 */
2465 	if (pnfs_mark_matching_lsegs_return(lo, &lo->plh_return_segs, range, 0) != -EBUSY) {
2466 		const struct cred *cred;
2467 		nfs4_stateid stateid;
2468 		enum pnfs_iomode iomode;
2469 
2470 		return_now = pnfs_prepare_layoutreturn(lo, &stateid, &cred, &iomode);
2471 		spin_unlock(&inode->i_lock);
2472 		if (return_now)
2473 			pnfs_send_layoutreturn(lo, &stateid, &cred, iomode, false);
2474 	} else {
2475 		spin_unlock(&inode->i_lock);
2476 		nfs_commit_inode(inode, 0);
2477 	}
2478 }
2479 
2480 void pnfs_error_mark_layout_for_return(struct inode *inode,
2481 				       struct pnfs_layout_segment *lseg)
2482 {
2483 	struct pnfs_layout_range range = {
2484 		.iomode = lseg->pls_range.iomode,
2485 		.offset = 0,
2486 		.length = NFS4_MAX_UINT64,
2487 	};
2488 
2489 	pnfs_mark_layout_for_return(inode, &range);
2490 }
2491 EXPORT_SYMBOL_GPL(pnfs_error_mark_layout_for_return);
2492 
2493 static bool
2494 pnfs_layout_can_be_returned(struct pnfs_layout_hdr *lo)
2495 {
2496 	return pnfs_layout_is_valid(lo) &&
2497 		!test_bit(NFS_LAYOUT_INODE_FREEING, &lo->plh_flags) &&
2498 		!test_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
2499 }
2500 
2501 static struct pnfs_layout_segment *
2502 pnfs_find_first_lseg(struct pnfs_layout_hdr *lo,
2503 		     const struct pnfs_layout_range *range,
2504 		     enum pnfs_iomode iomode)
2505 {
2506 	struct pnfs_layout_segment *lseg;
2507 
2508 	list_for_each_entry(lseg, &lo->plh_segs, pls_list) {
2509 		if (!test_bit(NFS_LSEG_VALID, &lseg->pls_flags))
2510 			continue;
2511 		if (test_bit(NFS_LSEG_LAYOUTRETURN, &lseg->pls_flags))
2512 			continue;
2513 		if (lseg->pls_range.iomode != iomode && iomode != IOMODE_ANY)
2514 			continue;
2515 		if (pnfs_lseg_range_intersecting(&lseg->pls_range, range))
2516 			return lseg;
2517 	}
2518 	return NULL;
2519 }
2520 
2521 /* Find open file states whose mode matches that of the range */
2522 static bool
2523 pnfs_should_return_unused_layout(struct pnfs_layout_hdr *lo,
2524 				 const struct pnfs_layout_range *range)
2525 {
2526 	struct list_head *head;
2527 	struct nfs_open_context *ctx;
2528 	fmode_t mode = 0;
2529 
2530 	if (!pnfs_layout_can_be_returned(lo) ||
2531 	    !pnfs_find_first_lseg(lo, range, range->iomode))
2532 		return false;
2533 
2534 	head = &NFS_I(lo->plh_inode)->open_files;
2535 	list_for_each_entry_rcu(ctx, head, list) {
2536 		if (ctx->state)
2537 			mode |= ctx->state->state & (FMODE_READ|FMODE_WRITE);
2538 	}
2539 
2540 	switch (range->iomode) {
2541 	default:
2542 		break;
2543 	case IOMODE_READ:
2544 		mode &= ~FMODE_WRITE;
2545 		break;
2546 	case IOMODE_RW:
2547 		if (pnfs_find_first_lseg(lo, range, IOMODE_READ))
2548 			mode &= ~FMODE_READ;
2549 	}
2550 	return mode == 0;
2551 }
2552 
2553 static int
2554 pnfs_layout_return_unused_byserver(struct nfs_server *server, void *data)
2555 {
2556 	const struct pnfs_layout_range *range = data;
2557 	struct pnfs_layout_hdr *lo;
2558 	struct inode *inode;
2559 restart:
2560 	rcu_read_lock();
2561 	list_for_each_entry_rcu(lo, &server->layouts, plh_layouts) {
2562 		if (!pnfs_layout_can_be_returned(lo) ||
2563 		    test_bit(NFS_LAYOUT_RETURN_REQUESTED, &lo->plh_flags))
2564 			continue;
2565 		inode = lo->plh_inode;
2566 		spin_lock(&inode->i_lock);
2567 		if (!pnfs_should_return_unused_layout(lo, range)) {
2568 			spin_unlock(&inode->i_lock);
2569 			continue;
2570 		}
2571 		spin_unlock(&inode->i_lock);
2572 		inode = pnfs_grab_inode_layout_hdr(lo);
2573 		if (!inode)
2574 			continue;
2575 		rcu_read_unlock();
2576 		pnfs_mark_layout_for_return(inode, range);
2577 		iput(inode);
2578 		cond_resched();
2579 		goto restart;
2580 	}
2581 	rcu_read_unlock();
2582 	return 0;
2583 }
2584 
2585 void
2586 pnfs_layout_return_unused_byclid(struct nfs_client *clp,
2587 				 enum pnfs_iomode iomode)
2588 {
2589 	struct pnfs_layout_range range = {
2590 		.iomode = iomode,
2591 		.offset = 0,
2592 		.length = NFS4_MAX_UINT64,
2593 	};
2594 
2595 	nfs_client_for_each_server(clp, pnfs_layout_return_unused_byserver,
2596 			&range);
2597 }
2598 
2599 void
2600 pnfs_generic_pg_check_layout(struct nfs_pageio_descriptor *pgio)
2601 {
2602 	if (pgio->pg_lseg == NULL ||
2603 	    test_bit(NFS_LSEG_VALID, &pgio->pg_lseg->pls_flags))
2604 		return;
2605 	pnfs_put_lseg(pgio->pg_lseg);
2606 	pgio->pg_lseg = NULL;
2607 }
2608 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_layout);
2609 
2610 /*
2611  * Check for any intersection between the request and the pgio->pg_lseg,
2612  * and if none, put this pgio->pg_lseg away.
2613  */
2614 void
2615 pnfs_generic_pg_check_range(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2616 {
2617 	if (pgio->pg_lseg && !pnfs_lseg_request_intersecting(pgio->pg_lseg, req)) {
2618 		pnfs_put_lseg(pgio->pg_lseg);
2619 		pgio->pg_lseg = NULL;
2620 	}
2621 }
2622 EXPORT_SYMBOL_GPL(pnfs_generic_pg_check_range);
2623 
2624 void
2625 pnfs_generic_pg_init_read(struct nfs_pageio_descriptor *pgio, struct nfs_page *req)
2626 {
2627 	u64 rd_size = req->wb_bytes;
2628 
2629 	pnfs_generic_pg_check_layout(pgio);
2630 	pnfs_generic_pg_check_range(pgio, req);
2631 	if (pgio->pg_lseg == NULL) {
2632 		if (pgio->pg_dreq == NULL)
2633 			rd_size = i_size_read(pgio->pg_inode) - req_offset(req);
2634 		else
2635 			rd_size = nfs_dreq_bytes_left(pgio->pg_dreq);
2636 
2637 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2638 						   nfs_req_openctx(req),
2639 						   req_offset(req),
2640 						   rd_size,
2641 						   IOMODE_READ,
2642 						   false,
2643 						   GFP_KERNEL);
2644 		if (IS_ERR(pgio->pg_lseg)) {
2645 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2646 			pgio->pg_lseg = NULL;
2647 			return;
2648 		}
2649 	}
2650 	/* If no lseg, fall back to read through mds */
2651 	if (pgio->pg_lseg == NULL)
2652 		nfs_pageio_reset_read_mds(pgio);
2653 
2654 }
2655 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_read);
2656 
2657 void
2658 pnfs_generic_pg_init_write(struct nfs_pageio_descriptor *pgio,
2659 			   struct nfs_page *req, u64 wb_size)
2660 {
2661 	pnfs_generic_pg_check_layout(pgio);
2662 	pnfs_generic_pg_check_range(pgio, req);
2663 	if (pgio->pg_lseg == NULL) {
2664 		pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
2665 						   nfs_req_openctx(req),
2666 						   req_offset(req),
2667 						   wb_size,
2668 						   IOMODE_RW,
2669 						   false,
2670 						   GFP_KERNEL);
2671 		if (IS_ERR(pgio->pg_lseg)) {
2672 			pgio->pg_error = PTR_ERR(pgio->pg_lseg);
2673 			pgio->pg_lseg = NULL;
2674 			return;
2675 		}
2676 	}
2677 	/* If no lseg, fall back to write through mds */
2678 	if (pgio->pg_lseg == NULL)
2679 		nfs_pageio_reset_write_mds(pgio);
2680 }
2681 EXPORT_SYMBOL_GPL(pnfs_generic_pg_init_write);
2682 
2683 void
2684 pnfs_generic_pg_cleanup(struct nfs_pageio_descriptor *desc)
2685 {
2686 	if (desc->pg_lseg) {
2687 		pnfs_put_lseg(desc->pg_lseg);
2688 		desc->pg_lseg = NULL;
2689 	}
2690 }
2691 EXPORT_SYMBOL_GPL(pnfs_generic_pg_cleanup);
2692 
2693 /*
2694  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
2695  * of bytes (maximum @req->wb_bytes) that can be coalesced.
2696  */
2697 size_t
2698 pnfs_generic_pg_test(struct nfs_pageio_descriptor *pgio,
2699 		     struct nfs_page *prev, struct nfs_page *req)
2700 {
2701 	unsigned int size;
2702 	u64 seg_end, req_start, seg_left;
2703 
2704 	size = nfs_generic_pg_test(pgio, prev, req);
2705 	if (!size)
2706 		return 0;
2707 
2708 	/*
2709 	 * 'size' contains the number of bytes left in the current page (up
2710 	 * to the original size asked for in @req->wb_bytes).
2711 	 *
2712 	 * Calculate how many bytes are left in the layout segment
2713 	 * and if there are less bytes than 'size', return that instead.
2714 	 *
2715 	 * Please also note that 'end_offset' is actually the offset of the
2716 	 * first byte that lies outside the pnfs_layout_range. FIXME?
2717 	 *
2718 	 */
2719 	if (pgio->pg_lseg) {
2720 		seg_end = pnfs_end_offset(pgio->pg_lseg->pls_range.offset,
2721 				     pgio->pg_lseg->pls_range.length);
2722 		req_start = req_offset(req);
2723 
2724 		/* start of request is past the last byte of this segment */
2725 		if (req_start >= seg_end)
2726 			return 0;
2727 
2728 		/* adjust 'size' iff there are fewer bytes left in the
2729 		 * segment than what nfs_generic_pg_test returned */
2730 		seg_left = seg_end - req_start;
2731 		if (seg_left < size)
2732 			size = (unsigned int)seg_left;
2733 	}
2734 
2735 	return size;
2736 }
2737 EXPORT_SYMBOL_GPL(pnfs_generic_pg_test);
2738 
2739 int pnfs_write_done_resend_to_mds(struct nfs_pgio_header *hdr)
2740 {
2741 	struct nfs_pageio_descriptor pgio;
2742 
2743 	/* Resend all requests through the MDS */
2744 	nfs_pageio_init_write(&pgio, hdr->inode, FLUSH_STABLE, true,
2745 			      hdr->completion_ops);
2746 	set_bit(NFS_CONTEXT_RESEND_WRITES, &hdr->args.context->flags);
2747 	return nfs_pageio_resend(&pgio, hdr);
2748 }
2749 EXPORT_SYMBOL_GPL(pnfs_write_done_resend_to_mds);
2750 
2751 static void pnfs_ld_handle_write_error(struct nfs_pgio_header *hdr)
2752 {
2753 
2754 	dprintk("pnfs write error = %d\n", hdr->pnfs_error);
2755 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2756 	    PNFS_LAYOUTRET_ON_ERROR) {
2757 		pnfs_return_layout(hdr->inode);
2758 	}
2759 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2760 		hdr->task.tk_status = pnfs_write_done_resend_to_mds(hdr);
2761 }
2762 
2763 /*
2764  * Called by non rpc-based layout drivers
2765  */
2766 void pnfs_ld_write_done(struct nfs_pgio_header *hdr)
2767 {
2768 	if (likely(!hdr->pnfs_error)) {
2769 		pnfs_set_layoutcommit(hdr->inode, hdr->lseg,
2770 				hdr->mds_offset + hdr->res.count);
2771 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2772 	}
2773 	trace_nfs4_pnfs_write(hdr, hdr->pnfs_error);
2774 	if (unlikely(hdr->pnfs_error))
2775 		pnfs_ld_handle_write_error(hdr);
2776 	hdr->mds_ops->rpc_release(hdr);
2777 }
2778 EXPORT_SYMBOL_GPL(pnfs_ld_write_done);
2779 
2780 static void
2781 pnfs_write_through_mds(struct nfs_pageio_descriptor *desc,
2782 		struct nfs_pgio_header *hdr)
2783 {
2784 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2785 
2786 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2787 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2788 		nfs_pageio_reset_write_mds(desc);
2789 		mirror->pg_recoalesce = 1;
2790 	}
2791 	hdr->completion_ops->completion(hdr);
2792 }
2793 
2794 static enum pnfs_try_status
2795 pnfs_try_to_write_data(struct nfs_pgio_header *hdr,
2796 			const struct rpc_call_ops *call_ops,
2797 			struct pnfs_layout_segment *lseg,
2798 			int how)
2799 {
2800 	struct inode *inode = hdr->inode;
2801 	enum pnfs_try_status trypnfs;
2802 	struct nfs_server *nfss = NFS_SERVER(inode);
2803 
2804 	hdr->mds_ops = call_ops;
2805 
2806 	dprintk("%s: Writing ino:%lu %u@%llu (how %d)\n", __func__,
2807 		inode->i_ino, hdr->args.count, hdr->args.offset, how);
2808 	trypnfs = nfss->pnfs_curr_ld->write_pagelist(hdr, how);
2809 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2810 		nfs_inc_stats(inode, NFSIOS_PNFS_WRITE);
2811 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2812 	return trypnfs;
2813 }
2814 
2815 static void
2816 pnfs_do_write(struct nfs_pageio_descriptor *desc,
2817 	      struct nfs_pgio_header *hdr, int how)
2818 {
2819 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2820 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2821 	enum pnfs_try_status trypnfs;
2822 
2823 	trypnfs = pnfs_try_to_write_data(hdr, call_ops, lseg, how);
2824 	switch (trypnfs) {
2825 	case PNFS_NOT_ATTEMPTED:
2826 		pnfs_write_through_mds(desc, hdr);
2827 	case PNFS_ATTEMPTED:
2828 		break;
2829 	case PNFS_TRY_AGAIN:
2830 		/* cleanup hdr and prepare to redo pnfs */
2831 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2832 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2833 			list_splice_init(&hdr->pages, &mirror->pg_list);
2834 			mirror->pg_recoalesce = 1;
2835 		}
2836 		hdr->mds_ops->rpc_release(hdr);
2837 	}
2838 }
2839 
2840 static void pnfs_writehdr_free(struct nfs_pgio_header *hdr)
2841 {
2842 	pnfs_put_lseg(hdr->lseg);
2843 	nfs_pgio_header_free(hdr);
2844 }
2845 
2846 int
2847 pnfs_generic_pg_writepages(struct nfs_pageio_descriptor *desc)
2848 {
2849 	struct nfs_pgio_header *hdr;
2850 	int ret;
2851 
2852 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2853 	if (!hdr) {
2854 		desc->pg_error = -ENOMEM;
2855 		return desc->pg_error;
2856 	}
2857 	nfs_pgheader_init(desc, hdr, pnfs_writehdr_free);
2858 
2859 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
2860 	ret = nfs_generic_pgio(desc, hdr);
2861 	if (!ret)
2862 		pnfs_do_write(desc, hdr, desc->pg_ioflags);
2863 
2864 	return ret;
2865 }
2866 EXPORT_SYMBOL_GPL(pnfs_generic_pg_writepages);
2867 
2868 int pnfs_read_done_resend_to_mds(struct nfs_pgio_header *hdr)
2869 {
2870 	struct nfs_pageio_descriptor pgio;
2871 
2872 	/* Resend all requests through the MDS */
2873 	nfs_pageio_init_read(&pgio, hdr->inode, true, hdr->completion_ops);
2874 	return nfs_pageio_resend(&pgio, hdr);
2875 }
2876 EXPORT_SYMBOL_GPL(pnfs_read_done_resend_to_mds);
2877 
2878 static void pnfs_ld_handle_read_error(struct nfs_pgio_header *hdr)
2879 {
2880 	dprintk("pnfs read error = %d\n", hdr->pnfs_error);
2881 	if (NFS_SERVER(hdr->inode)->pnfs_curr_ld->flags &
2882 	    PNFS_LAYOUTRET_ON_ERROR) {
2883 		pnfs_return_layout(hdr->inode);
2884 	}
2885 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags))
2886 		hdr->task.tk_status = pnfs_read_done_resend_to_mds(hdr);
2887 }
2888 
2889 /*
2890  * Called by non rpc-based layout drivers
2891  */
2892 void pnfs_ld_read_done(struct nfs_pgio_header *hdr)
2893 {
2894 	if (likely(!hdr->pnfs_error))
2895 		hdr->mds_ops->rpc_call_done(&hdr->task, hdr);
2896 	trace_nfs4_pnfs_read(hdr, hdr->pnfs_error);
2897 	if (unlikely(hdr->pnfs_error))
2898 		pnfs_ld_handle_read_error(hdr);
2899 	hdr->mds_ops->rpc_release(hdr);
2900 }
2901 EXPORT_SYMBOL_GPL(pnfs_ld_read_done);
2902 
2903 static void
2904 pnfs_read_through_mds(struct nfs_pageio_descriptor *desc,
2905 		struct nfs_pgio_header *hdr)
2906 {
2907 	struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2908 
2909 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2910 		list_splice_tail_init(&hdr->pages, &mirror->pg_list);
2911 		nfs_pageio_reset_read_mds(desc);
2912 		mirror->pg_recoalesce = 1;
2913 	}
2914 	hdr->completion_ops->completion(hdr);
2915 }
2916 
2917 /*
2918  * Call the appropriate parallel I/O subsystem read function.
2919  */
2920 static enum pnfs_try_status
2921 pnfs_try_to_read_data(struct nfs_pgio_header *hdr,
2922 		       const struct rpc_call_ops *call_ops,
2923 		       struct pnfs_layout_segment *lseg)
2924 {
2925 	struct inode *inode = hdr->inode;
2926 	struct nfs_server *nfss = NFS_SERVER(inode);
2927 	enum pnfs_try_status trypnfs;
2928 
2929 	hdr->mds_ops = call_ops;
2930 
2931 	dprintk("%s: Reading ino:%lu %u@%llu\n",
2932 		__func__, inode->i_ino, hdr->args.count, hdr->args.offset);
2933 
2934 	trypnfs = nfss->pnfs_curr_ld->read_pagelist(hdr);
2935 	if (trypnfs != PNFS_NOT_ATTEMPTED)
2936 		nfs_inc_stats(inode, NFSIOS_PNFS_READ);
2937 	dprintk("%s End (trypnfs:%d)\n", __func__, trypnfs);
2938 	return trypnfs;
2939 }
2940 
2941 /* Resend all requests through pnfs. */
2942 void pnfs_read_resend_pnfs(struct nfs_pgio_header *hdr,
2943 			   unsigned int mirror_idx)
2944 {
2945 	struct nfs_pageio_descriptor pgio;
2946 
2947 	if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2948 		/* Prevent deadlocks with layoutreturn! */
2949 		pnfs_put_lseg(hdr->lseg);
2950 		hdr->lseg = NULL;
2951 
2952 		nfs_pageio_init_read(&pgio, hdr->inode, false,
2953 					hdr->completion_ops);
2954 		pgio.pg_mirror_idx = mirror_idx;
2955 		hdr->task.tk_status = nfs_pageio_resend(&pgio, hdr);
2956 	}
2957 }
2958 EXPORT_SYMBOL_GPL(pnfs_read_resend_pnfs);
2959 
2960 static void
2961 pnfs_do_read(struct nfs_pageio_descriptor *desc, struct nfs_pgio_header *hdr)
2962 {
2963 	const struct rpc_call_ops *call_ops = desc->pg_rpc_callops;
2964 	struct pnfs_layout_segment *lseg = desc->pg_lseg;
2965 	enum pnfs_try_status trypnfs;
2966 
2967 	trypnfs = pnfs_try_to_read_data(hdr, call_ops, lseg);
2968 	switch (trypnfs) {
2969 	case PNFS_NOT_ATTEMPTED:
2970 		pnfs_read_through_mds(desc, hdr);
2971 	case PNFS_ATTEMPTED:
2972 		break;
2973 	case PNFS_TRY_AGAIN:
2974 		/* cleanup hdr and prepare to redo pnfs */
2975 		if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
2976 			struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
2977 			list_splice_init(&hdr->pages, &mirror->pg_list);
2978 			mirror->pg_recoalesce = 1;
2979 		}
2980 		hdr->mds_ops->rpc_release(hdr);
2981 	}
2982 }
2983 
2984 static void pnfs_readhdr_free(struct nfs_pgio_header *hdr)
2985 {
2986 	pnfs_put_lseg(hdr->lseg);
2987 	nfs_pgio_header_free(hdr);
2988 }
2989 
2990 int
2991 pnfs_generic_pg_readpages(struct nfs_pageio_descriptor *desc)
2992 {
2993 	struct nfs_pgio_header *hdr;
2994 	int ret;
2995 
2996 	hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
2997 	if (!hdr) {
2998 		desc->pg_error = -ENOMEM;
2999 		return desc->pg_error;
3000 	}
3001 	nfs_pgheader_init(desc, hdr, pnfs_readhdr_free);
3002 	hdr->lseg = pnfs_get_lseg(desc->pg_lseg);
3003 	ret = nfs_generic_pgio(desc, hdr);
3004 	if (!ret)
3005 		pnfs_do_read(desc, hdr);
3006 	return ret;
3007 }
3008 EXPORT_SYMBOL_GPL(pnfs_generic_pg_readpages);
3009 
3010 static void pnfs_clear_layoutcommitting(struct inode *inode)
3011 {
3012 	unsigned long *bitlock = &NFS_I(inode)->flags;
3013 
3014 	clear_bit_unlock(NFS_INO_LAYOUTCOMMITTING, bitlock);
3015 	smp_mb__after_atomic();
3016 	wake_up_bit(bitlock, NFS_INO_LAYOUTCOMMITTING);
3017 }
3018 
3019 /*
3020  * There can be multiple RW segments.
3021  */
3022 static void pnfs_list_write_lseg(struct inode *inode, struct list_head *listp)
3023 {
3024 	struct pnfs_layout_segment *lseg;
3025 
3026 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
3027 		if (lseg->pls_range.iomode == IOMODE_RW &&
3028 		    test_and_clear_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags))
3029 			list_add(&lseg->pls_lc_list, listp);
3030 	}
3031 }
3032 
3033 static void pnfs_list_write_lseg_done(struct inode *inode, struct list_head *listp)
3034 {
3035 	struct pnfs_layout_segment *lseg, *tmp;
3036 
3037 	/* Matched by references in pnfs_set_layoutcommit */
3038 	list_for_each_entry_safe(lseg, tmp, listp, pls_lc_list) {
3039 		list_del_init(&lseg->pls_lc_list);
3040 		pnfs_put_lseg(lseg);
3041 	}
3042 
3043 	pnfs_clear_layoutcommitting(inode);
3044 }
3045 
3046 void pnfs_set_lo_fail(struct pnfs_layout_segment *lseg)
3047 {
3048 	pnfs_layout_io_set_failed(lseg->pls_layout, lseg->pls_range.iomode);
3049 }
3050 EXPORT_SYMBOL_GPL(pnfs_set_lo_fail);
3051 
3052 void
3053 pnfs_set_layoutcommit(struct inode *inode, struct pnfs_layout_segment *lseg,
3054 		loff_t end_pos)
3055 {
3056 	struct nfs_inode *nfsi = NFS_I(inode);
3057 	bool mark_as_dirty = false;
3058 
3059 	spin_lock(&inode->i_lock);
3060 	if (!test_and_set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags)) {
3061 		nfsi->layout->plh_lwb = end_pos;
3062 		mark_as_dirty = true;
3063 		dprintk("%s: Set layoutcommit for inode %lu ",
3064 			__func__, inode->i_ino);
3065 	} else if (end_pos > nfsi->layout->plh_lwb)
3066 		nfsi->layout->plh_lwb = end_pos;
3067 	if (!test_and_set_bit(NFS_LSEG_LAYOUTCOMMIT, &lseg->pls_flags)) {
3068 		/* references matched in nfs4_layoutcommit_release */
3069 		pnfs_get_lseg(lseg);
3070 	}
3071 	spin_unlock(&inode->i_lock);
3072 	dprintk("%s: lseg %p end_pos %llu\n",
3073 		__func__, lseg, nfsi->layout->plh_lwb);
3074 
3075 	/* if pnfs_layoutcommit_inode() runs between inode locks, the next one
3076 	 * will be a noop because NFS_INO_LAYOUTCOMMIT will not be set */
3077 	if (mark_as_dirty)
3078 		mark_inode_dirty_sync(inode);
3079 }
3080 EXPORT_SYMBOL_GPL(pnfs_set_layoutcommit);
3081 
3082 void pnfs_cleanup_layoutcommit(struct nfs4_layoutcommit_data *data)
3083 {
3084 	struct nfs_server *nfss = NFS_SERVER(data->args.inode);
3085 
3086 	if (nfss->pnfs_curr_ld->cleanup_layoutcommit)
3087 		nfss->pnfs_curr_ld->cleanup_layoutcommit(data);
3088 	pnfs_list_write_lseg_done(data->args.inode, &data->lseg_list);
3089 }
3090 
3091 /*
3092  * For the LAYOUT4_NFSV4_1_FILES layout type, NFS_DATA_SYNC WRITEs and
3093  * NFS_UNSTABLE WRITEs with a COMMIT to data servers must store enough
3094  * data to disk to allow the server to recover the data if it crashes.
3095  * LAYOUTCOMMIT is only needed when the NFL4_UFLG_COMMIT_THRU_MDS flag
3096  * is off, and a COMMIT is sent to a data server, or
3097  * if WRITEs to a data server return NFS_DATA_SYNC.
3098  */
3099 int
3100 pnfs_layoutcommit_inode(struct inode *inode, bool sync)
3101 {
3102 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3103 	struct nfs4_layoutcommit_data *data;
3104 	struct nfs_inode *nfsi = NFS_I(inode);
3105 	loff_t end_pos;
3106 	int status;
3107 
3108 	if (!pnfs_layoutcommit_outstanding(inode))
3109 		return 0;
3110 
3111 	dprintk("--> %s inode %lu\n", __func__, inode->i_ino);
3112 
3113 	status = -EAGAIN;
3114 	if (test_and_set_bit(NFS_INO_LAYOUTCOMMITTING, &nfsi->flags)) {
3115 		if (!sync)
3116 			goto out;
3117 		status = wait_on_bit_lock_action(&nfsi->flags,
3118 				NFS_INO_LAYOUTCOMMITTING,
3119 				nfs_wait_bit_killable,
3120 				TASK_KILLABLE);
3121 		if (status)
3122 			goto out;
3123 	}
3124 
3125 	status = -ENOMEM;
3126 	/* Note kzalloc ensures data->res.seq_res.sr_slot == NULL */
3127 	data = kzalloc(sizeof(*data), GFP_NOFS);
3128 	if (!data)
3129 		goto clear_layoutcommitting;
3130 
3131 	status = 0;
3132 	spin_lock(&inode->i_lock);
3133 	if (!test_and_clear_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags))
3134 		goto out_unlock;
3135 
3136 	INIT_LIST_HEAD(&data->lseg_list);
3137 	pnfs_list_write_lseg(inode, &data->lseg_list);
3138 
3139 	end_pos = nfsi->layout->plh_lwb;
3140 
3141 	nfs4_stateid_copy(&data->args.stateid, &nfsi->layout->plh_stateid);
3142 	data->cred = get_cred(nfsi->layout->plh_lc_cred);
3143 	spin_unlock(&inode->i_lock);
3144 
3145 	data->args.inode = inode;
3146 	nfs_fattr_init(&data->fattr);
3147 	data->args.bitmask = NFS_SERVER(inode)->cache_consistency_bitmask;
3148 	data->res.fattr = &data->fattr;
3149 	if (end_pos != 0)
3150 		data->args.lastbytewritten = end_pos - 1;
3151 	else
3152 		data->args.lastbytewritten = U64_MAX;
3153 	data->res.server = NFS_SERVER(inode);
3154 
3155 	if (ld->prepare_layoutcommit) {
3156 		status = ld->prepare_layoutcommit(&data->args);
3157 		if (status) {
3158 			put_cred(data->cred);
3159 			spin_lock(&inode->i_lock);
3160 			set_bit(NFS_INO_LAYOUTCOMMIT, &nfsi->flags);
3161 			if (end_pos > nfsi->layout->plh_lwb)
3162 				nfsi->layout->plh_lwb = end_pos;
3163 			goto out_unlock;
3164 		}
3165 	}
3166 
3167 
3168 	status = nfs4_proc_layoutcommit(data, sync);
3169 out:
3170 	if (status)
3171 		mark_inode_dirty_sync(inode);
3172 	dprintk("<-- %s status %d\n", __func__, status);
3173 	return status;
3174 out_unlock:
3175 	spin_unlock(&inode->i_lock);
3176 	kfree(data);
3177 clear_layoutcommitting:
3178 	pnfs_clear_layoutcommitting(inode);
3179 	goto out;
3180 }
3181 EXPORT_SYMBOL_GPL(pnfs_layoutcommit_inode);
3182 
3183 int
3184 pnfs_generic_sync(struct inode *inode, bool datasync)
3185 {
3186 	return pnfs_layoutcommit_inode(inode, true);
3187 }
3188 EXPORT_SYMBOL_GPL(pnfs_generic_sync);
3189 
3190 struct nfs4_threshold *pnfs_mdsthreshold_alloc(void)
3191 {
3192 	struct nfs4_threshold *thp;
3193 
3194 	thp = kzalloc(sizeof(*thp), GFP_NOFS);
3195 	if (!thp) {
3196 		dprintk("%s mdsthreshold allocation failed\n", __func__);
3197 		return NULL;
3198 	}
3199 	return thp;
3200 }
3201 
3202 #if IS_ENABLED(CONFIG_NFS_V4_2)
3203 int
3204 pnfs_report_layoutstat(struct inode *inode, gfp_t gfp_flags)
3205 {
3206 	struct pnfs_layoutdriver_type *ld = NFS_SERVER(inode)->pnfs_curr_ld;
3207 	struct nfs_server *server = NFS_SERVER(inode);
3208 	struct nfs_inode *nfsi = NFS_I(inode);
3209 	struct nfs42_layoutstat_data *data;
3210 	struct pnfs_layout_hdr *hdr;
3211 	int status = 0;
3212 
3213 	if (!pnfs_enabled_sb(server) || !ld->prepare_layoutstats)
3214 		goto out;
3215 
3216 	if (!nfs_server_capable(inode, NFS_CAP_LAYOUTSTATS))
3217 		goto out;
3218 
3219 	if (test_and_set_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags))
3220 		goto out;
3221 
3222 	spin_lock(&inode->i_lock);
3223 	if (!NFS_I(inode)->layout) {
3224 		spin_unlock(&inode->i_lock);
3225 		goto out_clear_layoutstats;
3226 	}
3227 	hdr = NFS_I(inode)->layout;
3228 	pnfs_get_layout_hdr(hdr);
3229 	spin_unlock(&inode->i_lock);
3230 
3231 	data = kzalloc(sizeof(*data), gfp_flags);
3232 	if (!data) {
3233 		status = -ENOMEM;
3234 		goto out_put;
3235 	}
3236 
3237 	data->args.fh = NFS_FH(inode);
3238 	data->args.inode = inode;
3239 	status = ld->prepare_layoutstats(&data->args);
3240 	if (status)
3241 		goto out_free;
3242 
3243 	status = nfs42_proc_layoutstats_generic(NFS_SERVER(inode), data);
3244 
3245 out:
3246 	dprintk("%s returns %d\n", __func__, status);
3247 	return status;
3248 
3249 out_free:
3250 	kfree(data);
3251 out_put:
3252 	pnfs_put_layout_hdr(hdr);
3253 out_clear_layoutstats:
3254 	smp_mb__before_atomic();
3255 	clear_bit(NFS_INO_LAYOUTSTATS, &nfsi->flags);
3256 	smp_mb__after_atomic();
3257 	goto out;
3258 }
3259 EXPORT_SYMBOL_GPL(pnfs_report_layoutstat);
3260 #endif
3261 
3262 unsigned int layoutstats_timer;
3263 module_param(layoutstats_timer, uint, 0644);
3264 EXPORT_SYMBOL_GPL(layoutstats_timer);
3265